1 //===- llvm/CodeGen/GlobalISel/Utils.cpp -------------------------*- C++ -*-==// 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 //===----------------------------------------------------------------------===// 8 /// \file This file implements the utility functions used by the GlobalISel 9 /// pipeline. 10 //===----------------------------------------------------------------------===// 11 12 #include "llvm/CodeGen/GlobalISel/Utils.h" 13 #include "llvm/ADT/APFloat.h" 14 #include "llvm/ADT/APInt.h" 15 #include "llvm/CodeGen/CodeGenCommonISel.h" 16 #include "llvm/CodeGen/GlobalISel/GISelChangeObserver.h" 17 #include "llvm/CodeGen/GlobalISel/GISelKnownBits.h" 18 #include "llvm/CodeGen/GlobalISel/GenericMachineInstrs.h" 19 #include "llvm/CodeGen/GlobalISel/LostDebugLocObserver.h" 20 #include "llvm/CodeGen/GlobalISel/MIPatternMatch.h" 21 #include "llvm/CodeGen/MachineInstr.h" 22 #include "llvm/CodeGen/MachineInstrBuilder.h" 23 #include "llvm/CodeGen/MachineOptimizationRemarkEmitter.h" 24 #include "llvm/CodeGen/MachineRegisterInfo.h" 25 #include "llvm/CodeGen/MachineSizeOpts.h" 26 #include "llvm/CodeGen/RegisterBankInfo.h" 27 #include "llvm/CodeGen/StackProtector.h" 28 #include "llvm/CodeGen/TargetInstrInfo.h" 29 #include "llvm/CodeGen/TargetLowering.h" 30 #include "llvm/CodeGen/TargetPassConfig.h" 31 #include "llvm/CodeGen/TargetRegisterInfo.h" 32 #include "llvm/IR/Constants.h" 33 #include "llvm/Target/TargetMachine.h" 34 #include "llvm/Transforms/Utils/SizeOpts.h" 35 #include <numeric> 36 #include <optional> 37 38 #define DEBUG_TYPE "globalisel-utils" 39 40 using namespace llvm; 41 using namespace MIPatternMatch; 42 43 Register llvm::constrainRegToClass(MachineRegisterInfo &MRI, 44 const TargetInstrInfo &TII, 45 const RegisterBankInfo &RBI, Register Reg, 46 const TargetRegisterClass &RegClass) { 47 if (!RBI.constrainGenericRegister(Reg, RegClass, MRI)) 48 return MRI.createVirtualRegister(&RegClass); 49 50 return Reg; 51 } 52 53 Register llvm::constrainOperandRegClass( 54 const MachineFunction &MF, const TargetRegisterInfo &TRI, 55 MachineRegisterInfo &MRI, const TargetInstrInfo &TII, 56 const RegisterBankInfo &RBI, MachineInstr &InsertPt, 57 const TargetRegisterClass &RegClass, MachineOperand &RegMO) { 58 Register Reg = RegMO.getReg(); 59 // Assume physical registers are properly constrained. 60 assert(Register::isVirtualRegister(Reg) && "PhysReg not implemented"); 61 62 // Save the old register class to check whether 63 // the change notifications will be required. 64 // TODO: A better approach would be to pass 65 // the observers to constrainRegToClass(). 66 auto *OldRegClass = MRI.getRegClassOrNull(Reg); 67 Register ConstrainedReg = constrainRegToClass(MRI, TII, RBI, Reg, RegClass); 68 // If we created a new virtual register because the class is not compatible 69 // then create a copy between the new and the old register. 70 if (ConstrainedReg != Reg) { 71 MachineBasicBlock::iterator InsertIt(&InsertPt); 72 MachineBasicBlock &MBB = *InsertPt.getParent(); 73 // FIXME: The copy needs to have the classes constrained for its operands. 74 // Use operand's regbank to get the class for old register (Reg). 75 if (RegMO.isUse()) { 76 BuildMI(MBB, InsertIt, InsertPt.getDebugLoc(), 77 TII.get(TargetOpcode::COPY), ConstrainedReg) 78 .addReg(Reg); 79 } else { 80 assert(RegMO.isDef() && "Must be a definition"); 81 BuildMI(MBB, std::next(InsertIt), InsertPt.getDebugLoc(), 82 TII.get(TargetOpcode::COPY), Reg) 83 .addReg(ConstrainedReg); 84 } 85 if (GISelChangeObserver *Observer = MF.getObserver()) { 86 Observer->changingInstr(*RegMO.getParent()); 87 } 88 RegMO.setReg(ConstrainedReg); 89 if (GISelChangeObserver *Observer = MF.getObserver()) { 90 Observer->changedInstr(*RegMO.getParent()); 91 } 92 } else if (OldRegClass != MRI.getRegClassOrNull(Reg)) { 93 if (GISelChangeObserver *Observer = MF.getObserver()) { 94 if (!RegMO.isDef()) { 95 MachineInstr *RegDef = MRI.getVRegDef(Reg); 96 Observer->changedInstr(*RegDef); 97 } 98 Observer->changingAllUsesOfReg(MRI, Reg); 99 Observer->finishedChangingAllUsesOfReg(); 100 } 101 } 102 return ConstrainedReg; 103 } 104 105 Register llvm::constrainOperandRegClass( 106 const MachineFunction &MF, const TargetRegisterInfo &TRI, 107 MachineRegisterInfo &MRI, const TargetInstrInfo &TII, 108 const RegisterBankInfo &RBI, MachineInstr &InsertPt, const MCInstrDesc &II, 109 MachineOperand &RegMO, unsigned OpIdx) { 110 Register Reg = RegMO.getReg(); 111 // Assume physical registers are properly constrained. 112 assert(Register::isVirtualRegister(Reg) && "PhysReg not implemented"); 113 114 const TargetRegisterClass *OpRC = TII.getRegClass(II, OpIdx, &TRI, MF); 115 // Some of the target independent instructions, like COPY, may not impose any 116 // register class constraints on some of their operands: If it's a use, we can 117 // skip constraining as the instruction defining the register would constrain 118 // it. 119 120 if (OpRC) { 121 // Obtain the RC from incoming regbank if it is a proper sub-class. Operands 122 // can have multiple regbanks for a superclass that combine different 123 // register types (E.g., AMDGPU's VGPR and AGPR). The regbank ambiguity 124 // resolved by targets during regbankselect should not be overridden. 125 if (const auto *SubRC = TRI.getCommonSubClass( 126 OpRC, TRI.getConstrainedRegClassForOperand(RegMO, MRI))) 127 OpRC = SubRC; 128 129 OpRC = TRI.getAllocatableClass(OpRC); 130 } 131 132 if (!OpRC) { 133 assert((!isTargetSpecificOpcode(II.getOpcode()) || RegMO.isUse()) && 134 "Register class constraint is required unless either the " 135 "instruction is target independent or the operand is a use"); 136 // FIXME: Just bailing out like this here could be not enough, unless we 137 // expect the users of this function to do the right thing for PHIs and 138 // COPY: 139 // v1 = COPY v0 140 // v2 = COPY v1 141 // v1 here may end up not being constrained at all. Please notice that to 142 // reproduce the issue we likely need a destination pattern of a selection 143 // rule producing such extra copies, not just an input GMIR with them as 144 // every existing target using selectImpl handles copies before calling it 145 // and they never reach this function. 146 return Reg; 147 } 148 return constrainOperandRegClass(MF, TRI, MRI, TII, RBI, InsertPt, *OpRC, 149 RegMO); 150 } 151 152 bool llvm::constrainSelectedInstRegOperands(MachineInstr &I, 153 const TargetInstrInfo &TII, 154 const TargetRegisterInfo &TRI, 155 const RegisterBankInfo &RBI) { 156 assert(!isPreISelGenericOpcode(I.getOpcode()) && 157 "A selected instruction is expected"); 158 MachineBasicBlock &MBB = *I.getParent(); 159 MachineFunction &MF = *MBB.getParent(); 160 MachineRegisterInfo &MRI = MF.getRegInfo(); 161 162 for (unsigned OpI = 0, OpE = I.getNumExplicitOperands(); OpI != OpE; ++OpI) { 163 MachineOperand &MO = I.getOperand(OpI); 164 165 // There's nothing to be done on non-register operands. 166 if (!MO.isReg()) 167 continue; 168 169 LLVM_DEBUG(dbgs() << "Converting operand: " << MO << '\n'); 170 assert(MO.isReg() && "Unsupported non-reg operand"); 171 172 Register Reg = MO.getReg(); 173 // Physical registers don't need to be constrained. 174 if (Register::isPhysicalRegister(Reg)) 175 continue; 176 177 // Register operands with a value of 0 (e.g. predicate operands) don't need 178 // to be constrained. 179 if (Reg == 0) 180 continue; 181 182 // If the operand is a vreg, we should constrain its regclass, and only 183 // insert COPYs if that's impossible. 184 // constrainOperandRegClass does that for us. 185 constrainOperandRegClass(MF, TRI, MRI, TII, RBI, I, I.getDesc(), MO, OpI); 186 187 // Tie uses to defs as indicated in MCInstrDesc if this hasn't already been 188 // done. 189 if (MO.isUse()) { 190 int DefIdx = I.getDesc().getOperandConstraint(OpI, MCOI::TIED_TO); 191 if (DefIdx != -1 && !I.isRegTiedToUseOperand(DefIdx)) 192 I.tieOperands(DefIdx, OpI); 193 } 194 } 195 return true; 196 } 197 198 bool llvm::canReplaceReg(Register DstReg, Register SrcReg, 199 MachineRegisterInfo &MRI) { 200 // Give up if either DstReg or SrcReg is a physical register. 201 if (DstReg.isPhysical() || SrcReg.isPhysical()) 202 return false; 203 // Give up if the types don't match. 204 if (MRI.getType(DstReg) != MRI.getType(SrcReg)) 205 return false; 206 // Replace if either DstReg has no constraints or the register 207 // constraints match. 208 return !MRI.getRegClassOrRegBank(DstReg) || 209 MRI.getRegClassOrRegBank(DstReg) == MRI.getRegClassOrRegBank(SrcReg); 210 } 211 212 bool llvm::isTriviallyDead(const MachineInstr &MI, 213 const MachineRegisterInfo &MRI) { 214 // FIXME: This logical is mostly duplicated with 215 // DeadMachineInstructionElim::isDead. Why is LOCAL_ESCAPE not considered in 216 // MachineInstr::isLabel? 217 218 // Don't delete frame allocation labels. 219 if (MI.getOpcode() == TargetOpcode::LOCAL_ESCAPE) 220 return false; 221 // LIFETIME markers should be preserved even if they seem dead. 222 if (MI.getOpcode() == TargetOpcode::LIFETIME_START || 223 MI.getOpcode() == TargetOpcode::LIFETIME_END) 224 return false; 225 226 // If we can move an instruction, we can remove it. Otherwise, it has 227 // a side-effect of some sort. 228 bool SawStore = false; 229 if (!MI.isSafeToMove(/*AA=*/nullptr, SawStore) && !MI.isPHI()) 230 return false; 231 232 // Instructions without side-effects are dead iff they only define dead vregs. 233 for (const auto &MO : MI.operands()) { 234 if (!MO.isReg() || !MO.isDef()) 235 continue; 236 237 Register Reg = MO.getReg(); 238 if (Register::isPhysicalRegister(Reg) || !MRI.use_nodbg_empty(Reg)) 239 return false; 240 } 241 return true; 242 } 243 244 static void reportGISelDiagnostic(DiagnosticSeverity Severity, 245 MachineFunction &MF, 246 const TargetPassConfig &TPC, 247 MachineOptimizationRemarkEmitter &MORE, 248 MachineOptimizationRemarkMissed &R) { 249 bool IsFatal = Severity == DS_Error && 250 TPC.isGlobalISelAbortEnabled(); 251 // Print the function name explicitly if we don't have a debug location (which 252 // makes the diagnostic less useful) or if we're going to emit a raw error. 253 if (!R.getLocation().isValid() || IsFatal) 254 R << (" (in function: " + MF.getName() + ")").str(); 255 256 if (IsFatal) 257 report_fatal_error(Twine(R.getMsg())); 258 else 259 MORE.emit(R); 260 } 261 262 void llvm::reportGISelWarning(MachineFunction &MF, const TargetPassConfig &TPC, 263 MachineOptimizationRemarkEmitter &MORE, 264 MachineOptimizationRemarkMissed &R) { 265 reportGISelDiagnostic(DS_Warning, MF, TPC, MORE, R); 266 } 267 268 void llvm::reportGISelFailure(MachineFunction &MF, const TargetPassConfig &TPC, 269 MachineOptimizationRemarkEmitter &MORE, 270 MachineOptimizationRemarkMissed &R) { 271 MF.getProperties().set(MachineFunctionProperties::Property::FailedISel); 272 reportGISelDiagnostic(DS_Error, MF, TPC, MORE, R); 273 } 274 275 void llvm::reportGISelFailure(MachineFunction &MF, const TargetPassConfig &TPC, 276 MachineOptimizationRemarkEmitter &MORE, 277 const char *PassName, StringRef Msg, 278 const MachineInstr &MI) { 279 MachineOptimizationRemarkMissed R(PassName, "GISelFailure: ", 280 MI.getDebugLoc(), MI.getParent()); 281 R << Msg; 282 // Printing MI is expensive; only do it if expensive remarks are enabled. 283 if (TPC.isGlobalISelAbortEnabled() || MORE.allowExtraAnalysis(PassName)) 284 R << ": " << ore::MNV("Inst", MI); 285 reportGISelFailure(MF, TPC, MORE, R); 286 } 287 288 std::optional<APInt> llvm::getIConstantVRegVal(Register VReg, 289 const MachineRegisterInfo &MRI) { 290 std::optional<ValueAndVReg> ValAndVReg = getIConstantVRegValWithLookThrough( 291 VReg, MRI, /*LookThroughInstrs*/ false); 292 assert((!ValAndVReg || ValAndVReg->VReg == VReg) && 293 "Value found while looking through instrs"); 294 if (!ValAndVReg) 295 return std::nullopt; 296 return ValAndVReg->Value; 297 } 298 299 std::optional<int64_t> 300 llvm::getIConstantVRegSExtVal(Register VReg, const MachineRegisterInfo &MRI) { 301 std::optional<APInt> Val = getIConstantVRegVal(VReg, MRI); 302 if (Val && Val->getBitWidth() <= 64) 303 return Val->getSExtValue(); 304 return std::nullopt; 305 } 306 307 namespace { 308 309 typedef std::function<bool(const MachineInstr *)> IsOpcodeFn; 310 typedef std::function<std::optional<APInt>(const MachineInstr *MI)> GetAPCstFn; 311 312 std::optional<ValueAndVReg> getConstantVRegValWithLookThrough( 313 Register VReg, const MachineRegisterInfo &MRI, IsOpcodeFn IsConstantOpcode, 314 GetAPCstFn getAPCstValue, bool LookThroughInstrs = true, 315 bool LookThroughAnyExt = false) { 316 SmallVector<std::pair<unsigned, unsigned>, 4> SeenOpcodes; 317 MachineInstr *MI; 318 319 while ((MI = MRI.getVRegDef(VReg)) && !IsConstantOpcode(MI) && 320 LookThroughInstrs) { 321 switch (MI->getOpcode()) { 322 case TargetOpcode::G_ANYEXT: 323 if (!LookThroughAnyExt) 324 return std::nullopt; 325 [[fallthrough]]; 326 case TargetOpcode::G_TRUNC: 327 case TargetOpcode::G_SEXT: 328 case TargetOpcode::G_ZEXT: 329 SeenOpcodes.push_back(std::make_pair( 330 MI->getOpcode(), 331 MRI.getType(MI->getOperand(0).getReg()).getSizeInBits())); 332 VReg = MI->getOperand(1).getReg(); 333 break; 334 case TargetOpcode::COPY: 335 VReg = MI->getOperand(1).getReg(); 336 if (Register::isPhysicalRegister(VReg)) 337 return std::nullopt; 338 break; 339 case TargetOpcode::G_INTTOPTR: 340 VReg = MI->getOperand(1).getReg(); 341 break; 342 default: 343 return std::nullopt; 344 } 345 } 346 if (!MI || !IsConstantOpcode(MI)) 347 return std::nullopt; 348 349 std::optional<APInt> MaybeVal = getAPCstValue(MI); 350 if (!MaybeVal) 351 return std::nullopt; 352 APInt &Val = *MaybeVal; 353 while (!SeenOpcodes.empty()) { 354 std::pair<unsigned, unsigned> OpcodeAndSize = SeenOpcodes.pop_back_val(); 355 switch (OpcodeAndSize.first) { 356 case TargetOpcode::G_TRUNC: 357 Val = Val.trunc(OpcodeAndSize.second); 358 break; 359 case TargetOpcode::G_ANYEXT: 360 case TargetOpcode::G_SEXT: 361 Val = Val.sext(OpcodeAndSize.second); 362 break; 363 case TargetOpcode::G_ZEXT: 364 Val = Val.zext(OpcodeAndSize.second); 365 break; 366 } 367 } 368 369 return ValueAndVReg{Val, VReg}; 370 } 371 372 bool isIConstant(const MachineInstr *MI) { 373 if (!MI) 374 return false; 375 return MI->getOpcode() == TargetOpcode::G_CONSTANT; 376 } 377 378 bool isFConstant(const MachineInstr *MI) { 379 if (!MI) 380 return false; 381 return MI->getOpcode() == TargetOpcode::G_FCONSTANT; 382 } 383 384 bool isAnyConstant(const MachineInstr *MI) { 385 if (!MI) 386 return false; 387 unsigned Opc = MI->getOpcode(); 388 return Opc == TargetOpcode::G_CONSTANT || Opc == TargetOpcode::G_FCONSTANT; 389 } 390 391 std::optional<APInt> getCImmAsAPInt(const MachineInstr *MI) { 392 const MachineOperand &CstVal = MI->getOperand(1); 393 if (CstVal.isCImm()) 394 return CstVal.getCImm()->getValue(); 395 return std::nullopt; 396 } 397 398 std::optional<APInt> getCImmOrFPImmAsAPInt(const MachineInstr *MI) { 399 const MachineOperand &CstVal = MI->getOperand(1); 400 if (CstVal.isCImm()) 401 return CstVal.getCImm()->getValue(); 402 if (CstVal.isFPImm()) 403 return CstVal.getFPImm()->getValueAPF().bitcastToAPInt(); 404 return std::nullopt; 405 } 406 407 } // end anonymous namespace 408 409 std::optional<ValueAndVReg> llvm::getIConstantVRegValWithLookThrough( 410 Register VReg, const MachineRegisterInfo &MRI, bool LookThroughInstrs) { 411 return getConstantVRegValWithLookThrough(VReg, MRI, isIConstant, 412 getCImmAsAPInt, LookThroughInstrs); 413 } 414 415 std::optional<ValueAndVReg> llvm::getAnyConstantVRegValWithLookThrough( 416 Register VReg, const MachineRegisterInfo &MRI, bool LookThroughInstrs, 417 bool LookThroughAnyExt) { 418 return getConstantVRegValWithLookThrough( 419 VReg, MRI, isAnyConstant, getCImmOrFPImmAsAPInt, LookThroughInstrs, 420 LookThroughAnyExt); 421 } 422 423 std::optional<FPValueAndVReg> llvm::getFConstantVRegValWithLookThrough( 424 Register VReg, const MachineRegisterInfo &MRI, bool LookThroughInstrs) { 425 auto Reg = getConstantVRegValWithLookThrough( 426 VReg, MRI, isFConstant, getCImmOrFPImmAsAPInt, LookThroughInstrs); 427 if (!Reg) 428 return std::nullopt; 429 return FPValueAndVReg{getConstantFPVRegVal(Reg->VReg, MRI)->getValueAPF(), 430 Reg->VReg}; 431 } 432 433 const ConstantFP * 434 llvm::getConstantFPVRegVal(Register VReg, const MachineRegisterInfo &MRI) { 435 MachineInstr *MI = MRI.getVRegDef(VReg); 436 if (TargetOpcode::G_FCONSTANT != MI->getOpcode()) 437 return nullptr; 438 return MI->getOperand(1).getFPImm(); 439 } 440 441 std::optional<DefinitionAndSourceRegister> 442 llvm::getDefSrcRegIgnoringCopies(Register Reg, const MachineRegisterInfo &MRI) { 443 Register DefSrcReg = Reg; 444 auto *DefMI = MRI.getVRegDef(Reg); 445 auto DstTy = MRI.getType(DefMI->getOperand(0).getReg()); 446 if (!DstTy.isValid()) 447 return std::nullopt; 448 unsigned Opc = DefMI->getOpcode(); 449 while (Opc == TargetOpcode::COPY || isPreISelGenericOptimizationHint(Opc)) { 450 Register SrcReg = DefMI->getOperand(1).getReg(); 451 auto SrcTy = MRI.getType(SrcReg); 452 if (!SrcTy.isValid()) 453 break; 454 DefMI = MRI.getVRegDef(SrcReg); 455 DefSrcReg = SrcReg; 456 Opc = DefMI->getOpcode(); 457 } 458 return DefinitionAndSourceRegister{DefMI, DefSrcReg}; 459 } 460 461 MachineInstr *llvm::getDefIgnoringCopies(Register Reg, 462 const MachineRegisterInfo &MRI) { 463 std::optional<DefinitionAndSourceRegister> DefSrcReg = 464 getDefSrcRegIgnoringCopies(Reg, MRI); 465 return DefSrcReg ? DefSrcReg->MI : nullptr; 466 } 467 468 Register llvm::getSrcRegIgnoringCopies(Register Reg, 469 const MachineRegisterInfo &MRI) { 470 std::optional<DefinitionAndSourceRegister> DefSrcReg = 471 getDefSrcRegIgnoringCopies(Reg, MRI); 472 return DefSrcReg ? DefSrcReg->Reg : Register(); 473 } 474 475 MachineInstr *llvm::getOpcodeDef(unsigned Opcode, Register Reg, 476 const MachineRegisterInfo &MRI) { 477 MachineInstr *DefMI = getDefIgnoringCopies(Reg, MRI); 478 return DefMI && DefMI->getOpcode() == Opcode ? DefMI : nullptr; 479 } 480 481 APFloat llvm::getAPFloatFromSize(double Val, unsigned Size) { 482 if (Size == 32) 483 return APFloat(float(Val)); 484 if (Size == 64) 485 return APFloat(Val); 486 if (Size != 16) 487 llvm_unreachable("Unsupported FPConstant size"); 488 bool Ignored; 489 APFloat APF(Val); 490 APF.convert(APFloat::IEEEhalf(), APFloat::rmNearestTiesToEven, &Ignored); 491 return APF; 492 } 493 494 std::optional<APInt> llvm::ConstantFoldBinOp(unsigned Opcode, 495 const Register Op1, 496 const Register Op2, 497 const MachineRegisterInfo &MRI) { 498 auto MaybeOp2Cst = getAnyConstantVRegValWithLookThrough(Op2, MRI, false); 499 if (!MaybeOp2Cst) 500 return std::nullopt; 501 502 auto MaybeOp1Cst = getAnyConstantVRegValWithLookThrough(Op1, MRI, false); 503 if (!MaybeOp1Cst) 504 return std::nullopt; 505 506 const APInt &C1 = MaybeOp1Cst->Value; 507 const APInt &C2 = MaybeOp2Cst->Value; 508 switch (Opcode) { 509 default: 510 break; 511 case TargetOpcode::G_ADD: 512 case TargetOpcode::G_PTR_ADD: 513 return C1 + C2; 514 case TargetOpcode::G_AND: 515 return C1 & C2; 516 case TargetOpcode::G_ASHR: 517 return C1.ashr(C2); 518 case TargetOpcode::G_LSHR: 519 return C1.lshr(C2); 520 case TargetOpcode::G_MUL: 521 return C1 * C2; 522 case TargetOpcode::G_OR: 523 return C1 | C2; 524 case TargetOpcode::G_SHL: 525 return C1 << C2; 526 case TargetOpcode::G_SUB: 527 return C1 - C2; 528 case TargetOpcode::G_XOR: 529 return C1 ^ C2; 530 case TargetOpcode::G_UDIV: 531 if (!C2.getBoolValue()) 532 break; 533 return C1.udiv(C2); 534 case TargetOpcode::G_SDIV: 535 if (!C2.getBoolValue()) 536 break; 537 return C1.sdiv(C2); 538 case TargetOpcode::G_UREM: 539 if (!C2.getBoolValue()) 540 break; 541 return C1.urem(C2); 542 case TargetOpcode::G_SREM: 543 if (!C2.getBoolValue()) 544 break; 545 return C1.srem(C2); 546 case TargetOpcode::G_SMIN: 547 return APIntOps::smin(C1, C2); 548 case TargetOpcode::G_SMAX: 549 return APIntOps::smax(C1, C2); 550 case TargetOpcode::G_UMIN: 551 return APIntOps::umin(C1, C2); 552 case TargetOpcode::G_UMAX: 553 return APIntOps::umax(C1, C2); 554 } 555 556 return std::nullopt; 557 } 558 559 std::optional<APFloat> 560 llvm::ConstantFoldFPBinOp(unsigned Opcode, const Register Op1, 561 const Register Op2, const MachineRegisterInfo &MRI) { 562 const ConstantFP *Op2Cst = getConstantFPVRegVal(Op2, MRI); 563 if (!Op2Cst) 564 return std::nullopt; 565 566 const ConstantFP *Op1Cst = getConstantFPVRegVal(Op1, MRI); 567 if (!Op1Cst) 568 return std::nullopt; 569 570 APFloat C1 = Op1Cst->getValueAPF(); 571 const APFloat &C2 = Op2Cst->getValueAPF(); 572 switch (Opcode) { 573 case TargetOpcode::G_FADD: 574 C1.add(C2, APFloat::rmNearestTiesToEven); 575 return C1; 576 case TargetOpcode::G_FSUB: 577 C1.subtract(C2, APFloat::rmNearestTiesToEven); 578 return C1; 579 case TargetOpcode::G_FMUL: 580 C1.multiply(C2, APFloat::rmNearestTiesToEven); 581 return C1; 582 case TargetOpcode::G_FDIV: 583 C1.divide(C2, APFloat::rmNearestTiesToEven); 584 return C1; 585 case TargetOpcode::G_FREM: 586 C1.mod(C2); 587 return C1; 588 case TargetOpcode::G_FCOPYSIGN: 589 C1.copySign(C2); 590 return C1; 591 case TargetOpcode::G_FMINNUM: 592 return minnum(C1, C2); 593 case TargetOpcode::G_FMAXNUM: 594 return maxnum(C1, C2); 595 case TargetOpcode::G_FMINIMUM: 596 return minimum(C1, C2); 597 case TargetOpcode::G_FMAXIMUM: 598 return maximum(C1, C2); 599 case TargetOpcode::G_FMINNUM_IEEE: 600 case TargetOpcode::G_FMAXNUM_IEEE: 601 // FIXME: These operations were unfortunately named. fminnum/fmaxnum do not 602 // follow the IEEE behavior for signaling nans and follow libm's fmin/fmax, 603 // and currently there isn't a nice wrapper in APFloat for the version with 604 // correct snan handling. 605 break; 606 default: 607 break; 608 } 609 610 return std::nullopt; 611 } 612 613 SmallVector<APInt> 614 llvm::ConstantFoldVectorBinop(unsigned Opcode, const Register Op1, 615 const Register Op2, 616 const MachineRegisterInfo &MRI) { 617 auto *SrcVec2 = getOpcodeDef<GBuildVector>(Op2, MRI); 618 if (!SrcVec2) 619 return SmallVector<APInt>(); 620 621 auto *SrcVec1 = getOpcodeDef<GBuildVector>(Op1, MRI); 622 if (!SrcVec1) 623 return SmallVector<APInt>(); 624 625 SmallVector<APInt> FoldedElements; 626 for (unsigned Idx = 0, E = SrcVec1->getNumSources(); Idx < E; ++Idx) { 627 auto MaybeCst = ConstantFoldBinOp(Opcode, SrcVec1->getSourceReg(Idx), 628 SrcVec2->getSourceReg(Idx), MRI); 629 if (!MaybeCst) 630 return SmallVector<APInt>(); 631 FoldedElements.push_back(*MaybeCst); 632 } 633 return FoldedElements; 634 } 635 636 bool llvm::isKnownNeverNaN(Register Val, const MachineRegisterInfo &MRI, 637 bool SNaN) { 638 const MachineInstr *DefMI = MRI.getVRegDef(Val); 639 if (!DefMI) 640 return false; 641 642 const TargetMachine& TM = DefMI->getMF()->getTarget(); 643 if (DefMI->getFlag(MachineInstr::FmNoNans) || TM.Options.NoNaNsFPMath) 644 return true; 645 646 // If the value is a constant, we can obviously see if it is a NaN or not. 647 if (const ConstantFP *FPVal = getConstantFPVRegVal(Val, MRI)) { 648 return !FPVal->getValueAPF().isNaN() || 649 (SNaN && !FPVal->getValueAPF().isSignaling()); 650 } 651 652 if (DefMI->getOpcode() == TargetOpcode::G_BUILD_VECTOR) { 653 for (const auto &Op : DefMI->uses()) 654 if (!isKnownNeverNaN(Op.getReg(), MRI, SNaN)) 655 return false; 656 return true; 657 } 658 659 switch (DefMI->getOpcode()) { 660 default: 661 break; 662 case TargetOpcode::G_FADD: 663 case TargetOpcode::G_FSUB: 664 case TargetOpcode::G_FMUL: 665 case TargetOpcode::G_FDIV: 666 case TargetOpcode::G_FREM: 667 case TargetOpcode::G_FSIN: 668 case TargetOpcode::G_FCOS: 669 case TargetOpcode::G_FMA: 670 case TargetOpcode::G_FMAD: 671 if (SNaN) 672 return true; 673 674 // TODO: Need isKnownNeverInfinity 675 return false; 676 case TargetOpcode::G_FMINNUM_IEEE: 677 case TargetOpcode::G_FMAXNUM_IEEE: { 678 if (SNaN) 679 return true; 680 // This can return a NaN if either operand is an sNaN, or if both operands 681 // are NaN. 682 return (isKnownNeverNaN(DefMI->getOperand(1).getReg(), MRI) && 683 isKnownNeverSNaN(DefMI->getOperand(2).getReg(), MRI)) || 684 (isKnownNeverSNaN(DefMI->getOperand(1).getReg(), MRI) && 685 isKnownNeverNaN(DefMI->getOperand(2).getReg(), MRI)); 686 } 687 case TargetOpcode::G_FMINNUM: 688 case TargetOpcode::G_FMAXNUM: { 689 // Only one needs to be known not-nan, since it will be returned if the 690 // other ends up being one. 691 return isKnownNeverNaN(DefMI->getOperand(1).getReg(), MRI, SNaN) || 692 isKnownNeverNaN(DefMI->getOperand(2).getReg(), MRI, SNaN); 693 } 694 } 695 696 if (SNaN) { 697 // FP operations quiet. For now, just handle the ones inserted during 698 // legalization. 699 switch (DefMI->getOpcode()) { 700 case TargetOpcode::G_FPEXT: 701 case TargetOpcode::G_FPTRUNC: 702 case TargetOpcode::G_FCANONICALIZE: 703 return true; 704 default: 705 return false; 706 } 707 } 708 709 return false; 710 } 711 712 Align llvm::inferAlignFromPtrInfo(MachineFunction &MF, 713 const MachinePointerInfo &MPO) { 714 auto PSV = MPO.V.dyn_cast<const PseudoSourceValue *>(); 715 if (auto FSPV = dyn_cast_or_null<FixedStackPseudoSourceValue>(PSV)) { 716 MachineFrameInfo &MFI = MF.getFrameInfo(); 717 return commonAlignment(MFI.getObjectAlign(FSPV->getFrameIndex()), 718 MPO.Offset); 719 } 720 721 if (const Value *V = MPO.V.dyn_cast<const Value *>()) { 722 const Module *M = MF.getFunction().getParent(); 723 return V->getPointerAlignment(M->getDataLayout()); 724 } 725 726 return Align(1); 727 } 728 729 Register llvm::getFunctionLiveInPhysReg(MachineFunction &MF, 730 const TargetInstrInfo &TII, 731 MCRegister PhysReg, 732 const TargetRegisterClass &RC, 733 const DebugLoc &DL, LLT RegTy) { 734 MachineBasicBlock &EntryMBB = MF.front(); 735 MachineRegisterInfo &MRI = MF.getRegInfo(); 736 Register LiveIn = MRI.getLiveInVirtReg(PhysReg); 737 if (LiveIn) { 738 MachineInstr *Def = MRI.getVRegDef(LiveIn); 739 if (Def) { 740 // FIXME: Should the verifier check this is in the entry block? 741 assert(Def->getParent() == &EntryMBB && "live-in copy not in entry block"); 742 return LiveIn; 743 } 744 745 // It's possible the incoming argument register and copy was added during 746 // lowering, but later deleted due to being/becoming dead. If this happens, 747 // re-insert the copy. 748 } else { 749 // The live in register was not present, so add it. 750 LiveIn = MF.addLiveIn(PhysReg, &RC); 751 if (RegTy.isValid()) 752 MRI.setType(LiveIn, RegTy); 753 } 754 755 BuildMI(EntryMBB, EntryMBB.begin(), DL, TII.get(TargetOpcode::COPY), LiveIn) 756 .addReg(PhysReg); 757 if (!EntryMBB.isLiveIn(PhysReg)) 758 EntryMBB.addLiveIn(PhysReg); 759 return LiveIn; 760 } 761 762 std::optional<APInt> llvm::ConstantFoldExtOp(unsigned Opcode, 763 const Register Op1, uint64_t Imm, 764 const MachineRegisterInfo &MRI) { 765 auto MaybeOp1Cst = getIConstantVRegVal(Op1, MRI); 766 if (MaybeOp1Cst) { 767 switch (Opcode) { 768 default: 769 break; 770 case TargetOpcode::G_SEXT_INREG: { 771 LLT Ty = MRI.getType(Op1); 772 return MaybeOp1Cst->trunc(Imm).sext(Ty.getScalarSizeInBits()); 773 } 774 } 775 } 776 return std::nullopt; 777 } 778 779 std::optional<APFloat> 780 llvm::ConstantFoldIntToFloat(unsigned Opcode, LLT DstTy, Register Src, 781 const MachineRegisterInfo &MRI) { 782 assert(Opcode == TargetOpcode::G_SITOFP || Opcode == TargetOpcode::G_UITOFP); 783 if (auto MaybeSrcVal = getIConstantVRegVal(Src, MRI)) { 784 APFloat DstVal(getFltSemanticForLLT(DstTy)); 785 DstVal.convertFromAPInt(*MaybeSrcVal, Opcode == TargetOpcode::G_SITOFP, 786 APFloat::rmNearestTiesToEven); 787 return DstVal; 788 } 789 return std::nullopt; 790 } 791 792 std::optional<SmallVector<unsigned>> 793 llvm::ConstantFoldCTLZ(Register Src, const MachineRegisterInfo &MRI) { 794 LLT Ty = MRI.getType(Src); 795 SmallVector<unsigned> FoldedCTLZs; 796 auto tryFoldScalar = [&](Register R) -> std::optional<unsigned> { 797 auto MaybeCst = getIConstantVRegVal(R, MRI); 798 if (!MaybeCst) 799 return std::nullopt; 800 return MaybeCst->countLeadingZeros(); 801 }; 802 if (Ty.isVector()) { 803 // Try to constant fold each element. 804 auto *BV = getOpcodeDef<GBuildVector>(Src, MRI); 805 if (!BV) 806 return std::nullopt; 807 for (unsigned SrcIdx = 0; SrcIdx < BV->getNumSources(); ++SrcIdx) { 808 if (auto MaybeFold = tryFoldScalar(BV->getSourceReg(SrcIdx))) { 809 FoldedCTLZs.emplace_back(*MaybeFold); 810 continue; 811 } 812 return std::nullopt; 813 } 814 return FoldedCTLZs; 815 } 816 if (auto MaybeCst = tryFoldScalar(Src)) { 817 FoldedCTLZs.emplace_back(*MaybeCst); 818 return FoldedCTLZs; 819 } 820 return std::nullopt; 821 } 822 823 bool llvm::isKnownToBeAPowerOfTwo(Register Reg, const MachineRegisterInfo &MRI, 824 GISelKnownBits *KB) { 825 std::optional<DefinitionAndSourceRegister> DefSrcReg = 826 getDefSrcRegIgnoringCopies(Reg, MRI); 827 if (!DefSrcReg) 828 return false; 829 830 const MachineInstr &MI = *DefSrcReg->MI; 831 const LLT Ty = MRI.getType(Reg); 832 833 switch (MI.getOpcode()) { 834 case TargetOpcode::G_CONSTANT: { 835 unsigned BitWidth = Ty.getScalarSizeInBits(); 836 const ConstantInt *CI = MI.getOperand(1).getCImm(); 837 return CI->getValue().zextOrTrunc(BitWidth).isPowerOf2(); 838 } 839 case TargetOpcode::G_SHL: { 840 // A left-shift of a constant one will have exactly one bit set because 841 // shifting the bit off the end is undefined. 842 843 // TODO: Constant splat 844 if (auto ConstLHS = getIConstantVRegVal(MI.getOperand(1).getReg(), MRI)) { 845 if (*ConstLHS == 1) 846 return true; 847 } 848 849 break; 850 } 851 case TargetOpcode::G_LSHR: { 852 if (auto ConstLHS = getIConstantVRegVal(MI.getOperand(1).getReg(), MRI)) { 853 if (ConstLHS->isSignMask()) 854 return true; 855 } 856 857 break; 858 } 859 case TargetOpcode::G_BUILD_VECTOR: { 860 // TODO: Probably should have a recursion depth guard since you could have 861 // bitcasted vector elements. 862 for (const MachineOperand &MO : llvm::drop_begin(MI.operands())) 863 if (!isKnownToBeAPowerOfTwo(MO.getReg(), MRI, KB)) 864 return false; 865 866 return true; 867 } 868 case TargetOpcode::G_BUILD_VECTOR_TRUNC: { 869 // Only handle constants since we would need to know if number of leading 870 // zeros is greater than the truncation amount. 871 const unsigned BitWidth = Ty.getScalarSizeInBits(); 872 for (const MachineOperand &MO : llvm::drop_begin(MI.operands())) { 873 auto Const = getIConstantVRegVal(MO.getReg(), MRI); 874 if (!Const || !Const->zextOrTrunc(BitWidth).isPowerOf2()) 875 return false; 876 } 877 878 return true; 879 } 880 default: 881 break; 882 } 883 884 if (!KB) 885 return false; 886 887 // More could be done here, though the above checks are enough 888 // to handle some common cases. 889 890 // Fall back to computeKnownBits to catch other known cases. 891 KnownBits Known = KB->getKnownBits(Reg); 892 return (Known.countMaxPopulation() == 1) && (Known.countMinPopulation() == 1); 893 } 894 895 void llvm::getSelectionDAGFallbackAnalysisUsage(AnalysisUsage &AU) { 896 AU.addPreserved<StackProtector>(); 897 } 898 899 LLT llvm::getLCMType(LLT OrigTy, LLT TargetTy) { 900 const unsigned OrigSize = OrigTy.getSizeInBits(); 901 const unsigned TargetSize = TargetTy.getSizeInBits(); 902 903 if (OrigSize == TargetSize) 904 return OrigTy; 905 906 if (OrigTy.isVector()) { 907 const LLT OrigElt = OrigTy.getElementType(); 908 909 if (TargetTy.isVector()) { 910 const LLT TargetElt = TargetTy.getElementType(); 911 912 if (OrigElt.getSizeInBits() == TargetElt.getSizeInBits()) { 913 int GCDElts = 914 std::gcd(OrigTy.getNumElements(), TargetTy.getNumElements()); 915 // Prefer the original element type. 916 ElementCount Mul = OrigTy.getElementCount() * TargetTy.getNumElements(); 917 return LLT::vector(Mul.divideCoefficientBy(GCDElts), 918 OrigTy.getElementType()); 919 } 920 } else { 921 if (OrigElt.getSizeInBits() == TargetSize) 922 return OrigTy; 923 } 924 925 unsigned LCMSize = std::lcm(OrigSize, TargetSize); 926 return LLT::fixed_vector(LCMSize / OrigElt.getSizeInBits(), OrigElt); 927 } 928 929 if (TargetTy.isVector()) { 930 unsigned LCMSize = std::lcm(OrigSize, TargetSize); 931 return LLT::fixed_vector(LCMSize / OrigSize, OrigTy); 932 } 933 934 unsigned LCMSize = std::lcm(OrigSize, TargetSize); 935 936 // Preserve pointer types. 937 if (LCMSize == OrigSize) 938 return OrigTy; 939 if (LCMSize == TargetSize) 940 return TargetTy; 941 942 return LLT::scalar(LCMSize); 943 } 944 945 LLT llvm::getCoverTy(LLT OrigTy, LLT TargetTy) { 946 if (!OrigTy.isVector() || !TargetTy.isVector() || OrigTy == TargetTy || 947 (OrigTy.getScalarSizeInBits() != TargetTy.getScalarSizeInBits())) 948 return getLCMType(OrigTy, TargetTy); 949 950 unsigned OrigTyNumElts = OrigTy.getNumElements(); 951 unsigned TargetTyNumElts = TargetTy.getNumElements(); 952 if (OrigTyNumElts % TargetTyNumElts == 0) 953 return OrigTy; 954 955 unsigned NumElts = alignTo(OrigTyNumElts, TargetTyNumElts); 956 return LLT::scalarOrVector(ElementCount::getFixed(NumElts), 957 OrigTy.getElementType()); 958 } 959 960 LLT llvm::getGCDType(LLT OrigTy, LLT TargetTy) { 961 const unsigned OrigSize = OrigTy.getSizeInBits(); 962 const unsigned TargetSize = TargetTy.getSizeInBits(); 963 964 if (OrigSize == TargetSize) 965 return OrigTy; 966 967 if (OrigTy.isVector()) { 968 LLT OrigElt = OrigTy.getElementType(); 969 if (TargetTy.isVector()) { 970 LLT TargetElt = TargetTy.getElementType(); 971 if (OrigElt.getSizeInBits() == TargetElt.getSizeInBits()) { 972 int GCD = std::gcd(OrigTy.getNumElements(), TargetTy.getNumElements()); 973 return LLT::scalarOrVector(ElementCount::getFixed(GCD), OrigElt); 974 } 975 } else { 976 // If the source is a vector of pointers, return a pointer element. 977 if (OrigElt.getSizeInBits() == TargetSize) 978 return OrigElt; 979 } 980 981 unsigned GCD = std::gcd(OrigSize, TargetSize); 982 if (GCD == OrigElt.getSizeInBits()) 983 return OrigElt; 984 985 // If we can't produce the original element type, we have to use a smaller 986 // scalar. 987 if (GCD < OrigElt.getSizeInBits()) 988 return LLT::scalar(GCD); 989 return LLT::fixed_vector(GCD / OrigElt.getSizeInBits(), OrigElt); 990 } 991 992 if (TargetTy.isVector()) { 993 // Try to preserve the original element type. 994 LLT TargetElt = TargetTy.getElementType(); 995 if (TargetElt.getSizeInBits() == OrigSize) 996 return OrigTy; 997 } 998 999 unsigned GCD = std::gcd(OrigSize, TargetSize); 1000 return LLT::scalar(GCD); 1001 } 1002 1003 std::optional<int> llvm::getSplatIndex(MachineInstr &MI) { 1004 assert(MI.getOpcode() == TargetOpcode::G_SHUFFLE_VECTOR && 1005 "Only G_SHUFFLE_VECTOR can have a splat index!"); 1006 ArrayRef<int> Mask = MI.getOperand(3).getShuffleMask(); 1007 auto FirstDefinedIdx = find_if(Mask, [](int Elt) { return Elt >= 0; }); 1008 1009 // If all elements are undefined, this shuffle can be considered a splat. 1010 // Return 0 for better potential for callers to simplify. 1011 if (FirstDefinedIdx == Mask.end()) 1012 return 0; 1013 1014 // Make sure all remaining elements are either undef or the same 1015 // as the first non-undef value. 1016 int SplatValue = *FirstDefinedIdx; 1017 if (any_of(make_range(std::next(FirstDefinedIdx), Mask.end()), 1018 [&SplatValue](int Elt) { return Elt >= 0 && Elt != SplatValue; })) 1019 return std::nullopt; 1020 1021 return SplatValue; 1022 } 1023 1024 static bool isBuildVectorOp(unsigned Opcode) { 1025 return Opcode == TargetOpcode::G_BUILD_VECTOR || 1026 Opcode == TargetOpcode::G_BUILD_VECTOR_TRUNC; 1027 } 1028 1029 namespace { 1030 1031 std::optional<ValueAndVReg> getAnyConstantSplat(Register VReg, 1032 const MachineRegisterInfo &MRI, 1033 bool AllowUndef) { 1034 MachineInstr *MI = getDefIgnoringCopies(VReg, MRI); 1035 if (!MI) 1036 return std::nullopt; 1037 1038 if (!isBuildVectorOp(MI->getOpcode())) 1039 return std::nullopt; 1040 1041 std::optional<ValueAndVReg> SplatValAndReg; 1042 for (MachineOperand &Op : MI->uses()) { 1043 Register Element = Op.getReg(); 1044 auto ElementValAndReg = 1045 getAnyConstantVRegValWithLookThrough(Element, MRI, true, true); 1046 1047 // If AllowUndef, treat undef as value that will result in a constant splat. 1048 if (!ElementValAndReg) { 1049 if (AllowUndef && isa<GImplicitDef>(MRI.getVRegDef(Element))) 1050 continue; 1051 return std::nullopt; 1052 } 1053 1054 // Record splat value 1055 if (!SplatValAndReg) 1056 SplatValAndReg = ElementValAndReg; 1057 1058 // Different constant then the one already recorded, not a constant splat. 1059 if (SplatValAndReg->Value != ElementValAndReg->Value) 1060 return std::nullopt; 1061 } 1062 1063 return SplatValAndReg; 1064 } 1065 1066 } // end anonymous namespace 1067 1068 bool llvm::isBuildVectorConstantSplat(const Register Reg, 1069 const MachineRegisterInfo &MRI, 1070 int64_t SplatValue, bool AllowUndef) { 1071 if (auto SplatValAndReg = getAnyConstantSplat(Reg, MRI, AllowUndef)) 1072 return mi_match(SplatValAndReg->VReg, MRI, m_SpecificICst(SplatValue)); 1073 return false; 1074 } 1075 1076 bool llvm::isBuildVectorConstantSplat(const MachineInstr &MI, 1077 const MachineRegisterInfo &MRI, 1078 int64_t SplatValue, bool AllowUndef) { 1079 return isBuildVectorConstantSplat(MI.getOperand(0).getReg(), MRI, SplatValue, 1080 AllowUndef); 1081 } 1082 1083 std::optional<APInt> 1084 llvm::getIConstantSplatVal(const Register Reg, const MachineRegisterInfo &MRI) { 1085 if (auto SplatValAndReg = 1086 getAnyConstantSplat(Reg, MRI, /* AllowUndef */ false)) { 1087 std::optional<ValueAndVReg> ValAndVReg = 1088 getIConstantVRegValWithLookThrough(SplatValAndReg->VReg, MRI); 1089 return ValAndVReg->Value; 1090 } 1091 1092 return std::nullopt; 1093 } 1094 1095 std::optional<APInt> 1096 llvm::getIConstantSplatVal(const MachineInstr &MI, 1097 const MachineRegisterInfo &MRI) { 1098 return getIConstantSplatVal(MI.getOperand(0).getReg(), MRI); 1099 } 1100 1101 std::optional<int64_t> 1102 llvm::getIConstantSplatSExtVal(const Register Reg, 1103 const MachineRegisterInfo &MRI) { 1104 if (auto SplatValAndReg = 1105 getAnyConstantSplat(Reg, MRI, /* AllowUndef */ false)) 1106 return getIConstantVRegSExtVal(SplatValAndReg->VReg, MRI); 1107 return std::nullopt; 1108 } 1109 1110 std::optional<int64_t> 1111 llvm::getIConstantSplatSExtVal(const MachineInstr &MI, 1112 const MachineRegisterInfo &MRI) { 1113 return getIConstantSplatSExtVal(MI.getOperand(0).getReg(), MRI); 1114 } 1115 1116 std::optional<FPValueAndVReg> 1117 llvm::getFConstantSplat(Register VReg, const MachineRegisterInfo &MRI, 1118 bool AllowUndef) { 1119 if (auto SplatValAndReg = getAnyConstantSplat(VReg, MRI, AllowUndef)) 1120 return getFConstantVRegValWithLookThrough(SplatValAndReg->VReg, MRI); 1121 return std::nullopt; 1122 } 1123 1124 bool llvm::isBuildVectorAllZeros(const MachineInstr &MI, 1125 const MachineRegisterInfo &MRI, 1126 bool AllowUndef) { 1127 return isBuildVectorConstantSplat(MI, MRI, 0, AllowUndef); 1128 } 1129 1130 bool llvm::isBuildVectorAllOnes(const MachineInstr &MI, 1131 const MachineRegisterInfo &MRI, 1132 bool AllowUndef) { 1133 return isBuildVectorConstantSplat(MI, MRI, -1, AllowUndef); 1134 } 1135 1136 std::optional<RegOrConstant> 1137 llvm::getVectorSplat(const MachineInstr &MI, const MachineRegisterInfo &MRI) { 1138 unsigned Opc = MI.getOpcode(); 1139 if (!isBuildVectorOp(Opc)) 1140 return std::nullopt; 1141 if (auto Splat = getIConstantSplatSExtVal(MI, MRI)) 1142 return RegOrConstant(*Splat); 1143 auto Reg = MI.getOperand(1).getReg(); 1144 if (any_of(make_range(MI.operands_begin() + 2, MI.operands_end()), 1145 [&Reg](const MachineOperand &Op) { return Op.getReg() != Reg; })) 1146 return std::nullopt; 1147 return RegOrConstant(Reg); 1148 } 1149 1150 static bool isConstantScalar(const MachineInstr &MI, 1151 const MachineRegisterInfo &MRI, 1152 bool AllowFP = true, 1153 bool AllowOpaqueConstants = true) { 1154 switch (MI.getOpcode()) { 1155 case TargetOpcode::G_CONSTANT: 1156 case TargetOpcode::G_IMPLICIT_DEF: 1157 return true; 1158 case TargetOpcode::G_FCONSTANT: 1159 return AllowFP; 1160 case TargetOpcode::G_GLOBAL_VALUE: 1161 case TargetOpcode::G_FRAME_INDEX: 1162 case TargetOpcode::G_BLOCK_ADDR: 1163 case TargetOpcode::G_JUMP_TABLE: 1164 return AllowOpaqueConstants; 1165 default: 1166 return false; 1167 } 1168 } 1169 1170 bool llvm::isConstantOrConstantVector(MachineInstr &MI, 1171 const MachineRegisterInfo &MRI) { 1172 Register Def = MI.getOperand(0).getReg(); 1173 if (auto C = getIConstantVRegValWithLookThrough(Def, MRI)) 1174 return true; 1175 GBuildVector *BV = dyn_cast<GBuildVector>(&MI); 1176 if (!BV) 1177 return false; 1178 for (unsigned SrcIdx = 0; SrcIdx < BV->getNumSources(); ++SrcIdx) { 1179 if (getIConstantVRegValWithLookThrough(BV->getSourceReg(SrcIdx), MRI) || 1180 getOpcodeDef<GImplicitDef>(BV->getSourceReg(SrcIdx), MRI)) 1181 continue; 1182 return false; 1183 } 1184 return true; 1185 } 1186 1187 bool llvm::isConstantOrConstantVector(const MachineInstr &MI, 1188 const MachineRegisterInfo &MRI, 1189 bool AllowFP, bool AllowOpaqueConstants) { 1190 if (isConstantScalar(MI, MRI, AllowFP, AllowOpaqueConstants)) 1191 return true; 1192 1193 if (!isBuildVectorOp(MI.getOpcode())) 1194 return false; 1195 1196 const unsigned NumOps = MI.getNumOperands(); 1197 for (unsigned I = 1; I != NumOps; ++I) { 1198 const MachineInstr *ElementDef = MRI.getVRegDef(MI.getOperand(I).getReg()); 1199 if (!isConstantScalar(*ElementDef, MRI, AllowFP, AllowOpaqueConstants)) 1200 return false; 1201 } 1202 1203 return true; 1204 } 1205 1206 std::optional<APInt> 1207 llvm::isConstantOrConstantSplatVector(MachineInstr &MI, 1208 const MachineRegisterInfo &MRI) { 1209 Register Def = MI.getOperand(0).getReg(); 1210 if (auto C = getIConstantVRegValWithLookThrough(Def, MRI)) 1211 return C->Value; 1212 auto MaybeCst = getIConstantSplatSExtVal(MI, MRI); 1213 if (!MaybeCst) 1214 return std::nullopt; 1215 const unsigned ScalarSize = MRI.getType(Def).getScalarSizeInBits(); 1216 return APInt(ScalarSize, *MaybeCst, true); 1217 } 1218 1219 bool llvm::isNullOrNullSplat(const MachineInstr &MI, 1220 const MachineRegisterInfo &MRI, bool AllowUndefs) { 1221 switch (MI.getOpcode()) { 1222 case TargetOpcode::G_IMPLICIT_DEF: 1223 return AllowUndefs; 1224 case TargetOpcode::G_CONSTANT: 1225 return MI.getOperand(1).getCImm()->isNullValue(); 1226 case TargetOpcode::G_FCONSTANT: { 1227 const ConstantFP *FPImm = MI.getOperand(1).getFPImm(); 1228 return FPImm->isZero() && !FPImm->isNegative(); 1229 } 1230 default: 1231 if (!AllowUndefs) // TODO: isBuildVectorAllZeros assumes undef is OK already 1232 return false; 1233 return isBuildVectorAllZeros(MI, MRI); 1234 } 1235 } 1236 1237 bool llvm::isAllOnesOrAllOnesSplat(const MachineInstr &MI, 1238 const MachineRegisterInfo &MRI, 1239 bool AllowUndefs) { 1240 switch (MI.getOpcode()) { 1241 case TargetOpcode::G_IMPLICIT_DEF: 1242 return AllowUndefs; 1243 case TargetOpcode::G_CONSTANT: 1244 return MI.getOperand(1).getCImm()->isAllOnesValue(); 1245 default: 1246 if (!AllowUndefs) // TODO: isBuildVectorAllOnes assumes undef is OK already 1247 return false; 1248 return isBuildVectorAllOnes(MI, MRI); 1249 } 1250 } 1251 1252 bool llvm::matchUnaryPredicate( 1253 const MachineRegisterInfo &MRI, Register Reg, 1254 std::function<bool(const Constant *ConstVal)> Match, bool AllowUndefs) { 1255 1256 const MachineInstr *Def = getDefIgnoringCopies(Reg, MRI); 1257 if (AllowUndefs && Def->getOpcode() == TargetOpcode::G_IMPLICIT_DEF) 1258 return Match(nullptr); 1259 1260 // TODO: Also handle fconstant 1261 if (Def->getOpcode() == TargetOpcode::G_CONSTANT) 1262 return Match(Def->getOperand(1).getCImm()); 1263 1264 if (Def->getOpcode() != TargetOpcode::G_BUILD_VECTOR) 1265 return false; 1266 1267 for (unsigned I = 1, E = Def->getNumOperands(); I != E; ++I) { 1268 Register SrcElt = Def->getOperand(I).getReg(); 1269 const MachineInstr *SrcDef = getDefIgnoringCopies(SrcElt, MRI); 1270 if (AllowUndefs && SrcDef->getOpcode() == TargetOpcode::G_IMPLICIT_DEF) { 1271 if (!Match(nullptr)) 1272 return false; 1273 continue; 1274 } 1275 1276 if (SrcDef->getOpcode() != TargetOpcode::G_CONSTANT || 1277 !Match(SrcDef->getOperand(1).getCImm())) 1278 return false; 1279 } 1280 1281 return true; 1282 } 1283 1284 bool llvm::isConstTrueVal(const TargetLowering &TLI, int64_t Val, bool IsVector, 1285 bool IsFP) { 1286 switch (TLI.getBooleanContents(IsVector, IsFP)) { 1287 case TargetLowering::UndefinedBooleanContent: 1288 return Val & 0x1; 1289 case TargetLowering::ZeroOrOneBooleanContent: 1290 return Val == 1; 1291 case TargetLowering::ZeroOrNegativeOneBooleanContent: 1292 return Val == -1; 1293 } 1294 llvm_unreachable("Invalid boolean contents"); 1295 } 1296 1297 bool llvm::isConstFalseVal(const TargetLowering &TLI, int64_t Val, 1298 bool IsVector, bool IsFP) { 1299 switch (TLI.getBooleanContents(IsVector, IsFP)) { 1300 case TargetLowering::UndefinedBooleanContent: 1301 return ~Val & 0x1; 1302 case TargetLowering::ZeroOrOneBooleanContent: 1303 case TargetLowering::ZeroOrNegativeOneBooleanContent: 1304 return Val == 0; 1305 } 1306 llvm_unreachable("Invalid boolean contents"); 1307 } 1308 1309 int64_t llvm::getICmpTrueVal(const TargetLowering &TLI, bool IsVector, 1310 bool IsFP) { 1311 switch (TLI.getBooleanContents(IsVector, IsFP)) { 1312 case TargetLowering::UndefinedBooleanContent: 1313 case TargetLowering::ZeroOrOneBooleanContent: 1314 return 1; 1315 case TargetLowering::ZeroOrNegativeOneBooleanContent: 1316 return -1; 1317 } 1318 llvm_unreachable("Invalid boolean contents"); 1319 } 1320 1321 bool llvm::shouldOptForSize(const MachineBasicBlock &MBB, 1322 ProfileSummaryInfo *PSI, BlockFrequencyInfo *BFI) { 1323 const auto &F = MBB.getParent()->getFunction(); 1324 return F.hasOptSize() || F.hasMinSize() || 1325 llvm::shouldOptimizeForSize(MBB.getBasicBlock(), PSI, BFI); 1326 } 1327 1328 void llvm::saveUsesAndErase(MachineInstr &MI, MachineRegisterInfo &MRI, 1329 LostDebugLocObserver *LocObserver, 1330 SmallInstListTy &DeadInstChain) { 1331 for (MachineOperand &Op : MI.uses()) { 1332 if (Op.isReg() && Op.getReg().isVirtual()) 1333 DeadInstChain.insert(MRI.getVRegDef(Op.getReg())); 1334 } 1335 LLVM_DEBUG(dbgs() << MI << "Is dead; erasing.\n"); 1336 DeadInstChain.remove(&MI); 1337 MI.eraseFromParent(); 1338 if (LocObserver) 1339 LocObserver->checkpoint(false); 1340 } 1341 1342 void llvm::eraseInstrs(ArrayRef<MachineInstr *> DeadInstrs, 1343 MachineRegisterInfo &MRI, 1344 LostDebugLocObserver *LocObserver) { 1345 SmallInstListTy DeadInstChain; 1346 for (MachineInstr *MI : DeadInstrs) 1347 saveUsesAndErase(*MI, MRI, LocObserver, DeadInstChain); 1348 1349 while (!DeadInstChain.empty()) { 1350 MachineInstr *Inst = DeadInstChain.pop_back_val(); 1351 if (!isTriviallyDead(*Inst, MRI)) 1352 continue; 1353 saveUsesAndErase(*Inst, MRI, LocObserver, DeadInstChain); 1354 } 1355 } 1356 1357 void llvm::eraseInstr(MachineInstr &MI, MachineRegisterInfo &MRI, 1358 LostDebugLocObserver *LocObserver) { 1359 return eraseInstrs({&MI}, MRI, LocObserver); 1360 } 1361 1362 void llvm::salvageDebugInfo(const MachineRegisterInfo &MRI, MachineInstr &MI) { 1363 for (auto &Def : MI.defs()) { 1364 assert(Def.isReg() && "Must be a reg"); 1365 1366 SmallVector<MachineOperand *, 16> DbgUsers; 1367 for (auto &MOUse : MRI.use_operands(Def.getReg())) { 1368 MachineInstr *DbgValue = MOUse.getParent(); 1369 // Ignore partially formed DBG_VALUEs. 1370 if (DbgValue->isNonListDebugValue() && DbgValue->getNumOperands() == 4) { 1371 DbgUsers.push_back(&MOUse); 1372 } 1373 } 1374 1375 if (!DbgUsers.empty()) { 1376 salvageDebugInfoForDbgValue(MRI, MI, DbgUsers); 1377 } 1378 } 1379 } 1380