diff --git a/llvm/include/llvm/CodeGen/TargetInstrInfo.h b/llvm/include/llvm/CodeGen/TargetInstrInfo.h
--- a/llvm/include/llvm/CodeGen/TargetInstrInfo.h
+++ b/llvm/include/llvm/CodeGen/TargetInstrInfo.h
@@ -1044,16 +1044,6 @@
return isCopyInstrImpl(MI);
}
- bool isFullCopyInstr(const MachineInstr &MI) const {
- auto DestSrc = isCopyInstr(MI);
- if (!DestSrc)
- return false;
-
- const MachineOperand *DestRegOp = DestSrc->Destination;
- const MachineOperand *SrcRegOp = DestSrc->Source;
- return !DestRegOp->getSubReg() && !SrcRegOp->getSubReg();
- }
-
/// If the specific machine instruction is an instruction that adds an
/// immediate value and a physical register, and stores the result in
/// the given physical register \c Reg, return a pair of the source
@@ -1968,13 +1958,6 @@
return false;
}
- /// Allows targets to use appropriate copy instruction while spilitting live
- /// range of a register in register allocation.
- virtual unsigned getLiveRangeSplitOpcode(Register Reg,
- const MachineFunction &MF) const {
- return TargetOpcode::COPY;
- }
-
/// During PHI eleimination lets target to make necessary checks and
/// insert the copy to the PHI destination register in a target specific
/// manner.
diff --git a/llvm/lib/CodeGen/CalcSpillWeights.cpp b/llvm/lib/CodeGen/CalcSpillWeights.cpp
--- a/llvm/lib/CodeGen/CalcSpillWeights.cpp
+++ b/llvm/lib/CodeGen/CalcSpillWeights.cpp
@@ -97,7 +97,7 @@
// Trace copies introduced by live range splitting. The inline
// spiller can rematerialize through these copies, so the spill
// weight must reflect this.
- while (TII.isFullCopyInstr(*MI)) {
+ while (MI->isFullCopy()) {
// The copy destination must match the interval register.
if (MI->getOperand(0).getReg() != Reg)
return false;
@@ -224,16 +224,7 @@
continue;
NumInstr++;
- bool identityCopy = false;
- auto DestSrc = TII.isCopyInstr(*MI);
- if (DestSrc) {
- const MachineOperand *DestRegOp = DestSrc->Destination;
- const MachineOperand *SrcRegOp = DestSrc->Source;
- identityCopy = DestRegOp->getReg() == SrcRegOp->getReg() &&
- DestRegOp->getSubReg() == SrcRegOp->getSubReg();
- }
-
- if (identityCopy || MI->isImplicitDef())
+ if (MI->isIdentityCopy() || MI->isImplicitDef())
continue;
if (!Visited.insert(MI).second)
continue;
@@ -267,7 +258,7 @@
}
// Get allocation hints from copies.
- if (!TII.isCopyInstr(*MI))
+ if (!MI->isCopy())
continue;
Register HintReg = copyHint(MI, LI.reg(), TRI, MRI);
if (!HintReg)
diff --git a/llvm/lib/CodeGen/InlineSpiller.cpp b/llvm/lib/CodeGen/InlineSpiller.cpp
--- a/llvm/lib/CodeGen/InlineSpiller.cpp
+++ b/llvm/lib/CodeGen/InlineSpiller.cpp
@@ -256,11 +256,11 @@
// This minimizes register pressure and maximizes the store-to-load distance for
// spill slots which can be important in tight loops.
-/// isFullCopyOf - If MI is a COPY to or from Reg, return the other register,
-/// otherwise return 0.
-static Register isCopyOf(const MachineInstr &MI, Register Reg,
- const TargetInstrInfo &TII) {
- if (!TII.isCopyInstr(MI))
+/// If MI is a COPY to or from Reg, return the other register, otherwise return
+/// 0.
+static Register isCopyOf(const MachineInstr &MI, Register Reg) {
+ assert(!MI.isBundled());
+ if (!MI.isCopy())
return Register();
const MachineOperand &DstOp = MI.getOperand(0);
@@ -277,10 +277,9 @@
}
/// Check for a copy bundle as formed by SplitKit.
-static Register isCopyOfBundle(const MachineInstr &FirstMI, Register Reg,
- const TargetInstrInfo &TII) {
+static Register isCopyOfBundle(const MachineInstr &FirstMI, Register Reg) {
if (!FirstMI.isBundled())
- return isCopyOf(FirstMI, Reg, TII);
+ return isCopyOf(FirstMI, Reg);
assert(!FirstMI.isBundledWithPred() && FirstMI.isBundledWithSucc() &&
"expected to see first instruction in bundle");
@@ -289,12 +288,11 @@
MachineBasicBlock::const_instr_iterator I = FirstMI.getIterator();
while (I->isBundledWithSucc()) {
const MachineInstr &MI = *I;
- auto CopyInst = TII.isCopyInstr(MI);
- if (!CopyInst)
+ if (!MI.isCopy())
return Register();
- const MachineOperand &DstOp = *CopyInst->Destination;
- const MachineOperand &SrcOp = *CopyInst->Source;
+ const MachineOperand &DstOp = MI.getOperand(0);
+ const MachineOperand &SrcOp = MI.getOperand(1);
if (DstOp.getReg() == Reg) {
if (!SnipReg)
SnipReg = SrcOp.getReg();
@@ -360,7 +358,7 @@
MachineInstr &MI = *RI++;
// Allow copies to/from Reg.
- if (isCopyOfBundle(MI, Reg, TII))
+ if (isCopyOfBundle(MI, Reg))
continue;
// Allow stack slot loads.
@@ -398,7 +396,7 @@
return;
for (MachineInstr &MI : llvm::make_early_inc_range(MRI.reg_bundles(Reg))) {
- Register SnipReg = isCopyOfBundle(MI, Reg, TII);
+ Register SnipReg = isCopyOfBundle(MI, Reg);
if (!isSibling(SnipReg))
continue;
LiveInterval &SnipLI = LIS.getInterval(SnipReg);
@@ -521,14 +519,14 @@
// Find all spills and copies of VNI.
for (MachineInstr &MI :
llvm::make_early_inc_range(MRI.use_nodbg_bundles(Reg))) {
- if (!MI.mayStore() && !TII.isCopyInstr(MI))
+ if (!MI.isCopy() && !MI.mayStore())
continue;
SlotIndex Idx = LIS.getInstructionIndex(MI);
if (LI->getVNInfoAt(Idx) != VNI)
continue;
// Follow sibling copies down the dominator tree.
- if (Register DstReg = isCopyOfBundle(MI, Reg, TII)) {
+ if (Register DstReg = isCopyOfBundle(MI, Reg)) {
if (isSibling(DstReg)) {
LiveInterval &DstLI = LIS.getInterval(DstReg);
VNInfo *DstVNI = DstLI.getVNInfoAt(Idx.getRegSlot());
@@ -872,7 +870,7 @@
if (Ops.back().first != MI || MI->isBundled())
return false;
- bool WasCopy = TII.isCopyInstr(*MI).has_value();
+ bool WasCopy = MI->isCopy();
Register ImpReg;
// TII::foldMemoryOperand will do what we need here for statepoint
@@ -1157,7 +1155,7 @@
Idx = VNI->def;
// Check for a sibling copy.
- Register SibReg = isCopyOfBundle(MI, Reg, TII);
+ Register SibReg = isCopyOfBundle(MI, Reg);
if (SibReg && isSibling(SibReg)) {
// This may actually be a copy between snippets.
if (isRegToSpill(SibReg)) {
diff --git a/llvm/lib/CodeGen/LiveRangeEdit.cpp b/llvm/lib/CodeGen/LiveRangeEdit.cpp
--- a/llvm/lib/CodeGen/LiveRangeEdit.cpp
+++ b/llvm/lib/CodeGen/LiveRangeEdit.cpp
@@ -352,8 +352,7 @@
// unlikely to change anything. We typically don't want to shrink the
// PIC base register that has lots of uses everywhere.
// Always shrink COPY uses that probably come from live range splitting.
- if ((MI->readsVirtualRegister(Reg) &&
- (MO.isDef() || TII.isCopyInstr(*MI))) ||
+ if ((MI->readsVirtualRegister(Reg) && (MI->isCopy() || MO.isDef())) ||
(MO.readsReg() && (MRI.hasOneNonDBGUse(Reg) || useIsKill(LI, MO))))
ToShrink.insert(&LI);
else if (MO.readsReg())
diff --git a/llvm/lib/CodeGen/LiveRangeShrink.cpp b/llvm/lib/CodeGen/LiveRangeShrink.cpp
--- a/llvm/lib/CodeGen/LiveRangeShrink.cpp
+++ b/llvm/lib/CodeGen/LiveRangeShrink.cpp
@@ -23,7 +23,6 @@
#include "llvm/CodeGen/MachineInstr.h"
#include "llvm/CodeGen/MachineOperand.h"
#include "llvm/CodeGen/MachineRegisterInfo.h"
-#include "llvm/CodeGen/TargetInstrInfo.h"
#include "llvm/InitializePasses.h"
#include "llvm/Pass.h"
#include "llvm/Support/Debug.h"
@@ -110,7 +109,6 @@
return false;
MachineRegisterInfo &MRI = MF.getRegInfo();
- const TargetInstrInfo &TII = *MF.getSubtarget().getInstrInfo();
LLVM_DEBUG(dbgs() << "**** Analysing " << MF.getName() << '\n');
@@ -199,7 +197,7 @@
// is because it needs more accurate model to handle register
// pressure correctly.
MachineInstr &DefInstr = *MRI.def_instr_begin(Reg);
- if (!TII.isCopyInstr(DefInstr))
+ if (!DefInstr.isCopy())
NumEligibleUse++;
Insert = FindDominatedInstruction(DefInstr, Insert, IOM);
} else {
diff --git a/llvm/lib/CodeGen/RegAllocGreedy.cpp b/llvm/lib/CodeGen/RegAllocGreedy.cpp
--- a/llvm/lib/CodeGen/RegAllocGreedy.cpp
+++ b/llvm/lib/CodeGen/RegAllocGreedy.cpp
@@ -1282,12 +1282,10 @@
/// VirtReg.
static bool readsLaneSubset(const MachineRegisterInfo &MRI,
const MachineInstr *MI, const LiveInterval &VirtReg,
- const TargetRegisterInfo *TRI, SlotIndex Use,
- const TargetInstrInfo *TII) {
+ const TargetRegisterInfo *TRI, SlotIndex Use) {
// Early check the common case.
- auto DestSrc = TII->isCopyInstr(*MI);
- if (DestSrc &&
- DestSrc->Destination->getSubReg() == DestSrc->Source->getSubReg())
+ if (MI->isCopy() &&
+ MI->getOperand(0).getSubReg() == MI->getOperand(1).getSubReg())
return false;
// FIXME: We're only considering uses, but should be consider defs too?
@@ -1346,14 +1344,14 @@
// the allocation.
for (const SlotIndex Use : Uses) {
if (const MachineInstr *MI = Indexes->getInstructionFromIndex(Use)) {
- if (TII->isFullCopyInstr(*MI) ||
+ if (MI->isFullCopy() ||
(SplitSubClass &&
SuperRCNumAllocatableRegs ==
getNumAllocatableRegsForConstraints(MI, VirtReg.reg(), SuperRC,
TII, TRI, RegClassInfo)) ||
// TODO: Handle split for subranges with subclass constraints?
(!SplitSubClass && VirtReg.hasSubRanges() &&
- !readsLaneSubset(*MRI, MI, VirtReg, TRI, Use, TII))) {
+ !readsLaneSubset(*MRI, MI, VirtReg, TRI, Use))) {
LLVM_DEBUG(dbgs() << " skip:\t" << Use << '\t' << *MI);
continue;
}
@@ -2140,7 +2138,7 @@
/// \p Out is not cleared before being populated.
void RAGreedy::collectHintInfo(Register Reg, HintsInfo &Out) {
for (const MachineInstr &Instr : MRI->reg_nodbg_instructions(Reg)) {
- if (!TII->isFullCopyInstr(Instr))
+ if (!Instr.isFullCopy())
continue;
// Look for the other end of the copy.
Register OtherReg = Instr.getOperand(0).getReg();
@@ -2455,10 +2453,9 @@
MI.getOpcode() == TargetOpcode::STATEPOINT;
};
for (MachineInstr &MI : MBB) {
- auto DestSrc = TII->isCopyInstr(MI);
- if (DestSrc) {
- const MachineOperand &Dest = *DestSrc->Destination;
- const MachineOperand &Src = *DestSrc->Source;
+ if (MI.isCopy()) {
+ const MachineOperand &Dest = MI.getOperand(0);
+ const MachineOperand &Src = MI.getOperand(1);
Register SrcReg = Src.getReg();
Register DestReg = Dest.getReg();
// Only count `COPY`s with a virtual register as source or destination.
diff --git a/llvm/lib/CodeGen/SplitKit.h b/llvm/lib/CodeGen/SplitKit.h
--- a/llvm/lib/CodeGen/SplitKit.h
+++ b/llvm/lib/CodeGen/SplitKit.h
@@ -428,11 +428,8 @@
bool Late, unsigned RegIdx);
SlotIndex buildSingleSubRegCopy(Register FromReg, Register ToReg,
- MachineBasicBlock &MB,
- MachineBasicBlock::iterator InsertBefore,
- unsigned SubIdx, LiveInterval &DestLI,
- bool Late, SlotIndex Def,
- const MCInstrDesc &Desc);
+ MachineBasicBlock &MB, MachineBasicBlock::iterator InsertBefore,
+ unsigned SubIdx, LiveInterval &DestLI, bool Late, SlotIndex Def);
public:
/// Create a new SplitEditor for editing the LiveInterval analyzed by SA.
diff --git a/llvm/lib/CodeGen/SplitKit.cpp b/llvm/lib/CodeGen/SplitKit.cpp
--- a/llvm/lib/CodeGen/SplitKit.cpp
+++ b/llvm/lib/CodeGen/SplitKit.cpp
@@ -514,10 +514,10 @@
VFP = ValueForcePair(nullptr, true);
}
-SlotIndex SplitEditor::buildSingleSubRegCopy(
- Register FromReg, Register ToReg, MachineBasicBlock &MBB,
- MachineBasicBlock::iterator InsertBefore, unsigned SubIdx,
- LiveInterval &DestLI, bool Late, SlotIndex Def, const MCInstrDesc &Desc) {
+SlotIndex SplitEditor::buildSingleSubRegCopy(Register FromReg, Register ToReg,
+ MachineBasicBlock &MBB, MachineBasicBlock::iterator InsertBefore,
+ unsigned SubIdx, LiveInterval &DestLI, bool Late, SlotIndex Def) {
+ const MCInstrDesc &Desc = TII.get(TargetOpcode::COPY);
bool FirstCopy = !Def.isValid();
MachineInstr *CopyMI = BuildMI(MBB, InsertBefore, DebugLoc(), Desc)
.addReg(ToReg, RegState::Define | getUndefRegState(FirstCopy)
@@ -536,8 +536,7 @@
SlotIndex SplitEditor::buildCopy(Register FromReg, Register ToReg,
LaneBitmask LaneMask, MachineBasicBlock &MBB,
MachineBasicBlock::iterator InsertBefore, bool Late, unsigned RegIdx) {
- const MCInstrDesc &Desc =
- TII.get(TII.getLiveRangeSplitOpcode(FromReg, *MBB.getParent()));
+ const MCInstrDesc &Desc = TII.get(TargetOpcode::COPY);
SlotIndexes &Indexes = *LIS.getSlotIndexes();
if (LaneMask.all() || LaneMask == MRI.getMaxLaneMaskForVReg(FromReg)) {
// The full vreg is copied.
@@ -565,7 +564,7 @@
SlotIndex Def;
for (unsigned BestIdx : SubIndexes) {
Def = buildSingleSubRegCopy(FromReg, ToReg, MBB, InsertBefore, BestIdx,
- DestLI, Late, Def, Desc);
+ DestLI, Late, Def);
}
BumpPtrAllocator &Allocator = LIS.getVNInfoAllocator();
@@ -1585,9 +1584,7 @@
if (BI.LiveIn && BI.LiveOut)
return true;
// No point in isolating a copy. It has no register class constraints.
- MachineInstr *MI = LIS.getInstructionFromIndex(BI.FirstInstr);
- bool copyLike = TII.isCopyInstr(*MI) || MI->isSubregToReg();
- if (copyLike)
+ if (LIS.getInstructionFromIndex(BI.FirstInstr)->isCopyLike())
return false;
// Finally, don't isolate an end point that was created by earlier splits.
return isOriginalEndpoint(BI.FirstInstr);
diff --git a/llvm/lib/CodeGen/TargetInstrInfo.cpp b/llvm/lib/CodeGen/TargetInstrInfo.cpp
--- a/llvm/lib/CodeGen/TargetInstrInfo.cpp
+++ b/llvm/lib/CodeGen/TargetInstrInfo.cpp
@@ -440,9 +440,8 @@
// If the COPY instruction in MI can be folded to a stack operation, return
// the register class to use.
static const TargetRegisterClass *canFoldCopy(const MachineInstr &MI,
- const TargetInstrInfo &TII,
unsigned FoldIdx) {
- assert(TII.isCopyInstr(MI) && "MI must be a COPY instruction");
+ assert(MI.isCopy() && "MI must be a COPY instruction");
if (MI.getNumOperands() != 2)
return nullptr;
assert(FoldIdx<2 && "FoldIdx refers no nonexistent operand");
@@ -631,10 +630,10 @@
}
// Straight COPY may fold as load/store.
- if (!isCopyInstr(MI) || Ops.size() != 1)
+ if (!MI.isCopy() || Ops.size() != 1)
return nullptr;
- const TargetRegisterClass *RC = canFoldCopy(MI, *this, Ops[0]);
+ const TargetRegisterClass *RC = canFoldCopy(MI, Ops[0]);
if (!RC)
return nullptr;
diff --git a/llvm/lib/Target/AMDGPU/AMDGPU.h b/llvm/lib/Target/AMDGPU/AMDGPU.h
--- a/llvm/lib/Target/AMDGPU/AMDGPU.h
+++ b/llvm/lib/Target/AMDGPU/AMDGPU.h
@@ -41,7 +41,6 @@
FunctionPass *createSIOptimizeExecMaskingPreRAPass();
FunctionPass *createSIOptimizeVGPRLiveRangePass();
FunctionPass *createSIFixSGPRCopiesPass();
-FunctionPass *createLowerWWMCopiesPass();
FunctionPass *createSIMemoryLegalizerPass();
FunctionPass *createSIInsertWaitcntsPass();
FunctionPass *createSIPreAllocateWWMRegsPass();
@@ -145,9 +144,6 @@
void initializeSIFixVGPRCopiesPass(PassRegistry &);
extern char &SIFixVGPRCopiesID;
-void initializeSILowerWWMCopiesPass(PassRegistry &);
-extern char &SILowerWWMCopiesID;
-
void initializeSILowerI1CopiesPass(PassRegistry &);
extern char &SILowerI1CopiesID;
diff --git a/llvm/lib/Target/AMDGPU/AMDGPUTargetMachine.cpp b/llvm/lib/Target/AMDGPU/AMDGPUTargetMachine.cpp
--- a/llvm/lib/Target/AMDGPU/AMDGPUTargetMachine.cpp
+++ b/llvm/lib/Target/AMDGPU/AMDGPUTargetMachine.cpp
@@ -364,7 +364,6 @@
initializeAMDGPUDAGToDAGISelPass(*PR);
initializeGCNDPPCombinePass(*PR);
initializeSILowerI1CopiesPass(*PR);
- initializeSILowerWWMCopiesPass(*PR);
initializeSILowerSGPRSpillsPass(*PR);
initializeSIFixSGPRCopiesPass(*PR);
initializeSIFixVGPRCopiesPass(*PR);
@@ -1297,7 +1296,6 @@
}
bool GCNPassConfig::addPreRewrite() {
- addPass(&SILowerWWMCopiesID);
if (EnableRegReassign)
addPass(&GCNNSAReassignID);
return true;
@@ -1352,8 +1350,6 @@
addPass(&SILowerSGPRSpillsID);
addPass(createVGPRAllocPass(false));
-
- addPass(&SILowerWWMCopiesID);
return true;
}
diff --git a/llvm/lib/Target/AMDGPU/CMakeLists.txt b/llvm/lib/Target/AMDGPU/CMakeLists.txt
--- a/llvm/lib/Target/AMDGPU/CMakeLists.txt
+++ b/llvm/lib/Target/AMDGPU/CMakeLists.txt
@@ -145,7 +145,6 @@
SILoadStoreOptimizer.cpp
SILowerControlFlow.cpp
SILowerI1Copies.cpp
- SILowerWWMCopies.cpp
SILowerSGPRSpills.cpp
SIMachineFunctionInfo.cpp
SIMachineScheduler.cpp
diff --git a/llvm/lib/Target/AMDGPU/SIFrameLowering.cpp b/llvm/lib/Target/AMDGPU/SIFrameLowering.cpp
--- a/llvm/lib/Target/AMDGPU/SIFrameLowering.cpp
+++ b/llvm/lib/Target/AMDGPU/SIFrameLowering.cpp
@@ -99,10 +99,10 @@
SGPR, PrologEpilogSGPRSaveRestoreInfo(
SGPRSaveKind::SPILL_TO_VGPR_LANE, FI));
- LLVM_DEBUG(auto Spill = MFI->getSGPRSpillToPhysicalVGPRLanes(FI).front();
- dbgs() << printReg(SGPR, TRI) << " requires fallback spill to "
- << printReg(Spill.VGPR, TRI) << ':' << Spill.Lane
- << '\n';);
+ LLVM_DEBUG(
+ auto Spill = MFI->getPrologEpilogSGPRSpillToVGPRLanes(FI).front();
+ dbgs() << printReg(SGPR, TRI) << " requires fallback spill to "
+ << printReg(Spill.VGPR, TRI) << ':' << Spill.Lane << '\n';);
} else {
// Remove dead <FI> index
MF.getFrameInfo().RemoveStackObject(FI);
@@ -264,7 +264,7 @@
assert(MFI.getStackID(FI) == TargetStackID::SGPRSpill);
ArrayRef<SIRegisterInfo::SpilledReg> Spill =
- FuncInfo->getSGPRSpillToPhysicalVGPRLanes(FI);
+ FuncInfo->getPrologEpilogSGPRSpillToVGPRLanes(FI);
assert(Spill.size() == NumSubRegs);
for (unsigned I = 0; I < NumSubRegs; ++I) {
@@ -309,7 +309,7 @@
void restoreFromVGPRLane(const int FI) {
assert(MFI.getStackID(FI) == TargetStackID::SGPRSpill);
ArrayRef<SIRegisterInfo::SpilledReg> Spill =
- FuncInfo->getSGPRSpillToPhysicalVGPRLanes(FI);
+ FuncInfo->getPrologEpilogSGPRSpillToVGPRLanes(FI);
assert(Spill.size() == NumSubRegs);
for (unsigned I = 0; I < NumSubRegs; ++I) {
@@ -1353,8 +1353,8 @@
if (FuncInfo->allocateVGPRSpillToAGPR(MF, FI,
TRI->isAGPR(MRI, VReg))) {
assert(RS != nullptr);
- RS->enterBasicBlockEnd(MBB);
- RS->backward(MI);
+ // FIXME: change to enterBasicBlockEnd()
+ RS->enterBasicBlock(MBB);
TRI->eliminateFrameIndex(MI, 0, FIOp, RS);
SpillFIs.set(FI);
continue;
diff --git a/llvm/lib/Target/AMDGPU/SIInstrInfo.h b/llvm/lib/Target/AMDGPU/SIInstrInfo.h
--- a/llvm/lib/Target/AMDGPU/SIInstrInfo.h
+++ b/llvm/lib/Target/AMDGPU/SIInstrInfo.h
@@ -170,12 +170,6 @@
Register findUsedSGPR(const MachineInstr &MI, int OpIndices[3]) const;
protected:
- /// If the specific machine instruction is a instruction that moves/copies
- /// value from one register to another register return destination and source
- /// registers as machine operands.
- std::optional<DestSourcePair>
- isCopyInstrImpl(const MachineInstr &MI) const override;
-
bool swapSourceModifiers(MachineInstr &MI,
MachineOperand &Src0, unsigned Src0OpName,
MachineOperand &Src1, unsigned Src1OpName) const;
@@ -833,7 +827,7 @@
}
bool isVGPRCopy(const MachineInstr &MI) const {
- assert(isCopyInstr(MI));
+ assert(MI.isCopy());
Register Dest = MI.getOperand(0).getReg();
const MachineFunction &MF = *MI.getParent()->getParent();
const MachineRegisterInfo &MRI = MF.getRegInfo();
@@ -903,7 +897,7 @@
if (OpIdx >= MI.getDesc().NumOperands)
return false;
- if (isCopyInstr(MI)) {
+ if (MI.isCopy()) {
unsigned Size = getOpSize(MI, OpIdx);
assert(Size == 8 || Size == 4);
@@ -952,12 +946,12 @@
void insertScratchExecCopy(MachineFunction &MF, MachineBasicBlock &MBB,
MachineBasicBlock::iterator MBBI,
- const DebugLoc &DL, Register Reg, bool IsSCCLive,
- SlotIndexes *Indexes = nullptr) const;
+ const DebugLoc &DL, Register Reg,
+ bool IsSCCLive) const;
void restoreExec(MachineFunction &MF, MachineBasicBlock &MBB,
MachineBasicBlock::iterator MBBI, const DebugLoc &DL,
- Register Reg, SlotIndexes *Indexes = nullptr) const;
+ Register Reg) const;
/// Return the correct register class for \p OpNo. For target-specific
/// instructions, this will return the register class that has been defined
@@ -1149,9 +1143,6 @@
CreateTargetMIHazardRecognizer(const InstrItineraryData *II,
const ScheduleDAGMI *DAG) const override;
- unsigned getLiveRangeSplitOpcode(Register Reg,
- const MachineFunction &MF) const override;
-
bool isBasicBlockPrologue(const MachineInstr &MI) const override;
MachineInstr *createPHIDestinationCopy(MachineBasicBlock &MBB,
diff --git a/llvm/lib/Target/AMDGPU/SIInstrInfo.cpp b/llvm/lib/Target/AMDGPU/SIInstrInfo.cpp
--- a/llvm/lib/Target/AMDGPU/SIInstrInfo.cpp
+++ b/llvm/lib/Target/AMDGPU/SIInstrInfo.cpp
@@ -2414,14 +2414,6 @@
return std::pair(Split[0], Split[1]);
}
-std::optional<DestSourcePair>
-SIInstrInfo::isCopyInstrImpl(const MachineInstr &MI) const {
- if (MI.getOpcode() == AMDGPU::WWM_COPY)
- return DestSourcePair{MI.getOperand(0), MI.getOperand(1)};
-
- return std::nullopt;
-}
-
bool SIInstrInfo::swapSourceModifiers(MachineInstr &MI,
MachineOperand &Src0,
unsigned Src0OpName,
@@ -3088,7 +3080,6 @@
case AMDGPU::S_MOV_B32:
case AMDGPU::S_MOV_B64:
case AMDGPU::COPY:
- case AMDGPU::WWM_COPY:
case AMDGPU::V_ACCVGPR_WRITE_B32_e64:
case AMDGPU::V_ACCVGPR_READ_B32_e64:
case AMDGPU::V_ACCVGPR_MOV_B32:
@@ -4978,8 +4969,7 @@
MachineBasicBlock &MBB,
MachineBasicBlock::iterator MBBI,
const DebugLoc &DL, Register Reg,
- bool IsSCCLive,
- SlotIndexes *Indexes) const {
+ bool IsSCCLive) const {
const GCNSubtarget &ST = MF.getSubtarget<GCNSubtarget>();
const SIInstrInfo *TII = ST.getInstrInfo();
bool IsWave32 = ST.isWave32();
@@ -4989,34 +4979,23 @@
// the single instruction S_OR_SAVEEXEC that clobbers SCC.
unsigned MovOpc = IsWave32 ? AMDGPU::S_MOV_B32 : AMDGPU::S_MOV_B64;
MCRegister Exec = IsWave32 ? AMDGPU::EXEC_LO : AMDGPU::EXEC;
- auto StoreExecMI = BuildMI(MBB, MBBI, DL, TII->get(MovOpc), Reg)
- .addReg(Exec, RegState::Kill);
- auto FlipExecMI = BuildMI(MBB, MBBI, DL, TII->get(MovOpc), Exec).addImm(-1);
- if (Indexes) {
- Indexes->insertMachineInstrInMaps(*StoreExecMI);
- Indexes->insertMachineInstrInMaps(*FlipExecMI);
- }
+ BuildMI(MBB, MBBI, DL, TII->get(MovOpc), Reg).addReg(Exec, RegState::Kill);
+ BuildMI(MBB, MBBI, DL, TII->get(MovOpc), Exec).addImm(-1);
} else {
const unsigned OrSaveExec =
IsWave32 ? AMDGPU::S_OR_SAVEEXEC_B32 : AMDGPU::S_OR_SAVEEXEC_B64;
auto SaveExec =
BuildMI(MBB, MBBI, DL, TII->get(OrSaveExec), Reg).addImm(-1);
SaveExec->getOperand(3).setIsDead(); // Mark SCC as dead.
- if (Indexes)
- Indexes->insertMachineInstrInMaps(*SaveExec);
}
}
void SIInstrInfo::restoreExec(MachineFunction &MF, MachineBasicBlock &MBB,
MachineBasicBlock::iterator MBBI,
- const DebugLoc &DL, Register Reg,
- SlotIndexes *Indexes) const {
+ const DebugLoc &DL, Register Reg) const {
unsigned ExecMov = isWave32() ? AMDGPU::S_MOV_B32 : AMDGPU::S_MOV_B64;
MCRegister Exec = isWave32() ? AMDGPU::EXEC_LO : AMDGPU::EXEC;
- auto ExecRestoreMI =
- BuildMI(MBB, MBBI, DL, get(ExecMov), Exec).addReg(Reg, RegState::Kill);
- if (Indexes)
- Indexes->insertMachineInstrInMaps(*ExecRestoreMI);
+ BuildMI(MBB, MBBI, DL, get(ExecMov), Exec).addReg(Reg, RegState::Kill);
}
static const TargetRegisterClass *
@@ -8001,16 +7980,6 @@
return ArrayRef(TargetFlags);
}
-unsigned SIInstrInfo::getLiveRangeSplitOpcode(Register SrcReg,
- const MachineFunction &MF) const {
- const SIMachineFunctionInfo *MFI = MF.getInfo<SIMachineFunctionInfo>();
- assert(SrcReg.isVirtual());
- if (MFI->checkFlag(SrcReg, AMDGPU::VirtRegFlag::WWM_REG))
- return AMDGPU::WWM_COPY;
-
- return AMDGPU::COPY;
-}
-
bool SIInstrInfo::isBasicBlockPrologue(const MachineInstr &MI) const {
return !MI.isTerminator() && MI.getOpcode() != AMDGPU::COPY &&
MI.modifiesRegister(AMDGPU::EXEC, &RI);
@@ -8562,7 +8531,7 @@
// A similar issue also exists with spilling and reloading $exec registers.
//
// To prevent that, constrain the %0 register class here.
- if (isFullCopyInstr(MI)) {
+ if (MI.isFullCopy()) {
Register DstReg = MI.getOperand(0).getReg();
Register SrcReg = MI.getOperand(1).getReg();
if ((DstReg.isVirtual() || SrcReg.isVirtual()) &&
@@ -8659,7 +8628,7 @@
if (opcode == AMDGPU::V_READLANE_B32 || opcode == AMDGPU::V_READFIRSTLANE_B32)
return InstructionUniformity::AlwaysUniform;
- if (isCopyInstr(MI)) {
+ if (MI.isCopy()) {
const MachineOperand &srcOp = MI.getOperand(1);
if (srcOp.isReg() && srcOp.getReg().isPhysical()) {
const TargetRegisterClass *regClass =
diff --git a/llvm/lib/Target/AMDGPU/SIInstructions.td b/llvm/lib/Target/AMDGPU/SIInstructions.td
--- a/llvm/lib/Target/AMDGPU/SIInstructions.td
+++ b/llvm/lib/Target/AMDGPU/SIInstructions.td
@@ -172,13 +172,6 @@
} // End let hasSideEffects = 0, mayLoad = 0, mayStore = 0, Uses = [EXEC]
-def WWM_COPY : SPseudoInstSI <
- (outs unknown:$dst), (ins unknown:$src)> {
- let hasSideEffects = 0;
- let isAsCheapAsAMove = 1;
- let isConvergent = 1;
-}
-
def ENTER_STRICT_WWM : SPseudoInstSI <(outs SReg_1:$sdst), (ins i64imm:$src0)> {
let Uses = [EXEC];
let Defs = [EXEC, SCC];
diff --git a/llvm/lib/Target/AMDGPU/SILowerSGPRSpills.cpp b/llvm/lib/Target/AMDGPU/SILowerSGPRSpills.cpp
--- a/llvm/lib/Target/AMDGPU/SILowerSGPRSpills.cpp
+++ b/llvm/lib/Target/AMDGPU/SILowerSGPRSpills.cpp
@@ -50,9 +50,7 @@
SILowerSGPRSpills() : MachineFunctionPass(ID) {}
void calculateSaveRestoreBlocks(MachineFunction &MF);
- bool spillCalleeSavedRegs(MachineFunction &MF,
- SmallVectorImpl<int> &CalleeSavedFIs);
- void extendWWMVirtRegLiveness(MachineFunction &MF, LiveIntervals *LIS);
+ bool spillCalleeSavedRegs(MachineFunction &MF);
bool runOnMachineFunction(MachineFunction &MF) override;
@@ -60,13 +58,6 @@
AU.setPreservesAll();
MachineFunctionPass::getAnalysisUsage(AU);
}
-
- MachineFunctionProperties getClearedProperties() const override {
- // SILowerSGPRSpills introduces new Virtual VGPRs for spilling SGPRs.
- return MachineFunctionProperties()
- .set(MachineFunctionProperties::Property::IsSSA)
- .set(MachineFunctionProperties::Property::NoVRegs);
- }
};
} // end anonymous namespace
@@ -206,8 +197,7 @@
EntryBB.sortUniqueLiveIns();
}
-bool SILowerSGPRSpills::spillCalleeSavedRegs(
- MachineFunction &MF, SmallVectorImpl<int> &CalleeSavedFIs) {
+bool SILowerSGPRSpills::spillCalleeSavedRegs(MachineFunction &MF) {
MachineRegisterInfo &MRI = MF.getRegInfo();
const Function &F = MF.getFunction();
const GCNSubtarget &ST = MF.getSubtarget<GCNSubtarget>();
@@ -238,7 +228,6 @@
TRI->getSpillAlign(*RC), true);
CSI.push_back(CalleeSavedInfo(Reg, JunkFI));
- CalleeSavedFIs.push_back(JunkFI);
}
}
@@ -259,50 +248,6 @@
return false;
}
-void SILowerSGPRSpills::extendWWMVirtRegLiveness(MachineFunction &MF,
- LiveIntervals *LIS) {
- // TODO: This is a workaround to avoid the unmodelled liveness computed with
- // whole-wave virtual registers when allocated together with the regular VGPR
- // virtual registers. Presently, the liveness computed during the regalloc is
- // only uniform (or single lane aware) and it doesn't take account of the
- // divergent control flow that exists for our GPUs. Since the WWM registers
- // can modify inactive lanes, the wave-aware liveness should be computed for
- // the virtual registers to accurately plot their interferences. Without
- // having the divergent CFG for the function, it is difficult to implement the
- // wave-aware liveness info. Until then, we conservatively extend the liveness
- // of the wwm registers into the entire function so that they won't be reused
- // without first spilling/splitting their liveranges.
- SIMachineFunctionInfo *MFI = MF.getInfo<SIMachineFunctionInfo>();
-
- // Insert the IMPLICIT_DEF for the wwm-registers in the entry blocks.
- for (auto Reg : MFI->getSGPRSpillVGPRs()) {
- for (MachineBasicBlock *SaveBlock : SaveBlocks) {
- MachineBasicBlock::iterator InsertBefore = SaveBlock->begin();
- auto MIB = BuildMI(*SaveBlock, *InsertBefore, InsertBefore->getDebugLoc(),
- TII->get(AMDGPU::IMPLICIT_DEF), Reg);
- MFI->setFlag(Reg, AMDGPU::VirtRegFlag::WWM_REG);
- if (LIS) {
- LIS->InsertMachineInstrInMaps(*MIB);
- }
- }
- }
-
- // Insert the KILL in the return blocks to extend their liveness untill the
- // end of function. Insert a separate KILL for each VGPR.
- for (MachineBasicBlock *RestoreBlock : RestoreBlocks) {
- MachineBasicBlock::iterator InsertBefore =
- RestoreBlock->getFirstTerminator();
- for (auto Reg : MFI->getSGPRSpillVGPRs()) {
- auto MIB =
- BuildMI(*RestoreBlock, *InsertBefore, InsertBefore->getDebugLoc(),
- TII->get(TargetOpcode::KILL));
- MIB.addReg(Reg);
- if (LIS)
- LIS->InsertMachineInstrInMaps(*MIB);
- }
- }
-}
-
bool SILowerSGPRSpills::runOnMachineFunction(MachineFunction &MF) {
const GCNSubtarget &ST = MF.getSubtarget<GCNSubtarget>();
TII = ST.getInstrInfo();
@@ -316,8 +261,7 @@
// First, expose any CSR SGPR spills. This is mostly the same as what PEI
// does, but somewhat simpler.
calculateSaveRestoreBlocks(MF);
- SmallVector<int> CalleeSavedFIs;
- bool HasCSRs = spillCalleeSavedRegs(MF, CalleeSavedFIs);
+ bool HasCSRs = spillCalleeSavedRegs(MF);
MachineFrameInfo &MFI = MF.getFrameInfo();
MachineRegisterInfo &MRI = MF.getRegInfo();
@@ -331,7 +275,6 @@
bool MadeChange = false;
bool NewReservedRegs = false;
- bool SpilledToVirtVGPRLanes = false;
// TODO: CSR VGPRs will never be spilled to AGPRs. These can probably be
// handled as SpilledToReg in regular PrologEpilogInserter.
@@ -354,53 +297,23 @@
int FI = TII->getNamedOperand(MI, AMDGPU::OpName::addr)->getIndex();
assert(MFI.getStackID(FI) == TargetStackID::SGPRSpill);
-
- bool IsCalleeSaveSGPRSpill =
- std::find(CalleeSavedFIs.begin(), CalleeSavedFIs.end(), FI) !=
- CalleeSavedFIs.end();
- if (IsCalleeSaveSGPRSpill) {
- // Spill callee-saved SGPRs into physical VGPR lanes.
-
- // TODO: This is to ensure the CFIs are static for efficient frame
- // unwinding in the debugger. Spilling them into virtual VGPR lanes
- // involve regalloc to allocate the physical VGPRs and that might
- // cause intermediate spill/split of such liveranges for successful
- // allocation. This would result in broken CFI encoding unless the
- // regalloc aware CFI generation to insert new CFIs along with the
- // intermediate spills is implemented. There is no such support
- // currently exist in the LLVM compiler.
- if (FuncInfo->allocateSGPRSpillToVGPRLane(MF, FI, true)) {
- NewReservedRegs = true;
- bool Spilled = TRI->eliminateSGPRToVGPRSpillFrameIndex(
- MI, FI, nullptr, Indexes, LIS, true);
- if (!Spilled)
- llvm_unreachable(
- "failed to spill SGPR to physical VGPR lane when allocated");
- }
- } else {
- if (FuncInfo->allocateSGPRSpillToVGPRLane(MF, FI)) {
- bool Spilled = TRI->eliminateSGPRToVGPRSpillFrameIndex(
- MI, FI, nullptr, Indexes, LIS);
- if (!Spilled)
- llvm_unreachable(
- "failed to spill SGPR to virtual VGPR lane when allocated");
- SpillFIs.set(FI);
- SpilledToVirtVGPRLanes = true;
- }
+ if (FuncInfo->allocateSGPRSpillToVGPRLane(MF, FI)) {
+ NewReservedRegs = true;
+ bool Spilled = TRI->eliminateSGPRToVGPRSpillFrameIndex(
+ MI, FI, nullptr, Indexes, LIS);
+ (void)Spilled;
+ assert(Spilled && "failed to spill SGPR to VGPR when allocated");
+ SpillFIs.set(FI);
}
}
}
- if (SpilledToVirtVGPRLanes) {
- extendWWMVirtRegLiveness(MF, LIS);
- if (LIS) {
- // Compute the LiveInterval for the newly created virtual registers.
- for (auto Reg : FuncInfo->getSGPRSpillVGPRs())
- LIS->createAndComputeVirtRegInterval(Reg);
- }
- }
-
+ // FIXME: Adding to live-ins redundant with reserving registers.
for (MachineBasicBlock &MBB : MF) {
+ for (auto Reg : FuncInfo->getSGPRSpillVGPRs())
+ MBB.addLiveIn(Reg);
+ MBB.sortUniqueLiveIns();
+
// FIXME: The dead frame indices are replaced with a null register from
// the debug value instructions. We should instead, update it with the
// correct register value. But not sure the register value alone is
@@ -421,10 +334,6 @@
// lane".
FuncInfo->removeDeadFrameIndices(MFI, /*ResetSGPRSpillStackIDs*/ false);
- MadeChange = true;
- }
-
- if (SpilledToVirtVGPRLanes) {
const TargetRegisterClass *RC = TRI->getWaveMaskRegClass();
// Shift back the reserved SGPR for EXEC copy into the lowest range.
// This SGPR is reserved to handle the whole-wave spill/copy operations
@@ -433,21 +342,20 @@
if (UnusedLowSGPR && TRI->getHWRegIndex(UnusedLowSGPR) <
TRI->getHWRegIndex(FuncInfo->getSGPRForEXECCopy()))
FuncInfo->setSGPRForEXECCopy(UnusedLowSGPR);
+
+ MadeChange = true;
} else {
- // No SGPR spills to virtual VGPR lanes and hence there won't be any WWM
- // spills/copies. Reset the SGPR reserved for EXEC copy.
+ // No SGPR spills and hence there won't be any WWM spills/copies. Reset the
+ // SGPR reserved for EXEC copy.
FuncInfo->setSGPRForEXECCopy(AMDGPU::NoRegister);
}
SaveBlocks.clear();
RestoreBlocks.clear();
- // Updated the reserved registers with any physical VGPRs added for SGPR
- // spills.
- if (NewReservedRegs) {
- for (Register Reg : FuncInfo->getWWMReservedRegs())
- MRI.reserveReg(Reg, TRI);
- }
+ // Updated the reserved registers with any VGPRs added for SGPR spills.
+ if (NewReservedRegs)
+ MRI.freezeReservedRegs(MF);
return MadeChange;
}
diff --git a/llvm/lib/Target/AMDGPU/SILowerWWMCopies.cpp b/llvm/lib/Target/AMDGPU/SILowerWWMCopies.cpp
deleted file mode 100644
--- a/llvm/lib/Target/AMDGPU/SILowerWWMCopies.cpp
+++ /dev/null
@@ -1,141 +0,0 @@
-//===-- SILowerWWMCopies.cpp - Lower Copies after regalloc ---===//
-//
-// Part of the LLVM Project, under the Apache License v2.0 with LLVM Exceptions.
-// See https://llvm.org/LICENSE.txt for license information.
-// SPDX-License-Identifier: Apache-2.0 WITH LLVM-exception
-//
-//===----------------------------------------------------------------------===//
-//
-/// \file
-/// Lowering the WWM_COPY instructions for various register classes.
-/// AMDGPU target generates WWM_COPY instruction to differentiate WWM
-/// copy from COPY. This pass generates the necessary exec mask manipulation
-/// instructions to replicate 'Whole Wave Mode' and lowers WWM_COPY back to
-/// COPY.
-//
-//===----------------------------------------------------------------------===//
-
-#include "AMDGPU.h"
-#include "GCNSubtarget.h"
-#include "MCTargetDesc/AMDGPUMCTargetDesc.h"
-#include "SIMachineFunctionInfo.h"
-#include "llvm/CodeGen/LiveIntervals.h"
-#include "llvm/CodeGen/MachineFunctionPass.h"
-#include "llvm/CodeGen/VirtRegMap.h"
-#include "llvm/InitializePasses.h"
-
-using namespace llvm;
-
-#define DEBUG_TYPE "si-lower-wwm-copies"
-
-namespace {
-
-class SILowerWWMCopies : public MachineFunctionPass {
-public:
- static char ID;
-
- SILowerWWMCopies() : MachineFunctionPass(ID) {
- initializeSILowerWWMCopiesPass(*PassRegistry::getPassRegistry());
- }
-
- bool runOnMachineFunction(MachineFunction &MF) override;
-
- StringRef getPassName() const override { return "SI Lower WWM Copies"; }
-
- void getAnalysisUsage(AnalysisUsage &AU) const override {
- AU.setPreservesAll();
- MachineFunctionPass::getAnalysisUsage(AU);
- }
-
-private:
- bool isSCCLiveAtMI(const MachineInstr &MI);
- void addToWWMSpills(MachineFunction &MF, Register Reg);
-
- LiveIntervals *LIS;
- SlotIndexes *Indexes;
- VirtRegMap *VRM;
- const SIRegisterInfo *TRI;
- const MachineRegisterInfo *MRI;
- SIMachineFunctionInfo *MFI;
-};
-
-} // End anonymous namespace.
-
-INITIALIZE_PASS_BEGIN(SILowerWWMCopies, DEBUG_TYPE, "SI Lower WWM Copies",
- false, false)
-INITIALIZE_PASS_DEPENDENCY(LiveIntervals)
-INITIALIZE_PASS_DEPENDENCY(VirtRegMap)
-INITIALIZE_PASS_END(SILowerWWMCopies, DEBUG_TYPE, "SI Lower WWM Copies", false,
- false)
-
-char SILowerWWMCopies::ID = 0;
-
-char &llvm::SILowerWWMCopiesID = SILowerWWMCopies::ID;
-
-bool SILowerWWMCopies::isSCCLiveAtMI(const MachineInstr &MI) {
- // We can't determine the liveness info if LIS isn't available. Early return
- // in that case and always assume SCC is live.
- if (!LIS)
- return true;
-
- LiveRange &LR =
- LIS->getRegUnit(*MCRegUnitIterator(MCRegister::from(AMDGPU::SCC), TRI));
- SlotIndex Idx = LIS->getInstructionIndex(MI);
- return LR.liveAt(Idx);
-}
-
-// If \p Reg is assigned with a physical VGPR, add the latter into wwm-spills
-// for preserving its entire lanes at function prolog/epilog.
-void SILowerWWMCopies::addToWWMSpills(MachineFunction &MF, Register Reg) {
- if (Reg.isPhysical())
- return;
-
- Register PhysReg = VRM->getPhys(Reg);
- assert(PhysReg != VirtRegMap::NO_PHYS_REG &&
- "should have allocated a physical register");
-
- MFI->allocateWWMSpill(MF, PhysReg);
-}
-
-bool SILowerWWMCopies::runOnMachineFunction(MachineFunction &MF) {
- const GCNSubtarget &ST = MF.getSubtarget<GCNSubtarget>();
- const SIInstrInfo *TII = ST.getInstrInfo();
-
- MFI = MF.getInfo<SIMachineFunctionInfo>();
- LIS = getAnalysisIfAvailable<LiveIntervals>();
- Indexes = getAnalysisIfAvailable<SlotIndexes>();
- VRM = getAnalysisIfAvailable<VirtRegMap>();
- TRI = ST.getRegisterInfo();
- MRI = &MF.getRegInfo();
-
- if (!MFI->hasVRegFlags())
- return false;
-
- bool Changed = false;
- for (MachineBasicBlock &MBB : MF) {
- for (MachineInstr &MI : MBB) {
- if (MI.getOpcode() != AMDGPU::WWM_COPY)
- continue;
-
- // TODO: Club adjacent WWM ops between same exec save/restore
- assert(TII->isVGPRCopy(MI));
-
- // For WWM vector copies, manipulate the exec mask around the copy
- // instruction.
- const DebugLoc &DL = MI.getDebugLoc();
- MachineBasicBlock::iterator InsertPt = MI.getIterator();
- Register RegForExecCopy = MFI->getSGPRForEXECCopy();
- TII->insertScratchExecCopy(MF, MBB, InsertPt, DL, RegForExecCopy,
- isSCCLiveAtMI(MI), Indexes);
- TII->restoreExec(MF, MBB, ++InsertPt, DL, RegForExecCopy, Indexes);
- addToWWMSpills(MF, MI.getOperand(0).getReg());
- LLVM_DEBUG(dbgs() << "WWM copy manipulation for " << MI);
-
- // Lower WWM_COPY back to COPY
- MI.setDesc(TII->get(AMDGPU::COPY));
- Changed |= true;
- }
- }
-
- return Changed;
-}
diff --git a/llvm/lib/Target/AMDGPU/SIMachineFunctionInfo.h b/llvm/lib/Target/AMDGPU/SIMachineFunctionInfo.h
--- a/llvm/lib/Target/AMDGPU/SIMachineFunctionInfo.h
+++ b/llvm/lib/Target/AMDGPU/SIMachineFunctionInfo.h
@@ -496,16 +496,15 @@
};
private:
- // To track virtual VGPR + lane index for each subregister of the SGPR spilled
- // to frameindex key during SILowerSGPRSpills pass.
+ // To track VGPR + lane index for each subregister of the SGPR spilled to
+ // frameindex key during SILowerSGPRSpills pass.
+ DenseMap<int, std::vector<SIRegisterInfo::SpilledReg>> SGPRSpillToVGPRLanes;
+ // To track VGPR + lane index for spilling special SGPRs like Frame Pointer
+ // identified during PrologEpilogInserter.
DenseMap<int, std::vector<SIRegisterInfo::SpilledReg>>
- SGPRSpillsToVirtualVGPRLanes;
- // To track physical VGPR + lane index for CSR SGPR spills and special SGPRs
- // like Frame Pointer identified during PrologEpilogInserter.
- DenseMap<int, std::vector<SIRegisterInfo::SpilledReg>>
- SGPRSpillsToPhysicalVGPRLanes;
- unsigned NumVirtualVGPRSpillLanes = 0;
- unsigned NumPhysicalVGPRSpillLanes = 0;
+ PrologEpilogSGPRSpillToVGPRLanes;
+ unsigned NumVGPRSpillLanes = 0;
+ unsigned NumVGPRPrologEpilogSpillLanes = 0;
SmallVector<Register, 2> SpillVGPRs;
using WWMSpillsMap = MapVector<Register, int>;
// To track the registers used in instructions that can potentially modify the
@@ -549,10 +548,10 @@
private:
Register VGPRForAGPRCopy;
- bool allocateVirtualVGPRForSGPRSpills(MachineFunction &MF, int FI,
- unsigned LaneIndex);
- bool allocatePhysicalVGPRForSGPRSpills(MachineFunction &MF, int FI,
- unsigned LaneIndex);
+ bool allocateVGPRForSGPRSpills(MachineFunction &MF, int FI,
+ unsigned LaneIndex);
+ bool allocateVGPRForPrologEpilogSGPRSpills(MachineFunction &MF, int FI,
+ unsigned LaneIndex);
public:
Register getVGPRForAGPRCopy() const {
@@ -584,9 +583,9 @@
SIModeRegisterDefaults getMode() const { return Mode; }
ArrayRef<SIRegisterInfo::SpilledReg>
- getSGPRSpillToVirtualVGPRLanes(int FrameIndex) const {
- auto I = SGPRSpillsToVirtualVGPRLanes.find(FrameIndex);
- return (I == SGPRSpillsToVirtualVGPRLanes.end())
+ getSGPRSpillToVGPRLanes(int FrameIndex) const {
+ auto I = SGPRSpillToVGPRLanes.find(FrameIndex);
+ return (I == SGPRSpillToVGPRLanes.end())
? ArrayRef<SIRegisterInfo::SpilledReg>()
: ArrayRef(I->second);
}
@@ -648,9 +647,9 @@
}
ArrayRef<SIRegisterInfo::SpilledReg>
- getSGPRSpillToPhysicalVGPRLanes(int FrameIndex) const {
- auto I = SGPRSpillsToPhysicalVGPRLanes.find(FrameIndex);
- return (I == SGPRSpillsToPhysicalVGPRLanes.end())
+ getPrologEpilogSGPRSpillToVGPRLanes(int FrameIndex) const {
+ auto I = PrologEpilogSGPRSpillToVGPRLanes.find(FrameIndex);
+ return (I == PrologEpilogSGPRSpillToVGPRLanes.end())
? ArrayRef<SIRegisterInfo::SpilledReg>()
: ArrayRef(I->second);
}
@@ -668,8 +667,6 @@
return VRegFlags.inBounds(Reg) && VRegFlags[Reg] & Flag;
}
- bool hasVRegFlags() { return VRegFlags.size(); }
-
void allocateWWMSpill(MachineFunction &MF, Register VGPR, uint64_t Size = 4,
Align Alignment = Align(4));
diff --git a/llvm/lib/Target/AMDGPU/SIMachineFunctionInfo.cpp b/llvm/lib/Target/AMDGPU/SIMachineFunctionInfo.cpp
--- a/llvm/lib/Target/AMDGPU/SIMachineFunctionInfo.cpp
+++ b/llvm/lib/Target/AMDGPU/SIMachineFunctionInfo.cpp
@@ -314,23 +314,37 @@
return false;
}
-bool SIMachineFunctionInfo::allocateVirtualVGPRForSGPRSpills(
- MachineFunction &MF, int FI, unsigned LaneIndex) {
+bool SIMachineFunctionInfo::allocateVGPRForSGPRSpills(MachineFunction &MF,
+ int FI,
+ unsigned LaneIndex) {
+ const GCNSubtarget &ST = MF.getSubtarget<GCNSubtarget>();
+ const SIRegisterInfo *TRI = ST.getRegisterInfo();
MachineRegisterInfo &MRI = MF.getRegInfo();
Register LaneVGPR;
if (!LaneIndex) {
- LaneVGPR = MRI.createVirtualRegister(&AMDGPU::VGPR_32RegClass);
+ LaneVGPR = TRI->findUnusedRegister(MRI, &AMDGPU::VGPR_32RegClass, MF);
+ if (LaneVGPR == AMDGPU::NoRegister) {
+ // We have no VGPRs left for spilling SGPRs. Reset because we will not
+ // partially spill the SGPR to VGPRs.
+ SGPRSpillToVGPRLanes.erase(FI);
+ return false;
+ }
+
SpillVGPRs.push_back(LaneVGPR);
+ // Add this register as live-in to all blocks to avoid machine verifier
+ // complaining about use of an undefined physical register.
+ for (MachineBasicBlock &BB : MF)
+ BB.addLiveIn(LaneVGPR);
} else {
LaneVGPR = SpillVGPRs.back();
}
- SGPRSpillsToVirtualVGPRLanes[FI].push_back(
+ SGPRSpillToVGPRLanes[FI].push_back(
SIRegisterInfo::SpilledReg(LaneVGPR, LaneIndex));
return true;
}
-bool SIMachineFunctionInfo::allocatePhysicalVGPRForSGPRSpills(
+bool SIMachineFunctionInfo::allocateVGPRForPrologEpilogSGPRSpills(
MachineFunction &MF, int FI, unsigned LaneIndex) {
const GCNSubtarget &ST = MF.getSubtarget<GCNSubtarget>();
const SIRegisterInfo *TRI = ST.getRegisterInfo();
@@ -341,21 +355,16 @@
if (LaneVGPR == AMDGPU::NoRegister) {
// We have no VGPRs left for spilling SGPRs. Reset because we will not
// partially spill the SGPR to VGPRs.
- SGPRSpillsToPhysicalVGPRLanes.erase(FI);
+ PrologEpilogSGPRSpillToVGPRLanes.erase(FI);
return false;
}
allocateWWMSpill(MF, LaneVGPR);
- reserveWWMRegister(LaneVGPR);
- for (MachineBasicBlock &MBB : MF) {
- MBB.addLiveIn(LaneVGPR);
- MBB.sortUniqueLiveIns();
- }
} else {
- LaneVGPR = WWMReservedRegs.back();
+ LaneVGPR = WWMSpills.back().first;
}
- SGPRSpillsToPhysicalVGPRLanes[FI].push_back(
+ PrologEpilogSGPRSpillToVGPRLanes[FI].push_back(
SIRegisterInfo::SpilledReg(LaneVGPR, LaneIndex));
return true;
}
@@ -364,8 +373,8 @@
int FI,
bool IsPrologEpilog) {
std::vector<SIRegisterInfo::SpilledReg> &SpillLanes =
- IsPrologEpilog ? SGPRSpillsToPhysicalVGPRLanes[FI]
- : SGPRSpillsToVirtualVGPRLanes[FI];
+ IsPrologEpilog ? PrologEpilogSGPRSpillToVGPRLanes[FI]
+ : SGPRSpillToVGPRLanes[FI];
// This has already been allocated.
if (!SpillLanes.empty())
@@ -386,14 +395,15 @@
"not spilling SGPRs to VGPRs");
unsigned &NumSpillLanes =
- IsPrologEpilog ? NumPhysicalVGPRSpillLanes : NumVirtualVGPRSpillLanes;
+ IsPrologEpilog ? NumVGPRPrologEpilogSpillLanes : NumVGPRSpillLanes;
for (unsigned I = 0; I < NumLanes; ++I, ++NumSpillLanes) {
unsigned LaneIndex = (NumSpillLanes % WaveSize);
- bool Allocated = IsPrologEpilog
- ? allocatePhysicalVGPRForSGPRSpills(MF, FI, LaneIndex)
- : allocateVirtualVGPRForSGPRSpills(MF, FI, LaneIndex);
+ bool Allocated =
+ IsPrologEpilog
+ ? allocateVGPRForPrologEpilogSGPRSpills(MF, FI, LaneIndex)
+ : allocateVGPRForSGPRSpills(MF, FI, LaneIndex);
if (!Allocated) {
NumSpillLanes -= I;
return false;
@@ -474,25 +484,16 @@
bool SIMachineFunctionInfo::removeDeadFrameIndices(
MachineFrameInfo &MFI, bool ResetSGPRSpillStackIDs) {
- // Remove dead frame indices from function frame, however keep FP & BP since
- // spills for them haven't been inserted yet. And also make sure to remove the
- // frame indices from `SGPRSpillsToVirtualVGPRLanes` data structure,
- // otherwise, it could result in an unexpected side effect and bug, in case of
- // any re-mapping of freed frame indices by later pass(es) like "stack slot
+ // Remove dead frame indices from function frame. And also make sure to remove
+ // the frame indices from `SGPRSpillToVGPRLanes` data structure, otherwise, it
+ // could result in an unexpected side effect and bug, in case of any
+ // re-mapping of freed frame indices by later pass(es) like "stack slot
// coloring".
- for (auto &R : make_early_inc_range(SGPRSpillsToVirtualVGPRLanes)) {
+ for (auto &R : make_early_inc_range(SGPRSpillToVGPRLanes)) {
MFI.RemoveStackObject(R.first);
- SGPRSpillsToVirtualVGPRLanes.erase(R.first);
+ SGPRSpillToVGPRLanes.erase(R.first);
}
- // Remove the dead frame indices of CSR SGPRs which are spilled to physical
- // VGPR lanes during SILowerSGPRSpills pass.
- if (!ResetSGPRSpillStackIDs) {
- for (auto &R : make_early_inc_range(SGPRSpillsToPhysicalVGPRLanes)) {
- MFI.RemoveStackObject(R.first);
- SGPRSpillsToPhysicalVGPRLanes.erase(R.first);
- }
- }
bool HaveSGPRToMemory = false;
if (ResetSGPRSpillStackIDs) {
diff --git a/llvm/lib/Target/AMDGPU/SIRegisterInfo.h b/llvm/lib/Target/AMDGPU/SIRegisterInfo.h
--- a/llvm/lib/Target/AMDGPU/SIRegisterInfo.h
+++ b/llvm/lib/Target/AMDGPU/SIRegisterInfo.h
@@ -142,17 +142,14 @@
void buildVGPRSpillLoadStore(SGPRSpillBuilder &SB, int Index, int Offset,
bool IsLoad, bool IsKill = true) const;
- /// If \p OnlyToVGPR is true, this will only succeed if this manages to find a
- /// free VGPR lane to spill.
+ /// If \p OnlyToVGPR is true, this will only succeed if this
bool spillSGPR(MachineBasicBlock::iterator MI, int FI, RegScavenger *RS,
SlotIndexes *Indexes = nullptr, LiveIntervals *LIS = nullptr,
- bool OnlyToVGPR = false,
- bool SpillToPhysVGPRLane = false) const;
+ bool OnlyToVGPR = false) const;
bool restoreSGPR(MachineBasicBlock::iterator MI, int FI, RegScavenger *RS,
SlotIndexes *Indexes = nullptr, LiveIntervals *LIS = nullptr,
- bool OnlyToVGPR = false,
- bool SpillToPhysVGPRLane = false) const;
+ bool OnlyToVGPR = false) const;
bool spillEmergencySGPR(MachineBasicBlock::iterator MI,
MachineBasicBlock &RestoreMBB, Register SGPR,
@@ -166,10 +163,10 @@
unsigned FIOperandNum,
RegScavenger *RS) const override;
- bool eliminateSGPRToVGPRSpillFrameIndex(
- MachineBasicBlock::iterator MI, int FI, RegScavenger *RS,
- SlotIndexes *Indexes = nullptr, LiveIntervals *LIS = nullptr,
- bool SpillToPhysVGPRLane = false) const;
+ bool eliminateSGPRToVGPRSpillFrameIndex(MachineBasicBlock::iterator MI,
+ int FI, RegScavenger *RS,
+ SlotIndexes *Indexes = nullptr,
+ LiveIntervals *LIS = nullptr) const;
StringRef getRegAsmName(MCRegister Reg) const override;
diff --git a/llvm/lib/Target/AMDGPU/SIRegisterInfo.cpp b/llvm/lib/Target/AMDGPU/SIRegisterInfo.cpp
--- a/llvm/lib/Target/AMDGPU/SIRegisterInfo.cpp
+++ b/llvm/lib/Target/AMDGPU/SIRegisterInfo.cpp
@@ -712,6 +712,9 @@
for (MCPhysReg Reg : MFI->getVGPRSpillAGPRs())
reserveRegisterTuples(Reserved, Reg);
+ for (auto Reg : MFI->getSGPRSpillVGPRs())
+ reserveRegisterTuples(Reserved, Reg);
+
return Reserved;
}
@@ -1733,13 +1736,10 @@
bool SIRegisterInfo::spillSGPR(MachineBasicBlock::iterator MI, int Index,
RegScavenger *RS, SlotIndexes *Indexes,
- LiveIntervals *LIS, bool OnlyToVGPR,
- bool SpillToPhysVGPRLane) const {
+ LiveIntervals *LIS, bool OnlyToVGPR) const {
SGPRSpillBuilder SB(*this, *ST.getInstrInfo(), isWave32, MI, Index, RS);
- ArrayRef<SpilledReg> VGPRSpills =
- SpillToPhysVGPRLane ? SB.MFI.getSGPRSpillToPhysicalVGPRLanes(Index)
- : SB.MFI.getSGPRSpillToVirtualVGPRLanes(Index);
+ ArrayRef<SpilledReg> VGPRSpills = SB.MFI.getSGPRSpillToVGPRLanes(Index);
bool SpillToVGPR = !VGPRSpills.empty();
if (OnlyToVGPR && !SpillToVGPR)
return false;
@@ -1856,13 +1856,10 @@
bool SIRegisterInfo::restoreSGPR(MachineBasicBlock::iterator MI, int Index,
RegScavenger *RS, SlotIndexes *Indexes,
- LiveIntervals *LIS, bool OnlyToVGPR,
- bool SpillToPhysVGPRLane) const {
+ LiveIntervals *LIS, bool OnlyToVGPR) const {
SGPRSpillBuilder SB(*this, *ST.getInstrInfo(), isWave32, MI, Index, RS);
- ArrayRef<SpilledReg> VGPRSpills =
- SpillToPhysVGPRLane ? SB.MFI.getSGPRSpillToPhysicalVGPRLanes(Index)
- : SB.MFI.getSGPRSpillToVirtualVGPRLanes(Index);
+ ArrayRef<SpilledReg> VGPRSpills = SB.MFI.getSGPRSpillToVGPRLanes(Index);
bool SpillToVGPR = !VGPRSpills.empty();
if (OnlyToVGPR && !SpillToVGPR)
return false;
@@ -2008,7 +2005,7 @@
/// handled.
bool SIRegisterInfo::eliminateSGPRToVGPRSpillFrameIndex(
MachineBasicBlock::iterator MI, int FI, RegScavenger *RS,
- SlotIndexes *Indexes, LiveIntervals *LIS, bool SpillToPhysVGPRLane) const {
+ SlotIndexes *Indexes, LiveIntervals *LIS) const {
switch (MI->getOpcode()) {
case AMDGPU::SI_SPILL_S1024_SAVE:
case AMDGPU::SI_SPILL_S512_SAVE:
@@ -2024,7 +2021,7 @@
case AMDGPU::SI_SPILL_S96_SAVE:
case AMDGPU::SI_SPILL_S64_SAVE:
case AMDGPU::SI_SPILL_S32_SAVE:
- return spillSGPR(MI, FI, RS, Indexes, LIS, true, SpillToPhysVGPRLane);
+ return spillSGPR(MI, FI, RS, Indexes, LIS, true);
case AMDGPU::SI_SPILL_S1024_RESTORE:
case AMDGPU::SI_SPILL_S512_RESTORE:
case AMDGPU::SI_SPILL_S384_RESTORE:
@@ -2039,7 +2036,7 @@
case AMDGPU::SI_SPILL_S96_RESTORE:
case AMDGPU::SI_SPILL_S64_RESTORE:
case AMDGPU::SI_SPILL_S32_RESTORE:
- return restoreSGPR(MI, FI, RS, Indexes, LIS, true, SpillToPhysVGPRLane);
+ return restoreSGPR(MI, FI, RS, Indexes, LIS, true);
default:
llvm_unreachable("not an SGPR spill instruction");
}
diff --git a/llvm/test/CodeGen/AMDGPU/GlobalISel/assert-align.ll b/llvm/test/CodeGen/AMDGPU/GlobalISel/assert-align.ll
--- a/llvm/test/CodeGen/AMDGPU/GlobalISel/assert-align.ll
+++ b/llvm/test/CodeGen/AMDGPU/GlobalISel/assert-align.ll
@@ -11,12 +11,13 @@
; CHECK-NEXT: s_mov_b32 s33, s32
; CHECK-NEXT: s_or_saveexec_b64 s[18:19], -1
; CHECK-NEXT: buffer_store_dword v40, off, s[0:3], s33 ; 4-byte Folded Spill
+; CHECK-NEXT: buffer_store_dword v41, off, s[0:3], s33 offset:4 ; 4-byte Folded Spill
; CHECK-NEXT: s_mov_b64 exec, s[18:19]
-; CHECK-NEXT: v_writelane_b32 v40, s16, 2
; CHECK-NEXT: s_addk_i32 s32, 0x400
; CHECK-NEXT: v_writelane_b32 v40, s30, 0
; CHECK-NEXT: v_mov_b32_e32 v0, 0
; CHECK-NEXT: v_mov_b32_e32 v1, 0
+; CHECK-NEXT: v_writelane_b32 v41, s16, 0
; CHECK-NEXT: v_writelane_b32 v40, s31, 1
; CHECK-NEXT: s_getpc_b64 s[16:17]
; CHECK-NEXT: s_add_u32 s16, s16, ext@rel32@lo+4
@@ -27,9 +28,10 @@
; CHECK-NEXT: s_waitcnt vmcnt(0)
; CHECK-NEXT: v_readlane_b32 s31, v40, 1
; CHECK-NEXT: v_readlane_b32 s30, v40, 0
-; CHECK-NEXT: v_readlane_b32 s4, v40, 2
+; CHECK-NEXT: v_readlane_b32 s4, v41, 0
; CHECK-NEXT: s_or_saveexec_b64 s[6:7], -1
; CHECK-NEXT: buffer_load_dword v40, off, s[0:3], s33 ; 4-byte Folded Reload
+; CHECK-NEXT: buffer_load_dword v41, off, s[0:3], s33 offset:4 ; 4-byte Folded Reload
; CHECK-NEXT: s_mov_b64 exec, s[6:7]
; CHECK-NEXT: s_addk_i32 s32, 0xfc00
; CHECK-NEXT: s_mov_b32 s33, s4
diff --git a/llvm/test/CodeGen/AMDGPU/GlobalISel/call-outgoing-stack-args.ll b/llvm/test/CodeGen/AMDGPU/GlobalISel/call-outgoing-stack-args.ll
--- a/llvm/test/CodeGen/AMDGPU/GlobalISel/call-outgoing-stack-args.ll
+++ b/llvm/test/CodeGen/AMDGPU/GlobalISel/call-outgoing-stack-args.ll
@@ -228,17 +228,18 @@
; MUBUF-NEXT: s_mov_b32 s33, s32
; MUBUF-NEXT: s_or_saveexec_b64 s[6:7], -1
; MUBUF-NEXT: buffer_store_dword v40, off, s[0:3], s33 ; 4-byte Folded Spill
+; MUBUF-NEXT: buffer_store_dword v41, off, s[0:3], s33 offset:4 ; 4-byte Folded Spill
; MUBUF-NEXT: s_mov_b64 exec, s[6:7]
; MUBUF-NEXT: s_addk_i32 s32, 0x400
; MUBUF-NEXT: v_mov_b32_e32 v0, 9
; MUBUF-NEXT: buffer_store_dword v0, off, s[0:3], s32 offset:4
; MUBUF-NEXT: v_mov_b32_e32 v0, 10
-; MUBUF-NEXT: v_writelane_b32 v40, s4, 2
; MUBUF-NEXT: buffer_store_dword v0, off, s[0:3], s32 offset:8
; MUBUF-NEXT: v_mov_b32_e32 v0, 11
; MUBUF-NEXT: v_writelane_b32 v40, s30, 0
; MUBUF-NEXT: buffer_store_dword v0, off, s[0:3], s32 offset:12
; MUBUF-NEXT: v_mov_b32_e32 v0, 12
+; MUBUF-NEXT: v_writelane_b32 v41, s4, 0
; MUBUF-NEXT: v_writelane_b32 v40, s31, 1
; MUBUF-NEXT: buffer_store_dword v0, off, s[0:3], s32 offset:16
; MUBUF-NEXT: s_getpc_b64 s[4:5]
@@ -247,9 +248,10 @@
; MUBUF-NEXT: s_swappc_b64 s[30:31], s[4:5]
; MUBUF-NEXT: v_readlane_b32 s31, v40, 1
; MUBUF-NEXT: v_readlane_b32 s30, v40, 0
-; MUBUF-NEXT: v_readlane_b32 s4, v40, 2
+; MUBUF-NEXT: v_readlane_b32 s4, v41, 0
; MUBUF-NEXT: s_or_saveexec_b64 s[6:7], -1
; MUBUF-NEXT: buffer_load_dword v40, off, s[0:3], s33 ; 4-byte Folded Reload
+; MUBUF-NEXT: buffer_load_dword v41, off, s[0:3], s33 offset:4 ; 4-byte Folded Reload
; MUBUF-NEXT: s_mov_b64 exec, s[6:7]
; MUBUF-NEXT: s_addk_i32 s32, 0xfc00
; MUBUF-NEXT: s_mov_b32 s33, s4
@@ -263,9 +265,10 @@
; FLATSCR-NEXT: s_mov_b32 s33, s32
; FLATSCR-NEXT: s_or_saveexec_b64 s[2:3], -1
; FLATSCR-NEXT: scratch_store_dword off, v40, s33 ; 4-byte Folded Spill
+; FLATSCR-NEXT: scratch_store_dword off, v41, s33 offset:4 ; 4-byte Folded Spill
; FLATSCR-NEXT: s_mov_b64 exec, s[2:3]
; FLATSCR-NEXT: s_add_i32 s32, s32, 16
-; FLATSCR-NEXT: v_writelane_b32 v40, s0, 2
+; FLATSCR-NEXT: v_writelane_b32 v41, s0, 0
; FLATSCR-NEXT: s_add_u32 s0, s32, 4
; FLATSCR-NEXT: v_mov_b32_e32 v0, 9
; FLATSCR-NEXT: scratch_store_dword off, v0, s0
@@ -286,9 +289,10 @@
; FLATSCR-NEXT: s_swappc_b64 s[30:31], s[0:1]
; FLATSCR-NEXT: v_readlane_b32 s31, v40, 1
; FLATSCR-NEXT: v_readlane_b32 s30, v40, 0
-; FLATSCR-NEXT: v_readlane_b32 s0, v40, 2
+; FLATSCR-NEXT: v_readlane_b32 s0, v41, 0
; FLATSCR-NEXT: s_or_saveexec_b64 s[2:3], -1
; FLATSCR-NEXT: scratch_load_dword v40, off, s33 ; 4-byte Folded Reload
+; FLATSCR-NEXT: scratch_load_dword v41, off, s33 offset:4 ; 4-byte Folded Reload
; FLATSCR-NEXT: s_mov_b64 exec, s[2:3]
; FLATSCR-NEXT: s_add_i32 s32, s32, -16
; FLATSCR-NEXT: s_mov_b32 s33, s0
@@ -306,12 +310,13 @@
; MUBUF-NEXT: s_mov_b32 s33, s32
; MUBUF-NEXT: s_or_saveexec_b64 s[6:7], -1
; MUBUF-NEXT: buffer_store_dword v40, off, s[0:3], s33 ; 4-byte Folded Spill
+; MUBUF-NEXT: buffer_store_dword v41, off, s[0:3], s33 offset:4 ; 4-byte Folded Spill
; MUBUF-NEXT: s_mov_b64 exec, s[6:7]
; MUBUF-NEXT: buffer_load_dword v1, v0, s[0:3], 0 offen
; MUBUF-NEXT: buffer_load_dword v2, v0, s[0:3], 0 offen offset:4
; MUBUF-NEXT: s_addk_i32 s32, 0x400
-; MUBUF-NEXT: v_writelane_b32 v40, s4, 2
; MUBUF-NEXT: v_writelane_b32 v40, s30, 0
+; MUBUF-NEXT: v_writelane_b32 v41, s4, 0
; MUBUF-NEXT: v_writelane_b32 v40, s31, 1
; MUBUF-NEXT: s_getpc_b64 s[4:5]
; MUBUF-NEXT: s_add_u32 s4, s4, external_void_func_byval@rel32@lo+4
@@ -372,9 +377,10 @@
; MUBUF-NEXT: s_swappc_b64 s[30:31], s[4:5]
; MUBUF-NEXT: v_readlane_b32 s31, v40, 1
; MUBUF-NEXT: v_readlane_b32 s30, v40, 0
-; MUBUF-NEXT: v_readlane_b32 s4, v40, 2
+; MUBUF-NEXT: v_readlane_b32 s4, v41, 0
; MUBUF-NEXT: s_or_saveexec_b64 s[6:7], -1
; MUBUF-NEXT: buffer_load_dword v40, off, s[0:3], s33 ; 4-byte Folded Reload
+; MUBUF-NEXT: buffer_load_dword v41, off, s[0:3], s33 offset:4 ; 4-byte Folded Reload
; MUBUF-NEXT: s_mov_b64 exec, s[6:7]
; MUBUF-NEXT: s_addk_i32 s32, 0xfc00
; MUBUF-NEXT: s_mov_b32 s33, s4
@@ -388,11 +394,12 @@
; FLATSCR-NEXT: s_mov_b32 s33, s32
; FLATSCR-NEXT: s_or_saveexec_b64 s[2:3], -1
; FLATSCR-NEXT: scratch_store_dword off, v40, s33 ; 4-byte Folded Spill
+; FLATSCR-NEXT: scratch_store_dword off, v41, s33 offset:4 ; 4-byte Folded Spill
; FLATSCR-NEXT: s_mov_b64 exec, s[2:3]
; FLATSCR-NEXT: scratch_load_dwordx2 v[1:2], v0, off
; FLATSCR-NEXT: s_add_i32 s32, s32, 16
; FLATSCR-NEXT: v_add_u32_e32 v3, 8, v0
-; FLATSCR-NEXT: v_writelane_b32 v40, s0, 2
+; FLATSCR-NEXT: v_writelane_b32 v41, s0, 0
; FLATSCR-NEXT: s_add_u32 s0, s32, 8
; FLATSCR-NEXT: v_writelane_b32 v40, s30, 0
; FLATSCR-NEXT: s_add_u32 s2, s32, 56
@@ -437,9 +444,10 @@
; FLATSCR-NEXT: s_swappc_b64 s[30:31], s[0:1]
; FLATSCR-NEXT: v_readlane_b32 s31, v40, 1
; FLATSCR-NEXT: v_readlane_b32 s30, v40, 0
-; FLATSCR-NEXT: v_readlane_b32 s0, v40, 2
+; FLATSCR-NEXT: v_readlane_b32 s0, v41, 0
; FLATSCR-NEXT: s_or_saveexec_b64 s[2:3], -1
; FLATSCR-NEXT: scratch_load_dword v40, off, s33 ; 4-byte Folded Reload
+; FLATSCR-NEXT: scratch_load_dword v41, off, s33 offset:4 ; 4-byte Folded Reload
; FLATSCR-NEXT: s_mov_b64 exec, s[2:3]
; FLATSCR-NEXT: s_add_i32 s32, s32, -16
; FLATSCR-NEXT: s_mov_b32 s33, s0
diff --git a/llvm/test/CodeGen/AMDGPU/GlobalISel/image-waterfall-loop-O0.ll b/llvm/test/CodeGen/AMDGPU/GlobalISel/image-waterfall-loop-O0.ll
--- a/llvm/test/CodeGen/AMDGPU/GlobalISel/image-waterfall-loop-O0.ll
+++ b/llvm/test/CodeGen/AMDGPU/GlobalISel/image-waterfall-loop-O0.ll
@@ -12,39 +12,32 @@
; CHECK: ; %bb.0: ; %bb
; CHECK-NEXT: s_waitcnt vmcnt(0) expcnt(0) lgkmcnt(0)
; CHECK-NEXT: s_xor_saveexec_b32 s4, -1
-; CHECK-NEXT: buffer_store_dword v0, off, s[0:3], s32 offset:80 ; 4-byte Folded Spill
-; CHECK-NEXT: buffer_store_dword v2, off, s[0:3], s32 offset:84 ; 4-byte Folded Spill
+; CHECK-NEXT: buffer_store_dword v8, off, s[0:3], s32 offset:76 ; 4-byte Folded Spill
; CHECK-NEXT: s_mov_b32 exec_lo, s4
-; CHECK-NEXT: ; implicit-def: $vgpr8
-; CHECK-NEXT: v_mov_b32_e32 v8, v0
-; CHECK-NEXT: s_or_saveexec_b32 s21, -1
-; CHECK-NEXT: buffer_load_dword v0, off, s[0:3], s32 ; 4-byte Folded Reload
-; CHECK-NEXT: s_mov_b32 exec_lo, s21
-; CHECK-NEXT: buffer_store_dword v8, off, s[0:3], s32 offset:72 ; 4-byte Folded Spill
+; CHECK-NEXT: buffer_store_dword v0, off, s[0:3], s32 offset:68 ; 4-byte Folded Spill
; CHECK-NEXT: v_mov_b32_e32 v15, v1
-; CHECK-NEXT: buffer_load_dword v1, off, s[0:3], s32 offset:72 ; 4-byte Folded Reload
-; CHECK-NEXT: buffer_store_dword v15, off, s[0:3], s32 offset:68 ; 4-byte Folded Spill
+; CHECK-NEXT: buffer_store_dword v15, off, s[0:3], s32 offset:64 ; 4-byte Folded Spill
; CHECK-NEXT: v_mov_b32_e32 v14, v2
-; CHECK-NEXT: buffer_store_dword v14, off, s[0:3], s32 offset:64 ; 4-byte Folded Spill
+; CHECK-NEXT: buffer_store_dword v14, off, s[0:3], s32 offset:60 ; 4-byte Folded Spill
; CHECK-NEXT: v_mov_b32_e32 v13, v3
-; CHECK-NEXT: buffer_store_dword v13, off, s[0:3], s32 offset:60 ; 4-byte Folded Spill
+; CHECK-NEXT: buffer_store_dword v13, off, s[0:3], s32 offset:56 ; 4-byte Folded Spill
; CHECK-NEXT: v_mov_b32_e32 v12, v4
-; CHECK-NEXT: buffer_store_dword v12, off, s[0:3], s32 offset:56 ; 4-byte Folded Spill
+; CHECK-NEXT: buffer_store_dword v12, off, s[0:3], s32 offset:52 ; 4-byte Folded Spill
; CHECK-NEXT: v_mov_b32_e32 v11, v5
-; CHECK-NEXT: buffer_store_dword v11, off, s[0:3], s32 offset:52 ; 4-byte Folded Spill
+; CHECK-NEXT: buffer_store_dword v11, off, s[0:3], s32 offset:48 ; 4-byte Folded Spill
; CHECK-NEXT: v_mov_b32_e32 v10, v6
-; CHECK-NEXT: buffer_store_dword v10, off, s[0:3], s32 offset:48 ; 4-byte Folded Spill
+; CHECK-NEXT: buffer_store_dword v10, off, s[0:3], s32 offset:44 ; 4-byte Folded Spill
; CHECK-NEXT: v_mov_b32_e32 v9, v7
-; CHECK-NEXT: buffer_store_dword v9, off, s[0:3], s32 offset:44 ; 4-byte Folded Spill
-; CHECK-NEXT: ; kill: def $vgpr1 killed $vgpr1 def $vgpr1_vgpr2_vgpr3_vgpr4_vgpr5_vgpr6_vgpr7_vgpr8 killed $exec
-; CHECK-NEXT: v_mov_b32_e32 v2, v15
-; CHECK-NEXT: v_mov_b32_e32 v3, v14
-; CHECK-NEXT: v_mov_b32_e32 v4, v13
-; CHECK-NEXT: v_mov_b32_e32 v5, v12
-; CHECK-NEXT: v_mov_b32_e32 v6, v11
-; CHECK-NEXT: v_mov_b32_e32 v7, v10
-; CHECK-NEXT: v_mov_b32_e32 v8, v9
-; CHECK-NEXT: s_waitcnt vmcnt(0)
+; CHECK-NEXT: buffer_store_dword v9, off, s[0:3], s32 offset:40 ; 4-byte Folded Spill
+; CHECK-NEXT: ; kill: def $vgpr0 killed $vgpr0 def $vgpr0_vgpr1_vgpr2_vgpr3_vgpr4_vgpr5_vgpr6_vgpr7 killed $exec
+; CHECK-NEXT: v_mov_b32_e32 v1, v15
+; CHECK-NEXT: v_mov_b32_e32 v2, v14
+; CHECK-NEXT: v_mov_b32_e32 v3, v13
+; CHECK-NEXT: v_mov_b32_e32 v4, v12
+; CHECK-NEXT: v_mov_b32_e32 v5, v11
+; CHECK-NEXT: v_mov_b32_e32 v6, v10
+; CHECK-NEXT: v_mov_b32_e32 v7, v9
+; CHECK-NEXT: buffer_store_dword v0, off, s[0:3], s32 offset:8 ; 4-byte Folded Spill
; CHECK-NEXT: buffer_store_dword v1, off, s[0:3], s32 offset:12 ; 4-byte Folded Spill
; CHECK-NEXT: buffer_store_dword v2, off, s[0:3], s32 offset:16 ; 4-byte Folded Spill
; CHECK-NEXT: buffer_store_dword v3, off, s[0:3], s32 offset:20 ; 4-byte Folded Spill
@@ -52,40 +45,34 @@
; CHECK-NEXT: buffer_store_dword v5, off, s[0:3], s32 offset:28 ; 4-byte Folded Spill
; CHECK-NEXT: buffer_store_dword v6, off, s[0:3], s32 offset:32 ; 4-byte Folded Spill
; CHECK-NEXT: buffer_store_dword v7, off, s[0:3], s32 offset:36 ; 4-byte Folded Spill
-; CHECK-NEXT: buffer_store_dword v8, off, s[0:3], s32 offset:40 ; 4-byte Folded Spill
; CHECK-NEXT: s_mov_b32 s8, 0
; CHECK-NEXT: s_mov_b32 s4, s8
; CHECK-NEXT: s_mov_b32 s5, s8
; CHECK-NEXT: s_mov_b32 s6, s8
; CHECK-NEXT: s_mov_b32 s7, s8
-; CHECK-NEXT: v_writelane_b32 v0, s4, 0
-; CHECK-NEXT: v_writelane_b32 v0, s5, 1
-; CHECK-NEXT: v_writelane_b32 v0, s6, 2
-; CHECK-NEXT: v_writelane_b32 v0, s7, 3
+; CHECK-NEXT: v_writelane_b32 v8, s4, 0
+; CHECK-NEXT: v_writelane_b32 v8, s5, 1
+; CHECK-NEXT: v_writelane_b32 v8, s6, 2
+; CHECK-NEXT: v_writelane_b32 v8, s7, 3
; CHECK-NEXT: s_mov_b32 s6, 0
; CHECK-NEXT: s_mov_b32 s4, s6
; CHECK-NEXT: s_mov_b32 s5, s6
-; CHECK-NEXT: v_mov_b32_e32 v1, s4
-; CHECK-NEXT: v_mov_b32_e32 v2, s5
+; CHECK-NEXT: v_mov_b32_e32 v0, s4
+; CHECK-NEXT: v_mov_b32_e32 v1, s5
+; CHECK-NEXT: buffer_store_dword v0, off, s[0:3], s32 ; 4-byte Folded Spill
; CHECK-NEXT: buffer_store_dword v1, off, s[0:3], s32 offset:4 ; 4-byte Folded Spill
-; CHECK-NEXT: buffer_store_dword v2, off, s[0:3], s32 offset:8 ; 4-byte Folded Spill
; CHECK-NEXT: s_mov_b32 s4, exec_lo
-; CHECK-NEXT: v_writelane_b32 v0, s4, 4
-; CHECK-NEXT: s_or_saveexec_b32 s21, -1
-; CHECK-NEXT: buffer_store_dword v0, off, s[0:3], s32 ; 4-byte Folded Spill
-; CHECK-NEXT: s_mov_b32 exec_lo, s21
+; CHECK-NEXT: v_writelane_b32 v8, s4, 4
; CHECK-NEXT: .LBB0_1: ; =>This Inner Loop Header: Depth=1
-; CHECK-NEXT: s_or_saveexec_b32 s21, -1
-; CHECK-NEXT: buffer_load_dword v0, off, s[0:3], s32 ; 4-byte Folded Reload
-; CHECK-NEXT: s_mov_b32 exec_lo, s21
-; CHECK-NEXT: buffer_load_dword v9, off, s[0:3], s32 offset:12 ; 4-byte Folded Reload
-; CHECK-NEXT: buffer_load_dword v10, off, s[0:3], s32 offset:16 ; 4-byte Folded Reload
-; CHECK-NEXT: buffer_load_dword v11, off, s[0:3], s32 offset:20 ; 4-byte Folded Reload
-; CHECK-NEXT: buffer_load_dword v12, off, s[0:3], s32 offset:24 ; 4-byte Folded Reload
-; CHECK-NEXT: buffer_load_dword v13, off, s[0:3], s32 offset:28 ; 4-byte Folded Reload
-; CHECK-NEXT: buffer_load_dword v14, off, s[0:3], s32 offset:32 ; 4-byte Folded Reload
-; CHECK-NEXT: buffer_load_dword v15, off, s[0:3], s32 offset:36 ; 4-byte Folded Reload
-; CHECK-NEXT: buffer_load_dword v16, off, s[0:3], s32 offset:40 ; 4-byte Folded Reload
+; CHECK-NEXT: buffer_load_dword v9, off, s[0:3], s32 offset:8 ; 4-byte Folded Reload
+; CHECK-NEXT: buffer_load_dword v10, off, s[0:3], s32 offset:12 ; 4-byte Folded Reload
+; CHECK-NEXT: buffer_load_dword v11, off, s[0:3], s32 offset:16 ; 4-byte Folded Reload
+; CHECK-NEXT: buffer_load_dword v12, off, s[0:3], s32 offset:20 ; 4-byte Folded Reload
+; CHECK-NEXT: buffer_load_dword v13, off, s[0:3], s32 offset:24 ; 4-byte Folded Reload
+; CHECK-NEXT: buffer_load_dword v14, off, s[0:3], s32 offset:28 ; 4-byte Folded Reload
+; CHECK-NEXT: buffer_load_dword v15, off, s[0:3], s32 offset:32 ; 4-byte Folded Reload
+; CHECK-NEXT: buffer_load_dword v16, off, s[0:3], s32 offset:36 ; 4-byte Folded Reload
+; CHECK-NEXT: buffer_load_dword v0, off, s[0:3], s32 offset:40 ; 4-byte Folded Reload
; CHECK-NEXT: buffer_load_dword v1, off, s[0:3], s32 offset:44 ; 4-byte Folded Reload
; CHECK-NEXT: buffer_load_dword v2, off, s[0:3], s32 offset:48 ; 4-byte Folded Reload
; CHECK-NEXT: buffer_load_dword v3, off, s[0:3], s32 offset:52 ; 4-byte Folded Reload
@@ -93,16 +80,15 @@
; CHECK-NEXT: buffer_load_dword v5, off, s[0:3], s32 offset:60 ; 4-byte Folded Reload
; CHECK-NEXT: buffer_load_dword v6, off, s[0:3], s32 offset:64 ; 4-byte Folded Reload
; CHECK-NEXT: buffer_load_dword v7, off, s[0:3], s32 offset:68 ; 4-byte Folded Reload
-; CHECK-NEXT: buffer_load_dword v8, off, s[0:3], s32 offset:72 ; 4-byte Folded Reload
; CHECK-NEXT: s_waitcnt vmcnt(0)
-; CHECK-NEXT: v_readfirstlane_b32 s12, v8
-; CHECK-NEXT: v_readfirstlane_b32 s10, v7
-; CHECK-NEXT: v_readfirstlane_b32 s9, v6
-; CHECK-NEXT: v_readfirstlane_b32 s8, v5
-; CHECK-NEXT: v_readfirstlane_b32 s7, v4
-; CHECK-NEXT: v_readfirstlane_b32 s6, v3
-; CHECK-NEXT: v_readfirstlane_b32 s5, v2
-; CHECK-NEXT: v_readfirstlane_b32 s4, v1
+; CHECK-NEXT: v_readfirstlane_b32 s12, v7
+; CHECK-NEXT: v_readfirstlane_b32 s10, v6
+; CHECK-NEXT: v_readfirstlane_b32 s9, v5
+; CHECK-NEXT: v_readfirstlane_b32 s8, v4
+; CHECK-NEXT: v_readfirstlane_b32 s7, v3
+; CHECK-NEXT: v_readfirstlane_b32 s6, v2
+; CHECK-NEXT: v_readfirstlane_b32 s5, v1
+; CHECK-NEXT: v_readfirstlane_b32 s4, v0
; CHECK-NEXT: ; kill: def $sgpr12 killed $sgpr12 def $sgpr12_sgpr13_sgpr14_sgpr15_sgpr16_sgpr17_sgpr18_sgpr19
; CHECK-NEXT: s_mov_b32 s13, s10
; CHECK-NEXT: s_mov_b32 s14, s9
@@ -111,84 +97,68 @@
; CHECK-NEXT: s_mov_b32 s17, s6
; CHECK-NEXT: s_mov_b32 s18, s5
; CHECK-NEXT: s_mov_b32 s19, s4
-; CHECK-NEXT: v_writelane_b32 v0, s12, 5
-; CHECK-NEXT: v_writelane_b32 v0, s13, 6
-; CHECK-NEXT: v_writelane_b32 v0, s14, 7
-; CHECK-NEXT: v_writelane_b32 v0, s15, 8
-; CHECK-NEXT: v_writelane_b32 v0, s16, 9
-; CHECK-NEXT: v_writelane_b32 v0, s17, 10
-; CHECK-NEXT: v_writelane_b32 v0, s18, 11
-; CHECK-NEXT: v_writelane_b32 v0, s19, 12
-; CHECK-NEXT: v_mov_b32_e32 v7, v9
-; CHECK-NEXT: v_mov_b32_e32 v8, v10
-; CHECK-NEXT: v_mov_b32_e32 v5, v11
-; CHECK-NEXT: v_mov_b32_e32 v6, v12
-; CHECK-NEXT: v_mov_b32_e32 v3, v13
-; CHECK-NEXT: v_mov_b32_e32 v4, v14
-; CHECK-NEXT: v_mov_b32_e32 v1, v15
-; CHECK-NEXT: v_mov_b32_e32 v2, v16
+; CHECK-NEXT: v_writelane_b32 v8, s12, 5
+; CHECK-NEXT: v_writelane_b32 v8, s13, 6
+; CHECK-NEXT: v_writelane_b32 v8, s14, 7
+; CHECK-NEXT: v_writelane_b32 v8, s15, 8
+; CHECK-NEXT: v_writelane_b32 v8, s16, 9
+; CHECK-NEXT: v_writelane_b32 v8, s17, 10
+; CHECK-NEXT: v_writelane_b32 v8, s18, 11
+; CHECK-NEXT: v_writelane_b32 v8, s19, 12
+; CHECK-NEXT: v_mov_b32_e32 v6, v9
+; CHECK-NEXT: v_mov_b32_e32 v7, v10
+; CHECK-NEXT: v_mov_b32_e32 v4, v11
+; CHECK-NEXT: v_mov_b32_e32 v5, v12
+; CHECK-NEXT: v_mov_b32_e32 v2, v13
+; CHECK-NEXT: v_mov_b32_e32 v3, v14
+; CHECK-NEXT: v_mov_b32_e32 v0, v15
+; CHECK-NEXT: v_mov_b32_e32 v1, v16
; CHECK-NEXT: s_mov_b64 s[4:5], s[12:13]
; CHECK-NEXT: s_mov_b64 s[10:11], s[14:15]
; CHECK-NEXT: s_mov_b64 s[8:9], s[16:17]
; CHECK-NEXT: s_mov_b64 s[6:7], s[18:19]
-; CHECK-NEXT: v_cmp_eq_u64_e64 s4, s[4:5], v[7:8]
-; CHECK-NEXT: v_cmp_eq_u64_e64 s5, s[10:11], v[5:6]
+; CHECK-NEXT: v_cmp_eq_u64_e64 s4, s[4:5], v[6:7]
+; CHECK-NEXT: v_cmp_eq_u64_e64 s5, s[10:11], v[4:5]
; CHECK-NEXT: s_and_b32 s4, s4, s5
-; CHECK-NEXT: v_cmp_eq_u64_e64 s5, s[8:9], v[3:4]
+; CHECK-NEXT: v_cmp_eq_u64_e64 s5, s[8:9], v[2:3]
; CHECK-NEXT: s_and_b32 s4, s4, s5
-; CHECK-NEXT: v_cmp_eq_u64_e64 s5, s[6:7], v[1:2]
+; CHECK-NEXT: v_cmp_eq_u64_e64 s5, s[6:7], v[0:1]
; CHECK-NEXT: s_and_b32 s4, s4, s5
; CHECK-NEXT: s_and_saveexec_b32 s4, s4
-; CHECK-NEXT: v_writelane_b32 v0, s4, 13
-; CHECK-NEXT: s_or_saveexec_b32 s21, -1
-; CHECK-NEXT: buffer_store_dword v0, off, s[0:3], s32 ; 4-byte Folded Spill
-; CHECK-NEXT: s_mov_b32 exec_lo, s21
+; CHECK-NEXT: v_writelane_b32 v8, s4, 13
; CHECK-NEXT: ; %bb.2: ; in Loop: Header=BB0_1 Depth=1
-; CHECK-NEXT: buffer_load_dword v0, off, s[0:3], s32 offset:4 ; 4-byte Folded Reload
-; CHECK-NEXT: buffer_load_dword v1, off, s[0:3], s32 offset:8 ; 4-byte Folded Reload
-; CHECK-NEXT: s_or_saveexec_b32 s21, -1
-; CHECK-NEXT: buffer_load_dword v2, off, s[0:3], s32 ; 4-byte Folded Reload
-; CHECK-NEXT: s_mov_b32 exec_lo, s21
+; CHECK-NEXT: buffer_load_dword v0, off, s[0:3], s32 ; 4-byte Folded Reload
+; CHECK-NEXT: buffer_load_dword v1, off, s[0:3], s32 offset:4 ; 4-byte Folded Reload
+; CHECK-NEXT: v_readlane_b32 s4, v8, 13
+; CHECK-NEXT: v_readlane_b32 s8, v8, 5
+; CHECK-NEXT: v_readlane_b32 s9, v8, 6
+; CHECK-NEXT: v_readlane_b32 s10, v8, 7
+; CHECK-NEXT: v_readlane_b32 s11, v8, 8
+; CHECK-NEXT: v_readlane_b32 s12, v8, 9
+; CHECK-NEXT: v_readlane_b32 s13, v8, 10
+; CHECK-NEXT: v_readlane_b32 s14, v8, 11
+; CHECK-NEXT: v_readlane_b32 s15, v8, 12
+; CHECK-NEXT: v_readlane_b32 s16, v8, 0
+; CHECK-NEXT: v_readlane_b32 s17, v8, 1
+; CHECK-NEXT: v_readlane_b32 s18, v8, 2
+; CHECK-NEXT: v_readlane_b32 s19, v8, 3
; CHECK-NEXT: s_waitcnt vmcnt(0)
-; CHECK-NEXT: v_readlane_b32 s4, v2, 13
-; CHECK-NEXT: v_readlane_b32 s8, v2, 5
-; CHECK-NEXT: v_readlane_b32 s9, v2, 6
-; CHECK-NEXT: v_readlane_b32 s10, v2, 7
-; CHECK-NEXT: v_readlane_b32 s11, v2, 8
-; CHECK-NEXT: v_readlane_b32 s12, v2, 9
-; CHECK-NEXT: v_readlane_b32 s13, v2, 10
-; CHECK-NEXT: v_readlane_b32 s14, v2, 11
-; CHECK-NEXT: v_readlane_b32 s15, v2, 12
-; CHECK-NEXT: v_readlane_b32 s16, v2, 0
-; CHECK-NEXT: v_readlane_b32 s17, v2, 1
-; CHECK-NEXT: v_readlane_b32 s18, v2, 2
-; CHECK-NEXT: v_readlane_b32 s19, v2, 3
; CHECK-NEXT: image_sample v0, v[0:1], s[8:15], s[16:19] dmask:0x1 dim:SQ_RSRC_IMG_2D
; CHECK-NEXT: s_waitcnt vmcnt(0)
-; CHECK-NEXT: buffer_store_dword v0, off, s[0:3], s32 offset:76 ; 4-byte Folded Spill
+; CHECK-NEXT: buffer_store_dword v0, off, s[0:3], s32 offset:72 ; 4-byte Folded Spill
; CHECK-NEXT: s_xor_b32 exec_lo, exec_lo, s4
; CHECK-NEXT: s_cbranch_execnz .LBB0_1
; CHECK-NEXT: ; %bb.3:
-; CHECK-NEXT: s_or_saveexec_b32 s21, -1
-; CHECK-NEXT: buffer_load_dword v0, off, s[0:3], s32 ; 4-byte Folded Reload
-; CHECK-NEXT: s_mov_b32 exec_lo, s21
-; CHECK-NEXT: s_waitcnt vmcnt(0)
-; CHECK-NEXT: v_readlane_b32 s4, v0, 4
+; CHECK-NEXT: v_readlane_b32 s4, v8, 4
; CHECK-NEXT: s_mov_b32 exec_lo, s4
; CHECK-NEXT: ; %bb.4:
-; CHECK-NEXT: s_or_saveexec_b32 s21, -1
-; CHECK-NEXT: buffer_load_dword v4, off, s[0:3], s32 ; 4-byte Folded Reload
-; CHECK-NEXT: s_mov_b32 exec_lo, s21
-; CHECK-NEXT: buffer_load_dword v0, off, s[0:3], s32 offset:76 ; 4-byte Folded Reload
+; CHECK-NEXT: buffer_load_dword v0, off, s[0:3], s32 offset:72 ; 4-byte Folded Reload
; CHECK-NEXT: ; implicit-def: $sgpr4
; CHECK-NEXT: v_mov_b32_e32 v1, s4
; CHECK-NEXT: v_mov_b32_e32 v2, s4
; CHECK-NEXT: v_mov_b32_e32 v3, s4
-; CHECK-NEXT: ; kill: killed $vgpr4
; CHECK-NEXT: s_xor_saveexec_b32 s4, -1
-; CHECK-NEXT: s_waitcnt vmcnt(0)
-; CHECK-NEXT: buffer_load_dword v0, off, s[0:3], s32 offset:80 ; 4-byte Folded Reload
-; CHECK-NEXT: buffer_load_dword v2, off, s[0:3], s32 offset:84 ; 4-byte Folded Reload
+; CHECK-NEXT: buffer_load_dword v8, off, s[0:3], s32 offset:76 ; 4-byte Folded Reload
; CHECK-NEXT: s_mov_b32 exec_lo, s4
; CHECK-NEXT: s_waitcnt vmcnt(0)
; CHECK-NEXT: s_setpc_b64 s[30:31]
diff --git a/llvm/test/CodeGen/AMDGPU/GlobalISel/localizer.ll b/llvm/test/CodeGen/AMDGPU/GlobalISel/localizer.ll
--- a/llvm/test/CodeGen/AMDGPU/GlobalISel/localizer.ll
+++ b/llvm/test/CodeGen/AMDGPU/GlobalISel/localizer.ll
@@ -236,21 +236,23 @@
; GFX9-NEXT: s_mov_b32 s33, s32
; GFX9-NEXT: s_or_saveexec_b64 s[18:19], -1
; GFX9-NEXT: buffer_store_dword v40, off, s[0:3], s33 ; 4-byte Folded Spill
+; GFX9-NEXT: buffer_store_dword v41, off, s[0:3], s33 offset:4 ; 4-byte Folded Spill
; GFX9-NEXT: s_mov_b64 exec, s[18:19]
; GFX9-NEXT: v_mov_b32_e32 v0, 0
; GFX9-NEXT: v_mov_b32_e32 v1, 0
; GFX9-NEXT: global_load_dword v0, v[0:1], off glc
; GFX9-NEXT: s_waitcnt vmcnt(0)
-; GFX9-NEXT: v_writelane_b32 v40, s16, 2
; GFX9-NEXT: v_writelane_b32 v40, s30, 0
+; GFX9-NEXT: v_writelane_b32 v41, s16, 0
; GFX9-NEXT: s_addk_i32 s32, 0x400
; GFX9-NEXT: v_writelane_b32 v40, s31, 1
; GFX9-NEXT: s_swappc_b64 s[30:31], 0
; GFX9-NEXT: v_readlane_b32 s31, v40, 1
; GFX9-NEXT: v_readlane_b32 s30, v40, 0
-; GFX9-NEXT: v_readlane_b32 s4, v40, 2
+; GFX9-NEXT: v_readlane_b32 s4, v41, 0
; GFX9-NEXT: s_or_saveexec_b64 s[6:7], -1
; GFX9-NEXT: buffer_load_dword v40, off, s[0:3], s33 ; 4-byte Folded Reload
+; GFX9-NEXT: buffer_load_dword v41, off, s[0:3], s33 offset:4 ; 4-byte Folded Reload
; GFX9-NEXT: s_mov_b64 exec, s[6:7]
; GFX9-NEXT: s_addk_i32 s32, 0xfc00
; GFX9-NEXT: s_mov_b32 s33, s4
diff --git a/llvm/test/CodeGen/AMDGPU/abi-attribute-hints-undefined-behavior.ll b/llvm/test/CodeGen/AMDGPU/abi-attribute-hints-undefined-behavior.ll
--- a/llvm/test/CodeGen/AMDGPU/abi-attribute-hints-undefined-behavior.ll
+++ b/llvm/test/CodeGen/AMDGPU/abi-attribute-hints-undefined-behavior.ll
@@ -21,10 +21,11 @@
; FIXEDABI-NEXT: s_mov_b32 s33, s32
; FIXEDABI-NEXT: s_or_saveexec_b64 s[18:19], -1
; FIXEDABI-NEXT: buffer_store_dword v40, off, s[0:3], s33 ; 4-byte Folded Spill
+; FIXEDABI-NEXT: buffer_store_dword v41, off, s[0:3], s33 offset:4 ; 4-byte Folded Spill
; FIXEDABI-NEXT: s_mov_b64 exec, s[18:19]
-; FIXEDABI-NEXT: v_writelane_b32 v40, s16, 2
; FIXEDABI-NEXT: s_addk_i32 s32, 0x400
; FIXEDABI-NEXT: v_writelane_b32 v40, s30, 0
+; FIXEDABI-NEXT: v_writelane_b32 v41, s16, 0
; FIXEDABI-NEXT: v_writelane_b32 v40, s31, 1
; FIXEDABI-NEXT: s_getpc_b64 s[16:17]
; FIXEDABI-NEXT: s_add_u32 s16, s16, requires_all_inputs@rel32@lo+4
@@ -32,9 +33,10 @@
; FIXEDABI-NEXT: s_swappc_b64 s[30:31], s[16:17]
; FIXEDABI-NEXT: v_readlane_b32 s31, v40, 1
; FIXEDABI-NEXT: v_readlane_b32 s30, v40, 0
-; FIXEDABI-NEXT: v_readlane_b32 s4, v40, 2
+; FIXEDABI-NEXT: v_readlane_b32 s4, v41, 0
; FIXEDABI-NEXT: s_or_saveexec_b64 s[6:7], -1
; FIXEDABI-NEXT: buffer_load_dword v40, off, s[0:3], s33 ; 4-byte Folded Reload
+; FIXEDABI-NEXT: buffer_load_dword v41, off, s[0:3], s33 offset:4 ; 4-byte Folded Reload
; FIXEDABI-NEXT: s_mov_b64 exec, s[6:7]
; FIXEDABI-NEXT: s_addk_i32 s32, 0xfc00
; FIXEDABI-NEXT: s_mov_b32 s33, s4
diff --git a/llvm/test/CodeGen/AMDGPU/amdpal-callable.ll b/llvm/test/CodeGen/AMDGPU/amdpal-callable.ll
--- a/llvm/test/CodeGen/AMDGPU/amdpal-callable.ll
+++ b/llvm/test/CodeGen/AMDGPU/amdpal-callable.ll
@@ -178,19 +178,19 @@
; GFX8-NEXT: .sgpr_count: 0x28{{$}}
; GFX9-NEXT: .sgpr_count: 0x2c{{$}}
; GCN-NEXT: .stack_frame_size_in_bytes: 0x10{{$}}
-; GCN-NEXT: .vgpr_count: 0x2b{{$}}
+; GCN-NEXT: .vgpr_count: 0x2c{{$}}
; GCN-NEXT: no_stack_extern_call_many_args:
; GCN-NEXT: .lds_size: 0{{$}}
; GFX8-NEXT: .sgpr_count: 0x28{{$}}
; GFX9-NEXT: .sgpr_count: 0x2c{{$}}
; GCN-NEXT: .stack_frame_size_in_bytes: 0x90{{$}}
-; GCN-NEXT: .vgpr_count: 0x2b{{$}}
+; GCN-NEXT: .vgpr_count: 0x2c{{$}}
; GCN-NEXT: no_stack_indirect_call:
; GCN-NEXT: .lds_size: 0{{$}}
; GFX8-NEXT: .sgpr_count: 0x28{{$}}
; GFX9-NEXT: .sgpr_count: 0x2c{{$}}
; GCN-NEXT: .stack_frame_size_in_bytes: 0x10{{$}}
-; GCN-NEXT: .vgpr_count: 0x2b{{$}}
+; GCN-NEXT: .vgpr_count: 0x2c{{$}}
; GCN-NEXT: simple_lds:
; GCN-NEXT: .lds_size: 0x100{{$}}
; GCN-NEXT: .sgpr_count: 0x20{{$}}
@@ -200,7 +200,7 @@
; GCN-NEXT: .lds_size: 0x100{{$}}
; GCN-NEXT: .sgpr_count: 0x28{{$}}
; GCN-NEXT: .stack_frame_size_in_bytes: 0x10{{$}}
-; GCN-NEXT: .vgpr_count: 0x29{{$}}
+; GCN-NEXT: .vgpr_count: 0x2a{{$}}
; GCN-NEXT: simple_stack:
; GCN-NEXT: .lds_size: 0{{$}}
; GCN-NEXT: .sgpr_count: 0x21{{$}}
@@ -216,16 +216,16 @@
; GFX8-NEXT: .sgpr_count: 0x28{{$}}
; GFX9-NEXT: .sgpr_count: 0x2c{{$}}
; GCN-NEXT: .stack_frame_size_in_bytes: 0x20{{$}}
-; GCN-NEXT: .vgpr_count: 0x2b{{$}}
+; GCN-NEXT: .vgpr_count: 0x2c{{$}}
; GCN-NEXT: simple_stack_indirect_call:
; GCN-NEXT: .lds_size: 0{{$}}
; GFX8-NEXT: .sgpr_count: 0x28{{$}}
; GFX9-NEXT: .sgpr_count: 0x2c{{$}}
-; GCN-NEXT: .stack_frame_size_in_bytes: 0x20{{$}}
-; GCN-NEXT: .vgpr_count: 0x2b{{$}}
+; GCN-NEXT: .stack_frame_size_in_bytes: 0x30{{$}}
+; GCN-NEXT: .vgpr_count: 0x2c{{$}}
; GCN-NEXT: simple_stack_recurse:
; GCN-NEXT: .lds_size: 0{{$}}
; GCN-NEXT: .sgpr_count: 0x28{{$}}
; GCN-NEXT: .stack_frame_size_in_bytes: 0x20{{$}}
-; GCN-NEXT: .vgpr_count: 0x2a{{$}}
+; GCN-NEXT: .vgpr_count: 0x2b{{$}}
; GCN-NEXT: ...
diff --git a/llvm/test/CodeGen/AMDGPU/bf16.ll b/llvm/test/CodeGen/AMDGPU/bf16.ll
--- a/llvm/test/CodeGen/AMDGPU/bf16.ll
+++ b/llvm/test/CodeGen/AMDGPU/bf16.ll
@@ -1371,7 +1371,7 @@
; GCN-LABEL: test_call:
; GCN: ; %bb.0: ; %entry
; GCN-NEXT: s_waitcnt vmcnt(0) expcnt(0) lgkmcnt(0)
-; GCN-NEXT: s_mov_b32 s8, s33
+; GCN-NEXT: s_mov_b32 s10, s33
; GCN-NEXT: s_mov_b32 s33, s32
; GCN-NEXT: s_xor_saveexec_b64 s[4:5], -1
; GCN-NEXT: buffer_store_dword v2, off, s[0:3], s33 ; 4-byte Folded Spill
@@ -1395,14 +1395,14 @@
; GCN-NEXT: buffer_load_dword v2, off, s[0:3], s33 ; 4-byte Folded Reload
; GCN-NEXT: s_mov_b64 exec, s[4:5]
; GCN-NEXT: s_addk_i32 s32, 0xfc00
-; GCN-NEXT: s_mov_b32 s33, s8
+; GCN-NEXT: s_mov_b32 s33, s10
; GCN-NEXT: s_waitcnt vmcnt(0) expcnt(0)
; GCN-NEXT: s_setpc_b64 s[30:31]
;
; GFX7-LABEL: test_call:
; GFX7: ; %bb.0: ; %entry
; GFX7-NEXT: s_waitcnt vmcnt(0) expcnt(0) lgkmcnt(0)
-; GFX7-NEXT: s_mov_b32 s8, s33
+; GFX7-NEXT: s_mov_b32 s10, s33
; GFX7-NEXT: s_mov_b32 s33, s32
; GFX7-NEXT: s_xor_saveexec_b64 s[4:5], -1
; GFX7-NEXT: buffer_store_dword v2, off, s[0:3], s33 ; 4-byte Folded Spill
@@ -1425,14 +1425,14 @@
; GFX7-NEXT: buffer_load_dword v2, off, s[0:3], s33 ; 4-byte Folded Reload
; GFX7-NEXT: s_mov_b64 exec, s[4:5]
; GFX7-NEXT: s_addk_i32 s32, 0xfc00
-; GFX7-NEXT: s_mov_b32 s33, s8
+; GFX7-NEXT: s_mov_b32 s33, s10
; GFX7-NEXT: s_waitcnt vmcnt(0)
; GFX7-NEXT: s_setpc_b64 s[30:31]
;
; GFX8-LABEL: test_call:
; GFX8: ; %bb.0: ; %entry
; GFX8-NEXT: s_waitcnt vmcnt(0) expcnt(0) lgkmcnt(0)
-; GFX8-NEXT: s_mov_b32 s6, s33
+; GFX8-NEXT: s_mov_b32 s8, s33
; GFX8-NEXT: s_mov_b32 s33, s32
; GFX8-NEXT: s_xor_saveexec_b64 s[4:5], -1
; GFX8-NEXT: buffer_store_dword v2, off, s[0:3], s33 ; 4-byte Folded Spill
@@ -1455,14 +1455,14 @@
; GFX8-NEXT: buffer_load_dword v2, off, s[0:3], s33 ; 4-byte Folded Reload
; GFX8-NEXT: s_mov_b64 exec, s[4:5]
; GFX8-NEXT: s_addk_i32 s32, 0xfc00
-; GFX8-NEXT: s_mov_b32 s33, s6
+; GFX8-NEXT: s_mov_b32 s33, s8
; GFX8-NEXT: s_waitcnt vmcnt(0)
; GFX8-NEXT: s_setpc_b64 s[30:31]
;
; GFX9-LABEL: test_call:
; GFX9: ; %bb.0: ; %entry
; GFX9-NEXT: s_waitcnt vmcnt(0) expcnt(0) lgkmcnt(0)
-; GFX9-NEXT: s_mov_b32 s6, s33
+; GFX9-NEXT: s_mov_b32 s8, s33
; GFX9-NEXT: s_mov_b32 s33, s32
; GFX9-NEXT: s_xor_saveexec_b64 s[4:5], -1
; GFX9-NEXT: buffer_store_dword v2, off, s[0:3], s33 ; 4-byte Folded Spill
@@ -1484,14 +1484,14 @@
; GFX9-NEXT: buffer_load_dword v2, off, s[0:3], s33 ; 4-byte Folded Reload
; GFX9-NEXT: s_mov_b64 exec, s[4:5]
; GFX9-NEXT: s_addk_i32 s32, 0xfc00
-; GFX9-NEXT: s_mov_b32 s33, s6
+; GFX9-NEXT: s_mov_b32 s33, s8
; GFX9-NEXT: s_waitcnt vmcnt(0)
; GFX9-NEXT: s_setpc_b64 s[30:31]
;
; GFX10-LABEL: test_call:
; GFX10: ; %bb.0: ; %entry
; GFX10-NEXT: s_waitcnt vmcnt(0) expcnt(0) lgkmcnt(0)
-; GFX10-NEXT: s_mov_b32 s6, s33
+; GFX10-NEXT: s_mov_b32 s7, s33
; GFX10-NEXT: s_mov_b32 s33, s32
; GFX10-NEXT: s_xor_saveexec_b32 s4, -1
; GFX10-NEXT: buffer_store_dword v2, off, s[0:3], s33 ; 4-byte Folded Spill
@@ -1515,7 +1515,7 @@
; GFX10-NEXT: s_waitcnt_depctr 0xffe3
; GFX10-NEXT: s_mov_b32 exec_lo, s4
; GFX10-NEXT: s_addk_i32 s32, 0xfe00
-; GFX10-NEXT: s_mov_b32 s33, s6
+; GFX10-NEXT: s_mov_b32 s33, s7
; GFX10-NEXT: s_waitcnt vmcnt(0)
; GFX10-NEXT: s_setpc_b64 s[30:31]
entry:
@@ -1528,7 +1528,7 @@
; GCN-LABEL: test_call_v2bf16:
; GCN: ; %bb.0: ; %entry
; GCN-NEXT: s_waitcnt vmcnt(0) expcnt(0) lgkmcnt(0)
-; GCN-NEXT: s_mov_b32 s8, s33
+; GCN-NEXT: s_mov_b32 s10, s33
; GCN-NEXT: s_mov_b32 s33, s32
; GCN-NEXT: s_xor_saveexec_b64 s[4:5], -1
; GCN-NEXT: buffer_store_dword v3, off, s[0:3], s33 ; 4-byte Folded Spill
@@ -1556,14 +1556,14 @@
; GCN-NEXT: buffer_load_dword v3, off, s[0:3], s33 ; 4-byte Folded Reload
; GCN-NEXT: s_mov_b64 exec, s[4:5]
; GCN-NEXT: s_addk_i32 s32, 0xfc00
-; GCN-NEXT: s_mov_b32 s33, s8
+; GCN-NEXT: s_mov_b32 s33, s10
; GCN-NEXT: s_waitcnt vmcnt(0) expcnt(0)
; GCN-NEXT: s_setpc_b64 s[30:31]
;
; GFX7-LABEL: test_call_v2bf16:
; GFX7: ; %bb.0: ; %entry
; GFX7-NEXT: s_waitcnt vmcnt(0) expcnt(0) lgkmcnt(0)
-; GFX7-NEXT: s_mov_b32 s8, s33
+; GFX7-NEXT: s_mov_b32 s10, s33
; GFX7-NEXT: s_mov_b32 s33, s32
; GFX7-NEXT: s_xor_saveexec_b64 s[4:5], -1
; GFX7-NEXT: buffer_store_dword v3, off, s[0:3], s33 ; 4-byte Folded Spill
@@ -1590,14 +1590,14 @@
; GFX7-NEXT: buffer_load_dword v3, off, s[0:3], s33 ; 4-byte Folded Reload
; GFX7-NEXT: s_mov_b64 exec, s[4:5]
; GFX7-NEXT: s_addk_i32 s32, 0xfc00
-; GFX7-NEXT: s_mov_b32 s33, s8
+; GFX7-NEXT: s_mov_b32 s33, s10
; GFX7-NEXT: s_waitcnt vmcnt(0)
; GFX7-NEXT: s_setpc_b64 s[30:31]
;
; GFX8-LABEL: test_call_v2bf16:
; GFX8: ; %bb.0: ; %entry
; GFX8-NEXT: s_waitcnt vmcnt(0) expcnt(0) lgkmcnt(0)
-; GFX8-NEXT: s_mov_b32 s6, s33
+; GFX8-NEXT: s_mov_b32 s8, s33
; GFX8-NEXT: s_mov_b32 s33, s32
; GFX8-NEXT: s_xor_saveexec_b64 s[4:5], -1
; GFX8-NEXT: buffer_store_dword v2, off, s[0:3], s33 ; 4-byte Folded Spill
@@ -1619,14 +1619,14 @@
; GFX8-NEXT: buffer_load_dword v2, off, s[0:3], s33 ; 4-byte Folded Reload
; GFX8-NEXT: s_mov_b64 exec, s[4:5]
; GFX8-NEXT: s_addk_i32 s32, 0xfc00
-; GFX8-NEXT: s_mov_b32 s33, s6
+; GFX8-NEXT: s_mov_b32 s33, s8
; GFX8-NEXT: s_waitcnt vmcnt(0)
; GFX8-NEXT: s_setpc_b64 s[30:31]
;
; GFX9-LABEL: test_call_v2bf16:
; GFX9: ; %bb.0: ; %entry
; GFX9-NEXT: s_waitcnt vmcnt(0) expcnt(0) lgkmcnt(0)
-; GFX9-NEXT: s_mov_b32 s6, s33
+; GFX9-NEXT: s_mov_b32 s8, s33
; GFX9-NEXT: s_mov_b32 s33, s32
; GFX9-NEXT: s_xor_saveexec_b64 s[4:5], -1
; GFX9-NEXT: buffer_store_dword v2, off, s[0:3], s33 ; 4-byte Folded Spill
@@ -1648,14 +1648,14 @@
; GFX9-NEXT: buffer_load_dword v2, off, s[0:3], s33 ; 4-byte Folded Reload
; GFX9-NEXT: s_mov_b64 exec, s[4:5]
; GFX9-NEXT: s_addk_i32 s32, 0xfc00
-; GFX9-NEXT: s_mov_b32 s33, s6
+; GFX9-NEXT: s_mov_b32 s33, s8
; GFX9-NEXT: s_waitcnt vmcnt(0)
; GFX9-NEXT: s_setpc_b64 s[30:31]
;
; GFX10-LABEL: test_call_v2bf16:
; GFX10: ; %bb.0: ; %entry
; GFX10-NEXT: s_waitcnt vmcnt(0) expcnt(0) lgkmcnt(0)
-; GFX10-NEXT: s_mov_b32 s6, s33
+; GFX10-NEXT: s_mov_b32 s7, s33
; GFX10-NEXT: s_mov_b32 s33, s32
; GFX10-NEXT: s_xor_saveexec_b32 s4, -1
; GFX10-NEXT: buffer_store_dword v2, off, s[0:3], s33 ; 4-byte Folded Spill
@@ -1679,7 +1679,7 @@
; GFX10-NEXT: s_waitcnt_depctr 0xffe3
; GFX10-NEXT: s_mov_b32 exec_lo, s4
; GFX10-NEXT: s_addk_i32 s32, 0xfe00
-; GFX10-NEXT: s_mov_b32 s33, s6
+; GFX10-NEXT: s_mov_b32 s33, s7
; GFX10-NEXT: s_waitcnt vmcnt(0)
; GFX10-NEXT: s_setpc_b64 s[30:31]
entry:
@@ -1692,7 +1692,7 @@
; GCN-LABEL: test_call_v3bf16:
; GCN: ; %bb.0: ; %entry
; GCN-NEXT: s_waitcnt vmcnt(0) expcnt(0) lgkmcnt(0)
-; GCN-NEXT: s_mov_b32 s8, s33
+; GCN-NEXT: s_mov_b32 s10, s33
; GCN-NEXT: s_mov_b32 s33, s32
; GCN-NEXT: s_xor_saveexec_b64 s[4:5], -1
; GCN-NEXT: buffer_store_dword v4, off, s[0:3], s33 ; 4-byte Folded Spill
@@ -1721,14 +1721,14 @@
; GCN-NEXT: buffer_load_dword v4, off, s[0:3], s33 ; 4-byte Folded Reload
; GCN-NEXT: s_mov_b64 exec, s[4:5]
; GCN-NEXT: s_addk_i32 s32, 0xfc00
-; GCN-NEXT: s_mov_b32 s33, s8
+; GCN-NEXT: s_mov_b32 s33, s10
; GCN-NEXT: s_waitcnt vmcnt(0) expcnt(0)
; GCN-NEXT: s_setpc_b64 s[30:31]
;
; GFX7-LABEL: test_call_v3bf16:
; GFX7: ; %bb.0: ; %entry
; GFX7-NEXT: s_waitcnt vmcnt(0) expcnt(0) lgkmcnt(0)
-; GFX7-NEXT: s_mov_b32 s8, s33
+; GFX7-NEXT: s_mov_b32 s10, s33
; GFX7-NEXT: s_mov_b32 s33, s32
; GFX7-NEXT: s_xor_saveexec_b64 s[4:5], -1
; GFX7-NEXT: buffer_store_dword v4, off, s[0:3], s33 ; 4-byte Folded Spill
@@ -1756,14 +1756,14 @@
; GFX7-NEXT: buffer_load_dword v4, off, s[0:3], s33 ; 4-byte Folded Reload
; GFX7-NEXT: s_mov_b64 exec, s[4:5]
; GFX7-NEXT: s_addk_i32 s32, 0xfc00
-; GFX7-NEXT: s_mov_b32 s33, s8
+; GFX7-NEXT: s_mov_b32 s33, s10
; GFX7-NEXT: s_waitcnt vmcnt(0)
; GFX7-NEXT: s_setpc_b64 s[30:31]
;
; GFX8-LABEL: test_call_v3bf16:
; GFX8: ; %bb.0: ; %entry
; GFX8-NEXT: s_waitcnt vmcnt(0) expcnt(0) lgkmcnt(0)
-; GFX8-NEXT: s_mov_b32 s6, s33
+; GFX8-NEXT: s_mov_b32 s8, s33
; GFX8-NEXT: s_mov_b32 s33, s32
; GFX8-NEXT: s_xor_saveexec_b64 s[4:5], -1
; GFX8-NEXT: buffer_store_dword v3, off, s[0:3], s33 ; 4-byte Folded Spill
@@ -1789,14 +1789,14 @@
; GFX8-NEXT: buffer_load_dword v3, off, s[0:3], s33 ; 4-byte Folded Reload
; GFX8-NEXT: s_mov_b64 exec, s[4:5]
; GFX8-NEXT: s_addk_i32 s32, 0xfc00
-; GFX8-NEXT: s_mov_b32 s33, s6
+; GFX8-NEXT: s_mov_b32 s33, s8
; GFX8-NEXT: s_waitcnt vmcnt(0)
; GFX8-NEXT: s_setpc_b64 s[30:31]
;
; GFX9-LABEL: test_call_v3bf16:
; GFX9: ; %bb.0: ; %entry
; GFX9-NEXT: s_waitcnt vmcnt(0) expcnt(0) lgkmcnt(0)
-; GFX9-NEXT: s_mov_b32 s6, s33
+; GFX9-NEXT: s_mov_b32 s8, s33
; GFX9-NEXT: s_mov_b32 s33, s32
; GFX9-NEXT: s_xor_saveexec_b64 s[4:5], -1
; GFX9-NEXT: buffer_store_dword v3, off, s[0:3], s33 ; 4-byte Folded Spill
@@ -1824,14 +1824,14 @@
; GFX9-NEXT: buffer_load_dword v3, off, s[0:3], s33 ; 4-byte Folded Reload
; GFX9-NEXT: s_mov_b64 exec, s[4:5]
; GFX9-NEXT: s_addk_i32 s32, 0xfc00
-; GFX9-NEXT: s_mov_b32 s33, s6
+; GFX9-NEXT: s_mov_b32 s33, s8
; GFX9-NEXT: s_waitcnt vmcnt(0)
; GFX9-NEXT: s_setpc_b64 s[30:31]
;
; GFX10-LABEL: test_call_v3bf16:
; GFX10: ; %bb.0: ; %entry
; GFX10-NEXT: s_waitcnt vmcnt(0) expcnt(0) lgkmcnt(0)
-; GFX10-NEXT: s_mov_b32 s6, s33
+; GFX10-NEXT: s_mov_b32 s7, s33
; GFX10-NEXT: s_mov_b32 s33, s32
; GFX10-NEXT: s_xor_saveexec_b32 s4, -1
; GFX10-NEXT: buffer_store_dword v3, off, s[0:3], s33 ; 4-byte Folded Spill
@@ -1860,7 +1860,7 @@
; GFX10-NEXT: s_waitcnt_depctr 0xffe3
; GFX10-NEXT: s_mov_b32 exec_lo, s4
; GFX10-NEXT: s_addk_i32 s32, 0xfe00
-; GFX10-NEXT: s_mov_b32 s33, s6
+; GFX10-NEXT: s_mov_b32 s33, s7
; GFX10-NEXT: s_waitcnt vmcnt(0)
; GFX10-NEXT: s_setpc_b64 s[30:31]
entry:
@@ -1873,7 +1873,7 @@
; GCN-LABEL: test_call_v4bf16:
; GCN: ; %bb.0: ; %entry
; GCN-NEXT: s_waitcnt vmcnt(0) expcnt(0) lgkmcnt(0)
-; GCN-NEXT: s_mov_b32 s8, s33
+; GCN-NEXT: s_mov_b32 s10, s33
; GCN-NEXT: s_mov_b32 s33, s32
; GCN-NEXT: s_xor_saveexec_b64 s[4:5], -1
; GCN-NEXT: buffer_store_dword v5, off, s[0:3], s33 ; 4-byte Folded Spill
@@ -1909,14 +1909,14 @@
; GCN-NEXT: buffer_load_dword v5, off, s[0:3], s33 ; 4-byte Folded Reload
; GCN-NEXT: s_mov_b64 exec, s[4:5]
; GCN-NEXT: s_addk_i32 s32, 0xfc00
-; GCN-NEXT: s_mov_b32 s33, s8
+; GCN-NEXT: s_mov_b32 s33, s10
; GCN-NEXT: s_waitcnt vmcnt(0) expcnt(0)
; GCN-NEXT: s_setpc_b64 s[30:31]
;
; GFX7-LABEL: test_call_v4bf16:
; GFX7: ; %bb.0: ; %entry
; GFX7-NEXT: s_waitcnt vmcnt(0) expcnt(0) lgkmcnt(0)
-; GFX7-NEXT: s_mov_b32 s8, s33
+; GFX7-NEXT: s_mov_b32 s10, s33
; GFX7-NEXT: s_mov_b32 s33, s32
; GFX7-NEXT: s_xor_saveexec_b64 s[4:5], -1
; GFX7-NEXT: buffer_store_dword v5, off, s[0:3], s33 ; 4-byte Folded Spill
@@ -1951,14 +1951,14 @@
; GFX7-NEXT: buffer_load_dword v5, off, s[0:3], s33 ; 4-byte Folded Reload
; GFX7-NEXT: s_mov_b64 exec, s[4:5]
; GFX7-NEXT: s_addk_i32 s32, 0xfc00
-; GFX7-NEXT: s_mov_b32 s33, s8
+; GFX7-NEXT: s_mov_b32 s33, s10
; GFX7-NEXT: s_waitcnt vmcnt(0)
; GFX7-NEXT: s_setpc_b64 s[30:31]
;
; GFX8-LABEL: test_call_v4bf16:
; GFX8: ; %bb.0: ; %entry
; GFX8-NEXT: s_waitcnt vmcnt(0) expcnt(0) lgkmcnt(0)
-; GFX8-NEXT: s_mov_b32 s6, s33
+; GFX8-NEXT: s_mov_b32 s8, s33
; GFX8-NEXT: s_mov_b32 s33, s32
; GFX8-NEXT: s_xor_saveexec_b64 s[4:5], -1
; GFX8-NEXT: buffer_store_dword v3, off, s[0:3], s33 ; 4-byte Folded Spill
@@ -1991,14 +1991,14 @@
; GFX8-NEXT: buffer_load_dword v3, off, s[0:3], s33 ; 4-byte Folded Reload
; GFX8-NEXT: s_mov_b64 exec, s[4:5]
; GFX8-NEXT: s_addk_i32 s32, 0xfc00
-; GFX8-NEXT: s_mov_b32 s33, s6
+; GFX8-NEXT: s_mov_b32 s33, s8
; GFX8-NEXT: s_waitcnt vmcnt(0)
; GFX8-NEXT: s_setpc_b64 s[30:31]
;
; GFX9-LABEL: test_call_v4bf16:
; GFX9: ; %bb.0: ; %entry
; GFX9-NEXT: s_waitcnt vmcnt(0) expcnt(0) lgkmcnt(0)
-; GFX9-NEXT: s_mov_b32 s6, s33
+; GFX9-NEXT: s_mov_b32 s8, s33
; GFX9-NEXT: s_mov_b32 s33, s32
; GFX9-NEXT: s_xor_saveexec_b64 s[4:5], -1
; GFX9-NEXT: buffer_store_dword v3, off, s[0:3], s33 ; 4-byte Folded Spill
@@ -2026,14 +2026,14 @@
; GFX9-NEXT: buffer_load_dword v3, off, s[0:3], s33 ; 4-byte Folded Reload
; GFX9-NEXT: s_mov_b64 exec, s[4:5]
; GFX9-NEXT: s_addk_i32 s32, 0xfc00
-; GFX9-NEXT: s_mov_b32 s33, s6
+; GFX9-NEXT: s_mov_b32 s33, s8
; GFX9-NEXT: s_waitcnt vmcnt(0)
; GFX9-NEXT: s_setpc_b64 s[30:31]
;
; GFX10-LABEL: test_call_v4bf16:
; GFX10: ; %bb.0: ; %entry
; GFX10-NEXT: s_waitcnt vmcnt(0) expcnt(0) lgkmcnt(0)
-; GFX10-NEXT: s_mov_b32 s6, s33
+; GFX10-NEXT: s_mov_b32 s7, s33
; GFX10-NEXT: s_mov_b32 s33, s32
; GFX10-NEXT: s_xor_saveexec_b32 s4, -1
; GFX10-NEXT: buffer_store_dword v3, off, s[0:3], s33 ; 4-byte Folded Spill
@@ -2063,7 +2063,7 @@
; GFX10-NEXT: s_waitcnt_depctr 0xffe3
; GFX10-NEXT: s_mov_b32 exec_lo, s4
; GFX10-NEXT: s_addk_i32 s32, 0xfe00
-; GFX10-NEXT: s_mov_b32 s33, s6
+; GFX10-NEXT: s_mov_b32 s33, s7
; GFX10-NEXT: s_waitcnt vmcnt(0)
; GFX10-NEXT: s_setpc_b64 s[30:31]
entry:
@@ -2076,7 +2076,7 @@
; GCN-LABEL: test_call_v8bf16:
; GCN: ; %bb.0: ; %entry
; GCN-NEXT: s_waitcnt vmcnt(0) expcnt(0) lgkmcnt(0)
-; GCN-NEXT: s_mov_b32 s8, s33
+; GCN-NEXT: s_mov_b32 s10, s33
; GCN-NEXT: s_mov_b32 s33, s32
; GCN-NEXT: s_xor_saveexec_b64 s[4:5], -1
; GCN-NEXT: buffer_store_dword v9, off, s[0:3], s33 ; 4-byte Folded Spill
@@ -2128,14 +2128,14 @@
; GCN-NEXT: buffer_load_dword v9, off, s[0:3], s33 ; 4-byte Folded Reload
; GCN-NEXT: s_mov_b64 exec, s[4:5]
; GCN-NEXT: s_addk_i32 s32, 0xfc00
-; GCN-NEXT: s_mov_b32 s33, s8
+; GCN-NEXT: s_mov_b32 s33, s10
; GCN-NEXT: s_waitcnt vmcnt(0) expcnt(0)
; GCN-NEXT: s_setpc_b64 s[30:31]
;
; GFX7-LABEL: test_call_v8bf16:
; GFX7: ; %bb.0: ; %entry
; GFX7-NEXT: s_waitcnt vmcnt(0) expcnt(0) lgkmcnt(0)
-; GFX7-NEXT: s_mov_b32 s8, s33
+; GFX7-NEXT: s_mov_b32 s10, s33
; GFX7-NEXT: s_mov_b32 s33, s32
; GFX7-NEXT: s_xor_saveexec_b64 s[4:5], -1
; GFX7-NEXT: buffer_store_dword v9, off, s[0:3], s33 ; 4-byte Folded Spill
@@ -2186,14 +2186,14 @@
; GFX7-NEXT: buffer_load_dword v9, off, s[0:3], s33 ; 4-byte Folded Reload
; GFX7-NEXT: s_mov_b64 exec, s[4:5]
; GFX7-NEXT: s_addk_i32 s32, 0xfc00
-; GFX7-NEXT: s_mov_b32 s33, s8
+; GFX7-NEXT: s_mov_b32 s33, s10
; GFX7-NEXT: s_waitcnt vmcnt(0)
; GFX7-NEXT: s_setpc_b64 s[30:31]
;
; GFX8-LABEL: test_call_v8bf16:
; GFX8: ; %bb.0: ; %entry
; GFX8-NEXT: s_waitcnt vmcnt(0) expcnt(0) lgkmcnt(0)
-; GFX8-NEXT: s_mov_b32 s6, s33
+; GFX8-NEXT: s_mov_b32 s8, s33
; GFX8-NEXT: s_mov_b32 s33, s32
; GFX8-NEXT: s_xor_saveexec_b64 s[4:5], -1
; GFX8-NEXT: buffer_store_dword v5, off, s[0:3], s33 ; 4-byte Folded Spill
@@ -2240,14 +2240,14 @@
; GFX8-NEXT: buffer_load_dword v5, off, s[0:3], s33 ; 4-byte Folded Reload
; GFX8-NEXT: s_mov_b64 exec, s[4:5]
; GFX8-NEXT: s_addk_i32 s32, 0xfc00
-; GFX8-NEXT: s_mov_b32 s33, s6
+; GFX8-NEXT: s_mov_b32 s33, s8
; GFX8-NEXT: s_waitcnt vmcnt(0)
; GFX8-NEXT: s_setpc_b64 s[30:31]
;
; GFX9-LABEL: test_call_v8bf16:
; GFX9: ; %bb.0: ; %entry
; GFX9-NEXT: s_waitcnt vmcnt(0) expcnt(0) lgkmcnt(0)
-; GFX9-NEXT: s_mov_b32 s6, s33
+; GFX9-NEXT: s_mov_b32 s8, s33
; GFX9-NEXT: s_mov_b32 s33, s32
; GFX9-NEXT: s_xor_saveexec_b64 s[4:5], -1
; GFX9-NEXT: buffer_store_dword v5, off, s[0:3], s33 ; 4-byte Folded Spill
@@ -2283,14 +2283,14 @@
; GFX9-NEXT: buffer_load_dword v5, off, s[0:3], s33 ; 4-byte Folded Reload
; GFX9-NEXT: s_mov_b64 exec, s[4:5]
; GFX9-NEXT: s_addk_i32 s32, 0xfc00
-; GFX9-NEXT: s_mov_b32 s33, s6
+; GFX9-NEXT: s_mov_b32 s33, s8
; GFX9-NEXT: s_waitcnt vmcnt(0)
; GFX9-NEXT: s_setpc_b64 s[30:31]
;
; GFX10-LABEL: test_call_v8bf16:
; GFX10: ; %bb.0: ; %entry
; GFX10-NEXT: s_waitcnt vmcnt(0) expcnt(0) lgkmcnt(0)
-; GFX10-NEXT: s_mov_b32 s6, s33
+; GFX10-NEXT: s_mov_b32 s7, s33
; GFX10-NEXT: s_mov_b32 s33, s32
; GFX10-NEXT: s_xor_saveexec_b32 s4, -1
; GFX10-NEXT: buffer_store_dword v5, off, s[0:3], s33 ; 4-byte Folded Spill
@@ -2328,7 +2328,7 @@
; GFX10-NEXT: s_waitcnt_depctr 0xffe3
; GFX10-NEXT: s_mov_b32 exec_lo, s4
; GFX10-NEXT: s_addk_i32 s32, 0xfe00
-; GFX10-NEXT: s_mov_b32 s33, s6
+; GFX10-NEXT: s_mov_b32 s33, s7
; GFX10-NEXT: s_waitcnt vmcnt(0)
; GFX10-NEXT: s_setpc_b64 s[30:31]
entry:
@@ -2341,7 +2341,7 @@
; GCN-LABEL: test_call_v16bf16:
; GCN: ; %bb.0: ; %entry
; GCN-NEXT: s_waitcnt vmcnt(0) expcnt(0) lgkmcnt(0)
-; GCN-NEXT: s_mov_b32 s8, s33
+; GCN-NEXT: s_mov_b32 s10, s33
; GCN-NEXT: s_mov_b32 s33, s32
; GCN-NEXT: s_xor_saveexec_b64 s[4:5], -1
; GCN-NEXT: buffer_store_dword v17, off, s[0:3], s33 ; 4-byte Folded Spill
@@ -2425,14 +2425,14 @@
; GCN-NEXT: buffer_load_dword v17, off, s[0:3], s33 ; 4-byte Folded Reload
; GCN-NEXT: s_mov_b64 exec, s[4:5]
; GCN-NEXT: s_addk_i32 s32, 0xfc00
-; GCN-NEXT: s_mov_b32 s33, s8
+; GCN-NEXT: s_mov_b32 s33, s10
; GCN-NEXT: s_waitcnt vmcnt(0) expcnt(0)
; GCN-NEXT: s_setpc_b64 s[30:31]
;
; GFX7-LABEL: test_call_v16bf16:
; GFX7: ; %bb.0: ; %entry
; GFX7-NEXT: s_waitcnt vmcnt(0) expcnt(0) lgkmcnt(0)
-; GFX7-NEXT: s_mov_b32 s8, s33
+; GFX7-NEXT: s_mov_b32 s10, s33
; GFX7-NEXT: s_mov_b32 s33, s32
; GFX7-NEXT: s_xor_saveexec_b64 s[4:5], -1
; GFX7-NEXT: buffer_store_dword v17, off, s[0:3], s33 ; 4-byte Folded Spill
@@ -2515,14 +2515,14 @@
; GFX7-NEXT: buffer_load_dword v17, off, s[0:3], s33 ; 4-byte Folded Reload
; GFX7-NEXT: s_mov_b64 exec, s[4:5]
; GFX7-NEXT: s_addk_i32 s32, 0xfc00
-; GFX7-NEXT: s_mov_b32 s33, s8
+; GFX7-NEXT: s_mov_b32 s33, s10
; GFX7-NEXT: s_waitcnt vmcnt(0)
; GFX7-NEXT: s_setpc_b64 s[30:31]
;
; GFX8-LABEL: test_call_v16bf16:
; GFX8: ; %bb.0: ; %entry
; GFX8-NEXT: s_waitcnt vmcnt(0) expcnt(0) lgkmcnt(0)
-; GFX8-NEXT: s_mov_b32 s6, s33
+; GFX8-NEXT: s_mov_b32 s8, s33
; GFX8-NEXT: s_mov_b32 s33, s32
; GFX8-NEXT: s_xor_saveexec_b64 s[4:5], -1
; GFX8-NEXT: buffer_store_dword v9, off, s[0:3], s33 ; 4-byte Folded Spill
@@ -2597,14 +2597,14 @@
; GFX8-NEXT: buffer_load_dword v9, off, s[0:3], s33 ; 4-byte Folded Reload
; GFX8-NEXT: s_mov_b64 exec, s[4:5]
; GFX8-NEXT: s_addk_i32 s32, 0xfc00
-; GFX8-NEXT: s_mov_b32 s33, s6
+; GFX8-NEXT: s_mov_b32 s33, s8
; GFX8-NEXT: s_waitcnt vmcnt(0)
; GFX8-NEXT: s_setpc_b64 s[30:31]
;
; GFX9-LABEL: test_call_v16bf16:
; GFX9: ; %bb.0: ; %entry
; GFX9-NEXT: s_waitcnt vmcnt(0) expcnt(0) lgkmcnt(0)
-; GFX9-NEXT: s_mov_b32 s6, s33
+; GFX9-NEXT: s_mov_b32 s8, s33
; GFX9-NEXT: s_mov_b32 s33, s32
; GFX9-NEXT: s_xor_saveexec_b64 s[4:5], -1
; GFX9-NEXT: buffer_store_dword v9, off, s[0:3], s33 ; 4-byte Folded Spill
@@ -2656,14 +2656,14 @@
; GFX9-NEXT: buffer_load_dword v9, off, s[0:3], s33 ; 4-byte Folded Reload
; GFX9-NEXT: s_mov_b64 exec, s[4:5]
; GFX9-NEXT: s_addk_i32 s32, 0xfc00
-; GFX9-NEXT: s_mov_b32 s33, s6
+; GFX9-NEXT: s_mov_b32 s33, s8
; GFX9-NEXT: s_waitcnt vmcnt(0)
; GFX9-NEXT: s_setpc_b64 s[30:31]
;
; GFX10-LABEL: test_call_v16bf16:
; GFX10: ; %bb.0: ; %entry
; GFX10-NEXT: s_waitcnt vmcnt(0) expcnt(0) lgkmcnt(0)
-; GFX10-NEXT: s_mov_b32 s6, s33
+; GFX10-NEXT: s_mov_b32 s7, s33
; GFX10-NEXT: s_mov_b32 s33, s32
; GFX10-NEXT: s_xor_saveexec_b32 s4, -1
; GFX10-NEXT: buffer_store_dword v9, off, s[0:3], s33 ; 4-byte Folded Spill
@@ -2717,7 +2717,7 @@
; GFX10-NEXT: s_waitcnt_depctr 0xffe3
; GFX10-NEXT: s_mov_b32 exec_lo, s4
; GFX10-NEXT: s_addk_i32 s32, 0xfe00
-; GFX10-NEXT: s_mov_b32 s33, s6
+; GFX10-NEXT: s_mov_b32 s33, s7
; GFX10-NEXT: s_waitcnt vmcnt(0)
; GFX10-NEXT: s_setpc_b64 s[30:31]
entry:
diff --git a/llvm/test/CodeGen/AMDGPU/call-alias-register-usage-agpr.ll b/llvm/test/CodeGen/AMDGPU/call-alias-register-usage-agpr.ll
--- a/llvm/test/CodeGen/AMDGPU/call-alias-register-usage-agpr.ll
+++ b/llvm/test/CodeGen/AMDGPU/call-alias-register-usage-agpr.ll
@@ -8,12 +8,12 @@
@alias = hidden alias void (), ptr @aliasee_default
; ALL-LABEL: {{^}}kernel:
-; GFX908: .amdhsa_next_free_vgpr 32
-; GFX908-NEXT: .amdhsa_next_free_sgpr 36
+; GFX908: .amdhsa_next_free_vgpr 41
+; GFX908-NEXT: .amdhsa_next_free_sgpr 33
-; GFX90A: .amdhsa_next_free_vgpr 65
-; GFX90A-NEXT: .amdhsa_next_free_sgpr 36
-; GFX90A-NEXT: .amdhsa_accum_offset 32
+; GFX90A: .amdhsa_next_free_vgpr 71
+; GFX90A-NEXT: .amdhsa_next_free_sgpr 33
+; GFX90A-NEXT: .amdhsa_accum_offset 44
define amdgpu_kernel void @kernel() #0 {
bb:
call void @alias() #2
diff --git a/llvm/test/CodeGen/AMDGPU/call-alias-register-usage0.ll b/llvm/test/CodeGen/AMDGPU/call-alias-register-usage0.ll
--- a/llvm/test/CodeGen/AMDGPU/call-alias-register-usage0.ll
+++ b/llvm/test/CodeGen/AMDGPU/call-alias-register-usage0.ll
@@ -8,7 +8,7 @@
; CHECK-LABEL: {{^}}kernel0:
; CHECK: .amdhsa_next_free_vgpr 53
-; CHECK-NEXT: .amdhsa_next_free_sgpr 36
+; CHECK-NEXT: .amdhsa_next_free_sgpr 33
define amdgpu_kernel void @kernel0() #0 {
bb:
call void @alias0() #2
diff --git a/llvm/test/CodeGen/AMDGPU/call-alias-register-usage1.ll b/llvm/test/CodeGen/AMDGPU/call-alias-register-usage1.ll
--- a/llvm/test/CodeGen/AMDGPU/call-alias-register-usage1.ll
+++ b/llvm/test/CodeGen/AMDGPU/call-alias-register-usage1.ll
@@ -9,8 +9,8 @@
; The parent kernel has a higher VGPR usage than the possible callees.
; CHECK-LABEL: {{^}}kernel1:
-; CHECK: .amdhsa_next_free_vgpr 41
-; CHECK-NEXT: .amdhsa_next_free_sgpr 36
+; CHECK: .amdhsa_next_free_vgpr 42
+; CHECK-NEXT: .amdhsa_next_free_sgpr 33
define amdgpu_kernel void @kernel1() #0 {
bb:
call void asm sideeffect "; clobber v40 ", "~{v40}"()
diff --git a/llvm/test/CodeGen/AMDGPU/call-alias-register-usage2.ll b/llvm/test/CodeGen/AMDGPU/call-alias-register-usage2.ll
--- a/llvm/test/CodeGen/AMDGPU/call-alias-register-usage2.ll
+++ b/llvm/test/CodeGen/AMDGPU/call-alias-register-usage2.ll
@@ -8,7 +8,7 @@
; CHECK-LABEL: {{^}}kernel2:
; CHECK: .amdhsa_next_free_vgpr 53
-; CHECK-NEXT: .amdhsa_next_free_sgpr 36
+; CHECK-NEXT: .amdhsa_next_free_sgpr 33
define amdgpu_kernel void @kernel2() #0 {
bb:
call void @alias2() #2
diff --git a/llvm/test/CodeGen/AMDGPU/call-alias-register-usage3.ll b/llvm/test/CodeGen/AMDGPU/call-alias-register-usage3.ll
--- a/llvm/test/CodeGen/AMDGPU/call-alias-register-usage3.ll
+++ b/llvm/test/CodeGen/AMDGPU/call-alias-register-usage3.ll
@@ -8,7 +8,7 @@
; CHECK-LABEL: {{^}}kernel3:
; CHECK: .amdhsa_next_free_vgpr 253
-; CHECK-NEXT: .amdhsa_next_free_sgpr 36
+; CHECK-NEXT: .amdhsa_next_free_sgpr 33
define amdgpu_kernel void @kernel3() #0 {
bb:
call void @alias3() #2
diff --git a/llvm/test/CodeGen/AMDGPU/call-argument-types.ll b/llvm/test/CodeGen/AMDGPU/call-argument-types.ll
--- a/llvm/test/CodeGen/AMDGPU/call-argument-types.ll
+++ b/llvm/test/CodeGen/AMDGPU/call-argument-types.ll
@@ -5823,6 +5823,7 @@
; VI-NEXT: s_mov_b32 s33, s32
; VI-NEXT: s_or_saveexec_b64 s[8:9], -1
; VI-NEXT: buffer_store_dword v40, off, s[0:3], s33 ; 4-byte Folded Spill
+; VI-NEXT: buffer_store_dword v41, off, s[0:3], s33 offset:4 ; 4-byte Folded Spill
; VI-NEXT: s_mov_b64 exec, s[8:9]
; VI-NEXT: s_addk_i32 s32, 0x400
; VI-NEXT: v_mov_b32_e32 v0, 11
@@ -5832,7 +5833,6 @@
; VI-NEXT: v_mov_b32_e32 v0, 13
; VI-NEXT: buffer_store_dword v0, off, s[0:3], s32 offset:8
; VI-NEXT: v_mov_b32_e32 v0, 14
-; VI-NEXT: v_writelane_b32 v40, s4, 2
; VI-NEXT: buffer_store_dword v0, off, s[0:3], s32 offset:12
; VI-NEXT: v_mov_b32_e32 v0, 15
; VI-NEXT: v_writelane_b32 v40, s30, 0
@@ -5868,6 +5868,7 @@
; VI-NEXT: v_mov_b32_e32 v28, 9
; VI-NEXT: v_mov_b32_e32 v29, 9
; VI-NEXT: v_mov_b32_e32 v30, 10
+; VI-NEXT: v_writelane_b32 v41, s4, 0
; VI-NEXT: v_writelane_b32 v40, s31, 1
; VI-NEXT: s_getpc_b64 s[4:5]
; VI-NEXT: s_add_u32 s4, s4, external_void_func_12xv3i32@rel32@lo+4
@@ -5875,9 +5876,10 @@
; VI-NEXT: s_swappc_b64 s[30:31], s[4:5]
; VI-NEXT: v_readlane_b32 s31, v40, 1
; VI-NEXT: v_readlane_b32 s30, v40, 0
-; VI-NEXT: v_readlane_b32 s4, v40, 2
+; VI-NEXT: v_readlane_b32 s4, v41, 0
; VI-NEXT: s_or_saveexec_b64 s[6:7], -1
; VI-NEXT: buffer_load_dword v40, off, s[0:3], s33 ; 4-byte Folded Reload
+; VI-NEXT: buffer_load_dword v41, off, s[0:3], s33 offset:4 ; 4-byte Folded Reload
; VI-NEXT: s_mov_b64 exec, s[6:7]
; VI-NEXT: s_addk_i32 s32, 0xfc00
; VI-NEXT: s_mov_b32 s33, s4
@@ -5891,6 +5893,7 @@
; CI-NEXT: s_mov_b32 s33, s32
; CI-NEXT: s_or_saveexec_b64 s[8:9], -1
; CI-NEXT: buffer_store_dword v40, off, s[0:3], s33 ; 4-byte Folded Spill
+; CI-NEXT: buffer_store_dword v41, off, s[0:3], s33 offset:4 ; 4-byte Folded Spill
; CI-NEXT: s_mov_b64 exec, s[8:9]
; CI-NEXT: s_addk_i32 s32, 0x400
; CI-NEXT: v_mov_b32_e32 v0, 11
@@ -5900,7 +5903,6 @@
; CI-NEXT: v_mov_b32_e32 v0, 13
; CI-NEXT: buffer_store_dword v0, off, s[0:3], s32 offset:8
; CI-NEXT: v_mov_b32_e32 v0, 14
-; CI-NEXT: v_writelane_b32 v40, s4, 2
; CI-NEXT: buffer_store_dword v0, off, s[0:3], s32 offset:12
; CI-NEXT: v_mov_b32_e32 v0, 15
; CI-NEXT: v_writelane_b32 v40, s30, 0
@@ -5936,6 +5938,7 @@
; CI-NEXT: v_mov_b32_e32 v28, 9
; CI-NEXT: v_mov_b32_e32 v29, 9
; CI-NEXT: v_mov_b32_e32 v30, 10
+; CI-NEXT: v_writelane_b32 v41, s4, 0
; CI-NEXT: v_writelane_b32 v40, s31, 1
; CI-NEXT: s_getpc_b64 s[4:5]
; CI-NEXT: s_add_u32 s4, s4, external_void_func_12xv3i32@rel32@lo+4
@@ -5943,9 +5946,10 @@
; CI-NEXT: s_swappc_b64 s[30:31], s[4:5]
; CI-NEXT: v_readlane_b32 s31, v40, 1
; CI-NEXT: v_readlane_b32 s30, v40, 0
-; CI-NEXT: v_readlane_b32 s4, v40, 2
+; CI-NEXT: v_readlane_b32 s4, v41, 0
; CI-NEXT: s_or_saveexec_b64 s[6:7], -1
; CI-NEXT: buffer_load_dword v40, off, s[0:3], s33 ; 4-byte Folded Reload
+; CI-NEXT: buffer_load_dword v41, off, s[0:3], s33 offset:4 ; 4-byte Folded Reload
; CI-NEXT: s_mov_b64 exec, s[6:7]
; CI-NEXT: s_addk_i32 s32, 0xfc00
; CI-NEXT: s_mov_b32 s33, s4
@@ -5959,6 +5963,7 @@
; GFX9-NEXT: s_mov_b32 s33, s32
; GFX9-NEXT: s_or_saveexec_b64 s[8:9], -1
; GFX9-NEXT: buffer_store_dword v40, off, s[0:3], s33 ; 4-byte Folded Spill
+; GFX9-NEXT: buffer_store_dword v41, off, s[0:3], s33 offset:4 ; 4-byte Folded Spill
; GFX9-NEXT: s_mov_b64 exec, s[8:9]
; GFX9-NEXT: s_addk_i32 s32, 0x400
; GFX9-NEXT: v_mov_b32_e32 v0, 11
@@ -5968,7 +5973,6 @@
; GFX9-NEXT: v_mov_b32_e32 v0, 13
; GFX9-NEXT: buffer_store_dword v0, off, s[0:3], s32 offset:8
; GFX9-NEXT: v_mov_b32_e32 v0, 14
-; GFX9-NEXT: v_writelane_b32 v40, s4, 2
; GFX9-NEXT: buffer_store_dword v0, off, s[0:3], s32 offset:12
; GFX9-NEXT: v_mov_b32_e32 v0, 15
; GFX9-NEXT: v_writelane_b32 v40, s30, 0
@@ -6004,6 +6008,7 @@
; GFX9-NEXT: v_mov_b32_e32 v28, 9
; GFX9-NEXT: v_mov_b32_e32 v29, 9
; GFX9-NEXT: v_mov_b32_e32 v30, 10
+; GFX9-NEXT: v_writelane_b32 v41, s4, 0
; GFX9-NEXT: v_writelane_b32 v40, s31, 1
; GFX9-NEXT: s_getpc_b64 s[4:5]
; GFX9-NEXT: s_add_u32 s4, s4, external_void_func_12xv3i32@rel32@lo+4
@@ -6011,9 +6016,10 @@
; GFX9-NEXT: s_swappc_b64 s[30:31], s[4:5]
; GFX9-NEXT: v_readlane_b32 s31, v40, 1
; GFX9-NEXT: v_readlane_b32 s30, v40, 0
-; GFX9-NEXT: v_readlane_b32 s4, v40, 2
+; GFX9-NEXT: v_readlane_b32 s4, v41, 0
; GFX9-NEXT: s_or_saveexec_b64 s[6:7], -1
; GFX9-NEXT: buffer_load_dword v40, off, s[0:3], s33 ; 4-byte Folded Reload
+; GFX9-NEXT: buffer_load_dword v41, off, s[0:3], s33 offset:4 ; 4-byte Folded Reload
; GFX9-NEXT: s_mov_b64 exec, s[6:7]
; GFX9-NEXT: s_addk_i32 s32, 0xfc00
; GFX9-NEXT: s_mov_b32 s33, s4
@@ -6026,15 +6032,17 @@
; GFX11-NEXT: s_mov_b32 s0, s33
; GFX11-NEXT: s_mov_b32 s33, s32
; GFX11-NEXT: s_or_saveexec_b32 s1, -1
-; GFX11-NEXT: scratch_store_b32 off, v40, s33 ; 4-byte Folded Spill
+; GFX11-NEXT: s_clause 0x1
+; GFX11-NEXT: scratch_store_b32 off, v40, s33
+; GFX11-NEXT: scratch_store_b32 off, v41, s33 offset:4
; GFX11-NEXT: s_mov_b32 exec_lo, s1
-; GFX11-NEXT: v_writelane_b32 v40, s0, 2
; GFX11-NEXT: v_dual_mov_b32 v0, 11 :: v_dual_mov_b32 v1, 12
; GFX11-NEXT: v_dual_mov_b32 v2, 13 :: v_dual_mov_b32 v3, 14
; GFX11-NEXT: v_mov_b32_e32 v4, 15
; GFX11-NEXT: s_add_i32 s32, s32, 16
-; GFX11-NEXT: v_writelane_b32 v40, s30, 0
+; GFX11-NEXT: v_writelane_b32 v41, s0, 0
; GFX11-NEXT: s_add_i32 s0, s32, 16
+; GFX11-NEXT: v_writelane_b32 v40, s30, 0
; GFX11-NEXT: scratch_store_b128 off, v[0:3], s32
; GFX11-NEXT: scratch_store_b32 off, v4, s0
; GFX11-NEXT: v_dual_mov_b32 v1, 0 :: v_dual_mov_b32 v0, 0
@@ -6061,9 +6069,11 @@
; GFX11-NEXT: s_swappc_b64 s[30:31], s[0:1]
; GFX11-NEXT: v_readlane_b32 s31, v40, 1
; GFX11-NEXT: v_readlane_b32 s30, v40, 0
-; GFX11-NEXT: v_readlane_b32 s0, v40, 2
+; GFX11-NEXT: v_readlane_b32 s0, v41, 0
; GFX11-NEXT: s_or_saveexec_b32 s1, -1
-; GFX11-NEXT: scratch_load_b32 v40, off, s33 ; 4-byte Folded Reload
+; GFX11-NEXT: s_clause 0x1
+; GFX11-NEXT: scratch_load_b32 v40, off, s33
+; GFX11-NEXT: scratch_load_b32 v41, off, s33 offset:4
; GFX11-NEXT: s_mov_b32 exec_lo, s1
; GFX11-NEXT: s_add_i32 s32, s32, -16
; GFX11-NEXT: s_mov_b32 s33, s0
@@ -6077,6 +6087,7 @@
; HSA-NEXT: s_mov_b32 s33, s32
; HSA-NEXT: s_or_saveexec_b64 s[8:9], -1
; HSA-NEXT: buffer_store_dword v40, off, s[0:3], s33 ; 4-byte Folded Spill
+; HSA-NEXT: buffer_store_dword v41, off, s[0:3], s33 offset:4 ; 4-byte Folded Spill
; HSA-NEXT: s_mov_b64 exec, s[8:9]
; HSA-NEXT: s_addk_i32 s32, 0x400
; HSA-NEXT: v_mov_b32_e32 v0, 11
@@ -6086,7 +6097,6 @@
; HSA-NEXT: v_mov_b32_e32 v0, 13
; HSA-NEXT: buffer_store_dword v0, off, s[0:3], s32 offset:8
; HSA-NEXT: v_mov_b32_e32 v0, 14
-; HSA-NEXT: v_writelane_b32 v40, s4, 2
; HSA-NEXT: buffer_store_dword v0, off, s[0:3], s32 offset:12
; HSA-NEXT: v_mov_b32_e32 v0, 15
; HSA-NEXT: v_writelane_b32 v40, s30, 0
@@ -6122,6 +6132,7 @@
; HSA-NEXT: v_mov_b32_e32 v28, 9
; HSA-NEXT: v_mov_b32_e32 v29, 9
; HSA-NEXT: v_mov_b32_e32 v30, 10
+; HSA-NEXT: v_writelane_b32 v41, s4, 0
; HSA-NEXT: v_writelane_b32 v40, s31, 1
; HSA-NEXT: s_getpc_b64 s[4:5]
; HSA-NEXT: s_add_u32 s4, s4, external_void_func_12xv3i32@rel32@lo+4
@@ -6129,9 +6140,10 @@
; HSA-NEXT: s_swappc_b64 s[30:31], s[4:5]
; HSA-NEXT: v_readlane_b32 s31, v40, 1
; HSA-NEXT: v_readlane_b32 s30, v40, 0
-; HSA-NEXT: v_readlane_b32 s4, v40, 2
+; HSA-NEXT: v_readlane_b32 s4, v41, 0
; HSA-NEXT: s_or_saveexec_b64 s[6:7], -1
; HSA-NEXT: buffer_load_dword v40, off, s[0:3], s33 ; 4-byte Folded Reload
+; HSA-NEXT: buffer_load_dword v41, off, s[0:3], s33 offset:4 ; 4-byte Folded Reload
; HSA-NEXT: s_mov_b64 exec, s[6:7]
; HSA-NEXT: s_addk_i32 s32, 0xfc00
; HSA-NEXT: s_mov_b32 s33, s4
@@ -6162,6 +6174,7 @@
; VI-NEXT: s_mov_b32 s33, s32
; VI-NEXT: s_or_saveexec_b64 s[8:9], -1
; VI-NEXT: buffer_store_dword v40, off, s[0:3], s33 ; 4-byte Folded Spill
+; VI-NEXT: buffer_store_dword v41, off, s[0:3], s33 offset:4 ; 4-byte Folded Spill
; VI-NEXT: s_mov_b64 exec, s[8:9]
; VI-NEXT: s_addk_i32 s32, 0x400
; VI-NEXT: v_mov_b32_e32 v0, 0x41300000
@@ -6171,7 +6184,6 @@
; VI-NEXT: v_mov_b32_e32 v0, 0x41500000
; VI-NEXT: buffer_store_dword v0, off, s[0:3], s32 offset:8
; VI-NEXT: v_mov_b32_e32 v0, 0x41600000
-; VI-NEXT: v_writelane_b32 v40, s4, 2
; VI-NEXT: buffer_store_dword v0, off, s[0:3], s32 offset:12
; VI-NEXT: v_mov_b32_e32 v0, 0x41700000
; VI-NEXT: v_writelane_b32 v40, s30, 0
@@ -6207,6 +6219,7 @@
; VI-NEXT: v_mov_b32_e32 v28, 0x41100000
; VI-NEXT: v_mov_b32_e32 v29, 0x41100000
; VI-NEXT: v_mov_b32_e32 v30, 0x41200000
+; VI-NEXT: v_writelane_b32 v41, s4, 0
; VI-NEXT: v_writelane_b32 v40, s31, 1
; VI-NEXT: s_getpc_b64 s[4:5]
; VI-NEXT: s_add_u32 s4, s4, external_void_func_12xv3f32@rel32@lo+4
@@ -6214,9 +6227,10 @@
; VI-NEXT: s_swappc_b64 s[30:31], s[4:5]
; VI-NEXT: v_readlane_b32 s31, v40, 1
; VI-NEXT: v_readlane_b32 s30, v40, 0
-; VI-NEXT: v_readlane_b32 s4, v40, 2
+; VI-NEXT: v_readlane_b32 s4, v41, 0
; VI-NEXT: s_or_saveexec_b64 s[6:7], -1
; VI-NEXT: buffer_load_dword v40, off, s[0:3], s33 ; 4-byte Folded Reload
+; VI-NEXT: buffer_load_dword v41, off, s[0:3], s33 offset:4 ; 4-byte Folded Reload
; VI-NEXT: s_mov_b64 exec, s[6:7]
; VI-NEXT: s_addk_i32 s32, 0xfc00
; VI-NEXT: s_mov_b32 s33, s4
@@ -6230,6 +6244,7 @@
; CI-NEXT: s_mov_b32 s33, s32
; CI-NEXT: s_or_saveexec_b64 s[8:9], -1
; CI-NEXT: buffer_store_dword v40, off, s[0:3], s33 ; 4-byte Folded Spill
+; CI-NEXT: buffer_store_dword v41, off, s[0:3], s33 offset:4 ; 4-byte Folded Spill
; CI-NEXT: s_mov_b64 exec, s[8:9]
; CI-NEXT: s_addk_i32 s32, 0x400
; CI-NEXT: v_mov_b32_e32 v0, 0x41300000
@@ -6239,7 +6254,6 @@
; CI-NEXT: v_mov_b32_e32 v0, 0x41500000
; CI-NEXT: buffer_store_dword v0, off, s[0:3], s32 offset:8
; CI-NEXT: v_mov_b32_e32 v0, 0x41600000
-; CI-NEXT: v_writelane_b32 v40, s4, 2
; CI-NEXT: buffer_store_dword v0, off, s[0:3], s32 offset:12
; CI-NEXT: v_mov_b32_e32 v0, 0x41700000
; CI-NEXT: v_writelane_b32 v40, s30, 0
@@ -6275,6 +6289,7 @@
; CI-NEXT: v_mov_b32_e32 v28, 0x41100000
; CI-NEXT: v_mov_b32_e32 v29, 0x41100000
; CI-NEXT: v_mov_b32_e32 v30, 0x41200000
+; CI-NEXT: v_writelane_b32 v41, s4, 0
; CI-NEXT: v_writelane_b32 v40, s31, 1
; CI-NEXT: s_getpc_b64 s[4:5]
; CI-NEXT: s_add_u32 s4, s4, external_void_func_12xv3f32@rel32@lo+4
@@ -6282,9 +6297,10 @@
; CI-NEXT: s_swappc_b64 s[30:31], s[4:5]
; CI-NEXT: v_readlane_b32 s31, v40, 1
; CI-NEXT: v_readlane_b32 s30, v40, 0
-; CI-NEXT: v_readlane_b32 s4, v40, 2
+; CI-NEXT: v_readlane_b32 s4, v41, 0
; CI-NEXT: s_or_saveexec_b64 s[6:7], -1
; CI-NEXT: buffer_load_dword v40, off, s[0:3], s33 ; 4-byte Folded Reload
+; CI-NEXT: buffer_load_dword v41, off, s[0:3], s33 offset:4 ; 4-byte Folded Reload
; CI-NEXT: s_mov_b64 exec, s[6:7]
; CI-NEXT: s_addk_i32 s32, 0xfc00
; CI-NEXT: s_mov_b32 s33, s4
@@ -6298,6 +6314,7 @@
; GFX9-NEXT: s_mov_b32 s33, s32
; GFX9-NEXT: s_or_saveexec_b64 s[8:9], -1
; GFX9-NEXT: buffer_store_dword v40, off, s[0:3], s33 ; 4-byte Folded Spill
+; GFX9-NEXT: buffer_store_dword v41, off, s[0:3], s33 offset:4 ; 4-byte Folded Spill
; GFX9-NEXT: s_mov_b64 exec, s[8:9]
; GFX9-NEXT: s_addk_i32 s32, 0x400
; GFX9-NEXT: v_mov_b32_e32 v0, 0x41300000
@@ -6307,7 +6324,6 @@
; GFX9-NEXT: v_mov_b32_e32 v0, 0x41500000
; GFX9-NEXT: buffer_store_dword v0, off, s[0:3], s32 offset:8
; GFX9-NEXT: v_mov_b32_e32 v0, 0x41600000
-; GFX9-NEXT: v_writelane_b32 v40, s4, 2
; GFX9-NEXT: buffer_store_dword v0, off, s[0:3], s32 offset:12
; GFX9-NEXT: v_mov_b32_e32 v0, 0x41700000
; GFX9-NEXT: v_writelane_b32 v40, s30, 0
@@ -6343,6 +6359,7 @@
; GFX9-NEXT: v_mov_b32_e32 v28, 0x41100000
; GFX9-NEXT: v_mov_b32_e32 v29, 0x41100000
; GFX9-NEXT: v_mov_b32_e32 v30, 0x41200000
+; GFX9-NEXT: v_writelane_b32 v41, s4, 0
; GFX9-NEXT: v_writelane_b32 v40, s31, 1
; GFX9-NEXT: s_getpc_b64 s[4:5]
; GFX9-NEXT: s_add_u32 s4, s4, external_void_func_12xv3f32@rel32@lo+4
@@ -6350,9 +6367,10 @@
; GFX9-NEXT: s_swappc_b64 s[30:31], s[4:5]
; GFX9-NEXT: v_readlane_b32 s31, v40, 1
; GFX9-NEXT: v_readlane_b32 s30, v40, 0
-; GFX9-NEXT: v_readlane_b32 s4, v40, 2
+; GFX9-NEXT: v_readlane_b32 s4, v41, 0
; GFX9-NEXT: s_or_saveexec_b64 s[6:7], -1
; GFX9-NEXT: buffer_load_dword v40, off, s[0:3], s33 ; 4-byte Folded Reload
+; GFX9-NEXT: buffer_load_dword v41, off, s[0:3], s33 offset:4 ; 4-byte Folded Reload
; GFX9-NEXT: s_mov_b64 exec, s[6:7]
; GFX9-NEXT: s_addk_i32 s32, 0xfc00
; GFX9-NEXT: s_mov_b32 s33, s4
@@ -6365,17 +6383,19 @@
; GFX11-NEXT: s_mov_b32 s0, s33
; GFX11-NEXT: s_mov_b32 s33, s32
; GFX11-NEXT: s_or_saveexec_b32 s1, -1
-; GFX11-NEXT: scratch_store_b32 off, v40, s33 ; 4-byte Folded Spill
+; GFX11-NEXT: s_clause 0x1
+; GFX11-NEXT: scratch_store_b32 off, v40, s33
+; GFX11-NEXT: scratch_store_b32 off, v41, s33 offset:4
; GFX11-NEXT: s_mov_b32 exec_lo, s1
-; GFX11-NEXT: v_writelane_b32 v40, s0, 2
; GFX11-NEXT: v_mov_b32_e32 v0, 0x41300000
; GFX11-NEXT: v_mov_b32_e32 v1, 0x41400000
; GFX11-NEXT: v_mov_b32_e32 v2, 0x41500000
; GFX11-NEXT: v_mov_b32_e32 v3, 0x41600000
; GFX11-NEXT: v_dual_mov_b32 v4, 0x41700000 :: v_dual_mov_b32 v5, 1.0
; GFX11-NEXT: s_add_i32 s32, s32, 16
-; GFX11-NEXT: v_writelane_b32 v40, s30, 0
+; GFX11-NEXT: v_writelane_b32 v41, s0, 0
; GFX11-NEXT: s_add_i32 s0, s32, 16
+; GFX11-NEXT: v_writelane_b32 v40, s30, 0
; GFX11-NEXT: scratch_store_b128 off, v[0:3], s32
; GFX11-NEXT: scratch_store_b32 off, v4, s0
; GFX11-NEXT: v_dual_mov_b32 v0, 0 :: v_dual_mov_b32 v1, 0
@@ -6404,9 +6424,11 @@
; GFX11-NEXT: s_swappc_b64 s[30:31], s[0:1]
; GFX11-NEXT: v_readlane_b32 s31, v40, 1
; GFX11-NEXT: v_readlane_b32 s30, v40, 0
-; GFX11-NEXT: v_readlane_b32 s0, v40, 2
+; GFX11-NEXT: v_readlane_b32 s0, v41, 0
; GFX11-NEXT: s_or_saveexec_b32 s1, -1
-; GFX11-NEXT: scratch_load_b32 v40, off, s33 ; 4-byte Folded Reload
+; GFX11-NEXT: s_clause 0x1
+; GFX11-NEXT: scratch_load_b32 v40, off, s33
+; GFX11-NEXT: scratch_load_b32 v41, off, s33 offset:4
; GFX11-NEXT: s_mov_b32 exec_lo, s1
; GFX11-NEXT: s_add_i32 s32, s32, -16
; GFX11-NEXT: s_mov_b32 s33, s0
@@ -6420,6 +6442,7 @@
; HSA-NEXT: s_mov_b32 s33, s32
; HSA-NEXT: s_or_saveexec_b64 s[8:9], -1
; HSA-NEXT: buffer_store_dword v40, off, s[0:3], s33 ; 4-byte Folded Spill
+; HSA-NEXT: buffer_store_dword v41, off, s[0:3], s33 offset:4 ; 4-byte Folded Spill
; HSA-NEXT: s_mov_b64 exec, s[8:9]
; HSA-NEXT: s_addk_i32 s32, 0x400
; HSA-NEXT: v_mov_b32_e32 v0, 0x41300000
@@ -6429,7 +6452,6 @@
; HSA-NEXT: v_mov_b32_e32 v0, 0x41500000
; HSA-NEXT: buffer_store_dword v0, off, s[0:3], s32 offset:8
; HSA-NEXT: v_mov_b32_e32 v0, 0x41600000
-; HSA-NEXT: v_writelane_b32 v40, s4, 2
; HSA-NEXT: buffer_store_dword v0, off, s[0:3], s32 offset:12
; HSA-NEXT: v_mov_b32_e32 v0, 0x41700000
; HSA-NEXT: v_writelane_b32 v40, s30, 0
@@ -6465,6 +6487,7 @@
; HSA-NEXT: v_mov_b32_e32 v28, 0x41100000
; HSA-NEXT: v_mov_b32_e32 v29, 0x41100000
; HSA-NEXT: v_mov_b32_e32 v30, 0x41200000
+; HSA-NEXT: v_writelane_b32 v41, s4, 0
; HSA-NEXT: v_writelane_b32 v40, s31, 1
; HSA-NEXT: s_getpc_b64 s[4:5]
; HSA-NEXT: s_add_u32 s4, s4, external_void_func_12xv3f32@rel32@lo+4
@@ -6472,9 +6495,10 @@
; HSA-NEXT: s_swappc_b64 s[30:31], s[4:5]
; HSA-NEXT: v_readlane_b32 s31, v40, 1
; HSA-NEXT: v_readlane_b32 s30, v40, 0
-; HSA-NEXT: v_readlane_b32 s4, v40, 2
+; HSA-NEXT: v_readlane_b32 s4, v41, 0
; HSA-NEXT: s_or_saveexec_b64 s[6:7], -1
; HSA-NEXT: buffer_load_dword v40, off, s[0:3], s33 ; 4-byte Folded Reload
+; HSA-NEXT: buffer_load_dword v41, off, s[0:3], s33 offset:4 ; 4-byte Folded Reload
; HSA-NEXT: s_mov_b64 exec, s[6:7]
; HSA-NEXT: s_addk_i32 s32, 0xfc00
; HSA-NEXT: s_mov_b32 s33, s4
@@ -6505,6 +6529,7 @@
; VI-NEXT: s_mov_b32 s33, s32
; VI-NEXT: s_or_saveexec_b64 s[8:9], -1
; VI-NEXT: buffer_store_dword v40, off, s[0:3], s33 ; 4-byte Folded Spill
+; VI-NEXT: buffer_store_dword v41, off, s[0:3], s33 offset:4 ; 4-byte Folded Spill
; VI-NEXT: s_mov_b64 exec, s[8:9]
; VI-NEXT: s_addk_i32 s32, 0x400
; VI-NEXT: v_mov_b32_e32 v0, 7
@@ -6522,7 +6547,6 @@
; VI-NEXT: v_mov_b32_e32 v0, 13
; VI-NEXT: buffer_store_dword v0, off, s[0:3], s32 offset:24
; VI-NEXT: v_mov_b32_e32 v0, 14
-; VI-NEXT: v_writelane_b32 v40, s4, 2
; VI-NEXT: buffer_store_dword v0, off, s[0:3], s32 offset:28
; VI-NEXT: v_mov_b32_e32 v0, 15
; VI-NEXT: v_writelane_b32 v40, s30, 0
@@ -6558,6 +6582,7 @@
; VI-NEXT: v_mov_b32_e32 v28, 5
; VI-NEXT: v_mov_b32_e32 v29, 5
; VI-NEXT: v_mov_b32_e32 v30, 6
+; VI-NEXT: v_writelane_b32 v41, s4, 0
; VI-NEXT: v_writelane_b32 v40, s31, 1
; VI-NEXT: s_getpc_b64 s[4:5]
; VI-NEXT: s_add_u32 s4, s4, external_void_func_8xv5i32@rel32@lo+4
@@ -6565,9 +6590,10 @@
; VI-NEXT: s_swappc_b64 s[30:31], s[4:5]
; VI-NEXT: v_readlane_b32 s31, v40, 1
; VI-NEXT: v_readlane_b32 s30, v40, 0
-; VI-NEXT: v_readlane_b32 s4, v40, 2
+; VI-NEXT: v_readlane_b32 s4, v41, 0
; VI-NEXT: s_or_saveexec_b64 s[6:7], -1
; VI-NEXT: buffer_load_dword v40, off, s[0:3], s33 ; 4-byte Folded Reload
+; VI-NEXT: buffer_load_dword v41, off, s[0:3], s33 offset:4 ; 4-byte Folded Reload
; VI-NEXT: s_mov_b64 exec, s[6:7]
; VI-NEXT: s_addk_i32 s32, 0xfc00
; VI-NEXT: s_mov_b32 s33, s4
@@ -6581,6 +6607,7 @@
; CI-NEXT: s_mov_b32 s33, s32
; CI-NEXT: s_or_saveexec_b64 s[8:9], -1
; CI-NEXT: buffer_store_dword v40, off, s[0:3], s33 ; 4-byte Folded Spill
+; CI-NEXT: buffer_store_dword v41, off, s[0:3], s33 offset:4 ; 4-byte Folded Spill
; CI-NEXT: s_mov_b64 exec, s[8:9]
; CI-NEXT: s_addk_i32 s32, 0x400
; CI-NEXT: v_mov_b32_e32 v0, 7
@@ -6598,7 +6625,6 @@
; CI-NEXT: v_mov_b32_e32 v0, 13
; CI-NEXT: buffer_store_dword v0, off, s[0:3], s32 offset:24
; CI-NEXT: v_mov_b32_e32 v0, 14
-; CI-NEXT: v_writelane_b32 v40, s4, 2
; CI-NEXT: buffer_store_dword v0, off, s[0:3], s32 offset:28
; CI-NEXT: v_mov_b32_e32 v0, 15
; CI-NEXT: v_writelane_b32 v40, s30, 0
@@ -6634,6 +6660,7 @@
; CI-NEXT: v_mov_b32_e32 v28, 5
; CI-NEXT: v_mov_b32_e32 v29, 5
; CI-NEXT: v_mov_b32_e32 v30, 6
+; CI-NEXT: v_writelane_b32 v41, s4, 0
; CI-NEXT: v_writelane_b32 v40, s31, 1
; CI-NEXT: s_getpc_b64 s[4:5]
; CI-NEXT: s_add_u32 s4, s4, external_void_func_8xv5i32@rel32@lo+4
@@ -6641,9 +6668,10 @@
; CI-NEXT: s_swappc_b64 s[30:31], s[4:5]
; CI-NEXT: v_readlane_b32 s31, v40, 1
; CI-NEXT: v_readlane_b32 s30, v40, 0
-; CI-NEXT: v_readlane_b32 s4, v40, 2
+; CI-NEXT: v_readlane_b32 s4, v41, 0
; CI-NEXT: s_or_saveexec_b64 s[6:7], -1
; CI-NEXT: buffer_load_dword v40, off, s[0:3], s33 ; 4-byte Folded Reload
+; CI-NEXT: buffer_load_dword v41, off, s[0:3], s33 offset:4 ; 4-byte Folded Reload
; CI-NEXT: s_mov_b64 exec, s[6:7]
; CI-NEXT: s_addk_i32 s32, 0xfc00
; CI-NEXT: s_mov_b32 s33, s4
@@ -6657,6 +6685,7 @@
; GFX9-NEXT: s_mov_b32 s33, s32
; GFX9-NEXT: s_or_saveexec_b64 s[8:9], -1
; GFX9-NEXT: buffer_store_dword v40, off, s[0:3], s33 ; 4-byte Folded Spill
+; GFX9-NEXT: buffer_store_dword v41, off, s[0:3], s33 offset:4 ; 4-byte Folded Spill
; GFX9-NEXT: s_mov_b64 exec, s[8:9]
; GFX9-NEXT: s_addk_i32 s32, 0x400
; GFX9-NEXT: v_mov_b32_e32 v0, 7
@@ -6674,7 +6703,6 @@
; GFX9-NEXT: v_mov_b32_e32 v0, 13
; GFX9-NEXT: buffer_store_dword v0, off, s[0:3], s32 offset:24
; GFX9-NEXT: v_mov_b32_e32 v0, 14
-; GFX9-NEXT: v_writelane_b32 v40, s4, 2
; GFX9-NEXT: buffer_store_dword v0, off, s[0:3], s32 offset:28
; GFX9-NEXT: v_mov_b32_e32 v0, 15
; GFX9-NEXT: v_writelane_b32 v40, s30, 0
@@ -6710,6 +6738,7 @@
; GFX9-NEXT: v_mov_b32_e32 v28, 5
; GFX9-NEXT: v_mov_b32_e32 v29, 5
; GFX9-NEXT: v_mov_b32_e32 v30, 6
+; GFX9-NEXT: v_writelane_b32 v41, s4, 0
; GFX9-NEXT: v_writelane_b32 v40, s31, 1
; GFX9-NEXT: s_getpc_b64 s[4:5]
; GFX9-NEXT: s_add_u32 s4, s4, external_void_func_8xv5i32@rel32@lo+4
@@ -6717,9 +6746,10 @@
; GFX9-NEXT: s_swappc_b64 s[30:31], s[4:5]
; GFX9-NEXT: v_readlane_b32 s31, v40, 1
; GFX9-NEXT: v_readlane_b32 s30, v40, 0
-; GFX9-NEXT: v_readlane_b32 s4, v40, 2
+; GFX9-NEXT: v_readlane_b32 s4, v41, 0
; GFX9-NEXT: s_or_saveexec_b64 s[6:7], -1
; GFX9-NEXT: buffer_load_dword v40, off, s[0:3], s33 ; 4-byte Folded Reload
+; GFX9-NEXT: buffer_load_dword v41, off, s[0:3], s33 offset:4 ; 4-byte Folded Reload
; GFX9-NEXT: s_mov_b64 exec, s[6:7]
; GFX9-NEXT: s_addk_i32 s32, 0xfc00
; GFX9-NEXT: s_mov_b32 s33, s4
@@ -6732,15 +6762,17 @@
; GFX11-NEXT: s_mov_b32 s0, s33
; GFX11-NEXT: s_mov_b32 s33, s32
; GFX11-NEXT: s_or_saveexec_b32 s1, -1
-; GFX11-NEXT: scratch_store_b32 off, v40, s33 ; 4-byte Folded Spill
+; GFX11-NEXT: s_clause 0x1
+; GFX11-NEXT: scratch_store_b32 off, v40, s33
+; GFX11-NEXT: scratch_store_b32 off, v41, s33 offset:4
; GFX11-NEXT: s_mov_b32 exec_lo, s1
-; GFX11-NEXT: v_writelane_b32 v40, s0, 2
; GFX11-NEXT: v_dual_mov_b32 v0, 7 :: v_dual_mov_b32 v1, 8
; GFX11-NEXT: v_dual_mov_b32 v2, 9 :: v_dual_mov_b32 v3, 10
; GFX11-NEXT: v_dual_mov_b32 v8, 15 :: v_dual_mov_b32 v5, 12
; GFX11-NEXT: s_add_i32 s32, s32, 16
; GFX11-NEXT: v_dual_mov_b32 v4, 11 :: v_dual_mov_b32 v7, 14
; GFX11-NEXT: v_mov_b32_e32 v6, 13
+; GFX11-NEXT: v_writelane_b32 v41, s0, 0
; GFX11-NEXT: s_add_i32 s0, s32, 32
; GFX11-NEXT: s_add_i32 s1, s32, 16
; GFX11-NEXT: v_writelane_b32 v40, s30, 0
@@ -6772,9 +6804,11 @@
; GFX11-NEXT: s_swappc_b64 s[30:31], s[0:1]
; GFX11-NEXT: v_readlane_b32 s31, v40, 1
; GFX11-NEXT: v_readlane_b32 s30, v40, 0
-; GFX11-NEXT: v_readlane_b32 s0, v40, 2
+; GFX11-NEXT: v_readlane_b32 s0, v41, 0
; GFX11-NEXT: s_or_saveexec_b32 s1, -1
-; GFX11-NEXT: scratch_load_b32 v40, off, s33 ; 4-byte Folded Reload
+; GFX11-NEXT: s_clause 0x1
+; GFX11-NEXT: scratch_load_b32 v40, off, s33
+; GFX11-NEXT: scratch_load_b32 v41, off, s33 offset:4
; GFX11-NEXT: s_mov_b32 exec_lo, s1
; GFX11-NEXT: s_add_i32 s32, s32, -16
; GFX11-NEXT: s_mov_b32 s33, s0
@@ -6788,6 +6822,7 @@
; HSA-NEXT: s_mov_b32 s33, s32
; HSA-NEXT: s_or_saveexec_b64 s[8:9], -1
; HSA-NEXT: buffer_store_dword v40, off, s[0:3], s33 ; 4-byte Folded Spill
+; HSA-NEXT: buffer_store_dword v41, off, s[0:3], s33 offset:4 ; 4-byte Folded Spill
; HSA-NEXT: s_mov_b64 exec, s[8:9]
; HSA-NEXT: s_addk_i32 s32, 0x400
; HSA-NEXT: v_mov_b32_e32 v0, 7
@@ -6805,7 +6840,6 @@
; HSA-NEXT: v_mov_b32_e32 v0, 13
; HSA-NEXT: buffer_store_dword v0, off, s[0:3], s32 offset:24
; HSA-NEXT: v_mov_b32_e32 v0, 14
-; HSA-NEXT: v_writelane_b32 v40, s4, 2
; HSA-NEXT: buffer_store_dword v0, off, s[0:3], s32 offset:28
; HSA-NEXT: v_mov_b32_e32 v0, 15
; HSA-NEXT: v_writelane_b32 v40, s30, 0
@@ -6841,6 +6875,7 @@
; HSA-NEXT: v_mov_b32_e32 v28, 5
; HSA-NEXT: v_mov_b32_e32 v29, 5
; HSA-NEXT: v_mov_b32_e32 v30, 6
+; HSA-NEXT: v_writelane_b32 v41, s4, 0
; HSA-NEXT: v_writelane_b32 v40, s31, 1
; HSA-NEXT: s_getpc_b64 s[4:5]
; HSA-NEXT: s_add_u32 s4, s4, external_void_func_8xv5i32@rel32@lo+4
@@ -6848,9 +6883,10 @@
; HSA-NEXT: s_swappc_b64 s[30:31], s[4:5]
; HSA-NEXT: v_readlane_b32 s31, v40, 1
; HSA-NEXT: v_readlane_b32 s30, v40, 0
-; HSA-NEXT: v_readlane_b32 s4, v40, 2
+; HSA-NEXT: v_readlane_b32 s4, v41, 0
; HSA-NEXT: s_or_saveexec_b64 s[6:7], -1
; HSA-NEXT: buffer_load_dword v40, off, s[0:3], s33 ; 4-byte Folded Reload
+; HSA-NEXT: buffer_load_dword v41, off, s[0:3], s33 offset:4 ; 4-byte Folded Reload
; HSA-NEXT: s_mov_b64 exec, s[6:7]
; HSA-NEXT: s_addk_i32 s32, 0xfc00
; HSA-NEXT: s_mov_b32 s33, s4
@@ -6877,6 +6913,7 @@
; VI-NEXT: s_mov_b32 s33, s32
; VI-NEXT: s_or_saveexec_b64 s[8:9], -1
; VI-NEXT: buffer_store_dword v40, off, s[0:3], s33 ; 4-byte Folded Spill
+; VI-NEXT: buffer_store_dword v41, off, s[0:3], s33 offset:4 ; 4-byte Folded Spill
; VI-NEXT: s_mov_b64 exec, s[8:9]
; VI-NEXT: s_addk_i32 s32, 0x400
; VI-NEXT: v_mov_b32_e32 v0, 0x40e00000
@@ -6894,7 +6931,6 @@
; VI-NEXT: v_mov_b32_e32 v0, 0x41500000
; VI-NEXT: buffer_store_dword v0, off, s[0:3], s32 offset:24
; VI-NEXT: v_mov_b32_e32 v0, 0x41600000
-; VI-NEXT: v_writelane_b32 v40, s4, 2
; VI-NEXT: buffer_store_dword v0, off, s[0:3], s32 offset:28
; VI-NEXT: v_mov_b32_e32 v0, 0x41700000
; VI-NEXT: v_writelane_b32 v40, s30, 0
@@ -6930,6 +6966,7 @@
; VI-NEXT: v_mov_b32_e32 v28, 0x40a00000
; VI-NEXT: v_mov_b32_e32 v29, 0x40a00000
; VI-NEXT: v_mov_b32_e32 v30, 0x40c00000
+; VI-NEXT: v_writelane_b32 v41, s4, 0
; VI-NEXT: v_writelane_b32 v40, s31, 1
; VI-NEXT: s_getpc_b64 s[4:5]
; VI-NEXT: s_add_u32 s4, s4, external_void_func_8xv5f32@rel32@lo+4
@@ -6937,9 +6974,10 @@
; VI-NEXT: s_swappc_b64 s[30:31], s[4:5]
; VI-NEXT: v_readlane_b32 s31, v40, 1
; VI-NEXT: v_readlane_b32 s30, v40, 0
-; VI-NEXT: v_readlane_b32 s4, v40, 2
+; VI-NEXT: v_readlane_b32 s4, v41, 0
; VI-NEXT: s_or_saveexec_b64 s[6:7], -1
; VI-NEXT: buffer_load_dword v40, off, s[0:3], s33 ; 4-byte Folded Reload
+; VI-NEXT: buffer_load_dword v41, off, s[0:3], s33 offset:4 ; 4-byte Folded Reload
; VI-NEXT: s_mov_b64 exec, s[6:7]
; VI-NEXT: s_addk_i32 s32, 0xfc00
; VI-NEXT: s_mov_b32 s33, s4
@@ -6953,6 +6991,7 @@
; CI-NEXT: s_mov_b32 s33, s32
; CI-NEXT: s_or_saveexec_b64 s[8:9], -1
; CI-NEXT: buffer_store_dword v40, off, s[0:3], s33 ; 4-byte Folded Spill
+; CI-NEXT: buffer_store_dword v41, off, s[0:3], s33 offset:4 ; 4-byte Folded Spill
; CI-NEXT: s_mov_b64 exec, s[8:9]
; CI-NEXT: s_addk_i32 s32, 0x400
; CI-NEXT: v_mov_b32_e32 v0, 0x40e00000
@@ -6970,7 +7009,6 @@
; CI-NEXT: v_mov_b32_e32 v0, 0x41500000
; CI-NEXT: buffer_store_dword v0, off, s[0:3], s32 offset:24
; CI-NEXT: v_mov_b32_e32 v0, 0x41600000
-; CI-NEXT: v_writelane_b32 v40, s4, 2
; CI-NEXT: buffer_store_dword v0, off, s[0:3], s32 offset:28
; CI-NEXT: v_mov_b32_e32 v0, 0x41700000
; CI-NEXT: v_writelane_b32 v40, s30, 0
@@ -7006,6 +7044,7 @@
; CI-NEXT: v_mov_b32_e32 v28, 0x40a00000
; CI-NEXT: v_mov_b32_e32 v29, 0x40a00000
; CI-NEXT: v_mov_b32_e32 v30, 0x40c00000
+; CI-NEXT: v_writelane_b32 v41, s4, 0
; CI-NEXT: v_writelane_b32 v40, s31, 1
; CI-NEXT: s_getpc_b64 s[4:5]
; CI-NEXT: s_add_u32 s4, s4, external_void_func_8xv5f32@rel32@lo+4
@@ -7013,9 +7052,10 @@
; CI-NEXT: s_swappc_b64 s[30:31], s[4:5]
; CI-NEXT: v_readlane_b32 s31, v40, 1
; CI-NEXT: v_readlane_b32 s30, v40, 0
-; CI-NEXT: v_readlane_b32 s4, v40, 2
+; CI-NEXT: v_readlane_b32 s4, v41, 0
; CI-NEXT: s_or_saveexec_b64 s[6:7], -1
; CI-NEXT: buffer_load_dword v40, off, s[0:3], s33 ; 4-byte Folded Reload
+; CI-NEXT: buffer_load_dword v41, off, s[0:3], s33 offset:4 ; 4-byte Folded Reload
; CI-NEXT: s_mov_b64 exec, s[6:7]
; CI-NEXT: s_addk_i32 s32, 0xfc00
; CI-NEXT: s_mov_b32 s33, s4
@@ -7029,6 +7069,7 @@
; GFX9-NEXT: s_mov_b32 s33, s32
; GFX9-NEXT: s_or_saveexec_b64 s[8:9], -1
; GFX9-NEXT: buffer_store_dword v40, off, s[0:3], s33 ; 4-byte Folded Spill
+; GFX9-NEXT: buffer_store_dword v41, off, s[0:3], s33 offset:4 ; 4-byte Folded Spill
; GFX9-NEXT: s_mov_b64 exec, s[8:9]
; GFX9-NEXT: s_addk_i32 s32, 0x400
; GFX9-NEXT: v_mov_b32_e32 v0, 0x40e00000
@@ -7046,7 +7087,6 @@
; GFX9-NEXT: v_mov_b32_e32 v0, 0x41500000
; GFX9-NEXT: buffer_store_dword v0, off, s[0:3], s32 offset:24
; GFX9-NEXT: v_mov_b32_e32 v0, 0x41600000
-; GFX9-NEXT: v_writelane_b32 v40, s4, 2
; GFX9-NEXT: buffer_store_dword v0, off, s[0:3], s32 offset:28
; GFX9-NEXT: v_mov_b32_e32 v0, 0x41700000
; GFX9-NEXT: v_writelane_b32 v40, s30, 0
@@ -7082,6 +7122,7 @@
; GFX9-NEXT: v_mov_b32_e32 v28, 0x40a00000
; GFX9-NEXT: v_mov_b32_e32 v29, 0x40a00000
; GFX9-NEXT: v_mov_b32_e32 v30, 0x40c00000
+; GFX9-NEXT: v_writelane_b32 v41, s4, 0
; GFX9-NEXT: v_writelane_b32 v40, s31, 1
; GFX9-NEXT: s_getpc_b64 s[4:5]
; GFX9-NEXT: s_add_u32 s4, s4, external_void_func_8xv5f32@rel32@lo+4
@@ -7089,9 +7130,10 @@
; GFX9-NEXT: s_swappc_b64 s[30:31], s[4:5]
; GFX9-NEXT: v_readlane_b32 s31, v40, 1
; GFX9-NEXT: v_readlane_b32 s30, v40, 0
-; GFX9-NEXT: v_readlane_b32 s4, v40, 2
+; GFX9-NEXT: v_readlane_b32 s4, v41, 0
; GFX9-NEXT: s_or_saveexec_b64 s[6:7], -1
; GFX9-NEXT: buffer_load_dword v40, off, s[0:3], s33 ; 4-byte Folded Reload
+; GFX9-NEXT: buffer_load_dword v41, off, s[0:3], s33 offset:4 ; 4-byte Folded Reload
; GFX9-NEXT: s_mov_b64 exec, s[6:7]
; GFX9-NEXT: s_addk_i32 s32, 0xfc00
; GFX9-NEXT: s_mov_b32 s33, s4
@@ -7104,9 +7146,10 @@
; GFX11-NEXT: s_mov_b32 s0, s33
; GFX11-NEXT: s_mov_b32 s33, s32
; GFX11-NEXT: s_or_saveexec_b32 s1, -1
-; GFX11-NEXT: scratch_store_b32 off, v40, s33 ; 4-byte Folded Spill
+; GFX11-NEXT: s_clause 0x1
+; GFX11-NEXT: scratch_store_b32 off, v40, s33
+; GFX11-NEXT: scratch_store_b32 off, v41, s33 offset:4
; GFX11-NEXT: s_mov_b32 exec_lo, s1
-; GFX11-NEXT: v_writelane_b32 v40, s0, 2
; GFX11-NEXT: v_mov_b32_e32 v0, 0x40e00000
; GFX11-NEXT: v_mov_b32_e32 v1, 0x41000000
; GFX11-NEXT: v_mov_b32_e32 v2, 0x41100000
@@ -7117,6 +7160,7 @@
; GFX11-NEXT: v_mov_b32_e32 v5, 0x41400000
; GFX11-NEXT: v_dual_mov_b32 v6, 0x41500000 :: v_dual_mov_b32 v9, 1.0
; GFX11-NEXT: v_mov_b32_e32 v7, 0x41600000
+; GFX11-NEXT: v_writelane_b32 v41, s0, 0
; GFX11-NEXT: s_add_i32 s0, s32, 32
; GFX11-NEXT: s_add_i32 s1, s32, 16
; GFX11-NEXT: v_writelane_b32 v40, s30, 0
@@ -7147,9 +7191,11 @@
; GFX11-NEXT: s_swappc_b64 s[30:31], s[0:1]
; GFX11-NEXT: v_readlane_b32 s31, v40, 1
; GFX11-NEXT: v_readlane_b32 s30, v40, 0
-; GFX11-NEXT: v_readlane_b32 s0, v40, 2
+; GFX11-NEXT: v_readlane_b32 s0, v41, 0
; GFX11-NEXT: s_or_saveexec_b32 s1, -1
-; GFX11-NEXT: scratch_load_b32 v40, off, s33 ; 4-byte Folded Reload
+; GFX11-NEXT: s_clause 0x1
+; GFX11-NEXT: scratch_load_b32 v40, off, s33
+; GFX11-NEXT: scratch_load_b32 v41, off, s33 offset:4
; GFX11-NEXT: s_mov_b32 exec_lo, s1
; GFX11-NEXT: s_add_i32 s32, s32, -16
; GFX11-NEXT: s_mov_b32 s33, s0
@@ -7163,6 +7209,7 @@
; HSA-NEXT: s_mov_b32 s33, s32
; HSA-NEXT: s_or_saveexec_b64 s[8:9], -1
; HSA-NEXT: buffer_store_dword v40, off, s[0:3], s33 ; 4-byte Folded Spill
+; HSA-NEXT: buffer_store_dword v41, off, s[0:3], s33 offset:4 ; 4-byte Folded Spill
; HSA-NEXT: s_mov_b64 exec, s[8:9]
; HSA-NEXT: s_addk_i32 s32, 0x400
; HSA-NEXT: v_mov_b32_e32 v0, 0x40e00000
@@ -7180,7 +7227,6 @@
; HSA-NEXT: v_mov_b32_e32 v0, 0x41500000
; HSA-NEXT: buffer_store_dword v0, off, s[0:3], s32 offset:24
; HSA-NEXT: v_mov_b32_e32 v0, 0x41600000
-; HSA-NEXT: v_writelane_b32 v40, s4, 2
; HSA-NEXT: buffer_store_dword v0, off, s[0:3], s32 offset:28
; HSA-NEXT: v_mov_b32_e32 v0, 0x41700000
; HSA-NEXT: v_writelane_b32 v40, s30, 0
@@ -7216,6 +7262,7 @@
; HSA-NEXT: v_mov_b32_e32 v28, 0x40a00000
; HSA-NEXT: v_mov_b32_e32 v29, 0x40a00000
; HSA-NEXT: v_mov_b32_e32 v30, 0x40c00000
+; HSA-NEXT: v_writelane_b32 v41, s4, 0
; HSA-NEXT: v_writelane_b32 v40, s31, 1
; HSA-NEXT: s_getpc_b64 s[4:5]
; HSA-NEXT: s_add_u32 s4, s4, external_void_func_8xv5f32@rel32@lo+4
@@ -7223,9 +7270,10 @@
; HSA-NEXT: s_swappc_b64 s[30:31], s[4:5]
; HSA-NEXT: v_readlane_b32 s31, v40, 1
; HSA-NEXT: v_readlane_b32 s30, v40, 0
-; HSA-NEXT: v_readlane_b32 s4, v40, 2
+; HSA-NEXT: v_readlane_b32 s4, v41, 0
; HSA-NEXT: s_or_saveexec_b64 s[6:7], -1
; HSA-NEXT: buffer_load_dword v40, off, s[0:3], s33 ; 4-byte Folded Reload
+; HSA-NEXT: buffer_load_dword v41, off, s[0:3], s33 offset:4 ; 4-byte Folded Reload
; HSA-NEXT: s_mov_b64 exec, s[6:7]
; HSA-NEXT: s_addk_i32 s32, 0xfc00
; HSA-NEXT: s_mov_b32 s33, s4
diff --git a/llvm/test/CodeGen/AMDGPU/call-graph-register-usage.ll b/llvm/test/CodeGen/AMDGPU/call-graph-register-usage.ll
--- a/llvm/test/CodeGen/AMDGPU/call-graph-register-usage.ll
+++ b/llvm/test/CodeGen/AMDGPU/call-graph-register-usage.ll
@@ -15,17 +15,17 @@
; GCN-LABEL: {{^}}indirect_use_vcc:
; GCN: s_mov_b32 s4, s33
-; GCN: v_writelane_b32 v40, s4, 2
+; GCN: v_writelane_b32 v41, s4, 0
; GCN: v_writelane_b32 v40, s30, 0
; GCN: v_writelane_b32 v40, s31, 1
; GCN: s_swappc_b64
; GCN: v_readlane_b32 s31, v40, 1
; GCN: v_readlane_b32 s30, v40, 0
-; GCN: v_readlane_b32 s4, v40, 2
+; GCN: v_readlane_b32 s4, v41, 0
; GCN: s_mov_b32 s33, s4
; GCN: s_setpc_b64 s[30:31]
; GCN: ; NumSgprs: 36
-; GCN: ; NumVgprs: 41
+; GCN: ; NumVgprs: 42
define void @indirect_use_vcc() #1 {
call void @use_vcc()
ret void
@@ -36,7 +36,7 @@
; CI: ; NumSgprs: 38
; VI-NOBUG: ; NumSgprs: 40
; VI-BUG: ; NumSgprs: 96
-; GCN: ; NumVgprs: 41
+; GCN: ; NumVgprs: 42
define amdgpu_kernel void @indirect_2level_use_vcc_kernel(ptr addrspace(1) %out) #0 {
call void @indirect_use_vcc()
ret void
@@ -54,7 +54,7 @@
; GCN-LABEL: {{^}}indirect_use_flat_scratch:
; CI: ; NumSgprs: 38
; VI: ; NumSgprs: 40
-; GCN: ; NumVgprs: 41
+; GCN: ; NumVgprs: 42
define void @indirect_use_flat_scratch() #1 {
call void @use_flat_scratch()
ret void
@@ -65,7 +65,7 @@
; CI: ; NumSgprs: 38
; VI-NOBUG: ; NumSgprs: 40
; VI-BUG: ; NumSgprs: 96
-; GCN: ; NumVgprs: 41
+; GCN: ; NumVgprs: 42
define amdgpu_kernel void @indirect_2level_use_flat_scratch_kernel(ptr addrspace(1) %out) #0 {
call void @indirect_use_flat_scratch()
ret void
@@ -80,7 +80,7 @@
}
; GCN-LABEL: {{^}}indirect_use_10_vgpr:
-; GCN: ; NumVgprs: 41
+; GCN: ; NumVgprs: 42
define void @indirect_use_10_vgpr() #0 {
call void @use_10_vgpr()
ret void
@@ -88,7 +88,7 @@
; GCN-LABEL: {{^}}indirect_2_level_use_10_vgpr:
; GCN: is_dynamic_callstack = 0
-; GCN: ; NumVgprs: 41
+; GCN: ; NumVgprs: 42
define amdgpu_kernel void @indirect_2_level_use_10_vgpr() #0 {
call void @indirect_use_10_vgpr()
ret void
diff --git a/llvm/test/CodeGen/AMDGPU/call-preserved-registers.ll b/llvm/test/CodeGen/AMDGPU/call-preserved-registers.ll
--- a/llvm/test/CodeGen/AMDGPU/call-preserved-registers.ll
+++ b/llvm/test/CodeGen/AMDGPU/call-preserved-registers.ll
@@ -24,11 +24,13 @@
; GCN-LABEL: {{^}}test_func_call_external_void_func_void_clobber_s30_s31_call_external_void_func_void:
; GCN: s_mov_b32 [[FP_SCRATCH_COPY:s[0-9]+]], s33
; MUBUF: buffer_store_dword
+; MUBUF: buffer_store_dword
+; FLATSCR: scratch_store_dword
; FLATSCR: scratch_store_dword
-; GCN: v_writelane_b32 v40, [[FP_SCRATCH_COPY]], 4
; GCN: v_writelane_b32 v40, s30, 0
; GCN: v_writelane_b32 v40, s31, 1
; GCN: v_writelane_b32 v40, s34, 2
+; GCN: v_writelane_b32 v41, [[FP_SCRATCH_COPY]], 0
; GCN: v_writelane_b32 v40, s35, 3
; GCN: s_swappc_b64
@@ -42,8 +44,10 @@
; FLATSCR-DAG: v_readlane_b32 s31, v40, 1
; FLATSCR-DAG: v_readlane_b32 s30, v40, 0
-; GCN: v_readlane_b32 [[FP_SCRATCH_COPY:s[0-9]+]], v40, 4
+; GCN: v_readlane_b32 [[FP_SCRATCH_COPY:s[0-9]+]], v41, 0
; MUBUF: buffer_load_dword
+; MUBUF: buffer_load_dword
+; FLATSCR: scratch_load_dword
; FLATSCR: scratch_load_dword
; GCN: s_mov_b32 s33, [[FP_SCRATCH_COPY]]
; GCN: s_setpc_b64 s[30:31]
@@ -58,17 +62,21 @@
; GCN: s_mov_b32 [[FP_SCRATCH_COPY:s[0-9]+]], s33
; GCN: s_mov_b32 s33, s32
; MUBUF: buffer_store_dword v40
+; MUBUF: buffer_store_dword v41
; FLATSCR: scratch_store_dword off, v40
-; GCN: v_writelane_b32 v40, [[FP_SCRATCH_COPY]], 4
+; FLATSCR: scratch_store_dword off, v41
; MUBUF: s_addk_i32 s32, 0x400
; FLATSCR: s_add_i32 s32, s32, 16
+; GCN: v_writelane_b32 v41, [[FP_SCRATCH_COPY]], 0
; GCN: s_swappc_b64
; GCN-NEXT: s_swappc_b64
-; GCN: v_readlane_b32 [[FP_SCRATCH_COPY:s[0-9]+]], v40, 4
+; GCN: v_readlane_b32 [[FP_SCRATCH_COPY:s[0-9]+]], v41, 0
; MUBUF: buffer_load_dword v40
+; MUBUF: buffer_load_dword v41
; FLATSCR: scratch_load_dword v40
+; FLATSCR: scratch_load_dword v41
; GCN: s_mov_b32 s33, [[FP_SCRATCH_COPY]]
define void @test_func_call_external_void_funcx2() #0 {
call void @external_void_func_void()
diff --git a/llvm/test/CodeGen/AMDGPU/callee-frame-setup.ll b/llvm/test/CodeGen/AMDGPU/callee-frame-setup.ll
--- a/llvm/test/CodeGen/AMDGPU/callee-frame-setup.ll
+++ b/llvm/test/CodeGen/AMDGPU/callee-frame-setup.ll
@@ -89,13 +89,15 @@
; GCN-NEXT: s_mov_b32 s33, s32
; GCN-NEXT: s_or_saveexec_b64 [[COPY_EXEC0:s\[[0-9]+:[0-9]+\]]], -1{{$}}
; MUBUF-NEXT: buffer_store_dword [[CSR_VGPR:v[0-9]+]], off, s[0:3], s33 offset:4 ; 4-byte Folded Spill
+; MUBUF-NEXT: buffer_store_dword [[CSR_VGPR_1:v[0-9]+]], off, s[0:3], s33 offset:8 ; 4-byte Folded Spill
; FLATSCR-NEXT: scratch_store_dword off, [[CSR_VGPR:v[0-9]+]], s33 offset:4 ; 4-byte Folded Spill
+; FLATSCR-NEXT: scratch_store_dword off, [[CSR_VGPR_1:v[0-9]+]], s33 offset:8 ; 4-byte Folded Spill
; GCN-NEXT: s_mov_b64 exec, [[COPY_EXEC0]]
-; GCN: v_writelane_b32 [[CSR_VGPR]], [[FP_SCRATCH_COPY]], 2
; MUBUF-DAG: s_addk_i32 s32, 0x400{{$}}
; FLATSCR-DAG: s_add_i32 s32, s32, 16{{$}}
; GCN-DAG: v_writelane_b32 [[CSR_VGPR]], s30,
; GCN-DAG: v_mov_b32_e32 [[ZERO:v[0-9]+]], 0{{$}}
+; GCN: v_writelane_b32 [[CSR_VGPR_1]], [[FP_SCRATCH_COPY]], 0
; GCN-DAG: v_writelane_b32 [[CSR_VGPR]], s31,
; MUBUF-DAG: buffer_store_dword [[ZERO]], off, s[0:3], s33{{$}}
@@ -106,10 +108,12 @@
; GCN-DAG: v_readlane_b32 s30, [[CSR_VGPR]]
; GCN-DAG: v_readlane_b32 s31, [[CSR_VGPR]]
-; GCN-NEXT: v_readlane_b32 [[FP_SCRATCH_COPY:s[0-9]+]], [[CSR_VGPR]], 2
+; GCN-NEXT: v_readlane_b32 [[FP_SCRATCH_COPY:s[0-9]+]], [[CSR_VGPR_1]], 0
; GCN-NEXT: s_or_saveexec_b64 [[COPY_EXEC1:s\[[0-9]+:[0-9]+\]]], -1{{$}}
; MUBUF-NEXT: buffer_load_dword [[CSR_VGPR]], off, s[0:3], s33 offset:4 ; 4-byte Folded Reload
+; MUBUF-NEXT: buffer_load_dword [[CSR_VGPR_1]], off, s[0:3], s33 offset:8 ; 4-byte Folded Reload
; FLATSCR-NEXT: scratch_load_dword [[CSR_VGPR]], off, s33 offset:4 ; 4-byte Folded Reload
+; FLATSCR-NEXT: scratch_load_dword [[CSR_VGPR_1]], off, s33 offset:8 ; 4-byte Folded Reload
; GCN-NEXT: s_mov_b64 exec, [[COPY_EXEC1]]
; MUBUF: s_addk_i32 s32, 0xfc00{{$}}
; FLATSCR: s_add_i32 s32, s32, -16{{$}}
@@ -136,11 +140,13 @@
; GCN-NEXT: s_mov_b32 s33, s32
; GCN-NEXT: s_or_saveexec_b64 [[COPY_EXEC0:s\[[0-9]+:[0-9]+\]]], -1{{$}}
; MUBUF-NEXT: buffer_store_dword [[CSR_VGPR:v[0-9]+]], off, s[0:3], s33 ; 4-byte Folded Spill
+; MUBUF-NEXT: buffer_store_dword [[CSR_VGPR_1:v[0-9]+]], off, s[0:3], s33 offset:4 ; 4-byte Folded Spill
; FLATSCR-NEXT: scratch_store_dword off, [[CSR_VGPR:v[0-9]+]], s33 ; 4-byte Folded Spill
+; FLATSCR-NEXT: scratch_store_dword off, [[CSR_VGPR_1:v[0-9]+]], s33 offset:4 ; 4-byte Folded Spill
; GCN-NEXT: s_mov_b64 exec, [[COPY_EXEC0]]
; MUBUF-DAG: s_addk_i32 s32, 0x400
; FLATSCR-DAG: s_add_i32 s32, s32, 16
-; GCN-DAG: v_writelane_b32 [[CSR_VGPR]], [[FP_SCRATCH_COPY]], [[FP_SPILL_LANE:[0-9]+]]
+; GCN-DAG: v_writelane_b32 [[CSR_VGPR_1]], [[FP_SCRATCH_COPY]], [[FP_SPILL_LANE:[0-9]+]]
; GCN-DAG: v_writelane_b32 [[CSR_VGPR]], s30, 0
; GCN-DAG: v_writelane_b32 [[CSR_VGPR]], s31, 1
@@ -149,10 +155,12 @@
; GCN-DAG: v_readlane_b32 s30, [[CSR_VGPR]], 0
; GCN-DAG: v_readlane_b32 s31, [[CSR_VGPR]], 1
-; GCN-NEXT: v_readlane_b32 [[FP_SCRATCH_COPY:s[0-9]+]], [[CSR_VGPR]], [[FP_SPILL_LANE]]
+; GCN-NEXT: v_readlane_b32 [[FP_SCRATCH_COPY:s[0-9]+]], [[CSR_VGPR_1]], [[FP_SPILL_LANE]]
; GCN-NEXT: s_or_saveexec_b64 [[COPY_EXEC1:s\[[0-9]+:[0-9]+\]]], -1{{$}}
; MUBUF-NEXT: buffer_load_dword [[CSR_VGPR]], off, s[0:3], s33 ; 4-byte Folded Reload
+; MUBUF-NEXT: buffer_load_dword [[CSR_VGPR_1]], off, s[0:3], s33 offset:4 ; 4-byte Folded Reload
; FLATSCR-NEXT: scratch_load_dword [[CSR_VGPR]], off, s33 ; 4-byte Folded Reload
+; FLATSCR-NEXT: scratch_load_dword [[CSR_VGPR_1]], off, s33 offset:4 ; 4-byte Folded Reload
; GCN-NEXT: s_mov_b64 exec, [[COPY_EXEC1]]
; MUBUF: s_addk_i32 s32, 0xfc00
; FLATSCR: s_add_i32 s32, s32, -16
@@ -623,14 +631,14 @@
; Make sure that the FP save happens after restoring exec from the same
; register.
; GCN-LABEL: {{^}}callee_need_to_spill_fp_to_reg:
-; FLATSCR: s_mov_b32 [[FP_SCRATCH_COPY:s[0-9]+]], s33
+; FLATSCR: s_mov_b32 s2, s33
; FLATSCR: s_mov_b32 s33, s32
; GCN-NOT: v_writelane_b32 v40, s33
-; FLATSCR: s_or_saveexec_b64 [[COPY_EXEC0:s\[[0-9]+:[0-9]+\]]], -1{{$}}
-; FLATSCR: s_mov_b64 exec, [[COPY_EXEC0]]
-; FLATSCR: s_or_saveexec_b64 [[COPY_EXEC1:s\[[0-9]+:[0-9]+\]]], -1{{$}}
+; FLATSCR: s_or_saveexec_b64 s[4:5], -1
+; FLATSCR: s_mov_b64 exec, s[4:5]
+; FLATSCR: s_or_saveexec_b64 s[4:5], -1
; GCN-NOT: v_readlane_b32 s33, v40
-; FLATSCR: s_mov_b32 s33, [[FP_SCRATCH_COPY]]
+; FLATSCR: s_mov_b32 s33, s2
; GCN: s_setpc_b64
define void @callee_need_to_spill_fp_to_reg() #1 {
call void asm sideeffect "; clobber nonpreserved SGPRs and 64 CSRs",
diff --git a/llvm/test/CodeGen/AMDGPU/cf-loop-on-constant.ll b/llvm/test/CodeGen/AMDGPU/cf-loop-on-constant.ll
--- a/llvm/test/CodeGen/AMDGPU/cf-loop-on-constant.ll
+++ b/llvm/test/CodeGen/AMDGPU/cf-loop-on-constant.ll
@@ -30,13 +30,6 @@
;
; GCN_DBG-LABEL: test_loop:
; GCN_DBG: ; %bb.0: ; %entry
-; GCN_DBG-NEXT: s_mov_b32 s12, SCRATCH_RSRC_DWORD0
-; GCN_DBG-NEXT: s_mov_b32 s13, SCRATCH_RSRC_DWORD1
-; GCN_DBG-NEXT: s_mov_b32 s14, -1
-; GCN_DBG-NEXT: s_mov_b32 s15, 0xe8f000
-; GCN_DBG-NEXT: s_add_u32 s12, s12, s11
-; GCN_DBG-NEXT: s_addc_u32 s13, s13, 0
-; GCN_DBG-NEXT: ; implicit-def: $vgpr0
; GCN_DBG-NEXT: s_load_dword s0, s[4:5], 0x9
; GCN_DBG-NEXT: s_waitcnt lgkmcnt(0)
; GCN_DBG-NEXT: v_writelane_b32 v0, s0, 0
@@ -46,25 +39,11 @@
; GCN_DBG-NEXT: s_waitcnt lgkmcnt(0)
; GCN_DBG-NEXT: s_cmp_lg_u32 s1, s2
; GCN_DBG-NEXT: v_writelane_b32 v0, s0, 1
-; GCN_DBG-NEXT: s_mov_b64 s[6:7], exec
-; GCN_DBG-NEXT: s_mov_b64 exec, -1
-; GCN_DBG-NEXT: buffer_store_dword v0, off, s[12:15], 0 offset:4 ; 4-byte Folded Spill
-; GCN_DBG-NEXT: s_mov_b64 exec, s[6:7]
; GCN_DBG-NEXT: s_cbranch_scc1 .LBB0_2
; GCN_DBG-NEXT: ; %bb.1: ; %for.exit
-; GCN_DBG-NEXT: s_or_saveexec_b64 s[6:7], -1
-; GCN_DBG-NEXT: s_waitcnt expcnt(0)
-; GCN_DBG-NEXT: buffer_load_dword v0, off, s[12:15], 0 offset:4 ; 4-byte Folded Reload
-; GCN_DBG-NEXT: s_mov_b64 exec, s[6:7]
-; GCN_DBG-NEXT: ; kill: killed $vgpr0
; GCN_DBG-NEXT: s_endpgm
; GCN_DBG-NEXT: .LBB0_2: ; %for.body
; GCN_DBG-NEXT: ; =>This Inner Loop Header: Depth=1
-; GCN_DBG-NEXT: s_or_saveexec_b64 s[6:7], -1
-; GCN_DBG-NEXT: s_waitcnt expcnt(0)
-; GCN_DBG-NEXT: buffer_load_dword v0, off, s[12:15], 0 offset:4 ; 4-byte Folded Reload
-; GCN_DBG-NEXT: s_mov_b64 exec, s[6:7]
-; GCN_DBG-NEXT: s_waitcnt vmcnt(0)
; GCN_DBG-NEXT: v_readlane_b32 s0, v0, 1
; GCN_DBG-NEXT: v_readlane_b32 s2, v0, 0
; GCN_DBG-NEXT: s_mov_b32 s1, 2
@@ -86,16 +65,8 @@
; GCN_DBG-NEXT: s_mov_b64 s[2:3], -1
; GCN_DBG-NEXT: s_and_b64 vcc, exec, s[2:3]
; GCN_DBG-NEXT: v_writelane_b32 v0, s0, 1
-; GCN_DBG-NEXT: s_or_saveexec_b64 s[6:7], -1
-; GCN_DBG-NEXT: buffer_store_dword v0, off, s[12:15], 0 offset:4 ; 4-byte Folded Spill
-; GCN_DBG-NEXT: s_mov_b64 exec, s[6:7]
; GCN_DBG-NEXT: s_cbranch_vccnz .LBB0_2
; GCN_DBG-NEXT: ; %bb.3: ; %DummyReturnBlock
-; GCN_DBG-NEXT: s_or_saveexec_b64 s[6:7], -1
-; GCN_DBG-NEXT: s_waitcnt expcnt(0)
-; GCN_DBG-NEXT: buffer_load_dword v0, off, s[12:15], 0 offset:4 ; 4-byte Folded Reload
-; GCN_DBG-NEXT: s_mov_b64 exec, s[6:7]
-; GCN_DBG-NEXT: ; kill: killed $vgpr0
; GCN_DBG-NEXT: s_endpgm
entry:
%cmp = icmp eq i32 %n, -1
@@ -138,36 +109,16 @@
;
; GCN_DBG-LABEL: loop_const_true:
; GCN_DBG: ; %bb.0: ; %entry
-; GCN_DBG-NEXT: s_mov_b32 s12, SCRATCH_RSRC_DWORD0
-; GCN_DBG-NEXT: s_mov_b32 s13, SCRATCH_RSRC_DWORD1
-; GCN_DBG-NEXT: s_mov_b32 s14, -1
-; GCN_DBG-NEXT: s_mov_b32 s15, 0xe8f000
-; GCN_DBG-NEXT: s_add_u32 s12, s12, s11
-; GCN_DBG-NEXT: s_addc_u32 s13, s13, 0
-; GCN_DBG-NEXT: ; implicit-def: $vgpr0
; GCN_DBG-NEXT: s_load_dword s0, s[4:5], 0x9
; GCN_DBG-NEXT: s_waitcnt lgkmcnt(0)
; GCN_DBG-NEXT: v_writelane_b32 v0, s0, 0
; GCN_DBG-NEXT: s_mov_b32 s0, 0
; GCN_DBG-NEXT: v_writelane_b32 v0, s0, 1
-; GCN_DBG-NEXT: s_or_saveexec_b64 s[6:7], -1
-; GCN_DBG-NEXT: buffer_store_dword v0, off, s[12:15], 0 offset:4 ; 4-byte Folded Spill
-; GCN_DBG-NEXT: s_mov_b64 exec, s[6:7]
; GCN_DBG-NEXT: s_branch .LBB1_2
; GCN_DBG-NEXT: .LBB1_1: ; %for.exit
-; GCN_DBG-NEXT: s_or_saveexec_b64 s[6:7], -1
-; GCN_DBG-NEXT: s_waitcnt expcnt(0)
-; GCN_DBG-NEXT: buffer_load_dword v0, off, s[12:15], 0 offset:4 ; 4-byte Folded Reload
-; GCN_DBG-NEXT: s_mov_b64 exec, s[6:7]
-; GCN_DBG-NEXT: ; kill: killed $vgpr0
; GCN_DBG-NEXT: s_endpgm
; GCN_DBG-NEXT: .LBB1_2: ; %for.body
; GCN_DBG-NEXT: ; =>This Inner Loop Header: Depth=1
-; GCN_DBG-NEXT: s_or_saveexec_b64 s[6:7], -1
-; GCN_DBG-NEXT: s_waitcnt expcnt(0)
-; GCN_DBG-NEXT: buffer_load_dword v0, off, s[12:15], 0 offset:4 ; 4-byte Folded Reload
-; GCN_DBG-NEXT: s_mov_b64 exec, s[6:7]
-; GCN_DBG-NEXT: s_waitcnt vmcnt(0)
; GCN_DBG-NEXT: v_readlane_b32 s0, v0, 1
; GCN_DBG-NEXT: v_readlane_b32 s2, v0, 0
; GCN_DBG-NEXT: s_mov_b32 s1, 2
@@ -189,9 +140,6 @@
; GCN_DBG-NEXT: s_mov_b64 s[2:3], 0
; GCN_DBG-NEXT: s_and_b64 vcc, exec, s[2:3]
; GCN_DBG-NEXT: v_writelane_b32 v0, s0, 1
-; GCN_DBG-NEXT: s_or_saveexec_b64 s[6:7], -1
-; GCN_DBG-NEXT: buffer_store_dword v0, off, s[12:15], 0 offset:4 ; 4-byte Folded Spill
-; GCN_DBG-NEXT: s_mov_b64 exec, s[6:7]
; GCN_DBG-NEXT: s_cbranch_vccnz .LBB1_1
; GCN_DBG-NEXT: s_branch .LBB1_2
entry:
@@ -226,36 +174,16 @@
;
; GCN_DBG-LABEL: loop_const_false:
; GCN_DBG: ; %bb.0: ; %entry
-; GCN_DBG-NEXT: s_mov_b32 s12, SCRATCH_RSRC_DWORD0
-; GCN_DBG-NEXT: s_mov_b32 s13, SCRATCH_RSRC_DWORD1
-; GCN_DBG-NEXT: s_mov_b32 s14, -1
-; GCN_DBG-NEXT: s_mov_b32 s15, 0xe8f000
-; GCN_DBG-NEXT: s_add_u32 s12, s12, s11
-; GCN_DBG-NEXT: s_addc_u32 s13, s13, 0
-; GCN_DBG-NEXT: ; implicit-def: $vgpr0
; GCN_DBG-NEXT: s_load_dword s0, s[4:5], 0x9
; GCN_DBG-NEXT: s_waitcnt lgkmcnt(0)
; GCN_DBG-NEXT: v_writelane_b32 v0, s0, 0
; GCN_DBG-NEXT: s_mov_b32 s0, 0
; GCN_DBG-NEXT: v_writelane_b32 v0, s0, 1
-; GCN_DBG-NEXT: s_or_saveexec_b64 s[6:7], -1
-; GCN_DBG-NEXT: buffer_store_dword v0, off, s[12:15], 0 offset:4 ; 4-byte Folded Spill
-; GCN_DBG-NEXT: s_mov_b64 exec, s[6:7]
; GCN_DBG-NEXT: s_branch .LBB2_2
; GCN_DBG-NEXT: .LBB2_1: ; %for.exit
-; GCN_DBG-NEXT: s_or_saveexec_b64 s[6:7], -1
-; GCN_DBG-NEXT: s_waitcnt expcnt(0)
-; GCN_DBG-NEXT: buffer_load_dword v0, off, s[12:15], 0 offset:4 ; 4-byte Folded Reload
-; GCN_DBG-NEXT: s_mov_b64 exec, s[6:7]
-; GCN_DBG-NEXT: ; kill: killed $vgpr0
; GCN_DBG-NEXT: s_endpgm
; GCN_DBG-NEXT: .LBB2_2: ; %for.body
; GCN_DBG-NEXT: ; =>This Inner Loop Header: Depth=1
-; GCN_DBG-NEXT: s_or_saveexec_b64 s[6:7], -1
-; GCN_DBG-NEXT: s_waitcnt expcnt(0)
-; GCN_DBG-NEXT: buffer_load_dword v0, off, s[12:15], 0 offset:4 ; 4-byte Folded Reload
-; GCN_DBG-NEXT: s_mov_b64 exec, s[6:7]
-; GCN_DBG-NEXT: s_waitcnt vmcnt(0)
; GCN_DBG-NEXT: v_readlane_b32 s0, v0, 1
; GCN_DBG-NEXT: v_readlane_b32 s2, v0, 0
; GCN_DBG-NEXT: s_mov_b32 s1, 2
@@ -277,9 +205,6 @@
; GCN_DBG-NEXT: s_mov_b64 s[2:3], -1
; GCN_DBG-NEXT: s_and_b64 vcc, exec, s[2:3]
; GCN_DBG-NEXT: v_writelane_b32 v0, s0, 1
-; GCN_DBG-NEXT: s_or_saveexec_b64 s[6:7], -1
-; GCN_DBG-NEXT: buffer_store_dword v0, off, s[12:15], 0 offset:4 ; 4-byte Folded Spill
-; GCN_DBG-NEXT: s_mov_b64 exec, s[6:7]
; GCN_DBG-NEXT: s_cbranch_vccnz .LBB2_1
; GCN_DBG-NEXT: s_branch .LBB2_2
entry:
@@ -315,36 +240,16 @@
;
; GCN_DBG-LABEL: loop_const_undef:
; GCN_DBG: ; %bb.0: ; %entry
-; GCN_DBG-NEXT: s_mov_b32 s12, SCRATCH_RSRC_DWORD0
-; GCN_DBG-NEXT: s_mov_b32 s13, SCRATCH_RSRC_DWORD1
-; GCN_DBG-NEXT: s_mov_b32 s14, -1
-; GCN_DBG-NEXT: s_mov_b32 s15, 0xe8f000
-; GCN_DBG-NEXT: s_add_u32 s12, s12, s11
-; GCN_DBG-NEXT: s_addc_u32 s13, s13, 0
-; GCN_DBG-NEXT: ; implicit-def: $vgpr0
; GCN_DBG-NEXT: s_load_dword s0, s[4:5], 0x9
; GCN_DBG-NEXT: s_waitcnt lgkmcnt(0)
; GCN_DBG-NEXT: v_writelane_b32 v0, s0, 0
; GCN_DBG-NEXT: s_mov_b32 s0, 0
; GCN_DBG-NEXT: v_writelane_b32 v0, s0, 1
-; GCN_DBG-NEXT: s_or_saveexec_b64 s[6:7], -1
-; GCN_DBG-NEXT: buffer_store_dword v0, off, s[12:15], 0 offset:4 ; 4-byte Folded Spill
-; GCN_DBG-NEXT: s_mov_b64 exec, s[6:7]
; GCN_DBG-NEXT: s_branch .LBB3_2
; GCN_DBG-NEXT: .LBB3_1: ; %for.exit
-; GCN_DBG-NEXT: s_or_saveexec_b64 s[6:7], -1
-; GCN_DBG-NEXT: s_waitcnt expcnt(0)
-; GCN_DBG-NEXT: buffer_load_dword v0, off, s[12:15], 0 offset:4 ; 4-byte Folded Reload
-; GCN_DBG-NEXT: s_mov_b64 exec, s[6:7]
-; GCN_DBG-NEXT: ; kill: killed $vgpr0
; GCN_DBG-NEXT: s_endpgm
; GCN_DBG-NEXT: .LBB3_2: ; %for.body
; GCN_DBG-NEXT: ; =>This Inner Loop Header: Depth=1
-; GCN_DBG-NEXT: s_or_saveexec_b64 s[6:7], -1
-; GCN_DBG-NEXT: s_waitcnt expcnt(0)
-; GCN_DBG-NEXT: buffer_load_dword v0, off, s[12:15], 0 offset:4 ; 4-byte Folded Reload
-; GCN_DBG-NEXT: s_mov_b64 exec, s[6:7]
-; GCN_DBG-NEXT: s_waitcnt vmcnt(0)
; GCN_DBG-NEXT: v_readlane_b32 s0, v0, 1
; GCN_DBG-NEXT: v_readlane_b32 s2, v0, 0
; GCN_DBG-NEXT: s_mov_b32 s1, 2
@@ -364,9 +269,6 @@
; GCN_DBG-NEXT: s_mov_b32 s1, 1
; GCN_DBG-NEXT: s_add_i32 s0, s0, s1
; GCN_DBG-NEXT: v_writelane_b32 v0, s0, 1
-; GCN_DBG-NEXT: s_or_saveexec_b64 s[6:7], -1
-; GCN_DBG-NEXT: buffer_store_dword v0, off, s[12:15], 0 offset:4 ; 4-byte Folded Spill
-; GCN_DBG-NEXT: s_mov_b64 exec, s[6:7]
; GCN_DBG-NEXT: s_cbranch_scc1 .LBB3_1
; GCN_DBG-NEXT: s_branch .LBB3_2
entry:
@@ -416,13 +318,6 @@
;
; GCN_DBG-LABEL: loop_arg_0:
; GCN_DBG: ; %bb.0: ; %entry
-; GCN_DBG-NEXT: s_mov_b32 s12, SCRATCH_RSRC_DWORD0
-; GCN_DBG-NEXT: s_mov_b32 s13, SCRATCH_RSRC_DWORD1
-; GCN_DBG-NEXT: s_mov_b32 s14, -1
-; GCN_DBG-NEXT: s_mov_b32 s15, 0xe8f000
-; GCN_DBG-NEXT: s_add_u32 s12, s12, s11
-; GCN_DBG-NEXT: s_addc_u32 s13, s13, 0
-; GCN_DBG-NEXT: ; implicit-def: $vgpr0
; GCN_DBG-NEXT: s_load_dword s0, s[4:5], 0x9
; GCN_DBG-NEXT: s_waitcnt lgkmcnt(0)
; GCN_DBG-NEXT: v_writelane_b32 v0, s0, 0
@@ -440,24 +335,11 @@
; GCN_DBG-NEXT: v_writelane_b32 v0, s1, 2
; GCN_DBG-NEXT: s_mov_b32 s0, 0
; GCN_DBG-NEXT: v_writelane_b32 v0, s0, 3
-; GCN_DBG-NEXT: s_or_saveexec_b64 s[6:7], -1
-; GCN_DBG-NEXT: buffer_store_dword v0, off, s[12:15], 0 offset:4 ; 4-byte Folded Spill
-; GCN_DBG-NEXT: s_mov_b64 exec, s[6:7]
; GCN_DBG-NEXT: s_branch .LBB4_2
; GCN_DBG-NEXT: .LBB4_1: ; %for.exit
-; GCN_DBG-NEXT: s_or_saveexec_b64 s[6:7], -1
-; GCN_DBG-NEXT: s_waitcnt expcnt(0)
-; GCN_DBG-NEXT: buffer_load_dword v0, off, s[12:15], 0 offset:4 ; 4-byte Folded Reload
-; GCN_DBG-NEXT: s_mov_b64 exec, s[6:7]
-; GCN_DBG-NEXT: ; kill: killed $vgpr0
; GCN_DBG-NEXT: s_endpgm
; GCN_DBG-NEXT: .LBB4_2: ; %for.body
; GCN_DBG-NEXT: ; =>This Inner Loop Header: Depth=1
-; GCN_DBG-NEXT: s_or_saveexec_b64 s[6:7], -1
-; GCN_DBG-NEXT: s_waitcnt expcnt(0)
-; GCN_DBG-NEXT: buffer_load_dword v0, off, s[12:15], 0 offset:4 ; 4-byte Folded Reload
-; GCN_DBG-NEXT: s_mov_b64 exec, s[6:7]
-; GCN_DBG-NEXT: s_waitcnt vmcnt(0)
; GCN_DBG-NEXT: v_readlane_b32 s0, v0, 3
; GCN_DBG-NEXT: v_readlane_b32 s2, v0, 1
; GCN_DBG-NEXT: v_readlane_b32 s3, v0, 2
@@ -480,9 +362,6 @@
; GCN_DBG-NEXT: s_add_i32 s0, s0, s1
; GCN_DBG-NEXT: s_and_b64 vcc, exec, s[2:3]
; GCN_DBG-NEXT: v_writelane_b32 v0, s0, 3
-; GCN_DBG-NEXT: s_or_saveexec_b64 s[6:7], -1
-; GCN_DBG-NEXT: buffer_store_dword v0, off, s[12:15], 0 offset:4 ; 4-byte Folded Spill
-; GCN_DBG-NEXT: s_mov_b64 exec, s[6:7]
; GCN_DBG-NEXT: s_cbranch_vccnz .LBB4_1
; GCN_DBG-NEXT: s_branch .LBB4_2
entry:
diff --git a/llvm/test/CodeGen/AMDGPU/collapse-endcf.ll b/llvm/test/CodeGen/AMDGPU/collapse-endcf.ll
--- a/llvm/test/CodeGen/AMDGPU/collapse-endcf.ll
+++ b/llvm/test/CodeGen/AMDGPU/collapse-endcf.ll
@@ -48,37 +48,24 @@
; GCN-O0-NEXT: s_mov_b32 s15, 0xe8f000
; GCN-O0-NEXT: s_add_u32 s12, s12, s11
; GCN-O0-NEXT: s_addc_u32 s13, s13, 0
-; GCN-O0-NEXT: ; implicit-def: $vgpr1
-; GCN-O0-NEXT: v_mov_b32_e32 v1, v0
-; GCN-O0-NEXT: s_or_saveexec_b64 s[8:9], -1
-; GCN-O0-NEXT: buffer_load_dword v0, off, s[12:15], 0 offset:4 ; 4-byte Folded Reload
-; GCN-O0-NEXT: s_mov_b64 exec, s[8:9]
; GCN-O0-NEXT: s_load_dwordx2 s[0:1], s[4:5], 0x9
-; GCN-O0-NEXT: s_waitcnt vmcnt(0) lgkmcnt(0)
-; GCN-O0-NEXT: v_writelane_b32 v0, s0, 0
-; GCN-O0-NEXT: v_writelane_b32 v0, s1, 1
-; GCN-O0-NEXT: v_mov_b32_e32 v2, v1
-; GCN-O0-NEXT: buffer_store_dword v2, off, s[12:15], 0 offset:8 ; 4-byte Folded Spill
+; GCN-O0-NEXT: s_waitcnt lgkmcnt(0)
+; GCN-O0-NEXT: v_writelane_b32 v1, s0, 0
+; GCN-O0-NEXT: v_writelane_b32 v1, s1, 1
+; GCN-O0-NEXT: v_mov_b32_e32 v2, v0
+; GCN-O0-NEXT: buffer_store_dword v2, off, s[12:15], 0 offset:4 ; 4-byte Folded Spill
; GCN-O0-NEXT: s_mov_b32 s0, 1
-; GCN-O0-NEXT: v_cmp_gt_u32_e64 s[2:3], v1, s0
+; GCN-O0-NEXT: v_cmp_gt_u32_e64 s[2:3], v0, s0
; GCN-O0-NEXT: s_mov_b64 s[0:1], exec
-; GCN-O0-NEXT: v_writelane_b32 v0, s0, 2
-; GCN-O0-NEXT: v_writelane_b32 v0, s1, 3
-; GCN-O0-NEXT: s_or_saveexec_b64 s[8:9], -1
-; GCN-O0-NEXT: buffer_store_dword v0, off, s[12:15], 0 offset:4 ; 4-byte Folded Spill
-; GCN-O0-NEXT: s_mov_b64 exec, s[8:9]
+; GCN-O0-NEXT: v_writelane_b32 v1, s0, 2
+; GCN-O0-NEXT: v_writelane_b32 v1, s1, 3
; GCN-O0-NEXT: s_and_b64 s[0:1], s[0:1], s[2:3]
; GCN-O0-NEXT: s_mov_b64 exec, s[0:1]
; GCN-O0-NEXT: s_cbranch_execz .LBB0_4
; GCN-O0-NEXT: ; %bb.1: ; %bb.outer.then
-; GCN-O0-NEXT: s_or_saveexec_b64 s[8:9], -1
-; GCN-O0-NEXT: s_waitcnt expcnt(0)
; GCN-O0-NEXT: buffer_load_dword v0, off, s[12:15], 0 offset:4 ; 4-byte Folded Reload
-; GCN-O0-NEXT: s_mov_b64 exec, s[8:9]
-; GCN-O0-NEXT: buffer_load_dword v1, off, s[12:15], 0 offset:8 ; 4-byte Folded Reload
-; GCN-O0-NEXT: s_waitcnt vmcnt(1)
-; GCN-O0-NEXT: v_readlane_b32 s4, v0, 0
-; GCN-O0-NEXT: v_readlane_b32 s5, v0, 1
+; GCN-O0-NEXT: v_readlane_b32 s4, v1, 0
+; GCN-O0-NEXT: v_readlane_b32 s5, v1, 1
; GCN-O0-NEXT: s_mov_b32 s2, 0xf000
; GCN-O0-NEXT: s_mov_b32 s0, 0
; GCN-O0-NEXT: ; kill: def $sgpr0 killed $sgpr0 def $sgpr0_sgpr1
@@ -86,69 +73,54 @@
; GCN-O0-NEXT: ; kill: def $sgpr4_sgpr5 killed $sgpr4_sgpr5 def $sgpr4_sgpr5_sgpr6_sgpr7
; GCN-O0-NEXT: s_mov_b64 s[6:7], s[0:1]
; GCN-O0-NEXT: s_waitcnt vmcnt(0)
-; GCN-O0-NEXT: v_ashrrev_i32_e64 v4, 31, v1
-; GCN-O0-NEXT: v_mov_b32_e32 v2, v1
+; GCN-O0-NEXT: v_ashrrev_i32_e64 v4, 31, v0
+; GCN-O0-NEXT: s_waitcnt expcnt(0)
+; GCN-O0-NEXT: v_mov_b32_e32 v2, v0
; GCN-O0-NEXT: v_mov_b32_e32 v3, v4
; GCN-O0-NEXT: s_mov_b32 s0, 2
; GCN-O0-NEXT: v_lshl_b64 v[3:4], v[2:3], s0
; GCN-O0-NEXT: v_mov_b32_e32 v2, 0
; GCN-O0-NEXT: buffer_store_dword v2, v[3:4], s[4:7], 0 addr64
-; GCN-O0-NEXT: v_cmp_ne_u32_e64 s[2:3], v1, s0
+; GCN-O0-NEXT: v_cmp_ne_u32_e64 s[2:3], v0, s0
; GCN-O0-NEXT: s_mov_b64 s[0:1], exec
-; GCN-O0-NEXT: v_writelane_b32 v0, s0, 4
-; GCN-O0-NEXT: v_writelane_b32 v0, s1, 5
-; GCN-O0-NEXT: s_or_saveexec_b64 s[8:9], -1
-; GCN-O0-NEXT: buffer_store_dword v0, off, s[12:15], 0 offset:4 ; 4-byte Folded Spill
-; GCN-O0-NEXT: s_mov_b64 exec, s[8:9]
+; GCN-O0-NEXT: v_writelane_b32 v1, s0, 4
+; GCN-O0-NEXT: v_writelane_b32 v1, s1, 5
; GCN-O0-NEXT: s_and_b64 s[0:1], s[0:1], s[2:3]
; GCN-O0-NEXT: s_mov_b64 exec, s[0:1]
; GCN-O0-NEXT: s_cbranch_execz .LBB0_3
; GCN-O0-NEXT: ; %bb.2: ; %bb.inner.then
-; GCN-O0-NEXT: buffer_load_dword v1, off, s[12:15], 0 offset:8 ; 4-byte Folded Reload
-; GCN-O0-NEXT: s_or_saveexec_b64 s[8:9], -1
; GCN-O0-NEXT: s_waitcnt expcnt(0)
-; GCN-O0-NEXT: buffer_load_dword v0, off, s[12:15], 0 offset:4 ; 4-byte Folded Reload
-; GCN-O0-NEXT: s_mov_b64 exec, s[8:9]
-; GCN-O0-NEXT: s_waitcnt vmcnt(0)
-; GCN-O0-NEXT: v_readlane_b32 s0, v0, 0
-; GCN-O0-NEXT: v_readlane_b32 s1, v0, 1
+; GCN-O0-NEXT: buffer_load_dword v2, off, s[12:15], 0 offset:4 ; 4-byte Folded Reload
+; GCN-O0-NEXT: v_readlane_b32 s0, v1, 0
+; GCN-O0-NEXT: v_readlane_b32 s1, v1, 1
; GCN-O0-NEXT: v_mov_b32_e32 v0, 1
-; GCN-O0-NEXT: v_add_i32_e64 v1, s[2:3], v1, v0
-; GCN-O0-NEXT: v_ashrrev_i32_e64 v3, 31, v1
-; GCN-O0-NEXT: ; kill: def $vgpr1 killed $vgpr1 def $vgpr1_vgpr2 killed $exec
-; GCN-O0-NEXT: v_mov_b32_e32 v2, v3
+; GCN-O0-NEXT: s_waitcnt vmcnt(0)
+; GCN-O0-NEXT: v_add_i32_e64 v2, s[2:3], v2, v0
+; GCN-O0-NEXT: v_ashrrev_i32_e64 v4, 31, v2
+; GCN-O0-NEXT: ; kill: def $vgpr2 killed $vgpr2 def $vgpr2_vgpr3 killed $exec
+; GCN-O0-NEXT: v_mov_b32_e32 v3, v4
; GCN-O0-NEXT: s_mov_b32 s2, 2
-; GCN-O0-NEXT: v_lshl_b64 v[1:2], v[1:2], s2
+; GCN-O0-NEXT: v_lshl_b64 v[2:3], v[2:3], s2
; GCN-O0-NEXT: s_mov_b32 s2, 0xf000
; GCN-O0-NEXT: s_mov_b32 s4, 0
; GCN-O0-NEXT: ; kill: def $sgpr4 killed $sgpr4 def $sgpr4_sgpr5
; GCN-O0-NEXT: s_mov_b32 s5, s2
; GCN-O0-NEXT: ; kill: def $sgpr0_sgpr1 killed $sgpr0_sgpr1 def $sgpr0_sgpr1_sgpr2_sgpr3
; GCN-O0-NEXT: s_mov_b64 s[2:3], s[4:5]
-; GCN-O0-NEXT: buffer_store_dword v0, v[1:2], s[0:3], 0 addr64
+; GCN-O0-NEXT: buffer_store_dword v0, v[2:3], s[0:3], 0 addr64
; GCN-O0-NEXT: .LBB0_3: ; %Flow
-; GCN-O0-NEXT: s_or_saveexec_b64 s[8:9], -1
-; GCN-O0-NEXT: s_waitcnt expcnt(0)
-; GCN-O0-NEXT: buffer_load_dword v0, off, s[12:15], 0 offset:4 ; 4-byte Folded Reload
-; GCN-O0-NEXT: s_mov_b64 exec, s[8:9]
-; GCN-O0-NEXT: s_waitcnt vmcnt(0)
-; GCN-O0-NEXT: v_readlane_b32 s0, v0, 4
-; GCN-O0-NEXT: v_readlane_b32 s1, v0, 5
+; GCN-O0-NEXT: v_readlane_b32 s0, v1, 4
+; GCN-O0-NEXT: v_readlane_b32 s1, v1, 5
; GCN-O0-NEXT: s_or_b64 exec, exec, s[0:1]
; GCN-O0-NEXT: .LBB0_4: ; %bb.outer.end
-; GCN-O0-NEXT: s_or_saveexec_b64 s[8:9], -1
-; GCN-O0-NEXT: s_waitcnt expcnt(0)
-; GCN-O0-NEXT: buffer_load_dword v0, off, s[12:15], 0 offset:4 ; 4-byte Folded Reload
-; GCN-O0-NEXT: s_mov_b64 exec, s[8:9]
-; GCN-O0-NEXT: s_waitcnt vmcnt(0)
-; GCN-O0-NEXT: v_readlane_b32 s0, v0, 2
-; GCN-O0-NEXT: v_readlane_b32 s1, v0, 3
+; GCN-O0-NEXT: v_readlane_b32 s0, v1, 2
+; GCN-O0-NEXT: v_readlane_b32 s1, v1, 3
; GCN-O0-NEXT: s_or_b64 exec, exec, s[0:1]
+; GCN-O0-NEXT: s_waitcnt expcnt(0)
; GCN-O0-NEXT: v_mov_b32_e32 v2, 3
-; GCN-O0-NEXT: v_mov_b32_e32 v1, 0
+; GCN-O0-NEXT: v_mov_b32_e32 v0, 0
; GCN-O0-NEXT: s_mov_b32 m0, -1
-; GCN-O0-NEXT: ds_write_b32 v1, v2
-; GCN-O0-NEXT: ; kill: killed $vgpr0
+; GCN-O0-NEXT: ds_write_b32 v0, v2
; GCN-O0-NEXT: s_endpgm
bb:
%tmp = tail call i32 @llvm.amdgcn.workitem.id.x()
@@ -221,37 +193,24 @@
; GCN-O0-NEXT: s_mov_b32 s15, 0xe8f000
; GCN-O0-NEXT: s_add_u32 s12, s12, s11
; GCN-O0-NEXT: s_addc_u32 s13, s13, 0
-; GCN-O0-NEXT: ; implicit-def: $vgpr1
-; GCN-O0-NEXT: v_mov_b32_e32 v1, v0
-; GCN-O0-NEXT: s_or_saveexec_b64 s[8:9], -1
-; GCN-O0-NEXT: buffer_load_dword v0, off, s[12:15], 0 offset:4 ; 4-byte Folded Reload
-; GCN-O0-NEXT: s_mov_b64 exec, s[8:9]
; GCN-O0-NEXT: s_load_dwordx2 s[0:1], s[4:5], 0x9
-; GCN-O0-NEXT: s_waitcnt vmcnt(0) lgkmcnt(0)
-; GCN-O0-NEXT: v_writelane_b32 v0, s0, 0
-; GCN-O0-NEXT: v_writelane_b32 v0, s1, 1
-; GCN-O0-NEXT: v_mov_b32_e32 v2, v1
-; GCN-O0-NEXT: buffer_store_dword v2, off, s[12:15], 0 offset:8 ; 4-byte Folded Spill
+; GCN-O0-NEXT: s_waitcnt lgkmcnt(0)
+; GCN-O0-NEXT: v_writelane_b32 v1, s0, 0
+; GCN-O0-NEXT: v_writelane_b32 v1, s1, 1
+; GCN-O0-NEXT: v_mov_b32_e32 v2, v0
+; GCN-O0-NEXT: buffer_store_dword v2, off, s[12:15], 0 offset:4 ; 4-byte Folded Spill
; GCN-O0-NEXT: s_mov_b32 s0, 1
-; GCN-O0-NEXT: v_cmp_gt_u32_e64 s[2:3], v1, s0
+; GCN-O0-NEXT: v_cmp_gt_u32_e64 s[2:3], v0, s0
; GCN-O0-NEXT: s_mov_b64 s[0:1], exec
-; GCN-O0-NEXT: v_writelane_b32 v0, s0, 2
-; GCN-O0-NEXT: v_writelane_b32 v0, s1, 3
-; GCN-O0-NEXT: s_or_saveexec_b64 s[8:9], -1
-; GCN-O0-NEXT: buffer_store_dword v0, off, s[12:15], 0 offset:4 ; 4-byte Folded Spill
-; GCN-O0-NEXT: s_mov_b64 exec, s[8:9]
+; GCN-O0-NEXT: v_writelane_b32 v1, s0, 2
+; GCN-O0-NEXT: v_writelane_b32 v1, s1, 3
; GCN-O0-NEXT: s_and_b64 s[0:1], s[0:1], s[2:3]
; GCN-O0-NEXT: s_mov_b64 exec, s[0:1]
; GCN-O0-NEXT: s_cbranch_execz .LBB1_3
; GCN-O0-NEXT: ; %bb.1: ; %bb.outer.then
-; GCN-O0-NEXT: s_or_saveexec_b64 s[8:9], -1
-; GCN-O0-NEXT: s_waitcnt expcnt(0)
; GCN-O0-NEXT: buffer_load_dword v0, off, s[12:15], 0 offset:4 ; 4-byte Folded Reload
-; GCN-O0-NEXT: s_mov_b64 exec, s[8:9]
-; GCN-O0-NEXT: buffer_load_dword v1, off, s[12:15], 0 offset:8 ; 4-byte Folded Reload
-; GCN-O0-NEXT: s_waitcnt vmcnt(1)
-; GCN-O0-NEXT: v_readlane_b32 s4, v0, 0
-; GCN-O0-NEXT: v_readlane_b32 s5, v0, 1
+; GCN-O0-NEXT: v_readlane_b32 s4, v1, 0
+; GCN-O0-NEXT: v_readlane_b32 s5, v1, 1
; GCN-O0-NEXT: s_mov_b32 s2, 0xf000
; GCN-O0-NEXT: s_mov_b32 s0, 0
; GCN-O0-NEXT: ; kill: def $sgpr0 killed $sgpr0 def $sgpr0_sgpr1
@@ -259,92 +218,76 @@
; GCN-O0-NEXT: ; kill: def $sgpr4_sgpr5 killed $sgpr4_sgpr5 def $sgpr4_sgpr5_sgpr6_sgpr7
; GCN-O0-NEXT: s_mov_b64 s[6:7], s[0:1]
; GCN-O0-NEXT: s_waitcnt vmcnt(0)
-; GCN-O0-NEXT: v_ashrrev_i32_e64 v4, 31, v1
-; GCN-O0-NEXT: v_mov_b32_e32 v2, v1
+; GCN-O0-NEXT: v_ashrrev_i32_e64 v4, 31, v0
+; GCN-O0-NEXT: s_waitcnt expcnt(0)
+; GCN-O0-NEXT: v_mov_b32_e32 v2, v0
; GCN-O0-NEXT: v_mov_b32_e32 v3, v4
; GCN-O0-NEXT: s_mov_b32 s0, 2
; GCN-O0-NEXT: v_lshl_b64 v[3:4], v[2:3], s0
; GCN-O0-NEXT: v_mov_b32_e32 v2, 0
; GCN-O0-NEXT: buffer_store_dword v2, v[3:4], s[4:7], 0 addr64
-; GCN-O0-NEXT: v_cmp_ne_u32_e64 s[2:3], v1, s0
+; GCN-O0-NEXT: v_cmp_ne_u32_e64 s[2:3], v0, s0
; GCN-O0-NEXT: s_mov_b64 s[0:1], exec
-; GCN-O0-NEXT: v_writelane_b32 v0, s0, 4
-; GCN-O0-NEXT: v_writelane_b32 v0, s1, 5
-; GCN-O0-NEXT: s_or_saveexec_b64 s[8:9], -1
-; GCN-O0-NEXT: buffer_store_dword v0, off, s[12:15], 0 offset:4 ; 4-byte Folded Spill
-; GCN-O0-NEXT: s_mov_b64 exec, s[8:9]
+; GCN-O0-NEXT: v_writelane_b32 v1, s0, 4
+; GCN-O0-NEXT: v_writelane_b32 v1, s1, 5
; GCN-O0-NEXT: s_and_b64 s[0:1], s[0:1], s[2:3]
; GCN-O0-NEXT: s_mov_b64 exec, s[0:1]
; GCN-O0-NEXT: s_cbranch_execz .LBB1_4
; GCN-O0-NEXT: ; %bb.2: ; %bb.inner.then
-; GCN-O0-NEXT: buffer_load_dword v1, off, s[12:15], 0 offset:8 ; 4-byte Folded Reload
-; GCN-O0-NEXT: s_or_saveexec_b64 s[8:9], -1
; GCN-O0-NEXT: s_waitcnt expcnt(0)
-; GCN-O0-NEXT: buffer_load_dword v0, off, s[12:15], 0 offset:4 ; 4-byte Folded Reload
-; GCN-O0-NEXT: s_mov_b64 exec, s[8:9]
-; GCN-O0-NEXT: s_waitcnt vmcnt(0)
-; GCN-O0-NEXT: v_readlane_b32 s0, v0, 0
-; GCN-O0-NEXT: v_readlane_b32 s1, v0, 1
+; GCN-O0-NEXT: buffer_load_dword v2, off, s[12:15], 0 offset:4 ; 4-byte Folded Reload
+; GCN-O0-NEXT: v_readlane_b32 s0, v1, 0
+; GCN-O0-NEXT: v_readlane_b32 s1, v1, 1
; GCN-O0-NEXT: v_mov_b32_e32 v0, 1
-; GCN-O0-NEXT: v_add_i32_e64 v1, s[2:3], v1, v0
-; GCN-O0-NEXT: v_ashrrev_i32_e64 v3, 31, v1
-; GCN-O0-NEXT: ; kill: def $vgpr1 killed $vgpr1 def $vgpr1_vgpr2 killed $exec
-; GCN-O0-NEXT: v_mov_b32_e32 v2, v3
+; GCN-O0-NEXT: s_waitcnt vmcnt(0)
+; GCN-O0-NEXT: v_add_i32_e64 v2, s[2:3], v2, v0
+; GCN-O0-NEXT: v_ashrrev_i32_e64 v4, 31, v2
+; GCN-O0-NEXT: ; kill: def $vgpr2 killed $vgpr2 def $vgpr2_vgpr3 killed $exec
+; GCN-O0-NEXT: v_mov_b32_e32 v3, v4
; GCN-O0-NEXT: s_mov_b32 s2, 2
-; GCN-O0-NEXT: v_lshl_b64 v[1:2], v[1:2], s2
+; GCN-O0-NEXT: v_lshl_b64 v[2:3], v[2:3], s2
; GCN-O0-NEXT: s_mov_b32 s2, 0xf000
; GCN-O0-NEXT: s_mov_b32 s4, 0
; GCN-O0-NEXT: ; kill: def $sgpr4 killed $sgpr4 def $sgpr4_sgpr5
; GCN-O0-NEXT: s_mov_b32 s5, s2
; GCN-O0-NEXT: ; kill: def $sgpr0_sgpr1 killed $sgpr0_sgpr1 def $sgpr0_sgpr1_sgpr2_sgpr3
; GCN-O0-NEXT: s_mov_b64 s[2:3], s[4:5]
-; GCN-O0-NEXT: buffer_store_dword v0, v[1:2], s[0:3], 0 addr64
+; GCN-O0-NEXT: buffer_store_dword v0, v[2:3], s[0:3], 0 addr64
; GCN-O0-NEXT: s_branch .LBB1_4
; GCN-O0-NEXT: .LBB1_3: ; %Flow
-; GCN-O0-NEXT: s_or_saveexec_b64 s[8:9], -1
-; GCN-O0-NEXT: s_waitcnt expcnt(0)
-; GCN-O0-NEXT: buffer_load_dword v0, off, s[12:15], 0 offset:4 ; 4-byte Folded Reload
-; GCN-O0-NEXT: s_mov_b64 exec, s[8:9]
-; GCN-O0-NEXT: s_waitcnt vmcnt(0)
-; GCN-O0-NEXT: v_readlane_b32 s0, v0, 2
-; GCN-O0-NEXT: v_readlane_b32 s1, v0, 3
+; GCN-O0-NEXT: v_readlane_b32 s0, v1, 2
+; GCN-O0-NEXT: v_readlane_b32 s1, v1, 3
; GCN-O0-NEXT: s_or_b64 exec, exec, s[0:1]
; GCN-O0-NEXT: s_branch .LBB1_5
; GCN-O0-NEXT: .LBB1_4: ; %bb.inner.end
-; GCN-O0-NEXT: buffer_load_dword v1, off, s[12:15], 0 offset:8 ; 4-byte Folded Reload
-; GCN-O0-NEXT: s_or_saveexec_b64 s[8:9], -1
; GCN-O0-NEXT: s_waitcnt expcnt(0)
-; GCN-O0-NEXT: buffer_load_dword v0, off, s[12:15], 0 offset:4 ; 4-byte Folded Reload
-; GCN-O0-NEXT: s_mov_b64 exec, s[8:9]
-; GCN-O0-NEXT: s_waitcnt vmcnt(0)
-; GCN-O0-NEXT: v_readlane_b32 s2, v0, 4
-; GCN-O0-NEXT: v_readlane_b32 s3, v0, 5
+; GCN-O0-NEXT: buffer_load_dword v2, off, s[12:15], 0 offset:4 ; 4-byte Folded Reload
+; GCN-O0-NEXT: v_readlane_b32 s2, v1, 4
+; GCN-O0-NEXT: v_readlane_b32 s3, v1, 5
; GCN-O0-NEXT: s_or_b64 exec, exec, s[2:3]
-; GCN-O0-NEXT: v_readlane_b32 s0, v0, 0
-; GCN-O0-NEXT: v_readlane_b32 s1, v0, 1
+; GCN-O0-NEXT: v_readlane_b32 s0, v1, 0
+; GCN-O0-NEXT: v_readlane_b32 s1, v1, 1
; GCN-O0-NEXT: v_mov_b32_e32 v0, 2
-; GCN-O0-NEXT: v_add_i32_e64 v1, s[2:3], v1, v0
-; GCN-O0-NEXT: v_ashrrev_i32_e64 v3, 31, v1
-; GCN-O0-NEXT: ; kill: def $vgpr1 killed $vgpr1 def $vgpr1_vgpr2 killed $exec
-; GCN-O0-NEXT: v_mov_b32_e32 v2, v3
-; GCN-O0-NEXT: v_lshl_b64 v[1:2], v[1:2], v0
+; GCN-O0-NEXT: s_waitcnt vmcnt(0)
+; GCN-O0-NEXT: v_add_i32_e64 v2, s[2:3], v2, v0
+; GCN-O0-NEXT: v_ashrrev_i32_e64 v4, 31, v2
+; GCN-O0-NEXT: ; kill: def $vgpr2 killed $vgpr2 def $vgpr2_vgpr3 killed $exec
+; GCN-O0-NEXT: v_mov_b32_e32 v3, v4
+; GCN-O0-NEXT: v_lshl_b64 v[2:3], v[2:3], v0
; GCN-O0-NEXT: s_mov_b32 s2, 0xf000
; GCN-O0-NEXT: s_mov_b32 s4, 0
; GCN-O0-NEXT: ; kill: def $sgpr4 killed $sgpr4 def $sgpr4_sgpr5
; GCN-O0-NEXT: s_mov_b32 s5, s2
; GCN-O0-NEXT: ; kill: def $sgpr0_sgpr1 killed $sgpr0_sgpr1 def $sgpr0_sgpr1_sgpr2_sgpr3
; GCN-O0-NEXT: s_mov_b64 s[2:3], s[4:5]
-; GCN-O0-NEXT: buffer_store_dword v0, v[1:2], s[0:3], 0 addr64
+; GCN-O0-NEXT: buffer_store_dword v0, v[2:3], s[0:3], 0 addr64
; GCN-O0-NEXT: s_branch .LBB1_3
; GCN-O0-NEXT: .LBB1_5: ; %bb.outer.end
-; GCN-O0-NEXT: s_or_saveexec_b64 s[8:9], -1
-; GCN-O0-NEXT: buffer_load_dword v0, off, s[12:15], 0 offset:4 ; 4-byte Folded Reload
-; GCN-O0-NEXT: s_mov_b64 exec, s[8:9]
+; GCN-O0-NEXT: s_waitcnt expcnt(0)
; GCN-O0-NEXT: v_mov_b32_e32 v2, 3
-; GCN-O0-NEXT: v_mov_b32_e32 v1, 0
+; GCN-O0-NEXT: v_mov_b32_e32 v0, 0
; GCN-O0-NEXT: s_mov_b32 m0, -1
-; GCN-O0-NEXT: ds_write_b32 v1, v2
-; GCN-O0-NEXT: ; kill: killed $vgpr0
+; GCN-O0-NEXT: ds_write_b32 v0, v2
; GCN-O0-NEXT: s_endpgm
bb:
%tmp = tail call i32 @llvm.amdgcn.workitem.id.x()
@@ -430,19 +373,13 @@
; GCN-O0-NEXT: s_mov_b32 s15, 0xe8f000
; GCN-O0-NEXT: s_add_u32 s12, s12, s11
; GCN-O0-NEXT: s_addc_u32 s13, s13, 0
-; GCN-O0-NEXT: ; implicit-def: $vgpr1
-; GCN-O0-NEXT: v_mov_b32_e32 v1, v0
-; GCN-O0-NEXT: s_or_saveexec_b64 s[6:7], -1
-; GCN-O0-NEXT: buffer_load_dword v0, off, s[12:15], 0 offset:4 ; 4-byte Folded Reload
-; GCN-O0-NEXT: s_mov_b64 exec, s[6:7]
; GCN-O0-NEXT: s_load_dwordx2 s[0:1], s[4:5], 0x9
; GCN-O0-NEXT: s_waitcnt lgkmcnt(0)
; GCN-O0-NEXT: s_mov_b64 s[2:3], s[0:1]
-; GCN-O0-NEXT: s_waitcnt vmcnt(0)
-; GCN-O0-NEXT: v_writelane_b32 v0, s2, 0
-; GCN-O0-NEXT: v_writelane_b32 v0, s3, 1
-; GCN-O0-NEXT: v_mov_b32_e32 v2, v1
-; GCN-O0-NEXT: buffer_store_dword v2, off, s[12:15], 0 offset:8 ; 4-byte Folded Spill
+; GCN-O0-NEXT: v_writelane_b32 v1, s2, 0
+; GCN-O0-NEXT: v_writelane_b32 v1, s3, 1
+; GCN-O0-NEXT: v_mov_b32_e32 v2, v0
+; GCN-O0-NEXT: buffer_store_dword v2, off, s[12:15], 0 offset:4 ; 4-byte Folded Spill
; GCN-O0-NEXT: s_mov_b32 s2, 0xf000
; GCN-O0-NEXT: s_mov_b32 s4, 0
; GCN-O0-NEXT: ; kill: def $sgpr4 killed $sgpr4 def $sgpr4_sgpr5
@@ -453,130 +390,96 @@
; GCN-O0-NEXT: ; implicit-def: $sgpr4
; GCN-O0-NEXT: v_mov_b32_e32 v4, 0
; GCN-O0-NEXT: s_waitcnt expcnt(0)
-; GCN-O0-NEXT: v_mov_b32_e32 v2, v1
+; GCN-O0-NEXT: v_mov_b32_e32 v2, v0
; GCN-O0-NEXT: v_mov_b32_e32 v3, v4
; GCN-O0-NEXT: s_mov_b32 s4, 2
; GCN-O0-NEXT: v_lshl_b64 v[3:4], v[2:3], s4
; GCN-O0-NEXT: v_mov_b32_e32 v2, 0
; GCN-O0-NEXT: buffer_store_dword v2, v[3:4], s[0:3], 0 addr64
; GCN-O0-NEXT: s_mov_b32 s0, 1
-; GCN-O0-NEXT: v_cmp_gt_u32_e64 s[2:3], v1, s0
+; GCN-O0-NEXT: v_cmp_gt_u32_e64 s[2:3], v0, s0
; GCN-O0-NEXT: s_mov_b64 s[0:1], exec
-; GCN-O0-NEXT: v_writelane_b32 v0, s0, 2
-; GCN-O0-NEXT: v_writelane_b32 v0, s1, 3
-; GCN-O0-NEXT: s_or_saveexec_b64 s[6:7], -1
-; GCN-O0-NEXT: buffer_store_dword v0, off, s[12:15], 0 offset:4 ; 4-byte Folded Spill
-; GCN-O0-NEXT: s_mov_b64 exec, s[6:7]
+; GCN-O0-NEXT: v_writelane_b32 v1, s0, 2
+; GCN-O0-NEXT: v_writelane_b32 v1, s1, 3
; GCN-O0-NEXT: s_and_b64 s[0:1], s[0:1], s[2:3]
; GCN-O0-NEXT: s_mov_b64 exec, s[0:1]
; GCN-O0-NEXT: s_cbranch_execz .LBB2_6
; GCN-O0-NEXT: ; %bb.1: ; %bb.outer.then
-; GCN-O0-NEXT: s_or_saveexec_b64 s[6:7], -1
-; GCN-O0-NEXT: s_waitcnt expcnt(0)
; GCN-O0-NEXT: buffer_load_dword v0, off, s[12:15], 0 offset:4 ; 4-byte Folded Reload
-; GCN-O0-NEXT: s_mov_b64 exec, s[6:7]
-; GCN-O0-NEXT: buffer_load_dword v1, off, s[12:15], 0 offset:8 ; 4-byte Folded Reload
; GCN-O0-NEXT: s_mov_b32 s0, 2
; GCN-O0-NEXT: s_waitcnt vmcnt(0)
-; GCN-O0-NEXT: v_cmp_ne_u32_e64 s[0:1], v1, s0
+; GCN-O0-NEXT: v_cmp_ne_u32_e64 s[0:1], v0, s0
; GCN-O0-NEXT: s_mov_b64 s[2:3], exec
; GCN-O0-NEXT: s_and_b64 s[0:1], s[2:3], s[0:1]
; GCN-O0-NEXT: s_xor_b64 s[2:3], s[0:1], s[2:3]
-; GCN-O0-NEXT: v_writelane_b32 v0, s2, 4
-; GCN-O0-NEXT: v_writelane_b32 v0, s3, 5
-; GCN-O0-NEXT: s_or_saveexec_b64 s[6:7], -1
-; GCN-O0-NEXT: buffer_store_dword v0, off, s[12:15], 0 offset:4 ; 4-byte Folded Spill
-; GCN-O0-NEXT: s_mov_b64 exec, s[6:7]
+; GCN-O0-NEXT: v_writelane_b32 v1, s2, 4
+; GCN-O0-NEXT: v_writelane_b32 v1, s3, 5
; GCN-O0-NEXT: s_mov_b64 exec, s[0:1]
; GCN-O0-NEXT: s_cbranch_execz .LBB2_2
; GCN-O0-NEXT: s_branch .LBB2_4
; GCN-O0-NEXT: .LBB2_2: ; %Flow
-; GCN-O0-NEXT: s_or_saveexec_b64 s[6:7], -1
-; GCN-O0-NEXT: s_waitcnt expcnt(0)
-; GCN-O0-NEXT: buffer_load_dword v0, off, s[12:15], 0 offset:4 ; 4-byte Folded Reload
-; GCN-O0-NEXT: s_mov_b64 exec, s[6:7]
-; GCN-O0-NEXT: s_waitcnt vmcnt(0)
-; GCN-O0-NEXT: v_readlane_b32 s0, v0, 4
-; GCN-O0-NEXT: v_readlane_b32 s1, v0, 5
+; GCN-O0-NEXT: v_readlane_b32 s0, v1, 4
+; GCN-O0-NEXT: v_readlane_b32 s1, v1, 5
; GCN-O0-NEXT: s_or_saveexec_b64 s[0:1], s[0:1]
; GCN-O0-NEXT: s_and_b64 s[0:1], exec, s[0:1]
-; GCN-O0-NEXT: v_writelane_b32 v0, s0, 6
-; GCN-O0-NEXT: v_writelane_b32 v0, s1, 7
-; GCN-O0-NEXT: s_or_saveexec_b64 s[6:7], -1
-; GCN-O0-NEXT: buffer_store_dword v0, off, s[12:15], 0 offset:4 ; 4-byte Folded Spill
-; GCN-O0-NEXT: s_mov_b64 exec, s[6:7]
+; GCN-O0-NEXT: v_writelane_b32 v1, s0, 6
+; GCN-O0-NEXT: v_writelane_b32 v1, s1, 7
; GCN-O0-NEXT: s_xor_b64 exec, exec, s[0:1]
; GCN-O0-NEXT: s_cbranch_execz .LBB2_5
; GCN-O0-NEXT: ; %bb.3: ; %bb.then
-; GCN-O0-NEXT: buffer_load_dword v1, off, s[12:15], 0 offset:8 ; 4-byte Folded Reload
-; GCN-O0-NEXT: s_or_saveexec_b64 s[6:7], -1
; GCN-O0-NEXT: s_waitcnt expcnt(0)
-; GCN-O0-NEXT: buffer_load_dword v0, off, s[12:15], 0 offset:4 ; 4-byte Folded Reload
-; GCN-O0-NEXT: s_mov_b64 exec, s[6:7]
-; GCN-O0-NEXT: s_waitcnt vmcnt(0)
-; GCN-O0-NEXT: v_readlane_b32 s0, v0, 0
-; GCN-O0-NEXT: v_readlane_b32 s1, v0, 1
+; GCN-O0-NEXT: buffer_load_dword v2, off, s[12:15], 0 offset:4 ; 4-byte Folded Reload
+; GCN-O0-NEXT: v_readlane_b32 s0, v1, 0
+; GCN-O0-NEXT: v_readlane_b32 s1, v1, 1
; GCN-O0-NEXT: v_mov_b32_e32 v0, 1
-; GCN-O0-NEXT: v_add_i32_e64 v1, s[2:3], v1, v0
-; GCN-O0-NEXT: v_ashrrev_i32_e64 v3, 31, v1
-; GCN-O0-NEXT: ; kill: def $vgpr1 killed $vgpr1 def $vgpr1_vgpr2 killed $exec
-; GCN-O0-NEXT: v_mov_b32_e32 v2, v3
+; GCN-O0-NEXT: s_waitcnt vmcnt(0)
+; GCN-O0-NEXT: v_add_i32_e64 v2, s[2:3], v2, v0
+; GCN-O0-NEXT: v_ashrrev_i32_e64 v4, 31, v2
+; GCN-O0-NEXT: ; kill: def $vgpr2 killed $vgpr2 def $vgpr2_vgpr3 killed $exec
+; GCN-O0-NEXT: v_mov_b32_e32 v3, v4
; GCN-O0-NEXT: s_mov_b32 s2, 2
-; GCN-O0-NEXT: v_lshl_b64 v[1:2], v[1:2], s2
+; GCN-O0-NEXT: v_lshl_b64 v[2:3], v[2:3], s2
; GCN-O0-NEXT: s_mov_b32 s2, 0xf000
; GCN-O0-NEXT: s_mov_b32 s4, 0
; GCN-O0-NEXT: ; kill: def $sgpr4 killed $sgpr4 def $sgpr4_sgpr5
; GCN-O0-NEXT: s_mov_b32 s5, s2
; GCN-O0-NEXT: ; kill: def $sgpr0_sgpr1 killed $sgpr0_sgpr1 def $sgpr0_sgpr1_sgpr2_sgpr3
; GCN-O0-NEXT: s_mov_b64 s[2:3], s[4:5]
-; GCN-O0-NEXT: buffer_store_dword v0, v[1:2], s[0:3], 0 addr64
+; GCN-O0-NEXT: buffer_store_dword v0, v[2:3], s[0:3], 0 addr64
; GCN-O0-NEXT: s_branch .LBB2_5
; GCN-O0-NEXT: .LBB2_4: ; %bb.else
-; GCN-O0-NEXT: buffer_load_dword v1, off, s[12:15], 0 offset:8 ; 4-byte Folded Reload
-; GCN-O0-NEXT: s_or_saveexec_b64 s[6:7], -1
; GCN-O0-NEXT: s_waitcnt expcnt(0)
-; GCN-O0-NEXT: buffer_load_dword v0, off, s[12:15], 0 offset:4 ; 4-byte Folded Reload
-; GCN-O0-NEXT: s_mov_b64 exec, s[6:7]
-; GCN-O0-NEXT: s_waitcnt vmcnt(0)
-; GCN-O0-NEXT: v_readlane_b32 s0, v0, 0
-; GCN-O0-NEXT: v_readlane_b32 s1, v0, 1
+; GCN-O0-NEXT: buffer_load_dword v2, off, s[12:15], 0 offset:4 ; 4-byte Folded Reload
+; GCN-O0-NEXT: v_readlane_b32 s0, v1, 0
+; GCN-O0-NEXT: v_readlane_b32 s1, v1, 1
; GCN-O0-NEXT: v_mov_b32_e32 v0, 2
-; GCN-O0-NEXT: v_add_i32_e64 v1, s[2:3], v1, v0
-; GCN-O0-NEXT: v_ashrrev_i32_e64 v3, 31, v1
-; GCN-O0-NEXT: ; kill: def $vgpr1 killed $vgpr1 def $vgpr1_vgpr2 killed $exec
-; GCN-O0-NEXT: v_mov_b32_e32 v2, v3
-; GCN-O0-NEXT: v_lshl_b64 v[1:2], v[1:2], v0
+; GCN-O0-NEXT: s_waitcnt vmcnt(0)
+; GCN-O0-NEXT: v_add_i32_e64 v2, s[2:3], v2, v0
+; GCN-O0-NEXT: v_ashrrev_i32_e64 v4, 31, v2
+; GCN-O0-NEXT: ; kill: def $vgpr2 killed $vgpr2 def $vgpr2_vgpr3 killed $exec
+; GCN-O0-NEXT: v_mov_b32_e32 v3, v4
+; GCN-O0-NEXT: v_lshl_b64 v[2:3], v[2:3], v0
; GCN-O0-NEXT: s_mov_b32 s2, 0xf000
; GCN-O0-NEXT: s_mov_b32 s4, 0
; GCN-O0-NEXT: ; kill: def $sgpr4 killed $sgpr4 def $sgpr4_sgpr5
; GCN-O0-NEXT: s_mov_b32 s5, s2
; GCN-O0-NEXT: ; kill: def $sgpr0_sgpr1 killed $sgpr0_sgpr1 def $sgpr0_sgpr1_sgpr2_sgpr3
; GCN-O0-NEXT: s_mov_b64 s[2:3], s[4:5]
-; GCN-O0-NEXT: buffer_store_dword v0, v[1:2], s[0:3], 0 addr64
+; GCN-O0-NEXT: buffer_store_dword v0, v[2:3], s[0:3], 0 addr64
; GCN-O0-NEXT: s_branch .LBB2_2
; GCN-O0-NEXT: .LBB2_5: ; %Flow1
-; GCN-O0-NEXT: s_or_saveexec_b64 s[6:7], -1
-; GCN-O0-NEXT: s_waitcnt expcnt(0)
-; GCN-O0-NEXT: buffer_load_dword v0, off, s[12:15], 0 offset:4 ; 4-byte Folded Reload
-; GCN-O0-NEXT: s_mov_b64 exec, s[6:7]
-; GCN-O0-NEXT: s_waitcnt vmcnt(0)
-; GCN-O0-NEXT: v_readlane_b32 s0, v0, 6
-; GCN-O0-NEXT: v_readlane_b32 s1, v0, 7
+; GCN-O0-NEXT: v_readlane_b32 s0, v1, 6
+; GCN-O0-NEXT: v_readlane_b32 s1, v1, 7
; GCN-O0-NEXT: s_or_b64 exec, exec, s[0:1]
; GCN-O0-NEXT: .LBB2_6: ; %bb.outer.end
-; GCN-O0-NEXT: s_or_saveexec_b64 s[6:7], -1
-; GCN-O0-NEXT: s_waitcnt expcnt(0)
-; GCN-O0-NEXT: buffer_load_dword v0, off, s[12:15], 0 offset:4 ; 4-byte Folded Reload
-; GCN-O0-NEXT: s_mov_b64 exec, s[6:7]
-; GCN-O0-NEXT: s_waitcnt vmcnt(0)
-; GCN-O0-NEXT: v_readlane_b32 s0, v0, 2
-; GCN-O0-NEXT: v_readlane_b32 s1, v0, 3
+; GCN-O0-NEXT: v_readlane_b32 s0, v1, 2
+; GCN-O0-NEXT: v_readlane_b32 s1, v1, 3
; GCN-O0-NEXT: s_or_b64 exec, exec, s[0:1]
+; GCN-O0-NEXT: s_waitcnt expcnt(0)
; GCN-O0-NEXT: v_mov_b32_e32 v2, 3
-; GCN-O0-NEXT: v_mov_b32_e32 v1, 0
+; GCN-O0-NEXT: v_mov_b32_e32 v0, 0
; GCN-O0-NEXT: s_mov_b32 m0, -1
-; GCN-O0-NEXT: ds_write_b32 v1, v2
-; GCN-O0-NEXT: ; kill: killed $vgpr0
+; GCN-O0-NEXT: ds_write_b32 v0, v2
; GCN-O0-NEXT: s_endpgm
bb:
%tmp = tail call i32 @llvm.amdgcn.workitem.id.x()
@@ -676,19 +579,14 @@
; GCN-O0-NEXT: s_mov_b32 s15, 0xe8f000
; GCN-O0-NEXT: s_add_u32 s12, s12, s11
; GCN-O0-NEXT: s_addc_u32 s13, s13, 0
-; GCN-O0-NEXT: ; implicit-def: $vgpr1
-; GCN-O0-NEXT: v_mov_b32_e32 v1, v0
-; GCN-O0-NEXT: s_or_saveexec_b64 s[8:9], -1
-; GCN-O0-NEXT: buffer_load_dword v0, off, s[12:15], 0 offset:4 ; 4-byte Folded Reload
-; GCN-O0-NEXT: s_mov_b64 exec, s[8:9]
; GCN-O0-NEXT: s_load_dwordx2 s[4:5], s[4:5], 0x9
-; GCN-O0-NEXT: v_mov_b32_e32 v2, v1
-; GCN-O0-NEXT: buffer_store_dword v2, off, s[12:15], 0 offset:16 ; 4-byte Folded Spill
+; GCN-O0-NEXT: v_mov_b32_e32 v2, v0
+; GCN-O0-NEXT: buffer_store_dword v2, off, s[12:15], 0 offset:12 ; 4-byte Folded Spill
; GCN-O0-NEXT: s_mov_b32 s0, 0
; GCN-O0-NEXT: ; implicit-def: $sgpr0
; GCN-O0-NEXT: v_mov_b32_e32 v4, 0
; GCN-O0-NEXT: s_waitcnt expcnt(0)
-; GCN-O0-NEXT: v_mov_b32_e32 v2, v1
+; GCN-O0-NEXT: v_mov_b32_e32 v2, v0
; GCN-O0-NEXT: v_mov_b32_e32 v3, v4
; GCN-O0-NEXT: s_mov_b32 s0, 2
; GCN-O0-NEXT: s_mov_b32 s1, s0
@@ -703,9 +601,9 @@
; GCN-O0-NEXT: v_addc_u32_e64 v2, s[2:3], v2, v6, s[2:3]
; GCN-O0-NEXT: ; kill: def $vgpr5 killed $vgpr5 def $vgpr5_vgpr6 killed $exec
; GCN-O0-NEXT: v_mov_b32_e32 v6, v2
-; GCN-O0-NEXT: buffer_store_dword v5, off, s[12:15], 0 offset:8 ; 4-byte Folded Spill
+; GCN-O0-NEXT: buffer_store_dword v5, off, s[12:15], 0 offset:4 ; 4-byte Folded Spill
; GCN-O0-NEXT: s_waitcnt vmcnt(0)
-; GCN-O0-NEXT: buffer_store_dword v6, off, s[12:15], 0 offset:12 ; 4-byte Folded Spill
+; GCN-O0-NEXT: buffer_store_dword v6, off, s[12:15], 0 offset:8 ; 4-byte Folded Spill
; GCN-O0-NEXT: s_mov_b32 s1, 0xf000
; GCN-O0-NEXT: s_mov_b32 s2, 0
; GCN-O0-NEXT: ; kill: def $sgpr2 killed $sgpr2 def $sgpr2_sgpr3
@@ -714,43 +612,29 @@
; GCN-O0-NEXT: s_mov_b64 s[6:7], s[2:3]
; GCN-O0-NEXT: v_mov_b32_e32 v2, 0
; GCN-O0-NEXT: buffer_store_dword v2, v[3:4], s[4:7], 0 addr64
-; GCN-O0-NEXT: v_cmp_lt_u32_e64 s[0:1], v1, s0
+; GCN-O0-NEXT: v_cmp_lt_u32_e64 s[0:1], v0, s0
; GCN-O0-NEXT: s_mov_b64 s[2:3], exec
; GCN-O0-NEXT: s_and_b64 s[0:1], s[2:3], s[0:1]
; GCN-O0-NEXT: s_xor_b64 s[2:3], s[0:1], s[2:3]
-; GCN-O0-NEXT: v_writelane_b32 v0, s2, 0
-; GCN-O0-NEXT: v_writelane_b32 v0, s3, 1
-; GCN-O0-NEXT: s_or_saveexec_b64 s[8:9], -1
-; GCN-O0-NEXT: buffer_store_dword v0, off, s[12:15], 0 offset:4 ; 4-byte Folded Spill
-; GCN-O0-NEXT: s_mov_b64 exec, s[8:9]
+; GCN-O0-NEXT: v_writelane_b32 v1, s2, 0
+; GCN-O0-NEXT: v_writelane_b32 v1, s3, 1
; GCN-O0-NEXT: s_mov_b64 exec, s[0:1]
; GCN-O0-NEXT: s_cbranch_execz .LBB3_1
; GCN-O0-NEXT: s_branch .LBB3_4
; GCN-O0-NEXT: .LBB3_1: ; %Flow2
-; GCN-O0-NEXT: s_or_saveexec_b64 s[8:9], -1
-; GCN-O0-NEXT: s_waitcnt expcnt(0)
-; GCN-O0-NEXT: buffer_load_dword v0, off, s[12:15], 0 offset:4 ; 4-byte Folded Reload
-; GCN-O0-NEXT: s_mov_b64 exec, s[8:9]
-; GCN-O0-NEXT: s_waitcnt vmcnt(0)
-; GCN-O0-NEXT: v_readlane_b32 s0, v0, 0
-; GCN-O0-NEXT: v_readlane_b32 s1, v0, 1
+; GCN-O0-NEXT: v_readlane_b32 s0, v1, 0
+; GCN-O0-NEXT: v_readlane_b32 s1, v1, 1
; GCN-O0-NEXT: s_or_saveexec_b64 s[0:1], s[0:1]
; GCN-O0-NEXT: s_and_b64 s[0:1], exec, s[0:1]
-; GCN-O0-NEXT: v_writelane_b32 v0, s0, 2
-; GCN-O0-NEXT: v_writelane_b32 v0, s1, 3
-; GCN-O0-NEXT: s_or_saveexec_b64 s[8:9], -1
-; GCN-O0-NEXT: buffer_store_dword v0, off, s[12:15], 0 offset:4 ; 4-byte Folded Spill
-; GCN-O0-NEXT: s_mov_b64 exec, s[8:9]
+; GCN-O0-NEXT: v_writelane_b32 v1, s0, 2
+; GCN-O0-NEXT: v_writelane_b32 v1, s1, 3
; GCN-O0-NEXT: s_xor_b64 exec, exec, s[0:1]
; GCN-O0-NEXT: s_cbranch_execz .LBB3_8
; GCN-O0-NEXT: ; %bb.2: ; %bb.outer.then
-; GCN-O0-NEXT: s_or_saveexec_b64 s[8:9], -1
; GCN-O0-NEXT: s_waitcnt expcnt(0)
-; GCN-O0-NEXT: buffer_load_dword v0, off, s[12:15], 0 offset:4 ; 4-byte Folded Reload
-; GCN-O0-NEXT: s_mov_b64 exec, s[8:9]
-; GCN-O0-NEXT: buffer_load_dword v1, off, s[12:15], 0 offset:16 ; 4-byte Folded Reload
-; GCN-O0-NEXT: buffer_load_dword v3, off, s[12:15], 0 offset:8 ; 4-byte Folded Reload
-; GCN-O0-NEXT: buffer_load_dword v4, off, s[12:15], 0 offset:12 ; 4-byte Folded Reload
+; GCN-O0-NEXT: buffer_load_dword v0, off, s[12:15], 0 offset:12 ; 4-byte Folded Reload
+; GCN-O0-NEXT: buffer_load_dword v3, off, s[12:15], 0 offset:4 ; 4-byte Folded Reload
+; GCN-O0-NEXT: buffer_load_dword v4, off, s[12:15], 0 offset:8 ; 4-byte Folded Reload
; GCN-O0-NEXT: s_mov_b32 s0, 0xf000
; GCN-O0-NEXT: s_mov_b32 s2, 0
; GCN-O0-NEXT: s_mov_b32 s4, s2
@@ -763,20 +647,17 @@
; GCN-O0-NEXT: s_waitcnt vmcnt(0)
; GCN-O0-NEXT: buffer_store_dword v2, v[3:4], s[0:3], 0 addr64 offset:4
; GCN-O0-NEXT: s_mov_b32 s0, 2
-; GCN-O0-NEXT: v_cmp_eq_u32_e64 s[2:3], v1, s0
+; GCN-O0-NEXT: v_cmp_eq_u32_e64 s[2:3], v0, s0
; GCN-O0-NEXT: s_mov_b64 s[0:1], exec
-; GCN-O0-NEXT: v_writelane_b32 v0, s0, 4
-; GCN-O0-NEXT: v_writelane_b32 v0, s1, 5
-; GCN-O0-NEXT: s_or_saveexec_b64 s[8:9], -1
-; GCN-O0-NEXT: buffer_store_dword v0, off, s[12:15], 0 offset:4 ; 4-byte Folded Spill
-; GCN-O0-NEXT: s_mov_b64 exec, s[8:9]
+; GCN-O0-NEXT: v_writelane_b32 v1, s0, 4
+; GCN-O0-NEXT: v_writelane_b32 v1, s1, 5
; GCN-O0-NEXT: s_and_b64 s[0:1], s[0:1], s[2:3]
; GCN-O0-NEXT: s_mov_b64 exec, s[0:1]
; GCN-O0-NEXT: s_cbranch_execz .LBB3_7
; GCN-O0-NEXT: ; %bb.3: ; %bb.inner.then
-; GCN-O0-NEXT: s_waitcnt expcnt(1)
-; GCN-O0-NEXT: buffer_load_dword v1, off, s[12:15], 0 offset:8 ; 4-byte Folded Reload
-; GCN-O0-NEXT: buffer_load_dword v2, off, s[12:15], 0 offset:12 ; 4-byte Folded Reload
+; GCN-O0-NEXT: s_waitcnt expcnt(0)
+; GCN-O0-NEXT: buffer_load_dword v2, off, s[12:15], 0 offset:4 ; 4-byte Folded Reload
+; GCN-O0-NEXT: buffer_load_dword v3, off, s[12:15], 0 offset:8 ; 4-byte Folded Reload
; GCN-O0-NEXT: s_mov_b32 s0, 0xf000
; GCN-O0-NEXT: s_mov_b32 s2, 0
; GCN-O0-NEXT: s_mov_b32 s4, s2
@@ -785,19 +666,14 @@
; GCN-O0-NEXT: s_mov_b32 s1, s2
; GCN-O0-NEXT: ; kill: def $sgpr0_sgpr1 killed $sgpr0_sgpr1 def $sgpr0_sgpr1_sgpr2_sgpr3
; GCN-O0-NEXT: s_mov_b64 s[2:3], s[4:5]
-; GCN-O0-NEXT: s_waitcnt expcnt(0)
; GCN-O0-NEXT: v_mov_b32_e32 v0, 2
; GCN-O0-NEXT: s_waitcnt vmcnt(0)
-; GCN-O0-NEXT: buffer_store_dword v0, v[1:2], s[0:3], 0 addr64 offset:8
+; GCN-O0-NEXT: buffer_store_dword v0, v[2:3], s[0:3], 0 addr64 offset:8
; GCN-O0-NEXT: s_branch .LBB3_7
; GCN-O0-NEXT: .LBB3_4: ; %bb.outer.else
-; GCN-O0-NEXT: s_or_saveexec_b64 s[8:9], -1
-; GCN-O0-NEXT: s_waitcnt expcnt(0)
-; GCN-O0-NEXT: buffer_load_dword v0, off, s[12:15], 0 offset:4 ; 4-byte Folded Reload
-; GCN-O0-NEXT: s_mov_b64 exec, s[8:9]
-; GCN-O0-NEXT: buffer_load_dword v1, off, s[12:15], 0 offset:16 ; 4-byte Folded Reload
-; GCN-O0-NEXT: buffer_load_dword v3, off, s[12:15], 0 offset:8 ; 4-byte Folded Reload
-; GCN-O0-NEXT: buffer_load_dword v4, off, s[12:15], 0 offset:12 ; 4-byte Folded Reload
+; GCN-O0-NEXT: buffer_load_dword v0, off, s[12:15], 0 offset:12 ; 4-byte Folded Reload
+; GCN-O0-NEXT: buffer_load_dword v3, off, s[12:15], 0 offset:4 ; 4-byte Folded Reload
+; GCN-O0-NEXT: buffer_load_dword v4, off, s[12:15], 0 offset:8 ; 4-byte Folded Reload
; GCN-O0-NEXT: s_mov_b32 s1, 0xf000
; GCN-O0-NEXT: s_mov_b32 s0, 0
; GCN-O0-NEXT: s_mov_b32 s2, s0
@@ -806,23 +682,21 @@
; GCN-O0-NEXT: s_mov_b32 s5, s0
; GCN-O0-NEXT: ; kill: def $sgpr4_sgpr5 killed $sgpr4_sgpr5 def $sgpr4_sgpr5_sgpr6_sgpr7
; GCN-O0-NEXT: s_mov_b64 s[6:7], s[2:3]
+; GCN-O0-NEXT: s_waitcnt expcnt(0)
; GCN-O0-NEXT: v_mov_b32_e32 v2, 3
; GCN-O0-NEXT: s_waitcnt vmcnt(0)
; GCN-O0-NEXT: buffer_store_dword v2, v[3:4], s[4:7], 0 addr64 offset:12
-; GCN-O0-NEXT: v_cmp_eq_u32_e64 s[2:3], v1, s0
+; GCN-O0-NEXT: v_cmp_eq_u32_e64 s[2:3], v0, s0
; GCN-O0-NEXT: s_mov_b64 s[0:1], exec
-; GCN-O0-NEXT: v_writelane_b32 v0, s0, 6
-; GCN-O0-NEXT: v_writelane_b32 v0, s1, 7
-; GCN-O0-NEXT: s_or_saveexec_b64 s[8:9], -1
-; GCN-O0-NEXT: buffer_store_dword v0, off, s[12:15], 0 offset:4 ; 4-byte Folded Spill
-; GCN-O0-NEXT: s_mov_b64 exec, s[8:9]
+; GCN-O0-NEXT: v_writelane_b32 v1, s0, 6
+; GCN-O0-NEXT: v_writelane_b32 v1, s1, 7
; GCN-O0-NEXT: s_and_b64 s[0:1], s[0:1], s[2:3]
; GCN-O0-NEXT: s_mov_b64 exec, s[0:1]
; GCN-O0-NEXT: s_cbranch_execz .LBB3_6
; GCN-O0-NEXT: ; %bb.5: ; %bb.inner.then2
-; GCN-O0-NEXT: s_waitcnt expcnt(1)
-; GCN-O0-NEXT: buffer_load_dword v1, off, s[12:15], 0 offset:8 ; 4-byte Folded Reload
-; GCN-O0-NEXT: buffer_load_dword v2, off, s[12:15], 0 offset:12 ; 4-byte Folded Reload
+; GCN-O0-NEXT: s_waitcnt expcnt(0)
+; GCN-O0-NEXT: buffer_load_dword v2, off, s[12:15], 0 offset:4 ; 4-byte Folded Reload
+; GCN-O0-NEXT: buffer_load_dword v3, off, s[12:15], 0 offset:8 ; 4-byte Folded Reload
; GCN-O0-NEXT: s_mov_b32 s0, 0xf000
; GCN-O0-NEXT: s_mov_b32 s2, 0
; GCN-O0-NEXT: s_mov_b32 s4, s2
@@ -831,43 +705,27 @@
; GCN-O0-NEXT: s_mov_b32 s1, s2
; GCN-O0-NEXT: ; kill: def $sgpr0_sgpr1 killed $sgpr0_sgpr1 def $sgpr0_sgpr1_sgpr2_sgpr3
; GCN-O0-NEXT: s_mov_b64 s[2:3], s[4:5]
-; GCN-O0-NEXT: s_waitcnt expcnt(0)
; GCN-O0-NEXT: v_mov_b32_e32 v0, 4
; GCN-O0-NEXT: s_waitcnt vmcnt(0)
-; GCN-O0-NEXT: buffer_store_dword v0, v[1:2], s[0:3], 0 addr64 offset:16
+; GCN-O0-NEXT: buffer_store_dword v0, v[2:3], s[0:3], 0 addr64 offset:16
; GCN-O0-NEXT: .LBB3_6: ; %Flow
-; GCN-O0-NEXT: s_or_saveexec_b64 s[8:9], -1
-; GCN-O0-NEXT: s_waitcnt expcnt(0)
-; GCN-O0-NEXT: buffer_load_dword v0, off, s[12:15], 0 offset:4 ; 4-byte Folded Reload
-; GCN-O0-NEXT: s_mov_b64 exec, s[8:9]
-; GCN-O0-NEXT: s_waitcnt vmcnt(0)
-; GCN-O0-NEXT: v_readlane_b32 s0, v0, 6
-; GCN-O0-NEXT: v_readlane_b32 s1, v0, 7
+; GCN-O0-NEXT: v_readlane_b32 s0, v1, 6
+; GCN-O0-NEXT: v_readlane_b32 s1, v1, 7
; GCN-O0-NEXT: s_or_b64 exec, exec, s[0:1]
; GCN-O0-NEXT: s_branch .LBB3_1
; GCN-O0-NEXT: .LBB3_7: ; %Flow1
-; GCN-O0-NEXT: s_or_saveexec_b64 s[8:9], -1
-; GCN-O0-NEXT: s_waitcnt expcnt(0)
-; GCN-O0-NEXT: buffer_load_dword v0, off, s[12:15], 0 offset:4 ; 4-byte Folded Reload
-; GCN-O0-NEXT: s_mov_b64 exec, s[8:9]
-; GCN-O0-NEXT: s_waitcnt vmcnt(0)
-; GCN-O0-NEXT: v_readlane_b32 s0, v0, 4
-; GCN-O0-NEXT: v_readlane_b32 s1, v0, 5
+; GCN-O0-NEXT: v_readlane_b32 s0, v1, 4
+; GCN-O0-NEXT: v_readlane_b32 s1, v1, 5
; GCN-O0-NEXT: s_or_b64 exec, exec, s[0:1]
; GCN-O0-NEXT: .LBB3_8: ; %bb.outer.end
-; GCN-O0-NEXT: s_or_saveexec_b64 s[8:9], -1
-; GCN-O0-NEXT: s_waitcnt expcnt(0)
-; GCN-O0-NEXT: buffer_load_dword v0, off, s[12:15], 0 offset:4 ; 4-byte Folded Reload
-; GCN-O0-NEXT: s_mov_b64 exec, s[8:9]
-; GCN-O0-NEXT: s_waitcnt vmcnt(0)
-; GCN-O0-NEXT: v_readlane_b32 s0, v0, 2
-; GCN-O0-NEXT: v_readlane_b32 s1, v0, 3
+; GCN-O0-NEXT: v_readlane_b32 s0, v1, 2
+; GCN-O0-NEXT: v_readlane_b32 s1, v1, 3
; GCN-O0-NEXT: s_or_b64 exec, exec, s[0:1]
+; GCN-O0-NEXT: s_waitcnt expcnt(0)
; GCN-O0-NEXT: v_mov_b32_e32 v2, 3
-; GCN-O0-NEXT: v_mov_b32_e32 v1, 0
+; GCN-O0-NEXT: v_mov_b32_e32 v0, 0
; GCN-O0-NEXT: s_mov_b32 m0, -1
-; GCN-O0-NEXT: ds_write_b32 v1, v2
-; GCN-O0-NEXT: ; kill: killed $vgpr0
+; GCN-O0-NEXT: ds_write_b32 v0, v2
; GCN-O0-NEXT: s_endpgm
bb:
%tmp = tail call i32 @llvm.amdgcn.workitem.id.x()
@@ -931,35 +789,23 @@
; GCN-O0-NEXT: s_mov_b32 s15, 0xe8f000
; GCN-O0-NEXT: s_add_u32 s12, s12, s11
; GCN-O0-NEXT: s_addc_u32 s13, s13, 0
-; GCN-O0-NEXT: ; implicit-def: $vgpr1
-; GCN-O0-NEXT: v_mov_b32_e32 v1, v0
-; GCN-O0-NEXT: s_or_saveexec_b64 s[6:7], -1
-; GCN-O0-NEXT: buffer_load_dword v0, off, s[12:15], 0 offset:4 ; 4-byte Folded Reload
-; GCN-O0-NEXT: s_mov_b64 exec, s[6:7]
; GCN-O0-NEXT: s_load_dwordx2 s[0:1], s[4:5], 0x9
-; GCN-O0-NEXT: s_waitcnt vmcnt(0) lgkmcnt(0)
-; GCN-O0-NEXT: v_writelane_b32 v0, s0, 0
-; GCN-O0-NEXT: v_writelane_b32 v0, s1, 1
-; GCN-O0-NEXT: v_mov_b32_e32 v2, v1
-; GCN-O0-NEXT: buffer_store_dword v2, off, s[12:15], 0 offset:8 ; 4-byte Folded Spill
+; GCN-O0-NEXT: s_waitcnt lgkmcnt(0)
+; GCN-O0-NEXT: v_writelane_b32 v1, s0, 0
+; GCN-O0-NEXT: v_writelane_b32 v1, s1, 1
+; GCN-O0-NEXT: v_mov_b32_e32 v2, v0
+; GCN-O0-NEXT: buffer_store_dword v2, off, s[12:15], 0 offset:4 ; 4-byte Folded Spill
; GCN-O0-NEXT: s_mov_b32 s0, 1
-; GCN-O0-NEXT: v_cmp_gt_u32_e64 s[2:3], v1, s0
+; GCN-O0-NEXT: v_cmp_gt_u32_e64 s[2:3], v0, s0
; GCN-O0-NEXT: s_mov_b64 s[0:1], exec
-; GCN-O0-NEXT: v_writelane_b32 v0, s0, 2
-; GCN-O0-NEXT: v_writelane_b32 v0, s1, 3
-; GCN-O0-NEXT: s_or_saveexec_b64 s[6:7], -1
-; GCN-O0-NEXT: buffer_store_dword v0, off, s[12:15], 0 offset:4 ; 4-byte Folded Spill
-; GCN-O0-NEXT: s_mov_b64 exec, s[6:7]
+; GCN-O0-NEXT: v_writelane_b32 v1, s0, 2
+; GCN-O0-NEXT: v_writelane_b32 v1, s1, 3
; GCN-O0-NEXT: s_and_b64 s[0:1], s[0:1], s[2:3]
; GCN-O0-NEXT: s_mov_b64 exec, s[0:1]
; GCN-O0-NEXT: s_cbranch_execz .LBB4_2
; GCN-O0-NEXT: ; %bb.1: ; %bb.then
; GCN-O0-NEXT: s_waitcnt expcnt(0)
-; GCN-O0-NEXT: buffer_load_dword v0, off, s[12:15], 0 offset:8 ; 4-byte Folded Reload
-; GCN-O0-NEXT: s_or_saveexec_b64 s[6:7], -1
-; GCN-O0-NEXT: buffer_load_dword v1, off, s[12:15], 0 offset:4 ; 4-byte Folded Reload
-; GCN-O0-NEXT: s_mov_b64 exec, s[6:7]
-; GCN-O0-NEXT: s_waitcnt vmcnt(0)
+; GCN-O0-NEXT: buffer_load_dword v2, off, s[12:15], 0 offset:4 ; 4-byte Folded Reload
; GCN-O0-NEXT: v_readlane_b32 s0, v1, 0
; GCN-O0-NEXT: v_readlane_b32 s1, v1, 1
; GCN-O0-NEXT: s_mov_b32 s2, 0xf000
@@ -968,24 +814,20 @@
; GCN-O0-NEXT: s_mov_b32 s5, s2
; GCN-O0-NEXT: ; kill: def $sgpr0_sgpr1 killed $sgpr0_sgpr1 def $sgpr0_sgpr1_sgpr2_sgpr3
; GCN-O0-NEXT: s_mov_b64 s[2:3], s[4:5]
-; GCN-O0-NEXT: v_ashrrev_i32_e64 v2, 31, v0
-; GCN-O0-NEXT: ; kill: def $vgpr0 killed $vgpr0 def $vgpr0_vgpr1 killed $exec
-; GCN-O0-NEXT: v_mov_b32_e32 v1, v2
+; GCN-O0-NEXT: s_waitcnt vmcnt(0)
+; GCN-O0-NEXT: v_ashrrev_i32_e64 v0, 31, v2
+; GCN-O0-NEXT: ; kill: def $vgpr2 killed $vgpr2 def $vgpr2_vgpr3 killed $exec
+; GCN-O0-NEXT: v_mov_b32_e32 v3, v0
; GCN-O0-NEXT: s_mov_b32 s4, 2
-; GCN-O0-NEXT: v_lshl_b64 v[1:2], v[0:1], s4
+; GCN-O0-NEXT: v_lshl_b64 v[2:3], v[2:3], s4
; GCN-O0-NEXT: v_mov_b32_e32 v0, 0
-; GCN-O0-NEXT: buffer_store_dword v0, v[1:2], s[0:3], 0 addr64
+; GCN-O0-NEXT: buffer_store_dword v0, v[2:3], s[0:3], 0 addr64
; GCN-O0-NEXT: .LBB4_2: ; %bb.end
-; GCN-O0-NEXT: s_or_saveexec_b64 s[6:7], -1
-; GCN-O0-NEXT: s_waitcnt expcnt(0)
-; GCN-O0-NEXT: buffer_load_dword v0, off, s[12:15], 0 offset:4 ; 4-byte Folded Reload
-; GCN-O0-NEXT: s_mov_b64 exec, s[6:7]
-; GCN-O0-NEXT: s_waitcnt vmcnt(0)
-; GCN-O0-NEXT: v_readlane_b32 s0, v0, 2
-; GCN-O0-NEXT: v_readlane_b32 s1, v0, 3
+; GCN-O0-NEXT: v_readlane_b32 s0, v1, 2
+; GCN-O0-NEXT: v_readlane_b32 s1, v1, 3
; GCN-O0-NEXT: s_or_b64 exec, exec, s[0:1]
+; GCN-O0-NEXT: s_waitcnt vmcnt(0) expcnt(0)
; GCN-O0-NEXT: s_barrier
-; GCN-O0-NEXT: ; kill: killed $vgpr0
; GCN-O0-NEXT: s_endpgm
bb:
%tmp = tail call i32 @llvm.amdgcn.workitem.id.x()
@@ -1077,127 +919,96 @@
; GCN-O0: ; %bb.0: ; %bb
; GCN-O0-NEXT: s_waitcnt vmcnt(0) expcnt(0) lgkmcnt(0)
; GCN-O0-NEXT: s_xor_saveexec_b64 s[4:5], -1
-; GCN-O0-NEXT: buffer_store_dword v0, off, s[0:3], s32 offset:72 ; 4-byte Folded Spill
-; GCN-O0-NEXT: buffer_store_dword v4, off, s[0:3], s32 offset:76 ; 4-byte Folded Spill
+; GCN-O0-NEXT: buffer_store_dword v1, off, s[0:3], s32 offset:68 ; 4-byte Folded Spill
; GCN-O0-NEXT: s_mov_b64 exec, s[4:5]
-; GCN-O0-NEXT: ; implicit-def: $vgpr1
-; GCN-O0-NEXT: v_mov_b32_e32 v1, v0
-; GCN-O0-NEXT: s_or_saveexec_b64 s[14:15], -1
-; GCN-O0-NEXT: s_waitcnt expcnt(1)
-; GCN-O0-NEXT: buffer_load_dword v0, off, s[0:3], s32 ; 4-byte Folded Reload
-; GCN-O0-NEXT: s_mov_b64 exec, s[14:15]
-; GCN-O0-NEXT: buffer_store_dword v1, off, s[0:3], s32 offset:4 ; 4-byte Folded Spill
+; GCN-O0-NEXT: buffer_store_dword v0, off, s[0:3], s32 ; 4-byte Folded Spill
; GCN-O0-NEXT: s_mov_b64 s[4:5], 0
; GCN-O0-NEXT: s_mov_b64 s[6:7], s[4:5]
-; GCN-O0-NEXT: s_waitcnt vmcnt(1)
-; GCN-O0-NEXT: v_writelane_b32 v0, s6, 0
-; GCN-O0-NEXT: v_writelane_b32 v0, s7, 1
-; GCN-O0-NEXT: v_writelane_b32 v0, s4, 2
-; GCN-O0-NEXT: v_writelane_b32 v0, s5, 3
-; GCN-O0-NEXT: s_or_saveexec_b64 s[14:15], -1
-; GCN-O0-NEXT: buffer_store_dword v0, off, s[0:3], s32 ; 4-byte Folded Spill
-; GCN-O0-NEXT: s_mov_b64 exec, s[14:15]
+; GCN-O0-NEXT: s_waitcnt expcnt(1)
+; GCN-O0-NEXT: v_writelane_b32 v1, s6, 0
+; GCN-O0-NEXT: v_writelane_b32 v1, s7, 1
+; GCN-O0-NEXT: v_writelane_b32 v1, s4, 2
+; GCN-O0-NEXT: v_writelane_b32 v1, s5, 3
; GCN-O0-NEXT: .LBB5_1: ; %bb1
; GCN-O0-NEXT: ; =>This Inner Loop Header: Depth=1
-; GCN-O0-NEXT: s_or_saveexec_b64 s[14:15], -1
; GCN-O0-NEXT: s_waitcnt expcnt(0)
; GCN-O0-NEXT: buffer_load_dword v0, off, s[0:3], s32 ; 4-byte Folded Reload
-; GCN-O0-NEXT: s_mov_b64 exec, s[14:15]
-; GCN-O0-NEXT: buffer_load_dword v1, off, s[0:3], s32 offset:4 ; 4-byte Folded Reload
-; GCN-O0-NEXT: s_waitcnt vmcnt(1)
-; GCN-O0-NEXT: v_readlane_b32 s8, v0, 2
-; GCN-O0-NEXT: v_readlane_b32 s9, v0, 3
-; GCN-O0-NEXT: v_readlane_b32 s6, v0, 0
-; GCN-O0-NEXT: v_readlane_b32 s7, v0, 1
-; GCN-O0-NEXT: v_writelane_b32 v0, s6, 4
-; GCN-O0-NEXT: v_writelane_b32 v0, s7, 5
+; GCN-O0-NEXT: v_readlane_b32 s8, v1, 2
+; GCN-O0-NEXT: v_readlane_b32 s9, v1, 3
+; GCN-O0-NEXT: v_readlane_b32 s6, v1, 0
+; GCN-O0-NEXT: v_readlane_b32 s7, v1, 1
+; GCN-O0-NEXT: v_writelane_b32 v1, s6, 4
+; GCN-O0-NEXT: v_writelane_b32 v1, s7, 5
; GCN-O0-NEXT: s_mov_b32 s4, 0x207
; GCN-O0-NEXT: s_waitcnt vmcnt(0)
-; GCN-O0-NEXT: v_cmp_lt_i32_e64 s[4:5], v1, s4
+; GCN-O0-NEXT: v_cmp_lt_i32_e64 s[4:5], v0, s4
; GCN-O0-NEXT: s_or_b64 s[4:5], s[4:5], s[8:9]
-; GCN-O0-NEXT: v_writelane_b32 v0, s4, 6
-; GCN-O0-NEXT: v_writelane_b32 v0, s5, 7
-; GCN-O0-NEXT: v_writelane_b32 v0, s6, 0
-; GCN-O0-NEXT: v_writelane_b32 v0, s7, 1
+; GCN-O0-NEXT: v_writelane_b32 v1, s4, 6
+; GCN-O0-NEXT: v_writelane_b32 v1, s5, 7
+; GCN-O0-NEXT: v_writelane_b32 v1, s6, 0
+; GCN-O0-NEXT: v_writelane_b32 v1, s7, 1
; GCN-O0-NEXT: s_mov_b64 s[6:7], s[4:5]
-; GCN-O0-NEXT: v_writelane_b32 v0, s6, 2
-; GCN-O0-NEXT: v_writelane_b32 v0, s7, 3
-; GCN-O0-NEXT: s_or_saveexec_b64 s[14:15], -1
-; GCN-O0-NEXT: buffer_store_dword v0, off, s[0:3], s32 ; 4-byte Folded Spill
-; GCN-O0-NEXT: s_mov_b64 exec, s[14:15]
+; GCN-O0-NEXT: v_writelane_b32 v1, s6, 2
+; GCN-O0-NEXT: v_writelane_b32 v1, s7, 3
; GCN-O0-NEXT: s_andn2_b64 exec, exec, s[4:5]
; GCN-O0-NEXT: s_cbranch_execnz .LBB5_1
; GCN-O0-NEXT: ; %bb.2: ; %bb2
; GCN-O0-NEXT: ; in Loop: Header=BB5_1 Depth=1
-; GCN-O0-NEXT: s_or_saveexec_b64 s[14:15], -1
-; GCN-O0-NEXT: s_waitcnt expcnt(0)
; GCN-O0-NEXT: buffer_load_dword v0, off, s[0:3], s32 ; 4-byte Folded Reload
-; GCN-O0-NEXT: s_mov_b64 exec, s[14:15]
-; GCN-O0-NEXT: buffer_load_dword v1, off, s[0:3], s32 offset:4 ; 4-byte Folded Reload
-; GCN-O0-NEXT: s_waitcnt vmcnt(1)
-; GCN-O0-NEXT: v_readlane_b32 s4, v0, 6
-; GCN-O0-NEXT: v_readlane_b32 s5, v0, 7
+; GCN-O0-NEXT: v_readlane_b32 s4, v1, 6
+; GCN-O0-NEXT: v_readlane_b32 s5, v1, 7
; GCN-O0-NEXT: s_or_b64 exec, exec, s[4:5]
; GCN-O0-NEXT: s_mov_b32 s6, 0
; GCN-O0-NEXT: s_waitcnt vmcnt(0)
-; GCN-O0-NEXT: v_cmp_ne_u32_e64 s[4:5], v1, s6
-; GCN-O0-NEXT: v_cmp_eq_u32_e64 s[6:7], v1, s6
-; GCN-O0-NEXT: v_writelane_b32 v0, s4, 8
-; GCN-O0-NEXT: v_writelane_b32 v0, s5, 9
+; GCN-O0-NEXT: v_cmp_ne_u32_e64 s[4:5], v0, s6
+; GCN-O0-NEXT: v_cmp_eq_u32_e64 s[6:7], v0, s6
+; GCN-O0-NEXT: v_writelane_b32 v1, s4, 8
+; GCN-O0-NEXT: v_writelane_b32 v1, s5, 9
; GCN-O0-NEXT: s_mov_b32 s4, 0
; GCN-O0-NEXT: s_mov_b32 s8, s4
; GCN-O0-NEXT: s_mov_b32 s9, s4
; GCN-O0-NEXT: s_mov_b32 s10, s4
; GCN-O0-NEXT: s_mov_b32 s11, s4
-; GCN-O0-NEXT: v_mov_b32_e32 v1, s8
-; GCN-O0-NEXT: v_mov_b32_e32 v2, s9
-; GCN-O0-NEXT: v_mov_b32_e32 v3, s10
-; GCN-O0-NEXT: v_mov_b32_e32 v4, s11
-; GCN-O0-NEXT: buffer_store_dword v1, off, s[0:3], s32 offset:8 ; 4-byte Folded Spill
+; GCN-O0-NEXT: v_mov_b32_e32 v2, s8
+; GCN-O0-NEXT: v_mov_b32_e32 v3, s9
+; GCN-O0-NEXT: v_mov_b32_e32 v4, s10
+; GCN-O0-NEXT: v_mov_b32_e32 v5, s11
+; GCN-O0-NEXT: buffer_store_dword v2, off, s[0:3], s32 offset:4 ; 4-byte Folded Spill
; GCN-O0-NEXT: s_waitcnt vmcnt(0)
-; GCN-O0-NEXT: buffer_store_dword v2, off, s[0:3], s32 offset:12 ; 4-byte Folded Spill
-; GCN-O0-NEXT: buffer_store_dword v3, off, s[0:3], s32 offset:16 ; 4-byte Folded Spill
-; GCN-O0-NEXT: buffer_store_dword v4, off, s[0:3], s32 offset:20 ; 4-byte Folded Spill
+; GCN-O0-NEXT: buffer_store_dword v3, off, s[0:3], s32 offset:8 ; 4-byte Folded Spill
+; GCN-O0-NEXT: buffer_store_dword v4, off, s[0:3], s32 offset:12 ; 4-byte Folded Spill
+; GCN-O0-NEXT: buffer_store_dword v5, off, s[0:3], s32 offset:16 ; 4-byte Folded Spill
; GCN-O0-NEXT: s_mov_b64 s[4:5], exec
-; GCN-O0-NEXT: v_writelane_b32 v0, s4, 10
-; GCN-O0-NEXT: v_writelane_b32 v0, s5, 11
-; GCN-O0-NEXT: s_or_saveexec_b64 s[14:15], -1
-; GCN-O0-NEXT: buffer_store_dword v0, off, s[0:3], s32 ; 4-byte Folded Spill
-; GCN-O0-NEXT: s_mov_b64 exec, s[14:15]
+; GCN-O0-NEXT: v_writelane_b32 v1, s4, 10
+; GCN-O0-NEXT: v_writelane_b32 v1, s5, 11
; GCN-O0-NEXT: s_and_b64 s[4:5], s[4:5], s[6:7]
; GCN-O0-NEXT: s_mov_b64 exec, s[4:5]
; GCN-O0-NEXT: s_cbranch_execz .LBB5_5
; GCN-O0-NEXT: ; %bb.3: ; %bb4
; GCN-O0-NEXT: ; in Loop: Header=BB5_1 Depth=1
-; GCN-O0-NEXT: s_or_saveexec_b64 s[14:15], -1
-; GCN-O0-NEXT: s_waitcnt expcnt(0)
-; GCN-O0-NEXT: buffer_load_dword v0, off, s[0:3], s32 ; 4-byte Folded Reload
-; GCN-O0-NEXT: s_mov_b64 exec, s[14:15]
; GCN-O0-NEXT: ; implicit-def: $sgpr4
-; GCN-O0-NEXT: v_mov_b32_e32 v1, s4
-; GCN-O0-NEXT: buffer_load_dword v1, v1, s[0:3], 0 offen
+; GCN-O0-NEXT: v_mov_b32_e32 v0, s4
+; GCN-O0-NEXT: buffer_load_dword v0, v0, s[0:3], 0 offen
; GCN-O0-NEXT: s_mov_b32 s4, 0
; GCN-O0-NEXT: s_waitcnt vmcnt(0)
-; GCN-O0-NEXT: v_cmp_lt_f32_e64 s[6:7], v1, s4
+; GCN-O0-NEXT: v_cmp_lt_f32_e64 s[6:7], v0, s4
; GCN-O0-NEXT: s_mov_b32 s8, s4
; GCN-O0-NEXT: s_mov_b32 s9, s4
; GCN-O0-NEXT: s_mov_b32 s10, s4
; GCN-O0-NEXT: s_mov_b32 s11, s4
-; GCN-O0-NEXT: v_mov_b32_e32 v1, s8
-; GCN-O0-NEXT: v_mov_b32_e32 v2, s9
-; GCN-O0-NEXT: v_mov_b32_e32 v3, s10
-; GCN-O0-NEXT: v_mov_b32_e32 v4, s11
-; GCN-O0-NEXT: buffer_store_dword v1, off, s[0:3], s32 offset:24 ; 4-byte Folded Spill
+; GCN-O0-NEXT: s_waitcnt expcnt(0)
+; GCN-O0-NEXT: v_mov_b32_e32 v2, s8
+; GCN-O0-NEXT: v_mov_b32_e32 v3, s9
+; GCN-O0-NEXT: v_mov_b32_e32 v4, s10
+; GCN-O0-NEXT: v_mov_b32_e32 v5, s11
+; GCN-O0-NEXT: buffer_store_dword v2, off, s[0:3], s32 offset:20 ; 4-byte Folded Spill
; GCN-O0-NEXT: s_waitcnt vmcnt(0)
-; GCN-O0-NEXT: buffer_store_dword v2, off, s[0:3], s32 offset:28 ; 4-byte Folded Spill
-; GCN-O0-NEXT: buffer_store_dword v3, off, s[0:3], s32 offset:32 ; 4-byte Folded Spill
-; GCN-O0-NEXT: buffer_store_dword v4, off, s[0:3], s32 offset:36 ; 4-byte Folded Spill
+; GCN-O0-NEXT: buffer_store_dword v3, off, s[0:3], s32 offset:24 ; 4-byte Folded Spill
+; GCN-O0-NEXT: buffer_store_dword v4, off, s[0:3], s32 offset:28 ; 4-byte Folded Spill
+; GCN-O0-NEXT: buffer_store_dword v5, off, s[0:3], s32 offset:32 ; 4-byte Folded Spill
; GCN-O0-NEXT: s_mov_b64 s[4:5], exec
-; GCN-O0-NEXT: v_writelane_b32 v0, s4, 12
-; GCN-O0-NEXT: v_writelane_b32 v0, s5, 13
-; GCN-O0-NEXT: s_or_saveexec_b64 s[14:15], -1
-; GCN-O0-NEXT: buffer_store_dword v0, off, s[0:3], s32 ; 4-byte Folded Spill
-; GCN-O0-NEXT: s_mov_b64 exec, s[14:15]
+; GCN-O0-NEXT: v_writelane_b32 v1, s4, 12
+; GCN-O0-NEXT: v_writelane_b32 v1, s5, 13
; GCN-O0-NEXT: s_and_b64 s[4:5], s[4:5], s[6:7]
; GCN-O0-NEXT: s_mov_b64 exec, s[4:5]
; GCN-O0-NEXT: s_cbranch_execz .LBB5_6
@@ -1215,173 +1026,135 @@
; GCN-O0-NEXT: s_mov_b32 s6, s9
; GCN-O0-NEXT: s_mov_b32 s7, s8
; GCN-O0-NEXT: s_waitcnt expcnt(0)
-; GCN-O0-NEXT: v_mov_b32_e32 v0, s4
-; GCN-O0-NEXT: v_mov_b32_e32 v1, s5
-; GCN-O0-NEXT: v_mov_b32_e32 v2, s6
-; GCN-O0-NEXT: v_mov_b32_e32 v3, s7
-; GCN-O0-NEXT: buffer_store_dword v0, off, s[0:3], s32 offset:24 ; 4-byte Folded Spill
+; GCN-O0-NEXT: v_mov_b32_e32 v2, s4
+; GCN-O0-NEXT: v_mov_b32_e32 v3, s5
+; GCN-O0-NEXT: v_mov_b32_e32 v4, s6
+; GCN-O0-NEXT: v_mov_b32_e32 v5, s7
+; GCN-O0-NEXT: buffer_store_dword v2, off, s[0:3], s32 offset:20 ; 4-byte Folded Spill
; GCN-O0-NEXT: s_waitcnt vmcnt(0)
-; GCN-O0-NEXT: buffer_store_dword v1, off, s[0:3], s32 offset:28 ; 4-byte Folded Spill
-; GCN-O0-NEXT: buffer_store_dword v2, off, s[0:3], s32 offset:32 ; 4-byte Folded Spill
-; GCN-O0-NEXT: buffer_store_dword v3, off, s[0:3], s32 offset:36 ; 4-byte Folded Spill
+; GCN-O0-NEXT: buffer_store_dword v3, off, s[0:3], s32 offset:24 ; 4-byte Folded Spill
+; GCN-O0-NEXT: buffer_store_dword v4, off, s[0:3], s32 offset:28 ; 4-byte Folded Spill
+; GCN-O0-NEXT: buffer_store_dword v5, off, s[0:3], s32 offset:32 ; 4-byte Folded Spill
; GCN-O0-NEXT: s_branch .LBB5_6
; GCN-O0-NEXT: .LBB5_5: ; %Flow2
; GCN-O0-NEXT: ; in Loop: Header=BB5_1 Depth=1
; GCN-O0-NEXT: s_waitcnt expcnt(0)
-; GCN-O0-NEXT: buffer_load_dword v0, off, s[0:3], s32 offset:8 ; 4-byte Folded Reload
-; GCN-O0-NEXT: buffer_load_dword v1, off, s[0:3], s32 offset:12 ; 4-byte Folded Reload
-; GCN-O0-NEXT: buffer_load_dword v2, off, s[0:3], s32 offset:16 ; 4-byte Folded Reload
-; GCN-O0-NEXT: buffer_load_dword v3, off, s[0:3], s32 offset:20 ; 4-byte Folded Reload
-; GCN-O0-NEXT: s_or_saveexec_b64 s[14:15], -1
-; GCN-O0-NEXT: buffer_load_dword v4, off, s[0:3], s32 ; 4-byte Folded Reload
-; GCN-O0-NEXT: s_mov_b64 exec, s[14:15]
-; GCN-O0-NEXT: s_waitcnt vmcnt(0)
-; GCN-O0-NEXT: v_readlane_b32 s4, v4, 10
-; GCN-O0-NEXT: v_readlane_b32 s5, v4, 11
+; GCN-O0-NEXT: buffer_load_dword v2, off, s[0:3], s32 offset:4 ; 4-byte Folded Reload
+; GCN-O0-NEXT: buffer_load_dword v3, off, s[0:3], s32 offset:8 ; 4-byte Folded Reload
+; GCN-O0-NEXT: buffer_load_dword v4, off, s[0:3], s32 offset:12 ; 4-byte Folded Reload
+; GCN-O0-NEXT: buffer_load_dword v5, off, s[0:3], s32 offset:16 ; 4-byte Folded Reload
+; GCN-O0-NEXT: v_readlane_b32 s4, v1, 10
+; GCN-O0-NEXT: v_readlane_b32 s5, v1, 11
; GCN-O0-NEXT: s_or_b64 exec, exec, s[4:5]
-; GCN-O0-NEXT: buffer_store_dword v0, off, s[0:3], s32 offset:40 ; 4-byte Folded Spill
; GCN-O0-NEXT: s_waitcnt vmcnt(0)
-; GCN-O0-NEXT: buffer_store_dword v1, off, s[0:3], s32 offset:44 ; 4-byte Folded Spill
-; GCN-O0-NEXT: buffer_store_dword v2, off, s[0:3], s32 offset:48 ; 4-byte Folded Spill
-; GCN-O0-NEXT: buffer_store_dword v3, off, s[0:3], s32 offset:52 ; 4-byte Folded Spill
+; GCN-O0-NEXT: buffer_store_dword v2, off, s[0:3], s32 offset:36 ; 4-byte Folded Spill
+; GCN-O0-NEXT: s_waitcnt vmcnt(0)
+; GCN-O0-NEXT: buffer_store_dword v3, off, s[0:3], s32 offset:40 ; 4-byte Folded Spill
+; GCN-O0-NEXT: buffer_store_dword v4, off, s[0:3], s32 offset:44 ; 4-byte Folded Spill
+; GCN-O0-NEXT: buffer_store_dword v5, off, s[0:3], s32 offset:48 ; 4-byte Folded Spill
; GCN-O0-NEXT: s_branch .LBB5_7
; GCN-O0-NEXT: .LBB5_6: ; %Flow
; GCN-O0-NEXT: ; in Loop: Header=BB5_1 Depth=1
; GCN-O0-NEXT: s_waitcnt expcnt(0)
-; GCN-O0-NEXT: buffer_load_dword v0, off, s[0:3], s32 offset:24 ; 4-byte Folded Reload
-; GCN-O0-NEXT: buffer_load_dword v1, off, s[0:3], s32 offset:28 ; 4-byte Folded Reload
-; GCN-O0-NEXT: buffer_load_dword v2, off, s[0:3], s32 offset:32 ; 4-byte Folded Reload
-; GCN-O0-NEXT: buffer_load_dword v3, off, s[0:3], s32 offset:36 ; 4-byte Folded Reload
-; GCN-O0-NEXT: s_or_saveexec_b64 s[14:15], -1
-; GCN-O0-NEXT: buffer_load_dword v4, off, s[0:3], s32 ; 4-byte Folded Reload
-; GCN-O0-NEXT: s_mov_b64 exec, s[14:15]
-; GCN-O0-NEXT: s_waitcnt vmcnt(0)
-; GCN-O0-NEXT: v_readlane_b32 s4, v4, 12
-; GCN-O0-NEXT: v_readlane_b32 s5, v4, 13
+; GCN-O0-NEXT: buffer_load_dword v2, off, s[0:3], s32 offset:20 ; 4-byte Folded Reload
+; GCN-O0-NEXT: buffer_load_dword v3, off, s[0:3], s32 offset:24 ; 4-byte Folded Reload
+; GCN-O0-NEXT: buffer_load_dword v4, off, s[0:3], s32 offset:28 ; 4-byte Folded Reload
+; GCN-O0-NEXT: buffer_load_dword v5, off, s[0:3], s32 offset:32 ; 4-byte Folded Reload
+; GCN-O0-NEXT: v_readlane_b32 s4, v1, 12
+; GCN-O0-NEXT: v_readlane_b32 s5, v1, 13
; GCN-O0-NEXT: s_or_b64 exec, exec, s[4:5]
-; GCN-O0-NEXT: buffer_store_dword v0, off, s[0:3], s32 offset:8 ; 4-byte Folded Spill
; GCN-O0-NEXT: s_waitcnt vmcnt(0)
-; GCN-O0-NEXT: buffer_store_dword v1, off, s[0:3], s32 offset:12 ; 4-byte Folded Spill
-; GCN-O0-NEXT: buffer_store_dword v2, off, s[0:3], s32 offset:16 ; 4-byte Folded Spill
-; GCN-O0-NEXT: buffer_store_dword v3, off, s[0:3], s32 offset:20 ; 4-byte Folded Spill
+; GCN-O0-NEXT: buffer_store_dword v2, off, s[0:3], s32 offset:4 ; 4-byte Folded Spill
+; GCN-O0-NEXT: s_waitcnt vmcnt(0)
+; GCN-O0-NEXT: buffer_store_dword v3, off, s[0:3], s32 offset:8 ; 4-byte Folded Spill
+; GCN-O0-NEXT: buffer_store_dword v4, off, s[0:3], s32 offset:12 ; 4-byte Folded Spill
+; GCN-O0-NEXT: buffer_store_dword v5, off, s[0:3], s32 offset:16 ; 4-byte Folded Spill
; GCN-O0-NEXT: s_branch .LBB5_5
; GCN-O0-NEXT: .LBB5_7: ; %bb10
; GCN-O0-NEXT: ; in Loop: Header=BB5_1 Depth=1
-; GCN-O0-NEXT: s_or_saveexec_b64 s[14:15], -1
-; GCN-O0-NEXT: s_waitcnt expcnt(3)
-; GCN-O0-NEXT: buffer_load_dword v0, off, s[0:3], s32 ; 4-byte Folded Reload
-; GCN-O0-NEXT: s_mov_b64 exec, s[14:15]
-; GCN-O0-NEXT: s_waitcnt vmcnt(0)
-; GCN-O0-NEXT: v_readlane_b32 s6, v0, 8
-; GCN-O0-NEXT: v_readlane_b32 s7, v0, 9
+; GCN-O0-NEXT: v_readlane_b32 s6, v1, 8
+; GCN-O0-NEXT: v_readlane_b32 s7, v1, 9
; GCN-O0-NEXT: s_mov_b64 s[4:5], -1
-; GCN-O0-NEXT: v_writelane_b32 v0, s4, 14
-; GCN-O0-NEXT: v_writelane_b32 v0, s5, 15
+; GCN-O0-NEXT: v_writelane_b32 v1, s4, 14
+; GCN-O0-NEXT: v_writelane_b32 v1, s5, 15
; GCN-O0-NEXT: s_mov_b64 s[4:5], exec
-; GCN-O0-NEXT: v_writelane_b32 v0, s4, 16
-; GCN-O0-NEXT: v_writelane_b32 v0, s5, 17
-; GCN-O0-NEXT: s_or_saveexec_b64 s[14:15], -1
-; GCN-O0-NEXT: buffer_store_dword v0, off, s[0:3], s32 ; 4-byte Folded Spill
-; GCN-O0-NEXT: s_mov_b64 exec, s[14:15]
+; GCN-O0-NEXT: v_writelane_b32 v1, s4, 16
+; GCN-O0-NEXT: v_writelane_b32 v1, s5, 17
; GCN-O0-NEXT: s_and_b64 s[4:5], s[4:5], s[6:7]
; GCN-O0-NEXT: s_mov_b64 exec, s[4:5]
; GCN-O0-NEXT: s_cbranch_execz .LBB5_9
; GCN-O0-NEXT: ; %bb.8: ; %Flow1
; GCN-O0-NEXT: ; in Loop: Header=BB5_1 Depth=1
-; GCN-O0-NEXT: s_or_saveexec_b64 s[14:15], -1
-; GCN-O0-NEXT: s_waitcnt expcnt(0)
-; GCN-O0-NEXT: buffer_load_dword v0, off, s[0:3], s32 ; 4-byte Folded Reload
-; GCN-O0-NEXT: s_mov_b64 exec, s[14:15]
; GCN-O0-NEXT: s_mov_b64 s[4:5], 0
; GCN-O0-NEXT: s_xor_b64 s[4:5], exec, -1
-; GCN-O0-NEXT: s_waitcnt vmcnt(0)
-; GCN-O0-NEXT: v_writelane_b32 v0, s4, 14
-; GCN-O0-NEXT: v_writelane_b32 v0, s5, 15
-; GCN-O0-NEXT: s_or_saveexec_b64 s[14:15], -1
-; GCN-O0-NEXT: buffer_store_dword v0, off, s[0:3], s32 ; 4-byte Folded Spill
-; GCN-O0-NEXT: s_mov_b64 exec, s[14:15]
+; GCN-O0-NEXT: v_writelane_b32 v1, s4, 14
+; GCN-O0-NEXT: v_writelane_b32 v1, s5, 15
; GCN-O0-NEXT: .LBB5_9: ; %Flow3
; GCN-O0-NEXT: ; in Loop: Header=BB5_1 Depth=1
; GCN-O0-NEXT: s_waitcnt expcnt(0)
-; GCN-O0-NEXT: buffer_load_dword v0, off, s[0:3], s32 offset:40 ; 4-byte Folded Reload
-; GCN-O0-NEXT: buffer_load_dword v1, off, s[0:3], s32 offset:44 ; 4-byte Folded Reload
-; GCN-O0-NEXT: buffer_load_dword v2, off, s[0:3], s32 offset:48 ; 4-byte Folded Reload
-; GCN-O0-NEXT: buffer_load_dword v3, off, s[0:3], s32 offset:52 ; 4-byte Folded Reload
-; GCN-O0-NEXT: s_or_saveexec_b64 s[14:15], -1
-; GCN-O0-NEXT: buffer_load_dword v4, off, s[0:3], s32 ; 4-byte Folded Reload
-; GCN-O0-NEXT: s_mov_b64 exec, s[14:15]
-; GCN-O0-NEXT: s_waitcnt vmcnt(0)
-; GCN-O0-NEXT: v_readlane_b32 s8, v4, 16
-; GCN-O0-NEXT: v_readlane_b32 s9, v4, 17
+; GCN-O0-NEXT: buffer_load_dword v2, off, s[0:3], s32 offset:36 ; 4-byte Folded Reload
+; GCN-O0-NEXT: buffer_load_dword v3, off, s[0:3], s32 offset:40 ; 4-byte Folded Reload
+; GCN-O0-NEXT: buffer_load_dword v4, off, s[0:3], s32 offset:44 ; 4-byte Folded Reload
+; GCN-O0-NEXT: buffer_load_dword v5, off, s[0:3], s32 offset:48 ; 4-byte Folded Reload
+; GCN-O0-NEXT: v_readlane_b32 s8, v1, 16
+; GCN-O0-NEXT: v_readlane_b32 s9, v1, 17
; GCN-O0-NEXT: s_or_b64 exec, exec, s[8:9]
-; GCN-O0-NEXT: v_readlane_b32 s6, v4, 4
-; GCN-O0-NEXT: v_readlane_b32 s7, v4, 5
-; GCN-O0-NEXT: v_readlane_b32 s4, v4, 14
-; GCN-O0-NEXT: v_readlane_b32 s5, v4, 15
+; GCN-O0-NEXT: v_readlane_b32 s6, v1, 4
+; GCN-O0-NEXT: v_readlane_b32 s7, v1, 5
+; GCN-O0-NEXT: v_readlane_b32 s4, v1, 14
+; GCN-O0-NEXT: v_readlane_b32 s5, v1, 15
; GCN-O0-NEXT: s_and_b64 s[4:5], exec, s[4:5]
; GCN-O0-NEXT: s_or_b64 s[4:5], s[4:5], s[6:7]
; GCN-O0-NEXT: s_mov_b64 s[6:7], 0
; GCN-O0-NEXT: s_mov_b64 s[8:9], s[4:5]
-; GCN-O0-NEXT: v_writelane_b32 v4, s8, 0
-; GCN-O0-NEXT: v_writelane_b32 v4, s9, 1
-; GCN-O0-NEXT: v_writelane_b32 v4, s6, 2
-; GCN-O0-NEXT: v_writelane_b32 v4, s7, 3
+; GCN-O0-NEXT: v_writelane_b32 v1, s8, 0
+; GCN-O0-NEXT: v_writelane_b32 v1, s9, 1
+; GCN-O0-NEXT: v_writelane_b32 v1, s6, 2
+; GCN-O0-NEXT: v_writelane_b32 v1, s7, 3
; GCN-O0-NEXT: s_mov_b64 s[6:7], s[4:5]
-; GCN-O0-NEXT: v_writelane_b32 v4, s6, 18
-; GCN-O0-NEXT: v_writelane_b32 v4, s7, 19
-; GCN-O0-NEXT: s_or_saveexec_b64 s[14:15], -1
-; GCN-O0-NEXT: buffer_store_dword v4, off, s[0:3], s32 ; 4-byte Folded Spill
-; GCN-O0-NEXT: s_mov_b64 exec, s[14:15]
-; GCN-O0-NEXT: buffer_store_dword v0, off, s[0:3], s32 offset:56 ; 4-byte Folded Spill
+; GCN-O0-NEXT: v_writelane_b32 v1, s6, 18
+; GCN-O0-NEXT: v_writelane_b32 v1, s7, 19
; GCN-O0-NEXT: s_waitcnt vmcnt(0)
-; GCN-O0-NEXT: buffer_store_dword v1, off, s[0:3], s32 offset:60 ; 4-byte Folded Spill
-; GCN-O0-NEXT: buffer_store_dword v2, off, s[0:3], s32 offset:64 ; 4-byte Folded Spill
-; GCN-O0-NEXT: buffer_store_dword v3, off, s[0:3], s32 offset:68 ; 4-byte Folded Spill
+; GCN-O0-NEXT: buffer_store_dword v2, off, s[0:3], s32 offset:52 ; 4-byte Folded Spill
+; GCN-O0-NEXT: s_waitcnt vmcnt(0)
+; GCN-O0-NEXT: buffer_store_dword v3, off, s[0:3], s32 offset:56 ; 4-byte Folded Spill
+; GCN-O0-NEXT: buffer_store_dword v4, off, s[0:3], s32 offset:60 ; 4-byte Folded Spill
+; GCN-O0-NEXT: buffer_store_dword v5, off, s[0:3], s32 offset:64 ; 4-byte Folded Spill
; GCN-O0-NEXT: s_andn2_b64 exec, exec, s[4:5]
; GCN-O0-NEXT: s_cbranch_execnz .LBB5_1
; GCN-O0-NEXT: ; %bb.10: ; %bb12
-; GCN-O0-NEXT: s_or_saveexec_b64 s[14:15], -1
-; GCN-O0-NEXT: s_waitcnt expcnt(3)
-; GCN-O0-NEXT: buffer_load_dword v0, off, s[0:3], s32 ; 4-byte Folded Reload
-; GCN-O0-NEXT: s_mov_b64 exec, s[14:15]
-; GCN-O0-NEXT: s_waitcnt vmcnt(0)
-; GCN-O0-NEXT: v_readlane_b32 s4, v0, 18
-; GCN-O0-NEXT: v_readlane_b32 s5, v0, 19
+; GCN-O0-NEXT: v_readlane_b32 s4, v1, 18
+; GCN-O0-NEXT: v_readlane_b32 s5, v1, 19
; GCN-O0-NEXT: s_or_b64 exec, exec, s[4:5]
; GCN-O0-NEXT: ; %bb.11: ; %bb12
-; GCN-O0-NEXT: s_or_saveexec_b64 s[14:15], -1
-; GCN-O0-NEXT: buffer_load_dword v0, off, s[0:3], s32 ; 4-byte Folded Reload
-; GCN-O0-NEXT: s_mov_b64 exec, s[14:15]
; GCN-O0-NEXT: s_waitcnt expcnt(0)
-; GCN-O0-NEXT: buffer_load_dword v1, off, s[0:3], s32 offset:56 ; 4-byte Folded Reload
-; GCN-O0-NEXT: buffer_load_dword v2, off, s[0:3], s32 offset:60 ; 4-byte Folded Reload
-; GCN-O0-NEXT: buffer_load_dword v3, off, s[0:3], s32 offset:64 ; 4-byte Folded Reload
-; GCN-O0-NEXT: buffer_load_dword v4, off, s[0:3], s32 offset:68 ; 4-byte Folded Reload
+; GCN-O0-NEXT: buffer_load_dword v2, off, s[0:3], s32 offset:52 ; 4-byte Folded Reload
+; GCN-O0-NEXT: buffer_load_dword v3, off, s[0:3], s32 offset:56 ; 4-byte Folded Reload
+; GCN-O0-NEXT: buffer_load_dword v4, off, s[0:3], s32 offset:60 ; 4-byte Folded Reload
+; GCN-O0-NEXT: buffer_load_dword v5, off, s[0:3], s32 offset:64 ; 4-byte Folded Reload
; GCN-O0-NEXT: s_waitcnt vmcnt(0)
-; GCN-O0-NEXT: v_mov_b32_e32 v5, v4
+; GCN-O0-NEXT: v_mov_b32_e32 v0, v5
; GCN-O0-NEXT: ; implicit-def: $sgpr4
; GCN-O0-NEXT: v_mov_b32_e32 v6, s4
-; GCN-O0-NEXT: buffer_store_dword v5, v6, s[0:3], 0 offen
+; GCN-O0-NEXT: buffer_store_dword v0, v6, s[0:3], 0 offen
; GCN-O0-NEXT: s_waitcnt vmcnt(0) expcnt(0)
-; GCN-O0-NEXT: v_mov_b32_e32 v5, v3
+; GCN-O0-NEXT: v_mov_b32_e32 v0, v4
; GCN-O0-NEXT: ; implicit-def: $sgpr4
; GCN-O0-NEXT: v_mov_b32_e32 v6, s4
-; GCN-O0-NEXT: buffer_store_dword v5, v6, s[0:3], 0 offen
+; GCN-O0-NEXT: buffer_store_dword v0, v6, s[0:3], 0 offen
; GCN-O0-NEXT: s_waitcnt vmcnt(0) expcnt(0)
-; GCN-O0-NEXT: v_mov_b32_e32 v5, v2
+; GCN-O0-NEXT: v_mov_b32_e32 v0, v3
; GCN-O0-NEXT: ; implicit-def: $sgpr4
; GCN-O0-NEXT: v_mov_b32_e32 v6, s4
-; GCN-O0-NEXT: buffer_store_dword v5, v6, s[0:3], 0 offen
-; GCN-O0-NEXT: s_waitcnt vmcnt(0)
-; GCN-O0-NEXT: ; kill: def $vgpr1 killed $vgpr1 killed $vgpr1_vgpr2_vgpr3_vgpr4 killed $exec
+; GCN-O0-NEXT: buffer_store_dword v0, v6, s[0:3], 0 offen
+; GCN-O0-NEXT: s_waitcnt vmcnt(0) expcnt(0)
+; GCN-O0-NEXT: v_mov_b32_e32 v0, v2
; GCN-O0-NEXT: ; implicit-def: $sgpr4
; GCN-O0-NEXT: v_mov_b32_e32 v2, s4
-; GCN-O0-NEXT: buffer_store_dword v1, v2, s[0:3], 0 offen
+; GCN-O0-NEXT: buffer_store_dword v0, v2, s[0:3], 0 offen
; GCN-O0-NEXT: s_waitcnt vmcnt(0)
-; GCN-O0-NEXT: ; kill: killed $vgpr0
; GCN-O0-NEXT: s_xor_saveexec_b64 s[4:5], -1
-; GCN-O0-NEXT: buffer_load_dword v0, off, s[0:3], s32 offset:72 ; 4-byte Folded Reload
-; GCN-O0-NEXT: buffer_load_dword v4, off, s[0:3], s32 offset:76 ; 4-byte Folded Reload
+; GCN-O0-NEXT: buffer_load_dword v1, off, s[0:3], s32 offset:68 ; 4-byte Folded Reload
; GCN-O0-NEXT: s_mov_b64 exec, s[4:5]
; GCN-O0-NEXT: s_waitcnt vmcnt(0) expcnt(0)
; GCN-O0-NEXT: s_setpc_b64 s[30:31]
diff --git a/llvm/test/CodeGen/AMDGPU/control-flow-fastregalloc.ll b/llvm/test/CodeGen/AMDGPU/control-flow-fastregalloc.ll
--- a/llvm/test/CodeGen/AMDGPU/control-flow-fastregalloc.ll
+++ b/llvm/test/CodeGen/AMDGPU/control-flow-fastregalloc.ll
@@ -10,7 +10,7 @@
; GCN-LABEL: {{^}}divergent_if_endif:
-; VGPR: workitem_private_segment_byte_size = 16{{$}}
+; VGPR: workitem_private_segment_byte_size = 12{{$}}
; GCN: {{^}}; %bb.0:
@@ -19,7 +19,7 @@
; Spill load
; GCN: buffer_store_dword [[LOAD0]], off, s[0:3], 0 offset:[[LOAD0_OFFSET:[0-9]+]] ; 4-byte Folded Spill
-; GCN: v_cmp_eq_u32_e64 [[CMP0:s\[[0-9]+:[0-9]\]]], v{{[0-9]+}}, s{{[0-9]+}}
+; GCN: v_cmp_eq_u32_e64 [[CMP0:s\[[0-9]+:[0-9]\]]], v0, s{{[0-9]+}}
; Spill saved exec
; GCN: s_mov_b64 s[[[SAVEEXEC_LO:[0-9]+]]:[[SAVEEXEC_HI:[0-9]+]]], exec
@@ -82,13 +82,13 @@
}
; GCN-LABEL: {{^}}divergent_loop:
-; VGPR: workitem_private_segment_byte_size = 20{{$}}
+; VGPR: workitem_private_segment_byte_size = 16{{$}}
; GCN: {{^}}; %bb.0:
; GCN-DAG: s_mov_b32 m0, -1
; GCN-DAG: v_mov_b32_e32 [[PTR0:v[0-9]+]], 0{{$}}
; GCN: ds_read_b32 [[LOAD0:v[0-9]+]], [[PTR0]]
-; GCN: v_cmp_eq_u32_e64 [[CMP0:s\[[0-9]+:[0-9]+\]]], v{{[0-9]+}}, s{{[0-9]+}}
+; GCN: v_cmp_eq_u32_e64 [[CMP0:s\[[0-9]+:[0-9]+\]]], v0, s{{[0-9]+}}
; Spill load
; GCN: buffer_store_dword [[LOAD0]], off, s[0:3], 0 offset:[[LOAD0_OFFSET:[0-9]+]] ; 4-byte Folded Spill
@@ -166,7 +166,7 @@
; GCN: buffer_store_dword [[LOAD0]], off, s[0:3], 0 offset:[[LOAD0_OFFSET:[0-9]+]] ; 4-byte Folded Spill
; GCN: s_mov_b32 [[ZERO:s[0-9]+]], 0
-; GCN: v_cmp_ne_u32_e64 [[CMP0:s\[[0-9]+:[0-9]\]]], v{{[0-9]+}}, [[ZERO]]
+; GCN: v_cmp_ne_u32_e64 [[CMP0:s\[[0-9]+:[0-9]\]]], v0, [[ZERO]]
; GCN: s_mov_b64 s[[[SAVEEXEC_LO:[0-9]+]]:[[SAVEEXEC_HI:[0-9]+]]], exec
; GCN: s_and_b64 s[[[ANDEXEC_LO:[0-9]+]]:[[ANDEXEC_HI:[0-9]+]]], s[[[SAVEEXEC_LO:[0-9]+]]:[[SAVEEXEC_HI:[0-9]+]]], [[CMP0]]
@@ -175,7 +175,6 @@
; Spill saved exec
; VGPR: v_writelane_b32 [[SPILL_VGPR:v[0-9]+]], s[[SAVEEXEC_LO]], [[SAVEEXEC_LO_LANE:[0-9]+]]
; VGPR: v_writelane_b32 [[SPILL_VGPR]], s[[SAVEEXEC_HI]], [[SAVEEXEC_HI_LANE:[0-9]+]]
-; VGPR: buffer_store_dword [[SPILL_VGPR]], off, s[0:3], 0 offset:[[VREG_SAVE_RESTORE_OFFSET:[0-9]+]] ; 4-byte Folded Spill
; VMEM: v_writelane_b32 v[[V_SAVEEXEC:[0-9]+]], s[[SAVEEXEC_LO]], 0
; VMEM: v_writelane_b32 v[[V_SAVEEXEC]], s[[SAVEEXEC_HI]], 1
@@ -188,7 +187,6 @@
; GCN-NEXT: s_branch [[ELSE:.LBB[0-9]+_[0-9]+]]
; GCN: [[FLOW]]: ; %Flow
-; VGPR: buffer_load_dword [[SPILL_VGPR:v[0-9]+]], off, s[0:3], 0 offset:[[VREG_SAVE_RESTORE_OFFSET]] ; 4-byte Folded Reload
; GCN: buffer_load_dword [[FLOW_VAL:v[0-9]+]], off, s[0:3], 0 offset:[[FLOW_VAL_OFFSET:[0-9]+]] ; 4-byte Folded Reload
; VGPR: v_readlane_b32 s[[FLOW_S_RELOAD_SAVEEXEC_LO:[0-9]+]], [[SPILL_VGPR]], [[SAVEEXEC_LO_LANE]]
; VGPR: v_readlane_b32 s[[FLOW_S_RELOAD_SAVEEXEC_HI:[0-9]+]], [[SPILL_VGPR]], [[SAVEEXEC_HI_LANE]]
diff --git a/llvm/test/CodeGen/AMDGPU/cross-block-use-is-not-abi-copy.ll b/llvm/test/CodeGen/AMDGPU/cross-block-use-is-not-abi-copy.ll
--- a/llvm/test/CodeGen/AMDGPU/cross-block-use-is-not-abi-copy.ll
+++ b/llvm/test/CodeGen/AMDGPU/cross-block-use-is-not-abi-copy.ll
@@ -31,10 +31,11 @@
; GCN-NEXT: s_mov_b32 s33, s32
; GCN-NEXT: s_or_saveexec_b64 s[18:19], -1
; GCN-NEXT: buffer_store_dword v40, off, s[0:3], s33 ; 4-byte Folded Spill
+; GCN-NEXT: buffer_store_dword v41, off, s[0:3], s33 offset:4 ; 4-byte Folded Spill
; GCN-NEXT: s_mov_b64 exec, s[18:19]
-; GCN-NEXT: v_writelane_b32 v40, s16, 2
; GCN-NEXT: s_addk_i32 s32, 0x400
; GCN-NEXT: v_writelane_b32 v40, s30, 0
+; GCN-NEXT: v_writelane_b32 v41, s16, 0
; GCN-NEXT: v_writelane_b32 v40, s31, 1
; GCN-NEXT: s_getpc_b64 s[16:17]
; GCN-NEXT: s_add_u32 s16, s16, func_v2f32@rel32@lo+4
@@ -42,9 +43,10 @@
; GCN-NEXT: s_swappc_b64 s[30:31], s[16:17]
; GCN-NEXT: v_readlane_b32 s31, v40, 1
; GCN-NEXT: v_readlane_b32 s30, v40, 0
-; GCN-NEXT: v_readlane_b32 s4, v40, 2
+; GCN-NEXT: v_readlane_b32 s4, v41, 0
; GCN-NEXT: s_or_saveexec_b64 s[6:7], -1
; GCN-NEXT: buffer_load_dword v40, off, s[0:3], s33 ; 4-byte Folded Reload
+; GCN-NEXT: buffer_load_dword v41, off, s[0:3], s33 offset:4 ; 4-byte Folded Reload
; GCN-NEXT: s_mov_b64 exec, s[6:7]
; GCN-NEXT: s_addk_i32 s32, 0xfc00
; GCN-NEXT: s_mov_b32 s33, s4
@@ -67,10 +69,11 @@
; GCN-NEXT: s_mov_b32 s33, s32
; GCN-NEXT: s_or_saveexec_b64 s[18:19], -1
; GCN-NEXT: buffer_store_dword v40, off, s[0:3], s33 ; 4-byte Folded Spill
+; GCN-NEXT: buffer_store_dword v41, off, s[0:3], s33 offset:4 ; 4-byte Folded Spill
; GCN-NEXT: s_mov_b64 exec, s[18:19]
-; GCN-NEXT: v_writelane_b32 v40, s16, 2
; GCN-NEXT: s_addk_i32 s32, 0x400
; GCN-NEXT: v_writelane_b32 v40, s30, 0
+; GCN-NEXT: v_writelane_b32 v41, s16, 0
; GCN-NEXT: v_writelane_b32 v40, s31, 1
; GCN-NEXT: s_getpc_b64 s[16:17]
; GCN-NEXT: s_add_u32 s16, s16, func_v3f32@rel32@lo+4
@@ -78,9 +81,10 @@
; GCN-NEXT: s_swappc_b64 s[30:31], s[16:17]
; GCN-NEXT: v_readlane_b32 s31, v40, 1
; GCN-NEXT: v_readlane_b32 s30, v40, 0
-; GCN-NEXT: v_readlane_b32 s4, v40, 2
+; GCN-NEXT: v_readlane_b32 s4, v41, 0
; GCN-NEXT: s_or_saveexec_b64 s[6:7], -1
; GCN-NEXT: buffer_load_dword v40, off, s[0:3], s33 ; 4-byte Folded Reload
+; GCN-NEXT: buffer_load_dword v41, off, s[0:3], s33 offset:4 ; 4-byte Folded Reload
; GCN-NEXT: s_mov_b64 exec, s[6:7]
; GCN-NEXT: s_addk_i32 s32, 0xfc00
; GCN-NEXT: s_mov_b32 s33, s4
@@ -103,10 +107,11 @@
; GCN-NEXT: s_mov_b32 s33, s32
; GCN-NEXT: s_or_saveexec_b64 s[18:19], -1
; GCN-NEXT: buffer_store_dword v40, off, s[0:3], s33 ; 4-byte Folded Spill
+; GCN-NEXT: buffer_store_dword v41, off, s[0:3], s33 offset:4 ; 4-byte Folded Spill
; GCN-NEXT: s_mov_b64 exec, s[18:19]
-; GCN-NEXT: v_writelane_b32 v40, s16, 2
; GCN-NEXT: s_addk_i32 s32, 0x400
; GCN-NEXT: v_writelane_b32 v40, s30, 0
+; GCN-NEXT: v_writelane_b32 v41, s16, 0
; GCN-NEXT: v_writelane_b32 v40, s31, 1
; GCN-NEXT: s_getpc_b64 s[16:17]
; GCN-NEXT: s_add_u32 s16, s16, func_v4f16@rel32@lo+4
@@ -114,9 +119,10 @@
; GCN-NEXT: s_swappc_b64 s[30:31], s[16:17]
; GCN-NEXT: v_readlane_b32 s31, v40, 1
; GCN-NEXT: v_readlane_b32 s30, v40, 0
-; GCN-NEXT: v_readlane_b32 s4, v40, 2
+; GCN-NEXT: v_readlane_b32 s4, v41, 0
; GCN-NEXT: s_or_saveexec_b64 s[6:7], -1
; GCN-NEXT: buffer_load_dword v40, off, s[0:3], s33 ; 4-byte Folded Reload
+; GCN-NEXT: buffer_load_dword v41, off, s[0:3], s33 offset:4 ; 4-byte Folded Reload
; GCN-NEXT: s_mov_b64 exec, s[6:7]
; GCN-NEXT: s_addk_i32 s32, 0xfc00
; GCN-NEXT: s_mov_b32 s33, s4
@@ -139,10 +145,11 @@
; GCN-NEXT: s_mov_b32 s33, s32
; GCN-NEXT: s_or_saveexec_b64 s[18:19], -1
; GCN-NEXT: buffer_store_dword v40, off, s[0:3], s33 ; 4-byte Folded Spill
+; GCN-NEXT: buffer_store_dword v41, off, s[0:3], s33 offset:4 ; 4-byte Folded Spill
; GCN-NEXT: s_mov_b64 exec, s[18:19]
-; GCN-NEXT: v_writelane_b32 v40, s16, 2
; GCN-NEXT: s_addk_i32 s32, 0x400
; GCN-NEXT: v_writelane_b32 v40, s30, 0
+; GCN-NEXT: v_writelane_b32 v41, s16, 0
; GCN-NEXT: v_writelane_b32 v40, s31, 1
; GCN-NEXT: s_getpc_b64 s[16:17]
; GCN-NEXT: s_add_u32 s16, s16, func_struct@rel32@lo+4
@@ -151,9 +158,10 @@
; GCN-NEXT: v_mov_b32_e32 v1, v4
; GCN-NEXT: v_readlane_b32 s31, v40, 1
; GCN-NEXT: v_readlane_b32 s30, v40, 0
-; GCN-NEXT: v_readlane_b32 s4, v40, 2
+; GCN-NEXT: v_readlane_b32 s4, v41, 0
; GCN-NEXT: s_or_saveexec_b64 s[6:7], -1
; GCN-NEXT: buffer_load_dword v40, off, s[0:3], s33 ; 4-byte Folded Reload
+; GCN-NEXT: buffer_load_dword v41, off, s[0:3], s33 offset:4 ; 4-byte Folded Reload
; GCN-NEXT: s_mov_b64 exec, s[6:7]
; GCN-NEXT: s_addk_i32 s32, 0xfc00
; GCN-NEXT: s_mov_b32 s33, s4
diff --git a/llvm/test/CodeGen/AMDGPU/dwarf-multi-register-use-crash.ll b/llvm/test/CodeGen/AMDGPU/dwarf-multi-register-use-crash.ll
--- a/llvm/test/CodeGen/AMDGPU/dwarf-multi-register-use-crash.ll
+++ b/llvm/test/CodeGen/AMDGPU/dwarf-multi-register-use-crash.ll
@@ -20,8 +20,8 @@
; CHECK-NEXT: s_mov_b32 s33, s32
; CHECK-NEXT: s_or_saveexec_b64 s[18:19], -1
; CHECK-NEXT: buffer_store_dword v40, off, s[0:3], s33 offset:4 ; 4-byte Folded Spill
+; CHECK-NEXT: buffer_store_dword v42, off, s[0:3], s33 offset:8 ; 4-byte Folded Spill
; CHECK-NEXT: s_mov_b64 exec, s[18:19]
-; CHECK-NEXT: v_writelane_b32 v40, s16, 16
; CHECK-NEXT: v_writelane_b32 v40, s30, 0
; CHECK-NEXT: v_writelane_b32 v40, s31, 1
; CHECK-NEXT: v_writelane_b32 v40, s34, 2
@@ -48,6 +48,7 @@
; CHECK-NEXT: v_writelane_b32 v40, s47, 15
; CHECK-NEXT: s_load_dwordx2 s[46:47], s[4:5], 0x0
; CHECK-NEXT: s_mov_b64 s[4:5], s[40:41]
+; CHECK-NEXT: v_writelane_b32 v42, s16, 0
; CHECK-NEXT: buffer_store_dword v41, off, s[0:3], s33 ; 4-byte Folded Spill
; CHECK-NEXT: v_mov_b32_e32 v41, v31
; CHECK-NEXT: s_mov_b32 s42, s15
@@ -91,9 +92,10 @@
; CHECK-NEXT: v_readlane_b32 s34, v40, 2
; CHECK-NEXT: v_readlane_b32 s31, v40, 1
; CHECK-NEXT: v_readlane_b32 s30, v40, 0
-; CHECK-NEXT: v_readlane_b32 s4, v40, 16
+; CHECK-NEXT: v_readlane_b32 s4, v42, 0
; CHECK-NEXT: s_or_saveexec_b64 s[6:7], -1
; CHECK-NEXT: buffer_load_dword v40, off, s[0:3], s33 offset:4 ; 4-byte Folded Reload
+; CHECK-NEXT: buffer_load_dword v42, off, s[0:3], s33 offset:8 ; 4-byte Folded Reload
; CHECK-NEXT: s_mov_b64 exec, s[6:7]
; CHECK-NEXT: s_addk_i32 s32, 0xfc00
; CHECK-NEXT: s_mov_b32 s33, s4
diff --git a/llvm/test/CodeGen/AMDGPU/extend-wwm-virt-reg-liveness.mir b/llvm/test/CodeGen/AMDGPU/extend-wwm-virt-reg-liveness.mir
deleted file mode 100644
--- a/llvm/test/CodeGen/AMDGPU/extend-wwm-virt-reg-liveness.mir
+++ /dev/null
@@ -1,279 +0,0 @@
-# NOTE: Assertions have been autogenerated by utils/update_mir_test_checks.py
-# RUN: llc -mtriple=amdgcn-amd-amdhsa -mcpu=gfx906 -start-before=si-lower-sgpr-spills -stop-after=virtregrewriter,1 -verify-machineinstrs %s -o - | FileCheck -check-prefix=GCN %s
-
-# Tests to check the conservative lieness extension for the wwm registers during SGPR spill lowering.
-
-# Even though the VGPR can be shared for the wwm-operand (writelane/readlane get inserted for the SGPR spills)
-# and the regular operand (%0), they get different registers as we conservatively extend the liveness of the
-# wwm-operands.
----
-name: test_single_block
-tracksRegLiveness: true
-frameInfo:
- maxAlignment: 4
-stack:
- - { id: 0, type: spill-slot, size: 4, alignment: 4, stack-id: sgpr-spill }
-machineFunctionInfo:
- isEntryFunction: false
- scratchRSrcReg: '$sgpr0_sgpr1_sgpr2_sgpr3'
- stackPtrOffsetReg: '$sgpr32'
- frameOffsetReg: '$sgpr33'
- hasSpilledSGPRs: true
-body: |
- bb.0:
- liveins: $sgpr4, $vgpr2_vgpr3
- ; GCN-LABEL: name: test_single_block
- ; GCN: liveins: $sgpr4, $vgpr2_vgpr3
- ; GCN-NEXT: {{ $}}
- ; GCN-NEXT: renamable $vgpr0 = IMPLICIT_DEF
- ; GCN-NEXT: renamable $vgpr0 = V_WRITELANE_B32 $sgpr4, 0, killed $vgpr0
- ; GCN-NEXT: S_NOP 0
- ; GCN-NEXT: $sgpr4 = V_READLANE_B32 $vgpr0, 0
- ; GCN-NEXT: renamable $vgpr1 = V_MOV_B32_e32 20, implicit $exec
- ; GCN-NEXT: GLOBAL_STORE_DWORD $vgpr2_vgpr3, killed renamable $vgpr1, 0, 0, implicit $exec
- ; GCN-NEXT: KILL killed renamable $vgpr0
- ; GCN-NEXT: SI_RETURN
- SI_SPILL_S32_SAVE killed $sgpr4, %stack.0, implicit $exec, implicit $sgpr0_sgpr1_sgpr2_sgpr3, implicit $sgpr32
- S_NOP 0
- renamable $sgpr4 = SI_SPILL_S32_RESTORE %stack.0, implicit $exec, implicit $sgpr0_sgpr1_sgpr2_sgpr3, implicit $sgpr32
- %0:vgpr_32 = V_MOV_B32_e32 20, implicit $exec
- GLOBAL_STORE_DWORD $vgpr2_vgpr3, %0:vgpr_32, 0, 0, implicit $exec
- SI_RETURN
-...
-
-# Due to the presence of wwm-operand in the divergent flow, the regular variable (%0) shouldn't get the same register
-# allocated for the wwm-operand in writelane/readlane when the SGPR spill is lowered.
-
----
-name: test_if_else
-tracksRegLiveness: true
-frameInfo:
- maxAlignment: 4
-stack:
- - { id: 0, type: spill-slot, size: 4, alignment: 4, stack-id: sgpr-spill }
-machineFunctionInfo:
- isEntryFunction: false
- scratchRSrcReg: '$sgpr0_sgpr1_sgpr2_sgpr3'
- stackPtrOffsetReg: '$sgpr32'
- frameOffsetReg: '$sgpr33'
- hasSpilledSGPRs: true
-body: |
- ; GCN-LABEL: name: test_if_else
- ; GCN: bb.0:
- ; GCN-NEXT: successors: %bb.1(0x80000000)
- ; GCN-NEXT: liveins: $sgpr6, $sgpr10_sgpr11
- ; GCN-NEXT: {{ $}}
- ; GCN-NEXT: renamable $vgpr0 = IMPLICIT_DEF
- ; GCN-NEXT: S_BRANCH %bb.1
- ; GCN-NEXT: {{ $}}
- ; GCN-NEXT: bb.1:
- ; GCN-NEXT: successors: %bb.3(0x40000000), %bb.2(0x40000000)
- ; GCN-NEXT: liveins: $sgpr6, $vgpr0, $sgpr10_sgpr11
- ; GCN-NEXT: {{ $}}
- ; GCN-NEXT: renamable $vgpr1 = V_MOV_B32_e32 10, implicit $exec
- ; GCN-NEXT: S_CBRANCH_EXECZ %bb.3, implicit $exec
- ; GCN-NEXT: {{ $}}
- ; GCN-NEXT: bb.2:
- ; GCN-NEXT: successors: %bb.3(0x80000000)
- ; GCN-NEXT: liveins: $sgpr6, $vgpr0, $sgpr10_sgpr11
- ; GCN-NEXT: {{ $}}
- ; GCN-NEXT: renamable $vgpr0 = V_WRITELANE_B32 $sgpr6, 0, killed $vgpr0
- ; GCN-NEXT: S_NOP 0
- ; GCN-NEXT: $sgpr6 = V_READLANE_B32 $vgpr0, 0
- ; GCN-NEXT: renamable $vgpr1 = V_MOV_B32_e32 20, implicit $exec
- ; GCN-NEXT: S_BRANCH %bb.3
- ; GCN-NEXT: {{ $}}
- ; GCN-NEXT: bb.3:
- ; GCN-NEXT: liveins: $vgpr0, $vgpr1, $sgpr10_sgpr11
- ; GCN-NEXT: {{ $}}
- ; GCN-NEXT: $sgpr5 = V_READFIRSTLANE_B32 killed $vgpr1, implicit $exec
- ; GCN-NEXT: S_STORE_DWORD_IMM $sgpr5, $sgpr10_sgpr11, 0, 0
- ; GCN-NEXT: KILL killed renamable $vgpr0
- ; GCN-NEXT: SI_RETURN
- bb.0:
- liveins: $sgpr6, $sgpr10_sgpr11
- S_BRANCH %bb.1
- bb.1:
- liveins: $sgpr6, $sgpr10_sgpr11
- %0:vgpr_32 = V_MOV_B32_e32 10, implicit $exec
- S_CBRANCH_EXECZ %bb.3, implicit $exec
- bb.2:
- liveins: $sgpr6, $sgpr10_sgpr11
- SI_SPILL_S32_SAVE killed $sgpr6, %stack.0, implicit $exec, implicit $sgpr0_sgpr1_sgpr2_sgpr3, implicit $sgpr32
- S_NOP 0
- renamable $sgpr6 = SI_SPILL_S32_RESTORE %stack.0, implicit $exec, implicit $sgpr0_sgpr1_sgpr2_sgpr3, implicit $sgpr32
- %0:vgpr_32 = V_MOV_B32_e32 20, implicit $exec
- S_BRANCH %bb.3
- bb.3:
- liveins: $sgpr10_sgpr11
- $sgpr5 = V_READFIRSTLANE_B32 %0:vgpr_32, implicit $exec
- S_STORE_DWORD_IMM $sgpr5, $sgpr10_sgpr11, 0, 0
- SI_RETURN
-...
-
-# The wwm-register usage outside the loop should have the interference marked with
-# all the regular virtual registers used in the test. The divergent loop index value (%1)
-# can actually share the same VGPR as the wwm-operand. But since we extend the liveness of
-# the wwm operand, an interference will always exist between them.
-
----
-name: test_loop
-tracksRegLiveness: true
-frameInfo:
- maxAlignment: 4
-stack:
- - { id: 0, type: spill-slot, size: 4, alignment: 4, stack-id: sgpr-spill }
-machineFunctionInfo:
- isEntryFunction: false
- scratchRSrcReg: '$sgpr0_sgpr1_sgpr2_sgpr3'
- stackPtrOffsetReg: '$sgpr32'
- frameOffsetReg: '$sgpr33'
- hasSpilledSGPRs: true
-body: |
- ; GCN-LABEL: name: test_loop
- ; GCN: bb.0:
- ; GCN-NEXT: successors: %bb.2(0x40000000), %bb.1(0x40000000)
- ; GCN-NEXT: liveins: $sgpr4, $sgpr10_sgpr11
- ; GCN-NEXT: {{ $}}
- ; GCN-NEXT: renamable $vgpr0 = IMPLICIT_DEF
- ; GCN-NEXT: renamable $vgpr1 = V_MOV_B32_e32 10, implicit $exec
- ; GCN-NEXT: S_CBRANCH_EXECZ %bb.2, implicit $exec
- ; GCN-NEXT: {{ $}}
- ; GCN-NEXT: bb.1:
- ; GCN-NEXT: successors: %bb.2(0x80000000)
- ; GCN-NEXT: liveins: $sgpr4, $vgpr0, $sgpr10_sgpr11
- ; GCN-NEXT: {{ $}}
- ; GCN-NEXT: renamable $vgpr0 = V_WRITELANE_B32 $sgpr4, 0, killed $vgpr0
- ; GCN-NEXT: S_NOP 0
- ; GCN-NEXT: $sgpr4 = V_READLANE_B32 $vgpr0, 0
- ; GCN-NEXT: renamable $vgpr1 = V_MOV_B32_e32 20, implicit $exec
- ; GCN-NEXT: S_BRANCH %bb.2
- ; GCN-NEXT: {{ $}}
- ; GCN-NEXT: bb.2:
- ; GCN-NEXT: successors: %bb.3(0x80000000)
- ; GCN-NEXT: liveins: $sgpr4, $vgpr0, $vgpr1, $sgpr10_sgpr11
- ; GCN-NEXT: {{ $}}
- ; GCN-NEXT: S_STORE_DWORD_IMM $sgpr4, $sgpr10_sgpr11, 0, 0
- ; GCN-NEXT: $sgpr5 = V_READFIRSTLANE_B32 killed $vgpr1, implicit $exec
- ; GCN-NEXT: S_STORE_DWORD_IMM $sgpr5, $sgpr10_sgpr11, 0, 4
- ; GCN-NEXT: renamable $vgpr1 = V_MOV_B32_e32 5, implicit $exec
- ; GCN-NEXT: S_CBRANCH_EXECZ %bb.3, implicit $exec
- ; GCN-NEXT: S_BRANCH %bb.3
- ; GCN-NEXT: {{ $}}
- ; GCN-NEXT: bb.3:
- ; GCN-NEXT: successors: %bb.5(0x40000000), %bb.4(0x40000000)
- ; GCN-NEXT: liveins: $vgpr0, $vgpr1
- ; GCN-NEXT: {{ $}}
- ; GCN-NEXT: $vcc = V_CMP_EQ_U32_e64 0, $vgpr1, implicit $exec
- ; GCN-NEXT: $sgpr6_sgpr7 = S_AND_SAVEEXEC_B64 $vcc, implicit-def $exec, implicit-def $scc, implicit $exec
- ; GCN-NEXT: S_CBRANCH_SCC1 %bb.5, implicit $scc
- ; GCN-NEXT: {{ $}}
- ; GCN-NEXT: bb.4:
- ; GCN-NEXT: successors: %bb.3(0x80000000)
- ; GCN-NEXT: liveins: $vgpr0, $vgpr1, $sgpr6_sgpr7
- ; GCN-NEXT: {{ $}}
- ; GCN-NEXT: renamable $vgpr1 = V_SUB_U32_e32 1, killed $vgpr1, implicit $exec
- ; GCN-NEXT: renamable $vgpr1 = V_MOV_B32_e32 killed $vgpr1, implicit $exec
- ; GCN-NEXT: S_BRANCH %bb.3
- ; GCN-NEXT: {{ $}}
- ; GCN-NEXT: bb.5:
- ; GCN-NEXT: liveins: $vgpr0, $sgpr6_sgpr7
- ; GCN-NEXT: {{ $}}
- ; GCN-NEXT: $exec = S_OR_B64 $exec, $sgpr6_sgpr7, implicit-def $scc
- ; GCN-NEXT: KILL killed renamable $vgpr0
- ; GCN-NEXT: SI_RETURN
- bb.0:
- liveins: $sgpr4, $sgpr10_sgpr11
- %0:vgpr_32 = V_MOV_B32_e32 10, implicit $exec
- S_CBRANCH_EXECZ %bb.2, implicit $exec
- bb.1:
- liveins: $sgpr4, $sgpr10_sgpr11
- SI_SPILL_S32_SAVE killed $sgpr4, %stack.0, implicit $exec, implicit $sgpr0_sgpr1_sgpr2_sgpr3, implicit $sgpr32
- S_NOP 0
- renamable $sgpr4 = SI_SPILL_S32_RESTORE %stack.0, implicit $exec, implicit $sgpr0_sgpr1_sgpr2_sgpr3, implicit $sgpr32
- %0:vgpr_32 = V_MOV_B32_e32 20, implicit $exec
- S_BRANCH %bb.2
- bb.2:
- liveins: $sgpr4, $sgpr10_sgpr11
- S_STORE_DWORD_IMM $sgpr4, $sgpr10_sgpr11, 0, 0
- $sgpr5 = V_READFIRSTLANE_B32 %0:vgpr_32, implicit $exec
- S_STORE_DWORD_IMM $sgpr5, $sgpr10_sgpr11, 0, 4
- %1:vgpr_32 = V_MOV_B32_e32 5, implicit $exec
- S_CBRANCH_EXECZ %bb.3, implicit $exec
- S_BRANCH %bb.3
- bb.3:
- $vcc = V_CMP_EQ_U32_e64 0, %1:vgpr_32, implicit $exec
- $sgpr6_sgpr7 = S_AND_SAVEEXEC_B64 $vcc, implicit-def $exec, implicit-def $scc, implicit $exec
- S_CBRANCH_SCC1 %bb.5, implicit $scc
- bb.4:
- liveins: $sgpr6_sgpr7
- %2:vgpr_32 = V_SUB_U32_e32 1, %1:vgpr_32, implicit $exec
- %1:vgpr_32 = V_MOV_B32_e32 %2:vgpr_32, implicit $exec
- S_BRANCH %bb.3
- bb.5:
- liveins: $sgpr6_sgpr7
- $exec = S_OR_B64 $exec, $sgpr6_sgpr7, implicit-def $scc
- SI_RETURN
-...
-
-# There must be one KILL instruction for the wwm-operand in every return block.
-# Due to that, the wwm-register allocated should be different from the ones
-# allocated for the regular virtual registers.
-
----
-name: test_multiple_return_blocks
-tracksRegLiveness: true
-frameInfo:
- maxAlignment: 4
-stack:
- - { id: 0, type: spill-slot, size: 4, alignment: 4, stack-id: sgpr-spill }
-machineFunctionInfo:
- isEntryFunction: false
- scratchRSrcReg: '$sgpr0_sgpr1_sgpr2_sgpr3'
- stackPtrOffsetReg: '$sgpr32'
- frameOffsetReg: '$sgpr33'
- hasSpilledSGPRs: true
-body: |
- ; GCN-LABEL: name: test_multiple_return_blocks
- ; GCN: bb.0:
- ; GCN-NEXT: successors: %bb.2(0x40000000), %bb.1(0x40000000)
- ; GCN-NEXT: liveins: $sgpr4, $vgpr2_vgpr3
- ; GCN-NEXT: {{ $}}
- ; GCN-NEXT: renamable $vgpr0 = IMPLICIT_DEF
- ; GCN-NEXT: S_CBRANCH_EXECZ %bb.2, implicit $exec
- ; GCN-NEXT: {{ $}}
- ; GCN-NEXT: bb.1:
- ; GCN-NEXT: liveins: $sgpr4, $vgpr0, $vgpr2_vgpr3
- ; GCN-NEXT: {{ $}}
- ; GCN-NEXT: renamable $vgpr0 = V_WRITELANE_B32 $sgpr4, 0, killed $vgpr0
- ; GCN-NEXT: S_NOP 0
- ; GCN-NEXT: $sgpr4 = V_READLANE_B32 $vgpr0, 0
- ; GCN-NEXT: renamable $vgpr1 = V_MOV_B32_e32 10, implicit $exec
- ; GCN-NEXT: GLOBAL_STORE_DWORD $vgpr2_vgpr3, killed renamable $vgpr1, 0, 0, implicit $exec
- ; GCN-NEXT: KILL killed renamable $vgpr0
- ; GCN-NEXT: SI_RETURN
- ; GCN-NEXT: {{ $}}
- ; GCN-NEXT: bb.2:
- ; GCN-NEXT: liveins: $vgpr0, $vgpr2_vgpr3
- ; GCN-NEXT: {{ $}}
- ; GCN-NEXT: renamable $vgpr1 = V_MOV_B32_e32 20, implicit $exec
- ; GCN-NEXT: GLOBAL_STORE_DWORD $vgpr2_vgpr3, killed renamable $vgpr1, 0, 0, implicit $exec
- ; GCN-NEXT: KILL killed renamable $vgpr0
- ; GCN-NEXT: SI_RETURN
- bb.0:
- liveins: $sgpr4, $vgpr2_vgpr3
- S_CBRANCH_EXECZ %bb.2, implicit $exec
- bb.1:
- liveins: $sgpr4, $vgpr2_vgpr3
- SI_SPILL_S32_SAVE killed $sgpr4, %stack.0, implicit $exec, implicit $sgpr0_sgpr1_sgpr2_sgpr3, implicit $sgpr32
- S_NOP 0
- renamable $sgpr4 = SI_SPILL_S32_RESTORE %stack.0, implicit $exec, implicit $sgpr0_sgpr1_sgpr2_sgpr3, implicit $sgpr32
- %0:vgpr_32 = V_MOV_B32_e32 10, implicit $exec
- GLOBAL_STORE_DWORD $vgpr2_vgpr3, %0:vgpr_32, 0, 0, implicit $exec
- SI_RETURN
- bb.2:
- liveins: $vgpr2_vgpr3
- %1:vgpr_32 = V_MOV_B32_e32 20, implicit $exec
- GLOBAL_STORE_DWORD $vgpr2_vgpr3, %1:vgpr_32, 0, 0, implicit $exec
- SI_RETURN
-...
diff --git a/llvm/test/CodeGen/AMDGPU/fix-frame-reg-in-custom-csr-spills.ll b/llvm/test/CodeGen/AMDGPU/fix-frame-reg-in-custom-csr-spills.ll
--- a/llvm/test/CodeGen/AMDGPU/fix-frame-reg-in-custom-csr-spills.ll
+++ b/llvm/test/CodeGen/AMDGPU/fix-frame-reg-in-custom-csr-spills.ll
@@ -16,9 +16,10 @@
; GCN-NEXT: s_mov_b32 s33, s32
; GCN-NEXT: s_or_saveexec_b64 s[18:19], -1
; GCN-NEXT: buffer_store_dword v42, off, s[0:3], s33 offset:8 ; 4-byte Folded Spill
+; GCN-NEXT: buffer_store_dword v43, off, s[0:3], s33 offset:12 ; 4-byte Folded Spill
; GCN-NEXT: s_mov_b64 exec, s[18:19]
-; GCN-NEXT: s_addk_i32 s32, 0x400
-; GCN-NEXT: v_writelane_b32 v42, s16, 2
+; GCN-NEXT: s_addk_i32 s32, 0x800
+; GCN-NEXT: v_writelane_b32 v43, s16, 0
; GCN-NEXT: s_getpc_b64 s[16:17]
; GCN-NEXT: s_add_u32 s16, s16, extern_func@gotpcrel32@lo+4
; GCN-NEXT: s_addc_u32 s17, s17, extern_func@gotpcrel32@hi+12
@@ -38,11 +39,12 @@
; GCN-NEXT: buffer_load_dword v40, off, s[0:3], s33 offset:4 ; 4-byte Folded Reload
; GCN-NEXT: v_readlane_b32 s31, v42, 1
; GCN-NEXT: v_readlane_b32 s30, v42, 0
-; GCN-NEXT: v_readlane_b32 s4, v42, 2
+; GCN-NEXT: v_readlane_b32 s4, v43, 0
; GCN-NEXT: s_or_saveexec_b64 s[6:7], -1
; GCN-NEXT: buffer_load_dword v42, off, s[0:3], s33 offset:8 ; 4-byte Folded Reload
+; GCN-NEXT: buffer_load_dword v43, off, s[0:3], s33 offset:12 ; 4-byte Folded Reload
; GCN-NEXT: s_mov_b64 exec, s[6:7]
-; GCN-NEXT: s_addk_i32 s32, 0xfc00
+; GCN-NEXT: s_addk_i32 s32, 0xf800
; GCN-NEXT: s_mov_b32 s33, s4
; GCN-NEXT: s_waitcnt vmcnt(0)
; GCN-NEXT: s_setpc_b64 s[30:31]
diff --git a/llvm/test/CodeGen/AMDGPU/flat-scratch-init.ll b/llvm/test/CodeGen/AMDGPU/flat-scratch-init.ll
--- a/llvm/test/CodeGen/AMDGPU/flat-scratch-init.ll
+++ b/llvm/test/CodeGen/AMDGPU/flat-scratch-init.ll
@@ -117,15 +117,21 @@
; FLAT_SCR_OPT-NEXT: s_setreg_b32 hwreg(HW_REG_FLAT_SCR_LO), s2
; FLAT_SCR_OPT-NEXT: s_setreg_b32 hwreg(HW_REG_FLAT_SCR_HI), s3
; FLAT_SCR_OPT-NEXT: s_load_dwordx2 s[2:3], s[0:1], 0x0
-; FLAT_SCR_OPT-NEXT: ; implicit-def: $vgpr0
+; FLAT_SCR_OPT-NEXT: s_mov_b32 s4, exec_lo
+; FLAT_SCR_OPT-NEXT: s_mov_b32 exec_lo, 3
+; FLAT_SCR_OPT-NEXT: s_mov_b32 s5, 0
+; FLAT_SCR_OPT-NEXT: scratch_store_dword off, v0, s5
; FLAT_SCR_OPT-NEXT: s_waitcnt lgkmcnt(0)
; FLAT_SCR_OPT-NEXT: v_writelane_b32 v0, s2, 0
+; FLAT_SCR_OPT-NEXT: s_mov_b32 s5, 4
; FLAT_SCR_OPT-NEXT: v_writelane_b32 v0, s3, 1
-; FLAT_SCR_OPT-NEXT: s_or_saveexec_b32 s105, -1
-; FLAT_SCR_OPT-NEXT: s_mov_b32 s2, 4
-; FLAT_SCR_OPT-NEXT: scratch_store_dword off, v0, s2 ; 4-byte Folded Spill
+; FLAT_SCR_OPT-NEXT: scratch_store_dword off, v0, s5 ; 4-byte Folded Spill
; FLAT_SCR_OPT-NEXT: s_waitcnt_depctr 0xffe3
-; FLAT_SCR_OPT-NEXT: s_mov_b32 exec_lo, s105
+; FLAT_SCR_OPT-NEXT: s_mov_b32 s5, 0
+; FLAT_SCR_OPT-NEXT: scratch_load_dword v0, off, s5
+; FLAT_SCR_OPT-NEXT: s_waitcnt vmcnt(0)
+; FLAT_SCR_OPT-NEXT: s_waitcnt_depctr 0xffe3
+; FLAT_SCR_OPT-NEXT: s_mov_b32 exec_lo, s4
; FLAT_SCR_OPT-NEXT: s_load_dword vcc_lo, s[0:1], 0x8
; FLAT_SCR_OPT-NEXT: ; kill: killed $sgpr0_sgpr1
; FLAT_SCR_OPT-NEXT: ;;#ASMSTART
@@ -222,31 +228,44 @@
; FLAT_SCR_OPT-NEXT: ;;#ASMEND
; FLAT_SCR_OPT-NEXT: ;;#ASMSTART
; FLAT_SCR_OPT-NEXT: ;;#ASMEND
-; FLAT_SCR_OPT-NEXT: s_or_saveexec_b32 s105, -1
-; FLAT_SCR_OPT-NEXT: s_mov_b32 s0, 4
-; FLAT_SCR_OPT-NEXT: scratch_load_dword v1, off, s0 ; 4-byte Folded Reload
+; FLAT_SCR_OPT-NEXT: s_mov_b32 s2, exec_lo
+; FLAT_SCR_OPT-NEXT: s_mov_b32 exec_lo, 3
+; FLAT_SCR_OPT-NEXT: s_mov_b32 s3, 0
+; FLAT_SCR_OPT-NEXT: scratch_store_dword off, v1, s3
+; FLAT_SCR_OPT-NEXT: s_waitcnt_depctr 0xffe3
+; FLAT_SCR_OPT-NEXT: s_mov_b32 s3, 4
+; FLAT_SCR_OPT-NEXT: scratch_load_dword v1, off, s3 ; 4-byte Folded Reload
; FLAT_SCR_OPT-NEXT: s_waitcnt_depctr 0xffe3
-; FLAT_SCR_OPT-NEXT: s_mov_b32 exec_lo, s105
+; FLAT_SCR_OPT-NEXT: s_mov_b32 s3, 0
; FLAT_SCR_OPT-NEXT: s_waitcnt vmcnt(0)
; FLAT_SCR_OPT-NEXT: v_readlane_b32 s0, v1, 0
; FLAT_SCR_OPT-NEXT: v_readlane_b32 s1, v1, 1
-; FLAT_SCR_OPT-NEXT: v_mov_b32_e32 v2, 0
-; FLAT_SCR_OPT-NEXT: ; kill: killed $vgpr1
-; FLAT_SCR_OPT-NEXT: global_store_dword v2, v0, s[0:1]
+; FLAT_SCR_OPT-NEXT: scratch_load_dword v1, off, s3
+; FLAT_SCR_OPT-NEXT: s_waitcnt vmcnt(0)
+; FLAT_SCR_OPT-NEXT: s_waitcnt_depctr 0xffe3
+; FLAT_SCR_OPT-NEXT: s_mov_b32 exec_lo, s2
+; FLAT_SCR_OPT-NEXT: v_mov_b32_e32 v1, 0
+; FLAT_SCR_OPT-NEXT: global_store_dword v1, v0, s[0:1]
; FLAT_SCR_OPT-NEXT: s_endpgm
;
; FLAT_SCR_ARCH-LABEL: test:
; FLAT_SCR_ARCH: ; %bb.0:
; FLAT_SCR_ARCH-NEXT: s_load_dwordx2 s[2:3], s[0:1], 0x0
-; FLAT_SCR_ARCH-NEXT: ; implicit-def: $vgpr0
+; FLAT_SCR_ARCH-NEXT: s_mov_b32 s4, exec_lo
+; FLAT_SCR_ARCH-NEXT: s_mov_b32 exec_lo, 3
+; FLAT_SCR_ARCH-NEXT: s_mov_b32 s5, 0
+; FLAT_SCR_ARCH-NEXT: scratch_store_dword off, v0, s5
; FLAT_SCR_ARCH-NEXT: s_waitcnt lgkmcnt(0)
; FLAT_SCR_ARCH-NEXT: v_writelane_b32 v0, s2, 0
+; FLAT_SCR_ARCH-NEXT: s_mov_b32 s5, 4
; FLAT_SCR_ARCH-NEXT: v_writelane_b32 v0, s3, 1
-; FLAT_SCR_ARCH-NEXT: s_or_saveexec_b32 s105, -1
-; FLAT_SCR_ARCH-NEXT: s_mov_b32 s2, 4
-; FLAT_SCR_ARCH-NEXT: scratch_store_dword off, v0, s2 ; 4-byte Folded Spill
+; FLAT_SCR_ARCH-NEXT: scratch_store_dword off, v0, s5 ; 4-byte Folded Spill
; FLAT_SCR_ARCH-NEXT: s_waitcnt_depctr 0xffe3
-; FLAT_SCR_ARCH-NEXT: s_mov_b32 exec_lo, s105
+; FLAT_SCR_ARCH-NEXT: s_mov_b32 s5, 0
+; FLAT_SCR_ARCH-NEXT: scratch_load_dword v0, off, s5
+; FLAT_SCR_ARCH-NEXT: s_waitcnt vmcnt(0)
+; FLAT_SCR_ARCH-NEXT: s_waitcnt_depctr 0xffe3
+; FLAT_SCR_ARCH-NEXT: s_mov_b32 exec_lo, s4
; FLAT_SCR_ARCH-NEXT: s_load_dword vcc_lo, s[0:1], 0x8
; FLAT_SCR_ARCH-NEXT: ; kill: killed $sgpr0_sgpr1
; FLAT_SCR_ARCH-NEXT: ;;#ASMSTART
@@ -343,17 +362,24 @@
; FLAT_SCR_ARCH-NEXT: ;;#ASMEND
; FLAT_SCR_ARCH-NEXT: ;;#ASMSTART
; FLAT_SCR_ARCH-NEXT: ;;#ASMEND
-; FLAT_SCR_ARCH-NEXT: s_or_saveexec_b32 s105, -1
-; FLAT_SCR_ARCH-NEXT: s_mov_b32 s0, 4
-; FLAT_SCR_ARCH-NEXT: scratch_load_dword v1, off, s0 ; 4-byte Folded Reload
+; FLAT_SCR_ARCH-NEXT: s_mov_b32 s2, exec_lo
+; FLAT_SCR_ARCH-NEXT: s_mov_b32 exec_lo, 3
+; FLAT_SCR_ARCH-NEXT: s_mov_b32 s3, 0
+; FLAT_SCR_ARCH-NEXT: scratch_store_dword off, v1, s3
+; FLAT_SCR_ARCH-NEXT: s_waitcnt_depctr 0xffe3
+; FLAT_SCR_ARCH-NEXT: s_mov_b32 s3, 4
+; FLAT_SCR_ARCH-NEXT: scratch_load_dword v1, off, s3 ; 4-byte Folded Reload
; FLAT_SCR_ARCH-NEXT: s_waitcnt_depctr 0xffe3
-; FLAT_SCR_ARCH-NEXT: s_mov_b32 exec_lo, s105
+; FLAT_SCR_ARCH-NEXT: s_mov_b32 s3, 0
; FLAT_SCR_ARCH-NEXT: s_waitcnt vmcnt(0)
; FLAT_SCR_ARCH-NEXT: v_readlane_b32 s0, v1, 0
; FLAT_SCR_ARCH-NEXT: v_readlane_b32 s1, v1, 1
-; FLAT_SCR_ARCH-NEXT: v_mov_b32_e32 v2, 0
-; FLAT_SCR_ARCH-NEXT: ; kill: killed $vgpr1
-; FLAT_SCR_ARCH-NEXT: global_store_dword v2, v0, s[0:1]
+; FLAT_SCR_ARCH-NEXT: scratch_load_dword v1, off, s3
+; FLAT_SCR_ARCH-NEXT: s_waitcnt vmcnt(0)
+; FLAT_SCR_ARCH-NEXT: s_waitcnt_depctr 0xffe3
+; FLAT_SCR_ARCH-NEXT: s_mov_b32 exec_lo, s2
+; FLAT_SCR_ARCH-NEXT: v_mov_b32_e32 v1, 0
+; FLAT_SCR_ARCH-NEXT: global_store_dword v1, v0, s[0:1]
; FLAT_SCR_ARCH-NEXT: s_endpgm
call void asm sideeffect "", "~{s[0:7]}" ()
call void asm sideeffect "", "~{s[8:15]}" ()
diff --git a/llvm/test/CodeGen/AMDGPU/fold-reload-into-exec.mir b/llvm/test/CodeGen/AMDGPU/fold-reload-into-exec.mir
--- a/llvm/test/CodeGen/AMDGPU/fold-reload-into-exec.mir
+++ b/llvm/test/CodeGen/AMDGPU/fold-reload-into-exec.mir
@@ -12,13 +12,14 @@
body: |
bb.0:
; CHECK-LABEL: name: merge_sgpr_spill_into_copy_from_exec_lo
- ; CHECK: renamable $vgpr0 = IMPLICIT_DEF
+ ; CHECK: liveins: $vgpr0
+ ; CHECK-NEXT: {{ $}}
; CHECK-NEXT: S_NOP 0, implicit-def $exec_lo
; CHECK-NEXT: $sgpr0 = S_MOV_B32 $exec_lo
- ; CHECK-NEXT: renamable $vgpr0 = V_WRITELANE_B32 killed $sgpr0, 0, killed $vgpr0
+ ; CHECK-NEXT: $vgpr0 = V_WRITELANE_B32 killed $sgpr0, 0, $vgpr0
; CHECK-NEXT: $sgpr0 = V_READLANE_B32 $vgpr0, 0
; CHECK-NEXT: S_NOP 0, implicit-def dead renamable $sgpr1, implicit-def dead renamable $sgpr0, implicit killed renamable $sgpr0
- ; CHECK-NEXT: $sgpr0 = V_READLANE_B32 killed $vgpr0, 0
+ ; CHECK-NEXT: $sgpr0 = V_READLANE_B32 $vgpr0, 0
; CHECK-NEXT: $exec_lo = S_MOV_B32 killed $sgpr0
; CHECK-NEXT: S_SENDMSG 0, implicit $m0, implicit $exec
S_NOP 0, implicit-def $exec_lo
@@ -37,13 +38,14 @@
body: |
bb.0:
; CHECK-LABEL: name: merge_sgpr_spill_into_copy_from_exec_hi
- ; CHECK: renamable $vgpr0 = IMPLICIT_DEF
+ ; CHECK: liveins: $vgpr0
+ ; CHECK-NEXT: {{ $}}
; CHECK-NEXT: S_NOP 0, implicit-def $exec_hi
; CHECK-NEXT: $sgpr0 = S_MOV_B32 $exec_hi
- ; CHECK-NEXT: renamable $vgpr0 = V_WRITELANE_B32 killed $sgpr0, 0, killed $vgpr0
+ ; CHECK-NEXT: $vgpr0 = V_WRITELANE_B32 killed $sgpr0, 0, $vgpr0
; CHECK-NEXT: $sgpr0 = V_READLANE_B32 $vgpr0, 0
; CHECK-NEXT: S_NOP 0, implicit-def dead renamable $sgpr1, implicit-def dead renamable $sgpr0, implicit killed renamable $sgpr0
- ; CHECK-NEXT: $sgpr0 = V_READLANE_B32 killed $vgpr0, 0
+ ; CHECK-NEXT: $sgpr0 = V_READLANE_B32 $vgpr0, 0
; CHECK-NEXT: $exec_hi = S_MOV_B32 killed $sgpr0
; CHECK-NEXT: S_SENDMSG 0, implicit $m0, implicit $exec
S_NOP 0, implicit-def $exec_hi
@@ -62,16 +64,17 @@
body: |
bb.0:
; CHECK-LABEL: name: merge_sgpr_spill_into_copy_from_exec
- ; CHECK: renamable $vgpr0 = IMPLICIT_DEF
+ ; CHECK: liveins: $vgpr0
+ ; CHECK-NEXT: {{ $}}
; CHECK-NEXT: S_NOP 0, implicit-def $exec
; CHECK-NEXT: $sgpr0_sgpr1 = S_MOV_B64 $exec
- ; CHECK-NEXT: renamable $vgpr0 = V_WRITELANE_B32 killed $sgpr0, 0, killed $vgpr0, implicit-def $sgpr0_sgpr1, implicit $sgpr0_sgpr1
- ; CHECK-NEXT: renamable $vgpr0 = V_WRITELANE_B32 killed $sgpr1, 1, killed $vgpr0, implicit $sgpr0_sgpr1
+ ; CHECK-NEXT: $vgpr0 = V_WRITELANE_B32 killed $sgpr0, 0, $vgpr0, implicit-def $sgpr0_sgpr1, implicit $sgpr0_sgpr1
+ ; CHECK-NEXT: $vgpr0 = V_WRITELANE_B32 killed $sgpr1, 1, $vgpr0, implicit $sgpr0_sgpr1
; CHECK-NEXT: $sgpr0 = V_READLANE_B32 $vgpr0, 0, implicit-def $sgpr0_sgpr1
; CHECK-NEXT: $sgpr1 = V_READLANE_B32 $vgpr0, 1
; CHECK-NEXT: S_NOP 0, implicit-def dead renamable $sgpr2_sgpr3, implicit-def dead renamable $sgpr0_sgpr1, implicit killed renamable $sgpr0_sgpr1
; CHECK-NEXT: $sgpr0 = V_READLANE_B32 $vgpr0, 0, implicit-def $sgpr0_sgpr1
- ; CHECK-NEXT: $sgpr1 = V_READLANE_B32 killed $vgpr0, 1
+ ; CHECK-NEXT: $sgpr1 = V_READLANE_B32 $vgpr0, 1
; CHECK-NEXT: $exec = S_MOV_B64 killed $sgpr0_sgpr1
; CHECK-NEXT: S_SENDMSG 0, implicit $m0, implicit $exec
S_NOP 0, implicit-def $exec
@@ -93,12 +96,13 @@
body: |
bb.0:
; CHECK-LABEL: name: reload_sgpr_spill_into_copy_to_exec_lo
- ; CHECK: renamable $vgpr0 = IMPLICIT_DEF
+ ; CHECK: liveins: $vgpr0
+ ; CHECK-NEXT: {{ $}}
; CHECK-NEXT: S_NOP 0, implicit-def renamable $sgpr0, implicit-def dead renamable $sgpr1, implicit-def $exec_lo
- ; CHECK-NEXT: renamable $vgpr0 = V_WRITELANE_B32 killed $sgpr0, 0, killed $vgpr0
+ ; CHECK-NEXT: $vgpr0 = V_WRITELANE_B32 killed $sgpr0, 0, $vgpr0
; CHECK-NEXT: $sgpr0 = V_READLANE_B32 $vgpr0, 0
; CHECK-NEXT: S_NOP 0, implicit killed renamable $sgpr0, implicit-def dead renamable $sgpr1, implicit-def dead renamable $sgpr0
- ; CHECK-NEXT: $sgpr0 = V_READLANE_B32 killed $vgpr0, 0
+ ; CHECK-NEXT: $sgpr0 = V_READLANE_B32 $vgpr0, 0
; CHECK-NEXT: $exec_lo = S_MOV_B32 killed $sgpr0
; CHECK-NEXT: S_SENDMSG 0, implicit $m0, implicit $exec
S_NOP 0, implicit-def %0:sreg_32, implicit-def %1:sreg_32, implicit-def $exec_lo
@@ -116,12 +120,13 @@
body: |
bb.0:
; CHECK-LABEL: name: reload_sgpr_spill_into_copy_to_exec_hi
- ; CHECK: renamable $vgpr0 = IMPLICIT_DEF
+ ; CHECK: liveins: $vgpr0
+ ; CHECK-NEXT: {{ $}}
; CHECK-NEXT: S_NOP 0, implicit-def renamable $sgpr0, implicit-def dead renamable $sgpr1, implicit-def $exec_hi
- ; CHECK-NEXT: renamable $vgpr0 = V_WRITELANE_B32 killed $sgpr0, 0, killed $vgpr0
+ ; CHECK-NEXT: $vgpr0 = V_WRITELANE_B32 killed $sgpr0, 0, $vgpr0
; CHECK-NEXT: $sgpr0 = V_READLANE_B32 $vgpr0, 0
; CHECK-NEXT: S_NOP 0, implicit killed renamable $sgpr0, implicit-def dead renamable $sgpr1, implicit-def dead renamable $sgpr0
- ; CHECK-NEXT: $sgpr0 = V_READLANE_B32 killed $vgpr0, 0
+ ; CHECK-NEXT: $sgpr0 = V_READLANE_B32 $vgpr0, 0
; CHECK-NEXT: $exec_hi = S_MOV_B32 killed $sgpr0
; CHECK-NEXT: S_SENDMSG 0, implicit $m0, implicit $exec
S_NOP 0, implicit-def %0:sreg_32, implicit-def %1:sreg_32, implicit-def $exec_hi
@@ -139,15 +144,16 @@
body: |
bb.0:
; CHECK-LABEL: name: reload_sgpr_spill_into_copy_to_exec
- ; CHECK: renamable $vgpr0 = IMPLICIT_DEF
+ ; CHECK: liveins: $vgpr0
+ ; CHECK-NEXT: {{ $}}
; CHECK-NEXT: S_NOP 0, implicit-def renamable $sgpr0_sgpr1, implicit-def dead renamable $sgpr2_sgpr3, implicit-def $exec
- ; CHECK-NEXT: renamable $vgpr0 = V_WRITELANE_B32 killed $sgpr0, 0, killed $vgpr0, implicit-def $sgpr0_sgpr1, implicit $sgpr0_sgpr1
- ; CHECK-NEXT: renamable $vgpr0 = V_WRITELANE_B32 killed $sgpr1, 1, killed $vgpr0, implicit $sgpr0_sgpr1
+ ; CHECK-NEXT: $vgpr0 = V_WRITELANE_B32 killed $sgpr0, 0, $vgpr0, implicit-def $sgpr0_sgpr1, implicit $sgpr0_sgpr1
+ ; CHECK-NEXT: $vgpr0 = V_WRITELANE_B32 killed $sgpr1, 1, $vgpr0, implicit $sgpr0_sgpr1
; CHECK-NEXT: $sgpr0 = V_READLANE_B32 $vgpr0, 0, implicit-def $sgpr0_sgpr1
; CHECK-NEXT: $sgpr1 = V_READLANE_B32 $vgpr0, 1
; CHECK-NEXT: S_NOP 0, implicit killed renamable $sgpr0_sgpr1, implicit-def dead renamable $sgpr2_sgpr3, implicit-def dead renamable $sgpr0_sgpr1
; CHECK-NEXT: $sgpr0 = V_READLANE_B32 $vgpr0, 0, implicit-def $sgpr0_sgpr1
- ; CHECK-NEXT: $sgpr1 = V_READLANE_B32 killed $vgpr0, 1
+ ; CHECK-NEXT: $sgpr1 = V_READLANE_B32 $vgpr0, 1
; CHECK-NEXT: $exec = S_MOV_B64 killed $sgpr0_sgpr1
; CHECK-NEXT: S_SENDMSG 0, implicit $m0, implicit $exec
S_NOP 0, implicit-def %0:sreg_64, implicit-def %1:sreg_64, implicit-def $exec
diff --git a/llvm/test/CodeGen/AMDGPU/fold-reload-into-m0.mir b/llvm/test/CodeGen/AMDGPU/fold-reload-into-m0.mir
--- a/llvm/test/CodeGen/AMDGPU/fold-reload-into-m0.mir
+++ b/llvm/test/CodeGen/AMDGPU/fold-reload-into-m0.mir
@@ -13,13 +13,14 @@
bb.0:
; CHECK-LABEL: name: merge_sgpr_spill_into_copy_from_m0
- ; CHECK: renamable $vgpr0 = IMPLICIT_DEF
+ ; CHECK: liveins: $vgpr0
+ ; CHECK-NEXT: {{ $}}
; CHECK-NEXT: S_NOP 0, implicit-def $m0
; CHECK-NEXT: $sgpr0 = S_MOV_B32 $m0
- ; CHECK-NEXT: renamable $vgpr0 = V_WRITELANE_B32 killed $sgpr0, 0, killed $vgpr0
+ ; CHECK-NEXT: $vgpr0 = V_WRITELANE_B32 killed $sgpr0, 0, $vgpr0
; CHECK-NEXT: $sgpr0 = V_READLANE_B32 $vgpr0, 0
; CHECK-NEXT: S_NOP 0, implicit-def dead renamable $sgpr1, implicit-def dead renamable $sgpr0, implicit killed renamable $sgpr0
- ; CHECK-NEXT: $sgpr0 = V_READLANE_B32 killed $vgpr0, 0
+ ; CHECK-NEXT: $sgpr0 = V_READLANE_B32 $vgpr0, 0
; CHECK-NEXT: $m0 = S_MOV_B32 killed $sgpr0
; CHECK-NEXT: S_NOP 0
; CHECK-NEXT: S_SENDMSG 0, implicit $m0, implicit $exec
@@ -43,12 +44,13 @@
bb.0:
; CHECK-LABEL: name: reload_sgpr_spill_into_copy_to_m0
- ; CHECK: renamable $vgpr0 = IMPLICIT_DEF
+ ; CHECK: liveins: $vgpr0
+ ; CHECK-NEXT: {{ $}}
; CHECK-NEXT: S_NOP 0, implicit-def renamable $sgpr0, implicit-def dead renamable $sgpr1, implicit-def $m0
- ; CHECK-NEXT: renamable $vgpr0 = V_WRITELANE_B32 killed $sgpr0, 0, killed $vgpr0
+ ; CHECK-NEXT: $vgpr0 = V_WRITELANE_B32 killed $sgpr0, 0, $vgpr0
; CHECK-NEXT: $sgpr0 = V_READLANE_B32 $vgpr0, 0
; CHECK-NEXT: S_NOP 0, implicit killed renamable $sgpr0, implicit-def dead renamable $sgpr1, implicit-def dead renamable $sgpr0
- ; CHECK-NEXT: $sgpr0 = V_READLANE_B32 killed $vgpr0, 0
+ ; CHECK-NEXT: $sgpr0 = V_READLANE_B32 $vgpr0, 0
; CHECK-NEXT: $m0 = S_MOV_B32 killed $sgpr0
; CHECK-NEXT: S_NOP 0
; CHECK-NEXT: S_SENDMSG 0, implicit $m0, implicit $exec
diff --git a/llvm/test/CodeGen/AMDGPU/frame-setup-without-sgpr-to-vgpr-spills.ll b/llvm/test/CodeGen/AMDGPU/frame-setup-without-sgpr-to-vgpr-spills.ll
--- a/llvm/test/CodeGen/AMDGPU/frame-setup-without-sgpr-to-vgpr-spills.ll
+++ b/llvm/test/CodeGen/AMDGPU/frame-setup-without-sgpr-to-vgpr-spills.ll
@@ -14,11 +14,12 @@
; SPILL-TO-VGPR-NEXT: s_mov_b32 s33, s32
; SPILL-TO-VGPR-NEXT: s_or_saveexec_b64 s[8:9], -1
; SPILL-TO-VGPR-NEXT: buffer_store_dword v40, off, s[0:3], s33 offset:4 ; 4-byte Folded Spill
+; SPILL-TO-VGPR-NEXT: buffer_store_dword v41, off, s[0:3], s33 offset:8 ; 4-byte Folded Spill
; SPILL-TO-VGPR-NEXT: s_mov_b64 exec, s[8:9]
-; SPILL-TO-VGPR-NEXT: v_writelane_b32 v40, s4, 2
; SPILL-TO-VGPR-NEXT: s_addk_i32 s32, 0x400
; SPILL-TO-VGPR-NEXT: v_writelane_b32 v40, s30, 0
; SPILL-TO-VGPR-NEXT: v_mov_b32_e32 v0, 0
+; SPILL-TO-VGPR-NEXT: v_writelane_b32 v41, s4, 0
; SPILL-TO-VGPR-NEXT: v_writelane_b32 v40, s31, 1
; SPILL-TO-VGPR-NEXT: buffer_store_dword v0, off, s[0:3], s33
; SPILL-TO-VGPR-NEXT: s_waitcnt vmcnt(0)
@@ -28,9 +29,10 @@
; SPILL-TO-VGPR-NEXT: s_swappc_b64 s[30:31], s[4:5]
; SPILL-TO-VGPR-NEXT: v_readlane_b32 s31, v40, 1
; SPILL-TO-VGPR-NEXT: v_readlane_b32 s30, v40, 0
-; SPILL-TO-VGPR-NEXT: v_readlane_b32 s4, v40, 2
+; SPILL-TO-VGPR-NEXT: v_readlane_b32 s4, v41, 0
; SPILL-TO-VGPR-NEXT: s_or_saveexec_b64 s[6:7], -1
; SPILL-TO-VGPR-NEXT: buffer_load_dword v40, off, s[0:3], s33 offset:4 ; 4-byte Folded Reload
+; SPILL-TO-VGPR-NEXT: buffer_load_dword v41, off, s[0:3], s33 offset:8 ; 4-byte Folded Reload
; SPILL-TO-VGPR-NEXT: s_mov_b64 exec, s[6:7]
; SPILL-TO-VGPR-NEXT: s_addk_i32 s32, 0xfc00
; SPILL-TO-VGPR-NEXT: s_mov_b32 s33, s4
diff --git a/llvm/test/CodeGen/AMDGPU/gfx-call-non-gfx-func.ll b/llvm/test/CodeGen/AMDGPU/gfx-call-non-gfx-func.ll
--- a/llvm/test/CodeGen/AMDGPU/gfx-call-non-gfx-func.ll
+++ b/llvm/test/CodeGen/AMDGPU/gfx-call-non-gfx-func.ll
@@ -8,7 +8,7 @@
; SDAG-LABEL: gfx_func:
; SDAG: ; %bb.0:
; SDAG-NEXT: s_waitcnt vmcnt(0) expcnt(0) lgkmcnt(0)
-; SDAG-NEXT: s_mov_b32 s36, s33
+; SDAG-NEXT: s_mov_b32 s38, s33
; SDAG-NEXT: s_mov_b32 s33, s32
; SDAG-NEXT: s_or_saveexec_b64 s[34:35], -1
; SDAG-NEXT: buffer_store_dword v40, off, s[0:3], s33 ; 4-byte Folded Spill
@@ -81,14 +81,14 @@
; SDAG-NEXT: buffer_load_dword v40, off, s[0:3], s33 ; 4-byte Folded Reload
; SDAG-NEXT: s_mov_b64 exec, s[34:35]
; SDAG-NEXT: s_addk_i32 s32, 0xfc00
-; SDAG-NEXT: s_mov_b32 s33, s36
+; SDAG-NEXT: s_mov_b32 s33, s38
; SDAG-NEXT: s_waitcnt vmcnt(0)
; SDAG-NEXT: s_setpc_b64 s[30:31]
;
; GISEL-LABEL: gfx_func:
; GISEL: ; %bb.0:
; GISEL-NEXT: s_waitcnt vmcnt(0) expcnt(0) lgkmcnt(0)
-; GISEL-NEXT: s_mov_b32 s36, s33
+; GISEL-NEXT: s_mov_b32 s38, s33
; GISEL-NEXT: s_mov_b32 s33, s32
; GISEL-NEXT: s_or_saveexec_b64 s[34:35], -1
; GISEL-NEXT: buffer_store_dword v40, off, s[0:3], s33 ; 4-byte Folded Spill
@@ -161,7 +161,7 @@
; GISEL-NEXT: buffer_load_dword v40, off, s[0:3], s33 ; 4-byte Folded Reload
; GISEL-NEXT: s_mov_b64 exec, s[34:35]
; GISEL-NEXT: s_addk_i32 s32, 0xfc00
-; GISEL-NEXT: s_mov_b32 s33, s36
+; GISEL-NEXT: s_mov_b32 s33, s38
; GISEL-NEXT: s_waitcnt vmcnt(0)
; GISEL-NEXT: s_setpc_b64 s[30:31]
call void @extern_c_func()
diff --git a/llvm/test/CodeGen/AMDGPU/gfx-callable-argument-types.ll b/llvm/test/CodeGen/AMDGPU/gfx-callable-argument-types.ll
--- a/llvm/test/CodeGen/AMDGPU/gfx-callable-argument-types.ll
+++ b/llvm/test/CodeGen/AMDGPU/gfx-callable-argument-types.ll
@@ -101,11 +101,12 @@
; GFX9-NEXT: s_mov_b32 s33, s32
; GFX9-NEXT: s_or_saveexec_b64 s[36:37], -1
; GFX9-NEXT: buffer_store_dword v40, off, s[0:3], s33 ; 4-byte Folded Spill
+; GFX9-NEXT: buffer_store_dword v41, off, s[0:3], s33 offset:4 ; 4-byte Folded Spill
; GFX9-NEXT: s_mov_b64 exec, s[36:37]
-; GFX9-NEXT: v_writelane_b32 v40, s34, 2
; GFX9-NEXT: s_addk_i32 s32, 0x400
; GFX9-NEXT: v_writelane_b32 v40, s30, 0
; GFX9-NEXT: v_mov_b32_e32 v0, 1
+; GFX9-NEXT: v_writelane_b32 v41, s34, 0
; GFX9-NEXT: v_writelane_b32 v40, s31, 1
; GFX9-NEXT: s_getpc_b64 s[34:35]
; GFX9-NEXT: s_add_u32 s34, s34, external_void_func_i1@rel32@lo+4
@@ -114,9 +115,10 @@
; GFX9-NEXT: s_swappc_b64 s[30:31], s[34:35]
; GFX9-NEXT: v_readlane_b32 s31, v40, 1
; GFX9-NEXT: v_readlane_b32 s30, v40, 0
-; GFX9-NEXT: v_readlane_b32 s34, v40, 2
+; GFX9-NEXT: v_readlane_b32 s34, v41, 0
; GFX9-NEXT: s_or_saveexec_b64 s[36:37], -1
; GFX9-NEXT: buffer_load_dword v40, off, s[0:3], s33 ; 4-byte Folded Reload
+; GFX9-NEXT: buffer_load_dword v41, off, s[0:3], s33 offset:4 ; 4-byte Folded Reload
; GFX9-NEXT: s_mov_b64 exec, s[36:37]
; GFX9-NEXT: s_addk_i32 s32, 0xfc00
; GFX9-NEXT: s_mov_b32 s33, s34
@@ -130,23 +132,26 @@
; GFX10-NEXT: s_mov_b32 s33, s32
; GFX10-NEXT: s_or_saveexec_b32 s35, -1
; GFX10-NEXT: buffer_store_dword v40, off, s[0:3], s33 ; 4-byte Folded Spill
+; GFX10-NEXT: buffer_store_dword v41, off, s[0:3], s33 offset:4 ; 4-byte Folded Spill
; GFX10-NEXT: s_waitcnt_depctr 0xffe3
; GFX10-NEXT: s_mov_b32 exec_lo, s35
-; GFX10-NEXT: v_writelane_b32 v40, s34, 2
+; GFX10-NEXT: v_writelane_b32 v40, s30, 0
; GFX10-NEXT: v_mov_b32_e32 v0, 1
; GFX10-NEXT: s_addk_i32 s32, 0x200
+; GFX10-NEXT: v_writelane_b32 v41, s34, 0
; GFX10-NEXT: s_getpc_b64 s[34:35]
; GFX10-NEXT: s_add_u32 s34, s34, external_void_func_i1@rel32@lo+4
; GFX10-NEXT: s_addc_u32 s35, s35, external_void_func_i1@rel32@hi+12
-; GFX10-NEXT: v_writelane_b32 v40, s30, 0
-; GFX10-NEXT: buffer_store_byte v0, off, s[0:3], s32
; GFX10-NEXT: v_writelane_b32 v40, s31, 1
+; GFX10-NEXT: buffer_store_byte v0, off, s[0:3], s32
; GFX10-NEXT: s_swappc_b64 s[30:31], s[34:35]
; GFX10-NEXT: v_readlane_b32 s31, v40, 1
; GFX10-NEXT: v_readlane_b32 s30, v40, 0
-; GFX10-NEXT: v_readlane_b32 s34, v40, 2
+; GFX10-NEXT: v_readlane_b32 s34, v41, 0
; GFX10-NEXT: s_or_saveexec_b32 s35, -1
-; GFX10-NEXT: buffer_load_dword v40, off, s[0:3], s33 ; 4-byte Folded Reload
+; GFX10-NEXT: s_clause 0x1
+; GFX10-NEXT: buffer_load_dword v40, off, s[0:3], s33
+; GFX10-NEXT: buffer_load_dword v41, off, s[0:3], s33 offset:4
; GFX10-NEXT: s_waitcnt_depctr 0xffe3
; GFX10-NEXT: s_mov_b32 exec_lo, s35
; GFX10-NEXT: s_addk_i32 s32, 0xfe00
@@ -160,24 +165,27 @@
; GFX11-NEXT: s_mov_b32 s0, s33
; GFX11-NEXT: s_mov_b32 s33, s32
; GFX11-NEXT: s_or_saveexec_b32 s1, -1
-; GFX11-NEXT: scratch_store_b32 off, v40, s33 ; 4-byte Folded Spill
+; GFX11-NEXT: s_clause 0x1
+; GFX11-NEXT: scratch_store_b32 off, v40, s33
+; GFX11-NEXT: scratch_store_b32 off, v41, s33 offset:4
; GFX11-NEXT: s_mov_b32 exec_lo, s1
-; GFX11-NEXT: v_writelane_b32 v40, s0, 2
+; GFX11-NEXT: v_writelane_b32 v40, s30, 0
; GFX11-NEXT: v_mov_b32_e32 v0, 1
; GFX11-NEXT: s_add_i32 s32, s32, 16
+; GFX11-NEXT: v_writelane_b32 v41, s0, 0
; GFX11-NEXT: s_getpc_b64 s[0:1]
; GFX11-NEXT: s_add_u32 s0, s0, external_void_func_i1@rel32@lo+4
; GFX11-NEXT: s_addc_u32 s1, s1, external_void_func_i1@rel32@hi+12
-; GFX11-NEXT: v_writelane_b32 v40, s30, 0
-; GFX11-NEXT: scratch_store_b8 off, v0, s32
; GFX11-NEXT: v_writelane_b32 v40, s31, 1
+; GFX11-NEXT: scratch_store_b8 off, v0, s32
; GFX11-NEXT: s_swappc_b64 s[30:31], s[0:1]
-; GFX11-NEXT: s_delay_alu instid0(VALU_DEP_1)
; GFX11-NEXT: v_readlane_b32 s31, v40, 1
; GFX11-NEXT: v_readlane_b32 s30, v40, 0
-; GFX11-NEXT: v_readlane_b32 s0, v40, 2
+; GFX11-NEXT: v_readlane_b32 s0, v41, 0
; GFX11-NEXT: s_or_saveexec_b32 s1, -1
-; GFX11-NEXT: scratch_load_b32 v40, off, s33 ; 4-byte Folded Reload
+; GFX11-NEXT: s_clause 0x1
+; GFX11-NEXT: scratch_load_b32 v40, off, s33
+; GFX11-NEXT: scratch_load_b32 v41, off, s33 offset:4
; GFX11-NEXT: s_mov_b32 exec_lo, s1
; GFX11-NEXT: s_add_i32 s32, s32, -16
; GFX11-NEXT: s_mov_b32 s33, s0
@@ -191,23 +199,26 @@
; GFX10-SCRATCH-NEXT: s_mov_b32 s33, s32
; GFX10-SCRATCH-NEXT: s_or_saveexec_b32 s1, -1
; GFX10-SCRATCH-NEXT: scratch_store_dword off, v40, s33 ; 4-byte Folded Spill
+; GFX10-SCRATCH-NEXT: scratch_store_dword off, v41, s33 offset:4 ; 4-byte Folded Spill
; GFX10-SCRATCH-NEXT: s_waitcnt_depctr 0xffe3
; GFX10-SCRATCH-NEXT: s_mov_b32 exec_lo, s1
-; GFX10-SCRATCH-NEXT: v_writelane_b32 v40, s0, 2
+; GFX10-SCRATCH-NEXT: v_writelane_b32 v40, s30, 0
; GFX10-SCRATCH-NEXT: v_mov_b32_e32 v0, 1
; GFX10-SCRATCH-NEXT: s_add_i32 s32, s32, 16
+; GFX10-SCRATCH-NEXT: v_writelane_b32 v41, s0, 0
; GFX10-SCRATCH-NEXT: s_getpc_b64 s[0:1]
; GFX10-SCRATCH-NEXT: s_add_u32 s0, s0, external_void_func_i1@rel32@lo+4
; GFX10-SCRATCH-NEXT: s_addc_u32 s1, s1, external_void_func_i1@rel32@hi+12
-; GFX10-SCRATCH-NEXT: v_writelane_b32 v40, s30, 0
-; GFX10-SCRATCH-NEXT: scratch_store_byte off, v0, s32
; GFX10-SCRATCH-NEXT: v_writelane_b32 v40, s31, 1
+; GFX10-SCRATCH-NEXT: scratch_store_byte off, v0, s32
; GFX10-SCRATCH-NEXT: s_swappc_b64 s[30:31], s[0:1]
; GFX10-SCRATCH-NEXT: v_readlane_b32 s31, v40, 1
; GFX10-SCRATCH-NEXT: v_readlane_b32 s30, v40, 0
-; GFX10-SCRATCH-NEXT: v_readlane_b32 s0, v40, 2
+; GFX10-SCRATCH-NEXT: v_readlane_b32 s0, v41, 0
; GFX10-SCRATCH-NEXT: s_or_saveexec_b32 s1, -1
-; GFX10-SCRATCH-NEXT: scratch_load_dword v40, off, s33 ; 4-byte Folded Reload
+; GFX10-SCRATCH-NEXT: s_clause 0x1
+; GFX10-SCRATCH-NEXT: scratch_load_dword v40, off, s33
+; GFX10-SCRATCH-NEXT: scratch_load_dword v41, off, s33 offset:4
; GFX10-SCRATCH-NEXT: s_waitcnt_depctr 0xffe3
; GFX10-SCRATCH-NEXT: s_mov_b32 exec_lo, s1
; GFX10-SCRATCH-NEXT: s_add_i32 s32, s32, -16
@@ -226,12 +237,13 @@
; GFX9-NEXT: s_mov_b32 s33, s32
; GFX9-NEXT: s_or_saveexec_b64 s[36:37], -1
; GFX9-NEXT: buffer_store_dword v40, off, s[0:3], s33 ; 4-byte Folded Spill
+; GFX9-NEXT: buffer_store_dword v41, off, s[0:3], s33 offset:4 ; 4-byte Folded Spill
; GFX9-NEXT: s_mov_b64 exec, s[36:37]
; GFX9-NEXT: global_load_ubyte v0, v[0:1], off glc
; GFX9-NEXT: s_waitcnt vmcnt(0)
-; GFX9-NEXT: v_writelane_b32 v40, s34, 2
; GFX9-NEXT: s_addk_i32 s32, 0x400
; GFX9-NEXT: v_writelane_b32 v40, s30, 0
+; GFX9-NEXT: v_writelane_b32 v41, s34, 0
; GFX9-NEXT: v_writelane_b32 v40, s31, 1
; GFX9-NEXT: s_getpc_b64 s[34:35]
; GFX9-NEXT: s_add_u32 s34, s34, external_void_func_i1_signext@rel32@lo+4
@@ -241,9 +253,10 @@
; GFX9-NEXT: s_swappc_b64 s[30:31], s[34:35]
; GFX9-NEXT: v_readlane_b32 s31, v40, 1
; GFX9-NEXT: v_readlane_b32 s30, v40, 0
-; GFX9-NEXT: v_readlane_b32 s34, v40, 2
+; GFX9-NEXT: v_readlane_b32 s34, v41, 0
; GFX9-NEXT: s_or_saveexec_b64 s[36:37], -1
; GFX9-NEXT: buffer_load_dword v40, off, s[0:3], s33 ; 4-byte Folded Reload
+; GFX9-NEXT: buffer_load_dword v41, off, s[0:3], s33 offset:4 ; 4-byte Folded Reload
; GFX9-NEXT: s_mov_b64 exec, s[36:37]
; GFX9-NEXT: s_addk_i32 s32, 0xfc00
; GFX9-NEXT: s_mov_b32 s33, s34
@@ -257,25 +270,28 @@
; GFX10-NEXT: s_mov_b32 s33, s32
; GFX10-NEXT: s_or_saveexec_b32 s35, -1
; GFX10-NEXT: buffer_store_dword v40, off, s[0:3], s33 ; 4-byte Folded Spill
+; GFX10-NEXT: buffer_store_dword v41, off, s[0:3], s33 offset:4 ; 4-byte Folded Spill
; GFX10-NEXT: s_waitcnt_depctr 0xffe3
; GFX10-NEXT: s_mov_b32 exec_lo, s35
; GFX10-NEXT: global_load_ubyte v0, v[0:1], off glc dlc
; GFX10-NEXT: s_waitcnt vmcnt(0)
-; GFX10-NEXT: v_writelane_b32 v40, s34, 2
+; GFX10-NEXT: v_writelane_b32 v40, s30, 0
; GFX10-NEXT: s_addk_i32 s32, 0x200
+; GFX10-NEXT: v_writelane_b32 v41, s34, 0
; GFX10-NEXT: s_getpc_b64 s[34:35]
; GFX10-NEXT: s_add_u32 s34, s34, external_void_func_i1_signext@rel32@lo+4
; GFX10-NEXT: s_addc_u32 s35, s35, external_void_func_i1_signext@rel32@hi+12
-; GFX10-NEXT: v_writelane_b32 v40, s30, 0
; GFX10-NEXT: v_writelane_b32 v40, s31, 1
; GFX10-NEXT: v_and_b32_e32 v0, 1, v0
; GFX10-NEXT: buffer_store_byte v0, off, s[0:3], s32
; GFX10-NEXT: s_swappc_b64 s[30:31], s[34:35]
; GFX10-NEXT: v_readlane_b32 s31, v40, 1
; GFX10-NEXT: v_readlane_b32 s30, v40, 0
-; GFX10-NEXT: v_readlane_b32 s34, v40, 2
+; GFX10-NEXT: v_readlane_b32 s34, v41, 0
; GFX10-NEXT: s_or_saveexec_b32 s35, -1
-; GFX10-NEXT: buffer_load_dword v40, off, s[0:3], s33 ; 4-byte Folded Reload
+; GFX10-NEXT: s_clause 0x1
+; GFX10-NEXT: buffer_load_dword v40, off, s[0:3], s33
+; GFX10-NEXT: buffer_load_dword v41, off, s[0:3], s33 offset:4
; GFX10-NEXT: s_waitcnt_depctr 0xffe3
; GFX10-NEXT: s_mov_b32 exec_lo, s35
; GFX10-NEXT: s_addk_i32 s32, 0xfe00
@@ -289,25 +305,29 @@
; GFX11-NEXT: s_mov_b32 s0, s33
; GFX11-NEXT: s_mov_b32 s33, s32
; GFX11-NEXT: s_or_saveexec_b32 s1, -1
-; GFX11-NEXT: scratch_store_b32 off, v40, s33 ; 4-byte Folded Spill
+; GFX11-NEXT: s_clause 0x1
+; GFX11-NEXT: scratch_store_b32 off, v40, s33
+; GFX11-NEXT: scratch_store_b32 off, v41, s33 offset:4
; GFX11-NEXT: s_mov_b32 exec_lo, s1
; GFX11-NEXT: global_load_u8 v0, v[0:1], off glc dlc
; GFX11-NEXT: s_waitcnt vmcnt(0)
-; GFX11-NEXT: v_writelane_b32 v40, s0, 2
+; GFX11-NEXT: v_writelane_b32 v40, s30, 0
; GFX11-NEXT: s_add_i32 s32, s32, 16
+; GFX11-NEXT: v_writelane_b32 v41, s0, 0
; GFX11-NEXT: s_getpc_b64 s[0:1]
; GFX11-NEXT: s_add_u32 s0, s0, external_void_func_i1_signext@rel32@lo+4
; GFX11-NEXT: s_addc_u32 s1, s1, external_void_func_i1_signext@rel32@hi+12
-; GFX11-NEXT: v_writelane_b32 v40, s30, 0
; GFX11-NEXT: v_writelane_b32 v40, s31, 1
; GFX11-NEXT: v_and_b32_e32 v0, 1, v0
; GFX11-NEXT: scratch_store_b8 off, v0, s32
; GFX11-NEXT: s_swappc_b64 s[30:31], s[0:1]
; GFX11-NEXT: v_readlane_b32 s31, v40, 1
; GFX11-NEXT: v_readlane_b32 s30, v40, 0
-; GFX11-NEXT: v_readlane_b32 s0, v40, 2
+; GFX11-NEXT: v_readlane_b32 s0, v41, 0
; GFX11-NEXT: s_or_saveexec_b32 s1, -1
-; GFX11-NEXT: scratch_load_b32 v40, off, s33 ; 4-byte Folded Reload
+; GFX11-NEXT: s_clause 0x1
+; GFX11-NEXT: scratch_load_b32 v40, off, s33
+; GFX11-NEXT: scratch_load_b32 v41, off, s33 offset:4
; GFX11-NEXT: s_mov_b32 exec_lo, s1
; GFX11-NEXT: s_add_i32 s32, s32, -16
; GFX11-NEXT: s_mov_b32 s33, s0
@@ -321,25 +341,28 @@
; GFX10-SCRATCH-NEXT: s_mov_b32 s33, s32
; GFX10-SCRATCH-NEXT: s_or_saveexec_b32 s1, -1
; GFX10-SCRATCH-NEXT: scratch_store_dword off, v40, s33 ; 4-byte Folded Spill
+; GFX10-SCRATCH-NEXT: scratch_store_dword off, v41, s33 offset:4 ; 4-byte Folded Spill
; GFX10-SCRATCH-NEXT: s_waitcnt_depctr 0xffe3
; GFX10-SCRATCH-NEXT: s_mov_b32 exec_lo, s1
; GFX10-SCRATCH-NEXT: global_load_ubyte v0, v[0:1], off glc dlc
; GFX10-SCRATCH-NEXT: s_waitcnt vmcnt(0)
-; GFX10-SCRATCH-NEXT: v_writelane_b32 v40, s0, 2
+; GFX10-SCRATCH-NEXT: v_writelane_b32 v40, s30, 0
; GFX10-SCRATCH-NEXT: s_add_i32 s32, s32, 16
+; GFX10-SCRATCH-NEXT: v_writelane_b32 v41, s0, 0
; GFX10-SCRATCH-NEXT: s_getpc_b64 s[0:1]
; GFX10-SCRATCH-NEXT: s_add_u32 s0, s0, external_void_func_i1_signext@rel32@lo+4
; GFX10-SCRATCH-NEXT: s_addc_u32 s1, s1, external_void_func_i1_signext@rel32@hi+12
-; GFX10-SCRATCH-NEXT: v_writelane_b32 v40, s30, 0
; GFX10-SCRATCH-NEXT: v_writelane_b32 v40, s31, 1
; GFX10-SCRATCH-NEXT: v_and_b32_e32 v0, 1, v0
; GFX10-SCRATCH-NEXT: scratch_store_byte off, v0, s32
; GFX10-SCRATCH-NEXT: s_swappc_b64 s[30:31], s[0:1]
; GFX10-SCRATCH-NEXT: v_readlane_b32 s31, v40, 1
; GFX10-SCRATCH-NEXT: v_readlane_b32 s30, v40, 0
-; GFX10-SCRATCH-NEXT: v_readlane_b32 s0, v40, 2
+; GFX10-SCRATCH-NEXT: v_readlane_b32 s0, v41, 0
; GFX10-SCRATCH-NEXT: s_or_saveexec_b32 s1, -1
-; GFX10-SCRATCH-NEXT: scratch_load_dword v40, off, s33 ; 4-byte Folded Reload
+; GFX10-SCRATCH-NEXT: s_clause 0x1
+; GFX10-SCRATCH-NEXT: scratch_load_dword v40, off, s33
+; GFX10-SCRATCH-NEXT: scratch_load_dword v41, off, s33 offset:4
; GFX10-SCRATCH-NEXT: s_waitcnt_depctr 0xffe3
; GFX10-SCRATCH-NEXT: s_mov_b32 exec_lo, s1
; GFX10-SCRATCH-NEXT: s_add_i32 s32, s32, -16
@@ -359,12 +382,13 @@
; GFX9-NEXT: s_mov_b32 s33, s32
; GFX9-NEXT: s_or_saveexec_b64 s[36:37], -1
; GFX9-NEXT: buffer_store_dword v40, off, s[0:3], s33 ; 4-byte Folded Spill
+; GFX9-NEXT: buffer_store_dword v41, off, s[0:3], s33 offset:4 ; 4-byte Folded Spill
; GFX9-NEXT: s_mov_b64 exec, s[36:37]
; GFX9-NEXT: global_load_ubyte v0, v[0:1], off glc
; GFX9-NEXT: s_waitcnt vmcnt(0)
-; GFX9-NEXT: v_writelane_b32 v40, s34, 2
; GFX9-NEXT: s_addk_i32 s32, 0x400
; GFX9-NEXT: v_writelane_b32 v40, s30, 0
+; GFX9-NEXT: v_writelane_b32 v41, s34, 0
; GFX9-NEXT: v_writelane_b32 v40, s31, 1
; GFX9-NEXT: s_getpc_b64 s[34:35]
; GFX9-NEXT: s_add_u32 s34, s34, external_void_func_i1_zeroext@rel32@lo+4
@@ -374,9 +398,10 @@
; GFX9-NEXT: s_swappc_b64 s[30:31], s[34:35]
; GFX9-NEXT: v_readlane_b32 s31, v40, 1
; GFX9-NEXT: v_readlane_b32 s30, v40, 0
-; GFX9-NEXT: v_readlane_b32 s34, v40, 2
+; GFX9-NEXT: v_readlane_b32 s34, v41, 0
; GFX9-NEXT: s_or_saveexec_b64 s[36:37], -1
; GFX9-NEXT: buffer_load_dword v40, off, s[0:3], s33 ; 4-byte Folded Reload
+; GFX9-NEXT: buffer_load_dword v41, off, s[0:3], s33 offset:4 ; 4-byte Folded Reload
; GFX9-NEXT: s_mov_b64 exec, s[36:37]
; GFX9-NEXT: s_addk_i32 s32, 0xfc00
; GFX9-NEXT: s_mov_b32 s33, s34
@@ -390,25 +415,28 @@
; GFX10-NEXT: s_mov_b32 s33, s32
; GFX10-NEXT: s_or_saveexec_b32 s35, -1
; GFX10-NEXT: buffer_store_dword v40, off, s[0:3], s33 ; 4-byte Folded Spill
+; GFX10-NEXT: buffer_store_dword v41, off, s[0:3], s33 offset:4 ; 4-byte Folded Spill
; GFX10-NEXT: s_waitcnt_depctr 0xffe3
; GFX10-NEXT: s_mov_b32 exec_lo, s35
; GFX10-NEXT: global_load_ubyte v0, v[0:1], off glc dlc
; GFX10-NEXT: s_waitcnt vmcnt(0)
-; GFX10-NEXT: v_writelane_b32 v40, s34, 2
+; GFX10-NEXT: v_writelane_b32 v40, s30, 0
; GFX10-NEXT: s_addk_i32 s32, 0x200
+; GFX10-NEXT: v_writelane_b32 v41, s34, 0
; GFX10-NEXT: s_getpc_b64 s[34:35]
; GFX10-NEXT: s_add_u32 s34, s34, external_void_func_i1_zeroext@rel32@lo+4
; GFX10-NEXT: s_addc_u32 s35, s35, external_void_func_i1_zeroext@rel32@hi+12
-; GFX10-NEXT: v_writelane_b32 v40, s30, 0
; GFX10-NEXT: v_writelane_b32 v40, s31, 1
; GFX10-NEXT: v_and_b32_e32 v0, 1, v0
; GFX10-NEXT: buffer_store_byte v0, off, s[0:3], s32
; GFX10-NEXT: s_swappc_b64 s[30:31], s[34:35]
; GFX10-NEXT: v_readlane_b32 s31, v40, 1
; GFX10-NEXT: v_readlane_b32 s30, v40, 0
-; GFX10-NEXT: v_readlane_b32 s34, v40, 2
+; GFX10-NEXT: v_readlane_b32 s34, v41, 0
; GFX10-NEXT: s_or_saveexec_b32 s35, -1
-; GFX10-NEXT: buffer_load_dword v40, off, s[0:3], s33 ; 4-byte Folded Reload
+; GFX10-NEXT: s_clause 0x1
+; GFX10-NEXT: buffer_load_dword v40, off, s[0:3], s33
+; GFX10-NEXT: buffer_load_dword v41, off, s[0:3], s33 offset:4
; GFX10-NEXT: s_waitcnt_depctr 0xffe3
; GFX10-NEXT: s_mov_b32 exec_lo, s35
; GFX10-NEXT: s_addk_i32 s32, 0xfe00
@@ -422,25 +450,29 @@
; GFX11-NEXT: s_mov_b32 s0, s33
; GFX11-NEXT: s_mov_b32 s33, s32
; GFX11-NEXT: s_or_saveexec_b32 s1, -1
-; GFX11-NEXT: scratch_store_b32 off, v40, s33 ; 4-byte Folded Spill
+; GFX11-NEXT: s_clause 0x1
+; GFX11-NEXT: scratch_store_b32 off, v40, s33
+; GFX11-NEXT: scratch_store_b32 off, v41, s33 offset:4
; GFX11-NEXT: s_mov_b32 exec_lo, s1
; GFX11-NEXT: global_load_u8 v0, v[0:1], off glc dlc
; GFX11-NEXT: s_waitcnt vmcnt(0)
-; GFX11-NEXT: v_writelane_b32 v40, s0, 2
+; GFX11-NEXT: v_writelane_b32 v40, s30, 0
; GFX11-NEXT: s_add_i32 s32, s32, 16
+; GFX11-NEXT: v_writelane_b32 v41, s0, 0
; GFX11-NEXT: s_getpc_b64 s[0:1]
; GFX11-NEXT: s_add_u32 s0, s0, external_void_func_i1_zeroext@rel32@lo+4
; GFX11-NEXT: s_addc_u32 s1, s1, external_void_func_i1_zeroext@rel32@hi+12
-; GFX11-NEXT: v_writelane_b32 v40, s30, 0
; GFX11-NEXT: v_writelane_b32 v40, s31, 1
; GFX11-NEXT: v_and_b32_e32 v0, 1, v0
; GFX11-NEXT: scratch_store_b8 off, v0, s32
; GFX11-NEXT: s_swappc_b64 s[30:31], s[0:1]
; GFX11-NEXT: v_readlane_b32 s31, v40, 1
; GFX11-NEXT: v_readlane_b32 s30, v40, 0
-; GFX11-NEXT: v_readlane_b32 s0, v40, 2
+; GFX11-NEXT: v_readlane_b32 s0, v41, 0
; GFX11-NEXT: s_or_saveexec_b32 s1, -1
-; GFX11-NEXT: scratch_load_b32 v40, off, s33 ; 4-byte Folded Reload
+; GFX11-NEXT: s_clause 0x1
+; GFX11-NEXT: scratch_load_b32 v40, off, s33
+; GFX11-NEXT: scratch_load_b32 v41, off, s33 offset:4
; GFX11-NEXT: s_mov_b32 exec_lo, s1
; GFX11-NEXT: s_add_i32 s32, s32, -16
; GFX11-NEXT: s_mov_b32 s33, s0
@@ -454,25 +486,28 @@
; GFX10-SCRATCH-NEXT: s_mov_b32 s33, s32
; GFX10-SCRATCH-NEXT: s_or_saveexec_b32 s1, -1
; GFX10-SCRATCH-NEXT: scratch_store_dword off, v40, s33 ; 4-byte Folded Spill
+; GFX10-SCRATCH-NEXT: scratch_store_dword off, v41, s33 offset:4 ; 4-byte Folded Spill
; GFX10-SCRATCH-NEXT: s_waitcnt_depctr 0xffe3
; GFX10-SCRATCH-NEXT: s_mov_b32 exec_lo, s1
; GFX10-SCRATCH-NEXT: global_load_ubyte v0, v[0:1], off glc dlc
; GFX10-SCRATCH-NEXT: s_waitcnt vmcnt(0)
-; GFX10-SCRATCH-NEXT: v_writelane_b32 v40, s0, 2
+; GFX10-SCRATCH-NEXT: v_writelane_b32 v40, s30, 0
; GFX10-SCRATCH-NEXT: s_add_i32 s32, s32, 16
+; GFX10-SCRATCH-NEXT: v_writelane_b32 v41, s0, 0
; GFX10-SCRATCH-NEXT: s_getpc_b64 s[0:1]
; GFX10-SCRATCH-NEXT: s_add_u32 s0, s0, external_void_func_i1_zeroext@rel32@lo+4
; GFX10-SCRATCH-NEXT: s_addc_u32 s1, s1, external_void_func_i1_zeroext@rel32@hi+12
-; GFX10-SCRATCH-NEXT: v_writelane_b32 v40, s30, 0
; GFX10-SCRATCH-NEXT: v_writelane_b32 v40, s31, 1
; GFX10-SCRATCH-NEXT: v_and_b32_e32 v0, 1, v0
; GFX10-SCRATCH-NEXT: scratch_store_byte off, v0, s32
; GFX10-SCRATCH-NEXT: s_swappc_b64 s[30:31], s[0:1]
; GFX10-SCRATCH-NEXT: v_readlane_b32 s31, v40, 1
; GFX10-SCRATCH-NEXT: v_readlane_b32 s30, v40, 0
-; GFX10-SCRATCH-NEXT: v_readlane_b32 s0, v40, 2
+; GFX10-SCRATCH-NEXT: v_readlane_b32 s0, v41, 0
; GFX10-SCRATCH-NEXT: s_or_saveexec_b32 s1, -1
-; GFX10-SCRATCH-NEXT: scratch_load_dword v40, off, s33 ; 4-byte Folded Reload
+; GFX10-SCRATCH-NEXT: s_clause 0x1
+; GFX10-SCRATCH-NEXT: scratch_load_dword v40, off, s33
+; GFX10-SCRATCH-NEXT: scratch_load_dword v41, off, s33 offset:4
; GFX10-SCRATCH-NEXT: s_waitcnt_depctr 0xffe3
; GFX10-SCRATCH-NEXT: s_mov_b32 exec_lo, s1
; GFX10-SCRATCH-NEXT: s_add_i32 s32, s32, -16
@@ -492,11 +527,12 @@
; GFX9-NEXT: s_mov_b32 s33, s32
; GFX9-NEXT: s_or_saveexec_b64 s[36:37], -1
; GFX9-NEXT: buffer_store_dword v40, off, s[0:3], s33 ; 4-byte Folded Spill
+; GFX9-NEXT: buffer_store_dword v41, off, s[0:3], s33 offset:4 ; 4-byte Folded Spill
; GFX9-NEXT: s_mov_b64 exec, s[36:37]
-; GFX9-NEXT: v_writelane_b32 v40, s34, 2
; GFX9-NEXT: s_addk_i32 s32, 0x400
; GFX9-NEXT: v_writelane_b32 v40, s30, 0
; GFX9-NEXT: v_mov_b32_e32 v0, 0x7b
+; GFX9-NEXT: v_writelane_b32 v41, s34, 0
; GFX9-NEXT: v_writelane_b32 v40, s31, 1
; GFX9-NEXT: s_getpc_b64 s[34:35]
; GFX9-NEXT: s_add_u32 s34, s34, external_void_func_i8@rel32@lo+4
@@ -504,9 +540,10 @@
; GFX9-NEXT: s_swappc_b64 s[30:31], s[34:35]
; GFX9-NEXT: v_readlane_b32 s31, v40, 1
; GFX9-NEXT: v_readlane_b32 s30, v40, 0
-; GFX9-NEXT: v_readlane_b32 s34, v40, 2
+; GFX9-NEXT: v_readlane_b32 s34, v41, 0
; GFX9-NEXT: s_or_saveexec_b64 s[36:37], -1
; GFX9-NEXT: buffer_load_dword v40, off, s[0:3], s33 ; 4-byte Folded Reload
+; GFX9-NEXT: buffer_load_dword v41, off, s[0:3], s33 offset:4 ; 4-byte Folded Reload
; GFX9-NEXT: s_mov_b64 exec, s[36:37]
; GFX9-NEXT: s_addk_i32 s32, 0xfc00
; GFX9-NEXT: s_mov_b32 s33, s34
@@ -520,22 +557,25 @@
; GFX10-NEXT: s_mov_b32 s33, s32
; GFX10-NEXT: s_or_saveexec_b32 s35, -1
; GFX10-NEXT: buffer_store_dword v40, off, s[0:3], s33 ; 4-byte Folded Spill
+; GFX10-NEXT: buffer_store_dword v41, off, s[0:3], s33 offset:4 ; 4-byte Folded Spill
; GFX10-NEXT: s_waitcnt_depctr 0xffe3
; GFX10-NEXT: s_mov_b32 exec_lo, s35
-; GFX10-NEXT: v_writelane_b32 v40, s34, 2
+; GFX10-NEXT: v_writelane_b32 v40, s30, 0
; GFX10-NEXT: v_mov_b32_e32 v0, 0x7b
; GFX10-NEXT: s_addk_i32 s32, 0x200
+; GFX10-NEXT: v_writelane_b32 v41, s34, 0
; GFX10-NEXT: s_getpc_b64 s[34:35]
; GFX10-NEXT: s_add_u32 s34, s34, external_void_func_i8@rel32@lo+4
; GFX10-NEXT: s_addc_u32 s35, s35, external_void_func_i8@rel32@hi+12
-; GFX10-NEXT: v_writelane_b32 v40, s30, 0
; GFX10-NEXT: v_writelane_b32 v40, s31, 1
; GFX10-NEXT: s_swappc_b64 s[30:31], s[34:35]
; GFX10-NEXT: v_readlane_b32 s31, v40, 1
; GFX10-NEXT: v_readlane_b32 s30, v40, 0
-; GFX10-NEXT: v_readlane_b32 s34, v40, 2
+; GFX10-NEXT: v_readlane_b32 s34, v41, 0
; GFX10-NEXT: s_or_saveexec_b32 s35, -1
-; GFX10-NEXT: buffer_load_dword v40, off, s[0:3], s33 ; 4-byte Folded Reload
+; GFX10-NEXT: s_clause 0x1
+; GFX10-NEXT: buffer_load_dword v40, off, s[0:3], s33
+; GFX10-NEXT: buffer_load_dword v41, off, s[0:3], s33 offset:4
; GFX10-NEXT: s_waitcnt_depctr 0xffe3
; GFX10-NEXT: s_mov_b32 exec_lo, s35
; GFX10-NEXT: s_addk_i32 s32, 0xfe00
@@ -549,23 +589,27 @@
; GFX11-NEXT: s_mov_b32 s0, s33
; GFX11-NEXT: s_mov_b32 s33, s32
; GFX11-NEXT: s_or_saveexec_b32 s1, -1
-; GFX11-NEXT: scratch_store_b32 off, v40, s33 ; 4-byte Folded Spill
+; GFX11-NEXT: s_clause 0x1
+; GFX11-NEXT: scratch_store_b32 off, v40, s33
+; GFX11-NEXT: scratch_store_b32 off, v41, s33 offset:4
; GFX11-NEXT: s_mov_b32 exec_lo, s1
-; GFX11-NEXT: v_writelane_b32 v40, s0, 2
+; GFX11-NEXT: v_writelane_b32 v40, s30, 0
; GFX11-NEXT: v_mov_b32_e32 v0, 0x7b
; GFX11-NEXT: s_add_i32 s32, s32, 16
+; GFX11-NEXT: v_writelane_b32 v41, s0, 0
; GFX11-NEXT: s_getpc_b64 s[0:1]
; GFX11-NEXT: s_add_u32 s0, s0, external_void_func_i8@rel32@lo+4
; GFX11-NEXT: s_addc_u32 s1, s1, external_void_func_i8@rel32@hi+12
-; GFX11-NEXT: v_writelane_b32 v40, s30, 0
; GFX11-NEXT: v_writelane_b32 v40, s31, 1
; GFX11-NEXT: s_swappc_b64 s[30:31], s[0:1]
; GFX11-NEXT: s_delay_alu instid0(VALU_DEP_1)
; GFX11-NEXT: v_readlane_b32 s31, v40, 1
; GFX11-NEXT: v_readlane_b32 s30, v40, 0
-; GFX11-NEXT: v_readlane_b32 s0, v40, 2
+; GFX11-NEXT: v_readlane_b32 s0, v41, 0
; GFX11-NEXT: s_or_saveexec_b32 s1, -1
-; GFX11-NEXT: scratch_load_b32 v40, off, s33 ; 4-byte Folded Reload
+; GFX11-NEXT: s_clause 0x1
+; GFX11-NEXT: scratch_load_b32 v40, off, s33
+; GFX11-NEXT: scratch_load_b32 v41, off, s33 offset:4
; GFX11-NEXT: s_mov_b32 exec_lo, s1
; GFX11-NEXT: s_add_i32 s32, s32, -16
; GFX11-NEXT: s_mov_b32 s33, s0
@@ -579,22 +623,25 @@
; GFX10-SCRATCH-NEXT: s_mov_b32 s33, s32
; GFX10-SCRATCH-NEXT: s_or_saveexec_b32 s1, -1
; GFX10-SCRATCH-NEXT: scratch_store_dword off, v40, s33 ; 4-byte Folded Spill
+; GFX10-SCRATCH-NEXT: scratch_store_dword off, v41, s33 offset:4 ; 4-byte Folded Spill
; GFX10-SCRATCH-NEXT: s_waitcnt_depctr 0xffe3
; GFX10-SCRATCH-NEXT: s_mov_b32 exec_lo, s1
-; GFX10-SCRATCH-NEXT: v_writelane_b32 v40, s0, 2
+; GFX10-SCRATCH-NEXT: v_writelane_b32 v40, s30, 0
; GFX10-SCRATCH-NEXT: v_mov_b32_e32 v0, 0x7b
; GFX10-SCRATCH-NEXT: s_add_i32 s32, s32, 16
+; GFX10-SCRATCH-NEXT: v_writelane_b32 v41, s0, 0
; GFX10-SCRATCH-NEXT: s_getpc_b64 s[0:1]
; GFX10-SCRATCH-NEXT: s_add_u32 s0, s0, external_void_func_i8@rel32@lo+4
; GFX10-SCRATCH-NEXT: s_addc_u32 s1, s1, external_void_func_i8@rel32@hi+12
-; GFX10-SCRATCH-NEXT: v_writelane_b32 v40, s30, 0
; GFX10-SCRATCH-NEXT: v_writelane_b32 v40, s31, 1
; GFX10-SCRATCH-NEXT: s_swappc_b64 s[30:31], s[0:1]
; GFX10-SCRATCH-NEXT: v_readlane_b32 s31, v40, 1
; GFX10-SCRATCH-NEXT: v_readlane_b32 s30, v40, 0
-; GFX10-SCRATCH-NEXT: v_readlane_b32 s0, v40, 2
+; GFX10-SCRATCH-NEXT: v_readlane_b32 s0, v41, 0
; GFX10-SCRATCH-NEXT: s_or_saveexec_b32 s1, -1
-; GFX10-SCRATCH-NEXT: scratch_load_dword v40, off, s33 ; 4-byte Folded Reload
+; GFX10-SCRATCH-NEXT: s_clause 0x1
+; GFX10-SCRATCH-NEXT: scratch_load_dword v40, off, s33
+; GFX10-SCRATCH-NEXT: scratch_load_dword v41, off, s33 offset:4
; GFX10-SCRATCH-NEXT: s_waitcnt_depctr 0xffe3
; GFX10-SCRATCH-NEXT: s_mov_b32 exec_lo, s1
; GFX10-SCRATCH-NEXT: s_add_i32 s32, s32, -16
@@ -613,12 +660,13 @@
; GFX9-NEXT: s_mov_b32 s33, s32
; GFX9-NEXT: s_or_saveexec_b64 s[36:37], -1
; GFX9-NEXT: buffer_store_dword v40, off, s[0:3], s33 ; 4-byte Folded Spill
+; GFX9-NEXT: buffer_store_dword v41, off, s[0:3], s33 offset:4 ; 4-byte Folded Spill
; GFX9-NEXT: s_mov_b64 exec, s[36:37]
; GFX9-NEXT: global_load_sbyte v0, v[0:1], off glc
; GFX9-NEXT: s_waitcnt vmcnt(0)
-; GFX9-NEXT: v_writelane_b32 v40, s34, 2
; GFX9-NEXT: s_addk_i32 s32, 0x400
; GFX9-NEXT: v_writelane_b32 v40, s30, 0
+; GFX9-NEXT: v_writelane_b32 v41, s34, 0
; GFX9-NEXT: v_writelane_b32 v40, s31, 1
; GFX9-NEXT: s_getpc_b64 s[34:35]
; GFX9-NEXT: s_add_u32 s34, s34, external_void_func_i8_signext@rel32@lo+4
@@ -626,9 +674,10 @@
; GFX9-NEXT: s_swappc_b64 s[30:31], s[34:35]
; GFX9-NEXT: v_readlane_b32 s31, v40, 1
; GFX9-NEXT: v_readlane_b32 s30, v40, 0
-; GFX9-NEXT: v_readlane_b32 s34, v40, 2
+; GFX9-NEXT: v_readlane_b32 s34, v41, 0
; GFX9-NEXT: s_or_saveexec_b64 s[36:37], -1
; GFX9-NEXT: buffer_load_dword v40, off, s[0:3], s33 ; 4-byte Folded Reload
+; GFX9-NEXT: buffer_load_dword v41, off, s[0:3], s33 offset:4 ; 4-byte Folded Reload
; GFX9-NEXT: s_mov_b64 exec, s[36:37]
; GFX9-NEXT: s_addk_i32 s32, 0xfc00
; GFX9-NEXT: s_mov_b32 s33, s34
@@ -642,23 +691,26 @@
; GFX10-NEXT: s_mov_b32 s33, s32
; GFX10-NEXT: s_or_saveexec_b32 s35, -1
; GFX10-NEXT: buffer_store_dword v40, off, s[0:3], s33 ; 4-byte Folded Spill
+; GFX10-NEXT: buffer_store_dword v41, off, s[0:3], s33 offset:4 ; 4-byte Folded Spill
; GFX10-NEXT: s_waitcnt_depctr 0xffe3
; GFX10-NEXT: s_mov_b32 exec_lo, s35
; GFX10-NEXT: global_load_sbyte v0, v[0:1], off glc dlc
; GFX10-NEXT: s_waitcnt vmcnt(0)
-; GFX10-NEXT: v_writelane_b32 v40, s34, 2
+; GFX10-NEXT: v_writelane_b32 v40, s30, 0
; GFX10-NEXT: s_addk_i32 s32, 0x200
+; GFX10-NEXT: v_writelane_b32 v41, s34, 0
; GFX10-NEXT: s_getpc_b64 s[34:35]
; GFX10-NEXT: s_add_u32 s34, s34, external_void_func_i8_signext@rel32@lo+4
; GFX10-NEXT: s_addc_u32 s35, s35, external_void_func_i8_signext@rel32@hi+12
-; GFX10-NEXT: v_writelane_b32 v40, s30, 0
; GFX10-NEXT: v_writelane_b32 v40, s31, 1
; GFX10-NEXT: s_swappc_b64 s[30:31], s[34:35]
; GFX10-NEXT: v_readlane_b32 s31, v40, 1
; GFX10-NEXT: v_readlane_b32 s30, v40, 0
-; GFX10-NEXT: v_readlane_b32 s34, v40, 2
+; GFX10-NEXT: v_readlane_b32 s34, v41, 0
; GFX10-NEXT: s_or_saveexec_b32 s35, -1
-; GFX10-NEXT: buffer_load_dword v40, off, s[0:3], s33 ; 4-byte Folded Reload
+; GFX10-NEXT: s_clause 0x1
+; GFX10-NEXT: buffer_load_dword v40, off, s[0:3], s33
+; GFX10-NEXT: buffer_load_dword v41, off, s[0:3], s33 offset:4
; GFX10-NEXT: s_waitcnt_depctr 0xffe3
; GFX10-NEXT: s_mov_b32 exec_lo, s35
; GFX10-NEXT: s_addk_i32 s32, 0xfe00
@@ -672,24 +724,28 @@
; GFX11-NEXT: s_mov_b32 s0, s33
; GFX11-NEXT: s_mov_b32 s33, s32
; GFX11-NEXT: s_or_saveexec_b32 s1, -1
-; GFX11-NEXT: scratch_store_b32 off, v40, s33 ; 4-byte Folded Spill
+; GFX11-NEXT: s_clause 0x1
+; GFX11-NEXT: scratch_store_b32 off, v40, s33
+; GFX11-NEXT: scratch_store_b32 off, v41, s33 offset:4
; GFX11-NEXT: s_mov_b32 exec_lo, s1
; GFX11-NEXT: global_load_i8 v0, v[0:1], off glc dlc
; GFX11-NEXT: s_waitcnt vmcnt(0)
-; GFX11-NEXT: v_writelane_b32 v40, s0, 2
+; GFX11-NEXT: v_writelane_b32 v40, s30, 0
; GFX11-NEXT: s_add_i32 s32, s32, 16
+; GFX11-NEXT: v_writelane_b32 v41, s0, 0
; GFX11-NEXT: s_getpc_b64 s[0:1]
; GFX11-NEXT: s_add_u32 s0, s0, external_void_func_i8_signext@rel32@lo+4
; GFX11-NEXT: s_addc_u32 s1, s1, external_void_func_i8_signext@rel32@hi+12
-; GFX11-NEXT: v_writelane_b32 v40, s30, 0
; GFX11-NEXT: v_writelane_b32 v40, s31, 1
; GFX11-NEXT: s_swappc_b64 s[30:31], s[0:1]
; GFX11-NEXT: s_delay_alu instid0(VALU_DEP_1)
; GFX11-NEXT: v_readlane_b32 s31, v40, 1
; GFX11-NEXT: v_readlane_b32 s30, v40, 0
-; GFX11-NEXT: v_readlane_b32 s0, v40, 2
+; GFX11-NEXT: v_readlane_b32 s0, v41, 0
; GFX11-NEXT: s_or_saveexec_b32 s1, -1
-; GFX11-NEXT: scratch_load_b32 v40, off, s33 ; 4-byte Folded Reload
+; GFX11-NEXT: s_clause 0x1
+; GFX11-NEXT: scratch_load_b32 v40, off, s33
+; GFX11-NEXT: scratch_load_b32 v41, off, s33 offset:4
; GFX11-NEXT: s_mov_b32 exec_lo, s1
; GFX11-NEXT: s_add_i32 s32, s32, -16
; GFX11-NEXT: s_mov_b32 s33, s0
@@ -703,23 +759,26 @@
; GFX10-SCRATCH-NEXT: s_mov_b32 s33, s32
; GFX10-SCRATCH-NEXT: s_or_saveexec_b32 s1, -1
; GFX10-SCRATCH-NEXT: scratch_store_dword off, v40, s33 ; 4-byte Folded Spill
+; GFX10-SCRATCH-NEXT: scratch_store_dword off, v41, s33 offset:4 ; 4-byte Folded Spill
; GFX10-SCRATCH-NEXT: s_waitcnt_depctr 0xffe3
; GFX10-SCRATCH-NEXT: s_mov_b32 exec_lo, s1
; GFX10-SCRATCH-NEXT: global_load_sbyte v0, v[0:1], off glc dlc
; GFX10-SCRATCH-NEXT: s_waitcnt vmcnt(0)
-; GFX10-SCRATCH-NEXT: v_writelane_b32 v40, s0, 2
+; GFX10-SCRATCH-NEXT: v_writelane_b32 v40, s30, 0
; GFX10-SCRATCH-NEXT: s_add_i32 s32, s32, 16
+; GFX10-SCRATCH-NEXT: v_writelane_b32 v41, s0, 0
; GFX10-SCRATCH-NEXT: s_getpc_b64 s[0:1]
; GFX10-SCRATCH-NEXT: s_add_u32 s0, s0, external_void_func_i8_signext@rel32@lo+4
; GFX10-SCRATCH-NEXT: s_addc_u32 s1, s1, external_void_func_i8_signext@rel32@hi+12
-; GFX10-SCRATCH-NEXT: v_writelane_b32 v40, s30, 0
; GFX10-SCRATCH-NEXT: v_writelane_b32 v40, s31, 1
; GFX10-SCRATCH-NEXT: s_swappc_b64 s[30:31], s[0:1]
; GFX10-SCRATCH-NEXT: v_readlane_b32 s31, v40, 1
; GFX10-SCRATCH-NEXT: v_readlane_b32 s30, v40, 0
-; GFX10-SCRATCH-NEXT: v_readlane_b32 s0, v40, 2
+; GFX10-SCRATCH-NEXT: v_readlane_b32 s0, v41, 0
; GFX10-SCRATCH-NEXT: s_or_saveexec_b32 s1, -1
-; GFX10-SCRATCH-NEXT: scratch_load_dword v40, off, s33 ; 4-byte Folded Reload
+; GFX10-SCRATCH-NEXT: s_clause 0x1
+; GFX10-SCRATCH-NEXT: scratch_load_dword v40, off, s33
+; GFX10-SCRATCH-NEXT: scratch_load_dword v41, off, s33 offset:4
; GFX10-SCRATCH-NEXT: s_waitcnt_depctr 0xffe3
; GFX10-SCRATCH-NEXT: s_mov_b32 exec_lo, s1
; GFX10-SCRATCH-NEXT: s_add_i32 s32, s32, -16
@@ -739,12 +798,13 @@
; GFX9-NEXT: s_mov_b32 s33, s32
; GFX9-NEXT: s_or_saveexec_b64 s[36:37], -1
; GFX9-NEXT: buffer_store_dword v40, off, s[0:3], s33 ; 4-byte Folded Spill
+; GFX9-NEXT: buffer_store_dword v41, off, s[0:3], s33 offset:4 ; 4-byte Folded Spill
; GFX9-NEXT: s_mov_b64 exec, s[36:37]
; GFX9-NEXT: global_load_ubyte v0, v[0:1], off glc
; GFX9-NEXT: s_waitcnt vmcnt(0)
-; GFX9-NEXT: v_writelane_b32 v40, s34, 2
; GFX9-NEXT: s_addk_i32 s32, 0x400
; GFX9-NEXT: v_writelane_b32 v40, s30, 0
+; GFX9-NEXT: v_writelane_b32 v41, s34, 0
; GFX9-NEXT: v_writelane_b32 v40, s31, 1
; GFX9-NEXT: s_getpc_b64 s[34:35]
; GFX9-NEXT: s_add_u32 s34, s34, external_void_func_i8_zeroext@rel32@lo+4
@@ -752,9 +812,10 @@
; GFX9-NEXT: s_swappc_b64 s[30:31], s[34:35]
; GFX9-NEXT: v_readlane_b32 s31, v40, 1
; GFX9-NEXT: v_readlane_b32 s30, v40, 0
-; GFX9-NEXT: v_readlane_b32 s34, v40, 2
+; GFX9-NEXT: v_readlane_b32 s34, v41, 0
; GFX9-NEXT: s_or_saveexec_b64 s[36:37], -1
; GFX9-NEXT: buffer_load_dword v40, off, s[0:3], s33 ; 4-byte Folded Reload
+; GFX9-NEXT: buffer_load_dword v41, off, s[0:3], s33 offset:4 ; 4-byte Folded Reload
; GFX9-NEXT: s_mov_b64 exec, s[36:37]
; GFX9-NEXT: s_addk_i32 s32, 0xfc00
; GFX9-NEXT: s_mov_b32 s33, s34
@@ -768,23 +829,26 @@
; GFX10-NEXT: s_mov_b32 s33, s32
; GFX10-NEXT: s_or_saveexec_b32 s35, -1
; GFX10-NEXT: buffer_store_dword v40, off, s[0:3], s33 ; 4-byte Folded Spill
+; GFX10-NEXT: buffer_store_dword v41, off, s[0:3], s33 offset:4 ; 4-byte Folded Spill
; GFX10-NEXT: s_waitcnt_depctr 0xffe3
; GFX10-NEXT: s_mov_b32 exec_lo, s35
; GFX10-NEXT: global_load_ubyte v0, v[0:1], off glc dlc
; GFX10-NEXT: s_waitcnt vmcnt(0)
-; GFX10-NEXT: v_writelane_b32 v40, s34, 2
+; GFX10-NEXT: v_writelane_b32 v40, s30, 0
; GFX10-NEXT: s_addk_i32 s32, 0x200
+; GFX10-NEXT: v_writelane_b32 v41, s34, 0
; GFX10-NEXT: s_getpc_b64 s[34:35]
; GFX10-NEXT: s_add_u32 s34, s34, external_void_func_i8_zeroext@rel32@lo+4
; GFX10-NEXT: s_addc_u32 s35, s35, external_void_func_i8_zeroext@rel32@hi+12
-; GFX10-NEXT: v_writelane_b32 v40, s30, 0
; GFX10-NEXT: v_writelane_b32 v40, s31, 1
; GFX10-NEXT: s_swappc_b64 s[30:31], s[34:35]
; GFX10-NEXT: v_readlane_b32 s31, v40, 1
; GFX10-NEXT: v_readlane_b32 s30, v40, 0
-; GFX10-NEXT: v_readlane_b32 s34, v40, 2
+; GFX10-NEXT: v_readlane_b32 s34, v41, 0
; GFX10-NEXT: s_or_saveexec_b32 s35, -1
-; GFX10-NEXT: buffer_load_dword v40, off, s[0:3], s33 ; 4-byte Folded Reload
+; GFX10-NEXT: s_clause 0x1
+; GFX10-NEXT: buffer_load_dword v40, off, s[0:3], s33
+; GFX10-NEXT: buffer_load_dword v41, off, s[0:3], s33 offset:4
; GFX10-NEXT: s_waitcnt_depctr 0xffe3
; GFX10-NEXT: s_mov_b32 exec_lo, s35
; GFX10-NEXT: s_addk_i32 s32, 0xfe00
@@ -798,24 +862,28 @@
; GFX11-NEXT: s_mov_b32 s0, s33
; GFX11-NEXT: s_mov_b32 s33, s32
; GFX11-NEXT: s_or_saveexec_b32 s1, -1
-; GFX11-NEXT: scratch_store_b32 off, v40, s33 ; 4-byte Folded Spill
+; GFX11-NEXT: s_clause 0x1
+; GFX11-NEXT: scratch_store_b32 off, v40, s33
+; GFX11-NEXT: scratch_store_b32 off, v41, s33 offset:4
; GFX11-NEXT: s_mov_b32 exec_lo, s1
; GFX11-NEXT: global_load_u8 v0, v[0:1], off glc dlc
; GFX11-NEXT: s_waitcnt vmcnt(0)
-; GFX11-NEXT: v_writelane_b32 v40, s0, 2
+; GFX11-NEXT: v_writelane_b32 v40, s30, 0
; GFX11-NEXT: s_add_i32 s32, s32, 16
+; GFX11-NEXT: v_writelane_b32 v41, s0, 0
; GFX11-NEXT: s_getpc_b64 s[0:1]
; GFX11-NEXT: s_add_u32 s0, s0, external_void_func_i8_zeroext@rel32@lo+4
; GFX11-NEXT: s_addc_u32 s1, s1, external_void_func_i8_zeroext@rel32@hi+12
-; GFX11-NEXT: v_writelane_b32 v40, s30, 0
; GFX11-NEXT: v_writelane_b32 v40, s31, 1
; GFX11-NEXT: s_swappc_b64 s[30:31], s[0:1]
; GFX11-NEXT: s_delay_alu instid0(VALU_DEP_1)
; GFX11-NEXT: v_readlane_b32 s31, v40, 1
; GFX11-NEXT: v_readlane_b32 s30, v40, 0
-; GFX11-NEXT: v_readlane_b32 s0, v40, 2
+; GFX11-NEXT: v_readlane_b32 s0, v41, 0
; GFX11-NEXT: s_or_saveexec_b32 s1, -1
-; GFX11-NEXT: scratch_load_b32 v40, off, s33 ; 4-byte Folded Reload
+; GFX11-NEXT: s_clause 0x1
+; GFX11-NEXT: scratch_load_b32 v40, off, s33
+; GFX11-NEXT: scratch_load_b32 v41, off, s33 offset:4
; GFX11-NEXT: s_mov_b32 exec_lo, s1
; GFX11-NEXT: s_add_i32 s32, s32, -16
; GFX11-NEXT: s_mov_b32 s33, s0
@@ -829,23 +897,26 @@
; GFX10-SCRATCH-NEXT: s_mov_b32 s33, s32
; GFX10-SCRATCH-NEXT: s_or_saveexec_b32 s1, -1
; GFX10-SCRATCH-NEXT: scratch_store_dword off, v40, s33 ; 4-byte Folded Spill
+; GFX10-SCRATCH-NEXT: scratch_store_dword off, v41, s33 offset:4 ; 4-byte Folded Spill
; GFX10-SCRATCH-NEXT: s_waitcnt_depctr 0xffe3
; GFX10-SCRATCH-NEXT: s_mov_b32 exec_lo, s1
; GFX10-SCRATCH-NEXT: global_load_ubyte v0, v[0:1], off glc dlc
; GFX10-SCRATCH-NEXT: s_waitcnt vmcnt(0)
-; GFX10-SCRATCH-NEXT: v_writelane_b32 v40, s0, 2
+; GFX10-SCRATCH-NEXT: v_writelane_b32 v40, s30, 0
; GFX10-SCRATCH-NEXT: s_add_i32 s32, s32, 16
+; GFX10-SCRATCH-NEXT: v_writelane_b32 v41, s0, 0
; GFX10-SCRATCH-NEXT: s_getpc_b64 s[0:1]
; GFX10-SCRATCH-NEXT: s_add_u32 s0, s0, external_void_func_i8_zeroext@rel32@lo+4
; GFX10-SCRATCH-NEXT: s_addc_u32 s1, s1, external_void_func_i8_zeroext@rel32@hi+12
-; GFX10-SCRATCH-NEXT: v_writelane_b32 v40, s30, 0
; GFX10-SCRATCH-NEXT: v_writelane_b32 v40, s31, 1
; GFX10-SCRATCH-NEXT: s_swappc_b64 s[30:31], s[0:1]
; GFX10-SCRATCH-NEXT: v_readlane_b32 s31, v40, 1
; GFX10-SCRATCH-NEXT: v_readlane_b32 s30, v40, 0
-; GFX10-SCRATCH-NEXT: v_readlane_b32 s0, v40, 2
+; GFX10-SCRATCH-NEXT: v_readlane_b32 s0, v41, 0
; GFX10-SCRATCH-NEXT: s_or_saveexec_b32 s1, -1
-; GFX10-SCRATCH-NEXT: scratch_load_dword v40, off, s33 ; 4-byte Folded Reload
+; GFX10-SCRATCH-NEXT: s_clause 0x1
+; GFX10-SCRATCH-NEXT: scratch_load_dword v40, off, s33
+; GFX10-SCRATCH-NEXT: scratch_load_dword v41, off, s33 offset:4
; GFX10-SCRATCH-NEXT: s_waitcnt_depctr 0xffe3
; GFX10-SCRATCH-NEXT: s_mov_b32 exec_lo, s1
; GFX10-SCRATCH-NEXT: s_add_i32 s32, s32, -16
@@ -865,11 +936,12 @@
; GFX9-NEXT: s_mov_b32 s33, s32
; GFX9-NEXT: s_or_saveexec_b64 s[36:37], -1
; GFX9-NEXT: buffer_store_dword v40, off, s[0:3], s33 ; 4-byte Folded Spill
+; GFX9-NEXT: buffer_store_dword v41, off, s[0:3], s33 offset:4 ; 4-byte Folded Spill
; GFX9-NEXT: s_mov_b64 exec, s[36:37]
-; GFX9-NEXT: v_writelane_b32 v40, s34, 2
; GFX9-NEXT: s_addk_i32 s32, 0x400
; GFX9-NEXT: v_writelane_b32 v40, s30, 0
; GFX9-NEXT: v_mov_b32_e32 v0, 0x7b
+; GFX9-NEXT: v_writelane_b32 v41, s34, 0
; GFX9-NEXT: v_writelane_b32 v40, s31, 1
; GFX9-NEXT: s_getpc_b64 s[34:35]
; GFX9-NEXT: s_add_u32 s34, s34, external_void_func_i16@rel32@lo+4
@@ -877,9 +949,10 @@
; GFX9-NEXT: s_swappc_b64 s[30:31], s[34:35]
; GFX9-NEXT: v_readlane_b32 s31, v40, 1
; GFX9-NEXT: v_readlane_b32 s30, v40, 0
-; GFX9-NEXT: v_readlane_b32 s34, v40, 2
+; GFX9-NEXT: v_readlane_b32 s34, v41, 0
; GFX9-NEXT: s_or_saveexec_b64 s[36:37], -1
; GFX9-NEXT: buffer_load_dword v40, off, s[0:3], s33 ; 4-byte Folded Reload
+; GFX9-NEXT: buffer_load_dword v41, off, s[0:3], s33 offset:4 ; 4-byte Folded Reload
; GFX9-NEXT: s_mov_b64 exec, s[36:37]
; GFX9-NEXT: s_addk_i32 s32, 0xfc00
; GFX9-NEXT: s_mov_b32 s33, s34
@@ -893,22 +966,25 @@
; GFX10-NEXT: s_mov_b32 s33, s32
; GFX10-NEXT: s_or_saveexec_b32 s35, -1
; GFX10-NEXT: buffer_store_dword v40, off, s[0:3], s33 ; 4-byte Folded Spill
+; GFX10-NEXT: buffer_store_dword v41, off, s[0:3], s33 offset:4 ; 4-byte Folded Spill
; GFX10-NEXT: s_waitcnt_depctr 0xffe3
; GFX10-NEXT: s_mov_b32 exec_lo, s35
-; GFX10-NEXT: v_writelane_b32 v40, s34, 2
+; GFX10-NEXT: v_writelane_b32 v40, s30, 0
; GFX10-NEXT: v_mov_b32_e32 v0, 0x7b
; GFX10-NEXT: s_addk_i32 s32, 0x200
+; GFX10-NEXT: v_writelane_b32 v41, s34, 0
; GFX10-NEXT: s_getpc_b64 s[34:35]
; GFX10-NEXT: s_add_u32 s34, s34, external_void_func_i16@rel32@lo+4
; GFX10-NEXT: s_addc_u32 s35, s35, external_void_func_i16@rel32@hi+12
-; GFX10-NEXT: v_writelane_b32 v40, s30, 0
; GFX10-NEXT: v_writelane_b32 v40, s31, 1
; GFX10-NEXT: s_swappc_b64 s[30:31], s[34:35]
; GFX10-NEXT: v_readlane_b32 s31, v40, 1
; GFX10-NEXT: v_readlane_b32 s30, v40, 0
-; GFX10-NEXT: v_readlane_b32 s34, v40, 2
+; GFX10-NEXT: v_readlane_b32 s34, v41, 0
; GFX10-NEXT: s_or_saveexec_b32 s35, -1
-; GFX10-NEXT: buffer_load_dword v40, off, s[0:3], s33 ; 4-byte Folded Reload
+; GFX10-NEXT: s_clause 0x1
+; GFX10-NEXT: buffer_load_dword v40, off, s[0:3], s33
+; GFX10-NEXT: buffer_load_dword v41, off, s[0:3], s33 offset:4
; GFX10-NEXT: s_waitcnt_depctr 0xffe3
; GFX10-NEXT: s_mov_b32 exec_lo, s35
; GFX10-NEXT: s_addk_i32 s32, 0xfe00
@@ -922,23 +998,27 @@
; GFX11-NEXT: s_mov_b32 s0, s33
; GFX11-NEXT: s_mov_b32 s33, s32
; GFX11-NEXT: s_or_saveexec_b32 s1, -1
-; GFX11-NEXT: scratch_store_b32 off, v40, s33 ; 4-byte Folded Spill
+; GFX11-NEXT: s_clause 0x1
+; GFX11-NEXT: scratch_store_b32 off, v40, s33
+; GFX11-NEXT: scratch_store_b32 off, v41, s33 offset:4
; GFX11-NEXT: s_mov_b32 exec_lo, s1
-; GFX11-NEXT: v_writelane_b32 v40, s0, 2
+; GFX11-NEXT: v_writelane_b32 v40, s30, 0
; GFX11-NEXT: v_mov_b32_e32 v0, 0x7b
; GFX11-NEXT: s_add_i32 s32, s32, 16
+; GFX11-NEXT: v_writelane_b32 v41, s0, 0
; GFX11-NEXT: s_getpc_b64 s[0:1]
; GFX11-NEXT: s_add_u32 s0, s0, external_void_func_i16@rel32@lo+4
; GFX11-NEXT: s_addc_u32 s1, s1, external_void_func_i16@rel32@hi+12
-; GFX11-NEXT: v_writelane_b32 v40, s30, 0
; GFX11-NEXT: v_writelane_b32 v40, s31, 1
; GFX11-NEXT: s_swappc_b64 s[30:31], s[0:1]
; GFX11-NEXT: s_delay_alu instid0(VALU_DEP_1)
; GFX11-NEXT: v_readlane_b32 s31, v40, 1
; GFX11-NEXT: v_readlane_b32 s30, v40, 0
-; GFX11-NEXT: v_readlane_b32 s0, v40, 2
+; GFX11-NEXT: v_readlane_b32 s0, v41, 0
; GFX11-NEXT: s_or_saveexec_b32 s1, -1
-; GFX11-NEXT: scratch_load_b32 v40, off, s33 ; 4-byte Folded Reload
+; GFX11-NEXT: s_clause 0x1
+; GFX11-NEXT: scratch_load_b32 v40, off, s33
+; GFX11-NEXT: scratch_load_b32 v41, off, s33 offset:4
; GFX11-NEXT: s_mov_b32 exec_lo, s1
; GFX11-NEXT: s_add_i32 s32, s32, -16
; GFX11-NEXT: s_mov_b32 s33, s0
@@ -952,22 +1032,25 @@
; GFX10-SCRATCH-NEXT: s_mov_b32 s33, s32
; GFX10-SCRATCH-NEXT: s_or_saveexec_b32 s1, -1
; GFX10-SCRATCH-NEXT: scratch_store_dword off, v40, s33 ; 4-byte Folded Spill
+; GFX10-SCRATCH-NEXT: scratch_store_dword off, v41, s33 offset:4 ; 4-byte Folded Spill
; GFX10-SCRATCH-NEXT: s_waitcnt_depctr 0xffe3
; GFX10-SCRATCH-NEXT: s_mov_b32 exec_lo, s1
-; GFX10-SCRATCH-NEXT: v_writelane_b32 v40, s0, 2
+; GFX10-SCRATCH-NEXT: v_writelane_b32 v40, s30, 0
; GFX10-SCRATCH-NEXT: v_mov_b32_e32 v0, 0x7b
; GFX10-SCRATCH-NEXT: s_add_i32 s32, s32, 16
+; GFX10-SCRATCH-NEXT: v_writelane_b32 v41, s0, 0
; GFX10-SCRATCH-NEXT: s_getpc_b64 s[0:1]
; GFX10-SCRATCH-NEXT: s_add_u32 s0, s0, external_void_func_i16@rel32@lo+4
; GFX10-SCRATCH-NEXT: s_addc_u32 s1, s1, external_void_func_i16@rel32@hi+12
-; GFX10-SCRATCH-NEXT: v_writelane_b32 v40, s30, 0
; GFX10-SCRATCH-NEXT: v_writelane_b32 v40, s31, 1
; GFX10-SCRATCH-NEXT: s_swappc_b64 s[30:31], s[0:1]
; GFX10-SCRATCH-NEXT: v_readlane_b32 s31, v40, 1
; GFX10-SCRATCH-NEXT: v_readlane_b32 s30, v40, 0
-; GFX10-SCRATCH-NEXT: v_readlane_b32 s0, v40, 2
+; GFX10-SCRATCH-NEXT: v_readlane_b32 s0, v41, 0
; GFX10-SCRATCH-NEXT: s_or_saveexec_b32 s1, -1
-; GFX10-SCRATCH-NEXT: scratch_load_dword v40, off, s33 ; 4-byte Folded Reload
+; GFX10-SCRATCH-NEXT: s_clause 0x1
+; GFX10-SCRATCH-NEXT: scratch_load_dword v40, off, s33
+; GFX10-SCRATCH-NEXT: scratch_load_dword v41, off, s33 offset:4
; GFX10-SCRATCH-NEXT: s_waitcnt_depctr 0xffe3
; GFX10-SCRATCH-NEXT: s_mov_b32 exec_lo, s1
; GFX10-SCRATCH-NEXT: s_add_i32 s32, s32, -16
@@ -986,12 +1069,13 @@
; GFX9-NEXT: s_mov_b32 s33, s32
; GFX9-NEXT: s_or_saveexec_b64 s[36:37], -1
; GFX9-NEXT: buffer_store_dword v40, off, s[0:3], s33 ; 4-byte Folded Spill
+; GFX9-NEXT: buffer_store_dword v41, off, s[0:3], s33 offset:4 ; 4-byte Folded Spill
; GFX9-NEXT: s_mov_b64 exec, s[36:37]
; GFX9-NEXT: global_load_ushort v0, v[0:1], off glc
; GFX9-NEXT: s_waitcnt vmcnt(0)
-; GFX9-NEXT: v_writelane_b32 v40, s34, 2
; GFX9-NEXT: s_addk_i32 s32, 0x400
; GFX9-NEXT: v_writelane_b32 v40, s30, 0
+; GFX9-NEXT: v_writelane_b32 v41, s34, 0
; GFX9-NEXT: v_writelane_b32 v40, s31, 1
; GFX9-NEXT: s_getpc_b64 s[34:35]
; GFX9-NEXT: s_add_u32 s34, s34, external_void_func_i16_signext@rel32@lo+4
@@ -999,9 +1083,10 @@
; GFX9-NEXT: s_swappc_b64 s[30:31], s[34:35]
; GFX9-NEXT: v_readlane_b32 s31, v40, 1
; GFX9-NEXT: v_readlane_b32 s30, v40, 0
-; GFX9-NEXT: v_readlane_b32 s34, v40, 2
+; GFX9-NEXT: v_readlane_b32 s34, v41, 0
; GFX9-NEXT: s_or_saveexec_b64 s[36:37], -1
; GFX9-NEXT: buffer_load_dword v40, off, s[0:3], s33 ; 4-byte Folded Reload
+; GFX9-NEXT: buffer_load_dword v41, off, s[0:3], s33 offset:4 ; 4-byte Folded Reload
; GFX9-NEXT: s_mov_b64 exec, s[36:37]
; GFX9-NEXT: s_addk_i32 s32, 0xfc00
; GFX9-NEXT: s_mov_b32 s33, s34
@@ -1015,23 +1100,26 @@
; GFX10-NEXT: s_mov_b32 s33, s32
; GFX10-NEXT: s_or_saveexec_b32 s35, -1
; GFX10-NEXT: buffer_store_dword v40, off, s[0:3], s33 ; 4-byte Folded Spill
+; GFX10-NEXT: buffer_store_dword v41, off, s[0:3], s33 offset:4 ; 4-byte Folded Spill
; GFX10-NEXT: s_waitcnt_depctr 0xffe3
; GFX10-NEXT: s_mov_b32 exec_lo, s35
; GFX10-NEXT: global_load_ushort v0, v[0:1], off glc dlc
; GFX10-NEXT: s_waitcnt vmcnt(0)
-; GFX10-NEXT: v_writelane_b32 v40, s34, 2
+; GFX10-NEXT: v_writelane_b32 v40, s30, 0
; GFX10-NEXT: s_addk_i32 s32, 0x200
+; GFX10-NEXT: v_writelane_b32 v41, s34, 0
; GFX10-NEXT: s_getpc_b64 s[34:35]
; GFX10-NEXT: s_add_u32 s34, s34, external_void_func_i16_signext@rel32@lo+4
; GFX10-NEXT: s_addc_u32 s35, s35, external_void_func_i16_signext@rel32@hi+12
-; GFX10-NEXT: v_writelane_b32 v40, s30, 0
; GFX10-NEXT: v_writelane_b32 v40, s31, 1
; GFX10-NEXT: s_swappc_b64 s[30:31], s[34:35]
; GFX10-NEXT: v_readlane_b32 s31, v40, 1
; GFX10-NEXT: v_readlane_b32 s30, v40, 0
-; GFX10-NEXT: v_readlane_b32 s34, v40, 2
+; GFX10-NEXT: v_readlane_b32 s34, v41, 0
; GFX10-NEXT: s_or_saveexec_b32 s35, -1
-; GFX10-NEXT: buffer_load_dword v40, off, s[0:3], s33 ; 4-byte Folded Reload
+; GFX10-NEXT: s_clause 0x1
+; GFX10-NEXT: buffer_load_dword v40, off, s[0:3], s33
+; GFX10-NEXT: buffer_load_dword v41, off, s[0:3], s33 offset:4
; GFX10-NEXT: s_waitcnt_depctr 0xffe3
; GFX10-NEXT: s_mov_b32 exec_lo, s35
; GFX10-NEXT: s_addk_i32 s32, 0xfe00
@@ -1045,24 +1133,28 @@
; GFX11-NEXT: s_mov_b32 s0, s33
; GFX11-NEXT: s_mov_b32 s33, s32
; GFX11-NEXT: s_or_saveexec_b32 s1, -1
-; GFX11-NEXT: scratch_store_b32 off, v40, s33 ; 4-byte Folded Spill
+; GFX11-NEXT: s_clause 0x1
+; GFX11-NEXT: scratch_store_b32 off, v40, s33
+; GFX11-NEXT: scratch_store_b32 off, v41, s33 offset:4
; GFX11-NEXT: s_mov_b32 exec_lo, s1
; GFX11-NEXT: global_load_u16 v0, v[0:1], off glc dlc
; GFX11-NEXT: s_waitcnt vmcnt(0)
-; GFX11-NEXT: v_writelane_b32 v40, s0, 2
+; GFX11-NEXT: v_writelane_b32 v40, s30, 0
; GFX11-NEXT: s_add_i32 s32, s32, 16
+; GFX11-NEXT: v_writelane_b32 v41, s0, 0
; GFX11-NEXT: s_getpc_b64 s[0:1]
; GFX11-NEXT: s_add_u32 s0, s0, external_void_func_i16_signext@rel32@lo+4
; GFX11-NEXT: s_addc_u32 s1, s1, external_void_func_i16_signext@rel32@hi+12
-; GFX11-NEXT: v_writelane_b32 v40, s30, 0
; GFX11-NEXT: v_writelane_b32 v40, s31, 1
; GFX11-NEXT: s_swappc_b64 s[30:31], s[0:1]
; GFX11-NEXT: s_delay_alu instid0(VALU_DEP_1)
; GFX11-NEXT: v_readlane_b32 s31, v40, 1
; GFX11-NEXT: v_readlane_b32 s30, v40, 0
-; GFX11-NEXT: v_readlane_b32 s0, v40, 2
+; GFX11-NEXT: v_readlane_b32 s0, v41, 0
; GFX11-NEXT: s_or_saveexec_b32 s1, -1
-; GFX11-NEXT: scratch_load_b32 v40, off, s33 ; 4-byte Folded Reload
+; GFX11-NEXT: s_clause 0x1
+; GFX11-NEXT: scratch_load_b32 v40, off, s33
+; GFX11-NEXT: scratch_load_b32 v41, off, s33 offset:4
; GFX11-NEXT: s_mov_b32 exec_lo, s1
; GFX11-NEXT: s_add_i32 s32, s32, -16
; GFX11-NEXT: s_mov_b32 s33, s0
@@ -1076,23 +1168,26 @@
; GFX10-SCRATCH-NEXT: s_mov_b32 s33, s32
; GFX10-SCRATCH-NEXT: s_or_saveexec_b32 s1, -1
; GFX10-SCRATCH-NEXT: scratch_store_dword off, v40, s33 ; 4-byte Folded Spill
+; GFX10-SCRATCH-NEXT: scratch_store_dword off, v41, s33 offset:4 ; 4-byte Folded Spill
; GFX10-SCRATCH-NEXT: s_waitcnt_depctr 0xffe3
; GFX10-SCRATCH-NEXT: s_mov_b32 exec_lo, s1
; GFX10-SCRATCH-NEXT: global_load_ushort v0, v[0:1], off glc dlc
; GFX10-SCRATCH-NEXT: s_waitcnt vmcnt(0)
-; GFX10-SCRATCH-NEXT: v_writelane_b32 v40, s0, 2
+; GFX10-SCRATCH-NEXT: v_writelane_b32 v40, s30, 0
; GFX10-SCRATCH-NEXT: s_add_i32 s32, s32, 16
+; GFX10-SCRATCH-NEXT: v_writelane_b32 v41, s0, 0
; GFX10-SCRATCH-NEXT: s_getpc_b64 s[0:1]
; GFX10-SCRATCH-NEXT: s_add_u32 s0, s0, external_void_func_i16_signext@rel32@lo+4
; GFX10-SCRATCH-NEXT: s_addc_u32 s1, s1, external_void_func_i16_signext@rel32@hi+12
-; GFX10-SCRATCH-NEXT: v_writelane_b32 v40, s30, 0
; GFX10-SCRATCH-NEXT: v_writelane_b32 v40, s31, 1
; GFX10-SCRATCH-NEXT: s_swappc_b64 s[30:31], s[0:1]
; GFX10-SCRATCH-NEXT: v_readlane_b32 s31, v40, 1
; GFX10-SCRATCH-NEXT: v_readlane_b32 s30, v40, 0
-; GFX10-SCRATCH-NEXT: v_readlane_b32 s0, v40, 2
+; GFX10-SCRATCH-NEXT: v_readlane_b32 s0, v41, 0
; GFX10-SCRATCH-NEXT: s_or_saveexec_b32 s1, -1
-; GFX10-SCRATCH-NEXT: scratch_load_dword v40, off, s33 ; 4-byte Folded Reload
+; GFX10-SCRATCH-NEXT: s_clause 0x1
+; GFX10-SCRATCH-NEXT: scratch_load_dword v40, off, s33
+; GFX10-SCRATCH-NEXT: scratch_load_dword v41, off, s33 offset:4
; GFX10-SCRATCH-NEXT: s_waitcnt_depctr 0xffe3
; GFX10-SCRATCH-NEXT: s_mov_b32 exec_lo, s1
; GFX10-SCRATCH-NEXT: s_add_i32 s32, s32, -16
@@ -1112,12 +1207,13 @@
; GFX9-NEXT: s_mov_b32 s33, s32
; GFX9-NEXT: s_or_saveexec_b64 s[36:37], -1
; GFX9-NEXT: buffer_store_dword v40, off, s[0:3], s33 ; 4-byte Folded Spill
+; GFX9-NEXT: buffer_store_dword v41, off, s[0:3], s33 offset:4 ; 4-byte Folded Spill
; GFX9-NEXT: s_mov_b64 exec, s[36:37]
; GFX9-NEXT: global_load_ushort v0, v[0:1], off glc
; GFX9-NEXT: s_waitcnt vmcnt(0)
-; GFX9-NEXT: v_writelane_b32 v40, s34, 2
; GFX9-NEXT: s_addk_i32 s32, 0x400
; GFX9-NEXT: v_writelane_b32 v40, s30, 0
+; GFX9-NEXT: v_writelane_b32 v41, s34, 0
; GFX9-NEXT: v_writelane_b32 v40, s31, 1
; GFX9-NEXT: s_getpc_b64 s[34:35]
; GFX9-NEXT: s_add_u32 s34, s34, external_void_func_i16_zeroext@rel32@lo+4
@@ -1125,9 +1221,10 @@
; GFX9-NEXT: s_swappc_b64 s[30:31], s[34:35]
; GFX9-NEXT: v_readlane_b32 s31, v40, 1
; GFX9-NEXT: v_readlane_b32 s30, v40, 0
-; GFX9-NEXT: v_readlane_b32 s34, v40, 2
+; GFX9-NEXT: v_readlane_b32 s34, v41, 0
; GFX9-NEXT: s_or_saveexec_b64 s[36:37], -1
; GFX9-NEXT: buffer_load_dword v40, off, s[0:3], s33 ; 4-byte Folded Reload
+; GFX9-NEXT: buffer_load_dword v41, off, s[0:3], s33 offset:4 ; 4-byte Folded Reload
; GFX9-NEXT: s_mov_b64 exec, s[36:37]
; GFX9-NEXT: s_addk_i32 s32, 0xfc00
; GFX9-NEXT: s_mov_b32 s33, s34
@@ -1141,23 +1238,26 @@
; GFX10-NEXT: s_mov_b32 s33, s32
; GFX10-NEXT: s_or_saveexec_b32 s35, -1
; GFX10-NEXT: buffer_store_dword v40, off, s[0:3], s33 ; 4-byte Folded Spill
+; GFX10-NEXT: buffer_store_dword v41, off, s[0:3], s33 offset:4 ; 4-byte Folded Spill
; GFX10-NEXT: s_waitcnt_depctr 0xffe3
; GFX10-NEXT: s_mov_b32 exec_lo, s35
; GFX10-NEXT: global_load_ushort v0, v[0:1], off glc dlc
; GFX10-NEXT: s_waitcnt vmcnt(0)
-; GFX10-NEXT: v_writelane_b32 v40, s34, 2
+; GFX10-NEXT: v_writelane_b32 v40, s30, 0
; GFX10-NEXT: s_addk_i32 s32, 0x200
+; GFX10-NEXT: v_writelane_b32 v41, s34, 0
; GFX10-NEXT: s_getpc_b64 s[34:35]
; GFX10-NEXT: s_add_u32 s34, s34, external_void_func_i16_zeroext@rel32@lo+4
; GFX10-NEXT: s_addc_u32 s35, s35, external_void_func_i16_zeroext@rel32@hi+12
-; GFX10-NEXT: v_writelane_b32 v40, s30, 0
; GFX10-NEXT: v_writelane_b32 v40, s31, 1
; GFX10-NEXT: s_swappc_b64 s[30:31], s[34:35]
; GFX10-NEXT: v_readlane_b32 s31, v40, 1
; GFX10-NEXT: v_readlane_b32 s30, v40, 0
-; GFX10-NEXT: v_readlane_b32 s34, v40, 2
+; GFX10-NEXT: v_readlane_b32 s34, v41, 0
; GFX10-NEXT: s_or_saveexec_b32 s35, -1
-; GFX10-NEXT: buffer_load_dword v40, off, s[0:3], s33 ; 4-byte Folded Reload
+; GFX10-NEXT: s_clause 0x1
+; GFX10-NEXT: buffer_load_dword v40, off, s[0:3], s33
+; GFX10-NEXT: buffer_load_dword v41, off, s[0:3], s33 offset:4
; GFX10-NEXT: s_waitcnt_depctr 0xffe3
; GFX10-NEXT: s_mov_b32 exec_lo, s35
; GFX10-NEXT: s_addk_i32 s32, 0xfe00
@@ -1171,24 +1271,28 @@
; GFX11-NEXT: s_mov_b32 s0, s33
; GFX11-NEXT: s_mov_b32 s33, s32
; GFX11-NEXT: s_or_saveexec_b32 s1, -1
-; GFX11-NEXT: scratch_store_b32 off, v40, s33 ; 4-byte Folded Spill
+; GFX11-NEXT: s_clause 0x1
+; GFX11-NEXT: scratch_store_b32 off, v40, s33
+; GFX11-NEXT: scratch_store_b32 off, v41, s33 offset:4
; GFX11-NEXT: s_mov_b32 exec_lo, s1
; GFX11-NEXT: global_load_u16 v0, v[0:1], off glc dlc
; GFX11-NEXT: s_waitcnt vmcnt(0)
-; GFX11-NEXT: v_writelane_b32 v40, s0, 2
+; GFX11-NEXT: v_writelane_b32 v40, s30, 0
; GFX11-NEXT: s_add_i32 s32, s32, 16
+; GFX11-NEXT: v_writelane_b32 v41, s0, 0
; GFX11-NEXT: s_getpc_b64 s[0:1]
; GFX11-NEXT: s_add_u32 s0, s0, external_void_func_i16_zeroext@rel32@lo+4
; GFX11-NEXT: s_addc_u32 s1, s1, external_void_func_i16_zeroext@rel32@hi+12
-; GFX11-NEXT: v_writelane_b32 v40, s30, 0
; GFX11-NEXT: v_writelane_b32 v40, s31, 1
; GFX11-NEXT: s_swappc_b64 s[30:31], s[0:1]
; GFX11-NEXT: s_delay_alu instid0(VALU_DEP_1)
; GFX11-NEXT: v_readlane_b32 s31, v40, 1
; GFX11-NEXT: v_readlane_b32 s30, v40, 0
-; GFX11-NEXT: v_readlane_b32 s0, v40, 2
+; GFX11-NEXT: v_readlane_b32 s0, v41, 0
; GFX11-NEXT: s_or_saveexec_b32 s1, -1
-; GFX11-NEXT: scratch_load_b32 v40, off, s33 ; 4-byte Folded Reload
+; GFX11-NEXT: s_clause 0x1
+; GFX11-NEXT: scratch_load_b32 v40, off, s33
+; GFX11-NEXT: scratch_load_b32 v41, off, s33 offset:4
; GFX11-NEXT: s_mov_b32 exec_lo, s1
; GFX11-NEXT: s_add_i32 s32, s32, -16
; GFX11-NEXT: s_mov_b32 s33, s0
@@ -1202,23 +1306,26 @@
; GFX10-SCRATCH-NEXT: s_mov_b32 s33, s32
; GFX10-SCRATCH-NEXT: s_or_saveexec_b32 s1, -1
; GFX10-SCRATCH-NEXT: scratch_store_dword off, v40, s33 ; 4-byte Folded Spill
+; GFX10-SCRATCH-NEXT: scratch_store_dword off, v41, s33 offset:4 ; 4-byte Folded Spill
; GFX10-SCRATCH-NEXT: s_waitcnt_depctr 0xffe3
; GFX10-SCRATCH-NEXT: s_mov_b32 exec_lo, s1
; GFX10-SCRATCH-NEXT: global_load_ushort v0, v[0:1], off glc dlc
; GFX10-SCRATCH-NEXT: s_waitcnt vmcnt(0)
-; GFX10-SCRATCH-NEXT: v_writelane_b32 v40, s0, 2
+; GFX10-SCRATCH-NEXT: v_writelane_b32 v40, s30, 0
; GFX10-SCRATCH-NEXT: s_add_i32 s32, s32, 16
+; GFX10-SCRATCH-NEXT: v_writelane_b32 v41, s0, 0
; GFX10-SCRATCH-NEXT: s_getpc_b64 s[0:1]
; GFX10-SCRATCH-NEXT: s_add_u32 s0, s0, external_void_func_i16_zeroext@rel32@lo+4
; GFX10-SCRATCH-NEXT: s_addc_u32 s1, s1, external_void_func_i16_zeroext@rel32@hi+12
-; GFX10-SCRATCH-NEXT: v_writelane_b32 v40, s30, 0
; GFX10-SCRATCH-NEXT: v_writelane_b32 v40, s31, 1
; GFX10-SCRATCH-NEXT: s_swappc_b64 s[30:31], s[0:1]
; GFX10-SCRATCH-NEXT: v_readlane_b32 s31, v40, 1
; GFX10-SCRATCH-NEXT: v_readlane_b32 s30, v40, 0
-; GFX10-SCRATCH-NEXT: v_readlane_b32 s0, v40, 2
+; GFX10-SCRATCH-NEXT: v_readlane_b32 s0, v41, 0
; GFX10-SCRATCH-NEXT: s_or_saveexec_b32 s1, -1
-; GFX10-SCRATCH-NEXT: scratch_load_dword v40, off, s33 ; 4-byte Folded Reload
+; GFX10-SCRATCH-NEXT: s_clause 0x1
+; GFX10-SCRATCH-NEXT: scratch_load_dword v40, off, s33
+; GFX10-SCRATCH-NEXT: scratch_load_dword v41, off, s33 offset:4
; GFX10-SCRATCH-NEXT: s_waitcnt_depctr 0xffe3
; GFX10-SCRATCH-NEXT: s_mov_b32 exec_lo, s1
; GFX10-SCRATCH-NEXT: s_add_i32 s32, s32, -16
@@ -1238,11 +1345,12 @@
; GFX9-NEXT: s_mov_b32 s33, s32
; GFX9-NEXT: s_or_saveexec_b64 s[36:37], -1
; GFX9-NEXT: buffer_store_dword v40, off, s[0:3], s33 ; 4-byte Folded Spill
+; GFX9-NEXT: buffer_store_dword v41, off, s[0:3], s33 offset:4 ; 4-byte Folded Spill
; GFX9-NEXT: s_mov_b64 exec, s[36:37]
-; GFX9-NEXT: v_writelane_b32 v40, s34, 2
; GFX9-NEXT: s_addk_i32 s32, 0x400
; GFX9-NEXT: v_writelane_b32 v40, s30, 0
; GFX9-NEXT: v_mov_b32_e32 v0, 42
+; GFX9-NEXT: v_writelane_b32 v41, s34, 0
; GFX9-NEXT: v_writelane_b32 v40, s31, 1
; GFX9-NEXT: s_getpc_b64 s[34:35]
; GFX9-NEXT: s_add_u32 s34, s34, external_void_func_i32@rel32@lo+4
@@ -1250,9 +1358,10 @@
; GFX9-NEXT: s_swappc_b64 s[30:31], s[34:35]
; GFX9-NEXT: v_readlane_b32 s31, v40, 1
; GFX9-NEXT: v_readlane_b32 s30, v40, 0
-; GFX9-NEXT: v_readlane_b32 s34, v40, 2
+; GFX9-NEXT: v_readlane_b32 s34, v41, 0
; GFX9-NEXT: s_or_saveexec_b64 s[36:37], -1
; GFX9-NEXT: buffer_load_dword v40, off, s[0:3], s33 ; 4-byte Folded Reload
+; GFX9-NEXT: buffer_load_dword v41, off, s[0:3], s33 offset:4 ; 4-byte Folded Reload
; GFX9-NEXT: s_mov_b64 exec, s[36:37]
; GFX9-NEXT: s_addk_i32 s32, 0xfc00
; GFX9-NEXT: s_mov_b32 s33, s34
@@ -1266,22 +1375,25 @@
; GFX10-NEXT: s_mov_b32 s33, s32
; GFX10-NEXT: s_or_saveexec_b32 s35, -1
; GFX10-NEXT: buffer_store_dword v40, off, s[0:3], s33 ; 4-byte Folded Spill
+; GFX10-NEXT: buffer_store_dword v41, off, s[0:3], s33 offset:4 ; 4-byte Folded Spill
; GFX10-NEXT: s_waitcnt_depctr 0xffe3
; GFX10-NEXT: s_mov_b32 exec_lo, s35
-; GFX10-NEXT: v_writelane_b32 v40, s34, 2
+; GFX10-NEXT: v_writelane_b32 v40, s30, 0
; GFX10-NEXT: v_mov_b32_e32 v0, 42
; GFX10-NEXT: s_addk_i32 s32, 0x200
+; GFX10-NEXT: v_writelane_b32 v41, s34, 0
; GFX10-NEXT: s_getpc_b64 s[34:35]
; GFX10-NEXT: s_add_u32 s34, s34, external_void_func_i32@rel32@lo+4
; GFX10-NEXT: s_addc_u32 s35, s35, external_void_func_i32@rel32@hi+12
-; GFX10-NEXT: v_writelane_b32 v40, s30, 0
; GFX10-NEXT: v_writelane_b32 v40, s31, 1
; GFX10-NEXT: s_swappc_b64 s[30:31], s[34:35]
; GFX10-NEXT: v_readlane_b32 s31, v40, 1
; GFX10-NEXT: v_readlane_b32 s30, v40, 0
-; GFX10-NEXT: v_readlane_b32 s34, v40, 2
+; GFX10-NEXT: v_readlane_b32 s34, v41, 0
; GFX10-NEXT: s_or_saveexec_b32 s35, -1
-; GFX10-NEXT: buffer_load_dword v40, off, s[0:3], s33 ; 4-byte Folded Reload
+; GFX10-NEXT: s_clause 0x1
+; GFX10-NEXT: buffer_load_dword v40, off, s[0:3], s33
+; GFX10-NEXT: buffer_load_dword v41, off, s[0:3], s33 offset:4
; GFX10-NEXT: s_waitcnt_depctr 0xffe3
; GFX10-NEXT: s_mov_b32 exec_lo, s35
; GFX10-NEXT: s_addk_i32 s32, 0xfe00
@@ -1295,23 +1407,27 @@
; GFX11-NEXT: s_mov_b32 s0, s33
; GFX11-NEXT: s_mov_b32 s33, s32
; GFX11-NEXT: s_or_saveexec_b32 s1, -1
-; GFX11-NEXT: scratch_store_b32 off, v40, s33 ; 4-byte Folded Spill
+; GFX11-NEXT: s_clause 0x1
+; GFX11-NEXT: scratch_store_b32 off, v40, s33
+; GFX11-NEXT: scratch_store_b32 off, v41, s33 offset:4
; GFX11-NEXT: s_mov_b32 exec_lo, s1
-; GFX11-NEXT: v_writelane_b32 v40, s0, 2
+; GFX11-NEXT: v_writelane_b32 v40, s30, 0
; GFX11-NEXT: v_mov_b32_e32 v0, 42
; GFX11-NEXT: s_add_i32 s32, s32, 16
+; GFX11-NEXT: v_writelane_b32 v41, s0, 0
; GFX11-NEXT: s_getpc_b64 s[0:1]
; GFX11-NEXT: s_add_u32 s0, s0, external_void_func_i32@rel32@lo+4
; GFX11-NEXT: s_addc_u32 s1, s1, external_void_func_i32@rel32@hi+12
-; GFX11-NEXT: v_writelane_b32 v40, s30, 0
; GFX11-NEXT: v_writelane_b32 v40, s31, 1
; GFX11-NEXT: s_swappc_b64 s[30:31], s[0:1]
; GFX11-NEXT: s_delay_alu instid0(VALU_DEP_1)
; GFX11-NEXT: v_readlane_b32 s31, v40, 1
; GFX11-NEXT: v_readlane_b32 s30, v40, 0
-; GFX11-NEXT: v_readlane_b32 s0, v40, 2
+; GFX11-NEXT: v_readlane_b32 s0, v41, 0
; GFX11-NEXT: s_or_saveexec_b32 s1, -1
-; GFX11-NEXT: scratch_load_b32 v40, off, s33 ; 4-byte Folded Reload
+; GFX11-NEXT: s_clause 0x1
+; GFX11-NEXT: scratch_load_b32 v40, off, s33
+; GFX11-NEXT: scratch_load_b32 v41, off, s33 offset:4
; GFX11-NEXT: s_mov_b32 exec_lo, s1
; GFX11-NEXT: s_add_i32 s32, s32, -16
; GFX11-NEXT: s_mov_b32 s33, s0
@@ -1325,22 +1441,25 @@
; GFX10-SCRATCH-NEXT: s_mov_b32 s33, s32
; GFX10-SCRATCH-NEXT: s_or_saveexec_b32 s1, -1
; GFX10-SCRATCH-NEXT: scratch_store_dword off, v40, s33 ; 4-byte Folded Spill
+; GFX10-SCRATCH-NEXT: scratch_store_dword off, v41, s33 offset:4 ; 4-byte Folded Spill
; GFX10-SCRATCH-NEXT: s_waitcnt_depctr 0xffe3
; GFX10-SCRATCH-NEXT: s_mov_b32 exec_lo, s1
-; GFX10-SCRATCH-NEXT: v_writelane_b32 v40, s0, 2
+; GFX10-SCRATCH-NEXT: v_writelane_b32 v40, s30, 0
; GFX10-SCRATCH-NEXT: v_mov_b32_e32 v0, 42
; GFX10-SCRATCH-NEXT: s_add_i32 s32, s32, 16
+; GFX10-SCRATCH-NEXT: v_writelane_b32 v41, s0, 0
; GFX10-SCRATCH-NEXT: s_getpc_b64 s[0:1]
; GFX10-SCRATCH-NEXT: s_add_u32 s0, s0, external_void_func_i32@rel32@lo+4
; GFX10-SCRATCH-NEXT: s_addc_u32 s1, s1, external_void_func_i32@rel32@hi+12
-; GFX10-SCRATCH-NEXT: v_writelane_b32 v40, s30, 0
; GFX10-SCRATCH-NEXT: v_writelane_b32 v40, s31, 1
; GFX10-SCRATCH-NEXT: s_swappc_b64 s[30:31], s[0:1]
; GFX10-SCRATCH-NEXT: v_readlane_b32 s31, v40, 1
; GFX10-SCRATCH-NEXT: v_readlane_b32 s30, v40, 0
-; GFX10-SCRATCH-NEXT: v_readlane_b32 s0, v40, 2
+; GFX10-SCRATCH-NEXT: v_readlane_b32 s0, v41, 0
; GFX10-SCRATCH-NEXT: s_or_saveexec_b32 s1, -1
-; GFX10-SCRATCH-NEXT: scratch_load_dword v40, off, s33 ; 4-byte Folded Reload
+; GFX10-SCRATCH-NEXT: s_clause 0x1
+; GFX10-SCRATCH-NEXT: scratch_load_dword v40, off, s33
+; GFX10-SCRATCH-NEXT: scratch_load_dword v41, off, s33 offset:4
; GFX10-SCRATCH-NEXT: s_waitcnt_depctr 0xffe3
; GFX10-SCRATCH-NEXT: s_mov_b32 exec_lo, s1
; GFX10-SCRATCH-NEXT: s_add_i32 s32, s32, -16
@@ -1359,12 +1478,13 @@
; GFX9-NEXT: s_mov_b32 s33, s32
; GFX9-NEXT: s_or_saveexec_b64 s[36:37], -1
; GFX9-NEXT: buffer_store_dword v40, off, s[0:3], s33 ; 4-byte Folded Spill
+; GFX9-NEXT: buffer_store_dword v41, off, s[0:3], s33 offset:4 ; 4-byte Folded Spill
; GFX9-NEXT: s_mov_b64 exec, s[36:37]
-; GFX9-NEXT: v_writelane_b32 v40, s34, 2
; GFX9-NEXT: s_addk_i32 s32, 0x400
; GFX9-NEXT: v_writelane_b32 v40, s30, 0
; GFX9-NEXT: v_mov_b32_e32 v0, 0x7b
; GFX9-NEXT: v_mov_b32_e32 v1, 0
+; GFX9-NEXT: v_writelane_b32 v41, s34, 0
; GFX9-NEXT: v_writelane_b32 v40, s31, 1
; GFX9-NEXT: s_getpc_b64 s[34:35]
; GFX9-NEXT: s_add_u32 s34, s34, external_void_func_i64@rel32@lo+4
@@ -1372,9 +1492,10 @@
; GFX9-NEXT: s_swappc_b64 s[30:31], s[34:35]
; GFX9-NEXT: v_readlane_b32 s31, v40, 1
; GFX9-NEXT: v_readlane_b32 s30, v40, 0
-; GFX9-NEXT: v_readlane_b32 s34, v40, 2
+; GFX9-NEXT: v_readlane_b32 s34, v41, 0
; GFX9-NEXT: s_or_saveexec_b64 s[36:37], -1
; GFX9-NEXT: buffer_load_dword v40, off, s[0:3], s33 ; 4-byte Folded Reload
+; GFX9-NEXT: buffer_load_dword v41, off, s[0:3], s33 offset:4 ; 4-byte Folded Reload
; GFX9-NEXT: s_mov_b64 exec, s[36:37]
; GFX9-NEXT: s_addk_i32 s32, 0xfc00
; GFX9-NEXT: s_mov_b32 s33, s34
@@ -1388,23 +1509,26 @@
; GFX10-NEXT: s_mov_b32 s33, s32
; GFX10-NEXT: s_or_saveexec_b32 s35, -1
; GFX10-NEXT: buffer_store_dword v40, off, s[0:3], s33 ; 4-byte Folded Spill
+; GFX10-NEXT: buffer_store_dword v41, off, s[0:3], s33 offset:4 ; 4-byte Folded Spill
; GFX10-NEXT: s_waitcnt_depctr 0xffe3
; GFX10-NEXT: s_mov_b32 exec_lo, s35
-; GFX10-NEXT: v_writelane_b32 v40, s34, 2
+; GFX10-NEXT: v_writelane_b32 v40, s30, 0
; GFX10-NEXT: v_mov_b32_e32 v0, 0x7b
; GFX10-NEXT: v_mov_b32_e32 v1, 0
; GFX10-NEXT: s_addk_i32 s32, 0x200
+; GFX10-NEXT: v_writelane_b32 v41, s34, 0
+; GFX10-NEXT: v_writelane_b32 v40, s31, 1
; GFX10-NEXT: s_getpc_b64 s[34:35]
; GFX10-NEXT: s_add_u32 s34, s34, external_void_func_i64@rel32@lo+4
; GFX10-NEXT: s_addc_u32 s35, s35, external_void_func_i64@rel32@hi+12
-; GFX10-NEXT: v_writelane_b32 v40, s30, 0
-; GFX10-NEXT: v_writelane_b32 v40, s31, 1
; GFX10-NEXT: s_swappc_b64 s[30:31], s[34:35]
; GFX10-NEXT: v_readlane_b32 s31, v40, 1
; GFX10-NEXT: v_readlane_b32 s30, v40, 0
-; GFX10-NEXT: v_readlane_b32 s34, v40, 2
+; GFX10-NEXT: v_readlane_b32 s34, v41, 0
; GFX10-NEXT: s_or_saveexec_b32 s35, -1
-; GFX10-NEXT: buffer_load_dword v40, off, s[0:3], s33 ; 4-byte Folded Reload
+; GFX10-NEXT: s_clause 0x1
+; GFX10-NEXT: buffer_load_dword v40, off, s[0:3], s33
+; GFX10-NEXT: buffer_load_dword v41, off, s[0:3], s33 offset:4
; GFX10-NEXT: s_waitcnt_depctr 0xffe3
; GFX10-NEXT: s_mov_b32 exec_lo, s35
; GFX10-NEXT: s_addk_i32 s32, 0xfe00
@@ -1418,23 +1542,27 @@
; GFX11-NEXT: s_mov_b32 s0, s33
; GFX11-NEXT: s_mov_b32 s33, s32
; GFX11-NEXT: s_or_saveexec_b32 s1, -1
-; GFX11-NEXT: scratch_store_b32 off, v40, s33 ; 4-byte Folded Spill
+; GFX11-NEXT: s_clause 0x1
+; GFX11-NEXT: scratch_store_b32 off, v40, s33
+; GFX11-NEXT: scratch_store_b32 off, v41, s33 offset:4
; GFX11-NEXT: s_mov_b32 exec_lo, s1
-; GFX11-NEXT: v_writelane_b32 v40, s0, 2
+; GFX11-NEXT: v_writelane_b32 v40, s30, 0
; GFX11-NEXT: v_dual_mov_b32 v0, 0x7b :: v_dual_mov_b32 v1, 0
; GFX11-NEXT: s_add_i32 s32, s32, 16
+; GFX11-NEXT: v_writelane_b32 v41, s0, 0
+; GFX11-NEXT: v_writelane_b32 v40, s31, 1
; GFX11-NEXT: s_getpc_b64 s[0:1]
; GFX11-NEXT: s_add_u32 s0, s0, external_void_func_i64@rel32@lo+4
; GFX11-NEXT: s_addc_u32 s1, s1, external_void_func_i64@rel32@hi+12
-; GFX11-NEXT: v_writelane_b32 v40, s30, 0
-; GFX11-NEXT: v_writelane_b32 v40, s31, 1
+; GFX11-NEXT: s_delay_alu instid0(SALU_CYCLE_1)
; GFX11-NEXT: s_swappc_b64 s[30:31], s[0:1]
-; GFX11-NEXT: s_delay_alu instid0(VALU_DEP_1)
; GFX11-NEXT: v_readlane_b32 s31, v40, 1
; GFX11-NEXT: v_readlane_b32 s30, v40, 0
-; GFX11-NEXT: v_readlane_b32 s0, v40, 2
+; GFX11-NEXT: v_readlane_b32 s0, v41, 0
; GFX11-NEXT: s_or_saveexec_b32 s1, -1
-; GFX11-NEXT: scratch_load_b32 v40, off, s33 ; 4-byte Folded Reload
+; GFX11-NEXT: s_clause 0x1
+; GFX11-NEXT: scratch_load_b32 v40, off, s33
+; GFX11-NEXT: scratch_load_b32 v41, off, s33 offset:4
; GFX11-NEXT: s_mov_b32 exec_lo, s1
; GFX11-NEXT: s_add_i32 s32, s32, -16
; GFX11-NEXT: s_mov_b32 s33, s0
@@ -1448,24 +1576,27 @@
; GFX10-SCRATCH-NEXT: s_mov_b32 s33, s32
; GFX10-SCRATCH-NEXT: s_or_saveexec_b32 s1, -1
; GFX10-SCRATCH-NEXT: scratch_store_dword off, v40, s33 ; 4-byte Folded Spill
+; GFX10-SCRATCH-NEXT: scratch_store_dword off, v41, s33 offset:4 ; 4-byte Folded Spill
; GFX10-SCRATCH-NEXT: s_waitcnt_depctr 0xffe3
; GFX10-SCRATCH-NEXT: s_mov_b32 exec_lo, s1
-; GFX10-SCRATCH-NEXT: v_writelane_b32 v40, s0, 2
+; GFX10-SCRATCH-NEXT: v_writelane_b32 v40, s30, 0
; GFX10-SCRATCH-NEXT: v_mov_b32_e32 v0, 0x7b
; GFX10-SCRATCH-NEXT: v_mov_b32_e32 v1, 0
; GFX10-SCRATCH-NEXT: s_add_i32 s32, s32, 16
+; GFX10-SCRATCH-NEXT: v_writelane_b32 v41, s0, 0
+; GFX10-SCRATCH-NEXT: v_writelane_b32 v40, s31, 1
; GFX10-SCRATCH-NEXT: s_getpc_b64 s[0:1]
; GFX10-SCRATCH-NEXT: s_add_u32 s0, s0, external_void_func_i64@rel32@lo+4
; GFX10-SCRATCH-NEXT: s_addc_u32 s1, s1, external_void_func_i64@rel32@hi+12
-; GFX10-SCRATCH-NEXT: v_writelane_b32 v40, s30, 0
-; GFX10-SCRATCH-NEXT: v_writelane_b32 v40, s31, 1
; GFX10-SCRATCH-NEXT: s_swappc_b64 s[30:31], s[0:1]
; GFX10-SCRATCH-NEXT: v_readlane_b32 s31, v40, 1
; GFX10-SCRATCH-NEXT: v_readlane_b32 s30, v40, 0
-; GFX10-SCRATCH-NEXT: v_readlane_b32 s0, v40, 2
+; GFX10-SCRATCH-NEXT: v_readlane_b32 s0, v41, 0
; GFX10-SCRATCH-NEXT: s_or_saveexec_b32 s1, -1
-; GFX10-SCRATCH-NEXT: scratch_load_dword v40, off, s33 ; 4-byte Folded Reload
-; GFX10-SCRATCH-NEXT: s_waitcnt_depctr 0xffe3
+; GFX10-SCRATCH-NEXT: s_clause 0x1
+; GFX10-SCRATCH-NEXT: scratch_load_dword v40, off, s33
+; GFX10-SCRATCH-NEXT: scratch_load_dword v41, off, s33 offset:4
+; GFX10-SCRATCH-NEXT: s_waitcnt_depctr 0xffe3
; GFX10-SCRATCH-NEXT: s_mov_b32 exec_lo, s1
; GFX10-SCRATCH-NEXT: s_add_i32 s32, s32, -16
; GFX10-SCRATCH-NEXT: s_mov_b32 s33, s0
@@ -1483,13 +1614,14 @@
; GFX9-NEXT: s_mov_b32 s33, s32
; GFX9-NEXT: s_or_saveexec_b64 s[36:37], -1
; GFX9-NEXT: buffer_store_dword v40, off, s[0:3], s33 ; 4-byte Folded Spill
+; GFX9-NEXT: buffer_store_dword v41, off, s[0:3], s33 offset:4 ; 4-byte Folded Spill
; GFX9-NEXT: s_mov_b64 exec, s[36:37]
; GFX9-NEXT: v_mov_b32_e32 v0, 0
; GFX9-NEXT: v_mov_b32_e32 v1, 0
; GFX9-NEXT: global_load_dwordx4 v[0:3], v[0:1], off
-; GFX9-NEXT: v_writelane_b32 v40, s34, 2
; GFX9-NEXT: s_addk_i32 s32, 0x400
; GFX9-NEXT: v_writelane_b32 v40, s30, 0
+; GFX9-NEXT: v_writelane_b32 v41, s34, 0
; GFX9-NEXT: v_writelane_b32 v40, s31, 1
; GFX9-NEXT: s_getpc_b64 s[34:35]
; GFX9-NEXT: s_add_u32 s34, s34, external_void_func_v2i64@rel32@lo+4
@@ -1497,9 +1629,10 @@
; GFX9-NEXT: s_swappc_b64 s[30:31], s[34:35]
; GFX9-NEXT: v_readlane_b32 s31, v40, 1
; GFX9-NEXT: v_readlane_b32 s30, v40, 0
-; GFX9-NEXT: v_readlane_b32 s34, v40, 2
+; GFX9-NEXT: v_readlane_b32 s34, v41, 0
; GFX9-NEXT: s_or_saveexec_b64 s[36:37], -1
; GFX9-NEXT: buffer_load_dword v40, off, s[0:3], s33 ; 4-byte Folded Reload
+; GFX9-NEXT: buffer_load_dword v41, off, s[0:3], s33 offset:4 ; 4-byte Folded Reload
; GFX9-NEXT: s_mov_b64 exec, s[36:37]
; GFX9-NEXT: s_addk_i32 s32, 0xfc00
; GFX9-NEXT: s_mov_b32 s33, s34
@@ -1513,24 +1646,27 @@
; GFX10-NEXT: s_mov_b32 s33, s32
; GFX10-NEXT: s_or_saveexec_b32 s35, -1
; GFX10-NEXT: buffer_store_dword v40, off, s[0:3], s33 ; 4-byte Folded Spill
+; GFX10-NEXT: buffer_store_dword v41, off, s[0:3], s33 offset:4 ; 4-byte Folded Spill
; GFX10-NEXT: s_waitcnt_depctr 0xffe3
; GFX10-NEXT: s_mov_b32 exec_lo, s35
; GFX10-NEXT: v_mov_b32_e32 v0, 0
; GFX10-NEXT: v_mov_b32_e32 v1, 0
-; GFX10-NEXT: v_writelane_b32 v40, s34, 2
+; GFX10-NEXT: v_writelane_b32 v40, s30, 0
; GFX10-NEXT: s_addk_i32 s32, 0x200
+; GFX10-NEXT: v_writelane_b32 v41, s34, 0
; GFX10-NEXT: s_getpc_b64 s[34:35]
; GFX10-NEXT: s_add_u32 s34, s34, external_void_func_v2i64@rel32@lo+4
; GFX10-NEXT: s_addc_u32 s35, s35, external_void_func_v2i64@rel32@hi+12
; GFX10-NEXT: global_load_dwordx4 v[0:3], v[0:1], off
-; GFX10-NEXT: v_writelane_b32 v40, s30, 0
; GFX10-NEXT: v_writelane_b32 v40, s31, 1
; GFX10-NEXT: s_swappc_b64 s[30:31], s[34:35]
; GFX10-NEXT: v_readlane_b32 s31, v40, 1
; GFX10-NEXT: v_readlane_b32 s30, v40, 0
-; GFX10-NEXT: v_readlane_b32 s34, v40, 2
+; GFX10-NEXT: v_readlane_b32 s34, v41, 0
; GFX10-NEXT: s_or_saveexec_b32 s35, -1
-; GFX10-NEXT: buffer_load_dword v40, off, s[0:3], s33 ; 4-byte Folded Reload
+; GFX10-NEXT: s_clause 0x1
+; GFX10-NEXT: buffer_load_dword v40, off, s[0:3], s33
+; GFX10-NEXT: buffer_load_dword v41, off, s[0:3], s33 offset:4
; GFX10-NEXT: s_waitcnt_depctr 0xffe3
; GFX10-NEXT: s_mov_b32 exec_lo, s35
; GFX10-NEXT: s_addk_i32 s32, 0xfe00
@@ -1544,25 +1680,29 @@
; GFX11-NEXT: s_mov_b32 s0, s33
; GFX11-NEXT: s_mov_b32 s33, s32
; GFX11-NEXT: s_or_saveexec_b32 s1, -1
-; GFX11-NEXT: scratch_store_b32 off, v40, s33 ; 4-byte Folded Spill
+; GFX11-NEXT: s_clause 0x1
+; GFX11-NEXT: scratch_store_b32 off, v40, s33
+; GFX11-NEXT: scratch_store_b32 off, v41, s33 offset:4
; GFX11-NEXT: s_mov_b32 exec_lo, s1
; GFX11-NEXT: v_mov_b32_e32 v0, 0
; GFX11-NEXT: v_mov_b32_e32 v1, 0
-; GFX11-NEXT: v_writelane_b32 v40, s0, 2
+; GFX11-NEXT: v_writelane_b32 v40, s30, 0
; GFX11-NEXT: s_add_i32 s32, s32, 16
+; GFX11-NEXT: v_writelane_b32 v41, s0, 0
; GFX11-NEXT: s_getpc_b64 s[0:1]
; GFX11-NEXT: s_add_u32 s0, s0, external_void_func_v2i64@rel32@lo+4
; GFX11-NEXT: s_addc_u32 s1, s1, external_void_func_v2i64@rel32@hi+12
; GFX11-NEXT: global_load_b128 v[0:3], v[0:1], off
-; GFX11-NEXT: v_writelane_b32 v40, s30, 0
; GFX11-NEXT: v_writelane_b32 v40, s31, 1
; GFX11-NEXT: s_swappc_b64 s[30:31], s[0:1]
; GFX11-NEXT: s_delay_alu instid0(VALU_DEP_1)
; GFX11-NEXT: v_readlane_b32 s31, v40, 1
; GFX11-NEXT: v_readlane_b32 s30, v40, 0
-; GFX11-NEXT: v_readlane_b32 s0, v40, 2
+; GFX11-NEXT: v_readlane_b32 s0, v41, 0
; GFX11-NEXT: s_or_saveexec_b32 s1, -1
-; GFX11-NEXT: scratch_load_b32 v40, off, s33 ; 4-byte Folded Reload
+; GFX11-NEXT: s_clause 0x1
+; GFX11-NEXT: scratch_load_b32 v40, off, s33
+; GFX11-NEXT: scratch_load_b32 v41, off, s33 offset:4
; GFX11-NEXT: s_mov_b32 exec_lo, s1
; GFX11-NEXT: s_add_i32 s32, s32, -16
; GFX11-NEXT: s_mov_b32 s33, s0
@@ -1576,24 +1716,27 @@
; GFX10-SCRATCH-NEXT: s_mov_b32 s33, s32
; GFX10-SCRATCH-NEXT: s_or_saveexec_b32 s1, -1
; GFX10-SCRATCH-NEXT: scratch_store_dword off, v40, s33 ; 4-byte Folded Spill
+; GFX10-SCRATCH-NEXT: scratch_store_dword off, v41, s33 offset:4 ; 4-byte Folded Spill
; GFX10-SCRATCH-NEXT: s_waitcnt_depctr 0xffe3
; GFX10-SCRATCH-NEXT: s_mov_b32 exec_lo, s1
; GFX10-SCRATCH-NEXT: v_mov_b32_e32 v0, 0
; GFX10-SCRATCH-NEXT: v_mov_b32_e32 v1, 0
-; GFX10-SCRATCH-NEXT: v_writelane_b32 v40, s0, 2
+; GFX10-SCRATCH-NEXT: v_writelane_b32 v40, s30, 0
; GFX10-SCRATCH-NEXT: s_add_i32 s32, s32, 16
+; GFX10-SCRATCH-NEXT: v_writelane_b32 v41, s0, 0
; GFX10-SCRATCH-NEXT: s_getpc_b64 s[0:1]
; GFX10-SCRATCH-NEXT: s_add_u32 s0, s0, external_void_func_v2i64@rel32@lo+4
; GFX10-SCRATCH-NEXT: s_addc_u32 s1, s1, external_void_func_v2i64@rel32@hi+12
; GFX10-SCRATCH-NEXT: global_load_dwordx4 v[0:3], v[0:1], off
-; GFX10-SCRATCH-NEXT: v_writelane_b32 v40, s30, 0
; GFX10-SCRATCH-NEXT: v_writelane_b32 v40, s31, 1
; GFX10-SCRATCH-NEXT: s_swappc_b64 s[30:31], s[0:1]
; GFX10-SCRATCH-NEXT: v_readlane_b32 s31, v40, 1
; GFX10-SCRATCH-NEXT: v_readlane_b32 s30, v40, 0
-; GFX10-SCRATCH-NEXT: v_readlane_b32 s0, v40, 2
+; GFX10-SCRATCH-NEXT: v_readlane_b32 s0, v41, 0
; GFX10-SCRATCH-NEXT: s_or_saveexec_b32 s1, -1
-; GFX10-SCRATCH-NEXT: scratch_load_dword v40, off, s33 ; 4-byte Folded Reload
+; GFX10-SCRATCH-NEXT: s_clause 0x1
+; GFX10-SCRATCH-NEXT: scratch_load_dword v40, off, s33
+; GFX10-SCRATCH-NEXT: scratch_load_dword v41, off, s33 offset:4
; GFX10-SCRATCH-NEXT: s_waitcnt_depctr 0xffe3
; GFX10-SCRATCH-NEXT: s_mov_b32 exec_lo, s1
; GFX10-SCRATCH-NEXT: s_add_i32 s32, s32, -16
@@ -1613,14 +1756,15 @@
; GFX9-NEXT: s_mov_b32 s33, s32
; GFX9-NEXT: s_or_saveexec_b64 s[36:37], -1
; GFX9-NEXT: buffer_store_dword v40, off, s[0:3], s33 ; 4-byte Folded Spill
+; GFX9-NEXT: buffer_store_dword v41, off, s[0:3], s33 offset:4 ; 4-byte Folded Spill
; GFX9-NEXT: s_mov_b64 exec, s[36:37]
-; GFX9-NEXT: v_writelane_b32 v40, s34, 2
; GFX9-NEXT: s_addk_i32 s32, 0x400
; GFX9-NEXT: v_writelane_b32 v40, s30, 0
; GFX9-NEXT: v_mov_b32_e32 v0, 1
; GFX9-NEXT: v_mov_b32_e32 v1, 2
; GFX9-NEXT: v_mov_b32_e32 v2, 3
; GFX9-NEXT: v_mov_b32_e32 v3, 4
+; GFX9-NEXT: v_writelane_b32 v41, s34, 0
; GFX9-NEXT: v_writelane_b32 v40, s31, 1
; GFX9-NEXT: s_getpc_b64 s[34:35]
; GFX9-NEXT: s_add_u32 s34, s34, external_void_func_v2i64@rel32@lo+4
@@ -1628,9 +1772,10 @@
; GFX9-NEXT: s_swappc_b64 s[30:31], s[34:35]
; GFX9-NEXT: v_readlane_b32 s31, v40, 1
; GFX9-NEXT: v_readlane_b32 s30, v40, 0
-; GFX9-NEXT: v_readlane_b32 s34, v40, 2
+; GFX9-NEXT: v_readlane_b32 s34, v41, 0
; GFX9-NEXT: s_or_saveexec_b64 s[36:37], -1
; GFX9-NEXT: buffer_load_dword v40, off, s[0:3], s33 ; 4-byte Folded Reload
+; GFX9-NEXT: buffer_load_dword v41, off, s[0:3], s33 offset:4 ; 4-byte Folded Reload
; GFX9-NEXT: s_mov_b64 exec, s[36:37]
; GFX9-NEXT: s_addk_i32 s32, 0xfc00
; GFX9-NEXT: s_mov_b32 s33, s34
@@ -1644,25 +1789,28 @@
; GFX10-NEXT: s_mov_b32 s33, s32
; GFX10-NEXT: s_or_saveexec_b32 s35, -1
; GFX10-NEXT: buffer_store_dword v40, off, s[0:3], s33 ; 4-byte Folded Spill
+; GFX10-NEXT: buffer_store_dword v41, off, s[0:3], s33 offset:4 ; 4-byte Folded Spill
; GFX10-NEXT: s_waitcnt_depctr 0xffe3
; GFX10-NEXT: s_mov_b32 exec_lo, s35
-; GFX10-NEXT: v_writelane_b32 v40, s34, 2
+; GFX10-NEXT: v_writelane_b32 v40, s30, 0
; GFX10-NEXT: v_mov_b32_e32 v0, 1
; GFX10-NEXT: v_mov_b32_e32 v1, 2
; GFX10-NEXT: v_mov_b32_e32 v2, 3
; GFX10-NEXT: v_mov_b32_e32 v3, 4
-; GFX10-NEXT: v_writelane_b32 v40, s30, 0
; GFX10-NEXT: s_addk_i32 s32, 0x200
+; GFX10-NEXT: v_writelane_b32 v41, s34, 0
+; GFX10-NEXT: v_writelane_b32 v40, s31, 1
; GFX10-NEXT: s_getpc_b64 s[34:35]
; GFX10-NEXT: s_add_u32 s34, s34, external_void_func_v2i64@rel32@lo+4
; GFX10-NEXT: s_addc_u32 s35, s35, external_void_func_v2i64@rel32@hi+12
-; GFX10-NEXT: v_writelane_b32 v40, s31, 1
; GFX10-NEXT: s_swappc_b64 s[30:31], s[34:35]
; GFX10-NEXT: v_readlane_b32 s31, v40, 1
; GFX10-NEXT: v_readlane_b32 s30, v40, 0
-; GFX10-NEXT: v_readlane_b32 s34, v40, 2
+; GFX10-NEXT: v_readlane_b32 s34, v41, 0
; GFX10-NEXT: s_or_saveexec_b32 s35, -1
-; GFX10-NEXT: buffer_load_dword v40, off, s[0:3], s33 ; 4-byte Folded Reload
+; GFX10-NEXT: s_clause 0x1
+; GFX10-NEXT: buffer_load_dword v40, off, s[0:3], s33
+; GFX10-NEXT: buffer_load_dword v41, off, s[0:3], s33 offset:4
; GFX10-NEXT: s_waitcnt_depctr 0xffe3
; GFX10-NEXT: s_mov_b32 exec_lo, s35
; GFX10-NEXT: s_addk_i32 s32, 0xfe00
@@ -1676,24 +1824,28 @@
; GFX11-NEXT: s_mov_b32 s0, s33
; GFX11-NEXT: s_mov_b32 s33, s32
; GFX11-NEXT: s_or_saveexec_b32 s1, -1
-; GFX11-NEXT: scratch_store_b32 off, v40, s33 ; 4-byte Folded Spill
+; GFX11-NEXT: s_clause 0x1
+; GFX11-NEXT: scratch_store_b32 off, v40, s33
+; GFX11-NEXT: scratch_store_b32 off, v41, s33 offset:4
; GFX11-NEXT: s_mov_b32 exec_lo, s1
-; GFX11-NEXT: v_writelane_b32 v40, s0, 2
+; GFX11-NEXT: v_writelane_b32 v40, s30, 0
; GFX11-NEXT: v_dual_mov_b32 v0, 1 :: v_dual_mov_b32 v1, 2
; GFX11-NEXT: v_dual_mov_b32 v2, 3 :: v_dual_mov_b32 v3, 4
-; GFX11-NEXT: v_writelane_b32 v40, s30, 0
; GFX11-NEXT: s_add_i32 s32, s32, 16
+; GFX11-NEXT: v_writelane_b32 v41, s0, 0
+; GFX11-NEXT: v_writelane_b32 v40, s31, 1
; GFX11-NEXT: s_getpc_b64 s[0:1]
; GFX11-NEXT: s_add_u32 s0, s0, external_void_func_v2i64@rel32@lo+4
; GFX11-NEXT: s_addc_u32 s1, s1, external_void_func_v2i64@rel32@hi+12
-; GFX11-NEXT: v_writelane_b32 v40, s31, 1
+; GFX11-NEXT: s_delay_alu instid0(SALU_CYCLE_1)
; GFX11-NEXT: s_swappc_b64 s[30:31], s[0:1]
-; GFX11-NEXT: s_delay_alu instid0(VALU_DEP_1)
; GFX11-NEXT: v_readlane_b32 s31, v40, 1
; GFX11-NEXT: v_readlane_b32 s30, v40, 0
-; GFX11-NEXT: v_readlane_b32 s0, v40, 2
+; GFX11-NEXT: v_readlane_b32 s0, v41, 0
; GFX11-NEXT: s_or_saveexec_b32 s1, -1
-; GFX11-NEXT: scratch_load_b32 v40, off, s33 ; 4-byte Folded Reload
+; GFX11-NEXT: s_clause 0x1
+; GFX11-NEXT: scratch_load_b32 v40, off, s33
+; GFX11-NEXT: scratch_load_b32 v41, off, s33 offset:4
; GFX11-NEXT: s_mov_b32 exec_lo, s1
; GFX11-NEXT: s_add_i32 s32, s32, -16
; GFX11-NEXT: s_mov_b32 s33, s0
@@ -1707,25 +1859,28 @@
; GFX10-SCRATCH-NEXT: s_mov_b32 s33, s32
; GFX10-SCRATCH-NEXT: s_or_saveexec_b32 s1, -1
; GFX10-SCRATCH-NEXT: scratch_store_dword off, v40, s33 ; 4-byte Folded Spill
+; GFX10-SCRATCH-NEXT: scratch_store_dword off, v41, s33 offset:4 ; 4-byte Folded Spill
; GFX10-SCRATCH-NEXT: s_waitcnt_depctr 0xffe3
; GFX10-SCRATCH-NEXT: s_mov_b32 exec_lo, s1
-; GFX10-SCRATCH-NEXT: v_writelane_b32 v40, s0, 2
+; GFX10-SCRATCH-NEXT: v_writelane_b32 v40, s30, 0
; GFX10-SCRATCH-NEXT: v_mov_b32_e32 v0, 1
; GFX10-SCRATCH-NEXT: v_mov_b32_e32 v1, 2
; GFX10-SCRATCH-NEXT: v_mov_b32_e32 v2, 3
; GFX10-SCRATCH-NEXT: v_mov_b32_e32 v3, 4
-; GFX10-SCRATCH-NEXT: v_writelane_b32 v40, s30, 0
; GFX10-SCRATCH-NEXT: s_add_i32 s32, s32, 16
+; GFX10-SCRATCH-NEXT: v_writelane_b32 v41, s0, 0
+; GFX10-SCRATCH-NEXT: v_writelane_b32 v40, s31, 1
; GFX10-SCRATCH-NEXT: s_getpc_b64 s[0:1]
; GFX10-SCRATCH-NEXT: s_add_u32 s0, s0, external_void_func_v2i64@rel32@lo+4
; GFX10-SCRATCH-NEXT: s_addc_u32 s1, s1, external_void_func_v2i64@rel32@hi+12
-; GFX10-SCRATCH-NEXT: v_writelane_b32 v40, s31, 1
; GFX10-SCRATCH-NEXT: s_swappc_b64 s[30:31], s[0:1]
; GFX10-SCRATCH-NEXT: v_readlane_b32 s31, v40, 1
; GFX10-SCRATCH-NEXT: v_readlane_b32 s30, v40, 0
-; GFX10-SCRATCH-NEXT: v_readlane_b32 s0, v40, 2
+; GFX10-SCRATCH-NEXT: v_readlane_b32 s0, v41, 0
; GFX10-SCRATCH-NEXT: s_or_saveexec_b32 s1, -1
-; GFX10-SCRATCH-NEXT: scratch_load_dword v40, off, s33 ; 4-byte Folded Reload
+; GFX10-SCRATCH-NEXT: s_clause 0x1
+; GFX10-SCRATCH-NEXT: scratch_load_dword v40, off, s33
+; GFX10-SCRATCH-NEXT: scratch_load_dword v41, off, s33 offset:4
; GFX10-SCRATCH-NEXT: s_waitcnt_depctr 0xffe3
; GFX10-SCRATCH-NEXT: s_mov_b32 exec_lo, s1
; GFX10-SCRATCH-NEXT: s_add_i32 s32, s32, -16
@@ -1744,15 +1899,16 @@
; GFX9-NEXT: s_mov_b32 s33, s32
; GFX9-NEXT: s_or_saveexec_b64 s[36:37], -1
; GFX9-NEXT: buffer_store_dword v40, off, s[0:3], s33 ; 4-byte Folded Spill
+; GFX9-NEXT: buffer_store_dword v41, off, s[0:3], s33 offset:4 ; 4-byte Folded Spill
; GFX9-NEXT: s_mov_b64 exec, s[36:37]
; GFX9-NEXT: v_mov_b32_e32 v0, 0
; GFX9-NEXT: v_mov_b32_e32 v1, 0
; GFX9-NEXT: global_load_dwordx4 v[0:3], v[0:1], off
-; GFX9-NEXT: v_writelane_b32 v40, s34, 2
; GFX9-NEXT: s_addk_i32 s32, 0x400
; GFX9-NEXT: v_writelane_b32 v40, s30, 0
; GFX9-NEXT: v_mov_b32_e32 v4, 1
; GFX9-NEXT: v_mov_b32_e32 v5, 2
+; GFX9-NEXT: v_writelane_b32 v41, s34, 0
; GFX9-NEXT: v_writelane_b32 v40, s31, 1
; GFX9-NEXT: s_getpc_b64 s[34:35]
; GFX9-NEXT: s_add_u32 s34, s34, external_void_func_v3i64@rel32@lo+4
@@ -1760,9 +1916,10 @@
; GFX9-NEXT: s_swappc_b64 s[30:31], s[34:35]
; GFX9-NEXT: v_readlane_b32 s31, v40, 1
; GFX9-NEXT: v_readlane_b32 s30, v40, 0
-; GFX9-NEXT: v_readlane_b32 s34, v40, 2
+; GFX9-NEXT: v_readlane_b32 s34, v41, 0
; GFX9-NEXT: s_or_saveexec_b64 s[36:37], -1
; GFX9-NEXT: buffer_load_dword v40, off, s[0:3], s33 ; 4-byte Folded Reload
+; GFX9-NEXT: buffer_load_dword v41, off, s[0:3], s33 offset:4 ; 4-byte Folded Reload
; GFX9-NEXT: s_mov_b64 exec, s[36:37]
; GFX9-NEXT: s_addk_i32 s32, 0xfc00
; GFX9-NEXT: s_mov_b32 s33, s34
@@ -1776,26 +1933,29 @@
; GFX10-NEXT: s_mov_b32 s33, s32
; GFX10-NEXT: s_or_saveexec_b32 s35, -1
; GFX10-NEXT: buffer_store_dword v40, off, s[0:3], s33 ; 4-byte Folded Spill
+; GFX10-NEXT: buffer_store_dword v41, off, s[0:3], s33 offset:4 ; 4-byte Folded Spill
; GFX10-NEXT: s_waitcnt_depctr 0xffe3
; GFX10-NEXT: s_mov_b32 exec_lo, s35
; GFX10-NEXT: v_mov_b32_e32 v0, 0
; GFX10-NEXT: v_mov_b32_e32 v1, 0
-; GFX10-NEXT: v_writelane_b32 v40, s34, 2
+; GFX10-NEXT: v_writelane_b32 v40, s30, 0
; GFX10-NEXT: v_mov_b32_e32 v4, 1
; GFX10-NEXT: v_mov_b32_e32 v5, 2
; GFX10-NEXT: s_addk_i32 s32, 0x200
; GFX10-NEXT: global_load_dwordx4 v[0:3], v[0:1], off
-; GFX10-NEXT: v_writelane_b32 v40, s30, 0
+; GFX10-NEXT: v_writelane_b32 v41, s34, 0
+; GFX10-NEXT: v_writelane_b32 v40, s31, 1
; GFX10-NEXT: s_getpc_b64 s[34:35]
; GFX10-NEXT: s_add_u32 s34, s34, external_void_func_v3i64@rel32@lo+4
; GFX10-NEXT: s_addc_u32 s35, s35, external_void_func_v3i64@rel32@hi+12
-; GFX10-NEXT: v_writelane_b32 v40, s31, 1
; GFX10-NEXT: s_swappc_b64 s[30:31], s[34:35]
; GFX10-NEXT: v_readlane_b32 s31, v40, 1
; GFX10-NEXT: v_readlane_b32 s30, v40, 0
-; GFX10-NEXT: v_readlane_b32 s34, v40, 2
+; GFX10-NEXT: v_readlane_b32 s34, v41, 0
; GFX10-NEXT: s_or_saveexec_b32 s35, -1
-; GFX10-NEXT: buffer_load_dword v40, off, s[0:3], s33 ; 4-byte Folded Reload
+; GFX10-NEXT: s_clause 0x1
+; GFX10-NEXT: buffer_load_dword v40, off, s[0:3], s33
+; GFX10-NEXT: buffer_load_dword v41, off, s[0:3], s33 offset:4
; GFX10-NEXT: s_waitcnt_depctr 0xffe3
; GFX10-NEXT: s_mov_b32 exec_lo, s35
; GFX10-NEXT: s_addk_i32 s32, 0xfe00
@@ -1809,25 +1969,29 @@
; GFX11-NEXT: s_mov_b32 s0, s33
; GFX11-NEXT: s_mov_b32 s33, s32
; GFX11-NEXT: s_or_saveexec_b32 s1, -1
-; GFX11-NEXT: scratch_store_b32 off, v40, s33 ; 4-byte Folded Spill
+; GFX11-NEXT: s_clause 0x1
+; GFX11-NEXT: scratch_store_b32 off, v40, s33
+; GFX11-NEXT: scratch_store_b32 off, v41, s33 offset:4
; GFX11-NEXT: s_mov_b32 exec_lo, s1
; GFX11-NEXT: v_dual_mov_b32 v0, 0 :: v_dual_mov_b32 v5, 2
; GFX11-NEXT: v_dual_mov_b32 v1, 0 :: v_dual_mov_b32 v4, 1
-; GFX11-NEXT: v_writelane_b32 v40, s0, 2
+; GFX11-NEXT: v_writelane_b32 v40, s30, 0
; GFX11-NEXT: s_add_i32 s32, s32, 16
+; GFX11-NEXT: v_writelane_b32 v41, s0, 0
+; GFX11-NEXT: global_load_b128 v[0:3], v[0:1], off
; GFX11-NEXT: s_getpc_b64 s[0:1]
; GFX11-NEXT: s_add_u32 s0, s0, external_void_func_v3i64@rel32@lo+4
; GFX11-NEXT: s_addc_u32 s1, s1, external_void_func_v3i64@rel32@hi+12
-; GFX11-NEXT: global_load_b128 v[0:3], v[0:1], off
-; GFX11-NEXT: v_writelane_b32 v40, s30, 0
; GFX11-NEXT: v_writelane_b32 v40, s31, 1
; GFX11-NEXT: s_swappc_b64 s[30:31], s[0:1]
; GFX11-NEXT: s_delay_alu instid0(VALU_DEP_1)
; GFX11-NEXT: v_readlane_b32 s31, v40, 1
; GFX11-NEXT: v_readlane_b32 s30, v40, 0
-; GFX11-NEXT: v_readlane_b32 s0, v40, 2
+; GFX11-NEXT: v_readlane_b32 s0, v41, 0
; GFX11-NEXT: s_or_saveexec_b32 s1, -1
-; GFX11-NEXT: scratch_load_b32 v40, off, s33 ; 4-byte Folded Reload
+; GFX11-NEXT: s_clause 0x1
+; GFX11-NEXT: scratch_load_b32 v40, off, s33
+; GFX11-NEXT: scratch_load_b32 v41, off, s33 offset:4
; GFX11-NEXT: s_mov_b32 exec_lo, s1
; GFX11-NEXT: s_add_i32 s32, s32, -16
; GFX11-NEXT: s_mov_b32 s33, s0
@@ -1841,26 +2005,29 @@
; GFX10-SCRATCH-NEXT: s_mov_b32 s33, s32
; GFX10-SCRATCH-NEXT: s_or_saveexec_b32 s1, -1
; GFX10-SCRATCH-NEXT: scratch_store_dword off, v40, s33 ; 4-byte Folded Spill
+; GFX10-SCRATCH-NEXT: scratch_store_dword off, v41, s33 offset:4 ; 4-byte Folded Spill
; GFX10-SCRATCH-NEXT: s_waitcnt_depctr 0xffe3
; GFX10-SCRATCH-NEXT: s_mov_b32 exec_lo, s1
; GFX10-SCRATCH-NEXT: v_mov_b32_e32 v0, 0
; GFX10-SCRATCH-NEXT: v_mov_b32_e32 v1, 0
-; GFX10-SCRATCH-NEXT: v_writelane_b32 v40, s0, 2
+; GFX10-SCRATCH-NEXT: v_writelane_b32 v40, s30, 0
; GFX10-SCRATCH-NEXT: v_mov_b32_e32 v4, 1
; GFX10-SCRATCH-NEXT: v_mov_b32_e32 v5, 2
; GFX10-SCRATCH-NEXT: s_add_i32 s32, s32, 16
; GFX10-SCRATCH-NEXT: global_load_dwordx4 v[0:3], v[0:1], off
-; GFX10-SCRATCH-NEXT: v_writelane_b32 v40, s30, 0
+; GFX10-SCRATCH-NEXT: v_writelane_b32 v41, s0, 0
+; GFX10-SCRATCH-NEXT: v_writelane_b32 v40, s31, 1
; GFX10-SCRATCH-NEXT: s_getpc_b64 s[0:1]
; GFX10-SCRATCH-NEXT: s_add_u32 s0, s0, external_void_func_v3i64@rel32@lo+4
; GFX10-SCRATCH-NEXT: s_addc_u32 s1, s1, external_void_func_v3i64@rel32@hi+12
-; GFX10-SCRATCH-NEXT: v_writelane_b32 v40, s31, 1
; GFX10-SCRATCH-NEXT: s_swappc_b64 s[30:31], s[0:1]
; GFX10-SCRATCH-NEXT: v_readlane_b32 s31, v40, 1
; GFX10-SCRATCH-NEXT: v_readlane_b32 s30, v40, 0
-; GFX10-SCRATCH-NEXT: v_readlane_b32 s0, v40, 2
+; GFX10-SCRATCH-NEXT: v_readlane_b32 s0, v41, 0
; GFX10-SCRATCH-NEXT: s_or_saveexec_b32 s1, -1
-; GFX10-SCRATCH-NEXT: scratch_load_dword v40, off, s33 ; 4-byte Folded Reload
+; GFX10-SCRATCH-NEXT: s_clause 0x1
+; GFX10-SCRATCH-NEXT: scratch_load_dword v40, off, s33
+; GFX10-SCRATCH-NEXT: scratch_load_dword v41, off, s33 offset:4
; GFX10-SCRATCH-NEXT: s_waitcnt_depctr 0xffe3
; GFX10-SCRATCH-NEXT: s_mov_b32 exec_lo, s1
; GFX10-SCRATCH-NEXT: s_add_i32 s32, s32, -16
@@ -1882,17 +2049,18 @@
; GFX9-NEXT: s_mov_b32 s33, s32
; GFX9-NEXT: s_or_saveexec_b64 s[36:37], -1
; GFX9-NEXT: buffer_store_dword v40, off, s[0:3], s33 ; 4-byte Folded Spill
+; GFX9-NEXT: buffer_store_dword v41, off, s[0:3], s33 offset:4 ; 4-byte Folded Spill
; GFX9-NEXT: s_mov_b64 exec, s[36:37]
; GFX9-NEXT: v_mov_b32_e32 v0, 0
; GFX9-NEXT: v_mov_b32_e32 v1, 0
; GFX9-NEXT: global_load_dwordx4 v[0:3], v[0:1], off
-; GFX9-NEXT: v_writelane_b32 v40, s34, 2
; GFX9-NEXT: s_addk_i32 s32, 0x400
; GFX9-NEXT: v_writelane_b32 v40, s30, 0
; GFX9-NEXT: v_mov_b32_e32 v4, 1
; GFX9-NEXT: v_mov_b32_e32 v5, 2
; GFX9-NEXT: v_mov_b32_e32 v6, 3
; GFX9-NEXT: v_mov_b32_e32 v7, 4
+; GFX9-NEXT: v_writelane_b32 v41, s34, 0
; GFX9-NEXT: v_writelane_b32 v40, s31, 1
; GFX9-NEXT: s_getpc_b64 s[34:35]
; GFX9-NEXT: s_add_u32 s34, s34, external_void_func_v4i64@rel32@lo+4
@@ -1900,9 +2068,10 @@
; GFX9-NEXT: s_swappc_b64 s[30:31], s[34:35]
; GFX9-NEXT: v_readlane_b32 s31, v40, 1
; GFX9-NEXT: v_readlane_b32 s30, v40, 0
-; GFX9-NEXT: v_readlane_b32 s34, v40, 2
+; GFX9-NEXT: v_readlane_b32 s34, v41, 0
; GFX9-NEXT: s_or_saveexec_b64 s[36:37], -1
; GFX9-NEXT: buffer_load_dword v40, off, s[0:3], s33 ; 4-byte Folded Reload
+; GFX9-NEXT: buffer_load_dword v41, off, s[0:3], s33 offset:4 ; 4-byte Folded Reload
; GFX9-NEXT: s_mov_b64 exec, s[36:37]
; GFX9-NEXT: s_addk_i32 s32, 0xfc00
; GFX9-NEXT: s_mov_b32 s33, s34
@@ -1916,28 +2085,31 @@
; GFX10-NEXT: s_mov_b32 s33, s32
; GFX10-NEXT: s_or_saveexec_b32 s35, -1
; GFX10-NEXT: buffer_store_dword v40, off, s[0:3], s33 ; 4-byte Folded Spill
+; GFX10-NEXT: buffer_store_dword v41, off, s[0:3], s33 offset:4 ; 4-byte Folded Spill
; GFX10-NEXT: s_waitcnt_depctr 0xffe3
; GFX10-NEXT: s_mov_b32 exec_lo, s35
; GFX10-NEXT: v_mov_b32_e32 v0, 0
; GFX10-NEXT: v_mov_b32_e32 v1, 0
-; GFX10-NEXT: v_writelane_b32 v40, s34, 2
+; GFX10-NEXT: v_writelane_b32 v40, s30, 0
; GFX10-NEXT: v_mov_b32_e32 v4, 1
; GFX10-NEXT: v_mov_b32_e32 v5, 2
; GFX10-NEXT: v_mov_b32_e32 v6, 3
; GFX10-NEXT: global_load_dwordx4 v[0:3], v[0:1], off
-; GFX10-NEXT: v_writelane_b32 v40, s30, 0
; GFX10-NEXT: v_mov_b32_e32 v7, 4
; GFX10-NEXT: s_addk_i32 s32, 0x200
+; GFX10-NEXT: v_writelane_b32 v41, s34, 0
+; GFX10-NEXT: v_writelane_b32 v40, s31, 1
; GFX10-NEXT: s_getpc_b64 s[34:35]
; GFX10-NEXT: s_add_u32 s34, s34, external_void_func_v4i64@rel32@lo+4
; GFX10-NEXT: s_addc_u32 s35, s35, external_void_func_v4i64@rel32@hi+12
-; GFX10-NEXT: v_writelane_b32 v40, s31, 1
; GFX10-NEXT: s_swappc_b64 s[30:31], s[34:35]
; GFX10-NEXT: v_readlane_b32 s31, v40, 1
; GFX10-NEXT: v_readlane_b32 s30, v40, 0
-; GFX10-NEXT: v_readlane_b32 s34, v40, 2
+; GFX10-NEXT: v_readlane_b32 s34, v41, 0
; GFX10-NEXT: s_or_saveexec_b32 s35, -1
-; GFX10-NEXT: buffer_load_dword v40, off, s[0:3], s33 ; 4-byte Folded Reload
+; GFX10-NEXT: s_clause 0x1
+; GFX10-NEXT: buffer_load_dword v40, off, s[0:3], s33
+; GFX10-NEXT: buffer_load_dword v41, off, s[0:3], s33 offset:4
; GFX10-NEXT: s_waitcnt_depctr 0xffe3
; GFX10-NEXT: s_mov_b32 exec_lo, s35
; GFX10-NEXT: s_addk_i32 s32, 0xfe00
@@ -1951,26 +2123,30 @@
; GFX11-NEXT: s_mov_b32 s0, s33
; GFX11-NEXT: s_mov_b32 s33, s32
; GFX11-NEXT: s_or_saveexec_b32 s1, -1
-; GFX11-NEXT: scratch_store_b32 off, v40, s33 ; 4-byte Folded Spill
+; GFX11-NEXT: s_clause 0x1
+; GFX11-NEXT: scratch_store_b32 off, v40, s33
+; GFX11-NEXT: scratch_store_b32 off, v41, s33 offset:4
; GFX11-NEXT: s_mov_b32 exec_lo, s1
; GFX11-NEXT: v_dual_mov_b32 v0, 0 :: v_dual_mov_b32 v5, 2
; GFX11-NEXT: v_dual_mov_b32 v1, 0 :: v_dual_mov_b32 v4, 1
-; GFX11-NEXT: v_writelane_b32 v40, s0, 2
+; GFX11-NEXT: v_writelane_b32 v40, s30, 0
; GFX11-NEXT: v_dual_mov_b32 v6, 3 :: v_dual_mov_b32 v7, 4
; GFX11-NEXT: global_load_b128 v[0:3], v[0:1], off
; GFX11-NEXT: s_add_i32 s32, s32, 16
-; GFX11-NEXT: v_writelane_b32 v40, s30, 0
+; GFX11-NEXT: v_writelane_b32 v41, s0, 0
+; GFX11-NEXT: v_writelane_b32 v40, s31, 1
; GFX11-NEXT: s_getpc_b64 s[0:1]
; GFX11-NEXT: s_add_u32 s0, s0, external_void_func_v4i64@rel32@lo+4
; GFX11-NEXT: s_addc_u32 s1, s1, external_void_func_v4i64@rel32@hi+12
-; GFX11-NEXT: v_writelane_b32 v40, s31, 1
+; GFX11-NEXT: s_delay_alu instid0(SALU_CYCLE_1)
; GFX11-NEXT: s_swappc_b64 s[30:31], s[0:1]
-; GFX11-NEXT: s_delay_alu instid0(VALU_DEP_1)
; GFX11-NEXT: v_readlane_b32 s31, v40, 1
; GFX11-NEXT: v_readlane_b32 s30, v40, 0
-; GFX11-NEXT: v_readlane_b32 s0, v40, 2
+; GFX11-NEXT: v_readlane_b32 s0, v41, 0
; GFX11-NEXT: s_or_saveexec_b32 s1, -1
-; GFX11-NEXT: scratch_load_b32 v40, off, s33 ; 4-byte Folded Reload
+; GFX11-NEXT: s_clause 0x1
+; GFX11-NEXT: scratch_load_b32 v40, off, s33
+; GFX11-NEXT: scratch_load_b32 v41, off, s33 offset:4
; GFX11-NEXT: s_mov_b32 exec_lo, s1
; GFX11-NEXT: s_add_i32 s32, s32, -16
; GFX11-NEXT: s_mov_b32 s33, s0
@@ -1984,28 +2160,31 @@
; GFX10-SCRATCH-NEXT: s_mov_b32 s33, s32
; GFX10-SCRATCH-NEXT: s_or_saveexec_b32 s1, -1
; GFX10-SCRATCH-NEXT: scratch_store_dword off, v40, s33 ; 4-byte Folded Spill
+; GFX10-SCRATCH-NEXT: scratch_store_dword off, v41, s33 offset:4 ; 4-byte Folded Spill
; GFX10-SCRATCH-NEXT: s_waitcnt_depctr 0xffe3
; GFX10-SCRATCH-NEXT: s_mov_b32 exec_lo, s1
; GFX10-SCRATCH-NEXT: v_mov_b32_e32 v0, 0
; GFX10-SCRATCH-NEXT: v_mov_b32_e32 v1, 0
-; GFX10-SCRATCH-NEXT: v_writelane_b32 v40, s0, 2
+; GFX10-SCRATCH-NEXT: v_writelane_b32 v40, s30, 0
; GFX10-SCRATCH-NEXT: v_mov_b32_e32 v4, 1
; GFX10-SCRATCH-NEXT: v_mov_b32_e32 v5, 2
; GFX10-SCRATCH-NEXT: v_mov_b32_e32 v6, 3
; GFX10-SCRATCH-NEXT: global_load_dwordx4 v[0:3], v[0:1], off
-; GFX10-SCRATCH-NEXT: v_writelane_b32 v40, s30, 0
; GFX10-SCRATCH-NEXT: v_mov_b32_e32 v7, 4
; GFX10-SCRATCH-NEXT: s_add_i32 s32, s32, 16
+; GFX10-SCRATCH-NEXT: v_writelane_b32 v41, s0, 0
+; GFX10-SCRATCH-NEXT: v_writelane_b32 v40, s31, 1
; GFX10-SCRATCH-NEXT: s_getpc_b64 s[0:1]
; GFX10-SCRATCH-NEXT: s_add_u32 s0, s0, external_void_func_v4i64@rel32@lo+4
; GFX10-SCRATCH-NEXT: s_addc_u32 s1, s1, external_void_func_v4i64@rel32@hi+12
-; GFX10-SCRATCH-NEXT: v_writelane_b32 v40, s31, 1
; GFX10-SCRATCH-NEXT: s_swappc_b64 s[30:31], s[0:1]
; GFX10-SCRATCH-NEXT: v_readlane_b32 s31, v40, 1
; GFX10-SCRATCH-NEXT: v_readlane_b32 s30, v40, 0
-; GFX10-SCRATCH-NEXT: v_readlane_b32 s0, v40, 2
+; GFX10-SCRATCH-NEXT: v_readlane_b32 s0, v41, 0
; GFX10-SCRATCH-NEXT: s_or_saveexec_b32 s1, -1
-; GFX10-SCRATCH-NEXT: scratch_load_dword v40, off, s33 ; 4-byte Folded Reload
+; GFX10-SCRATCH-NEXT: s_clause 0x1
+; GFX10-SCRATCH-NEXT: scratch_load_dword v40, off, s33
+; GFX10-SCRATCH-NEXT: scratch_load_dword v41, off, s33 offset:4
; GFX10-SCRATCH-NEXT: s_waitcnt_depctr 0xffe3
; GFX10-SCRATCH-NEXT: s_mov_b32 exec_lo, s1
; GFX10-SCRATCH-NEXT: s_add_i32 s32, s32, -16
@@ -2026,11 +2205,12 @@
; GFX9-NEXT: s_mov_b32 s33, s32
; GFX9-NEXT: s_or_saveexec_b64 s[36:37], -1
; GFX9-NEXT: buffer_store_dword v40, off, s[0:3], s33 ; 4-byte Folded Spill
+; GFX9-NEXT: buffer_store_dword v41, off, s[0:3], s33 offset:4 ; 4-byte Folded Spill
; GFX9-NEXT: s_mov_b64 exec, s[36:37]
-; GFX9-NEXT: v_writelane_b32 v40, s34, 2
; GFX9-NEXT: s_addk_i32 s32, 0x400
; GFX9-NEXT: v_writelane_b32 v40, s30, 0
; GFX9-NEXT: v_mov_b32_e32 v0, 0x4400
+; GFX9-NEXT: v_writelane_b32 v41, s34, 0
; GFX9-NEXT: v_writelane_b32 v40, s31, 1
; GFX9-NEXT: s_getpc_b64 s[34:35]
; GFX9-NEXT: s_add_u32 s34, s34, external_void_func_f16@rel32@lo+4
@@ -2038,9 +2218,10 @@
; GFX9-NEXT: s_swappc_b64 s[30:31], s[34:35]
; GFX9-NEXT: v_readlane_b32 s31, v40, 1
; GFX9-NEXT: v_readlane_b32 s30, v40, 0
-; GFX9-NEXT: v_readlane_b32 s34, v40, 2
+; GFX9-NEXT: v_readlane_b32 s34, v41, 0
; GFX9-NEXT: s_or_saveexec_b64 s[36:37], -1
; GFX9-NEXT: buffer_load_dword v40, off, s[0:3], s33 ; 4-byte Folded Reload
+; GFX9-NEXT: buffer_load_dword v41, off, s[0:3], s33 offset:4 ; 4-byte Folded Reload
; GFX9-NEXT: s_mov_b64 exec, s[36:37]
; GFX9-NEXT: s_addk_i32 s32, 0xfc00
; GFX9-NEXT: s_mov_b32 s33, s34
@@ -2054,22 +2235,25 @@
; GFX10-NEXT: s_mov_b32 s33, s32
; GFX10-NEXT: s_or_saveexec_b32 s35, -1
; GFX10-NEXT: buffer_store_dword v40, off, s[0:3], s33 ; 4-byte Folded Spill
+; GFX10-NEXT: buffer_store_dword v41, off, s[0:3], s33 offset:4 ; 4-byte Folded Spill
; GFX10-NEXT: s_waitcnt_depctr 0xffe3
; GFX10-NEXT: s_mov_b32 exec_lo, s35
-; GFX10-NEXT: v_writelane_b32 v40, s34, 2
+; GFX10-NEXT: v_writelane_b32 v40, s30, 0
; GFX10-NEXT: v_mov_b32_e32 v0, 0x4400
; GFX10-NEXT: s_addk_i32 s32, 0x200
+; GFX10-NEXT: v_writelane_b32 v41, s34, 0
; GFX10-NEXT: s_getpc_b64 s[34:35]
; GFX10-NEXT: s_add_u32 s34, s34, external_void_func_f16@rel32@lo+4
; GFX10-NEXT: s_addc_u32 s35, s35, external_void_func_f16@rel32@hi+12
-; GFX10-NEXT: v_writelane_b32 v40, s30, 0
; GFX10-NEXT: v_writelane_b32 v40, s31, 1
; GFX10-NEXT: s_swappc_b64 s[30:31], s[34:35]
; GFX10-NEXT: v_readlane_b32 s31, v40, 1
; GFX10-NEXT: v_readlane_b32 s30, v40, 0
-; GFX10-NEXT: v_readlane_b32 s34, v40, 2
+; GFX10-NEXT: v_readlane_b32 s34, v41, 0
; GFX10-NEXT: s_or_saveexec_b32 s35, -1
-; GFX10-NEXT: buffer_load_dword v40, off, s[0:3], s33 ; 4-byte Folded Reload
+; GFX10-NEXT: s_clause 0x1
+; GFX10-NEXT: buffer_load_dword v40, off, s[0:3], s33
+; GFX10-NEXT: buffer_load_dword v41, off, s[0:3], s33 offset:4
; GFX10-NEXT: s_waitcnt_depctr 0xffe3
; GFX10-NEXT: s_mov_b32 exec_lo, s35
; GFX10-NEXT: s_addk_i32 s32, 0xfe00
@@ -2083,23 +2267,27 @@
; GFX11-NEXT: s_mov_b32 s0, s33
; GFX11-NEXT: s_mov_b32 s33, s32
; GFX11-NEXT: s_or_saveexec_b32 s1, -1
-; GFX11-NEXT: scratch_store_b32 off, v40, s33 ; 4-byte Folded Spill
+; GFX11-NEXT: s_clause 0x1
+; GFX11-NEXT: scratch_store_b32 off, v40, s33
+; GFX11-NEXT: scratch_store_b32 off, v41, s33 offset:4
; GFX11-NEXT: s_mov_b32 exec_lo, s1
-; GFX11-NEXT: v_writelane_b32 v40, s0, 2
+; GFX11-NEXT: v_writelane_b32 v40, s30, 0
; GFX11-NEXT: v_mov_b32_e32 v0, 0x4400
; GFX11-NEXT: s_add_i32 s32, s32, 16
+; GFX11-NEXT: v_writelane_b32 v41, s0, 0
; GFX11-NEXT: s_getpc_b64 s[0:1]
; GFX11-NEXT: s_add_u32 s0, s0, external_void_func_f16@rel32@lo+4
; GFX11-NEXT: s_addc_u32 s1, s1, external_void_func_f16@rel32@hi+12
-; GFX11-NEXT: v_writelane_b32 v40, s30, 0
; GFX11-NEXT: v_writelane_b32 v40, s31, 1
; GFX11-NEXT: s_swappc_b64 s[30:31], s[0:1]
; GFX11-NEXT: s_delay_alu instid0(VALU_DEP_1)
; GFX11-NEXT: v_readlane_b32 s31, v40, 1
; GFX11-NEXT: v_readlane_b32 s30, v40, 0
-; GFX11-NEXT: v_readlane_b32 s0, v40, 2
+; GFX11-NEXT: v_readlane_b32 s0, v41, 0
; GFX11-NEXT: s_or_saveexec_b32 s1, -1
-; GFX11-NEXT: scratch_load_b32 v40, off, s33 ; 4-byte Folded Reload
+; GFX11-NEXT: s_clause 0x1
+; GFX11-NEXT: scratch_load_b32 v40, off, s33
+; GFX11-NEXT: scratch_load_b32 v41, off, s33 offset:4
; GFX11-NEXT: s_mov_b32 exec_lo, s1
; GFX11-NEXT: s_add_i32 s32, s32, -16
; GFX11-NEXT: s_mov_b32 s33, s0
@@ -2113,22 +2301,25 @@
; GFX10-SCRATCH-NEXT: s_mov_b32 s33, s32
; GFX10-SCRATCH-NEXT: s_or_saveexec_b32 s1, -1
; GFX10-SCRATCH-NEXT: scratch_store_dword off, v40, s33 ; 4-byte Folded Spill
+; GFX10-SCRATCH-NEXT: scratch_store_dword off, v41, s33 offset:4 ; 4-byte Folded Spill
; GFX10-SCRATCH-NEXT: s_waitcnt_depctr 0xffe3
; GFX10-SCRATCH-NEXT: s_mov_b32 exec_lo, s1
-; GFX10-SCRATCH-NEXT: v_writelane_b32 v40, s0, 2
+; GFX10-SCRATCH-NEXT: v_writelane_b32 v40, s30, 0
; GFX10-SCRATCH-NEXT: v_mov_b32_e32 v0, 0x4400
; GFX10-SCRATCH-NEXT: s_add_i32 s32, s32, 16
+; GFX10-SCRATCH-NEXT: v_writelane_b32 v41, s0, 0
; GFX10-SCRATCH-NEXT: s_getpc_b64 s[0:1]
; GFX10-SCRATCH-NEXT: s_add_u32 s0, s0, external_void_func_f16@rel32@lo+4
; GFX10-SCRATCH-NEXT: s_addc_u32 s1, s1, external_void_func_f16@rel32@hi+12
-; GFX10-SCRATCH-NEXT: v_writelane_b32 v40, s30, 0
; GFX10-SCRATCH-NEXT: v_writelane_b32 v40, s31, 1
; GFX10-SCRATCH-NEXT: s_swappc_b64 s[30:31], s[0:1]
; GFX10-SCRATCH-NEXT: v_readlane_b32 s31, v40, 1
; GFX10-SCRATCH-NEXT: v_readlane_b32 s30, v40, 0
-; GFX10-SCRATCH-NEXT: v_readlane_b32 s0, v40, 2
+; GFX10-SCRATCH-NEXT: v_readlane_b32 s0, v41, 0
; GFX10-SCRATCH-NEXT: s_or_saveexec_b32 s1, -1
-; GFX10-SCRATCH-NEXT: scratch_load_dword v40, off, s33 ; 4-byte Folded Reload
+; GFX10-SCRATCH-NEXT: s_clause 0x1
+; GFX10-SCRATCH-NEXT: scratch_load_dword v40, off, s33
+; GFX10-SCRATCH-NEXT: scratch_load_dword v41, off, s33 offset:4
; GFX10-SCRATCH-NEXT: s_waitcnt_depctr 0xffe3
; GFX10-SCRATCH-NEXT: s_mov_b32 exec_lo, s1
; GFX10-SCRATCH-NEXT: s_add_i32 s32, s32, -16
@@ -2147,11 +2338,12 @@
; GFX9-NEXT: s_mov_b32 s33, s32
; GFX9-NEXT: s_or_saveexec_b64 s[36:37], -1
; GFX9-NEXT: buffer_store_dword v40, off, s[0:3], s33 ; 4-byte Folded Spill
+; GFX9-NEXT: buffer_store_dword v41, off, s[0:3], s33 offset:4 ; 4-byte Folded Spill
; GFX9-NEXT: s_mov_b64 exec, s[36:37]
-; GFX9-NEXT: v_writelane_b32 v40, s34, 2
; GFX9-NEXT: s_addk_i32 s32, 0x400
; GFX9-NEXT: v_writelane_b32 v40, s30, 0
; GFX9-NEXT: v_mov_b32_e32 v0, 4.0
+; GFX9-NEXT: v_writelane_b32 v41, s34, 0
; GFX9-NEXT: v_writelane_b32 v40, s31, 1
; GFX9-NEXT: s_getpc_b64 s[34:35]
; GFX9-NEXT: s_add_u32 s34, s34, external_void_func_f32@rel32@lo+4
@@ -2159,9 +2351,10 @@
; GFX9-NEXT: s_swappc_b64 s[30:31], s[34:35]
; GFX9-NEXT: v_readlane_b32 s31, v40, 1
; GFX9-NEXT: v_readlane_b32 s30, v40, 0
-; GFX9-NEXT: v_readlane_b32 s34, v40, 2
+; GFX9-NEXT: v_readlane_b32 s34, v41, 0
; GFX9-NEXT: s_or_saveexec_b64 s[36:37], -1
; GFX9-NEXT: buffer_load_dword v40, off, s[0:3], s33 ; 4-byte Folded Reload
+; GFX9-NEXT: buffer_load_dword v41, off, s[0:3], s33 offset:4 ; 4-byte Folded Reload
; GFX9-NEXT: s_mov_b64 exec, s[36:37]
; GFX9-NEXT: s_addk_i32 s32, 0xfc00
; GFX9-NEXT: s_mov_b32 s33, s34
@@ -2175,22 +2368,25 @@
; GFX10-NEXT: s_mov_b32 s33, s32
; GFX10-NEXT: s_or_saveexec_b32 s35, -1
; GFX10-NEXT: buffer_store_dword v40, off, s[0:3], s33 ; 4-byte Folded Spill
+; GFX10-NEXT: buffer_store_dword v41, off, s[0:3], s33 offset:4 ; 4-byte Folded Spill
; GFX10-NEXT: s_waitcnt_depctr 0xffe3
; GFX10-NEXT: s_mov_b32 exec_lo, s35
-; GFX10-NEXT: v_writelane_b32 v40, s34, 2
+; GFX10-NEXT: v_writelane_b32 v40, s30, 0
; GFX10-NEXT: v_mov_b32_e32 v0, 4.0
; GFX10-NEXT: s_addk_i32 s32, 0x200
+; GFX10-NEXT: v_writelane_b32 v41, s34, 0
; GFX10-NEXT: s_getpc_b64 s[34:35]
; GFX10-NEXT: s_add_u32 s34, s34, external_void_func_f32@rel32@lo+4
; GFX10-NEXT: s_addc_u32 s35, s35, external_void_func_f32@rel32@hi+12
-; GFX10-NEXT: v_writelane_b32 v40, s30, 0
; GFX10-NEXT: v_writelane_b32 v40, s31, 1
; GFX10-NEXT: s_swappc_b64 s[30:31], s[34:35]
; GFX10-NEXT: v_readlane_b32 s31, v40, 1
; GFX10-NEXT: v_readlane_b32 s30, v40, 0
-; GFX10-NEXT: v_readlane_b32 s34, v40, 2
+; GFX10-NEXT: v_readlane_b32 s34, v41, 0
; GFX10-NEXT: s_or_saveexec_b32 s35, -1
-; GFX10-NEXT: buffer_load_dword v40, off, s[0:3], s33 ; 4-byte Folded Reload
+; GFX10-NEXT: s_clause 0x1
+; GFX10-NEXT: buffer_load_dword v40, off, s[0:3], s33
+; GFX10-NEXT: buffer_load_dword v41, off, s[0:3], s33 offset:4
; GFX10-NEXT: s_waitcnt_depctr 0xffe3
; GFX10-NEXT: s_mov_b32 exec_lo, s35
; GFX10-NEXT: s_addk_i32 s32, 0xfe00
@@ -2204,23 +2400,27 @@
; GFX11-NEXT: s_mov_b32 s0, s33
; GFX11-NEXT: s_mov_b32 s33, s32
; GFX11-NEXT: s_or_saveexec_b32 s1, -1
-; GFX11-NEXT: scratch_store_b32 off, v40, s33 ; 4-byte Folded Spill
+; GFX11-NEXT: s_clause 0x1
+; GFX11-NEXT: scratch_store_b32 off, v40, s33
+; GFX11-NEXT: scratch_store_b32 off, v41, s33 offset:4
; GFX11-NEXT: s_mov_b32 exec_lo, s1
-; GFX11-NEXT: v_writelane_b32 v40, s0, 2
+; GFX11-NEXT: v_writelane_b32 v40, s30, 0
; GFX11-NEXT: v_mov_b32_e32 v0, 4.0
; GFX11-NEXT: s_add_i32 s32, s32, 16
+; GFX11-NEXT: v_writelane_b32 v41, s0, 0
; GFX11-NEXT: s_getpc_b64 s[0:1]
; GFX11-NEXT: s_add_u32 s0, s0, external_void_func_f32@rel32@lo+4
; GFX11-NEXT: s_addc_u32 s1, s1, external_void_func_f32@rel32@hi+12
-; GFX11-NEXT: v_writelane_b32 v40, s30, 0
; GFX11-NEXT: v_writelane_b32 v40, s31, 1
; GFX11-NEXT: s_swappc_b64 s[30:31], s[0:1]
; GFX11-NEXT: s_delay_alu instid0(VALU_DEP_1)
; GFX11-NEXT: v_readlane_b32 s31, v40, 1
; GFX11-NEXT: v_readlane_b32 s30, v40, 0
-; GFX11-NEXT: v_readlane_b32 s0, v40, 2
+; GFX11-NEXT: v_readlane_b32 s0, v41, 0
; GFX11-NEXT: s_or_saveexec_b32 s1, -1
-; GFX11-NEXT: scratch_load_b32 v40, off, s33 ; 4-byte Folded Reload
+; GFX11-NEXT: s_clause 0x1
+; GFX11-NEXT: scratch_load_b32 v40, off, s33
+; GFX11-NEXT: scratch_load_b32 v41, off, s33 offset:4
; GFX11-NEXT: s_mov_b32 exec_lo, s1
; GFX11-NEXT: s_add_i32 s32, s32, -16
; GFX11-NEXT: s_mov_b32 s33, s0
@@ -2234,22 +2434,25 @@
; GFX10-SCRATCH-NEXT: s_mov_b32 s33, s32
; GFX10-SCRATCH-NEXT: s_or_saveexec_b32 s1, -1
; GFX10-SCRATCH-NEXT: scratch_store_dword off, v40, s33 ; 4-byte Folded Spill
+; GFX10-SCRATCH-NEXT: scratch_store_dword off, v41, s33 offset:4 ; 4-byte Folded Spill
; GFX10-SCRATCH-NEXT: s_waitcnt_depctr 0xffe3
; GFX10-SCRATCH-NEXT: s_mov_b32 exec_lo, s1
-; GFX10-SCRATCH-NEXT: v_writelane_b32 v40, s0, 2
+; GFX10-SCRATCH-NEXT: v_writelane_b32 v40, s30, 0
; GFX10-SCRATCH-NEXT: v_mov_b32_e32 v0, 4.0
; GFX10-SCRATCH-NEXT: s_add_i32 s32, s32, 16
+; GFX10-SCRATCH-NEXT: v_writelane_b32 v41, s0, 0
; GFX10-SCRATCH-NEXT: s_getpc_b64 s[0:1]
; GFX10-SCRATCH-NEXT: s_add_u32 s0, s0, external_void_func_f32@rel32@lo+4
; GFX10-SCRATCH-NEXT: s_addc_u32 s1, s1, external_void_func_f32@rel32@hi+12
-; GFX10-SCRATCH-NEXT: v_writelane_b32 v40, s30, 0
; GFX10-SCRATCH-NEXT: v_writelane_b32 v40, s31, 1
; GFX10-SCRATCH-NEXT: s_swappc_b64 s[30:31], s[0:1]
; GFX10-SCRATCH-NEXT: v_readlane_b32 s31, v40, 1
; GFX10-SCRATCH-NEXT: v_readlane_b32 s30, v40, 0
-; GFX10-SCRATCH-NEXT: v_readlane_b32 s0, v40, 2
+; GFX10-SCRATCH-NEXT: v_readlane_b32 s0, v41, 0
; GFX10-SCRATCH-NEXT: s_or_saveexec_b32 s1, -1
-; GFX10-SCRATCH-NEXT: scratch_load_dword v40, off, s33 ; 4-byte Folded Reload
+; GFX10-SCRATCH-NEXT: s_clause 0x1
+; GFX10-SCRATCH-NEXT: scratch_load_dword v40, off, s33
+; GFX10-SCRATCH-NEXT: scratch_load_dword v41, off, s33 offset:4
; GFX10-SCRATCH-NEXT: s_waitcnt_depctr 0xffe3
; GFX10-SCRATCH-NEXT: s_mov_b32 exec_lo, s1
; GFX10-SCRATCH-NEXT: s_add_i32 s32, s32, -16
@@ -2268,12 +2471,13 @@
; GFX9-NEXT: s_mov_b32 s33, s32
; GFX9-NEXT: s_or_saveexec_b64 s[36:37], -1
; GFX9-NEXT: buffer_store_dword v40, off, s[0:3], s33 ; 4-byte Folded Spill
+; GFX9-NEXT: buffer_store_dword v41, off, s[0:3], s33 offset:4 ; 4-byte Folded Spill
; GFX9-NEXT: s_mov_b64 exec, s[36:37]
-; GFX9-NEXT: v_writelane_b32 v40, s34, 2
; GFX9-NEXT: s_addk_i32 s32, 0x400
; GFX9-NEXT: v_writelane_b32 v40, s30, 0
; GFX9-NEXT: v_mov_b32_e32 v0, 1.0
; GFX9-NEXT: v_mov_b32_e32 v1, 2.0
+; GFX9-NEXT: v_writelane_b32 v41, s34, 0
; GFX9-NEXT: v_writelane_b32 v40, s31, 1
; GFX9-NEXT: s_getpc_b64 s[34:35]
; GFX9-NEXT: s_add_u32 s34, s34, external_void_func_v2f32@rel32@lo+4
@@ -2281,9 +2485,10 @@
; GFX9-NEXT: s_swappc_b64 s[30:31], s[34:35]
; GFX9-NEXT: v_readlane_b32 s31, v40, 1
; GFX9-NEXT: v_readlane_b32 s30, v40, 0
-; GFX9-NEXT: v_readlane_b32 s34, v40, 2
+; GFX9-NEXT: v_readlane_b32 s34, v41, 0
; GFX9-NEXT: s_or_saveexec_b64 s[36:37], -1
; GFX9-NEXT: buffer_load_dword v40, off, s[0:3], s33 ; 4-byte Folded Reload
+; GFX9-NEXT: buffer_load_dword v41, off, s[0:3], s33 offset:4 ; 4-byte Folded Reload
; GFX9-NEXT: s_mov_b64 exec, s[36:37]
; GFX9-NEXT: s_addk_i32 s32, 0xfc00
; GFX9-NEXT: s_mov_b32 s33, s34
@@ -2297,23 +2502,26 @@
; GFX10-NEXT: s_mov_b32 s33, s32
; GFX10-NEXT: s_or_saveexec_b32 s35, -1
; GFX10-NEXT: buffer_store_dword v40, off, s[0:3], s33 ; 4-byte Folded Spill
+; GFX10-NEXT: buffer_store_dword v41, off, s[0:3], s33 offset:4 ; 4-byte Folded Spill
; GFX10-NEXT: s_waitcnt_depctr 0xffe3
; GFX10-NEXT: s_mov_b32 exec_lo, s35
-; GFX10-NEXT: v_writelane_b32 v40, s34, 2
+; GFX10-NEXT: v_writelane_b32 v40, s30, 0
; GFX10-NEXT: v_mov_b32_e32 v0, 1.0
; GFX10-NEXT: v_mov_b32_e32 v1, 2.0
; GFX10-NEXT: s_addk_i32 s32, 0x200
+; GFX10-NEXT: v_writelane_b32 v41, s34, 0
+; GFX10-NEXT: v_writelane_b32 v40, s31, 1
; GFX10-NEXT: s_getpc_b64 s[34:35]
; GFX10-NEXT: s_add_u32 s34, s34, external_void_func_v2f32@rel32@lo+4
; GFX10-NEXT: s_addc_u32 s35, s35, external_void_func_v2f32@rel32@hi+12
-; GFX10-NEXT: v_writelane_b32 v40, s30, 0
-; GFX10-NEXT: v_writelane_b32 v40, s31, 1
; GFX10-NEXT: s_swappc_b64 s[30:31], s[34:35]
; GFX10-NEXT: v_readlane_b32 s31, v40, 1
; GFX10-NEXT: v_readlane_b32 s30, v40, 0
-; GFX10-NEXT: v_readlane_b32 s34, v40, 2
+; GFX10-NEXT: v_readlane_b32 s34, v41, 0
; GFX10-NEXT: s_or_saveexec_b32 s35, -1
-; GFX10-NEXT: buffer_load_dword v40, off, s[0:3], s33 ; 4-byte Folded Reload
+; GFX10-NEXT: s_clause 0x1
+; GFX10-NEXT: buffer_load_dword v40, off, s[0:3], s33
+; GFX10-NEXT: buffer_load_dword v41, off, s[0:3], s33 offset:4
; GFX10-NEXT: s_waitcnt_depctr 0xffe3
; GFX10-NEXT: s_mov_b32 exec_lo, s35
; GFX10-NEXT: s_addk_i32 s32, 0xfe00
@@ -2327,23 +2535,27 @@
; GFX11-NEXT: s_mov_b32 s0, s33
; GFX11-NEXT: s_mov_b32 s33, s32
; GFX11-NEXT: s_or_saveexec_b32 s1, -1
-; GFX11-NEXT: scratch_store_b32 off, v40, s33 ; 4-byte Folded Spill
+; GFX11-NEXT: s_clause 0x1
+; GFX11-NEXT: scratch_store_b32 off, v40, s33
+; GFX11-NEXT: scratch_store_b32 off, v41, s33 offset:4
; GFX11-NEXT: s_mov_b32 exec_lo, s1
-; GFX11-NEXT: v_writelane_b32 v40, s0, 2
+; GFX11-NEXT: v_writelane_b32 v40, s30, 0
; GFX11-NEXT: v_dual_mov_b32 v0, 1.0 :: v_dual_mov_b32 v1, 2.0
; GFX11-NEXT: s_add_i32 s32, s32, 16
+; GFX11-NEXT: v_writelane_b32 v41, s0, 0
+; GFX11-NEXT: v_writelane_b32 v40, s31, 1
; GFX11-NEXT: s_getpc_b64 s[0:1]
; GFX11-NEXT: s_add_u32 s0, s0, external_void_func_v2f32@rel32@lo+4
; GFX11-NEXT: s_addc_u32 s1, s1, external_void_func_v2f32@rel32@hi+12
-; GFX11-NEXT: v_writelane_b32 v40, s30, 0
-; GFX11-NEXT: v_writelane_b32 v40, s31, 1
+; GFX11-NEXT: s_delay_alu instid0(SALU_CYCLE_1)
; GFX11-NEXT: s_swappc_b64 s[30:31], s[0:1]
-; GFX11-NEXT: s_delay_alu instid0(VALU_DEP_1)
; GFX11-NEXT: v_readlane_b32 s31, v40, 1
; GFX11-NEXT: v_readlane_b32 s30, v40, 0
-; GFX11-NEXT: v_readlane_b32 s0, v40, 2
+; GFX11-NEXT: v_readlane_b32 s0, v41, 0
; GFX11-NEXT: s_or_saveexec_b32 s1, -1
-; GFX11-NEXT: scratch_load_b32 v40, off, s33 ; 4-byte Folded Reload
+; GFX11-NEXT: s_clause 0x1
+; GFX11-NEXT: scratch_load_b32 v40, off, s33
+; GFX11-NEXT: scratch_load_b32 v41, off, s33 offset:4
; GFX11-NEXT: s_mov_b32 exec_lo, s1
; GFX11-NEXT: s_add_i32 s32, s32, -16
; GFX11-NEXT: s_mov_b32 s33, s0
@@ -2357,23 +2569,26 @@
; GFX10-SCRATCH-NEXT: s_mov_b32 s33, s32
; GFX10-SCRATCH-NEXT: s_or_saveexec_b32 s1, -1
; GFX10-SCRATCH-NEXT: scratch_store_dword off, v40, s33 ; 4-byte Folded Spill
+; GFX10-SCRATCH-NEXT: scratch_store_dword off, v41, s33 offset:4 ; 4-byte Folded Spill
; GFX10-SCRATCH-NEXT: s_waitcnt_depctr 0xffe3
; GFX10-SCRATCH-NEXT: s_mov_b32 exec_lo, s1
-; GFX10-SCRATCH-NEXT: v_writelane_b32 v40, s0, 2
+; GFX10-SCRATCH-NEXT: v_writelane_b32 v40, s30, 0
; GFX10-SCRATCH-NEXT: v_mov_b32_e32 v0, 1.0
; GFX10-SCRATCH-NEXT: v_mov_b32_e32 v1, 2.0
; GFX10-SCRATCH-NEXT: s_add_i32 s32, s32, 16
+; GFX10-SCRATCH-NEXT: v_writelane_b32 v41, s0, 0
+; GFX10-SCRATCH-NEXT: v_writelane_b32 v40, s31, 1
; GFX10-SCRATCH-NEXT: s_getpc_b64 s[0:1]
; GFX10-SCRATCH-NEXT: s_add_u32 s0, s0, external_void_func_v2f32@rel32@lo+4
; GFX10-SCRATCH-NEXT: s_addc_u32 s1, s1, external_void_func_v2f32@rel32@hi+12
-; GFX10-SCRATCH-NEXT: v_writelane_b32 v40, s30, 0
-; GFX10-SCRATCH-NEXT: v_writelane_b32 v40, s31, 1
; GFX10-SCRATCH-NEXT: s_swappc_b64 s[30:31], s[0:1]
; GFX10-SCRATCH-NEXT: v_readlane_b32 s31, v40, 1
; GFX10-SCRATCH-NEXT: v_readlane_b32 s30, v40, 0
-; GFX10-SCRATCH-NEXT: v_readlane_b32 s0, v40, 2
+; GFX10-SCRATCH-NEXT: v_readlane_b32 s0, v41, 0
; GFX10-SCRATCH-NEXT: s_or_saveexec_b32 s1, -1
-; GFX10-SCRATCH-NEXT: scratch_load_dword v40, off, s33 ; 4-byte Folded Reload
+; GFX10-SCRATCH-NEXT: s_clause 0x1
+; GFX10-SCRATCH-NEXT: scratch_load_dword v40, off, s33
+; GFX10-SCRATCH-NEXT: scratch_load_dword v41, off, s33 offset:4
; GFX10-SCRATCH-NEXT: s_waitcnt_depctr 0xffe3
; GFX10-SCRATCH-NEXT: s_mov_b32 exec_lo, s1
; GFX10-SCRATCH-NEXT: s_add_i32 s32, s32, -16
@@ -2392,13 +2607,14 @@
; GFX9-NEXT: s_mov_b32 s33, s32
; GFX9-NEXT: s_or_saveexec_b64 s[36:37], -1
; GFX9-NEXT: buffer_store_dword v40, off, s[0:3], s33 ; 4-byte Folded Spill
+; GFX9-NEXT: buffer_store_dword v41, off, s[0:3], s33 offset:4 ; 4-byte Folded Spill
; GFX9-NEXT: s_mov_b64 exec, s[36:37]
-; GFX9-NEXT: v_writelane_b32 v40, s34, 2
; GFX9-NEXT: s_addk_i32 s32, 0x400
; GFX9-NEXT: v_writelane_b32 v40, s30, 0
; GFX9-NEXT: v_mov_b32_e32 v0, 1.0
; GFX9-NEXT: v_mov_b32_e32 v1, 2.0
; GFX9-NEXT: v_mov_b32_e32 v2, 4.0
+; GFX9-NEXT: v_writelane_b32 v41, s34, 0
; GFX9-NEXT: v_writelane_b32 v40, s31, 1
; GFX9-NEXT: s_getpc_b64 s[34:35]
; GFX9-NEXT: s_add_u32 s34, s34, external_void_func_v3f32@rel32@lo+4
@@ -2406,9 +2622,10 @@
; GFX9-NEXT: s_swappc_b64 s[30:31], s[34:35]
; GFX9-NEXT: v_readlane_b32 s31, v40, 1
; GFX9-NEXT: v_readlane_b32 s30, v40, 0
-; GFX9-NEXT: v_readlane_b32 s34, v40, 2
+; GFX9-NEXT: v_readlane_b32 s34, v41, 0
; GFX9-NEXT: s_or_saveexec_b64 s[36:37], -1
; GFX9-NEXT: buffer_load_dword v40, off, s[0:3], s33 ; 4-byte Folded Reload
+; GFX9-NEXT: buffer_load_dword v41, off, s[0:3], s33 offset:4 ; 4-byte Folded Reload
; GFX9-NEXT: s_mov_b64 exec, s[36:37]
; GFX9-NEXT: s_addk_i32 s32, 0xfc00
; GFX9-NEXT: s_mov_b32 s33, s34
@@ -2422,24 +2639,27 @@
; GFX10-NEXT: s_mov_b32 s33, s32
; GFX10-NEXT: s_or_saveexec_b32 s35, -1
; GFX10-NEXT: buffer_store_dword v40, off, s[0:3], s33 ; 4-byte Folded Spill
+; GFX10-NEXT: buffer_store_dword v41, off, s[0:3], s33 offset:4 ; 4-byte Folded Spill
; GFX10-NEXT: s_waitcnt_depctr 0xffe3
; GFX10-NEXT: s_mov_b32 exec_lo, s35
-; GFX10-NEXT: v_writelane_b32 v40, s34, 2
+; GFX10-NEXT: v_writelane_b32 v40, s30, 0
; GFX10-NEXT: v_mov_b32_e32 v0, 1.0
; GFX10-NEXT: v_mov_b32_e32 v1, 2.0
; GFX10-NEXT: v_mov_b32_e32 v2, 4.0
; GFX10-NEXT: s_addk_i32 s32, 0x200
-; GFX10-NEXT: v_writelane_b32 v40, s30, 0
+; GFX10-NEXT: v_writelane_b32 v41, s34, 0
+; GFX10-NEXT: v_writelane_b32 v40, s31, 1
; GFX10-NEXT: s_getpc_b64 s[34:35]
; GFX10-NEXT: s_add_u32 s34, s34, external_void_func_v3f32@rel32@lo+4
; GFX10-NEXT: s_addc_u32 s35, s35, external_void_func_v3f32@rel32@hi+12
-; GFX10-NEXT: v_writelane_b32 v40, s31, 1
; GFX10-NEXT: s_swappc_b64 s[30:31], s[34:35]
; GFX10-NEXT: v_readlane_b32 s31, v40, 1
; GFX10-NEXT: v_readlane_b32 s30, v40, 0
-; GFX10-NEXT: v_readlane_b32 s34, v40, 2
+; GFX10-NEXT: v_readlane_b32 s34, v41, 0
; GFX10-NEXT: s_or_saveexec_b32 s35, -1
-; GFX10-NEXT: buffer_load_dword v40, off, s[0:3], s33 ; 4-byte Folded Reload
+; GFX10-NEXT: s_clause 0x1
+; GFX10-NEXT: buffer_load_dword v40, off, s[0:3], s33
+; GFX10-NEXT: buffer_load_dword v41, off, s[0:3], s33 offset:4
; GFX10-NEXT: s_waitcnt_depctr 0xffe3
; GFX10-NEXT: s_mov_b32 exec_lo, s35
; GFX10-NEXT: s_addk_i32 s32, 0xfe00
@@ -2453,24 +2673,28 @@
; GFX11-NEXT: s_mov_b32 s0, s33
; GFX11-NEXT: s_mov_b32 s33, s32
; GFX11-NEXT: s_or_saveexec_b32 s1, -1
-; GFX11-NEXT: scratch_store_b32 off, v40, s33 ; 4-byte Folded Spill
+; GFX11-NEXT: s_clause 0x1
+; GFX11-NEXT: scratch_store_b32 off, v40, s33
+; GFX11-NEXT: scratch_store_b32 off, v41, s33 offset:4
; GFX11-NEXT: s_mov_b32 exec_lo, s1
-; GFX11-NEXT: v_writelane_b32 v40, s0, 2
+; GFX11-NEXT: v_writelane_b32 v40, s30, 0
; GFX11-NEXT: v_dual_mov_b32 v0, 1.0 :: v_dual_mov_b32 v1, 2.0
; GFX11-NEXT: v_mov_b32_e32 v2, 4.0
; GFX11-NEXT: s_add_i32 s32, s32, 16
-; GFX11-NEXT: v_writelane_b32 v40, s30, 0
+; GFX11-NEXT: v_writelane_b32 v41, s0, 0
+; GFX11-NEXT: v_writelane_b32 v40, s31, 1
; GFX11-NEXT: s_getpc_b64 s[0:1]
; GFX11-NEXT: s_add_u32 s0, s0, external_void_func_v3f32@rel32@lo+4
; GFX11-NEXT: s_addc_u32 s1, s1, external_void_func_v3f32@rel32@hi+12
-; GFX11-NEXT: v_writelane_b32 v40, s31, 1
+; GFX11-NEXT: s_delay_alu instid0(SALU_CYCLE_1)
; GFX11-NEXT: s_swappc_b64 s[30:31], s[0:1]
-; GFX11-NEXT: s_delay_alu instid0(VALU_DEP_1)
; GFX11-NEXT: v_readlane_b32 s31, v40, 1
; GFX11-NEXT: v_readlane_b32 s30, v40, 0
-; GFX11-NEXT: v_readlane_b32 s0, v40, 2
+; GFX11-NEXT: v_readlane_b32 s0, v41, 0
; GFX11-NEXT: s_or_saveexec_b32 s1, -1
-; GFX11-NEXT: scratch_load_b32 v40, off, s33 ; 4-byte Folded Reload
+; GFX11-NEXT: s_clause 0x1
+; GFX11-NEXT: scratch_load_b32 v40, off, s33
+; GFX11-NEXT: scratch_load_b32 v41, off, s33 offset:4
; GFX11-NEXT: s_mov_b32 exec_lo, s1
; GFX11-NEXT: s_add_i32 s32, s32, -16
; GFX11-NEXT: s_mov_b32 s33, s0
@@ -2484,24 +2708,27 @@
; GFX10-SCRATCH-NEXT: s_mov_b32 s33, s32
; GFX10-SCRATCH-NEXT: s_or_saveexec_b32 s1, -1
; GFX10-SCRATCH-NEXT: scratch_store_dword off, v40, s33 ; 4-byte Folded Spill
+; GFX10-SCRATCH-NEXT: scratch_store_dword off, v41, s33 offset:4 ; 4-byte Folded Spill
; GFX10-SCRATCH-NEXT: s_waitcnt_depctr 0xffe3
; GFX10-SCRATCH-NEXT: s_mov_b32 exec_lo, s1
-; GFX10-SCRATCH-NEXT: v_writelane_b32 v40, s0, 2
+; GFX10-SCRATCH-NEXT: v_writelane_b32 v40, s30, 0
; GFX10-SCRATCH-NEXT: v_mov_b32_e32 v0, 1.0
; GFX10-SCRATCH-NEXT: v_mov_b32_e32 v1, 2.0
; GFX10-SCRATCH-NEXT: v_mov_b32_e32 v2, 4.0
; GFX10-SCRATCH-NEXT: s_add_i32 s32, s32, 16
-; GFX10-SCRATCH-NEXT: v_writelane_b32 v40, s30, 0
+; GFX10-SCRATCH-NEXT: v_writelane_b32 v41, s0, 0
+; GFX10-SCRATCH-NEXT: v_writelane_b32 v40, s31, 1
; GFX10-SCRATCH-NEXT: s_getpc_b64 s[0:1]
; GFX10-SCRATCH-NEXT: s_add_u32 s0, s0, external_void_func_v3f32@rel32@lo+4
; GFX10-SCRATCH-NEXT: s_addc_u32 s1, s1, external_void_func_v3f32@rel32@hi+12
-; GFX10-SCRATCH-NEXT: v_writelane_b32 v40, s31, 1
; GFX10-SCRATCH-NEXT: s_swappc_b64 s[30:31], s[0:1]
; GFX10-SCRATCH-NEXT: v_readlane_b32 s31, v40, 1
; GFX10-SCRATCH-NEXT: v_readlane_b32 s30, v40, 0
-; GFX10-SCRATCH-NEXT: v_readlane_b32 s0, v40, 2
+; GFX10-SCRATCH-NEXT: v_readlane_b32 s0, v41, 0
; GFX10-SCRATCH-NEXT: s_or_saveexec_b32 s1, -1
-; GFX10-SCRATCH-NEXT: scratch_load_dword v40, off, s33 ; 4-byte Folded Reload
+; GFX10-SCRATCH-NEXT: s_clause 0x1
+; GFX10-SCRATCH-NEXT: scratch_load_dword v40, off, s33
+; GFX10-SCRATCH-NEXT: scratch_load_dword v41, off, s33 offset:4
; GFX10-SCRATCH-NEXT: s_waitcnt_depctr 0xffe3
; GFX10-SCRATCH-NEXT: s_mov_b32 exec_lo, s1
; GFX10-SCRATCH-NEXT: s_add_i32 s32, s32, -16
@@ -2520,8 +2747,8 @@
; GFX9-NEXT: s_mov_b32 s33, s32
; GFX9-NEXT: s_or_saveexec_b64 s[36:37], -1
; GFX9-NEXT: buffer_store_dword v40, off, s[0:3], s33 ; 4-byte Folded Spill
+; GFX9-NEXT: buffer_store_dword v41, off, s[0:3], s33 offset:4 ; 4-byte Folded Spill
; GFX9-NEXT: s_mov_b64 exec, s[36:37]
-; GFX9-NEXT: v_writelane_b32 v40, s34, 2
; GFX9-NEXT: s_addk_i32 s32, 0x400
; GFX9-NEXT: v_writelane_b32 v40, s30, 0
; GFX9-NEXT: v_mov_b32_e32 v0, 1.0
@@ -2529,6 +2756,7 @@
; GFX9-NEXT: v_mov_b32_e32 v2, 4.0
; GFX9-NEXT: v_mov_b32_e32 v3, -1.0
; GFX9-NEXT: v_mov_b32_e32 v4, 0.5
+; GFX9-NEXT: v_writelane_b32 v41, s34, 0
; GFX9-NEXT: v_writelane_b32 v40, s31, 1
; GFX9-NEXT: s_getpc_b64 s[34:35]
; GFX9-NEXT: s_add_u32 s34, s34, external_void_func_v5f32@rel32@lo+4
@@ -2536,9 +2764,10 @@
; GFX9-NEXT: s_swappc_b64 s[30:31], s[34:35]
; GFX9-NEXT: v_readlane_b32 s31, v40, 1
; GFX9-NEXT: v_readlane_b32 s30, v40, 0
-; GFX9-NEXT: v_readlane_b32 s34, v40, 2
+; GFX9-NEXT: v_readlane_b32 s34, v41, 0
; GFX9-NEXT: s_or_saveexec_b64 s[36:37], -1
; GFX9-NEXT: buffer_load_dword v40, off, s[0:3], s33 ; 4-byte Folded Reload
+; GFX9-NEXT: buffer_load_dword v41, off, s[0:3], s33 offset:4 ; 4-byte Folded Reload
; GFX9-NEXT: s_mov_b64 exec, s[36:37]
; GFX9-NEXT: s_addk_i32 s32, 0xfc00
; GFX9-NEXT: s_mov_b32 s33, s34
@@ -2552,26 +2781,29 @@
; GFX10-NEXT: s_mov_b32 s33, s32
; GFX10-NEXT: s_or_saveexec_b32 s35, -1
; GFX10-NEXT: buffer_store_dword v40, off, s[0:3], s33 ; 4-byte Folded Spill
+; GFX10-NEXT: buffer_store_dword v41, off, s[0:3], s33 offset:4 ; 4-byte Folded Spill
; GFX10-NEXT: s_waitcnt_depctr 0xffe3
; GFX10-NEXT: s_mov_b32 exec_lo, s35
-; GFX10-NEXT: v_writelane_b32 v40, s34, 2
+; GFX10-NEXT: v_writelane_b32 v40, s30, 0
; GFX10-NEXT: v_mov_b32_e32 v0, 1.0
; GFX10-NEXT: v_mov_b32_e32 v1, 2.0
; GFX10-NEXT: v_mov_b32_e32 v2, 4.0
; GFX10-NEXT: v_mov_b32_e32 v3, -1.0
-; GFX10-NEXT: v_writelane_b32 v40, s30, 0
; GFX10-NEXT: v_mov_b32_e32 v4, 0.5
; GFX10-NEXT: s_addk_i32 s32, 0x200
+; GFX10-NEXT: v_writelane_b32 v41, s34, 0
+; GFX10-NEXT: v_writelane_b32 v40, s31, 1
; GFX10-NEXT: s_getpc_b64 s[34:35]
; GFX10-NEXT: s_add_u32 s34, s34, external_void_func_v5f32@rel32@lo+4
; GFX10-NEXT: s_addc_u32 s35, s35, external_void_func_v5f32@rel32@hi+12
-; GFX10-NEXT: v_writelane_b32 v40, s31, 1
; GFX10-NEXT: s_swappc_b64 s[30:31], s[34:35]
; GFX10-NEXT: v_readlane_b32 s31, v40, 1
; GFX10-NEXT: v_readlane_b32 s30, v40, 0
-; GFX10-NEXT: v_readlane_b32 s34, v40, 2
+; GFX10-NEXT: v_readlane_b32 s34, v41, 0
; GFX10-NEXT: s_or_saveexec_b32 s35, -1
-; GFX10-NEXT: buffer_load_dword v40, off, s[0:3], s33 ; 4-byte Folded Reload
+; GFX10-NEXT: s_clause 0x1
+; GFX10-NEXT: buffer_load_dword v40, off, s[0:3], s33
+; GFX10-NEXT: buffer_load_dword v41, off, s[0:3], s33 offset:4
; GFX10-NEXT: s_waitcnt_depctr 0xffe3
; GFX10-NEXT: s_mov_b32 exec_lo, s35
; GFX10-NEXT: s_addk_i32 s32, 0xfe00
@@ -2585,25 +2817,29 @@
; GFX11-NEXT: s_mov_b32 s0, s33
; GFX11-NEXT: s_mov_b32 s33, s32
; GFX11-NEXT: s_or_saveexec_b32 s1, -1
-; GFX11-NEXT: scratch_store_b32 off, v40, s33 ; 4-byte Folded Spill
+; GFX11-NEXT: s_clause 0x1
+; GFX11-NEXT: scratch_store_b32 off, v40, s33
+; GFX11-NEXT: scratch_store_b32 off, v41, s33 offset:4
; GFX11-NEXT: s_mov_b32 exec_lo, s1
-; GFX11-NEXT: v_writelane_b32 v40, s0, 2
+; GFX11-NEXT: v_writelane_b32 v40, s30, 0
; GFX11-NEXT: v_dual_mov_b32 v0, 1.0 :: v_dual_mov_b32 v1, 2.0
; GFX11-NEXT: v_dual_mov_b32 v2, 4.0 :: v_dual_mov_b32 v3, -1.0
-; GFX11-NEXT: v_writelane_b32 v40, s30, 0
; GFX11-NEXT: v_mov_b32_e32 v4, 0.5
; GFX11-NEXT: s_add_i32 s32, s32, 16
+; GFX11-NEXT: v_writelane_b32 v41, s0, 0
+; GFX11-NEXT: v_writelane_b32 v40, s31, 1
; GFX11-NEXT: s_getpc_b64 s[0:1]
; GFX11-NEXT: s_add_u32 s0, s0, external_void_func_v5f32@rel32@lo+4
; GFX11-NEXT: s_addc_u32 s1, s1, external_void_func_v5f32@rel32@hi+12
-; GFX11-NEXT: v_writelane_b32 v40, s31, 1
+; GFX11-NEXT: s_delay_alu instid0(SALU_CYCLE_1)
; GFX11-NEXT: s_swappc_b64 s[30:31], s[0:1]
-; GFX11-NEXT: s_delay_alu instid0(VALU_DEP_1)
; GFX11-NEXT: v_readlane_b32 s31, v40, 1
; GFX11-NEXT: v_readlane_b32 s30, v40, 0
-; GFX11-NEXT: v_readlane_b32 s0, v40, 2
+; GFX11-NEXT: v_readlane_b32 s0, v41, 0
; GFX11-NEXT: s_or_saveexec_b32 s1, -1
-; GFX11-NEXT: scratch_load_b32 v40, off, s33 ; 4-byte Folded Reload
+; GFX11-NEXT: s_clause 0x1
+; GFX11-NEXT: scratch_load_b32 v40, off, s33
+; GFX11-NEXT: scratch_load_b32 v41, off, s33 offset:4
; GFX11-NEXT: s_mov_b32 exec_lo, s1
; GFX11-NEXT: s_add_i32 s32, s32, -16
; GFX11-NEXT: s_mov_b32 s33, s0
@@ -2617,26 +2853,29 @@
; GFX10-SCRATCH-NEXT: s_mov_b32 s33, s32
; GFX10-SCRATCH-NEXT: s_or_saveexec_b32 s1, -1
; GFX10-SCRATCH-NEXT: scratch_store_dword off, v40, s33 ; 4-byte Folded Spill
+; GFX10-SCRATCH-NEXT: scratch_store_dword off, v41, s33 offset:4 ; 4-byte Folded Spill
; GFX10-SCRATCH-NEXT: s_waitcnt_depctr 0xffe3
; GFX10-SCRATCH-NEXT: s_mov_b32 exec_lo, s1
-; GFX10-SCRATCH-NEXT: v_writelane_b32 v40, s0, 2
+; GFX10-SCRATCH-NEXT: v_writelane_b32 v40, s30, 0
; GFX10-SCRATCH-NEXT: v_mov_b32_e32 v0, 1.0
; GFX10-SCRATCH-NEXT: v_mov_b32_e32 v1, 2.0
; GFX10-SCRATCH-NEXT: v_mov_b32_e32 v2, 4.0
; GFX10-SCRATCH-NEXT: v_mov_b32_e32 v3, -1.0
-; GFX10-SCRATCH-NEXT: v_writelane_b32 v40, s30, 0
; GFX10-SCRATCH-NEXT: v_mov_b32_e32 v4, 0.5
; GFX10-SCRATCH-NEXT: s_add_i32 s32, s32, 16
+; GFX10-SCRATCH-NEXT: v_writelane_b32 v41, s0, 0
+; GFX10-SCRATCH-NEXT: v_writelane_b32 v40, s31, 1
; GFX10-SCRATCH-NEXT: s_getpc_b64 s[0:1]
; GFX10-SCRATCH-NEXT: s_add_u32 s0, s0, external_void_func_v5f32@rel32@lo+4
; GFX10-SCRATCH-NEXT: s_addc_u32 s1, s1, external_void_func_v5f32@rel32@hi+12
-; GFX10-SCRATCH-NEXT: v_writelane_b32 v40, s31, 1
; GFX10-SCRATCH-NEXT: s_swappc_b64 s[30:31], s[0:1]
; GFX10-SCRATCH-NEXT: v_readlane_b32 s31, v40, 1
; GFX10-SCRATCH-NEXT: v_readlane_b32 s30, v40, 0
-; GFX10-SCRATCH-NEXT: v_readlane_b32 s0, v40, 2
+; GFX10-SCRATCH-NEXT: v_readlane_b32 s0, v41, 0
; GFX10-SCRATCH-NEXT: s_or_saveexec_b32 s1, -1
-; GFX10-SCRATCH-NEXT: scratch_load_dword v40, off, s33 ; 4-byte Folded Reload
+; GFX10-SCRATCH-NEXT: s_clause 0x1
+; GFX10-SCRATCH-NEXT: scratch_load_dword v40, off, s33
+; GFX10-SCRATCH-NEXT: scratch_load_dword v41, off, s33 offset:4
; GFX10-SCRATCH-NEXT: s_waitcnt_depctr 0xffe3
; GFX10-SCRATCH-NEXT: s_mov_b32 exec_lo, s1
; GFX10-SCRATCH-NEXT: s_add_i32 s32, s32, -16
@@ -2655,12 +2894,13 @@
; GFX9-NEXT: s_mov_b32 s33, s32
; GFX9-NEXT: s_or_saveexec_b64 s[36:37], -1
; GFX9-NEXT: buffer_store_dword v40, off, s[0:3], s33 ; 4-byte Folded Spill
+; GFX9-NEXT: buffer_store_dword v41, off, s[0:3], s33 offset:4 ; 4-byte Folded Spill
; GFX9-NEXT: s_mov_b64 exec, s[36:37]
-; GFX9-NEXT: v_writelane_b32 v40, s34, 2
; GFX9-NEXT: s_addk_i32 s32, 0x400
; GFX9-NEXT: v_writelane_b32 v40, s30, 0
; GFX9-NEXT: v_mov_b32_e32 v0, 0
; GFX9-NEXT: v_mov_b32_e32 v1, 0x40100000
+; GFX9-NEXT: v_writelane_b32 v41, s34, 0
; GFX9-NEXT: v_writelane_b32 v40, s31, 1
; GFX9-NEXT: s_getpc_b64 s[34:35]
; GFX9-NEXT: s_add_u32 s34, s34, external_void_func_f64@rel32@lo+4
@@ -2668,9 +2908,10 @@
; GFX9-NEXT: s_swappc_b64 s[30:31], s[34:35]
; GFX9-NEXT: v_readlane_b32 s31, v40, 1
; GFX9-NEXT: v_readlane_b32 s30, v40, 0
-; GFX9-NEXT: v_readlane_b32 s34, v40, 2
+; GFX9-NEXT: v_readlane_b32 s34, v41, 0
; GFX9-NEXT: s_or_saveexec_b64 s[36:37], -1
; GFX9-NEXT: buffer_load_dword v40, off, s[0:3], s33 ; 4-byte Folded Reload
+; GFX9-NEXT: buffer_load_dword v41, off, s[0:3], s33 offset:4 ; 4-byte Folded Reload
; GFX9-NEXT: s_mov_b64 exec, s[36:37]
; GFX9-NEXT: s_addk_i32 s32, 0xfc00
; GFX9-NEXT: s_mov_b32 s33, s34
@@ -2684,23 +2925,26 @@
; GFX10-NEXT: s_mov_b32 s33, s32
; GFX10-NEXT: s_or_saveexec_b32 s35, -1
; GFX10-NEXT: buffer_store_dword v40, off, s[0:3], s33 ; 4-byte Folded Spill
+; GFX10-NEXT: buffer_store_dword v41, off, s[0:3], s33 offset:4 ; 4-byte Folded Spill
; GFX10-NEXT: s_waitcnt_depctr 0xffe3
; GFX10-NEXT: s_mov_b32 exec_lo, s35
-; GFX10-NEXT: v_writelane_b32 v40, s34, 2
+; GFX10-NEXT: v_writelane_b32 v40, s30, 0
; GFX10-NEXT: v_mov_b32_e32 v0, 0
; GFX10-NEXT: v_mov_b32_e32 v1, 0x40100000
; GFX10-NEXT: s_addk_i32 s32, 0x200
+; GFX10-NEXT: v_writelane_b32 v41, s34, 0
+; GFX10-NEXT: v_writelane_b32 v40, s31, 1
; GFX10-NEXT: s_getpc_b64 s[34:35]
; GFX10-NEXT: s_add_u32 s34, s34, external_void_func_f64@rel32@lo+4
; GFX10-NEXT: s_addc_u32 s35, s35, external_void_func_f64@rel32@hi+12
-; GFX10-NEXT: v_writelane_b32 v40, s30, 0
-; GFX10-NEXT: v_writelane_b32 v40, s31, 1
; GFX10-NEXT: s_swappc_b64 s[30:31], s[34:35]
; GFX10-NEXT: v_readlane_b32 s31, v40, 1
; GFX10-NEXT: v_readlane_b32 s30, v40, 0
-; GFX10-NEXT: v_readlane_b32 s34, v40, 2
+; GFX10-NEXT: v_readlane_b32 s34, v41, 0
; GFX10-NEXT: s_or_saveexec_b32 s35, -1
-; GFX10-NEXT: buffer_load_dword v40, off, s[0:3], s33 ; 4-byte Folded Reload
+; GFX10-NEXT: s_clause 0x1
+; GFX10-NEXT: buffer_load_dword v40, off, s[0:3], s33
+; GFX10-NEXT: buffer_load_dword v41, off, s[0:3], s33 offset:4
; GFX10-NEXT: s_waitcnt_depctr 0xffe3
; GFX10-NEXT: s_mov_b32 exec_lo, s35
; GFX10-NEXT: s_addk_i32 s32, 0xfe00
@@ -2714,23 +2958,27 @@
; GFX11-NEXT: s_mov_b32 s0, s33
; GFX11-NEXT: s_mov_b32 s33, s32
; GFX11-NEXT: s_or_saveexec_b32 s1, -1
-; GFX11-NEXT: scratch_store_b32 off, v40, s33 ; 4-byte Folded Spill
+; GFX11-NEXT: s_clause 0x1
+; GFX11-NEXT: scratch_store_b32 off, v40, s33
+; GFX11-NEXT: scratch_store_b32 off, v41, s33 offset:4
; GFX11-NEXT: s_mov_b32 exec_lo, s1
-; GFX11-NEXT: v_writelane_b32 v40, s0, 2
+; GFX11-NEXT: v_writelane_b32 v40, s30, 0
; GFX11-NEXT: v_dual_mov_b32 v0, 0 :: v_dual_mov_b32 v1, 0x40100000
; GFX11-NEXT: s_add_i32 s32, s32, 16
+; GFX11-NEXT: v_writelane_b32 v41, s0, 0
+; GFX11-NEXT: v_writelane_b32 v40, s31, 1
; GFX11-NEXT: s_getpc_b64 s[0:1]
; GFX11-NEXT: s_add_u32 s0, s0, external_void_func_f64@rel32@lo+4
; GFX11-NEXT: s_addc_u32 s1, s1, external_void_func_f64@rel32@hi+12
-; GFX11-NEXT: v_writelane_b32 v40, s30, 0
-; GFX11-NEXT: v_writelane_b32 v40, s31, 1
+; GFX11-NEXT: s_delay_alu instid0(SALU_CYCLE_1)
; GFX11-NEXT: s_swappc_b64 s[30:31], s[0:1]
-; GFX11-NEXT: s_delay_alu instid0(VALU_DEP_1)
; GFX11-NEXT: v_readlane_b32 s31, v40, 1
; GFX11-NEXT: v_readlane_b32 s30, v40, 0
-; GFX11-NEXT: v_readlane_b32 s0, v40, 2
+; GFX11-NEXT: v_readlane_b32 s0, v41, 0
; GFX11-NEXT: s_or_saveexec_b32 s1, -1
-; GFX11-NEXT: scratch_load_b32 v40, off, s33 ; 4-byte Folded Reload
+; GFX11-NEXT: s_clause 0x1
+; GFX11-NEXT: scratch_load_b32 v40, off, s33
+; GFX11-NEXT: scratch_load_b32 v41, off, s33 offset:4
; GFX11-NEXT: s_mov_b32 exec_lo, s1
; GFX11-NEXT: s_add_i32 s32, s32, -16
; GFX11-NEXT: s_mov_b32 s33, s0
@@ -2744,23 +2992,26 @@
; GFX10-SCRATCH-NEXT: s_mov_b32 s33, s32
; GFX10-SCRATCH-NEXT: s_or_saveexec_b32 s1, -1
; GFX10-SCRATCH-NEXT: scratch_store_dword off, v40, s33 ; 4-byte Folded Spill
+; GFX10-SCRATCH-NEXT: scratch_store_dword off, v41, s33 offset:4 ; 4-byte Folded Spill
; GFX10-SCRATCH-NEXT: s_waitcnt_depctr 0xffe3
; GFX10-SCRATCH-NEXT: s_mov_b32 exec_lo, s1
-; GFX10-SCRATCH-NEXT: v_writelane_b32 v40, s0, 2
+; GFX10-SCRATCH-NEXT: v_writelane_b32 v40, s30, 0
; GFX10-SCRATCH-NEXT: v_mov_b32_e32 v0, 0
; GFX10-SCRATCH-NEXT: v_mov_b32_e32 v1, 0x40100000
; GFX10-SCRATCH-NEXT: s_add_i32 s32, s32, 16
+; GFX10-SCRATCH-NEXT: v_writelane_b32 v41, s0, 0
+; GFX10-SCRATCH-NEXT: v_writelane_b32 v40, s31, 1
; GFX10-SCRATCH-NEXT: s_getpc_b64 s[0:1]
; GFX10-SCRATCH-NEXT: s_add_u32 s0, s0, external_void_func_f64@rel32@lo+4
; GFX10-SCRATCH-NEXT: s_addc_u32 s1, s1, external_void_func_f64@rel32@hi+12
-; GFX10-SCRATCH-NEXT: v_writelane_b32 v40, s30, 0
-; GFX10-SCRATCH-NEXT: v_writelane_b32 v40, s31, 1
; GFX10-SCRATCH-NEXT: s_swappc_b64 s[30:31], s[0:1]
; GFX10-SCRATCH-NEXT: v_readlane_b32 s31, v40, 1
; GFX10-SCRATCH-NEXT: v_readlane_b32 s30, v40, 0
-; GFX10-SCRATCH-NEXT: v_readlane_b32 s0, v40, 2
+; GFX10-SCRATCH-NEXT: v_readlane_b32 s0, v41, 0
; GFX10-SCRATCH-NEXT: s_or_saveexec_b32 s1, -1
-; GFX10-SCRATCH-NEXT: scratch_load_dword v40, off, s33 ; 4-byte Folded Reload
+; GFX10-SCRATCH-NEXT: s_clause 0x1
+; GFX10-SCRATCH-NEXT: scratch_load_dword v40, off, s33
+; GFX10-SCRATCH-NEXT: scratch_load_dword v41, off, s33 offset:4
; GFX10-SCRATCH-NEXT: s_waitcnt_depctr 0xffe3
; GFX10-SCRATCH-NEXT: s_mov_b32 exec_lo, s1
; GFX10-SCRATCH-NEXT: s_add_i32 s32, s32, -16
@@ -2779,14 +3030,15 @@
; GFX9-NEXT: s_mov_b32 s33, s32
; GFX9-NEXT: s_or_saveexec_b64 s[36:37], -1
; GFX9-NEXT: buffer_store_dword v40, off, s[0:3], s33 ; 4-byte Folded Spill
+; GFX9-NEXT: buffer_store_dword v41, off, s[0:3], s33 offset:4 ; 4-byte Folded Spill
; GFX9-NEXT: s_mov_b64 exec, s[36:37]
-; GFX9-NEXT: v_writelane_b32 v40, s34, 2
; GFX9-NEXT: s_addk_i32 s32, 0x400
; GFX9-NEXT: v_writelane_b32 v40, s30, 0
; GFX9-NEXT: v_mov_b32_e32 v0, 0
; GFX9-NEXT: v_mov_b32_e32 v1, 2.0
; GFX9-NEXT: v_mov_b32_e32 v2, 0
; GFX9-NEXT: v_mov_b32_e32 v3, 0x40100000
+; GFX9-NEXT: v_writelane_b32 v41, s34, 0
; GFX9-NEXT: v_writelane_b32 v40, s31, 1
; GFX9-NEXT: s_getpc_b64 s[34:35]
; GFX9-NEXT: s_add_u32 s34, s34, external_void_func_v2f64@rel32@lo+4
@@ -2794,9 +3046,10 @@
; GFX9-NEXT: s_swappc_b64 s[30:31], s[34:35]
; GFX9-NEXT: v_readlane_b32 s31, v40, 1
; GFX9-NEXT: v_readlane_b32 s30, v40, 0
-; GFX9-NEXT: v_readlane_b32 s34, v40, 2
+; GFX9-NEXT: v_readlane_b32 s34, v41, 0
; GFX9-NEXT: s_or_saveexec_b64 s[36:37], -1
; GFX9-NEXT: buffer_load_dword v40, off, s[0:3], s33 ; 4-byte Folded Reload
+; GFX9-NEXT: buffer_load_dword v41, off, s[0:3], s33 offset:4 ; 4-byte Folded Reload
; GFX9-NEXT: s_mov_b64 exec, s[36:37]
; GFX9-NEXT: s_addk_i32 s32, 0xfc00
; GFX9-NEXT: s_mov_b32 s33, s34
@@ -2810,25 +3063,28 @@
; GFX10-NEXT: s_mov_b32 s33, s32
; GFX10-NEXT: s_or_saveexec_b32 s35, -1
; GFX10-NEXT: buffer_store_dword v40, off, s[0:3], s33 ; 4-byte Folded Spill
+; GFX10-NEXT: buffer_store_dword v41, off, s[0:3], s33 offset:4 ; 4-byte Folded Spill
; GFX10-NEXT: s_waitcnt_depctr 0xffe3
; GFX10-NEXT: s_mov_b32 exec_lo, s35
-; GFX10-NEXT: v_writelane_b32 v40, s34, 2
+; GFX10-NEXT: v_writelane_b32 v40, s30, 0
; GFX10-NEXT: v_mov_b32_e32 v0, 0
; GFX10-NEXT: v_mov_b32_e32 v1, 2.0
; GFX10-NEXT: v_mov_b32_e32 v2, 0
; GFX10-NEXT: v_mov_b32_e32 v3, 0x40100000
-; GFX10-NEXT: v_writelane_b32 v40, s30, 0
; GFX10-NEXT: s_addk_i32 s32, 0x200
+; GFX10-NEXT: v_writelane_b32 v41, s34, 0
+; GFX10-NEXT: v_writelane_b32 v40, s31, 1
; GFX10-NEXT: s_getpc_b64 s[34:35]
; GFX10-NEXT: s_add_u32 s34, s34, external_void_func_v2f64@rel32@lo+4
; GFX10-NEXT: s_addc_u32 s35, s35, external_void_func_v2f64@rel32@hi+12
-; GFX10-NEXT: v_writelane_b32 v40, s31, 1
; GFX10-NEXT: s_swappc_b64 s[30:31], s[34:35]
; GFX10-NEXT: v_readlane_b32 s31, v40, 1
; GFX10-NEXT: v_readlane_b32 s30, v40, 0
-; GFX10-NEXT: v_readlane_b32 s34, v40, 2
+; GFX10-NEXT: v_readlane_b32 s34, v41, 0
; GFX10-NEXT: s_or_saveexec_b32 s35, -1
-; GFX10-NEXT: buffer_load_dword v40, off, s[0:3], s33 ; 4-byte Folded Reload
+; GFX10-NEXT: s_clause 0x1
+; GFX10-NEXT: buffer_load_dword v40, off, s[0:3], s33
+; GFX10-NEXT: buffer_load_dword v41, off, s[0:3], s33 offset:4
; GFX10-NEXT: s_waitcnt_depctr 0xffe3
; GFX10-NEXT: s_mov_b32 exec_lo, s35
; GFX10-NEXT: s_addk_i32 s32, 0xfe00
@@ -2842,24 +3098,28 @@
; GFX11-NEXT: s_mov_b32 s0, s33
; GFX11-NEXT: s_mov_b32 s33, s32
; GFX11-NEXT: s_or_saveexec_b32 s1, -1
-; GFX11-NEXT: scratch_store_b32 off, v40, s33 ; 4-byte Folded Spill
+; GFX11-NEXT: s_clause 0x1
+; GFX11-NEXT: scratch_store_b32 off, v40, s33
+; GFX11-NEXT: scratch_store_b32 off, v41, s33 offset:4
; GFX11-NEXT: s_mov_b32 exec_lo, s1
-; GFX11-NEXT: v_writelane_b32 v40, s0, 2
+; GFX11-NEXT: v_writelane_b32 v40, s30, 0
; GFX11-NEXT: v_dual_mov_b32 v0, 0 :: v_dual_mov_b32 v1, 2.0
; GFX11-NEXT: v_dual_mov_b32 v2, 0 :: v_dual_mov_b32 v3, 0x40100000
-; GFX11-NEXT: v_writelane_b32 v40, s30, 0
; GFX11-NEXT: s_add_i32 s32, s32, 16
+; GFX11-NEXT: v_writelane_b32 v41, s0, 0
+; GFX11-NEXT: v_writelane_b32 v40, s31, 1
; GFX11-NEXT: s_getpc_b64 s[0:1]
; GFX11-NEXT: s_add_u32 s0, s0, external_void_func_v2f64@rel32@lo+4
; GFX11-NEXT: s_addc_u32 s1, s1, external_void_func_v2f64@rel32@hi+12
-; GFX11-NEXT: v_writelane_b32 v40, s31, 1
+; GFX11-NEXT: s_delay_alu instid0(SALU_CYCLE_1)
; GFX11-NEXT: s_swappc_b64 s[30:31], s[0:1]
-; GFX11-NEXT: s_delay_alu instid0(VALU_DEP_1)
; GFX11-NEXT: v_readlane_b32 s31, v40, 1
; GFX11-NEXT: v_readlane_b32 s30, v40, 0
-; GFX11-NEXT: v_readlane_b32 s0, v40, 2
+; GFX11-NEXT: v_readlane_b32 s0, v41, 0
; GFX11-NEXT: s_or_saveexec_b32 s1, -1
-; GFX11-NEXT: scratch_load_b32 v40, off, s33 ; 4-byte Folded Reload
+; GFX11-NEXT: s_clause 0x1
+; GFX11-NEXT: scratch_load_b32 v40, off, s33
+; GFX11-NEXT: scratch_load_b32 v41, off, s33 offset:4
; GFX11-NEXT: s_mov_b32 exec_lo, s1
; GFX11-NEXT: s_add_i32 s32, s32, -16
; GFX11-NEXT: s_mov_b32 s33, s0
@@ -2873,25 +3133,28 @@
; GFX10-SCRATCH-NEXT: s_mov_b32 s33, s32
; GFX10-SCRATCH-NEXT: s_or_saveexec_b32 s1, -1
; GFX10-SCRATCH-NEXT: scratch_store_dword off, v40, s33 ; 4-byte Folded Spill
+; GFX10-SCRATCH-NEXT: scratch_store_dword off, v41, s33 offset:4 ; 4-byte Folded Spill
; GFX10-SCRATCH-NEXT: s_waitcnt_depctr 0xffe3
; GFX10-SCRATCH-NEXT: s_mov_b32 exec_lo, s1
-; GFX10-SCRATCH-NEXT: v_writelane_b32 v40, s0, 2
+; GFX10-SCRATCH-NEXT: v_writelane_b32 v40, s30, 0
; GFX10-SCRATCH-NEXT: v_mov_b32_e32 v0, 0
; GFX10-SCRATCH-NEXT: v_mov_b32_e32 v1, 2.0
; GFX10-SCRATCH-NEXT: v_mov_b32_e32 v2, 0
; GFX10-SCRATCH-NEXT: v_mov_b32_e32 v3, 0x40100000
-; GFX10-SCRATCH-NEXT: v_writelane_b32 v40, s30, 0
; GFX10-SCRATCH-NEXT: s_add_i32 s32, s32, 16
+; GFX10-SCRATCH-NEXT: v_writelane_b32 v41, s0, 0
+; GFX10-SCRATCH-NEXT: v_writelane_b32 v40, s31, 1
; GFX10-SCRATCH-NEXT: s_getpc_b64 s[0:1]
; GFX10-SCRATCH-NEXT: s_add_u32 s0, s0, external_void_func_v2f64@rel32@lo+4
; GFX10-SCRATCH-NEXT: s_addc_u32 s1, s1, external_void_func_v2f64@rel32@hi+12
-; GFX10-SCRATCH-NEXT: v_writelane_b32 v40, s31, 1
; GFX10-SCRATCH-NEXT: s_swappc_b64 s[30:31], s[0:1]
; GFX10-SCRATCH-NEXT: v_readlane_b32 s31, v40, 1
; GFX10-SCRATCH-NEXT: v_readlane_b32 s30, v40, 0
-; GFX10-SCRATCH-NEXT: v_readlane_b32 s0, v40, 2
+; GFX10-SCRATCH-NEXT: v_readlane_b32 s0, v41, 0
; GFX10-SCRATCH-NEXT: s_or_saveexec_b32 s1, -1
-; GFX10-SCRATCH-NEXT: scratch_load_dword v40, off, s33 ; 4-byte Folded Reload
+; GFX10-SCRATCH-NEXT: s_clause 0x1
+; GFX10-SCRATCH-NEXT: scratch_load_dword v40, off, s33
+; GFX10-SCRATCH-NEXT: scratch_load_dword v41, off, s33 offset:4
; GFX10-SCRATCH-NEXT: s_waitcnt_depctr 0xffe3
; GFX10-SCRATCH-NEXT: s_mov_b32 exec_lo, s1
; GFX10-SCRATCH-NEXT: s_add_i32 s32, s32, -16
@@ -2910,8 +3173,8 @@
; GFX9-NEXT: s_mov_b32 s33, s32
; GFX9-NEXT: s_or_saveexec_b64 s[36:37], -1
; GFX9-NEXT: buffer_store_dword v40, off, s[0:3], s33 ; 4-byte Folded Spill
+; GFX9-NEXT: buffer_store_dword v41, off, s[0:3], s33 offset:4 ; 4-byte Folded Spill
; GFX9-NEXT: s_mov_b64 exec, s[36:37]
-; GFX9-NEXT: v_writelane_b32 v40, s34, 2
; GFX9-NEXT: s_addk_i32 s32, 0x400
; GFX9-NEXT: v_writelane_b32 v40, s30, 0
; GFX9-NEXT: v_mov_b32_e32 v0, 0
@@ -2920,6 +3183,7 @@
; GFX9-NEXT: v_mov_b32_e32 v3, 0x40100000
; GFX9-NEXT: v_mov_b32_e32 v4, 0
; GFX9-NEXT: v_mov_b32_e32 v5, 0x40200000
+; GFX9-NEXT: v_writelane_b32 v41, s34, 0
; GFX9-NEXT: v_writelane_b32 v40, s31, 1
; GFX9-NEXT: s_getpc_b64 s[34:35]
; GFX9-NEXT: s_add_u32 s34, s34, external_void_func_v3f64@rel32@lo+4
@@ -2927,9 +3191,10 @@
; GFX9-NEXT: s_swappc_b64 s[30:31], s[34:35]
; GFX9-NEXT: v_readlane_b32 s31, v40, 1
; GFX9-NEXT: v_readlane_b32 s30, v40, 0
-; GFX9-NEXT: v_readlane_b32 s34, v40, 2
+; GFX9-NEXT: v_readlane_b32 s34, v41, 0
; GFX9-NEXT: s_or_saveexec_b64 s[36:37], -1
; GFX9-NEXT: buffer_load_dword v40, off, s[0:3], s33 ; 4-byte Folded Reload
+; GFX9-NEXT: buffer_load_dword v41, off, s[0:3], s33 offset:4 ; 4-byte Folded Reload
; GFX9-NEXT: s_mov_b64 exec, s[36:37]
; GFX9-NEXT: s_addk_i32 s32, 0xfc00
; GFX9-NEXT: s_mov_b32 s33, s34
@@ -2943,27 +3208,30 @@
; GFX10-NEXT: s_mov_b32 s33, s32
; GFX10-NEXT: s_or_saveexec_b32 s35, -1
; GFX10-NEXT: buffer_store_dword v40, off, s[0:3], s33 ; 4-byte Folded Spill
+; GFX10-NEXT: buffer_store_dword v41, off, s[0:3], s33 offset:4 ; 4-byte Folded Spill
; GFX10-NEXT: s_waitcnt_depctr 0xffe3
; GFX10-NEXT: s_mov_b32 exec_lo, s35
-; GFX10-NEXT: v_writelane_b32 v40, s34, 2
+; GFX10-NEXT: v_writelane_b32 v40, s30, 0
; GFX10-NEXT: v_mov_b32_e32 v0, 0
; GFX10-NEXT: v_mov_b32_e32 v1, 2.0
; GFX10-NEXT: v_mov_b32_e32 v2, 0
; GFX10-NEXT: v_mov_b32_e32 v3, 0x40100000
-; GFX10-NEXT: v_writelane_b32 v40, s30, 0
; GFX10-NEXT: v_mov_b32_e32 v4, 0
; GFX10-NEXT: v_mov_b32_e32 v5, 0x40200000
; GFX10-NEXT: s_addk_i32 s32, 0x200
+; GFX10-NEXT: v_writelane_b32 v41, s34, 0
+; GFX10-NEXT: v_writelane_b32 v40, s31, 1
; GFX10-NEXT: s_getpc_b64 s[34:35]
; GFX10-NEXT: s_add_u32 s34, s34, external_void_func_v3f64@rel32@lo+4
; GFX10-NEXT: s_addc_u32 s35, s35, external_void_func_v3f64@rel32@hi+12
-; GFX10-NEXT: v_writelane_b32 v40, s31, 1
; GFX10-NEXT: s_swappc_b64 s[30:31], s[34:35]
; GFX10-NEXT: v_readlane_b32 s31, v40, 1
; GFX10-NEXT: v_readlane_b32 s30, v40, 0
-; GFX10-NEXT: v_readlane_b32 s34, v40, 2
+; GFX10-NEXT: v_readlane_b32 s34, v41, 0
; GFX10-NEXT: s_or_saveexec_b32 s35, -1
-; GFX10-NEXT: buffer_load_dword v40, off, s[0:3], s33 ; 4-byte Folded Reload
+; GFX10-NEXT: s_clause 0x1
+; GFX10-NEXT: buffer_load_dword v40, off, s[0:3], s33
+; GFX10-NEXT: buffer_load_dword v41, off, s[0:3], s33 offset:4
; GFX10-NEXT: s_waitcnt_depctr 0xffe3
; GFX10-NEXT: s_mov_b32 exec_lo, s35
; GFX10-NEXT: s_addk_i32 s32, 0xfe00
@@ -2977,25 +3245,29 @@
; GFX11-NEXT: s_mov_b32 s0, s33
; GFX11-NEXT: s_mov_b32 s33, s32
; GFX11-NEXT: s_or_saveexec_b32 s1, -1
-; GFX11-NEXT: scratch_store_b32 off, v40, s33 ; 4-byte Folded Spill
+; GFX11-NEXT: s_clause 0x1
+; GFX11-NEXT: scratch_store_b32 off, v40, s33
+; GFX11-NEXT: scratch_store_b32 off, v41, s33 offset:4
; GFX11-NEXT: s_mov_b32 exec_lo, s1
-; GFX11-NEXT: v_writelane_b32 v40, s0, 2
+; GFX11-NEXT: v_writelane_b32 v40, s30, 0
; GFX11-NEXT: v_dual_mov_b32 v0, 0 :: v_dual_mov_b32 v1, 2.0
; GFX11-NEXT: v_dual_mov_b32 v2, 0 :: v_dual_mov_b32 v3, 0x40100000
-; GFX11-NEXT: v_writelane_b32 v40, s30, 0
; GFX11-NEXT: v_dual_mov_b32 v4, 0 :: v_dual_mov_b32 v5, 0x40200000
; GFX11-NEXT: s_add_i32 s32, s32, 16
+; GFX11-NEXT: v_writelane_b32 v41, s0, 0
+; GFX11-NEXT: v_writelane_b32 v40, s31, 1
; GFX11-NEXT: s_getpc_b64 s[0:1]
; GFX11-NEXT: s_add_u32 s0, s0, external_void_func_v3f64@rel32@lo+4
; GFX11-NEXT: s_addc_u32 s1, s1, external_void_func_v3f64@rel32@hi+12
-; GFX11-NEXT: v_writelane_b32 v40, s31, 1
+; GFX11-NEXT: s_delay_alu instid0(SALU_CYCLE_1)
; GFX11-NEXT: s_swappc_b64 s[30:31], s[0:1]
-; GFX11-NEXT: s_delay_alu instid0(VALU_DEP_1)
; GFX11-NEXT: v_readlane_b32 s31, v40, 1
; GFX11-NEXT: v_readlane_b32 s30, v40, 0
-; GFX11-NEXT: v_readlane_b32 s0, v40, 2
+; GFX11-NEXT: v_readlane_b32 s0, v41, 0
; GFX11-NEXT: s_or_saveexec_b32 s1, -1
-; GFX11-NEXT: scratch_load_b32 v40, off, s33 ; 4-byte Folded Reload
+; GFX11-NEXT: s_clause 0x1
+; GFX11-NEXT: scratch_load_b32 v40, off, s33
+; GFX11-NEXT: scratch_load_b32 v41, off, s33 offset:4
; GFX11-NEXT: s_mov_b32 exec_lo, s1
; GFX11-NEXT: s_add_i32 s32, s32, -16
; GFX11-NEXT: s_mov_b32 s33, s0
@@ -3009,27 +3281,30 @@
; GFX10-SCRATCH-NEXT: s_mov_b32 s33, s32
; GFX10-SCRATCH-NEXT: s_or_saveexec_b32 s1, -1
; GFX10-SCRATCH-NEXT: scratch_store_dword off, v40, s33 ; 4-byte Folded Spill
+; GFX10-SCRATCH-NEXT: scratch_store_dword off, v41, s33 offset:4 ; 4-byte Folded Spill
; GFX10-SCRATCH-NEXT: s_waitcnt_depctr 0xffe3
; GFX10-SCRATCH-NEXT: s_mov_b32 exec_lo, s1
-; GFX10-SCRATCH-NEXT: v_writelane_b32 v40, s0, 2
+; GFX10-SCRATCH-NEXT: v_writelane_b32 v40, s30, 0
; GFX10-SCRATCH-NEXT: v_mov_b32_e32 v0, 0
; GFX10-SCRATCH-NEXT: v_mov_b32_e32 v1, 2.0
; GFX10-SCRATCH-NEXT: v_mov_b32_e32 v2, 0
; GFX10-SCRATCH-NEXT: v_mov_b32_e32 v3, 0x40100000
-; GFX10-SCRATCH-NEXT: v_writelane_b32 v40, s30, 0
; GFX10-SCRATCH-NEXT: v_mov_b32_e32 v4, 0
; GFX10-SCRATCH-NEXT: v_mov_b32_e32 v5, 0x40200000
; GFX10-SCRATCH-NEXT: s_add_i32 s32, s32, 16
+; GFX10-SCRATCH-NEXT: v_writelane_b32 v41, s0, 0
+; GFX10-SCRATCH-NEXT: v_writelane_b32 v40, s31, 1
; GFX10-SCRATCH-NEXT: s_getpc_b64 s[0:1]
; GFX10-SCRATCH-NEXT: s_add_u32 s0, s0, external_void_func_v3f64@rel32@lo+4
; GFX10-SCRATCH-NEXT: s_addc_u32 s1, s1, external_void_func_v3f64@rel32@hi+12
-; GFX10-SCRATCH-NEXT: v_writelane_b32 v40, s31, 1
; GFX10-SCRATCH-NEXT: s_swappc_b64 s[30:31], s[0:1]
; GFX10-SCRATCH-NEXT: v_readlane_b32 s31, v40, 1
; GFX10-SCRATCH-NEXT: v_readlane_b32 s30, v40, 0
-; GFX10-SCRATCH-NEXT: v_readlane_b32 s0, v40, 2
+; GFX10-SCRATCH-NEXT: v_readlane_b32 s0, v41, 0
; GFX10-SCRATCH-NEXT: s_or_saveexec_b32 s1, -1
-; GFX10-SCRATCH-NEXT: scratch_load_dword v40, off, s33 ; 4-byte Folded Reload
+; GFX10-SCRATCH-NEXT: s_clause 0x1
+; GFX10-SCRATCH-NEXT: scratch_load_dword v40, off, s33
+; GFX10-SCRATCH-NEXT: scratch_load_dword v41, off, s33 offset:4
; GFX10-SCRATCH-NEXT: s_waitcnt_depctr 0xffe3
; GFX10-SCRATCH-NEXT: s_mov_b32 exec_lo, s1
; GFX10-SCRATCH-NEXT: s_add_i32 s32, s32, -16
@@ -3048,11 +3323,12 @@
; GFX9-NEXT: s_mov_b32 s33, s32
; GFX9-NEXT: s_or_saveexec_b64 s[36:37], -1
; GFX9-NEXT: buffer_store_dword v40, off, s[0:3], s33 ; 4-byte Folded Spill
+; GFX9-NEXT: buffer_store_dword v41, off, s[0:3], s33 offset:4 ; 4-byte Folded Spill
; GFX9-NEXT: s_mov_b64 exec, s[36:37]
; GFX9-NEXT: global_load_dword v0, v[0:1], off
-; GFX9-NEXT: v_writelane_b32 v40, s34, 2
; GFX9-NEXT: s_addk_i32 s32, 0x400
; GFX9-NEXT: v_writelane_b32 v40, s30, 0
+; GFX9-NEXT: v_writelane_b32 v41, s34, 0
; GFX9-NEXT: v_writelane_b32 v40, s31, 1
; GFX9-NEXT: s_getpc_b64 s[34:35]
; GFX9-NEXT: s_add_u32 s34, s34, external_void_func_v2i16@rel32@lo+4
@@ -3060,9 +3336,10 @@
; GFX9-NEXT: s_swappc_b64 s[30:31], s[34:35]
; GFX9-NEXT: v_readlane_b32 s31, v40, 1
; GFX9-NEXT: v_readlane_b32 s30, v40, 0
-; GFX9-NEXT: v_readlane_b32 s34, v40, 2
+; GFX9-NEXT: v_readlane_b32 s34, v41, 0
; GFX9-NEXT: s_or_saveexec_b64 s[36:37], -1
; GFX9-NEXT: buffer_load_dword v40, off, s[0:3], s33 ; 4-byte Folded Reload
+; GFX9-NEXT: buffer_load_dword v41, off, s[0:3], s33 offset:4 ; 4-byte Folded Reload
; GFX9-NEXT: s_mov_b64 exec, s[36:37]
; GFX9-NEXT: s_addk_i32 s32, 0xfc00
; GFX9-NEXT: s_mov_b32 s33, s34
@@ -3076,22 +3353,25 @@
; GFX10-NEXT: s_mov_b32 s33, s32
; GFX10-NEXT: s_or_saveexec_b32 s35, -1
; GFX10-NEXT: buffer_store_dword v40, off, s[0:3], s33 ; 4-byte Folded Spill
+; GFX10-NEXT: buffer_store_dword v41, off, s[0:3], s33 offset:4 ; 4-byte Folded Spill
; GFX10-NEXT: s_waitcnt_depctr 0xffe3
; GFX10-NEXT: s_mov_b32 exec_lo, s35
; GFX10-NEXT: global_load_dword v0, v[0:1], off
-; GFX10-NEXT: v_writelane_b32 v40, s34, 2
+; GFX10-NEXT: v_writelane_b32 v40, s30, 0
; GFX10-NEXT: s_addk_i32 s32, 0x200
+; GFX10-NEXT: v_writelane_b32 v41, s34, 0
; GFX10-NEXT: s_getpc_b64 s[34:35]
; GFX10-NEXT: s_add_u32 s34, s34, external_void_func_v2i16@rel32@lo+4
; GFX10-NEXT: s_addc_u32 s35, s35, external_void_func_v2i16@rel32@hi+12
-; GFX10-NEXT: v_writelane_b32 v40, s30, 0
; GFX10-NEXT: v_writelane_b32 v40, s31, 1
; GFX10-NEXT: s_swappc_b64 s[30:31], s[34:35]
; GFX10-NEXT: v_readlane_b32 s31, v40, 1
; GFX10-NEXT: v_readlane_b32 s30, v40, 0
-; GFX10-NEXT: v_readlane_b32 s34, v40, 2
+; GFX10-NEXT: v_readlane_b32 s34, v41, 0
; GFX10-NEXT: s_or_saveexec_b32 s35, -1
-; GFX10-NEXT: buffer_load_dword v40, off, s[0:3], s33 ; 4-byte Folded Reload
+; GFX10-NEXT: s_clause 0x1
+; GFX10-NEXT: buffer_load_dword v40, off, s[0:3], s33
+; GFX10-NEXT: buffer_load_dword v41, off, s[0:3], s33 offset:4
; GFX10-NEXT: s_waitcnt_depctr 0xffe3
; GFX10-NEXT: s_mov_b32 exec_lo, s35
; GFX10-NEXT: s_addk_i32 s32, 0xfe00
@@ -3105,23 +3385,27 @@
; GFX11-NEXT: s_mov_b32 s0, s33
; GFX11-NEXT: s_mov_b32 s33, s32
; GFX11-NEXT: s_or_saveexec_b32 s1, -1
-; GFX11-NEXT: scratch_store_b32 off, v40, s33 ; 4-byte Folded Spill
+; GFX11-NEXT: s_clause 0x1
+; GFX11-NEXT: scratch_store_b32 off, v40, s33
+; GFX11-NEXT: scratch_store_b32 off, v41, s33 offset:4
; GFX11-NEXT: s_mov_b32 exec_lo, s1
; GFX11-NEXT: global_load_b32 v0, v[0:1], off
-; GFX11-NEXT: v_writelane_b32 v40, s0, 2
+; GFX11-NEXT: v_writelane_b32 v40, s30, 0
; GFX11-NEXT: s_add_i32 s32, s32, 16
+; GFX11-NEXT: v_writelane_b32 v41, s0, 0
; GFX11-NEXT: s_getpc_b64 s[0:1]
; GFX11-NEXT: s_add_u32 s0, s0, external_void_func_v2i16@rel32@lo+4
; GFX11-NEXT: s_addc_u32 s1, s1, external_void_func_v2i16@rel32@hi+12
-; GFX11-NEXT: v_writelane_b32 v40, s30, 0
; GFX11-NEXT: v_writelane_b32 v40, s31, 1
; GFX11-NEXT: s_swappc_b64 s[30:31], s[0:1]
; GFX11-NEXT: s_delay_alu instid0(VALU_DEP_1)
; GFX11-NEXT: v_readlane_b32 s31, v40, 1
; GFX11-NEXT: v_readlane_b32 s30, v40, 0
-; GFX11-NEXT: v_readlane_b32 s0, v40, 2
+; GFX11-NEXT: v_readlane_b32 s0, v41, 0
; GFX11-NEXT: s_or_saveexec_b32 s1, -1
-; GFX11-NEXT: scratch_load_b32 v40, off, s33 ; 4-byte Folded Reload
+; GFX11-NEXT: s_clause 0x1
+; GFX11-NEXT: scratch_load_b32 v40, off, s33
+; GFX11-NEXT: scratch_load_b32 v41, off, s33 offset:4
; GFX11-NEXT: s_mov_b32 exec_lo, s1
; GFX11-NEXT: s_add_i32 s32, s32, -16
; GFX11-NEXT: s_mov_b32 s33, s0
@@ -3135,22 +3419,25 @@
; GFX10-SCRATCH-NEXT: s_mov_b32 s33, s32
; GFX10-SCRATCH-NEXT: s_or_saveexec_b32 s1, -1
; GFX10-SCRATCH-NEXT: scratch_store_dword off, v40, s33 ; 4-byte Folded Spill
+; GFX10-SCRATCH-NEXT: scratch_store_dword off, v41, s33 offset:4 ; 4-byte Folded Spill
; GFX10-SCRATCH-NEXT: s_waitcnt_depctr 0xffe3
; GFX10-SCRATCH-NEXT: s_mov_b32 exec_lo, s1
; GFX10-SCRATCH-NEXT: global_load_dword v0, v[0:1], off
-; GFX10-SCRATCH-NEXT: v_writelane_b32 v40, s0, 2
+; GFX10-SCRATCH-NEXT: v_writelane_b32 v40, s30, 0
; GFX10-SCRATCH-NEXT: s_add_i32 s32, s32, 16
+; GFX10-SCRATCH-NEXT: v_writelane_b32 v41, s0, 0
; GFX10-SCRATCH-NEXT: s_getpc_b64 s[0:1]
; GFX10-SCRATCH-NEXT: s_add_u32 s0, s0, external_void_func_v2i16@rel32@lo+4
; GFX10-SCRATCH-NEXT: s_addc_u32 s1, s1, external_void_func_v2i16@rel32@hi+12
-; GFX10-SCRATCH-NEXT: v_writelane_b32 v40, s30, 0
; GFX10-SCRATCH-NEXT: v_writelane_b32 v40, s31, 1
; GFX10-SCRATCH-NEXT: s_swappc_b64 s[30:31], s[0:1]
; GFX10-SCRATCH-NEXT: v_readlane_b32 s31, v40, 1
; GFX10-SCRATCH-NEXT: v_readlane_b32 s30, v40, 0
-; GFX10-SCRATCH-NEXT: v_readlane_b32 s0, v40, 2
+; GFX10-SCRATCH-NEXT: v_readlane_b32 s0, v41, 0
; GFX10-SCRATCH-NEXT: s_or_saveexec_b32 s1, -1
-; GFX10-SCRATCH-NEXT: scratch_load_dword v40, off, s33 ; 4-byte Folded Reload
+; GFX10-SCRATCH-NEXT: s_clause 0x1
+; GFX10-SCRATCH-NEXT: scratch_load_dword v40, off, s33
+; GFX10-SCRATCH-NEXT: scratch_load_dword v41, off, s33 offset:4
; GFX10-SCRATCH-NEXT: s_waitcnt_depctr 0xffe3
; GFX10-SCRATCH-NEXT: s_mov_b32 exec_lo, s1
; GFX10-SCRATCH-NEXT: s_add_i32 s32, s32, -16
@@ -3170,11 +3457,12 @@
; GFX9-NEXT: s_mov_b32 s33, s32
; GFX9-NEXT: s_or_saveexec_b64 s[36:37], -1
; GFX9-NEXT: buffer_store_dword v40, off, s[0:3], s33 ; 4-byte Folded Spill
+; GFX9-NEXT: buffer_store_dword v41, off, s[0:3], s33 offset:4 ; 4-byte Folded Spill
; GFX9-NEXT: s_mov_b64 exec, s[36:37]
; GFX9-NEXT: global_load_dwordx2 v[0:1], v[0:1], off
-; GFX9-NEXT: v_writelane_b32 v40, s34, 2
; GFX9-NEXT: s_addk_i32 s32, 0x400
; GFX9-NEXT: v_writelane_b32 v40, s30, 0
+; GFX9-NEXT: v_writelane_b32 v41, s34, 0
; GFX9-NEXT: v_writelane_b32 v40, s31, 1
; GFX9-NEXT: s_getpc_b64 s[34:35]
; GFX9-NEXT: s_add_u32 s34, s34, external_void_func_v3i16@rel32@lo+4
@@ -3182,9 +3470,10 @@
; GFX9-NEXT: s_swappc_b64 s[30:31], s[34:35]
; GFX9-NEXT: v_readlane_b32 s31, v40, 1
; GFX9-NEXT: v_readlane_b32 s30, v40, 0
-; GFX9-NEXT: v_readlane_b32 s34, v40, 2
+; GFX9-NEXT: v_readlane_b32 s34, v41, 0
; GFX9-NEXT: s_or_saveexec_b64 s[36:37], -1
; GFX9-NEXT: buffer_load_dword v40, off, s[0:3], s33 ; 4-byte Folded Reload
+; GFX9-NEXT: buffer_load_dword v41, off, s[0:3], s33 offset:4 ; 4-byte Folded Reload
; GFX9-NEXT: s_mov_b64 exec, s[36:37]
; GFX9-NEXT: s_addk_i32 s32, 0xfc00
; GFX9-NEXT: s_mov_b32 s33, s34
@@ -3198,22 +3487,25 @@
; GFX10-NEXT: s_mov_b32 s33, s32
; GFX10-NEXT: s_or_saveexec_b32 s35, -1
; GFX10-NEXT: buffer_store_dword v40, off, s[0:3], s33 ; 4-byte Folded Spill
+; GFX10-NEXT: buffer_store_dword v41, off, s[0:3], s33 offset:4 ; 4-byte Folded Spill
; GFX10-NEXT: s_waitcnt_depctr 0xffe3
; GFX10-NEXT: s_mov_b32 exec_lo, s35
; GFX10-NEXT: global_load_dwordx2 v[0:1], v[0:1], off
-; GFX10-NEXT: v_writelane_b32 v40, s34, 2
+; GFX10-NEXT: v_writelane_b32 v40, s30, 0
; GFX10-NEXT: s_addk_i32 s32, 0x200
+; GFX10-NEXT: v_writelane_b32 v41, s34, 0
; GFX10-NEXT: s_getpc_b64 s[34:35]
; GFX10-NEXT: s_add_u32 s34, s34, external_void_func_v3i16@rel32@lo+4
; GFX10-NEXT: s_addc_u32 s35, s35, external_void_func_v3i16@rel32@hi+12
-; GFX10-NEXT: v_writelane_b32 v40, s30, 0
; GFX10-NEXT: v_writelane_b32 v40, s31, 1
; GFX10-NEXT: s_swappc_b64 s[30:31], s[34:35]
; GFX10-NEXT: v_readlane_b32 s31, v40, 1
; GFX10-NEXT: v_readlane_b32 s30, v40, 0
-; GFX10-NEXT: v_readlane_b32 s34, v40, 2
+; GFX10-NEXT: v_readlane_b32 s34, v41, 0
; GFX10-NEXT: s_or_saveexec_b32 s35, -1
-; GFX10-NEXT: buffer_load_dword v40, off, s[0:3], s33 ; 4-byte Folded Reload
+; GFX10-NEXT: s_clause 0x1
+; GFX10-NEXT: buffer_load_dword v40, off, s[0:3], s33
+; GFX10-NEXT: buffer_load_dword v41, off, s[0:3], s33 offset:4
; GFX10-NEXT: s_waitcnt_depctr 0xffe3
; GFX10-NEXT: s_mov_b32 exec_lo, s35
; GFX10-NEXT: s_addk_i32 s32, 0xfe00
@@ -3227,23 +3519,27 @@
; GFX11-NEXT: s_mov_b32 s0, s33
; GFX11-NEXT: s_mov_b32 s33, s32
; GFX11-NEXT: s_or_saveexec_b32 s1, -1
-; GFX11-NEXT: scratch_store_b32 off, v40, s33 ; 4-byte Folded Spill
+; GFX11-NEXT: s_clause 0x1
+; GFX11-NEXT: scratch_store_b32 off, v40, s33
+; GFX11-NEXT: scratch_store_b32 off, v41, s33 offset:4
; GFX11-NEXT: s_mov_b32 exec_lo, s1
; GFX11-NEXT: global_load_b64 v[0:1], v[0:1], off
-; GFX11-NEXT: v_writelane_b32 v40, s0, 2
+; GFX11-NEXT: v_writelane_b32 v40, s30, 0
; GFX11-NEXT: s_add_i32 s32, s32, 16
+; GFX11-NEXT: v_writelane_b32 v41, s0, 0
; GFX11-NEXT: s_getpc_b64 s[0:1]
; GFX11-NEXT: s_add_u32 s0, s0, external_void_func_v3i16@rel32@lo+4
; GFX11-NEXT: s_addc_u32 s1, s1, external_void_func_v3i16@rel32@hi+12
-; GFX11-NEXT: v_writelane_b32 v40, s30, 0
; GFX11-NEXT: v_writelane_b32 v40, s31, 1
; GFX11-NEXT: s_swappc_b64 s[30:31], s[0:1]
; GFX11-NEXT: s_delay_alu instid0(VALU_DEP_1)
; GFX11-NEXT: v_readlane_b32 s31, v40, 1
; GFX11-NEXT: v_readlane_b32 s30, v40, 0
-; GFX11-NEXT: v_readlane_b32 s0, v40, 2
+; GFX11-NEXT: v_readlane_b32 s0, v41, 0
; GFX11-NEXT: s_or_saveexec_b32 s1, -1
-; GFX11-NEXT: scratch_load_b32 v40, off, s33 ; 4-byte Folded Reload
+; GFX11-NEXT: s_clause 0x1
+; GFX11-NEXT: scratch_load_b32 v40, off, s33
+; GFX11-NEXT: scratch_load_b32 v41, off, s33 offset:4
; GFX11-NEXT: s_mov_b32 exec_lo, s1
; GFX11-NEXT: s_add_i32 s32, s32, -16
; GFX11-NEXT: s_mov_b32 s33, s0
@@ -3257,22 +3553,25 @@
; GFX10-SCRATCH-NEXT: s_mov_b32 s33, s32
; GFX10-SCRATCH-NEXT: s_or_saveexec_b32 s1, -1
; GFX10-SCRATCH-NEXT: scratch_store_dword off, v40, s33 ; 4-byte Folded Spill
+; GFX10-SCRATCH-NEXT: scratch_store_dword off, v41, s33 offset:4 ; 4-byte Folded Spill
; GFX10-SCRATCH-NEXT: s_waitcnt_depctr 0xffe3
; GFX10-SCRATCH-NEXT: s_mov_b32 exec_lo, s1
; GFX10-SCRATCH-NEXT: global_load_dwordx2 v[0:1], v[0:1], off
-; GFX10-SCRATCH-NEXT: v_writelane_b32 v40, s0, 2
+; GFX10-SCRATCH-NEXT: v_writelane_b32 v40, s30, 0
; GFX10-SCRATCH-NEXT: s_add_i32 s32, s32, 16
+; GFX10-SCRATCH-NEXT: v_writelane_b32 v41, s0, 0
; GFX10-SCRATCH-NEXT: s_getpc_b64 s[0:1]
; GFX10-SCRATCH-NEXT: s_add_u32 s0, s0, external_void_func_v3i16@rel32@lo+4
; GFX10-SCRATCH-NEXT: s_addc_u32 s1, s1, external_void_func_v3i16@rel32@hi+12
-; GFX10-SCRATCH-NEXT: v_writelane_b32 v40, s30, 0
; GFX10-SCRATCH-NEXT: v_writelane_b32 v40, s31, 1
; GFX10-SCRATCH-NEXT: s_swappc_b64 s[30:31], s[0:1]
; GFX10-SCRATCH-NEXT: v_readlane_b32 s31, v40, 1
; GFX10-SCRATCH-NEXT: v_readlane_b32 s30, v40, 0
-; GFX10-SCRATCH-NEXT: v_readlane_b32 s0, v40, 2
+; GFX10-SCRATCH-NEXT: v_readlane_b32 s0, v41, 0
; GFX10-SCRATCH-NEXT: s_or_saveexec_b32 s1, -1
-; GFX10-SCRATCH-NEXT: scratch_load_dword v40, off, s33 ; 4-byte Folded Reload
+; GFX10-SCRATCH-NEXT: s_clause 0x1
+; GFX10-SCRATCH-NEXT: scratch_load_dword v40, off, s33
+; GFX10-SCRATCH-NEXT: scratch_load_dword v41, off, s33 offset:4
; GFX10-SCRATCH-NEXT: s_waitcnt_depctr 0xffe3
; GFX10-SCRATCH-NEXT: s_mov_b32 exec_lo, s1
; GFX10-SCRATCH-NEXT: s_add_i32 s32, s32, -16
@@ -3292,11 +3591,12 @@
; GFX9-NEXT: s_mov_b32 s33, s32
; GFX9-NEXT: s_or_saveexec_b64 s[36:37], -1
; GFX9-NEXT: buffer_store_dword v40, off, s[0:3], s33 ; 4-byte Folded Spill
+; GFX9-NEXT: buffer_store_dword v41, off, s[0:3], s33 offset:4 ; 4-byte Folded Spill
; GFX9-NEXT: s_mov_b64 exec, s[36:37]
; GFX9-NEXT: global_load_dwordx2 v[0:1], v[0:1], off
-; GFX9-NEXT: v_writelane_b32 v40, s34, 2
; GFX9-NEXT: s_addk_i32 s32, 0x400
; GFX9-NEXT: v_writelane_b32 v40, s30, 0
+; GFX9-NEXT: v_writelane_b32 v41, s34, 0
; GFX9-NEXT: v_writelane_b32 v40, s31, 1
; GFX9-NEXT: s_getpc_b64 s[34:35]
; GFX9-NEXT: s_add_u32 s34, s34, external_void_func_v3f16@rel32@lo+4
@@ -3304,9 +3604,10 @@
; GFX9-NEXT: s_swappc_b64 s[30:31], s[34:35]
; GFX9-NEXT: v_readlane_b32 s31, v40, 1
; GFX9-NEXT: v_readlane_b32 s30, v40, 0
-; GFX9-NEXT: v_readlane_b32 s34, v40, 2
+; GFX9-NEXT: v_readlane_b32 s34, v41, 0
; GFX9-NEXT: s_or_saveexec_b64 s[36:37], -1
; GFX9-NEXT: buffer_load_dword v40, off, s[0:3], s33 ; 4-byte Folded Reload
+; GFX9-NEXT: buffer_load_dword v41, off, s[0:3], s33 offset:4 ; 4-byte Folded Reload
; GFX9-NEXT: s_mov_b64 exec, s[36:37]
; GFX9-NEXT: s_addk_i32 s32, 0xfc00
; GFX9-NEXT: s_mov_b32 s33, s34
@@ -3320,22 +3621,25 @@
; GFX10-NEXT: s_mov_b32 s33, s32
; GFX10-NEXT: s_or_saveexec_b32 s35, -1
; GFX10-NEXT: buffer_store_dword v40, off, s[0:3], s33 ; 4-byte Folded Spill
+; GFX10-NEXT: buffer_store_dword v41, off, s[0:3], s33 offset:4 ; 4-byte Folded Spill
; GFX10-NEXT: s_waitcnt_depctr 0xffe3
; GFX10-NEXT: s_mov_b32 exec_lo, s35
; GFX10-NEXT: global_load_dwordx2 v[0:1], v[0:1], off
-; GFX10-NEXT: v_writelane_b32 v40, s34, 2
+; GFX10-NEXT: v_writelane_b32 v40, s30, 0
; GFX10-NEXT: s_addk_i32 s32, 0x200
+; GFX10-NEXT: v_writelane_b32 v41, s34, 0
; GFX10-NEXT: s_getpc_b64 s[34:35]
; GFX10-NEXT: s_add_u32 s34, s34, external_void_func_v3f16@rel32@lo+4
; GFX10-NEXT: s_addc_u32 s35, s35, external_void_func_v3f16@rel32@hi+12
-; GFX10-NEXT: v_writelane_b32 v40, s30, 0
; GFX10-NEXT: v_writelane_b32 v40, s31, 1
; GFX10-NEXT: s_swappc_b64 s[30:31], s[34:35]
; GFX10-NEXT: v_readlane_b32 s31, v40, 1
; GFX10-NEXT: v_readlane_b32 s30, v40, 0
-; GFX10-NEXT: v_readlane_b32 s34, v40, 2
+; GFX10-NEXT: v_readlane_b32 s34, v41, 0
; GFX10-NEXT: s_or_saveexec_b32 s35, -1
-; GFX10-NEXT: buffer_load_dword v40, off, s[0:3], s33 ; 4-byte Folded Reload
+; GFX10-NEXT: s_clause 0x1
+; GFX10-NEXT: buffer_load_dword v40, off, s[0:3], s33
+; GFX10-NEXT: buffer_load_dword v41, off, s[0:3], s33 offset:4
; GFX10-NEXT: s_waitcnt_depctr 0xffe3
; GFX10-NEXT: s_mov_b32 exec_lo, s35
; GFX10-NEXT: s_addk_i32 s32, 0xfe00
@@ -3349,23 +3653,27 @@
; GFX11-NEXT: s_mov_b32 s0, s33
; GFX11-NEXT: s_mov_b32 s33, s32
; GFX11-NEXT: s_or_saveexec_b32 s1, -1
-; GFX11-NEXT: scratch_store_b32 off, v40, s33 ; 4-byte Folded Spill
+; GFX11-NEXT: s_clause 0x1
+; GFX11-NEXT: scratch_store_b32 off, v40, s33
+; GFX11-NEXT: scratch_store_b32 off, v41, s33 offset:4
; GFX11-NEXT: s_mov_b32 exec_lo, s1
; GFX11-NEXT: global_load_b64 v[0:1], v[0:1], off
-; GFX11-NEXT: v_writelane_b32 v40, s0, 2
+; GFX11-NEXT: v_writelane_b32 v40, s30, 0
; GFX11-NEXT: s_add_i32 s32, s32, 16
+; GFX11-NEXT: v_writelane_b32 v41, s0, 0
; GFX11-NEXT: s_getpc_b64 s[0:1]
; GFX11-NEXT: s_add_u32 s0, s0, external_void_func_v3f16@rel32@lo+4
; GFX11-NEXT: s_addc_u32 s1, s1, external_void_func_v3f16@rel32@hi+12
-; GFX11-NEXT: v_writelane_b32 v40, s30, 0
; GFX11-NEXT: v_writelane_b32 v40, s31, 1
; GFX11-NEXT: s_swappc_b64 s[30:31], s[0:1]
; GFX11-NEXT: s_delay_alu instid0(VALU_DEP_1)
; GFX11-NEXT: v_readlane_b32 s31, v40, 1
; GFX11-NEXT: v_readlane_b32 s30, v40, 0
-; GFX11-NEXT: v_readlane_b32 s0, v40, 2
+; GFX11-NEXT: v_readlane_b32 s0, v41, 0
; GFX11-NEXT: s_or_saveexec_b32 s1, -1
-; GFX11-NEXT: scratch_load_b32 v40, off, s33 ; 4-byte Folded Reload
+; GFX11-NEXT: s_clause 0x1
+; GFX11-NEXT: scratch_load_b32 v40, off, s33
+; GFX11-NEXT: scratch_load_b32 v41, off, s33 offset:4
; GFX11-NEXT: s_mov_b32 exec_lo, s1
; GFX11-NEXT: s_add_i32 s32, s32, -16
; GFX11-NEXT: s_mov_b32 s33, s0
@@ -3379,22 +3687,25 @@
; GFX10-SCRATCH-NEXT: s_mov_b32 s33, s32
; GFX10-SCRATCH-NEXT: s_or_saveexec_b32 s1, -1
; GFX10-SCRATCH-NEXT: scratch_store_dword off, v40, s33 ; 4-byte Folded Spill
+; GFX10-SCRATCH-NEXT: scratch_store_dword off, v41, s33 offset:4 ; 4-byte Folded Spill
; GFX10-SCRATCH-NEXT: s_waitcnt_depctr 0xffe3
; GFX10-SCRATCH-NEXT: s_mov_b32 exec_lo, s1
; GFX10-SCRATCH-NEXT: global_load_dwordx2 v[0:1], v[0:1], off
-; GFX10-SCRATCH-NEXT: v_writelane_b32 v40, s0, 2
+; GFX10-SCRATCH-NEXT: v_writelane_b32 v40, s30, 0
; GFX10-SCRATCH-NEXT: s_add_i32 s32, s32, 16
+; GFX10-SCRATCH-NEXT: v_writelane_b32 v41, s0, 0
; GFX10-SCRATCH-NEXT: s_getpc_b64 s[0:1]
; GFX10-SCRATCH-NEXT: s_add_u32 s0, s0, external_void_func_v3f16@rel32@lo+4
; GFX10-SCRATCH-NEXT: s_addc_u32 s1, s1, external_void_func_v3f16@rel32@hi+12
-; GFX10-SCRATCH-NEXT: v_writelane_b32 v40, s30, 0
; GFX10-SCRATCH-NEXT: v_writelane_b32 v40, s31, 1
; GFX10-SCRATCH-NEXT: s_swappc_b64 s[30:31], s[0:1]
; GFX10-SCRATCH-NEXT: v_readlane_b32 s31, v40, 1
; GFX10-SCRATCH-NEXT: v_readlane_b32 s30, v40, 0
-; GFX10-SCRATCH-NEXT: v_readlane_b32 s0, v40, 2
+; GFX10-SCRATCH-NEXT: v_readlane_b32 s0, v41, 0
; GFX10-SCRATCH-NEXT: s_or_saveexec_b32 s1, -1
-; GFX10-SCRATCH-NEXT: scratch_load_dword v40, off, s33 ; 4-byte Folded Reload
+; GFX10-SCRATCH-NEXT: s_clause 0x1
+; GFX10-SCRATCH-NEXT: scratch_load_dword v40, off, s33
+; GFX10-SCRATCH-NEXT: scratch_load_dword v41, off, s33 offset:4
; GFX10-SCRATCH-NEXT: s_waitcnt_depctr 0xffe3
; GFX10-SCRATCH-NEXT: s_mov_b32 exec_lo, s1
; GFX10-SCRATCH-NEXT: s_add_i32 s32, s32, -16
@@ -3414,12 +3725,13 @@
; GFX9-NEXT: s_mov_b32 s33, s32
; GFX9-NEXT: s_or_saveexec_b64 s[36:37], -1
; GFX9-NEXT: buffer_store_dword v40, off, s[0:3], s33 ; 4-byte Folded Spill
+; GFX9-NEXT: buffer_store_dword v41, off, s[0:3], s33 offset:4 ; 4-byte Folded Spill
; GFX9-NEXT: s_mov_b64 exec, s[36:37]
-; GFX9-NEXT: v_writelane_b32 v40, s34, 2
; GFX9-NEXT: s_addk_i32 s32, 0x400
; GFX9-NEXT: v_writelane_b32 v40, s30, 0
; GFX9-NEXT: v_mov_b32_e32 v0, 0x20001
; GFX9-NEXT: v_mov_b32_e32 v1, 3
+; GFX9-NEXT: v_writelane_b32 v41, s34, 0
; GFX9-NEXT: v_writelane_b32 v40, s31, 1
; GFX9-NEXT: s_getpc_b64 s[34:35]
; GFX9-NEXT: s_add_u32 s34, s34, external_void_func_v3i16@rel32@lo+4
@@ -3427,9 +3739,10 @@
; GFX9-NEXT: s_swappc_b64 s[30:31], s[34:35]
; GFX9-NEXT: v_readlane_b32 s31, v40, 1
; GFX9-NEXT: v_readlane_b32 s30, v40, 0
-; GFX9-NEXT: v_readlane_b32 s34, v40, 2
+; GFX9-NEXT: v_readlane_b32 s34, v41, 0
; GFX9-NEXT: s_or_saveexec_b64 s[36:37], -1
; GFX9-NEXT: buffer_load_dword v40, off, s[0:3], s33 ; 4-byte Folded Reload
+; GFX9-NEXT: buffer_load_dword v41, off, s[0:3], s33 offset:4 ; 4-byte Folded Reload
; GFX9-NEXT: s_mov_b64 exec, s[36:37]
; GFX9-NEXT: s_addk_i32 s32, 0xfc00
; GFX9-NEXT: s_mov_b32 s33, s34
@@ -3443,23 +3756,26 @@
; GFX10-NEXT: s_mov_b32 s33, s32
; GFX10-NEXT: s_or_saveexec_b32 s35, -1
; GFX10-NEXT: buffer_store_dword v40, off, s[0:3], s33 ; 4-byte Folded Spill
+; GFX10-NEXT: buffer_store_dword v41, off, s[0:3], s33 offset:4 ; 4-byte Folded Spill
; GFX10-NEXT: s_waitcnt_depctr 0xffe3
; GFX10-NEXT: s_mov_b32 exec_lo, s35
-; GFX10-NEXT: v_writelane_b32 v40, s34, 2
+; GFX10-NEXT: v_writelane_b32 v40, s30, 0
; GFX10-NEXT: v_mov_b32_e32 v0, 0x20001
; GFX10-NEXT: v_mov_b32_e32 v1, 3
; GFX10-NEXT: s_addk_i32 s32, 0x200
+; GFX10-NEXT: v_writelane_b32 v41, s34, 0
+; GFX10-NEXT: v_writelane_b32 v40, s31, 1
; GFX10-NEXT: s_getpc_b64 s[34:35]
; GFX10-NEXT: s_add_u32 s34, s34, external_void_func_v3i16@rel32@lo+4
; GFX10-NEXT: s_addc_u32 s35, s35, external_void_func_v3i16@rel32@hi+12
-; GFX10-NEXT: v_writelane_b32 v40, s30, 0
-; GFX10-NEXT: v_writelane_b32 v40, s31, 1
; GFX10-NEXT: s_swappc_b64 s[30:31], s[34:35]
; GFX10-NEXT: v_readlane_b32 s31, v40, 1
; GFX10-NEXT: v_readlane_b32 s30, v40, 0
-; GFX10-NEXT: v_readlane_b32 s34, v40, 2
+; GFX10-NEXT: v_readlane_b32 s34, v41, 0
; GFX10-NEXT: s_or_saveexec_b32 s35, -1
-; GFX10-NEXT: buffer_load_dword v40, off, s[0:3], s33 ; 4-byte Folded Reload
+; GFX10-NEXT: s_clause 0x1
+; GFX10-NEXT: buffer_load_dword v40, off, s[0:3], s33
+; GFX10-NEXT: buffer_load_dword v41, off, s[0:3], s33 offset:4
; GFX10-NEXT: s_waitcnt_depctr 0xffe3
; GFX10-NEXT: s_mov_b32 exec_lo, s35
; GFX10-NEXT: s_addk_i32 s32, 0xfe00
@@ -3473,23 +3789,27 @@
; GFX11-NEXT: s_mov_b32 s0, s33
; GFX11-NEXT: s_mov_b32 s33, s32
; GFX11-NEXT: s_or_saveexec_b32 s1, -1
-; GFX11-NEXT: scratch_store_b32 off, v40, s33 ; 4-byte Folded Spill
+; GFX11-NEXT: s_clause 0x1
+; GFX11-NEXT: scratch_store_b32 off, v40, s33
+; GFX11-NEXT: scratch_store_b32 off, v41, s33 offset:4
; GFX11-NEXT: s_mov_b32 exec_lo, s1
-; GFX11-NEXT: v_writelane_b32 v40, s0, 2
+; GFX11-NEXT: v_writelane_b32 v40, s30, 0
; GFX11-NEXT: v_dual_mov_b32 v0, 0x20001 :: v_dual_mov_b32 v1, 3
; GFX11-NEXT: s_add_i32 s32, s32, 16
+; GFX11-NEXT: v_writelane_b32 v41, s0, 0
+; GFX11-NEXT: v_writelane_b32 v40, s31, 1
; GFX11-NEXT: s_getpc_b64 s[0:1]
; GFX11-NEXT: s_add_u32 s0, s0, external_void_func_v3i16@rel32@lo+4
; GFX11-NEXT: s_addc_u32 s1, s1, external_void_func_v3i16@rel32@hi+12
-; GFX11-NEXT: v_writelane_b32 v40, s30, 0
-; GFX11-NEXT: v_writelane_b32 v40, s31, 1
+; GFX11-NEXT: s_delay_alu instid0(SALU_CYCLE_1)
; GFX11-NEXT: s_swappc_b64 s[30:31], s[0:1]
-; GFX11-NEXT: s_delay_alu instid0(VALU_DEP_1)
; GFX11-NEXT: v_readlane_b32 s31, v40, 1
; GFX11-NEXT: v_readlane_b32 s30, v40, 0
-; GFX11-NEXT: v_readlane_b32 s0, v40, 2
+; GFX11-NEXT: v_readlane_b32 s0, v41, 0
; GFX11-NEXT: s_or_saveexec_b32 s1, -1
-; GFX11-NEXT: scratch_load_b32 v40, off, s33 ; 4-byte Folded Reload
+; GFX11-NEXT: s_clause 0x1
+; GFX11-NEXT: scratch_load_b32 v40, off, s33
+; GFX11-NEXT: scratch_load_b32 v41, off, s33 offset:4
; GFX11-NEXT: s_mov_b32 exec_lo, s1
; GFX11-NEXT: s_add_i32 s32, s32, -16
; GFX11-NEXT: s_mov_b32 s33, s0
@@ -3503,23 +3823,26 @@
; GFX10-SCRATCH-NEXT: s_mov_b32 s33, s32
; GFX10-SCRATCH-NEXT: s_or_saveexec_b32 s1, -1
; GFX10-SCRATCH-NEXT: scratch_store_dword off, v40, s33 ; 4-byte Folded Spill
+; GFX10-SCRATCH-NEXT: scratch_store_dword off, v41, s33 offset:4 ; 4-byte Folded Spill
; GFX10-SCRATCH-NEXT: s_waitcnt_depctr 0xffe3
; GFX10-SCRATCH-NEXT: s_mov_b32 exec_lo, s1
-; GFX10-SCRATCH-NEXT: v_writelane_b32 v40, s0, 2
+; GFX10-SCRATCH-NEXT: v_writelane_b32 v40, s30, 0
; GFX10-SCRATCH-NEXT: v_mov_b32_e32 v0, 0x20001
; GFX10-SCRATCH-NEXT: v_mov_b32_e32 v1, 3
; GFX10-SCRATCH-NEXT: s_add_i32 s32, s32, 16
+; GFX10-SCRATCH-NEXT: v_writelane_b32 v41, s0, 0
+; GFX10-SCRATCH-NEXT: v_writelane_b32 v40, s31, 1
; GFX10-SCRATCH-NEXT: s_getpc_b64 s[0:1]
; GFX10-SCRATCH-NEXT: s_add_u32 s0, s0, external_void_func_v3i16@rel32@lo+4
; GFX10-SCRATCH-NEXT: s_addc_u32 s1, s1, external_void_func_v3i16@rel32@hi+12
-; GFX10-SCRATCH-NEXT: v_writelane_b32 v40, s30, 0
-; GFX10-SCRATCH-NEXT: v_writelane_b32 v40, s31, 1
; GFX10-SCRATCH-NEXT: s_swappc_b64 s[30:31], s[0:1]
; GFX10-SCRATCH-NEXT: v_readlane_b32 s31, v40, 1
; GFX10-SCRATCH-NEXT: v_readlane_b32 s30, v40, 0
-; GFX10-SCRATCH-NEXT: v_readlane_b32 s0, v40, 2
+; GFX10-SCRATCH-NEXT: v_readlane_b32 s0, v41, 0
; GFX10-SCRATCH-NEXT: s_or_saveexec_b32 s1, -1
-; GFX10-SCRATCH-NEXT: scratch_load_dword v40, off, s33 ; 4-byte Folded Reload
+; GFX10-SCRATCH-NEXT: s_clause 0x1
+; GFX10-SCRATCH-NEXT: scratch_load_dword v40, off, s33
+; GFX10-SCRATCH-NEXT: scratch_load_dword v41, off, s33 offset:4
; GFX10-SCRATCH-NEXT: s_waitcnt_depctr 0xffe3
; GFX10-SCRATCH-NEXT: s_mov_b32 exec_lo, s1
; GFX10-SCRATCH-NEXT: s_add_i32 s32, s32, -16
@@ -3538,12 +3861,13 @@
; GFX9-NEXT: s_mov_b32 s33, s32
; GFX9-NEXT: s_or_saveexec_b64 s[36:37], -1
; GFX9-NEXT: buffer_store_dword v40, off, s[0:3], s33 ; 4-byte Folded Spill
+; GFX9-NEXT: buffer_store_dword v41, off, s[0:3], s33 offset:4 ; 4-byte Folded Spill
; GFX9-NEXT: s_mov_b64 exec, s[36:37]
-; GFX9-NEXT: v_writelane_b32 v40, s34, 2
; GFX9-NEXT: s_addk_i32 s32, 0x400
; GFX9-NEXT: v_writelane_b32 v40, s30, 0
; GFX9-NEXT: v_mov_b32_e32 v0, 0x40003c00
; GFX9-NEXT: v_mov_b32_e32 v1, 0x4400
+; GFX9-NEXT: v_writelane_b32 v41, s34, 0
; GFX9-NEXT: v_writelane_b32 v40, s31, 1
; GFX9-NEXT: s_getpc_b64 s[34:35]
; GFX9-NEXT: s_add_u32 s34, s34, external_void_func_v3f16@rel32@lo+4
@@ -3551,9 +3875,10 @@
; GFX9-NEXT: s_swappc_b64 s[30:31], s[34:35]
; GFX9-NEXT: v_readlane_b32 s31, v40, 1
; GFX9-NEXT: v_readlane_b32 s30, v40, 0
-; GFX9-NEXT: v_readlane_b32 s34, v40, 2
+; GFX9-NEXT: v_readlane_b32 s34, v41, 0
; GFX9-NEXT: s_or_saveexec_b64 s[36:37], -1
; GFX9-NEXT: buffer_load_dword v40, off, s[0:3], s33 ; 4-byte Folded Reload
+; GFX9-NEXT: buffer_load_dword v41, off, s[0:3], s33 offset:4 ; 4-byte Folded Reload
; GFX9-NEXT: s_mov_b64 exec, s[36:37]
; GFX9-NEXT: s_addk_i32 s32, 0xfc00
; GFX9-NEXT: s_mov_b32 s33, s34
@@ -3567,23 +3892,26 @@
; GFX10-NEXT: s_mov_b32 s33, s32
; GFX10-NEXT: s_or_saveexec_b32 s35, -1
; GFX10-NEXT: buffer_store_dword v40, off, s[0:3], s33 ; 4-byte Folded Spill
+; GFX10-NEXT: buffer_store_dword v41, off, s[0:3], s33 offset:4 ; 4-byte Folded Spill
; GFX10-NEXT: s_waitcnt_depctr 0xffe3
; GFX10-NEXT: s_mov_b32 exec_lo, s35
-; GFX10-NEXT: v_writelane_b32 v40, s34, 2
+; GFX10-NEXT: v_writelane_b32 v40, s30, 0
; GFX10-NEXT: v_mov_b32_e32 v0, 0x40003c00
; GFX10-NEXT: v_mov_b32_e32 v1, 0x4400
; GFX10-NEXT: s_addk_i32 s32, 0x200
+; GFX10-NEXT: v_writelane_b32 v41, s34, 0
+; GFX10-NEXT: v_writelane_b32 v40, s31, 1
; GFX10-NEXT: s_getpc_b64 s[34:35]
; GFX10-NEXT: s_add_u32 s34, s34, external_void_func_v3f16@rel32@lo+4
; GFX10-NEXT: s_addc_u32 s35, s35, external_void_func_v3f16@rel32@hi+12
-; GFX10-NEXT: v_writelane_b32 v40, s30, 0
-; GFX10-NEXT: v_writelane_b32 v40, s31, 1
; GFX10-NEXT: s_swappc_b64 s[30:31], s[34:35]
; GFX10-NEXT: v_readlane_b32 s31, v40, 1
; GFX10-NEXT: v_readlane_b32 s30, v40, 0
-; GFX10-NEXT: v_readlane_b32 s34, v40, 2
+; GFX10-NEXT: v_readlane_b32 s34, v41, 0
; GFX10-NEXT: s_or_saveexec_b32 s35, -1
-; GFX10-NEXT: buffer_load_dword v40, off, s[0:3], s33 ; 4-byte Folded Reload
+; GFX10-NEXT: s_clause 0x1
+; GFX10-NEXT: buffer_load_dword v40, off, s[0:3], s33
+; GFX10-NEXT: buffer_load_dword v41, off, s[0:3], s33 offset:4
; GFX10-NEXT: s_waitcnt_depctr 0xffe3
; GFX10-NEXT: s_mov_b32 exec_lo, s35
; GFX10-NEXT: s_addk_i32 s32, 0xfe00
@@ -3597,24 +3925,28 @@
; GFX11-NEXT: s_mov_b32 s0, s33
; GFX11-NEXT: s_mov_b32 s33, s32
; GFX11-NEXT: s_or_saveexec_b32 s1, -1
-; GFX11-NEXT: scratch_store_b32 off, v40, s33 ; 4-byte Folded Spill
+; GFX11-NEXT: s_clause 0x1
+; GFX11-NEXT: scratch_store_b32 off, v40, s33
+; GFX11-NEXT: scratch_store_b32 off, v41, s33 offset:4
; GFX11-NEXT: s_mov_b32 exec_lo, s1
-; GFX11-NEXT: v_writelane_b32 v40, s0, 2
+; GFX11-NEXT: v_writelane_b32 v40, s30, 0
; GFX11-NEXT: v_mov_b32_e32 v0, 0x40003c00
; GFX11-NEXT: v_mov_b32_e32 v1, 0x4400
; GFX11-NEXT: s_add_i32 s32, s32, 16
+; GFX11-NEXT: v_writelane_b32 v41, s0, 0
+; GFX11-NEXT: v_writelane_b32 v40, s31, 1
; GFX11-NEXT: s_getpc_b64 s[0:1]
; GFX11-NEXT: s_add_u32 s0, s0, external_void_func_v3f16@rel32@lo+4
; GFX11-NEXT: s_addc_u32 s1, s1, external_void_func_v3f16@rel32@hi+12
-; GFX11-NEXT: v_writelane_b32 v40, s30, 0
-; GFX11-NEXT: v_writelane_b32 v40, s31, 1
+; GFX11-NEXT: s_delay_alu instid0(SALU_CYCLE_1)
; GFX11-NEXT: s_swappc_b64 s[30:31], s[0:1]
-; GFX11-NEXT: s_delay_alu instid0(VALU_DEP_1)
; GFX11-NEXT: v_readlane_b32 s31, v40, 1
; GFX11-NEXT: v_readlane_b32 s30, v40, 0
-; GFX11-NEXT: v_readlane_b32 s0, v40, 2
+; GFX11-NEXT: v_readlane_b32 s0, v41, 0
; GFX11-NEXT: s_or_saveexec_b32 s1, -1
-; GFX11-NEXT: scratch_load_b32 v40, off, s33 ; 4-byte Folded Reload
+; GFX11-NEXT: s_clause 0x1
+; GFX11-NEXT: scratch_load_b32 v40, off, s33
+; GFX11-NEXT: scratch_load_b32 v41, off, s33 offset:4
; GFX11-NEXT: s_mov_b32 exec_lo, s1
; GFX11-NEXT: s_add_i32 s32, s32, -16
; GFX11-NEXT: s_mov_b32 s33, s0
@@ -3628,23 +3960,26 @@
; GFX10-SCRATCH-NEXT: s_mov_b32 s33, s32
; GFX10-SCRATCH-NEXT: s_or_saveexec_b32 s1, -1
; GFX10-SCRATCH-NEXT: scratch_store_dword off, v40, s33 ; 4-byte Folded Spill
+; GFX10-SCRATCH-NEXT: scratch_store_dword off, v41, s33 offset:4 ; 4-byte Folded Spill
; GFX10-SCRATCH-NEXT: s_waitcnt_depctr 0xffe3
; GFX10-SCRATCH-NEXT: s_mov_b32 exec_lo, s1
-; GFX10-SCRATCH-NEXT: v_writelane_b32 v40, s0, 2
+; GFX10-SCRATCH-NEXT: v_writelane_b32 v40, s30, 0
; GFX10-SCRATCH-NEXT: v_mov_b32_e32 v0, 0x40003c00
; GFX10-SCRATCH-NEXT: v_mov_b32_e32 v1, 0x4400
; GFX10-SCRATCH-NEXT: s_add_i32 s32, s32, 16
+; GFX10-SCRATCH-NEXT: v_writelane_b32 v41, s0, 0
+; GFX10-SCRATCH-NEXT: v_writelane_b32 v40, s31, 1
; GFX10-SCRATCH-NEXT: s_getpc_b64 s[0:1]
; GFX10-SCRATCH-NEXT: s_add_u32 s0, s0, external_void_func_v3f16@rel32@lo+4
; GFX10-SCRATCH-NEXT: s_addc_u32 s1, s1, external_void_func_v3f16@rel32@hi+12
-; GFX10-SCRATCH-NEXT: v_writelane_b32 v40, s30, 0
-; GFX10-SCRATCH-NEXT: v_writelane_b32 v40, s31, 1
; GFX10-SCRATCH-NEXT: s_swappc_b64 s[30:31], s[0:1]
; GFX10-SCRATCH-NEXT: v_readlane_b32 s31, v40, 1
; GFX10-SCRATCH-NEXT: v_readlane_b32 s30, v40, 0
-; GFX10-SCRATCH-NEXT: v_readlane_b32 s0, v40, 2
+; GFX10-SCRATCH-NEXT: v_readlane_b32 s0, v41, 0
; GFX10-SCRATCH-NEXT: s_or_saveexec_b32 s1, -1
-; GFX10-SCRATCH-NEXT: scratch_load_dword v40, off, s33 ; 4-byte Folded Reload
+; GFX10-SCRATCH-NEXT: s_clause 0x1
+; GFX10-SCRATCH-NEXT: scratch_load_dword v40, off, s33
+; GFX10-SCRATCH-NEXT: scratch_load_dword v41, off, s33 offset:4
; GFX10-SCRATCH-NEXT: s_waitcnt_depctr 0xffe3
; GFX10-SCRATCH-NEXT: s_mov_b32 exec_lo, s1
; GFX10-SCRATCH-NEXT: s_add_i32 s32, s32, -16
@@ -3663,11 +3998,12 @@
; GFX9-NEXT: s_mov_b32 s33, s32
; GFX9-NEXT: s_or_saveexec_b64 s[36:37], -1
; GFX9-NEXT: buffer_store_dword v40, off, s[0:3], s33 ; 4-byte Folded Spill
+; GFX9-NEXT: buffer_store_dword v41, off, s[0:3], s33 offset:4 ; 4-byte Folded Spill
; GFX9-NEXT: s_mov_b64 exec, s[36:37]
; GFX9-NEXT: global_load_dwordx2 v[0:1], v[0:1], off
-; GFX9-NEXT: v_writelane_b32 v40, s34, 2
; GFX9-NEXT: s_addk_i32 s32, 0x400
; GFX9-NEXT: v_writelane_b32 v40, s30, 0
+; GFX9-NEXT: v_writelane_b32 v41, s34, 0
; GFX9-NEXT: v_writelane_b32 v40, s31, 1
; GFX9-NEXT: s_getpc_b64 s[34:35]
; GFX9-NEXT: s_add_u32 s34, s34, external_void_func_v4i16@rel32@lo+4
@@ -3675,9 +4011,10 @@
; GFX9-NEXT: s_swappc_b64 s[30:31], s[34:35]
; GFX9-NEXT: v_readlane_b32 s31, v40, 1
; GFX9-NEXT: v_readlane_b32 s30, v40, 0
-; GFX9-NEXT: v_readlane_b32 s34, v40, 2
+; GFX9-NEXT: v_readlane_b32 s34, v41, 0
; GFX9-NEXT: s_or_saveexec_b64 s[36:37], -1
; GFX9-NEXT: buffer_load_dword v40, off, s[0:3], s33 ; 4-byte Folded Reload
+; GFX9-NEXT: buffer_load_dword v41, off, s[0:3], s33 offset:4 ; 4-byte Folded Reload
; GFX9-NEXT: s_mov_b64 exec, s[36:37]
; GFX9-NEXT: s_addk_i32 s32, 0xfc00
; GFX9-NEXT: s_mov_b32 s33, s34
@@ -3691,22 +4028,25 @@
; GFX10-NEXT: s_mov_b32 s33, s32
; GFX10-NEXT: s_or_saveexec_b32 s35, -1
; GFX10-NEXT: buffer_store_dword v40, off, s[0:3], s33 ; 4-byte Folded Spill
+; GFX10-NEXT: buffer_store_dword v41, off, s[0:3], s33 offset:4 ; 4-byte Folded Spill
; GFX10-NEXT: s_waitcnt_depctr 0xffe3
; GFX10-NEXT: s_mov_b32 exec_lo, s35
; GFX10-NEXT: global_load_dwordx2 v[0:1], v[0:1], off
-; GFX10-NEXT: v_writelane_b32 v40, s34, 2
+; GFX10-NEXT: v_writelane_b32 v40, s30, 0
; GFX10-NEXT: s_addk_i32 s32, 0x200
+; GFX10-NEXT: v_writelane_b32 v41, s34, 0
; GFX10-NEXT: s_getpc_b64 s[34:35]
; GFX10-NEXT: s_add_u32 s34, s34, external_void_func_v4i16@rel32@lo+4
; GFX10-NEXT: s_addc_u32 s35, s35, external_void_func_v4i16@rel32@hi+12
-; GFX10-NEXT: v_writelane_b32 v40, s30, 0
; GFX10-NEXT: v_writelane_b32 v40, s31, 1
; GFX10-NEXT: s_swappc_b64 s[30:31], s[34:35]
; GFX10-NEXT: v_readlane_b32 s31, v40, 1
; GFX10-NEXT: v_readlane_b32 s30, v40, 0
-; GFX10-NEXT: v_readlane_b32 s34, v40, 2
+; GFX10-NEXT: v_readlane_b32 s34, v41, 0
; GFX10-NEXT: s_or_saveexec_b32 s35, -1
-; GFX10-NEXT: buffer_load_dword v40, off, s[0:3], s33 ; 4-byte Folded Reload
+; GFX10-NEXT: s_clause 0x1
+; GFX10-NEXT: buffer_load_dword v40, off, s[0:3], s33
+; GFX10-NEXT: buffer_load_dword v41, off, s[0:3], s33 offset:4
; GFX10-NEXT: s_waitcnt_depctr 0xffe3
; GFX10-NEXT: s_mov_b32 exec_lo, s35
; GFX10-NEXT: s_addk_i32 s32, 0xfe00
@@ -3720,23 +4060,27 @@
; GFX11-NEXT: s_mov_b32 s0, s33
; GFX11-NEXT: s_mov_b32 s33, s32
; GFX11-NEXT: s_or_saveexec_b32 s1, -1
-; GFX11-NEXT: scratch_store_b32 off, v40, s33 ; 4-byte Folded Spill
+; GFX11-NEXT: s_clause 0x1
+; GFX11-NEXT: scratch_store_b32 off, v40, s33
+; GFX11-NEXT: scratch_store_b32 off, v41, s33 offset:4
; GFX11-NEXT: s_mov_b32 exec_lo, s1
; GFX11-NEXT: global_load_b64 v[0:1], v[0:1], off
-; GFX11-NEXT: v_writelane_b32 v40, s0, 2
+; GFX11-NEXT: v_writelane_b32 v40, s30, 0
; GFX11-NEXT: s_add_i32 s32, s32, 16
+; GFX11-NEXT: v_writelane_b32 v41, s0, 0
; GFX11-NEXT: s_getpc_b64 s[0:1]
; GFX11-NEXT: s_add_u32 s0, s0, external_void_func_v4i16@rel32@lo+4
; GFX11-NEXT: s_addc_u32 s1, s1, external_void_func_v4i16@rel32@hi+12
-; GFX11-NEXT: v_writelane_b32 v40, s30, 0
; GFX11-NEXT: v_writelane_b32 v40, s31, 1
; GFX11-NEXT: s_swappc_b64 s[30:31], s[0:1]
; GFX11-NEXT: s_delay_alu instid0(VALU_DEP_1)
; GFX11-NEXT: v_readlane_b32 s31, v40, 1
; GFX11-NEXT: v_readlane_b32 s30, v40, 0
-; GFX11-NEXT: v_readlane_b32 s0, v40, 2
+; GFX11-NEXT: v_readlane_b32 s0, v41, 0
; GFX11-NEXT: s_or_saveexec_b32 s1, -1
-; GFX11-NEXT: scratch_load_b32 v40, off, s33 ; 4-byte Folded Reload
+; GFX11-NEXT: s_clause 0x1
+; GFX11-NEXT: scratch_load_b32 v40, off, s33
+; GFX11-NEXT: scratch_load_b32 v41, off, s33 offset:4
; GFX11-NEXT: s_mov_b32 exec_lo, s1
; GFX11-NEXT: s_add_i32 s32, s32, -16
; GFX11-NEXT: s_mov_b32 s33, s0
@@ -3750,22 +4094,25 @@
; GFX10-SCRATCH-NEXT: s_mov_b32 s33, s32
; GFX10-SCRATCH-NEXT: s_or_saveexec_b32 s1, -1
; GFX10-SCRATCH-NEXT: scratch_store_dword off, v40, s33 ; 4-byte Folded Spill
+; GFX10-SCRATCH-NEXT: scratch_store_dword off, v41, s33 offset:4 ; 4-byte Folded Spill
; GFX10-SCRATCH-NEXT: s_waitcnt_depctr 0xffe3
; GFX10-SCRATCH-NEXT: s_mov_b32 exec_lo, s1
; GFX10-SCRATCH-NEXT: global_load_dwordx2 v[0:1], v[0:1], off
-; GFX10-SCRATCH-NEXT: v_writelane_b32 v40, s0, 2
+; GFX10-SCRATCH-NEXT: v_writelane_b32 v40, s30, 0
; GFX10-SCRATCH-NEXT: s_add_i32 s32, s32, 16
+; GFX10-SCRATCH-NEXT: v_writelane_b32 v41, s0, 0
; GFX10-SCRATCH-NEXT: s_getpc_b64 s[0:1]
; GFX10-SCRATCH-NEXT: s_add_u32 s0, s0, external_void_func_v4i16@rel32@lo+4
; GFX10-SCRATCH-NEXT: s_addc_u32 s1, s1, external_void_func_v4i16@rel32@hi+12
-; GFX10-SCRATCH-NEXT: v_writelane_b32 v40, s30, 0
; GFX10-SCRATCH-NEXT: v_writelane_b32 v40, s31, 1
; GFX10-SCRATCH-NEXT: s_swappc_b64 s[30:31], s[0:1]
; GFX10-SCRATCH-NEXT: v_readlane_b32 s31, v40, 1
; GFX10-SCRATCH-NEXT: v_readlane_b32 s30, v40, 0
-; GFX10-SCRATCH-NEXT: v_readlane_b32 s0, v40, 2
+; GFX10-SCRATCH-NEXT: v_readlane_b32 s0, v41, 0
; GFX10-SCRATCH-NEXT: s_or_saveexec_b32 s1, -1
-; GFX10-SCRATCH-NEXT: scratch_load_dword v40, off, s33 ; 4-byte Folded Reload
+; GFX10-SCRATCH-NEXT: s_clause 0x1
+; GFX10-SCRATCH-NEXT: scratch_load_dword v40, off, s33
+; GFX10-SCRATCH-NEXT: scratch_load_dword v41, off, s33 offset:4
; GFX10-SCRATCH-NEXT: s_waitcnt_depctr 0xffe3
; GFX10-SCRATCH-NEXT: s_mov_b32 exec_lo, s1
; GFX10-SCRATCH-NEXT: s_add_i32 s32, s32, -16
@@ -3785,12 +4132,13 @@
; GFX9-NEXT: s_mov_b32 s33, s32
; GFX9-NEXT: s_or_saveexec_b64 s[36:37], -1
; GFX9-NEXT: buffer_store_dword v40, off, s[0:3], s33 ; 4-byte Folded Spill
+; GFX9-NEXT: buffer_store_dword v41, off, s[0:3], s33 offset:4 ; 4-byte Folded Spill
; GFX9-NEXT: s_mov_b64 exec, s[36:37]
-; GFX9-NEXT: v_writelane_b32 v40, s34, 2
; GFX9-NEXT: s_addk_i32 s32, 0x400
; GFX9-NEXT: v_writelane_b32 v40, s30, 0
; GFX9-NEXT: v_mov_b32_e32 v0, 0x20001
; GFX9-NEXT: v_mov_b32_e32 v1, 0x40003
+; GFX9-NEXT: v_writelane_b32 v41, s34, 0
; GFX9-NEXT: v_writelane_b32 v40, s31, 1
; GFX9-NEXT: s_getpc_b64 s[34:35]
; GFX9-NEXT: s_add_u32 s34, s34, external_void_func_v4i16@rel32@lo+4
@@ -3798,9 +4146,10 @@
; GFX9-NEXT: s_swappc_b64 s[30:31], s[34:35]
; GFX9-NEXT: v_readlane_b32 s31, v40, 1
; GFX9-NEXT: v_readlane_b32 s30, v40, 0
-; GFX9-NEXT: v_readlane_b32 s34, v40, 2
+; GFX9-NEXT: v_readlane_b32 s34, v41, 0
; GFX9-NEXT: s_or_saveexec_b64 s[36:37], -1
; GFX9-NEXT: buffer_load_dword v40, off, s[0:3], s33 ; 4-byte Folded Reload
+; GFX9-NEXT: buffer_load_dword v41, off, s[0:3], s33 offset:4 ; 4-byte Folded Reload
; GFX9-NEXT: s_mov_b64 exec, s[36:37]
; GFX9-NEXT: s_addk_i32 s32, 0xfc00
; GFX9-NEXT: s_mov_b32 s33, s34
@@ -3814,23 +4163,26 @@
; GFX10-NEXT: s_mov_b32 s33, s32
; GFX10-NEXT: s_or_saveexec_b32 s35, -1
; GFX10-NEXT: buffer_store_dword v40, off, s[0:3], s33 ; 4-byte Folded Spill
+; GFX10-NEXT: buffer_store_dword v41, off, s[0:3], s33 offset:4 ; 4-byte Folded Spill
; GFX10-NEXT: s_waitcnt_depctr 0xffe3
; GFX10-NEXT: s_mov_b32 exec_lo, s35
-; GFX10-NEXT: v_writelane_b32 v40, s34, 2
+; GFX10-NEXT: v_writelane_b32 v40, s30, 0
; GFX10-NEXT: v_mov_b32_e32 v0, 0x20001
; GFX10-NEXT: v_mov_b32_e32 v1, 0x40003
; GFX10-NEXT: s_addk_i32 s32, 0x200
+; GFX10-NEXT: v_writelane_b32 v41, s34, 0
+; GFX10-NEXT: v_writelane_b32 v40, s31, 1
; GFX10-NEXT: s_getpc_b64 s[34:35]
; GFX10-NEXT: s_add_u32 s34, s34, external_void_func_v4i16@rel32@lo+4
; GFX10-NEXT: s_addc_u32 s35, s35, external_void_func_v4i16@rel32@hi+12
-; GFX10-NEXT: v_writelane_b32 v40, s30, 0
-; GFX10-NEXT: v_writelane_b32 v40, s31, 1
; GFX10-NEXT: s_swappc_b64 s[30:31], s[34:35]
; GFX10-NEXT: v_readlane_b32 s31, v40, 1
; GFX10-NEXT: v_readlane_b32 s30, v40, 0
-; GFX10-NEXT: v_readlane_b32 s34, v40, 2
+; GFX10-NEXT: v_readlane_b32 s34, v41, 0
; GFX10-NEXT: s_or_saveexec_b32 s35, -1
-; GFX10-NEXT: buffer_load_dword v40, off, s[0:3], s33 ; 4-byte Folded Reload
+; GFX10-NEXT: s_clause 0x1
+; GFX10-NEXT: buffer_load_dword v40, off, s[0:3], s33
+; GFX10-NEXT: buffer_load_dword v41, off, s[0:3], s33 offset:4
; GFX10-NEXT: s_waitcnt_depctr 0xffe3
; GFX10-NEXT: s_mov_b32 exec_lo, s35
; GFX10-NEXT: s_addk_i32 s32, 0xfe00
@@ -3844,24 +4196,28 @@
; GFX11-NEXT: s_mov_b32 s0, s33
; GFX11-NEXT: s_mov_b32 s33, s32
; GFX11-NEXT: s_or_saveexec_b32 s1, -1
-; GFX11-NEXT: scratch_store_b32 off, v40, s33 ; 4-byte Folded Spill
+; GFX11-NEXT: s_clause 0x1
+; GFX11-NEXT: scratch_store_b32 off, v40, s33
+; GFX11-NEXT: scratch_store_b32 off, v41, s33 offset:4
; GFX11-NEXT: s_mov_b32 exec_lo, s1
-; GFX11-NEXT: v_writelane_b32 v40, s0, 2
+; GFX11-NEXT: v_writelane_b32 v40, s30, 0
; GFX11-NEXT: v_mov_b32_e32 v0, 0x20001
; GFX11-NEXT: v_mov_b32_e32 v1, 0x40003
; GFX11-NEXT: s_add_i32 s32, s32, 16
+; GFX11-NEXT: v_writelane_b32 v41, s0, 0
+; GFX11-NEXT: v_writelane_b32 v40, s31, 1
; GFX11-NEXT: s_getpc_b64 s[0:1]
; GFX11-NEXT: s_add_u32 s0, s0, external_void_func_v4i16@rel32@lo+4
; GFX11-NEXT: s_addc_u32 s1, s1, external_void_func_v4i16@rel32@hi+12
-; GFX11-NEXT: v_writelane_b32 v40, s30, 0
-; GFX11-NEXT: v_writelane_b32 v40, s31, 1
+; GFX11-NEXT: s_delay_alu instid0(SALU_CYCLE_1)
; GFX11-NEXT: s_swappc_b64 s[30:31], s[0:1]
-; GFX11-NEXT: s_delay_alu instid0(VALU_DEP_1)
; GFX11-NEXT: v_readlane_b32 s31, v40, 1
; GFX11-NEXT: v_readlane_b32 s30, v40, 0
-; GFX11-NEXT: v_readlane_b32 s0, v40, 2
+; GFX11-NEXT: v_readlane_b32 s0, v41, 0
; GFX11-NEXT: s_or_saveexec_b32 s1, -1
-; GFX11-NEXT: scratch_load_b32 v40, off, s33 ; 4-byte Folded Reload
+; GFX11-NEXT: s_clause 0x1
+; GFX11-NEXT: scratch_load_b32 v40, off, s33
+; GFX11-NEXT: scratch_load_b32 v41, off, s33 offset:4
; GFX11-NEXT: s_mov_b32 exec_lo, s1
; GFX11-NEXT: s_add_i32 s32, s32, -16
; GFX11-NEXT: s_mov_b32 s33, s0
@@ -3875,23 +4231,26 @@
; GFX10-SCRATCH-NEXT: s_mov_b32 s33, s32
; GFX10-SCRATCH-NEXT: s_or_saveexec_b32 s1, -1
; GFX10-SCRATCH-NEXT: scratch_store_dword off, v40, s33 ; 4-byte Folded Spill
+; GFX10-SCRATCH-NEXT: scratch_store_dword off, v41, s33 offset:4 ; 4-byte Folded Spill
; GFX10-SCRATCH-NEXT: s_waitcnt_depctr 0xffe3
; GFX10-SCRATCH-NEXT: s_mov_b32 exec_lo, s1
-; GFX10-SCRATCH-NEXT: v_writelane_b32 v40, s0, 2
+; GFX10-SCRATCH-NEXT: v_writelane_b32 v40, s30, 0
; GFX10-SCRATCH-NEXT: v_mov_b32_e32 v0, 0x20001
; GFX10-SCRATCH-NEXT: v_mov_b32_e32 v1, 0x40003
; GFX10-SCRATCH-NEXT: s_add_i32 s32, s32, 16
+; GFX10-SCRATCH-NEXT: v_writelane_b32 v41, s0, 0
+; GFX10-SCRATCH-NEXT: v_writelane_b32 v40, s31, 1
; GFX10-SCRATCH-NEXT: s_getpc_b64 s[0:1]
; GFX10-SCRATCH-NEXT: s_add_u32 s0, s0, external_void_func_v4i16@rel32@lo+4
; GFX10-SCRATCH-NEXT: s_addc_u32 s1, s1, external_void_func_v4i16@rel32@hi+12
-; GFX10-SCRATCH-NEXT: v_writelane_b32 v40, s30, 0
-; GFX10-SCRATCH-NEXT: v_writelane_b32 v40, s31, 1
; GFX10-SCRATCH-NEXT: s_swappc_b64 s[30:31], s[0:1]
; GFX10-SCRATCH-NEXT: v_readlane_b32 s31, v40, 1
; GFX10-SCRATCH-NEXT: v_readlane_b32 s30, v40, 0
-; GFX10-SCRATCH-NEXT: v_readlane_b32 s0, v40, 2
+; GFX10-SCRATCH-NEXT: v_readlane_b32 s0, v41, 0
; GFX10-SCRATCH-NEXT: s_or_saveexec_b32 s1, -1
-; GFX10-SCRATCH-NEXT: scratch_load_dword v40, off, s33 ; 4-byte Folded Reload
+; GFX10-SCRATCH-NEXT: s_clause 0x1
+; GFX10-SCRATCH-NEXT: scratch_load_dword v40, off, s33
+; GFX10-SCRATCH-NEXT: scratch_load_dword v41, off, s33 offset:4
; GFX10-SCRATCH-NEXT: s_waitcnt_depctr 0xffe3
; GFX10-SCRATCH-NEXT: s_mov_b32 exec_lo, s1
; GFX10-SCRATCH-NEXT: s_add_i32 s32, s32, -16
@@ -3910,11 +4269,12 @@
; GFX9-NEXT: s_mov_b32 s33, s32
; GFX9-NEXT: s_or_saveexec_b64 s[36:37], -1
; GFX9-NEXT: buffer_store_dword v40, off, s[0:3], s33 ; 4-byte Folded Spill
+; GFX9-NEXT: buffer_store_dword v41, off, s[0:3], s33 offset:4 ; 4-byte Folded Spill
; GFX9-NEXT: s_mov_b64 exec, s[36:37]
; GFX9-NEXT: global_load_dword v0, v[0:1], off
-; GFX9-NEXT: v_writelane_b32 v40, s34, 2
; GFX9-NEXT: s_addk_i32 s32, 0x400
; GFX9-NEXT: v_writelane_b32 v40, s30, 0
+; GFX9-NEXT: v_writelane_b32 v41, s34, 0
; GFX9-NEXT: v_writelane_b32 v40, s31, 1
; GFX9-NEXT: s_getpc_b64 s[34:35]
; GFX9-NEXT: s_add_u32 s34, s34, external_void_func_v2f16@rel32@lo+4
@@ -3922,9 +4282,10 @@
; GFX9-NEXT: s_swappc_b64 s[30:31], s[34:35]
; GFX9-NEXT: v_readlane_b32 s31, v40, 1
; GFX9-NEXT: v_readlane_b32 s30, v40, 0
-; GFX9-NEXT: v_readlane_b32 s34, v40, 2
+; GFX9-NEXT: v_readlane_b32 s34, v41, 0
; GFX9-NEXT: s_or_saveexec_b64 s[36:37], -1
; GFX9-NEXT: buffer_load_dword v40, off, s[0:3], s33 ; 4-byte Folded Reload
+; GFX9-NEXT: buffer_load_dword v41, off, s[0:3], s33 offset:4 ; 4-byte Folded Reload
; GFX9-NEXT: s_mov_b64 exec, s[36:37]
; GFX9-NEXT: s_addk_i32 s32, 0xfc00
; GFX9-NEXT: s_mov_b32 s33, s34
@@ -3938,22 +4299,25 @@
; GFX10-NEXT: s_mov_b32 s33, s32
; GFX10-NEXT: s_or_saveexec_b32 s35, -1
; GFX10-NEXT: buffer_store_dword v40, off, s[0:3], s33 ; 4-byte Folded Spill
+; GFX10-NEXT: buffer_store_dword v41, off, s[0:3], s33 offset:4 ; 4-byte Folded Spill
; GFX10-NEXT: s_waitcnt_depctr 0xffe3
; GFX10-NEXT: s_mov_b32 exec_lo, s35
; GFX10-NEXT: global_load_dword v0, v[0:1], off
-; GFX10-NEXT: v_writelane_b32 v40, s34, 2
+; GFX10-NEXT: v_writelane_b32 v40, s30, 0
; GFX10-NEXT: s_addk_i32 s32, 0x200
+; GFX10-NEXT: v_writelane_b32 v41, s34, 0
; GFX10-NEXT: s_getpc_b64 s[34:35]
; GFX10-NEXT: s_add_u32 s34, s34, external_void_func_v2f16@rel32@lo+4
; GFX10-NEXT: s_addc_u32 s35, s35, external_void_func_v2f16@rel32@hi+12
-; GFX10-NEXT: v_writelane_b32 v40, s30, 0
; GFX10-NEXT: v_writelane_b32 v40, s31, 1
; GFX10-NEXT: s_swappc_b64 s[30:31], s[34:35]
; GFX10-NEXT: v_readlane_b32 s31, v40, 1
; GFX10-NEXT: v_readlane_b32 s30, v40, 0
-; GFX10-NEXT: v_readlane_b32 s34, v40, 2
+; GFX10-NEXT: v_readlane_b32 s34, v41, 0
; GFX10-NEXT: s_or_saveexec_b32 s35, -1
-; GFX10-NEXT: buffer_load_dword v40, off, s[0:3], s33 ; 4-byte Folded Reload
+; GFX10-NEXT: s_clause 0x1
+; GFX10-NEXT: buffer_load_dword v40, off, s[0:3], s33
+; GFX10-NEXT: buffer_load_dword v41, off, s[0:3], s33 offset:4
; GFX10-NEXT: s_waitcnt_depctr 0xffe3
; GFX10-NEXT: s_mov_b32 exec_lo, s35
; GFX10-NEXT: s_addk_i32 s32, 0xfe00
@@ -3967,23 +4331,27 @@
; GFX11-NEXT: s_mov_b32 s0, s33
; GFX11-NEXT: s_mov_b32 s33, s32
; GFX11-NEXT: s_or_saveexec_b32 s1, -1
-; GFX11-NEXT: scratch_store_b32 off, v40, s33 ; 4-byte Folded Spill
+; GFX11-NEXT: s_clause 0x1
+; GFX11-NEXT: scratch_store_b32 off, v40, s33
+; GFX11-NEXT: scratch_store_b32 off, v41, s33 offset:4
; GFX11-NEXT: s_mov_b32 exec_lo, s1
; GFX11-NEXT: global_load_b32 v0, v[0:1], off
-; GFX11-NEXT: v_writelane_b32 v40, s0, 2
+; GFX11-NEXT: v_writelane_b32 v40, s30, 0
; GFX11-NEXT: s_add_i32 s32, s32, 16
+; GFX11-NEXT: v_writelane_b32 v41, s0, 0
; GFX11-NEXT: s_getpc_b64 s[0:1]
; GFX11-NEXT: s_add_u32 s0, s0, external_void_func_v2f16@rel32@lo+4
; GFX11-NEXT: s_addc_u32 s1, s1, external_void_func_v2f16@rel32@hi+12
-; GFX11-NEXT: v_writelane_b32 v40, s30, 0
; GFX11-NEXT: v_writelane_b32 v40, s31, 1
; GFX11-NEXT: s_swappc_b64 s[30:31], s[0:1]
; GFX11-NEXT: s_delay_alu instid0(VALU_DEP_1)
; GFX11-NEXT: v_readlane_b32 s31, v40, 1
; GFX11-NEXT: v_readlane_b32 s30, v40, 0
-; GFX11-NEXT: v_readlane_b32 s0, v40, 2
+; GFX11-NEXT: v_readlane_b32 s0, v41, 0
; GFX11-NEXT: s_or_saveexec_b32 s1, -1
-; GFX11-NEXT: scratch_load_b32 v40, off, s33 ; 4-byte Folded Reload
+; GFX11-NEXT: s_clause 0x1
+; GFX11-NEXT: scratch_load_b32 v40, off, s33
+; GFX11-NEXT: scratch_load_b32 v41, off, s33 offset:4
; GFX11-NEXT: s_mov_b32 exec_lo, s1
; GFX11-NEXT: s_add_i32 s32, s32, -16
; GFX11-NEXT: s_mov_b32 s33, s0
@@ -3997,22 +4365,25 @@
; GFX10-SCRATCH-NEXT: s_mov_b32 s33, s32
; GFX10-SCRATCH-NEXT: s_or_saveexec_b32 s1, -1
; GFX10-SCRATCH-NEXT: scratch_store_dword off, v40, s33 ; 4-byte Folded Spill
+; GFX10-SCRATCH-NEXT: scratch_store_dword off, v41, s33 offset:4 ; 4-byte Folded Spill
; GFX10-SCRATCH-NEXT: s_waitcnt_depctr 0xffe3
; GFX10-SCRATCH-NEXT: s_mov_b32 exec_lo, s1
; GFX10-SCRATCH-NEXT: global_load_dword v0, v[0:1], off
-; GFX10-SCRATCH-NEXT: v_writelane_b32 v40, s0, 2
+; GFX10-SCRATCH-NEXT: v_writelane_b32 v40, s30, 0
; GFX10-SCRATCH-NEXT: s_add_i32 s32, s32, 16
+; GFX10-SCRATCH-NEXT: v_writelane_b32 v41, s0, 0
; GFX10-SCRATCH-NEXT: s_getpc_b64 s[0:1]
; GFX10-SCRATCH-NEXT: s_add_u32 s0, s0, external_void_func_v2f16@rel32@lo+4
; GFX10-SCRATCH-NEXT: s_addc_u32 s1, s1, external_void_func_v2f16@rel32@hi+12
-; GFX10-SCRATCH-NEXT: v_writelane_b32 v40, s30, 0
; GFX10-SCRATCH-NEXT: v_writelane_b32 v40, s31, 1
; GFX10-SCRATCH-NEXT: s_swappc_b64 s[30:31], s[0:1]
; GFX10-SCRATCH-NEXT: v_readlane_b32 s31, v40, 1
; GFX10-SCRATCH-NEXT: v_readlane_b32 s30, v40, 0
-; GFX10-SCRATCH-NEXT: v_readlane_b32 s0, v40, 2
+; GFX10-SCRATCH-NEXT: v_readlane_b32 s0, v41, 0
; GFX10-SCRATCH-NEXT: s_or_saveexec_b32 s1, -1
-; GFX10-SCRATCH-NEXT: scratch_load_dword v40, off, s33 ; 4-byte Folded Reload
+; GFX10-SCRATCH-NEXT: s_clause 0x1
+; GFX10-SCRATCH-NEXT: scratch_load_dword v40, off, s33
+; GFX10-SCRATCH-NEXT: scratch_load_dword v41, off, s33 offset:4
; GFX10-SCRATCH-NEXT: s_waitcnt_depctr 0xffe3
; GFX10-SCRATCH-NEXT: s_mov_b32 exec_lo, s1
; GFX10-SCRATCH-NEXT: s_add_i32 s32, s32, -16
@@ -4032,11 +4403,12 @@
; GFX9-NEXT: s_mov_b32 s33, s32
; GFX9-NEXT: s_or_saveexec_b64 s[36:37], -1
; GFX9-NEXT: buffer_store_dword v40, off, s[0:3], s33 ; 4-byte Folded Spill
+; GFX9-NEXT: buffer_store_dword v41, off, s[0:3], s33 offset:4 ; 4-byte Folded Spill
; GFX9-NEXT: s_mov_b64 exec, s[36:37]
; GFX9-NEXT: global_load_dwordx2 v[0:1], v[0:1], off
-; GFX9-NEXT: v_writelane_b32 v40, s34, 2
; GFX9-NEXT: s_addk_i32 s32, 0x400
; GFX9-NEXT: v_writelane_b32 v40, s30, 0
+; GFX9-NEXT: v_writelane_b32 v41, s34, 0
; GFX9-NEXT: v_writelane_b32 v40, s31, 1
; GFX9-NEXT: s_getpc_b64 s[34:35]
; GFX9-NEXT: s_add_u32 s34, s34, external_void_func_v2i32@rel32@lo+4
@@ -4044,9 +4416,10 @@
; GFX9-NEXT: s_swappc_b64 s[30:31], s[34:35]
; GFX9-NEXT: v_readlane_b32 s31, v40, 1
; GFX9-NEXT: v_readlane_b32 s30, v40, 0
-; GFX9-NEXT: v_readlane_b32 s34, v40, 2
+; GFX9-NEXT: v_readlane_b32 s34, v41, 0
; GFX9-NEXT: s_or_saveexec_b64 s[36:37], -1
; GFX9-NEXT: buffer_load_dword v40, off, s[0:3], s33 ; 4-byte Folded Reload
+; GFX9-NEXT: buffer_load_dword v41, off, s[0:3], s33 offset:4 ; 4-byte Folded Reload
; GFX9-NEXT: s_mov_b64 exec, s[36:37]
; GFX9-NEXT: s_addk_i32 s32, 0xfc00
; GFX9-NEXT: s_mov_b32 s33, s34
@@ -4060,22 +4433,25 @@
; GFX10-NEXT: s_mov_b32 s33, s32
; GFX10-NEXT: s_or_saveexec_b32 s35, -1
; GFX10-NEXT: buffer_store_dword v40, off, s[0:3], s33 ; 4-byte Folded Spill
+; GFX10-NEXT: buffer_store_dword v41, off, s[0:3], s33 offset:4 ; 4-byte Folded Spill
; GFX10-NEXT: s_waitcnt_depctr 0xffe3
; GFX10-NEXT: s_mov_b32 exec_lo, s35
; GFX10-NEXT: global_load_dwordx2 v[0:1], v[0:1], off
-; GFX10-NEXT: v_writelane_b32 v40, s34, 2
+; GFX10-NEXT: v_writelane_b32 v40, s30, 0
; GFX10-NEXT: s_addk_i32 s32, 0x200
+; GFX10-NEXT: v_writelane_b32 v41, s34, 0
; GFX10-NEXT: s_getpc_b64 s[34:35]
; GFX10-NEXT: s_add_u32 s34, s34, external_void_func_v2i32@rel32@lo+4
; GFX10-NEXT: s_addc_u32 s35, s35, external_void_func_v2i32@rel32@hi+12
-; GFX10-NEXT: v_writelane_b32 v40, s30, 0
; GFX10-NEXT: v_writelane_b32 v40, s31, 1
; GFX10-NEXT: s_swappc_b64 s[30:31], s[34:35]
; GFX10-NEXT: v_readlane_b32 s31, v40, 1
; GFX10-NEXT: v_readlane_b32 s30, v40, 0
-; GFX10-NEXT: v_readlane_b32 s34, v40, 2
+; GFX10-NEXT: v_readlane_b32 s34, v41, 0
; GFX10-NEXT: s_or_saveexec_b32 s35, -1
-; GFX10-NEXT: buffer_load_dword v40, off, s[0:3], s33 ; 4-byte Folded Reload
+; GFX10-NEXT: s_clause 0x1
+; GFX10-NEXT: buffer_load_dword v40, off, s[0:3], s33
+; GFX10-NEXT: buffer_load_dword v41, off, s[0:3], s33 offset:4
; GFX10-NEXT: s_waitcnt_depctr 0xffe3
; GFX10-NEXT: s_mov_b32 exec_lo, s35
; GFX10-NEXT: s_addk_i32 s32, 0xfe00
@@ -4089,23 +4465,27 @@
; GFX11-NEXT: s_mov_b32 s0, s33
; GFX11-NEXT: s_mov_b32 s33, s32
; GFX11-NEXT: s_or_saveexec_b32 s1, -1
-; GFX11-NEXT: scratch_store_b32 off, v40, s33 ; 4-byte Folded Spill
+; GFX11-NEXT: s_clause 0x1
+; GFX11-NEXT: scratch_store_b32 off, v40, s33
+; GFX11-NEXT: scratch_store_b32 off, v41, s33 offset:4
; GFX11-NEXT: s_mov_b32 exec_lo, s1
; GFX11-NEXT: global_load_b64 v[0:1], v[0:1], off
-; GFX11-NEXT: v_writelane_b32 v40, s0, 2
+; GFX11-NEXT: v_writelane_b32 v40, s30, 0
; GFX11-NEXT: s_add_i32 s32, s32, 16
+; GFX11-NEXT: v_writelane_b32 v41, s0, 0
; GFX11-NEXT: s_getpc_b64 s[0:1]
; GFX11-NEXT: s_add_u32 s0, s0, external_void_func_v2i32@rel32@lo+4
; GFX11-NEXT: s_addc_u32 s1, s1, external_void_func_v2i32@rel32@hi+12
-; GFX11-NEXT: v_writelane_b32 v40, s30, 0
; GFX11-NEXT: v_writelane_b32 v40, s31, 1
; GFX11-NEXT: s_swappc_b64 s[30:31], s[0:1]
; GFX11-NEXT: s_delay_alu instid0(VALU_DEP_1)
; GFX11-NEXT: v_readlane_b32 s31, v40, 1
; GFX11-NEXT: v_readlane_b32 s30, v40, 0
-; GFX11-NEXT: v_readlane_b32 s0, v40, 2
+; GFX11-NEXT: v_readlane_b32 s0, v41, 0
; GFX11-NEXT: s_or_saveexec_b32 s1, -1
-; GFX11-NEXT: scratch_load_b32 v40, off, s33 ; 4-byte Folded Reload
+; GFX11-NEXT: s_clause 0x1
+; GFX11-NEXT: scratch_load_b32 v40, off, s33
+; GFX11-NEXT: scratch_load_b32 v41, off, s33 offset:4
; GFX11-NEXT: s_mov_b32 exec_lo, s1
; GFX11-NEXT: s_add_i32 s32, s32, -16
; GFX11-NEXT: s_mov_b32 s33, s0
@@ -4119,22 +4499,25 @@
; GFX10-SCRATCH-NEXT: s_mov_b32 s33, s32
; GFX10-SCRATCH-NEXT: s_or_saveexec_b32 s1, -1
; GFX10-SCRATCH-NEXT: scratch_store_dword off, v40, s33 ; 4-byte Folded Spill
+; GFX10-SCRATCH-NEXT: scratch_store_dword off, v41, s33 offset:4 ; 4-byte Folded Spill
; GFX10-SCRATCH-NEXT: s_waitcnt_depctr 0xffe3
; GFX10-SCRATCH-NEXT: s_mov_b32 exec_lo, s1
; GFX10-SCRATCH-NEXT: global_load_dwordx2 v[0:1], v[0:1], off
-; GFX10-SCRATCH-NEXT: v_writelane_b32 v40, s0, 2
+; GFX10-SCRATCH-NEXT: v_writelane_b32 v40, s30, 0
; GFX10-SCRATCH-NEXT: s_add_i32 s32, s32, 16
+; GFX10-SCRATCH-NEXT: v_writelane_b32 v41, s0, 0
; GFX10-SCRATCH-NEXT: s_getpc_b64 s[0:1]
; GFX10-SCRATCH-NEXT: s_add_u32 s0, s0, external_void_func_v2i32@rel32@lo+4
; GFX10-SCRATCH-NEXT: s_addc_u32 s1, s1, external_void_func_v2i32@rel32@hi+12
-; GFX10-SCRATCH-NEXT: v_writelane_b32 v40, s30, 0
; GFX10-SCRATCH-NEXT: v_writelane_b32 v40, s31, 1
; GFX10-SCRATCH-NEXT: s_swappc_b64 s[30:31], s[0:1]
; GFX10-SCRATCH-NEXT: v_readlane_b32 s31, v40, 1
; GFX10-SCRATCH-NEXT: v_readlane_b32 s30, v40, 0
-; GFX10-SCRATCH-NEXT: v_readlane_b32 s0, v40, 2
+; GFX10-SCRATCH-NEXT: v_readlane_b32 s0, v41, 0
; GFX10-SCRATCH-NEXT: s_or_saveexec_b32 s1, -1
-; GFX10-SCRATCH-NEXT: scratch_load_dword v40, off, s33 ; 4-byte Folded Reload
+; GFX10-SCRATCH-NEXT: s_clause 0x1
+; GFX10-SCRATCH-NEXT: scratch_load_dword v40, off, s33
+; GFX10-SCRATCH-NEXT: scratch_load_dword v41, off, s33 offset:4
; GFX10-SCRATCH-NEXT: s_waitcnt_depctr 0xffe3
; GFX10-SCRATCH-NEXT: s_mov_b32 exec_lo, s1
; GFX10-SCRATCH-NEXT: s_add_i32 s32, s32, -16
@@ -4154,12 +4537,13 @@
; GFX9-NEXT: s_mov_b32 s33, s32
; GFX9-NEXT: s_or_saveexec_b64 s[36:37], -1
; GFX9-NEXT: buffer_store_dword v40, off, s[0:3], s33 ; 4-byte Folded Spill
+; GFX9-NEXT: buffer_store_dword v41, off, s[0:3], s33 offset:4 ; 4-byte Folded Spill
; GFX9-NEXT: s_mov_b64 exec, s[36:37]
-; GFX9-NEXT: v_writelane_b32 v40, s34, 2
; GFX9-NEXT: s_addk_i32 s32, 0x400
; GFX9-NEXT: v_writelane_b32 v40, s30, 0
; GFX9-NEXT: v_mov_b32_e32 v0, 1
; GFX9-NEXT: v_mov_b32_e32 v1, 2
+; GFX9-NEXT: v_writelane_b32 v41, s34, 0
; GFX9-NEXT: v_writelane_b32 v40, s31, 1
; GFX9-NEXT: s_getpc_b64 s[34:35]
; GFX9-NEXT: s_add_u32 s34, s34, external_void_func_v2i32@rel32@lo+4
@@ -4167,9 +4551,10 @@
; GFX9-NEXT: s_swappc_b64 s[30:31], s[34:35]
; GFX9-NEXT: v_readlane_b32 s31, v40, 1
; GFX9-NEXT: v_readlane_b32 s30, v40, 0
-; GFX9-NEXT: v_readlane_b32 s34, v40, 2
+; GFX9-NEXT: v_readlane_b32 s34, v41, 0
; GFX9-NEXT: s_or_saveexec_b64 s[36:37], -1
; GFX9-NEXT: buffer_load_dword v40, off, s[0:3], s33 ; 4-byte Folded Reload
+; GFX9-NEXT: buffer_load_dword v41, off, s[0:3], s33 offset:4 ; 4-byte Folded Reload
; GFX9-NEXT: s_mov_b64 exec, s[36:37]
; GFX9-NEXT: s_addk_i32 s32, 0xfc00
; GFX9-NEXT: s_mov_b32 s33, s34
@@ -4183,23 +4568,26 @@
; GFX10-NEXT: s_mov_b32 s33, s32
; GFX10-NEXT: s_or_saveexec_b32 s35, -1
; GFX10-NEXT: buffer_store_dword v40, off, s[0:3], s33 ; 4-byte Folded Spill
+; GFX10-NEXT: buffer_store_dword v41, off, s[0:3], s33 offset:4 ; 4-byte Folded Spill
; GFX10-NEXT: s_waitcnt_depctr 0xffe3
; GFX10-NEXT: s_mov_b32 exec_lo, s35
-; GFX10-NEXT: v_writelane_b32 v40, s34, 2
+; GFX10-NEXT: v_writelane_b32 v40, s30, 0
; GFX10-NEXT: v_mov_b32_e32 v0, 1
; GFX10-NEXT: v_mov_b32_e32 v1, 2
; GFX10-NEXT: s_addk_i32 s32, 0x200
+; GFX10-NEXT: v_writelane_b32 v41, s34, 0
+; GFX10-NEXT: v_writelane_b32 v40, s31, 1
; GFX10-NEXT: s_getpc_b64 s[34:35]
; GFX10-NEXT: s_add_u32 s34, s34, external_void_func_v2i32@rel32@lo+4
; GFX10-NEXT: s_addc_u32 s35, s35, external_void_func_v2i32@rel32@hi+12
-; GFX10-NEXT: v_writelane_b32 v40, s30, 0
-; GFX10-NEXT: v_writelane_b32 v40, s31, 1
; GFX10-NEXT: s_swappc_b64 s[30:31], s[34:35]
; GFX10-NEXT: v_readlane_b32 s31, v40, 1
; GFX10-NEXT: v_readlane_b32 s30, v40, 0
-; GFX10-NEXT: v_readlane_b32 s34, v40, 2
+; GFX10-NEXT: v_readlane_b32 s34, v41, 0
; GFX10-NEXT: s_or_saveexec_b32 s35, -1
-; GFX10-NEXT: buffer_load_dword v40, off, s[0:3], s33 ; 4-byte Folded Reload
+; GFX10-NEXT: s_clause 0x1
+; GFX10-NEXT: buffer_load_dword v40, off, s[0:3], s33
+; GFX10-NEXT: buffer_load_dword v41, off, s[0:3], s33 offset:4
; GFX10-NEXT: s_waitcnt_depctr 0xffe3
; GFX10-NEXT: s_mov_b32 exec_lo, s35
; GFX10-NEXT: s_addk_i32 s32, 0xfe00
@@ -4213,23 +4601,27 @@
; GFX11-NEXT: s_mov_b32 s0, s33
; GFX11-NEXT: s_mov_b32 s33, s32
; GFX11-NEXT: s_or_saveexec_b32 s1, -1
-; GFX11-NEXT: scratch_store_b32 off, v40, s33 ; 4-byte Folded Spill
+; GFX11-NEXT: s_clause 0x1
+; GFX11-NEXT: scratch_store_b32 off, v40, s33
+; GFX11-NEXT: scratch_store_b32 off, v41, s33 offset:4
; GFX11-NEXT: s_mov_b32 exec_lo, s1
-; GFX11-NEXT: v_writelane_b32 v40, s0, 2
+; GFX11-NEXT: v_writelane_b32 v40, s30, 0
; GFX11-NEXT: v_dual_mov_b32 v0, 1 :: v_dual_mov_b32 v1, 2
; GFX11-NEXT: s_add_i32 s32, s32, 16
+; GFX11-NEXT: v_writelane_b32 v41, s0, 0
+; GFX11-NEXT: v_writelane_b32 v40, s31, 1
; GFX11-NEXT: s_getpc_b64 s[0:1]
; GFX11-NEXT: s_add_u32 s0, s0, external_void_func_v2i32@rel32@lo+4
; GFX11-NEXT: s_addc_u32 s1, s1, external_void_func_v2i32@rel32@hi+12
-; GFX11-NEXT: v_writelane_b32 v40, s30, 0
-; GFX11-NEXT: v_writelane_b32 v40, s31, 1
+; GFX11-NEXT: s_delay_alu instid0(SALU_CYCLE_1)
; GFX11-NEXT: s_swappc_b64 s[30:31], s[0:1]
-; GFX11-NEXT: s_delay_alu instid0(VALU_DEP_1)
; GFX11-NEXT: v_readlane_b32 s31, v40, 1
; GFX11-NEXT: v_readlane_b32 s30, v40, 0
-; GFX11-NEXT: v_readlane_b32 s0, v40, 2
+; GFX11-NEXT: v_readlane_b32 s0, v41, 0
; GFX11-NEXT: s_or_saveexec_b32 s1, -1
-; GFX11-NEXT: scratch_load_b32 v40, off, s33 ; 4-byte Folded Reload
+; GFX11-NEXT: s_clause 0x1
+; GFX11-NEXT: scratch_load_b32 v40, off, s33
+; GFX11-NEXT: scratch_load_b32 v41, off, s33 offset:4
; GFX11-NEXT: s_mov_b32 exec_lo, s1
; GFX11-NEXT: s_add_i32 s32, s32, -16
; GFX11-NEXT: s_mov_b32 s33, s0
@@ -4243,23 +4635,26 @@
; GFX10-SCRATCH-NEXT: s_mov_b32 s33, s32
; GFX10-SCRATCH-NEXT: s_or_saveexec_b32 s1, -1
; GFX10-SCRATCH-NEXT: scratch_store_dword off, v40, s33 ; 4-byte Folded Spill
+; GFX10-SCRATCH-NEXT: scratch_store_dword off, v41, s33 offset:4 ; 4-byte Folded Spill
; GFX10-SCRATCH-NEXT: s_waitcnt_depctr 0xffe3
; GFX10-SCRATCH-NEXT: s_mov_b32 exec_lo, s1
-; GFX10-SCRATCH-NEXT: v_writelane_b32 v40, s0, 2
+; GFX10-SCRATCH-NEXT: v_writelane_b32 v40, s30, 0
; GFX10-SCRATCH-NEXT: v_mov_b32_e32 v0, 1
; GFX10-SCRATCH-NEXT: v_mov_b32_e32 v1, 2
; GFX10-SCRATCH-NEXT: s_add_i32 s32, s32, 16
+; GFX10-SCRATCH-NEXT: v_writelane_b32 v41, s0, 0
+; GFX10-SCRATCH-NEXT: v_writelane_b32 v40, s31, 1
; GFX10-SCRATCH-NEXT: s_getpc_b64 s[0:1]
; GFX10-SCRATCH-NEXT: s_add_u32 s0, s0, external_void_func_v2i32@rel32@lo+4
; GFX10-SCRATCH-NEXT: s_addc_u32 s1, s1, external_void_func_v2i32@rel32@hi+12
-; GFX10-SCRATCH-NEXT: v_writelane_b32 v40, s30, 0
-; GFX10-SCRATCH-NEXT: v_writelane_b32 v40, s31, 1
; GFX10-SCRATCH-NEXT: s_swappc_b64 s[30:31], s[0:1]
; GFX10-SCRATCH-NEXT: v_readlane_b32 s31, v40, 1
; GFX10-SCRATCH-NEXT: v_readlane_b32 s30, v40, 0
-; GFX10-SCRATCH-NEXT: v_readlane_b32 s0, v40, 2
+; GFX10-SCRATCH-NEXT: v_readlane_b32 s0, v41, 0
; GFX10-SCRATCH-NEXT: s_or_saveexec_b32 s1, -1
-; GFX10-SCRATCH-NEXT: scratch_load_dword v40, off, s33 ; 4-byte Folded Reload
+; GFX10-SCRATCH-NEXT: s_clause 0x1
+; GFX10-SCRATCH-NEXT: scratch_load_dword v40, off, s33
+; GFX10-SCRATCH-NEXT: scratch_load_dword v41, off, s33 offset:4
; GFX10-SCRATCH-NEXT: s_waitcnt_depctr 0xffe3
; GFX10-SCRATCH-NEXT: s_mov_b32 exec_lo, s1
; GFX10-SCRATCH-NEXT: s_add_i32 s32, s32, -16
@@ -4278,13 +4673,14 @@
; GFX9-NEXT: s_mov_b32 s33, s32
; GFX9-NEXT: s_or_saveexec_b64 s[36:37], -1
; GFX9-NEXT: buffer_store_dword v40, off, s[0:3], s33 ; 4-byte Folded Spill
+; GFX9-NEXT: buffer_store_dword v41, off, s[0:3], s33 offset:4 ; 4-byte Folded Spill
; GFX9-NEXT: s_mov_b64 exec, s[36:37]
-; GFX9-NEXT: v_writelane_b32 v40, s34, 2
; GFX9-NEXT: s_addk_i32 s32, 0x400
; GFX9-NEXT: v_writelane_b32 v40, s30, 0
; GFX9-NEXT: v_mov_b32_e32 v0, 3
; GFX9-NEXT: v_mov_b32_e32 v1, 4
; GFX9-NEXT: v_mov_b32_e32 v2, 5
+; GFX9-NEXT: v_writelane_b32 v41, s34, 0
; GFX9-NEXT: v_writelane_b32 v40, s31, 1
; GFX9-NEXT: s_getpc_b64 s[34:35]
; GFX9-NEXT: s_add_u32 s34, s34, external_void_func_v3i32@rel32@lo+4
@@ -4292,9 +4688,10 @@
; GFX9-NEXT: s_swappc_b64 s[30:31], s[34:35]
; GFX9-NEXT: v_readlane_b32 s31, v40, 1
; GFX9-NEXT: v_readlane_b32 s30, v40, 0
-; GFX9-NEXT: v_readlane_b32 s34, v40, 2
+; GFX9-NEXT: v_readlane_b32 s34, v41, 0
; GFX9-NEXT: s_or_saveexec_b64 s[36:37], -1
; GFX9-NEXT: buffer_load_dword v40, off, s[0:3], s33 ; 4-byte Folded Reload
+; GFX9-NEXT: buffer_load_dword v41, off, s[0:3], s33 offset:4 ; 4-byte Folded Reload
; GFX9-NEXT: s_mov_b64 exec, s[36:37]
; GFX9-NEXT: s_addk_i32 s32, 0xfc00
; GFX9-NEXT: s_mov_b32 s33, s34
@@ -4308,24 +4705,27 @@
; GFX10-NEXT: s_mov_b32 s33, s32
; GFX10-NEXT: s_or_saveexec_b32 s35, -1
; GFX10-NEXT: buffer_store_dword v40, off, s[0:3], s33 ; 4-byte Folded Spill
+; GFX10-NEXT: buffer_store_dword v41, off, s[0:3], s33 offset:4 ; 4-byte Folded Spill
; GFX10-NEXT: s_waitcnt_depctr 0xffe3
; GFX10-NEXT: s_mov_b32 exec_lo, s35
-; GFX10-NEXT: v_writelane_b32 v40, s34, 2
+; GFX10-NEXT: v_writelane_b32 v40, s30, 0
; GFX10-NEXT: v_mov_b32_e32 v0, 3
; GFX10-NEXT: v_mov_b32_e32 v1, 4
; GFX10-NEXT: v_mov_b32_e32 v2, 5
; GFX10-NEXT: s_addk_i32 s32, 0x200
-; GFX10-NEXT: v_writelane_b32 v40, s30, 0
+; GFX10-NEXT: v_writelane_b32 v41, s34, 0
+; GFX10-NEXT: v_writelane_b32 v40, s31, 1
; GFX10-NEXT: s_getpc_b64 s[34:35]
; GFX10-NEXT: s_add_u32 s34, s34, external_void_func_v3i32@rel32@lo+4
; GFX10-NEXT: s_addc_u32 s35, s35, external_void_func_v3i32@rel32@hi+12
-; GFX10-NEXT: v_writelane_b32 v40, s31, 1
; GFX10-NEXT: s_swappc_b64 s[30:31], s[34:35]
; GFX10-NEXT: v_readlane_b32 s31, v40, 1
; GFX10-NEXT: v_readlane_b32 s30, v40, 0
-; GFX10-NEXT: v_readlane_b32 s34, v40, 2
+; GFX10-NEXT: v_readlane_b32 s34, v41, 0
; GFX10-NEXT: s_or_saveexec_b32 s35, -1
-; GFX10-NEXT: buffer_load_dword v40, off, s[0:3], s33 ; 4-byte Folded Reload
+; GFX10-NEXT: s_clause 0x1
+; GFX10-NEXT: buffer_load_dword v40, off, s[0:3], s33
+; GFX10-NEXT: buffer_load_dword v41, off, s[0:3], s33 offset:4
; GFX10-NEXT: s_waitcnt_depctr 0xffe3
; GFX10-NEXT: s_mov_b32 exec_lo, s35
; GFX10-NEXT: s_addk_i32 s32, 0xfe00
@@ -4339,24 +4739,28 @@
; GFX11-NEXT: s_mov_b32 s0, s33
; GFX11-NEXT: s_mov_b32 s33, s32
; GFX11-NEXT: s_or_saveexec_b32 s1, -1
-; GFX11-NEXT: scratch_store_b32 off, v40, s33 ; 4-byte Folded Spill
+; GFX11-NEXT: s_clause 0x1
+; GFX11-NEXT: scratch_store_b32 off, v40, s33
+; GFX11-NEXT: scratch_store_b32 off, v41, s33 offset:4
; GFX11-NEXT: s_mov_b32 exec_lo, s1
-; GFX11-NEXT: v_writelane_b32 v40, s0, 2
+; GFX11-NEXT: v_writelane_b32 v40, s30, 0
; GFX11-NEXT: v_dual_mov_b32 v0, 3 :: v_dual_mov_b32 v1, 4
; GFX11-NEXT: v_mov_b32_e32 v2, 5
; GFX11-NEXT: s_add_i32 s32, s32, 16
-; GFX11-NEXT: v_writelane_b32 v40, s30, 0
+; GFX11-NEXT: v_writelane_b32 v41, s0, 0
+; GFX11-NEXT: v_writelane_b32 v40, s31, 1
; GFX11-NEXT: s_getpc_b64 s[0:1]
; GFX11-NEXT: s_add_u32 s0, s0, external_void_func_v3i32@rel32@lo+4
; GFX11-NEXT: s_addc_u32 s1, s1, external_void_func_v3i32@rel32@hi+12
-; GFX11-NEXT: v_writelane_b32 v40, s31, 1
+; GFX11-NEXT: s_delay_alu instid0(SALU_CYCLE_1)
; GFX11-NEXT: s_swappc_b64 s[30:31], s[0:1]
-; GFX11-NEXT: s_delay_alu instid0(VALU_DEP_1)
; GFX11-NEXT: v_readlane_b32 s31, v40, 1
; GFX11-NEXT: v_readlane_b32 s30, v40, 0
-; GFX11-NEXT: v_readlane_b32 s0, v40, 2
+; GFX11-NEXT: v_readlane_b32 s0, v41, 0
; GFX11-NEXT: s_or_saveexec_b32 s1, -1
-; GFX11-NEXT: scratch_load_b32 v40, off, s33 ; 4-byte Folded Reload
+; GFX11-NEXT: s_clause 0x1
+; GFX11-NEXT: scratch_load_b32 v40, off, s33
+; GFX11-NEXT: scratch_load_b32 v41, off, s33 offset:4
; GFX11-NEXT: s_mov_b32 exec_lo, s1
; GFX11-NEXT: s_add_i32 s32, s32, -16
; GFX11-NEXT: s_mov_b32 s33, s0
@@ -4370,24 +4774,27 @@
; GFX10-SCRATCH-NEXT: s_mov_b32 s33, s32
; GFX10-SCRATCH-NEXT: s_or_saveexec_b32 s1, -1
; GFX10-SCRATCH-NEXT: scratch_store_dword off, v40, s33 ; 4-byte Folded Spill
+; GFX10-SCRATCH-NEXT: scratch_store_dword off, v41, s33 offset:4 ; 4-byte Folded Spill
; GFX10-SCRATCH-NEXT: s_waitcnt_depctr 0xffe3
; GFX10-SCRATCH-NEXT: s_mov_b32 exec_lo, s1
-; GFX10-SCRATCH-NEXT: v_writelane_b32 v40, s0, 2
+; GFX10-SCRATCH-NEXT: v_writelane_b32 v40, s30, 0
; GFX10-SCRATCH-NEXT: v_mov_b32_e32 v0, 3
; GFX10-SCRATCH-NEXT: v_mov_b32_e32 v1, 4
; GFX10-SCRATCH-NEXT: v_mov_b32_e32 v2, 5
; GFX10-SCRATCH-NEXT: s_add_i32 s32, s32, 16
-; GFX10-SCRATCH-NEXT: v_writelane_b32 v40, s30, 0
+; GFX10-SCRATCH-NEXT: v_writelane_b32 v41, s0, 0
+; GFX10-SCRATCH-NEXT: v_writelane_b32 v40, s31, 1
; GFX10-SCRATCH-NEXT: s_getpc_b64 s[0:1]
; GFX10-SCRATCH-NEXT: s_add_u32 s0, s0, external_void_func_v3i32@rel32@lo+4
; GFX10-SCRATCH-NEXT: s_addc_u32 s1, s1, external_void_func_v3i32@rel32@hi+12
-; GFX10-SCRATCH-NEXT: v_writelane_b32 v40, s31, 1
; GFX10-SCRATCH-NEXT: s_swappc_b64 s[30:31], s[0:1]
; GFX10-SCRATCH-NEXT: v_readlane_b32 s31, v40, 1
; GFX10-SCRATCH-NEXT: v_readlane_b32 s30, v40, 0
-; GFX10-SCRATCH-NEXT: v_readlane_b32 s0, v40, 2
+; GFX10-SCRATCH-NEXT: v_readlane_b32 s0, v41, 0
; GFX10-SCRATCH-NEXT: s_or_saveexec_b32 s1, -1
-; GFX10-SCRATCH-NEXT: scratch_load_dword v40, off, s33 ; 4-byte Folded Reload
+; GFX10-SCRATCH-NEXT: s_clause 0x1
+; GFX10-SCRATCH-NEXT: scratch_load_dword v40, off, s33
+; GFX10-SCRATCH-NEXT: scratch_load_dword v41, off, s33 offset:4
; GFX10-SCRATCH-NEXT: s_waitcnt_depctr 0xffe3
; GFX10-SCRATCH-NEXT: s_mov_b32 exec_lo, s1
; GFX10-SCRATCH-NEXT: s_add_i32 s32, s32, -16
@@ -4406,14 +4813,15 @@
; GFX9-NEXT: s_mov_b32 s33, s32
; GFX9-NEXT: s_or_saveexec_b64 s[36:37], -1
; GFX9-NEXT: buffer_store_dword v40, off, s[0:3], s33 ; 4-byte Folded Spill
+; GFX9-NEXT: buffer_store_dword v41, off, s[0:3], s33 offset:4 ; 4-byte Folded Spill
; GFX9-NEXT: s_mov_b64 exec, s[36:37]
-; GFX9-NEXT: v_writelane_b32 v40, s34, 2
; GFX9-NEXT: s_addk_i32 s32, 0x400
; GFX9-NEXT: v_writelane_b32 v40, s30, 0
; GFX9-NEXT: v_mov_b32_e32 v0, 3
; GFX9-NEXT: v_mov_b32_e32 v1, 4
; GFX9-NEXT: v_mov_b32_e32 v2, 5
; GFX9-NEXT: v_mov_b32_e32 v3, 6
+; GFX9-NEXT: v_writelane_b32 v41, s34, 0
; GFX9-NEXT: v_writelane_b32 v40, s31, 1
; GFX9-NEXT: s_getpc_b64 s[34:35]
; GFX9-NEXT: s_add_u32 s34, s34, external_void_func_v3i32_i32@rel32@lo+4
@@ -4421,9 +4829,10 @@
; GFX9-NEXT: s_swappc_b64 s[30:31], s[34:35]
; GFX9-NEXT: v_readlane_b32 s31, v40, 1
; GFX9-NEXT: v_readlane_b32 s30, v40, 0
-; GFX9-NEXT: v_readlane_b32 s34, v40, 2
+; GFX9-NEXT: v_readlane_b32 s34, v41, 0
; GFX9-NEXT: s_or_saveexec_b64 s[36:37], -1
; GFX9-NEXT: buffer_load_dword v40, off, s[0:3], s33 ; 4-byte Folded Reload
+; GFX9-NEXT: buffer_load_dword v41, off, s[0:3], s33 offset:4 ; 4-byte Folded Reload
; GFX9-NEXT: s_mov_b64 exec, s[36:37]
; GFX9-NEXT: s_addk_i32 s32, 0xfc00
; GFX9-NEXT: s_mov_b32 s33, s34
@@ -4437,25 +4846,28 @@
; GFX10-NEXT: s_mov_b32 s33, s32
; GFX10-NEXT: s_or_saveexec_b32 s35, -1
; GFX10-NEXT: buffer_store_dword v40, off, s[0:3], s33 ; 4-byte Folded Spill
+; GFX10-NEXT: buffer_store_dword v41, off, s[0:3], s33 offset:4 ; 4-byte Folded Spill
; GFX10-NEXT: s_waitcnt_depctr 0xffe3
; GFX10-NEXT: s_mov_b32 exec_lo, s35
-; GFX10-NEXT: v_writelane_b32 v40, s34, 2
+; GFX10-NEXT: v_writelane_b32 v40, s30, 0
; GFX10-NEXT: v_mov_b32_e32 v0, 3
; GFX10-NEXT: v_mov_b32_e32 v1, 4
; GFX10-NEXT: v_mov_b32_e32 v2, 5
; GFX10-NEXT: v_mov_b32_e32 v3, 6
-; GFX10-NEXT: v_writelane_b32 v40, s30, 0
; GFX10-NEXT: s_addk_i32 s32, 0x200
+; GFX10-NEXT: v_writelane_b32 v41, s34, 0
+; GFX10-NEXT: v_writelane_b32 v40, s31, 1
; GFX10-NEXT: s_getpc_b64 s[34:35]
; GFX10-NEXT: s_add_u32 s34, s34, external_void_func_v3i32_i32@rel32@lo+4
; GFX10-NEXT: s_addc_u32 s35, s35, external_void_func_v3i32_i32@rel32@hi+12
-; GFX10-NEXT: v_writelane_b32 v40, s31, 1
; GFX10-NEXT: s_swappc_b64 s[30:31], s[34:35]
; GFX10-NEXT: v_readlane_b32 s31, v40, 1
; GFX10-NEXT: v_readlane_b32 s30, v40, 0
-; GFX10-NEXT: v_readlane_b32 s34, v40, 2
+; GFX10-NEXT: v_readlane_b32 s34, v41, 0
; GFX10-NEXT: s_or_saveexec_b32 s35, -1
-; GFX10-NEXT: buffer_load_dword v40, off, s[0:3], s33 ; 4-byte Folded Reload
+; GFX10-NEXT: s_clause 0x1
+; GFX10-NEXT: buffer_load_dword v40, off, s[0:3], s33
+; GFX10-NEXT: buffer_load_dword v41, off, s[0:3], s33 offset:4
; GFX10-NEXT: s_waitcnt_depctr 0xffe3
; GFX10-NEXT: s_mov_b32 exec_lo, s35
; GFX10-NEXT: s_addk_i32 s32, 0xfe00
@@ -4469,24 +4881,28 @@
; GFX11-NEXT: s_mov_b32 s0, s33
; GFX11-NEXT: s_mov_b32 s33, s32
; GFX11-NEXT: s_or_saveexec_b32 s1, -1
-; GFX11-NEXT: scratch_store_b32 off, v40, s33 ; 4-byte Folded Spill
+; GFX11-NEXT: s_clause 0x1
+; GFX11-NEXT: scratch_store_b32 off, v40, s33
+; GFX11-NEXT: scratch_store_b32 off, v41, s33 offset:4
; GFX11-NEXT: s_mov_b32 exec_lo, s1
-; GFX11-NEXT: v_writelane_b32 v40, s0, 2
+; GFX11-NEXT: v_writelane_b32 v40, s30, 0
; GFX11-NEXT: v_dual_mov_b32 v0, 3 :: v_dual_mov_b32 v1, 4
; GFX11-NEXT: v_dual_mov_b32 v2, 5 :: v_dual_mov_b32 v3, 6
-; GFX11-NEXT: v_writelane_b32 v40, s30, 0
; GFX11-NEXT: s_add_i32 s32, s32, 16
+; GFX11-NEXT: v_writelane_b32 v41, s0, 0
+; GFX11-NEXT: v_writelane_b32 v40, s31, 1
; GFX11-NEXT: s_getpc_b64 s[0:1]
; GFX11-NEXT: s_add_u32 s0, s0, external_void_func_v3i32_i32@rel32@lo+4
; GFX11-NEXT: s_addc_u32 s1, s1, external_void_func_v3i32_i32@rel32@hi+12
-; GFX11-NEXT: v_writelane_b32 v40, s31, 1
+; GFX11-NEXT: s_delay_alu instid0(SALU_CYCLE_1)
; GFX11-NEXT: s_swappc_b64 s[30:31], s[0:1]
-; GFX11-NEXT: s_delay_alu instid0(VALU_DEP_1)
; GFX11-NEXT: v_readlane_b32 s31, v40, 1
; GFX11-NEXT: v_readlane_b32 s30, v40, 0
-; GFX11-NEXT: v_readlane_b32 s0, v40, 2
+; GFX11-NEXT: v_readlane_b32 s0, v41, 0
; GFX11-NEXT: s_or_saveexec_b32 s1, -1
-; GFX11-NEXT: scratch_load_b32 v40, off, s33 ; 4-byte Folded Reload
+; GFX11-NEXT: s_clause 0x1
+; GFX11-NEXT: scratch_load_b32 v40, off, s33
+; GFX11-NEXT: scratch_load_b32 v41, off, s33 offset:4
; GFX11-NEXT: s_mov_b32 exec_lo, s1
; GFX11-NEXT: s_add_i32 s32, s32, -16
; GFX11-NEXT: s_mov_b32 s33, s0
@@ -4500,25 +4916,28 @@
; GFX10-SCRATCH-NEXT: s_mov_b32 s33, s32
; GFX10-SCRATCH-NEXT: s_or_saveexec_b32 s1, -1
; GFX10-SCRATCH-NEXT: scratch_store_dword off, v40, s33 ; 4-byte Folded Spill
+; GFX10-SCRATCH-NEXT: scratch_store_dword off, v41, s33 offset:4 ; 4-byte Folded Spill
; GFX10-SCRATCH-NEXT: s_waitcnt_depctr 0xffe3
; GFX10-SCRATCH-NEXT: s_mov_b32 exec_lo, s1
-; GFX10-SCRATCH-NEXT: v_writelane_b32 v40, s0, 2
+; GFX10-SCRATCH-NEXT: v_writelane_b32 v40, s30, 0
; GFX10-SCRATCH-NEXT: v_mov_b32_e32 v0, 3
; GFX10-SCRATCH-NEXT: v_mov_b32_e32 v1, 4
; GFX10-SCRATCH-NEXT: v_mov_b32_e32 v2, 5
; GFX10-SCRATCH-NEXT: v_mov_b32_e32 v3, 6
-; GFX10-SCRATCH-NEXT: v_writelane_b32 v40, s30, 0
; GFX10-SCRATCH-NEXT: s_add_i32 s32, s32, 16
+; GFX10-SCRATCH-NEXT: v_writelane_b32 v41, s0, 0
+; GFX10-SCRATCH-NEXT: v_writelane_b32 v40, s31, 1
; GFX10-SCRATCH-NEXT: s_getpc_b64 s[0:1]
; GFX10-SCRATCH-NEXT: s_add_u32 s0, s0, external_void_func_v3i32_i32@rel32@lo+4
; GFX10-SCRATCH-NEXT: s_addc_u32 s1, s1, external_void_func_v3i32_i32@rel32@hi+12
-; GFX10-SCRATCH-NEXT: v_writelane_b32 v40, s31, 1
; GFX10-SCRATCH-NEXT: s_swappc_b64 s[30:31], s[0:1]
; GFX10-SCRATCH-NEXT: v_readlane_b32 s31, v40, 1
; GFX10-SCRATCH-NEXT: v_readlane_b32 s30, v40, 0
-; GFX10-SCRATCH-NEXT: v_readlane_b32 s0, v40, 2
+; GFX10-SCRATCH-NEXT: v_readlane_b32 s0, v41, 0
; GFX10-SCRATCH-NEXT: s_or_saveexec_b32 s1, -1
-; GFX10-SCRATCH-NEXT: scratch_load_dword v40, off, s33 ; 4-byte Folded Reload
+; GFX10-SCRATCH-NEXT: s_clause 0x1
+; GFX10-SCRATCH-NEXT: scratch_load_dword v40, off, s33
+; GFX10-SCRATCH-NEXT: scratch_load_dword v41, off, s33 offset:4
; GFX10-SCRATCH-NEXT: s_waitcnt_depctr 0xffe3
; GFX10-SCRATCH-NEXT: s_mov_b32 exec_lo, s1
; GFX10-SCRATCH-NEXT: s_add_i32 s32, s32, -16
@@ -4537,11 +4956,12 @@
; GFX9-NEXT: s_mov_b32 s33, s32
; GFX9-NEXT: s_or_saveexec_b64 s[36:37], -1
; GFX9-NEXT: buffer_store_dword v40, off, s[0:3], s33 ; 4-byte Folded Spill
+; GFX9-NEXT: buffer_store_dword v41, off, s[0:3], s33 offset:4 ; 4-byte Folded Spill
; GFX9-NEXT: s_mov_b64 exec, s[36:37]
; GFX9-NEXT: global_load_dwordx4 v[0:3], v[0:1], off
-; GFX9-NEXT: v_writelane_b32 v40, s34, 2
; GFX9-NEXT: s_addk_i32 s32, 0x400
; GFX9-NEXT: v_writelane_b32 v40, s30, 0
+; GFX9-NEXT: v_writelane_b32 v41, s34, 0
; GFX9-NEXT: v_writelane_b32 v40, s31, 1
; GFX9-NEXT: s_getpc_b64 s[34:35]
; GFX9-NEXT: s_add_u32 s34, s34, external_void_func_v4i32@rel32@lo+4
@@ -4549,9 +4969,10 @@
; GFX9-NEXT: s_swappc_b64 s[30:31], s[34:35]
; GFX9-NEXT: v_readlane_b32 s31, v40, 1
; GFX9-NEXT: v_readlane_b32 s30, v40, 0
-; GFX9-NEXT: v_readlane_b32 s34, v40, 2
+; GFX9-NEXT: v_readlane_b32 s34, v41, 0
; GFX9-NEXT: s_or_saveexec_b64 s[36:37], -1
; GFX9-NEXT: buffer_load_dword v40, off, s[0:3], s33 ; 4-byte Folded Reload
+; GFX9-NEXT: buffer_load_dword v41, off, s[0:3], s33 offset:4 ; 4-byte Folded Reload
; GFX9-NEXT: s_mov_b64 exec, s[36:37]
; GFX9-NEXT: s_addk_i32 s32, 0xfc00
; GFX9-NEXT: s_mov_b32 s33, s34
@@ -4565,22 +4986,25 @@
; GFX10-NEXT: s_mov_b32 s33, s32
; GFX10-NEXT: s_or_saveexec_b32 s35, -1
; GFX10-NEXT: buffer_store_dword v40, off, s[0:3], s33 ; 4-byte Folded Spill
+; GFX10-NEXT: buffer_store_dword v41, off, s[0:3], s33 offset:4 ; 4-byte Folded Spill
; GFX10-NEXT: s_waitcnt_depctr 0xffe3
; GFX10-NEXT: s_mov_b32 exec_lo, s35
; GFX10-NEXT: global_load_dwordx4 v[0:3], v[0:1], off
-; GFX10-NEXT: v_writelane_b32 v40, s34, 2
+; GFX10-NEXT: v_writelane_b32 v40, s30, 0
; GFX10-NEXT: s_addk_i32 s32, 0x200
+; GFX10-NEXT: v_writelane_b32 v41, s34, 0
; GFX10-NEXT: s_getpc_b64 s[34:35]
; GFX10-NEXT: s_add_u32 s34, s34, external_void_func_v4i32@rel32@lo+4
; GFX10-NEXT: s_addc_u32 s35, s35, external_void_func_v4i32@rel32@hi+12
-; GFX10-NEXT: v_writelane_b32 v40, s30, 0
; GFX10-NEXT: v_writelane_b32 v40, s31, 1
; GFX10-NEXT: s_swappc_b64 s[30:31], s[34:35]
; GFX10-NEXT: v_readlane_b32 s31, v40, 1
; GFX10-NEXT: v_readlane_b32 s30, v40, 0
-; GFX10-NEXT: v_readlane_b32 s34, v40, 2
+; GFX10-NEXT: v_readlane_b32 s34, v41, 0
; GFX10-NEXT: s_or_saveexec_b32 s35, -1
-; GFX10-NEXT: buffer_load_dword v40, off, s[0:3], s33 ; 4-byte Folded Reload
+; GFX10-NEXT: s_clause 0x1
+; GFX10-NEXT: buffer_load_dword v40, off, s[0:3], s33
+; GFX10-NEXT: buffer_load_dword v41, off, s[0:3], s33 offset:4
; GFX10-NEXT: s_waitcnt_depctr 0xffe3
; GFX10-NEXT: s_mov_b32 exec_lo, s35
; GFX10-NEXT: s_addk_i32 s32, 0xfe00
@@ -4594,23 +5018,27 @@
; GFX11-NEXT: s_mov_b32 s0, s33
; GFX11-NEXT: s_mov_b32 s33, s32
; GFX11-NEXT: s_or_saveexec_b32 s1, -1
-; GFX11-NEXT: scratch_store_b32 off, v40, s33 ; 4-byte Folded Spill
+; GFX11-NEXT: s_clause 0x1
+; GFX11-NEXT: scratch_store_b32 off, v40, s33
+; GFX11-NEXT: scratch_store_b32 off, v41, s33 offset:4
; GFX11-NEXT: s_mov_b32 exec_lo, s1
; GFX11-NEXT: global_load_b128 v[0:3], v[0:1], off
-; GFX11-NEXT: v_writelane_b32 v40, s0, 2
+; GFX11-NEXT: v_writelane_b32 v40, s30, 0
; GFX11-NEXT: s_add_i32 s32, s32, 16
+; GFX11-NEXT: v_writelane_b32 v41, s0, 0
; GFX11-NEXT: s_getpc_b64 s[0:1]
; GFX11-NEXT: s_add_u32 s0, s0, external_void_func_v4i32@rel32@lo+4
; GFX11-NEXT: s_addc_u32 s1, s1, external_void_func_v4i32@rel32@hi+12
-; GFX11-NEXT: v_writelane_b32 v40, s30, 0
; GFX11-NEXT: v_writelane_b32 v40, s31, 1
; GFX11-NEXT: s_swappc_b64 s[30:31], s[0:1]
; GFX11-NEXT: s_delay_alu instid0(VALU_DEP_1)
; GFX11-NEXT: v_readlane_b32 s31, v40, 1
; GFX11-NEXT: v_readlane_b32 s30, v40, 0
-; GFX11-NEXT: v_readlane_b32 s0, v40, 2
+; GFX11-NEXT: v_readlane_b32 s0, v41, 0
; GFX11-NEXT: s_or_saveexec_b32 s1, -1
-; GFX11-NEXT: scratch_load_b32 v40, off, s33 ; 4-byte Folded Reload
+; GFX11-NEXT: s_clause 0x1
+; GFX11-NEXT: scratch_load_b32 v40, off, s33
+; GFX11-NEXT: scratch_load_b32 v41, off, s33 offset:4
; GFX11-NEXT: s_mov_b32 exec_lo, s1
; GFX11-NEXT: s_add_i32 s32, s32, -16
; GFX11-NEXT: s_mov_b32 s33, s0
@@ -4624,22 +5052,25 @@
; GFX10-SCRATCH-NEXT: s_mov_b32 s33, s32
; GFX10-SCRATCH-NEXT: s_or_saveexec_b32 s1, -1
; GFX10-SCRATCH-NEXT: scratch_store_dword off, v40, s33 ; 4-byte Folded Spill
+; GFX10-SCRATCH-NEXT: scratch_store_dword off, v41, s33 offset:4 ; 4-byte Folded Spill
; GFX10-SCRATCH-NEXT: s_waitcnt_depctr 0xffe3
; GFX10-SCRATCH-NEXT: s_mov_b32 exec_lo, s1
; GFX10-SCRATCH-NEXT: global_load_dwordx4 v[0:3], v[0:1], off
-; GFX10-SCRATCH-NEXT: v_writelane_b32 v40, s0, 2
+; GFX10-SCRATCH-NEXT: v_writelane_b32 v40, s30, 0
; GFX10-SCRATCH-NEXT: s_add_i32 s32, s32, 16
+; GFX10-SCRATCH-NEXT: v_writelane_b32 v41, s0, 0
; GFX10-SCRATCH-NEXT: s_getpc_b64 s[0:1]
; GFX10-SCRATCH-NEXT: s_add_u32 s0, s0, external_void_func_v4i32@rel32@lo+4
; GFX10-SCRATCH-NEXT: s_addc_u32 s1, s1, external_void_func_v4i32@rel32@hi+12
-; GFX10-SCRATCH-NEXT: v_writelane_b32 v40, s30, 0
; GFX10-SCRATCH-NEXT: v_writelane_b32 v40, s31, 1
; GFX10-SCRATCH-NEXT: s_swappc_b64 s[30:31], s[0:1]
; GFX10-SCRATCH-NEXT: v_readlane_b32 s31, v40, 1
; GFX10-SCRATCH-NEXT: v_readlane_b32 s30, v40, 0
-; GFX10-SCRATCH-NEXT: v_readlane_b32 s0, v40, 2
+; GFX10-SCRATCH-NEXT: v_readlane_b32 s0, v41, 0
; GFX10-SCRATCH-NEXT: s_or_saveexec_b32 s1, -1
-; GFX10-SCRATCH-NEXT: scratch_load_dword v40, off, s33 ; 4-byte Folded Reload
+; GFX10-SCRATCH-NEXT: s_clause 0x1
+; GFX10-SCRATCH-NEXT: scratch_load_dword v40, off, s33
+; GFX10-SCRATCH-NEXT: scratch_load_dword v41, off, s33 offset:4
; GFX10-SCRATCH-NEXT: s_waitcnt_depctr 0xffe3
; GFX10-SCRATCH-NEXT: s_mov_b32 exec_lo, s1
; GFX10-SCRATCH-NEXT: s_add_i32 s32, s32, -16
@@ -4659,14 +5090,15 @@
; GFX9-NEXT: s_mov_b32 s33, s32
; GFX9-NEXT: s_or_saveexec_b64 s[36:37], -1
; GFX9-NEXT: buffer_store_dword v40, off, s[0:3], s33 ; 4-byte Folded Spill
+; GFX9-NEXT: buffer_store_dword v41, off, s[0:3], s33 offset:4 ; 4-byte Folded Spill
; GFX9-NEXT: s_mov_b64 exec, s[36:37]
-; GFX9-NEXT: v_writelane_b32 v40, s34, 2
; GFX9-NEXT: s_addk_i32 s32, 0x400
; GFX9-NEXT: v_writelane_b32 v40, s30, 0
; GFX9-NEXT: v_mov_b32_e32 v0, 1
; GFX9-NEXT: v_mov_b32_e32 v1, 2
; GFX9-NEXT: v_mov_b32_e32 v2, 3
; GFX9-NEXT: v_mov_b32_e32 v3, 4
+; GFX9-NEXT: v_writelane_b32 v41, s34, 0
; GFX9-NEXT: v_writelane_b32 v40, s31, 1
; GFX9-NEXT: s_getpc_b64 s[34:35]
; GFX9-NEXT: s_add_u32 s34, s34, external_void_func_v4i32@rel32@lo+4
@@ -4674,9 +5106,10 @@
; GFX9-NEXT: s_swappc_b64 s[30:31], s[34:35]
; GFX9-NEXT: v_readlane_b32 s31, v40, 1
; GFX9-NEXT: v_readlane_b32 s30, v40, 0
-; GFX9-NEXT: v_readlane_b32 s34, v40, 2
+; GFX9-NEXT: v_readlane_b32 s34, v41, 0
; GFX9-NEXT: s_or_saveexec_b64 s[36:37], -1
; GFX9-NEXT: buffer_load_dword v40, off, s[0:3], s33 ; 4-byte Folded Reload
+; GFX9-NEXT: buffer_load_dword v41, off, s[0:3], s33 offset:4 ; 4-byte Folded Reload
; GFX9-NEXT: s_mov_b64 exec, s[36:37]
; GFX9-NEXT: s_addk_i32 s32, 0xfc00
; GFX9-NEXT: s_mov_b32 s33, s34
@@ -4690,25 +5123,28 @@
; GFX10-NEXT: s_mov_b32 s33, s32
; GFX10-NEXT: s_or_saveexec_b32 s35, -1
; GFX10-NEXT: buffer_store_dword v40, off, s[0:3], s33 ; 4-byte Folded Spill
+; GFX10-NEXT: buffer_store_dword v41, off, s[0:3], s33 offset:4 ; 4-byte Folded Spill
; GFX10-NEXT: s_waitcnt_depctr 0xffe3
; GFX10-NEXT: s_mov_b32 exec_lo, s35
-; GFX10-NEXT: v_writelane_b32 v40, s34, 2
+; GFX10-NEXT: v_writelane_b32 v40, s30, 0
; GFX10-NEXT: v_mov_b32_e32 v0, 1
; GFX10-NEXT: v_mov_b32_e32 v1, 2
; GFX10-NEXT: v_mov_b32_e32 v2, 3
; GFX10-NEXT: v_mov_b32_e32 v3, 4
-; GFX10-NEXT: v_writelane_b32 v40, s30, 0
; GFX10-NEXT: s_addk_i32 s32, 0x200
+; GFX10-NEXT: v_writelane_b32 v41, s34, 0
+; GFX10-NEXT: v_writelane_b32 v40, s31, 1
; GFX10-NEXT: s_getpc_b64 s[34:35]
; GFX10-NEXT: s_add_u32 s34, s34, external_void_func_v4i32@rel32@lo+4
; GFX10-NEXT: s_addc_u32 s35, s35, external_void_func_v4i32@rel32@hi+12
-; GFX10-NEXT: v_writelane_b32 v40, s31, 1
; GFX10-NEXT: s_swappc_b64 s[30:31], s[34:35]
; GFX10-NEXT: v_readlane_b32 s31, v40, 1
; GFX10-NEXT: v_readlane_b32 s30, v40, 0
-; GFX10-NEXT: v_readlane_b32 s34, v40, 2
+; GFX10-NEXT: v_readlane_b32 s34, v41, 0
; GFX10-NEXT: s_or_saveexec_b32 s35, -1
-; GFX10-NEXT: buffer_load_dword v40, off, s[0:3], s33 ; 4-byte Folded Reload
+; GFX10-NEXT: s_clause 0x1
+; GFX10-NEXT: buffer_load_dword v40, off, s[0:3], s33
+; GFX10-NEXT: buffer_load_dword v41, off, s[0:3], s33 offset:4
; GFX10-NEXT: s_waitcnt_depctr 0xffe3
; GFX10-NEXT: s_mov_b32 exec_lo, s35
; GFX10-NEXT: s_addk_i32 s32, 0xfe00
@@ -4722,24 +5158,28 @@
; GFX11-NEXT: s_mov_b32 s0, s33
; GFX11-NEXT: s_mov_b32 s33, s32
; GFX11-NEXT: s_or_saveexec_b32 s1, -1
-; GFX11-NEXT: scratch_store_b32 off, v40, s33 ; 4-byte Folded Spill
+; GFX11-NEXT: s_clause 0x1
+; GFX11-NEXT: scratch_store_b32 off, v40, s33
+; GFX11-NEXT: scratch_store_b32 off, v41, s33 offset:4
; GFX11-NEXT: s_mov_b32 exec_lo, s1
-; GFX11-NEXT: v_writelane_b32 v40, s0, 2
+; GFX11-NEXT: v_writelane_b32 v40, s30, 0
; GFX11-NEXT: v_dual_mov_b32 v0, 1 :: v_dual_mov_b32 v1, 2
; GFX11-NEXT: v_dual_mov_b32 v2, 3 :: v_dual_mov_b32 v3, 4
-; GFX11-NEXT: v_writelane_b32 v40, s30, 0
; GFX11-NEXT: s_add_i32 s32, s32, 16
+; GFX11-NEXT: v_writelane_b32 v41, s0, 0
+; GFX11-NEXT: v_writelane_b32 v40, s31, 1
; GFX11-NEXT: s_getpc_b64 s[0:1]
; GFX11-NEXT: s_add_u32 s0, s0, external_void_func_v4i32@rel32@lo+4
; GFX11-NEXT: s_addc_u32 s1, s1, external_void_func_v4i32@rel32@hi+12
-; GFX11-NEXT: v_writelane_b32 v40, s31, 1
+; GFX11-NEXT: s_delay_alu instid0(SALU_CYCLE_1)
; GFX11-NEXT: s_swappc_b64 s[30:31], s[0:1]
-; GFX11-NEXT: s_delay_alu instid0(VALU_DEP_1)
; GFX11-NEXT: v_readlane_b32 s31, v40, 1
; GFX11-NEXT: v_readlane_b32 s30, v40, 0
-; GFX11-NEXT: v_readlane_b32 s0, v40, 2
+; GFX11-NEXT: v_readlane_b32 s0, v41, 0
; GFX11-NEXT: s_or_saveexec_b32 s1, -1
-; GFX11-NEXT: scratch_load_b32 v40, off, s33 ; 4-byte Folded Reload
+; GFX11-NEXT: s_clause 0x1
+; GFX11-NEXT: scratch_load_b32 v40, off, s33
+; GFX11-NEXT: scratch_load_b32 v41, off, s33 offset:4
; GFX11-NEXT: s_mov_b32 exec_lo, s1
; GFX11-NEXT: s_add_i32 s32, s32, -16
; GFX11-NEXT: s_mov_b32 s33, s0
@@ -4753,25 +5193,28 @@
; GFX10-SCRATCH-NEXT: s_mov_b32 s33, s32
; GFX10-SCRATCH-NEXT: s_or_saveexec_b32 s1, -1
; GFX10-SCRATCH-NEXT: scratch_store_dword off, v40, s33 ; 4-byte Folded Spill
+; GFX10-SCRATCH-NEXT: scratch_store_dword off, v41, s33 offset:4 ; 4-byte Folded Spill
; GFX10-SCRATCH-NEXT: s_waitcnt_depctr 0xffe3
; GFX10-SCRATCH-NEXT: s_mov_b32 exec_lo, s1
-; GFX10-SCRATCH-NEXT: v_writelane_b32 v40, s0, 2
+; GFX10-SCRATCH-NEXT: v_writelane_b32 v40, s30, 0
; GFX10-SCRATCH-NEXT: v_mov_b32_e32 v0, 1
; GFX10-SCRATCH-NEXT: v_mov_b32_e32 v1, 2
; GFX10-SCRATCH-NEXT: v_mov_b32_e32 v2, 3
; GFX10-SCRATCH-NEXT: v_mov_b32_e32 v3, 4
-; GFX10-SCRATCH-NEXT: v_writelane_b32 v40, s30, 0
; GFX10-SCRATCH-NEXT: s_add_i32 s32, s32, 16
+; GFX10-SCRATCH-NEXT: v_writelane_b32 v41, s0, 0
+; GFX10-SCRATCH-NEXT: v_writelane_b32 v40, s31, 1
; GFX10-SCRATCH-NEXT: s_getpc_b64 s[0:1]
; GFX10-SCRATCH-NEXT: s_add_u32 s0, s0, external_void_func_v4i32@rel32@lo+4
; GFX10-SCRATCH-NEXT: s_addc_u32 s1, s1, external_void_func_v4i32@rel32@hi+12
-; GFX10-SCRATCH-NEXT: v_writelane_b32 v40, s31, 1
; GFX10-SCRATCH-NEXT: s_swappc_b64 s[30:31], s[0:1]
; GFX10-SCRATCH-NEXT: v_readlane_b32 s31, v40, 1
; GFX10-SCRATCH-NEXT: v_readlane_b32 s30, v40, 0
-; GFX10-SCRATCH-NEXT: v_readlane_b32 s0, v40, 2
+; GFX10-SCRATCH-NEXT: v_readlane_b32 s0, v41, 0
; GFX10-SCRATCH-NEXT: s_or_saveexec_b32 s1, -1
-; GFX10-SCRATCH-NEXT: scratch_load_dword v40, off, s33 ; 4-byte Folded Reload
+; GFX10-SCRATCH-NEXT: s_clause 0x1
+; GFX10-SCRATCH-NEXT: scratch_load_dword v40, off, s33
+; GFX10-SCRATCH-NEXT: scratch_load_dword v41, off, s33 offset:4
; GFX10-SCRATCH-NEXT: s_waitcnt_depctr 0xffe3
; GFX10-SCRATCH-NEXT: s_mov_b32 exec_lo, s1
; GFX10-SCRATCH-NEXT: s_add_i32 s32, s32, -16
@@ -4790,8 +5233,8 @@
; GFX9-NEXT: s_mov_b32 s33, s32
; GFX9-NEXT: s_or_saveexec_b64 s[36:37], -1
; GFX9-NEXT: buffer_store_dword v40, off, s[0:3], s33 ; 4-byte Folded Spill
+; GFX9-NEXT: buffer_store_dword v41, off, s[0:3], s33 offset:4 ; 4-byte Folded Spill
; GFX9-NEXT: s_mov_b64 exec, s[36:37]
-; GFX9-NEXT: v_writelane_b32 v40, s34, 2
; GFX9-NEXT: s_addk_i32 s32, 0x400
; GFX9-NEXT: v_writelane_b32 v40, s30, 0
; GFX9-NEXT: v_mov_b32_e32 v0, 1
@@ -4799,6 +5242,7 @@
; GFX9-NEXT: v_mov_b32_e32 v2, 3
; GFX9-NEXT: v_mov_b32_e32 v3, 4
; GFX9-NEXT: v_mov_b32_e32 v4, 5
+; GFX9-NEXT: v_writelane_b32 v41, s34, 0
; GFX9-NEXT: v_writelane_b32 v40, s31, 1
; GFX9-NEXT: s_getpc_b64 s[34:35]
; GFX9-NEXT: s_add_u32 s34, s34, external_void_func_v5i32@rel32@lo+4
@@ -4806,9 +5250,10 @@
; GFX9-NEXT: s_swappc_b64 s[30:31], s[34:35]
; GFX9-NEXT: v_readlane_b32 s31, v40, 1
; GFX9-NEXT: v_readlane_b32 s30, v40, 0
-; GFX9-NEXT: v_readlane_b32 s34, v40, 2
+; GFX9-NEXT: v_readlane_b32 s34, v41, 0
; GFX9-NEXT: s_or_saveexec_b64 s[36:37], -1
; GFX9-NEXT: buffer_load_dword v40, off, s[0:3], s33 ; 4-byte Folded Reload
+; GFX9-NEXT: buffer_load_dword v41, off, s[0:3], s33 offset:4 ; 4-byte Folded Reload
; GFX9-NEXT: s_mov_b64 exec, s[36:37]
; GFX9-NEXT: s_addk_i32 s32, 0xfc00
; GFX9-NEXT: s_mov_b32 s33, s34
@@ -4822,26 +5267,29 @@
; GFX10-NEXT: s_mov_b32 s33, s32
; GFX10-NEXT: s_or_saveexec_b32 s35, -1
; GFX10-NEXT: buffer_store_dword v40, off, s[0:3], s33 ; 4-byte Folded Spill
+; GFX10-NEXT: buffer_store_dword v41, off, s[0:3], s33 offset:4 ; 4-byte Folded Spill
; GFX10-NEXT: s_waitcnt_depctr 0xffe3
; GFX10-NEXT: s_mov_b32 exec_lo, s35
-; GFX10-NEXT: v_writelane_b32 v40, s34, 2
+; GFX10-NEXT: v_writelane_b32 v40, s30, 0
; GFX10-NEXT: v_mov_b32_e32 v0, 1
; GFX10-NEXT: v_mov_b32_e32 v1, 2
; GFX10-NEXT: v_mov_b32_e32 v2, 3
; GFX10-NEXT: v_mov_b32_e32 v3, 4
-; GFX10-NEXT: v_writelane_b32 v40, s30, 0
; GFX10-NEXT: v_mov_b32_e32 v4, 5
; GFX10-NEXT: s_addk_i32 s32, 0x200
+; GFX10-NEXT: v_writelane_b32 v41, s34, 0
+; GFX10-NEXT: v_writelane_b32 v40, s31, 1
; GFX10-NEXT: s_getpc_b64 s[34:35]
; GFX10-NEXT: s_add_u32 s34, s34, external_void_func_v5i32@rel32@lo+4
; GFX10-NEXT: s_addc_u32 s35, s35, external_void_func_v5i32@rel32@hi+12
-; GFX10-NEXT: v_writelane_b32 v40, s31, 1
; GFX10-NEXT: s_swappc_b64 s[30:31], s[34:35]
; GFX10-NEXT: v_readlane_b32 s31, v40, 1
; GFX10-NEXT: v_readlane_b32 s30, v40, 0
-; GFX10-NEXT: v_readlane_b32 s34, v40, 2
+; GFX10-NEXT: v_readlane_b32 s34, v41, 0
; GFX10-NEXT: s_or_saveexec_b32 s35, -1
-; GFX10-NEXT: buffer_load_dword v40, off, s[0:3], s33 ; 4-byte Folded Reload
+; GFX10-NEXT: s_clause 0x1
+; GFX10-NEXT: buffer_load_dword v40, off, s[0:3], s33
+; GFX10-NEXT: buffer_load_dword v41, off, s[0:3], s33 offset:4
; GFX10-NEXT: s_waitcnt_depctr 0xffe3
; GFX10-NEXT: s_mov_b32 exec_lo, s35
; GFX10-NEXT: s_addk_i32 s32, 0xfe00
@@ -4855,25 +5303,29 @@
; GFX11-NEXT: s_mov_b32 s0, s33
; GFX11-NEXT: s_mov_b32 s33, s32
; GFX11-NEXT: s_or_saveexec_b32 s1, -1
-; GFX11-NEXT: scratch_store_b32 off, v40, s33 ; 4-byte Folded Spill
+; GFX11-NEXT: s_clause 0x1
+; GFX11-NEXT: scratch_store_b32 off, v40, s33
+; GFX11-NEXT: scratch_store_b32 off, v41, s33 offset:4
; GFX11-NEXT: s_mov_b32 exec_lo, s1
-; GFX11-NEXT: v_writelane_b32 v40, s0, 2
+; GFX11-NEXT: v_writelane_b32 v40, s30, 0
; GFX11-NEXT: v_dual_mov_b32 v0, 1 :: v_dual_mov_b32 v1, 2
; GFX11-NEXT: v_dual_mov_b32 v2, 3 :: v_dual_mov_b32 v3, 4
-; GFX11-NEXT: v_writelane_b32 v40, s30, 0
; GFX11-NEXT: v_mov_b32_e32 v4, 5
; GFX11-NEXT: s_add_i32 s32, s32, 16
+; GFX11-NEXT: v_writelane_b32 v41, s0, 0
+; GFX11-NEXT: v_writelane_b32 v40, s31, 1
; GFX11-NEXT: s_getpc_b64 s[0:1]
; GFX11-NEXT: s_add_u32 s0, s0, external_void_func_v5i32@rel32@lo+4
; GFX11-NEXT: s_addc_u32 s1, s1, external_void_func_v5i32@rel32@hi+12
-; GFX11-NEXT: v_writelane_b32 v40, s31, 1
+; GFX11-NEXT: s_delay_alu instid0(SALU_CYCLE_1)
; GFX11-NEXT: s_swappc_b64 s[30:31], s[0:1]
-; GFX11-NEXT: s_delay_alu instid0(VALU_DEP_1)
; GFX11-NEXT: v_readlane_b32 s31, v40, 1
; GFX11-NEXT: v_readlane_b32 s30, v40, 0
-; GFX11-NEXT: v_readlane_b32 s0, v40, 2
+; GFX11-NEXT: v_readlane_b32 s0, v41, 0
; GFX11-NEXT: s_or_saveexec_b32 s1, -1
-; GFX11-NEXT: scratch_load_b32 v40, off, s33 ; 4-byte Folded Reload
+; GFX11-NEXT: s_clause 0x1
+; GFX11-NEXT: scratch_load_b32 v40, off, s33
+; GFX11-NEXT: scratch_load_b32 v41, off, s33 offset:4
; GFX11-NEXT: s_mov_b32 exec_lo, s1
; GFX11-NEXT: s_add_i32 s32, s32, -16
; GFX11-NEXT: s_mov_b32 s33, s0
@@ -4887,26 +5339,29 @@
; GFX10-SCRATCH-NEXT: s_mov_b32 s33, s32
; GFX10-SCRATCH-NEXT: s_or_saveexec_b32 s1, -1
; GFX10-SCRATCH-NEXT: scratch_store_dword off, v40, s33 ; 4-byte Folded Spill
+; GFX10-SCRATCH-NEXT: scratch_store_dword off, v41, s33 offset:4 ; 4-byte Folded Spill
; GFX10-SCRATCH-NEXT: s_waitcnt_depctr 0xffe3
; GFX10-SCRATCH-NEXT: s_mov_b32 exec_lo, s1
-; GFX10-SCRATCH-NEXT: v_writelane_b32 v40, s0, 2
+; GFX10-SCRATCH-NEXT: v_writelane_b32 v40, s30, 0
; GFX10-SCRATCH-NEXT: v_mov_b32_e32 v0, 1
; GFX10-SCRATCH-NEXT: v_mov_b32_e32 v1, 2
; GFX10-SCRATCH-NEXT: v_mov_b32_e32 v2, 3
; GFX10-SCRATCH-NEXT: v_mov_b32_e32 v3, 4
-; GFX10-SCRATCH-NEXT: v_writelane_b32 v40, s30, 0
; GFX10-SCRATCH-NEXT: v_mov_b32_e32 v4, 5
; GFX10-SCRATCH-NEXT: s_add_i32 s32, s32, 16
+; GFX10-SCRATCH-NEXT: v_writelane_b32 v41, s0, 0
+; GFX10-SCRATCH-NEXT: v_writelane_b32 v40, s31, 1
; GFX10-SCRATCH-NEXT: s_getpc_b64 s[0:1]
; GFX10-SCRATCH-NEXT: s_add_u32 s0, s0, external_void_func_v5i32@rel32@lo+4
; GFX10-SCRATCH-NEXT: s_addc_u32 s1, s1, external_void_func_v5i32@rel32@hi+12
-; GFX10-SCRATCH-NEXT: v_writelane_b32 v40, s31, 1
; GFX10-SCRATCH-NEXT: s_swappc_b64 s[30:31], s[0:1]
; GFX10-SCRATCH-NEXT: v_readlane_b32 s31, v40, 1
; GFX10-SCRATCH-NEXT: v_readlane_b32 s30, v40, 0
-; GFX10-SCRATCH-NEXT: v_readlane_b32 s0, v40, 2
+; GFX10-SCRATCH-NEXT: v_readlane_b32 s0, v41, 0
; GFX10-SCRATCH-NEXT: s_or_saveexec_b32 s1, -1
-; GFX10-SCRATCH-NEXT: scratch_load_dword v40, off, s33 ; 4-byte Folded Reload
+; GFX10-SCRATCH-NEXT: s_clause 0x1
+; GFX10-SCRATCH-NEXT: scratch_load_dword v40, off, s33
+; GFX10-SCRATCH-NEXT: scratch_load_dword v41, off, s33 offset:4
; GFX10-SCRATCH-NEXT: s_waitcnt_depctr 0xffe3
; GFX10-SCRATCH-NEXT: s_mov_b32 exec_lo, s1
; GFX10-SCRATCH-NEXT: s_add_i32 s32, s32, -16
@@ -4925,8 +5380,9 @@
; GFX9-NEXT: s_mov_b32 s33, s32
; GFX9-NEXT: s_or_saveexec_b64 s[36:37], -1
; GFX9-NEXT: buffer_store_dword v40, off, s[0:3], s33 ; 4-byte Folded Spill
+; GFX9-NEXT: buffer_store_dword v41, off, s[0:3], s33 offset:4 ; 4-byte Folded Spill
; GFX9-NEXT: s_mov_b64 exec, s[36:37]
-; GFX9-NEXT: v_writelane_b32 v40, s34, 2
+; GFX9-NEXT: v_writelane_b32 v41, s34, 0
; GFX9-NEXT: s_load_dwordx2 s[34:35], s[34:35], 0x0
; GFX9-NEXT: v_mov_b32_e32 v8, 0
; GFX9-NEXT: s_addk_i32 s32, 0x400
@@ -4941,9 +5397,10 @@
; GFX9-NEXT: s_swappc_b64 s[30:31], s[34:35]
; GFX9-NEXT: v_readlane_b32 s31, v40, 1
; GFX9-NEXT: v_readlane_b32 s30, v40, 0
-; GFX9-NEXT: v_readlane_b32 s34, v40, 2
+; GFX9-NEXT: v_readlane_b32 s34, v41, 0
; GFX9-NEXT: s_or_saveexec_b64 s[36:37], -1
; GFX9-NEXT: buffer_load_dword v40, off, s[0:3], s33 ; 4-byte Folded Reload
+; GFX9-NEXT: buffer_load_dword v41, off, s[0:3], s33 offset:4 ; 4-byte Folded Reload
; GFX9-NEXT: s_mov_b64 exec, s[36:37]
; GFX9-NEXT: s_addk_i32 s32, 0xfc00
; GFX9-NEXT: s_mov_b32 s33, s34
@@ -4957,27 +5414,30 @@
; GFX10-NEXT: s_mov_b32 s33, s32
; GFX10-NEXT: s_or_saveexec_b32 s35, -1
; GFX10-NEXT: buffer_store_dword v40, off, s[0:3], s33 ; 4-byte Folded Spill
+; GFX10-NEXT: buffer_store_dword v41, off, s[0:3], s33 offset:4 ; 4-byte Folded Spill
; GFX10-NEXT: s_waitcnt_depctr 0xffe3
; GFX10-NEXT: s_mov_b32 exec_lo, s35
-; GFX10-NEXT: v_writelane_b32 v40, s34, 2
+; GFX10-NEXT: v_writelane_b32 v41, s34, 0
; GFX10-NEXT: s_load_dwordx2 s[34:35], s[34:35], 0x0
; GFX10-NEXT: v_mov_b32_e32 v8, 0
+; GFX10-NEXT: v_writelane_b32 v40, s30, 0
; GFX10-NEXT: s_addk_i32 s32, 0x200
; GFX10-NEXT: s_waitcnt lgkmcnt(0)
; GFX10-NEXT: s_clause 0x1
; GFX10-NEXT: global_load_dwordx4 v[0:3], v8, s[34:35]
; GFX10-NEXT: global_load_dwordx4 v[4:7], v8, s[34:35] offset:16
-; GFX10-NEXT: v_writelane_b32 v40, s30, 0
+; GFX10-NEXT: v_writelane_b32 v40, s31, 1
; GFX10-NEXT: s_getpc_b64 s[34:35]
; GFX10-NEXT: s_add_u32 s34, s34, external_void_func_v8i32@rel32@lo+4
; GFX10-NEXT: s_addc_u32 s35, s35, external_void_func_v8i32@rel32@hi+12
-; GFX10-NEXT: v_writelane_b32 v40, s31, 1
; GFX10-NEXT: s_swappc_b64 s[30:31], s[34:35]
; GFX10-NEXT: v_readlane_b32 s31, v40, 1
; GFX10-NEXT: v_readlane_b32 s30, v40, 0
-; GFX10-NEXT: v_readlane_b32 s34, v40, 2
+; GFX10-NEXT: v_readlane_b32 s34, v41, 0
; GFX10-NEXT: s_or_saveexec_b32 s35, -1
-; GFX10-NEXT: buffer_load_dword v40, off, s[0:3], s33 ; 4-byte Folded Reload
+; GFX10-NEXT: s_clause 0x1
+; GFX10-NEXT: buffer_load_dword v40, off, s[0:3], s33
+; GFX10-NEXT: buffer_load_dword v41, off, s[0:3], s33 offset:4
; GFX10-NEXT: s_waitcnt_depctr 0xffe3
; GFX10-NEXT: s_mov_b32 exec_lo, s35
; GFX10-NEXT: s_addk_i32 s32, 0xfe00
@@ -4991,28 +5451,32 @@
; GFX11-NEXT: s_mov_b32 s0, s33
; GFX11-NEXT: s_mov_b32 s33, s32
; GFX11-NEXT: s_or_saveexec_b32 s1, -1
-; GFX11-NEXT: scratch_store_b32 off, v40, s33 ; 4-byte Folded Spill
+; GFX11-NEXT: s_clause 0x1
+; GFX11-NEXT: scratch_store_b32 off, v40, s33
+; GFX11-NEXT: scratch_store_b32 off, v41, s33 offset:4
; GFX11-NEXT: s_mov_b32 exec_lo, s1
-; GFX11-NEXT: v_writelane_b32 v40, s0, 2
+; GFX11-NEXT: v_writelane_b32 v41, s0, 0
; GFX11-NEXT: s_load_b64 s[0:1], s[0:1], 0x0
; GFX11-NEXT: v_mov_b32_e32 v4, 0
+; GFX11-NEXT: v_writelane_b32 v40, s30, 0
; GFX11-NEXT: s_add_i32 s32, s32, 16
; GFX11-NEXT: s_waitcnt lgkmcnt(0)
; GFX11-NEXT: s_clause 0x1
; GFX11-NEXT: global_load_b128 v[0:3], v4, s[0:1]
; GFX11-NEXT: global_load_b128 v[4:7], v4, s[0:1] offset:16
-; GFX11-NEXT: v_writelane_b32 v40, s30, 0
+; GFX11-NEXT: v_writelane_b32 v40, s31, 1
; GFX11-NEXT: s_getpc_b64 s[0:1]
; GFX11-NEXT: s_add_u32 s0, s0, external_void_func_v8i32@rel32@lo+4
; GFX11-NEXT: s_addc_u32 s1, s1, external_void_func_v8i32@rel32@hi+12
-; GFX11-NEXT: v_writelane_b32 v40, s31, 1
+; GFX11-NEXT: s_delay_alu instid0(SALU_CYCLE_1)
; GFX11-NEXT: s_swappc_b64 s[30:31], s[0:1]
-; GFX11-NEXT: s_delay_alu instid0(VALU_DEP_1)
; GFX11-NEXT: v_readlane_b32 s31, v40, 1
; GFX11-NEXT: v_readlane_b32 s30, v40, 0
-; GFX11-NEXT: v_readlane_b32 s0, v40, 2
+; GFX11-NEXT: v_readlane_b32 s0, v41, 0
; GFX11-NEXT: s_or_saveexec_b32 s1, -1
-; GFX11-NEXT: scratch_load_b32 v40, off, s33 ; 4-byte Folded Reload
+; GFX11-NEXT: s_clause 0x1
+; GFX11-NEXT: scratch_load_b32 v40, off, s33
+; GFX11-NEXT: scratch_load_b32 v41, off, s33 offset:4
; GFX11-NEXT: s_mov_b32 exec_lo, s1
; GFX11-NEXT: s_add_i32 s32, s32, -16
; GFX11-NEXT: s_mov_b32 s33, s0
@@ -5026,27 +5490,30 @@
; GFX10-SCRATCH-NEXT: s_mov_b32 s33, s32
; GFX10-SCRATCH-NEXT: s_or_saveexec_b32 s1, -1
; GFX10-SCRATCH-NEXT: scratch_store_dword off, v40, s33 ; 4-byte Folded Spill
+; GFX10-SCRATCH-NEXT: scratch_store_dword off, v41, s33 offset:4 ; 4-byte Folded Spill
; GFX10-SCRATCH-NEXT: s_waitcnt_depctr 0xffe3
; GFX10-SCRATCH-NEXT: s_mov_b32 exec_lo, s1
-; GFX10-SCRATCH-NEXT: v_writelane_b32 v40, s0, 2
+; GFX10-SCRATCH-NEXT: v_writelane_b32 v41, s0, 0
; GFX10-SCRATCH-NEXT: s_load_dwordx2 s[0:1], s[0:1], 0x0
; GFX10-SCRATCH-NEXT: v_mov_b32_e32 v8, 0
+; GFX10-SCRATCH-NEXT: v_writelane_b32 v40, s30, 0
; GFX10-SCRATCH-NEXT: s_add_i32 s32, s32, 16
; GFX10-SCRATCH-NEXT: s_waitcnt lgkmcnt(0)
; GFX10-SCRATCH-NEXT: s_clause 0x1
; GFX10-SCRATCH-NEXT: global_load_dwordx4 v[0:3], v8, s[0:1]
; GFX10-SCRATCH-NEXT: global_load_dwordx4 v[4:7], v8, s[0:1] offset:16
-; GFX10-SCRATCH-NEXT: v_writelane_b32 v40, s30, 0
+; GFX10-SCRATCH-NEXT: v_writelane_b32 v40, s31, 1
; GFX10-SCRATCH-NEXT: s_getpc_b64 s[0:1]
; GFX10-SCRATCH-NEXT: s_add_u32 s0, s0, external_void_func_v8i32@rel32@lo+4
; GFX10-SCRATCH-NEXT: s_addc_u32 s1, s1, external_void_func_v8i32@rel32@hi+12
-; GFX10-SCRATCH-NEXT: v_writelane_b32 v40, s31, 1
; GFX10-SCRATCH-NEXT: s_swappc_b64 s[30:31], s[0:1]
; GFX10-SCRATCH-NEXT: v_readlane_b32 s31, v40, 1
; GFX10-SCRATCH-NEXT: v_readlane_b32 s30, v40, 0
-; GFX10-SCRATCH-NEXT: v_readlane_b32 s0, v40, 2
+; GFX10-SCRATCH-NEXT: v_readlane_b32 s0, v41, 0
; GFX10-SCRATCH-NEXT: s_or_saveexec_b32 s1, -1
-; GFX10-SCRATCH-NEXT: scratch_load_dword v40, off, s33 ; 4-byte Folded Reload
+; GFX10-SCRATCH-NEXT: s_clause 0x1
+; GFX10-SCRATCH-NEXT: scratch_load_dword v40, off, s33
+; GFX10-SCRATCH-NEXT: scratch_load_dword v41, off, s33 offset:4
; GFX10-SCRATCH-NEXT: s_waitcnt_depctr 0xffe3
; GFX10-SCRATCH-NEXT: s_mov_b32 exec_lo, s1
; GFX10-SCRATCH-NEXT: s_add_i32 s32, s32, -16
@@ -5067,8 +5534,8 @@
; GFX9-NEXT: s_mov_b32 s33, s32
; GFX9-NEXT: s_or_saveexec_b64 s[36:37], -1
; GFX9-NEXT: buffer_store_dword v40, off, s[0:3], s33 ; 4-byte Folded Spill
+; GFX9-NEXT: buffer_store_dword v41, off, s[0:3], s33 offset:4 ; 4-byte Folded Spill
; GFX9-NEXT: s_mov_b64 exec, s[36:37]
-; GFX9-NEXT: v_writelane_b32 v40, s34, 2
; GFX9-NEXT: s_addk_i32 s32, 0x400
; GFX9-NEXT: v_writelane_b32 v40, s30, 0
; GFX9-NEXT: v_mov_b32_e32 v0, 1
@@ -5079,6 +5546,7 @@
; GFX9-NEXT: v_mov_b32_e32 v5, 6
; GFX9-NEXT: v_mov_b32_e32 v6, 7
; GFX9-NEXT: v_mov_b32_e32 v7, 8
+; GFX9-NEXT: v_writelane_b32 v41, s34, 0
; GFX9-NEXT: v_writelane_b32 v40, s31, 1
; GFX9-NEXT: s_getpc_b64 s[34:35]
; GFX9-NEXT: s_add_u32 s34, s34, external_void_func_v8i32@rel32@lo+4
@@ -5086,9 +5554,10 @@
; GFX9-NEXT: s_swappc_b64 s[30:31], s[34:35]
; GFX9-NEXT: v_readlane_b32 s31, v40, 1
; GFX9-NEXT: v_readlane_b32 s30, v40, 0
-; GFX9-NEXT: v_readlane_b32 s34, v40, 2
+; GFX9-NEXT: v_readlane_b32 s34, v41, 0
; GFX9-NEXT: s_or_saveexec_b64 s[36:37], -1
; GFX9-NEXT: buffer_load_dword v40, off, s[0:3], s33 ; 4-byte Folded Reload
+; GFX9-NEXT: buffer_load_dword v41, off, s[0:3], s33 offset:4 ; 4-byte Folded Reload
; GFX9-NEXT: s_mov_b64 exec, s[36:37]
; GFX9-NEXT: s_addk_i32 s32, 0xfc00
; GFX9-NEXT: s_mov_b32 s33, s34
@@ -5102,19 +5571,20 @@
; GFX10-NEXT: s_mov_b32 s33, s32
; GFX10-NEXT: s_or_saveexec_b32 s35, -1
; GFX10-NEXT: buffer_store_dword v40, off, s[0:3], s33 ; 4-byte Folded Spill
+; GFX10-NEXT: buffer_store_dword v41, off, s[0:3], s33 offset:4 ; 4-byte Folded Spill
; GFX10-NEXT: s_waitcnt_depctr 0xffe3
; GFX10-NEXT: s_mov_b32 exec_lo, s35
-; GFX10-NEXT: v_writelane_b32 v40, s34, 2
+; GFX10-NEXT: v_writelane_b32 v40, s30, 0
; GFX10-NEXT: v_mov_b32_e32 v0, 1
; GFX10-NEXT: v_mov_b32_e32 v1, 2
; GFX10-NEXT: v_mov_b32_e32 v2, 3
; GFX10-NEXT: v_mov_b32_e32 v3, 4
-; GFX10-NEXT: v_writelane_b32 v40, s30, 0
; GFX10-NEXT: v_mov_b32_e32 v4, 5
; GFX10-NEXT: v_mov_b32_e32 v5, 6
; GFX10-NEXT: v_mov_b32_e32 v6, 7
; GFX10-NEXT: v_mov_b32_e32 v7, 8
; GFX10-NEXT: s_addk_i32 s32, 0x200
+; GFX10-NEXT: v_writelane_b32 v41, s34, 0
; GFX10-NEXT: v_writelane_b32 v40, s31, 1
; GFX10-NEXT: s_getpc_b64 s[34:35]
; GFX10-NEXT: s_add_u32 s34, s34, external_void_func_v8i32@rel32@lo+4
@@ -5122,9 +5592,11 @@
; GFX10-NEXT: s_swappc_b64 s[30:31], s[34:35]
; GFX10-NEXT: v_readlane_b32 s31, v40, 1
; GFX10-NEXT: v_readlane_b32 s30, v40, 0
-; GFX10-NEXT: v_readlane_b32 s34, v40, 2
+; GFX10-NEXT: v_readlane_b32 s34, v41, 0
; GFX10-NEXT: s_or_saveexec_b32 s35, -1
-; GFX10-NEXT: buffer_load_dword v40, off, s[0:3], s33 ; 4-byte Folded Reload
+; GFX10-NEXT: s_clause 0x1
+; GFX10-NEXT: buffer_load_dword v40, off, s[0:3], s33
+; GFX10-NEXT: buffer_load_dword v41, off, s[0:3], s33 offset:4
; GFX10-NEXT: s_waitcnt_depctr 0xffe3
; GFX10-NEXT: s_mov_b32 exec_lo, s35
; GFX10-NEXT: s_addk_i32 s32, 0xfe00
@@ -5138,15 +5610,17 @@
; GFX11-NEXT: s_mov_b32 s0, s33
; GFX11-NEXT: s_mov_b32 s33, s32
; GFX11-NEXT: s_or_saveexec_b32 s1, -1
-; GFX11-NEXT: scratch_store_b32 off, v40, s33 ; 4-byte Folded Spill
+; GFX11-NEXT: s_clause 0x1
+; GFX11-NEXT: scratch_store_b32 off, v40, s33
+; GFX11-NEXT: scratch_store_b32 off, v41, s33 offset:4
; GFX11-NEXT: s_mov_b32 exec_lo, s1
-; GFX11-NEXT: v_writelane_b32 v40, s0, 2
+; GFX11-NEXT: v_writelane_b32 v40, s30, 0
; GFX11-NEXT: v_dual_mov_b32 v0, 1 :: v_dual_mov_b32 v1, 2
; GFX11-NEXT: v_dual_mov_b32 v2, 3 :: v_dual_mov_b32 v3, 4
-; GFX11-NEXT: v_writelane_b32 v40, s30, 0
; GFX11-NEXT: v_dual_mov_b32 v4, 5 :: v_dual_mov_b32 v5, 6
; GFX11-NEXT: v_dual_mov_b32 v6, 7 :: v_dual_mov_b32 v7, 8
; GFX11-NEXT: s_add_i32 s32, s32, 16
+; GFX11-NEXT: v_writelane_b32 v41, s0, 0
; GFX11-NEXT: v_writelane_b32 v40, s31, 1
; GFX11-NEXT: s_getpc_b64 s[0:1]
; GFX11-NEXT: s_add_u32 s0, s0, external_void_func_v8i32@rel32@lo+4
@@ -5155,9 +5629,11 @@
; GFX11-NEXT: s_swappc_b64 s[30:31], s[0:1]
; GFX11-NEXT: v_readlane_b32 s31, v40, 1
; GFX11-NEXT: v_readlane_b32 s30, v40, 0
-; GFX11-NEXT: v_readlane_b32 s0, v40, 2
+; GFX11-NEXT: v_readlane_b32 s0, v41, 0
; GFX11-NEXT: s_or_saveexec_b32 s1, -1
-; GFX11-NEXT: scratch_load_b32 v40, off, s33 ; 4-byte Folded Reload
+; GFX11-NEXT: s_clause 0x1
+; GFX11-NEXT: scratch_load_b32 v40, off, s33
+; GFX11-NEXT: scratch_load_b32 v41, off, s33 offset:4
; GFX11-NEXT: s_mov_b32 exec_lo, s1
; GFX11-NEXT: s_add_i32 s32, s32, -16
; GFX11-NEXT: s_mov_b32 s33, s0
@@ -5171,19 +5647,20 @@
; GFX10-SCRATCH-NEXT: s_mov_b32 s33, s32
; GFX10-SCRATCH-NEXT: s_or_saveexec_b32 s1, -1
; GFX10-SCRATCH-NEXT: scratch_store_dword off, v40, s33 ; 4-byte Folded Spill
+; GFX10-SCRATCH-NEXT: scratch_store_dword off, v41, s33 offset:4 ; 4-byte Folded Spill
; GFX10-SCRATCH-NEXT: s_waitcnt_depctr 0xffe3
; GFX10-SCRATCH-NEXT: s_mov_b32 exec_lo, s1
-; GFX10-SCRATCH-NEXT: v_writelane_b32 v40, s0, 2
+; GFX10-SCRATCH-NEXT: v_writelane_b32 v40, s30, 0
; GFX10-SCRATCH-NEXT: v_mov_b32_e32 v0, 1
; GFX10-SCRATCH-NEXT: v_mov_b32_e32 v1, 2
; GFX10-SCRATCH-NEXT: v_mov_b32_e32 v2, 3
; GFX10-SCRATCH-NEXT: v_mov_b32_e32 v3, 4
-; GFX10-SCRATCH-NEXT: v_writelane_b32 v40, s30, 0
; GFX10-SCRATCH-NEXT: v_mov_b32_e32 v4, 5
; GFX10-SCRATCH-NEXT: v_mov_b32_e32 v5, 6
; GFX10-SCRATCH-NEXT: v_mov_b32_e32 v6, 7
; GFX10-SCRATCH-NEXT: v_mov_b32_e32 v7, 8
; GFX10-SCRATCH-NEXT: s_add_i32 s32, s32, 16
+; GFX10-SCRATCH-NEXT: v_writelane_b32 v41, s0, 0
; GFX10-SCRATCH-NEXT: v_writelane_b32 v40, s31, 1
; GFX10-SCRATCH-NEXT: s_getpc_b64 s[0:1]
; GFX10-SCRATCH-NEXT: s_add_u32 s0, s0, external_void_func_v8i32@rel32@lo+4
@@ -5191,9 +5668,11 @@
; GFX10-SCRATCH-NEXT: s_swappc_b64 s[30:31], s[0:1]
; GFX10-SCRATCH-NEXT: v_readlane_b32 s31, v40, 1
; GFX10-SCRATCH-NEXT: v_readlane_b32 s30, v40, 0
-; GFX10-SCRATCH-NEXT: v_readlane_b32 s0, v40, 2
+; GFX10-SCRATCH-NEXT: v_readlane_b32 s0, v41, 0
; GFX10-SCRATCH-NEXT: s_or_saveexec_b32 s1, -1
-; GFX10-SCRATCH-NEXT: scratch_load_dword v40, off, s33 ; 4-byte Folded Reload
+; GFX10-SCRATCH-NEXT: s_clause 0x1
+; GFX10-SCRATCH-NEXT: scratch_load_dword v40, off, s33
+; GFX10-SCRATCH-NEXT: scratch_load_dword v41, off, s33 offset:4
; GFX10-SCRATCH-NEXT: s_waitcnt_depctr 0xffe3
; GFX10-SCRATCH-NEXT: s_mov_b32 exec_lo, s1
; GFX10-SCRATCH-NEXT: s_add_i32 s32, s32, -16
@@ -5212,8 +5691,9 @@
; GFX9-NEXT: s_mov_b32 s33, s32
; GFX9-NEXT: s_or_saveexec_b64 s[36:37], -1
; GFX9-NEXT: buffer_store_dword v40, off, s[0:3], s33 ; 4-byte Folded Spill
+; GFX9-NEXT: buffer_store_dword v41, off, s[0:3], s33 offset:4 ; 4-byte Folded Spill
; GFX9-NEXT: s_mov_b64 exec, s[36:37]
-; GFX9-NEXT: v_writelane_b32 v40, s34, 2
+; GFX9-NEXT: v_writelane_b32 v41, s34, 0
; GFX9-NEXT: s_load_dwordx2 s[34:35], s[34:35], 0x0
; GFX9-NEXT: v_mov_b32_e32 v16, 0
; GFX9-NEXT: s_addk_i32 s32, 0x400
@@ -5230,9 +5710,10 @@
; GFX9-NEXT: s_swappc_b64 s[30:31], s[34:35]
; GFX9-NEXT: v_readlane_b32 s31, v40, 1
; GFX9-NEXT: v_readlane_b32 s30, v40, 0
-; GFX9-NEXT: v_readlane_b32 s34, v40, 2
+; GFX9-NEXT: v_readlane_b32 s34, v41, 0
; GFX9-NEXT: s_or_saveexec_b64 s[36:37], -1
; GFX9-NEXT: buffer_load_dword v40, off, s[0:3], s33 ; 4-byte Folded Reload
+; GFX9-NEXT: buffer_load_dword v41, off, s[0:3], s33 offset:4 ; 4-byte Folded Reload
; GFX9-NEXT: s_mov_b64 exec, s[36:37]
; GFX9-NEXT: s_addk_i32 s32, 0xfc00
; GFX9-NEXT: s_mov_b32 s33, s34
@@ -5246,11 +5727,13 @@
; GFX10-NEXT: s_mov_b32 s33, s32
; GFX10-NEXT: s_or_saveexec_b32 s35, -1
; GFX10-NEXT: buffer_store_dword v40, off, s[0:3], s33 ; 4-byte Folded Spill
+; GFX10-NEXT: buffer_store_dword v41, off, s[0:3], s33 offset:4 ; 4-byte Folded Spill
; GFX10-NEXT: s_waitcnt_depctr 0xffe3
; GFX10-NEXT: s_mov_b32 exec_lo, s35
-; GFX10-NEXT: v_writelane_b32 v40, s34, 2
+; GFX10-NEXT: v_writelane_b32 v41, s34, 0
; GFX10-NEXT: s_load_dwordx2 s[34:35], s[34:35], 0x0
; GFX10-NEXT: v_mov_b32_e32 v16, 0
+; GFX10-NEXT: v_writelane_b32 v40, s30, 0
; GFX10-NEXT: s_addk_i32 s32, 0x200
; GFX10-NEXT: s_waitcnt lgkmcnt(0)
; GFX10-NEXT: s_clause 0x3
@@ -5258,17 +5741,18 @@
; GFX10-NEXT: global_load_dwordx4 v[4:7], v16, s[34:35] offset:16
; GFX10-NEXT: global_load_dwordx4 v[8:11], v16, s[34:35] offset:32
; GFX10-NEXT: global_load_dwordx4 v[12:15], v16, s[34:35] offset:48
-; GFX10-NEXT: v_writelane_b32 v40, s30, 0
+; GFX10-NEXT: v_writelane_b32 v40, s31, 1
; GFX10-NEXT: s_getpc_b64 s[34:35]
; GFX10-NEXT: s_add_u32 s34, s34, external_void_func_v16i32@rel32@lo+4
; GFX10-NEXT: s_addc_u32 s35, s35, external_void_func_v16i32@rel32@hi+12
-; GFX10-NEXT: v_writelane_b32 v40, s31, 1
; GFX10-NEXT: s_swappc_b64 s[30:31], s[34:35]
; GFX10-NEXT: v_readlane_b32 s31, v40, 1
; GFX10-NEXT: v_readlane_b32 s30, v40, 0
-; GFX10-NEXT: v_readlane_b32 s34, v40, 2
+; GFX10-NEXT: v_readlane_b32 s34, v41, 0
; GFX10-NEXT: s_or_saveexec_b32 s35, -1
-; GFX10-NEXT: buffer_load_dword v40, off, s[0:3], s33 ; 4-byte Folded Reload
+; GFX10-NEXT: s_clause 0x1
+; GFX10-NEXT: buffer_load_dword v40, off, s[0:3], s33
+; GFX10-NEXT: buffer_load_dword v41, off, s[0:3], s33 offset:4
; GFX10-NEXT: s_waitcnt_depctr 0xffe3
; GFX10-NEXT: s_mov_b32 exec_lo, s35
; GFX10-NEXT: s_addk_i32 s32, 0xfe00
@@ -5282,11 +5766,14 @@
; GFX11-NEXT: s_mov_b32 s0, s33
; GFX11-NEXT: s_mov_b32 s33, s32
; GFX11-NEXT: s_or_saveexec_b32 s1, -1
-; GFX11-NEXT: scratch_store_b32 off, v40, s33 ; 4-byte Folded Spill
+; GFX11-NEXT: s_clause 0x1
+; GFX11-NEXT: scratch_store_b32 off, v40, s33
+; GFX11-NEXT: scratch_store_b32 off, v41, s33 offset:4
; GFX11-NEXT: s_mov_b32 exec_lo, s1
-; GFX11-NEXT: v_writelane_b32 v40, s0, 2
+; GFX11-NEXT: v_writelane_b32 v41, s0, 0
; GFX11-NEXT: s_load_b64 s[0:1], s[0:1], 0x0
; GFX11-NEXT: v_mov_b32_e32 v12, 0
+; GFX11-NEXT: v_writelane_b32 v40, s30, 0
; GFX11-NEXT: s_add_i32 s32, s32, 16
; GFX11-NEXT: s_waitcnt lgkmcnt(0)
; GFX11-NEXT: s_clause 0x3
@@ -5294,18 +5781,19 @@
; GFX11-NEXT: global_load_b128 v[4:7], v12, s[0:1] offset:16
; GFX11-NEXT: global_load_b128 v[8:11], v12, s[0:1] offset:32
; GFX11-NEXT: global_load_b128 v[12:15], v12, s[0:1] offset:48
-; GFX11-NEXT: v_writelane_b32 v40, s30, 0
+; GFX11-NEXT: v_writelane_b32 v40, s31, 1
; GFX11-NEXT: s_getpc_b64 s[0:1]
; GFX11-NEXT: s_add_u32 s0, s0, external_void_func_v16i32@rel32@lo+4
; GFX11-NEXT: s_addc_u32 s1, s1, external_void_func_v16i32@rel32@hi+12
-; GFX11-NEXT: v_writelane_b32 v40, s31, 1
+; GFX11-NEXT: s_delay_alu instid0(SALU_CYCLE_1)
; GFX11-NEXT: s_swappc_b64 s[30:31], s[0:1]
-; GFX11-NEXT: s_delay_alu instid0(VALU_DEP_1)
; GFX11-NEXT: v_readlane_b32 s31, v40, 1
; GFX11-NEXT: v_readlane_b32 s30, v40, 0
-; GFX11-NEXT: v_readlane_b32 s0, v40, 2
+; GFX11-NEXT: v_readlane_b32 s0, v41, 0
; GFX11-NEXT: s_or_saveexec_b32 s1, -1
-; GFX11-NEXT: scratch_load_b32 v40, off, s33 ; 4-byte Folded Reload
+; GFX11-NEXT: s_clause 0x1
+; GFX11-NEXT: scratch_load_b32 v40, off, s33
+; GFX11-NEXT: scratch_load_b32 v41, off, s33 offset:4
; GFX11-NEXT: s_mov_b32 exec_lo, s1
; GFX11-NEXT: s_add_i32 s32, s32, -16
; GFX11-NEXT: s_mov_b32 s33, s0
@@ -5319,11 +5807,13 @@
; GFX10-SCRATCH-NEXT: s_mov_b32 s33, s32
; GFX10-SCRATCH-NEXT: s_or_saveexec_b32 s1, -1
; GFX10-SCRATCH-NEXT: scratch_store_dword off, v40, s33 ; 4-byte Folded Spill
+; GFX10-SCRATCH-NEXT: scratch_store_dword off, v41, s33 offset:4 ; 4-byte Folded Spill
; GFX10-SCRATCH-NEXT: s_waitcnt_depctr 0xffe3
; GFX10-SCRATCH-NEXT: s_mov_b32 exec_lo, s1
-; GFX10-SCRATCH-NEXT: v_writelane_b32 v40, s0, 2
+; GFX10-SCRATCH-NEXT: v_writelane_b32 v41, s0, 0
; GFX10-SCRATCH-NEXT: s_load_dwordx2 s[0:1], s[0:1], 0x0
; GFX10-SCRATCH-NEXT: v_mov_b32_e32 v16, 0
+; GFX10-SCRATCH-NEXT: v_writelane_b32 v40, s30, 0
; GFX10-SCRATCH-NEXT: s_add_i32 s32, s32, 16
; GFX10-SCRATCH-NEXT: s_waitcnt lgkmcnt(0)
; GFX10-SCRATCH-NEXT: s_clause 0x3
@@ -5331,17 +5821,18 @@
; GFX10-SCRATCH-NEXT: global_load_dwordx4 v[4:7], v16, s[0:1] offset:16
; GFX10-SCRATCH-NEXT: global_load_dwordx4 v[8:11], v16, s[0:1] offset:32
; GFX10-SCRATCH-NEXT: global_load_dwordx4 v[12:15], v16, s[0:1] offset:48
-; GFX10-SCRATCH-NEXT: v_writelane_b32 v40, s30, 0
+; GFX10-SCRATCH-NEXT: v_writelane_b32 v40, s31, 1
; GFX10-SCRATCH-NEXT: s_getpc_b64 s[0:1]
; GFX10-SCRATCH-NEXT: s_add_u32 s0, s0, external_void_func_v16i32@rel32@lo+4
; GFX10-SCRATCH-NEXT: s_addc_u32 s1, s1, external_void_func_v16i32@rel32@hi+12
-; GFX10-SCRATCH-NEXT: v_writelane_b32 v40, s31, 1
; GFX10-SCRATCH-NEXT: s_swappc_b64 s[30:31], s[0:1]
; GFX10-SCRATCH-NEXT: v_readlane_b32 s31, v40, 1
; GFX10-SCRATCH-NEXT: v_readlane_b32 s30, v40, 0
-; GFX10-SCRATCH-NEXT: v_readlane_b32 s0, v40, 2
+; GFX10-SCRATCH-NEXT: v_readlane_b32 s0, v41, 0
; GFX10-SCRATCH-NEXT: s_or_saveexec_b32 s1, -1
-; GFX10-SCRATCH-NEXT: scratch_load_dword v40, off, s33 ; 4-byte Folded Reload
+; GFX10-SCRATCH-NEXT: s_clause 0x1
+; GFX10-SCRATCH-NEXT: scratch_load_dword v40, off, s33
+; GFX10-SCRATCH-NEXT: scratch_load_dword v41, off, s33 offset:4
; GFX10-SCRATCH-NEXT: s_waitcnt_depctr 0xffe3
; GFX10-SCRATCH-NEXT: s_mov_b32 exec_lo, s1
; GFX10-SCRATCH-NEXT: s_add_i32 s32, s32, -16
@@ -5362,8 +5853,9 @@
; GFX9-NEXT: s_mov_b32 s33, s32
; GFX9-NEXT: s_or_saveexec_b64 s[36:37], -1
; GFX9-NEXT: buffer_store_dword v40, off, s[0:3], s33 ; 4-byte Folded Spill
+; GFX9-NEXT: buffer_store_dword v41, off, s[0:3], s33 offset:4 ; 4-byte Folded Spill
; GFX9-NEXT: s_mov_b64 exec, s[36:37]
-; GFX9-NEXT: v_writelane_b32 v40, s34, 2
+; GFX9-NEXT: v_writelane_b32 v41, s34, 0
; GFX9-NEXT: s_load_dwordx2 s[34:35], s[34:35], 0x0
; GFX9-NEXT: v_mov_b32_e32 v28, 0
; GFX9-NEXT: s_addk_i32 s32, 0x400
@@ -5385,9 +5877,10 @@
; GFX9-NEXT: s_swappc_b64 s[30:31], s[34:35]
; GFX9-NEXT: v_readlane_b32 s31, v40, 1
; GFX9-NEXT: v_readlane_b32 s30, v40, 0
-; GFX9-NEXT: v_readlane_b32 s34, v40, 2
+; GFX9-NEXT: v_readlane_b32 s34, v41, 0
; GFX9-NEXT: s_or_saveexec_b64 s[36:37], -1
; GFX9-NEXT: buffer_load_dword v40, off, s[0:3], s33 ; 4-byte Folded Reload
+; GFX9-NEXT: buffer_load_dword v41, off, s[0:3], s33 offset:4 ; 4-byte Folded Reload
; GFX9-NEXT: s_mov_b64 exec, s[36:37]
; GFX9-NEXT: s_addk_i32 s32, 0xfc00
; GFX9-NEXT: s_mov_b32 s33, s34
@@ -5401,11 +5894,13 @@
; GFX10-NEXT: s_mov_b32 s33, s32
; GFX10-NEXT: s_or_saveexec_b32 s35, -1
; GFX10-NEXT: buffer_store_dword v40, off, s[0:3], s33 ; 4-byte Folded Spill
+; GFX10-NEXT: buffer_store_dword v41, off, s[0:3], s33 offset:4 ; 4-byte Folded Spill
; GFX10-NEXT: s_waitcnt_depctr 0xffe3
; GFX10-NEXT: s_mov_b32 exec_lo, s35
-; GFX10-NEXT: v_writelane_b32 v40, s34, 2
+; GFX10-NEXT: v_writelane_b32 v41, s34, 0
; GFX10-NEXT: s_load_dwordx2 s[34:35], s[34:35], 0x0
; GFX10-NEXT: v_mov_b32_e32 v32, 0
+; GFX10-NEXT: v_writelane_b32 v40, s30, 0
; GFX10-NEXT: s_addk_i32 s32, 0x200
; GFX10-NEXT: s_waitcnt lgkmcnt(0)
; GFX10-NEXT: s_clause 0x7
@@ -5417,17 +5912,18 @@
; GFX10-NEXT: global_load_dwordx4 v[20:23], v32, s[34:35] offset:80
; GFX10-NEXT: global_load_dwordx4 v[24:27], v32, s[34:35] offset:96
; GFX10-NEXT: global_load_dwordx4 v[28:31], v32, s[34:35] offset:112
-; GFX10-NEXT: v_writelane_b32 v40, s30, 0
+; GFX10-NEXT: v_writelane_b32 v40, s31, 1
; GFX10-NEXT: s_getpc_b64 s[34:35]
; GFX10-NEXT: s_add_u32 s34, s34, external_void_func_v32i32@rel32@lo+4
; GFX10-NEXT: s_addc_u32 s35, s35, external_void_func_v32i32@rel32@hi+12
-; GFX10-NEXT: v_writelane_b32 v40, s31, 1
; GFX10-NEXT: s_swappc_b64 s[30:31], s[34:35]
; GFX10-NEXT: v_readlane_b32 s31, v40, 1
; GFX10-NEXT: v_readlane_b32 s30, v40, 0
-; GFX10-NEXT: v_readlane_b32 s34, v40, 2
+; GFX10-NEXT: v_readlane_b32 s34, v41, 0
; GFX10-NEXT: s_or_saveexec_b32 s35, -1
-; GFX10-NEXT: buffer_load_dword v40, off, s[0:3], s33 ; 4-byte Folded Reload
+; GFX10-NEXT: s_clause 0x1
+; GFX10-NEXT: buffer_load_dword v40, off, s[0:3], s33
+; GFX10-NEXT: buffer_load_dword v41, off, s[0:3], s33 offset:4
; GFX10-NEXT: s_waitcnt_depctr 0xffe3
; GFX10-NEXT: s_mov_b32 exec_lo, s35
; GFX10-NEXT: s_addk_i32 s32, 0xfe00
@@ -5441,11 +5937,14 @@
; GFX11-NEXT: s_mov_b32 s0, s33
; GFX11-NEXT: s_mov_b32 s33, s32
; GFX11-NEXT: s_or_saveexec_b32 s1, -1
-; GFX11-NEXT: scratch_store_b32 off, v40, s33 ; 4-byte Folded Spill
+; GFX11-NEXT: s_clause 0x1
+; GFX11-NEXT: scratch_store_b32 off, v40, s33
+; GFX11-NEXT: scratch_store_b32 off, v41, s33 offset:4
; GFX11-NEXT: s_mov_b32 exec_lo, s1
-; GFX11-NEXT: v_writelane_b32 v40, s0, 2
+; GFX11-NEXT: v_writelane_b32 v41, s0, 0
; GFX11-NEXT: s_load_b64 s[0:1], s[0:1], 0x0
; GFX11-NEXT: v_mov_b32_e32 v28, 0
+; GFX11-NEXT: v_writelane_b32 v40, s30, 0
; GFX11-NEXT: s_add_i32 s32, s32, 16
; GFX11-NEXT: s_waitcnt lgkmcnt(0)
; GFX11-NEXT: s_clause 0x7
@@ -5457,18 +5956,19 @@
; GFX11-NEXT: global_load_b128 v[20:23], v28, s[0:1] offset:80
; GFX11-NEXT: global_load_b128 v[24:27], v28, s[0:1] offset:96
; GFX11-NEXT: global_load_b128 v[28:31], v28, s[0:1] offset:112
-; GFX11-NEXT: v_writelane_b32 v40, s30, 0
+; GFX11-NEXT: v_writelane_b32 v40, s31, 1
; GFX11-NEXT: s_getpc_b64 s[0:1]
; GFX11-NEXT: s_add_u32 s0, s0, external_void_func_v32i32@rel32@lo+4
; GFX11-NEXT: s_addc_u32 s1, s1, external_void_func_v32i32@rel32@hi+12
-; GFX11-NEXT: v_writelane_b32 v40, s31, 1
+; GFX11-NEXT: s_delay_alu instid0(SALU_CYCLE_1)
; GFX11-NEXT: s_swappc_b64 s[30:31], s[0:1]
-; GFX11-NEXT: s_delay_alu instid0(VALU_DEP_1)
; GFX11-NEXT: v_readlane_b32 s31, v40, 1
; GFX11-NEXT: v_readlane_b32 s30, v40, 0
-; GFX11-NEXT: v_readlane_b32 s0, v40, 2
+; GFX11-NEXT: v_readlane_b32 s0, v41, 0
; GFX11-NEXT: s_or_saveexec_b32 s1, -1
-; GFX11-NEXT: scratch_load_b32 v40, off, s33 ; 4-byte Folded Reload
+; GFX11-NEXT: s_clause 0x1
+; GFX11-NEXT: scratch_load_b32 v40, off, s33
+; GFX11-NEXT: scratch_load_b32 v41, off, s33 offset:4
; GFX11-NEXT: s_mov_b32 exec_lo, s1
; GFX11-NEXT: s_add_i32 s32, s32, -16
; GFX11-NEXT: s_mov_b32 s33, s0
@@ -5482,11 +5982,13 @@
; GFX10-SCRATCH-NEXT: s_mov_b32 s33, s32
; GFX10-SCRATCH-NEXT: s_or_saveexec_b32 s1, -1
; GFX10-SCRATCH-NEXT: scratch_store_dword off, v40, s33 ; 4-byte Folded Spill
+; GFX10-SCRATCH-NEXT: scratch_store_dword off, v41, s33 offset:4 ; 4-byte Folded Spill
; GFX10-SCRATCH-NEXT: s_waitcnt_depctr 0xffe3
; GFX10-SCRATCH-NEXT: s_mov_b32 exec_lo, s1
-; GFX10-SCRATCH-NEXT: v_writelane_b32 v40, s0, 2
+; GFX10-SCRATCH-NEXT: v_writelane_b32 v41, s0, 0
; GFX10-SCRATCH-NEXT: s_load_dwordx2 s[0:1], s[0:1], 0x0
; GFX10-SCRATCH-NEXT: v_mov_b32_e32 v32, 0
+; GFX10-SCRATCH-NEXT: v_writelane_b32 v40, s30, 0
; GFX10-SCRATCH-NEXT: s_add_i32 s32, s32, 16
; GFX10-SCRATCH-NEXT: s_waitcnt lgkmcnt(0)
; GFX10-SCRATCH-NEXT: s_clause 0x7
@@ -5498,17 +6000,18 @@
; GFX10-SCRATCH-NEXT: global_load_dwordx4 v[20:23], v32, s[0:1] offset:80
; GFX10-SCRATCH-NEXT: global_load_dwordx4 v[24:27], v32, s[0:1] offset:96
; GFX10-SCRATCH-NEXT: global_load_dwordx4 v[28:31], v32, s[0:1] offset:112
-; GFX10-SCRATCH-NEXT: v_writelane_b32 v40, s30, 0
+; GFX10-SCRATCH-NEXT: v_writelane_b32 v40, s31, 1
; GFX10-SCRATCH-NEXT: s_getpc_b64 s[0:1]
; GFX10-SCRATCH-NEXT: s_add_u32 s0, s0, external_void_func_v32i32@rel32@lo+4
; GFX10-SCRATCH-NEXT: s_addc_u32 s1, s1, external_void_func_v32i32@rel32@hi+12
-; GFX10-SCRATCH-NEXT: v_writelane_b32 v40, s31, 1
; GFX10-SCRATCH-NEXT: s_swappc_b64 s[30:31], s[0:1]
; GFX10-SCRATCH-NEXT: v_readlane_b32 s31, v40, 1
; GFX10-SCRATCH-NEXT: v_readlane_b32 s30, v40, 0
-; GFX10-SCRATCH-NEXT: v_readlane_b32 s0, v40, 2
+; GFX10-SCRATCH-NEXT: v_readlane_b32 s0, v41, 0
; GFX10-SCRATCH-NEXT: s_or_saveexec_b32 s1, -1
-; GFX10-SCRATCH-NEXT: scratch_load_dword v40, off, s33 ; 4-byte Folded Reload
+; GFX10-SCRATCH-NEXT: s_clause 0x1
+; GFX10-SCRATCH-NEXT: scratch_load_dword v40, off, s33
+; GFX10-SCRATCH-NEXT: scratch_load_dword v41, off, s33 offset:4
; GFX10-SCRATCH-NEXT: s_waitcnt_depctr 0xffe3
; GFX10-SCRATCH-NEXT: s_mov_b32 exec_lo, s1
; GFX10-SCRATCH-NEXT: s_add_i32 s32, s32, -16
@@ -5529,8 +6032,9 @@
; GFX9-NEXT: s_mov_b32 s33, s32
; GFX9-NEXT: s_or_saveexec_b64 s[36:37], -1
; GFX9-NEXT: buffer_store_dword v40, off, s[0:3], s33 ; 4-byte Folded Spill
+; GFX9-NEXT: buffer_store_dword v41, off, s[0:3], s33 offset:4 ; 4-byte Folded Spill
; GFX9-NEXT: s_mov_b64 exec, s[36:37]
-; GFX9-NEXT: v_writelane_b32 v40, s34, 2
+; GFX9-NEXT: v_writelane_b32 v41, s34, 0
; GFX9-NEXT: s_load_dwordx2 s[34:35], s[34:35], 0x0
; GFX9-NEXT: v_mov_b32_e32 v28, 0
; GFX9-NEXT: global_load_dword v32, v[0:1], off
@@ -5555,9 +6059,10 @@
; GFX9-NEXT: s_swappc_b64 s[30:31], s[34:35]
; GFX9-NEXT: v_readlane_b32 s31, v40, 1
; GFX9-NEXT: v_readlane_b32 s30, v40, 0
-; GFX9-NEXT: v_readlane_b32 s34, v40, 2
+; GFX9-NEXT: v_readlane_b32 s34, v41, 0
; GFX9-NEXT: s_or_saveexec_b64 s[36:37], -1
; GFX9-NEXT: buffer_load_dword v40, off, s[0:3], s33 ; 4-byte Folded Reload
+; GFX9-NEXT: buffer_load_dword v41, off, s[0:3], s33 offset:4 ; 4-byte Folded Reload
; GFX9-NEXT: s_mov_b64 exec, s[36:37]
; GFX9-NEXT: s_addk_i32 s32, 0xfc00
; GFX9-NEXT: s_mov_b32 s33, s34
@@ -5571,11 +6076,13 @@
; GFX10-NEXT: s_mov_b32 s33, s32
; GFX10-NEXT: s_or_saveexec_b32 s35, -1
; GFX10-NEXT: buffer_store_dword v40, off, s[0:3], s33 ; 4-byte Folded Spill
+; GFX10-NEXT: buffer_store_dword v41, off, s[0:3], s33 offset:4 ; 4-byte Folded Spill
; GFX10-NEXT: s_waitcnt_depctr 0xffe3
; GFX10-NEXT: s_mov_b32 exec_lo, s35
-; GFX10-NEXT: v_writelane_b32 v40, s34, 2
+; GFX10-NEXT: v_writelane_b32 v41, s34, 0
; GFX10-NEXT: s_load_dwordx2 s[34:35], s[34:35], 0x0
; GFX10-NEXT: v_mov_b32_e32 v32, 0
+; GFX10-NEXT: v_writelane_b32 v40, s30, 0
; GFX10-NEXT: s_addk_i32 s32, 0x200
; GFX10-NEXT: global_load_dword v33, v[0:1], off
; GFX10-NEXT: s_waitcnt lgkmcnt(0)
@@ -5588,19 +6095,20 @@
; GFX10-NEXT: global_load_dwordx4 v[20:23], v32, s[34:35] offset:80
; GFX10-NEXT: global_load_dwordx4 v[24:27], v32, s[34:35] offset:96
; GFX10-NEXT: global_load_dwordx4 v[28:31], v32, s[34:35] offset:112
-; GFX10-NEXT: v_writelane_b32 v40, s30, 0
+; GFX10-NEXT: v_writelane_b32 v40, s31, 1
; GFX10-NEXT: s_getpc_b64 s[34:35]
; GFX10-NEXT: s_add_u32 s34, s34, external_void_func_v32i32_i32@rel32@lo+4
; GFX10-NEXT: s_addc_u32 s35, s35, external_void_func_v32i32_i32@rel32@hi+12
-; GFX10-NEXT: v_writelane_b32 v40, s31, 1
; GFX10-NEXT: s_waitcnt vmcnt(8)
; GFX10-NEXT: buffer_store_dword v33, off, s[0:3], s32
; GFX10-NEXT: s_swappc_b64 s[30:31], s[34:35]
; GFX10-NEXT: v_readlane_b32 s31, v40, 1
; GFX10-NEXT: v_readlane_b32 s30, v40, 0
-; GFX10-NEXT: v_readlane_b32 s34, v40, 2
+; GFX10-NEXT: v_readlane_b32 s34, v41, 0
; GFX10-NEXT: s_or_saveexec_b32 s35, -1
-; GFX10-NEXT: buffer_load_dword v40, off, s[0:3], s33 ; 4-byte Folded Reload
+; GFX10-NEXT: s_clause 0x1
+; GFX10-NEXT: buffer_load_dword v40, off, s[0:3], s33
+; GFX10-NEXT: buffer_load_dword v41, off, s[0:3], s33 offset:4
; GFX10-NEXT: s_waitcnt_depctr 0xffe3
; GFX10-NEXT: s_mov_b32 exec_lo, s35
; GFX10-NEXT: s_addk_i32 s32, 0xfe00
@@ -5614,11 +6122,14 @@
; GFX11-NEXT: s_mov_b32 s0, s33
; GFX11-NEXT: s_mov_b32 s33, s32
; GFX11-NEXT: s_or_saveexec_b32 s1, -1
-; GFX11-NEXT: scratch_store_b32 off, v40, s33 ; 4-byte Folded Spill
+; GFX11-NEXT: s_clause 0x1
+; GFX11-NEXT: scratch_store_b32 off, v40, s33
+; GFX11-NEXT: scratch_store_b32 off, v41, s33 offset:4
; GFX11-NEXT: s_mov_b32 exec_lo, s1
-; GFX11-NEXT: v_writelane_b32 v40, s0, 2
+; GFX11-NEXT: v_writelane_b32 v41, s0, 0
; GFX11-NEXT: s_load_b64 s[0:1], s[0:1], 0x0
; GFX11-NEXT: v_mov_b32_e32 v28, 0
+; GFX11-NEXT: v_writelane_b32 v40, s30, 0
; GFX11-NEXT: s_add_i32 s32, s32, 16
; GFX11-NEXT: global_load_b32 v32, v[0:1], off
; GFX11-NEXT: s_waitcnt lgkmcnt(0)
@@ -5631,19 +6142,20 @@
; GFX11-NEXT: global_load_b128 v[20:23], v28, s[0:1] offset:80
; GFX11-NEXT: global_load_b128 v[24:27], v28, s[0:1] offset:96
; GFX11-NEXT: global_load_b128 v[28:31], v28, s[0:1] offset:112
-; GFX11-NEXT: v_writelane_b32 v40, s30, 0
+; GFX11-NEXT: v_writelane_b32 v40, s31, 1
; GFX11-NEXT: s_getpc_b64 s[0:1]
; GFX11-NEXT: s_add_u32 s0, s0, external_void_func_v32i32_i32@rel32@lo+4
; GFX11-NEXT: s_addc_u32 s1, s1, external_void_func_v32i32_i32@rel32@hi+12
-; GFX11-NEXT: v_writelane_b32 v40, s31, 1
; GFX11-NEXT: s_waitcnt vmcnt(8)
; GFX11-NEXT: scratch_store_b32 off, v32, s32
; GFX11-NEXT: s_swappc_b64 s[30:31], s[0:1]
; GFX11-NEXT: v_readlane_b32 s31, v40, 1
; GFX11-NEXT: v_readlane_b32 s30, v40, 0
-; GFX11-NEXT: v_readlane_b32 s0, v40, 2
+; GFX11-NEXT: v_readlane_b32 s0, v41, 0
; GFX11-NEXT: s_or_saveexec_b32 s1, -1
-; GFX11-NEXT: scratch_load_b32 v40, off, s33 ; 4-byte Folded Reload
+; GFX11-NEXT: s_clause 0x1
+; GFX11-NEXT: scratch_load_b32 v40, off, s33
+; GFX11-NEXT: scratch_load_b32 v41, off, s33 offset:4
; GFX11-NEXT: s_mov_b32 exec_lo, s1
; GFX11-NEXT: s_add_i32 s32, s32, -16
; GFX11-NEXT: s_mov_b32 s33, s0
@@ -5657,11 +6169,13 @@
; GFX10-SCRATCH-NEXT: s_mov_b32 s33, s32
; GFX10-SCRATCH-NEXT: s_or_saveexec_b32 s1, -1
; GFX10-SCRATCH-NEXT: scratch_store_dword off, v40, s33 ; 4-byte Folded Spill
+; GFX10-SCRATCH-NEXT: scratch_store_dword off, v41, s33 offset:4 ; 4-byte Folded Spill
; GFX10-SCRATCH-NEXT: s_waitcnt_depctr 0xffe3
; GFX10-SCRATCH-NEXT: s_mov_b32 exec_lo, s1
-; GFX10-SCRATCH-NEXT: v_writelane_b32 v40, s0, 2
+; GFX10-SCRATCH-NEXT: v_writelane_b32 v41, s0, 0
; GFX10-SCRATCH-NEXT: s_load_dwordx2 s[0:1], s[0:1], 0x0
; GFX10-SCRATCH-NEXT: v_mov_b32_e32 v32, 0
+; GFX10-SCRATCH-NEXT: v_writelane_b32 v40, s30, 0
; GFX10-SCRATCH-NEXT: s_add_i32 s32, s32, 16
; GFX10-SCRATCH-NEXT: global_load_dword v33, v[0:1], off
; GFX10-SCRATCH-NEXT: s_waitcnt lgkmcnt(0)
@@ -5674,19 +6188,20 @@
; GFX10-SCRATCH-NEXT: global_load_dwordx4 v[20:23], v32, s[0:1] offset:80
; GFX10-SCRATCH-NEXT: global_load_dwordx4 v[24:27], v32, s[0:1] offset:96
; GFX10-SCRATCH-NEXT: global_load_dwordx4 v[28:31], v32, s[0:1] offset:112
-; GFX10-SCRATCH-NEXT: v_writelane_b32 v40, s30, 0
+; GFX10-SCRATCH-NEXT: v_writelane_b32 v40, s31, 1
; GFX10-SCRATCH-NEXT: s_getpc_b64 s[0:1]
; GFX10-SCRATCH-NEXT: s_add_u32 s0, s0, external_void_func_v32i32_i32@rel32@lo+4
; GFX10-SCRATCH-NEXT: s_addc_u32 s1, s1, external_void_func_v32i32_i32@rel32@hi+12
-; GFX10-SCRATCH-NEXT: v_writelane_b32 v40, s31, 1
; GFX10-SCRATCH-NEXT: s_waitcnt vmcnt(8)
; GFX10-SCRATCH-NEXT: scratch_store_dword off, v33, s32
; GFX10-SCRATCH-NEXT: s_swappc_b64 s[30:31], s[0:1]
; GFX10-SCRATCH-NEXT: v_readlane_b32 s31, v40, 1
; GFX10-SCRATCH-NEXT: v_readlane_b32 s30, v40, 0
-; GFX10-SCRATCH-NEXT: v_readlane_b32 s0, v40, 2
+; GFX10-SCRATCH-NEXT: v_readlane_b32 s0, v41, 0
; GFX10-SCRATCH-NEXT: s_or_saveexec_b32 s1, -1
-; GFX10-SCRATCH-NEXT: scratch_load_dword v40, off, s33 ; 4-byte Folded Reload
+; GFX10-SCRATCH-NEXT: s_clause 0x1
+; GFX10-SCRATCH-NEXT: scratch_load_dword v40, off, s33
+; GFX10-SCRATCH-NEXT: scratch_load_dword v41, off, s33 offset:4
; GFX10-SCRATCH-NEXT: s_waitcnt_depctr 0xffe3
; GFX10-SCRATCH-NEXT: s_mov_b32 exec_lo, s1
; GFX10-SCRATCH-NEXT: s_add_i32 s32, s32, -16
@@ -5708,14 +6223,15 @@
; GFX9-NEXT: s_mov_b32 s33, s32
; GFX9-NEXT: s_or_saveexec_b64 s[36:37], -1
; GFX9-NEXT: buffer_store_dword v40, off, s[0:3], s33 offset:8 ; 4-byte Folded Spill
+; GFX9-NEXT: buffer_store_dword v43, off, s[0:3], s33 offset:12 ; 4-byte Folded Spill
; GFX9-NEXT: s_mov_b64 exec, s[36:37]
-; GFX9-NEXT: v_writelane_b32 v40, s34, 2
-; GFX9-NEXT: s_addk_i32 s32, 0x400
+; GFX9-NEXT: s_addk_i32 s32, 0x800
; GFX9-NEXT: buffer_store_dword v41, off, s[0:3], s33 offset:4 ; 4-byte Folded Spill
; GFX9-NEXT: buffer_store_dword v42, off, s[0:3], s33 ; 4-byte Folded Spill
; GFX9-NEXT: v_writelane_b32 v40, s30, 0
; GFX9-NEXT: v_mov_b32_e32 v41, v0
; GFX9-NEXT: v_mov_b32_e32 v0, 42
+; GFX9-NEXT: v_writelane_b32 v43, s34, 0
; GFX9-NEXT: v_writelane_b32 v40, s31, 1
; GFX9-NEXT: v_mov_b32_e32 v42, v1
; GFX9-NEXT: s_getpc_b64 s[34:35]
@@ -5728,11 +6244,12 @@
; GFX9-NEXT: buffer_load_dword v41, off, s[0:3], s33 offset:4 ; 4-byte Folded Reload
; GFX9-NEXT: v_readlane_b32 s31, v40, 1
; GFX9-NEXT: v_readlane_b32 s30, v40, 0
-; GFX9-NEXT: v_readlane_b32 s34, v40, 2
+; GFX9-NEXT: v_readlane_b32 s34, v43, 0
; GFX9-NEXT: s_or_saveexec_b64 s[36:37], -1
; GFX9-NEXT: buffer_load_dword v40, off, s[0:3], s33 offset:8 ; 4-byte Folded Reload
+; GFX9-NEXT: buffer_load_dword v43, off, s[0:3], s33 offset:12 ; 4-byte Folded Reload
; GFX9-NEXT: s_mov_b64 exec, s[36:37]
-; GFX9-NEXT: s_addk_i32 s32, 0xfc00
+; GFX9-NEXT: s_addk_i32 s32, 0xf800
; GFX9-NEXT: s_mov_b32 s33, s34
; GFX9-NEXT: s_waitcnt vmcnt(0)
; GFX9-NEXT: s_setpc_b64 s[30:31]
@@ -5744,20 +6261,21 @@
; GFX10-NEXT: s_mov_b32 s33, s32
; GFX10-NEXT: s_or_saveexec_b32 s35, -1
; GFX10-NEXT: buffer_store_dword v40, off, s[0:3], s33 offset:8 ; 4-byte Folded Spill
+; GFX10-NEXT: buffer_store_dword v43, off, s[0:3], s33 offset:12 ; 4-byte Folded Spill
; GFX10-NEXT: s_waitcnt_depctr 0xffe3
; GFX10-NEXT: s_mov_b32 exec_lo, s35
-; GFX10-NEXT: v_writelane_b32 v40, s34, 2
; GFX10-NEXT: buffer_store_dword v41, off, s[0:3], s33 offset:4 ; 4-byte Folded Spill
; GFX10-NEXT: buffer_store_dword v42, off, s[0:3], s33 ; 4-byte Folded Spill
+; GFX10-NEXT: v_writelane_b32 v40, s30, 0
; GFX10-NEXT: v_mov_b32_e32 v41, v0
; GFX10-NEXT: v_mov_b32_e32 v0, 42
-; GFX10-NEXT: s_addk_i32 s32, 0x200
-; GFX10-NEXT: v_writelane_b32 v40, s30, 0
+; GFX10-NEXT: s_addk_i32 s32, 0x400
+; GFX10-NEXT: v_writelane_b32 v43, s34, 0
+; GFX10-NEXT: v_writelane_b32 v40, s31, 1
; GFX10-NEXT: v_mov_b32_e32 v42, v1
; GFX10-NEXT: s_getpc_b64 s[34:35]
; GFX10-NEXT: s_add_u32 s34, s34, external_i32_func_i32@rel32@lo+4
; GFX10-NEXT: s_addc_u32 s35, s35, external_i32_func_i32@rel32@hi+12
-; GFX10-NEXT: v_writelane_b32 v40, s31, 1
; GFX10-NEXT: s_swappc_b64 s[30:31], s[34:35]
; GFX10-NEXT: global_store_dword v[41:42], v0, off
; GFX10-NEXT: s_waitcnt_vscnt null, 0x0
@@ -5766,12 +6284,14 @@
; GFX10-NEXT: buffer_load_dword v41, off, s[0:3], s33 offset:4
; GFX10-NEXT: v_readlane_b32 s31, v40, 1
; GFX10-NEXT: v_readlane_b32 s30, v40, 0
-; GFX10-NEXT: v_readlane_b32 s34, v40, 2
+; GFX10-NEXT: v_readlane_b32 s34, v43, 0
; GFX10-NEXT: s_or_saveexec_b32 s35, -1
-; GFX10-NEXT: buffer_load_dword v40, off, s[0:3], s33 offset:8 ; 4-byte Folded Reload
+; GFX10-NEXT: s_clause 0x1
+; GFX10-NEXT: buffer_load_dword v40, off, s[0:3], s33 offset:8
+; GFX10-NEXT: buffer_load_dword v43, off, s[0:3], s33 offset:12
; GFX10-NEXT: s_waitcnt_depctr 0xffe3
; GFX10-NEXT: s_mov_b32 exec_lo, s35
-; GFX10-NEXT: s_addk_i32 s32, 0xfe00
+; GFX10-NEXT: s_addk_i32 s32, 0xfc00
; GFX10-NEXT: s_mov_b32 s33, s34
; GFX10-NEXT: s_waitcnt vmcnt(0)
; GFX10-NEXT: s_setpc_b64 s[30:31]
@@ -5782,20 +6302,23 @@
; GFX11-NEXT: s_mov_b32 s0, s33
; GFX11-NEXT: s_mov_b32 s33, s32
; GFX11-NEXT: s_or_saveexec_b32 s1, -1
-; GFX11-NEXT: scratch_store_b32 off, v40, s33 offset:8 ; 4-byte Folded Spill
+; GFX11-NEXT: s_clause 0x1
+; GFX11-NEXT: scratch_store_b32 off, v40, s33 offset:8
+; GFX11-NEXT: scratch_store_b32 off, v43, s33 offset:12
; GFX11-NEXT: s_mov_b32 exec_lo, s1
-; GFX11-NEXT: v_writelane_b32 v40, s0, 2
; GFX11-NEXT: s_clause 0x1
; GFX11-NEXT: scratch_store_b32 off, v41, s33 offset:4
; GFX11-NEXT: scratch_store_b32 off, v42, s33
+; GFX11-NEXT: v_writelane_b32 v40, s30, 0
; GFX11-NEXT: v_dual_mov_b32 v42, v1 :: v_dual_mov_b32 v41, v0
; GFX11-NEXT: v_mov_b32_e32 v0, 42
-; GFX11-NEXT: v_writelane_b32 v40, s30, 0
-; GFX11-NEXT: s_add_i32 s32, s32, 16
+; GFX11-NEXT: s_add_i32 s32, s32, 32
+; GFX11-NEXT: v_writelane_b32 v43, s0, 0
+; GFX11-NEXT: v_writelane_b32 v40, s31, 1
; GFX11-NEXT: s_getpc_b64 s[0:1]
; GFX11-NEXT: s_add_u32 s0, s0, external_i32_func_i32@rel32@lo+4
; GFX11-NEXT: s_addc_u32 s1, s1, external_i32_func_i32@rel32@hi+12
-; GFX11-NEXT: v_writelane_b32 v40, s31, 1
+; GFX11-NEXT: s_delay_alu instid0(SALU_CYCLE_1)
; GFX11-NEXT: s_swappc_b64 s[30:31], s[0:1]
; GFX11-NEXT: global_store_b32 v[41:42], v0, off dlc
; GFX11-NEXT: s_waitcnt_vscnt null, 0x0
@@ -5804,11 +6327,13 @@
; GFX11-NEXT: scratch_load_b32 v41, off, s33 offset:4
; GFX11-NEXT: v_readlane_b32 s31, v40, 1
; GFX11-NEXT: v_readlane_b32 s30, v40, 0
-; GFX11-NEXT: v_readlane_b32 s0, v40, 2
+; GFX11-NEXT: v_readlane_b32 s0, v43, 0
; GFX11-NEXT: s_or_saveexec_b32 s1, -1
-; GFX11-NEXT: scratch_load_b32 v40, off, s33 offset:8 ; 4-byte Folded Reload
+; GFX11-NEXT: s_clause 0x1
+; GFX11-NEXT: scratch_load_b32 v40, off, s33 offset:8
+; GFX11-NEXT: scratch_load_b32 v43, off, s33 offset:12
; GFX11-NEXT: s_mov_b32 exec_lo, s1
-; GFX11-NEXT: s_add_i32 s32, s32, -16
+; GFX11-NEXT: s_addk_i32 s32, 0xffe0
; GFX11-NEXT: s_mov_b32 s33, s0
; GFX11-NEXT: s_waitcnt vmcnt(0)
; GFX11-NEXT: s_setpc_b64 s[30:31]
@@ -5820,20 +6345,21 @@
; GFX10-SCRATCH-NEXT: s_mov_b32 s33, s32
; GFX10-SCRATCH-NEXT: s_or_saveexec_b32 s1, -1
; GFX10-SCRATCH-NEXT: scratch_store_dword off, v40, s33 offset:8 ; 4-byte Folded Spill
+; GFX10-SCRATCH-NEXT: scratch_store_dword off, v43, s33 offset:12 ; 4-byte Folded Spill
; GFX10-SCRATCH-NEXT: s_waitcnt_depctr 0xffe3
; GFX10-SCRATCH-NEXT: s_mov_b32 exec_lo, s1
-; GFX10-SCRATCH-NEXT: v_writelane_b32 v40, s0, 2
; GFX10-SCRATCH-NEXT: scratch_store_dword off, v41, s33 offset:4 ; 4-byte Folded Spill
; GFX10-SCRATCH-NEXT: scratch_store_dword off, v42, s33 ; 4-byte Folded Spill
+; GFX10-SCRATCH-NEXT: v_writelane_b32 v40, s30, 0
; GFX10-SCRATCH-NEXT: v_mov_b32_e32 v41, v0
; GFX10-SCRATCH-NEXT: v_mov_b32_e32 v0, 42
-; GFX10-SCRATCH-NEXT: s_add_i32 s32, s32, 16
-; GFX10-SCRATCH-NEXT: v_writelane_b32 v40, s30, 0
+; GFX10-SCRATCH-NEXT: s_add_i32 s32, s32, 32
+; GFX10-SCRATCH-NEXT: v_writelane_b32 v43, s0, 0
+; GFX10-SCRATCH-NEXT: v_writelane_b32 v40, s31, 1
; GFX10-SCRATCH-NEXT: v_mov_b32_e32 v42, v1
; GFX10-SCRATCH-NEXT: s_getpc_b64 s[0:1]
; GFX10-SCRATCH-NEXT: s_add_u32 s0, s0, external_i32_func_i32@rel32@lo+4
; GFX10-SCRATCH-NEXT: s_addc_u32 s1, s1, external_i32_func_i32@rel32@hi+12
-; GFX10-SCRATCH-NEXT: v_writelane_b32 v40, s31, 1
; GFX10-SCRATCH-NEXT: s_swappc_b64 s[30:31], s[0:1]
; GFX10-SCRATCH-NEXT: global_store_dword v[41:42], v0, off
; GFX10-SCRATCH-NEXT: s_waitcnt_vscnt null, 0x0
@@ -5842,12 +6368,14 @@
; GFX10-SCRATCH-NEXT: scratch_load_dword v41, off, s33 offset:4
; GFX10-SCRATCH-NEXT: v_readlane_b32 s31, v40, 1
; GFX10-SCRATCH-NEXT: v_readlane_b32 s30, v40, 0
-; GFX10-SCRATCH-NEXT: v_readlane_b32 s0, v40, 2
+; GFX10-SCRATCH-NEXT: v_readlane_b32 s0, v43, 0
; GFX10-SCRATCH-NEXT: s_or_saveexec_b32 s1, -1
-; GFX10-SCRATCH-NEXT: scratch_load_dword v40, off, s33 offset:8 ; 4-byte Folded Reload
+; GFX10-SCRATCH-NEXT: s_clause 0x1
+; GFX10-SCRATCH-NEXT: scratch_load_dword v40, off, s33 offset:8
+; GFX10-SCRATCH-NEXT: scratch_load_dword v43, off, s33 offset:12
; GFX10-SCRATCH-NEXT: s_waitcnt_depctr 0xffe3
; GFX10-SCRATCH-NEXT: s_mov_b32 exec_lo, s1
-; GFX10-SCRATCH-NEXT: s_add_i32 s32, s32, -16
+; GFX10-SCRATCH-NEXT: s_addk_i32 s32, 0xffe0
; GFX10-SCRATCH-NEXT: s_mov_b32 s33, s0
; GFX10-SCRATCH-NEXT: s_waitcnt vmcnt(0)
; GFX10-SCRATCH-NEXT: s_setpc_b64 s[30:31]
@@ -5864,8 +6392,9 @@
; GFX9-NEXT: s_mov_b32 s33, s32
; GFX9-NEXT: s_or_saveexec_b64 s[36:37], -1
; GFX9-NEXT: buffer_store_dword v40, off, s[0:3], s33 ; 4-byte Folded Spill
+; GFX9-NEXT: buffer_store_dword v41, off, s[0:3], s33 offset:4 ; 4-byte Folded Spill
; GFX9-NEXT: s_mov_b64 exec, s[36:37]
-; GFX9-NEXT: v_writelane_b32 v40, s34, 2
+; GFX9-NEXT: v_writelane_b32 v41, s34, 0
; GFX9-NEXT: s_load_dwordx2 s[34:35], s[34:35], 0x0
; GFX9-NEXT: v_mov_b32_e32 v2, 0
; GFX9-NEXT: s_addk_i32 s32, 0x400
@@ -5880,9 +6409,10 @@
; GFX9-NEXT: s_swappc_b64 s[30:31], s[34:35]
; GFX9-NEXT: v_readlane_b32 s31, v40, 1
; GFX9-NEXT: v_readlane_b32 s30, v40, 0
-; GFX9-NEXT: v_readlane_b32 s34, v40, 2
+; GFX9-NEXT: v_readlane_b32 s34, v41, 0
; GFX9-NEXT: s_or_saveexec_b64 s[36:37], -1
; GFX9-NEXT: buffer_load_dword v40, off, s[0:3], s33 ; 4-byte Folded Reload
+; GFX9-NEXT: buffer_load_dword v41, off, s[0:3], s33 offset:4 ; 4-byte Folded Reload
; GFX9-NEXT: s_mov_b64 exec, s[36:37]
; GFX9-NEXT: s_addk_i32 s32, 0xfc00
; GFX9-NEXT: s_mov_b32 s33, s34
@@ -5896,27 +6426,30 @@
; GFX10-NEXT: s_mov_b32 s33, s32
; GFX10-NEXT: s_or_saveexec_b32 s35, -1
; GFX10-NEXT: buffer_store_dword v40, off, s[0:3], s33 ; 4-byte Folded Spill
+; GFX10-NEXT: buffer_store_dword v41, off, s[0:3], s33 offset:4 ; 4-byte Folded Spill
; GFX10-NEXT: s_waitcnt_depctr 0xffe3
; GFX10-NEXT: s_mov_b32 exec_lo, s35
-; GFX10-NEXT: v_writelane_b32 v40, s34, 2
+; GFX10-NEXT: v_writelane_b32 v41, s34, 0
; GFX10-NEXT: s_load_dwordx2 s[34:35], s[34:35], 0x0
; GFX10-NEXT: v_mov_b32_e32 v2, 0
+; GFX10-NEXT: v_writelane_b32 v40, s30, 0
; GFX10-NEXT: s_addk_i32 s32, 0x200
; GFX10-NEXT: s_waitcnt lgkmcnt(0)
; GFX10-NEXT: s_clause 0x1
; GFX10-NEXT: global_load_ubyte v0, v2, s[34:35]
; GFX10-NEXT: global_load_dword v1, v2, s[34:35] offset:4
-; GFX10-NEXT: v_writelane_b32 v40, s30, 0
+; GFX10-NEXT: v_writelane_b32 v40, s31, 1
; GFX10-NEXT: s_getpc_b64 s[34:35]
; GFX10-NEXT: s_add_u32 s34, s34, external_void_func_struct_i8_i32@rel32@lo+4
; GFX10-NEXT: s_addc_u32 s35, s35, external_void_func_struct_i8_i32@rel32@hi+12
-; GFX10-NEXT: v_writelane_b32 v40, s31, 1
; GFX10-NEXT: s_swappc_b64 s[30:31], s[34:35]
; GFX10-NEXT: v_readlane_b32 s31, v40, 1
; GFX10-NEXT: v_readlane_b32 s30, v40, 0
-; GFX10-NEXT: v_readlane_b32 s34, v40, 2
+; GFX10-NEXT: v_readlane_b32 s34, v41, 0
; GFX10-NEXT: s_or_saveexec_b32 s35, -1
-; GFX10-NEXT: buffer_load_dword v40, off, s[0:3], s33 ; 4-byte Folded Reload
+; GFX10-NEXT: s_clause 0x1
+; GFX10-NEXT: buffer_load_dword v40, off, s[0:3], s33
+; GFX10-NEXT: buffer_load_dword v41, off, s[0:3], s33 offset:4
; GFX10-NEXT: s_waitcnt_depctr 0xffe3
; GFX10-NEXT: s_mov_b32 exec_lo, s35
; GFX10-NEXT: s_addk_i32 s32, 0xfe00
@@ -5930,28 +6463,32 @@
; GFX11-NEXT: s_mov_b32 s0, s33
; GFX11-NEXT: s_mov_b32 s33, s32
; GFX11-NEXT: s_or_saveexec_b32 s1, -1
-; GFX11-NEXT: scratch_store_b32 off, v40, s33 ; 4-byte Folded Spill
+; GFX11-NEXT: s_clause 0x1
+; GFX11-NEXT: scratch_store_b32 off, v40, s33
+; GFX11-NEXT: scratch_store_b32 off, v41, s33 offset:4
; GFX11-NEXT: s_mov_b32 exec_lo, s1
-; GFX11-NEXT: v_writelane_b32 v40, s0, 2
+; GFX11-NEXT: v_writelane_b32 v41, s0, 0
; GFX11-NEXT: s_load_b64 s[0:1], s[0:1], 0x0
; GFX11-NEXT: v_mov_b32_e32 v1, 0
+; GFX11-NEXT: v_writelane_b32 v40, s30, 0
; GFX11-NEXT: s_add_i32 s32, s32, 16
; GFX11-NEXT: s_waitcnt lgkmcnt(0)
; GFX11-NEXT: s_clause 0x1
; GFX11-NEXT: global_load_u8 v0, v1, s[0:1]
; GFX11-NEXT: global_load_b32 v1, v1, s[0:1] offset:4
-; GFX11-NEXT: v_writelane_b32 v40, s30, 0
+; GFX11-NEXT: v_writelane_b32 v40, s31, 1
; GFX11-NEXT: s_getpc_b64 s[0:1]
; GFX11-NEXT: s_add_u32 s0, s0, external_void_func_struct_i8_i32@rel32@lo+4
; GFX11-NEXT: s_addc_u32 s1, s1, external_void_func_struct_i8_i32@rel32@hi+12
-; GFX11-NEXT: v_writelane_b32 v40, s31, 1
+; GFX11-NEXT: s_delay_alu instid0(SALU_CYCLE_1)
; GFX11-NEXT: s_swappc_b64 s[30:31], s[0:1]
-; GFX11-NEXT: s_delay_alu instid0(VALU_DEP_1)
; GFX11-NEXT: v_readlane_b32 s31, v40, 1
; GFX11-NEXT: v_readlane_b32 s30, v40, 0
-; GFX11-NEXT: v_readlane_b32 s0, v40, 2
+; GFX11-NEXT: v_readlane_b32 s0, v41, 0
; GFX11-NEXT: s_or_saveexec_b32 s1, -1
-; GFX11-NEXT: scratch_load_b32 v40, off, s33 ; 4-byte Folded Reload
+; GFX11-NEXT: s_clause 0x1
+; GFX11-NEXT: scratch_load_b32 v40, off, s33
+; GFX11-NEXT: scratch_load_b32 v41, off, s33 offset:4
; GFX11-NEXT: s_mov_b32 exec_lo, s1
; GFX11-NEXT: s_add_i32 s32, s32, -16
; GFX11-NEXT: s_mov_b32 s33, s0
@@ -5965,27 +6502,30 @@
; GFX10-SCRATCH-NEXT: s_mov_b32 s33, s32
; GFX10-SCRATCH-NEXT: s_or_saveexec_b32 s1, -1
; GFX10-SCRATCH-NEXT: scratch_store_dword off, v40, s33 ; 4-byte Folded Spill
+; GFX10-SCRATCH-NEXT: scratch_store_dword off, v41, s33 offset:4 ; 4-byte Folded Spill
; GFX10-SCRATCH-NEXT: s_waitcnt_depctr 0xffe3
; GFX10-SCRATCH-NEXT: s_mov_b32 exec_lo, s1
-; GFX10-SCRATCH-NEXT: v_writelane_b32 v40, s0, 2
+; GFX10-SCRATCH-NEXT: v_writelane_b32 v41, s0, 0
; GFX10-SCRATCH-NEXT: s_load_dwordx2 s[0:1], s[0:1], 0x0
; GFX10-SCRATCH-NEXT: v_mov_b32_e32 v2, 0
+; GFX10-SCRATCH-NEXT: v_writelane_b32 v40, s30, 0
; GFX10-SCRATCH-NEXT: s_add_i32 s32, s32, 16
; GFX10-SCRATCH-NEXT: s_waitcnt lgkmcnt(0)
; GFX10-SCRATCH-NEXT: s_clause 0x1
; GFX10-SCRATCH-NEXT: global_load_ubyte v0, v2, s[0:1]
; GFX10-SCRATCH-NEXT: global_load_dword v1, v2, s[0:1] offset:4
-; GFX10-SCRATCH-NEXT: v_writelane_b32 v40, s30, 0
+; GFX10-SCRATCH-NEXT: v_writelane_b32 v40, s31, 1
; GFX10-SCRATCH-NEXT: s_getpc_b64 s[0:1]
; GFX10-SCRATCH-NEXT: s_add_u32 s0, s0, external_void_func_struct_i8_i32@rel32@lo+4
; GFX10-SCRATCH-NEXT: s_addc_u32 s1, s1, external_void_func_struct_i8_i32@rel32@hi+12
-; GFX10-SCRATCH-NEXT: v_writelane_b32 v40, s31, 1
; GFX10-SCRATCH-NEXT: s_swappc_b64 s[30:31], s[0:1]
; GFX10-SCRATCH-NEXT: v_readlane_b32 s31, v40, 1
; GFX10-SCRATCH-NEXT: v_readlane_b32 s30, v40, 0
-; GFX10-SCRATCH-NEXT: v_readlane_b32 s0, v40, 2
+; GFX10-SCRATCH-NEXT: v_readlane_b32 s0, v41, 0
; GFX10-SCRATCH-NEXT: s_or_saveexec_b32 s1, -1
-; GFX10-SCRATCH-NEXT: scratch_load_dword v40, off, s33 ; 4-byte Folded Reload
+; GFX10-SCRATCH-NEXT: s_clause 0x1
+; GFX10-SCRATCH-NEXT: scratch_load_dword v40, off, s33
+; GFX10-SCRATCH-NEXT: scratch_load_dword v41, off, s33 offset:4
; GFX10-SCRATCH-NEXT: s_waitcnt_depctr 0xffe3
; GFX10-SCRATCH-NEXT: s_mov_b32 exec_lo, s1
; GFX10-SCRATCH-NEXT: s_add_i32 s32, s32, -16
@@ -6006,15 +6546,16 @@
; GFX9-NEXT: s_mov_b32 s33, s32
; GFX9-NEXT: s_or_saveexec_b64 s[36:37], -1
; GFX9-NEXT: buffer_store_dword v40, off, s[0:3], s33 offset:8 ; 4-byte Folded Spill
+; GFX9-NEXT: buffer_store_dword v41, off, s[0:3], s33 offset:12 ; 4-byte Folded Spill
; GFX9-NEXT: s_mov_b64 exec, s[36:37]
; GFX9-NEXT: v_mov_b32_e32 v0, 3
-; GFX9-NEXT: v_writelane_b32 v40, s34, 2
; GFX9-NEXT: buffer_store_byte v0, off, s[0:3], s33
; GFX9-NEXT: v_mov_b32_e32 v0, 8
-; GFX9-NEXT: s_addk_i32 s32, 0x400
+; GFX9-NEXT: s_addk_i32 s32, 0x800
; GFX9-NEXT: v_writelane_b32 v40, s30, 0
; GFX9-NEXT: buffer_store_dword v0, off, s[0:3], s33 offset:4
; GFX9-NEXT: v_lshrrev_b32_e64 v0, 6, s33
+; GFX9-NEXT: v_writelane_b32 v41, s34, 0
; GFX9-NEXT: v_writelane_b32 v40, s31, 1
; GFX9-NEXT: s_getpc_b64 s[34:35]
; GFX9-NEXT: s_add_u32 s34, s34, external_void_func_byval_struct_i8_i32@rel32@lo+4
@@ -6022,11 +6563,12 @@
; GFX9-NEXT: s_swappc_b64 s[30:31], s[34:35]
; GFX9-NEXT: v_readlane_b32 s31, v40, 1
; GFX9-NEXT: v_readlane_b32 s30, v40, 0
-; GFX9-NEXT: v_readlane_b32 s34, v40, 2
+; GFX9-NEXT: v_readlane_b32 s34, v41, 0
; GFX9-NEXT: s_or_saveexec_b64 s[36:37], -1
; GFX9-NEXT: buffer_load_dword v40, off, s[0:3], s33 offset:8 ; 4-byte Folded Reload
+; GFX9-NEXT: buffer_load_dword v41, off, s[0:3], s33 offset:12 ; 4-byte Folded Reload
; GFX9-NEXT: s_mov_b64 exec, s[36:37]
-; GFX9-NEXT: s_addk_i32 s32, 0xfc00
+; GFX9-NEXT: s_addk_i32 s32, 0xf800
; GFX9-NEXT: s_mov_b32 s33, s34
; GFX9-NEXT: s_waitcnt vmcnt(0)
; GFX9-NEXT: s_setpc_b64 s[30:31]
@@ -6038,29 +6580,32 @@
; GFX10-NEXT: s_mov_b32 s33, s32
; GFX10-NEXT: s_or_saveexec_b32 s35, -1
; GFX10-NEXT: buffer_store_dword v40, off, s[0:3], s33 offset:8 ; 4-byte Folded Spill
+; GFX10-NEXT: buffer_store_dword v41, off, s[0:3], s33 offset:12 ; 4-byte Folded Spill
; GFX10-NEXT: s_waitcnt_depctr 0xffe3
; GFX10-NEXT: s_mov_b32 exec_lo, s35
-; GFX10-NEXT: v_writelane_b32 v40, s34, 2
; GFX10-NEXT: v_mov_b32_e32 v0, 3
; GFX10-NEXT: v_mov_b32_e32 v1, 8
-; GFX10-NEXT: s_addk_i32 s32, 0x200
-; GFX10-NEXT: s_getpc_b64 s[34:35]
-; GFX10-NEXT: s_add_u32 s34, s34, external_void_func_byval_struct_i8_i32@rel32@lo+4
-; GFX10-NEXT: s_addc_u32 s35, s35, external_void_func_byval_struct_i8_i32@rel32@hi+12
; GFX10-NEXT: v_writelane_b32 v40, s30, 0
+; GFX10-NEXT: s_addk_i32 s32, 0x400
+; GFX10-NEXT: v_writelane_b32 v41, s34, 0
; GFX10-NEXT: buffer_store_byte v0, off, s[0:3], s33
; GFX10-NEXT: buffer_store_dword v1, off, s[0:3], s33 offset:4
; GFX10-NEXT: v_lshrrev_b32_e64 v0, 5, s33
; GFX10-NEXT: v_writelane_b32 v40, s31, 1
+; GFX10-NEXT: s_getpc_b64 s[34:35]
+; GFX10-NEXT: s_add_u32 s34, s34, external_void_func_byval_struct_i8_i32@rel32@lo+4
+; GFX10-NEXT: s_addc_u32 s35, s35, external_void_func_byval_struct_i8_i32@rel32@hi+12
; GFX10-NEXT: s_swappc_b64 s[30:31], s[34:35]
; GFX10-NEXT: v_readlane_b32 s31, v40, 1
; GFX10-NEXT: v_readlane_b32 s30, v40, 0
-; GFX10-NEXT: v_readlane_b32 s34, v40, 2
+; GFX10-NEXT: v_readlane_b32 s34, v41, 0
; GFX10-NEXT: s_or_saveexec_b32 s35, -1
-; GFX10-NEXT: buffer_load_dword v40, off, s[0:3], s33 offset:8 ; 4-byte Folded Reload
+; GFX10-NEXT: s_clause 0x1
+; GFX10-NEXT: buffer_load_dword v40, off, s[0:3], s33 offset:8
+; GFX10-NEXT: buffer_load_dword v41, off, s[0:3], s33 offset:12
; GFX10-NEXT: s_waitcnt_depctr 0xffe3
; GFX10-NEXT: s_mov_b32 exec_lo, s35
-; GFX10-NEXT: s_addk_i32 s32, 0xfe00
+; GFX10-NEXT: s_addk_i32 s32, 0xfc00
; GFX10-NEXT: s_mov_b32 s33, s34
; GFX10-NEXT: s_waitcnt vmcnt(0)
; GFX10-NEXT: s_setpc_b64 s[30:31]
@@ -6071,29 +6616,33 @@
; GFX11-NEXT: s_mov_b32 s0, s33
; GFX11-NEXT: s_mov_b32 s33, s32
; GFX11-NEXT: s_or_saveexec_b32 s1, -1
-; GFX11-NEXT: scratch_store_b32 off, v40, s33 offset:8 ; 4-byte Folded Spill
+; GFX11-NEXT: s_clause 0x1
+; GFX11-NEXT: scratch_store_b32 off, v40, s33 offset:8
+; GFX11-NEXT: scratch_store_b32 off, v41, s33 offset:12
; GFX11-NEXT: s_mov_b32 exec_lo, s1
-; GFX11-NEXT: v_writelane_b32 v40, s0, 2
; GFX11-NEXT: v_dual_mov_b32 v0, 3 :: v_dual_mov_b32 v1, 8
-; GFX11-NEXT: s_add_i32 s32, s32, 16
-; GFX11-NEXT: s_getpc_b64 s[0:1]
-; GFX11-NEXT: s_add_u32 s0, s0, external_void_func_byval_struct_i8_i32@rel32@lo+4
-; GFX11-NEXT: s_addc_u32 s1, s1, external_void_func_byval_struct_i8_i32@rel32@hi+12
; GFX11-NEXT: v_writelane_b32 v40, s30, 0
+; GFX11-NEXT: s_add_i32 s32, s32, 32
+; GFX11-NEXT: v_writelane_b32 v41, s0, 0
; GFX11-NEXT: s_clause 0x1
; GFX11-NEXT: scratch_store_b8 off, v0, s33
; GFX11-NEXT: scratch_store_b32 off, v1, s33 offset:4
; GFX11-NEXT: v_mov_b32_e32 v0, s33
; GFX11-NEXT: v_writelane_b32 v40, s31, 1
+; GFX11-NEXT: s_getpc_b64 s[0:1]
+; GFX11-NEXT: s_add_u32 s0, s0, external_void_func_byval_struct_i8_i32@rel32@lo+4
+; GFX11-NEXT: s_addc_u32 s1, s1, external_void_func_byval_struct_i8_i32@rel32@hi+12
+; GFX11-NEXT: s_delay_alu instid0(SALU_CYCLE_1)
; GFX11-NEXT: s_swappc_b64 s[30:31], s[0:1]
-; GFX11-NEXT: s_delay_alu instid0(VALU_DEP_1)
; GFX11-NEXT: v_readlane_b32 s31, v40, 1
; GFX11-NEXT: v_readlane_b32 s30, v40, 0
-; GFX11-NEXT: v_readlane_b32 s0, v40, 2
+; GFX11-NEXT: v_readlane_b32 s0, v41, 0
; GFX11-NEXT: s_or_saveexec_b32 s1, -1
-; GFX11-NEXT: scratch_load_b32 v40, off, s33 offset:8 ; 4-byte Folded Reload
+; GFX11-NEXT: s_clause 0x1
+; GFX11-NEXT: scratch_load_b32 v40, off, s33 offset:8
+; GFX11-NEXT: scratch_load_b32 v41, off, s33 offset:12
; GFX11-NEXT: s_mov_b32 exec_lo, s1
-; GFX11-NEXT: s_add_i32 s32, s32, -16
+; GFX11-NEXT: s_addk_i32 s32, 0xffe0
; GFX11-NEXT: s_mov_b32 s33, s0
; GFX11-NEXT: s_waitcnt vmcnt(0)
; GFX11-NEXT: s_setpc_b64 s[30:31]
@@ -6105,29 +6654,32 @@
; GFX10-SCRATCH-NEXT: s_mov_b32 s33, s32
; GFX10-SCRATCH-NEXT: s_or_saveexec_b32 s1, -1
; GFX10-SCRATCH-NEXT: scratch_store_dword off, v40, s33 offset:8 ; 4-byte Folded Spill
+; GFX10-SCRATCH-NEXT: scratch_store_dword off, v41, s33 offset:12 ; 4-byte Folded Spill
; GFX10-SCRATCH-NEXT: s_waitcnt_depctr 0xffe3
; GFX10-SCRATCH-NEXT: s_mov_b32 exec_lo, s1
-; GFX10-SCRATCH-NEXT: v_writelane_b32 v40, s0, 2
; GFX10-SCRATCH-NEXT: v_mov_b32_e32 v0, 3
; GFX10-SCRATCH-NEXT: v_mov_b32_e32 v1, 8
-; GFX10-SCRATCH-NEXT: s_add_i32 s32, s32, 16
-; GFX10-SCRATCH-NEXT: s_getpc_b64 s[0:1]
-; GFX10-SCRATCH-NEXT: s_add_u32 s0, s0, external_void_func_byval_struct_i8_i32@rel32@lo+4
-; GFX10-SCRATCH-NEXT: s_addc_u32 s1, s1, external_void_func_byval_struct_i8_i32@rel32@hi+12
; GFX10-SCRATCH-NEXT: v_writelane_b32 v40, s30, 0
+; GFX10-SCRATCH-NEXT: s_add_i32 s32, s32, 32
+; GFX10-SCRATCH-NEXT: v_writelane_b32 v41, s0, 0
; GFX10-SCRATCH-NEXT: scratch_store_byte off, v0, s33
; GFX10-SCRATCH-NEXT: scratch_store_dword off, v1, s33 offset:4
; GFX10-SCRATCH-NEXT: v_mov_b32_e32 v0, s33
; GFX10-SCRATCH-NEXT: v_writelane_b32 v40, s31, 1
+; GFX10-SCRATCH-NEXT: s_getpc_b64 s[0:1]
+; GFX10-SCRATCH-NEXT: s_add_u32 s0, s0, external_void_func_byval_struct_i8_i32@rel32@lo+4
+; GFX10-SCRATCH-NEXT: s_addc_u32 s1, s1, external_void_func_byval_struct_i8_i32@rel32@hi+12
; GFX10-SCRATCH-NEXT: s_swappc_b64 s[30:31], s[0:1]
; GFX10-SCRATCH-NEXT: v_readlane_b32 s31, v40, 1
; GFX10-SCRATCH-NEXT: v_readlane_b32 s30, v40, 0
-; GFX10-SCRATCH-NEXT: v_readlane_b32 s0, v40, 2
+; GFX10-SCRATCH-NEXT: v_readlane_b32 s0, v41, 0
; GFX10-SCRATCH-NEXT: s_or_saveexec_b32 s1, -1
-; GFX10-SCRATCH-NEXT: scratch_load_dword v40, off, s33 offset:8 ; 4-byte Folded Reload
+; GFX10-SCRATCH-NEXT: s_clause 0x1
+; GFX10-SCRATCH-NEXT: scratch_load_dword v40, off, s33 offset:8
+; GFX10-SCRATCH-NEXT: scratch_load_dword v41, off, s33 offset:12
; GFX10-SCRATCH-NEXT: s_waitcnt_depctr 0xffe3
; GFX10-SCRATCH-NEXT: s_mov_b32 exec_lo, s1
-; GFX10-SCRATCH-NEXT: s_add_i32 s32, s32, -16
+; GFX10-SCRATCH-NEXT: s_addk_i32 s32, 0xffe0
; GFX10-SCRATCH-NEXT: s_mov_b32 s33, s0
; GFX10-SCRATCH-NEXT: s_waitcnt vmcnt(0)
; GFX10-SCRATCH-NEXT: s_setpc_b64 s[30:31]
@@ -6148,17 +6700,18 @@
; GFX9-NEXT: s_mov_b32 s33, s32
; GFX9-NEXT: s_or_saveexec_b64 s[36:37], -1
; GFX9-NEXT: buffer_store_dword v40, off, s[0:3], s33 offset:16 ; 4-byte Folded Spill
+; GFX9-NEXT: buffer_store_dword v41, off, s[0:3], s33 offset:20 ; 4-byte Folded Spill
; GFX9-NEXT: s_mov_b64 exec, s[36:37]
; GFX9-NEXT: v_mov_b32_e32 v0, 3
; GFX9-NEXT: buffer_store_byte v0, off, s[0:3], s33
; GFX9-NEXT: v_mov_b32_e32 v0, 8
-; GFX9-NEXT: v_writelane_b32 v40, s34, 2
; GFX9-NEXT: buffer_store_dword v0, off, s[0:3], s33 offset:4
; GFX9-NEXT: v_lshrrev_b32_e64 v0, 6, s33
; GFX9-NEXT: s_addk_i32 s32, 0x800
; GFX9-NEXT: v_writelane_b32 v40, s30, 0
; GFX9-NEXT: v_add_u32_e32 v0, 8, v0
; GFX9-NEXT: v_lshrrev_b32_e64 v1, 6, s33
+; GFX9-NEXT: v_writelane_b32 v41, s34, 0
; GFX9-NEXT: v_writelane_b32 v40, s31, 1
; GFX9-NEXT: s_getpc_b64 s[34:35]
; GFX9-NEXT: s_add_u32 s34, s34, external_void_func_sret_struct_i8_i32_byval_struct_i8_i32@rel32@lo+4
@@ -6168,7 +6721,7 @@
; GFX9-NEXT: buffer_load_dword v1, off, s[0:3], s33 offset:12
; GFX9-NEXT: v_readlane_b32 s31, v40, 1
; GFX9-NEXT: v_readlane_b32 s30, v40, 0
-; GFX9-NEXT: v_readlane_b32 s34, v40, 2
+; GFX9-NEXT: v_readlane_b32 s34, v41, 0
; GFX9-NEXT: s_waitcnt vmcnt(0)
; GFX9-NEXT: global_store_byte v[0:1], v0, off
; GFX9-NEXT: s_waitcnt vmcnt(0)
@@ -6176,6 +6729,7 @@
; GFX9-NEXT: s_waitcnt vmcnt(0)
; GFX9-NEXT: s_or_saveexec_b64 s[36:37], -1
; GFX9-NEXT: buffer_load_dword v40, off, s[0:3], s33 offset:16 ; 4-byte Folded Reload
+; GFX9-NEXT: buffer_load_dword v41, off, s[0:3], s33 offset:20 ; 4-byte Folded Reload
; GFX9-NEXT: s_mov_b64 exec, s[36:37]
; GFX9-NEXT: s_addk_i32 s32, 0xf800
; GFX9-NEXT: s_mov_b32 s33, s34
@@ -6189,36 +6743,39 @@
; GFX10-NEXT: s_mov_b32 s33, s32
; GFX10-NEXT: s_or_saveexec_b32 s35, -1
; GFX10-NEXT: buffer_store_dword v40, off, s[0:3], s33 offset:16 ; 4-byte Folded Spill
+; GFX10-NEXT: buffer_store_dword v41, off, s[0:3], s33 offset:20 ; 4-byte Folded Spill
; GFX10-NEXT: s_waitcnt_depctr 0xffe3
; GFX10-NEXT: s_mov_b32 exec_lo, s35
; GFX10-NEXT: v_mov_b32_e32 v0, 3
; GFX10-NEXT: v_mov_b32_e32 v1, 8
-; GFX10-NEXT: v_writelane_b32 v40, s34, 2
+; GFX10-NEXT: v_writelane_b32 v40, s30, 0
; GFX10-NEXT: s_addk_i32 s32, 0x400
-; GFX10-NEXT: s_getpc_b64 s[34:35]
-; GFX10-NEXT: s_add_u32 s34, s34, external_void_func_sret_struct_i8_i32_byval_struct_i8_i32@rel32@lo+4
-; GFX10-NEXT: s_addc_u32 s35, s35, external_void_func_sret_struct_i8_i32_byval_struct_i8_i32@rel32@hi+12
+; GFX10-NEXT: v_writelane_b32 v41, s34, 0
; GFX10-NEXT: buffer_store_byte v0, off, s[0:3], s33
; GFX10-NEXT: buffer_store_dword v1, off, s[0:3], s33 offset:4
; GFX10-NEXT: v_lshrrev_b32_e64 v0, 5, s33
-; GFX10-NEXT: v_writelane_b32 v40, s30, 0
; GFX10-NEXT: v_lshrrev_b32_e64 v1, 5, s33
-; GFX10-NEXT: v_add_nc_u32_e32 v0, 8, v0
; GFX10-NEXT: v_writelane_b32 v40, s31, 1
+; GFX10-NEXT: s_getpc_b64 s[34:35]
+; GFX10-NEXT: s_add_u32 s34, s34, external_void_func_sret_struct_i8_i32_byval_struct_i8_i32@rel32@lo+4
+; GFX10-NEXT: s_addc_u32 s35, s35, external_void_func_sret_struct_i8_i32_byval_struct_i8_i32@rel32@hi+12
+; GFX10-NEXT: v_add_nc_u32_e32 v0, 8, v0
; GFX10-NEXT: s_swappc_b64 s[30:31], s[34:35]
; GFX10-NEXT: s_clause 0x1
; GFX10-NEXT: buffer_load_ubyte v0, off, s[0:3], s33 offset:8
; GFX10-NEXT: buffer_load_dword v1, off, s[0:3], s33 offset:12
; GFX10-NEXT: v_readlane_b32 s31, v40, 1
; GFX10-NEXT: v_readlane_b32 s30, v40, 0
-; GFX10-NEXT: v_readlane_b32 s34, v40, 2
+; GFX10-NEXT: v_readlane_b32 s34, v41, 0
; GFX10-NEXT: s_waitcnt vmcnt(0)
; GFX10-NEXT: global_store_byte v[0:1], v0, off
; GFX10-NEXT: s_waitcnt_vscnt null, 0x0
; GFX10-NEXT: global_store_dword v[0:1], v1, off
; GFX10-NEXT: s_waitcnt_vscnt null, 0x0
; GFX10-NEXT: s_or_saveexec_b32 s35, -1
-; GFX10-NEXT: buffer_load_dword v40, off, s[0:3], s33 offset:16 ; 4-byte Folded Reload
+; GFX10-NEXT: s_clause 0x1
+; GFX10-NEXT: buffer_load_dword v40, off, s[0:3], s33 offset:16
+; GFX10-NEXT: buffer_load_dword v41, off, s[0:3], s33 offset:20
; GFX10-NEXT: s_waitcnt_depctr 0xffe3
; GFX10-NEXT: s_mov_b32 exec_lo, s35
; GFX10-NEXT: s_addk_i32 s32, 0xfc00
@@ -6232,11 +6789,13 @@
; GFX11-NEXT: s_mov_b32 s0, s33
; GFX11-NEXT: s_mov_b32 s33, s32
; GFX11-NEXT: s_or_saveexec_b32 s1, -1
-; GFX11-NEXT: scratch_store_b32 off, v40, s33 offset:16 ; 4-byte Folded Spill
+; GFX11-NEXT: s_clause 0x1
+; GFX11-NEXT: scratch_store_b32 off, v40, s33 offset:16
+; GFX11-NEXT: scratch_store_b32 off, v41, s33 offset:20
; GFX11-NEXT: s_mov_b32 exec_lo, s1
-; GFX11-NEXT: v_writelane_b32 v40, s0, 2
; GFX11-NEXT: v_dual_mov_b32 v0, 3 :: v_dual_mov_b32 v1, 8
; GFX11-NEXT: s_add_i32 s32, s32, 32
+; GFX11-NEXT: v_writelane_b32 v41, s0, 0
; GFX11-NEXT: s_getpc_b64 s[0:1]
; GFX11-NEXT: s_add_u32 s0, s0, external_void_func_sret_struct_i8_i32_byval_struct_i8_i32@rel32@lo+4
; GFX11-NEXT: s_addc_u32 s1, s1, external_void_func_sret_struct_i8_i32_byval_struct_i8_i32@rel32@hi+12
@@ -6253,14 +6812,16 @@
; GFX11-NEXT: scratch_load_b32 v1, off, s33 offset:12
; GFX11-NEXT: v_readlane_b32 s31, v40, 1
; GFX11-NEXT: v_readlane_b32 s30, v40, 0
-; GFX11-NEXT: v_readlane_b32 s0, v40, 2
+; GFX11-NEXT: v_readlane_b32 s0, v41, 0
; GFX11-NEXT: s_waitcnt vmcnt(0)
; GFX11-NEXT: global_store_b8 v[0:1], v0, off dlc
; GFX11-NEXT: s_waitcnt_vscnt null, 0x0
; GFX11-NEXT: global_store_b32 v[0:1], v1, off dlc
; GFX11-NEXT: s_waitcnt_vscnt null, 0x0
; GFX11-NEXT: s_or_saveexec_b32 s1, -1
-; GFX11-NEXT: scratch_load_b32 v40, off, s33 offset:16 ; 4-byte Folded Reload
+; GFX11-NEXT: s_clause 0x1
+; GFX11-NEXT: scratch_load_b32 v40, off, s33 offset:16
+; GFX11-NEXT: scratch_load_b32 v41, off, s33 offset:20
; GFX11-NEXT: s_mov_b32 exec_lo, s1
; GFX11-NEXT: s_addk_i32 s32, 0xffe0
; GFX11-NEXT: s_mov_b32 s33, s0
@@ -6274,12 +6835,13 @@
; GFX10-SCRATCH-NEXT: s_mov_b32 s33, s32
; GFX10-SCRATCH-NEXT: s_or_saveexec_b32 s1, -1
; GFX10-SCRATCH-NEXT: scratch_store_dword off, v40, s33 offset:16 ; 4-byte Folded Spill
+; GFX10-SCRATCH-NEXT: scratch_store_dword off, v41, s33 offset:20 ; 4-byte Folded Spill
; GFX10-SCRATCH-NEXT: s_waitcnt_depctr 0xffe3
; GFX10-SCRATCH-NEXT: s_mov_b32 exec_lo, s1
-; GFX10-SCRATCH-NEXT: v_writelane_b32 v40, s0, 2
; GFX10-SCRATCH-NEXT: v_mov_b32_e32 v0, 3
; GFX10-SCRATCH-NEXT: s_add_i32 s32, s32, 32
; GFX10-SCRATCH-NEXT: v_mov_b32_e32 v1, 8
+; GFX10-SCRATCH-NEXT: v_writelane_b32 v41, s0, 0
; GFX10-SCRATCH-NEXT: s_getpc_b64 s[0:1]
; GFX10-SCRATCH-NEXT: s_add_u32 s0, s0, external_void_func_sret_struct_i8_i32_byval_struct_i8_i32@rel32@lo+4
; GFX10-SCRATCH-NEXT: s_addc_u32 s1, s1, external_void_func_sret_struct_i8_i32_byval_struct_i8_i32@rel32@hi+12
@@ -6296,14 +6858,16 @@
; GFX10-SCRATCH-NEXT: scratch_load_dword v1, off, s33 offset:12
; GFX10-SCRATCH-NEXT: v_readlane_b32 s31, v40, 1
; GFX10-SCRATCH-NEXT: v_readlane_b32 s30, v40, 0
-; GFX10-SCRATCH-NEXT: v_readlane_b32 s0, v40, 2
+; GFX10-SCRATCH-NEXT: v_readlane_b32 s0, v41, 0
; GFX10-SCRATCH-NEXT: s_waitcnt vmcnt(0)
; GFX10-SCRATCH-NEXT: global_store_byte v[0:1], v0, off
; GFX10-SCRATCH-NEXT: s_waitcnt_vscnt null, 0x0
; GFX10-SCRATCH-NEXT: global_store_dword v[0:1], v1, off
; GFX10-SCRATCH-NEXT: s_waitcnt_vscnt null, 0x0
; GFX10-SCRATCH-NEXT: s_or_saveexec_b32 s1, -1
-; GFX10-SCRATCH-NEXT: scratch_load_dword v40, off, s33 offset:16 ; 4-byte Folded Reload
+; GFX10-SCRATCH-NEXT: s_clause 0x1
+; GFX10-SCRATCH-NEXT: scratch_load_dword v40, off, s33 offset:16
+; GFX10-SCRATCH-NEXT: scratch_load_dword v41, off, s33 offset:20
; GFX10-SCRATCH-NEXT: s_waitcnt_depctr 0xffe3
; GFX10-SCRATCH-NEXT: s_mov_b32 exec_lo, s1
; GFX10-SCRATCH-NEXT: s_addk_i32 s32, 0xffe0
@@ -6335,8 +6899,9 @@
; GFX9-NEXT: s_mov_b32 s33, s32
; GFX9-NEXT: s_or_saveexec_b64 s[36:37], -1
; GFX9-NEXT: buffer_store_dword v40, off, s[0:3], s33 ; 4-byte Folded Spill
+; GFX9-NEXT: buffer_store_dword v41, off, s[0:3], s33 offset:4 ; 4-byte Folded Spill
; GFX9-NEXT: s_mov_b64 exec, s[36:37]
-; GFX9-NEXT: v_writelane_b32 v40, s34, 2
+; GFX9-NEXT: v_writelane_b32 v41, s34, 0
; GFX9-NEXT: s_load_dwordx2 s[34:35], s[34:35], 0x0
; GFX9-NEXT: v_mov_b32_e32 v0, 0
; GFX9-NEXT: s_addk_i32 s32, 0x400
@@ -6369,9 +6934,10 @@
; GFX9-NEXT: s_swappc_b64 s[30:31], s[34:35]
; GFX9-NEXT: v_readlane_b32 s31, v40, 1
; GFX9-NEXT: v_readlane_b32 s30, v40, 0
-; GFX9-NEXT: v_readlane_b32 s34, v40, 2
+; GFX9-NEXT: v_readlane_b32 s34, v41, 0
; GFX9-NEXT: s_or_saveexec_b64 s[36:37], -1
; GFX9-NEXT: buffer_load_dword v40, off, s[0:3], s33 ; 4-byte Folded Reload
+; GFX9-NEXT: buffer_load_dword v41, off, s[0:3], s33 offset:4 ; 4-byte Folded Reload
; GFX9-NEXT: s_mov_b64 exec, s[36:37]
; GFX9-NEXT: s_addk_i32 s32, 0xfc00
; GFX9-NEXT: s_mov_b32 s33, s34
@@ -6385,13 +6951,14 @@
; GFX10-NEXT: s_mov_b32 s33, s32
; GFX10-NEXT: s_or_saveexec_b32 s35, -1
; GFX10-NEXT: buffer_store_dword v40, off, s[0:3], s33 ; 4-byte Folded Spill
+; GFX10-NEXT: buffer_store_dword v41, off, s[0:3], s33 offset:4 ; 4-byte Folded Spill
; GFX10-NEXT: s_waitcnt_depctr 0xffe3
; GFX10-NEXT: s_mov_b32 exec_lo, s35
-; GFX10-NEXT: v_writelane_b32 v40, s34, 2
+; GFX10-NEXT: v_writelane_b32 v41, s34, 0
; GFX10-NEXT: s_load_dwordx2 s[34:35], s[34:35], 0x0
; GFX10-NEXT: v_mov_b32_e32 v0, 0
-; GFX10-NEXT: s_addk_i32 s32, 0x200
; GFX10-NEXT: v_writelane_b32 v40, s30, 0
+; GFX10-NEXT: s_addk_i32 s32, 0x200
; GFX10-NEXT: v_writelane_b32 v40, s31, 1
; GFX10-NEXT: s_waitcnt lgkmcnt(0)
; GFX10-NEXT: global_load_dwordx4 v[0:3], v0, s[34:35]
@@ -6421,9 +6988,11 @@
; GFX10-NEXT: s_swappc_b64 s[30:31], s[34:35]
; GFX10-NEXT: v_readlane_b32 s31, v40, 1
; GFX10-NEXT: v_readlane_b32 s30, v40, 0
-; GFX10-NEXT: v_readlane_b32 s34, v40, 2
+; GFX10-NEXT: v_readlane_b32 s34, v41, 0
; GFX10-NEXT: s_or_saveexec_b32 s35, -1
-; GFX10-NEXT: buffer_load_dword v40, off, s[0:3], s33 ; 4-byte Folded Reload
+; GFX10-NEXT: s_clause 0x1
+; GFX10-NEXT: buffer_load_dword v40, off, s[0:3], s33
+; GFX10-NEXT: buffer_load_dword v41, off, s[0:3], s33 offset:4
; GFX10-NEXT: s_waitcnt_depctr 0xffe3
; GFX10-NEXT: s_mov_b32 exec_lo, s35
; GFX10-NEXT: s_addk_i32 s32, 0xfe00
@@ -6437,13 +7006,15 @@
; GFX11-NEXT: s_mov_b32 s0, s33
; GFX11-NEXT: s_mov_b32 s33, s32
; GFX11-NEXT: s_or_saveexec_b32 s1, -1
-; GFX11-NEXT: scratch_store_b32 off, v40, s33 ; 4-byte Folded Spill
+; GFX11-NEXT: s_clause 0x1
+; GFX11-NEXT: scratch_store_b32 off, v40, s33
+; GFX11-NEXT: scratch_store_b32 off, v41, s33 offset:4
; GFX11-NEXT: s_mov_b32 exec_lo, s1
-; GFX11-NEXT: v_writelane_b32 v40, s0, 2
+; GFX11-NEXT: v_writelane_b32 v41, s0, 0
; GFX11-NEXT: s_load_b64 s[0:1], s[0:1], 0x0
; GFX11-NEXT: v_mov_b32_e32 v0, 0
-; GFX11-NEXT: s_add_i32 s32, s32, 16
; GFX11-NEXT: v_writelane_b32 v40, s30, 0
+; GFX11-NEXT: s_add_i32 s32, s32, 16
; GFX11-NEXT: v_writelane_b32 v40, s31, 1
; GFX11-NEXT: s_waitcnt lgkmcnt(0)
; GFX11-NEXT: global_load_b128 v[0:3], v0, s[0:1]
@@ -6470,9 +7041,11 @@
; GFX11-NEXT: s_swappc_b64 s[30:31], s[0:1]
; GFX11-NEXT: v_readlane_b32 s31, v40, 1
; GFX11-NEXT: v_readlane_b32 s30, v40, 0
-; GFX11-NEXT: v_readlane_b32 s0, v40, 2
+; GFX11-NEXT: v_readlane_b32 s0, v41, 0
; GFX11-NEXT: s_or_saveexec_b32 s1, -1
-; GFX11-NEXT: scratch_load_b32 v40, off, s33 ; 4-byte Folded Reload
+; GFX11-NEXT: s_clause 0x1
+; GFX11-NEXT: scratch_load_b32 v40, off, s33
+; GFX11-NEXT: scratch_load_b32 v41, off, s33 offset:4
; GFX11-NEXT: s_mov_b32 exec_lo, s1
; GFX11-NEXT: s_add_i32 s32, s32, -16
; GFX11-NEXT: s_mov_b32 s33, s0
@@ -6486,13 +7059,14 @@
; GFX10-SCRATCH-NEXT: s_mov_b32 s33, s32
; GFX10-SCRATCH-NEXT: s_or_saveexec_b32 s1, -1
; GFX10-SCRATCH-NEXT: scratch_store_dword off, v40, s33 ; 4-byte Folded Spill
+; GFX10-SCRATCH-NEXT: scratch_store_dword off, v41, s33 offset:4 ; 4-byte Folded Spill
; GFX10-SCRATCH-NEXT: s_waitcnt_depctr 0xffe3
; GFX10-SCRATCH-NEXT: s_mov_b32 exec_lo, s1
-; GFX10-SCRATCH-NEXT: v_writelane_b32 v40, s0, 2
+; GFX10-SCRATCH-NEXT: v_writelane_b32 v41, s0, 0
; GFX10-SCRATCH-NEXT: s_load_dwordx2 s[0:1], s[0:1], 0x0
; GFX10-SCRATCH-NEXT: v_mov_b32_e32 v0, 0
-; GFX10-SCRATCH-NEXT: s_add_i32 s32, s32, 16
; GFX10-SCRATCH-NEXT: v_writelane_b32 v40, s30, 0
+; GFX10-SCRATCH-NEXT: s_add_i32 s32, s32, 16
; GFX10-SCRATCH-NEXT: v_writelane_b32 v40, s31, 1
; GFX10-SCRATCH-NEXT: s_waitcnt lgkmcnt(0)
; GFX10-SCRATCH-NEXT: global_load_dwordx4 v[0:3], v0, s[0:1]
@@ -6522,9 +7096,11 @@
; GFX10-SCRATCH-NEXT: s_swappc_b64 s[30:31], s[0:1]
; GFX10-SCRATCH-NEXT: v_readlane_b32 s31, v40, 1
; GFX10-SCRATCH-NEXT: v_readlane_b32 s30, v40, 0
-; GFX10-SCRATCH-NEXT: v_readlane_b32 s0, v40, 2
+; GFX10-SCRATCH-NEXT: v_readlane_b32 s0, v41, 0
; GFX10-SCRATCH-NEXT: s_or_saveexec_b32 s1, -1
-; GFX10-SCRATCH-NEXT: scratch_load_dword v40, off, s33 ; 4-byte Folded Reload
+; GFX10-SCRATCH-NEXT: s_clause 0x1
+; GFX10-SCRATCH-NEXT: scratch_load_dword v40, off, s33
+; GFX10-SCRATCH-NEXT: scratch_load_dword v41, off, s33 offset:4
; GFX10-SCRATCH-NEXT: s_waitcnt_depctr 0xffe3
; GFX10-SCRATCH-NEXT: s_mov_b32 exec_lo, s1
; GFX10-SCRATCH-NEXT: s_add_i32 s32, s32, -16
@@ -6541,7 +7117,7 @@
; GFX9-LABEL: tail_call_byval_align16:
; GFX9: ; %bb.0: ; %entry
; GFX9-NEXT: s_waitcnt vmcnt(0) expcnt(0) lgkmcnt(0)
-; GFX9-NEXT: s_mov_b32 s6, s33
+; GFX9-NEXT: s_mov_b32 s8, s33
; GFX9-NEXT: s_mov_b32 s33, s32
; GFX9-NEXT: s_or_saveexec_b64 s[4:5], -1
; GFX9-NEXT: buffer_store_dword v40, off, s[0:3], s33 offset:24 ; 4-byte Folded Spill
@@ -6626,14 +7202,14 @@
; GFX9-NEXT: buffer_load_dword v40, off, s[0:3], s33 offset:24 ; 4-byte Folded Reload
; GFX9-NEXT: s_mov_b64 exec, s[4:5]
; GFX9-NEXT: s_addk_i32 s32, 0xf800
-; GFX9-NEXT: s_mov_b32 s33, s6
+; GFX9-NEXT: s_mov_b32 s33, s8
; GFX9-NEXT: s_waitcnt vmcnt(0)
; GFX9-NEXT: s_setpc_b64 s[30:31]
;
; GFX10-LABEL: tail_call_byval_align16:
; GFX10: ; %bb.0: ; %entry
; GFX10-NEXT: s_waitcnt vmcnt(0) expcnt(0) lgkmcnt(0)
-; GFX10-NEXT: s_mov_b32 s6, s33
+; GFX10-NEXT: s_mov_b32 s7, s33
; GFX10-NEXT: s_mov_b32 s33, s32
; GFX10-NEXT: s_or_saveexec_b32 s4, -1
; GFX10-NEXT: buffer_store_dword v40, off, s[0:3], s33 offset:24 ; 4-byte Folded Spill
@@ -6721,14 +7297,14 @@
; GFX10-NEXT: s_waitcnt_depctr 0xffe3
; GFX10-NEXT: s_mov_b32 exec_lo, s4
; GFX10-NEXT: s_addk_i32 s32, 0xfc00
-; GFX10-NEXT: s_mov_b32 s33, s6
+; GFX10-NEXT: s_mov_b32 s33, s7
; GFX10-NEXT: s_waitcnt vmcnt(0)
; GFX10-NEXT: s_setpc_b64 s[30:31]
;
; GFX11-LABEL: tail_call_byval_align16:
; GFX11: ; %bb.0: ; %entry
; GFX11-NEXT: s_waitcnt vmcnt(0) expcnt(0) lgkmcnt(0)
-; GFX11-NEXT: s_mov_b32 s4, s33
+; GFX11-NEXT: s_mov_b32 s5, s33
; GFX11-NEXT: s_mov_b32 s33, s32
; GFX11-NEXT: s_or_saveexec_b32 s0, -1
; GFX11-NEXT: scratch_store_b32 off, v40, s33 offset:24 ; 4-byte Folded Spill
@@ -6811,14 +7387,14 @@
; GFX11-NEXT: scratch_load_b32 v40, off, s33 offset:24 ; 4-byte Folded Reload
; GFX11-NEXT: s_mov_b32 exec_lo, s0
; GFX11-NEXT: s_addk_i32 s32, 0xffe0
-; GFX11-NEXT: s_mov_b32 s33, s4
+; GFX11-NEXT: s_mov_b32 s33, s5
; GFX11-NEXT: s_waitcnt vmcnt(0)
; GFX11-NEXT: s_setpc_b64 s[30:31]
;
; GFX10-SCRATCH-LABEL: tail_call_byval_align16:
; GFX10-SCRATCH: ; %bb.0: ; %entry
; GFX10-SCRATCH-NEXT: s_waitcnt vmcnt(0) expcnt(0) lgkmcnt(0)
-; GFX10-SCRATCH-NEXT: s_mov_b32 s4, s33
+; GFX10-SCRATCH-NEXT: s_mov_b32 s5, s33
; GFX10-SCRATCH-NEXT: s_mov_b32 s33, s32
; GFX10-SCRATCH-NEXT: s_or_saveexec_b32 s0, -1
; GFX10-SCRATCH-NEXT: scratch_store_dword off, v40, s33 offset:24 ; 4-byte Folded Spill
@@ -6903,7 +7479,7 @@
; GFX10-SCRATCH-NEXT: s_waitcnt_depctr 0xffe3
; GFX10-SCRATCH-NEXT: s_mov_b32 exec_lo, s0
; GFX10-SCRATCH-NEXT: s_addk_i32 s32, 0xffe0
-; GFX10-SCRATCH-NEXT: s_mov_b32 s33, s4
+; GFX10-SCRATCH-NEXT: s_mov_b32 s33, s5
; GFX10-SCRATCH-NEXT: s_waitcnt vmcnt(0)
; GFX10-SCRATCH-NEXT: s_setpc_b64 s[30:31]
entry:
@@ -6921,11 +7497,12 @@
; GFX9-NEXT: s_mov_b32 s33, s32
; GFX9-NEXT: s_or_saveexec_b64 s[36:37], -1
; GFX9-NEXT: buffer_store_dword v40, off, s[0:3], s33 ; 4-byte Folded Spill
+; GFX9-NEXT: buffer_store_dword v41, off, s[0:3], s33 offset:4 ; 4-byte Folded Spill
; GFX9-NEXT: s_mov_b64 exec, s[36:37]
-; GFX9-NEXT: v_writelane_b32 v40, s34, 2
; GFX9-NEXT: s_addk_i32 s32, 0x400
; GFX9-NEXT: v_writelane_b32 v40, s30, 0
; GFX9-NEXT: v_mov_b32_e32 v0, 1
+; GFX9-NEXT: v_writelane_b32 v41, s34, 0
; GFX9-NEXT: v_writelane_b32 v40, s31, 1
; GFX9-NEXT: s_getpc_b64 s[34:35]
; GFX9-NEXT: s_add_u32 s34, s34, external_void_func_i1_inreg@rel32@lo+4
@@ -6934,9 +7511,10 @@
; GFX9-NEXT: s_swappc_b64 s[30:31], s[34:35]
; GFX9-NEXT: v_readlane_b32 s31, v40, 1
; GFX9-NEXT: v_readlane_b32 s30, v40, 0
-; GFX9-NEXT: v_readlane_b32 s34, v40, 2
+; GFX9-NEXT: v_readlane_b32 s34, v41, 0
; GFX9-NEXT: s_or_saveexec_b64 s[36:37], -1
; GFX9-NEXT: buffer_load_dword v40, off, s[0:3], s33 ; 4-byte Folded Reload
+; GFX9-NEXT: buffer_load_dword v41, off, s[0:3], s33 offset:4 ; 4-byte Folded Reload
; GFX9-NEXT: s_mov_b64 exec, s[36:37]
; GFX9-NEXT: s_addk_i32 s32, 0xfc00
; GFX9-NEXT: s_mov_b32 s33, s34
@@ -6950,23 +7528,26 @@
; GFX10-NEXT: s_mov_b32 s33, s32
; GFX10-NEXT: s_or_saveexec_b32 s35, -1
; GFX10-NEXT: buffer_store_dword v40, off, s[0:3], s33 ; 4-byte Folded Spill
+; GFX10-NEXT: buffer_store_dword v41, off, s[0:3], s33 offset:4 ; 4-byte Folded Spill
; GFX10-NEXT: s_waitcnt_depctr 0xffe3
; GFX10-NEXT: s_mov_b32 exec_lo, s35
-; GFX10-NEXT: v_writelane_b32 v40, s34, 2
+; GFX10-NEXT: v_writelane_b32 v40, s30, 0
; GFX10-NEXT: v_mov_b32_e32 v0, 1
; GFX10-NEXT: s_addk_i32 s32, 0x200
+; GFX10-NEXT: v_writelane_b32 v41, s34, 0
; GFX10-NEXT: s_getpc_b64 s[34:35]
; GFX10-NEXT: s_add_u32 s34, s34, external_void_func_i1_inreg@rel32@lo+4
; GFX10-NEXT: s_addc_u32 s35, s35, external_void_func_i1_inreg@rel32@hi+12
-; GFX10-NEXT: v_writelane_b32 v40, s30, 0
-; GFX10-NEXT: buffer_store_byte v0, off, s[0:3], s32
; GFX10-NEXT: v_writelane_b32 v40, s31, 1
+; GFX10-NEXT: buffer_store_byte v0, off, s[0:3], s32
; GFX10-NEXT: s_swappc_b64 s[30:31], s[34:35]
; GFX10-NEXT: v_readlane_b32 s31, v40, 1
; GFX10-NEXT: v_readlane_b32 s30, v40, 0
-; GFX10-NEXT: v_readlane_b32 s34, v40, 2
+; GFX10-NEXT: v_readlane_b32 s34, v41, 0
; GFX10-NEXT: s_or_saveexec_b32 s35, -1
-; GFX10-NEXT: buffer_load_dword v40, off, s[0:3], s33 ; 4-byte Folded Reload
+; GFX10-NEXT: s_clause 0x1
+; GFX10-NEXT: buffer_load_dword v40, off, s[0:3], s33
+; GFX10-NEXT: buffer_load_dword v41, off, s[0:3], s33 offset:4
; GFX10-NEXT: s_waitcnt_depctr 0xffe3
; GFX10-NEXT: s_mov_b32 exec_lo, s35
; GFX10-NEXT: s_addk_i32 s32, 0xfe00
@@ -6980,24 +7561,27 @@
; GFX11-NEXT: s_mov_b32 s0, s33
; GFX11-NEXT: s_mov_b32 s33, s32
; GFX11-NEXT: s_or_saveexec_b32 s1, -1
-; GFX11-NEXT: scratch_store_b32 off, v40, s33 ; 4-byte Folded Spill
+; GFX11-NEXT: s_clause 0x1
+; GFX11-NEXT: scratch_store_b32 off, v40, s33
+; GFX11-NEXT: scratch_store_b32 off, v41, s33 offset:4
; GFX11-NEXT: s_mov_b32 exec_lo, s1
-; GFX11-NEXT: v_writelane_b32 v40, s0, 2
+; GFX11-NEXT: v_writelane_b32 v40, s30, 0
; GFX11-NEXT: v_mov_b32_e32 v0, 1
; GFX11-NEXT: s_add_i32 s32, s32, 16
+; GFX11-NEXT: v_writelane_b32 v41, s0, 0
; GFX11-NEXT: s_getpc_b64 s[0:1]
; GFX11-NEXT: s_add_u32 s0, s0, external_void_func_i1_inreg@rel32@lo+4
; GFX11-NEXT: s_addc_u32 s1, s1, external_void_func_i1_inreg@rel32@hi+12
-; GFX11-NEXT: v_writelane_b32 v40, s30, 0
-; GFX11-NEXT: scratch_store_b8 off, v0, s32
; GFX11-NEXT: v_writelane_b32 v40, s31, 1
+; GFX11-NEXT: scratch_store_b8 off, v0, s32
; GFX11-NEXT: s_swappc_b64 s[30:31], s[0:1]
-; GFX11-NEXT: s_delay_alu instid0(VALU_DEP_1)
; GFX11-NEXT: v_readlane_b32 s31, v40, 1
; GFX11-NEXT: v_readlane_b32 s30, v40, 0
-; GFX11-NEXT: v_readlane_b32 s0, v40, 2
+; GFX11-NEXT: v_readlane_b32 s0, v41, 0
; GFX11-NEXT: s_or_saveexec_b32 s1, -1
-; GFX11-NEXT: scratch_load_b32 v40, off, s33 ; 4-byte Folded Reload
+; GFX11-NEXT: s_clause 0x1
+; GFX11-NEXT: scratch_load_b32 v40, off, s33
+; GFX11-NEXT: scratch_load_b32 v41, off, s33 offset:4
; GFX11-NEXT: s_mov_b32 exec_lo, s1
; GFX11-NEXT: s_add_i32 s32, s32, -16
; GFX11-NEXT: s_mov_b32 s33, s0
@@ -7011,23 +7595,26 @@
; GFX10-SCRATCH-NEXT: s_mov_b32 s33, s32
; GFX10-SCRATCH-NEXT: s_or_saveexec_b32 s1, -1
; GFX10-SCRATCH-NEXT: scratch_store_dword off, v40, s33 ; 4-byte Folded Spill
+; GFX10-SCRATCH-NEXT: scratch_store_dword off, v41, s33 offset:4 ; 4-byte Folded Spill
; GFX10-SCRATCH-NEXT: s_waitcnt_depctr 0xffe3
; GFX10-SCRATCH-NEXT: s_mov_b32 exec_lo, s1
-; GFX10-SCRATCH-NEXT: v_writelane_b32 v40, s0, 2
+; GFX10-SCRATCH-NEXT: v_writelane_b32 v40, s30, 0
; GFX10-SCRATCH-NEXT: v_mov_b32_e32 v0, 1
; GFX10-SCRATCH-NEXT: s_add_i32 s32, s32, 16
+; GFX10-SCRATCH-NEXT: v_writelane_b32 v41, s0, 0
; GFX10-SCRATCH-NEXT: s_getpc_b64 s[0:1]
; GFX10-SCRATCH-NEXT: s_add_u32 s0, s0, external_void_func_i1_inreg@rel32@lo+4
; GFX10-SCRATCH-NEXT: s_addc_u32 s1, s1, external_void_func_i1_inreg@rel32@hi+12
-; GFX10-SCRATCH-NEXT: v_writelane_b32 v40, s30, 0
-; GFX10-SCRATCH-NEXT: scratch_store_byte off, v0, s32
; GFX10-SCRATCH-NEXT: v_writelane_b32 v40, s31, 1
+; GFX10-SCRATCH-NEXT: scratch_store_byte off, v0, s32
; GFX10-SCRATCH-NEXT: s_swappc_b64 s[30:31], s[0:1]
; GFX10-SCRATCH-NEXT: v_readlane_b32 s31, v40, 1
; GFX10-SCRATCH-NEXT: v_readlane_b32 s30, v40, 0
-; GFX10-SCRATCH-NEXT: v_readlane_b32 s0, v40, 2
+; GFX10-SCRATCH-NEXT: v_readlane_b32 s0, v41, 0
; GFX10-SCRATCH-NEXT: s_or_saveexec_b32 s1, -1
-; GFX10-SCRATCH-NEXT: scratch_load_dword v40, off, s33 ; 4-byte Folded Reload
+; GFX10-SCRATCH-NEXT: s_clause 0x1
+; GFX10-SCRATCH-NEXT: scratch_load_dword v40, off, s33
+; GFX10-SCRATCH-NEXT: scratch_load_dword v41, off, s33 offset:4
; GFX10-SCRATCH-NEXT: s_waitcnt_depctr 0xffe3
; GFX10-SCRATCH-NEXT: s_mov_b32 exec_lo, s1
; GFX10-SCRATCH-NEXT: s_add_i32 s32, s32, -16
@@ -7046,12 +7633,13 @@
; GFX9-NEXT: s_mov_b32 s33, s32
; GFX9-NEXT: s_or_saveexec_b64 s[36:37], -1
; GFX9-NEXT: buffer_store_dword v40, off, s[0:3], s33 ; 4-byte Folded Spill
+; GFX9-NEXT: buffer_store_dword v41, off, s[0:3], s33 offset:4 ; 4-byte Folded Spill
; GFX9-NEXT: s_mov_b64 exec, s[36:37]
-; GFX9-NEXT: v_writelane_b32 v40, s34, 3
; GFX9-NEXT: v_writelane_b32 v40, s4, 0
; GFX9-NEXT: s_addk_i32 s32, 0x400
; GFX9-NEXT: v_writelane_b32 v40, s30, 1
; GFX9-NEXT: s_movk_i32 s4, 0x7b
+; GFX9-NEXT: v_writelane_b32 v41, s34, 0
; GFX9-NEXT: v_writelane_b32 v40, s31, 2
; GFX9-NEXT: s_getpc_b64 s[34:35]
; GFX9-NEXT: s_add_u32 s34, s34, external_void_func_i8_inreg@rel32@lo+4
@@ -7060,9 +7648,10 @@
; GFX9-NEXT: v_readlane_b32 s31, v40, 2
; GFX9-NEXT: v_readlane_b32 s30, v40, 1
; GFX9-NEXT: v_readlane_b32 s4, v40, 0
-; GFX9-NEXT: v_readlane_b32 s34, v40, 3
+; GFX9-NEXT: v_readlane_b32 s34, v41, 0
; GFX9-NEXT: s_or_saveexec_b64 s[36:37], -1
; GFX9-NEXT: buffer_load_dword v40, off, s[0:3], s33 ; 4-byte Folded Reload
+; GFX9-NEXT: buffer_load_dword v41, off, s[0:3], s33 offset:4 ; 4-byte Folded Reload
; GFX9-NEXT: s_mov_b64 exec, s[36:37]
; GFX9-NEXT: s_addk_i32 s32, 0xfc00
; GFX9-NEXT: s_mov_b32 s33, s34
@@ -7076,24 +7665,27 @@
; GFX10-NEXT: s_mov_b32 s33, s32
; GFX10-NEXT: s_or_saveexec_b32 s35, -1
; GFX10-NEXT: buffer_store_dword v40, off, s[0:3], s33 ; 4-byte Folded Spill
+; GFX10-NEXT: buffer_store_dword v41, off, s[0:3], s33 offset:4 ; 4-byte Folded Spill
; GFX10-NEXT: s_waitcnt_depctr 0xffe3
; GFX10-NEXT: s_mov_b32 exec_lo, s35
-; GFX10-NEXT: v_writelane_b32 v40, s34, 3
+; GFX10-NEXT: v_writelane_b32 v40, s4, 0
+; GFX10-NEXT: s_movk_i32 s4, 0x7b
; GFX10-NEXT: s_addk_i32 s32, 0x200
+; GFX10-NEXT: v_writelane_b32 v41, s34, 0
; GFX10-NEXT: s_getpc_b64 s[34:35]
; GFX10-NEXT: s_add_u32 s34, s34, external_void_func_i8_inreg@rel32@lo+4
; GFX10-NEXT: s_addc_u32 s35, s35, external_void_func_i8_inreg@rel32@hi+12
-; GFX10-NEXT: v_writelane_b32 v40, s4, 0
-; GFX10-NEXT: s_movk_i32 s4, 0x7b
; GFX10-NEXT: v_writelane_b32 v40, s30, 1
; GFX10-NEXT: v_writelane_b32 v40, s31, 2
; GFX10-NEXT: s_swappc_b64 s[30:31], s[34:35]
; GFX10-NEXT: v_readlane_b32 s31, v40, 2
; GFX10-NEXT: v_readlane_b32 s30, v40, 1
; GFX10-NEXT: v_readlane_b32 s4, v40, 0
-; GFX10-NEXT: v_readlane_b32 s34, v40, 3
+; GFX10-NEXT: v_readlane_b32 s34, v41, 0
; GFX10-NEXT: s_or_saveexec_b32 s35, -1
-; GFX10-NEXT: buffer_load_dword v40, off, s[0:3], s33 ; 4-byte Folded Reload
+; GFX10-NEXT: s_clause 0x1
+; GFX10-NEXT: buffer_load_dword v40, off, s[0:3], s33
+; GFX10-NEXT: buffer_load_dword v41, off, s[0:3], s33 offset:4
; GFX10-NEXT: s_waitcnt_depctr 0xffe3
; GFX10-NEXT: s_mov_b32 exec_lo, s35
; GFX10-NEXT: s_addk_i32 s32, 0xfe00
@@ -7107,15 +7699,17 @@
; GFX11-NEXT: s_mov_b32 s0, s33
; GFX11-NEXT: s_mov_b32 s33, s32
; GFX11-NEXT: s_or_saveexec_b32 s1, -1
-; GFX11-NEXT: scratch_store_b32 off, v40, s33 ; 4-byte Folded Spill
+; GFX11-NEXT: s_clause 0x1
+; GFX11-NEXT: scratch_store_b32 off, v40, s33
+; GFX11-NEXT: scratch_store_b32 off, v41, s33 offset:4
; GFX11-NEXT: s_mov_b32 exec_lo, s1
-; GFX11-NEXT: v_writelane_b32 v40, s0, 3
+; GFX11-NEXT: v_writelane_b32 v40, s4, 0
+; GFX11-NEXT: s_movk_i32 s4, 0x7b
; GFX11-NEXT: s_add_i32 s32, s32, 16
+; GFX11-NEXT: v_writelane_b32 v41, s0, 0
; GFX11-NEXT: s_getpc_b64 s[0:1]
; GFX11-NEXT: s_add_u32 s0, s0, external_void_func_i8_inreg@rel32@lo+4
; GFX11-NEXT: s_addc_u32 s1, s1, external_void_func_i8_inreg@rel32@hi+12
-; GFX11-NEXT: v_writelane_b32 v40, s4, 0
-; GFX11-NEXT: s_movk_i32 s4, 0x7b
; GFX11-NEXT: v_writelane_b32 v40, s30, 1
; GFX11-NEXT: v_writelane_b32 v40, s31, 2
; GFX11-NEXT: s_swappc_b64 s[30:31], s[0:1]
@@ -7123,9 +7717,11 @@
; GFX11-NEXT: v_readlane_b32 s31, v40, 2
; GFX11-NEXT: v_readlane_b32 s30, v40, 1
; GFX11-NEXT: v_readlane_b32 s4, v40, 0
-; GFX11-NEXT: v_readlane_b32 s0, v40, 3
+; GFX11-NEXT: v_readlane_b32 s0, v41, 0
; GFX11-NEXT: s_or_saveexec_b32 s1, -1
-; GFX11-NEXT: scratch_load_b32 v40, off, s33 ; 4-byte Folded Reload
+; GFX11-NEXT: s_clause 0x1
+; GFX11-NEXT: scratch_load_b32 v40, off, s33
+; GFX11-NEXT: scratch_load_b32 v41, off, s33 offset:4
; GFX11-NEXT: s_mov_b32 exec_lo, s1
; GFX11-NEXT: s_add_i32 s32, s32, -16
; GFX11-NEXT: s_mov_b32 s33, s0
@@ -7139,24 +7735,27 @@
; GFX10-SCRATCH-NEXT: s_mov_b32 s33, s32
; GFX10-SCRATCH-NEXT: s_or_saveexec_b32 s1, -1
; GFX10-SCRATCH-NEXT: scratch_store_dword off, v40, s33 ; 4-byte Folded Spill
+; GFX10-SCRATCH-NEXT: scratch_store_dword off, v41, s33 offset:4 ; 4-byte Folded Spill
; GFX10-SCRATCH-NEXT: s_waitcnt_depctr 0xffe3
; GFX10-SCRATCH-NEXT: s_mov_b32 exec_lo, s1
-; GFX10-SCRATCH-NEXT: v_writelane_b32 v40, s0, 3
+; GFX10-SCRATCH-NEXT: v_writelane_b32 v40, s4, 0
+; GFX10-SCRATCH-NEXT: s_movk_i32 s4, 0x7b
; GFX10-SCRATCH-NEXT: s_add_i32 s32, s32, 16
+; GFX10-SCRATCH-NEXT: v_writelane_b32 v41, s0, 0
; GFX10-SCRATCH-NEXT: s_getpc_b64 s[0:1]
; GFX10-SCRATCH-NEXT: s_add_u32 s0, s0, external_void_func_i8_inreg@rel32@lo+4
; GFX10-SCRATCH-NEXT: s_addc_u32 s1, s1, external_void_func_i8_inreg@rel32@hi+12
-; GFX10-SCRATCH-NEXT: v_writelane_b32 v40, s4, 0
-; GFX10-SCRATCH-NEXT: s_movk_i32 s4, 0x7b
; GFX10-SCRATCH-NEXT: v_writelane_b32 v40, s30, 1
; GFX10-SCRATCH-NEXT: v_writelane_b32 v40, s31, 2
; GFX10-SCRATCH-NEXT: s_swappc_b64 s[30:31], s[0:1]
; GFX10-SCRATCH-NEXT: v_readlane_b32 s31, v40, 2
; GFX10-SCRATCH-NEXT: v_readlane_b32 s30, v40, 1
; GFX10-SCRATCH-NEXT: v_readlane_b32 s4, v40, 0
-; GFX10-SCRATCH-NEXT: v_readlane_b32 s0, v40, 3
+; GFX10-SCRATCH-NEXT: v_readlane_b32 s0, v41, 0
; GFX10-SCRATCH-NEXT: s_or_saveexec_b32 s1, -1
-; GFX10-SCRATCH-NEXT: scratch_load_dword v40, off, s33 ; 4-byte Folded Reload
+; GFX10-SCRATCH-NEXT: s_clause 0x1
+; GFX10-SCRATCH-NEXT: scratch_load_dword v40, off, s33
+; GFX10-SCRATCH-NEXT: scratch_load_dword v41, off, s33 offset:4
; GFX10-SCRATCH-NEXT: s_waitcnt_depctr 0xffe3
; GFX10-SCRATCH-NEXT: s_mov_b32 exec_lo, s1
; GFX10-SCRATCH-NEXT: s_add_i32 s32, s32, -16
@@ -7175,12 +7774,13 @@
; GFX9-NEXT: s_mov_b32 s33, s32
; GFX9-NEXT: s_or_saveexec_b64 s[36:37], -1
; GFX9-NEXT: buffer_store_dword v40, off, s[0:3], s33 ; 4-byte Folded Spill
+; GFX9-NEXT: buffer_store_dword v41, off, s[0:3], s33 offset:4 ; 4-byte Folded Spill
; GFX9-NEXT: s_mov_b64 exec, s[36:37]
-; GFX9-NEXT: v_writelane_b32 v40, s34, 3
; GFX9-NEXT: v_writelane_b32 v40, s4, 0
; GFX9-NEXT: s_addk_i32 s32, 0x400
; GFX9-NEXT: v_writelane_b32 v40, s30, 1
; GFX9-NEXT: s_movk_i32 s4, 0x7b
+; GFX9-NEXT: v_writelane_b32 v41, s34, 0
; GFX9-NEXT: v_writelane_b32 v40, s31, 2
; GFX9-NEXT: s_getpc_b64 s[34:35]
; GFX9-NEXT: s_add_u32 s34, s34, external_void_func_i16_inreg@rel32@lo+4
@@ -7189,9 +7789,10 @@
; GFX9-NEXT: v_readlane_b32 s31, v40, 2
; GFX9-NEXT: v_readlane_b32 s30, v40, 1
; GFX9-NEXT: v_readlane_b32 s4, v40, 0
-; GFX9-NEXT: v_readlane_b32 s34, v40, 3
+; GFX9-NEXT: v_readlane_b32 s34, v41, 0
; GFX9-NEXT: s_or_saveexec_b64 s[36:37], -1
; GFX9-NEXT: buffer_load_dword v40, off, s[0:3], s33 ; 4-byte Folded Reload
+; GFX9-NEXT: buffer_load_dword v41, off, s[0:3], s33 offset:4 ; 4-byte Folded Reload
; GFX9-NEXT: s_mov_b64 exec, s[36:37]
; GFX9-NEXT: s_addk_i32 s32, 0xfc00
; GFX9-NEXT: s_mov_b32 s33, s34
@@ -7205,24 +7806,27 @@
; GFX10-NEXT: s_mov_b32 s33, s32
; GFX10-NEXT: s_or_saveexec_b32 s35, -1
; GFX10-NEXT: buffer_store_dword v40, off, s[0:3], s33 ; 4-byte Folded Spill
+; GFX10-NEXT: buffer_store_dword v41, off, s[0:3], s33 offset:4 ; 4-byte Folded Spill
; GFX10-NEXT: s_waitcnt_depctr 0xffe3
; GFX10-NEXT: s_mov_b32 exec_lo, s35
-; GFX10-NEXT: v_writelane_b32 v40, s34, 3
+; GFX10-NEXT: v_writelane_b32 v40, s4, 0
+; GFX10-NEXT: s_movk_i32 s4, 0x7b
; GFX10-NEXT: s_addk_i32 s32, 0x200
+; GFX10-NEXT: v_writelane_b32 v41, s34, 0
; GFX10-NEXT: s_getpc_b64 s[34:35]
; GFX10-NEXT: s_add_u32 s34, s34, external_void_func_i16_inreg@rel32@lo+4
; GFX10-NEXT: s_addc_u32 s35, s35, external_void_func_i16_inreg@rel32@hi+12
-; GFX10-NEXT: v_writelane_b32 v40, s4, 0
-; GFX10-NEXT: s_movk_i32 s4, 0x7b
; GFX10-NEXT: v_writelane_b32 v40, s30, 1
; GFX10-NEXT: v_writelane_b32 v40, s31, 2
; GFX10-NEXT: s_swappc_b64 s[30:31], s[34:35]
; GFX10-NEXT: v_readlane_b32 s31, v40, 2
; GFX10-NEXT: v_readlane_b32 s30, v40, 1
; GFX10-NEXT: v_readlane_b32 s4, v40, 0
-; GFX10-NEXT: v_readlane_b32 s34, v40, 3
+; GFX10-NEXT: v_readlane_b32 s34, v41, 0
; GFX10-NEXT: s_or_saveexec_b32 s35, -1
-; GFX10-NEXT: buffer_load_dword v40, off, s[0:3], s33 ; 4-byte Folded Reload
+; GFX10-NEXT: s_clause 0x1
+; GFX10-NEXT: buffer_load_dword v40, off, s[0:3], s33
+; GFX10-NEXT: buffer_load_dword v41, off, s[0:3], s33 offset:4
; GFX10-NEXT: s_waitcnt_depctr 0xffe3
; GFX10-NEXT: s_mov_b32 exec_lo, s35
; GFX10-NEXT: s_addk_i32 s32, 0xfe00
@@ -7236,15 +7840,17 @@
; GFX11-NEXT: s_mov_b32 s0, s33
; GFX11-NEXT: s_mov_b32 s33, s32
; GFX11-NEXT: s_or_saveexec_b32 s1, -1
-; GFX11-NEXT: scratch_store_b32 off, v40, s33 ; 4-byte Folded Spill
+; GFX11-NEXT: s_clause 0x1
+; GFX11-NEXT: scratch_store_b32 off, v40, s33
+; GFX11-NEXT: scratch_store_b32 off, v41, s33 offset:4
; GFX11-NEXT: s_mov_b32 exec_lo, s1
-; GFX11-NEXT: v_writelane_b32 v40, s0, 3
+; GFX11-NEXT: v_writelane_b32 v40, s4, 0
+; GFX11-NEXT: s_movk_i32 s4, 0x7b
; GFX11-NEXT: s_add_i32 s32, s32, 16
+; GFX11-NEXT: v_writelane_b32 v41, s0, 0
; GFX11-NEXT: s_getpc_b64 s[0:1]
; GFX11-NEXT: s_add_u32 s0, s0, external_void_func_i16_inreg@rel32@lo+4
; GFX11-NEXT: s_addc_u32 s1, s1, external_void_func_i16_inreg@rel32@hi+12
-; GFX11-NEXT: v_writelane_b32 v40, s4, 0
-; GFX11-NEXT: s_movk_i32 s4, 0x7b
; GFX11-NEXT: v_writelane_b32 v40, s30, 1
; GFX11-NEXT: v_writelane_b32 v40, s31, 2
; GFX11-NEXT: s_swappc_b64 s[30:31], s[0:1]
@@ -7252,9 +7858,11 @@
; GFX11-NEXT: v_readlane_b32 s31, v40, 2
; GFX11-NEXT: v_readlane_b32 s30, v40, 1
; GFX11-NEXT: v_readlane_b32 s4, v40, 0
-; GFX11-NEXT: v_readlane_b32 s0, v40, 3
+; GFX11-NEXT: v_readlane_b32 s0, v41, 0
; GFX11-NEXT: s_or_saveexec_b32 s1, -1
-; GFX11-NEXT: scratch_load_b32 v40, off, s33 ; 4-byte Folded Reload
+; GFX11-NEXT: s_clause 0x1
+; GFX11-NEXT: scratch_load_b32 v40, off, s33
+; GFX11-NEXT: scratch_load_b32 v41, off, s33 offset:4
; GFX11-NEXT: s_mov_b32 exec_lo, s1
; GFX11-NEXT: s_add_i32 s32, s32, -16
; GFX11-NEXT: s_mov_b32 s33, s0
@@ -7268,24 +7876,27 @@
; GFX10-SCRATCH-NEXT: s_mov_b32 s33, s32
; GFX10-SCRATCH-NEXT: s_or_saveexec_b32 s1, -1
; GFX10-SCRATCH-NEXT: scratch_store_dword off, v40, s33 ; 4-byte Folded Spill
+; GFX10-SCRATCH-NEXT: scratch_store_dword off, v41, s33 offset:4 ; 4-byte Folded Spill
; GFX10-SCRATCH-NEXT: s_waitcnt_depctr 0xffe3
; GFX10-SCRATCH-NEXT: s_mov_b32 exec_lo, s1
-; GFX10-SCRATCH-NEXT: v_writelane_b32 v40, s0, 3
+; GFX10-SCRATCH-NEXT: v_writelane_b32 v40, s4, 0
+; GFX10-SCRATCH-NEXT: s_movk_i32 s4, 0x7b
; GFX10-SCRATCH-NEXT: s_add_i32 s32, s32, 16
+; GFX10-SCRATCH-NEXT: v_writelane_b32 v41, s0, 0
; GFX10-SCRATCH-NEXT: s_getpc_b64 s[0:1]
; GFX10-SCRATCH-NEXT: s_add_u32 s0, s0, external_void_func_i16_inreg@rel32@lo+4
; GFX10-SCRATCH-NEXT: s_addc_u32 s1, s1, external_void_func_i16_inreg@rel32@hi+12
-; GFX10-SCRATCH-NEXT: v_writelane_b32 v40, s4, 0
-; GFX10-SCRATCH-NEXT: s_movk_i32 s4, 0x7b
; GFX10-SCRATCH-NEXT: v_writelane_b32 v40, s30, 1
; GFX10-SCRATCH-NEXT: v_writelane_b32 v40, s31, 2
; GFX10-SCRATCH-NEXT: s_swappc_b64 s[30:31], s[0:1]
; GFX10-SCRATCH-NEXT: v_readlane_b32 s31, v40, 2
; GFX10-SCRATCH-NEXT: v_readlane_b32 s30, v40, 1
; GFX10-SCRATCH-NEXT: v_readlane_b32 s4, v40, 0
-; GFX10-SCRATCH-NEXT: v_readlane_b32 s0, v40, 3
+; GFX10-SCRATCH-NEXT: v_readlane_b32 s0, v41, 0
; GFX10-SCRATCH-NEXT: s_or_saveexec_b32 s1, -1
-; GFX10-SCRATCH-NEXT: scratch_load_dword v40, off, s33 ; 4-byte Folded Reload
+; GFX10-SCRATCH-NEXT: s_clause 0x1
+; GFX10-SCRATCH-NEXT: scratch_load_dword v40, off, s33
+; GFX10-SCRATCH-NEXT: scratch_load_dword v41, off, s33 offset:4
; GFX10-SCRATCH-NEXT: s_waitcnt_depctr 0xffe3
; GFX10-SCRATCH-NEXT: s_mov_b32 exec_lo, s1
; GFX10-SCRATCH-NEXT: s_add_i32 s32, s32, -16
@@ -7304,12 +7915,13 @@
; GFX9-NEXT: s_mov_b32 s33, s32
; GFX9-NEXT: s_or_saveexec_b64 s[36:37], -1
; GFX9-NEXT: buffer_store_dword v40, off, s[0:3], s33 ; 4-byte Folded Spill
+; GFX9-NEXT: buffer_store_dword v41, off, s[0:3], s33 offset:4 ; 4-byte Folded Spill
; GFX9-NEXT: s_mov_b64 exec, s[36:37]
-; GFX9-NEXT: v_writelane_b32 v40, s34, 3
; GFX9-NEXT: v_writelane_b32 v40, s4, 0
; GFX9-NEXT: s_addk_i32 s32, 0x400
; GFX9-NEXT: v_writelane_b32 v40, s30, 1
; GFX9-NEXT: s_mov_b32 s4, 42
+; GFX9-NEXT: v_writelane_b32 v41, s34, 0
; GFX9-NEXT: v_writelane_b32 v40, s31, 2
; GFX9-NEXT: s_getpc_b64 s[34:35]
; GFX9-NEXT: s_add_u32 s34, s34, external_void_func_i32_inreg@rel32@lo+4
@@ -7318,9 +7930,10 @@
; GFX9-NEXT: v_readlane_b32 s31, v40, 2
; GFX9-NEXT: v_readlane_b32 s30, v40, 1
; GFX9-NEXT: v_readlane_b32 s4, v40, 0
-; GFX9-NEXT: v_readlane_b32 s34, v40, 3
+; GFX9-NEXT: v_readlane_b32 s34, v41, 0
; GFX9-NEXT: s_or_saveexec_b64 s[36:37], -1
; GFX9-NEXT: buffer_load_dword v40, off, s[0:3], s33 ; 4-byte Folded Reload
+; GFX9-NEXT: buffer_load_dword v41, off, s[0:3], s33 offset:4 ; 4-byte Folded Reload
; GFX9-NEXT: s_mov_b64 exec, s[36:37]
; GFX9-NEXT: s_addk_i32 s32, 0xfc00
; GFX9-NEXT: s_mov_b32 s33, s34
@@ -7334,24 +7947,27 @@
; GFX10-NEXT: s_mov_b32 s33, s32
; GFX10-NEXT: s_or_saveexec_b32 s35, -1
; GFX10-NEXT: buffer_store_dword v40, off, s[0:3], s33 ; 4-byte Folded Spill
+; GFX10-NEXT: buffer_store_dword v41, off, s[0:3], s33 offset:4 ; 4-byte Folded Spill
; GFX10-NEXT: s_waitcnt_depctr 0xffe3
; GFX10-NEXT: s_mov_b32 exec_lo, s35
-; GFX10-NEXT: v_writelane_b32 v40, s34, 3
+; GFX10-NEXT: v_writelane_b32 v40, s4, 0
+; GFX10-NEXT: s_mov_b32 s4, 42
; GFX10-NEXT: s_addk_i32 s32, 0x200
+; GFX10-NEXT: v_writelane_b32 v41, s34, 0
; GFX10-NEXT: s_getpc_b64 s[34:35]
; GFX10-NEXT: s_add_u32 s34, s34, external_void_func_i32_inreg@rel32@lo+4
; GFX10-NEXT: s_addc_u32 s35, s35, external_void_func_i32_inreg@rel32@hi+12
-; GFX10-NEXT: v_writelane_b32 v40, s4, 0
-; GFX10-NEXT: s_mov_b32 s4, 42
; GFX10-NEXT: v_writelane_b32 v40, s30, 1
; GFX10-NEXT: v_writelane_b32 v40, s31, 2
; GFX10-NEXT: s_swappc_b64 s[30:31], s[34:35]
; GFX10-NEXT: v_readlane_b32 s31, v40, 2
; GFX10-NEXT: v_readlane_b32 s30, v40, 1
; GFX10-NEXT: v_readlane_b32 s4, v40, 0
-; GFX10-NEXT: v_readlane_b32 s34, v40, 3
+; GFX10-NEXT: v_readlane_b32 s34, v41, 0
; GFX10-NEXT: s_or_saveexec_b32 s35, -1
-; GFX10-NEXT: buffer_load_dword v40, off, s[0:3], s33 ; 4-byte Folded Reload
+; GFX10-NEXT: s_clause 0x1
+; GFX10-NEXT: buffer_load_dword v40, off, s[0:3], s33
+; GFX10-NEXT: buffer_load_dword v41, off, s[0:3], s33 offset:4
; GFX10-NEXT: s_waitcnt_depctr 0xffe3
; GFX10-NEXT: s_mov_b32 exec_lo, s35
; GFX10-NEXT: s_addk_i32 s32, 0xfe00
@@ -7365,15 +7981,17 @@
; GFX11-NEXT: s_mov_b32 s0, s33
; GFX11-NEXT: s_mov_b32 s33, s32
; GFX11-NEXT: s_or_saveexec_b32 s1, -1
-; GFX11-NEXT: scratch_store_b32 off, v40, s33 ; 4-byte Folded Spill
+; GFX11-NEXT: s_clause 0x1
+; GFX11-NEXT: scratch_store_b32 off, v40, s33
+; GFX11-NEXT: scratch_store_b32 off, v41, s33 offset:4
; GFX11-NEXT: s_mov_b32 exec_lo, s1
-; GFX11-NEXT: v_writelane_b32 v40, s0, 3
+; GFX11-NEXT: v_writelane_b32 v40, s4, 0
+; GFX11-NEXT: s_mov_b32 s4, 42
; GFX11-NEXT: s_add_i32 s32, s32, 16
+; GFX11-NEXT: v_writelane_b32 v41, s0, 0
; GFX11-NEXT: s_getpc_b64 s[0:1]
; GFX11-NEXT: s_add_u32 s0, s0, external_void_func_i32_inreg@rel32@lo+4
; GFX11-NEXT: s_addc_u32 s1, s1, external_void_func_i32_inreg@rel32@hi+12
-; GFX11-NEXT: v_writelane_b32 v40, s4, 0
-; GFX11-NEXT: s_mov_b32 s4, 42
; GFX11-NEXT: v_writelane_b32 v40, s30, 1
; GFX11-NEXT: v_writelane_b32 v40, s31, 2
; GFX11-NEXT: s_swappc_b64 s[30:31], s[0:1]
@@ -7381,9 +7999,11 @@
; GFX11-NEXT: v_readlane_b32 s31, v40, 2
; GFX11-NEXT: v_readlane_b32 s30, v40, 1
; GFX11-NEXT: v_readlane_b32 s4, v40, 0
-; GFX11-NEXT: v_readlane_b32 s0, v40, 3
+; GFX11-NEXT: v_readlane_b32 s0, v41, 0
; GFX11-NEXT: s_or_saveexec_b32 s1, -1
-; GFX11-NEXT: scratch_load_b32 v40, off, s33 ; 4-byte Folded Reload
+; GFX11-NEXT: s_clause 0x1
+; GFX11-NEXT: scratch_load_b32 v40, off, s33
+; GFX11-NEXT: scratch_load_b32 v41, off, s33 offset:4
; GFX11-NEXT: s_mov_b32 exec_lo, s1
; GFX11-NEXT: s_add_i32 s32, s32, -16
; GFX11-NEXT: s_mov_b32 s33, s0
@@ -7397,24 +8017,27 @@
; GFX10-SCRATCH-NEXT: s_mov_b32 s33, s32
; GFX10-SCRATCH-NEXT: s_or_saveexec_b32 s1, -1
; GFX10-SCRATCH-NEXT: scratch_store_dword off, v40, s33 ; 4-byte Folded Spill
+; GFX10-SCRATCH-NEXT: scratch_store_dword off, v41, s33 offset:4 ; 4-byte Folded Spill
; GFX10-SCRATCH-NEXT: s_waitcnt_depctr 0xffe3
; GFX10-SCRATCH-NEXT: s_mov_b32 exec_lo, s1
-; GFX10-SCRATCH-NEXT: v_writelane_b32 v40, s0, 3
+; GFX10-SCRATCH-NEXT: v_writelane_b32 v40, s4, 0
+; GFX10-SCRATCH-NEXT: s_mov_b32 s4, 42
; GFX10-SCRATCH-NEXT: s_add_i32 s32, s32, 16
+; GFX10-SCRATCH-NEXT: v_writelane_b32 v41, s0, 0
; GFX10-SCRATCH-NEXT: s_getpc_b64 s[0:1]
; GFX10-SCRATCH-NEXT: s_add_u32 s0, s0, external_void_func_i32_inreg@rel32@lo+4
; GFX10-SCRATCH-NEXT: s_addc_u32 s1, s1, external_void_func_i32_inreg@rel32@hi+12
-; GFX10-SCRATCH-NEXT: v_writelane_b32 v40, s4, 0
-; GFX10-SCRATCH-NEXT: s_mov_b32 s4, 42
; GFX10-SCRATCH-NEXT: v_writelane_b32 v40, s30, 1
; GFX10-SCRATCH-NEXT: v_writelane_b32 v40, s31, 2
; GFX10-SCRATCH-NEXT: s_swappc_b64 s[30:31], s[0:1]
; GFX10-SCRATCH-NEXT: v_readlane_b32 s31, v40, 2
; GFX10-SCRATCH-NEXT: v_readlane_b32 s30, v40, 1
; GFX10-SCRATCH-NEXT: v_readlane_b32 s4, v40, 0
-; GFX10-SCRATCH-NEXT: v_readlane_b32 s0, v40, 3
+; GFX10-SCRATCH-NEXT: v_readlane_b32 s0, v41, 0
; GFX10-SCRATCH-NEXT: s_or_saveexec_b32 s1, -1
-; GFX10-SCRATCH-NEXT: scratch_load_dword v40, off, s33 ; 4-byte Folded Reload
+; GFX10-SCRATCH-NEXT: s_clause 0x1
+; GFX10-SCRATCH-NEXT: scratch_load_dword v40, off, s33
+; GFX10-SCRATCH-NEXT: scratch_load_dword v41, off, s33 offset:4
; GFX10-SCRATCH-NEXT: s_waitcnt_depctr 0xffe3
; GFX10-SCRATCH-NEXT: s_mov_b32 exec_lo, s1
; GFX10-SCRATCH-NEXT: s_add_i32 s32, s32, -16
@@ -7433,14 +8056,15 @@
; GFX9-NEXT: s_mov_b32 s33, s32
; GFX9-NEXT: s_or_saveexec_b64 s[36:37], -1
; GFX9-NEXT: buffer_store_dword v40, off, s[0:3], s33 ; 4-byte Folded Spill
+; GFX9-NEXT: buffer_store_dword v41, off, s[0:3], s33 offset:4 ; 4-byte Folded Spill
; GFX9-NEXT: s_mov_b64 exec, s[36:37]
-; GFX9-NEXT: v_writelane_b32 v40, s34, 4
; GFX9-NEXT: v_writelane_b32 v40, s4, 0
; GFX9-NEXT: v_writelane_b32 v40, s5, 1
; GFX9-NEXT: s_addk_i32 s32, 0x400
; GFX9-NEXT: v_writelane_b32 v40, s30, 2
; GFX9-NEXT: s_movk_i32 s4, 0x7b
; GFX9-NEXT: s_mov_b32 s5, 0
+; GFX9-NEXT: v_writelane_b32 v41, s34, 0
; GFX9-NEXT: v_writelane_b32 v40, s31, 3
; GFX9-NEXT: s_getpc_b64 s[34:35]
; GFX9-NEXT: s_add_u32 s34, s34, external_void_func_i64_inreg@rel32@lo+4
@@ -7450,9 +8074,10 @@
; GFX9-NEXT: v_readlane_b32 s30, v40, 2
; GFX9-NEXT: v_readlane_b32 s5, v40, 1
; GFX9-NEXT: v_readlane_b32 s4, v40, 0
-; GFX9-NEXT: v_readlane_b32 s34, v40, 4
+; GFX9-NEXT: v_readlane_b32 s34, v41, 0
; GFX9-NEXT: s_or_saveexec_b64 s[36:37], -1
; GFX9-NEXT: buffer_load_dword v40, off, s[0:3], s33 ; 4-byte Folded Reload
+; GFX9-NEXT: buffer_load_dword v41, off, s[0:3], s33 offset:4 ; 4-byte Folded Reload
; GFX9-NEXT: s_mov_b64 exec, s[36:37]
; GFX9-NEXT: s_addk_i32 s32, 0xfc00
; GFX9-NEXT: s_mov_b32 s33, s34
@@ -7466,15 +8091,16 @@
; GFX10-NEXT: s_mov_b32 s33, s32
; GFX10-NEXT: s_or_saveexec_b32 s35, -1
; GFX10-NEXT: buffer_store_dword v40, off, s[0:3], s33 ; 4-byte Folded Spill
+; GFX10-NEXT: buffer_store_dword v41, off, s[0:3], s33 offset:4 ; 4-byte Folded Spill
; GFX10-NEXT: s_waitcnt_depctr 0xffe3
; GFX10-NEXT: s_mov_b32 exec_lo, s35
-; GFX10-NEXT: v_writelane_b32 v40, s34, 4
+; GFX10-NEXT: v_writelane_b32 v40, s4, 0
+; GFX10-NEXT: s_movk_i32 s4, 0x7b
; GFX10-NEXT: s_addk_i32 s32, 0x200
+; GFX10-NEXT: v_writelane_b32 v41, s34, 0
; GFX10-NEXT: s_getpc_b64 s[34:35]
; GFX10-NEXT: s_add_u32 s34, s34, external_void_func_i64_inreg@rel32@lo+4
; GFX10-NEXT: s_addc_u32 s35, s35, external_void_func_i64_inreg@rel32@hi+12
-; GFX10-NEXT: v_writelane_b32 v40, s4, 0
-; GFX10-NEXT: s_movk_i32 s4, 0x7b
; GFX10-NEXT: v_writelane_b32 v40, s5, 1
; GFX10-NEXT: s_mov_b32 s5, 0
; GFX10-NEXT: v_writelane_b32 v40, s30, 2
@@ -7484,9 +8110,11 @@
; GFX10-NEXT: v_readlane_b32 s30, v40, 2
; GFX10-NEXT: v_readlane_b32 s5, v40, 1
; GFX10-NEXT: v_readlane_b32 s4, v40, 0
-; GFX10-NEXT: v_readlane_b32 s34, v40, 4
+; GFX10-NEXT: v_readlane_b32 s34, v41, 0
; GFX10-NEXT: s_or_saveexec_b32 s35, -1
-; GFX10-NEXT: buffer_load_dword v40, off, s[0:3], s33 ; 4-byte Folded Reload
+; GFX10-NEXT: s_clause 0x1
+; GFX10-NEXT: buffer_load_dword v40, off, s[0:3], s33
+; GFX10-NEXT: buffer_load_dword v41, off, s[0:3], s33 offset:4
; GFX10-NEXT: s_waitcnt_depctr 0xffe3
; GFX10-NEXT: s_mov_b32 exec_lo, s35
; GFX10-NEXT: s_addk_i32 s32, 0xfe00
@@ -7500,15 +8128,17 @@
; GFX11-NEXT: s_mov_b32 s0, s33
; GFX11-NEXT: s_mov_b32 s33, s32
; GFX11-NEXT: s_or_saveexec_b32 s1, -1
-; GFX11-NEXT: scratch_store_b32 off, v40, s33 ; 4-byte Folded Spill
+; GFX11-NEXT: s_clause 0x1
+; GFX11-NEXT: scratch_store_b32 off, v40, s33
+; GFX11-NEXT: scratch_store_b32 off, v41, s33 offset:4
; GFX11-NEXT: s_mov_b32 exec_lo, s1
-; GFX11-NEXT: v_writelane_b32 v40, s0, 4
+; GFX11-NEXT: v_writelane_b32 v40, s4, 0
+; GFX11-NEXT: s_movk_i32 s4, 0x7b
; GFX11-NEXT: s_add_i32 s32, s32, 16
+; GFX11-NEXT: v_writelane_b32 v41, s0, 0
; GFX11-NEXT: s_getpc_b64 s[0:1]
; GFX11-NEXT: s_add_u32 s0, s0, external_void_func_i64_inreg@rel32@lo+4
; GFX11-NEXT: s_addc_u32 s1, s1, external_void_func_i64_inreg@rel32@hi+12
-; GFX11-NEXT: v_writelane_b32 v40, s4, 0
-; GFX11-NEXT: s_movk_i32 s4, 0x7b
; GFX11-NEXT: v_writelane_b32 v40, s5, 1
; GFX11-NEXT: s_mov_b32 s5, 0
; GFX11-NEXT: v_writelane_b32 v40, s30, 2
@@ -7519,9 +8149,11 @@
; GFX11-NEXT: v_readlane_b32 s30, v40, 2
; GFX11-NEXT: v_readlane_b32 s5, v40, 1
; GFX11-NEXT: v_readlane_b32 s4, v40, 0
-; GFX11-NEXT: v_readlane_b32 s0, v40, 4
+; GFX11-NEXT: v_readlane_b32 s0, v41, 0
; GFX11-NEXT: s_or_saveexec_b32 s1, -1
-; GFX11-NEXT: scratch_load_b32 v40, off, s33 ; 4-byte Folded Reload
+; GFX11-NEXT: s_clause 0x1
+; GFX11-NEXT: scratch_load_b32 v40, off, s33
+; GFX11-NEXT: scratch_load_b32 v41, off, s33 offset:4
; GFX11-NEXT: s_mov_b32 exec_lo, s1
; GFX11-NEXT: s_add_i32 s32, s32, -16
; GFX11-NEXT: s_mov_b32 s33, s0
@@ -7535,15 +8167,16 @@
; GFX10-SCRATCH-NEXT: s_mov_b32 s33, s32
; GFX10-SCRATCH-NEXT: s_or_saveexec_b32 s1, -1
; GFX10-SCRATCH-NEXT: scratch_store_dword off, v40, s33 ; 4-byte Folded Spill
+; GFX10-SCRATCH-NEXT: scratch_store_dword off, v41, s33 offset:4 ; 4-byte Folded Spill
; GFX10-SCRATCH-NEXT: s_waitcnt_depctr 0xffe3
; GFX10-SCRATCH-NEXT: s_mov_b32 exec_lo, s1
-; GFX10-SCRATCH-NEXT: v_writelane_b32 v40, s0, 4
+; GFX10-SCRATCH-NEXT: v_writelane_b32 v40, s4, 0
+; GFX10-SCRATCH-NEXT: s_movk_i32 s4, 0x7b
; GFX10-SCRATCH-NEXT: s_add_i32 s32, s32, 16
+; GFX10-SCRATCH-NEXT: v_writelane_b32 v41, s0, 0
; GFX10-SCRATCH-NEXT: s_getpc_b64 s[0:1]
; GFX10-SCRATCH-NEXT: s_add_u32 s0, s0, external_void_func_i64_inreg@rel32@lo+4
; GFX10-SCRATCH-NEXT: s_addc_u32 s1, s1, external_void_func_i64_inreg@rel32@hi+12
-; GFX10-SCRATCH-NEXT: v_writelane_b32 v40, s4, 0
-; GFX10-SCRATCH-NEXT: s_movk_i32 s4, 0x7b
; GFX10-SCRATCH-NEXT: v_writelane_b32 v40, s5, 1
; GFX10-SCRATCH-NEXT: s_mov_b32 s5, 0
; GFX10-SCRATCH-NEXT: v_writelane_b32 v40, s30, 2
@@ -7553,9 +8186,11 @@
; GFX10-SCRATCH-NEXT: v_readlane_b32 s30, v40, 2
; GFX10-SCRATCH-NEXT: v_readlane_b32 s5, v40, 1
; GFX10-SCRATCH-NEXT: v_readlane_b32 s4, v40, 0
-; GFX10-SCRATCH-NEXT: v_readlane_b32 s0, v40, 4
+; GFX10-SCRATCH-NEXT: v_readlane_b32 s0, v41, 0
; GFX10-SCRATCH-NEXT: s_or_saveexec_b32 s1, -1
-; GFX10-SCRATCH-NEXT: scratch_load_dword v40, off, s33 ; 4-byte Folded Reload
+; GFX10-SCRATCH-NEXT: s_clause 0x1
+; GFX10-SCRATCH-NEXT: scratch_load_dword v40, off, s33
+; GFX10-SCRATCH-NEXT: scratch_load_dword v41, off, s33 offset:4
; GFX10-SCRATCH-NEXT: s_waitcnt_depctr 0xffe3
; GFX10-SCRATCH-NEXT: s_mov_b32 exec_lo, s1
; GFX10-SCRATCH-NEXT: s_add_i32 s32, s32, -16
@@ -7574,10 +8209,11 @@
; GFX9-NEXT: s_mov_b32 s33, s32
; GFX9-NEXT: s_or_saveexec_b64 s[36:37], -1
; GFX9-NEXT: buffer_store_dword v40, off, s[0:3], s33 ; 4-byte Folded Spill
+; GFX9-NEXT: buffer_store_dword v41, off, s[0:3], s33 offset:4 ; 4-byte Folded Spill
; GFX9-NEXT: s_mov_b64 exec, s[36:37]
-; GFX9-NEXT: v_writelane_b32 v40, s34, 6
; GFX9-NEXT: v_writelane_b32 v40, s4, 0
; GFX9-NEXT: v_writelane_b32 v40, s5, 1
+; GFX9-NEXT: v_writelane_b32 v41, s34, 0
; GFX9-NEXT: v_writelane_b32 v40, s6, 2
; GFX9-NEXT: s_mov_b64 s[34:35], 0
; GFX9-NEXT: v_writelane_b32 v40, s7, 3
@@ -7595,9 +8231,10 @@
; GFX9-NEXT: v_readlane_b32 s6, v40, 2
; GFX9-NEXT: v_readlane_b32 s5, v40, 1
; GFX9-NEXT: v_readlane_b32 s4, v40, 0
-; GFX9-NEXT: v_readlane_b32 s34, v40, 6
+; GFX9-NEXT: v_readlane_b32 s34, v41, 0
; GFX9-NEXT: s_or_saveexec_b64 s[36:37], -1
; GFX9-NEXT: buffer_load_dword v40, off, s[0:3], s33 ; 4-byte Folded Reload
+; GFX9-NEXT: buffer_load_dword v41, off, s[0:3], s33 offset:4 ; 4-byte Folded Reload
; GFX9-NEXT: s_mov_b64 exec, s[36:37]
; GFX9-NEXT: s_addk_i32 s32, 0xfc00
; GFX9-NEXT: s_mov_b32 s33, s34
@@ -7611,12 +8248,13 @@
; GFX10-NEXT: s_mov_b32 s33, s32
; GFX10-NEXT: s_or_saveexec_b32 s35, -1
; GFX10-NEXT: buffer_store_dword v40, off, s[0:3], s33 ; 4-byte Folded Spill
+; GFX10-NEXT: buffer_store_dword v41, off, s[0:3], s33 offset:4 ; 4-byte Folded Spill
; GFX10-NEXT: s_waitcnt_depctr 0xffe3
; GFX10-NEXT: s_mov_b32 exec_lo, s35
-; GFX10-NEXT: v_writelane_b32 v40, s34, 6
+; GFX10-NEXT: v_writelane_b32 v40, s4, 0
+; GFX10-NEXT: v_writelane_b32 v41, s34, 0
; GFX10-NEXT: s_mov_b64 s[34:35], 0
; GFX10-NEXT: s_addk_i32 s32, 0x200
-; GFX10-NEXT: v_writelane_b32 v40, s4, 0
; GFX10-NEXT: v_writelane_b32 v40, s5, 1
; GFX10-NEXT: v_writelane_b32 v40, s6, 2
; GFX10-NEXT: v_writelane_b32 v40, s7, 3
@@ -7633,9 +8271,11 @@
; GFX10-NEXT: v_readlane_b32 s6, v40, 2
; GFX10-NEXT: v_readlane_b32 s5, v40, 1
; GFX10-NEXT: v_readlane_b32 s4, v40, 0
-; GFX10-NEXT: v_readlane_b32 s34, v40, 6
+; GFX10-NEXT: v_readlane_b32 s34, v41, 0
; GFX10-NEXT: s_or_saveexec_b32 s35, -1
-; GFX10-NEXT: buffer_load_dword v40, off, s[0:3], s33 ; 4-byte Folded Reload
+; GFX10-NEXT: s_clause 0x1
+; GFX10-NEXT: buffer_load_dword v40, off, s[0:3], s33
+; GFX10-NEXT: buffer_load_dword v41, off, s[0:3], s33 offset:4
; GFX10-NEXT: s_waitcnt_depctr 0xffe3
; GFX10-NEXT: s_mov_b32 exec_lo, s35
; GFX10-NEXT: s_addk_i32 s32, 0xfe00
@@ -7649,12 +8289,14 @@
; GFX11-NEXT: s_mov_b32 s0, s33
; GFX11-NEXT: s_mov_b32 s33, s32
; GFX11-NEXT: s_or_saveexec_b32 s1, -1
-; GFX11-NEXT: scratch_store_b32 off, v40, s33 ; 4-byte Folded Spill
+; GFX11-NEXT: s_clause 0x1
+; GFX11-NEXT: scratch_store_b32 off, v40, s33
+; GFX11-NEXT: scratch_store_b32 off, v41, s33 offset:4
; GFX11-NEXT: s_mov_b32 exec_lo, s1
-; GFX11-NEXT: v_writelane_b32 v40, s0, 6
+; GFX11-NEXT: v_writelane_b32 v40, s4, 0
+; GFX11-NEXT: v_writelane_b32 v41, s0, 0
; GFX11-NEXT: s_mov_b64 s[0:1], 0
; GFX11-NEXT: s_add_i32 s32, s32, 16
-; GFX11-NEXT: v_writelane_b32 v40, s4, 0
; GFX11-NEXT: v_writelane_b32 v40, s5, 1
; GFX11-NEXT: v_writelane_b32 v40, s6, 2
; GFX11-NEXT: v_writelane_b32 v40, s7, 3
@@ -7672,9 +8314,11 @@
; GFX11-NEXT: v_readlane_b32 s6, v40, 2
; GFX11-NEXT: v_readlane_b32 s5, v40, 1
; GFX11-NEXT: v_readlane_b32 s4, v40, 0
-; GFX11-NEXT: v_readlane_b32 s0, v40, 6
+; GFX11-NEXT: v_readlane_b32 s0, v41, 0
; GFX11-NEXT: s_or_saveexec_b32 s1, -1
-; GFX11-NEXT: scratch_load_b32 v40, off, s33 ; 4-byte Folded Reload
+; GFX11-NEXT: s_clause 0x1
+; GFX11-NEXT: scratch_load_b32 v40, off, s33
+; GFX11-NEXT: scratch_load_b32 v41, off, s33 offset:4
; GFX11-NEXT: s_mov_b32 exec_lo, s1
; GFX11-NEXT: s_add_i32 s32, s32, -16
; GFX11-NEXT: s_mov_b32 s33, s0
@@ -7688,12 +8332,13 @@
; GFX10-SCRATCH-NEXT: s_mov_b32 s33, s32
; GFX10-SCRATCH-NEXT: s_or_saveexec_b32 s1, -1
; GFX10-SCRATCH-NEXT: scratch_store_dword off, v40, s33 ; 4-byte Folded Spill
+; GFX10-SCRATCH-NEXT: scratch_store_dword off, v41, s33 offset:4 ; 4-byte Folded Spill
; GFX10-SCRATCH-NEXT: s_waitcnt_depctr 0xffe3
; GFX10-SCRATCH-NEXT: s_mov_b32 exec_lo, s1
-; GFX10-SCRATCH-NEXT: v_writelane_b32 v40, s0, 6
+; GFX10-SCRATCH-NEXT: v_writelane_b32 v40, s4, 0
+; GFX10-SCRATCH-NEXT: v_writelane_b32 v41, s0, 0
; GFX10-SCRATCH-NEXT: s_mov_b64 s[0:1], 0
; GFX10-SCRATCH-NEXT: s_add_i32 s32, s32, 16
-; GFX10-SCRATCH-NEXT: v_writelane_b32 v40, s4, 0
; GFX10-SCRATCH-NEXT: v_writelane_b32 v40, s5, 1
; GFX10-SCRATCH-NEXT: v_writelane_b32 v40, s6, 2
; GFX10-SCRATCH-NEXT: v_writelane_b32 v40, s7, 3
@@ -7710,9 +8355,11 @@
; GFX10-SCRATCH-NEXT: v_readlane_b32 s6, v40, 2
; GFX10-SCRATCH-NEXT: v_readlane_b32 s5, v40, 1
; GFX10-SCRATCH-NEXT: v_readlane_b32 s4, v40, 0
-; GFX10-SCRATCH-NEXT: v_readlane_b32 s0, v40, 6
+; GFX10-SCRATCH-NEXT: v_readlane_b32 s0, v41, 0
; GFX10-SCRATCH-NEXT: s_or_saveexec_b32 s1, -1
-; GFX10-SCRATCH-NEXT: scratch_load_dword v40, off, s33 ; 4-byte Folded Reload
+; GFX10-SCRATCH-NEXT: s_clause 0x1
+; GFX10-SCRATCH-NEXT: scratch_load_dword v40, off, s33
+; GFX10-SCRATCH-NEXT: scratch_load_dword v41, off, s33 offset:4
; GFX10-SCRATCH-NEXT: s_waitcnt_depctr 0xffe3
; GFX10-SCRATCH-NEXT: s_mov_b32 exec_lo, s1
; GFX10-SCRATCH-NEXT: s_add_i32 s32, s32, -16
@@ -7732,8 +8379,8 @@
; GFX9-NEXT: s_mov_b32 s33, s32
; GFX9-NEXT: s_or_saveexec_b64 s[36:37], -1
; GFX9-NEXT: buffer_store_dword v40, off, s[0:3], s33 ; 4-byte Folded Spill
+; GFX9-NEXT: buffer_store_dword v41, off, s[0:3], s33 offset:4 ; 4-byte Folded Spill
; GFX9-NEXT: s_mov_b64 exec, s[36:37]
-; GFX9-NEXT: v_writelane_b32 v40, s34, 6
; GFX9-NEXT: v_writelane_b32 v40, s4, 0
; GFX9-NEXT: v_writelane_b32 v40, s5, 1
; GFX9-NEXT: v_writelane_b32 v40, s6, 2
@@ -7744,6 +8391,7 @@
; GFX9-NEXT: s_mov_b32 s5, 2
; GFX9-NEXT: s_mov_b32 s6, 3
; GFX9-NEXT: s_mov_b32 s7, 4
+; GFX9-NEXT: v_writelane_b32 v41, s34, 0
; GFX9-NEXT: v_writelane_b32 v40, s31, 5
; GFX9-NEXT: s_getpc_b64 s[34:35]
; GFX9-NEXT: s_add_u32 s34, s34, external_void_func_v2i64_inreg@rel32@lo+4
@@ -7755,9 +8403,10 @@
; GFX9-NEXT: v_readlane_b32 s6, v40, 2
; GFX9-NEXT: v_readlane_b32 s5, v40, 1
; GFX9-NEXT: v_readlane_b32 s4, v40, 0
-; GFX9-NEXT: v_readlane_b32 s34, v40, 6
+; GFX9-NEXT: v_readlane_b32 s34, v41, 0
; GFX9-NEXT: s_or_saveexec_b64 s[36:37], -1
; GFX9-NEXT: buffer_load_dword v40, off, s[0:3], s33 ; 4-byte Folded Reload
+; GFX9-NEXT: buffer_load_dword v41, off, s[0:3], s33 offset:4 ; 4-byte Folded Reload
; GFX9-NEXT: s_mov_b64 exec, s[36:37]
; GFX9-NEXT: s_addk_i32 s32, 0xfc00
; GFX9-NEXT: s_mov_b32 s33, s34
@@ -7771,15 +8420,16 @@
; GFX10-NEXT: s_mov_b32 s33, s32
; GFX10-NEXT: s_or_saveexec_b32 s35, -1
; GFX10-NEXT: buffer_store_dword v40, off, s[0:3], s33 ; 4-byte Folded Spill
+; GFX10-NEXT: buffer_store_dword v41, off, s[0:3], s33 offset:4 ; 4-byte Folded Spill
; GFX10-NEXT: s_waitcnt_depctr 0xffe3
; GFX10-NEXT: s_mov_b32 exec_lo, s35
-; GFX10-NEXT: v_writelane_b32 v40, s34, 6
+; GFX10-NEXT: v_writelane_b32 v40, s4, 0
+; GFX10-NEXT: s_mov_b32 s4, 1
; GFX10-NEXT: s_addk_i32 s32, 0x200
+; GFX10-NEXT: v_writelane_b32 v41, s34, 0
; GFX10-NEXT: s_getpc_b64 s[34:35]
; GFX10-NEXT: s_add_u32 s34, s34, external_void_func_v2i64_inreg@rel32@lo+4
; GFX10-NEXT: s_addc_u32 s35, s35, external_void_func_v2i64_inreg@rel32@hi+12
-; GFX10-NEXT: v_writelane_b32 v40, s4, 0
-; GFX10-NEXT: s_mov_b32 s4, 1
; GFX10-NEXT: v_writelane_b32 v40, s5, 1
; GFX10-NEXT: s_mov_b32 s5, 2
; GFX10-NEXT: v_writelane_b32 v40, s6, 2
@@ -7795,9 +8445,11 @@
; GFX10-NEXT: v_readlane_b32 s6, v40, 2
; GFX10-NEXT: v_readlane_b32 s5, v40, 1
; GFX10-NEXT: v_readlane_b32 s4, v40, 0
-; GFX10-NEXT: v_readlane_b32 s34, v40, 6
+; GFX10-NEXT: v_readlane_b32 s34, v41, 0
; GFX10-NEXT: s_or_saveexec_b32 s35, -1
-; GFX10-NEXT: buffer_load_dword v40, off, s[0:3], s33 ; 4-byte Folded Reload
+; GFX10-NEXT: s_clause 0x1
+; GFX10-NEXT: buffer_load_dword v40, off, s[0:3], s33
+; GFX10-NEXT: buffer_load_dword v41, off, s[0:3], s33 offset:4
; GFX10-NEXT: s_waitcnt_depctr 0xffe3
; GFX10-NEXT: s_mov_b32 exec_lo, s35
; GFX10-NEXT: s_addk_i32 s32, 0xfe00
@@ -7811,15 +8463,17 @@
; GFX11-NEXT: s_mov_b32 s0, s33
; GFX11-NEXT: s_mov_b32 s33, s32
; GFX11-NEXT: s_or_saveexec_b32 s1, -1
-; GFX11-NEXT: scratch_store_b32 off, v40, s33 ; 4-byte Folded Spill
+; GFX11-NEXT: s_clause 0x1
+; GFX11-NEXT: scratch_store_b32 off, v40, s33
+; GFX11-NEXT: scratch_store_b32 off, v41, s33 offset:4
; GFX11-NEXT: s_mov_b32 exec_lo, s1
-; GFX11-NEXT: v_writelane_b32 v40, s0, 6
+; GFX11-NEXT: v_writelane_b32 v40, s4, 0
+; GFX11-NEXT: s_mov_b32 s4, 1
; GFX11-NEXT: s_add_i32 s32, s32, 16
+; GFX11-NEXT: v_writelane_b32 v41, s0, 0
; GFX11-NEXT: s_getpc_b64 s[0:1]
; GFX11-NEXT: s_add_u32 s0, s0, external_void_func_v2i64_inreg@rel32@lo+4
; GFX11-NEXT: s_addc_u32 s1, s1, external_void_func_v2i64_inreg@rel32@hi+12
-; GFX11-NEXT: v_writelane_b32 v40, s4, 0
-; GFX11-NEXT: s_mov_b32 s4, 1
; GFX11-NEXT: v_writelane_b32 v40, s5, 1
; GFX11-NEXT: s_mov_b32 s5, 2
; GFX11-NEXT: v_writelane_b32 v40, s6, 2
@@ -7836,9 +8490,11 @@
; GFX11-NEXT: v_readlane_b32 s6, v40, 2
; GFX11-NEXT: v_readlane_b32 s5, v40, 1
; GFX11-NEXT: v_readlane_b32 s4, v40, 0
-; GFX11-NEXT: v_readlane_b32 s0, v40, 6
+; GFX11-NEXT: v_readlane_b32 s0, v41, 0
; GFX11-NEXT: s_or_saveexec_b32 s1, -1
-; GFX11-NEXT: scratch_load_b32 v40, off, s33 ; 4-byte Folded Reload
+; GFX11-NEXT: s_clause 0x1
+; GFX11-NEXT: scratch_load_b32 v40, off, s33
+; GFX11-NEXT: scratch_load_b32 v41, off, s33 offset:4
; GFX11-NEXT: s_mov_b32 exec_lo, s1
; GFX11-NEXT: s_add_i32 s32, s32, -16
; GFX11-NEXT: s_mov_b32 s33, s0
@@ -7852,15 +8508,16 @@
; GFX10-SCRATCH-NEXT: s_mov_b32 s33, s32
; GFX10-SCRATCH-NEXT: s_or_saveexec_b32 s1, -1
; GFX10-SCRATCH-NEXT: scratch_store_dword off, v40, s33 ; 4-byte Folded Spill
+; GFX10-SCRATCH-NEXT: scratch_store_dword off, v41, s33 offset:4 ; 4-byte Folded Spill
; GFX10-SCRATCH-NEXT: s_waitcnt_depctr 0xffe3
; GFX10-SCRATCH-NEXT: s_mov_b32 exec_lo, s1
-; GFX10-SCRATCH-NEXT: v_writelane_b32 v40, s0, 6
+; GFX10-SCRATCH-NEXT: v_writelane_b32 v40, s4, 0
+; GFX10-SCRATCH-NEXT: s_mov_b32 s4, 1
; GFX10-SCRATCH-NEXT: s_add_i32 s32, s32, 16
+; GFX10-SCRATCH-NEXT: v_writelane_b32 v41, s0, 0
; GFX10-SCRATCH-NEXT: s_getpc_b64 s[0:1]
; GFX10-SCRATCH-NEXT: s_add_u32 s0, s0, external_void_func_v2i64_inreg@rel32@lo+4
; GFX10-SCRATCH-NEXT: s_addc_u32 s1, s1, external_void_func_v2i64_inreg@rel32@hi+12
-; GFX10-SCRATCH-NEXT: v_writelane_b32 v40, s4, 0
-; GFX10-SCRATCH-NEXT: s_mov_b32 s4, 1
; GFX10-SCRATCH-NEXT: v_writelane_b32 v40, s5, 1
; GFX10-SCRATCH-NEXT: s_mov_b32 s5, 2
; GFX10-SCRATCH-NEXT: v_writelane_b32 v40, s6, 2
@@ -7876,9 +8533,11 @@
; GFX10-SCRATCH-NEXT: v_readlane_b32 s6, v40, 2
; GFX10-SCRATCH-NEXT: v_readlane_b32 s5, v40, 1
; GFX10-SCRATCH-NEXT: v_readlane_b32 s4, v40, 0
-; GFX10-SCRATCH-NEXT: v_readlane_b32 s0, v40, 6
+; GFX10-SCRATCH-NEXT: v_readlane_b32 s0, v41, 0
; GFX10-SCRATCH-NEXT: s_or_saveexec_b32 s1, -1
-; GFX10-SCRATCH-NEXT: scratch_load_dword v40, off, s33 ; 4-byte Folded Reload
+; GFX10-SCRATCH-NEXT: s_clause 0x1
+; GFX10-SCRATCH-NEXT: scratch_load_dword v40, off, s33
+; GFX10-SCRATCH-NEXT: scratch_load_dword v41, off, s33 offset:4
; GFX10-SCRATCH-NEXT: s_waitcnt_depctr 0xffe3
; GFX10-SCRATCH-NEXT: s_mov_b32 exec_lo, s1
; GFX10-SCRATCH-NEXT: s_add_i32 s32, s32, -16
@@ -7897,10 +8556,11 @@
; GFX9-NEXT: s_mov_b32 s33, s32
; GFX9-NEXT: s_or_saveexec_b64 s[36:37], -1
; GFX9-NEXT: buffer_store_dword v40, off, s[0:3], s33 ; 4-byte Folded Spill
+; GFX9-NEXT: buffer_store_dword v41, off, s[0:3], s33 offset:4 ; 4-byte Folded Spill
; GFX9-NEXT: s_mov_b64 exec, s[36:37]
-; GFX9-NEXT: v_writelane_b32 v40, s34, 8
; GFX9-NEXT: v_writelane_b32 v40, s4, 0
; GFX9-NEXT: v_writelane_b32 v40, s5, 1
+; GFX9-NEXT: v_writelane_b32 v41, s34, 0
; GFX9-NEXT: v_writelane_b32 v40, s6, 2
; GFX9-NEXT: s_mov_b64 s[34:35], 0
; GFX9-NEXT: v_writelane_b32 v40, s7, 3
@@ -7924,9 +8584,10 @@
; GFX9-NEXT: v_readlane_b32 s6, v40, 2
; GFX9-NEXT: v_readlane_b32 s5, v40, 1
; GFX9-NEXT: v_readlane_b32 s4, v40, 0
-; GFX9-NEXT: v_readlane_b32 s34, v40, 8
+; GFX9-NEXT: v_readlane_b32 s34, v41, 0
; GFX9-NEXT: s_or_saveexec_b64 s[36:37], -1
; GFX9-NEXT: buffer_load_dword v40, off, s[0:3], s33 ; 4-byte Folded Reload
+; GFX9-NEXT: buffer_load_dword v41, off, s[0:3], s33 offset:4 ; 4-byte Folded Reload
; GFX9-NEXT: s_mov_b64 exec, s[36:37]
; GFX9-NEXT: s_addk_i32 s32, 0xfc00
; GFX9-NEXT: s_mov_b32 s33, s34
@@ -7940,12 +8601,13 @@
; GFX10-NEXT: s_mov_b32 s33, s32
; GFX10-NEXT: s_or_saveexec_b32 s35, -1
; GFX10-NEXT: buffer_store_dword v40, off, s[0:3], s33 ; 4-byte Folded Spill
+; GFX10-NEXT: buffer_store_dword v41, off, s[0:3], s33 offset:4 ; 4-byte Folded Spill
; GFX10-NEXT: s_waitcnt_depctr 0xffe3
; GFX10-NEXT: s_mov_b32 exec_lo, s35
-; GFX10-NEXT: v_writelane_b32 v40, s34, 8
+; GFX10-NEXT: v_writelane_b32 v40, s4, 0
+; GFX10-NEXT: v_writelane_b32 v41, s34, 0
; GFX10-NEXT: s_mov_b64 s[34:35], 0
; GFX10-NEXT: s_addk_i32 s32, 0x200
-; GFX10-NEXT: v_writelane_b32 v40, s4, 0
; GFX10-NEXT: v_writelane_b32 v40, s5, 1
; GFX10-NEXT: v_writelane_b32 v40, s6, 2
; GFX10-NEXT: v_writelane_b32 v40, s7, 3
@@ -7968,9 +8630,11 @@
; GFX10-NEXT: v_readlane_b32 s6, v40, 2
; GFX10-NEXT: v_readlane_b32 s5, v40, 1
; GFX10-NEXT: v_readlane_b32 s4, v40, 0
-; GFX10-NEXT: v_readlane_b32 s34, v40, 8
+; GFX10-NEXT: v_readlane_b32 s34, v41, 0
; GFX10-NEXT: s_or_saveexec_b32 s35, -1
-; GFX10-NEXT: buffer_load_dword v40, off, s[0:3], s33 ; 4-byte Folded Reload
+; GFX10-NEXT: s_clause 0x1
+; GFX10-NEXT: buffer_load_dword v40, off, s[0:3], s33
+; GFX10-NEXT: buffer_load_dword v41, off, s[0:3], s33 offset:4
; GFX10-NEXT: s_waitcnt_depctr 0xffe3
; GFX10-NEXT: s_mov_b32 exec_lo, s35
; GFX10-NEXT: s_addk_i32 s32, 0xfe00
@@ -7984,12 +8648,14 @@
; GFX11-NEXT: s_mov_b32 s0, s33
; GFX11-NEXT: s_mov_b32 s33, s32
; GFX11-NEXT: s_or_saveexec_b32 s1, -1
-; GFX11-NEXT: scratch_store_b32 off, v40, s33 ; 4-byte Folded Spill
+; GFX11-NEXT: s_clause 0x1
+; GFX11-NEXT: scratch_store_b32 off, v40, s33
+; GFX11-NEXT: scratch_store_b32 off, v41, s33 offset:4
; GFX11-NEXT: s_mov_b32 exec_lo, s1
-; GFX11-NEXT: v_writelane_b32 v40, s0, 8
+; GFX11-NEXT: v_writelane_b32 v40, s4, 0
+; GFX11-NEXT: v_writelane_b32 v41, s0, 0
; GFX11-NEXT: s_mov_b64 s[0:1], 0
; GFX11-NEXT: s_add_i32 s32, s32, 16
-; GFX11-NEXT: v_writelane_b32 v40, s4, 0
; GFX11-NEXT: v_writelane_b32 v40, s5, 1
; GFX11-NEXT: v_writelane_b32 v40, s6, 2
; GFX11-NEXT: v_writelane_b32 v40, s7, 3
@@ -8013,9 +8679,11 @@
; GFX11-NEXT: v_readlane_b32 s6, v40, 2
; GFX11-NEXT: v_readlane_b32 s5, v40, 1
; GFX11-NEXT: v_readlane_b32 s4, v40, 0
-; GFX11-NEXT: v_readlane_b32 s0, v40, 8
+; GFX11-NEXT: v_readlane_b32 s0, v41, 0
; GFX11-NEXT: s_or_saveexec_b32 s1, -1
-; GFX11-NEXT: scratch_load_b32 v40, off, s33 ; 4-byte Folded Reload
+; GFX11-NEXT: s_clause 0x1
+; GFX11-NEXT: scratch_load_b32 v40, off, s33
+; GFX11-NEXT: scratch_load_b32 v41, off, s33 offset:4
; GFX11-NEXT: s_mov_b32 exec_lo, s1
; GFX11-NEXT: s_add_i32 s32, s32, -16
; GFX11-NEXT: s_mov_b32 s33, s0
@@ -8029,12 +8697,13 @@
; GFX10-SCRATCH-NEXT: s_mov_b32 s33, s32
; GFX10-SCRATCH-NEXT: s_or_saveexec_b32 s1, -1
; GFX10-SCRATCH-NEXT: scratch_store_dword off, v40, s33 ; 4-byte Folded Spill
+; GFX10-SCRATCH-NEXT: scratch_store_dword off, v41, s33 offset:4 ; 4-byte Folded Spill
; GFX10-SCRATCH-NEXT: s_waitcnt_depctr 0xffe3
; GFX10-SCRATCH-NEXT: s_mov_b32 exec_lo, s1
-; GFX10-SCRATCH-NEXT: v_writelane_b32 v40, s0, 8
+; GFX10-SCRATCH-NEXT: v_writelane_b32 v40, s4, 0
+; GFX10-SCRATCH-NEXT: v_writelane_b32 v41, s0, 0
; GFX10-SCRATCH-NEXT: s_mov_b64 s[0:1], 0
; GFX10-SCRATCH-NEXT: s_add_i32 s32, s32, 16
-; GFX10-SCRATCH-NEXT: v_writelane_b32 v40, s4, 0
; GFX10-SCRATCH-NEXT: v_writelane_b32 v40, s5, 1
; GFX10-SCRATCH-NEXT: v_writelane_b32 v40, s6, 2
; GFX10-SCRATCH-NEXT: v_writelane_b32 v40, s7, 3
@@ -8057,9 +8726,11 @@
; GFX10-SCRATCH-NEXT: v_readlane_b32 s6, v40, 2
; GFX10-SCRATCH-NEXT: v_readlane_b32 s5, v40, 1
; GFX10-SCRATCH-NEXT: v_readlane_b32 s4, v40, 0
-; GFX10-SCRATCH-NEXT: v_readlane_b32 s0, v40, 8
+; GFX10-SCRATCH-NEXT: v_readlane_b32 s0, v41, 0
; GFX10-SCRATCH-NEXT: s_or_saveexec_b32 s1, -1
-; GFX10-SCRATCH-NEXT: scratch_load_dword v40, off, s33 ; 4-byte Folded Reload
+; GFX10-SCRATCH-NEXT: s_clause 0x1
+; GFX10-SCRATCH-NEXT: scratch_load_dword v40, off, s33
+; GFX10-SCRATCH-NEXT: scratch_load_dword v41, off, s33 offset:4
; GFX10-SCRATCH-NEXT: s_waitcnt_depctr 0xffe3
; GFX10-SCRATCH-NEXT: s_mov_b32 exec_lo, s1
; GFX10-SCRATCH-NEXT: s_add_i32 s32, s32, -16
@@ -8081,11 +8752,12 @@
; GFX9-NEXT: s_mov_b32 s33, s32
; GFX9-NEXT: s_or_saveexec_b64 s[36:37], -1
; GFX9-NEXT: buffer_store_dword v40, off, s[0:3], s33 ; 4-byte Folded Spill
+; GFX9-NEXT: buffer_store_dword v41, off, s[0:3], s33 offset:4 ; 4-byte Folded Spill
; GFX9-NEXT: s_mov_b64 exec, s[36:37]
-; GFX9-NEXT: v_writelane_b32 v40, s34, 10
; GFX9-NEXT: v_writelane_b32 v40, s4, 0
; GFX9-NEXT: v_writelane_b32 v40, s5, 1
; GFX9-NEXT: v_writelane_b32 v40, s6, 2
+; GFX9-NEXT: v_writelane_b32 v41, s34, 0
; GFX9-NEXT: v_writelane_b32 v40, s7, 3
; GFX9-NEXT: s_mov_b64 s[34:35], 0
; GFX9-NEXT: v_writelane_b32 v40, s8, 4
@@ -8114,9 +8786,10 @@
; GFX9-NEXT: v_readlane_b32 s6, v40, 2
; GFX9-NEXT: v_readlane_b32 s5, v40, 1
; GFX9-NEXT: v_readlane_b32 s4, v40, 0
-; GFX9-NEXT: v_readlane_b32 s34, v40, 10
+; GFX9-NEXT: v_readlane_b32 s34, v41, 0
; GFX9-NEXT: s_or_saveexec_b64 s[36:37], -1
; GFX9-NEXT: buffer_load_dword v40, off, s[0:3], s33 ; 4-byte Folded Reload
+; GFX9-NEXT: buffer_load_dword v41, off, s[0:3], s33 offset:4 ; 4-byte Folded Reload
; GFX9-NEXT: s_mov_b64 exec, s[36:37]
; GFX9-NEXT: s_addk_i32 s32, 0xfc00
; GFX9-NEXT: s_mov_b32 s33, s34
@@ -8130,12 +8803,13 @@
; GFX10-NEXT: s_mov_b32 s33, s32
; GFX10-NEXT: s_or_saveexec_b32 s35, -1
; GFX10-NEXT: buffer_store_dword v40, off, s[0:3], s33 ; 4-byte Folded Spill
+; GFX10-NEXT: buffer_store_dword v41, off, s[0:3], s33 offset:4 ; 4-byte Folded Spill
; GFX10-NEXT: s_waitcnt_depctr 0xffe3
; GFX10-NEXT: s_mov_b32 exec_lo, s35
-; GFX10-NEXT: v_writelane_b32 v40, s34, 10
+; GFX10-NEXT: v_writelane_b32 v40, s4, 0
+; GFX10-NEXT: v_writelane_b32 v41, s34, 0
; GFX10-NEXT: s_mov_b64 s[34:35], 0
; GFX10-NEXT: s_addk_i32 s32, 0x200
-; GFX10-NEXT: v_writelane_b32 v40, s4, 0
; GFX10-NEXT: v_writelane_b32 v40, s5, 1
; GFX10-NEXT: v_writelane_b32 v40, s6, 2
; GFX10-NEXT: v_writelane_b32 v40, s7, 3
@@ -8164,9 +8838,11 @@
; GFX10-NEXT: v_readlane_b32 s6, v40, 2
; GFX10-NEXT: v_readlane_b32 s5, v40, 1
; GFX10-NEXT: v_readlane_b32 s4, v40, 0
-; GFX10-NEXT: v_readlane_b32 s34, v40, 10
+; GFX10-NEXT: v_readlane_b32 s34, v41, 0
; GFX10-NEXT: s_or_saveexec_b32 s35, -1
-; GFX10-NEXT: buffer_load_dword v40, off, s[0:3], s33 ; 4-byte Folded Reload
+; GFX10-NEXT: s_clause 0x1
+; GFX10-NEXT: buffer_load_dword v40, off, s[0:3], s33
+; GFX10-NEXT: buffer_load_dword v41, off, s[0:3], s33 offset:4
; GFX10-NEXT: s_waitcnt_depctr 0xffe3
; GFX10-NEXT: s_mov_b32 exec_lo, s35
; GFX10-NEXT: s_addk_i32 s32, 0xfe00
@@ -8180,12 +8856,14 @@
; GFX11-NEXT: s_mov_b32 s0, s33
; GFX11-NEXT: s_mov_b32 s33, s32
; GFX11-NEXT: s_or_saveexec_b32 s1, -1
-; GFX11-NEXT: scratch_store_b32 off, v40, s33 ; 4-byte Folded Spill
+; GFX11-NEXT: s_clause 0x1
+; GFX11-NEXT: scratch_store_b32 off, v40, s33
+; GFX11-NEXT: scratch_store_b32 off, v41, s33 offset:4
; GFX11-NEXT: s_mov_b32 exec_lo, s1
-; GFX11-NEXT: v_writelane_b32 v40, s0, 10
+; GFX11-NEXT: v_writelane_b32 v40, s4, 0
+; GFX11-NEXT: v_writelane_b32 v41, s0, 0
; GFX11-NEXT: s_mov_b64 s[0:1], 0
; GFX11-NEXT: s_add_i32 s32, s32, 16
-; GFX11-NEXT: v_writelane_b32 v40, s4, 0
; GFX11-NEXT: v_writelane_b32 v40, s5, 1
; GFX11-NEXT: v_writelane_b32 v40, s6, 2
; GFX11-NEXT: v_writelane_b32 v40, s7, 3
@@ -8215,9 +8893,11 @@
; GFX11-NEXT: v_readlane_b32 s6, v40, 2
; GFX11-NEXT: v_readlane_b32 s5, v40, 1
; GFX11-NEXT: v_readlane_b32 s4, v40, 0
-; GFX11-NEXT: v_readlane_b32 s0, v40, 10
+; GFX11-NEXT: v_readlane_b32 s0, v41, 0
; GFX11-NEXT: s_or_saveexec_b32 s1, -1
-; GFX11-NEXT: scratch_load_b32 v40, off, s33 ; 4-byte Folded Reload
+; GFX11-NEXT: s_clause 0x1
+; GFX11-NEXT: scratch_load_b32 v40, off, s33
+; GFX11-NEXT: scratch_load_b32 v41, off, s33 offset:4
; GFX11-NEXT: s_mov_b32 exec_lo, s1
; GFX11-NEXT: s_add_i32 s32, s32, -16
; GFX11-NEXT: s_mov_b32 s33, s0
@@ -8231,12 +8911,13 @@
; GFX10-SCRATCH-NEXT: s_mov_b32 s33, s32
; GFX10-SCRATCH-NEXT: s_or_saveexec_b32 s1, -1
; GFX10-SCRATCH-NEXT: scratch_store_dword off, v40, s33 ; 4-byte Folded Spill
+; GFX10-SCRATCH-NEXT: scratch_store_dword off, v41, s33 offset:4 ; 4-byte Folded Spill
; GFX10-SCRATCH-NEXT: s_waitcnt_depctr 0xffe3
; GFX10-SCRATCH-NEXT: s_mov_b32 exec_lo, s1
-; GFX10-SCRATCH-NEXT: v_writelane_b32 v40, s0, 10
+; GFX10-SCRATCH-NEXT: v_writelane_b32 v40, s4, 0
+; GFX10-SCRATCH-NEXT: v_writelane_b32 v41, s0, 0
; GFX10-SCRATCH-NEXT: s_mov_b64 s[0:1], 0
; GFX10-SCRATCH-NEXT: s_add_i32 s32, s32, 16
-; GFX10-SCRATCH-NEXT: v_writelane_b32 v40, s4, 0
; GFX10-SCRATCH-NEXT: v_writelane_b32 v40, s5, 1
; GFX10-SCRATCH-NEXT: v_writelane_b32 v40, s6, 2
; GFX10-SCRATCH-NEXT: v_writelane_b32 v40, s7, 3
@@ -8265,9 +8946,11 @@
; GFX10-SCRATCH-NEXT: v_readlane_b32 s6, v40, 2
; GFX10-SCRATCH-NEXT: v_readlane_b32 s5, v40, 1
; GFX10-SCRATCH-NEXT: v_readlane_b32 s4, v40, 0
-; GFX10-SCRATCH-NEXT: v_readlane_b32 s0, v40, 10
+; GFX10-SCRATCH-NEXT: v_readlane_b32 s0, v41, 0
; GFX10-SCRATCH-NEXT: s_or_saveexec_b32 s1, -1
-; GFX10-SCRATCH-NEXT: scratch_load_dword v40, off, s33 ; 4-byte Folded Reload
+; GFX10-SCRATCH-NEXT: s_clause 0x1
+; GFX10-SCRATCH-NEXT: scratch_load_dword v40, off, s33
+; GFX10-SCRATCH-NEXT: scratch_load_dword v41, off, s33 offset:4
; GFX10-SCRATCH-NEXT: s_waitcnt_depctr 0xffe3
; GFX10-SCRATCH-NEXT: s_mov_b32 exec_lo, s1
; GFX10-SCRATCH-NEXT: s_add_i32 s32, s32, -16
@@ -8288,12 +8971,13 @@
; GFX9-NEXT: s_mov_b32 s33, s32
; GFX9-NEXT: s_or_saveexec_b64 s[36:37], -1
; GFX9-NEXT: buffer_store_dword v40, off, s[0:3], s33 ; 4-byte Folded Spill
+; GFX9-NEXT: buffer_store_dword v41, off, s[0:3], s33 offset:4 ; 4-byte Folded Spill
; GFX9-NEXT: s_mov_b64 exec, s[36:37]
-; GFX9-NEXT: v_writelane_b32 v40, s34, 3
; GFX9-NEXT: v_writelane_b32 v40, s4, 0
; GFX9-NEXT: s_addk_i32 s32, 0x400
; GFX9-NEXT: v_writelane_b32 v40, s30, 1
; GFX9-NEXT: s_movk_i32 s4, 0x4400
+; GFX9-NEXT: v_writelane_b32 v41, s34, 0
; GFX9-NEXT: v_writelane_b32 v40, s31, 2
; GFX9-NEXT: s_getpc_b64 s[34:35]
; GFX9-NEXT: s_add_u32 s34, s34, external_void_func_f16_inreg@rel32@lo+4
@@ -8302,9 +8986,10 @@
; GFX9-NEXT: v_readlane_b32 s31, v40, 2
; GFX9-NEXT: v_readlane_b32 s30, v40, 1
; GFX9-NEXT: v_readlane_b32 s4, v40, 0
-; GFX9-NEXT: v_readlane_b32 s34, v40, 3
+; GFX9-NEXT: v_readlane_b32 s34, v41, 0
; GFX9-NEXT: s_or_saveexec_b64 s[36:37], -1
; GFX9-NEXT: buffer_load_dword v40, off, s[0:3], s33 ; 4-byte Folded Reload
+; GFX9-NEXT: buffer_load_dword v41, off, s[0:3], s33 offset:4 ; 4-byte Folded Reload
; GFX9-NEXT: s_mov_b64 exec, s[36:37]
; GFX9-NEXT: s_addk_i32 s32, 0xfc00
; GFX9-NEXT: s_mov_b32 s33, s34
@@ -8318,24 +9003,27 @@
; GFX10-NEXT: s_mov_b32 s33, s32
; GFX10-NEXT: s_or_saveexec_b32 s35, -1
; GFX10-NEXT: buffer_store_dword v40, off, s[0:3], s33 ; 4-byte Folded Spill
+; GFX10-NEXT: buffer_store_dword v41, off, s[0:3], s33 offset:4 ; 4-byte Folded Spill
; GFX10-NEXT: s_waitcnt_depctr 0xffe3
; GFX10-NEXT: s_mov_b32 exec_lo, s35
-; GFX10-NEXT: v_writelane_b32 v40, s34, 3
+; GFX10-NEXT: v_writelane_b32 v40, s4, 0
+; GFX10-NEXT: s_movk_i32 s4, 0x4400
; GFX10-NEXT: s_addk_i32 s32, 0x200
+; GFX10-NEXT: v_writelane_b32 v41, s34, 0
; GFX10-NEXT: s_getpc_b64 s[34:35]
; GFX10-NEXT: s_add_u32 s34, s34, external_void_func_f16_inreg@rel32@lo+4
; GFX10-NEXT: s_addc_u32 s35, s35, external_void_func_f16_inreg@rel32@hi+12
-; GFX10-NEXT: v_writelane_b32 v40, s4, 0
-; GFX10-NEXT: s_movk_i32 s4, 0x4400
; GFX10-NEXT: v_writelane_b32 v40, s30, 1
; GFX10-NEXT: v_writelane_b32 v40, s31, 2
; GFX10-NEXT: s_swappc_b64 s[30:31], s[34:35]
; GFX10-NEXT: v_readlane_b32 s31, v40, 2
; GFX10-NEXT: v_readlane_b32 s30, v40, 1
; GFX10-NEXT: v_readlane_b32 s4, v40, 0
-; GFX10-NEXT: v_readlane_b32 s34, v40, 3
+; GFX10-NEXT: v_readlane_b32 s34, v41, 0
; GFX10-NEXT: s_or_saveexec_b32 s35, -1
-; GFX10-NEXT: buffer_load_dword v40, off, s[0:3], s33 ; 4-byte Folded Reload
+; GFX10-NEXT: s_clause 0x1
+; GFX10-NEXT: buffer_load_dword v40, off, s[0:3], s33
+; GFX10-NEXT: buffer_load_dword v41, off, s[0:3], s33 offset:4
; GFX10-NEXT: s_waitcnt_depctr 0xffe3
; GFX10-NEXT: s_mov_b32 exec_lo, s35
; GFX10-NEXT: s_addk_i32 s32, 0xfe00
@@ -8349,15 +9037,17 @@
; GFX11-NEXT: s_mov_b32 s0, s33
; GFX11-NEXT: s_mov_b32 s33, s32
; GFX11-NEXT: s_or_saveexec_b32 s1, -1
-; GFX11-NEXT: scratch_store_b32 off, v40, s33 ; 4-byte Folded Spill
+; GFX11-NEXT: s_clause 0x1
+; GFX11-NEXT: scratch_store_b32 off, v40, s33
+; GFX11-NEXT: scratch_store_b32 off, v41, s33 offset:4
; GFX11-NEXT: s_mov_b32 exec_lo, s1
-; GFX11-NEXT: v_writelane_b32 v40, s0, 3
+; GFX11-NEXT: v_writelane_b32 v40, s4, 0
+; GFX11-NEXT: s_movk_i32 s4, 0x4400
; GFX11-NEXT: s_add_i32 s32, s32, 16
+; GFX11-NEXT: v_writelane_b32 v41, s0, 0
; GFX11-NEXT: s_getpc_b64 s[0:1]
; GFX11-NEXT: s_add_u32 s0, s0, external_void_func_f16_inreg@rel32@lo+4
; GFX11-NEXT: s_addc_u32 s1, s1, external_void_func_f16_inreg@rel32@hi+12
-; GFX11-NEXT: v_writelane_b32 v40, s4, 0
-; GFX11-NEXT: s_movk_i32 s4, 0x4400
; GFX11-NEXT: v_writelane_b32 v40, s30, 1
; GFX11-NEXT: v_writelane_b32 v40, s31, 2
; GFX11-NEXT: s_swappc_b64 s[30:31], s[0:1]
@@ -8365,9 +9055,11 @@
; GFX11-NEXT: v_readlane_b32 s31, v40, 2
; GFX11-NEXT: v_readlane_b32 s30, v40, 1
; GFX11-NEXT: v_readlane_b32 s4, v40, 0
-; GFX11-NEXT: v_readlane_b32 s0, v40, 3
+; GFX11-NEXT: v_readlane_b32 s0, v41, 0
; GFX11-NEXT: s_or_saveexec_b32 s1, -1
-; GFX11-NEXT: scratch_load_b32 v40, off, s33 ; 4-byte Folded Reload
+; GFX11-NEXT: s_clause 0x1
+; GFX11-NEXT: scratch_load_b32 v40, off, s33
+; GFX11-NEXT: scratch_load_b32 v41, off, s33 offset:4
; GFX11-NEXT: s_mov_b32 exec_lo, s1
; GFX11-NEXT: s_add_i32 s32, s32, -16
; GFX11-NEXT: s_mov_b32 s33, s0
@@ -8381,24 +9073,27 @@
; GFX10-SCRATCH-NEXT: s_mov_b32 s33, s32
; GFX10-SCRATCH-NEXT: s_or_saveexec_b32 s1, -1
; GFX10-SCRATCH-NEXT: scratch_store_dword off, v40, s33 ; 4-byte Folded Spill
+; GFX10-SCRATCH-NEXT: scratch_store_dword off, v41, s33 offset:4 ; 4-byte Folded Spill
; GFX10-SCRATCH-NEXT: s_waitcnt_depctr 0xffe3
; GFX10-SCRATCH-NEXT: s_mov_b32 exec_lo, s1
-; GFX10-SCRATCH-NEXT: v_writelane_b32 v40, s0, 3
+; GFX10-SCRATCH-NEXT: v_writelane_b32 v40, s4, 0
+; GFX10-SCRATCH-NEXT: s_movk_i32 s4, 0x4400
; GFX10-SCRATCH-NEXT: s_add_i32 s32, s32, 16
+; GFX10-SCRATCH-NEXT: v_writelane_b32 v41, s0, 0
; GFX10-SCRATCH-NEXT: s_getpc_b64 s[0:1]
; GFX10-SCRATCH-NEXT: s_add_u32 s0, s0, external_void_func_f16_inreg@rel32@lo+4
; GFX10-SCRATCH-NEXT: s_addc_u32 s1, s1, external_void_func_f16_inreg@rel32@hi+12
-; GFX10-SCRATCH-NEXT: v_writelane_b32 v40, s4, 0
-; GFX10-SCRATCH-NEXT: s_movk_i32 s4, 0x4400
; GFX10-SCRATCH-NEXT: v_writelane_b32 v40, s30, 1
; GFX10-SCRATCH-NEXT: v_writelane_b32 v40, s31, 2
; GFX10-SCRATCH-NEXT: s_swappc_b64 s[30:31], s[0:1]
; GFX10-SCRATCH-NEXT: v_readlane_b32 s31, v40, 2
; GFX10-SCRATCH-NEXT: v_readlane_b32 s30, v40, 1
; GFX10-SCRATCH-NEXT: v_readlane_b32 s4, v40, 0
-; GFX10-SCRATCH-NEXT: v_readlane_b32 s0, v40, 3
+; GFX10-SCRATCH-NEXT: v_readlane_b32 s0, v41, 0
; GFX10-SCRATCH-NEXT: s_or_saveexec_b32 s1, -1
-; GFX10-SCRATCH-NEXT: scratch_load_dword v40, off, s33 ; 4-byte Folded Reload
+; GFX10-SCRATCH-NEXT: s_clause 0x1
+; GFX10-SCRATCH-NEXT: scratch_load_dword v40, off, s33
+; GFX10-SCRATCH-NEXT: scratch_load_dword v41, off, s33 offset:4
; GFX10-SCRATCH-NEXT: s_waitcnt_depctr 0xffe3
; GFX10-SCRATCH-NEXT: s_mov_b32 exec_lo, s1
; GFX10-SCRATCH-NEXT: s_add_i32 s32, s32, -16
@@ -8417,12 +9112,13 @@
; GFX9-NEXT: s_mov_b32 s33, s32
; GFX9-NEXT: s_or_saveexec_b64 s[36:37], -1
; GFX9-NEXT: buffer_store_dword v40, off, s[0:3], s33 ; 4-byte Folded Spill
+; GFX9-NEXT: buffer_store_dword v41, off, s[0:3], s33 offset:4 ; 4-byte Folded Spill
; GFX9-NEXT: s_mov_b64 exec, s[36:37]
-; GFX9-NEXT: v_writelane_b32 v40, s34, 3
; GFX9-NEXT: v_writelane_b32 v40, s4, 0
; GFX9-NEXT: s_addk_i32 s32, 0x400
; GFX9-NEXT: v_writelane_b32 v40, s30, 1
; GFX9-NEXT: s_mov_b32 s4, 4.0
+; GFX9-NEXT: v_writelane_b32 v41, s34, 0
; GFX9-NEXT: v_writelane_b32 v40, s31, 2
; GFX9-NEXT: s_getpc_b64 s[34:35]
; GFX9-NEXT: s_add_u32 s34, s34, external_void_func_f32_inreg@rel32@lo+4
@@ -8431,9 +9127,10 @@
; GFX9-NEXT: v_readlane_b32 s31, v40, 2
; GFX9-NEXT: v_readlane_b32 s30, v40, 1
; GFX9-NEXT: v_readlane_b32 s4, v40, 0
-; GFX9-NEXT: v_readlane_b32 s34, v40, 3
+; GFX9-NEXT: v_readlane_b32 s34, v41, 0
; GFX9-NEXT: s_or_saveexec_b64 s[36:37], -1
; GFX9-NEXT: buffer_load_dword v40, off, s[0:3], s33 ; 4-byte Folded Reload
+; GFX9-NEXT: buffer_load_dword v41, off, s[0:3], s33 offset:4 ; 4-byte Folded Reload
; GFX9-NEXT: s_mov_b64 exec, s[36:37]
; GFX9-NEXT: s_addk_i32 s32, 0xfc00
; GFX9-NEXT: s_mov_b32 s33, s34
@@ -8447,24 +9144,27 @@
; GFX10-NEXT: s_mov_b32 s33, s32
; GFX10-NEXT: s_or_saveexec_b32 s35, -1
; GFX10-NEXT: buffer_store_dword v40, off, s[0:3], s33 ; 4-byte Folded Spill
+; GFX10-NEXT: buffer_store_dword v41, off, s[0:3], s33 offset:4 ; 4-byte Folded Spill
; GFX10-NEXT: s_waitcnt_depctr 0xffe3
; GFX10-NEXT: s_mov_b32 exec_lo, s35
-; GFX10-NEXT: v_writelane_b32 v40, s34, 3
+; GFX10-NEXT: v_writelane_b32 v40, s4, 0
+; GFX10-NEXT: s_mov_b32 s4, 4.0
; GFX10-NEXT: s_addk_i32 s32, 0x200
+; GFX10-NEXT: v_writelane_b32 v41, s34, 0
; GFX10-NEXT: s_getpc_b64 s[34:35]
; GFX10-NEXT: s_add_u32 s34, s34, external_void_func_f32_inreg@rel32@lo+4
; GFX10-NEXT: s_addc_u32 s35, s35, external_void_func_f32_inreg@rel32@hi+12
-; GFX10-NEXT: v_writelane_b32 v40, s4, 0
-; GFX10-NEXT: s_mov_b32 s4, 4.0
; GFX10-NEXT: v_writelane_b32 v40, s30, 1
; GFX10-NEXT: v_writelane_b32 v40, s31, 2
; GFX10-NEXT: s_swappc_b64 s[30:31], s[34:35]
; GFX10-NEXT: v_readlane_b32 s31, v40, 2
; GFX10-NEXT: v_readlane_b32 s30, v40, 1
; GFX10-NEXT: v_readlane_b32 s4, v40, 0
-; GFX10-NEXT: v_readlane_b32 s34, v40, 3
+; GFX10-NEXT: v_readlane_b32 s34, v41, 0
; GFX10-NEXT: s_or_saveexec_b32 s35, -1
-; GFX10-NEXT: buffer_load_dword v40, off, s[0:3], s33 ; 4-byte Folded Reload
+; GFX10-NEXT: s_clause 0x1
+; GFX10-NEXT: buffer_load_dword v40, off, s[0:3], s33
+; GFX10-NEXT: buffer_load_dword v41, off, s[0:3], s33 offset:4
; GFX10-NEXT: s_waitcnt_depctr 0xffe3
; GFX10-NEXT: s_mov_b32 exec_lo, s35
; GFX10-NEXT: s_addk_i32 s32, 0xfe00
@@ -8478,15 +9178,17 @@
; GFX11-NEXT: s_mov_b32 s0, s33
; GFX11-NEXT: s_mov_b32 s33, s32
; GFX11-NEXT: s_or_saveexec_b32 s1, -1
-; GFX11-NEXT: scratch_store_b32 off, v40, s33 ; 4-byte Folded Spill
+; GFX11-NEXT: s_clause 0x1
+; GFX11-NEXT: scratch_store_b32 off, v40, s33
+; GFX11-NEXT: scratch_store_b32 off, v41, s33 offset:4
; GFX11-NEXT: s_mov_b32 exec_lo, s1
-; GFX11-NEXT: v_writelane_b32 v40, s0, 3
+; GFX11-NEXT: v_writelane_b32 v40, s4, 0
+; GFX11-NEXT: s_mov_b32 s4, 4.0
; GFX11-NEXT: s_add_i32 s32, s32, 16
+; GFX11-NEXT: v_writelane_b32 v41, s0, 0
; GFX11-NEXT: s_getpc_b64 s[0:1]
; GFX11-NEXT: s_add_u32 s0, s0, external_void_func_f32_inreg@rel32@lo+4
; GFX11-NEXT: s_addc_u32 s1, s1, external_void_func_f32_inreg@rel32@hi+12
-; GFX11-NEXT: v_writelane_b32 v40, s4, 0
-; GFX11-NEXT: s_mov_b32 s4, 4.0
; GFX11-NEXT: v_writelane_b32 v40, s30, 1
; GFX11-NEXT: v_writelane_b32 v40, s31, 2
; GFX11-NEXT: s_swappc_b64 s[30:31], s[0:1]
@@ -8494,9 +9196,11 @@
; GFX11-NEXT: v_readlane_b32 s31, v40, 2
; GFX11-NEXT: v_readlane_b32 s30, v40, 1
; GFX11-NEXT: v_readlane_b32 s4, v40, 0
-; GFX11-NEXT: v_readlane_b32 s0, v40, 3
+; GFX11-NEXT: v_readlane_b32 s0, v41, 0
; GFX11-NEXT: s_or_saveexec_b32 s1, -1
-; GFX11-NEXT: scratch_load_b32 v40, off, s33 ; 4-byte Folded Reload
+; GFX11-NEXT: s_clause 0x1
+; GFX11-NEXT: scratch_load_b32 v40, off, s33
+; GFX11-NEXT: scratch_load_b32 v41, off, s33 offset:4
; GFX11-NEXT: s_mov_b32 exec_lo, s1
; GFX11-NEXT: s_add_i32 s32, s32, -16
; GFX11-NEXT: s_mov_b32 s33, s0
@@ -8510,24 +9214,27 @@
; GFX10-SCRATCH-NEXT: s_mov_b32 s33, s32
; GFX10-SCRATCH-NEXT: s_or_saveexec_b32 s1, -1
; GFX10-SCRATCH-NEXT: scratch_store_dword off, v40, s33 ; 4-byte Folded Spill
+; GFX10-SCRATCH-NEXT: scratch_store_dword off, v41, s33 offset:4 ; 4-byte Folded Spill
; GFX10-SCRATCH-NEXT: s_waitcnt_depctr 0xffe3
; GFX10-SCRATCH-NEXT: s_mov_b32 exec_lo, s1
-; GFX10-SCRATCH-NEXT: v_writelane_b32 v40, s0, 3
+; GFX10-SCRATCH-NEXT: v_writelane_b32 v40, s4, 0
+; GFX10-SCRATCH-NEXT: s_mov_b32 s4, 4.0
; GFX10-SCRATCH-NEXT: s_add_i32 s32, s32, 16
+; GFX10-SCRATCH-NEXT: v_writelane_b32 v41, s0, 0
; GFX10-SCRATCH-NEXT: s_getpc_b64 s[0:1]
; GFX10-SCRATCH-NEXT: s_add_u32 s0, s0, external_void_func_f32_inreg@rel32@lo+4
; GFX10-SCRATCH-NEXT: s_addc_u32 s1, s1, external_void_func_f32_inreg@rel32@hi+12
-; GFX10-SCRATCH-NEXT: v_writelane_b32 v40, s4, 0
-; GFX10-SCRATCH-NEXT: s_mov_b32 s4, 4.0
; GFX10-SCRATCH-NEXT: v_writelane_b32 v40, s30, 1
; GFX10-SCRATCH-NEXT: v_writelane_b32 v40, s31, 2
; GFX10-SCRATCH-NEXT: s_swappc_b64 s[30:31], s[0:1]
; GFX10-SCRATCH-NEXT: v_readlane_b32 s31, v40, 2
; GFX10-SCRATCH-NEXT: v_readlane_b32 s30, v40, 1
; GFX10-SCRATCH-NEXT: v_readlane_b32 s4, v40, 0
-; GFX10-SCRATCH-NEXT: v_readlane_b32 s0, v40, 3
+; GFX10-SCRATCH-NEXT: v_readlane_b32 s0, v41, 0
; GFX10-SCRATCH-NEXT: s_or_saveexec_b32 s1, -1
-; GFX10-SCRATCH-NEXT: scratch_load_dword v40, off, s33 ; 4-byte Folded Reload
+; GFX10-SCRATCH-NEXT: s_clause 0x1
+; GFX10-SCRATCH-NEXT: scratch_load_dword v40, off, s33
+; GFX10-SCRATCH-NEXT: scratch_load_dword v41, off, s33 offset:4
; GFX10-SCRATCH-NEXT: s_waitcnt_depctr 0xffe3
; GFX10-SCRATCH-NEXT: s_mov_b32 exec_lo, s1
; GFX10-SCRATCH-NEXT: s_add_i32 s32, s32, -16
@@ -8546,14 +9253,15 @@
; GFX9-NEXT: s_mov_b32 s33, s32
; GFX9-NEXT: s_or_saveexec_b64 s[36:37], -1
; GFX9-NEXT: buffer_store_dword v40, off, s[0:3], s33 ; 4-byte Folded Spill
+; GFX9-NEXT: buffer_store_dword v41, off, s[0:3], s33 offset:4 ; 4-byte Folded Spill
; GFX9-NEXT: s_mov_b64 exec, s[36:37]
-; GFX9-NEXT: v_writelane_b32 v40, s34, 4
; GFX9-NEXT: v_writelane_b32 v40, s4, 0
; GFX9-NEXT: v_writelane_b32 v40, s5, 1
; GFX9-NEXT: s_addk_i32 s32, 0x400
; GFX9-NEXT: v_writelane_b32 v40, s30, 2
; GFX9-NEXT: s_mov_b32 s4, 1.0
; GFX9-NEXT: s_mov_b32 s5, 2.0
+; GFX9-NEXT: v_writelane_b32 v41, s34, 0
; GFX9-NEXT: v_writelane_b32 v40, s31, 3
; GFX9-NEXT: s_getpc_b64 s[34:35]
; GFX9-NEXT: s_add_u32 s34, s34, external_void_func_v2f32_inreg@rel32@lo+4
@@ -8563,9 +9271,10 @@
; GFX9-NEXT: v_readlane_b32 s30, v40, 2
; GFX9-NEXT: v_readlane_b32 s5, v40, 1
; GFX9-NEXT: v_readlane_b32 s4, v40, 0
-; GFX9-NEXT: v_readlane_b32 s34, v40, 4
+; GFX9-NEXT: v_readlane_b32 s34, v41, 0
; GFX9-NEXT: s_or_saveexec_b64 s[36:37], -1
; GFX9-NEXT: buffer_load_dword v40, off, s[0:3], s33 ; 4-byte Folded Reload
+; GFX9-NEXT: buffer_load_dword v41, off, s[0:3], s33 offset:4 ; 4-byte Folded Reload
; GFX9-NEXT: s_mov_b64 exec, s[36:37]
; GFX9-NEXT: s_addk_i32 s32, 0xfc00
; GFX9-NEXT: s_mov_b32 s33, s34
@@ -8579,15 +9288,16 @@
; GFX10-NEXT: s_mov_b32 s33, s32
; GFX10-NEXT: s_or_saveexec_b32 s35, -1
; GFX10-NEXT: buffer_store_dword v40, off, s[0:3], s33 ; 4-byte Folded Spill
+; GFX10-NEXT: buffer_store_dword v41, off, s[0:3], s33 offset:4 ; 4-byte Folded Spill
; GFX10-NEXT: s_waitcnt_depctr 0xffe3
; GFX10-NEXT: s_mov_b32 exec_lo, s35
-; GFX10-NEXT: v_writelane_b32 v40, s34, 4
+; GFX10-NEXT: v_writelane_b32 v40, s4, 0
+; GFX10-NEXT: s_mov_b32 s4, 1.0
; GFX10-NEXT: s_addk_i32 s32, 0x200
+; GFX10-NEXT: v_writelane_b32 v41, s34, 0
; GFX10-NEXT: s_getpc_b64 s[34:35]
; GFX10-NEXT: s_add_u32 s34, s34, external_void_func_v2f32_inreg@rel32@lo+4
; GFX10-NEXT: s_addc_u32 s35, s35, external_void_func_v2f32_inreg@rel32@hi+12
-; GFX10-NEXT: v_writelane_b32 v40, s4, 0
-; GFX10-NEXT: s_mov_b32 s4, 1.0
; GFX10-NEXT: v_writelane_b32 v40, s5, 1
; GFX10-NEXT: s_mov_b32 s5, 2.0
; GFX10-NEXT: v_writelane_b32 v40, s30, 2
@@ -8597,9 +9307,11 @@
; GFX10-NEXT: v_readlane_b32 s30, v40, 2
; GFX10-NEXT: v_readlane_b32 s5, v40, 1
; GFX10-NEXT: v_readlane_b32 s4, v40, 0
-; GFX10-NEXT: v_readlane_b32 s34, v40, 4
+; GFX10-NEXT: v_readlane_b32 s34, v41, 0
; GFX10-NEXT: s_or_saveexec_b32 s35, -1
-; GFX10-NEXT: buffer_load_dword v40, off, s[0:3], s33 ; 4-byte Folded Reload
+; GFX10-NEXT: s_clause 0x1
+; GFX10-NEXT: buffer_load_dword v40, off, s[0:3], s33
+; GFX10-NEXT: buffer_load_dword v41, off, s[0:3], s33 offset:4
; GFX10-NEXT: s_waitcnt_depctr 0xffe3
; GFX10-NEXT: s_mov_b32 exec_lo, s35
; GFX10-NEXT: s_addk_i32 s32, 0xfe00
@@ -8613,15 +9325,17 @@
; GFX11-NEXT: s_mov_b32 s0, s33
; GFX11-NEXT: s_mov_b32 s33, s32
; GFX11-NEXT: s_or_saveexec_b32 s1, -1
-; GFX11-NEXT: scratch_store_b32 off, v40, s33 ; 4-byte Folded Spill
+; GFX11-NEXT: s_clause 0x1
+; GFX11-NEXT: scratch_store_b32 off, v40, s33
+; GFX11-NEXT: scratch_store_b32 off, v41, s33 offset:4
; GFX11-NEXT: s_mov_b32 exec_lo, s1
-; GFX11-NEXT: v_writelane_b32 v40, s0, 4
+; GFX11-NEXT: v_writelane_b32 v40, s4, 0
+; GFX11-NEXT: s_mov_b32 s4, 1.0
; GFX11-NEXT: s_add_i32 s32, s32, 16
+; GFX11-NEXT: v_writelane_b32 v41, s0, 0
; GFX11-NEXT: s_getpc_b64 s[0:1]
; GFX11-NEXT: s_add_u32 s0, s0, external_void_func_v2f32_inreg@rel32@lo+4
; GFX11-NEXT: s_addc_u32 s1, s1, external_void_func_v2f32_inreg@rel32@hi+12
-; GFX11-NEXT: v_writelane_b32 v40, s4, 0
-; GFX11-NEXT: s_mov_b32 s4, 1.0
; GFX11-NEXT: v_writelane_b32 v40, s5, 1
; GFX11-NEXT: s_mov_b32 s5, 2.0
; GFX11-NEXT: v_writelane_b32 v40, s30, 2
@@ -8632,9 +9346,11 @@
; GFX11-NEXT: v_readlane_b32 s30, v40, 2
; GFX11-NEXT: v_readlane_b32 s5, v40, 1
; GFX11-NEXT: v_readlane_b32 s4, v40, 0
-; GFX11-NEXT: v_readlane_b32 s0, v40, 4
+; GFX11-NEXT: v_readlane_b32 s0, v41, 0
; GFX11-NEXT: s_or_saveexec_b32 s1, -1
-; GFX11-NEXT: scratch_load_b32 v40, off, s33 ; 4-byte Folded Reload
+; GFX11-NEXT: s_clause 0x1
+; GFX11-NEXT: scratch_load_b32 v40, off, s33
+; GFX11-NEXT: scratch_load_b32 v41, off, s33 offset:4
; GFX11-NEXT: s_mov_b32 exec_lo, s1
; GFX11-NEXT: s_add_i32 s32, s32, -16
; GFX11-NEXT: s_mov_b32 s33, s0
@@ -8648,15 +9364,16 @@
; GFX10-SCRATCH-NEXT: s_mov_b32 s33, s32
; GFX10-SCRATCH-NEXT: s_or_saveexec_b32 s1, -1
; GFX10-SCRATCH-NEXT: scratch_store_dword off, v40, s33 ; 4-byte Folded Spill
+; GFX10-SCRATCH-NEXT: scratch_store_dword off, v41, s33 offset:4 ; 4-byte Folded Spill
; GFX10-SCRATCH-NEXT: s_waitcnt_depctr 0xffe3
; GFX10-SCRATCH-NEXT: s_mov_b32 exec_lo, s1
-; GFX10-SCRATCH-NEXT: v_writelane_b32 v40, s0, 4
+; GFX10-SCRATCH-NEXT: v_writelane_b32 v40, s4, 0
+; GFX10-SCRATCH-NEXT: s_mov_b32 s4, 1.0
; GFX10-SCRATCH-NEXT: s_add_i32 s32, s32, 16
+; GFX10-SCRATCH-NEXT: v_writelane_b32 v41, s0, 0
; GFX10-SCRATCH-NEXT: s_getpc_b64 s[0:1]
; GFX10-SCRATCH-NEXT: s_add_u32 s0, s0, external_void_func_v2f32_inreg@rel32@lo+4
; GFX10-SCRATCH-NEXT: s_addc_u32 s1, s1, external_void_func_v2f32_inreg@rel32@hi+12
-; GFX10-SCRATCH-NEXT: v_writelane_b32 v40, s4, 0
-; GFX10-SCRATCH-NEXT: s_mov_b32 s4, 1.0
; GFX10-SCRATCH-NEXT: v_writelane_b32 v40, s5, 1
; GFX10-SCRATCH-NEXT: s_mov_b32 s5, 2.0
; GFX10-SCRATCH-NEXT: v_writelane_b32 v40, s30, 2
@@ -8666,9 +9383,11 @@
; GFX10-SCRATCH-NEXT: v_readlane_b32 s30, v40, 2
; GFX10-SCRATCH-NEXT: v_readlane_b32 s5, v40, 1
; GFX10-SCRATCH-NEXT: v_readlane_b32 s4, v40, 0
-; GFX10-SCRATCH-NEXT: v_readlane_b32 s0, v40, 4
+; GFX10-SCRATCH-NEXT: v_readlane_b32 s0, v41, 0
; GFX10-SCRATCH-NEXT: s_or_saveexec_b32 s1, -1
-; GFX10-SCRATCH-NEXT: scratch_load_dword v40, off, s33 ; 4-byte Folded Reload
+; GFX10-SCRATCH-NEXT: s_clause 0x1
+; GFX10-SCRATCH-NEXT: scratch_load_dword v40, off, s33
+; GFX10-SCRATCH-NEXT: scratch_load_dword v41, off, s33 offset:4
; GFX10-SCRATCH-NEXT: s_waitcnt_depctr 0xffe3
; GFX10-SCRATCH-NEXT: s_mov_b32 exec_lo, s1
; GFX10-SCRATCH-NEXT: s_add_i32 s32, s32, -16
@@ -8687,8 +9406,8 @@
; GFX9-NEXT: s_mov_b32 s33, s32
; GFX9-NEXT: s_or_saveexec_b64 s[36:37], -1
; GFX9-NEXT: buffer_store_dword v40, off, s[0:3], s33 ; 4-byte Folded Spill
+; GFX9-NEXT: buffer_store_dword v41, off, s[0:3], s33 offset:4 ; 4-byte Folded Spill
; GFX9-NEXT: s_mov_b64 exec, s[36:37]
-; GFX9-NEXT: v_writelane_b32 v40, s34, 5
; GFX9-NEXT: v_writelane_b32 v40, s4, 0
; GFX9-NEXT: v_writelane_b32 v40, s5, 1
; GFX9-NEXT: v_writelane_b32 v40, s6, 2
@@ -8697,6 +9416,7 @@
; GFX9-NEXT: s_mov_b32 s4, 1.0
; GFX9-NEXT: s_mov_b32 s5, 2.0
; GFX9-NEXT: s_mov_b32 s6, 4.0
+; GFX9-NEXT: v_writelane_b32 v41, s34, 0
; GFX9-NEXT: v_writelane_b32 v40, s31, 4
; GFX9-NEXT: s_getpc_b64 s[34:35]
; GFX9-NEXT: s_add_u32 s34, s34, external_void_func_v3f32_inreg@rel32@lo+4
@@ -8707,9 +9427,10 @@
; GFX9-NEXT: v_readlane_b32 s6, v40, 2
; GFX9-NEXT: v_readlane_b32 s5, v40, 1
; GFX9-NEXT: v_readlane_b32 s4, v40, 0
-; GFX9-NEXT: v_readlane_b32 s34, v40, 5
+; GFX9-NEXT: v_readlane_b32 s34, v41, 0
; GFX9-NEXT: s_or_saveexec_b64 s[36:37], -1
; GFX9-NEXT: buffer_load_dword v40, off, s[0:3], s33 ; 4-byte Folded Reload
+; GFX9-NEXT: buffer_load_dword v41, off, s[0:3], s33 offset:4 ; 4-byte Folded Reload
; GFX9-NEXT: s_mov_b64 exec, s[36:37]
; GFX9-NEXT: s_addk_i32 s32, 0xfc00
; GFX9-NEXT: s_mov_b32 s33, s34
@@ -8723,15 +9444,16 @@
; GFX10-NEXT: s_mov_b32 s33, s32
; GFX10-NEXT: s_or_saveexec_b32 s35, -1
; GFX10-NEXT: buffer_store_dword v40, off, s[0:3], s33 ; 4-byte Folded Spill
+; GFX10-NEXT: buffer_store_dword v41, off, s[0:3], s33 offset:4 ; 4-byte Folded Spill
; GFX10-NEXT: s_waitcnt_depctr 0xffe3
; GFX10-NEXT: s_mov_b32 exec_lo, s35
-; GFX10-NEXT: v_writelane_b32 v40, s34, 5
+; GFX10-NEXT: v_writelane_b32 v40, s4, 0
+; GFX10-NEXT: s_mov_b32 s4, 1.0
; GFX10-NEXT: s_addk_i32 s32, 0x200
+; GFX10-NEXT: v_writelane_b32 v41, s34, 0
; GFX10-NEXT: s_getpc_b64 s[34:35]
; GFX10-NEXT: s_add_u32 s34, s34, external_void_func_v3f32_inreg@rel32@lo+4
; GFX10-NEXT: s_addc_u32 s35, s35, external_void_func_v3f32_inreg@rel32@hi+12
-; GFX10-NEXT: v_writelane_b32 v40, s4, 0
-; GFX10-NEXT: s_mov_b32 s4, 1.0
; GFX10-NEXT: v_writelane_b32 v40, s5, 1
; GFX10-NEXT: s_mov_b32 s5, 2.0
; GFX10-NEXT: v_writelane_b32 v40, s6, 2
@@ -8744,9 +9466,11 @@
; GFX10-NEXT: v_readlane_b32 s6, v40, 2
; GFX10-NEXT: v_readlane_b32 s5, v40, 1
; GFX10-NEXT: v_readlane_b32 s4, v40, 0
-; GFX10-NEXT: v_readlane_b32 s34, v40, 5
+; GFX10-NEXT: v_readlane_b32 s34, v41, 0
; GFX10-NEXT: s_or_saveexec_b32 s35, -1
-; GFX10-NEXT: buffer_load_dword v40, off, s[0:3], s33 ; 4-byte Folded Reload
+; GFX10-NEXT: s_clause 0x1
+; GFX10-NEXT: buffer_load_dword v40, off, s[0:3], s33
+; GFX10-NEXT: buffer_load_dword v41, off, s[0:3], s33 offset:4
; GFX10-NEXT: s_waitcnt_depctr 0xffe3
; GFX10-NEXT: s_mov_b32 exec_lo, s35
; GFX10-NEXT: s_addk_i32 s32, 0xfe00
@@ -8760,15 +9484,17 @@
; GFX11-NEXT: s_mov_b32 s0, s33
; GFX11-NEXT: s_mov_b32 s33, s32
; GFX11-NEXT: s_or_saveexec_b32 s1, -1
-; GFX11-NEXT: scratch_store_b32 off, v40, s33 ; 4-byte Folded Spill
+; GFX11-NEXT: s_clause 0x1
+; GFX11-NEXT: scratch_store_b32 off, v40, s33
+; GFX11-NEXT: scratch_store_b32 off, v41, s33 offset:4
; GFX11-NEXT: s_mov_b32 exec_lo, s1
-; GFX11-NEXT: v_writelane_b32 v40, s0, 5
+; GFX11-NEXT: v_writelane_b32 v40, s4, 0
+; GFX11-NEXT: s_mov_b32 s4, 1.0
; GFX11-NEXT: s_add_i32 s32, s32, 16
+; GFX11-NEXT: v_writelane_b32 v41, s0, 0
; GFX11-NEXT: s_getpc_b64 s[0:1]
; GFX11-NEXT: s_add_u32 s0, s0, external_void_func_v3f32_inreg@rel32@lo+4
; GFX11-NEXT: s_addc_u32 s1, s1, external_void_func_v3f32_inreg@rel32@hi+12
-; GFX11-NEXT: v_writelane_b32 v40, s4, 0
-; GFX11-NEXT: s_mov_b32 s4, 1.0
; GFX11-NEXT: v_writelane_b32 v40, s5, 1
; GFX11-NEXT: s_mov_b32 s5, 2.0
; GFX11-NEXT: v_writelane_b32 v40, s6, 2
@@ -8782,9 +9508,11 @@
; GFX11-NEXT: v_readlane_b32 s6, v40, 2
; GFX11-NEXT: v_readlane_b32 s5, v40, 1
; GFX11-NEXT: v_readlane_b32 s4, v40, 0
-; GFX11-NEXT: v_readlane_b32 s0, v40, 5
+; GFX11-NEXT: v_readlane_b32 s0, v41, 0
; GFX11-NEXT: s_or_saveexec_b32 s1, -1
-; GFX11-NEXT: scratch_load_b32 v40, off, s33 ; 4-byte Folded Reload
+; GFX11-NEXT: s_clause 0x1
+; GFX11-NEXT: scratch_load_b32 v40, off, s33
+; GFX11-NEXT: scratch_load_b32 v41, off, s33 offset:4
; GFX11-NEXT: s_mov_b32 exec_lo, s1
; GFX11-NEXT: s_add_i32 s32, s32, -16
; GFX11-NEXT: s_mov_b32 s33, s0
@@ -8798,15 +9526,16 @@
; GFX10-SCRATCH-NEXT: s_mov_b32 s33, s32
; GFX10-SCRATCH-NEXT: s_or_saveexec_b32 s1, -1
; GFX10-SCRATCH-NEXT: scratch_store_dword off, v40, s33 ; 4-byte Folded Spill
+; GFX10-SCRATCH-NEXT: scratch_store_dword off, v41, s33 offset:4 ; 4-byte Folded Spill
; GFX10-SCRATCH-NEXT: s_waitcnt_depctr 0xffe3
; GFX10-SCRATCH-NEXT: s_mov_b32 exec_lo, s1
-; GFX10-SCRATCH-NEXT: v_writelane_b32 v40, s0, 5
+; GFX10-SCRATCH-NEXT: v_writelane_b32 v40, s4, 0
+; GFX10-SCRATCH-NEXT: s_mov_b32 s4, 1.0
; GFX10-SCRATCH-NEXT: s_add_i32 s32, s32, 16
+; GFX10-SCRATCH-NEXT: v_writelane_b32 v41, s0, 0
; GFX10-SCRATCH-NEXT: s_getpc_b64 s[0:1]
; GFX10-SCRATCH-NEXT: s_add_u32 s0, s0, external_void_func_v3f32_inreg@rel32@lo+4
; GFX10-SCRATCH-NEXT: s_addc_u32 s1, s1, external_void_func_v3f32_inreg@rel32@hi+12
-; GFX10-SCRATCH-NEXT: v_writelane_b32 v40, s4, 0
-; GFX10-SCRATCH-NEXT: s_mov_b32 s4, 1.0
; GFX10-SCRATCH-NEXT: v_writelane_b32 v40, s5, 1
; GFX10-SCRATCH-NEXT: s_mov_b32 s5, 2.0
; GFX10-SCRATCH-NEXT: v_writelane_b32 v40, s6, 2
@@ -8819,9 +9548,11 @@
; GFX10-SCRATCH-NEXT: v_readlane_b32 s6, v40, 2
; GFX10-SCRATCH-NEXT: v_readlane_b32 s5, v40, 1
; GFX10-SCRATCH-NEXT: v_readlane_b32 s4, v40, 0
-; GFX10-SCRATCH-NEXT: v_readlane_b32 s0, v40, 5
+; GFX10-SCRATCH-NEXT: v_readlane_b32 s0, v41, 0
; GFX10-SCRATCH-NEXT: s_or_saveexec_b32 s1, -1
-; GFX10-SCRATCH-NEXT: scratch_load_dword v40, off, s33 ; 4-byte Folded Reload
+; GFX10-SCRATCH-NEXT: s_clause 0x1
+; GFX10-SCRATCH-NEXT: scratch_load_dword v40, off, s33
+; GFX10-SCRATCH-NEXT: scratch_load_dword v41, off, s33 offset:4
; GFX10-SCRATCH-NEXT: s_waitcnt_depctr 0xffe3
; GFX10-SCRATCH-NEXT: s_mov_b32 exec_lo, s1
; GFX10-SCRATCH-NEXT: s_add_i32 s32, s32, -16
@@ -8840,8 +9571,8 @@
; GFX9-NEXT: s_mov_b32 s33, s32
; GFX9-NEXT: s_or_saveexec_b64 s[36:37], -1
; GFX9-NEXT: buffer_store_dword v40, off, s[0:3], s33 ; 4-byte Folded Spill
+; GFX9-NEXT: buffer_store_dword v41, off, s[0:3], s33 offset:4 ; 4-byte Folded Spill
; GFX9-NEXT: s_mov_b64 exec, s[36:37]
-; GFX9-NEXT: v_writelane_b32 v40, s34, 7
; GFX9-NEXT: v_writelane_b32 v40, s4, 0
; GFX9-NEXT: v_writelane_b32 v40, s5, 1
; GFX9-NEXT: v_writelane_b32 v40, s6, 2
@@ -8854,6 +9585,7 @@
; GFX9-NEXT: s_mov_b32 s6, 4.0
; GFX9-NEXT: s_mov_b32 s7, -1.0
; GFX9-NEXT: s_mov_b32 s8, 0.5
+; GFX9-NEXT: v_writelane_b32 v41, s34, 0
; GFX9-NEXT: v_writelane_b32 v40, s31, 6
; GFX9-NEXT: s_getpc_b64 s[34:35]
; GFX9-NEXT: s_add_u32 s34, s34, external_void_func_v5f32_inreg@rel32@lo+4
@@ -8866,9 +9598,10 @@
; GFX9-NEXT: v_readlane_b32 s6, v40, 2
; GFX9-NEXT: v_readlane_b32 s5, v40, 1
; GFX9-NEXT: v_readlane_b32 s4, v40, 0
-; GFX9-NEXT: v_readlane_b32 s34, v40, 7
+; GFX9-NEXT: v_readlane_b32 s34, v41, 0
; GFX9-NEXT: s_or_saveexec_b64 s[36:37], -1
; GFX9-NEXT: buffer_load_dword v40, off, s[0:3], s33 ; 4-byte Folded Reload
+; GFX9-NEXT: buffer_load_dword v41, off, s[0:3], s33 offset:4 ; 4-byte Folded Reload
; GFX9-NEXT: s_mov_b64 exec, s[36:37]
; GFX9-NEXT: s_addk_i32 s32, 0xfc00
; GFX9-NEXT: s_mov_b32 s33, s34
@@ -8882,15 +9615,16 @@
; GFX10-NEXT: s_mov_b32 s33, s32
; GFX10-NEXT: s_or_saveexec_b32 s35, -1
; GFX10-NEXT: buffer_store_dword v40, off, s[0:3], s33 ; 4-byte Folded Spill
+; GFX10-NEXT: buffer_store_dword v41, off, s[0:3], s33 offset:4 ; 4-byte Folded Spill
; GFX10-NEXT: s_waitcnt_depctr 0xffe3
; GFX10-NEXT: s_mov_b32 exec_lo, s35
-; GFX10-NEXT: v_writelane_b32 v40, s34, 7
+; GFX10-NEXT: v_writelane_b32 v40, s4, 0
+; GFX10-NEXT: s_mov_b32 s4, 1.0
; GFX10-NEXT: s_addk_i32 s32, 0x200
+; GFX10-NEXT: v_writelane_b32 v41, s34, 0
; GFX10-NEXT: s_getpc_b64 s[34:35]
; GFX10-NEXT: s_add_u32 s34, s34, external_void_func_v5f32_inreg@rel32@lo+4
; GFX10-NEXT: s_addc_u32 s35, s35, external_void_func_v5f32_inreg@rel32@hi+12
-; GFX10-NEXT: v_writelane_b32 v40, s4, 0
-; GFX10-NEXT: s_mov_b32 s4, 1.0
; GFX10-NEXT: v_writelane_b32 v40, s5, 1
; GFX10-NEXT: s_mov_b32 s5, 2.0
; GFX10-NEXT: v_writelane_b32 v40, s6, 2
@@ -8909,9 +9643,11 @@
; GFX10-NEXT: v_readlane_b32 s6, v40, 2
; GFX10-NEXT: v_readlane_b32 s5, v40, 1
; GFX10-NEXT: v_readlane_b32 s4, v40, 0
-; GFX10-NEXT: v_readlane_b32 s34, v40, 7
+; GFX10-NEXT: v_readlane_b32 s34, v41, 0
; GFX10-NEXT: s_or_saveexec_b32 s35, -1
-; GFX10-NEXT: buffer_load_dword v40, off, s[0:3], s33 ; 4-byte Folded Reload
+; GFX10-NEXT: s_clause 0x1
+; GFX10-NEXT: buffer_load_dword v40, off, s[0:3], s33
+; GFX10-NEXT: buffer_load_dword v41, off, s[0:3], s33 offset:4
; GFX10-NEXT: s_waitcnt_depctr 0xffe3
; GFX10-NEXT: s_mov_b32 exec_lo, s35
; GFX10-NEXT: s_addk_i32 s32, 0xfe00
@@ -8925,15 +9661,17 @@
; GFX11-NEXT: s_mov_b32 s0, s33
; GFX11-NEXT: s_mov_b32 s33, s32
; GFX11-NEXT: s_or_saveexec_b32 s1, -1
-; GFX11-NEXT: scratch_store_b32 off, v40, s33 ; 4-byte Folded Spill
+; GFX11-NEXT: s_clause 0x1
+; GFX11-NEXT: scratch_store_b32 off, v40, s33
+; GFX11-NEXT: scratch_store_b32 off, v41, s33 offset:4
; GFX11-NEXT: s_mov_b32 exec_lo, s1
-; GFX11-NEXT: v_writelane_b32 v40, s0, 7
+; GFX11-NEXT: v_writelane_b32 v40, s4, 0
+; GFX11-NEXT: s_mov_b32 s4, 1.0
; GFX11-NEXT: s_add_i32 s32, s32, 16
+; GFX11-NEXT: v_writelane_b32 v41, s0, 0
; GFX11-NEXT: s_getpc_b64 s[0:1]
; GFX11-NEXT: s_add_u32 s0, s0, external_void_func_v5f32_inreg@rel32@lo+4
; GFX11-NEXT: s_addc_u32 s1, s1, external_void_func_v5f32_inreg@rel32@hi+12
-; GFX11-NEXT: v_writelane_b32 v40, s4, 0
-; GFX11-NEXT: s_mov_b32 s4, 1.0
; GFX11-NEXT: v_writelane_b32 v40, s5, 1
; GFX11-NEXT: s_mov_b32 s5, 2.0
; GFX11-NEXT: v_writelane_b32 v40, s6, 2
@@ -8953,9 +9691,11 @@
; GFX11-NEXT: v_readlane_b32 s6, v40, 2
; GFX11-NEXT: v_readlane_b32 s5, v40, 1
; GFX11-NEXT: v_readlane_b32 s4, v40, 0
-; GFX11-NEXT: v_readlane_b32 s0, v40, 7
+; GFX11-NEXT: v_readlane_b32 s0, v41, 0
; GFX11-NEXT: s_or_saveexec_b32 s1, -1
-; GFX11-NEXT: scratch_load_b32 v40, off, s33 ; 4-byte Folded Reload
+; GFX11-NEXT: s_clause 0x1
+; GFX11-NEXT: scratch_load_b32 v40, off, s33
+; GFX11-NEXT: scratch_load_b32 v41, off, s33 offset:4
; GFX11-NEXT: s_mov_b32 exec_lo, s1
; GFX11-NEXT: s_add_i32 s32, s32, -16
; GFX11-NEXT: s_mov_b32 s33, s0
@@ -8969,15 +9709,16 @@
; GFX10-SCRATCH-NEXT: s_mov_b32 s33, s32
; GFX10-SCRATCH-NEXT: s_or_saveexec_b32 s1, -1
; GFX10-SCRATCH-NEXT: scratch_store_dword off, v40, s33 ; 4-byte Folded Spill
+; GFX10-SCRATCH-NEXT: scratch_store_dword off, v41, s33 offset:4 ; 4-byte Folded Spill
; GFX10-SCRATCH-NEXT: s_waitcnt_depctr 0xffe3
; GFX10-SCRATCH-NEXT: s_mov_b32 exec_lo, s1
-; GFX10-SCRATCH-NEXT: v_writelane_b32 v40, s0, 7
+; GFX10-SCRATCH-NEXT: v_writelane_b32 v40, s4, 0
+; GFX10-SCRATCH-NEXT: s_mov_b32 s4, 1.0
; GFX10-SCRATCH-NEXT: s_add_i32 s32, s32, 16
+; GFX10-SCRATCH-NEXT: v_writelane_b32 v41, s0, 0
; GFX10-SCRATCH-NEXT: s_getpc_b64 s[0:1]
; GFX10-SCRATCH-NEXT: s_add_u32 s0, s0, external_void_func_v5f32_inreg@rel32@lo+4
; GFX10-SCRATCH-NEXT: s_addc_u32 s1, s1, external_void_func_v5f32_inreg@rel32@hi+12
-; GFX10-SCRATCH-NEXT: v_writelane_b32 v40, s4, 0
-; GFX10-SCRATCH-NEXT: s_mov_b32 s4, 1.0
; GFX10-SCRATCH-NEXT: v_writelane_b32 v40, s5, 1
; GFX10-SCRATCH-NEXT: s_mov_b32 s5, 2.0
; GFX10-SCRATCH-NEXT: v_writelane_b32 v40, s6, 2
@@ -8996,9 +9737,11 @@
; GFX10-SCRATCH-NEXT: v_readlane_b32 s6, v40, 2
; GFX10-SCRATCH-NEXT: v_readlane_b32 s5, v40, 1
; GFX10-SCRATCH-NEXT: v_readlane_b32 s4, v40, 0
-; GFX10-SCRATCH-NEXT: v_readlane_b32 s0, v40, 7
+; GFX10-SCRATCH-NEXT: v_readlane_b32 s0, v41, 0
; GFX10-SCRATCH-NEXT: s_or_saveexec_b32 s1, -1
-; GFX10-SCRATCH-NEXT: scratch_load_dword v40, off, s33 ; 4-byte Folded Reload
+; GFX10-SCRATCH-NEXT: s_clause 0x1
+; GFX10-SCRATCH-NEXT: scratch_load_dword v40, off, s33
+; GFX10-SCRATCH-NEXT: scratch_load_dword v41, off, s33 offset:4
; GFX10-SCRATCH-NEXT: s_waitcnt_depctr 0xffe3
; GFX10-SCRATCH-NEXT: s_mov_b32 exec_lo, s1
; GFX10-SCRATCH-NEXT: s_add_i32 s32, s32, -16
@@ -9017,14 +9760,15 @@
; GFX9-NEXT: s_mov_b32 s33, s32
; GFX9-NEXT: s_or_saveexec_b64 s[36:37], -1
; GFX9-NEXT: buffer_store_dword v40, off, s[0:3], s33 ; 4-byte Folded Spill
+; GFX9-NEXT: buffer_store_dword v41, off, s[0:3], s33 offset:4 ; 4-byte Folded Spill
; GFX9-NEXT: s_mov_b64 exec, s[36:37]
-; GFX9-NEXT: v_writelane_b32 v40, s34, 4
; GFX9-NEXT: v_writelane_b32 v40, s4, 0
; GFX9-NEXT: v_writelane_b32 v40, s5, 1
; GFX9-NEXT: s_addk_i32 s32, 0x400
; GFX9-NEXT: v_writelane_b32 v40, s30, 2
; GFX9-NEXT: s_mov_b32 s4, 0
; GFX9-NEXT: s_mov_b32 s5, 0x40100000
+; GFX9-NEXT: v_writelane_b32 v41, s34, 0
; GFX9-NEXT: v_writelane_b32 v40, s31, 3
; GFX9-NEXT: s_getpc_b64 s[34:35]
; GFX9-NEXT: s_add_u32 s34, s34, external_void_func_f64_inreg@rel32@lo+4
@@ -9034,9 +9778,10 @@
; GFX9-NEXT: v_readlane_b32 s30, v40, 2
; GFX9-NEXT: v_readlane_b32 s5, v40, 1
; GFX9-NEXT: v_readlane_b32 s4, v40, 0
-; GFX9-NEXT: v_readlane_b32 s34, v40, 4
+; GFX9-NEXT: v_readlane_b32 s34, v41, 0
; GFX9-NEXT: s_or_saveexec_b64 s[36:37], -1
; GFX9-NEXT: buffer_load_dword v40, off, s[0:3], s33 ; 4-byte Folded Reload
+; GFX9-NEXT: buffer_load_dword v41, off, s[0:3], s33 offset:4 ; 4-byte Folded Reload
; GFX9-NEXT: s_mov_b64 exec, s[36:37]
; GFX9-NEXT: s_addk_i32 s32, 0xfc00
; GFX9-NEXT: s_mov_b32 s33, s34
@@ -9050,15 +9795,16 @@
; GFX10-NEXT: s_mov_b32 s33, s32
; GFX10-NEXT: s_or_saveexec_b32 s35, -1
; GFX10-NEXT: buffer_store_dword v40, off, s[0:3], s33 ; 4-byte Folded Spill
+; GFX10-NEXT: buffer_store_dword v41, off, s[0:3], s33 offset:4 ; 4-byte Folded Spill
; GFX10-NEXT: s_waitcnt_depctr 0xffe3
; GFX10-NEXT: s_mov_b32 exec_lo, s35
-; GFX10-NEXT: v_writelane_b32 v40, s34, 4
+; GFX10-NEXT: v_writelane_b32 v40, s4, 0
+; GFX10-NEXT: s_mov_b32 s4, 0
; GFX10-NEXT: s_addk_i32 s32, 0x200
+; GFX10-NEXT: v_writelane_b32 v41, s34, 0
; GFX10-NEXT: s_getpc_b64 s[34:35]
; GFX10-NEXT: s_add_u32 s34, s34, external_void_func_f64_inreg@rel32@lo+4
; GFX10-NEXT: s_addc_u32 s35, s35, external_void_func_f64_inreg@rel32@hi+12
-; GFX10-NEXT: v_writelane_b32 v40, s4, 0
-; GFX10-NEXT: s_mov_b32 s4, 0
; GFX10-NEXT: v_writelane_b32 v40, s5, 1
; GFX10-NEXT: s_mov_b32 s5, 0x40100000
; GFX10-NEXT: v_writelane_b32 v40, s30, 2
@@ -9068,9 +9814,11 @@
; GFX10-NEXT: v_readlane_b32 s30, v40, 2
; GFX10-NEXT: v_readlane_b32 s5, v40, 1
; GFX10-NEXT: v_readlane_b32 s4, v40, 0
-; GFX10-NEXT: v_readlane_b32 s34, v40, 4
+; GFX10-NEXT: v_readlane_b32 s34, v41, 0
; GFX10-NEXT: s_or_saveexec_b32 s35, -1
-; GFX10-NEXT: buffer_load_dword v40, off, s[0:3], s33 ; 4-byte Folded Reload
+; GFX10-NEXT: s_clause 0x1
+; GFX10-NEXT: buffer_load_dword v40, off, s[0:3], s33
+; GFX10-NEXT: buffer_load_dword v41, off, s[0:3], s33 offset:4
; GFX10-NEXT: s_waitcnt_depctr 0xffe3
; GFX10-NEXT: s_mov_b32 exec_lo, s35
; GFX10-NEXT: s_addk_i32 s32, 0xfe00
@@ -9084,15 +9832,17 @@
; GFX11-NEXT: s_mov_b32 s0, s33
; GFX11-NEXT: s_mov_b32 s33, s32
; GFX11-NEXT: s_or_saveexec_b32 s1, -1
-; GFX11-NEXT: scratch_store_b32 off, v40, s33 ; 4-byte Folded Spill
+; GFX11-NEXT: s_clause 0x1
+; GFX11-NEXT: scratch_store_b32 off, v40, s33
+; GFX11-NEXT: scratch_store_b32 off, v41, s33 offset:4
; GFX11-NEXT: s_mov_b32 exec_lo, s1
-; GFX11-NEXT: v_writelane_b32 v40, s0, 4
+; GFX11-NEXT: v_writelane_b32 v40, s4, 0
+; GFX11-NEXT: s_mov_b32 s4, 0
; GFX11-NEXT: s_add_i32 s32, s32, 16
+; GFX11-NEXT: v_writelane_b32 v41, s0, 0
; GFX11-NEXT: s_getpc_b64 s[0:1]
; GFX11-NEXT: s_add_u32 s0, s0, external_void_func_f64_inreg@rel32@lo+4
; GFX11-NEXT: s_addc_u32 s1, s1, external_void_func_f64_inreg@rel32@hi+12
-; GFX11-NEXT: v_writelane_b32 v40, s4, 0
-; GFX11-NEXT: s_mov_b32 s4, 0
; GFX11-NEXT: v_writelane_b32 v40, s5, 1
; GFX11-NEXT: s_mov_b32 s5, 0x40100000
; GFX11-NEXT: v_writelane_b32 v40, s30, 2
@@ -9103,9 +9853,11 @@
; GFX11-NEXT: v_readlane_b32 s30, v40, 2
; GFX11-NEXT: v_readlane_b32 s5, v40, 1
; GFX11-NEXT: v_readlane_b32 s4, v40, 0
-; GFX11-NEXT: v_readlane_b32 s0, v40, 4
+; GFX11-NEXT: v_readlane_b32 s0, v41, 0
; GFX11-NEXT: s_or_saveexec_b32 s1, -1
-; GFX11-NEXT: scratch_load_b32 v40, off, s33 ; 4-byte Folded Reload
+; GFX11-NEXT: s_clause 0x1
+; GFX11-NEXT: scratch_load_b32 v40, off, s33
+; GFX11-NEXT: scratch_load_b32 v41, off, s33 offset:4
; GFX11-NEXT: s_mov_b32 exec_lo, s1
; GFX11-NEXT: s_add_i32 s32, s32, -16
; GFX11-NEXT: s_mov_b32 s33, s0
@@ -9119,15 +9871,16 @@
; GFX10-SCRATCH-NEXT: s_mov_b32 s33, s32
; GFX10-SCRATCH-NEXT: s_or_saveexec_b32 s1, -1
; GFX10-SCRATCH-NEXT: scratch_store_dword off, v40, s33 ; 4-byte Folded Spill
+; GFX10-SCRATCH-NEXT: scratch_store_dword off, v41, s33 offset:4 ; 4-byte Folded Spill
; GFX10-SCRATCH-NEXT: s_waitcnt_depctr 0xffe3
; GFX10-SCRATCH-NEXT: s_mov_b32 exec_lo, s1
-; GFX10-SCRATCH-NEXT: v_writelane_b32 v40, s0, 4
+; GFX10-SCRATCH-NEXT: v_writelane_b32 v40, s4, 0
+; GFX10-SCRATCH-NEXT: s_mov_b32 s4, 0
; GFX10-SCRATCH-NEXT: s_add_i32 s32, s32, 16
+; GFX10-SCRATCH-NEXT: v_writelane_b32 v41, s0, 0
; GFX10-SCRATCH-NEXT: s_getpc_b64 s[0:1]
; GFX10-SCRATCH-NEXT: s_add_u32 s0, s0, external_void_func_f64_inreg@rel32@lo+4
; GFX10-SCRATCH-NEXT: s_addc_u32 s1, s1, external_void_func_f64_inreg@rel32@hi+12
-; GFX10-SCRATCH-NEXT: v_writelane_b32 v40, s4, 0
-; GFX10-SCRATCH-NEXT: s_mov_b32 s4, 0
; GFX10-SCRATCH-NEXT: v_writelane_b32 v40, s5, 1
; GFX10-SCRATCH-NEXT: s_mov_b32 s5, 0x40100000
; GFX10-SCRATCH-NEXT: v_writelane_b32 v40, s30, 2
@@ -9137,9 +9890,11 @@
; GFX10-SCRATCH-NEXT: v_readlane_b32 s30, v40, 2
; GFX10-SCRATCH-NEXT: v_readlane_b32 s5, v40, 1
; GFX10-SCRATCH-NEXT: v_readlane_b32 s4, v40, 0
-; GFX10-SCRATCH-NEXT: v_readlane_b32 s0, v40, 4
+; GFX10-SCRATCH-NEXT: v_readlane_b32 s0, v41, 0
; GFX10-SCRATCH-NEXT: s_or_saveexec_b32 s1, -1
-; GFX10-SCRATCH-NEXT: scratch_load_dword v40, off, s33 ; 4-byte Folded Reload
+; GFX10-SCRATCH-NEXT: s_clause 0x1
+; GFX10-SCRATCH-NEXT: scratch_load_dword v40, off, s33
+; GFX10-SCRATCH-NEXT: scratch_load_dword v41, off, s33 offset:4
; GFX10-SCRATCH-NEXT: s_waitcnt_depctr 0xffe3
; GFX10-SCRATCH-NEXT: s_mov_b32 exec_lo, s1
; GFX10-SCRATCH-NEXT: s_add_i32 s32, s32, -16
@@ -9158,8 +9913,8 @@
; GFX9-NEXT: s_mov_b32 s33, s32
; GFX9-NEXT: s_or_saveexec_b64 s[36:37], -1
; GFX9-NEXT: buffer_store_dword v40, off, s[0:3], s33 ; 4-byte Folded Spill
+; GFX9-NEXT: buffer_store_dword v41, off, s[0:3], s33 offset:4 ; 4-byte Folded Spill
; GFX9-NEXT: s_mov_b64 exec, s[36:37]
-; GFX9-NEXT: v_writelane_b32 v40, s34, 6
; GFX9-NEXT: v_writelane_b32 v40, s4, 0
; GFX9-NEXT: v_writelane_b32 v40, s5, 1
; GFX9-NEXT: v_writelane_b32 v40, s6, 2
@@ -9170,6 +9925,7 @@
; GFX9-NEXT: s_mov_b32 s5, 2.0
; GFX9-NEXT: s_mov_b32 s6, 0
; GFX9-NEXT: s_mov_b32 s7, 0x40100000
+; GFX9-NEXT: v_writelane_b32 v41, s34, 0
; GFX9-NEXT: v_writelane_b32 v40, s31, 5
; GFX9-NEXT: s_getpc_b64 s[34:35]
; GFX9-NEXT: s_add_u32 s34, s34, external_void_func_v2f64_inreg@rel32@lo+4
@@ -9181,9 +9937,10 @@
; GFX9-NEXT: v_readlane_b32 s6, v40, 2
; GFX9-NEXT: v_readlane_b32 s5, v40, 1
; GFX9-NEXT: v_readlane_b32 s4, v40, 0
-; GFX9-NEXT: v_readlane_b32 s34, v40, 6
+; GFX9-NEXT: v_readlane_b32 s34, v41, 0
; GFX9-NEXT: s_or_saveexec_b64 s[36:37], -1
; GFX9-NEXT: buffer_load_dword v40, off, s[0:3], s33 ; 4-byte Folded Reload
+; GFX9-NEXT: buffer_load_dword v41, off, s[0:3], s33 offset:4 ; 4-byte Folded Reload
; GFX9-NEXT: s_mov_b64 exec, s[36:37]
; GFX9-NEXT: s_addk_i32 s32, 0xfc00
; GFX9-NEXT: s_mov_b32 s33, s34
@@ -9197,15 +9954,16 @@
; GFX10-NEXT: s_mov_b32 s33, s32
; GFX10-NEXT: s_or_saveexec_b32 s35, -1
; GFX10-NEXT: buffer_store_dword v40, off, s[0:3], s33 ; 4-byte Folded Spill
+; GFX10-NEXT: buffer_store_dword v41, off, s[0:3], s33 offset:4 ; 4-byte Folded Spill
; GFX10-NEXT: s_waitcnt_depctr 0xffe3
; GFX10-NEXT: s_mov_b32 exec_lo, s35
-; GFX10-NEXT: v_writelane_b32 v40, s34, 6
+; GFX10-NEXT: v_writelane_b32 v40, s4, 0
+; GFX10-NEXT: s_mov_b32 s4, 0
; GFX10-NEXT: s_addk_i32 s32, 0x200
+; GFX10-NEXT: v_writelane_b32 v41, s34, 0
; GFX10-NEXT: s_getpc_b64 s[34:35]
; GFX10-NEXT: s_add_u32 s34, s34, external_void_func_v2f64_inreg@rel32@lo+4
; GFX10-NEXT: s_addc_u32 s35, s35, external_void_func_v2f64_inreg@rel32@hi+12
-; GFX10-NEXT: v_writelane_b32 v40, s4, 0
-; GFX10-NEXT: s_mov_b32 s4, 0
; GFX10-NEXT: v_writelane_b32 v40, s5, 1
; GFX10-NEXT: s_mov_b32 s5, 2.0
; GFX10-NEXT: v_writelane_b32 v40, s6, 2
@@ -9221,9 +9979,11 @@
; GFX10-NEXT: v_readlane_b32 s6, v40, 2
; GFX10-NEXT: v_readlane_b32 s5, v40, 1
; GFX10-NEXT: v_readlane_b32 s4, v40, 0
-; GFX10-NEXT: v_readlane_b32 s34, v40, 6
+; GFX10-NEXT: v_readlane_b32 s34, v41, 0
; GFX10-NEXT: s_or_saveexec_b32 s35, -1
-; GFX10-NEXT: buffer_load_dword v40, off, s[0:3], s33 ; 4-byte Folded Reload
+; GFX10-NEXT: s_clause 0x1
+; GFX10-NEXT: buffer_load_dword v40, off, s[0:3], s33
+; GFX10-NEXT: buffer_load_dword v41, off, s[0:3], s33 offset:4
; GFX10-NEXT: s_waitcnt_depctr 0xffe3
; GFX10-NEXT: s_mov_b32 exec_lo, s35
; GFX10-NEXT: s_addk_i32 s32, 0xfe00
@@ -9237,15 +9997,17 @@
; GFX11-NEXT: s_mov_b32 s0, s33
; GFX11-NEXT: s_mov_b32 s33, s32
; GFX11-NEXT: s_or_saveexec_b32 s1, -1
-; GFX11-NEXT: scratch_store_b32 off, v40, s33 ; 4-byte Folded Spill
+; GFX11-NEXT: s_clause 0x1
+; GFX11-NEXT: scratch_store_b32 off, v40, s33
+; GFX11-NEXT: scratch_store_b32 off, v41, s33 offset:4
; GFX11-NEXT: s_mov_b32 exec_lo, s1
-; GFX11-NEXT: v_writelane_b32 v40, s0, 6
+; GFX11-NEXT: v_writelane_b32 v40, s4, 0
+; GFX11-NEXT: s_mov_b32 s4, 0
; GFX11-NEXT: s_add_i32 s32, s32, 16
+; GFX11-NEXT: v_writelane_b32 v41, s0, 0
; GFX11-NEXT: s_getpc_b64 s[0:1]
; GFX11-NEXT: s_add_u32 s0, s0, external_void_func_v2f64_inreg@rel32@lo+4
; GFX11-NEXT: s_addc_u32 s1, s1, external_void_func_v2f64_inreg@rel32@hi+12
-; GFX11-NEXT: v_writelane_b32 v40, s4, 0
-; GFX11-NEXT: s_mov_b32 s4, 0
; GFX11-NEXT: v_writelane_b32 v40, s5, 1
; GFX11-NEXT: s_mov_b32 s5, 2.0
; GFX11-NEXT: v_writelane_b32 v40, s6, 2
@@ -9262,9 +10024,11 @@
; GFX11-NEXT: v_readlane_b32 s6, v40, 2
; GFX11-NEXT: v_readlane_b32 s5, v40, 1
; GFX11-NEXT: v_readlane_b32 s4, v40, 0
-; GFX11-NEXT: v_readlane_b32 s0, v40, 6
+; GFX11-NEXT: v_readlane_b32 s0, v41, 0
; GFX11-NEXT: s_or_saveexec_b32 s1, -1
-; GFX11-NEXT: scratch_load_b32 v40, off, s33 ; 4-byte Folded Reload
+; GFX11-NEXT: s_clause 0x1
+; GFX11-NEXT: scratch_load_b32 v40, off, s33
+; GFX11-NEXT: scratch_load_b32 v41, off, s33 offset:4
; GFX11-NEXT: s_mov_b32 exec_lo, s1
; GFX11-NEXT: s_add_i32 s32, s32, -16
; GFX11-NEXT: s_mov_b32 s33, s0
@@ -9278,15 +10042,16 @@
; GFX10-SCRATCH-NEXT: s_mov_b32 s33, s32
; GFX10-SCRATCH-NEXT: s_or_saveexec_b32 s1, -1
; GFX10-SCRATCH-NEXT: scratch_store_dword off, v40, s33 ; 4-byte Folded Spill
+; GFX10-SCRATCH-NEXT: scratch_store_dword off, v41, s33 offset:4 ; 4-byte Folded Spill
; GFX10-SCRATCH-NEXT: s_waitcnt_depctr 0xffe3
; GFX10-SCRATCH-NEXT: s_mov_b32 exec_lo, s1
-; GFX10-SCRATCH-NEXT: v_writelane_b32 v40, s0, 6
+; GFX10-SCRATCH-NEXT: v_writelane_b32 v40, s4, 0
+; GFX10-SCRATCH-NEXT: s_mov_b32 s4, 0
; GFX10-SCRATCH-NEXT: s_add_i32 s32, s32, 16
+; GFX10-SCRATCH-NEXT: v_writelane_b32 v41, s0, 0
; GFX10-SCRATCH-NEXT: s_getpc_b64 s[0:1]
; GFX10-SCRATCH-NEXT: s_add_u32 s0, s0, external_void_func_v2f64_inreg@rel32@lo+4
; GFX10-SCRATCH-NEXT: s_addc_u32 s1, s1, external_void_func_v2f64_inreg@rel32@hi+12
-; GFX10-SCRATCH-NEXT: v_writelane_b32 v40, s4, 0
-; GFX10-SCRATCH-NEXT: s_mov_b32 s4, 0
; GFX10-SCRATCH-NEXT: v_writelane_b32 v40, s5, 1
; GFX10-SCRATCH-NEXT: s_mov_b32 s5, 2.0
; GFX10-SCRATCH-NEXT: v_writelane_b32 v40, s6, 2
@@ -9302,9 +10067,11 @@
; GFX10-SCRATCH-NEXT: v_readlane_b32 s6, v40, 2
; GFX10-SCRATCH-NEXT: v_readlane_b32 s5, v40, 1
; GFX10-SCRATCH-NEXT: v_readlane_b32 s4, v40, 0
-; GFX10-SCRATCH-NEXT: v_readlane_b32 s0, v40, 6
+; GFX10-SCRATCH-NEXT: v_readlane_b32 s0, v41, 0
; GFX10-SCRATCH-NEXT: s_or_saveexec_b32 s1, -1
-; GFX10-SCRATCH-NEXT: scratch_load_dword v40, off, s33 ; 4-byte Folded Reload
+; GFX10-SCRATCH-NEXT: s_clause 0x1
+; GFX10-SCRATCH-NEXT: scratch_load_dword v40, off, s33
+; GFX10-SCRATCH-NEXT: scratch_load_dword v41, off, s33 offset:4
; GFX10-SCRATCH-NEXT: s_waitcnt_depctr 0xffe3
; GFX10-SCRATCH-NEXT: s_mov_b32 exec_lo, s1
; GFX10-SCRATCH-NEXT: s_add_i32 s32, s32, -16
@@ -9323,8 +10090,8 @@
; GFX9-NEXT: s_mov_b32 s33, s32
; GFX9-NEXT: s_or_saveexec_b64 s[36:37], -1
; GFX9-NEXT: buffer_store_dword v40, off, s[0:3], s33 ; 4-byte Folded Spill
+; GFX9-NEXT: buffer_store_dword v41, off, s[0:3], s33 offset:4 ; 4-byte Folded Spill
; GFX9-NEXT: s_mov_b64 exec, s[36:37]
-; GFX9-NEXT: v_writelane_b32 v40, s34, 8
; GFX9-NEXT: v_writelane_b32 v40, s4, 0
; GFX9-NEXT: v_writelane_b32 v40, s5, 1
; GFX9-NEXT: v_writelane_b32 v40, s6, 2
@@ -9339,6 +10106,7 @@
; GFX9-NEXT: s_mov_b32 s7, 0x40100000
; GFX9-NEXT: s_mov_b32 s8, 0
; GFX9-NEXT: s_mov_b32 s9, 0x40200000
+; GFX9-NEXT: v_writelane_b32 v41, s34, 0
; GFX9-NEXT: v_writelane_b32 v40, s31, 7
; GFX9-NEXT: s_getpc_b64 s[34:35]
; GFX9-NEXT: s_add_u32 s34, s34, external_void_func_v3f64_inreg@rel32@lo+4
@@ -9352,9 +10120,10 @@
; GFX9-NEXT: v_readlane_b32 s6, v40, 2
; GFX9-NEXT: v_readlane_b32 s5, v40, 1
; GFX9-NEXT: v_readlane_b32 s4, v40, 0
-; GFX9-NEXT: v_readlane_b32 s34, v40, 8
+; GFX9-NEXT: v_readlane_b32 s34, v41, 0
; GFX9-NEXT: s_or_saveexec_b64 s[36:37], -1
; GFX9-NEXT: buffer_load_dword v40, off, s[0:3], s33 ; 4-byte Folded Reload
+; GFX9-NEXT: buffer_load_dword v41, off, s[0:3], s33 offset:4 ; 4-byte Folded Reload
; GFX9-NEXT: s_mov_b64 exec, s[36:37]
; GFX9-NEXT: s_addk_i32 s32, 0xfc00
; GFX9-NEXT: s_mov_b32 s33, s34
@@ -9368,15 +10137,16 @@
; GFX10-NEXT: s_mov_b32 s33, s32
; GFX10-NEXT: s_or_saveexec_b32 s35, -1
; GFX10-NEXT: buffer_store_dword v40, off, s[0:3], s33 ; 4-byte Folded Spill
+; GFX10-NEXT: buffer_store_dword v41, off, s[0:3], s33 offset:4 ; 4-byte Folded Spill
; GFX10-NEXT: s_waitcnt_depctr 0xffe3
; GFX10-NEXT: s_mov_b32 exec_lo, s35
-; GFX10-NEXT: v_writelane_b32 v40, s34, 8
+; GFX10-NEXT: v_writelane_b32 v40, s4, 0
+; GFX10-NEXT: s_mov_b32 s4, 0
; GFX10-NEXT: s_addk_i32 s32, 0x200
+; GFX10-NEXT: v_writelane_b32 v41, s34, 0
; GFX10-NEXT: s_getpc_b64 s[34:35]
; GFX10-NEXT: s_add_u32 s34, s34, external_void_func_v3f64_inreg@rel32@lo+4
; GFX10-NEXT: s_addc_u32 s35, s35, external_void_func_v3f64_inreg@rel32@hi+12
-; GFX10-NEXT: v_writelane_b32 v40, s4, 0
-; GFX10-NEXT: s_mov_b32 s4, 0
; GFX10-NEXT: v_writelane_b32 v40, s5, 1
; GFX10-NEXT: s_mov_b32 s5, 2.0
; GFX10-NEXT: v_writelane_b32 v40, s6, 2
@@ -9398,9 +10168,11 @@
; GFX10-NEXT: v_readlane_b32 s6, v40, 2
; GFX10-NEXT: v_readlane_b32 s5, v40, 1
; GFX10-NEXT: v_readlane_b32 s4, v40, 0
-; GFX10-NEXT: v_readlane_b32 s34, v40, 8
+; GFX10-NEXT: v_readlane_b32 s34, v41, 0
; GFX10-NEXT: s_or_saveexec_b32 s35, -1
-; GFX10-NEXT: buffer_load_dword v40, off, s[0:3], s33 ; 4-byte Folded Reload
+; GFX10-NEXT: s_clause 0x1
+; GFX10-NEXT: buffer_load_dword v40, off, s[0:3], s33
+; GFX10-NEXT: buffer_load_dword v41, off, s[0:3], s33 offset:4
; GFX10-NEXT: s_waitcnt_depctr 0xffe3
; GFX10-NEXT: s_mov_b32 exec_lo, s35
; GFX10-NEXT: s_addk_i32 s32, 0xfe00
@@ -9414,15 +10186,17 @@
; GFX11-NEXT: s_mov_b32 s0, s33
; GFX11-NEXT: s_mov_b32 s33, s32
; GFX11-NEXT: s_or_saveexec_b32 s1, -1
-; GFX11-NEXT: scratch_store_b32 off, v40, s33 ; 4-byte Folded Spill
+; GFX11-NEXT: s_clause 0x1
+; GFX11-NEXT: scratch_store_b32 off, v40, s33
+; GFX11-NEXT: scratch_store_b32 off, v41, s33 offset:4
; GFX11-NEXT: s_mov_b32 exec_lo, s1
-; GFX11-NEXT: v_writelane_b32 v40, s0, 8
+; GFX11-NEXT: v_writelane_b32 v40, s4, 0
+; GFX11-NEXT: s_mov_b32 s4, 0
; GFX11-NEXT: s_add_i32 s32, s32, 16
+; GFX11-NEXT: v_writelane_b32 v41, s0, 0
; GFX11-NEXT: s_getpc_b64 s[0:1]
; GFX11-NEXT: s_add_u32 s0, s0, external_void_func_v3f64_inreg@rel32@lo+4
; GFX11-NEXT: s_addc_u32 s1, s1, external_void_func_v3f64_inreg@rel32@hi+12
-; GFX11-NEXT: v_writelane_b32 v40, s4, 0
-; GFX11-NEXT: s_mov_b32 s4, 0
; GFX11-NEXT: v_writelane_b32 v40, s5, 1
; GFX11-NEXT: s_mov_b32 s5, 2.0
; GFX11-NEXT: v_writelane_b32 v40, s6, 2
@@ -9445,9 +10219,11 @@
; GFX11-NEXT: v_readlane_b32 s6, v40, 2
; GFX11-NEXT: v_readlane_b32 s5, v40, 1
; GFX11-NEXT: v_readlane_b32 s4, v40, 0
-; GFX11-NEXT: v_readlane_b32 s0, v40, 8
+; GFX11-NEXT: v_readlane_b32 s0, v41, 0
; GFX11-NEXT: s_or_saveexec_b32 s1, -1
-; GFX11-NEXT: scratch_load_b32 v40, off, s33 ; 4-byte Folded Reload
+; GFX11-NEXT: s_clause 0x1
+; GFX11-NEXT: scratch_load_b32 v40, off, s33
+; GFX11-NEXT: scratch_load_b32 v41, off, s33 offset:4
; GFX11-NEXT: s_mov_b32 exec_lo, s1
; GFX11-NEXT: s_add_i32 s32, s32, -16
; GFX11-NEXT: s_mov_b32 s33, s0
@@ -9461,15 +10237,16 @@
; GFX10-SCRATCH-NEXT: s_mov_b32 s33, s32
; GFX10-SCRATCH-NEXT: s_or_saveexec_b32 s1, -1
; GFX10-SCRATCH-NEXT: scratch_store_dword off, v40, s33 ; 4-byte Folded Spill
+; GFX10-SCRATCH-NEXT: scratch_store_dword off, v41, s33 offset:4 ; 4-byte Folded Spill
; GFX10-SCRATCH-NEXT: s_waitcnt_depctr 0xffe3
; GFX10-SCRATCH-NEXT: s_mov_b32 exec_lo, s1
-; GFX10-SCRATCH-NEXT: v_writelane_b32 v40, s0, 8
+; GFX10-SCRATCH-NEXT: v_writelane_b32 v40, s4, 0
+; GFX10-SCRATCH-NEXT: s_mov_b32 s4, 0
; GFX10-SCRATCH-NEXT: s_add_i32 s32, s32, 16
+; GFX10-SCRATCH-NEXT: v_writelane_b32 v41, s0, 0
; GFX10-SCRATCH-NEXT: s_getpc_b64 s[0:1]
; GFX10-SCRATCH-NEXT: s_add_u32 s0, s0, external_void_func_v3f64_inreg@rel32@lo+4
; GFX10-SCRATCH-NEXT: s_addc_u32 s1, s1, external_void_func_v3f64_inreg@rel32@hi+12
-; GFX10-SCRATCH-NEXT: v_writelane_b32 v40, s4, 0
-; GFX10-SCRATCH-NEXT: s_mov_b32 s4, 0
; GFX10-SCRATCH-NEXT: v_writelane_b32 v40, s5, 1
; GFX10-SCRATCH-NEXT: s_mov_b32 s5, 2.0
; GFX10-SCRATCH-NEXT: v_writelane_b32 v40, s6, 2
@@ -9491,9 +10268,11 @@
; GFX10-SCRATCH-NEXT: v_readlane_b32 s6, v40, 2
; GFX10-SCRATCH-NEXT: v_readlane_b32 s5, v40, 1
; GFX10-SCRATCH-NEXT: v_readlane_b32 s4, v40, 0
-; GFX10-SCRATCH-NEXT: v_readlane_b32 s0, v40, 8
+; GFX10-SCRATCH-NEXT: v_readlane_b32 s0, v41, 0
; GFX10-SCRATCH-NEXT: s_or_saveexec_b32 s1, -1
-; GFX10-SCRATCH-NEXT: scratch_load_dword v40, off, s33 ; 4-byte Folded Reload
+; GFX10-SCRATCH-NEXT: s_clause 0x1
+; GFX10-SCRATCH-NEXT: scratch_load_dword v40, off, s33
+; GFX10-SCRATCH-NEXT: scratch_load_dword v41, off, s33 offset:4
; GFX10-SCRATCH-NEXT: s_waitcnt_depctr 0xffe3
; GFX10-SCRATCH-NEXT: s_mov_b32 exec_lo, s1
; GFX10-SCRATCH-NEXT: s_add_i32 s32, s32, -16
@@ -9512,12 +10291,13 @@
; GFX9-NEXT: s_mov_b32 s33, s32
; GFX9-NEXT: s_or_saveexec_b64 s[36:37], -1
; GFX9-NEXT: buffer_store_dword v40, off, s[0:3], s33 ; 4-byte Folded Spill
+; GFX9-NEXT: buffer_store_dword v41, off, s[0:3], s33 offset:4 ; 4-byte Folded Spill
; GFX9-NEXT: s_mov_b64 exec, s[36:37]
-; GFX9-NEXT: v_writelane_b32 v40, s34, 3
; GFX9-NEXT: v_writelane_b32 v40, s4, 0
; GFX9-NEXT: s_load_dword s4, s[34:35], 0x0
; GFX9-NEXT: s_addk_i32 s32, 0x400
; GFX9-NEXT: v_writelane_b32 v40, s30, 1
+; GFX9-NEXT: v_writelane_b32 v41, s34, 0
; GFX9-NEXT: v_writelane_b32 v40, s31, 2
; GFX9-NEXT: s_getpc_b64 s[34:35]
; GFX9-NEXT: s_add_u32 s34, s34, external_void_func_v2i16_inreg@rel32@lo+4
@@ -9526,9 +10306,10 @@
; GFX9-NEXT: v_readlane_b32 s31, v40, 2
; GFX9-NEXT: v_readlane_b32 s30, v40, 1
; GFX9-NEXT: v_readlane_b32 s4, v40, 0
-; GFX9-NEXT: v_readlane_b32 s34, v40, 3
+; GFX9-NEXT: v_readlane_b32 s34, v41, 0
; GFX9-NEXT: s_or_saveexec_b64 s[36:37], -1
; GFX9-NEXT: buffer_load_dword v40, off, s[0:3], s33 ; 4-byte Folded Reload
+; GFX9-NEXT: buffer_load_dword v41, off, s[0:3], s33 offset:4 ; 4-byte Folded Reload
; GFX9-NEXT: s_mov_b64 exec, s[36:37]
; GFX9-NEXT: s_addk_i32 s32, 0xfc00
; GFX9-NEXT: s_mov_b32 s33, s34
@@ -9542,12 +10323,13 @@
; GFX10-NEXT: s_mov_b32 s33, s32
; GFX10-NEXT: s_or_saveexec_b32 s35, -1
; GFX10-NEXT: buffer_store_dword v40, off, s[0:3], s33 ; 4-byte Folded Spill
+; GFX10-NEXT: buffer_store_dword v41, off, s[0:3], s33 offset:4 ; 4-byte Folded Spill
; GFX10-NEXT: s_waitcnt_depctr 0xffe3
; GFX10-NEXT: s_mov_b32 exec_lo, s35
-; GFX10-NEXT: v_writelane_b32 v40, s34, 3
-; GFX10-NEXT: s_addk_i32 s32, 0x200
; GFX10-NEXT: v_writelane_b32 v40, s4, 0
; GFX10-NEXT: s_load_dword s4, s[34:35], 0x0
+; GFX10-NEXT: s_addk_i32 s32, 0x200
+; GFX10-NEXT: v_writelane_b32 v41, s34, 0
; GFX10-NEXT: s_getpc_b64 s[34:35]
; GFX10-NEXT: s_add_u32 s34, s34, external_void_func_v2i16_inreg@rel32@lo+4
; GFX10-NEXT: s_addc_u32 s35, s35, external_void_func_v2i16_inreg@rel32@hi+12
@@ -9557,9 +10339,11 @@
; GFX10-NEXT: v_readlane_b32 s31, v40, 2
; GFX10-NEXT: v_readlane_b32 s30, v40, 1
; GFX10-NEXT: v_readlane_b32 s4, v40, 0
-; GFX10-NEXT: v_readlane_b32 s34, v40, 3
+; GFX10-NEXT: v_readlane_b32 s34, v41, 0
; GFX10-NEXT: s_or_saveexec_b32 s35, -1
-; GFX10-NEXT: buffer_load_dword v40, off, s[0:3], s33 ; 4-byte Folded Reload
+; GFX10-NEXT: s_clause 0x1
+; GFX10-NEXT: buffer_load_dword v40, off, s[0:3], s33
+; GFX10-NEXT: buffer_load_dword v41, off, s[0:3], s33 offset:4
; GFX10-NEXT: s_waitcnt_depctr 0xffe3
; GFX10-NEXT: s_mov_b32 exec_lo, s35
; GFX10-NEXT: s_addk_i32 s32, 0xfe00
@@ -9573,12 +10357,14 @@
; GFX11-NEXT: s_mov_b32 s0, s33
; GFX11-NEXT: s_mov_b32 s33, s32
; GFX11-NEXT: s_or_saveexec_b32 s1, -1
-; GFX11-NEXT: scratch_store_b32 off, v40, s33 ; 4-byte Folded Spill
+; GFX11-NEXT: s_clause 0x1
+; GFX11-NEXT: scratch_store_b32 off, v40, s33
+; GFX11-NEXT: scratch_store_b32 off, v41, s33 offset:4
; GFX11-NEXT: s_mov_b32 exec_lo, s1
-; GFX11-NEXT: v_writelane_b32 v40, s0, 3
-; GFX11-NEXT: s_add_i32 s32, s32, 16
; GFX11-NEXT: v_writelane_b32 v40, s4, 0
; GFX11-NEXT: s_load_b32 s4, s[0:1], 0x0
+; GFX11-NEXT: s_add_i32 s32, s32, 16
+; GFX11-NEXT: v_writelane_b32 v41, s0, 0
; GFX11-NEXT: s_getpc_b64 s[0:1]
; GFX11-NEXT: s_add_u32 s0, s0, external_void_func_v2i16_inreg@rel32@lo+4
; GFX11-NEXT: s_addc_u32 s1, s1, external_void_func_v2i16_inreg@rel32@hi+12
@@ -9589,9 +10375,11 @@
; GFX11-NEXT: v_readlane_b32 s31, v40, 2
; GFX11-NEXT: v_readlane_b32 s30, v40, 1
; GFX11-NEXT: v_readlane_b32 s4, v40, 0
-; GFX11-NEXT: v_readlane_b32 s0, v40, 3
+; GFX11-NEXT: v_readlane_b32 s0, v41, 0
; GFX11-NEXT: s_or_saveexec_b32 s1, -1
-; GFX11-NEXT: scratch_load_b32 v40, off, s33 ; 4-byte Folded Reload
+; GFX11-NEXT: s_clause 0x1
+; GFX11-NEXT: scratch_load_b32 v40, off, s33
+; GFX11-NEXT: scratch_load_b32 v41, off, s33 offset:4
; GFX11-NEXT: s_mov_b32 exec_lo, s1
; GFX11-NEXT: s_add_i32 s32, s32, -16
; GFX11-NEXT: s_mov_b32 s33, s0
@@ -9605,12 +10393,13 @@
; GFX10-SCRATCH-NEXT: s_mov_b32 s33, s32
; GFX10-SCRATCH-NEXT: s_or_saveexec_b32 s1, -1
; GFX10-SCRATCH-NEXT: scratch_store_dword off, v40, s33 ; 4-byte Folded Spill
+; GFX10-SCRATCH-NEXT: scratch_store_dword off, v41, s33 offset:4 ; 4-byte Folded Spill
; GFX10-SCRATCH-NEXT: s_waitcnt_depctr 0xffe3
; GFX10-SCRATCH-NEXT: s_mov_b32 exec_lo, s1
-; GFX10-SCRATCH-NEXT: v_writelane_b32 v40, s0, 3
-; GFX10-SCRATCH-NEXT: s_add_i32 s32, s32, 16
; GFX10-SCRATCH-NEXT: v_writelane_b32 v40, s4, 0
; GFX10-SCRATCH-NEXT: s_load_dword s4, s[0:1], 0x0
+; GFX10-SCRATCH-NEXT: s_add_i32 s32, s32, 16
+; GFX10-SCRATCH-NEXT: v_writelane_b32 v41, s0, 0
; GFX10-SCRATCH-NEXT: s_getpc_b64 s[0:1]
; GFX10-SCRATCH-NEXT: s_add_u32 s0, s0, external_void_func_v2i16_inreg@rel32@lo+4
; GFX10-SCRATCH-NEXT: s_addc_u32 s1, s1, external_void_func_v2i16_inreg@rel32@hi+12
@@ -9620,9 +10409,11 @@
; GFX10-SCRATCH-NEXT: v_readlane_b32 s31, v40, 2
; GFX10-SCRATCH-NEXT: v_readlane_b32 s30, v40, 1
; GFX10-SCRATCH-NEXT: v_readlane_b32 s4, v40, 0
-; GFX10-SCRATCH-NEXT: v_readlane_b32 s0, v40, 3
+; GFX10-SCRATCH-NEXT: v_readlane_b32 s0, v41, 0
; GFX10-SCRATCH-NEXT: s_or_saveexec_b32 s1, -1
-; GFX10-SCRATCH-NEXT: scratch_load_dword v40, off, s33 ; 4-byte Folded Reload
+; GFX10-SCRATCH-NEXT: s_clause 0x1
+; GFX10-SCRATCH-NEXT: scratch_load_dword v40, off, s33
+; GFX10-SCRATCH-NEXT: scratch_load_dword v41, off, s33 offset:4
; GFX10-SCRATCH-NEXT: s_waitcnt_depctr 0xffe3
; GFX10-SCRATCH-NEXT: s_mov_b32 exec_lo, s1
; GFX10-SCRATCH-NEXT: s_add_i32 s32, s32, -16
@@ -9642,13 +10433,14 @@
; GFX9-NEXT: s_mov_b32 s33, s32
; GFX9-NEXT: s_or_saveexec_b64 s[36:37], -1
; GFX9-NEXT: buffer_store_dword v40, off, s[0:3], s33 ; 4-byte Folded Spill
+; GFX9-NEXT: buffer_store_dword v41, off, s[0:3], s33 offset:4 ; 4-byte Folded Spill
; GFX9-NEXT: s_mov_b64 exec, s[36:37]
-; GFX9-NEXT: v_writelane_b32 v40, s34, 4
; GFX9-NEXT: v_writelane_b32 v40, s4, 0
; GFX9-NEXT: v_writelane_b32 v40, s5, 1
; GFX9-NEXT: s_load_dwordx2 s[4:5], s[34:35], 0x0
; GFX9-NEXT: s_addk_i32 s32, 0x400
; GFX9-NEXT: v_writelane_b32 v40, s30, 2
+; GFX9-NEXT: v_writelane_b32 v41, s34, 0
; GFX9-NEXT: v_writelane_b32 v40, s31, 3
; GFX9-NEXT: s_getpc_b64 s[34:35]
; GFX9-NEXT: s_add_u32 s34, s34, external_void_func_v3i16_inreg@rel32@lo+4
@@ -9658,9 +10450,10 @@
; GFX9-NEXT: v_readlane_b32 s30, v40, 2
; GFX9-NEXT: v_readlane_b32 s5, v40, 1
; GFX9-NEXT: v_readlane_b32 s4, v40, 0
-; GFX9-NEXT: v_readlane_b32 s34, v40, 4
+; GFX9-NEXT: v_readlane_b32 s34, v41, 0
; GFX9-NEXT: s_or_saveexec_b64 s[36:37], -1
; GFX9-NEXT: buffer_load_dword v40, off, s[0:3], s33 ; 4-byte Folded Reload
+; GFX9-NEXT: buffer_load_dword v41, off, s[0:3], s33 offset:4 ; 4-byte Folded Reload
; GFX9-NEXT: s_mov_b64 exec, s[36:37]
; GFX9-NEXT: s_addk_i32 s32, 0xfc00
; GFX9-NEXT: s_mov_b32 s33, s34
@@ -9674,11 +10467,12 @@
; GFX10-NEXT: s_mov_b32 s33, s32
; GFX10-NEXT: s_or_saveexec_b32 s35, -1
; GFX10-NEXT: buffer_store_dword v40, off, s[0:3], s33 ; 4-byte Folded Spill
+; GFX10-NEXT: buffer_store_dword v41, off, s[0:3], s33 offset:4 ; 4-byte Folded Spill
; GFX10-NEXT: s_waitcnt_depctr 0xffe3
; GFX10-NEXT: s_mov_b32 exec_lo, s35
-; GFX10-NEXT: v_writelane_b32 v40, s34, 4
-; GFX10-NEXT: s_addk_i32 s32, 0x200
; GFX10-NEXT: v_writelane_b32 v40, s4, 0
+; GFX10-NEXT: s_addk_i32 s32, 0x200
+; GFX10-NEXT: v_writelane_b32 v41, s34, 0
; GFX10-NEXT: v_writelane_b32 v40, s5, 1
; GFX10-NEXT: s_load_dwordx2 s[4:5], s[34:35], 0x0
; GFX10-NEXT: s_getpc_b64 s[34:35]
@@ -9691,9 +10485,11 @@
; GFX10-NEXT: v_readlane_b32 s30, v40, 2
; GFX10-NEXT: v_readlane_b32 s5, v40, 1
; GFX10-NEXT: v_readlane_b32 s4, v40, 0
-; GFX10-NEXT: v_readlane_b32 s34, v40, 4
+; GFX10-NEXT: v_readlane_b32 s34, v41, 0
; GFX10-NEXT: s_or_saveexec_b32 s35, -1
-; GFX10-NEXT: buffer_load_dword v40, off, s[0:3], s33 ; 4-byte Folded Reload
+; GFX10-NEXT: s_clause 0x1
+; GFX10-NEXT: buffer_load_dword v40, off, s[0:3], s33
+; GFX10-NEXT: buffer_load_dword v41, off, s[0:3], s33 offset:4
; GFX10-NEXT: s_waitcnt_depctr 0xffe3
; GFX10-NEXT: s_mov_b32 exec_lo, s35
; GFX10-NEXT: s_addk_i32 s32, 0xfe00
@@ -9707,11 +10503,13 @@
; GFX11-NEXT: s_mov_b32 s0, s33
; GFX11-NEXT: s_mov_b32 s33, s32
; GFX11-NEXT: s_or_saveexec_b32 s1, -1
-; GFX11-NEXT: scratch_store_b32 off, v40, s33 ; 4-byte Folded Spill
+; GFX11-NEXT: s_clause 0x1
+; GFX11-NEXT: scratch_store_b32 off, v40, s33
+; GFX11-NEXT: scratch_store_b32 off, v41, s33 offset:4
; GFX11-NEXT: s_mov_b32 exec_lo, s1
-; GFX11-NEXT: v_writelane_b32 v40, s0, 4
-; GFX11-NEXT: s_add_i32 s32, s32, 16
; GFX11-NEXT: v_writelane_b32 v40, s4, 0
+; GFX11-NEXT: s_add_i32 s32, s32, 16
+; GFX11-NEXT: v_writelane_b32 v41, s0, 0
; GFX11-NEXT: v_writelane_b32 v40, s5, 1
; GFX11-NEXT: s_load_b64 s[4:5], s[0:1], 0x0
; GFX11-NEXT: s_getpc_b64 s[0:1]
@@ -9725,9 +10523,11 @@
; GFX11-NEXT: v_readlane_b32 s30, v40, 2
; GFX11-NEXT: v_readlane_b32 s5, v40, 1
; GFX11-NEXT: v_readlane_b32 s4, v40, 0
-; GFX11-NEXT: v_readlane_b32 s0, v40, 4
+; GFX11-NEXT: v_readlane_b32 s0, v41, 0
; GFX11-NEXT: s_or_saveexec_b32 s1, -1
-; GFX11-NEXT: scratch_load_b32 v40, off, s33 ; 4-byte Folded Reload
+; GFX11-NEXT: s_clause 0x1
+; GFX11-NEXT: scratch_load_b32 v40, off, s33
+; GFX11-NEXT: scratch_load_b32 v41, off, s33 offset:4
; GFX11-NEXT: s_mov_b32 exec_lo, s1
; GFX11-NEXT: s_add_i32 s32, s32, -16
; GFX11-NEXT: s_mov_b32 s33, s0
@@ -9741,11 +10541,12 @@
; GFX10-SCRATCH-NEXT: s_mov_b32 s33, s32
; GFX10-SCRATCH-NEXT: s_or_saveexec_b32 s1, -1
; GFX10-SCRATCH-NEXT: scratch_store_dword off, v40, s33 ; 4-byte Folded Spill
+; GFX10-SCRATCH-NEXT: scratch_store_dword off, v41, s33 offset:4 ; 4-byte Folded Spill
; GFX10-SCRATCH-NEXT: s_waitcnt_depctr 0xffe3
; GFX10-SCRATCH-NEXT: s_mov_b32 exec_lo, s1
-; GFX10-SCRATCH-NEXT: v_writelane_b32 v40, s0, 4
-; GFX10-SCRATCH-NEXT: s_add_i32 s32, s32, 16
; GFX10-SCRATCH-NEXT: v_writelane_b32 v40, s4, 0
+; GFX10-SCRATCH-NEXT: s_add_i32 s32, s32, 16
+; GFX10-SCRATCH-NEXT: v_writelane_b32 v41, s0, 0
; GFX10-SCRATCH-NEXT: v_writelane_b32 v40, s5, 1
; GFX10-SCRATCH-NEXT: s_load_dwordx2 s[4:5], s[0:1], 0x0
; GFX10-SCRATCH-NEXT: s_getpc_b64 s[0:1]
@@ -9758,9 +10559,11 @@
; GFX10-SCRATCH-NEXT: v_readlane_b32 s30, v40, 2
; GFX10-SCRATCH-NEXT: v_readlane_b32 s5, v40, 1
; GFX10-SCRATCH-NEXT: v_readlane_b32 s4, v40, 0
-; GFX10-SCRATCH-NEXT: v_readlane_b32 s0, v40, 4
+; GFX10-SCRATCH-NEXT: v_readlane_b32 s0, v41, 0
; GFX10-SCRATCH-NEXT: s_or_saveexec_b32 s1, -1
-; GFX10-SCRATCH-NEXT: scratch_load_dword v40, off, s33 ; 4-byte Folded Reload
+; GFX10-SCRATCH-NEXT: s_clause 0x1
+; GFX10-SCRATCH-NEXT: scratch_load_dword v40, off, s33
+; GFX10-SCRATCH-NEXT: scratch_load_dword v41, off, s33 offset:4
; GFX10-SCRATCH-NEXT: s_waitcnt_depctr 0xffe3
; GFX10-SCRATCH-NEXT: s_mov_b32 exec_lo, s1
; GFX10-SCRATCH-NEXT: s_add_i32 s32, s32, -16
@@ -9780,13 +10583,14 @@
; GFX9-NEXT: s_mov_b32 s33, s32
; GFX9-NEXT: s_or_saveexec_b64 s[36:37], -1
; GFX9-NEXT: buffer_store_dword v40, off, s[0:3], s33 ; 4-byte Folded Spill
+; GFX9-NEXT: buffer_store_dword v41, off, s[0:3], s33 offset:4 ; 4-byte Folded Spill
; GFX9-NEXT: s_mov_b64 exec, s[36:37]
-; GFX9-NEXT: v_writelane_b32 v40, s34, 4
; GFX9-NEXT: v_writelane_b32 v40, s4, 0
; GFX9-NEXT: v_writelane_b32 v40, s5, 1
; GFX9-NEXT: s_load_dwordx2 s[4:5], s[34:35], 0x0
; GFX9-NEXT: s_addk_i32 s32, 0x400
; GFX9-NEXT: v_writelane_b32 v40, s30, 2
+; GFX9-NEXT: v_writelane_b32 v41, s34, 0
; GFX9-NEXT: v_writelane_b32 v40, s31, 3
; GFX9-NEXT: s_getpc_b64 s[34:35]
; GFX9-NEXT: s_add_u32 s34, s34, external_void_func_v3f16_inreg@rel32@lo+4
@@ -9796,9 +10600,10 @@
; GFX9-NEXT: v_readlane_b32 s30, v40, 2
; GFX9-NEXT: v_readlane_b32 s5, v40, 1
; GFX9-NEXT: v_readlane_b32 s4, v40, 0
-; GFX9-NEXT: v_readlane_b32 s34, v40, 4
+; GFX9-NEXT: v_readlane_b32 s34, v41, 0
; GFX9-NEXT: s_or_saveexec_b64 s[36:37], -1
; GFX9-NEXT: buffer_load_dword v40, off, s[0:3], s33 ; 4-byte Folded Reload
+; GFX9-NEXT: buffer_load_dword v41, off, s[0:3], s33 offset:4 ; 4-byte Folded Reload
; GFX9-NEXT: s_mov_b64 exec, s[36:37]
; GFX9-NEXT: s_addk_i32 s32, 0xfc00
; GFX9-NEXT: s_mov_b32 s33, s34
@@ -9812,11 +10617,12 @@
; GFX10-NEXT: s_mov_b32 s33, s32
; GFX10-NEXT: s_or_saveexec_b32 s35, -1
; GFX10-NEXT: buffer_store_dword v40, off, s[0:3], s33 ; 4-byte Folded Spill
+; GFX10-NEXT: buffer_store_dword v41, off, s[0:3], s33 offset:4 ; 4-byte Folded Spill
; GFX10-NEXT: s_waitcnt_depctr 0xffe3
; GFX10-NEXT: s_mov_b32 exec_lo, s35
-; GFX10-NEXT: v_writelane_b32 v40, s34, 4
-; GFX10-NEXT: s_addk_i32 s32, 0x200
; GFX10-NEXT: v_writelane_b32 v40, s4, 0
+; GFX10-NEXT: s_addk_i32 s32, 0x200
+; GFX10-NEXT: v_writelane_b32 v41, s34, 0
; GFX10-NEXT: v_writelane_b32 v40, s5, 1
; GFX10-NEXT: s_load_dwordx2 s[4:5], s[34:35], 0x0
; GFX10-NEXT: s_getpc_b64 s[34:35]
@@ -9829,9 +10635,11 @@
; GFX10-NEXT: v_readlane_b32 s30, v40, 2
; GFX10-NEXT: v_readlane_b32 s5, v40, 1
; GFX10-NEXT: v_readlane_b32 s4, v40, 0
-; GFX10-NEXT: v_readlane_b32 s34, v40, 4
+; GFX10-NEXT: v_readlane_b32 s34, v41, 0
; GFX10-NEXT: s_or_saveexec_b32 s35, -1
-; GFX10-NEXT: buffer_load_dword v40, off, s[0:3], s33 ; 4-byte Folded Reload
+; GFX10-NEXT: s_clause 0x1
+; GFX10-NEXT: buffer_load_dword v40, off, s[0:3], s33
+; GFX10-NEXT: buffer_load_dword v41, off, s[0:3], s33 offset:4
; GFX10-NEXT: s_waitcnt_depctr 0xffe3
; GFX10-NEXT: s_mov_b32 exec_lo, s35
; GFX10-NEXT: s_addk_i32 s32, 0xfe00
@@ -9845,11 +10653,13 @@
; GFX11-NEXT: s_mov_b32 s0, s33
; GFX11-NEXT: s_mov_b32 s33, s32
; GFX11-NEXT: s_or_saveexec_b32 s1, -1
-; GFX11-NEXT: scratch_store_b32 off, v40, s33 ; 4-byte Folded Spill
+; GFX11-NEXT: s_clause 0x1
+; GFX11-NEXT: scratch_store_b32 off, v40, s33
+; GFX11-NEXT: scratch_store_b32 off, v41, s33 offset:4
; GFX11-NEXT: s_mov_b32 exec_lo, s1
-; GFX11-NEXT: v_writelane_b32 v40, s0, 4
-; GFX11-NEXT: s_add_i32 s32, s32, 16
; GFX11-NEXT: v_writelane_b32 v40, s4, 0
+; GFX11-NEXT: s_add_i32 s32, s32, 16
+; GFX11-NEXT: v_writelane_b32 v41, s0, 0
; GFX11-NEXT: v_writelane_b32 v40, s5, 1
; GFX11-NEXT: s_load_b64 s[4:5], s[0:1], 0x0
; GFX11-NEXT: s_getpc_b64 s[0:1]
@@ -9863,9 +10673,11 @@
; GFX11-NEXT: v_readlane_b32 s30, v40, 2
; GFX11-NEXT: v_readlane_b32 s5, v40, 1
; GFX11-NEXT: v_readlane_b32 s4, v40, 0
-; GFX11-NEXT: v_readlane_b32 s0, v40, 4
+; GFX11-NEXT: v_readlane_b32 s0, v41, 0
; GFX11-NEXT: s_or_saveexec_b32 s1, -1
-; GFX11-NEXT: scratch_load_b32 v40, off, s33 ; 4-byte Folded Reload
+; GFX11-NEXT: s_clause 0x1
+; GFX11-NEXT: scratch_load_b32 v40, off, s33
+; GFX11-NEXT: scratch_load_b32 v41, off, s33 offset:4
; GFX11-NEXT: s_mov_b32 exec_lo, s1
; GFX11-NEXT: s_add_i32 s32, s32, -16
; GFX11-NEXT: s_mov_b32 s33, s0
@@ -9879,11 +10691,12 @@
; GFX10-SCRATCH-NEXT: s_mov_b32 s33, s32
; GFX10-SCRATCH-NEXT: s_or_saveexec_b32 s1, -1
; GFX10-SCRATCH-NEXT: scratch_store_dword off, v40, s33 ; 4-byte Folded Spill
+; GFX10-SCRATCH-NEXT: scratch_store_dword off, v41, s33 offset:4 ; 4-byte Folded Spill
; GFX10-SCRATCH-NEXT: s_waitcnt_depctr 0xffe3
; GFX10-SCRATCH-NEXT: s_mov_b32 exec_lo, s1
-; GFX10-SCRATCH-NEXT: v_writelane_b32 v40, s0, 4
-; GFX10-SCRATCH-NEXT: s_add_i32 s32, s32, 16
; GFX10-SCRATCH-NEXT: v_writelane_b32 v40, s4, 0
+; GFX10-SCRATCH-NEXT: s_add_i32 s32, s32, 16
+; GFX10-SCRATCH-NEXT: v_writelane_b32 v41, s0, 0
; GFX10-SCRATCH-NEXT: v_writelane_b32 v40, s5, 1
; GFX10-SCRATCH-NEXT: s_load_dwordx2 s[4:5], s[0:1], 0x0
; GFX10-SCRATCH-NEXT: s_getpc_b64 s[0:1]
@@ -9896,9 +10709,11 @@
; GFX10-SCRATCH-NEXT: v_readlane_b32 s30, v40, 2
; GFX10-SCRATCH-NEXT: v_readlane_b32 s5, v40, 1
; GFX10-SCRATCH-NEXT: v_readlane_b32 s4, v40, 0
-; GFX10-SCRATCH-NEXT: v_readlane_b32 s0, v40, 4
+; GFX10-SCRATCH-NEXT: v_readlane_b32 s0, v41, 0
; GFX10-SCRATCH-NEXT: s_or_saveexec_b32 s1, -1
-; GFX10-SCRATCH-NEXT: scratch_load_dword v40, off, s33 ; 4-byte Folded Reload
+; GFX10-SCRATCH-NEXT: s_clause 0x1
+; GFX10-SCRATCH-NEXT: scratch_load_dword v40, off, s33
+; GFX10-SCRATCH-NEXT: scratch_load_dword v41, off, s33 offset:4
; GFX10-SCRATCH-NEXT: s_waitcnt_depctr 0xffe3
; GFX10-SCRATCH-NEXT: s_mov_b32 exec_lo, s1
; GFX10-SCRATCH-NEXT: s_add_i32 s32, s32, -16
@@ -9918,14 +10733,15 @@
; GFX9-NEXT: s_mov_b32 s33, s32
; GFX9-NEXT: s_or_saveexec_b64 s[36:37], -1
; GFX9-NEXT: buffer_store_dword v40, off, s[0:3], s33 ; 4-byte Folded Spill
+; GFX9-NEXT: buffer_store_dword v41, off, s[0:3], s33 offset:4 ; 4-byte Folded Spill
; GFX9-NEXT: s_mov_b64 exec, s[36:37]
-; GFX9-NEXT: v_writelane_b32 v40, s34, 4
; GFX9-NEXT: v_writelane_b32 v40, s4, 0
; GFX9-NEXT: v_writelane_b32 v40, s5, 1
; GFX9-NEXT: s_addk_i32 s32, 0x400
; GFX9-NEXT: v_writelane_b32 v40, s30, 2
; GFX9-NEXT: s_mov_b32 s4, 0x20001
; GFX9-NEXT: s_mov_b32 s5, 3
+; GFX9-NEXT: v_writelane_b32 v41, s34, 0
; GFX9-NEXT: v_writelane_b32 v40, s31, 3
; GFX9-NEXT: s_getpc_b64 s[34:35]
; GFX9-NEXT: s_add_u32 s34, s34, external_void_func_v3i16_inreg@rel32@lo+4
@@ -9935,9 +10751,10 @@
; GFX9-NEXT: v_readlane_b32 s30, v40, 2
; GFX9-NEXT: v_readlane_b32 s5, v40, 1
; GFX9-NEXT: v_readlane_b32 s4, v40, 0
-; GFX9-NEXT: v_readlane_b32 s34, v40, 4
+; GFX9-NEXT: v_readlane_b32 s34, v41, 0
; GFX9-NEXT: s_or_saveexec_b64 s[36:37], -1
; GFX9-NEXT: buffer_load_dword v40, off, s[0:3], s33 ; 4-byte Folded Reload
+; GFX9-NEXT: buffer_load_dword v41, off, s[0:3], s33 offset:4 ; 4-byte Folded Reload
; GFX9-NEXT: s_mov_b64 exec, s[36:37]
; GFX9-NEXT: s_addk_i32 s32, 0xfc00
; GFX9-NEXT: s_mov_b32 s33, s34
@@ -9951,15 +10768,16 @@
; GFX10-NEXT: s_mov_b32 s33, s32
; GFX10-NEXT: s_or_saveexec_b32 s35, -1
; GFX10-NEXT: buffer_store_dword v40, off, s[0:3], s33 ; 4-byte Folded Spill
+; GFX10-NEXT: buffer_store_dword v41, off, s[0:3], s33 offset:4 ; 4-byte Folded Spill
; GFX10-NEXT: s_waitcnt_depctr 0xffe3
; GFX10-NEXT: s_mov_b32 exec_lo, s35
-; GFX10-NEXT: v_writelane_b32 v40, s34, 4
+; GFX10-NEXT: v_writelane_b32 v40, s4, 0
+; GFX10-NEXT: s_mov_b32 s4, 0x20001
; GFX10-NEXT: s_addk_i32 s32, 0x200
+; GFX10-NEXT: v_writelane_b32 v41, s34, 0
; GFX10-NEXT: s_getpc_b64 s[34:35]
; GFX10-NEXT: s_add_u32 s34, s34, external_void_func_v3i16_inreg@rel32@lo+4
; GFX10-NEXT: s_addc_u32 s35, s35, external_void_func_v3i16_inreg@rel32@hi+12
-; GFX10-NEXT: v_writelane_b32 v40, s4, 0
-; GFX10-NEXT: s_mov_b32 s4, 0x20001
; GFX10-NEXT: v_writelane_b32 v40, s5, 1
; GFX10-NEXT: s_mov_b32 s5, 3
; GFX10-NEXT: v_writelane_b32 v40, s30, 2
@@ -9969,9 +10787,11 @@
; GFX10-NEXT: v_readlane_b32 s30, v40, 2
; GFX10-NEXT: v_readlane_b32 s5, v40, 1
; GFX10-NEXT: v_readlane_b32 s4, v40, 0
-; GFX10-NEXT: v_readlane_b32 s34, v40, 4
+; GFX10-NEXT: v_readlane_b32 s34, v41, 0
; GFX10-NEXT: s_or_saveexec_b32 s35, -1
-; GFX10-NEXT: buffer_load_dword v40, off, s[0:3], s33 ; 4-byte Folded Reload
+; GFX10-NEXT: s_clause 0x1
+; GFX10-NEXT: buffer_load_dword v40, off, s[0:3], s33
+; GFX10-NEXT: buffer_load_dword v41, off, s[0:3], s33 offset:4
; GFX10-NEXT: s_waitcnt_depctr 0xffe3
; GFX10-NEXT: s_mov_b32 exec_lo, s35
; GFX10-NEXT: s_addk_i32 s32, 0xfe00
@@ -9985,15 +10805,17 @@
; GFX11-NEXT: s_mov_b32 s0, s33
; GFX11-NEXT: s_mov_b32 s33, s32
; GFX11-NEXT: s_or_saveexec_b32 s1, -1
-; GFX11-NEXT: scratch_store_b32 off, v40, s33 ; 4-byte Folded Spill
+; GFX11-NEXT: s_clause 0x1
+; GFX11-NEXT: scratch_store_b32 off, v40, s33
+; GFX11-NEXT: scratch_store_b32 off, v41, s33 offset:4
; GFX11-NEXT: s_mov_b32 exec_lo, s1
-; GFX11-NEXT: v_writelane_b32 v40, s0, 4
+; GFX11-NEXT: v_writelane_b32 v40, s4, 0
+; GFX11-NEXT: s_mov_b32 s4, 0x20001
; GFX11-NEXT: s_add_i32 s32, s32, 16
+; GFX11-NEXT: v_writelane_b32 v41, s0, 0
; GFX11-NEXT: s_getpc_b64 s[0:1]
; GFX11-NEXT: s_add_u32 s0, s0, external_void_func_v3i16_inreg@rel32@lo+4
; GFX11-NEXT: s_addc_u32 s1, s1, external_void_func_v3i16_inreg@rel32@hi+12
-; GFX11-NEXT: v_writelane_b32 v40, s4, 0
-; GFX11-NEXT: s_mov_b32 s4, 0x20001
; GFX11-NEXT: v_writelane_b32 v40, s5, 1
; GFX11-NEXT: s_mov_b32 s5, 3
; GFX11-NEXT: v_writelane_b32 v40, s30, 2
@@ -10004,9 +10826,11 @@
; GFX11-NEXT: v_readlane_b32 s30, v40, 2
; GFX11-NEXT: v_readlane_b32 s5, v40, 1
; GFX11-NEXT: v_readlane_b32 s4, v40, 0
-; GFX11-NEXT: v_readlane_b32 s0, v40, 4
+; GFX11-NEXT: v_readlane_b32 s0, v41, 0
; GFX11-NEXT: s_or_saveexec_b32 s1, -1
-; GFX11-NEXT: scratch_load_b32 v40, off, s33 ; 4-byte Folded Reload
+; GFX11-NEXT: s_clause 0x1
+; GFX11-NEXT: scratch_load_b32 v40, off, s33
+; GFX11-NEXT: scratch_load_b32 v41, off, s33 offset:4
; GFX11-NEXT: s_mov_b32 exec_lo, s1
; GFX11-NEXT: s_add_i32 s32, s32, -16
; GFX11-NEXT: s_mov_b32 s33, s0
@@ -10020,15 +10844,16 @@
; GFX10-SCRATCH-NEXT: s_mov_b32 s33, s32
; GFX10-SCRATCH-NEXT: s_or_saveexec_b32 s1, -1
; GFX10-SCRATCH-NEXT: scratch_store_dword off, v40, s33 ; 4-byte Folded Spill
+; GFX10-SCRATCH-NEXT: scratch_store_dword off, v41, s33 offset:4 ; 4-byte Folded Spill
; GFX10-SCRATCH-NEXT: s_waitcnt_depctr 0xffe3
; GFX10-SCRATCH-NEXT: s_mov_b32 exec_lo, s1
-; GFX10-SCRATCH-NEXT: v_writelane_b32 v40, s0, 4
+; GFX10-SCRATCH-NEXT: v_writelane_b32 v40, s4, 0
+; GFX10-SCRATCH-NEXT: s_mov_b32 s4, 0x20001
; GFX10-SCRATCH-NEXT: s_add_i32 s32, s32, 16
+; GFX10-SCRATCH-NEXT: v_writelane_b32 v41, s0, 0
; GFX10-SCRATCH-NEXT: s_getpc_b64 s[0:1]
; GFX10-SCRATCH-NEXT: s_add_u32 s0, s0, external_void_func_v3i16_inreg@rel32@lo+4
; GFX10-SCRATCH-NEXT: s_addc_u32 s1, s1, external_void_func_v3i16_inreg@rel32@hi+12
-; GFX10-SCRATCH-NEXT: v_writelane_b32 v40, s4, 0
-; GFX10-SCRATCH-NEXT: s_mov_b32 s4, 0x20001
; GFX10-SCRATCH-NEXT: v_writelane_b32 v40, s5, 1
; GFX10-SCRATCH-NEXT: s_mov_b32 s5, 3
; GFX10-SCRATCH-NEXT: v_writelane_b32 v40, s30, 2
@@ -10038,9 +10863,11 @@
; GFX10-SCRATCH-NEXT: v_readlane_b32 s30, v40, 2
; GFX10-SCRATCH-NEXT: v_readlane_b32 s5, v40, 1
; GFX10-SCRATCH-NEXT: v_readlane_b32 s4, v40, 0
-; GFX10-SCRATCH-NEXT: v_readlane_b32 s0, v40, 4
+; GFX10-SCRATCH-NEXT: v_readlane_b32 s0, v41, 0
; GFX10-SCRATCH-NEXT: s_or_saveexec_b32 s1, -1
-; GFX10-SCRATCH-NEXT: scratch_load_dword v40, off, s33 ; 4-byte Folded Reload
+; GFX10-SCRATCH-NEXT: s_clause 0x1
+; GFX10-SCRATCH-NEXT: scratch_load_dword v40, off, s33
+; GFX10-SCRATCH-NEXT: scratch_load_dword v41, off, s33 offset:4
; GFX10-SCRATCH-NEXT: s_waitcnt_depctr 0xffe3
; GFX10-SCRATCH-NEXT: s_mov_b32 exec_lo, s1
; GFX10-SCRATCH-NEXT: s_add_i32 s32, s32, -16
@@ -10059,14 +10886,15 @@
; GFX9-NEXT: s_mov_b32 s33, s32
; GFX9-NEXT: s_or_saveexec_b64 s[36:37], -1
; GFX9-NEXT: buffer_store_dword v40, off, s[0:3], s33 ; 4-byte Folded Spill
+; GFX9-NEXT: buffer_store_dword v41, off, s[0:3], s33 offset:4 ; 4-byte Folded Spill
; GFX9-NEXT: s_mov_b64 exec, s[36:37]
-; GFX9-NEXT: v_writelane_b32 v40, s34, 4
; GFX9-NEXT: v_writelane_b32 v40, s4, 0
; GFX9-NEXT: v_writelane_b32 v40, s5, 1
; GFX9-NEXT: s_addk_i32 s32, 0x400
; GFX9-NEXT: v_writelane_b32 v40, s30, 2
; GFX9-NEXT: s_mov_b32 s4, 0x40003c00
; GFX9-NEXT: s_movk_i32 s5, 0x4400
+; GFX9-NEXT: v_writelane_b32 v41, s34, 0
; GFX9-NEXT: v_writelane_b32 v40, s31, 3
; GFX9-NEXT: s_getpc_b64 s[34:35]
; GFX9-NEXT: s_add_u32 s34, s34, external_void_func_v3f16_inreg@rel32@lo+4
@@ -10076,9 +10904,10 @@
; GFX9-NEXT: v_readlane_b32 s30, v40, 2
; GFX9-NEXT: v_readlane_b32 s5, v40, 1
; GFX9-NEXT: v_readlane_b32 s4, v40, 0
-; GFX9-NEXT: v_readlane_b32 s34, v40, 4
+; GFX9-NEXT: v_readlane_b32 s34, v41, 0
; GFX9-NEXT: s_or_saveexec_b64 s[36:37], -1
; GFX9-NEXT: buffer_load_dword v40, off, s[0:3], s33 ; 4-byte Folded Reload
+; GFX9-NEXT: buffer_load_dword v41, off, s[0:3], s33 offset:4 ; 4-byte Folded Reload
; GFX9-NEXT: s_mov_b64 exec, s[36:37]
; GFX9-NEXT: s_addk_i32 s32, 0xfc00
; GFX9-NEXT: s_mov_b32 s33, s34
@@ -10092,15 +10921,16 @@
; GFX10-NEXT: s_mov_b32 s33, s32
; GFX10-NEXT: s_or_saveexec_b32 s35, -1
; GFX10-NEXT: buffer_store_dword v40, off, s[0:3], s33 ; 4-byte Folded Spill
+; GFX10-NEXT: buffer_store_dword v41, off, s[0:3], s33 offset:4 ; 4-byte Folded Spill
; GFX10-NEXT: s_waitcnt_depctr 0xffe3
; GFX10-NEXT: s_mov_b32 exec_lo, s35
-; GFX10-NEXT: v_writelane_b32 v40, s34, 4
+; GFX10-NEXT: v_writelane_b32 v40, s4, 0
+; GFX10-NEXT: s_mov_b32 s4, 0x40003c00
; GFX10-NEXT: s_addk_i32 s32, 0x200
+; GFX10-NEXT: v_writelane_b32 v41, s34, 0
; GFX10-NEXT: s_getpc_b64 s[34:35]
; GFX10-NEXT: s_add_u32 s34, s34, external_void_func_v3f16_inreg@rel32@lo+4
; GFX10-NEXT: s_addc_u32 s35, s35, external_void_func_v3f16_inreg@rel32@hi+12
-; GFX10-NEXT: v_writelane_b32 v40, s4, 0
-; GFX10-NEXT: s_mov_b32 s4, 0x40003c00
; GFX10-NEXT: v_writelane_b32 v40, s5, 1
; GFX10-NEXT: s_movk_i32 s5, 0x4400
; GFX10-NEXT: v_writelane_b32 v40, s30, 2
@@ -10110,9 +10940,11 @@
; GFX10-NEXT: v_readlane_b32 s30, v40, 2
; GFX10-NEXT: v_readlane_b32 s5, v40, 1
; GFX10-NEXT: v_readlane_b32 s4, v40, 0
-; GFX10-NEXT: v_readlane_b32 s34, v40, 4
+; GFX10-NEXT: v_readlane_b32 s34, v41, 0
; GFX10-NEXT: s_or_saveexec_b32 s35, -1
-; GFX10-NEXT: buffer_load_dword v40, off, s[0:3], s33 ; 4-byte Folded Reload
+; GFX10-NEXT: s_clause 0x1
+; GFX10-NEXT: buffer_load_dword v40, off, s[0:3], s33
+; GFX10-NEXT: buffer_load_dword v41, off, s[0:3], s33 offset:4
; GFX10-NEXT: s_waitcnt_depctr 0xffe3
; GFX10-NEXT: s_mov_b32 exec_lo, s35
; GFX10-NEXT: s_addk_i32 s32, 0xfe00
@@ -10126,15 +10958,17 @@
; GFX11-NEXT: s_mov_b32 s0, s33
; GFX11-NEXT: s_mov_b32 s33, s32
; GFX11-NEXT: s_or_saveexec_b32 s1, -1
-; GFX11-NEXT: scratch_store_b32 off, v40, s33 ; 4-byte Folded Spill
+; GFX11-NEXT: s_clause 0x1
+; GFX11-NEXT: scratch_store_b32 off, v40, s33
+; GFX11-NEXT: scratch_store_b32 off, v41, s33 offset:4
; GFX11-NEXT: s_mov_b32 exec_lo, s1
-; GFX11-NEXT: v_writelane_b32 v40, s0, 4
+; GFX11-NEXT: v_writelane_b32 v40, s4, 0
+; GFX11-NEXT: s_mov_b32 s4, 0x40003c00
; GFX11-NEXT: s_add_i32 s32, s32, 16
+; GFX11-NEXT: v_writelane_b32 v41, s0, 0
; GFX11-NEXT: s_getpc_b64 s[0:1]
; GFX11-NEXT: s_add_u32 s0, s0, external_void_func_v3f16_inreg@rel32@lo+4
; GFX11-NEXT: s_addc_u32 s1, s1, external_void_func_v3f16_inreg@rel32@hi+12
-; GFX11-NEXT: v_writelane_b32 v40, s4, 0
-; GFX11-NEXT: s_mov_b32 s4, 0x40003c00
; GFX11-NEXT: v_writelane_b32 v40, s5, 1
; GFX11-NEXT: s_movk_i32 s5, 0x4400
; GFX11-NEXT: v_writelane_b32 v40, s30, 2
@@ -10145,9 +10979,11 @@
; GFX11-NEXT: v_readlane_b32 s30, v40, 2
; GFX11-NEXT: v_readlane_b32 s5, v40, 1
; GFX11-NEXT: v_readlane_b32 s4, v40, 0
-; GFX11-NEXT: v_readlane_b32 s0, v40, 4
+; GFX11-NEXT: v_readlane_b32 s0, v41, 0
; GFX11-NEXT: s_or_saveexec_b32 s1, -1
-; GFX11-NEXT: scratch_load_b32 v40, off, s33 ; 4-byte Folded Reload
+; GFX11-NEXT: s_clause 0x1
+; GFX11-NEXT: scratch_load_b32 v40, off, s33
+; GFX11-NEXT: scratch_load_b32 v41, off, s33 offset:4
; GFX11-NEXT: s_mov_b32 exec_lo, s1
; GFX11-NEXT: s_add_i32 s32, s32, -16
; GFX11-NEXT: s_mov_b32 s33, s0
@@ -10161,15 +10997,16 @@
; GFX10-SCRATCH-NEXT: s_mov_b32 s33, s32
; GFX10-SCRATCH-NEXT: s_or_saveexec_b32 s1, -1
; GFX10-SCRATCH-NEXT: scratch_store_dword off, v40, s33 ; 4-byte Folded Spill
+; GFX10-SCRATCH-NEXT: scratch_store_dword off, v41, s33 offset:4 ; 4-byte Folded Spill
; GFX10-SCRATCH-NEXT: s_waitcnt_depctr 0xffe3
; GFX10-SCRATCH-NEXT: s_mov_b32 exec_lo, s1
-; GFX10-SCRATCH-NEXT: v_writelane_b32 v40, s0, 4
+; GFX10-SCRATCH-NEXT: v_writelane_b32 v40, s4, 0
+; GFX10-SCRATCH-NEXT: s_mov_b32 s4, 0x40003c00
; GFX10-SCRATCH-NEXT: s_add_i32 s32, s32, 16
+; GFX10-SCRATCH-NEXT: v_writelane_b32 v41, s0, 0
; GFX10-SCRATCH-NEXT: s_getpc_b64 s[0:1]
; GFX10-SCRATCH-NEXT: s_add_u32 s0, s0, external_void_func_v3f16_inreg@rel32@lo+4
; GFX10-SCRATCH-NEXT: s_addc_u32 s1, s1, external_void_func_v3f16_inreg@rel32@hi+12
-; GFX10-SCRATCH-NEXT: v_writelane_b32 v40, s4, 0
-; GFX10-SCRATCH-NEXT: s_mov_b32 s4, 0x40003c00
; GFX10-SCRATCH-NEXT: v_writelane_b32 v40, s5, 1
; GFX10-SCRATCH-NEXT: s_movk_i32 s5, 0x4400
; GFX10-SCRATCH-NEXT: v_writelane_b32 v40, s30, 2
@@ -10179,9 +11016,11 @@
; GFX10-SCRATCH-NEXT: v_readlane_b32 s30, v40, 2
; GFX10-SCRATCH-NEXT: v_readlane_b32 s5, v40, 1
; GFX10-SCRATCH-NEXT: v_readlane_b32 s4, v40, 0
-; GFX10-SCRATCH-NEXT: v_readlane_b32 s0, v40, 4
+; GFX10-SCRATCH-NEXT: v_readlane_b32 s0, v41, 0
; GFX10-SCRATCH-NEXT: s_or_saveexec_b32 s1, -1
-; GFX10-SCRATCH-NEXT: scratch_load_dword v40, off, s33 ; 4-byte Folded Reload
+; GFX10-SCRATCH-NEXT: s_clause 0x1
+; GFX10-SCRATCH-NEXT: scratch_load_dword v40, off, s33
+; GFX10-SCRATCH-NEXT: scratch_load_dword v41, off, s33 offset:4
; GFX10-SCRATCH-NEXT: s_waitcnt_depctr 0xffe3
; GFX10-SCRATCH-NEXT: s_mov_b32 exec_lo, s1
; GFX10-SCRATCH-NEXT: s_add_i32 s32, s32, -16
@@ -10200,13 +11039,14 @@
; GFX9-NEXT: s_mov_b32 s33, s32
; GFX9-NEXT: s_or_saveexec_b64 s[36:37], -1
; GFX9-NEXT: buffer_store_dword v40, off, s[0:3], s33 ; 4-byte Folded Spill
+; GFX9-NEXT: buffer_store_dword v41, off, s[0:3], s33 offset:4 ; 4-byte Folded Spill
; GFX9-NEXT: s_mov_b64 exec, s[36:37]
-; GFX9-NEXT: v_writelane_b32 v40, s34, 4
; GFX9-NEXT: v_writelane_b32 v40, s4, 0
; GFX9-NEXT: v_writelane_b32 v40, s5, 1
; GFX9-NEXT: s_load_dwordx2 s[4:5], s[34:35], 0x0
; GFX9-NEXT: s_addk_i32 s32, 0x400
; GFX9-NEXT: v_writelane_b32 v40, s30, 2
+; GFX9-NEXT: v_writelane_b32 v41, s34, 0
; GFX9-NEXT: v_writelane_b32 v40, s31, 3
; GFX9-NEXT: s_getpc_b64 s[34:35]
; GFX9-NEXT: s_add_u32 s34, s34, external_void_func_v4i16_inreg@rel32@lo+4
@@ -10216,9 +11056,10 @@
; GFX9-NEXT: v_readlane_b32 s30, v40, 2
; GFX9-NEXT: v_readlane_b32 s5, v40, 1
; GFX9-NEXT: v_readlane_b32 s4, v40, 0
-; GFX9-NEXT: v_readlane_b32 s34, v40, 4
+; GFX9-NEXT: v_readlane_b32 s34, v41, 0
; GFX9-NEXT: s_or_saveexec_b64 s[36:37], -1
; GFX9-NEXT: buffer_load_dword v40, off, s[0:3], s33 ; 4-byte Folded Reload
+; GFX9-NEXT: buffer_load_dword v41, off, s[0:3], s33 offset:4 ; 4-byte Folded Reload
; GFX9-NEXT: s_mov_b64 exec, s[36:37]
; GFX9-NEXT: s_addk_i32 s32, 0xfc00
; GFX9-NEXT: s_mov_b32 s33, s34
@@ -10232,11 +11073,12 @@
; GFX10-NEXT: s_mov_b32 s33, s32
; GFX10-NEXT: s_or_saveexec_b32 s35, -1
; GFX10-NEXT: buffer_store_dword v40, off, s[0:3], s33 ; 4-byte Folded Spill
+; GFX10-NEXT: buffer_store_dword v41, off, s[0:3], s33 offset:4 ; 4-byte Folded Spill
; GFX10-NEXT: s_waitcnt_depctr 0xffe3
; GFX10-NEXT: s_mov_b32 exec_lo, s35
-; GFX10-NEXT: v_writelane_b32 v40, s34, 4
-; GFX10-NEXT: s_addk_i32 s32, 0x200
; GFX10-NEXT: v_writelane_b32 v40, s4, 0
+; GFX10-NEXT: s_addk_i32 s32, 0x200
+; GFX10-NEXT: v_writelane_b32 v41, s34, 0
; GFX10-NEXT: v_writelane_b32 v40, s5, 1
; GFX10-NEXT: s_load_dwordx2 s[4:5], s[34:35], 0x0
; GFX10-NEXT: s_getpc_b64 s[34:35]
@@ -10249,9 +11091,11 @@
; GFX10-NEXT: v_readlane_b32 s30, v40, 2
; GFX10-NEXT: v_readlane_b32 s5, v40, 1
; GFX10-NEXT: v_readlane_b32 s4, v40, 0
-; GFX10-NEXT: v_readlane_b32 s34, v40, 4
+; GFX10-NEXT: v_readlane_b32 s34, v41, 0
; GFX10-NEXT: s_or_saveexec_b32 s35, -1
-; GFX10-NEXT: buffer_load_dword v40, off, s[0:3], s33 ; 4-byte Folded Reload
+; GFX10-NEXT: s_clause 0x1
+; GFX10-NEXT: buffer_load_dword v40, off, s[0:3], s33
+; GFX10-NEXT: buffer_load_dword v41, off, s[0:3], s33 offset:4
; GFX10-NEXT: s_waitcnt_depctr 0xffe3
; GFX10-NEXT: s_mov_b32 exec_lo, s35
; GFX10-NEXT: s_addk_i32 s32, 0xfe00
@@ -10265,11 +11109,13 @@
; GFX11-NEXT: s_mov_b32 s0, s33
; GFX11-NEXT: s_mov_b32 s33, s32
; GFX11-NEXT: s_or_saveexec_b32 s1, -1
-; GFX11-NEXT: scratch_store_b32 off, v40, s33 ; 4-byte Folded Spill
+; GFX11-NEXT: s_clause 0x1
+; GFX11-NEXT: scratch_store_b32 off, v40, s33
+; GFX11-NEXT: scratch_store_b32 off, v41, s33 offset:4
; GFX11-NEXT: s_mov_b32 exec_lo, s1
-; GFX11-NEXT: v_writelane_b32 v40, s0, 4
-; GFX11-NEXT: s_add_i32 s32, s32, 16
; GFX11-NEXT: v_writelane_b32 v40, s4, 0
+; GFX11-NEXT: s_add_i32 s32, s32, 16
+; GFX11-NEXT: v_writelane_b32 v41, s0, 0
; GFX11-NEXT: v_writelane_b32 v40, s5, 1
; GFX11-NEXT: s_load_b64 s[4:5], s[0:1], 0x0
; GFX11-NEXT: s_getpc_b64 s[0:1]
@@ -10283,9 +11129,11 @@
; GFX11-NEXT: v_readlane_b32 s30, v40, 2
; GFX11-NEXT: v_readlane_b32 s5, v40, 1
; GFX11-NEXT: v_readlane_b32 s4, v40, 0
-; GFX11-NEXT: v_readlane_b32 s0, v40, 4
+; GFX11-NEXT: v_readlane_b32 s0, v41, 0
; GFX11-NEXT: s_or_saveexec_b32 s1, -1
-; GFX11-NEXT: scratch_load_b32 v40, off, s33 ; 4-byte Folded Reload
+; GFX11-NEXT: s_clause 0x1
+; GFX11-NEXT: scratch_load_b32 v40, off, s33
+; GFX11-NEXT: scratch_load_b32 v41, off, s33 offset:4
; GFX11-NEXT: s_mov_b32 exec_lo, s1
; GFX11-NEXT: s_add_i32 s32, s32, -16
; GFX11-NEXT: s_mov_b32 s33, s0
@@ -10299,11 +11147,12 @@
; GFX10-SCRATCH-NEXT: s_mov_b32 s33, s32
; GFX10-SCRATCH-NEXT: s_or_saveexec_b32 s1, -1
; GFX10-SCRATCH-NEXT: scratch_store_dword off, v40, s33 ; 4-byte Folded Spill
+; GFX10-SCRATCH-NEXT: scratch_store_dword off, v41, s33 offset:4 ; 4-byte Folded Spill
; GFX10-SCRATCH-NEXT: s_waitcnt_depctr 0xffe3
; GFX10-SCRATCH-NEXT: s_mov_b32 exec_lo, s1
-; GFX10-SCRATCH-NEXT: v_writelane_b32 v40, s0, 4
-; GFX10-SCRATCH-NEXT: s_add_i32 s32, s32, 16
; GFX10-SCRATCH-NEXT: v_writelane_b32 v40, s4, 0
+; GFX10-SCRATCH-NEXT: s_add_i32 s32, s32, 16
+; GFX10-SCRATCH-NEXT: v_writelane_b32 v41, s0, 0
; GFX10-SCRATCH-NEXT: v_writelane_b32 v40, s5, 1
; GFX10-SCRATCH-NEXT: s_load_dwordx2 s[4:5], s[0:1], 0x0
; GFX10-SCRATCH-NEXT: s_getpc_b64 s[0:1]
@@ -10316,9 +11165,11 @@
; GFX10-SCRATCH-NEXT: v_readlane_b32 s30, v40, 2
; GFX10-SCRATCH-NEXT: v_readlane_b32 s5, v40, 1
; GFX10-SCRATCH-NEXT: v_readlane_b32 s4, v40, 0
-; GFX10-SCRATCH-NEXT: v_readlane_b32 s0, v40, 4
+; GFX10-SCRATCH-NEXT: v_readlane_b32 s0, v41, 0
; GFX10-SCRATCH-NEXT: s_or_saveexec_b32 s1, -1
-; GFX10-SCRATCH-NEXT: scratch_load_dword v40, off, s33 ; 4-byte Folded Reload
+; GFX10-SCRATCH-NEXT: s_clause 0x1
+; GFX10-SCRATCH-NEXT: scratch_load_dword v40, off, s33
+; GFX10-SCRATCH-NEXT: scratch_load_dword v41, off, s33 offset:4
; GFX10-SCRATCH-NEXT: s_waitcnt_depctr 0xffe3
; GFX10-SCRATCH-NEXT: s_mov_b32 exec_lo, s1
; GFX10-SCRATCH-NEXT: s_add_i32 s32, s32, -16
@@ -10338,14 +11189,15 @@
; GFX9-NEXT: s_mov_b32 s33, s32
; GFX9-NEXT: s_or_saveexec_b64 s[36:37], -1
; GFX9-NEXT: buffer_store_dword v40, off, s[0:3], s33 ; 4-byte Folded Spill
+; GFX9-NEXT: buffer_store_dword v41, off, s[0:3], s33 offset:4 ; 4-byte Folded Spill
; GFX9-NEXT: s_mov_b64 exec, s[36:37]
-; GFX9-NEXT: v_writelane_b32 v40, s34, 4
; GFX9-NEXT: v_writelane_b32 v40, s4, 0
; GFX9-NEXT: v_writelane_b32 v40, s5, 1
; GFX9-NEXT: s_addk_i32 s32, 0x400
; GFX9-NEXT: v_writelane_b32 v40, s30, 2
; GFX9-NEXT: s_mov_b32 s4, 0x20001
; GFX9-NEXT: s_mov_b32 s5, 0x40003
+; GFX9-NEXT: v_writelane_b32 v41, s34, 0
; GFX9-NEXT: v_writelane_b32 v40, s31, 3
; GFX9-NEXT: s_getpc_b64 s[34:35]
; GFX9-NEXT: s_add_u32 s34, s34, external_void_func_v4i16_inreg@rel32@lo+4
@@ -10355,9 +11207,10 @@
; GFX9-NEXT: v_readlane_b32 s30, v40, 2
; GFX9-NEXT: v_readlane_b32 s5, v40, 1
; GFX9-NEXT: v_readlane_b32 s4, v40, 0
-; GFX9-NEXT: v_readlane_b32 s34, v40, 4
+; GFX9-NEXT: v_readlane_b32 s34, v41, 0
; GFX9-NEXT: s_or_saveexec_b64 s[36:37], -1
; GFX9-NEXT: buffer_load_dword v40, off, s[0:3], s33 ; 4-byte Folded Reload
+; GFX9-NEXT: buffer_load_dword v41, off, s[0:3], s33 offset:4 ; 4-byte Folded Reload
; GFX9-NEXT: s_mov_b64 exec, s[36:37]
; GFX9-NEXT: s_addk_i32 s32, 0xfc00
; GFX9-NEXT: s_mov_b32 s33, s34
@@ -10371,15 +11224,16 @@
; GFX10-NEXT: s_mov_b32 s33, s32
; GFX10-NEXT: s_or_saveexec_b32 s35, -1
; GFX10-NEXT: buffer_store_dword v40, off, s[0:3], s33 ; 4-byte Folded Spill
+; GFX10-NEXT: buffer_store_dword v41, off, s[0:3], s33 offset:4 ; 4-byte Folded Spill
; GFX10-NEXT: s_waitcnt_depctr 0xffe3
; GFX10-NEXT: s_mov_b32 exec_lo, s35
-; GFX10-NEXT: v_writelane_b32 v40, s34, 4
+; GFX10-NEXT: v_writelane_b32 v40, s4, 0
+; GFX10-NEXT: s_mov_b32 s4, 0x20001
; GFX10-NEXT: s_addk_i32 s32, 0x200
+; GFX10-NEXT: v_writelane_b32 v41, s34, 0
; GFX10-NEXT: s_getpc_b64 s[34:35]
; GFX10-NEXT: s_add_u32 s34, s34, external_void_func_v4i16_inreg@rel32@lo+4
; GFX10-NEXT: s_addc_u32 s35, s35, external_void_func_v4i16_inreg@rel32@hi+12
-; GFX10-NEXT: v_writelane_b32 v40, s4, 0
-; GFX10-NEXT: s_mov_b32 s4, 0x20001
; GFX10-NEXT: v_writelane_b32 v40, s5, 1
; GFX10-NEXT: s_mov_b32 s5, 0x40003
; GFX10-NEXT: v_writelane_b32 v40, s30, 2
@@ -10389,9 +11243,11 @@
; GFX10-NEXT: v_readlane_b32 s30, v40, 2
; GFX10-NEXT: v_readlane_b32 s5, v40, 1
; GFX10-NEXT: v_readlane_b32 s4, v40, 0
-; GFX10-NEXT: v_readlane_b32 s34, v40, 4
+; GFX10-NEXT: v_readlane_b32 s34, v41, 0
; GFX10-NEXT: s_or_saveexec_b32 s35, -1
-; GFX10-NEXT: buffer_load_dword v40, off, s[0:3], s33 ; 4-byte Folded Reload
+; GFX10-NEXT: s_clause 0x1
+; GFX10-NEXT: buffer_load_dword v40, off, s[0:3], s33
+; GFX10-NEXT: buffer_load_dword v41, off, s[0:3], s33 offset:4
; GFX10-NEXT: s_waitcnt_depctr 0xffe3
; GFX10-NEXT: s_mov_b32 exec_lo, s35
; GFX10-NEXT: s_addk_i32 s32, 0xfe00
@@ -10405,15 +11261,17 @@
; GFX11-NEXT: s_mov_b32 s0, s33
; GFX11-NEXT: s_mov_b32 s33, s32
; GFX11-NEXT: s_or_saveexec_b32 s1, -1
-; GFX11-NEXT: scratch_store_b32 off, v40, s33 ; 4-byte Folded Spill
+; GFX11-NEXT: s_clause 0x1
+; GFX11-NEXT: scratch_store_b32 off, v40, s33
+; GFX11-NEXT: scratch_store_b32 off, v41, s33 offset:4
; GFX11-NEXT: s_mov_b32 exec_lo, s1
-; GFX11-NEXT: v_writelane_b32 v40, s0, 4
+; GFX11-NEXT: v_writelane_b32 v40, s4, 0
+; GFX11-NEXT: s_mov_b32 s4, 0x20001
; GFX11-NEXT: s_add_i32 s32, s32, 16
+; GFX11-NEXT: v_writelane_b32 v41, s0, 0
; GFX11-NEXT: s_getpc_b64 s[0:1]
; GFX11-NEXT: s_add_u32 s0, s0, external_void_func_v4i16_inreg@rel32@lo+4
; GFX11-NEXT: s_addc_u32 s1, s1, external_void_func_v4i16_inreg@rel32@hi+12
-; GFX11-NEXT: v_writelane_b32 v40, s4, 0
-; GFX11-NEXT: s_mov_b32 s4, 0x20001
; GFX11-NEXT: v_writelane_b32 v40, s5, 1
; GFX11-NEXT: s_mov_b32 s5, 0x40003
; GFX11-NEXT: v_writelane_b32 v40, s30, 2
@@ -10424,9 +11282,11 @@
; GFX11-NEXT: v_readlane_b32 s30, v40, 2
; GFX11-NEXT: v_readlane_b32 s5, v40, 1
; GFX11-NEXT: v_readlane_b32 s4, v40, 0
-; GFX11-NEXT: v_readlane_b32 s0, v40, 4
+; GFX11-NEXT: v_readlane_b32 s0, v41, 0
; GFX11-NEXT: s_or_saveexec_b32 s1, -1
-; GFX11-NEXT: scratch_load_b32 v40, off, s33 ; 4-byte Folded Reload
+; GFX11-NEXT: s_clause 0x1
+; GFX11-NEXT: scratch_load_b32 v40, off, s33
+; GFX11-NEXT: scratch_load_b32 v41, off, s33 offset:4
; GFX11-NEXT: s_mov_b32 exec_lo, s1
; GFX11-NEXT: s_add_i32 s32, s32, -16
; GFX11-NEXT: s_mov_b32 s33, s0
@@ -10440,15 +11300,16 @@
; GFX10-SCRATCH-NEXT: s_mov_b32 s33, s32
; GFX10-SCRATCH-NEXT: s_or_saveexec_b32 s1, -1
; GFX10-SCRATCH-NEXT: scratch_store_dword off, v40, s33 ; 4-byte Folded Spill
+; GFX10-SCRATCH-NEXT: scratch_store_dword off, v41, s33 offset:4 ; 4-byte Folded Spill
; GFX10-SCRATCH-NEXT: s_waitcnt_depctr 0xffe3
; GFX10-SCRATCH-NEXT: s_mov_b32 exec_lo, s1
-; GFX10-SCRATCH-NEXT: v_writelane_b32 v40, s0, 4
+; GFX10-SCRATCH-NEXT: v_writelane_b32 v40, s4, 0
+; GFX10-SCRATCH-NEXT: s_mov_b32 s4, 0x20001
; GFX10-SCRATCH-NEXT: s_add_i32 s32, s32, 16
+; GFX10-SCRATCH-NEXT: v_writelane_b32 v41, s0, 0
; GFX10-SCRATCH-NEXT: s_getpc_b64 s[0:1]
; GFX10-SCRATCH-NEXT: s_add_u32 s0, s0, external_void_func_v4i16_inreg@rel32@lo+4
; GFX10-SCRATCH-NEXT: s_addc_u32 s1, s1, external_void_func_v4i16_inreg@rel32@hi+12
-; GFX10-SCRATCH-NEXT: v_writelane_b32 v40, s4, 0
-; GFX10-SCRATCH-NEXT: s_mov_b32 s4, 0x20001
; GFX10-SCRATCH-NEXT: v_writelane_b32 v40, s5, 1
; GFX10-SCRATCH-NEXT: s_mov_b32 s5, 0x40003
; GFX10-SCRATCH-NEXT: v_writelane_b32 v40, s30, 2
@@ -10458,9 +11319,11 @@
; GFX10-SCRATCH-NEXT: v_readlane_b32 s30, v40, 2
; GFX10-SCRATCH-NEXT: v_readlane_b32 s5, v40, 1
; GFX10-SCRATCH-NEXT: v_readlane_b32 s4, v40, 0
-; GFX10-SCRATCH-NEXT: v_readlane_b32 s0, v40, 4
+; GFX10-SCRATCH-NEXT: v_readlane_b32 s0, v41, 0
; GFX10-SCRATCH-NEXT: s_or_saveexec_b32 s1, -1
-; GFX10-SCRATCH-NEXT: scratch_load_dword v40, off, s33 ; 4-byte Folded Reload
+; GFX10-SCRATCH-NEXT: s_clause 0x1
+; GFX10-SCRATCH-NEXT: scratch_load_dword v40, off, s33
+; GFX10-SCRATCH-NEXT: scratch_load_dword v41, off, s33 offset:4
; GFX10-SCRATCH-NEXT: s_waitcnt_depctr 0xffe3
; GFX10-SCRATCH-NEXT: s_mov_b32 exec_lo, s1
; GFX10-SCRATCH-NEXT: s_add_i32 s32, s32, -16
@@ -10479,12 +11342,13 @@
; GFX9-NEXT: s_mov_b32 s33, s32
; GFX9-NEXT: s_or_saveexec_b64 s[36:37], -1
; GFX9-NEXT: buffer_store_dword v40, off, s[0:3], s33 ; 4-byte Folded Spill
+; GFX9-NEXT: buffer_store_dword v41, off, s[0:3], s33 offset:4 ; 4-byte Folded Spill
; GFX9-NEXT: s_mov_b64 exec, s[36:37]
-; GFX9-NEXT: v_writelane_b32 v40, s34, 3
; GFX9-NEXT: v_writelane_b32 v40, s4, 0
; GFX9-NEXT: s_load_dword s4, s[34:35], 0x0
; GFX9-NEXT: s_addk_i32 s32, 0x400
; GFX9-NEXT: v_writelane_b32 v40, s30, 1
+; GFX9-NEXT: v_writelane_b32 v41, s34, 0
; GFX9-NEXT: v_writelane_b32 v40, s31, 2
; GFX9-NEXT: s_getpc_b64 s[34:35]
; GFX9-NEXT: s_add_u32 s34, s34, external_void_func_v2f16_inreg@rel32@lo+4
@@ -10493,9 +11357,10 @@
; GFX9-NEXT: v_readlane_b32 s31, v40, 2
; GFX9-NEXT: v_readlane_b32 s30, v40, 1
; GFX9-NEXT: v_readlane_b32 s4, v40, 0
-; GFX9-NEXT: v_readlane_b32 s34, v40, 3
+; GFX9-NEXT: v_readlane_b32 s34, v41, 0
; GFX9-NEXT: s_or_saveexec_b64 s[36:37], -1
; GFX9-NEXT: buffer_load_dword v40, off, s[0:3], s33 ; 4-byte Folded Reload
+; GFX9-NEXT: buffer_load_dword v41, off, s[0:3], s33 offset:4 ; 4-byte Folded Reload
; GFX9-NEXT: s_mov_b64 exec, s[36:37]
; GFX9-NEXT: s_addk_i32 s32, 0xfc00
; GFX9-NEXT: s_mov_b32 s33, s34
@@ -10509,12 +11374,13 @@
; GFX10-NEXT: s_mov_b32 s33, s32
; GFX10-NEXT: s_or_saveexec_b32 s35, -1
; GFX10-NEXT: buffer_store_dword v40, off, s[0:3], s33 ; 4-byte Folded Spill
+; GFX10-NEXT: buffer_store_dword v41, off, s[0:3], s33 offset:4 ; 4-byte Folded Spill
; GFX10-NEXT: s_waitcnt_depctr 0xffe3
; GFX10-NEXT: s_mov_b32 exec_lo, s35
-; GFX10-NEXT: v_writelane_b32 v40, s34, 3
-; GFX10-NEXT: s_addk_i32 s32, 0x200
; GFX10-NEXT: v_writelane_b32 v40, s4, 0
; GFX10-NEXT: s_load_dword s4, s[34:35], 0x0
+; GFX10-NEXT: s_addk_i32 s32, 0x200
+; GFX10-NEXT: v_writelane_b32 v41, s34, 0
; GFX10-NEXT: s_getpc_b64 s[34:35]
; GFX10-NEXT: s_add_u32 s34, s34, external_void_func_v2f16_inreg@rel32@lo+4
; GFX10-NEXT: s_addc_u32 s35, s35, external_void_func_v2f16_inreg@rel32@hi+12
@@ -10524,9 +11390,11 @@
; GFX10-NEXT: v_readlane_b32 s31, v40, 2
; GFX10-NEXT: v_readlane_b32 s30, v40, 1
; GFX10-NEXT: v_readlane_b32 s4, v40, 0
-; GFX10-NEXT: v_readlane_b32 s34, v40, 3
+; GFX10-NEXT: v_readlane_b32 s34, v41, 0
; GFX10-NEXT: s_or_saveexec_b32 s35, -1
-; GFX10-NEXT: buffer_load_dword v40, off, s[0:3], s33 ; 4-byte Folded Reload
+; GFX10-NEXT: s_clause 0x1
+; GFX10-NEXT: buffer_load_dword v40, off, s[0:3], s33
+; GFX10-NEXT: buffer_load_dword v41, off, s[0:3], s33 offset:4
; GFX10-NEXT: s_waitcnt_depctr 0xffe3
; GFX10-NEXT: s_mov_b32 exec_lo, s35
; GFX10-NEXT: s_addk_i32 s32, 0xfe00
@@ -10540,12 +11408,14 @@
; GFX11-NEXT: s_mov_b32 s0, s33
; GFX11-NEXT: s_mov_b32 s33, s32
; GFX11-NEXT: s_or_saveexec_b32 s1, -1
-; GFX11-NEXT: scratch_store_b32 off, v40, s33 ; 4-byte Folded Spill
+; GFX11-NEXT: s_clause 0x1
+; GFX11-NEXT: scratch_store_b32 off, v40, s33
+; GFX11-NEXT: scratch_store_b32 off, v41, s33 offset:4
; GFX11-NEXT: s_mov_b32 exec_lo, s1
-; GFX11-NEXT: v_writelane_b32 v40, s0, 3
-; GFX11-NEXT: s_add_i32 s32, s32, 16
; GFX11-NEXT: v_writelane_b32 v40, s4, 0
; GFX11-NEXT: s_load_b32 s4, s[0:1], 0x0
+; GFX11-NEXT: s_add_i32 s32, s32, 16
+; GFX11-NEXT: v_writelane_b32 v41, s0, 0
; GFX11-NEXT: s_getpc_b64 s[0:1]
; GFX11-NEXT: s_add_u32 s0, s0, external_void_func_v2f16_inreg@rel32@lo+4
; GFX11-NEXT: s_addc_u32 s1, s1, external_void_func_v2f16_inreg@rel32@hi+12
@@ -10556,9 +11426,11 @@
; GFX11-NEXT: v_readlane_b32 s31, v40, 2
; GFX11-NEXT: v_readlane_b32 s30, v40, 1
; GFX11-NEXT: v_readlane_b32 s4, v40, 0
-; GFX11-NEXT: v_readlane_b32 s0, v40, 3
+; GFX11-NEXT: v_readlane_b32 s0, v41, 0
; GFX11-NEXT: s_or_saveexec_b32 s1, -1
-; GFX11-NEXT: scratch_load_b32 v40, off, s33 ; 4-byte Folded Reload
+; GFX11-NEXT: s_clause 0x1
+; GFX11-NEXT: scratch_load_b32 v40, off, s33
+; GFX11-NEXT: scratch_load_b32 v41, off, s33 offset:4
; GFX11-NEXT: s_mov_b32 exec_lo, s1
; GFX11-NEXT: s_add_i32 s32, s32, -16
; GFX11-NEXT: s_mov_b32 s33, s0
@@ -10572,12 +11444,13 @@
; GFX10-SCRATCH-NEXT: s_mov_b32 s33, s32
; GFX10-SCRATCH-NEXT: s_or_saveexec_b32 s1, -1
; GFX10-SCRATCH-NEXT: scratch_store_dword off, v40, s33 ; 4-byte Folded Spill
+; GFX10-SCRATCH-NEXT: scratch_store_dword off, v41, s33 offset:4 ; 4-byte Folded Spill
; GFX10-SCRATCH-NEXT: s_waitcnt_depctr 0xffe3
; GFX10-SCRATCH-NEXT: s_mov_b32 exec_lo, s1
-; GFX10-SCRATCH-NEXT: v_writelane_b32 v40, s0, 3
-; GFX10-SCRATCH-NEXT: s_add_i32 s32, s32, 16
; GFX10-SCRATCH-NEXT: v_writelane_b32 v40, s4, 0
; GFX10-SCRATCH-NEXT: s_load_dword s4, s[0:1], 0x0
+; GFX10-SCRATCH-NEXT: s_add_i32 s32, s32, 16
+; GFX10-SCRATCH-NEXT: v_writelane_b32 v41, s0, 0
; GFX10-SCRATCH-NEXT: s_getpc_b64 s[0:1]
; GFX10-SCRATCH-NEXT: s_add_u32 s0, s0, external_void_func_v2f16_inreg@rel32@lo+4
; GFX10-SCRATCH-NEXT: s_addc_u32 s1, s1, external_void_func_v2f16_inreg@rel32@hi+12
@@ -10587,9 +11460,11 @@
; GFX10-SCRATCH-NEXT: v_readlane_b32 s31, v40, 2
; GFX10-SCRATCH-NEXT: v_readlane_b32 s30, v40, 1
; GFX10-SCRATCH-NEXT: v_readlane_b32 s4, v40, 0
-; GFX10-SCRATCH-NEXT: v_readlane_b32 s0, v40, 3
+; GFX10-SCRATCH-NEXT: v_readlane_b32 s0, v41, 0
; GFX10-SCRATCH-NEXT: s_or_saveexec_b32 s1, -1
-; GFX10-SCRATCH-NEXT: scratch_load_dword v40, off, s33 ; 4-byte Folded Reload
+; GFX10-SCRATCH-NEXT: s_clause 0x1
+; GFX10-SCRATCH-NEXT: scratch_load_dword v40, off, s33
+; GFX10-SCRATCH-NEXT: scratch_load_dword v41, off, s33 offset:4
; GFX10-SCRATCH-NEXT: s_waitcnt_depctr 0xffe3
; GFX10-SCRATCH-NEXT: s_mov_b32 exec_lo, s1
; GFX10-SCRATCH-NEXT: s_add_i32 s32, s32, -16
@@ -10609,13 +11484,14 @@
; GFX9-NEXT: s_mov_b32 s33, s32
; GFX9-NEXT: s_or_saveexec_b64 s[36:37], -1
; GFX9-NEXT: buffer_store_dword v40, off, s[0:3], s33 ; 4-byte Folded Spill
+; GFX9-NEXT: buffer_store_dword v41, off, s[0:3], s33 offset:4 ; 4-byte Folded Spill
; GFX9-NEXT: s_mov_b64 exec, s[36:37]
-; GFX9-NEXT: v_writelane_b32 v40, s34, 4
; GFX9-NEXT: v_writelane_b32 v40, s4, 0
; GFX9-NEXT: v_writelane_b32 v40, s5, 1
; GFX9-NEXT: s_load_dwordx2 s[4:5], s[34:35], 0x0
; GFX9-NEXT: s_addk_i32 s32, 0x400
; GFX9-NEXT: v_writelane_b32 v40, s30, 2
+; GFX9-NEXT: v_writelane_b32 v41, s34, 0
; GFX9-NEXT: v_writelane_b32 v40, s31, 3
; GFX9-NEXT: s_getpc_b64 s[34:35]
; GFX9-NEXT: s_add_u32 s34, s34, external_void_func_v2i32_inreg@rel32@lo+4
@@ -10625,9 +11501,10 @@
; GFX9-NEXT: v_readlane_b32 s30, v40, 2
; GFX9-NEXT: v_readlane_b32 s5, v40, 1
; GFX9-NEXT: v_readlane_b32 s4, v40, 0
-; GFX9-NEXT: v_readlane_b32 s34, v40, 4
+; GFX9-NEXT: v_readlane_b32 s34, v41, 0
; GFX9-NEXT: s_or_saveexec_b64 s[36:37], -1
; GFX9-NEXT: buffer_load_dword v40, off, s[0:3], s33 ; 4-byte Folded Reload
+; GFX9-NEXT: buffer_load_dword v41, off, s[0:3], s33 offset:4 ; 4-byte Folded Reload
; GFX9-NEXT: s_mov_b64 exec, s[36:37]
; GFX9-NEXT: s_addk_i32 s32, 0xfc00
; GFX9-NEXT: s_mov_b32 s33, s34
@@ -10641,11 +11518,12 @@
; GFX10-NEXT: s_mov_b32 s33, s32
; GFX10-NEXT: s_or_saveexec_b32 s35, -1
; GFX10-NEXT: buffer_store_dword v40, off, s[0:3], s33 ; 4-byte Folded Spill
+; GFX10-NEXT: buffer_store_dword v41, off, s[0:3], s33 offset:4 ; 4-byte Folded Spill
; GFX10-NEXT: s_waitcnt_depctr 0xffe3
; GFX10-NEXT: s_mov_b32 exec_lo, s35
-; GFX10-NEXT: v_writelane_b32 v40, s34, 4
-; GFX10-NEXT: s_addk_i32 s32, 0x200
; GFX10-NEXT: v_writelane_b32 v40, s4, 0
+; GFX10-NEXT: s_addk_i32 s32, 0x200
+; GFX10-NEXT: v_writelane_b32 v41, s34, 0
; GFX10-NEXT: v_writelane_b32 v40, s5, 1
; GFX10-NEXT: s_load_dwordx2 s[4:5], s[34:35], 0x0
; GFX10-NEXT: s_getpc_b64 s[34:35]
@@ -10658,9 +11536,11 @@
; GFX10-NEXT: v_readlane_b32 s30, v40, 2
; GFX10-NEXT: v_readlane_b32 s5, v40, 1
; GFX10-NEXT: v_readlane_b32 s4, v40, 0
-; GFX10-NEXT: v_readlane_b32 s34, v40, 4
+; GFX10-NEXT: v_readlane_b32 s34, v41, 0
; GFX10-NEXT: s_or_saveexec_b32 s35, -1
-; GFX10-NEXT: buffer_load_dword v40, off, s[0:3], s33 ; 4-byte Folded Reload
+; GFX10-NEXT: s_clause 0x1
+; GFX10-NEXT: buffer_load_dword v40, off, s[0:3], s33
+; GFX10-NEXT: buffer_load_dword v41, off, s[0:3], s33 offset:4
; GFX10-NEXT: s_waitcnt_depctr 0xffe3
; GFX10-NEXT: s_mov_b32 exec_lo, s35
; GFX10-NEXT: s_addk_i32 s32, 0xfe00
@@ -10674,11 +11554,13 @@
; GFX11-NEXT: s_mov_b32 s0, s33
; GFX11-NEXT: s_mov_b32 s33, s32
; GFX11-NEXT: s_or_saveexec_b32 s1, -1
-; GFX11-NEXT: scratch_store_b32 off, v40, s33 ; 4-byte Folded Spill
+; GFX11-NEXT: s_clause 0x1
+; GFX11-NEXT: scratch_store_b32 off, v40, s33
+; GFX11-NEXT: scratch_store_b32 off, v41, s33 offset:4
; GFX11-NEXT: s_mov_b32 exec_lo, s1
-; GFX11-NEXT: v_writelane_b32 v40, s0, 4
-; GFX11-NEXT: s_add_i32 s32, s32, 16
; GFX11-NEXT: v_writelane_b32 v40, s4, 0
+; GFX11-NEXT: s_add_i32 s32, s32, 16
+; GFX11-NEXT: v_writelane_b32 v41, s0, 0
; GFX11-NEXT: v_writelane_b32 v40, s5, 1
; GFX11-NEXT: s_load_b64 s[4:5], s[0:1], 0x0
; GFX11-NEXT: s_getpc_b64 s[0:1]
@@ -10692,9 +11574,11 @@
; GFX11-NEXT: v_readlane_b32 s30, v40, 2
; GFX11-NEXT: v_readlane_b32 s5, v40, 1
; GFX11-NEXT: v_readlane_b32 s4, v40, 0
-; GFX11-NEXT: v_readlane_b32 s0, v40, 4
+; GFX11-NEXT: v_readlane_b32 s0, v41, 0
; GFX11-NEXT: s_or_saveexec_b32 s1, -1
-; GFX11-NEXT: scratch_load_b32 v40, off, s33 ; 4-byte Folded Reload
+; GFX11-NEXT: s_clause 0x1
+; GFX11-NEXT: scratch_load_b32 v40, off, s33
+; GFX11-NEXT: scratch_load_b32 v41, off, s33 offset:4
; GFX11-NEXT: s_mov_b32 exec_lo, s1
; GFX11-NEXT: s_add_i32 s32, s32, -16
; GFX11-NEXT: s_mov_b32 s33, s0
@@ -10708,11 +11592,12 @@
; GFX10-SCRATCH-NEXT: s_mov_b32 s33, s32
; GFX10-SCRATCH-NEXT: s_or_saveexec_b32 s1, -1
; GFX10-SCRATCH-NEXT: scratch_store_dword off, v40, s33 ; 4-byte Folded Spill
+; GFX10-SCRATCH-NEXT: scratch_store_dword off, v41, s33 offset:4 ; 4-byte Folded Spill
; GFX10-SCRATCH-NEXT: s_waitcnt_depctr 0xffe3
; GFX10-SCRATCH-NEXT: s_mov_b32 exec_lo, s1
-; GFX10-SCRATCH-NEXT: v_writelane_b32 v40, s0, 4
-; GFX10-SCRATCH-NEXT: s_add_i32 s32, s32, 16
; GFX10-SCRATCH-NEXT: v_writelane_b32 v40, s4, 0
+; GFX10-SCRATCH-NEXT: s_add_i32 s32, s32, 16
+; GFX10-SCRATCH-NEXT: v_writelane_b32 v41, s0, 0
; GFX10-SCRATCH-NEXT: v_writelane_b32 v40, s5, 1
; GFX10-SCRATCH-NEXT: s_load_dwordx2 s[4:5], s[0:1], 0x0
; GFX10-SCRATCH-NEXT: s_getpc_b64 s[0:1]
@@ -10725,9 +11610,11 @@
; GFX10-SCRATCH-NEXT: v_readlane_b32 s30, v40, 2
; GFX10-SCRATCH-NEXT: v_readlane_b32 s5, v40, 1
; GFX10-SCRATCH-NEXT: v_readlane_b32 s4, v40, 0
-; GFX10-SCRATCH-NEXT: v_readlane_b32 s0, v40, 4
+; GFX10-SCRATCH-NEXT: v_readlane_b32 s0, v41, 0
; GFX10-SCRATCH-NEXT: s_or_saveexec_b32 s1, -1
-; GFX10-SCRATCH-NEXT: scratch_load_dword v40, off, s33 ; 4-byte Folded Reload
+; GFX10-SCRATCH-NEXT: s_clause 0x1
+; GFX10-SCRATCH-NEXT: scratch_load_dword v40, off, s33
+; GFX10-SCRATCH-NEXT: scratch_load_dword v41, off, s33 offset:4
; GFX10-SCRATCH-NEXT: s_waitcnt_depctr 0xffe3
; GFX10-SCRATCH-NEXT: s_mov_b32 exec_lo, s1
; GFX10-SCRATCH-NEXT: s_add_i32 s32, s32, -16
@@ -10747,14 +11634,15 @@
; GFX9-NEXT: s_mov_b32 s33, s32
; GFX9-NEXT: s_or_saveexec_b64 s[36:37], -1
; GFX9-NEXT: buffer_store_dword v40, off, s[0:3], s33 ; 4-byte Folded Spill
+; GFX9-NEXT: buffer_store_dword v41, off, s[0:3], s33 offset:4 ; 4-byte Folded Spill
; GFX9-NEXT: s_mov_b64 exec, s[36:37]
-; GFX9-NEXT: v_writelane_b32 v40, s34, 4
; GFX9-NEXT: v_writelane_b32 v40, s4, 0
; GFX9-NEXT: v_writelane_b32 v40, s5, 1
; GFX9-NEXT: s_addk_i32 s32, 0x400
; GFX9-NEXT: v_writelane_b32 v40, s30, 2
; GFX9-NEXT: s_mov_b32 s4, 1
; GFX9-NEXT: s_mov_b32 s5, 2
+; GFX9-NEXT: v_writelane_b32 v41, s34, 0
; GFX9-NEXT: v_writelane_b32 v40, s31, 3
; GFX9-NEXT: s_getpc_b64 s[34:35]
; GFX9-NEXT: s_add_u32 s34, s34, external_void_func_v2i32_inreg@rel32@lo+4
@@ -10764,9 +11652,10 @@
; GFX9-NEXT: v_readlane_b32 s30, v40, 2
; GFX9-NEXT: v_readlane_b32 s5, v40, 1
; GFX9-NEXT: v_readlane_b32 s4, v40, 0
-; GFX9-NEXT: v_readlane_b32 s34, v40, 4
+; GFX9-NEXT: v_readlane_b32 s34, v41, 0
; GFX9-NEXT: s_or_saveexec_b64 s[36:37], -1
; GFX9-NEXT: buffer_load_dword v40, off, s[0:3], s33 ; 4-byte Folded Reload
+; GFX9-NEXT: buffer_load_dword v41, off, s[0:3], s33 offset:4 ; 4-byte Folded Reload
; GFX9-NEXT: s_mov_b64 exec, s[36:37]
; GFX9-NEXT: s_addk_i32 s32, 0xfc00
; GFX9-NEXT: s_mov_b32 s33, s34
@@ -10780,15 +11669,16 @@
; GFX10-NEXT: s_mov_b32 s33, s32
; GFX10-NEXT: s_or_saveexec_b32 s35, -1
; GFX10-NEXT: buffer_store_dword v40, off, s[0:3], s33 ; 4-byte Folded Spill
+; GFX10-NEXT: buffer_store_dword v41, off, s[0:3], s33 offset:4 ; 4-byte Folded Spill
; GFX10-NEXT: s_waitcnt_depctr 0xffe3
; GFX10-NEXT: s_mov_b32 exec_lo, s35
-; GFX10-NEXT: v_writelane_b32 v40, s34, 4
+; GFX10-NEXT: v_writelane_b32 v40, s4, 0
+; GFX10-NEXT: s_mov_b32 s4, 1
; GFX10-NEXT: s_addk_i32 s32, 0x200
+; GFX10-NEXT: v_writelane_b32 v41, s34, 0
; GFX10-NEXT: s_getpc_b64 s[34:35]
; GFX10-NEXT: s_add_u32 s34, s34, external_void_func_v2i32_inreg@rel32@lo+4
; GFX10-NEXT: s_addc_u32 s35, s35, external_void_func_v2i32_inreg@rel32@hi+12
-; GFX10-NEXT: v_writelane_b32 v40, s4, 0
-; GFX10-NEXT: s_mov_b32 s4, 1
; GFX10-NEXT: v_writelane_b32 v40, s5, 1
; GFX10-NEXT: s_mov_b32 s5, 2
; GFX10-NEXT: v_writelane_b32 v40, s30, 2
@@ -10798,9 +11688,11 @@
; GFX10-NEXT: v_readlane_b32 s30, v40, 2
; GFX10-NEXT: v_readlane_b32 s5, v40, 1
; GFX10-NEXT: v_readlane_b32 s4, v40, 0
-; GFX10-NEXT: v_readlane_b32 s34, v40, 4
+; GFX10-NEXT: v_readlane_b32 s34, v41, 0
; GFX10-NEXT: s_or_saveexec_b32 s35, -1
-; GFX10-NEXT: buffer_load_dword v40, off, s[0:3], s33 ; 4-byte Folded Reload
+; GFX10-NEXT: s_clause 0x1
+; GFX10-NEXT: buffer_load_dword v40, off, s[0:3], s33
+; GFX10-NEXT: buffer_load_dword v41, off, s[0:3], s33 offset:4
; GFX10-NEXT: s_waitcnt_depctr 0xffe3
; GFX10-NEXT: s_mov_b32 exec_lo, s35
; GFX10-NEXT: s_addk_i32 s32, 0xfe00
@@ -10814,15 +11706,17 @@
; GFX11-NEXT: s_mov_b32 s0, s33
; GFX11-NEXT: s_mov_b32 s33, s32
; GFX11-NEXT: s_or_saveexec_b32 s1, -1
-; GFX11-NEXT: scratch_store_b32 off, v40, s33 ; 4-byte Folded Spill
+; GFX11-NEXT: s_clause 0x1
+; GFX11-NEXT: scratch_store_b32 off, v40, s33
+; GFX11-NEXT: scratch_store_b32 off, v41, s33 offset:4
; GFX11-NEXT: s_mov_b32 exec_lo, s1
-; GFX11-NEXT: v_writelane_b32 v40, s0, 4
+; GFX11-NEXT: v_writelane_b32 v40, s4, 0
+; GFX11-NEXT: s_mov_b32 s4, 1
; GFX11-NEXT: s_add_i32 s32, s32, 16
+; GFX11-NEXT: v_writelane_b32 v41, s0, 0
; GFX11-NEXT: s_getpc_b64 s[0:1]
; GFX11-NEXT: s_add_u32 s0, s0, external_void_func_v2i32_inreg@rel32@lo+4
; GFX11-NEXT: s_addc_u32 s1, s1, external_void_func_v2i32_inreg@rel32@hi+12
-; GFX11-NEXT: v_writelane_b32 v40, s4, 0
-; GFX11-NEXT: s_mov_b32 s4, 1
; GFX11-NEXT: v_writelane_b32 v40, s5, 1
; GFX11-NEXT: s_mov_b32 s5, 2
; GFX11-NEXT: v_writelane_b32 v40, s30, 2
@@ -10833,9 +11727,11 @@
; GFX11-NEXT: v_readlane_b32 s30, v40, 2
; GFX11-NEXT: v_readlane_b32 s5, v40, 1
; GFX11-NEXT: v_readlane_b32 s4, v40, 0
-; GFX11-NEXT: v_readlane_b32 s0, v40, 4
+; GFX11-NEXT: v_readlane_b32 s0, v41, 0
; GFX11-NEXT: s_or_saveexec_b32 s1, -1
-; GFX11-NEXT: scratch_load_b32 v40, off, s33 ; 4-byte Folded Reload
+; GFX11-NEXT: s_clause 0x1
+; GFX11-NEXT: scratch_load_b32 v40, off, s33
+; GFX11-NEXT: scratch_load_b32 v41, off, s33 offset:4
; GFX11-NEXT: s_mov_b32 exec_lo, s1
; GFX11-NEXT: s_add_i32 s32, s32, -16
; GFX11-NEXT: s_mov_b32 s33, s0
@@ -10849,15 +11745,16 @@
; GFX10-SCRATCH-NEXT: s_mov_b32 s33, s32
; GFX10-SCRATCH-NEXT: s_or_saveexec_b32 s1, -1
; GFX10-SCRATCH-NEXT: scratch_store_dword off, v40, s33 ; 4-byte Folded Spill
+; GFX10-SCRATCH-NEXT: scratch_store_dword off, v41, s33 offset:4 ; 4-byte Folded Spill
; GFX10-SCRATCH-NEXT: s_waitcnt_depctr 0xffe3
; GFX10-SCRATCH-NEXT: s_mov_b32 exec_lo, s1
-; GFX10-SCRATCH-NEXT: v_writelane_b32 v40, s0, 4
+; GFX10-SCRATCH-NEXT: v_writelane_b32 v40, s4, 0
+; GFX10-SCRATCH-NEXT: s_mov_b32 s4, 1
; GFX10-SCRATCH-NEXT: s_add_i32 s32, s32, 16
+; GFX10-SCRATCH-NEXT: v_writelane_b32 v41, s0, 0
; GFX10-SCRATCH-NEXT: s_getpc_b64 s[0:1]
; GFX10-SCRATCH-NEXT: s_add_u32 s0, s0, external_void_func_v2i32_inreg@rel32@lo+4
; GFX10-SCRATCH-NEXT: s_addc_u32 s1, s1, external_void_func_v2i32_inreg@rel32@hi+12
-; GFX10-SCRATCH-NEXT: v_writelane_b32 v40, s4, 0
-; GFX10-SCRATCH-NEXT: s_mov_b32 s4, 1
; GFX10-SCRATCH-NEXT: v_writelane_b32 v40, s5, 1
; GFX10-SCRATCH-NEXT: s_mov_b32 s5, 2
; GFX10-SCRATCH-NEXT: v_writelane_b32 v40, s30, 2
@@ -10867,9 +11764,11 @@
; GFX10-SCRATCH-NEXT: v_readlane_b32 s30, v40, 2
; GFX10-SCRATCH-NEXT: v_readlane_b32 s5, v40, 1
; GFX10-SCRATCH-NEXT: v_readlane_b32 s4, v40, 0
-; GFX10-SCRATCH-NEXT: v_readlane_b32 s0, v40, 4
+; GFX10-SCRATCH-NEXT: v_readlane_b32 s0, v41, 0
; GFX10-SCRATCH-NEXT: s_or_saveexec_b32 s1, -1
-; GFX10-SCRATCH-NEXT: scratch_load_dword v40, off, s33 ; 4-byte Folded Reload
+; GFX10-SCRATCH-NEXT: s_clause 0x1
+; GFX10-SCRATCH-NEXT: scratch_load_dword v40, off, s33
+; GFX10-SCRATCH-NEXT: scratch_load_dword v41, off, s33 offset:4
; GFX10-SCRATCH-NEXT: s_waitcnt_depctr 0xffe3
; GFX10-SCRATCH-NEXT: s_mov_b32 exec_lo, s1
; GFX10-SCRATCH-NEXT: s_add_i32 s32, s32, -16
@@ -10888,8 +11787,8 @@
; GFX9-NEXT: s_mov_b32 s33, s32
; GFX9-NEXT: s_or_saveexec_b64 s[36:37], -1
; GFX9-NEXT: buffer_store_dword v40, off, s[0:3], s33 ; 4-byte Folded Spill
+; GFX9-NEXT: buffer_store_dword v41, off, s[0:3], s33 offset:4 ; 4-byte Folded Spill
; GFX9-NEXT: s_mov_b64 exec, s[36:37]
-; GFX9-NEXT: v_writelane_b32 v40, s34, 5
; GFX9-NEXT: v_writelane_b32 v40, s4, 0
; GFX9-NEXT: v_writelane_b32 v40, s5, 1
; GFX9-NEXT: v_writelane_b32 v40, s6, 2
@@ -10898,6 +11797,7 @@
; GFX9-NEXT: s_mov_b32 s4, 3
; GFX9-NEXT: s_mov_b32 s5, 4
; GFX9-NEXT: s_mov_b32 s6, 5
+; GFX9-NEXT: v_writelane_b32 v41, s34, 0
; GFX9-NEXT: v_writelane_b32 v40, s31, 4
; GFX9-NEXT: s_getpc_b64 s[34:35]
; GFX9-NEXT: s_add_u32 s34, s34, external_void_func_v3i32_inreg@rel32@lo+4
@@ -10908,9 +11808,10 @@
; GFX9-NEXT: v_readlane_b32 s6, v40, 2
; GFX9-NEXT: v_readlane_b32 s5, v40, 1
; GFX9-NEXT: v_readlane_b32 s4, v40, 0
-; GFX9-NEXT: v_readlane_b32 s34, v40, 5
+; GFX9-NEXT: v_readlane_b32 s34, v41, 0
; GFX9-NEXT: s_or_saveexec_b64 s[36:37], -1
; GFX9-NEXT: buffer_load_dword v40, off, s[0:3], s33 ; 4-byte Folded Reload
+; GFX9-NEXT: buffer_load_dword v41, off, s[0:3], s33 offset:4 ; 4-byte Folded Reload
; GFX9-NEXT: s_mov_b64 exec, s[36:37]
; GFX9-NEXT: s_addk_i32 s32, 0xfc00
; GFX9-NEXT: s_mov_b32 s33, s34
@@ -10924,15 +11825,16 @@
; GFX10-NEXT: s_mov_b32 s33, s32
; GFX10-NEXT: s_or_saveexec_b32 s35, -1
; GFX10-NEXT: buffer_store_dword v40, off, s[0:3], s33 ; 4-byte Folded Spill
+; GFX10-NEXT: buffer_store_dword v41, off, s[0:3], s33 offset:4 ; 4-byte Folded Spill
; GFX10-NEXT: s_waitcnt_depctr 0xffe3
; GFX10-NEXT: s_mov_b32 exec_lo, s35
-; GFX10-NEXT: v_writelane_b32 v40, s34, 5
+; GFX10-NEXT: v_writelane_b32 v40, s4, 0
+; GFX10-NEXT: s_mov_b32 s4, 3
; GFX10-NEXT: s_addk_i32 s32, 0x200
+; GFX10-NEXT: v_writelane_b32 v41, s34, 0
; GFX10-NEXT: s_getpc_b64 s[34:35]
; GFX10-NEXT: s_add_u32 s34, s34, external_void_func_v3i32_inreg@rel32@lo+4
; GFX10-NEXT: s_addc_u32 s35, s35, external_void_func_v3i32_inreg@rel32@hi+12
-; GFX10-NEXT: v_writelane_b32 v40, s4, 0
-; GFX10-NEXT: s_mov_b32 s4, 3
; GFX10-NEXT: v_writelane_b32 v40, s5, 1
; GFX10-NEXT: s_mov_b32 s5, 4
; GFX10-NEXT: v_writelane_b32 v40, s6, 2
@@ -10945,9 +11847,11 @@
; GFX10-NEXT: v_readlane_b32 s6, v40, 2
; GFX10-NEXT: v_readlane_b32 s5, v40, 1
; GFX10-NEXT: v_readlane_b32 s4, v40, 0
-; GFX10-NEXT: v_readlane_b32 s34, v40, 5
+; GFX10-NEXT: v_readlane_b32 s34, v41, 0
; GFX10-NEXT: s_or_saveexec_b32 s35, -1
-; GFX10-NEXT: buffer_load_dword v40, off, s[0:3], s33 ; 4-byte Folded Reload
+; GFX10-NEXT: s_clause 0x1
+; GFX10-NEXT: buffer_load_dword v40, off, s[0:3], s33
+; GFX10-NEXT: buffer_load_dword v41, off, s[0:3], s33 offset:4
; GFX10-NEXT: s_waitcnt_depctr 0xffe3
; GFX10-NEXT: s_mov_b32 exec_lo, s35
; GFX10-NEXT: s_addk_i32 s32, 0xfe00
@@ -10961,15 +11865,17 @@
; GFX11-NEXT: s_mov_b32 s0, s33
; GFX11-NEXT: s_mov_b32 s33, s32
; GFX11-NEXT: s_or_saveexec_b32 s1, -1
-; GFX11-NEXT: scratch_store_b32 off, v40, s33 ; 4-byte Folded Spill
+; GFX11-NEXT: s_clause 0x1
+; GFX11-NEXT: scratch_store_b32 off, v40, s33
+; GFX11-NEXT: scratch_store_b32 off, v41, s33 offset:4
; GFX11-NEXT: s_mov_b32 exec_lo, s1
-; GFX11-NEXT: v_writelane_b32 v40, s0, 5
+; GFX11-NEXT: v_writelane_b32 v40, s4, 0
+; GFX11-NEXT: s_mov_b32 s4, 3
; GFX11-NEXT: s_add_i32 s32, s32, 16
+; GFX11-NEXT: v_writelane_b32 v41, s0, 0
; GFX11-NEXT: s_getpc_b64 s[0:1]
; GFX11-NEXT: s_add_u32 s0, s0, external_void_func_v3i32_inreg@rel32@lo+4
; GFX11-NEXT: s_addc_u32 s1, s1, external_void_func_v3i32_inreg@rel32@hi+12
-; GFX11-NEXT: v_writelane_b32 v40, s4, 0
-; GFX11-NEXT: s_mov_b32 s4, 3
; GFX11-NEXT: v_writelane_b32 v40, s5, 1
; GFX11-NEXT: s_mov_b32 s5, 4
; GFX11-NEXT: v_writelane_b32 v40, s6, 2
@@ -10983,9 +11889,11 @@
; GFX11-NEXT: v_readlane_b32 s6, v40, 2
; GFX11-NEXT: v_readlane_b32 s5, v40, 1
; GFX11-NEXT: v_readlane_b32 s4, v40, 0
-; GFX11-NEXT: v_readlane_b32 s0, v40, 5
+; GFX11-NEXT: v_readlane_b32 s0, v41, 0
; GFX11-NEXT: s_or_saveexec_b32 s1, -1
-; GFX11-NEXT: scratch_load_b32 v40, off, s33 ; 4-byte Folded Reload
+; GFX11-NEXT: s_clause 0x1
+; GFX11-NEXT: scratch_load_b32 v40, off, s33
+; GFX11-NEXT: scratch_load_b32 v41, off, s33 offset:4
; GFX11-NEXT: s_mov_b32 exec_lo, s1
; GFX11-NEXT: s_add_i32 s32, s32, -16
; GFX11-NEXT: s_mov_b32 s33, s0
@@ -10999,15 +11907,16 @@
; GFX10-SCRATCH-NEXT: s_mov_b32 s33, s32
; GFX10-SCRATCH-NEXT: s_or_saveexec_b32 s1, -1
; GFX10-SCRATCH-NEXT: scratch_store_dword off, v40, s33 ; 4-byte Folded Spill
+; GFX10-SCRATCH-NEXT: scratch_store_dword off, v41, s33 offset:4 ; 4-byte Folded Spill
; GFX10-SCRATCH-NEXT: s_waitcnt_depctr 0xffe3
; GFX10-SCRATCH-NEXT: s_mov_b32 exec_lo, s1
-; GFX10-SCRATCH-NEXT: v_writelane_b32 v40, s0, 5
+; GFX10-SCRATCH-NEXT: v_writelane_b32 v40, s4, 0
+; GFX10-SCRATCH-NEXT: s_mov_b32 s4, 3
; GFX10-SCRATCH-NEXT: s_add_i32 s32, s32, 16
+; GFX10-SCRATCH-NEXT: v_writelane_b32 v41, s0, 0
; GFX10-SCRATCH-NEXT: s_getpc_b64 s[0:1]
; GFX10-SCRATCH-NEXT: s_add_u32 s0, s0, external_void_func_v3i32_inreg@rel32@lo+4
; GFX10-SCRATCH-NEXT: s_addc_u32 s1, s1, external_void_func_v3i32_inreg@rel32@hi+12
-; GFX10-SCRATCH-NEXT: v_writelane_b32 v40, s4, 0
-; GFX10-SCRATCH-NEXT: s_mov_b32 s4, 3
; GFX10-SCRATCH-NEXT: v_writelane_b32 v40, s5, 1
; GFX10-SCRATCH-NEXT: s_mov_b32 s5, 4
; GFX10-SCRATCH-NEXT: v_writelane_b32 v40, s6, 2
@@ -11020,9 +11929,11 @@
; GFX10-SCRATCH-NEXT: v_readlane_b32 s6, v40, 2
; GFX10-SCRATCH-NEXT: v_readlane_b32 s5, v40, 1
; GFX10-SCRATCH-NEXT: v_readlane_b32 s4, v40, 0
-; GFX10-SCRATCH-NEXT: v_readlane_b32 s0, v40, 5
+; GFX10-SCRATCH-NEXT: v_readlane_b32 s0, v41, 0
; GFX10-SCRATCH-NEXT: s_or_saveexec_b32 s1, -1
-; GFX10-SCRATCH-NEXT: scratch_load_dword v40, off, s33 ; 4-byte Folded Reload
+; GFX10-SCRATCH-NEXT: s_clause 0x1
+; GFX10-SCRATCH-NEXT: scratch_load_dword v40, off, s33
+; GFX10-SCRATCH-NEXT: scratch_load_dword v41, off, s33 offset:4
; GFX10-SCRATCH-NEXT: s_waitcnt_depctr 0xffe3
; GFX10-SCRATCH-NEXT: s_mov_b32 exec_lo, s1
; GFX10-SCRATCH-NEXT: s_add_i32 s32, s32, -16
@@ -11041,8 +11952,8 @@
; GFX9-NEXT: s_mov_b32 s33, s32
; GFX9-NEXT: s_or_saveexec_b64 s[36:37], -1
; GFX9-NEXT: buffer_store_dword v40, off, s[0:3], s33 ; 4-byte Folded Spill
+; GFX9-NEXT: buffer_store_dword v41, off, s[0:3], s33 offset:4 ; 4-byte Folded Spill
; GFX9-NEXT: s_mov_b64 exec, s[36:37]
-; GFX9-NEXT: v_writelane_b32 v40, s34, 6
; GFX9-NEXT: v_writelane_b32 v40, s4, 0
; GFX9-NEXT: v_writelane_b32 v40, s5, 1
; GFX9-NEXT: v_writelane_b32 v40, s6, 2
@@ -11053,6 +11964,7 @@
; GFX9-NEXT: s_mov_b32 s5, 4
; GFX9-NEXT: s_mov_b32 s6, 5
; GFX9-NEXT: s_mov_b32 s7, 6
+; GFX9-NEXT: v_writelane_b32 v41, s34, 0
; GFX9-NEXT: v_writelane_b32 v40, s31, 5
; GFX9-NEXT: s_getpc_b64 s[34:35]
; GFX9-NEXT: s_add_u32 s34, s34, external_void_func_v3i32_i32_inreg@rel32@lo+4
@@ -11064,9 +11976,10 @@
; GFX9-NEXT: v_readlane_b32 s6, v40, 2
; GFX9-NEXT: v_readlane_b32 s5, v40, 1
; GFX9-NEXT: v_readlane_b32 s4, v40, 0
-; GFX9-NEXT: v_readlane_b32 s34, v40, 6
+; GFX9-NEXT: v_readlane_b32 s34, v41, 0
; GFX9-NEXT: s_or_saveexec_b64 s[36:37], -1
; GFX9-NEXT: buffer_load_dword v40, off, s[0:3], s33 ; 4-byte Folded Reload
+; GFX9-NEXT: buffer_load_dword v41, off, s[0:3], s33 offset:4 ; 4-byte Folded Reload
; GFX9-NEXT: s_mov_b64 exec, s[36:37]
; GFX9-NEXT: s_addk_i32 s32, 0xfc00
; GFX9-NEXT: s_mov_b32 s33, s34
@@ -11080,15 +11993,16 @@
; GFX10-NEXT: s_mov_b32 s33, s32
; GFX10-NEXT: s_or_saveexec_b32 s35, -1
; GFX10-NEXT: buffer_store_dword v40, off, s[0:3], s33 ; 4-byte Folded Spill
+; GFX10-NEXT: buffer_store_dword v41, off, s[0:3], s33 offset:4 ; 4-byte Folded Spill
; GFX10-NEXT: s_waitcnt_depctr 0xffe3
; GFX10-NEXT: s_mov_b32 exec_lo, s35
-; GFX10-NEXT: v_writelane_b32 v40, s34, 6
+; GFX10-NEXT: v_writelane_b32 v40, s4, 0
+; GFX10-NEXT: s_mov_b32 s4, 3
; GFX10-NEXT: s_addk_i32 s32, 0x200
+; GFX10-NEXT: v_writelane_b32 v41, s34, 0
; GFX10-NEXT: s_getpc_b64 s[34:35]
; GFX10-NEXT: s_add_u32 s34, s34, external_void_func_v3i32_i32_inreg@rel32@lo+4
; GFX10-NEXT: s_addc_u32 s35, s35, external_void_func_v3i32_i32_inreg@rel32@hi+12
-; GFX10-NEXT: v_writelane_b32 v40, s4, 0
-; GFX10-NEXT: s_mov_b32 s4, 3
; GFX10-NEXT: v_writelane_b32 v40, s5, 1
; GFX10-NEXT: s_mov_b32 s5, 4
; GFX10-NEXT: v_writelane_b32 v40, s6, 2
@@ -11104,9 +12018,11 @@
; GFX10-NEXT: v_readlane_b32 s6, v40, 2
; GFX10-NEXT: v_readlane_b32 s5, v40, 1
; GFX10-NEXT: v_readlane_b32 s4, v40, 0
-; GFX10-NEXT: v_readlane_b32 s34, v40, 6
+; GFX10-NEXT: v_readlane_b32 s34, v41, 0
; GFX10-NEXT: s_or_saveexec_b32 s35, -1
-; GFX10-NEXT: buffer_load_dword v40, off, s[0:3], s33 ; 4-byte Folded Reload
+; GFX10-NEXT: s_clause 0x1
+; GFX10-NEXT: buffer_load_dword v40, off, s[0:3], s33
+; GFX10-NEXT: buffer_load_dword v41, off, s[0:3], s33 offset:4
; GFX10-NEXT: s_waitcnt_depctr 0xffe3
; GFX10-NEXT: s_mov_b32 exec_lo, s35
; GFX10-NEXT: s_addk_i32 s32, 0xfe00
@@ -11120,15 +12036,17 @@
; GFX11-NEXT: s_mov_b32 s0, s33
; GFX11-NEXT: s_mov_b32 s33, s32
; GFX11-NEXT: s_or_saveexec_b32 s1, -1
-; GFX11-NEXT: scratch_store_b32 off, v40, s33 ; 4-byte Folded Spill
+; GFX11-NEXT: s_clause 0x1
+; GFX11-NEXT: scratch_store_b32 off, v40, s33
+; GFX11-NEXT: scratch_store_b32 off, v41, s33 offset:4
; GFX11-NEXT: s_mov_b32 exec_lo, s1
-; GFX11-NEXT: v_writelane_b32 v40, s0, 6
+; GFX11-NEXT: v_writelane_b32 v40, s4, 0
+; GFX11-NEXT: s_mov_b32 s4, 3
; GFX11-NEXT: s_add_i32 s32, s32, 16
+; GFX11-NEXT: v_writelane_b32 v41, s0, 0
; GFX11-NEXT: s_getpc_b64 s[0:1]
; GFX11-NEXT: s_add_u32 s0, s0, external_void_func_v3i32_i32_inreg@rel32@lo+4
; GFX11-NEXT: s_addc_u32 s1, s1, external_void_func_v3i32_i32_inreg@rel32@hi+12
-; GFX11-NEXT: v_writelane_b32 v40, s4, 0
-; GFX11-NEXT: s_mov_b32 s4, 3
; GFX11-NEXT: v_writelane_b32 v40, s5, 1
; GFX11-NEXT: s_mov_b32 s5, 4
; GFX11-NEXT: v_writelane_b32 v40, s6, 2
@@ -11145,9 +12063,11 @@
; GFX11-NEXT: v_readlane_b32 s6, v40, 2
; GFX11-NEXT: v_readlane_b32 s5, v40, 1
; GFX11-NEXT: v_readlane_b32 s4, v40, 0
-; GFX11-NEXT: v_readlane_b32 s0, v40, 6
+; GFX11-NEXT: v_readlane_b32 s0, v41, 0
; GFX11-NEXT: s_or_saveexec_b32 s1, -1
-; GFX11-NEXT: scratch_load_b32 v40, off, s33 ; 4-byte Folded Reload
+; GFX11-NEXT: s_clause 0x1
+; GFX11-NEXT: scratch_load_b32 v40, off, s33
+; GFX11-NEXT: scratch_load_b32 v41, off, s33 offset:4
; GFX11-NEXT: s_mov_b32 exec_lo, s1
; GFX11-NEXT: s_add_i32 s32, s32, -16
; GFX11-NEXT: s_mov_b32 s33, s0
@@ -11161,15 +12081,16 @@
; GFX10-SCRATCH-NEXT: s_mov_b32 s33, s32
; GFX10-SCRATCH-NEXT: s_or_saveexec_b32 s1, -1
; GFX10-SCRATCH-NEXT: scratch_store_dword off, v40, s33 ; 4-byte Folded Spill
+; GFX10-SCRATCH-NEXT: scratch_store_dword off, v41, s33 offset:4 ; 4-byte Folded Spill
; GFX10-SCRATCH-NEXT: s_waitcnt_depctr 0xffe3
; GFX10-SCRATCH-NEXT: s_mov_b32 exec_lo, s1
-; GFX10-SCRATCH-NEXT: v_writelane_b32 v40, s0, 6
+; GFX10-SCRATCH-NEXT: v_writelane_b32 v40, s4, 0
+; GFX10-SCRATCH-NEXT: s_mov_b32 s4, 3
; GFX10-SCRATCH-NEXT: s_add_i32 s32, s32, 16
+; GFX10-SCRATCH-NEXT: v_writelane_b32 v41, s0, 0
; GFX10-SCRATCH-NEXT: s_getpc_b64 s[0:1]
; GFX10-SCRATCH-NEXT: s_add_u32 s0, s0, external_void_func_v3i32_i32_inreg@rel32@lo+4
; GFX10-SCRATCH-NEXT: s_addc_u32 s1, s1, external_void_func_v3i32_i32_inreg@rel32@hi+12
-; GFX10-SCRATCH-NEXT: v_writelane_b32 v40, s4, 0
-; GFX10-SCRATCH-NEXT: s_mov_b32 s4, 3
; GFX10-SCRATCH-NEXT: v_writelane_b32 v40, s5, 1
; GFX10-SCRATCH-NEXT: s_mov_b32 s5, 4
; GFX10-SCRATCH-NEXT: v_writelane_b32 v40, s6, 2
@@ -11185,9 +12106,11 @@
; GFX10-SCRATCH-NEXT: v_readlane_b32 s6, v40, 2
; GFX10-SCRATCH-NEXT: v_readlane_b32 s5, v40, 1
; GFX10-SCRATCH-NEXT: v_readlane_b32 s4, v40, 0
-; GFX10-SCRATCH-NEXT: v_readlane_b32 s0, v40, 6
+; GFX10-SCRATCH-NEXT: v_readlane_b32 s0, v41, 0
; GFX10-SCRATCH-NEXT: s_or_saveexec_b32 s1, -1
-; GFX10-SCRATCH-NEXT: scratch_load_dword v40, off, s33 ; 4-byte Folded Reload
+; GFX10-SCRATCH-NEXT: s_clause 0x1
+; GFX10-SCRATCH-NEXT: scratch_load_dword v40, off, s33
+; GFX10-SCRATCH-NEXT: scratch_load_dword v41, off, s33 offset:4
; GFX10-SCRATCH-NEXT: s_waitcnt_depctr 0xffe3
; GFX10-SCRATCH-NEXT: s_mov_b32 exec_lo, s1
; GFX10-SCRATCH-NEXT: s_add_i32 s32, s32, -16
@@ -11206,8 +12129,8 @@
; GFX9-NEXT: s_mov_b32 s33, s32
; GFX9-NEXT: s_or_saveexec_b64 s[36:37], -1
; GFX9-NEXT: buffer_store_dword v40, off, s[0:3], s33 ; 4-byte Folded Spill
+; GFX9-NEXT: buffer_store_dword v41, off, s[0:3], s33 offset:4 ; 4-byte Folded Spill
; GFX9-NEXT: s_mov_b64 exec, s[36:37]
-; GFX9-NEXT: v_writelane_b32 v40, s34, 6
; GFX9-NEXT: v_writelane_b32 v40, s4, 0
; GFX9-NEXT: v_writelane_b32 v40, s5, 1
; GFX9-NEXT: v_writelane_b32 v40, s6, 2
@@ -11215,6 +12138,7 @@
; GFX9-NEXT: s_load_dwordx4 s[4:7], s[34:35], 0x0
; GFX9-NEXT: s_addk_i32 s32, 0x400
; GFX9-NEXT: v_writelane_b32 v40, s30, 4
+; GFX9-NEXT: v_writelane_b32 v41, s34, 0
; GFX9-NEXT: v_writelane_b32 v40, s31, 5
; GFX9-NEXT: s_getpc_b64 s[34:35]
; GFX9-NEXT: s_add_u32 s34, s34, external_void_func_v4i32_inreg@rel32@lo+4
@@ -11226,9 +12150,10 @@
; GFX9-NEXT: v_readlane_b32 s6, v40, 2
; GFX9-NEXT: v_readlane_b32 s5, v40, 1
; GFX9-NEXT: v_readlane_b32 s4, v40, 0
-; GFX9-NEXT: v_readlane_b32 s34, v40, 6
+; GFX9-NEXT: v_readlane_b32 s34, v41, 0
; GFX9-NEXT: s_or_saveexec_b64 s[36:37], -1
; GFX9-NEXT: buffer_load_dword v40, off, s[0:3], s33 ; 4-byte Folded Reload
+; GFX9-NEXT: buffer_load_dword v41, off, s[0:3], s33 offset:4 ; 4-byte Folded Reload
; GFX9-NEXT: s_mov_b64 exec, s[36:37]
; GFX9-NEXT: s_addk_i32 s32, 0xfc00
; GFX9-NEXT: s_mov_b32 s33, s34
@@ -11242,11 +12167,12 @@
; GFX10-NEXT: s_mov_b32 s33, s32
; GFX10-NEXT: s_or_saveexec_b32 s35, -1
; GFX10-NEXT: buffer_store_dword v40, off, s[0:3], s33 ; 4-byte Folded Spill
+; GFX10-NEXT: buffer_store_dword v41, off, s[0:3], s33 offset:4 ; 4-byte Folded Spill
; GFX10-NEXT: s_waitcnt_depctr 0xffe3
; GFX10-NEXT: s_mov_b32 exec_lo, s35
-; GFX10-NEXT: v_writelane_b32 v40, s34, 6
-; GFX10-NEXT: s_addk_i32 s32, 0x200
; GFX10-NEXT: v_writelane_b32 v40, s4, 0
+; GFX10-NEXT: s_addk_i32 s32, 0x200
+; GFX10-NEXT: v_writelane_b32 v41, s34, 0
; GFX10-NEXT: v_writelane_b32 v40, s5, 1
; GFX10-NEXT: v_writelane_b32 v40, s6, 2
; GFX10-NEXT: v_writelane_b32 v40, s7, 3
@@ -11263,9 +12189,11 @@
; GFX10-NEXT: v_readlane_b32 s6, v40, 2
; GFX10-NEXT: v_readlane_b32 s5, v40, 1
; GFX10-NEXT: v_readlane_b32 s4, v40, 0
-; GFX10-NEXT: v_readlane_b32 s34, v40, 6
+; GFX10-NEXT: v_readlane_b32 s34, v41, 0
; GFX10-NEXT: s_or_saveexec_b32 s35, -1
-; GFX10-NEXT: buffer_load_dword v40, off, s[0:3], s33 ; 4-byte Folded Reload
+; GFX10-NEXT: s_clause 0x1
+; GFX10-NEXT: buffer_load_dword v40, off, s[0:3], s33
+; GFX10-NEXT: buffer_load_dword v41, off, s[0:3], s33 offset:4
; GFX10-NEXT: s_waitcnt_depctr 0xffe3
; GFX10-NEXT: s_mov_b32 exec_lo, s35
; GFX10-NEXT: s_addk_i32 s32, 0xfe00
@@ -11279,11 +12207,13 @@
; GFX11-NEXT: s_mov_b32 s0, s33
; GFX11-NEXT: s_mov_b32 s33, s32
; GFX11-NEXT: s_or_saveexec_b32 s1, -1
-; GFX11-NEXT: scratch_store_b32 off, v40, s33 ; 4-byte Folded Spill
+; GFX11-NEXT: s_clause 0x1
+; GFX11-NEXT: scratch_store_b32 off, v40, s33
+; GFX11-NEXT: scratch_store_b32 off, v41, s33 offset:4
; GFX11-NEXT: s_mov_b32 exec_lo, s1
-; GFX11-NEXT: v_writelane_b32 v40, s0, 6
-; GFX11-NEXT: s_add_i32 s32, s32, 16
; GFX11-NEXT: v_writelane_b32 v40, s4, 0
+; GFX11-NEXT: s_add_i32 s32, s32, 16
+; GFX11-NEXT: v_writelane_b32 v41, s0, 0
; GFX11-NEXT: v_writelane_b32 v40, s5, 1
; GFX11-NEXT: v_writelane_b32 v40, s6, 2
; GFX11-NEXT: v_writelane_b32 v40, s7, 3
@@ -11301,9 +12231,11 @@
; GFX11-NEXT: v_readlane_b32 s6, v40, 2
; GFX11-NEXT: v_readlane_b32 s5, v40, 1
; GFX11-NEXT: v_readlane_b32 s4, v40, 0
-; GFX11-NEXT: v_readlane_b32 s0, v40, 6
+; GFX11-NEXT: v_readlane_b32 s0, v41, 0
; GFX11-NEXT: s_or_saveexec_b32 s1, -1
-; GFX11-NEXT: scratch_load_b32 v40, off, s33 ; 4-byte Folded Reload
+; GFX11-NEXT: s_clause 0x1
+; GFX11-NEXT: scratch_load_b32 v40, off, s33
+; GFX11-NEXT: scratch_load_b32 v41, off, s33 offset:4
; GFX11-NEXT: s_mov_b32 exec_lo, s1
; GFX11-NEXT: s_add_i32 s32, s32, -16
; GFX11-NEXT: s_mov_b32 s33, s0
@@ -11317,11 +12249,12 @@
; GFX10-SCRATCH-NEXT: s_mov_b32 s33, s32
; GFX10-SCRATCH-NEXT: s_or_saveexec_b32 s1, -1
; GFX10-SCRATCH-NEXT: scratch_store_dword off, v40, s33 ; 4-byte Folded Spill
+; GFX10-SCRATCH-NEXT: scratch_store_dword off, v41, s33 offset:4 ; 4-byte Folded Spill
; GFX10-SCRATCH-NEXT: s_waitcnt_depctr 0xffe3
; GFX10-SCRATCH-NEXT: s_mov_b32 exec_lo, s1
-; GFX10-SCRATCH-NEXT: v_writelane_b32 v40, s0, 6
-; GFX10-SCRATCH-NEXT: s_add_i32 s32, s32, 16
; GFX10-SCRATCH-NEXT: v_writelane_b32 v40, s4, 0
+; GFX10-SCRATCH-NEXT: s_add_i32 s32, s32, 16
+; GFX10-SCRATCH-NEXT: v_writelane_b32 v41, s0, 0
; GFX10-SCRATCH-NEXT: v_writelane_b32 v40, s5, 1
; GFX10-SCRATCH-NEXT: v_writelane_b32 v40, s6, 2
; GFX10-SCRATCH-NEXT: v_writelane_b32 v40, s7, 3
@@ -11338,9 +12271,11 @@
; GFX10-SCRATCH-NEXT: v_readlane_b32 s6, v40, 2
; GFX10-SCRATCH-NEXT: v_readlane_b32 s5, v40, 1
; GFX10-SCRATCH-NEXT: v_readlane_b32 s4, v40, 0
-; GFX10-SCRATCH-NEXT: v_readlane_b32 s0, v40, 6
+; GFX10-SCRATCH-NEXT: v_readlane_b32 s0, v41, 0
; GFX10-SCRATCH-NEXT: s_or_saveexec_b32 s1, -1
-; GFX10-SCRATCH-NEXT: scratch_load_dword v40, off, s33 ; 4-byte Folded Reload
+; GFX10-SCRATCH-NEXT: s_clause 0x1
+; GFX10-SCRATCH-NEXT: scratch_load_dword v40, off, s33
+; GFX10-SCRATCH-NEXT: scratch_load_dword v41, off, s33 offset:4
; GFX10-SCRATCH-NEXT: s_waitcnt_depctr 0xffe3
; GFX10-SCRATCH-NEXT: s_mov_b32 exec_lo, s1
; GFX10-SCRATCH-NEXT: s_add_i32 s32, s32, -16
@@ -11360,8 +12295,8 @@
; GFX9-NEXT: s_mov_b32 s33, s32
; GFX9-NEXT: s_or_saveexec_b64 s[36:37], -1
; GFX9-NEXT: buffer_store_dword v40, off, s[0:3], s33 ; 4-byte Folded Spill
+; GFX9-NEXT: buffer_store_dword v41, off, s[0:3], s33 offset:4 ; 4-byte Folded Spill
; GFX9-NEXT: s_mov_b64 exec, s[36:37]
-; GFX9-NEXT: v_writelane_b32 v40, s34, 6
; GFX9-NEXT: v_writelane_b32 v40, s4, 0
; GFX9-NEXT: v_writelane_b32 v40, s5, 1
; GFX9-NEXT: v_writelane_b32 v40, s6, 2
@@ -11372,6 +12307,7 @@
; GFX9-NEXT: s_mov_b32 s5, 2
; GFX9-NEXT: s_mov_b32 s6, 3
; GFX9-NEXT: s_mov_b32 s7, 4
+; GFX9-NEXT: v_writelane_b32 v41, s34, 0
; GFX9-NEXT: v_writelane_b32 v40, s31, 5
; GFX9-NEXT: s_getpc_b64 s[34:35]
; GFX9-NEXT: s_add_u32 s34, s34, external_void_func_v4i32_inreg@rel32@lo+4
@@ -11383,9 +12319,10 @@
; GFX9-NEXT: v_readlane_b32 s6, v40, 2
; GFX9-NEXT: v_readlane_b32 s5, v40, 1
; GFX9-NEXT: v_readlane_b32 s4, v40, 0
-; GFX9-NEXT: v_readlane_b32 s34, v40, 6
+; GFX9-NEXT: v_readlane_b32 s34, v41, 0
; GFX9-NEXT: s_or_saveexec_b64 s[36:37], -1
; GFX9-NEXT: buffer_load_dword v40, off, s[0:3], s33 ; 4-byte Folded Reload
+; GFX9-NEXT: buffer_load_dword v41, off, s[0:3], s33 offset:4 ; 4-byte Folded Reload
; GFX9-NEXT: s_mov_b64 exec, s[36:37]
; GFX9-NEXT: s_addk_i32 s32, 0xfc00
; GFX9-NEXT: s_mov_b32 s33, s34
@@ -11399,15 +12336,16 @@
; GFX10-NEXT: s_mov_b32 s33, s32
; GFX10-NEXT: s_or_saveexec_b32 s35, -1
; GFX10-NEXT: buffer_store_dword v40, off, s[0:3], s33 ; 4-byte Folded Spill
+; GFX10-NEXT: buffer_store_dword v41, off, s[0:3], s33 offset:4 ; 4-byte Folded Spill
; GFX10-NEXT: s_waitcnt_depctr 0xffe3
; GFX10-NEXT: s_mov_b32 exec_lo, s35
-; GFX10-NEXT: v_writelane_b32 v40, s34, 6
+; GFX10-NEXT: v_writelane_b32 v40, s4, 0
+; GFX10-NEXT: s_mov_b32 s4, 1
; GFX10-NEXT: s_addk_i32 s32, 0x200
+; GFX10-NEXT: v_writelane_b32 v41, s34, 0
; GFX10-NEXT: s_getpc_b64 s[34:35]
; GFX10-NEXT: s_add_u32 s34, s34, external_void_func_v4i32_inreg@rel32@lo+4
; GFX10-NEXT: s_addc_u32 s35, s35, external_void_func_v4i32_inreg@rel32@hi+12
-; GFX10-NEXT: v_writelane_b32 v40, s4, 0
-; GFX10-NEXT: s_mov_b32 s4, 1
; GFX10-NEXT: v_writelane_b32 v40, s5, 1
; GFX10-NEXT: s_mov_b32 s5, 2
; GFX10-NEXT: v_writelane_b32 v40, s6, 2
@@ -11423,9 +12361,11 @@
; GFX10-NEXT: v_readlane_b32 s6, v40, 2
; GFX10-NEXT: v_readlane_b32 s5, v40, 1
; GFX10-NEXT: v_readlane_b32 s4, v40, 0
-; GFX10-NEXT: v_readlane_b32 s34, v40, 6
+; GFX10-NEXT: v_readlane_b32 s34, v41, 0
; GFX10-NEXT: s_or_saveexec_b32 s35, -1
-; GFX10-NEXT: buffer_load_dword v40, off, s[0:3], s33 ; 4-byte Folded Reload
+; GFX10-NEXT: s_clause 0x1
+; GFX10-NEXT: buffer_load_dword v40, off, s[0:3], s33
+; GFX10-NEXT: buffer_load_dword v41, off, s[0:3], s33 offset:4
; GFX10-NEXT: s_waitcnt_depctr 0xffe3
; GFX10-NEXT: s_mov_b32 exec_lo, s35
; GFX10-NEXT: s_addk_i32 s32, 0xfe00
@@ -11439,15 +12379,17 @@
; GFX11-NEXT: s_mov_b32 s0, s33
; GFX11-NEXT: s_mov_b32 s33, s32
; GFX11-NEXT: s_or_saveexec_b32 s1, -1
-; GFX11-NEXT: scratch_store_b32 off, v40, s33 ; 4-byte Folded Spill
+; GFX11-NEXT: s_clause 0x1
+; GFX11-NEXT: scratch_store_b32 off, v40, s33
+; GFX11-NEXT: scratch_store_b32 off, v41, s33 offset:4
; GFX11-NEXT: s_mov_b32 exec_lo, s1
-; GFX11-NEXT: v_writelane_b32 v40, s0, 6
+; GFX11-NEXT: v_writelane_b32 v40, s4, 0
+; GFX11-NEXT: s_mov_b32 s4, 1
; GFX11-NEXT: s_add_i32 s32, s32, 16
+; GFX11-NEXT: v_writelane_b32 v41, s0, 0
; GFX11-NEXT: s_getpc_b64 s[0:1]
; GFX11-NEXT: s_add_u32 s0, s0, external_void_func_v4i32_inreg@rel32@lo+4
; GFX11-NEXT: s_addc_u32 s1, s1, external_void_func_v4i32_inreg@rel32@hi+12
-; GFX11-NEXT: v_writelane_b32 v40, s4, 0
-; GFX11-NEXT: s_mov_b32 s4, 1
; GFX11-NEXT: v_writelane_b32 v40, s5, 1
; GFX11-NEXT: s_mov_b32 s5, 2
; GFX11-NEXT: v_writelane_b32 v40, s6, 2
@@ -11464,9 +12406,11 @@
; GFX11-NEXT: v_readlane_b32 s6, v40, 2
; GFX11-NEXT: v_readlane_b32 s5, v40, 1
; GFX11-NEXT: v_readlane_b32 s4, v40, 0
-; GFX11-NEXT: v_readlane_b32 s0, v40, 6
+; GFX11-NEXT: v_readlane_b32 s0, v41, 0
; GFX11-NEXT: s_or_saveexec_b32 s1, -1
-; GFX11-NEXT: scratch_load_b32 v40, off, s33 ; 4-byte Folded Reload
+; GFX11-NEXT: s_clause 0x1
+; GFX11-NEXT: scratch_load_b32 v40, off, s33
+; GFX11-NEXT: scratch_load_b32 v41, off, s33 offset:4
; GFX11-NEXT: s_mov_b32 exec_lo, s1
; GFX11-NEXT: s_add_i32 s32, s32, -16
; GFX11-NEXT: s_mov_b32 s33, s0
@@ -11480,15 +12424,16 @@
; GFX10-SCRATCH-NEXT: s_mov_b32 s33, s32
; GFX10-SCRATCH-NEXT: s_or_saveexec_b32 s1, -1
; GFX10-SCRATCH-NEXT: scratch_store_dword off, v40, s33 ; 4-byte Folded Spill
+; GFX10-SCRATCH-NEXT: scratch_store_dword off, v41, s33 offset:4 ; 4-byte Folded Spill
; GFX10-SCRATCH-NEXT: s_waitcnt_depctr 0xffe3
; GFX10-SCRATCH-NEXT: s_mov_b32 exec_lo, s1
-; GFX10-SCRATCH-NEXT: v_writelane_b32 v40, s0, 6
+; GFX10-SCRATCH-NEXT: v_writelane_b32 v40, s4, 0
+; GFX10-SCRATCH-NEXT: s_mov_b32 s4, 1
; GFX10-SCRATCH-NEXT: s_add_i32 s32, s32, 16
+; GFX10-SCRATCH-NEXT: v_writelane_b32 v41, s0, 0
; GFX10-SCRATCH-NEXT: s_getpc_b64 s[0:1]
; GFX10-SCRATCH-NEXT: s_add_u32 s0, s0, external_void_func_v4i32_inreg@rel32@lo+4
; GFX10-SCRATCH-NEXT: s_addc_u32 s1, s1, external_void_func_v4i32_inreg@rel32@hi+12
-; GFX10-SCRATCH-NEXT: v_writelane_b32 v40, s4, 0
-; GFX10-SCRATCH-NEXT: s_mov_b32 s4, 1
; GFX10-SCRATCH-NEXT: v_writelane_b32 v40, s5, 1
; GFX10-SCRATCH-NEXT: s_mov_b32 s5, 2
; GFX10-SCRATCH-NEXT: v_writelane_b32 v40, s6, 2
@@ -11504,9 +12449,11 @@
; GFX10-SCRATCH-NEXT: v_readlane_b32 s6, v40, 2
; GFX10-SCRATCH-NEXT: v_readlane_b32 s5, v40, 1
; GFX10-SCRATCH-NEXT: v_readlane_b32 s4, v40, 0
-; GFX10-SCRATCH-NEXT: v_readlane_b32 s0, v40, 6
+; GFX10-SCRATCH-NEXT: v_readlane_b32 s0, v41, 0
; GFX10-SCRATCH-NEXT: s_or_saveexec_b32 s1, -1
-; GFX10-SCRATCH-NEXT: scratch_load_dword v40, off, s33 ; 4-byte Folded Reload
+; GFX10-SCRATCH-NEXT: s_clause 0x1
+; GFX10-SCRATCH-NEXT: scratch_load_dword v40, off, s33
+; GFX10-SCRATCH-NEXT: scratch_load_dword v41, off, s33 offset:4
; GFX10-SCRATCH-NEXT: s_waitcnt_depctr 0xffe3
; GFX10-SCRATCH-NEXT: s_mov_b32 exec_lo, s1
; GFX10-SCRATCH-NEXT: s_add_i32 s32, s32, -16
@@ -11525,8 +12472,8 @@
; GFX9-NEXT: s_mov_b32 s33, s32
; GFX9-NEXT: s_or_saveexec_b64 s[36:37], -1
; GFX9-NEXT: buffer_store_dword v40, off, s[0:3], s33 ; 4-byte Folded Spill
+; GFX9-NEXT: buffer_store_dword v41, off, s[0:3], s33 offset:4 ; 4-byte Folded Spill
; GFX9-NEXT: s_mov_b64 exec, s[36:37]
-; GFX9-NEXT: v_writelane_b32 v40, s34, 7
; GFX9-NEXT: v_writelane_b32 v40, s4, 0
; GFX9-NEXT: v_writelane_b32 v40, s5, 1
; GFX9-NEXT: v_writelane_b32 v40, s6, 2
@@ -11539,6 +12486,7 @@
; GFX9-NEXT: s_mov_b32 s6, 3
; GFX9-NEXT: s_mov_b32 s7, 4
; GFX9-NEXT: s_mov_b32 s8, 5
+; GFX9-NEXT: v_writelane_b32 v41, s34, 0
; GFX9-NEXT: v_writelane_b32 v40, s31, 6
; GFX9-NEXT: s_getpc_b64 s[34:35]
; GFX9-NEXT: s_add_u32 s34, s34, external_void_func_v5i32_inreg@rel32@lo+4
@@ -11551,9 +12499,10 @@
; GFX9-NEXT: v_readlane_b32 s6, v40, 2
; GFX9-NEXT: v_readlane_b32 s5, v40, 1
; GFX9-NEXT: v_readlane_b32 s4, v40, 0
-; GFX9-NEXT: v_readlane_b32 s34, v40, 7
+; GFX9-NEXT: v_readlane_b32 s34, v41, 0
; GFX9-NEXT: s_or_saveexec_b64 s[36:37], -1
; GFX9-NEXT: buffer_load_dword v40, off, s[0:3], s33 ; 4-byte Folded Reload
+; GFX9-NEXT: buffer_load_dword v41, off, s[0:3], s33 offset:4 ; 4-byte Folded Reload
; GFX9-NEXT: s_mov_b64 exec, s[36:37]
; GFX9-NEXT: s_addk_i32 s32, 0xfc00
; GFX9-NEXT: s_mov_b32 s33, s34
@@ -11567,15 +12516,16 @@
; GFX10-NEXT: s_mov_b32 s33, s32
; GFX10-NEXT: s_or_saveexec_b32 s35, -1
; GFX10-NEXT: buffer_store_dword v40, off, s[0:3], s33 ; 4-byte Folded Spill
+; GFX10-NEXT: buffer_store_dword v41, off, s[0:3], s33 offset:4 ; 4-byte Folded Spill
; GFX10-NEXT: s_waitcnt_depctr 0xffe3
; GFX10-NEXT: s_mov_b32 exec_lo, s35
-; GFX10-NEXT: v_writelane_b32 v40, s34, 7
+; GFX10-NEXT: v_writelane_b32 v40, s4, 0
+; GFX10-NEXT: s_mov_b32 s4, 1
; GFX10-NEXT: s_addk_i32 s32, 0x200
+; GFX10-NEXT: v_writelane_b32 v41, s34, 0
; GFX10-NEXT: s_getpc_b64 s[34:35]
; GFX10-NEXT: s_add_u32 s34, s34, external_void_func_v5i32_inreg@rel32@lo+4
; GFX10-NEXT: s_addc_u32 s35, s35, external_void_func_v5i32_inreg@rel32@hi+12
-; GFX10-NEXT: v_writelane_b32 v40, s4, 0
-; GFX10-NEXT: s_mov_b32 s4, 1
; GFX10-NEXT: v_writelane_b32 v40, s5, 1
; GFX10-NEXT: s_mov_b32 s5, 2
; GFX10-NEXT: v_writelane_b32 v40, s6, 2
@@ -11594,9 +12544,11 @@
; GFX10-NEXT: v_readlane_b32 s6, v40, 2
; GFX10-NEXT: v_readlane_b32 s5, v40, 1
; GFX10-NEXT: v_readlane_b32 s4, v40, 0
-; GFX10-NEXT: v_readlane_b32 s34, v40, 7
+; GFX10-NEXT: v_readlane_b32 s34, v41, 0
; GFX10-NEXT: s_or_saveexec_b32 s35, -1
-; GFX10-NEXT: buffer_load_dword v40, off, s[0:3], s33 ; 4-byte Folded Reload
+; GFX10-NEXT: s_clause 0x1
+; GFX10-NEXT: buffer_load_dword v40, off, s[0:3], s33
+; GFX10-NEXT: buffer_load_dword v41, off, s[0:3], s33 offset:4
; GFX10-NEXT: s_waitcnt_depctr 0xffe3
; GFX10-NEXT: s_mov_b32 exec_lo, s35
; GFX10-NEXT: s_addk_i32 s32, 0xfe00
@@ -11610,15 +12562,17 @@
; GFX11-NEXT: s_mov_b32 s0, s33
; GFX11-NEXT: s_mov_b32 s33, s32
; GFX11-NEXT: s_or_saveexec_b32 s1, -1
-; GFX11-NEXT: scratch_store_b32 off, v40, s33 ; 4-byte Folded Spill
+; GFX11-NEXT: s_clause 0x1
+; GFX11-NEXT: scratch_store_b32 off, v40, s33
+; GFX11-NEXT: scratch_store_b32 off, v41, s33 offset:4
; GFX11-NEXT: s_mov_b32 exec_lo, s1
-; GFX11-NEXT: v_writelane_b32 v40, s0, 7
+; GFX11-NEXT: v_writelane_b32 v40, s4, 0
+; GFX11-NEXT: s_mov_b32 s4, 1
; GFX11-NEXT: s_add_i32 s32, s32, 16
+; GFX11-NEXT: v_writelane_b32 v41, s0, 0
; GFX11-NEXT: s_getpc_b64 s[0:1]
; GFX11-NEXT: s_add_u32 s0, s0, external_void_func_v5i32_inreg@rel32@lo+4
; GFX11-NEXT: s_addc_u32 s1, s1, external_void_func_v5i32_inreg@rel32@hi+12
-; GFX11-NEXT: v_writelane_b32 v40, s4, 0
-; GFX11-NEXT: s_mov_b32 s4, 1
; GFX11-NEXT: v_writelane_b32 v40, s5, 1
; GFX11-NEXT: s_mov_b32 s5, 2
; GFX11-NEXT: v_writelane_b32 v40, s6, 2
@@ -11638,9 +12592,11 @@
; GFX11-NEXT: v_readlane_b32 s6, v40, 2
; GFX11-NEXT: v_readlane_b32 s5, v40, 1
; GFX11-NEXT: v_readlane_b32 s4, v40, 0
-; GFX11-NEXT: v_readlane_b32 s0, v40, 7
+; GFX11-NEXT: v_readlane_b32 s0, v41, 0
; GFX11-NEXT: s_or_saveexec_b32 s1, -1
-; GFX11-NEXT: scratch_load_b32 v40, off, s33 ; 4-byte Folded Reload
+; GFX11-NEXT: s_clause 0x1
+; GFX11-NEXT: scratch_load_b32 v40, off, s33
+; GFX11-NEXT: scratch_load_b32 v41, off, s33 offset:4
; GFX11-NEXT: s_mov_b32 exec_lo, s1
; GFX11-NEXT: s_add_i32 s32, s32, -16
; GFX11-NEXT: s_mov_b32 s33, s0
@@ -11654,15 +12610,16 @@
; GFX10-SCRATCH-NEXT: s_mov_b32 s33, s32
; GFX10-SCRATCH-NEXT: s_or_saveexec_b32 s1, -1
; GFX10-SCRATCH-NEXT: scratch_store_dword off, v40, s33 ; 4-byte Folded Spill
+; GFX10-SCRATCH-NEXT: scratch_store_dword off, v41, s33 offset:4 ; 4-byte Folded Spill
; GFX10-SCRATCH-NEXT: s_waitcnt_depctr 0xffe3
; GFX10-SCRATCH-NEXT: s_mov_b32 exec_lo, s1
-; GFX10-SCRATCH-NEXT: v_writelane_b32 v40, s0, 7
+; GFX10-SCRATCH-NEXT: v_writelane_b32 v40, s4, 0
+; GFX10-SCRATCH-NEXT: s_mov_b32 s4, 1
; GFX10-SCRATCH-NEXT: s_add_i32 s32, s32, 16
+; GFX10-SCRATCH-NEXT: v_writelane_b32 v41, s0, 0
; GFX10-SCRATCH-NEXT: s_getpc_b64 s[0:1]
; GFX10-SCRATCH-NEXT: s_add_u32 s0, s0, external_void_func_v5i32_inreg@rel32@lo+4
; GFX10-SCRATCH-NEXT: s_addc_u32 s1, s1, external_void_func_v5i32_inreg@rel32@hi+12
-; GFX10-SCRATCH-NEXT: v_writelane_b32 v40, s4, 0
-; GFX10-SCRATCH-NEXT: s_mov_b32 s4, 1
; GFX10-SCRATCH-NEXT: v_writelane_b32 v40, s5, 1
; GFX10-SCRATCH-NEXT: s_mov_b32 s5, 2
; GFX10-SCRATCH-NEXT: v_writelane_b32 v40, s6, 2
@@ -11681,9 +12638,11 @@
; GFX10-SCRATCH-NEXT: v_readlane_b32 s6, v40, 2
; GFX10-SCRATCH-NEXT: v_readlane_b32 s5, v40, 1
; GFX10-SCRATCH-NEXT: v_readlane_b32 s4, v40, 0
-; GFX10-SCRATCH-NEXT: v_readlane_b32 s0, v40, 7
+; GFX10-SCRATCH-NEXT: v_readlane_b32 s0, v41, 0
; GFX10-SCRATCH-NEXT: s_or_saveexec_b32 s1, -1
-; GFX10-SCRATCH-NEXT: scratch_load_dword v40, off, s33 ; 4-byte Folded Reload
+; GFX10-SCRATCH-NEXT: s_clause 0x1
+; GFX10-SCRATCH-NEXT: scratch_load_dword v40, off, s33
+; GFX10-SCRATCH-NEXT: scratch_load_dword v41, off, s33 offset:4
; GFX10-SCRATCH-NEXT: s_waitcnt_depctr 0xffe3
; GFX10-SCRATCH-NEXT: s_mov_b32 exec_lo, s1
; GFX10-SCRATCH-NEXT: s_add_i32 s32, s32, -16
@@ -11702,10 +12661,11 @@
; GFX9-NEXT: s_mov_b32 s33, s32
; GFX9-NEXT: s_or_saveexec_b64 s[36:37], -1
; GFX9-NEXT: buffer_store_dword v40, off, s[0:3], s33 ; 4-byte Folded Spill
+; GFX9-NEXT: buffer_store_dword v41, off, s[0:3], s33 offset:4 ; 4-byte Folded Spill
; GFX9-NEXT: s_mov_b64 exec, s[36:37]
-; GFX9-NEXT: v_writelane_b32 v40, s34, 10
; GFX9-NEXT: v_writelane_b32 v40, s4, 0
; GFX9-NEXT: v_writelane_b32 v40, s5, 1
+; GFX9-NEXT: v_writelane_b32 v41, s34, 0
; GFX9-NEXT: v_writelane_b32 v40, s6, 2
; GFX9-NEXT: s_load_dwordx2 s[34:35], s[34:35], 0x0
; GFX9-NEXT: v_writelane_b32 v40, s7, 3
@@ -11732,9 +12692,10 @@
; GFX9-NEXT: v_readlane_b32 s6, v40, 2
; GFX9-NEXT: v_readlane_b32 s5, v40, 1
; GFX9-NEXT: v_readlane_b32 s4, v40, 0
-; GFX9-NEXT: v_readlane_b32 s34, v40, 10
+; GFX9-NEXT: v_readlane_b32 s34, v41, 0
; GFX9-NEXT: s_or_saveexec_b64 s[36:37], -1
; GFX9-NEXT: buffer_load_dword v40, off, s[0:3], s33 ; 4-byte Folded Reload
+; GFX9-NEXT: buffer_load_dword v41, off, s[0:3], s33 offset:4 ; 4-byte Folded Reload
; GFX9-NEXT: s_mov_b64 exec, s[36:37]
; GFX9-NEXT: s_addk_i32 s32, 0xfc00
; GFX9-NEXT: s_mov_b32 s33, s34
@@ -11748,12 +12709,13 @@
; GFX10-NEXT: s_mov_b32 s33, s32
; GFX10-NEXT: s_or_saveexec_b32 s35, -1
; GFX10-NEXT: buffer_store_dword v40, off, s[0:3], s33 ; 4-byte Folded Spill
+; GFX10-NEXT: buffer_store_dword v41, off, s[0:3], s33 offset:4 ; 4-byte Folded Spill
; GFX10-NEXT: s_waitcnt_depctr 0xffe3
; GFX10-NEXT: s_mov_b32 exec_lo, s35
-; GFX10-NEXT: v_writelane_b32 v40, s34, 10
+; GFX10-NEXT: v_writelane_b32 v40, s4, 0
+; GFX10-NEXT: v_writelane_b32 v41, s34, 0
; GFX10-NEXT: s_load_dwordx2 s[34:35], s[34:35], 0x0
; GFX10-NEXT: s_addk_i32 s32, 0x200
-; GFX10-NEXT: v_writelane_b32 v40, s4, 0
; GFX10-NEXT: v_writelane_b32 v40, s5, 1
; GFX10-NEXT: v_writelane_b32 v40, s6, 2
; GFX10-NEXT: v_writelane_b32 v40, s7, 3
@@ -11779,9 +12741,11 @@
; GFX10-NEXT: v_readlane_b32 s6, v40, 2
; GFX10-NEXT: v_readlane_b32 s5, v40, 1
; GFX10-NEXT: v_readlane_b32 s4, v40, 0
-; GFX10-NEXT: v_readlane_b32 s34, v40, 10
+; GFX10-NEXT: v_readlane_b32 s34, v41, 0
; GFX10-NEXT: s_or_saveexec_b32 s35, -1
-; GFX10-NEXT: buffer_load_dword v40, off, s[0:3], s33 ; 4-byte Folded Reload
+; GFX10-NEXT: s_clause 0x1
+; GFX10-NEXT: buffer_load_dword v40, off, s[0:3], s33
+; GFX10-NEXT: buffer_load_dword v41, off, s[0:3], s33 offset:4
; GFX10-NEXT: s_waitcnt_depctr 0xffe3
; GFX10-NEXT: s_mov_b32 exec_lo, s35
; GFX10-NEXT: s_addk_i32 s32, 0xfe00
@@ -11795,12 +12759,14 @@
; GFX11-NEXT: s_mov_b32 s0, s33
; GFX11-NEXT: s_mov_b32 s33, s32
; GFX11-NEXT: s_or_saveexec_b32 s1, -1
-; GFX11-NEXT: scratch_store_b32 off, v40, s33 ; 4-byte Folded Spill
+; GFX11-NEXT: s_clause 0x1
+; GFX11-NEXT: scratch_store_b32 off, v40, s33
+; GFX11-NEXT: scratch_store_b32 off, v41, s33 offset:4
; GFX11-NEXT: s_mov_b32 exec_lo, s1
-; GFX11-NEXT: v_writelane_b32 v40, s0, 10
+; GFX11-NEXT: v_writelane_b32 v40, s4, 0
+; GFX11-NEXT: v_writelane_b32 v41, s0, 0
; GFX11-NEXT: s_load_b64 s[0:1], s[0:1], 0x0
; GFX11-NEXT: s_add_i32 s32, s32, 16
-; GFX11-NEXT: v_writelane_b32 v40, s4, 0
; GFX11-NEXT: v_writelane_b32 v40, s5, 1
; GFX11-NEXT: v_writelane_b32 v40, s6, 2
; GFX11-NEXT: v_writelane_b32 v40, s7, 3
@@ -11827,9 +12793,11 @@
; GFX11-NEXT: v_readlane_b32 s6, v40, 2
; GFX11-NEXT: v_readlane_b32 s5, v40, 1
; GFX11-NEXT: v_readlane_b32 s4, v40, 0
-; GFX11-NEXT: v_readlane_b32 s0, v40, 10
+; GFX11-NEXT: v_readlane_b32 s0, v41, 0
; GFX11-NEXT: s_or_saveexec_b32 s1, -1
-; GFX11-NEXT: scratch_load_b32 v40, off, s33 ; 4-byte Folded Reload
+; GFX11-NEXT: s_clause 0x1
+; GFX11-NEXT: scratch_load_b32 v40, off, s33
+; GFX11-NEXT: scratch_load_b32 v41, off, s33 offset:4
; GFX11-NEXT: s_mov_b32 exec_lo, s1
; GFX11-NEXT: s_add_i32 s32, s32, -16
; GFX11-NEXT: s_mov_b32 s33, s0
@@ -11843,12 +12811,13 @@
; GFX10-SCRATCH-NEXT: s_mov_b32 s33, s32
; GFX10-SCRATCH-NEXT: s_or_saveexec_b32 s1, -1
; GFX10-SCRATCH-NEXT: scratch_store_dword off, v40, s33 ; 4-byte Folded Spill
+; GFX10-SCRATCH-NEXT: scratch_store_dword off, v41, s33 offset:4 ; 4-byte Folded Spill
; GFX10-SCRATCH-NEXT: s_waitcnt_depctr 0xffe3
; GFX10-SCRATCH-NEXT: s_mov_b32 exec_lo, s1
-; GFX10-SCRATCH-NEXT: v_writelane_b32 v40, s0, 10
+; GFX10-SCRATCH-NEXT: v_writelane_b32 v40, s4, 0
+; GFX10-SCRATCH-NEXT: v_writelane_b32 v41, s0, 0
; GFX10-SCRATCH-NEXT: s_load_dwordx2 s[0:1], s[0:1], 0x0
; GFX10-SCRATCH-NEXT: s_add_i32 s32, s32, 16
-; GFX10-SCRATCH-NEXT: v_writelane_b32 v40, s4, 0
; GFX10-SCRATCH-NEXT: v_writelane_b32 v40, s5, 1
; GFX10-SCRATCH-NEXT: v_writelane_b32 v40, s6, 2
; GFX10-SCRATCH-NEXT: v_writelane_b32 v40, s7, 3
@@ -11874,9 +12843,11 @@
; GFX10-SCRATCH-NEXT: v_readlane_b32 s6, v40, 2
; GFX10-SCRATCH-NEXT: v_readlane_b32 s5, v40, 1
; GFX10-SCRATCH-NEXT: v_readlane_b32 s4, v40, 0
-; GFX10-SCRATCH-NEXT: v_readlane_b32 s0, v40, 10
+; GFX10-SCRATCH-NEXT: v_readlane_b32 s0, v41, 0
; GFX10-SCRATCH-NEXT: s_or_saveexec_b32 s1, -1
-; GFX10-SCRATCH-NEXT: scratch_load_dword v40, off, s33 ; 4-byte Folded Reload
+; GFX10-SCRATCH-NEXT: s_clause 0x1
+; GFX10-SCRATCH-NEXT: scratch_load_dword v40, off, s33
+; GFX10-SCRATCH-NEXT: scratch_load_dword v41, off, s33 offset:4
; GFX10-SCRATCH-NEXT: s_waitcnt_depctr 0xffe3
; GFX10-SCRATCH-NEXT: s_mov_b32 exec_lo, s1
; GFX10-SCRATCH-NEXT: s_add_i32 s32, s32, -16
@@ -11897,8 +12868,8 @@
; GFX9-NEXT: s_mov_b32 s33, s32
; GFX9-NEXT: s_or_saveexec_b64 s[36:37], -1
; GFX9-NEXT: buffer_store_dword v40, off, s[0:3], s33 ; 4-byte Folded Spill
+; GFX9-NEXT: buffer_store_dword v41, off, s[0:3], s33 offset:4 ; 4-byte Folded Spill
; GFX9-NEXT: s_mov_b64 exec, s[36:37]
-; GFX9-NEXT: v_writelane_b32 v40, s34, 10
; GFX9-NEXT: v_writelane_b32 v40, s4, 0
; GFX9-NEXT: v_writelane_b32 v40, s5, 1
; GFX9-NEXT: v_writelane_b32 v40, s6, 2
@@ -11917,6 +12888,7 @@
; GFX9-NEXT: s_mov_b32 s9, 6
; GFX9-NEXT: s_mov_b32 s10, 7
; GFX9-NEXT: s_mov_b32 s11, 8
+; GFX9-NEXT: v_writelane_b32 v41, s34, 0
; GFX9-NEXT: v_writelane_b32 v40, s31, 9
; GFX9-NEXT: s_getpc_b64 s[34:35]
; GFX9-NEXT: s_add_u32 s34, s34, external_void_func_v8i32_inreg@rel32@lo+4
@@ -11932,9 +12904,10 @@
; GFX9-NEXT: v_readlane_b32 s6, v40, 2
; GFX9-NEXT: v_readlane_b32 s5, v40, 1
; GFX9-NEXT: v_readlane_b32 s4, v40, 0
-; GFX9-NEXT: v_readlane_b32 s34, v40, 10
+; GFX9-NEXT: v_readlane_b32 s34, v41, 0
; GFX9-NEXT: s_or_saveexec_b64 s[36:37], -1
; GFX9-NEXT: buffer_load_dword v40, off, s[0:3], s33 ; 4-byte Folded Reload
+; GFX9-NEXT: buffer_load_dword v41, off, s[0:3], s33 offset:4 ; 4-byte Folded Reload
; GFX9-NEXT: s_mov_b64 exec, s[36:37]
; GFX9-NEXT: s_addk_i32 s32, 0xfc00
; GFX9-NEXT: s_mov_b32 s33, s34
@@ -11948,15 +12921,16 @@
; GFX10-NEXT: s_mov_b32 s33, s32
; GFX10-NEXT: s_or_saveexec_b32 s35, -1
; GFX10-NEXT: buffer_store_dword v40, off, s[0:3], s33 ; 4-byte Folded Spill
+; GFX10-NEXT: buffer_store_dword v41, off, s[0:3], s33 offset:4 ; 4-byte Folded Spill
; GFX10-NEXT: s_waitcnt_depctr 0xffe3
; GFX10-NEXT: s_mov_b32 exec_lo, s35
-; GFX10-NEXT: v_writelane_b32 v40, s34, 10
+; GFX10-NEXT: v_writelane_b32 v40, s4, 0
+; GFX10-NEXT: s_mov_b32 s4, 1
; GFX10-NEXT: s_addk_i32 s32, 0x200
+; GFX10-NEXT: v_writelane_b32 v41, s34, 0
; GFX10-NEXT: s_getpc_b64 s[34:35]
; GFX10-NEXT: s_add_u32 s34, s34, external_void_func_v8i32_inreg@rel32@lo+4
; GFX10-NEXT: s_addc_u32 s35, s35, external_void_func_v8i32_inreg@rel32@hi+12
-; GFX10-NEXT: v_writelane_b32 v40, s4, 0
-; GFX10-NEXT: s_mov_b32 s4, 1
; GFX10-NEXT: v_writelane_b32 v40, s5, 1
; GFX10-NEXT: s_mov_b32 s5, 2
; GFX10-NEXT: v_writelane_b32 v40, s6, 2
@@ -11984,9 +12958,11 @@
; GFX10-NEXT: v_readlane_b32 s6, v40, 2
; GFX10-NEXT: v_readlane_b32 s5, v40, 1
; GFX10-NEXT: v_readlane_b32 s4, v40, 0
-; GFX10-NEXT: v_readlane_b32 s34, v40, 10
+; GFX10-NEXT: v_readlane_b32 s34, v41, 0
; GFX10-NEXT: s_or_saveexec_b32 s35, -1
-; GFX10-NEXT: buffer_load_dword v40, off, s[0:3], s33 ; 4-byte Folded Reload
+; GFX10-NEXT: s_clause 0x1
+; GFX10-NEXT: buffer_load_dword v40, off, s[0:3], s33
+; GFX10-NEXT: buffer_load_dword v41, off, s[0:3], s33 offset:4
; GFX10-NEXT: s_waitcnt_depctr 0xffe3
; GFX10-NEXT: s_mov_b32 exec_lo, s35
; GFX10-NEXT: s_addk_i32 s32, 0xfe00
@@ -12000,15 +12976,17 @@
; GFX11-NEXT: s_mov_b32 s0, s33
; GFX11-NEXT: s_mov_b32 s33, s32
; GFX11-NEXT: s_or_saveexec_b32 s1, -1
-; GFX11-NEXT: scratch_store_b32 off, v40, s33 ; 4-byte Folded Spill
+; GFX11-NEXT: s_clause 0x1
+; GFX11-NEXT: scratch_store_b32 off, v40, s33
+; GFX11-NEXT: scratch_store_b32 off, v41, s33 offset:4
; GFX11-NEXT: s_mov_b32 exec_lo, s1
-; GFX11-NEXT: v_writelane_b32 v40, s0, 10
+; GFX11-NEXT: v_writelane_b32 v40, s4, 0
+; GFX11-NEXT: s_mov_b32 s4, 1
; GFX11-NEXT: s_add_i32 s32, s32, 16
+; GFX11-NEXT: v_writelane_b32 v41, s0, 0
; GFX11-NEXT: s_getpc_b64 s[0:1]
; GFX11-NEXT: s_add_u32 s0, s0, external_void_func_v8i32_inreg@rel32@lo+4
; GFX11-NEXT: s_addc_u32 s1, s1, external_void_func_v8i32_inreg@rel32@hi+12
-; GFX11-NEXT: v_writelane_b32 v40, s4, 0
-; GFX11-NEXT: s_mov_b32 s4, 1
; GFX11-NEXT: v_writelane_b32 v40, s5, 1
; GFX11-NEXT: s_mov_b32 s5, 2
; GFX11-NEXT: v_writelane_b32 v40, s6, 2
@@ -12037,9 +13015,11 @@
; GFX11-NEXT: v_readlane_b32 s6, v40, 2
; GFX11-NEXT: v_readlane_b32 s5, v40, 1
; GFX11-NEXT: v_readlane_b32 s4, v40, 0
-; GFX11-NEXT: v_readlane_b32 s0, v40, 10
+; GFX11-NEXT: v_readlane_b32 s0, v41, 0
; GFX11-NEXT: s_or_saveexec_b32 s1, -1
-; GFX11-NEXT: scratch_load_b32 v40, off, s33 ; 4-byte Folded Reload
+; GFX11-NEXT: s_clause 0x1
+; GFX11-NEXT: scratch_load_b32 v40, off, s33
+; GFX11-NEXT: scratch_load_b32 v41, off, s33 offset:4
; GFX11-NEXT: s_mov_b32 exec_lo, s1
; GFX11-NEXT: s_add_i32 s32, s32, -16
; GFX11-NEXT: s_mov_b32 s33, s0
@@ -12053,15 +13033,16 @@
; GFX10-SCRATCH-NEXT: s_mov_b32 s33, s32
; GFX10-SCRATCH-NEXT: s_or_saveexec_b32 s1, -1
; GFX10-SCRATCH-NEXT: scratch_store_dword off, v40, s33 ; 4-byte Folded Spill
+; GFX10-SCRATCH-NEXT: scratch_store_dword off, v41, s33 offset:4 ; 4-byte Folded Spill
; GFX10-SCRATCH-NEXT: s_waitcnt_depctr 0xffe3
; GFX10-SCRATCH-NEXT: s_mov_b32 exec_lo, s1
-; GFX10-SCRATCH-NEXT: v_writelane_b32 v40, s0, 10
+; GFX10-SCRATCH-NEXT: v_writelane_b32 v40, s4, 0
+; GFX10-SCRATCH-NEXT: s_mov_b32 s4, 1
; GFX10-SCRATCH-NEXT: s_add_i32 s32, s32, 16
+; GFX10-SCRATCH-NEXT: v_writelane_b32 v41, s0, 0
; GFX10-SCRATCH-NEXT: s_getpc_b64 s[0:1]
; GFX10-SCRATCH-NEXT: s_add_u32 s0, s0, external_void_func_v8i32_inreg@rel32@lo+4
; GFX10-SCRATCH-NEXT: s_addc_u32 s1, s1, external_void_func_v8i32_inreg@rel32@hi+12
-; GFX10-SCRATCH-NEXT: v_writelane_b32 v40, s4, 0
-; GFX10-SCRATCH-NEXT: s_mov_b32 s4, 1
; GFX10-SCRATCH-NEXT: v_writelane_b32 v40, s5, 1
; GFX10-SCRATCH-NEXT: s_mov_b32 s5, 2
; GFX10-SCRATCH-NEXT: v_writelane_b32 v40, s6, 2
@@ -12089,9 +13070,11 @@
; GFX10-SCRATCH-NEXT: v_readlane_b32 s6, v40, 2
; GFX10-SCRATCH-NEXT: v_readlane_b32 s5, v40, 1
; GFX10-SCRATCH-NEXT: v_readlane_b32 s4, v40, 0
-; GFX10-SCRATCH-NEXT: v_readlane_b32 s0, v40, 10
+; GFX10-SCRATCH-NEXT: v_readlane_b32 s0, v41, 0
; GFX10-SCRATCH-NEXT: s_or_saveexec_b32 s1, -1
-; GFX10-SCRATCH-NEXT: scratch_load_dword v40, off, s33 ; 4-byte Folded Reload
+; GFX10-SCRATCH-NEXT: s_clause 0x1
+; GFX10-SCRATCH-NEXT: scratch_load_dword v40, off, s33
+; GFX10-SCRATCH-NEXT: scratch_load_dword v41, off, s33 offset:4
; GFX10-SCRATCH-NEXT: s_waitcnt_depctr 0xffe3
; GFX10-SCRATCH-NEXT: s_mov_b32 exec_lo, s1
; GFX10-SCRATCH-NEXT: s_add_i32 s32, s32, -16
@@ -12110,8 +13093,8 @@
; GFX9-NEXT: s_mov_b32 s33, s32
; GFX9-NEXT: s_or_saveexec_b64 s[36:37], -1
; GFX9-NEXT: buffer_store_dword v40, off, s[0:3], s33 ; 4-byte Folded Spill
+; GFX9-NEXT: buffer_store_dword v41, off, s[0:3], s33 offset:4 ; 4-byte Folded Spill
; GFX9-NEXT: s_mov_b64 exec, s[36:37]
-; GFX9-NEXT: v_writelane_b32 v40, s34, 18
; GFX9-NEXT: v_writelane_b32 v40, s4, 0
; GFX9-NEXT: v_writelane_b32 v40, s5, 1
; GFX9-NEXT: v_writelane_b32 v40, s6, 2
@@ -12122,6 +13105,7 @@
; GFX9-NEXT: v_writelane_b32 v40, s11, 7
; GFX9-NEXT: v_writelane_b32 v40, s12, 8
; GFX9-NEXT: v_writelane_b32 v40, s13, 9
+; GFX9-NEXT: v_writelane_b32 v41, s34, 0
; GFX9-NEXT: v_writelane_b32 v40, s14, 10
; GFX9-NEXT: s_load_dwordx2 s[34:35], s[34:35], 0x0
; GFX9-NEXT: v_writelane_b32 v40, s15, 11
@@ -12156,9 +13140,10 @@
; GFX9-NEXT: v_readlane_b32 s6, v40, 2
; GFX9-NEXT: v_readlane_b32 s5, v40, 1
; GFX9-NEXT: v_readlane_b32 s4, v40, 0
-; GFX9-NEXT: v_readlane_b32 s34, v40, 18
+; GFX9-NEXT: v_readlane_b32 s34, v41, 0
; GFX9-NEXT: s_or_saveexec_b64 s[36:37], -1
; GFX9-NEXT: buffer_load_dword v40, off, s[0:3], s33 ; 4-byte Folded Reload
+; GFX9-NEXT: buffer_load_dword v41, off, s[0:3], s33 offset:4 ; 4-byte Folded Reload
; GFX9-NEXT: s_mov_b64 exec, s[36:37]
; GFX9-NEXT: s_addk_i32 s32, 0xfc00
; GFX9-NEXT: s_mov_b32 s33, s34
@@ -12172,12 +13157,13 @@
; GFX10-NEXT: s_mov_b32 s33, s32
; GFX10-NEXT: s_or_saveexec_b32 s35, -1
; GFX10-NEXT: buffer_store_dword v40, off, s[0:3], s33 ; 4-byte Folded Spill
+; GFX10-NEXT: buffer_store_dword v41, off, s[0:3], s33 offset:4 ; 4-byte Folded Spill
; GFX10-NEXT: s_waitcnt_depctr 0xffe3
; GFX10-NEXT: s_mov_b32 exec_lo, s35
-; GFX10-NEXT: v_writelane_b32 v40, s34, 18
+; GFX10-NEXT: v_writelane_b32 v40, s4, 0
+; GFX10-NEXT: v_writelane_b32 v41, s34, 0
; GFX10-NEXT: s_load_dwordx2 s[34:35], s[34:35], 0x0
; GFX10-NEXT: s_addk_i32 s32, 0x200
-; GFX10-NEXT: v_writelane_b32 v40, s4, 0
; GFX10-NEXT: v_writelane_b32 v40, s5, 1
; GFX10-NEXT: v_writelane_b32 v40, s6, 2
; GFX10-NEXT: v_writelane_b32 v40, s7, 3
@@ -12219,9 +13205,11 @@
; GFX10-NEXT: v_readlane_b32 s6, v40, 2
; GFX10-NEXT: v_readlane_b32 s5, v40, 1
; GFX10-NEXT: v_readlane_b32 s4, v40, 0
-; GFX10-NEXT: v_readlane_b32 s34, v40, 18
+; GFX10-NEXT: v_readlane_b32 s34, v41, 0
; GFX10-NEXT: s_or_saveexec_b32 s35, -1
-; GFX10-NEXT: buffer_load_dword v40, off, s[0:3], s33 ; 4-byte Folded Reload
+; GFX10-NEXT: s_clause 0x1
+; GFX10-NEXT: buffer_load_dword v40, off, s[0:3], s33
+; GFX10-NEXT: buffer_load_dword v41, off, s[0:3], s33 offset:4
; GFX10-NEXT: s_waitcnt_depctr 0xffe3
; GFX10-NEXT: s_mov_b32 exec_lo, s35
; GFX10-NEXT: s_addk_i32 s32, 0xfe00
@@ -12235,12 +13223,14 @@
; GFX11-NEXT: s_mov_b32 s0, s33
; GFX11-NEXT: s_mov_b32 s33, s32
; GFX11-NEXT: s_or_saveexec_b32 s1, -1
-; GFX11-NEXT: scratch_store_b32 off, v40, s33 ; 4-byte Folded Spill
+; GFX11-NEXT: s_clause 0x1
+; GFX11-NEXT: scratch_store_b32 off, v40, s33
+; GFX11-NEXT: scratch_store_b32 off, v41, s33 offset:4
; GFX11-NEXT: s_mov_b32 exec_lo, s1
-; GFX11-NEXT: v_writelane_b32 v40, s0, 18
+; GFX11-NEXT: v_writelane_b32 v40, s4, 0
+; GFX11-NEXT: v_writelane_b32 v41, s0, 0
; GFX11-NEXT: s_load_b64 s[0:1], s[0:1], 0x0
; GFX11-NEXT: s_add_i32 s32, s32, 16
-; GFX11-NEXT: v_writelane_b32 v40, s4, 0
; GFX11-NEXT: v_writelane_b32 v40, s5, 1
; GFX11-NEXT: v_writelane_b32 v40, s6, 2
; GFX11-NEXT: v_writelane_b32 v40, s7, 3
@@ -12283,9 +13273,11 @@
; GFX11-NEXT: v_readlane_b32 s6, v40, 2
; GFX11-NEXT: v_readlane_b32 s5, v40, 1
; GFX11-NEXT: v_readlane_b32 s4, v40, 0
-; GFX11-NEXT: v_readlane_b32 s0, v40, 18
+; GFX11-NEXT: v_readlane_b32 s0, v41, 0
; GFX11-NEXT: s_or_saveexec_b32 s1, -1
-; GFX11-NEXT: scratch_load_b32 v40, off, s33 ; 4-byte Folded Reload
+; GFX11-NEXT: s_clause 0x1
+; GFX11-NEXT: scratch_load_b32 v40, off, s33
+; GFX11-NEXT: scratch_load_b32 v41, off, s33 offset:4
; GFX11-NEXT: s_mov_b32 exec_lo, s1
; GFX11-NEXT: s_add_i32 s32, s32, -16
; GFX11-NEXT: s_mov_b32 s33, s0
@@ -12299,12 +13291,13 @@
; GFX10-SCRATCH-NEXT: s_mov_b32 s33, s32
; GFX10-SCRATCH-NEXT: s_or_saveexec_b32 s1, -1
; GFX10-SCRATCH-NEXT: scratch_store_dword off, v40, s33 ; 4-byte Folded Spill
+; GFX10-SCRATCH-NEXT: scratch_store_dword off, v41, s33 offset:4 ; 4-byte Folded Spill
; GFX10-SCRATCH-NEXT: s_waitcnt_depctr 0xffe3
; GFX10-SCRATCH-NEXT: s_mov_b32 exec_lo, s1
-; GFX10-SCRATCH-NEXT: v_writelane_b32 v40, s0, 18
+; GFX10-SCRATCH-NEXT: v_writelane_b32 v40, s4, 0
+; GFX10-SCRATCH-NEXT: v_writelane_b32 v41, s0, 0
; GFX10-SCRATCH-NEXT: s_load_dwordx2 s[0:1], s[0:1], 0x0
; GFX10-SCRATCH-NEXT: s_add_i32 s32, s32, 16
-; GFX10-SCRATCH-NEXT: v_writelane_b32 v40, s4, 0
; GFX10-SCRATCH-NEXT: v_writelane_b32 v40, s5, 1
; GFX10-SCRATCH-NEXT: v_writelane_b32 v40, s6, 2
; GFX10-SCRATCH-NEXT: v_writelane_b32 v40, s7, 3
@@ -12346,9 +13339,11 @@
; GFX10-SCRATCH-NEXT: v_readlane_b32 s6, v40, 2
; GFX10-SCRATCH-NEXT: v_readlane_b32 s5, v40, 1
; GFX10-SCRATCH-NEXT: v_readlane_b32 s4, v40, 0
-; GFX10-SCRATCH-NEXT: v_readlane_b32 s0, v40, 18
+; GFX10-SCRATCH-NEXT: v_readlane_b32 s0, v41, 0
; GFX10-SCRATCH-NEXT: s_or_saveexec_b32 s1, -1
-; GFX10-SCRATCH-NEXT: scratch_load_dword v40, off, s33 ; 4-byte Folded Reload
+; GFX10-SCRATCH-NEXT: s_clause 0x1
+; GFX10-SCRATCH-NEXT: scratch_load_dword v40, off, s33
+; GFX10-SCRATCH-NEXT: scratch_load_dword v41, off, s33 offset:4
; GFX10-SCRATCH-NEXT: s_waitcnt_depctr 0xffe3
; GFX10-SCRATCH-NEXT: s_mov_b32 exec_lo, s1
; GFX10-SCRATCH-NEXT: s_add_i32 s32, s32, -16
@@ -12369,8 +13364,8 @@
; GFX9-NEXT: s_mov_b32 s33, s32
; GFX9-NEXT: s_or_saveexec_b64 s[36:37], -1
; GFX9-NEXT: buffer_store_dword v40, off, s[0:3], s33 ; 4-byte Folded Spill
+; GFX9-NEXT: buffer_store_dword v41, off, s[0:3], s33 offset:4 ; 4-byte Folded Spill
; GFX9-NEXT: s_mov_b64 exec, s[36:37]
-; GFX9-NEXT: v_writelane_b32 v40, s34, 28
; GFX9-NEXT: v_writelane_b32 v40, s4, 0
; GFX9-NEXT: v_writelane_b32 v40, s5, 1
; GFX9-NEXT: v_writelane_b32 v40, s6, 2
@@ -12386,6 +13381,7 @@
; GFX9-NEXT: v_writelane_b32 v40, s16, 12
; GFX9-NEXT: v_writelane_b32 v40, s17, 13
; GFX9-NEXT: v_writelane_b32 v40, s18, 14
+; GFX9-NEXT: v_writelane_b32 v41, s34, 0
; GFX9-NEXT: v_writelane_b32 v40, s19, 15
; GFX9-NEXT: s_load_dwordx2 s[34:35], s[34:35], 0x0
; GFX9-NEXT: v_writelane_b32 v40, s20, 16
@@ -12459,9 +13455,10 @@
; GFX9-NEXT: v_readlane_b32 s6, v40, 2
; GFX9-NEXT: v_readlane_b32 s5, v40, 1
; GFX9-NEXT: v_readlane_b32 s4, v40, 0
-; GFX9-NEXT: v_readlane_b32 s34, v40, 28
+; GFX9-NEXT: v_readlane_b32 s34, v41, 0
; GFX9-NEXT: s_or_saveexec_b64 s[36:37], -1
; GFX9-NEXT: buffer_load_dword v40, off, s[0:3], s33 ; 4-byte Folded Reload
+; GFX9-NEXT: buffer_load_dword v41, off, s[0:3], s33 offset:4 ; 4-byte Folded Reload
; GFX9-NEXT: s_mov_b64 exec, s[36:37]
; GFX9-NEXT: s_addk_i32 s32, 0xfc00
; GFX9-NEXT: s_mov_b32 s33, s34
@@ -12475,12 +13472,13 @@
; GFX10-NEXT: s_mov_b32 s33, s32
; GFX10-NEXT: s_or_saveexec_b32 s35, -1
; GFX10-NEXT: buffer_store_dword v40, off, s[0:3], s33 ; 4-byte Folded Spill
+; GFX10-NEXT: buffer_store_dword v41, off, s[0:3], s33 offset:4 ; 4-byte Folded Spill
; GFX10-NEXT: s_waitcnt_depctr 0xffe3
; GFX10-NEXT: s_mov_b32 exec_lo, s35
-; GFX10-NEXT: v_writelane_b32 v40, s34, 28
+; GFX10-NEXT: v_writelane_b32 v40, s4, 0
+; GFX10-NEXT: v_writelane_b32 v41, s34, 0
; GFX10-NEXT: s_load_dwordx2 s[34:35], s[34:35], 0x0
; GFX10-NEXT: s_addk_i32 s32, 0x200
-; GFX10-NEXT: v_writelane_b32 v40, s4, 0
; GFX10-NEXT: v_writelane_b32 v40, s5, 1
; GFX10-NEXT: v_writelane_b32 v40, s6, 2
; GFX10-NEXT: v_writelane_b32 v40, s7, 3
@@ -12567,9 +13565,11 @@
; GFX10-NEXT: v_readlane_b32 s6, v40, 2
; GFX10-NEXT: v_readlane_b32 s5, v40, 1
; GFX10-NEXT: v_readlane_b32 s4, v40, 0
-; GFX10-NEXT: v_readlane_b32 s34, v40, 28
+; GFX10-NEXT: v_readlane_b32 s34, v41, 0
; GFX10-NEXT: s_or_saveexec_b32 s35, -1
-; GFX10-NEXT: buffer_load_dword v40, off, s[0:3], s33 ; 4-byte Folded Reload
+; GFX10-NEXT: s_clause 0x1
+; GFX10-NEXT: buffer_load_dword v40, off, s[0:3], s33
+; GFX10-NEXT: buffer_load_dword v41, off, s[0:3], s33 offset:4
; GFX10-NEXT: s_waitcnt_depctr 0xffe3
; GFX10-NEXT: s_mov_b32 exec_lo, s35
; GFX10-NEXT: s_addk_i32 s32, 0xfe00
@@ -12583,14 +13583,16 @@
; GFX11-NEXT: s_mov_b32 s0, s33
; GFX11-NEXT: s_mov_b32 s33, s32
; GFX11-NEXT: s_or_saveexec_b32 s1, -1
-; GFX11-NEXT: scratch_store_b32 off, v40, s33 ; 4-byte Folded Spill
+; GFX11-NEXT: s_clause 0x1
+; GFX11-NEXT: scratch_store_b32 off, v40, s33
+; GFX11-NEXT: scratch_store_b32 off, v41, s33 offset:4
; GFX11-NEXT: s_mov_b32 exec_lo, s1
-; GFX11-NEXT: v_writelane_b32 v40, s0, 28
+; GFX11-NEXT: v_writelane_b32 v40, s4, 0
+; GFX11-NEXT: v_writelane_b32 v41, s0, 0
; GFX11-NEXT: s_load_b64 s[0:1], s[0:1], 0x0
; GFX11-NEXT: s_add_i32 s32, s32, 16
; GFX11-NEXT: s_delay_alu instid0(SALU_CYCLE_1)
; GFX11-NEXT: s_add_i32 s2, s32, 16
-; GFX11-NEXT: v_writelane_b32 v40, s4, 0
; GFX11-NEXT: v_writelane_b32 v40, s5, 1
; GFX11-NEXT: v_writelane_b32 v40, s6, 2
; GFX11-NEXT: v_writelane_b32 v40, s7, 3
@@ -12671,9 +13673,11 @@
; GFX11-NEXT: v_readlane_b32 s6, v40, 2
; GFX11-NEXT: v_readlane_b32 s5, v40, 1
; GFX11-NEXT: v_readlane_b32 s4, v40, 0
-; GFX11-NEXT: v_readlane_b32 s0, v40, 28
+; GFX11-NEXT: v_readlane_b32 s0, v41, 0
; GFX11-NEXT: s_or_saveexec_b32 s1, -1
-; GFX11-NEXT: scratch_load_b32 v40, off, s33 ; 4-byte Folded Reload
+; GFX11-NEXT: s_clause 0x1
+; GFX11-NEXT: scratch_load_b32 v40, off, s33
+; GFX11-NEXT: scratch_load_b32 v41, off, s33 offset:4
; GFX11-NEXT: s_mov_b32 exec_lo, s1
; GFX11-NEXT: s_add_i32 s32, s32, -16
; GFX11-NEXT: s_mov_b32 s33, s0
@@ -12687,13 +13691,14 @@
; GFX10-SCRATCH-NEXT: s_mov_b32 s33, s32
; GFX10-SCRATCH-NEXT: s_or_saveexec_b32 s1, -1
; GFX10-SCRATCH-NEXT: scratch_store_dword off, v40, s33 ; 4-byte Folded Spill
+; GFX10-SCRATCH-NEXT: scratch_store_dword off, v41, s33 offset:4 ; 4-byte Folded Spill
; GFX10-SCRATCH-NEXT: s_waitcnt_depctr 0xffe3
; GFX10-SCRATCH-NEXT: s_mov_b32 exec_lo, s1
-; GFX10-SCRATCH-NEXT: v_writelane_b32 v40, s0, 28
+; GFX10-SCRATCH-NEXT: v_writelane_b32 v40, s4, 0
+; GFX10-SCRATCH-NEXT: v_writelane_b32 v41, s0, 0
; GFX10-SCRATCH-NEXT: s_load_dwordx2 s[0:1], s[0:1], 0x0
; GFX10-SCRATCH-NEXT: s_add_i32 s32, s32, 16
; GFX10-SCRATCH-NEXT: s_add_i32 s2, s32, 16
-; GFX10-SCRATCH-NEXT: v_writelane_b32 v40, s4, 0
; GFX10-SCRATCH-NEXT: v_writelane_b32 v40, s5, 1
; GFX10-SCRATCH-NEXT: v_writelane_b32 v40, s6, 2
; GFX10-SCRATCH-NEXT: v_writelane_b32 v40, s7, 3
@@ -12776,9 +13781,11 @@
; GFX10-SCRATCH-NEXT: v_readlane_b32 s6, v40, 2
; GFX10-SCRATCH-NEXT: v_readlane_b32 s5, v40, 1
; GFX10-SCRATCH-NEXT: v_readlane_b32 s4, v40, 0
-; GFX10-SCRATCH-NEXT: v_readlane_b32 s0, v40, 28
+; GFX10-SCRATCH-NEXT: v_readlane_b32 s0, v41, 0
; GFX10-SCRATCH-NEXT: s_or_saveexec_b32 s1, -1
-; GFX10-SCRATCH-NEXT: scratch_load_dword v40, off, s33 ; 4-byte Folded Reload
+; GFX10-SCRATCH-NEXT: s_clause 0x1
+; GFX10-SCRATCH-NEXT: scratch_load_dword v40, off, s33
+; GFX10-SCRATCH-NEXT: scratch_load_dword v41, off, s33 offset:4
; GFX10-SCRATCH-NEXT: s_waitcnt_depctr 0xffe3
; GFX10-SCRATCH-NEXT: s_mov_b32 exec_lo, s1
; GFX10-SCRATCH-NEXT: s_add_i32 s32, s32, -16
@@ -12799,8 +13806,8 @@
; GFX9-NEXT: s_mov_b32 s33, s32
; GFX9-NEXT: s_or_saveexec_b64 s[36:37], -1
; GFX9-NEXT: buffer_store_dword v40, off, s[0:3], s33 ; 4-byte Folded Spill
+; GFX9-NEXT: buffer_store_dword v41, off, s[0:3], s33 offset:4 ; 4-byte Folded Spill
; GFX9-NEXT: s_mov_b64 exec, s[36:37]
-; GFX9-NEXT: v_writelane_b32 v40, s34, 28
; GFX9-NEXT: v_writelane_b32 v40, s4, 0
; GFX9-NEXT: v_writelane_b32 v40, s5, 1
; GFX9-NEXT: v_writelane_b32 v40, s6, 2
@@ -12815,6 +13822,7 @@
; GFX9-NEXT: v_writelane_b32 v40, s15, 11
; GFX9-NEXT: v_writelane_b32 v40, s16, 12
; GFX9-NEXT: v_writelane_b32 v40, s17, 13
+; GFX9-NEXT: v_writelane_b32 v41, s34, 0
; GFX9-NEXT: v_writelane_b32 v40, s18, 14
; GFX9-NEXT: s_load_dwordx2 s[34:35], s[34:35], 0x0
; GFX9-NEXT: v_writelane_b32 v40, s19, 15
@@ -12894,9 +13902,10 @@
; GFX9-NEXT: v_readlane_b32 s6, v40, 2
; GFX9-NEXT: v_readlane_b32 s5, v40, 1
; GFX9-NEXT: v_readlane_b32 s4, v40, 0
-; GFX9-NEXT: v_readlane_b32 s34, v40, 28
+; GFX9-NEXT: v_readlane_b32 s34, v41, 0
; GFX9-NEXT: s_or_saveexec_b64 s[36:37], -1
; GFX9-NEXT: buffer_load_dword v40, off, s[0:3], s33 ; 4-byte Folded Reload
+; GFX9-NEXT: buffer_load_dword v41, off, s[0:3], s33 offset:4 ; 4-byte Folded Reload
; GFX9-NEXT: s_mov_b64 exec, s[36:37]
; GFX9-NEXT: s_addk_i32 s32, 0xfc00
; GFX9-NEXT: s_mov_b32 s33, s34
@@ -12910,12 +13919,13 @@
; GFX10-NEXT: s_mov_b32 s33, s32
; GFX10-NEXT: s_or_saveexec_b32 s35, -1
; GFX10-NEXT: buffer_store_dword v40, off, s[0:3], s33 ; 4-byte Folded Spill
+; GFX10-NEXT: buffer_store_dword v41, off, s[0:3], s33 offset:4 ; 4-byte Folded Spill
; GFX10-NEXT: s_waitcnt_depctr 0xffe3
; GFX10-NEXT: s_mov_b32 exec_lo, s35
-; GFX10-NEXT: v_writelane_b32 v40, s34, 28
+; GFX10-NEXT: v_writelane_b32 v40, s4, 0
+; GFX10-NEXT: v_writelane_b32 v41, s34, 0
; GFX10-NEXT: s_load_dwordx2 s[34:35], s[34:35], 0x0
; GFX10-NEXT: s_addk_i32 s32, 0x200
-; GFX10-NEXT: v_writelane_b32 v40, s4, 0
; GFX10-NEXT: v_writelane_b32 v40, s5, 1
; GFX10-NEXT: v_writelane_b32 v40, s6, 2
; GFX10-NEXT: v_writelane_b32 v40, s7, 3
@@ -13007,9 +14017,11 @@
; GFX10-NEXT: v_readlane_b32 s6, v40, 2
; GFX10-NEXT: v_readlane_b32 s5, v40, 1
; GFX10-NEXT: v_readlane_b32 s4, v40, 0
-; GFX10-NEXT: v_readlane_b32 s34, v40, 28
+; GFX10-NEXT: v_readlane_b32 s34, v41, 0
; GFX10-NEXT: s_or_saveexec_b32 s35, -1
-; GFX10-NEXT: buffer_load_dword v40, off, s[0:3], s33 ; 4-byte Folded Reload
+; GFX10-NEXT: s_clause 0x1
+; GFX10-NEXT: buffer_load_dword v40, off, s[0:3], s33
+; GFX10-NEXT: buffer_load_dword v41, off, s[0:3], s33 offset:4
; GFX10-NEXT: s_waitcnt_depctr 0xffe3
; GFX10-NEXT: s_mov_b32 exec_lo, s35
; GFX10-NEXT: s_addk_i32 s32, 0xfe00
@@ -13023,12 +14035,14 @@
; GFX11-NEXT: s_mov_b32 s0, s33
; GFX11-NEXT: s_mov_b32 s33, s32
; GFX11-NEXT: s_or_saveexec_b32 s1, -1
-; GFX11-NEXT: scratch_store_b32 off, v40, s33 ; 4-byte Folded Spill
+; GFX11-NEXT: s_clause 0x1
+; GFX11-NEXT: scratch_store_b32 off, v40, s33
+; GFX11-NEXT: scratch_store_b32 off, v41, s33 offset:4
; GFX11-NEXT: s_mov_b32 exec_lo, s1
-; GFX11-NEXT: v_writelane_b32 v40, s0, 28
+; GFX11-NEXT: v_writelane_b32 v40, s4, 0
+; GFX11-NEXT: v_writelane_b32 v41, s0, 0
; GFX11-NEXT: s_load_b64 s[0:1], s[0:1], 0x0
; GFX11-NEXT: s_add_i32 s32, s32, 16
-; GFX11-NEXT: v_writelane_b32 v40, s4, 0
; GFX11-NEXT: v_writelane_b32 v40, s5, 1
; GFX11-NEXT: v_writelane_b32 v40, s6, 2
; GFX11-NEXT: v_writelane_b32 v40, s7, 3
@@ -13114,9 +14128,11 @@
; GFX11-NEXT: v_readlane_b32 s6, v40, 2
; GFX11-NEXT: v_readlane_b32 s5, v40, 1
; GFX11-NEXT: v_readlane_b32 s4, v40, 0
-; GFX11-NEXT: v_readlane_b32 s0, v40, 28
+; GFX11-NEXT: v_readlane_b32 s0, v41, 0
; GFX11-NEXT: s_or_saveexec_b32 s1, -1
-; GFX11-NEXT: scratch_load_b32 v40, off, s33 ; 4-byte Folded Reload
+; GFX11-NEXT: s_clause 0x1
+; GFX11-NEXT: scratch_load_b32 v40, off, s33
+; GFX11-NEXT: scratch_load_b32 v41, off, s33 offset:4
; GFX11-NEXT: s_mov_b32 exec_lo, s1
; GFX11-NEXT: s_add_i32 s32, s32, -16
; GFX11-NEXT: s_mov_b32 s33, s0
@@ -13130,12 +14146,13 @@
; GFX10-SCRATCH-NEXT: s_mov_b32 s33, s32
; GFX10-SCRATCH-NEXT: s_or_saveexec_b32 s1, -1
; GFX10-SCRATCH-NEXT: scratch_store_dword off, v40, s33 ; 4-byte Folded Spill
+; GFX10-SCRATCH-NEXT: scratch_store_dword off, v41, s33 offset:4 ; 4-byte Folded Spill
; GFX10-SCRATCH-NEXT: s_waitcnt_depctr 0xffe3
; GFX10-SCRATCH-NEXT: s_mov_b32 exec_lo, s1
-; GFX10-SCRATCH-NEXT: v_writelane_b32 v40, s0, 28
+; GFX10-SCRATCH-NEXT: v_writelane_b32 v40, s4, 0
+; GFX10-SCRATCH-NEXT: v_writelane_b32 v41, s0, 0
; GFX10-SCRATCH-NEXT: s_load_dwordx2 s[0:1], s[0:1], 0x0
; GFX10-SCRATCH-NEXT: s_add_i32 s32, s32, 16
-; GFX10-SCRATCH-NEXT: v_writelane_b32 v40, s4, 0
; GFX10-SCRATCH-NEXT: v_writelane_b32 v40, s5, 1
; GFX10-SCRATCH-NEXT: v_writelane_b32 v40, s6, 2
; GFX10-SCRATCH-NEXT: v_writelane_b32 v40, s7, 3
@@ -13225,9 +14242,11 @@
; GFX10-SCRATCH-NEXT: v_readlane_b32 s6, v40, 2
; GFX10-SCRATCH-NEXT: v_readlane_b32 s5, v40, 1
; GFX10-SCRATCH-NEXT: v_readlane_b32 s4, v40, 0
-; GFX10-SCRATCH-NEXT: v_readlane_b32 s0, v40, 28
+; GFX10-SCRATCH-NEXT: v_readlane_b32 s0, v41, 0
; GFX10-SCRATCH-NEXT: s_or_saveexec_b32 s1, -1
-; GFX10-SCRATCH-NEXT: scratch_load_dword v40, off, s33 ; 4-byte Folded Reload
+; GFX10-SCRATCH-NEXT: s_clause 0x1
+; GFX10-SCRATCH-NEXT: scratch_load_dword v40, off, s33
+; GFX10-SCRATCH-NEXT: scratch_load_dword v41, off, s33 offset:4
; GFX10-SCRATCH-NEXT: s_waitcnt_depctr 0xffe3
; GFX10-SCRATCH-NEXT: s_mov_b32 exec_lo, s1
; GFX10-SCRATCH-NEXT: s_add_i32 s32, s32, -16
@@ -13249,12 +14268,13 @@
; GFX9-NEXT: s_mov_b32 s33, s32
; GFX9-NEXT: s_or_saveexec_b64 s[36:37], -1
; GFX9-NEXT: buffer_store_dword v40, off, s[0:3], s33 offset:8 ; 4-byte Folded Spill
+; GFX9-NEXT: buffer_store_dword v41, off, s[0:3], s33 offset:12 ; 4-byte Folded Spill
; GFX9-NEXT: s_mov_b64 exec, s[36:37]
; GFX9-NEXT: buffer_load_dword v32, off, s[0:3], s33
; GFX9-NEXT: buffer_load_dword v33, off, s[0:3], s33 offset:4
-; GFX9-NEXT: v_writelane_b32 v40, s34, 2
-; GFX9-NEXT: s_addk_i32 s32, 0x400
+; GFX9-NEXT: s_addk_i32 s32, 0x800
; GFX9-NEXT: v_writelane_b32 v40, s30, 0
+; GFX9-NEXT: v_writelane_b32 v41, s34, 0
; GFX9-NEXT: v_writelane_b32 v40, s31, 1
; GFX9-NEXT: s_getpc_b64 s[34:35]
; GFX9-NEXT: s_add_u32 s34, s34, stack_passed_f64_arg@rel32@lo+4
@@ -13266,11 +14286,12 @@
; GFX9-NEXT: s_swappc_b64 s[30:31], s[34:35]
; GFX9-NEXT: v_readlane_b32 s31, v40, 1
; GFX9-NEXT: v_readlane_b32 s30, v40, 0
-; GFX9-NEXT: v_readlane_b32 s34, v40, 2
+; GFX9-NEXT: v_readlane_b32 s34, v41, 0
; GFX9-NEXT: s_or_saveexec_b64 s[36:37], -1
; GFX9-NEXT: buffer_load_dword v40, off, s[0:3], s33 offset:8 ; 4-byte Folded Reload
+; GFX9-NEXT: buffer_load_dword v41, off, s[0:3], s33 offset:12 ; 4-byte Folded Reload
; GFX9-NEXT: s_mov_b64 exec, s[36:37]
-; GFX9-NEXT: s_addk_i32 s32, 0xfc00
+; GFX9-NEXT: s_addk_i32 s32, 0xf800
; GFX9-NEXT: s_mov_b32 s33, s34
; GFX9-NEXT: s_waitcnt vmcnt(0)
; GFX9-NEXT: s_setpc_b64 s[30:31]
@@ -13282,13 +14303,15 @@
; GFX10-NEXT: s_mov_b32 s33, s32
; GFX10-NEXT: s_or_saveexec_b32 s35, -1
; GFX10-NEXT: buffer_store_dword v40, off, s[0:3], s33 offset:8 ; 4-byte Folded Spill
+; GFX10-NEXT: buffer_store_dword v41, off, s[0:3], s33 offset:12 ; 4-byte Folded Spill
; GFX10-NEXT: s_waitcnt_depctr 0xffe3
; GFX10-NEXT: s_mov_b32 exec_lo, s35
; GFX10-NEXT: s_clause 0x1
; GFX10-NEXT: buffer_load_dword v32, off, s[0:3], s33
; GFX10-NEXT: buffer_load_dword v33, off, s[0:3], s33 offset:4
-; GFX10-NEXT: v_writelane_b32 v40, s34, 2
-; GFX10-NEXT: s_addk_i32 s32, 0x200
+; GFX10-NEXT: v_writelane_b32 v40, s30, 0
+; GFX10-NEXT: s_addk_i32 s32, 0x400
+; GFX10-NEXT: v_writelane_b32 v41, s34, 0
; GFX10-NEXT: s_getpc_b64 s[34:35]
; GFX10-NEXT: s_add_u32 s34, s34, stack_passed_f64_arg@rel32@lo+4
; GFX10-NEXT: s_addc_u32 s35, s35, stack_passed_f64_arg@rel32@hi+12
@@ -13296,17 +14319,18 @@
; GFX10-NEXT: buffer_store_dword v32, off, s[0:3], s32
; GFX10-NEXT: s_waitcnt vmcnt(0)
; GFX10-NEXT: buffer_store_dword v33, off, s[0:3], s32 offset:4
-; GFX10-NEXT: v_writelane_b32 v40, s30, 0
; GFX10-NEXT: v_writelane_b32 v40, s31, 1
; GFX10-NEXT: s_swappc_b64 s[30:31], s[34:35]
; GFX10-NEXT: v_readlane_b32 s31, v40, 1
; GFX10-NEXT: v_readlane_b32 s30, v40, 0
-; GFX10-NEXT: v_readlane_b32 s34, v40, 2
+; GFX10-NEXT: v_readlane_b32 s34, v41, 0
; GFX10-NEXT: s_or_saveexec_b32 s35, -1
-; GFX10-NEXT: buffer_load_dword v40, off, s[0:3], s33 offset:8 ; 4-byte Folded Reload
+; GFX10-NEXT: s_clause 0x1
+; GFX10-NEXT: buffer_load_dword v40, off, s[0:3], s33 offset:8
+; GFX10-NEXT: buffer_load_dword v41, off, s[0:3], s33 offset:12
; GFX10-NEXT: s_waitcnt_depctr 0xffe3
; GFX10-NEXT: s_mov_b32 exec_lo, s35
-; GFX10-NEXT: s_addk_i32 s32, 0xfe00
+; GFX10-NEXT: s_addk_i32 s32, 0xfc00
; GFX10-NEXT: s_mov_b32 s33, s34
; GFX10-NEXT: s_waitcnt vmcnt(0)
; GFX10-NEXT: s_setpc_b64 s[30:31]
@@ -13317,26 +14341,30 @@
; GFX11-NEXT: s_mov_b32 s0, s33
; GFX11-NEXT: s_mov_b32 s33, s32
; GFX11-NEXT: s_or_saveexec_b32 s1, -1
-; GFX11-NEXT: scratch_store_b32 off, v40, s33 offset:8 ; 4-byte Folded Spill
+; GFX11-NEXT: s_clause 0x1
+; GFX11-NEXT: scratch_store_b32 off, v40, s33 offset:8
+; GFX11-NEXT: scratch_store_b32 off, v41, s33 offset:12
; GFX11-NEXT: s_mov_b32 exec_lo, s1
; GFX11-NEXT: scratch_load_b64 v[32:33], off, s33
-; GFX11-NEXT: v_writelane_b32 v40, s0, 2
-; GFX11-NEXT: s_add_i32 s32, s32, 16
+; GFX11-NEXT: v_writelane_b32 v40, s30, 0
+; GFX11-NEXT: s_add_i32 s32, s32, 32
+; GFX11-NEXT: v_writelane_b32 v41, s0, 0
; GFX11-NEXT: s_getpc_b64 s[0:1]
; GFX11-NEXT: s_add_u32 s0, s0, stack_passed_f64_arg@rel32@lo+4
; GFX11-NEXT: s_addc_u32 s1, s1, stack_passed_f64_arg@rel32@hi+12
-; GFX11-NEXT: v_writelane_b32 v40, s30, 0
; GFX11-NEXT: v_writelane_b32 v40, s31, 1
; GFX11-NEXT: s_waitcnt vmcnt(0)
; GFX11-NEXT: scratch_store_b64 off, v[32:33], s32
; GFX11-NEXT: s_swappc_b64 s[30:31], s[0:1]
; GFX11-NEXT: v_readlane_b32 s31, v40, 1
; GFX11-NEXT: v_readlane_b32 s30, v40, 0
-; GFX11-NEXT: v_readlane_b32 s0, v40, 2
+; GFX11-NEXT: v_readlane_b32 s0, v41, 0
; GFX11-NEXT: s_or_saveexec_b32 s1, -1
-; GFX11-NEXT: scratch_load_b32 v40, off, s33 offset:8 ; 4-byte Folded Reload
+; GFX11-NEXT: s_clause 0x1
+; GFX11-NEXT: scratch_load_b32 v40, off, s33 offset:8
+; GFX11-NEXT: scratch_load_b32 v41, off, s33 offset:12
; GFX11-NEXT: s_mov_b32 exec_lo, s1
-; GFX11-NEXT: s_add_i32 s32, s32, -16
+; GFX11-NEXT: s_addk_i32 s32, 0xffe0
; GFX11-NEXT: s_mov_b32 s33, s0
; GFX11-NEXT: s_waitcnt vmcnt(0)
; GFX11-NEXT: s_setpc_b64 s[30:31]
@@ -13348,27 +14376,30 @@
; GFX10-SCRATCH-NEXT: s_mov_b32 s33, s32
; GFX10-SCRATCH-NEXT: s_or_saveexec_b32 s1, -1
; GFX10-SCRATCH-NEXT: scratch_store_dword off, v40, s33 offset:8 ; 4-byte Folded Spill
+; GFX10-SCRATCH-NEXT: scratch_store_dword off, v41, s33 offset:12 ; 4-byte Folded Spill
; GFX10-SCRATCH-NEXT: s_waitcnt_depctr 0xffe3
; GFX10-SCRATCH-NEXT: s_mov_b32 exec_lo, s1
; GFX10-SCRATCH-NEXT: scratch_load_dwordx2 v[32:33], off, s33
-; GFX10-SCRATCH-NEXT: v_writelane_b32 v40, s0, 2
-; GFX10-SCRATCH-NEXT: s_add_i32 s32, s32, 16
+; GFX10-SCRATCH-NEXT: v_writelane_b32 v40, s30, 0
+; GFX10-SCRATCH-NEXT: s_add_i32 s32, s32, 32
+; GFX10-SCRATCH-NEXT: v_writelane_b32 v41, s0, 0
; GFX10-SCRATCH-NEXT: s_getpc_b64 s[0:1]
; GFX10-SCRATCH-NEXT: s_add_u32 s0, s0, stack_passed_f64_arg@rel32@lo+4
; GFX10-SCRATCH-NEXT: s_addc_u32 s1, s1, stack_passed_f64_arg@rel32@hi+12
-; GFX10-SCRATCH-NEXT: v_writelane_b32 v40, s30, 0
; GFX10-SCRATCH-NEXT: v_writelane_b32 v40, s31, 1
; GFX10-SCRATCH-NEXT: s_waitcnt vmcnt(0)
; GFX10-SCRATCH-NEXT: scratch_store_dwordx2 off, v[32:33], s32
; GFX10-SCRATCH-NEXT: s_swappc_b64 s[30:31], s[0:1]
; GFX10-SCRATCH-NEXT: v_readlane_b32 s31, v40, 1
; GFX10-SCRATCH-NEXT: v_readlane_b32 s30, v40, 0
-; GFX10-SCRATCH-NEXT: v_readlane_b32 s0, v40, 2
+; GFX10-SCRATCH-NEXT: v_readlane_b32 s0, v41, 0
; GFX10-SCRATCH-NEXT: s_or_saveexec_b32 s1, -1
-; GFX10-SCRATCH-NEXT: scratch_load_dword v40, off, s33 offset:8 ; 4-byte Folded Reload
+; GFX10-SCRATCH-NEXT: s_clause 0x1
+; GFX10-SCRATCH-NEXT: scratch_load_dword v40, off, s33 offset:8
+; GFX10-SCRATCH-NEXT: scratch_load_dword v41, off, s33 offset:12
; GFX10-SCRATCH-NEXT: s_waitcnt_depctr 0xffe3
; GFX10-SCRATCH-NEXT: s_mov_b32 exec_lo, s1
-; GFX10-SCRATCH-NEXT: s_add_i32 s32, s32, -16
+; GFX10-SCRATCH-NEXT: s_addk_i32 s32, 0xffe0
; GFX10-SCRATCH-NEXT: s_mov_b32 s33, s0
; GFX10-SCRATCH-NEXT: s_waitcnt vmcnt(0)
; GFX10-SCRATCH-NEXT: s_setpc_b64 s[30:31]
@@ -13385,6 +14416,7 @@
; GFX9-NEXT: s_mov_b32 s33, s32
; GFX9-NEXT: s_or_saveexec_b64 s[36:37], -1
; GFX9-NEXT: buffer_store_dword v40, off, s[0:3], s33 ; 4-byte Folded Spill
+; GFX9-NEXT: buffer_store_dword v41, off, s[0:3], s33 offset:4 ; 4-byte Folded Spill
; GFX9-NEXT: s_mov_b64 exec, s[36:37]
; GFX9-NEXT: s_addk_i32 s32, 0x400
; GFX9-NEXT: v_mov_b32_e32 v0, 12
@@ -13392,7 +14424,6 @@
; GFX9-NEXT: v_mov_b32_e32 v0, 13
; GFX9-NEXT: buffer_store_dword v0, off, s[0:3], s32 offset:4
; GFX9-NEXT: v_mov_b32_e32 v0, 14
-; GFX9-NEXT: v_writelane_b32 v40, s34, 2
; GFX9-NEXT: buffer_store_dword v0, off, s[0:3], s32 offset:8
; GFX9-NEXT: v_mov_b32_e32 v0, 15
; GFX9-NEXT: v_writelane_b32 v40, s30, 0
@@ -13429,6 +14460,7 @@
; GFX9-NEXT: v_mov_b32_e32 v29, 9
; GFX9-NEXT: v_mov_b32_e32 v30, 10
; GFX9-NEXT: v_mov_b32_e32 v31, 11
+; GFX9-NEXT: v_writelane_b32 v41, s34, 0
; GFX9-NEXT: v_writelane_b32 v40, s31, 1
; GFX9-NEXT: s_getpc_b64 s[34:35]
; GFX9-NEXT: s_add_u32 s34, s34, external_void_func_12xv3i32@rel32@lo+4
@@ -13436,9 +14468,10 @@
; GFX9-NEXT: s_swappc_b64 s[30:31], s[34:35]
; GFX9-NEXT: v_readlane_b32 s31, v40, 1
; GFX9-NEXT: v_readlane_b32 s30, v40, 0
-; GFX9-NEXT: v_readlane_b32 s34, v40, 2
+; GFX9-NEXT: v_readlane_b32 s34, v41, 0
; GFX9-NEXT: s_or_saveexec_b64 s[36:37], -1
; GFX9-NEXT: buffer_load_dword v40, off, s[0:3], s33 ; 4-byte Folded Reload
+; GFX9-NEXT: buffer_load_dword v41, off, s[0:3], s33 offset:4 ; 4-byte Folded Reload
; GFX9-NEXT: s_mov_b64 exec, s[36:37]
; GFX9-NEXT: s_addk_i32 s32, 0xfc00
; GFX9-NEXT: s_mov_b32 s33, s34
@@ -13452,9 +14485,9 @@
; GFX10-NEXT: s_mov_b32 s33, s32
; GFX10-NEXT: s_or_saveexec_b32 s35, -1
; GFX10-NEXT: buffer_store_dword v40, off, s[0:3], s33 ; 4-byte Folded Spill
+; GFX10-NEXT: buffer_store_dword v41, off, s[0:3], s33 offset:4 ; 4-byte Folded Spill
; GFX10-NEXT: s_waitcnt_depctr 0xffe3
; GFX10-NEXT: s_mov_b32 exec_lo, s35
-; GFX10-NEXT: v_writelane_b32 v40, s34, 2
; GFX10-NEXT: v_mov_b32_e32 v0, 12
; GFX10-NEXT: v_mov_b32_e32 v1, 13
; GFX10-NEXT: v_mov_b32_e32 v2, 14
@@ -13497,6 +14530,7 @@
; GFX10-NEXT: v_mov_b32_e32 v29, 9
; GFX10-NEXT: v_mov_b32_e32 v30, 10
; GFX10-NEXT: v_mov_b32_e32 v31, 11
+; GFX10-NEXT: v_writelane_b32 v41, s34, 0
; GFX10-NEXT: v_writelane_b32 v40, s31, 1
; GFX10-NEXT: s_getpc_b64 s[34:35]
; GFX10-NEXT: s_add_u32 s34, s34, external_void_func_12xv3i32@rel32@lo+4
@@ -13504,9 +14538,11 @@
; GFX10-NEXT: s_swappc_b64 s[30:31], s[34:35]
; GFX10-NEXT: v_readlane_b32 s31, v40, 1
; GFX10-NEXT: v_readlane_b32 s30, v40, 0
-; GFX10-NEXT: v_readlane_b32 s34, v40, 2
+; GFX10-NEXT: v_readlane_b32 s34, v41, 0
; GFX10-NEXT: s_or_saveexec_b32 s35, -1
-; GFX10-NEXT: buffer_load_dword v40, off, s[0:3], s33 ; 4-byte Folded Reload
+; GFX10-NEXT: s_clause 0x1
+; GFX10-NEXT: buffer_load_dword v40, off, s[0:3], s33
+; GFX10-NEXT: buffer_load_dword v41, off, s[0:3], s33 offset:4
; GFX10-NEXT: s_waitcnt_depctr 0xffe3
; GFX10-NEXT: s_mov_b32 exec_lo, s35
; GFX10-NEXT: s_addk_i32 s32, 0xfe00
@@ -13520,9 +14556,10 @@
; GFX11-NEXT: s_mov_b32 s0, s33
; GFX11-NEXT: s_mov_b32 s33, s32
; GFX11-NEXT: s_or_saveexec_b32 s1, -1
-; GFX11-NEXT: scratch_store_b32 off, v40, s33 ; 4-byte Folded Spill
+; GFX11-NEXT: s_clause 0x1
+; GFX11-NEXT: scratch_store_b32 off, v40, s33
+; GFX11-NEXT: scratch_store_b32 off, v41, s33 offset:4
; GFX11-NEXT: s_mov_b32 exec_lo, s1
-; GFX11-NEXT: v_writelane_b32 v40, s0, 2
; GFX11-NEXT: v_dual_mov_b32 v0, 12 :: v_dual_mov_b32 v1, 13
; GFX11-NEXT: v_dual_mov_b32 v2, 14 :: v_dual_mov_b32 v3, 15
; GFX11-NEXT: s_add_i32 s32, s32, 16
@@ -13544,6 +14581,7 @@
; GFX11-NEXT: v_dual_mov_b32 v26, 8 :: v_dual_mov_b32 v27, 9
; GFX11-NEXT: v_dual_mov_b32 v28, 9 :: v_dual_mov_b32 v29, 9
; GFX11-NEXT: v_dual_mov_b32 v30, 10 :: v_dual_mov_b32 v31, 11
+; GFX11-NEXT: v_writelane_b32 v41, s0, 0
; GFX11-NEXT: v_writelane_b32 v40, s31, 1
; GFX11-NEXT: s_getpc_b64 s[0:1]
; GFX11-NEXT: s_add_u32 s0, s0, external_void_func_12xv3i32@rel32@lo+4
@@ -13552,9 +14590,11 @@
; GFX11-NEXT: s_swappc_b64 s[30:31], s[0:1]
; GFX11-NEXT: v_readlane_b32 s31, v40, 1
; GFX11-NEXT: v_readlane_b32 s30, v40, 0
-; GFX11-NEXT: v_readlane_b32 s0, v40, 2
+; GFX11-NEXT: v_readlane_b32 s0, v41, 0
; GFX11-NEXT: s_or_saveexec_b32 s1, -1
-; GFX11-NEXT: scratch_load_b32 v40, off, s33 ; 4-byte Folded Reload
+; GFX11-NEXT: s_clause 0x1
+; GFX11-NEXT: scratch_load_b32 v40, off, s33
+; GFX11-NEXT: scratch_load_b32 v41, off, s33 offset:4
; GFX11-NEXT: s_mov_b32 exec_lo, s1
; GFX11-NEXT: s_add_i32 s32, s32, -16
; GFX11-NEXT: s_mov_b32 s33, s0
@@ -13568,9 +14608,9 @@
; GFX10-SCRATCH-NEXT: s_mov_b32 s33, s32
; GFX10-SCRATCH-NEXT: s_or_saveexec_b32 s1, -1
; GFX10-SCRATCH-NEXT: scratch_store_dword off, v40, s33 ; 4-byte Folded Spill
+; GFX10-SCRATCH-NEXT: scratch_store_dword off, v41, s33 offset:4 ; 4-byte Folded Spill
; GFX10-SCRATCH-NEXT: s_waitcnt_depctr 0xffe3
; GFX10-SCRATCH-NEXT: s_mov_b32 exec_lo, s1
-; GFX10-SCRATCH-NEXT: v_writelane_b32 v40, s0, 2
; GFX10-SCRATCH-NEXT: v_mov_b32_e32 v0, 12
; GFX10-SCRATCH-NEXT: v_mov_b32_e32 v1, 13
; GFX10-SCRATCH-NEXT: v_mov_b32_e32 v2, 14
@@ -13610,6 +14650,7 @@
; GFX10-SCRATCH-NEXT: v_mov_b32_e32 v29, 9
; GFX10-SCRATCH-NEXT: v_mov_b32_e32 v30, 10
; GFX10-SCRATCH-NEXT: v_mov_b32_e32 v31, 11
+; GFX10-SCRATCH-NEXT: v_writelane_b32 v41, s0, 0
; GFX10-SCRATCH-NEXT: v_writelane_b32 v40, s31, 1
; GFX10-SCRATCH-NEXT: s_getpc_b64 s[0:1]
; GFX10-SCRATCH-NEXT: s_add_u32 s0, s0, external_void_func_12xv3i32@rel32@lo+4
@@ -13617,9 +14658,11 @@
; GFX10-SCRATCH-NEXT: s_swappc_b64 s[30:31], s[0:1]
; GFX10-SCRATCH-NEXT: v_readlane_b32 s31, v40, 1
; GFX10-SCRATCH-NEXT: v_readlane_b32 s30, v40, 0
-; GFX10-SCRATCH-NEXT: v_readlane_b32 s0, v40, 2
+; GFX10-SCRATCH-NEXT: v_readlane_b32 s0, v41, 0
; GFX10-SCRATCH-NEXT: s_or_saveexec_b32 s1, -1
-; GFX10-SCRATCH-NEXT: scratch_load_dword v40, off, s33 ; 4-byte Folded Reload
+; GFX10-SCRATCH-NEXT: s_clause 0x1
+; GFX10-SCRATCH-NEXT: scratch_load_dword v40, off, s33
+; GFX10-SCRATCH-NEXT: scratch_load_dword v41, off, s33 offset:4
; GFX10-SCRATCH-NEXT: s_waitcnt_depctr 0xffe3
; GFX10-SCRATCH-NEXT: s_mov_b32 exec_lo, s1
; GFX10-SCRATCH-NEXT: s_add_i32 s32, s32, -16
@@ -13651,6 +14694,7 @@
; GFX9-NEXT: s_mov_b32 s33, s32
; GFX9-NEXT: s_or_saveexec_b64 s[36:37], -1
; GFX9-NEXT: buffer_store_dword v40, off, s[0:3], s33 ; 4-byte Folded Spill
+; GFX9-NEXT: buffer_store_dword v41, off, s[0:3], s33 offset:4 ; 4-byte Folded Spill
; GFX9-NEXT: s_mov_b64 exec, s[36:37]
; GFX9-NEXT: s_addk_i32 s32, 0x400
; GFX9-NEXT: v_mov_b32_e32 v0, 8
@@ -13666,7 +14710,6 @@
; GFX9-NEXT: v_mov_b32_e32 v0, 13
; GFX9-NEXT: buffer_store_dword v0, off, s[0:3], s32 offset:20
; GFX9-NEXT: v_mov_b32_e32 v0, 14
-; GFX9-NEXT: v_writelane_b32 v40, s34, 2
; GFX9-NEXT: buffer_store_dword v0, off, s[0:3], s32 offset:24
; GFX9-NEXT: v_mov_b32_e32 v0, 15
; GFX9-NEXT: v_writelane_b32 v40, s30, 0
@@ -13703,6 +14746,7 @@
; GFX9-NEXT: v_mov_b32_e32 v29, 5
; GFX9-NEXT: v_mov_b32_e32 v30, 6
; GFX9-NEXT: v_mov_b32_e32 v31, 7
+; GFX9-NEXT: v_writelane_b32 v41, s34, 0
; GFX9-NEXT: v_writelane_b32 v40, s31, 1
; GFX9-NEXT: s_getpc_b64 s[34:35]
; GFX9-NEXT: s_add_u32 s34, s34, external_void_func_8xv5i32@rel32@lo+4
@@ -13710,9 +14754,10 @@
; GFX9-NEXT: s_swappc_b64 s[30:31], s[34:35]
; GFX9-NEXT: v_readlane_b32 s31, v40, 1
; GFX9-NEXT: v_readlane_b32 s30, v40, 0
-; GFX9-NEXT: v_readlane_b32 s34, v40, 2
+; GFX9-NEXT: v_readlane_b32 s34, v41, 0
; GFX9-NEXT: s_or_saveexec_b64 s[36:37], -1
; GFX9-NEXT: buffer_load_dword v40, off, s[0:3], s33 ; 4-byte Folded Reload
+; GFX9-NEXT: buffer_load_dword v41, off, s[0:3], s33 offset:4 ; 4-byte Folded Reload
; GFX9-NEXT: s_mov_b64 exec, s[36:37]
; GFX9-NEXT: s_addk_i32 s32, 0xfc00
; GFX9-NEXT: s_mov_b32 s33, s34
@@ -13726,20 +14771,20 @@
; GFX10-NEXT: s_mov_b32 s33, s32
; GFX10-NEXT: s_or_saveexec_b32 s35, -1
; GFX10-NEXT: buffer_store_dword v40, off, s[0:3], s33 ; 4-byte Folded Spill
+; GFX10-NEXT: buffer_store_dword v41, off, s[0:3], s33 offset:4 ; 4-byte Folded Spill
; GFX10-NEXT: s_waitcnt_depctr 0xffe3
; GFX10-NEXT: s_mov_b32 exec_lo, s35
; GFX10-NEXT: v_mov_b32_e32 v0, 8
; GFX10-NEXT: v_mov_b32_e32 v1, 9
; GFX10-NEXT: v_mov_b32_e32 v2, 10
; GFX10-NEXT: s_addk_i32 s32, 0x200
-; GFX10-NEXT: v_writelane_b32 v40, s34, 2
+; GFX10-NEXT: v_mov_b32_e32 v3, 14
; GFX10-NEXT: buffer_store_dword v0, off, s[0:3], s32
; GFX10-NEXT: buffer_store_dword v1, off, s[0:3], s32 offset:4
; GFX10-NEXT: buffer_store_dword v2, off, s[0:3], s32 offset:8
; GFX10-NEXT: v_mov_b32_e32 v0, 11
; GFX10-NEXT: v_mov_b32_e32 v1, 12
; GFX10-NEXT: v_mov_b32_e32 v2, 13
-; GFX10-NEXT: v_mov_b32_e32 v3, 14
; GFX10-NEXT: v_mov_b32_e32 v4, 15
; GFX10-NEXT: v_writelane_b32 v40, s30, 0
; GFX10-NEXT: buffer_store_dword v0, off, s[0:3], s32 offset:12
@@ -13779,6 +14824,7 @@
; GFX10-NEXT: v_mov_b32_e32 v29, 5
; GFX10-NEXT: v_mov_b32_e32 v30, 6
; GFX10-NEXT: v_mov_b32_e32 v31, 7
+; GFX10-NEXT: v_writelane_b32 v41, s34, 0
; GFX10-NEXT: v_writelane_b32 v40, s31, 1
; GFX10-NEXT: s_getpc_b64 s[34:35]
; GFX10-NEXT: s_add_u32 s34, s34, external_void_func_8xv5i32@rel32@lo+4
@@ -13786,9 +14832,11 @@
; GFX10-NEXT: s_swappc_b64 s[30:31], s[34:35]
; GFX10-NEXT: v_readlane_b32 s31, v40, 1
; GFX10-NEXT: v_readlane_b32 s30, v40, 0
-; GFX10-NEXT: v_readlane_b32 s34, v40, 2
+; GFX10-NEXT: v_readlane_b32 s34, v41, 0
; GFX10-NEXT: s_or_saveexec_b32 s35, -1
-; GFX10-NEXT: buffer_load_dword v40, off, s[0:3], s33 ; 4-byte Folded Reload
+; GFX10-NEXT: s_clause 0x1
+; GFX10-NEXT: buffer_load_dword v40, off, s[0:3], s33
+; GFX10-NEXT: buffer_load_dword v41, off, s[0:3], s33 offset:4
; GFX10-NEXT: s_waitcnt_depctr 0xffe3
; GFX10-NEXT: s_mov_b32 exec_lo, s35
; GFX10-NEXT: s_addk_i32 s32, 0xfe00
@@ -13802,16 +14850,18 @@
; GFX11-NEXT: s_mov_b32 s0, s33
; GFX11-NEXT: s_mov_b32 s33, s32
; GFX11-NEXT: s_or_saveexec_b32 s1, -1
-; GFX11-NEXT: scratch_store_b32 off, v40, s33 ; 4-byte Folded Spill
+; GFX11-NEXT: s_clause 0x1
+; GFX11-NEXT: scratch_store_b32 off, v40, s33
+; GFX11-NEXT: scratch_store_b32 off, v41, s33 offset:4
; GFX11-NEXT: s_mov_b32 exec_lo, s1
-; GFX11-NEXT: v_writelane_b32 v40, s0, 2
; GFX11-NEXT: v_dual_mov_b32 v0, 8 :: v_dual_mov_b32 v1, 9
; GFX11-NEXT: v_dual_mov_b32 v2, 10 :: v_dual_mov_b32 v3, 11
; GFX11-NEXT: v_dual_mov_b32 v4, 12 :: v_dual_mov_b32 v5, 13
; GFX11-NEXT: v_dual_mov_b32 v6, 14 :: v_dual_mov_b32 v7, 15
; GFX11-NEXT: s_add_i32 s32, s32, 16
-; GFX11-NEXT: v_writelane_b32 v40, s30, 0
+; GFX11-NEXT: v_writelane_b32 v41, s0, 0
; GFX11-NEXT: s_add_i32 s0, s32, 16
+; GFX11-NEXT: v_writelane_b32 v40, s30, 0
; GFX11-NEXT: scratch_store_b128 off, v[0:3], s32
; GFX11-NEXT: scratch_store_b128 off, v[4:7], s0
; GFX11-NEXT: v_dual_mov_b32 v0, 0 :: v_dual_mov_b32 v1, 0
@@ -13838,9 +14888,11 @@
; GFX11-NEXT: s_swappc_b64 s[30:31], s[0:1]
; GFX11-NEXT: v_readlane_b32 s31, v40, 1
; GFX11-NEXT: v_readlane_b32 s30, v40, 0
-; GFX11-NEXT: v_readlane_b32 s0, v40, 2
+; GFX11-NEXT: v_readlane_b32 s0, v41, 0
; GFX11-NEXT: s_or_saveexec_b32 s1, -1
-; GFX11-NEXT: scratch_load_b32 v40, off, s33 ; 4-byte Folded Reload
+; GFX11-NEXT: s_clause 0x1
+; GFX11-NEXT: scratch_load_b32 v40, off, s33
+; GFX11-NEXT: scratch_load_b32 v41, off, s33 offset:4
; GFX11-NEXT: s_mov_b32 exec_lo, s1
; GFX11-NEXT: s_add_i32 s32, s32, -16
; GFX11-NEXT: s_mov_b32 s33, s0
@@ -13854,9 +14906,9 @@
; GFX10-SCRATCH-NEXT: s_mov_b32 s33, s32
; GFX10-SCRATCH-NEXT: s_or_saveexec_b32 s1, -1
; GFX10-SCRATCH-NEXT: scratch_store_dword off, v40, s33 ; 4-byte Folded Spill
+; GFX10-SCRATCH-NEXT: scratch_store_dword off, v41, s33 offset:4 ; 4-byte Folded Spill
; GFX10-SCRATCH-NEXT: s_waitcnt_depctr 0xffe3
; GFX10-SCRATCH-NEXT: s_mov_b32 exec_lo, s1
-; GFX10-SCRATCH-NEXT: v_writelane_b32 v40, s0, 2
; GFX10-SCRATCH-NEXT: v_mov_b32_e32 v0, 8
; GFX10-SCRATCH-NEXT: v_mov_b32_e32 v1, 9
; GFX10-SCRATCH-NEXT: v_mov_b32_e32 v2, 10
@@ -13866,8 +14918,9 @@
; GFX10-SCRATCH-NEXT: v_mov_b32_e32 v6, 14
; GFX10-SCRATCH-NEXT: v_mov_b32_e32 v7, 15
; GFX10-SCRATCH-NEXT: s_add_i32 s32, s32, 16
-; GFX10-SCRATCH-NEXT: v_writelane_b32 v40, s30, 0
+; GFX10-SCRATCH-NEXT: v_writelane_b32 v41, s0, 0
; GFX10-SCRATCH-NEXT: s_add_i32 s0, s32, 16
+; GFX10-SCRATCH-NEXT: v_writelane_b32 v40, s30, 0
; GFX10-SCRATCH-NEXT: scratch_store_dwordx4 off, v[0:3], s32
; GFX10-SCRATCH-NEXT: scratch_store_dwordx4 off, v[4:7], s0
; GFX10-SCRATCH-NEXT: v_mov_b32_e32 v0, 0
@@ -13909,9 +14962,11 @@
; GFX10-SCRATCH-NEXT: s_swappc_b64 s[30:31], s[0:1]
; GFX10-SCRATCH-NEXT: v_readlane_b32 s31, v40, 1
; GFX10-SCRATCH-NEXT: v_readlane_b32 s30, v40, 0
-; GFX10-SCRATCH-NEXT: v_readlane_b32 s0, v40, 2
+; GFX10-SCRATCH-NEXT: v_readlane_b32 s0, v41, 0
; GFX10-SCRATCH-NEXT: s_or_saveexec_b32 s1, -1
-; GFX10-SCRATCH-NEXT: scratch_load_dword v40, off, s33 ; 4-byte Folded Reload
+; GFX10-SCRATCH-NEXT: s_clause 0x1
+; GFX10-SCRATCH-NEXT: scratch_load_dword v40, off, s33
+; GFX10-SCRATCH-NEXT: scratch_load_dword v41, off, s33 offset:4
; GFX10-SCRATCH-NEXT: s_waitcnt_depctr 0xffe3
; GFX10-SCRATCH-NEXT: s_mov_b32 exec_lo, s1
; GFX10-SCRATCH-NEXT: s_add_i32 s32, s32, -16
@@ -13939,6 +14994,7 @@
; GFX9-NEXT: s_mov_b32 s33, s32
; GFX9-NEXT: s_or_saveexec_b64 s[36:37], -1
; GFX9-NEXT: buffer_store_dword v40, off, s[0:3], s33 ; 4-byte Folded Spill
+; GFX9-NEXT: buffer_store_dword v41, off, s[0:3], s33 offset:4 ; 4-byte Folded Spill
; GFX9-NEXT: s_mov_b64 exec, s[36:37]
; GFX9-NEXT: s_addk_i32 s32, 0x400
; GFX9-NEXT: v_mov_b32_e32 v0, 0x41000000
@@ -13954,7 +15010,6 @@
; GFX9-NEXT: v_mov_b32_e32 v0, 0x41500000
; GFX9-NEXT: buffer_store_dword v0, off, s[0:3], s32 offset:20
; GFX9-NEXT: v_mov_b32_e32 v0, 0x41600000
-; GFX9-NEXT: v_writelane_b32 v40, s34, 2
; GFX9-NEXT: buffer_store_dword v0, off, s[0:3], s32 offset:24
; GFX9-NEXT: v_mov_b32_e32 v0, 0x41700000
; GFX9-NEXT: v_writelane_b32 v40, s30, 0
@@ -13991,6 +15046,7 @@
; GFX9-NEXT: v_mov_b32_e32 v29, 0x40a00000
; GFX9-NEXT: v_mov_b32_e32 v30, 0x40c00000
; GFX9-NEXT: v_mov_b32_e32 v31, 0x40e00000
+; GFX9-NEXT: v_writelane_b32 v41, s34, 0
; GFX9-NEXT: v_writelane_b32 v40, s31, 1
; GFX9-NEXT: s_getpc_b64 s[34:35]
; GFX9-NEXT: s_add_u32 s34, s34, external_void_func_8xv5f32@rel32@lo+4
@@ -13998,9 +15054,10 @@
; GFX9-NEXT: s_swappc_b64 s[30:31], s[34:35]
; GFX9-NEXT: v_readlane_b32 s31, v40, 1
; GFX9-NEXT: v_readlane_b32 s30, v40, 0
-; GFX9-NEXT: v_readlane_b32 s34, v40, 2
+; GFX9-NEXT: v_readlane_b32 s34, v41, 0
; GFX9-NEXT: s_or_saveexec_b64 s[36:37], -1
; GFX9-NEXT: buffer_load_dword v40, off, s[0:3], s33 ; 4-byte Folded Reload
+; GFX9-NEXT: buffer_load_dword v41, off, s[0:3], s33 offset:4 ; 4-byte Folded Reload
; GFX9-NEXT: s_mov_b64 exec, s[36:37]
; GFX9-NEXT: s_addk_i32 s32, 0xfc00
; GFX9-NEXT: s_mov_b32 s33, s34
@@ -14014,20 +15071,20 @@
; GFX10-NEXT: s_mov_b32 s33, s32
; GFX10-NEXT: s_or_saveexec_b32 s35, -1
; GFX10-NEXT: buffer_store_dword v40, off, s[0:3], s33 ; 4-byte Folded Spill
+; GFX10-NEXT: buffer_store_dword v41, off, s[0:3], s33 offset:4 ; 4-byte Folded Spill
; GFX10-NEXT: s_waitcnt_depctr 0xffe3
; GFX10-NEXT: s_mov_b32 exec_lo, s35
; GFX10-NEXT: v_mov_b32_e32 v0, 0x41000000
; GFX10-NEXT: v_mov_b32_e32 v1, 0x41100000
; GFX10-NEXT: v_mov_b32_e32 v2, 0x41200000
; GFX10-NEXT: s_addk_i32 s32, 0x200
-; GFX10-NEXT: v_writelane_b32 v40, s34, 2
+; GFX10-NEXT: v_mov_b32_e32 v3, 0x41600000
; GFX10-NEXT: buffer_store_dword v0, off, s[0:3], s32
; GFX10-NEXT: buffer_store_dword v1, off, s[0:3], s32 offset:4
; GFX10-NEXT: buffer_store_dword v2, off, s[0:3], s32 offset:8
; GFX10-NEXT: v_mov_b32_e32 v0, 0x41300000
; GFX10-NEXT: v_mov_b32_e32 v1, 0x41400000
; GFX10-NEXT: v_mov_b32_e32 v2, 0x41500000
-; GFX10-NEXT: v_mov_b32_e32 v3, 0x41600000
; GFX10-NEXT: v_mov_b32_e32 v4, 0x41700000
; GFX10-NEXT: v_writelane_b32 v40, s30, 0
; GFX10-NEXT: buffer_store_dword v0, off, s[0:3], s32 offset:12
@@ -14067,6 +15124,7 @@
; GFX10-NEXT: v_mov_b32_e32 v29, 0x40a00000
; GFX10-NEXT: v_mov_b32_e32 v30, 0x40c00000
; GFX10-NEXT: v_mov_b32_e32 v31, 0x40e00000
+; GFX10-NEXT: v_writelane_b32 v41, s34, 0
; GFX10-NEXT: v_writelane_b32 v40, s31, 1
; GFX10-NEXT: s_getpc_b64 s[34:35]
; GFX10-NEXT: s_add_u32 s34, s34, external_void_func_8xv5f32@rel32@lo+4
@@ -14074,9 +15132,11 @@
; GFX10-NEXT: s_swappc_b64 s[30:31], s[34:35]
; GFX10-NEXT: v_readlane_b32 s31, v40, 1
; GFX10-NEXT: v_readlane_b32 s30, v40, 0
-; GFX10-NEXT: v_readlane_b32 s34, v40, 2
+; GFX10-NEXT: v_readlane_b32 s34, v41, 0
; GFX10-NEXT: s_or_saveexec_b32 s35, -1
-; GFX10-NEXT: buffer_load_dword v40, off, s[0:3], s33 ; 4-byte Folded Reload
+; GFX10-NEXT: s_clause 0x1
+; GFX10-NEXT: buffer_load_dword v40, off, s[0:3], s33
+; GFX10-NEXT: buffer_load_dword v41, off, s[0:3], s33 offset:4
; GFX10-NEXT: s_waitcnt_depctr 0xffe3
; GFX10-NEXT: s_mov_b32 exec_lo, s35
; GFX10-NEXT: s_addk_i32 s32, 0xfe00
@@ -14090,9 +15150,10 @@
; GFX11-NEXT: s_mov_b32 s0, s33
; GFX11-NEXT: s_mov_b32 s33, s32
; GFX11-NEXT: s_or_saveexec_b32 s1, -1
-; GFX11-NEXT: scratch_store_b32 off, v40, s33 ; 4-byte Folded Spill
+; GFX11-NEXT: s_clause 0x1
+; GFX11-NEXT: scratch_store_b32 off, v40, s33
+; GFX11-NEXT: scratch_store_b32 off, v41, s33 offset:4
; GFX11-NEXT: s_mov_b32 exec_lo, s1
-; GFX11-NEXT: v_writelane_b32 v40, s0, 2
; GFX11-NEXT: v_mov_b32_e32 v0, 0x41000000
; GFX11-NEXT: v_mov_b32_e32 v1, 0x41100000
; GFX11-NEXT: v_mov_b32_e32 v2, 0x41200000
@@ -14102,8 +15163,9 @@
; GFX11-NEXT: v_mov_b32_e32 v6, 0x41600000
; GFX11-NEXT: v_mov_b32_e32 v7, 0x41700000
; GFX11-NEXT: s_add_i32 s32, s32, 16
-; GFX11-NEXT: v_writelane_b32 v40, s30, 0
+; GFX11-NEXT: v_writelane_b32 v41, s0, 0
; GFX11-NEXT: s_add_i32 s0, s32, 16
+; GFX11-NEXT: v_writelane_b32 v40, s30, 0
; GFX11-NEXT: scratch_store_b128 off, v[0:3], s32
; GFX11-NEXT: scratch_store_b128 off, v[4:7], s0
; GFX11-NEXT: v_mov_b32_e32 v6, 1.0
@@ -14132,9 +15194,11 @@
; GFX11-NEXT: s_swappc_b64 s[30:31], s[0:1]
; GFX11-NEXT: v_readlane_b32 s31, v40, 1
; GFX11-NEXT: v_readlane_b32 s30, v40, 0
-; GFX11-NEXT: v_readlane_b32 s0, v40, 2
+; GFX11-NEXT: v_readlane_b32 s0, v41, 0
; GFX11-NEXT: s_or_saveexec_b32 s1, -1
-; GFX11-NEXT: scratch_load_b32 v40, off, s33 ; 4-byte Folded Reload
+; GFX11-NEXT: s_clause 0x1
+; GFX11-NEXT: scratch_load_b32 v40, off, s33
+; GFX11-NEXT: scratch_load_b32 v41, off, s33 offset:4
; GFX11-NEXT: s_mov_b32 exec_lo, s1
; GFX11-NEXT: s_add_i32 s32, s32, -16
; GFX11-NEXT: s_mov_b32 s33, s0
@@ -14148,9 +15212,9 @@
; GFX10-SCRATCH-NEXT: s_mov_b32 s33, s32
; GFX10-SCRATCH-NEXT: s_or_saveexec_b32 s1, -1
; GFX10-SCRATCH-NEXT: scratch_store_dword off, v40, s33 ; 4-byte Folded Spill
+; GFX10-SCRATCH-NEXT: scratch_store_dword off, v41, s33 offset:4 ; 4-byte Folded Spill
; GFX10-SCRATCH-NEXT: s_waitcnt_depctr 0xffe3
; GFX10-SCRATCH-NEXT: s_mov_b32 exec_lo, s1
-; GFX10-SCRATCH-NEXT: v_writelane_b32 v40, s0, 2
; GFX10-SCRATCH-NEXT: v_mov_b32_e32 v0, 0x41000000
; GFX10-SCRATCH-NEXT: v_mov_b32_e32 v1, 0x41100000
; GFX10-SCRATCH-NEXT: v_mov_b32_e32 v2, 0x41200000
@@ -14160,8 +15224,9 @@
; GFX10-SCRATCH-NEXT: v_mov_b32_e32 v6, 0x41600000
; GFX10-SCRATCH-NEXT: v_mov_b32_e32 v7, 0x41700000
; GFX10-SCRATCH-NEXT: s_add_i32 s32, s32, 16
-; GFX10-SCRATCH-NEXT: v_writelane_b32 v40, s30, 0
+; GFX10-SCRATCH-NEXT: v_writelane_b32 v41, s0, 0
; GFX10-SCRATCH-NEXT: s_add_i32 s0, s32, 16
+; GFX10-SCRATCH-NEXT: v_writelane_b32 v40, s30, 0
; GFX10-SCRATCH-NEXT: scratch_store_dwordx4 off, v[0:3], s32
; GFX10-SCRATCH-NEXT: scratch_store_dwordx4 off, v[4:7], s0
; GFX10-SCRATCH-NEXT: v_mov_b32_e32 v0, 0
@@ -14203,9 +15268,11 @@
; GFX10-SCRATCH-NEXT: s_swappc_b64 s[30:31], s[0:1]
; GFX10-SCRATCH-NEXT: v_readlane_b32 s31, v40, 1
; GFX10-SCRATCH-NEXT: v_readlane_b32 s30, v40, 0
-; GFX10-SCRATCH-NEXT: v_readlane_b32 s0, v40, 2
+; GFX10-SCRATCH-NEXT: v_readlane_b32 s0, v41, 0
; GFX10-SCRATCH-NEXT: s_or_saveexec_b32 s1, -1
-; GFX10-SCRATCH-NEXT: scratch_load_dword v40, off, s33 ; 4-byte Folded Reload
+; GFX10-SCRATCH-NEXT: s_clause 0x1
+; GFX10-SCRATCH-NEXT: scratch_load_dword v40, off, s33
+; GFX10-SCRATCH-NEXT: scratch_load_dword v41, off, s33 offset:4
; GFX10-SCRATCH-NEXT: s_waitcnt_depctr 0xffe3
; GFX10-SCRATCH-NEXT: s_mov_b32 exec_lo, s1
; GFX10-SCRATCH-NEXT: s_add_i32 s32, s32, -16
diff --git a/llvm/test/CodeGen/AMDGPU/gfx-callable-preserved-registers.ll b/llvm/test/CodeGen/AMDGPU/gfx-callable-preserved-registers.ll
--- a/llvm/test/CodeGen/AMDGPU/gfx-callable-preserved-registers.ll
+++ b/llvm/test/CodeGen/AMDGPU/gfx-callable-preserved-registers.ll
@@ -13,12 +13,13 @@
; GFX9-NEXT: s_mov_b32 s33, s32
; GFX9-NEXT: s_or_saveexec_b64 s[36:37], -1
; GFX9-NEXT: buffer_store_dword v40, off, s[0:3], s33 ; 4-byte Folded Spill
+; GFX9-NEXT: buffer_store_dword v41, off, s[0:3], s33 offset:4 ; 4-byte Folded Spill
; GFX9-NEXT: s_mov_b64 exec, s[36:37]
-; GFX9-NEXT: v_writelane_b32 v40, s34, 4
; GFX9-NEXT: v_writelane_b32 v40, s4, 0
; GFX9-NEXT: v_writelane_b32 v40, s5, 1
; GFX9-NEXT: s_addk_i32 s32, 0x400
; GFX9-NEXT: v_writelane_b32 v40, s30, 2
+; GFX9-NEXT: v_writelane_b32 v41, s34, 0
; GFX9-NEXT: v_writelane_b32 v40, s31, 3
; GFX9-NEXT: s_getpc_b64 s[4:5]
; GFX9-NEXT: s_add_u32 s4, s4, external_void_func_void@rel32@lo+4
@@ -31,9 +32,10 @@
; GFX9-NEXT: v_readlane_b32 s30, v40, 2
; GFX9-NEXT: v_readlane_b32 s5, v40, 1
; GFX9-NEXT: v_readlane_b32 s4, v40, 0
-; GFX9-NEXT: v_readlane_b32 s34, v40, 4
+; GFX9-NEXT: v_readlane_b32 s34, v41, 0
; GFX9-NEXT: s_or_saveexec_b64 s[36:37], -1
; GFX9-NEXT: buffer_load_dword v40, off, s[0:3], s33 ; 4-byte Folded Reload
+; GFX9-NEXT: buffer_load_dword v41, off, s[0:3], s33 offset:4 ; 4-byte Folded Reload
; GFX9-NEXT: s_mov_b64 exec, s[36:37]
; GFX9-NEXT: s_addk_i32 s32, 0xfc00
; GFX9-NEXT: s_mov_b32 s33, s34
@@ -47,11 +49,12 @@
; GFX10-NEXT: s_mov_b32 s33, s32
; GFX10-NEXT: s_or_saveexec_b32 s35, -1
; GFX10-NEXT: buffer_store_dword v40, off, s[0:3], s33 ; 4-byte Folded Spill
+; GFX10-NEXT: buffer_store_dword v41, off, s[0:3], s33 offset:4 ; 4-byte Folded Spill
; GFX10-NEXT: s_waitcnt_depctr 0xffe3
; GFX10-NEXT: s_mov_b32 exec_lo, s35
-; GFX10-NEXT: v_writelane_b32 v40, s34, 4
-; GFX10-NEXT: s_addk_i32 s32, 0x200
; GFX10-NEXT: v_writelane_b32 v40, s4, 0
+; GFX10-NEXT: s_addk_i32 s32, 0x200
+; GFX10-NEXT: v_writelane_b32 v41, s34, 0
; GFX10-NEXT: v_writelane_b32 v40, s5, 1
; GFX10-NEXT: s_getpc_b64 s[4:5]
; GFX10-NEXT: s_add_u32 s4, s4, external_void_func_void@rel32@lo+4
@@ -66,9 +69,11 @@
; GFX10-NEXT: v_readlane_b32 s30, v40, 2
; GFX10-NEXT: v_readlane_b32 s5, v40, 1
; GFX10-NEXT: v_readlane_b32 s4, v40, 0
-; GFX10-NEXT: v_readlane_b32 s34, v40, 4
+; GFX10-NEXT: v_readlane_b32 s34, v41, 0
; GFX10-NEXT: s_or_saveexec_b32 s35, -1
-; GFX10-NEXT: buffer_load_dword v40, off, s[0:3], s33 ; 4-byte Folded Reload
+; GFX10-NEXT: s_clause 0x1
+; GFX10-NEXT: buffer_load_dword v40, off, s[0:3], s33
+; GFX10-NEXT: buffer_load_dword v41, off, s[0:3], s33 offset:4
; GFX10-NEXT: s_waitcnt_depctr 0xffe3
; GFX10-NEXT: s_mov_b32 exec_lo, s35
; GFX10-NEXT: s_addk_i32 s32, 0xfe00
@@ -82,11 +87,13 @@
; GFX11-NEXT: s_mov_b32 s0, s33
; GFX11-NEXT: s_mov_b32 s33, s32
; GFX11-NEXT: s_or_saveexec_b32 s1, -1
-; GFX11-NEXT: scratch_store_b32 off, v40, s33 ; 4-byte Folded Spill
+; GFX11-NEXT: s_clause 0x1
+; GFX11-NEXT: scratch_store_b32 off, v40, s33
+; GFX11-NEXT: scratch_store_b32 off, v41, s33 offset:4
; GFX11-NEXT: s_mov_b32 exec_lo, s1
-; GFX11-NEXT: v_writelane_b32 v40, s0, 4
-; GFX11-NEXT: s_add_i32 s32, s32, 16
; GFX11-NEXT: v_writelane_b32 v40, s4, 0
+; GFX11-NEXT: s_add_i32 s32, s32, 16
+; GFX11-NEXT: v_writelane_b32 v41, s0, 0
; GFX11-NEXT: v_writelane_b32 v40, s5, 1
; GFX11-NEXT: s_getpc_b64 s[4:5]
; GFX11-NEXT: s_add_u32 s4, s4, external_void_func_void@rel32@lo+4
@@ -102,9 +109,11 @@
; GFX11-NEXT: v_readlane_b32 s30, v40, 2
; GFX11-NEXT: v_readlane_b32 s5, v40, 1
; GFX11-NEXT: v_readlane_b32 s4, v40, 0
-; GFX11-NEXT: v_readlane_b32 s0, v40, 4
+; GFX11-NEXT: v_readlane_b32 s0, v41, 0
; GFX11-NEXT: s_or_saveexec_b32 s1, -1
-; GFX11-NEXT: scratch_load_b32 v40, off, s33 ; 4-byte Folded Reload
+; GFX11-NEXT: s_clause 0x1
+; GFX11-NEXT: scratch_load_b32 v40, off, s33
+; GFX11-NEXT: scratch_load_b32 v41, off, s33 offset:4
; GFX11-NEXT: s_mov_b32 exec_lo, s1
; GFX11-NEXT: s_add_i32 s32, s32, -16
; GFX11-NEXT: s_mov_b32 s33, s0
@@ -174,9 +183,9 @@
; GFX11-LABEL: void_func_void_clobber_s28_s29:
; GFX11: ; %bb.0:
; GFX11-NEXT: s_waitcnt vmcnt(0) expcnt(0) lgkmcnt(0)
-; GFX11-NEXT: s_xor_saveexec_b32 s0, -1
+; GFX11-NEXT: s_xor_saveexec_b32 s1, -1
; GFX11-NEXT: scratch_store_b32 off, v0, s32 ; 4-byte Folded Spill
-; GFX11-NEXT: s_mov_b32 exec_lo, s0
+; GFX11-NEXT: s_mov_b32 exec_lo, s1
; GFX11-NEXT: v_writelane_b32 v0, s28, 0
; GFX11-NEXT: v_writelane_b32 v0, s29, 1
; GFX11-NEXT: v_writelane_b32 v0, s30, 2
@@ -192,9 +201,9 @@
; GFX11-NEXT: v_readlane_b32 s30, v0, 2
; GFX11-NEXT: v_readlane_b32 s29, v0, 1
; GFX11-NEXT: v_readlane_b32 s28, v0, 0
-; GFX11-NEXT: s_xor_saveexec_b32 s0, -1
+; GFX11-NEXT: s_xor_saveexec_b32 s1, -1
; GFX11-NEXT: scratch_load_b32 v0, off, s32 ; 4-byte Folded Reload
-; GFX11-NEXT: s_mov_b32 exec_lo, s0
+; GFX11-NEXT: s_mov_b32 exec_lo, s1
; GFX11-NEXT: s_waitcnt vmcnt(0)
; GFX11-NEXT: s_setpc_b64 s[30:31]
call void asm sideeffect "; clobber", "~{s[30:31]}"() #0
@@ -210,11 +219,12 @@
; GFX9-NEXT: s_mov_b32 s33, s32
; GFX9-NEXT: s_or_saveexec_b64 s[36:37], -1
; GFX9-NEXT: buffer_store_dword v40, off, s[0:3], s33 ; 4-byte Folded Spill
+; GFX9-NEXT: buffer_store_dword v41, off, s[0:3], s33 offset:4 ; 4-byte Folded Spill
; GFX9-NEXT: s_mov_b64 exec, s[36:37]
-; GFX9-NEXT: v_writelane_b32 v40, s34, 3
; GFX9-NEXT: v_writelane_b32 v40, s4, 0
; GFX9-NEXT: s_addk_i32 s32, 0x400
; GFX9-NEXT: v_writelane_b32 v40, s30, 1
+; GFX9-NEXT: v_writelane_b32 v41, s34, 0
; GFX9-NEXT: v_writelane_b32 v40, s31, 2
; GFX9-NEXT: ;;#ASMSTART
; GFX9-NEXT: ; def s31
@@ -231,9 +241,10 @@
; GFX9-NEXT: v_readlane_b32 s31, v40, 2
; GFX9-NEXT: v_readlane_b32 s30, v40, 1
; GFX9-NEXT: v_readlane_b32 s4, v40, 0
-; GFX9-NEXT: v_readlane_b32 s34, v40, 3
+; GFX9-NEXT: v_readlane_b32 s34, v41, 0
; GFX9-NEXT: s_or_saveexec_b64 s[36:37], -1
; GFX9-NEXT: buffer_load_dword v40, off, s[0:3], s33 ; 4-byte Folded Reload
+; GFX9-NEXT: buffer_load_dword v41, off, s[0:3], s33 offset:4 ; 4-byte Folded Reload
; GFX9-NEXT: s_mov_b64 exec, s[36:37]
; GFX9-NEXT: s_addk_i32 s32, 0xfc00
; GFX9-NEXT: s_mov_b32 s33, s34
@@ -247,14 +258,15 @@
; GFX10-NEXT: s_mov_b32 s33, s32
; GFX10-NEXT: s_or_saveexec_b32 s35, -1
; GFX10-NEXT: buffer_store_dword v40, off, s[0:3], s33 ; 4-byte Folded Spill
+; GFX10-NEXT: buffer_store_dword v41, off, s[0:3], s33 offset:4 ; 4-byte Folded Spill
; GFX10-NEXT: s_waitcnt_depctr 0xffe3
; GFX10-NEXT: s_mov_b32 exec_lo, s35
-; GFX10-NEXT: v_writelane_b32 v40, s34, 3
+; GFX10-NEXT: v_writelane_b32 v40, s4, 0
; GFX10-NEXT: s_addk_i32 s32, 0x200
+; GFX10-NEXT: v_writelane_b32 v41, s34, 0
; GFX10-NEXT: s_getpc_b64 s[34:35]
; GFX10-NEXT: s_add_u32 s34, s34, external_void_func_void@rel32@lo+4
; GFX10-NEXT: s_addc_u32 s35, s35, external_void_func_void@rel32@hi+12
-; GFX10-NEXT: v_writelane_b32 v40, s4, 0
; GFX10-NEXT: v_writelane_b32 v40, s30, 1
; GFX10-NEXT: v_writelane_b32 v40, s31, 2
; GFX10-NEXT: ;;#ASMSTART
@@ -269,9 +281,11 @@
; GFX10-NEXT: v_readlane_b32 s31, v40, 2
; GFX10-NEXT: v_readlane_b32 s30, v40, 1
; GFX10-NEXT: v_readlane_b32 s4, v40, 0
-; GFX10-NEXT: v_readlane_b32 s34, v40, 3
+; GFX10-NEXT: v_readlane_b32 s34, v41, 0
; GFX10-NEXT: s_or_saveexec_b32 s35, -1
-; GFX10-NEXT: buffer_load_dword v40, off, s[0:3], s33 ; 4-byte Folded Reload
+; GFX10-NEXT: s_clause 0x1
+; GFX10-NEXT: buffer_load_dword v40, off, s[0:3], s33
+; GFX10-NEXT: buffer_load_dword v41, off, s[0:3], s33 offset:4
; GFX10-NEXT: s_waitcnt_depctr 0xffe3
; GFX10-NEXT: s_mov_b32 exec_lo, s35
; GFX10-NEXT: s_addk_i32 s32, 0xfe00
@@ -285,14 +299,16 @@
; GFX11-NEXT: s_mov_b32 s0, s33
; GFX11-NEXT: s_mov_b32 s33, s32
; GFX11-NEXT: s_or_saveexec_b32 s1, -1
-; GFX11-NEXT: scratch_store_b32 off, v40, s33 ; 4-byte Folded Spill
+; GFX11-NEXT: s_clause 0x1
+; GFX11-NEXT: scratch_store_b32 off, v40, s33
+; GFX11-NEXT: scratch_store_b32 off, v41, s33 offset:4
; GFX11-NEXT: s_mov_b32 exec_lo, s1
-; GFX11-NEXT: v_writelane_b32 v40, s0, 3
+; GFX11-NEXT: v_writelane_b32 v40, s4, 0
; GFX11-NEXT: s_add_i32 s32, s32, 16
+; GFX11-NEXT: v_writelane_b32 v41, s0, 0
; GFX11-NEXT: s_getpc_b64 s[0:1]
; GFX11-NEXT: s_add_u32 s0, s0, external_void_func_void@rel32@lo+4
; GFX11-NEXT: s_addc_u32 s1, s1, external_void_func_void@rel32@hi+12
-; GFX11-NEXT: v_writelane_b32 v40, s4, 0
; GFX11-NEXT: v_writelane_b32 v40, s30, 1
; GFX11-NEXT: v_writelane_b32 v40, s31, 2
; GFX11-NEXT: ;;#ASMSTART
@@ -307,9 +323,11 @@
; GFX11-NEXT: v_readlane_b32 s31, v40, 2
; GFX11-NEXT: v_readlane_b32 s30, v40, 1
; GFX11-NEXT: v_readlane_b32 s4, v40, 0
-; GFX11-NEXT: v_readlane_b32 s0, v40, 3
+; GFX11-NEXT: v_readlane_b32 s0, v41, 0
; GFX11-NEXT: s_or_saveexec_b32 s1, -1
-; GFX11-NEXT: scratch_load_b32 v40, off, s33 ; 4-byte Folded Reload
+; GFX11-NEXT: s_clause 0x1
+; GFX11-NEXT: scratch_load_b32 v40, off, s33
+; GFX11-NEXT: scratch_load_b32 v41, off, s33 offset:4
; GFX11-NEXT: s_mov_b32 exec_lo, s1
; GFX11-NEXT: s_add_i32 s32, s32, -16
; GFX11-NEXT: s_mov_b32 s33, s0
@@ -329,10 +347,11 @@
; GFX9-NEXT: s_mov_b32 s33, s32
; GFX9-NEXT: s_or_saveexec_b64 s[36:37], -1
; GFX9-NEXT: buffer_store_dword v40, off, s[0:3], s33 offset:4 ; 4-byte Folded Spill
+; GFX9-NEXT: buffer_store_dword v42, off, s[0:3], s33 offset:8 ; 4-byte Folded Spill
; GFX9-NEXT: s_mov_b64 exec, s[36:37]
-; GFX9-NEXT: v_writelane_b32 v40, s34, 2
; GFX9-NEXT: s_addk_i32 s32, 0x400
; GFX9-NEXT: v_writelane_b32 v40, s30, 0
+; GFX9-NEXT: v_writelane_b32 v42, s34, 0
; GFX9-NEXT: buffer_store_dword v41, off, s[0:3], s33 ; 4-byte Folded Spill
; GFX9-NEXT: v_writelane_b32 v40, s31, 1
; GFX9-NEXT: ;;#ASMSTART
@@ -350,9 +369,10 @@
; GFX9-NEXT: buffer_load_dword v41, off, s[0:3], s33 ; 4-byte Folded Reload
; GFX9-NEXT: v_readlane_b32 s31, v40, 1
; GFX9-NEXT: v_readlane_b32 s30, v40, 0
-; GFX9-NEXT: v_readlane_b32 s34, v40, 2
+; GFX9-NEXT: v_readlane_b32 s34, v42, 0
; GFX9-NEXT: s_or_saveexec_b64 s[36:37], -1
; GFX9-NEXT: buffer_load_dword v40, off, s[0:3], s33 offset:4 ; 4-byte Folded Reload
+; GFX9-NEXT: buffer_load_dword v42, off, s[0:3], s33 offset:8 ; 4-byte Folded Reload
; GFX9-NEXT: s_mov_b64 exec, s[36:37]
; GFX9-NEXT: s_addk_i32 s32, 0xfc00
; GFX9-NEXT: s_mov_b32 s33, s34
@@ -366,20 +386,21 @@
; GFX10-NEXT: s_mov_b32 s33, s32
; GFX10-NEXT: s_or_saveexec_b32 s35, -1
; GFX10-NEXT: buffer_store_dword v40, off, s[0:3], s33 offset:4 ; 4-byte Folded Spill
+; GFX10-NEXT: buffer_store_dword v42, off, s[0:3], s33 offset:8 ; 4-byte Folded Spill
; GFX10-NEXT: s_waitcnt_depctr 0xffe3
; GFX10-NEXT: s_mov_b32 exec_lo, s35
-; GFX10-NEXT: v_writelane_b32 v40, s34, 2
+; GFX10-NEXT: v_writelane_b32 v40, s30, 0
; GFX10-NEXT: s_addk_i32 s32, 0x200
+; GFX10-NEXT: v_writelane_b32 v42, s34, 0
; GFX10-NEXT: buffer_store_dword v41, off, s[0:3], s33 ; 4-byte Folded Spill
; GFX10-NEXT: ;;#ASMSTART
; GFX10-NEXT: ; def v31
; GFX10-NEXT: ;;#ASMEND
+; GFX10-NEXT: v_writelane_b32 v40, s31, 1
; GFX10-NEXT: v_mov_b32_e32 v41, v31
-; GFX10-NEXT: v_writelane_b32 v40, s30, 0
; GFX10-NEXT: s_getpc_b64 s[34:35]
; GFX10-NEXT: s_add_u32 s34, s34, external_void_func_void@rel32@lo+4
; GFX10-NEXT: s_addc_u32 s35, s35, external_void_func_void@rel32@hi+12
-; GFX10-NEXT: v_writelane_b32 v40, s31, 1
; GFX10-NEXT: s_swappc_b64 s[30:31], s[34:35]
; GFX10-NEXT: v_mov_b32_e32 v31, v41
; GFX10-NEXT: ;;#ASMSTART
@@ -388,9 +409,11 @@
; GFX10-NEXT: buffer_load_dword v41, off, s[0:3], s33 ; 4-byte Folded Reload
; GFX10-NEXT: v_readlane_b32 s31, v40, 1
; GFX10-NEXT: v_readlane_b32 s30, v40, 0
-; GFX10-NEXT: v_readlane_b32 s34, v40, 2
+; GFX10-NEXT: v_readlane_b32 s34, v42, 0
; GFX10-NEXT: s_or_saveexec_b32 s35, -1
-; GFX10-NEXT: buffer_load_dword v40, off, s[0:3], s33 offset:4 ; 4-byte Folded Reload
+; GFX10-NEXT: s_clause 0x1
+; GFX10-NEXT: buffer_load_dword v40, off, s[0:3], s33 offset:4
+; GFX10-NEXT: buffer_load_dword v42, off, s[0:3], s33 offset:8
; GFX10-NEXT: s_waitcnt_depctr 0xffe3
; GFX10-NEXT: s_mov_b32 exec_lo, s35
; GFX10-NEXT: s_addk_i32 s32, 0xfe00
@@ -404,20 +427,23 @@
; GFX11-NEXT: s_mov_b32 s0, s33
; GFX11-NEXT: s_mov_b32 s33, s32
; GFX11-NEXT: s_or_saveexec_b32 s1, -1
-; GFX11-NEXT: scratch_store_b32 off, v40, s33 offset:4 ; 4-byte Folded Spill
+; GFX11-NEXT: s_clause 0x1
+; GFX11-NEXT: scratch_store_b32 off, v40, s33 offset:4
+; GFX11-NEXT: scratch_store_b32 off, v42, s33 offset:8
; GFX11-NEXT: s_mov_b32 exec_lo, s1
-; GFX11-NEXT: v_writelane_b32 v40, s0, 2
+; GFX11-NEXT: v_writelane_b32 v40, s30, 0
; GFX11-NEXT: s_add_i32 s32, s32, 16
+; GFX11-NEXT: v_writelane_b32 v42, s0, 0
; GFX11-NEXT: scratch_store_b32 off, v41, s33 ; 4-byte Folded Spill
; GFX11-NEXT: ;;#ASMSTART
; GFX11-NEXT: ; def v31
; GFX11-NEXT: ;;#ASMEND
+; GFX11-NEXT: v_writelane_b32 v40, s31, 1
; GFX11-NEXT: v_mov_b32_e32 v41, v31
-; GFX11-NEXT: v_writelane_b32 v40, s30, 0
; GFX11-NEXT: s_getpc_b64 s[0:1]
; GFX11-NEXT: s_add_u32 s0, s0, external_void_func_void@rel32@lo+4
; GFX11-NEXT: s_addc_u32 s1, s1, external_void_func_void@rel32@hi+12
-; GFX11-NEXT: v_writelane_b32 v40, s31, 1
+; GFX11-NEXT: s_delay_alu instid0(SALU_CYCLE_1)
; GFX11-NEXT: s_swappc_b64 s[30:31], s[0:1]
; GFX11-NEXT: v_mov_b32_e32 v31, v41
; GFX11-NEXT: ;;#ASMSTART
@@ -426,9 +452,11 @@
; GFX11-NEXT: scratch_load_b32 v41, off, s33 ; 4-byte Folded Reload
; GFX11-NEXT: v_readlane_b32 s31, v40, 1
; GFX11-NEXT: v_readlane_b32 s30, v40, 0
-; GFX11-NEXT: v_readlane_b32 s0, v40, 2
+; GFX11-NEXT: v_readlane_b32 s0, v42, 0
; GFX11-NEXT: s_or_saveexec_b32 s1, -1
-; GFX11-NEXT: scratch_load_b32 v40, off, s33 offset:4 ; 4-byte Folded Reload
+; GFX11-NEXT: s_clause 0x1
+; GFX11-NEXT: scratch_load_b32 v40, off, s33 offset:4
+; GFX11-NEXT: scratch_load_b32 v42, off, s33 offset:8
; GFX11-NEXT: s_mov_b32 exec_lo, s1
; GFX11-NEXT: s_add_i32 s32, s32, -16
; GFX11-NEXT: s_mov_b32 s33, s0
@@ -449,11 +477,12 @@
; GFX9-NEXT: s_mov_b32 s33, s32
; GFX9-NEXT: s_or_saveexec_b64 s[36:37], -1
; GFX9-NEXT: buffer_store_dword v40, off, s[0:3], s33 ; 4-byte Folded Spill
+; GFX9-NEXT: buffer_store_dword v41, off, s[0:3], s33 offset:4 ; 4-byte Folded Spill
; GFX9-NEXT: s_mov_b64 exec, s[36:37]
-; GFX9-NEXT: v_writelane_b32 v40, s34, 3
; GFX9-NEXT: v_writelane_b32 v40, s4, 0
; GFX9-NEXT: s_addk_i32 s32, 0x400
; GFX9-NEXT: v_writelane_b32 v40, s30, 1
+; GFX9-NEXT: v_writelane_b32 v41, s34, 0
; GFX9-NEXT: v_writelane_b32 v40, s31, 2
; GFX9-NEXT: ;;#ASMSTART
; GFX9-NEXT: ; def s33
@@ -470,9 +499,10 @@
; GFX9-NEXT: v_readlane_b32 s31, v40, 2
; GFX9-NEXT: v_readlane_b32 s30, v40, 1
; GFX9-NEXT: v_readlane_b32 s4, v40, 0
-; GFX9-NEXT: v_readlane_b32 s34, v40, 3
+; GFX9-NEXT: v_readlane_b32 s34, v41, 0
; GFX9-NEXT: s_or_saveexec_b64 s[36:37], -1
; GFX9-NEXT: buffer_load_dword v40, off, s[0:3], s33 ; 4-byte Folded Reload
+; GFX9-NEXT: buffer_load_dword v41, off, s[0:3], s33 offset:4 ; 4-byte Folded Reload
; GFX9-NEXT: s_mov_b64 exec, s[36:37]
; GFX9-NEXT: s_addk_i32 s32, 0xfc00
; GFX9-NEXT: s_mov_b32 s33, s34
@@ -486,19 +516,20 @@
; GFX10-NEXT: s_mov_b32 s33, s32
; GFX10-NEXT: s_or_saveexec_b32 s35, -1
; GFX10-NEXT: buffer_store_dword v40, off, s[0:3], s33 ; 4-byte Folded Spill
+; GFX10-NEXT: buffer_store_dword v41, off, s[0:3], s33 offset:4 ; 4-byte Folded Spill
; GFX10-NEXT: s_waitcnt_depctr 0xffe3
; GFX10-NEXT: s_mov_b32 exec_lo, s35
-; GFX10-NEXT: v_writelane_b32 v40, s34, 3
+; GFX10-NEXT: v_writelane_b32 v40, s4, 0
; GFX10-NEXT: s_addk_i32 s32, 0x200
+; GFX10-NEXT: v_writelane_b32 v41, s34, 0
; GFX10-NEXT: s_getpc_b64 s[34:35]
; GFX10-NEXT: s_add_u32 s34, s34, external_void_func_void@rel32@lo+4
; GFX10-NEXT: s_addc_u32 s35, s35, external_void_func_void@rel32@hi+12
; GFX10-NEXT: ;;#ASMSTART
; GFX10-NEXT: ; def s33
; GFX10-NEXT: ;;#ASMEND
-; GFX10-NEXT: v_writelane_b32 v40, s4, 0
-; GFX10-NEXT: s_mov_b32 s4, s33
; GFX10-NEXT: v_writelane_b32 v40, s30, 1
+; GFX10-NEXT: s_mov_b32 s4, s33
; GFX10-NEXT: v_writelane_b32 v40, s31, 2
; GFX10-NEXT: s_swappc_b64 s[30:31], s[34:35]
; GFX10-NEXT: s_mov_b32 s33, s4
@@ -508,9 +539,11 @@
; GFX10-NEXT: v_readlane_b32 s31, v40, 2
; GFX10-NEXT: v_readlane_b32 s30, v40, 1
; GFX10-NEXT: v_readlane_b32 s4, v40, 0
-; GFX10-NEXT: v_readlane_b32 s34, v40, 3
+; GFX10-NEXT: v_readlane_b32 s34, v41, 0
; GFX10-NEXT: s_or_saveexec_b32 s35, -1
-; GFX10-NEXT: buffer_load_dword v40, off, s[0:3], s33 ; 4-byte Folded Reload
+; GFX10-NEXT: s_clause 0x1
+; GFX10-NEXT: buffer_load_dword v40, off, s[0:3], s33
+; GFX10-NEXT: buffer_load_dword v41, off, s[0:3], s33 offset:4
; GFX10-NEXT: s_waitcnt_depctr 0xffe3
; GFX10-NEXT: s_mov_b32 exec_lo, s35
; GFX10-NEXT: s_addk_i32 s32, 0xfe00
@@ -524,19 +557,21 @@
; GFX11-NEXT: s_mov_b32 s0, s33
; GFX11-NEXT: s_mov_b32 s33, s32
; GFX11-NEXT: s_or_saveexec_b32 s1, -1
-; GFX11-NEXT: scratch_store_b32 off, v40, s33 ; 4-byte Folded Spill
+; GFX11-NEXT: s_clause 0x1
+; GFX11-NEXT: scratch_store_b32 off, v40, s33
+; GFX11-NEXT: scratch_store_b32 off, v41, s33 offset:4
; GFX11-NEXT: s_mov_b32 exec_lo, s1
-; GFX11-NEXT: v_writelane_b32 v40, s0, 3
+; GFX11-NEXT: v_writelane_b32 v40, s4, 0
; GFX11-NEXT: s_add_i32 s32, s32, 16
+; GFX11-NEXT: v_writelane_b32 v41, s0, 0
; GFX11-NEXT: s_getpc_b64 s[0:1]
; GFX11-NEXT: s_add_u32 s0, s0, external_void_func_void@rel32@lo+4
; GFX11-NEXT: s_addc_u32 s1, s1, external_void_func_void@rel32@hi+12
; GFX11-NEXT: ;;#ASMSTART
; GFX11-NEXT: ; def s33
; GFX11-NEXT: ;;#ASMEND
-; GFX11-NEXT: v_writelane_b32 v40, s4, 0
-; GFX11-NEXT: s_mov_b32 s4, s33
; GFX11-NEXT: v_writelane_b32 v40, s30, 1
+; GFX11-NEXT: s_mov_b32 s4, s33
; GFX11-NEXT: v_writelane_b32 v40, s31, 2
; GFX11-NEXT: s_swappc_b64 s[30:31], s[0:1]
; GFX11-NEXT: s_mov_b32 s33, s4
@@ -547,9 +582,11 @@
; GFX11-NEXT: v_readlane_b32 s31, v40, 2
; GFX11-NEXT: v_readlane_b32 s30, v40, 1
; GFX11-NEXT: v_readlane_b32 s4, v40, 0
-; GFX11-NEXT: v_readlane_b32 s0, v40, 3
+; GFX11-NEXT: v_readlane_b32 s0, v41, 0
; GFX11-NEXT: s_or_saveexec_b32 s1, -1
-; GFX11-NEXT: scratch_load_b32 v40, off, s33 ; 4-byte Folded Reload
+; GFX11-NEXT: s_clause 0x1
+; GFX11-NEXT: scratch_load_b32 v40, off, s33
+; GFX11-NEXT: scratch_load_b32 v41, off, s33 offset:4
; GFX11-NEXT: s_mov_b32 exec_lo, s1
; GFX11-NEXT: s_add_i32 s32, s32, -16
; GFX11-NEXT: s_mov_b32 s33, s0
@@ -569,9 +606,10 @@
; GFX9-NEXT: s_mov_b32 s33, s32
; GFX9-NEXT: s_or_saveexec_b64 s[36:37], -1
; GFX9-NEXT: buffer_store_dword v40, off, s[0:3], s33 ; 4-byte Folded Spill
+; GFX9-NEXT: buffer_store_dword v41, off, s[0:3], s33 offset:4 ; 4-byte Folded Spill
; GFX9-NEXT: s_mov_b64 exec, s[36:37]
-; GFX9-NEXT: v_writelane_b32 v40, s34, 3
; GFX9-NEXT: v_writelane_b32 v40, s4, 0
+; GFX9-NEXT: v_writelane_b32 v41, s34, 0
; GFX9-NEXT: s_addk_i32 s32, 0x400
; GFX9-NEXT: v_writelane_b32 v40, s30, 1
; GFX9-NEXT: ;;#ASMSTART
@@ -590,9 +628,10 @@
; GFX9-NEXT: v_readlane_b32 s31, v40, 2
; GFX9-NEXT: v_readlane_b32 s30, v40, 1
; GFX9-NEXT: v_readlane_b32 s4, v40, 0
-; GFX9-NEXT: v_readlane_b32 s34, v40, 3
+; GFX9-NEXT: v_readlane_b32 s34, v41, 0
; GFX9-NEXT: s_or_saveexec_b64 s[36:37], -1
; GFX9-NEXT: buffer_load_dword v40, off, s[0:3], s33 ; 4-byte Folded Reload
+; GFX9-NEXT: buffer_load_dword v41, off, s[0:3], s33 offset:4 ; 4-byte Folded Reload
; GFX9-NEXT: s_mov_b64 exec, s[36:37]
; GFX9-NEXT: s_addk_i32 s32, 0xfc00
; GFX9-NEXT: s_mov_b32 s33, s34
@@ -606,19 +645,20 @@
; GFX10-NEXT: s_mov_b32 s33, s32
; GFX10-NEXT: s_or_saveexec_b32 s35, -1
; GFX10-NEXT: buffer_store_dword v40, off, s[0:3], s33 ; 4-byte Folded Spill
+; GFX10-NEXT: buffer_store_dword v41, off, s[0:3], s33 offset:4 ; 4-byte Folded Spill
; GFX10-NEXT: s_waitcnt_depctr 0xffe3
; GFX10-NEXT: s_mov_b32 exec_lo, s35
-; GFX10-NEXT: v_writelane_b32 v40, s34, 3
+; GFX10-NEXT: v_writelane_b32 v40, s4, 0
; GFX10-NEXT: s_addk_i32 s32, 0x200
+; GFX10-NEXT: v_writelane_b32 v41, s34, 0
; GFX10-NEXT: s_getpc_b64 s[36:37]
; GFX10-NEXT: s_add_u32 s36, s36, external_void_func_void@rel32@lo+4
; GFX10-NEXT: s_addc_u32 s37, s37, external_void_func_void@rel32@hi+12
; GFX10-NEXT: ;;#ASMSTART
; GFX10-NEXT: ; def s34
; GFX10-NEXT: ;;#ASMEND
-; GFX10-NEXT: v_writelane_b32 v40, s4, 0
-; GFX10-NEXT: s_mov_b32 s4, s34
; GFX10-NEXT: v_writelane_b32 v40, s30, 1
+; GFX10-NEXT: s_mov_b32 s4, s34
; GFX10-NEXT: v_writelane_b32 v40, s31, 2
; GFX10-NEXT: s_swappc_b64 s[30:31], s[36:37]
; GFX10-NEXT: s_mov_b32 s34, s4
@@ -628,9 +668,11 @@
; GFX10-NEXT: v_readlane_b32 s31, v40, 2
; GFX10-NEXT: v_readlane_b32 s30, v40, 1
; GFX10-NEXT: v_readlane_b32 s4, v40, 0
-; GFX10-NEXT: v_readlane_b32 s34, v40, 3
+; GFX10-NEXT: v_readlane_b32 s34, v41, 0
; GFX10-NEXT: s_or_saveexec_b32 s35, -1
-; GFX10-NEXT: buffer_load_dword v40, off, s[0:3], s33 ; 4-byte Folded Reload
+; GFX10-NEXT: s_clause 0x1
+; GFX10-NEXT: buffer_load_dword v40, off, s[0:3], s33
+; GFX10-NEXT: buffer_load_dword v41, off, s[0:3], s33 offset:4
; GFX10-NEXT: s_waitcnt_depctr 0xffe3
; GFX10-NEXT: s_mov_b32 exec_lo, s35
; GFX10-NEXT: s_addk_i32 s32, 0xfe00
@@ -644,19 +686,21 @@
; GFX11-NEXT: s_mov_b32 s0, s33
; GFX11-NEXT: s_mov_b32 s33, s32
; GFX11-NEXT: s_or_saveexec_b32 s1, -1
-; GFX11-NEXT: scratch_store_b32 off, v40, s33 ; 4-byte Folded Spill
+; GFX11-NEXT: s_clause 0x1
+; GFX11-NEXT: scratch_store_b32 off, v40, s33
+; GFX11-NEXT: scratch_store_b32 off, v41, s33 offset:4
; GFX11-NEXT: s_mov_b32 exec_lo, s1
-; GFX11-NEXT: v_writelane_b32 v40, s0, 3
+; GFX11-NEXT: v_writelane_b32 v40, s4, 0
; GFX11-NEXT: s_add_i32 s32, s32, 16
+; GFX11-NEXT: v_writelane_b32 v41, s0, 0
; GFX11-NEXT: s_getpc_b64 s[0:1]
; GFX11-NEXT: s_add_u32 s0, s0, external_void_func_void@rel32@lo+4
; GFX11-NEXT: s_addc_u32 s1, s1, external_void_func_void@rel32@hi+12
; GFX11-NEXT: ;;#ASMSTART
; GFX11-NEXT: ; def s34
; GFX11-NEXT: ;;#ASMEND
-; GFX11-NEXT: v_writelane_b32 v40, s4, 0
-; GFX11-NEXT: s_mov_b32 s4, s34
; GFX11-NEXT: v_writelane_b32 v40, s30, 1
+; GFX11-NEXT: s_mov_b32 s4, s34
; GFX11-NEXT: v_writelane_b32 v40, s31, 2
; GFX11-NEXT: s_swappc_b64 s[30:31], s[0:1]
; GFX11-NEXT: s_mov_b32 s34, s4
@@ -667,9 +711,11 @@
; GFX11-NEXT: v_readlane_b32 s31, v40, 2
; GFX11-NEXT: v_readlane_b32 s30, v40, 1
; GFX11-NEXT: v_readlane_b32 s4, v40, 0
-; GFX11-NEXT: v_readlane_b32 s0, v40, 3
+; GFX11-NEXT: v_readlane_b32 s0, v41, 0
; GFX11-NEXT: s_or_saveexec_b32 s1, -1
-; GFX11-NEXT: scratch_load_b32 v40, off, s33 ; 4-byte Folded Reload
+; GFX11-NEXT: s_clause 0x1
+; GFX11-NEXT: scratch_load_b32 v40, off, s33
+; GFX11-NEXT: scratch_load_b32 v41, off, s33 offset:4
; GFX11-NEXT: s_mov_b32 exec_lo, s1
; GFX11-NEXT: s_add_i32 s32, s32, -16
; GFX11-NEXT: s_mov_b32 s33, s0
@@ -689,10 +735,11 @@
; GFX9-NEXT: s_mov_b32 s33, s32
; GFX9-NEXT: s_or_saveexec_b64 s[36:37], -1
; GFX9-NEXT: buffer_store_dword v41, off, s[0:3], s33 offset:4 ; 4-byte Folded Spill
+; GFX9-NEXT: buffer_store_dword v42, off, s[0:3], s33 offset:8 ; 4-byte Folded Spill
; GFX9-NEXT: s_mov_b64 exec, s[36:37]
-; GFX9-NEXT: v_writelane_b32 v41, s34, 2
; GFX9-NEXT: s_addk_i32 s32, 0x400
; GFX9-NEXT: v_writelane_b32 v41, s30, 0
+; GFX9-NEXT: v_writelane_b32 v42, s34, 0
; GFX9-NEXT: buffer_store_dword v40, off, s[0:3], s33 ; 4-byte Folded Spill
; GFX9-NEXT: v_writelane_b32 v41, s31, 1
; GFX9-NEXT: ;;#ASMSTART
@@ -708,9 +755,10 @@
; GFX9-NEXT: buffer_load_dword v40, off, s[0:3], s33 ; 4-byte Folded Reload
; GFX9-NEXT: v_readlane_b32 s31, v41, 1
; GFX9-NEXT: v_readlane_b32 s30, v41, 0
-; GFX9-NEXT: v_readlane_b32 s34, v41, 2
+; GFX9-NEXT: v_readlane_b32 s34, v42, 0
; GFX9-NEXT: s_or_saveexec_b64 s[36:37], -1
; GFX9-NEXT: buffer_load_dword v41, off, s[0:3], s33 offset:4 ; 4-byte Folded Reload
+; GFX9-NEXT: buffer_load_dword v42, off, s[0:3], s33 offset:8 ; 4-byte Folded Reload
; GFX9-NEXT: s_mov_b64 exec, s[36:37]
; GFX9-NEXT: s_addk_i32 s32, 0xfc00
; GFX9-NEXT: s_mov_b32 s33, s34
@@ -724,19 +772,20 @@
; GFX10-NEXT: s_mov_b32 s33, s32
; GFX10-NEXT: s_or_saveexec_b32 s35, -1
; GFX10-NEXT: buffer_store_dword v41, off, s[0:3], s33 offset:4 ; 4-byte Folded Spill
+; GFX10-NEXT: buffer_store_dword v42, off, s[0:3], s33 offset:8 ; 4-byte Folded Spill
; GFX10-NEXT: s_waitcnt_depctr 0xffe3
; GFX10-NEXT: s_mov_b32 exec_lo, s35
-; GFX10-NEXT: v_writelane_b32 v41, s34, 2
+; GFX10-NEXT: v_writelane_b32 v41, s30, 0
; GFX10-NEXT: s_addk_i32 s32, 0x200
+; GFX10-NEXT: v_writelane_b32 v42, s34, 0
; GFX10-NEXT: buffer_store_dword v40, off, s[0:3], s33 ; 4-byte Folded Spill
; GFX10-NEXT: ;;#ASMSTART
; GFX10-NEXT: ; def v40
; GFX10-NEXT: ;;#ASMEND
+; GFX10-NEXT: v_writelane_b32 v41, s31, 1
; GFX10-NEXT: s_getpc_b64 s[34:35]
; GFX10-NEXT: s_add_u32 s34, s34, external_void_func_void@rel32@lo+4
; GFX10-NEXT: s_addc_u32 s35, s35, external_void_func_void@rel32@hi+12
-; GFX10-NEXT: v_writelane_b32 v41, s30, 0
-; GFX10-NEXT: v_writelane_b32 v41, s31, 1
; GFX10-NEXT: s_swappc_b64 s[30:31], s[34:35]
; GFX10-NEXT: ;;#ASMSTART
; GFX10-NEXT: ; use v40
@@ -744,9 +793,11 @@
; GFX10-NEXT: buffer_load_dword v40, off, s[0:3], s33 ; 4-byte Folded Reload
; GFX10-NEXT: v_readlane_b32 s31, v41, 1
; GFX10-NEXT: v_readlane_b32 s30, v41, 0
-; GFX10-NEXT: v_readlane_b32 s34, v41, 2
+; GFX10-NEXT: v_readlane_b32 s34, v42, 0
; GFX10-NEXT: s_or_saveexec_b32 s35, -1
-; GFX10-NEXT: buffer_load_dword v41, off, s[0:3], s33 offset:4 ; 4-byte Folded Reload
+; GFX10-NEXT: s_clause 0x1
+; GFX10-NEXT: buffer_load_dword v41, off, s[0:3], s33 offset:4
+; GFX10-NEXT: buffer_load_dword v42, off, s[0:3], s33 offset:8
; GFX10-NEXT: s_waitcnt_depctr 0xffe3
; GFX10-NEXT: s_mov_b32 exec_lo, s35
; GFX10-NEXT: s_addk_i32 s32, 0xfe00
@@ -760,19 +811,22 @@
; GFX11-NEXT: s_mov_b32 s0, s33
; GFX11-NEXT: s_mov_b32 s33, s32
; GFX11-NEXT: s_or_saveexec_b32 s1, -1
-; GFX11-NEXT: scratch_store_b32 off, v41, s33 offset:4 ; 4-byte Folded Spill
+; GFX11-NEXT: s_clause 0x1
+; GFX11-NEXT: scratch_store_b32 off, v41, s33 offset:4
+; GFX11-NEXT: scratch_store_b32 off, v42, s33 offset:8
; GFX11-NEXT: s_mov_b32 exec_lo, s1
-; GFX11-NEXT: v_writelane_b32 v41, s0, 2
+; GFX11-NEXT: v_writelane_b32 v41, s30, 0
; GFX11-NEXT: s_add_i32 s32, s32, 16
+; GFX11-NEXT: v_writelane_b32 v42, s0, 0
; GFX11-NEXT: scratch_store_b32 off, v40, s33 ; 4-byte Folded Spill
; GFX11-NEXT: ;;#ASMSTART
; GFX11-NEXT: ; def v40
; GFX11-NEXT: ;;#ASMEND
+; GFX11-NEXT: v_writelane_b32 v41, s31, 1
; GFX11-NEXT: s_getpc_b64 s[0:1]
; GFX11-NEXT: s_add_u32 s0, s0, external_void_func_void@rel32@lo+4
; GFX11-NEXT: s_addc_u32 s1, s1, external_void_func_void@rel32@hi+12
-; GFX11-NEXT: v_writelane_b32 v41, s30, 0
-; GFX11-NEXT: v_writelane_b32 v41, s31, 1
+; GFX11-NEXT: s_delay_alu instid0(SALU_CYCLE_1)
; GFX11-NEXT: s_swappc_b64 s[30:31], s[0:1]
; GFX11-NEXT: ;;#ASMSTART
; GFX11-NEXT: ; use v40
@@ -780,9 +834,11 @@
; GFX11-NEXT: scratch_load_b32 v40, off, s33 ; 4-byte Folded Reload
; GFX11-NEXT: v_readlane_b32 s31, v41, 1
; GFX11-NEXT: v_readlane_b32 s30, v41, 0
-; GFX11-NEXT: v_readlane_b32 s0, v41, 2
+; GFX11-NEXT: v_readlane_b32 s0, v42, 0
; GFX11-NEXT: s_or_saveexec_b32 s1, -1
-; GFX11-NEXT: scratch_load_b32 v41, off, s33 offset:4 ; 4-byte Folded Reload
+; GFX11-NEXT: s_clause 0x1
+; GFX11-NEXT: scratch_load_b32 v41, off, s33 offset:4
+; GFX11-NEXT: scratch_load_b32 v42, off, s33 offset:8
; GFX11-NEXT: s_mov_b32 exec_lo, s1
; GFX11-NEXT: s_add_i32 s32, s32, -16
; GFX11-NEXT: s_mov_b32 s33, s0
@@ -798,54 +854,54 @@
; GFX9-LABEL: void_func_void_clobber_s33:
; GFX9: ; %bb.0:
; GFX9-NEXT: s_waitcnt vmcnt(0) expcnt(0) lgkmcnt(0)
-; GFX9-NEXT: s_xor_saveexec_b64 s[4:5], -1
+; GFX9-NEXT: s_xor_saveexec_b64 s[6:7], -1
; GFX9-NEXT: buffer_store_dword v0, off, s[0:3], s32 ; 4-byte Folded Spill
-; GFX9-NEXT: s_mov_b64 exec, s[4:5]
+; GFX9-NEXT: s_mov_b64 exec, s[6:7]
; GFX9-NEXT: v_writelane_b32 v0, s33, 0
; GFX9-NEXT: ;;#ASMSTART
; GFX9-NEXT: ; clobber
; GFX9-NEXT: ;;#ASMEND
; GFX9-NEXT: v_readlane_b32 s33, v0, 0
-; GFX9-NEXT: s_xor_saveexec_b64 s[4:5], -1
+; GFX9-NEXT: s_xor_saveexec_b64 s[6:7], -1
; GFX9-NEXT: buffer_load_dword v0, off, s[0:3], s32 ; 4-byte Folded Reload
-; GFX9-NEXT: s_mov_b64 exec, s[4:5]
+; GFX9-NEXT: s_mov_b64 exec, s[6:7]
; GFX9-NEXT: s_waitcnt vmcnt(0)
; GFX9-NEXT: s_setpc_b64 s[30:31]
;
; GFX10-LABEL: void_func_void_clobber_s33:
; GFX10: ; %bb.0:
; GFX10-NEXT: s_waitcnt vmcnt(0) expcnt(0) lgkmcnt(0)
-; GFX10-NEXT: s_xor_saveexec_b32 s4, -1
+; GFX10-NEXT: s_xor_saveexec_b32 s5, -1
; GFX10-NEXT: buffer_store_dword v0, off, s[0:3], s32 ; 4-byte Folded Spill
; GFX10-NEXT: s_waitcnt_depctr 0xffe3
-; GFX10-NEXT: s_mov_b32 exec_lo, s4
+; GFX10-NEXT: s_mov_b32 exec_lo, s5
; GFX10-NEXT: v_writelane_b32 v0, s33, 0
; GFX10-NEXT: ;;#ASMSTART
; GFX10-NEXT: ; clobber
; GFX10-NEXT: ;;#ASMEND
; GFX10-NEXT: v_readlane_b32 s33, v0, 0
-; GFX10-NEXT: s_xor_saveexec_b32 s4, -1
+; GFX10-NEXT: s_xor_saveexec_b32 s5, -1
; GFX10-NEXT: buffer_load_dword v0, off, s[0:3], s32 ; 4-byte Folded Reload
; GFX10-NEXT: s_waitcnt_depctr 0xffe3
-; GFX10-NEXT: s_mov_b32 exec_lo, s4
+; GFX10-NEXT: s_mov_b32 exec_lo, s5
; GFX10-NEXT: s_waitcnt vmcnt(0)
; GFX10-NEXT: s_setpc_b64 s[30:31]
;
; GFX11-LABEL: void_func_void_clobber_s33:
; GFX11: ; %bb.0:
; GFX11-NEXT: s_waitcnt vmcnt(0) expcnt(0) lgkmcnt(0)
-; GFX11-NEXT: s_xor_saveexec_b32 s0, -1
+; GFX11-NEXT: s_xor_saveexec_b32 s1, -1
; GFX11-NEXT: scratch_store_b32 off, v0, s32 ; 4-byte Folded Spill
-; GFX11-NEXT: s_mov_b32 exec_lo, s0
+; GFX11-NEXT: s_mov_b32 exec_lo, s1
; GFX11-NEXT: v_writelane_b32 v0, s33, 0
; GFX11-NEXT: ;;#ASMSTART
; GFX11-NEXT: ; clobber
; GFX11-NEXT: ;;#ASMEND
; GFX11-NEXT: s_delay_alu instid0(VALU_DEP_1)
; GFX11-NEXT: v_readlane_b32 s33, v0, 0
-; GFX11-NEXT: s_xor_saveexec_b32 s0, -1
+; GFX11-NEXT: s_xor_saveexec_b32 s1, -1
; GFX11-NEXT: scratch_load_b32 v0, off, s32 ; 4-byte Folded Reload
-; GFX11-NEXT: s_mov_b32 exec_lo, s0
+; GFX11-NEXT: s_mov_b32 exec_lo, s1
; GFX11-NEXT: s_waitcnt vmcnt(0)
; GFX11-NEXT: s_setpc_b64 s[30:31]
call void asm sideeffect "; clobber", "~{s33}"() #0
@@ -856,54 +912,54 @@
; GFX9-LABEL: void_func_void_clobber_s34:
; GFX9: ; %bb.0:
; GFX9-NEXT: s_waitcnt vmcnt(0) expcnt(0) lgkmcnt(0)
-; GFX9-NEXT: s_xor_saveexec_b64 s[4:5], -1
+; GFX9-NEXT: s_xor_saveexec_b64 s[6:7], -1
; GFX9-NEXT: buffer_store_dword v0, off, s[0:3], s32 ; 4-byte Folded Spill
-; GFX9-NEXT: s_mov_b64 exec, s[4:5]
+; GFX9-NEXT: s_mov_b64 exec, s[6:7]
; GFX9-NEXT: v_writelane_b32 v0, s34, 0
; GFX9-NEXT: ;;#ASMSTART
; GFX9-NEXT: ; clobber
; GFX9-NEXT: ;;#ASMEND
; GFX9-NEXT: v_readlane_b32 s34, v0, 0
-; GFX9-NEXT: s_xor_saveexec_b64 s[4:5], -1
+; GFX9-NEXT: s_xor_saveexec_b64 s[6:7], -1
; GFX9-NEXT: buffer_load_dword v0, off, s[0:3], s32 ; 4-byte Folded Reload
-; GFX9-NEXT: s_mov_b64 exec, s[4:5]
+; GFX9-NEXT: s_mov_b64 exec, s[6:7]
; GFX9-NEXT: s_waitcnt vmcnt(0)
; GFX9-NEXT: s_setpc_b64 s[30:31]
;
; GFX10-LABEL: void_func_void_clobber_s34:
; GFX10: ; %bb.0:
; GFX10-NEXT: s_waitcnt vmcnt(0) expcnt(0) lgkmcnt(0)
-; GFX10-NEXT: s_xor_saveexec_b32 s4, -1
+; GFX10-NEXT: s_xor_saveexec_b32 s5, -1
; GFX10-NEXT: buffer_store_dword v0, off, s[0:3], s32 ; 4-byte Folded Spill
; GFX10-NEXT: s_waitcnt_depctr 0xffe3
-; GFX10-NEXT: s_mov_b32 exec_lo, s4
+; GFX10-NEXT: s_mov_b32 exec_lo, s5
; GFX10-NEXT: v_writelane_b32 v0, s34, 0
; GFX10-NEXT: ;;#ASMSTART
; GFX10-NEXT: ; clobber
; GFX10-NEXT: ;;#ASMEND
; GFX10-NEXT: v_readlane_b32 s34, v0, 0
-; GFX10-NEXT: s_xor_saveexec_b32 s4, -1
+; GFX10-NEXT: s_xor_saveexec_b32 s5, -1
; GFX10-NEXT: buffer_load_dword v0, off, s[0:3], s32 ; 4-byte Folded Reload
; GFX10-NEXT: s_waitcnt_depctr 0xffe3
-; GFX10-NEXT: s_mov_b32 exec_lo, s4
+; GFX10-NEXT: s_mov_b32 exec_lo, s5
; GFX10-NEXT: s_waitcnt vmcnt(0)
; GFX10-NEXT: s_setpc_b64 s[30:31]
;
; GFX11-LABEL: void_func_void_clobber_s34:
; GFX11: ; %bb.0:
; GFX11-NEXT: s_waitcnt vmcnt(0) expcnt(0) lgkmcnt(0)
-; GFX11-NEXT: s_xor_saveexec_b32 s0, -1
+; GFX11-NEXT: s_xor_saveexec_b32 s1, -1
; GFX11-NEXT: scratch_store_b32 off, v0, s32 ; 4-byte Folded Spill
-; GFX11-NEXT: s_mov_b32 exec_lo, s0
+; GFX11-NEXT: s_mov_b32 exec_lo, s1
; GFX11-NEXT: v_writelane_b32 v0, s34, 0
; GFX11-NEXT: ;;#ASMSTART
; GFX11-NEXT: ; clobber
; GFX11-NEXT: ;;#ASMEND
; GFX11-NEXT: s_delay_alu instid0(VALU_DEP_1)
; GFX11-NEXT: v_readlane_b32 s34, v0, 0
-; GFX11-NEXT: s_xor_saveexec_b32 s0, -1
+; GFX11-NEXT: s_xor_saveexec_b32 s1, -1
; GFX11-NEXT: scratch_load_b32 v0, off, s32 ; 4-byte Folded Reload
-; GFX11-NEXT: s_mov_b32 exec_lo, s0
+; GFX11-NEXT: s_mov_b32 exec_lo, s1
; GFX11-NEXT: s_waitcnt vmcnt(0)
; GFX11-NEXT: s_setpc_b64 s[30:31]
call void asm sideeffect "; clobber", "~{s34}"() #0
@@ -918,10 +974,11 @@
; GFX9-NEXT: s_mov_b32 s33, s32
; GFX9-NEXT: s_or_saveexec_b64 s[36:37], -1
; GFX9-NEXT: buffer_store_dword v40, off, s[0:3], s33 ; 4-byte Folded Spill
+; GFX9-NEXT: buffer_store_dword v41, off, s[0:3], s33 offset:4 ; 4-byte Folded Spill
; GFX9-NEXT: s_mov_b64 exec, s[36:37]
-; GFX9-NEXT: v_writelane_b32 v40, s34, 2
; GFX9-NEXT: s_addk_i32 s32, 0x400
; GFX9-NEXT: v_writelane_b32 v40, s30, 0
+; GFX9-NEXT: v_writelane_b32 v41, s34, 0
; GFX9-NEXT: v_writelane_b32 v40, s31, 1
; GFX9-NEXT: s_getpc_b64 s[34:35]
; GFX9-NEXT: s_add_u32 s34, s34, void_func_void_clobber_s33@rel32@lo+4
@@ -929,9 +986,10 @@
; GFX9-NEXT: s_swappc_b64 s[30:31], s[34:35]
; GFX9-NEXT: v_readlane_b32 s31, v40, 1
; GFX9-NEXT: v_readlane_b32 s30, v40, 0
-; GFX9-NEXT: v_readlane_b32 s34, v40, 2
+; GFX9-NEXT: v_readlane_b32 s34, v41, 0
; GFX9-NEXT: s_or_saveexec_b64 s[36:37], -1
; GFX9-NEXT: buffer_load_dword v40, off, s[0:3], s33 ; 4-byte Folded Reload
+; GFX9-NEXT: buffer_load_dword v41, off, s[0:3], s33 offset:4 ; 4-byte Folded Reload
; GFX9-NEXT: s_mov_b64 exec, s[36:37]
; GFX9-NEXT: s_addk_i32 s32, 0xfc00
; GFX9-NEXT: s_mov_b32 s33, s34
@@ -945,21 +1003,24 @@
; GFX10-NEXT: s_mov_b32 s33, s32
; GFX10-NEXT: s_or_saveexec_b32 s35, -1
; GFX10-NEXT: buffer_store_dword v40, off, s[0:3], s33 ; 4-byte Folded Spill
+; GFX10-NEXT: buffer_store_dword v41, off, s[0:3], s33 offset:4 ; 4-byte Folded Spill
; GFX10-NEXT: s_waitcnt_depctr 0xffe3
; GFX10-NEXT: s_mov_b32 exec_lo, s35
-; GFX10-NEXT: v_writelane_b32 v40, s34, 2
+; GFX10-NEXT: v_writelane_b32 v40, s30, 0
; GFX10-NEXT: s_addk_i32 s32, 0x200
+; GFX10-NEXT: v_writelane_b32 v41, s34, 0
; GFX10-NEXT: s_getpc_b64 s[34:35]
; GFX10-NEXT: s_add_u32 s34, s34, void_func_void_clobber_s33@rel32@lo+4
; GFX10-NEXT: s_addc_u32 s35, s35, void_func_void_clobber_s33@rel32@hi+12
-; GFX10-NEXT: v_writelane_b32 v40, s30, 0
; GFX10-NEXT: v_writelane_b32 v40, s31, 1
; GFX10-NEXT: s_swappc_b64 s[30:31], s[34:35]
; GFX10-NEXT: v_readlane_b32 s31, v40, 1
; GFX10-NEXT: v_readlane_b32 s30, v40, 0
-; GFX10-NEXT: v_readlane_b32 s34, v40, 2
+; GFX10-NEXT: v_readlane_b32 s34, v41, 0
; GFX10-NEXT: s_or_saveexec_b32 s35, -1
-; GFX10-NEXT: buffer_load_dword v40, off, s[0:3], s33 ; 4-byte Folded Reload
+; GFX10-NEXT: s_clause 0x1
+; GFX10-NEXT: buffer_load_dword v40, off, s[0:3], s33
+; GFX10-NEXT: buffer_load_dword v41, off, s[0:3], s33 offset:4
; GFX10-NEXT: s_waitcnt_depctr 0xffe3
; GFX10-NEXT: s_mov_b32 exec_lo, s35
; GFX10-NEXT: s_addk_i32 s32, 0xfe00
@@ -973,22 +1034,26 @@
; GFX11-NEXT: s_mov_b32 s0, s33
; GFX11-NEXT: s_mov_b32 s33, s32
; GFX11-NEXT: s_or_saveexec_b32 s1, -1
-; GFX11-NEXT: scratch_store_b32 off, v40, s33 ; 4-byte Folded Spill
+; GFX11-NEXT: s_clause 0x1
+; GFX11-NEXT: scratch_store_b32 off, v40, s33
+; GFX11-NEXT: scratch_store_b32 off, v41, s33 offset:4
; GFX11-NEXT: s_mov_b32 exec_lo, s1
-; GFX11-NEXT: v_writelane_b32 v40, s0, 2
+; GFX11-NEXT: v_writelane_b32 v40, s30, 0
; GFX11-NEXT: s_add_i32 s32, s32, 16
+; GFX11-NEXT: v_writelane_b32 v41, s0, 0
; GFX11-NEXT: s_getpc_b64 s[0:1]
; GFX11-NEXT: s_add_u32 s0, s0, void_func_void_clobber_s33@rel32@lo+4
; GFX11-NEXT: s_addc_u32 s1, s1, void_func_void_clobber_s33@rel32@hi+12
-; GFX11-NEXT: v_writelane_b32 v40, s30, 0
; GFX11-NEXT: v_writelane_b32 v40, s31, 1
; GFX11-NEXT: s_swappc_b64 s[30:31], s[0:1]
; GFX11-NEXT: s_delay_alu instid0(VALU_DEP_1)
; GFX11-NEXT: v_readlane_b32 s31, v40, 1
; GFX11-NEXT: v_readlane_b32 s30, v40, 0
-; GFX11-NEXT: v_readlane_b32 s0, v40, 2
+; GFX11-NEXT: v_readlane_b32 s0, v41, 0
; GFX11-NEXT: s_or_saveexec_b32 s1, -1
-; GFX11-NEXT: scratch_load_b32 v40, off, s33 ; 4-byte Folded Reload
+; GFX11-NEXT: s_clause 0x1
+; GFX11-NEXT: scratch_load_b32 v40, off, s33
+; GFX11-NEXT: scratch_load_b32 v41, off, s33 offset:4
; GFX11-NEXT: s_mov_b32 exec_lo, s1
; GFX11-NEXT: s_add_i32 s32, s32, -16
; GFX11-NEXT: s_mov_b32 s33, s0
@@ -1006,10 +1071,11 @@
; GFX9-NEXT: s_mov_b32 s33, s32
; GFX9-NEXT: s_or_saveexec_b64 s[36:37], -1
; GFX9-NEXT: buffer_store_dword v40, off, s[0:3], s33 ; 4-byte Folded Spill
+; GFX9-NEXT: buffer_store_dword v41, off, s[0:3], s33 offset:4 ; 4-byte Folded Spill
; GFX9-NEXT: s_mov_b64 exec, s[36:37]
-; GFX9-NEXT: v_writelane_b32 v40, s34, 2
; GFX9-NEXT: s_addk_i32 s32, 0x400
; GFX9-NEXT: v_writelane_b32 v40, s30, 0
+; GFX9-NEXT: v_writelane_b32 v41, s34, 0
; GFX9-NEXT: v_writelane_b32 v40, s31, 1
; GFX9-NEXT: s_getpc_b64 s[34:35]
; GFX9-NEXT: s_add_u32 s34, s34, void_func_void_clobber_s34@rel32@lo+4
@@ -1017,9 +1083,10 @@
; GFX9-NEXT: s_swappc_b64 s[30:31], s[34:35]
; GFX9-NEXT: v_readlane_b32 s31, v40, 1
; GFX9-NEXT: v_readlane_b32 s30, v40, 0
-; GFX9-NEXT: v_readlane_b32 s34, v40, 2
+; GFX9-NEXT: v_readlane_b32 s34, v41, 0
; GFX9-NEXT: s_or_saveexec_b64 s[36:37], -1
; GFX9-NEXT: buffer_load_dword v40, off, s[0:3], s33 ; 4-byte Folded Reload
+; GFX9-NEXT: buffer_load_dword v41, off, s[0:3], s33 offset:4 ; 4-byte Folded Reload
; GFX9-NEXT: s_mov_b64 exec, s[36:37]
; GFX9-NEXT: s_addk_i32 s32, 0xfc00
; GFX9-NEXT: s_mov_b32 s33, s34
@@ -1033,21 +1100,24 @@
; GFX10-NEXT: s_mov_b32 s33, s32
; GFX10-NEXT: s_or_saveexec_b32 s35, -1
; GFX10-NEXT: buffer_store_dword v40, off, s[0:3], s33 ; 4-byte Folded Spill
+; GFX10-NEXT: buffer_store_dword v41, off, s[0:3], s33 offset:4 ; 4-byte Folded Spill
; GFX10-NEXT: s_waitcnt_depctr 0xffe3
; GFX10-NEXT: s_mov_b32 exec_lo, s35
-; GFX10-NEXT: v_writelane_b32 v40, s34, 2
+; GFX10-NEXT: v_writelane_b32 v40, s30, 0
; GFX10-NEXT: s_addk_i32 s32, 0x200
+; GFX10-NEXT: v_writelane_b32 v41, s34, 0
; GFX10-NEXT: s_getpc_b64 s[34:35]
; GFX10-NEXT: s_add_u32 s34, s34, void_func_void_clobber_s34@rel32@lo+4
; GFX10-NEXT: s_addc_u32 s35, s35, void_func_void_clobber_s34@rel32@hi+12
-; GFX10-NEXT: v_writelane_b32 v40, s30, 0
; GFX10-NEXT: v_writelane_b32 v40, s31, 1
; GFX10-NEXT: s_swappc_b64 s[30:31], s[34:35]
; GFX10-NEXT: v_readlane_b32 s31, v40, 1
; GFX10-NEXT: v_readlane_b32 s30, v40, 0
-; GFX10-NEXT: v_readlane_b32 s34, v40, 2
+; GFX10-NEXT: v_readlane_b32 s34, v41, 0
; GFX10-NEXT: s_or_saveexec_b32 s35, -1
-; GFX10-NEXT: buffer_load_dword v40, off, s[0:3], s33 ; 4-byte Folded Reload
+; GFX10-NEXT: s_clause 0x1
+; GFX10-NEXT: buffer_load_dword v40, off, s[0:3], s33
+; GFX10-NEXT: buffer_load_dword v41, off, s[0:3], s33 offset:4
; GFX10-NEXT: s_waitcnt_depctr 0xffe3
; GFX10-NEXT: s_mov_b32 exec_lo, s35
; GFX10-NEXT: s_addk_i32 s32, 0xfe00
@@ -1061,22 +1131,26 @@
; GFX11-NEXT: s_mov_b32 s0, s33
; GFX11-NEXT: s_mov_b32 s33, s32
; GFX11-NEXT: s_or_saveexec_b32 s1, -1
-; GFX11-NEXT: scratch_store_b32 off, v40, s33 ; 4-byte Folded Spill
+; GFX11-NEXT: s_clause 0x1
+; GFX11-NEXT: scratch_store_b32 off, v40, s33
+; GFX11-NEXT: scratch_store_b32 off, v41, s33 offset:4
; GFX11-NEXT: s_mov_b32 exec_lo, s1
-; GFX11-NEXT: v_writelane_b32 v40, s0, 2
+; GFX11-NEXT: v_writelane_b32 v40, s30, 0
; GFX11-NEXT: s_add_i32 s32, s32, 16
+; GFX11-NEXT: v_writelane_b32 v41, s0, 0
; GFX11-NEXT: s_getpc_b64 s[0:1]
; GFX11-NEXT: s_add_u32 s0, s0, void_func_void_clobber_s34@rel32@lo+4
; GFX11-NEXT: s_addc_u32 s1, s1, void_func_void_clobber_s34@rel32@hi+12
-; GFX11-NEXT: v_writelane_b32 v40, s30, 0
; GFX11-NEXT: v_writelane_b32 v40, s31, 1
; GFX11-NEXT: s_swappc_b64 s[30:31], s[0:1]
; GFX11-NEXT: s_delay_alu instid0(VALU_DEP_1)
; GFX11-NEXT: v_readlane_b32 s31, v40, 1
; GFX11-NEXT: v_readlane_b32 s30, v40, 0
-; GFX11-NEXT: v_readlane_b32 s0, v40, 2
+; GFX11-NEXT: v_readlane_b32 s0, v41, 0
; GFX11-NEXT: s_or_saveexec_b32 s1, -1
-; GFX11-NEXT: scratch_load_b32 v40, off, s33 ; 4-byte Folded Reload
+; GFX11-NEXT: s_clause 0x1
+; GFX11-NEXT: scratch_load_b32 v40, off, s33
+; GFX11-NEXT: scratch_load_b32 v41, off, s33 offset:4
; GFX11-NEXT: s_mov_b32 exec_lo, s1
; GFX11-NEXT: s_add_i32 s32, s32, -16
; GFX11-NEXT: s_mov_b32 s33, s0
@@ -1094,11 +1168,12 @@
; GFX9-NEXT: s_mov_b32 s33, s32
; GFX9-NEXT: s_or_saveexec_b64 s[36:37], -1
; GFX9-NEXT: buffer_store_dword v40, off, s[0:3], s33 ; 4-byte Folded Spill
+; GFX9-NEXT: buffer_store_dword v41, off, s[0:3], s33 offset:4 ; 4-byte Folded Spill
; GFX9-NEXT: s_mov_b64 exec, s[36:37]
-; GFX9-NEXT: v_writelane_b32 v40, s34, 3
; GFX9-NEXT: v_writelane_b32 v40, s4, 0
; GFX9-NEXT: s_addk_i32 s32, 0x400
; GFX9-NEXT: v_writelane_b32 v40, s30, 1
+; GFX9-NEXT: v_writelane_b32 v41, s34, 0
; GFX9-NEXT: v_writelane_b32 v40, s31, 2
; GFX9-NEXT: ;;#ASMSTART
; GFX9-NEXT: ; def s40
@@ -1114,9 +1189,10 @@
; GFX9-NEXT: v_readlane_b32 s31, v40, 2
; GFX9-NEXT: v_readlane_b32 s30, v40, 1
; GFX9-NEXT: v_readlane_b32 s4, v40, 0
-; GFX9-NEXT: v_readlane_b32 s34, v40, 3
+; GFX9-NEXT: v_readlane_b32 s34, v41, 0
; GFX9-NEXT: s_or_saveexec_b64 s[36:37], -1
; GFX9-NEXT: buffer_load_dword v40, off, s[0:3], s33 ; 4-byte Folded Reload
+; GFX9-NEXT: buffer_load_dword v41, off, s[0:3], s33 offset:4 ; 4-byte Folded Reload
; GFX9-NEXT: s_mov_b64 exec, s[36:37]
; GFX9-NEXT: s_addk_i32 s32, 0xfc00
; GFX9-NEXT: s_mov_b32 s33, s34
@@ -1130,19 +1206,20 @@
; GFX10-NEXT: s_mov_b32 s33, s32
; GFX10-NEXT: s_or_saveexec_b32 s35, -1
; GFX10-NEXT: buffer_store_dword v40, off, s[0:3], s33 ; 4-byte Folded Spill
+; GFX10-NEXT: buffer_store_dword v41, off, s[0:3], s33 offset:4 ; 4-byte Folded Spill
; GFX10-NEXT: s_waitcnt_depctr 0xffe3
; GFX10-NEXT: s_mov_b32 exec_lo, s35
-; GFX10-NEXT: v_writelane_b32 v40, s34, 3
+; GFX10-NEXT: v_writelane_b32 v40, s4, 0
; GFX10-NEXT: s_addk_i32 s32, 0x200
+; GFX10-NEXT: v_writelane_b32 v41, s34, 0
; GFX10-NEXT: s_getpc_b64 s[34:35]
; GFX10-NEXT: s_add_u32 s34, s34, external_void_func_void@rel32@lo+4
; GFX10-NEXT: s_addc_u32 s35, s35, external_void_func_void@rel32@hi+12
; GFX10-NEXT: ;;#ASMSTART
; GFX10-NEXT: ; def s40
; GFX10-NEXT: ;;#ASMEND
-; GFX10-NEXT: v_writelane_b32 v40, s4, 0
-; GFX10-NEXT: s_mov_b32 s4, s40
; GFX10-NEXT: v_writelane_b32 v40, s30, 1
+; GFX10-NEXT: s_mov_b32 s4, s40
; GFX10-NEXT: v_writelane_b32 v40, s31, 2
; GFX10-NEXT: s_swappc_b64 s[30:31], s[34:35]
; GFX10-NEXT: ;;#ASMSTART
@@ -1151,9 +1228,11 @@
; GFX10-NEXT: v_readlane_b32 s31, v40, 2
; GFX10-NEXT: v_readlane_b32 s30, v40, 1
; GFX10-NEXT: v_readlane_b32 s4, v40, 0
-; GFX10-NEXT: v_readlane_b32 s34, v40, 3
+; GFX10-NEXT: v_readlane_b32 s34, v41, 0
; GFX10-NEXT: s_or_saveexec_b32 s35, -1
-; GFX10-NEXT: buffer_load_dword v40, off, s[0:3], s33 ; 4-byte Folded Reload
+; GFX10-NEXT: s_clause 0x1
+; GFX10-NEXT: buffer_load_dword v40, off, s[0:3], s33
+; GFX10-NEXT: buffer_load_dword v41, off, s[0:3], s33 offset:4
; GFX10-NEXT: s_waitcnt_depctr 0xffe3
; GFX10-NEXT: s_mov_b32 exec_lo, s35
; GFX10-NEXT: s_addk_i32 s32, 0xfe00
@@ -1167,19 +1246,21 @@
; GFX11-NEXT: s_mov_b32 s0, s33
; GFX11-NEXT: s_mov_b32 s33, s32
; GFX11-NEXT: s_or_saveexec_b32 s1, -1
-; GFX11-NEXT: scratch_store_b32 off, v40, s33 ; 4-byte Folded Spill
+; GFX11-NEXT: s_clause 0x1
+; GFX11-NEXT: scratch_store_b32 off, v40, s33
+; GFX11-NEXT: scratch_store_b32 off, v41, s33 offset:4
; GFX11-NEXT: s_mov_b32 exec_lo, s1
-; GFX11-NEXT: v_writelane_b32 v40, s0, 3
+; GFX11-NEXT: v_writelane_b32 v40, s4, 0
; GFX11-NEXT: s_add_i32 s32, s32, 16
+; GFX11-NEXT: v_writelane_b32 v41, s0, 0
; GFX11-NEXT: s_getpc_b64 s[0:1]
; GFX11-NEXT: s_add_u32 s0, s0, external_void_func_void@rel32@lo+4
; GFX11-NEXT: s_addc_u32 s1, s1, external_void_func_void@rel32@hi+12
; GFX11-NEXT: ;;#ASMSTART
; GFX11-NEXT: ; def s40
; GFX11-NEXT: ;;#ASMEND
-; GFX11-NEXT: v_writelane_b32 v40, s4, 0
-; GFX11-NEXT: s_mov_b32 s4, s40
; GFX11-NEXT: v_writelane_b32 v40, s30, 1
+; GFX11-NEXT: s_mov_b32 s4, s40
; GFX11-NEXT: v_writelane_b32 v40, s31, 2
; GFX11-NEXT: s_swappc_b64 s[30:31], s[0:1]
; GFX11-NEXT: ;;#ASMSTART
@@ -1189,9 +1270,11 @@
; GFX11-NEXT: v_readlane_b32 s31, v40, 2
; GFX11-NEXT: v_readlane_b32 s30, v40, 1
; GFX11-NEXT: v_readlane_b32 s4, v40, 0
-; GFX11-NEXT: v_readlane_b32 s0, v40, 3
+; GFX11-NEXT: v_readlane_b32 s0, v41, 0
; GFX11-NEXT: s_or_saveexec_b32 s1, -1
-; GFX11-NEXT: scratch_load_b32 v40, off, s33 ; 4-byte Folded Reload
+; GFX11-NEXT: s_clause 0x1
+; GFX11-NEXT: scratch_load_b32 v40, off, s33
+; GFX11-NEXT: scratch_load_b32 v41, off, s33 offset:4
; GFX11-NEXT: s_mov_b32 exec_lo, s1
; GFX11-NEXT: s_add_i32 s32, s32, -16
; GFX11-NEXT: s_mov_b32 s33, s0
@@ -1211,11 +1294,12 @@
; GFX9-NEXT: s_mov_b32 s33, s32
; GFX9-NEXT: s_or_saveexec_b64 s[36:37], -1
; GFX9-NEXT: buffer_store_dword v40, off, s[0:3], s33 offset:4 ; 4-byte Folded Spill
+; GFX9-NEXT: buffer_store_dword v42, off, s[0:3], s33 offset:8 ; 4-byte Folded Spill
; GFX9-NEXT: s_mov_b64 exec, s[36:37]
-; GFX9-NEXT: v_writelane_b32 v40, s34, 3
; GFX9-NEXT: v_writelane_b32 v40, s4, 0
; GFX9-NEXT: s_addk_i32 s32, 0x400
; GFX9-NEXT: v_writelane_b32 v40, s30, 1
+; GFX9-NEXT: v_writelane_b32 v42, s34, 0
; GFX9-NEXT: buffer_store_dword v41, off, s[0:3], s33 ; 4-byte Folded Spill
; GFX9-NEXT: v_writelane_b32 v40, s31, 2
; GFX9-NEXT: ;;#ASMSTART
@@ -1240,9 +1324,10 @@
; GFX9-NEXT: v_readlane_b32 s31, v40, 2
; GFX9-NEXT: v_readlane_b32 s30, v40, 1
; GFX9-NEXT: v_readlane_b32 s4, v40, 0
-; GFX9-NEXT: v_readlane_b32 s34, v40, 3
+; GFX9-NEXT: v_readlane_b32 s34, v42, 0
; GFX9-NEXT: s_or_saveexec_b64 s[36:37], -1
; GFX9-NEXT: buffer_load_dword v40, off, s[0:3], s33 offset:4 ; 4-byte Folded Reload
+; GFX9-NEXT: buffer_load_dword v42, off, s[0:3], s33 offset:8 ; 4-byte Folded Reload
; GFX9-NEXT: s_mov_b64 exec, s[36:37]
; GFX9-NEXT: s_addk_i32 s32, 0xfc00
; GFX9-NEXT: s_mov_b32 s33, s34
@@ -1256,24 +1341,25 @@
; GFX10-NEXT: s_mov_b32 s33, s32
; GFX10-NEXT: s_or_saveexec_b32 s35, -1
; GFX10-NEXT: buffer_store_dword v40, off, s[0:3], s33 offset:4 ; 4-byte Folded Spill
+; GFX10-NEXT: buffer_store_dword v42, off, s[0:3], s33 offset:8 ; 4-byte Folded Spill
; GFX10-NEXT: s_waitcnt_depctr 0xffe3
; GFX10-NEXT: s_mov_b32 exec_lo, s35
-; GFX10-NEXT: v_writelane_b32 v40, s34, 3
+; GFX10-NEXT: v_writelane_b32 v40, s4, 0
; GFX10-NEXT: s_addk_i32 s32, 0x200
+; GFX10-NEXT: v_writelane_b32 v42, s34, 0
; GFX10-NEXT: buffer_store_dword v41, off, s[0:3], s33 ; 4-byte Folded Spill
; GFX10-NEXT: ;;#ASMSTART
; GFX10-NEXT: ; def s40
; GFX10-NEXT: ;;#ASMEND
+; GFX10-NEXT: v_writelane_b32 v40, s30, 1
+; GFX10-NEXT: s_mov_b32 s4, s40
; GFX10-NEXT: ;;#ASMSTART
; GFX10-NEXT: ; def v32
; GFX10-NEXT: ;;#ASMEND
-; GFX10-NEXT: v_writelane_b32 v40, s4, 0
-; GFX10-NEXT: s_mov_b32 s4, s40
; GFX10-NEXT: v_mov_b32_e32 v41, v32
; GFX10-NEXT: s_getpc_b64 s[34:35]
; GFX10-NEXT: s_add_u32 s34, s34, external_void_func_void@rel32@lo+4
; GFX10-NEXT: s_addc_u32 s35, s35, external_void_func_void@rel32@hi+12
-; GFX10-NEXT: v_writelane_b32 v40, s30, 1
; GFX10-NEXT: v_writelane_b32 v40, s31, 2
; GFX10-NEXT: s_swappc_b64 s[30:31], s[34:35]
; GFX10-NEXT: ;;#ASMSTART
@@ -1286,9 +1372,11 @@
; GFX10-NEXT: v_readlane_b32 s31, v40, 2
; GFX10-NEXT: v_readlane_b32 s30, v40, 1
; GFX10-NEXT: v_readlane_b32 s4, v40, 0
-; GFX10-NEXT: v_readlane_b32 s34, v40, 3
+; GFX10-NEXT: v_readlane_b32 s34, v42, 0
; GFX10-NEXT: s_or_saveexec_b32 s35, -1
-; GFX10-NEXT: buffer_load_dword v40, off, s[0:3], s33 offset:4 ; 4-byte Folded Reload
+; GFX10-NEXT: s_clause 0x1
+; GFX10-NEXT: buffer_load_dword v40, off, s[0:3], s33 offset:4
+; GFX10-NEXT: buffer_load_dword v42, off, s[0:3], s33 offset:8
; GFX10-NEXT: s_waitcnt_depctr 0xffe3
; GFX10-NEXT: s_mov_b32 exec_lo, s35
; GFX10-NEXT: s_addk_i32 s32, 0xfe00
@@ -1302,24 +1390,26 @@
; GFX11-NEXT: s_mov_b32 s0, s33
; GFX11-NEXT: s_mov_b32 s33, s32
; GFX11-NEXT: s_or_saveexec_b32 s1, -1
-; GFX11-NEXT: scratch_store_b32 off, v40, s33 offset:4 ; 4-byte Folded Spill
+; GFX11-NEXT: s_clause 0x1
+; GFX11-NEXT: scratch_store_b32 off, v40, s33 offset:4
+; GFX11-NEXT: scratch_store_b32 off, v42, s33 offset:8
; GFX11-NEXT: s_mov_b32 exec_lo, s1
-; GFX11-NEXT: v_writelane_b32 v40, s0, 3
+; GFX11-NEXT: v_writelane_b32 v40, s4, 0
; GFX11-NEXT: s_add_i32 s32, s32, 16
+; GFX11-NEXT: v_writelane_b32 v42, s0, 0
; GFX11-NEXT: scratch_store_b32 off, v41, s33 ; 4-byte Folded Spill
; GFX11-NEXT: ;;#ASMSTART
; GFX11-NEXT: ; def s40
; GFX11-NEXT: ;;#ASMEND
+; GFX11-NEXT: v_writelane_b32 v40, s30, 1
+; GFX11-NEXT: s_mov_b32 s4, s40
; GFX11-NEXT: ;;#ASMSTART
; GFX11-NEXT: ; def v32
; GFX11-NEXT: ;;#ASMEND
-; GFX11-NEXT: v_writelane_b32 v40, s4, 0
-; GFX11-NEXT: s_mov_b32 s4, s40
; GFX11-NEXT: v_mov_b32_e32 v41, v32
; GFX11-NEXT: s_getpc_b64 s[0:1]
; GFX11-NEXT: s_add_u32 s0, s0, external_void_func_void@rel32@lo+4
; GFX11-NEXT: s_addc_u32 s1, s1, external_void_func_void@rel32@hi+12
-; GFX11-NEXT: v_writelane_b32 v40, s30, 1
; GFX11-NEXT: v_writelane_b32 v40, s31, 2
; GFX11-NEXT: s_swappc_b64 s[30:31], s[0:1]
; GFX11-NEXT: ;;#ASMSTART
@@ -1332,9 +1422,11 @@
; GFX11-NEXT: v_readlane_b32 s31, v40, 2
; GFX11-NEXT: v_readlane_b32 s30, v40, 1
; GFX11-NEXT: v_readlane_b32 s4, v40, 0
-; GFX11-NEXT: v_readlane_b32 s0, v40, 3
+; GFX11-NEXT: v_readlane_b32 s0, v42, 0
; GFX11-NEXT: s_or_saveexec_b32 s1, -1
-; GFX11-NEXT: scratch_load_b32 v40, off, s33 offset:4 ; 4-byte Folded Reload
+; GFX11-NEXT: s_clause 0x1
+; GFX11-NEXT: scratch_load_b32 v40, off, s33 offset:4
+; GFX11-NEXT: scratch_load_b32 v42, off, s33 offset:8
; GFX11-NEXT: s_mov_b32 exec_lo, s1
; GFX11-NEXT: s_add_i32 s32, s32, -16
; GFX11-NEXT: s_mov_b32 s33, s0
diff --git a/llvm/test/CodeGen/AMDGPU/gfx-callable-return-types.ll b/llvm/test/CodeGen/AMDGPU/gfx-callable-return-types.ll
--- a/llvm/test/CodeGen/AMDGPU/gfx-callable-return-types.ll
+++ b/llvm/test/CodeGen/AMDGPU/gfx-callable-return-types.ll
@@ -79,7 +79,7 @@
; GFX11-LABEL: call_i1:
; GFX11: ; %bb.0: ; %entry
; GFX11-NEXT: s_waitcnt vmcnt(0) expcnt(0) lgkmcnt(0)
-; GFX11-NEXT: s_mov_b32 s2, s33
+; GFX11-NEXT: s_mov_b32 s3, s33
; GFX11-NEXT: s_mov_b32 s33, s32
; GFX11-NEXT: s_xor_saveexec_b32 s0, -1
; GFX11-NEXT: scratch_store_b32 off, v1, s33 ; 4-byte Folded Spill
@@ -100,7 +100,7 @@
; GFX11-NEXT: scratch_load_b32 v1, off, s33 ; 4-byte Folded Reload
; GFX11-NEXT: s_mov_b32 exec_lo, s0
; GFX11-NEXT: s_add_i32 s32, s32, -16
-; GFX11-NEXT: s_mov_b32 s33, s2
+; GFX11-NEXT: s_mov_b32 s33, s3
; GFX11-NEXT: s_waitcnt vmcnt(0)
; GFX11-NEXT: s_setpc_b64 s[30:31]
entry:
@@ -184,7 +184,7 @@
; GFX11-LABEL: call_i16:
; GFX11: ; %bb.0: ; %entry
; GFX11-NEXT: s_waitcnt vmcnt(0) expcnt(0) lgkmcnt(0)
-; GFX11-NEXT: s_mov_b32 s2, s33
+; GFX11-NEXT: s_mov_b32 s3, s33
; GFX11-NEXT: s_mov_b32 s33, s32
; GFX11-NEXT: s_xor_saveexec_b32 s0, -1
; GFX11-NEXT: scratch_store_b32 off, v1, s33 ; 4-byte Folded Spill
@@ -205,7 +205,7 @@
; GFX11-NEXT: scratch_load_b32 v1, off, s33 ; 4-byte Folded Reload
; GFX11-NEXT: s_mov_b32 exec_lo, s0
; GFX11-NEXT: s_add_i32 s32, s32, -16
-; GFX11-NEXT: s_mov_b32 s33, s2
+; GFX11-NEXT: s_mov_b32 s33, s3
; GFX11-NEXT: s_waitcnt vmcnt(0)
; GFX11-NEXT: s_setpc_b64 s[30:31]
entry:
@@ -289,7 +289,7 @@
; GFX11-LABEL: call_2xi16:
; GFX11: ; %bb.0: ; %entry
; GFX11-NEXT: s_waitcnt vmcnt(0) expcnt(0) lgkmcnt(0)
-; GFX11-NEXT: s_mov_b32 s2, s33
+; GFX11-NEXT: s_mov_b32 s3, s33
; GFX11-NEXT: s_mov_b32 s33, s32
; GFX11-NEXT: s_xor_saveexec_b32 s0, -1
; GFX11-NEXT: scratch_store_b32 off, v1, s33 ; 4-byte Folded Spill
@@ -310,7 +310,7 @@
; GFX11-NEXT: scratch_load_b32 v1, off, s33 ; 4-byte Folded Reload
; GFX11-NEXT: s_mov_b32 exec_lo, s0
; GFX11-NEXT: s_add_i32 s32, s32, -16
-; GFX11-NEXT: s_mov_b32 s33, s2
+; GFX11-NEXT: s_mov_b32 s33, s3
; GFX11-NEXT: s_waitcnt vmcnt(0)
; GFX11-NEXT: s_setpc_b64 s[30:31]
entry:
@@ -402,7 +402,7 @@
; GFX11-LABEL: call_3xi16:
; GFX11: ; %bb.0: ; %entry
; GFX11-NEXT: s_waitcnt vmcnt(0) expcnt(0) lgkmcnt(0)
-; GFX11-NEXT: s_mov_b32 s2, s33
+; GFX11-NEXT: s_mov_b32 s3, s33
; GFX11-NEXT: s_mov_b32 s33, s32
; GFX11-NEXT: s_xor_saveexec_b32 s0, -1
; GFX11-NEXT: scratch_store_b32 off, v2, s33 ; 4-byte Folded Spill
@@ -423,7 +423,7 @@
; GFX11-NEXT: scratch_load_b32 v2, off, s33 ; 4-byte Folded Reload
; GFX11-NEXT: s_mov_b32 exec_lo, s0
; GFX11-NEXT: s_add_i32 s32, s32, -16
-; GFX11-NEXT: s_mov_b32 s33, s2
+; GFX11-NEXT: s_mov_b32 s33, s3
; GFX11-NEXT: s_waitcnt vmcnt(0)
; GFX11-NEXT: s_setpc_b64 s[30:31]
entry:
@@ -890,7 +890,7 @@
; GFX11-LABEL: call_100xi32:
; GFX11: ; %bb.0: ; %entry
; GFX11-NEXT: s_waitcnt vmcnt(0) expcnt(0) lgkmcnt(0)
-; GFX11-NEXT: s_mov_b32 s2, s33
+; GFX11-NEXT: s_mov_b32 s3, s33
; GFX11-NEXT: s_mov_b32 s33, s32
; GFX11-NEXT: s_xor_saveexec_b32 s0, -1
; GFX11-NEXT: scratch_store_b32 off, v100, s33 offset:128 ; 4-byte Folded Spill
@@ -976,7 +976,7 @@
; GFX11-NEXT: scratch_load_b32 v100, off, s33 offset:128 ; 4-byte Folded Reload
; GFX11-NEXT: s_mov_b32 exec_lo, s0
; GFX11-NEXT: s_addk_i32 s32, 0xff70
-; GFX11-NEXT: s_mov_b32 s33, s2
+; GFX11-NEXT: s_mov_b32 s33, s3
; GFX11-NEXT: s_waitcnt vmcnt(0)
; GFX11-NEXT: s_setpc_b64 s[30:31]
entry:
@@ -2835,7 +2835,7 @@
; GFX9-NEXT: s_add_i32 s33, s32, 0x7fc0
; GFX9-NEXT: s_and_b32 s33, s33, 0xffff8000
; GFX9-NEXT: s_xor_saveexec_b64 s[34:35], -1
-; GFX9-NEXT: buffer_store_dword v33, off, s[0:3], s33 offset:1536 ; 4-byte Folded Spill
+; GFX9-NEXT: buffer_store_dword v33, off, s[0:3], s33 offset:1568 ; 4-byte Folded Spill
; GFX9-NEXT: s_mov_b64 exec, s[34:35]
; GFX9-NEXT: s_add_i32 s32, s32, 0x28000
; GFX9-NEXT: s_getpc_b64 s[34:35]
@@ -3001,28 +3001,28 @@
; GFX9-NEXT: buffer_load_dword v23, off, s[0:3], s33 offset:600
; GFX9-NEXT: buffer_load_dword v31, off, s[0:3], s33 offset:604
; GFX9-NEXT: s_waitcnt vmcnt(0)
-; GFX9-NEXT: buffer_store_dword v31, off, s[0:3], s33 offset:1568 ; 4-byte Folded Spill
+; GFX9-NEXT: buffer_store_dword v31, off, s[0:3], s33 offset:1564 ; 4-byte Folded Spill
; GFX9-NEXT: buffer_load_dword v31, off, s[0:3], s33 offset:608
; GFX9-NEXT: s_waitcnt vmcnt(0)
-; GFX9-NEXT: buffer_store_dword v31, off, s[0:3], s33 offset:1564 ; 4-byte Folded Spill
+; GFX9-NEXT: buffer_store_dword v31, off, s[0:3], s33 offset:1560 ; 4-byte Folded Spill
; GFX9-NEXT: buffer_load_dword v31, off, s[0:3], s33 offset:612
; GFX9-NEXT: s_waitcnt vmcnt(0)
-; GFX9-NEXT: buffer_store_dword v31, off, s[0:3], s33 offset:1560 ; 4-byte Folded Spill
+; GFX9-NEXT: buffer_store_dword v31, off, s[0:3], s33 offset:1556 ; 4-byte Folded Spill
; GFX9-NEXT: buffer_load_dword v31, off, s[0:3], s33 offset:616
; GFX9-NEXT: s_waitcnt vmcnt(0)
-; GFX9-NEXT: buffer_store_dword v31, off, s[0:3], s33 offset:1556 ; 4-byte Folded Spill
+; GFX9-NEXT: buffer_store_dword v31, off, s[0:3], s33 offset:1552 ; 4-byte Folded Spill
; GFX9-NEXT: buffer_load_dword v31, off, s[0:3], s33 offset:620
; GFX9-NEXT: s_waitcnt vmcnt(0)
-; GFX9-NEXT: buffer_store_dword v31, off, s[0:3], s33 offset:1552 ; 4-byte Folded Spill
+; GFX9-NEXT: buffer_store_dword v31, off, s[0:3], s33 offset:1548 ; 4-byte Folded Spill
; GFX9-NEXT: buffer_load_dword v31, off, s[0:3], s33 offset:624
; GFX9-NEXT: s_waitcnt vmcnt(0)
-; GFX9-NEXT: buffer_store_dword v31, off, s[0:3], s33 offset:1548 ; 4-byte Folded Spill
+; GFX9-NEXT: buffer_store_dword v31, off, s[0:3], s33 offset:1544 ; 4-byte Folded Spill
; GFX9-NEXT: buffer_load_dword v31, off, s[0:3], s33 offset:628
; GFX9-NEXT: s_waitcnt vmcnt(0)
-; GFX9-NEXT: buffer_store_dword v31, off, s[0:3], s33 offset:1544 ; 4-byte Folded Spill
+; GFX9-NEXT: buffer_store_dword v31, off, s[0:3], s33 offset:1540 ; 4-byte Folded Spill
; GFX9-NEXT: buffer_load_dword v31, off, s[0:3], s33 offset:632
; GFX9-NEXT: s_waitcnt vmcnt(0)
-; GFX9-NEXT: buffer_store_dword v31, off, s[0:3], s33 offset:1540 ; 4-byte Folded Spill
+; GFX9-NEXT: buffer_store_dword v31, off, s[0:3], s33 offset:1536 ; 4-byte Folded Spill
; GFX9-NEXT: buffer_store_dword v0, off, s[0:3], s32
; GFX9-NEXT: buffer_store_dword v1, off, s[0:3], s32 offset:4
; GFX9-NEXT: buffer_store_dword v32, off, s[0:3], s32 offset:8
@@ -3072,14 +3072,14 @@
; GFX9-NEXT: v_mov_b32_e32 v6, v28
; GFX9-NEXT: v_mov_b32_e32 v7, v29
; GFX9-NEXT: v_mov_b32_e32 v8, v30
-; GFX9-NEXT: buffer_load_dword v24, off, s[0:3], s33 offset:1568 ; 4-byte Folded Reload
-; GFX9-NEXT: buffer_load_dword v25, off, s[0:3], s33 offset:1564 ; 4-byte Folded Reload
-; GFX9-NEXT: buffer_load_dword v26, off, s[0:3], s33 offset:1560 ; 4-byte Folded Reload
-; GFX9-NEXT: buffer_load_dword v27, off, s[0:3], s33 offset:1556 ; 4-byte Folded Reload
-; GFX9-NEXT: buffer_load_dword v28, off, s[0:3], s33 offset:1552 ; 4-byte Folded Reload
-; GFX9-NEXT: buffer_load_dword v29, off, s[0:3], s33 offset:1548 ; 4-byte Folded Reload
-; GFX9-NEXT: buffer_load_dword v30, off, s[0:3], s33 offset:1544 ; 4-byte Folded Reload
-; GFX9-NEXT: buffer_load_dword v31, off, s[0:3], s33 offset:1540 ; 4-byte Folded Reload
+; GFX9-NEXT: buffer_load_dword v24, off, s[0:3], s33 offset:1564 ; 4-byte Folded Reload
+; GFX9-NEXT: buffer_load_dword v25, off, s[0:3], s33 offset:1560 ; 4-byte Folded Reload
+; GFX9-NEXT: buffer_load_dword v26, off, s[0:3], s33 offset:1556 ; 4-byte Folded Reload
+; GFX9-NEXT: buffer_load_dword v27, off, s[0:3], s33 offset:1552 ; 4-byte Folded Reload
+; GFX9-NEXT: buffer_load_dword v28, off, s[0:3], s33 offset:1548 ; 4-byte Folded Reload
+; GFX9-NEXT: buffer_load_dword v29, off, s[0:3], s33 offset:1544 ; 4-byte Folded Reload
+; GFX9-NEXT: buffer_load_dword v30, off, s[0:3], s33 offset:1540 ; 4-byte Folded Reload
+; GFX9-NEXT: buffer_load_dword v31, off, s[0:3], s33 offset:1536 ; 4-byte Folded Reload
; GFX9-NEXT: v_lshrrev_b32_e64 v0, 6, s33
; GFX9-NEXT: v_add_u32_e32 v0, 0x400, v0
; GFX9-NEXT: v_mov_b32_e32 v1, 42
@@ -3103,7 +3103,7 @@
; GFX9-NEXT: v_readlane_b32 s31, v33, 1
; GFX9-NEXT: v_readlane_b32 s30, v33, 0
; GFX9-NEXT: s_xor_saveexec_b64 s[34:35], -1
-; GFX9-NEXT: buffer_load_dword v33, off, s[0:3], s33 offset:1536 ; 4-byte Folded Reload
+; GFX9-NEXT: buffer_load_dword v33, off, s[0:3], s33 offset:1568 ; 4-byte Folded Reload
; GFX9-NEXT: s_mov_b64 exec, s[34:35]
; GFX9-NEXT: s_add_i32 s32, s32, 0xfffd8000
; GFX9-NEXT: s_mov_b32 s33, s36
@@ -3117,7 +3117,7 @@
; GFX10-NEXT: s_add_i32 s33, s32, 0x3fe0
; GFX10-NEXT: s_and_b32 s33, s33, 0xffffc000
; GFX10-NEXT: s_or_saveexec_b32 s34, -1
-; GFX10-NEXT: buffer_store_dword v40, off, s[0:3], s33 offset:1536 ; 4-byte Folded Spill
+; GFX10-NEXT: buffer_store_dword v40, off, s[0:3], s33 offset:1568 ; 4-byte Folded Spill
; GFX10-NEXT: s_waitcnt_depctr 0xffe3
; GFX10-NEXT: s_mov_b32 exec_lo, s34
; GFX10-NEXT: s_add_i32 s32, s32, 0x14000
@@ -3262,28 +3262,28 @@
; GFX10-NEXT: buffer_load_dword v8, off, s[0:3], s33 offset:796
; GFX10-NEXT: buffer_load_dword v0, off, s[0:3], s33 offset:516
; GFX10-NEXT: s_waitcnt vmcnt(0)
-; GFX10-NEXT: buffer_store_dword v0, off, s[0:3], s33 offset:1540 ; 4-byte Folded Spill
+; GFX10-NEXT: buffer_store_dword v0, off, s[0:3], s33 offset:1536 ; 4-byte Folded Spill
; GFX10-NEXT: buffer_load_dword v0, off, s[0:3], s33 offset:520
; GFX10-NEXT: s_waitcnt vmcnt(0)
-; GFX10-NEXT: buffer_store_dword v0, off, s[0:3], s33 offset:1544 ; 4-byte Folded Spill
+; GFX10-NEXT: buffer_store_dword v0, off, s[0:3], s33 offset:1540 ; 4-byte Folded Spill
; GFX10-NEXT: buffer_load_dword v0, off, s[0:3], s33 offset:524
; GFX10-NEXT: s_waitcnt vmcnt(0)
-; GFX10-NEXT: buffer_store_dword v0, off, s[0:3], s33 offset:1548 ; 4-byte Folded Spill
+; GFX10-NEXT: buffer_store_dword v0, off, s[0:3], s33 offset:1544 ; 4-byte Folded Spill
; GFX10-NEXT: buffer_load_dword v0, off, s[0:3], s33 offset:528
; GFX10-NEXT: s_waitcnt vmcnt(0)
-; GFX10-NEXT: buffer_store_dword v0, off, s[0:3], s33 offset:1552 ; 4-byte Folded Spill
+; GFX10-NEXT: buffer_store_dword v0, off, s[0:3], s33 offset:1548 ; 4-byte Folded Spill
; GFX10-NEXT: buffer_load_dword v0, off, s[0:3], s33 offset:532
; GFX10-NEXT: s_waitcnt vmcnt(0)
-; GFX10-NEXT: buffer_store_dword v0, off, s[0:3], s33 offset:1556 ; 4-byte Folded Spill
+; GFX10-NEXT: buffer_store_dword v0, off, s[0:3], s33 offset:1552 ; 4-byte Folded Spill
; GFX10-NEXT: buffer_load_dword v0, off, s[0:3], s33 offset:536
; GFX10-NEXT: s_waitcnt vmcnt(0)
-; GFX10-NEXT: buffer_store_dword v0, off, s[0:3], s33 offset:1560 ; 4-byte Folded Spill
+; GFX10-NEXT: buffer_store_dword v0, off, s[0:3], s33 offset:1556 ; 4-byte Folded Spill
; GFX10-NEXT: buffer_load_dword v0, off, s[0:3], s33 offset:540
; GFX10-NEXT: s_waitcnt vmcnt(0)
-; GFX10-NEXT: buffer_store_dword v0, off, s[0:3], s33 offset:1564 ; 4-byte Folded Spill
+; GFX10-NEXT: buffer_store_dword v0, off, s[0:3], s33 offset:1560 ; 4-byte Folded Spill
; GFX10-NEXT: buffer_load_dword v0, off, s[0:3], s33 offset:544
; GFX10-NEXT: s_waitcnt vmcnt(0)
-; GFX10-NEXT: buffer_store_dword v0, off, s[0:3], s33 offset:1568 ; 4-byte Folded Spill
+; GFX10-NEXT: buffer_store_dword v0, off, s[0:3], s33 offset:1564 ; 4-byte Folded Spill
; GFX10-NEXT: s_clause 0x15
; GFX10-NEXT: buffer_load_dword v10, off, s[0:3], s33 offset:548
; GFX10-NEXT: buffer_load_dword v11, off, s[0:3], s33 offset:552
@@ -3350,14 +3350,14 @@
; GFX10-NEXT: buffer_store_dword v7, off, s[0:3], s32 offset:156
; GFX10-NEXT: buffer_store_dword v8, off, s[0:3], s32 offset:160
; GFX10-NEXT: s_clause 0x7
-; GFX10-NEXT: buffer_load_dword v2, off, s[0:3], s33 offset:1540
-; GFX10-NEXT: buffer_load_dword v3, off, s[0:3], s33 offset:1544
-; GFX10-NEXT: buffer_load_dword v4, off, s[0:3], s33 offset:1548
-; GFX10-NEXT: buffer_load_dword v5, off, s[0:3], s33 offset:1552
-; GFX10-NEXT: buffer_load_dword v6, off, s[0:3], s33 offset:1556
-; GFX10-NEXT: buffer_load_dword v7, off, s[0:3], s33 offset:1560
-; GFX10-NEXT: buffer_load_dword v8, off, s[0:3], s33 offset:1564
-; GFX10-NEXT: buffer_load_dword v9, off, s[0:3], s33 offset:1568
+; GFX10-NEXT: buffer_load_dword v2, off, s[0:3], s33 offset:1536
+; GFX10-NEXT: buffer_load_dword v3, off, s[0:3], s33 offset:1540
+; GFX10-NEXT: buffer_load_dword v4, off, s[0:3], s33 offset:1544
+; GFX10-NEXT: buffer_load_dword v5, off, s[0:3], s33 offset:1548
+; GFX10-NEXT: buffer_load_dword v6, off, s[0:3], s33 offset:1552
+; GFX10-NEXT: buffer_load_dword v7, off, s[0:3], s33 offset:1556
+; GFX10-NEXT: buffer_load_dword v8, off, s[0:3], s33 offset:1560
+; GFX10-NEXT: buffer_load_dword v9, off, s[0:3], s33 offset:1564
; GFX10-NEXT: v_lshrrev_b32_e64 v0, 5, s33
; GFX10-NEXT: v_mov_b32_e32 v1, 42
; GFX10-NEXT: v_add_nc_u32_e32 v0, 0x400, v0
@@ -3381,7 +3381,7 @@
; GFX10-NEXT: v_readlane_b32 s31, v40, 1
; GFX10-NEXT: v_readlane_b32 s30, v40, 0
; GFX10-NEXT: s_or_saveexec_b32 s34, -1
-; GFX10-NEXT: buffer_load_dword v40, off, s[0:3], s33 offset:1536 ; 4-byte Folded Reload
+; GFX10-NEXT: buffer_load_dword v40, off, s[0:3], s33 offset:1568 ; 4-byte Folded Reload
; GFX10-NEXT: s_waitcnt_depctr 0xffe3
; GFX10-NEXT: s_mov_b32 exec_lo, s34
; GFX10-NEXT: s_add_i32 s32, s32, 0xfffec000
@@ -3397,7 +3397,7 @@
; GFX11-NEXT: s_delay_alu instid0(SALU_CYCLE_1)
; GFX11-NEXT: s_and_b32 s33, s33, 0xfffffe00
; GFX11-NEXT: s_or_saveexec_b32 s0, -1
-; GFX11-NEXT: scratch_store_b32 off, v40, s33 offset:1536 ; 4-byte Folded Spill
+; GFX11-NEXT: scratch_store_b32 off, v40, s33 offset:1600 ; 4-byte Folded Spill
; GFX11-NEXT: s_mov_b32 exec_lo, s0
; GFX11-NEXT: s_mov_b32 s0, 0
; GFX11-NEXT: v_mov_b32_e32 v4, 0
@@ -3495,7 +3495,7 @@
; GFX11-NEXT: s_waitcnt vmcnt(4)
; GFX11-NEXT: v_dual_mov_b32 v44, v62 :: v_dual_mov_b32 v57, v16
; GFX11-NEXT: s_waitcnt vmcnt(0)
-; GFX11-NEXT: scratch_store_b128 off, v[12:15], s33 offset:1588 ; 16-byte Folded Spill
+; GFX11-NEXT: scratch_store_b128 off, v[12:15], s33 offset:1584 ; 16-byte Folded Spill
; GFX11-NEXT: s_clause 0x3
; GFX11-NEXT: scratch_load_b128 v[12:15], off, s33 offset:528
; GFX11-NEXT: scratch_load_b128 v[20:23], off, s33 offset:544
@@ -3511,13 +3511,13 @@
; GFX11-NEXT: s_waitcnt vmcnt(1)
; GFX11-NEXT: v_dual_mov_b32 v10, v21 :: v_dual_mov_b32 v15, v26
; GFX11-NEXT: s_waitcnt vmcnt(0)
-; GFX11-NEXT: scratch_store_b128 off, v[28:31], s33 offset:1572 ; 16-byte Folded Spill
+; GFX11-NEXT: scratch_store_b128 off, v[28:31], s33 offset:1568 ; 16-byte Folded Spill
; GFX11-NEXT: scratch_load_b128 v[28:31], off, s33 offset:592
; GFX11-NEXT: s_waitcnt vmcnt(0)
-; GFX11-NEXT: scratch_store_b128 off, v[28:31], s33 offset:1556 ; 16-byte Folded Spill
+; GFX11-NEXT: scratch_store_b128 off, v[28:31], s33 offset:1552 ; 16-byte Folded Spill
; GFX11-NEXT: scratch_load_b128 v[28:31], off, s33 offset:608
; GFX11-NEXT: s_waitcnt vmcnt(0)
-; GFX11-NEXT: scratch_store_b128 off, v[28:31], s33 offset:1540 ; 16-byte Folded Spill
+; GFX11-NEXT: scratch_store_b128 off, v[28:31], s33 offset:1536 ; 16-byte Folded Spill
; GFX11-NEXT: scratch_store_b128 off, v[32:35], s32
; GFX11-NEXT: v_mov_b32_e32 v32, v36
; GFX11-NEXT: v_dual_mov_b32 v33, v48 :: v_dual_mov_b32 v34, v49
@@ -3560,13 +3560,13 @@
; GFX11-NEXT: scratch_store_b128 off, v[36:39], s0
; GFX11-NEXT: s_add_i32 s0, s32, 16
; GFX11-NEXT: scratch_store_b128 off, v[32:35], s0
-; GFX11-NEXT: scratch_load_b128 v[1:4], off, s33 offset:1588 ; 16-byte Folded Reload
+; GFX11-NEXT: scratch_load_b128 v[1:4], off, s33 offset:1584 ; 16-byte Folded Reload
; GFX11-NEXT: s_waitcnt vmcnt(0)
; GFX11-NEXT: v_mov_b32_e32 v1, 42
; GFX11-NEXT: s_clause 0x2
-; GFX11-NEXT: scratch_load_b128 v[17:20], off, s33 offset:1572
-; GFX11-NEXT: scratch_load_b128 v[21:24], off, s33 offset:1556
-; GFX11-NEXT: scratch_load_b128 v[25:28], off, s33 offset:1540
+; GFX11-NEXT: scratch_load_b128 v[17:20], off, s33 offset:1568
+; GFX11-NEXT: scratch_load_b128 v[21:24], off, s33 offset:1552
+; GFX11-NEXT: scratch_load_b128 v[25:28], off, s33 offset:1536
; GFX11-NEXT: s_add_i32 s0, s33, 0x400
; GFX11-NEXT: s_delay_alu instid0(SALU_CYCLE_1)
; GFX11-NEXT: v_mov_b32_e32 v0, s0
@@ -3590,7 +3590,7 @@
; GFX11-NEXT: v_readlane_b32 s31, v40, 1
; GFX11-NEXT: v_readlane_b32 s30, v40, 0
; GFX11-NEXT: s_or_saveexec_b32 s0, -1
-; GFX11-NEXT: scratch_load_b32 v40, off, s33 offset:1536 ; 4-byte Folded Reload
+; GFX11-NEXT: scratch_load_b32 v40, off, s33 offset:1600 ; 4-byte Folded Reload
; GFX11-NEXT: s_mov_b32 exec_lo, s0
; GFX11-NEXT: s_addk_i32 s32, 0xf600
; GFX11-NEXT: s_mov_b32 s33, s45
diff --git a/llvm/test/CodeGen/AMDGPU/indirect-call.ll b/llvm/test/CodeGen/AMDGPU/indirect-call.ll
--- a/llvm/test/CodeGen/AMDGPU/indirect-call.ll
+++ b/llvm/test/CodeGen/AMDGPU/indirect-call.ll
@@ -397,8 +397,9 @@
; GCN-NEXT: s_mov_b32 s33, s32
; GCN-NEXT: s_or_saveexec_b64 s[18:19], -1
; GCN-NEXT: buffer_store_dword v40, off, s[0:3], s33 ; 4-byte Folded Spill
+; GCN-NEXT: buffer_store_dword v41, off, s[0:3], s33 offset:4 ; 4-byte Folded Spill
; GCN-NEXT: s_mov_b64 exec, s[18:19]
-; GCN-NEXT: v_writelane_b32 v40, s16, 18
+; GCN-NEXT: v_writelane_b32 v41, s16, 0
; GCN-NEXT: s_addk_i32 s32, 0x400
; GCN-NEXT: v_writelane_b32 v40, s30, 0
; GCN-NEXT: v_writelane_b32 v40, s31, 1
@@ -465,9 +466,10 @@
; GCN-NEXT: v_readlane_b32 s34, v40, 2
; GCN-NEXT: v_readlane_b32 s31, v40, 1
; GCN-NEXT: v_readlane_b32 s30, v40, 0
-; GCN-NEXT: v_readlane_b32 s4, v40, 18
+; GCN-NEXT: v_readlane_b32 s4, v41, 0
; GCN-NEXT: s_or_saveexec_b64 s[6:7], -1
; GCN-NEXT: buffer_load_dword v40, off, s[0:3], s33 ; 4-byte Folded Reload
+; GCN-NEXT: buffer_load_dword v41, off, s[0:3], s33 offset:4 ; 4-byte Folded Reload
; GCN-NEXT: s_mov_b64 exec, s[6:7]
; GCN-NEXT: s_addk_i32 s32, 0xfc00
; GCN-NEXT: s_mov_b32 s33, s4
@@ -481,8 +483,9 @@
; GISEL-NEXT: s_mov_b32 s33, s32
; GISEL-NEXT: s_or_saveexec_b64 s[18:19], -1
; GISEL-NEXT: buffer_store_dword v40, off, s[0:3], s33 ; 4-byte Folded Spill
+; GISEL-NEXT: buffer_store_dword v41, off, s[0:3], s33 offset:4 ; 4-byte Folded Spill
; GISEL-NEXT: s_mov_b64 exec, s[18:19]
-; GISEL-NEXT: v_writelane_b32 v40, s16, 18
+; GISEL-NEXT: v_writelane_b32 v41, s16, 0
; GISEL-NEXT: s_addk_i32 s32, 0x400
; GISEL-NEXT: v_writelane_b32 v40, s30, 0
; GISEL-NEXT: v_writelane_b32 v40, s31, 1
@@ -549,9 +552,10 @@
; GISEL-NEXT: v_readlane_b32 s34, v40, 2
; GISEL-NEXT: v_readlane_b32 s31, v40, 1
; GISEL-NEXT: v_readlane_b32 s30, v40, 0
-; GISEL-NEXT: v_readlane_b32 s4, v40, 18
+; GISEL-NEXT: v_readlane_b32 s4, v41, 0
; GISEL-NEXT: s_or_saveexec_b64 s[6:7], -1
; GISEL-NEXT: buffer_load_dword v40, off, s[0:3], s33 ; 4-byte Folded Reload
+; GISEL-NEXT: buffer_load_dword v41, off, s[0:3], s33 offset:4 ; 4-byte Folded Reload
; GISEL-NEXT: s_mov_b64 exec, s[6:7]
; GISEL-NEXT: s_addk_i32 s32, 0xfc00
; GISEL-NEXT: s_mov_b32 s33, s4
@@ -569,8 +573,9 @@
; GCN-NEXT: s_mov_b32 s33, s32
; GCN-NEXT: s_or_saveexec_b64 s[18:19], -1
; GCN-NEXT: buffer_store_dword v40, off, s[0:3], s33 ; 4-byte Folded Spill
+; GCN-NEXT: buffer_store_dword v41, off, s[0:3], s33 offset:4 ; 4-byte Folded Spill
; GCN-NEXT: s_mov_b64 exec, s[18:19]
-; GCN-NEXT: v_writelane_b32 v40, s16, 18
+; GCN-NEXT: v_writelane_b32 v41, s16, 0
; GCN-NEXT: s_addk_i32 s32, 0x400
; GCN-NEXT: v_writelane_b32 v40, s30, 0
; GCN-NEXT: v_writelane_b32 v40, s31, 1
@@ -640,9 +645,10 @@
; GCN-NEXT: v_readlane_b32 s34, v40, 2
; GCN-NEXT: v_readlane_b32 s31, v40, 1
; GCN-NEXT: v_readlane_b32 s30, v40, 0
-; GCN-NEXT: v_readlane_b32 s4, v40, 18
+; GCN-NEXT: v_readlane_b32 s4, v41, 0
; GCN-NEXT: s_or_saveexec_b64 s[6:7], -1
; GCN-NEXT: buffer_load_dword v40, off, s[0:3], s33 ; 4-byte Folded Reload
+; GCN-NEXT: buffer_load_dword v41, off, s[0:3], s33 offset:4 ; 4-byte Folded Reload
; GCN-NEXT: s_mov_b64 exec, s[6:7]
; GCN-NEXT: s_addk_i32 s32, 0xfc00
; GCN-NEXT: s_mov_b32 s33, s4
@@ -656,8 +662,9 @@
; GISEL-NEXT: s_mov_b32 s33, s32
; GISEL-NEXT: s_or_saveexec_b64 s[18:19], -1
; GISEL-NEXT: buffer_store_dword v40, off, s[0:3], s33 ; 4-byte Folded Spill
+; GISEL-NEXT: buffer_store_dword v41, off, s[0:3], s33 offset:4 ; 4-byte Folded Spill
; GISEL-NEXT: s_mov_b64 exec, s[18:19]
-; GISEL-NEXT: v_writelane_b32 v40, s16, 18
+; GISEL-NEXT: v_writelane_b32 v41, s16, 0
; GISEL-NEXT: s_addk_i32 s32, 0x400
; GISEL-NEXT: v_writelane_b32 v40, s30, 0
; GISEL-NEXT: v_writelane_b32 v40, s31, 1
@@ -725,9 +732,10 @@
; GISEL-NEXT: v_readlane_b32 s34, v40, 2
; GISEL-NEXT: v_readlane_b32 s31, v40, 1
; GISEL-NEXT: v_readlane_b32 s30, v40, 0
-; GISEL-NEXT: v_readlane_b32 s4, v40, 18
+; GISEL-NEXT: v_readlane_b32 s4, v41, 0
; GISEL-NEXT: s_or_saveexec_b64 s[6:7], -1
; GISEL-NEXT: buffer_load_dword v40, off, s[0:3], s33 ; 4-byte Folded Reload
+; GISEL-NEXT: buffer_load_dword v41, off, s[0:3], s33 offset:4 ; 4-byte Folded Reload
; GISEL-NEXT: s_mov_b64 exec, s[6:7]
; GISEL-NEXT: s_addk_i32 s32, 0xfc00
; GISEL-NEXT: s_mov_b32 s33, s4
@@ -745,8 +753,9 @@
; GCN-NEXT: s_mov_b32 s33, s32
; GCN-NEXT: s_or_saveexec_b64 s[18:19], -1
; GCN-NEXT: buffer_store_dword v40, off, s[0:3], s33 ; 4-byte Folded Spill
+; GCN-NEXT: buffer_store_dword v41, off, s[0:3], s33 offset:4 ; 4-byte Folded Spill
; GCN-NEXT: s_mov_b64 exec, s[18:19]
-; GCN-NEXT: v_writelane_b32 v40, s16, 18
+; GCN-NEXT: v_writelane_b32 v41, s16, 0
; GCN-NEXT: s_addk_i32 s32, 0x400
; GCN-NEXT: v_writelane_b32 v40, s30, 0
; GCN-NEXT: v_writelane_b32 v40, s31, 1
@@ -815,9 +824,10 @@
; GCN-NEXT: v_readlane_b32 s34, v40, 2
; GCN-NEXT: v_readlane_b32 s31, v40, 1
; GCN-NEXT: v_readlane_b32 s30, v40, 0
-; GCN-NEXT: v_readlane_b32 s4, v40, 18
+; GCN-NEXT: v_readlane_b32 s4, v41, 0
; GCN-NEXT: s_or_saveexec_b64 s[6:7], -1
; GCN-NEXT: buffer_load_dword v40, off, s[0:3], s33 ; 4-byte Folded Reload
+; GCN-NEXT: buffer_load_dword v41, off, s[0:3], s33 offset:4 ; 4-byte Folded Reload
; GCN-NEXT: s_mov_b64 exec, s[6:7]
; GCN-NEXT: s_addk_i32 s32, 0xfc00
; GCN-NEXT: s_mov_b32 s33, s4
@@ -831,8 +841,9 @@
; GISEL-NEXT: s_mov_b32 s33, s32
; GISEL-NEXT: s_or_saveexec_b64 s[18:19], -1
; GISEL-NEXT: buffer_store_dword v40, off, s[0:3], s33 ; 4-byte Folded Spill
+; GISEL-NEXT: buffer_store_dword v41, off, s[0:3], s33 offset:4 ; 4-byte Folded Spill
; GISEL-NEXT: s_mov_b64 exec, s[18:19]
-; GISEL-NEXT: v_writelane_b32 v40, s16, 18
+; GISEL-NEXT: v_writelane_b32 v41, s16, 0
; GISEL-NEXT: s_addk_i32 s32, 0x400
; GISEL-NEXT: v_writelane_b32 v40, s30, 0
; GISEL-NEXT: v_writelane_b32 v40, s31, 1
@@ -901,9 +912,10 @@
; GISEL-NEXT: v_readlane_b32 s34, v40, 2
; GISEL-NEXT: v_readlane_b32 s31, v40, 1
; GISEL-NEXT: v_readlane_b32 s30, v40, 0
-; GISEL-NEXT: v_readlane_b32 s4, v40, 18
+; GISEL-NEXT: v_readlane_b32 s4, v41, 0
; GISEL-NEXT: s_or_saveexec_b64 s[6:7], -1
; GISEL-NEXT: buffer_load_dword v40, off, s[0:3], s33 ; 4-byte Folded Reload
+; GISEL-NEXT: buffer_load_dword v41, off, s[0:3], s33 offset:4 ; 4-byte Folded Reload
; GISEL-NEXT: s_mov_b64 exec, s[6:7]
; GISEL-NEXT: s_addk_i32 s32, 0xfc00
; GISEL-NEXT: s_mov_b32 s33, s4
@@ -922,8 +934,9 @@
; GCN-NEXT: s_mov_b32 s33, s32
; GCN-NEXT: s_or_saveexec_b64 s[18:19], -1
; GCN-NEXT: buffer_store_dword v40, off, s[0:3], s33 ; 4-byte Folded Spill
+; GCN-NEXT: buffer_store_dword v41, off, s[0:3], s33 offset:4 ; 4-byte Folded Spill
; GCN-NEXT: s_mov_b64 exec, s[18:19]
-; GCN-NEXT: v_writelane_b32 v40, s16, 20
+; GCN-NEXT: v_writelane_b32 v41, s16, 0
; GCN-NEXT: s_addk_i32 s32, 0x400
; GCN-NEXT: v_writelane_b32 v40, s30, 0
; GCN-NEXT: v_writelane_b32 v40, s31, 1
@@ -1001,9 +1014,10 @@
; GCN-NEXT: v_readlane_b32 s34, v40, 2
; GCN-NEXT: v_readlane_b32 s31, v40, 1
; GCN-NEXT: v_readlane_b32 s30, v40, 0
-; GCN-NEXT: v_readlane_b32 s4, v40, 20
+; GCN-NEXT: v_readlane_b32 s4, v41, 0
; GCN-NEXT: s_or_saveexec_b64 s[6:7], -1
; GCN-NEXT: buffer_load_dword v40, off, s[0:3], s33 ; 4-byte Folded Reload
+; GCN-NEXT: buffer_load_dword v41, off, s[0:3], s33 offset:4 ; 4-byte Folded Reload
; GCN-NEXT: s_mov_b64 exec, s[6:7]
; GCN-NEXT: s_addk_i32 s32, 0xfc00
; GCN-NEXT: s_mov_b32 s33, s4
@@ -1017,8 +1031,9 @@
; GISEL-NEXT: s_mov_b32 s33, s32
; GISEL-NEXT: s_or_saveexec_b64 s[18:19], -1
; GISEL-NEXT: buffer_store_dword v40, off, s[0:3], s33 ; 4-byte Folded Spill
+; GISEL-NEXT: buffer_store_dword v41, off, s[0:3], s33 offset:4 ; 4-byte Folded Spill
; GISEL-NEXT: s_mov_b64 exec, s[18:19]
-; GISEL-NEXT: v_writelane_b32 v40, s16, 20
+; GISEL-NEXT: v_writelane_b32 v41, s16, 0
; GISEL-NEXT: s_addk_i32 s32, 0x400
; GISEL-NEXT: v_writelane_b32 v40, s30, 0
; GISEL-NEXT: v_writelane_b32 v40, s31, 1
@@ -1096,9 +1111,10 @@
; GISEL-NEXT: v_readlane_b32 s34, v40, 2
; GISEL-NEXT: v_readlane_b32 s31, v40, 1
; GISEL-NEXT: v_readlane_b32 s30, v40, 0
-; GISEL-NEXT: v_readlane_b32 s4, v40, 20
+; GISEL-NEXT: v_readlane_b32 s4, v41, 0
; GISEL-NEXT: s_or_saveexec_b64 s[6:7], -1
; GISEL-NEXT: buffer_load_dword v40, off, s[0:3], s33 ; 4-byte Folded Reload
+; GISEL-NEXT: buffer_load_dword v41, off, s[0:3], s33 offset:4 ; 4-byte Folded Reload
; GISEL-NEXT: s_mov_b64 exec, s[6:7]
; GISEL-NEXT: s_addk_i32 s32, 0xfc00
; GISEL-NEXT: s_mov_b32 s33, s4
@@ -1311,7 +1327,7 @@
; GCN-LABEL: test_indirect_call_vgpr_ptr_arg_and_reuse:
; GCN: ; %bb.0:
; GCN-NEXT: s_waitcnt vmcnt(0) expcnt(0) lgkmcnt(0)
-; GCN-NEXT: s_mov_b32 s10, s33
+; GCN-NEXT: s_mov_b32 s12, s33
; GCN-NEXT: s_mov_b32 s33, s32
; GCN-NEXT: s_or_saveexec_b64 s[4:5], -1
; GCN-NEXT: buffer_store_dword v40, off, s[0:3], s33 offset:4 ; 4-byte Folded Spill
@@ -1402,14 +1418,14 @@
; GCN-NEXT: buffer_load_dword v40, off, s[0:3], s33 offset:4 ; 4-byte Folded Reload
; GCN-NEXT: s_mov_b64 exec, s[4:5]
; GCN-NEXT: s_addk_i32 s32, 0xfc00
-; GCN-NEXT: s_mov_b32 s33, s10
+; GCN-NEXT: s_mov_b32 s33, s12
; GCN-NEXT: s_waitcnt vmcnt(0)
; GCN-NEXT: s_setpc_b64 s[30:31]
;
; GISEL-LABEL: test_indirect_call_vgpr_ptr_arg_and_reuse:
; GISEL: ; %bb.0:
; GISEL-NEXT: s_waitcnt vmcnt(0) expcnt(0) lgkmcnt(0)
-; GISEL-NEXT: s_mov_b32 s10, s33
+; GISEL-NEXT: s_mov_b32 s12, s33
; GISEL-NEXT: s_mov_b32 s33, s32
; GISEL-NEXT: s_or_saveexec_b64 s[4:5], -1
; GISEL-NEXT: buffer_store_dword v40, off, s[0:3], s33 offset:4 ; 4-byte Folded Spill
@@ -1500,7 +1516,7 @@
; GISEL-NEXT: buffer_load_dword v40, off, s[0:3], s33 offset:4 ; 4-byte Folded Reload
; GISEL-NEXT: s_mov_b64 exec, s[4:5]
; GISEL-NEXT: s_addk_i32 s32, 0xfc00
-; GISEL-NEXT: s_mov_b32 s33, s10
+; GISEL-NEXT: s_mov_b32 s33, s12
; GISEL-NEXT: s_waitcnt vmcnt(0)
; GISEL-NEXT: s_setpc_b64 s[30:31]
call amdgpu_gfx void %fptr(i32 %i)
@@ -1515,7 +1531,7 @@
; GCN-LABEL: test_indirect_call_vgpr_ptr_arg_and_return:
; GCN: ; %bb.0:
; GCN-NEXT: s_waitcnt vmcnt(0) expcnt(0) lgkmcnt(0)
-; GCN-NEXT: s_mov_b32 s10, s33
+; GCN-NEXT: s_mov_b32 s12, s33
; GCN-NEXT: s_mov_b32 s33, s32
; GCN-NEXT: s_or_saveexec_b64 s[4:5], -1
; GCN-NEXT: buffer_store_dword v40, off, s[0:3], s33 ; 4-byte Folded Spill
@@ -1604,14 +1620,14 @@
; GCN-NEXT: buffer_load_dword v40, off, s[0:3], s33 ; 4-byte Folded Reload
; GCN-NEXT: s_mov_b64 exec, s[4:5]
; GCN-NEXT: s_addk_i32 s32, 0xfc00
-; GCN-NEXT: s_mov_b32 s33, s10
+; GCN-NEXT: s_mov_b32 s33, s12
; GCN-NEXT: s_waitcnt vmcnt(0)
; GCN-NEXT: s_setpc_b64 s[30:31]
;
; GISEL-LABEL: test_indirect_call_vgpr_ptr_arg_and_return:
; GISEL: ; %bb.0:
; GISEL-NEXT: s_waitcnt vmcnt(0) expcnt(0) lgkmcnt(0)
-; GISEL-NEXT: s_mov_b32 s10, s33
+; GISEL-NEXT: s_mov_b32 s12, s33
; GISEL-NEXT: s_mov_b32 s33, s32
; GISEL-NEXT: s_or_saveexec_b64 s[4:5], -1
; GISEL-NEXT: buffer_store_dword v40, off, s[0:3], s33 ; 4-byte Folded Spill
@@ -1700,7 +1716,7 @@
; GISEL-NEXT: buffer_load_dword v40, off, s[0:3], s33 ; 4-byte Folded Reload
; GISEL-NEXT: s_mov_b64 exec, s[4:5]
; GISEL-NEXT: s_addk_i32 s32, 0xfc00
-; GISEL-NEXT: s_mov_b32 s33, s10
+; GISEL-NEXT: s_mov_b32 s33, s12
; GISEL-NEXT: s_waitcnt vmcnt(0)
; GISEL-NEXT: s_setpc_b64 s[30:31]
%ret = call amdgpu_gfx i32 %fptr(i32 %i)
@@ -1712,7 +1728,7 @@
; GCN-LABEL: test_indirect_tail_call_vgpr_ptr:
; GCN: ; %bb.0:
; GCN-NEXT: s_waitcnt vmcnt(0) expcnt(0) lgkmcnt(0)
-; GCN-NEXT: s_mov_b32 s10, s33
+; GCN-NEXT: s_mov_b32 s12, s33
; GCN-NEXT: s_mov_b32 s33, s32
; GCN-NEXT: s_or_saveexec_b64 s[4:5], -1
; GCN-NEXT: buffer_store_dword v40, off, s[0:3], s33 ; 4-byte Folded Spill
@@ -1798,14 +1814,14 @@
; GCN-NEXT: buffer_load_dword v40, off, s[0:3], s33 ; 4-byte Folded Reload
; GCN-NEXT: s_mov_b64 exec, s[4:5]
; GCN-NEXT: s_addk_i32 s32, 0xfc00
-; GCN-NEXT: s_mov_b32 s33, s10
+; GCN-NEXT: s_mov_b32 s33, s12
; GCN-NEXT: s_waitcnt vmcnt(0)
; GCN-NEXT: s_setpc_b64 s[30:31]
;
; GISEL-LABEL: test_indirect_tail_call_vgpr_ptr:
; GISEL: ; %bb.0:
; GISEL-NEXT: s_waitcnt vmcnt(0) expcnt(0) lgkmcnt(0)
-; GISEL-NEXT: s_mov_b32 s10, s33
+; GISEL-NEXT: s_mov_b32 s12, s33
; GISEL-NEXT: s_mov_b32 s33, s32
; GISEL-NEXT: s_or_saveexec_b64 s[4:5], -1
; GISEL-NEXT: buffer_store_dword v40, off, s[0:3], s33 ; 4-byte Folded Spill
@@ -1891,7 +1907,7 @@
; GISEL-NEXT: buffer_load_dword v40, off, s[0:3], s33 ; 4-byte Folded Reload
; GISEL-NEXT: s_mov_b64 exec, s[4:5]
; GISEL-NEXT: s_addk_i32 s32, 0xfc00
-; GISEL-NEXT: s_mov_b32 s33, s10
+; GISEL-NEXT: s_mov_b32 s33, s12
; GISEL-NEXT: s_waitcnt vmcnt(0)
; GISEL-NEXT: s_setpc_b64 s[30:31]
tail call amdgpu_gfx void %fptr()
diff --git a/llvm/test/CodeGen/AMDGPU/insert-delay-alu-bug.ll b/llvm/test/CodeGen/AMDGPU/insert-delay-alu-bug.ll
--- a/llvm/test/CodeGen/AMDGPU/insert-delay-alu-bug.ll
+++ b/llvm/test/CodeGen/AMDGPU/insert-delay-alu-bug.ll
@@ -7,7 +7,7 @@
; GFX11-LABEL: f0:
; GFX11: ; %bb.0: ; %bb
; GFX11-NEXT: s_waitcnt vmcnt(0) expcnt(0) lgkmcnt(0)
-; GFX11-NEXT: s_mov_b32 s2, s33
+; GFX11-NEXT: s_mov_b32 s3, s33
; GFX11-NEXT: s_mov_b32 s33, s32
; GFX11-NEXT: s_xor_saveexec_b32 s0, -1
; GFX11-NEXT: scratch_store_b32 off, v4, s33 ; 4-byte Folded Spill
@@ -28,7 +28,7 @@
; GFX11-NEXT: scratch_load_b32 v4, off, s33 ; 4-byte Folded Reload
; GFX11-NEXT: s_mov_b32 exec_lo, s0
; GFX11-NEXT: s_add_i32 s32, s32, -16
-; GFX11-NEXT: s_mov_b32 s33, s2
+; GFX11-NEXT: s_mov_b32 s33, s3
; GFX11-NEXT: s_waitcnt vmcnt(0)
; GFX11-NEXT: s_setpc_b64 s[30:31]
bb:
@@ -53,20 +53,20 @@
; GFX11-LABEL: f2:
; GFX11: ; %bb.0: ; %bb
; GFX11-NEXT: s_mov_b64 s[16:17], s[4:5]
+; GFX11-NEXT: s_mov_b64 s[10:11], s[6:7]
+; GFX11-NEXT: s_mov_b64 s[6:7], s[2:3]
+; GFX11-NEXT: s_load_b32 s2, s[16:17], 0x24
; GFX11-NEXT: v_mov_b32_e32 v31, v0
-; GFX11-NEXT: s_load_b32 s24, s[16:17], 0x24
; GFX11-NEXT: s_mov_b32 s18, s14
; GFX11-NEXT: s_mov_b32 s12, s13
-; GFX11-NEXT: s_mov_b64 s[10:11], s[6:7]
-; GFX11-NEXT: v_and_b32_e32 v0, 0x3ff, v31
-; GFX11-NEXT: s_mov_b64 s[6:7], s[2:3]
; GFX11-NEXT: s_mov_b64 s[4:5], s[0:1]
-; GFX11-NEXT: s_mov_b32 s3, 0
+; GFX11-NEXT: s_mov_b32 s20, 0
+; GFX11-NEXT: v_and_b32_e32 v0, 0x3ff, v31
; GFX11-NEXT: s_mov_b32 s0, -1
; GFX11-NEXT: s_mov_b32 s19, exec_lo
; GFX11-NEXT: s_mov_b32 s32, 0
; GFX11-NEXT: s_waitcnt lgkmcnt(0)
-; GFX11-NEXT: v_mul_lo_u32 v0, s24, v0
+; GFX11-NEXT: v_mul_lo_u32 v0, s2, v0
; GFX11-NEXT: s_delay_alu instid0(VALU_DEP_1)
; GFX11-NEXT: v_cmpx_eq_u32_e32 0, v0
; GFX11-NEXT: s_cbranch_execz .LBB2_13
@@ -74,7 +74,7 @@
; GFX11-NEXT: s_load_b128 s[20:23], s[16:17], 0x2c
; GFX11-NEXT: s_waitcnt lgkmcnt(0)
; GFX11-NEXT: s_bitcmp1_b32 s21, 0
-; GFX11-NEXT: s_cselect_b32 s25, -1, 0
+; GFX11-NEXT: s_cselect_b32 s24, -1, 0
; GFX11-NEXT: s_bitcmp0_b32 s21, 0
; GFX11-NEXT: s_mov_b32 s21, 0
; GFX11-NEXT: s_cbranch_scc0 .LBB2_3
@@ -90,41 +90,40 @@
; GFX11-NEXT: s_waitcnt lgkmcnt(0)
; GFX11-NEXT: s_swappc_b64 s[30:31], s[0:1]
; GFX11-NEXT: s_mov_b32 s1, -1
-; GFX11-NEXT: s_and_not1_b32 vcc_lo, exec_lo, s3
-; GFX11-NEXT: s_cbranch_vccz .LBB2_4
+; GFX11-NEXT: s_cbranch_execz .LBB2_4
; GFX11-NEXT: s_branch .LBB2_12
; GFX11-NEXT: .LBB2_3:
; GFX11-NEXT: s_mov_b32 s1, 0
; GFX11-NEXT: .LBB2_4: ; %bb16
-; GFX11-NEXT: s_load_b32 s2, s[16:17], 0x54
+; GFX11-NEXT: s_load_b32 s3, s[16:17], 0x54
; GFX11-NEXT: s_bitcmp1_b32 s23, 0
; GFX11-NEXT: s_cselect_b32 s0, -1, 0
-; GFX11-NEXT: s_and_b32 s3, s23, 1
+; GFX11-NEXT: s_and_b32 s9, s23, 1
; GFX11-NEXT: s_waitcnt lgkmcnt(0)
-; GFX11-NEXT: s_bitcmp1_b32 s2, 0
-; GFX11-NEXT: s_mov_b32 s2, -1
+; GFX11-NEXT: s_bitcmp1_b32 s3, 0
+; GFX11-NEXT: s_mov_b32 s3, -1
; GFX11-NEXT: s_cselect_b32 s8, -1, 0
-; GFX11-NEXT: s_cmp_eq_u32 s3, 0
+; GFX11-NEXT: s_cmp_eq_u32 s9, 0
; GFX11-NEXT: s_cbranch_scc0 .LBB2_8
; GFX11-NEXT: ; %bb.5: ; %bb18.preheader
; GFX11-NEXT: s_load_b128 s[28:31], s[16:17], 0x44
; GFX11-NEXT: v_mov_b32_e32 v1, 0
; GFX11-NEXT: s_waitcnt lgkmcnt(0)
-; GFX11-NEXT: s_mul_hi_u32 s2, s29, s28
-; GFX11-NEXT: s_mul_i32 s3, s29, s28
-; GFX11-NEXT: s_delay_alu instid0(SALU_CYCLE_1) | instskip(SKIP_1) | instid1(VALU_DEP_1)
-; GFX11-NEXT: v_alignbit_b32 v0, s2, s3, 1
-; GFX11-NEXT: s_mov_b32 s3, 0
-; GFX11-NEXT: v_readfirstlane_b32 s2, v0
-; GFX11-NEXT: v_cndmask_b32_e64 v0, 0, 1, s25
+; GFX11-NEXT: s_mul_hi_u32 s3, s29, s28
+; GFX11-NEXT: s_mul_i32 s9, s29, s28
+; GFX11-NEXT: s_delay_alu instid0(SALU_CYCLE_1) | instskip(NEXT) | instid1(VALU_DEP_1)
+; GFX11-NEXT: v_alignbit_b32 v0, s3, s9, 1
+; GFX11-NEXT: v_readfirstlane_b32 s3, v0
+; GFX11-NEXT: v_cndmask_b32_e64 v0, 0, 1, s24
; GFX11-NEXT: s_delay_alu instid0(VALU_DEP_2) | instskip(NEXT) | instid1(SALU_CYCLE_1)
-; GFX11-NEXT: s_or_b32 s2, s2, 1
-; GFX11-NEXT: s_lshr_b32 s2, s2, s30
-; GFX11-NEXT: s_delay_alu instid0(SALU_CYCLE_1) | instskip(NEXT) | instid1(SALU_CYCLE_1)
-; GFX11-NEXT: s_mul_i32 s2, s2, s22
-; GFX11-NEXT: s_mul_i32 s2, s2, s20
-; GFX11-NEXT: s_delay_alu instid0(SALU_CYCLE_1) | instskip(NEXT) | instid1(SALU_CYCLE_1)
-; GFX11-NEXT: s_or_b32 s2, s24, s2
+; GFX11-NEXT: s_or_b32 s3, s3, 1
+; GFX11-NEXT: s_lshr_b32 s3, s3, s30
+; GFX11-NEXT: s_delay_alu instid0(SALU_CYCLE_1) | instskip(SKIP_2) | instid1(SALU_CYCLE_1)
+; GFX11-NEXT: s_mul_i32 s9, s3, s22
+; GFX11-NEXT: s_mov_b32 s3, 0
+; GFX11-NEXT: s_mul_i32 s9, s9, s20
+; GFX11-NEXT: s_or_b32 s2, s2, s9
+; GFX11-NEXT: s_delay_alu instid0(SALU_CYCLE_1)
; GFX11-NEXT: s_lshl_b64 s[22:23], s[2:3], 1
; GFX11-NEXT: global_load_u16 v2, v1, s[22:23]
; GFX11-NEXT: s_waitcnt vmcnt(0)
@@ -152,10 +151,10 @@
; GFX11-NEXT: s_or_b32 s3, s2, s3
; GFX11-NEXT: s_cbranch_vccz .LBB2_6
; GFX11-NEXT: ; %bb.7: ; %Flow
-; GFX11-NEXT: s_mov_b32 s2, 0
+; GFX11-NEXT: s_mov_b32 s3, 0
; GFX11-NEXT: .LBB2_8: ; %Flow12
; GFX11-NEXT: s_delay_alu instid0(SALU_CYCLE_1)
-; GFX11-NEXT: s_and_b32 vcc_lo, exec_lo, s2
+; GFX11-NEXT: s_and_b32 vcc_lo, exec_lo, s3
; GFX11-NEXT: s_cbranch_vccz .LBB2_12
; GFX11-NEXT: ; %bb.9:
; GFX11-NEXT: s_xor_b32 s0, s8, -1
@@ -167,11 +166,11 @@
; GFX11-NEXT: ; %bb.11: ; %Flow6
; GFX11-NEXT: s_mov_b32 s21, -1
; GFX11-NEXT: .LBB2_12: ; %Flow11
-; GFX11-NEXT: s_and_b32 s3, s1, exec_lo
+; GFX11-NEXT: s_and_b32 s20, s1, exec_lo
; GFX11-NEXT: s_or_not1_b32 s0, s21, exec_lo
; GFX11-NEXT: .LBB2_13: ; %Flow9
; GFX11-NEXT: s_or_b32 exec_lo, exec_lo, s19
-; GFX11-NEXT: s_and_saveexec_b32 s19, s0
+; GFX11-NEXT: s_and_saveexec_b32 s2, s0
; GFX11-NEXT: s_cbranch_execz .LBB2_15
; GFX11-NEXT: ; %bb.14: ; %bb43
; GFX11-NEXT: s_add_u32 s8, s16, 0x58
@@ -184,10 +183,10 @@
; GFX11-NEXT: s_mov_b32 s14, s15
; GFX11-NEXT: s_waitcnt lgkmcnt(0)
; GFX11-NEXT: s_swappc_b64 s[30:31], s[0:1]
-; GFX11-NEXT: s_or_b32 s3, s3, exec_lo
+; GFX11-NEXT: s_or_b32 s20, s20, exec_lo
; GFX11-NEXT: .LBB2_15: ; %Flow14
-; GFX11-NEXT: s_or_b32 exec_lo, exec_lo, s19
-; GFX11-NEXT: s_and_saveexec_b32 s0, s3
+; GFX11-NEXT: s_or_b32 exec_lo, exec_lo, s2
+; GFX11-NEXT: s_and_saveexec_b32 s0, s20
; GFX11-NEXT: ; %bb.16: ; %UnifiedUnreachableBlock
; GFX11-NEXT: ; divergent unreachable
; GFX11-NEXT: ; %bb.17: ; %UnifiedReturnBlock
diff --git a/llvm/test/CodeGen/AMDGPU/kernel-vgpr-spill-mubuf-with-voffset.ll b/llvm/test/CodeGen/AMDGPU/kernel-vgpr-spill-mubuf-with-voffset.ll
--- a/llvm/test/CodeGen/AMDGPU/kernel-vgpr-spill-mubuf-with-voffset.ll
+++ b/llvm/test/CodeGen/AMDGPU/kernel-vgpr-spill-mubuf-with-voffset.ll
@@ -13,30 +13,17 @@
; CHECK-NEXT: s_addc_u32 flat_scratch_hi, s13, 0
; CHECK-NEXT: s_add_u32 s0, s0, s17
; CHECK-NEXT: s_addc_u32 s1, s1, 0
-; CHECK-NEXT: ; implicit-def: $vgpr3
-; CHECK-NEXT: v_writelane_b32 v3, s16, 0
-; CHECK-NEXT: s_or_saveexec_b64 s[34:35], -1
-; CHECK-NEXT: s_add_i32 s12, s33, 0x100200
-; CHECK-NEXT: buffer_store_dword v3, off, s[0:3], s12 ; 4-byte Folded Spill
-; CHECK-NEXT: s_mov_b64 exec, s[34:35]
+; CHECK-NEXT: v_writelane_b32 v40, s16, 0
; CHECK-NEXT: s_mov_b32 s13, s15
; CHECK-NEXT: s_mov_b32 s12, s14
-; CHECK-NEXT: v_readlane_b32 s14, v3, 0
+; CHECK-NEXT: v_readlane_b32 s14, v40, 0
; CHECK-NEXT: s_mov_b64 s[16:17], s[8:9]
; CHECK-NEXT: v_mov_b32_e32 v3, v2
; CHECK-NEXT: v_mov_b32_e32 v2, v1
; CHECK-NEXT: v_mov_b32_e32 v1, v0
-; CHECK-NEXT: s_or_saveexec_b64 s[34:35], -1
-; CHECK-NEXT: s_add_i32 s8, s33, 0x100200
-; CHECK-NEXT: buffer_load_dword v0, off, s[0:3], s8 ; 4-byte Folded Reload
-; CHECK-NEXT: s_mov_b64 exec, s[34:35]
; CHECK-NEXT: s_load_dword s8, s[16:17], 0x0
-; CHECK-NEXT: s_waitcnt vmcnt(0) lgkmcnt(0)
-; CHECK-NEXT: v_writelane_b32 v0, s8, 1
-; CHECK-NEXT: s_or_saveexec_b64 s[34:35], -1
-; CHECK-NEXT: s_add_i32 s8, s33, 0x100200
-; CHECK-NEXT: buffer_store_dword v0, off, s[0:3], s8 ; 4-byte Folded Spill
-; CHECK-NEXT: s_mov_b64 exec, s[34:35]
+; CHECK-NEXT: s_waitcnt lgkmcnt(0)
+; CHECK-NEXT: v_writelane_b32 v40, s8, 1
; CHECK-NEXT: ;;#ASMSTART
; CHECK-NEXT: ; def vgpr10
; CHECK-NEXT: ;;#ASMEND
@@ -69,14 +56,9 @@
; CHECK-NEXT: s_mov_b64 s[2:3], s[22:23]
; CHECK-NEXT: s_waitcnt lgkmcnt(0)
; CHECK-NEXT: s_swappc_b64 s[30:31], s[16:17]
-; CHECK-NEXT: s_or_saveexec_b64 s[34:35], -1
-; CHECK-NEXT: s_add_i32 s4, s33, 0x100200
-; CHECK-NEXT: buffer_load_dword v0, off, s[0:3], s4 ; 4-byte Folded Reload
-; CHECK-NEXT: s_mov_b64 exec, s[34:35]
; CHECK-NEXT: s_add_i32 s4, s33, 0x100100
; CHECK-NEXT: buffer_load_dword v10, off, s[0:3], s4 ; 4-byte Folded Reload
-; CHECK-NEXT: s_waitcnt vmcnt(1)
-; CHECK-NEXT: v_readlane_b32 s4, v0, 1
+; CHECK-NEXT: v_readlane_b32 s4, v40, 1
; CHECK-NEXT: s_mov_b32 s5, 0
; CHECK-NEXT: s_cmp_eq_u32 s4, s5
; CHECK-NEXT: v_mov_b32_e32 v0, 0x4000
@@ -84,24 +66,14 @@
; CHECK-NEXT: buffer_store_dword v10, v0, s[0:3], s33 offen ; 4-byte Folded Spill
; CHECK-NEXT: s_cbranch_scc1 .LBB0_2
; CHECK-NEXT: ; %bb.1: ; %store
-; CHECK-NEXT: s_or_saveexec_b64 s[34:35], -1
-; CHECK-NEXT: s_add_i32 s4, s33, 0x100200
-; CHECK-NEXT: buffer_load_dword v0, off, s[0:3], s4 ; 4-byte Folded Reload
-; CHECK-NEXT: s_mov_b64 exec, s[34:35]
; CHECK-NEXT: s_add_i32 s4, s33, 0x100000
-; CHECK-NEXT: buffer_load_dword v2, off, s[0:3], s4 ; 4-byte Folded Reload
+; CHECK-NEXT: buffer_load_dword v1, off, s[0:3], s4 ; 4-byte Folded Reload
; CHECK-NEXT: ; implicit-def: $sgpr4
-; CHECK-NEXT: v_mov_b32_e32 v1, s4
+; CHECK-NEXT: v_mov_b32_e32 v0, s4
; CHECK-NEXT: s_waitcnt vmcnt(0)
-; CHECK-NEXT: ds_write_b32 v1, v2
-; CHECK-NEXT: ; kill: killed $vgpr0
+; CHECK-NEXT: ds_write_b32 v0, v1
; CHECK-NEXT: s_endpgm
; CHECK-NEXT: .LBB0_2: ; %end
-; CHECK-NEXT: s_or_saveexec_b64 s[34:35], -1
-; CHECK-NEXT: s_add_i32 s4, s33, 0x100200
-; CHECK-NEXT: buffer_load_dword v0, off, s[0:3], s4 ; 4-byte Folded Reload
-; CHECK-NEXT: s_mov_b64 exec, s[34:35]
-; CHECK-NEXT: ; kill: killed $vgpr0
; CHECK-NEXT: s_endpgm
%arr = alloca < 1339 x i32>, align 8192, addrspace(5)
%cmp = icmp ne i32 %val, 0
diff --git a/llvm/test/CodeGen/AMDGPU/llc-pipeline.ll b/llvm/test/CodeGen/AMDGPU/llc-pipeline.ll
--- a/llvm/test/CodeGen/AMDGPU/llc-pipeline.ll
+++ b/llvm/test/CodeGen/AMDGPU/llc-pipeline.ll
@@ -116,7 +116,6 @@
; GCN-O0-NEXT: Fast Register Allocator
; GCN-O0-NEXT: SI lower SGPR spill instructions
; GCN-O0-NEXT: Fast Register Allocator
-; GCN-O0-NEXT: SI Lower WWM Copies
; GCN-O0-NEXT: SI Fix VGPR copies
; GCN-O0-NEXT: Remove Redundant DEBUG_VALUE analysis
; GCN-O0-NEXT: Fixup Statepoint Caller Saved
@@ -361,7 +360,6 @@
; GCN-O1-NEXT: Virtual Register Map
; GCN-O1-NEXT: Live Register Matrix
; GCN-O1-NEXT: Greedy Register Allocator
-; GCN-O1-NEXT: SI Lower WWM Copies
; GCN-O1-NEXT: GCN NSA Reassign
; GCN-O1-NEXT: Virtual Register Rewriter
; GCN-O1-NEXT: Stack Slot Coloring
@@ -655,7 +653,6 @@
; GCN-O1-OPTS-NEXT: Virtual Register Map
; GCN-O1-OPTS-NEXT: Live Register Matrix
; GCN-O1-OPTS-NEXT: Greedy Register Allocator
-; GCN-O1-OPTS-NEXT: SI Lower WWM Copies
; GCN-O1-OPTS-NEXT: GCN NSA Reassign
; GCN-O1-OPTS-NEXT: Virtual Register Rewriter
; GCN-O1-OPTS-NEXT: Stack Slot Coloring
@@ -959,7 +956,6 @@
; GCN-O2-NEXT: Virtual Register Map
; GCN-O2-NEXT: Live Register Matrix
; GCN-O2-NEXT: Greedy Register Allocator
-; GCN-O2-NEXT: SI Lower WWM Copies
; GCN-O2-NEXT: GCN NSA Reassign
; GCN-O2-NEXT: Virtual Register Rewriter
; GCN-O2-NEXT: Stack Slot Coloring
@@ -1275,7 +1271,6 @@
; GCN-O3-NEXT: Virtual Register Map
; GCN-O3-NEXT: Live Register Matrix
; GCN-O3-NEXT: Greedy Register Allocator
-; GCN-O3-NEXT: SI Lower WWM Copies
; GCN-O3-NEXT: GCN NSA Reassign
; GCN-O3-NEXT: Virtual Register Rewriter
; GCN-O3-NEXT: Stack Slot Coloring
diff --git a/llvm/test/CodeGen/AMDGPU/mubuf-legalize-operands-non-ptr-intrinsics.ll b/llvm/test/CodeGen/AMDGPU/mubuf-legalize-operands-non-ptr-intrinsics.ll
--- a/llvm/test/CodeGen/AMDGPU/mubuf-legalize-operands-non-ptr-intrinsics.ll
+++ b/llvm/test/CodeGen/AMDGPU/mubuf-legalize-operands-non-ptr-intrinsics.ll
@@ -141,108 +141,82 @@
; W64-O0: ; %bb.0:
; W64-O0-NEXT: s_waitcnt vmcnt(0) expcnt(0) lgkmcnt(0)
; W64-O0-NEXT: s_xor_saveexec_b64 s[4:5], -1
-; W64-O0-NEXT: buffer_store_dword v0, off, s[0:3], s32 offset:28 ; 4-byte Folded Spill
-; W64-O0-NEXT: buffer_store_dword v1, off, s[0:3], s32 offset:32 ; 4-byte Folded Spill
+; W64-O0-NEXT: buffer_store_dword v5, off, s[0:3], s32 offset:24 ; 4-byte Folded Spill
; W64-O0-NEXT: s_mov_b64 exec, s[4:5]
-; W64-O0-NEXT: ; implicit-def: $vgpr5
-; W64-O0-NEXT: buffer_store_dword v4, off, s[0:3], s32 offset:20 ; 4-byte Folded Spill
-; W64-O0-NEXT: v_mov_b32_e32 v5, v3
+; W64-O0-NEXT: buffer_store_dword v4, off, s[0:3], s32 offset:16 ; 4-byte Folded Spill
+; W64-O0-NEXT: v_mov_b32_e32 v4, v3
; W64-O0-NEXT: v_mov_b32_e32 v6, v2
; W64-O0-NEXT: v_mov_b32_e32 v7, v1
-; W64-O0-NEXT: v_mov_b32_e32 v1, v0
-; W64-O0-NEXT: s_or_saveexec_b64 s[16:17], -1
-; W64-O0-NEXT: buffer_load_dword v0, off, s[0:3], s32 ; 4-byte Folded Reload
-; W64-O0-NEXT: s_mov_b64 exec, s[16:17]
; W64-O0-NEXT: ; implicit-def: $sgpr4
; W64-O0-NEXT: ; implicit-def: $sgpr4
; W64-O0-NEXT: ; implicit-def: $sgpr4
; W64-O0-NEXT: ; implicit-def: $sgpr4
-; W64-O0-NEXT: ; kill: def $vgpr1 killed $vgpr1 def $vgpr1_vgpr2_vgpr3_vgpr4 killed $exec
-; W64-O0-NEXT: v_mov_b32_e32 v2, v7
-; W64-O0-NEXT: v_mov_b32_e32 v3, v6
-; W64-O0-NEXT: v_mov_b32_e32 v4, v5
-; W64-O0-NEXT: buffer_store_dword v1, off, s[0:3], s32 offset:4 ; 4-byte Folded Spill
+; W64-O0-NEXT: ; kill: def $vgpr0 killed $vgpr0 def $vgpr0_vgpr1_vgpr2_vgpr3 killed $exec
+; W64-O0-NEXT: v_mov_b32_e32 v1, v7
+; W64-O0-NEXT: v_mov_b32_e32 v2, v6
+; W64-O0-NEXT: v_mov_b32_e32 v3, v4
+; W64-O0-NEXT: buffer_store_dword v0, off, s[0:3], s32 ; 4-byte Folded Spill
; W64-O0-NEXT: s_waitcnt vmcnt(0)
+; W64-O0-NEXT: buffer_store_dword v1, off, s[0:3], s32 offset:4 ; 4-byte Folded Spill
; W64-O0-NEXT: buffer_store_dword v2, off, s[0:3], s32 offset:8 ; 4-byte Folded Spill
; W64-O0-NEXT: buffer_store_dword v3, off, s[0:3], s32 offset:12 ; 4-byte Folded Spill
-; W64-O0-NEXT: buffer_store_dword v4, off, s[0:3], s32 offset:16 ; 4-byte Folded Spill
; W64-O0-NEXT: ; implicit-def: $sgpr4_sgpr5_sgpr6_sgpr7
; W64-O0-NEXT: s_mov_b32 s4, 0
-; W64-O0-NEXT: v_writelane_b32 v0, s4, 0
+; W64-O0-NEXT: v_writelane_b32 v5, s4, 0
; W64-O0-NEXT: s_mov_b64 s[4:5], exec
-; W64-O0-NEXT: v_writelane_b32 v0, s4, 1
-; W64-O0-NEXT: v_writelane_b32 v0, s5, 2
-; W64-O0-NEXT: s_or_saveexec_b64 s[16:17], -1
-; W64-O0-NEXT: buffer_store_dword v0, off, s[0:3], s32 ; 4-byte Folded Spill
-; W64-O0-NEXT: s_mov_b64 exec, s[16:17]
+; W64-O0-NEXT: v_writelane_b32 v5, s4, 1
+; W64-O0-NEXT: v_writelane_b32 v5, s5, 2
; W64-O0-NEXT: .LBB0_1: ; =>This Inner Loop Header: Depth=1
-; W64-O0-NEXT: s_or_saveexec_b64 s[16:17], -1
; W64-O0-NEXT: buffer_load_dword v0, off, s[0:3], s32 ; 4-byte Folded Reload
-; W64-O0-NEXT: s_mov_b64 exec, s[16:17]
; W64-O0-NEXT: buffer_load_dword v1, off, s[0:3], s32 offset:4 ; 4-byte Folded Reload
; W64-O0-NEXT: buffer_load_dword v2, off, s[0:3], s32 offset:8 ; 4-byte Folded Reload
; W64-O0-NEXT: buffer_load_dword v3, off, s[0:3], s32 offset:12 ; 4-byte Folded Reload
-; W64-O0-NEXT: buffer_load_dword v4, off, s[0:3], s32 offset:16 ; 4-byte Folded Reload
; W64-O0-NEXT: s_waitcnt vmcnt(0)
-; W64-O0-NEXT: v_readfirstlane_b32 s8, v1
-; W64-O0-NEXT: v_readfirstlane_b32 s12, v2
+; W64-O0-NEXT: v_readfirstlane_b32 s8, v0
+; W64-O0-NEXT: v_readfirstlane_b32 s12, v1
; W64-O0-NEXT: s_mov_b32 s4, s8
; W64-O0-NEXT: s_mov_b32 s5, s12
-; W64-O0-NEXT: v_cmp_eq_u64_e64 s[4:5], s[4:5], v[1:2]
-; W64-O0-NEXT: v_readfirstlane_b32 s7, v3
-; W64-O0-NEXT: v_readfirstlane_b32 s6, v4
+; W64-O0-NEXT: v_cmp_eq_u64_e64 s[4:5], s[4:5], v[0:1]
+; W64-O0-NEXT: v_readfirstlane_b32 s7, v2
+; W64-O0-NEXT: v_readfirstlane_b32 s6, v3
; W64-O0-NEXT: s_mov_b32 s10, s7
; W64-O0-NEXT: s_mov_b32 s11, s6
-; W64-O0-NEXT: v_cmp_eq_u64_e64 s[10:11], s[10:11], v[3:4]
+; W64-O0-NEXT: v_cmp_eq_u64_e64 s[10:11], s[10:11], v[2:3]
; W64-O0-NEXT: s_and_b64 s[4:5], s[4:5], s[10:11]
; W64-O0-NEXT: ; kill: def $sgpr8 killed $sgpr8 def $sgpr8_sgpr9_sgpr10_sgpr11
; W64-O0-NEXT: s_mov_b32 s9, s12
; W64-O0-NEXT: s_mov_b32 s10, s7
; W64-O0-NEXT: s_mov_b32 s11, s6
-; W64-O0-NEXT: v_writelane_b32 v0, s8, 3
-; W64-O0-NEXT: v_writelane_b32 v0, s9, 4
-; W64-O0-NEXT: v_writelane_b32 v0, s10, 5
-; W64-O0-NEXT: v_writelane_b32 v0, s11, 6
+; W64-O0-NEXT: v_writelane_b32 v5, s8, 3
+; W64-O0-NEXT: v_writelane_b32 v5, s9, 4
+; W64-O0-NEXT: v_writelane_b32 v5, s10, 5
+; W64-O0-NEXT: v_writelane_b32 v5, s11, 6
; W64-O0-NEXT: s_and_saveexec_b64 s[4:5], s[4:5]
-; W64-O0-NEXT: v_writelane_b32 v0, s4, 7
-; W64-O0-NEXT: v_writelane_b32 v0, s5, 8
-; W64-O0-NEXT: s_or_saveexec_b64 s[16:17], -1
-; W64-O0-NEXT: buffer_store_dword v0, off, s[0:3], s32 ; 4-byte Folded Spill
-; W64-O0-NEXT: s_mov_b64 exec, s[16:17]
+; W64-O0-NEXT: v_writelane_b32 v5, s4, 7
+; W64-O0-NEXT: v_writelane_b32 v5, s5, 8
; W64-O0-NEXT: ; %bb.2: ; in Loop: Header=BB0_1 Depth=1
-; W64-O0-NEXT: buffer_load_dword v0, off, s[0:3], s32 offset:20 ; 4-byte Folded Reload
-; W64-O0-NEXT: s_or_saveexec_b64 s[16:17], -1
-; W64-O0-NEXT: buffer_load_dword v1, off, s[0:3], s32 ; 4-byte Folded Reload
-; W64-O0-NEXT: s_mov_b64 exec, s[16:17]
+; W64-O0-NEXT: buffer_load_dword v0, off, s[0:3], s32 offset:16 ; 4-byte Folded Reload
+; W64-O0-NEXT: v_readlane_b32 s4, v5, 7
+; W64-O0-NEXT: v_readlane_b32 s5, v5, 8
+; W64-O0-NEXT: v_readlane_b32 s8, v5, 3
+; W64-O0-NEXT: v_readlane_b32 s9, v5, 4
+; W64-O0-NEXT: v_readlane_b32 s10, v5, 5
+; W64-O0-NEXT: v_readlane_b32 s11, v5, 6
+; W64-O0-NEXT: v_readlane_b32 s6, v5, 0
; W64-O0-NEXT: s_waitcnt vmcnt(0)
-; W64-O0-NEXT: v_readlane_b32 s4, v1, 7
-; W64-O0-NEXT: v_readlane_b32 s5, v1, 8
-; W64-O0-NEXT: v_readlane_b32 s8, v1, 3
-; W64-O0-NEXT: v_readlane_b32 s9, v1, 4
-; W64-O0-NEXT: v_readlane_b32 s10, v1, 5
-; W64-O0-NEXT: v_readlane_b32 s11, v1, 6
-; W64-O0-NEXT: v_readlane_b32 s6, v1, 0
-; W64-O0-NEXT: s_nop 4
+; W64-O0-NEXT: s_nop 3
; W64-O0-NEXT: buffer_load_format_x v0, v0, s[8:11], s6 idxen
; W64-O0-NEXT: s_waitcnt vmcnt(0)
-; W64-O0-NEXT: buffer_store_dword v0, off, s[0:3], s32 offset:24 ; 4-byte Folded Spill
+; W64-O0-NEXT: buffer_store_dword v0, off, s[0:3], s32 offset:20 ; 4-byte Folded Spill
; W64-O0-NEXT: s_xor_b64 exec, exec, s[4:5]
; W64-O0-NEXT: s_cbranch_execnz .LBB0_1
; W64-O0-NEXT: ; %bb.3:
-; W64-O0-NEXT: s_or_saveexec_b64 s[16:17], -1
-; W64-O0-NEXT: buffer_load_dword v1, off, s[0:3], s32 ; 4-byte Folded Reload
-; W64-O0-NEXT: s_mov_b64 exec, s[16:17]
-; W64-O0-NEXT: buffer_load_dword v0, off, s[0:3], s32 offset:24 ; 4-byte Folded Reload
-; W64-O0-NEXT: s_waitcnt vmcnt(1)
-; W64-O0-NEXT: v_readlane_b32 s4, v1, 1
-; W64-O0-NEXT: v_readlane_b32 s5, v1, 2
+; W64-O0-NEXT: buffer_load_dword v0, off, s[0:3], s32 offset:20 ; 4-byte Folded Reload
+; W64-O0-NEXT: v_readlane_b32 s4, v5, 1
+; W64-O0-NEXT: v_readlane_b32 s5, v5, 2
; W64-O0-NEXT: s_mov_b64 exec, s[4:5]
-; W64-O0-NEXT: ; kill: killed $vgpr1
; W64-O0-NEXT: s_xor_saveexec_b64 s[4:5], -1
-; W64-O0-NEXT: s_waitcnt vmcnt(0)
-; W64-O0-NEXT: buffer_load_dword v0, off, s[0:3], s32 offset:28 ; 4-byte Folded Reload
-; W64-O0-NEXT: s_nop 0
-; W64-O0-NEXT: buffer_load_dword v1, off, s[0:3], s32 offset:32 ; 4-byte Folded Reload
+; W64-O0-NEXT: buffer_load_dword v5, off, s[0:3], s32 offset:24 ; 4-byte Folded Reload
; W64-O0-NEXT: s_mov_b64 exec, s[4:5]
; W64-O0-NEXT: s_waitcnt vmcnt(0)
; W64-O0-NEXT: s_setpc_b64 s[30:31]
@@ -494,221 +468,185 @@
; W64-O0: ; %bb.0: ; %entry
; W64-O0-NEXT: s_waitcnt vmcnt(0) expcnt(0) lgkmcnt(0)
; W64-O0-NEXT: s_xor_saveexec_b64 s[4:5], -1
-; W64-O0-NEXT: buffer_store_dword v0, off, s[0:3], s32 offset:76 ; 4-byte Folded Spill
-; W64-O0-NEXT: buffer_store_dword v1, off, s[0:3], s32 offset:80 ; 4-byte Folded Spill
+; W64-O0-NEXT: buffer_store_dword v13, off, s[0:3], s32 offset:76 ; 4-byte Folded Spill
; W64-O0-NEXT: s_mov_b64 exec, s[4:5]
-; W64-O0-NEXT: ; implicit-def: $vgpr13
-; W64-O0-NEXT: buffer_store_dword v11, off, s[0:3], s32 offset:56 ; 4-byte Folded Spill
-; W64-O0-NEXT: buffer_store_dword v9, off, s[0:3], s32 offset:52 ; 4-byte Folded Spill
+; W64-O0-NEXT: buffer_store_dword v11, off, s[0:3], s32 offset:52 ; 4-byte Folded Spill
+; W64-O0-NEXT: buffer_store_dword v9, off, s[0:3], s32 offset:48 ; 4-byte Folded Spill
; W64-O0-NEXT: buffer_store_dword v8, off, s[0:3], s32 offset:64 ; 4-byte Folded Spill
-; W64-O0-NEXT: buffer_store_dword v7, off, s[0:3], s32 offset:60 ; 4-byte Folded Spill
-; W64-O0-NEXT: v_mov_b32_e32 v13, v4
+; W64-O0-NEXT: buffer_store_dword v7, off, s[0:3], s32 offset:56 ; 4-byte Folded Spill
+; W64-O0-NEXT: buffer_store_dword v6, off, s[0:3], s32 offset:60 ; 4-byte Folded Spill
+; W64-O0-NEXT: v_mov_b32_e32 v14, v4
; W64-O0-NEXT: buffer_load_dword v4, off, s[0:3], s32 offset:60 ; 4-byte Folded Reload
-; W64-O0-NEXT: v_mov_b32_e32 v7, v3
-; W64-O0-NEXT: v_mov_b32_e32 v8, v2
-; W64-O0-NEXT: buffer_load_dword v2, off, s[0:3], s32 offset:56 ; 4-byte Folded Reload
-; W64-O0-NEXT: v_mov_b32_e32 v9, v1
+; W64-O0-NEXT: v_mov_b32_e32 v6, v3
+; W64-O0-NEXT: buffer_load_dword v3, off, s[0:3], s32 offset:56 ; 4-byte Folded Reload
+; W64-O0-NEXT: v_mov_b32_e32 v7, v2
+; W64-O0-NEXT: v_mov_b32_e32 v8, v1
; W64-O0-NEXT: buffer_load_dword v1, off, s[0:3], s32 offset:52 ; 4-byte Folded Reload
-; W64-O0-NEXT: v_mov_b32_e32 v3, v0
-; W64-O0-NEXT: s_or_saveexec_b64 s[16:17], -1
-; W64-O0-NEXT: buffer_load_dword v0, off, s[0:3], s32 ; 4-byte Folded Reload
-; W64-O0-NEXT: s_mov_b64 exec, s[16:17]
+; W64-O0-NEXT: v_mov_b32_e32 v2, v0
+; W64-O0-NEXT: buffer_load_dword v0, off, s[0:3], s32 offset:48 ; 4-byte Folded Reload
; W64-O0-NEXT: ; implicit-def: $sgpr4
; W64-O0-NEXT: ; implicit-def: $sgpr4
; W64-O0-NEXT: ; implicit-def: $sgpr4
; W64-O0-NEXT: ; implicit-def: $sgpr4
-; W64-O0-NEXT: ; kill: def $vgpr13 killed $vgpr13 def $vgpr13_vgpr14_vgpr15_vgpr16 killed $exec
-; W64-O0-NEXT: v_mov_b32_e32 v14, v5
-; W64-O0-NEXT: v_mov_b32_e32 v15, v6
+; W64-O0-NEXT: ; kill: def $vgpr14 killed $vgpr14 def $vgpr14_vgpr15_vgpr16_vgpr17 killed $exec
+; W64-O0-NEXT: v_mov_b32_e32 v15, v5
; W64-O0-NEXT: s_waitcnt vmcnt(3)
; W64-O0-NEXT: v_mov_b32_e32 v16, v4
-; W64-O0-NEXT: buffer_store_dword v13, off, s[0:3], s32 offset:36 ; 4-byte Folded Spill
+; W64-O0-NEXT: s_waitcnt vmcnt(2)
+; W64-O0-NEXT: v_mov_b32_e32 v17, v3
+; W64-O0-NEXT: buffer_store_dword v14, off, s[0:3], s32 offset:32 ; 4-byte Folded Spill
; W64-O0-NEXT: s_waitcnt vmcnt(0)
-; W64-O0-NEXT: buffer_store_dword v14, off, s[0:3], s32 offset:40 ; 4-byte Folded Spill
-; W64-O0-NEXT: buffer_store_dword v15, off, s[0:3], s32 offset:44 ; 4-byte Folded Spill
-; W64-O0-NEXT: buffer_store_dword v16, off, s[0:3], s32 offset:48 ; 4-byte Folded Spill
+; W64-O0-NEXT: buffer_store_dword v15, off, s[0:3], s32 offset:36 ; 4-byte Folded Spill
+; W64-O0-NEXT: buffer_store_dword v16, off, s[0:3], s32 offset:40 ; 4-byte Folded Spill
+; W64-O0-NEXT: buffer_store_dword v17, off, s[0:3], s32 offset:44 ; 4-byte Folded Spill
; W64-O0-NEXT: ; implicit-def: $sgpr4
; W64-O0-NEXT: ; implicit-def: $sgpr4
; W64-O0-NEXT: ; implicit-def: $sgpr4
; W64-O0-NEXT: ; implicit-def: $sgpr4
-; W64-O0-NEXT: ; kill: def $vgpr3 killed $vgpr3 def $vgpr3_vgpr4_vgpr5_vgpr6 killed $exec
-; W64-O0-NEXT: v_mov_b32_e32 v4, v9
-; W64-O0-NEXT: v_mov_b32_e32 v5, v8
-; W64-O0-NEXT: v_mov_b32_e32 v6, v7
-; W64-O0-NEXT: buffer_store_dword v3, off, s[0:3], s32 offset:20 ; 4-byte Folded Spill
+; W64-O0-NEXT: ; kill: def $vgpr2 killed $vgpr2 def $vgpr2_vgpr3_vgpr4_vgpr5 killed $exec
+; W64-O0-NEXT: v_mov_b32_e32 v3, v8
+; W64-O0-NEXT: v_mov_b32_e32 v4, v7
+; W64-O0-NEXT: v_mov_b32_e32 v5, v6
+; W64-O0-NEXT: buffer_store_dword v2, off, s[0:3], s32 offset:16 ; 4-byte Folded Spill
; W64-O0-NEXT: s_waitcnt vmcnt(0)
+; W64-O0-NEXT: buffer_store_dword v3, off, s[0:3], s32 offset:20 ; 4-byte Folded Spill
; W64-O0-NEXT: buffer_store_dword v4, off, s[0:3], s32 offset:24 ; 4-byte Folded Spill
; W64-O0-NEXT: buffer_store_dword v5, off, s[0:3], s32 offset:28 ; 4-byte Folded Spill
-; W64-O0-NEXT: buffer_store_dword v6, off, s[0:3], s32 offset:32 ; 4-byte Folded Spill
; W64-O0-NEXT: ; implicit-def: $sgpr4
; W64-O0-NEXT: ; implicit-def: $sgpr4
-; W64-O0-NEXT: ; kill: def $vgpr2 killed $vgpr2 def $vgpr2_vgpr3 killed $exec
-; W64-O0-NEXT: v_mov_b32_e32 v3, v12
-; W64-O0-NEXT: buffer_store_dword v2, off, s[0:3], s32 offset:12 ; 4-byte Folded Spill
+; W64-O0-NEXT: ; kill: def $vgpr1 killed $vgpr1 def $vgpr1_vgpr2 killed $exec
+; W64-O0-NEXT: v_mov_b32_e32 v2, v12
+; W64-O0-NEXT: buffer_store_dword v1, off, s[0:3], s32 offset:8 ; 4-byte Folded Spill
; W64-O0-NEXT: s_waitcnt vmcnt(0)
-; W64-O0-NEXT: buffer_store_dword v3, off, s[0:3], s32 offset:16 ; 4-byte Folded Spill
+; W64-O0-NEXT: buffer_store_dword v2, off, s[0:3], s32 offset:12 ; 4-byte Folded Spill
; W64-O0-NEXT: ; implicit-def: $sgpr4
; W64-O0-NEXT: ; implicit-def: $sgpr4
-; W64-O0-NEXT: ; kill: def $vgpr1 killed $vgpr1 def $vgpr1_vgpr2 killed $exec
-; W64-O0-NEXT: v_mov_b32_e32 v2, v10
-; W64-O0-NEXT: buffer_store_dword v1, off, s[0:3], s32 offset:4 ; 4-byte Folded Spill
+; W64-O0-NEXT: ; kill: def $vgpr0 killed $vgpr0 def $vgpr0_vgpr1 killed $exec
+; W64-O0-NEXT: v_mov_b32_e32 v1, v10
+; W64-O0-NEXT: buffer_store_dword v0, off, s[0:3], s32 ; 4-byte Folded Spill
; W64-O0-NEXT: s_waitcnt vmcnt(0)
-; W64-O0-NEXT: buffer_store_dword v2, off, s[0:3], s32 offset:8 ; 4-byte Folded Spill
+; W64-O0-NEXT: buffer_store_dword v1, off, s[0:3], s32 offset:4 ; 4-byte Folded Spill
; W64-O0-NEXT: ; implicit-def: $sgpr4_sgpr5
; W64-O0-NEXT: ; implicit-def: $sgpr4_sgpr5
; W64-O0-NEXT: ; implicit-def: $sgpr4_sgpr5_sgpr6_sgpr7
; W64-O0-NEXT: ; implicit-def: $sgpr4_sgpr5_sgpr6_sgpr7
; W64-O0-NEXT: s_mov_b32 s4, 0
-; W64-O0-NEXT: v_writelane_b32 v0, s4, 0
+; W64-O0-NEXT: v_writelane_b32 v13, s4, 0
; W64-O0-NEXT: s_mov_b64 s[4:5], exec
-; W64-O0-NEXT: v_writelane_b32 v0, s4, 1
-; W64-O0-NEXT: v_writelane_b32 v0, s5, 2
-; W64-O0-NEXT: s_or_saveexec_b64 s[16:17], -1
-; W64-O0-NEXT: buffer_store_dword v0, off, s[0:3], s32 ; 4-byte Folded Spill
-; W64-O0-NEXT: s_mov_b64 exec, s[16:17]
+; W64-O0-NEXT: v_writelane_b32 v13, s4, 1
+; W64-O0-NEXT: v_writelane_b32 v13, s5, 2
; W64-O0-NEXT: .LBB1_1: ; =>This Inner Loop Header: Depth=1
-; W64-O0-NEXT: s_or_saveexec_b64 s[16:17], -1
-; W64-O0-NEXT: buffer_load_dword v0, off, s[0:3], s32 ; 4-byte Folded Reload
-; W64-O0-NEXT: s_mov_b64 exec, s[16:17]
+; W64-O0-NEXT: buffer_load_dword v0, off, s[0:3], s32 offset:16 ; 4-byte Folded Reload
; W64-O0-NEXT: buffer_load_dword v1, off, s[0:3], s32 offset:20 ; 4-byte Folded Reload
; W64-O0-NEXT: buffer_load_dword v2, off, s[0:3], s32 offset:24 ; 4-byte Folded Reload
; W64-O0-NEXT: buffer_load_dword v3, off, s[0:3], s32 offset:28 ; 4-byte Folded Reload
-; W64-O0-NEXT: buffer_load_dword v4, off, s[0:3], s32 offset:32 ; 4-byte Folded Reload
; W64-O0-NEXT: s_waitcnt vmcnt(0)
-; W64-O0-NEXT: v_readfirstlane_b32 s8, v1
-; W64-O0-NEXT: v_readfirstlane_b32 s12, v2
+; W64-O0-NEXT: v_readfirstlane_b32 s8, v0
+; W64-O0-NEXT: v_readfirstlane_b32 s12, v1
; W64-O0-NEXT: s_mov_b32 s4, s8
; W64-O0-NEXT: s_mov_b32 s5, s12
-; W64-O0-NEXT: v_cmp_eq_u64_e64 s[4:5], s[4:5], v[1:2]
-; W64-O0-NEXT: v_readfirstlane_b32 s7, v3
-; W64-O0-NEXT: v_readfirstlane_b32 s6, v4
+; W64-O0-NEXT: v_cmp_eq_u64_e64 s[4:5], s[4:5], v[0:1]
+; W64-O0-NEXT: v_readfirstlane_b32 s7, v2
+; W64-O0-NEXT: v_readfirstlane_b32 s6, v3
; W64-O0-NEXT: s_mov_b32 s10, s7
; W64-O0-NEXT: s_mov_b32 s11, s6
-; W64-O0-NEXT: v_cmp_eq_u64_e64 s[10:11], s[10:11], v[3:4]
+; W64-O0-NEXT: v_cmp_eq_u64_e64 s[10:11], s[10:11], v[2:3]
; W64-O0-NEXT: s_and_b64 s[4:5], s[4:5], s[10:11]
; W64-O0-NEXT: ; kill: def $sgpr8 killed $sgpr8 def $sgpr8_sgpr9_sgpr10_sgpr11
; W64-O0-NEXT: s_mov_b32 s9, s12
; W64-O0-NEXT: s_mov_b32 s10, s7
; W64-O0-NEXT: s_mov_b32 s11, s6
-; W64-O0-NEXT: v_writelane_b32 v0, s8, 3
-; W64-O0-NEXT: v_writelane_b32 v0, s9, 4
-; W64-O0-NEXT: v_writelane_b32 v0, s10, 5
-; W64-O0-NEXT: v_writelane_b32 v0, s11, 6
+; W64-O0-NEXT: v_writelane_b32 v13, s8, 3
+; W64-O0-NEXT: v_writelane_b32 v13, s9, 4
+; W64-O0-NEXT: v_writelane_b32 v13, s10, 5
+; W64-O0-NEXT: v_writelane_b32 v13, s11, 6
; W64-O0-NEXT: s_and_saveexec_b64 s[4:5], s[4:5]
-; W64-O0-NEXT: v_writelane_b32 v0, s4, 7
-; W64-O0-NEXT: v_writelane_b32 v0, s5, 8
-; W64-O0-NEXT: s_or_saveexec_b64 s[16:17], -1
-; W64-O0-NEXT: buffer_store_dword v0, off, s[0:3], s32 ; 4-byte Folded Spill
-; W64-O0-NEXT: s_mov_b64 exec, s[16:17]
+; W64-O0-NEXT: v_writelane_b32 v13, s4, 7
+; W64-O0-NEXT: v_writelane_b32 v13, s5, 8
; W64-O0-NEXT: ; %bb.2: ; in Loop: Header=BB1_1 Depth=1
; W64-O0-NEXT: buffer_load_dword v0, off, s[0:3], s32 offset:64 ; 4-byte Folded Reload
-; W64-O0-NEXT: s_or_saveexec_b64 s[16:17], -1
-; W64-O0-NEXT: buffer_load_dword v1, off, s[0:3], s32 ; 4-byte Folded Reload
-; W64-O0-NEXT: s_mov_b64 exec, s[16:17]
+; W64-O0-NEXT: v_readlane_b32 s4, v13, 7
+; W64-O0-NEXT: v_readlane_b32 s5, v13, 8
+; W64-O0-NEXT: v_readlane_b32 s8, v13, 3
+; W64-O0-NEXT: v_readlane_b32 s9, v13, 4
+; W64-O0-NEXT: v_readlane_b32 s10, v13, 5
+; W64-O0-NEXT: v_readlane_b32 s11, v13, 6
+; W64-O0-NEXT: v_readlane_b32 s6, v13, 0
; W64-O0-NEXT: s_waitcnt vmcnt(0)
-; W64-O0-NEXT: v_readlane_b32 s4, v1, 7
-; W64-O0-NEXT: v_readlane_b32 s5, v1, 8
-; W64-O0-NEXT: v_readlane_b32 s8, v1, 3
-; W64-O0-NEXT: v_readlane_b32 s9, v1, 4
-; W64-O0-NEXT: v_readlane_b32 s10, v1, 5
-; W64-O0-NEXT: v_readlane_b32 s11, v1, 6
-; W64-O0-NEXT: v_readlane_b32 s6, v1, 0
-; W64-O0-NEXT: s_nop 4
+; W64-O0-NEXT: s_nop 3
; W64-O0-NEXT: buffer_load_format_x v0, v0, s[8:11], s6 idxen
; W64-O0-NEXT: s_waitcnt vmcnt(0)
; W64-O0-NEXT: buffer_store_dword v0, off, s[0:3], s32 offset:68 ; 4-byte Folded Spill
; W64-O0-NEXT: s_xor_b64 exec, exec, s[4:5]
; W64-O0-NEXT: s_cbranch_execnz .LBB1_1
; W64-O0-NEXT: ; %bb.3:
-; W64-O0-NEXT: s_or_saveexec_b64 s[16:17], -1
-; W64-O0-NEXT: buffer_load_dword v0, off, s[0:3], s32 ; 4-byte Folded Reload
-; W64-O0-NEXT: s_mov_b64 exec, s[16:17]
-; W64-O0-NEXT: s_waitcnt vmcnt(0)
-; W64-O0-NEXT: v_readlane_b32 s4, v0, 1
-; W64-O0-NEXT: v_readlane_b32 s5, v0, 2
+; W64-O0-NEXT: v_readlane_b32 s4, v13, 1
+; W64-O0-NEXT: v_readlane_b32 s5, v13, 2
; W64-O0-NEXT: s_mov_b64 exec, s[4:5]
; W64-O0-NEXT: s_mov_b64 s[4:5], exec
-; W64-O0-NEXT: v_writelane_b32 v0, s4, 9
-; W64-O0-NEXT: v_writelane_b32 v0, s5, 10
-; W64-O0-NEXT: s_or_saveexec_b64 s[16:17], -1
-; W64-O0-NEXT: buffer_store_dword v0, off, s[0:3], s32 ; 4-byte Folded Spill
-; W64-O0-NEXT: s_mov_b64 exec, s[16:17]
+; W64-O0-NEXT: v_writelane_b32 v13, s4, 9
+; W64-O0-NEXT: v_writelane_b32 v13, s5, 10
; W64-O0-NEXT: .LBB1_4: ; =>This Inner Loop Header: Depth=1
-; W64-O0-NEXT: s_or_saveexec_b64 s[16:17], -1
-; W64-O0-NEXT: buffer_load_dword v0, off, s[0:3], s32 ; 4-byte Folded Reload
-; W64-O0-NEXT: s_mov_b64 exec, s[16:17]
+; W64-O0-NEXT: buffer_load_dword v0, off, s[0:3], s32 offset:32 ; 4-byte Folded Reload
; W64-O0-NEXT: buffer_load_dword v1, off, s[0:3], s32 offset:36 ; 4-byte Folded Reload
; W64-O0-NEXT: buffer_load_dword v2, off, s[0:3], s32 offset:40 ; 4-byte Folded Reload
; W64-O0-NEXT: buffer_load_dword v3, off, s[0:3], s32 offset:44 ; 4-byte Folded Reload
-; W64-O0-NEXT: buffer_load_dword v4, off, s[0:3], s32 offset:48 ; 4-byte Folded Reload
; W64-O0-NEXT: s_waitcnt vmcnt(0)
-; W64-O0-NEXT: v_readfirstlane_b32 s8, v1
-; W64-O0-NEXT: v_readfirstlane_b32 s12, v2
+; W64-O0-NEXT: v_readfirstlane_b32 s8, v0
+; W64-O0-NEXT: v_readfirstlane_b32 s12, v1
; W64-O0-NEXT: s_mov_b32 s4, s8
; W64-O0-NEXT: s_mov_b32 s5, s12
-; W64-O0-NEXT: v_cmp_eq_u64_e64 s[4:5], s[4:5], v[1:2]
-; W64-O0-NEXT: v_readfirstlane_b32 s7, v3
-; W64-O0-NEXT: v_readfirstlane_b32 s6, v4
+; W64-O0-NEXT: v_cmp_eq_u64_e64 s[4:5], s[4:5], v[0:1]
+; W64-O0-NEXT: v_readfirstlane_b32 s7, v2
+; W64-O0-NEXT: v_readfirstlane_b32 s6, v3
; W64-O0-NEXT: s_mov_b32 s10, s7
; W64-O0-NEXT: s_mov_b32 s11, s6
-; W64-O0-NEXT: v_cmp_eq_u64_e64 s[10:11], s[10:11], v[3:4]
+; W64-O0-NEXT: v_cmp_eq_u64_e64 s[10:11], s[10:11], v[2:3]
; W64-O0-NEXT: s_and_b64 s[4:5], s[4:5], s[10:11]
; W64-O0-NEXT: ; kill: def $sgpr8 killed $sgpr8 def $sgpr8_sgpr9_sgpr10_sgpr11
; W64-O0-NEXT: s_mov_b32 s9, s12
; W64-O0-NEXT: s_mov_b32 s10, s7
; W64-O0-NEXT: s_mov_b32 s11, s6
-; W64-O0-NEXT: v_writelane_b32 v0, s8, 11
-; W64-O0-NEXT: v_writelane_b32 v0, s9, 12
-; W64-O0-NEXT: v_writelane_b32 v0, s10, 13
-; W64-O0-NEXT: v_writelane_b32 v0, s11, 14
+; W64-O0-NEXT: v_writelane_b32 v13, s8, 11
+; W64-O0-NEXT: v_writelane_b32 v13, s9, 12
+; W64-O0-NEXT: v_writelane_b32 v13, s10, 13
+; W64-O0-NEXT: v_writelane_b32 v13, s11, 14
; W64-O0-NEXT: s_and_saveexec_b64 s[4:5], s[4:5]
-; W64-O0-NEXT: v_writelane_b32 v0, s4, 15
-; W64-O0-NEXT: v_writelane_b32 v0, s5, 16
-; W64-O0-NEXT: s_or_saveexec_b64 s[16:17], -1
-; W64-O0-NEXT: buffer_store_dword v0, off, s[0:3], s32 ; 4-byte Folded Spill
-; W64-O0-NEXT: s_mov_b64 exec, s[16:17]
+; W64-O0-NEXT: v_writelane_b32 v13, s4, 15
+; W64-O0-NEXT: v_writelane_b32 v13, s5, 16
; W64-O0-NEXT: ; %bb.5: ; in Loop: Header=BB1_4 Depth=1
; W64-O0-NEXT: buffer_load_dword v0, off, s[0:3], s32 offset:64 ; 4-byte Folded Reload
-; W64-O0-NEXT: s_or_saveexec_b64 s[16:17], -1
-; W64-O0-NEXT: buffer_load_dword v1, off, s[0:3], s32 ; 4-byte Folded Reload
-; W64-O0-NEXT: s_mov_b64 exec, s[16:17]
+; W64-O0-NEXT: v_readlane_b32 s4, v13, 15
+; W64-O0-NEXT: v_readlane_b32 s5, v13, 16
+; W64-O0-NEXT: v_readlane_b32 s8, v13, 11
+; W64-O0-NEXT: v_readlane_b32 s9, v13, 12
+; W64-O0-NEXT: v_readlane_b32 s10, v13, 13
+; W64-O0-NEXT: v_readlane_b32 s11, v13, 14
+; W64-O0-NEXT: v_readlane_b32 s6, v13, 0
; W64-O0-NEXT: s_waitcnt vmcnt(0)
-; W64-O0-NEXT: v_readlane_b32 s4, v1, 15
-; W64-O0-NEXT: v_readlane_b32 s5, v1, 16
-; W64-O0-NEXT: v_readlane_b32 s8, v1, 11
-; W64-O0-NEXT: v_readlane_b32 s9, v1, 12
-; W64-O0-NEXT: v_readlane_b32 s10, v1, 13
-; W64-O0-NEXT: v_readlane_b32 s11, v1, 14
-; W64-O0-NEXT: v_readlane_b32 s6, v1, 0
-; W64-O0-NEXT: s_nop 4
+; W64-O0-NEXT: s_nop 3
; W64-O0-NEXT: buffer_load_format_x v0, v0, s[8:11], s6 idxen
; W64-O0-NEXT: s_waitcnt vmcnt(0)
; W64-O0-NEXT: buffer_store_dword v0, off, s[0:3], s32 offset:72 ; 4-byte Folded Spill
; W64-O0-NEXT: s_xor_b64 exec, exec, s[4:5]
; W64-O0-NEXT: s_cbranch_execnz .LBB1_4
; W64-O0-NEXT: ; %bb.6:
-; W64-O0-NEXT: s_or_saveexec_b64 s[16:17], -1
-; W64-O0-NEXT: buffer_load_dword v0, off, s[0:3], s32 ; 4-byte Folded Reload
-; W64-O0-NEXT: s_mov_b64 exec, s[16:17]
-; W64-O0-NEXT: buffer_load_dword v3, off, s[0:3], s32 offset:72 ; 4-byte Folded Reload
+; W64-O0-NEXT: buffer_load_dword v2, off, s[0:3], s32 offset:72 ; 4-byte Folded Reload
+; W64-O0-NEXT: buffer_load_dword v0, off, s[0:3], s32 offset:8 ; 4-byte Folded Reload
; W64-O0-NEXT: buffer_load_dword v1, off, s[0:3], s32 offset:12 ; 4-byte Folded Reload
-; W64-O0-NEXT: buffer_load_dword v2, off, s[0:3], s32 offset:16 ; 4-byte Folded Reload
-; W64-O0-NEXT: buffer_load_dword v6, off, s[0:3], s32 offset:68 ; 4-byte Folded Reload
+; W64-O0-NEXT: buffer_load_dword v5, off, s[0:3], s32 offset:68 ; 4-byte Folded Reload
+; W64-O0-NEXT: buffer_load_dword v3, off, s[0:3], s32 ; 4-byte Folded Reload
; W64-O0-NEXT: buffer_load_dword v4, off, s[0:3], s32 offset:4 ; 4-byte Folded Reload
-; W64-O0-NEXT: buffer_load_dword v5, off, s[0:3], s32 offset:8 ; 4-byte Folded Reload
-; W64-O0-NEXT: s_waitcnt vmcnt(6)
-; W64-O0-NEXT: v_readlane_b32 s4, v0, 9
-; W64-O0-NEXT: v_readlane_b32 s5, v0, 10
+; W64-O0-NEXT: v_readlane_b32 s4, v13, 9
+; W64-O0-NEXT: v_readlane_b32 s5, v13, 10
; W64-O0-NEXT: s_mov_b64 exec, s[4:5]
; W64-O0-NEXT: s_waitcnt vmcnt(0)
-; W64-O0-NEXT: global_store_dword v[4:5], v6, off
+; W64-O0-NEXT: global_store_dword v[3:4], v5, off
; W64-O0-NEXT: s_waitcnt vmcnt(0)
-; W64-O0-NEXT: global_store_dword v[1:2], v3, off
+; W64-O0-NEXT: global_store_dword v[0:1], v2, off
; W64-O0-NEXT: s_waitcnt vmcnt(0)
-; W64-O0-NEXT: ; kill: killed $vgpr0
; W64-O0-NEXT: s_xor_saveexec_b64 s[4:5], -1
-; W64-O0-NEXT: buffer_load_dword v0, off, s[0:3], s32 offset:76 ; 4-byte Folded Reload
-; W64-O0-NEXT: buffer_load_dword v1, off, s[0:3], s32 offset:80 ; 4-byte Folded Reload
+; W64-O0-NEXT: buffer_load_dword v13, off, s[0:3], s32 offset:76 ; 4-byte Folded Reload
; W64-O0-NEXT: s_mov_b64 exec, s[4:5]
; W64-O0-NEXT: s_waitcnt vmcnt(0)
; W64-O0-NEXT: s_setpc_b64 s[30:31]
@@ -1016,250 +954,204 @@
; W64-O0: ; %bb.0: ; %entry
; W64-O0-NEXT: s_waitcnt vmcnt(0) expcnt(0) lgkmcnt(0)
; W64-O0-NEXT: s_xor_saveexec_b64 s[4:5], -1
-; W64-O0-NEXT: buffer_store_dword v0, off, s[0:3], s32 offset:76 ; 4-byte Folded Spill
-; W64-O0-NEXT: buffer_store_dword v1, off, s[0:3], s32 offset:80 ; 4-byte Folded Spill
+; W64-O0-NEXT: buffer_store_dword v8, off, s[0:3], s32 offset:72 ; 4-byte Folded Spill
; W64-O0-NEXT: s_mov_b64 exec, s[4:5]
-; W64-O0-NEXT: ; implicit-def: $vgpr8
-; W64-O0-NEXT: buffer_store_dword v31, off, s[0:3], s32 offset:56 ; 4-byte Folded Spill
-; W64-O0-NEXT: buffer_store_dword v11, off, s[0:3], s32 offset:52 ; 4-byte Folded Spill
-; W64-O0-NEXT: v_mov_b32_e32 v8, v6
-; W64-O0-NEXT: v_mov_b32_e32 v9, v5
-; W64-O0-NEXT: buffer_load_dword v5, off, s[0:3], s32 offset:52 ; 4-byte Folded Reload
+; W64-O0-NEXT: buffer_store_dword v31, off, s[0:3], s32 offset:52 ; 4-byte Folded Spill
+; W64-O0-NEXT: buffer_store_dword v11, off, s[0:3], s32 offset:48 ; 4-byte Folded Spill
+; W64-O0-NEXT: v_mov_b32_e32 v9, v7
+; W64-O0-NEXT: v_mov_b32_e32 v10, v6
+; W64-O0-NEXT: v_mov_b32_e32 v11, v5
+; W64-O0-NEXT: v_mov_b32_e32 v5, v4
+; W64-O0-NEXT: buffer_load_dword v4, off, s[0:3], s32 offset:48 ; 4-byte Folded Reload
; W64-O0-NEXT: s_nop 0
-; W64-O0-NEXT: buffer_store_dword v4, off, s[0:3], s32 offset:48 ; 4-byte Folded Spill
-; W64-O0-NEXT: v_mov_b32_e32 v10, v3
-; W64-O0-NEXT: v_mov_b32_e32 v11, v2
-; W64-O0-NEXT: v_mov_b32_e32 v13, v1
-; W64-O0-NEXT: buffer_load_dword v1, off, s[0:3], s32 offset:48 ; 4-byte Folded Reload
-; W64-O0-NEXT: v_mov_b32_e32 v6, v0
-; W64-O0-NEXT: s_or_saveexec_b64 s[16:17], -1
-; W64-O0-NEXT: buffer_load_dword v0, off, s[0:3], s32 ; 4-byte Folded Reload
-; W64-O0-NEXT: s_mov_b64 exec, s[16:17]
+; W64-O0-NEXT: buffer_store_dword v5, off, s[0:3], s32 offset:44 ; 4-byte Folded Spill
+; W64-O0-NEXT: v_mov_b32_e32 v5, v3
+; W64-O0-NEXT: v_mov_b32_e32 v6, v2
+; W64-O0-NEXT: v_mov_b32_e32 v7, v1
+; W64-O0-NEXT: v_mov_b32_e32 v13, v0
+; W64-O0-NEXT: buffer_load_dword v0, off, s[0:3], s32 offset:44 ; 4-byte Folded Reload
; W64-O0-NEXT: ; implicit-def: $sgpr4
; W64-O0-NEXT: ; implicit-def: $sgpr4
; W64-O0-NEXT: ; implicit-def: $sgpr4
; W64-O0-NEXT: ; implicit-def: $sgpr4
-; W64-O0-NEXT: ; kill: def $vgpr1 killed $vgpr1 def $vgpr1_vgpr2_vgpr3_vgpr4 killed $exec
-; W64-O0-NEXT: v_mov_b32_e32 v2, v9
-; W64-O0-NEXT: v_mov_b32_e32 v3, v8
-; W64-O0-NEXT: v_mov_b32_e32 v4, v7
+; W64-O0-NEXT: ; kill: def $vgpr0 killed $vgpr0 def $vgpr0_vgpr1_vgpr2_vgpr3 killed $exec
+; W64-O0-NEXT: v_mov_b32_e32 v1, v11
+; W64-O0-NEXT: v_mov_b32_e32 v2, v10
+; W64-O0-NEXT: v_mov_b32_e32 v3, v9
; W64-O0-NEXT: ; implicit-def: $sgpr4
; W64-O0-NEXT: ; implicit-def: $sgpr4
; W64-O0-NEXT: ; implicit-def: $sgpr4
; W64-O0-NEXT: ; implicit-def: $sgpr4
-; W64-O0-NEXT: ; kill: def $vgpr6 killed $vgpr6 def $vgpr6_vgpr7_vgpr8_vgpr9 killed $exec
-; W64-O0-NEXT: v_mov_b32_e32 v7, v13
-; W64-O0-NEXT: v_mov_b32_e32 v8, v11
-; W64-O0-NEXT: v_mov_b32_e32 v9, v10
-; W64-O0-NEXT: buffer_store_dword v6, off, s[0:3], s32 offset:32 ; 4-byte Folded Spill
+; W64-O0-NEXT: ; kill: def $vgpr13 killed $vgpr13 def $vgpr13_vgpr14_vgpr15_vgpr16 killed $exec
+; W64-O0-NEXT: v_mov_b32_e32 v14, v7
+; W64-O0-NEXT: v_mov_b32_e32 v15, v6
+; W64-O0-NEXT: v_mov_b32_e32 v16, v5
+; W64-O0-NEXT: buffer_store_dword v13, off, s[0:3], s32 offset:28 ; 4-byte Folded Spill
; W64-O0-NEXT: s_waitcnt vmcnt(0)
-; W64-O0-NEXT: buffer_store_dword v7, off, s[0:3], s32 offset:36 ; 4-byte Folded Spill
-; W64-O0-NEXT: buffer_store_dword v8, off, s[0:3], s32 offset:40 ; 4-byte Folded Spill
-; W64-O0-NEXT: buffer_store_dword v9, off, s[0:3], s32 offset:44 ; 4-byte Folded Spill
+; W64-O0-NEXT: buffer_store_dword v14, off, s[0:3], s32 offset:32 ; 4-byte Folded Spill
+; W64-O0-NEXT: buffer_store_dword v15, off, s[0:3], s32 offset:36 ; 4-byte Folded Spill
+; W64-O0-NEXT: buffer_store_dword v16, off, s[0:3], s32 offset:40 ; 4-byte Folded Spill
; W64-O0-NEXT: ; implicit-def: $sgpr4
; W64-O0-NEXT: ; implicit-def: $sgpr4
-; W64-O0-NEXT: ; kill: def $vgpr5 killed $vgpr5 def $vgpr5_vgpr6 killed $exec
-; W64-O0-NEXT: v_mov_b32_e32 v6, v12
+; W64-O0-NEXT: ; kill: def $vgpr4 killed $vgpr4 def $vgpr4_vgpr5 killed $exec
+; W64-O0-NEXT: v_mov_b32_e32 v5, v12
+; W64-O0-NEXT: buffer_store_dword v4, off, s[0:3], s32 offset:20 ; 4-byte Folded Spill
+; W64-O0-NEXT: s_waitcnt vmcnt(0)
; W64-O0-NEXT: buffer_store_dword v5, off, s[0:3], s32 offset:24 ; 4-byte Folded Spill
+; W64-O0-NEXT: buffer_store_dword v0, off, s[0:3], s32 offset:4 ; 4-byte Folded Spill
; W64-O0-NEXT: s_waitcnt vmcnt(0)
-; W64-O0-NEXT: buffer_store_dword v6, off, s[0:3], s32 offset:28 ; 4-byte Folded Spill
; W64-O0-NEXT: buffer_store_dword v1, off, s[0:3], s32 offset:8 ; 4-byte Folded Spill
-; W64-O0-NEXT: s_waitcnt vmcnt(0)
; W64-O0-NEXT: buffer_store_dword v2, off, s[0:3], s32 offset:12 ; 4-byte Folded Spill
; W64-O0-NEXT: buffer_store_dword v3, off, s[0:3], s32 offset:16 ; 4-byte Folded Spill
-; W64-O0-NEXT: buffer_store_dword v4, off, s[0:3], s32 offset:20 ; 4-byte Folded Spill
; W64-O0-NEXT: ; implicit-def: $sgpr4_sgpr5_sgpr6_sgpr7
; W64-O0-NEXT: ;;#ASMSTART
; W64-O0-NEXT: s_mov_b32 s4, 17
; W64-O0-NEXT: ;;#ASMEND
; W64-O0-NEXT: s_mov_b32 s5, s4
-; W64-O0-NEXT: v_writelane_b32 v0, s5, 0
+; W64-O0-NEXT: v_writelane_b32 v8, s5, 0
; W64-O0-NEXT: s_mov_b32 s5, 0
-; W64-O0-NEXT: v_writelane_b32 v0, s5, 1
-; W64-O0-NEXT: v_mov_b32_e32 v1, s4
-; W64-O0-NEXT: buffer_store_dword v1, off, s[0:3], s32 offset:4 ; 4-byte Folded Spill
-; W64-O0-NEXT: s_mov_b64 s[4:5], exec
-; W64-O0-NEXT: v_writelane_b32 v0, s4, 2
-; W64-O0-NEXT: v_writelane_b32 v0, s5, 3
-; W64-O0-NEXT: s_or_saveexec_b64 s[16:17], -1
+; W64-O0-NEXT: v_writelane_b32 v8, s5, 1
+; W64-O0-NEXT: v_mov_b32_e32 v0, s4
; W64-O0-NEXT: buffer_store_dword v0, off, s[0:3], s32 ; 4-byte Folded Spill
-; W64-O0-NEXT: s_mov_b64 exec, s[16:17]
+; W64-O0-NEXT: s_mov_b64 s[4:5], exec
+; W64-O0-NEXT: v_writelane_b32 v8, s4, 2
+; W64-O0-NEXT: v_writelane_b32 v8, s5, 3
; W64-O0-NEXT: .LBB2_1: ; =>This Inner Loop Header: Depth=1
-; W64-O0-NEXT: s_or_saveexec_b64 s[16:17], -1
-; W64-O0-NEXT: buffer_load_dword v0, off, s[0:3], s32 ; 4-byte Folded Reload
-; W64-O0-NEXT: s_mov_b64 exec, s[16:17]
+; W64-O0-NEXT: buffer_load_dword v0, off, s[0:3], s32 offset:28 ; 4-byte Folded Reload
; W64-O0-NEXT: buffer_load_dword v1, off, s[0:3], s32 offset:32 ; 4-byte Folded Reload
; W64-O0-NEXT: buffer_load_dword v2, off, s[0:3], s32 offset:36 ; 4-byte Folded Reload
; W64-O0-NEXT: buffer_load_dword v3, off, s[0:3], s32 offset:40 ; 4-byte Folded Reload
-; W64-O0-NEXT: buffer_load_dword v4, off, s[0:3], s32 offset:44 ; 4-byte Folded Reload
; W64-O0-NEXT: s_waitcnt vmcnt(0)
-; W64-O0-NEXT: v_readfirstlane_b32 s8, v1
-; W64-O0-NEXT: v_readfirstlane_b32 s12, v2
+; W64-O0-NEXT: v_readfirstlane_b32 s8, v0
+; W64-O0-NEXT: v_readfirstlane_b32 s12, v1
; W64-O0-NEXT: s_mov_b32 s4, s8
; W64-O0-NEXT: s_mov_b32 s5, s12
-; W64-O0-NEXT: v_cmp_eq_u64_e64 s[4:5], s[4:5], v[1:2]
-; W64-O0-NEXT: v_readfirstlane_b32 s7, v3
-; W64-O0-NEXT: v_readfirstlane_b32 s6, v4
+; W64-O0-NEXT: v_cmp_eq_u64_e64 s[4:5], s[4:5], v[0:1]
+; W64-O0-NEXT: v_readfirstlane_b32 s7, v2
+; W64-O0-NEXT: v_readfirstlane_b32 s6, v3
; W64-O0-NEXT: s_mov_b32 s10, s7
; W64-O0-NEXT: s_mov_b32 s11, s6
-; W64-O0-NEXT: v_cmp_eq_u64_e64 s[10:11], s[10:11], v[3:4]
+; W64-O0-NEXT: v_cmp_eq_u64_e64 s[10:11], s[10:11], v[2:3]
; W64-O0-NEXT: s_and_b64 s[4:5], s[4:5], s[10:11]
; W64-O0-NEXT: ; kill: def $sgpr8 killed $sgpr8 def $sgpr8_sgpr9_sgpr10_sgpr11
; W64-O0-NEXT: s_mov_b32 s9, s12
; W64-O0-NEXT: s_mov_b32 s10, s7
; W64-O0-NEXT: s_mov_b32 s11, s6
-; W64-O0-NEXT: v_writelane_b32 v0, s8, 4
-; W64-O0-NEXT: v_writelane_b32 v0, s9, 5
-; W64-O0-NEXT: v_writelane_b32 v0, s10, 6
-; W64-O0-NEXT: v_writelane_b32 v0, s11, 7
+; W64-O0-NEXT: v_writelane_b32 v8, s8, 4
+; W64-O0-NEXT: v_writelane_b32 v8, s9, 5
+; W64-O0-NEXT: v_writelane_b32 v8, s10, 6
+; W64-O0-NEXT: v_writelane_b32 v8, s11, 7
; W64-O0-NEXT: s_and_saveexec_b64 s[4:5], s[4:5]
-; W64-O0-NEXT: v_writelane_b32 v0, s4, 8
-; W64-O0-NEXT: v_writelane_b32 v0, s5, 9
-; W64-O0-NEXT: s_or_saveexec_b64 s[16:17], -1
-; W64-O0-NEXT: buffer_store_dword v0, off, s[0:3], s32 ; 4-byte Folded Spill
-; W64-O0-NEXT: s_mov_b64 exec, s[16:17]
+; W64-O0-NEXT: v_writelane_b32 v8, s4, 8
+; W64-O0-NEXT: v_writelane_b32 v8, s5, 9
; W64-O0-NEXT: ; %bb.2: ; in Loop: Header=BB2_1 Depth=1
-; W64-O0-NEXT: buffer_load_dword v0, off, s[0:3], s32 offset:4 ; 4-byte Folded Reload
-; W64-O0-NEXT: s_or_saveexec_b64 s[16:17], -1
-; W64-O0-NEXT: buffer_load_dword v1, off, s[0:3], s32 ; 4-byte Folded Reload
-; W64-O0-NEXT: s_mov_b64 exec, s[16:17]
+; W64-O0-NEXT: buffer_load_dword v0, off, s[0:3], s32 ; 4-byte Folded Reload
+; W64-O0-NEXT: v_readlane_b32 s4, v8, 8
+; W64-O0-NEXT: v_readlane_b32 s5, v8, 9
+; W64-O0-NEXT: v_readlane_b32 s8, v8, 4
+; W64-O0-NEXT: v_readlane_b32 s9, v8, 5
+; W64-O0-NEXT: v_readlane_b32 s10, v8, 6
+; W64-O0-NEXT: v_readlane_b32 s11, v8, 7
+; W64-O0-NEXT: v_readlane_b32 s6, v8, 1
; W64-O0-NEXT: s_waitcnt vmcnt(0)
-; W64-O0-NEXT: v_readlane_b32 s4, v1, 8
-; W64-O0-NEXT: v_readlane_b32 s5, v1, 9
-; W64-O0-NEXT: v_readlane_b32 s8, v1, 4
-; W64-O0-NEXT: v_readlane_b32 s9, v1, 5
-; W64-O0-NEXT: v_readlane_b32 s10, v1, 6
-; W64-O0-NEXT: v_readlane_b32 s11, v1, 7
-; W64-O0-NEXT: v_readlane_b32 s6, v1, 1
-; W64-O0-NEXT: s_nop 4
+; W64-O0-NEXT: s_nop 3
; W64-O0-NEXT: buffer_load_format_x v0, v0, s[8:11], s6 idxen
; W64-O0-NEXT: s_waitcnt vmcnt(0)
-; W64-O0-NEXT: buffer_store_dword v0, off, s[0:3], s32 offset:60 ; 4-byte Folded Spill
+; W64-O0-NEXT: buffer_store_dword v0, off, s[0:3], s32 offset:56 ; 4-byte Folded Spill
; W64-O0-NEXT: s_xor_b64 exec, exec, s[4:5]
; W64-O0-NEXT: s_cbranch_execnz .LBB2_1
; W64-O0-NEXT: ; %bb.3:
-; W64-O0-NEXT: s_or_saveexec_b64 s[16:17], -1
-; W64-O0-NEXT: buffer_load_dword v0, off, s[0:3], s32 ; 4-byte Folded Reload
-; W64-O0-NEXT: s_mov_b64 exec, s[16:17]
-; W64-O0-NEXT: buffer_load_dword v1, off, s[0:3], s32 offset:60 ; 4-byte Folded Reload
-; W64-O0-NEXT: buffer_load_dword v2, off, s[0:3], s32 offset:56 ; 4-byte Folded Reload
-; W64-O0-NEXT: s_waitcnt vmcnt(2)
-; W64-O0-NEXT: v_readlane_b32 s6, v0, 2
-; W64-O0-NEXT: v_readlane_b32 s7, v0, 3
+; W64-O0-NEXT: buffer_load_dword v0, off, s[0:3], s32 offset:56 ; 4-byte Folded Reload
+; W64-O0-NEXT: buffer_load_dword v1, off, s[0:3], s32 offset:52 ; 4-byte Folded Reload
+; W64-O0-NEXT: v_readlane_b32 s6, v8, 2
+; W64-O0-NEXT: v_readlane_b32 s7, v8, 3
; W64-O0-NEXT: s_mov_b64 exec, s[6:7]
-; W64-O0-NEXT: v_readlane_b32 s4, v0, 1
+; W64-O0-NEXT: v_readlane_b32 s4, v8, 1
; W64-O0-NEXT: s_mov_b32 s5, 0x3ff
; W64-O0-NEXT: s_waitcnt vmcnt(0)
-; W64-O0-NEXT: v_and_b32_e64 v2, v2, s5
-; W64-O0-NEXT: v_cmp_eq_u32_e64 s[6:7], v2, s4
-; W64-O0-NEXT: buffer_store_dword v1, off, s[0:3], s32 offset:64 ; 4-byte Folded Spill
+; W64-O0-NEXT: v_and_b32_e64 v1, v1, s5
+; W64-O0-NEXT: v_cmp_eq_u32_e64 s[6:7], v1, s4
+; W64-O0-NEXT: buffer_store_dword v0, off, s[0:3], s32 offset:60 ; 4-byte Folded Spill
; W64-O0-NEXT: s_mov_b64 s[4:5], exec
-; W64-O0-NEXT: v_writelane_b32 v0, s4, 10
-; W64-O0-NEXT: v_writelane_b32 v0, s5, 11
-; W64-O0-NEXT: s_or_saveexec_b64 s[16:17], -1
-; W64-O0-NEXT: buffer_store_dword v0, off, s[0:3], s32 ; 4-byte Folded Spill
-; W64-O0-NEXT: s_mov_b64 exec, s[16:17]
+; W64-O0-NEXT: v_writelane_b32 v8, s4, 10
+; W64-O0-NEXT: v_writelane_b32 v8, s5, 11
; W64-O0-NEXT: s_and_b64 s[4:5], s[4:5], s[6:7]
; W64-O0-NEXT: s_mov_b64 exec, s[4:5]
; W64-O0-NEXT: s_cbranch_execz .LBB2_8
; W64-O0-NEXT: ; %bb.4: ; %bb1
-; W64-O0-NEXT: s_or_saveexec_b64 s[16:17], -1
-; W64-O0-NEXT: buffer_load_dword v0, off, s[0:3], s32 ; 4-byte Folded Reload
-; W64-O0-NEXT: s_mov_b64 exec, s[16:17]
-; W64-O0-NEXT: s_waitcnt vmcnt(0)
-; W64-O0-NEXT: v_readlane_b32 s4, v0, 0
+; W64-O0-NEXT: v_readlane_b32 s4, v8, 0
; W64-O0-NEXT: s_mov_b32 s5, 0
-; W64-O0-NEXT: v_writelane_b32 v0, s5, 12
-; W64-O0-NEXT: v_mov_b32_e32 v1, s4
-; W64-O0-NEXT: buffer_store_dword v1, off, s[0:3], s32 offset:68 ; 4-byte Folded Spill
+; W64-O0-NEXT: v_writelane_b32 v8, s5, 12
+; W64-O0-NEXT: v_mov_b32_e32 v0, s4
+; W64-O0-NEXT: buffer_store_dword v0, off, s[0:3], s32 offset:64 ; 4-byte Folded Spill
; W64-O0-NEXT: ; implicit-def: $sgpr4_sgpr5_sgpr6_sgpr7
; W64-O0-NEXT: s_mov_b64 s[4:5], exec
-; W64-O0-NEXT: v_writelane_b32 v0, s4, 13
-; W64-O0-NEXT: v_writelane_b32 v0, s5, 14
-; W64-O0-NEXT: s_or_saveexec_b64 s[16:17], -1
-; W64-O0-NEXT: buffer_store_dword v0, off, s[0:3], s32 ; 4-byte Folded Spill
-; W64-O0-NEXT: s_mov_b64 exec, s[16:17]
+; W64-O0-NEXT: v_writelane_b32 v8, s4, 13
+; W64-O0-NEXT: v_writelane_b32 v8, s5, 14
; W64-O0-NEXT: .LBB2_5: ; =>This Inner Loop Header: Depth=1
-; W64-O0-NEXT: s_or_saveexec_b64 s[16:17], -1
-; W64-O0-NEXT: buffer_load_dword v0, off, s[0:3], s32 ; 4-byte Folded Reload
-; W64-O0-NEXT: s_mov_b64 exec, s[16:17]
+; W64-O0-NEXT: buffer_load_dword v0, off, s[0:3], s32 offset:4 ; 4-byte Folded Reload
; W64-O0-NEXT: buffer_load_dword v1, off, s[0:3], s32 offset:8 ; 4-byte Folded Reload
; W64-O0-NEXT: buffer_load_dword v2, off, s[0:3], s32 offset:12 ; 4-byte Folded Reload
; W64-O0-NEXT: buffer_load_dword v3, off, s[0:3], s32 offset:16 ; 4-byte Folded Reload
-; W64-O0-NEXT: buffer_load_dword v4, off, s[0:3], s32 offset:20 ; 4-byte Folded Reload
; W64-O0-NEXT: s_waitcnt vmcnt(0)
-; W64-O0-NEXT: v_readfirstlane_b32 s8, v1
-; W64-O0-NEXT: v_readfirstlane_b32 s12, v2
+; W64-O0-NEXT: v_readfirstlane_b32 s8, v0
+; W64-O0-NEXT: v_readfirstlane_b32 s12, v1
; W64-O0-NEXT: s_mov_b32 s4, s8
; W64-O0-NEXT: s_mov_b32 s5, s12
-; W64-O0-NEXT: v_cmp_eq_u64_e64 s[4:5], s[4:5], v[1:2]
-; W64-O0-NEXT: v_readfirstlane_b32 s7, v3
-; W64-O0-NEXT: v_readfirstlane_b32 s6, v4
+; W64-O0-NEXT: v_cmp_eq_u64_e64 s[4:5], s[4:5], v[0:1]
+; W64-O0-NEXT: v_readfirstlane_b32 s7, v2
+; W64-O0-NEXT: v_readfirstlane_b32 s6, v3
; W64-O0-NEXT: s_mov_b32 s10, s7
; W64-O0-NEXT: s_mov_b32 s11, s6
-; W64-O0-NEXT: v_cmp_eq_u64_e64 s[10:11], s[10:11], v[3:4]
+; W64-O0-NEXT: v_cmp_eq_u64_e64 s[10:11], s[10:11], v[2:3]
; W64-O0-NEXT: s_and_b64 s[4:5], s[4:5], s[10:11]
; W64-O0-NEXT: ; kill: def $sgpr8 killed $sgpr8 def $sgpr8_sgpr9_sgpr10_sgpr11
; W64-O0-NEXT: s_mov_b32 s9, s12
; W64-O0-NEXT: s_mov_b32 s10, s7
; W64-O0-NEXT: s_mov_b32 s11, s6
-; W64-O0-NEXT: v_writelane_b32 v0, s8, 15
-; W64-O0-NEXT: v_writelane_b32 v0, s9, 16
-; W64-O0-NEXT: v_writelane_b32 v0, s10, 17
-; W64-O0-NEXT: v_writelane_b32 v0, s11, 18
+; W64-O0-NEXT: v_writelane_b32 v8, s8, 15
+; W64-O0-NEXT: v_writelane_b32 v8, s9, 16
+; W64-O0-NEXT: v_writelane_b32 v8, s10, 17
+; W64-O0-NEXT: v_writelane_b32 v8, s11, 18
; W64-O0-NEXT: s_and_saveexec_b64 s[4:5], s[4:5]
-; W64-O0-NEXT: v_writelane_b32 v0, s4, 19
-; W64-O0-NEXT: v_writelane_b32 v0, s5, 20
-; W64-O0-NEXT: s_or_saveexec_b64 s[16:17], -1
-; W64-O0-NEXT: buffer_store_dword v0, off, s[0:3], s32 ; 4-byte Folded Spill
-; W64-O0-NEXT: s_mov_b64 exec, s[16:17]
+; W64-O0-NEXT: v_writelane_b32 v8, s4, 19
+; W64-O0-NEXT: v_writelane_b32 v8, s5, 20
; W64-O0-NEXT: ; %bb.6: ; in Loop: Header=BB2_5 Depth=1
-; W64-O0-NEXT: buffer_load_dword v0, off, s[0:3], s32 offset:68 ; 4-byte Folded Reload
-; W64-O0-NEXT: s_or_saveexec_b64 s[16:17], -1
-; W64-O0-NEXT: buffer_load_dword v1, off, s[0:3], s32 ; 4-byte Folded Reload
-; W64-O0-NEXT: s_mov_b64 exec, s[16:17]
+; W64-O0-NEXT: buffer_load_dword v0, off, s[0:3], s32 offset:64 ; 4-byte Folded Reload
+; W64-O0-NEXT: v_readlane_b32 s4, v8, 19
+; W64-O0-NEXT: v_readlane_b32 s5, v8, 20
+; W64-O0-NEXT: v_readlane_b32 s8, v8, 15
+; W64-O0-NEXT: v_readlane_b32 s9, v8, 16
+; W64-O0-NEXT: v_readlane_b32 s10, v8, 17
+; W64-O0-NEXT: v_readlane_b32 s11, v8, 18
+; W64-O0-NEXT: v_readlane_b32 s6, v8, 12
; W64-O0-NEXT: s_waitcnt vmcnt(0)
-; W64-O0-NEXT: v_readlane_b32 s4, v1, 19
-; W64-O0-NEXT: v_readlane_b32 s5, v1, 20
-; W64-O0-NEXT: v_readlane_b32 s8, v1, 15
-; W64-O0-NEXT: v_readlane_b32 s9, v1, 16
-; W64-O0-NEXT: v_readlane_b32 s10, v1, 17
-; W64-O0-NEXT: v_readlane_b32 s11, v1, 18
-; W64-O0-NEXT: v_readlane_b32 s6, v1, 12
-; W64-O0-NEXT: s_nop 4
+; W64-O0-NEXT: s_nop 3
; W64-O0-NEXT: buffer_load_format_x v0, v0, s[8:11], s6 idxen
; W64-O0-NEXT: s_waitcnt vmcnt(0)
-; W64-O0-NEXT: buffer_store_dword v0, off, s[0:3], s32 offset:72 ; 4-byte Folded Spill
+; W64-O0-NEXT: buffer_store_dword v0, off, s[0:3], s32 offset:68 ; 4-byte Folded Spill
; W64-O0-NEXT: s_xor_b64 exec, exec, s[4:5]
; W64-O0-NEXT: s_cbranch_execnz .LBB2_5
; W64-O0-NEXT: ; %bb.7:
-; W64-O0-NEXT: buffer_load_dword v0, off, s[0:3], s32 offset:72 ; 4-byte Folded Reload
-; W64-O0-NEXT: s_or_saveexec_b64 s[16:17], -1
-; W64-O0-NEXT: buffer_load_dword v1, off, s[0:3], s32 ; 4-byte Folded Reload
-; W64-O0-NEXT: s_mov_b64 exec, s[16:17]
-; W64-O0-NEXT: s_waitcnt vmcnt(0)
-; W64-O0-NEXT: v_readlane_b32 s4, v1, 13
-; W64-O0-NEXT: v_readlane_b32 s5, v1, 14
+; W64-O0-NEXT: buffer_load_dword v0, off, s[0:3], s32 offset:68 ; 4-byte Folded Reload
+; W64-O0-NEXT: v_readlane_b32 s4, v8, 13
+; W64-O0-NEXT: v_readlane_b32 s5, v8, 14
; W64-O0-NEXT: s_mov_b64 exec, s[4:5]
-; W64-O0-NEXT: buffer_store_dword v0, off, s[0:3], s32 offset:64 ; 4-byte Folded Spill
+; W64-O0-NEXT: s_waitcnt vmcnt(0)
+; W64-O0-NEXT: buffer_store_dword v0, off, s[0:3], s32 offset:60 ; 4-byte Folded Spill
; W64-O0-NEXT: .LBB2_8: ; %bb2
-; W64-O0-NEXT: s_or_saveexec_b64 s[16:17], -1
-; W64-O0-NEXT: buffer_load_dword v0, off, s[0:3], s32 ; 4-byte Folded Reload
-; W64-O0-NEXT: s_mov_b64 exec, s[16:17]
+; W64-O0-NEXT: buffer_load_dword v0, off, s[0:3], s32 offset:20 ; 4-byte Folded Reload
+; W64-O0-NEXT: s_nop 0
; W64-O0-NEXT: buffer_load_dword v1, off, s[0:3], s32 offset:24 ; 4-byte Folded Reload
-; W64-O0-NEXT: buffer_load_dword v2, off, s[0:3], s32 offset:28 ; 4-byte Folded Reload
-; W64-O0-NEXT: buffer_load_dword v3, off, s[0:3], s32 offset:64 ; 4-byte Folded Reload
-; W64-O0-NEXT: s_waitcnt vmcnt(3)
-; W64-O0-NEXT: v_readlane_b32 s4, v0, 10
-; W64-O0-NEXT: v_readlane_b32 s5, v0, 11
+; W64-O0-NEXT: buffer_load_dword v2, off, s[0:3], s32 offset:60 ; 4-byte Folded Reload
+; W64-O0-NEXT: v_readlane_b32 s4, v8, 10
+; W64-O0-NEXT: v_readlane_b32 s5, v8, 11
; W64-O0-NEXT: s_or_b64 exec, exec, s[4:5]
; W64-O0-NEXT: s_waitcnt vmcnt(0)
-; W64-O0-NEXT: global_store_dword v[1:2], v3, off
+; W64-O0-NEXT: global_store_dword v[0:1], v2, off
; W64-O0-NEXT: s_waitcnt vmcnt(0)
-; W64-O0-NEXT: ; kill: killed $vgpr0
; W64-O0-NEXT: s_xor_saveexec_b64 s[4:5], -1
-; W64-O0-NEXT: buffer_load_dword v0, off, s[0:3], s32 offset:76 ; 4-byte Folded Reload
-; W64-O0-NEXT: buffer_load_dword v1, off, s[0:3], s32 offset:80 ; 4-byte Folded Reload
+; W64-O0-NEXT: buffer_load_dword v8, off, s[0:3], s32 offset:72 ; 4-byte Folded Reload
; W64-O0-NEXT: s_mov_b64 exec, s[4:5]
; W64-O0-NEXT: s_waitcnt vmcnt(0)
; W64-O0-NEXT: s_setpc_b64 s[30:31]
diff --git a/llvm/test/CodeGen/AMDGPU/mubuf-legalize-operands.ll b/llvm/test/CodeGen/AMDGPU/mubuf-legalize-operands.ll
--- a/llvm/test/CodeGen/AMDGPU/mubuf-legalize-operands.ll
+++ b/llvm/test/CodeGen/AMDGPU/mubuf-legalize-operands.ll
@@ -140,123 +140,94 @@
; W64-O0: ; %bb.0:
; W64-O0-NEXT: s_waitcnt vmcnt(0) expcnt(0) lgkmcnt(0)
; W64-O0-NEXT: s_xor_saveexec_b64 s[4:5], -1
-; W64-O0-NEXT: buffer_store_dword v0, off, s[0:3], s32 offset:32 ; 4-byte Folded Spill
-; W64-O0-NEXT: buffer_store_dword v1, off, s[0:3], s32 offset:36 ; 4-byte Folded Spill
+; W64-O0-NEXT: buffer_store_dword v5, off, s[0:3], s32 offset:24 ; 4-byte Folded Spill
; W64-O0-NEXT: s_mov_b64 exec, s[4:5]
-; W64-O0-NEXT: ; implicit-def: $vgpr5
-; W64-O0-NEXT: buffer_store_dword v4, off, s[0:3], s32 offset:24 ; 4-byte Folded Spill
-; W64-O0-NEXT: buffer_store_dword v3, off, s[0:3], s32 offset:20 ; 4-byte Folded Spill
+; W64-O0-NEXT: buffer_store_dword v4, off, s[0:3], s32 offset:16 ; 4-byte Folded Spill
; W64-O0-NEXT: v_mov_b32_e32 v6, v2
-; W64-O0-NEXT: buffer_load_dword v2, off, s[0:3], s32 offset:20 ; 4-byte Folded Reload
-; W64-O0-NEXT: v_mov_b32_e32 v3, v1
-; W64-O0-NEXT: v_mov_b32_e32 v1, v0
-; W64-O0-NEXT: s_or_saveexec_b64 s[16:17], -1
-; W64-O0-NEXT: buffer_load_dword v0, off, s[0:3], s32 ; 4-byte Folded Reload
-; W64-O0-NEXT: s_mov_b64 exec, s[16:17]
+; W64-O0-NEXT: v_mov_b32_e32 v2, v1
; W64-O0-NEXT: ; implicit-def: $sgpr4
; W64-O0-NEXT: ; implicit-def: $sgpr4
; W64-O0-NEXT: ; kill: def $vgpr6 killed $vgpr6 def $vgpr6_vgpr7 killed $exec
-; W64-O0-NEXT: s_waitcnt vmcnt(1)
-; W64-O0-NEXT: v_mov_b32_e32 v7, v2
-; W64-O0-NEXT: v_mov_b32_e32 v5, v7
+; W64-O0-NEXT: v_mov_b32_e32 v7, v3
+; W64-O0-NEXT: v_mov_b32_e32 v4, v7
; W64-O0-NEXT: ; kill: def $vgpr6 killed $vgpr6 killed $vgpr6_vgpr7 killed $exec
; W64-O0-NEXT: ; implicit-def: $sgpr4
; W64-O0-NEXT: ; implicit-def: $sgpr4
-; W64-O0-NEXT: ; kill: def $vgpr1 killed $vgpr1 def $vgpr1_vgpr2 killed $exec
-; W64-O0-NEXT: v_mov_b32_e32 v2, v3
-; W64-O0-NEXT: v_mov_b32_e32 v7, v2
-; W64-O0-NEXT: ; kill: def $vgpr1 killed $vgpr1 killed $vgpr1_vgpr2 killed $exec
+; W64-O0-NEXT: ; kill: def $vgpr0 killed $vgpr0 def $vgpr0_vgpr1 killed $exec
+; W64-O0-NEXT: v_mov_b32_e32 v1, v2
+; W64-O0-NEXT: v_mov_b32_e32 v7, v1
+; W64-O0-NEXT: ; kill: def $vgpr0 killed $vgpr0 killed $vgpr0_vgpr1 killed $exec
; W64-O0-NEXT: ; implicit-def: $sgpr4
; W64-O0-NEXT: ; implicit-def: $sgpr4
; W64-O0-NEXT: ; implicit-def: $sgpr4
; W64-O0-NEXT: ; implicit-def: $sgpr4
-; W64-O0-NEXT: ; kill: def $vgpr1 killed $vgpr1 def $vgpr1_vgpr2_vgpr3_vgpr4 killed $exec
-; W64-O0-NEXT: v_mov_b32_e32 v2, v7
-; W64-O0-NEXT: v_mov_b32_e32 v3, v6
-; W64-O0-NEXT: v_mov_b32_e32 v4, v5
-; W64-O0-NEXT: buffer_store_dword v1, off, s[0:3], s32 offset:4 ; 4-byte Folded Spill
+; W64-O0-NEXT: ; kill: def $vgpr0 killed $vgpr0 def $vgpr0_vgpr1_vgpr2_vgpr3 killed $exec
+; W64-O0-NEXT: v_mov_b32_e32 v1, v7
+; W64-O0-NEXT: v_mov_b32_e32 v2, v6
+; W64-O0-NEXT: v_mov_b32_e32 v3, v4
+; W64-O0-NEXT: buffer_store_dword v0, off, s[0:3], s32 ; 4-byte Folded Spill
; W64-O0-NEXT: s_waitcnt vmcnt(0)
+; W64-O0-NEXT: buffer_store_dword v1, off, s[0:3], s32 offset:4 ; 4-byte Folded Spill
; W64-O0-NEXT: buffer_store_dword v2, off, s[0:3], s32 offset:8 ; 4-byte Folded Spill
; W64-O0-NEXT: buffer_store_dword v3, off, s[0:3], s32 offset:12 ; 4-byte Folded Spill
-; W64-O0-NEXT: buffer_store_dword v4, off, s[0:3], s32 offset:16 ; 4-byte Folded Spill
; W64-O0-NEXT: ; implicit-def: $sgpr4_sgpr5
; W64-O0-NEXT: ; implicit-def: $sgpr4_sgpr5
; W64-O0-NEXT: s_mov_b32 s4, 0
-; W64-O0-NEXT: v_writelane_b32 v0, s4, 0
+; W64-O0-NEXT: v_writelane_b32 v5, s4, 0
; W64-O0-NEXT: s_mov_b64 s[4:5], exec
-; W64-O0-NEXT: v_writelane_b32 v0, s4, 1
-; W64-O0-NEXT: v_writelane_b32 v0, s5, 2
-; W64-O0-NEXT: s_or_saveexec_b64 s[16:17], -1
-; W64-O0-NEXT: buffer_store_dword v0, off, s[0:3], s32 ; 4-byte Folded Spill
-; W64-O0-NEXT: s_mov_b64 exec, s[16:17]
+; W64-O0-NEXT: v_writelane_b32 v5, s4, 1
+; W64-O0-NEXT: v_writelane_b32 v5, s5, 2
; W64-O0-NEXT: .LBB0_1: ; =>This Inner Loop Header: Depth=1
-; W64-O0-NEXT: s_or_saveexec_b64 s[16:17], -1
; W64-O0-NEXT: buffer_load_dword v0, off, s[0:3], s32 ; 4-byte Folded Reload
-; W64-O0-NEXT: s_mov_b64 exec, s[16:17]
; W64-O0-NEXT: buffer_load_dword v1, off, s[0:3], s32 offset:4 ; 4-byte Folded Reload
; W64-O0-NEXT: buffer_load_dword v2, off, s[0:3], s32 offset:8 ; 4-byte Folded Reload
; W64-O0-NEXT: buffer_load_dword v3, off, s[0:3], s32 offset:12 ; 4-byte Folded Reload
-; W64-O0-NEXT: buffer_load_dword v4, off, s[0:3], s32 offset:16 ; 4-byte Folded Reload
; W64-O0-NEXT: s_waitcnt vmcnt(0)
-; W64-O0-NEXT: v_readfirstlane_b32 s8, v1
-; W64-O0-NEXT: v_readfirstlane_b32 s12, v2
+; W64-O0-NEXT: v_readfirstlane_b32 s8, v0
+; W64-O0-NEXT: v_readfirstlane_b32 s12, v1
; W64-O0-NEXT: s_mov_b32 s4, s8
; W64-O0-NEXT: s_mov_b32 s5, s12
-; W64-O0-NEXT: v_cmp_eq_u64_e64 s[4:5], s[4:5], v[1:2]
-; W64-O0-NEXT: v_readfirstlane_b32 s7, v3
-; W64-O0-NEXT: v_readfirstlane_b32 s6, v4
+; W64-O0-NEXT: v_cmp_eq_u64_e64 s[4:5], s[4:5], v[0:1]
+; W64-O0-NEXT: v_readfirstlane_b32 s7, v2
+; W64-O0-NEXT: v_readfirstlane_b32 s6, v3
; W64-O0-NEXT: s_mov_b32 s10, s7
; W64-O0-NEXT: s_mov_b32 s11, s6
-; W64-O0-NEXT: v_cmp_eq_u64_e64 s[10:11], s[10:11], v[3:4]
+; W64-O0-NEXT: v_cmp_eq_u64_e64 s[10:11], s[10:11], v[2:3]
; W64-O0-NEXT: s_and_b64 s[4:5], s[4:5], s[10:11]
; W64-O0-NEXT: ; kill: def $sgpr8 killed $sgpr8 def $sgpr8_sgpr9_sgpr10_sgpr11
; W64-O0-NEXT: s_mov_b32 s9, s12
; W64-O0-NEXT: s_mov_b32 s10, s7
; W64-O0-NEXT: s_mov_b32 s11, s6
-; W64-O0-NEXT: v_writelane_b32 v0, s8, 3
-; W64-O0-NEXT: v_writelane_b32 v0, s9, 4
-; W64-O0-NEXT: v_writelane_b32 v0, s10, 5
-; W64-O0-NEXT: v_writelane_b32 v0, s11, 6
+; W64-O0-NEXT: v_writelane_b32 v5, s8, 3
+; W64-O0-NEXT: v_writelane_b32 v5, s9, 4
+; W64-O0-NEXT: v_writelane_b32 v5, s10, 5
+; W64-O0-NEXT: v_writelane_b32 v5, s11, 6
; W64-O0-NEXT: s_and_saveexec_b64 s[4:5], s[4:5]
-; W64-O0-NEXT: v_writelane_b32 v0, s4, 7
-; W64-O0-NEXT: v_writelane_b32 v0, s5, 8
-; W64-O0-NEXT: s_or_saveexec_b64 s[16:17], -1
-; W64-O0-NEXT: buffer_store_dword v0, off, s[0:3], s32 ; 4-byte Folded Spill
-; W64-O0-NEXT: s_mov_b64 exec, s[16:17]
+; W64-O0-NEXT: v_writelane_b32 v5, s4, 7
+; W64-O0-NEXT: v_writelane_b32 v5, s5, 8
; W64-O0-NEXT: ; %bb.2: ; in Loop: Header=BB0_1 Depth=1
-; W64-O0-NEXT: buffer_load_dword v0, off, s[0:3], s32 offset:24 ; 4-byte Folded Reload
-; W64-O0-NEXT: s_or_saveexec_b64 s[16:17], -1
-; W64-O0-NEXT: buffer_load_dword v1, off, s[0:3], s32 ; 4-byte Folded Reload
-; W64-O0-NEXT: s_mov_b64 exec, s[16:17]
+; W64-O0-NEXT: buffer_load_dword v0, off, s[0:3], s32 offset:16 ; 4-byte Folded Reload
+; W64-O0-NEXT: v_readlane_b32 s4, v5, 7
+; W64-O0-NEXT: v_readlane_b32 s5, v5, 8
+; W64-O0-NEXT: v_readlane_b32 s8, v5, 3
+; W64-O0-NEXT: v_readlane_b32 s9, v5, 4
+; W64-O0-NEXT: v_readlane_b32 s10, v5, 5
+; W64-O0-NEXT: v_readlane_b32 s11, v5, 6
+; W64-O0-NEXT: v_readlane_b32 s6, v5, 0
; W64-O0-NEXT: s_waitcnt vmcnt(0)
-; W64-O0-NEXT: v_readlane_b32 s4, v1, 7
-; W64-O0-NEXT: v_readlane_b32 s5, v1, 8
-; W64-O0-NEXT: v_readlane_b32 s8, v1, 3
-; W64-O0-NEXT: v_readlane_b32 s9, v1, 4
-; W64-O0-NEXT: v_readlane_b32 s10, v1, 5
-; W64-O0-NEXT: v_readlane_b32 s11, v1, 6
-; W64-O0-NEXT: v_readlane_b32 s6, v1, 0
-; W64-O0-NEXT: s_nop 4
+; W64-O0-NEXT: s_nop 3
; W64-O0-NEXT: buffer_load_format_x v0, v0, s[8:11], s6 idxen
; W64-O0-NEXT: s_waitcnt vmcnt(0)
-; W64-O0-NEXT: buffer_store_dword v0, off, s[0:3], s32 offset:28 ; 4-byte Folded Spill
+; W64-O0-NEXT: buffer_store_dword v0, off, s[0:3], s32 offset:20 ; 4-byte Folded Spill
; W64-O0-NEXT: s_xor_b64 exec, exec, s[4:5]
; W64-O0-NEXT: s_cbranch_execnz .LBB0_1
; W64-O0-NEXT: ; %bb.3:
-; W64-O0-NEXT: s_or_saveexec_b64 s[16:17], -1
-; W64-O0-NEXT: buffer_load_dword v1, off, s[0:3], s32 ; 4-byte Folded Reload
-; W64-O0-NEXT: s_mov_b64 exec, s[16:17]
-; W64-O0-NEXT: buffer_load_dword v0, off, s[0:3], s32 offset:28 ; 4-byte Folded Reload
-; W64-O0-NEXT: s_waitcnt vmcnt(1)
-; W64-O0-NEXT: v_readlane_b32 s4, v1, 1
-; W64-O0-NEXT: v_readlane_b32 s5, v1, 2
+; W64-O0-NEXT: buffer_load_dword v0, off, s[0:3], s32 offset:20 ; 4-byte Folded Reload
+; W64-O0-NEXT: v_readlane_b32 s4, v5, 1
+; W64-O0-NEXT: v_readlane_b32 s5, v5, 2
; W64-O0-NEXT: s_mov_b64 exec, s[4:5]
-; W64-O0-NEXT: ; kill: killed $vgpr1
; W64-O0-NEXT: s_xor_saveexec_b64 s[4:5], -1
-; W64-O0-NEXT: s_waitcnt vmcnt(0)
-; W64-O0-NEXT: buffer_load_dword v0, off, s[0:3], s32 offset:32 ; 4-byte Folded Reload
-; W64-O0-NEXT: s_nop 0
-; W64-O0-NEXT: buffer_load_dword v1, off, s[0:3], s32 offset:36 ; 4-byte Folded Reload
+; W64-O0-NEXT: buffer_load_dword v5, off, s[0:3], s32 offset:24 ; 4-byte Folded Reload
; W64-O0-NEXT: s_mov_b64 exec, s[4:5]
; W64-O0-NEXT: s_waitcnt vmcnt(0)
; W64-O0-NEXT: s_setpc_b64 s[30:31]
@@ -508,89 +479,86 @@
; W64-O0: ; %bb.0: ; %entry
; W64-O0-NEXT: s_waitcnt vmcnt(0) expcnt(0) lgkmcnt(0)
; W64-O0-NEXT: s_xor_saveexec_b64 s[4:5], -1
-; W64-O0-NEXT: buffer_store_dword v0, off, s[0:3], s32 offset:72 ; 4-byte Folded Spill
-; W64-O0-NEXT: buffer_store_dword v1, off, s[0:3], s32 offset:76 ; 4-byte Folded Spill
+; W64-O0-NEXT: buffer_store_dword v13, off, s[0:3], s32 offset:72 ; 4-byte Folded Spill
; W64-O0-NEXT: s_mov_b64 exec, s[4:5]
-; W64-O0-NEXT: ; implicit-def: $vgpr13
-; W64-O0-NEXT: buffer_store_dword v11, off, s[0:3], s32 offset:56 ; 4-byte Folded Spill
-; W64-O0-NEXT: buffer_store_dword v9, off, s[0:3], s32 offset:52 ; 4-byte Folded Spill
+; W64-O0-NEXT: buffer_store_dword v11, off, s[0:3], s32 offset:52 ; 4-byte Folded Spill
+; W64-O0-NEXT: buffer_store_dword v9, off, s[0:3], s32 offset:48 ; 4-byte Folded Spill
; W64-O0-NEXT: buffer_store_dword v8, off, s[0:3], s32 offset:60 ; 4-byte Folded Spill
-; W64-O0-NEXT: v_mov_b32_e32 v14, v6
-; W64-O0-NEXT: v_mov_b32_e32 v9, v5
-; W64-O0-NEXT: v_mov_b32_e32 v13, v4
-; W64-O0-NEXT: v_mov_b32_e32 v4, v3
-; W64-O0-NEXT: v_mov_b32_e32 v8, v2
-; W64-O0-NEXT: buffer_load_dword v2, off, s[0:3], s32 offset:56 ; 4-byte Folded Reload
-; W64-O0-NEXT: v_mov_b32_e32 v5, v1
+; W64-O0-NEXT: buffer_store_dword v7, off, s[0:3], s32 offset:56 ; 4-byte Folded Spill
+; W64-O0-NEXT: v_mov_b32_e32 v15, v6
+; W64-O0-NEXT: v_mov_b32_e32 v8, v5
+; W64-O0-NEXT: buffer_load_dword v5, off, s[0:3], s32 offset:56 ; 4-byte Folded Reload
+; W64-O0-NEXT: v_mov_b32_e32 v14, v4
+; W64-O0-NEXT: v_mov_b32_e32 v7, v2
+; W64-O0-NEXT: v_mov_b32_e32 v4, v1
; W64-O0-NEXT: buffer_load_dword v1, off, s[0:3], s32 offset:52 ; 4-byte Folded Reload
-; W64-O0-NEXT: v_mov_b32_e32 v3, v0
-; W64-O0-NEXT: s_or_saveexec_b64 s[16:17], -1
-; W64-O0-NEXT: buffer_load_dword v0, off, s[0:3], s32 ; 4-byte Folded Reload
-; W64-O0-NEXT: s_mov_b64 exec, s[16:17]
+; W64-O0-NEXT: v_mov_b32_e32 v2, v0
+; W64-O0-NEXT: buffer_load_dword v0, off, s[0:3], s32 offset:48 ; 4-byte Folded Reload
; W64-O0-NEXT: ; implicit-def: $sgpr4
; W64-O0-NEXT: ; implicit-def: $sgpr4
-; W64-O0-NEXT: ; kill: def $vgpr14 killed $vgpr14 def $vgpr14_vgpr15 killed $exec
-; W64-O0-NEXT: v_mov_b32_e32 v15, v7
+; W64-O0-NEXT: ; kill: def $vgpr15 killed $vgpr15 def $vgpr15_vgpr16 killed $exec
+; W64-O0-NEXT: s_waitcnt vmcnt(2)
+; W64-O0-NEXT: v_mov_b32_e32 v16, v5
+; W64-O0-NEXT: v_mov_b32_e32 v5, v16
; W64-O0-NEXT: v_mov_b32_e32 v6, v15
-; W64-O0-NEXT: v_mov_b32_e32 v7, v14
; W64-O0-NEXT: ; implicit-def: $sgpr4
; W64-O0-NEXT: ; implicit-def: $sgpr4
-; W64-O0-NEXT: ; kill: def $vgpr13 killed $vgpr13 def $vgpr13_vgpr14 killed $exec
-; W64-O0-NEXT: v_mov_b32_e32 v14, v9
-; W64-O0-NEXT: v_mov_b32_e32 v9, v14
-; W64-O0-NEXT: ; kill: def $vgpr13 killed $vgpr13 killed $vgpr13_vgpr14 killed $exec
+; W64-O0-NEXT: ; kill: def $vgpr14 killed $vgpr14 def $vgpr14_vgpr15 killed $exec
+; W64-O0-NEXT: v_mov_b32_e32 v15, v8
+; W64-O0-NEXT: v_mov_b32_e32 v8, v15
+; W64-O0-NEXT: ; kill: def $vgpr14 killed $vgpr14 killed $vgpr14_vgpr15 killed $exec
; W64-O0-NEXT: ; implicit-def: $sgpr4
; W64-O0-NEXT: ; implicit-def: $sgpr4
; W64-O0-NEXT: ; implicit-def: $sgpr4
; W64-O0-NEXT: ; implicit-def: $sgpr4
-; W64-O0-NEXT: ; kill: def $vgpr13 killed $vgpr13 def $vgpr13_vgpr14_vgpr15_vgpr16 killed $exec
-; W64-O0-NEXT: v_mov_b32_e32 v14, v9
-; W64-O0-NEXT: v_mov_b32_e32 v15, v7
+; W64-O0-NEXT: ; kill: def $vgpr14 killed $vgpr14 def $vgpr14_vgpr15_vgpr16_vgpr17 killed $exec
+; W64-O0-NEXT: v_mov_b32_e32 v15, v8
; W64-O0-NEXT: v_mov_b32_e32 v16, v6
-; W64-O0-NEXT: buffer_store_dword v13, off, s[0:3], s32 offset:36 ; 4-byte Folded Spill
+; W64-O0-NEXT: v_mov_b32_e32 v17, v5
+; W64-O0-NEXT: buffer_store_dword v14, off, s[0:3], s32 offset:32 ; 4-byte Folded Spill
; W64-O0-NEXT: s_waitcnt vmcnt(0)
-; W64-O0-NEXT: buffer_store_dword v14, off, s[0:3], s32 offset:40 ; 4-byte Folded Spill
-; W64-O0-NEXT: buffer_store_dword v15, off, s[0:3], s32 offset:44 ; 4-byte Folded Spill
-; W64-O0-NEXT: buffer_store_dword v16, off, s[0:3], s32 offset:48 ; 4-byte Folded Spill
+; W64-O0-NEXT: buffer_store_dword v15, off, s[0:3], s32 offset:36 ; 4-byte Folded Spill
+; W64-O0-NEXT: buffer_store_dword v16, off, s[0:3], s32 offset:40 ; 4-byte Folded Spill
+; W64-O0-NEXT: buffer_store_dword v17, off, s[0:3], s32 offset:44 ; 4-byte Folded Spill
; W64-O0-NEXT: ; implicit-def: $sgpr4
; W64-O0-NEXT: ; implicit-def: $sgpr4
-; W64-O0-NEXT: ; kill: def $vgpr8 killed $vgpr8 def $vgpr8_vgpr9 killed $exec
-; W64-O0-NEXT: v_mov_b32_e32 v9, v4
-; W64-O0-NEXT: v_mov_b32_e32 v7, v9
-; W64-O0-NEXT: ; kill: def $vgpr8 killed $vgpr8 killed $vgpr8_vgpr9 killed $exec
+; W64-O0-NEXT: ; kill: def $vgpr7 killed $vgpr7 def $vgpr7_vgpr8 killed $exec
+; W64-O0-NEXT: v_mov_b32_e32 v8, v3
+; W64-O0-NEXT: v_mov_b32_e32 v6, v8
+; W64-O0-NEXT: ; kill: def $vgpr7 killed $vgpr7 killed $vgpr7_vgpr8 killed $exec
; W64-O0-NEXT: ; implicit-def: $sgpr4
; W64-O0-NEXT: ; implicit-def: $sgpr4
-; W64-O0-NEXT: ; kill: def $vgpr3 killed $vgpr3 def $vgpr3_vgpr4 killed $exec
-; W64-O0-NEXT: v_mov_b32_e32 v4, v5
-; W64-O0-NEXT: v_mov_b32_e32 v9, v4
-; W64-O0-NEXT: ; kill: def $vgpr3 killed $vgpr3 killed $vgpr3_vgpr4 killed $exec
+; W64-O0-NEXT: ; kill: def $vgpr2 killed $vgpr2 def $vgpr2_vgpr3 killed $exec
+; W64-O0-NEXT: v_mov_b32_e32 v3, v4
+; W64-O0-NEXT: v_mov_b32_e32 v8, v3
+; W64-O0-NEXT: ; kill: def $vgpr2 killed $vgpr2 killed $vgpr2_vgpr3 killed $exec
; W64-O0-NEXT: ; implicit-def: $sgpr4
; W64-O0-NEXT: ; implicit-def: $sgpr4
; W64-O0-NEXT: ; implicit-def: $sgpr4
; W64-O0-NEXT: ; implicit-def: $sgpr4
-; W64-O0-NEXT: ; kill: def $vgpr3 killed $vgpr3 def $vgpr3_vgpr4_vgpr5_vgpr6 killed $exec
-; W64-O0-NEXT: v_mov_b32_e32 v4, v9
-; W64-O0-NEXT: v_mov_b32_e32 v5, v8
-; W64-O0-NEXT: v_mov_b32_e32 v6, v7
-; W64-O0-NEXT: buffer_store_dword v3, off, s[0:3], s32 offset:20 ; 4-byte Folded Spill
+; W64-O0-NEXT: ; kill: def $vgpr2 killed $vgpr2 def $vgpr2_vgpr3_vgpr4_vgpr5 killed $exec
+; W64-O0-NEXT: v_mov_b32_e32 v3, v8
+; W64-O0-NEXT: v_mov_b32_e32 v4, v7
+; W64-O0-NEXT: v_mov_b32_e32 v5, v6
+; W64-O0-NEXT: buffer_store_dword v2, off, s[0:3], s32 offset:16 ; 4-byte Folded Spill
; W64-O0-NEXT: s_waitcnt vmcnt(0)
+; W64-O0-NEXT: buffer_store_dword v3, off, s[0:3], s32 offset:20 ; 4-byte Folded Spill
; W64-O0-NEXT: buffer_store_dword v4, off, s[0:3], s32 offset:24 ; 4-byte Folded Spill
; W64-O0-NEXT: buffer_store_dword v5, off, s[0:3], s32 offset:28 ; 4-byte Folded Spill
-; W64-O0-NEXT: buffer_store_dword v6, off, s[0:3], s32 offset:32 ; 4-byte Folded Spill
; W64-O0-NEXT: ; implicit-def: $sgpr4
; W64-O0-NEXT: ; implicit-def: $sgpr4
-; W64-O0-NEXT: ; kill: def $vgpr2 killed $vgpr2 def $vgpr2_vgpr3 killed $exec
-; W64-O0-NEXT: v_mov_b32_e32 v3, v12
-; W64-O0-NEXT: buffer_store_dword v2, off, s[0:3], s32 offset:12 ; 4-byte Folded Spill
+; W64-O0-NEXT: ; kill: def $vgpr1 killed $vgpr1 def $vgpr1_vgpr2 killed $exec
+; W64-O0-NEXT: v_mov_b32_e32 v2, v12
+; W64-O0-NEXT: buffer_store_dword v1, off, s[0:3], s32 offset:8 ; 4-byte Folded Spill
; W64-O0-NEXT: s_waitcnt vmcnt(0)
-; W64-O0-NEXT: buffer_store_dword v3, off, s[0:3], s32 offset:16 ; 4-byte Folded Spill
+; W64-O0-NEXT: buffer_store_dword v2, off, s[0:3], s32 offset:12 ; 4-byte Folded Spill
; W64-O0-NEXT: ; implicit-def: $sgpr4
; W64-O0-NEXT: ; implicit-def: $sgpr4
-; W64-O0-NEXT: ; kill: def $vgpr1 killed $vgpr1 def $vgpr1_vgpr2 killed $exec
-; W64-O0-NEXT: v_mov_b32_e32 v2, v10
-; W64-O0-NEXT: buffer_store_dword v1, off, s[0:3], s32 offset:4 ; 4-byte Folded Spill
+; W64-O0-NEXT: ; kill: def $vgpr0 killed $vgpr0 def $vgpr0_vgpr1 killed $exec
+; W64-O0-NEXT: v_mov_b32_e32 v1, v10
+; W64-O0-NEXT: buffer_store_dword v0, off, s[0:3], s32 ; 4-byte Folded Spill
; W64-O0-NEXT: s_waitcnt vmcnt(0)
-; W64-O0-NEXT: buffer_store_dword v2, off, s[0:3], s32 offset:8 ; 4-byte Folded Spill
+; W64-O0-NEXT: buffer_store_dword v1, off, s[0:3], s32 offset:4 ; 4-byte Folded Spill
; W64-O0-NEXT: ; implicit-def: $sgpr4_sgpr5
; W64-O0-NEXT: ; implicit-def: $sgpr4_sgpr5
; W64-O0-NEXT: ; implicit-def: $sgpr4_sgpr5
@@ -598,156 +566,122 @@
; W64-O0-NEXT: ; implicit-def: $sgpr4_sgpr5
; W64-O0-NEXT: ; implicit-def: $sgpr4_sgpr5
; W64-O0-NEXT: s_mov_b32 s4, 0
-; W64-O0-NEXT: v_writelane_b32 v0, s4, 0
+; W64-O0-NEXT: v_writelane_b32 v13, s4, 0
; W64-O0-NEXT: s_mov_b64 s[4:5], exec
-; W64-O0-NEXT: v_writelane_b32 v0, s4, 1
-; W64-O0-NEXT: v_writelane_b32 v0, s5, 2
-; W64-O0-NEXT: s_or_saveexec_b64 s[16:17], -1
-; W64-O0-NEXT: buffer_store_dword v0, off, s[0:3], s32 ; 4-byte Folded Spill
-; W64-O0-NEXT: s_mov_b64 exec, s[16:17]
+; W64-O0-NEXT: v_writelane_b32 v13, s4, 1
+; W64-O0-NEXT: v_writelane_b32 v13, s5, 2
; W64-O0-NEXT: .LBB1_1: ; =>This Inner Loop Header: Depth=1
-; W64-O0-NEXT: s_or_saveexec_b64 s[16:17], -1
-; W64-O0-NEXT: buffer_load_dword v0, off, s[0:3], s32 ; 4-byte Folded Reload
-; W64-O0-NEXT: s_mov_b64 exec, s[16:17]
+; W64-O0-NEXT: buffer_load_dword v0, off, s[0:3], s32 offset:16 ; 4-byte Folded Reload
; W64-O0-NEXT: buffer_load_dword v1, off, s[0:3], s32 offset:20 ; 4-byte Folded Reload
; W64-O0-NEXT: buffer_load_dword v2, off, s[0:3], s32 offset:24 ; 4-byte Folded Reload
; W64-O0-NEXT: buffer_load_dword v3, off, s[0:3], s32 offset:28 ; 4-byte Folded Reload
-; W64-O0-NEXT: buffer_load_dword v4, off, s[0:3], s32 offset:32 ; 4-byte Folded Reload
; W64-O0-NEXT: s_waitcnt vmcnt(0)
-; W64-O0-NEXT: v_readfirstlane_b32 s8, v1
-; W64-O0-NEXT: v_readfirstlane_b32 s12, v2
+; W64-O0-NEXT: v_readfirstlane_b32 s8, v0
+; W64-O0-NEXT: v_readfirstlane_b32 s12, v1
; W64-O0-NEXT: s_mov_b32 s4, s8
; W64-O0-NEXT: s_mov_b32 s5, s12
-; W64-O0-NEXT: v_cmp_eq_u64_e64 s[4:5], s[4:5], v[1:2]
-; W64-O0-NEXT: v_readfirstlane_b32 s7, v3
-; W64-O0-NEXT: v_readfirstlane_b32 s6, v4
+; W64-O0-NEXT: v_cmp_eq_u64_e64 s[4:5], s[4:5], v[0:1]
+; W64-O0-NEXT: v_readfirstlane_b32 s7, v2
+; W64-O0-NEXT: v_readfirstlane_b32 s6, v3
; W64-O0-NEXT: s_mov_b32 s10, s7
; W64-O0-NEXT: s_mov_b32 s11, s6
-; W64-O0-NEXT: v_cmp_eq_u64_e64 s[10:11], s[10:11], v[3:4]
+; W64-O0-NEXT: v_cmp_eq_u64_e64 s[10:11], s[10:11], v[2:3]
; W64-O0-NEXT: s_and_b64 s[4:5], s[4:5], s[10:11]
; W64-O0-NEXT: ; kill: def $sgpr8 killed $sgpr8 def $sgpr8_sgpr9_sgpr10_sgpr11
; W64-O0-NEXT: s_mov_b32 s9, s12
; W64-O0-NEXT: s_mov_b32 s10, s7
; W64-O0-NEXT: s_mov_b32 s11, s6
-; W64-O0-NEXT: v_writelane_b32 v0, s8, 3
-; W64-O0-NEXT: v_writelane_b32 v0, s9, 4
-; W64-O0-NEXT: v_writelane_b32 v0, s10, 5
-; W64-O0-NEXT: v_writelane_b32 v0, s11, 6
+; W64-O0-NEXT: v_writelane_b32 v13, s8, 3
+; W64-O0-NEXT: v_writelane_b32 v13, s9, 4
+; W64-O0-NEXT: v_writelane_b32 v13, s10, 5
+; W64-O0-NEXT: v_writelane_b32 v13, s11, 6
; W64-O0-NEXT: s_and_saveexec_b64 s[4:5], s[4:5]
-; W64-O0-NEXT: v_writelane_b32 v0, s4, 7
-; W64-O0-NEXT: v_writelane_b32 v0, s5, 8
-; W64-O0-NEXT: s_or_saveexec_b64 s[16:17], -1
-; W64-O0-NEXT: buffer_store_dword v0, off, s[0:3], s32 ; 4-byte Folded Spill
-; W64-O0-NEXT: s_mov_b64 exec, s[16:17]
+; W64-O0-NEXT: v_writelane_b32 v13, s4, 7
+; W64-O0-NEXT: v_writelane_b32 v13, s5, 8
; W64-O0-NEXT: ; %bb.2: ; in Loop: Header=BB1_1 Depth=1
; W64-O0-NEXT: buffer_load_dword v0, off, s[0:3], s32 offset:60 ; 4-byte Folded Reload
-; W64-O0-NEXT: s_or_saveexec_b64 s[16:17], -1
-; W64-O0-NEXT: buffer_load_dword v1, off, s[0:3], s32 ; 4-byte Folded Reload
-; W64-O0-NEXT: s_mov_b64 exec, s[16:17]
+; W64-O0-NEXT: v_readlane_b32 s4, v13, 7
+; W64-O0-NEXT: v_readlane_b32 s5, v13, 8
+; W64-O0-NEXT: v_readlane_b32 s8, v13, 3
+; W64-O0-NEXT: v_readlane_b32 s9, v13, 4
+; W64-O0-NEXT: v_readlane_b32 s10, v13, 5
+; W64-O0-NEXT: v_readlane_b32 s11, v13, 6
+; W64-O0-NEXT: v_readlane_b32 s6, v13, 0
; W64-O0-NEXT: s_waitcnt vmcnt(0)
-; W64-O0-NEXT: v_readlane_b32 s4, v1, 7
-; W64-O0-NEXT: v_readlane_b32 s5, v1, 8
-; W64-O0-NEXT: v_readlane_b32 s8, v1, 3
-; W64-O0-NEXT: v_readlane_b32 s9, v1, 4
-; W64-O0-NEXT: v_readlane_b32 s10, v1, 5
-; W64-O0-NEXT: v_readlane_b32 s11, v1, 6
-; W64-O0-NEXT: v_readlane_b32 s6, v1, 0
-; W64-O0-NEXT: s_nop 4
+; W64-O0-NEXT: s_nop 3
; W64-O0-NEXT: buffer_load_format_x v0, v0, s[8:11], s6 idxen
; W64-O0-NEXT: s_waitcnt vmcnt(0)
; W64-O0-NEXT: buffer_store_dword v0, off, s[0:3], s32 offset:64 ; 4-byte Folded Spill
; W64-O0-NEXT: s_xor_b64 exec, exec, s[4:5]
; W64-O0-NEXT: s_cbranch_execnz .LBB1_1
; W64-O0-NEXT: ; %bb.3:
-; W64-O0-NEXT: s_or_saveexec_b64 s[16:17], -1
-; W64-O0-NEXT: buffer_load_dword v0, off, s[0:3], s32 ; 4-byte Folded Reload
-; W64-O0-NEXT: s_mov_b64 exec, s[16:17]
-; W64-O0-NEXT: s_waitcnt vmcnt(0)
-; W64-O0-NEXT: v_readlane_b32 s4, v0, 1
-; W64-O0-NEXT: v_readlane_b32 s5, v0, 2
+; W64-O0-NEXT: v_readlane_b32 s4, v13, 1
+; W64-O0-NEXT: v_readlane_b32 s5, v13, 2
; W64-O0-NEXT: s_mov_b64 exec, s[4:5]
; W64-O0-NEXT: s_mov_b64 s[4:5], exec
-; W64-O0-NEXT: v_writelane_b32 v0, s4, 9
-; W64-O0-NEXT: v_writelane_b32 v0, s5, 10
-; W64-O0-NEXT: s_or_saveexec_b64 s[16:17], -1
-; W64-O0-NEXT: buffer_store_dword v0, off, s[0:3], s32 ; 4-byte Folded Spill
-; W64-O0-NEXT: s_mov_b64 exec, s[16:17]
+; W64-O0-NEXT: v_writelane_b32 v13, s4, 9
+; W64-O0-NEXT: v_writelane_b32 v13, s5, 10
; W64-O0-NEXT: .LBB1_4: ; =>This Inner Loop Header: Depth=1
-; W64-O0-NEXT: s_or_saveexec_b64 s[16:17], -1
-; W64-O0-NEXT: buffer_load_dword v0, off, s[0:3], s32 ; 4-byte Folded Reload
-; W64-O0-NEXT: s_mov_b64 exec, s[16:17]
+; W64-O0-NEXT: buffer_load_dword v0, off, s[0:3], s32 offset:32 ; 4-byte Folded Reload
; W64-O0-NEXT: buffer_load_dword v1, off, s[0:3], s32 offset:36 ; 4-byte Folded Reload
; W64-O0-NEXT: buffer_load_dword v2, off, s[0:3], s32 offset:40 ; 4-byte Folded Reload
; W64-O0-NEXT: buffer_load_dword v3, off, s[0:3], s32 offset:44 ; 4-byte Folded Reload
-; W64-O0-NEXT: buffer_load_dword v4, off, s[0:3], s32 offset:48 ; 4-byte Folded Reload
; W64-O0-NEXT: s_waitcnt vmcnt(0)
-; W64-O0-NEXT: v_readfirstlane_b32 s8, v1
-; W64-O0-NEXT: v_readfirstlane_b32 s12, v2
+; W64-O0-NEXT: v_readfirstlane_b32 s8, v0
+; W64-O0-NEXT: v_readfirstlane_b32 s12, v1
; W64-O0-NEXT: s_mov_b32 s4, s8
; W64-O0-NEXT: s_mov_b32 s5, s12
-; W64-O0-NEXT: v_cmp_eq_u64_e64 s[4:5], s[4:5], v[1:2]
-; W64-O0-NEXT: v_readfirstlane_b32 s7, v3
-; W64-O0-NEXT: v_readfirstlane_b32 s6, v4
+; W64-O0-NEXT: v_cmp_eq_u64_e64 s[4:5], s[4:5], v[0:1]
+; W64-O0-NEXT: v_readfirstlane_b32 s7, v2
+; W64-O0-NEXT: v_readfirstlane_b32 s6, v3
; W64-O0-NEXT: s_mov_b32 s10, s7
; W64-O0-NEXT: s_mov_b32 s11, s6
-; W64-O0-NEXT: v_cmp_eq_u64_e64 s[10:11], s[10:11], v[3:4]
+; W64-O0-NEXT: v_cmp_eq_u64_e64 s[10:11], s[10:11], v[2:3]
; W64-O0-NEXT: s_and_b64 s[4:5], s[4:5], s[10:11]
; W64-O0-NEXT: ; kill: def $sgpr8 killed $sgpr8 def $sgpr8_sgpr9_sgpr10_sgpr11
; W64-O0-NEXT: s_mov_b32 s9, s12
; W64-O0-NEXT: s_mov_b32 s10, s7
; W64-O0-NEXT: s_mov_b32 s11, s6
-; W64-O0-NEXT: v_writelane_b32 v0, s8, 11
-; W64-O0-NEXT: v_writelane_b32 v0, s9, 12
-; W64-O0-NEXT: v_writelane_b32 v0, s10, 13
-; W64-O0-NEXT: v_writelane_b32 v0, s11, 14
+; W64-O0-NEXT: v_writelane_b32 v13, s8, 11
+; W64-O0-NEXT: v_writelane_b32 v13, s9, 12
+; W64-O0-NEXT: v_writelane_b32 v13, s10, 13
+; W64-O0-NEXT: v_writelane_b32 v13, s11, 14
; W64-O0-NEXT: s_and_saveexec_b64 s[4:5], s[4:5]
-; W64-O0-NEXT: v_writelane_b32 v0, s4, 15
-; W64-O0-NEXT: v_writelane_b32 v0, s5, 16
-; W64-O0-NEXT: s_or_saveexec_b64 s[16:17], -1
-; W64-O0-NEXT: buffer_store_dword v0, off, s[0:3], s32 ; 4-byte Folded Spill
-; W64-O0-NEXT: s_mov_b64 exec, s[16:17]
+; W64-O0-NEXT: v_writelane_b32 v13, s4, 15
+; W64-O0-NEXT: v_writelane_b32 v13, s5, 16
; W64-O0-NEXT: ; %bb.5: ; in Loop: Header=BB1_4 Depth=1
; W64-O0-NEXT: buffer_load_dword v0, off, s[0:3], s32 offset:60 ; 4-byte Folded Reload
-; W64-O0-NEXT: s_or_saveexec_b64 s[16:17], -1
-; W64-O0-NEXT: buffer_load_dword v1, off, s[0:3], s32 ; 4-byte Folded Reload
-; W64-O0-NEXT: s_mov_b64 exec, s[16:17]
+; W64-O0-NEXT: v_readlane_b32 s4, v13, 15
+; W64-O0-NEXT: v_readlane_b32 s5, v13, 16
+; W64-O0-NEXT: v_readlane_b32 s8, v13, 11
+; W64-O0-NEXT: v_readlane_b32 s9, v13, 12
+; W64-O0-NEXT: v_readlane_b32 s10, v13, 13
+; W64-O0-NEXT: v_readlane_b32 s11, v13, 14
+; W64-O0-NEXT: v_readlane_b32 s6, v13, 0
; W64-O0-NEXT: s_waitcnt vmcnt(0)
-; W64-O0-NEXT: v_readlane_b32 s4, v1, 15
-; W64-O0-NEXT: v_readlane_b32 s5, v1, 16
-; W64-O0-NEXT: v_readlane_b32 s8, v1, 11
-; W64-O0-NEXT: v_readlane_b32 s9, v1, 12
-; W64-O0-NEXT: v_readlane_b32 s10, v1, 13
-; W64-O0-NEXT: v_readlane_b32 s11, v1, 14
-; W64-O0-NEXT: v_readlane_b32 s6, v1, 0
-; W64-O0-NEXT: s_nop 4
+; W64-O0-NEXT: s_nop 3
; W64-O0-NEXT: buffer_load_format_x v0, v0, s[8:11], s6 idxen
; W64-O0-NEXT: s_waitcnt vmcnt(0)
; W64-O0-NEXT: buffer_store_dword v0, off, s[0:3], s32 offset:68 ; 4-byte Folded Spill
; W64-O0-NEXT: s_xor_b64 exec, exec, s[4:5]
; W64-O0-NEXT: s_cbranch_execnz .LBB1_4
; W64-O0-NEXT: ; %bb.6:
-; W64-O0-NEXT: s_or_saveexec_b64 s[16:17], -1
-; W64-O0-NEXT: buffer_load_dword v0, off, s[0:3], s32 ; 4-byte Folded Reload
-; W64-O0-NEXT: s_mov_b64 exec, s[16:17]
-; W64-O0-NEXT: buffer_load_dword v3, off, s[0:3], s32 offset:68 ; 4-byte Folded Reload
+; W64-O0-NEXT: buffer_load_dword v2, off, s[0:3], s32 offset:68 ; 4-byte Folded Reload
+; W64-O0-NEXT: buffer_load_dword v0, off, s[0:3], s32 offset:8 ; 4-byte Folded Reload
; W64-O0-NEXT: buffer_load_dword v1, off, s[0:3], s32 offset:12 ; 4-byte Folded Reload
-; W64-O0-NEXT: buffer_load_dword v2, off, s[0:3], s32 offset:16 ; 4-byte Folded Reload
-; W64-O0-NEXT: buffer_load_dword v6, off, s[0:3], s32 offset:64 ; 4-byte Folded Reload
+; W64-O0-NEXT: buffer_load_dword v5, off, s[0:3], s32 offset:64 ; 4-byte Folded Reload
+; W64-O0-NEXT: buffer_load_dword v3, off, s[0:3], s32 ; 4-byte Folded Reload
; W64-O0-NEXT: buffer_load_dword v4, off, s[0:3], s32 offset:4 ; 4-byte Folded Reload
-; W64-O0-NEXT: buffer_load_dword v5, off, s[0:3], s32 offset:8 ; 4-byte Folded Reload
-; W64-O0-NEXT: s_waitcnt vmcnt(6)
-; W64-O0-NEXT: v_readlane_b32 s4, v0, 9
-; W64-O0-NEXT: v_readlane_b32 s5, v0, 10
+; W64-O0-NEXT: v_readlane_b32 s4, v13, 9
+; W64-O0-NEXT: v_readlane_b32 s5, v13, 10
; W64-O0-NEXT: s_mov_b64 exec, s[4:5]
; W64-O0-NEXT: s_waitcnt vmcnt(0)
-; W64-O0-NEXT: global_store_dword v[4:5], v6, off
+; W64-O0-NEXT: global_store_dword v[3:4], v5, off
; W64-O0-NEXT: s_waitcnt vmcnt(0)
-; W64-O0-NEXT: global_store_dword v[1:2], v3, off
+; W64-O0-NEXT: global_store_dword v[0:1], v2, off
; W64-O0-NEXT: s_waitcnt vmcnt(0)
-; W64-O0-NEXT: ; kill: killed $vgpr0
; W64-O0-NEXT: s_xor_saveexec_b64 s[4:5], -1
-; W64-O0-NEXT: buffer_load_dword v0, off, s[0:3], s32 offset:72 ; 4-byte Folded Reload
-; W64-O0-NEXT: buffer_load_dword v1, off, s[0:3], s32 offset:76 ; 4-byte Folded Reload
+; W64-O0-NEXT: buffer_load_dword v13, off, s[0:3], s32 offset:72 ; 4-byte Folded Reload
; W64-O0-NEXT: s_mov_b64 exec, s[4:5]
; W64-O0-NEXT: s_waitcnt vmcnt(0)
; W64-O0-NEXT: s_setpc_b64 s[30:31]
@@ -1055,288 +989,239 @@
; W64-O0: ; %bb.0: ; %entry
; W64-O0-NEXT: s_waitcnt vmcnt(0) expcnt(0) lgkmcnt(0)
; W64-O0-NEXT: s_xor_saveexec_b64 s[4:5], -1
-; W64-O0-NEXT: buffer_store_dword v0, off, s[0:3], s32 offset:100 ; 4-byte Folded Spill
-; W64-O0-NEXT: buffer_store_dword v1, off, s[0:3], s32 offset:104 ; 4-byte Folded Spill
+; W64-O0-NEXT: buffer_store_dword v8, off, s[0:3], s32 offset:96 ; 4-byte Folded Spill
; W64-O0-NEXT: s_mov_b64 exec, s[4:5]
-; W64-O0-NEXT: ; implicit-def: $vgpr8
-; W64-O0-NEXT: buffer_store_dword v31, off, s[0:3], s32 offset:64 ; 4-byte Folded Spill
-; W64-O0-NEXT: buffer_store_dword v11, off, s[0:3], s32 offset:48 ; 4-byte Folded Spill
-; W64-O0-NEXT: buffer_store_dword v6, off, s[0:3], s32 offset:60 ; 4-byte Folded Spill
-; W64-O0-NEXT: v_mov_b32_e32 v6, v5
-; W64-O0-NEXT: buffer_load_dword v5, off, s[0:3], s32 offset:60 ; 4-byte Folded Reload
+; W64-O0-NEXT: buffer_store_dword v31, off, s[0:3], s32 offset:60 ; 4-byte Folded Spill
+; W64-O0-NEXT: buffer_store_dword v11, off, s[0:3], s32 offset:44 ; 4-byte Folded Spill
+; W64-O0-NEXT: buffer_store_dword v6, off, s[0:3], s32 offset:56 ; 4-byte Folded Spill
+; W64-O0-NEXT: buffer_store_dword v5, off, s[0:3], s32 offset:48 ; 4-byte Folded Spill
+; W64-O0-NEXT: v_mov_b32_e32 v5, v4
+; W64-O0-NEXT: buffer_load_dword v4, off, s[0:3], s32 offset:56 ; 4-byte Folded Reload
; W64-O0-NEXT: s_nop 0
-; W64-O0-NEXT: buffer_store_dword v6, off, s[0:3], s32 offset:52 ; 4-byte Folded Spill
-; W64-O0-NEXT: buffer_store_dword v4, off, s[0:3], s32 offset:56 ; 4-byte Folded Spill
-; W64-O0-NEXT: v_mov_b32_e32 v4, v3
-; W64-O0-NEXT: buffer_load_dword v3, off, s[0:3], s32 offset:56 ; 4-byte Folded Reload
-; W64-O0-NEXT: v_mov_b32_e32 v13, v2
+; W64-O0-NEXT: buffer_store_dword v5, off, s[0:3], s32 offset:52 ; 4-byte Folded Spill
+; W64-O0-NEXT: v_mov_b32_e32 v10, v2
; W64-O0-NEXT: buffer_load_dword v2, off, s[0:3], s32 offset:52 ; 4-byte Folded Reload
-; W64-O0-NEXT: v_mov_b32_e32 v10, v1
+; W64-O0-NEXT: v_mov_b32_e32 v6, v1
; W64-O0-NEXT: buffer_load_dword v1, off, s[0:3], s32 offset:48 ; 4-byte Folded Reload
-; W64-O0-NEXT: v_mov_b32_e32 v8, v0
-; W64-O0-NEXT: s_or_saveexec_b64 s[16:17], -1
-; W64-O0-NEXT: buffer_load_dword v0, off, s[0:3], s32 ; 4-byte Folded Reload
-; W64-O0-NEXT: s_mov_b64 exec, s[16:17]
+; W64-O0-NEXT: v_mov_b32_e32 v9, v0
+; W64-O0-NEXT: buffer_load_dword v0, off, s[0:3], s32 offset:44 ; 4-byte Folded Reload
; W64-O0-NEXT: ; implicit-def: $sgpr4
; W64-O0-NEXT: ; implicit-def: $sgpr4
-; W64-O0-NEXT: ; kill: def $vgpr13 killed $vgpr13 def $vgpr13_vgpr14 killed $exec
-; W64-O0-NEXT: v_mov_b32_e32 v14, v4
-; W64-O0-NEXT: v_mov_b32_e32 v4, v14
-; W64-O0-NEXT: v_mov_b32_e32 v6, v13
+; W64-O0-NEXT: ; kill: def $vgpr10 killed $vgpr10 def $vgpr10_vgpr11 killed $exec
+; W64-O0-NEXT: v_mov_b32_e32 v11, v3
+; W64-O0-NEXT: v_mov_b32_e32 v3, v11
+; W64-O0-NEXT: v_mov_b32_e32 v5, v10
; W64-O0-NEXT: ; implicit-def: $sgpr4
; W64-O0-NEXT: ; implicit-def: $sgpr4
-; W64-O0-NEXT: ; kill: def $vgpr8 killed $vgpr8 def $vgpr8_vgpr9 killed $exec
-; W64-O0-NEXT: v_mov_b32_e32 v9, v10
+; W64-O0-NEXT: ; kill: def $vgpr9 killed $vgpr9 def $vgpr9_vgpr10 killed $exec
+; W64-O0-NEXT: v_mov_b32_e32 v10, v6
+; W64-O0-NEXT: v_mov_b32_e32 v6, v10
; W64-O0-NEXT: v_mov_b32_e32 v13, v9
-; W64-O0-NEXT: ; kill: def $vgpr8 killed $vgpr8 killed $vgpr8_vgpr9 killed $exec
; W64-O0-NEXT: ; implicit-def: $sgpr4
; W64-O0-NEXT: ; implicit-def: $sgpr4
; W64-O0-NEXT: ; implicit-def: $sgpr4
; W64-O0-NEXT: ; implicit-def: $sgpr4
-; W64-O0-NEXT: ; kill: def $vgpr8 killed $vgpr8 def $vgpr8_vgpr9_vgpr10_vgpr11 killed $exec
-; W64-O0-NEXT: v_mov_b32_e32 v9, v13
-; W64-O0-NEXT: v_mov_b32_e32 v10, v6
-; W64-O0-NEXT: v_mov_b32_e32 v11, v4
-; W64-O0-NEXT: buffer_store_dword v8, off, s[0:3], s32 offset:32 ; 4-byte Folded Spill
+; W64-O0-NEXT: ; kill: def $vgpr13 killed $vgpr13 def $vgpr13_vgpr14_vgpr15_vgpr16 killed $exec
+; W64-O0-NEXT: v_mov_b32_e32 v14, v6
+; W64-O0-NEXT: v_mov_b32_e32 v15, v5
+; W64-O0-NEXT: v_mov_b32_e32 v16, v3
+; W64-O0-NEXT: buffer_store_dword v13, off, s[0:3], s32 offset:28 ; 4-byte Folded Spill
; W64-O0-NEXT: s_waitcnt vmcnt(0)
-; W64-O0-NEXT: buffer_store_dword v9, off, s[0:3], s32 offset:36 ; 4-byte Folded Spill
-; W64-O0-NEXT: buffer_store_dword v10, off, s[0:3], s32 offset:40 ; 4-byte Folded Spill
-; W64-O0-NEXT: buffer_store_dword v11, off, s[0:3], s32 offset:44 ; 4-byte Folded Spill
+; W64-O0-NEXT: buffer_store_dword v14, off, s[0:3], s32 offset:32 ; 4-byte Folded Spill
+; W64-O0-NEXT: buffer_store_dword v15, off, s[0:3], s32 offset:36 ; 4-byte Folded Spill
+; W64-O0-NEXT: buffer_store_dword v16, off, s[0:3], s32 offset:40 ; 4-byte Folded Spill
; W64-O0-NEXT: ; implicit-def: $sgpr4
; W64-O0-NEXT: ; implicit-def: $sgpr4
-; W64-O0-NEXT: ; kill: def $vgpr5 killed $vgpr5 def $vgpr5_vgpr6 killed $exec
-; W64-O0-NEXT: v_mov_b32_e32 v6, v7
+; W64-O0-NEXT: ; kill: def $vgpr4 killed $vgpr4 def $vgpr4_vgpr5 killed $exec
+; W64-O0-NEXT: v_mov_b32_e32 v5, v7
; W64-O0-NEXT: ; implicit-def: $sgpr4
; W64-O0-NEXT: ; implicit-def: $sgpr4
-; W64-O0-NEXT: ; kill: def $vgpr3 killed $vgpr3 def $vgpr3_vgpr4 killed $exec
-; W64-O0-NEXT: v_mov_b32_e32 v4, v2
+; W64-O0-NEXT: ; kill: def $vgpr2 killed $vgpr2 def $vgpr2_vgpr3 killed $exec
+; W64-O0-NEXT: v_mov_b32_e32 v3, v1
; W64-O0-NEXT: ; implicit-def: $sgpr4
; W64-O0-NEXT: ; implicit-def: $sgpr4
-; W64-O0-NEXT: ; kill: def $vgpr1 killed $vgpr1 def $vgpr1_vgpr2 killed $exec
-; W64-O0-NEXT: v_mov_b32_e32 v2, v12
+; W64-O0-NEXT: ; kill: def $vgpr0 killed $vgpr0 def $vgpr0_vgpr1 killed $exec
+; W64-O0-NEXT: v_mov_b32_e32 v1, v12
+; W64-O0-NEXT: buffer_store_dword v4, off, s[0:3], s32 offset:20 ; 4-byte Folded Spill
+; W64-O0-NEXT: s_waitcnt vmcnt(0)
; W64-O0-NEXT: buffer_store_dword v5, off, s[0:3], s32 offset:24 ; 4-byte Folded Spill
+; W64-O0-NEXT: buffer_store_dword v2, off, s[0:3], s32 offset:12 ; 4-byte Folded Spill
; W64-O0-NEXT: s_waitcnt vmcnt(0)
-; W64-O0-NEXT: buffer_store_dword v6, off, s[0:3], s32 offset:28 ; 4-byte Folded Spill
; W64-O0-NEXT: buffer_store_dword v3, off, s[0:3], s32 offset:16 ; 4-byte Folded Spill
-; W64-O0-NEXT: s_waitcnt vmcnt(0)
-; W64-O0-NEXT: buffer_store_dword v4, off, s[0:3], s32 offset:20 ; 4-byte Folded Spill
; W64-O0-NEXT: ; implicit-def: $sgpr4_sgpr5
; W64-O0-NEXT: ; implicit-def: $sgpr4_sgpr5
-; W64-O0-NEXT: buffer_store_dword v1, off, s[0:3], s32 offset:8 ; 4-byte Folded Spill
+; W64-O0-NEXT: buffer_store_dword v0, off, s[0:3], s32 offset:4 ; 4-byte Folded Spill
; W64-O0-NEXT: s_waitcnt vmcnt(0)
-; W64-O0-NEXT: buffer_store_dword v2, off, s[0:3], s32 offset:12 ; 4-byte Folded Spill
+; W64-O0-NEXT: buffer_store_dword v1, off, s[0:3], s32 offset:8 ; 4-byte Folded Spill
; W64-O0-NEXT: ;;#ASMSTART
; W64-O0-NEXT: s_mov_b32 s4, 17
; W64-O0-NEXT: ;;#ASMEND
; W64-O0-NEXT: s_mov_b32 s5, s4
-; W64-O0-NEXT: v_writelane_b32 v0, s5, 0
+; W64-O0-NEXT: v_writelane_b32 v8, s5, 0
; W64-O0-NEXT: s_mov_b32 s5, 0
-; W64-O0-NEXT: v_writelane_b32 v0, s5, 1
-; W64-O0-NEXT: v_mov_b32_e32 v1, s4
-; W64-O0-NEXT: buffer_store_dword v1, off, s[0:3], s32 offset:4 ; 4-byte Folded Spill
-; W64-O0-NEXT: s_mov_b64 s[4:5], exec
-; W64-O0-NEXT: v_writelane_b32 v0, s4, 2
-; W64-O0-NEXT: v_writelane_b32 v0, s5, 3
-; W64-O0-NEXT: s_or_saveexec_b64 s[16:17], -1
+; W64-O0-NEXT: v_writelane_b32 v8, s5, 1
+; W64-O0-NEXT: v_mov_b32_e32 v0, s4
; W64-O0-NEXT: buffer_store_dword v0, off, s[0:3], s32 ; 4-byte Folded Spill
-; W64-O0-NEXT: s_mov_b64 exec, s[16:17]
+; W64-O0-NEXT: s_mov_b64 s[4:5], exec
+; W64-O0-NEXT: v_writelane_b32 v8, s4, 2
+; W64-O0-NEXT: v_writelane_b32 v8, s5, 3
; W64-O0-NEXT: .LBB2_1: ; =>This Inner Loop Header: Depth=1
-; W64-O0-NEXT: s_or_saveexec_b64 s[16:17], -1
-; W64-O0-NEXT: buffer_load_dword v0, off, s[0:3], s32 ; 4-byte Folded Reload
-; W64-O0-NEXT: s_mov_b64 exec, s[16:17]
+; W64-O0-NEXT: buffer_load_dword v0, off, s[0:3], s32 offset:28 ; 4-byte Folded Reload
; W64-O0-NEXT: buffer_load_dword v1, off, s[0:3], s32 offset:32 ; 4-byte Folded Reload
; W64-O0-NEXT: buffer_load_dword v2, off, s[0:3], s32 offset:36 ; 4-byte Folded Reload
; W64-O0-NEXT: buffer_load_dword v3, off, s[0:3], s32 offset:40 ; 4-byte Folded Reload
-; W64-O0-NEXT: buffer_load_dword v4, off, s[0:3], s32 offset:44 ; 4-byte Folded Reload
; W64-O0-NEXT: s_waitcnt vmcnt(0)
-; W64-O0-NEXT: v_readfirstlane_b32 s8, v1
-; W64-O0-NEXT: v_readfirstlane_b32 s12, v2
+; W64-O0-NEXT: v_readfirstlane_b32 s8, v0
+; W64-O0-NEXT: v_readfirstlane_b32 s12, v1
; W64-O0-NEXT: s_mov_b32 s4, s8
; W64-O0-NEXT: s_mov_b32 s5, s12
-; W64-O0-NEXT: v_cmp_eq_u64_e64 s[4:5], s[4:5], v[1:2]
-; W64-O0-NEXT: v_readfirstlane_b32 s7, v3
-; W64-O0-NEXT: v_readfirstlane_b32 s6, v4
+; W64-O0-NEXT: v_cmp_eq_u64_e64 s[4:5], s[4:5], v[0:1]
+; W64-O0-NEXT: v_readfirstlane_b32 s7, v2
+; W64-O0-NEXT: v_readfirstlane_b32 s6, v3
; W64-O0-NEXT: s_mov_b32 s10, s7
; W64-O0-NEXT: s_mov_b32 s11, s6
-; W64-O0-NEXT: v_cmp_eq_u64_e64 s[10:11], s[10:11], v[3:4]
+; W64-O0-NEXT: v_cmp_eq_u64_e64 s[10:11], s[10:11], v[2:3]
; W64-O0-NEXT: s_and_b64 s[4:5], s[4:5], s[10:11]
; W64-O0-NEXT: ; kill: def $sgpr8 killed $sgpr8 def $sgpr8_sgpr9_sgpr10_sgpr11
; W64-O0-NEXT: s_mov_b32 s9, s12
; W64-O0-NEXT: s_mov_b32 s10, s7
; W64-O0-NEXT: s_mov_b32 s11, s6
-; W64-O0-NEXT: v_writelane_b32 v0, s8, 4
-; W64-O0-NEXT: v_writelane_b32 v0, s9, 5
-; W64-O0-NEXT: v_writelane_b32 v0, s10, 6
-; W64-O0-NEXT: v_writelane_b32 v0, s11, 7
+; W64-O0-NEXT: v_writelane_b32 v8, s8, 4
+; W64-O0-NEXT: v_writelane_b32 v8, s9, 5
+; W64-O0-NEXT: v_writelane_b32 v8, s10, 6
+; W64-O0-NEXT: v_writelane_b32 v8, s11, 7
; W64-O0-NEXT: s_and_saveexec_b64 s[4:5], s[4:5]
-; W64-O0-NEXT: v_writelane_b32 v0, s4, 8
-; W64-O0-NEXT: v_writelane_b32 v0, s5, 9
-; W64-O0-NEXT: s_or_saveexec_b64 s[16:17], -1
-; W64-O0-NEXT: buffer_store_dword v0, off, s[0:3], s32 ; 4-byte Folded Spill
-; W64-O0-NEXT: s_mov_b64 exec, s[16:17]
+; W64-O0-NEXT: v_writelane_b32 v8, s4, 8
+; W64-O0-NEXT: v_writelane_b32 v8, s5, 9
; W64-O0-NEXT: ; %bb.2: ; in Loop: Header=BB2_1 Depth=1
-; W64-O0-NEXT: buffer_load_dword v0, off, s[0:3], s32 offset:4 ; 4-byte Folded Reload
-; W64-O0-NEXT: s_or_saveexec_b64 s[16:17], -1
-; W64-O0-NEXT: buffer_load_dword v1, off, s[0:3], s32 ; 4-byte Folded Reload
-; W64-O0-NEXT: s_mov_b64 exec, s[16:17]
+; W64-O0-NEXT: buffer_load_dword v0, off, s[0:3], s32 ; 4-byte Folded Reload
+; W64-O0-NEXT: v_readlane_b32 s4, v8, 8
+; W64-O0-NEXT: v_readlane_b32 s5, v8, 9
+; W64-O0-NEXT: v_readlane_b32 s8, v8, 4
+; W64-O0-NEXT: v_readlane_b32 s9, v8, 5
+; W64-O0-NEXT: v_readlane_b32 s10, v8, 6
+; W64-O0-NEXT: v_readlane_b32 s11, v8, 7
+; W64-O0-NEXT: v_readlane_b32 s6, v8, 1
; W64-O0-NEXT: s_waitcnt vmcnt(0)
-; W64-O0-NEXT: v_readlane_b32 s4, v1, 8
-; W64-O0-NEXT: v_readlane_b32 s5, v1, 9
-; W64-O0-NEXT: v_readlane_b32 s8, v1, 4
-; W64-O0-NEXT: v_readlane_b32 s9, v1, 5
-; W64-O0-NEXT: v_readlane_b32 s10, v1, 6
-; W64-O0-NEXT: v_readlane_b32 s11, v1, 7
-; W64-O0-NEXT: v_readlane_b32 s6, v1, 1
-; W64-O0-NEXT: s_nop 4
+; W64-O0-NEXT: s_nop 3
; W64-O0-NEXT: buffer_load_format_x v0, v0, s[8:11], s6 idxen
; W64-O0-NEXT: s_waitcnt vmcnt(0)
-; W64-O0-NEXT: buffer_store_dword v0, off, s[0:3], s32 offset:68 ; 4-byte Folded Spill
+; W64-O0-NEXT: buffer_store_dword v0, off, s[0:3], s32 offset:64 ; 4-byte Folded Spill
; W64-O0-NEXT: s_xor_b64 exec, exec, s[4:5]
; W64-O0-NEXT: s_cbranch_execnz .LBB2_1
; W64-O0-NEXT: ; %bb.3:
-; W64-O0-NEXT: s_or_saveexec_b64 s[16:17], -1
-; W64-O0-NEXT: buffer_load_dword v0, off, s[0:3], s32 ; 4-byte Folded Reload
-; W64-O0-NEXT: s_mov_b64 exec, s[16:17]
-; W64-O0-NEXT: buffer_load_dword v1, off, s[0:3], s32 offset:68 ; 4-byte Folded Reload
-; W64-O0-NEXT: buffer_load_dword v2, off, s[0:3], s32 offset:64 ; 4-byte Folded Reload
-; W64-O0-NEXT: s_waitcnt vmcnt(2)
-; W64-O0-NEXT: v_readlane_b32 s6, v0, 2
-; W64-O0-NEXT: v_readlane_b32 s7, v0, 3
+; W64-O0-NEXT: buffer_load_dword v0, off, s[0:3], s32 offset:64 ; 4-byte Folded Reload
+; W64-O0-NEXT: buffer_load_dword v1, off, s[0:3], s32 offset:60 ; 4-byte Folded Reload
+; W64-O0-NEXT: v_readlane_b32 s6, v8, 2
+; W64-O0-NEXT: v_readlane_b32 s7, v8, 3
; W64-O0-NEXT: s_mov_b64 exec, s[6:7]
-; W64-O0-NEXT: v_readlane_b32 s4, v0, 1
+; W64-O0-NEXT: v_readlane_b32 s4, v8, 1
; W64-O0-NEXT: s_mov_b32 s5, 0x3ff
; W64-O0-NEXT: s_waitcnt vmcnt(0)
-; W64-O0-NEXT: v_and_b32_e64 v2, v2, s5
-; W64-O0-NEXT: v_cmp_eq_u32_e64 s[6:7], v2, s4
-; W64-O0-NEXT: buffer_store_dword v1, off, s[0:3], s32 offset:72 ; 4-byte Folded Spill
+; W64-O0-NEXT: v_and_b32_e64 v1, v1, s5
+; W64-O0-NEXT: v_cmp_eq_u32_e64 s[6:7], v1, s4
+; W64-O0-NEXT: buffer_store_dword v0, off, s[0:3], s32 offset:68 ; 4-byte Folded Spill
; W64-O0-NEXT: s_mov_b64 s[4:5], exec
-; W64-O0-NEXT: v_writelane_b32 v0, s4, 10
-; W64-O0-NEXT: v_writelane_b32 v0, s5, 11
-; W64-O0-NEXT: s_or_saveexec_b64 s[16:17], -1
-; W64-O0-NEXT: buffer_store_dword v0, off, s[0:3], s32 ; 4-byte Folded Spill
-; W64-O0-NEXT: s_mov_b64 exec, s[16:17]
+; W64-O0-NEXT: v_writelane_b32 v8, s4, 10
+; W64-O0-NEXT: v_writelane_b32 v8, s5, 11
; W64-O0-NEXT: s_and_b64 s[4:5], s[4:5], s[6:7]
; W64-O0-NEXT: s_mov_b64 exec, s[4:5]
; W64-O0-NEXT: s_cbranch_execz .LBB2_8
; W64-O0-NEXT: ; %bb.4: ; %bb1
-; W64-O0-NEXT: s_or_saveexec_b64 s[16:17], -1
-; W64-O0-NEXT: buffer_load_dword v0, off, s[0:3], s32 ; 4-byte Folded Reload
-; W64-O0-NEXT: s_mov_b64 exec, s[16:17]
+; W64-O0-NEXT: buffer_load_dword v1, off, s[0:3], s32 offset:20 ; 4-byte Folded Reload
; W64-O0-NEXT: buffer_load_dword v2, off, s[0:3], s32 offset:24 ; 4-byte Folded Reload
-; W64-O0-NEXT: buffer_load_dword v3, off, s[0:3], s32 offset:28 ; 4-byte Folded Reload
+; W64-O0-NEXT: buffer_load_dword v3, off, s[0:3], s32 offset:12 ; 4-byte Folded Reload
; W64-O0-NEXT: buffer_load_dword v4, off, s[0:3], s32 offset:16 ; 4-byte Folded Reload
-; W64-O0-NEXT: buffer_load_dword v5, off, s[0:3], s32 offset:20 ; 4-byte Folded Reload
-; W64-O0-NEXT: s_waitcnt vmcnt(4)
-; W64-O0-NEXT: v_readlane_b32 s4, v0, 0
+; W64-O0-NEXT: v_readlane_b32 s4, v8, 0
; W64-O0-NEXT: s_waitcnt vmcnt(0)
-; W64-O0-NEXT: v_mov_b32_e32 v7, v5
-; W64-O0-NEXT: v_mov_b32_e32 v1, v4
-; W64-O0-NEXT: v_mov_b32_e32 v5, v3
-; W64-O0-NEXT: v_mov_b32_e32 v6, v2
+; W64-O0-NEXT: v_mov_b32_e32 v6, v4
+; W64-O0-NEXT: v_mov_b32_e32 v0, v3
+; W64-O0-NEXT: v_mov_b32_e32 v4, v2
+; W64-O0-NEXT: v_mov_b32_e32 v5, v1
; W64-O0-NEXT: ; implicit-def: $sgpr5
; W64-O0-NEXT: ; implicit-def: $sgpr5
; W64-O0-NEXT: ; implicit-def: $sgpr5
; W64-O0-NEXT: ; implicit-def: $sgpr5
-; W64-O0-NEXT: ; kill: def $vgpr1 killed $vgpr1 def $vgpr1_vgpr2_vgpr3_vgpr4 killed $exec
-; W64-O0-NEXT: v_mov_b32_e32 v2, v7
-; W64-O0-NEXT: v_mov_b32_e32 v3, v6
-; W64-O0-NEXT: v_mov_b32_e32 v4, v5
-; W64-O0-NEXT: buffer_store_dword v1, off, s[0:3], s32 offset:80 ; 4-byte Folded Spill
+; W64-O0-NEXT: ; kill: def $vgpr0 killed $vgpr0 def $vgpr0_vgpr1_vgpr2_vgpr3 killed $exec
+; W64-O0-NEXT: v_mov_b32_e32 v1, v6
+; W64-O0-NEXT: v_mov_b32_e32 v2, v5
+; W64-O0-NEXT: v_mov_b32_e32 v3, v4
+; W64-O0-NEXT: buffer_store_dword v0, off, s[0:3], s32 offset:76 ; 4-byte Folded Spill
; W64-O0-NEXT: s_waitcnt vmcnt(0)
+; W64-O0-NEXT: buffer_store_dword v1, off, s[0:3], s32 offset:80 ; 4-byte Folded Spill
; W64-O0-NEXT: buffer_store_dword v2, off, s[0:3], s32 offset:84 ; 4-byte Folded Spill
; W64-O0-NEXT: buffer_store_dword v3, off, s[0:3], s32 offset:88 ; 4-byte Folded Spill
-; W64-O0-NEXT: buffer_store_dword v4, off, s[0:3], s32 offset:92 ; 4-byte Folded Spill
; W64-O0-NEXT: s_mov_b32 s5, 0
-; W64-O0-NEXT: v_writelane_b32 v0, s5, 12
-; W64-O0-NEXT: v_mov_b32_e32 v1, s4
-; W64-O0-NEXT: buffer_store_dword v1, off, s[0:3], s32 offset:76 ; 4-byte Folded Spill
+; W64-O0-NEXT: v_writelane_b32 v8, s5, 12
+; W64-O0-NEXT: v_mov_b32_e32 v0, s4
+; W64-O0-NEXT: buffer_store_dword v0, off, s[0:3], s32 offset:72 ; 4-byte Folded Spill
; W64-O0-NEXT: s_mov_b64 s[4:5], exec
-; W64-O0-NEXT: v_writelane_b32 v0, s4, 13
-; W64-O0-NEXT: v_writelane_b32 v0, s5, 14
-; W64-O0-NEXT: s_or_saveexec_b64 s[16:17], -1
-; W64-O0-NEXT: buffer_store_dword v0, off, s[0:3], s32 ; 4-byte Folded Spill
-; W64-O0-NEXT: s_mov_b64 exec, s[16:17]
+; W64-O0-NEXT: v_writelane_b32 v8, s4, 13
+; W64-O0-NEXT: v_writelane_b32 v8, s5, 14
; W64-O0-NEXT: .LBB2_5: ; =>This Inner Loop Header: Depth=1
-; W64-O0-NEXT: s_or_saveexec_b64 s[16:17], -1
-; W64-O0-NEXT: buffer_load_dword v0, off, s[0:3], s32 ; 4-byte Folded Reload
-; W64-O0-NEXT: s_mov_b64 exec, s[16:17]
+; W64-O0-NEXT: buffer_load_dword v0, off, s[0:3], s32 offset:76 ; 4-byte Folded Reload
; W64-O0-NEXT: buffer_load_dword v1, off, s[0:3], s32 offset:80 ; 4-byte Folded Reload
; W64-O0-NEXT: buffer_load_dword v2, off, s[0:3], s32 offset:84 ; 4-byte Folded Reload
; W64-O0-NEXT: buffer_load_dword v3, off, s[0:3], s32 offset:88 ; 4-byte Folded Reload
-; W64-O0-NEXT: buffer_load_dword v4, off, s[0:3], s32 offset:92 ; 4-byte Folded Reload
; W64-O0-NEXT: s_waitcnt vmcnt(0)
-; W64-O0-NEXT: v_readfirstlane_b32 s8, v1
-; W64-O0-NEXT: v_readfirstlane_b32 s12, v2
+; W64-O0-NEXT: v_readfirstlane_b32 s8, v0
+; W64-O0-NEXT: v_readfirstlane_b32 s12, v1
; W64-O0-NEXT: s_mov_b32 s4, s8
; W64-O0-NEXT: s_mov_b32 s5, s12
-; W64-O0-NEXT: v_cmp_eq_u64_e64 s[4:5], s[4:5], v[1:2]
-; W64-O0-NEXT: v_readfirstlane_b32 s7, v3
-; W64-O0-NEXT: v_readfirstlane_b32 s6, v4
+; W64-O0-NEXT: v_cmp_eq_u64_e64 s[4:5], s[4:5], v[0:1]
+; W64-O0-NEXT: v_readfirstlane_b32 s7, v2
+; W64-O0-NEXT: v_readfirstlane_b32 s6, v3
; W64-O0-NEXT: s_mov_b32 s10, s7
; W64-O0-NEXT: s_mov_b32 s11, s6
-; W64-O0-NEXT: v_cmp_eq_u64_e64 s[10:11], s[10:11], v[3:4]
+; W64-O0-NEXT: v_cmp_eq_u64_e64 s[10:11], s[10:11], v[2:3]
; W64-O0-NEXT: s_and_b64 s[4:5], s[4:5], s[10:11]
; W64-O0-NEXT: ; kill: def $sgpr8 killed $sgpr8 def $sgpr8_sgpr9_sgpr10_sgpr11
; W64-O0-NEXT: s_mov_b32 s9, s12
; W64-O0-NEXT: s_mov_b32 s10, s7
; W64-O0-NEXT: s_mov_b32 s11, s6
-; W64-O0-NEXT: v_writelane_b32 v0, s8, 15
-; W64-O0-NEXT: v_writelane_b32 v0, s9, 16
-; W64-O0-NEXT: v_writelane_b32 v0, s10, 17
-; W64-O0-NEXT: v_writelane_b32 v0, s11, 18
+; W64-O0-NEXT: v_writelane_b32 v8, s8, 15
+; W64-O0-NEXT: v_writelane_b32 v8, s9, 16
+; W64-O0-NEXT: v_writelane_b32 v8, s10, 17
+; W64-O0-NEXT: v_writelane_b32 v8, s11, 18
; W64-O0-NEXT: s_and_saveexec_b64 s[4:5], s[4:5]
-; W64-O0-NEXT: v_writelane_b32 v0, s4, 19
-; W64-O0-NEXT: v_writelane_b32 v0, s5, 20
-; W64-O0-NEXT: s_or_saveexec_b64 s[16:17], -1
-; W64-O0-NEXT: buffer_store_dword v0, off, s[0:3], s32 ; 4-byte Folded Spill
-; W64-O0-NEXT: s_mov_b64 exec, s[16:17]
+; W64-O0-NEXT: v_writelane_b32 v8, s4, 19
+; W64-O0-NEXT: v_writelane_b32 v8, s5, 20
; W64-O0-NEXT: ; %bb.6: ; in Loop: Header=BB2_5 Depth=1
-; W64-O0-NEXT: buffer_load_dword v0, off, s[0:3], s32 offset:76 ; 4-byte Folded Reload
-; W64-O0-NEXT: s_or_saveexec_b64 s[16:17], -1
-; W64-O0-NEXT: buffer_load_dword v1, off, s[0:3], s32 ; 4-byte Folded Reload
-; W64-O0-NEXT: s_mov_b64 exec, s[16:17]
+; W64-O0-NEXT: buffer_load_dword v0, off, s[0:3], s32 offset:72 ; 4-byte Folded Reload
+; W64-O0-NEXT: v_readlane_b32 s4, v8, 19
+; W64-O0-NEXT: v_readlane_b32 s5, v8, 20
+; W64-O0-NEXT: v_readlane_b32 s8, v8, 15
+; W64-O0-NEXT: v_readlane_b32 s9, v8, 16
+; W64-O0-NEXT: v_readlane_b32 s10, v8, 17
+; W64-O0-NEXT: v_readlane_b32 s11, v8, 18
+; W64-O0-NEXT: v_readlane_b32 s6, v8, 12
; W64-O0-NEXT: s_waitcnt vmcnt(0)
-; W64-O0-NEXT: v_readlane_b32 s4, v1, 19
-; W64-O0-NEXT: v_readlane_b32 s5, v1, 20
-; W64-O0-NEXT: v_readlane_b32 s8, v1, 15
-; W64-O0-NEXT: v_readlane_b32 s9, v1, 16
-; W64-O0-NEXT: v_readlane_b32 s10, v1, 17
-; W64-O0-NEXT: v_readlane_b32 s11, v1, 18
-; W64-O0-NEXT: v_readlane_b32 s6, v1, 12
-; W64-O0-NEXT: s_nop 4
+; W64-O0-NEXT: s_nop 3
; W64-O0-NEXT: buffer_load_format_x v0, v0, s[8:11], s6 idxen
; W64-O0-NEXT: s_waitcnt vmcnt(0)
-; W64-O0-NEXT: buffer_store_dword v0, off, s[0:3], s32 offset:96 ; 4-byte Folded Spill
+; W64-O0-NEXT: buffer_store_dword v0, off, s[0:3], s32 offset:92 ; 4-byte Folded Spill
; W64-O0-NEXT: s_xor_b64 exec, exec, s[4:5]
; W64-O0-NEXT: s_cbranch_execnz .LBB2_5
; W64-O0-NEXT: ; %bb.7:
-; W64-O0-NEXT: buffer_load_dword v0, off, s[0:3], s32 offset:96 ; 4-byte Folded Reload
-; W64-O0-NEXT: s_or_saveexec_b64 s[16:17], -1
-; W64-O0-NEXT: buffer_load_dword v1, off, s[0:3], s32 ; 4-byte Folded Reload
-; W64-O0-NEXT: s_mov_b64 exec, s[16:17]
-; W64-O0-NEXT: s_waitcnt vmcnt(0)
-; W64-O0-NEXT: v_readlane_b32 s4, v1, 13
-; W64-O0-NEXT: v_readlane_b32 s5, v1, 14
+; W64-O0-NEXT: buffer_load_dword v0, off, s[0:3], s32 offset:92 ; 4-byte Folded Reload
+; W64-O0-NEXT: v_readlane_b32 s4, v8, 13
+; W64-O0-NEXT: v_readlane_b32 s5, v8, 14
; W64-O0-NEXT: s_mov_b64 exec, s[4:5]
-; W64-O0-NEXT: buffer_store_dword v0, off, s[0:3], s32 offset:72 ; 4-byte Folded Spill
+; W64-O0-NEXT: s_waitcnt vmcnt(0)
+; W64-O0-NEXT: buffer_store_dword v0, off, s[0:3], s32 offset:68 ; 4-byte Folded Spill
; W64-O0-NEXT: .LBB2_8: ; %bb2
-; W64-O0-NEXT: s_or_saveexec_b64 s[16:17], -1
-; W64-O0-NEXT: buffer_load_dword v0, off, s[0:3], s32 ; 4-byte Folded Reload
-; W64-O0-NEXT: s_mov_b64 exec, s[16:17]
+; W64-O0-NEXT: buffer_load_dword v0, off, s[0:3], s32 offset:4 ; 4-byte Folded Reload
+; W64-O0-NEXT: s_nop 0
; W64-O0-NEXT: buffer_load_dword v1, off, s[0:3], s32 offset:8 ; 4-byte Folded Reload
-; W64-O0-NEXT: buffer_load_dword v2, off, s[0:3], s32 offset:12 ; 4-byte Folded Reload
-; W64-O0-NEXT: buffer_load_dword v3, off, s[0:3], s32 offset:72 ; 4-byte Folded Reload
-; W64-O0-NEXT: s_waitcnt vmcnt(3)
-; W64-O0-NEXT: v_readlane_b32 s4, v0, 10
-; W64-O0-NEXT: v_readlane_b32 s5, v0, 11
+; W64-O0-NEXT: buffer_load_dword v2, off, s[0:3], s32 offset:68 ; 4-byte Folded Reload
+; W64-O0-NEXT: v_readlane_b32 s4, v8, 10
+; W64-O0-NEXT: v_readlane_b32 s5, v8, 11
; W64-O0-NEXT: s_or_b64 exec, exec, s[4:5]
; W64-O0-NEXT: s_waitcnt vmcnt(0)
-; W64-O0-NEXT: global_store_dword v[1:2], v3, off
+; W64-O0-NEXT: global_store_dword v[0:1], v2, off
; W64-O0-NEXT: s_waitcnt vmcnt(0)
-; W64-O0-NEXT: ; kill: killed $vgpr0
; W64-O0-NEXT: s_xor_saveexec_b64 s[4:5], -1
-; W64-O0-NEXT: buffer_load_dword v0, off, s[0:3], s32 offset:100 ; 4-byte Folded Reload
-; W64-O0-NEXT: buffer_load_dword v1, off, s[0:3], s32 offset:104 ; 4-byte Folded Reload
+; W64-O0-NEXT: buffer_load_dword v8, off, s[0:3], s32 offset:96 ; 4-byte Folded Reload
; W64-O0-NEXT: s_mov_b64 exec, s[4:5]
; W64-O0-NEXT: s_waitcnt vmcnt(0)
; W64-O0-NEXT: s_setpc_b64 s[30:31]
diff --git a/llvm/test/CodeGen/AMDGPU/mul24-pass-ordering.ll b/llvm/test/CodeGen/AMDGPU/mul24-pass-ordering.ll
--- a/llvm/test/CodeGen/AMDGPU/mul24-pass-ordering.ll
+++ b/llvm/test/CodeGen/AMDGPU/mul24-pass-ordering.ll
@@ -192,12 +192,13 @@
; GFX9-NEXT: s_mov_b32 s33, s32
; GFX9-NEXT: s_or_saveexec_b64 s[6:7], -1
; GFX9-NEXT: buffer_store_dword v40, off, s[0:3], s33 offset:12 ; 4-byte Folded Spill
+; GFX9-NEXT: buffer_store_dword v44, off, s[0:3], s33 offset:16 ; 4-byte Folded Spill
; GFX9-NEXT: s_mov_b64 exec, s[6:7]
-; GFX9-NEXT: v_writelane_b32 v40, s4, 5
; GFX9-NEXT: v_writelane_b32 v40, s30, 0
; GFX9-NEXT: v_writelane_b32 v40, s31, 1
; GFX9-NEXT: s_addk_i32 s32, 0x800
; GFX9-NEXT: v_writelane_b32 v40, s34, 2
+; GFX9-NEXT: v_writelane_b32 v44, s4, 0
; GFX9-NEXT: v_writelane_b32 v40, s36, 3
; GFX9-NEXT: s_getpc_b64 s[4:5]
; GFX9-NEXT: s_add_u32 s4, s4, foo@gotpcrel32@lo+4
@@ -229,9 +230,10 @@
; GFX9-NEXT: v_readlane_b32 s34, v40, 2
; GFX9-NEXT: v_readlane_b32 s31, v40, 1
; GFX9-NEXT: v_readlane_b32 s30, v40, 0
-; GFX9-NEXT: v_readlane_b32 s4, v40, 5
+; GFX9-NEXT: v_readlane_b32 s4, v44, 0
; GFX9-NEXT: s_or_saveexec_b64 s[6:7], -1
; GFX9-NEXT: buffer_load_dword v40, off, s[0:3], s33 offset:12 ; 4-byte Folded Reload
+; GFX9-NEXT: buffer_load_dword v44, off, s[0:3], s33 offset:16 ; 4-byte Folded Reload
; GFX9-NEXT: s_mov_b64 exec, s[6:7]
; GFX9-NEXT: s_addk_i32 s32, 0xf800
; GFX9-NEXT: s_mov_b32 s33, s4
diff --git a/llvm/test/CodeGen/AMDGPU/need-fp-from-vgpr-spills.ll b/llvm/test/CodeGen/AMDGPU/need-fp-from-vgpr-spills.ll
--- a/llvm/test/CodeGen/AMDGPU/need-fp-from-vgpr-spills.ll
+++ b/llvm/test/CodeGen/AMDGPU/need-fp-from-vgpr-spills.ll
@@ -27,7 +27,7 @@
; CHECK-LABEL: csr_vgpr_spill_fp_callee:
; CHECK: ; %bb.0: ; %bb
; CHECK-NEXT: s_waitcnt vmcnt(0) expcnt(0) lgkmcnt(0)
-; CHECK-NEXT: s_mov_b32 s18, s33
+; CHECK-NEXT: s_mov_b32 s24, s33
; CHECK-NEXT: s_mov_b32 s33, s32
; CHECK-NEXT: s_xor_saveexec_b64 s[16:17], -1
; CHECK-NEXT: buffer_store_dword v1, off, s[0:3], s33 offset:4 ; 4-byte Folded Spill
@@ -54,7 +54,7 @@
; CHECK-NEXT: buffer_load_dword v1, off, s[0:3], s33 offset:4 ; 4-byte Folded Reload
; CHECK-NEXT: s_mov_b64 exec, s[4:5]
; CHECK-NEXT: s_add_i32 s32, s32, 0xfffffc00
-; CHECK-NEXT: s_mov_b32 s33, s18
+; CHECK-NEXT: s_mov_b32 s33, s24
; CHECK-NEXT: s_waitcnt vmcnt(0)
; CHECK-NEXT: s_setpc_b64 s[30:31]
bb:
@@ -66,16 +66,12 @@
define amdgpu_kernel void @kernel_call() {
; CHECK-LABEL: kernel_call:
; CHECK: ; %bb.0: ; %bb
-; CHECK-NEXT: s_mov_b32 s32, 0x400
+; CHECK-NEXT: s_mov_b32 s32, 0
; CHECK-NEXT: s_add_u32 flat_scratch_lo, s12, s17
; CHECK-NEXT: s_addc_u32 flat_scratch_hi, s13, 0
; CHECK-NEXT: s_add_u32 s0, s0, s17
; CHECK-NEXT: s_addc_u32 s1, s1, 0
-; CHECK-NEXT: ; implicit-def: $vgpr3
; CHECK-NEXT: v_writelane_b32 v3, s16, 0
-; CHECK-NEXT: s_or_saveexec_b64 s[24:25], -1
-; CHECK-NEXT: buffer_store_dword v3, off, s[0:3], 0 offset:4 ; 4-byte Folded Spill
-; CHECK-NEXT: s_mov_b64 exec, s[24:25]
; CHECK-NEXT: s_mov_b32 s13, s15
; CHECK-NEXT: s_mov_b32 s12, s14
; CHECK-NEXT: v_readlane_b32 s14, v3, 0
@@ -93,10 +89,6 @@
; CHECK-NEXT: s_mov_b64 s[0:1], s[20:21]
; CHECK-NEXT: s_mov_b64 s[2:3], s[22:23]
; CHECK-NEXT: s_swappc_b64 s[30:31], s[16:17]
-; CHECK-NEXT: s_or_saveexec_b64 s[24:25], -1
-; CHECK-NEXT: buffer_load_dword v0, off, s[0:3], 0 offset:4 ; 4-byte Folded Reload
-; CHECK-NEXT: s_mov_b64 exec, s[24:25]
-; CHECK-NEXT: ; kill: killed $vgpr0
; CHECK-NEXT: s_endpgm
bb:
tail call fastcc void @csr_vgpr_spill_fp_callee()
@@ -121,9 +113,9 @@
; CHECK-NEXT: s_addc_u32 s17, s17, callee_has_fp@rel32@hi+12
; CHECK-NEXT: v_readlane_b32 s33, v1, 0
; CHECK-NEXT: buffer_load_dword v40, off, s[0:3], s32 ; 4-byte Folded Reload
-; CHECK-NEXT: s_xor_saveexec_b64 s[18:19], -1
+; CHECK-NEXT: s_xor_saveexec_b64 s[20:21], -1
; CHECK-NEXT: buffer_load_dword v1, off, s[0:3], s32 offset:4 ; 4-byte Folded Reload
-; CHECK-NEXT: s_mov_b64 exec, s[18:19]
+; CHECK-NEXT: s_mov_b64 exec, s[20:21]
; CHECK-NEXT: s_setpc_b64 s[16:17]
bb:
call void asm sideeffect "; clobber csr v40", "~{v40}"()
@@ -134,16 +126,12 @@
define amdgpu_kernel void @kernel_tailcall() {
; CHECK-LABEL: kernel_tailcall:
; CHECK: ; %bb.0: ; %bb
-; CHECK-NEXT: s_mov_b32 s32, 0x400
+; CHECK-NEXT: s_mov_b32 s32, 0
; CHECK-NEXT: s_add_u32 flat_scratch_lo, s12, s17
; CHECK-NEXT: s_addc_u32 flat_scratch_hi, s13, 0
; CHECK-NEXT: s_add_u32 s0, s0, s17
; CHECK-NEXT: s_addc_u32 s1, s1, 0
-; CHECK-NEXT: ; implicit-def: $vgpr3
; CHECK-NEXT: v_writelane_b32 v3, s16, 0
-; CHECK-NEXT: s_or_saveexec_b64 s[24:25], -1
-; CHECK-NEXT: buffer_store_dword v3, off, s[0:3], 0 offset:4 ; 4-byte Folded Spill
-; CHECK-NEXT: s_mov_b64 exec, s[24:25]
; CHECK-NEXT: s_mov_b32 s13, s15
; CHECK-NEXT: s_mov_b32 s12, s14
; CHECK-NEXT: v_readlane_b32 s14, v3, 0
@@ -161,10 +149,6 @@
; CHECK-NEXT: s_mov_b64 s[0:1], s[20:21]
; CHECK-NEXT: s_mov_b64 s[2:3], s[22:23]
; CHECK-NEXT: s_swappc_b64 s[30:31], s[16:17]
-; CHECK-NEXT: s_or_saveexec_b64 s[24:25], -1
-; CHECK-NEXT: buffer_load_dword v0, off, s[0:3], 0 offset:4 ; 4-byte Folded Reload
-; CHECK-NEXT: s_mov_b64 exec, s[24:25]
-; CHECK-NEXT: ; kill: killed $vgpr0
; CHECK-NEXT: s_endpgm
bb:
tail call fastcc void @csr_vgpr_spill_fp_tailcall_callee()
@@ -188,7 +172,7 @@
; CHECK-LABEL: caller_save_vgpr_spill_fp_tail_call:
; CHECK: ; %bb.0: ; %entry
; CHECK-NEXT: s_waitcnt vmcnt(0) expcnt(0) lgkmcnt(0)
-; CHECK-NEXT: s_mov_b32 s18, s33
+; CHECK-NEXT: s_mov_b32 s24, s33
; CHECK-NEXT: s_mov_b32 s33, s32
; CHECK-NEXT: s_xor_saveexec_b64 s[16:17], -1
; CHECK-NEXT: buffer_store_dword v1, off, s[0:3], s33 ; 4-byte Folded Spill
@@ -210,7 +194,7 @@
; CHECK-NEXT: buffer_load_dword v1, off, s[0:3], s33 ; 4-byte Folded Reload
; CHECK-NEXT: s_mov_b64 exec, s[4:5]
; CHECK-NEXT: s_add_i32 s32, s32, 0xfffffc00
-; CHECK-NEXT: s_mov_b32 s33, s18
+; CHECK-NEXT: s_mov_b32 s33, s24
; CHECK-NEXT: s_waitcnt vmcnt(0)
; CHECK-NEXT: s_setpc_b64 s[30:31]
entry:
@@ -222,7 +206,7 @@
; CHECK-LABEL: caller_save_vgpr_spill_fp:
; CHECK: ; %bb.0: ; %entry
; CHECK-NEXT: s_waitcnt vmcnt(0) expcnt(0) lgkmcnt(0)
-; CHECK-NEXT: s_mov_b32 s19, s33
+; CHECK-NEXT: s_mov_b32 s25, s33
; CHECK-NEXT: s_mov_b32 s33, s32
; CHECK-NEXT: s_xor_saveexec_b64 s[16:17], -1
; CHECK-NEXT: buffer_store_dword v2, off, s[0:3], s33 ; 4-byte Folded Spill
@@ -244,7 +228,7 @@
; CHECK-NEXT: buffer_load_dword v2, off, s[0:3], s33 ; 4-byte Folded Reload
; CHECK-NEXT: s_mov_b64 exec, s[4:5]
; CHECK-NEXT: s_add_i32 s32, s32, 0xfffffc00
-; CHECK-NEXT: s_mov_b32 s33, s19
+; CHECK-NEXT: s_mov_b32 s33, s25
; CHECK-NEXT: s_waitcnt vmcnt(0)
; CHECK-NEXT: s_setpc_b64 s[30:31]
entry:
@@ -255,16 +239,12 @@
define protected amdgpu_kernel void @kernel() {
; CHECK-LABEL: kernel:
; CHECK: ; %bb.0: ; %entry
-; CHECK-NEXT: s_mov_b32 s32, 0x400
+; CHECK-NEXT: s_mov_b32 s32, 0
; CHECK-NEXT: s_add_u32 flat_scratch_lo, s12, s17
; CHECK-NEXT: s_addc_u32 flat_scratch_hi, s13, 0
; CHECK-NEXT: s_add_u32 s0, s0, s17
; CHECK-NEXT: s_addc_u32 s1, s1, 0
-; CHECK-NEXT: ; implicit-def: $vgpr3
; CHECK-NEXT: v_writelane_b32 v3, s16, 0
-; CHECK-NEXT: s_or_saveexec_b64 s[24:25], -1
-; CHECK-NEXT: buffer_store_dword v3, off, s[0:3], 0 offset:4 ; 4-byte Folded Spill
-; CHECK-NEXT: s_mov_b64 exec, s[24:25]
; CHECK-NEXT: s_mov_b32 s13, s15
; CHECK-NEXT: s_mov_b32 s12, s14
; CHECK-NEXT: v_readlane_b32 s14, v3, 0
@@ -282,11 +262,6 @@
; CHECK-NEXT: s_mov_b64 s[0:1], s[20:21]
; CHECK-NEXT: s_mov_b64 s[2:3], s[22:23]
; CHECK-NEXT: s_swappc_b64 s[30:31], s[16:17]
-; CHECK-NEXT: ; kill: def $vgpr1 killed $vgpr0 killed $exec
-; CHECK-NEXT: s_or_saveexec_b64 s[24:25], -1
-; CHECK-NEXT: buffer_load_dword v0, off, s[0:3], 0 offset:4 ; 4-byte Folded Reload
-; CHECK-NEXT: s_mov_b64 exec, s[24:25]
-; CHECK-NEXT: ; kill: killed $vgpr0
; CHECK-NEXT: s_endpgm
entry:
%call = call i32 @caller_save_vgpr_spill_fp()
diff --git a/llvm/test/CodeGen/AMDGPU/nested-calls.ll b/llvm/test/CodeGen/AMDGPU/nested-calls.ll
--- a/llvm/test/CodeGen/AMDGPU/nested-calls.ll
+++ b/llvm/test/CodeGen/AMDGPU/nested-calls.ll
@@ -15,8 +15,9 @@
; GCN-NEXT: s_mov_b32 s33, s32
; GCN-NEXT: s_or_saveexec_b64 [[COPY_EXEC0:s\[[0-9]+:[0-9]+\]]], -1{{$}}
; GCN-NEXT: buffer_store_dword v40, off, s[0:3], s33 ; 4-byte Folded Spill
+; GCN-NEXT: buffer_store_dword v41, off, s[0:3], s33 offset:4 ; 4-byte Folded Spill
; GCN-NEXT: s_mov_b64 exec, [[COPY_EXEC0]]
-; GCN-DAG: v_writelane_b32 v40, [[FP_SCRATCH_COPY]], 2
+; GCN-DAG: v_writelane_b32 v41, [[FP_SCRATCH_COPY]], 0
; GCN-DAG: v_writelane_b32 v40, s30, 0
; GCN-DAG: v_writelane_b32 v40, s31, 1
@@ -25,9 +26,10 @@
; GCN: v_readlane_b32 s31, v40, 1
; GCN: v_readlane_b32 s30, v40, 0
-; GCN-NEXT: v_readlane_b32 [[FP_SCRATCH_COPY:s[0-9]+]], v40, 2
+; GCN-NEXT: v_readlane_b32 [[FP_SCRATCH_COPY:s[0-9]+]], v41, 0
; GCN: s_or_saveexec_b64 [[COPY_EXEC1:s\[[0-9]+:[0-9]+\]]], -1{{$}}
; GCN-NEXT: buffer_load_dword v40, off, s[0:3], s33 ; 4-byte Folded Reload
+; GCN-NEXT: buffer_load_dword v41, off, s[0:3], s33 offset:4 ; 4-byte Folded Reload
; GCN-NEXT: s_mov_b64 exec, [[COPY_EXEC1]]
; GCN-NEXT: s_addk_i32 s32, 0xfc00
; GCN-NEXT: s_mov_b32 s33, [[FP_SCRATCH_COPY]]
diff --git a/llvm/test/CodeGen/AMDGPU/no-source-locations-in-prologue.ll b/llvm/test/CodeGen/AMDGPU/no-source-locations-in-prologue.ll
--- a/llvm/test/CodeGen/AMDGPU/no-source-locations-in-prologue.ll
+++ b/llvm/test/CodeGen/AMDGPU/no-source-locations-in-prologue.ll
@@ -17,8 +17,9 @@
; CHECK-NEXT: s_mov_b32 s33, s32
; CHECK-NEXT: s_or_saveexec_b64 s[18:19], -1
; CHECK-NEXT: buffer_store_dword v40, off, s[0:3], s33 ; 4-byte Folded Spill
+; CHECK-NEXT: buffer_store_dword v41, off, s[0:3], s33 offset:4 ; 4-byte Folded Spill
; CHECK-NEXT: s_mov_b64 exec, s[18:19]
-; CHECK-NEXT: v_writelane_b32 v40, s16, 2
+; CHECK-NEXT: v_writelane_b32 v41, s16, 0
; CHECK-NEXT: s_add_i32 s32, s32, 0x400
; CHECK-NEXT: v_writelane_b32 v40, s30, 0
; CHECK-NEXT: v_writelane_b32 v40, s31, 1
@@ -38,11 +39,12 @@
; CHECK-NEXT: .loc 0 32 1 ; lane-info.cpp:32:1
; CHECK-NEXT: v_readlane_b32 s31, v40, 1
; CHECK-NEXT: v_readlane_b32 s30, v40, 0
-; CHECK-NEXT: v_readlane_b32 s4, v40, 2
+; CHECK-NEXT: v_readlane_b32 s4, v41, 0
; CHECK-NEXT: s_or_saveexec_b64 s[6:7], -1
; CHECK-NEXT: buffer_load_dword v40, off, s[0:3], s33 ; 4-byte Folded Reload
+; CHECK-NEXT: buffer_load_dword v41, off, s[0:3], s33 offset:4 ; 4-byte Folded Reload
; CHECK-NEXT: s_mov_b64 exec, s[6:7]
-; CHECK-NEXT: .loc 0 32 1 epilogue_begin is_stmt 0 ; lane-info.cpp:32:1
+; CHECK-NEXT: .loc 0 32 1 epilogue_begin is_stmt 0 ; lane-info.cpp:32:1
; CHECK-NEXT: s_add_i32 s32, s32, 0xfffffc00
; CHECK-NEXT: s_mov_b32 s33, s4
; CHECK-NEXT: s_waitcnt vmcnt(0)
diff --git a/llvm/test/CodeGen/AMDGPU/partial-sgpr-to-vgpr-spills.ll b/llvm/test/CodeGen/AMDGPU/partial-sgpr-to-vgpr-spills.ll
--- a/llvm/test/CodeGen/AMDGPU/partial-sgpr-to-vgpr-spills.ll
+++ b/llvm/test/CodeGen/AMDGPU/partial-sgpr-to-vgpr-spills.ll
@@ -11,107 +11,95 @@
define amdgpu_kernel void @spill_sgprs_to_multiple_vgprs(ptr addrspace(1) %out, i32 %in) #0 {
; GCN-LABEL: spill_sgprs_to_multiple_vgprs:
; GCN: ; %bb.0:
-; GCN-NEXT: s_mov_b32 s92, SCRATCH_RSRC_DWORD0
-; GCN-NEXT: s_mov_b32 s93, SCRATCH_RSRC_DWORD1
-; GCN-NEXT: s_mov_b32 s94, -1
-; GCN-NEXT: s_mov_b32 s95, 0xe8f000
-; GCN-NEXT: s_add_u32 s92, s92, s11
-; GCN-NEXT: s_addc_u32 s93, s93, 0
-; GCN-NEXT: ; implicit-def: $vgpr0
-; GCN-NEXT: ; implicit-def: $vgpr1
-; GCN-NEXT: ; implicit-def: $vgpr2
; GCN-NEXT: s_load_dword s0, s[4:5], 0xb
; GCN-NEXT: ;;#ASMSTART
; GCN-NEXT: ; def s[4:11]
; GCN-NEXT: ;;#ASMEND
-; GCN-NEXT: v_writelane_b32 v2, s4, 0
-; GCN-NEXT: v_writelane_b32 v2, s5, 1
-; GCN-NEXT: v_writelane_b32 v2, s6, 2
-; GCN-NEXT: v_writelane_b32 v2, s7, 3
-; GCN-NEXT: v_writelane_b32 v2, s8, 4
-; GCN-NEXT: v_writelane_b32 v2, s9, 5
-; GCN-NEXT: v_writelane_b32 v2, s10, 6
-; GCN-NEXT: v_writelane_b32 v2, s11, 7
+; GCN-NEXT: v_writelane_b32 v0, s4, 0
+; GCN-NEXT: v_writelane_b32 v0, s5, 1
+; GCN-NEXT: v_writelane_b32 v0, s6, 2
+; GCN-NEXT: v_writelane_b32 v0, s7, 3
+; GCN-NEXT: v_writelane_b32 v0, s8, 4
+; GCN-NEXT: v_writelane_b32 v0, s9, 5
+; GCN-NEXT: v_writelane_b32 v0, s10, 6
+; GCN-NEXT: v_writelane_b32 v0, s11, 7
; GCN-NEXT: ;;#ASMSTART
; GCN-NEXT: ; def s[4:11]
; GCN-NEXT: ;;#ASMEND
-; GCN-NEXT: v_writelane_b32 v2, s4, 8
-; GCN-NEXT: v_writelane_b32 v2, s5, 9
-; GCN-NEXT: v_writelane_b32 v2, s6, 10
-; GCN-NEXT: v_writelane_b32 v2, s7, 11
-; GCN-NEXT: v_writelane_b32 v2, s8, 12
-; GCN-NEXT: v_writelane_b32 v2, s9, 13
-; GCN-NEXT: v_writelane_b32 v2, s10, 14
-; GCN-NEXT: v_writelane_b32 v2, s11, 15
+; GCN-NEXT: v_writelane_b32 v0, s4, 8
+; GCN-NEXT: v_writelane_b32 v0, s5, 9
+; GCN-NEXT: v_writelane_b32 v0, s6, 10
+; GCN-NEXT: v_writelane_b32 v0, s7, 11
+; GCN-NEXT: v_writelane_b32 v0, s8, 12
+; GCN-NEXT: v_writelane_b32 v0, s9, 13
+; GCN-NEXT: v_writelane_b32 v0, s10, 14
+; GCN-NEXT: v_writelane_b32 v0, s11, 15
; GCN-NEXT: ;;#ASMSTART
; GCN-NEXT: ; def s[4:11]
; GCN-NEXT: ;;#ASMEND
-; GCN-NEXT: v_writelane_b32 v2, s4, 16
-; GCN-NEXT: v_writelane_b32 v2, s5, 17
-; GCN-NEXT: v_writelane_b32 v2, s6, 18
-; GCN-NEXT: v_writelane_b32 v2, s7, 19
-; GCN-NEXT: v_writelane_b32 v2, s8, 20
-; GCN-NEXT: v_writelane_b32 v2, s9, 21
-; GCN-NEXT: v_writelane_b32 v2, s10, 22
-; GCN-NEXT: v_writelane_b32 v2, s11, 23
+; GCN-NEXT: v_writelane_b32 v0, s4, 16
+; GCN-NEXT: v_writelane_b32 v0, s5, 17
+; GCN-NEXT: v_writelane_b32 v0, s6, 18
+; GCN-NEXT: v_writelane_b32 v0, s7, 19
+; GCN-NEXT: v_writelane_b32 v0, s8, 20
+; GCN-NEXT: v_writelane_b32 v0, s9, 21
+; GCN-NEXT: v_writelane_b32 v0, s10, 22
+; GCN-NEXT: v_writelane_b32 v0, s11, 23
; GCN-NEXT: ;;#ASMSTART
; GCN-NEXT: ; def s[4:11]
; GCN-NEXT: ;;#ASMEND
-; GCN-NEXT: v_writelane_b32 v2, s4, 24
-; GCN-NEXT: v_writelane_b32 v2, s5, 25
-; GCN-NEXT: v_writelane_b32 v2, s6, 26
-; GCN-NEXT: v_writelane_b32 v2, s7, 27
-; GCN-NEXT: v_writelane_b32 v2, s8, 28
-; GCN-NEXT: v_writelane_b32 v2, s9, 29
-; GCN-NEXT: v_writelane_b32 v2, s10, 30
-; GCN-NEXT: v_writelane_b32 v2, s11, 31
+; GCN-NEXT: v_writelane_b32 v0, s4, 24
+; GCN-NEXT: v_writelane_b32 v0, s5, 25
+; GCN-NEXT: v_writelane_b32 v0, s6, 26
+; GCN-NEXT: v_writelane_b32 v0, s7, 27
+; GCN-NEXT: v_writelane_b32 v0, s8, 28
+; GCN-NEXT: v_writelane_b32 v0, s9, 29
+; GCN-NEXT: v_writelane_b32 v0, s10, 30
+; GCN-NEXT: v_writelane_b32 v0, s11, 31
; GCN-NEXT: ;;#ASMSTART
; GCN-NEXT: ; def s[4:11]
; GCN-NEXT: ;;#ASMEND
-; GCN-NEXT: v_writelane_b32 v2, s4, 32
-; GCN-NEXT: v_writelane_b32 v2, s5, 33
-; GCN-NEXT: v_writelane_b32 v2, s6, 34
-; GCN-NEXT: v_writelane_b32 v2, s7, 35
-; GCN-NEXT: v_writelane_b32 v2, s8, 36
-; GCN-NEXT: v_writelane_b32 v2, s9, 37
-; GCN-NEXT: v_writelane_b32 v2, s10, 38
-; GCN-NEXT: v_writelane_b32 v2, s11, 39
+; GCN-NEXT: v_writelane_b32 v0, s4, 32
+; GCN-NEXT: v_writelane_b32 v0, s5, 33
+; GCN-NEXT: v_writelane_b32 v0, s6, 34
+; GCN-NEXT: v_writelane_b32 v0, s7, 35
+; GCN-NEXT: v_writelane_b32 v0, s8, 36
+; GCN-NEXT: v_writelane_b32 v0, s9, 37
+; GCN-NEXT: v_writelane_b32 v0, s10, 38
+; GCN-NEXT: v_writelane_b32 v0, s11, 39
; GCN-NEXT: ;;#ASMSTART
; GCN-NEXT: ; def s[4:11]
; GCN-NEXT: ;;#ASMEND
-; GCN-NEXT: v_writelane_b32 v2, s4, 40
-; GCN-NEXT: v_writelane_b32 v2, s5, 41
-; GCN-NEXT: v_writelane_b32 v2, s6, 42
-; GCN-NEXT: v_writelane_b32 v2, s7, 43
-; GCN-NEXT: v_writelane_b32 v2, s8, 44
-; GCN-NEXT: v_writelane_b32 v2, s9, 45
-; GCN-NEXT: v_writelane_b32 v2, s10, 46
-; GCN-NEXT: v_writelane_b32 v2, s11, 47
+; GCN-NEXT: v_writelane_b32 v0, s4, 40
+; GCN-NEXT: v_writelane_b32 v0, s5, 41
+; GCN-NEXT: v_writelane_b32 v0, s6, 42
+; GCN-NEXT: v_writelane_b32 v0, s7, 43
+; GCN-NEXT: v_writelane_b32 v0, s8, 44
+; GCN-NEXT: v_writelane_b32 v0, s9, 45
+; GCN-NEXT: v_writelane_b32 v0, s10, 46
+; GCN-NEXT: v_writelane_b32 v0, s11, 47
; GCN-NEXT: ;;#ASMSTART
; GCN-NEXT: ; def s[4:11]
; GCN-NEXT: ;;#ASMEND
-; GCN-NEXT: v_writelane_b32 v2, s4, 48
-; GCN-NEXT: v_writelane_b32 v2, s5, 49
-; GCN-NEXT: v_writelane_b32 v2, s6, 50
-; GCN-NEXT: v_writelane_b32 v2, s7, 51
-; GCN-NEXT: v_writelane_b32 v2, s8, 52
-; GCN-NEXT: v_writelane_b32 v2, s9, 53
-; GCN-NEXT: v_writelane_b32 v2, s10, 54
-; GCN-NEXT: v_writelane_b32 v2, s11, 55
+; GCN-NEXT: v_writelane_b32 v0, s4, 48
+; GCN-NEXT: v_writelane_b32 v0, s5, 49
+; GCN-NEXT: v_writelane_b32 v0, s6, 50
+; GCN-NEXT: v_writelane_b32 v0, s7, 51
+; GCN-NEXT: v_writelane_b32 v0, s8, 52
+; GCN-NEXT: v_writelane_b32 v0, s9, 53
+; GCN-NEXT: v_writelane_b32 v0, s10, 54
+; GCN-NEXT: v_writelane_b32 v0, s11, 55
; GCN-NEXT: ;;#ASMSTART
; GCN-NEXT: ; def s[4:11]
; GCN-NEXT: ;;#ASMEND
-; GCN-NEXT: v_writelane_b32 v2, s4, 56
-; GCN-NEXT: v_writelane_b32 v2, s5, 57
-; GCN-NEXT: v_writelane_b32 v2, s6, 58
-; GCN-NEXT: v_writelane_b32 v2, s7, 59
-; GCN-NEXT: v_writelane_b32 v2, s8, 60
-; GCN-NEXT: v_writelane_b32 v2, s9, 61
-; GCN-NEXT: v_writelane_b32 v2, s10, 62
-; GCN-NEXT: v_writelane_b32 v2, s11, 63
-; GCN-NEXT: s_or_saveexec_b64 s[34:35], -1
-; GCN-NEXT: buffer_store_dword v2, off, s[92:95], 0 offset:12 ; 4-byte Folded Spill
-; GCN-NEXT: s_mov_b64 exec, s[34:35]
+; GCN-NEXT: v_writelane_b32 v0, s4, 56
+; GCN-NEXT: v_writelane_b32 v0, s5, 57
+; GCN-NEXT: v_writelane_b32 v0, s6, 58
+; GCN-NEXT: v_writelane_b32 v0, s7, 59
+; GCN-NEXT: v_writelane_b32 v0, s8, 60
+; GCN-NEXT: v_writelane_b32 v0, s9, 61
+; GCN-NEXT: v_writelane_b32 v0, s10, 62
+; GCN-NEXT: v_writelane_b32 v0, s11, 63
; GCN-NEXT: ;;#ASMSTART
; GCN-NEXT: ; def s[4:11]
; GCN-NEXT: ;;#ASMEND
@@ -200,192 +188,176 @@
; GCN-NEXT: v_writelane_b32 v1, s9, 61
; GCN-NEXT: v_writelane_b32 v1, s10, 62
; GCN-NEXT: v_writelane_b32 v1, s11, 63
-; GCN-NEXT: s_or_saveexec_b64 s[34:35], -1
-; GCN-NEXT: buffer_store_dword v1, off, s[92:95], 0 offset:8 ; 4-byte Folded Spill
-; GCN-NEXT: s_mov_b64 exec, s[34:35]
; GCN-NEXT: ;;#ASMSTART
; GCN-NEXT: ; def s[4:11]
; GCN-NEXT: ;;#ASMEND
-; GCN-NEXT: v_writelane_b32 v0, s4, 0
-; GCN-NEXT: v_writelane_b32 v0, s5, 1
-; GCN-NEXT: v_writelane_b32 v0, s6, 2
-; GCN-NEXT: v_writelane_b32 v0, s7, 3
-; GCN-NEXT: v_writelane_b32 v0, s8, 4
-; GCN-NEXT: v_writelane_b32 v0, s9, 5
-; GCN-NEXT: v_writelane_b32 v0, s10, 6
-; GCN-NEXT: v_writelane_b32 v0, s11, 7
-; GCN-NEXT: s_or_saveexec_b64 s[34:35], -1
-; GCN-NEXT: buffer_store_dword v0, off, s[92:95], 0 offset:4 ; 4-byte Folded Spill
-; GCN-NEXT: s_mov_b64 exec, s[34:35]
+; GCN-NEXT: v_writelane_b32 v2, s4, 0
+; GCN-NEXT: v_writelane_b32 v2, s5, 1
+; GCN-NEXT: v_writelane_b32 v2, s6, 2
+; GCN-NEXT: v_writelane_b32 v2, s7, 3
+; GCN-NEXT: v_writelane_b32 v2, s8, 4
+; GCN-NEXT: v_writelane_b32 v2, s9, 5
+; GCN-NEXT: v_writelane_b32 v2, s10, 6
+; GCN-NEXT: v_writelane_b32 v2, s11, 7
; GCN-NEXT: s_mov_b32 s1, 0
; GCN-NEXT: s_waitcnt lgkmcnt(0)
; GCN-NEXT: s_cmp_lg_u32 s0, s1
; GCN-NEXT: s_cbranch_scc1 .LBB0_2
; GCN-NEXT: ; %bb.1: ; %bb0
-; GCN-NEXT: s_or_saveexec_b64 s[34:35], -1
-; GCN-NEXT: buffer_load_dword v0, off, s[92:95], 0 offset:4 ; 4-byte Folded Reload
-; GCN-NEXT: s_mov_b64 exec, s[34:35]
-; GCN-NEXT: s_or_saveexec_b64 s[34:35], -1
-; GCN-NEXT: buffer_load_dword v1, off, s[92:95], 0 offset:12 ; 4-byte Folded Reload
-; GCN-NEXT: s_mov_b64 exec, s[34:35]
-; GCN-NEXT: s_or_saveexec_b64 s[34:35], -1
-; GCN-NEXT: buffer_load_dword v2, off, s[92:95], 0 offset:8 ; 4-byte Folded Reload
-; GCN-NEXT: s_mov_b64 exec, s[34:35]
-; GCN-NEXT: s_waitcnt vmcnt(0)
-; GCN-NEXT: v_readlane_b32 s8, v2, 56
-; GCN-NEXT: v_readlane_b32 s9, v2, 57
-; GCN-NEXT: v_readlane_b32 s10, v2, 58
-; GCN-NEXT: v_readlane_b32 s11, v2, 59
-; GCN-NEXT: v_readlane_b32 s12, v2, 60
-; GCN-NEXT: v_readlane_b32 s13, v2, 61
-; GCN-NEXT: v_readlane_b32 s14, v2, 62
-; GCN-NEXT: v_readlane_b32 s15, v2, 63
-; GCN-NEXT: v_readlane_b32 s16, v2, 48
-; GCN-NEXT: v_readlane_b32 s17, v2, 49
-; GCN-NEXT: v_readlane_b32 s18, v2, 50
-; GCN-NEXT: v_readlane_b32 s19, v2, 51
-; GCN-NEXT: v_readlane_b32 s20, v2, 52
-; GCN-NEXT: v_readlane_b32 s21, v2, 53
-; GCN-NEXT: v_readlane_b32 s22, v2, 54
-; GCN-NEXT: v_readlane_b32 s23, v2, 55
-; GCN-NEXT: v_readlane_b32 s24, v2, 40
-; GCN-NEXT: v_readlane_b32 s25, v2, 41
-; GCN-NEXT: v_readlane_b32 s26, v2, 42
-; GCN-NEXT: v_readlane_b32 s27, v2, 43
-; GCN-NEXT: v_readlane_b32 s28, v2, 44
-; GCN-NEXT: v_readlane_b32 s29, v2, 45
-; GCN-NEXT: v_readlane_b32 s30, v2, 46
-; GCN-NEXT: v_readlane_b32 s31, v2, 47
-; GCN-NEXT: v_readlane_b32 s36, v2, 32
-; GCN-NEXT: v_readlane_b32 s37, v2, 33
-; GCN-NEXT: v_readlane_b32 s38, v2, 34
-; GCN-NEXT: v_readlane_b32 s39, v2, 35
-; GCN-NEXT: v_readlane_b32 s40, v2, 36
-; GCN-NEXT: v_readlane_b32 s41, v2, 37
-; GCN-NEXT: v_readlane_b32 s42, v2, 38
-; GCN-NEXT: v_readlane_b32 s43, v2, 39
-; GCN-NEXT: v_readlane_b32 s44, v2, 24
-; GCN-NEXT: v_readlane_b32 s45, v2, 25
-; GCN-NEXT: v_readlane_b32 s46, v2, 26
-; GCN-NEXT: v_readlane_b32 s47, v2, 27
-; GCN-NEXT: v_readlane_b32 s48, v2, 28
-; GCN-NEXT: v_readlane_b32 s49, v2, 29
-; GCN-NEXT: v_readlane_b32 s50, v2, 30
-; GCN-NEXT: v_readlane_b32 s51, v2, 31
-; GCN-NEXT: v_readlane_b32 s52, v2, 16
-; GCN-NEXT: v_readlane_b32 s53, v2, 17
-; GCN-NEXT: v_readlane_b32 s54, v2, 18
-; GCN-NEXT: v_readlane_b32 s55, v2, 19
-; GCN-NEXT: v_readlane_b32 s56, v2, 20
-; GCN-NEXT: v_readlane_b32 s57, v2, 21
-; GCN-NEXT: v_readlane_b32 s58, v2, 22
-; GCN-NEXT: v_readlane_b32 s59, v2, 23
-; GCN-NEXT: v_readlane_b32 s60, v2, 8
-; GCN-NEXT: v_readlane_b32 s61, v2, 9
-; GCN-NEXT: v_readlane_b32 s62, v2, 10
-; GCN-NEXT: v_readlane_b32 s63, v2, 11
-; GCN-NEXT: v_readlane_b32 s64, v2, 12
-; GCN-NEXT: v_readlane_b32 s65, v2, 13
-; GCN-NEXT: v_readlane_b32 s66, v2, 14
-; GCN-NEXT: v_readlane_b32 s67, v2, 15
-; GCN-NEXT: v_readlane_b32 s68, v2, 0
-; GCN-NEXT: v_readlane_b32 s69, v2, 1
-; GCN-NEXT: v_readlane_b32 s70, v2, 2
-; GCN-NEXT: v_readlane_b32 s71, v2, 3
-; GCN-NEXT: v_readlane_b32 s72, v2, 4
-; GCN-NEXT: v_readlane_b32 s73, v2, 5
-; GCN-NEXT: v_readlane_b32 s74, v2, 6
-; GCN-NEXT: v_readlane_b32 s75, v2, 7
-; GCN-NEXT: v_readlane_b32 s76, v1, 56
-; GCN-NEXT: v_readlane_b32 s77, v1, 57
-; GCN-NEXT: v_readlane_b32 s78, v1, 58
-; GCN-NEXT: v_readlane_b32 s79, v1, 59
-; GCN-NEXT: v_readlane_b32 s80, v1, 60
-; GCN-NEXT: v_readlane_b32 s81, v1, 61
-; GCN-NEXT: v_readlane_b32 s82, v1, 62
-; GCN-NEXT: v_readlane_b32 s83, v1, 63
-; GCN-NEXT: v_readlane_b32 s84, v1, 48
-; GCN-NEXT: v_readlane_b32 s85, v1, 49
-; GCN-NEXT: v_readlane_b32 s86, v1, 50
-; GCN-NEXT: v_readlane_b32 s87, v1, 51
-; GCN-NEXT: v_readlane_b32 s88, v1, 52
-; GCN-NEXT: v_readlane_b32 s89, v1, 53
-; GCN-NEXT: v_readlane_b32 s90, v1, 54
-; GCN-NEXT: v_readlane_b32 s91, v1, 55
-; GCN-NEXT: v_readlane_b32 s0, v1, 0
-; GCN-NEXT: v_readlane_b32 s1, v1, 1
-; GCN-NEXT: v_readlane_b32 s2, v1, 2
-; GCN-NEXT: v_readlane_b32 s3, v1, 3
-; GCN-NEXT: v_readlane_b32 s4, v1, 4
-; GCN-NEXT: v_readlane_b32 s5, v1, 5
-; GCN-NEXT: v_readlane_b32 s6, v1, 6
-; GCN-NEXT: v_readlane_b32 s7, v1, 7
+; GCN-NEXT: v_readlane_b32 s8, v1, 56
+; GCN-NEXT: v_readlane_b32 s9, v1, 57
+; GCN-NEXT: v_readlane_b32 s10, v1, 58
+; GCN-NEXT: v_readlane_b32 s11, v1, 59
+; GCN-NEXT: v_readlane_b32 s12, v1, 60
+; GCN-NEXT: v_readlane_b32 s13, v1, 61
+; GCN-NEXT: v_readlane_b32 s14, v1, 62
+; GCN-NEXT: v_readlane_b32 s15, v1, 63
+; GCN-NEXT: v_readlane_b32 s16, v1, 48
+; GCN-NEXT: v_readlane_b32 s17, v1, 49
+; GCN-NEXT: v_readlane_b32 s18, v1, 50
+; GCN-NEXT: v_readlane_b32 s19, v1, 51
+; GCN-NEXT: v_readlane_b32 s20, v1, 52
+; GCN-NEXT: v_readlane_b32 s21, v1, 53
+; GCN-NEXT: v_readlane_b32 s22, v1, 54
+; GCN-NEXT: v_readlane_b32 s23, v1, 55
+; GCN-NEXT: v_readlane_b32 s24, v1, 40
+; GCN-NEXT: v_readlane_b32 s25, v1, 41
+; GCN-NEXT: v_readlane_b32 s26, v1, 42
+; GCN-NEXT: v_readlane_b32 s27, v1, 43
+; GCN-NEXT: v_readlane_b32 s28, v1, 44
+; GCN-NEXT: v_readlane_b32 s29, v1, 45
+; GCN-NEXT: v_readlane_b32 s30, v1, 46
+; GCN-NEXT: v_readlane_b32 s31, v1, 47
+; GCN-NEXT: v_readlane_b32 s36, v1, 32
+; GCN-NEXT: v_readlane_b32 s37, v1, 33
+; GCN-NEXT: v_readlane_b32 s38, v1, 34
+; GCN-NEXT: v_readlane_b32 s39, v1, 35
+; GCN-NEXT: v_readlane_b32 s40, v1, 36
+; GCN-NEXT: v_readlane_b32 s41, v1, 37
+; GCN-NEXT: v_readlane_b32 s42, v1, 38
+; GCN-NEXT: v_readlane_b32 s43, v1, 39
+; GCN-NEXT: v_readlane_b32 s44, v1, 24
+; GCN-NEXT: v_readlane_b32 s45, v1, 25
+; GCN-NEXT: v_readlane_b32 s46, v1, 26
+; GCN-NEXT: v_readlane_b32 s47, v1, 27
+; GCN-NEXT: v_readlane_b32 s48, v1, 28
+; GCN-NEXT: v_readlane_b32 s49, v1, 29
+; GCN-NEXT: v_readlane_b32 s50, v1, 30
+; GCN-NEXT: v_readlane_b32 s51, v1, 31
+; GCN-NEXT: v_readlane_b32 s52, v1, 16
+; GCN-NEXT: v_readlane_b32 s53, v1, 17
+; GCN-NEXT: v_readlane_b32 s54, v1, 18
+; GCN-NEXT: v_readlane_b32 s55, v1, 19
+; GCN-NEXT: v_readlane_b32 s56, v1, 20
+; GCN-NEXT: v_readlane_b32 s57, v1, 21
+; GCN-NEXT: v_readlane_b32 s58, v1, 22
+; GCN-NEXT: v_readlane_b32 s59, v1, 23
+; GCN-NEXT: v_readlane_b32 s60, v1, 8
+; GCN-NEXT: v_readlane_b32 s61, v1, 9
+; GCN-NEXT: v_readlane_b32 s62, v1, 10
+; GCN-NEXT: v_readlane_b32 s63, v1, 11
+; GCN-NEXT: v_readlane_b32 s64, v1, 12
+; GCN-NEXT: v_readlane_b32 s65, v1, 13
+; GCN-NEXT: v_readlane_b32 s66, v1, 14
+; GCN-NEXT: v_readlane_b32 s67, v1, 15
+; GCN-NEXT: v_readlane_b32 s68, v1, 0
+; GCN-NEXT: v_readlane_b32 s69, v1, 1
+; GCN-NEXT: v_readlane_b32 s70, v1, 2
+; GCN-NEXT: v_readlane_b32 s71, v1, 3
+; GCN-NEXT: v_readlane_b32 s72, v1, 4
+; GCN-NEXT: v_readlane_b32 s73, v1, 5
+; GCN-NEXT: v_readlane_b32 s74, v1, 6
+; GCN-NEXT: v_readlane_b32 s75, v1, 7
+; GCN-NEXT: v_readlane_b32 s76, v0, 56
+; GCN-NEXT: v_readlane_b32 s77, v0, 57
+; GCN-NEXT: v_readlane_b32 s78, v0, 58
+; GCN-NEXT: v_readlane_b32 s79, v0, 59
+; GCN-NEXT: v_readlane_b32 s80, v0, 60
+; GCN-NEXT: v_readlane_b32 s81, v0, 61
+; GCN-NEXT: v_readlane_b32 s82, v0, 62
+; GCN-NEXT: v_readlane_b32 s83, v0, 63
+; GCN-NEXT: v_readlane_b32 s84, v0, 48
+; GCN-NEXT: v_readlane_b32 s85, v0, 49
+; GCN-NEXT: v_readlane_b32 s86, v0, 50
+; GCN-NEXT: v_readlane_b32 s87, v0, 51
+; GCN-NEXT: v_readlane_b32 s88, v0, 52
+; GCN-NEXT: v_readlane_b32 s89, v0, 53
+; GCN-NEXT: v_readlane_b32 s90, v0, 54
+; GCN-NEXT: v_readlane_b32 s91, v0, 55
+; GCN-NEXT: v_readlane_b32 s0, v0, 0
+; GCN-NEXT: v_readlane_b32 s1, v0, 1
+; GCN-NEXT: v_readlane_b32 s2, v0, 2
+; GCN-NEXT: v_readlane_b32 s3, v0, 3
+; GCN-NEXT: v_readlane_b32 s4, v0, 4
+; GCN-NEXT: v_readlane_b32 s5, v0, 5
+; GCN-NEXT: v_readlane_b32 s6, v0, 6
+; GCN-NEXT: v_readlane_b32 s7, v0, 7
; GCN-NEXT: ;;#ASMSTART
; GCN-NEXT: ; use s[0:7]
; GCN-NEXT: ;;#ASMEND
-; GCN-NEXT: v_readlane_b32 s0, v1, 8
-; GCN-NEXT: v_readlane_b32 s1, v1, 9
-; GCN-NEXT: v_readlane_b32 s2, v1, 10
-; GCN-NEXT: v_readlane_b32 s3, v1, 11
-; GCN-NEXT: v_readlane_b32 s4, v1, 12
-; GCN-NEXT: v_readlane_b32 s5, v1, 13
-; GCN-NEXT: v_readlane_b32 s6, v1, 14
-; GCN-NEXT: v_readlane_b32 s7, v1, 15
+; GCN-NEXT: v_readlane_b32 s0, v0, 8
+; GCN-NEXT: v_readlane_b32 s1, v0, 9
+; GCN-NEXT: v_readlane_b32 s2, v0, 10
+; GCN-NEXT: v_readlane_b32 s3, v0, 11
+; GCN-NEXT: v_readlane_b32 s4, v0, 12
+; GCN-NEXT: v_readlane_b32 s5, v0, 13
+; GCN-NEXT: v_readlane_b32 s6, v0, 14
+; GCN-NEXT: v_readlane_b32 s7, v0, 15
; GCN-NEXT: ;;#ASMSTART
; GCN-NEXT: ; use s[0:7]
; GCN-NEXT: ;;#ASMEND
-; GCN-NEXT: v_readlane_b32 s0, v1, 16
-; GCN-NEXT: v_readlane_b32 s1, v1, 17
-; GCN-NEXT: v_readlane_b32 s2, v1, 18
-; GCN-NEXT: v_readlane_b32 s3, v1, 19
-; GCN-NEXT: v_readlane_b32 s4, v1, 20
-; GCN-NEXT: v_readlane_b32 s5, v1, 21
-; GCN-NEXT: v_readlane_b32 s6, v1, 22
-; GCN-NEXT: v_readlane_b32 s7, v1, 23
+; GCN-NEXT: v_readlane_b32 s0, v0, 16
+; GCN-NEXT: v_readlane_b32 s1, v0, 17
+; GCN-NEXT: v_readlane_b32 s2, v0, 18
+; GCN-NEXT: v_readlane_b32 s3, v0, 19
+; GCN-NEXT: v_readlane_b32 s4, v0, 20
+; GCN-NEXT: v_readlane_b32 s5, v0, 21
+; GCN-NEXT: v_readlane_b32 s6, v0, 22
+; GCN-NEXT: v_readlane_b32 s7, v0, 23
; GCN-NEXT: ;;#ASMSTART
; GCN-NEXT: ; use s[0:7]
; GCN-NEXT: ;;#ASMEND
-; GCN-NEXT: v_readlane_b32 s0, v1, 24
-; GCN-NEXT: v_readlane_b32 s1, v1, 25
-; GCN-NEXT: v_readlane_b32 s2, v1, 26
-; GCN-NEXT: v_readlane_b32 s3, v1, 27
-; GCN-NEXT: v_readlane_b32 s4, v1, 28
-; GCN-NEXT: v_readlane_b32 s5, v1, 29
-; GCN-NEXT: v_readlane_b32 s6, v1, 30
-; GCN-NEXT: v_readlane_b32 s7, v1, 31
+; GCN-NEXT: v_readlane_b32 s0, v0, 24
+; GCN-NEXT: v_readlane_b32 s1, v0, 25
+; GCN-NEXT: v_readlane_b32 s2, v0, 26
+; GCN-NEXT: v_readlane_b32 s3, v0, 27
+; GCN-NEXT: v_readlane_b32 s4, v0, 28
+; GCN-NEXT: v_readlane_b32 s5, v0, 29
+; GCN-NEXT: v_readlane_b32 s6, v0, 30
+; GCN-NEXT: v_readlane_b32 s7, v0, 31
; GCN-NEXT: ;;#ASMSTART
; GCN-NEXT: ; use s[0:7]
; GCN-NEXT: ;;#ASMEND
-; GCN-NEXT: v_readlane_b32 s0, v1, 32
-; GCN-NEXT: v_readlane_b32 s1, v1, 33
-; GCN-NEXT: v_readlane_b32 s2, v1, 34
-; GCN-NEXT: v_readlane_b32 s3, v1, 35
-; GCN-NEXT: v_readlane_b32 s4, v1, 36
-; GCN-NEXT: v_readlane_b32 s5, v1, 37
-; GCN-NEXT: v_readlane_b32 s6, v1, 38
-; GCN-NEXT: v_readlane_b32 s7, v1, 39
+; GCN-NEXT: v_readlane_b32 s0, v0, 32
+; GCN-NEXT: v_readlane_b32 s1, v0, 33
+; GCN-NEXT: v_readlane_b32 s2, v0, 34
+; GCN-NEXT: v_readlane_b32 s3, v0, 35
+; GCN-NEXT: v_readlane_b32 s4, v0, 36
+; GCN-NEXT: v_readlane_b32 s5, v0, 37
+; GCN-NEXT: v_readlane_b32 s6, v0, 38
+; GCN-NEXT: v_readlane_b32 s7, v0, 39
; GCN-NEXT: ;;#ASMSTART
; GCN-NEXT: ; use s[0:7]
; GCN-NEXT: ;;#ASMEND
-; GCN-NEXT: v_readlane_b32 s0, v1, 40
-; GCN-NEXT: v_readlane_b32 s1, v1, 41
-; GCN-NEXT: v_readlane_b32 s2, v1, 42
-; GCN-NEXT: v_readlane_b32 s3, v1, 43
-; GCN-NEXT: v_readlane_b32 s4, v1, 44
-; GCN-NEXT: v_readlane_b32 s5, v1, 45
-; GCN-NEXT: v_readlane_b32 s6, v1, 46
-; GCN-NEXT: v_readlane_b32 s7, v1, 47
+; GCN-NEXT: v_readlane_b32 s0, v0, 40
+; GCN-NEXT: v_readlane_b32 s1, v0, 41
+; GCN-NEXT: v_readlane_b32 s2, v0, 42
+; GCN-NEXT: v_readlane_b32 s3, v0, 43
+; GCN-NEXT: v_readlane_b32 s4, v0, 44
+; GCN-NEXT: v_readlane_b32 s5, v0, 45
+; GCN-NEXT: v_readlane_b32 s6, v0, 46
+; GCN-NEXT: v_readlane_b32 s7, v0, 47
; GCN-NEXT: ;;#ASMSTART
; GCN-NEXT: ; use s[0:7]
; GCN-NEXT: ;;#ASMEND
-; GCN-NEXT: v_readlane_b32 s0, v0, 0
-; GCN-NEXT: v_readlane_b32 s1, v0, 1
-; GCN-NEXT: v_readlane_b32 s2, v0, 2
-; GCN-NEXT: v_readlane_b32 s3, v0, 3
-; GCN-NEXT: v_readlane_b32 s4, v0, 4
-; GCN-NEXT: v_readlane_b32 s5, v0, 5
-; GCN-NEXT: v_readlane_b32 s6, v0, 6
-; GCN-NEXT: v_readlane_b32 s7, v0, 7
+; GCN-NEXT: v_readlane_b32 s0, v2, 0
+; GCN-NEXT: v_readlane_b32 s1, v2, 1
+; GCN-NEXT: v_readlane_b32 s2, v2, 2
+; GCN-NEXT: v_readlane_b32 s3, v2, 3
+; GCN-NEXT: v_readlane_b32 s4, v2, 4
+; GCN-NEXT: v_readlane_b32 s5, v2, 5
+; GCN-NEXT: v_readlane_b32 s6, v2, 6
+; GCN-NEXT: v_readlane_b32 s7, v2, 7
; GCN-NEXT: ;;#ASMSTART
; GCN-NEXT: ; use s[84:91]
; GCN-NEXT: ;;#ASMEND
@@ -420,18 +392,6 @@
; GCN-NEXT: ; use s[0:7]
; GCN-NEXT: ;;#ASMEND
; GCN-NEXT: .LBB0_2: ; %ret
-; GCN-NEXT: s_or_saveexec_b64 s[34:35], -1
-; GCN-NEXT: buffer_load_dword v0, off, s[92:95], 0 offset:4 ; 4-byte Folded Reload
-; GCN-NEXT: s_mov_b64 exec, s[34:35]
-; GCN-NEXT: s_or_saveexec_b64 s[34:35], -1
-; GCN-NEXT: buffer_load_dword v1, off, s[92:95], 0 offset:8 ; 4-byte Folded Reload
-; GCN-NEXT: s_mov_b64 exec, s[34:35]
-; GCN-NEXT: s_or_saveexec_b64 s[34:35], -1
-; GCN-NEXT: buffer_load_dword v2, off, s[92:95], 0 offset:12 ; 4-byte Folded Reload
-; GCN-NEXT: s_mov_b64 exec, s[34:35]
-; GCN-NEXT: ; kill: killed $vgpr2
-; GCN-NEXT: ; kill: killed $vgpr1
-; GCN-NEXT: ; kill: killed $vgpr0
; GCN-NEXT: s_endpgm
%wide.sgpr0 = call <8 x i32> asm sideeffect "; def $0", "=s" () #0
%wide.sgpr1 = call <8 x i32> asm sideeffect "; def $0", "=s" () #0
@@ -482,125 +442,104 @@
define amdgpu_kernel void @split_sgpr_spill_2_vgprs(ptr addrspace(1) %out, i32 %in) #1 {
; GCN-LABEL: split_sgpr_spill_2_vgprs:
; GCN: ; %bb.0:
-; GCN-NEXT: s_mov_b32 s52, SCRATCH_RSRC_DWORD0
-; GCN-NEXT: s_mov_b32 s53, SCRATCH_RSRC_DWORD1
-; GCN-NEXT: s_mov_b32 s54, -1
-; GCN-NEXT: s_mov_b32 s55, 0xe8f000
-; GCN-NEXT: s_add_u32 s52, s52, s11
-; GCN-NEXT: s_addc_u32 s53, s53, 0
-; GCN-NEXT: ; implicit-def: $vgpr0
-; GCN-NEXT: ; implicit-def: $vgpr1
; GCN-NEXT: s_load_dword s0, s[4:5], 0xb
; GCN-NEXT: ;;#ASMSTART
; GCN-NEXT: ; def s[4:19]
; GCN-NEXT: ;;#ASMEND
-; GCN-NEXT: v_writelane_b32 v1, s4, 0
-; GCN-NEXT: v_writelane_b32 v1, s5, 1
-; GCN-NEXT: v_writelane_b32 v1, s6, 2
-; GCN-NEXT: v_writelane_b32 v1, s7, 3
-; GCN-NEXT: v_writelane_b32 v1, s8, 4
-; GCN-NEXT: v_writelane_b32 v1, s9, 5
-; GCN-NEXT: v_writelane_b32 v1, s10, 6
-; GCN-NEXT: v_writelane_b32 v1, s11, 7
-; GCN-NEXT: v_writelane_b32 v1, s12, 8
-; GCN-NEXT: v_writelane_b32 v1, s13, 9
-; GCN-NEXT: v_writelane_b32 v1, s14, 10
-; GCN-NEXT: v_writelane_b32 v1, s15, 11
-; GCN-NEXT: v_writelane_b32 v1, s16, 12
-; GCN-NEXT: v_writelane_b32 v1, s17, 13
-; GCN-NEXT: v_writelane_b32 v1, s18, 14
-; GCN-NEXT: v_writelane_b32 v1, s19, 15
+; GCN-NEXT: v_writelane_b32 v0, s4, 0
+; GCN-NEXT: v_writelane_b32 v0, s5, 1
+; GCN-NEXT: v_writelane_b32 v0, s6, 2
+; GCN-NEXT: v_writelane_b32 v0, s7, 3
+; GCN-NEXT: v_writelane_b32 v0, s8, 4
+; GCN-NEXT: v_writelane_b32 v0, s9, 5
+; GCN-NEXT: v_writelane_b32 v0, s10, 6
+; GCN-NEXT: v_writelane_b32 v0, s11, 7
+; GCN-NEXT: v_writelane_b32 v0, s12, 8
+; GCN-NEXT: v_writelane_b32 v0, s13, 9
+; GCN-NEXT: v_writelane_b32 v0, s14, 10
+; GCN-NEXT: v_writelane_b32 v0, s15, 11
+; GCN-NEXT: v_writelane_b32 v0, s16, 12
+; GCN-NEXT: v_writelane_b32 v0, s17, 13
+; GCN-NEXT: v_writelane_b32 v0, s18, 14
+; GCN-NEXT: v_writelane_b32 v0, s19, 15
; GCN-NEXT: ;;#ASMSTART
; GCN-NEXT: ; def s[4:19]
; GCN-NEXT: ;;#ASMEND
-; GCN-NEXT: v_writelane_b32 v1, s4, 16
-; GCN-NEXT: v_writelane_b32 v1, s5, 17
-; GCN-NEXT: v_writelane_b32 v1, s6, 18
-; GCN-NEXT: v_writelane_b32 v1, s7, 19
-; GCN-NEXT: v_writelane_b32 v1, s8, 20
-; GCN-NEXT: v_writelane_b32 v1, s9, 21
-; GCN-NEXT: v_writelane_b32 v1, s10, 22
-; GCN-NEXT: v_writelane_b32 v1, s11, 23
-; GCN-NEXT: v_writelane_b32 v1, s12, 24
-; GCN-NEXT: v_writelane_b32 v1, s13, 25
-; GCN-NEXT: v_writelane_b32 v1, s14, 26
-; GCN-NEXT: v_writelane_b32 v1, s15, 27
-; GCN-NEXT: v_writelane_b32 v1, s16, 28
-; GCN-NEXT: v_writelane_b32 v1, s17, 29
-; GCN-NEXT: v_writelane_b32 v1, s18, 30
-; GCN-NEXT: v_writelane_b32 v1, s19, 31
+; GCN-NEXT: v_writelane_b32 v0, s4, 16
+; GCN-NEXT: v_writelane_b32 v0, s5, 17
+; GCN-NEXT: v_writelane_b32 v0, s6, 18
+; GCN-NEXT: v_writelane_b32 v0, s7, 19
+; GCN-NEXT: v_writelane_b32 v0, s8, 20
+; GCN-NEXT: v_writelane_b32 v0, s9, 21
+; GCN-NEXT: v_writelane_b32 v0, s10, 22
+; GCN-NEXT: v_writelane_b32 v0, s11, 23
+; GCN-NEXT: v_writelane_b32 v0, s12, 24
+; GCN-NEXT: v_writelane_b32 v0, s13, 25
+; GCN-NEXT: v_writelane_b32 v0, s14, 26
+; GCN-NEXT: v_writelane_b32 v0, s15, 27
+; GCN-NEXT: v_writelane_b32 v0, s16, 28
+; GCN-NEXT: v_writelane_b32 v0, s17, 29
+; GCN-NEXT: v_writelane_b32 v0, s18, 30
+; GCN-NEXT: v_writelane_b32 v0, s19, 31
; GCN-NEXT: ;;#ASMSTART
; GCN-NEXT: ; def s[4:19]
; GCN-NEXT: ;;#ASMEND
-; GCN-NEXT: v_writelane_b32 v1, s4, 32
-; GCN-NEXT: v_writelane_b32 v1, s5, 33
-; GCN-NEXT: v_writelane_b32 v1, s6, 34
-; GCN-NEXT: v_writelane_b32 v1, s7, 35
-; GCN-NEXT: v_writelane_b32 v1, s8, 36
-; GCN-NEXT: v_writelane_b32 v1, s9, 37
-; GCN-NEXT: v_writelane_b32 v1, s10, 38
-; GCN-NEXT: v_writelane_b32 v1, s11, 39
-; GCN-NEXT: v_writelane_b32 v1, s12, 40
-; GCN-NEXT: v_writelane_b32 v1, s13, 41
-; GCN-NEXT: v_writelane_b32 v1, s14, 42
-; GCN-NEXT: v_writelane_b32 v1, s15, 43
-; GCN-NEXT: v_writelane_b32 v1, s16, 44
-; GCN-NEXT: v_writelane_b32 v1, s17, 45
-; GCN-NEXT: v_writelane_b32 v1, s18, 46
-; GCN-NEXT: v_writelane_b32 v1, s19, 47
+; GCN-NEXT: v_writelane_b32 v0, s4, 32
+; GCN-NEXT: v_writelane_b32 v0, s5, 33
+; GCN-NEXT: v_writelane_b32 v0, s6, 34
+; GCN-NEXT: v_writelane_b32 v0, s7, 35
+; GCN-NEXT: v_writelane_b32 v0, s8, 36
+; GCN-NEXT: v_writelane_b32 v0, s9, 37
+; GCN-NEXT: v_writelane_b32 v0, s10, 38
+; GCN-NEXT: v_writelane_b32 v0, s11, 39
+; GCN-NEXT: v_writelane_b32 v0, s12, 40
+; GCN-NEXT: v_writelane_b32 v0, s13, 41
+; GCN-NEXT: v_writelane_b32 v0, s14, 42
+; GCN-NEXT: v_writelane_b32 v0, s15, 43
+; GCN-NEXT: v_writelane_b32 v0, s16, 44
+; GCN-NEXT: v_writelane_b32 v0, s17, 45
+; GCN-NEXT: v_writelane_b32 v0, s18, 46
+; GCN-NEXT: v_writelane_b32 v0, s19, 47
; GCN-NEXT: ;;#ASMSTART
; GCN-NEXT: ; def s[4:19]
; GCN-NEXT: ;;#ASMEND
-; GCN-NEXT: v_writelane_b32 v1, s4, 48
-; GCN-NEXT: v_writelane_b32 v1, s5, 49
-; GCN-NEXT: v_writelane_b32 v1, s6, 50
-; GCN-NEXT: v_writelane_b32 v1, s7, 51
-; GCN-NEXT: v_writelane_b32 v1, s8, 52
-; GCN-NEXT: v_writelane_b32 v1, s9, 53
-; GCN-NEXT: v_writelane_b32 v1, s10, 54
-; GCN-NEXT: v_writelane_b32 v1, s11, 55
-; GCN-NEXT: v_writelane_b32 v1, s12, 56
-; GCN-NEXT: v_writelane_b32 v1, s13, 57
-; GCN-NEXT: v_writelane_b32 v1, s14, 58
-; GCN-NEXT: v_writelane_b32 v1, s15, 59
-; GCN-NEXT: v_writelane_b32 v1, s16, 60
-; GCN-NEXT: v_writelane_b32 v1, s17, 61
-; GCN-NEXT: v_writelane_b32 v1, s18, 62
-; GCN-NEXT: v_writelane_b32 v1, s19, 63
-; GCN-NEXT: s_or_saveexec_b64 s[28:29], -1
-; GCN-NEXT: buffer_store_dword v1, off, s[52:55], 0 offset:8 ; 4-byte Folded Spill
-; GCN-NEXT: s_mov_b64 exec, s[28:29]
+; GCN-NEXT: v_writelane_b32 v0, s4, 48
+; GCN-NEXT: v_writelane_b32 v0, s5, 49
+; GCN-NEXT: v_writelane_b32 v0, s6, 50
+; GCN-NEXT: v_writelane_b32 v0, s7, 51
+; GCN-NEXT: v_writelane_b32 v0, s8, 52
+; GCN-NEXT: v_writelane_b32 v0, s9, 53
+; GCN-NEXT: v_writelane_b32 v0, s10, 54
+; GCN-NEXT: v_writelane_b32 v0, s11, 55
+; GCN-NEXT: v_writelane_b32 v0, s12, 56
+; GCN-NEXT: v_writelane_b32 v0, s13, 57
+; GCN-NEXT: v_writelane_b32 v0, s14, 58
+; GCN-NEXT: v_writelane_b32 v0, s15, 59
+; GCN-NEXT: v_writelane_b32 v0, s16, 60
+; GCN-NEXT: v_writelane_b32 v0, s17, 61
+; GCN-NEXT: v_writelane_b32 v0, s18, 62
+; GCN-NEXT: v_writelane_b32 v0, s19, 63
; GCN-NEXT: ;;#ASMSTART
; GCN-NEXT: ; def s[4:11]
; GCN-NEXT: ;;#ASMEND
-; GCN-NEXT: v_writelane_b32 v0, s4, 0
-; GCN-NEXT: v_writelane_b32 v0, s5, 1
-; GCN-NEXT: v_writelane_b32 v0, s6, 2
-; GCN-NEXT: v_writelane_b32 v0, s7, 3
-; GCN-NEXT: v_writelane_b32 v0, s8, 4
-; GCN-NEXT: v_writelane_b32 v0, s9, 5
-; GCN-NEXT: v_writelane_b32 v0, s10, 6
-; GCN-NEXT: v_writelane_b32 v0, s11, 7
+; GCN-NEXT: v_writelane_b32 v1, s4, 0
+; GCN-NEXT: v_writelane_b32 v1, s5, 1
+; GCN-NEXT: v_writelane_b32 v1, s6, 2
+; GCN-NEXT: v_writelane_b32 v1, s7, 3
+; GCN-NEXT: v_writelane_b32 v1, s8, 4
+; GCN-NEXT: v_writelane_b32 v1, s9, 5
+; GCN-NEXT: v_writelane_b32 v1, s10, 6
+; GCN-NEXT: v_writelane_b32 v1, s11, 7
; GCN-NEXT: ;;#ASMSTART
; GCN-NEXT: ; def s[2:3]
; GCN-NEXT: ;;#ASMEND
-; GCN-NEXT: v_writelane_b32 v0, s2, 8
-; GCN-NEXT: v_writelane_b32 v0, s3, 9
-; GCN-NEXT: s_or_saveexec_b64 s[28:29], -1
-; GCN-NEXT: buffer_store_dword v0, off, s[52:55], 0 offset:4 ; 4-byte Folded Spill
-; GCN-NEXT: s_mov_b64 exec, s[28:29]
+; GCN-NEXT: v_writelane_b32 v1, s2, 8
+; GCN-NEXT: v_writelane_b32 v1, s3, 9
; GCN-NEXT: s_mov_b32 s1, 0
; GCN-NEXT: s_waitcnt lgkmcnt(0)
; GCN-NEXT: s_cmp_lg_u32 s0, s1
; GCN-NEXT: s_cbranch_scc1 .LBB1_2
; GCN-NEXT: ; %bb.1: ; %bb0
-; GCN-NEXT: s_or_saveexec_b64 s[28:29], -1
-; GCN-NEXT: buffer_load_dword v0, off, s[52:55], 0 offset:8 ; 4-byte Folded Reload
-; GCN-NEXT: s_mov_b64 exec, s[28:29]
-; GCN-NEXT: s_or_saveexec_b64 s[28:29], -1
-; GCN-NEXT: buffer_load_dword v1, off, s[52:55], 0 offset:4 ; 4-byte Folded Reload
-; GCN-NEXT: s_mov_b64 exec, s[28:29]
-; GCN-NEXT: s_waitcnt vmcnt(0)
; GCN-NEXT: v_readlane_b32 s16, v1, 8
; GCN-NEXT: v_readlane_b32 s17, v1, 9
; GCN-NEXT: v_readlane_b32 s20, v1, 0
@@ -694,14 +633,6 @@
; GCN-NEXT: ; use s[0:15]
; GCN-NEXT: ;;#ASMEND
; GCN-NEXT: .LBB1_2: ; %ret
-; GCN-NEXT: s_or_saveexec_b64 s[28:29], -1
-; GCN-NEXT: buffer_load_dword v0, off, s[52:55], 0 offset:4 ; 4-byte Folded Reload
-; GCN-NEXT: s_mov_b64 exec, s[28:29]
-; GCN-NEXT: s_or_saveexec_b64 s[28:29], -1
-; GCN-NEXT: buffer_load_dword v1, off, s[52:55], 0 offset:8 ; 4-byte Folded Reload
-; GCN-NEXT: s_mov_b64 exec, s[28:29]
-; GCN-NEXT: ; kill: killed $vgpr1
-; GCN-NEXT: ; kill: killed $vgpr0
; GCN-NEXT: s_endpgm
%wide.sgpr0 = call <16 x i32> asm sideeffect "; def $0", "=s" () #0
%wide.sgpr1 = call <16 x i32> asm sideeffect "; def $0", "=s" () #0
@@ -738,17 +669,9 @@
; GCN-NEXT: s_mov_b32 s55, 0xe8f000
; GCN-NEXT: s_add_u32 s52, s52, s11
; GCN-NEXT: s_addc_u32 s53, s53, 0
-; GCN-NEXT: ; implicit-def: $vgpr0
-; GCN-NEXT: ; implicit-def: $vgpr0
; GCN-NEXT: s_load_dword s0, s[4:5], 0xb
; GCN-NEXT: ;;#ASMSTART
; GCN-NEXT: ;;#ASMEND
-; GCN-NEXT: s_or_saveexec_b64 s[34:35], -1
-; GCN-NEXT: buffer_load_dword v1, off, s[52:55], 0 offset:8 ; 4-byte Folded Reload
-; GCN-NEXT: s_mov_b64 exec, s[34:35]
-; GCN-NEXT: s_or_saveexec_b64 s[34:35], -1
-; GCN-NEXT: buffer_load_dword v0, off, s[52:55], 0 offset:4 ; 4-byte Folded Reload
-; GCN-NEXT: s_mov_b64 exec, s[34:35]
; GCN-NEXT: ;;#ASMSTART
; GCN-NEXT: ;;#ASMEND
; GCN-NEXT: ;;#ASMSTART
@@ -762,176 +685,176 @@
; GCN-NEXT: ;;#ASMSTART
; GCN-NEXT: ; def s[4:19]
; GCN-NEXT: ;;#ASMEND
-; GCN-NEXT: s_waitcnt vmcnt(1)
-; GCN-NEXT: v_writelane_b32 v1, s4, 0
-; GCN-NEXT: v_writelane_b32 v1, s5, 1
-; GCN-NEXT: v_writelane_b32 v1, s6, 2
-; GCN-NEXT: v_writelane_b32 v1, s7, 3
-; GCN-NEXT: v_writelane_b32 v1, s8, 4
-; GCN-NEXT: v_writelane_b32 v1, s9, 5
-; GCN-NEXT: v_writelane_b32 v1, s10, 6
-; GCN-NEXT: v_writelane_b32 v1, s11, 7
-; GCN-NEXT: v_writelane_b32 v1, s12, 8
-; GCN-NEXT: v_writelane_b32 v1, s13, 9
-; GCN-NEXT: v_writelane_b32 v1, s14, 10
-; GCN-NEXT: v_writelane_b32 v1, s15, 11
-; GCN-NEXT: v_writelane_b32 v1, s16, 12
-; GCN-NEXT: v_writelane_b32 v1, s17, 13
-; GCN-NEXT: v_writelane_b32 v1, s18, 14
-; GCN-NEXT: v_writelane_b32 v1, s19, 15
+; GCN-NEXT: v_writelane_b32 v31, s4, 0
+; GCN-NEXT: v_writelane_b32 v31, s5, 1
+; GCN-NEXT: v_writelane_b32 v31, s6, 2
+; GCN-NEXT: v_writelane_b32 v31, s7, 3
+; GCN-NEXT: v_writelane_b32 v31, s8, 4
+; GCN-NEXT: v_writelane_b32 v31, s9, 5
+; GCN-NEXT: v_writelane_b32 v31, s10, 6
+; GCN-NEXT: v_writelane_b32 v31, s11, 7
+; GCN-NEXT: v_writelane_b32 v31, s12, 8
+; GCN-NEXT: v_writelane_b32 v31, s13, 9
+; GCN-NEXT: v_writelane_b32 v31, s14, 10
+; GCN-NEXT: v_writelane_b32 v31, s15, 11
+; GCN-NEXT: v_writelane_b32 v31, s16, 12
+; GCN-NEXT: v_writelane_b32 v31, s17, 13
+; GCN-NEXT: v_writelane_b32 v31, s18, 14
+; GCN-NEXT: v_writelane_b32 v31, s19, 15
; GCN-NEXT: ;;#ASMSTART
; GCN-NEXT: ; def s[4:19]
; GCN-NEXT: ;;#ASMEND
-; GCN-NEXT: v_writelane_b32 v1, s4, 16
-; GCN-NEXT: v_writelane_b32 v1, s5, 17
-; GCN-NEXT: v_writelane_b32 v1, s6, 18
-; GCN-NEXT: v_writelane_b32 v1, s7, 19
-; GCN-NEXT: v_writelane_b32 v1, s8, 20
-; GCN-NEXT: v_writelane_b32 v1, s9, 21
-; GCN-NEXT: v_writelane_b32 v1, s10, 22
-; GCN-NEXT: v_writelane_b32 v1, s11, 23
-; GCN-NEXT: v_writelane_b32 v1, s12, 24
-; GCN-NEXT: v_writelane_b32 v1, s13, 25
-; GCN-NEXT: v_writelane_b32 v1, s14, 26
-; GCN-NEXT: v_writelane_b32 v1, s15, 27
-; GCN-NEXT: v_writelane_b32 v1, s16, 28
-; GCN-NEXT: v_writelane_b32 v1, s17, 29
-; GCN-NEXT: v_writelane_b32 v1, s18, 30
-; GCN-NEXT: v_writelane_b32 v1, s19, 31
+; GCN-NEXT: v_writelane_b32 v31, s4, 16
+; GCN-NEXT: v_writelane_b32 v31, s5, 17
+; GCN-NEXT: v_writelane_b32 v31, s6, 18
+; GCN-NEXT: v_writelane_b32 v31, s7, 19
+; GCN-NEXT: v_writelane_b32 v31, s8, 20
+; GCN-NEXT: v_writelane_b32 v31, s9, 21
+; GCN-NEXT: v_writelane_b32 v31, s10, 22
+; GCN-NEXT: v_writelane_b32 v31, s11, 23
+; GCN-NEXT: v_writelane_b32 v31, s12, 24
+; GCN-NEXT: v_writelane_b32 v31, s13, 25
+; GCN-NEXT: v_writelane_b32 v31, s14, 26
+; GCN-NEXT: v_writelane_b32 v31, s15, 27
+; GCN-NEXT: v_writelane_b32 v31, s16, 28
+; GCN-NEXT: v_writelane_b32 v31, s17, 29
+; GCN-NEXT: v_writelane_b32 v31, s18, 30
+; GCN-NEXT: v_writelane_b32 v31, s19, 31
; GCN-NEXT: ;;#ASMSTART
; GCN-NEXT: ; def s[4:19]
; GCN-NEXT: ;;#ASMEND
-; GCN-NEXT: v_writelane_b32 v1, s4, 32
-; GCN-NEXT: v_writelane_b32 v1, s5, 33
-; GCN-NEXT: v_writelane_b32 v1, s6, 34
-; GCN-NEXT: v_writelane_b32 v1, s7, 35
-; GCN-NEXT: v_writelane_b32 v1, s8, 36
-; GCN-NEXT: v_writelane_b32 v1, s9, 37
-; GCN-NEXT: v_writelane_b32 v1, s10, 38
-; GCN-NEXT: v_writelane_b32 v1, s11, 39
-; GCN-NEXT: v_writelane_b32 v1, s12, 40
-; GCN-NEXT: v_writelane_b32 v1, s13, 41
-; GCN-NEXT: v_writelane_b32 v1, s14, 42
-; GCN-NEXT: v_writelane_b32 v1, s15, 43
-; GCN-NEXT: v_writelane_b32 v1, s16, 44
-; GCN-NEXT: v_writelane_b32 v1, s17, 45
-; GCN-NEXT: v_writelane_b32 v1, s18, 46
-; GCN-NEXT: v_writelane_b32 v1, s19, 47
+; GCN-NEXT: v_writelane_b32 v31, s4, 32
+; GCN-NEXT: v_writelane_b32 v31, s5, 33
+; GCN-NEXT: v_writelane_b32 v31, s6, 34
+; GCN-NEXT: v_writelane_b32 v31, s7, 35
+; GCN-NEXT: v_writelane_b32 v31, s8, 36
+; GCN-NEXT: v_writelane_b32 v31, s9, 37
+; GCN-NEXT: v_writelane_b32 v31, s10, 38
+; GCN-NEXT: v_writelane_b32 v31, s11, 39
+; GCN-NEXT: v_writelane_b32 v31, s12, 40
+; GCN-NEXT: v_writelane_b32 v31, s13, 41
+; GCN-NEXT: v_writelane_b32 v31, s14, 42
+; GCN-NEXT: v_writelane_b32 v31, s15, 43
+; GCN-NEXT: v_writelane_b32 v31, s16, 44
+; GCN-NEXT: v_writelane_b32 v31, s17, 45
+; GCN-NEXT: v_writelane_b32 v31, s18, 46
+; GCN-NEXT: v_writelane_b32 v31, s19, 47
; GCN-NEXT: ;;#ASMSTART
; GCN-NEXT: ; def s[4:19]
; GCN-NEXT: ;;#ASMEND
-; GCN-NEXT: v_writelane_b32 v1, s4, 48
-; GCN-NEXT: v_writelane_b32 v1, s5, 49
-; GCN-NEXT: v_writelane_b32 v1, s6, 50
-; GCN-NEXT: v_writelane_b32 v1, s7, 51
-; GCN-NEXT: v_writelane_b32 v1, s8, 52
-; GCN-NEXT: v_writelane_b32 v1, s9, 53
-; GCN-NEXT: v_writelane_b32 v1, s10, 54
-; GCN-NEXT: v_writelane_b32 v1, s11, 55
-; GCN-NEXT: v_writelane_b32 v1, s12, 56
-; GCN-NEXT: v_writelane_b32 v1, s13, 57
-; GCN-NEXT: v_writelane_b32 v1, s14, 58
-; GCN-NEXT: v_writelane_b32 v1, s15, 59
-; GCN-NEXT: v_writelane_b32 v1, s16, 60
-; GCN-NEXT: v_writelane_b32 v1, s17, 61
-; GCN-NEXT: v_writelane_b32 v1, s18, 62
-; GCN-NEXT: v_writelane_b32 v1, s19, 63
-; GCN-NEXT: s_or_saveexec_b64 s[34:35], -1
-; GCN-NEXT: buffer_store_dword v1, off, s[52:55], 0 offset:8 ; 4-byte Folded Spill
-; GCN-NEXT: s_mov_b64 exec, s[34:35]
+; GCN-NEXT: v_writelane_b32 v31, s4, 48
+; GCN-NEXT: v_writelane_b32 v31, s5, 49
+; GCN-NEXT: v_writelane_b32 v31, s6, 50
+; GCN-NEXT: v_writelane_b32 v31, s7, 51
+; GCN-NEXT: v_writelane_b32 v31, s8, 52
+; GCN-NEXT: v_writelane_b32 v31, s9, 53
+; GCN-NEXT: v_writelane_b32 v31, s10, 54
+; GCN-NEXT: v_writelane_b32 v31, s11, 55
+; GCN-NEXT: v_writelane_b32 v31, s12, 56
+; GCN-NEXT: v_writelane_b32 v31, s13, 57
+; GCN-NEXT: v_writelane_b32 v31, s14, 58
+; GCN-NEXT: v_writelane_b32 v31, s15, 59
+; GCN-NEXT: v_writelane_b32 v31, s16, 60
+; GCN-NEXT: v_writelane_b32 v31, s17, 61
+; GCN-NEXT: v_writelane_b32 v31, s18, 62
+; GCN-NEXT: v_writelane_b32 v31, s19, 63
; GCN-NEXT: ;;#ASMSTART
; GCN-NEXT: ; def s[2:3]
; GCN-NEXT: ;;#ASMEND
-; GCN-NEXT: s_waitcnt vmcnt(1)
+; GCN-NEXT: s_mov_b64 s[4:5], exec
+; GCN-NEXT: s_mov_b64 exec, 3
+; GCN-NEXT: buffer_store_dword v0, off, s[52:55], 0
; GCN-NEXT: v_writelane_b32 v0, s2, 0
; GCN-NEXT: v_writelane_b32 v0, s3, 1
-; GCN-NEXT: s_or_saveexec_b64 s[34:35], -1
; GCN-NEXT: buffer_store_dword v0, off, s[52:55], 0 offset:4 ; 4-byte Folded Spill
-; GCN-NEXT: s_mov_b64 exec, s[34:35]
+; GCN-NEXT: buffer_load_dword v0, off, s[52:55], 0
+; GCN-NEXT: s_waitcnt vmcnt(0)
+; GCN-NEXT: s_mov_b64 exec, s[4:5]
; GCN-NEXT: s_mov_b32 s1, 0
; GCN-NEXT: s_waitcnt lgkmcnt(0)
; GCN-NEXT: s_cmp_lg_u32 s0, s1
; GCN-NEXT: s_cbranch_scc1 .LBB2_2
; GCN-NEXT: ; %bb.1: ; %bb0
-; GCN-NEXT: s_or_saveexec_b64 s[34:35], -1
-; GCN-NEXT: buffer_load_dword v0, off, s[52:55], 0 offset:4 ; 4-byte Folded Reload
-; GCN-NEXT: s_mov_b64 exec, s[34:35]
-; GCN-NEXT: s_or_saveexec_b64 s[34:35], -1
-; GCN-NEXT: buffer_load_dword v1, off, s[52:55], 0 offset:8 ; 4-byte Folded Reload
-; GCN-NEXT: s_mov_b64 exec, s[34:35]
-; GCN-NEXT: s_waitcnt vmcnt(0)
-; GCN-NEXT: v_readlane_b32 s36, v1, 32
-; GCN-NEXT: v_readlane_b32 s37, v1, 33
-; GCN-NEXT: v_readlane_b32 s38, v1, 34
-; GCN-NEXT: v_readlane_b32 s39, v1, 35
-; GCN-NEXT: v_readlane_b32 s40, v1, 36
-; GCN-NEXT: v_readlane_b32 s41, v1, 37
-; GCN-NEXT: v_readlane_b32 s42, v1, 38
-; GCN-NEXT: v_readlane_b32 s43, v1, 39
-; GCN-NEXT: v_readlane_b32 s44, v1, 40
-; GCN-NEXT: v_readlane_b32 s45, v1, 41
-; GCN-NEXT: v_readlane_b32 s46, v1, 42
-; GCN-NEXT: v_readlane_b32 s47, v1, 43
-; GCN-NEXT: v_readlane_b32 s48, v1, 44
-; GCN-NEXT: v_readlane_b32 s49, v1, 45
-; GCN-NEXT: v_readlane_b32 s50, v1, 46
-; GCN-NEXT: v_readlane_b32 s51, v1, 47
-; GCN-NEXT: v_readlane_b32 s0, v1, 16
-; GCN-NEXT: v_readlane_b32 s1, v1, 17
-; GCN-NEXT: v_readlane_b32 s2, v1, 18
-; GCN-NEXT: v_readlane_b32 s3, v1, 19
-; GCN-NEXT: v_readlane_b32 s4, v1, 20
-; GCN-NEXT: v_readlane_b32 s5, v1, 21
-; GCN-NEXT: v_readlane_b32 s6, v1, 22
-; GCN-NEXT: v_readlane_b32 s7, v1, 23
-; GCN-NEXT: v_readlane_b32 s8, v1, 24
-; GCN-NEXT: v_readlane_b32 s9, v1, 25
-; GCN-NEXT: v_readlane_b32 s10, v1, 26
-; GCN-NEXT: v_readlane_b32 s11, v1, 27
-; GCN-NEXT: v_readlane_b32 s12, v1, 28
-; GCN-NEXT: v_readlane_b32 s13, v1, 29
-; GCN-NEXT: v_readlane_b32 s14, v1, 30
-; GCN-NEXT: v_readlane_b32 s15, v1, 31
-; GCN-NEXT: v_readlane_b32 s16, v1, 0
-; GCN-NEXT: v_readlane_b32 s17, v1, 1
-; GCN-NEXT: v_readlane_b32 s18, v1, 2
-; GCN-NEXT: v_readlane_b32 s19, v1, 3
-; GCN-NEXT: v_readlane_b32 s20, v1, 4
-; GCN-NEXT: v_readlane_b32 s21, v1, 5
-; GCN-NEXT: v_readlane_b32 s22, v1, 6
-; GCN-NEXT: v_readlane_b32 s23, v1, 7
-; GCN-NEXT: v_readlane_b32 s24, v1, 8
-; GCN-NEXT: v_readlane_b32 s25, v1, 9
-; GCN-NEXT: v_readlane_b32 s26, v1, 10
-; GCN-NEXT: v_readlane_b32 s27, v1, 11
-; GCN-NEXT: v_readlane_b32 s28, v1, 12
-; GCN-NEXT: v_readlane_b32 s29, v1, 13
-; GCN-NEXT: v_readlane_b32 s30, v1, 14
-; GCN-NEXT: v_readlane_b32 s31, v1, 15
+; GCN-NEXT: v_readlane_b32 s36, v31, 32
+; GCN-NEXT: v_readlane_b32 s37, v31, 33
+; GCN-NEXT: v_readlane_b32 s38, v31, 34
+; GCN-NEXT: v_readlane_b32 s39, v31, 35
+; GCN-NEXT: v_readlane_b32 s40, v31, 36
+; GCN-NEXT: v_readlane_b32 s41, v31, 37
+; GCN-NEXT: v_readlane_b32 s42, v31, 38
+; GCN-NEXT: v_readlane_b32 s43, v31, 39
+; GCN-NEXT: v_readlane_b32 s44, v31, 40
+; GCN-NEXT: v_readlane_b32 s45, v31, 41
+; GCN-NEXT: v_readlane_b32 s46, v31, 42
+; GCN-NEXT: v_readlane_b32 s47, v31, 43
+; GCN-NEXT: v_readlane_b32 s48, v31, 44
+; GCN-NEXT: v_readlane_b32 s49, v31, 45
+; GCN-NEXT: v_readlane_b32 s50, v31, 46
+; GCN-NEXT: v_readlane_b32 s51, v31, 47
+; GCN-NEXT: v_readlane_b32 s0, v31, 16
+; GCN-NEXT: v_readlane_b32 s1, v31, 17
+; GCN-NEXT: v_readlane_b32 s2, v31, 18
+; GCN-NEXT: v_readlane_b32 s3, v31, 19
+; GCN-NEXT: v_readlane_b32 s4, v31, 20
+; GCN-NEXT: v_readlane_b32 s5, v31, 21
+; GCN-NEXT: v_readlane_b32 s6, v31, 22
+; GCN-NEXT: v_readlane_b32 s7, v31, 23
+; GCN-NEXT: v_readlane_b32 s8, v31, 24
+; GCN-NEXT: v_readlane_b32 s9, v31, 25
+; GCN-NEXT: v_readlane_b32 s10, v31, 26
+; GCN-NEXT: v_readlane_b32 s11, v31, 27
+; GCN-NEXT: v_readlane_b32 s12, v31, 28
+; GCN-NEXT: v_readlane_b32 s13, v31, 29
+; GCN-NEXT: v_readlane_b32 s14, v31, 30
+; GCN-NEXT: v_readlane_b32 s15, v31, 31
+; GCN-NEXT: v_readlane_b32 s16, v31, 0
+; GCN-NEXT: v_readlane_b32 s17, v31, 1
+; GCN-NEXT: v_readlane_b32 s18, v31, 2
+; GCN-NEXT: v_readlane_b32 s19, v31, 3
+; GCN-NEXT: v_readlane_b32 s20, v31, 4
+; GCN-NEXT: v_readlane_b32 s21, v31, 5
+; GCN-NEXT: v_readlane_b32 s22, v31, 6
+; GCN-NEXT: v_readlane_b32 s23, v31, 7
+; GCN-NEXT: v_readlane_b32 s24, v31, 8
+; GCN-NEXT: v_readlane_b32 s25, v31, 9
+; GCN-NEXT: v_readlane_b32 s26, v31, 10
+; GCN-NEXT: v_readlane_b32 s27, v31, 11
+; GCN-NEXT: v_readlane_b32 s28, v31, 12
+; GCN-NEXT: v_readlane_b32 s29, v31, 13
+; GCN-NEXT: v_readlane_b32 s30, v31, 14
+; GCN-NEXT: v_readlane_b32 s31, v31, 15
; GCN-NEXT: ;;#ASMSTART
; GCN-NEXT: ; use s[16:31]
; GCN-NEXT: ;;#ASMEND
; GCN-NEXT: ;;#ASMSTART
; GCN-NEXT: ; use s[0:15]
; GCN-NEXT: ;;#ASMEND
-; GCN-NEXT: v_readlane_b32 s4, v1, 48
-; GCN-NEXT: v_readlane_b32 s5, v1, 49
-; GCN-NEXT: v_readlane_b32 s6, v1, 50
-; GCN-NEXT: v_readlane_b32 s7, v1, 51
-; GCN-NEXT: v_readlane_b32 s8, v1, 52
-; GCN-NEXT: v_readlane_b32 s9, v1, 53
-; GCN-NEXT: v_readlane_b32 s10, v1, 54
-; GCN-NEXT: v_readlane_b32 s11, v1, 55
-; GCN-NEXT: v_readlane_b32 s12, v1, 56
-; GCN-NEXT: v_readlane_b32 s13, v1, 57
-; GCN-NEXT: v_readlane_b32 s14, v1, 58
-; GCN-NEXT: v_readlane_b32 s15, v1, 59
-; GCN-NEXT: v_readlane_b32 s16, v1, 60
-; GCN-NEXT: v_readlane_b32 s17, v1, 61
-; GCN-NEXT: v_readlane_b32 s18, v1, 62
-; GCN-NEXT: v_readlane_b32 s19, v1, 63
+; GCN-NEXT: v_readlane_b32 s4, v31, 48
+; GCN-NEXT: v_readlane_b32 s5, v31, 49
+; GCN-NEXT: v_readlane_b32 s6, v31, 50
+; GCN-NEXT: v_readlane_b32 s7, v31, 51
+; GCN-NEXT: v_readlane_b32 s8, v31, 52
+; GCN-NEXT: v_readlane_b32 s9, v31, 53
+; GCN-NEXT: v_readlane_b32 s10, v31, 54
+; GCN-NEXT: v_readlane_b32 s11, v31, 55
+; GCN-NEXT: v_readlane_b32 s12, v31, 56
+; GCN-NEXT: v_readlane_b32 s13, v31, 57
+; GCN-NEXT: v_readlane_b32 s14, v31, 58
+; GCN-NEXT: v_readlane_b32 s15, v31, 59
+; GCN-NEXT: v_readlane_b32 s16, v31, 60
+; GCN-NEXT: v_readlane_b32 s17, v31, 61
+; GCN-NEXT: v_readlane_b32 s18, v31, 62
+; GCN-NEXT: v_readlane_b32 s19, v31, 63
+; GCN-NEXT: s_mov_b64 s[2:3], exec
+; GCN-NEXT: s_mov_b64 exec, 3
+; GCN-NEXT: buffer_store_dword v0, off, s[52:55], 0
+; GCN-NEXT: buffer_load_dword v0, off, s[52:55], 0 offset:4 ; 4-byte Folded Reload
+; GCN-NEXT: s_waitcnt vmcnt(0)
; GCN-NEXT: v_readlane_b32 s0, v0, 0
; GCN-NEXT: v_readlane_b32 s1, v0, 1
+; GCN-NEXT: buffer_load_dword v0, off, s[52:55], 0
+; GCN-NEXT: s_waitcnt vmcnt(0)
+; GCN-NEXT: s_mov_b64 exec, s[2:3]
; GCN-NEXT: ;;#ASMSTART
; GCN-NEXT: ; use s[36:51]
; GCN-NEXT: ;;#ASMEND
@@ -942,14 +865,6 @@
; GCN-NEXT: ; use s[0:1]
; GCN-NEXT: ;;#ASMEND
; GCN-NEXT: .LBB2_2: ; %ret
-; GCN-NEXT: s_or_saveexec_b64 s[34:35], -1
-; GCN-NEXT: buffer_load_dword v0, off, s[52:55], 0 offset:4 ; 4-byte Folded Reload
-; GCN-NEXT: s_mov_b64 exec, s[34:35]
-; GCN-NEXT: s_or_saveexec_b64 s[34:35], -1
-; GCN-NEXT: buffer_load_dword v1, off, s[52:55], 0 offset:8 ; 4-byte Folded Reload
-; GCN-NEXT: s_mov_b64 exec, s[34:35]
-; GCN-NEXT: ; kill: killed $vgpr1
-; GCN-NEXT: ; kill: killed $vgpr0
; GCN-NEXT: s_endpgm
call void asm sideeffect "", "~{v[0:7]}" () #0
call void asm sideeffect "", "~{v[8:15]}" () #0
@@ -989,17 +904,9 @@
; GCN-NEXT: s_mov_b32 s55, 0xe8f000
; GCN-NEXT: s_add_u32 s52, s52, s11
; GCN-NEXT: s_addc_u32 s53, s53, 0
-; GCN-NEXT: ; implicit-def: $vgpr0
-; GCN-NEXT: ; implicit-def: $vgpr0
; GCN-NEXT: s_load_dword s0, s[4:5], 0x9
; GCN-NEXT: ;;#ASMSTART
; GCN-NEXT: ;;#ASMEND
-; GCN-NEXT: s_or_saveexec_b64 s[34:35], -1
-; GCN-NEXT: buffer_load_dword v1, off, s[52:55], 0 offset:8 ; 4-byte Folded Reload
-; GCN-NEXT: s_mov_b64 exec, s[34:35]
-; GCN-NEXT: s_or_saveexec_b64 s[34:35], -1
-; GCN-NEXT: buffer_load_dword v0, off, s[52:55], 0 offset:4 ; 4-byte Folded Reload
-; GCN-NEXT: s_mov_b64 exec, s[34:35]
; GCN-NEXT: ;;#ASMSTART
; GCN-NEXT: ;;#ASMEND
; GCN-NEXT: ;;#ASMSTART
@@ -1013,152 +920,144 @@
; GCN-NEXT: ;;#ASMSTART
; GCN-NEXT: ; def s[4:19]
; GCN-NEXT: ;;#ASMEND
-; GCN-NEXT: s_waitcnt vmcnt(1)
-; GCN-NEXT: v_writelane_b32 v1, s4, 0
-; GCN-NEXT: v_writelane_b32 v1, s5, 1
-; GCN-NEXT: v_writelane_b32 v1, s6, 2
-; GCN-NEXT: v_writelane_b32 v1, s7, 3
-; GCN-NEXT: v_writelane_b32 v1, s8, 4
-; GCN-NEXT: v_writelane_b32 v1, s9, 5
-; GCN-NEXT: v_writelane_b32 v1, s10, 6
-; GCN-NEXT: v_writelane_b32 v1, s11, 7
-; GCN-NEXT: v_writelane_b32 v1, s12, 8
-; GCN-NEXT: v_writelane_b32 v1, s13, 9
-; GCN-NEXT: v_writelane_b32 v1, s14, 10
-; GCN-NEXT: v_writelane_b32 v1, s15, 11
-; GCN-NEXT: v_writelane_b32 v1, s16, 12
-; GCN-NEXT: v_writelane_b32 v1, s17, 13
-; GCN-NEXT: v_writelane_b32 v1, s18, 14
-; GCN-NEXT: v_writelane_b32 v1, s19, 15
+; GCN-NEXT: v_writelane_b32 v31, s4, 0
+; GCN-NEXT: v_writelane_b32 v31, s5, 1
+; GCN-NEXT: v_writelane_b32 v31, s6, 2
+; GCN-NEXT: v_writelane_b32 v31, s7, 3
+; GCN-NEXT: v_writelane_b32 v31, s8, 4
+; GCN-NEXT: v_writelane_b32 v31, s9, 5
+; GCN-NEXT: v_writelane_b32 v31, s10, 6
+; GCN-NEXT: v_writelane_b32 v31, s11, 7
+; GCN-NEXT: v_writelane_b32 v31, s12, 8
+; GCN-NEXT: v_writelane_b32 v31, s13, 9
+; GCN-NEXT: v_writelane_b32 v31, s14, 10
+; GCN-NEXT: v_writelane_b32 v31, s15, 11
+; GCN-NEXT: v_writelane_b32 v31, s16, 12
+; GCN-NEXT: v_writelane_b32 v31, s17, 13
+; GCN-NEXT: v_writelane_b32 v31, s18, 14
+; GCN-NEXT: v_writelane_b32 v31, s19, 15
; GCN-NEXT: ;;#ASMSTART
; GCN-NEXT: ; def s[4:19]
; GCN-NEXT: ;;#ASMEND
-; GCN-NEXT: v_writelane_b32 v1, s4, 16
-; GCN-NEXT: v_writelane_b32 v1, s5, 17
-; GCN-NEXT: v_writelane_b32 v1, s6, 18
-; GCN-NEXT: v_writelane_b32 v1, s7, 19
-; GCN-NEXT: v_writelane_b32 v1, s8, 20
-; GCN-NEXT: v_writelane_b32 v1, s9, 21
-; GCN-NEXT: v_writelane_b32 v1, s10, 22
-; GCN-NEXT: v_writelane_b32 v1, s11, 23
-; GCN-NEXT: v_writelane_b32 v1, s12, 24
-; GCN-NEXT: v_writelane_b32 v1, s13, 25
-; GCN-NEXT: v_writelane_b32 v1, s14, 26
-; GCN-NEXT: v_writelane_b32 v1, s15, 27
-; GCN-NEXT: v_writelane_b32 v1, s16, 28
-; GCN-NEXT: v_writelane_b32 v1, s17, 29
-; GCN-NEXT: v_writelane_b32 v1, s18, 30
-; GCN-NEXT: v_writelane_b32 v1, s19, 31
+; GCN-NEXT: v_writelane_b32 v31, s4, 16
+; GCN-NEXT: v_writelane_b32 v31, s5, 17
+; GCN-NEXT: v_writelane_b32 v31, s6, 18
+; GCN-NEXT: v_writelane_b32 v31, s7, 19
+; GCN-NEXT: v_writelane_b32 v31, s8, 20
+; GCN-NEXT: v_writelane_b32 v31, s9, 21
+; GCN-NEXT: v_writelane_b32 v31, s10, 22
+; GCN-NEXT: v_writelane_b32 v31, s11, 23
+; GCN-NEXT: v_writelane_b32 v31, s12, 24
+; GCN-NEXT: v_writelane_b32 v31, s13, 25
+; GCN-NEXT: v_writelane_b32 v31, s14, 26
+; GCN-NEXT: v_writelane_b32 v31, s15, 27
+; GCN-NEXT: v_writelane_b32 v31, s16, 28
+; GCN-NEXT: v_writelane_b32 v31, s17, 29
+; GCN-NEXT: v_writelane_b32 v31, s18, 30
+; GCN-NEXT: v_writelane_b32 v31, s19, 31
; GCN-NEXT: ;;#ASMSTART
; GCN-NEXT: ; def s[4:19]
; GCN-NEXT: ;;#ASMEND
-; GCN-NEXT: v_writelane_b32 v1, s4, 32
-; GCN-NEXT: v_writelane_b32 v1, s5, 33
-; GCN-NEXT: v_writelane_b32 v1, s6, 34
-; GCN-NEXT: v_writelane_b32 v1, s7, 35
-; GCN-NEXT: v_writelane_b32 v1, s8, 36
-; GCN-NEXT: v_writelane_b32 v1, s9, 37
-; GCN-NEXT: v_writelane_b32 v1, s10, 38
-; GCN-NEXT: v_writelane_b32 v1, s11, 39
-; GCN-NEXT: v_writelane_b32 v1, s12, 40
-; GCN-NEXT: v_writelane_b32 v1, s13, 41
-; GCN-NEXT: v_writelane_b32 v1, s14, 42
-; GCN-NEXT: v_writelane_b32 v1, s15, 43
-; GCN-NEXT: v_writelane_b32 v1, s16, 44
-; GCN-NEXT: v_writelane_b32 v1, s17, 45
-; GCN-NEXT: v_writelane_b32 v1, s18, 46
-; GCN-NEXT: v_writelane_b32 v1, s19, 47
+; GCN-NEXT: v_writelane_b32 v31, s4, 32
+; GCN-NEXT: v_writelane_b32 v31, s5, 33
+; GCN-NEXT: v_writelane_b32 v31, s6, 34
+; GCN-NEXT: v_writelane_b32 v31, s7, 35
+; GCN-NEXT: v_writelane_b32 v31, s8, 36
+; GCN-NEXT: v_writelane_b32 v31, s9, 37
+; GCN-NEXT: v_writelane_b32 v31, s10, 38
+; GCN-NEXT: v_writelane_b32 v31, s11, 39
+; GCN-NEXT: v_writelane_b32 v31, s12, 40
+; GCN-NEXT: v_writelane_b32 v31, s13, 41
+; GCN-NEXT: v_writelane_b32 v31, s14, 42
+; GCN-NEXT: v_writelane_b32 v31, s15, 43
+; GCN-NEXT: v_writelane_b32 v31, s16, 44
+; GCN-NEXT: v_writelane_b32 v31, s17, 45
+; GCN-NEXT: v_writelane_b32 v31, s18, 46
+; GCN-NEXT: v_writelane_b32 v31, s19, 47
; GCN-NEXT: ;;#ASMSTART
; GCN-NEXT: ; def s[4:19]
; GCN-NEXT: ;;#ASMEND
-; GCN-NEXT: v_writelane_b32 v1, s4, 48
-; GCN-NEXT: v_writelane_b32 v1, s5, 49
-; GCN-NEXT: v_writelane_b32 v1, s6, 50
-; GCN-NEXT: v_writelane_b32 v1, s7, 51
-; GCN-NEXT: v_writelane_b32 v1, s8, 52
-; GCN-NEXT: v_writelane_b32 v1, s9, 53
-; GCN-NEXT: v_writelane_b32 v1, s10, 54
-; GCN-NEXT: v_writelane_b32 v1, s11, 55
-; GCN-NEXT: v_writelane_b32 v1, s12, 56
-; GCN-NEXT: v_writelane_b32 v1, s13, 57
-; GCN-NEXT: v_writelane_b32 v1, s14, 58
-; GCN-NEXT: v_writelane_b32 v1, s15, 59
-; GCN-NEXT: v_writelane_b32 v1, s16, 60
-; GCN-NEXT: v_writelane_b32 v1, s17, 61
-; GCN-NEXT: v_writelane_b32 v1, s18, 62
-; GCN-NEXT: v_writelane_b32 v1, s19, 63
-; GCN-NEXT: s_or_saveexec_b64 s[34:35], -1
-; GCN-NEXT: buffer_store_dword v1, off, s[52:55], 0 offset:8 ; 4-byte Folded Spill
-; GCN-NEXT: s_mov_b64 exec, s[34:35]
+; GCN-NEXT: v_writelane_b32 v31, s4, 48
+; GCN-NEXT: v_writelane_b32 v31, s5, 49
+; GCN-NEXT: v_writelane_b32 v31, s6, 50
+; GCN-NEXT: v_writelane_b32 v31, s7, 51
+; GCN-NEXT: v_writelane_b32 v31, s8, 52
+; GCN-NEXT: v_writelane_b32 v31, s9, 53
+; GCN-NEXT: v_writelane_b32 v31, s10, 54
+; GCN-NEXT: v_writelane_b32 v31, s11, 55
+; GCN-NEXT: v_writelane_b32 v31, s12, 56
+; GCN-NEXT: v_writelane_b32 v31, s13, 57
+; GCN-NEXT: v_writelane_b32 v31, s14, 58
+; GCN-NEXT: v_writelane_b32 v31, s15, 59
+; GCN-NEXT: v_writelane_b32 v31, s16, 60
+; GCN-NEXT: v_writelane_b32 v31, s17, 61
+; GCN-NEXT: v_writelane_b32 v31, s18, 62
+; GCN-NEXT: v_writelane_b32 v31, s19, 63
; GCN-NEXT: ;;#ASMSTART
; GCN-NEXT: ; def s[2:3]
; GCN-NEXT: ;;#ASMEND
-; GCN-NEXT: s_waitcnt vmcnt(1)
+; GCN-NEXT: s_mov_b64 s[4:5], exec
+; GCN-NEXT: s_mov_b64 exec, 3
+; GCN-NEXT: buffer_store_dword v0, off, s[52:55], 0
; GCN-NEXT: v_writelane_b32 v0, s2, 0
; GCN-NEXT: v_writelane_b32 v0, s3, 1
-; GCN-NEXT: s_or_saveexec_b64 s[34:35], -1
; GCN-NEXT: buffer_store_dword v0, off, s[52:55], 0 offset:4 ; 4-byte Folded Spill
-; GCN-NEXT: s_mov_b64 exec, s[34:35]
+; GCN-NEXT: buffer_load_dword v0, off, s[52:55], 0
+; GCN-NEXT: s_waitcnt vmcnt(0)
+; GCN-NEXT: s_mov_b64 exec, s[4:5]
; GCN-NEXT: s_mov_b32 s1, 0
; GCN-NEXT: s_waitcnt lgkmcnt(0)
; GCN-NEXT: s_cmp_lg_u32 s0, s1
; GCN-NEXT: s_cbranch_scc1 .LBB3_2
; GCN-NEXT: ; %bb.1: ; %bb0
-; GCN-NEXT: s_or_saveexec_b64 s[34:35], -1
-; GCN-NEXT: buffer_load_dword v1, off, s[52:55], 0 offset:4 ; 4-byte Folded Reload
-; GCN-NEXT: s_mov_b64 exec, s[34:35]
-; GCN-NEXT: s_or_saveexec_b64 s[34:35], -1
-; GCN-NEXT: buffer_load_dword v2, off, s[52:55], 0 offset:8 ; 4-byte Folded Reload
-; GCN-NEXT: s_mov_b64 exec, s[34:35]
-; GCN-NEXT: s_waitcnt vmcnt(0)
-; GCN-NEXT: v_readlane_b32 s36, v2, 32
-; GCN-NEXT: v_readlane_b32 s37, v2, 33
-; GCN-NEXT: v_readlane_b32 s38, v2, 34
-; GCN-NEXT: v_readlane_b32 s39, v2, 35
-; GCN-NEXT: v_readlane_b32 s40, v2, 36
-; GCN-NEXT: v_readlane_b32 s41, v2, 37
-; GCN-NEXT: v_readlane_b32 s42, v2, 38
-; GCN-NEXT: v_readlane_b32 s43, v2, 39
-; GCN-NEXT: v_readlane_b32 s44, v2, 40
-; GCN-NEXT: v_readlane_b32 s45, v2, 41
-; GCN-NEXT: v_readlane_b32 s46, v2, 42
-; GCN-NEXT: v_readlane_b32 s47, v2, 43
-; GCN-NEXT: v_readlane_b32 s48, v2, 44
-; GCN-NEXT: v_readlane_b32 s49, v2, 45
-; GCN-NEXT: v_readlane_b32 s50, v2, 46
-; GCN-NEXT: v_readlane_b32 s51, v2, 47
-; GCN-NEXT: v_readlane_b32 s0, v2, 16
-; GCN-NEXT: v_readlane_b32 s1, v2, 17
-; GCN-NEXT: v_readlane_b32 s2, v2, 18
-; GCN-NEXT: v_readlane_b32 s3, v2, 19
-; GCN-NEXT: v_readlane_b32 s4, v2, 20
-; GCN-NEXT: v_readlane_b32 s5, v2, 21
-; GCN-NEXT: v_readlane_b32 s6, v2, 22
-; GCN-NEXT: v_readlane_b32 s7, v2, 23
-; GCN-NEXT: v_readlane_b32 s8, v2, 24
-; GCN-NEXT: v_readlane_b32 s9, v2, 25
-; GCN-NEXT: v_readlane_b32 s10, v2, 26
-; GCN-NEXT: v_readlane_b32 s11, v2, 27
-; GCN-NEXT: v_readlane_b32 s12, v2, 28
-; GCN-NEXT: v_readlane_b32 s13, v2, 29
-; GCN-NEXT: v_readlane_b32 s14, v2, 30
-; GCN-NEXT: v_readlane_b32 s15, v2, 31
-; GCN-NEXT: v_readlane_b32 s16, v2, 0
-; GCN-NEXT: v_readlane_b32 s17, v2, 1
-; GCN-NEXT: v_readlane_b32 s18, v2, 2
-; GCN-NEXT: v_readlane_b32 s19, v2, 3
-; GCN-NEXT: v_readlane_b32 s20, v2, 4
-; GCN-NEXT: v_readlane_b32 s21, v2, 5
-; GCN-NEXT: v_readlane_b32 s22, v2, 6
-; GCN-NEXT: v_readlane_b32 s23, v2, 7
-; GCN-NEXT: v_readlane_b32 s24, v2, 8
-; GCN-NEXT: v_readlane_b32 s25, v2, 9
-; GCN-NEXT: v_readlane_b32 s26, v2, 10
-; GCN-NEXT: v_readlane_b32 s27, v2, 11
-; GCN-NEXT: v_readlane_b32 s28, v2, 12
-; GCN-NEXT: v_readlane_b32 s29, v2, 13
-; GCN-NEXT: v_readlane_b32 s30, v2, 14
-; GCN-NEXT: v_readlane_b32 s31, v2, 15
+; GCN-NEXT: v_readlane_b32 s36, v31, 32
+; GCN-NEXT: v_readlane_b32 s37, v31, 33
+; GCN-NEXT: v_readlane_b32 s38, v31, 34
+; GCN-NEXT: v_readlane_b32 s39, v31, 35
+; GCN-NEXT: v_readlane_b32 s40, v31, 36
+; GCN-NEXT: v_readlane_b32 s41, v31, 37
+; GCN-NEXT: v_readlane_b32 s42, v31, 38
+; GCN-NEXT: v_readlane_b32 s43, v31, 39
+; GCN-NEXT: v_readlane_b32 s44, v31, 40
+; GCN-NEXT: v_readlane_b32 s45, v31, 41
+; GCN-NEXT: v_readlane_b32 s46, v31, 42
+; GCN-NEXT: v_readlane_b32 s47, v31, 43
+; GCN-NEXT: v_readlane_b32 s48, v31, 44
+; GCN-NEXT: v_readlane_b32 s49, v31, 45
+; GCN-NEXT: v_readlane_b32 s50, v31, 46
+; GCN-NEXT: v_readlane_b32 s51, v31, 47
+; GCN-NEXT: v_readlane_b32 s0, v31, 16
+; GCN-NEXT: v_readlane_b32 s1, v31, 17
+; GCN-NEXT: v_readlane_b32 s2, v31, 18
+; GCN-NEXT: v_readlane_b32 s3, v31, 19
+; GCN-NEXT: v_readlane_b32 s4, v31, 20
+; GCN-NEXT: v_readlane_b32 s5, v31, 21
+; GCN-NEXT: v_readlane_b32 s6, v31, 22
+; GCN-NEXT: v_readlane_b32 s7, v31, 23
+; GCN-NEXT: v_readlane_b32 s8, v31, 24
+; GCN-NEXT: v_readlane_b32 s9, v31, 25
+; GCN-NEXT: v_readlane_b32 s10, v31, 26
+; GCN-NEXT: v_readlane_b32 s11, v31, 27
+; GCN-NEXT: v_readlane_b32 s12, v31, 28
+; GCN-NEXT: v_readlane_b32 s13, v31, 29
+; GCN-NEXT: v_readlane_b32 s14, v31, 30
+; GCN-NEXT: v_readlane_b32 s15, v31, 31
+; GCN-NEXT: v_readlane_b32 s16, v31, 0
+; GCN-NEXT: v_readlane_b32 s17, v31, 1
+; GCN-NEXT: v_readlane_b32 s18, v31, 2
+; GCN-NEXT: v_readlane_b32 s19, v31, 3
+; GCN-NEXT: v_readlane_b32 s20, v31, 4
+; GCN-NEXT: v_readlane_b32 s21, v31, 5
+; GCN-NEXT: v_readlane_b32 s22, v31, 6
+; GCN-NEXT: v_readlane_b32 s23, v31, 7
+; GCN-NEXT: v_readlane_b32 s24, v31, 8
+; GCN-NEXT: v_readlane_b32 s25, v31, 9
+; GCN-NEXT: v_readlane_b32 s26, v31, 10
+; GCN-NEXT: v_readlane_b32 s27, v31, 11
+; GCN-NEXT: v_readlane_b32 s28, v31, 12
+; GCN-NEXT: v_readlane_b32 s29, v31, 13
+; GCN-NEXT: v_readlane_b32 s30, v31, 14
+; GCN-NEXT: v_readlane_b32 s31, v31, 15
; GCN-NEXT: ;;#ASMSTART
; GCN-NEXT: ; def v0
; GCN-NEXT: ;;#ASMEND
@@ -1168,24 +1067,32 @@
; GCN-NEXT: ;;#ASMSTART
; GCN-NEXT: ; use s[0:15]
; GCN-NEXT: ;;#ASMEND
-; GCN-NEXT: v_readlane_b32 s4, v2, 48
-; GCN-NEXT: v_readlane_b32 s5, v2, 49
-; GCN-NEXT: v_readlane_b32 s6, v2, 50
-; GCN-NEXT: v_readlane_b32 s7, v2, 51
-; GCN-NEXT: v_readlane_b32 s8, v2, 52
-; GCN-NEXT: v_readlane_b32 s9, v2, 53
-; GCN-NEXT: v_readlane_b32 s10, v2, 54
-; GCN-NEXT: v_readlane_b32 s11, v2, 55
-; GCN-NEXT: v_readlane_b32 s12, v2, 56
-; GCN-NEXT: v_readlane_b32 s13, v2, 57
-; GCN-NEXT: v_readlane_b32 s14, v2, 58
-; GCN-NEXT: v_readlane_b32 s15, v2, 59
-; GCN-NEXT: v_readlane_b32 s16, v2, 60
-; GCN-NEXT: v_readlane_b32 s17, v2, 61
-; GCN-NEXT: v_readlane_b32 s18, v2, 62
-; GCN-NEXT: v_readlane_b32 s19, v2, 63
+; GCN-NEXT: v_readlane_b32 s4, v31, 48
+; GCN-NEXT: v_readlane_b32 s5, v31, 49
+; GCN-NEXT: v_readlane_b32 s6, v31, 50
+; GCN-NEXT: v_readlane_b32 s7, v31, 51
+; GCN-NEXT: v_readlane_b32 s8, v31, 52
+; GCN-NEXT: v_readlane_b32 s9, v31, 53
+; GCN-NEXT: v_readlane_b32 s10, v31, 54
+; GCN-NEXT: v_readlane_b32 s11, v31, 55
+; GCN-NEXT: v_readlane_b32 s12, v31, 56
+; GCN-NEXT: v_readlane_b32 s13, v31, 57
+; GCN-NEXT: v_readlane_b32 s14, v31, 58
+; GCN-NEXT: v_readlane_b32 s15, v31, 59
+; GCN-NEXT: v_readlane_b32 s16, v31, 60
+; GCN-NEXT: v_readlane_b32 s17, v31, 61
+; GCN-NEXT: v_readlane_b32 s18, v31, 62
+; GCN-NEXT: v_readlane_b32 s19, v31, 63
+; GCN-NEXT: s_mov_b64 s[2:3], exec
+; GCN-NEXT: s_mov_b64 exec, 3
+; GCN-NEXT: buffer_store_dword v1, off, s[52:55], 0
+; GCN-NEXT: buffer_load_dword v1, off, s[52:55], 0 offset:4 ; 4-byte Folded Reload
+; GCN-NEXT: s_waitcnt vmcnt(0)
; GCN-NEXT: v_readlane_b32 s0, v1, 0
; GCN-NEXT: v_readlane_b32 s1, v1, 1
+; GCN-NEXT: buffer_load_dword v1, off, s[52:55], 0
+; GCN-NEXT: s_waitcnt vmcnt(0)
+; GCN-NEXT: s_mov_b64 exec, s[2:3]
; GCN-NEXT: ;;#ASMSTART
; GCN-NEXT: ; use s[36:51]
; GCN-NEXT: ;;#ASMEND
@@ -1199,14 +1106,6 @@
; GCN-NEXT: ; use v0
; GCN-NEXT: ;;#ASMEND
; GCN-NEXT: .LBB3_2: ; %ret
-; GCN-NEXT: s_or_saveexec_b64 s[34:35], -1
-; GCN-NEXT: buffer_load_dword v0, off, s[52:55], 0 offset:4 ; 4-byte Folded Reload
-; GCN-NEXT: s_mov_b64 exec, s[34:35]
-; GCN-NEXT: s_or_saveexec_b64 s[34:35], -1
-; GCN-NEXT: buffer_load_dword v1, off, s[52:55], 0 offset:8 ; 4-byte Folded Reload
-; GCN-NEXT: s_mov_b64 exec, s[34:35]
-; GCN-NEXT: ; kill: killed $vgpr1
-; GCN-NEXT: ; kill: killed $vgpr0
; GCN-NEXT: s_endpgm
call void asm sideeffect "", "~{v[0:7]}" () #0
call void asm sideeffect "", "~{v[8:15]}" () #0
diff --git a/llvm/test/CodeGen/AMDGPU/preserve-wwm-copy-dst-reg.ll b/llvm/test/CodeGen/AMDGPU/preserve-wwm-copy-dst-reg.ll
deleted file mode 100644
--- a/llvm/test/CodeGen/AMDGPU/preserve-wwm-copy-dst-reg.ll
+++ /dev/null
@@ -1,816 +0,0 @@
-; NOTE: Assertions have been autogenerated by utils/update_llc_test_checks.py
-; RUN: llc -mtriple=amdgcn-amd-amdhsa -mcpu=gfx906 -verify-machineinstrs < %s | FileCheck -check-prefix=GFX906 %s
-; RUN: llc -mtriple=amdgcn-amd-amdhsa -mcpu=gfx908 -verify-machineinstrs < %s | FileCheck -check-prefix=GFX908 %s
-
-; Due to high register pressure, regalloc would split the liverange of wwm VGPR register used for SGPR spills
-; and introduce a copy. The copy should be of whole-wave with exec mask manipulation around it.
-; FIXME: The destination register involved in the whole-wave copy should be considered for preserving all the lanes
-; with a spill/restore at function prolog/epilog. The copy might otherwise clobber its inactive lanes unwantedly.
-define void @preserve_wwm_copy_dstreg(ptr %parg0, ptr %parg1, ptr %parg2) #0 {
-; GFX906-LABEL: preserve_wwm_copy_dstreg:
-; GFX906: ; %bb.0:
-; GFX906-NEXT: s_waitcnt vmcnt(0) expcnt(0) lgkmcnt(0)
-; GFX906-NEXT: s_mov_b32 s16, s33
-; GFX906-NEXT: s_mov_b32 s33, s32
-; GFX906-NEXT: s_xor_saveexec_b64 s[18:19], -1
-; GFX906-NEXT: buffer_store_dword v33, off, s[0:3], s33 offset:144 ; 4-byte Folded Spill
-; GFX906-NEXT: buffer_store_dword v2, off, s[0:3], s33 offset:152 ; 4-byte Folded Spill
-; GFX906-NEXT: s_mov_b64 exec, -1
-; GFX906-NEXT: buffer_store_dword v41, off, s[0:3], s33 ; 4-byte Folded Spill
-; GFX906-NEXT: buffer_store_dword v40, off, s[0:3], s33 offset:148 ; 4-byte Folded Spill
-; GFX906-NEXT: s_mov_b64 exec, s[18:19]
-; GFX906-NEXT: s_mov_b32 s21, s15
-; GFX906-NEXT: ; implicit-def: $vgpr2
-; GFX906-NEXT: s_mov_b32 s22, s14
-; GFX906-NEXT: v_writelane_b32 v2, s21, 0
-; GFX906-NEXT: v_writelane_b32 v2, s22, 1
-; GFX906-NEXT: s_mov_b32 s23, s13
-; GFX906-NEXT: v_writelane_b32 v2, s23, 2
-; GFX906-NEXT: s_mov_b32 s24, s12
-; GFX906-NEXT: v_writelane_b32 v2, s24, 3
-; GFX906-NEXT: s_mov_b64 s[26:27], s[10:11]
-; GFX906-NEXT: v_writelane_b32 v2, s26, 4
-; GFX906-NEXT: v_writelane_b32 v2, s27, 5
-; GFX906-NEXT: v_writelane_b32 v2, s8, 6
-; GFX906-NEXT: v_writelane_b32 v2, s9, 7
-; GFX906-NEXT: v_writelane_b32 v2, s6, 8
-; GFX906-NEXT: v_writelane_b32 v41, s16, 2
-; GFX906-NEXT: v_writelane_b32 v2, s7, 9
-; GFX906-NEXT: v_writelane_b32 v41, s30, 0
-; GFX906-NEXT: v_writelane_b32 v2, s4, 10
-; GFX906-NEXT: s_addk_i32 s32, 0x2800
-; GFX906-NEXT: v_writelane_b32 v41, s31, 1
-; GFX906-NEXT: v_mov_b32_e32 v32, v31
-; GFX906-NEXT: v_writelane_b32 v2, s5, 11
-; GFX906-NEXT: s_or_saveexec_b64 s[34:35], -1
-; GFX906-NEXT: v_mov_b32_e32 v33, v2
-; GFX906-NEXT: s_mov_b64 exec, s[34:35]
-; GFX906-NEXT: buffer_store_dword v0, off, s[0:3], s33 offset:8 ; 4-byte Folded Spill
-; GFX906-NEXT: s_waitcnt vmcnt(0)
-; GFX906-NEXT: buffer_store_dword v1, off, s[0:3], s33 offset:12 ; 4-byte Folded Spill
-; GFX906-NEXT: ;;#ASMSTART
-; GFX906-NEXT: ; def v[0:31]
-; GFX906-NEXT: ;;#ASMEND
-; GFX906-NEXT: buffer_store_dword v0, off, s[0:3], s33 offset:16 ; 4-byte Folded Spill
-; GFX906-NEXT: s_waitcnt vmcnt(0)
-; GFX906-NEXT: buffer_store_dword v1, off, s[0:3], s33 offset:20 ; 4-byte Folded Spill
-; GFX906-NEXT: buffer_store_dword v2, off, s[0:3], s33 offset:24 ; 4-byte Folded Spill
-; GFX906-NEXT: buffer_store_dword v3, off, s[0:3], s33 offset:28 ; 4-byte Folded Spill
-; GFX906-NEXT: buffer_store_dword v4, off, s[0:3], s33 offset:32 ; 4-byte Folded Spill
-; GFX906-NEXT: buffer_store_dword v5, off, s[0:3], s33 offset:36 ; 4-byte Folded Spill
-; GFX906-NEXT: buffer_store_dword v6, off, s[0:3], s33 offset:40 ; 4-byte Folded Spill
-; GFX906-NEXT: buffer_store_dword v7, off, s[0:3], s33 offset:44 ; 4-byte Folded Spill
-; GFX906-NEXT: buffer_store_dword v8, off, s[0:3], s33 offset:48 ; 4-byte Folded Spill
-; GFX906-NEXT: buffer_store_dword v9, off, s[0:3], s33 offset:52 ; 4-byte Folded Spill
-; GFX906-NEXT: buffer_store_dword v10, off, s[0:3], s33 offset:56 ; 4-byte Folded Spill
-; GFX906-NEXT: buffer_store_dword v11, off, s[0:3], s33 offset:60 ; 4-byte Folded Spill
-; GFX906-NEXT: buffer_store_dword v12, off, s[0:3], s33 offset:64 ; 4-byte Folded Spill
-; GFX906-NEXT: buffer_store_dword v13, off, s[0:3], s33 offset:68 ; 4-byte Folded Spill
-; GFX906-NEXT: buffer_store_dword v14, off, s[0:3], s33 offset:72 ; 4-byte Folded Spill
-; GFX906-NEXT: buffer_store_dword v15, off, s[0:3], s33 offset:76 ; 4-byte Folded Spill
-; GFX906-NEXT: buffer_store_dword v16, off, s[0:3], s33 offset:80 ; 4-byte Folded Spill
-; GFX906-NEXT: buffer_store_dword v17, off, s[0:3], s33 offset:84 ; 4-byte Folded Spill
-; GFX906-NEXT: buffer_store_dword v18, off, s[0:3], s33 offset:88 ; 4-byte Folded Spill
-; GFX906-NEXT: buffer_store_dword v19, off, s[0:3], s33 offset:92 ; 4-byte Folded Spill
-; GFX906-NEXT: buffer_store_dword v20, off, s[0:3], s33 offset:96 ; 4-byte Folded Spill
-; GFX906-NEXT: buffer_store_dword v21, off, s[0:3], s33 offset:100 ; 4-byte Folded Spill
-; GFX906-NEXT: buffer_store_dword v22, off, s[0:3], s33 offset:104 ; 4-byte Folded Spill
-; GFX906-NEXT: buffer_store_dword v23, off, s[0:3], s33 offset:108 ; 4-byte Folded Spill
-; GFX906-NEXT: buffer_store_dword v24, off, s[0:3], s33 offset:112 ; 4-byte Folded Spill
-; GFX906-NEXT: buffer_store_dword v25, off, s[0:3], s33 offset:116 ; 4-byte Folded Spill
-; GFX906-NEXT: buffer_store_dword v26, off, s[0:3], s33 offset:120 ; 4-byte Folded Spill
-; GFX906-NEXT: buffer_store_dword v27, off, s[0:3], s33 offset:124 ; 4-byte Folded Spill
-; GFX906-NEXT: buffer_store_dword v28, off, s[0:3], s33 offset:128 ; 4-byte Folded Spill
-; GFX906-NEXT: buffer_store_dword v29, off, s[0:3], s33 offset:132 ; 4-byte Folded Spill
-; GFX906-NEXT: buffer_store_dword v30, off, s[0:3], s33 offset:136 ; 4-byte Folded Spill
-; GFX906-NEXT: buffer_store_dword v31, off, s[0:3], s33 offset:140 ; 4-byte Folded Spill
-; GFX906-NEXT: ;;#ASMSTART
-; GFX906-NEXT: ; def v40
-; GFX906-NEXT: ;;#ASMEND
-; GFX906-NEXT: ;;#ASMSTART
-; GFX906-NEXT: ; def s11
-; GFX906-NEXT: ;;#ASMEND
-; GFX906-NEXT: s_or_saveexec_b64 s[34:35], -1
-; GFX906-NEXT: v_mov_b32_e32 v40, v33
-; GFX906-NEXT: s_mov_b64 exec, s[34:35]
-; GFX906-NEXT: v_writelane_b32 v40, s11, 12
-; GFX906-NEXT: ;;#ASMSTART
-; GFX906-NEXT: ; def s12
-; GFX906-NEXT: ;;#ASMEND
-; GFX906-NEXT: v_writelane_b32 v40, s12, 13
-; GFX906-NEXT: ;;#ASMSTART
-; GFX906-NEXT: ; def s13
-; GFX906-NEXT: ;;#ASMEND
-; GFX906-NEXT: v_writelane_b32 v40, s13, 14
-; GFX906-NEXT: ;;#ASMSTART
-; GFX906-NEXT: ; def s14
-; GFX906-NEXT: ;;#ASMEND
-; GFX906-NEXT: v_writelane_b32 v40, s14, 15
-; GFX906-NEXT: ;;#ASMSTART
-; GFX906-NEXT: ; def s15
-; GFX906-NEXT: ;;#ASMEND
-; GFX906-NEXT: v_writelane_b32 v40, s15, 16
-; GFX906-NEXT: s_getpc_b64 s[10:11]
-; GFX906-NEXT: s_add_u32 s10, s10, foo@gotpcrel32@lo+4
-; GFX906-NEXT: s_addc_u32 s11, s11, foo@gotpcrel32@hi+12
-; GFX906-NEXT: ;;#ASMSTART
-; GFX906-NEXT: ; def s16
-; GFX906-NEXT: ;;#ASMEND
-; GFX906-NEXT: v_writelane_b32 v40, s16, 17
-; GFX906-NEXT: s_load_dwordx2 s[10:11], s[10:11], 0x0
-; GFX906-NEXT: ;;#ASMSTART
-; GFX906-NEXT: ; def s17
-; GFX906-NEXT: ;;#ASMEND
-; GFX906-NEXT: v_writelane_b32 v40, s17, 18
-; GFX906-NEXT: ;;#ASMSTART
-; GFX906-NEXT: ; def s18
-; GFX906-NEXT: ;;#ASMEND
-; GFX906-NEXT: v_writelane_b32 v40, s18, 19
-; GFX906-NEXT: ;;#ASMSTART
-; GFX906-NEXT: ; def s19
-; GFX906-NEXT: ;;#ASMEND
-; GFX906-NEXT: v_writelane_b32 v40, s19, 20
-; GFX906-NEXT: ;;#ASMSTART
-; GFX906-NEXT: ; def s20
-; GFX906-NEXT: ;;#ASMEND
-; GFX906-NEXT: v_writelane_b32 v40, s20, 21
-; GFX906-NEXT: s_waitcnt lgkmcnt(0)
-; GFX906-NEXT: v_writelane_b32 v40, s10, 22
-; GFX906-NEXT: v_writelane_b32 v40, s11, 23
-; GFX906-NEXT: s_or_saveexec_b64 s[34:35], -1
-; GFX906-NEXT: s_mov_b64 exec, s[34:35]
-; GFX906-NEXT: v_readlane_b32 s16, v40, 22
-; GFX906-NEXT: s_mov_b32 s12, s24
-; GFX906-NEXT: s_mov_b32 s13, s23
-; GFX906-NEXT: s_mov_b32 s14, s22
-; GFX906-NEXT: v_mov_b32_e32 v31, v32
-; GFX906-NEXT: s_mov_b32 s15, s21
-; GFX906-NEXT: s_mov_b64 s[10:11], s[26:27]
-; GFX906-NEXT: v_readlane_b32 s17, v40, 23
-; GFX906-NEXT: buffer_store_dword v32, off, s[0:3], s33 offset:4 ; 4-byte Folded Spill
-; GFX906-NEXT: s_swappc_b64 s[30:31], s[16:17]
-; GFX906-NEXT: s_or_saveexec_b64 s[34:35], -1
-; GFX906-NEXT: s_mov_b64 exec, s[34:35]
-; GFX906-NEXT: v_readlane_b32 s11, v40, 12
-; GFX906-NEXT: ;;#ASMSTART
-; GFX906-NEXT: ; use s11
-; GFX906-NEXT: ;;#ASMEND
-; GFX906-NEXT: v_readlane_b32 s12, v40, 13
-; GFX906-NEXT: ;;#ASMSTART
-; GFX906-NEXT: ; use s12
-; GFX906-NEXT: ;;#ASMEND
-; GFX906-NEXT: v_readlane_b32 s13, v40, 14
-; GFX906-NEXT: ;;#ASMSTART
-; GFX906-NEXT: ; use s13
-; GFX906-NEXT: ;;#ASMEND
-; GFX906-NEXT: v_readlane_b32 s14, v40, 15
-; GFX906-NEXT: ;;#ASMSTART
-; GFX906-NEXT: ; use s14
-; GFX906-NEXT: ;;#ASMEND
-; GFX906-NEXT: v_readlane_b32 s15, v40, 16
-; GFX906-NEXT: ;;#ASMSTART
-; GFX906-NEXT: ; use s15
-; GFX906-NEXT: ;;#ASMEND
-; GFX906-NEXT: v_readlane_b32 s16, v40, 17
-; GFX906-NEXT: ;;#ASMSTART
-; GFX906-NEXT: ; use s16
-; GFX906-NEXT: ;;#ASMEND
-; GFX906-NEXT: v_readlane_b32 s17, v40, 18
-; GFX906-NEXT: ;;#ASMSTART
-; GFX906-NEXT: ; use s17
-; GFX906-NEXT: ;;#ASMEND
-; GFX906-NEXT: v_readlane_b32 s18, v40, 19
-; GFX906-NEXT: ;;#ASMSTART
-; GFX906-NEXT: ; use s18
-; GFX906-NEXT: ;;#ASMEND
-; GFX906-NEXT: v_readlane_b32 s19, v40, 20
-; GFX906-NEXT: ;;#ASMSTART
-; GFX906-NEXT: ; use s19
-; GFX906-NEXT: ;;#ASMEND
-; GFX906-NEXT: v_readlane_b32 s20, v40, 21
-; GFX906-NEXT: ;;#ASMSTART
-; GFX906-NEXT: ; use s20
-; GFX906-NEXT: ;;#ASMEND
-; GFX906-NEXT: ;;#ASMSTART
-; GFX906-NEXT: ; def s21
-; GFX906-NEXT: ;;#ASMEND
-; GFX906-NEXT: ;;#ASMSTART
-; GFX906-NEXT: ; def s22
-; GFX906-NEXT: ;;#ASMEND
-; GFX906-NEXT: ;;#ASMSTART
-; GFX906-NEXT: ; def s23
-; GFX906-NEXT: ;;#ASMEND
-; GFX906-NEXT: ;;#ASMSTART
-; GFX906-NEXT: ; def s24
-; GFX906-NEXT: ;;#ASMEND
-; GFX906-NEXT: ;;#ASMSTART
-; GFX906-NEXT: ; def s25
-; GFX906-NEXT: ;;#ASMEND
-; GFX906-NEXT: ;;#ASMSTART
-; GFX906-NEXT: ; def s26
-; GFX906-NEXT: ;;#ASMEND
-; GFX906-NEXT: ;;#ASMSTART
-; GFX906-NEXT: ; def s27
-; GFX906-NEXT: ;;#ASMEND
-; GFX906-NEXT: ;;#ASMSTART
-; GFX906-NEXT: ; def s28
-; GFX906-NEXT: ;;#ASMEND
-; GFX906-NEXT: ;;#ASMSTART
-; GFX906-NEXT: ; def s29
-; GFX906-NEXT: ;;#ASMEND
-; GFX906-NEXT: buffer_load_dword v31, off, s[0:3], s33 offset:4 ; 4-byte Folded Reload
-; GFX906-NEXT: v_writelane_b32 v40, s21, 24
-; GFX906-NEXT: v_writelane_b32 v40, s22, 25
-; GFX906-NEXT: v_writelane_b32 v40, s23, 26
-; GFX906-NEXT: v_writelane_b32 v40, s24, 27
-; GFX906-NEXT: v_writelane_b32 v40, s25, 28
-; GFX906-NEXT: v_writelane_b32 v40, s26, 29
-; GFX906-NEXT: v_writelane_b32 v40, s27, 30
-; GFX906-NEXT: v_writelane_b32 v40, s28, 31
-; GFX906-NEXT: v_writelane_b32 v40, s29, 32
-; GFX906-NEXT: v_readlane_b32 s4, v40, 10
-; GFX906-NEXT: v_readlane_b32 s6, v40, 8
-; GFX906-NEXT: v_readlane_b32 s8, v40, 6
-; GFX906-NEXT: v_readlane_b32 s10, v40, 4
-; GFX906-NEXT: v_readlane_b32 s16, v40, 22
-; GFX906-NEXT: v_readlane_b32 s12, v40, 3
-; GFX906-NEXT: v_readlane_b32 s13, v40, 2
-; GFX906-NEXT: v_readlane_b32 s14, v40, 1
-; GFX906-NEXT: v_readlane_b32 s15, v40, 0
-; GFX906-NEXT: v_readlane_b32 s5, v40, 11
-; GFX906-NEXT: v_readlane_b32 s7, v40, 9
-; GFX906-NEXT: v_readlane_b32 s9, v40, 7
-; GFX906-NEXT: v_readlane_b32 s11, v40, 5
-; GFX906-NEXT: v_readlane_b32 s17, v40, 23
-; GFX906-NEXT: s_or_saveexec_b64 s[34:35], -1
-; GFX906-NEXT: s_mov_b64 exec, s[34:35]
-; GFX906-NEXT: s_swappc_b64 s[30:31], s[16:17]
-; GFX906-NEXT: s_or_saveexec_b64 s[34:35], -1
-; GFX906-NEXT: s_mov_b64 exec, s[34:35]
-; GFX906-NEXT: v_readlane_b32 s21, v40, 24
-; GFX906-NEXT: ;;#ASMSTART
-; GFX906-NEXT: ; use s21
-; GFX906-NEXT: ;;#ASMEND
-; GFX906-NEXT: v_readlane_b32 s22, v40, 25
-; GFX906-NEXT: ;;#ASMSTART
-; GFX906-NEXT: ; use s22
-; GFX906-NEXT: ;;#ASMEND
-; GFX906-NEXT: v_readlane_b32 s23, v40, 26
-; GFX906-NEXT: ;;#ASMSTART
-; GFX906-NEXT: ; use s23
-; GFX906-NEXT: ;;#ASMEND
-; GFX906-NEXT: v_readlane_b32 s24, v40, 27
-; GFX906-NEXT: ;;#ASMSTART
-; GFX906-NEXT: ; use s24
-; GFX906-NEXT: ;;#ASMEND
-; GFX906-NEXT: v_readlane_b32 s25, v40, 28
-; GFX906-NEXT: ;;#ASMSTART
-; GFX906-NEXT: ; use s25
-; GFX906-NEXT: ;;#ASMEND
-; GFX906-NEXT: v_readlane_b32 s26, v40, 29
-; GFX906-NEXT: ;;#ASMSTART
-; GFX906-NEXT: ; use s26
-; GFX906-NEXT: ;;#ASMEND
-; GFX906-NEXT: v_readlane_b32 s27, v40, 30
-; GFX906-NEXT: ;;#ASMSTART
-; GFX906-NEXT: ; use s27
-; GFX906-NEXT: ;;#ASMEND
-; GFX906-NEXT: v_readlane_b32 s28, v40, 31
-; GFX906-NEXT: ;;#ASMSTART
-; GFX906-NEXT: ; use s28
-; GFX906-NEXT: ;;#ASMEND
-; GFX906-NEXT: v_readlane_b32 s29, v40, 32
-; GFX906-NEXT: ;;#ASMSTART
-; GFX906-NEXT: ; use s29
-; GFX906-NEXT: ;;#ASMEND
-; GFX906-NEXT: buffer_load_dword v31, off, s[0:3], s33 offset:4 ; 4-byte Folded Reload
-; GFX906-NEXT: v_readlane_b32 s4, v40, 10
-; GFX906-NEXT: v_readlane_b32 s6, v40, 8
-; GFX906-NEXT: v_readlane_b32 s8, v40, 6
-; GFX906-NEXT: v_readlane_b32 s10, v40, 4
-; GFX906-NEXT: v_readlane_b32 s16, v40, 22
-; GFX906-NEXT: v_readlane_b32 s5, v40, 11
-; GFX906-NEXT: v_readlane_b32 s7, v40, 9
-; GFX906-NEXT: v_readlane_b32 s9, v40, 7
-; GFX906-NEXT: v_readlane_b32 s11, v40, 5
-; GFX906-NEXT: v_readlane_b32 s12, v40, 3
-; GFX906-NEXT: v_readlane_b32 s13, v40, 2
-; GFX906-NEXT: v_readlane_b32 s14, v40, 1
-; GFX906-NEXT: v_readlane_b32 s15, v40, 0
-; GFX906-NEXT: v_readlane_b32 s17, v40, 23
-; GFX906-NEXT: s_or_saveexec_b64 s[34:35], -1
-; GFX906-NEXT: s_mov_b64 exec, s[34:35]
-; GFX906-NEXT: s_swappc_b64 s[30:31], s[16:17]
-; GFX906-NEXT: buffer_load_dword v0, off, s[0:3], s33 offset:8 ; 4-byte Folded Reload
-; GFX906-NEXT: buffer_load_dword v1, off, s[0:3], s33 offset:12 ; 4-byte Folded Reload
-; GFX906-NEXT: buffer_load_dword v2, off, s[0:3], s33 offset:16 ; 4-byte Folded Reload
-; GFX906-NEXT: buffer_load_dword v3, off, s[0:3], s33 offset:20 ; 4-byte Folded Reload
-; GFX906-NEXT: buffer_load_dword v4, off, s[0:3], s33 offset:24 ; 4-byte Folded Reload
-; GFX906-NEXT: buffer_load_dword v5, off, s[0:3], s33 offset:28 ; 4-byte Folded Reload
-; GFX906-NEXT: buffer_load_dword v6, off, s[0:3], s33 offset:32 ; 4-byte Folded Reload
-; GFX906-NEXT: buffer_load_dword v7, off, s[0:3], s33 offset:36 ; 4-byte Folded Reload
-; GFX906-NEXT: buffer_load_dword v8, off, s[0:3], s33 offset:40 ; 4-byte Folded Reload
-; GFX906-NEXT: buffer_load_dword v9, off, s[0:3], s33 offset:44 ; 4-byte Folded Reload
-; GFX906-NEXT: buffer_load_dword v10, off, s[0:3], s33 offset:48 ; 4-byte Folded Reload
-; GFX906-NEXT: buffer_load_dword v11, off, s[0:3], s33 offset:52 ; 4-byte Folded Reload
-; GFX906-NEXT: buffer_load_dword v12, off, s[0:3], s33 offset:56 ; 4-byte Folded Reload
-; GFX906-NEXT: buffer_load_dword v13, off, s[0:3], s33 offset:60 ; 4-byte Folded Reload
-; GFX906-NEXT: buffer_load_dword v14, off, s[0:3], s33 offset:64 ; 4-byte Folded Reload
-; GFX906-NEXT: buffer_load_dword v15, off, s[0:3], s33 offset:68 ; 4-byte Folded Reload
-; GFX906-NEXT: buffer_load_dword v16, off, s[0:3], s33 offset:72 ; 4-byte Folded Reload
-; GFX906-NEXT: buffer_load_dword v17, off, s[0:3], s33 offset:76 ; 4-byte Folded Reload
-; GFX906-NEXT: buffer_load_dword v18, off, s[0:3], s33 offset:80 ; 4-byte Folded Reload
-; GFX906-NEXT: buffer_load_dword v19, off, s[0:3], s33 offset:84 ; 4-byte Folded Reload
-; GFX906-NEXT: buffer_load_dword v20, off, s[0:3], s33 offset:88 ; 4-byte Folded Reload
-; GFX906-NEXT: buffer_load_dword v21, off, s[0:3], s33 offset:92 ; 4-byte Folded Reload
-; GFX906-NEXT: buffer_load_dword v22, off, s[0:3], s33 offset:96 ; 4-byte Folded Reload
-; GFX906-NEXT: buffer_load_dword v23, off, s[0:3], s33 offset:100 ; 4-byte Folded Reload
-; GFX906-NEXT: buffer_load_dword v24, off, s[0:3], s33 offset:104 ; 4-byte Folded Reload
-; GFX906-NEXT: buffer_load_dword v25, off, s[0:3], s33 offset:108 ; 4-byte Folded Reload
-; GFX906-NEXT: buffer_load_dword v26, off, s[0:3], s33 offset:112 ; 4-byte Folded Reload
-; GFX906-NEXT: buffer_load_dword v27, off, s[0:3], s33 offset:116 ; 4-byte Folded Reload
-; GFX906-NEXT: buffer_load_dword v28, off, s[0:3], s33 offset:120 ; 4-byte Folded Reload
-; GFX906-NEXT: buffer_load_dword v29, off, s[0:3], s33 offset:124 ; 4-byte Folded Reload
-; GFX906-NEXT: buffer_load_dword v30, off, s[0:3], s33 offset:128 ; 4-byte Folded Reload
-; GFX906-NEXT: buffer_load_dword v31, off, s[0:3], s33 offset:132 ; 4-byte Folded Reload
-; GFX906-NEXT: buffer_load_dword v32, off, s[0:3], s33 offset:136 ; 4-byte Folded Reload
-; GFX906-NEXT: buffer_load_dword v33, off, s[0:3], s33 offset:140 ; 4-byte Folded Reload
-; GFX906-NEXT: v_readlane_b32 s31, v41, 1
-; GFX906-NEXT: v_readlane_b32 s30, v41, 0
-; GFX906-NEXT: ; kill: killed $vgpr40
-; GFX906-NEXT: v_readlane_b32 s4, v41, 2
-; GFX906-NEXT: s_waitcnt vmcnt(0)
-; GFX906-NEXT: flat_store_dwordx4 v[0:1], v[30:33] offset:112
-; GFX906-NEXT: s_waitcnt vmcnt(0)
-; GFX906-NEXT: flat_store_dwordx4 v[0:1], v[26:29] offset:96
-; GFX906-NEXT: s_waitcnt vmcnt(0)
-; GFX906-NEXT: flat_store_dwordx4 v[0:1], v[22:25] offset:80
-; GFX906-NEXT: s_waitcnt vmcnt(0)
-; GFX906-NEXT: flat_store_dwordx4 v[0:1], v[18:21] offset:64
-; GFX906-NEXT: s_waitcnt vmcnt(0)
-; GFX906-NEXT: flat_store_dwordx4 v[0:1], v[14:17] offset:48
-; GFX906-NEXT: s_waitcnt vmcnt(0)
-; GFX906-NEXT: flat_store_dwordx4 v[0:1], v[10:13] offset:32
-; GFX906-NEXT: s_waitcnt vmcnt(0)
-; GFX906-NEXT: flat_store_dwordx4 v[0:1], v[6:9] offset:16
-; GFX906-NEXT: s_waitcnt vmcnt(0)
-; GFX906-NEXT: flat_store_dwordx4 v[0:1], v[2:5]
-; GFX906-NEXT: s_waitcnt vmcnt(0)
-; GFX906-NEXT: s_xor_saveexec_b64 s[6:7], -1
-; GFX906-NEXT: buffer_load_dword v33, off, s[0:3], s33 offset:144 ; 4-byte Folded Reload
-; GFX906-NEXT: buffer_load_dword v2, off, s[0:3], s33 offset:152 ; 4-byte Folded Reload
-; GFX906-NEXT: s_mov_b64 exec, -1
-; GFX906-NEXT: buffer_load_dword v41, off, s[0:3], s33 ; 4-byte Folded Reload
-; GFX906-NEXT: buffer_load_dword v40, off, s[0:3], s33 offset:148 ; 4-byte Folded Reload
-; GFX906-NEXT: s_mov_b64 exec, s[6:7]
-; GFX906-NEXT: s_addk_i32 s32, 0xd800
-; GFX906-NEXT: s_mov_b32 s33, s4
-; GFX906-NEXT: s_waitcnt vmcnt(0) lgkmcnt(0)
-; GFX906-NEXT: s_setpc_b64 s[30:31]
-;
-; GFX908-LABEL: preserve_wwm_copy_dstreg:
-; GFX908: ; %bb.0:
-; GFX908-NEXT: s_waitcnt vmcnt(0) expcnt(0) lgkmcnt(0)
-; GFX908-NEXT: s_mov_b32 s16, s33
-; GFX908-NEXT: s_mov_b32 s33, s32
-; GFX908-NEXT: s_xor_saveexec_b64 s[18:19], -1
-; GFX908-NEXT: buffer_store_dword v33, off, s[0:3], s33 offset:148 ; 4-byte Folded Spill
-; GFX908-NEXT: buffer_store_dword v2, off, s[0:3], s33 offset:156 ; 4-byte Folded Spill
-; GFX908-NEXT: s_mov_b64 exec, -1
-; GFX908-NEXT: buffer_store_dword v40, off, s[0:3], s33 offset:152 ; 4-byte Folded Spill
-; GFX908-NEXT: s_mov_b64 exec, s[18:19]
-; GFX908-NEXT: v_mov_b32_e32 v3, s16
-; GFX908-NEXT: buffer_store_dword v3, off, s[0:3], s33 offset:160 ; 4-byte Folded Spill
-; GFX908-NEXT: s_addk_i32 s32, 0x2c00
-; GFX908-NEXT: s_mov_b64 s[16:17], exec
-; GFX908-NEXT: s_mov_b64 exec, 1
-; GFX908-NEXT: buffer_store_dword v2, off, s[0:3], s33 offset:164
-; GFX908-NEXT: v_writelane_b32 v2, s30, 0
-; GFX908-NEXT: buffer_store_dword v2, off, s[0:3], s33 ; 4-byte Folded Spill
-; GFX908-NEXT: buffer_load_dword v2, off, s[0:3], s33 offset:164
-; GFX908-NEXT: s_waitcnt vmcnt(0)
-; GFX908-NEXT: s_mov_b64 exec, s[16:17]
-; GFX908-NEXT: s_mov_b64 s[16:17], exec
-; GFX908-NEXT: s_mov_b64 exec, 1
-; GFX908-NEXT: buffer_store_dword v2, off, s[0:3], s33 offset:164
-; GFX908-NEXT: v_writelane_b32 v2, s31, 0
-; GFX908-NEXT: buffer_store_dword v2, off, s[0:3], s33 offset:4 ; 4-byte Folded Spill
-; GFX908-NEXT: buffer_load_dword v2, off, s[0:3], s33 offset:164
-; GFX908-NEXT: s_waitcnt vmcnt(0)
-; GFX908-NEXT: s_mov_b64 exec, s[16:17]
-; GFX908-NEXT: s_mov_b32 s21, s15
-; GFX908-NEXT: ; implicit-def: $vgpr2
-; GFX908-NEXT: s_mov_b32 s22, s14
-; GFX908-NEXT: v_writelane_b32 v2, s21, 0
-; GFX908-NEXT: v_writelane_b32 v2, s22, 1
-; GFX908-NEXT: s_mov_b32 s23, s13
-; GFX908-NEXT: v_writelane_b32 v2, s23, 2
-; GFX908-NEXT: s_mov_b32 s24, s12
-; GFX908-NEXT: v_writelane_b32 v2, s24, 3
-; GFX908-NEXT: s_mov_b64 s[26:27], s[10:11]
-; GFX908-NEXT: v_writelane_b32 v2, s26, 4
-; GFX908-NEXT: v_writelane_b32 v2, s27, 5
-; GFX908-NEXT: v_writelane_b32 v2, s8, 6
-; GFX908-NEXT: v_writelane_b32 v2, s9, 7
-; GFX908-NEXT: v_writelane_b32 v2, s6, 8
-; GFX908-NEXT: v_writelane_b32 v2, s7, 9
-; GFX908-NEXT: v_writelane_b32 v2, s4, 10
-; GFX908-NEXT: v_mov_b32_e32 v32, v31
-; GFX908-NEXT: v_writelane_b32 v2, s5, 11
-; GFX908-NEXT: s_or_saveexec_b64 s[34:35], -1
-; GFX908-NEXT: v_mov_b32_e32 v33, v2
-; GFX908-NEXT: s_mov_b64 exec, s[34:35]
-; GFX908-NEXT: buffer_store_dword v0, off, s[0:3], s33 offset:12 ; 4-byte Folded Spill
-; GFX908-NEXT: s_waitcnt vmcnt(0)
-; GFX908-NEXT: buffer_store_dword v1, off, s[0:3], s33 offset:16 ; 4-byte Folded Spill
-; GFX908-NEXT: ;;#ASMSTART
-; GFX908-NEXT: ; def v[0:31]
-; GFX908-NEXT: ;;#ASMEND
-; GFX908-NEXT: buffer_store_dword v0, off, s[0:3], s33 offset:20 ; 4-byte Folded Spill
-; GFX908-NEXT: s_waitcnt vmcnt(0)
-; GFX908-NEXT: buffer_store_dword v1, off, s[0:3], s33 offset:24 ; 4-byte Folded Spill
-; GFX908-NEXT: buffer_store_dword v2, off, s[0:3], s33 offset:28 ; 4-byte Folded Spill
-; GFX908-NEXT: buffer_store_dword v3, off, s[0:3], s33 offset:32 ; 4-byte Folded Spill
-; GFX908-NEXT: buffer_store_dword v4, off, s[0:3], s33 offset:36 ; 4-byte Folded Spill
-; GFX908-NEXT: buffer_store_dword v5, off, s[0:3], s33 offset:40 ; 4-byte Folded Spill
-; GFX908-NEXT: buffer_store_dword v6, off, s[0:3], s33 offset:44 ; 4-byte Folded Spill
-; GFX908-NEXT: buffer_store_dword v7, off, s[0:3], s33 offset:48 ; 4-byte Folded Spill
-; GFX908-NEXT: buffer_store_dword v8, off, s[0:3], s33 offset:52 ; 4-byte Folded Spill
-; GFX908-NEXT: buffer_store_dword v9, off, s[0:3], s33 offset:56 ; 4-byte Folded Spill
-; GFX908-NEXT: buffer_store_dword v10, off, s[0:3], s33 offset:60 ; 4-byte Folded Spill
-; GFX908-NEXT: buffer_store_dword v11, off, s[0:3], s33 offset:64 ; 4-byte Folded Spill
-; GFX908-NEXT: buffer_store_dword v12, off, s[0:3], s33 offset:68 ; 4-byte Folded Spill
-; GFX908-NEXT: buffer_store_dword v13, off, s[0:3], s33 offset:72 ; 4-byte Folded Spill
-; GFX908-NEXT: buffer_store_dword v14, off, s[0:3], s33 offset:76 ; 4-byte Folded Spill
-; GFX908-NEXT: buffer_store_dword v15, off, s[0:3], s33 offset:80 ; 4-byte Folded Spill
-; GFX908-NEXT: buffer_store_dword v16, off, s[0:3], s33 offset:84 ; 4-byte Folded Spill
-; GFX908-NEXT: buffer_store_dword v17, off, s[0:3], s33 offset:88 ; 4-byte Folded Spill
-; GFX908-NEXT: buffer_store_dword v18, off, s[0:3], s33 offset:92 ; 4-byte Folded Spill
-; GFX908-NEXT: buffer_store_dword v19, off, s[0:3], s33 offset:96 ; 4-byte Folded Spill
-; GFX908-NEXT: buffer_store_dword v20, off, s[0:3], s33 offset:100 ; 4-byte Folded Spill
-; GFX908-NEXT: buffer_store_dword v21, off, s[0:3], s33 offset:104 ; 4-byte Folded Spill
-; GFX908-NEXT: buffer_store_dword v22, off, s[0:3], s33 offset:108 ; 4-byte Folded Spill
-; GFX908-NEXT: buffer_store_dword v23, off, s[0:3], s33 offset:112 ; 4-byte Folded Spill
-; GFX908-NEXT: buffer_store_dword v24, off, s[0:3], s33 offset:116 ; 4-byte Folded Spill
-; GFX908-NEXT: buffer_store_dword v25, off, s[0:3], s33 offset:120 ; 4-byte Folded Spill
-; GFX908-NEXT: buffer_store_dword v26, off, s[0:3], s33 offset:124 ; 4-byte Folded Spill
-; GFX908-NEXT: buffer_store_dword v27, off, s[0:3], s33 offset:128 ; 4-byte Folded Spill
-; GFX908-NEXT: buffer_store_dword v28, off, s[0:3], s33 offset:132 ; 4-byte Folded Spill
-; GFX908-NEXT: buffer_store_dword v29, off, s[0:3], s33 offset:136 ; 4-byte Folded Spill
-; GFX908-NEXT: buffer_store_dword v30, off, s[0:3], s33 offset:140 ; 4-byte Folded Spill
-; GFX908-NEXT: buffer_store_dword v31, off, s[0:3], s33 offset:144 ; 4-byte Folded Spill
-; GFX908-NEXT: ;;#ASMSTART
-; GFX908-NEXT: ; def v40
-; GFX908-NEXT: ;;#ASMEND
-; GFX908-NEXT: ;;#ASMSTART
-; GFX908-NEXT: ; def s11
-; GFX908-NEXT: ;;#ASMEND
-; GFX908-NEXT: s_or_saveexec_b64 s[34:35], -1
-; GFX908-NEXT: v_mov_b32_e32 v40, v33
-; GFX908-NEXT: s_mov_b64 exec, s[34:35]
-; GFX908-NEXT: v_writelane_b32 v40, s11, 12
-; GFX908-NEXT: ;;#ASMSTART
-; GFX908-NEXT: ; def s12
-; GFX908-NEXT: ;;#ASMEND
-; GFX908-NEXT: v_writelane_b32 v40, s12, 13
-; GFX908-NEXT: ;;#ASMSTART
-; GFX908-NEXT: ; def s13
-; GFX908-NEXT: ;;#ASMEND
-; GFX908-NEXT: v_writelane_b32 v40, s13, 14
-; GFX908-NEXT: ;;#ASMSTART
-; GFX908-NEXT: ; def s14
-; GFX908-NEXT: ;;#ASMEND
-; GFX908-NEXT: v_writelane_b32 v40, s14, 15
-; GFX908-NEXT: ;;#ASMSTART
-; GFX908-NEXT: ; def s15
-; GFX908-NEXT: ;;#ASMEND
-; GFX908-NEXT: v_writelane_b32 v40, s15, 16
-; GFX908-NEXT: s_getpc_b64 s[10:11]
-; GFX908-NEXT: s_add_u32 s10, s10, foo@gotpcrel32@lo+4
-; GFX908-NEXT: s_addc_u32 s11, s11, foo@gotpcrel32@hi+12
-; GFX908-NEXT: ;;#ASMSTART
-; GFX908-NEXT: ; def s16
-; GFX908-NEXT: ;;#ASMEND
-; GFX908-NEXT: v_writelane_b32 v40, s16, 17
-; GFX908-NEXT: s_load_dwordx2 s[10:11], s[10:11], 0x0
-; GFX908-NEXT: ;;#ASMSTART
-; GFX908-NEXT: ; def s17
-; GFX908-NEXT: ;;#ASMEND
-; GFX908-NEXT: v_writelane_b32 v40, s17, 18
-; GFX908-NEXT: ;;#ASMSTART
-; GFX908-NEXT: ; def s18
-; GFX908-NEXT: ;;#ASMEND
-; GFX908-NEXT: v_writelane_b32 v40, s18, 19
-; GFX908-NEXT: ;;#ASMSTART
-; GFX908-NEXT: ; def s19
-; GFX908-NEXT: ;;#ASMEND
-; GFX908-NEXT: v_writelane_b32 v40, s19, 20
-; GFX908-NEXT: ;;#ASMSTART
-; GFX908-NEXT: ; def s20
-; GFX908-NEXT: ;;#ASMEND
-; GFX908-NEXT: v_writelane_b32 v40, s20, 21
-; GFX908-NEXT: s_waitcnt lgkmcnt(0)
-; GFX908-NEXT: v_writelane_b32 v40, s10, 22
-; GFX908-NEXT: v_writelane_b32 v40, s11, 23
-; GFX908-NEXT: s_or_saveexec_b64 s[34:35], -1
-; GFX908-NEXT: s_mov_b64 exec, s[34:35]
-; GFX908-NEXT: v_readlane_b32 s16, v40, 22
-; GFX908-NEXT: s_mov_b32 s12, s24
-; GFX908-NEXT: s_mov_b32 s13, s23
-; GFX908-NEXT: s_mov_b32 s14, s22
-; GFX908-NEXT: v_mov_b32_e32 v31, v32
-; GFX908-NEXT: s_mov_b32 s15, s21
-; GFX908-NEXT: s_mov_b64 s[10:11], s[26:27]
-; GFX908-NEXT: v_readlane_b32 s17, v40, 23
-; GFX908-NEXT: buffer_store_dword v32, off, s[0:3], s33 offset:8 ; 4-byte Folded Spill
-; GFX908-NEXT: s_swappc_b64 s[30:31], s[16:17]
-; GFX908-NEXT: s_or_saveexec_b64 s[34:35], -1
-; GFX908-NEXT: s_mov_b64 exec, s[34:35]
-; GFX908-NEXT: v_readlane_b32 s11, v40, 12
-; GFX908-NEXT: ;;#ASMSTART
-; GFX908-NEXT: ; use s11
-; GFX908-NEXT: ;;#ASMEND
-; GFX908-NEXT: v_readlane_b32 s12, v40, 13
-; GFX908-NEXT: ;;#ASMSTART
-; GFX908-NEXT: ; use s12
-; GFX908-NEXT: ;;#ASMEND
-; GFX908-NEXT: v_readlane_b32 s13, v40, 14
-; GFX908-NEXT: ;;#ASMSTART
-; GFX908-NEXT: ; use s13
-; GFX908-NEXT: ;;#ASMEND
-; GFX908-NEXT: v_readlane_b32 s14, v40, 15
-; GFX908-NEXT: ;;#ASMSTART
-; GFX908-NEXT: ; use s14
-; GFX908-NEXT: ;;#ASMEND
-; GFX908-NEXT: v_readlane_b32 s15, v40, 16
-; GFX908-NEXT: ;;#ASMSTART
-; GFX908-NEXT: ; use s15
-; GFX908-NEXT: ;;#ASMEND
-; GFX908-NEXT: v_readlane_b32 s16, v40, 17
-; GFX908-NEXT: ;;#ASMSTART
-; GFX908-NEXT: ; use s16
-; GFX908-NEXT: ;;#ASMEND
-; GFX908-NEXT: v_readlane_b32 s17, v40, 18
-; GFX908-NEXT: ;;#ASMSTART
-; GFX908-NEXT: ; use s17
-; GFX908-NEXT: ;;#ASMEND
-; GFX908-NEXT: v_readlane_b32 s18, v40, 19
-; GFX908-NEXT: ;;#ASMSTART
-; GFX908-NEXT: ; use s18
-; GFX908-NEXT: ;;#ASMEND
-; GFX908-NEXT: v_readlane_b32 s19, v40, 20
-; GFX908-NEXT: ;;#ASMSTART
-; GFX908-NEXT: ; use s19
-; GFX908-NEXT: ;;#ASMEND
-; GFX908-NEXT: v_readlane_b32 s20, v40, 21
-; GFX908-NEXT: ;;#ASMSTART
-; GFX908-NEXT: ; use s20
-; GFX908-NEXT: ;;#ASMEND
-; GFX908-NEXT: ;;#ASMSTART
-; GFX908-NEXT: ; def s21
-; GFX908-NEXT: ;;#ASMEND
-; GFX908-NEXT: ;;#ASMSTART
-; GFX908-NEXT: ; def s22
-; GFX908-NEXT: ;;#ASMEND
-; GFX908-NEXT: ;;#ASMSTART
-; GFX908-NEXT: ; def s23
-; GFX908-NEXT: ;;#ASMEND
-; GFX908-NEXT: ;;#ASMSTART
-; GFX908-NEXT: ; def s24
-; GFX908-NEXT: ;;#ASMEND
-; GFX908-NEXT: ;;#ASMSTART
-; GFX908-NEXT: ; def s25
-; GFX908-NEXT: ;;#ASMEND
-; GFX908-NEXT: ;;#ASMSTART
-; GFX908-NEXT: ; def s26
-; GFX908-NEXT: ;;#ASMEND
-; GFX908-NEXT: ;;#ASMSTART
-; GFX908-NEXT: ; def s27
-; GFX908-NEXT: ;;#ASMEND
-; GFX908-NEXT: ;;#ASMSTART
-; GFX908-NEXT: ; def s28
-; GFX908-NEXT: ;;#ASMEND
-; GFX908-NEXT: ;;#ASMSTART
-; GFX908-NEXT: ; def s29
-; GFX908-NEXT: ;;#ASMEND
-; GFX908-NEXT: buffer_load_dword v31, off, s[0:3], s33 offset:8 ; 4-byte Folded Reload
-; GFX908-NEXT: v_writelane_b32 v40, s21, 24
-; GFX908-NEXT: v_writelane_b32 v40, s22, 25
-; GFX908-NEXT: v_writelane_b32 v40, s23, 26
-; GFX908-NEXT: v_writelane_b32 v40, s24, 27
-; GFX908-NEXT: v_writelane_b32 v40, s25, 28
-; GFX908-NEXT: v_writelane_b32 v40, s26, 29
-; GFX908-NEXT: v_writelane_b32 v40, s27, 30
-; GFX908-NEXT: v_writelane_b32 v40, s28, 31
-; GFX908-NEXT: v_writelane_b32 v40, s29, 32
-; GFX908-NEXT: v_readlane_b32 s4, v40, 10
-; GFX908-NEXT: v_readlane_b32 s6, v40, 8
-; GFX908-NEXT: v_readlane_b32 s8, v40, 6
-; GFX908-NEXT: v_readlane_b32 s10, v40, 4
-; GFX908-NEXT: v_readlane_b32 s16, v40, 22
-; GFX908-NEXT: v_readlane_b32 s12, v40, 3
-; GFX908-NEXT: v_readlane_b32 s13, v40, 2
-; GFX908-NEXT: v_readlane_b32 s14, v40, 1
-; GFX908-NEXT: v_readlane_b32 s15, v40, 0
-; GFX908-NEXT: v_readlane_b32 s5, v40, 11
-; GFX908-NEXT: v_readlane_b32 s7, v40, 9
-; GFX908-NEXT: v_readlane_b32 s9, v40, 7
-; GFX908-NEXT: v_readlane_b32 s11, v40, 5
-; GFX908-NEXT: v_readlane_b32 s17, v40, 23
-; GFX908-NEXT: s_or_saveexec_b64 s[34:35], -1
-; GFX908-NEXT: s_mov_b64 exec, s[34:35]
-; GFX908-NEXT: s_swappc_b64 s[30:31], s[16:17]
-; GFX908-NEXT: s_or_saveexec_b64 s[34:35], -1
-; GFX908-NEXT: s_mov_b64 exec, s[34:35]
-; GFX908-NEXT: v_readlane_b32 s21, v40, 24
-; GFX908-NEXT: ;;#ASMSTART
-; GFX908-NEXT: ; use s21
-; GFX908-NEXT: ;;#ASMEND
-; GFX908-NEXT: v_readlane_b32 s22, v40, 25
-; GFX908-NEXT: ;;#ASMSTART
-; GFX908-NEXT: ; use s22
-; GFX908-NEXT: ;;#ASMEND
-; GFX908-NEXT: v_readlane_b32 s23, v40, 26
-; GFX908-NEXT: ;;#ASMSTART
-; GFX908-NEXT: ; use s23
-; GFX908-NEXT: ;;#ASMEND
-; GFX908-NEXT: v_readlane_b32 s24, v40, 27
-; GFX908-NEXT: ;;#ASMSTART
-; GFX908-NEXT: ; use s24
-; GFX908-NEXT: ;;#ASMEND
-; GFX908-NEXT: v_readlane_b32 s25, v40, 28
-; GFX908-NEXT: ;;#ASMSTART
-; GFX908-NEXT: ; use s25
-; GFX908-NEXT: ;;#ASMEND
-; GFX908-NEXT: v_readlane_b32 s26, v40, 29
-; GFX908-NEXT: ;;#ASMSTART
-; GFX908-NEXT: ; use s26
-; GFX908-NEXT: ;;#ASMEND
-; GFX908-NEXT: v_readlane_b32 s27, v40, 30
-; GFX908-NEXT: ;;#ASMSTART
-; GFX908-NEXT: ; use s27
-; GFX908-NEXT: ;;#ASMEND
-; GFX908-NEXT: v_readlane_b32 s28, v40, 31
-; GFX908-NEXT: ;;#ASMSTART
-; GFX908-NEXT: ; use s28
-; GFX908-NEXT: ;;#ASMEND
-; GFX908-NEXT: v_readlane_b32 s29, v40, 32
-; GFX908-NEXT: ;;#ASMSTART
-; GFX908-NEXT: ; use s29
-; GFX908-NEXT: ;;#ASMEND
-; GFX908-NEXT: buffer_load_dword v31, off, s[0:3], s33 offset:8 ; 4-byte Folded Reload
-; GFX908-NEXT: v_readlane_b32 s4, v40, 10
-; GFX908-NEXT: v_readlane_b32 s6, v40, 8
-; GFX908-NEXT: v_readlane_b32 s8, v40, 6
-; GFX908-NEXT: v_readlane_b32 s10, v40, 4
-; GFX908-NEXT: v_readlane_b32 s16, v40, 22
-; GFX908-NEXT: v_readlane_b32 s5, v40, 11
-; GFX908-NEXT: v_readlane_b32 s7, v40, 9
-; GFX908-NEXT: v_readlane_b32 s9, v40, 7
-; GFX908-NEXT: v_readlane_b32 s11, v40, 5
-; GFX908-NEXT: v_readlane_b32 s12, v40, 3
-; GFX908-NEXT: v_readlane_b32 s13, v40, 2
-; GFX908-NEXT: v_readlane_b32 s14, v40, 1
-; GFX908-NEXT: v_readlane_b32 s15, v40, 0
-; GFX908-NEXT: v_readlane_b32 s17, v40, 23
-; GFX908-NEXT: s_or_saveexec_b64 s[34:35], -1
-; GFX908-NEXT: s_mov_b64 exec, s[34:35]
-; GFX908-NEXT: s_swappc_b64 s[30:31], s[16:17]
-; GFX908-NEXT: buffer_load_dword v0, off, s[0:3], s33 offset:12 ; 4-byte Folded Reload
-; GFX908-NEXT: buffer_load_dword v1, off, s[0:3], s33 offset:16 ; 4-byte Folded Reload
-; GFX908-NEXT: buffer_load_dword v2, off, s[0:3], s33 offset:20 ; 4-byte Folded Reload
-; GFX908-NEXT: buffer_load_dword v3, off, s[0:3], s33 offset:24 ; 4-byte Folded Reload
-; GFX908-NEXT: buffer_load_dword v4, off, s[0:3], s33 offset:28 ; 4-byte Folded Reload
-; GFX908-NEXT: buffer_load_dword v5, off, s[0:3], s33 offset:32 ; 4-byte Folded Reload
-; GFX908-NEXT: buffer_load_dword v6, off, s[0:3], s33 offset:36 ; 4-byte Folded Reload
-; GFX908-NEXT: buffer_load_dword v7, off, s[0:3], s33 offset:40 ; 4-byte Folded Reload
-; GFX908-NEXT: buffer_load_dword v8, off, s[0:3], s33 offset:44 ; 4-byte Folded Reload
-; GFX908-NEXT: buffer_load_dword v9, off, s[0:3], s33 offset:48 ; 4-byte Folded Reload
-; GFX908-NEXT: buffer_load_dword v10, off, s[0:3], s33 offset:52 ; 4-byte Folded Reload
-; GFX908-NEXT: buffer_load_dword v11, off, s[0:3], s33 offset:56 ; 4-byte Folded Reload
-; GFX908-NEXT: buffer_load_dword v12, off, s[0:3], s33 offset:60 ; 4-byte Folded Reload
-; GFX908-NEXT: buffer_load_dword v13, off, s[0:3], s33 offset:64 ; 4-byte Folded Reload
-; GFX908-NEXT: buffer_load_dword v14, off, s[0:3], s33 offset:68 ; 4-byte Folded Reload
-; GFX908-NEXT: buffer_load_dword v15, off, s[0:3], s33 offset:72 ; 4-byte Folded Reload
-; GFX908-NEXT: buffer_load_dword v16, off, s[0:3], s33 offset:76 ; 4-byte Folded Reload
-; GFX908-NEXT: buffer_load_dword v17, off, s[0:3], s33 offset:80 ; 4-byte Folded Reload
-; GFX908-NEXT: buffer_load_dword v18, off, s[0:3], s33 offset:84 ; 4-byte Folded Reload
-; GFX908-NEXT: buffer_load_dword v19, off, s[0:3], s33 offset:88 ; 4-byte Folded Reload
-; GFX908-NEXT: buffer_load_dword v20, off, s[0:3], s33 offset:92 ; 4-byte Folded Reload
-; GFX908-NEXT: buffer_load_dword v21, off, s[0:3], s33 offset:96 ; 4-byte Folded Reload
-; GFX908-NEXT: buffer_load_dword v22, off, s[0:3], s33 offset:100 ; 4-byte Folded Reload
-; GFX908-NEXT: buffer_load_dword v23, off, s[0:3], s33 offset:104 ; 4-byte Folded Reload
-; GFX908-NEXT: buffer_load_dword v24, off, s[0:3], s33 offset:108 ; 4-byte Folded Reload
-; GFX908-NEXT: buffer_load_dword v25, off, s[0:3], s33 offset:112 ; 4-byte Folded Reload
-; GFX908-NEXT: buffer_load_dword v26, off, s[0:3], s33 offset:116 ; 4-byte Folded Reload
-; GFX908-NEXT: buffer_load_dword v27, off, s[0:3], s33 offset:120 ; 4-byte Folded Reload
-; GFX908-NEXT: buffer_load_dword v28, off, s[0:3], s33 offset:124 ; 4-byte Folded Reload
-; GFX908-NEXT: buffer_load_dword v29, off, s[0:3], s33 offset:128 ; 4-byte Folded Reload
-; GFX908-NEXT: buffer_load_dword v30, off, s[0:3], s33 offset:132 ; 4-byte Folded Reload
-; GFX908-NEXT: buffer_load_dword v31, off, s[0:3], s33 offset:136 ; 4-byte Folded Reload
-; GFX908-NEXT: buffer_load_dword v32, off, s[0:3], s33 offset:140 ; 4-byte Folded Reload
-; GFX908-NEXT: buffer_load_dword v33, off, s[0:3], s33 offset:144 ; 4-byte Folded Reload
-; GFX908-NEXT: s_mov_b64 s[4:5], exec
-; GFX908-NEXT: s_waitcnt vmcnt(0)
-; GFX908-NEXT: flat_store_dwordx4 v[0:1], v[30:33] offset:112
-; GFX908-NEXT: s_waitcnt vmcnt(0)
-; GFX908-NEXT: flat_store_dwordx4 v[0:1], v[26:29] offset:96
-; GFX908-NEXT: s_waitcnt vmcnt(0)
-; GFX908-NEXT: flat_store_dwordx4 v[0:1], v[22:25] offset:80
-; GFX908-NEXT: s_waitcnt vmcnt(0)
-; GFX908-NEXT: flat_store_dwordx4 v[0:1], v[18:21] offset:64
-; GFX908-NEXT: s_waitcnt vmcnt(0)
-; GFX908-NEXT: flat_store_dwordx4 v[0:1], v[14:17] offset:48
-; GFX908-NEXT: s_waitcnt vmcnt(0)
-; GFX908-NEXT: flat_store_dwordx4 v[0:1], v[10:13] offset:32
-; GFX908-NEXT: s_waitcnt vmcnt(0)
-; GFX908-NEXT: flat_store_dwordx4 v[0:1], v[6:9] offset:16
-; GFX908-NEXT: s_waitcnt vmcnt(0)
-; GFX908-NEXT: flat_store_dwordx4 v[0:1], v[2:5]
-; GFX908-NEXT: s_waitcnt vmcnt(0)
-; GFX908-NEXT: s_mov_b64 exec, 1
-; GFX908-NEXT: buffer_store_dword v0, off, s[0:3], s33 offset:164
-; GFX908-NEXT: buffer_load_dword v0, off, s[0:3], s33 offset:4 ; 4-byte Folded Reload
-; GFX908-NEXT: s_waitcnt vmcnt(0)
-; GFX908-NEXT: v_readlane_b32 s31, v0, 0
-; GFX908-NEXT: buffer_load_dword v0, off, s[0:3], s33 offset:164
-; GFX908-NEXT: s_waitcnt vmcnt(0)
-; GFX908-NEXT: s_mov_b64 exec, s[4:5]
-; GFX908-NEXT: s_mov_b64 s[4:5], exec
-; GFX908-NEXT: s_mov_b64 exec, 1
-; GFX908-NEXT: buffer_store_dword v0, off, s[0:3], s33 offset:164
-; GFX908-NEXT: buffer_load_dword v0, off, s[0:3], s33 ; 4-byte Folded Reload
-; GFX908-NEXT: s_waitcnt vmcnt(0)
-; GFX908-NEXT: v_readlane_b32 s30, v0, 0
-; GFX908-NEXT: buffer_load_dword v0, off, s[0:3], s33 offset:164
-; GFX908-NEXT: s_waitcnt vmcnt(0)
-; GFX908-NEXT: s_mov_b64 exec, s[4:5]
-; GFX908-NEXT: buffer_load_dword v0, off, s[0:3], s33 offset:160 ; 4-byte Folded Reload
-; GFX908-NEXT: ; kill: killed $vgpr40
-; GFX908-NEXT: s_waitcnt vmcnt(0)
-; GFX908-NEXT: v_readfirstlane_b32 s4, v0
-; GFX908-NEXT: s_xor_saveexec_b64 s[6:7], -1
-; GFX908-NEXT: buffer_load_dword v33, off, s[0:3], s33 offset:148 ; 4-byte Folded Reload
-; GFX908-NEXT: buffer_load_dword v2, off, s[0:3], s33 offset:156 ; 4-byte Folded Reload
-; GFX908-NEXT: s_mov_b64 exec, -1
-; GFX908-NEXT: buffer_load_dword v40, off, s[0:3], s33 offset:152 ; 4-byte Folded Reload
-; GFX908-NEXT: s_mov_b64 exec, s[6:7]
-; GFX908-NEXT: s_addk_i32 s32, 0xd400
-; GFX908-NEXT: s_mov_b32 s33, s4
-; GFX908-NEXT: s_waitcnt vmcnt(0) lgkmcnt(0)
-; GFX908-NEXT: s_setpc_b64 s[30:31]
- %vreg0 = call <32 x float> asm sideeffect "; def $0", "=v"()
- %v40 = call i32 asm sideeffect "; def $0","=${v40}"()
-
- %s11 = call i32 asm sideeffect "; def $0","=${s11}"()
- %s12 = call i32 asm sideeffect "; def $0","=${s12}"()
- %s13 = call i32 asm sideeffect "; def $0","=${s13}"()
- %s14 = call i32 asm sideeffect "; def $0","=${s14}"()
- %s15 = call i32 asm sideeffect "; def $0","=${s15}"()
- %s16 = call i32 asm sideeffect "; def $0","=${s16}"()
- %s17 = call i32 asm sideeffect "; def $0","=${s17}"()
- %s18 = call i32 asm sideeffect "; def $0","=${s18}"()
- %s19 = call i32 asm sideeffect "; def $0","=${s19}"()
- %s20 = call i32 asm sideeffect "; def $0","=${s20}"()
- call void @foo()
- call void asm sideeffect "; use $0","${s11}"(i32 %s11)
- call void asm sideeffect "; use $0","${s12}"(i32 %s12)
- call void asm sideeffect "; use $0","${s13}"(i32 %s13)
- call void asm sideeffect "; use $0","${s14}"(i32 %s14)
- call void asm sideeffect "; use $0","${s15}"(i32 %s15)
- call void asm sideeffect "; use $0","${s16}"(i32 %s16)
- call void asm sideeffect "; use $0","${s17}"(i32 %s17)
- call void asm sideeffect "; use $0","${s18}"(i32 %s18)
- call void asm sideeffect "; use $0","${s19}"(i32 %s19)
- call void asm sideeffect "; use $0","${s20}"(i32 %s20)
-
- %s21 = call i32 asm sideeffect "; def $0","=${s21}"()
- %s22 = call i32 asm sideeffect "; def $0","=${s22}"()
- %s23 = call i32 asm sideeffect "; def $0","=${s23}"()
- %s24 = call i32 asm sideeffect "; def $0","=${s24}"()
- %s25 = call i32 asm sideeffect "; def $0","=${s25}"()
- %s26 = call i32 asm sideeffect "; def $0","=${s26}"()
- %s27 = call i32 asm sideeffect "; def $0","=${s27}"()
- %s28 = call i32 asm sideeffect "; def $0","=${s28}"()
- %s29 = call i32 asm sideeffect "; def $0","=${s29}"()
- call void @foo()
- call void asm sideeffect "; use $0","${s21}"(i32 %s21)
- call void asm sideeffect "; use $0","${s22}"(i32 %s22)
- call void asm sideeffect "; use $0","${s23}"(i32 %s23)
- call void asm sideeffect "; use $0","${s24}"(i32 %s24)
- call void asm sideeffect "; use $0","${s25}"(i32 %s25)
- call void asm sideeffect "; use $0","${s26}"(i32 %s26)
- call void asm sideeffect "; use $0","${s27}"(i32 %s27)
- call void asm sideeffect "; use $0","${s28}"(i32 %s28)
- call void asm sideeffect "; use $0","${s29}"(i32 %s29)
-
- call void @foo()
-
- store volatile <32 x float> %vreg0, ptr %parg0
-
- ret void
-}
-
-declare void @foo()
-
-attributes #0 = { "amdgpu-num-vgpr"="42" "amdgpu-num-sgpr"="40"}
diff --git a/llvm/test/CodeGen/AMDGPU/scc-clobbered-sgpr-to-vmem-spill.ll b/llvm/test/CodeGen/AMDGPU/scc-clobbered-sgpr-to-vmem-spill.ll
--- a/llvm/test/CodeGen/AMDGPU/scc-clobbered-sgpr-to-vmem-spill.ll
+++ b/llvm/test/CodeGen/AMDGPU/scc-clobbered-sgpr-to-vmem-spill.ll
@@ -1,381 +1,22 @@
-; NOTE: Assertions have been autogenerated by utils/update_llc_test_checks.py UTC_ARGS: --version 2
-; RUN: llc -mtriple=amdgcn--amdhsa -mcpu=gfx900 -verify-machineinstrs < %s | FileCheck %s
+; RUN: not --crash llc -mtriple=amdgcn--amdhsa -mcpu=gfx900 -verify-machineinstrs -o /dev/null %s 2>&1 | FileCheck %s
-; This was a negative test to catch an extreme case when all options are exhausted
-; while trying to spill SGPRs to memory. After we enabled SGPR spills into virtual VGPRs
-; the edge case won't arise and the test would always compile.
+; This ends up needing to spill SGPRs to memory, and also does not
+; have any free SGPRs available to save the exec mask when doing so.
+; The register scavenger also needs to use the emergency stack slot,
+; which tries to place the scavenged register restore instruction as
+; far the block as possible, near the terminator. This places a
+; restore instruction between the condition and the conditional
+; branch, which gets expanded into a sequence involving s_not_b64 on
+; the exec mask, clobbering SCC value before the branch. We probably
+; have to stop relying on being able to flip and restore the exec
+; mask, and always require a free SGPR for saving exec.
+; CHECK: *** Bad machine code: Using an undefined physical register ***
+; CHECK-NEXT: - function: kernel0
+; CHECK-NEXT: - basic block: %bb.0
+; CHECK-NEXT: - instruction: S_CBRANCH_SCC1 %bb.2, implicit killed $scc
+; CHECK-NEXT: - operand 1: implicit killed $scc
define amdgpu_kernel void @kernel0(ptr addrspace(1) %out, i32 %in) #1 {
-; CHECK-LABEL: kernel0:
-; CHECK: ; %bb.0:
-; CHECK-NEXT: ; implicit-def: $vgpr23
-; CHECK-NEXT: ;;#ASMSTART
-; CHECK-NEXT: ;;#ASMEND
-; CHECK-NEXT: ;;#ASMSTART
-; CHECK-NEXT: ;;#ASMEND
-; CHECK-NEXT: ;;#ASMSTART
-; CHECK-NEXT: ;;#ASMEND
-; CHECK-NEXT: ;;#ASMSTART
-; CHECK-NEXT: ;;#ASMEND
-; CHECK-NEXT: ;;#ASMSTART
-; CHECK-NEXT: ;;#ASMEND
-; CHECK-NEXT: ;;#ASMSTART
-; CHECK-NEXT: ; def s[2:3]
-; CHECK-NEXT: ;;#ASMEND
-; CHECK-NEXT: v_writelane_b32 v23, s2, 0
-; CHECK-NEXT: s_load_dword s0, s[4:5], 0x8
-; CHECK-NEXT: v_writelane_b32 v23, s3, 1
-; CHECK-NEXT: ;;#ASMSTART
-; CHECK-NEXT: ; def s[4:7]
-; CHECK-NEXT: ;;#ASMEND
-; CHECK-NEXT: v_writelane_b32 v23, s4, 2
-; CHECK-NEXT: v_writelane_b32 v23, s5, 3
-; CHECK-NEXT: v_writelane_b32 v23, s6, 4
-; CHECK-NEXT: v_writelane_b32 v23, s7, 5
-; CHECK-NEXT: ;;#ASMSTART
-; CHECK-NEXT: ; def s[4:11]
-; CHECK-NEXT: ;;#ASMEND
-; CHECK-NEXT: v_writelane_b32 v23, s4, 6
-; CHECK-NEXT: v_writelane_b32 v23, s5, 7
-; CHECK-NEXT: v_writelane_b32 v23, s6, 8
-; CHECK-NEXT: v_writelane_b32 v23, s7, 9
-; CHECK-NEXT: v_writelane_b32 v23, s8, 10
-; CHECK-NEXT: v_writelane_b32 v23, s9, 11
-; CHECK-NEXT: v_writelane_b32 v23, s10, 12
-; CHECK-NEXT: v_writelane_b32 v23, s11, 13
-; CHECK-NEXT: ;;#ASMSTART
-; CHECK-NEXT: ; def s[4:19]
-; CHECK-NEXT: ;;#ASMEND
-; CHECK-NEXT: v_writelane_b32 v23, s4, 14
-; CHECK-NEXT: v_writelane_b32 v23, s5, 15
-; CHECK-NEXT: v_writelane_b32 v23, s6, 16
-; CHECK-NEXT: v_writelane_b32 v23, s7, 17
-; CHECK-NEXT: v_writelane_b32 v23, s8, 18
-; CHECK-NEXT: v_writelane_b32 v23, s9, 19
-; CHECK-NEXT: v_writelane_b32 v23, s10, 20
-; CHECK-NEXT: v_writelane_b32 v23, s11, 21
-; CHECK-NEXT: v_writelane_b32 v23, s12, 22
-; CHECK-NEXT: v_writelane_b32 v23, s13, 23
-; CHECK-NEXT: v_writelane_b32 v23, s14, 24
-; CHECK-NEXT: v_writelane_b32 v23, s15, 25
-; CHECK-NEXT: v_writelane_b32 v23, s16, 26
-; CHECK-NEXT: v_writelane_b32 v23, s17, 27
-; CHECK-NEXT: v_writelane_b32 v23, s18, 28
-; CHECK-NEXT: v_writelane_b32 v23, s19, 29
-; CHECK-NEXT: ;;#ASMSTART
-; CHECK-NEXT: ; def s[2:3]
-; CHECK-NEXT: ;;#ASMEND
-; CHECK-NEXT: v_writelane_b32 v23, s2, 30
-; CHECK-NEXT: v_writelane_b32 v23, s3, 31
-; CHECK-NEXT: ;;#ASMSTART
-; CHECK-NEXT: ; def s[4:7]
-; CHECK-NEXT: ;;#ASMEND
-; CHECK-NEXT: v_writelane_b32 v23, s4, 32
-; CHECK-NEXT: v_writelane_b32 v23, s5, 33
-; CHECK-NEXT: v_writelane_b32 v23, s6, 34
-; CHECK-NEXT: v_writelane_b32 v23, s7, 35
-; CHECK-NEXT: ;;#ASMSTART
-; CHECK-NEXT: ; def s[4:11]
-; CHECK-NEXT: ;;#ASMEND
-; CHECK-NEXT: v_writelane_b32 v23, s4, 36
-; CHECK-NEXT: v_writelane_b32 v23, s5, 37
-; CHECK-NEXT: v_writelane_b32 v23, s6, 38
-; CHECK-NEXT: v_writelane_b32 v23, s7, 39
-; CHECK-NEXT: v_writelane_b32 v23, s8, 40
-; CHECK-NEXT: v_writelane_b32 v23, s9, 41
-; CHECK-NEXT: v_writelane_b32 v23, s10, 42
-; CHECK-NEXT: v_writelane_b32 v23, s11, 43
-; CHECK-NEXT: s_waitcnt lgkmcnt(0)
-; CHECK-NEXT: s_cmp_lg_u32 s0, 0
-; CHECK-NEXT: ;;#ASMSTART
-; CHECK-NEXT: ; def s[16:31]
-; CHECK-NEXT: ;;#ASMEND
-; CHECK-NEXT: ;;#ASMSTART
-; CHECK-NEXT: ; def s[52:53]
-; CHECK-NEXT: ;;#ASMEND
-; CHECK-NEXT: ;;#ASMSTART
-; CHECK-NEXT: ; def s[48:51]
-; CHECK-NEXT: ;;#ASMEND
-; CHECK-NEXT: ;;#ASMSTART
-; CHECK-NEXT: ; def s[36:43]
-; CHECK-NEXT: ;;#ASMEND
-; CHECK-NEXT: ;;#ASMSTART
-; CHECK-NEXT: ; def s[0:15]
-; CHECK-NEXT: ;;#ASMEND
-; CHECK-NEXT: v_writelane_b32 v23, s0, 44
-; CHECK-NEXT: v_writelane_b32 v23, s1, 45
-; CHECK-NEXT: v_writelane_b32 v23, s2, 46
-; CHECK-NEXT: v_writelane_b32 v23, s3, 47
-; CHECK-NEXT: v_writelane_b32 v23, s4, 48
-; CHECK-NEXT: v_writelane_b32 v23, s5, 49
-; CHECK-NEXT: v_writelane_b32 v23, s6, 50
-; CHECK-NEXT: v_writelane_b32 v23, s7, 51
-; CHECK-NEXT: v_writelane_b32 v23, s8, 52
-; CHECK-NEXT: v_writelane_b32 v23, s9, 53
-; CHECK-NEXT: v_writelane_b32 v23, s10, 54
-; CHECK-NEXT: v_writelane_b32 v23, s11, 55
-; CHECK-NEXT: v_writelane_b32 v23, s12, 56
-; CHECK-NEXT: v_writelane_b32 v23, s13, 57
-; CHECK-NEXT: v_writelane_b32 v23, s14, 58
-; CHECK-NEXT: ; implicit-def: $vgpr0
-; CHECK-NEXT: v_writelane_b32 v23, s15, 59
-; CHECK-NEXT: ;;#ASMSTART
-; CHECK-NEXT: ; def s[34:35]
-; CHECK-NEXT: ;;#ASMEND
-; CHECK-NEXT: ;;#ASMSTART
-; CHECK-NEXT: ; def s[44:47]
-; CHECK-NEXT: ;;#ASMEND
-; CHECK-NEXT: ;;#ASMSTART
-; CHECK-NEXT: ; def s[0:7]
-; CHECK-NEXT: ;;#ASMEND
-; CHECK-NEXT: v_writelane_b32 v23, s0, 60
-; CHECK-NEXT: v_writelane_b32 v0, s4, 0
-; CHECK-NEXT: v_writelane_b32 v23, s1, 61
-; CHECK-NEXT: v_writelane_b32 v0, s5, 1
-; CHECK-NEXT: v_writelane_b32 v23, s2, 62
-; CHECK-NEXT: v_writelane_b32 v0, s6, 2
-; CHECK-NEXT: v_writelane_b32 v23, s3, 63
-; CHECK-NEXT: v_writelane_b32 v0, s7, 3
-; CHECK-NEXT: ;;#ASMSTART
-; CHECK-NEXT: ; def s[0:15]
-; CHECK-NEXT: ;;#ASMEND
-; CHECK-NEXT: v_writelane_b32 v0, s0, 4
-; CHECK-NEXT: v_writelane_b32 v0, s1, 5
-; CHECK-NEXT: v_writelane_b32 v0, s2, 6
-; CHECK-NEXT: v_writelane_b32 v0, s3, 7
-; CHECK-NEXT: v_writelane_b32 v0, s4, 8
-; CHECK-NEXT: v_writelane_b32 v0, s5, 9
-; CHECK-NEXT: v_writelane_b32 v0, s6, 10
-; CHECK-NEXT: v_writelane_b32 v0, s7, 11
-; CHECK-NEXT: v_writelane_b32 v0, s8, 12
-; CHECK-NEXT: v_writelane_b32 v0, s9, 13
-; CHECK-NEXT: v_writelane_b32 v0, s10, 14
-; CHECK-NEXT: v_writelane_b32 v0, s11, 15
-; CHECK-NEXT: v_writelane_b32 v0, s12, 16
-; CHECK-NEXT: v_writelane_b32 v0, s13, 17
-; CHECK-NEXT: v_writelane_b32 v0, s14, 18
-; CHECK-NEXT: v_writelane_b32 v0, s15, 19
-; CHECK-NEXT: ;;#ASMSTART
-; CHECK-NEXT: ; def s[54:55]
-; CHECK-NEXT: ;;#ASMEND
-; CHECK-NEXT: ;;#ASMSTART
-; CHECK-NEXT: ; def s[0:3]
-; CHECK-NEXT: ;;#ASMEND
-; CHECK-NEXT: v_writelane_b32 v0, s0, 20
-; CHECK-NEXT: v_writelane_b32 v0, s1, 21
-; CHECK-NEXT: v_writelane_b32 v0, s2, 22
-; CHECK-NEXT: v_writelane_b32 v0, s3, 23
-; CHECK-NEXT: ;;#ASMSTART
-; CHECK-NEXT: ; def s[0:7]
-; CHECK-NEXT: ;;#ASMEND
-; CHECK-NEXT: v_writelane_b32 v0, s0, 24
-; CHECK-NEXT: v_writelane_b32 v0, s1, 25
-; CHECK-NEXT: v_writelane_b32 v0, s2, 26
-; CHECK-NEXT: v_writelane_b32 v0, s3, 27
-; CHECK-NEXT: v_writelane_b32 v0, s4, 28
-; CHECK-NEXT: v_writelane_b32 v0, s5, 29
-; CHECK-NEXT: v_writelane_b32 v0, s6, 30
-; CHECK-NEXT: v_writelane_b32 v0, s7, 31
-; CHECK-NEXT: ;;#ASMSTART
-; CHECK-NEXT: ; def s[0:15]
-; CHECK-NEXT: ;;#ASMEND
-; CHECK-NEXT: v_writelane_b32 v0, s0, 32
-; CHECK-NEXT: v_writelane_b32 v0, s1, 33
-; CHECK-NEXT: v_writelane_b32 v0, s2, 34
-; CHECK-NEXT: v_writelane_b32 v0, s3, 35
-; CHECK-NEXT: v_writelane_b32 v0, s4, 36
-; CHECK-NEXT: v_writelane_b32 v0, s5, 37
-; CHECK-NEXT: v_writelane_b32 v0, s6, 38
-; CHECK-NEXT: v_writelane_b32 v0, s7, 39
-; CHECK-NEXT: v_writelane_b32 v0, s8, 40
-; CHECK-NEXT: v_writelane_b32 v0, s9, 41
-; CHECK-NEXT: v_writelane_b32 v0, s10, 42
-; CHECK-NEXT: v_writelane_b32 v0, s11, 43
-; CHECK-NEXT: v_writelane_b32 v0, s12, 44
-; CHECK-NEXT: v_writelane_b32 v0, s13, 45
-; CHECK-NEXT: v_writelane_b32 v0, s14, 46
-; CHECK-NEXT: v_writelane_b32 v0, s15, 47
-; CHECK-NEXT: s_cbranch_scc0 .LBB0_2
-; CHECK-NEXT: ; %bb.1: ; %ret
-; CHECK-NEXT: ; kill: killed $vgpr23
-; CHECK-NEXT: ; kill: killed $vgpr0
-; CHECK-NEXT: s_endpgm
-; CHECK-NEXT: .LBB0_2: ; %bb0
-; CHECK-NEXT: v_readlane_b32 s0, v23, 0
-; CHECK-NEXT: v_readlane_b32 s1, v23, 1
-; CHECK-NEXT: ;;#ASMSTART
-; CHECK-NEXT: ; use s[0:1]
-; CHECK-NEXT: ;;#ASMEND
-; CHECK-NEXT: v_readlane_b32 s0, v23, 2
-; CHECK-NEXT: v_readlane_b32 s1, v23, 3
-; CHECK-NEXT: v_readlane_b32 s2, v23, 4
-; CHECK-NEXT: v_readlane_b32 s3, v23, 5
-; CHECK-NEXT: ;;#ASMSTART
-; CHECK-NEXT: ; use s[0:3]
-; CHECK-NEXT: ;;#ASMEND
-; CHECK-NEXT: v_readlane_b32 s0, v23, 6
-; CHECK-NEXT: v_readlane_b32 s1, v23, 7
-; CHECK-NEXT: v_readlane_b32 s2, v23, 8
-; CHECK-NEXT: v_readlane_b32 s3, v23, 9
-; CHECK-NEXT: v_readlane_b32 s4, v23, 10
-; CHECK-NEXT: v_readlane_b32 s5, v23, 11
-; CHECK-NEXT: v_readlane_b32 s6, v23, 12
-; CHECK-NEXT: v_readlane_b32 s7, v23, 13
-; CHECK-NEXT: ;;#ASMSTART
-; CHECK-NEXT: ; use s[0:7]
-; CHECK-NEXT: ;;#ASMEND
-; CHECK-NEXT: v_readlane_b32 s0, v23, 14
-; CHECK-NEXT: v_readlane_b32 s1, v23, 15
-; CHECK-NEXT: v_readlane_b32 s2, v23, 16
-; CHECK-NEXT: v_readlane_b32 s3, v23, 17
-; CHECK-NEXT: v_readlane_b32 s4, v23, 18
-; CHECK-NEXT: v_readlane_b32 s5, v23, 19
-; CHECK-NEXT: v_readlane_b32 s6, v23, 20
-; CHECK-NEXT: v_readlane_b32 s7, v23, 21
-; CHECK-NEXT: v_readlane_b32 s8, v23, 22
-; CHECK-NEXT: v_readlane_b32 s9, v23, 23
-; CHECK-NEXT: v_readlane_b32 s10, v23, 24
-; CHECK-NEXT: v_readlane_b32 s11, v23, 25
-; CHECK-NEXT: v_readlane_b32 s12, v23, 26
-; CHECK-NEXT: v_readlane_b32 s13, v23, 27
-; CHECK-NEXT: v_readlane_b32 s14, v23, 28
-; CHECK-NEXT: v_readlane_b32 s15, v23, 29
-; CHECK-NEXT: ;;#ASMSTART
-; CHECK-NEXT: ; use s[0:15]
-; CHECK-NEXT: ;;#ASMEND
-; CHECK-NEXT: v_readlane_b32 s0, v23, 30
-; CHECK-NEXT: v_readlane_b32 s1, v23, 31
-; CHECK-NEXT: ;;#ASMSTART
-; CHECK-NEXT: ; use s[0:1]
-; CHECK-NEXT: ;;#ASMEND
-; CHECK-NEXT: v_readlane_b32 s0, v23, 32
-; CHECK-NEXT: v_readlane_b32 s1, v23, 33
-; CHECK-NEXT: v_readlane_b32 s2, v23, 34
-; CHECK-NEXT: v_readlane_b32 s3, v23, 35
-; CHECK-NEXT: ;;#ASMSTART
-; CHECK-NEXT: ; use s[0:3]
-; CHECK-NEXT: ;;#ASMEND
-; CHECK-NEXT: v_readlane_b32 s0, v23, 36
-; CHECK-NEXT: v_readlane_b32 s1, v23, 37
-; CHECK-NEXT: v_readlane_b32 s2, v23, 38
-; CHECK-NEXT: v_readlane_b32 s3, v23, 39
-; CHECK-NEXT: v_readlane_b32 s4, v23, 40
-; CHECK-NEXT: v_readlane_b32 s5, v23, 41
-; CHECK-NEXT: v_readlane_b32 s6, v23, 42
-; CHECK-NEXT: v_readlane_b32 s7, v23, 43
-; CHECK-NEXT: ;;#ASMSTART
-; CHECK-NEXT: ; use s[0:7]
-; CHECK-NEXT: ;;#ASMEND
-; CHECK-NEXT: v_readlane_b32 s0, v23, 44
-; CHECK-NEXT: v_readlane_b32 s1, v23, 45
-; CHECK-NEXT: v_readlane_b32 s2, v23, 46
-; CHECK-NEXT: v_readlane_b32 s3, v23, 47
-; CHECK-NEXT: v_readlane_b32 s4, v23, 48
-; CHECK-NEXT: v_readlane_b32 s5, v23, 49
-; CHECK-NEXT: v_readlane_b32 s6, v23, 50
-; CHECK-NEXT: v_readlane_b32 s7, v23, 51
-; CHECK-NEXT: ;;#ASMSTART
-; CHECK-NEXT: ; use s[16:31]
-; CHECK-NEXT: ;;#ASMEND
-; CHECK-NEXT: ;;#ASMSTART
-; CHECK-NEXT: ; use s[52:53]
-; CHECK-NEXT: ;;#ASMEND
-; CHECK-NEXT: ;;#ASMSTART
-; CHECK-NEXT: ; use s[48:51]
-; CHECK-NEXT: ;;#ASMEND
-; CHECK-NEXT: ;;#ASMSTART
-; CHECK-NEXT: ; use s[36:43]
-; CHECK-NEXT: ;;#ASMEND
-; CHECK-NEXT: v_readlane_b32 s8, v23, 52
-; CHECK-NEXT: v_readlane_b32 s9, v23, 53
-; CHECK-NEXT: v_readlane_b32 s10, v23, 54
-; CHECK-NEXT: v_readlane_b32 s11, v23, 55
-; CHECK-NEXT: v_readlane_b32 s12, v23, 56
-; CHECK-NEXT: v_readlane_b32 s13, v23, 57
-; CHECK-NEXT: v_readlane_b32 s14, v23, 58
-; CHECK-NEXT: v_readlane_b32 s15, v23, 59
-; CHECK-NEXT: ;;#ASMSTART
-; CHECK-NEXT: ; use s[0:15]
-; CHECK-NEXT: ;;#ASMEND
-; CHECK-NEXT: v_readlane_b32 s0, v23, 60
-; CHECK-NEXT: v_readlane_b32 s1, v23, 61
-; CHECK-NEXT: v_readlane_b32 s2, v23, 62
-; CHECK-NEXT: v_readlane_b32 s3, v23, 63
-; CHECK-NEXT: v_readlane_b32 s4, v0, 0
-; CHECK-NEXT: v_readlane_b32 s5, v0, 1
-; CHECK-NEXT: v_readlane_b32 s6, v0, 2
-; CHECK-NEXT: v_readlane_b32 s7, v0, 3
-; CHECK-NEXT: ;;#ASMSTART
-; CHECK-NEXT: ; use s[34:35]
-; CHECK-NEXT: ;;#ASMEND
-; CHECK-NEXT: ;;#ASMSTART
-; CHECK-NEXT: ; use s[44:47]
-; CHECK-NEXT: ;;#ASMEND
-; CHECK-NEXT: ;;#ASMSTART
-; CHECK-NEXT: ; use s[0:7]
-; CHECK-NEXT: ;;#ASMEND
-; CHECK-NEXT: v_readlane_b32 s0, v0, 4
-; CHECK-NEXT: v_readlane_b32 s1, v0, 5
-; CHECK-NEXT: v_readlane_b32 s2, v0, 6
-; CHECK-NEXT: v_readlane_b32 s3, v0, 7
-; CHECK-NEXT: v_readlane_b32 s4, v0, 8
-; CHECK-NEXT: v_readlane_b32 s5, v0, 9
-; CHECK-NEXT: v_readlane_b32 s6, v0, 10
-; CHECK-NEXT: v_readlane_b32 s7, v0, 11
-; CHECK-NEXT: v_readlane_b32 s8, v0, 12
-; CHECK-NEXT: v_readlane_b32 s9, v0, 13
-; CHECK-NEXT: v_readlane_b32 s10, v0, 14
-; CHECK-NEXT: v_readlane_b32 s11, v0, 15
-; CHECK-NEXT: v_readlane_b32 s12, v0, 16
-; CHECK-NEXT: v_readlane_b32 s13, v0, 17
-; CHECK-NEXT: v_readlane_b32 s14, v0, 18
-; CHECK-NEXT: v_readlane_b32 s15, v0, 19
-; CHECK-NEXT: ;;#ASMSTART
-; CHECK-NEXT: ; use s[0:15]
-; CHECK-NEXT: ;;#ASMEND
-; CHECK-NEXT: v_readlane_b32 s0, v0, 20
-; CHECK-NEXT: v_readlane_b32 s1, v0, 21
-; CHECK-NEXT: v_readlane_b32 s2, v0, 22
-; CHECK-NEXT: v_readlane_b32 s3, v0, 23
-; CHECK-NEXT: ;;#ASMSTART
-; CHECK-NEXT: ; use s[54:55]
-; CHECK-NEXT: ;;#ASMEND
-; CHECK-NEXT: ;;#ASMSTART
-; CHECK-NEXT: ; use s[0:3]
-; CHECK-NEXT: ;;#ASMEND
-; CHECK-NEXT: v_readlane_b32 s0, v0, 24
-; CHECK-NEXT: v_readlane_b32 s1, v0, 25
-; CHECK-NEXT: v_readlane_b32 s2, v0, 26
-; CHECK-NEXT: v_readlane_b32 s3, v0, 27
-; CHECK-NEXT: v_readlane_b32 s4, v0, 28
-; CHECK-NEXT: v_readlane_b32 s5, v0, 29
-; CHECK-NEXT: v_readlane_b32 s6, v0, 30
-; CHECK-NEXT: v_readlane_b32 s7, v0, 31
-; CHECK-NEXT: ;;#ASMSTART
-; CHECK-NEXT: ; use s[0:7]
-; CHECK-NEXT: ;;#ASMEND
-; CHECK-NEXT: v_readlane_b32 s0, v0, 32
-; CHECK-NEXT: v_readlane_b32 s1, v0, 33
-; CHECK-NEXT: v_readlane_b32 s2, v0, 34
-; CHECK-NEXT: v_readlane_b32 s3, v0, 35
-; CHECK-NEXT: v_readlane_b32 s4, v0, 36
-; CHECK-NEXT: v_readlane_b32 s5, v0, 37
-; CHECK-NEXT: v_readlane_b32 s6, v0, 38
-; CHECK-NEXT: v_readlane_b32 s7, v0, 39
-; CHECK-NEXT: v_readlane_b32 s8, v0, 40
-; CHECK-NEXT: v_readlane_b32 s9, v0, 41
-; CHECK-NEXT: v_readlane_b32 s10, v0, 42
-; CHECK-NEXT: v_readlane_b32 s11, v0, 43
-; CHECK-NEXT: v_readlane_b32 s12, v0, 44
-; CHECK-NEXT: v_readlane_b32 s13, v0, 45
-; CHECK-NEXT: v_readlane_b32 s14, v0, 46
-; CHECK-NEXT: v_readlane_b32 s15, v0, 47
-; CHECK-NEXT: ;;#ASMSTART
-; CHECK-NEXT: ; use s[0:15]
-; CHECK-NEXT: ;;#ASMEND
-; CHECK-NEXT: ; kill: killed $vgpr23
-; CHECK-NEXT: ; kill: killed $vgpr0
-; CHECK-NEXT: s_endpgm
call void asm sideeffect "", "~{v[0:7]}" () #0
call void asm sideeffect "", "~{v[8:15]}" () #0
call void asm sideeffect "", "~{v[16:19]}"() #0
diff --git a/llvm/test/CodeGen/AMDGPU/sgpr-regalloc-flags.ll b/llvm/test/CodeGen/AMDGPU/sgpr-regalloc-flags.ll
--- a/llvm/test/CodeGen/AMDGPU/sgpr-regalloc-flags.ll
+++ b/llvm/test/CodeGen/AMDGPU/sgpr-regalloc-flags.ll
@@ -21,7 +21,6 @@
; DEFAULT-NEXT: Virtual Register Map
; DEFAULT-NEXT: Live Register Matrix
; DEFAULT-NEXT: Greedy Register Allocator
-; DEFAULT-NEXT: SI Lower WWM Copies
; DEFAULT-NEXT: GCN NSA Reassign
; DEFAULT-NEXT: Virtual Register Rewriter
; DEFAULT-NEXT: Stack Slot Coloring
@@ -29,7 +28,6 @@
; O0: Fast Register Allocator
; O0-NEXT: SI lower SGPR spill instructions
; O0-NEXT: Fast Register Allocator
-; O0-NEXT: SI Lower WWM Copies
; O0-NEXT: SI Fix VGPR copies
@@ -51,7 +49,6 @@
; BASIC-DEFAULT-NEXT: Lazy Machine Block Frequency Analysis
; BASIC-DEFAULT-NEXT: Machine Optimization Remark Emitter
; BASIC-DEFAULT-NEXT: Greedy Register Allocator
-; BASIC-DEFAULT-NEXT: SI Lower WWM Copies
; BASIC-DEFAULT-NEXT: GCN NSA Reassign
; BASIC-DEFAULT-NEXT: Virtual Register Rewriter
; BASIC-DEFAULT-NEXT: Stack Slot Coloring
@@ -64,7 +61,6 @@
; DEFAULT-BASIC-NEXT: Virtual Register Map
; DEFAULT-BASIC-NEXT: Live Register Matrix
; DEFAULT-BASIC-NEXT: Basic Register Allocator
-; DEFAULT-BASIC-NEXT: SI Lower WWM Copies
; DEFAULT-BASIC-NEXT: GCN NSA Reassign
; DEFAULT-BASIC-NEXT: Virtual Register Rewriter
; DEFAULT-BASIC-NEXT: Stack Slot Coloring
@@ -83,7 +79,6 @@
; BASIC-BASIC-NEXT: Virtual Register Map
; BASIC-BASIC-NEXT: Live Register Matrix
; BASIC-BASIC-NEXT: Basic Register Allocator
-; BASIC-BASIC-NEXT: SI Lower WWM Copies
; BASIC-BASIC-NEXT: GCN NSA Reassign
; BASIC-BASIC-NEXT: Virtual Register Rewriter
; BASIC-BASIC-NEXT: Stack Slot Coloring
diff --git a/llvm/test/CodeGen/AMDGPU/sgpr-spill-dead-frame-in-dbg-value.mir b/llvm/test/CodeGen/AMDGPU/sgpr-spill-dead-frame-in-dbg-value.mir
--- a/llvm/test/CodeGen/AMDGPU/sgpr-spill-dead-frame-in-dbg-value.mir
+++ b/llvm/test/CodeGen/AMDGPU/sgpr-spill-dead-frame-in-dbg-value.mir
@@ -1,5 +1,4 @@
-# RUN: llc -mtriple=amdgcn-amd-amdhsa -mcpu=gfx908 -amdgpu-spill-sgpr-to-vgpr=true -verify-machineinstrs -run-pass=si-lower-sgpr-spills -o - %s | FileCheck -check-prefix=SGPR_SPILL %s
-# RUN: llc -mtriple=amdgcn-amd-amdhsa -mcpu=gfx908 -amdgpu-spill-sgpr-to-vgpr=true -verify-machineinstrs --start-before=si-lower-sgpr-spills --stop-after=prologepilog -o - %s | FileCheck -check-prefix=PEI %s
+# RUN: llc -mtriple=amdgcn-amd-amdhsa -mcpu=gfx908 -amdgpu-spill-sgpr-to-vgpr=true -verify-machineinstrs -run-pass=si-lower-sgpr-spills,prologepilog -o - %s | FileCheck %s
# After handling the SGPR spill to VGPR in SILowerSGPRSpills pass, replace the dead frame index in the DBG_VALUE instruction with reg 0.
# Otherwise, the test would crash during PEI while trying to replace the dead frame index.
@@ -40,33 +39,13 @@
workGroupIDX: { reg: '$sgpr8' }
privateSegmentWaveByteOffset: { reg: '$sgpr9' }
body: |
- ; SGPR_SPILL-LABEL: name: test
- ; SGPR_SPILL: bb.0:
- ; SGPR_SPILL-NEXT: successors: %bb.1(0x80000000)
- ; SGPR_SPILL-NEXT: {{ $}}
- ; SGPR_SPILL-NEXT: [[DEF:%[0-9]+]]:vgpr_32 = IMPLICIT_DEF
- ; SGPR_SPILL-NEXT: renamable $sgpr10 = IMPLICIT_DEF
- ; SGPR_SPILL-NEXT: [[V_WRITELANE_B32_:%[0-9]+]]:vgpr_32 = V_WRITELANE_B32 killed $sgpr10, 0, [[V_WRITELANE_B32_]]
- ; SGPR_SPILL-NEXT: DBG_VALUE $noreg, 0
- ; SGPR_SPILL-NEXT: {{ $}}
- ; SGPR_SPILL-NEXT: bb.1:
- ; SGPR_SPILL-NEXT: $sgpr10 = V_READLANE_B32 [[V_WRITELANE_B32_]], 0
- ; SGPR_SPILL-NEXT: KILL [[V_WRITELANE_B32_]]
- ; SGPR_SPILL-NEXT: S_ENDPGM 0
- ; PEI-LABEL: name: test
- ; PEI: bb.0:
- ; PEI-NEXT: successors: %bb.1(0x80000000)
- ; PEI-NEXT: {{ $}}
- ; PEI-NEXT: renamable $vgpr0 = IMPLICIT_DEF
- ; PEI-NEXT: renamable $sgpr10 = IMPLICIT_DEF
- ; PEI-NEXT: renamable $vgpr0 = V_WRITELANE_B32 killed $sgpr10, 0, killed $vgpr0
- ; PEI-NEXT: {{ $}}
- ; PEI-NEXT: bb.1:
- ; PEI-NEXT: liveins: $vgpr0
- ; PEI-NEXT: {{ $}}
- ; PEI-NEXT: $sgpr10 = V_READLANE_B32 $vgpr0, 0
- ; PEI-NEXT: KILL killed renamable $vgpr0
- ; PEI-NEXT: S_ENDPGM 0
+ ; CHECK-LABEL: name: test
+ ; CHECK: bb.0:
+ ; CHECK: $vgpr0 = V_WRITELANE_B32 killed $sgpr10, 0, $vgpr0
+ ; CHECK: DBG_VALUE $noreg, 0
+ ; CHECK: bb.1:
+ ; CHECK: $sgpr10 = V_READLANE_B32 $vgpr0, 0
+ ; CHECK: S_ENDPGM 0
bb.0:
renamable $sgpr10 = IMPLICIT_DEF
SI_SPILL_S32_SAVE killed $sgpr10, %stack.0, implicit $exec, implicit $sgpr96_sgpr97_sgpr98_sgpr99, implicit $sgpr32
diff --git a/llvm/test/CodeGen/AMDGPU/sgpr-spill-fi-skip-processing-stack-arg-dbg-value.mir b/llvm/test/CodeGen/AMDGPU/sgpr-spill-fi-skip-processing-stack-arg-dbg-value.mir
--- a/llvm/test/CodeGen/AMDGPU/sgpr-spill-fi-skip-processing-stack-arg-dbg-value.mir
+++ b/llvm/test/CodeGen/AMDGPU/sgpr-spill-fi-skip-processing-stack-arg-dbg-value.mir
@@ -1,4 +1,4 @@
-# RUN: llc -mtriple=amdgcn-amd-amdhsa -mcpu=gfx908 -amdgpu-spill-sgpr-to-vgpr=true -verify-machineinstrs -run-pass=si-lower-sgpr-spills -o - %s | FileCheck %s
+# RUN: llc -mtriple=amdgcn-amd-amdhsa -mcpu=gfx908 -amdgpu-spill-sgpr-to-vgpr=true -verify-machineinstrs -run-pass=si-lower-sgpr-spills,prologepilog -o - %s | FileCheck %s
# After handling the SGPR spill to VGPR in SILowerSGPRSpills pass, we replace the dead frame index in the DBG_VALUE instruction with reg 0.
# Skip looking for frame indices in the debug value instruction for incoming arguments passed via stack. The test would crash otherwise.
@@ -45,7 +45,7 @@
body: |
; CHECK-LABEL: name: test
; CHECK: bb.0:
- ; CHECK: DBG_VALUE
+ ; CHECK: DBG_VALUE $noreg, 0
bb.0:
renamable $sgpr10 = IMPLICIT_DEF
SI_SPILL_S32_SAVE killed $sgpr10, %stack.0, implicit $exec, implicit $sgpr96_sgpr97_sgpr98_sgpr99, implicit $sgpr32
diff --git a/llvm/test/CodeGen/AMDGPU/sgpr-spill-no-vgprs.ll b/llvm/test/CodeGen/AMDGPU/sgpr-spill-no-vgprs.ll
--- a/llvm/test/CodeGen/AMDGPU/sgpr-spill-no-vgprs.ll
+++ b/llvm/test/CodeGen/AMDGPU/sgpr-spill-no-vgprs.ll
@@ -9,17 +9,9 @@
; GCN: ; %bb.0:
; GCN-NEXT: s_add_u32 s0, s0, s15
; GCN-NEXT: s_addc_u32 s1, s1, 0
-; GCN-NEXT: ; implicit-def: $vgpr0
-; GCN-NEXT: ; implicit-def: $vgpr0
; GCN-NEXT: s_load_dword s4, s[8:9], 0x2
; GCN-NEXT: ;;#ASMSTART
; GCN-NEXT: ;;#ASMEND
-; GCN-NEXT: s_or_saveexec_b64 s[24:25], -1
-; GCN-NEXT: buffer_load_dword v1, off, s[0:3], 0 offset:8 ; 4-byte Folded Reload
-; GCN-NEXT: s_mov_b64 exec, s[24:25]
-; GCN-NEXT: s_or_saveexec_b64 s[24:25], -1
-; GCN-NEXT: buffer_load_dword v0, off, s[0:3], 0 offset:4 ; 4-byte Folded Reload
-; GCN-NEXT: s_mov_b64 exec, s[24:25]
; GCN-NEXT: ;;#ASMSTART
; GCN-NEXT: ;;#ASMEND
; GCN-NEXT: ;;#ASMSTART
@@ -31,179 +23,179 @@
; GCN-NEXT: ;;#ASMSTART
; GCN-NEXT: ; def s[8:23]
; GCN-NEXT: ;;#ASMEND
-; GCN-NEXT: s_waitcnt vmcnt(1)
-; GCN-NEXT: v_writelane_b32 v1, s8, 0
-; GCN-NEXT: v_writelane_b32 v1, s9, 1
-; GCN-NEXT: v_writelane_b32 v1, s10, 2
-; GCN-NEXT: v_writelane_b32 v1, s11, 3
-; GCN-NEXT: v_writelane_b32 v1, s12, 4
-; GCN-NEXT: v_writelane_b32 v1, s13, 5
-; GCN-NEXT: v_writelane_b32 v1, s14, 6
-; GCN-NEXT: v_writelane_b32 v1, s15, 7
-; GCN-NEXT: v_writelane_b32 v1, s16, 8
-; GCN-NEXT: v_writelane_b32 v1, s17, 9
-; GCN-NEXT: v_writelane_b32 v1, s18, 10
-; GCN-NEXT: v_writelane_b32 v1, s19, 11
-; GCN-NEXT: v_writelane_b32 v1, s20, 12
-; GCN-NEXT: v_writelane_b32 v1, s21, 13
-; GCN-NEXT: v_writelane_b32 v1, s22, 14
-; GCN-NEXT: v_writelane_b32 v1, s23, 15
+; GCN-NEXT: v_writelane_b32 v23, s8, 0
+; GCN-NEXT: v_writelane_b32 v23, s9, 1
+; GCN-NEXT: v_writelane_b32 v23, s10, 2
+; GCN-NEXT: v_writelane_b32 v23, s11, 3
+; GCN-NEXT: v_writelane_b32 v23, s12, 4
+; GCN-NEXT: v_writelane_b32 v23, s13, 5
+; GCN-NEXT: v_writelane_b32 v23, s14, 6
+; GCN-NEXT: v_writelane_b32 v23, s15, 7
+; GCN-NEXT: v_writelane_b32 v23, s16, 8
+; GCN-NEXT: v_writelane_b32 v23, s17, 9
+; GCN-NEXT: v_writelane_b32 v23, s18, 10
+; GCN-NEXT: v_writelane_b32 v23, s19, 11
+; GCN-NEXT: v_writelane_b32 v23, s20, 12
+; GCN-NEXT: v_writelane_b32 v23, s21, 13
+; GCN-NEXT: v_writelane_b32 v23, s22, 14
+; GCN-NEXT: v_writelane_b32 v23, s23, 15
; GCN-NEXT: ;;#ASMSTART
; GCN-NEXT: ; def s[8:23]
; GCN-NEXT: ;;#ASMEND
-; GCN-NEXT: v_writelane_b32 v1, s8, 16
-; GCN-NEXT: v_writelane_b32 v1, s9, 17
-; GCN-NEXT: v_writelane_b32 v1, s10, 18
-; GCN-NEXT: v_writelane_b32 v1, s11, 19
-; GCN-NEXT: v_writelane_b32 v1, s12, 20
-; GCN-NEXT: v_writelane_b32 v1, s13, 21
-; GCN-NEXT: v_writelane_b32 v1, s14, 22
-; GCN-NEXT: v_writelane_b32 v1, s15, 23
-; GCN-NEXT: v_writelane_b32 v1, s16, 24
-; GCN-NEXT: v_writelane_b32 v1, s17, 25
-; GCN-NEXT: v_writelane_b32 v1, s18, 26
-; GCN-NEXT: v_writelane_b32 v1, s19, 27
-; GCN-NEXT: v_writelane_b32 v1, s20, 28
-; GCN-NEXT: v_writelane_b32 v1, s21, 29
-; GCN-NEXT: v_writelane_b32 v1, s22, 30
-; GCN-NEXT: v_writelane_b32 v1, s23, 31
+; GCN-NEXT: v_writelane_b32 v23, s8, 16
+; GCN-NEXT: v_writelane_b32 v23, s9, 17
+; GCN-NEXT: v_writelane_b32 v23, s10, 18
+; GCN-NEXT: v_writelane_b32 v23, s11, 19
+; GCN-NEXT: v_writelane_b32 v23, s12, 20
+; GCN-NEXT: v_writelane_b32 v23, s13, 21
+; GCN-NEXT: v_writelane_b32 v23, s14, 22
+; GCN-NEXT: v_writelane_b32 v23, s15, 23
+; GCN-NEXT: v_writelane_b32 v23, s16, 24
+; GCN-NEXT: v_writelane_b32 v23, s17, 25
+; GCN-NEXT: v_writelane_b32 v23, s18, 26
+; GCN-NEXT: v_writelane_b32 v23, s19, 27
+; GCN-NEXT: v_writelane_b32 v23, s20, 28
+; GCN-NEXT: v_writelane_b32 v23, s21, 29
+; GCN-NEXT: v_writelane_b32 v23, s22, 30
+; GCN-NEXT: v_writelane_b32 v23, s23, 31
; GCN-NEXT: ;;#ASMSTART
; GCN-NEXT: ; def s[8:23]
; GCN-NEXT: ;;#ASMEND
-; GCN-NEXT: v_writelane_b32 v1, s8, 32
-; GCN-NEXT: v_writelane_b32 v1, s9, 33
-; GCN-NEXT: v_writelane_b32 v1, s10, 34
-; GCN-NEXT: v_writelane_b32 v1, s11, 35
-; GCN-NEXT: v_writelane_b32 v1, s12, 36
-; GCN-NEXT: v_writelane_b32 v1, s13, 37
-; GCN-NEXT: v_writelane_b32 v1, s14, 38
-; GCN-NEXT: v_writelane_b32 v1, s15, 39
-; GCN-NEXT: v_writelane_b32 v1, s16, 40
-; GCN-NEXT: v_writelane_b32 v1, s17, 41
-; GCN-NEXT: v_writelane_b32 v1, s18, 42
-; GCN-NEXT: v_writelane_b32 v1, s19, 43
-; GCN-NEXT: v_writelane_b32 v1, s20, 44
-; GCN-NEXT: v_writelane_b32 v1, s21, 45
-; GCN-NEXT: v_writelane_b32 v1, s22, 46
-; GCN-NEXT: v_writelane_b32 v1, s23, 47
+; GCN-NEXT: v_writelane_b32 v23, s8, 32
+; GCN-NEXT: v_writelane_b32 v23, s9, 33
+; GCN-NEXT: v_writelane_b32 v23, s10, 34
+; GCN-NEXT: v_writelane_b32 v23, s11, 35
+; GCN-NEXT: v_writelane_b32 v23, s12, 36
+; GCN-NEXT: v_writelane_b32 v23, s13, 37
+; GCN-NEXT: v_writelane_b32 v23, s14, 38
+; GCN-NEXT: v_writelane_b32 v23, s15, 39
+; GCN-NEXT: v_writelane_b32 v23, s16, 40
+; GCN-NEXT: v_writelane_b32 v23, s17, 41
+; GCN-NEXT: v_writelane_b32 v23, s18, 42
+; GCN-NEXT: v_writelane_b32 v23, s19, 43
+; GCN-NEXT: v_writelane_b32 v23, s20, 44
+; GCN-NEXT: v_writelane_b32 v23, s21, 45
+; GCN-NEXT: v_writelane_b32 v23, s22, 46
+; GCN-NEXT: v_writelane_b32 v23, s23, 47
; GCN-NEXT: ;;#ASMSTART
; GCN-NEXT: ; def s[8:23]
; GCN-NEXT: ;;#ASMEND
-; GCN-NEXT: v_writelane_b32 v1, s8, 48
-; GCN-NEXT: v_writelane_b32 v1, s9, 49
-; GCN-NEXT: v_writelane_b32 v1, s10, 50
-; GCN-NEXT: v_writelane_b32 v1, s11, 51
-; GCN-NEXT: v_writelane_b32 v1, s12, 52
-; GCN-NEXT: v_writelane_b32 v1, s13, 53
-; GCN-NEXT: v_writelane_b32 v1, s14, 54
-; GCN-NEXT: v_writelane_b32 v1, s15, 55
-; GCN-NEXT: v_writelane_b32 v1, s16, 56
-; GCN-NEXT: v_writelane_b32 v1, s17, 57
-; GCN-NEXT: v_writelane_b32 v1, s18, 58
-; GCN-NEXT: v_writelane_b32 v1, s19, 59
-; GCN-NEXT: v_writelane_b32 v1, s20, 60
-; GCN-NEXT: v_writelane_b32 v1, s21, 61
-; GCN-NEXT: v_writelane_b32 v1, s22, 62
-; GCN-NEXT: v_writelane_b32 v1, s23, 63
-; GCN-NEXT: s_or_saveexec_b64 s[24:25], -1
-; GCN-NEXT: buffer_store_dword v1, off, s[0:3], 0 offset:8 ; 4-byte Folded Spill
-; GCN-NEXT: s_mov_b64 exec, s[24:25]
+; GCN-NEXT: v_writelane_b32 v23, s8, 48
+; GCN-NEXT: v_writelane_b32 v23, s9, 49
+; GCN-NEXT: v_writelane_b32 v23, s10, 50
+; GCN-NEXT: v_writelane_b32 v23, s11, 51
+; GCN-NEXT: v_writelane_b32 v23, s12, 52
+; GCN-NEXT: v_writelane_b32 v23, s13, 53
+; GCN-NEXT: v_writelane_b32 v23, s14, 54
+; GCN-NEXT: v_writelane_b32 v23, s15, 55
+; GCN-NEXT: v_writelane_b32 v23, s16, 56
+; GCN-NEXT: v_writelane_b32 v23, s17, 57
+; GCN-NEXT: v_writelane_b32 v23, s18, 58
+; GCN-NEXT: v_writelane_b32 v23, s19, 59
+; GCN-NEXT: v_writelane_b32 v23, s20, 60
+; GCN-NEXT: v_writelane_b32 v23, s21, 61
+; GCN-NEXT: v_writelane_b32 v23, s22, 62
+; GCN-NEXT: v_writelane_b32 v23, s23, 63
; GCN-NEXT: ;;#ASMSTART
; GCN-NEXT: ; def s[6:7]
; GCN-NEXT: ;;#ASMEND
-; GCN-NEXT: s_waitcnt vmcnt(1)
+; GCN-NEXT: s_mov_b64 s[8:9], exec
+; GCN-NEXT: s_mov_b64 exec, 3
+; GCN-NEXT: buffer_store_dword v0, off, s[0:3], 0
; GCN-NEXT: v_writelane_b32 v0, s6, 0
; GCN-NEXT: v_writelane_b32 v0, s7, 1
-; GCN-NEXT: s_or_saveexec_b64 s[24:25], -1
; GCN-NEXT: buffer_store_dword v0, off, s[0:3], 0 offset:4 ; 4-byte Folded Spill
-; GCN-NEXT: s_mov_b64 exec, s[24:25]
+; GCN-NEXT: buffer_load_dword v0, off, s[0:3], 0
+; GCN-NEXT: s_waitcnt vmcnt(0)
+; GCN-NEXT: s_mov_b64 exec, s[8:9]
; GCN-NEXT: s_mov_b32 s5, 0
; GCN-NEXT: s_waitcnt lgkmcnt(0)
; GCN-NEXT: s_cmp_lg_u32 s4, s5
; GCN-NEXT: s_cbranch_scc1 .LBB0_2
; GCN-NEXT: ; %bb.1: ; %bb0
-; GCN-NEXT: s_or_saveexec_b64 s[24:25], -1
-; GCN-NEXT: buffer_load_dword v0, off, s[0:3], 0 offset:4 ; 4-byte Folded Reload
-; GCN-NEXT: s_mov_b64 exec, s[24:25]
-; GCN-NEXT: s_or_saveexec_b64 s[24:25], -1
-; GCN-NEXT: buffer_load_dword v1, off, s[0:3], 0 offset:8 ; 4-byte Folded Reload
-; GCN-NEXT: s_mov_b64 exec, s[24:25]
-; GCN-NEXT: s_waitcnt vmcnt(0)
-; GCN-NEXT: v_readlane_b32 s4, v1, 0
-; GCN-NEXT: v_readlane_b32 s5, v1, 1
-; GCN-NEXT: v_readlane_b32 s6, v1, 2
-; GCN-NEXT: v_readlane_b32 s7, v1, 3
-; GCN-NEXT: v_readlane_b32 s8, v1, 4
-; GCN-NEXT: v_readlane_b32 s9, v1, 5
-; GCN-NEXT: v_readlane_b32 s10, v1, 6
-; GCN-NEXT: v_readlane_b32 s11, v1, 7
-; GCN-NEXT: v_readlane_b32 s12, v1, 8
-; GCN-NEXT: v_readlane_b32 s13, v1, 9
-; GCN-NEXT: v_readlane_b32 s14, v1, 10
-; GCN-NEXT: v_readlane_b32 s15, v1, 11
-; GCN-NEXT: v_readlane_b32 s16, v1, 12
-; GCN-NEXT: v_readlane_b32 s17, v1, 13
-; GCN-NEXT: v_readlane_b32 s18, v1, 14
-; GCN-NEXT: v_readlane_b32 s19, v1, 15
+; GCN-NEXT: v_readlane_b32 s4, v23, 0
+; GCN-NEXT: v_readlane_b32 s5, v23, 1
+; GCN-NEXT: v_readlane_b32 s6, v23, 2
+; GCN-NEXT: v_readlane_b32 s7, v23, 3
+; GCN-NEXT: v_readlane_b32 s8, v23, 4
+; GCN-NEXT: v_readlane_b32 s9, v23, 5
+; GCN-NEXT: v_readlane_b32 s10, v23, 6
+; GCN-NEXT: v_readlane_b32 s11, v23, 7
+; GCN-NEXT: v_readlane_b32 s12, v23, 8
+; GCN-NEXT: v_readlane_b32 s13, v23, 9
+; GCN-NEXT: v_readlane_b32 s14, v23, 10
+; GCN-NEXT: v_readlane_b32 s15, v23, 11
+; GCN-NEXT: v_readlane_b32 s16, v23, 12
+; GCN-NEXT: v_readlane_b32 s17, v23, 13
+; GCN-NEXT: v_readlane_b32 s18, v23, 14
+; GCN-NEXT: v_readlane_b32 s19, v23, 15
; GCN-NEXT: ;;#ASMSTART
; GCN-NEXT: ; use s[4:19]
; GCN-NEXT: ;;#ASMEND
-; GCN-NEXT: v_readlane_b32 s4, v1, 16
-; GCN-NEXT: v_readlane_b32 s5, v1, 17
-; GCN-NEXT: v_readlane_b32 s6, v1, 18
-; GCN-NEXT: v_readlane_b32 s7, v1, 19
-; GCN-NEXT: v_readlane_b32 s8, v1, 20
-; GCN-NEXT: v_readlane_b32 s9, v1, 21
-; GCN-NEXT: v_readlane_b32 s10, v1, 22
-; GCN-NEXT: v_readlane_b32 s11, v1, 23
-; GCN-NEXT: v_readlane_b32 s12, v1, 24
-; GCN-NEXT: v_readlane_b32 s13, v1, 25
-; GCN-NEXT: v_readlane_b32 s14, v1, 26
-; GCN-NEXT: v_readlane_b32 s15, v1, 27
-; GCN-NEXT: v_readlane_b32 s16, v1, 28
-; GCN-NEXT: v_readlane_b32 s17, v1, 29
-; GCN-NEXT: v_readlane_b32 s18, v1, 30
-; GCN-NEXT: v_readlane_b32 s19, v1, 31
+; GCN-NEXT: v_readlane_b32 s4, v23, 16
+; GCN-NEXT: v_readlane_b32 s5, v23, 17
+; GCN-NEXT: v_readlane_b32 s6, v23, 18
+; GCN-NEXT: v_readlane_b32 s7, v23, 19
+; GCN-NEXT: v_readlane_b32 s8, v23, 20
+; GCN-NEXT: v_readlane_b32 s9, v23, 21
+; GCN-NEXT: v_readlane_b32 s10, v23, 22
+; GCN-NEXT: v_readlane_b32 s11, v23, 23
+; GCN-NEXT: v_readlane_b32 s12, v23, 24
+; GCN-NEXT: v_readlane_b32 s13, v23, 25
+; GCN-NEXT: v_readlane_b32 s14, v23, 26
+; GCN-NEXT: v_readlane_b32 s15, v23, 27
+; GCN-NEXT: v_readlane_b32 s16, v23, 28
+; GCN-NEXT: v_readlane_b32 s17, v23, 29
+; GCN-NEXT: v_readlane_b32 s18, v23, 30
+; GCN-NEXT: v_readlane_b32 s19, v23, 31
; GCN-NEXT: ;;#ASMSTART
; GCN-NEXT: ; use s[4:19]
; GCN-NEXT: ;;#ASMEND
-; GCN-NEXT: v_readlane_b32 s4, v1, 32
-; GCN-NEXT: v_readlane_b32 s5, v1, 33
-; GCN-NEXT: v_readlane_b32 s6, v1, 34
-; GCN-NEXT: v_readlane_b32 s7, v1, 35
-; GCN-NEXT: v_readlane_b32 s8, v1, 36
-; GCN-NEXT: v_readlane_b32 s9, v1, 37
-; GCN-NEXT: v_readlane_b32 s10, v1, 38
-; GCN-NEXT: v_readlane_b32 s11, v1, 39
-; GCN-NEXT: v_readlane_b32 s12, v1, 40
-; GCN-NEXT: v_readlane_b32 s13, v1, 41
-; GCN-NEXT: v_readlane_b32 s14, v1, 42
-; GCN-NEXT: v_readlane_b32 s15, v1, 43
-; GCN-NEXT: v_readlane_b32 s16, v1, 44
-; GCN-NEXT: v_readlane_b32 s17, v1, 45
-; GCN-NEXT: v_readlane_b32 s18, v1, 46
-; GCN-NEXT: v_readlane_b32 s19, v1, 47
+; GCN-NEXT: v_readlane_b32 s4, v23, 32
+; GCN-NEXT: v_readlane_b32 s5, v23, 33
+; GCN-NEXT: v_readlane_b32 s6, v23, 34
+; GCN-NEXT: v_readlane_b32 s7, v23, 35
+; GCN-NEXT: v_readlane_b32 s8, v23, 36
+; GCN-NEXT: v_readlane_b32 s9, v23, 37
+; GCN-NEXT: v_readlane_b32 s10, v23, 38
+; GCN-NEXT: v_readlane_b32 s11, v23, 39
+; GCN-NEXT: v_readlane_b32 s12, v23, 40
+; GCN-NEXT: v_readlane_b32 s13, v23, 41
+; GCN-NEXT: v_readlane_b32 s14, v23, 42
+; GCN-NEXT: v_readlane_b32 s15, v23, 43
+; GCN-NEXT: v_readlane_b32 s16, v23, 44
+; GCN-NEXT: v_readlane_b32 s17, v23, 45
+; GCN-NEXT: v_readlane_b32 s18, v23, 46
+; GCN-NEXT: v_readlane_b32 s19, v23, 47
; GCN-NEXT: ;;#ASMSTART
; GCN-NEXT: ; use s[4:19]
; GCN-NEXT: ;;#ASMEND
-; GCN-NEXT: v_readlane_b32 s8, v1, 48
-; GCN-NEXT: v_readlane_b32 s9, v1, 49
-; GCN-NEXT: v_readlane_b32 s10, v1, 50
-; GCN-NEXT: v_readlane_b32 s11, v1, 51
-; GCN-NEXT: v_readlane_b32 s12, v1, 52
-; GCN-NEXT: v_readlane_b32 s13, v1, 53
-; GCN-NEXT: v_readlane_b32 s14, v1, 54
-; GCN-NEXT: v_readlane_b32 s15, v1, 55
-; GCN-NEXT: v_readlane_b32 s16, v1, 56
-; GCN-NEXT: v_readlane_b32 s17, v1, 57
-; GCN-NEXT: v_readlane_b32 s18, v1, 58
-; GCN-NEXT: v_readlane_b32 s19, v1, 59
-; GCN-NEXT: v_readlane_b32 s20, v1, 60
-; GCN-NEXT: v_readlane_b32 s21, v1, 61
-; GCN-NEXT: v_readlane_b32 s22, v1, 62
-; GCN-NEXT: v_readlane_b32 s23, v1, 63
+; GCN-NEXT: v_readlane_b32 s8, v23, 48
+; GCN-NEXT: v_readlane_b32 s9, v23, 49
+; GCN-NEXT: v_readlane_b32 s10, v23, 50
+; GCN-NEXT: v_readlane_b32 s11, v23, 51
+; GCN-NEXT: v_readlane_b32 s12, v23, 52
+; GCN-NEXT: v_readlane_b32 s13, v23, 53
+; GCN-NEXT: v_readlane_b32 s14, v23, 54
+; GCN-NEXT: v_readlane_b32 s15, v23, 55
+; GCN-NEXT: v_readlane_b32 s16, v23, 56
+; GCN-NEXT: v_readlane_b32 s17, v23, 57
+; GCN-NEXT: v_readlane_b32 s18, v23, 58
+; GCN-NEXT: v_readlane_b32 s19, v23, 59
+; GCN-NEXT: v_readlane_b32 s20, v23, 60
+; GCN-NEXT: v_readlane_b32 s21, v23, 61
+; GCN-NEXT: v_readlane_b32 s22, v23, 62
+; GCN-NEXT: v_readlane_b32 s23, v23, 63
+; GCN-NEXT: s_mov_b64 s[6:7], exec
+; GCN-NEXT: s_mov_b64 exec, 3
+; GCN-NEXT: buffer_store_dword v0, off, s[0:3], 0
+; GCN-NEXT: buffer_load_dword v0, off, s[0:3], 0 offset:4 ; 4-byte Folded Reload
+; GCN-NEXT: s_waitcnt vmcnt(0)
; GCN-NEXT: v_readlane_b32 s4, v0, 0
; GCN-NEXT: v_readlane_b32 s5, v0, 1
+; GCN-NEXT: buffer_load_dword v0, off, s[0:3], 0
+; GCN-NEXT: s_waitcnt vmcnt(0)
+; GCN-NEXT: s_mov_b64 exec, s[6:7]
; GCN-NEXT: ;;#ASMSTART
; GCN-NEXT: ; use s[8:23]
; GCN-NEXT: ;;#ASMEND
@@ -211,14 +203,6 @@
; GCN-NEXT: ; use s[4:5]
; GCN-NEXT: ;;#ASMEND
; GCN-NEXT: .LBB0_2: ; %ret
-; GCN-NEXT: s_or_saveexec_b64 s[24:25], -1
-; GCN-NEXT: buffer_load_dword v0, off, s[0:3], 0 offset:4 ; 4-byte Folded Reload
-; GCN-NEXT: s_mov_b64 exec, s[24:25]
-; GCN-NEXT: s_or_saveexec_b64 s[24:25], -1
-; GCN-NEXT: buffer_load_dword v1, off, s[0:3], 0 offset:8 ; 4-byte Folded Reload
-; GCN-NEXT: s_mov_b64 exec, s[24:25]
-; GCN-NEXT: ; kill: killed $vgpr1
-; GCN-NEXT: ; kill: killed $vgpr0
; GCN-NEXT: s_endpgm
call void asm sideeffect "", "~{v[0:7]}" () #0
call void asm sideeffect "", "~{v[8:15]}" () #0
diff --git a/llvm/test/CodeGen/AMDGPU/sgpr-spill-partially-undef.mir b/llvm/test/CodeGen/AMDGPU/sgpr-spill-partially-undef.mir
--- a/llvm/test/CodeGen/AMDGPU/sgpr-spill-partially-undef.mir
+++ b/llvm/test/CodeGen/AMDGPU/sgpr-spill-partially-undef.mir
@@ -20,11 +20,10 @@
liveins: $sgpr4
; CHECK-LABEL: name: sgpr_spill_s64_undef_high32
- ; CHECK: liveins: $sgpr4
+ ; CHECK: liveins: $sgpr4, $vgpr0
; CHECK-NEXT: {{ $}}
- ; CHECK-NEXT: [[DEF:%[0-9]+]]:vgpr_32 = IMPLICIT_DEF
- ; CHECK-NEXT: [[V_WRITELANE_B32_:%[0-9]+]]:vgpr_32 = V_WRITELANE_B32 $sgpr4, 0, [[V_WRITELANE_B32_]], implicit-def $sgpr4_sgpr5, implicit $sgpr4_sgpr5
- ; CHECK-NEXT: [[V_WRITELANE_B32_1:%[0-9]+]]:vgpr_32 = V_WRITELANE_B32 $sgpr5, 1, [[V_WRITELANE_B32_1]], implicit $sgpr4_sgpr5
+ ; CHECK-NEXT: $vgpr0 = V_WRITELANE_B32 $sgpr4, 0, $vgpr0, implicit-def $sgpr4_sgpr5, implicit $sgpr4_sgpr5
+ ; CHECK-NEXT: $vgpr0 = V_WRITELANE_B32 $sgpr5, 1, $vgpr0, implicit $sgpr4_sgpr5
SI_SPILL_S64_SAVE renamable $sgpr4_sgpr5, %stack.0, implicit $exec, implicit $sgpr96_sgpr97_sgpr98_sgpr99, implicit $sgpr32 :: (store (s64) into %stack.0, align 4, addrspace 5)
...
@@ -46,11 +45,10 @@
liveins: $sgpr5
; CHECK-LABEL: name: sgpr_spill_s64_undef_low32
- ; CHECK: liveins: $sgpr5
+ ; CHECK: liveins: $sgpr5, $vgpr0
; CHECK-NEXT: {{ $}}
- ; CHECK-NEXT: [[DEF:%[0-9]+]]:vgpr_32 = IMPLICIT_DEF
- ; CHECK-NEXT: [[V_WRITELANE_B32_:%[0-9]+]]:vgpr_32 = V_WRITELANE_B32 $sgpr4, 0, [[V_WRITELANE_B32_]], implicit-def $sgpr4_sgpr5, implicit $sgpr4_sgpr5
- ; CHECK-NEXT: [[V_WRITELANE_B32_1:%[0-9]+]]:vgpr_32 = V_WRITELANE_B32 $sgpr5, 1, [[V_WRITELANE_B32_1]], implicit $sgpr4_sgpr5
+ ; CHECK-NEXT: $vgpr0 = V_WRITELANE_B32 $sgpr4, 0, $vgpr0, implicit-def $sgpr4_sgpr5, implicit $sgpr4_sgpr5
+ ; CHECK-NEXT: $vgpr0 = V_WRITELANE_B32 $sgpr5, 1, $vgpr0, implicit $sgpr4_sgpr5
SI_SPILL_S64_SAVE renamable $sgpr4_sgpr5, %stack.0, implicit $exec, implicit $sgpr96_sgpr97_sgpr98_sgpr99, implicit $sgpr32 :: (store (s64) into %stack.0, align 4, addrspace 5)
...
diff --git a/llvm/test/CodeGen/AMDGPU/sgpr-spill-update-only-slot-indexes.ll b/llvm/test/CodeGen/AMDGPU/sgpr-spill-update-only-slot-indexes.ll
--- a/llvm/test/CodeGen/AMDGPU/sgpr-spill-update-only-slot-indexes.ll
+++ b/llvm/test/CodeGen/AMDGPU/sgpr-spill-update-only-slot-indexes.ll
@@ -13,40 +13,32 @@
; GCN-NEXT: s_mov_b32 s37, SCRATCH_RSRC_DWORD1
; GCN-NEXT: s_mov_b32 s38, -1
; GCN-NEXT: s_mov_b32 s39, 0xe00000
-; GCN-NEXT: ; implicit-def: $vgpr3
+; GCN-NEXT: v_writelane_b32 v40, s4, 0
; GCN-NEXT: s_add_u32 s36, s36, s11
-; GCN-NEXT: v_writelane_b32 v3, s4, 0
-; GCN-NEXT: s_movk_i32 s32, 0x400
+; GCN-NEXT: v_writelane_b32 v40, s5, 1
; GCN-NEXT: s_addc_u32 s37, s37, 0
-; GCN-NEXT: s_mov_b32 s14, s10
+; GCN-NEXT: s_mov_b64 s[4:5], s[0:1]
+; GCN-NEXT: v_readlane_b32 s0, v40, 0
; GCN-NEXT: s_mov_b32 s13, s9
; GCN-NEXT: s_mov_b32 s12, s8
-; GCN-NEXT: s_mov_b64 s[10:11], s[6:7]
-; GCN-NEXT: v_writelane_b32 v3, s5, 1
-; GCN-NEXT: s_or_saveexec_b64 s[34:35], -1
-; GCN-NEXT: buffer_store_dword v3, off, s[36:39], 0 offset:4 ; 4-byte Folded Spill
-; GCN-NEXT: s_mov_b64 exec, s[34:35]
-; GCN-NEXT: s_mov_b64 s[4:5], s[0:1]
-; GCN-NEXT: v_readlane_b32 s0, v3, 0
-; GCN-NEXT: v_readlane_b32 s1, v3, 1
+; GCN-NEXT: v_readlane_b32 s1, v40, 1
; GCN-NEXT: s_add_u32 s8, s0, 36
; GCN-NEXT: s_addc_u32 s9, s1, 0
; GCN-NEXT: s_getpc_b64 s[0:1]
; GCN-NEXT: s_add_u32 s0, s0, foo@gotpcrel32@lo+4
; GCN-NEXT: s_addc_u32 s1, s1, foo@gotpcrel32@hi+12
; GCN-NEXT: s_load_dwordx2 s[16:17], s[0:1], 0x0
+; GCN-NEXT: s_mov_b32 s14, s10
+; GCN-NEXT: s_mov_b64 s[10:11], s[6:7]
; GCN-NEXT: s_mov_b64 s[6:7], s[2:3]
; GCN-NEXT: v_lshlrev_b32_e32 v2, 20, v2
; GCN-NEXT: v_lshlrev_b32_e32 v1, 10, v1
; GCN-NEXT: s_mov_b64 s[0:1], s[36:37]
; GCN-NEXT: v_or3_b32 v31, v0, v1, v2
; GCN-NEXT: s_mov_b64 s[2:3], s[38:39]
+; GCN-NEXT: s_mov_b32 s32, 0
; GCN-NEXT: s_waitcnt lgkmcnt(0)
; GCN-NEXT: s_swappc_b64 s[30:31], s[16:17]
-; GCN-NEXT: s_or_saveexec_b64 s[34:35], -1
-; GCN-NEXT: buffer_load_dword v0, off, s[36:39], 0 offset:4 ; 4-byte Folded Reload
-; GCN-NEXT: s_mov_b64 exec, s[34:35]
-; GCN-NEXT: ; kill: killed $vgpr0
; GCN-NEXT: s_endpgm
call void @foo()
ret void
diff --git a/llvm/test/CodeGen/AMDGPU/sgpr-spill-vmem-large-frame.mir b/llvm/test/CodeGen/AMDGPU/sgpr-spill-vmem-large-frame.mir
--- a/llvm/test/CodeGen/AMDGPU/sgpr-spill-vmem-large-frame.mir
+++ b/llvm/test/CodeGen/AMDGPU/sgpr-spill-vmem-large-frame.mir
@@ -1,5 +1,5 @@
# NOTE: Assertions have been autogenerated by utils/update_mir_test_checks.py
-# RUN: llc -mtriple=amdgcn-amd-amdhsa -mcpu=gfx908 -amdgpu-spill-sgpr-to-vgpr=false -verify-machineinstrs -start-before=si-lower-sgpr-spills -stop-after=prologepilog -o - %s | FileCheck %s
+# RUN: llc -mtriple=amdgcn-amd-amdhsa -mcpu=gfx908 -amdgpu-spill-sgpr-to-vgpr=false -verify-machineinstrs -run-pass=si-lower-sgpr-spills,prologepilog -o - %s | FileCheck %s
# Check that we allocate 2 emergency stack slots if we're spilling
# SGPRs to memory and potentially have an offset larger than fits in
@@ -29,7 +29,7 @@
; CHECK-NEXT: $sgpr4_sgpr5 = S_MOV_B64 $exec
; CHECK-NEXT: $exec = S_MOV_B64 1, implicit-def $vgpr1
; CHECK-NEXT: BUFFER_STORE_DWORD_OFFSET killed $vgpr1, $sgpr0_sgpr1_sgpr2_sgpr3, $sgpr32, 0, 0, 0, implicit $exec :: (store (s32) into %stack.2, addrspace 5)
- ; CHECK-NEXT: $vgpr1 = V_WRITELANE_B32 $sgpr10, 0, undef $vgpr1
+ ; CHECK-NEXT: $vgpr1 = V_WRITELANE_B32 killed $sgpr10, 0, undef $vgpr1
; CHECK-NEXT: BUFFER_STORE_DWORD_OFFSET killed $vgpr1, $sgpr0_sgpr1_sgpr2_sgpr3, $sgpr32, 8, 0, 0, implicit $exec :: (store (s32) into %stack.0, addrspace 5)
; CHECK-NEXT: $vgpr1 = BUFFER_LOAD_DWORD_OFFSET $sgpr0_sgpr1_sgpr2_sgpr3, $sgpr32, 0, 0, 0, implicit $exec :: (load (s32) from %stack.2, addrspace 5)
; CHECK-NEXT: $exec = S_MOV_B64 killed $sgpr4_sgpr5, implicit killed $vgpr1
diff --git a/llvm/test/CodeGen/AMDGPU/sgpr-spills-split-regalloc.ll b/llvm/test/CodeGen/AMDGPU/sgpr-spills-split-regalloc.ll
--- a/llvm/test/CodeGen/AMDGPU/sgpr-spills-split-regalloc.ll
+++ b/llvm/test/CodeGen/AMDGPU/sgpr-spills-split-regalloc.ll
@@ -16,10 +16,10 @@
; GCN-LABEL: spill_sgpr_with_no_lower_vgpr_available:
; GCN: ; %bb.0:
; GCN-NEXT: s_waitcnt vmcnt(0) expcnt(0) lgkmcnt(0)
-; GCN-NEXT: s_mov_b32 s18, s33
+; GCN-NEXT: s_mov_b32 s24, s33
; GCN-NEXT: s_mov_b32 s33, s32
; GCN-NEXT: s_or_saveexec_b64 s[16:17], -1
-; GCN-NEXT: buffer_store_dword v255, off, s[0:3], s33 offset:448 ; 4-byte Folded Spill
+; GCN-NEXT: buffer_store_dword v255, off, s[0:3], s33 offset:452 ; 4-byte Folded Spill
; GCN-NEXT: s_mov_b64 exec, s[16:17]
; GCN-NEXT: s_add_i32 s32, s32, 0x7400
; GCN-NEXT: buffer_store_dword v40, off, s[0:3], s33 offset:440 ; 4-byte Folded Spill
@@ -135,13 +135,13 @@
; GCN-NEXT: buffer_store_dword v254, off, s[0:3], s33 ; 4-byte Folded Spill
; GCN-NEXT: v_writelane_b32 v255, s30, 0
; GCN-NEXT: v_writelane_b32 v255, s31, 1
-; GCN-NEXT: buffer_store_dword v31, off, s[0:3], s33 offset:452 ; 4-byte Folded Spill
+; GCN-NEXT: buffer_store_dword v31, off, s[0:3], s33 offset:448 ; 4-byte Folded Spill
; GCN-NEXT: v_mov_b32_e32 v0, 0
; GCN-NEXT: buffer_store_dword v0, off, s[0:3], s33 offset:444
; GCN-NEXT: s_waitcnt vmcnt(0)
; GCN-NEXT: ;;#ASMSTART
; GCN-NEXT: ;;#ASMEND
-; GCN-NEXT: buffer_load_dword v31, off, s[0:3], s33 offset:452 ; 4-byte Folded Reload
+; GCN-NEXT: buffer_load_dword v31, off, s[0:3], s33 offset:448 ; 4-byte Folded Reload
; GCN-NEXT: s_getpc_b64 s[16:17]
; GCN-NEXT: s_add_u32 s16, s16, child_function@gotpcrel32@lo+4
; GCN-NEXT: s_addc_u32 s17, s17, child_function@gotpcrel32@hi+12
@@ -266,10 +266,10 @@
; GCN-NEXT: buffer_load_dword v41, off, s[0:3], s33 offset:436 ; 4-byte Folded Reload
; GCN-NEXT: buffer_load_dword v40, off, s[0:3], s33 offset:440 ; 4-byte Folded Reload
; GCN-NEXT: s_or_saveexec_b64 s[4:5], -1
-; GCN-NEXT: buffer_load_dword v255, off, s[0:3], s33 offset:448 ; 4-byte Folded Reload
+; GCN-NEXT: buffer_load_dword v255, off, s[0:3], s33 offset:452 ; 4-byte Folded Reload
; GCN-NEXT: s_mov_b64 exec, s[4:5]
; GCN-NEXT: s_add_i32 s32, s32, 0xffff8c00
-; GCN-NEXT: s_mov_b32 s33, s18
+; GCN-NEXT: s_mov_b32 s33, s24
; GCN-NEXT: s_waitcnt vmcnt(0)
; GCN-NEXT: s_setpc_b64 s[30:31]
%alloca = alloca i32, align 4, addrspace(5)
@@ -310,10 +310,10 @@
; GCN-LABEL: spill_to_lowest_available_vgpr:
; GCN: ; %bb.0:
; GCN-NEXT: s_waitcnt vmcnt(0) expcnt(0) lgkmcnt(0)
-; GCN-NEXT: s_mov_b32 s18, s33
+; GCN-NEXT: s_mov_b32 s24, s33
; GCN-NEXT: s_mov_b32 s33, s32
; GCN-NEXT: s_or_saveexec_b64 s[16:17], -1
-; GCN-NEXT: buffer_store_dword v254, off, s[0:3], s33 offset:444 ; 4-byte Folded Spill
+; GCN-NEXT: buffer_store_dword v254, off, s[0:3], s33 offset:448 ; 4-byte Folded Spill
; GCN-NEXT: s_mov_b64 exec, s[16:17]
; GCN-NEXT: s_add_i32 s32, s32, 0x7400
; GCN-NEXT: buffer_store_dword v40, off, s[0:3], s33 offset:436 ; 4-byte Folded Spill
@@ -428,13 +428,13 @@
; GCN-NEXT: buffer_store_dword v253, off, s[0:3], s33 ; 4-byte Folded Spill
; GCN-NEXT: v_writelane_b32 v254, s30, 0
; GCN-NEXT: v_writelane_b32 v254, s31, 1
-; GCN-NEXT: buffer_store_dword v31, off, s[0:3], s33 offset:448 ; 4-byte Folded Spill
+; GCN-NEXT: buffer_store_dword v31, off, s[0:3], s33 offset:444 ; 4-byte Folded Spill
; GCN-NEXT: v_mov_b32_e32 v0, 0
; GCN-NEXT: buffer_store_dword v0, off, s[0:3], s33 offset:440
; GCN-NEXT: s_waitcnt vmcnt(0)
; GCN-NEXT: ;;#ASMSTART
; GCN-NEXT: ;;#ASMEND
-; GCN-NEXT: buffer_load_dword v31, off, s[0:3], s33 offset:448 ; 4-byte Folded Reload
+; GCN-NEXT: buffer_load_dword v31, off, s[0:3], s33 offset:444 ; 4-byte Folded Reload
; GCN-NEXT: s_getpc_b64 s[16:17]
; GCN-NEXT: s_add_u32 s16, s16, child_function@gotpcrel32@lo+4
; GCN-NEXT: s_addc_u32 s17, s17, child_function@gotpcrel32@hi+12
@@ -558,10 +558,10 @@
; GCN-NEXT: buffer_load_dword v41, off, s[0:3], s33 offset:432 ; 4-byte Folded Reload
; GCN-NEXT: buffer_load_dword v40, off, s[0:3], s33 offset:436 ; 4-byte Folded Reload
; GCN-NEXT: s_or_saveexec_b64 s[4:5], -1
-; GCN-NEXT: buffer_load_dword v254, off, s[0:3], s33 offset:444 ; 4-byte Folded Reload
+; GCN-NEXT: buffer_load_dword v254, off, s[0:3], s33 offset:448 ; 4-byte Folded Reload
; GCN-NEXT: s_mov_b64 exec, s[4:5]
; GCN-NEXT: s_add_i32 s32, s32, 0xffff8c00
-; GCN-NEXT: s_mov_b32 s33, s18
+; GCN-NEXT: s_mov_b32 s33, s24
; GCN-NEXT: s_waitcnt vmcnt(0)
; GCN-NEXT: s_setpc_b64 s[30:31]
%alloca = alloca i32, align 4, addrspace(5)
@@ -602,8 +602,8 @@
; GCN-LABEL: spill_sgpr_with_sgpr_uses:
; GCN: ; %bb.0:
; GCN-NEXT: s_waitcnt vmcnt(0) expcnt(0) lgkmcnt(0)
-; GCN-NEXT: s_xor_saveexec_b64 s[4:5], -1
-; GCN-NEXT: buffer_store_dword v0, off, s[0:3], s32 offset:448 ; 4-byte Folded Spill
+; GCN-NEXT: s_or_saveexec_b64 s[4:5], -1
+; GCN-NEXT: buffer_store_dword v254, off, s[0:3], s32 offset:444 ; 4-byte Folded Spill
; GCN-NEXT: s_mov_b64 exec, s[4:5]
; GCN-NEXT: buffer_store_dword v40, off, s[0:3], s32 offset:436 ; 4-byte Folded Spill
; GCN-NEXT: buffer_store_dword v41, off, s[0:3], s32 offset:432 ; 4-byte Folded Spill
@@ -715,38 +715,22 @@
; GCN-NEXT: buffer_store_dword v251, off, s[0:3], s32 offset:8 ; 4-byte Folded Spill
; GCN-NEXT: buffer_store_dword v252, off, s[0:3], s32 offset:4 ; 4-byte Folded Spill
; GCN-NEXT: buffer_store_dword v253, off, s[0:3], s32 ; 4-byte Folded Spill
-; GCN-NEXT: ; implicit-def: $vgpr0
; GCN-NEXT: v_mov_b32_e32 v0, 0
; GCN-NEXT: buffer_store_dword v0, off, s[0:3], s32 offset:440
; GCN-NEXT: s_waitcnt vmcnt(0)
; GCN-NEXT: ;;#ASMSTART
; GCN-NEXT: ;;#ASMEND
-; GCN-NEXT: s_or_saveexec_b64 s[8:9], -1
-; GCN-NEXT: buffer_load_dword v0, off, s[0:3], s32 offset:444 ; 4-byte Folded Reload
-; GCN-NEXT: s_mov_b64 exec, s[8:9]
; GCN-NEXT: ;;#ASMSTART
; GCN-NEXT: ; def s4
; GCN-NEXT: ;;#ASMEND
-; GCN-NEXT: s_waitcnt vmcnt(0)
-; GCN-NEXT: v_writelane_b32 v0, s4, 0
-; GCN-NEXT: s_or_saveexec_b64 s[8:9], -1
-; GCN-NEXT: buffer_store_dword v0, off, s[0:3], s32 offset:444 ; 4-byte Folded Spill
-; GCN-NEXT: s_mov_b64 exec, s[8:9]
+; GCN-NEXT: v_writelane_b32 v254, s4, 0
; GCN-NEXT: s_cbranch_scc1 .LBB3_2
; GCN-NEXT: ; %bb.1: ; %bb0
-; GCN-NEXT: s_or_saveexec_b64 s[8:9], -1
-; GCN-NEXT: buffer_load_dword v0, off, s[0:3], s32 offset:444 ; 4-byte Folded Reload
-; GCN-NEXT: s_mov_b64 exec, s[8:9]
-; GCN-NEXT: s_waitcnt vmcnt(0)
-; GCN-NEXT: v_readlane_b32 s4, v0, 0
+; GCN-NEXT: v_readlane_b32 s4, v254, 0
; GCN-NEXT: ;;#ASMSTART
; GCN-NEXT: ; use s4
; GCN-NEXT: ;;#ASMEND
; GCN-NEXT: .LBB3_2: ; %ret
-; GCN-NEXT: s_or_saveexec_b64 s[8:9], -1
-; GCN-NEXT: buffer_load_dword v0, off, s[0:3], s32 offset:444 ; 4-byte Folded Reload
-; GCN-NEXT: s_mov_b64 exec, s[8:9]
-; GCN-NEXT: ; kill: killed $vgpr0
; GCN-NEXT: buffer_load_dword v253, off, s[0:3], s32 ; 4-byte Folded Reload
; GCN-NEXT: buffer_load_dword v252, off, s[0:3], s32 offset:4 ; 4-byte Folded Reload
; GCN-NEXT: buffer_load_dword v251, off, s[0:3], s32 offset:8 ; 4-byte Folded Reload
@@ -857,8 +841,8 @@
; GCN-NEXT: buffer_load_dword v42, off, s[0:3], s32 offset:428 ; 4-byte Folded Reload
; GCN-NEXT: buffer_load_dword v41, off, s[0:3], s32 offset:432 ; 4-byte Folded Reload
; GCN-NEXT: buffer_load_dword v40, off, s[0:3], s32 offset:436 ; 4-byte Folded Reload
-; GCN-NEXT: s_xor_saveexec_b64 s[4:5], -1
-; GCN-NEXT: buffer_load_dword v0, off, s[0:3], s32 offset:448 ; 4-byte Folded Reload
+; GCN-NEXT: s_or_saveexec_b64 s[4:5], -1
+; GCN-NEXT: buffer_load_dword v254, off, s[0:3], s32 offset:444 ; 4-byte Folded Reload
; GCN-NEXT: s_mov_b64 exec, s[4:5]
; GCN-NEXT: s_waitcnt vmcnt(0)
; GCN-NEXT: s_setpc_b64 s[30:31]
@@ -1183,7 +1167,7 @@
; GCN: ; %bb.0:
; GCN-NEXT: s_waitcnt vmcnt(0) expcnt(0) lgkmcnt(0)
; GCN-NEXT: s_xor_saveexec_b64 s[4:5], -1
-; GCN-NEXT: buffer_store_dword v4, off, s[0:3], s32 offset:448 ; 4-byte Folded Spill
+; GCN-NEXT: buffer_store_dword v4, off, s[0:3], s32 offset:464 ; 4-byte Folded Spill
; GCN-NEXT: s_mov_b64 exec, s[4:5]
; GCN-NEXT: buffer_store_dword v40, off, s[0:3], s32 offset:444 ; 4-byte Folded Spill
; GCN-NEXT: buffer_store_dword v41, off, s[0:3], s32 offset:440 ; 4-byte Folded Spill
@@ -1315,17 +1299,17 @@
; GCN-NEXT: ; implicit-def: $sgpr4_sgpr5
; GCN-NEXT: flat_load_dwordx4 v[5:8], v[2:3]
; GCN-NEXT: s_waitcnt vmcnt(0)
-; GCN-NEXT: buffer_store_dword v5, off, s[0:3], s32 offset:452 ; 4-byte Folded Spill
+; GCN-NEXT: buffer_store_dword v5, off, s[0:3], s32 offset:448 ; 4-byte Folded Spill
; GCN-NEXT: s_waitcnt vmcnt(0)
-; GCN-NEXT: buffer_store_dword v6, off, s[0:3], s32 offset:456 ; 4-byte Folded Spill
-; GCN-NEXT: buffer_store_dword v7, off, s[0:3], s32 offset:460 ; 4-byte Folded Spill
-; GCN-NEXT: buffer_store_dword v8, off, s[0:3], s32 offset:464 ; 4-byte Folded Spill
+; GCN-NEXT: buffer_store_dword v6, off, s[0:3], s32 offset:452 ; 4-byte Folded Spill
+; GCN-NEXT: buffer_store_dword v7, off, s[0:3], s32 offset:456 ; 4-byte Folded Spill
+; GCN-NEXT: buffer_store_dword v8, off, s[0:3], s32 offset:460 ; 4-byte Folded Spill
; GCN-NEXT: ;;#ASMSTART
; GCN-NEXT: ;;#ASMEND
-; GCN-NEXT: buffer_load_dword v5, off, s[0:3], s32 offset:452 ; 4-byte Folded Reload
-; GCN-NEXT: buffer_load_dword v6, off, s[0:3], s32 offset:456 ; 4-byte Folded Reload
-; GCN-NEXT: buffer_load_dword v7, off, s[0:3], s32 offset:460 ; 4-byte Folded Reload
-; GCN-NEXT: buffer_load_dword v8, off, s[0:3], s32 offset:464 ; 4-byte Folded Reload
+; GCN-NEXT: buffer_load_dword v5, off, s[0:3], s32 offset:448 ; 4-byte Folded Reload
+; GCN-NEXT: buffer_load_dword v6, off, s[0:3], s32 offset:452 ; 4-byte Folded Reload
+; GCN-NEXT: buffer_load_dword v7, off, s[0:3], s32 offset:456 ; 4-byte Folded Reload
+; GCN-NEXT: buffer_load_dword v8, off, s[0:3], s32 offset:460 ; 4-byte Folded Reload
; GCN-NEXT: ;;#ASMSTART
; GCN-NEXT: ;;#ASMEND
; GCN-NEXT: s_waitcnt vmcnt(0)
@@ -1447,7 +1431,7 @@
; GCN-NEXT: buffer_load_dword v41, off, s[0:3], s32 offset:440 ; 4-byte Folded Reload
; GCN-NEXT: buffer_load_dword v40, off, s[0:3], s32 offset:444 ; 4-byte Folded Reload
; GCN-NEXT: s_xor_saveexec_b64 s[4:5], -1
-; GCN-NEXT: buffer_load_dword v4, off, s[0:3], s32 offset:448 ; 4-byte Folded Reload
+; GCN-NEXT: buffer_load_dword v4, off, s[0:3], s32 offset:464 ; 4-byte Folded Reload
; GCN-NEXT: s_mov_b64 exec, s[4:5]
; GCN-NEXT: s_waitcnt vmcnt(0)
; GCN-NEXT: s_setpc_b64 s[30:31]
@@ -1528,7 +1512,7 @@
; GCN-LABEL: spill_sgpr_no_free_vgpr_ipra:
; GCN: ; %bb.0:
; GCN-NEXT: s_waitcnt vmcnt(0) expcnt(0) lgkmcnt(0)
-; GCN-NEXT: s_mov_b32 s18, s33
+; GCN-NEXT: s_mov_b32 s24, s33
; GCN-NEXT: s_mov_b32 s33, s32
; GCN-NEXT: s_add_i32 s32, s32, 0x7400
; GCN-NEXT: buffer_store_dword v40, off, s[0:3], s33 offset:444 ; 4-byte Folded Spill
@@ -1798,7 +1782,7 @@
; GCN-NEXT: buffer_load_dword v41, off, s[0:3], s33 offset:440 ; 4-byte Folded Reload
; GCN-NEXT: buffer_load_dword v40, off, s[0:3], s33 offset:444 ; 4-byte Folded Reload
; GCN-NEXT: s_add_i32 s32, s32, 0xffff8c00
-; GCN-NEXT: s_mov_b32 s33, s18
+; GCN-NEXT: s_mov_b32 s33, s24
; GCN-NEXT: s_waitcnt vmcnt(0)
; GCN-NEXT: s_setpc_b64 s[30:31]
call void @child_function_ipra()
diff --git a/llvm/test/CodeGen/AMDGPU/si-spill-sgpr-stack.ll b/llvm/test/CodeGen/AMDGPU/si-spill-sgpr-stack.ll
--- a/llvm/test/CodeGen/AMDGPU/si-spill-sgpr-stack.ll
+++ b/llvm/test/CodeGen/AMDGPU/si-spill-sgpr-stack.ll
@@ -6,17 +6,17 @@
; ALL: s_mov_b32 s[[HI:[0-9]+]], 0xe80000
; Make sure we are handling hazards correctly.
-; SGPR: v_mov_b32_e32 v0, vcc_lo
-; SGPR-NEXT: s_or_saveexec_b64 [[EXEC_COPY:s\[[0-9]+:[0-9]+\]]], -1
-; SGPR-NEXT: buffer_load_dword [[VHI:v[0-9]+]], off, s[{{[0-9]+:[0-9]+}}], 0 offset:4 ; 4-byte Folded Reload
-; SGPR-NEXT: s_mov_b64 exec, [[EXEC_COPY]]
+; SGPR: buffer_load_dword [[VHI:v[0-9]+]], off, s[{{[0-9]+:[0-9]+}}], 0 offset:4
; SGPR-NEXT: s_waitcnt vmcnt(0)
; SGPR-NEXT: v_readlane_b32 s{{[0-9]+}}, [[VHI]], 0
; SGPR-NEXT: v_readlane_b32 s{{[0-9]+}}, [[VHI]], 1
; SGPR-NEXT: v_readlane_b32 s{{[0-9]+}}, [[VHI]], 2
; SGPR-NEXT: v_readlane_b32 s[[HI:[0-9]+]], [[VHI]], 3
-; SGPR-NEXT: ; kill: killed $vgpr1
-; SGPR-NEXT: s_nop 4
+; SGPR-NEXT: buffer_load_dword [[VHI]], off, s[96:99], 0
+; SGPR-NEXT: s_waitcnt vmcnt(0)
+; SGPR-NEXT: s_mov_b64 exec, s[4:5]
+; SGPR-NEXT: s_nop 1
+; SGPR-NEXT: buffer_store_dword v0, off, s[0:3], 0
; ALL: s_endpgm
define amdgpu_kernel void @test(ptr addrspace(1) %out, i32 %in) {
diff --git a/llvm/test/CodeGen/AMDGPU/sibling-call.ll b/llvm/test/CodeGen/AMDGPU/sibling-call.ll
--- a/llvm/test/CodeGen/AMDGPU/sibling-call.ll
+++ b/llvm/test/CodeGen/AMDGPU/sibling-call.ll
@@ -202,9 +202,10 @@
; GCN-NEXT: s_mov_b32 s33, s32
; GCN-NEXT: s_or_saveexec_b64 s{{\[[0-9]+:[0-9]+\]}}, -1
; GCN-NEXT: buffer_store_dword [[CSRV:v[0-9]+]], off, s[0:3], s33 offset:8 ; 4-byte Folded Spill
+; GCN-NEXT: buffer_store_dword [[CSRV_1:v[0-9]+]], off, s[0:3], s33 offset:12 ; 4-byte Folded Spill
; GCN-NEXT: s_mov_b64 exec
-; GCN-DAG: s_addk_i32 s32, 0x400
-; GCN: v_writelane_b32 [[CSRV]], [[FP_SCRATCH_COPY]], 2
+; GCN-DAG: s_addk_i32 s32, 0x800
+; GCN: v_writelane_b32 [[CSRV_1]], [[FP_SCRATCH_COPY]], 0
; GCN-DAG: s_getpc_b64 s[4:5]
; GCN-DAG: s_add_u32 s4, s4, i32_fastcc_i32_i32@gotpcrel32@lo+4
@@ -226,11 +227,12 @@
; GCN-NEXT: s_addc_u32 s5, s5, sibling_call_i32_fastcc_i32_i32@rel32@hi+12
; GCN-NEXT: v_readlane_b32 s31, [[CSRV]], 1
; GCN-NEXT: v_readlane_b32 s30, [[CSRV]], 0
-; GCN-NEXT: v_readlane_b32 [[FP_SCRATCH_COPY:s[0-9]+]], [[CSRV]], 2
+; GCN-NEXT: v_readlane_b32 [[FP_SCRATCH_COPY:s[0-9]+]], [[CSRV_1]], 0
; GCN-NEXT: s_or_saveexec_b64 s[8:9], -1
; GCN-NEXT: buffer_load_dword [[CSRV]], off, s[0:3], s33 offset:8 ; 4-byte Folded Reload
+; GCN-NEXT: buffer_load_dword [[CSRV_1]], off, s[0:3], s33 offset:12 ; 4-byte Folded Reload
; GCN-NEXT: s_mov_b64 exec, s[8:9]
-; GCN-NEXT: s_addk_i32 s32, 0xfc00
+; GCN-NEXT: s_addk_i32 s32, 0xf800
; GCN-NEXT: s_mov_b32 s33, [[FP_SCRATCH_COPY]]
; GCN-NEXT: s_setpc_b64 s[4:5]
define fastcc i32 @sibling_call_i32_fastcc_i32_i32_other_call(i32 %a, i32 %b, i32 %c) #1 {
diff --git a/llvm/test/CodeGen/AMDGPU/snippet-copy-bundle-regression.mir b/llvm/test/CodeGen/AMDGPU/snippet-copy-bundle-regression.mir
--- a/llvm/test/CodeGen/AMDGPU/snippet-copy-bundle-regression.mir
+++ b/llvm/test/CodeGen/AMDGPU/snippet-copy-bundle-regression.mir
@@ -33,7 +33,6 @@
; CHECK-NEXT: successors: %bb.2(0x40000000), %bb.1(0x40000000)
; CHECK-NEXT: liveins: $sgpr14, $sgpr15, $sgpr16, $vgpr0, $vgpr1, $vgpr2, $sgpr4_sgpr5, $sgpr6_sgpr7, $sgpr8_sgpr9, $sgpr10_sgpr11
; CHECK-NEXT: {{ $}}
- ; CHECK-NEXT: renamable $vgpr1 = IMPLICIT_DEF
; CHECK-NEXT: renamable $sgpr34_sgpr35 = IMPLICIT_DEF
; CHECK-NEXT: dead renamable $vgpr0 = IMPLICIT_DEF
; CHECK-NEXT: renamable $sgpr41 = IMPLICIT_DEF
@@ -42,15 +41,7 @@
; CHECK-NEXT: renamable $sgpr44_sgpr45_sgpr46_sgpr47_sgpr48_sgpr49_sgpr50_sgpr51 = S_LOAD_DWORDX8_IMM renamable $sgpr38_sgpr39, 0, 0 :: (dereferenceable invariant load (s256), align 16, addrspace 4)
; CHECK-NEXT: dead renamable $sgpr4 = S_LOAD_DWORD_IMM renamable $sgpr38_sgpr39, 48, 0 :: (dereferenceable invariant load (s32), align 16, addrspace 4)
; CHECK-NEXT: renamable $sgpr4_sgpr5_sgpr6_sgpr7_sgpr8_sgpr9_sgpr10_sgpr11 = S_LOAD_DWORDX8_IMM renamable $sgpr38_sgpr39, 56, 0 :: (dereferenceable invariant load (s256), align 8, addrspace 4)
- ; CHECK-NEXT: renamable $vgpr1 = V_WRITELANE_B32 $sgpr4, 0, killed $vgpr1, implicit-def $sgpr4_sgpr5_sgpr6_sgpr7_sgpr8_sgpr9_sgpr10_sgpr11, implicit $sgpr4_sgpr5_sgpr6_sgpr7_sgpr8_sgpr9_sgpr10_sgpr11
- ; CHECK-NEXT: renamable $vgpr1 = V_WRITELANE_B32 $sgpr5, 1, killed $vgpr1
- ; CHECK-NEXT: renamable $vgpr1 = V_WRITELANE_B32 $sgpr6, 2, killed $vgpr1
- ; CHECK-NEXT: renamable $vgpr1 = V_WRITELANE_B32 $sgpr7, 3, killed $vgpr1
- ; CHECK-NEXT: renamable $vgpr1 = V_WRITELANE_B32 $sgpr8, 4, killed $vgpr1
- ; CHECK-NEXT: renamable $vgpr1 = V_WRITELANE_B32 $sgpr9, 5, killed $vgpr1
- ; CHECK-NEXT: renamable $vgpr1 = V_WRITELANE_B32 $sgpr10, 6, killed $vgpr1
- ; CHECK-NEXT: renamable $vgpr1 = V_WRITELANE_B32 killed $sgpr11, 7, killed $vgpr1, implicit killed $sgpr4_sgpr5_sgpr6_sgpr7_sgpr8_sgpr9_sgpr10_sgpr11
- ; CHECK-NEXT: SI_SPILL_WWM_V32_SAVE killed $vgpr1, %stack.1, $sgpr32, 0, implicit $exec :: (store (s32) into %stack.1, addrspace 5)
+ ; CHECK-NEXT: SI_SPILL_S256_SAVE killed renamable $sgpr4_sgpr5_sgpr6_sgpr7_sgpr8_sgpr9_sgpr10_sgpr11, %stack.0, implicit $exec, implicit $sgpr32 :: (store (s256) into %stack.0, align 4, addrspace 5)
; CHECK-NEXT: dead renamable $sgpr4_sgpr5 = S_LOAD_DWORDX2_IMM renamable $sgpr44_sgpr45, 0, 0 :: (invariant load (s64), align 16, addrspace 4)
; CHECK-NEXT: ADJCALLSTACKUP 0, 0, implicit-def dead $scc, implicit-def $sgpr32, implicit $sgpr32
; CHECK-NEXT: $vgpr1 = COPY renamable $sgpr51
@@ -63,50 +54,30 @@
; CHECK-NEXT: successors: %bb.3(0x80000000)
; CHECK-NEXT: liveins: $sgpr34_sgpr35, $sgpr36_sgpr37, $sgpr38_sgpr39, $sgpr44_sgpr45_sgpr46_sgpr47_sgpr48_sgpr49_sgpr50_sgpr51:0x000000000000FC00
; CHECK-NEXT: {{ $}}
- ; CHECK-NEXT: renamable $vgpr1 = SI_SPILL_WWM_V32_RESTORE %stack.1, $sgpr32, 0, implicit $exec :: (load (s32) from %stack.1, addrspace 5)
- ; CHECK-NEXT: $sgpr4 = V_READLANE_B32 $vgpr1, 0, implicit-def $sgpr4_sgpr5_sgpr6_sgpr7_sgpr8_sgpr9_sgpr10_sgpr11
- ; CHECK-NEXT: $sgpr5 = V_READLANE_B32 $vgpr1, 1
- ; CHECK-NEXT: $sgpr6 = V_READLANE_B32 $vgpr1, 2
- ; CHECK-NEXT: $sgpr7 = V_READLANE_B32 $vgpr1, 3
- ; CHECK-NEXT: $sgpr8 = V_READLANE_B32 $vgpr1, 4
- ; CHECK-NEXT: $sgpr9 = V_READLANE_B32 $vgpr1, 5
- ; CHECK-NEXT: $sgpr10 = V_READLANE_B32 $vgpr1, 6
- ; CHECK-NEXT: $sgpr11 = V_READLANE_B32 $vgpr1, 7
- ; CHECK-NEXT: $noreg = S_OR_SAVEEXEC_B64 -1, implicit-def $exec, implicit-def dead $scc, implicit $exec
- ; CHECK-NEXT: $exec = S_MOV_B64 killed $noreg
+ ; CHECK-NEXT: renamable $sgpr4_sgpr5_sgpr6_sgpr7_sgpr8_sgpr9_sgpr10_sgpr11 = SI_SPILL_S256_RESTORE %stack.0, implicit $exec, implicit $sgpr32 :: (load (s256) from %stack.0, align 4, addrspace 5)
; CHECK-NEXT: S_BRANCH %bb.3
; CHECK-NEXT: {{ $}}
; CHECK-NEXT: bb.2:
; CHECK-NEXT: successors: %bb.3(0x80000000)
; CHECK-NEXT: liveins: $sgpr34_sgpr35, $sgpr36_sgpr37, $sgpr38_sgpr39, $sgpr44_sgpr45_sgpr46_sgpr47_sgpr48_sgpr49_sgpr50_sgpr51:0x000000000000FC00
; CHECK-NEXT: {{ $}}
- ; CHECK-NEXT: renamable $vgpr1 = SI_SPILL_WWM_V32_RESTORE %stack.1, $sgpr32, 0, implicit $exec :: (load (s32) from %stack.1, addrspace 5)
- ; CHECK-NEXT: $sgpr4 = V_READLANE_B32 $vgpr1, 0, implicit-def $sgpr4_sgpr5_sgpr6_sgpr7_sgpr8_sgpr9_sgpr10_sgpr11
- ; CHECK-NEXT: $sgpr5 = V_READLANE_B32 $vgpr1, 1
- ; CHECK-NEXT: $sgpr6 = V_READLANE_B32 $vgpr1, 2
- ; CHECK-NEXT: $sgpr7 = V_READLANE_B32 $vgpr1, 3
- ; CHECK-NEXT: $sgpr8 = V_READLANE_B32 $vgpr1, 4
- ; CHECK-NEXT: $sgpr9 = V_READLANE_B32 $vgpr1, 5
- ; CHECK-NEXT: $sgpr10 = V_READLANE_B32 $vgpr1, 6
- ; CHECK-NEXT: $sgpr11 = V_READLANE_B32 $vgpr1, 7
+ ; CHECK-NEXT: renamable $sgpr4_sgpr5_sgpr6_sgpr7_sgpr8_sgpr9_sgpr10_sgpr11 = SI_SPILL_S256_RESTORE %stack.0, implicit $exec, implicit $sgpr32 :: (load (s256) from %stack.0, align 4, addrspace 5)
; CHECK-NEXT: S_CMP_LG_U64 renamable $sgpr4_sgpr5, 0, implicit-def $scc
- ; CHECK-NEXT: $noreg = S_OR_SAVEEXEC_B64 -1, implicit-def $exec, implicit-def dead $scc, implicit $exec
- ; CHECK-NEXT: $exec = S_MOV_B64 killed $noreg
; CHECK-NEXT: {{ $}}
; CHECK-NEXT: bb.3:
; CHECK-NEXT: successors: %bb.5(0x40000000), %bb.4(0x40000000)
- ; CHECK-NEXT: liveins: $vgpr1, $sgpr34_sgpr35, $sgpr36_sgpr37, $sgpr38_sgpr39, $sgpr4_sgpr5_sgpr6_sgpr7_sgpr8_sgpr9_sgpr10_sgpr11:0x00000000000003F0, $sgpr44_sgpr45_sgpr46_sgpr47_sgpr48_sgpr49_sgpr50_sgpr51:0x000000000000FC00
+ ; CHECK-NEXT: liveins: $sgpr34_sgpr35, $sgpr36_sgpr37, $sgpr38_sgpr39, $sgpr4_sgpr5_sgpr6_sgpr7_sgpr8_sgpr9_sgpr10_sgpr11:0x00000000000003F0, $sgpr44_sgpr45_sgpr46_sgpr47_sgpr48_sgpr49_sgpr50_sgpr51:0x000000000000FC00
; CHECK-NEXT: {{ $}}
; CHECK-NEXT: S_CBRANCH_VCCZ %bb.5, implicit undef $vcc
; CHECK-NEXT: {{ $}}
; CHECK-NEXT: bb.4:
; CHECK-NEXT: successors: %bb.5(0x80000000)
- ; CHECK-NEXT: liveins: $vgpr1, $sgpr34_sgpr35, $sgpr36_sgpr37, $sgpr38_sgpr39, $sgpr4_sgpr5_sgpr6_sgpr7_sgpr8_sgpr9_sgpr10_sgpr11:0x00000000000003F0, $sgpr44_sgpr45_sgpr46_sgpr47_sgpr48_sgpr49_sgpr50_sgpr51:0x000000000000FC00
+ ; CHECK-NEXT: liveins: $sgpr34_sgpr35, $sgpr36_sgpr37, $sgpr38_sgpr39, $sgpr4_sgpr5_sgpr6_sgpr7_sgpr8_sgpr9_sgpr10_sgpr11:0x00000000000003F0, $sgpr44_sgpr45_sgpr46_sgpr47_sgpr48_sgpr49_sgpr50_sgpr51:0x000000000000FC00
; CHECK-NEXT: {{ $}}
; CHECK-NEXT: S_CMP_EQ_U32 renamable $sgpr8, 0, implicit-def $scc
; CHECK-NEXT: {{ $}}
; CHECK-NEXT: bb.5:
- ; CHECK-NEXT: liveins: $vgpr1, $sgpr34_sgpr35, $sgpr36_sgpr37, $sgpr38_sgpr39, $sgpr4_sgpr5_sgpr6_sgpr7_sgpr8_sgpr9_sgpr10_sgpr11:0x00000000000000F0, $sgpr44_sgpr45_sgpr46_sgpr47_sgpr48_sgpr49_sgpr50_sgpr51:0x000000000000FC00
+ ; CHECK-NEXT: liveins: $sgpr34_sgpr35, $sgpr36_sgpr37, $sgpr38_sgpr39, $sgpr4_sgpr5_sgpr6_sgpr7_sgpr8_sgpr9_sgpr10_sgpr11:0x00000000000000F0, $sgpr44_sgpr45_sgpr46_sgpr47_sgpr48_sgpr49_sgpr50_sgpr51:0x000000000000FC00
; CHECK-NEXT: {{ $}}
; CHECK-NEXT: dead renamable $sgpr4_sgpr5 = S_LOAD_DWORDX2_IMM killed renamable $sgpr38_sgpr39, 40, 0 :: (dereferenceable invariant load (s64), addrspace 4)
; CHECK-NEXT: GLOBAL_STORE_DWORD_SADDR undef renamable $vgpr0, undef renamable $vgpr0, killed renamable $sgpr6_sgpr7, 0, 0, implicit $exec :: (store (s32), addrspace 1)
@@ -116,7 +87,6 @@
; CHECK-NEXT: $sgpr6_sgpr7 = COPY killed renamable $sgpr36_sgpr37
; CHECK-NEXT: $sgpr10_sgpr11 = COPY killed renamable $sgpr34_sgpr35
; CHECK-NEXT: ADJCALLSTACKDOWN 0, 0, implicit-def dead $scc, implicit-def $sgpr32, implicit $sgpr32
- ; CHECK-NEXT: KILL killed renamable $vgpr1
; CHECK-NEXT: S_ENDPGM 0
bb.0:
liveins: $vgpr0, $vgpr1, $vgpr2, $sgpr4_sgpr5, $sgpr6_sgpr7, $sgpr8_sgpr9, $sgpr10_sgpr11, $sgpr14, $sgpr15, $sgpr16
diff --git a/llvm/test/CodeGen/AMDGPU/spill-csr-frame-ptr-reg-copy.ll b/llvm/test/CodeGen/AMDGPU/spill-csr-frame-ptr-reg-copy.ll
--- a/llvm/test/CodeGen/AMDGPU/spill-csr-frame-ptr-reg-copy.ll
+++ b/llvm/test/CodeGen/AMDGPU/spill-csr-frame-ptr-reg-copy.ll
@@ -2,22 +2,20 @@
; GCN-LABEL: {{^}}spill_csr_s5_copy:
; GCN: s_mov_b32 [[FP_SCRATCH_COPY:s[0-9]+]], s33
-; GCN: s_xor_saveexec_b64
-; GCN-NEXT: buffer_store_dword v0, off, s[0:3], s33 offset:12 ; 4-byte Folded Spill
-; GCN-NEXT: s_mov_b64 exec, -1
+; GCN: s_or_saveexec_b64
; GCN-NEXT: buffer_store_dword v40, off, s[0:3], s33 offset:4 ; 4-byte Folded Spill
+; GCN-NEXT: buffer_store_dword v41, off, s[0:3], s33 offset:8 ; 4-byte Folded Spill
; GCN-NEXT: s_mov_b64 exec
-; GCN: v_writelane_b32 v40, [[FP_SCRATCH_COPY]], 4
+; GCN: v_writelane_b32 v41, [[FP_SCRATCH_COPY]], 0
; GCN: s_swappc_b64
; GCN: v_mov_b32_e32 [[K:v[0-9]+]], 9
; GCN: buffer_store_dword [[K]], off, s[0:3], s33{{$}}
-; GCN: v_readlane_b32 [[FP_SCRATCH_COPY:s[0-9]+]], v40, 4
-; GCN: s_xor_saveexec_b64
-; GCN-NEXT: buffer_load_dword v0, off, s[0:3], s33 offset:12 ; 4-byte Folded Reload
-; GCN-NEXT: s_mov_b64 exec, -1
+; GCN: v_readlane_b32 [[FP_SCRATCH_COPY:s[0-9]+]], v41, 0
+; GCN: s_or_saveexec_b64
; GCN-NEXT: buffer_load_dword v40, off, s[0:3], s33 offset:4 ; 4-byte Folded Reload
+; GCN-NEXT: buffer_load_dword v41, off, s[0:3], s33 offset:8 ; 4-byte Folded Reload
; GCN: s_mov_b64 exec
; GCN: s_mov_b32 s33, [[FP_SCRATCH_COPY]]
; GCN: s_setpc_b64
diff --git a/llvm/test/CodeGen/AMDGPU/spill-reg-tuple-super-reg-use.mir b/llvm/test/CodeGen/AMDGPU/spill-reg-tuple-super-reg-use.mir
--- a/llvm/test/CodeGen/AMDGPU/spill-reg-tuple-super-reg-use.mir
+++ b/llvm/test/CodeGen/AMDGPU/spill-reg-tuple-super-reg-use.mir
@@ -1,5 +1,5 @@
# NOTE: Assertions have been autogenerated by utils/update_mir_test_checks.py
-# RUN: llc -march=amdgcn -mcpu=gfx900 -start-before=si-lower-sgpr-spills -stop-after=prologepilog -verify-machineinstrs %s -o - | FileCheck -check-prefix=GCN %s
+# RUN: llc -march=amdgcn -mcpu=gfx900 -run-pass=si-lower-sgpr-spills,prologepilog,machine-cp -verify-machineinstrs %s -o - | FileCheck -check-prefix=GCN %s
# Make sure the initial first $sgpr1 = COPY $sgpr2 copy is not deleted
# by the copy propagation after lowering the spill.
@@ -25,14 +25,12 @@
; GCN-NEXT: $sgpr8_sgpr9 = S_XOR_SAVEEXEC_B64 -1, implicit-def $exec, implicit-def dead $scc, implicit $exec
; GCN-NEXT: BUFFER_STORE_DWORD_OFFSET $vgpr0, $sgpr100_sgpr101_sgpr102_sgpr103, $sgpr32, 0, 0, 0, implicit $exec :: (store (s32) into %stack.1, addrspace 5)
; GCN-NEXT: $exec = S_MOV_B64 killed $sgpr8_sgpr9
- ; GCN-NEXT: renamable $vgpr0 = IMPLICIT_DEF
; GCN-NEXT: renamable $sgpr1 = COPY $sgpr2
- ; GCN-NEXT: renamable $vgpr0 = V_WRITELANE_B32 $sgpr0, 0, killed $vgpr0, implicit-def $sgpr0_sgpr1_sgpr2_sgpr3, implicit $sgpr0_sgpr1_sgpr2_sgpr3
- ; GCN-NEXT: renamable $vgpr0 = V_WRITELANE_B32 $sgpr1, 1, killed $vgpr0
- ; GCN-NEXT: renamable $vgpr0 = V_WRITELANE_B32 $sgpr2, 2, killed $vgpr0
- ; GCN-NEXT: renamable $vgpr0 = V_WRITELANE_B32 $sgpr3, 3, killed $vgpr0, implicit $sgpr0_sgpr1_sgpr2_sgpr3
- ; GCN-NEXT: renamable $sgpr8 = COPY renamable $sgpr1
- ; GCN-NEXT: KILL killed renamable $vgpr0
+ ; GCN-NEXT: $vgpr0 = V_WRITELANE_B32 $sgpr0, 0, $vgpr0, implicit-def $sgpr0_sgpr1_sgpr2_sgpr3, implicit $sgpr0_sgpr1_sgpr2_sgpr3
+ ; GCN-NEXT: $vgpr0 = V_WRITELANE_B32 $sgpr1, 1, $vgpr0
+ ; GCN-NEXT: $vgpr0 = V_WRITELANE_B32 $sgpr2, 2, $vgpr0
+ ; GCN-NEXT: $vgpr0 = V_WRITELANE_B32 $sgpr3, 3, $vgpr0, implicit $sgpr0_sgpr1_sgpr2_sgpr3
+ ; GCN-NEXT: renamable $sgpr8 = COPY killed renamable $sgpr1
; GCN-NEXT: $sgpr0_sgpr1 = S_XOR_SAVEEXEC_B64 -1, implicit-def $exec, implicit-def dead $scc, implicit $exec
; GCN-NEXT: $vgpr0 = BUFFER_LOAD_DWORD_OFFSET $sgpr100_sgpr101_sgpr102_sgpr103, $sgpr32, 0, 0, 0, implicit $exec :: (load (s32) from %stack.1, addrspace 5)
; GCN-NEXT: $exec = S_MOV_B64 killed $sgpr0_sgpr1
@@ -64,13 +62,11 @@
; GCN-NEXT: $sgpr8_sgpr9 = S_XOR_SAVEEXEC_B64 -1, implicit-def $exec, implicit-def dead $scc, implicit $exec
; GCN-NEXT: BUFFER_STORE_DWORD_OFFSET $vgpr0, $sgpr100_sgpr101_sgpr102_sgpr103, $sgpr32, 0, 0, 0, implicit $exec :: (store (s32) into %stack.1, addrspace 5)
; GCN-NEXT: $exec = S_MOV_B64 killed $sgpr8_sgpr9
- ; GCN-NEXT: renamable $vgpr0 = IMPLICIT_DEF
; GCN-NEXT: renamable $sgpr1 = COPY $sgpr2
- ; GCN-NEXT: renamable $vgpr0 = V_WRITELANE_B32 $sgpr0, 0, killed $vgpr0, implicit-def $sgpr0_sgpr1_sgpr2_sgpr3, implicit $sgpr0_sgpr1_sgpr2_sgpr3
- ; GCN-NEXT: renamable $vgpr0 = V_WRITELANE_B32 $sgpr1, 1, killed $vgpr0
- ; GCN-NEXT: renamable $vgpr0 = V_WRITELANE_B32 $sgpr2, 2, killed $vgpr0
- ; GCN-NEXT: renamable $vgpr0 = V_WRITELANE_B32 $sgpr3, 3, killed $vgpr0, implicit $sgpr0_sgpr1_sgpr2_sgpr3
- ; GCN-NEXT: KILL killed renamable $vgpr0
+ ; GCN-NEXT: $vgpr0 = V_WRITELANE_B32 $sgpr0, 0, $vgpr0, implicit-def $sgpr0_sgpr1_sgpr2_sgpr3, implicit $sgpr0_sgpr1_sgpr2_sgpr3
+ ; GCN-NEXT: $vgpr0 = V_WRITELANE_B32 $sgpr1, 1, $vgpr0
+ ; GCN-NEXT: $vgpr0 = V_WRITELANE_B32 $sgpr2, 2, $vgpr0
+ ; GCN-NEXT: $vgpr0 = V_WRITELANE_B32 killed $sgpr3, 3, $vgpr0, implicit killed $sgpr0_sgpr1_sgpr2_sgpr3
; GCN-NEXT: $sgpr0_sgpr1 = S_XOR_SAVEEXEC_B64 -1, implicit-def $exec, implicit-def dead $scc, implicit $exec
; GCN-NEXT: $vgpr0 = BUFFER_LOAD_DWORD_OFFSET $sgpr100_sgpr101_sgpr102_sgpr103, $sgpr32, 0, 0, 0, implicit $exec :: (load (s32) from %stack.1, addrspace 5)
; GCN-NEXT: $exec = S_MOV_B64 killed $sgpr0_sgpr1
@@ -97,12 +93,12 @@
; GCN-LABEL: name: spill_vgpr128_use_subreg
; GCN: liveins: $vgpr0, $vgpr1, $vgpr2, $vgpr3, $vgpr4, $vgpr5, $vgpr6, $vgpr7
; GCN-NEXT: {{ $}}
- ; GCN-NEXT: renamable $vgpr1 = COPY $vgpr2, implicit $exec
+ ; GCN-NEXT: renamable $vgpr1 = COPY $vgpr2
; GCN-NEXT: BUFFER_STORE_DWORD_OFFSET $vgpr0, $sgpr100_sgpr101_sgpr102_sgpr103, $sgpr32, 0, 0, 0, implicit $exec, implicit-def $vgpr0_vgpr1_vgpr2_vgpr3, implicit $vgpr0_vgpr1_vgpr2_vgpr3 :: (store (s32) into %stack.0, addrspace 5)
; GCN-NEXT: BUFFER_STORE_DWORD_OFFSET $vgpr1, $sgpr100_sgpr101_sgpr102_sgpr103, $sgpr32, 4, 0, 0, implicit $exec :: (store (s32) into %stack.0 + 4, addrspace 5)
; GCN-NEXT: BUFFER_STORE_DWORD_OFFSET $vgpr2, $sgpr100_sgpr101_sgpr102_sgpr103, $sgpr32, 8, 0, 0, implicit $exec :: (store (s32) into %stack.0 + 8, addrspace 5)
; GCN-NEXT: BUFFER_STORE_DWORD_OFFSET $vgpr3, $sgpr100_sgpr101_sgpr102_sgpr103, $sgpr32, 12, 0, 0, implicit $exec, implicit $vgpr0_vgpr1_vgpr2_vgpr3 :: (store (s32) into %stack.0 + 12, addrspace 5)
- ; GCN-NEXT: renamable $vgpr8 = COPY $vgpr2, implicit $exec
+ ; GCN-NEXT: renamable $vgpr8 = COPY killed renamable $vgpr1
; GCN-NEXT: S_ENDPGM 0, implicit $vgpr8
renamable $vgpr1 = COPY $vgpr2
SI_SPILL_V128_SAVE renamable $vgpr0_vgpr1_vgpr2_vgpr3, %stack.0, $sgpr32, 0, implicit $exec :: (store (s128) into %stack.0, align 4, addrspace 5)
@@ -127,11 +123,11 @@
; GCN-LABEL: name: spill_vgpr128_use_kill
; GCN: liveins: $vgpr0, $vgpr1, $vgpr2, $vgpr3, $vgpr4, $vgpr5, $vgpr6, $vgpr7
; GCN-NEXT: {{ $}}
- ; GCN-NEXT: renamable $vgpr1 = COPY $vgpr2, implicit $exec
- ; GCN-NEXT: BUFFER_STORE_DWORD_OFFSET $vgpr0, $sgpr100_sgpr101_sgpr102_sgpr103, $sgpr32, 0, 0, 0, implicit $exec, implicit-def $vgpr0_vgpr1_vgpr2_vgpr3, implicit $vgpr0_vgpr1_vgpr2_vgpr3 :: (store (s32) into %stack.0, addrspace 5)
- ; GCN-NEXT: BUFFER_STORE_DWORD_OFFSET $vgpr1, $sgpr100_sgpr101_sgpr102_sgpr103, $sgpr32, 4, 0, 0, implicit $exec :: (store (s32) into %stack.0 + 4, addrspace 5)
- ; GCN-NEXT: BUFFER_STORE_DWORD_OFFSET $vgpr2, $sgpr100_sgpr101_sgpr102_sgpr103, $sgpr32, 8, 0, 0, implicit $exec :: (store (s32) into %stack.0 + 8, addrspace 5)
- ; GCN-NEXT: BUFFER_STORE_DWORD_OFFSET $vgpr3, $sgpr100_sgpr101_sgpr102_sgpr103, $sgpr32, 12, 0, 0, implicit $exec, implicit $vgpr0_vgpr1_vgpr2_vgpr3 :: (store (s32) into %stack.0 + 12, addrspace 5)
+ ; GCN-NEXT: renamable $vgpr1 = COPY $vgpr2
+ ; GCN-NEXT: BUFFER_STORE_DWORD_OFFSET killed $vgpr0, $sgpr100_sgpr101_sgpr102_sgpr103, $sgpr32, 0, 0, 0, implicit $exec, implicit-def $vgpr0_vgpr1_vgpr2_vgpr3, implicit $vgpr0_vgpr1_vgpr2_vgpr3 :: (store (s32) into %stack.0, addrspace 5)
+ ; GCN-NEXT: BUFFER_STORE_DWORD_OFFSET killed $vgpr1, $sgpr100_sgpr101_sgpr102_sgpr103, $sgpr32, 4, 0, 0, implicit $exec :: (store (s32) into %stack.0 + 4, addrspace 5)
+ ; GCN-NEXT: BUFFER_STORE_DWORD_OFFSET killed $vgpr2, $sgpr100_sgpr101_sgpr102_sgpr103, $sgpr32, 8, 0, 0, implicit $exec :: (store (s32) into %stack.0 + 8, addrspace 5)
+ ; GCN-NEXT: BUFFER_STORE_DWORD_OFFSET killed $vgpr3, $sgpr100_sgpr101_sgpr102_sgpr103, $sgpr32, 12, 0, 0, implicit $exec, implicit killed $vgpr0_vgpr1_vgpr2_vgpr3 :: (store (s32) into %stack.0 + 12, addrspace 5)
; GCN-NEXT: S_ENDPGM 0
renamable $vgpr1 = COPY $vgpr2
SI_SPILL_V128_SAVE renamable killed $vgpr0_vgpr1_vgpr2_vgpr3, %stack.0, $sgpr32, 0, implicit $exec :: (store (s128) into %stack.0, align 4, addrspace 5)
diff --git a/llvm/test/CodeGen/AMDGPU/spill-sgpr-to-virtual-vgpr.mir b/llvm/test/CodeGen/AMDGPU/spill-sgpr-to-virtual-vgpr.mir
deleted file mode 100644
--- a/llvm/test/CodeGen/AMDGPU/spill-sgpr-to-virtual-vgpr.mir
+++ /dev/null
@@ -1,319 +0,0 @@
-# NOTE: Assertions have been autogenerated by utils/update_mir_test_checks.py
-# RUN: llc -mtriple=amdgcn-amd-amdhsa -mcpu=gfx906 -run-pass=si-lower-sgpr-spills -verify-machineinstrs %s -o - | FileCheck -check-prefix=GCN %s
-
-# A simple SGPR spill. Implicit def for lane VGPR should be inserted just before the spill instruction.
----
-name: sgpr32_spill
-tracksRegLiveness: true
-frameInfo:
- maxAlignment: 4
-stack:
- - { id: 0, type: spill-slot, size: 4, alignment: 4, stack-id: sgpr-spill }
-machineFunctionInfo:
- isEntryFunction: false
- scratchRSrcReg: '$sgpr0_sgpr1_sgpr2_sgpr3'
- stackPtrOffsetReg: '$sgpr32'
- frameOffsetReg: '$sgpr33'
- hasSpilledSGPRs: true
-body: |
- bb.0:
- liveins: $sgpr30_sgpr31, $sgpr10
- ; GCN-LABEL: name: sgpr32_spill
- ; GCN: liveins: $sgpr30_sgpr31, $sgpr10
- ; GCN-NEXT: {{ $}}
- ; GCN-NEXT: [[DEF:%[0-9]+]]:vgpr_32 = IMPLICIT_DEF
- ; GCN-NEXT: S_NOP 0
- ; GCN-NEXT: [[V_WRITELANE_B32_:%[0-9]+]]:vgpr_32 = V_WRITELANE_B32 killed $sgpr10, 0, [[V_WRITELANE_B32_]]
- ; GCN-NEXT: $sgpr10 = V_READLANE_B32 [[V_WRITELANE_B32_]], 0
- ; GCN-NEXT: S_SETPC_B64 $sgpr30_sgpr31
- S_NOP 0
- SI_SPILL_S32_SAVE killed $sgpr10, %stack.0, implicit $exec, implicit $sgpr0_sgpr1_sgpr2_sgpr3, implicit $sgpr32
- renamable $sgpr10 = SI_SPILL_S32_RESTORE %stack.0, implicit $exec, implicit $sgpr0_sgpr1_sgpr2_sgpr3, implicit $sgpr32
- S_SETPC_B64 $sgpr30_sgpr31
-...
-
-# Needed an additional virtual lane register as the lanes of current register are fully occupied while spilling a wide SGPR tuple.
-# There must be two implicit def for the two lane VGPRs.
-
----
-name: sgpr_spill_lane_crossover
-tracksRegLiveness: true
-frameInfo:
- maxAlignment: 4
-stack:
- - { id: 0, type: spill-slot, size: 4, alignment: 4, stack-id: sgpr-spill }
- - { id: 1, type: spill-slot, size: 128, alignment: 4, stack-id: sgpr-spill }
-machineFunctionInfo:
- isEntryFunction: false
- scratchRSrcReg: '$sgpr0_sgpr1_sgpr2_sgpr3'
- stackPtrOffsetReg: '$sgpr32'
- frameOffsetReg: '$sgpr33'
- hasSpilledSGPRs: true
-body: |
- bb.0:
- liveins: $sgpr30_sgpr31, $sgpr10, $sgpr64_sgpr65_sgpr66_sgpr67_sgpr68_sgpr69_sgpr70_sgpr71, $sgpr72_sgpr73_sgpr74_sgpr75_sgpr76_sgpr77_sgpr78_sgpr79, $sgpr80_sgpr81_sgpr82_sgpr83_sgpr84_sgpr85_sgpr86_sgpr87, $sgpr88_sgpr89_sgpr90_sgpr91_sgpr92_sgpr93_sgpr94_sgpr95
- ; GCN-LABEL: name: sgpr_spill_lane_crossover
- ; GCN: liveins: $sgpr10, $sgpr64, $sgpr65, $sgpr66, $sgpr67, $sgpr68, $sgpr69, $sgpr70, $sgpr71, $sgpr72, $sgpr73, $sgpr74, $sgpr75, $sgpr76, $sgpr77, $sgpr78, $sgpr79, $sgpr80, $sgpr81, $sgpr82, $sgpr83, $sgpr84, $sgpr85, $sgpr86, $sgpr87, $sgpr88, $sgpr89, $sgpr90, $sgpr91, $sgpr92, $sgpr93, $sgpr94, $sgpr95, $vgpr0, $sgpr30_sgpr31, $sgpr64_sgpr65_sgpr66_sgpr67_sgpr68_sgpr69_sgpr70_sgpr71, $sgpr72_sgpr73_sgpr74_sgpr75_sgpr76_sgpr77_sgpr78_sgpr79, $sgpr80_sgpr81_sgpr82_sgpr83_sgpr84_sgpr85_sgpr86_sgpr87, $sgpr88_sgpr89_sgpr90_sgpr91_sgpr92_sgpr93_sgpr94_sgpr95
- ; GCN-NEXT: {{ $}}
- ; GCN-NEXT: [[DEF:%[0-9]+]]:vgpr_32 = IMPLICIT_DEF
- ; GCN-NEXT: $vgpr0 = V_WRITELANE_B32 killed $sgpr64, 0, $vgpr0
- ; GCN-NEXT: $vgpr0 = V_WRITELANE_B32 killed $sgpr65, 1, $vgpr0
- ; GCN-NEXT: $vgpr0 = V_WRITELANE_B32 killed $sgpr66, 2, $vgpr0
- ; GCN-NEXT: $vgpr0 = V_WRITELANE_B32 killed $sgpr67, 3, $vgpr0
- ; GCN-NEXT: $vgpr0 = V_WRITELANE_B32 killed $sgpr68, 4, $vgpr0
- ; GCN-NEXT: $vgpr0 = V_WRITELANE_B32 killed $sgpr69, 5, $vgpr0
- ; GCN-NEXT: $vgpr0 = V_WRITELANE_B32 killed $sgpr70, 6, $vgpr0
- ; GCN-NEXT: $vgpr0 = V_WRITELANE_B32 killed $sgpr71, 7, $vgpr0
- ; GCN-NEXT: $vgpr0 = V_WRITELANE_B32 killed $sgpr72, 8, $vgpr0
- ; GCN-NEXT: $vgpr0 = V_WRITELANE_B32 killed $sgpr73, 9, $vgpr0
- ; GCN-NEXT: $vgpr0 = V_WRITELANE_B32 killed $sgpr74, 10, $vgpr0
- ; GCN-NEXT: $vgpr0 = V_WRITELANE_B32 killed $sgpr75, 11, $vgpr0
- ; GCN-NEXT: $vgpr0 = V_WRITELANE_B32 killed $sgpr76, 12, $vgpr0
- ; GCN-NEXT: $vgpr0 = V_WRITELANE_B32 killed $sgpr77, 13, $vgpr0
- ; GCN-NEXT: $vgpr0 = V_WRITELANE_B32 killed $sgpr78, 14, $vgpr0
- ; GCN-NEXT: $vgpr0 = V_WRITELANE_B32 killed $sgpr79, 15, $vgpr0
- ; GCN-NEXT: $vgpr0 = V_WRITELANE_B32 killed $sgpr80, 16, $vgpr0
- ; GCN-NEXT: $vgpr0 = V_WRITELANE_B32 killed $sgpr81, 17, $vgpr0
- ; GCN-NEXT: $vgpr0 = V_WRITELANE_B32 killed $sgpr82, 18, $vgpr0
- ; GCN-NEXT: $vgpr0 = V_WRITELANE_B32 killed $sgpr83, 19, $vgpr0
- ; GCN-NEXT: $vgpr0 = V_WRITELANE_B32 killed $sgpr84, 20, $vgpr0
- ; GCN-NEXT: $vgpr0 = V_WRITELANE_B32 killed $sgpr85, 21, $vgpr0
- ; GCN-NEXT: $vgpr0 = V_WRITELANE_B32 killed $sgpr86, 22, $vgpr0
- ; GCN-NEXT: $vgpr0 = V_WRITELANE_B32 killed $sgpr87, 23, $vgpr0
- ; GCN-NEXT: $vgpr0 = V_WRITELANE_B32 killed $sgpr88, 24, $vgpr0
- ; GCN-NEXT: $vgpr0 = V_WRITELANE_B32 killed $sgpr89, 25, $vgpr0
- ; GCN-NEXT: $vgpr0 = V_WRITELANE_B32 killed $sgpr90, 26, $vgpr0
- ; GCN-NEXT: $vgpr0 = V_WRITELANE_B32 killed $sgpr91, 27, $vgpr0
- ; GCN-NEXT: $vgpr0 = V_WRITELANE_B32 killed $sgpr92, 28, $vgpr0
- ; GCN-NEXT: $vgpr0 = V_WRITELANE_B32 killed $sgpr93, 29, $vgpr0
- ; GCN-NEXT: $vgpr0 = V_WRITELANE_B32 killed $sgpr94, 30, $vgpr0
- ; GCN-NEXT: $vgpr0 = V_WRITELANE_B32 killed $sgpr95, 31, $vgpr0
- ; GCN-NEXT: S_NOP 0
- ; GCN-NEXT: [[V_WRITELANE_B32_:%[0-9]+]]:vgpr_32 = V_WRITELANE_B32 killed $sgpr10, 0, [[V_WRITELANE_B32_]]
- ; GCN-NEXT: [[V_WRITELANE_B32_1:%[0-9]+]]:vgpr_32 = V_WRITELANE_B32 $sgpr64, 1, [[V_WRITELANE_B32_1]], implicit-def $sgpr64_sgpr65_sgpr66_sgpr67_sgpr68_sgpr69_sgpr70_sgpr71_sgpr72_sgpr73_sgpr74_sgpr75_sgpr76_sgpr77_sgpr78_sgpr79_sgpr80_sgpr81_sgpr82_sgpr83_sgpr84_sgpr85_sgpr86_sgpr87_sgpr88_sgpr89_sgpr90_sgpr91_sgpr92_sgpr93_sgpr94_sgpr95, implicit $sgpr64_sgpr65_sgpr66_sgpr67_sgpr68_sgpr69_sgpr70_sgpr71_sgpr72_sgpr73_sgpr74_sgpr75_sgpr76_sgpr77_sgpr78_sgpr79_sgpr80_sgpr81_sgpr82_sgpr83_sgpr84_sgpr85_sgpr86_sgpr87_sgpr88_sgpr89_sgpr90_sgpr91_sgpr92_sgpr93_sgpr94_sgpr95
- ; GCN-NEXT: [[V_WRITELANE_B32_1:%[0-9]+]]:vgpr_32 = V_WRITELANE_B32 $sgpr65, 2, [[V_WRITELANE_B32_1]]
- ; GCN-NEXT: [[V_WRITELANE_B32_1:%[0-9]+]]:vgpr_32 = V_WRITELANE_B32 $sgpr66, 3, [[V_WRITELANE_B32_1]]
- ; GCN-NEXT: [[V_WRITELANE_B32_1:%[0-9]+]]:vgpr_32 = V_WRITELANE_B32 $sgpr67, 4, [[V_WRITELANE_B32_1]]
- ; GCN-NEXT: [[V_WRITELANE_B32_1:%[0-9]+]]:vgpr_32 = V_WRITELANE_B32 $sgpr68, 5, [[V_WRITELANE_B32_1]]
- ; GCN-NEXT: [[V_WRITELANE_B32_1:%[0-9]+]]:vgpr_32 = V_WRITELANE_B32 $sgpr69, 6, [[V_WRITELANE_B32_1]]
- ; GCN-NEXT: [[V_WRITELANE_B32_1:%[0-9]+]]:vgpr_32 = V_WRITELANE_B32 $sgpr70, 7, [[V_WRITELANE_B32_1]]
- ; GCN-NEXT: [[V_WRITELANE_B32_1:%[0-9]+]]:vgpr_32 = V_WRITELANE_B32 $sgpr71, 8, [[V_WRITELANE_B32_1]]
- ; GCN-NEXT: [[V_WRITELANE_B32_1:%[0-9]+]]:vgpr_32 = V_WRITELANE_B32 $sgpr72, 9, [[V_WRITELANE_B32_1]]
- ; GCN-NEXT: [[V_WRITELANE_B32_1:%[0-9]+]]:vgpr_32 = V_WRITELANE_B32 $sgpr73, 10, [[V_WRITELANE_B32_1]]
- ; GCN-NEXT: [[V_WRITELANE_B32_1:%[0-9]+]]:vgpr_32 = V_WRITELANE_B32 $sgpr74, 11, [[V_WRITELANE_B32_1]]
- ; GCN-NEXT: [[V_WRITELANE_B32_1:%[0-9]+]]:vgpr_32 = V_WRITELANE_B32 $sgpr75, 12, [[V_WRITELANE_B32_1]]
- ; GCN-NEXT: [[V_WRITELANE_B32_1:%[0-9]+]]:vgpr_32 = V_WRITELANE_B32 $sgpr76, 13, [[V_WRITELANE_B32_1]]
- ; GCN-NEXT: [[V_WRITELANE_B32_1:%[0-9]+]]:vgpr_32 = V_WRITELANE_B32 $sgpr77, 14, [[V_WRITELANE_B32_1]]
- ; GCN-NEXT: [[V_WRITELANE_B32_1:%[0-9]+]]:vgpr_32 = V_WRITELANE_B32 $sgpr78, 15, [[V_WRITELANE_B32_1]]
- ; GCN-NEXT: [[V_WRITELANE_B32_1:%[0-9]+]]:vgpr_32 = V_WRITELANE_B32 $sgpr79, 16, [[V_WRITELANE_B32_1]]
- ; GCN-NEXT: [[V_WRITELANE_B32_1:%[0-9]+]]:vgpr_32 = V_WRITELANE_B32 $sgpr80, 17, [[V_WRITELANE_B32_1]]
- ; GCN-NEXT: [[V_WRITELANE_B32_1:%[0-9]+]]:vgpr_32 = V_WRITELANE_B32 $sgpr81, 18, [[V_WRITELANE_B32_1]]
- ; GCN-NEXT: [[V_WRITELANE_B32_1:%[0-9]+]]:vgpr_32 = V_WRITELANE_B32 $sgpr82, 19, [[V_WRITELANE_B32_1]]
- ; GCN-NEXT: [[V_WRITELANE_B32_1:%[0-9]+]]:vgpr_32 = V_WRITELANE_B32 $sgpr83, 20, [[V_WRITELANE_B32_1]]
- ; GCN-NEXT: [[V_WRITELANE_B32_1:%[0-9]+]]:vgpr_32 = V_WRITELANE_B32 $sgpr84, 21, [[V_WRITELANE_B32_1]]
- ; GCN-NEXT: [[V_WRITELANE_B32_1:%[0-9]+]]:vgpr_32 = V_WRITELANE_B32 $sgpr85, 22, [[V_WRITELANE_B32_1]]
- ; GCN-NEXT: [[V_WRITELANE_B32_1:%[0-9]+]]:vgpr_32 = V_WRITELANE_B32 $sgpr86, 23, [[V_WRITELANE_B32_1]]
- ; GCN-NEXT: [[V_WRITELANE_B32_1:%[0-9]+]]:vgpr_32 = V_WRITELANE_B32 $sgpr87, 24, [[V_WRITELANE_B32_1]]
- ; GCN-NEXT: [[V_WRITELANE_B32_1:%[0-9]+]]:vgpr_32 = V_WRITELANE_B32 $sgpr88, 25, [[V_WRITELANE_B32_1]]
- ; GCN-NEXT: [[V_WRITELANE_B32_1:%[0-9]+]]:vgpr_32 = V_WRITELANE_B32 $sgpr89, 26, [[V_WRITELANE_B32_1]]
- ; GCN-NEXT: [[V_WRITELANE_B32_1:%[0-9]+]]:vgpr_32 = V_WRITELANE_B32 $sgpr90, 27, [[V_WRITELANE_B32_1]]
- ; GCN-NEXT: [[V_WRITELANE_B32_1:%[0-9]+]]:vgpr_32 = V_WRITELANE_B32 $sgpr91, 28, [[V_WRITELANE_B32_1]]
- ; GCN-NEXT: [[V_WRITELANE_B32_1:%[0-9]+]]:vgpr_32 = V_WRITELANE_B32 $sgpr92, 29, [[V_WRITELANE_B32_1]]
- ; GCN-NEXT: [[V_WRITELANE_B32_1:%[0-9]+]]:vgpr_32 = V_WRITELANE_B32 $sgpr93, 30, [[V_WRITELANE_B32_1]]
- ; GCN-NEXT: [[V_WRITELANE_B32_1:%[0-9]+]]:vgpr_32 = V_WRITELANE_B32 $sgpr94, 31, [[V_WRITELANE_B32_1]]
- ; GCN-NEXT: [[V_WRITELANE_B32_1:%[0-9]+]]:vgpr_32 = V_WRITELANE_B32 killed $sgpr95, 32, [[V_WRITELANE_B32_1]], implicit killed $sgpr64_sgpr65_sgpr66_sgpr67_sgpr68_sgpr69_sgpr70_sgpr71_sgpr72_sgpr73_sgpr74_sgpr75_sgpr76_sgpr77_sgpr78_sgpr79_sgpr80_sgpr81_sgpr82_sgpr83_sgpr84_sgpr85_sgpr86_sgpr87_sgpr88_sgpr89_sgpr90_sgpr91_sgpr92_sgpr93_sgpr94_sgpr95
- ; GCN-NEXT: S_NOP 0
- ; GCN-NEXT: $sgpr64 = V_READLANE_B32 [[V_WRITELANE_B32_1]], 1, implicit-def $sgpr64_sgpr65_sgpr66_sgpr67_sgpr68_sgpr69_sgpr70_sgpr71_sgpr72_sgpr73_sgpr74_sgpr75_sgpr76_sgpr77_sgpr78_sgpr79_sgpr80_sgpr81_sgpr82_sgpr83_sgpr84_sgpr85_sgpr86_sgpr87_sgpr88_sgpr89_sgpr90_sgpr91_sgpr92_sgpr93_sgpr94_sgpr95
- ; GCN-NEXT: $sgpr65 = V_READLANE_B32 [[V_WRITELANE_B32_1]], 2
- ; GCN-NEXT: $sgpr66 = V_READLANE_B32 [[V_WRITELANE_B32_1]], 3
- ; GCN-NEXT: $sgpr67 = V_READLANE_B32 [[V_WRITELANE_B32_1]], 4
- ; GCN-NEXT: $sgpr68 = V_READLANE_B32 [[V_WRITELANE_B32_1]], 5
- ; GCN-NEXT: $sgpr69 = V_READLANE_B32 [[V_WRITELANE_B32_1]], 6
- ; GCN-NEXT: $sgpr70 = V_READLANE_B32 [[V_WRITELANE_B32_1]], 7
- ; GCN-NEXT: $sgpr71 = V_READLANE_B32 [[V_WRITELANE_B32_1]], 8
- ; GCN-NEXT: $sgpr72 = V_READLANE_B32 [[V_WRITELANE_B32_1]], 9
- ; GCN-NEXT: $sgpr73 = V_READLANE_B32 [[V_WRITELANE_B32_1]], 10
- ; GCN-NEXT: $sgpr74 = V_READLANE_B32 [[V_WRITELANE_B32_1]], 11
- ; GCN-NEXT: $sgpr75 = V_READLANE_B32 [[V_WRITELANE_B32_1]], 12
- ; GCN-NEXT: $sgpr76 = V_READLANE_B32 [[V_WRITELANE_B32_1]], 13
- ; GCN-NEXT: $sgpr77 = V_READLANE_B32 [[V_WRITELANE_B32_1]], 14
- ; GCN-NEXT: $sgpr78 = V_READLANE_B32 [[V_WRITELANE_B32_1]], 15
- ; GCN-NEXT: $sgpr79 = V_READLANE_B32 [[V_WRITELANE_B32_1]], 16
- ; GCN-NEXT: $sgpr80 = V_READLANE_B32 [[V_WRITELANE_B32_1]], 17
- ; GCN-NEXT: $sgpr81 = V_READLANE_B32 [[V_WRITELANE_B32_1]], 18
- ; GCN-NEXT: $sgpr82 = V_READLANE_B32 [[V_WRITELANE_B32_1]], 19
- ; GCN-NEXT: $sgpr83 = V_READLANE_B32 [[V_WRITELANE_B32_1]], 20
- ; GCN-NEXT: $sgpr84 = V_READLANE_B32 [[V_WRITELANE_B32_1]], 21
- ; GCN-NEXT: $sgpr85 = V_READLANE_B32 [[V_WRITELANE_B32_1]], 22
- ; GCN-NEXT: $sgpr86 = V_READLANE_B32 [[V_WRITELANE_B32_1]], 23
- ; GCN-NEXT: $sgpr87 = V_READLANE_B32 [[V_WRITELANE_B32_1]], 24
- ; GCN-NEXT: $sgpr88 = V_READLANE_B32 [[V_WRITELANE_B32_1]], 25
- ; GCN-NEXT: $sgpr89 = V_READLANE_B32 [[V_WRITELANE_B32_1]], 26
- ; GCN-NEXT: $sgpr90 = V_READLANE_B32 [[V_WRITELANE_B32_1]], 27
- ; GCN-NEXT: $sgpr91 = V_READLANE_B32 [[V_WRITELANE_B32_1]], 28
- ; GCN-NEXT: $sgpr92 = V_READLANE_B32 [[V_WRITELANE_B32_1]], 29
- ; GCN-NEXT: $sgpr93 = V_READLANE_B32 [[V_WRITELANE_B32_1]], 30
- ; GCN-NEXT: $sgpr94 = V_READLANE_B32 [[V_WRITELANE_B32_1]], 31
- ; GCN-NEXT: $sgpr95 = V_READLANE_B32 [[V_WRITELANE_B32_1]], 32
- ; GCN-NEXT: $sgpr10 = V_READLANE_B32 [[V_WRITELANE_B32_1]], 0
- ; GCN-NEXT: S_SETPC_B64 $sgpr30_sgpr31
- S_NOP 0
- SI_SPILL_S32_SAVE killed $sgpr10, %stack.0, implicit $exec, implicit $sgpr0_sgpr1_sgpr2_sgpr3, implicit $sgpr32
- SI_SPILL_S1024_SAVE killed $sgpr64_sgpr65_sgpr66_sgpr67_sgpr68_sgpr69_sgpr70_sgpr71_sgpr72_sgpr73_sgpr74_sgpr75_sgpr76_sgpr77_sgpr78_sgpr79_sgpr80_sgpr81_sgpr82_sgpr83_sgpr84_sgpr85_sgpr86_sgpr87_sgpr88_sgpr89_sgpr90_sgpr91_sgpr92_sgpr93_sgpr94_sgpr95, %stack.1, implicit $exec, implicit $sgpr0_sgpr1_sgpr2_sgpr3, implicit $sgpr32
- S_NOP 0
- renamable $sgpr64_sgpr65_sgpr66_sgpr67_sgpr68_sgpr69_sgpr70_sgpr71_sgpr72_sgpr73_sgpr74_sgpr75_sgpr76_sgpr77_sgpr78_sgpr79_sgpr80_sgpr81_sgpr82_sgpr83_sgpr84_sgpr85_sgpr86_sgpr87_sgpr88_sgpr89_sgpr90_sgpr91_sgpr92_sgpr93_sgpr94_sgpr95 = SI_SPILL_S1024_RESTORE %stack.1, implicit $exec, implicit $sgpr0_sgpr1_sgpr2_sgpr3, implicit $sgpr32
- renamable $sgpr10 = SI_SPILL_S32_RESTORE %stack.0, implicit $exec, implicit $sgpr0_sgpr1_sgpr2_sgpr3, implicit $sgpr32
- S_SETPC_B64 $sgpr30_sgpr31
-...
-
-# The implicit def for the lane VGPR should be inserted at the common dominator block (the entry block here).
-
----
-name: lane_vgpr_implicit_def_at_common_dominator_block
-tracksRegLiveness: true
-frameInfo:
- maxAlignment: 4
-stack:
- - { id: 0, type: spill-slot, size: 4, alignment: 4, stack-id: sgpr-spill }
-machineFunctionInfo:
- isEntryFunction: false
- scratchRSrcReg: '$sgpr0_sgpr1_sgpr2_sgpr3'
- stackPtrOffsetReg: '$sgpr32'
- frameOffsetReg: '$sgpr33'
- hasSpilledSGPRs: true
-body: |
- ; GCN-LABEL: name: lane_vgpr_implicit_def_at_common_dominator_block
- ; GCN: bb.0:
- ; GCN-NEXT: successors: %bb.2(0x40000000), %bb.1(0x40000000)
- ; GCN-NEXT: liveins: $sgpr10, $sgpr11, $sgpr30_sgpr31
- ; GCN-NEXT: {{ $}}
- ; GCN-NEXT: [[DEF:%[0-9]+]]:vgpr_32 = IMPLICIT_DEF
- ; GCN-NEXT: S_NOP 0
- ; GCN-NEXT: S_CMP_EQ_U32 $sgpr11, 0, implicit-def $scc
- ; GCN-NEXT: S_CBRANCH_SCC1 %bb.2, implicit killed $scc
- ; GCN-NEXT: {{ $}}
- ; GCN-NEXT: bb.1:
- ; GCN-NEXT: successors: %bb.3(0x80000000)
- ; GCN-NEXT: liveins: $sgpr10, $sgpr30_sgpr31
- ; GCN-NEXT: {{ $}}
- ; GCN-NEXT: $sgpr10 = S_MOV_B32 10
- ; GCN-NEXT: [[V_WRITELANE_B32_:%[0-9]+]]:vgpr_32 = V_WRITELANE_B32 killed $sgpr10, 0, [[V_WRITELANE_B32_]]
- ; GCN-NEXT: S_BRANCH %bb.3
- ; GCN-NEXT: {{ $}}
- ; GCN-NEXT: bb.2:
- ; GCN-NEXT: successors: %bb.3(0x80000000)
- ; GCN-NEXT: liveins: $sgpr10, $sgpr30_sgpr31
- ; GCN-NEXT: {{ $}}
- ; GCN-NEXT: $sgpr10 = S_MOV_B32 20
- ; GCN-NEXT: [[V_WRITELANE_B32_1:%[0-9]+]]:vgpr_32 = V_WRITELANE_B32 killed $sgpr10, 0, [[V_WRITELANE_B32_1]]
- ; GCN-NEXT: S_BRANCH %bb.3
- ; GCN-NEXT: {{ $}}
- ; GCN-NEXT: bb.3:
- ; GCN-NEXT: liveins: $sgpr10, $sgpr30_sgpr31
- ; GCN-NEXT: {{ $}}
- ; GCN-NEXT: $sgpr10 = V_READLANE_B32 [[V_WRITELANE_B32_1]], 0
- ; GCN-NEXT: S_SETPC_B64 $sgpr30_sgpr31, implicit $sgpr10
- bb.0:
- liveins: $sgpr10, $sgpr11, $sgpr30_sgpr31
- S_NOP 0
- S_CMP_EQ_U32 $sgpr11, 0, implicit-def $scc
- S_CBRANCH_SCC1 %bb.2, implicit killed $scc
- bb.1:
- liveins: $sgpr10, $sgpr30_sgpr31
- $sgpr10 = S_MOV_B32 10
- SI_SPILL_S32_SAVE killed $sgpr10, %stack.0, implicit $exec, implicit $sgpr0_sgpr1_sgpr2_sgpr3, implicit $sgpr32
- S_BRANCH %bb.3
- bb.2:
- liveins: $sgpr10, $sgpr30_sgpr31
- $sgpr10 = S_MOV_B32 20
- SI_SPILL_S32_SAVE killed $sgpr10, %stack.0, implicit $exec, implicit $sgpr0_sgpr1_sgpr2_sgpr3, implicit $sgpr32
- S_BRANCH %bb.3
- bb.3:
- liveins: $sgpr10, $sgpr30_sgpr31
- renamable $sgpr10 = SI_SPILL_S32_RESTORE %stack.0, implicit $exec, implicit $sgpr0_sgpr1_sgpr2_sgpr3, implicit $sgpr32
- S_SETPC_B64 $sgpr30_sgpr31, implicit $sgpr10
-...
-
-# The common dominator block is visited only at the end. The insertion point was initially identified to the
-# terminator instruction in the dominator block which later becomes the point where a spill get inserted in the same block.
-
----
-name: dominator_block_follows_the_successors_bbs
-tracksRegLiveness: true
-frameInfo:
- maxAlignment: 4
-stack:
- - { id: 0, type: spill-slot, size: 4, alignment: 4, stack-id: sgpr-spill }
-machineFunctionInfo:
- isEntryFunction: false
- scratchRSrcReg: '$sgpr0_sgpr1_sgpr2_sgpr3'
- stackPtrOffsetReg: '$sgpr32'
- frameOffsetReg: '$sgpr33'
- hasSpilledSGPRs: true
-body: |
- ; GCN-LABEL: name: dominator_block_follows_the_successors_bbs
- ; GCN: bb.0:
- ; GCN-NEXT: successors: %bb.3(0x80000000)
- ; GCN-NEXT: liveins: $sgpr10, $sgpr11, $sgpr30_sgpr31
- ; GCN-NEXT: {{ $}}
- ; GCN-NEXT: [[DEF:%[0-9]+]]:vgpr_32 = IMPLICIT_DEF
- ; GCN-NEXT: S_NOP 0
- ; GCN-NEXT: S_BRANCH %bb.3
- ; GCN-NEXT: {{ $}}
- ; GCN-NEXT: bb.1:
- ; GCN-NEXT: successors: %bb.2(0x80000000)
- ; GCN-NEXT: liveins: $sgpr10, $sgpr30_sgpr31
- ; GCN-NEXT: {{ $}}
- ; GCN-NEXT: $sgpr10 = V_READLANE_B32 [[DEF]], 0
- ; GCN-NEXT: $sgpr10 = S_ADD_I32 $sgpr10, 15, implicit-def dead $scc
- ; GCN-NEXT: S_BRANCH %bb.2
- ; GCN-NEXT: {{ $}}
- ; GCN-NEXT: bb.2:
- ; GCN-NEXT: successors: %bb.3(0x80000000)
- ; GCN-NEXT: liveins: $sgpr10, $sgpr30_sgpr31
- ; GCN-NEXT: {{ $}}
- ; GCN-NEXT: $sgpr10 = V_READLANE_B32 [[DEF]], 0
- ; GCN-NEXT: $sgpr10 = S_ADD_I32 $sgpr10, 20, implicit-def dead $scc
- ; GCN-NEXT: S_BRANCH %bb.3
- ; GCN-NEXT: {{ $}}
- ; GCN-NEXT: bb.3:
- ; GCN-NEXT: successors: %bb.2(0x40000000), %bb.1(0x40000000)
- ; GCN-NEXT: liveins: $sgpr10, $sgpr11, $sgpr30_sgpr31
- ; GCN-NEXT: {{ $}}
- ; GCN-NEXT: $sgpr10 = S_MOV_B32 10
- ; GCN-NEXT: [[V_WRITELANE_B32_:%[0-9]+]]:vgpr_32 = V_WRITELANE_B32 killed $sgpr10, 0, [[V_WRITELANE_B32_]]
- ; GCN-NEXT: S_CMP_EQ_U32 $sgpr11, 0, implicit-def $scc
- ; GCN-NEXT: S_CBRANCH_SCC1 %bb.2, implicit killed $scc
- ; GCN-NEXT: S_BRANCH %bb.1
- ; GCN-NEXT: {{ $}}
- ; GCN-NEXT: bb.4:
- ; GCN-NEXT: liveins: $sgpr10, $sgpr30_sgpr31
- ; GCN-NEXT: {{ $}}
- ; GCN-NEXT: S_NOP 0
- ; GCN-NEXT: S_SETPC_B64 $sgpr30_sgpr31, implicit $sgpr10
- bb.0:
- liveins: $sgpr10, $sgpr11, $sgpr30_sgpr31
- S_NOP 0
- S_BRANCH %bb.3
- bb.1:
- liveins: $sgpr10, $sgpr30_sgpr31
- renamable $sgpr10 = SI_SPILL_S32_RESTORE %stack.0, implicit $exec, implicit $sgpr0_sgpr1_sgpr2_sgpr3, implicit $sgpr32
- $sgpr10 = S_ADD_I32 $sgpr10, 15, implicit-def dead $scc
- S_BRANCH %bb.2
- bb.2:
- liveins: $sgpr10, $sgpr30_sgpr31
- renamable $sgpr10 = SI_SPILL_S32_RESTORE %stack.0, implicit $exec, implicit $sgpr0_sgpr1_sgpr2_sgpr3, implicit $sgpr32
- $sgpr10 = S_ADD_I32 $sgpr10, 20, implicit-def dead $scc
- S_BRANCH %bb.3
- bb.3:
- liveins: $sgpr10, $sgpr11, $sgpr30_sgpr31
- $sgpr10 = S_MOV_B32 10
- SI_SPILL_S32_SAVE killed $sgpr10, %stack.0, implicit $exec, implicit $sgpr0_sgpr1_sgpr2_sgpr3, implicit $sgpr32
- S_CMP_EQ_U32 $sgpr11, 0, implicit-def $scc
- S_CBRANCH_SCC1 %bb.2, implicit killed $scc
- S_BRANCH %bb.1
- bb.4:
- liveins: $sgpr10, $sgpr30_sgpr31
- S_NOP 0
- S_SETPC_B64 $sgpr30_sgpr31, implicit $sgpr10
-...
diff --git a/llvm/test/CodeGen/AMDGPU/spill-vgpr-to-agpr-update-regscavenger.ll b/llvm/test/CodeGen/AMDGPU/spill-vgpr-to-agpr-update-regscavenger.ll
deleted file mode 100644
--- a/llvm/test/CodeGen/AMDGPU/spill-vgpr-to-agpr-update-regscavenger.ll
+++ /dev/null
@@ -1,92 +0,0 @@
-; NOTE: Assertions have been autogenerated by utils/update_llc_test_checks.py
-; RUN: llc -mtriple=amdgcn-amd-amdhsa -mcpu=gfx908 -O0 -verify-machineinstrs -o - %s | FileCheck %s
-
-; Regression test for `processFunctionBeforeFrameFinalized`:
-; Check that it correctly updates RegisterScavenger so we
-; don't end up with bad machine code due to using undefined
-; physical registers.
-
-define void @test() {
-; CHECK-LABEL: test:
-; CHECK: ; %bb.0: ; %bb.0
-; CHECK-NEXT: s_waitcnt vmcnt(0) expcnt(0) lgkmcnt(0)
-; CHECK-NEXT: s_xor_saveexec_b64 s[4:5], -1
-; CHECK-NEXT: buffer_store_dword v0, off, s[0:3], s32 ; 4-byte Folded Spill
-; CHECK-NEXT: s_mov_b64 exec, s[4:5]
-; CHECK-NEXT: ; implicit-def: $vgpr0
-; CHECK-NEXT: .LBB0_1: ; %bb.1
-; CHECK-NEXT: ; =>This Inner Loop Header: Depth=1
-; CHECK-NEXT: s_cbranch_scc1 .LBB0_3
-; CHECK-NEXT: ; %bb.2: ; %bb.2
-; CHECK-NEXT: ; in Loop: Header=BB0_1 Depth=1
-; CHECK-NEXT: .LBB0_3: ; %bb.3
-; CHECK-NEXT: ; in Loop: Header=BB0_1 Depth=1
-; CHECK-NEXT: s_or_saveexec_b64 s[10:11], -1
-; CHECK-NEXT: v_accvgpr_read_b32 v0, a0 ; Reload Reuse
-; CHECK-NEXT: s_mov_b64 exec, s[10:11]
-; CHECK-NEXT: ; implicit-def: $sgpr4
-; CHECK-NEXT: v_mov_b32_e32 v1, s4
-; CHECK-NEXT: v_readfirstlane_b32 s6, v1
-; CHECK-NEXT: s_mov_b64 s[4:5], -1
-; CHECK-NEXT: s_mov_b32 s7, 0
-; CHECK-NEXT: s_cmp_eq_u32 s6, s7
-; CHECK-NEXT: v_writelane_b32 v0, s4, 0
-; CHECK-NEXT: v_writelane_b32 v0, s5, 1
-; CHECK-NEXT: s_mov_b64 s[10:11], exec
-; CHECK-NEXT: s_mov_b64 exec, -1
-; CHECK-NEXT: v_accvgpr_write_b32 a0, v0 ; Reload Reuse
-; CHECK-NEXT: s_mov_b64 exec, s[10:11]
-; CHECK-NEXT: s_cbranch_scc1 .LBB0_5
-; CHECK-NEXT: ; %bb.4: ; %bb.4
-; CHECK-NEXT: ; in Loop: Header=BB0_1 Depth=1
-; CHECK-NEXT: s_or_saveexec_b64 s[10:11], -1
-; CHECK-NEXT: v_accvgpr_read_b32 v0, a0 ; Reload Reuse
-; CHECK-NEXT: s_mov_b64 exec, s[10:11]
-; CHECK-NEXT: s_mov_b64 s[4:5], 0
-; CHECK-NEXT: v_writelane_b32 v0, s4, 0
-; CHECK-NEXT: v_writelane_b32 v0, s5, 1
-; CHECK-NEXT: s_or_saveexec_b64 s[10:11], -1
-; CHECK-NEXT: s_nop 0
-; CHECK-NEXT: v_accvgpr_write_b32 a0, v0 ; Reload Reuse
-; CHECK-NEXT: s_mov_b64 exec, s[10:11]
-; CHECK-NEXT: .LBB0_5: ; %Flow
-; CHECK-NEXT: ; in Loop: Header=BB0_1 Depth=1
-; CHECK-NEXT: s_or_saveexec_b64 s[10:11], -1
-; CHECK-NEXT: s_nop 0
-; CHECK-NEXT: v_accvgpr_read_b32 v0, a0 ; Reload Reuse
-; CHECK-NEXT: s_mov_b64 exec, s[10:11]
-; CHECK-NEXT: v_readlane_b32 s4, v0, 0
-; CHECK-NEXT: v_readlane_b32 s5, v0, 1
-; CHECK-NEXT: v_cndmask_b32_e64 v0, 0, 1, s[4:5]
-; CHECK-NEXT: s_mov_b32 s4, 1
-; CHECK-NEXT: ; implicit-def: $sgpr5
-; CHECK-NEXT: v_cmp_ne_u32_e64 s[4:5], v0, s4
-; CHECK-NEXT: s_and_b64 vcc, exec, s[4:5]
-; CHECK-NEXT: s_cbranch_vccnz .LBB0_1
-; CHECK-NEXT: ; %bb.6: ; %bb.5
-; CHECK-NEXT: s_or_saveexec_b64 s[10:11], -1
-; CHECK-NEXT: v_accvgpr_read_b32 v0, a0 ; Reload Reuse
-; CHECK-NEXT: s_mov_b64 exec, s[10:11]
-; CHECK-NEXT: ; kill: killed $vgpr0
-; CHECK-NEXT: s_xor_saveexec_b64 s[4:5], -1
-; CHECK-NEXT: buffer_load_dword v0, off, s[0:3], s32 ; 4-byte Folded Reload
-; CHECK-NEXT: s_mov_b64 exec, s[4:5]
-; CHECK-NEXT: s_waitcnt vmcnt(0)
-; CHECK-NEXT: s_setpc_b64 s[30:31]
-bb.0:
- br label %bb.1
-bb.1: ; preds = %bb.4, %bb.0
- br i1 poison, label %bb.2, label %bb.3
-bb.2: ; preds = %bb.1
- br label %bb.3
-bb.3: ; preds = %bb.2, %bb.1
- %call = tail call i32 @llvm.amdgcn.readfirstlane(i32 poison)
- %cmp = icmp eq i32 %call, 0
- br i1 %cmp, label %bb.5, label %bb.4
-bb.4: ; preds = %bb.3
- br label %bb.1
-bb.5: ; preds = %bb.3
- ret void
-}
-
-declare i32 @llvm.amdgcn.readfirstlane(i32)
diff --git a/llvm/test/CodeGen/AMDGPU/spill-writelane-vgprs.ll b/llvm/test/CodeGen/AMDGPU/spill-writelane-vgprs.ll
--- a/llvm/test/CodeGen/AMDGPU/spill-writelane-vgprs.ll
+++ b/llvm/test/CodeGen/AMDGPU/spill-writelane-vgprs.ll
@@ -10,16 +10,16 @@
; GCN-LABEL: sgpr_spill_writelane:
; GCN: ; %bb.0:
; GCN-NEXT: s_waitcnt vmcnt(0) expcnt(0) lgkmcnt(0)
-; GCN-NEXT: s_xor_saveexec_b64 s[4:5], -1
+; GCN-NEXT: s_xor_saveexec_b64 s[6:7], -1
; GCN-NEXT: buffer_store_dword v0, off, s[0:3], s32 ; 4-byte Folded Spill
-; GCN-NEXT: s_mov_b64 exec, s[4:5]
+; GCN-NEXT: s_mov_b64 exec, s[6:7]
; GCN-NEXT: v_writelane_b32 v0, s35, 0
; GCN-NEXT: ;;#ASMSTART
; GCN-NEXT: ;;#ASMEND
; GCN-NEXT: v_readlane_b32 s35, v0, 0
-; GCN-NEXT: s_xor_saveexec_b64 s[4:5], -1
+; GCN-NEXT: s_xor_saveexec_b64 s[6:7], -1
; GCN-NEXT: buffer_load_dword v0, off, s[0:3], s32 ; 4-byte Folded Reload
-; GCN-NEXT: s_mov_b64 exec, s[4:5]
+; GCN-NEXT: s_mov_b64 exec, s[6:7]
; GCN-NEXT: s_waitcnt vmcnt(0)
; GCN-NEXT: s_setpc_b64 s[30:31]
call void asm sideeffect "", "~{s35}"()
diff --git a/llvm/test/CodeGen/AMDGPU/spill192.mir b/llvm/test/CodeGen/AMDGPU/spill192.mir
--- a/llvm/test/CodeGen/AMDGPU/spill192.mir
+++ b/llvm/test/CodeGen/AMDGPU/spill192.mir
@@ -32,29 +32,32 @@
; EXPANDED-LABEL: name: spill_restore_sgpr192
; EXPANDED: bb.0:
; EXPANDED-NEXT: successors: %bb.1(0x80000000)
+ ; EXPANDED-NEXT: liveins: $vgpr0
; EXPANDED-NEXT: {{ $}}
- ; EXPANDED-NEXT: [[DEF:%[0-9]+]]:vgpr_32 = IMPLICIT_DEF
; EXPANDED-NEXT: S_NOP 0, implicit-def renamable $sgpr4_sgpr5_sgpr6_sgpr7_sgpr8_sgpr9
- ; EXPANDED-NEXT: [[V_WRITELANE_B32_:%[0-9]+]]:vgpr_32 = V_WRITELANE_B32 $sgpr4, 0, [[V_WRITELANE_B32_]], implicit-def $sgpr4_sgpr5_sgpr6_sgpr7_sgpr8_sgpr9, implicit $sgpr4_sgpr5_sgpr6_sgpr7_sgpr8_sgpr9
- ; EXPANDED-NEXT: [[V_WRITELANE_B32_1:%[0-9]+]]:vgpr_32 = V_WRITELANE_B32 $sgpr5, 1, [[V_WRITELANE_B32_1]]
- ; EXPANDED-NEXT: [[V_WRITELANE_B32_1:%[0-9]+]]:vgpr_32 = V_WRITELANE_B32 $sgpr6, 2, [[V_WRITELANE_B32_1]]
- ; EXPANDED-NEXT: [[V_WRITELANE_B32_1:%[0-9]+]]:vgpr_32 = V_WRITELANE_B32 $sgpr7, 3, [[V_WRITELANE_B32_1]]
- ; EXPANDED-NEXT: [[V_WRITELANE_B32_1:%[0-9]+]]:vgpr_32 = V_WRITELANE_B32 $sgpr8, 4, [[V_WRITELANE_B32_1]]
- ; EXPANDED-NEXT: [[V_WRITELANE_B32_1:%[0-9]+]]:vgpr_32 = V_WRITELANE_B32 killed $sgpr9, 5, [[V_WRITELANE_B32_1]], implicit killed $sgpr4_sgpr5_sgpr6_sgpr7_sgpr8_sgpr9
+ ; EXPANDED-NEXT: $vgpr0 = V_WRITELANE_B32 $sgpr4, 0, $vgpr0, implicit-def $sgpr4_sgpr5_sgpr6_sgpr7_sgpr8_sgpr9, implicit $sgpr4_sgpr5_sgpr6_sgpr7_sgpr8_sgpr9
+ ; EXPANDED-NEXT: $vgpr0 = V_WRITELANE_B32 $sgpr5, 1, $vgpr0
+ ; EXPANDED-NEXT: $vgpr0 = V_WRITELANE_B32 $sgpr6, 2, $vgpr0
+ ; EXPANDED-NEXT: $vgpr0 = V_WRITELANE_B32 $sgpr7, 3, $vgpr0
+ ; EXPANDED-NEXT: $vgpr0 = V_WRITELANE_B32 $sgpr8, 4, $vgpr0
+ ; EXPANDED-NEXT: $vgpr0 = V_WRITELANE_B32 killed $sgpr9, 5, $vgpr0, implicit killed $sgpr4_sgpr5_sgpr6_sgpr7_sgpr8_sgpr9
; EXPANDED-NEXT: S_CBRANCH_SCC1 %bb.1, implicit undef $scc
; EXPANDED-NEXT: {{ $}}
; EXPANDED-NEXT: bb.1:
; EXPANDED-NEXT: successors: %bb.2(0x80000000)
+ ; EXPANDED-NEXT: liveins: $vgpr0
; EXPANDED-NEXT: {{ $}}
; EXPANDED-NEXT: S_NOP 1
; EXPANDED-NEXT: {{ $}}
; EXPANDED-NEXT: bb.2:
- ; EXPANDED-NEXT: $sgpr4 = V_READLANE_B32 [[V_WRITELANE_B32_1]], 0, implicit-def $sgpr4_sgpr5_sgpr6_sgpr7_sgpr8_sgpr9
- ; EXPANDED-NEXT: $sgpr5 = V_READLANE_B32 [[V_WRITELANE_B32_1]], 1
- ; EXPANDED-NEXT: $sgpr6 = V_READLANE_B32 [[V_WRITELANE_B32_1]], 2
- ; EXPANDED-NEXT: $sgpr7 = V_READLANE_B32 [[V_WRITELANE_B32_1]], 3
- ; EXPANDED-NEXT: $sgpr8 = V_READLANE_B32 [[V_WRITELANE_B32_1]], 4
- ; EXPANDED-NEXT: $sgpr9 = V_READLANE_B32 [[V_WRITELANE_B32_1]], 5
+ ; EXPANDED-NEXT: liveins: $vgpr0
+ ; EXPANDED-NEXT: {{ $}}
+ ; EXPANDED-NEXT: $sgpr4 = V_READLANE_B32 $vgpr0, 0, implicit-def $sgpr4_sgpr5_sgpr6_sgpr7_sgpr8_sgpr9
+ ; EXPANDED-NEXT: $sgpr5 = V_READLANE_B32 $vgpr0, 1
+ ; EXPANDED-NEXT: $sgpr6 = V_READLANE_B32 $vgpr0, 2
+ ; EXPANDED-NEXT: $sgpr7 = V_READLANE_B32 $vgpr0, 3
+ ; EXPANDED-NEXT: $sgpr8 = V_READLANE_B32 $vgpr0, 4
+ ; EXPANDED-NEXT: $sgpr9 = V_READLANE_B32 $vgpr0, 5
; EXPANDED-NEXT: S_NOP 0, implicit killed renamable $sgpr4_sgpr5_sgpr6_sgpr7_sgpr8_sgpr9
bb.0:
S_NOP 0, implicit-def %0:sgpr_192
diff --git a/llvm/test/CodeGen/AMDGPU/spill224.mir b/llvm/test/CodeGen/AMDGPU/spill224.mir
--- a/llvm/test/CodeGen/AMDGPU/spill224.mir
+++ b/llvm/test/CodeGen/AMDGPU/spill224.mir
@@ -30,31 +30,34 @@
; EXPANDED-LABEL: name: spill_restore_sgpr224
; EXPANDED: bb.0:
; EXPANDED-NEXT: successors: %bb.1(0x80000000)
+ ; EXPANDED-NEXT: liveins: $vgpr0
; EXPANDED-NEXT: {{ $}}
- ; EXPANDED-NEXT: [[DEF:%[0-9]+]]:vgpr_32 = IMPLICIT_DEF
; EXPANDED-NEXT: S_NOP 0, implicit-def renamable $sgpr4_sgpr5_sgpr6_sgpr7_sgpr8_sgpr9_sgpr10
- ; EXPANDED-NEXT: [[V_WRITELANE_B32_:%[0-9]+]]:vgpr_32 = V_WRITELANE_B32 $sgpr4, 0, [[V_WRITELANE_B32_]], implicit-def $sgpr4_sgpr5_sgpr6_sgpr7_sgpr8_sgpr9_sgpr10, implicit $sgpr4_sgpr5_sgpr6_sgpr7_sgpr8_sgpr9_sgpr10
- ; EXPANDED-NEXT: [[V_WRITELANE_B32_1:%[0-9]+]]:vgpr_32 = V_WRITELANE_B32 $sgpr5, 1, [[V_WRITELANE_B32_1]]
- ; EXPANDED-NEXT: [[V_WRITELANE_B32_1:%[0-9]+]]:vgpr_32 = V_WRITELANE_B32 $sgpr6, 2, [[V_WRITELANE_B32_1]]
- ; EXPANDED-NEXT: [[V_WRITELANE_B32_1:%[0-9]+]]:vgpr_32 = V_WRITELANE_B32 $sgpr7, 3, [[V_WRITELANE_B32_1]]
- ; EXPANDED-NEXT: [[V_WRITELANE_B32_1:%[0-9]+]]:vgpr_32 = V_WRITELANE_B32 $sgpr8, 4, [[V_WRITELANE_B32_1]]
- ; EXPANDED-NEXT: [[V_WRITELANE_B32_1:%[0-9]+]]:vgpr_32 = V_WRITELANE_B32 $sgpr9, 5, [[V_WRITELANE_B32_1]]
- ; EXPANDED-NEXT: [[V_WRITELANE_B32_1:%[0-9]+]]:vgpr_32 = V_WRITELANE_B32 killed $sgpr10, 6, [[V_WRITELANE_B32_1]], implicit killed $sgpr4_sgpr5_sgpr6_sgpr7_sgpr8_sgpr9_sgpr10
+ ; EXPANDED-NEXT: $vgpr0 = V_WRITELANE_B32 $sgpr4, 0, $vgpr0, implicit-def $sgpr4_sgpr5_sgpr6_sgpr7_sgpr8_sgpr9_sgpr10, implicit $sgpr4_sgpr5_sgpr6_sgpr7_sgpr8_sgpr9_sgpr10
+ ; EXPANDED-NEXT: $vgpr0 = V_WRITELANE_B32 $sgpr5, 1, $vgpr0
+ ; EXPANDED-NEXT: $vgpr0 = V_WRITELANE_B32 $sgpr6, 2, $vgpr0
+ ; EXPANDED-NEXT: $vgpr0 = V_WRITELANE_B32 $sgpr7, 3, $vgpr0
+ ; EXPANDED-NEXT: $vgpr0 = V_WRITELANE_B32 $sgpr8, 4, $vgpr0
+ ; EXPANDED-NEXT: $vgpr0 = V_WRITELANE_B32 $sgpr9, 5, $vgpr0
+ ; EXPANDED-NEXT: $vgpr0 = V_WRITELANE_B32 killed $sgpr10, 6, $vgpr0, implicit killed $sgpr4_sgpr5_sgpr6_sgpr7_sgpr8_sgpr9_sgpr10
; EXPANDED-NEXT: S_CBRANCH_SCC1 %bb.1, implicit undef $scc
; EXPANDED-NEXT: {{ $}}
; EXPANDED-NEXT: bb.1:
; EXPANDED-NEXT: successors: %bb.2(0x80000000)
+ ; EXPANDED-NEXT: liveins: $vgpr0
; EXPANDED-NEXT: {{ $}}
; EXPANDED-NEXT: S_NOP 1
; EXPANDED-NEXT: {{ $}}
; EXPANDED-NEXT: bb.2:
- ; EXPANDED-NEXT: $sgpr4 = V_READLANE_B32 [[V_WRITELANE_B32_1]], 0, implicit-def $sgpr4_sgpr5_sgpr6_sgpr7_sgpr8_sgpr9_sgpr10
- ; EXPANDED-NEXT: $sgpr5 = V_READLANE_B32 [[V_WRITELANE_B32_1]], 1
- ; EXPANDED-NEXT: $sgpr6 = V_READLANE_B32 [[V_WRITELANE_B32_1]], 2
- ; EXPANDED-NEXT: $sgpr7 = V_READLANE_B32 [[V_WRITELANE_B32_1]], 3
- ; EXPANDED-NEXT: $sgpr8 = V_READLANE_B32 [[V_WRITELANE_B32_1]], 4
- ; EXPANDED-NEXT: $sgpr9 = V_READLANE_B32 [[V_WRITELANE_B32_1]], 5
- ; EXPANDED-NEXT: $sgpr10 = V_READLANE_B32 [[V_WRITELANE_B32_1]], 6
+ ; EXPANDED-NEXT: liveins: $vgpr0
+ ; EXPANDED-NEXT: {{ $}}
+ ; EXPANDED-NEXT: $sgpr4 = V_READLANE_B32 $vgpr0, 0, implicit-def $sgpr4_sgpr5_sgpr6_sgpr7_sgpr8_sgpr9_sgpr10
+ ; EXPANDED-NEXT: $sgpr5 = V_READLANE_B32 $vgpr0, 1
+ ; EXPANDED-NEXT: $sgpr6 = V_READLANE_B32 $vgpr0, 2
+ ; EXPANDED-NEXT: $sgpr7 = V_READLANE_B32 $vgpr0, 3
+ ; EXPANDED-NEXT: $sgpr8 = V_READLANE_B32 $vgpr0, 4
+ ; EXPANDED-NEXT: $sgpr9 = V_READLANE_B32 $vgpr0, 5
+ ; EXPANDED-NEXT: $sgpr10 = V_READLANE_B32 $vgpr0, 6
; EXPANDED-NEXT: S_NOP 0, implicit killed renamable $sgpr4_sgpr5_sgpr6_sgpr7_sgpr8_sgpr9_sgpr10
bb.0:
S_NOP 0, implicit-def %0:sgpr_224
diff --git a/llvm/test/CodeGen/AMDGPU/spill288.mir b/llvm/test/CodeGen/AMDGPU/spill288.mir
--- a/llvm/test/CodeGen/AMDGPU/spill288.mir
+++ b/llvm/test/CodeGen/AMDGPU/spill288.mir
@@ -30,35 +30,38 @@
; EXPANDED-LABEL: name: spill_restore_sgpr288
; EXPANDED: bb.0:
; EXPANDED-NEXT: successors: %bb.1(0x80000000)
+ ; EXPANDED-NEXT: liveins: $vgpr0
; EXPANDED-NEXT: {{ $}}
- ; EXPANDED-NEXT: [[DEF:%[0-9]+]]:vgpr_32 = IMPLICIT_DEF
; EXPANDED-NEXT: S_NOP 0, implicit-def renamable $sgpr4_sgpr5_sgpr6_sgpr7_sgpr8_sgpr9_sgpr10_sgpr11_sgpr12
- ; EXPANDED-NEXT: [[V_WRITELANE_B32_:%[0-9]+]]:vgpr_32 = V_WRITELANE_B32 $sgpr4, 0, [[V_WRITELANE_B32_]], implicit-def $sgpr4_sgpr5_sgpr6_sgpr7_sgpr8_sgpr9_sgpr10_sgpr11_sgpr12, implicit $sgpr4_sgpr5_sgpr6_sgpr7_sgpr8_sgpr9_sgpr10_sgpr11_sgpr12
- ; EXPANDED-NEXT: [[V_WRITELANE_B32_1:%[0-9]+]]:vgpr_32 = V_WRITELANE_B32 $sgpr5, 1, [[V_WRITELANE_B32_1]]
- ; EXPANDED-NEXT: [[V_WRITELANE_B32_1:%[0-9]+]]:vgpr_32 = V_WRITELANE_B32 $sgpr6, 2, [[V_WRITELANE_B32_1]]
- ; EXPANDED-NEXT: [[V_WRITELANE_B32_1:%[0-9]+]]:vgpr_32 = V_WRITELANE_B32 $sgpr7, 3, [[V_WRITELANE_B32_1]]
- ; EXPANDED-NEXT: [[V_WRITELANE_B32_1:%[0-9]+]]:vgpr_32 = V_WRITELANE_B32 $sgpr8, 4, [[V_WRITELANE_B32_1]]
- ; EXPANDED-NEXT: [[V_WRITELANE_B32_1:%[0-9]+]]:vgpr_32 = V_WRITELANE_B32 $sgpr9, 5, [[V_WRITELANE_B32_1]]
- ; EXPANDED-NEXT: [[V_WRITELANE_B32_1:%[0-9]+]]:vgpr_32 = V_WRITELANE_B32 $sgpr10, 6, [[V_WRITELANE_B32_1]]
- ; EXPANDED-NEXT: [[V_WRITELANE_B32_1:%[0-9]+]]:vgpr_32 = V_WRITELANE_B32 $sgpr11, 7, [[V_WRITELANE_B32_1]]
- ; EXPANDED-NEXT: [[V_WRITELANE_B32_1:%[0-9]+]]:vgpr_32 = V_WRITELANE_B32 killed $sgpr12, 8, [[V_WRITELANE_B32_1]], implicit killed $sgpr4_sgpr5_sgpr6_sgpr7_sgpr8_sgpr9_sgpr10_sgpr11_sgpr12
+ ; EXPANDED-NEXT: $vgpr0 = V_WRITELANE_B32 $sgpr4, 0, $vgpr0, implicit-def $sgpr4_sgpr5_sgpr6_sgpr7_sgpr8_sgpr9_sgpr10_sgpr11_sgpr12, implicit $sgpr4_sgpr5_sgpr6_sgpr7_sgpr8_sgpr9_sgpr10_sgpr11_sgpr12
+ ; EXPANDED-NEXT: $vgpr0 = V_WRITELANE_B32 $sgpr5, 1, $vgpr0
+ ; EXPANDED-NEXT: $vgpr0 = V_WRITELANE_B32 $sgpr6, 2, $vgpr0
+ ; EXPANDED-NEXT: $vgpr0 = V_WRITELANE_B32 $sgpr7, 3, $vgpr0
+ ; EXPANDED-NEXT: $vgpr0 = V_WRITELANE_B32 $sgpr8, 4, $vgpr0
+ ; EXPANDED-NEXT: $vgpr0 = V_WRITELANE_B32 $sgpr9, 5, $vgpr0
+ ; EXPANDED-NEXT: $vgpr0 = V_WRITELANE_B32 $sgpr10, 6, $vgpr0
+ ; EXPANDED-NEXT: $vgpr0 = V_WRITELANE_B32 $sgpr11, 7, $vgpr0
+ ; EXPANDED-NEXT: $vgpr0 = V_WRITELANE_B32 killed $sgpr12, 8, $vgpr0, implicit killed $sgpr4_sgpr5_sgpr6_sgpr7_sgpr8_sgpr9_sgpr10_sgpr11_sgpr12
; EXPANDED-NEXT: S_CBRANCH_SCC1 %bb.1, implicit undef $scc
; EXPANDED-NEXT: {{ $}}
; EXPANDED-NEXT: bb.1:
; EXPANDED-NEXT: successors: %bb.2(0x80000000)
+ ; EXPANDED-NEXT: liveins: $vgpr0
; EXPANDED-NEXT: {{ $}}
; EXPANDED-NEXT: S_NOP 1
; EXPANDED-NEXT: {{ $}}
; EXPANDED-NEXT: bb.2:
- ; EXPANDED-NEXT: $sgpr4 = V_READLANE_B32 [[V_WRITELANE_B32_1]], 0, implicit-def $sgpr4_sgpr5_sgpr6_sgpr7_sgpr8_sgpr9_sgpr10_sgpr11_sgpr12
- ; EXPANDED-NEXT: $sgpr5 = V_READLANE_B32 [[V_WRITELANE_B32_1]], 1
- ; EXPANDED-NEXT: $sgpr6 = V_READLANE_B32 [[V_WRITELANE_B32_1]], 2
- ; EXPANDED-NEXT: $sgpr7 = V_READLANE_B32 [[V_WRITELANE_B32_1]], 3
- ; EXPANDED-NEXT: $sgpr8 = V_READLANE_B32 [[V_WRITELANE_B32_1]], 4
- ; EXPANDED-NEXT: $sgpr9 = V_READLANE_B32 [[V_WRITELANE_B32_1]], 5
- ; EXPANDED-NEXT: $sgpr10 = V_READLANE_B32 [[V_WRITELANE_B32_1]], 6
- ; EXPANDED-NEXT: $sgpr11 = V_READLANE_B32 [[V_WRITELANE_B32_1]], 7
- ; EXPANDED-NEXT: $sgpr12 = V_READLANE_B32 [[V_WRITELANE_B32_1]], 8
+ ; EXPANDED-NEXT: liveins: $vgpr0
+ ; EXPANDED-NEXT: {{ $}}
+ ; EXPANDED-NEXT: $sgpr4 = V_READLANE_B32 $vgpr0, 0, implicit-def $sgpr4_sgpr5_sgpr6_sgpr7_sgpr8_sgpr9_sgpr10_sgpr11_sgpr12
+ ; EXPANDED-NEXT: $sgpr5 = V_READLANE_B32 $vgpr0, 1
+ ; EXPANDED-NEXT: $sgpr6 = V_READLANE_B32 $vgpr0, 2
+ ; EXPANDED-NEXT: $sgpr7 = V_READLANE_B32 $vgpr0, 3
+ ; EXPANDED-NEXT: $sgpr8 = V_READLANE_B32 $vgpr0, 4
+ ; EXPANDED-NEXT: $sgpr9 = V_READLANE_B32 $vgpr0, 5
+ ; EXPANDED-NEXT: $sgpr10 = V_READLANE_B32 $vgpr0, 6
+ ; EXPANDED-NEXT: $sgpr11 = V_READLANE_B32 $vgpr0, 7
+ ; EXPANDED-NEXT: $sgpr12 = V_READLANE_B32 $vgpr0, 8
; EXPANDED-NEXT: S_NOP 0, implicit killed renamable $sgpr4_sgpr5_sgpr6_sgpr7_sgpr8_sgpr9_sgpr10_sgpr11_sgpr12
bb.0:
S_NOP 0, implicit-def %0:sgpr_288
diff --git a/llvm/test/CodeGen/AMDGPU/spill320.mir b/llvm/test/CodeGen/AMDGPU/spill320.mir
--- a/llvm/test/CodeGen/AMDGPU/spill320.mir
+++ b/llvm/test/CodeGen/AMDGPU/spill320.mir
@@ -30,37 +30,40 @@
; EXPANDED-LABEL: name: spill_restore_sgpr320
; EXPANDED: bb.0:
; EXPANDED-NEXT: successors: %bb.1(0x80000000)
+ ; EXPANDED-NEXT: liveins: $vgpr0
; EXPANDED-NEXT: {{ $}}
- ; EXPANDED-NEXT: [[DEF:%[0-9]+]]:vgpr_32 = IMPLICIT_DEF
; EXPANDED-NEXT: S_NOP 0, implicit-def renamable $sgpr4_sgpr5_sgpr6_sgpr7_sgpr8_sgpr9_sgpr10_sgpr11_sgpr12_sgpr13
- ; EXPANDED-NEXT: [[V_WRITELANE_B32_:%[0-9]+]]:vgpr_32 = V_WRITELANE_B32 $sgpr4, 0, [[V_WRITELANE_B32_]], implicit-def $sgpr4_sgpr5_sgpr6_sgpr7_sgpr8_sgpr9_sgpr10_sgpr11_sgpr12_sgpr13, implicit $sgpr4_sgpr5_sgpr6_sgpr7_sgpr8_sgpr9_sgpr10_sgpr11_sgpr12_sgpr13
- ; EXPANDED-NEXT: [[V_WRITELANE_B32_1:%[0-9]+]]:vgpr_32 = V_WRITELANE_B32 $sgpr5, 1, [[V_WRITELANE_B32_1]]
- ; EXPANDED-NEXT: [[V_WRITELANE_B32_1:%[0-9]+]]:vgpr_32 = V_WRITELANE_B32 $sgpr6, 2, [[V_WRITELANE_B32_1]]
- ; EXPANDED-NEXT: [[V_WRITELANE_B32_1:%[0-9]+]]:vgpr_32 = V_WRITELANE_B32 $sgpr7, 3, [[V_WRITELANE_B32_1]]
- ; EXPANDED-NEXT: [[V_WRITELANE_B32_1:%[0-9]+]]:vgpr_32 = V_WRITELANE_B32 $sgpr8, 4, [[V_WRITELANE_B32_1]]
- ; EXPANDED-NEXT: [[V_WRITELANE_B32_1:%[0-9]+]]:vgpr_32 = V_WRITELANE_B32 $sgpr9, 5, [[V_WRITELANE_B32_1]]
- ; EXPANDED-NEXT: [[V_WRITELANE_B32_1:%[0-9]+]]:vgpr_32 = V_WRITELANE_B32 $sgpr10, 6, [[V_WRITELANE_B32_1]]
- ; EXPANDED-NEXT: [[V_WRITELANE_B32_1:%[0-9]+]]:vgpr_32 = V_WRITELANE_B32 $sgpr11, 7, [[V_WRITELANE_B32_1]]
- ; EXPANDED-NEXT: [[V_WRITELANE_B32_1:%[0-9]+]]:vgpr_32 = V_WRITELANE_B32 $sgpr12, 8, [[V_WRITELANE_B32_1]]
- ; EXPANDED-NEXT: [[V_WRITELANE_B32_1:%[0-9]+]]:vgpr_32 = V_WRITELANE_B32 killed $sgpr13, 9, [[V_WRITELANE_B32_1]], implicit killed $sgpr4_sgpr5_sgpr6_sgpr7_sgpr8_sgpr9_sgpr10_sgpr11_sgpr12_sgpr13
+ ; EXPANDED-NEXT: $vgpr0 = V_WRITELANE_B32 $sgpr4, 0, $vgpr0, implicit-def $sgpr4_sgpr5_sgpr6_sgpr7_sgpr8_sgpr9_sgpr10_sgpr11_sgpr12_sgpr13, implicit $sgpr4_sgpr5_sgpr6_sgpr7_sgpr8_sgpr9_sgpr10_sgpr11_sgpr12_sgpr13
+ ; EXPANDED-NEXT: $vgpr0 = V_WRITELANE_B32 $sgpr5, 1, $vgpr0
+ ; EXPANDED-NEXT: $vgpr0 = V_WRITELANE_B32 $sgpr6, 2, $vgpr0
+ ; EXPANDED-NEXT: $vgpr0 = V_WRITELANE_B32 $sgpr7, 3, $vgpr0
+ ; EXPANDED-NEXT: $vgpr0 = V_WRITELANE_B32 $sgpr8, 4, $vgpr0
+ ; EXPANDED-NEXT: $vgpr0 = V_WRITELANE_B32 $sgpr9, 5, $vgpr0
+ ; EXPANDED-NEXT: $vgpr0 = V_WRITELANE_B32 $sgpr10, 6, $vgpr0
+ ; EXPANDED-NEXT: $vgpr0 = V_WRITELANE_B32 $sgpr11, 7, $vgpr0
+ ; EXPANDED-NEXT: $vgpr0 = V_WRITELANE_B32 $sgpr12, 8, $vgpr0
+ ; EXPANDED-NEXT: $vgpr0 = V_WRITELANE_B32 killed $sgpr13, 9, $vgpr0, implicit killed $sgpr4_sgpr5_sgpr6_sgpr7_sgpr8_sgpr9_sgpr10_sgpr11_sgpr12_sgpr13
; EXPANDED-NEXT: S_CBRANCH_SCC1 %bb.1, implicit undef $scc
; EXPANDED-NEXT: {{ $}}
; EXPANDED-NEXT: bb.1:
; EXPANDED-NEXT: successors: %bb.2(0x80000000)
+ ; EXPANDED-NEXT: liveins: $vgpr0
; EXPANDED-NEXT: {{ $}}
; EXPANDED-NEXT: S_NOP 1
; EXPANDED-NEXT: {{ $}}
; EXPANDED-NEXT: bb.2:
- ; EXPANDED-NEXT: $sgpr4 = V_READLANE_B32 [[V_WRITELANE_B32_1]], 0, implicit-def $sgpr4_sgpr5_sgpr6_sgpr7_sgpr8_sgpr9_sgpr10_sgpr11_sgpr12_sgpr13
- ; EXPANDED-NEXT: $sgpr5 = V_READLANE_B32 [[V_WRITELANE_B32_1]], 1
- ; EXPANDED-NEXT: $sgpr6 = V_READLANE_B32 [[V_WRITELANE_B32_1]], 2
- ; EXPANDED-NEXT: $sgpr7 = V_READLANE_B32 [[V_WRITELANE_B32_1]], 3
- ; EXPANDED-NEXT: $sgpr8 = V_READLANE_B32 [[V_WRITELANE_B32_1]], 4
- ; EXPANDED-NEXT: $sgpr9 = V_READLANE_B32 [[V_WRITELANE_B32_1]], 5
- ; EXPANDED-NEXT: $sgpr10 = V_READLANE_B32 [[V_WRITELANE_B32_1]], 6
- ; EXPANDED-NEXT: $sgpr11 = V_READLANE_B32 [[V_WRITELANE_B32_1]], 7
- ; EXPANDED-NEXT: $sgpr12 = V_READLANE_B32 [[V_WRITELANE_B32_1]], 8
- ; EXPANDED-NEXT: $sgpr13 = V_READLANE_B32 [[V_WRITELANE_B32_1]], 9
+ ; EXPANDED-NEXT: liveins: $vgpr0
+ ; EXPANDED-NEXT: {{ $}}
+ ; EXPANDED-NEXT: $sgpr4 = V_READLANE_B32 $vgpr0, 0, implicit-def $sgpr4_sgpr5_sgpr6_sgpr7_sgpr8_sgpr9_sgpr10_sgpr11_sgpr12_sgpr13
+ ; EXPANDED-NEXT: $sgpr5 = V_READLANE_B32 $vgpr0, 1
+ ; EXPANDED-NEXT: $sgpr6 = V_READLANE_B32 $vgpr0, 2
+ ; EXPANDED-NEXT: $sgpr7 = V_READLANE_B32 $vgpr0, 3
+ ; EXPANDED-NEXT: $sgpr8 = V_READLANE_B32 $vgpr0, 4
+ ; EXPANDED-NEXT: $sgpr9 = V_READLANE_B32 $vgpr0, 5
+ ; EXPANDED-NEXT: $sgpr10 = V_READLANE_B32 $vgpr0, 6
+ ; EXPANDED-NEXT: $sgpr11 = V_READLANE_B32 $vgpr0, 7
+ ; EXPANDED-NEXT: $sgpr12 = V_READLANE_B32 $vgpr0, 8
+ ; EXPANDED-NEXT: $sgpr13 = V_READLANE_B32 $vgpr0, 9
; EXPANDED-NEXT: S_NOP 0, implicit killed renamable $sgpr4_sgpr5_sgpr6_sgpr7_sgpr8_sgpr9_sgpr10_sgpr11_sgpr12_sgpr13
bb.0:
S_NOP 0, implicit-def %0:sgpr_320
diff --git a/llvm/test/CodeGen/AMDGPU/spill352.mir b/llvm/test/CodeGen/AMDGPU/spill352.mir
--- a/llvm/test/CodeGen/AMDGPU/spill352.mir
+++ b/llvm/test/CodeGen/AMDGPU/spill352.mir
@@ -30,39 +30,42 @@
; EXPANDED-LABEL: name: spill_restore_sgpr352
; EXPANDED: bb.0:
; EXPANDED-NEXT: successors: %bb.1(0x80000000)
+ ; EXPANDED-NEXT: liveins: $vgpr0
; EXPANDED-NEXT: {{ $}}
- ; EXPANDED-NEXT: [[DEF:%[0-9]+]]:vgpr_32 = IMPLICIT_DEF
; EXPANDED-NEXT: S_NOP 0, implicit-def renamable $sgpr4_sgpr5_sgpr6_sgpr7_sgpr8_sgpr9_sgpr10_sgpr11_sgpr12_sgpr13_sgpr14
- ; EXPANDED-NEXT: [[V_WRITELANE_B32_:%[0-9]+]]:vgpr_32 = V_WRITELANE_B32 $sgpr4, 0, [[V_WRITELANE_B32_]], implicit-def $sgpr4_sgpr5_sgpr6_sgpr7_sgpr8_sgpr9_sgpr10_sgpr11_sgpr12_sgpr13_sgpr14, implicit $sgpr4_sgpr5_sgpr6_sgpr7_sgpr8_sgpr9_sgpr10_sgpr11_sgpr12_sgpr13_sgpr14
- ; EXPANDED-NEXT: [[V_WRITELANE_B32_1:%[0-9]+]]:vgpr_32 = V_WRITELANE_B32 $sgpr5, 1, [[V_WRITELANE_B32_1]]
- ; EXPANDED-NEXT: [[V_WRITELANE_B32_1:%[0-9]+]]:vgpr_32 = V_WRITELANE_B32 $sgpr6, 2, [[V_WRITELANE_B32_1]]
- ; EXPANDED-NEXT: [[V_WRITELANE_B32_1:%[0-9]+]]:vgpr_32 = V_WRITELANE_B32 $sgpr7, 3, [[V_WRITELANE_B32_1]]
- ; EXPANDED-NEXT: [[V_WRITELANE_B32_1:%[0-9]+]]:vgpr_32 = V_WRITELANE_B32 $sgpr8, 4, [[V_WRITELANE_B32_1]]
- ; EXPANDED-NEXT: [[V_WRITELANE_B32_1:%[0-9]+]]:vgpr_32 = V_WRITELANE_B32 $sgpr9, 5, [[V_WRITELANE_B32_1]]
- ; EXPANDED-NEXT: [[V_WRITELANE_B32_1:%[0-9]+]]:vgpr_32 = V_WRITELANE_B32 $sgpr10, 6, [[V_WRITELANE_B32_1]]
- ; EXPANDED-NEXT: [[V_WRITELANE_B32_1:%[0-9]+]]:vgpr_32 = V_WRITELANE_B32 $sgpr11, 7, [[V_WRITELANE_B32_1]]
- ; EXPANDED-NEXT: [[V_WRITELANE_B32_1:%[0-9]+]]:vgpr_32 = V_WRITELANE_B32 $sgpr12, 8, [[V_WRITELANE_B32_1]]
- ; EXPANDED-NEXT: [[V_WRITELANE_B32_1:%[0-9]+]]:vgpr_32 = V_WRITELANE_B32 $sgpr13, 9, [[V_WRITELANE_B32_1]]
- ; EXPANDED-NEXT: [[V_WRITELANE_B32_1:%[0-9]+]]:vgpr_32 = V_WRITELANE_B32 killed $sgpr14, 10, [[V_WRITELANE_B32_1]], implicit killed $sgpr4_sgpr5_sgpr6_sgpr7_sgpr8_sgpr9_sgpr10_sgpr11_sgpr12_sgpr13_sgpr14
+ ; EXPANDED-NEXT: $vgpr0 = V_WRITELANE_B32 $sgpr4, 0, $vgpr0, implicit-def $sgpr4_sgpr5_sgpr6_sgpr7_sgpr8_sgpr9_sgpr10_sgpr11_sgpr12_sgpr13_sgpr14, implicit $sgpr4_sgpr5_sgpr6_sgpr7_sgpr8_sgpr9_sgpr10_sgpr11_sgpr12_sgpr13_sgpr14
+ ; EXPANDED-NEXT: $vgpr0 = V_WRITELANE_B32 $sgpr5, 1, $vgpr0
+ ; EXPANDED-NEXT: $vgpr0 = V_WRITELANE_B32 $sgpr6, 2, $vgpr0
+ ; EXPANDED-NEXT: $vgpr0 = V_WRITELANE_B32 $sgpr7, 3, $vgpr0
+ ; EXPANDED-NEXT: $vgpr0 = V_WRITELANE_B32 $sgpr8, 4, $vgpr0
+ ; EXPANDED-NEXT: $vgpr0 = V_WRITELANE_B32 $sgpr9, 5, $vgpr0
+ ; EXPANDED-NEXT: $vgpr0 = V_WRITELANE_B32 $sgpr10, 6, $vgpr0
+ ; EXPANDED-NEXT: $vgpr0 = V_WRITELANE_B32 $sgpr11, 7, $vgpr0
+ ; EXPANDED-NEXT: $vgpr0 = V_WRITELANE_B32 $sgpr12, 8, $vgpr0
+ ; EXPANDED-NEXT: $vgpr0 = V_WRITELANE_B32 $sgpr13, 9, $vgpr0
+ ; EXPANDED-NEXT: $vgpr0 = V_WRITELANE_B32 killed $sgpr14, 10, $vgpr0, implicit killed $sgpr4_sgpr5_sgpr6_sgpr7_sgpr8_sgpr9_sgpr10_sgpr11_sgpr12_sgpr13_sgpr14
; EXPANDED-NEXT: S_CBRANCH_SCC1 %bb.1, implicit undef $scc
; EXPANDED-NEXT: {{ $}}
; EXPANDED-NEXT: bb.1:
; EXPANDED-NEXT: successors: %bb.2(0x80000000)
+ ; EXPANDED-NEXT: liveins: $vgpr0
; EXPANDED-NEXT: {{ $}}
; EXPANDED-NEXT: S_NOP 1
; EXPANDED-NEXT: {{ $}}
; EXPANDED-NEXT: bb.2:
- ; EXPANDED-NEXT: $sgpr4 = V_READLANE_B32 [[V_WRITELANE_B32_1]], 0, implicit-def $sgpr4_sgpr5_sgpr6_sgpr7_sgpr8_sgpr9_sgpr10_sgpr11_sgpr12_sgpr13_sgpr14
- ; EXPANDED-NEXT: $sgpr5 = V_READLANE_B32 [[V_WRITELANE_B32_1]], 1
- ; EXPANDED-NEXT: $sgpr6 = V_READLANE_B32 [[V_WRITELANE_B32_1]], 2
- ; EXPANDED-NEXT: $sgpr7 = V_READLANE_B32 [[V_WRITELANE_B32_1]], 3
- ; EXPANDED-NEXT: $sgpr8 = V_READLANE_B32 [[V_WRITELANE_B32_1]], 4
- ; EXPANDED-NEXT: $sgpr9 = V_READLANE_B32 [[V_WRITELANE_B32_1]], 5
- ; EXPANDED-NEXT: $sgpr10 = V_READLANE_B32 [[V_WRITELANE_B32_1]], 6
- ; EXPANDED-NEXT: $sgpr11 = V_READLANE_B32 [[V_WRITELANE_B32_1]], 7
- ; EXPANDED-NEXT: $sgpr12 = V_READLANE_B32 [[V_WRITELANE_B32_1]], 8
- ; EXPANDED-NEXT: $sgpr13 = V_READLANE_B32 [[V_WRITELANE_B32_1]], 9
- ; EXPANDED-NEXT: $sgpr14 = V_READLANE_B32 [[V_WRITELANE_B32_1]], 10
+ ; EXPANDED-NEXT: liveins: $vgpr0
+ ; EXPANDED-NEXT: {{ $}}
+ ; EXPANDED-NEXT: $sgpr4 = V_READLANE_B32 $vgpr0, 0, implicit-def $sgpr4_sgpr5_sgpr6_sgpr7_sgpr8_sgpr9_sgpr10_sgpr11_sgpr12_sgpr13_sgpr14
+ ; EXPANDED-NEXT: $sgpr5 = V_READLANE_B32 $vgpr0, 1
+ ; EXPANDED-NEXT: $sgpr6 = V_READLANE_B32 $vgpr0, 2
+ ; EXPANDED-NEXT: $sgpr7 = V_READLANE_B32 $vgpr0, 3
+ ; EXPANDED-NEXT: $sgpr8 = V_READLANE_B32 $vgpr0, 4
+ ; EXPANDED-NEXT: $sgpr9 = V_READLANE_B32 $vgpr0, 5
+ ; EXPANDED-NEXT: $sgpr10 = V_READLANE_B32 $vgpr0, 6
+ ; EXPANDED-NEXT: $sgpr11 = V_READLANE_B32 $vgpr0, 7
+ ; EXPANDED-NEXT: $sgpr12 = V_READLANE_B32 $vgpr0, 8
+ ; EXPANDED-NEXT: $sgpr13 = V_READLANE_B32 $vgpr0, 9
+ ; EXPANDED-NEXT: $sgpr14 = V_READLANE_B32 $vgpr0, 10
; EXPANDED-NEXT: S_NOP 0, implicit killed renamable $sgpr4_sgpr5_sgpr6_sgpr7_sgpr8_sgpr9_sgpr10_sgpr11_sgpr12_sgpr13_sgpr14
bb.0:
S_NOP 0, implicit-def %0:sgpr_352
diff --git a/llvm/test/CodeGen/AMDGPU/spill384.mir b/llvm/test/CodeGen/AMDGPU/spill384.mir
--- a/llvm/test/CodeGen/AMDGPU/spill384.mir
+++ b/llvm/test/CodeGen/AMDGPU/spill384.mir
@@ -30,41 +30,44 @@
; EXPANDED-LABEL: name: spill_restore_sgpr384
; EXPANDED: bb.0:
; EXPANDED-NEXT: successors: %bb.1(0x80000000)
+ ; EXPANDED-NEXT: liveins: $vgpr0
; EXPANDED-NEXT: {{ $}}
- ; EXPANDED-NEXT: [[DEF:%[0-9]+]]:vgpr_32 = IMPLICIT_DEF
; EXPANDED-NEXT: S_NOP 0, implicit-def renamable $sgpr4_sgpr5_sgpr6_sgpr7_sgpr8_sgpr9_sgpr10_sgpr11_sgpr12_sgpr13_sgpr14_sgpr15
- ; EXPANDED-NEXT: [[V_WRITELANE_B32_:%[0-9]+]]:vgpr_32 = V_WRITELANE_B32 $sgpr4, 0, [[V_WRITELANE_B32_]], implicit-def $sgpr4_sgpr5_sgpr6_sgpr7_sgpr8_sgpr9_sgpr10_sgpr11_sgpr12_sgpr13_sgpr14_sgpr15, implicit $sgpr4_sgpr5_sgpr6_sgpr7_sgpr8_sgpr9_sgpr10_sgpr11_sgpr12_sgpr13_sgpr14_sgpr15
- ; EXPANDED-NEXT: [[V_WRITELANE_B32_1:%[0-9]+]]:vgpr_32 = V_WRITELANE_B32 $sgpr5, 1, [[V_WRITELANE_B32_1]]
- ; EXPANDED-NEXT: [[V_WRITELANE_B32_1:%[0-9]+]]:vgpr_32 = V_WRITELANE_B32 $sgpr6, 2, [[V_WRITELANE_B32_1]]
- ; EXPANDED-NEXT: [[V_WRITELANE_B32_1:%[0-9]+]]:vgpr_32 = V_WRITELANE_B32 $sgpr7, 3, [[V_WRITELANE_B32_1]]
- ; EXPANDED-NEXT: [[V_WRITELANE_B32_1:%[0-9]+]]:vgpr_32 = V_WRITELANE_B32 $sgpr8, 4, [[V_WRITELANE_B32_1]]
- ; EXPANDED-NEXT: [[V_WRITELANE_B32_1:%[0-9]+]]:vgpr_32 = V_WRITELANE_B32 $sgpr9, 5, [[V_WRITELANE_B32_1]]
- ; EXPANDED-NEXT: [[V_WRITELANE_B32_1:%[0-9]+]]:vgpr_32 = V_WRITELANE_B32 $sgpr10, 6, [[V_WRITELANE_B32_1]]
- ; EXPANDED-NEXT: [[V_WRITELANE_B32_1:%[0-9]+]]:vgpr_32 = V_WRITELANE_B32 $sgpr11, 7, [[V_WRITELANE_B32_1]]
- ; EXPANDED-NEXT: [[V_WRITELANE_B32_1:%[0-9]+]]:vgpr_32 = V_WRITELANE_B32 $sgpr12, 8, [[V_WRITELANE_B32_1]]
- ; EXPANDED-NEXT: [[V_WRITELANE_B32_1:%[0-9]+]]:vgpr_32 = V_WRITELANE_B32 $sgpr13, 9, [[V_WRITELANE_B32_1]]
- ; EXPANDED-NEXT: [[V_WRITELANE_B32_1:%[0-9]+]]:vgpr_32 = V_WRITELANE_B32 $sgpr14, 10, [[V_WRITELANE_B32_1]]
- ; EXPANDED-NEXT: [[V_WRITELANE_B32_1:%[0-9]+]]:vgpr_32 = V_WRITELANE_B32 killed $sgpr15, 11, [[V_WRITELANE_B32_1]], implicit killed $sgpr4_sgpr5_sgpr6_sgpr7_sgpr8_sgpr9_sgpr10_sgpr11_sgpr12_sgpr13_sgpr14_sgpr15
+ ; EXPANDED-NEXT: $vgpr0 = V_WRITELANE_B32 $sgpr4, 0, $vgpr0, implicit-def $sgpr4_sgpr5_sgpr6_sgpr7_sgpr8_sgpr9_sgpr10_sgpr11_sgpr12_sgpr13_sgpr14_sgpr15, implicit $sgpr4_sgpr5_sgpr6_sgpr7_sgpr8_sgpr9_sgpr10_sgpr11_sgpr12_sgpr13_sgpr14_sgpr15
+ ; EXPANDED-NEXT: $vgpr0 = V_WRITELANE_B32 $sgpr5, 1, $vgpr0
+ ; EXPANDED-NEXT: $vgpr0 = V_WRITELANE_B32 $sgpr6, 2, $vgpr0
+ ; EXPANDED-NEXT: $vgpr0 = V_WRITELANE_B32 $sgpr7, 3, $vgpr0
+ ; EXPANDED-NEXT: $vgpr0 = V_WRITELANE_B32 $sgpr8, 4, $vgpr0
+ ; EXPANDED-NEXT: $vgpr0 = V_WRITELANE_B32 $sgpr9, 5, $vgpr0
+ ; EXPANDED-NEXT: $vgpr0 = V_WRITELANE_B32 $sgpr10, 6, $vgpr0
+ ; EXPANDED-NEXT: $vgpr0 = V_WRITELANE_B32 $sgpr11, 7, $vgpr0
+ ; EXPANDED-NEXT: $vgpr0 = V_WRITELANE_B32 $sgpr12, 8, $vgpr0
+ ; EXPANDED-NEXT: $vgpr0 = V_WRITELANE_B32 $sgpr13, 9, $vgpr0
+ ; EXPANDED-NEXT: $vgpr0 = V_WRITELANE_B32 $sgpr14, 10, $vgpr0
+ ; EXPANDED-NEXT: $vgpr0 = V_WRITELANE_B32 killed $sgpr15, 11, $vgpr0, implicit killed $sgpr4_sgpr5_sgpr6_sgpr7_sgpr8_sgpr9_sgpr10_sgpr11_sgpr12_sgpr13_sgpr14_sgpr15
; EXPANDED-NEXT: S_CBRANCH_SCC1 %bb.1, implicit undef $scc
; EXPANDED-NEXT: {{ $}}
; EXPANDED-NEXT: bb.1:
; EXPANDED-NEXT: successors: %bb.2(0x80000000)
+ ; EXPANDED-NEXT: liveins: $vgpr0
; EXPANDED-NEXT: {{ $}}
; EXPANDED-NEXT: S_NOP 1
; EXPANDED-NEXT: {{ $}}
; EXPANDED-NEXT: bb.2:
- ; EXPANDED-NEXT: $sgpr4 = V_READLANE_B32 [[V_WRITELANE_B32_1]], 0, implicit-def $sgpr4_sgpr5_sgpr6_sgpr7_sgpr8_sgpr9_sgpr10_sgpr11_sgpr12_sgpr13_sgpr14_sgpr15
- ; EXPANDED-NEXT: $sgpr5 = V_READLANE_B32 [[V_WRITELANE_B32_1]], 1
- ; EXPANDED-NEXT: $sgpr6 = V_READLANE_B32 [[V_WRITELANE_B32_1]], 2
- ; EXPANDED-NEXT: $sgpr7 = V_READLANE_B32 [[V_WRITELANE_B32_1]], 3
- ; EXPANDED-NEXT: $sgpr8 = V_READLANE_B32 [[V_WRITELANE_B32_1]], 4
- ; EXPANDED-NEXT: $sgpr9 = V_READLANE_B32 [[V_WRITELANE_B32_1]], 5
- ; EXPANDED-NEXT: $sgpr10 = V_READLANE_B32 [[V_WRITELANE_B32_1]], 6
- ; EXPANDED-NEXT: $sgpr11 = V_READLANE_B32 [[V_WRITELANE_B32_1]], 7
- ; EXPANDED-NEXT: $sgpr12 = V_READLANE_B32 [[V_WRITELANE_B32_1]], 8
- ; EXPANDED-NEXT: $sgpr13 = V_READLANE_B32 [[V_WRITELANE_B32_1]], 9
- ; EXPANDED-NEXT: $sgpr14 = V_READLANE_B32 [[V_WRITELANE_B32_1]], 10
- ; EXPANDED-NEXT: $sgpr15 = V_READLANE_B32 [[V_WRITELANE_B32_1]], 11
+ ; EXPANDED-NEXT: liveins: $vgpr0
+ ; EXPANDED-NEXT: {{ $}}
+ ; EXPANDED-NEXT: $sgpr4 = V_READLANE_B32 $vgpr0, 0, implicit-def $sgpr4_sgpr5_sgpr6_sgpr7_sgpr8_sgpr9_sgpr10_sgpr11_sgpr12_sgpr13_sgpr14_sgpr15
+ ; EXPANDED-NEXT: $sgpr5 = V_READLANE_B32 $vgpr0, 1
+ ; EXPANDED-NEXT: $sgpr6 = V_READLANE_B32 $vgpr0, 2
+ ; EXPANDED-NEXT: $sgpr7 = V_READLANE_B32 $vgpr0, 3
+ ; EXPANDED-NEXT: $sgpr8 = V_READLANE_B32 $vgpr0, 4
+ ; EXPANDED-NEXT: $sgpr9 = V_READLANE_B32 $vgpr0, 5
+ ; EXPANDED-NEXT: $sgpr10 = V_READLANE_B32 $vgpr0, 6
+ ; EXPANDED-NEXT: $sgpr11 = V_READLANE_B32 $vgpr0, 7
+ ; EXPANDED-NEXT: $sgpr12 = V_READLANE_B32 $vgpr0, 8
+ ; EXPANDED-NEXT: $sgpr13 = V_READLANE_B32 $vgpr0, 9
+ ; EXPANDED-NEXT: $sgpr14 = V_READLANE_B32 $vgpr0, 10
+ ; EXPANDED-NEXT: $sgpr15 = V_READLANE_B32 $vgpr0, 11
; EXPANDED-NEXT: S_NOP 0, implicit killed renamable $sgpr4_sgpr5_sgpr6_sgpr7_sgpr8_sgpr9_sgpr10_sgpr11_sgpr12_sgpr13_sgpr14_sgpr15
bb.0:
S_NOP 0, implicit-def %0:sgpr_384
diff --git a/llvm/test/CodeGen/AMDGPU/stack-realign.ll b/llvm/test/CodeGen/AMDGPU/stack-realign.ll
--- a/llvm/test/CodeGen/AMDGPU/stack-realign.ll
+++ b/llvm/test/CodeGen/AMDGPU/stack-realign.ll
@@ -161,24 +161,26 @@
; GCN-NEXT: s_and_b32 s33, [[SCRATCH_REG]], 0xffff0000
; GCN-NEXT: s_or_saveexec_b64 s[18:19], -1
; GCN-NEXT: buffer_store_dword [[VGPR_REG:v[0-9]+]], off, s[0:3], s33 offset:1028 ; 4-byte Folded Spill
+; GCN-NEXT: buffer_store_dword [[VGPR_REG_1:v[0-9]+]], off, s[0:3], s33 offset:1032 ; 4-byte Folded Spill
; GCN-NEXT: s_mov_b64 exec, s[18:19]
; GCN-NEXT: v_mov_b32_e32 v32, 0
-; GCN-DAG: v_writelane_b32 [[VGPR_REG]], s34, 3
+; GCN-DAG: v_writelane_b32 [[VGPR_REG_1]], s34, 1
; GCN: s_mov_b32 s34, s32
; GCN: buffer_store_dword v32, off, s[0:3], s33 offset:1024
; GCN-NEXT: s_waitcnt vmcnt(0)
; GCN-NEXT: buffer_load_dword v{{[0-9]+}}, off, s[0:3], s34
; GCN-DAG: s_add_i32 s32, s32, 0x30000
-; GCN: v_writelane_b32 [[VGPR_REG]], [[FP_SCRATCH_COPY]], 2
+; GCN: v_writelane_b32 [[VGPR_REG_1]], [[FP_SCRATCH_COPY]], 0
; GCN: buffer_store_dword v{{[0-9]+}}, off, s[0:3], s32
; GCN: s_swappc_b64 s[30:31],
; GCN: v_readlane_b32 s31, [[VGPR_REG]], 1
; GCN: v_readlane_b32 s30, [[VGPR_REG]], 0
-; GCN-NEXT: v_readlane_b32 s34, [[VGPR_REG]], 3
-; GCN-NEXT: v_readlane_b32 [[FP_SCRATCH_COPY:s[0-9]+]], [[VGPR_REG]], 2
+; GCN-NEXT: v_readlane_b32 s34, [[VGPR_REG_1]], 1
+; GCN-NEXT: v_readlane_b32 [[FP_SCRATCH_COPY:s[0-9]+]], [[VGPR_REG_1]], 0
; GCN-NEXT: s_or_saveexec_b64 s[6:7], -1
; GCN-NEXT: buffer_load_dword [[VGPR_REG]], off, s[0:3], s33 offset:1028 ; 4-byte Folded Reload
+; GCN-NEXT: buffer_load_dword [[VGPR_REG_1]], off, s[0:3], s33 offset:1032 ; 4-byte Folded Reload
; GCN-NEXT: s_mov_b64 exec, s[6:7]
; GCN-NEXT: s_add_i32 s32, s32, 0xfffd0000
; GCN-NEXT: s_mov_b32 s33, [[FP_SCRATCH_COPY]]
diff --git a/llvm/test/CodeGen/AMDGPU/swdev380865.ll b/llvm/test/CodeGen/AMDGPU/swdev380865.ll
--- a/llvm/test/CodeGen/AMDGPU/swdev380865.ll
+++ b/llvm/test/CodeGen/AMDGPU/swdev380865.ll
@@ -16,96 +16,94 @@
; CHECK: ; %bb.0: ; %entry
; CHECK-NEXT: s_mov_b64 s[0:1], 0
; CHECK-NEXT: s_load_dword s2, s[0:1], 0x0
-; CHECK-NEXT: ; implicit-def: $vgpr2
; CHECK-NEXT: ; kill: killed $sgpr0_sgpr1
; CHECK-NEXT: s_mov_b32 s7, 0x401c0000
; CHECK-NEXT: s_mov_b32 s5, 0x40280000
; CHECK-NEXT: s_waitcnt lgkmcnt(0)
-; CHECK-NEXT: v_writelane_b32 v2, s2, 0
+; CHECK-NEXT: v_writelane_b32 v0, s2, 0
; CHECK-NEXT: s_load_dwordx2 s[2:3], s[0:1], 0x0
; CHECK-NEXT: s_mov_b32 s0, 0
; CHECK-NEXT: s_mov_b32 s1, 0x40140000
; CHECK-NEXT: s_mov_b32 s1, 0x40180000
-; CHECK-NEXT: v_writelane_b32 v2, s0, 1
-; CHECK-NEXT: v_writelane_b32 v2, s1, 2
+; CHECK-NEXT: v_writelane_b32 v0, s0, 1
+; CHECK-NEXT: v_writelane_b32 v0, s1, 2
; CHECK-NEXT: s_mov_b32 s1, 0x40220000
-; CHECK-NEXT: v_writelane_b32 v2, s0, 3
-; CHECK-NEXT: v_writelane_b32 v2, s1, 4
+; CHECK-NEXT: v_writelane_b32 v0, s0, 3
+; CHECK-NEXT: v_writelane_b32 v0, s1, 4
; CHECK-NEXT: s_mov_b32 s1, 0x40240000
-; CHECK-NEXT: v_writelane_b32 v2, s0, 5
-; CHECK-NEXT: v_writelane_b32 v2, s1, 6
+; CHECK-NEXT: v_writelane_b32 v0, s0, 5
+; CHECK-NEXT: v_writelane_b32 v0, s1, 6
; CHECK-NEXT: s_mov_b32 s1, 0x40260000
-; CHECK-NEXT: v_writelane_b32 v2, s0, 7
+; CHECK-NEXT: v_writelane_b32 v0, s0, 7
; CHECK-NEXT: s_waitcnt lgkmcnt(0)
-; CHECK-NEXT: v_mov_b32_e32 v0, s2
-; CHECK-NEXT: v_writelane_b32 v2, s1, 8
-; CHECK-NEXT: v_mov_b32_e32 v1, s3
+; CHECK-NEXT: v_mov_b32_e32 v1, s2
+; CHECK-NEXT: v_writelane_b32 v0, s1, 8
+; CHECK-NEXT: v_mov_b32_e32 v2, s3
; CHECK-NEXT: .LBB0_1: ; %for.cond4.preheader
; CHECK-NEXT: ; =>This Inner Loop Header: Depth=1
-; CHECK-NEXT: v_add_f64 v[0:1], v[0:1], 0
+; CHECK-NEXT: v_add_f64 v[1:2], v[1:2], 0
; CHECK-NEXT: s_mov_b32 s2, 0
; CHECK-NEXT: s_mov_b32 s3, 0x40140000
-; CHECK-NEXT: v_writelane_b32 v2, s6, 9
-; CHECK-NEXT: v_writelane_b32 v2, s7, 10
-; CHECK-NEXT: v_writelane_b32 v2, s0, 11
-; CHECK-NEXT: v_readlane_b32 s6, v2, 1
-; CHECK-NEXT: v_readlane_b32 s7, v2, 2
-; CHECK-NEXT: v_add_f64 v[0:1], v[0:1], s[2:3]
+; CHECK-NEXT: v_writelane_b32 v0, s6, 9
+; CHECK-NEXT: v_writelane_b32 v0, s7, 10
+; CHECK-NEXT: v_writelane_b32 v0, s0, 11
+; CHECK-NEXT: v_readlane_b32 s6, v0, 1
+; CHECK-NEXT: v_readlane_b32 s7, v0, 2
+; CHECK-NEXT: v_add_f64 v[1:2], v[1:2], s[2:3]
; CHECK-NEXT: s_mov_b32 s1, s7
; CHECK-NEXT: s_mov_b32 s0, s2
-; CHECK-NEXT: v_writelane_b32 v2, s6, 1
-; CHECK-NEXT: v_writelane_b32 v2, s7, 2
-; CHECK-NEXT: v_readlane_b32 s6, v2, 9
-; CHECK-NEXT: v_readlane_b32 s7, v2, 10
+; CHECK-NEXT: v_writelane_b32 v0, s6, 1
+; CHECK-NEXT: v_writelane_b32 v0, s7, 2
+; CHECK-NEXT: v_readlane_b32 s6, v0, 9
+; CHECK-NEXT: v_readlane_b32 s7, v0, 10
; CHECK-NEXT: s_mov_b32 s6, s2
-; CHECK-NEXT: v_add_f64 v[0:1], v[0:1], s[0:1]
-; CHECK-NEXT: v_readlane_b32 s0, v2, 3
-; CHECK-NEXT: v_readlane_b32 s1, v2, 4
+; CHECK-NEXT: v_add_f64 v[1:2], v[1:2], s[0:1]
+; CHECK-NEXT: v_readlane_b32 s0, v0, 3
+; CHECK-NEXT: v_readlane_b32 s1, v0, 4
; CHECK-NEXT: s_mov_b32 s3, s1
; CHECK-NEXT: s_mov_b32 s0, 0
; CHECK-NEXT: s_mov_b32 s1, 0x40140000
; CHECK-NEXT: s_mov_b32 s2, s0
; CHECK-NEXT: s_mov_b32 s1, s3
-; CHECK-NEXT: v_add_f64 v[0:1], v[0:1], s[6:7]
-; CHECK-NEXT: v_writelane_b32 v2, s0, 3
-; CHECK-NEXT: v_writelane_b32 v2, s1, 4
-; CHECK-NEXT: v_readlane_b32 s0, v2, 5
-; CHECK-NEXT: v_readlane_b32 s1, v2, 6
-; CHECK-NEXT: v_add_f64 v[0:1], v[0:1], s[2:3]
+; CHECK-NEXT: v_add_f64 v[1:2], v[1:2], s[6:7]
+; CHECK-NEXT: v_writelane_b32 v0, s0, 3
+; CHECK-NEXT: v_writelane_b32 v0, s1, 4
+; CHECK-NEXT: v_readlane_b32 s0, v0, 5
+; CHECK-NEXT: v_readlane_b32 s1, v0, 6
+; CHECK-NEXT: v_add_f64 v[1:2], v[1:2], s[2:3]
; CHECK-NEXT: s_mov_b32 s3, s1
; CHECK-NEXT: s_mov_b32 s0, 0
; CHECK-NEXT: s_mov_b32 s1, 0x40140000
; CHECK-NEXT: s_mov_b32 s2, s0
; CHECK-NEXT: s_mov_b32 s1, s3
-; CHECK-NEXT: v_writelane_b32 v2, s0, 5
-; CHECK-NEXT: v_writelane_b32 v2, s1, 6
-; CHECK-NEXT: v_add_f64 v[0:1], v[0:1], s[2:3]
-; CHECK-NEXT: v_readlane_b32 s0, v2, 7
-; CHECK-NEXT: v_readlane_b32 s1, v2, 8
+; CHECK-NEXT: v_writelane_b32 v0, s0, 5
+; CHECK-NEXT: v_writelane_b32 v0, s1, 6
+; CHECK-NEXT: v_add_f64 v[1:2], v[1:2], s[2:3]
+; CHECK-NEXT: v_readlane_b32 s0, v0, 7
+; CHECK-NEXT: v_readlane_b32 s1, v0, 8
; CHECK-NEXT: s_mov_b32 s3, s1
; CHECK-NEXT: s_mov_b32 s0, 0
; CHECK-NEXT: s_mov_b32 s1, 0x40140000
; CHECK-NEXT: s_mov_b32 s2, s0
; CHECK-NEXT: s_mov_b32 s1, s3
-; CHECK-NEXT: v_add_f64 v[0:1], v[0:1], s[2:3]
-; CHECK-NEXT: v_writelane_b32 v2, s0, 7
-; CHECK-NEXT: v_writelane_b32 v2, s1, 8
+; CHECK-NEXT: v_add_f64 v[1:2], v[1:2], s[2:3]
+; CHECK-NEXT: v_writelane_b32 v0, s0, 7
+; CHECK-NEXT: v_writelane_b32 v0, s1, 8
; CHECK-NEXT: s_mov_b32 s0, 0
; CHECK-NEXT: s_mov_b32 s1, 0x40140000
; CHECK-NEXT: s_mov_b32 s4, s0
-; CHECK-NEXT: v_readlane_b32 s0, v2, 0
-; CHECK-NEXT: v_readlane_b32 s2, v2, 11
-; CHECK-NEXT: v_add_f64 v[0:1], v[0:1], s[4:5]
+; CHECK-NEXT: v_readlane_b32 s0, v0, 0
+; CHECK-NEXT: v_readlane_b32 s2, v0, 11
+; CHECK-NEXT: v_add_f64 v[1:2], v[1:2], s[4:5]
; CHECK-NEXT: s_add_i32 s2, s2, s0
-; CHECK-NEXT: v_writelane_b32 v2, s2, 11
-; CHECK-NEXT: v_readlane_b32 s0, v2, 11
+; CHECK-NEXT: v_writelane_b32 v0, s2, 11
+; CHECK-NEXT: v_readlane_b32 s0, v0, 11
; CHECK-NEXT: s_cmpk_lt_i32 s0, 0xa00
; CHECK-NEXT: s_cbranch_scc1 .LBB0_1
; CHECK-NEXT: ; %bb.2: ; %for.cond.cleanup.loopexit
; CHECK-NEXT: v_mov_b32_e32 v3, 0
; CHECK-NEXT: v_mov_b32_e32 v4, 0
-; CHECK-NEXT: global_store_dwordx2 v[3:4], v[0:1], off
-; CHECK-NEXT: ; kill: killed $vgpr2
+; CHECK-NEXT: global_store_dwordx2 v[3:4], v[1:2], off
; CHECK-NEXT: s_endpgm
entry:
%0 = load i32, ptr addrspace(4) null, align 4
diff --git a/llvm/test/CodeGen/AMDGPU/tuple-allocation-failure.ll b/llvm/test/CodeGen/AMDGPU/tuple-allocation-failure.ll
--- a/llvm/test/CodeGen/AMDGPU/tuple-allocation-failure.ll
+++ b/llvm/test/CodeGen/AMDGPU/tuple-allocation-failure.ll
@@ -12,9 +12,10 @@
; GFX90A-NEXT: s_mov_b32 s33, s32
; GFX90A-NEXT: s_or_saveexec_b64 s[18:19], -1
; GFX90A-NEXT: buffer_store_dword v40, off, s[0:3], s33 ; 4-byte Folded Spill
+; GFX90A-NEXT: buffer_store_dword v41, off, s[0:3], s33 offset:4 ; 4-byte Folded Spill
; GFX90A-NEXT: s_mov_b64 exec, s[18:19]
; GFX90A-NEXT: s_addk_i32 s32, 0x400
-; GFX90A-NEXT: v_writelane_b32 v40, s16, 2
+; GFX90A-NEXT: v_writelane_b32 v41, s16, 0
; GFX90A-NEXT: s_getpc_b64 s[16:17]
; GFX90A-NEXT: s_add_u32 s16, s16, wobble@gotpcrel32@lo+4
; GFX90A-NEXT: s_addc_u32 s17, s17, wobble@gotpcrel32@hi+12
@@ -34,12 +35,12 @@
; GLOBALNESS1-NEXT: s_mov_b64 s[36:37], s[6:7]
; GLOBALNESS1-NEXT: s_load_dwordx4 s[76:79], s[8:9], 0x0
; GLOBALNESS1-NEXT: s_load_dword s6, s[8:9], 0x14
-; GLOBALNESS1-NEXT: v_mov_b32_e32 v41, v0
-; GLOBALNESS1-NEXT: v_mov_b32_e32 v42, 0
+; GLOBALNESS1-NEXT: v_mov_b32_e32 v42, v0
+; GLOBALNESS1-NEXT: v_mov_b32_e32 v40, 0
; GLOBALNESS1-NEXT: v_pk_mov_b32 v[0:1], 0, 0
-; GLOBALNESS1-NEXT: global_store_dword v[0:1], v42, off
+; GLOBALNESS1-NEXT: global_store_dword v[0:1], v40, off
; GLOBALNESS1-NEXT: s_waitcnt lgkmcnt(0)
-; GLOBALNESS1-NEXT: global_load_dword v0, v42, s[76:77]
+; GLOBALNESS1-NEXT: global_load_dword v0, v40, s[76:77]
; GLOBALNESS1-NEXT: s_mov_b64 s[40:41], s[4:5]
; GLOBALNESS1-NEXT: s_load_dwordx2 s[4:5], s[8:9], 0x18
; GLOBALNESS1-NEXT: s_load_dword s7, s[8:9], 0x20
@@ -47,10 +48,10 @@
; GLOBALNESS1-NEXT: s_addc_u32 flat_scratch_hi, s13, 0
; GLOBALNESS1-NEXT: s_add_u32 s0, s0, s17
; GLOBALNESS1-NEXT: s_addc_u32 s1, s1, 0
-; GLOBALNESS1-NEXT: v_mov_b32_e32 v43, 0x40994400
+; GLOBALNESS1-NEXT: v_mov_b32_e32 v41, 0x40994400
; GLOBALNESS1-NEXT: s_bitcmp1_b32 s78, 0
; GLOBALNESS1-NEXT: s_waitcnt lgkmcnt(0)
-; GLOBALNESS1-NEXT: v_cmp_ngt_f64_e32 vcc, s[4:5], v[42:43]
+; GLOBALNESS1-NEXT: v_cmp_ngt_f64_e32 vcc, s[4:5], v[40:41]
; GLOBALNESS1-NEXT: v_cmp_ngt_f64_e64 s[4:5], s[4:5], 0
; GLOBALNESS1-NEXT: v_cndmask_b32_e64 v2, 0, 1, s[4:5]
; GLOBALNESS1-NEXT: s_cselect_b64 s[4:5], -1, 0
@@ -116,9 +117,9 @@
; GLOBALNESS1-NEXT: ; =>This Loop Header: Depth=1
; GLOBALNESS1-NEXT: ; Child Loop BB1_15 Depth 2
; GLOBALNESS1-NEXT: v_pk_mov_b32 v[0:1], s[74:75], s[74:75] op_sel:[0,1]
-; GLOBALNESS1-NEXT: flat_load_dword v40, v[0:1]
+; GLOBALNESS1-NEXT: flat_load_dword v43, v[0:1]
; GLOBALNESS1-NEXT: s_add_u32 s8, s38, 40
-; GLOBALNESS1-NEXT: buffer_store_dword v42, off, s[0:3], 0
+; GLOBALNESS1-NEXT: buffer_store_dword v40, off, s[0:3], 0
; GLOBALNESS1-NEXT: flat_load_dword v46, v[0:1]
; GLOBALNESS1-NEXT: s_addc_u32 s9, s39, 0
; GLOBALNESS1-NEXT: s_mov_b64 s[4:5], s[40:41]
@@ -127,7 +128,7 @@
; GLOBALNESS1-NEXT: s_mov_b32 s12, s72
; GLOBALNESS1-NEXT: s_mov_b32 s13, s71
; GLOBALNESS1-NEXT: s_mov_b32 s14, s70
-; GLOBALNESS1-NEXT: v_mov_b32_e32 v31, v41
+; GLOBALNESS1-NEXT: v_mov_b32_e32 v31, v42
; GLOBALNESS1-NEXT: s_waitcnt lgkmcnt(0)
; GLOBALNESS1-NEXT: s_swappc_b64 s[30:31], s[76:77]
; GLOBALNESS1-NEXT: s_and_b64 vcc, exec, s[46:47]
@@ -170,13 +171,13 @@
; GLOBALNESS1-NEXT: s_cbranch_vccnz .LBB1_12
; GLOBALNESS1-NEXT: ; %bb.11: ; %bb39.i
; GLOBALNESS1-NEXT: ; in Loop: Header=BB1_4 Depth=1
-; GLOBALNESS1-NEXT: v_mov_b32_e32 v43, v42
+; GLOBALNESS1-NEXT: v_mov_b32_e32 v41, v40
; GLOBALNESS1-NEXT: v_pk_mov_b32 v[2:3], 0, 0
-; GLOBALNESS1-NEXT: global_store_dwordx2 v[2:3], v[42:43], off
+; GLOBALNESS1-NEXT: global_store_dwordx2 v[2:3], v[40:41], off
; GLOBALNESS1-NEXT: .LBB1_12: ; %bb44.lr.ph.i
; GLOBALNESS1-NEXT: ; in Loop: Header=BB1_4 Depth=1
; GLOBALNESS1-NEXT: v_cmp_ne_u32_e32 vcc, 0, v46
-; GLOBALNESS1-NEXT: v_cndmask_b32_e32 v2, 0, v40, vcc
+; GLOBALNESS1-NEXT: v_cndmask_b32_e32 v2, 0, v43, vcc
; GLOBALNESS1-NEXT: s_waitcnt vmcnt(0)
; GLOBALNESS1-NEXT: v_cmp_nlt_f64_e64 s[64:65], 0, v[0:1]
; GLOBALNESS1-NEXT: v_cmp_eq_u32_e64 s[66:67], 0, v2
@@ -223,7 +224,7 @@
; GLOBALNESS1-NEXT: s_mov_b32 s12, s72
; GLOBALNESS1-NEXT: s_mov_b32 s13, s71
; GLOBALNESS1-NEXT: s_mov_b32 s14, s70
-; GLOBALNESS1-NEXT: v_mov_b32_e32 v31, v41
+; GLOBALNESS1-NEXT: v_mov_b32_e32 v31, v42
; GLOBALNESS1-NEXT: s_swappc_b64 s[30:31], s[76:77]
; GLOBALNESS1-NEXT: v_pk_mov_b32 v[46:47], 0, 0
; GLOBALNESS1-NEXT: s_mov_b64 s[4:5], s[40:41]
@@ -233,15 +234,15 @@
; GLOBALNESS1-NEXT: s_mov_b32 s12, s72
; GLOBALNESS1-NEXT: s_mov_b32 s13, s71
; GLOBALNESS1-NEXT: s_mov_b32 s14, s70
-; GLOBALNESS1-NEXT: v_mov_b32_e32 v31, v41
+; GLOBALNESS1-NEXT: v_mov_b32_e32 v31, v42
; GLOBALNESS1-NEXT: global_store_dwordx2 v[46:47], v[44:45], off
; GLOBALNESS1-NEXT: s_swappc_b64 s[30:31], s[76:77]
; GLOBALNESS1-NEXT: s_and_saveexec_b64 s[4:5], s[66:67]
; GLOBALNESS1-NEXT: s_cbranch_execz .LBB1_13
; GLOBALNESS1-NEXT: ; %bb.22: ; %bb62.i
; GLOBALNESS1-NEXT: ; in Loop: Header=BB1_15 Depth=2
-; GLOBALNESS1-NEXT: v_mov_b32_e32 v43, v42
-; GLOBALNESS1-NEXT: global_store_dwordx2 v[46:47], v[42:43], off
+; GLOBALNESS1-NEXT: v_mov_b32_e32 v41, v40
+; GLOBALNESS1-NEXT: global_store_dwordx2 v[46:47], v[40:41], off
; GLOBALNESS1-NEXT: s_branch .LBB1_13
; GLOBALNESS1-NEXT: .LBB1_23: ; %LeafBlock
; GLOBALNESS1-NEXT: ; in Loop: Header=BB1_4 Depth=1
@@ -268,15 +269,15 @@
; GLOBALNESS1-NEXT: s_cbranch_vccnz .LBB1_1
; GLOBALNESS1-NEXT: ; %bb.28: ; %bb69.i
; GLOBALNESS1-NEXT: ; in Loop: Header=BB1_4 Depth=1
-; GLOBALNESS1-NEXT: v_mov_b32_e32 v43, v42
+; GLOBALNESS1-NEXT: v_mov_b32_e32 v41, v40
; GLOBALNESS1-NEXT: v_pk_mov_b32 v[2:3], 0, 0
-; GLOBALNESS1-NEXT: global_store_dwordx2 v[2:3], v[42:43], off
+; GLOBALNESS1-NEXT: global_store_dwordx2 v[2:3], v[40:41], off
; GLOBALNESS1-NEXT: s_branch .LBB1_1
; GLOBALNESS1-NEXT: .LBB1_29: ; %bb73.i
; GLOBALNESS1-NEXT: ; in Loop: Header=BB1_4 Depth=1
-; GLOBALNESS1-NEXT: v_mov_b32_e32 v43, v42
+; GLOBALNESS1-NEXT: v_mov_b32_e32 v41, v40
; GLOBALNESS1-NEXT: v_pk_mov_b32 v[2:3], 0, 0
-; GLOBALNESS1-NEXT: global_store_dwordx2 v[2:3], v[42:43], off
+; GLOBALNESS1-NEXT: global_store_dwordx2 v[2:3], v[40:41], off
; GLOBALNESS1-NEXT: s_branch .LBB1_2
; GLOBALNESS1-NEXT: .LBB1_30: ; %loop.exit.guard
; GLOBALNESS1-NEXT: s_andn2_b64 vcc, exec, s[4:5]
@@ -291,7 +292,7 @@
; GLOBALNESS1-NEXT: s_mov_b32 s12, s72
; GLOBALNESS1-NEXT: s_mov_b32 s13, s71
; GLOBALNESS1-NEXT: s_mov_b32 s14, s70
-; GLOBALNESS1-NEXT: v_mov_b32_e32 v31, v41
+; GLOBALNESS1-NEXT: v_mov_b32_e32 v31, v42
; GLOBALNESS1-NEXT: s_getpc_b64 s[16:17]
; GLOBALNESS1-NEXT: s_add_u32 s16, s16, widget@rel32@lo+4
; GLOBALNESS1-NEXT: s_addc_u32 s17, s17, widget@rel32@hi+12
@@ -309,7 +310,7 @@
; GLOBALNESS1-NEXT: s_mov_b32 s12, s72
; GLOBALNESS1-NEXT: s_mov_b32 s13, s71
; GLOBALNESS1-NEXT: s_mov_b32 s14, s70
-; GLOBALNESS1-NEXT: v_mov_b32_e32 v31, v41
+; GLOBALNESS1-NEXT: v_mov_b32_e32 v31, v42
; GLOBALNESS1-NEXT: s_getpc_b64 s[16:17]
; GLOBALNESS1-NEXT: s_add_u32 s16, s16, widget@rel32@lo+4
; GLOBALNESS1-NEXT: s_addc_u32 s17, s17, widget@rel32@hi+12
@@ -321,12 +322,12 @@
; GLOBALNESS0-NEXT: s_mov_b64 s[36:37], s[6:7]
; GLOBALNESS0-NEXT: s_load_dwordx4 s[72:75], s[8:9], 0x0
; GLOBALNESS0-NEXT: s_load_dword s6, s[8:9], 0x14
-; GLOBALNESS0-NEXT: v_mov_b32_e32 v41, v0
-; GLOBALNESS0-NEXT: v_mov_b32_e32 v42, 0
+; GLOBALNESS0-NEXT: v_mov_b32_e32 v42, v0
+; GLOBALNESS0-NEXT: v_mov_b32_e32 v40, 0
; GLOBALNESS0-NEXT: v_pk_mov_b32 v[0:1], 0, 0
-; GLOBALNESS0-NEXT: global_store_dword v[0:1], v42, off
+; GLOBALNESS0-NEXT: global_store_dword v[0:1], v40, off
; GLOBALNESS0-NEXT: s_waitcnt lgkmcnt(0)
-; GLOBALNESS0-NEXT: global_load_dword v0, v42, s[72:73]
+; GLOBALNESS0-NEXT: global_load_dword v0, v40, s[72:73]
; GLOBALNESS0-NEXT: s_mov_b64 s[40:41], s[4:5]
; GLOBALNESS0-NEXT: s_load_dwordx2 s[4:5], s[8:9], 0x18
; GLOBALNESS0-NEXT: s_load_dword s7, s[8:9], 0x20
@@ -334,10 +335,10 @@
; GLOBALNESS0-NEXT: s_addc_u32 flat_scratch_hi, s13, 0
; GLOBALNESS0-NEXT: s_add_u32 s0, s0, s17
; GLOBALNESS0-NEXT: s_addc_u32 s1, s1, 0
-; GLOBALNESS0-NEXT: v_mov_b32_e32 v43, 0x40994400
+; GLOBALNESS0-NEXT: v_mov_b32_e32 v41, 0x40994400
; GLOBALNESS0-NEXT: s_bitcmp1_b32 s74, 0
; GLOBALNESS0-NEXT: s_waitcnt lgkmcnt(0)
-; GLOBALNESS0-NEXT: v_cmp_ngt_f64_e32 vcc, s[4:5], v[42:43]
+; GLOBALNESS0-NEXT: v_cmp_ngt_f64_e32 vcc, s[4:5], v[40:41]
; GLOBALNESS0-NEXT: v_cmp_ngt_f64_e64 s[4:5], s[4:5], 0
; GLOBALNESS0-NEXT: v_cndmask_b32_e64 v2, 0, 1, s[4:5]
; GLOBALNESS0-NEXT: s_cselect_b64 s[4:5], -1, 0
@@ -403,9 +404,9 @@
; GLOBALNESS0-NEXT: ; =>This Loop Header: Depth=1
; GLOBALNESS0-NEXT: ; Child Loop BB1_15 Depth 2
; GLOBALNESS0-NEXT: v_pk_mov_b32 v[0:1], s[76:77], s[76:77] op_sel:[0,1]
-; GLOBALNESS0-NEXT: flat_load_dword v40, v[0:1]
+; GLOBALNESS0-NEXT: flat_load_dword v43, v[0:1]
; GLOBALNESS0-NEXT: s_add_u32 s8, s38, 40
-; GLOBALNESS0-NEXT: buffer_store_dword v42, off, s[0:3], 0
+; GLOBALNESS0-NEXT: buffer_store_dword v40, off, s[0:3], 0
; GLOBALNESS0-NEXT: flat_load_dword v46, v[0:1]
; GLOBALNESS0-NEXT: s_addc_u32 s9, s39, 0
; GLOBALNESS0-NEXT: s_mov_b64 s[4:5], s[40:41]
@@ -414,7 +415,7 @@
; GLOBALNESS0-NEXT: s_mov_b32 s12, s70
; GLOBALNESS0-NEXT: s_mov_b32 s13, s69
; GLOBALNESS0-NEXT: s_mov_b32 s14, s68
-; GLOBALNESS0-NEXT: v_mov_b32_e32 v31, v41
+; GLOBALNESS0-NEXT: v_mov_b32_e32 v31, v42
; GLOBALNESS0-NEXT: s_waitcnt lgkmcnt(0)
; GLOBALNESS0-NEXT: s_swappc_b64 s[30:31], s[78:79]
; GLOBALNESS0-NEXT: s_and_b64 vcc, exec, s[46:47]
@@ -457,13 +458,13 @@
; GLOBALNESS0-NEXT: s_cbranch_vccnz .LBB1_12
; GLOBALNESS0-NEXT: ; %bb.11: ; %bb39.i
; GLOBALNESS0-NEXT: ; in Loop: Header=BB1_4 Depth=1
-; GLOBALNESS0-NEXT: v_mov_b32_e32 v43, v42
+; GLOBALNESS0-NEXT: v_mov_b32_e32 v41, v40
; GLOBALNESS0-NEXT: v_pk_mov_b32 v[2:3], 0, 0
-; GLOBALNESS0-NEXT: global_store_dwordx2 v[2:3], v[42:43], off
+; GLOBALNESS0-NEXT: global_store_dwordx2 v[2:3], v[40:41], off
; GLOBALNESS0-NEXT: .LBB1_12: ; %bb44.lr.ph.i
; GLOBALNESS0-NEXT: ; in Loop: Header=BB1_4 Depth=1
; GLOBALNESS0-NEXT: v_cmp_ne_u32_e32 vcc, 0, v46
-; GLOBALNESS0-NEXT: v_cndmask_b32_e32 v2, 0, v40, vcc
+; GLOBALNESS0-NEXT: v_cndmask_b32_e32 v2, 0, v43, vcc
; GLOBALNESS0-NEXT: s_waitcnt vmcnt(0)
; GLOBALNESS0-NEXT: v_cmp_nlt_f64_e64 s[64:65], 0, v[0:1]
; GLOBALNESS0-NEXT: v_cmp_eq_u32_e64 s[66:67], 0, v2
@@ -510,7 +511,7 @@
; GLOBALNESS0-NEXT: s_mov_b32 s12, s70
; GLOBALNESS0-NEXT: s_mov_b32 s13, s69
; GLOBALNESS0-NEXT: s_mov_b32 s14, s68
-; GLOBALNESS0-NEXT: v_mov_b32_e32 v31, v41
+; GLOBALNESS0-NEXT: v_mov_b32_e32 v31, v42
; GLOBALNESS0-NEXT: s_swappc_b64 s[30:31], s[78:79]
; GLOBALNESS0-NEXT: v_pk_mov_b32 v[46:47], 0, 0
; GLOBALNESS0-NEXT: s_mov_b64 s[4:5], s[40:41]
@@ -520,15 +521,15 @@
; GLOBALNESS0-NEXT: s_mov_b32 s12, s70
; GLOBALNESS0-NEXT: s_mov_b32 s13, s69
; GLOBALNESS0-NEXT: s_mov_b32 s14, s68
-; GLOBALNESS0-NEXT: v_mov_b32_e32 v31, v41
+; GLOBALNESS0-NEXT: v_mov_b32_e32 v31, v42
; GLOBALNESS0-NEXT: global_store_dwordx2 v[46:47], v[44:45], off
; GLOBALNESS0-NEXT: s_swappc_b64 s[30:31], s[78:79]
; GLOBALNESS0-NEXT: s_and_saveexec_b64 s[4:5], s[66:67]
; GLOBALNESS0-NEXT: s_cbranch_execz .LBB1_13
; GLOBALNESS0-NEXT: ; %bb.22: ; %bb62.i
; GLOBALNESS0-NEXT: ; in Loop: Header=BB1_15 Depth=2
-; GLOBALNESS0-NEXT: v_mov_b32_e32 v43, v42
-; GLOBALNESS0-NEXT: global_store_dwordx2 v[46:47], v[42:43], off
+; GLOBALNESS0-NEXT: v_mov_b32_e32 v41, v40
+; GLOBALNESS0-NEXT: global_store_dwordx2 v[46:47], v[40:41], off
; GLOBALNESS0-NEXT: s_branch .LBB1_13
; GLOBALNESS0-NEXT: .LBB1_23: ; %LeafBlock
; GLOBALNESS0-NEXT: ; in Loop: Header=BB1_4 Depth=1
@@ -555,15 +556,15 @@
; GLOBALNESS0-NEXT: s_cbranch_vccnz .LBB1_1
; GLOBALNESS0-NEXT: ; %bb.28: ; %bb69.i
; GLOBALNESS0-NEXT: ; in Loop: Header=BB1_4 Depth=1
-; GLOBALNESS0-NEXT: v_mov_b32_e32 v43, v42
+; GLOBALNESS0-NEXT: v_mov_b32_e32 v41, v40
; GLOBALNESS0-NEXT: v_pk_mov_b32 v[2:3], 0, 0
-; GLOBALNESS0-NEXT: global_store_dwordx2 v[2:3], v[42:43], off
+; GLOBALNESS0-NEXT: global_store_dwordx2 v[2:3], v[40:41], off
; GLOBALNESS0-NEXT: s_branch .LBB1_1
; GLOBALNESS0-NEXT: .LBB1_29: ; %bb73.i
; GLOBALNESS0-NEXT: ; in Loop: Header=BB1_4 Depth=1
-; GLOBALNESS0-NEXT: v_mov_b32_e32 v43, v42
+; GLOBALNESS0-NEXT: v_mov_b32_e32 v41, v40
; GLOBALNESS0-NEXT: v_pk_mov_b32 v[2:3], 0, 0
-; GLOBALNESS0-NEXT: global_store_dwordx2 v[2:3], v[42:43], off
+; GLOBALNESS0-NEXT: global_store_dwordx2 v[2:3], v[40:41], off
; GLOBALNESS0-NEXT: s_branch .LBB1_2
; GLOBALNESS0-NEXT: .LBB1_30: ; %loop.exit.guard
; GLOBALNESS0-NEXT: s_andn2_b64 vcc, exec, s[4:5]
@@ -578,7 +579,7 @@
; GLOBALNESS0-NEXT: s_mov_b32 s12, s70
; GLOBALNESS0-NEXT: s_mov_b32 s13, s69
; GLOBALNESS0-NEXT: s_mov_b32 s14, s68
-; GLOBALNESS0-NEXT: v_mov_b32_e32 v31, v41
+; GLOBALNESS0-NEXT: v_mov_b32_e32 v31, v42
; GLOBALNESS0-NEXT: s_getpc_b64 s[16:17]
; GLOBALNESS0-NEXT: s_add_u32 s16, s16, widget@rel32@lo+4
; GLOBALNESS0-NEXT: s_addc_u32 s17, s17, widget@rel32@hi+12
@@ -596,7 +597,7 @@
; GLOBALNESS0-NEXT: s_mov_b32 s12, s70
; GLOBALNESS0-NEXT: s_mov_b32 s13, s69
; GLOBALNESS0-NEXT: s_mov_b32 s14, s68
-; GLOBALNESS0-NEXT: v_mov_b32_e32 v31, v41
+; GLOBALNESS0-NEXT: v_mov_b32_e32 v31, v42
; GLOBALNESS0-NEXT: s_getpc_b64 s[16:17]
; GLOBALNESS0-NEXT: s_add_u32 s16, s16, widget@rel32@lo+4
; GLOBALNESS0-NEXT: s_addc_u32 s17, s17, widget@rel32@hi+12
diff --git a/llvm/test/CodeGen/AMDGPU/unstructured-cfg-def-use-issue.ll b/llvm/test/CodeGen/AMDGPU/unstructured-cfg-def-use-issue.ll
--- a/llvm/test/CodeGen/AMDGPU/unstructured-cfg-def-use-issue.ll
+++ b/llvm/test/CodeGen/AMDGPU/unstructured-cfg-def-use-issue.ll
@@ -10,8 +10,9 @@
; GCN-NEXT: s_mov_b32 s33, s32
; GCN-NEXT: s_or_saveexec_b64 s[18:19], -1
; GCN-NEXT: buffer_store_dword v40, off, s[0:3], s33 offset:4 ; 4-byte Folded Spill
+; GCN-NEXT: buffer_store_dword v42, off, s[0:3], s33 offset:8 ; 4-byte Folded Spill
; GCN-NEXT: s_mov_b64 exec, s[18:19]
-; GCN-NEXT: v_writelane_b32 v40, s16, 16
+; GCN-NEXT: v_writelane_b32 v42, s16, 0
; GCN-NEXT: s_addk_i32 s32, 0x400
; GCN-NEXT: buffer_store_dword v41, off, s[0:3], s33 ; 4-byte Folded Spill
; GCN-NEXT: v_writelane_b32 v40, s30, 0
@@ -119,9 +120,10 @@
; GCN-NEXT: v_readlane_b32 s31, v40, 1
; GCN-NEXT: v_readlane_b32 s30, v40, 0
; GCN-NEXT: buffer_load_dword v41, off, s[0:3], s33 ; 4-byte Folded Reload
-; GCN-NEXT: v_readlane_b32 s4, v40, 16
+; GCN-NEXT: v_readlane_b32 s4, v42, 0
; GCN-NEXT: s_or_saveexec_b64 s[6:7], -1
; GCN-NEXT: buffer_load_dword v40, off, s[0:3], s33 offset:4 ; 4-byte Folded Reload
+; GCN-NEXT: buffer_load_dword v42, off, s[0:3], s33 offset:8 ; 4-byte Folded Reload
; GCN-NEXT: s_mov_b64 exec, s[6:7]
; GCN-NEXT: s_addk_i32 s32, 0xfc00
; GCN-NEXT: s_mov_b32 s33, s4
@@ -267,8 +269,9 @@
; GCN-NEXT: s_mov_b32 s33, s32
; GCN-NEXT: s_or_saveexec_b64 s[18:19], -1
; GCN-NEXT: buffer_store_dword v40, off, s[0:3], s33 offset:20 ; 4-byte Folded Spill
+; GCN-NEXT: buffer_store_dword v46, off, s[0:3], s33 offset:24 ; 4-byte Folded Spill
; GCN-NEXT: s_mov_b64 exec, s[18:19]
-; GCN-NEXT: v_writelane_b32 v40, s16, 28
+; GCN-NEXT: v_writelane_b32 v46, s16, 0
; GCN-NEXT: s_addk_i32 s32, 0x800
; GCN-NEXT: buffer_store_dword v41, off, s[0:3], s33 offset:16 ; 4-byte Folded Spill
; GCN-NEXT: buffer_store_dword v42, off, s[0:3], s33 offset:12 ; 4-byte Folded Spill
@@ -471,9 +474,10 @@
; GCN-NEXT: buffer_load_dword v43, off, s[0:3], s33 offset:8 ; 4-byte Folded Reload
; GCN-NEXT: buffer_load_dword v42, off, s[0:3], s33 offset:12 ; 4-byte Folded Reload
; GCN-NEXT: buffer_load_dword v41, off, s[0:3], s33 offset:16 ; 4-byte Folded Reload
-; GCN-NEXT: v_readlane_b32 s4, v40, 28
+; GCN-NEXT: v_readlane_b32 s4, v46, 0
; GCN-NEXT: s_or_saveexec_b64 s[6:7], -1
; GCN-NEXT: buffer_load_dword v40, off, s[0:3], s33 offset:20 ; 4-byte Folded Reload
+; GCN-NEXT: buffer_load_dword v46, off, s[0:3], s33 offset:24 ; 4-byte Folded Reload
; GCN-NEXT: s_mov_b64 exec, s[6:7]
; GCN-NEXT: s_addk_i32 s32, 0xf800
; GCN-NEXT: s_mov_b32 s33, s4
diff --git a/llvm/test/CodeGen/AMDGPU/vgpr-spill-placement-issue61083.ll b/llvm/test/CodeGen/AMDGPU/vgpr-spill-placement-issue61083.ll
--- a/llvm/test/CodeGen/AMDGPU/vgpr-spill-placement-issue61083.ll
+++ b/llvm/test/CodeGen/AMDGPU/vgpr-spill-placement-issue61083.ll
@@ -13,61 +13,47 @@
; CHECK: ; %bb.0: ; %bb
; CHECK-NEXT: s_add_u32 s0, s0, s15
; CHECK-NEXT: s_addc_u32 s1, s1, 0
-; CHECK-NEXT: ; implicit-def: $vgpr1
; CHECK-NEXT: v_mov_b32_e32 v2, v0
-; CHECK-NEXT: s_or_saveexec_b64 s[8:9], -1
-; CHECK-NEXT: buffer_load_dword v0, off, s[0:3], 0 offset:4 ; 4-byte Folded Reload
-; CHECK-NEXT: s_mov_b64 exec, s[8:9]
-; CHECK-NEXT: v_mov_b32_e32 v1, 0
-; CHECK-NEXT: global_load_ushort v3, v1, s[4:5] offset:4
+; CHECK-NEXT: v_mov_b32_e32 v0, 0
+; CHECK-NEXT: global_load_ushort v3, v0, s[4:5] offset:4
; CHECK-NEXT: s_waitcnt vmcnt(0)
-; CHECK-NEXT: buffer_store_dword v3, off, s[0:3], 0 offset:8 ; 4-byte Folded Spill
+; CHECK-NEXT: buffer_store_dword v3, off, s[0:3], 0 offset:4 ; 4-byte Folded Spill
; CHECK-NEXT: ; implicit-def: $sgpr4
; CHECK-NEXT: s_mov_b32 s4, 0
; CHECK-NEXT: v_cmp_eq_u32_e64 s[6:7], v2, s4
; CHECK-NEXT: s_mov_b32 s4, 0
; CHECK-NEXT: v_mov_b32_e32 v2, s4
-; CHECK-NEXT: ds_write_b8 v1, v2
+; CHECK-NEXT: ds_write_b8 v0, v2
; CHECK-NEXT: s_mov_b64 s[4:5], exec
-; CHECK-NEXT: v_writelane_b32 v0, s4, 0
-; CHECK-NEXT: v_writelane_b32 v0, s5, 1
-; CHECK-NEXT: s_or_saveexec_b64 s[8:9], -1
-; CHECK-NEXT: buffer_store_dword v0, off, s[0:3], 0 offset:4 ; 4-byte Folded Spill
-; CHECK-NEXT: s_mov_b64 exec, s[8:9]
+; CHECK-NEXT: v_writelane_b32 v1, s4, 0
+; CHECK-NEXT: v_writelane_b32 v1, s5, 1
; CHECK-NEXT: s_and_b64 s[4:5], s[4:5], s[6:7]
; CHECK-NEXT: s_mov_b64 exec, s[4:5]
; CHECK-NEXT: s_cbranch_execz .LBB0_2
; CHECK-NEXT: ; %bb.1: ; %bb193
; CHECK-NEXT: .LBB0_2: ; %bb194
-; CHECK-NEXT: buffer_load_dword v0, off, s[0:3], 0 offset:8 ; 4-byte Folded Reload
-; CHECK-NEXT: s_or_saveexec_b64 s[8:9], -1
-; CHECK-NEXT: buffer_load_dword v1, off, s[0:3], 0 offset:4 ; 4-byte Folded Reload
-; CHECK-NEXT: s_mov_b64 exec, s[8:9]
-; CHECK-NEXT: s_waitcnt vmcnt(0)
+; CHECK-NEXT: buffer_load_dword v0, off, s[0:3], 0 offset:4 ; 4-byte Folded Reload
; CHECK-NEXT: v_readlane_b32 s4, v1, 0
; CHECK-NEXT: v_readlane_b32 s5, v1, 1
; CHECK-NEXT: s_or_b64 exec, exec, s[4:5]
; CHECK-NEXT: s_mov_b32 s4, 0
+; CHECK-NEXT: s_waitcnt vmcnt(0)
; CHECK-NEXT: v_cmp_ne_u16_e64 s[4:5], v0, s4
; CHECK-NEXT: s_and_b64 vcc, exec, s[4:5]
; CHECK-NEXT: s_cbranch_vccnz .LBB0_4
; CHECK-NEXT: ; %bb.3: ; %bb201
-; CHECK-NEXT: buffer_load_dword v1, off, s[0:3], 0 offset:8 ; 4-byte Folded Reload
+; CHECK-NEXT: buffer_load_dword v2, off, s[0:3], 0 offset:4 ; 4-byte Folded Reload
; CHECK-NEXT: s_getpc_b64 s[4:5]
; CHECK-NEXT: s_add_u32 s4, s4, V2@rel32@lo+4
; CHECK-NEXT: s_addc_u32 s5, s5, V2@rel32@hi+12
; CHECK-NEXT: v_mov_b32_e32 v0, 0
; CHECK-NEXT: s_waitcnt vmcnt(0)
-; CHECK-NEXT: global_store_short v0, v1, s[4:5]
+; CHECK-NEXT: global_store_short v0, v2, s[4:5]
; CHECK-NEXT: s_waitcnt vmcnt(0) lgkmcnt(0)
; CHECK-NEXT: s_barrier
; CHECK-NEXT: s_trap 2
; CHECK-NEXT: ; divergent unreachable
; CHECK-NEXT: .LBB0_4: ; %UnifiedReturnBlock
-; CHECK-NEXT: s_or_saveexec_b64 s[8:9], -1
-; CHECK-NEXT: buffer_load_dword v0, off, s[0:3], 0 offset:4 ; 4-byte Folded Reload
-; CHECK-NEXT: s_mov_b64 exec, s[8:9]
-; CHECK-NEXT: ; kill: killed $vgpr0
; CHECK-NEXT: s_endpgm
bb:
%i10 = tail call i32 @llvm.amdgcn.workitem.id.x()
diff --git a/llvm/test/CodeGen/AMDGPU/vgpr-tuple-allocation.ll b/llvm/test/CodeGen/AMDGPU/vgpr-tuple-allocation.ll
--- a/llvm/test/CodeGen/AMDGPU/vgpr-tuple-allocation.ll
+++ b/llvm/test/CodeGen/AMDGPU/vgpr-tuple-allocation.ll
@@ -15,6 +15,7 @@
; GFX9-NEXT: s_mov_b32 s33, s32
; GFX9-NEXT: s_or_saveexec_b64 s[6:7], -1
; GFX9-NEXT: buffer_store_dword v40, off, s[0:3], s33 offset:16 ; 4-byte Folded Spill
+; GFX9-NEXT: buffer_store_dword v45, off, s[0:3], s33 offset:20 ; 4-byte Folded Spill
; GFX9-NEXT: s_mov_b64 exec, s[6:7]
; GFX9-NEXT: v_mov_b32_e32 v36, v16
; GFX9-NEXT: v_mov_b32_e32 v35, v15
@@ -35,7 +36,7 @@
; GFX9-NEXT: ;;#ASMEND
; GFX9-NEXT: image_gather4_c_b_cl v[41:44], v[32:36], s[4:11], s[4:7] dmask:0x1
; GFX9-NEXT: s_addk_i32 s32, 0x800
-; GFX9-NEXT: v_writelane_b32 v40, s4, 2
+; GFX9-NEXT: v_writelane_b32 v45, s4, 0
; GFX9-NEXT: s_getpc_b64 s[4:5]
; GFX9-NEXT: s_add_u32 s4, s4, extern_func@gotpcrel32@lo+4
; GFX9-NEXT: s_addc_u32 s5, s5, extern_func@gotpcrel32@hi+12
@@ -54,9 +55,10 @@
; GFX9-NEXT: buffer_load_dword v41, off, s[0:3], s33 offset:12 ; 4-byte Folded Reload
; GFX9-NEXT: v_readlane_b32 s31, v40, 1
; GFX9-NEXT: v_readlane_b32 s30, v40, 0
-; GFX9-NEXT: v_readlane_b32 s4, v40, 2
+; GFX9-NEXT: v_readlane_b32 s4, v45, 0
; GFX9-NEXT: s_or_saveexec_b64 s[6:7], -1
; GFX9-NEXT: buffer_load_dword v40, off, s[0:3], s33 offset:16 ; 4-byte Folded Reload
+; GFX9-NEXT: buffer_load_dword v45, off, s[0:3], s33 offset:20 ; 4-byte Folded Reload
; GFX9-NEXT: s_mov_b64 exec, s[6:7]
; GFX9-NEXT: s_addk_i32 s32, 0xf800
; GFX9-NEXT: s_mov_b32 s33, s4
@@ -70,6 +72,7 @@
; GFX10-NEXT: s_mov_b32 s33, s32
; GFX10-NEXT: s_or_saveexec_b32 s5, -1
; GFX10-NEXT: buffer_store_dword v40, off, s[0:3], s33 offset:16 ; 4-byte Folded Spill
+; GFX10-NEXT: buffer_store_dword v45, off, s[0:3], s33 offset:20 ; 4-byte Folded Spill
; GFX10-NEXT: s_waitcnt_depctr 0xffe3
; GFX10-NEXT: s_mov_b32 exec_lo, s5
; GFX10-NEXT: v_mov_b32_e32 v36, v16
@@ -91,12 +94,12 @@
; GFX10-NEXT: ;;#ASMEND
; GFX10-NEXT: image_gather4_c_b_cl v[41:44], v[32:36], s[4:11], s[4:7] dmask:0x1 dim:SQ_RSRC_IMG_2D
; GFX10-NEXT: s_addk_i32 s32, 0x400
-; GFX10-NEXT: v_writelane_b32 v40, s4, 2
+; GFX10-NEXT: v_writelane_b32 v45, s4, 0
; GFX10-NEXT: s_getpc_b64 s[4:5]
; GFX10-NEXT: s_add_u32 s4, s4, extern_func@gotpcrel32@lo+4
; GFX10-NEXT: s_addc_u32 s5, s5, extern_func@gotpcrel32@hi+12
-; GFX10-NEXT: s_load_dwordx2 s[4:5], s[4:5], 0x0
; GFX10-NEXT: v_writelane_b32 v40, s30, 0
+; GFX10-NEXT: s_load_dwordx2 s[4:5], s[4:5], 0x0
; GFX10-NEXT: v_writelane_b32 v40, s31, 1
; GFX10-NEXT: s_waitcnt lgkmcnt(0)
; GFX10-NEXT: s_swappc_b64 s[30:31], s[4:5]
@@ -111,9 +114,11 @@
; GFX10-NEXT: buffer_load_dword v41, off, s[0:3], s33 offset:12
; GFX10-NEXT: v_readlane_b32 s31, v40, 1
; GFX10-NEXT: v_readlane_b32 s30, v40, 0
-; GFX10-NEXT: v_readlane_b32 s4, v40, 2
+; GFX10-NEXT: v_readlane_b32 s4, v45, 0
; GFX10-NEXT: s_or_saveexec_b32 s5, -1
-; GFX10-NEXT: buffer_load_dword v40, off, s[0:3], s33 offset:16 ; 4-byte Folded Reload
+; GFX10-NEXT: s_clause 0x1
+; GFX10-NEXT: buffer_load_dword v40, off, s[0:3], s33 offset:16
+; GFX10-NEXT: buffer_load_dword v45, off, s[0:3], s33 offset:20
; GFX10-NEXT: s_waitcnt_depctr 0xffe3
; GFX10-NEXT: s_mov_b32 exec_lo, s5
; GFX10-NEXT: s_addk_i32 s32, 0xfc00
@@ -127,7 +132,9 @@
; GFX11-NEXT: s_mov_b32 s0, s33
; GFX11-NEXT: s_mov_b32 s33, s32
; GFX11-NEXT: s_or_saveexec_b32 s1, -1
-; GFX11-NEXT: scratch_store_b32 off, v40, s33 offset:16 ; 4-byte Folded Spill
+; GFX11-NEXT: s_clause 0x1
+; GFX11-NEXT: scratch_store_b32 off, v40, s33 offset:16
+; GFX11-NEXT: scratch_store_b32 off, v45, s33 offset:20
; GFX11-NEXT: s_mov_b32 exec_lo, s1
; GFX11-NEXT: v_dual_mov_b32 v36, v16 :: v_dual_mov_b32 v35, v15
; GFX11-NEXT: v_dual_mov_b32 v34, v14 :: v_dual_mov_b32 v33, v13
@@ -147,12 +154,12 @@
; GFX11-NEXT: ;;#ASMEND
; GFX11-NEXT: image_gather4_c_b_cl v[41:44], v[32:36], s[0:7], s[0:3] dmask:0x1 dim:SQ_RSRC_IMG_2D
; GFX11-NEXT: s_add_i32 s32, s32, 32
-; GFX11-NEXT: v_writelane_b32 v40, s0, 2
+; GFX11-NEXT: v_writelane_b32 v45, s0, 0
; GFX11-NEXT: s_getpc_b64 s[0:1]
; GFX11-NEXT: s_add_u32 s0, s0, extern_func@gotpcrel32@lo+4
; GFX11-NEXT: s_addc_u32 s1, s1, extern_func@gotpcrel32@hi+12
-; GFX11-NEXT: s_load_b64 s[0:1], s[0:1], 0x0
; GFX11-NEXT: v_writelane_b32 v40, s30, 0
+; GFX11-NEXT: s_load_b64 s[0:1], s[0:1], 0x0
; GFX11-NEXT: v_writelane_b32 v40, s31, 1
; GFX11-NEXT: s_waitcnt lgkmcnt(0)
; GFX11-NEXT: s_swappc_b64 s[30:31], s[0:1]
@@ -165,9 +172,11 @@
; GFX11-NEXT: scratch_load_b32 v41, off, s33 offset:12
; GFX11-NEXT: v_readlane_b32 s31, v40, 1
; GFX11-NEXT: v_readlane_b32 s30, v40, 0
-; GFX11-NEXT: v_readlane_b32 s0, v40, 2
+; GFX11-NEXT: v_readlane_b32 s0, v45, 0
; GFX11-NEXT: s_or_saveexec_b32 s1, -1
-; GFX11-NEXT: scratch_load_b32 v40, off, s33 offset:16 ; 4-byte Folded Reload
+; GFX11-NEXT: s_clause 0x1
+; GFX11-NEXT: scratch_load_b32 v40, off, s33 offset:16
+; GFX11-NEXT: scratch_load_b32 v45, off, s33 offset:20
; GFX11-NEXT: s_mov_b32 exec_lo, s1
; GFX11-NEXT: s_addk_i32 s32, 0xffe0
; GFX11-NEXT: s_mov_b32 s33, s0
@@ -205,6 +214,7 @@
; GFX9-NEXT: s_mov_b32 s33, s32
; GFX9-NEXT: s_or_saveexec_b64 s[6:7], -1
; GFX9-NEXT: buffer_store_dword v40, off, s[0:3], s33 offset:20 ; 4-byte Folded Spill
+; GFX9-NEXT: buffer_store_dword v46, off, s[0:3], s33 offset:24 ; 4-byte Folded Spill
; GFX9-NEXT: s_mov_b64 exec, s[6:7]
; GFX9-NEXT: buffer_store_dword v41, off, s[0:3], s33 offset:16 ; 4-byte Folded Spill
; GFX9-NEXT: buffer_store_dword v42, off, s[0:3], s33 offset:12 ; 4-byte Folded Spill
@@ -218,7 +228,7 @@
; GFX9-NEXT: v_mov_b32_e32 v41, v12
; GFX9-NEXT: image_gather4_c_b_cl v[0:3], v[41:45], s[4:11], s[4:7] dmask:0x1
; GFX9-NEXT: s_addk_i32 s32, 0x800
-; GFX9-NEXT: v_writelane_b32 v40, s4, 2
+; GFX9-NEXT: v_writelane_b32 v46, s4, 0
; GFX9-NEXT: s_getpc_b64 s[4:5]
; GFX9-NEXT: s_add_u32 s4, s4, extern_func@gotpcrel32@lo+4
; GFX9-NEXT: s_addc_u32 s5, s5, extern_func@gotpcrel32@hi+12
@@ -238,9 +248,10 @@
; GFX9-NEXT: buffer_load_dword v41, off, s[0:3], s33 offset:16 ; 4-byte Folded Reload
; GFX9-NEXT: v_readlane_b32 s31, v40, 1
; GFX9-NEXT: v_readlane_b32 s30, v40, 0
-; GFX9-NEXT: v_readlane_b32 s4, v40, 2
+; GFX9-NEXT: v_readlane_b32 s4, v46, 0
; GFX9-NEXT: s_or_saveexec_b64 s[6:7], -1
; GFX9-NEXT: buffer_load_dword v40, off, s[0:3], s33 offset:20 ; 4-byte Folded Reload
+; GFX9-NEXT: buffer_load_dword v46, off, s[0:3], s33 offset:24 ; 4-byte Folded Reload
; GFX9-NEXT: s_mov_b64 exec, s[6:7]
; GFX9-NEXT: s_addk_i32 s32, 0xf800
; GFX9-NEXT: s_mov_b32 s33, s4
@@ -254,6 +265,7 @@
; GFX10-NEXT: s_mov_b32 s33, s32
; GFX10-NEXT: s_or_saveexec_b32 s5, -1
; GFX10-NEXT: buffer_store_dword v40, off, s[0:3], s33 offset:20 ; 4-byte Folded Spill
+; GFX10-NEXT: buffer_store_dword v46, off, s[0:3], s33 offset:24 ; 4-byte Folded Spill
; GFX10-NEXT: s_waitcnt_depctr 0xffe3
; GFX10-NEXT: s_mov_b32 exec_lo, s5
; GFX10-NEXT: buffer_store_dword v41, off, s[0:3], s33 offset:16 ; 4-byte Folded Spill
@@ -263,18 +275,18 @@
; GFX10-NEXT: buffer_store_dword v45, off, s[0:3], s33 ; 4-byte Folded Spill
; GFX10-NEXT: image_gather4_c_b_cl v[0:3], v[12:16], s[4:11], s[4:7] dmask:0x1 dim:SQ_RSRC_IMG_2D
; GFX10-NEXT: s_addk_i32 s32, 0x400
-; GFX10-NEXT: v_writelane_b32 v40, s4, 2
+; GFX10-NEXT: v_writelane_b32 v46, s4, 0
; GFX10-NEXT: s_getpc_b64 s[4:5]
; GFX10-NEXT: s_add_u32 s4, s4, extern_func@gotpcrel32@lo+4
; GFX10-NEXT: s_addc_u32 s5, s5, extern_func@gotpcrel32@hi+12
-; GFX10-NEXT: v_mov_b32_e32 v41, v16
+; GFX10-NEXT: v_writelane_b32 v40, s30, 0
; GFX10-NEXT: s_load_dwordx2 s[4:5], s[4:5], 0x0
+; GFX10-NEXT: v_mov_b32_e32 v41, v16
; GFX10-NEXT: v_mov_b32_e32 v42, v15
-; GFX10-NEXT: v_writelane_b32 v40, s30, 0
; GFX10-NEXT: v_mov_b32_e32 v43, v14
+; GFX10-NEXT: v_writelane_b32 v40, s31, 1
; GFX10-NEXT: v_mov_b32_e32 v44, v13
; GFX10-NEXT: v_mov_b32_e32 v45, v12
-; GFX10-NEXT: v_writelane_b32 v40, s31, 1
; GFX10-NEXT: s_waitcnt vmcnt(0)
; GFX10-NEXT: global_store_dwordx4 v[0:1], v[0:3], off
; GFX10-NEXT: s_waitcnt lgkmcnt(0)
@@ -288,9 +300,11 @@
; GFX10-NEXT: buffer_load_dword v41, off, s[0:3], s33 offset:16
; GFX10-NEXT: v_readlane_b32 s31, v40, 1
; GFX10-NEXT: v_readlane_b32 s30, v40, 0
-; GFX10-NEXT: v_readlane_b32 s4, v40, 2
+; GFX10-NEXT: v_readlane_b32 s4, v46, 0
; GFX10-NEXT: s_or_saveexec_b32 s5, -1
-; GFX10-NEXT: buffer_load_dword v40, off, s[0:3], s33 offset:20 ; 4-byte Folded Reload
+; GFX10-NEXT: s_clause 0x1
+; GFX10-NEXT: buffer_load_dword v40, off, s[0:3], s33 offset:20
+; GFX10-NEXT: buffer_load_dword v46, off, s[0:3], s33 offset:24
; GFX10-NEXT: s_waitcnt_depctr 0xffe3
; GFX10-NEXT: s_mov_b32 exec_lo, s5
; GFX10-NEXT: s_addk_i32 s32, 0xfc00
@@ -304,7 +318,9 @@
; GFX11-NEXT: s_mov_b32 s0, s33
; GFX11-NEXT: s_mov_b32 s33, s32
; GFX11-NEXT: s_or_saveexec_b32 s1, -1
-; GFX11-NEXT: scratch_store_b32 off, v40, s33 offset:20 ; 4-byte Folded Spill
+; GFX11-NEXT: s_clause 0x1
+; GFX11-NEXT: scratch_store_b32 off, v40, s33 offset:20
+; GFX11-NEXT: scratch_store_b32 off, v46, s33 offset:24
; GFX11-NEXT: s_mov_b32 exec_lo, s1
; GFX11-NEXT: s_clause 0x4
; GFX11-NEXT: scratch_store_b32 off, v41, s33 offset:16
@@ -314,16 +330,16 @@
; GFX11-NEXT: scratch_store_b32 off, v45, s33
; GFX11-NEXT: image_gather4_c_b_cl v[0:3], v[12:16], s[0:7], s[0:3] dmask:0x1 dim:SQ_RSRC_IMG_2D
; GFX11-NEXT: s_add_i32 s32, s32, 32
-; GFX11-NEXT: v_writelane_b32 v40, s0, 2
+; GFX11-NEXT: v_writelane_b32 v46, s0, 0
; GFX11-NEXT: s_getpc_b64 s[0:1]
; GFX11-NEXT: s_add_u32 s0, s0, extern_func@gotpcrel32@lo+4
; GFX11-NEXT: s_addc_u32 s1, s1, extern_func@gotpcrel32@hi+12
-; GFX11-NEXT: v_dual_mov_b32 v41, v16 :: v_dual_mov_b32 v42, v15
-; GFX11-NEXT: s_load_b64 s[0:1], s[0:1], 0x0
; GFX11-NEXT: v_writelane_b32 v40, s30, 0
+; GFX11-NEXT: s_load_b64 s[0:1], s[0:1], 0x0
+; GFX11-NEXT: v_dual_mov_b32 v41, v16 :: v_dual_mov_b32 v42, v15
; GFX11-NEXT: v_dual_mov_b32 v43, v14 :: v_dual_mov_b32 v44, v13
-; GFX11-NEXT: v_mov_b32_e32 v45, v12
; GFX11-NEXT: v_writelane_b32 v40, s31, 1
+; GFX11-NEXT: v_mov_b32_e32 v45, v12
; GFX11-NEXT: s_waitcnt vmcnt(0)
; GFX11-NEXT: global_store_b128 v[0:1], v[0:3], off
; GFX11-NEXT: s_waitcnt lgkmcnt(0)
@@ -337,9 +353,11 @@
; GFX11-NEXT: scratch_load_b32 v41, off, s33 offset:16
; GFX11-NEXT: v_readlane_b32 s31, v40, 1
; GFX11-NEXT: v_readlane_b32 s30, v40, 0
-; GFX11-NEXT: v_readlane_b32 s0, v40, 2
+; GFX11-NEXT: v_readlane_b32 s0, v46, 0
; GFX11-NEXT: s_or_saveexec_b32 s1, -1
-; GFX11-NEXT: scratch_load_b32 v40, off, s33 offset:20 ; 4-byte Folded Reload
+; GFX11-NEXT: s_clause 0x1
+; GFX11-NEXT: scratch_load_b32 v40, off, s33 offset:20
+; GFX11-NEXT: scratch_load_b32 v46, off, s33 offset:24
; GFX11-NEXT: s_mov_b32 exec_lo, s1
; GFX11-NEXT: s_addk_i32 s32, 0xffe0
; GFX11-NEXT: s_mov_b32 s33, s0
diff --git a/llvm/test/CodeGen/AMDGPU/vgpr_constant_to_sgpr.ll b/llvm/test/CodeGen/AMDGPU/vgpr_constant_to_sgpr.ll
--- a/llvm/test/CodeGen/AMDGPU/vgpr_constant_to_sgpr.ll
+++ b/llvm/test/CodeGen/AMDGPU/vgpr_constant_to_sgpr.ll
@@ -9,21 +9,17 @@
define protected amdgpu_kernel void @kern(ptr %addr) !llvm.amdgcn.lds.kernel.id !0 {
; CHECK-LABEL: kern:
; CHECK: ; %bb.0:
-; CHECK-NEXT: s_mov_b32 s32, 0x200
+; CHECK-NEXT: s_mov_b32 s32, 0
; CHECK-NEXT: s_add_u32 s12, s12, s17
; CHECK-NEXT: s_addc_u32 s13, s13, 0
; CHECK-NEXT: s_setreg_b32 hwreg(HW_REG_FLAT_SCR_LO), s12
; CHECK-NEXT: s_setreg_b32 hwreg(HW_REG_FLAT_SCR_HI), s13
; CHECK-NEXT: s_add_u32 s0, s0, s17
; CHECK-NEXT: s_addc_u32 s1, s1, 0
-; CHECK-NEXT: ; implicit-def: $vgpr3
-; CHECK-NEXT: v_writelane_b32 v3, s16, 0
-; CHECK-NEXT: s_or_saveexec_b32 s33, -1
-; CHECK-NEXT: buffer_store_dword v3, off, s[0:3], 0 offset:4 ; 4-byte Folded Spill
-; CHECK-NEXT: s_mov_b32 exec_lo, s33
+; CHECK-NEXT: v_writelane_b32 v40, s16, 0
; CHECK-NEXT: s_mov_b32 s13, s15
; CHECK-NEXT: s_mov_b32 s12, s14
-; CHECK-NEXT: v_readlane_b32 s14, v3, 0
+; CHECK-NEXT: v_readlane_b32 s14, v40, 0
; CHECK-NEXT: s_mov_b64 s[16:17], s[8:9]
; CHECK-NEXT: s_load_dwordx2 s[8:9], s[16:17], 0x0
; CHECK-NEXT: v_mov_b32_e32 v5, 42
@@ -56,10 +52,6 @@
; CHECK-NEXT: s_mov_b64 s[2:3], s[22:23]
; CHECK-NEXT: s_waitcnt lgkmcnt(0)
; CHECK-NEXT: s_swappc_b64 s[30:31], s[16:17]
-; CHECK-NEXT: s_or_saveexec_b32 s33, -1
-; CHECK-NEXT: buffer_load_dword v0, off, s[0:3], 0 offset:4 ; 4-byte Folded Reload
-; CHECK-NEXT: s_mov_b32 exec_lo, s33
-; CHECK-NEXT: ; kill: killed $vgpr0
; CHECK-NEXT: s_endpgm
store i32 42, ptr %addr
call fastcc void @unknown_call()
diff --git a/llvm/test/CodeGen/AMDGPU/wave32.ll b/llvm/test/CodeGen/AMDGPU/wave32.ll
--- a/llvm/test/CodeGen/AMDGPU/wave32.ll
+++ b/llvm/test/CodeGen/AMDGPU/wave32.ll
@@ -2856,23 +2856,26 @@
; GFX1032-NEXT: s_mov_b32 s33, s32
; GFX1032-NEXT: s_or_saveexec_b32 s17, -1
; GFX1032-NEXT: buffer_store_dword v40, off, s[0:3], s33 ; 4-byte Folded Spill
+; GFX1032-NEXT: buffer_store_dword v41, off, s[0:3], s33 offset:4 ; 4-byte Folded Spill
; GFX1032-NEXT: s_waitcnt_depctr 0xffe3
; GFX1032-NEXT: s_mov_b32 exec_lo, s17
; GFX1032-NEXT: s_addk_i32 s32, 0x200
-; GFX1032-NEXT: v_writelane_b32 v40, s16, 2
+; GFX1032-NEXT: v_writelane_b32 v41, s16, 0
; GFX1032-NEXT: s_getpc_b64 s[16:17]
; GFX1032-NEXT: s_add_u32 s16, s16, external_void_func_void@gotpcrel32@lo+4
; GFX1032-NEXT: s_addc_u32 s17, s17, external_void_func_void@gotpcrel32@hi+12
-; GFX1032-NEXT: s_load_dwordx2 s[16:17], s[16:17], 0x0
; GFX1032-NEXT: v_writelane_b32 v40, s30, 0
+; GFX1032-NEXT: s_load_dwordx2 s[16:17], s[16:17], 0x0
; GFX1032-NEXT: v_writelane_b32 v40, s31, 1
; GFX1032-NEXT: s_waitcnt lgkmcnt(0)
; GFX1032-NEXT: s_swappc_b64 s[30:31], s[16:17]
; GFX1032-NEXT: v_readlane_b32 s31, v40, 1
; GFX1032-NEXT: v_readlane_b32 s30, v40, 0
-; GFX1032-NEXT: v_readlane_b32 s4, v40, 2
+; GFX1032-NEXT: v_readlane_b32 s4, v41, 0
; GFX1032-NEXT: s_or_saveexec_b32 s5, -1
-; GFX1032-NEXT: buffer_load_dword v40, off, s[0:3], s33 ; 4-byte Folded Reload
+; GFX1032-NEXT: s_clause 0x1
+; GFX1032-NEXT: buffer_load_dword v40, off, s[0:3], s33
+; GFX1032-NEXT: buffer_load_dword v41, off, s[0:3], s33 offset:4
; GFX1032-NEXT: s_waitcnt_depctr 0xffe3
; GFX1032-NEXT: s_mov_b32 exec_lo, s5
; GFX1032-NEXT: s_addk_i32 s32, 0xfe00
@@ -2887,23 +2890,26 @@
; GFX1064-NEXT: s_mov_b32 s33, s32
; GFX1064-NEXT: s_or_saveexec_b64 s[18:19], -1
; GFX1064-NEXT: buffer_store_dword v40, off, s[0:3], s33 ; 4-byte Folded Spill
+; GFX1064-NEXT: buffer_store_dword v41, off, s[0:3], s33 offset:4 ; 4-byte Folded Spill
; GFX1064-NEXT: s_waitcnt_depctr 0xffe3
; GFX1064-NEXT: s_mov_b64 exec, s[18:19]
; GFX1064-NEXT: s_addk_i32 s32, 0x400
-; GFX1064-NEXT: v_writelane_b32 v40, s16, 2
+; GFX1064-NEXT: v_writelane_b32 v41, s16, 0
; GFX1064-NEXT: s_getpc_b64 s[16:17]
; GFX1064-NEXT: s_add_u32 s16, s16, external_void_func_void@gotpcrel32@lo+4
; GFX1064-NEXT: s_addc_u32 s17, s17, external_void_func_void@gotpcrel32@hi+12
-; GFX1064-NEXT: s_load_dwordx2 s[16:17], s[16:17], 0x0
; GFX1064-NEXT: v_writelane_b32 v40, s30, 0
+; GFX1064-NEXT: s_load_dwordx2 s[16:17], s[16:17], 0x0
; GFX1064-NEXT: v_writelane_b32 v40, s31, 1
; GFX1064-NEXT: s_waitcnt lgkmcnt(0)
; GFX1064-NEXT: s_swappc_b64 s[30:31], s[16:17]
; GFX1064-NEXT: v_readlane_b32 s31, v40, 1
; GFX1064-NEXT: v_readlane_b32 s30, v40, 0
-; GFX1064-NEXT: v_readlane_b32 s4, v40, 2
+; GFX1064-NEXT: v_readlane_b32 s4, v41, 0
; GFX1064-NEXT: s_or_saveexec_b64 s[6:7], -1
-; GFX1064-NEXT: buffer_load_dword v40, off, s[0:3], s33 ; 4-byte Folded Reload
+; GFX1064-NEXT: s_clause 0x1
+; GFX1064-NEXT: buffer_load_dword v40, off, s[0:3], s33
+; GFX1064-NEXT: buffer_load_dword v41, off, s[0:3], s33 offset:4
; GFX1064-NEXT: s_waitcnt_depctr 0xffe3
; GFX1064-NEXT: s_mov_b64 exec, s[6:7]
; GFX1064-NEXT: s_addk_i32 s32, 0xfc00
diff --git a/llvm/test/CodeGen/AMDGPU/whole-wave-register-copy.ll b/llvm/test/CodeGen/AMDGPU/whole-wave-register-copy.ll
deleted file mode 100644
--- a/llvm/test/CodeGen/AMDGPU/whole-wave-register-copy.ll
+++ /dev/null
@@ -1,68 +0,0 @@
-; NOTE: Assertions have been autogenerated by utils/update_llc_test_checks.py
-; RUN: llc -mtriple=amdgcn-amd-amdhsa -mcpu=gfx90a -verify-machineinstrs < %s | FileCheck -check-prefix=GFX90A %s
-
-; The test forces a high vector register pressure and there won't be sufficient VGPRs to be allocated
-; for writelane/readlane SGPR spill instructions. Regalloc would split the vector register liverange
-; by introducing a copy to AGPR register. The VGPR store to AGPR (v_accvgpr_write_b32) and later the
-; restore from AGPR (v_accvgpr_read_b32) should be whole-wave operations and hence exec mask should be
-; manipulated to ensure all lanes are active when these instructions are executed.
-define void @vector_reg_liverange_split() #0 {
-; GFX90A-LABEL: vector_reg_liverange_split:
-; GFX90A: ; %bb.0:
-; GFX90A-NEXT: s_waitcnt vmcnt(0) expcnt(0) lgkmcnt(0)
-; GFX90A-NEXT: s_mov_b32 s16, s33
-; GFX90A-NEXT: s_mov_b32 s33, s32
-; GFX90A-NEXT: s_xor_saveexec_b64 s[18:19], -1
-; GFX90A-NEXT: buffer_store_dword v0, off, s[0:3], s33 offset:8 ; 4-byte Folded Spill
-; GFX90A-NEXT: s_mov_b64 exec, -1
-; GFX90A-NEXT: buffer_store_dword v40, off, s[0:3], s33 ; 4-byte Folded Spill
-; GFX90A-NEXT: buffer_store_dword a32, off, s[0:3], s33 offset:4 ; 4-byte Folded Spill
-; GFX90A-NEXT: s_mov_b64 exec, s[18:19]
-; GFX90A-NEXT: v_writelane_b32 v40, s16, 2
-; GFX90A-NEXT: ; implicit-def: $vgpr0
-; GFX90A-NEXT: v_writelane_b32 v40, s30, 0
-; GFX90A-NEXT: s_addk_i32 s32, 0x400
-; GFX90A-NEXT: v_writelane_b32 v40, s31, 1
-; GFX90A-NEXT: ;;#ASMSTART
-; GFX90A-NEXT: ; def s20
-; GFX90A-NEXT: ;;#ASMEND
-; GFX90A-NEXT: v_writelane_b32 v0, s20, 0
-; GFX90A-NEXT: s_or_saveexec_b64 s[28:29], -1
-; GFX90A-NEXT: v_accvgpr_write_b32 a32, v0
-; GFX90A-NEXT: s_mov_b64 exec, s[28:29]
-; GFX90A-NEXT: s_getpc_b64 s[16:17]
-; GFX90A-NEXT: s_add_u32 s16, s16, foo@gotpcrel32@lo+4
-; GFX90A-NEXT: s_addc_u32 s17, s17, foo@gotpcrel32@hi+12
-; GFX90A-NEXT: s_load_dwordx2 s[16:17], s[16:17], 0x0
-; GFX90A-NEXT: s_waitcnt lgkmcnt(0)
-; GFX90A-NEXT: s_swappc_b64 s[30:31], s[16:17]
-; GFX90A-NEXT: s_or_saveexec_b64 s[28:29], -1
-; GFX90A-NEXT: v_accvgpr_read_b32 v0, a32
-; GFX90A-NEXT: s_mov_b64 exec, s[28:29]
-; GFX90A-NEXT: v_readlane_b32 s20, v0, 0
-; GFX90A-NEXT: ;;#ASMSTART
-; GFX90A-NEXT: ; use s20
-; GFX90A-NEXT: ;;#ASMEND
-; GFX90A-NEXT: v_readlane_b32 s31, v40, 1
-; GFX90A-NEXT: v_readlane_b32 s30, v40, 0
-; GFX90A-NEXT: ; kill: killed $vgpr0
-; GFX90A-NEXT: v_readlane_b32 s4, v40, 2
-; GFX90A-NEXT: s_xor_saveexec_b64 s[6:7], -1
-; GFX90A-NEXT: buffer_load_dword v0, off, s[0:3], s33 offset:8 ; 4-byte Folded Reload
-; GFX90A-NEXT: s_mov_b64 exec, -1
-; GFX90A-NEXT: buffer_load_dword v40, off, s[0:3], s33 ; 4-byte Folded Reload
-; GFX90A-NEXT: buffer_load_dword a32, off, s[0:3], s33 offset:4 ; 4-byte Folded Reload
-; GFX90A-NEXT: s_mov_b64 exec, s[6:7]
-; GFX90A-NEXT: s_addk_i32 s32, 0xfc00
-; GFX90A-NEXT: s_mov_b32 s33, s4
-; GFX90A-NEXT: s_waitcnt vmcnt(0)
-; GFX90A-NEXT: s_setpc_b64 s[30:31]
- %s20 = call i32 asm sideeffect "; def $0","=${s20}"()
- call void @foo()
- call void asm sideeffect "; use $0","${s20}"(i32 %s20)
- ret void
-}
-
-declare void @foo()
-
-attributes #0 = { "amdgpu-num-vgpr"="41" "amdgpu-num-sgpr"="34"}
diff --git a/llvm/test/CodeGen/AMDGPU/whole-wave-register-spill.ll b/llvm/test/CodeGen/AMDGPU/whole-wave-register-spill.ll
deleted file mode 100644
--- a/llvm/test/CodeGen/AMDGPU/whole-wave-register-spill.ll
+++ /dev/null
@@ -1,140 +0,0 @@
-; NOTE: Assertions have been autogenerated by utils/update_llc_test_checks.py
-; RUN: llc -mtriple=amdgcn-amd-amdhsa -mcpu=gfx906 --verify-machineinstrs -o - %s | FileCheck -check-prefix=GCN %s
-; RUN: llc -mtriple=amdgcn-amd-amdhsa -mcpu=gfx906 -O0 --verify-machineinstrs -o - %s | FileCheck -check-prefix=GCN-O0 %s
-
-; Test whole-wave register spilling.
-
-; In this testcase, the return address registers, PC value (SGPR30_SGPR31) and the scratch SGPR used in
-; the inline asm statements should be preserved across the call. Since the test limits the VGPR numbers,
-; the PC will be spilled to the only available CSR VGPR (VGPR40) as we spill CSR SGPRs including the PC
-; directly to the physical VGPR lane to correctly generate the CFIs. The SGPR20 will get spilled to the
-; virtual VGPR lane and that would be allocated by regalloc. Since there is no free VGPR to allocate, RA
-; must spill a scratch VGPR. The writelane/readlane instructions that spill/restore SGPRs into/from VGPR
-; are whole-wave operations and hence the VGPRs involved in such operations require whole-wave spilling.
-
-define void @test() #0 {
-; GCN-LABEL: test:
-; GCN: ; %bb.0:
-; GCN-NEXT: s_waitcnt vmcnt(0) expcnt(0) lgkmcnt(0)
-; GCN-NEXT: s_mov_b32 s16, s33
-; GCN-NEXT: s_mov_b32 s33, s32
-; GCN-NEXT: s_xor_saveexec_b64 s[18:19], -1
-; GCN-NEXT: buffer_store_dword v0, off, s[0:3], s33 offset:8 ; 4-byte Folded Spill
-; GCN-NEXT: buffer_store_dword v1, off, s[0:3], s33 offset:12 ; 4-byte Folded Spill
-; GCN-NEXT: s_mov_b64 exec, -1
-; GCN-NEXT: buffer_store_dword v40, off, s[0:3], s33 ; 4-byte Folded Spill
-; GCN-NEXT: s_mov_b64 exec, s[18:19]
-; GCN-NEXT: v_writelane_b32 v40, s28, 2
-; GCN-NEXT: v_writelane_b32 v40, s29, 3
-; GCN-NEXT: v_writelane_b32 v40, s16, 4
-; GCN-NEXT: ; implicit-def: $vgpr0
-; GCN-NEXT: v_writelane_b32 v40, s30, 0
-; GCN-NEXT: s_addk_i32 s32, 0x800
-; GCN-NEXT: v_writelane_b32 v40, s31, 1
-; GCN-NEXT: ;;#ASMSTART
-; GCN-NEXT: ; def s16
-; GCN-NEXT: ;;#ASMEND
-; GCN-NEXT: v_writelane_b32 v0, s16, 0
-; GCN-NEXT: s_or_saveexec_b64 s[28:29], -1
-; GCN-NEXT: buffer_store_dword v0, off, s[0:3], s33 offset:4 ; 4-byte Folded Spill
-; GCN-NEXT: s_mov_b64 exec, s[28:29]
-; GCN-NEXT: s_getpc_b64 s[16:17]
-; GCN-NEXT: s_add_u32 s16, s16, ext_func@gotpcrel32@lo+4
-; GCN-NEXT: s_addc_u32 s17, s17, ext_func@gotpcrel32@hi+12
-; GCN-NEXT: s_load_dwordx2 s[16:17], s[16:17], 0x0
-; GCN-NEXT: s_waitcnt lgkmcnt(0)
-; GCN-NEXT: s_swappc_b64 s[30:31], s[16:17]
-; GCN-NEXT: s_or_saveexec_b64 s[28:29], -1
-; GCN-NEXT: buffer_load_dword v1, off, s[0:3], s33 offset:4 ; 4-byte Folded Reload
-; GCN-NEXT: s_mov_b64 exec, s[28:29]
-; GCN-NEXT: s_waitcnt vmcnt(0)
-; GCN-NEXT: v_readlane_b32 s4, v1, 0
-; GCN-NEXT: v_mov_b32_e32 v0, s4
-; GCN-NEXT: global_store_dword v[0:1], v0, off
-; GCN-NEXT: s_waitcnt vmcnt(0)
-; GCN-NEXT: v_readlane_b32 s31, v40, 1
-; GCN-NEXT: v_readlane_b32 s30, v40, 0
-; GCN-NEXT: ; kill: killed $vgpr1
-; GCN-NEXT: v_readlane_b32 s28, v40, 2
-; GCN-NEXT: v_readlane_b32 s29, v40, 3
-; GCN-NEXT: v_readlane_b32 s4, v40, 4
-; GCN-NEXT: s_xor_saveexec_b64 s[6:7], -1
-; GCN-NEXT: buffer_load_dword v0, off, s[0:3], s33 offset:8 ; 4-byte Folded Reload
-; GCN-NEXT: buffer_load_dword v1, off, s[0:3], s33 offset:12 ; 4-byte Folded Reload
-; GCN-NEXT: s_mov_b64 exec, -1
-; GCN-NEXT: buffer_load_dword v40, off, s[0:3], s33 ; 4-byte Folded Reload
-; GCN-NEXT: s_mov_b64 exec, s[6:7]
-; GCN-NEXT: s_addk_i32 s32, 0xf800
-; GCN-NEXT: s_mov_b32 s33, s4
-; GCN-NEXT: s_waitcnt vmcnt(0)
-; GCN-NEXT: s_setpc_b64 s[30:31]
-;
-; GCN-O0-LABEL: test:
-; GCN-O0: ; %bb.0:
-; GCN-O0-NEXT: s_waitcnt vmcnt(0) expcnt(0) lgkmcnt(0)
-; GCN-O0-NEXT: s_mov_b32 s16, s33
-; GCN-O0-NEXT: s_mov_b32 s33, s32
-; GCN-O0-NEXT: s_xor_saveexec_b64 s[18:19], -1
-; GCN-O0-NEXT: buffer_store_dword v0, off, s[0:3], s33 offset:8 ; 4-byte Folded Spill
-; GCN-O0-NEXT: s_mov_b64 exec, -1
-; GCN-O0-NEXT: buffer_store_dword v40, off, s[0:3], s33 ; 4-byte Folded Spill
-; GCN-O0-NEXT: s_mov_b64 exec, s[18:19]
-; GCN-O0-NEXT: v_writelane_b32 v40, s28, 2
-; GCN-O0-NEXT: v_writelane_b32 v40, s29, 3
-; GCN-O0-NEXT: v_writelane_b32 v40, s16, 4
-; GCN-O0-NEXT: s_add_i32 s32, s32, 0x400
-; GCN-O0-NEXT: ; implicit-def: $vgpr0
-; GCN-O0-NEXT: v_writelane_b32 v40, s30, 0
-; GCN-O0-NEXT: v_writelane_b32 v40, s31, 1
-; GCN-O0-NEXT: ;;#ASMSTART
-; GCN-O0-NEXT: ; def s16
-; GCN-O0-NEXT: ;;#ASMEND
-; GCN-O0-NEXT: v_writelane_b32 v0, s16, 0
-; GCN-O0-NEXT: s_or_saveexec_b64 s[28:29], -1
-; GCN-O0-NEXT: buffer_store_dword v0, off, s[0:3], s33 offset:4 ; 4-byte Folded Spill
-; GCN-O0-NEXT: s_mov_b64 exec, s[28:29]
-; GCN-O0-NEXT: s_getpc_b64 s[16:17]
-; GCN-O0-NEXT: s_add_u32 s16, s16, ext_func@gotpcrel32@lo+4
-; GCN-O0-NEXT: s_addc_u32 s17, s17, ext_func@gotpcrel32@hi+12
-; GCN-O0-NEXT: s_load_dwordx2 s[16:17], s[16:17], 0x0
-; GCN-O0-NEXT: s_mov_b64 s[22:23], s[2:3]
-; GCN-O0-NEXT: s_mov_b64 s[20:21], s[0:1]
-; GCN-O0-NEXT: s_mov_b64 s[0:1], s[20:21]
-; GCN-O0-NEXT: s_mov_b64 s[2:3], s[22:23]
-; GCN-O0-NEXT: s_waitcnt lgkmcnt(0)
-; GCN-O0-NEXT: s_swappc_b64 s[30:31], s[16:17]
-; GCN-O0-NEXT: s_or_saveexec_b64 s[28:29], -1
-; GCN-O0-NEXT: buffer_load_dword v0, off, s[0:3], s33 offset:4 ; 4-byte Folded Reload
-; GCN-O0-NEXT: s_mov_b64 exec, s[28:29]
-; GCN-O0-NEXT: s_waitcnt vmcnt(0)
-; GCN-O0-NEXT: v_readlane_b32 s4, v0, 0
-; GCN-O0-NEXT: ; implicit-def: $sgpr6_sgpr7
-; GCN-O0-NEXT: v_mov_b32_e32 v1, s6
-; GCN-O0-NEXT: v_mov_b32_e32 v2, s7
-; GCN-O0-NEXT: v_mov_b32_e32 v3, s4
-; GCN-O0-NEXT: global_store_dword v[1:2], v3, off
-; GCN-O0-NEXT: s_waitcnt vmcnt(0)
-; GCN-O0-NEXT: v_readlane_b32 s31, v40, 1
-; GCN-O0-NEXT: v_readlane_b32 s30, v40, 0
-; GCN-O0-NEXT: ; kill: killed $vgpr0
-; GCN-O0-NEXT: v_readlane_b32 s28, v40, 2
-; GCN-O0-NEXT: v_readlane_b32 s29, v40, 3
-; GCN-O0-NEXT: v_readlane_b32 s4, v40, 4
-; GCN-O0-NEXT: s_xor_saveexec_b64 s[6:7], -1
-; GCN-O0-NEXT: buffer_load_dword v0, off, s[0:3], s33 offset:8 ; 4-byte Folded Reload
-; GCN-O0-NEXT: s_mov_b64 exec, -1
-; GCN-O0-NEXT: buffer_load_dword v40, off, s[0:3], s33 ; 4-byte Folded Reload
-; GCN-O0-NEXT: s_mov_b64 exec, s[6:7]
-; GCN-O0-NEXT: s_add_i32 s32, s32, 0xfffffc00
-; GCN-O0-NEXT: s_mov_b32 s33, s4
-; GCN-O0-NEXT: s_waitcnt vmcnt(0)
-; GCN-O0-NEXT: s_setpc_b64 s[30:31]
- %sgpr = call i32 asm sideeffect "; def $0", "=s" () #0
- call void @ext_func()
- store volatile i32 %sgpr, ptr addrspace(1) undef
- ret void
-}
-
-declare void @ext_func();
-
-attributes #0 = { nounwind "amdgpu-num-vgpr"="41" "amdgpu-num-sgpr"="34"}
diff --git a/llvm/test/CodeGen/AMDGPU/wwm-reserved-spill.ll b/llvm/test/CodeGen/AMDGPU/wwm-reserved-spill.ll
--- a/llvm/test/CodeGen/AMDGPU/wwm-reserved-spill.ll
+++ b/llvm/test/CodeGen/AMDGPU/wwm-reserved-spill.ll
@@ -141,15 +141,10 @@
; GFX9-O0: ; %bb.0: ; %entry
; GFX9-O0-NEXT: s_waitcnt vmcnt(0) expcnt(0) lgkmcnt(0)
; GFX9-O0-NEXT: s_xor_saveexec_b64 s[34:35], -1
-; GFX9-O0-NEXT: buffer_store_dword v0, off, s[0:3], s32 offset:20 ; 4-byte Folded Spill
-; GFX9-O0-NEXT: buffer_store_dword v1, off, s[0:3], s32 offset:24 ; 4-byte Folded Spill
-; GFX9-O0-NEXT: buffer_store_dword v2, off, s[0:3], s32 offset:28 ; 4-byte Folded Spill
+; GFX9-O0-NEXT: buffer_store_dword v3, off, s[0:3], s32 offset:16 ; 4-byte Folded Spill
+; GFX9-O0-NEXT: buffer_store_dword v1, off, s[0:3], s32 offset:20 ; 4-byte Folded Spill
+; GFX9-O0-NEXT: buffer_store_dword v2, off, s[0:3], s32 offset:24 ; 4-byte Folded Spill
; GFX9-O0-NEXT: s_mov_b64 exec, s[34:35]
-; GFX9-O0-NEXT: ; implicit-def: $vgpr3
-; GFX9-O0-NEXT: v_mov_b32_e32 v3, v0
-; GFX9-O0-NEXT: s_or_saveexec_b64 s[46:47], -1
-; GFX9-O0-NEXT: buffer_load_dword v0, off, s[0:3], s32 ; 4-byte Folded Reload
-; GFX9-O0-NEXT: s_mov_b64 exec, s[46:47]
; GFX9-O0-NEXT: s_mov_b32 s40, s6
; GFX9-O0-NEXT: s_mov_b32 s34, s4
; GFX9-O0-NEXT: ; kill: def $sgpr40 killed $sgpr40 def $sgpr40_sgpr41
@@ -164,18 +159,17 @@
; GFX9-O0-NEXT: s_mov_b32 s37, s44
; GFX9-O0-NEXT: s_mov_b32 s38, s43
; GFX9-O0-NEXT: s_mov_b32 s39, s42
-; GFX9-O0-NEXT: s_waitcnt vmcnt(0)
-; GFX9-O0-NEXT: v_writelane_b32 v0, s40, 0
-; GFX9-O0-NEXT: v_writelane_b32 v0, s41, 1
-; GFX9-O0-NEXT: v_writelane_b32 v0, s34, 2
-; GFX9-O0-NEXT: v_writelane_b32 v0, s35, 3
+; GFX9-O0-NEXT: v_writelane_b32 v3, s40, 0
+; GFX9-O0-NEXT: v_writelane_b32 v3, s41, 1
+; GFX9-O0-NEXT: v_writelane_b32 v3, s34, 2
+; GFX9-O0-NEXT: v_writelane_b32 v3, s35, 3
; GFX9-O0-NEXT: s_mov_b32 s34, 0
; GFX9-O0-NEXT: s_nop 2
; GFX9-O0-NEXT: buffer_load_dwordx2 v[4:5], off, s[36:39], s34
; GFX9-O0-NEXT: s_waitcnt vmcnt(0)
-; GFX9-O0-NEXT: buffer_store_dword v4, off, s[0:3], s32 offset:12 ; 4-byte Folded Spill
+; GFX9-O0-NEXT: buffer_store_dword v4, off, s[0:3], s32 offset:8 ; 4-byte Folded Spill
; GFX9-O0-NEXT: s_waitcnt vmcnt(0)
-; GFX9-O0-NEXT: buffer_store_dword v5, off, s[0:3], s32 offset:16 ; 4-byte Folded Spill
+; GFX9-O0-NEXT: buffer_store_dword v5, off, s[0:3], s32 offset:12 ; 4-byte Folded Spill
; GFX9-O0-NEXT: ; implicit-def: $sgpr36_sgpr37
; GFX9-O0-NEXT: v_mov_b32_e32 v1, v4
; GFX9-O0-NEXT: s_not_b64 exec, exec
@@ -188,24 +182,21 @@
; GFX9-O0-NEXT: v_add_u32_e64 v1, v1, v2
; GFX9-O0-NEXT: s_mov_b64 exec, s[36:37]
; GFX9-O0-NEXT: v_mov_b32_e32 v4, v1
-; GFX9-O0-NEXT: buffer_store_dword v4, off, s[0:3], s32 offset:8 ; 4-byte Folded Spill
-; GFX9-O0-NEXT: v_cmp_eq_u32_e64 s[36:37], v3, s34
-; GFX9-O0-NEXT: v_mov_b32_e32 v3, s34
-; GFX9-O0-NEXT: buffer_store_dword v3, off, s[0:3], s32 offset:4 ; 4-byte Folded Spill
-; GFX9-O0-NEXT: s_mov_b64 s[34:35], exec
-; GFX9-O0-NEXT: v_writelane_b32 v0, s34, 4
-; GFX9-O0-NEXT: v_writelane_b32 v0, s35, 5
-; GFX9-O0-NEXT: s_or_saveexec_b64 s[46:47], -1
+; GFX9-O0-NEXT: buffer_store_dword v4, off, s[0:3], s32 offset:4 ; 4-byte Folded Spill
+; GFX9-O0-NEXT: v_cmp_eq_u32_e64 s[36:37], v0, s34
+; GFX9-O0-NEXT: v_mov_b32_e32 v0, s34
; GFX9-O0-NEXT: buffer_store_dword v0, off, s[0:3], s32 ; 4-byte Folded Spill
-; GFX9-O0-NEXT: s_mov_b64 exec, s[46:47]
+; GFX9-O0-NEXT: s_mov_b64 s[34:35], exec
+; GFX9-O0-NEXT: v_writelane_b32 v3, s34, 4
+; GFX9-O0-NEXT: v_writelane_b32 v3, s35, 5
; GFX9-O0-NEXT: s_and_b64 s[34:35], s[34:35], s[36:37]
; GFX9-O0-NEXT: s_mov_b64 exec, s[34:35]
; GFX9-O0-NEXT: s_cbranch_execz .LBB1_2
; GFX9-O0-NEXT: ; %bb.1: ; %if
-; GFX9-O0-NEXT: buffer_load_dword v3, off, s[0:3], s32 offset:12 ; 4-byte Folded Reload
-; GFX9-O0-NEXT: buffer_load_dword v4, off, s[0:3], s32 offset:16 ; 4-byte Folded Reload
+; GFX9-O0-NEXT: buffer_load_dword v4, off, s[0:3], s32 offset:8 ; 4-byte Folded Reload
+; GFX9-O0-NEXT: buffer_load_dword v5, off, s[0:3], s32 offset:12 ; 4-byte Folded Reload
; GFX9-O0-NEXT: s_waitcnt vmcnt(0)
-; GFX9-O0-NEXT: v_mov_b32_e32 v0, v4
+; GFX9-O0-NEXT: v_mov_b32_e32 v0, v5
; GFX9-O0-NEXT: s_or_saveexec_b64 s[34:35], -1
; GFX9-O0-NEXT: v_mov_b32_e32 v1, 0
; GFX9-O0-NEXT: s_mov_b64 exec, s[34:35]
@@ -218,28 +209,25 @@
; GFX9-O0-NEXT: v_add_u32_e64 v1, v2, v1
; GFX9-O0-NEXT: s_mov_b64 exec, s[34:35]
; GFX9-O0-NEXT: v_mov_b32_e32 v0, v1
-; GFX9-O0-NEXT: buffer_store_dword v0, off, s[0:3], s32 offset:4 ; 4-byte Folded Spill
+; GFX9-O0-NEXT: buffer_store_dword v0, off, s[0:3], s32 ; 4-byte Folded Spill
; GFX9-O0-NEXT: .LBB1_2: ; %merge
-; GFX9-O0-NEXT: s_or_saveexec_b64 s[46:47], -1
-; GFX9-O0-NEXT: buffer_load_dword v0, off, s[0:3], s32 ; 4-byte Folded Reload
-; GFX9-O0-NEXT: s_mov_b64 exec, s[46:47]
-; GFX9-O0-NEXT: buffer_load_dword v3, off, s[0:3], s32 offset:8 ; 4-byte Folded Reload
-; GFX9-O0-NEXT: buffer_load_dword v4, off, s[0:3], s32 offset:4 ; 4-byte Folded Reload
-; GFX9-O0-NEXT: s_waitcnt vmcnt(2)
-; GFX9-O0-NEXT: v_readlane_b32 s36, v0, 4
-; GFX9-O0-NEXT: v_readlane_b32 s37, v0, 5
+; GFX9-O0-NEXT: buffer_load_dword v0, off, s[0:3], s32 offset:4 ; 4-byte Folded Reload
+; GFX9-O0-NEXT: s_nop 0
+; GFX9-O0-NEXT: buffer_load_dword v4, off, s[0:3], s32 ; 4-byte Folded Reload
+; GFX9-O0-NEXT: v_readlane_b32 s36, v3, 4
+; GFX9-O0-NEXT: v_readlane_b32 s37, v3, 5
; GFX9-O0-NEXT: s_or_b64 exec, exec, s[36:37]
-; GFX9-O0-NEXT: v_readlane_b32 s38, v0, 0
-; GFX9-O0-NEXT: v_readlane_b32 s39, v0, 1
-; GFX9-O0-NEXT: v_readlane_b32 s34, v0, 2
-; GFX9-O0-NEXT: v_readlane_b32 s35, v0, 3
+; GFX9-O0-NEXT: v_readlane_b32 s38, v3, 0
+; GFX9-O0-NEXT: v_readlane_b32 s39, v3, 1
+; GFX9-O0-NEXT: v_readlane_b32 s34, v3, 2
+; GFX9-O0-NEXT: v_readlane_b32 s35, v3, 3
; GFX9-O0-NEXT: s_waitcnt vmcnt(0)
-; GFX9-O0-NEXT: v_cmp_eq_u32_e64 s[36:37], v3, v4
-; GFX9-O0-NEXT: v_cndmask_b32_e64 v3, 0, 1, s[36:37]
+; GFX9-O0-NEXT: v_cmp_eq_u32_e64 s[36:37], v0, v4
+; GFX9-O0-NEXT: v_cndmask_b32_e64 v0, 0, 1, s[36:37]
; GFX9-O0-NEXT: s_mov_b32 s36, 1
-; GFX9-O0-NEXT: v_lshlrev_b32_e64 v3, s36, v3
+; GFX9-O0-NEXT: v_lshlrev_b32_e64 v0, s36, v0
; GFX9-O0-NEXT: s_mov_b32 s36, 2
-; GFX9-O0-NEXT: v_and_b32_e64 v3, v3, s36
+; GFX9-O0-NEXT: v_and_b32_e64 v0, v0, s36
; GFX9-O0-NEXT: s_mov_b32 s40, s35
; GFX9-O0-NEXT: s_mov_b32 s36, s34
; GFX9-O0-NEXT: s_mov_b32 s34, s39
@@ -249,12 +237,11 @@
; GFX9-O0-NEXT: s_mov_b32 s38, s35
; GFX9-O0-NEXT: s_mov_b32 s39, s34
; GFX9-O0-NEXT: s_mov_b32 s34, 0
-; GFX9-O0-NEXT: buffer_store_dword v3, off, s[36:39], s34 offset:4
-; GFX9-O0-NEXT: ; kill: killed $vgpr0
+; GFX9-O0-NEXT: buffer_store_dword v0, off, s[36:39], s34 offset:4
; GFX9-O0-NEXT: s_xor_saveexec_b64 s[34:35], -1
-; GFX9-O0-NEXT: buffer_load_dword v0, off, s[0:3], s32 offset:20 ; 4-byte Folded Reload
-; GFX9-O0-NEXT: buffer_load_dword v1, off, s[0:3], s32 offset:24 ; 4-byte Folded Reload
-; GFX9-O0-NEXT: buffer_load_dword v2, off, s[0:3], s32 offset:28 ; 4-byte Folded Reload
+; GFX9-O0-NEXT: buffer_load_dword v3, off, s[0:3], s32 offset:16 ; 4-byte Folded Reload
+; GFX9-O0-NEXT: buffer_load_dword v1, off, s[0:3], s32 offset:20 ; 4-byte Folded Reload
+; GFX9-O0-NEXT: buffer_load_dword v2, off, s[0:3], s32 offset:24 ; 4-byte Folded Reload
; GFX9-O0-NEXT: s_mov_b64 exec, s[34:35]
; GFX9-O0-NEXT: s_waitcnt vmcnt(0)
; GFX9-O0-NEXT: s_setpc_b64 s[30:31]
@@ -559,28 +546,25 @@
; GFX9-O0-LABEL: strict_wwm_call_i64:
; GFX9-O0: ; %bb.0:
; GFX9-O0-NEXT: s_waitcnt vmcnt(0) expcnt(0) lgkmcnt(0)
-; GFX9-O0-NEXT: s_mov_b32 s48, s33
+; GFX9-O0-NEXT: s_mov_b32 s46, s33
; GFX9-O0-NEXT: s_mov_b32 s33, s32
; GFX9-O0-NEXT: s_xor_saveexec_b64 s[34:35], -1
; GFX9-O0-NEXT: buffer_store_dword v10, off, s[0:3], s33 ; 4-byte Folded Spill
-; GFX9-O0-NEXT: buffer_store_dword v0, off, s[0:3], s33 offset:8 ; 4-byte Folded Spill
-; GFX9-O0-NEXT: buffer_store_dword v6, off, s[0:3], s33 offset:12 ; 4-byte Folded Spill
-; GFX9-O0-NEXT: buffer_store_dword v8, off, s[0:3], s33 offset:16 ; 4-byte Folded Spill
+; GFX9-O0-NEXT: buffer_store_dword v8, off, s[0:3], s33 offset:4 ; 4-byte Folded Spill
; GFX9-O0-NEXT: s_waitcnt vmcnt(0)
-; GFX9-O0-NEXT: buffer_store_dword v9, off, s[0:3], s33 offset:20 ; 4-byte Folded Spill
-; GFX9-O0-NEXT: buffer_store_dword v2, off, s[0:3], s33 offset:24 ; 4-byte Folded Spill
-; GFX9-O0-NEXT: buffer_store_dword v3, off, s[0:3], s33 offset:28 ; 4-byte Folded Spill
+; GFX9-O0-NEXT: buffer_store_dword v9, off, s[0:3], s33 offset:8 ; 4-byte Folded Spill
+; GFX9-O0-NEXT: buffer_store_dword v2, off, s[0:3], s33 offset:12 ; 4-byte Folded Spill
+; GFX9-O0-NEXT: buffer_store_dword v3, off, s[0:3], s33 offset:16 ; 4-byte Folded Spill
; GFX9-O0-NEXT: s_waitcnt vmcnt(0)
-; GFX9-O0-NEXT: buffer_store_dword v4, off, s[0:3], s33 offset:32 ; 4-byte Folded Spill
-; GFX9-O0-NEXT: buffer_store_dword v3, off, s[0:3], s33 offset:36 ; 4-byte Folded Spill
-; GFX9-O0-NEXT: buffer_store_dword v2, off, s[0:3], s33 offset:40 ; 4-byte Folded Spill
+; GFX9-O0-NEXT: buffer_store_dword v4, off, s[0:3], s33 offset:20 ; 4-byte Folded Spill
+; GFX9-O0-NEXT: buffer_store_dword v3, off, s[0:3], s33 offset:24 ; 4-byte Folded Spill
+; GFX9-O0-NEXT: buffer_store_dword v2, off, s[0:3], s33 offset:28 ; 4-byte Folded Spill
; GFX9-O0-NEXT: s_waitcnt vmcnt(0)
-; GFX9-O0-NEXT: buffer_store_dword v3, off, s[0:3], s33 offset:44 ; 4-byte Folded Spill
-; GFX9-O0-NEXT: buffer_store_dword v4, off, s[0:3], s33 offset:48 ; 4-byte Folded Spill
-; GFX9-O0-NEXT: buffer_store_dword v5, off, s[0:3], s33 offset:52 ; 4-byte Folded Spill
+; GFX9-O0-NEXT: buffer_store_dword v3, off, s[0:3], s33 offset:32 ; 4-byte Folded Spill
+; GFX9-O0-NEXT: buffer_store_dword v4, off, s[0:3], s33 offset:36 ; 4-byte Folded Spill
+; GFX9-O0-NEXT: buffer_store_dword v5, off, s[0:3], s33 offset:40 ; 4-byte Folded Spill
; GFX9-O0-NEXT: s_mov_b64 exec, s[34:35]
-; GFX9-O0-NEXT: s_add_i32 s32, s32, 0x1000
-; GFX9-O0-NEXT: ; implicit-def: $vgpr0
+; GFX9-O0-NEXT: s_add_i32 s32, s32, 0xc00
; GFX9-O0-NEXT: v_writelane_b32 v10, s30, 0
; GFX9-O0-NEXT: v_writelane_b32 v10, s31, 1
; GFX9-O0-NEXT: s_mov_b32 s34, s8
@@ -598,10 +582,10 @@
; GFX9-O0-NEXT: s_mov_b32 s41, s45
; GFX9-O0-NEXT: s_mov_b32 s42, s44
; GFX9-O0-NEXT: s_mov_b32 s43, s35
-; GFX9-O0-NEXT: v_writelane_b32 v0, s40, 0
-; GFX9-O0-NEXT: v_writelane_b32 v0, s41, 1
-; GFX9-O0-NEXT: v_writelane_b32 v0, s42, 2
-; GFX9-O0-NEXT: v_writelane_b32 v0, s43, 3
+; GFX9-O0-NEXT: v_writelane_b32 v10, s40, 2
+; GFX9-O0-NEXT: v_writelane_b32 v10, s41, 3
+; GFX9-O0-NEXT: v_writelane_b32 v10, s42, 4
+; GFX9-O0-NEXT: v_writelane_b32 v10, s43, 5
; GFX9-O0-NEXT: ; kill: def $sgpr34 killed $sgpr34 def $sgpr34_sgpr35
; GFX9-O0-NEXT: s_mov_b32 s35, s9
; GFX9-O0-NEXT: ; kill: def $sgpr36_sgpr37 killed $sgpr34_sgpr35
@@ -613,11 +597,8 @@
; GFX9-O0-NEXT: v_mov_b32_e32 v9, s37
; GFX9-O0-NEXT: s_not_b64 exec, exec
; GFX9-O0-NEXT: s_or_saveexec_b64 s[34:35], -1
-; GFX9-O0-NEXT: v_writelane_b32 v0, s34, 4
-; GFX9-O0-NEXT: v_writelane_b32 v0, s35, 5
-; GFX9-O0-NEXT: s_or_saveexec_b64 s[46:47], -1
-; GFX9-O0-NEXT: buffer_store_dword v0, off, s[0:3], s33 offset:4 ; 4-byte Folded Spill
-; GFX9-O0-NEXT: s_mov_b64 exec, s[46:47]
+; GFX9-O0-NEXT: v_writelane_b32 v10, s34, 6
+; GFX9-O0-NEXT: v_writelane_b32 v10, s35, 7
; GFX9-O0-NEXT: v_mov_b32_e32 v2, v8
; GFX9-O0-NEXT: s_mov_b32 s34, 32
; GFX9-O0-NEXT: ; implicit-def: $sgpr36_sgpr37
@@ -634,20 +615,13 @@
; GFX9-O0-NEXT: v_mov_b32_e32 v1, v3
; GFX9-O0-NEXT: s_waitcnt lgkmcnt(0)
; GFX9-O0-NEXT: s_swappc_b64 s[30:31], s[34:35]
-; GFX9-O0-NEXT: s_or_saveexec_b64 s[46:47], -1
-; GFX9-O0-NEXT: buffer_load_dword v6, off, s[0:3], s33 offset:4 ; 4-byte Folded Reload
-; GFX9-O0-NEXT: s_mov_b64 exec, s[46:47]
-; GFX9-O0-NEXT: s_waitcnt vmcnt(0)
-; GFX9-O0-NEXT: v_readlane_b32 s34, v6, 4
-; GFX9-O0-NEXT: v_readlane_b32 s35, v6, 5
-; GFX9-O0-NEXT: v_readlane_b32 s36, v6, 0
-; GFX9-O0-NEXT: v_readlane_b32 s37, v6, 1
-; GFX9-O0-NEXT: v_readlane_b32 s38, v6, 2
-; GFX9-O0-NEXT: v_readlane_b32 s39, v6, 3
+; GFX9-O0-NEXT: v_readlane_b32 s34, v10, 6
+; GFX9-O0-NEXT: v_readlane_b32 s35, v10, 7
+; GFX9-O0-NEXT: v_readlane_b32 s36, v10, 2
+; GFX9-O0-NEXT: v_readlane_b32 s37, v10, 3
+; GFX9-O0-NEXT: v_readlane_b32 s38, v10, 4
+; GFX9-O0-NEXT: v_readlane_b32 s39, v10, 5
; GFX9-O0-NEXT: v_mov_b32_e32 v2, v0
-; GFX9-O0-NEXT: s_or_saveexec_b64 s[46:47], -1
-; GFX9-O0-NEXT: buffer_load_dword v0, off, s[0:3], s33 offset:4 ; 4-byte Folded Reload
-; GFX9-O0-NEXT: s_mov_b64 exec, s[46:47]
; GFX9-O0-NEXT: v_mov_b32_e32 v3, v1
; GFX9-O0-NEXT: ; implicit-def: $sgpr40
; GFX9-O0-NEXT: ; implicit-def: $sgpr40
@@ -656,30 +630,27 @@
; GFX9-O0-NEXT: v_add_co_u32_e64 v2, s[40:41], v2, v4
; GFX9-O0-NEXT: v_addc_co_u32_e64 v3, s[40:41], v3, v5, s[40:41]
; GFX9-O0-NEXT: s_mov_b64 exec, s[34:35]
-; GFX9-O0-NEXT: v_mov_b32_e32 v6, v2
-; GFX9-O0-NEXT: v_mov_b32_e32 v7, v3
+; GFX9-O0-NEXT: v_mov_b32_e32 v0, v2
+; GFX9-O0-NEXT: v_mov_b32_e32 v1, v3
; GFX9-O0-NEXT: s_mov_b32 s34, 0
-; GFX9-O0-NEXT: buffer_store_dwordx2 v[6:7], off, s[36:39], s34 offset:4
+; GFX9-O0-NEXT: buffer_store_dwordx2 v[0:1], off, s[36:39], s34 offset:4
; GFX9-O0-NEXT: v_readlane_b32 s31, v10, 1
; GFX9-O0-NEXT: v_readlane_b32 s30, v10, 0
-; GFX9-O0-NEXT: ; kill: killed $vgpr0
; GFX9-O0-NEXT: s_xor_saveexec_b64 s[34:35], -1
; GFX9-O0-NEXT: buffer_load_dword v10, off, s[0:3], s33 ; 4-byte Folded Reload
-; GFX9-O0-NEXT: buffer_load_dword v0, off, s[0:3], s33 offset:8 ; 4-byte Folded Reload
-; GFX9-O0-NEXT: buffer_load_dword v6, off, s[0:3], s33 offset:12 ; 4-byte Folded Reload
-; GFX9-O0-NEXT: buffer_load_dword v8, off, s[0:3], s33 offset:16 ; 4-byte Folded Reload
-; GFX9-O0-NEXT: buffer_load_dword v9, off, s[0:3], s33 offset:20 ; 4-byte Folded Reload
-; GFX9-O0-NEXT: buffer_load_dword v2, off, s[0:3], s33 offset:24 ; 4-byte Folded Reload
-; GFX9-O0-NEXT: buffer_load_dword v3, off, s[0:3], s33 offset:28 ; 4-byte Folded Reload
-; GFX9-O0-NEXT: buffer_load_dword v4, off, s[0:3], s33 offset:32 ; 4-byte Folded Reload
-; GFX9-O0-NEXT: buffer_load_dword v3, off, s[0:3], s33 offset:36 ; 4-byte Folded Reload
-; GFX9-O0-NEXT: buffer_load_dword v2, off, s[0:3], s33 offset:40 ; 4-byte Folded Reload
-; GFX9-O0-NEXT: buffer_load_dword v3, off, s[0:3], s33 offset:44 ; 4-byte Folded Reload
-; GFX9-O0-NEXT: buffer_load_dword v4, off, s[0:3], s33 offset:48 ; 4-byte Folded Reload
-; GFX9-O0-NEXT: buffer_load_dword v5, off, s[0:3], s33 offset:52 ; 4-byte Folded Reload
+; GFX9-O0-NEXT: buffer_load_dword v8, off, s[0:3], s33 offset:4 ; 4-byte Folded Reload
+; GFX9-O0-NEXT: buffer_load_dword v9, off, s[0:3], s33 offset:8 ; 4-byte Folded Reload
+; GFX9-O0-NEXT: buffer_load_dword v2, off, s[0:3], s33 offset:12 ; 4-byte Folded Reload
+; GFX9-O0-NEXT: buffer_load_dword v3, off, s[0:3], s33 offset:16 ; 4-byte Folded Reload
+; GFX9-O0-NEXT: buffer_load_dword v4, off, s[0:3], s33 offset:20 ; 4-byte Folded Reload
+; GFX9-O0-NEXT: buffer_load_dword v3, off, s[0:3], s33 offset:24 ; 4-byte Folded Reload
+; GFX9-O0-NEXT: buffer_load_dword v2, off, s[0:3], s33 offset:28 ; 4-byte Folded Reload
+; GFX9-O0-NEXT: buffer_load_dword v3, off, s[0:3], s33 offset:32 ; 4-byte Folded Reload
+; GFX9-O0-NEXT: buffer_load_dword v4, off, s[0:3], s33 offset:36 ; 4-byte Folded Reload
+; GFX9-O0-NEXT: buffer_load_dword v5, off, s[0:3], s33 offset:40 ; 4-byte Folded Reload
; GFX9-O0-NEXT: s_mov_b64 exec, s[34:35]
-; GFX9-O0-NEXT: s_add_i32 s32, s32, 0xfffff000
-; GFX9-O0-NEXT: s_mov_b32 s33, s48
+; GFX9-O0-NEXT: s_add_i32 s32, s32, 0xfffff400
+; GFX9-O0-NEXT: s_mov_b32 s33, s46
; GFX9-O0-NEXT: s_waitcnt vmcnt(0)
; GFX9-O0-NEXT: s_setpc_b64 s[30:31]
;
@@ -938,7 +909,7 @@
; GFX9-O0-NEXT: s_waitcnt vmcnt(0)
; GFX9-O0-NEXT: buffer_store_dword v41, off, s[0:3], s32 offset:240 ; 4-byte Folded Spill
; GFX9-O0-NEXT: s_mov_b64 exec, -1
-; GFX9-O0-NEXT: buffer_store_dword v42, off, s[0:3], s32 offset:52 ; 4-byte Folded Spill
+; GFX9-O0-NEXT: buffer_store_dword v42, off, s[0:3], s32 offset:200 ; 4-byte Folded Spill
; GFX9-O0-NEXT: s_mov_b64 exec, s[34:35]
; GFX9-O0-NEXT: buffer_store_dword v40, off, s[0:3], s32 offset:48 ; 4-byte Folded Spill
; GFX9-O0-NEXT: buffer_store_dword v41, off, s[0:3], s32 offset:44 ; 4-byte Folded Spill
@@ -951,35 +922,35 @@
; GFX9-O0-NEXT: v_writelane_b32 v42, s65, 1
; GFX9-O0-NEXT: v_writelane_b32 v42, s66, 2
; GFX9-O0-NEXT: v_writelane_b32 v42, s67, 3
-; GFX9-O0-NEXT: buffer_store_dword v10, off, s[0:3], s32 offset:72 ; 4-byte Folded Spill
-; GFX9-O0-NEXT: buffer_store_dword v9, off, s[0:3], s32 offset:64 ; 4-byte Folded Spill
-; GFX9-O0-NEXT: buffer_store_dword v8, off, s[0:3], s32 offset:80 ; 4-byte Folded Spill
-; GFX9-O0-NEXT: buffer_store_dword v7, off, s[0:3], s32 offset:68 ; 4-byte Folded Spill
-; GFX9-O0-NEXT: buffer_store_dword v6, off, s[0:3], s32 offset:88 ; 4-byte Folded Spill
-; GFX9-O0-NEXT: buffer_store_dword v5, off, s[0:3], s32 offset:76 ; 4-byte Folded Spill
-; GFX9-O0-NEXT: buffer_store_dword v4, off, s[0:3], s32 offset:96 ; 4-byte Folded Spill
-; GFX9-O0-NEXT: buffer_store_dword v3, off, s[0:3], s32 offset:84 ; 4-byte Folded Spill
-; GFX9-O0-NEXT: buffer_store_dword v2, off, s[0:3], s32 offset:56 ; 4-byte Folded Spill
-; GFX9-O0-NEXT: buffer_store_dword v1, off, s[0:3], s32 offset:92 ; 4-byte Folded Spill
-; GFX9-O0-NEXT: buffer_store_dword v0, off, s[0:3], s32 offset:60 ; 4-byte Folded Spill
+; GFX9-O0-NEXT: buffer_store_dword v10, off, s[0:3], s32 offset:68 ; 4-byte Folded Spill
+; GFX9-O0-NEXT: buffer_store_dword v9, off, s[0:3], s32 offset:60 ; 4-byte Folded Spill
+; GFX9-O0-NEXT: buffer_store_dword v8, off, s[0:3], s32 offset:76 ; 4-byte Folded Spill
+; GFX9-O0-NEXT: buffer_store_dword v7, off, s[0:3], s32 offset:64 ; 4-byte Folded Spill
+; GFX9-O0-NEXT: buffer_store_dword v6, off, s[0:3], s32 offset:84 ; 4-byte Folded Spill
+; GFX9-O0-NEXT: buffer_store_dword v5, off, s[0:3], s32 offset:72 ; 4-byte Folded Spill
+; GFX9-O0-NEXT: buffer_store_dword v4, off, s[0:3], s32 offset:92 ; 4-byte Folded Spill
+; GFX9-O0-NEXT: buffer_store_dword v3, off, s[0:3], s32 offset:80 ; 4-byte Folded Spill
+; GFX9-O0-NEXT: buffer_store_dword v2, off, s[0:3], s32 offset:52 ; 4-byte Folded Spill
+; GFX9-O0-NEXT: buffer_store_dword v1, off, s[0:3], s32 offset:88 ; 4-byte Folded Spill
+; GFX9-O0-NEXT: buffer_store_dword v0, off, s[0:3], s32 offset:56 ; 4-byte Folded Spill
; GFX9-O0-NEXT: buffer_load_dword v0, off, s[0:3], s32 offset:20
; GFX9-O0-NEXT: s_waitcnt vmcnt(0)
-; GFX9-O0-NEXT: buffer_store_dword v0, off, s[0:3], s32 offset:120 ; 4-byte Folded Spill
+; GFX9-O0-NEXT: buffer_store_dword v0, off, s[0:3], s32 offset:116 ; 4-byte Folded Spill
; GFX9-O0-NEXT: buffer_load_dword v0, off, s[0:3], s32 offset:16
; GFX9-O0-NEXT: s_waitcnt vmcnt(0)
-; GFX9-O0-NEXT: buffer_store_dword v0, off, s[0:3], s32 offset:116 ; 4-byte Folded Spill
+; GFX9-O0-NEXT: buffer_store_dword v0, off, s[0:3], s32 offset:112 ; 4-byte Folded Spill
; GFX9-O0-NEXT: buffer_load_dword v0, off, s[0:3], s32 offset:12
; GFX9-O0-NEXT: s_waitcnt vmcnt(0)
-; GFX9-O0-NEXT: buffer_store_dword v0, off, s[0:3], s32 offset:112 ; 4-byte Folded Spill
+; GFX9-O0-NEXT: buffer_store_dword v0, off, s[0:3], s32 offset:108 ; 4-byte Folded Spill
; GFX9-O0-NEXT: buffer_load_dword v0, off, s[0:3], s32 offset:8
; GFX9-O0-NEXT: s_waitcnt vmcnt(0)
-; GFX9-O0-NEXT: buffer_store_dword v0, off, s[0:3], s32 offset:108 ; 4-byte Folded Spill
+; GFX9-O0-NEXT: buffer_store_dword v0, off, s[0:3], s32 offset:104 ; 4-byte Folded Spill
; GFX9-O0-NEXT: buffer_load_dword v0, off, s[0:3], s32 offset:4
; GFX9-O0-NEXT: s_waitcnt vmcnt(0)
-; GFX9-O0-NEXT: buffer_store_dword v0, off, s[0:3], s32 offset:104 ; 4-byte Folded Spill
+; GFX9-O0-NEXT: buffer_store_dword v0, off, s[0:3], s32 offset:100 ; 4-byte Folded Spill
; GFX9-O0-NEXT: buffer_load_dword v0, off, s[0:3], s32
; GFX9-O0-NEXT: s_waitcnt vmcnt(0)
-; GFX9-O0-NEXT: buffer_store_dword v0, off, s[0:3], s32 offset:100 ; 4-byte Folded Spill
+; GFX9-O0-NEXT: buffer_store_dword v0, off, s[0:3], s32 offset:96 ; 4-byte Folded Spill
; GFX9-O0-NEXT: ; implicit-def: $sgpr34
; GFX9-O0-NEXT: ; implicit-def: $sgpr34
; GFX9-O0-NEXT: ; implicit-def: $sgpr34
@@ -989,145 +960,145 @@
; GFX9-O0-NEXT: v_mov_b32_e32 v0, s4
; GFX9-O0-NEXT: v_mov_b32_e32 v43, s5
; GFX9-O0-NEXT: v_mov_b32_e32 v1, s6
-; GFX9-O0-NEXT: buffer_store_dword v1, off, s[0:3], s32 offset:200 ; 4-byte Folded Spill
-; GFX9-O0-NEXT: v_mov_b32_e32 v1, s7
; GFX9-O0-NEXT: buffer_store_dword v1, off, s[0:3], s32 offset:196 ; 4-byte Folded Spill
-; GFX9-O0-NEXT: v_mov_b32_e32 v1, s8
+; GFX9-O0-NEXT: v_mov_b32_e32 v1, s7
; GFX9-O0-NEXT: buffer_store_dword v1, off, s[0:3], s32 offset:192 ; 4-byte Folded Spill
-; GFX9-O0-NEXT: v_mov_b32_e32 v1, s9
+; GFX9-O0-NEXT: v_mov_b32_e32 v1, s8
; GFX9-O0-NEXT: buffer_store_dword v1, off, s[0:3], s32 offset:188 ; 4-byte Folded Spill
-; GFX9-O0-NEXT: v_mov_b32_e32 v1, s10
+; GFX9-O0-NEXT: v_mov_b32_e32 v1, s9
; GFX9-O0-NEXT: buffer_store_dword v1, off, s[0:3], s32 offset:184 ; 4-byte Folded Spill
-; GFX9-O0-NEXT: v_mov_b32_e32 v1, s11
+; GFX9-O0-NEXT: v_mov_b32_e32 v1, s10
; GFX9-O0-NEXT: buffer_store_dword v1, off, s[0:3], s32 offset:180 ; 4-byte Folded Spill
-; GFX9-O0-NEXT: v_mov_b32_e32 v1, s12
+; GFX9-O0-NEXT: v_mov_b32_e32 v1, s11
; GFX9-O0-NEXT: buffer_store_dword v1, off, s[0:3], s32 offset:176 ; 4-byte Folded Spill
-; GFX9-O0-NEXT: v_mov_b32_e32 v1, s13
+; GFX9-O0-NEXT: v_mov_b32_e32 v1, s12
; GFX9-O0-NEXT: buffer_store_dword v1, off, s[0:3], s32 offset:172 ; 4-byte Folded Spill
-; GFX9-O0-NEXT: v_mov_b32_e32 v1, s14
+; GFX9-O0-NEXT: v_mov_b32_e32 v1, s13
; GFX9-O0-NEXT: buffer_store_dword v1, off, s[0:3], s32 offset:168 ; 4-byte Folded Spill
-; GFX9-O0-NEXT: v_mov_b32_e32 v1, s15
+; GFX9-O0-NEXT: v_mov_b32_e32 v1, s14
; GFX9-O0-NEXT: buffer_store_dword v1, off, s[0:3], s32 offset:164 ; 4-byte Folded Spill
-; GFX9-O0-NEXT: v_mov_b32_e32 v1, s16
+; GFX9-O0-NEXT: v_mov_b32_e32 v1, s15
; GFX9-O0-NEXT: buffer_store_dword v1, off, s[0:3], s32 offset:160 ; 4-byte Folded Spill
-; GFX9-O0-NEXT: v_mov_b32_e32 v1, s17
+; GFX9-O0-NEXT: v_mov_b32_e32 v1, s16
; GFX9-O0-NEXT: buffer_store_dword v1, off, s[0:3], s32 offset:156 ; 4-byte Folded Spill
-; GFX9-O0-NEXT: v_mov_b32_e32 v1, s18
+; GFX9-O0-NEXT: v_mov_b32_e32 v1, s17
; GFX9-O0-NEXT: buffer_store_dword v1, off, s[0:3], s32 offset:152 ; 4-byte Folded Spill
-; GFX9-O0-NEXT: v_mov_b32_e32 v1, s19
+; GFX9-O0-NEXT: v_mov_b32_e32 v1, s18
; GFX9-O0-NEXT: buffer_store_dword v1, off, s[0:3], s32 offset:148 ; 4-byte Folded Spill
-; GFX9-O0-NEXT: v_mov_b32_e32 v1, s20
+; GFX9-O0-NEXT: v_mov_b32_e32 v1, s19
; GFX9-O0-NEXT: buffer_store_dword v1, off, s[0:3], s32 offset:144 ; 4-byte Folded Spill
-; GFX9-O0-NEXT: v_mov_b32_e32 v1, s21
+; GFX9-O0-NEXT: v_mov_b32_e32 v1, s20
; GFX9-O0-NEXT: buffer_store_dword v1, off, s[0:3], s32 offset:140 ; 4-byte Folded Spill
-; GFX9-O0-NEXT: v_mov_b32_e32 v1, s22
+; GFX9-O0-NEXT: v_mov_b32_e32 v1, s21
; GFX9-O0-NEXT: buffer_store_dword v1, off, s[0:3], s32 offset:136 ; 4-byte Folded Spill
-; GFX9-O0-NEXT: v_mov_b32_e32 v1, s23
+; GFX9-O0-NEXT: v_mov_b32_e32 v1, s22
; GFX9-O0-NEXT: buffer_store_dword v1, off, s[0:3], s32 offset:132 ; 4-byte Folded Spill
-; GFX9-O0-NEXT: v_mov_b32_e32 v1, s24
+; GFX9-O0-NEXT: v_mov_b32_e32 v1, s23
; GFX9-O0-NEXT: buffer_store_dword v1, off, s[0:3], s32 offset:128 ; 4-byte Folded Spill
+; GFX9-O0-NEXT: v_mov_b32_e32 v1, s24
+; GFX9-O0-NEXT: buffer_store_dword v1, off, s[0:3], s32 offset:124 ; 4-byte Folded Spill
; GFX9-O0-NEXT: v_mov_b32_e32 v47, s25
; GFX9-O0-NEXT: v_mov_b32_e32 v46, s26
; GFX9-O0-NEXT: v_mov_b32_e32 v45, s27
; GFX9-O0-NEXT: v_mov_b32_e32 v44, s28
; GFX9-O0-NEXT: v_mov_b32_e32 v1, s29
-; GFX9-O0-NEXT: buffer_store_dword v1, off, s[0:3], s32 offset:124 ; 4-byte Folded Spill
+; GFX9-O0-NEXT: buffer_store_dword v1, off, s[0:3], s32 offset:120 ; 4-byte Folded Spill
; GFX9-O0-NEXT: ; kill: def $vgpr0 killed $vgpr0 def $vgpr0_vgpr1_vgpr2_vgpr3_vgpr4_vgpr5_vgpr6_vgpr7_vgpr8_vgpr9_vgpr10_vgpr11_vgpr12_vgpr13_vgpr14_vgpr15_vgpr16_vgpr17_vgpr18_vgpr19_vgpr20_vgpr21_vgpr22_vgpr23_vgpr24_vgpr25_vgpr26_vgpr27_vgpr28_vgpr29_vgpr30_vgpr31 killed $exec
; GFX9-O0-NEXT: v_mov_b32_e32 v1, v43
-; GFX9-O0-NEXT: buffer_load_dword v43, off, s[0:3], s32 offset:200 ; 4-byte Folded Reload
+; GFX9-O0-NEXT: buffer_load_dword v43, off, s[0:3], s32 offset:196 ; 4-byte Folded Reload
; GFX9-O0-NEXT: s_waitcnt vmcnt(0)
; GFX9-O0-NEXT: v_mov_b32_e32 v2, v43
-; GFX9-O0-NEXT: buffer_load_dword v43, off, s[0:3], s32 offset:196 ; 4-byte Folded Reload
+; GFX9-O0-NEXT: buffer_load_dword v43, off, s[0:3], s32 offset:192 ; 4-byte Folded Reload
; GFX9-O0-NEXT: s_waitcnt vmcnt(0)
; GFX9-O0-NEXT: v_mov_b32_e32 v3, v43
-; GFX9-O0-NEXT: buffer_load_dword v43, off, s[0:3], s32 offset:192 ; 4-byte Folded Reload
+; GFX9-O0-NEXT: buffer_load_dword v43, off, s[0:3], s32 offset:188 ; 4-byte Folded Reload
; GFX9-O0-NEXT: s_waitcnt vmcnt(0)
; GFX9-O0-NEXT: v_mov_b32_e32 v4, v43
-; GFX9-O0-NEXT: buffer_load_dword v43, off, s[0:3], s32 offset:188 ; 4-byte Folded Reload
+; GFX9-O0-NEXT: buffer_load_dword v43, off, s[0:3], s32 offset:184 ; 4-byte Folded Reload
; GFX9-O0-NEXT: s_waitcnt vmcnt(0)
; GFX9-O0-NEXT: v_mov_b32_e32 v5, v43
-; GFX9-O0-NEXT: buffer_load_dword v43, off, s[0:3], s32 offset:184 ; 4-byte Folded Reload
+; GFX9-O0-NEXT: buffer_load_dword v43, off, s[0:3], s32 offset:180 ; 4-byte Folded Reload
; GFX9-O0-NEXT: s_waitcnt vmcnt(0)
; GFX9-O0-NEXT: v_mov_b32_e32 v6, v43
-; GFX9-O0-NEXT: buffer_load_dword v43, off, s[0:3], s32 offset:180 ; 4-byte Folded Reload
+; GFX9-O0-NEXT: buffer_load_dword v43, off, s[0:3], s32 offset:176 ; 4-byte Folded Reload
; GFX9-O0-NEXT: s_waitcnt vmcnt(0)
; GFX9-O0-NEXT: v_mov_b32_e32 v7, v43
-; GFX9-O0-NEXT: buffer_load_dword v43, off, s[0:3], s32 offset:176 ; 4-byte Folded Reload
+; GFX9-O0-NEXT: buffer_load_dword v43, off, s[0:3], s32 offset:172 ; 4-byte Folded Reload
; GFX9-O0-NEXT: s_waitcnt vmcnt(0)
; GFX9-O0-NEXT: v_mov_b32_e32 v8, v43
-; GFX9-O0-NEXT: buffer_load_dword v43, off, s[0:3], s32 offset:172 ; 4-byte Folded Reload
+; GFX9-O0-NEXT: buffer_load_dword v43, off, s[0:3], s32 offset:168 ; 4-byte Folded Reload
; GFX9-O0-NEXT: s_waitcnt vmcnt(0)
; GFX9-O0-NEXT: v_mov_b32_e32 v9, v43
-; GFX9-O0-NEXT: buffer_load_dword v43, off, s[0:3], s32 offset:168 ; 4-byte Folded Reload
+; GFX9-O0-NEXT: buffer_load_dword v43, off, s[0:3], s32 offset:164 ; 4-byte Folded Reload
; GFX9-O0-NEXT: s_waitcnt vmcnt(0)
; GFX9-O0-NEXT: v_mov_b32_e32 v10, v43
-; GFX9-O0-NEXT: buffer_load_dword v43, off, s[0:3], s32 offset:164 ; 4-byte Folded Reload
+; GFX9-O0-NEXT: buffer_load_dword v43, off, s[0:3], s32 offset:160 ; 4-byte Folded Reload
; GFX9-O0-NEXT: s_waitcnt vmcnt(0)
; GFX9-O0-NEXT: v_mov_b32_e32 v11, v43
-; GFX9-O0-NEXT: buffer_load_dword v43, off, s[0:3], s32 offset:160 ; 4-byte Folded Reload
+; GFX9-O0-NEXT: buffer_load_dword v43, off, s[0:3], s32 offset:156 ; 4-byte Folded Reload
; GFX9-O0-NEXT: s_waitcnt vmcnt(0)
; GFX9-O0-NEXT: v_mov_b32_e32 v12, v43
-; GFX9-O0-NEXT: buffer_load_dword v43, off, s[0:3], s32 offset:156 ; 4-byte Folded Reload
+; GFX9-O0-NEXT: buffer_load_dword v43, off, s[0:3], s32 offset:152 ; 4-byte Folded Reload
; GFX9-O0-NEXT: s_waitcnt vmcnt(0)
; GFX9-O0-NEXT: v_mov_b32_e32 v13, v43
-; GFX9-O0-NEXT: buffer_load_dword v43, off, s[0:3], s32 offset:152 ; 4-byte Folded Reload
+; GFX9-O0-NEXT: buffer_load_dword v43, off, s[0:3], s32 offset:148 ; 4-byte Folded Reload
; GFX9-O0-NEXT: s_waitcnt vmcnt(0)
; GFX9-O0-NEXT: v_mov_b32_e32 v14, v43
-; GFX9-O0-NEXT: buffer_load_dword v43, off, s[0:3], s32 offset:148 ; 4-byte Folded Reload
+; GFX9-O0-NEXT: buffer_load_dword v43, off, s[0:3], s32 offset:144 ; 4-byte Folded Reload
; GFX9-O0-NEXT: s_waitcnt vmcnt(0)
; GFX9-O0-NEXT: v_mov_b32_e32 v15, v43
-; GFX9-O0-NEXT: buffer_load_dword v43, off, s[0:3], s32 offset:144 ; 4-byte Folded Reload
+; GFX9-O0-NEXT: buffer_load_dword v43, off, s[0:3], s32 offset:140 ; 4-byte Folded Reload
; GFX9-O0-NEXT: s_waitcnt vmcnt(0)
; GFX9-O0-NEXT: v_mov_b32_e32 v16, v43
-; GFX9-O0-NEXT: buffer_load_dword v43, off, s[0:3], s32 offset:140 ; 4-byte Folded Reload
+; GFX9-O0-NEXT: buffer_load_dword v43, off, s[0:3], s32 offset:136 ; 4-byte Folded Reload
; GFX9-O0-NEXT: s_waitcnt vmcnt(0)
; GFX9-O0-NEXT: v_mov_b32_e32 v17, v43
-; GFX9-O0-NEXT: buffer_load_dword v43, off, s[0:3], s32 offset:136 ; 4-byte Folded Reload
+; GFX9-O0-NEXT: buffer_load_dword v43, off, s[0:3], s32 offset:132 ; 4-byte Folded Reload
; GFX9-O0-NEXT: s_waitcnt vmcnt(0)
; GFX9-O0-NEXT: v_mov_b32_e32 v18, v43
-; GFX9-O0-NEXT: buffer_load_dword v43, off, s[0:3], s32 offset:132 ; 4-byte Folded Reload
+; GFX9-O0-NEXT: buffer_load_dword v43, off, s[0:3], s32 offset:128 ; 4-byte Folded Reload
; GFX9-O0-NEXT: s_waitcnt vmcnt(0)
; GFX9-O0-NEXT: v_mov_b32_e32 v19, v43
-; GFX9-O0-NEXT: buffer_load_dword v43, off, s[0:3], s32 offset:128 ; 4-byte Folded Reload
+; GFX9-O0-NEXT: buffer_load_dword v43, off, s[0:3], s32 offset:124 ; 4-byte Folded Reload
; GFX9-O0-NEXT: s_waitcnt vmcnt(0)
; GFX9-O0-NEXT: v_mov_b32_e32 v20, v43
-; GFX9-O0-NEXT: buffer_load_dword v43, off, s[0:3], s32 offset:124 ; 4-byte Folded Reload
+; GFX9-O0-NEXT: buffer_load_dword v43, off, s[0:3], s32 offset:120 ; 4-byte Folded Reload
; GFX9-O0-NEXT: v_mov_b32_e32 v21, v47
-; GFX9-O0-NEXT: buffer_load_dword v47, off, s[0:3], s32 offset:104 ; 4-byte Folded Reload
+; GFX9-O0-NEXT: buffer_load_dword v47, off, s[0:3], s32 offset:100 ; 4-byte Folded Reload
; GFX9-O0-NEXT: v_mov_b32_e32 v22, v46
-; GFX9-O0-NEXT: buffer_load_dword v46, off, s[0:3], s32 offset:108 ; 4-byte Folded Reload
+; GFX9-O0-NEXT: buffer_load_dword v46, off, s[0:3], s32 offset:104 ; 4-byte Folded Reload
; GFX9-O0-NEXT: v_mov_b32_e32 v23, v45
-; GFX9-O0-NEXT: buffer_load_dword v45, off, s[0:3], s32 offset:112 ; 4-byte Folded Reload
+; GFX9-O0-NEXT: buffer_load_dword v45, off, s[0:3], s32 offset:108 ; 4-byte Folded Reload
; GFX9-O0-NEXT: v_mov_b32_e32 v24, v44
-; GFX9-O0-NEXT: buffer_load_dword v44, off, s[0:3], s32 offset:116 ; 4-byte Folded Reload
+; GFX9-O0-NEXT: buffer_load_dword v44, off, s[0:3], s32 offset:112 ; 4-byte Folded Reload
; GFX9-O0-NEXT: s_waitcnt vmcnt(4)
; GFX9-O0-NEXT: v_mov_b32_e32 v25, v43
-; GFX9-O0-NEXT: buffer_load_dword v43, off, s[0:3], s32 offset:100 ; 4-byte Folded Reload
+; GFX9-O0-NEXT: buffer_load_dword v43, off, s[0:3], s32 offset:96 ; 4-byte Folded Reload
; GFX9-O0-NEXT: s_waitcnt vmcnt(0)
; GFX9-O0-NEXT: v_mov_b32_e32 v26, v43
-; GFX9-O0-NEXT: buffer_load_dword v43, off, s[0:3], s32 offset:120 ; 4-byte Folded Reload
+; GFX9-O0-NEXT: buffer_load_dword v43, off, s[0:3], s32 offset:116 ; 4-byte Folded Reload
; GFX9-O0-NEXT: v_mov_b32_e32 v27, v47
; GFX9-O0-NEXT: v_mov_b32_e32 v28, v46
; GFX9-O0-NEXT: v_mov_b32_e32 v29, v45
; GFX9-O0-NEXT: v_mov_b32_e32 v30, v44
; GFX9-O0-NEXT: ; kill: def $vgpr31 killed $vgpr43 killed $exec
-; GFX9-O0-NEXT: buffer_load_dword v31, off, s[0:3], s32 offset:120 ; 4-byte Folded Reload
-; GFX9-O0-NEXT: buffer_load_dword v30, off, s[0:3], s32 offset:116 ; 4-byte Folded Reload
-; GFX9-O0-NEXT: buffer_load_dword v29, off, s[0:3], s32 offset:112 ; 4-byte Folded Reload
-; GFX9-O0-NEXT: buffer_load_dword v28, off, s[0:3], s32 offset:108 ; 4-byte Folded Reload
-; GFX9-O0-NEXT: buffer_load_dword v27, off, s[0:3], s32 offset:104 ; 4-byte Folded Reload
-; GFX9-O0-NEXT: buffer_load_dword v26, off, s[0:3], s32 offset:100 ; 4-byte Folded Reload
-; GFX9-O0-NEXT: buffer_load_dword v11, off, s[0:3], s32 offset:96 ; 4-byte Folded Reload
-; GFX9-O0-NEXT: buffer_load_dword v10, off, s[0:3], s32 offset:92 ; 4-byte Folded Reload
-; GFX9-O0-NEXT: buffer_load_dword v9, off, s[0:3], s32 offset:88 ; 4-byte Folded Reload
-; GFX9-O0-NEXT: buffer_load_dword v8, off, s[0:3], s32 offset:84 ; 4-byte Folded Reload
-; GFX9-O0-NEXT: buffer_load_dword v7, off, s[0:3], s32 offset:80 ; 4-byte Folded Reload
-; GFX9-O0-NEXT: buffer_load_dword v6, off, s[0:3], s32 offset:76 ; 4-byte Folded Reload
-; GFX9-O0-NEXT: buffer_load_dword v5, off, s[0:3], s32 offset:72 ; 4-byte Folded Reload
-; GFX9-O0-NEXT: buffer_load_dword v4, off, s[0:3], s32 offset:68 ; 4-byte Folded Reload
-; GFX9-O0-NEXT: buffer_load_dword v2, off, s[0:3], s32 offset:64 ; 4-byte Folded Reload
-; GFX9-O0-NEXT: buffer_load_dword v1, off, s[0:3], s32 offset:60 ; 4-byte Folded Reload
-; GFX9-O0-NEXT: buffer_load_dword v0, off, s[0:3], s32 offset:56 ; 4-byte Folded Reload
+; GFX9-O0-NEXT: buffer_load_dword v31, off, s[0:3], s32 offset:116 ; 4-byte Folded Reload
+; GFX9-O0-NEXT: buffer_load_dword v30, off, s[0:3], s32 offset:112 ; 4-byte Folded Reload
+; GFX9-O0-NEXT: buffer_load_dword v29, off, s[0:3], s32 offset:108 ; 4-byte Folded Reload
+; GFX9-O0-NEXT: buffer_load_dword v28, off, s[0:3], s32 offset:104 ; 4-byte Folded Reload
+; GFX9-O0-NEXT: buffer_load_dword v27, off, s[0:3], s32 offset:100 ; 4-byte Folded Reload
+; GFX9-O0-NEXT: buffer_load_dword v26, off, s[0:3], s32 offset:96 ; 4-byte Folded Reload
+; GFX9-O0-NEXT: buffer_load_dword v11, off, s[0:3], s32 offset:92 ; 4-byte Folded Reload
+; GFX9-O0-NEXT: buffer_load_dword v10, off, s[0:3], s32 offset:88 ; 4-byte Folded Reload
+; GFX9-O0-NEXT: buffer_load_dword v9, off, s[0:3], s32 offset:84 ; 4-byte Folded Reload
+; GFX9-O0-NEXT: buffer_load_dword v8, off, s[0:3], s32 offset:80 ; 4-byte Folded Reload
+; GFX9-O0-NEXT: buffer_load_dword v7, off, s[0:3], s32 offset:76 ; 4-byte Folded Reload
+; GFX9-O0-NEXT: buffer_load_dword v6, off, s[0:3], s32 offset:72 ; 4-byte Folded Reload
+; GFX9-O0-NEXT: buffer_load_dword v5, off, s[0:3], s32 offset:68 ; 4-byte Folded Reload
+; GFX9-O0-NEXT: buffer_load_dword v4, off, s[0:3], s32 offset:64 ; 4-byte Folded Reload
+; GFX9-O0-NEXT: buffer_load_dword v2, off, s[0:3], s32 offset:60 ; 4-byte Folded Reload
+; GFX9-O0-NEXT: buffer_load_dword v1, off, s[0:3], s32 offset:56 ; 4-byte Folded Reload
+; GFX9-O0-NEXT: buffer_load_dword v0, off, s[0:3], s32 offset:52 ; 4-byte Folded Reload
; GFX9-O0-NEXT: ; implicit-def: $sgpr34
; GFX9-O0-NEXT: ; implicit-def: $sgpr34
; GFX9-O0-NEXT: ; kill: def $vgpr2 killed $vgpr2 def $vgpr2_vgpr3 killed $exec
@@ -1256,7 +1227,7 @@
; GFX9-O0-NEXT: buffer_load_dword v40, off, s[0:3], s32 offset:236 ; 4-byte Folded Reload
; GFX9-O0-NEXT: buffer_load_dword v41, off, s[0:3], s32 offset:240 ; 4-byte Folded Reload
; GFX9-O0-NEXT: s_mov_b64 exec, -1
-; GFX9-O0-NEXT: buffer_load_dword v42, off, s[0:3], s32 offset:52 ; 4-byte Folded Reload
+; GFX9-O0-NEXT: buffer_load_dword v42, off, s[0:3], s32 offset:200 ; 4-byte Folded Reload
; GFX9-O0-NEXT: s_mov_b64 exec, s[34:35]
; GFX9-O0-NEXT: s_waitcnt vmcnt(0)
; GFX9-O0-NEXT: s_setpc_b64 s[30:31]
diff --git a/llvm/test/CodeGen/AMDGPU/wwm-reserved.ll b/llvm/test/CodeGen/AMDGPU/wwm-reserved.ll
--- a/llvm/test/CodeGen/AMDGPU/wwm-reserved.ll
+++ b/llvm/test/CodeGen/AMDGPU/wwm-reserved.ll
@@ -125,8 +125,7 @@
; GFX9-O0: v_add_u32_e64 v3, v3, v6
%tmp136 = add i32 %tmp134, %tmp107
%tmp137 = tail call i32 @llvm.amdgcn.wwm.i32(i32 %tmp136)
-; GFX9-O0: buffer_store_dword v1
-; GFX9-O3: buffer_store_dword v0
+; GFX9: buffer_store_dword v0
call void @llvm.amdgcn.raw.ptr.buffer.store.i32(i32 %tmp137, ptr addrspace(8) %tmp14, i32 4, i32 0, i32 0)
ret void
}
@@ -321,8 +320,7 @@
; GFX9-O0: v_add_u32_e64 v3, v3, v6
%tmp136 = add i32 %tmp134, %tmp107
%tmp137 = tail call i32 @llvm.amdgcn.strict.wwm.i32(i32 %tmp136)
-; GFX9-O0: buffer_store_dword v1
-; GFX9-O3: buffer_store_dword v0
+; GFX9: buffer_store_dword v0
call void @llvm.amdgcn.raw.ptr.buffer.store.i32(i32 %tmp137, ptr addrspace(8) %tmp14, i32 4, i32 0, i32 0)
ret void
}
@@ -408,4 +406,4 @@
declare void @llvm.amdgcn.raw.ptr.buffer.store.v2f32(<2 x float>, ptr addrspace(8), i32, i32, i32)
declare void @llvm.amdgcn.raw.ptr.buffer.store.v4f32(<4 x float>, ptr addrspace(8), i32, i32, i32)
declare <2 x i32> @llvm.amdgcn.s.buffer.load.v2i32(<4 x i32>, i32, i32)
-declare <4 x i32> @llvm.amdgcn.s.buffer.load.v4i32(<4 x i32>, i32, i32)
\ No newline at end of file
+declare <4 x i32> @llvm.amdgcn.s.buffer.load.v4i32(<4 x i32>, i32, i32)
diff --git a/llvm/test/CodeGen/MIR/AMDGPU/stack-id-assert.mir b/llvm/test/CodeGen/MIR/AMDGPU/stack-id-assert.mir
--- a/llvm/test/CodeGen/MIR/AMDGPU/stack-id-assert.mir
+++ b/llvm/test/CodeGen/MIR/AMDGPU/stack-id-assert.mir
@@ -3,7 +3,7 @@
# contains not dead objects only. So using objects IDs as offset in the storage
# caused out of bounds access.
-# RUN: llc -march=amdgcn -start-before=si-lower-sgpr-spills -stop-after=prologepilog -verify-machineinstrs -o - %s | FileCheck %s
+# RUN: llc -march=amdgcn -run-pass=si-lower-sgpr-spills,prologepilog -verify-machineinstrs -o - %s | FileCheck %s
# CHECK-LABEL: name: foo
# CHECK: {{^}}fixedStack: []
diff --git a/llvm/test/CodeGen/Mips/madd-msub.ll b/llvm/test/CodeGen/Mips/madd-msub.ll
--- a/llvm/test/CodeGen/Mips/madd-msub.ll
+++ b/llvm/test/CodeGen/Mips/madd-msub.ll
@@ -42,22 +42,22 @@
;
; 64-LABEL: madd1:
; 64: # %bb.0: # %entry
-; 64-NEXT: sll $4, $4, 0
-; 64-NEXT: sll $5, $5, 0
-; 64-NEXT: dmult $5, $4
+; 64-NEXT: sll $1, $4, 0
+; 64-NEXT: sll $2, $5, 0
+; 64-NEXT: dmult $2, $1
; 64-NEXT: mflo $1
-; 64-NEXT: sll $6, $6, 0
+; 64-NEXT: sll $2, $6, 0
; 64-NEXT: jr $ra
-; 64-NEXT: daddu $2, $1, $6
+; 64-NEXT: daddu $2, $1, $2
;
; 64R6-LABEL: madd1:
; 64R6: # %bb.0: # %entry
-; 64R6-NEXT: sll $4, $4, 0
-; 64R6-NEXT: sll $5, $5, 0
-; 64R6-NEXT: dmul $1, $5, $4
-; 64R6-NEXT: sll $6, $6, 0
+; 64R6-NEXT: sll $1, $4, 0
+; 64R6-NEXT: sll $2, $5, 0
+; 64R6-NEXT: dmul $1, $2, $1
+; 64R6-NEXT: sll $2, $6, 0
; 64R6-NEXT: jr $ra
-; 64R6-NEXT: daddu $2, $1, $6
+; 64R6-NEXT: daddu $2, $1, $2
;
; 16-LABEL: madd1:
; 16: # %bb.0: # %entry
@@ -173,18 +173,18 @@
;
; 64-LABEL: madd3:
; 64: # %bb.0: # %entry
-; 64-NEXT: sll $4, $4, 0
-; 64-NEXT: sll $5, $5, 0
-; 64-NEXT: dmult $5, $4
+; 64-NEXT: sll $1, $4, 0
+; 64-NEXT: sll $2, $5, 0
+; 64-NEXT: dmult $2, $1
; 64-NEXT: mflo $1
; 64-NEXT: jr $ra
; 64-NEXT: daddu $2, $1, $6
;
; 64R6-LABEL: madd3:
; 64R6: # %bb.0: # %entry
-; 64R6-NEXT: sll $4, $4, 0
-; 64R6-NEXT: sll $5, $5, 0
-; 64R6-NEXT: dmul $1, $5, $4
+; 64R6-NEXT: sll $1, $4, 0
+; 64R6-NEXT: sll $2, $5, 0
+; 64R6-NEXT: dmul $1, $2, $1
; 64R6-NEXT: jr $ra
; 64R6-NEXT: daddu $2, $1, $6
;
@@ -291,22 +291,22 @@
;
; 64-LABEL: msub1:
; 64: # %bb.0: # %entry
-; 64-NEXT: sll $4, $4, 0
-; 64-NEXT: sll $5, $5, 0
-; 64-NEXT: dmult $5, $4
+; 64-NEXT: sll $1, $4, 0
+; 64-NEXT: sll $2, $5, 0
+; 64-NEXT: dmult $2, $1
; 64-NEXT: mflo $1
-; 64-NEXT: sll $6, $6, 0
+; 64-NEXT: sll $2, $6, 0
; 64-NEXT: jr $ra
-; 64-NEXT: dsubu $2, $6, $1
+; 64-NEXT: dsubu $2, $2, $1
;
; 64R6-LABEL: msub1:
; 64R6: # %bb.0: # %entry
-; 64R6-NEXT: sll $4, $4, 0
-; 64R6-NEXT: sll $5, $5, 0
-; 64R6-NEXT: dmul $1, $5, $4
-; 64R6-NEXT: sll $6, $6, 0
+; 64R6-NEXT: sll $1, $4, 0
+; 64R6-NEXT: sll $2, $5, 0
+; 64R6-NEXT: dmul $1, $2, $1
+; 64R6-NEXT: sll $2, $6, 0
; 64R6-NEXT: jr $ra
-; 64R6-NEXT: dsubu $2, $6, $1
+; 64R6-NEXT: dsubu $2, $2, $1
;
; 16-LABEL: msub1:
; 16: # %bb.0: # %entry
@@ -424,18 +424,18 @@
;
; 64-LABEL: msub3:
; 64: # %bb.0: # %entry
-; 64-NEXT: sll $4, $4, 0
-; 64-NEXT: sll $5, $5, 0
-; 64-NEXT: dmult $5, $4
+; 64-NEXT: sll $1, $4, 0
+; 64-NEXT: sll $2, $5, 0
+; 64-NEXT: dmult $2, $1
; 64-NEXT: mflo $1
; 64-NEXT: jr $ra
; 64-NEXT: dsubu $2, $6, $1
;
; 64R6-LABEL: msub3:
; 64R6: # %bb.0: # %entry
-; 64R6-NEXT: sll $4, $4, 0
-; 64R6-NEXT: sll $5, $5, 0
-; 64R6-NEXT: dmul $1, $5, $4
+; 64R6-NEXT: sll $1, $4, 0
+; 64R6-NEXT: sll $2, $5, 0
+; 64R6-NEXT: dmul $1, $2, $1
; 64R6-NEXT: jr $ra
; 64R6-NEXT: dsubu $2, $6, $1
;
@@ -546,22 +546,22 @@
;
; 64-LABEL: msub5:
; 64: # %bb.0: # %entry
-; 64-NEXT: sll $4, $4, 0
-; 64-NEXT: sll $5, $5, 0
-; 64-NEXT: dmult $5, $4
+; 64-NEXT: sll $1, $4, 0
+; 64-NEXT: sll $2, $5, 0
+; 64-NEXT: dmult $2, $1
; 64-NEXT: mflo $1
-; 64-NEXT: sll $6, $6, 0
+; 64-NEXT: sll $2, $6, 0
; 64-NEXT: jr $ra
-; 64-NEXT: dsubu $2, $1, $6
+; 64-NEXT: dsubu $2, $1, $2
;
; 64R6-LABEL: msub5:
; 64R6: # %bb.0: # %entry
-; 64R6-NEXT: sll $4, $4, 0
-; 64R6-NEXT: sll $5, $5, 0
-; 64R6-NEXT: dmul $1, $5, $4
-; 64R6-NEXT: sll $6, $6, 0
+; 64R6-NEXT: sll $1, $4, 0
+; 64R6-NEXT: sll $2, $5, 0
+; 64R6-NEXT: dmul $1, $2, $1
+; 64R6-NEXT: sll $2, $6, 0
; 64R6-NEXT: jr $ra
-; 64R6-NEXT: dsubu $2, $1, $6
+; 64R6-NEXT: dsubu $2, $1, $2
;
; 16-LABEL: msub5:
; 16: # %bb.0: # %entry
diff --git a/llvm/test/CodeGen/X86/GlobalISel/add-ext.ll b/llvm/test/CodeGen/X86/GlobalISel/add-ext.ll
--- a/llvm/test/CodeGen/X86/GlobalISel/add-ext.ll
+++ b/llvm/test/CodeGen/X86/GlobalISel/add-ext.ll
@@ -205,8 +205,8 @@
; CHECK-NEXT: addq %rdi, %rcx
; CHECK-NEXT: movl (%rcx), %ecx
; CHECK-NEXT: addl (%rax), %ecx
-; CHECK-NEXT: movl %esi, %esi
-; CHECK-NEXT: imulq $4, %rsi, %rax
+; CHECK-NEXT: movl %esi, %eax
+; CHECK-NEXT: imulq $4, %rax, %rax
; CHECK-NEXT: addq %rdi, %rax
; CHECK-NEXT: movl %ecx, (%rax)
; CHECK-NEXT: retq
diff --git a/llvm/test/CodeGen/X86/dagcombine-cse.ll b/llvm/test/CodeGen/X86/dagcombine-cse.ll
--- a/llvm/test/CodeGen/X86/dagcombine-cse.ll
+++ b/llvm/test/CodeGen/X86/dagcombine-cse.ll
@@ -106,24 +106,24 @@
;
; X64-LABEL: square_high:
; X64: ## %bb.0: ## %entry
-; X64-NEXT: movl %esi, %esi
-; X64-NEXT: movq %rsi, %rax
+; X64-NEXT: movl %esi, %ecx
+; X64-NEXT: movq %rcx, %rax
; X64-NEXT: mulq %rdi
-; X64-NEXT: movq %rdx, %rcx
+; X64-NEXT: movq %rdx, %rsi
; X64-NEXT: movq %rax, %r8
; X64-NEXT: movq %rdi, %rax
; X64-NEXT: mulq %rdi
; X64-NEXT: addq %r8, %rdx
-; X64-NEXT: movq %rcx, %rax
+; X64-NEXT: movq %rsi, %rax
; X64-NEXT: adcq $0, %rax
; X64-NEXT: addq %rdx, %r8
-; X64-NEXT: adcq %rcx, %rax
-; X64-NEXT: imulq %rsi, %rsi
-; X64-NEXT: addq %rax, %rsi
-; X64-NEXT: shrdq $32, %rsi, %r8
-; X64-NEXT: shrq $32, %rsi
+; X64-NEXT: adcq %rsi, %rax
+; X64-NEXT: imulq %rcx, %rcx
+; X64-NEXT: addq %rax, %rcx
+; X64-NEXT: shrdq $32, %rcx, %r8
+; X64-NEXT: shrq $32, %rcx
; X64-NEXT: movq %r8, %rax
-; X64-NEXT: movq %rsi, %rdx
+; X64-NEXT: movq %rcx, %rdx
; X64-NEXT: retq
entry:
%conv = zext i96 %x to i192
diff --git a/llvm/test/CodeGen/X86/fold-and-shift-x86_64.ll b/llvm/test/CodeGen/X86/fold-and-shift-x86_64.ll
--- a/llvm/test/CodeGen/X86/fold-and-shift-x86_64.ll
+++ b/llvm/test/CodeGen/X86/fold-and-shift-x86_64.ll
@@ -34,8 +34,8 @@
define i8 @t3(ptr %X, i64 %i) {
; CHECK-LABEL: t3:
; CHECK: # %bb.0: # %entry
-; CHECK-NEXT: movl %esi, %esi
-; CHECK-NEXT: movzbl (%rdi,%rsi,4), %eax
+; CHECK-NEXT: movl %esi, %eax
+; CHECK-NEXT: movzbl (%rdi,%rax,4), %eax
; CHECK-NEXT: retq
entry:
diff --git a/llvm/test/CodeGen/X86/unfold-masked-merge-scalar-constmask-lowhigh.ll b/llvm/test/CodeGen/X86/unfold-masked-merge-scalar-constmask-lowhigh.ll
--- a/llvm/test/CodeGen/X86/unfold-masked-merge-scalar-constmask-lowhigh.ll
+++ b/llvm/test/CodeGen/X86/unfold-masked-merge-scalar-constmask-lowhigh.ll
@@ -78,18 +78,18 @@
define i64 @out64_constmask(i64 %x, i64 %y) {
; CHECK-NOBMI-LABEL: out64_constmask:
; CHECK-NOBMI: # %bb.0:
-; CHECK-NOBMI-NEXT: movl %edi, %edi
+; CHECK-NOBMI-NEXT: movl %edi, %ecx
; CHECK-NOBMI-NEXT: movabsq $-4294967296, %rax # imm = 0xFFFFFFFF00000000
; CHECK-NOBMI-NEXT: andq %rsi, %rax
-; CHECK-NOBMI-NEXT: orq %rdi, %rax
+; CHECK-NOBMI-NEXT: orq %rcx, %rax
; CHECK-NOBMI-NEXT: retq
;
; CHECK-BMI-LABEL: out64_constmask:
; CHECK-BMI: # %bb.0:
-; CHECK-BMI-NEXT: movl %edi, %edi
+; CHECK-BMI-NEXT: movl %edi, %ecx
; CHECK-BMI-NEXT: movabsq $-4294967296, %rax # imm = 0xFFFFFFFF00000000
; CHECK-BMI-NEXT: andq %rsi, %rax
-; CHECK-BMI-NEXT: orq %rdi, %rax
+; CHECK-BMI-NEXT: orq %rcx, %rax
; CHECK-BMI-NEXT: retq
%mx = and i64 %x, 4294967295
%my = and i64 %y, -4294967296