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 Old.ChangeToImmediate(Fold.ImmToFold); 304 return true; 305 } 306 307 if (Fold.isGlobal()) { 308 Old.ChangeToGA(Fold.OpToFold->getGlobal(), Fold.OpToFold->getOffset(), 309 Fold.OpToFold->getTargetFlags()); 310 return true; 311 } 312 313 if (Fold.isFI()) { 314 Old.ChangeToFrameIndex(Fold.FrameIndexToFold); 315 return true; 316 } 317 318 MachineOperand *New = Fold.OpToFold; 319 Old.substVirtReg(New->getReg(), New->getSubReg(), TRI); 320 Old.setIsUndef(New->isUndef()); 321 return true; 322 } 323 324 static bool isUseMIInFoldList(ArrayRef<FoldCandidate> FoldList, 325 const MachineInstr *MI) { 326 for (auto Candidate : FoldList) { 327 if (Candidate.UseMI == MI) 328 return true; 329 } 330 return false; 331 } 332 333 static void appendFoldCandidate(SmallVectorImpl<FoldCandidate> &FoldList, 334 MachineInstr *MI, unsigned OpNo, 335 MachineOperand *FoldOp, bool Commuted = false, 336 int ShrinkOp = -1) { 337 // Skip additional folding on the same operand. 338 for (FoldCandidate &Fold : FoldList) 339 if (Fold.UseMI == MI && Fold.UseOpNo == OpNo) 340 return; 341 LLVM_DEBUG(dbgs() << "Append " << (Commuted ? "commuted" : "normal") 342 << " operand " << OpNo << "\n " << *MI); 343 FoldList.emplace_back(MI, OpNo, FoldOp, Commuted, ShrinkOp); 344 } 345 346 static bool tryAddToFoldList(SmallVectorImpl<FoldCandidate> &FoldList, 347 MachineInstr *MI, unsigned OpNo, 348 MachineOperand *OpToFold, 349 const SIInstrInfo *TII) { 350 if (!TII->isOperandLegal(*MI, OpNo, OpToFold)) { 351 // Special case for v_mac_{f16, f32}_e64 if we are trying to fold into src2 352 unsigned Opc = MI->getOpcode(); 353 unsigned NewOpc = macToMad(Opc); 354 if (NewOpc != AMDGPU::INSTRUCTION_LIST_END) { 355 // Check if changing this to a v_mad_{f16, f32} instruction will allow us 356 // to fold the operand. 357 MI->setDesc(TII->get(NewOpc)); 358 bool FoldAsMAD = tryAddToFoldList(FoldList, MI, OpNo, OpToFold, TII); 359 if (FoldAsMAD) { 360 MI->untieRegOperand(OpNo); 361 return true; 362 } 363 MI->setDesc(TII->get(Opc)); 364 } 365 366 // Special case for s_setreg_b32 367 if (OpToFold->isImm()) { 368 unsigned ImmOpc = 0; 369 if (Opc == AMDGPU::S_SETREG_B32) 370 ImmOpc = AMDGPU::S_SETREG_IMM32_B32; 371 else if (Opc == AMDGPU::S_SETREG_B32_mode) 372 ImmOpc = AMDGPU::S_SETREG_IMM32_B32_mode; 373 if (ImmOpc) { 374 MI->setDesc(TII->get(ImmOpc)); 375 appendFoldCandidate(FoldList, MI, OpNo, OpToFold); 376 return true; 377 } 378 } 379 380 // If we are already folding into another operand of MI, then 381 // we can't commute the instruction, otherwise we risk making the 382 // other fold illegal. 383 if (isUseMIInFoldList(FoldList, MI)) 384 return false; 385 386 unsigned CommuteOpNo = OpNo; 387 388 // Operand is not legal, so try to commute the instruction to 389 // see if this makes it possible to fold. 390 unsigned CommuteIdx0 = TargetInstrInfo::CommuteAnyOperandIndex; 391 unsigned CommuteIdx1 = TargetInstrInfo::CommuteAnyOperandIndex; 392 bool CanCommute = TII->findCommutedOpIndices(*MI, CommuteIdx0, CommuteIdx1); 393 394 if (CanCommute) { 395 if (CommuteIdx0 == OpNo) 396 CommuteOpNo = CommuteIdx1; 397 else if (CommuteIdx1 == OpNo) 398 CommuteOpNo = CommuteIdx0; 399 } 400 401 402 // One of operands might be an Imm operand, and OpNo may refer to it after 403 // the call of commuteInstruction() below. Such situations are avoided 404 // here explicitly as OpNo must be a register operand to be a candidate 405 // for memory folding. 406 if (CanCommute && (!MI->getOperand(CommuteIdx0).isReg() || 407 !MI->getOperand(CommuteIdx1).isReg())) 408 return false; 409 410 if (!CanCommute || 411 !TII->commuteInstruction(*MI, false, CommuteIdx0, CommuteIdx1)) 412 return false; 413 414 if (!TII->isOperandLegal(*MI, CommuteOpNo, OpToFold)) { 415 if ((Opc == AMDGPU::V_ADD_CO_U32_e64 || 416 Opc == AMDGPU::V_SUB_CO_U32_e64 || 417 Opc == AMDGPU::V_SUBREV_CO_U32_e64) && // FIXME 418 (OpToFold->isImm() || OpToFold->isFI() || OpToFold->isGlobal())) { 419 MachineRegisterInfo &MRI = MI->getParent()->getParent()->getRegInfo(); 420 421 // Verify the other operand is a VGPR, otherwise we would violate the 422 // constant bus restriction. 423 unsigned OtherIdx = CommuteOpNo == CommuteIdx0 ? CommuteIdx1 : CommuteIdx0; 424 MachineOperand &OtherOp = MI->getOperand(OtherIdx); 425 if (!OtherOp.isReg() || 426 !TII->getRegisterInfo().isVGPR(MRI, OtherOp.getReg())) 427 return false; 428 429 assert(MI->getOperand(1).isDef()); 430 431 // Make sure to get the 32-bit version of the commuted opcode. 432 unsigned MaybeCommutedOpc = MI->getOpcode(); 433 int Op32 = AMDGPU::getVOPe32(MaybeCommutedOpc); 434 435 appendFoldCandidate(FoldList, MI, CommuteOpNo, OpToFold, true, Op32); 436 return true; 437 } 438 439 TII->commuteInstruction(*MI, false, CommuteIdx0, CommuteIdx1); 440 return false; 441 } 442 443 appendFoldCandidate(FoldList, MI, CommuteOpNo, OpToFold, true); 444 return true; 445 } 446 447 // Check the case where we might introduce a second constant operand to a 448 // scalar instruction 449 if (TII->isSALU(MI->getOpcode())) { 450 const MCInstrDesc &InstDesc = MI->getDesc(); 451 const MCOperandInfo &OpInfo = InstDesc.OpInfo[OpNo]; 452 const SIRegisterInfo &SRI = TII->getRegisterInfo(); 453 454 // Fine if the operand can be encoded as an inline constant 455 if (TII->isLiteralConstantLike(*OpToFold, OpInfo)) { 456 if (!SRI.opCanUseInlineConstant(OpInfo.OperandType) || 457 !TII->isInlineConstant(*OpToFold, OpInfo)) { 458 // Otherwise check for another constant 459 for (unsigned i = 0, e = InstDesc.getNumOperands(); i != e; ++i) { 460 auto &Op = MI->getOperand(i); 461 if (OpNo != i && 462 TII->isLiteralConstantLike(Op, OpInfo)) { 463 return false; 464 } 465 } 466 } 467 } 468 } 469 470 appendFoldCandidate(FoldList, MI, OpNo, OpToFold); 471 return true; 472 } 473 474 // If the use operand doesn't care about the value, this may be an operand only 475 // used for register indexing, in which case it is unsafe to fold. 476 static bool isUseSafeToFold(const SIInstrInfo *TII, 477 const MachineInstr &MI, 478 const MachineOperand &UseMO) { 479 if (UseMO.isUndef() || TII->isSDWA(MI)) 480 return false; 481 482 switch (MI.getOpcode()) { 483 case AMDGPU::V_MOV_B32_e32: 484 case AMDGPU::V_MOV_B32_e64: 485 case AMDGPU::V_MOV_B64_PSEUDO: 486 // Do not fold into an indirect mov. 487 return !MI.hasRegisterImplicitUseOperand(AMDGPU::M0); 488 } 489 490 return true; 491 //return !MI.hasRegisterImplicitUseOperand(UseMO.getReg()); 492 } 493 494 // Find a def of the UseReg, check if it is a reg_sequence and find initializers 495 // for each subreg, tracking it to foldable inline immediate if possible. 496 // Returns true on success. 497 static bool getRegSeqInit( 498 SmallVectorImpl<std::pair<MachineOperand*, unsigned>> &Defs, 499 Register UseReg, uint8_t OpTy, 500 const SIInstrInfo *TII, const MachineRegisterInfo &MRI) { 501 MachineInstr *Def = MRI.getVRegDef(UseReg); 502 if (!Def || !Def->isRegSequence()) 503 return false; 504 505 for (unsigned I = 1, E = Def->getNumExplicitOperands(); I < E; I += 2) { 506 MachineOperand *Sub = &Def->getOperand(I); 507 assert(Sub->isReg()); 508 509 for (MachineInstr *SubDef = MRI.getVRegDef(Sub->getReg()); 510 SubDef && Sub->isReg() && Sub->getReg().isVirtual() && 511 !Sub->getSubReg() && TII->isFoldableCopy(*SubDef); 512 SubDef = MRI.getVRegDef(Sub->getReg())) { 513 MachineOperand *Op = &SubDef->getOperand(1); 514 if (Op->isImm()) { 515 if (TII->isInlineConstant(*Op, OpTy)) 516 Sub = Op; 517 break; 518 } 519 if (!Op->isReg() || Op->getReg().isPhysical()) 520 break; 521 Sub = Op; 522 } 523 524 Defs.emplace_back(Sub, Def->getOperand(I + 1).getImm()); 525 } 526 527 return true; 528 } 529 530 static bool tryToFoldACImm(const SIInstrInfo *TII, 531 const MachineOperand &OpToFold, 532 MachineInstr *UseMI, 533 unsigned UseOpIdx, 534 SmallVectorImpl<FoldCandidate> &FoldList) { 535 const MCInstrDesc &Desc = UseMI->getDesc(); 536 const MCOperandInfo *OpInfo = Desc.OpInfo; 537 if (!OpInfo || UseOpIdx >= Desc.getNumOperands()) 538 return false; 539 540 uint8_t OpTy = OpInfo[UseOpIdx].OperandType; 541 if ((OpTy < AMDGPU::OPERAND_REG_INLINE_AC_FIRST || 542 OpTy > AMDGPU::OPERAND_REG_INLINE_AC_LAST) && 543 (OpTy < AMDGPU::OPERAND_REG_INLINE_C_FIRST || 544 OpTy > AMDGPU::OPERAND_REG_INLINE_C_LAST)) 545 return false; 546 547 if (OpToFold.isImm() && TII->isInlineConstant(OpToFold, OpTy) && 548 TII->isOperandLegal(*UseMI, UseOpIdx, &OpToFold)) { 549 UseMI->getOperand(UseOpIdx).ChangeToImmediate(OpToFold.getImm()); 550 return true; 551 } 552 553 if (!OpToFold.isReg()) 554 return false; 555 556 Register UseReg = OpToFold.getReg(); 557 if (!UseReg.isVirtual()) 558 return false; 559 560 if (isUseMIInFoldList(FoldList, UseMI)) 561 return false; 562 563 MachineRegisterInfo &MRI = UseMI->getParent()->getParent()->getRegInfo(); 564 565 // Maybe it is just a COPY of an immediate itself. 566 MachineInstr *Def = MRI.getVRegDef(UseReg); 567 MachineOperand &UseOp = UseMI->getOperand(UseOpIdx); 568 if (!UseOp.getSubReg() && Def && TII->isFoldableCopy(*Def)) { 569 MachineOperand &DefOp = Def->getOperand(1); 570 if (DefOp.isImm() && TII->isInlineConstant(DefOp, OpTy) && 571 TII->isOperandLegal(*UseMI, UseOpIdx, &DefOp)) { 572 UseMI->getOperand(UseOpIdx).ChangeToImmediate(DefOp.getImm()); 573 return true; 574 } 575 } 576 577 SmallVector<std::pair<MachineOperand*, unsigned>, 32> Defs; 578 if (!getRegSeqInit(Defs, UseReg, OpTy, TII, MRI)) 579 return false; 580 581 int32_t Imm; 582 for (unsigned I = 0, E = Defs.size(); I != E; ++I) { 583 const MachineOperand *Op = Defs[I].first; 584 if (!Op->isImm()) 585 return false; 586 587 auto SubImm = Op->getImm(); 588 if (!I) { 589 Imm = SubImm; 590 if (!TII->isInlineConstant(*Op, OpTy) || 591 !TII->isOperandLegal(*UseMI, UseOpIdx, Op)) 592 return false; 593 594 continue; 595 } 596 if (Imm != SubImm) 597 return false; // Can only fold splat constants 598 } 599 600 appendFoldCandidate(FoldList, UseMI, UseOpIdx, Defs[0].first); 601 return true; 602 } 603 604 void SIFoldOperands::foldOperand( 605 MachineOperand &OpToFold, 606 MachineInstr *UseMI, 607 int UseOpIdx, 608 SmallVectorImpl<FoldCandidate> &FoldList, 609 SmallVectorImpl<MachineInstr *> &CopiesToReplace) const { 610 const MachineOperand &UseOp = UseMI->getOperand(UseOpIdx); 611 612 if (!isUseSafeToFold(TII, *UseMI, UseOp)) 613 return; 614 615 // FIXME: Fold operands with subregs. 616 if (UseOp.isReg() && OpToFold.isReg()) { 617 if (UseOp.isImplicit() || UseOp.getSubReg() != AMDGPU::NoSubRegister) 618 return; 619 } 620 621 // Special case for REG_SEQUENCE: We can't fold literals into 622 // REG_SEQUENCE instructions, so we have to fold them into the 623 // uses of REG_SEQUENCE. 624 if (UseMI->isRegSequence()) { 625 Register RegSeqDstReg = UseMI->getOperand(0).getReg(); 626 unsigned RegSeqDstSubReg = UseMI->getOperand(UseOpIdx + 1).getImm(); 627 628 for (auto &RSUse : make_early_inc_range(MRI->use_nodbg_operands(RegSeqDstReg))) { 629 MachineInstr *RSUseMI = RSUse.getParent(); 630 631 if (tryToFoldACImm(TII, UseMI->getOperand(0), RSUseMI, 632 RSUseMI->getOperandNo(&RSUse), FoldList)) 633 continue; 634 635 if (RSUse.getSubReg() != RegSeqDstSubReg) 636 continue; 637 638 foldOperand(OpToFold, RSUseMI, RSUseMI->getOperandNo(&RSUse), FoldList, 639 CopiesToReplace); 640 } 641 642 return; 643 } 644 645 if (tryToFoldACImm(TII, OpToFold, UseMI, UseOpIdx, FoldList)) 646 return; 647 648 if (frameIndexMayFold(TII, *UseMI, UseOpIdx, OpToFold)) { 649 // Verify that this is a stack access. 650 // FIXME: Should probably use stack pseudos before frame lowering. 651 652 if (TII->isMUBUF(*UseMI)) { 653 if (TII->getNamedOperand(*UseMI, AMDGPU::OpName::srsrc)->getReg() != 654 MFI->getScratchRSrcReg()) 655 return; 656 657 // Ensure this is either relative to the current frame or the current 658 // wave. 659 MachineOperand &SOff = 660 *TII->getNamedOperand(*UseMI, AMDGPU::OpName::soffset); 661 if (!SOff.isImm() || SOff.getImm() != 0) 662 return; 663 } 664 665 // A frame index will resolve to a positive constant, so it should always be 666 // safe to fold the addressing mode, even pre-GFX9. 667 UseMI->getOperand(UseOpIdx).ChangeToFrameIndex(OpToFold.getIndex()); 668 669 if (TII->isFLATScratch(*UseMI) && 670 AMDGPU::getNamedOperandIdx(UseMI->getOpcode(), 671 AMDGPU::OpName::vaddr) != -1) { 672 unsigned NewOpc = AMDGPU::getFlatScratchInstSSfromSV(UseMI->getOpcode()); 673 UseMI->setDesc(TII->get(NewOpc)); 674 } 675 676 return; 677 } 678 679 bool FoldingImmLike = 680 OpToFold.isImm() || OpToFold.isFI() || OpToFold.isGlobal(); 681 682 if (FoldingImmLike && UseMI->isCopy()) { 683 Register DestReg = UseMI->getOperand(0).getReg(); 684 Register SrcReg = UseMI->getOperand(1).getReg(); 685 assert(SrcReg.isVirtual()); 686 687 const TargetRegisterClass *SrcRC = MRI->getRegClass(SrcReg); 688 689 // Don't fold into a copy to a physical register with the same class. Doing 690 // so would interfere with the register coalescer's logic which would avoid 691 // redundant initializations. 692 if (DestReg.isPhysical() && SrcRC->contains(DestReg)) 693 return; 694 695 const TargetRegisterClass *DestRC = TRI->getRegClassForReg(*MRI, DestReg); 696 if (!DestReg.isPhysical()) { 697 if (TRI->isSGPRClass(SrcRC) && TRI->hasVectorRegisters(DestRC)) { 698 SmallVector<FoldCandidate, 4> CopyUses; 699 for (auto &Use : MRI->use_nodbg_operands(DestReg)) { 700 // There's no point trying to fold into an implicit operand. 701 if (Use.isImplicit()) 702 continue; 703 704 CopyUses.emplace_back(Use.getParent(), 705 Use.getParent()->getOperandNo(&Use), 706 &UseMI->getOperand(1)); 707 } 708 for (auto &F : CopyUses) { 709 foldOperand(*F.OpToFold, F.UseMI, F.UseOpNo, FoldList, CopiesToReplace); 710 } 711 } 712 713 if (DestRC == &AMDGPU::AGPR_32RegClass && 714 TII->isInlineConstant(OpToFold, AMDGPU::OPERAND_REG_INLINE_C_INT32)) { 715 UseMI->setDesc(TII->get(AMDGPU::V_ACCVGPR_WRITE_B32_e64)); 716 UseMI->getOperand(1).ChangeToImmediate(OpToFold.getImm()); 717 CopiesToReplace.push_back(UseMI); 718 return; 719 } 720 } 721 722 // In order to fold immediates into copies, we need to change the 723 // copy to a MOV. 724 725 unsigned MovOp = TII->getMovOpcode(DestRC); 726 if (MovOp == AMDGPU::COPY) 727 return; 728 729 UseMI->setDesc(TII->get(MovOp)); 730 MachineInstr::mop_iterator ImpOpI = UseMI->implicit_operands().begin(); 731 MachineInstr::mop_iterator ImpOpE = UseMI->implicit_operands().end(); 732 while (ImpOpI != ImpOpE) { 733 MachineInstr::mop_iterator Tmp = ImpOpI; 734 ImpOpI++; 735 UseMI->RemoveOperand(UseMI->getOperandNo(Tmp)); 736 } 737 CopiesToReplace.push_back(UseMI); 738 } else { 739 if (UseMI->isCopy() && OpToFold.isReg() && 740 UseMI->getOperand(0).getReg().isVirtual() && 741 !UseMI->getOperand(1).getSubReg()) { 742 LLVM_DEBUG(dbgs() << "Folding " << OpToFold << "\n into " << *UseMI); 743 unsigned Size = TII->getOpSize(*UseMI, 1); 744 Register UseReg = OpToFold.getReg(); 745 UseMI->getOperand(1).setReg(UseReg); 746 UseMI->getOperand(1).setSubReg(OpToFold.getSubReg()); 747 UseMI->getOperand(1).setIsKill(false); 748 CopiesToReplace.push_back(UseMI); 749 OpToFold.setIsKill(false); 750 751 // That is very tricky to store a value into an AGPR. v_accvgpr_write_b32 752 // can only accept VGPR or inline immediate. Recreate a reg_sequence with 753 // its initializers right here, so we will rematerialize immediates and 754 // avoid copies via different reg classes. 755 SmallVector<std::pair<MachineOperand*, unsigned>, 32> Defs; 756 if (Size > 4 && TRI->isAGPR(*MRI, UseMI->getOperand(0).getReg()) && 757 getRegSeqInit(Defs, UseReg, AMDGPU::OPERAND_REG_INLINE_C_INT32, TII, 758 *MRI)) { 759 const DebugLoc &DL = UseMI->getDebugLoc(); 760 MachineBasicBlock &MBB = *UseMI->getParent(); 761 762 UseMI->setDesc(TII->get(AMDGPU::REG_SEQUENCE)); 763 for (unsigned I = UseMI->getNumOperands() - 1; I > 0; --I) 764 UseMI->RemoveOperand(I); 765 766 MachineInstrBuilder B(*MBB.getParent(), UseMI); 767 DenseMap<TargetInstrInfo::RegSubRegPair, Register> VGPRCopies; 768 SmallSetVector<TargetInstrInfo::RegSubRegPair, 32> SeenAGPRs; 769 for (unsigned I = 0; I < Size / 4; ++I) { 770 MachineOperand *Def = Defs[I].first; 771 TargetInstrInfo::RegSubRegPair CopyToVGPR; 772 if (Def->isImm() && 773 TII->isInlineConstant(*Def, AMDGPU::OPERAND_REG_INLINE_C_INT32)) { 774 int64_t Imm = Def->getImm(); 775 776 auto Tmp = MRI->createVirtualRegister(&AMDGPU::AGPR_32RegClass); 777 BuildMI(MBB, UseMI, DL, 778 TII->get(AMDGPU::V_ACCVGPR_WRITE_B32_e64), Tmp).addImm(Imm); 779 B.addReg(Tmp); 780 } else if (Def->isReg() && TRI->isAGPR(*MRI, Def->getReg())) { 781 auto Src = getRegSubRegPair(*Def); 782 Def->setIsKill(false); 783 if (!SeenAGPRs.insert(Src)) { 784 // We cannot build a reg_sequence out of the same registers, they 785 // must be copied. Better do it here before copyPhysReg() created 786 // several reads to do the AGPR->VGPR->AGPR copy. 787 CopyToVGPR = Src; 788 } else { 789 B.addReg(Src.Reg, Def->isUndef() ? RegState::Undef : 0, 790 Src.SubReg); 791 } 792 } else { 793 assert(Def->isReg()); 794 Def->setIsKill(false); 795 auto Src = getRegSubRegPair(*Def); 796 797 // Direct copy from SGPR to AGPR is not possible. To avoid creation 798 // of exploded copies SGPR->VGPR->AGPR in the copyPhysReg() later, 799 // create a copy here and track if we already have such a copy. 800 if (TRI->isSGPRReg(*MRI, Src.Reg)) { 801 CopyToVGPR = Src; 802 } else { 803 auto Tmp = MRI->createVirtualRegister(&AMDGPU::AGPR_32RegClass); 804 BuildMI(MBB, UseMI, DL, TII->get(AMDGPU::COPY), Tmp).add(*Def); 805 B.addReg(Tmp); 806 } 807 } 808 809 if (CopyToVGPR.Reg) { 810 Register Vgpr; 811 if (VGPRCopies.count(CopyToVGPR)) { 812 Vgpr = VGPRCopies[CopyToVGPR]; 813 } else { 814 Vgpr = MRI->createVirtualRegister(&AMDGPU::VGPR_32RegClass); 815 BuildMI(MBB, UseMI, DL, TII->get(AMDGPU::COPY), Vgpr).add(*Def); 816 VGPRCopies[CopyToVGPR] = Vgpr; 817 } 818 auto Tmp = MRI->createVirtualRegister(&AMDGPU::AGPR_32RegClass); 819 BuildMI(MBB, UseMI, DL, 820 TII->get(AMDGPU::V_ACCVGPR_WRITE_B32_e64), Tmp).addReg(Vgpr); 821 B.addReg(Tmp); 822 } 823 824 B.addImm(Defs[I].second); 825 } 826 LLVM_DEBUG(dbgs() << "Folded " << *UseMI); 827 return; 828 } 829 830 if (Size != 4) 831 return; 832 if (TRI->isAGPR(*MRI, UseMI->getOperand(0).getReg()) && 833 TRI->isVGPR(*MRI, UseMI->getOperand(1).getReg())) 834 UseMI->setDesc(TII->get(AMDGPU::V_ACCVGPR_WRITE_B32_e64)); 835 else if (TRI->isVGPR(*MRI, UseMI->getOperand(0).getReg()) && 836 TRI->isAGPR(*MRI, UseMI->getOperand(1).getReg())) 837 UseMI->setDesc(TII->get(AMDGPU::V_ACCVGPR_READ_B32_e64)); 838 else if (ST->hasGFX90AInsts() && 839 TRI->isAGPR(*MRI, UseMI->getOperand(0).getReg()) && 840 TRI->isAGPR(*MRI, UseMI->getOperand(1).getReg())) 841 UseMI->setDesc(TII->get(AMDGPU::V_ACCVGPR_MOV_B32)); 842 return; 843 } 844 845 unsigned UseOpc = UseMI->getOpcode(); 846 if (UseOpc == AMDGPU::V_READFIRSTLANE_B32 || 847 (UseOpc == AMDGPU::V_READLANE_B32 && 848 (int)UseOpIdx == 849 AMDGPU::getNamedOperandIdx(UseOpc, AMDGPU::OpName::src0))) { 850 // %vgpr = V_MOV_B32 imm 851 // %sgpr = V_READFIRSTLANE_B32 %vgpr 852 // => 853 // %sgpr = S_MOV_B32 imm 854 if (FoldingImmLike) { 855 if (execMayBeModifiedBeforeUse(*MRI, 856 UseMI->getOperand(UseOpIdx).getReg(), 857 *OpToFold.getParent(), 858 *UseMI)) 859 return; 860 861 UseMI->setDesc(TII->get(AMDGPU::S_MOV_B32)); 862 863 if (OpToFold.isImm()) 864 UseMI->getOperand(1).ChangeToImmediate(OpToFold.getImm()); 865 else 866 UseMI->getOperand(1).ChangeToFrameIndex(OpToFold.getIndex()); 867 UseMI->RemoveOperand(2); // Remove exec read (or src1 for readlane) 868 return; 869 } 870 871 if (OpToFold.isReg() && TRI->isSGPRReg(*MRI, OpToFold.getReg())) { 872 if (execMayBeModifiedBeforeUse(*MRI, 873 UseMI->getOperand(UseOpIdx).getReg(), 874 *OpToFold.getParent(), 875 *UseMI)) 876 return; 877 878 // %vgpr = COPY %sgpr0 879 // %sgpr1 = V_READFIRSTLANE_B32 %vgpr 880 // => 881 // %sgpr1 = COPY %sgpr0 882 UseMI->setDesc(TII->get(AMDGPU::COPY)); 883 UseMI->getOperand(1).setReg(OpToFold.getReg()); 884 UseMI->getOperand(1).setSubReg(OpToFold.getSubReg()); 885 UseMI->getOperand(1).setIsKill(false); 886 UseMI->RemoveOperand(2); // Remove exec read (or src1 for readlane) 887 return; 888 } 889 } 890 891 const MCInstrDesc &UseDesc = UseMI->getDesc(); 892 893 // Don't fold into target independent nodes. Target independent opcodes 894 // don't have defined register classes. 895 if (UseDesc.isVariadic() || 896 UseOp.isImplicit() || 897 UseDesc.OpInfo[UseOpIdx].RegClass == -1) 898 return; 899 } 900 901 if (!FoldingImmLike) { 902 tryAddToFoldList(FoldList, UseMI, UseOpIdx, &OpToFold, TII); 903 904 // FIXME: We could try to change the instruction from 64-bit to 32-bit 905 // to enable more folding opportunities. The shrink operands pass 906 // already does this. 907 return; 908 } 909 910 911 const MCInstrDesc &FoldDesc = OpToFold.getParent()->getDesc(); 912 const TargetRegisterClass *FoldRC = 913 TRI->getRegClass(FoldDesc.OpInfo[0].RegClass); 914 915 // Split 64-bit constants into 32-bits for folding. 916 if (UseOp.getSubReg() && AMDGPU::getRegBitWidth(FoldRC->getID()) == 64) { 917 Register UseReg = UseOp.getReg(); 918 const TargetRegisterClass *UseRC = MRI->getRegClass(UseReg); 919 920 if (AMDGPU::getRegBitWidth(UseRC->getID()) != 64) 921 return; 922 923 APInt Imm(64, OpToFold.getImm()); 924 if (UseOp.getSubReg() == AMDGPU::sub0) { 925 Imm = Imm.getLoBits(32); 926 } else { 927 assert(UseOp.getSubReg() == AMDGPU::sub1); 928 Imm = Imm.getHiBits(32); 929 } 930 931 MachineOperand ImmOp = MachineOperand::CreateImm(Imm.getSExtValue()); 932 tryAddToFoldList(FoldList, UseMI, UseOpIdx, &ImmOp, TII); 933 return; 934 } 935 936 937 938 tryAddToFoldList(FoldList, UseMI, UseOpIdx, &OpToFold, TII); 939 } 940 941 static bool evalBinaryInstruction(unsigned Opcode, int32_t &Result, 942 uint32_t LHS, uint32_t RHS) { 943 switch (Opcode) { 944 case AMDGPU::V_AND_B32_e64: 945 case AMDGPU::V_AND_B32_e32: 946 case AMDGPU::S_AND_B32: 947 Result = LHS & RHS; 948 return true; 949 case AMDGPU::V_OR_B32_e64: 950 case AMDGPU::V_OR_B32_e32: 951 case AMDGPU::S_OR_B32: 952 Result = LHS | RHS; 953 return true; 954 case AMDGPU::V_XOR_B32_e64: 955 case AMDGPU::V_XOR_B32_e32: 956 case AMDGPU::S_XOR_B32: 957 Result = LHS ^ RHS; 958 return true; 959 case AMDGPU::S_XNOR_B32: 960 Result = ~(LHS ^ RHS); 961 return true; 962 case AMDGPU::S_NAND_B32: 963 Result = ~(LHS & RHS); 964 return true; 965 case AMDGPU::S_NOR_B32: 966 Result = ~(LHS | RHS); 967 return true; 968 case AMDGPU::S_ANDN2_B32: 969 Result = LHS & ~RHS; 970 return true; 971 case AMDGPU::S_ORN2_B32: 972 Result = LHS | ~RHS; 973 return true; 974 case AMDGPU::V_LSHL_B32_e64: 975 case AMDGPU::V_LSHL_B32_e32: 976 case AMDGPU::S_LSHL_B32: 977 // The instruction ignores the high bits for out of bounds shifts. 978 Result = LHS << (RHS & 31); 979 return true; 980 case AMDGPU::V_LSHLREV_B32_e64: 981 case AMDGPU::V_LSHLREV_B32_e32: 982 Result = RHS << (LHS & 31); 983 return true; 984 case AMDGPU::V_LSHR_B32_e64: 985 case AMDGPU::V_LSHR_B32_e32: 986 case AMDGPU::S_LSHR_B32: 987 Result = LHS >> (RHS & 31); 988 return true; 989 case AMDGPU::V_LSHRREV_B32_e64: 990 case AMDGPU::V_LSHRREV_B32_e32: 991 Result = RHS >> (LHS & 31); 992 return true; 993 case AMDGPU::V_ASHR_I32_e64: 994 case AMDGPU::V_ASHR_I32_e32: 995 case AMDGPU::S_ASHR_I32: 996 Result = static_cast<int32_t>(LHS) >> (RHS & 31); 997 return true; 998 case AMDGPU::V_ASHRREV_I32_e64: 999 case AMDGPU::V_ASHRREV_I32_e32: 1000 Result = static_cast<int32_t>(RHS) >> (LHS & 31); 1001 return true; 1002 default: 1003 return false; 1004 } 1005 } 1006 1007 static unsigned getMovOpc(bool IsScalar) { 1008 return IsScalar ? AMDGPU::S_MOV_B32 : AMDGPU::V_MOV_B32_e32; 1009 } 1010 1011 /// Remove any leftover implicit operands from mutating the instruction. e.g. 1012 /// if we replace an s_and_b32 with a copy, we don't need the implicit scc def 1013 /// anymore. 1014 static void stripExtraCopyOperands(MachineInstr &MI) { 1015 const MCInstrDesc &Desc = MI.getDesc(); 1016 unsigned NumOps = Desc.getNumOperands() + 1017 Desc.getNumImplicitUses() + 1018 Desc.getNumImplicitDefs(); 1019 1020 for (unsigned I = MI.getNumOperands() - 1; I >= NumOps; --I) 1021 MI.RemoveOperand(I); 1022 } 1023 1024 static void mutateCopyOp(MachineInstr &MI, const MCInstrDesc &NewDesc) { 1025 MI.setDesc(NewDesc); 1026 stripExtraCopyOperands(MI); 1027 } 1028 1029 static MachineOperand *getImmOrMaterializedImm(MachineRegisterInfo &MRI, 1030 MachineOperand &Op) { 1031 if (Op.isReg()) { 1032 // If this has a subregister, it obviously is a register source. 1033 if (Op.getSubReg() != AMDGPU::NoSubRegister || !Op.getReg().isVirtual()) 1034 return &Op; 1035 1036 MachineInstr *Def = MRI.getVRegDef(Op.getReg()); 1037 if (Def && Def->isMoveImmediate()) { 1038 MachineOperand &ImmSrc = Def->getOperand(1); 1039 if (ImmSrc.isImm()) 1040 return &ImmSrc; 1041 } 1042 } 1043 1044 return &Op; 1045 } 1046 1047 // Try to simplify operations with a constant that may appear after instruction 1048 // selection. 1049 // TODO: See if a frame index with a fixed offset can fold. 1050 static bool tryConstantFoldOp(MachineRegisterInfo &MRI, const SIInstrInfo *TII, 1051 MachineInstr *MI) { 1052 unsigned Opc = MI->getOpcode(); 1053 1054 int Src0Idx = AMDGPU::getNamedOperandIdx(Opc, AMDGPU::OpName::src0); 1055 if (Src0Idx == -1) 1056 return false; 1057 MachineOperand *Src0 = getImmOrMaterializedImm(MRI, MI->getOperand(Src0Idx)); 1058 1059 if ((Opc == AMDGPU::V_NOT_B32_e64 || Opc == AMDGPU::V_NOT_B32_e32 || 1060 Opc == AMDGPU::S_NOT_B32) && 1061 Src0->isImm()) { 1062 MI->getOperand(1).ChangeToImmediate(~Src0->getImm()); 1063 mutateCopyOp(*MI, TII->get(getMovOpc(Opc == AMDGPU::S_NOT_B32))); 1064 return true; 1065 } 1066 1067 int Src1Idx = AMDGPU::getNamedOperandIdx(Opc, AMDGPU::OpName::src1); 1068 if (Src1Idx == -1) 1069 return false; 1070 MachineOperand *Src1 = getImmOrMaterializedImm(MRI, MI->getOperand(Src1Idx)); 1071 1072 if (!Src0->isImm() && !Src1->isImm()) 1073 return false; 1074 1075 // and k0, k1 -> v_mov_b32 (k0 & k1) 1076 // or k0, k1 -> v_mov_b32 (k0 | k1) 1077 // xor k0, k1 -> v_mov_b32 (k0 ^ k1) 1078 if (Src0->isImm() && Src1->isImm()) { 1079 int32_t NewImm; 1080 if (!evalBinaryInstruction(Opc, NewImm, Src0->getImm(), Src1->getImm())) 1081 return false; 1082 1083 const SIRegisterInfo &TRI = TII->getRegisterInfo(); 1084 bool IsSGPR = TRI.isSGPRReg(MRI, MI->getOperand(0).getReg()); 1085 1086 // Be careful to change the right operand, src0 may belong to a different 1087 // instruction. 1088 MI->getOperand(Src0Idx).ChangeToImmediate(NewImm); 1089 MI->RemoveOperand(Src1Idx); 1090 mutateCopyOp(*MI, TII->get(getMovOpc(IsSGPR))); 1091 return true; 1092 } 1093 1094 if (!MI->isCommutable()) 1095 return false; 1096 1097 if (Src0->isImm() && !Src1->isImm()) { 1098 std::swap(Src0, Src1); 1099 std::swap(Src0Idx, Src1Idx); 1100 } 1101 1102 int32_t Src1Val = static_cast<int32_t>(Src1->getImm()); 1103 if (Opc == AMDGPU::V_OR_B32_e64 || 1104 Opc == AMDGPU::V_OR_B32_e32 || 1105 Opc == AMDGPU::S_OR_B32) { 1106 if (Src1Val == 0) { 1107 // y = or x, 0 => y = copy x 1108 MI->RemoveOperand(Src1Idx); 1109 mutateCopyOp(*MI, TII->get(AMDGPU::COPY)); 1110 } else if (Src1Val == -1) { 1111 // y = or x, -1 => y = v_mov_b32 -1 1112 MI->RemoveOperand(Src1Idx); 1113 mutateCopyOp(*MI, TII->get(getMovOpc(Opc == AMDGPU::S_OR_B32))); 1114 } else 1115 return false; 1116 1117 return true; 1118 } 1119 1120 if (MI->getOpcode() == AMDGPU::V_AND_B32_e64 || 1121 MI->getOpcode() == AMDGPU::V_AND_B32_e32 || 1122 MI->getOpcode() == AMDGPU::S_AND_B32) { 1123 if (Src1Val == 0) { 1124 // y = and x, 0 => y = v_mov_b32 0 1125 MI->RemoveOperand(Src0Idx); 1126 mutateCopyOp(*MI, TII->get(getMovOpc(Opc == AMDGPU::S_AND_B32))); 1127 } else if (Src1Val == -1) { 1128 // y = and x, -1 => y = copy x 1129 MI->RemoveOperand(Src1Idx); 1130 mutateCopyOp(*MI, TII->get(AMDGPU::COPY)); 1131 stripExtraCopyOperands(*MI); 1132 } else 1133 return false; 1134 1135 return true; 1136 } 1137 1138 if (MI->getOpcode() == AMDGPU::V_XOR_B32_e64 || 1139 MI->getOpcode() == AMDGPU::V_XOR_B32_e32 || 1140 MI->getOpcode() == AMDGPU::S_XOR_B32) { 1141 if (Src1Val == 0) { 1142 // y = xor x, 0 => y = copy x 1143 MI->RemoveOperand(Src1Idx); 1144 mutateCopyOp(*MI, TII->get(AMDGPU::COPY)); 1145 return true; 1146 } 1147 } 1148 1149 return false; 1150 } 1151 1152 // Try to fold an instruction into a simpler one 1153 bool SIFoldOperands::tryFoldCndMask(MachineInstr &MI) const { 1154 unsigned Opc = MI.getOpcode(); 1155 if (Opc != AMDGPU::V_CNDMASK_B32_e32 && Opc != AMDGPU::V_CNDMASK_B32_e64 && 1156 Opc != AMDGPU::V_CNDMASK_B64_PSEUDO) 1157 return false; 1158 1159 MachineOperand *Src0 = TII->getNamedOperand(MI, AMDGPU::OpName::src0); 1160 MachineOperand *Src1 = TII->getNamedOperand(MI, AMDGPU::OpName::src1); 1161 if (!Src1->isIdenticalTo(*Src0)) { 1162 auto *Src0Imm = getImmOrMaterializedImm(*MRI, *Src0); 1163 auto *Src1Imm = getImmOrMaterializedImm(*MRI, *Src1); 1164 if (!Src1Imm->isIdenticalTo(*Src0Imm)) 1165 return false; 1166 } 1167 1168 int Src1ModIdx = 1169 AMDGPU::getNamedOperandIdx(Opc, AMDGPU::OpName::src1_modifiers); 1170 int Src0ModIdx = 1171 AMDGPU::getNamedOperandIdx(Opc, AMDGPU::OpName::src0_modifiers); 1172 if ((Src1ModIdx != -1 && MI.getOperand(Src1ModIdx).getImm() != 0) || 1173 (Src0ModIdx != -1 && MI.getOperand(Src0ModIdx).getImm() != 0)) 1174 return false; 1175 1176 LLVM_DEBUG(dbgs() << "Folded " << MI << " into "); 1177 auto &NewDesc = 1178 TII->get(Src0->isReg() ? (unsigned)AMDGPU::COPY : getMovOpc(false)); 1179 int Src2Idx = AMDGPU::getNamedOperandIdx(Opc, AMDGPU::OpName::src2); 1180 if (Src2Idx != -1) 1181 MI.RemoveOperand(Src2Idx); 1182 MI.RemoveOperand(AMDGPU::getNamedOperandIdx(Opc, AMDGPU::OpName::src1)); 1183 if (Src1ModIdx != -1) 1184 MI.RemoveOperand(Src1ModIdx); 1185 if (Src0ModIdx != -1) 1186 MI.RemoveOperand(Src0ModIdx); 1187 mutateCopyOp(MI, NewDesc); 1188 LLVM_DEBUG(dbgs() << MI); 1189 return true; 1190 } 1191 1192 bool SIFoldOperands::tryFoldZeroHighBits(MachineInstr &MI) const { 1193 if (MI.getOpcode() != AMDGPU::V_AND_B32_e64 && 1194 MI.getOpcode() != AMDGPU::V_AND_B32_e32) 1195 return false; 1196 1197 MachineOperand *Src0 = getImmOrMaterializedImm(*MRI, MI.getOperand(1)); 1198 if (!Src0->isImm() || Src0->getImm() != 0xffff) 1199 return false; 1200 1201 Register Src1 = MI.getOperand(2).getReg(); 1202 MachineInstr *SrcDef = MRI->getVRegDef(Src1); 1203 if (ST->zeroesHigh16BitsOfDest(SrcDef->getOpcode())) { 1204 Register Dst = MI.getOperand(0).getReg(); 1205 MRI->replaceRegWith(Dst, SrcDef->getOperand(0).getReg()); 1206 MI.eraseFromParent(); 1207 return true; 1208 } 1209 1210 return false; 1211 } 1212 1213 void SIFoldOperands::foldInstOperand(MachineInstr &MI, 1214 MachineOperand &OpToFold) const { 1215 // We need mutate the operands of new mov instructions to add implicit 1216 // uses of EXEC, but adding them invalidates the use_iterator, so defer 1217 // this. 1218 SmallVector<MachineInstr *, 4> CopiesToReplace; 1219 SmallVector<FoldCandidate, 4> FoldList; 1220 MachineOperand &Dst = MI.getOperand(0); 1221 1222 if (OpToFold.isImm()) { 1223 for (auto &UseMI : 1224 make_early_inc_range(MRI->use_nodbg_instructions(Dst.getReg()))) { 1225 // Folding the immediate may reveal operations that can be constant 1226 // folded or replaced with a copy. This can happen for example after 1227 // frame indices are lowered to constants or from splitting 64-bit 1228 // constants. 1229 // 1230 // We may also encounter cases where one or both operands are 1231 // immediates materialized into a register, which would ordinarily not 1232 // be folded due to multiple uses or operand constraints. 1233 if (tryConstantFoldOp(*MRI, TII, &UseMI)) 1234 LLVM_DEBUG(dbgs() << "Constant folded " << UseMI); 1235 } 1236 } 1237 1238 bool FoldingImm = OpToFold.isImm() || OpToFold.isFI() || OpToFold.isGlobal(); 1239 if (FoldingImm) { 1240 unsigned NumLiteralUses = 0; 1241 MachineOperand *NonInlineUse = nullptr; 1242 int NonInlineUseOpNo = -1; 1243 1244 for (auto &Use : 1245 make_early_inc_range(MRI->use_nodbg_operands(Dst.getReg()))) { 1246 MachineInstr *UseMI = Use.getParent(); 1247 unsigned OpNo = UseMI->getOperandNo(&Use); 1248 1249 // Try to fold any inline immediate uses, and then only fold other 1250 // constants if they have one use. 1251 // 1252 // The legality of the inline immediate must be checked based on the use 1253 // operand, not the defining instruction, because 32-bit instructions 1254 // with 32-bit inline immediate sources may be used to materialize 1255 // constants used in 16-bit operands. 1256 // 1257 // e.g. it is unsafe to fold: 1258 // s_mov_b32 s0, 1.0 // materializes 0x3f800000 1259 // v_add_f16 v0, v1, s0 // 1.0 f16 inline immediate sees 0x00003c00 1260 1261 // Folding immediates with more than one use will increase program size. 1262 // FIXME: This will also reduce register usage, which may be better 1263 // in some cases. A better heuristic is needed. 1264 if (isInlineConstantIfFolded(TII, *UseMI, OpNo, OpToFold)) { 1265 foldOperand(OpToFold, UseMI, OpNo, FoldList, CopiesToReplace); 1266 } else if (frameIndexMayFold(TII, *UseMI, OpNo, OpToFold)) { 1267 foldOperand(OpToFold, UseMI, OpNo, FoldList, CopiesToReplace); 1268 } else { 1269 if (++NumLiteralUses == 1) { 1270 NonInlineUse = &Use; 1271 NonInlineUseOpNo = OpNo; 1272 } 1273 } 1274 } 1275 1276 if (NumLiteralUses == 1) { 1277 MachineInstr *UseMI = NonInlineUse->getParent(); 1278 foldOperand(OpToFold, UseMI, NonInlineUseOpNo, FoldList, CopiesToReplace); 1279 } 1280 } else { 1281 // Folding register. 1282 SmallVector <MachineOperand *, 4> UsesToProcess; 1283 for (auto &Use : MRI->use_nodbg_operands(Dst.getReg())) 1284 UsesToProcess.push_back(&Use); 1285 for (auto U : UsesToProcess) { 1286 MachineInstr *UseMI = U->getParent(); 1287 1288 foldOperand(OpToFold, UseMI, UseMI->getOperandNo(U), 1289 FoldList, CopiesToReplace); 1290 } 1291 } 1292 1293 MachineFunction *MF = MI.getParent()->getParent(); 1294 // Make sure we add EXEC uses to any new v_mov instructions created. 1295 for (MachineInstr *Copy : CopiesToReplace) 1296 Copy->addImplicitDefUseOperands(*MF); 1297 1298 for (FoldCandidate &Fold : FoldList) { 1299 assert(!Fold.isReg() || Fold.OpToFold); 1300 if (Fold.isReg() && Fold.OpToFold->getReg().isVirtual()) { 1301 Register Reg = Fold.OpToFold->getReg(); 1302 MachineInstr *DefMI = Fold.OpToFold->getParent(); 1303 if (DefMI->readsRegister(AMDGPU::EXEC, TRI) && 1304 execMayBeModifiedBeforeUse(*MRI, Reg, *DefMI, *Fold.UseMI)) 1305 continue; 1306 } 1307 if (updateOperand(Fold, *TII, *TRI, *ST)) { 1308 // Clear kill flags. 1309 if (Fold.isReg()) { 1310 assert(Fold.OpToFold && Fold.OpToFold->isReg()); 1311 // FIXME: Probably shouldn't bother trying to fold if not an 1312 // SGPR. PeepholeOptimizer can eliminate redundant VGPR->VGPR 1313 // copies. 1314 MRI->clearKillFlags(Fold.OpToFold->getReg()); 1315 } 1316 LLVM_DEBUG(dbgs() << "Folded source from " << MI << " into OpNo " 1317 << static_cast<int>(Fold.UseOpNo) << " of " 1318 << *Fold.UseMI); 1319 } else if (Fold.isCommuted()) { 1320 // Restoring instruction's original operand order if fold has failed. 1321 TII->commuteInstruction(*Fold.UseMI, false); 1322 } 1323 } 1324 } 1325 1326 // Clamp patterns are canonically selected to v_max_* instructions, so only 1327 // handle them. 1328 const MachineOperand *SIFoldOperands::isClamp(const MachineInstr &MI) const { 1329 unsigned Op = MI.getOpcode(); 1330 switch (Op) { 1331 case AMDGPU::V_MAX_F32_e64: 1332 case AMDGPU::V_MAX_F16_e64: 1333 case AMDGPU::V_MAX_F64_e64: 1334 case AMDGPU::V_PK_MAX_F16: { 1335 if (!TII->getNamedOperand(MI, AMDGPU::OpName::clamp)->getImm()) 1336 return nullptr; 1337 1338 // Make sure sources are identical. 1339 const MachineOperand *Src0 = TII->getNamedOperand(MI, AMDGPU::OpName::src0); 1340 const MachineOperand *Src1 = TII->getNamedOperand(MI, AMDGPU::OpName::src1); 1341 if (!Src0->isReg() || !Src1->isReg() || 1342 Src0->getReg() != Src1->getReg() || 1343 Src0->getSubReg() != Src1->getSubReg() || 1344 Src0->getSubReg() != AMDGPU::NoSubRegister) 1345 return nullptr; 1346 1347 // Can't fold up if we have modifiers. 1348 if (TII->hasModifiersSet(MI, AMDGPU::OpName::omod)) 1349 return nullptr; 1350 1351 unsigned Src0Mods 1352 = TII->getNamedOperand(MI, AMDGPU::OpName::src0_modifiers)->getImm(); 1353 unsigned Src1Mods 1354 = TII->getNamedOperand(MI, AMDGPU::OpName::src1_modifiers)->getImm(); 1355 1356 // Having a 0 op_sel_hi would require swizzling the output in the source 1357 // instruction, which we can't do. 1358 unsigned UnsetMods = (Op == AMDGPU::V_PK_MAX_F16) ? SISrcMods::OP_SEL_1 1359 : 0u; 1360 if (Src0Mods != UnsetMods && Src1Mods != UnsetMods) 1361 return nullptr; 1362 return Src0; 1363 } 1364 default: 1365 return nullptr; 1366 } 1367 } 1368 1369 // FIXME: Clamp for v_mad_mixhi_f16 handled during isel. 1370 bool SIFoldOperands::tryFoldClamp(MachineInstr &MI) { 1371 const MachineOperand *ClampSrc = isClamp(MI); 1372 if (!ClampSrc || !MRI->hasOneNonDBGUser(ClampSrc->getReg())) 1373 return false; 1374 1375 MachineInstr *Def = MRI->getVRegDef(ClampSrc->getReg()); 1376 1377 // The type of clamp must be compatible. 1378 if (TII->getClampMask(*Def) != TII->getClampMask(MI)) 1379 return false; 1380 1381 MachineOperand *DefClamp = TII->getNamedOperand(*Def, AMDGPU::OpName::clamp); 1382 if (!DefClamp) 1383 return false; 1384 1385 LLVM_DEBUG(dbgs() << "Folding clamp " << *DefClamp << " into " << *Def); 1386 1387 // Clamp is applied after omod, so it is OK if omod is set. 1388 DefClamp->setImm(1); 1389 MRI->replaceRegWith(MI.getOperand(0).getReg(), Def->getOperand(0).getReg()); 1390 MI.eraseFromParent(); 1391 1392 // Use of output modifiers forces VOP3 encoding for a VOP2 mac/fmac 1393 // instruction, so we might as well convert it to the more flexible VOP3-only 1394 // mad/fma form. 1395 if (TII->convertToThreeAddress(*Def, nullptr, nullptr)) 1396 Def->eraseFromParent(); 1397 1398 return true; 1399 } 1400 1401 static int getOModValue(unsigned Opc, int64_t Val) { 1402 switch (Opc) { 1403 case AMDGPU::V_MUL_F64_e64: { 1404 switch (Val) { 1405 case 0x3fe0000000000000: // 0.5 1406 return SIOutMods::DIV2; 1407 case 0x4000000000000000: // 2.0 1408 return SIOutMods::MUL2; 1409 case 0x4010000000000000: // 4.0 1410 return SIOutMods::MUL4; 1411 default: 1412 return SIOutMods::NONE; 1413 } 1414 } 1415 case AMDGPU::V_MUL_F32_e64: { 1416 switch (static_cast<uint32_t>(Val)) { 1417 case 0x3f000000: // 0.5 1418 return SIOutMods::DIV2; 1419 case 0x40000000: // 2.0 1420 return SIOutMods::MUL2; 1421 case 0x40800000: // 4.0 1422 return SIOutMods::MUL4; 1423 default: 1424 return SIOutMods::NONE; 1425 } 1426 } 1427 case AMDGPU::V_MUL_F16_e64: { 1428 switch (static_cast<uint16_t>(Val)) { 1429 case 0x3800: // 0.5 1430 return SIOutMods::DIV2; 1431 case 0x4000: // 2.0 1432 return SIOutMods::MUL2; 1433 case 0x4400: // 4.0 1434 return SIOutMods::MUL4; 1435 default: 1436 return SIOutMods::NONE; 1437 } 1438 } 1439 default: 1440 llvm_unreachable("invalid mul opcode"); 1441 } 1442 } 1443 1444 // FIXME: Does this really not support denormals with f16? 1445 // FIXME: Does this need to check IEEE mode bit? SNaNs are generally not 1446 // handled, so will anything other than that break? 1447 std::pair<const MachineOperand *, int> 1448 SIFoldOperands::isOMod(const MachineInstr &MI) const { 1449 unsigned Op = MI.getOpcode(); 1450 switch (Op) { 1451 case AMDGPU::V_MUL_F64_e64: 1452 case AMDGPU::V_MUL_F32_e64: 1453 case AMDGPU::V_MUL_F16_e64: { 1454 // If output denormals are enabled, omod is ignored. 1455 if ((Op == AMDGPU::V_MUL_F32_e64 && MFI->getMode().FP32OutputDenormals) || 1456 ((Op == AMDGPU::V_MUL_F64_e64 || Op == AMDGPU::V_MUL_F16_e64) && 1457 MFI->getMode().FP64FP16OutputDenormals)) 1458 return std::make_pair(nullptr, SIOutMods::NONE); 1459 1460 const MachineOperand *RegOp = nullptr; 1461 const MachineOperand *ImmOp = nullptr; 1462 const MachineOperand *Src0 = TII->getNamedOperand(MI, AMDGPU::OpName::src0); 1463 const MachineOperand *Src1 = TII->getNamedOperand(MI, AMDGPU::OpName::src1); 1464 if (Src0->isImm()) { 1465 ImmOp = Src0; 1466 RegOp = Src1; 1467 } else if (Src1->isImm()) { 1468 ImmOp = Src1; 1469 RegOp = Src0; 1470 } else 1471 return std::make_pair(nullptr, SIOutMods::NONE); 1472 1473 int OMod = getOModValue(Op, ImmOp->getImm()); 1474 if (OMod == SIOutMods::NONE || 1475 TII->hasModifiersSet(MI, AMDGPU::OpName::src0_modifiers) || 1476 TII->hasModifiersSet(MI, AMDGPU::OpName::src1_modifiers) || 1477 TII->hasModifiersSet(MI, AMDGPU::OpName::omod) || 1478 TII->hasModifiersSet(MI, AMDGPU::OpName::clamp)) 1479 return std::make_pair(nullptr, SIOutMods::NONE); 1480 1481 return std::make_pair(RegOp, OMod); 1482 } 1483 case AMDGPU::V_ADD_F64_e64: 1484 case AMDGPU::V_ADD_F32_e64: 1485 case AMDGPU::V_ADD_F16_e64: { 1486 // If output denormals are enabled, omod is ignored. 1487 if ((Op == AMDGPU::V_ADD_F32_e64 && MFI->getMode().FP32OutputDenormals) || 1488 ((Op == AMDGPU::V_ADD_F64_e64 || Op == AMDGPU::V_ADD_F16_e64) && 1489 MFI->getMode().FP64FP16OutputDenormals)) 1490 return std::make_pair(nullptr, SIOutMods::NONE); 1491 1492 // Look through the DAGCombiner canonicalization fmul x, 2 -> fadd x, x 1493 const MachineOperand *Src0 = TII->getNamedOperand(MI, AMDGPU::OpName::src0); 1494 const MachineOperand *Src1 = TII->getNamedOperand(MI, AMDGPU::OpName::src1); 1495 1496 if (Src0->isReg() && Src1->isReg() && Src0->getReg() == Src1->getReg() && 1497 Src0->getSubReg() == Src1->getSubReg() && 1498 !TII->hasModifiersSet(MI, AMDGPU::OpName::src0_modifiers) && 1499 !TII->hasModifiersSet(MI, AMDGPU::OpName::src1_modifiers) && 1500 !TII->hasModifiersSet(MI, AMDGPU::OpName::clamp) && 1501 !TII->hasModifiersSet(MI, AMDGPU::OpName::omod)) 1502 return std::make_pair(Src0, SIOutMods::MUL2); 1503 1504 return std::make_pair(nullptr, SIOutMods::NONE); 1505 } 1506 default: 1507 return std::make_pair(nullptr, SIOutMods::NONE); 1508 } 1509 } 1510 1511 // FIXME: Does this need to check IEEE bit on function? 1512 bool SIFoldOperands::tryFoldOMod(MachineInstr &MI) { 1513 const MachineOperand *RegOp; 1514 int OMod; 1515 std::tie(RegOp, OMod) = isOMod(MI); 1516 if (OMod == SIOutMods::NONE || !RegOp->isReg() || 1517 RegOp->getSubReg() != AMDGPU::NoSubRegister || 1518 !MRI->hasOneNonDBGUser(RegOp->getReg())) 1519 return false; 1520 1521 MachineInstr *Def = MRI->getVRegDef(RegOp->getReg()); 1522 MachineOperand *DefOMod = TII->getNamedOperand(*Def, AMDGPU::OpName::omod); 1523 if (!DefOMod || DefOMod->getImm() != SIOutMods::NONE) 1524 return false; 1525 1526 // Clamp is applied after omod. If the source already has clamp set, don't 1527 // fold it. 1528 if (TII->hasModifiersSet(*Def, AMDGPU::OpName::clamp)) 1529 return false; 1530 1531 LLVM_DEBUG(dbgs() << "Folding omod " << MI << " into " << *Def); 1532 1533 DefOMod->setImm(OMod); 1534 MRI->replaceRegWith(MI.getOperand(0).getReg(), Def->getOperand(0).getReg()); 1535 MI.eraseFromParent(); 1536 1537 // Use of output modifiers forces VOP3 encoding for a VOP2 mac/fmac 1538 // instruction, so we might as well convert it to the more flexible VOP3-only 1539 // mad/fma form. 1540 if (TII->convertToThreeAddress(*Def, nullptr, nullptr)) 1541 Def->eraseFromParent(); 1542 1543 return true; 1544 } 1545 1546 // Try to fold a reg_sequence with vgpr output and agpr inputs into an 1547 // instruction which can take an agpr. So far that means a store. 1548 bool SIFoldOperands::tryFoldRegSequence(MachineInstr &MI) { 1549 assert(MI.isRegSequence()); 1550 auto Reg = MI.getOperand(0).getReg(); 1551 1552 if (!ST->hasGFX90AInsts() || !TRI->isVGPR(*MRI, Reg) || 1553 !MRI->hasOneNonDBGUse(Reg)) 1554 return false; 1555 1556 SmallVector<std::pair<MachineOperand*, unsigned>, 32> Defs; 1557 if (!getRegSeqInit(Defs, Reg, MCOI::OPERAND_REGISTER, TII, *MRI)) 1558 return false; 1559 1560 for (auto &Def : Defs) { 1561 const auto *Op = Def.first; 1562 if (!Op->isReg()) 1563 return false; 1564 if (TRI->isAGPR(*MRI, Op->getReg())) 1565 continue; 1566 // Maybe this is a COPY from AREG 1567 const MachineInstr *SubDef = MRI->getVRegDef(Op->getReg()); 1568 if (!SubDef || !SubDef->isCopy() || SubDef->getOperand(1).getSubReg()) 1569 return false; 1570 if (!TRI->isAGPR(*MRI, SubDef->getOperand(1).getReg())) 1571 return false; 1572 } 1573 1574 MachineOperand *Op = &*MRI->use_nodbg_begin(Reg); 1575 MachineInstr *UseMI = Op->getParent(); 1576 while (UseMI->isCopy() && !Op->getSubReg()) { 1577 Reg = UseMI->getOperand(0).getReg(); 1578 if (!TRI->isVGPR(*MRI, Reg) || !MRI->hasOneNonDBGUse(Reg)) 1579 return false; 1580 Op = &*MRI->use_nodbg_begin(Reg); 1581 UseMI = Op->getParent(); 1582 } 1583 1584 if (Op->getSubReg()) 1585 return false; 1586 1587 unsigned OpIdx = Op - &UseMI->getOperand(0); 1588 const MCInstrDesc &InstDesc = UseMI->getDesc(); 1589 const MCOperandInfo &OpInfo = InstDesc.OpInfo[OpIdx]; 1590 switch (OpInfo.RegClass) { 1591 case AMDGPU::AV_32RegClassID: LLVM_FALLTHROUGH; 1592 case AMDGPU::AV_64RegClassID: LLVM_FALLTHROUGH; 1593 case AMDGPU::AV_96RegClassID: LLVM_FALLTHROUGH; 1594 case AMDGPU::AV_128RegClassID: LLVM_FALLTHROUGH; 1595 case AMDGPU::AV_160RegClassID: 1596 break; 1597 default: 1598 return false; 1599 } 1600 1601 const auto *NewDstRC = TRI->getEquivalentAGPRClass(MRI->getRegClass(Reg)); 1602 auto Dst = MRI->createVirtualRegister(NewDstRC); 1603 auto RS = BuildMI(*MI.getParent(), MI, MI.getDebugLoc(), 1604 TII->get(AMDGPU::REG_SEQUENCE), Dst); 1605 1606 for (unsigned I = 0; I < Defs.size(); ++I) { 1607 MachineOperand *Def = Defs[I].first; 1608 Def->setIsKill(false); 1609 if (TRI->isAGPR(*MRI, Def->getReg())) { 1610 RS.add(*Def); 1611 } else { // This is a copy 1612 MachineInstr *SubDef = MRI->getVRegDef(Def->getReg()); 1613 SubDef->getOperand(1).setIsKill(false); 1614 RS.addReg(SubDef->getOperand(1).getReg(), 0, Def->getSubReg()); 1615 } 1616 RS.addImm(Defs[I].second); 1617 } 1618 1619 Op->setReg(Dst); 1620 if (!TII->isOperandLegal(*UseMI, OpIdx, Op)) { 1621 Op->setReg(Reg); 1622 RS->eraseFromParent(); 1623 return false; 1624 } 1625 1626 LLVM_DEBUG(dbgs() << "Folded " << *RS << " into " << *UseMI); 1627 1628 // Erase the REG_SEQUENCE eagerly, unless we followed a chain of COPY users, 1629 // in which case we can erase them all later in runOnMachineFunction. 1630 if (MRI->use_nodbg_empty(MI.getOperand(0).getReg())) 1631 MI.eraseFromParentAndMarkDBGValuesForRemoval(); 1632 return true; 1633 } 1634 1635 // Try to hoist an AGPR to VGPR copy out of the loop across a LCSSA PHI. 1636 // This should allow folding of an AGPR into a consumer which may support it. 1637 // I.e.: 1638 // 1639 // loop: // loop: 1640 // %1:vreg = COPY %0:areg // exit: 1641 // exit: => // %1:areg = PHI %0:areg, %loop 1642 // %2:vreg = PHI %1:vreg, %loop // %2:vreg = COPY %1:areg 1643 bool SIFoldOperands::tryFoldLCSSAPhi(MachineInstr &PHI) { 1644 assert(PHI.isPHI()); 1645 1646 if (PHI.getNumExplicitOperands() != 3) // Single input LCSSA PHI 1647 return false; 1648 1649 Register PhiIn = PHI.getOperand(1).getReg(); 1650 Register PhiOut = PHI.getOperand(0).getReg(); 1651 if (PHI.getOperand(1).getSubReg() || 1652 !TRI->isVGPR(*MRI, PhiIn) || !TRI->isVGPR(*MRI, PhiOut)) 1653 return false; 1654 1655 // A single use should not matter for correctness, but if it has another use 1656 // inside the loop we may perform copy twice in a worst case. 1657 if (!MRI->hasOneNonDBGUse(PhiIn)) 1658 return false; 1659 1660 MachineInstr *Copy = MRI->getVRegDef(PhiIn); 1661 if (!Copy || !Copy->isCopy()) 1662 return false; 1663 1664 Register CopyIn = Copy->getOperand(1).getReg(); 1665 if (!TRI->isAGPR(*MRI, CopyIn) || Copy->getOperand(1).getSubReg()) 1666 return false; 1667 1668 const TargetRegisterClass *ARC = MRI->getRegClass(CopyIn); 1669 Register NewReg = MRI->createVirtualRegister(ARC); 1670 PHI.getOperand(1).setReg(CopyIn); 1671 PHI.getOperand(0).setReg(NewReg); 1672 1673 MachineBasicBlock *MBB = PHI.getParent(); 1674 BuildMI(*MBB, MBB->getFirstNonPHI(), Copy->getDebugLoc(), 1675 TII->get(AMDGPU::COPY), PhiOut) 1676 .addReg(NewReg, RegState::Kill); 1677 Copy->eraseFromParent(); // We know this copy had a single use. 1678 1679 LLVM_DEBUG(dbgs() << "Folded " << PHI); 1680 1681 return true; 1682 } 1683 1684 // Attempt to convert VGPR load to an AGPR load. 1685 bool SIFoldOperands::tryFoldLoad(MachineInstr &MI) { 1686 assert(MI.mayLoad()); 1687 if (!ST->hasGFX90AInsts() || MI.getNumExplicitDefs() != 1) 1688 return false; 1689 1690 MachineOperand &Def = MI.getOperand(0); 1691 if (!Def.isDef()) 1692 return false; 1693 1694 Register DefReg = Def.getReg(); 1695 1696 if (DefReg.isPhysical() || !TRI->isVGPR(*MRI, DefReg)) 1697 return false; 1698 1699 SmallVector<const MachineInstr*, 8> Users; 1700 SmallVector<Register, 8> MoveRegs; 1701 for (const MachineInstr &I : MRI->use_nodbg_instructions(DefReg)) { 1702 Users.push_back(&I); 1703 } 1704 if (Users.empty()) 1705 return false; 1706 1707 // Check that all uses a copy to an agpr or a reg_sequence producing an agpr. 1708 while (!Users.empty()) { 1709 const MachineInstr *I = Users.pop_back_val(); 1710 if (!I->isCopy() && !I->isRegSequence()) 1711 return false; 1712 Register DstReg = I->getOperand(0).getReg(); 1713 if (TRI->isAGPR(*MRI, DstReg)) 1714 continue; 1715 MoveRegs.push_back(DstReg); 1716 for (const MachineInstr &U : MRI->use_nodbg_instructions(DstReg)) { 1717 Users.push_back(&U); 1718 } 1719 } 1720 1721 const TargetRegisterClass *RC = MRI->getRegClass(DefReg); 1722 MRI->setRegClass(DefReg, TRI->getEquivalentAGPRClass(RC)); 1723 if (!TII->isOperandLegal(MI, 0, &Def)) { 1724 MRI->setRegClass(DefReg, RC); 1725 return false; 1726 } 1727 1728 while (!MoveRegs.empty()) { 1729 Register Reg = MoveRegs.pop_back_val(); 1730 MRI->setRegClass(Reg, TRI->getEquivalentAGPRClass(MRI->getRegClass(Reg))); 1731 } 1732 1733 LLVM_DEBUG(dbgs() << "Folded " << MI); 1734 1735 return true; 1736 } 1737 1738 bool SIFoldOperands::runOnMachineFunction(MachineFunction &MF) { 1739 if (skipFunction(MF.getFunction())) 1740 return false; 1741 1742 MRI = &MF.getRegInfo(); 1743 ST = &MF.getSubtarget<GCNSubtarget>(); 1744 TII = ST->getInstrInfo(); 1745 TRI = &TII->getRegisterInfo(); 1746 MFI = MF.getInfo<SIMachineFunctionInfo>(); 1747 1748 // omod is ignored by hardware if IEEE bit is enabled. omod also does not 1749 // correctly handle signed zeros. 1750 // 1751 // FIXME: Also need to check strictfp 1752 bool IsIEEEMode = MFI->getMode().IEEE; 1753 bool HasNSZ = MFI->hasNoSignedZerosFPMath(); 1754 1755 for (MachineBasicBlock *MBB : depth_first(&MF)) { 1756 MachineOperand *CurrentKnownM0Val = nullptr; 1757 for (auto &MI : make_early_inc_range(*MBB)) { 1758 tryFoldCndMask(MI); 1759 1760 if (tryFoldZeroHighBits(MI)) 1761 continue; 1762 1763 if (MI.isRegSequence() && tryFoldRegSequence(MI)) 1764 continue; 1765 1766 if (MI.isPHI() && tryFoldLCSSAPhi(MI)) 1767 continue; 1768 1769 if (MI.mayLoad() && tryFoldLoad(MI)) 1770 continue; 1771 1772 if (!TII->isFoldableCopy(MI)) { 1773 // Saw an unknown clobber of m0, so we no longer know what it is. 1774 if (CurrentKnownM0Val && MI.modifiesRegister(AMDGPU::M0, TRI)) 1775 CurrentKnownM0Val = nullptr; 1776 1777 // TODO: Omod might be OK if there is NSZ only on the source 1778 // instruction, and not the omod multiply. 1779 if (IsIEEEMode || (!HasNSZ && !MI.getFlag(MachineInstr::FmNsz)) || 1780 !tryFoldOMod(MI)) 1781 tryFoldClamp(MI); 1782 1783 continue; 1784 } 1785 1786 // Specially track simple redefs of m0 to the same value in a block, so we 1787 // can erase the later ones. 1788 if (MI.getOperand(0).getReg() == AMDGPU::M0) { 1789 MachineOperand &NewM0Val = MI.getOperand(1); 1790 if (CurrentKnownM0Val && CurrentKnownM0Val->isIdenticalTo(NewM0Val)) { 1791 MI.eraseFromParent(); 1792 continue; 1793 } 1794 1795 // We aren't tracking other physical registers 1796 CurrentKnownM0Val = (NewM0Val.isReg() && NewM0Val.getReg().isPhysical()) ? 1797 nullptr : &NewM0Val; 1798 continue; 1799 } 1800 1801 MachineOperand &OpToFold = MI.getOperand(1); 1802 bool FoldingImm = 1803 OpToFold.isImm() || OpToFold.isFI() || OpToFold.isGlobal(); 1804 1805 // FIXME: We could also be folding things like TargetIndexes. 1806 if (!FoldingImm && !OpToFold.isReg()) 1807 continue; 1808 1809 if (OpToFold.isReg() && !OpToFold.getReg().isVirtual()) 1810 continue; 1811 1812 // Prevent folding operands backwards in the function. For example, 1813 // the COPY opcode must not be replaced by 1 in this example: 1814 // 1815 // %3 = COPY %vgpr0; VGPR_32:%3 1816 // ... 1817 // %vgpr0 = V_MOV_B32_e32 1, implicit %exec 1818 if (!MI.getOperand(0).getReg().isVirtual()) 1819 continue; 1820 1821 foldInstOperand(MI, OpToFold); 1822 1823 // If we managed to fold all uses of this copy then we might as well 1824 // delete it now. 1825 // The only reason we need to follow chains of copies here is that 1826 // tryFoldRegSequence looks forward through copies before folding a 1827 // REG_SEQUENCE into its eventual users. 1828 auto *InstToErase = &MI; 1829 while (MRI->use_nodbg_empty(InstToErase->getOperand(0).getReg())) { 1830 auto &SrcOp = InstToErase->getOperand(1); 1831 auto SrcReg = SrcOp.isReg() ? SrcOp.getReg() : Register(); 1832 InstToErase->eraseFromParentAndMarkDBGValuesForRemoval(); 1833 InstToErase = nullptr; 1834 if (!SrcReg || SrcReg.isPhysical()) 1835 break; 1836 InstToErase = MRI->getVRegDef(SrcReg); 1837 if (!InstToErase || !TII->isFoldableCopy(*InstToErase)) 1838 break; 1839 } 1840 if (InstToErase && InstToErase->isRegSequence() && 1841 MRI->use_nodbg_empty(InstToErase->getOperand(0).getReg())) 1842 InstToErase->eraseFromParentAndMarkDBGValuesForRemoval(); 1843 } 1844 } 1845 return true; 1846 } 1847