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/MIPatternMatch.h" 19 #include "llvm/CodeGen/GlobalISel/RegisterBankInfo.h" 20 #include "llvm/CodeGen/MachineInstr.h" 21 #include "llvm/CodeGen/MachineInstrBuilder.h" 22 #include "llvm/CodeGen/MachineOptimizationRemarkEmitter.h" 23 #include "llvm/CodeGen/MachineSizeOpts.h" 24 #include "llvm/CodeGen/MachineRegisterInfo.h" 25 #include "llvm/CodeGen/StackProtector.h" 26 #include "llvm/CodeGen/TargetInstrInfo.h" 27 #include "llvm/CodeGen/TargetLowering.h" 28 #include "llvm/CodeGen/TargetPassConfig.h" 29 #include "llvm/CodeGen/TargetRegisterInfo.h" 30 #include "llvm/IR/Constants.h" 31 #include "llvm/Target/TargetMachine.h" 32 33 #define DEBUG_TYPE "globalisel-utils" 34 35 using namespace llvm; 36 using namespace MIPatternMatch; 37 38 Register llvm::constrainRegToClass(MachineRegisterInfo &MRI, 39 const TargetInstrInfo &TII, 40 const RegisterBankInfo &RBI, Register Reg, 41 const TargetRegisterClass &RegClass) { 42 if (!RBI.constrainGenericRegister(Reg, RegClass, MRI)) 43 return MRI.createVirtualRegister(&RegClass); 44 45 return Reg; 46 } 47 48 Register llvm::constrainOperandRegClass( 49 const MachineFunction &MF, const TargetRegisterInfo &TRI, 50 MachineRegisterInfo &MRI, const TargetInstrInfo &TII, 51 const RegisterBankInfo &RBI, MachineInstr &InsertPt, 52 const TargetRegisterClass &RegClass, MachineOperand &RegMO) { 53 Register Reg = RegMO.getReg(); 54 // Assume physical registers are properly constrained. 55 assert(Register::isVirtualRegister(Reg) && "PhysReg not implemented"); 56 57 Register ConstrainedReg = constrainRegToClass(MRI, TII, RBI, Reg, RegClass); 58 // If we created a new virtual register because the class is not compatible 59 // then create a copy between the new and the old register. 60 if (ConstrainedReg != Reg) { 61 MachineBasicBlock::iterator InsertIt(&InsertPt); 62 MachineBasicBlock &MBB = *InsertPt.getParent(); 63 if (RegMO.isUse()) { 64 BuildMI(MBB, InsertIt, InsertPt.getDebugLoc(), 65 TII.get(TargetOpcode::COPY), ConstrainedReg) 66 .addReg(Reg); 67 } else { 68 assert(RegMO.isDef() && "Must be a definition"); 69 BuildMI(MBB, std::next(InsertIt), InsertPt.getDebugLoc(), 70 TII.get(TargetOpcode::COPY), Reg) 71 .addReg(ConstrainedReg); 72 } 73 if (GISelChangeObserver *Observer = MF.getObserver()) { 74 Observer->changingInstr(*RegMO.getParent()); 75 } 76 RegMO.setReg(ConstrainedReg); 77 if (GISelChangeObserver *Observer = MF.getObserver()) { 78 Observer->changedInstr(*RegMO.getParent()); 79 } 80 } else { 81 if (GISelChangeObserver *Observer = MF.getObserver()) { 82 if (!RegMO.isDef()) { 83 MachineInstr *RegDef = MRI.getVRegDef(Reg); 84 Observer->changedInstr(*RegDef); 85 } 86 Observer->changingAllUsesOfReg(MRI, Reg); 87 Observer->finishedChangingAllUsesOfReg(); 88 } 89 } 90 return ConstrainedReg; 91 } 92 93 Register llvm::constrainOperandRegClass( 94 const MachineFunction &MF, const TargetRegisterInfo &TRI, 95 MachineRegisterInfo &MRI, const TargetInstrInfo &TII, 96 const RegisterBankInfo &RBI, MachineInstr &InsertPt, const MCInstrDesc &II, 97 MachineOperand &RegMO, unsigned OpIdx) { 98 Register Reg = RegMO.getReg(); 99 // Assume physical registers are properly constrained. 100 assert(Register::isVirtualRegister(Reg) && "PhysReg not implemented"); 101 102 const TargetRegisterClass *RegClass = TII.getRegClass(II, OpIdx, &TRI, MF); 103 // Some of the target independent instructions, like COPY, may not impose any 104 // register class constraints on some of their operands: If it's a use, we can 105 // skip constraining as the instruction defining the register would constrain 106 // it. 107 108 // We can't constrain unallocatable register classes, because we can't create 109 // virtual registers for these classes, so we need to let targets handled this 110 // case. 111 if (RegClass && !RegClass->isAllocatable()) 112 RegClass = TRI.getConstrainedRegClassForOperand(RegMO, MRI); 113 114 if (!RegClass) { 115 assert((!isTargetSpecificOpcode(II.getOpcode()) || RegMO.isUse()) && 116 "Register class constraint is required unless either the " 117 "instruction is target independent or the operand is a use"); 118 // FIXME: Just bailing out like this here could be not enough, unless we 119 // expect the users of this function to do the right thing for PHIs and 120 // COPY: 121 // v1 = COPY v0 122 // v2 = COPY v1 123 // v1 here may end up not being constrained at all. Please notice that to 124 // reproduce the issue we likely need a destination pattern of a selection 125 // rule producing such extra copies, not just an input GMIR with them as 126 // every existing target using selectImpl handles copies before calling it 127 // and they never reach this function. 128 return Reg; 129 } 130 return constrainOperandRegClass(MF, TRI, MRI, TII, RBI, InsertPt, *RegClass, 131 RegMO); 132 } 133 134 bool llvm::constrainSelectedInstRegOperands(MachineInstr &I, 135 const TargetInstrInfo &TII, 136 const TargetRegisterInfo &TRI, 137 const RegisterBankInfo &RBI) { 138 assert(!isPreISelGenericOpcode(I.getOpcode()) && 139 "A selected instruction is expected"); 140 MachineBasicBlock &MBB = *I.getParent(); 141 MachineFunction &MF = *MBB.getParent(); 142 MachineRegisterInfo &MRI = MF.getRegInfo(); 143 144 for (unsigned OpI = 0, OpE = I.getNumExplicitOperands(); OpI != OpE; ++OpI) { 145 MachineOperand &MO = I.getOperand(OpI); 146 147 // There's nothing to be done on non-register operands. 148 if (!MO.isReg()) 149 continue; 150 151 LLVM_DEBUG(dbgs() << "Converting operand: " << MO << '\n'); 152 assert(MO.isReg() && "Unsupported non-reg operand"); 153 154 Register Reg = MO.getReg(); 155 // Physical registers don't need to be constrained. 156 if (Register::isPhysicalRegister(Reg)) 157 continue; 158 159 // Register operands with a value of 0 (e.g. predicate operands) don't need 160 // to be constrained. 161 if (Reg == 0) 162 continue; 163 164 // If the operand is a vreg, we should constrain its regclass, and only 165 // insert COPYs if that's impossible. 166 // constrainOperandRegClass does that for us. 167 constrainOperandRegClass(MF, TRI, MRI, TII, RBI, I, I.getDesc(), MO, OpI); 168 169 // Tie uses to defs as indicated in MCInstrDesc if this hasn't already been 170 // done. 171 if (MO.isUse()) { 172 int DefIdx = I.getDesc().getOperandConstraint(OpI, MCOI::TIED_TO); 173 if (DefIdx != -1 && !I.isRegTiedToUseOperand(DefIdx)) 174 I.tieOperands(DefIdx, OpI); 175 } 176 } 177 return true; 178 } 179 180 bool llvm::canReplaceReg(Register DstReg, Register SrcReg, 181 MachineRegisterInfo &MRI) { 182 // Give up if either DstReg or SrcReg is a physical register. 183 if (DstReg.isPhysical() || SrcReg.isPhysical()) 184 return false; 185 // Give up if the types don't match. 186 if (MRI.getType(DstReg) != MRI.getType(SrcReg)) 187 return false; 188 // Replace if either DstReg has no constraints or the register 189 // constraints match. 190 return !MRI.getRegClassOrRegBank(DstReg) || 191 MRI.getRegClassOrRegBank(DstReg) == MRI.getRegClassOrRegBank(SrcReg); 192 } 193 194 bool llvm::isTriviallyDead(const MachineInstr &MI, 195 const MachineRegisterInfo &MRI) { 196 // FIXME: This logical is mostly duplicated with 197 // DeadMachineInstructionElim::isDead. Why is LOCAL_ESCAPE not considered in 198 // MachineInstr::isLabel? 199 200 // Don't delete frame allocation labels. 201 if (MI.getOpcode() == TargetOpcode::LOCAL_ESCAPE) 202 return false; 203 // LIFETIME markers should be preserved even if they seem dead. 204 if (MI.getOpcode() == TargetOpcode::LIFETIME_START || 205 MI.getOpcode() == TargetOpcode::LIFETIME_END) 206 return false; 207 208 // If we can move an instruction, we can remove it. Otherwise, it has 209 // a side-effect of some sort. 210 bool SawStore = false; 211 if (!MI.isSafeToMove(/*AA=*/nullptr, SawStore) && !MI.isPHI()) 212 return false; 213 214 // Instructions without side-effects are dead iff they only define dead vregs. 215 for (auto &MO : MI.operands()) { 216 if (!MO.isReg() || !MO.isDef()) 217 continue; 218 219 Register Reg = MO.getReg(); 220 if (Register::isPhysicalRegister(Reg) || !MRI.use_nodbg_empty(Reg)) 221 return false; 222 } 223 return true; 224 } 225 226 static void reportGISelDiagnostic(DiagnosticSeverity Severity, 227 MachineFunction &MF, 228 const TargetPassConfig &TPC, 229 MachineOptimizationRemarkEmitter &MORE, 230 MachineOptimizationRemarkMissed &R) { 231 bool IsFatal = Severity == DS_Error && 232 TPC.isGlobalISelAbortEnabled(); 233 // Print the function name explicitly if we don't have a debug location (which 234 // makes the diagnostic less useful) or if we're going to emit a raw error. 235 if (!R.getLocation().isValid() || IsFatal) 236 R << (" (in function: " + MF.getName() + ")").str(); 237 238 if (IsFatal) 239 report_fatal_error(R.getMsg()); 240 else 241 MORE.emit(R); 242 } 243 244 void llvm::reportGISelWarning(MachineFunction &MF, const TargetPassConfig &TPC, 245 MachineOptimizationRemarkEmitter &MORE, 246 MachineOptimizationRemarkMissed &R) { 247 reportGISelDiagnostic(DS_Warning, MF, TPC, MORE, R); 248 } 249 250 void llvm::reportGISelFailure(MachineFunction &MF, const TargetPassConfig &TPC, 251 MachineOptimizationRemarkEmitter &MORE, 252 MachineOptimizationRemarkMissed &R) { 253 MF.getProperties().set(MachineFunctionProperties::Property::FailedISel); 254 reportGISelDiagnostic(DS_Error, MF, TPC, MORE, R); 255 } 256 257 void llvm::reportGISelFailure(MachineFunction &MF, const TargetPassConfig &TPC, 258 MachineOptimizationRemarkEmitter &MORE, 259 const char *PassName, StringRef Msg, 260 const MachineInstr &MI) { 261 MachineOptimizationRemarkMissed R(PassName, "GISelFailure: ", 262 MI.getDebugLoc(), MI.getParent()); 263 R << Msg; 264 // Printing MI is expensive; only do it if expensive remarks are enabled. 265 if (TPC.isGlobalISelAbortEnabled() || MORE.allowExtraAnalysis(PassName)) 266 R << ": " << ore::MNV("Inst", MI); 267 reportGISelFailure(MF, TPC, MORE, R); 268 } 269 270 Optional<APInt> llvm::getConstantVRegVal(Register VReg, 271 const MachineRegisterInfo &MRI) { 272 Optional<ValueAndVReg> ValAndVReg = 273 getConstantVRegValWithLookThrough(VReg, MRI, /*LookThroughInstrs*/ false); 274 assert((!ValAndVReg || ValAndVReg->VReg == VReg) && 275 "Value found while looking through instrs"); 276 if (!ValAndVReg) 277 return None; 278 return ValAndVReg->Value; 279 } 280 281 Optional<int64_t> llvm::getConstantVRegSExtVal(Register VReg, 282 const MachineRegisterInfo &MRI) { 283 Optional<APInt> Val = getConstantVRegVal(VReg, MRI); 284 if (Val && Val->getBitWidth() <= 64) 285 return Val->getSExtValue(); 286 return None; 287 } 288 289 Optional<ValueAndVReg> llvm::getConstantVRegValWithLookThrough( 290 Register VReg, const MachineRegisterInfo &MRI, bool LookThroughInstrs, 291 bool HandleFConstant, bool LookThroughAnyExt) { 292 SmallVector<std::pair<unsigned, unsigned>, 4> SeenOpcodes; 293 MachineInstr *MI; 294 auto IsConstantOpcode = [HandleFConstant](unsigned Opcode) { 295 return Opcode == TargetOpcode::G_CONSTANT || 296 (HandleFConstant && Opcode == TargetOpcode::G_FCONSTANT); 297 }; 298 auto GetImmediateValue = [HandleFConstant, 299 &MRI](const MachineInstr &MI) -> Optional<APInt> { 300 const MachineOperand &CstVal = MI.getOperand(1); 301 if (!CstVal.isImm() && !CstVal.isCImm() && 302 (!HandleFConstant || !CstVal.isFPImm())) 303 return None; 304 if (!CstVal.isFPImm()) { 305 unsigned BitWidth = 306 MRI.getType(MI.getOperand(0).getReg()).getSizeInBits(); 307 APInt Val = CstVal.isImm() ? APInt(BitWidth, CstVal.getImm()) 308 : CstVal.getCImm()->getValue(); 309 assert(Val.getBitWidth() == BitWidth && 310 "Value bitwidth doesn't match definition type"); 311 return Val; 312 } 313 return CstVal.getFPImm()->getValueAPF().bitcastToAPInt(); 314 }; 315 while ((MI = MRI.getVRegDef(VReg)) && !IsConstantOpcode(MI->getOpcode()) && 316 LookThroughInstrs) { 317 switch (MI->getOpcode()) { 318 case TargetOpcode::G_ANYEXT: 319 if (!LookThroughAnyExt) 320 return None; 321 LLVM_FALLTHROUGH; 322 case TargetOpcode::G_TRUNC: 323 case TargetOpcode::G_SEXT: 324 case TargetOpcode::G_ZEXT: 325 SeenOpcodes.push_back(std::make_pair( 326 MI->getOpcode(), 327 MRI.getType(MI->getOperand(0).getReg()).getSizeInBits())); 328 VReg = MI->getOperand(1).getReg(); 329 break; 330 case TargetOpcode::COPY: 331 VReg = MI->getOperand(1).getReg(); 332 if (Register::isPhysicalRegister(VReg)) 333 return None; 334 break; 335 case TargetOpcode::G_INTTOPTR: 336 VReg = MI->getOperand(1).getReg(); 337 break; 338 default: 339 return None; 340 } 341 } 342 if (!MI || !IsConstantOpcode(MI->getOpcode())) 343 return None; 344 345 Optional<APInt> MaybeVal = GetImmediateValue(*MI); 346 if (!MaybeVal) 347 return None; 348 APInt &Val = *MaybeVal; 349 while (!SeenOpcodes.empty()) { 350 std::pair<unsigned, unsigned> OpcodeAndSize = SeenOpcodes.pop_back_val(); 351 switch (OpcodeAndSize.first) { 352 case TargetOpcode::G_TRUNC: 353 Val = Val.trunc(OpcodeAndSize.second); 354 break; 355 case TargetOpcode::G_ANYEXT: 356 case TargetOpcode::G_SEXT: 357 Val = Val.sext(OpcodeAndSize.second); 358 break; 359 case TargetOpcode::G_ZEXT: 360 Val = Val.zext(OpcodeAndSize.second); 361 break; 362 } 363 } 364 365 return ValueAndVReg{Val, VReg}; 366 } 367 368 const ConstantFP * 369 llvm::getConstantFPVRegVal(Register VReg, const MachineRegisterInfo &MRI) { 370 MachineInstr *MI = MRI.getVRegDef(VReg); 371 if (TargetOpcode::G_FCONSTANT != MI->getOpcode()) 372 return nullptr; 373 return MI->getOperand(1).getFPImm(); 374 } 375 376 Optional<DefinitionAndSourceRegister> 377 llvm::getDefSrcRegIgnoringCopies(Register Reg, const MachineRegisterInfo &MRI) { 378 Register DefSrcReg = Reg; 379 auto *DefMI = MRI.getVRegDef(Reg); 380 auto DstTy = MRI.getType(DefMI->getOperand(0).getReg()); 381 if (!DstTy.isValid()) 382 return None; 383 unsigned Opc = DefMI->getOpcode(); 384 while (Opc == TargetOpcode::COPY || isPreISelGenericOptimizationHint(Opc)) { 385 Register SrcReg = DefMI->getOperand(1).getReg(); 386 auto SrcTy = MRI.getType(SrcReg); 387 if (!SrcTy.isValid()) 388 break; 389 DefMI = MRI.getVRegDef(SrcReg); 390 DefSrcReg = SrcReg; 391 Opc = DefMI->getOpcode(); 392 } 393 return DefinitionAndSourceRegister{DefMI, DefSrcReg}; 394 } 395 396 MachineInstr *llvm::getDefIgnoringCopies(Register Reg, 397 const MachineRegisterInfo &MRI) { 398 Optional<DefinitionAndSourceRegister> DefSrcReg = 399 getDefSrcRegIgnoringCopies(Reg, MRI); 400 return DefSrcReg ? DefSrcReg->MI : nullptr; 401 } 402 403 Register llvm::getSrcRegIgnoringCopies(Register Reg, 404 const MachineRegisterInfo &MRI) { 405 Optional<DefinitionAndSourceRegister> DefSrcReg = 406 getDefSrcRegIgnoringCopies(Reg, MRI); 407 return DefSrcReg ? DefSrcReg->Reg : Register(); 408 } 409 410 MachineInstr *llvm::getOpcodeDef(unsigned Opcode, Register Reg, 411 const MachineRegisterInfo &MRI) { 412 MachineInstr *DefMI = getDefIgnoringCopies(Reg, MRI); 413 return DefMI && DefMI->getOpcode() == Opcode ? DefMI : nullptr; 414 } 415 416 APFloat llvm::getAPFloatFromSize(double Val, unsigned Size) { 417 if (Size == 32) 418 return APFloat(float(Val)); 419 if (Size == 64) 420 return APFloat(Val); 421 if (Size != 16) 422 llvm_unreachable("Unsupported FPConstant size"); 423 bool Ignored; 424 APFloat APF(Val); 425 APF.convert(APFloat::IEEEhalf(), APFloat::rmNearestTiesToEven, &Ignored); 426 return APF; 427 } 428 429 Optional<APInt> llvm::ConstantFoldBinOp(unsigned Opcode, const Register Op1, 430 const Register Op2, 431 const MachineRegisterInfo &MRI) { 432 auto MaybeOp2Cst = getConstantVRegVal(Op2, MRI); 433 if (!MaybeOp2Cst) 434 return None; 435 436 auto MaybeOp1Cst = getConstantVRegVal(Op1, MRI); 437 if (!MaybeOp1Cst) 438 return None; 439 440 const APInt &C1 = *MaybeOp1Cst; 441 const APInt &C2 = *MaybeOp2Cst; 442 switch (Opcode) { 443 default: 444 break; 445 case TargetOpcode::G_ADD: 446 return C1 + C2; 447 case TargetOpcode::G_AND: 448 return C1 & C2; 449 case TargetOpcode::G_ASHR: 450 return C1.ashr(C2); 451 case TargetOpcode::G_LSHR: 452 return C1.lshr(C2); 453 case TargetOpcode::G_MUL: 454 return C1 * C2; 455 case TargetOpcode::G_OR: 456 return C1 | C2; 457 case TargetOpcode::G_SHL: 458 return C1 << C2; 459 case TargetOpcode::G_SUB: 460 return C1 - C2; 461 case TargetOpcode::G_XOR: 462 return C1 ^ C2; 463 case TargetOpcode::G_UDIV: 464 if (!C2.getBoolValue()) 465 break; 466 return C1.udiv(C2); 467 case TargetOpcode::G_SDIV: 468 if (!C2.getBoolValue()) 469 break; 470 return C1.sdiv(C2); 471 case TargetOpcode::G_UREM: 472 if (!C2.getBoolValue()) 473 break; 474 return C1.urem(C2); 475 case TargetOpcode::G_SREM: 476 if (!C2.getBoolValue()) 477 break; 478 return C1.srem(C2); 479 } 480 481 return None; 482 } 483 484 Optional<APFloat> llvm::ConstantFoldFPBinOp(unsigned Opcode, const Register Op1, 485 const Register Op2, 486 const MachineRegisterInfo &MRI) { 487 const ConstantFP *Op2Cst = getConstantFPVRegVal(Op2, MRI); 488 if (!Op2Cst) 489 return None; 490 491 const ConstantFP *Op1Cst = getConstantFPVRegVal(Op1, MRI); 492 if (!Op1Cst) 493 return None; 494 495 APFloat C1 = Op1Cst->getValueAPF(); 496 const APFloat &C2 = Op2Cst->getValueAPF(); 497 switch (Opcode) { 498 case TargetOpcode::G_FADD: 499 C1.add(C2, APFloat::rmNearestTiesToEven); 500 return C1; 501 case TargetOpcode::G_FSUB: 502 C1.subtract(C2, APFloat::rmNearestTiesToEven); 503 return C1; 504 case TargetOpcode::G_FMUL: 505 C1.multiply(C2, APFloat::rmNearestTiesToEven); 506 return C1; 507 case TargetOpcode::G_FDIV: 508 C1.divide(C2, APFloat::rmNearestTiesToEven); 509 return C1; 510 case TargetOpcode::G_FREM: 511 C1.mod(C2); 512 return C1; 513 case TargetOpcode::G_FCOPYSIGN: 514 C1.copySign(C2); 515 return C1; 516 case TargetOpcode::G_FMINNUM: 517 return minnum(C1, C2); 518 case TargetOpcode::G_FMAXNUM: 519 return maxnum(C1, C2); 520 case TargetOpcode::G_FMINIMUM: 521 return minimum(C1, C2); 522 case TargetOpcode::G_FMAXIMUM: 523 return maximum(C1, C2); 524 case TargetOpcode::G_FMINNUM_IEEE: 525 case TargetOpcode::G_FMAXNUM_IEEE: 526 // FIXME: These operations were unfortunately named. fminnum/fmaxnum do not 527 // follow the IEEE behavior for signaling nans and follow libm's fmin/fmax, 528 // and currently there isn't a nice wrapper in APFloat for the version with 529 // correct snan handling. 530 break; 531 default: 532 break; 533 } 534 535 return None; 536 } 537 538 bool llvm::isKnownNeverNaN(Register Val, const MachineRegisterInfo &MRI, 539 bool SNaN) { 540 const MachineInstr *DefMI = MRI.getVRegDef(Val); 541 if (!DefMI) 542 return false; 543 544 const TargetMachine& TM = DefMI->getMF()->getTarget(); 545 if (DefMI->getFlag(MachineInstr::FmNoNans) || TM.Options.NoNaNsFPMath) 546 return true; 547 548 // If the value is a constant, we can obviously see if it is a NaN or not. 549 if (const ConstantFP *FPVal = getConstantFPVRegVal(Val, MRI)) { 550 return !FPVal->getValueAPF().isNaN() || 551 (SNaN && !FPVal->getValueAPF().isSignaling()); 552 } 553 554 if (DefMI->getOpcode() == TargetOpcode::G_BUILD_VECTOR) { 555 for (const auto &Op : DefMI->uses()) 556 if (!isKnownNeverNaN(Op.getReg(), MRI, SNaN)) 557 return false; 558 return true; 559 } 560 561 switch (DefMI->getOpcode()) { 562 default: 563 break; 564 case TargetOpcode::G_FMINNUM_IEEE: 565 case TargetOpcode::G_FMAXNUM_IEEE: { 566 if (SNaN) 567 return true; 568 // This can return a NaN if either operand is an sNaN, or if both operands 569 // are NaN. 570 return (isKnownNeverNaN(DefMI->getOperand(1).getReg(), MRI) && 571 isKnownNeverSNaN(DefMI->getOperand(2).getReg(), MRI)) || 572 (isKnownNeverSNaN(DefMI->getOperand(1).getReg(), MRI) && 573 isKnownNeverNaN(DefMI->getOperand(2).getReg(), MRI)); 574 } 575 case TargetOpcode::G_FMINNUM: 576 case TargetOpcode::G_FMAXNUM: { 577 // Only one needs to be known not-nan, since it will be returned if the 578 // other ends up being one. 579 return isKnownNeverNaN(DefMI->getOperand(1).getReg(), MRI, SNaN) || 580 isKnownNeverNaN(DefMI->getOperand(2).getReg(), MRI, SNaN); 581 } 582 } 583 584 if (SNaN) { 585 // FP operations quiet. For now, just handle the ones inserted during 586 // legalization. 587 switch (DefMI->getOpcode()) { 588 case TargetOpcode::G_FPEXT: 589 case TargetOpcode::G_FPTRUNC: 590 case TargetOpcode::G_FCANONICALIZE: 591 return true; 592 default: 593 return false; 594 } 595 } 596 597 return false; 598 } 599 600 Align llvm::inferAlignFromPtrInfo(MachineFunction &MF, 601 const MachinePointerInfo &MPO) { 602 auto PSV = MPO.V.dyn_cast<const PseudoSourceValue *>(); 603 if (auto FSPV = dyn_cast_or_null<FixedStackPseudoSourceValue>(PSV)) { 604 MachineFrameInfo &MFI = MF.getFrameInfo(); 605 return commonAlignment(MFI.getObjectAlign(FSPV->getFrameIndex()), 606 MPO.Offset); 607 } 608 609 if (const Value *V = MPO.V.dyn_cast<const Value *>()) { 610 const Module *M = MF.getFunction().getParent(); 611 return V->getPointerAlignment(M->getDataLayout()); 612 } 613 614 return Align(1); 615 } 616 617 Register llvm::getFunctionLiveInPhysReg(MachineFunction &MF, 618 const TargetInstrInfo &TII, 619 MCRegister PhysReg, 620 const TargetRegisterClass &RC, 621 LLT RegTy) { 622 DebugLoc DL; // FIXME: Is no location the right choice? 623 MachineBasicBlock &EntryMBB = MF.front(); 624 MachineRegisterInfo &MRI = MF.getRegInfo(); 625 Register LiveIn = MRI.getLiveInVirtReg(PhysReg); 626 if (LiveIn) { 627 MachineInstr *Def = MRI.getVRegDef(LiveIn); 628 if (Def) { 629 // FIXME: Should the verifier check this is in the entry block? 630 assert(Def->getParent() == &EntryMBB && "live-in copy not in entry block"); 631 return LiveIn; 632 } 633 634 // It's possible the incoming argument register and copy was added during 635 // lowering, but later deleted due to being/becoming dead. If this happens, 636 // re-insert the copy. 637 } else { 638 // The live in register was not present, so add it. 639 LiveIn = MF.addLiveIn(PhysReg, &RC); 640 if (RegTy.isValid()) 641 MRI.setType(LiveIn, RegTy); 642 } 643 644 BuildMI(EntryMBB, EntryMBB.begin(), DL, TII.get(TargetOpcode::COPY), LiveIn) 645 .addReg(PhysReg); 646 if (!EntryMBB.isLiveIn(PhysReg)) 647 EntryMBB.addLiveIn(PhysReg); 648 return LiveIn; 649 } 650 651 Optional<APInt> llvm::ConstantFoldExtOp(unsigned Opcode, const Register Op1, 652 uint64_t Imm, 653 const MachineRegisterInfo &MRI) { 654 auto MaybeOp1Cst = getConstantVRegVal(Op1, MRI); 655 if (MaybeOp1Cst) { 656 switch (Opcode) { 657 default: 658 break; 659 case TargetOpcode::G_SEXT_INREG: { 660 LLT Ty = MRI.getType(Op1); 661 return MaybeOp1Cst->trunc(Imm).sext(Ty.getScalarSizeInBits()); 662 } 663 } 664 } 665 return None; 666 } 667 668 bool llvm::isKnownToBeAPowerOfTwo(Register Reg, const MachineRegisterInfo &MRI, 669 GISelKnownBits *KB) { 670 Optional<DefinitionAndSourceRegister> DefSrcReg = 671 getDefSrcRegIgnoringCopies(Reg, MRI); 672 if (!DefSrcReg) 673 return false; 674 675 const MachineInstr &MI = *DefSrcReg->MI; 676 const LLT Ty = MRI.getType(Reg); 677 678 switch (MI.getOpcode()) { 679 case TargetOpcode::G_CONSTANT: { 680 unsigned BitWidth = Ty.getScalarSizeInBits(); 681 const ConstantInt *CI = MI.getOperand(1).getCImm(); 682 return CI->getValue().zextOrTrunc(BitWidth).isPowerOf2(); 683 } 684 case TargetOpcode::G_SHL: { 685 // A left-shift of a constant one will have exactly one bit set because 686 // shifting the bit off the end is undefined. 687 688 // TODO: Constant splat 689 if (auto ConstLHS = getConstantVRegVal(MI.getOperand(1).getReg(), MRI)) { 690 if (*ConstLHS == 1) 691 return true; 692 } 693 694 break; 695 } 696 case TargetOpcode::G_LSHR: { 697 if (auto ConstLHS = getConstantVRegVal(MI.getOperand(1).getReg(), MRI)) { 698 if (ConstLHS->isSignMask()) 699 return true; 700 } 701 702 break; 703 } 704 case TargetOpcode::G_BUILD_VECTOR: { 705 // TODO: Probably should have a recursion depth guard since you could have 706 // bitcasted vector elements. 707 for (unsigned I = 1, E = MI.getNumOperands(); I != E; ++I) { 708 if (!isKnownToBeAPowerOfTwo(MI.getOperand(I).getReg(), MRI, KB)) 709 return false; 710 } 711 712 return true; 713 } 714 case TargetOpcode::G_BUILD_VECTOR_TRUNC: { 715 // Only handle constants since we would need to know if number of leading 716 // zeros is greater than the truncation amount. 717 const unsigned BitWidth = Ty.getScalarSizeInBits(); 718 for (unsigned I = 1, E = MI.getNumOperands(); I != E; ++I) { 719 auto Const = getConstantVRegVal(MI.getOperand(I).getReg(), MRI); 720 if (!Const || !Const->zextOrTrunc(BitWidth).isPowerOf2()) 721 return false; 722 } 723 724 return true; 725 } 726 default: 727 break; 728 } 729 730 if (!KB) 731 return false; 732 733 // More could be done here, though the above checks are enough 734 // to handle some common cases. 735 736 // Fall back to computeKnownBits to catch other known cases. 737 KnownBits Known = KB->getKnownBits(Reg); 738 return (Known.countMaxPopulation() == 1) && (Known.countMinPopulation() == 1); 739 } 740 741 void llvm::getSelectionDAGFallbackAnalysisUsage(AnalysisUsage &AU) { 742 AU.addPreserved<StackProtector>(); 743 } 744 745 static unsigned getLCMSize(unsigned OrigSize, unsigned TargetSize) { 746 unsigned Mul = OrigSize * TargetSize; 747 unsigned GCDSize = greatestCommonDivisor(OrigSize, TargetSize); 748 return Mul / GCDSize; 749 } 750 751 LLT llvm::getLCMType(LLT OrigTy, LLT TargetTy) { 752 const unsigned OrigSize = OrigTy.getSizeInBits(); 753 const unsigned TargetSize = TargetTy.getSizeInBits(); 754 755 if (OrigSize == TargetSize) 756 return OrigTy; 757 758 if (OrigTy.isVector()) { 759 const LLT OrigElt = OrigTy.getElementType(); 760 761 if (TargetTy.isVector()) { 762 const LLT TargetElt = TargetTy.getElementType(); 763 764 if (OrigElt.getSizeInBits() == TargetElt.getSizeInBits()) { 765 int GCDElts = greatestCommonDivisor(OrigTy.getNumElements(), 766 TargetTy.getNumElements()); 767 // Prefer the original element type. 768 int Mul = OrigTy.getNumElements() * TargetTy.getNumElements(); 769 return LLT::vector(Mul / GCDElts, OrigTy.getElementType()); 770 } 771 } else { 772 if (OrigElt.getSizeInBits() == TargetSize) 773 return OrigTy; 774 } 775 776 unsigned LCMSize = getLCMSize(OrigSize, TargetSize); 777 return LLT::vector(LCMSize / OrigElt.getSizeInBits(), OrigElt); 778 } 779 780 if (TargetTy.isVector()) { 781 unsigned LCMSize = getLCMSize(OrigSize, TargetSize); 782 return LLT::vector(LCMSize / OrigSize, OrigTy); 783 } 784 785 unsigned LCMSize = getLCMSize(OrigSize, TargetSize); 786 787 // Preserve pointer types. 788 if (LCMSize == OrigSize) 789 return OrigTy; 790 if (LCMSize == TargetSize) 791 return TargetTy; 792 793 return LLT::scalar(LCMSize); 794 } 795 796 LLT llvm::getGCDType(LLT OrigTy, LLT TargetTy) { 797 const unsigned OrigSize = OrigTy.getSizeInBits(); 798 const unsigned TargetSize = TargetTy.getSizeInBits(); 799 800 if (OrigSize == TargetSize) 801 return OrigTy; 802 803 if (OrigTy.isVector()) { 804 LLT OrigElt = OrigTy.getElementType(); 805 if (TargetTy.isVector()) { 806 LLT TargetElt = TargetTy.getElementType(); 807 if (OrigElt.getSizeInBits() == TargetElt.getSizeInBits()) { 808 int GCD = greatestCommonDivisor(OrigTy.getNumElements(), 809 TargetTy.getNumElements()); 810 return LLT::scalarOrVector(GCD, OrigElt); 811 } 812 } else { 813 // If the source is a vector of pointers, return a pointer element. 814 if (OrigElt.getSizeInBits() == TargetSize) 815 return OrigElt; 816 } 817 818 unsigned GCD = greatestCommonDivisor(OrigSize, TargetSize); 819 if (GCD == OrigElt.getSizeInBits()) 820 return OrigElt; 821 822 // If we can't produce the original element type, we have to use a smaller 823 // scalar. 824 if (GCD < OrigElt.getSizeInBits()) 825 return LLT::scalar(GCD); 826 return LLT::vector(GCD / OrigElt.getSizeInBits(), OrigElt); 827 } 828 829 if (TargetTy.isVector()) { 830 // Try to preserve the original element type. 831 LLT TargetElt = TargetTy.getElementType(); 832 if (TargetElt.getSizeInBits() == OrigSize) 833 return OrigTy; 834 } 835 836 unsigned GCD = greatestCommonDivisor(OrigSize, TargetSize); 837 return LLT::scalar(GCD); 838 } 839 840 Optional<int> llvm::getSplatIndex(MachineInstr &MI) { 841 assert(MI.getOpcode() == TargetOpcode::G_SHUFFLE_VECTOR && 842 "Only G_SHUFFLE_VECTOR can have a splat index!"); 843 ArrayRef<int> Mask = MI.getOperand(3).getShuffleMask(); 844 auto FirstDefinedIdx = find_if(Mask, [](int Elt) { return Elt >= 0; }); 845 846 // If all elements are undefined, this shuffle can be considered a splat. 847 // Return 0 for better potential for callers to simplify. 848 if (FirstDefinedIdx == Mask.end()) 849 return 0; 850 851 // Make sure all remaining elements are either undef or the same 852 // as the first non-undef value. 853 int SplatValue = *FirstDefinedIdx; 854 if (any_of(make_range(std::next(FirstDefinedIdx), Mask.end()), 855 [&SplatValue](int Elt) { return Elt >= 0 && Elt != SplatValue; })) 856 return None; 857 858 return SplatValue; 859 } 860 861 static bool isBuildVectorOp(unsigned Opcode) { 862 return Opcode == TargetOpcode::G_BUILD_VECTOR || 863 Opcode == TargetOpcode::G_BUILD_VECTOR_TRUNC; 864 } 865 866 // TODO: Handle mixed undef elements. 867 static bool isBuildVectorConstantSplat(const MachineInstr &MI, 868 const MachineRegisterInfo &MRI, 869 int64_t SplatValue) { 870 if (!isBuildVectorOp(MI.getOpcode())) 871 return false; 872 873 const unsigned NumOps = MI.getNumOperands(); 874 for (unsigned I = 1; I != NumOps; ++I) { 875 Register Element = MI.getOperand(I).getReg(); 876 if (!mi_match(Element, MRI, m_SpecificICst(SplatValue))) 877 return false; 878 } 879 880 return true; 881 } 882 883 Optional<int64_t> 884 llvm::getBuildVectorConstantSplat(const MachineInstr &MI, 885 const MachineRegisterInfo &MRI) { 886 if (!isBuildVectorOp(MI.getOpcode())) 887 return None; 888 889 const unsigned NumOps = MI.getNumOperands(); 890 Optional<int64_t> Scalar; 891 for (unsigned I = 1; I != NumOps; ++I) { 892 Register Element = MI.getOperand(I).getReg(); 893 int64_t ElementValue; 894 if (!mi_match(Element, MRI, m_ICst(ElementValue))) 895 return None; 896 if (!Scalar) 897 Scalar = ElementValue; 898 else if (*Scalar != ElementValue) 899 return None; 900 } 901 902 return Scalar; 903 } 904 905 bool llvm::isBuildVectorAllZeros(const MachineInstr &MI, 906 const MachineRegisterInfo &MRI) { 907 return isBuildVectorConstantSplat(MI, MRI, 0); 908 } 909 910 bool llvm::isBuildVectorAllOnes(const MachineInstr &MI, 911 const MachineRegisterInfo &MRI) { 912 return isBuildVectorConstantSplat(MI, MRI, -1); 913 } 914 915 Optional<RegOrConstant> llvm::getVectorSplat(const MachineInstr &MI, 916 const MachineRegisterInfo &MRI) { 917 unsigned Opc = MI.getOpcode(); 918 if (!isBuildVectorOp(Opc)) 919 return None; 920 if (auto Splat = getBuildVectorConstantSplat(MI, MRI)) 921 return RegOrConstant(*Splat); 922 auto Reg = MI.getOperand(1).getReg(); 923 if (any_of(make_range(MI.operands_begin() + 2, MI.operands_end()), 924 [&Reg](const MachineOperand &Op) { return Op.getReg() != Reg; })) 925 return None; 926 return RegOrConstant(Reg); 927 } 928 929 bool llvm::matchUnaryPredicate( 930 const MachineRegisterInfo &MRI, Register Reg, 931 std::function<bool(const Constant *ConstVal)> Match, bool AllowUndefs) { 932 933 const MachineInstr *Def = getDefIgnoringCopies(Reg, MRI); 934 if (AllowUndefs && Def->getOpcode() == TargetOpcode::G_IMPLICIT_DEF) 935 return Match(nullptr); 936 937 // TODO: Also handle fconstant 938 if (Def->getOpcode() == TargetOpcode::G_CONSTANT) 939 return Match(Def->getOperand(1).getCImm()); 940 941 if (Def->getOpcode() != TargetOpcode::G_BUILD_VECTOR) 942 return false; 943 944 for (unsigned I = 1, E = Def->getNumOperands(); I != E; ++I) { 945 Register SrcElt = Def->getOperand(I).getReg(); 946 const MachineInstr *SrcDef = getDefIgnoringCopies(SrcElt, MRI); 947 if (AllowUndefs && SrcDef->getOpcode() == TargetOpcode::G_IMPLICIT_DEF) { 948 if (!Match(nullptr)) 949 return false; 950 continue; 951 } 952 953 if (SrcDef->getOpcode() != TargetOpcode::G_CONSTANT || 954 !Match(SrcDef->getOperand(1).getCImm())) 955 return false; 956 } 957 958 return true; 959 } 960 961 bool llvm::isConstTrueVal(const TargetLowering &TLI, int64_t Val, bool IsVector, 962 bool IsFP) { 963 switch (TLI.getBooleanContents(IsVector, IsFP)) { 964 case TargetLowering::UndefinedBooleanContent: 965 return Val & 0x1; 966 case TargetLowering::ZeroOrOneBooleanContent: 967 return Val == 1; 968 case TargetLowering::ZeroOrNegativeOneBooleanContent: 969 return Val == -1; 970 } 971 llvm_unreachable("Invalid boolean contents"); 972 } 973 974 int64_t llvm::getICmpTrueVal(const TargetLowering &TLI, bool IsVector, 975 bool IsFP) { 976 switch (TLI.getBooleanContents(IsVector, IsFP)) { 977 case TargetLowering::UndefinedBooleanContent: 978 case TargetLowering::ZeroOrOneBooleanContent: 979 return 1; 980 case TargetLowering::ZeroOrNegativeOneBooleanContent: 981 return -1; 982 } 983 llvm_unreachable("Invalid boolean contents"); 984 } 985 986 bool llvm::shouldOptForSize(const MachineBasicBlock &MBB, 987 ProfileSummaryInfo *PSI, BlockFrequencyInfo *BFI) { 988 const auto &F = MBB.getParent()->getFunction(); 989 return F.hasOptSize() || F.hasMinSize() || 990 llvm::shouldOptimizeForSize(MBB.getBasicBlock(), PSI, BFI); 991 } 992 993 unsigned llvm::getIntrinsicID(const MachineInstr &MI) { 994 #ifndef NDEBUG 995 unsigned Opc = MI.getOpcode(); 996 assert(Opc == TargetOpcode::G_INTRINSIC || 997 Opc == TargetOpcode::G_INTRINSIC_W_SIDE_EFFECTS); 998 #endif 999 return MI.getOperand(MI.getNumExplicitDefs()).getIntrinsicID(); 1000 } 1001