1 //===- SIPeepholeSDWA.cpp - Peephole optimization for SDWA instructions ---===//
3 // The LLVM Compiler Infrastructure
5 // This file is distributed under the University of Illinois Open Source
6 // License. See LICENSE.TXT for details.
8 //===----------------------------------------------------------------------===//
10 /// \file This pass tries to apply several peephole SDWA patterns.
13 /// V_LSHRREV_B32_e32 %0, 16, %1
14 /// V_ADD_I32_e32 %2, %0, %3
15 /// V_LSHLREV_B32_e32 %4, 16, %2
18 /// V_ADD_I32_sdwa %4, %1, %3
19 /// dst_sel:WORD_1 dst_unused:UNUSED_PAD src0_sel:WORD_1 src1_sel:DWORD
21 //===----------------------------------------------------------------------===//
24 #include "AMDGPUSubtarget.h"
25 #include "SIDefines.h"
26 #include "SIInstrInfo.h"
27 #include "SIRegisterInfo.h"
28 #include "Utils/AMDGPUBaseInfo.h"
29 #include "llvm/ADT/None.h"
30 #include "llvm/ADT/Optional.h"
31 #include "llvm/ADT/STLExtras.h"
32 #include "llvm/ADT/SmallVector.h"
33 #include "llvm/ADT/Statistic.h"
34 #include "llvm/CodeGen/MachineBasicBlock.h"
35 #include "llvm/CodeGen/MachineFunction.h"
36 #include "llvm/CodeGen/MachineFunctionPass.h"
37 #include "llvm/CodeGen/MachineInstr.h"
38 #include "llvm/CodeGen/MachineInstrBuilder.h"
39 #include "llvm/CodeGen/MachineOperand.h"
40 #include "llvm/CodeGen/MachineRegisterInfo.h"
41 #include "llvm/CodeGen/TargetRegisterInfo.h"
42 #include "llvm/MC/LaneBitmask.h"
43 #include "llvm/MC/MCInstrDesc.h"
44 #include "llvm/Pass.h"
45 #include "llvm/Support/Debug.h"
46 #include "llvm/Support/raw_ostream.h"
51 #include <unordered_map>
55 #define DEBUG_TYPE "si-peephole-sdwa"
57 STATISTIC(NumSDWAPatternsFound, "Number of SDWA patterns found.");
58 STATISTIC(NumSDWAInstructionsPeepholed,
59 "Number of instruction converted to SDWA.");
66 class SIPeepholeSDWA : public MachineFunctionPass {
68 using SDWAOperandsVector = SmallVector<SDWAOperand *, 4>;
71 MachineRegisterInfo *MRI;
72 const SIRegisterInfo *TRI;
73 const SIInstrInfo *TII;
75 std::unordered_map<MachineInstr *, std::unique_ptr<SDWAOperand>> SDWAOperands;
76 std::unordered_map<MachineInstr *, SDWAOperandsVector> PotentialMatches;
77 SmallVector<MachineInstr *, 8> ConvertedInstructions;
79 Optional<int64_t> foldToImm(const MachineOperand &Op) const;
84 SIPeepholeSDWA() : MachineFunctionPass(ID) {
85 initializeSIPeepholeSDWAPass(*PassRegistry::getPassRegistry());
88 bool runOnMachineFunction(MachineFunction &MF) override;
89 void matchSDWAOperands(MachineFunction &MF);
90 std::unique_ptr<SDWAOperand> matchSDWAOperand(MachineInstr &MI);
91 bool isConvertibleToSDWA(const MachineInstr &MI, const SISubtarget &ST) const;
92 bool convertToSDWA(MachineInstr &MI, const SDWAOperandsVector &SDWAOperands);
93 void legalizeScalarOperands(MachineInstr &MI, const SISubtarget &ST) const;
95 StringRef getPassName() const override { return "SI Peephole SDWA"; }
97 void getAnalysisUsage(AnalysisUsage &AU) const override {
99 MachineFunctionPass::getAnalysisUsage(AU);
105 MachineOperand *Target; // Operand that would be used in converted instruction
106 MachineOperand *Replaced; // Operand that would be replace by Target
109 SDWAOperand(MachineOperand *TargetOp, MachineOperand *ReplacedOp)
110 : Target(TargetOp), Replaced(ReplacedOp) {
111 assert(Target->isReg());
112 assert(Replaced->isReg());
115 virtual ~SDWAOperand() = default;
117 virtual MachineInstr *potentialToConvert(const SIInstrInfo *TII) = 0;
118 virtual bool convertToSDWA(MachineInstr &MI, const SIInstrInfo *TII) = 0;
120 MachineOperand *getTargetOperand() const { return Target; }
121 MachineOperand *getReplacedOperand() const { return Replaced; }
122 MachineInstr *getParentInst() const { return Target->getParent(); }
124 MachineRegisterInfo *getMRI() const {
125 return &getParentInst()->getParent()->getParent()->getRegInfo();
128 #if !defined(NDEBUG) || defined(LLVM_ENABLE_DUMP)
129 virtual void print(raw_ostream& OS) const = 0;
130 void dump() const { print(dbgs()); }
134 using namespace AMDGPU::SDWA;
136 class SDWASrcOperand : public SDWAOperand {
144 SDWASrcOperand(MachineOperand *TargetOp, MachineOperand *ReplacedOp,
145 SdwaSel SrcSel_ = DWORD, bool Abs_ = false, bool Neg_ = false,
147 : SDWAOperand(TargetOp, ReplacedOp),
148 SrcSel(SrcSel_), Abs(Abs_), Neg(Neg_), Sext(Sext_) {}
150 MachineInstr *potentialToConvert(const SIInstrInfo *TII) override;
151 bool convertToSDWA(MachineInstr &MI, const SIInstrInfo *TII) override;
153 SdwaSel getSrcSel() const { return SrcSel; }
154 bool getAbs() const { return Abs; }
155 bool getNeg() const { return Neg; }
156 bool getSext() const { return Sext; }
158 uint64_t getSrcMods(const SIInstrInfo *TII,
159 const MachineOperand *SrcOp) const;
161 #if !defined(NDEBUG) || defined(LLVM_ENABLE_DUMP)
162 void print(raw_ostream& OS) const override;
166 class SDWADstOperand : public SDWAOperand {
173 SDWADstOperand(MachineOperand *TargetOp, MachineOperand *ReplacedOp,
174 SdwaSel DstSel_ = DWORD, DstUnused DstUn_ = UNUSED_PAD)
175 : SDWAOperand(TargetOp, ReplacedOp), DstSel(DstSel_), DstUn(DstUn_) {}
177 MachineInstr *potentialToConvert(const SIInstrInfo *TII) override;
178 bool convertToSDWA(MachineInstr &MI, const SIInstrInfo *TII) override;
180 SdwaSel getDstSel() const { return DstSel; }
181 DstUnused getDstUnused() const { return DstUn; }
183 #if !defined(NDEBUG) || defined(LLVM_ENABLE_DUMP)
184 void print(raw_ostream& OS) const override;
188 class SDWADstPreserveOperand : public SDWADstOperand {
190 MachineOperand *Preserve;
193 SDWADstPreserveOperand(MachineOperand *TargetOp, MachineOperand *ReplacedOp,
194 MachineOperand *PreserveOp, SdwaSel DstSel_ = DWORD)
195 : SDWADstOperand(TargetOp, ReplacedOp, DstSel_, UNUSED_PRESERVE),
196 Preserve(PreserveOp) {}
198 bool convertToSDWA(MachineInstr &MI, const SIInstrInfo *TII) override;
200 MachineOperand *getPreservedOperand() const { return Preserve; }
202 #if !defined(NDEBUG) || defined(LLVM_ENABLE_DUMP)
203 void print(raw_ostream& OS) const override;
207 } // end anonymous namespace
209 INITIALIZE_PASS(SIPeepholeSDWA, DEBUG_TYPE, "SI Peephole SDWA", false, false)
211 char SIPeepholeSDWA::ID = 0;
213 char &llvm::SIPeepholeSDWAID = SIPeepholeSDWA::ID;
215 FunctionPass *llvm::createSIPeepholeSDWAPass() {
216 return new SIPeepholeSDWA();
220 #if !defined(NDEBUG) || defined(LLVM_ENABLE_DUMP)
221 static raw_ostream& operator<<(raw_ostream &OS, const SdwaSel &Sel) {
223 case BYTE_0: OS << "BYTE_0"; break;
224 case BYTE_1: OS << "BYTE_1"; break;
225 case BYTE_2: OS << "BYTE_2"; break;
226 case BYTE_3: OS << "BYTE_3"; break;
227 case WORD_0: OS << "WORD_0"; break;
228 case WORD_1: OS << "WORD_1"; break;
229 case DWORD: OS << "DWORD"; break;
234 static raw_ostream& operator<<(raw_ostream &OS, const DstUnused &Un) {
236 case UNUSED_PAD: OS << "UNUSED_PAD"; break;
237 case UNUSED_SEXT: OS << "UNUSED_SEXT"; break;
238 case UNUSED_PRESERVE: OS << "UNUSED_PRESERVE"; break;
243 static raw_ostream& operator<<(raw_ostream &OS, const SDWAOperand &Operand) {
249 void SDWASrcOperand::print(raw_ostream& OS) const {
250 OS << "SDWA src: " << *getTargetOperand()
251 << " src_sel:" << getSrcSel()
252 << " abs:" << getAbs() << " neg:" << getNeg()
253 << " sext:" << getSext() << '\n';
257 void SDWADstOperand::print(raw_ostream& OS) const {
258 OS << "SDWA dst: " << *getTargetOperand()
259 << " dst_sel:" << getDstSel()
260 << " dst_unused:" << getDstUnused() << '\n';
264 void SDWADstPreserveOperand::print(raw_ostream& OS) const {
265 OS << "SDWA preserve dst: " << *getTargetOperand()
266 << " dst_sel:" << getDstSel()
267 << " preserve:" << *getPreservedOperand() << '\n';
272 static void copyRegOperand(MachineOperand &To, const MachineOperand &From) {
273 assert(To.isReg() && From.isReg());
274 To.setReg(From.getReg());
275 To.setSubReg(From.getSubReg());
276 To.setIsUndef(From.isUndef());
278 To.setIsKill(From.isKill());
280 To.setIsDead(From.isDead());
284 static bool isSameReg(const MachineOperand &LHS, const MachineOperand &RHS) {
285 return LHS.isReg() &&
287 LHS.getReg() == RHS.getReg() &&
288 LHS.getSubReg() == RHS.getSubReg();
291 static MachineOperand *findSingleRegUse(const MachineOperand *Reg,
292 const MachineRegisterInfo *MRI) {
293 if (!Reg->isReg() || !Reg->isDef())
296 MachineOperand *ResMO = nullptr;
297 for (MachineOperand &UseMO : MRI->use_nodbg_operands(Reg->getReg())) {
298 // If there exist use of subreg of Reg then return nullptr
299 if (!isSameReg(UseMO, *Reg))
302 // Check that there is only one instruction that uses Reg
305 } else if (ResMO->getParent() != UseMO.getParent()) {
313 static MachineOperand *findSingleRegDef(const MachineOperand *Reg,
314 const MachineRegisterInfo *MRI) {
318 MachineInstr *DefInstr = MRI->getUniqueVRegDef(Reg->getReg());
322 for (auto &DefMO : DefInstr->defs()) {
323 if (DefMO.isReg() && DefMO.getReg() == Reg->getReg())
327 // Ignore implicit defs.
331 uint64_t SDWASrcOperand::getSrcMods(const SIInstrInfo *TII,
332 const MachineOperand *SrcOp) const {
334 const auto *MI = SrcOp->getParent();
335 if (TII->getNamedOperand(*MI, AMDGPU::OpName::src0) == SrcOp) {
336 if (auto *Mod = TII->getNamedOperand(*MI, AMDGPU::OpName::src0_modifiers)) {
337 Mods = Mod->getImm();
339 } else if (TII->getNamedOperand(*MI, AMDGPU::OpName::src1) == SrcOp) {
340 if (auto *Mod = TII->getNamedOperand(*MI, AMDGPU::OpName::src1_modifiers)) {
341 Mods = Mod->getImm();
346 "Float and integer src modifiers can't be set simulteniously");
347 Mods |= Abs ? SISrcMods::ABS : 0;
348 Mods ^= Neg ? SISrcMods::NEG : 0;
350 Mods |= SISrcMods::SEXT;
356 MachineInstr *SDWASrcOperand::potentialToConvert(const SIInstrInfo *TII) {
357 // For SDWA src operand potential instruction is one that use register
358 // defined by parent instruction
359 MachineOperand *PotentialMO = findSingleRegUse(getReplacedOperand(), getMRI());
363 return PotentialMO->getParent();
366 bool SDWASrcOperand::convertToSDWA(MachineInstr &MI, const SIInstrInfo *TII) {
367 // Find operand in instruction that matches source operand and replace it with
368 // target operand. Set corresponding src_sel
370 MachineOperand *Src = TII->getNamedOperand(MI, AMDGPU::OpName::src0);
371 MachineOperand *SrcSel = TII->getNamedOperand(MI, AMDGPU::OpName::src0_sel);
372 MachineOperand *SrcMods =
373 TII->getNamedOperand(MI, AMDGPU::OpName::src0_modifiers);
374 assert(Src && (Src->isReg() || Src->isImm()));
375 if (!isSameReg(*Src, *getReplacedOperand())) {
376 // If this is not src0 then it should be src1
377 Src = TII->getNamedOperand(MI, AMDGPU::OpName::src1);
378 SrcSel = TII->getNamedOperand(MI, AMDGPU::OpName::src1_sel);
379 SrcMods = TII->getNamedOperand(MI, AMDGPU::OpName::src1_modifiers);
381 assert(Src && Src->isReg());
383 if ((MI.getOpcode() == AMDGPU::V_MAC_F16_sdwa ||
384 MI.getOpcode() == AMDGPU::V_MAC_F32_sdwa) &&
385 !isSameReg(*Src, *getReplacedOperand())) {
386 // In case of v_mac_f16/32_sdwa this pass can try to apply src operand to
387 // src2. This is not allowed.
391 assert(isSameReg(*Src, *getReplacedOperand()) && SrcSel && SrcMods);
393 copyRegOperand(*Src, *getTargetOperand());
394 SrcSel->setImm(getSrcSel());
395 SrcMods->setImm(getSrcMods(TII, Src));
396 getTargetOperand()->setIsKill(false);
400 MachineInstr *SDWADstOperand::potentialToConvert(const SIInstrInfo *TII) {
401 // For SDWA dst operand potential instruction is one that defines register
402 // that this operand uses
403 MachineRegisterInfo *MRI = getMRI();
404 MachineInstr *ParentMI = getParentInst();
406 MachineOperand *PotentialMO = findSingleRegDef(getReplacedOperand(), MRI);
410 // Check that ParentMI is the only instruction that uses replaced register
411 for (MachineInstr &UseInst : MRI->use_nodbg_instructions(PotentialMO->getReg())) {
412 if (&UseInst != ParentMI)
416 return PotentialMO->getParent();
419 bool SDWADstOperand::convertToSDWA(MachineInstr &MI, const SIInstrInfo *TII) {
420 // Replace vdst operand in MI with target operand. Set dst_sel and dst_unused
422 if ((MI.getOpcode() == AMDGPU::V_MAC_F16_sdwa ||
423 MI.getOpcode() == AMDGPU::V_MAC_F32_sdwa) &&
424 getDstSel() != AMDGPU::SDWA::DWORD) {
425 // v_mac_f16/32_sdwa allow dst_sel to be equal only to DWORD
429 MachineOperand *Operand = TII->getNamedOperand(MI, AMDGPU::OpName::vdst);
432 isSameReg(*Operand, *getReplacedOperand()));
433 copyRegOperand(*Operand, *getTargetOperand());
434 MachineOperand *DstSel= TII->getNamedOperand(MI, AMDGPU::OpName::dst_sel);
436 DstSel->setImm(getDstSel());
437 MachineOperand *DstUnused= TII->getNamedOperand(MI, AMDGPU::OpName::dst_unused);
439 DstUnused->setImm(getDstUnused());
441 // Remove original instruction because it would conflict with our new
442 // instruction by register definition
443 getParentInst()->eraseFromParent();
447 bool SDWADstPreserveOperand::convertToSDWA(MachineInstr &MI,
448 const SIInstrInfo *TII) {
449 // MI should be moved right before v_or_b32.
450 // For this we should clear all kill flags on uses of MI src-operands or else
451 // we can encounter problem with use of killed operand.
452 for (MachineOperand &MO : MI.uses()) {
455 getMRI()->clearKillFlags(MO.getReg());
458 // Move MI before v_or_b32
459 auto MBB = MI.getParent();
461 MBB->insert(getParentInst(), &MI);
463 // Add Implicit use of preserved register
464 MachineInstrBuilder MIB(*MBB->getParent(), MI);
465 MIB.addReg(getPreservedOperand()->getReg(),
466 RegState::ImplicitKill,
467 getPreservedOperand()->getSubReg());
469 // Tie dst to implicit use
470 MI.tieOperands(AMDGPU::getNamedOperandIdx(MI.getOpcode(), AMDGPU::OpName::vdst),
471 MI.getNumOperands() - 1);
473 // Convert MI as any other SDWADstOperand and remove v_or_b32
474 return SDWADstOperand::convertToSDWA(MI, TII);
477 Optional<int64_t> SIPeepholeSDWA::foldToImm(const MachineOperand &Op) const {
482 // If this is not immediate then it can be copy of immediate value, e.g.:
483 // %1 = S_MOV_B32 255;
485 for (const MachineOperand &Def : MRI->def_operands(Op.getReg())) {
486 if (!isSameReg(Op, Def))
489 const MachineInstr *DefInst = Def.getParent();
490 if (!TII->isFoldableCopy(*DefInst))
493 const MachineOperand &Copied = DefInst->getOperand(1);
497 return Copied.getImm();
504 std::unique_ptr<SDWAOperand>
505 SIPeepholeSDWA::matchSDWAOperand(MachineInstr &MI) {
506 unsigned Opcode = MI.getOpcode();
508 case AMDGPU::V_LSHRREV_B32_e32:
509 case AMDGPU::V_ASHRREV_I32_e32:
510 case AMDGPU::V_LSHLREV_B32_e32:
511 case AMDGPU::V_LSHRREV_B32_e64:
512 case AMDGPU::V_ASHRREV_I32_e64:
513 case AMDGPU::V_LSHLREV_B32_e64: {
514 // from: v_lshrrev_b32_e32 v1, 16/24, v0
515 // to SDWA src:v0 src_sel:WORD_1/BYTE_3
517 // from: v_ashrrev_i32_e32 v1, 16/24, v0
518 // to SDWA src:v0 src_sel:WORD_1/BYTE_3 sext:1
520 // from: v_lshlrev_b32_e32 v1, 16/24, v0
521 // to SDWA dst:v1 dst_sel:WORD_1/BYTE_3 dst_unused:UNUSED_PAD
522 MachineOperand *Src0 = TII->getNamedOperand(MI, AMDGPU::OpName::src0);
523 auto Imm = foldToImm(*Src0);
527 if (*Imm != 16 && *Imm != 24)
530 MachineOperand *Src1 = TII->getNamedOperand(MI, AMDGPU::OpName::src1);
531 MachineOperand *Dst = TII->getNamedOperand(MI, AMDGPU::OpName::vdst);
532 if (TRI->isPhysicalRegister(Src1->getReg()) ||
533 TRI->isPhysicalRegister(Dst->getReg()))
536 if (Opcode == AMDGPU::V_LSHLREV_B32_e32 ||
537 Opcode == AMDGPU::V_LSHLREV_B32_e64) {
538 return make_unique<SDWADstOperand>(
539 Dst, Src1, *Imm == 16 ? WORD_1 : BYTE_3, UNUSED_PAD);
541 return make_unique<SDWASrcOperand>(
542 Src1, Dst, *Imm == 16 ? WORD_1 : BYTE_3, false, false,
543 Opcode != AMDGPU::V_LSHRREV_B32_e32 &&
544 Opcode != AMDGPU::V_LSHRREV_B32_e64);
549 case AMDGPU::V_LSHRREV_B16_e32:
550 case AMDGPU::V_ASHRREV_I16_e32:
551 case AMDGPU::V_LSHLREV_B16_e32:
552 case AMDGPU::V_LSHRREV_B16_e64:
553 case AMDGPU::V_ASHRREV_I16_e64:
554 case AMDGPU::V_LSHLREV_B16_e64: {
555 // from: v_lshrrev_b16_e32 v1, 8, v0
556 // to SDWA src:v0 src_sel:BYTE_1
558 // from: v_ashrrev_i16_e32 v1, 8, v0
559 // to SDWA src:v0 src_sel:BYTE_1 sext:1
561 // from: v_lshlrev_b16_e32 v1, 8, v0
562 // to SDWA dst:v1 dst_sel:BYTE_1 dst_unused:UNUSED_PAD
563 MachineOperand *Src0 = TII->getNamedOperand(MI, AMDGPU::OpName::src0);
564 auto Imm = foldToImm(*Src0);
565 if (!Imm || *Imm != 8)
568 MachineOperand *Src1 = TII->getNamedOperand(MI, AMDGPU::OpName::src1);
569 MachineOperand *Dst = TII->getNamedOperand(MI, AMDGPU::OpName::vdst);
571 if (TRI->isPhysicalRegister(Src1->getReg()) ||
572 TRI->isPhysicalRegister(Dst->getReg()))
575 if (Opcode == AMDGPU::V_LSHLREV_B16_e32 ||
576 Opcode == AMDGPU::V_LSHLREV_B16_e64) {
577 return make_unique<SDWADstOperand>(Dst, Src1, BYTE_1, UNUSED_PAD);
579 return make_unique<SDWASrcOperand>(
580 Src1, Dst, BYTE_1, false, false,
581 Opcode != AMDGPU::V_LSHRREV_B16_e32 &&
582 Opcode != AMDGPU::V_LSHRREV_B16_e64);
587 case AMDGPU::V_BFE_I32:
588 case AMDGPU::V_BFE_U32: {
590 // from: v_bfe_u32 v1, v0, 8, 8
591 // to SDWA src:v0 src_sel:BYTE_1
593 // offset | width | src_sel
594 // ------------------------
603 MachineOperand *Src1 = TII->getNamedOperand(MI, AMDGPU::OpName::src1);
604 auto Offset = foldToImm(*Src1);
608 MachineOperand *Src2 = TII->getNamedOperand(MI, AMDGPU::OpName::src2);
609 auto Width = foldToImm(*Src2);
613 SdwaSel SrcSel = DWORD;
615 if (*Offset == 0 && *Width == 8)
617 else if (*Offset == 0 && *Width == 16)
619 else if (*Offset == 0 && *Width == 32)
621 else if (*Offset == 8 && *Width == 8)
623 else if (*Offset == 16 && *Width == 8)
625 else if (*Offset == 16 && *Width == 16)
627 else if (*Offset == 24 && *Width == 8)
632 MachineOperand *Src0 = TII->getNamedOperand(MI, AMDGPU::OpName::src0);
633 MachineOperand *Dst = TII->getNamedOperand(MI, AMDGPU::OpName::vdst);
635 if (TRI->isPhysicalRegister(Src0->getReg()) ||
636 TRI->isPhysicalRegister(Dst->getReg()))
639 return make_unique<SDWASrcOperand>(
640 Src0, Dst, SrcSel, false, false, Opcode != AMDGPU::V_BFE_U32);
643 case AMDGPU::V_AND_B32_e32:
644 case AMDGPU::V_AND_B32_e64: {
646 // from: v_and_b32_e32 v1, 0x0000ffff/0x000000ff, v0
647 // to SDWA src:v0 src_sel:WORD_0/BYTE_0
649 MachineOperand *Src0 = TII->getNamedOperand(MI, AMDGPU::OpName::src0);
650 MachineOperand *Src1 = TII->getNamedOperand(MI, AMDGPU::OpName::src1);
652 auto Imm = foldToImm(*Src0);
655 Imm = foldToImm(*Src1);
659 if (!Imm || (*Imm != 0x0000ffff && *Imm != 0x000000ff))
662 MachineOperand *Dst = TII->getNamedOperand(MI, AMDGPU::OpName::vdst);
664 if (TRI->isPhysicalRegister(Src1->getReg()) ||
665 TRI->isPhysicalRegister(Dst->getReg()))
668 return make_unique<SDWASrcOperand>(
669 ValSrc, Dst, *Imm == 0x0000ffff ? WORD_0 : BYTE_0);
672 case AMDGPU::V_OR_B32_e32:
673 case AMDGPU::V_OR_B32_e64: {
674 // Patterns for dst_unused:UNUSED_PRESERVE.
676 // v_add_f16_sdwa v0, v1, v2 dst_sel:WORD_1 dst_unused:UNUSED_PAD
677 // src1_sel:WORD_1 src2_sel:WORD1
678 // v_add_f16_e32 v3, v1, v2
679 // v_or_b32_e32 v4, v0, v3
680 // to SDWA preserve dst:v4 dst_sel:WORD_1 dst_unused:UNUSED_PRESERVE preserve:v3
682 // Check if one of operands of v_or_b32 is SDWA instruction
683 using CheckRetType = Optional<std::pair<MachineOperand *, MachineOperand *>>;
684 auto CheckOROperandsForSDWA =
685 [&](const MachineOperand *Op1, const MachineOperand *Op2) -> CheckRetType {
686 if (!Op1 || !Op1->isReg() || !Op2 || !Op2->isReg())
687 return CheckRetType(None);
689 MachineOperand *Op1Def = findSingleRegDef(Op1, MRI);
691 return CheckRetType(None);
693 MachineInstr *Op1Inst = Op1Def->getParent();
694 if (!TII->isSDWA(*Op1Inst))
695 return CheckRetType(None);
697 MachineOperand *Op2Def = findSingleRegDef(Op2, MRI);
699 return CheckRetType(None);
701 return CheckRetType(std::make_pair(Op1Def, Op2Def));
704 MachineOperand *OrSDWA = TII->getNamedOperand(MI, AMDGPU::OpName::src0);
705 MachineOperand *OrOther = TII->getNamedOperand(MI, AMDGPU::OpName::src1);
706 assert(OrSDWA && OrOther);
707 auto Res = CheckOROperandsForSDWA(OrSDWA, OrOther);
709 OrSDWA = TII->getNamedOperand(MI, AMDGPU::OpName::src1);
710 OrOther = TII->getNamedOperand(MI, AMDGPU::OpName::src0);
711 assert(OrSDWA && OrOther);
712 Res = CheckOROperandsForSDWA(OrSDWA, OrOther);
717 MachineOperand *OrSDWADef = Res->first;
718 MachineOperand *OrOtherDef = Res->second;
719 assert(OrSDWADef && OrOtherDef);
721 MachineInstr *SDWAInst = OrSDWADef->getParent();
722 MachineInstr *OtherInst = OrOtherDef->getParent();
724 // Check that OtherInstr is actually bitwise compatible with SDWAInst = their
725 // destination patterns don't overlap. Compatible instruction can be either
726 // regular instruction with compatible bitness or SDWA instruction with
728 // SDWAInst | OtherInst bitness / OtherInst dst_sel
729 // -----------------------------------------------------
731 // WORD_0 | no / BYTE_2/3, WORD_1
732 // WORD_1 | 8/16-bit instructions / BYTE_0/1, WORD_0
733 // BYTE_0 | no / BYTE_1/2/3, WORD_1
734 // BYTE_1 | 8-bit / BYTE_0/2/3, WORD_1
735 // BYTE_2 | 8/16-bit / BYTE_0/1/3. WORD_0
736 // BYTE_3 | 8/16/24-bit / BYTE_0/1/2, WORD_0
737 // E.g. if SDWAInst is v_add_f16_sdwa dst_sel:WORD_1 then v_add_f16 is OK
738 // but v_add_f32 is not.
740 // TODO: add support for non-SDWA instructions as OtherInst.
741 // For now this only works with SDWA instructions. For regular instructions
742 // there is no way to determine if instruction write only 8/16/24-bit out of
743 // full register size and all registers are at min 32-bit wide.
744 if (!TII->isSDWA(*OtherInst))
747 SdwaSel DstSel = static_cast<SdwaSel>(
748 TII->getNamedImmOperand(*SDWAInst, AMDGPU::OpName::dst_sel));;
749 SdwaSel OtherDstSel = static_cast<SdwaSel>(
750 TII->getNamedImmOperand(*OtherInst, AMDGPU::OpName::dst_sel));
752 bool DstSelAgree = false;
754 case WORD_0: DstSelAgree = ((OtherDstSel == BYTE_2) ||
755 (OtherDstSel == BYTE_3) ||
756 (OtherDstSel == WORD_1));
758 case WORD_1: DstSelAgree = ((OtherDstSel == BYTE_0) ||
759 (OtherDstSel == BYTE_1) ||
760 (OtherDstSel == WORD_0));
762 case BYTE_0: DstSelAgree = ((OtherDstSel == BYTE_1) ||
763 (OtherDstSel == BYTE_2) ||
764 (OtherDstSel == BYTE_3) ||
765 (OtherDstSel == WORD_1));
767 case BYTE_1: DstSelAgree = ((OtherDstSel == BYTE_0) ||
768 (OtherDstSel == BYTE_2) ||
769 (OtherDstSel == BYTE_3) ||
770 (OtherDstSel == WORD_1));
772 case BYTE_2: DstSelAgree = ((OtherDstSel == BYTE_0) ||
773 (OtherDstSel == BYTE_1) ||
774 (OtherDstSel == BYTE_3) ||
775 (OtherDstSel == WORD_0));
777 case BYTE_3: DstSelAgree = ((OtherDstSel == BYTE_0) ||
778 (OtherDstSel == BYTE_1) ||
779 (OtherDstSel == BYTE_2) ||
780 (OtherDstSel == WORD_0));
782 default: DstSelAgree = false;
788 // Also OtherInst dst_unused should be UNUSED_PAD
789 DstUnused OtherDstUnused = static_cast<DstUnused>(
790 TII->getNamedImmOperand(*OtherInst, AMDGPU::OpName::dst_unused));
791 if (OtherDstUnused != DstUnused::UNUSED_PAD)
794 // Create DstPreserveOperand
795 MachineOperand *OrDst = TII->getNamedOperand(MI, AMDGPU::OpName::vdst);
796 assert(OrDst && OrDst->isReg());
798 return make_unique<SDWADstPreserveOperand>(
799 OrDst, OrSDWADef, OrOtherDef, DstSel);
804 return std::unique_ptr<SDWAOperand>(nullptr);
807 void SIPeepholeSDWA::matchSDWAOperands(MachineFunction &MF) {
808 for (MachineBasicBlock &MBB : MF) {
809 for (MachineInstr &MI : MBB) {
810 if (auto Operand = matchSDWAOperand(MI)) {
811 DEBUG(dbgs() << "Match: " << MI << "To: " << *Operand << '\n');
812 SDWAOperands[&MI] = std::move(Operand);
813 ++NumSDWAPatternsFound;
819 bool SIPeepholeSDWA::isConvertibleToSDWA(const MachineInstr &MI,
820 const SISubtarget &ST) const {
821 // Check if this is already an SDWA instruction
822 unsigned Opc = MI.getOpcode();
823 if (TII->isSDWA(Opc))
826 // Check if this instruction has opcode that supports SDWA
827 if (AMDGPU::getSDWAOp(Opc) == -1)
828 Opc = AMDGPU::getVOPe32(Opc);
830 if (AMDGPU::getSDWAOp(Opc) == -1)
833 if (!ST.hasSDWAOmod() && TII->hasModifiersSet(MI, AMDGPU::OpName::omod))
836 if (TII->isVOPC(Opc)) {
837 if (!ST.hasSDWASdst()) {
838 const MachineOperand *SDst = TII->getNamedOperand(MI, AMDGPU::OpName::sdst);
839 if (SDst && SDst->getReg() != AMDGPU::VCC)
843 if (!ST.hasSDWAOutModsVOPC() &&
844 (TII->hasModifiersSet(MI, AMDGPU::OpName::clamp) ||
845 TII->hasModifiersSet(MI, AMDGPU::OpName::omod)))
848 } else if (TII->getNamedOperand(MI, AMDGPU::OpName::sdst) ||
849 !TII->getNamedOperand(MI, AMDGPU::OpName::vdst)) {
853 if (!ST.hasSDWAMac() && (Opc == AMDGPU::V_MAC_F16_e32 ||
854 Opc == AMDGPU::V_MAC_F32_e32))
860 bool SIPeepholeSDWA::convertToSDWA(MachineInstr &MI,
861 const SDWAOperandsVector &SDWAOperands) {
864 unsigned Opcode = MI.getOpcode();
865 if (TII->isSDWA(Opcode)) {
868 SDWAOpcode = AMDGPU::getSDWAOp(Opcode);
869 if (SDWAOpcode == -1)
870 SDWAOpcode = AMDGPU::getSDWAOp(AMDGPU::getVOPe32(Opcode));
872 assert(SDWAOpcode != -1);
874 const MCInstrDesc &SDWADesc = TII->get(SDWAOpcode);
876 // Create SDWA version of instruction MI and initialize its operands
877 MachineInstrBuilder SDWAInst =
878 BuildMI(*MI.getParent(), MI, MI.getDebugLoc(), SDWADesc);
880 // Copy dst, if it is present in original then should also be present in SDWA
881 MachineOperand *Dst = TII->getNamedOperand(MI, AMDGPU::OpName::vdst);
883 assert(AMDGPU::getNamedOperandIdx(SDWAOpcode, AMDGPU::OpName::vdst) != -1);
885 } else if ((Dst = TII->getNamedOperand(MI, AMDGPU::OpName::sdst))) {
887 AMDGPU::getNamedOperandIdx(SDWAOpcode, AMDGPU::OpName::sdst) != -1);
890 assert(AMDGPU::getNamedOperandIdx(SDWAOpcode, AMDGPU::OpName::sdst) != -1);
891 SDWAInst.addReg(AMDGPU::VCC, RegState::Define);
894 // Copy src0, initialize src0_modifiers. All sdwa instructions has src0 and
895 // src0_modifiers (except for v_nop_sdwa, but it can't get here)
896 MachineOperand *Src0 = TII->getNamedOperand(MI, AMDGPU::OpName::src0);
899 AMDGPU::getNamedOperandIdx(SDWAOpcode, AMDGPU::OpName::src0) != -1 &&
900 AMDGPU::getNamedOperandIdx(SDWAOpcode, AMDGPU::OpName::src0_modifiers) != -1);
901 if (auto *Mod = TII->getNamedOperand(MI, AMDGPU::OpName::src0_modifiers))
902 SDWAInst.addImm(Mod->getImm());
907 // Copy src1 if present, initialize src1_modifiers.
908 MachineOperand *Src1 = TII->getNamedOperand(MI, AMDGPU::OpName::src1);
911 AMDGPU::getNamedOperandIdx(SDWAOpcode, AMDGPU::OpName::src1) != -1 &&
912 AMDGPU::getNamedOperandIdx(SDWAOpcode, AMDGPU::OpName::src1_modifiers) != -1);
913 if (auto *Mod = TII->getNamedOperand(MI, AMDGPU::OpName::src1_modifiers))
914 SDWAInst.addImm(Mod->getImm());
920 if (SDWAOpcode == AMDGPU::V_MAC_F16_sdwa ||
921 SDWAOpcode == AMDGPU::V_MAC_F32_sdwa) {
922 // v_mac_f16/32 has additional src2 operand tied to vdst
923 MachineOperand *Src2 = TII->getNamedOperand(MI, AMDGPU::OpName::src2);
928 // Copy clamp if present, initialize otherwise
929 assert(AMDGPU::getNamedOperandIdx(SDWAOpcode, AMDGPU::OpName::clamp) != -1);
930 MachineOperand *Clamp = TII->getNamedOperand(MI, AMDGPU::OpName::clamp);
932 SDWAInst.add(*Clamp);
937 // Copy omod if present, initialize otherwise if needed
938 if (AMDGPU::getNamedOperandIdx(SDWAOpcode, AMDGPU::OpName::omod) != -1) {
939 MachineOperand *OMod = TII->getNamedOperand(MI, AMDGPU::OpName::omod);
947 // Copy dst_sel if present, initialize otherwise if needed
948 if (AMDGPU::getNamedOperandIdx(SDWAOpcode, AMDGPU::OpName::dst_sel) != -1) {
949 MachineOperand *DstSel = TII->getNamedOperand(MI, AMDGPU::OpName::dst_sel);
951 SDWAInst.add(*DstSel);
953 SDWAInst.addImm(AMDGPU::SDWA::SdwaSel::DWORD);
957 // Copy dst_unused if present, initialize otherwise if needed
958 if (AMDGPU::getNamedOperandIdx(SDWAOpcode, AMDGPU::OpName::dst_unused) != -1) {
959 MachineOperand *DstUnused = TII->getNamedOperand(MI, AMDGPU::OpName::dst_unused);
961 SDWAInst.add(*DstUnused);
963 SDWAInst.addImm(AMDGPU::SDWA::DstUnused::UNUSED_PAD);
967 // Copy src0_sel if present, initialize otherwise
968 assert(AMDGPU::getNamedOperandIdx(SDWAOpcode, AMDGPU::OpName::src0_sel) != -1);
969 MachineOperand *Src0Sel = TII->getNamedOperand(MI, AMDGPU::OpName::src0_sel);
971 SDWAInst.add(*Src0Sel);
973 SDWAInst.addImm(AMDGPU::SDWA::SdwaSel::DWORD);
976 // Copy src1_sel if present, initialize otherwise if needed
978 assert(AMDGPU::getNamedOperandIdx(SDWAOpcode, AMDGPU::OpName::src1_sel) != -1);
979 MachineOperand *Src1Sel = TII->getNamedOperand(MI, AMDGPU::OpName::src1_sel);
981 SDWAInst.add(*Src1Sel);
983 SDWAInst.addImm(AMDGPU::SDWA::SdwaSel::DWORD);
987 // Apply all sdwa operand pattenrs
988 bool Converted = false;
989 for (auto &Operand : SDWAOperands) {
990 // There should be no intesection between SDWA operands and potential MIs
992 // v_and_b32 v0, 0xff, v1 -> src:v1 sel:BYTE_0
993 // v_and_b32 v2, 0xff, v0 -> src:v0 sel:BYTE_0
994 // v_add_u32 v3, v4, v2
996 // In that example it is possible that we would fold 2nd instruction into 3rd
997 // (v_add_u32_sdwa) and then try to fold 1st instruction into 2nd (that was
998 // already destroyed). So if SDWAOperand is also a potential MI then do not
1000 if (PotentialMatches.count(Operand->getParentInst()) == 0)
1001 Converted |= Operand->convertToSDWA(*SDWAInst, TII);
1004 ConvertedInstructions.push_back(SDWAInst);
1006 SDWAInst->eraseFromParent();
1010 DEBUG(dbgs() << "Convert instruction:" << MI
1011 << "Into:" << *SDWAInst << '\n');
1012 ++NumSDWAInstructionsPeepholed;
1014 MI.eraseFromParent();
1018 // If an instruction was converted to SDWA it should not have immediates or SGPR
1019 // operands (allowed one SGPR on GFX9). Copy its scalar operands into VGPRs.
1020 void SIPeepholeSDWA::legalizeScalarOperands(MachineInstr &MI, const SISubtarget &ST) const {
1021 const MCInstrDesc &Desc = TII->get(MI.getOpcode());
1022 unsigned ConstantBusCount = 0;
1023 for (MachineOperand &Op : MI.explicit_uses()) {
1024 if (!Op.isImm() && !(Op.isReg() && !TRI->isVGPR(*MRI, Op.getReg())))
1027 unsigned I = MI.getOperandNo(&Op);
1028 if (Desc.OpInfo[I].RegClass == -1 ||
1029 !TRI->hasVGPRs(TRI->getRegClass(Desc.OpInfo[I].RegClass)))
1032 if (ST.hasSDWAScalar() && ConstantBusCount == 0 && Op.isReg() &&
1033 TRI->isSGPRReg(*MRI, Op.getReg())) {
1038 unsigned VGPR = MRI->createVirtualRegister(&AMDGPU::VGPR_32RegClass);
1039 auto Copy = BuildMI(*MI.getParent(), MI.getIterator(), MI.getDebugLoc(),
1040 TII->get(AMDGPU::V_MOV_B32_e32), VGPR);
1042 Copy.addImm(Op.getImm());
1043 else if (Op.isReg())
1044 Copy.addReg(Op.getReg(), Op.isKill() ? RegState::Kill : 0,
1046 Op.ChangeToRegister(VGPR, false);
1050 bool SIPeepholeSDWA::runOnMachineFunction(MachineFunction &MF) {
1051 const SISubtarget &ST = MF.getSubtarget<SISubtarget>();
1053 if (!ST.hasSDWA() || skipFunction(MF.getFunction()))
1056 MRI = &MF.getRegInfo();
1057 TRI = ST.getRegisterInfo();
1058 TII = ST.getInstrInfo();
1060 // Find all SDWA operands in MF.
1061 bool Changed = false;
1064 matchSDWAOperands(MF);
1066 for (const auto &OperandPair : SDWAOperands) {
1067 const auto &Operand = OperandPair.second;
1068 MachineInstr *PotentialMI = Operand->potentialToConvert(TII);
1069 if (PotentialMI && isConvertibleToSDWA(*PotentialMI, ST)) {
1070 PotentialMatches[PotentialMI].push_back(Operand.get());
1074 for (auto &PotentialPair : PotentialMatches) {
1075 MachineInstr &PotentialMI = *PotentialPair.first;
1076 convertToSDWA(PotentialMI, PotentialPair.second);
1079 PotentialMatches.clear();
1080 SDWAOperands.clear();
1082 Changed = !ConvertedInstructions.empty();
1087 while (!ConvertedInstructions.empty())
1088 legalizeScalarOperands(*ConvertedInstructions.pop_back_val(), ST);