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