1 //===-- SIFoldOperands.cpp - Fold operands --- ----------------------------===// 2 // 3 // Part of the LLVM Project, under the Apache License v2.0 with LLVM Exceptions. 4 // See https://llvm.org/LICENSE.txt for license information. 5 // SPDX-License-Identifier: Apache-2.0 WITH LLVM-exception 6 // 7 /// \file 8 //===----------------------------------------------------------------------===// 9 // 10 11 #include "AMDGPU.h" 12 #include "GCNSubtarget.h" 13 #include "MCTargetDesc/AMDGPUMCTargetDesc.h" 14 #include "SIMachineFunctionInfo.h" 15 #include "llvm/ADT/DepthFirstIterator.h" 16 #include "llvm/CodeGen/MachineFunctionPass.h" 17 18 #define DEBUG_TYPE "si-fold-operands" 19 using namespace llvm; 20 21 namespace { 22 23 struct FoldCandidate { 24 MachineInstr *UseMI; 25 union { 26 MachineOperand *OpToFold; 27 uint64_t ImmToFold; 28 int FrameIndexToFold; 29 }; 30 int ShrinkOpcode; 31 unsigned UseOpNo; 32 MachineOperand::MachineOperandType Kind; 33 bool Commuted; 34 35 FoldCandidate(MachineInstr *MI, unsigned OpNo, MachineOperand *FoldOp, 36 bool Commuted_ = false, 37 int ShrinkOp = -1) : 38 UseMI(MI), OpToFold(nullptr), ShrinkOpcode(ShrinkOp), UseOpNo(OpNo), 39 Kind(FoldOp->getType()), 40 Commuted(Commuted_) { 41 if (FoldOp->isImm()) { 42 ImmToFold = FoldOp->getImm(); 43 } else if (FoldOp->isFI()) { 44 FrameIndexToFold = FoldOp->getIndex(); 45 } else { 46 assert(FoldOp->isReg() || FoldOp->isGlobal()); 47 OpToFold = FoldOp; 48 } 49 } 50 51 bool isFI() const { 52 return Kind == MachineOperand::MO_FrameIndex; 53 } 54 55 bool isImm() const { 56 return Kind == MachineOperand::MO_Immediate; 57 } 58 59 bool isReg() const { 60 return Kind == MachineOperand::MO_Register; 61 } 62 63 bool isGlobal() const { return Kind == MachineOperand::MO_GlobalAddress; } 64 65 bool isCommuted() const { 66 return Commuted; 67 } 68 69 bool needsShrink() const { 70 return ShrinkOpcode != -1; 71 } 72 73 int getShrinkOpcode() const { 74 return ShrinkOpcode; 75 } 76 }; 77 78 class SIFoldOperands : public MachineFunctionPass { 79 public: 80 static char ID; 81 MachineRegisterInfo *MRI; 82 const SIInstrInfo *TII; 83 const SIRegisterInfo *TRI; 84 const GCNSubtarget *ST; 85 const SIMachineFunctionInfo *MFI; 86 87 void foldOperand(MachineOperand &OpToFold, 88 MachineInstr *UseMI, 89 int UseOpIdx, 90 SmallVectorImpl<FoldCandidate> &FoldList, 91 SmallVectorImpl<MachineInstr *> &CopiesToReplace) const; 92 93 bool tryFoldCndMask(MachineInstr &MI) const; 94 bool tryFoldZeroHighBits(MachineInstr &MI) const; 95 bool foldInstOperand(MachineInstr &MI, MachineOperand &OpToFold) const; 96 97 const MachineOperand *isClamp(const MachineInstr &MI) const; 98 bool tryFoldClamp(MachineInstr &MI); 99 100 std::pair<const MachineOperand *, int> isOMod(const MachineInstr &MI) const; 101 bool tryFoldOMod(MachineInstr &MI); 102 bool tryFoldRegSequence(MachineInstr &MI); 103 bool tryFoldLCSSAPhi(MachineInstr &MI); 104 bool tryFoldLoad(MachineInstr &MI); 105 106 public: 107 SIFoldOperands() : MachineFunctionPass(ID) { 108 initializeSIFoldOperandsPass(*PassRegistry::getPassRegistry()); 109 } 110 111 bool runOnMachineFunction(MachineFunction &MF) override; 112 113 StringRef getPassName() const override { return "SI Fold Operands"; } 114 115 void getAnalysisUsage(AnalysisUsage &AU) const override { 116 AU.setPreservesCFG(); 117 MachineFunctionPass::getAnalysisUsage(AU); 118 } 119 }; 120 121 } // End anonymous namespace. 122 123 INITIALIZE_PASS(SIFoldOperands, DEBUG_TYPE, 124 "SI Fold Operands", false, false) 125 126 char SIFoldOperands::ID = 0; 127 128 char &llvm::SIFoldOperandsID = SIFoldOperands::ID; 129 130 // Map multiply-accumulate opcode to corresponding multiply-add opcode if any. 131 static unsigned macToMad(unsigned Opc) { 132 switch (Opc) { 133 case AMDGPU::V_MAC_F32_e64: 134 return AMDGPU::V_MAD_F32_e64; 135 case AMDGPU::V_MAC_F16_e64: 136 return AMDGPU::V_MAD_F16_e64; 137 case AMDGPU::V_FMAC_F32_e64: 138 return AMDGPU::V_FMA_F32_e64; 139 case AMDGPU::V_FMAC_F16_e64: 140 return AMDGPU::V_FMA_F16_gfx9_e64; 141 case AMDGPU::V_FMAC_LEGACY_F32_e64: 142 return AMDGPU::V_FMA_LEGACY_F32_e64; 143 case AMDGPU::V_FMAC_F64_e64: 144 return AMDGPU::V_FMA_F64_e64; 145 } 146 return AMDGPU::INSTRUCTION_LIST_END; 147 } 148 149 // TODO: Add heuristic that the frame index might not fit in the addressing mode 150 // immediate offset to avoid materializing in loops. 151 static bool frameIndexMayFold(const SIInstrInfo *TII, 152 const MachineInstr &UseMI, 153 int OpNo, 154 const MachineOperand &OpToFold) { 155 if (!OpToFold.isFI()) 156 return false; 157 158 if (TII->isMUBUF(UseMI)) 159 return OpNo == AMDGPU::getNamedOperandIdx(UseMI.getOpcode(), 160 AMDGPU::OpName::vaddr); 161 if (!TII->isFLATScratch(UseMI)) 162 return false; 163 164 int SIdx = AMDGPU::getNamedOperandIdx(UseMI.getOpcode(), 165 AMDGPU::OpName::saddr); 166 if (OpNo == SIdx) 167 return true; 168 169 int VIdx = AMDGPU::getNamedOperandIdx(UseMI.getOpcode(), 170 AMDGPU::OpName::vaddr); 171 return OpNo == VIdx && SIdx == -1; 172 } 173 174 FunctionPass *llvm::createSIFoldOperandsPass() { 175 return new SIFoldOperands(); 176 } 177 178 static bool updateOperand(FoldCandidate &Fold, 179 const SIInstrInfo &TII, 180 const TargetRegisterInfo &TRI, 181 const GCNSubtarget &ST) { 182 MachineInstr *MI = Fold.UseMI; 183 MachineOperand &Old = MI->getOperand(Fold.UseOpNo); 184 assert(Old.isReg()); 185 186 if (Fold.isImm()) { 187 if (MI->getDesc().TSFlags & SIInstrFlags::IsPacked && 188 !(MI->getDesc().TSFlags & SIInstrFlags::IsMAI) && 189 (!ST.hasDOTOpSelHazard() || 190 !(MI->getDesc().TSFlags & SIInstrFlags::IsDOT)) && 191 AMDGPU::isFoldableLiteralV216(Fold.ImmToFold, 192 ST.hasInv2PiInlineImm())) { 193 // Set op_sel/op_sel_hi on this operand or bail out if op_sel is 194 // already set. 195 unsigned Opcode = MI->getOpcode(); 196 int OpNo = MI->getOperandNo(&Old); 197 int ModIdx = -1; 198 if (OpNo == AMDGPU::getNamedOperandIdx(Opcode, AMDGPU::OpName::src0)) 199 ModIdx = AMDGPU::OpName::src0_modifiers; 200 else if (OpNo == AMDGPU::getNamedOperandIdx(Opcode, AMDGPU::OpName::src1)) 201 ModIdx = AMDGPU::OpName::src1_modifiers; 202 else if (OpNo == AMDGPU::getNamedOperandIdx(Opcode, AMDGPU::OpName::src2)) 203 ModIdx = AMDGPU::OpName::src2_modifiers; 204 assert(ModIdx != -1); 205 ModIdx = AMDGPU::getNamedOperandIdx(Opcode, ModIdx); 206 MachineOperand &Mod = MI->getOperand(ModIdx); 207 unsigned Val = Mod.getImm(); 208 if (!(Val & SISrcMods::OP_SEL_0) && (Val & SISrcMods::OP_SEL_1)) { 209 // Only apply the following transformation if that operand requires 210 // a packed immediate. 211 switch (TII.get(Opcode).OpInfo[OpNo].OperandType) { 212 case AMDGPU::OPERAND_REG_IMM_V2FP16: 213 case AMDGPU::OPERAND_REG_IMM_V2INT16: 214 case AMDGPU::OPERAND_REG_INLINE_C_V2FP16: 215 case AMDGPU::OPERAND_REG_INLINE_C_V2INT16: 216 // If upper part is all zero we do not need op_sel_hi. 217 if (!isUInt<16>(Fold.ImmToFold)) { 218 if (!(Fold.ImmToFold & 0xffff)) { 219 Mod.setImm(Mod.getImm() | SISrcMods::OP_SEL_0); 220 Mod.setImm(Mod.getImm() & ~SISrcMods::OP_SEL_1); 221 Old.ChangeToImmediate((Fold.ImmToFold >> 16) & 0xffff); 222 return true; 223 } 224 Mod.setImm(Mod.getImm() & ~SISrcMods::OP_SEL_1); 225 Old.ChangeToImmediate(Fold.ImmToFold & 0xffff); 226 return true; 227 } 228 break; 229 default: 230 break; 231 } 232 } 233 } 234 } 235 236 if ((Fold.isImm() || Fold.isFI() || Fold.isGlobal()) && Fold.needsShrink()) { 237 MachineBasicBlock *MBB = MI->getParent(); 238 auto Liveness = MBB->computeRegisterLiveness(&TRI, AMDGPU::VCC, MI, 16); 239 if (Liveness != MachineBasicBlock::LQR_Dead) { 240 LLVM_DEBUG(dbgs() << "Not shrinking " << MI << " due to vcc liveness\n"); 241 return false; 242 } 243 244 MachineRegisterInfo &MRI = MBB->getParent()->getRegInfo(); 245 int Op32 = Fold.getShrinkOpcode(); 246 MachineOperand &Dst0 = MI->getOperand(0); 247 MachineOperand &Dst1 = MI->getOperand(1); 248 assert(Dst0.isDef() && Dst1.isDef()); 249 250 bool HaveNonDbgCarryUse = !MRI.use_nodbg_empty(Dst1.getReg()); 251 252 const TargetRegisterClass *Dst0RC = MRI.getRegClass(Dst0.getReg()); 253 Register NewReg0 = MRI.createVirtualRegister(Dst0RC); 254 255 MachineInstr *Inst32 = TII.buildShrunkInst(*MI, Op32); 256 257 if (HaveNonDbgCarryUse) { 258 BuildMI(*MBB, MI, MI->getDebugLoc(), TII.get(AMDGPU::COPY), Dst1.getReg()) 259 .addReg(AMDGPU::VCC, RegState::Kill); 260 } 261 262 // Keep the old instruction around to avoid breaking iterators, but 263 // replace it with a dummy instruction to remove uses. 264 // 265 // FIXME: We should not invert how this pass looks at operands to avoid 266 // this. Should track set of foldable movs instead of looking for uses 267 // when looking at a use. 268 Dst0.setReg(NewReg0); 269 for (unsigned I = MI->getNumOperands() - 1; I > 0; --I) 270 MI->removeOperand(I); 271 MI->setDesc(TII.get(AMDGPU::IMPLICIT_DEF)); 272 273 if (Fold.isCommuted()) 274 TII.commuteInstruction(*Inst32, false); 275 return true; 276 } 277 278 assert(!Fold.needsShrink() && "not handled"); 279 280 if (Fold.isImm()) { 281 if (Old.isTied()) { 282 int NewMFMAOpc = AMDGPU::getMFMAEarlyClobberOp(MI->getOpcode()); 283 if (NewMFMAOpc == -1) 284 return false; 285 MI->setDesc(TII.get(NewMFMAOpc)); 286 MI->untieRegOperand(0); 287 } 288 Old.ChangeToImmediate(Fold.ImmToFold); 289 return true; 290 } 291 292 if (Fold.isGlobal()) { 293 Old.ChangeToGA(Fold.OpToFold->getGlobal(), Fold.OpToFold->getOffset(), 294 Fold.OpToFold->getTargetFlags()); 295 return true; 296 } 297 298 if (Fold.isFI()) { 299 Old.ChangeToFrameIndex(Fold.FrameIndexToFold); 300 return true; 301 } 302 303 MachineOperand *New = Fold.OpToFold; 304 Old.substVirtReg(New->getReg(), New->getSubReg(), TRI); 305 Old.setIsUndef(New->isUndef()); 306 return true; 307 } 308 309 static bool isUseMIInFoldList(ArrayRef<FoldCandidate> FoldList, 310 const MachineInstr *MI) { 311 for (auto Candidate : FoldList) { 312 if (Candidate.UseMI == MI) 313 return true; 314 } 315 return false; 316 } 317 318 static void appendFoldCandidate(SmallVectorImpl<FoldCandidate> &FoldList, 319 MachineInstr *MI, unsigned OpNo, 320 MachineOperand *FoldOp, bool Commuted = false, 321 int ShrinkOp = -1) { 322 // Skip additional folding on the same operand. 323 for (FoldCandidate &Fold : FoldList) 324 if (Fold.UseMI == MI && Fold.UseOpNo == OpNo) 325 return; 326 LLVM_DEBUG(dbgs() << "Append " << (Commuted ? "commuted" : "normal") 327 << " operand " << OpNo << "\n " << *MI); 328 FoldList.emplace_back(MI, OpNo, FoldOp, Commuted, ShrinkOp); 329 } 330 331 static bool tryAddToFoldList(SmallVectorImpl<FoldCandidate> &FoldList, 332 MachineInstr *MI, unsigned OpNo, 333 MachineOperand *OpToFold, 334 const SIInstrInfo *TII) { 335 if (!TII->isOperandLegal(*MI, OpNo, OpToFold)) { 336 // Special case for v_mac_{f16, f32}_e64 if we are trying to fold into src2 337 unsigned Opc = MI->getOpcode(); 338 unsigned NewOpc = macToMad(Opc); 339 if (NewOpc != AMDGPU::INSTRUCTION_LIST_END) { 340 // Check if changing this to a v_mad_{f16, f32} instruction will allow us 341 // to fold the operand. 342 MI->setDesc(TII->get(NewOpc)); 343 bool FoldAsMAD = tryAddToFoldList(FoldList, MI, OpNo, OpToFold, TII); 344 if (FoldAsMAD) { 345 MI->untieRegOperand(OpNo); 346 return true; 347 } 348 MI->setDesc(TII->get(Opc)); 349 } 350 351 // Special case for s_setreg_b32 352 if (OpToFold->isImm()) { 353 unsigned ImmOpc = 0; 354 if (Opc == AMDGPU::S_SETREG_B32) 355 ImmOpc = AMDGPU::S_SETREG_IMM32_B32; 356 else if (Opc == AMDGPU::S_SETREG_B32_mode) 357 ImmOpc = AMDGPU::S_SETREG_IMM32_B32_mode; 358 if (ImmOpc) { 359 MI->setDesc(TII->get(ImmOpc)); 360 appendFoldCandidate(FoldList, MI, OpNo, OpToFold); 361 return true; 362 } 363 } 364 365 // If we are already folding into another operand of MI, then 366 // we can't commute the instruction, otherwise we risk making the 367 // other fold illegal. 368 if (isUseMIInFoldList(FoldList, MI)) 369 return false; 370 371 unsigned CommuteOpNo = OpNo; 372 373 // Operand is not legal, so try to commute the instruction to 374 // see if this makes it possible to fold. 375 unsigned CommuteIdx0 = TargetInstrInfo::CommuteAnyOperandIndex; 376 unsigned CommuteIdx1 = TargetInstrInfo::CommuteAnyOperandIndex; 377 bool CanCommute = TII->findCommutedOpIndices(*MI, CommuteIdx0, CommuteIdx1); 378 379 if (CanCommute) { 380 if (CommuteIdx0 == OpNo) 381 CommuteOpNo = CommuteIdx1; 382 else if (CommuteIdx1 == OpNo) 383 CommuteOpNo = CommuteIdx0; 384 } 385 386 387 // One of operands might be an Imm operand, and OpNo may refer to it after 388 // the call of commuteInstruction() below. Such situations are avoided 389 // here explicitly as OpNo must be a register operand to be a candidate 390 // for memory folding. 391 if (CanCommute && (!MI->getOperand(CommuteIdx0).isReg() || 392 !MI->getOperand(CommuteIdx1).isReg())) 393 return false; 394 395 if (!CanCommute || 396 !TII->commuteInstruction(*MI, false, CommuteIdx0, CommuteIdx1)) 397 return false; 398 399 if (!TII->isOperandLegal(*MI, CommuteOpNo, OpToFold)) { 400 if ((Opc == AMDGPU::V_ADD_CO_U32_e64 || 401 Opc == AMDGPU::V_SUB_CO_U32_e64 || 402 Opc == AMDGPU::V_SUBREV_CO_U32_e64) && // FIXME 403 (OpToFold->isImm() || OpToFold->isFI() || OpToFold->isGlobal())) { 404 MachineRegisterInfo &MRI = MI->getParent()->getParent()->getRegInfo(); 405 406 // Verify the other operand is a VGPR, otherwise we would violate the 407 // constant bus restriction. 408 unsigned OtherIdx = CommuteOpNo == CommuteIdx0 ? CommuteIdx1 : CommuteIdx0; 409 MachineOperand &OtherOp = MI->getOperand(OtherIdx); 410 if (!OtherOp.isReg() || 411 !TII->getRegisterInfo().isVGPR(MRI, OtherOp.getReg())) 412 return false; 413 414 assert(MI->getOperand(1).isDef()); 415 416 // Make sure to get the 32-bit version of the commuted opcode. 417 unsigned MaybeCommutedOpc = MI->getOpcode(); 418 int Op32 = AMDGPU::getVOPe32(MaybeCommutedOpc); 419 420 appendFoldCandidate(FoldList, MI, CommuteOpNo, OpToFold, true, Op32); 421 return true; 422 } 423 424 TII->commuteInstruction(*MI, false, CommuteIdx0, CommuteIdx1); 425 return false; 426 } 427 428 appendFoldCandidate(FoldList, MI, CommuteOpNo, OpToFold, true); 429 return true; 430 } 431 432 // Check the case where we might introduce a second constant operand to a 433 // scalar instruction 434 if (TII->isSALU(MI->getOpcode())) { 435 const MCInstrDesc &InstDesc = MI->getDesc(); 436 const MCOperandInfo &OpInfo = InstDesc.OpInfo[OpNo]; 437 const SIRegisterInfo &SRI = TII->getRegisterInfo(); 438 439 // Fine if the operand can be encoded as an inline constant 440 if (TII->isLiteralConstantLike(*OpToFold, OpInfo)) { 441 if (!SRI.opCanUseInlineConstant(OpInfo.OperandType) || 442 !TII->isInlineConstant(*OpToFold, OpInfo)) { 443 // Otherwise check for another constant 444 for (unsigned i = 0, e = InstDesc.getNumOperands(); i != e; ++i) { 445 auto &Op = MI->getOperand(i); 446 if (OpNo != i && 447 TII->isLiteralConstantLike(Op, OpInfo)) { 448 return false; 449 } 450 } 451 } 452 } 453 } 454 455 appendFoldCandidate(FoldList, MI, OpNo, OpToFold); 456 return true; 457 } 458 459 // If the use operand doesn't care about the value, this may be an operand only 460 // used for register indexing, in which case it is unsafe to fold. 461 static bool isUseSafeToFold(const SIInstrInfo *TII, 462 const MachineInstr &MI, 463 const MachineOperand &UseMO) { 464 if (UseMO.isUndef() || TII->isSDWA(MI)) 465 return false; 466 467 switch (MI.getOpcode()) { 468 case AMDGPU::V_MOV_B32_e32: 469 case AMDGPU::V_MOV_B32_e64: 470 case AMDGPU::V_MOV_B64_PSEUDO: 471 case AMDGPU::V_MOV_B64_e32: 472 case AMDGPU::V_MOV_B64_e64: 473 // Do not fold into an indirect mov. 474 return !MI.hasRegisterImplicitUseOperand(AMDGPU::M0); 475 } 476 477 return true; 478 //return !MI.hasRegisterImplicitUseOperand(UseMO.getReg()); 479 } 480 481 // Find a def of the UseReg, check if it is a reg_sequence and find initializers 482 // for each subreg, tracking it to foldable inline immediate if possible. 483 // Returns true on success. 484 static bool getRegSeqInit( 485 SmallVectorImpl<std::pair<MachineOperand*, unsigned>> &Defs, 486 Register UseReg, uint8_t OpTy, 487 const SIInstrInfo *TII, const MachineRegisterInfo &MRI) { 488 MachineInstr *Def = MRI.getVRegDef(UseReg); 489 if (!Def || !Def->isRegSequence()) 490 return false; 491 492 for (unsigned I = 1, E = Def->getNumExplicitOperands(); I < E; I += 2) { 493 MachineOperand *Sub = &Def->getOperand(I); 494 assert(Sub->isReg()); 495 496 for (MachineInstr *SubDef = MRI.getVRegDef(Sub->getReg()); 497 SubDef && Sub->isReg() && Sub->getReg().isVirtual() && 498 !Sub->getSubReg() && TII->isFoldableCopy(*SubDef); 499 SubDef = MRI.getVRegDef(Sub->getReg())) { 500 MachineOperand *Op = &SubDef->getOperand(1); 501 if (Op->isImm()) { 502 if (TII->isInlineConstant(*Op, OpTy)) 503 Sub = Op; 504 break; 505 } 506 if (!Op->isReg() || Op->getReg().isPhysical()) 507 break; 508 Sub = Op; 509 } 510 511 Defs.emplace_back(Sub, Def->getOperand(I + 1).getImm()); 512 } 513 514 return true; 515 } 516 517 static bool tryToFoldACImm(const SIInstrInfo *TII, 518 const MachineOperand &OpToFold, 519 MachineInstr *UseMI, 520 unsigned UseOpIdx, 521 SmallVectorImpl<FoldCandidate> &FoldList) { 522 const MCInstrDesc &Desc = UseMI->getDesc(); 523 const MCOperandInfo *OpInfo = Desc.OpInfo; 524 if (!OpInfo || UseOpIdx >= Desc.getNumOperands()) 525 return false; 526 527 uint8_t OpTy = OpInfo[UseOpIdx].OperandType; 528 if ((OpTy < AMDGPU::OPERAND_REG_INLINE_AC_FIRST || 529 OpTy > AMDGPU::OPERAND_REG_INLINE_AC_LAST) && 530 (OpTy < AMDGPU::OPERAND_REG_INLINE_C_FIRST || 531 OpTy > AMDGPU::OPERAND_REG_INLINE_C_LAST)) 532 return false; 533 534 if (OpToFold.isImm() && TII->isInlineConstant(OpToFold, OpTy) && 535 TII->isOperandLegal(*UseMI, UseOpIdx, &OpToFold)) { 536 UseMI->getOperand(UseOpIdx).ChangeToImmediate(OpToFold.getImm()); 537 return true; 538 } 539 540 if (!OpToFold.isReg()) 541 return false; 542 543 Register UseReg = OpToFold.getReg(); 544 if (!UseReg.isVirtual()) 545 return false; 546 547 if (isUseMIInFoldList(FoldList, UseMI)) 548 return false; 549 550 MachineRegisterInfo &MRI = UseMI->getParent()->getParent()->getRegInfo(); 551 552 // Maybe it is just a COPY of an immediate itself. 553 MachineInstr *Def = MRI.getVRegDef(UseReg); 554 MachineOperand &UseOp = UseMI->getOperand(UseOpIdx); 555 if (!UseOp.getSubReg() && Def && TII->isFoldableCopy(*Def)) { 556 MachineOperand &DefOp = Def->getOperand(1); 557 if (DefOp.isImm() && TII->isInlineConstant(DefOp, OpTy) && 558 TII->isOperandLegal(*UseMI, UseOpIdx, &DefOp)) { 559 UseMI->getOperand(UseOpIdx).ChangeToImmediate(DefOp.getImm()); 560 return true; 561 } 562 } 563 564 SmallVector<std::pair<MachineOperand*, unsigned>, 32> Defs; 565 if (!getRegSeqInit(Defs, UseReg, OpTy, TII, MRI)) 566 return false; 567 568 int32_t Imm; 569 for (unsigned I = 0, E = Defs.size(); I != E; ++I) { 570 const MachineOperand *Op = Defs[I].first; 571 if (!Op->isImm()) 572 return false; 573 574 auto SubImm = Op->getImm(); 575 if (!I) { 576 Imm = SubImm; 577 if (!TII->isInlineConstant(*Op, OpTy) || 578 !TII->isOperandLegal(*UseMI, UseOpIdx, Op)) 579 return false; 580 581 continue; 582 } 583 if (Imm != SubImm) 584 return false; // Can only fold splat constants 585 } 586 587 appendFoldCandidate(FoldList, UseMI, UseOpIdx, Defs[0].first); 588 return true; 589 } 590 591 void SIFoldOperands::foldOperand( 592 MachineOperand &OpToFold, 593 MachineInstr *UseMI, 594 int UseOpIdx, 595 SmallVectorImpl<FoldCandidate> &FoldList, 596 SmallVectorImpl<MachineInstr *> &CopiesToReplace) const { 597 const MachineOperand &UseOp = UseMI->getOperand(UseOpIdx); 598 599 if (!isUseSafeToFold(TII, *UseMI, UseOp)) 600 return; 601 602 // FIXME: Fold operands with subregs. 603 if (UseOp.isReg() && OpToFold.isReg()) { 604 if (UseOp.isImplicit() || UseOp.getSubReg() != AMDGPU::NoSubRegister) 605 return; 606 } 607 608 // Special case for REG_SEQUENCE: We can't fold literals into 609 // REG_SEQUENCE instructions, so we have to fold them into the 610 // uses of REG_SEQUENCE. 611 if (UseMI->isRegSequence()) { 612 Register RegSeqDstReg = UseMI->getOperand(0).getReg(); 613 unsigned RegSeqDstSubReg = UseMI->getOperand(UseOpIdx + 1).getImm(); 614 615 for (auto &RSUse : make_early_inc_range(MRI->use_nodbg_operands(RegSeqDstReg))) { 616 MachineInstr *RSUseMI = RSUse.getParent(); 617 618 if (tryToFoldACImm(TII, UseMI->getOperand(0), RSUseMI, 619 RSUseMI->getOperandNo(&RSUse), FoldList)) 620 continue; 621 622 if (RSUse.getSubReg() != RegSeqDstSubReg) 623 continue; 624 625 foldOperand(OpToFold, RSUseMI, RSUseMI->getOperandNo(&RSUse), FoldList, 626 CopiesToReplace); 627 } 628 629 return; 630 } 631 632 if (tryToFoldACImm(TII, OpToFold, UseMI, UseOpIdx, FoldList)) 633 return; 634 635 if (frameIndexMayFold(TII, *UseMI, UseOpIdx, OpToFold)) { 636 // Verify that this is a stack access. 637 // FIXME: Should probably use stack pseudos before frame lowering. 638 639 if (TII->isMUBUF(*UseMI)) { 640 if (TII->getNamedOperand(*UseMI, AMDGPU::OpName::srsrc)->getReg() != 641 MFI->getScratchRSrcReg()) 642 return; 643 644 // Ensure this is either relative to the current frame or the current 645 // wave. 646 MachineOperand &SOff = 647 *TII->getNamedOperand(*UseMI, AMDGPU::OpName::soffset); 648 if (!SOff.isImm() || SOff.getImm() != 0) 649 return; 650 } 651 652 // A frame index will resolve to a positive constant, so it should always be 653 // safe to fold the addressing mode, even pre-GFX9. 654 UseMI->getOperand(UseOpIdx).ChangeToFrameIndex(OpToFold.getIndex()); 655 656 if (TII->isFLATScratch(*UseMI) && 657 AMDGPU::getNamedOperandIdx(UseMI->getOpcode(), 658 AMDGPU::OpName::vaddr) != -1 && 659 AMDGPU::getNamedOperandIdx(UseMI->getOpcode(), 660 AMDGPU::OpName::saddr) == -1) { 661 unsigned NewOpc = AMDGPU::getFlatScratchInstSSfromSV(UseMI->getOpcode()); 662 UseMI->setDesc(TII->get(NewOpc)); 663 } 664 665 return; 666 } 667 668 bool FoldingImmLike = 669 OpToFold.isImm() || OpToFold.isFI() || OpToFold.isGlobal(); 670 671 if (FoldingImmLike && UseMI->isCopy()) { 672 Register DestReg = UseMI->getOperand(0).getReg(); 673 Register SrcReg = UseMI->getOperand(1).getReg(); 674 assert(SrcReg.isVirtual()); 675 676 const TargetRegisterClass *SrcRC = MRI->getRegClass(SrcReg); 677 678 // Don't fold into a copy to a physical register with the same class. Doing 679 // so would interfere with the register coalescer's logic which would avoid 680 // redundant initializations. 681 if (DestReg.isPhysical() && SrcRC->contains(DestReg)) 682 return; 683 684 const TargetRegisterClass *DestRC = TRI->getRegClassForReg(*MRI, DestReg); 685 if (!DestReg.isPhysical()) { 686 if (TRI->isSGPRClass(SrcRC) && TRI->hasVectorRegisters(DestRC)) { 687 SmallVector<FoldCandidate, 4> CopyUses; 688 for (auto &Use : MRI->use_nodbg_operands(DestReg)) { 689 // There's no point trying to fold into an implicit operand. 690 if (Use.isImplicit()) 691 continue; 692 693 CopyUses.emplace_back(Use.getParent(), 694 Use.getParent()->getOperandNo(&Use), 695 &UseMI->getOperand(1)); 696 } 697 for (auto &F : CopyUses) { 698 foldOperand(*F.OpToFold, F.UseMI, F.UseOpNo, FoldList, CopiesToReplace); 699 } 700 } 701 702 if (DestRC == &AMDGPU::AGPR_32RegClass && 703 TII->isInlineConstant(OpToFold, AMDGPU::OPERAND_REG_INLINE_C_INT32)) { 704 UseMI->setDesc(TII->get(AMDGPU::V_ACCVGPR_WRITE_B32_e64)); 705 UseMI->getOperand(1).ChangeToImmediate(OpToFold.getImm()); 706 CopiesToReplace.push_back(UseMI); 707 return; 708 } 709 } 710 711 // In order to fold immediates into copies, we need to change the 712 // copy to a MOV. 713 714 unsigned MovOp = TII->getMovOpcode(DestRC); 715 if (MovOp == AMDGPU::COPY) 716 return; 717 718 UseMI->setDesc(TII->get(MovOp)); 719 MachineInstr::mop_iterator ImpOpI = UseMI->implicit_operands().begin(); 720 MachineInstr::mop_iterator ImpOpE = UseMI->implicit_operands().end(); 721 while (ImpOpI != ImpOpE) { 722 MachineInstr::mop_iterator Tmp = ImpOpI; 723 ImpOpI++; 724 UseMI->removeOperand(UseMI->getOperandNo(Tmp)); 725 } 726 CopiesToReplace.push_back(UseMI); 727 } else { 728 if (UseMI->isCopy() && OpToFold.isReg() && 729 UseMI->getOperand(0).getReg().isVirtual() && 730 !UseMI->getOperand(1).getSubReg()) { 731 LLVM_DEBUG(dbgs() << "Folding " << OpToFold << "\n into " << *UseMI); 732 unsigned Size = TII->getOpSize(*UseMI, 1); 733 Register UseReg = OpToFold.getReg(); 734 UseMI->getOperand(1).setReg(UseReg); 735 UseMI->getOperand(1).setSubReg(OpToFold.getSubReg()); 736 UseMI->getOperand(1).setIsKill(false); 737 CopiesToReplace.push_back(UseMI); 738 OpToFold.setIsKill(false); 739 740 // Remove kill flags as kills may now be out of order with uses. 741 MRI->clearKillFlags(OpToFold.getReg()); 742 743 // That is very tricky to store a value into an AGPR. v_accvgpr_write_b32 744 // can only accept VGPR or inline immediate. Recreate a reg_sequence with 745 // its initializers right here, so we will rematerialize immediates and 746 // avoid copies via different reg classes. 747 SmallVector<std::pair<MachineOperand*, unsigned>, 32> Defs; 748 if (Size > 4 && TRI->isAGPR(*MRI, UseMI->getOperand(0).getReg()) && 749 getRegSeqInit(Defs, UseReg, AMDGPU::OPERAND_REG_INLINE_C_INT32, TII, 750 *MRI)) { 751 const DebugLoc &DL = UseMI->getDebugLoc(); 752 MachineBasicBlock &MBB = *UseMI->getParent(); 753 754 UseMI->setDesc(TII->get(AMDGPU::REG_SEQUENCE)); 755 for (unsigned I = UseMI->getNumOperands() - 1; I > 0; --I) 756 UseMI->removeOperand(I); 757 758 MachineInstrBuilder B(*MBB.getParent(), UseMI); 759 DenseMap<TargetInstrInfo::RegSubRegPair, Register> VGPRCopies; 760 SmallSetVector<TargetInstrInfo::RegSubRegPair, 32> SeenAGPRs; 761 for (unsigned I = 0; I < Size / 4; ++I) { 762 MachineOperand *Def = Defs[I].first; 763 TargetInstrInfo::RegSubRegPair CopyToVGPR; 764 if (Def->isImm() && 765 TII->isInlineConstant(*Def, AMDGPU::OPERAND_REG_INLINE_C_INT32)) { 766 int64_t Imm = Def->getImm(); 767 768 auto Tmp = MRI->createVirtualRegister(&AMDGPU::AGPR_32RegClass); 769 BuildMI(MBB, UseMI, DL, 770 TII->get(AMDGPU::V_ACCVGPR_WRITE_B32_e64), Tmp).addImm(Imm); 771 B.addReg(Tmp); 772 } else if (Def->isReg() && TRI->isAGPR(*MRI, Def->getReg())) { 773 auto Src = getRegSubRegPair(*Def); 774 Def->setIsKill(false); 775 if (!SeenAGPRs.insert(Src)) { 776 // We cannot build a reg_sequence out of the same registers, they 777 // must be copied. Better do it here before copyPhysReg() created 778 // several reads to do the AGPR->VGPR->AGPR copy. 779 CopyToVGPR = Src; 780 } else { 781 B.addReg(Src.Reg, Def->isUndef() ? RegState::Undef : 0, 782 Src.SubReg); 783 } 784 } else { 785 assert(Def->isReg()); 786 Def->setIsKill(false); 787 auto Src = getRegSubRegPair(*Def); 788 789 // Direct copy from SGPR to AGPR is not possible. To avoid creation 790 // of exploded copies SGPR->VGPR->AGPR in the copyPhysReg() later, 791 // create a copy here and track if we already have such a copy. 792 if (TRI->isSGPRReg(*MRI, Src.Reg)) { 793 CopyToVGPR = Src; 794 } else { 795 auto Tmp = MRI->createVirtualRegister(&AMDGPU::AGPR_32RegClass); 796 BuildMI(MBB, UseMI, DL, TII->get(AMDGPU::COPY), Tmp).add(*Def); 797 B.addReg(Tmp); 798 } 799 } 800 801 if (CopyToVGPR.Reg) { 802 Register Vgpr; 803 if (VGPRCopies.count(CopyToVGPR)) { 804 Vgpr = VGPRCopies[CopyToVGPR]; 805 } else { 806 Vgpr = MRI->createVirtualRegister(&AMDGPU::VGPR_32RegClass); 807 BuildMI(MBB, UseMI, DL, TII->get(AMDGPU::COPY), Vgpr).add(*Def); 808 VGPRCopies[CopyToVGPR] = Vgpr; 809 } 810 auto Tmp = MRI->createVirtualRegister(&AMDGPU::AGPR_32RegClass); 811 BuildMI(MBB, UseMI, DL, 812 TII->get(AMDGPU::V_ACCVGPR_WRITE_B32_e64), Tmp).addReg(Vgpr); 813 B.addReg(Tmp); 814 } 815 816 B.addImm(Defs[I].second); 817 } 818 LLVM_DEBUG(dbgs() << "Folded " << *UseMI); 819 return; 820 } 821 822 if (Size != 4) 823 return; 824 if (TRI->isAGPR(*MRI, UseMI->getOperand(0).getReg()) && 825 TRI->isVGPR(*MRI, UseMI->getOperand(1).getReg())) 826 UseMI->setDesc(TII->get(AMDGPU::V_ACCVGPR_WRITE_B32_e64)); 827 else if (TRI->isVGPR(*MRI, UseMI->getOperand(0).getReg()) && 828 TRI->isAGPR(*MRI, UseMI->getOperand(1).getReg())) 829 UseMI->setDesc(TII->get(AMDGPU::V_ACCVGPR_READ_B32_e64)); 830 else if (ST->hasGFX90AInsts() && 831 TRI->isAGPR(*MRI, UseMI->getOperand(0).getReg()) && 832 TRI->isAGPR(*MRI, UseMI->getOperand(1).getReg())) 833 UseMI->setDesc(TII->get(AMDGPU::V_ACCVGPR_MOV_B32)); 834 return; 835 } 836 837 unsigned UseOpc = UseMI->getOpcode(); 838 if (UseOpc == AMDGPU::V_READFIRSTLANE_B32 || 839 (UseOpc == AMDGPU::V_READLANE_B32 && 840 (int)UseOpIdx == 841 AMDGPU::getNamedOperandIdx(UseOpc, AMDGPU::OpName::src0))) { 842 // %vgpr = V_MOV_B32 imm 843 // %sgpr = V_READFIRSTLANE_B32 %vgpr 844 // => 845 // %sgpr = S_MOV_B32 imm 846 if (FoldingImmLike) { 847 if (execMayBeModifiedBeforeUse(*MRI, 848 UseMI->getOperand(UseOpIdx).getReg(), 849 *OpToFold.getParent(), 850 *UseMI)) 851 return; 852 853 UseMI->setDesc(TII->get(AMDGPU::S_MOV_B32)); 854 855 if (OpToFold.isImm()) 856 UseMI->getOperand(1).ChangeToImmediate(OpToFold.getImm()); 857 else 858 UseMI->getOperand(1).ChangeToFrameIndex(OpToFold.getIndex()); 859 UseMI->removeOperand(2); // Remove exec read (or src1 for readlane) 860 return; 861 } 862 863 if (OpToFold.isReg() && TRI->isSGPRReg(*MRI, OpToFold.getReg())) { 864 if (execMayBeModifiedBeforeUse(*MRI, 865 UseMI->getOperand(UseOpIdx).getReg(), 866 *OpToFold.getParent(), 867 *UseMI)) 868 return; 869 870 // %vgpr = COPY %sgpr0 871 // %sgpr1 = V_READFIRSTLANE_B32 %vgpr 872 // => 873 // %sgpr1 = COPY %sgpr0 874 UseMI->setDesc(TII->get(AMDGPU::COPY)); 875 UseMI->getOperand(1).setReg(OpToFold.getReg()); 876 UseMI->getOperand(1).setSubReg(OpToFold.getSubReg()); 877 UseMI->getOperand(1).setIsKill(false); 878 UseMI->removeOperand(2); // Remove exec read (or src1 for readlane) 879 return; 880 } 881 } 882 883 const MCInstrDesc &UseDesc = UseMI->getDesc(); 884 885 // Don't fold into target independent nodes. Target independent opcodes 886 // don't have defined register classes. 887 if (UseDesc.isVariadic() || 888 UseOp.isImplicit() || 889 UseDesc.OpInfo[UseOpIdx].RegClass == -1) 890 return; 891 } 892 893 if (!FoldingImmLike) { 894 if (OpToFold.isReg() && ST->needsAlignedVGPRs()) { 895 // Don't fold if OpToFold doesn't hold an aligned register. 896 const TargetRegisterClass *RC = 897 TRI->getRegClassForReg(*MRI, OpToFold.getReg()); 898 if (TRI->hasVectorRegisters(RC) && OpToFold.getSubReg()) { 899 unsigned SubReg = OpToFold.getSubReg(); 900 const TargetRegisterClass *SubRC = TRI->getSubRegClass(RC, SubReg); 901 RC = TRI->getCompatibleSubRegClass(RC, SubRC, SubReg); 902 if (RC) 903 RC = SubRC; 904 } 905 906 if (!RC || !TRI->isProperlyAlignedRC(*RC)) 907 return; 908 } 909 910 tryAddToFoldList(FoldList, UseMI, UseOpIdx, &OpToFold, TII); 911 912 // FIXME: We could try to change the instruction from 64-bit to 32-bit 913 // to enable more folding opportunities. The shrink operands pass 914 // already does this. 915 return; 916 } 917 918 919 const MCInstrDesc &FoldDesc = OpToFold.getParent()->getDesc(); 920 const TargetRegisterClass *FoldRC = 921 TRI->getRegClass(FoldDesc.OpInfo[0].RegClass); 922 923 // Split 64-bit constants into 32-bits for folding. 924 if (UseOp.getSubReg() && AMDGPU::getRegBitWidth(FoldRC->getID()) == 64) { 925 Register UseReg = UseOp.getReg(); 926 const TargetRegisterClass *UseRC = MRI->getRegClass(UseReg); 927 928 if (AMDGPU::getRegBitWidth(UseRC->getID()) != 64) 929 return; 930 931 APInt Imm(64, OpToFold.getImm()); 932 if (UseOp.getSubReg() == AMDGPU::sub0) { 933 Imm = Imm.getLoBits(32); 934 } else { 935 assert(UseOp.getSubReg() == AMDGPU::sub1); 936 Imm = Imm.getHiBits(32); 937 } 938 939 MachineOperand ImmOp = MachineOperand::CreateImm(Imm.getSExtValue()); 940 tryAddToFoldList(FoldList, UseMI, UseOpIdx, &ImmOp, TII); 941 return; 942 } 943 944 945 946 tryAddToFoldList(FoldList, UseMI, UseOpIdx, &OpToFold, TII); 947 } 948 949 static bool evalBinaryInstruction(unsigned Opcode, int32_t &Result, 950 uint32_t LHS, uint32_t RHS) { 951 switch (Opcode) { 952 case AMDGPU::V_AND_B32_e64: 953 case AMDGPU::V_AND_B32_e32: 954 case AMDGPU::S_AND_B32: 955 Result = LHS & RHS; 956 return true; 957 case AMDGPU::V_OR_B32_e64: 958 case AMDGPU::V_OR_B32_e32: 959 case AMDGPU::S_OR_B32: 960 Result = LHS | RHS; 961 return true; 962 case AMDGPU::V_XOR_B32_e64: 963 case AMDGPU::V_XOR_B32_e32: 964 case AMDGPU::S_XOR_B32: 965 Result = LHS ^ RHS; 966 return true; 967 case AMDGPU::S_XNOR_B32: 968 Result = ~(LHS ^ RHS); 969 return true; 970 case AMDGPU::S_NAND_B32: 971 Result = ~(LHS & RHS); 972 return true; 973 case AMDGPU::S_NOR_B32: 974 Result = ~(LHS | RHS); 975 return true; 976 case AMDGPU::S_ANDN2_B32: 977 Result = LHS & ~RHS; 978 return true; 979 case AMDGPU::S_ORN2_B32: 980 Result = LHS | ~RHS; 981 return true; 982 case AMDGPU::V_LSHL_B32_e64: 983 case AMDGPU::V_LSHL_B32_e32: 984 case AMDGPU::S_LSHL_B32: 985 // The instruction ignores the high bits for out of bounds shifts. 986 Result = LHS << (RHS & 31); 987 return true; 988 case AMDGPU::V_LSHLREV_B32_e64: 989 case AMDGPU::V_LSHLREV_B32_e32: 990 Result = RHS << (LHS & 31); 991 return true; 992 case AMDGPU::V_LSHR_B32_e64: 993 case AMDGPU::V_LSHR_B32_e32: 994 case AMDGPU::S_LSHR_B32: 995 Result = LHS >> (RHS & 31); 996 return true; 997 case AMDGPU::V_LSHRREV_B32_e64: 998 case AMDGPU::V_LSHRREV_B32_e32: 999 Result = RHS >> (LHS & 31); 1000 return true; 1001 case AMDGPU::V_ASHR_I32_e64: 1002 case AMDGPU::V_ASHR_I32_e32: 1003 case AMDGPU::S_ASHR_I32: 1004 Result = static_cast<int32_t>(LHS) >> (RHS & 31); 1005 return true; 1006 case AMDGPU::V_ASHRREV_I32_e64: 1007 case AMDGPU::V_ASHRREV_I32_e32: 1008 Result = static_cast<int32_t>(RHS) >> (LHS & 31); 1009 return true; 1010 default: 1011 return false; 1012 } 1013 } 1014 1015 static unsigned getMovOpc(bool IsScalar) { 1016 return IsScalar ? AMDGPU::S_MOV_B32 : AMDGPU::V_MOV_B32_e32; 1017 } 1018 1019 /// Remove any leftover implicit operands from mutating the instruction. e.g. 1020 /// if we replace an s_and_b32 with a copy, we don't need the implicit scc def 1021 /// anymore. 1022 static void stripExtraCopyOperands(MachineInstr &MI) { 1023 const MCInstrDesc &Desc = MI.getDesc(); 1024 unsigned NumOps = Desc.getNumOperands() + 1025 Desc.getNumImplicitUses() + 1026 Desc.getNumImplicitDefs(); 1027 1028 for (unsigned I = MI.getNumOperands() - 1; I >= NumOps; --I) 1029 MI.removeOperand(I); 1030 } 1031 1032 static void mutateCopyOp(MachineInstr &MI, const MCInstrDesc &NewDesc) { 1033 MI.setDesc(NewDesc); 1034 stripExtraCopyOperands(MI); 1035 } 1036 1037 static MachineOperand *getImmOrMaterializedImm(MachineRegisterInfo &MRI, 1038 MachineOperand &Op) { 1039 if (Op.isReg()) { 1040 // If this has a subregister, it obviously is a register source. 1041 if (Op.getSubReg() != AMDGPU::NoSubRegister || !Op.getReg().isVirtual()) 1042 return &Op; 1043 1044 MachineInstr *Def = MRI.getVRegDef(Op.getReg()); 1045 if (Def && Def->isMoveImmediate()) { 1046 MachineOperand &ImmSrc = Def->getOperand(1); 1047 if (ImmSrc.isImm()) 1048 return &ImmSrc; 1049 } 1050 } 1051 1052 return &Op; 1053 } 1054 1055 // Try to simplify operations with a constant that may appear after instruction 1056 // selection. 1057 // TODO: See if a frame index with a fixed offset can fold. 1058 static bool tryConstantFoldOp(MachineRegisterInfo &MRI, const SIInstrInfo *TII, 1059 MachineInstr *MI) { 1060 unsigned Opc = MI->getOpcode(); 1061 1062 int Src0Idx = AMDGPU::getNamedOperandIdx(Opc, AMDGPU::OpName::src0); 1063 if (Src0Idx == -1) 1064 return false; 1065 MachineOperand *Src0 = getImmOrMaterializedImm(MRI, MI->getOperand(Src0Idx)); 1066 1067 if ((Opc == AMDGPU::V_NOT_B32_e64 || Opc == AMDGPU::V_NOT_B32_e32 || 1068 Opc == AMDGPU::S_NOT_B32) && 1069 Src0->isImm()) { 1070 MI->getOperand(1).ChangeToImmediate(~Src0->getImm()); 1071 mutateCopyOp(*MI, TII->get(getMovOpc(Opc == AMDGPU::S_NOT_B32))); 1072 return true; 1073 } 1074 1075 int Src1Idx = AMDGPU::getNamedOperandIdx(Opc, AMDGPU::OpName::src1); 1076 if (Src1Idx == -1) 1077 return false; 1078 MachineOperand *Src1 = getImmOrMaterializedImm(MRI, MI->getOperand(Src1Idx)); 1079 1080 if (!Src0->isImm() && !Src1->isImm()) 1081 return false; 1082 1083 // and k0, k1 -> v_mov_b32 (k0 & k1) 1084 // or k0, k1 -> v_mov_b32 (k0 | k1) 1085 // xor k0, k1 -> v_mov_b32 (k0 ^ k1) 1086 if (Src0->isImm() && Src1->isImm()) { 1087 int32_t NewImm; 1088 if (!evalBinaryInstruction(Opc, NewImm, Src0->getImm(), Src1->getImm())) 1089 return false; 1090 1091 const SIRegisterInfo &TRI = TII->getRegisterInfo(); 1092 bool IsSGPR = TRI.isSGPRReg(MRI, MI->getOperand(0).getReg()); 1093 1094 // Be careful to change the right operand, src0 may belong to a different 1095 // instruction. 1096 MI->getOperand(Src0Idx).ChangeToImmediate(NewImm); 1097 MI->removeOperand(Src1Idx); 1098 mutateCopyOp(*MI, TII->get(getMovOpc(IsSGPR))); 1099 return true; 1100 } 1101 1102 if (!MI->isCommutable()) 1103 return false; 1104 1105 if (Src0->isImm() && !Src1->isImm()) { 1106 std::swap(Src0, Src1); 1107 std::swap(Src0Idx, Src1Idx); 1108 } 1109 1110 int32_t Src1Val = static_cast<int32_t>(Src1->getImm()); 1111 if (Opc == AMDGPU::V_OR_B32_e64 || 1112 Opc == AMDGPU::V_OR_B32_e32 || 1113 Opc == AMDGPU::S_OR_B32) { 1114 if (Src1Val == 0) { 1115 // y = or x, 0 => y = copy x 1116 MI->removeOperand(Src1Idx); 1117 mutateCopyOp(*MI, TII->get(AMDGPU::COPY)); 1118 } else if (Src1Val == -1) { 1119 // y = or x, -1 => y = v_mov_b32 -1 1120 MI->removeOperand(Src1Idx); 1121 mutateCopyOp(*MI, TII->get(getMovOpc(Opc == AMDGPU::S_OR_B32))); 1122 } else 1123 return false; 1124 1125 return true; 1126 } 1127 1128 if (MI->getOpcode() == AMDGPU::V_AND_B32_e64 || 1129 MI->getOpcode() == AMDGPU::V_AND_B32_e32 || 1130 MI->getOpcode() == AMDGPU::S_AND_B32) { 1131 if (Src1Val == 0) { 1132 // y = and x, 0 => y = v_mov_b32 0 1133 MI->removeOperand(Src0Idx); 1134 mutateCopyOp(*MI, TII->get(getMovOpc(Opc == AMDGPU::S_AND_B32))); 1135 } else if (Src1Val == -1) { 1136 // y = and x, -1 => y = copy x 1137 MI->removeOperand(Src1Idx); 1138 mutateCopyOp(*MI, TII->get(AMDGPU::COPY)); 1139 stripExtraCopyOperands(*MI); 1140 } else 1141 return false; 1142 1143 return true; 1144 } 1145 1146 if (MI->getOpcode() == AMDGPU::V_XOR_B32_e64 || 1147 MI->getOpcode() == AMDGPU::V_XOR_B32_e32 || 1148 MI->getOpcode() == AMDGPU::S_XOR_B32) { 1149 if (Src1Val == 0) { 1150 // y = xor x, 0 => y = copy x 1151 MI->removeOperand(Src1Idx); 1152 mutateCopyOp(*MI, TII->get(AMDGPU::COPY)); 1153 return true; 1154 } 1155 } 1156 1157 return false; 1158 } 1159 1160 // Try to fold an instruction into a simpler one 1161 bool SIFoldOperands::tryFoldCndMask(MachineInstr &MI) const { 1162 unsigned Opc = MI.getOpcode(); 1163 if (Opc != AMDGPU::V_CNDMASK_B32_e32 && Opc != AMDGPU::V_CNDMASK_B32_e64 && 1164 Opc != AMDGPU::V_CNDMASK_B64_PSEUDO) 1165 return false; 1166 1167 MachineOperand *Src0 = TII->getNamedOperand(MI, AMDGPU::OpName::src0); 1168 MachineOperand *Src1 = TII->getNamedOperand(MI, AMDGPU::OpName::src1); 1169 if (!Src1->isIdenticalTo(*Src0)) { 1170 auto *Src0Imm = getImmOrMaterializedImm(*MRI, *Src0); 1171 auto *Src1Imm = getImmOrMaterializedImm(*MRI, *Src1); 1172 if (!Src1Imm->isIdenticalTo(*Src0Imm)) 1173 return false; 1174 } 1175 1176 int Src1ModIdx = 1177 AMDGPU::getNamedOperandIdx(Opc, AMDGPU::OpName::src1_modifiers); 1178 int Src0ModIdx = 1179 AMDGPU::getNamedOperandIdx(Opc, AMDGPU::OpName::src0_modifiers); 1180 if ((Src1ModIdx != -1 && MI.getOperand(Src1ModIdx).getImm() != 0) || 1181 (Src0ModIdx != -1 && MI.getOperand(Src0ModIdx).getImm() != 0)) 1182 return false; 1183 1184 LLVM_DEBUG(dbgs() << "Folded " << MI << " into "); 1185 auto &NewDesc = 1186 TII->get(Src0->isReg() ? (unsigned)AMDGPU::COPY : getMovOpc(false)); 1187 int Src2Idx = AMDGPU::getNamedOperandIdx(Opc, AMDGPU::OpName::src2); 1188 if (Src2Idx != -1) 1189 MI.removeOperand(Src2Idx); 1190 MI.removeOperand(AMDGPU::getNamedOperandIdx(Opc, AMDGPU::OpName::src1)); 1191 if (Src1ModIdx != -1) 1192 MI.removeOperand(Src1ModIdx); 1193 if (Src0ModIdx != -1) 1194 MI.removeOperand(Src0ModIdx); 1195 mutateCopyOp(MI, NewDesc); 1196 LLVM_DEBUG(dbgs() << MI); 1197 return true; 1198 } 1199 1200 bool SIFoldOperands::tryFoldZeroHighBits(MachineInstr &MI) const { 1201 if (MI.getOpcode() != AMDGPU::V_AND_B32_e64 && 1202 MI.getOpcode() != AMDGPU::V_AND_B32_e32) 1203 return false; 1204 1205 MachineOperand *Src0 = getImmOrMaterializedImm(*MRI, MI.getOperand(1)); 1206 if (!Src0->isImm() || Src0->getImm() != 0xffff) 1207 return false; 1208 1209 Register Src1 = MI.getOperand(2).getReg(); 1210 MachineInstr *SrcDef = MRI->getVRegDef(Src1); 1211 if (ST->zeroesHigh16BitsOfDest(SrcDef->getOpcode())) { 1212 Register Dst = MI.getOperand(0).getReg(); 1213 MRI->replaceRegWith(Dst, SrcDef->getOperand(0).getReg()); 1214 MI.eraseFromParent(); 1215 return true; 1216 } 1217 1218 return false; 1219 } 1220 1221 bool SIFoldOperands::foldInstOperand(MachineInstr &MI, 1222 MachineOperand &OpToFold) const { 1223 // We need mutate the operands of new mov instructions to add implicit 1224 // uses of EXEC, but adding them invalidates the use_iterator, so defer 1225 // this. 1226 SmallVector<MachineInstr *, 4> CopiesToReplace; 1227 SmallVector<FoldCandidate, 4> FoldList; 1228 MachineOperand &Dst = MI.getOperand(0); 1229 bool Changed = false; 1230 1231 if (OpToFold.isImm()) { 1232 for (auto &UseMI : 1233 make_early_inc_range(MRI->use_nodbg_instructions(Dst.getReg()))) { 1234 // Folding the immediate may reveal operations that can be constant 1235 // folded or replaced with a copy. This can happen for example after 1236 // frame indices are lowered to constants or from splitting 64-bit 1237 // constants. 1238 // 1239 // We may also encounter cases where one or both operands are 1240 // immediates materialized into a register, which would ordinarily not 1241 // be folded due to multiple uses or operand constraints. 1242 if (tryConstantFoldOp(*MRI, TII, &UseMI)) { 1243 LLVM_DEBUG(dbgs() << "Constant folded " << UseMI); 1244 Changed = true; 1245 } 1246 } 1247 } 1248 1249 SmallVector<MachineOperand *, 4> UsesToProcess; 1250 for (auto &Use : MRI->use_nodbg_operands(Dst.getReg())) 1251 UsesToProcess.push_back(&Use); 1252 for (auto U : UsesToProcess) { 1253 MachineInstr *UseMI = U->getParent(); 1254 foldOperand(OpToFold, UseMI, UseMI->getOperandNo(U), FoldList, 1255 CopiesToReplace); 1256 } 1257 1258 if (CopiesToReplace.empty() && FoldList.empty()) 1259 return Changed; 1260 1261 MachineFunction *MF = MI.getParent()->getParent(); 1262 // Make sure we add EXEC uses to any new v_mov instructions created. 1263 for (MachineInstr *Copy : CopiesToReplace) 1264 Copy->addImplicitDefUseOperands(*MF); 1265 1266 for (FoldCandidate &Fold : FoldList) { 1267 assert(!Fold.isReg() || Fold.OpToFold); 1268 if (Fold.isReg() && Fold.OpToFold->getReg().isVirtual()) { 1269 Register Reg = Fold.OpToFold->getReg(); 1270 MachineInstr *DefMI = Fold.OpToFold->getParent(); 1271 if (DefMI->readsRegister(AMDGPU::EXEC, TRI) && 1272 execMayBeModifiedBeforeUse(*MRI, Reg, *DefMI, *Fold.UseMI)) 1273 continue; 1274 } 1275 if (updateOperand(Fold, *TII, *TRI, *ST)) { 1276 // Clear kill flags. 1277 if (Fold.isReg()) { 1278 assert(Fold.OpToFold && Fold.OpToFold->isReg()); 1279 // FIXME: Probably shouldn't bother trying to fold if not an 1280 // SGPR. PeepholeOptimizer can eliminate redundant VGPR->VGPR 1281 // copies. 1282 MRI->clearKillFlags(Fold.OpToFold->getReg()); 1283 } 1284 LLVM_DEBUG(dbgs() << "Folded source from " << MI << " into OpNo " 1285 << static_cast<int>(Fold.UseOpNo) << " of " 1286 << *Fold.UseMI); 1287 } else if (Fold.isCommuted()) { 1288 // Restoring instruction's original operand order if fold has failed. 1289 TII->commuteInstruction(*Fold.UseMI, false); 1290 } 1291 } 1292 return true; 1293 } 1294 1295 // Clamp patterns are canonically selected to v_max_* instructions, so only 1296 // handle them. 1297 const MachineOperand *SIFoldOperands::isClamp(const MachineInstr &MI) const { 1298 unsigned Op = MI.getOpcode(); 1299 switch (Op) { 1300 case AMDGPU::V_MAX_F32_e64: 1301 case AMDGPU::V_MAX_F16_e64: 1302 case AMDGPU::V_MAX_F64_e64: 1303 case AMDGPU::V_PK_MAX_F16: { 1304 if (!TII->getNamedOperand(MI, AMDGPU::OpName::clamp)->getImm()) 1305 return nullptr; 1306 1307 // Make sure sources are identical. 1308 const MachineOperand *Src0 = TII->getNamedOperand(MI, AMDGPU::OpName::src0); 1309 const MachineOperand *Src1 = TII->getNamedOperand(MI, AMDGPU::OpName::src1); 1310 if (!Src0->isReg() || !Src1->isReg() || 1311 Src0->getReg() != Src1->getReg() || 1312 Src0->getSubReg() != Src1->getSubReg() || 1313 Src0->getSubReg() != AMDGPU::NoSubRegister) 1314 return nullptr; 1315 1316 // Can't fold up if we have modifiers. 1317 if (TII->hasModifiersSet(MI, AMDGPU::OpName::omod)) 1318 return nullptr; 1319 1320 unsigned Src0Mods 1321 = TII->getNamedOperand(MI, AMDGPU::OpName::src0_modifiers)->getImm(); 1322 unsigned Src1Mods 1323 = TII->getNamedOperand(MI, AMDGPU::OpName::src1_modifiers)->getImm(); 1324 1325 // Having a 0 op_sel_hi would require swizzling the output in the source 1326 // instruction, which we can't do. 1327 unsigned UnsetMods = (Op == AMDGPU::V_PK_MAX_F16) ? SISrcMods::OP_SEL_1 1328 : 0u; 1329 if (Src0Mods != UnsetMods && Src1Mods != UnsetMods) 1330 return nullptr; 1331 return Src0; 1332 } 1333 default: 1334 return nullptr; 1335 } 1336 } 1337 1338 // FIXME: Clamp for v_mad_mixhi_f16 handled during isel. 1339 bool SIFoldOperands::tryFoldClamp(MachineInstr &MI) { 1340 const MachineOperand *ClampSrc = isClamp(MI); 1341 if (!ClampSrc || !MRI->hasOneNonDBGUser(ClampSrc->getReg())) 1342 return false; 1343 1344 MachineInstr *Def = MRI->getVRegDef(ClampSrc->getReg()); 1345 1346 // The type of clamp must be compatible. 1347 if (TII->getClampMask(*Def) != TII->getClampMask(MI)) 1348 return false; 1349 1350 MachineOperand *DefClamp = TII->getNamedOperand(*Def, AMDGPU::OpName::clamp); 1351 if (!DefClamp) 1352 return false; 1353 1354 LLVM_DEBUG(dbgs() << "Folding clamp " << *DefClamp << " into " << *Def); 1355 1356 // Clamp is applied after omod, so it is OK if omod is set. 1357 DefClamp->setImm(1); 1358 MRI->replaceRegWith(MI.getOperand(0).getReg(), Def->getOperand(0).getReg()); 1359 MI.eraseFromParent(); 1360 1361 // Use of output modifiers forces VOP3 encoding for a VOP2 mac/fmac 1362 // instruction, so we might as well convert it to the more flexible VOP3-only 1363 // mad/fma form. 1364 if (TII->convertToThreeAddress(*Def, nullptr, nullptr)) 1365 Def->eraseFromParent(); 1366 1367 return true; 1368 } 1369 1370 static int getOModValue(unsigned Opc, int64_t Val) { 1371 switch (Opc) { 1372 case AMDGPU::V_MUL_F64_e64: { 1373 switch (Val) { 1374 case 0x3fe0000000000000: // 0.5 1375 return SIOutMods::DIV2; 1376 case 0x4000000000000000: // 2.0 1377 return SIOutMods::MUL2; 1378 case 0x4010000000000000: // 4.0 1379 return SIOutMods::MUL4; 1380 default: 1381 return SIOutMods::NONE; 1382 } 1383 } 1384 case AMDGPU::V_MUL_F32_e64: { 1385 switch (static_cast<uint32_t>(Val)) { 1386 case 0x3f000000: // 0.5 1387 return SIOutMods::DIV2; 1388 case 0x40000000: // 2.0 1389 return SIOutMods::MUL2; 1390 case 0x40800000: // 4.0 1391 return SIOutMods::MUL4; 1392 default: 1393 return SIOutMods::NONE; 1394 } 1395 } 1396 case AMDGPU::V_MUL_F16_e64: { 1397 switch (static_cast<uint16_t>(Val)) { 1398 case 0x3800: // 0.5 1399 return SIOutMods::DIV2; 1400 case 0x4000: // 2.0 1401 return SIOutMods::MUL2; 1402 case 0x4400: // 4.0 1403 return SIOutMods::MUL4; 1404 default: 1405 return SIOutMods::NONE; 1406 } 1407 } 1408 default: 1409 llvm_unreachable("invalid mul opcode"); 1410 } 1411 } 1412 1413 // FIXME: Does this really not support denormals with f16? 1414 // FIXME: Does this need to check IEEE mode bit? SNaNs are generally not 1415 // handled, so will anything other than that break? 1416 std::pair<const MachineOperand *, int> 1417 SIFoldOperands::isOMod(const MachineInstr &MI) const { 1418 unsigned Op = MI.getOpcode(); 1419 switch (Op) { 1420 case AMDGPU::V_MUL_F64_e64: 1421 case AMDGPU::V_MUL_F32_e64: 1422 case AMDGPU::V_MUL_F16_e64: { 1423 // If output denormals are enabled, omod is ignored. 1424 if ((Op == AMDGPU::V_MUL_F32_e64 && MFI->getMode().FP32OutputDenormals) || 1425 ((Op == AMDGPU::V_MUL_F64_e64 || Op == AMDGPU::V_MUL_F16_e64) && 1426 MFI->getMode().FP64FP16OutputDenormals)) 1427 return std::make_pair(nullptr, SIOutMods::NONE); 1428 1429 const MachineOperand *RegOp = nullptr; 1430 const MachineOperand *ImmOp = nullptr; 1431 const MachineOperand *Src0 = TII->getNamedOperand(MI, AMDGPU::OpName::src0); 1432 const MachineOperand *Src1 = TII->getNamedOperand(MI, AMDGPU::OpName::src1); 1433 if (Src0->isImm()) { 1434 ImmOp = Src0; 1435 RegOp = Src1; 1436 } else if (Src1->isImm()) { 1437 ImmOp = Src1; 1438 RegOp = Src0; 1439 } else 1440 return std::make_pair(nullptr, SIOutMods::NONE); 1441 1442 int OMod = getOModValue(Op, ImmOp->getImm()); 1443 if (OMod == SIOutMods::NONE || 1444 TII->hasModifiersSet(MI, AMDGPU::OpName::src0_modifiers) || 1445 TII->hasModifiersSet(MI, AMDGPU::OpName::src1_modifiers) || 1446 TII->hasModifiersSet(MI, AMDGPU::OpName::omod) || 1447 TII->hasModifiersSet(MI, AMDGPU::OpName::clamp)) 1448 return std::make_pair(nullptr, SIOutMods::NONE); 1449 1450 return std::make_pair(RegOp, OMod); 1451 } 1452 case AMDGPU::V_ADD_F64_e64: 1453 case AMDGPU::V_ADD_F32_e64: 1454 case AMDGPU::V_ADD_F16_e64: { 1455 // If output denormals are enabled, omod is ignored. 1456 if ((Op == AMDGPU::V_ADD_F32_e64 && MFI->getMode().FP32OutputDenormals) || 1457 ((Op == AMDGPU::V_ADD_F64_e64 || Op == AMDGPU::V_ADD_F16_e64) && 1458 MFI->getMode().FP64FP16OutputDenormals)) 1459 return std::make_pair(nullptr, SIOutMods::NONE); 1460 1461 // Look through the DAGCombiner canonicalization fmul x, 2 -> fadd x, x 1462 const MachineOperand *Src0 = TII->getNamedOperand(MI, AMDGPU::OpName::src0); 1463 const MachineOperand *Src1 = TII->getNamedOperand(MI, AMDGPU::OpName::src1); 1464 1465 if (Src0->isReg() && Src1->isReg() && Src0->getReg() == Src1->getReg() && 1466 Src0->getSubReg() == Src1->getSubReg() && 1467 !TII->hasModifiersSet(MI, AMDGPU::OpName::src0_modifiers) && 1468 !TII->hasModifiersSet(MI, AMDGPU::OpName::src1_modifiers) && 1469 !TII->hasModifiersSet(MI, AMDGPU::OpName::clamp) && 1470 !TII->hasModifiersSet(MI, AMDGPU::OpName::omod)) 1471 return std::make_pair(Src0, SIOutMods::MUL2); 1472 1473 return std::make_pair(nullptr, SIOutMods::NONE); 1474 } 1475 default: 1476 return std::make_pair(nullptr, SIOutMods::NONE); 1477 } 1478 } 1479 1480 // FIXME: Does this need to check IEEE bit on function? 1481 bool SIFoldOperands::tryFoldOMod(MachineInstr &MI) { 1482 const MachineOperand *RegOp; 1483 int OMod; 1484 std::tie(RegOp, OMod) = isOMod(MI); 1485 if (OMod == SIOutMods::NONE || !RegOp->isReg() || 1486 RegOp->getSubReg() != AMDGPU::NoSubRegister || 1487 !MRI->hasOneNonDBGUser(RegOp->getReg())) 1488 return false; 1489 1490 MachineInstr *Def = MRI->getVRegDef(RegOp->getReg()); 1491 MachineOperand *DefOMod = TII->getNamedOperand(*Def, AMDGPU::OpName::omod); 1492 if (!DefOMod || DefOMod->getImm() != SIOutMods::NONE) 1493 return false; 1494 1495 // Clamp is applied after omod. If the source already has clamp set, don't 1496 // fold it. 1497 if (TII->hasModifiersSet(*Def, AMDGPU::OpName::clamp)) 1498 return false; 1499 1500 LLVM_DEBUG(dbgs() << "Folding omod " << MI << " into " << *Def); 1501 1502 DefOMod->setImm(OMod); 1503 MRI->replaceRegWith(MI.getOperand(0).getReg(), Def->getOperand(0).getReg()); 1504 MI.eraseFromParent(); 1505 1506 // Use of output modifiers forces VOP3 encoding for a VOP2 mac/fmac 1507 // instruction, so we might as well convert it to the more flexible VOP3-only 1508 // mad/fma form. 1509 if (TII->convertToThreeAddress(*Def, nullptr, nullptr)) 1510 Def->eraseFromParent(); 1511 1512 return true; 1513 } 1514 1515 // Try to fold a reg_sequence with vgpr output and agpr inputs into an 1516 // instruction which can take an agpr. So far that means a store. 1517 bool SIFoldOperands::tryFoldRegSequence(MachineInstr &MI) { 1518 assert(MI.isRegSequence()); 1519 auto Reg = MI.getOperand(0).getReg(); 1520 1521 if (!ST->hasGFX90AInsts() || !TRI->isVGPR(*MRI, Reg) || 1522 !MRI->hasOneNonDBGUse(Reg)) 1523 return false; 1524 1525 SmallVector<std::pair<MachineOperand*, unsigned>, 32> Defs; 1526 if (!getRegSeqInit(Defs, Reg, MCOI::OPERAND_REGISTER, TII, *MRI)) 1527 return false; 1528 1529 for (auto &Def : Defs) { 1530 const auto *Op = Def.first; 1531 if (!Op->isReg()) 1532 return false; 1533 if (TRI->isAGPR(*MRI, Op->getReg())) 1534 continue; 1535 // Maybe this is a COPY from AREG 1536 const MachineInstr *SubDef = MRI->getVRegDef(Op->getReg()); 1537 if (!SubDef || !SubDef->isCopy() || SubDef->getOperand(1).getSubReg()) 1538 return false; 1539 if (!TRI->isAGPR(*MRI, SubDef->getOperand(1).getReg())) 1540 return false; 1541 } 1542 1543 MachineOperand *Op = &*MRI->use_nodbg_begin(Reg); 1544 MachineInstr *UseMI = Op->getParent(); 1545 while (UseMI->isCopy() && !Op->getSubReg()) { 1546 Reg = UseMI->getOperand(0).getReg(); 1547 if (!TRI->isVGPR(*MRI, Reg) || !MRI->hasOneNonDBGUse(Reg)) 1548 return false; 1549 Op = &*MRI->use_nodbg_begin(Reg); 1550 UseMI = Op->getParent(); 1551 } 1552 1553 if (Op->getSubReg()) 1554 return false; 1555 1556 unsigned OpIdx = Op - &UseMI->getOperand(0); 1557 const MCInstrDesc &InstDesc = UseMI->getDesc(); 1558 const TargetRegisterClass *OpRC = 1559 TII->getRegClass(InstDesc, OpIdx, TRI, *MI.getMF()); 1560 if (!OpRC || !TRI->isVectorSuperClass(OpRC)) 1561 return false; 1562 1563 const auto *NewDstRC = TRI->getEquivalentAGPRClass(MRI->getRegClass(Reg)); 1564 auto Dst = MRI->createVirtualRegister(NewDstRC); 1565 auto RS = BuildMI(*MI.getParent(), MI, MI.getDebugLoc(), 1566 TII->get(AMDGPU::REG_SEQUENCE), Dst); 1567 1568 for (unsigned I = 0; I < Defs.size(); ++I) { 1569 MachineOperand *Def = Defs[I].first; 1570 Def->setIsKill(false); 1571 if (TRI->isAGPR(*MRI, Def->getReg())) { 1572 RS.add(*Def); 1573 } else { // This is a copy 1574 MachineInstr *SubDef = MRI->getVRegDef(Def->getReg()); 1575 SubDef->getOperand(1).setIsKill(false); 1576 RS.addReg(SubDef->getOperand(1).getReg(), 0, Def->getSubReg()); 1577 } 1578 RS.addImm(Defs[I].second); 1579 } 1580 1581 Op->setReg(Dst); 1582 if (!TII->isOperandLegal(*UseMI, OpIdx, Op)) { 1583 Op->setReg(Reg); 1584 RS->eraseFromParent(); 1585 return false; 1586 } 1587 1588 LLVM_DEBUG(dbgs() << "Folded " << *RS << " into " << *UseMI); 1589 1590 // Erase the REG_SEQUENCE eagerly, unless we followed a chain of COPY users, 1591 // in which case we can erase them all later in runOnMachineFunction. 1592 if (MRI->use_nodbg_empty(MI.getOperand(0).getReg())) 1593 MI.eraseFromParent(); 1594 return true; 1595 } 1596 1597 // Try to hoist an AGPR to VGPR copy out of the loop across a LCSSA PHI. 1598 // This should allow folding of an AGPR into a consumer which may support it. 1599 // I.e.: 1600 // 1601 // loop: // loop: 1602 // %1:vreg = COPY %0:areg // exit: 1603 // exit: => // %1:areg = PHI %0:areg, %loop 1604 // %2:vreg = PHI %1:vreg, %loop // %2:vreg = COPY %1:areg 1605 bool SIFoldOperands::tryFoldLCSSAPhi(MachineInstr &PHI) { 1606 assert(PHI.isPHI()); 1607 1608 if (PHI.getNumExplicitOperands() != 3) // Single input LCSSA PHI 1609 return false; 1610 1611 Register PhiIn = PHI.getOperand(1).getReg(); 1612 Register PhiOut = PHI.getOperand(0).getReg(); 1613 if (PHI.getOperand(1).getSubReg() || 1614 !TRI->isVGPR(*MRI, PhiIn) || !TRI->isVGPR(*MRI, PhiOut)) 1615 return false; 1616 1617 // A single use should not matter for correctness, but if it has another use 1618 // inside the loop we may perform copy twice in a worst case. 1619 if (!MRI->hasOneNonDBGUse(PhiIn)) 1620 return false; 1621 1622 MachineInstr *Copy = MRI->getVRegDef(PhiIn); 1623 if (!Copy || !Copy->isCopy()) 1624 return false; 1625 1626 Register CopyIn = Copy->getOperand(1).getReg(); 1627 if (!TRI->isAGPR(*MRI, CopyIn) || Copy->getOperand(1).getSubReg()) 1628 return false; 1629 1630 const TargetRegisterClass *ARC = MRI->getRegClass(CopyIn); 1631 Register NewReg = MRI->createVirtualRegister(ARC); 1632 PHI.getOperand(1).setReg(CopyIn); 1633 PHI.getOperand(0).setReg(NewReg); 1634 1635 MachineBasicBlock *MBB = PHI.getParent(); 1636 BuildMI(*MBB, MBB->getFirstNonPHI(), Copy->getDebugLoc(), 1637 TII->get(AMDGPU::COPY), PhiOut) 1638 .addReg(NewReg, RegState::Kill); 1639 Copy->eraseFromParent(); // We know this copy had a single use. 1640 1641 LLVM_DEBUG(dbgs() << "Folded " << PHI); 1642 1643 return true; 1644 } 1645 1646 // Attempt to convert VGPR load to an AGPR load. 1647 bool SIFoldOperands::tryFoldLoad(MachineInstr &MI) { 1648 assert(MI.mayLoad()); 1649 if (!ST->hasGFX90AInsts() || MI.getNumExplicitDefs() != 1) 1650 return false; 1651 1652 MachineOperand &Def = MI.getOperand(0); 1653 if (!Def.isDef()) 1654 return false; 1655 1656 Register DefReg = Def.getReg(); 1657 1658 if (DefReg.isPhysical() || !TRI->isVGPR(*MRI, DefReg)) 1659 return false; 1660 1661 SmallVector<const MachineInstr*, 8> Users; 1662 SmallVector<Register, 8> MoveRegs; 1663 for (const MachineInstr &I : MRI->use_nodbg_instructions(DefReg)) { 1664 Users.push_back(&I); 1665 } 1666 if (Users.empty()) 1667 return false; 1668 1669 // Check that all uses a copy to an agpr or a reg_sequence producing an agpr. 1670 while (!Users.empty()) { 1671 const MachineInstr *I = Users.pop_back_val(); 1672 if (!I->isCopy() && !I->isRegSequence()) 1673 return false; 1674 Register DstReg = I->getOperand(0).getReg(); 1675 if (TRI->isAGPR(*MRI, DstReg)) 1676 continue; 1677 MoveRegs.push_back(DstReg); 1678 for (const MachineInstr &U : MRI->use_nodbg_instructions(DstReg)) { 1679 Users.push_back(&U); 1680 } 1681 } 1682 1683 const TargetRegisterClass *RC = MRI->getRegClass(DefReg); 1684 MRI->setRegClass(DefReg, TRI->getEquivalentAGPRClass(RC)); 1685 if (!TII->isOperandLegal(MI, 0, &Def)) { 1686 MRI->setRegClass(DefReg, RC); 1687 return false; 1688 } 1689 1690 while (!MoveRegs.empty()) { 1691 Register Reg = MoveRegs.pop_back_val(); 1692 MRI->setRegClass(Reg, TRI->getEquivalentAGPRClass(MRI->getRegClass(Reg))); 1693 } 1694 1695 LLVM_DEBUG(dbgs() << "Folded " << MI); 1696 1697 return true; 1698 } 1699 1700 bool SIFoldOperands::runOnMachineFunction(MachineFunction &MF) { 1701 if (skipFunction(MF.getFunction())) 1702 return false; 1703 1704 MRI = &MF.getRegInfo(); 1705 ST = &MF.getSubtarget<GCNSubtarget>(); 1706 TII = ST->getInstrInfo(); 1707 TRI = &TII->getRegisterInfo(); 1708 MFI = MF.getInfo<SIMachineFunctionInfo>(); 1709 1710 // omod is ignored by hardware if IEEE bit is enabled. omod also does not 1711 // correctly handle signed zeros. 1712 // 1713 // FIXME: Also need to check strictfp 1714 bool IsIEEEMode = MFI->getMode().IEEE; 1715 bool HasNSZ = MFI->hasNoSignedZerosFPMath(); 1716 1717 bool Changed = false; 1718 for (MachineBasicBlock *MBB : depth_first(&MF)) { 1719 MachineOperand *CurrentKnownM0Val = nullptr; 1720 for (auto &MI : make_early_inc_range(*MBB)) { 1721 Changed |= tryFoldCndMask(MI); 1722 1723 if (tryFoldZeroHighBits(MI)) { 1724 Changed = true; 1725 continue; 1726 } 1727 1728 if (MI.isRegSequence() && tryFoldRegSequence(MI)) { 1729 Changed = true; 1730 continue; 1731 } 1732 1733 if (MI.isPHI() && tryFoldLCSSAPhi(MI)) { 1734 Changed = true; 1735 continue; 1736 } 1737 1738 if (MI.mayLoad() && tryFoldLoad(MI)) { 1739 Changed = true; 1740 continue; 1741 } 1742 1743 if (!TII->isFoldableCopy(MI)) { 1744 // Saw an unknown clobber of m0, so we no longer know what it is. 1745 if (CurrentKnownM0Val && MI.modifiesRegister(AMDGPU::M0, TRI)) 1746 CurrentKnownM0Val = nullptr; 1747 1748 // TODO: Omod might be OK if there is NSZ only on the source 1749 // instruction, and not the omod multiply. 1750 if (IsIEEEMode || (!HasNSZ && !MI.getFlag(MachineInstr::FmNsz)) || 1751 !tryFoldOMod(MI)) 1752 Changed |= tryFoldClamp(MI); 1753 1754 continue; 1755 } 1756 1757 // Specially track simple redefs of m0 to the same value in a block, so we 1758 // can erase the later ones. 1759 if (MI.getOperand(0).getReg() == AMDGPU::M0) { 1760 MachineOperand &NewM0Val = MI.getOperand(1); 1761 if (CurrentKnownM0Val && CurrentKnownM0Val->isIdenticalTo(NewM0Val)) { 1762 MI.eraseFromParent(); 1763 Changed = true; 1764 continue; 1765 } 1766 1767 // We aren't tracking other physical registers 1768 CurrentKnownM0Val = (NewM0Val.isReg() && NewM0Val.getReg().isPhysical()) ? 1769 nullptr : &NewM0Val; 1770 continue; 1771 } 1772 1773 MachineOperand &OpToFold = MI.getOperand(1); 1774 bool FoldingImm = 1775 OpToFold.isImm() || OpToFold.isFI() || OpToFold.isGlobal(); 1776 1777 // FIXME: We could also be folding things like TargetIndexes. 1778 if (!FoldingImm && !OpToFold.isReg()) 1779 continue; 1780 1781 if (OpToFold.isReg() && !OpToFold.getReg().isVirtual()) 1782 continue; 1783 1784 // Prevent folding operands backwards in the function. For example, 1785 // the COPY opcode must not be replaced by 1 in this example: 1786 // 1787 // %3 = COPY %vgpr0; VGPR_32:%3 1788 // ... 1789 // %vgpr0 = V_MOV_B32_e32 1, implicit %exec 1790 if (!MI.getOperand(0).getReg().isVirtual()) 1791 continue; 1792 1793 Changed |= foldInstOperand(MI, OpToFold); 1794 1795 // If we managed to fold all uses of this copy then we might as well 1796 // delete it now. 1797 // The only reason we need to follow chains of copies here is that 1798 // tryFoldRegSequence looks forward through copies before folding a 1799 // REG_SEQUENCE into its eventual users. 1800 auto *InstToErase = &MI; 1801 while (MRI->use_nodbg_empty(InstToErase->getOperand(0).getReg())) { 1802 auto &SrcOp = InstToErase->getOperand(1); 1803 auto SrcReg = SrcOp.isReg() ? SrcOp.getReg() : Register(); 1804 InstToErase->eraseFromParent(); 1805 Changed = true; 1806 InstToErase = nullptr; 1807 if (!SrcReg || SrcReg.isPhysical()) 1808 break; 1809 InstToErase = MRI->getVRegDef(SrcReg); 1810 if (!InstToErase || !TII->isFoldableCopy(*InstToErase)) 1811 break; 1812 } 1813 if (InstToErase && InstToErase->isRegSequence() && 1814 MRI->use_nodbg_empty(InstToErase->getOperand(0).getReg())) { 1815 InstToErase->eraseFromParent(); 1816 Changed = true; 1817 } 1818 } 1819 } 1820 return Changed; 1821 } 1822