[AMDGPU] SDWA: add support for PRESERVE into SDWA peephole. Add new merge SDWA preserve pass

[SamWot]

Summary

SDWA instructions support several values of dst_unused operand. One of this is UNUSED_PRESERVE. This value means that parts of destination register that are not wrote by SDWA instruction would not be changed. Currently SDWA peephole pass doesn't generate UNUSED_PRESERVE. It only generates UNUSED_PAD value that means that unused parts of dst register would be set to 0.
Big problem with UNUSED_PRESERVE is that by its nature it can't be represented in SSA form. PRESERVE assumes that register that it writes into was already wrote by some other instruction and our SDWA instruction keeps this value intact.
Another problem is that in AMDGPU backend smallest sub-reg is 32-bit wide and SDWA needs smaller so support for PRESERVE can't be done with subregs.
For those reasons support for PRESERVE for split into 2 major parts. First - changes in SDWA peephole pass that allows it to recognize patterns for PRESERVE and generate according instruction. This pass works on SSA machine code and generates SSA compatible code. Second part - new pass that works on non-SSA code and converts code generated by SDWA peephole into correct code.

1. Changes in SDWA peephole:

There were several changes in SDWA peephole.
a. First of all there was added new pattern to match for PRESERVE operand. This patterns looks for V_OR_B32 instruction with one of operands that is result of SDWA instruction. Second operand of V_OR_B32 should be instruction that is compatible bit-wise with SDWA instruction (there destination don't overlap). E.g. match:

v_add_f16_sdwa v0, v1, v2 dst_sel:WORD_1 dst_unused:UNUSED_PAD src1_sel:WORD_1 src2_sel:WORD_1
v_add_f16_e32 v3, v1, v2
v_or_b32_e32 v4, v0, v3

Into: SDWA preserve dst:v4 dst_sel:WORD_1 dst_unused:UNUSED_PRESERVE preserve:v3

Then this mathced SDWA preserve pattern is converted into SDWA with preserve. During conversion V_OR_B#@ instruction is replaced by SDWA instruction with dst_unused set to UNUSED_PRESERVE. Original instruction is removed. And new instruction gets additional implicit use-operand which is destination of second operand of V_OR_B32 (register that should be preserved):

v_add_f16_e32 v3, v1, v2
v_add_f16_sdwa v4, v1, v2 dst_sel:WORD_1 dst_unused:UNUSED_PRESERVE src1_sel:WORD_1 src2_sel:WORD_1, implicit v3

Problem with this match process is that currently it only works if both instructions were SDWA instructions. Reason is that to be able to match to instructions we should check that those two instructions are compatible to match - meaning that they write different parts of destination register. But currently there is no way to determine if regular instruction writes not whole destination register. E.g. we can't understand that V_ADD_F16 only writes low 16-bit of destination and high 16-bit are irrelevant. This can be determined only for SDWA instruction by looking into dst_sel operand. So for now this pattern only metch 2 SDWA instructions. Ability to match regular instructions would be added later.

b. Second big change in SDWA peephole pass is that now it tries to apply match patterns several times until it can't convert any new instruction. This is needed because (as said earlier) PRESERVE pattern need to match SDWA instructions but SDWA instruction apear (in most cases) only after SDWA peephole. So to be able to match PRESERVE pattern we first apply all other patterns that generate regular SDWA instructions and then on second try we apply PRESERVE pattern to SDWA instructions generated on first try.

2. New pass - merge SDWA preserve pass:

This pass is needed to convert SSA code generated by SDWA peephole pass into non-SSA correct code. It works after PHI-elimination pass where it is possible to generate non-SSA code.
This pass looks for SDWA instructions with dst_unused set to UNUSED_PRESERVE. In those instructions it looks for implicit register operand (which is added by SDWA peephole pass). This register is the one that should be reserved. Id such instruction is found then this pass changes destination register of this SDWA instruction to implicit register and creates copy from implicit register to original destination of SDWA instruction. E.g. instruction generated by SDWA peephole:

v_add_f16_sdwa v4, v1, v2 dst_sel:WORD_1 dst_unused:UNUSED_PRESERVE src1_sel:WORD_1 src2_sel:WORD_1, implicit v3

Would be converted into:

v_add_f16_sdwa v3, v1, v2 dst_sel:WORD_1 dst_unused:UNUSED_PRESERVE src1_sel:WORD_1 src2_sel:WORD_1, implicit v3
v_mov_b32 v4, v3

Putting it all together original sequence of instructions:

v_add_f16_sdwa v0, v1, v2 dst_sel:WORD_1 dst_unused:UNUSED_PAD src1_sel:WORD_1 src2_sel:WORD_1
v_add_f16_e32 v3, v1, v2
v_or_b32_e32 v4, v0, v3

Would be converted into:

v_add_f16_e32 v3, v1, v2
v_add_f16_sdwa v3, v1, v2 dst_sel:WORD_1 dst_unused:UNUSED_PRESERVE src1_sel:WORD_1 src2_sel:WORD_1, implicit v3
v_mov_b32 v4, v3

[arsenm]

I think we need to decide an overall strategy for dealing with instructions that only partially update the registers. GFX9 really complicated this issue by changing new instructions to preserve the high bits, and adding a control bit to some instructions to change the high bit behavior.

I started dealing with this to add the d16 loads and stores. We can still do this SSA by adding variants of the instructions with tied operands that preserve one half of the instructions, which would probably be less painful than adding another post-SSA pass that needs to deal with liveness. One issue is we still get suboptimal regalloc in some cases, so I'm debating adding new 16-bit subregister classes so subrange liveness tracking works.

[rampitec]

In D37817#869898, @arsenm wrote:

I think we need to decide an overall strategy for dealing with instructions that only partially update the registers. GFX9 really complicated this issue by changing new instructions to preserve the high bits, and adding a control bit to some instructions to change the high bit behavior.

I started dealing with this to add the d16 loads and stores. We can still do this SSA by adding variants of the instructions with tied operands that preserve one half of the instructions, which would probably be less painful than adding another post-SSA pass that needs to deal with liveness. One issue is we still get suboptimal regalloc in some cases, so I'm debating adding new 16-bit subregister classes so subrange liveness tracking works.

Do not we want to add bits to the instruction describing it preserves low or high half?
Adding new subregs would be quite painful, as we already have too much registers for RA and LIS to work fast and optimal.

[rampitec]

lib/Target/AMDGPU/AMDGPUTargetMachine.cpp

821

I do not think we need it under fast RA.

lib/Target/AMDGPU/SIMergeSDWAPreserve.cpp

115

We can have more than EXEC here. Check it is a virtual register instead, or even a VGPR?

lib/Target/AMDGPU/SIPeepholeSDWA.cpp

283

Such things shall be functions, function templates, anything but defines. In particular that is very hard to debug.

853

Needs cast from unsigned or use unsigned for SDWAOpcode/Opcode.

1052

This loop itself probably deserves a separate change.

test/CodeGen/AMDGPU/sdwa-merge-preserve.mir

1

Add -verify-machineinstrs to run lines.

test/CodeGen/AMDGPU/sdwa-preserve.mir

1

Add -verify-machineinstrs

[arsenm]

In D37817#870219, @rampitec wrote:

In D37817#869898, @arsenm wrote:

I think we need to decide an overall strategy for dealing with instructions that only partially update the registers. GFX9 really complicated this issue by changing new instructions to preserve the high bits, and adding a control bit to some instructions to change the high bit behavior.

I started dealing with this to add the d16 loads and stores. We can still do this SSA by adding variants of the instructions with tied operands that preserve one half of the instructions, which would probably be less painful than adding another post-SSA pass that needs to deal with liveness. One issue is we still get suboptimal regalloc in some cases, so I'm debating adding new 16-bit subregister classes so subrange liveness tracking works.

Do not we want to add bits to the instruction describing it preserves low or high half?
Adding new subregs would be quite painful, as we already have too much registers for RA and LIS to work fast and optimal.

That's another partial option, but won't solve the suboptimal RA. I think we need to try and see what the impact actually ends up being. These are more constrained since you sort of can't directly address the high component usually (i.e. the high component isn't actually separately allocatable).

[SamWot]

Resolved some issues

[arsenm]

In any case independent of sub register questions, I think this would be better off done in the existing pass by adding variants with tied operands. This is how I am handling this problem currently in D38070/D38071 for mad_mix

[rampitec]

In D37817#876063, @arsenm wrote:

In any case independent of sub register questions, I think this would be better off done in the existing pass by adding variants with tied operands. This is how I am handling this problem currently in D38070/D38071 for mad_mix

This is actually a better idea than using an implicit operand and having IR potentially broken in between of two passes.

[SamWot]

Removed SIMergeSDWAPreserve pass.
Use tied registers to achieve same results

[SamWot]

Ping

[arsenm]

lib/Target/AMDGPU/SIPeepholeSDWA.cpp

132

Probably should use LLVM_ENABLE_DUMP instead of NDEBUG. Also should add the matching dump() with LLVM_DUMP_METHOD.

292

use_nodbg_operands? Also space before :

293

C++ style comments

294–295

These various checks shouldn't be necessary in SSA. You can't have a def of a specific subregister (unless maybe there is a physical register which should probably be skipped anyway).

312–313

This seems to b be re-inventing MRI.getVRegDef/MRI.getUniqueVRegDef?

443

This whole loop is just MRI.clearKillFlags()

463–464

This is still manually tying the result operand. I was expecting another set of _sdwa opcodes with the preserve behavior with the tied operand statically known in the instruction definition. Manually tying this way is potentially hazardous because the verifier won't check it, and it makes it easier for another pass to accidentally drop the tied operand. I would expect there to be an InstrMapping table between the SDWA opcode and the SWDA with preserve set versions.

lib/Target/AMDGPU/SIRegisterInfo.cpp

1317–1318

How are these different from the various MCRegisterInfo functions for checking if registers alias or have subreg relationships?

[SamWot]

Fixed latests issues from arsenm

[SamWot]

lib/Target/AMDGPU/SIPeepholeSDWA.cpp

463–464

I thought that having separate instruction definition for just preserve case would be overkill. I didn't want to introduce another new kind of SDWA instruction that would only bloat already huge set of instruction definitions.
But if you think this would be better I will introduce new instructions definition.

[SamWot]

Ping.
Matt, what do you think about latest changes in reivew?

[arsenm]

At the very least there needs to be a verifier check for the tied operands if this is set. With the separate tied opcodes you get that for free

lib/Target/AMDGPU/SIPeepholeSDWA.cpp

988

Extra newline

[SamWot]

Added verification for tied register for UNUSED_PRESERVE

[arsenm]

lib/Target/AMDGPU/SIInstrInfo.cpp

2702–2705 ↗

(On Diff #124747)

Could use some of the stricter checks the normal verifier has, i.e.

else if (TargetRegisterInfo::isPhysicalRegister(MOTied.getReg()) &&
         MO->getReg() != MOTied.getReg())

[SamWot]

Stronger verification for UNUSED_PRESERVE

[arsenm]

LGTM

lib/Target/AMDGPU/AMDGPU.h

 FunctionPass *createR600ControlFlowFinalizer();
 FunctionPass *createAMDGPUCFGStructurizerPass();
 FunctionPass *createR600ISelDag(TargetMachine *TM, CodeGenOpt::Level OptLevel);
 
 // SI Passes
 FunctionPass *createSIAnnotateControlFlowPass();
 FunctionPass *createSIFoldOperandsPass();
 FunctionPass *createSIPeepholeSDWAPass();
+FunctionPass *createSIMergeSDWAPreservePass();
 FunctionPass *createSILowerI1CopiesPass();
 FunctionPass *createSIShrinkInstructionsPass();
 FunctionPass *createSILoadStoreOptimizerPass();
 FunctionPass *createSIWholeQuadModePass();
 FunctionPass *createSIFixControlFlowLiveIntervalsPass();
 FunctionPass *createSIOptimizeExecMaskingPreRAPass();
 FunctionPass *createSIFixSGPRCopiesPass();
 FunctionPass *createSIMemoryLegalizerPass();

 extern char &R600PacketizerID;
 
 void initializeSIFoldOperandsPass(PassRegistry &);
 extern char &SIFoldOperandsID;
 
 void initializeSIPeepholeSDWAPass(PassRegistry &);
 extern char &SIPeepholeSDWAID;
 
+void initializeSIMergeSDWAPreservePass(PassRegistry &);
+extern char &SIMergeSDWAPreserveID;
+
 void initializeSIShrinkInstructionsPass(PassRegistry&);
 extern char &SIShrinkInstructionsID;
 
 void initializeSIFixSGPRCopiesPass(PassRegistry &);
 extern char &SIFixSGPRCopiesID;
 
 void initializeSIFixVGPRCopiesPass(PassRegistry &);
 extern char &SIFixVGPRCopiesID;

lib/Target/AMDGPU/AMDGPUTargetMachine.cpp

extern "C" void LLVMInitializeAMDGPUTarget() {
   initializeR600ExpandSpecialInstrsPassPass(*PR);
   initializeR600VectorRegMergerPass(*PR);
   initializeAMDGPUDAGToDAGISelPass(*PR);
   initializeSILowerI1CopiesPass(*PR);
   initializeSIFixSGPRCopiesPass(*PR);
   initializeSIFixVGPRCopiesPass(*PR);
   initializeSIFoldOperandsPass(*PR);
   initializeSIPeepholeSDWAPass(*PR);
+  initializeSIMergeSDWAPreservePass(*PR);
   initializeSIShrinkInstructionsPass(*PR);
   initializeSIOptimizeExecMaskingPreRAPass(*PR);
   initializeSILoadStoreOptimizerPass(*PR);
   initializeAMDGPUAlwaysInlinePass(*PR);
   initializeAMDGPUAnnotateKernelFeaturesPass(*PR);
   initializeAMDGPUAnnotateUniformValuesPass(*PR);
   initializeAMDGPUArgumentUsageInfoPass(*PR);
   initializeAMDGPULowerIntrinsicsPass(*PR);
void GCNPassConfig::addFastRegAlloc(FunctionPass *RegAllocPass) {
   // 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);
 
+  // This must be run before TwoAddressInstructions, otherwise it will introduce
+  // a copy of merged register after SDWA instruction
+  if (EnableSDWAPeephole) {
+    insertPass(&SIFixWWMLivenessID, &SIMergeSDWAPreserveID, false);
+  }
+
   TargetPassConfig::addFastRegAlloc(RegAllocPass);
 }
 
 void GCNPassConfig::addOptimizedRegAlloc(FunctionPass *RegAllocPass) {
   insertPass(&MachineSchedulerID, &SIOptimizeExecMaskingPreRAID);
 
   // 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);
 
+  // This must be run before TwoAddressInstructions, otherwise it will introduce
+  // a copy of merged register after SDWA instruction
+  if (EnableSDWAPeephole) {
+    insertPass(&SIFixWWMLivenessID, &SIMergeSDWAPreserveID, false);
+  }
+
   TargetPassConfig::addOptimizedRegAlloc(RegAllocPass);
 }
 
 void GCNPassConfig::addPostRegAlloc() {
   addPass(&SIFixVGPRCopiesID);
   addPass(&SIOptimizeExecMaskingID);
   TargetPassConfig::addPostRegAlloc();
 }

lib/Target/AMDGPU/CMakeLists.txt

add_llvm_target(AMDGPUCodeGen
   SIInstrInfo.cpp
   SIISelLowering.cpp
   SILoadStoreOptimizer.cpp
   SILowerControlFlow.cpp
   SILowerI1Copies.cpp
   SIMachineFunctionInfo.cpp
   SIMachineScheduler.cpp
   SIMemoryLegalizer.cpp
+  SIMergeSDWAPreserve.cpp
   SIOptimizeExecMasking.cpp
   SIOptimizeExecMaskingPreRA.cpp
   SIPeepholeSDWA.cpp
   SIRegisterInfo.cpp
   SIShrinkInstructions.cpp
   SIWholeQuadMode.cpp
   )
 
 add_subdirectory(AsmParser)
 add_subdirectory(InstPrinter)
 add_subdirectory(Disassembler)
 add_subdirectory(TargetInfo)
 add_subdirectory(MCTargetDesc)
 add_subdirectory(Utils)

lib/Target/AMDGPU/SIMergeSDWAPreserve.cpp

+//===-- SIMergeSDWAPreserve.cpp - Merge SDWA dst_unused:PRESERVE operand --===//
+//
+//                     The LLVM Compiler Infrastructure
+//
+// This file is distributed under the University of Illinois Open Source
+// License. See LICENSE.TXT for details.
+//
+//===----------------------------------------------------------------------===//
+//
+/// \file This pass merges SDWA dst_unused:PRESERVE operand
+///
+/// E.g. original:
+///   %vreg2 = V_ADD_F16 %vreg0, %vreg1
+///   %vreg3 = V_ADD_F16_sdwa %vreg0, %vreg1
+///     dst_sel:WORD_1 dst_unused:UNUSED_RESERVE src0_sel:WORD_1 src1_sel:WORD_1
+///     %vreg2<imp-use>
+///
+/// Replace:
+///   %vreg2 = V_ADD_F16 %vreg0, %vreg1
+///   %vreg2 = V_ADD_F16_sdwa %vreg0, %vreg1
+///     dst_sel:WORD_1 dst_unused:UNUSED_RESERVE src0_sel:WORD_1 src1_sel:WORD_1
+///     %vreg3<imp-use>
+///   %vreg3 = V_MOV_B32 %vreg2
+///
+/// This pass should be placed after PHI elimination because it can't be
+/// implemented in SSA form
+//
+//===----------------------------------------------------------------------===//
+
+#include "AMDGPU.h"
+#include "AMDGPUSubtarget.h"
+#include "SIDefines.h"
+#include "SIInstrInfo.h"
+#include "llvm/ADT/Statistic.h"
+#include "llvm/CodeGen/MachineFunctionPass.h"
+#include "llvm/CodeGen/MachineInstrBuilder.h"
+
+using namespace llvm;
+
+#define DEBUG_TYPE "si-merge-sdwa-preserve"
+
+STATISTIC(NumMergeSDWAPreserve,
+          "Number of merges of SDWA dst_unused:PRESERVE.");
+
+namespace {
+
+  class SIMergeSDWAPreserve : public MachineFunctionPass {
+  public:
+    static char ID;
+
+    SIMergeSDWAPreserve() : MachineFunctionPass(ID) {
+      initializeSIMergeSDWAPreservePass(*PassRegistry::getPassRegistry());
+    }
+
+    bool runOnMachineFunction(MachineFunction &MF) override;
+
+    StringRef getPassName() const override { return "SI Merge SDWA Preserve"; }
+
+    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(SIMergeSDWAPreserve, DEBUG_TYPE, "SI Merge SDWA Preserve", false, false)
+
+char SIMergeSDWAPreserve::ID = 0;
+
+char &llvm::SIMergeSDWAPreserveID = SIMergeSDWAPreserve::ID;
+
+FunctionPass *llvm::createSIMergeSDWAPreservePass() {
+  return new SIMergeSDWAPreserve();
+}
+
+bool SIMergeSDWAPreserve::runOnMachineFunction(MachineFunction &MF) {
+  LiveIntervals *LIS = getAnalysisIfAvailable<LiveIntervals>();
+
+  const SISubtarget &ST = MF.getSubtarget<SISubtarget>();
+
+  if (!ST.hasSDWA())
+    return false;
+
+  bool Merged = false;
+
+  const MachineRegisterInfo *MRI = &MF.getRegInfo();
+  const SIRegisterInfo *TRI = ST.getRegisterInfo();
+  const SIInstrInfo *TII = ST.getInstrInfo();
+
+
+  for (MachineBasicBlock &MBB : MF) {
+    for (MachineInstr &MI : MBB) {
+      if (!TII->isSDWA(MI))
+        continue;
+
+      auto *DstUnused = TII->getNamedOperand(MI, AMDGPU::OpName::dst_unused);
+
+      if (!DstUnused || DstUnused->getImm() != AMDGPU::SDWA::UNUSED_PRESERVE)
+        continue;
+
+      // VOPC doesn't support PRESERVE
+      assert(!TII->isVOPC(AMDGPU::getBasicFromSDWAOp(MI.getOpcode())));
+
+      // Operand to copy should be last implicit operand in SDWA with
+      // preserve instruction
+      auto &SrcReg = MI.getOperand(MI.getNumOperands() - 1);
+      if (!SrcReg.isImplicit() || !SrcReg.isUse() ||
+          SrcReg.getReg() == AMDGPU::EXEC)
+        continue;
+
+
+      auto *DstReg = TII->getNamedOperand(MI, AMDGPU::OpName::vdst);
+      assert(DstReg &&
+             DstReg->isReg() &&
+             TRI->hasVGPRs(TII->getOpRegClass(MI, MI.getOperandNo(DstReg))));
+
+      // Check that MI is the only instruction that uses src register
+      for (MachineOperand &UseMO : MRI->use_operands(SrcReg.getReg())) {
+        if (TRI->isSubregOf(SrcReg, UseMO)) {
+          // TODO: in that case we should unfold this SDWA-with-PRESERVE back
+          // into v_or_b32 + SDWA-without-PRESERVE
+          assert(UseMO.getParent() == &MI);
+        }
+      }
+
+      MachineInstrBuilder Copy = BuildMI(MBB, MI.getNextNode(), MI.getDebugLoc(),
+                                         TII->get(AMDGPU::COPY),
+                                         DstReg->getReg())
+                                 .addReg(SrcReg.getReg());
+
+      DstReg->setReg(SrcReg.getReg());
+
+      if (LIS)
+        LIS->repairIntervalsInRange(&MBB, &MI, Copy.getInstr(),
+                                    { SrcReg.getReg(), DstReg->getReg() });
+
+      ++NumMergeSDWAPreserve;
+      Merged = true;
+    }
+  }
+
+  return Merged;
+}

lib/Target/AMDGPU/SIPeepholeSDWA.cpp

 
 STATISTIC(NumSDWAPatternsFound, "Number of SDWA patterns found.");
 STATISTIC(NumSDWAInstructionsPeepholed,
           "Number of instruction converted to SDWA.");
 
 namespace {
 
 class SDWAOperand;
+class SDWADstOperand;
 
 class SIPeepholeSDWA : public MachineFunctionPass {
 public:
   using SDWAOperandsVector = SmallVector<SDWAOperand *, 4>;
 
 private:
   MachineRegisterInfo *MRI;
   const SIRegisterInfo *TRI;
public:
   static char ID;
 
   SIPeepholeSDWA() : MachineFunctionPass(ID) {
     initializeSIPeepholeSDWAPass(*PassRegistry::getPassRegistry());
   }
 
   bool runOnMachineFunction(MachineFunction &MF) override;
   void matchSDWAOperands(MachineFunction &MF);
+  std::unique_ptr<SDWAOperand> matchSDWAOperand(MachineInstr &MI);
   bool isConvertibleToSDWA(const MachineInstr &MI, const SISubtarget &ST) const;
   bool convertToSDWA(MachineInstr &MI, const SDWAOperandsVector &SDWAOperands);
   void legalizeScalarOperands(MachineInstr &MI, const SISubtarget &ST) const;
 
   StringRef getPassName() const override { return "SI Peephole SDWA"; }
 
   void getAnalysisUsage(AnalysisUsage &AU) const override {
     AU.setPreservesCFG();
public:
 
   MachineOperand *getTargetOperand() const { return Target; }
   MachineOperand *getReplacedOperand() const { return Replaced; }
   MachineInstr *getParentInst() const { return Target->getParent(); }
 
   MachineRegisterInfo *getMRI() const {
     return &getParentInst()->getParent()->getParent()->getRegInfo();
   }
+
+  const SIRegisterInfo *getTRI() const {
+    return static_cast<const SIRegisterInfo *>(getMRI()->getTargetRegisterInfo());
+  }
+
+#ifndef NDEBUG
+  virtual void print(raw_ostream& OS) const = 0;
+#endif
 };
 
 using namespace AMDGPU::SDWA;
 
 class SDWASrcOperand : public SDWAOperand {
 private:
   SdwaSel SrcSel;
   bool Abs;
   bool Neg;
   bool Sext;
 
 public:
   SDWASrcOperand(MachineOperand *TargetOp, MachineOperand *ReplacedOp,
                  SdwaSel SrcSel_ = DWORD, bool Abs_ = false, bool Neg_ = false,
                  bool Sext_ = false)
-      : SDWAOperand(TargetOp, ReplacedOp), SrcSel(SrcSel_), Abs(Abs_),
-        Neg(Neg_), Sext(Sext_) {}
+      : SDWAOperand(TargetOp, ReplacedOp),
+        SrcSel(SrcSel_), Abs(Abs_), Neg(Neg_), Sext(Sext_) {}
 
   MachineInstr *potentialToConvert(const SIInstrInfo *TII) override;
   bool convertToSDWA(MachineInstr &MI, const SIInstrInfo *TII) override;
 
   SdwaSel getSrcSel() const { return SrcSel; }
   bool getAbs() const { return Abs; }
   bool getNeg() const { return Neg; }
   bool getSext() const { return Sext; }
 
   uint64_t getSrcMods(const SIInstrInfo *TII,
                       const MachineOperand *SrcOp) const;
+
+#ifndef NDEBUG
+  void print(raw_ostream& OS) const override;
+#endif
 };
 
 class SDWADstOperand : public SDWAOperand {
 private:
   SdwaSel DstSel;
   DstUnused DstUn;
 
 public:
+
   SDWADstOperand(MachineOperand *TargetOp, MachineOperand *ReplacedOp,
                  SdwaSel DstSel_ = DWORD, DstUnused DstUn_ = UNUSED_PAD)
-      : SDWAOperand(TargetOp, ReplacedOp), DstSel(DstSel_), DstUn(DstUn_) {}
 
   MachineInstr *potentialToConvert(const SIInstrInfo *TII) override;
   bool convertToSDWA(MachineInstr &MI, const SIInstrInfo *TII) override;
 
   SdwaSel getDstSel() const { return DstSel; }
   DstUnused getDstUnused() const { return DstUn; }
+
+#ifndef NDEBUG
+  void print(raw_ostream& OS) const override;
+#endif
+};
+
+class SDWADstPreserveOperand : public SDWADstOperand {
+private:
+  MachineOperand *Preserve;
+
+public:
+  SDWADstPreserveOperand(MachineOperand *TargetOp, MachineOperand *ReplacedOp,
+                         MachineOperand *PreserveOp, SdwaSel DstSel_ = DWORD)
+      : SDWADstOperand(TargetOp, ReplacedOp, DstSel_, UNUSED_PRESERVE),
+        Preserve(PreserveOp) {}
+
+  bool convertToSDWA(MachineInstr &MI, const SIInstrInfo *TII) override;
+
+  MachineOperand *getPreservedOperand() const { return Preserve; }
+
+#ifndef NDEBUG
+  void print(raw_ostream& OS) const override;
+#endif
 };
 
 } // end anonymous namespace
 
 INITIALIZE_PASS(SIPeepholeSDWA, DEBUG_TYPE, "SI Peephole SDWA", false, false)
 
 char SIPeepholeSDWA::ID = 0;
 
static raw_ostream& operator<<(raw_ostream &OS, const DstUnused &Un) {
   switch(Un) {
   case UNUSED_PAD: OS << "UNUSED_PAD"; break;
   case UNUSED_SEXT: OS << "UNUSED_SEXT"; break;
   case UNUSED_PRESERVE: OS << "UNUSED_PRESERVE"; break;
   }
   return OS;
 }
 
-static raw_ostream& operator<<(raw_ostream &OS, const SDWASrcOperand &Src) {
-  OS << "SDWA src: " << *Src.getTargetOperand()
-     << " src_sel:" << Src.getSrcSel()
-     << " abs:" << Src.getAbs() << " neg:" << Src.getNeg()
-     << " sext:" << Src.getSext() << '\n';
+static raw_ostream& operator<<(raw_ostream &OS, const SDWAOperand &Operand) {
+  Operand.print(OS);
   return OS;
 }
 
-static raw_ostream& operator<<(raw_ostream &OS, const SDWADstOperand &Dst) {
-  OS << "SDWA dst: " << *Dst.getTargetOperand()
-     << " dst_sel:" << Dst.getDstSel()
-     << " dst_unused:" << Dst.getDstUnused() << '\n';
-  return OS;
+void SDWASrcOperand::print(raw_ostream& OS) const {
+  OS << "SDWA src: " << *getTargetOperand()
+    << " src_sel:" << getSrcSel()
+    << " abs:" << getAbs() << " neg:" << getNeg()
+    << " sext:" << getSext() << '\n';
+}
+
+void SDWADstOperand::print(raw_ostream& OS) const {
+  OS << "SDWA dst: " << *getTargetOperand()
+    << " dst_sel:" << getDstSel()
+    << " dst_unused:" << getDstUnused() << '\n';
+}
+
+void SDWADstPreserveOperand::print(raw_ostream& OS) const {
+  OS << "SDWA preserve dst: " << *getTargetOperand()
+    << " dst_sel:" << getDstSel()
+    << " preserve:" << *getPreservedOperand() << '\n';
 }
+
 #endif
 
 static void copyRegOperand(MachineOperand &To, const MachineOperand &From) {
   assert(To.isReg() && From.isReg());
   To.setReg(From.getReg());
   To.setSubReg(From.getSubReg());
   To.setIsUndef(From.isUndef());
   if (To.isUse()) {
     To.setIsKill(From.isKill());
   } else {
     To.setIsDead(From.isDead());
   }
 }
 
-static bool isSameReg(const MachineOperand &LHS, const MachineOperand &RHS) {
-  return LHS.isReg() &&
-         RHS.isReg() &&
-         LHS.getReg() == RHS.getReg() &&
-         LHS.getSubReg() == RHS.getSubReg();
+#define FIND_SINGLE_REG_DEF_USE(NAME, FNAME)                                  \
+static MachineOperand *findSingleReg##NAME(const MachineOperand *Reg,         \
+                                           const MachineRegisterInfo *MRI) {  \
+  if (!Reg->isReg())                                                          \
+    return nullptr;                                                           \
+                                                                              \
+  auto TRI = static_cast<const SIRegisterInfo *>(                             \
+                                MRI->getTargetRegisterInfo());                \
+                                                                              \
+  MachineOperand *DefUseMO = nullptr;                                         \
+  for (MachineOperand &DefUse: MRI->FNAME(Reg->getReg())) {                   \
+    /* If this is def/use of another subreg of reg then do nothing */         \
+    if (!TRI->isSubregOf(DefUse, *Reg)                                        \
+        || !TRI->isSubregOf(*Reg, DefUse))                                    \
+      continue;                                                               \
+                                                                              \
+    /* If there exist def/use of superreg of reg then return nullptr*/        \
+    if (!TRI->isSameReg(DefUse, *Reg))                                        \
+      return nullptr;                                                         \
+                                                                              \
+    /* Check that UseMI is only instruction that defs/uses reg */             \
+    if (!DefUseMO) {                                                          \
+      DefUseMO = &DefUse;                                                     \
+    } else if (DefUseMO->getParent() != DefUse.getParent()) {                 \
+      return nullptr;                                                         \
+    }                                                                         \
+  }                                                                           \
+  return DefUseMO;                                                            \
 }
 
-static bool isSubregOf(const MachineOperand &SubReg,
-                       const MachineOperand &SuperReg,
-                       const TargetRegisterInfo *TRI) {
-
-  if (!SuperReg.isReg() || !SubReg.isReg())
-    return false;
+FIND_SINGLE_REG_DEF_USE(Use, use_operands)
+FIND_SINGLE_REG_DEF_USE(Def, def_operands)
 
-  if (isSameReg(SuperReg, SubReg))
-    return true;
-
-  if (SuperReg.getReg() != SubReg.getReg())
-    return false;
-
-  LaneBitmask SuperMask = TRI->getSubRegIndexLaneMask(SuperReg.getSubReg());
-  LaneBitmask SubMask = TRI->getSubRegIndexLaneMask(SubReg.getSubReg());
-  SuperMask |= ~SubMask;
-  return SuperMask.all();
-}
+#undef FIND_SINGLE_REG_DEF_USE
 
 uint64_t SDWASrcOperand::getSrcMods(const SIInstrInfo *TII,
                                     const MachineOperand *SrcOp) const {
   uint64_t Mods = 0;
   const auto *MI = SrcOp->getParent();
   if (TII->getNamedOperand(*MI, AMDGPU::OpName::src0) == SrcOp) {
     if (auto *Mod = TII->getNamedOperand(*MI, AMDGPU::OpName::src0_modifiers)) {
       Mods = Mod->getImm();
uint64_t SDWASrcOperand::getSrcMods(const SIInstrInfo *TII,
   }
 
   return Mods;
 }
 
 MachineInstr *SDWASrcOperand::potentialToConvert(const SIInstrInfo *TII) {
   // For SDWA src operand potential instruction is one that use register
   // defined by parent instruction
-  MachineRegisterInfo *MRI = getMRI();
-  MachineOperand *Replaced = getReplacedOperand();
-  assert(Replaced->isReg());
-
-  MachineInstr *PotentialMI = nullptr;
-  for (MachineOperand &PotentialMO : MRI->use_operands(Replaced->getReg())) {
-    // If this is use of another subreg of dst reg then do nothing
-    if (!isSubregOf(*Replaced, PotentialMO, MRI->getTargetRegisterInfo()))
-      continue;
-
-    // If there exist use of superreg of dst then we should not combine this
-    // opernad
-    if (!isSameReg(PotentialMO, *Replaced))
-      return nullptr;
+  MachineOperand *PotentialMO = findSingleRegUse(getReplacedOperand(), getMRI());
+  if (!PotentialMO)
 
-    // Check that PotentialMI is only instruction that uses dst reg
-    if (PotentialMI == nullptr) {
-      PotentialMI = PotentialMO.getParent();
-    } else if (PotentialMI != PotentialMO.getParent()) {
-      return nullptr;
-    }
-  }
-
-  return PotentialMI;
+  return PotentialMO->getParent();
 }
 
 bool SDWASrcOperand::convertToSDWA(MachineInstr &MI, const SIInstrInfo *TII) {
   // Find operand in instruction that matches source operand and replace it with
   // target operand. Set corresponding src_sel
 
+  const SIRegisterInfo *TRI = getTRI();
+
   MachineOperand *Src = TII->getNamedOperand(MI, AMDGPU::OpName::src0);
   MachineOperand *SrcSel = TII->getNamedOperand(MI, AMDGPU::OpName::src0_sel);
   MachineOperand *SrcMods =
       TII->getNamedOperand(MI, AMDGPU::OpName::src0_modifiers);
   assert(Src && (Src->isReg() || Src->isImm()));
-  if (!isSameReg(*Src, *getReplacedOperand())) {
+  if (!TRI->isSameReg(*Src, *getReplacedOperand())) {
     // If this is not src0 then it should be src1
     Src = TII->getNamedOperand(MI, AMDGPU::OpName::src1);
     SrcSel = TII->getNamedOperand(MI, AMDGPU::OpName::src1_sel);
     SrcMods = TII->getNamedOperand(MI, AMDGPU::OpName::src1_modifiers);
 
     assert(Src && Src->isReg());
 
     if ((MI.getOpcode() == AMDGPU::V_MAC_F16_sdwa ||
          MI.getOpcode() == AMDGPU::V_MAC_F32_sdwa) &&
-        !isSameReg(*Src, *getReplacedOperand())) {
+        !TRI->isSameReg(*Src, *getReplacedOperand())) {
       // In case of v_mac_f16/32_sdwa this pass can try to apply src operand to
       // src2. This is not allowed.
       return false;
     }
 
-    assert(isSameReg(*Src, *getReplacedOperand()) && SrcSel && SrcMods);
+    assert(TRI->isSameReg(*Src, *getReplacedOperand()) && SrcSel && SrcMods);
   }
   copyRegOperand(*Src, *getTargetOperand());
   SrcSel->setImm(getSrcSel());
   SrcMods->setImm(getSrcMods(TII, Src));
   getTargetOperand()->setIsKill(false);
   return true;
 }
 
 MachineInstr *SDWADstOperand::potentialToConvert(const SIInstrInfo *TII) {
   // For SDWA dst operand potential instruction is one that defines register
   // that this operand uses
   MachineRegisterInfo *MRI = getMRI();
   MachineInstr *ParentMI = getParentInst();
-  MachineOperand *Replaced = getReplacedOperand();
-  assert(Replaced->isReg());
-
-  for (MachineOperand &PotentialMO : MRI->def_operands(Replaced->getReg())) {
-    if (!isSubregOf(*Replaced, PotentialMO, MRI->getTargetRegisterInfo()))
-      continue;
 
-    if (!isSameReg(*Replaced, PotentialMO))
-      return nullptr;
+  MachineOperand *PotentialMO = findSingleRegDef(getReplacedOperand(), MRI);
+  if (!PotentialMO)
 
-    // Check that ParentMI is the only instruction that uses replaced register
-    for (MachineOperand &UseMO : MRI->use_operands(PotentialMO.getReg())) {
-      if (isSubregOf(UseMO, PotentialMO, MRI->getTargetRegisterInfo()) &&
-          UseMO.getParent() != ParentMI) {
-        return nullptr;
-      }
-    }
+  for (MachineOperand &UseMO : MRI->use_operands(PotentialMO->getReg())) {
+    if (getTRI()->isSubregOf(UseMO, *PotentialMO) &&
 
-    // Due to SSA this should be onle def of replaced register, so return it
-    return PotentialMO.getParent();
-  }
-
-  return nullptr;
+  return PotentialMO->getParent();
 }
 
 bool SDWADstOperand::convertToSDWA(MachineInstr &MI, const SIInstrInfo *TII) {
   // Replace vdst operand in MI with target operand. Set dst_sel and dst_unused
 
   if ((MI.getOpcode() == AMDGPU::V_MAC_F16_sdwa ||
        MI.getOpcode() == AMDGPU::V_MAC_F32_sdwa) &&
       getDstSel() != AMDGPU::SDWA::DWORD) {
     // v_mac_f16/32_sdwa allow dst_sel to be equal only to DWORD
     return false;
   }
 
   MachineOperand *Operand = TII->getNamedOperand(MI, AMDGPU::OpName::vdst);
   assert(Operand &&
          Operand->isReg() &&
-         isSameReg(*Operand, *getReplacedOperand()));
+         getTRI()->isSameReg(*Operand, *getReplacedOperand()));
   copyRegOperand(*Operand, *getTargetOperand());
   MachineOperand *DstSel= TII->getNamedOperand(MI, AMDGPU::OpName::dst_sel);
   assert(DstSel);
   DstSel->setImm(getDstSel());
   MachineOperand *DstUnused= TII->getNamedOperand(MI, AMDGPU::OpName::dst_unused);
   assert(DstUnused);
   DstUnused->setImm(getDstUnused());
 
   // Remove original instruction  because it would conflict with our new
   // instruction by register definition
   getParentInst()->eraseFromParent();
   return true;
 }
 
+bool SDWADstPreserveOperand::convertToSDWA(MachineInstr &MI,
+                                           const SIInstrInfo *TII) {
+  // MI should be moved right before v_or_b32.
+  // For this we should clear all kill flags on uses of MI src-operands or else
+  // we can encounter problem with use of killed operand.
+  for (MachineOperand &MO: MI.uses()) {
+    if (!MO.isReg())
+      continue;
+    for (MachineOperand &Use: getMRI()->use_operands(MO.getReg())) {
+      Use.setIsKill(false);
+    }
+  }
+
+  // Move MI before v_or_b32
+  auto MBB = MI.getParent();
+  MBB->remove(&MI);
+  MBB->insert(getParentInst(), &MI);
+
+  // Add Implicit use of preserved register
+  MachineInstrBuilder MIB(*MBB->getParent(), MI);
+  MIB.addReg(getPreservedOperand()->getReg(),
+             RegState::ImplicitKill,
+             getPreservedOperand()->getSubReg());
+
+  // Convert MI as any other SDWADstOperand and remove v_or_b32
+  return SDWADstOperand::convertToSDWA(MI, TII);
+}
+
 Optional<int64_t> SIPeepholeSDWA::foldToImm(const MachineOperand &Op) const {
   if (Op.isImm()) {
     return Op.getImm();
   }
 
   // If this is not immediate then it can be copy of immediate value, e.g.:
   // %vreg1<def> = S_MOV_B32 255;
   if (Op.isReg()) {
     for (const MachineOperand &Def : MRI->def_operands(Op.getReg())) {
-      if (!isSameReg(Op, Def))
+      if (!TRI->isSameReg(Op, Def))
         continue;
 
       const MachineInstr *DefInst = Def.getParent();
       if (!TII->isFoldableCopy(*DefInst))
         return None;
 
       const MachineOperand &Copied = DefInst->getOperand(1);
       if (!Copied.isImm())
         return None;
 
       return Copied.getImm();
     }
   }
 
   return None;
 }
 
-void SIPeepholeSDWA::matchSDWAOperands(MachineFunction &MF) {
-  for (MachineBasicBlock &MBB : MF) {
-    for (MachineInstr &MI : MBB) {
-      unsigned Opcode = MI.getOpcode();
-      switch (Opcode) {
-      case AMDGPU::V_LSHRREV_B32_e32:
-      case AMDGPU::V_ASHRREV_I32_e32:
-      case AMDGPU::V_LSHLREV_B32_e32:
-      case AMDGPU::V_LSHRREV_B32_e64:
-      case AMDGPU::V_ASHRREV_I32_e64:
-      case AMDGPU::V_LSHLREV_B32_e64: {
-        // from: v_lshrrev_b32_e32 v1, 16/24, v0
-        // to SDWA src:v0 src_sel:WORD_1/BYTE_3
+std::unique_ptr<SDWAOperand>
+SIPeepholeSDWA::matchSDWAOperand(MachineInstr &MI) {
 
-        // from: v_ashrrev_i32_e32 v1, 16/24, v0
-        // to SDWA src:v0 src_sel:WORD_1/BYTE_3 sext:1
 
-        // from: v_lshlrev_b32_e32 v1, 16/24, v0
-        // to SDWA dst:v1 dst_sel:WORD_1/BYTE_3 dst_unused:UNUSED_PAD
-        MachineOperand *Src0 = TII->getNamedOperand(MI, AMDGPU::OpName::src0);
-        auto Imm = foldToImm(*Src0);
-        if (!Imm)
-          break;
 
-        if (*Imm != 16 && *Imm != 24)
-          break;
 
-        MachineOperand *Src1 = TII->getNamedOperand(MI, AMDGPU::OpName::src1);
-        MachineOperand *Dst = TII->getNamedOperand(MI, AMDGPU::OpName::vdst);
-        if (TRI->isPhysicalRegister(Src1->getReg()) ||
-            TRI->isPhysicalRegister(Dst->getReg()))
-          break;
 
-        if (Opcode == AMDGPU::V_LSHLREV_B32_e32 ||
-            Opcode == AMDGPU::V_LSHLREV_B32_e64) {
-          auto SDWADst = make_unique<SDWADstOperand>(
-              Dst, Src1, *Imm == 16 ? WORD_1 : BYTE_3, UNUSED_PAD);
-          DEBUG(dbgs() << "Match: " << MI << "To: " << *SDWADst << '\n');
-          SDWAOperands[&MI] = std::move(SDWADst);
-          ++NumSDWAPatternsFound;
-        } else {
-          auto SDWASrc = make_unique<SDWASrcOperand>(
-              Src1, Dst, *Imm == 16 ? WORD_1 : BYTE_3, false, false,
-              Opcode != AMDGPU::V_LSHRREV_B32_e32 &&
-              Opcode != AMDGPU::V_LSHRREV_B32_e64);
-          DEBUG(dbgs() << "Match: " << MI << "To: " << *SDWASrc << '\n');
-          SDWAOperands[&MI] = std::move(SDWASrc);
-          ++NumSDWAPatternsFound;
-        }
-        break;
-      }
+      return make_unique<SDWADstOperand>(
+      return make_unique<SDWASrcOperand>(
 
-      case AMDGPU::V_LSHRREV_B16_e32:
-      case AMDGPU::V_ASHRREV_I16_e32:
-      case AMDGPU::V_LSHLREV_B16_e32:
-      case AMDGPU::V_LSHRREV_B16_e64:
-      case AMDGPU::V_ASHRREV_I16_e64:
-      case AMDGPU::V_LSHLREV_B16_e64: {
-        // from: v_lshrrev_b16_e32 v1, 8, v0
-        // to SDWA src:v0 src_sel:BYTE_1
 
-        // from: v_ashrrev_i16_e32 v1, 8, v0
-        // to SDWA src:v0 src_sel:BYTE_1 sext:1
 
-        // from: v_lshlrev_b16_e32 v1, 8, v0
-        // to SDWA dst:v1 dst_sel:BYTE_1 dst_unused:UNUSED_PAD
-        MachineOperand *Src0 = TII->getNamedOperand(MI, AMDGPU::OpName::src0);
-        auto Imm = foldToImm(*Src0);
-        if (!Imm || *Imm != 8)
-          break;
 
-        MachineOperand *Src1 = TII->getNamedOperand(MI, AMDGPU::OpName::src1);
-        MachineOperand *Dst = TII->getNamedOperand(MI, AMDGPU::OpName::vdst);
 
-        if (TRI->isPhysicalRegister(Src1->getReg()) ||
-            TRI->isPhysicalRegister(Dst->getReg()))
-          break;
 
-        if (Opcode == AMDGPU::V_LSHLREV_B16_e32 ||
-            Opcode == AMDGPU::V_LSHLREV_B16_e64) {
-          auto SDWADst =
-            make_unique<SDWADstOperand>(Dst, Src1, BYTE_1, UNUSED_PAD);
-          DEBUG(dbgs() << "Match: " << MI << "To: " << *SDWADst << '\n');
-          SDWAOperands[&MI] = std::move(SDWADst);
-          ++NumSDWAPatternsFound;
-        } else {
-          auto SDWASrc = make_unique<SDWASrcOperand>(
-              Src1, Dst, BYTE_1, false, false,
-              Opcode != AMDGPU::V_LSHRREV_B16_e32 &&
-              Opcode != AMDGPU::V_LSHRREV_B16_e64);
-          DEBUG(dbgs() << "Match: " << MI << "To: " << *SDWASrc << '\n');
-          SDWAOperands[&MI] = std::move(SDWASrc);
-          ++NumSDWAPatternsFound;
-        }
-        break;
-      }
+      return make_unique<SDWADstOperand>(Dst, Src1, BYTE_1, UNUSED_PAD);
+      return make_unique<SDWASrcOperand>(
 
-      case AMDGPU::V_BFE_I32:
-      case AMDGPU::V_BFE_U32: {
-        // e.g.:
-        // from: v_bfe_u32 v1, v0, 8, 8
-        // to SDWA src:v0 src_sel:BYTE_1
 
-        // offset | width | src_sel
-        // ------------------------
-        // 0      | 8     | BYTE_0
-        // 0      | 16    | WORD_0
-        // 0      | 32    | DWORD ?
-        // 8      | 8     | BYTE_1
-        // 16     | 8     | BYTE_2
-        // 16     | 16    | WORD_1
-        // 24     | 8     | BYTE_3
 
-        MachineOperand *Src1 = TII->getNamedOperand(MI, AMDGPU::OpName::src1);
-        auto Offset = foldToImm(*Src1);
-        if (!Offset)
-          break;
 
-        MachineOperand *Src2 = TII->getNamedOperand(MI, AMDGPU::OpName::src2);
-        auto Width = foldToImm(*Src2);
-        if (!Width)
-          break;
 
-        SdwaSel SrcSel = DWORD;
 
-        if (*Offset == 0 && *Width == 8)
-          SrcSel = BYTE_0;
-        else if (*Offset == 0 && *Width == 16)
-          SrcSel = WORD_0;
-        else if (*Offset == 0 && *Width == 32)
-          SrcSel = DWORD;
-        else if (*Offset == 8 && *Width == 8)
-          SrcSel = BYTE_1;
-        else if (*Offset == 16 && *Width == 8)
-          SrcSel = BYTE_2;
-        else if (*Offset == 16 && *Width == 16)
-          SrcSel = WORD_1;
-        else if (*Offset == 24 && *Width == 8)
-          SrcSel = BYTE_3;
-        else
-          break;
 
-        MachineOperand *Src0 = TII->getNamedOperand(MI, AMDGPU::OpName::src0);
-        MachineOperand *Dst = TII->getNamedOperand(MI, AMDGPU::OpName::vdst);
 
-        if (TRI->isPhysicalRegister(Src0->getReg()) ||
-            TRI->isPhysicalRegister(Dst->getReg()))
-          break;
 
-        auto SDWASrc = make_unique<SDWASrcOperand>(
-            Src0, Dst, SrcSel, false, false,
-            Opcode != AMDGPU::V_BFE_U32);
-        DEBUG(dbgs() << "Match: " << MI << "To: " << *SDWASrc << '\n');
-        SDWAOperands[&MI] = std::move(SDWASrc);
-        ++NumSDWAPatternsFound;
-        break;
-      }
-      case AMDGPU::V_AND_B32_e32:
-      case AMDGPU::V_AND_B32_e64: {
-        // e.g.:
-        // from: v_and_b32_e32 v1, 0x0000ffff/0x000000ff, v0
-        // to SDWA src:v0 src_sel:WORD_0/BYTE_0
+    return make_unique<SDWASrcOperand>(
+          Src0, Dst, SrcSel, false, false, Opcode != AMDGPU::V_BFE_U32);
+
 
-        MachineOperand *Src0 = TII->getNamedOperand(MI, AMDGPU::OpName::src0);
-        MachineOperand *Src1 = TII->getNamedOperand(MI, AMDGPU::OpName::src1);
-        auto ValSrc = Src1;
-        auto Imm = foldToImm(*Src0);
 
-        if (!Imm) {
-          Imm = foldToImm(*Src1);
-          ValSrc = Src0;
-        }
 
-        if (!Imm || (*Imm != 0x0000ffff && *Imm != 0x000000ff))
-          break;
 
-        MachineOperand *Dst = TII->getNamedOperand(MI, AMDGPU::OpName::vdst);
 
-        if (TRI->isPhysicalRegister(Src1->getReg()) ||
-            TRI->isPhysicalRegister(Dst->getReg()))
-          break;
 
-        auto SDWASrc = make_unique<SDWASrcOperand>(
-            ValSrc, Dst, *Imm == 0x0000ffff ? WORD_0 : BYTE_0);
-        DEBUG(dbgs() << "Match: " << MI << "To: " << *SDWASrc << '\n');
-        SDWAOperands[&MI] = std::move(SDWASrc);
-        ++NumSDWAPatternsFound;
+    return make_unique<SDWASrcOperand>(
+  }
+
+  case AMDGPU::V_OR_B32_e32:
+  case AMDGPU::V_OR_B32_e64: {
+    // Patterns for dst_unused:UNUSED_PRESERVE.
+    // e.g., from:
+    // v_add_f16_sdwa v0, v1, v2 dst_sel:WORD_1 dst_unused:UNUSED_PAD
+    //                           src1_sel:WORD_1 src2_sel:WORD1
+    // v_add_f16_e32 v3, v1, v2
+    // v_or_b32_e32 v4, v0, v3
+    // to SDWA preserve dst:v4 dst_sel:WORD_1 dst_unused:UNUSED_PRESERVE preserve:v3
+
+    // Check if one of operands of v_or_b32 is SDWA instruction
+    using CheckRetType = Optional<std::pair<MachineOperand *, MachineOperand *>>;
+    auto CheckOROperandsForSDWA =
+      [&](const MachineOperand *Op1, const MachineOperand *Op2) -> CheckRetType {
+        if (!Op1 || !Op1->isReg() || !Op2 || !Op2->isReg())
+          return CheckRetType(None);
+
+        MachineOperand *Op1Def = findSingleRegDef(Op1, MRI);
+        if (!Op1Def)
+          return CheckRetType(None);
+
+        MachineInstr *Op1Inst = Op1Def->getParent();
+        if (!TII->isSDWA(*Op1Inst))
+          return CheckRetType(None);
+
+        MachineOperand *Op2Def = findSingleRegDef(Op2, MRI);
+        if (!Op2Def)
+          return CheckRetType(None);
+
+        return CheckRetType(std::make_pair(Op1Def, Op2Def));
+      };
+
+    MachineOperand *OrSDWA = TII->getNamedOperand(MI, AMDGPU::OpName::src0);
+    MachineOperand *OrOther = TII->getNamedOperand(MI, AMDGPU::OpName::src1);
+    assert(OrSDWA && OrOther);
+    auto Res = CheckOROperandsForSDWA(OrSDWA, OrOther);
+    if (!Res) {
+      OrSDWA = TII->getNamedOperand(MI, AMDGPU::OpName::src1);
+      OrOther = TII->getNamedOperand(MI, AMDGPU::OpName::src0);
+      assert(OrSDWA && OrOther);
+      Res = CheckOROperandsForSDWA(OrSDWA, OrOther);
+      if (!Res)
         break;
-      }
+
+    MachineOperand *OrSDWADef = Res->first;
+    MachineOperand *OrOtherDef = Res->second;
+    assert(OrSDWADef && OrOtherDef);
+
+    MachineInstr *SDWAInst = OrSDWADef->getParent();
+    MachineInstr *OtherInst = OrOtherDef->getParent();
+
+    // Check that OtherInstr is actually bitwise compatible with SDWAInst = their
+    // destination patterns don't overlap. Compatible instruction can be either
+    // regular instruction with compatible bitness or SDWA instruction with
+    // correct dst_sel
+    // SDWAInst | OtherInst bitness / OtherInst dst_sel
+    // -----------------------------------------------------
+    // DWORD    | no                    / no
+    // WORD_0   | no                    / BYTE_2/3, WORD_1
+    // WORD_1   | 8/16-bit instructions / BYTE_0/1, WORD_0
+    // BYTE_0   | no                    / BYTE_1/2/3, WORD_1
+    // BYTE_1   | 8-bit                 / BYTE_0/2/3, WORD_1
+    // BYTE_2   | 8/16-bit              / BYTE_0/1/3. WORD_0
+    // BYTE_3   | 8/16/24-bit           / BYTE_0/1/2, WORD_0
+    // E.g. if SDWAInst is v_add_f16_sdwa dst_sel:WORD_1 then v_add_f16 is OK
+    // but v_add_f32 is not.
+
+    // TODO: add support for non-SDWA instructions as OtherInst.
+    // For now this only works with SDWA instructions. For regular instructions
+    // there is no way to determine if instruction write only 8/16/24-bit out of
+    // full register size and all registers are at min 32-bit wide.
+    if (!TII->isSDWA(*OtherInst))
+      break;
+
+    SdwaSel DstSel = static_cast<SdwaSel>(
+      TII->getNamedImmOperand(*SDWAInst, AMDGPU::OpName::dst_sel));;
+    SdwaSel OtherDstSel = static_cast<SdwaSel>(
+      TII->getNamedImmOperand(*OtherInst, AMDGPU::OpName::dst_sel));
+
+    bool DstSelAgree = false;
+    switch (DstSel) {
+    case WORD_0: DstSelAgree = ((OtherDstSel == BYTE_2) ||
+                                (OtherDstSel == BYTE_3) ||
+                                (OtherDstSel == WORD_1));
+      break;
+    case WORD_1: DstSelAgree = ((OtherDstSel == BYTE_0) ||
+                                (OtherDstSel == BYTE_1) ||
+                                (OtherDstSel == WORD_0));
+      break;
+    case BYTE_0: DstSelAgree = ((OtherDstSel == BYTE_1) ||
+                                (OtherDstSel == BYTE_2) ||
+                                (OtherDstSel == BYTE_3) ||
+                                (OtherDstSel == WORD_1));
+      break;
+    case BYTE_1: DstSelAgree = ((OtherDstSel == BYTE_0) ||
+                                (OtherDstSel == BYTE_2) ||
+                                (OtherDstSel == BYTE_3) ||
+                                (OtherDstSel == WORD_1));
+      break;
+    case BYTE_2: DstSelAgree = ((OtherDstSel == BYTE_0) ||
+                                (OtherDstSel == BYTE_1) ||
+                                (OtherDstSel == BYTE_3) ||
+                                (OtherDstSel == WORD_0));
+      break;
+    case BYTE_3: DstSelAgree = ((OtherDstSel == BYTE_0) ||
+                                (OtherDstSel == BYTE_1) ||
+                                (OtherDstSel == BYTE_2) ||
+                                (OtherDstSel == WORD_0));
+      break;
+    default: DstSelAgree = false;
+    }
+
+    if (!DstSelAgree)
+      break;
+
+    // Also OtherInst dst_unused should be UNUSED_PAD
+    DstUnused OtherDstUnused = static_cast<DstUnused>(
+      TII->getNamedImmOperand(*OtherInst, AMDGPU::OpName::dst_unused));
+    if (OtherDstUnused != DstUnused::UNUSED_PAD)
+      break;
+
+    // Create DstPreserveOperand
+    MachineOperand *OrDst = TII->getNamedOperand(MI, AMDGPU::OpName::vdst);
+    assert(OrDst && OrDst->isReg());
+
+    return make_unique<SDWADstPreserveOperand>(
+      OrDst, OrSDWADef, OrOtherDef, DstSel);
+
+  }
+  }
+
+  return std::unique_ptr<SDWAOperand>(nullptr);
+}
+
+void SIPeepholeSDWA::matchSDWAOperands(MachineFunction &MF) {
+  for (MachineBasicBlock &MBB : MF) {
+    for (MachineInstr &MI : MBB) {
+      if (auto Operand = matchSDWAOperand(MI)) {
+        DEBUG(dbgs() << "Match: " << MI << "To: " << *Operand << '\n');
+        SDWAOperands[&MI] = std::move(Operand);
+        ++NumSDWAPatternsFound;
       }
     }
   }
 }
 
 bool SIPeepholeSDWA::isConvertibleToSDWA(const MachineInstr &MI,
                                          const SISubtarget &ST) const {
-  // Check if this instruction has opcode that supports SDWA
+  // Check if this is already an SDWA instruction
   int Opc = MI.getOpcode();
+  if (TII->isSDWA(Opc))
+    return true;
+
+  // Check if this instruction has opcode that supports SDWA
   if (AMDGPU::getSDWAOp(Opc) == -1)
     Opc = AMDGPU::getVOPe32(Opc);
 
   if (Opc == -1 || AMDGPU::getSDWAOp(Opc) == -1)
     return false;
 
   if (!ST.hasSDWAOmod() && TII->hasModifiersSet(MI, AMDGPU::OpName::omod))
     return false;
if (!ST.hasSDWAMac() && (Opc == AMDGPU::V_MAC_F16_e32 ||
     return false;
 
   return true;
 }
 
 bool SIPeepholeSDWA::convertToSDWA(MachineInstr &MI,
                                    const SDWAOperandsVector &SDWAOperands) {
   // Convert to sdwa
-  int SDWAOpcode = AMDGPU::getSDWAOp(MI.getOpcode());
+  int SDWAOpcode = -1;
+  int Opcode = MI.getOpcode();
+  if (TII->isSDWA(Opcode))
+    SDWAOpcode = Opcode;
+  if (SDWAOpcode == -1)
+    SDWAOpcode = AMDGPU::getSDWAOp(Opcode);
   if (SDWAOpcode == -1)
-    SDWAOpcode = AMDGPU::getSDWAOp(AMDGPU::getVOPe32(MI.getOpcode()));
+    SDWAOpcode = AMDGPU::getSDWAOp(AMDGPU::getVOPe32(Opcode));
   assert(SDWAOpcode != -1);
 
   const MCInstrDesc &SDWADesc = TII->get(SDWAOpcode);
 
   // Create SDWA version of instruction MI and initialize its operands
   MachineInstrBuilder SDWAInst =
     BuildMI(*MI.getParent(), MI, MI.getDebugLoc(), SDWADesc);
 
if (AMDGPU::getNamedOperandIdx(SDWAOpcode, AMDGPU::OpName::omod) != -1) {
     MachineOperand *OMod = TII->getNamedOperand(MI, AMDGPU::OpName::omod);
     if (OMod) {
       SDWAInst.add(*OMod);
     } else {
       SDWAInst.addImm(0);
     }
   }
 
-  // Initialize dst_sel if present
+  // Copy dst_sel if present, initialize otherwise if needed
   if (AMDGPU::getNamedOperandIdx(SDWAOpcode, AMDGPU::OpName::dst_sel) != -1) {
+    MachineOperand *DstSel = TII->getNamedOperand(MI, AMDGPU::OpName::dst_sel);
+    if (DstSel) {
+      SDWAInst.add(*DstSel);
+    } else {
       SDWAInst.addImm(AMDGPU::SDWA::SdwaSel::DWORD);
     }
+  }
 
-  // Initialize dst_unused if present
+  // Copy dst_unused if present, initialize otherwise if needed
   if (AMDGPU::getNamedOperandIdx(SDWAOpcode, AMDGPU::OpName::dst_unused) != -1) {
+    MachineOperand *DstUnused = TII->getNamedOperand(MI, AMDGPU::OpName::dst_unused);
+    if (DstUnused) {
+      SDWAInst.add(*DstUnused);
+    } else {
       SDWAInst.addImm(AMDGPU::SDWA::DstUnused::UNUSED_PAD);
     }
+  }
 
-  // Initialize src0_sel
+  // Copy src0_sel if present, initialize otherwise
   assert(AMDGPU::getNamedOperandIdx(SDWAOpcode, AMDGPU::OpName::src0_sel) != -1);
+  MachineOperand *Src0Sel = TII->getNamedOperand(MI, AMDGPU::OpName::src0_sel);
+  if (Src0Sel) {
+    SDWAInst.add(*Src0Sel);
+  } else {
     SDWAInst.addImm(AMDGPU::SDWA::SdwaSel::DWORD);
+  }
 
-
-  // Initialize src1_sel if present
+  // Copy src1_sel if present, initialize otherwise if needed
   if (Src1) {
     assert(AMDGPU::getNamedOperandIdx(SDWAOpcode, AMDGPU::OpName::src1_sel) != -1);
+    MachineOperand *Src1Sel = TII->getNamedOperand(MI, AMDGPU::OpName::src1_sel);
+    if (Src1Sel) {
+      SDWAInst.add(*Src1Sel);
+    } else {
       SDWAInst.addImm(AMDGPU::SDWA::SdwaSel::DWORD);
     }
+  }
 
   // Apply all sdwa operand pattenrs
   bool Converted = false;
   for (auto &Operand : SDWAOperands) {
     // There should be no intesection between SDWA operands and potential MIs
     // e.g.:
     // v_and_b32 v0, 0xff, v1 -> src:v1 sel:BYTE_0
     // v_and_b32 v2, 0xff, v0 -> src:v0 sel:BYTE_0
     // v_add_u32 v3, v4, v2
     //
     // In that example it is possible that we would fold 2nd instruction into 3rd
     // (v_add_u32_sdwa) and then try to fold 1st instruction into 2nd (that was
     // already destroyed). So if SDWAOperand is also a potential MI then do not
     // apply it.
+
     if (PotentialMatches.count(Operand->getParentInst()) == 0)
       Converted |= Operand->convertToSDWA(*SDWAInst, TII);
   }
   if (Converted) {
     ConvertedInstructions.push_back(SDWAInst);
   } else {
     SDWAInst->eraseFromParent();
     return false;
bool SIPeepholeSDWA::runOnMachineFunction(MachineFunction &MF) {
   if (!ST.hasSDWA())
     return false;
 
   MRI = &MF.getRegInfo();
   TRI = ST.getRegisterInfo();
   TII = ST.getInstrInfo();
 
   // Find all SDWA operands in MF.
+  bool Changed = false;
+  bool Ret = false;
+  do {
     matchSDWAOperands(MF);
 
     for (const auto &OperandPair : SDWAOperands) {
       const auto &Operand = OperandPair.second;
       MachineInstr *PotentialMI = Operand->potentialToConvert(TII);
       if (PotentialMI && isConvertibleToSDWA(*PotentialMI, ST)) {
         PotentialMatches[PotentialMI].push_back(Operand.get());
       }
     }
 
     for (auto &PotentialPair : PotentialMatches) {
       MachineInstr &PotentialMI = *PotentialPair.first;
       convertToSDWA(PotentialMI, PotentialPair.second);
     }
 
     PotentialMatches.clear();
     SDWAOperands.clear();
 
-  bool Ret = !ConvertedInstructions.empty();
+    Changed = !ConvertedInstructions.empty();
+
+    if (Changed)
+      Ret = true;
+
     while (!ConvertedInstructions.empty())
       legalizeScalarOperands(*ConvertedInstructions.pop_back_val(), ST);
+  } while (Changed);
 
   return Ret;
 }

lib/Target/AMDGPU/SIRegisterInfo.h

const TargetRegisterClass *getEquivalentSGPRClass(
                                            const TargetRegisterClass *VRC) const;
 
   /// \returns The register class that is used for a sub-register of \p RC for
   /// the given \p SubIdx.  If \p SubIdx equals NoSubRegister, \p RC will
   /// be returned.
   const TargetRegisterClass *getSubRegClass(const TargetRegisterClass *RC,
                                             unsigned SubIdx) const;
 
+  bool isSameReg(const MachineOperand &LHS, const MachineOperand &RHS) const;
+
+  bool isSubregOf(const MachineOperand &SubReg,
+                  const MachineOperand &SuperReg) const;
+
   bool shouldRewriteCopySrc(const TargetRegisterClass *DefRC,
                             unsigned DefSubReg,
                             const TargetRegisterClass *SrcRC,
                             unsigned SrcSubReg) const override;
 
   /// \returns True if operands defined with this operand type can accept
   /// a literal constant (i.e. any 32-bit immediate).
   bool opCanUseLiteralConstant(unsigned OpType) const {

lib/Target/AMDGPU/SIRegisterInfo.cpp

case 8:
       return &AMDGPU::VReg_256RegClass;
     case 16: /* fall-through */
     default:
       llvm_unreachable("Invalid sub-register class size");
     }
   }
 }
 
+bool SIRegisterInfo::isSameReg(const MachineOperand &LHS,
+                               const MachineOperand &RHS) const {
+  return LHS.isReg() &&
+    RHS.isReg() &&
+    LHS.getReg() == RHS.getReg() &&
+    LHS.getSubReg() == RHS.getSubReg();
+}
+
+bool SIRegisterInfo::isSubregOf(const MachineOperand &SubReg,
+                                const MachineOperand &SuperReg) const {
+  if (!SuperReg.isReg() || !SubReg.isReg())
+    return false;
+
+  if (isSameReg(SuperReg, SubReg))
+    return true;
+
+  if (SuperReg.getReg() != SubReg.getReg())
+    return false;
+
+  LaneBitmask SuperMask = getSubRegIndexLaneMask(SuperReg.getSubReg());
+  LaneBitmask SubMask = getSubRegIndexLaneMask(SubReg.getSubReg());
+  SuperMask |= ~SubMask;
+  return SuperMask.all();
+}
+
 bool SIRegisterInfo::shouldRewriteCopySrc(
   const TargetRegisterClass *DefRC,
   unsigned DefSubReg,
   const TargetRegisterClass *SrcRC,
   unsigned SrcSubReg) const {
   // We want to prefer the smallest register class possible, so we don't want to
   // stop and rewrite on anything that looks like a subregister
   // extract. Operations mostly don't care about the super register class, so we

test/CodeGen/AMDGPU/fabs.f16.ll

 
 ; CI: v_cvt_f32_f16_e32
 ; CI: v_cvt_f32_f16_e32
 ; CI: v_mul_f32_e64 v{{[0-9]+}}, |v{{[0-9]+}}|, v{{[0-9]+}}
 ; CI: v_cvt_f16_f32
 ; CI: v_mul_f32_e64 v{{[0-9]+}}, |v{{[0-9]+}}|, v{{[0-9]+}}
 ; CI: v_cvt_f16_f32
 
-; VI: v_lshrrev_b32_e32 v{{[0-9]+}}, 16,
-; VI: v_mul_f16_sdwa v{{[0-9]+}}, |v{{[0-9]+}}|, v{{[0-9]+}} dst_sel:WORD_1 dst_unused:UNUSED_PAD src0_sel:WORD_1 src1_sel:DWORD
+; VI: v_mul_f16_sdwa v{{[0-9]+}}, |v{{[0-9]+}}|, v{{[0-9]+}} dst_sel:WORD_1 dst_unused:UNUSED_PAD src0_sel:WORD_1 src1_sel:WORD_1
 ; VI: v_mul_f16_e64 v{{[0-9]+}}, |v{{[0-9]+}}|, v{{[0-9]+}}
 
 ; GFX9: v_and_b32_e32 [[FABS:v[0-9]+]], 0x7fff7fff, [[VAL]]
 ; GFX9: v_pk_mul_f16 v{{[0-9]+}}, [[FABS]], v{{[0-9]+$}}
 define amdgpu_kernel void @v_fabs_fold_v2f16(<2 x half> addrspace(1)* %out, <2 x half> addrspace(1)* %in) #0 {
   %tid = call i32 @llvm.amdgcn.workitem.id.x()
   %gep = getelementptr <2 x half>, <2 x half> addrspace(1)* %in, i32 %tid
   %val = load <2 x half>, <2 x half> addrspace(1)* %gep

test/CodeGen/AMDGPU/fcanonicalize.f16.ll

 ; GCN: buffer_store_short [[REG]]
 define amdgpu_kernel void @test_fold_canonicalize_snan3_value_f16(half addrspace(1)* %out) #1 {
   %canonicalized = call half @llvm.canonicalize.f16(half 0xHFC01)
   store half %canonicalized, half addrspace(1)* %out
   ret void
 }
 
 ; GCN-LABEL: {{^}}v_test_canonicalize_var_v2f16:
-; VI-DAG: v_max_f16_sdwa [[REG0:v[0-9]+]], {{v[0-9]+}}, {{v[0-9]+}} dst_sel:WORD_1 dst_unused:UNUSED_PAD src0_sel:WORD_1 src1_sel:DWORD
+; VI-DAG: v_max_f16_sdwa [[REG0:v[0-9]+]], {{v[0-9]+}}, {{v[0-9]+}} dst_sel:WORD_1 dst_unused:UNUSED_PAD src0_sel:WORD_1 src1_sel:WORD_1
 ; VI-DAG: v_max_f16_e32 [[REG1:v[0-9]+]], {{v[0-9]+}}, {{v[0-9]+}}
 ; VI-NOT: v_and_b32
 
 ; GFX9: v_pk_max_f16 [[REG:v[0-9]+]], {{v[0-9]+}}, {{v[0-9]+$}}
 ; GFX9: buffer_store_dword [[REG]]
 define amdgpu_kernel void @v_test_canonicalize_var_v2f16(<2 x half> addrspace(1)* %out) #1 {
   %tid = call i32 @llvm.amdgcn.workitem.id.x()
   %gep = getelementptr <2 x half>, <2 x half> addrspace(1)* %out, i32 %tid
define amdgpu_kernel void @v_test_canonicalize_fabs_var_v2f16(<2 x half> addrspace(1)* %out) #1 {
   %val.fabs = call <2 x half> @llvm.fabs.v2f16(<2 x half> %val)
   %canonicalized = call <2 x half> @llvm.canonicalize.v2f16(<2 x half> %val.fabs)
   store <2 x half> %canonicalized, <2 x half> addrspace(1)* %out
   ret void
 }
 
 ; GCN-LABEL: {{^}}v_test_canonicalize_fneg_fabs_var_v2f16:
 ; VI-DAG: v_or_b32_e32 v{{[0-9]+}}, 0x80008000, v{{[0-9]+}}
-; VI-DAG: v_max_f16_sdwa [[REG0:v[0-9]+]], v{{[0-9]+}}, v{{[0-9]+}} dst_sel:WORD_1 dst_unused:UNUSED_PAD src0_sel:WORD_1 src1_sel:DWORD
+; VI-DAG: v_max_f16_sdwa [[REG0:v[0-9]+]], v{{[0-9]+}}, v{{[0-9]+}} dst_sel:WORD_1 dst_unused:UNUSED_PAD src0_sel:WORD_1 src1_sel:WORD_1
 ; VI-DAG: v_max_f16_e32 [[REG1:v[0-9]+]], v{{[0-9]+}}, v{{[0-9]+}}
 ; VI: v_or_b32
 
 ; GFX9: v_and_b32_e32 [[ABS:v[0-9]+]], 0x7fff7fff, v{{[0-9]+}}
 ; GFX9: v_pk_max_f16 [[REG:v[0-9]+]], [[ABS]], [[ABS]] neg_lo:[1,1] neg_hi:[1,1]{{$}}
 ; GCN: buffer_store_dword
 define amdgpu_kernel void @v_test_canonicalize_fneg_fabs_var_v2f16(<2 x half> addrspace(1)* %out) #1 {
   %tid = call i32 @llvm.amdgcn.workitem.id.x()
   %gep = getelementptr <2 x half>, <2 x half> addrspace(1)* %out, i32 %tid
   %val = load <2 x half>, <2 x half> addrspace(1)* %gep
   %val.fabs = call <2 x half> @llvm.fabs.v2f16(<2 x half> %val)
   %val.fabs.fneg = fsub <2 x half> <half -0.0, half -0.0>, %val.fabs
   %canonicalized = call <2 x half> @llvm.canonicalize.v2f16(<2 x half> %val.fabs.fneg)
   store <2 x half> %canonicalized, <2 x half> addrspace(1)* %out
   ret void
 }
 
 ; GCN-LABEL: {{^}}v_test_canonicalize_fneg_var_v2f16:
 ; VI:     v_xor_b32_e32 [[FNEG:v[0-9]+]], 0x80008000, v{{[0-9]+}}
-; VI:     v_lshrrev_b32_e32 [[FNEGHI:v[0-9]+]], 16, [[FNEG]]
-; VI-DAG: v_max_f16_sdwa [[REG1:v[0-9]+]], [[FNEG]], [[FNEGHI]] dst_sel:WORD_1 dst_unused:UNUSED_PAD src0_sel:WORD_1 src1_sel:DWORD
+; VI-DAG: v_max_f16_sdwa [[REG1:v[0-9]+]], [[FNEG]], [[FNEG]] dst_sel:WORD_1 dst_unused:UNUSED_PAD src0_sel:WORD_1 src1_sel:WORD_1
 ; VI-DAG: v_max_f16_e32 [[REG0:v[0-9]+]], [[FNEG]], [[FNEG]]
 ; VI-NOT: 0xffff
 
 ; GFX9: v_pk_max_f16 [[REG:v[0-9]+]], {{v[0-9]+}}, {{v[0-9]+}} neg_lo:[1,1] neg_hi:[1,1]{{$}}
 ; GFX9: buffer_store_dword [[REG]]
 define amdgpu_kernel void @v_test_canonicalize_fneg_var_v2f16(<2 x half> addrspace(1)* %out) #1 {
   %tid = call i32 @llvm.amdgcn.workitem.id.x()
   %gep = getelementptr <2 x half>, <2 x half> addrspace(1)* %out, i32 %tid

test/CodeGen/AMDGPU/fneg.f16.ll

 
 ; CI: v_cvt_f32_f16_e64 v{{[0-9]+}}, -v{{[0-9]+}}
 ; CI: v_cvt_f32_f16_e64 v{{[0-9]+}}, -v{{[0-9]+}}
 ; CI: v_mul_f32_e32 v{{[0-9]+}}, v{{[0-9]+}}, v{{[0-9]+}}
 ; CI: v_cvt_f16_f32
 ; CI: v_mul_f32_e32 v{{[0-9]+}}, v{{[0-9]+}}, v{{[0-9]+}}
 ; CI: v_cvt_f16_f32
 
-; VI: v_lshrrev_b32_e32 v{{[0-9]+}}, 16,
-; VI: v_mul_f16_sdwa v{{[0-9]+}}, -v{{[0-9]+}}, v{{[0-9]+}} dst_sel:WORD_1 dst_unused:UNUSED_PAD src0_sel:WORD_1 src1_sel:DWORD
+; VI: v_mul_f16_sdwa v{{[0-9]+}}, -v{{[0-9]+}}, v{{[0-9]+}} dst_sel:WORD_1 dst_unused:UNUSED_PAD src0_sel:WORD_1 src1_sel:WORD_1
 ; VI: v_mul_f16_e64 v{{[0-9]+}}, -v{{[0-9]+}}, v{{[0-9]+}}
 
 ; GFX9: v_pk_mul_f16 v{{[0-9]+}}, v{{[0-9]+}}, v{{[0-9]+}} neg_lo:[1,0] neg_hi:[1,0]{{$}}
 define amdgpu_kernel void @v_fneg_fold_v2f16(<2 x half> addrspace(1)* %out, <2 x half> addrspace(1)* %in) #0 {
   %val = load <2 x half>, <2 x half> addrspace(1)* %in
   %fsub = fsub <2 x half> <half -0.0, half -0.0>, %val
   %fmul = fmul <2 x half> %fsub, %val
   store <2 x half> %fmul, <2 x half> addrspace(1)* %out

test/CodeGen/AMDGPU/sdwa-merge-preserve.mir

+# RUN: llc -march=amdgcn -mcpu=fiji -run-pass=si-merge-sdwa-preserve -o - %s | FileCheck -check-prefix=SDWA %s
+# RUN: llc -march=amdgcn -mcpu=gfx900 -run-pass=si-merge-sdwa-preserve -o - %s | FileCheck -check-prefix=SDWA %s
+
+# SDWA-LABEL: {{^}}name: add_f16
+
+# SDWA: [[FIRST:%[0-9]+]] = FLAT_LOAD_DWORD %{{[0-9]+}}, 0, 0, 0, implicit %exec, implicit %flat_scr :: (load 4)
+# SDWA: [[SECOND:%[0-9]+]] = FLAT_LOAD_DWORD %{{[0-9]+}}, 0, 0, 0, implicit %exec, implicit %flat_scr :: (load 4)
+
+# SDWA: [[RES:%[0-9]+]] = V_ADD_F16_e64 0, [[FIRST]], 0, [[SECOND]], 0, 0, implicit %exec
+# SDWA: [[RES:%[0-9]+]] = V_ADD_F16_sdwa 0, [[FIRST]], 0, [[SECOND]], 0, 0, 5, 2, 5, 5, implicit %exec, implicit [[RES]]
+# SDWA: [[COPY:%[0-9]+]] = COPY [[RES]]
+
+# SDWA: FLAT_STORE_DWORD %{{[0-9]+}}, [[COPY:%[0-9]+]], 0, 0, 0, implicit %exec, implicit %flat_scr :: (store 4)
+
+---
+name:            add_f16
+tracksRegLiveness: true
+registers:       
+  - { id: 0, class: vreg_64 }
+  - { id: 1, class: vreg_64 }
+  - { id: 2, class: sreg_64 }
+  - { id: 3, class: vgpr_32 }
+  - { id: 4, class: vgpr_32 }
+  - { id: 5, class: sreg_32_xm0 }
+  - { id: 6, class: sreg_32 }
+  - { id: 7, class: sreg_32_xm0 }
+  - { id: 8, class: sreg_32 }
+  - { id: 9, class: vgpr_32 }
+  - { id: 10, class: vgpr_32 }
+  - { id: 11, class: vgpr_32 }
+  - { id: 12, class: sreg_32_xm0 }
+body:             |
+  bb.0:
+    liveins: %vgpr0_vgpr1, %vgpr2_vgpr3, %sgpr30_sgpr31
+  
+    %2 = COPY %sgpr30_sgpr31
+    %1 = COPY %vgpr2_vgpr3
+    %0 = COPY %vgpr0_vgpr1
+    ; This is needed to force MIR to treat this as non-SSA
+    %0 = COPY %vgpr0_vgpr1
+    %3 = FLAT_LOAD_DWORD %0, 0, 0, 0, implicit %exec, implicit %flat_scr :: (load 4)
+    %4 = FLAT_LOAD_DWORD %1, 0, 0, 0, implicit %exec, implicit %flat_scr :: (load 4)
+
+    %10 = V_ADD_F16_e64 0, %3, 0, %4, 0, 0, implicit %exec
+    %11 = V_ADD_F16_sdwa 0, %3, 0, %4, 0, 0, 5, 2, 5, 5, implicit %exec, implicit %10
+
+    FLAT_STORE_DWORD %0, %11, 0, 0, 0, implicit %exec, implicit %flat_scr :: (store 4)
+    %sgpr30_sgpr31 = COPY %2
+    S_SETPC_B64_return %sgpr30_sgpr31
+
+...
+
+# SDWA-LABEL: {{^}}name: add_f16_no_merge
+
+# SDWA: [[FIRST:%[0-9]+]] = FLAT_LOAD_DWORD %{{[0-9]+}}, 0, 0, 0, implicit %exec, implicit %flat_scr :: (load 4)
+# SDWA: [[SECOND:%[0-9]+]] = FLAT_LOAD_DWORD %{{[0-9]+}}, 0, 0, 0, implicit %exec, implicit %flat_scr :: (load 4)
+
+# SDWA: [[LOW:%[0-9]+]] = V_ADD_F16_e64 0, [[FIRST]], 0, [[SECOND]], 0, 0, implicit %exec
+# SDWA: [[HIGH:%[0-9]+]] = V_ADD_F16_sdwa 0, [[FIRST]], 0, [[SECOND]], 0, 0, 5, 2, 5, 5, implicit %exec
+
+# SDWA: FLAT_STORE_DWORD %{{[0-9]+}}, [[HIGH:%[0-9]+]], 0, 0, 0, implicit %exec, implicit %flat_scr :: (store 4)
+
+---
+name:            add_f16_no_merge
+tracksRegLiveness: true
+registers:       
+  - { id: 0, class: vreg_64 }
+  - { id: 1, class: vreg_64 }
+  - { id: 2, class: sreg_64 }
+  - { id: 3, class: vgpr_32 }
+  - { id: 4, class: vgpr_32 }
+  - { id: 5, class: sreg_32_xm0 }
+  - { id: 6, class: sreg_32 }
+  - { id: 7, class: sreg_32_xm0 }
+  - { id: 8, class: sreg_32 }
+  - { id: 9, class: vgpr_32 }
+  - { id: 10, class: vgpr_32 }
+  - { id: 11, class: vgpr_32 }
+  - { id: 12, class: sreg_32_xm0 }
+body:             |
+  bb.0:
+    liveins: %vgpr0_vgpr1, %vgpr2_vgpr3, %sgpr30_sgpr31
+  
+    %2 = COPY %sgpr30_sgpr31
+    %1 = COPY %vgpr2_vgpr3
+    %0 = COPY %vgpr0_vgpr1
+    ; This is needed to force MIR to treat this as non-SSA
+    %0 = COPY %vgpr0_vgpr1
+    %3 = FLAT_LOAD_DWORD %0, 0, 0, 0, implicit %exec, implicit %flat_scr :: (load 4)
+    %4 = FLAT_LOAD_DWORD %1, 0, 0, 0, implicit %exec, implicit %flat_scr :: (load 4)
+
+    %10 = V_ADD_F16_e64 0, %3, 0, %4, 0, 0, implicit %exec
+    %11 = V_ADD_F16_sdwa 0, %3, 0, %4, 0, 0, 5, 2, 5, 5, implicit %exec
+
+    FLAT_STORE_DWORD %0, %11, 0, 0, 0, implicit %exec, implicit %flat_scr :: (store 4)
+    %sgpr30_sgpr31 = COPY %2
+    S_SETPC_B64_return %sgpr30_sgpr31
+

test/CodeGen/AMDGPU/sdwa-peephole-instr.mir

bb.0:
     %sgpr30_sgpr31 = COPY %2
     S_SETPC_B64_return %sgpr30_sgpr31
 
 ...
 ---
 # GCN-LABEL: {{^}}name: vop2_instructions
 
 
-# VI: %{{[0-9]+}} = V_AND_B32_sdwa 0, %{{[0-9]+}}, 0, %{{[0-9]+}}, 0, 6, 0, 6, 5, implicit %exec
+# VI: %{{[0-9]+}} = V_AND_B32_sdwa 0, %{{[0-9]+}}, 0, %{{[0-9]+}}, 0, 5, 0, 6, 5, implicit %exec
 # VI: %{{[0-9]+}} = V_ADD_F32_sdwa 0, %{{[0-9]+}}, 0, %{{[0-9]+}}, 0, 0, 5, 0, 5, 1, implicit %exec
 # VI: %{{[0-9]+}} = V_SUB_F16_sdwa 0, %{{[0-9]+}}, 0, %{{[0-9]+}}, 0, 0, 6, 0, 5, 1, implicit %exec
 # VI: %{{[0-9]+}} = V_MAC_F32_sdwa 0, %{{[0-9]+}}, 0, %{{[0-9]+}}, %{{[0-9]+}}, 0, 0, 6, 0, 6, 1, implicit %exec
 # VI: %{{[0-9]+}} = V_MAC_F16_sdwa 0, %{{[0-9]+}}, 0, %{{[0-9]+}}, %{{[0-9]+}}, 0, 0, 6, 0, 5, 1, implicit %exec
 
-# GFX9: %{{[0-9]+}} = V_AND_B32_sdwa 0, %{{[0-9]+}}, 0, %{{[0-9]+}}, 0, 6, 0, 6, 5, implicit %exec
+# GFX9: %{{[0-9]+}} = V_AND_B32_sdwa 0, %{{[0-9]+}}, 0, %{{[0-9]+}}, 0, 5, 0, 6, 5, implicit %exec
 # GFX9: %{{[0-9]+}} = V_ADD_F32_sdwa 0, %{{[0-9]+}}, 0, %{{[0-9]+}}, 0, 0, 5, 0, 5, 1, implicit %exec
 # GFX9: %{{[0-9]+}} = V_SUB_F16_sdwa 0, %{{[0-9]+}}, 0, %{{[0-9]+}}, 0, 0, 6, 0, 5, 1, implicit %exec
 # GFX9: %{{[0-9]+}} = V_MAC_F32_e32 %{{[0-9]+}}, %{{[0-9]+}}, %{{[0-9]+}}, implicit %exec
 # GFX9: %{{[0-9]+}} = V_MAC_F16_e32 %{{[0-9]+}}, %{{[0-9]+}}, %{{[0-9]+}}, implicit %exec
 
 
 # VI: %{{[0-9]+}} = V_AND_B32_sdwa 0, %{{[0-9]+}}, 0, %{{[0-9]+}}, 0, 5, 0, 6, 5, implicit %exec
 # VI: %{{[0-9]+}} = V_ADD_F32_sdwa 0, %{{[0-9]+}}, 0, %{{[0-9]+}}, 0, 0, 5, 0, 5, 1, implicit %exec

test/CodeGen/AMDGPU/sdwa-preserve.mir

+# RUN: llc -march=amdgcn -mcpu=fiji -start-before=si-peephole-sdwa -o - %s | FileCheck -check-prefix=SDWA %s
+# RUN: llc -march=amdgcn -mcpu=gfx900 -start-before=si-peephole-sdwa -o - %s | FileCheck -check-prefix=SDWA %s
+
+# SDWA-LABEL: {{^}}add_f16_u32_preserve
+
+# SDWA: flat_load_dword [[FIRST:v[0-9]+]], v[{{[0-9]+}}:{{[0-9]+}}]
+# SDWA: flat_load_dword [[SECOND:v[0-9]+]], v[{{[0-9]+}}:{{[0-9]+}}]
+
+# SDWA: v_add_u32_sdwa [[RES:v[0-9]+]], [[FIRST]], [[SECOND]] dst_sel:WORD_1 dst_unused:UNUSED_PAD src0_sel:BYTE_1 src1_sel:BYTE_3
+# SDWA: v_add_f16_sdwa [[RES:v[0-9]+]], [[FIRST]], [[SECOND]] dst_sel:BYTE_1 dst_unused:UNUSED_PRESERVE src0_sel:WORD_0 src1_sel:WORD_1
+
+# SDWA: flat_store_dword v[{{[0-9]+}}:{{[0-9]+}}], [[RES]]
+
+---
+name:            add_f16_u32_preserve
+tracksRegLiveness: true
+registers:       
+  - { id: 0, class: vreg_64 }
+  - { id: 1, class: vreg_64 }
+  - { id: 2, class: sreg_64 }
+  - { id: 3, class: vgpr_32 }
+  - { id: 4, class: vgpr_32 }
+  - { id: 5, class: vgpr_32 }
+  - { id: 6, class: vgpr_32 }
+  - { id: 7, class: vgpr_32 }
+  - { id: 8, class: vgpr_32 }
+  - { id: 9, class: vgpr_32 }
+  - { id: 10, class: vgpr_32 }
+  - { id: 11, class: vgpr_32 }
+  - { id: 12, class: vgpr_32 }
+  - { id: 13, class: vgpr_32 }
+body:             |
+  bb.0:
+    liveins: %vgpr0_vgpr1, %vgpr2_vgpr3, %sgpr30_sgpr31
+  
+    %2 = COPY %sgpr30_sgpr31
+    %1 = COPY %vgpr2_vgpr3
+    %0 = COPY %vgpr0_vgpr1
+    %3 = FLAT_LOAD_DWORD %0, 0, 0, 0, implicit %exec, implicit %flat_scr :: (load 4)
+    %4 = FLAT_LOAD_DWORD %1, 0, 0, 0, implicit %exec, implicit %flat_scr :: (load 4)
+
+    %5 = V_AND_B32_e32 65535, %3, implicit %exec
+    %6 = V_LSHRREV_B32_e64 16, %4, implicit %exec
+    %7 = V_BFE_U32 %3, 8, 8, implicit %exec
+    %8 = V_LSHRREV_B32_e32 24, %4, implicit %exec
+
+    %9 = V_ADD_F16_e64 0, %5, 0, %6, 0, 0, implicit %exec
+    %10 = V_LSHLREV_B16_e64 8, %9, implicit %exec
+    %11 = V_ADD_U32_e64 %7, %8, implicit %exec
+    %12 = V_LSHLREV_B32_e64 16, %11, implicit %exec
+
+    %13 = V_OR_B32_e64 %10, %12, implicit %exec
+
+    FLAT_STORE_DWORD %0, %13, 0, 0, 0, implicit %exec, implicit %flat_scr :: (store 4)
+    %sgpr30_sgpr31 = COPY %2
+    S_SETPC_B64_return %sgpr30_sgpr31