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