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 void 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 // Wrapper around isInlineConstant that understands special cases when 150 // instruction types are replaced during operand folding. 151 static bool isInlineConstantIfFolded(const SIInstrInfo *TII, 152 const MachineInstr &UseMI, 153 unsigned OpNo, 154 const MachineOperand &OpToFold) { 155 if (TII->isInlineConstant(UseMI, OpNo, OpToFold)) 156 return true; 157 158 unsigned Opc = UseMI.getOpcode(); 159 unsigned NewOpc = macToMad(Opc); 160 if (NewOpc != AMDGPU::INSTRUCTION_LIST_END) { 161 // Special case for mac. Since this is replaced with mad when folded into 162 // src2, we need to check the legality for the final instruction. 163 int Src2Idx = AMDGPU::getNamedOperandIdx(Opc, AMDGPU::OpName::src2); 164 if (static_cast<int>(OpNo) == Src2Idx) { 165 const MCInstrDesc &MadDesc = TII->get(NewOpc); 166 return TII->isInlineConstant(OpToFold, MadDesc.OpInfo[OpNo].OperandType); 167 } 168 } 169 170 return false; 171 } 172 173 // TODO: Add heuristic that the frame index might not fit in the addressing mode 174 // immediate offset to avoid materializing in loops. 175 static bool frameIndexMayFold(const SIInstrInfo *TII, 176 const MachineInstr &UseMI, 177 int OpNo, 178 const MachineOperand &OpToFold) { 179 if (!OpToFold.isFI()) 180 return false; 181 182 if (TII->isMUBUF(UseMI)) 183 return OpNo == AMDGPU::getNamedOperandIdx(UseMI.getOpcode(), 184 AMDGPU::OpName::vaddr); 185 if (!TII->isFLATScratch(UseMI)) 186 return false; 187 188 int SIdx = AMDGPU::getNamedOperandIdx(UseMI.getOpcode(), 189 AMDGPU::OpName::saddr); 190 if (OpNo == SIdx) 191 return true; 192 193 int VIdx = AMDGPU::getNamedOperandIdx(UseMI.getOpcode(), 194 AMDGPU::OpName::vaddr); 195 return OpNo == VIdx && SIdx == -1; 196 } 197 198 FunctionPass *llvm::createSIFoldOperandsPass() { 199 return new SIFoldOperands(); 200 } 201 202 static bool updateOperand(FoldCandidate &Fold, 203 const SIInstrInfo &TII, 204 const TargetRegisterInfo &TRI, 205 const GCNSubtarget &ST) { 206 MachineInstr *MI = Fold.UseMI; 207 MachineOperand &Old = MI->getOperand(Fold.UseOpNo); 208 assert(Old.isReg()); 209 210 if (Fold.isImm()) { 211 if (MI->getDesc().TSFlags & SIInstrFlags::IsPacked && 212 !(MI->getDesc().TSFlags & SIInstrFlags::IsMAI) && 213 AMDGPU::isFoldableLiteralV216(Fold.ImmToFold, 214 ST.hasInv2PiInlineImm())) { 215 // Set op_sel/op_sel_hi on this operand or bail out if op_sel is 216 // already set. 217 unsigned Opcode = MI->getOpcode(); 218 int OpNo = MI->getOperandNo(&Old); 219 int ModIdx = -1; 220 if (OpNo == AMDGPU::getNamedOperandIdx(Opcode, AMDGPU::OpName::src0)) 221 ModIdx = AMDGPU::OpName::src0_modifiers; 222 else if (OpNo == AMDGPU::getNamedOperandIdx(Opcode, AMDGPU::OpName::src1)) 223 ModIdx = AMDGPU::OpName::src1_modifiers; 224 else if (OpNo == AMDGPU::getNamedOperandIdx(Opcode, AMDGPU::OpName::src2)) 225 ModIdx = AMDGPU::OpName::src2_modifiers; 226 assert(ModIdx != -1); 227 ModIdx = AMDGPU::getNamedOperandIdx(Opcode, ModIdx); 228 MachineOperand &Mod = MI->getOperand(ModIdx); 229 unsigned Val = Mod.getImm(); 230 if (!(Val & SISrcMods::OP_SEL_0) && (Val & SISrcMods::OP_SEL_1)) { 231 // Only apply the following transformation if that operand requires 232 // a packed immediate. 233 switch (TII.get(Opcode).OpInfo[OpNo].OperandType) { 234 case AMDGPU::OPERAND_REG_IMM_V2FP16: 235 case AMDGPU::OPERAND_REG_IMM_V2INT16: 236 case AMDGPU::OPERAND_REG_INLINE_C_V2FP16: 237 case AMDGPU::OPERAND_REG_INLINE_C_V2INT16: 238 // If upper part is all zero we do not need op_sel_hi. 239 if (!isUInt<16>(Fold.ImmToFold)) { 240 if (!(Fold.ImmToFold & 0xffff)) { 241 Mod.setImm(Mod.getImm() | SISrcMods::OP_SEL_0); 242 Mod.setImm(Mod.getImm() & ~SISrcMods::OP_SEL_1); 243 Old.ChangeToImmediate((Fold.ImmToFold >> 16) & 0xffff); 244 return true; 245 } 246 Mod.setImm(Mod.getImm() & ~SISrcMods::OP_SEL_1); 247 Old.ChangeToImmediate(Fold.ImmToFold & 0xffff); 248 return true; 249 } 250 break; 251 default: 252 break; 253 } 254 } 255 } 256 } 257 258 if ((Fold.isImm() || Fold.isFI() || Fold.isGlobal()) && Fold.needsShrink()) { 259 MachineBasicBlock *MBB = MI->getParent(); 260 auto Liveness = MBB->computeRegisterLiveness(&TRI, AMDGPU::VCC, MI, 16); 261 if (Liveness != MachineBasicBlock::LQR_Dead) { 262 LLVM_DEBUG(dbgs() << "Not shrinking " << MI << " due to vcc liveness\n"); 263 return false; 264 } 265 266 MachineRegisterInfo &MRI = MBB->getParent()->getRegInfo(); 267 int Op32 = Fold.getShrinkOpcode(); 268 MachineOperand &Dst0 = MI->getOperand(0); 269 MachineOperand &Dst1 = MI->getOperand(1); 270 assert(Dst0.isDef() && Dst1.isDef()); 271 272 bool HaveNonDbgCarryUse = !MRI.use_nodbg_empty(Dst1.getReg()); 273 274 const TargetRegisterClass *Dst0RC = MRI.getRegClass(Dst0.getReg()); 275 Register NewReg0 = MRI.createVirtualRegister(Dst0RC); 276 277 MachineInstr *Inst32 = TII.buildShrunkInst(*MI, Op32); 278 279 if (HaveNonDbgCarryUse) { 280 BuildMI(*MBB, MI, MI->getDebugLoc(), TII.get(AMDGPU::COPY), Dst1.getReg()) 281 .addReg(AMDGPU::VCC, RegState::Kill); 282 } 283 284 // Keep the old instruction around to avoid breaking iterators, but 285 // replace it with a dummy instruction to remove uses. 286 // 287 // FIXME: We should not invert how this pass looks at operands to avoid 288 // this. Should track set of foldable movs instead of looking for uses 289 // when looking at a use. 290 Dst0.setReg(NewReg0); 291 for (unsigned I = MI->getNumOperands() - 1; I > 0; --I) 292 MI->RemoveOperand(I); 293 MI->setDesc(TII.get(AMDGPU::IMPLICIT_DEF)); 294 295 if (Fold.isCommuted()) 296 TII.commuteInstruction(*Inst32, false); 297 return true; 298 } 299 300 assert(!Fold.needsShrink() && "not handled"); 301 302 if (Fold.isImm()) { 303 if (Old.isTied()) { 304 int NewMFMAOpc = AMDGPU::getMFMAEarlyClobberOp(MI->getOpcode()); 305 if (NewMFMAOpc == -1) 306 return false; 307 MI->setDesc(TII.get(NewMFMAOpc)); 308 MI->untieRegOperand(0); 309 } 310 Old.ChangeToImmediate(Fold.ImmToFold); 311 return true; 312 } 313 314 if (Fold.isGlobal()) { 315 Old.ChangeToGA(Fold.OpToFold->getGlobal(), Fold.OpToFold->getOffset(), 316 Fold.OpToFold->getTargetFlags()); 317 return true; 318 } 319 320 if (Fold.isFI()) { 321 Old.ChangeToFrameIndex(Fold.FrameIndexToFold); 322 return true; 323 } 324 325 MachineOperand *New = Fold.OpToFold; 326 Old.substVirtReg(New->getReg(), New->getSubReg(), TRI); 327 Old.setIsUndef(New->isUndef()); 328 return true; 329 } 330 331 static bool isUseMIInFoldList(ArrayRef<FoldCandidate> FoldList, 332 const MachineInstr *MI) { 333 for (auto Candidate : FoldList) { 334 if (Candidate.UseMI == MI) 335 return true; 336 } 337 return false; 338 } 339 340 static void appendFoldCandidate(SmallVectorImpl<FoldCandidate> &FoldList, 341 MachineInstr *MI, unsigned OpNo, 342 MachineOperand *FoldOp, bool Commuted = false, 343 int ShrinkOp = -1) { 344 // Skip additional folding on the same operand. 345 for (FoldCandidate &Fold : FoldList) 346 if (Fold.UseMI == MI && Fold.UseOpNo == OpNo) 347 return; 348 LLVM_DEBUG(dbgs() << "Append " << (Commuted ? "commuted" : "normal") 349 << " operand " << OpNo << "\n " << *MI); 350 FoldList.emplace_back(MI, OpNo, FoldOp, Commuted, ShrinkOp); 351 } 352 353 static bool tryAddToFoldList(SmallVectorImpl<FoldCandidate> &FoldList, 354 MachineInstr *MI, unsigned OpNo, 355 MachineOperand *OpToFold, 356 const SIInstrInfo *TII) { 357 if (!TII->isOperandLegal(*MI, OpNo, OpToFold)) { 358 // Special case for v_mac_{f16, f32}_e64 if we are trying to fold into src2 359 unsigned Opc = MI->getOpcode(); 360 unsigned NewOpc = macToMad(Opc); 361 if (NewOpc != AMDGPU::INSTRUCTION_LIST_END) { 362 // Check if changing this to a v_mad_{f16, f32} instruction will allow us 363 // to fold the operand. 364 MI->setDesc(TII->get(NewOpc)); 365 bool FoldAsMAD = tryAddToFoldList(FoldList, MI, OpNo, OpToFold, TII); 366 if (FoldAsMAD) { 367 MI->untieRegOperand(OpNo); 368 return true; 369 } 370 MI->setDesc(TII->get(Opc)); 371 } 372 373 // Special case for s_setreg_b32 374 if (OpToFold->isImm()) { 375 unsigned ImmOpc = 0; 376 if (Opc == AMDGPU::S_SETREG_B32) 377 ImmOpc = AMDGPU::S_SETREG_IMM32_B32; 378 else if (Opc == AMDGPU::S_SETREG_B32_mode) 379 ImmOpc = AMDGPU::S_SETREG_IMM32_B32_mode; 380 if (ImmOpc) { 381 MI->setDesc(TII->get(ImmOpc)); 382 appendFoldCandidate(FoldList, MI, OpNo, OpToFold); 383 return true; 384 } 385 } 386 387 // If we are already folding into another operand of MI, then 388 // we can't commute the instruction, otherwise we risk making the 389 // other fold illegal. 390 if (isUseMIInFoldList(FoldList, MI)) 391 return false; 392 393 unsigned CommuteOpNo = OpNo; 394 395 // Operand is not legal, so try to commute the instruction to 396 // see if this makes it possible to fold. 397 unsigned CommuteIdx0 = TargetInstrInfo::CommuteAnyOperandIndex; 398 unsigned CommuteIdx1 = TargetInstrInfo::CommuteAnyOperandIndex; 399 bool CanCommute = TII->findCommutedOpIndices(*MI, CommuteIdx0, CommuteIdx1); 400 401 if (CanCommute) { 402 if (CommuteIdx0 == OpNo) 403 CommuteOpNo = CommuteIdx1; 404 else if (CommuteIdx1 == OpNo) 405 CommuteOpNo = CommuteIdx0; 406 } 407 408 409 // One of operands might be an Imm operand, and OpNo may refer to it after 410 // the call of commuteInstruction() below. Such situations are avoided 411 // here explicitly as OpNo must be a register operand to be a candidate 412 // for memory folding. 413 if (CanCommute && (!MI->getOperand(CommuteIdx0).isReg() || 414 !MI->getOperand(CommuteIdx1).isReg())) 415 return false; 416 417 if (!CanCommute || 418 !TII->commuteInstruction(*MI, false, CommuteIdx0, CommuteIdx1)) 419 return false; 420 421 if (!TII->isOperandLegal(*MI, CommuteOpNo, OpToFold)) { 422 if ((Opc == AMDGPU::V_ADD_CO_U32_e64 || 423 Opc == AMDGPU::V_SUB_CO_U32_e64 || 424 Opc == AMDGPU::V_SUBREV_CO_U32_e64) && // FIXME 425 (OpToFold->isImm() || OpToFold->isFI() || OpToFold->isGlobal())) { 426 MachineRegisterInfo &MRI = MI->getParent()->getParent()->getRegInfo(); 427 428 // Verify the other operand is a VGPR, otherwise we would violate the 429 // constant bus restriction. 430 unsigned OtherIdx = CommuteOpNo == CommuteIdx0 ? CommuteIdx1 : CommuteIdx0; 431 MachineOperand &OtherOp = MI->getOperand(OtherIdx); 432 if (!OtherOp.isReg() || 433 !TII->getRegisterInfo().isVGPR(MRI, OtherOp.getReg())) 434 return false; 435 436 assert(MI->getOperand(1).isDef()); 437 438 // Make sure to get the 32-bit version of the commuted opcode. 439 unsigned MaybeCommutedOpc = MI->getOpcode(); 440 int Op32 = AMDGPU::getVOPe32(MaybeCommutedOpc); 441 442 appendFoldCandidate(FoldList, MI, CommuteOpNo, OpToFold, true, Op32); 443 return true; 444 } 445 446 TII->commuteInstruction(*MI, false, CommuteIdx0, CommuteIdx1); 447 return false; 448 } 449 450 appendFoldCandidate(FoldList, MI, CommuteOpNo, OpToFold, true); 451 return true; 452 } 453 454 // Check the case where we might introduce a second constant operand to a 455 // scalar instruction 456 if (TII->isSALU(MI->getOpcode())) { 457 const MCInstrDesc &InstDesc = MI->getDesc(); 458 const MCOperandInfo &OpInfo = InstDesc.OpInfo[OpNo]; 459 const SIRegisterInfo &SRI = TII->getRegisterInfo(); 460 461 // Fine if the operand can be encoded as an inline constant 462 if (TII->isLiteralConstantLike(*OpToFold, OpInfo)) { 463 if (!SRI.opCanUseInlineConstant(OpInfo.OperandType) || 464 !TII->isInlineConstant(*OpToFold, OpInfo)) { 465 // Otherwise check for another constant 466 for (unsigned i = 0, e = InstDesc.getNumOperands(); i != e; ++i) { 467 auto &Op = MI->getOperand(i); 468 if (OpNo != i && 469 TII->isLiteralConstantLike(Op, OpInfo)) { 470 return false; 471 } 472 } 473 } 474 } 475 } 476 477 appendFoldCandidate(FoldList, MI, OpNo, OpToFold); 478 return true; 479 } 480 481 // If the use operand doesn't care about the value, this may be an operand only 482 // used for register indexing, in which case it is unsafe to fold. 483 static bool isUseSafeToFold(const SIInstrInfo *TII, 484 const MachineInstr &MI, 485 const MachineOperand &UseMO) { 486 if (UseMO.isUndef() || TII->isSDWA(MI)) 487 return false; 488 489 switch (MI.getOpcode()) { 490 case AMDGPU::V_MOV_B32_e32: 491 case AMDGPU::V_MOV_B32_e64: 492 case AMDGPU::V_MOV_B64_PSEUDO: 493 // Do not fold into an indirect mov. 494 return !MI.hasRegisterImplicitUseOperand(AMDGPU::M0); 495 } 496 497 return true; 498 //return !MI.hasRegisterImplicitUseOperand(UseMO.getReg()); 499 } 500 501 // Find a def of the UseReg, check if it is a reg_sequence and find initializers 502 // for each subreg, tracking it to foldable inline immediate if possible. 503 // Returns true on success. 504 static bool getRegSeqInit( 505 SmallVectorImpl<std::pair<MachineOperand*, unsigned>> &Defs, 506 Register UseReg, uint8_t OpTy, 507 const SIInstrInfo *TII, const MachineRegisterInfo &MRI) { 508 MachineInstr *Def = MRI.getVRegDef(UseReg); 509 if (!Def || !Def->isRegSequence()) 510 return false; 511 512 for (unsigned I = 1, E = Def->getNumExplicitOperands(); I < E; I += 2) { 513 MachineOperand *Sub = &Def->getOperand(I); 514 assert(Sub->isReg()); 515 516 for (MachineInstr *SubDef = MRI.getVRegDef(Sub->getReg()); 517 SubDef && Sub->isReg() && Sub->getReg().isVirtual() && 518 !Sub->getSubReg() && TII->isFoldableCopy(*SubDef); 519 SubDef = MRI.getVRegDef(Sub->getReg())) { 520 MachineOperand *Op = &SubDef->getOperand(1); 521 if (Op->isImm()) { 522 if (TII->isInlineConstant(*Op, OpTy)) 523 Sub = Op; 524 break; 525 } 526 if (!Op->isReg() || Op->getReg().isPhysical()) 527 break; 528 Sub = Op; 529 } 530 531 Defs.emplace_back(Sub, Def->getOperand(I + 1).getImm()); 532 } 533 534 return true; 535 } 536 537 static bool tryToFoldACImm(const SIInstrInfo *TII, 538 const MachineOperand &OpToFold, 539 MachineInstr *UseMI, 540 unsigned UseOpIdx, 541 SmallVectorImpl<FoldCandidate> &FoldList) { 542 const MCInstrDesc &Desc = UseMI->getDesc(); 543 const MCOperandInfo *OpInfo = Desc.OpInfo; 544 if (!OpInfo || UseOpIdx >= Desc.getNumOperands()) 545 return false; 546 547 uint8_t OpTy = OpInfo[UseOpIdx].OperandType; 548 if ((OpTy < AMDGPU::OPERAND_REG_INLINE_AC_FIRST || 549 OpTy > AMDGPU::OPERAND_REG_INLINE_AC_LAST) && 550 (OpTy < AMDGPU::OPERAND_REG_INLINE_C_FIRST || 551 OpTy > AMDGPU::OPERAND_REG_INLINE_C_LAST)) 552 return false; 553 554 if (OpToFold.isImm() && TII->isInlineConstant(OpToFold, OpTy) && 555 TII->isOperandLegal(*UseMI, UseOpIdx, &OpToFold)) { 556 UseMI->getOperand(UseOpIdx).ChangeToImmediate(OpToFold.getImm()); 557 return true; 558 } 559 560 if (!OpToFold.isReg()) 561 return false; 562 563 Register UseReg = OpToFold.getReg(); 564 if (!UseReg.isVirtual()) 565 return false; 566 567 if (isUseMIInFoldList(FoldList, UseMI)) 568 return false; 569 570 MachineRegisterInfo &MRI = UseMI->getParent()->getParent()->getRegInfo(); 571 572 // Maybe it is just a COPY of an immediate itself. 573 MachineInstr *Def = MRI.getVRegDef(UseReg); 574 MachineOperand &UseOp = UseMI->getOperand(UseOpIdx); 575 if (!UseOp.getSubReg() && Def && TII->isFoldableCopy(*Def)) { 576 MachineOperand &DefOp = Def->getOperand(1); 577 if (DefOp.isImm() && TII->isInlineConstant(DefOp, OpTy) && 578 TII->isOperandLegal(*UseMI, UseOpIdx, &DefOp)) { 579 UseMI->getOperand(UseOpIdx).ChangeToImmediate(DefOp.getImm()); 580 return true; 581 } 582 } 583 584 SmallVector<std::pair<MachineOperand*, unsigned>, 32> Defs; 585 if (!getRegSeqInit(Defs, UseReg, OpTy, TII, MRI)) 586 return false; 587 588 int32_t Imm; 589 for (unsigned I = 0, E = Defs.size(); I != E; ++I) { 590 const MachineOperand *Op = Defs[I].first; 591 if (!Op->isImm()) 592 return false; 593 594 auto SubImm = Op->getImm(); 595 if (!I) { 596 Imm = SubImm; 597 if (!TII->isInlineConstant(*Op, OpTy) || 598 !TII->isOperandLegal(*UseMI, UseOpIdx, Op)) 599 return false; 600 601 continue; 602 } 603 if (Imm != SubImm) 604 return false; // Can only fold splat constants 605 } 606 607 appendFoldCandidate(FoldList, UseMI, UseOpIdx, Defs[0].first); 608 return true; 609 } 610 611 void SIFoldOperands::foldOperand( 612 MachineOperand &OpToFold, 613 MachineInstr *UseMI, 614 int UseOpIdx, 615 SmallVectorImpl<FoldCandidate> &FoldList, 616 SmallVectorImpl<MachineInstr *> &CopiesToReplace) const { 617 const MachineOperand &UseOp = UseMI->getOperand(UseOpIdx); 618 619 if (!isUseSafeToFold(TII, *UseMI, UseOp)) 620 return; 621 622 // FIXME: Fold operands with subregs. 623 if (UseOp.isReg() && OpToFold.isReg()) { 624 if (UseOp.isImplicit() || UseOp.getSubReg() != AMDGPU::NoSubRegister) 625 return; 626 } 627 628 // Special case for REG_SEQUENCE: We can't fold literals into 629 // REG_SEQUENCE instructions, so we have to fold them into the 630 // uses of REG_SEQUENCE. 631 if (UseMI->isRegSequence()) { 632 Register RegSeqDstReg = UseMI->getOperand(0).getReg(); 633 unsigned RegSeqDstSubReg = UseMI->getOperand(UseOpIdx + 1).getImm(); 634 635 for (auto &RSUse : make_early_inc_range(MRI->use_nodbg_operands(RegSeqDstReg))) { 636 MachineInstr *RSUseMI = RSUse.getParent(); 637 638 if (tryToFoldACImm(TII, UseMI->getOperand(0), RSUseMI, 639 RSUseMI->getOperandNo(&RSUse), FoldList)) 640 continue; 641 642 if (RSUse.getSubReg() != RegSeqDstSubReg) 643 continue; 644 645 foldOperand(OpToFold, RSUseMI, RSUseMI->getOperandNo(&RSUse), FoldList, 646 CopiesToReplace); 647 } 648 649 return; 650 } 651 652 if (tryToFoldACImm(TII, OpToFold, UseMI, UseOpIdx, FoldList)) 653 return; 654 655 if (frameIndexMayFold(TII, *UseMI, UseOpIdx, OpToFold)) { 656 // Verify that this is a stack access. 657 // FIXME: Should probably use stack pseudos before frame lowering. 658 659 if (TII->isMUBUF(*UseMI)) { 660 if (TII->getNamedOperand(*UseMI, AMDGPU::OpName::srsrc)->getReg() != 661 MFI->getScratchRSrcReg()) 662 return; 663 664 // Ensure this is either relative to the current frame or the current 665 // wave. 666 MachineOperand &SOff = 667 *TII->getNamedOperand(*UseMI, AMDGPU::OpName::soffset); 668 if (!SOff.isImm() || SOff.getImm() != 0) 669 return; 670 } 671 672 // A frame index will resolve to a positive constant, so it should always be 673 // safe to fold the addressing mode, even pre-GFX9. 674 UseMI->getOperand(UseOpIdx).ChangeToFrameIndex(OpToFold.getIndex()); 675 676 if (TII->isFLATScratch(*UseMI) && 677 AMDGPU::getNamedOperandIdx(UseMI->getOpcode(), 678 AMDGPU::OpName::vaddr) != -1) { 679 unsigned NewOpc = AMDGPU::getFlatScratchInstSSfromSV(UseMI->getOpcode()); 680 UseMI->setDesc(TII->get(NewOpc)); 681 } 682 683 return; 684 } 685 686 bool FoldingImmLike = 687 OpToFold.isImm() || OpToFold.isFI() || OpToFold.isGlobal(); 688 689 if (FoldingImmLike && UseMI->isCopy()) { 690 Register DestReg = UseMI->getOperand(0).getReg(); 691 Register SrcReg = UseMI->getOperand(1).getReg(); 692 assert(SrcReg.isVirtual()); 693 694 const TargetRegisterClass *SrcRC = MRI->getRegClass(SrcReg); 695 696 // Don't fold into a copy to a physical register with the same class. Doing 697 // so would interfere with the register coalescer's logic which would avoid 698 // redundant initializations. 699 if (DestReg.isPhysical() && SrcRC->contains(DestReg)) 700 return; 701 702 const TargetRegisterClass *DestRC = TRI->getRegClassForReg(*MRI, DestReg); 703 if (!DestReg.isPhysical()) { 704 if (TRI->isSGPRClass(SrcRC) && TRI->hasVectorRegisters(DestRC)) { 705 SmallVector<FoldCandidate, 4> CopyUses; 706 for (auto &Use : MRI->use_nodbg_operands(DestReg)) { 707 // There's no point trying to fold into an implicit operand. 708 if (Use.isImplicit()) 709 continue; 710 711 CopyUses.emplace_back(Use.getParent(), 712 Use.getParent()->getOperandNo(&Use), 713 &UseMI->getOperand(1)); 714 } 715 for (auto &F : CopyUses) { 716 foldOperand(*F.OpToFold, F.UseMI, F.UseOpNo, FoldList, CopiesToReplace); 717 } 718 } 719 720 if (DestRC == &AMDGPU::AGPR_32RegClass && 721 TII->isInlineConstant(OpToFold, AMDGPU::OPERAND_REG_INLINE_C_INT32)) { 722 UseMI->setDesc(TII->get(AMDGPU::V_ACCVGPR_WRITE_B32_e64)); 723 UseMI->getOperand(1).ChangeToImmediate(OpToFold.getImm()); 724 CopiesToReplace.push_back(UseMI); 725 return; 726 } 727 } 728 729 // In order to fold immediates into copies, we need to change the 730 // copy to a MOV. 731 732 unsigned MovOp = TII->getMovOpcode(DestRC); 733 if (MovOp == AMDGPU::COPY) 734 return; 735 736 UseMI->setDesc(TII->get(MovOp)); 737 MachineInstr::mop_iterator ImpOpI = UseMI->implicit_operands().begin(); 738 MachineInstr::mop_iterator ImpOpE = UseMI->implicit_operands().end(); 739 while (ImpOpI != ImpOpE) { 740 MachineInstr::mop_iterator Tmp = ImpOpI; 741 ImpOpI++; 742 UseMI->RemoveOperand(UseMI->getOperandNo(Tmp)); 743 } 744 CopiesToReplace.push_back(UseMI); 745 } else { 746 if (UseMI->isCopy() && OpToFold.isReg() && 747 UseMI->getOperand(0).getReg().isVirtual() && 748 !UseMI->getOperand(1).getSubReg()) { 749 LLVM_DEBUG(dbgs() << "Folding " << OpToFold << "\n into " << *UseMI); 750 unsigned Size = TII->getOpSize(*UseMI, 1); 751 Register UseReg = OpToFold.getReg(); 752 UseMI->getOperand(1).setReg(UseReg); 753 UseMI->getOperand(1).setSubReg(OpToFold.getSubReg()); 754 UseMI->getOperand(1).setIsKill(false); 755 CopiesToReplace.push_back(UseMI); 756 OpToFold.setIsKill(false); 757 758 // That is very tricky to store a value into an AGPR. v_accvgpr_write_b32 759 // can only accept VGPR or inline immediate. Recreate a reg_sequence with 760 // its initializers right here, so we will rematerialize immediates and 761 // avoid copies via different reg classes. 762 SmallVector<std::pair<MachineOperand*, unsigned>, 32> Defs; 763 if (Size > 4 && TRI->isAGPR(*MRI, UseMI->getOperand(0).getReg()) && 764 getRegSeqInit(Defs, UseReg, AMDGPU::OPERAND_REG_INLINE_C_INT32, TII, 765 *MRI)) { 766 const DebugLoc &DL = UseMI->getDebugLoc(); 767 MachineBasicBlock &MBB = *UseMI->getParent(); 768 769 UseMI->setDesc(TII->get(AMDGPU::REG_SEQUENCE)); 770 for (unsigned I = UseMI->getNumOperands() - 1; I > 0; --I) 771 UseMI->RemoveOperand(I); 772 773 MachineInstrBuilder B(*MBB.getParent(), UseMI); 774 DenseMap<TargetInstrInfo::RegSubRegPair, Register> VGPRCopies; 775 SmallSetVector<TargetInstrInfo::RegSubRegPair, 32> SeenAGPRs; 776 for (unsigned I = 0; I < Size / 4; ++I) { 777 MachineOperand *Def = Defs[I].first; 778 TargetInstrInfo::RegSubRegPair CopyToVGPR; 779 if (Def->isImm() && 780 TII->isInlineConstant(*Def, AMDGPU::OPERAND_REG_INLINE_C_INT32)) { 781 int64_t Imm = Def->getImm(); 782 783 auto Tmp = MRI->createVirtualRegister(&AMDGPU::AGPR_32RegClass); 784 BuildMI(MBB, UseMI, DL, 785 TII->get(AMDGPU::V_ACCVGPR_WRITE_B32_e64), Tmp).addImm(Imm); 786 B.addReg(Tmp); 787 } else if (Def->isReg() && TRI->isAGPR(*MRI, Def->getReg())) { 788 auto Src = getRegSubRegPair(*Def); 789 Def->setIsKill(false); 790 if (!SeenAGPRs.insert(Src)) { 791 // We cannot build a reg_sequence out of the same registers, they 792 // must be copied. Better do it here before copyPhysReg() created 793 // several reads to do the AGPR->VGPR->AGPR copy. 794 CopyToVGPR = Src; 795 } else { 796 B.addReg(Src.Reg, Def->isUndef() ? RegState::Undef : 0, 797 Src.SubReg); 798 } 799 } else { 800 assert(Def->isReg()); 801 Def->setIsKill(false); 802 auto Src = getRegSubRegPair(*Def); 803 804 // Direct copy from SGPR to AGPR is not possible. To avoid creation 805 // of exploded copies SGPR->VGPR->AGPR in the copyPhysReg() later, 806 // create a copy here and track if we already have such a copy. 807 if (TRI->isSGPRReg(*MRI, Src.Reg)) { 808 CopyToVGPR = Src; 809 } else { 810 auto Tmp = MRI->createVirtualRegister(&AMDGPU::AGPR_32RegClass); 811 BuildMI(MBB, UseMI, DL, TII->get(AMDGPU::COPY), Tmp).add(*Def); 812 B.addReg(Tmp); 813 } 814 } 815 816 if (CopyToVGPR.Reg) { 817 Register Vgpr; 818 if (VGPRCopies.count(CopyToVGPR)) { 819 Vgpr = VGPRCopies[CopyToVGPR]; 820 } else { 821 Vgpr = MRI->createVirtualRegister(&AMDGPU::VGPR_32RegClass); 822 BuildMI(MBB, UseMI, DL, TII->get(AMDGPU::COPY), Vgpr).add(*Def); 823 VGPRCopies[CopyToVGPR] = Vgpr; 824 } 825 auto Tmp = MRI->createVirtualRegister(&AMDGPU::AGPR_32RegClass); 826 BuildMI(MBB, UseMI, DL, 827 TII->get(AMDGPU::V_ACCVGPR_WRITE_B32_e64), Tmp).addReg(Vgpr); 828 B.addReg(Tmp); 829 } 830 831 B.addImm(Defs[I].second); 832 } 833 LLVM_DEBUG(dbgs() << "Folded " << *UseMI); 834 return; 835 } 836 837 if (Size != 4) 838 return; 839 if (TRI->isAGPR(*MRI, UseMI->getOperand(0).getReg()) && 840 TRI->isVGPR(*MRI, UseMI->getOperand(1).getReg())) 841 UseMI->setDesc(TII->get(AMDGPU::V_ACCVGPR_WRITE_B32_e64)); 842 else if (TRI->isVGPR(*MRI, UseMI->getOperand(0).getReg()) && 843 TRI->isAGPR(*MRI, UseMI->getOperand(1).getReg())) 844 UseMI->setDesc(TII->get(AMDGPU::V_ACCVGPR_READ_B32_e64)); 845 else if (ST->hasGFX90AInsts() && 846 TRI->isAGPR(*MRI, UseMI->getOperand(0).getReg()) && 847 TRI->isAGPR(*MRI, UseMI->getOperand(1).getReg())) 848 UseMI->setDesc(TII->get(AMDGPU::V_ACCVGPR_MOV_B32)); 849 return; 850 } 851 852 unsigned UseOpc = UseMI->getOpcode(); 853 if (UseOpc == AMDGPU::V_READFIRSTLANE_B32 || 854 (UseOpc == AMDGPU::V_READLANE_B32 && 855 (int)UseOpIdx == 856 AMDGPU::getNamedOperandIdx(UseOpc, AMDGPU::OpName::src0))) { 857 // %vgpr = V_MOV_B32 imm 858 // %sgpr = V_READFIRSTLANE_B32 %vgpr 859 // => 860 // %sgpr = S_MOV_B32 imm 861 if (FoldingImmLike) { 862 if (execMayBeModifiedBeforeUse(*MRI, 863 UseMI->getOperand(UseOpIdx).getReg(), 864 *OpToFold.getParent(), 865 *UseMI)) 866 return; 867 868 UseMI->setDesc(TII->get(AMDGPU::S_MOV_B32)); 869 870 if (OpToFold.isImm()) 871 UseMI->getOperand(1).ChangeToImmediate(OpToFold.getImm()); 872 else 873 UseMI->getOperand(1).ChangeToFrameIndex(OpToFold.getIndex()); 874 UseMI->RemoveOperand(2); // Remove exec read (or src1 for readlane) 875 return; 876 } 877 878 if (OpToFold.isReg() && TRI->isSGPRReg(*MRI, OpToFold.getReg())) { 879 if (execMayBeModifiedBeforeUse(*MRI, 880 UseMI->getOperand(UseOpIdx).getReg(), 881 *OpToFold.getParent(), 882 *UseMI)) 883 return; 884 885 // %vgpr = COPY %sgpr0 886 // %sgpr1 = V_READFIRSTLANE_B32 %vgpr 887 // => 888 // %sgpr1 = COPY %sgpr0 889 UseMI->setDesc(TII->get(AMDGPU::COPY)); 890 UseMI->getOperand(1).setReg(OpToFold.getReg()); 891 UseMI->getOperand(1).setSubReg(OpToFold.getSubReg()); 892 UseMI->getOperand(1).setIsKill(false); 893 UseMI->RemoveOperand(2); // Remove exec read (or src1 for readlane) 894 return; 895 } 896 } 897 898 const MCInstrDesc &UseDesc = UseMI->getDesc(); 899 900 // Don't fold into target independent nodes. Target independent opcodes 901 // don't have defined register classes. 902 if (UseDesc.isVariadic() || 903 UseOp.isImplicit() || 904 UseDesc.OpInfo[UseOpIdx].RegClass == -1) 905 return; 906 } 907 908 if (!FoldingImmLike) { 909 tryAddToFoldList(FoldList, UseMI, UseOpIdx, &OpToFold, TII); 910 911 // FIXME: We could try to change the instruction from 64-bit to 32-bit 912 // to enable more folding opportunities. The shrink operands pass 913 // already does this. 914 return; 915 } 916 917 918 const MCInstrDesc &FoldDesc = OpToFold.getParent()->getDesc(); 919 const TargetRegisterClass *FoldRC = 920 TRI->getRegClass(FoldDesc.OpInfo[0].RegClass); 921 922 // Split 64-bit constants into 32-bits for folding. 923 if (UseOp.getSubReg() && AMDGPU::getRegBitWidth(FoldRC->getID()) == 64) { 924 Register UseReg = UseOp.getReg(); 925 const TargetRegisterClass *UseRC = MRI->getRegClass(UseReg); 926 927 if (AMDGPU::getRegBitWidth(UseRC->getID()) != 64) 928 return; 929 930 APInt Imm(64, OpToFold.getImm()); 931 if (UseOp.getSubReg() == AMDGPU::sub0) { 932 Imm = Imm.getLoBits(32); 933 } else { 934 assert(UseOp.getSubReg() == AMDGPU::sub1); 935 Imm = Imm.getHiBits(32); 936 } 937 938 MachineOperand ImmOp = MachineOperand::CreateImm(Imm.getSExtValue()); 939 tryAddToFoldList(FoldList, UseMI, UseOpIdx, &ImmOp, TII); 940 return; 941 } 942 943 944 945 tryAddToFoldList(FoldList, UseMI, UseOpIdx, &OpToFold, TII); 946 } 947 948 static bool evalBinaryInstruction(unsigned Opcode, int32_t &Result, 949 uint32_t LHS, uint32_t RHS) { 950 switch (Opcode) { 951 case AMDGPU::V_AND_B32_e64: 952 case AMDGPU::V_AND_B32_e32: 953 case AMDGPU::S_AND_B32: 954 Result = LHS & RHS; 955 return true; 956 case AMDGPU::V_OR_B32_e64: 957 case AMDGPU::V_OR_B32_e32: 958 case AMDGPU::S_OR_B32: 959 Result = LHS | RHS; 960 return true; 961 case AMDGPU::V_XOR_B32_e64: 962 case AMDGPU::V_XOR_B32_e32: 963 case AMDGPU::S_XOR_B32: 964 Result = LHS ^ RHS; 965 return true; 966 case AMDGPU::S_XNOR_B32: 967 Result = ~(LHS ^ RHS); 968 return true; 969 case AMDGPU::S_NAND_B32: 970 Result = ~(LHS & RHS); 971 return true; 972 case AMDGPU::S_NOR_B32: 973 Result = ~(LHS | RHS); 974 return true; 975 case AMDGPU::S_ANDN2_B32: 976 Result = LHS & ~RHS; 977 return true; 978 case AMDGPU::S_ORN2_B32: 979 Result = LHS | ~RHS; 980 return true; 981 case AMDGPU::V_LSHL_B32_e64: 982 case AMDGPU::V_LSHL_B32_e32: 983 case AMDGPU::S_LSHL_B32: 984 // The instruction ignores the high bits for out of bounds shifts. 985 Result = LHS << (RHS & 31); 986 return true; 987 case AMDGPU::V_LSHLREV_B32_e64: 988 case AMDGPU::V_LSHLREV_B32_e32: 989 Result = RHS << (LHS & 31); 990 return true; 991 case AMDGPU::V_LSHR_B32_e64: 992 case AMDGPU::V_LSHR_B32_e32: 993 case AMDGPU::S_LSHR_B32: 994 Result = LHS >> (RHS & 31); 995 return true; 996 case AMDGPU::V_LSHRREV_B32_e64: 997 case AMDGPU::V_LSHRREV_B32_e32: 998 Result = RHS >> (LHS & 31); 999 return true; 1000 case AMDGPU::V_ASHR_I32_e64: 1001 case AMDGPU::V_ASHR_I32_e32: 1002 case AMDGPU::S_ASHR_I32: 1003 Result = static_cast<int32_t>(LHS) >> (RHS & 31); 1004 return true; 1005 case AMDGPU::V_ASHRREV_I32_e64: 1006 case AMDGPU::V_ASHRREV_I32_e32: 1007 Result = static_cast<int32_t>(RHS) >> (LHS & 31); 1008 return true; 1009 default: 1010 return false; 1011 } 1012 } 1013 1014 static unsigned getMovOpc(bool IsScalar) { 1015 return IsScalar ? AMDGPU::S_MOV_B32 : AMDGPU::V_MOV_B32_e32; 1016 } 1017 1018 /// Remove any leftover implicit operands from mutating the instruction. e.g. 1019 /// if we replace an s_and_b32 with a copy, we don't need the implicit scc def 1020 /// anymore. 1021 static void stripExtraCopyOperands(MachineInstr &MI) { 1022 const MCInstrDesc &Desc = MI.getDesc(); 1023 unsigned NumOps = Desc.getNumOperands() + 1024 Desc.getNumImplicitUses() + 1025 Desc.getNumImplicitDefs(); 1026 1027 for (unsigned I = MI.getNumOperands() - 1; I >= NumOps; --I) 1028 MI.RemoveOperand(I); 1029 } 1030 1031 static void mutateCopyOp(MachineInstr &MI, const MCInstrDesc &NewDesc) { 1032 MI.setDesc(NewDesc); 1033 stripExtraCopyOperands(MI); 1034 } 1035 1036 static MachineOperand *getImmOrMaterializedImm(MachineRegisterInfo &MRI, 1037 MachineOperand &Op) { 1038 if (Op.isReg()) { 1039 // If this has a subregister, it obviously is a register source. 1040 if (Op.getSubReg() != AMDGPU::NoSubRegister || !Op.getReg().isVirtual()) 1041 return &Op; 1042 1043 MachineInstr *Def = MRI.getVRegDef(Op.getReg()); 1044 if (Def && Def->isMoveImmediate()) { 1045 MachineOperand &ImmSrc = Def->getOperand(1); 1046 if (ImmSrc.isImm()) 1047 return &ImmSrc; 1048 } 1049 } 1050 1051 return &Op; 1052 } 1053 1054 // Try to simplify operations with a constant that may appear after instruction 1055 // selection. 1056 // TODO: See if a frame index with a fixed offset can fold. 1057 static bool tryConstantFoldOp(MachineRegisterInfo &MRI, const SIInstrInfo *TII, 1058 MachineInstr *MI) { 1059 unsigned Opc = MI->getOpcode(); 1060 1061 int Src0Idx = AMDGPU::getNamedOperandIdx(Opc, AMDGPU::OpName::src0); 1062 if (Src0Idx == -1) 1063 return false; 1064 MachineOperand *Src0 = getImmOrMaterializedImm(MRI, MI->getOperand(Src0Idx)); 1065 1066 if ((Opc == AMDGPU::V_NOT_B32_e64 || Opc == AMDGPU::V_NOT_B32_e32 || 1067 Opc == AMDGPU::S_NOT_B32) && 1068 Src0->isImm()) { 1069 MI->getOperand(1).ChangeToImmediate(~Src0->getImm()); 1070 mutateCopyOp(*MI, TII->get(getMovOpc(Opc == AMDGPU::S_NOT_B32))); 1071 return true; 1072 } 1073 1074 int Src1Idx = AMDGPU::getNamedOperandIdx(Opc, AMDGPU::OpName::src1); 1075 if (Src1Idx == -1) 1076 return false; 1077 MachineOperand *Src1 = getImmOrMaterializedImm(MRI, MI->getOperand(Src1Idx)); 1078 1079 if (!Src0->isImm() && !Src1->isImm()) 1080 return false; 1081 1082 // and k0, k1 -> v_mov_b32 (k0 & k1) 1083 // or k0, k1 -> v_mov_b32 (k0 | k1) 1084 // xor k0, k1 -> v_mov_b32 (k0 ^ k1) 1085 if (Src0->isImm() && Src1->isImm()) { 1086 int32_t NewImm; 1087 if (!evalBinaryInstruction(Opc, NewImm, Src0->getImm(), Src1->getImm())) 1088 return false; 1089 1090 const SIRegisterInfo &TRI = TII->getRegisterInfo(); 1091 bool IsSGPR = TRI.isSGPRReg(MRI, MI->getOperand(0).getReg()); 1092 1093 // Be careful to change the right operand, src0 may belong to a different 1094 // instruction. 1095 MI->getOperand(Src0Idx).ChangeToImmediate(NewImm); 1096 MI->RemoveOperand(Src1Idx); 1097 mutateCopyOp(*MI, TII->get(getMovOpc(IsSGPR))); 1098 return true; 1099 } 1100 1101 if (!MI->isCommutable()) 1102 return false; 1103 1104 if (Src0->isImm() && !Src1->isImm()) { 1105 std::swap(Src0, Src1); 1106 std::swap(Src0Idx, Src1Idx); 1107 } 1108 1109 int32_t Src1Val = static_cast<int32_t>(Src1->getImm()); 1110 if (Opc == AMDGPU::V_OR_B32_e64 || 1111 Opc == AMDGPU::V_OR_B32_e32 || 1112 Opc == AMDGPU::S_OR_B32) { 1113 if (Src1Val == 0) { 1114 // y = or x, 0 => y = copy x 1115 MI->RemoveOperand(Src1Idx); 1116 mutateCopyOp(*MI, TII->get(AMDGPU::COPY)); 1117 } else if (Src1Val == -1) { 1118 // y = or x, -1 => y = v_mov_b32 -1 1119 MI->RemoveOperand(Src1Idx); 1120 mutateCopyOp(*MI, TII->get(getMovOpc(Opc == AMDGPU::S_OR_B32))); 1121 } else 1122 return false; 1123 1124 return true; 1125 } 1126 1127 if (MI->getOpcode() == AMDGPU::V_AND_B32_e64 || 1128 MI->getOpcode() == AMDGPU::V_AND_B32_e32 || 1129 MI->getOpcode() == AMDGPU::S_AND_B32) { 1130 if (Src1Val == 0) { 1131 // y = and x, 0 => y = v_mov_b32 0 1132 MI->RemoveOperand(Src0Idx); 1133 mutateCopyOp(*MI, TII->get(getMovOpc(Opc == AMDGPU::S_AND_B32))); 1134 } else if (Src1Val == -1) { 1135 // y = and x, -1 => y = copy x 1136 MI->RemoveOperand(Src1Idx); 1137 mutateCopyOp(*MI, TII->get(AMDGPU::COPY)); 1138 stripExtraCopyOperands(*MI); 1139 } else 1140 return false; 1141 1142 return true; 1143 } 1144 1145 if (MI->getOpcode() == AMDGPU::V_XOR_B32_e64 || 1146 MI->getOpcode() == AMDGPU::V_XOR_B32_e32 || 1147 MI->getOpcode() == AMDGPU::S_XOR_B32) { 1148 if (Src1Val == 0) { 1149 // y = xor x, 0 => y = copy x 1150 MI->RemoveOperand(Src1Idx); 1151 mutateCopyOp(*MI, TII->get(AMDGPU::COPY)); 1152 return true; 1153 } 1154 } 1155 1156 return false; 1157 } 1158 1159 // Try to fold an instruction into a simpler one 1160 bool SIFoldOperands::tryFoldCndMask(MachineInstr &MI) const { 1161 unsigned Opc = MI.getOpcode(); 1162 if (Opc != AMDGPU::V_CNDMASK_B32_e32 && Opc != AMDGPU::V_CNDMASK_B32_e64 && 1163 Opc != AMDGPU::V_CNDMASK_B64_PSEUDO) 1164 return false; 1165 1166 MachineOperand *Src0 = TII->getNamedOperand(MI, AMDGPU::OpName::src0); 1167 MachineOperand *Src1 = TII->getNamedOperand(MI, AMDGPU::OpName::src1); 1168 if (!Src1->isIdenticalTo(*Src0)) { 1169 auto *Src0Imm = getImmOrMaterializedImm(*MRI, *Src0); 1170 auto *Src1Imm = getImmOrMaterializedImm(*MRI, *Src1); 1171 if (!Src1Imm->isIdenticalTo(*Src0Imm)) 1172 return false; 1173 } 1174 1175 int Src1ModIdx = 1176 AMDGPU::getNamedOperandIdx(Opc, AMDGPU::OpName::src1_modifiers); 1177 int Src0ModIdx = 1178 AMDGPU::getNamedOperandIdx(Opc, AMDGPU::OpName::src0_modifiers); 1179 if ((Src1ModIdx != -1 && MI.getOperand(Src1ModIdx).getImm() != 0) || 1180 (Src0ModIdx != -1 && MI.getOperand(Src0ModIdx).getImm() != 0)) 1181 return false; 1182 1183 LLVM_DEBUG(dbgs() << "Folded " << MI << " into "); 1184 auto &NewDesc = 1185 TII->get(Src0->isReg() ? (unsigned)AMDGPU::COPY : getMovOpc(false)); 1186 int Src2Idx = AMDGPU::getNamedOperandIdx(Opc, AMDGPU::OpName::src2); 1187 if (Src2Idx != -1) 1188 MI.RemoveOperand(Src2Idx); 1189 MI.RemoveOperand(AMDGPU::getNamedOperandIdx(Opc, AMDGPU::OpName::src1)); 1190 if (Src1ModIdx != -1) 1191 MI.RemoveOperand(Src1ModIdx); 1192 if (Src0ModIdx != -1) 1193 MI.RemoveOperand(Src0ModIdx); 1194 mutateCopyOp(MI, NewDesc); 1195 LLVM_DEBUG(dbgs() << MI); 1196 return true; 1197 } 1198 1199 bool SIFoldOperands::tryFoldZeroHighBits(MachineInstr &MI) const { 1200 if (MI.getOpcode() != AMDGPU::V_AND_B32_e64 && 1201 MI.getOpcode() != AMDGPU::V_AND_B32_e32) 1202 return false; 1203 1204 MachineOperand *Src0 = getImmOrMaterializedImm(*MRI, MI.getOperand(1)); 1205 if (!Src0->isImm() || Src0->getImm() != 0xffff) 1206 return false; 1207 1208 Register Src1 = MI.getOperand(2).getReg(); 1209 MachineInstr *SrcDef = MRI->getVRegDef(Src1); 1210 if (ST->zeroesHigh16BitsOfDest(SrcDef->getOpcode())) { 1211 Register Dst = MI.getOperand(0).getReg(); 1212 MRI->replaceRegWith(Dst, SrcDef->getOperand(0).getReg()); 1213 MI.eraseFromParent(); 1214 return true; 1215 } 1216 1217 return false; 1218 } 1219 1220 void SIFoldOperands::foldInstOperand(MachineInstr &MI, 1221 MachineOperand &OpToFold) const { 1222 // We need mutate the operands of new mov instructions to add implicit 1223 // uses of EXEC, but adding them invalidates the use_iterator, so defer 1224 // this. 1225 SmallVector<MachineInstr *, 4> CopiesToReplace; 1226 SmallVector<FoldCandidate, 4> FoldList; 1227 MachineOperand &Dst = MI.getOperand(0); 1228 1229 if (OpToFold.isImm()) { 1230 for (auto &UseMI : 1231 make_early_inc_range(MRI->use_nodbg_instructions(Dst.getReg()))) { 1232 // Folding the immediate may reveal operations that can be constant 1233 // folded or replaced with a copy. This can happen for example after 1234 // frame indices are lowered to constants or from splitting 64-bit 1235 // constants. 1236 // 1237 // We may also encounter cases where one or both operands are 1238 // immediates materialized into a register, which would ordinarily not 1239 // be folded due to multiple uses or operand constraints. 1240 if (tryConstantFoldOp(*MRI, TII, &UseMI)) 1241 LLVM_DEBUG(dbgs() << "Constant folded " << UseMI); 1242 } 1243 } 1244 1245 bool FoldingImm = OpToFold.isImm() || OpToFold.isFI() || OpToFold.isGlobal(); 1246 if (FoldingImm) { 1247 unsigned NumLiteralUses = 0; 1248 MachineOperand *NonInlineUse = nullptr; 1249 int NonInlineUseOpNo = -1; 1250 1251 for (auto &Use : 1252 make_early_inc_range(MRI->use_nodbg_operands(Dst.getReg()))) { 1253 MachineInstr *UseMI = Use.getParent(); 1254 unsigned OpNo = UseMI->getOperandNo(&Use); 1255 1256 // Try to fold any inline immediate uses, and then only fold other 1257 // constants if they have one use. 1258 // 1259 // The legality of the inline immediate must be checked based on the use 1260 // operand, not the defining instruction, because 32-bit instructions 1261 // with 32-bit inline immediate sources may be used to materialize 1262 // constants used in 16-bit operands. 1263 // 1264 // e.g. it is unsafe to fold: 1265 // s_mov_b32 s0, 1.0 // materializes 0x3f800000 1266 // v_add_f16 v0, v1, s0 // 1.0 f16 inline immediate sees 0x00003c00 1267 1268 // Folding immediates with more than one use will increase program size. 1269 // FIXME: This will also reduce register usage, which may be better 1270 // in some cases. A better heuristic is needed. 1271 if (isInlineConstantIfFolded(TII, *UseMI, OpNo, OpToFold)) { 1272 foldOperand(OpToFold, UseMI, OpNo, FoldList, CopiesToReplace); 1273 } else if (frameIndexMayFold(TII, *UseMI, OpNo, OpToFold)) { 1274 foldOperand(OpToFold, UseMI, OpNo, FoldList, CopiesToReplace); 1275 } else { 1276 if (++NumLiteralUses == 1) { 1277 NonInlineUse = &Use; 1278 NonInlineUseOpNo = OpNo; 1279 } 1280 } 1281 } 1282 1283 if (NumLiteralUses == 1) { 1284 MachineInstr *UseMI = NonInlineUse->getParent(); 1285 foldOperand(OpToFold, UseMI, NonInlineUseOpNo, FoldList, CopiesToReplace); 1286 } 1287 } else { 1288 // Folding register. 1289 SmallVector <MachineOperand *, 4> UsesToProcess; 1290 for (auto &Use : MRI->use_nodbg_operands(Dst.getReg())) 1291 UsesToProcess.push_back(&Use); 1292 for (auto U : UsesToProcess) { 1293 MachineInstr *UseMI = U->getParent(); 1294 1295 foldOperand(OpToFold, UseMI, UseMI->getOperandNo(U), 1296 FoldList, CopiesToReplace); 1297 } 1298 } 1299 1300 MachineFunction *MF = MI.getParent()->getParent(); 1301 // Make sure we add EXEC uses to any new v_mov instructions created. 1302 for (MachineInstr *Copy : CopiesToReplace) 1303 Copy->addImplicitDefUseOperands(*MF); 1304 1305 for (FoldCandidate &Fold : FoldList) { 1306 assert(!Fold.isReg() || Fold.OpToFold); 1307 if (Fold.isReg() && Fold.OpToFold->getReg().isVirtual()) { 1308 Register Reg = Fold.OpToFold->getReg(); 1309 MachineInstr *DefMI = Fold.OpToFold->getParent(); 1310 if (DefMI->readsRegister(AMDGPU::EXEC, TRI) && 1311 execMayBeModifiedBeforeUse(*MRI, Reg, *DefMI, *Fold.UseMI)) 1312 continue; 1313 } 1314 if (updateOperand(Fold, *TII, *TRI, *ST)) { 1315 // Clear kill flags. 1316 if (Fold.isReg()) { 1317 assert(Fold.OpToFold && Fold.OpToFold->isReg()); 1318 // FIXME: Probably shouldn't bother trying to fold if not an 1319 // SGPR. PeepholeOptimizer can eliminate redundant VGPR->VGPR 1320 // copies. 1321 MRI->clearKillFlags(Fold.OpToFold->getReg()); 1322 } 1323 LLVM_DEBUG(dbgs() << "Folded source from " << MI << " into OpNo " 1324 << static_cast<int>(Fold.UseOpNo) << " of " 1325 << *Fold.UseMI); 1326 } else if (Fold.isCommuted()) { 1327 // Restoring instruction's original operand order if fold has failed. 1328 TII->commuteInstruction(*Fold.UseMI, false); 1329 } 1330 } 1331 } 1332 1333 // Clamp patterns are canonically selected to v_max_* instructions, so only 1334 // handle them. 1335 const MachineOperand *SIFoldOperands::isClamp(const MachineInstr &MI) const { 1336 unsigned Op = MI.getOpcode(); 1337 switch (Op) { 1338 case AMDGPU::V_MAX_F32_e64: 1339 case AMDGPU::V_MAX_F16_e64: 1340 case AMDGPU::V_MAX_F64_e64: 1341 case AMDGPU::V_PK_MAX_F16: { 1342 if (!TII->getNamedOperand(MI, AMDGPU::OpName::clamp)->getImm()) 1343 return nullptr; 1344 1345 // Make sure sources are identical. 1346 const MachineOperand *Src0 = TII->getNamedOperand(MI, AMDGPU::OpName::src0); 1347 const MachineOperand *Src1 = TII->getNamedOperand(MI, AMDGPU::OpName::src1); 1348 if (!Src0->isReg() || !Src1->isReg() || 1349 Src0->getReg() != Src1->getReg() || 1350 Src0->getSubReg() != Src1->getSubReg() || 1351 Src0->getSubReg() != AMDGPU::NoSubRegister) 1352 return nullptr; 1353 1354 // Can't fold up if we have modifiers. 1355 if (TII->hasModifiersSet(MI, AMDGPU::OpName::omod)) 1356 return nullptr; 1357 1358 unsigned Src0Mods 1359 = TII->getNamedOperand(MI, AMDGPU::OpName::src0_modifiers)->getImm(); 1360 unsigned Src1Mods 1361 = TII->getNamedOperand(MI, AMDGPU::OpName::src1_modifiers)->getImm(); 1362 1363 // Having a 0 op_sel_hi would require swizzling the output in the source 1364 // instruction, which we can't do. 1365 unsigned UnsetMods = (Op == AMDGPU::V_PK_MAX_F16) ? SISrcMods::OP_SEL_1 1366 : 0u; 1367 if (Src0Mods != UnsetMods && Src1Mods != UnsetMods) 1368 return nullptr; 1369 return Src0; 1370 } 1371 default: 1372 return nullptr; 1373 } 1374 } 1375 1376 // FIXME: Clamp for v_mad_mixhi_f16 handled during isel. 1377 bool SIFoldOperands::tryFoldClamp(MachineInstr &MI) { 1378 const MachineOperand *ClampSrc = isClamp(MI); 1379 if (!ClampSrc || !MRI->hasOneNonDBGUser(ClampSrc->getReg())) 1380 return false; 1381 1382 MachineInstr *Def = MRI->getVRegDef(ClampSrc->getReg()); 1383 1384 // The type of clamp must be compatible. 1385 if (TII->getClampMask(*Def) != TII->getClampMask(MI)) 1386 return false; 1387 1388 MachineOperand *DefClamp = TII->getNamedOperand(*Def, AMDGPU::OpName::clamp); 1389 if (!DefClamp) 1390 return false; 1391 1392 LLVM_DEBUG(dbgs() << "Folding clamp " << *DefClamp << " into " << *Def); 1393 1394 // Clamp is applied after omod, so it is OK if omod is set. 1395 DefClamp->setImm(1); 1396 MRI->replaceRegWith(MI.getOperand(0).getReg(), Def->getOperand(0).getReg()); 1397 MI.eraseFromParent(); 1398 1399 // Use of output modifiers forces VOP3 encoding for a VOP2 mac/fmac 1400 // instruction, so we might as well convert it to the more flexible VOP3-only 1401 // mad/fma form. 1402 if (TII->convertToThreeAddress(*Def, nullptr, nullptr)) 1403 Def->eraseFromParent(); 1404 1405 return true; 1406 } 1407 1408 static int getOModValue(unsigned Opc, int64_t Val) { 1409 switch (Opc) { 1410 case AMDGPU::V_MUL_F64_e64: { 1411 switch (Val) { 1412 case 0x3fe0000000000000: // 0.5 1413 return SIOutMods::DIV2; 1414 case 0x4000000000000000: // 2.0 1415 return SIOutMods::MUL2; 1416 case 0x4010000000000000: // 4.0 1417 return SIOutMods::MUL4; 1418 default: 1419 return SIOutMods::NONE; 1420 } 1421 } 1422 case AMDGPU::V_MUL_F32_e64: { 1423 switch (static_cast<uint32_t>(Val)) { 1424 case 0x3f000000: // 0.5 1425 return SIOutMods::DIV2; 1426 case 0x40000000: // 2.0 1427 return SIOutMods::MUL2; 1428 case 0x40800000: // 4.0 1429 return SIOutMods::MUL4; 1430 default: 1431 return SIOutMods::NONE; 1432 } 1433 } 1434 case AMDGPU::V_MUL_F16_e64: { 1435 switch (static_cast<uint16_t>(Val)) { 1436 case 0x3800: // 0.5 1437 return SIOutMods::DIV2; 1438 case 0x4000: // 2.0 1439 return SIOutMods::MUL2; 1440 case 0x4400: // 4.0 1441 return SIOutMods::MUL4; 1442 default: 1443 return SIOutMods::NONE; 1444 } 1445 } 1446 default: 1447 llvm_unreachable("invalid mul opcode"); 1448 } 1449 } 1450 1451 // FIXME: Does this really not support denormals with f16? 1452 // FIXME: Does this need to check IEEE mode bit? SNaNs are generally not 1453 // handled, so will anything other than that break? 1454 std::pair<const MachineOperand *, int> 1455 SIFoldOperands::isOMod(const MachineInstr &MI) const { 1456 unsigned Op = MI.getOpcode(); 1457 switch (Op) { 1458 case AMDGPU::V_MUL_F64_e64: 1459 case AMDGPU::V_MUL_F32_e64: 1460 case AMDGPU::V_MUL_F16_e64: { 1461 // If output denormals are enabled, omod is ignored. 1462 if ((Op == AMDGPU::V_MUL_F32_e64 && MFI->getMode().FP32OutputDenormals) || 1463 ((Op == AMDGPU::V_MUL_F64_e64 || Op == AMDGPU::V_MUL_F16_e64) && 1464 MFI->getMode().FP64FP16OutputDenormals)) 1465 return std::make_pair(nullptr, SIOutMods::NONE); 1466 1467 const MachineOperand *RegOp = nullptr; 1468 const MachineOperand *ImmOp = nullptr; 1469 const MachineOperand *Src0 = TII->getNamedOperand(MI, AMDGPU::OpName::src0); 1470 const MachineOperand *Src1 = TII->getNamedOperand(MI, AMDGPU::OpName::src1); 1471 if (Src0->isImm()) { 1472 ImmOp = Src0; 1473 RegOp = Src1; 1474 } else if (Src1->isImm()) { 1475 ImmOp = Src1; 1476 RegOp = Src0; 1477 } else 1478 return std::make_pair(nullptr, SIOutMods::NONE); 1479 1480 int OMod = getOModValue(Op, ImmOp->getImm()); 1481 if (OMod == SIOutMods::NONE || 1482 TII->hasModifiersSet(MI, AMDGPU::OpName::src0_modifiers) || 1483 TII->hasModifiersSet(MI, AMDGPU::OpName::src1_modifiers) || 1484 TII->hasModifiersSet(MI, AMDGPU::OpName::omod) || 1485 TII->hasModifiersSet(MI, AMDGPU::OpName::clamp)) 1486 return std::make_pair(nullptr, SIOutMods::NONE); 1487 1488 return std::make_pair(RegOp, OMod); 1489 } 1490 case AMDGPU::V_ADD_F64_e64: 1491 case AMDGPU::V_ADD_F32_e64: 1492 case AMDGPU::V_ADD_F16_e64: { 1493 // If output denormals are enabled, omod is ignored. 1494 if ((Op == AMDGPU::V_ADD_F32_e64 && MFI->getMode().FP32OutputDenormals) || 1495 ((Op == AMDGPU::V_ADD_F64_e64 || Op == AMDGPU::V_ADD_F16_e64) && 1496 MFI->getMode().FP64FP16OutputDenormals)) 1497 return std::make_pair(nullptr, SIOutMods::NONE); 1498 1499 // Look through the DAGCombiner canonicalization fmul x, 2 -> fadd x, x 1500 const MachineOperand *Src0 = TII->getNamedOperand(MI, AMDGPU::OpName::src0); 1501 const MachineOperand *Src1 = TII->getNamedOperand(MI, AMDGPU::OpName::src1); 1502 1503 if (Src0->isReg() && Src1->isReg() && Src0->getReg() == Src1->getReg() && 1504 Src0->getSubReg() == Src1->getSubReg() && 1505 !TII->hasModifiersSet(MI, AMDGPU::OpName::src0_modifiers) && 1506 !TII->hasModifiersSet(MI, AMDGPU::OpName::src1_modifiers) && 1507 !TII->hasModifiersSet(MI, AMDGPU::OpName::clamp) && 1508 !TII->hasModifiersSet(MI, AMDGPU::OpName::omod)) 1509 return std::make_pair(Src0, SIOutMods::MUL2); 1510 1511 return std::make_pair(nullptr, SIOutMods::NONE); 1512 } 1513 default: 1514 return std::make_pair(nullptr, SIOutMods::NONE); 1515 } 1516 } 1517 1518 // FIXME: Does this need to check IEEE bit on function? 1519 bool SIFoldOperands::tryFoldOMod(MachineInstr &MI) { 1520 const MachineOperand *RegOp; 1521 int OMod; 1522 std::tie(RegOp, OMod) = isOMod(MI); 1523 if (OMod == SIOutMods::NONE || !RegOp->isReg() || 1524 RegOp->getSubReg() != AMDGPU::NoSubRegister || 1525 !MRI->hasOneNonDBGUser(RegOp->getReg())) 1526 return false; 1527 1528 MachineInstr *Def = MRI->getVRegDef(RegOp->getReg()); 1529 MachineOperand *DefOMod = TII->getNamedOperand(*Def, AMDGPU::OpName::omod); 1530 if (!DefOMod || DefOMod->getImm() != SIOutMods::NONE) 1531 return false; 1532 1533 // Clamp is applied after omod. If the source already has clamp set, don't 1534 // fold it. 1535 if (TII->hasModifiersSet(*Def, AMDGPU::OpName::clamp)) 1536 return false; 1537 1538 LLVM_DEBUG(dbgs() << "Folding omod " << MI << " into " << *Def); 1539 1540 DefOMod->setImm(OMod); 1541 MRI->replaceRegWith(MI.getOperand(0).getReg(), Def->getOperand(0).getReg()); 1542 MI.eraseFromParent(); 1543 1544 // Use of output modifiers forces VOP3 encoding for a VOP2 mac/fmac 1545 // instruction, so we might as well convert it to the more flexible VOP3-only 1546 // mad/fma form. 1547 if (TII->convertToThreeAddress(*Def, nullptr, nullptr)) 1548 Def->eraseFromParent(); 1549 1550 return true; 1551 } 1552 1553 // Try to fold a reg_sequence with vgpr output and agpr inputs into an 1554 // instruction which can take an agpr. So far that means a store. 1555 bool SIFoldOperands::tryFoldRegSequence(MachineInstr &MI) { 1556 assert(MI.isRegSequence()); 1557 auto Reg = MI.getOperand(0).getReg(); 1558 1559 if (!ST->hasGFX90AInsts() || !TRI->isVGPR(*MRI, Reg) || 1560 !MRI->hasOneNonDBGUse(Reg)) 1561 return false; 1562 1563 SmallVector<std::pair<MachineOperand*, unsigned>, 32> Defs; 1564 if (!getRegSeqInit(Defs, Reg, MCOI::OPERAND_REGISTER, TII, *MRI)) 1565 return false; 1566 1567 for (auto &Def : Defs) { 1568 const auto *Op = Def.first; 1569 if (!Op->isReg()) 1570 return false; 1571 if (TRI->isAGPR(*MRI, Op->getReg())) 1572 continue; 1573 // Maybe this is a COPY from AREG 1574 const MachineInstr *SubDef = MRI->getVRegDef(Op->getReg()); 1575 if (!SubDef || !SubDef->isCopy() || SubDef->getOperand(1).getSubReg()) 1576 return false; 1577 if (!TRI->isAGPR(*MRI, SubDef->getOperand(1).getReg())) 1578 return false; 1579 } 1580 1581 MachineOperand *Op = &*MRI->use_nodbg_begin(Reg); 1582 MachineInstr *UseMI = Op->getParent(); 1583 while (UseMI->isCopy() && !Op->getSubReg()) { 1584 Reg = UseMI->getOperand(0).getReg(); 1585 if (!TRI->isVGPR(*MRI, Reg) || !MRI->hasOneNonDBGUse(Reg)) 1586 return false; 1587 Op = &*MRI->use_nodbg_begin(Reg); 1588 UseMI = Op->getParent(); 1589 } 1590 1591 if (Op->getSubReg()) 1592 return false; 1593 1594 unsigned OpIdx = Op - &UseMI->getOperand(0); 1595 const MCInstrDesc &InstDesc = UseMI->getDesc(); 1596 if (!TRI->isVectorSuperClass( 1597 TRI->getRegClass(InstDesc.OpInfo[OpIdx].RegClass))) 1598 return false; 1599 1600 const auto *NewDstRC = TRI->getEquivalentAGPRClass(MRI->getRegClass(Reg)); 1601 auto Dst = MRI->createVirtualRegister(NewDstRC); 1602 auto RS = BuildMI(*MI.getParent(), MI, MI.getDebugLoc(), 1603 TII->get(AMDGPU::REG_SEQUENCE), Dst); 1604 1605 for (unsigned I = 0; I < Defs.size(); ++I) { 1606 MachineOperand *Def = Defs[I].first; 1607 Def->setIsKill(false); 1608 if (TRI->isAGPR(*MRI, Def->getReg())) { 1609 RS.add(*Def); 1610 } else { // This is a copy 1611 MachineInstr *SubDef = MRI->getVRegDef(Def->getReg()); 1612 SubDef->getOperand(1).setIsKill(false); 1613 RS.addReg(SubDef->getOperand(1).getReg(), 0, Def->getSubReg()); 1614 } 1615 RS.addImm(Defs[I].second); 1616 } 1617 1618 Op->setReg(Dst); 1619 if (!TII->isOperandLegal(*UseMI, OpIdx, Op)) { 1620 Op->setReg(Reg); 1621 RS->eraseFromParent(); 1622 return false; 1623 } 1624 1625 LLVM_DEBUG(dbgs() << "Folded " << *RS << " into " << *UseMI); 1626 1627 // Erase the REG_SEQUENCE eagerly, unless we followed a chain of COPY users, 1628 // in which case we can erase them all later in runOnMachineFunction. 1629 if (MRI->use_nodbg_empty(MI.getOperand(0).getReg())) 1630 MI.eraseFromParent(); 1631 return true; 1632 } 1633 1634 // Try to hoist an AGPR to VGPR copy out of the loop across a LCSSA PHI. 1635 // This should allow folding of an AGPR into a consumer which may support it. 1636 // I.e.: 1637 // 1638 // loop: // loop: 1639 // %1:vreg = COPY %0:areg // exit: 1640 // exit: => // %1:areg = PHI %0:areg, %loop 1641 // %2:vreg = PHI %1:vreg, %loop // %2:vreg = COPY %1:areg 1642 bool SIFoldOperands::tryFoldLCSSAPhi(MachineInstr &PHI) { 1643 assert(PHI.isPHI()); 1644 1645 if (PHI.getNumExplicitOperands() != 3) // Single input LCSSA PHI 1646 return false; 1647 1648 Register PhiIn = PHI.getOperand(1).getReg(); 1649 Register PhiOut = PHI.getOperand(0).getReg(); 1650 if (PHI.getOperand(1).getSubReg() || 1651 !TRI->isVGPR(*MRI, PhiIn) || !TRI->isVGPR(*MRI, PhiOut)) 1652 return false; 1653 1654 // A single use should not matter for correctness, but if it has another use 1655 // inside the loop we may perform copy twice in a worst case. 1656 if (!MRI->hasOneNonDBGUse(PhiIn)) 1657 return false; 1658 1659 MachineInstr *Copy = MRI->getVRegDef(PhiIn); 1660 if (!Copy || !Copy->isCopy()) 1661 return false; 1662 1663 Register CopyIn = Copy->getOperand(1).getReg(); 1664 if (!TRI->isAGPR(*MRI, CopyIn) || Copy->getOperand(1).getSubReg()) 1665 return false; 1666 1667 const TargetRegisterClass *ARC = MRI->getRegClass(CopyIn); 1668 Register NewReg = MRI->createVirtualRegister(ARC); 1669 PHI.getOperand(1).setReg(CopyIn); 1670 PHI.getOperand(0).setReg(NewReg); 1671 1672 MachineBasicBlock *MBB = PHI.getParent(); 1673 BuildMI(*MBB, MBB->getFirstNonPHI(), Copy->getDebugLoc(), 1674 TII->get(AMDGPU::COPY), PhiOut) 1675 .addReg(NewReg, RegState::Kill); 1676 Copy->eraseFromParent(); // We know this copy had a single use. 1677 1678 LLVM_DEBUG(dbgs() << "Folded " << PHI); 1679 1680 return true; 1681 } 1682 1683 // Attempt to convert VGPR load to an AGPR load. 1684 bool SIFoldOperands::tryFoldLoad(MachineInstr &MI) { 1685 assert(MI.mayLoad()); 1686 if (!ST->hasGFX90AInsts() || MI.getNumExplicitDefs() != 1) 1687 return false; 1688 1689 MachineOperand &Def = MI.getOperand(0); 1690 if (!Def.isDef()) 1691 return false; 1692 1693 Register DefReg = Def.getReg(); 1694 1695 if (DefReg.isPhysical() || !TRI->isVGPR(*MRI, DefReg)) 1696 return false; 1697 1698 SmallVector<const MachineInstr*, 8> Users; 1699 SmallVector<Register, 8> MoveRegs; 1700 for (const MachineInstr &I : MRI->use_nodbg_instructions(DefReg)) { 1701 Users.push_back(&I); 1702 } 1703 if (Users.empty()) 1704 return false; 1705 1706 // Check that all uses a copy to an agpr or a reg_sequence producing an agpr. 1707 while (!Users.empty()) { 1708 const MachineInstr *I = Users.pop_back_val(); 1709 if (!I->isCopy() && !I->isRegSequence()) 1710 return false; 1711 Register DstReg = I->getOperand(0).getReg(); 1712 if (TRI->isAGPR(*MRI, DstReg)) 1713 continue; 1714 MoveRegs.push_back(DstReg); 1715 for (const MachineInstr &U : MRI->use_nodbg_instructions(DstReg)) { 1716 Users.push_back(&U); 1717 } 1718 } 1719 1720 const TargetRegisterClass *RC = MRI->getRegClass(DefReg); 1721 MRI->setRegClass(DefReg, TRI->getEquivalentAGPRClass(RC)); 1722 if (!TII->isOperandLegal(MI, 0, &Def)) { 1723 MRI->setRegClass(DefReg, RC); 1724 return false; 1725 } 1726 1727 while (!MoveRegs.empty()) { 1728 Register Reg = MoveRegs.pop_back_val(); 1729 MRI->setRegClass(Reg, TRI->getEquivalentAGPRClass(MRI->getRegClass(Reg))); 1730 } 1731 1732 LLVM_DEBUG(dbgs() << "Folded " << MI); 1733 1734 return true; 1735 } 1736 1737 bool SIFoldOperands::runOnMachineFunction(MachineFunction &MF) { 1738 if (skipFunction(MF.getFunction())) 1739 return false; 1740 1741 MRI = &MF.getRegInfo(); 1742 ST = &MF.getSubtarget<GCNSubtarget>(); 1743 TII = ST->getInstrInfo(); 1744 TRI = &TII->getRegisterInfo(); 1745 MFI = MF.getInfo<SIMachineFunctionInfo>(); 1746 1747 // omod is ignored by hardware if IEEE bit is enabled. omod also does not 1748 // correctly handle signed zeros. 1749 // 1750 // FIXME: Also need to check strictfp 1751 bool IsIEEEMode = MFI->getMode().IEEE; 1752 bool HasNSZ = MFI->hasNoSignedZerosFPMath(); 1753 1754 for (MachineBasicBlock *MBB : depth_first(&MF)) { 1755 MachineOperand *CurrentKnownM0Val = nullptr; 1756 for (auto &MI : make_early_inc_range(*MBB)) { 1757 tryFoldCndMask(MI); 1758 1759 if (tryFoldZeroHighBits(MI)) 1760 continue; 1761 1762 if (MI.isRegSequence() && tryFoldRegSequence(MI)) 1763 continue; 1764 1765 if (MI.isPHI() && tryFoldLCSSAPhi(MI)) 1766 continue; 1767 1768 if (MI.mayLoad() && tryFoldLoad(MI)) 1769 continue; 1770 1771 if (!TII->isFoldableCopy(MI)) { 1772 // Saw an unknown clobber of m0, so we no longer know what it is. 1773 if (CurrentKnownM0Val && MI.modifiesRegister(AMDGPU::M0, TRI)) 1774 CurrentKnownM0Val = nullptr; 1775 1776 // TODO: Omod might be OK if there is NSZ only on the source 1777 // instruction, and not the omod multiply. 1778 if (IsIEEEMode || (!HasNSZ && !MI.getFlag(MachineInstr::FmNsz)) || 1779 !tryFoldOMod(MI)) 1780 tryFoldClamp(MI); 1781 1782 continue; 1783 } 1784 1785 // Specially track simple redefs of m0 to the same value in a block, so we 1786 // can erase the later ones. 1787 if (MI.getOperand(0).getReg() == AMDGPU::M0) { 1788 MachineOperand &NewM0Val = MI.getOperand(1); 1789 if (CurrentKnownM0Val && CurrentKnownM0Val->isIdenticalTo(NewM0Val)) { 1790 MI.eraseFromParent(); 1791 continue; 1792 } 1793 1794 // We aren't tracking other physical registers 1795 CurrentKnownM0Val = (NewM0Val.isReg() && NewM0Val.getReg().isPhysical()) ? 1796 nullptr : &NewM0Val; 1797 continue; 1798 } 1799 1800 MachineOperand &OpToFold = MI.getOperand(1); 1801 bool FoldingImm = 1802 OpToFold.isImm() || OpToFold.isFI() || OpToFold.isGlobal(); 1803 1804 // FIXME: We could also be folding things like TargetIndexes. 1805 if (!FoldingImm && !OpToFold.isReg()) 1806 continue; 1807 1808 if (OpToFold.isReg() && !OpToFold.getReg().isVirtual()) 1809 continue; 1810 1811 // Prevent folding operands backwards in the function. For example, 1812 // the COPY opcode must not be replaced by 1 in this example: 1813 // 1814 // %3 = COPY %vgpr0; VGPR_32:%3 1815 // ... 1816 // %vgpr0 = V_MOV_B32_e32 1, implicit %exec 1817 if (!MI.getOperand(0).getReg().isVirtual()) 1818 continue; 1819 1820 foldInstOperand(MI, OpToFold); 1821 1822 // If we managed to fold all uses of this copy then we might as well 1823 // delete it now. 1824 // The only reason we need to follow chains of copies here is that 1825 // tryFoldRegSequence looks forward through copies before folding a 1826 // REG_SEQUENCE into its eventual users. 1827 auto *InstToErase = &MI; 1828 while (MRI->use_nodbg_empty(InstToErase->getOperand(0).getReg())) { 1829 auto &SrcOp = InstToErase->getOperand(1); 1830 auto SrcReg = SrcOp.isReg() ? SrcOp.getReg() : Register(); 1831 InstToErase->eraseFromParent(); 1832 InstToErase = nullptr; 1833 if (!SrcReg || SrcReg.isPhysical()) 1834 break; 1835 InstToErase = MRI->getVRegDef(SrcReg); 1836 if (!InstToErase || !TII->isFoldableCopy(*InstToErase)) 1837 break; 1838 } 1839 if (InstToErase && InstToErase->isRegSequence() && 1840 MRI->use_nodbg_empty(InstToErase->getOperand(0).getReg())) 1841 InstToErase->eraseFromParent(); 1842 } 1843 } 1844 return true; 1845 } 1846