[AMDGPU] Change FLAT SADDR to VADDR form in moveToVALU

[rampitec]

Summary

Instead of legalizing saddr operand with a readfirstlane
when address is moved from SGPR to VGPR we can just
change the opcode.

[Joe_Nash]

llvm/lib/Target/AMDGPU/SIInstrInfo.cpp

5037

I don't understand why this assert is true. Does some previous check guarantee that? Other than that LGTM, but please wait for @arsenm

[rampitec]

llvm/lib/Target/AMDGPU/SIInstrInfo.cpp

5037

All flat global SADDR instructions have $vaddr component and other instructions cannot get to this point as getGlobalVaddrOp() will return -1. This changes in the followup D101408 which can also process flat scratch.

[Joe_Nash]

llvm/lib/Target/AMDGPU/SIInstrInfo.cpp

5037

I see it now, thanks.

[rampitec]

Renamed function, added comments to helpers.

[rampitec]

Remove dead def because it is simple, do not leave it to DCE.

[rampitec]

Fixed broken operand use list after moveOperands().

[arsenm]

llvm/lib/Target/AMDGPU/SIInstrInfo.cpp

5049–5063

At this point wouldn't it be simpler to just create a fresh new instruction and delete the old one?

[rampitec]

llvm/lib/Target/AMDGPU/SIInstrInfo.cpp

5049–5063

Unfortunately callers expect iterator to be intact. Otherwise probably yes, although it would not be less code.

[arsenm]

llvm/lib/Target/AMDGPU/SIInstrInfo.cpp

5049–5060

Probably should add a comment explaining why this nightmare is here

5049–5063

I think in some contexts in SIFixSGPRCopies the original instruction is erased, but it's a bit of a mess

llvm/test/CodeGen/AMDGPU/move-load-addr-to-valu.mir

35

Can you also add an IR test where this matters? I'm not understanding why the DAG divergence analysis would let this happen

[rampitec]

llvm/test/CodeGen/AMDGPU/move-load-addr-to-valu.mir

35

The test case shall be more than 100 instructions long, this is the threshold for memory dependency analysis to give up on the "noclobber" check. The divergence analysis tells it is ok, it is uniform. But noclobber is not set, so the address registers are not known not to be clobberd and we are here.

It really happens in large basic blocks. Probably it can also happen if we just had no SALU instructions to do some computations, so ended up with VALU which was pripagated, but that is not the case I have started with. I.e. it is really uniform, and readfirstlane is valid, but we don't want to issue it.

I am not really sure we want a test of that size. It will be more than obscure in the first place.

[foad]

llvm/lib/Target/AMDGPU/SIInstrInfo.cpp

5027

Use Inst.getMF() ?

5050

Redundant assert, you've just tested this condition.

5055–5056

Are these last two lines really necessary? Hasn't MRI.moveOperands already handled this?

5062

VAddrDef can't be 0 here, that was already checked earlier.

llvm/test/CodeGen/AMDGPU/move-load-addr-to-valu.mir

35

You could test with -memdep-block-scan-limit=1 ?

[rampitec]

Addressed review comments. One thing left is a reasonable IR test.

llvm/lib/Target/AMDGPU/SIInstrInfo.cpp

5055–5056

It did not. I have added the comment.

5062

Right, this is a part of D101408. Removed from this patch.

llvm/test/CodeGen/AMDGPU/move-load-addr-to-valu.mir

35

I will check if I can produce a reasonable IR test.

[rampitec]

Added IR test.

[arsenm]

LGTM with test comment

llvm/test/CodeGen/AMDGPU/global-load-saddr-to-vaddr.ll

2 ↗

(On Diff #342016)

Add a comment explaining what this is testing

21–23 ↗

(On Diff #342016)

Oh, this is the case where the loop index/pointer is really uniform, but forced to be a vector branch. It's just better to do the VALU address computation than do it scalar and copy with readfirstlane

llvm/test/CodeGen/AMDGPU/move-load-addr-to-valu.mir

35

Whether or not this can be converted to SMRD is a different problem

[rampitec]

Added comment to the test.

llvm/test/CodeGen/AMDGPU/global-load-saddr-to-vaddr.ll

21–23 ↗

(On Diff #342016)

Yes, the memory dependency tracing was another case related to the opposite situation. This is just because we are producing VGPRs with a uniform value for the pointer.

llvm/lib/Target/AMDGPU/SIInstrInfo.h

public:
 
   /// Legalize all operands in this instruction.  This function may create new
   /// instructions and control-flow around \p MI.  If present, \p MDT is
   /// updated.
   /// \returns A new basic block that contains \p MI if new blocks were created.
   MachineBasicBlock *
   legalizeOperands(MachineInstr &MI, MachineDominatorTree *MDT = nullptr) const;
 
+  /// Change SADDR form of a FLAT \p Inst to its VADDR form if saddr operand
+  /// was moved to VGPR. \returns true if succeeded.
+  bool moveFlatAddrToVGPR(MachineInstr &Inst) const;
+
   /// Replace this instruction's opcode with the equivalent VALU
   /// opcode.  This function will also move the users of \p MI to the
   /// VALU if necessary. If present, \p MDT is updated.
   MachineBasicBlock *moveToVALU(MachineInstr &MI,
                                 MachineDominatorTree *MDT = nullptr) const;
 
   void insertNoop(MachineBasicBlock &MBB,
                   MachineBasicBlock::iterator MI) const override;
namespace AMDGPU {
   int getMUBUFNoLdsInst(uint16_t Opcode);
 
   LLVM_READONLY
   int getAtomicNoRetOp(uint16_t Opcode);
 
   LLVM_READONLY
   int getSOPKOp(uint16_t Opcode);
 
+  /// \returns SADDR form of a FLAT Global instruction given an \p Opcode
+  /// of a VADDR form.
   LLVM_READONLY
   int getGlobalSaddrOp(uint16_t Opcode);
 
+  /// \returns VADDR form of a FLAT Global instruction given an \p Opcode
+  /// of a SADDR form.
+  LLVM_READONLY
+  int getGlobalVaddrOp(uint16_t Opcode);
+
   LLVM_READONLY
   int getVCMPXNoSDstOp(uint16_t Opcode);
 
   LLVM_READONLY
   int getFlatScratchInstSTfromSS(uint16_t Opcode);
 
   LLVM_READONLY
   int getFlatScratchInstSSfromSV(uint16_t Opcode);

llvm/lib/Target/AMDGPU/SIInstrInfo.cpp

void SIInstrInfo::legalizeOperandsSMRD(MachineRegisterInfo &MRI,
   }
   MachineOperand *SOff = getNamedOperand(MI, AMDGPU::OpName::soff);
   if (SOff && !RI.isSGPRClass(MRI.getRegClass(SOff->getReg()))) {
     Register SGPR = readlaneVGPRToSGPR(SOff->getReg(), MI, MRI);
     SOff->setReg(SGPR);
   }
 }
 
+bool SIInstrInfo::moveFlatAddrToVGPR(MachineInstr &Inst) const {
+  unsigned Opc = Inst.getOpcode();
+  int OldSAddrIdx = AMDGPU::getNamedOperandIdx(Opc, AMDGPU::OpName::saddr);
+  if (OldSAddrIdx < 0)
+    return false;
+
+  assert(isSegmentSpecificFLAT(Inst));
+
+  int NewOpc = AMDGPU::getGlobalVaddrOp(Opc);
+  if (NewOpc < 0)
+    return false;
+
+  MachineRegisterInfo &MRI = Inst.getParent()->getParent()->getRegInfo();
+  MachineOperand &SAddr = Inst.getOperand(OldSAddrIdx);
+  if (RI.isSGPRReg(MRI, SAddr.getReg()))
+    return false;
+
+  int NewVAddrIdx = AMDGPU::getNamedOperandIdx(NewOpc, AMDGPU::OpName::vaddr);
+  if (NewVAddrIdx < 0)
+    return false;
+
+  int OldVAddrIdx = AMDGPU::getNamedOperandIdx(Opc, AMDGPU::OpName::vaddr);
+  assert(OldVAddrIdx >= 0);
+
+  // Check vaddr, it shall be zero
+  MachineOperand &VAddr = Inst.getOperand(OldVAddrIdx);
+  MachineInstr *VAddrDef = MRI.getUniqueVRegDef(VAddr.getReg());
+  if (!VAddrDef || VAddrDef->getOpcode() != AMDGPU::V_MOV_B32_e32 ||
+      !VAddrDef->getOperand(1).isImm() || VAddrDef->getOperand(1).getImm() != 0)
+    return false;
+
+  const MCInstrDesc &NewDesc = get(NewOpc);
+  Inst.setDesc(NewDesc);
+
+  if (OldVAddrIdx == NewVAddrIdx) {
+    assert(OldVAddrIdx == NewVAddrIdx);
+    MachineOperand &NewVAddr = Inst.getOperand(NewVAddrIdx);
+    MRI.removeRegOperandFromUseList(&NewVAddr);
+    MRI.moveOperands(&NewVAddr, &SAddr, 1);
+    Inst.RemoveOperand(OldSAddrIdx);
+    MRI.removeRegOperandFromUseList(&NewVAddr);
+    MRI.addRegOperandToUseList(&NewVAddr);
+  } else {
+    assert(OldSAddrIdx == NewVAddrIdx);
+    Inst.RemoveOperand(OldVAddrIdx);
+  }
+
+  if (VAddrDef && MRI.use_nodbg_empty(VAddrDef->getOperand(0).getReg()))
+    VAddrDef->eraseFromParent();
+
+  return true;
+}
+
 // FIXME: Remove this when SelectionDAG is obsoleted.
 void SIInstrInfo::legalizeOperandsFLAT(MachineRegisterInfo &MRI,
                                        MachineInstr &MI) const {
   if (!isSegmentSpecificFLAT(MI))
     return;
 
   // Fixup SGPR operands in VGPRs. We only select these when the DAG divergence
   // thinks they are uniform, so a readfirstlane should be valid.
   MachineOperand *SAddr = getNamedOperand(MI, AMDGPU::OpName::saddr);
   if (!SAddr || RI.isSGPRClass(MRI.getRegClass(SAddr->getReg())))
     return;
 
+  if (moveFlatAddrToVGPR(MI))
+    return;
+
   Register ToSGPR = readlaneVGPRToSGPR(SAddr->getReg(), MI, MRI);
   SAddr->setReg(ToSGPR);
 }
 
 void SIInstrInfo::legalizeGenericOperand(MachineBasicBlock &InsertMBB,
                                          MachineBasicBlock::iterator I,
                                          const TargetRegisterClass *DstRC,
                                          MachineOperand &Op,

llvm/lib/Target/AMDGPU/SIInstrInfo.td

 def getGlobalSaddrOp : InstrMapping {
   let FilterClass = "GlobalSaddrTable";
   let RowFields = ["SaddrOp"];
   let ColFields = ["IsSaddr"];
   let KeyCol = ["0"];
   let ValueCols = [["1"]];
 }
 
+// Maps a GLOBAL SADDR to its VADDR form.
+def getGlobalVaddrOp : InstrMapping {
+  let FilterClass = "GlobalSaddrTable";
+  let RowFields = ["SaddrOp"];
+  let ColFields = ["IsSaddr"];
+  let KeyCol = ["1"];
+  let ValueCols = [["0"]];
+}
+
 // Maps a v_cmpx opcode with sdst to opcode without sdst.
 def getVCMPXNoSDstOp : InstrMapping {
   let FilterClass = "VCMPXNoSDstTable";
   let RowFields = ["NoSDstOp"];
   let ColFields = ["HasSDst"];
   let KeyCol = ["1"];
   let ValueCols = [["0"]];
 }

llvm/test/CodeGen/AMDGPU/move-load-addr-to-valu.mir

+# NOTE: Assertions have been autogenerated by utils/update_mir_test_checks.py
+# RUN: llc -march=amdgcn -mcpu=gfx900 -run-pass=si-fix-sgpr-copies -verify-machineinstrs -o - %s | FileCheck --check-prefix=GCN %s
+
+---
+name:            global_load_saddr_to_valu
+tracksRegLiveness: true
+body:             |
+  ; GCN-LABEL: name: global_load_saddr_to_valu
+  ; GCN: bb.0:
+  ; GCN:   successors: %bb.1(0x80000000)
+  ; GCN:   liveins: $vgpr0_vgpr1
+  ; GCN:   [[COPY:%[0-9]+]]:vreg_64 = COPY $vgpr0_vgpr1
+  ; GCN: bb.1:
+  ; GCN:   successors: %bb.1(0x40000000), %bb.2(0x40000000)
+  ; GCN:   [[PHI:%[0-9]+]]:vreg_64 = PHI [[COPY]], %bb.0, %7, %bb.1
+  ; GCN:   [[GLOBAL_LOAD_DWORD:%[0-9]+]]:vgpr_32 = GLOBAL_LOAD_DWORD [[PHI]], 0, 0, implicit $exec
+  ; GCN:   [[COPY1:%[0-9]+]]:vgpr_32 = COPY [[PHI]].sub0
+  ; GCN:   [[COPY2:%[0-9]+]]:vgpr_32 = COPY [[PHI]].sub1
+  ; GCN:   [[V_AND_B32_e64_:%[0-9]+]]:vgpr_32 = V_AND_B32_e64 [[COPY1]], 1, implicit $exec
+  ; GCN:   [[V_AND_B32_e64_1:%[0-9]+]]:vgpr_32 = V_AND_B32_e64 [[COPY2]], 0, implicit $exec
+  ; GCN:   [[REG_SEQUENCE:%[0-9]+]]:vreg_64 = REG_SEQUENCE [[V_AND_B32_e64_]], %subreg.sub0, [[V_AND_B32_e64_1]], %subreg.sub1
+  ; GCN:   V_CMP_NE_U64_e32 0, [[REG_SEQUENCE]], implicit-def $vcc, implicit $exec
+  ; GCN:   [[COPY3:%[0-9]+]]:vreg_64 = COPY [[REG_SEQUENCE]], implicit $exec
+  ; GCN:   $vcc = S_AND_B64 $exec, $vcc, implicit-def $scc
+  ; GCN:   S_CBRANCH_VCCNZ %bb.1, implicit $vcc
+  ; GCN: bb.2:
+  ; GCN:   S_ENDPGM 0
+  bb.0:
+    liveins: $vgpr0_vgpr1
+      %0:sreg_64 = COPY $vgpr0_vgpr1
+
+  bb.1:
+      %1:sreg_64 = PHI %0, %bb.0, %2, %bb.1
+      %3:vgpr_32 = V_MOV_B32_e32 0, implicit $exec
+      %4:vgpr_32 = GLOBAL_LOAD_DWORD_SADDR %1, %3, 0, 0, implicit $exec
+      %2:sreg_64 = S_AND_B64 %1, 1, implicit-def $scc
+      S_CMP_LG_U64 %2, 0, implicit-def $scc
+      S_CBRANCH_SCC1 %bb.1, implicit $scc
+
+  bb.2:
+      S_ENDPGM 0
+...
+
+---
+name:            global_load_saddr_to_valu_non_zero_vaddr
+tracksRegLiveness: true
+body:             |
+  ; GCN-LABEL: name: global_load_saddr_to_valu_non_zero_vaddr
+  ; GCN: bb.0:
+  ; GCN:   successors: %bb.1(0x80000000)
+  ; GCN:   liveins: $vgpr0_vgpr1
+  ; GCN:   [[COPY:%[0-9]+]]:vreg_64 = COPY $vgpr0_vgpr1
+  ; GCN: bb.1:
+  ; GCN:   successors: %bb.1(0x40000000), %bb.2(0x40000000)
+  ; GCN:   [[PHI:%[0-9]+]]:vreg_64 = PHI [[COPY]], %bb.0, %7, %bb.1
+  ; GCN:   [[V_MOV_B32_e32_:%[0-9]+]]:vgpr_32 = V_MOV_B32_e32 1, implicit $exec
+  ; GCN:   [[V_READFIRSTLANE_B32_:%[0-9]+]]:sgpr_32 = V_READFIRSTLANE_B32 [[PHI]].sub0, implicit $exec
+  ; GCN:   [[V_READFIRSTLANE_B32_1:%[0-9]+]]:sgpr_32 = V_READFIRSTLANE_B32 [[PHI]].sub1, implicit $exec
+  ; GCN:   [[REG_SEQUENCE:%[0-9]+]]:sreg_64 = REG_SEQUENCE [[V_READFIRSTLANE_B32_]], %subreg.sub0, [[V_READFIRSTLANE_B32_1]], %subreg.sub1
+  ; GCN:   [[GLOBAL_LOAD_DWORD_SADDR:%[0-9]+]]:vgpr_32 = GLOBAL_LOAD_DWORD_SADDR [[REG_SEQUENCE]], [[V_MOV_B32_e32_]], 0, 0, implicit $exec
+  ; GCN:   [[COPY1:%[0-9]+]]:vgpr_32 = COPY [[PHI]].sub0
+  ; GCN:   [[COPY2:%[0-9]+]]:vgpr_32 = COPY [[PHI]].sub1
+  ; GCN:   [[V_AND_B32_e64_:%[0-9]+]]:vgpr_32 = V_AND_B32_e64 [[COPY1]], 1, implicit $exec
+  ; GCN:   [[V_AND_B32_e64_1:%[0-9]+]]:vgpr_32 = V_AND_B32_e64 [[COPY2]], 0, implicit $exec
+  ; GCN:   [[REG_SEQUENCE1:%[0-9]+]]:vreg_64 = REG_SEQUENCE [[V_AND_B32_e64_]], %subreg.sub0, [[V_AND_B32_e64_1]], %subreg.sub1
+  ; GCN:   V_CMP_NE_U64_e32 0, [[REG_SEQUENCE1]], implicit-def $vcc, implicit $exec
+  ; GCN:   [[COPY3:%[0-9]+]]:vreg_64 = COPY [[REG_SEQUENCE1]], implicit $exec
+  ; GCN:   $vcc = S_AND_B64 $exec, $vcc, implicit-def $scc
+  ; GCN:   S_CBRANCH_VCCNZ %bb.1, implicit $vcc
+  ; GCN: bb.2:
+  ; GCN:   S_ENDPGM 0
+  bb.0:
+    liveins: $vgpr0_vgpr1
+      %0:sreg_64 = COPY $vgpr0_vgpr1
+
+  bb.1:
+      %1:sreg_64 = PHI %0, %bb.0, %2, %bb.1
+      %3:vgpr_32 = V_MOV_B32_e32 1, implicit $exec
+      %4:vgpr_32 = GLOBAL_LOAD_DWORD_SADDR %1, %3, 0, 0, implicit $exec
+      %2:sreg_64 = S_AND_B64 %1, 1, implicit-def $scc
+      S_CMP_LG_U64 %2, 0, implicit-def $scc
+      S_CBRANCH_SCC1 %bb.1, implicit $scc
+
+  bb.2:
+      S_ENDPGM 0
+...
+
+
+---
+name:            global_load_saddr_to_valu_undef_vaddr
+tracksRegLiveness: true
+body:             |
+  ; GCN-LABEL: name: global_load_saddr_to_valu_undef_vaddr
+  ; GCN: bb.0:
+  ; GCN:   successors: %bb.1(0x80000000)
+  ; GCN:   liveins: $vgpr0_vgpr1
+  ; GCN:   [[COPY:%[0-9]+]]:vreg_64 = COPY $vgpr0_vgpr1
+  ; GCN: bb.1:
+  ; GCN:   successors: %bb.1(0x40000000), %bb.2(0x40000000)
+  ; GCN:   [[PHI:%[0-9]+]]:vreg_64 = PHI [[COPY]], %bb.0, %7, %bb.1
+  ; GCN:   [[V_READFIRSTLANE_B32_:%[0-9]+]]:sgpr_32 = V_READFIRSTLANE_B32 [[PHI]].sub0, implicit $exec
+  ; GCN:   [[V_READFIRSTLANE_B32_1:%[0-9]+]]:sgpr_32 = V_READFIRSTLANE_B32 [[PHI]].sub1, implicit $exec
+  ; GCN:   [[REG_SEQUENCE:%[0-9]+]]:sreg_64 = REG_SEQUENCE [[V_READFIRSTLANE_B32_]], %subreg.sub0, [[V_READFIRSTLANE_B32_1]], %subreg.sub1
+  ; GCN:   [[GLOBAL_LOAD_DWORD_SADDR:%[0-9]+]]:vgpr_32 = GLOBAL_LOAD_DWORD_SADDR [[REG_SEQUENCE]], undef %4:vgpr_32, 0, 0, implicit $exec
+  ; GCN:   [[COPY1:%[0-9]+]]:vgpr_32 = COPY [[PHI]].sub0
+  ; GCN:   [[COPY2:%[0-9]+]]:vgpr_32 = COPY [[PHI]].sub1
+  ; GCN:   [[V_AND_B32_e64_:%[0-9]+]]:vgpr_32 = V_AND_B32_e64 [[COPY1]], 1, implicit $exec
+  ; GCN:   [[V_AND_B32_e64_1:%[0-9]+]]:vgpr_32 = V_AND_B32_e64 [[COPY2]], 0, implicit $exec
+  ; GCN:   [[REG_SEQUENCE1:%[0-9]+]]:vreg_64 = REG_SEQUENCE [[V_AND_B32_e64_]], %subreg.sub0, [[V_AND_B32_e64_1]], %subreg.sub1
+  ; GCN:   V_CMP_NE_U64_e32 0, [[REG_SEQUENCE1]], implicit-def $vcc, implicit $exec
+  ; GCN:   [[COPY3:%[0-9]+]]:vreg_64 = COPY [[REG_SEQUENCE1]], implicit $exec
+  ; GCN:   $vcc = S_AND_B64 $exec, $vcc, implicit-def $scc
+  ; GCN:   S_CBRANCH_VCCNZ %bb.1, implicit $vcc
+  ; GCN: bb.2:
+  ; GCN:   S_ENDPGM 0
+  bb.0:
+    liveins: $vgpr0_vgpr1
+      %0:sreg_64 = COPY $vgpr0_vgpr1
+
+  bb.1:
+      %1:sreg_64 = PHI %0, %bb.0, %2, %bb.1
+      %4:vgpr_32 = GLOBAL_LOAD_DWORD_SADDR %1, undef %3:vgpr_32, 0, 0, implicit $exec
+      %2:sreg_64 = S_AND_B64 %1, 1, implicit-def $scc
+      S_CMP_LG_U64 %2, 0, implicit-def $scc
+      S_CBRANCH_SCC1 %bb.1, implicit $scc
+
+  bb.2:
+      S_ENDPGM 0
+...
+
+---
+name:            global_store_saddr_to_valu
+tracksRegLiveness: true
+body:             |
+  ; GCN-LABEL: name: global_store_saddr_to_valu
+  ; GCN: bb.0:
+  ; GCN:   successors: %bb.1(0x80000000)
+  ; GCN:   liveins: $vgpr0_vgpr1
+  ; GCN:   [[COPY:%[0-9]+]]:vreg_64 = COPY $vgpr0_vgpr1
+  ; GCN: bb.1:
+  ; GCN:   successors: %bb.1(0x40000000), %bb.2(0x40000000)
+  ; GCN:   [[PHI:%[0-9]+]]:vreg_64 = PHI [[COPY]], %bb.0, %7, %bb.1
+  ; GCN:   [[DEF:%[0-9]+]]:vgpr_32 = IMPLICIT_DEF
+  ; GCN:   GLOBAL_STORE_DWORD [[PHI]], [[DEF]], 0, 0, implicit $exec
+  ; GCN:   [[COPY1:%[0-9]+]]:vgpr_32 = COPY [[PHI]].sub0
+  ; GCN:   [[COPY2:%[0-9]+]]:vgpr_32 = COPY [[PHI]].sub1
+  ; GCN:   [[V_AND_B32_e64_:%[0-9]+]]:vgpr_32 = V_AND_B32_e64 [[COPY1]], 1, implicit $exec
+  ; GCN:   [[V_AND_B32_e64_1:%[0-9]+]]:vgpr_32 = V_AND_B32_e64 [[COPY2]], 0, implicit $exec
+  ; GCN:   [[REG_SEQUENCE:%[0-9]+]]:vreg_64 = REG_SEQUENCE [[V_AND_B32_e64_]], %subreg.sub0, [[V_AND_B32_e64_1]], %subreg.sub1
+  ; GCN:   V_CMP_NE_U64_e32 0, [[REG_SEQUENCE]], implicit-def $vcc, implicit $exec
+  ; GCN:   [[COPY3:%[0-9]+]]:vreg_64 = COPY [[REG_SEQUENCE]], implicit $exec
+  ; GCN:   $vcc = S_AND_B64 $exec, $vcc, implicit-def $scc
+  ; GCN:   S_CBRANCH_VCCNZ %bb.1, implicit $vcc
+  ; GCN: bb.2:
+  ; GCN:   S_ENDPGM 0
+  bb.0:
+    liveins: $vgpr0_vgpr1
+      %0:sreg_64 = COPY $vgpr0_vgpr1
+
+  bb.1:
+      %1:sreg_64 = PHI %0, %bb.0, %2, %bb.1
+      %3:vgpr_32 = V_MOV_B32_e32 0, implicit $exec
+      %4:vgpr_32 = IMPLICIT_DEF
+      GLOBAL_STORE_DWORD_SADDR %3, %4, %1, 0, 0, implicit $exec
+      %2:sreg_64 = S_AND_B64 %1, 1, implicit-def $scc
+      S_CMP_LG_U64 %2, 0, implicit-def $scc
+      S_CBRANCH_SCC1 %bb.1, implicit $scc
+
+  bb.2:
+      S_ENDPGM 0
+...
+
+---
+name:            global_addtid_load_saddr_to_valu
+tracksRegLiveness: true
+body:             |
+  ; GCN-LABEL: name: global_addtid_load_saddr_to_valu
+  ; GCN: bb.0:
+  ; GCN:   successors: %bb.1(0x80000000)
+  ; GCN:   liveins: $vgpr0_vgpr1
+  ; GCN:   [[COPY:%[0-9]+]]:vreg_64 = COPY $vgpr0_vgpr1
+  ; GCN: bb.1:
+  ; GCN:   successors: %bb.1(0x40000000), %bb.2(0x40000000)
+  ; GCN:   [[PHI:%[0-9]+]]:vreg_64 = PHI [[COPY]], %bb.0, %6, %bb.1
+  ; GCN:   [[V_READFIRSTLANE_B32_:%[0-9]+]]:sgpr_32 = V_READFIRSTLANE_B32 [[PHI]].sub0, implicit $exec
+  ; GCN:   [[V_READFIRSTLANE_B32_1:%[0-9]+]]:sgpr_32 = V_READFIRSTLANE_B32 [[PHI]].sub1, implicit $exec
+  ; GCN:   [[REG_SEQUENCE:%[0-9]+]]:sreg_64 = REG_SEQUENCE [[V_READFIRSTLANE_B32_]], %subreg.sub0, [[V_READFIRSTLANE_B32_1]], %subreg.sub1
+  ; GCN:   [[GLOBAL_LOAD_DWORD_ADDTID_SADDR:%[0-9]+]]:vgpr_32 = GLOBAL_LOAD_DWORD_ADDTID_SADDR [[REG_SEQUENCE]], 0, 0, implicit $exec
+  ; GCN:   [[COPY1:%[0-9]+]]:vgpr_32 = COPY [[PHI]].sub0
+  ; GCN:   [[COPY2:%[0-9]+]]:vgpr_32 = COPY [[PHI]].sub1
+  ; GCN:   [[V_AND_B32_e64_:%[0-9]+]]:vgpr_32 = V_AND_B32_e64 [[COPY1]], 1, implicit $exec
+  ; GCN:   [[V_AND_B32_e64_1:%[0-9]+]]:vgpr_32 = V_AND_B32_e64 [[COPY2]], 0, implicit $exec
+  ; GCN:   [[REG_SEQUENCE1:%[0-9]+]]:vreg_64 = REG_SEQUENCE [[V_AND_B32_e64_]], %subreg.sub0, [[V_AND_B32_e64_1]], %subreg.sub1
+  ; GCN:   V_CMP_NE_U64_e32 0, [[REG_SEQUENCE1]], implicit-def $vcc, implicit $exec
+  ; GCN:   [[COPY3:%[0-9]+]]:vreg_64 = COPY [[REG_SEQUENCE1]], implicit $exec
+  ; GCN:   $vcc = S_AND_B64 $exec, $vcc, implicit-def $scc
+  ; GCN:   S_CBRANCH_VCCNZ %bb.1, implicit $vcc
+  ; GCN: bb.2:
+  ; GCN:   S_ENDPGM 0
+  bb.0:
+    liveins: $vgpr0_vgpr1
+      %0:sreg_64 = COPY $vgpr0_vgpr1
+
+  bb.1:
+      %1:sreg_64 = PHI %0, %bb.0, %2, %bb.1
+      %4:vgpr_32 = GLOBAL_LOAD_DWORD_ADDTID_SADDR %1, 0, 0, implicit $exec
+      %2:sreg_64 = S_AND_B64 %1, 1, implicit-def $scc
+      S_CMP_LG_U64 %2, 0, implicit-def $scc
+      S_CBRANCH_SCC1 %bb.1, implicit $scc
+
+  bb.2:
+      S_ENDPGM 0
+...
+
+---
+name:            global_store_addtid_saddr_to_valu
+tracksRegLiveness: true
+body:             |
+  ; GCN-LABEL: name: global_store_addtid_saddr_to_valu
+  ; GCN: bb.0:
+  ; GCN:   successors: %bb.1(0x80000000)
+  ; GCN:   liveins: $vgpr0_vgpr1
+  ; GCN:   [[COPY:%[0-9]+]]:vreg_64 = COPY $vgpr0_vgpr1
+  ; GCN: bb.1:
+  ; GCN:   successors: %bb.1(0x40000000), %bb.2(0x40000000)
+  ; GCN:   [[PHI:%[0-9]+]]:vreg_64 = PHI [[COPY]], %bb.0, %6, %bb.1
+  ; GCN:   [[DEF:%[0-9]+]]:vgpr_32 = IMPLICIT_DEF
+  ; GCN:   [[V_READFIRSTLANE_B32_:%[0-9]+]]:sgpr_32 = V_READFIRSTLANE_B32 [[PHI]].sub0, implicit $exec
+  ; GCN:   [[V_READFIRSTLANE_B32_1:%[0-9]+]]:sgpr_32 = V_READFIRSTLANE_B32 [[PHI]].sub1, implicit $exec
+  ; GCN:   [[REG_SEQUENCE:%[0-9]+]]:sreg_64 = REG_SEQUENCE [[V_READFIRSTLANE_B32_]], %subreg.sub0, [[V_READFIRSTLANE_B32_1]], %subreg.sub1
+  ; GCN:   GLOBAL_STORE_DWORD_ADDTID_SADDR [[DEF]], [[REG_SEQUENCE]], 0, 0, implicit $exec
+  ; GCN:   [[COPY1:%[0-9]+]]:vgpr_32 = COPY [[PHI]].sub0
+  ; GCN:   [[COPY2:%[0-9]+]]:vgpr_32 = COPY [[PHI]].sub1
+  ; GCN:   [[V_AND_B32_e64_:%[0-9]+]]:vgpr_32 = V_AND_B32_e64 [[COPY1]], 1, implicit $exec
+  ; GCN:   [[V_AND_B32_e64_1:%[0-9]+]]:vgpr_32 = V_AND_B32_e64 [[COPY2]], 0, implicit $exec
+  ; GCN:   [[REG_SEQUENCE1:%[0-9]+]]:vreg_64 = REG_SEQUENCE [[V_AND_B32_e64_]], %subreg.sub0, [[V_AND_B32_e64_1]], %subreg.sub1
+  ; GCN:   V_CMP_NE_U64_e32 0, [[REG_SEQUENCE1]], implicit-def $vcc, implicit $exec
+  ; GCN:   [[COPY3:%[0-9]+]]:vreg_64 = COPY [[REG_SEQUENCE1]], implicit $exec
+  ; GCN:   $vcc = S_AND_B64 $exec, $vcc, implicit-def $scc
+  ; GCN:   S_CBRANCH_VCCNZ %bb.1, implicit $vcc
+  ; GCN: bb.2:
+  ; GCN:   S_ENDPGM 0
+  bb.0:
+    liveins: $vgpr0_vgpr1
+      %0:sreg_64 = COPY $vgpr0_vgpr1
+
+  bb.1:
+      %1:sreg_64 = PHI %0, %bb.0, %2, %bb.1
+      %4:vgpr_32 = IMPLICIT_DEF
+      GLOBAL_STORE_DWORD_ADDTID_SADDR %4, %1, 0, 0, implicit $exec
+      %2:sreg_64 = S_AND_B64 %1, 1, implicit-def $scc
+      S_CMP_LG_U64 %2, 0, implicit-def $scc
+      S_CBRANCH_SCC1 %bb.1, implicit $scc
+
+  bb.2:
+      S_ENDPGM 0
+...
+
+---
+name:            global_atomic_noret_saddr_to_valu
+tracksRegLiveness: true
+body:             |
+  ; GCN-LABEL: name: global_atomic_noret_saddr_to_valu
+  ; GCN: bb.0:
+  ; GCN:   successors: %bb.1(0x80000000)
+  ; GCN:   liveins: $vgpr0_vgpr1
+  ; GCN:   [[COPY:%[0-9]+]]:vreg_64 = COPY $vgpr0_vgpr1
+  ; GCN: bb.1:
+  ; GCN:   successors: %bb.1(0x40000000), %bb.2(0x40000000)
+  ; GCN:   [[PHI:%[0-9]+]]:vreg_64 = PHI [[COPY]], %bb.0, %6, %bb.1
+  ; GCN:   [[V_MOV_B32_e32_:%[0-9]+]]:vgpr_32 = V_MOV_B32_e32 0, implicit $exec
+  ; GCN:   GLOBAL_ATOMIC_ADD [[PHI]], [[V_MOV_B32_e32_]], 0, 0, implicit $exec
+  ; GCN:   [[COPY1:%[0-9]+]]:vgpr_32 = COPY [[PHI]].sub0
+  ; GCN:   [[COPY2:%[0-9]+]]:vgpr_32 = COPY [[PHI]].sub1
+  ; GCN:   [[V_AND_B32_e64_:%[0-9]+]]:vgpr_32 = V_AND_B32_e64 [[COPY1]], 1, implicit $exec
+  ; GCN:   [[V_AND_B32_e64_1:%[0-9]+]]:vgpr_32 = V_AND_B32_e64 [[COPY2]], 0, implicit $exec
+  ; GCN:   [[REG_SEQUENCE:%[0-9]+]]:vreg_64 = REG_SEQUENCE [[V_AND_B32_e64_]], %subreg.sub0, [[V_AND_B32_e64_1]], %subreg.sub1
+  ; GCN:   V_CMP_NE_U64_e32 0, [[REG_SEQUENCE]], implicit-def $vcc, implicit $exec
+  ; GCN:   [[COPY3:%[0-9]+]]:vreg_64 = COPY [[REG_SEQUENCE]], implicit $exec
+  ; GCN:   $vcc = S_AND_B64 $exec, $vcc, implicit-def $scc
+  ; GCN:   S_CBRANCH_VCCNZ %bb.1, implicit $vcc
+  ; GCN: bb.2:
+  ; GCN:   S_ENDPGM 0
+  bb.0:
+    liveins: $vgpr0_vgpr1
+      %0:sreg_64 = COPY $vgpr0_vgpr1
+
+  bb.1:
+      %1:sreg_64 = PHI %0, %bb.0, %2, %bb.1
+      %3:vgpr_32 = V_MOV_B32_e32 0, implicit $exec
+      GLOBAL_ATOMIC_ADD_SADDR %3, %3, %1, 0, 0, implicit $exec
+      %2:sreg_64 = S_AND_B64 %1, 1, implicit-def $scc
+      S_CMP_LG_U64 %2, 0, implicit-def $scc
+      S_CBRANCH_SCC1 %bb.1, implicit $scc
+
+  bb.2:
+      S_ENDPGM 0
+...
+
+---
+name:            global_atomic_rtn_saddr_to_valu
+tracksRegLiveness: true
+body:             |
+  ; GCN-LABEL: name: global_atomic_rtn_saddr_to_valu
+  ; GCN: bb.0:
+  ; GCN:   successors: %bb.1(0x80000000)
+  ; GCN:   liveins: $vgpr0_vgpr1
+  ; GCN:   [[COPY:%[0-9]+]]:vreg_64 = COPY $vgpr0_vgpr1
+  ; GCN: bb.1:
+  ; GCN:   successors: %bb.1(0x40000000), %bb.2(0x40000000)
+  ; GCN:   [[PHI:%[0-9]+]]:vreg_64 = PHI [[COPY]], %bb.0, %7, %bb.1
+  ; GCN:   [[V_MOV_B32_e32_:%[0-9]+]]:vgpr_32 = V_MOV_B32_e32 0, implicit $exec
+  ; GCN:   [[GLOBAL_ATOMIC_ADD_RTN:%[0-9]+]]:vgpr_32 = GLOBAL_ATOMIC_ADD_RTN [[PHI]], [[V_MOV_B32_e32_]], 0, 0, implicit $exec
+  ; GCN:   [[COPY1:%[0-9]+]]:vgpr_32 = COPY [[PHI]].sub0
+  ; GCN:   [[COPY2:%[0-9]+]]:vgpr_32 = COPY [[PHI]].sub1
+  ; GCN:   [[V_AND_B32_e64_:%[0-9]+]]:vgpr_32 = V_AND_B32_e64 [[COPY1]], 1, implicit $exec
+  ; GCN:   [[V_AND_B32_e64_1:%[0-9]+]]:vgpr_32 = V_AND_B32_e64 [[COPY2]], 0, implicit $exec
+  ; GCN:   [[REG_SEQUENCE:%[0-9]+]]:vreg_64 = REG_SEQUENCE [[V_AND_B32_e64_]], %subreg.sub0, [[V_AND_B32_e64_1]], %subreg.sub1
+  ; GCN:   V_CMP_NE_U64_e32 0, [[REG_SEQUENCE]], implicit-def $vcc, implicit $exec
+  ; GCN:   [[COPY3:%[0-9]+]]:vreg_64 = COPY [[REG_SEQUENCE]], implicit $exec
+  ; GCN:   $vcc = S_AND_B64 $exec, $vcc, implicit-def $scc
+  ; GCN:   S_CBRANCH_VCCNZ %bb.1, implicit $vcc
+  ; GCN: bb.2:
+  ; GCN:   S_ENDPGM 0
+  bb.0:
+    liveins: $vgpr0_vgpr1
+      %0:sreg_64 = COPY $vgpr0_vgpr1
+
+  bb.1:
+      %1:sreg_64 = PHI %0, %bb.0, %2, %bb.1
+      %3:vgpr_32 = V_MOV_B32_e32 0, implicit $exec
+      %4:vgpr_32 = GLOBAL_ATOMIC_ADD_SADDR_RTN %3, %3, %1, 0, 0, implicit $exec
+      %2:sreg_64 = S_AND_B64 %1, 1, implicit-def $scc
+      S_CMP_LG_U64 %2, 0, implicit-def $scc
+      S_CBRANCH_SCC1 %bb.1, implicit $scc
+
+  bb.2:
+      S_ENDPGM 0
+...
+
+---
+name:            scratch_load_saddr_to_valu
+tracksRegLiveness: true
+body:             |
+  ; GCN-LABEL: name: scratch_load_saddr_to_valu
+  ; GCN: bb.0:
+  ; GCN:   successors: %bb.1(0x80000000)
+  ; GCN:   liveins: $vgpr0
+  ; GCN:   [[COPY:%[0-9]+]]:vgpr_32 = COPY $vgpr0
+  ; GCN: bb.1:
+  ; GCN:   successors: %bb.1(0x40000000), %bb.2(0x40000000)
+  ; GCN:   [[PHI:%[0-9]+]]:vgpr_32 = PHI [[COPY]], %bb.0, %6, %bb.1
+  ; GCN:   [[V_READFIRSTLANE_B32_:%[0-9]+]]:sgpr_32 = V_READFIRSTLANE_B32 [[PHI]], implicit $exec
+  ; GCN:   [[SCRATCH_LOAD_DWORD_SADDR:%[0-9]+]]:vgpr_32 = SCRATCH_LOAD_DWORD_SADDR [[V_READFIRSTLANE_B32_]], 0, 0, implicit $exec, implicit $flat_scr
+  ; GCN:   [[V_AND_B32_e64_:%[0-9]+]]:vgpr_32 = V_AND_B32_e64 [[PHI]], 1, implicit $exec
+  ; GCN:   V_CMP_NE_U32_e32 0, [[V_AND_B32_e64_]], implicit-def $vcc, implicit $exec
+  ; GCN:   [[COPY1:%[0-9]+]]:vgpr_32 = COPY [[V_AND_B32_e64_]], implicit $exec
+  ; GCN:   $vcc = S_AND_B64 $exec, $vcc, implicit-def $scc
+  ; GCN:   S_CBRANCH_VCCNZ %bb.1, implicit $vcc
+  ; GCN: bb.2:
+  ; GCN:   S_ENDPGM 0
+  bb.0:
+    liveins: $vgpr0
+      %0:sgpr_32 = COPY $vgpr0
+
+  bb.1:
+      %1:sgpr_32 = PHI %0, %bb.0, %2, %bb.1
+      %4:vgpr_32 = SCRATCH_LOAD_DWORD_SADDR %1, 0, 0, implicit $exec, implicit $flat_scr
+      %2:sgpr_32 = S_AND_B32 %1, 1, implicit-def $scc
+      S_CMP_LG_U32 %2, 0, implicit-def $scc
+      S_CBRANCH_SCC1 %bb.1, implicit $scc
+
+  bb.2:
+      S_ENDPGM 0
+...
+
+---
+name:            scratch_store_saddr_to_valu
+tracksRegLiveness: true
+body:             |
+  ; GCN-LABEL: name: scratch_store_saddr_to_valu
+  ; GCN: bb.0:
+  ; GCN:   successors: %bb.1(0x80000000)
+  ; GCN:   liveins: $vgpr0
+  ; GCN:   [[COPY:%[0-9]+]]:vgpr_32 = COPY $vgpr0
+  ; GCN: bb.1:
+  ; GCN:   successors: %bb.1(0x40000000), %bb.2(0x40000000)
+  ; GCN:   [[PHI:%[0-9]+]]:vgpr_32 = PHI [[COPY]], %bb.0, %6, %bb.1
+  ; GCN:   [[DEF:%[0-9]+]]:vgpr_32 = IMPLICIT_DEF
+  ; GCN:   [[V_READFIRSTLANE_B32_:%[0-9]+]]:sgpr_32 = V_READFIRSTLANE_B32 [[PHI]], implicit $exec
+  ; GCN:   SCRATCH_STORE_DWORD_SADDR [[DEF]], [[V_READFIRSTLANE_B32_]], 0, 0, implicit $exec, implicit $flat_scr
+  ; GCN:   [[V_AND_B32_e64_:%[0-9]+]]:vgpr_32 = V_AND_B32_e64 [[PHI]], 1, implicit $exec
+  ; GCN:   V_CMP_NE_U32_e32 0, [[V_AND_B32_e64_]], implicit-def $vcc, implicit $exec
+  ; GCN:   [[COPY1:%[0-9]+]]:vgpr_32 = COPY [[V_AND_B32_e64_]], implicit $exec
+  ; GCN:   $vcc = S_AND_B64 $exec, $vcc, implicit-def $scc
+  ; GCN:   S_CBRANCH_VCCNZ %bb.1, implicit $vcc
+  ; GCN: bb.2:
+  ; GCN:   S_ENDPGM 0
+  bb.0:
+    liveins: $vgpr0
+      %0:sgpr_32 = COPY $vgpr0
+
+  bb.1:
+      %1:sgpr_32 = PHI %0, %bb.0, %2, %bb.1
+      %4:vgpr_32 = IMPLICIT_DEF
+      SCRATCH_STORE_DWORD_SADDR %4, %1, 0, 0, implicit $exec, implicit $flat_scr
+      %2:sgpr_32 = S_AND_B32 %1, 1, implicit-def $scc
+      S_CMP_LG_U32 %2, 0, implicit-def $scc
+      S_CBRANCH_SCC1 %bb.1, implicit $scc
+
+  bb.2:
+      S_ENDPGM 0
+...