[AMDGPU] Add support for Whole Wavefront Mode

[cwabbott]

Summary

Whole Wavefront Wode (WWM) is similar to WQM, except that all of the
lanes are always enabled, regardless of control flow. This is required
for implementing wavefront reductions in non-uniform control flow, where
we need to use the inactive lanes to propagate intermediate results, so
they need to be enabled. We need to propagate WWM to uses (unless
they're explicitly marked as exact) so that they also propagate
intermediate results correctly. We do the analysis and exec mask munging
during the WQM pass, since there are interactions with WQM for things
that require both WQM and WWM. For simplicity, WWM is entirely
block-local -- blocks are never WWM on entry or exit of a block, and WWM
is not propagated to the block level. This means that computations
involving WWM cannot involve control flow, but we only ever plan to use
WWM for a few limited purposes (none of which involve control flow)
anyways.

Shaders can ask for WWM using the @llvm.amdgcn.wwm intrinsic. There
isn't yet a way to turn WWM off -- that will be added in a future
change.

Finally, it turns out that turning on inactive lanes causes a number of
problems with register allocation. While the best long-term solution
seems like teaching LLVM's register allocator about predication, for now
we need to add some hacks to prevent ourselves from getting into trouble
due to constraints that aren't currently expressed in LLVM. For the gory
details, see the comments at the top of SIFixWWMLiveness.cpp.

[tpr]

Looks good to me, but will leave to Nicolai to review.

[nhaehnle]

Looks mostly good, but I do have a couple of comments.

lib/Target/AMDGPU/SIFixWWMLiveness.cpp

27

*don't

101–102

Duplicate function call.

119

Upper case variable names.

169

Variable names are upper case in LLVM style.

184

Variable names are upper case in LLVM style.

lib/Target/AMDGPU/SIWholeQuadMode.cpp

146

LLVM style is upper-case variable names.

623–629

This could be an S_OR_SAVEEXEC_B64.

802–803

This should be simpler with MI->setDesc on the previous patch :)

test/CodeGen/AMDGPU/wqm.ll

127–185

Maybe I have jetlag blindness, but those two tests seem to be the same.

Also, there should be a test where the wwm computation does use some value from the predecessor block (e.g. use %hi as the offset in the load).

[cwabbott]

test/CodeGen/AMDGPU/wqm.ll

127–185

They're not, since the second one uses lacks the iadd after the llvm.amdgcn.wwm intrinsic. Instead, the intrinsic result is passed directly to the phi. Without the early clobber on the WWM intrinsic, we would coalesce the resulting copy with the other phi sources, and the WWM write would overwrite the inactive channels that should be 0. There's a similar thing going on with the above, but this is easier to verify in a less fragile way -- we just need to make sure that a v_mov_b32 gets emitted after the WWM computation. I'm currently jetlagged too, but I think that's why I added this test.

Good point about the additional test though.

[cwabbott]

lib/Target/AMDGPU/SIWholeQuadMode.cpp

623–629

This is a little tricky, since S_OR_SAVEEXEC_B64 clobbers SCC while the S_MOV_B64 doesn't, so we need to be more careful about where we put it. But I didn't like how I was handling prepareInsertion() in the face of WWM anyways, so I've refactored it to make using S_OR_SAVEEXEC_B64 possible.

[cwabbott]

• Fix style issues.
• Rebased on using setDesc() in lowerCopyInstrs()
• Reworked logic for determining safe points to transition where SCC isn't live, to make using s_or_saveexec_b64 for entering WWM possible.
• Added tests for WWM across multiple basic blocks and correct SCC handling. For the latter, I couldn't figure out how to make an IR test that exercised that path, so I made a MIR test for it.

[nhaehnle]

One remaining style nitpick. LGTM with that fixed.

lib/Target/AMDGPU/SIFixWWMLiveness.cpp

184

Gentle reminder :)

test/CodeGen/AMDGPU/wqm.ll

127–185

Wow, that's subtle stuff. Thanks for the explanation!

[arsenm]

lib/Target/AMDGPU/SIWholeQuadMode.cpp

623–629

I think this pass should probably not be producing the save exec instructions directly. We re-form these later and this may be a problem with -O0

633

Formatting

test/CodeGen/AMDGPU/wqm.ll

104

An -O0 run line would be nice

[arsenm]

lib/Target/AMDGPU/SIISelLowering.cpp

3300–3301

Shouldn't this have a chain?

test/CodeGen/AMDGPU/wqm.ll

123

Need some additional tests with wwm using different types

[arsenm]

lib/Target/AMDGPU/SIInstructions.td

133

I think you should keep the out name to match others, i.e. $sdst

lib/Target/AMDGPU/SIWholeQuadMode.cpp

713

Why remove this? It should definitely exit early for compute

test/CodeGen/AMDGPU/wqm.mir

36–55

You can delete all of this

[cwabbott]

lib/Target/AMDGPU/SIISelLowering.cpp

3300–3301

No, because it's really just a copy instruction that happens to also enable WWM for its sources -- it doesn't have any side-effects.

lib/Target/AMDGPU/SIWholeQuadMode.cpp

623–629

The existing pass already uses S_AND_SAVEEXEC_B64, so I don't see a reason not to use it here; I was just trying to be consistent. If we want to let that pass to handle it, it should probably be a separate change. Also, because of that, any problem with -O0 is probably an already-existing bug (although I doubt this pass got much testing with -O0 before), so I don't want to block anything on that. I can try adding an -O0 line to the test though.

713

Because OpenGL compute shaders (and geometry shaders, and tesselation shaders etc.) can all use WWM for doing non-uniform wavefront reductions, so we at least need to scan the program for WWM instructions. ROCm will probably want it at some point too.

[cwabbott]

test/CodeGen/AMDGPU/wqm.ll

104

I tried adding that, but one of the pre-existing functions in the test assert-fails. Something about it wanting to use scratch with -O0, but it isn't set up correctly. I'd be inclined to punt on this, since it's not related to WWM.

[cwabbott]

• fix style issues
• add test for WWM with integers
• remove unneccesary stuff from MIR test

[cwabbott]

• rename dest of EXIT_WWM to be consistent

[cwabbott]

Fix one last style issue.

[cwabbott]

Remove unneccesary stuff from fix-wwm-liveness.mir, based on Matt's comment for wqm.mir.

include/llvm/IR/IntrinsicsAMDGPU.td

 
 
 // Copies the source value to the destination value, with the guarantee that
 // the source value is computed as if the entire program were executed in WQM.
 def int_amdgcn_wqm : Intrinsic<[llvm_any_ty],
   [LLVMMatchType<0>], [IntrNoMem, IntrSpeculatable]
 >;
 
+// Copies the active channels of the source value to the destination value,
+// with the guarantee that the source value is computed as if the entire
+// program were executed in Whole Wavefront Mode, i.e. with all channels
+// enabled, with a few exceptions: - Phi nodes with require WWM return an
+// undefined value.
+def int_amdgcn_wwm : Intrinsic<[llvm_any_ty],
+  [LLVMMatchType<0>], [IntrNoMem, IntrSpeculatable]
+>;
+
 //===----------------------------------------------------------------------===//
 // CI+ Intrinsics
 //===----------------------------------------------------------------------===//
 
 def int_amdgcn_s_dcache_inv_vol :
   GCCBuiltin<"__builtin_amdgcn_s_dcache_inv_vol">,
   Intrinsic<[], [], []>;
 

lib/Target/AMDGPU/AMDGPU.h

 FunctionPass *createSIShrinkInstructionsPass();
 FunctionPass *createSILoadStoreOptimizerPass();
 FunctionPass *createSIWholeQuadModePass();
 FunctionPass *createSIFixControlFlowLiveIntervalsPass();
 FunctionPass *createSIFixSGPRCopiesPass();
 FunctionPass *createSIDebuggerInsertNopsPass();
 FunctionPass *createSIInsertWaitsPass();
 FunctionPass *createSIInsertWaitcntsPass();
+FunctionPass *createSIFixWWMLivenessPass();
 FunctionPass *createAMDGPUCodeGenPreparePass();
 FunctionPass *createAMDGPUMachineCFGStructurizerPass();
 
 void initializeAMDGPUMachineCFGStructurizerPass(PassRegistry&);
 extern char &AMDGPUMachineCFGStructurizerID;
 
 void initializeAMDGPUAlwaysInlinePass(PassRegistry&);
 

 extern char &SILowerControlFlowID;
 
 void initializeSIInsertSkipsPass(PassRegistry &);
 extern char &SIInsertSkipsPassID;
 
 void initializeSIOptimizeExecMaskingPass(PassRegistry &);
 extern char &SIOptimizeExecMaskingID;
 
+void initializeSIFixWWMLivenessPass(PassRegistry &);
+extern char &SIFixWWMLivenessID;
+
 // Passes common to R600 and SI
 FunctionPass *createAMDGPUPromoteAlloca();
 void initializeAMDGPUPromoteAllocaPass(PassRegistry&);
 extern char &AMDGPUPromoteAllocaID;
 
 Pass *createAMDGPUStructurizeCFGPass();
 FunctionPass *createAMDGPUISelDag(TargetMachine &TM,
                                   CodeGenOpt::Level OptLevel);

lib/Target/AMDGPU/AMDGPUTargetMachine.cpp

extern "C" void LLVMInitializeAMDGPUTarget() {
   initializeSIAnnotateControlFlowPass(*PR);
   initializeSIInsertWaitsPass(*PR);
   initializeSIInsertWaitcntsPass(*PR);
   initializeSIWholeQuadModePass(*PR);
   initializeSILowerControlFlowPass(*PR);
   initializeSIInsertSkipsPass(*PR);
   initializeSIDebuggerInsertNopsPass(*PR);
   initializeSIOptimizeExecMaskingPass(*PR);
+  initializeSIFixWWMLivenessPass(*PR);
   initializeAMDGPUUnifyDivergentExitNodesPass(*PR);
   initializeAMDGPUAAWrapperPassPass(*PR);
 }
 
 static std::unique_ptr<TargetLoweringObjectFile> createTLOF(const Triple &TT) {
   return llvm::make_unique<AMDGPUTargetObjectFile>();
 }
 
void GCNPassConfig::addFastRegAlloc(FunctionPass *RegAllocPass) {
   // FIXME: We have to disable the verifier here because of PHIElimination +
   // TwoAddressInstructions disabling it.
 
   // This must be run immediately after phi elimination and before
   // TwoAddressInstructions, otherwise the processing of the tied operand of
   // SI_ELSE will introduce a copy of the tied operand source after the else.
   insertPass(&PHIEliminationID, &SILowerControlFlowID, false);
 
+  // This must be run after SILowerControlFlow, since it needs to use the
+  // machine-level CFG, but before register allocation.
+  insertPass(&SILowerControlFlowID, &SIFixWWMLivenessID, false);
+
   TargetPassConfig::addFastRegAlloc(RegAllocPass);
 }
 
 void GCNPassConfig::addOptimizedRegAlloc(FunctionPass *RegAllocPass) {
   // This needs to be run directly before register allocation because earlier
   // passes might recompute live intervals.
   insertPass(&MachineSchedulerID, &SIFixControlFlowLiveIntervalsID);
 
   // This must be run immediately after phi elimination and before
   // TwoAddressInstructions, otherwise the processing of the tied operand of
   // SI_ELSE will introduce a copy of the tied operand source after the else.
   insertPass(&PHIEliminationID, &SILowerControlFlowID, false);
 
+  // This must be run after SILowerControlFlow, since it needs to use the
+  // machine-level CFG, but before register allocation.
+  insertPass(&SILowerControlFlowID, &SIFixWWMLivenessID, false);
+
   TargetPassConfig::addOptimizedRegAlloc(RegAllocPass);
 }
 
 void GCNPassConfig::addPostRegAlloc() {
   addPass(&SIFixVGPRCopiesID);
   addPass(&SIOptimizeExecMaskingID);
   TargetPassConfig::addPostRegAlloc();
 }

lib/Target/AMDGPU/CMakeLists.txt

add_llvm_target(AMDGPUCodeGen
   R600OptimizeVectorRegisters.cpp
   R600Packetizer.cpp
   R600RegisterInfo.cpp
   SIAnnotateControlFlow.cpp
   SIDebuggerInsertNops.cpp
   SIFixControlFlowLiveIntervals.cpp
   SIFixSGPRCopies.cpp
   SIFixVGPRCopies.cpp
+  SIFixWWMLiveness.cpp
   SIFoldOperands.cpp
   SIFrameLowering.cpp
   SIInsertSkips.cpp
   SIInsertWaits.cpp
   SIInsertWaitcnts.cpp
   SIInstrInfo.cpp
   SIISelLowering.cpp
   SILoadStoreOptimizer.cpp

lib/Target/AMDGPU/SIFixSGPRCopies.cpp

for (MachineFunction::iterator BI = MF.begin(), BE = MF.end();
     for (MachineBasicBlock::iterator I = MBB.begin(), E = MBB.end();
          I != E; ++I) {
       MachineInstr &MI = *I;
 
       switch (MI.getOpcode()) {
       default:
         continue;
       case AMDGPU::COPY:
-      case AMDGPU::WQM: {
+      case AMDGPU::WQM:
+      case AMDGPU::WWM: {
         // If the destination register is a physical register there isn't really
         // much we can do to fix this.
         if (!TargetRegisterInfo::isVirtualRegister(MI.getOperand(0).getReg()))
           continue;
 
         const TargetRegisterClass *SrcRC, *DstRC;
         std::tie(SrcRC, DstRC) = getCopyRegClasses(MI, *TRI, MRI);
         if (isVGPRToSGPRCopy(SrcRC, DstRC, *TRI)) {

lib/Target/AMDGPU/SIFixWWMLiveness.cpp

+//===-- SIFixWWMLiveness.cpp - Fix WWM live intervals ---------===//
+//
+//                     The LLVM Compiler Infrastructure
+//
+// This file is distributed under the University of Illinois Open Source
+// License. See LICENSE.TXT for details.
+//
+//===----------------------------------------------------------------------===//
+//
+/// \file
+/// \brief Computations in WWM can overwrite values in inactive channels for
+/// variables that the register allocator thinks are dead. This pass adds fake
+/// uses of those variables to WWM instructions to make sure that they aren't
+/// overwritten.
+///
+/// As an example, consider this snippet:
+/// %vgpr0 = V_MOV_B32_e32 0.0
+/// if (...) {
+///   %vgpr1 = ...
+///   %vgpr2 = WWM %vgpr1<kill>
+///   ... = %vgpr2<kill>
+///   %vgpr0 = V_MOV_B32_e32 1.0
+/// }
+/// ... = %vgpr0
+///
+/// The live intervals of %vgpr0 don't overlap with those of %vgpr1. Normally,
+/// we can safely allocate %vgpr0 and %vgpr1 in the same register, since
+/// writing %vgpr1 would only write to channels that would be clobbered by the
+/// second write to %vgpr0 anyways. But if %vgpr1 is written with WWM enabled,
+/// it would clobber even the inactive channels for which the if-condition is
+/// false, for which %vgpr0 is supposed to be 0. This pass adds an implicit use
+/// of %vgpr0 to the WWM instruction to make sure they aren't allocated to the
+/// same register.
+///
+/// In general, we need to figure out what registers might have their inactive
+/// channels which are eventually used accidentally clobbered by a WWM
+/// instruction. We approximate this using two conditions:
+///
+/// 1. A definition of the variable reaches the WWM instruction.
+/// 2. The variable would be live at the WWM instruction if all its defs were
+/// partial defs (i.e. considered as a use), ignoring normal uses.
+///
+/// If a register matches both conditions, then we add an implicit use of it to
+/// the WWM instruction. Condition #2 is the heart of the matter: every
+/// definition is really a partial definition, since every VALU instruction is
+/// implicitly predicated.  We can usually ignore this, but WWM forces us not
+/// to. Condition #1 prevents false positives if the variable is undefined at
+/// the WWM instruction anyways. This is overly conservative in certain cases,
+/// especially in uniform control flow, but this is a workaround anyways until
+/// LLVM gains the notion of predicated uses and definitions of variables.
+///
+//===----------------------------------------------------------------------===//
+
+#include "AMDGPU.h"
+#include "AMDGPUSubtarget.h"
+#include "SIInstrInfo.h"
+#include "SIRegisterInfo.h"
+#include "llvm/ADT/DepthFirstIterator.h"
+#include "llvm/ADT/SparseBitVector.h"
+#include "llvm/CodeGen/LiveIntervalAnalysis.h"
+#include "llvm/CodeGen/MachineFunctionPass.h"
+#include "llvm/CodeGen/Passes.h"
+#include "llvm/Target/TargetRegisterInfo.h"
+
+using namespace llvm;
+
+#define DEBUG_TYPE "si-fix-wwm-liveness"
+
+namespace {
+
+class SIFixWWMLiveness : public MachineFunctionPass {
+private:
+  LiveIntervals *LIS = nullptr;
+  const SIRegisterInfo *TRI;
+  MachineRegisterInfo *MRI;
+
+public:
+  static char ID;
+
+  SIFixWWMLiveness() : MachineFunctionPass(ID) {
+    initializeSIFixWWMLivenessPass(*PassRegistry::getPassRegistry());
+  }
+
+  bool runOnMachineFunction(MachineFunction &MF) override;
+
+  bool runOnWWMInstruction(MachineInstr &MI);
+
+  void addDefs(const MachineInstr &MI, SparseBitVector<> &set);
+
+  StringRef getPassName() const override { return "SI Fix WWM Liveness"; }
+
+  void getAnalysisUsage(AnalysisUsage &AU) const override {
+    // Should preserve the same set that TwoAddressInstructions does.
+    AU.addPreserved<SlotIndexes>();
+    AU.addPreserved<LiveIntervals>();
+    AU.addPreservedID(LiveVariablesID);
+    AU.addPreservedID(MachineLoopInfoID);
+    AU.addPreservedID(MachineDominatorsID);
+    AU.setPreservesCFG();
+    MachineFunctionPass::getAnalysisUsage(AU);
+  }
+};
+
+} // End anonymous namespace.
+
+INITIALIZE_PASS(SIFixWWMLiveness, DEBUG_TYPE,
+                "SI fix WWM liveness", false, false)
+
+char SIFixWWMLiveness::ID = 0;
+
+char &llvm::SIFixWWMLivenessID = SIFixWWMLiveness::ID;
+
+FunctionPass *llvm::createSIFixWWMLivenessPass() {
+  return new SIFixWWMLiveness();
+}
+
+void SIFixWWMLiveness::addDefs(const MachineInstr &MI, SparseBitVector<> &Regs)
+{
+  for (const MachineOperand &Op : MI.defs()) {
+    if (Op.isReg()) {
+      unsigned Reg = Op.getReg();
+      if (TRI->isVGPR(*MRI, Reg))
+        Regs.set(Reg);
+    }
+  }
+}
+
+bool SIFixWWMLiveness::runOnWWMInstruction(MachineInstr &WWM) {
+  MachineBasicBlock *MBB = WWM.getParent();
+
+  // Compute the registers that are live out of MI by figuring out which defs
+  // are reachable from MI.
+  SparseBitVector<> LiveOut;
+
+  for (auto II = MachineBasicBlock::iterator(WWM), IE =
+       MBB->end(); II != IE; ++II) {
+    addDefs(*II, LiveOut);
+  }
+
+  for (df_iterator<MachineBasicBlock *> I = ++df_begin(MBB),
+                                        E = df_end(MBB);
+       I != E; ++I) {
+    for (const MachineInstr &MI : **I) {
+      addDefs(MI, LiveOut);
+    }
+  }
+
+  // Compute the registers that reach MI.
+  SparseBitVector<> Reachable;
+
+  for (auto II = ++MachineBasicBlock::reverse_iterator(WWM), IE =
+       MBB->rend(); II != IE; ++II) {
+    addDefs(*II, Reachable);
+  }
+
+  for (idf_iterator<MachineBasicBlock *> I = ++idf_begin(MBB),
+                                         E = idf_end(MBB);
+       I != E; ++I) {
+    for (const MachineInstr &MI : **I) {
+      addDefs(MI, Reachable);
+    }
+  }
+
+  // find the intersection, and add implicit uses.
+  LiveOut &= Reachable;
+
+  bool Modified = false;
+  for (unsigned Reg : LiveOut) {
+    WWM.addOperand(MachineOperand::CreateReg(Reg, false, /*isImp=*/true));
+    if (LIS) {
+      // FIXME: is there a better way to update the live interval?
+      LIS->removeInterval(Reg);
+      LIS->createAndComputeVirtRegInterval(Reg);
+    }
+    Modified = true;
+  }
+
+  return Modified;
+}
+
+bool SIFixWWMLiveness::runOnMachineFunction(MachineFunction &MF) {
+  bool modified = false;
+
+  // This doesn't actually need LiveIntervals, but we can preserve them.
+  LIS = getAnalysisIfAvailable<LiveIntervals>();
+
+  const SISubtarget &ST = MF.getSubtarget<SISubtarget>();
+  const SIInstrInfo *TII = ST.getInstrInfo();
+
+  TRI = &TII->getRegisterInfo();
+  MRI = &MF.getRegInfo();
+
+  for (MachineBasicBlock &MBB : MF) {
+    for (MachineInstr &MI : MBB) {
+      if (MI.getOpcode() == AMDGPU::EXIT_WWM) {
+        modified |= runOnWWMInstruction(MI);
+      }
+    }
+  }
+
+  return modified;
+}

lib/Target/AMDGPU/SIISelLowering.cpp

SDValue Node = DAG.getNode(AMDGPUISD::CVT_PKRTZ_F16_F32, DL, MVT::i32,
                                Op.getOperand(1), Op.getOperand(2));
     return DAG.getNode(ISD::BITCAST, DL, VT, Node);
   }
   case Intrinsic::amdgcn_wqm: {
     SDValue Src = Op.getOperand(1);
     return SDValue(DAG.getMachineNode(AMDGPU::WQM, DL, Src.getValueType(), Src),
                    0);
   }
+  case Intrinsic::amdgcn_wwm: {
+    SDValue Src = Op.getOperand(1);
+    return SDValue(DAG.getMachineNode(AMDGPU::WWM, DL, Src.getValueType(), Src),
+                   0);
+  }
   default:
     return Op;
   }
 }
 
 SDValue SITargetLowering::LowerINTRINSIC_W_CHAIN(SDValue Op,
                                                  SelectionDAG &DAG) const {
   unsigned IntrID = cast<ConstantSDNode>(Op.getOperand(1))->getZExtValue();

lib/Target/AMDGPU/SIInstrInfo.cpp

else
       MIB.add(MI.getOperand(2));
 
     Bundler.append(MIB);
     llvm::finalizeBundle(MBB, Bundler.begin());
 
     MI.eraseFromParent();
     break;
   }
+  case AMDGPU::EXIT_WWM: {
+    // This only gets its own opcode so that SIFixWWMLiveness can tell when WWM
+    // is exited.
+    MI.setDesc(get(AMDGPU::S_MOV_B64));
+    break;
+  }
   }
   return true;
 }
 
 bool SIInstrInfo::swapSourceModifiers(MachineInstr &MI,
                                       MachineOperand &Src0,
                                       unsigned Src0OpName,
                                       MachineOperand &Src1,

 unsigned SIInstrInfo::getVALUOp(const MachineInstr &MI) {
   switch (MI.getOpcode()) {
   default: return AMDGPU::INSTRUCTION_LIST_END;
   case AMDGPU::REG_SEQUENCE: return AMDGPU::REG_SEQUENCE;
   case AMDGPU::COPY: return AMDGPU::COPY;
   case AMDGPU::PHI: return AMDGPU::PHI;
   case AMDGPU::INSERT_SUBREG: return AMDGPU::INSERT_SUBREG;
   case AMDGPU::WQM: return AMDGPU::WQM;
+  case AMDGPU::WWM: return AMDGPU::WWM;
   case AMDGPU::S_MOV_B32:
     return MI.getOperand(1).isReg() ?
            AMDGPU::COPY : AMDGPU::V_MOV_B32_e32;
   case AMDGPU::S_ADD_I32:
   case AMDGPU::S_ADD_U32: return AMDGPU::V_ADD_I32_e32;
   case AMDGPU::S_ADDC_U32: return AMDGPU::V_ADDC_U32_e32;
   case AMDGPU::S_SUB_I32:
   case AMDGPU::S_SUB_U32: return AMDGPU::V_SUB_I32_e32;
const TargetRegisterClass *SIInstrInfo::getDestEquivalentVGPRClass(
   // For target instructions, getOpRegClass just returns the virtual register
   // class associated with the operand, so we need to find an equivalent VGPR
   // register class in order to move the instruction to the VALU.
   case AMDGPU::COPY:
   case AMDGPU::PHI:
   case AMDGPU::REG_SEQUENCE:
   case AMDGPU::INSERT_SUBREG:
   case AMDGPU::WQM:
+  case AMDGPU::WWM:
     if (RI.hasVGPRs(NewDstRC))
       return nullptr;
 
     NewDstRC = RI.getEquivalentVGPRClass(NewDstRC);
     if (!NewDstRC)
       return nullptr;
     return NewDstRC;
   default:

lib/Target/AMDGPU/SIInstructions.td

def V_CNDMASK_B64_PSEUDO : VOP3Common <(outs VReg_64:$vdst),
   let usesCustomInserter = 1;
 }
 
 // 64-bit vector move instruction.  This is mainly used by the SIFoldOperands
 // pass to enable folding of inline immediates.
 def V_MOV_B64_PSEUDO : VPseudoInstSI <(outs VReg_64:$vdst),
                                       (ins VSrc_b64:$src0)>;
 
-// Pseudoinstruction for @llvm.amdgcn.wqm. It is turned into a copy
-// after the WQM pass processes them.
+// Pseudoinstruction for @llvm.amdgcn.wqm. It is turned into a copy after the
+// WQM pass processes it.
 def WQM : PseudoInstSI <(outs unknown:$vdst), (ins unknown:$src0)>;
 
+// Pseudoinstruction for @llvm.amdgcn.wwm. It is turned into a copy post-RA, so
+// that the @earlyclobber is respected. The @earlyclobber is to make sure that
+// the instruction that defines $src0 (which is run in WWM) doesn't
+// accidentally clobber inactive channels of $vdst.
+let Constraints = "@earlyclobber $vdst" in {
+def WWM : PseudoInstSI <(outs unknown:$vdst), (ins unknown:$src0)>;
+}
+
 } // End let hasSideEffects = 0, mayLoad = 0, mayStore = 0, Uses = [EXEC]
 
+def EXIT_WWM : SPseudoInstSI <(outs SReg_64:$sdst), (ins SReg_64:$src0)> {
+  let hasSideEffects = 0;
+  let mayLoad = 0;
+  let mayStore = 0;
+}
+
 let usesCustomInserter = 1, SALU = 1 in {
 def GET_GROUPSTATICSIZE : PseudoInstSI <(outs SReg_32:$sdst), (ins),
   [(set SReg_32:$sdst, (int_amdgcn_groupstaticsize))]>;
 } // End let usesCustomInserter = 1, SALU = 1
 
 def S_MOV_B64_term : PseudoInstSI<(outs SReg_64:$dst),
    (ins SSrc_b64:$src0)> {
   let SALU = 1;

lib/Target/AMDGPU/SIWholeQuadMode.cpp

 //===-- SIWholeQuadMode.cpp - enter and suspend whole quad mode -----------===//
 //
 //                     The LLVM Compiler Infrastructure
 //
 // This file is distributed under the University of Illinois Open Source
 // License. See LICENSE.TXT for details.
 //
 //===----------------------------------------------------------------------===//
 //
 /// \file
 /// \brief This pass adds instructions to enable whole quad mode for pixel
-/// shaders.
+/// shaders, and whole wavefront mode for all programs.
 ///
 /// Whole quad mode is required for derivative computations, but it interferes
 /// with shader side effects (stores and atomics). This pass is run on the
 /// scheduled machine IR but before register coalescing, so that machine SSA is
 /// available for analysis. It ensures that WQM is enabled when necessary, but
 /// disabled around stores and atomics.
 ///
 /// When necessary, this pass creates a function prolog
 ///
 ///   S_MOV_B64 LiveMask, EXEC
 ///   S_WQM_B64 EXEC, EXEC
 ///
 /// to enter WQM at the top of the function and surrounds blocks of Exact
 /// instructions by
 ///
 ///   S_AND_SAVEEXEC_B64 Tmp, LiveMask
 ///   ...
 ///   S_MOV_B64 EXEC, Tmp
 ///
+/// We also compute when a sequence of instructions requires Whole Wavefront
+/// Mode (WWM) and insert instructions to save and restore it:
+///
+/// S_OR_SAVEEXEC_B64 Tmp, -1
+/// ...
+/// S_MOV_B64 EXEC, Tmp
+///
 /// In order to avoid excessive switching during sequences of Exact
 /// instructions, the pass first analyzes which instructions must be run in WQM
 /// (aka which instructions produce values that lead to derivative
 /// computations).
 ///
 /// Basic blocks are always exited in WQM as long as some successor needs WQM.
 ///
 /// There is room for improvement given better control flow analysis:

 using namespace llvm;
 
 #define DEBUG_TYPE "si-wqm"
 
 namespace {
 
 enum {
   StateWQM = 0x1,
-  StateExact = 0x2,
+  StateWWM = 0x2,
+  StateExact = 0x4,
 };
 
 struct PrintState {
 public:
   int State;
 
   explicit PrintState(int State) : State(State) {}
 };
 
 static raw_ostream &operator<<(raw_ostream &OS, const PrintState &PS) {
   if (PS.State & StateWQM)
     OS << "WQM";
-  if (PS.State & StateExact) {
+  if (PS.State & StateWWM) {
     if (PS.State & StateWQM)
       OS << '|';
+    OS << "WWM";
+  }
+  if (PS.State & StateExact) {
+    if (PS.State & (StateWQM | StateWWM))
+      OS << '|';
     OS << "Exact";
   }
 
   return OS;
 }
 
 struct InstrInfo {
   char Needs = 0;
struct WorkItem {
 
   WorkItem() = default;
   WorkItem(MachineBasicBlock *MBB) : MBB(MBB) {}
   WorkItem(MachineInstr *MI) : MI(MI) {}
 };
 
 class SIWholeQuadMode : public MachineFunctionPass {
 private:
+  CallingConv::ID CallingConv;
   const SIInstrInfo *TII;
   const SIRegisterInfo *TRI;
   MachineRegisterInfo *MRI;
   LiveIntervals *LIS;
 
   DenseMap<const MachineInstr *, InstrInfo> Instructions;
   DenseMap<MachineBasicBlock *, BlockInfo> Blocks;
   SmallVector<MachineInstr *, 1> LiveMaskQueries;
private:
   MachineBasicBlock::iterator
   prepareInsertion(MachineBasicBlock &MBB, MachineBasicBlock::iterator First,
                    MachineBasicBlock::iterator Last, bool PreferLast,
                    bool SaveSCC);
   void toExact(MachineBasicBlock &MBB, MachineBasicBlock::iterator Before,
                unsigned SaveWQM, unsigned LiveMaskReg);
   void toWQM(MachineBasicBlock &MBB, MachineBasicBlock::iterator Before,
              unsigned SavedWQM);
+  void toWWM(MachineBasicBlock &MBB, MachineBasicBlock::iterator Before,
+             unsigned SaveOrig);
+  void fromWWM(MachineBasicBlock &MBB, MachineBasicBlock::iterator Before,
+               unsigned SavedOrig);
   void processBlock(MachineBasicBlock &MBB, unsigned LiveMaskReg, bool isEntry);
 
   void lowerLiveMaskQueries(unsigned LiveMaskReg);
   void lowerCopyInstrs();
 
 public:
   static char ID;
 
for (const auto &BII : Blocks) {
     }
   }
 }
 
 void SIWholeQuadMode::markInstruction(MachineInstr &MI, char Flag,
                                       std::vector<WorkItem> &Worklist) {
   InstrInfo &II = Instructions[&MI];
 
-  assert(Flag == StateWQM);
+  assert(!(Flag & StateExact) && Flag != 0);
 
   // Remove any disabled states from the flag. The user that required it gets
   // an undefined value in the helper lanes. For example, this can happen if
   // the result of an atomic is used by instruction that requires WQM, where
   // ignoring the request for WQM is correct as per the relevant specs.
   Flag &= ~II.Disabled;
 
   // Ignore if the flag is already encompassed by the existing needs, or we
   // just disabled everything.
   if ((II.Needs & Flag) == Flag)
     return;
 
   II.Needs |= Flag;
   Worklist.push_back(&MI);
 }
 
 /// Mark all instructions defining the uses in \p MI with \p Flag.
 void SIWholeQuadMode::markInstructionUses(const MachineInstr &MI, char Flag,
                                           std::vector<WorkItem> &Worklist) {
-  assert(Flag == StateWQM);
   for (const MachineOperand &Use : MI.uses()) {
     if (!Use.isReg() || !Use.isUse())
       continue;
 
     unsigned Reg = Use.getReg();
 
     // Handle physical registers that we need to track; this is mostly relevant
     // for VCC, which can appear as the (implicit) input of a uniform branch,
for (auto BI = RPOT.begin(), BE = RPOT.end(); BI != BE; ++BI) {
     BlockInfo &BBI = Blocks[&MBB];
 
     for (auto II = MBB.begin(), IE = MBB.end(); II != IE; ++II) {
       MachineInstr &MI = *II;
       InstrInfo &III = Instructions[&MI];
       unsigned Opcode = MI.getOpcode();
       char Flags = 0;
 
-      if (TII->isDS(Opcode)) {
+      if (TII->isDS(Opcode) && CallingConv == CallingConv::AMDGPU_PS) {
         Flags = StateWQM;
       } else if (TII->isWQM(Opcode)) {
         // Sampling instructions don't need to produce results for all pixels
         // in a quad, they just require all inputs of a quad to have been
         // computed for derivatives.
         markInstructionUses(MI, StateWQM, Worklist);
         GlobalFlags |= StateWQM;
         continue;
       } else if (Opcode == AMDGPU::WQM) {
         // The WQM intrinsic requires its output to have all the helper lanes
         // correct, so we need it to be in WQM.
         Flags = StateWQM;
         LowerToCopyInstrs.push_back(&MI);
+      } else if (Opcode == AMDGPU::WWM) {
+        // The WWM intrinsic doesn't make the same guarantee, and plus it needs
+        // to be executed in WQM or Exact so that its copy doesn't clobber
+        // inactive lanes.
+        markInstructionUses(MI, StateWWM, Worklist);
+        GlobalFlags |= StateWWM;
+        LowerToCopyInstrs.push_back(&MI);
+        continue;
       } else if (TII->isDisableWQM(MI)) {
         BBI.Needs |= StateExact;
         if (!(BBI.InNeeds & StateExact)) {
           BBI.InNeeds |= StateExact;
           Worklist.push_back(&MBB);
         }
         GlobalFlags |= StateExact;
-        III.Disabled = StateWQM;
+        III.Disabled = StateWQM | StateWWM;
         continue;
       } else {
         if (Opcode == AMDGPU::SI_PS_LIVE) {
           LiveMaskQueries.push_back(&MI);
         } else if (WQMOutputs) {
           // The function is in machine SSA form, which means that physical
           // VGPRs correspond to shader inputs and outputs. Inputs are
           // only used, outputs are only defined.
if (II.Needs & StateWQM) {
     if (!(BI.InNeeds & StateWQM)) {
       BI.InNeeds |= StateWQM;
       Worklist.push_back(MBB);
     }
   }
 
   // Propagate backwards within block
   if (MachineInstr *PrevMI = MI.getPrevNode()) {
-    char InNeeds = II.Needs | II.OutNeeds;
+    char InNeeds = (II.Needs & ~StateWWM) | II.OutNeeds;
     if (!PrevMI->isPHI()) {
       InstrInfo &PrevII = Instructions[PrevMI];
       if ((PrevII.OutNeeds | InNeeds) != PrevII.OutNeeds) {
         PrevII.OutNeeds |= InNeeds;
         Worklist.push_back(PrevMI);
       }
     }
   }
if (SavedWQM) {
     MI = BuildMI(MBB, Before, DebugLoc(), TII->get(AMDGPU::S_WQM_B64),
                  AMDGPU::EXEC)
              .addReg(AMDGPU::EXEC);
   }
 
   LIS->InsertMachineInstrInMaps(*MI);
 }
 
+void SIWholeQuadMode::toWWM(MachineBasicBlock &MBB,
+                            MachineBasicBlock::iterator Before,
+                            unsigned SaveOrig) {
+  MachineInstr *MI;
+
+  assert(SaveOrig);
+  MI = BuildMI(MBB, Before, DebugLoc(), TII->get(AMDGPU::S_OR_SAVEEXEC_B64),
+               SaveOrig)
+           .addImm(-1);
+  LIS->InsertMachineInstrInMaps(*MI);
+}
+
+void SIWholeQuadMode::fromWWM(MachineBasicBlock &MBB,
+                              MachineBasicBlock::iterator Before,
+                              unsigned SavedOrig) {
+  MachineInstr *MI;
+
+  assert(SavedOrig);
+  MI = BuildMI(MBB, Before, DebugLoc(), TII->get(AMDGPU::EXIT_WWM), AMDGPU::EXEC)
+           .addReg(SavedOrig);
+  LIS->InsertMachineInstrInMaps(*MI);
+}
+
 void SIWholeQuadMode::processBlock(MachineBasicBlock &MBB, unsigned LiveMaskReg,
                                    bool isEntry) {
   auto BII = Blocks.find(&MBB);
   if (BII == Blocks.end())
     return;
 
   const BlockInfo &BI = BII->second;
 
-  if (!(BI.InNeeds & StateWQM))
-    return;
-
   // This is a non-entry block that is WQM throughout, so no need to do
   // anything.
-  if (!isEntry && !(BI.Needs & StateExact) && BI.OutNeeds != StateExact)
+  if (!isEntry && BI.Needs == StateWQM && BI.OutNeeds != StateExact)
     return;
 
   DEBUG(dbgs() << "\nProcessing block BB#" << MBB.getNumber() << ":\n");
 
   unsigned SavedWQMReg = 0;
+  unsigned SavedNonWWMReg = 0;
   bool WQMFromExec = isEntry;
-  char State = isEntry ? StateExact : StateWQM;
+  char State = (isEntry || !(BI.InNeeds & StateWQM)) ? StateExact : StateWQM;
+  char NonWWMState = 0;
 
   auto II = MBB.getFirstNonPHI(), IE = MBB.end();
   if (isEntry)
     ++II; // Skip the instruction that saves LiveMask
 
-  MachineBasicBlock::iterator First = IE;
+  // This stores the first instruction where it's safe to switch from WQM to
+  // Exact or vice versa.
+  MachineBasicBlock::iterator FirstWQM = IE;
+
+  // This stores the first instruction where it's safe to switch from WWM to
+  // Exact/WQM or to switch to WWM. It must always be the same as, or after,
+  // FirstWQM since if it's safe to switch to/from WWM, it must be safe to
+  // switch to/from WQM as well.
+  MachineBasicBlock::iterator FirstWWM = IE;
   for (;;) {
     MachineBasicBlock::iterator Next = II;
-    char Needs = StateExact | StateWQM;
+    char Needs = StateExact | StateWQM; // WWM is disabled by default
     char OutNeeds = 0;
 
-    if (First == IE)
-      First = II;
+    if (FirstWQM == IE)
+      FirstWQM = II;
 
+    if (FirstWWM == IE)
+      FirstWWM = II;
+
+    // First, figure out the allowed states (Needs) based on the propagated
+    // flags.
     if (II != IE) {
       MachineInstr &MI = *II;
 
       if (requiresCorrectState(MI)) {
         auto III = Instructions.find(&MI);
         if (III != Instructions.end()) {
-          if (III->second.Needs & StateWQM)
+          if (III->second.Needs & StateWWM)
+            Needs = StateWWM;
+          else if (III->second.Needs & StateWQM)
             Needs = StateWQM;
           else
             Needs &= ~III->second.Disabled;
           OutNeeds = III->second.OutNeeds;
         }
+      } else {
+        // If the instruction doesn't actually need a correct EXEC, then we can
+        // safely leave WWM enabled.
+        Needs = StateExact | StateWQM | StateWWM;
       }
 
       if (MI.isTerminator() && OutNeeds == StateExact)
         Needs = StateExact;
 
       if (MI.getOpcode() == AMDGPU::SI_ELSE && BI.OutNeeds == StateExact)
         MI.getOperand(3).setImm(1);
 
       ++Next;
     } else {
       // End of basic block
       if (BI.OutNeeds & StateWQM)
         Needs = StateWQM;
       else if (BI.OutNeeds == StateExact)
         Needs = StateExact;
       else
         Needs = StateWQM | StateExact;
     }
 
+    // Now, transition if necessary.
     if (!(Needs & State)) {
+      MachineBasicBlock::iterator First;
+      if (State == StateWWM || Needs == StateWWM) {
+        // We must switch to or from WWM
+        First = FirstWWM;
+      } else {
+        // We only need to switch to/from WQM, so we can use FirstWQM
+        First = FirstWQM;
+      }
+
       MachineBasicBlock::iterator Before =
           prepareInsertion(MBB, First, II, Needs == StateWQM,
                            Needs == StateExact || WQMFromExec);
 
-      if (Needs == StateExact) {
+      if (State == StateWWM) {
+        assert(SavedNonWWMReg);
+        fromWWM(MBB, Before, SavedNonWWMReg);
+        State = NonWWMState;
+      }
+
+      if (Needs == StateWWM) {
+        NonWWMState = State;
+        SavedNonWWMReg = MRI->createVirtualRegister(&AMDGPU::SReg_64RegClass);
+        toWWM(MBB, Before, SavedNonWWMReg);
+        State = StateWWM;
+      } else {
+        if (State == StateWQM && (Needs & StateExact) && !(Needs & StateWQM)) {
           if (!WQMFromExec && (OutNeeds & StateWQM))
             SavedWQMReg = MRI->createVirtualRegister(&AMDGPU::SReg_64RegClass);
 
           toExact(MBB, Before, SavedWQMReg, LiveMaskReg);
           State = StateExact;
-      } else {
-        assert(Needs == StateWQM);
+        } else if (State == StateExact && (Needs & StateWQM) &&
+                   !(Needs & StateExact)) {
           assert(WQMFromExec == (SavedWQMReg == 0));
 
           toWQM(MBB, Before, SavedWQMReg);
 
           if (SavedWQMReg) {
             LIS->createAndComputeVirtRegInterval(SavedWQMReg);
             SavedWQMReg = 0;
           }
           State = StateWQM;
+        } else {
+          // We can get here if we transitioned from WWM to a non-WWM state that
+          // already matches our needs, but we shouldn't need to do anything.
+          assert(Needs & State);
+        }
       }
-
-      First = IE;
     }
 
+    if (Needs != (StateExact | StateWQM | StateWWM)) {
       if (Needs != (StateExact | StateWQM))
-      First = IE;
+        FirstWQM = IE;
+      FirstWWM = IE;
+    }
 
     if (II == IE)
       break;
     II = Next;
   }
 }
 
 void SIWholeQuadMode::lowerLiveMaskQueries(unsigned LiveMaskReg) {
   for (MachineInstr *MI : LiveMaskQueries) {
     const DebugLoc &DL = MI->getDebugLoc();
     unsigned Dest = MI->getOperand(0).getReg();
     MachineInstr *Copy =
         BuildMI(*MI->getParent(), MI, DL, TII->get(AMDGPU::COPY), Dest)
             .addReg(LiveMaskReg);
 
     LIS->ReplaceMachineInstrInMaps(*MI, *Copy);
     MI->eraseFromParent();
   }
 }
 
 void SIWholeQuadMode::lowerCopyInstrs() {
   for (MachineInstr *MI : LowerToCopyInstrs)
     MI->setDesc(TII->get(AMDGPU::COPY));
 }
 
 bool SIWholeQuadMode::runOnMachineFunction(MachineFunction &MF) {
-  if (MF.getFunction()->getCallingConv() != CallingConv::AMDGPU_PS)
-    return false;
-
   Instructions.clear();
   Blocks.clear();
   LiveMaskQueries.clear();
   LowerToCopyInstrs.clear();
+  CallingConv = MF.getFunction()->getCallingConv();
 
   const SISubtarget &ST = MF.getSubtarget<SISubtarget>();
 
   TII = ST.getInstrInfo();
   TRI = &TII->getRegisterInfo();
   MRI = &MF.getRegInfo();
   LIS = &getAnalysis<LiveIntervals>();
 
   char GlobalFlags = analyzeFunction(MF);
+  unsigned LiveMaskReg = 0;
   if (!(GlobalFlags & StateWQM)) {
     lowerLiveMaskQueries(AMDGPU::EXEC);
+    if (!(GlobalFlags & StateWWM))
       return !LiveMaskQueries.empty();
-  }
-
+  } else {
     // Store a copy of the original live mask when required
-  unsigned LiveMaskReg = 0;
-  {
     MachineBasicBlock &Entry = MF.front();
     MachineBasicBlock::iterator EntryMI = Entry.getFirstNonPHI();
 
     if (GlobalFlags & StateExact || !LiveMaskQueries.empty()) {
       LiveMaskReg = MRI->createVirtualRegister(&AMDGPU::SReg_64RegClass);
       MachineInstr *MI = BuildMI(Entry, EntryMI, DebugLoc(),
                                  TII->get(AMDGPU::COPY), LiveMaskReg)
                              .addReg(AMDGPU::EXEC);
       LIS->InsertMachineInstrInMaps(*MI);
     }
 
+    lowerLiveMaskQueries(LiveMaskReg);
+
     if (GlobalFlags == StateWQM) {
       // For a shader that needs only WQM, we can just set it once.
       BuildMI(Entry, EntryMI, DebugLoc(), TII->get(AMDGPU::S_WQM_B64),
               AMDGPU::EXEC)
           .addReg(AMDGPU::EXEC);
 
-      lowerLiveMaskQueries(LiveMaskReg);
       lowerCopyInstrs();
       // EntryMI may become invalid here
       return true;
     }
   }
 
   DEBUG(printInfo());
 
-  lowerLiveMaskQueries(LiveMaskReg);
   lowerCopyInstrs();
 
   // Handle the general case
   for (auto BII : Blocks)
     processBlock(*BII.first, LiveMaskReg, BII.first == &*MF.begin());
 
   // Physical registers like SCC aren't tracked by default anyway, so just
   // removing the ranges we computed is the simplest option for maintaining
   // the analysis results.
   LIS->removeRegUnit(*MCRegUnitIterator(AMDGPU::SCC, TRI));
 
   return true;
 }

test/CodeGen/AMDGPU/fix-wwm-liveness.mir

+# RUN: llc -march=amdgcn -verify-machineinstrs -run-pass si-fix-wwm-liveness -o -  %s | FileCheck %s
+#CHECK: %exec = EXIT_WWM killed %19, implicit %21
+
+---
+name:            test_wwm_liveness
+alignment:       0
+exposesReturnsTwice: false
+legalized:       false
+regBankSelected: false
+selected:        false
+tracksRegLiveness: true
+registers:       
+  - { id: 0, class: sreg_64, preferred-register: '' }
+  - { id: 1, class: sgpr_32, preferred-register: '' }
+  - { id: 2, class: sgpr_32, preferred-register: '' }
+  - { id: 3, class: vgpr_32, preferred-register: '' }
+  - { id: 4, class: vgpr_32, preferred-register: '' }
+  - { id: 5, class: vgpr_32, preferred-register: '' }
+  - { id: 6, class: vgpr_32, preferred-register: '' }
+  - { id: 7, class: vgpr_32, preferred-register: '' }
+  - { id: 8, class: sreg_64, preferred-register: '%vcc' }
+  - { id: 9, class: sreg_64, preferred-register: '' }
+  - { id: 10, class: sreg_32_xm0, preferred-register: '' }
+  - { id: 11, class: sreg_64, preferred-register: '' }
+  - { id: 12, class: sreg_32_xm0, preferred-register: '' }
+  - { id: 13, class: sreg_32_xm0, preferred-register: '' }
+  - { id: 14, class: sreg_32_xm0, preferred-register: '' }
+  - { id: 15, class: sreg_128, preferred-register: '' }
+  - { id: 16, class: vgpr_32, preferred-register: '' }
+  - { id: 17, class: vgpr_32, preferred-register: '' }
+  - { id: 18, class: vgpr_32, preferred-register: '' }
+  - { id: 19, class: sreg_64, preferred-register: '' }
+  - { id: 20, class: sreg_64, preferred-register: '' }
+  - { id: 21, class: vgpr_32, preferred-register: '' }
+  - { id: 22, class: sreg_64, preferred-register: '' }
+  - { id: 23, class: sreg_64, preferred-register: '' }
+liveins:         
+frameInfo:       
+  isFrameAddressTaken: false
+  isReturnAddressTaken: false
+  hasStackMap:     false
+  hasPatchPoint:   false
+  stackSize:       0
+  offsetAdjustment: 0
+  maxAlignment:    0
+  adjustsStack:    false
+  hasCalls:        false
+  stackProtector:  ''
+  maxCallFrameSize: 4294967295
+  hasOpaqueSPAdjustment: false
+  hasVAStart:      false
+  hasMustTailInVarArgFunc: false
+  savePoint:       ''
+  restorePoint:    ''
+fixedStack:      
+stack:           
+constants:       
+body:             |
+  bb.0:
+    successors: %bb.1(0x40000000), %bb.2(0x40000000)
+  
+    %21 = V_MOV_B32_e32 0, implicit %exec
+    %5 = V_MBCNT_LO_U32_B32_e64 -1, 0, implicit %exec
+    %6 = V_MBCNT_HI_U32_B32_e32 -1, killed %5, implicit %exec
+    %8 = V_CMP_GT_U32_e64 32, killed %6, implicit %exec
+    %22 = COPY %exec, implicit-def %exec
+    %23 = S_AND_B64 %22, %8, implicit-def dead %scc
+    %0 = S_XOR_B64 %23, %22, implicit-def dead %scc
+    %exec = S_MOV_B64_term killed %23
+    SI_MASK_BRANCH %bb.2, implicit %exec
+    S_BRANCH %bb.1
+  
+  bb.1:
+    successors: %bb.2(0x80000000)
+  
+    %13 = S_MOV_B32 61440
+    %14 = S_MOV_B32 -1
+    %15 = REG_SEQUENCE undef %12, 1, undef %10, 2, killed %14, 3, killed %13, 4
+    %19 = COPY %exec
+    %exec = S_MOV_B64 -1
+    %16 = BUFFER_LOAD_DWORD_OFFSET %15, 0, 0, 0, 0, 0, implicit %exec :: (volatile load 4)
+    %17 = V_ADD_F32_e32 1065353216, killed %16, implicit %exec
+    %exec = EXIT_WWM killed %19
+    %21 = V_MOV_B32_e32 1, implicit %exec
+    early-clobber %18 = WWM killed %17, implicit %exec
+    BUFFER_STORE_DWORD_OFFSET killed %18, killed %15, 0, 0, 0, 0, 0, implicit %exec :: (store 4)
+  
+  bb.2:
+    %exec = S_OR_B64 %exec, killed %0, implicit-def %scc
+    %vgpr0 = COPY killed %21
+    SI_RETURN_TO_EPILOG killed %vgpr0
+
+...

test/CodeGen/AMDGPU/wqm.ll

 ;CHECK: s_wqm_b64 exec, exec
 ;CHECK: buffer_load_dword
 ;CHECK: buffer_load_dword
 ;CHECK: v_add_f32_e32
 define amdgpu_ps float @test6(i32 inreg %idx0, i32 inreg %idx1) {
 main_body:
   %src0 = call float @llvm.amdgcn.buffer.load.f32(<4 x i32> undef, i32 %idx0, i32 0, i1 0, i1 0)
   %src1 = call float @llvm.amdgcn.buffer.load.f32(<4 x i32> undef, i32 %idx1, i32 0, i1 0, i1 0)
   %out = fadd float %src0, %src1
   %out.0 = bitcast float %out to i32
   %out.1 = call i32 @llvm.amdgcn.wqm.i32(i32 %out.0)
   %out.2 = bitcast i32 %out.1 to float
   ret float %out.2
 }
 
+; Check that WWM is triggered by the wwm intrinsic.
+;
+;CHECK-LABEL: {{^}}test_wwm1:
+;CHECK: s_or_saveexec_b64 s{{\[[0-9]+:[0-9]+\]}}, -1
+;CHECK: buffer_load_dword
+;CHECK: buffer_load_dword
+;CHECK: v_add_f32_e32
+define amdgpu_ps float @test_wwm1(i32 inreg %idx0, i32 inreg %idx1) {
+main_body:
+  %src0 = call float @llvm.amdgcn.buffer.load.f32(<4 x i32> undef, i32 %idx0, i32 0, i1 0, i1 0)
+  %src1 = call float @llvm.amdgcn.buffer.load.f32(<4 x i32> undef, i32 %idx1, i32 0, i1 0, i1 0)
+  %out = fadd float %src0, %src1
+  %out.0 = call float @llvm.amdgcn.wwm.f32(float %out)
+  ret float %out.0
+}
+
+; Same as above, but with an integer type.
+;
+;CHECK-LABEL: {{^}}test_wwm2:
+;CHECK: s_or_saveexec_b64 s{{\[[0-9]+:[0-9]+\]}}, -1
+;CHECK: buffer_load_dword
+;CHECK: buffer_load_dword
+;CHECK: v_add_i32_e32
+define amdgpu_ps float @test_wwm2(i32 inreg %idx0, i32 inreg %idx1) {
+main_body:
+  %src0 = call float @llvm.amdgcn.buffer.load.f32(<4 x i32> undef, i32 %idx0, i32 0, i1 0, i1 0)
+  %src1 = call float @llvm.amdgcn.buffer.load.f32(<4 x i32> undef, i32 %idx1, i32 0, i1 0, i1 0)
+  %src0.0 = bitcast float %src0 to i32
+  %src1.0 = bitcast float %src1 to i32
+  %out = add i32 %src0.0, %src1.0
+  %out.0 = call i32 @llvm.amdgcn.wwm.i32(i32 %out)
+  %out.1 = bitcast i32 %out.0 to float
+  ret float %out.1
+}
+
+; Check that we don't leave WWM on for computations that don't require WWM,
+; since that will lead clobbering things that aren't supposed to be clobbered
+; in cases like this.
+;
+;CHECK-LABEL: {{^}}test_wwm3:
+;CHECK: s_or_saveexec_b64 [[ORIG:s\[[0-9]+:[0-9]+\]]], -1
+;CHECK: buffer_load_dword
+;CHECK: v_add_f32_e32
+;CHECK: s_mov_b64 exec, [[ORIG]]
+;CHECK: v_add_f32_e32
+define amdgpu_ps float @test_wwm3(i32 inreg %idx) {
+main_body:
+  ; use mbcnt to make sure the branch is divergent
+  %lo = call i32 @llvm.amdgcn.mbcnt.lo(i32 -1, i32 0)
+  %hi = call i32 @llvm.amdgcn.mbcnt.hi(i32 -1, i32 %lo)
+  %cc = icmp uge i32 %hi, 32
+  br i1 %cc, label %endif, label %if
+
+if:
+  %src = call float @llvm.amdgcn.buffer.load.f32(<4 x i32> undef, i32 %idx, i32 0, i1 0, i1 0)
+  %out = fadd float %src, %src
+  %out.0 = call float @llvm.amdgcn.wwm.f32(float %out)
+  %out.1 = fadd float %src, %out.0
+  br label %endif
+
+endif:
+  %out.2 = phi float [ %out.1, %if ], [ 0.0, %main_body ]
+  ret float %out.2
+}
+
+; Check that WWM writes aren't coalesced with non-WWM writes, since the WWM
+; write could clobber disabled channels in the non-WWM one.
+;
+;CHECK-LABEL: {{^}}test_wwm4:
+;CHECK: s_or_saveexec_b64 [[ORIG:s\[[0-9]+:[0-9]+\]]], -1
+;CHECK: buffer_load_dword
+;CHECK: v_add_f32_e32
+;CHECK: s_mov_b64 exec, [[ORIG]]
+;CHECK-NEXT: v_mov_b32_e32
+define amdgpu_ps float @test_wwm4(i32 inreg %idx) {
+main_body:
+  ; use mbcnt to make sure the branch is divergent
+  %lo = call i32 @llvm.amdgcn.mbcnt.lo(i32 -1, i32 0)
+  %hi = call i32 @llvm.amdgcn.mbcnt.hi(i32 -1, i32 %lo)
+  %cc = icmp uge i32 %hi, 32
+  br i1 %cc, label %endif, label %if
+
+if:
+  %src = call float @llvm.amdgcn.buffer.load.f32(<4 x i32> undef, i32 %idx, i32 0, i1 0, i1 0)
+  %out = fadd float %src, %src
+  %out.0 = call float @llvm.amdgcn.wwm.f32(float %out)
+  br label %endif
+
+endif:
+  %out.1 = phi float [ %out.0, %if ], [ 0.0, %main_body ]
+  ret float %out.1
+}
+
+; Make sure the transition from Exact to WWM then WQM works properly.
+;
+;CHECK-LABEL: {{^}}test_wwm5:
+;CHECK: buffer_load_dword
+;CHECK: buffer_store_dword
+;CHECK: s_or_saveexec_b64 [[ORIG:s\[[0-9]+:[0-9]+\]]], -1
+;CHECK: buffer_load_dword
+;CHECK: v_add_f32_e32
+;CHECK: s_mov_b64 exec, [[ORIG]]
+;CHECK: s_wqm_b64 exec, exec
+define amdgpu_ps float @test_wwm5(i32 inreg %idx0, i32 inreg %idx1) {
+main_body:
+  %src0 = call float @llvm.amdgcn.buffer.load.f32(<4 x i32> undef, i32 %idx0, i32 0, i1 0, i1 0)
+  call void @llvm.amdgcn.buffer.store.f32(float %src0, <4 x i32> undef, i32 %idx0, i32 0, i1 0, i1 0)
+  %src1 = call float @llvm.amdgcn.buffer.load.f32(<4 x i32> undef, i32 %idx1, i32 0, i1 0, i1 0)
+  %temp = fadd float %src1, %src1
+  %temp.0 = call float @llvm.amdgcn.wwm.f32(float %temp)
+  %out = fadd float %temp.0, %temp.0
+  %out.0 = call float @llvm.amdgcn.wqm.f32(float %out)
+  ret float %out.0
+}
+
+; Check that WWM is turned on correctly across basic block boundaries.
+;
+;CHECK-LABEL: {{^}}test_wwm6:
+;CHECK: s_or_saveexec_b64 [[ORIG:s\[[0-9]+:[0-9]+\]]], -1
+;SI-CHECK: buffer_load_dword
+;VI-CHECK: flat_load_dword
+;CHECK: s_mov_b64 exec, [[ORIG]]
+;CHECK: %if
+;CHECK: s_or_saveexec_b64 [[ORIG2:s\[[0-9]+:[0-9]+\]]], -1
+;SI-CHECK: buffer_load_dword
+;VI-CHECK: flat_load_dword
+;CHECK: v_add_f32_e32
+;CHECK: s_mov_b64 exec, [[ORIG2]]
+define amdgpu_ps float @test_wwm6() {
+main_body:
+  %src0 = load volatile float, float addrspace(1)* undef
+  ; use mbcnt to make sure the branch is divergent
+  %lo = call i32 @llvm.amdgcn.mbcnt.lo(i32 -1, i32 0)
+  %hi = call i32 @llvm.amdgcn.mbcnt.hi(i32 -1, i32 %lo)
+  %cc = icmp uge i32 %hi, 32
+  br i1 %cc, label %endif, label %if
+
+if:
+  %src1 = load volatile float, float addrspace(1)* undef
+  %out = fadd float %src0, %src1
+  %out.0 = call float @llvm.amdgcn.wwm.f32(float %out)
+  br label %endif
+
+endif:
+  %out.1 = phi float [ %out.0, %if ], [ 0.0, %main_body ]
+  ret float %out.1
+}
+
 ; Check a case of one branch of an if-else requiring WQM, the other requiring
 ; exact.
 ;
 ; Note: In this particular case, the save-and-restore could be avoided if the
 ; analysis understood that the two branches of the if-else are mutually
 ; exclusive.
 ;
 ;CHECK-LABEL: {{^}}test_control_flow_0:

 declare <4 x float> @llvm.amdgcn.image.load.v4f32.v4i32.v8i32(<4 x i32>, <8 x i32>, i32, i1, i1, i1, i1) #3
 declare float @llvm.amdgcn.buffer.load.f32(<4 x i32>, i32, i32, i1, i1) #3
 declare <4 x float> @llvm.amdgcn.image.sample.v4f32.f32.v8i32(float, <8 x i32>, <4 x i32>, i32, i1, i1, i1, i1, i1) #3
 declare <4 x float> @llvm.amdgcn.image.sample.v4f32.v2f32.v8i32(<2 x float>, <8 x i32>, <4 x i32>, i32, i1, i1, i1, i1, i1) #3
 declare <4 x float> @llvm.amdgcn.image.sample.v4f32.v4f32.v8i32(<4 x float>, <8 x i32>, <4 x i32>, i32, i1, i1, i1, i1, i1) #3
 declare void @llvm.AMDGPU.kill(float) #1
 declare float @llvm.amdgcn.wqm.f32(float) #3
 declare i32 @llvm.amdgcn.wqm.i32(i32) #3
+declare float @llvm.amdgcn.wwm.f32(float) #3
+declare i32 @llvm.amdgcn.wwm.i32(i32) #3
+declare i32 @llvm.amdgcn.mbcnt.lo(i32, i32) #3
+declare i32 @llvm.amdgcn.mbcnt.hi(i32, i32) #3
 
 attributes #1 = { nounwind }
 attributes #2 = { nounwind readonly }
 attributes #3 = { nounwind readnone }
 attributes #4 = { "amdgpu-ps-wqm-outputs" }

test/CodeGen/AMDGPU/wqm.mir

+# RUN: llc -march=amdgcn -verify-machineinstrs -run-pass si-wqm -o -  %s | FileCheck %s
+
+---
+# Check for awareness that s_or_saveexec_b64 clobbers SCC
+#
+#CHECK: S_OR_SAVEEXEC_B64
+#CHECK: S_CMP_LT_I32
+#CHECK: S_CSELECT_B32
+name:            test_wwm_scc
+alignment:       0
+exposesReturnsTwice: false
+legalized:       false
+regBankSelected: false
+selected:        false
+tracksRegLiveness: true
+registers:       
+  - { id: 0, class: sgpr_32, preferred-register: '' }
+  - { id: 1, class: sgpr_32, preferred-register: '' }
+  - { id: 2, class: sgpr_32, preferred-register: '' }
+  - { id: 3, class: vgpr_32, preferred-register: '' }
+  - { id: 4, class: vgpr_32, preferred-register: '' }
+  - { id: 5, class: sgpr_32, preferred-register: '' }
+  - { id: 6, class: vgpr_32, preferred-register: '' }
+  - { id: 7, class: vgpr_32, preferred-register: '' }
+  - { id: 8, class: sreg_32_xm0, preferred-register: '' }
+  - { id: 9, class: sreg_32, preferred-register: '' }
+  - { id: 10, class: sreg_32, preferred-register: '' }
+  - { id: 11, class: vgpr_32, preferred-register: '' }
+  - { id: 12, class: vgpr_32, preferred-register: '' }
+liveins:         
+  - { reg: '%sgpr0', virtual-reg: '%0' }
+  - { reg: '%sgpr1', virtual-reg: '%1' }
+  - { reg: '%sgpr2', virtual-reg: '%2' }
+  - { reg: '%vgpr0', virtual-reg: '%3' }
+body:             |
+  bb.0:
+    liveins: %sgpr0, %sgpr1, %sgpr2, %vgpr0
+  
+    %3 = COPY %vgpr0
+    %2 = COPY %sgpr2
+    %1 = COPY %sgpr1
+    %0 = COPY %sgpr0
+    S_CMP_LT_I32 0, %0, implicit-def %scc
+    %12 = V_ADD_I32_e32 %3, %3, implicit-def %vcc, implicit %exec
+    %5 = S_CSELECT_B32 %2, %1, implicit %scc
+    %11 = V_ADD_I32_e32 %5, %12, implicit-def %vcc, implicit %exec
+    %vgpr0 = WWM %11, implicit %exec
+    SI_RETURN_TO_EPILOG %vgpr0
+
+...