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/Analysis/ValueTracking.h" 16 #include "llvm/CodeGen/CodeGenCommonISel.h" 17 #include "llvm/CodeGen/GlobalISel/GISelChangeObserver.h" 18 #include "llvm/CodeGen/GlobalISel/GISelKnownBits.h" 19 #include "llvm/CodeGen/GlobalISel/GenericMachineInstrs.h" 20 #include "llvm/CodeGen/GlobalISel/LostDebugLocObserver.h" 21 #include "llvm/CodeGen/GlobalISel/MIPatternMatch.h" 22 #include "llvm/CodeGen/GlobalISel/MachineIRBuilder.h" 23 #include "llvm/CodeGen/MachineInstr.h" 24 #include "llvm/CodeGen/MachineInstrBuilder.h" 25 #include "llvm/CodeGen/MachineOptimizationRemarkEmitter.h" 26 #include "llvm/CodeGen/MachineRegisterInfo.h" 27 #include "llvm/CodeGen/MachineSizeOpts.h" 28 #include "llvm/CodeGen/RegisterBankInfo.h" 29 #include "llvm/CodeGen/StackProtector.h" 30 #include "llvm/CodeGen/TargetInstrInfo.h" 31 #include "llvm/CodeGen/TargetLowering.h" 32 #include "llvm/CodeGen/TargetOpcodes.h" 33 #include "llvm/CodeGen/TargetPassConfig.h" 34 #include "llvm/CodeGen/TargetRegisterInfo.h" 35 #include "llvm/IR/Constants.h" 36 #include "llvm/Target/TargetMachine.h" 37 #include "llvm/Transforms/Utils/SizeOpts.h" 38 #include <numeric> 39 #include <optional> 40 41 #define DEBUG_TYPE "globalisel-utils" 42 43 using namespace llvm; 44 using namespace MIPatternMatch; 45 46 Register llvm::constrainRegToClass(MachineRegisterInfo &MRI, 47 const TargetInstrInfo &TII, 48 const RegisterBankInfo &RBI, Register Reg, 49 const TargetRegisterClass &RegClass) { 50 if (!RBI.constrainGenericRegister(Reg, RegClass, MRI)) 51 return MRI.createVirtualRegister(&RegClass); 52 53 return Reg; 54 } 55 56 Register llvm::constrainOperandRegClass( 57 const MachineFunction &MF, const TargetRegisterInfo &TRI, 58 MachineRegisterInfo &MRI, const TargetInstrInfo &TII, 59 const RegisterBankInfo &RBI, MachineInstr &InsertPt, 60 const TargetRegisterClass &RegClass, MachineOperand &RegMO) { 61 Register Reg = RegMO.getReg(); 62 // Assume physical registers are properly constrained. 63 assert(Reg.isVirtual() && "PhysReg not implemented"); 64 65 // Save the old register class to check whether 66 // the change notifications will be required. 67 // TODO: A better approach would be to pass 68 // the observers to constrainRegToClass(). 69 auto *OldRegClass = MRI.getRegClassOrNull(Reg); 70 Register ConstrainedReg = constrainRegToClass(MRI, TII, RBI, Reg, RegClass); 71 // If we created a new virtual register because the class is not compatible 72 // then create a copy between the new and the old register. 73 if (ConstrainedReg != Reg) { 74 MachineBasicBlock::iterator InsertIt(&InsertPt); 75 MachineBasicBlock &MBB = *InsertPt.getParent(); 76 // FIXME: The copy needs to have the classes constrained for its operands. 77 // Use operand's regbank to get the class for old register (Reg). 78 if (RegMO.isUse()) { 79 BuildMI(MBB, InsertIt, InsertPt.getDebugLoc(), 80 TII.get(TargetOpcode::COPY), ConstrainedReg) 81 .addReg(Reg); 82 } else { 83 assert(RegMO.isDef() && "Must be a definition"); 84 BuildMI(MBB, std::next(InsertIt), InsertPt.getDebugLoc(), 85 TII.get(TargetOpcode::COPY), Reg) 86 .addReg(ConstrainedReg); 87 } 88 if (GISelChangeObserver *Observer = MF.getObserver()) { 89 Observer->changingInstr(*RegMO.getParent()); 90 } 91 RegMO.setReg(ConstrainedReg); 92 if (GISelChangeObserver *Observer = MF.getObserver()) { 93 Observer->changedInstr(*RegMO.getParent()); 94 } 95 } else if (OldRegClass != MRI.getRegClassOrNull(Reg)) { 96 if (GISelChangeObserver *Observer = MF.getObserver()) { 97 if (!RegMO.isDef()) { 98 MachineInstr *RegDef = MRI.getVRegDef(Reg); 99 Observer->changedInstr(*RegDef); 100 } 101 Observer->changingAllUsesOfReg(MRI, Reg); 102 Observer->finishedChangingAllUsesOfReg(); 103 } 104 } 105 return ConstrainedReg; 106 } 107 108 Register llvm::constrainOperandRegClass( 109 const MachineFunction &MF, const TargetRegisterInfo &TRI, 110 MachineRegisterInfo &MRI, const TargetInstrInfo &TII, 111 const RegisterBankInfo &RBI, MachineInstr &InsertPt, const MCInstrDesc &II, 112 MachineOperand &RegMO, unsigned OpIdx) { 113 Register Reg = RegMO.getReg(); 114 // Assume physical registers are properly constrained. 115 assert(Reg.isVirtual() && "PhysReg not implemented"); 116 117 const TargetRegisterClass *OpRC = TII.getRegClass(II, OpIdx, &TRI, MF); 118 // Some of the target independent instructions, like COPY, may not impose any 119 // register class constraints on some of their operands: If it's a use, we can 120 // skip constraining as the instruction defining the register would constrain 121 // it. 122 123 if (OpRC) { 124 // Obtain the RC from incoming regbank if it is a proper sub-class. Operands 125 // can have multiple regbanks for a superclass that combine different 126 // register types (E.g., AMDGPU's VGPR and AGPR). The regbank ambiguity 127 // resolved by targets during regbankselect should not be overridden. 128 if (const auto *SubRC = TRI.getCommonSubClass( 129 OpRC, TRI.getConstrainedRegClassForOperand(RegMO, MRI))) 130 OpRC = SubRC; 131 132 OpRC = TRI.getAllocatableClass(OpRC); 133 } 134 135 if (!OpRC) { 136 assert((!isTargetSpecificOpcode(II.getOpcode()) || RegMO.isUse()) && 137 "Register class constraint is required unless either the " 138 "instruction is target independent or the operand is a use"); 139 // FIXME: Just bailing out like this here could be not enough, unless we 140 // expect the users of this function to do the right thing for PHIs and 141 // COPY: 142 // v1 = COPY v0 143 // v2 = COPY v1 144 // v1 here may end up not being constrained at all. Please notice that to 145 // reproduce the issue we likely need a destination pattern of a selection 146 // rule producing such extra copies, not just an input GMIR with them as 147 // every existing target using selectImpl handles copies before calling it 148 // and they never reach this function. 149 return Reg; 150 } 151 return constrainOperandRegClass(MF, TRI, MRI, TII, RBI, InsertPt, *OpRC, 152 RegMO); 153 } 154 155 bool llvm::constrainSelectedInstRegOperands(MachineInstr &I, 156 const TargetInstrInfo &TII, 157 const TargetRegisterInfo &TRI, 158 const RegisterBankInfo &RBI) { 159 assert(!isPreISelGenericOpcode(I.getOpcode()) && 160 "A selected instruction is expected"); 161 MachineBasicBlock &MBB = *I.getParent(); 162 MachineFunction &MF = *MBB.getParent(); 163 MachineRegisterInfo &MRI = MF.getRegInfo(); 164 165 for (unsigned OpI = 0, OpE = I.getNumExplicitOperands(); OpI != OpE; ++OpI) { 166 MachineOperand &MO = I.getOperand(OpI); 167 168 // There's nothing to be done on non-register operands. 169 if (!MO.isReg()) 170 continue; 171 172 LLVM_DEBUG(dbgs() << "Converting operand: " << MO << '\n'); 173 assert(MO.isReg() && "Unsupported non-reg operand"); 174 175 Register Reg = MO.getReg(); 176 // Physical registers don't need to be constrained. 177 if (Reg.isPhysical()) 178 continue; 179 180 // Register operands with a value of 0 (e.g. predicate operands) don't need 181 // to be constrained. 182 if (Reg == 0) 183 continue; 184 185 // If the operand is a vreg, we should constrain its regclass, and only 186 // insert COPYs if that's impossible. 187 // constrainOperandRegClass does that for us. 188 constrainOperandRegClass(MF, TRI, MRI, TII, RBI, I, I.getDesc(), MO, OpI); 189 190 // Tie uses to defs as indicated in MCInstrDesc if this hasn't already been 191 // done. 192 if (MO.isUse()) { 193 int DefIdx = I.getDesc().getOperandConstraint(OpI, MCOI::TIED_TO); 194 if (DefIdx != -1 && !I.isRegTiedToUseOperand(DefIdx)) 195 I.tieOperands(DefIdx, OpI); 196 } 197 } 198 return true; 199 } 200 201 bool llvm::canReplaceReg(Register DstReg, Register SrcReg, 202 MachineRegisterInfo &MRI) { 203 // Give up if either DstReg or SrcReg is a physical register. 204 if (DstReg.isPhysical() || SrcReg.isPhysical()) 205 return false; 206 // Give up if the types don't match. 207 if (MRI.getType(DstReg) != MRI.getType(SrcReg)) 208 return false; 209 // Replace if either DstReg has no constraints or the register 210 // constraints match. 211 const auto &DstRBC = MRI.getRegClassOrRegBank(DstReg); 212 if (!DstRBC || DstRBC == MRI.getRegClassOrRegBank(SrcReg)) 213 return true; 214 215 // Otherwise match if the Src is already a regclass that is covered by the Dst 216 // RegBank. 217 return DstRBC.is<const RegisterBank *>() && MRI.getRegClassOrNull(SrcReg) && 218 DstRBC.get<const RegisterBank *>()->covers( 219 *MRI.getRegClassOrNull(SrcReg)); 220 } 221 222 bool llvm::isTriviallyDead(const MachineInstr &MI, 223 const MachineRegisterInfo &MRI) { 224 // FIXME: This logical is mostly duplicated with 225 // DeadMachineInstructionElim::isDead. Why is LOCAL_ESCAPE not considered in 226 // MachineInstr::isLabel? 227 228 // Don't delete frame allocation labels. 229 if (MI.getOpcode() == TargetOpcode::LOCAL_ESCAPE) 230 return false; 231 // LIFETIME markers should be preserved even if they seem dead. 232 if (MI.getOpcode() == TargetOpcode::LIFETIME_START || 233 MI.getOpcode() == TargetOpcode::LIFETIME_END) 234 return false; 235 236 // If we can move an instruction, we can remove it. Otherwise, it has 237 // a side-effect of some sort. 238 bool SawStore = false; 239 if (!MI.isSafeToMove(/*AA=*/nullptr, SawStore) && !MI.isPHI()) 240 return false; 241 242 // Instructions without side-effects are dead iff they only define dead vregs. 243 for (const auto &MO : MI.all_defs()) { 244 Register Reg = MO.getReg(); 245 if (Reg.isPhysical() || !MRI.use_nodbg_empty(Reg)) 246 return false; 247 } 248 return true; 249 } 250 251 static void reportGISelDiagnostic(DiagnosticSeverity Severity, 252 MachineFunction &MF, 253 const TargetPassConfig &TPC, 254 MachineOptimizationRemarkEmitter &MORE, 255 MachineOptimizationRemarkMissed &R) { 256 bool IsFatal = Severity == DS_Error && 257 TPC.isGlobalISelAbortEnabled(); 258 // Print the function name explicitly if we don't have a debug location (which 259 // makes the diagnostic less useful) or if we're going to emit a raw error. 260 if (!R.getLocation().isValid() || IsFatal) 261 R << (" (in function: " + MF.getName() + ")").str(); 262 263 if (IsFatal) 264 report_fatal_error(Twine(R.getMsg())); 265 else 266 MORE.emit(R); 267 } 268 269 void llvm::reportGISelWarning(MachineFunction &MF, const TargetPassConfig &TPC, 270 MachineOptimizationRemarkEmitter &MORE, 271 MachineOptimizationRemarkMissed &R) { 272 reportGISelDiagnostic(DS_Warning, MF, TPC, MORE, R); 273 } 274 275 void llvm::reportGISelFailure(MachineFunction &MF, const TargetPassConfig &TPC, 276 MachineOptimizationRemarkEmitter &MORE, 277 MachineOptimizationRemarkMissed &R) { 278 MF.getProperties().set(MachineFunctionProperties::Property::FailedISel); 279 reportGISelDiagnostic(DS_Error, MF, TPC, MORE, R); 280 } 281 282 void llvm::reportGISelFailure(MachineFunction &MF, const TargetPassConfig &TPC, 283 MachineOptimizationRemarkEmitter &MORE, 284 const char *PassName, StringRef Msg, 285 const MachineInstr &MI) { 286 MachineOptimizationRemarkMissed R(PassName, "GISelFailure: ", 287 MI.getDebugLoc(), MI.getParent()); 288 R << Msg; 289 // Printing MI is expensive; only do it if expensive remarks are enabled. 290 if (TPC.isGlobalISelAbortEnabled() || MORE.allowExtraAnalysis(PassName)) 291 R << ": " << ore::MNV("Inst", MI); 292 reportGISelFailure(MF, TPC, MORE, R); 293 } 294 295 std::optional<APInt> llvm::getIConstantVRegVal(Register VReg, 296 const MachineRegisterInfo &MRI) { 297 std::optional<ValueAndVReg> ValAndVReg = getIConstantVRegValWithLookThrough( 298 VReg, MRI, /*LookThroughInstrs*/ false); 299 assert((!ValAndVReg || ValAndVReg->VReg == VReg) && 300 "Value found while looking through instrs"); 301 if (!ValAndVReg) 302 return std::nullopt; 303 return ValAndVReg->Value; 304 } 305 306 std::optional<int64_t> 307 llvm::getIConstantVRegSExtVal(Register VReg, const MachineRegisterInfo &MRI) { 308 std::optional<APInt> Val = getIConstantVRegVal(VReg, MRI); 309 if (Val && Val->getBitWidth() <= 64) 310 return Val->getSExtValue(); 311 return std::nullopt; 312 } 313 314 namespace { 315 316 // This function is used in many places, and as such, it has some 317 // micro-optimizations to try and make it as fast as it can be. 318 // 319 // - We use template arguments to avoid an indirect call caused by passing a 320 // function_ref/std::function 321 // - GetAPCstValue does not return std::optional<APInt> as that's expensive. 322 // Instead it returns true/false and places the result in a pre-constructed 323 // APInt. 324 // 325 // Please change this function carefully and benchmark your changes. 326 template <bool (*IsConstantOpcode)(const MachineInstr *), 327 bool (*GetAPCstValue)(const MachineInstr *MI, APInt &)> 328 std::optional<ValueAndVReg> 329 getConstantVRegValWithLookThrough(Register VReg, const MachineRegisterInfo &MRI, 330 bool LookThroughInstrs = true, 331 bool LookThroughAnyExt = false) { 332 SmallVector<std::pair<unsigned, unsigned>, 4> SeenOpcodes; 333 MachineInstr *MI; 334 335 while ((MI = MRI.getVRegDef(VReg)) && !IsConstantOpcode(MI) && 336 LookThroughInstrs) { 337 switch (MI->getOpcode()) { 338 case TargetOpcode::G_ANYEXT: 339 if (!LookThroughAnyExt) 340 return std::nullopt; 341 [[fallthrough]]; 342 case TargetOpcode::G_TRUNC: 343 case TargetOpcode::G_SEXT: 344 case TargetOpcode::G_ZEXT: 345 SeenOpcodes.push_back(std::make_pair( 346 MI->getOpcode(), 347 MRI.getType(MI->getOperand(0).getReg()).getSizeInBits())); 348 VReg = MI->getOperand(1).getReg(); 349 break; 350 case TargetOpcode::COPY: 351 VReg = MI->getOperand(1).getReg(); 352 if (VReg.isPhysical()) 353 return std::nullopt; 354 break; 355 case TargetOpcode::G_INTTOPTR: 356 VReg = MI->getOperand(1).getReg(); 357 break; 358 default: 359 return std::nullopt; 360 } 361 } 362 if (!MI || !IsConstantOpcode(MI)) 363 return std::nullopt; 364 365 APInt Val; 366 if (!GetAPCstValue(MI, Val)) 367 return std::nullopt; 368 for (auto &Pair : reverse(SeenOpcodes)) { 369 switch (Pair.first) { 370 case TargetOpcode::G_TRUNC: 371 Val = Val.trunc(Pair.second); 372 break; 373 case TargetOpcode::G_ANYEXT: 374 case TargetOpcode::G_SEXT: 375 Val = Val.sext(Pair.second); 376 break; 377 case TargetOpcode::G_ZEXT: 378 Val = Val.zext(Pair.second); 379 break; 380 } 381 } 382 383 return ValueAndVReg{std::move(Val), VReg}; 384 } 385 386 bool isIConstant(const MachineInstr *MI) { 387 if (!MI) 388 return false; 389 return MI->getOpcode() == TargetOpcode::G_CONSTANT; 390 } 391 392 bool isFConstant(const MachineInstr *MI) { 393 if (!MI) 394 return false; 395 return MI->getOpcode() == TargetOpcode::G_FCONSTANT; 396 } 397 398 bool isAnyConstant(const MachineInstr *MI) { 399 if (!MI) 400 return false; 401 unsigned Opc = MI->getOpcode(); 402 return Opc == TargetOpcode::G_CONSTANT || Opc == TargetOpcode::G_FCONSTANT; 403 } 404 405 bool getCImmAsAPInt(const MachineInstr *MI, APInt &Result) { 406 const MachineOperand &CstVal = MI->getOperand(1); 407 if (!CstVal.isCImm()) 408 return false; 409 Result = CstVal.getCImm()->getValue(); 410 return true; 411 } 412 413 bool getCImmOrFPImmAsAPInt(const MachineInstr *MI, APInt &Result) { 414 const MachineOperand &CstVal = MI->getOperand(1); 415 if (CstVal.isCImm()) 416 Result = CstVal.getCImm()->getValue(); 417 else if (CstVal.isFPImm()) 418 Result = CstVal.getFPImm()->getValueAPF().bitcastToAPInt(); 419 else 420 return false; 421 return true; 422 } 423 424 } // end anonymous namespace 425 426 std::optional<ValueAndVReg> llvm::getIConstantVRegValWithLookThrough( 427 Register VReg, const MachineRegisterInfo &MRI, bool LookThroughInstrs) { 428 return getConstantVRegValWithLookThrough<isIConstant, getCImmAsAPInt>( 429 VReg, MRI, LookThroughInstrs); 430 } 431 432 std::optional<ValueAndVReg> llvm::getAnyConstantVRegValWithLookThrough( 433 Register VReg, const MachineRegisterInfo &MRI, bool LookThroughInstrs, 434 bool LookThroughAnyExt) { 435 return getConstantVRegValWithLookThrough<isAnyConstant, 436 getCImmOrFPImmAsAPInt>( 437 VReg, MRI, LookThroughInstrs, LookThroughAnyExt); 438 } 439 440 std::optional<FPValueAndVReg> llvm::getFConstantVRegValWithLookThrough( 441 Register VReg, const MachineRegisterInfo &MRI, bool LookThroughInstrs) { 442 auto Reg = 443 getConstantVRegValWithLookThrough<isFConstant, getCImmOrFPImmAsAPInt>( 444 VReg, MRI, LookThroughInstrs); 445 if (!Reg) 446 return std::nullopt; 447 return FPValueAndVReg{getConstantFPVRegVal(Reg->VReg, MRI)->getValueAPF(), 448 Reg->VReg}; 449 } 450 451 const ConstantFP * 452 llvm::getConstantFPVRegVal(Register VReg, const MachineRegisterInfo &MRI) { 453 MachineInstr *MI = MRI.getVRegDef(VReg); 454 if (TargetOpcode::G_FCONSTANT != MI->getOpcode()) 455 return nullptr; 456 return MI->getOperand(1).getFPImm(); 457 } 458 459 std::optional<DefinitionAndSourceRegister> 460 llvm::getDefSrcRegIgnoringCopies(Register Reg, const MachineRegisterInfo &MRI) { 461 Register DefSrcReg = Reg; 462 auto *DefMI = MRI.getVRegDef(Reg); 463 auto DstTy = MRI.getType(DefMI->getOperand(0).getReg()); 464 if (!DstTy.isValid()) 465 return std::nullopt; 466 unsigned Opc = DefMI->getOpcode(); 467 while (Opc == TargetOpcode::COPY || isPreISelGenericOptimizationHint(Opc)) { 468 Register SrcReg = DefMI->getOperand(1).getReg(); 469 auto SrcTy = MRI.getType(SrcReg); 470 if (!SrcTy.isValid()) 471 break; 472 DefMI = MRI.getVRegDef(SrcReg); 473 DefSrcReg = SrcReg; 474 Opc = DefMI->getOpcode(); 475 } 476 return DefinitionAndSourceRegister{DefMI, DefSrcReg}; 477 } 478 479 MachineInstr *llvm::getDefIgnoringCopies(Register Reg, 480 const MachineRegisterInfo &MRI) { 481 std::optional<DefinitionAndSourceRegister> DefSrcReg = 482 getDefSrcRegIgnoringCopies(Reg, MRI); 483 return DefSrcReg ? DefSrcReg->MI : nullptr; 484 } 485 486 Register llvm::getSrcRegIgnoringCopies(Register Reg, 487 const MachineRegisterInfo &MRI) { 488 std::optional<DefinitionAndSourceRegister> DefSrcReg = 489 getDefSrcRegIgnoringCopies(Reg, MRI); 490 return DefSrcReg ? DefSrcReg->Reg : Register(); 491 } 492 493 void llvm::extractParts(Register Reg, LLT Ty, int NumParts, 494 SmallVectorImpl<Register> &VRegs, 495 MachineIRBuilder &MIRBuilder, 496 MachineRegisterInfo &MRI) { 497 for (int i = 0; i < NumParts; ++i) 498 VRegs.push_back(MRI.createGenericVirtualRegister(Ty)); 499 MIRBuilder.buildUnmerge(VRegs, Reg); 500 } 501 502 bool llvm::extractParts(Register Reg, LLT RegTy, LLT MainTy, LLT &LeftoverTy, 503 SmallVectorImpl<Register> &VRegs, 504 SmallVectorImpl<Register> &LeftoverRegs, 505 MachineIRBuilder &MIRBuilder, 506 MachineRegisterInfo &MRI) { 507 assert(!LeftoverTy.isValid() && "this is an out argument"); 508 509 unsigned RegSize = RegTy.getSizeInBits(); 510 unsigned MainSize = MainTy.getSizeInBits(); 511 unsigned NumParts = RegSize / MainSize; 512 unsigned LeftoverSize = RegSize - NumParts * MainSize; 513 514 // Use an unmerge when possible. 515 if (LeftoverSize == 0) { 516 for (unsigned I = 0; I < NumParts; ++I) 517 VRegs.push_back(MRI.createGenericVirtualRegister(MainTy)); 518 MIRBuilder.buildUnmerge(VRegs, Reg); 519 return true; 520 } 521 522 // Try to use unmerge for irregular vector split where possible 523 // For example when splitting a <6 x i32> into <4 x i32> with <2 x i32> 524 // leftover, it becomes: 525 // <2 x i32> %2, <2 x i32>%3, <2 x i32> %4 = G_UNMERGE_VALUE <6 x i32> %1 526 // <4 x i32> %5 = G_CONCAT_VECTOR <2 x i32> %2, <2 x i32> %3 527 if (RegTy.isVector() && MainTy.isVector()) { 528 unsigned RegNumElts = RegTy.getNumElements(); 529 unsigned MainNumElts = MainTy.getNumElements(); 530 unsigned LeftoverNumElts = RegNumElts % MainNumElts; 531 // If can unmerge to LeftoverTy, do it 532 if (MainNumElts % LeftoverNumElts == 0 && 533 RegNumElts % LeftoverNumElts == 0 && 534 RegTy.getScalarSizeInBits() == MainTy.getScalarSizeInBits() && 535 LeftoverNumElts > 1) { 536 LeftoverTy = 537 LLT::fixed_vector(LeftoverNumElts, RegTy.getScalarSizeInBits()); 538 539 // Unmerge the SrcReg to LeftoverTy vectors 540 SmallVector<Register, 4> UnmergeValues; 541 extractParts(Reg, LeftoverTy, RegNumElts / LeftoverNumElts, UnmergeValues, 542 MIRBuilder, MRI); 543 544 // Find how many LeftoverTy makes one MainTy 545 unsigned LeftoverPerMain = MainNumElts / LeftoverNumElts; 546 unsigned NumOfLeftoverVal = 547 ((RegNumElts % MainNumElts) / LeftoverNumElts); 548 549 // Create as many MainTy as possible using unmerged value 550 SmallVector<Register, 4> MergeValues; 551 for (unsigned I = 0; I < UnmergeValues.size() - NumOfLeftoverVal; I++) { 552 MergeValues.push_back(UnmergeValues[I]); 553 if (MergeValues.size() == LeftoverPerMain) { 554 VRegs.push_back( 555 MIRBuilder.buildMergeLikeInstr(MainTy, MergeValues).getReg(0)); 556 MergeValues.clear(); 557 } 558 } 559 // Populate LeftoverRegs with the leftovers 560 for (unsigned I = UnmergeValues.size() - NumOfLeftoverVal; 561 I < UnmergeValues.size(); I++) { 562 LeftoverRegs.push_back(UnmergeValues[I]); 563 } 564 return true; 565 } 566 } 567 // Perform irregular split. Leftover is last element of RegPieces. 568 if (MainTy.isVector()) { 569 SmallVector<Register, 8> RegPieces; 570 extractVectorParts(Reg, MainTy.getNumElements(), RegPieces, MIRBuilder, 571 MRI); 572 for (unsigned i = 0; i < RegPieces.size() - 1; ++i) 573 VRegs.push_back(RegPieces[i]); 574 LeftoverRegs.push_back(RegPieces[RegPieces.size() - 1]); 575 LeftoverTy = MRI.getType(LeftoverRegs[0]); 576 return true; 577 } 578 579 LeftoverTy = LLT::scalar(LeftoverSize); 580 // For irregular sizes, extract the individual parts. 581 for (unsigned I = 0; I != NumParts; ++I) { 582 Register NewReg = MRI.createGenericVirtualRegister(MainTy); 583 VRegs.push_back(NewReg); 584 MIRBuilder.buildExtract(NewReg, Reg, MainSize * I); 585 } 586 587 for (unsigned Offset = MainSize * NumParts; Offset < RegSize; 588 Offset += LeftoverSize) { 589 Register NewReg = MRI.createGenericVirtualRegister(LeftoverTy); 590 LeftoverRegs.push_back(NewReg); 591 MIRBuilder.buildExtract(NewReg, Reg, Offset); 592 } 593 594 return true; 595 } 596 597 void llvm::extractVectorParts(Register Reg, unsigned NumElts, 598 SmallVectorImpl<Register> &VRegs, 599 MachineIRBuilder &MIRBuilder, 600 MachineRegisterInfo &MRI) { 601 LLT RegTy = MRI.getType(Reg); 602 assert(RegTy.isVector() && "Expected a vector type"); 603 604 LLT EltTy = RegTy.getElementType(); 605 LLT NarrowTy = (NumElts == 1) ? EltTy : LLT::fixed_vector(NumElts, EltTy); 606 unsigned RegNumElts = RegTy.getNumElements(); 607 unsigned LeftoverNumElts = RegNumElts % NumElts; 608 unsigned NumNarrowTyPieces = RegNumElts / NumElts; 609 610 // Perfect split without leftover 611 if (LeftoverNumElts == 0) 612 return extractParts(Reg, NarrowTy, NumNarrowTyPieces, VRegs, MIRBuilder, 613 MRI); 614 615 // Irregular split. Provide direct access to all elements for artifact 616 // combiner using unmerge to elements. Then build vectors with NumElts 617 // elements. Remaining element(s) will be (used to build vector) Leftover. 618 SmallVector<Register, 8> Elts; 619 extractParts(Reg, EltTy, RegNumElts, Elts, MIRBuilder, MRI); 620 621 unsigned Offset = 0; 622 // Requested sub-vectors of NarrowTy. 623 for (unsigned i = 0; i < NumNarrowTyPieces; ++i, Offset += NumElts) { 624 ArrayRef<Register> Pieces(&Elts[Offset], NumElts); 625 VRegs.push_back(MIRBuilder.buildMergeLikeInstr(NarrowTy, Pieces).getReg(0)); 626 } 627 628 // Leftover element(s). 629 if (LeftoverNumElts == 1) { 630 VRegs.push_back(Elts[Offset]); 631 } else { 632 LLT LeftoverTy = LLT::fixed_vector(LeftoverNumElts, EltTy); 633 ArrayRef<Register> Pieces(&Elts[Offset], LeftoverNumElts); 634 VRegs.push_back( 635 MIRBuilder.buildMergeLikeInstr(LeftoverTy, Pieces).getReg(0)); 636 } 637 } 638 639 MachineInstr *llvm::getOpcodeDef(unsigned Opcode, Register Reg, 640 const MachineRegisterInfo &MRI) { 641 MachineInstr *DefMI = getDefIgnoringCopies(Reg, MRI); 642 return DefMI && DefMI->getOpcode() == Opcode ? DefMI : nullptr; 643 } 644 645 APFloat llvm::getAPFloatFromSize(double Val, unsigned Size) { 646 if (Size == 32) 647 return APFloat(float(Val)); 648 if (Size == 64) 649 return APFloat(Val); 650 if (Size != 16) 651 llvm_unreachable("Unsupported FPConstant size"); 652 bool Ignored; 653 APFloat APF(Val); 654 APF.convert(APFloat::IEEEhalf(), APFloat::rmNearestTiesToEven, &Ignored); 655 return APF; 656 } 657 658 std::optional<APInt> llvm::ConstantFoldBinOp(unsigned Opcode, 659 const Register Op1, 660 const Register Op2, 661 const MachineRegisterInfo &MRI) { 662 auto MaybeOp2Cst = getAnyConstantVRegValWithLookThrough(Op2, MRI, false); 663 if (!MaybeOp2Cst) 664 return std::nullopt; 665 666 auto MaybeOp1Cst = getAnyConstantVRegValWithLookThrough(Op1, MRI, false); 667 if (!MaybeOp1Cst) 668 return std::nullopt; 669 670 const APInt &C1 = MaybeOp1Cst->Value; 671 const APInt &C2 = MaybeOp2Cst->Value; 672 switch (Opcode) { 673 default: 674 break; 675 case TargetOpcode::G_ADD: 676 return C1 + C2; 677 case TargetOpcode::G_PTR_ADD: 678 // Types can be of different width here. 679 // Result needs to be the same width as C1, so trunc or sext C2. 680 return C1 + C2.sextOrTrunc(C1.getBitWidth()); 681 case TargetOpcode::G_AND: 682 return C1 & C2; 683 case TargetOpcode::G_ASHR: 684 return C1.ashr(C2); 685 case TargetOpcode::G_LSHR: 686 return C1.lshr(C2); 687 case TargetOpcode::G_MUL: 688 return C1 * C2; 689 case TargetOpcode::G_OR: 690 return C1 | C2; 691 case TargetOpcode::G_SHL: 692 return C1 << C2; 693 case TargetOpcode::G_SUB: 694 return C1 - C2; 695 case TargetOpcode::G_XOR: 696 return C1 ^ C2; 697 case TargetOpcode::G_UDIV: 698 if (!C2.getBoolValue()) 699 break; 700 return C1.udiv(C2); 701 case TargetOpcode::G_SDIV: 702 if (!C2.getBoolValue()) 703 break; 704 return C1.sdiv(C2); 705 case TargetOpcode::G_UREM: 706 if (!C2.getBoolValue()) 707 break; 708 return C1.urem(C2); 709 case TargetOpcode::G_SREM: 710 if (!C2.getBoolValue()) 711 break; 712 return C1.srem(C2); 713 case TargetOpcode::G_SMIN: 714 return APIntOps::smin(C1, C2); 715 case TargetOpcode::G_SMAX: 716 return APIntOps::smax(C1, C2); 717 case TargetOpcode::G_UMIN: 718 return APIntOps::umin(C1, C2); 719 case TargetOpcode::G_UMAX: 720 return APIntOps::umax(C1, C2); 721 } 722 723 return std::nullopt; 724 } 725 726 std::optional<APFloat> 727 llvm::ConstantFoldFPBinOp(unsigned Opcode, const Register Op1, 728 const Register Op2, const MachineRegisterInfo &MRI) { 729 const ConstantFP *Op2Cst = getConstantFPVRegVal(Op2, MRI); 730 if (!Op2Cst) 731 return std::nullopt; 732 733 const ConstantFP *Op1Cst = getConstantFPVRegVal(Op1, MRI); 734 if (!Op1Cst) 735 return std::nullopt; 736 737 APFloat C1 = Op1Cst->getValueAPF(); 738 const APFloat &C2 = Op2Cst->getValueAPF(); 739 switch (Opcode) { 740 case TargetOpcode::G_FADD: 741 C1.add(C2, APFloat::rmNearestTiesToEven); 742 return C1; 743 case TargetOpcode::G_FSUB: 744 C1.subtract(C2, APFloat::rmNearestTiesToEven); 745 return C1; 746 case TargetOpcode::G_FMUL: 747 C1.multiply(C2, APFloat::rmNearestTiesToEven); 748 return C1; 749 case TargetOpcode::G_FDIV: 750 C1.divide(C2, APFloat::rmNearestTiesToEven); 751 return C1; 752 case TargetOpcode::G_FREM: 753 C1.mod(C2); 754 return C1; 755 case TargetOpcode::G_FCOPYSIGN: 756 C1.copySign(C2); 757 return C1; 758 case TargetOpcode::G_FMINNUM: 759 return minnum(C1, C2); 760 case TargetOpcode::G_FMAXNUM: 761 return maxnum(C1, C2); 762 case TargetOpcode::G_FMINIMUM: 763 return minimum(C1, C2); 764 case TargetOpcode::G_FMAXIMUM: 765 return maximum(C1, C2); 766 case TargetOpcode::G_FMINNUM_IEEE: 767 case TargetOpcode::G_FMAXNUM_IEEE: 768 // FIXME: These operations were unfortunately named. fminnum/fmaxnum do not 769 // follow the IEEE behavior for signaling nans and follow libm's fmin/fmax, 770 // and currently there isn't a nice wrapper in APFloat for the version with 771 // correct snan handling. 772 break; 773 default: 774 break; 775 } 776 777 return std::nullopt; 778 } 779 780 SmallVector<APInt> 781 llvm::ConstantFoldVectorBinop(unsigned Opcode, const Register Op1, 782 const Register Op2, 783 const MachineRegisterInfo &MRI) { 784 auto *SrcVec2 = getOpcodeDef<GBuildVector>(Op2, MRI); 785 if (!SrcVec2) 786 return SmallVector<APInt>(); 787 788 auto *SrcVec1 = getOpcodeDef<GBuildVector>(Op1, MRI); 789 if (!SrcVec1) 790 return SmallVector<APInt>(); 791 792 SmallVector<APInt> FoldedElements; 793 for (unsigned Idx = 0, E = SrcVec1->getNumSources(); Idx < E; ++Idx) { 794 auto MaybeCst = ConstantFoldBinOp(Opcode, SrcVec1->getSourceReg(Idx), 795 SrcVec2->getSourceReg(Idx), MRI); 796 if (!MaybeCst) 797 return SmallVector<APInt>(); 798 FoldedElements.push_back(*MaybeCst); 799 } 800 return FoldedElements; 801 } 802 803 bool llvm::isKnownNeverNaN(Register Val, const MachineRegisterInfo &MRI, 804 bool SNaN) { 805 const MachineInstr *DefMI = MRI.getVRegDef(Val); 806 if (!DefMI) 807 return false; 808 809 const TargetMachine& TM = DefMI->getMF()->getTarget(); 810 if (DefMI->getFlag(MachineInstr::FmNoNans) || TM.Options.NoNaNsFPMath) 811 return true; 812 813 // If the value is a constant, we can obviously see if it is a NaN or not. 814 if (const ConstantFP *FPVal = getConstantFPVRegVal(Val, MRI)) { 815 return !FPVal->getValueAPF().isNaN() || 816 (SNaN && !FPVal->getValueAPF().isSignaling()); 817 } 818 819 if (DefMI->getOpcode() == TargetOpcode::G_BUILD_VECTOR) { 820 for (const auto &Op : DefMI->uses()) 821 if (!isKnownNeverNaN(Op.getReg(), MRI, SNaN)) 822 return false; 823 return true; 824 } 825 826 switch (DefMI->getOpcode()) { 827 default: 828 break; 829 case TargetOpcode::G_FADD: 830 case TargetOpcode::G_FSUB: 831 case TargetOpcode::G_FMUL: 832 case TargetOpcode::G_FDIV: 833 case TargetOpcode::G_FREM: 834 case TargetOpcode::G_FSIN: 835 case TargetOpcode::G_FCOS: 836 case TargetOpcode::G_FMA: 837 case TargetOpcode::G_FMAD: 838 if (SNaN) 839 return true; 840 841 // TODO: Need isKnownNeverInfinity 842 return false; 843 case TargetOpcode::G_FMINNUM_IEEE: 844 case TargetOpcode::G_FMAXNUM_IEEE: { 845 if (SNaN) 846 return true; 847 // This can return a NaN if either operand is an sNaN, or if both operands 848 // are NaN. 849 return (isKnownNeverNaN(DefMI->getOperand(1).getReg(), MRI) && 850 isKnownNeverSNaN(DefMI->getOperand(2).getReg(), MRI)) || 851 (isKnownNeverSNaN(DefMI->getOperand(1).getReg(), MRI) && 852 isKnownNeverNaN(DefMI->getOperand(2).getReg(), MRI)); 853 } 854 case TargetOpcode::G_FMINNUM: 855 case TargetOpcode::G_FMAXNUM: { 856 // Only one needs to be known not-nan, since it will be returned if the 857 // other ends up being one. 858 return isKnownNeverNaN(DefMI->getOperand(1).getReg(), MRI, SNaN) || 859 isKnownNeverNaN(DefMI->getOperand(2).getReg(), MRI, SNaN); 860 } 861 } 862 863 if (SNaN) { 864 // FP operations quiet. For now, just handle the ones inserted during 865 // legalization. 866 switch (DefMI->getOpcode()) { 867 case TargetOpcode::G_FPEXT: 868 case TargetOpcode::G_FPTRUNC: 869 case TargetOpcode::G_FCANONICALIZE: 870 return true; 871 default: 872 return false; 873 } 874 } 875 876 return false; 877 } 878 879 Align llvm::inferAlignFromPtrInfo(MachineFunction &MF, 880 const MachinePointerInfo &MPO) { 881 auto PSV = dyn_cast_if_present<const PseudoSourceValue *>(MPO.V); 882 if (auto FSPV = dyn_cast_or_null<FixedStackPseudoSourceValue>(PSV)) { 883 MachineFrameInfo &MFI = MF.getFrameInfo(); 884 return commonAlignment(MFI.getObjectAlign(FSPV->getFrameIndex()), 885 MPO.Offset); 886 } 887 888 if (const Value *V = dyn_cast_if_present<const Value *>(MPO.V)) { 889 const Module *M = MF.getFunction().getParent(); 890 return V->getPointerAlignment(M->getDataLayout()); 891 } 892 893 return Align(1); 894 } 895 896 Register llvm::getFunctionLiveInPhysReg(MachineFunction &MF, 897 const TargetInstrInfo &TII, 898 MCRegister PhysReg, 899 const TargetRegisterClass &RC, 900 const DebugLoc &DL, LLT RegTy) { 901 MachineBasicBlock &EntryMBB = MF.front(); 902 MachineRegisterInfo &MRI = MF.getRegInfo(); 903 Register LiveIn = MRI.getLiveInVirtReg(PhysReg); 904 if (LiveIn) { 905 MachineInstr *Def = MRI.getVRegDef(LiveIn); 906 if (Def) { 907 // FIXME: Should the verifier check this is in the entry block? 908 assert(Def->getParent() == &EntryMBB && "live-in copy not in entry block"); 909 return LiveIn; 910 } 911 912 // It's possible the incoming argument register and copy was added during 913 // lowering, but later deleted due to being/becoming dead. If this happens, 914 // re-insert the copy. 915 } else { 916 // The live in register was not present, so add it. 917 LiveIn = MF.addLiveIn(PhysReg, &RC); 918 if (RegTy.isValid()) 919 MRI.setType(LiveIn, RegTy); 920 } 921 922 BuildMI(EntryMBB, EntryMBB.begin(), DL, TII.get(TargetOpcode::COPY), LiveIn) 923 .addReg(PhysReg); 924 if (!EntryMBB.isLiveIn(PhysReg)) 925 EntryMBB.addLiveIn(PhysReg); 926 return LiveIn; 927 } 928 929 std::optional<APInt> llvm::ConstantFoldExtOp(unsigned Opcode, 930 const Register Op1, uint64_t Imm, 931 const MachineRegisterInfo &MRI) { 932 auto MaybeOp1Cst = getIConstantVRegVal(Op1, MRI); 933 if (MaybeOp1Cst) { 934 switch (Opcode) { 935 default: 936 break; 937 case TargetOpcode::G_SEXT_INREG: { 938 LLT Ty = MRI.getType(Op1); 939 return MaybeOp1Cst->trunc(Imm).sext(Ty.getScalarSizeInBits()); 940 } 941 } 942 } 943 return std::nullopt; 944 } 945 946 std::optional<APInt> llvm::ConstantFoldCastOp(unsigned Opcode, LLT DstTy, 947 const Register Op0, 948 const MachineRegisterInfo &MRI) { 949 std::optional<APInt> Val = getIConstantVRegVal(Op0, MRI); 950 if (!Val) 951 return Val; 952 953 const unsigned DstSize = DstTy.getScalarSizeInBits(); 954 955 switch (Opcode) { 956 case TargetOpcode::G_SEXT: 957 return Val->sext(DstSize); 958 case TargetOpcode::G_ZEXT: 959 case TargetOpcode::G_ANYEXT: 960 // TODO: DAG considers target preference when constant folding any_extend. 961 return Val->zext(DstSize); 962 default: 963 break; 964 } 965 966 llvm_unreachable("unexpected cast opcode to constant fold"); 967 } 968 969 std::optional<APFloat> 970 llvm::ConstantFoldIntToFloat(unsigned Opcode, LLT DstTy, Register Src, 971 const MachineRegisterInfo &MRI) { 972 assert(Opcode == TargetOpcode::G_SITOFP || Opcode == TargetOpcode::G_UITOFP); 973 if (auto MaybeSrcVal = getIConstantVRegVal(Src, MRI)) { 974 APFloat DstVal(getFltSemanticForLLT(DstTy)); 975 DstVal.convertFromAPInt(*MaybeSrcVal, Opcode == TargetOpcode::G_SITOFP, 976 APFloat::rmNearestTiesToEven); 977 return DstVal; 978 } 979 return std::nullopt; 980 } 981 982 std::optional<SmallVector<unsigned>> 983 llvm::ConstantFoldCountZeros(Register Src, const MachineRegisterInfo &MRI, 984 std::function<unsigned(APInt)> CB) { 985 LLT Ty = MRI.getType(Src); 986 SmallVector<unsigned> FoldedCTLZs; 987 auto tryFoldScalar = [&](Register R) -> std::optional<unsigned> { 988 auto MaybeCst = getIConstantVRegVal(R, MRI); 989 if (!MaybeCst) 990 return std::nullopt; 991 return CB(*MaybeCst); 992 }; 993 if (Ty.isVector()) { 994 // Try to constant fold each element. 995 auto *BV = getOpcodeDef<GBuildVector>(Src, MRI); 996 if (!BV) 997 return std::nullopt; 998 for (unsigned SrcIdx = 0; SrcIdx < BV->getNumSources(); ++SrcIdx) { 999 if (auto MaybeFold = tryFoldScalar(BV->getSourceReg(SrcIdx))) { 1000 FoldedCTLZs.emplace_back(*MaybeFold); 1001 continue; 1002 } 1003 return std::nullopt; 1004 } 1005 return FoldedCTLZs; 1006 } 1007 if (auto MaybeCst = tryFoldScalar(Src)) { 1008 FoldedCTLZs.emplace_back(*MaybeCst); 1009 return FoldedCTLZs; 1010 } 1011 return std::nullopt; 1012 } 1013 1014 std::optional<SmallVector<APInt>> 1015 llvm::ConstantFoldICmp(unsigned Pred, const Register Op1, const Register Op2, 1016 const MachineRegisterInfo &MRI) { 1017 LLT Ty = MRI.getType(Op1); 1018 if (Ty != MRI.getType(Op2)) 1019 return std::nullopt; 1020 1021 auto TryFoldScalar = [&MRI, Pred](Register LHS, 1022 Register RHS) -> std::optional<APInt> { 1023 auto LHSCst = getIConstantVRegVal(LHS, MRI); 1024 auto RHSCst = getIConstantVRegVal(RHS, MRI); 1025 if (!LHSCst || !RHSCst) 1026 return std::nullopt; 1027 1028 switch (Pred) { 1029 case CmpInst::Predicate::ICMP_EQ: 1030 return APInt(/*numBits=*/1, LHSCst->eq(*RHSCst)); 1031 case CmpInst::Predicate::ICMP_NE: 1032 return APInt(/*numBits=*/1, LHSCst->ne(*RHSCst)); 1033 case CmpInst::Predicate::ICMP_UGT: 1034 return APInt(/*numBits=*/1, LHSCst->ugt(*RHSCst)); 1035 case CmpInst::Predicate::ICMP_UGE: 1036 return APInt(/*numBits=*/1, LHSCst->uge(*RHSCst)); 1037 case CmpInst::Predicate::ICMP_ULT: 1038 return APInt(/*numBits=*/1, LHSCst->ult(*RHSCst)); 1039 case CmpInst::Predicate::ICMP_ULE: 1040 return APInt(/*numBits=*/1, LHSCst->ule(*RHSCst)); 1041 case CmpInst::Predicate::ICMP_SGT: 1042 return APInt(/*numBits=*/1, LHSCst->sgt(*RHSCst)); 1043 case CmpInst::Predicate::ICMP_SGE: 1044 return APInt(/*numBits=*/1, LHSCst->sge(*RHSCst)); 1045 case CmpInst::Predicate::ICMP_SLT: 1046 return APInt(/*numBits=*/1, LHSCst->slt(*RHSCst)); 1047 case CmpInst::Predicate::ICMP_SLE: 1048 return APInt(/*numBits=*/1, LHSCst->sle(*RHSCst)); 1049 default: 1050 return std::nullopt; 1051 } 1052 }; 1053 1054 SmallVector<APInt> FoldedICmps; 1055 1056 if (Ty.isVector()) { 1057 // Try to constant fold each element. 1058 auto *BV1 = getOpcodeDef<GBuildVector>(Op1, MRI); 1059 auto *BV2 = getOpcodeDef<GBuildVector>(Op2, MRI); 1060 if (!BV1 || !BV2) 1061 return std::nullopt; 1062 assert(BV1->getNumSources() == BV2->getNumSources() && "Invalid vectors"); 1063 for (unsigned I = 0; I < BV1->getNumSources(); ++I) { 1064 if (auto MaybeFold = 1065 TryFoldScalar(BV1->getSourceReg(I), BV2->getSourceReg(I))) { 1066 FoldedICmps.emplace_back(*MaybeFold); 1067 continue; 1068 } 1069 return std::nullopt; 1070 } 1071 return FoldedICmps; 1072 } 1073 1074 if (auto MaybeCst = TryFoldScalar(Op1, Op2)) { 1075 FoldedICmps.emplace_back(*MaybeCst); 1076 return FoldedICmps; 1077 } 1078 1079 return std::nullopt; 1080 } 1081 1082 bool llvm::isKnownToBeAPowerOfTwo(Register Reg, const MachineRegisterInfo &MRI, 1083 GISelKnownBits *KB) { 1084 std::optional<DefinitionAndSourceRegister> DefSrcReg = 1085 getDefSrcRegIgnoringCopies(Reg, MRI); 1086 if (!DefSrcReg) 1087 return false; 1088 1089 const MachineInstr &MI = *DefSrcReg->MI; 1090 const LLT Ty = MRI.getType(Reg); 1091 1092 switch (MI.getOpcode()) { 1093 case TargetOpcode::G_CONSTANT: { 1094 unsigned BitWidth = Ty.getScalarSizeInBits(); 1095 const ConstantInt *CI = MI.getOperand(1).getCImm(); 1096 return CI->getValue().zextOrTrunc(BitWidth).isPowerOf2(); 1097 } 1098 case TargetOpcode::G_SHL: { 1099 // A left-shift of a constant one will have exactly one bit set because 1100 // shifting the bit off the end is undefined. 1101 1102 // TODO: Constant splat 1103 if (auto ConstLHS = getIConstantVRegVal(MI.getOperand(1).getReg(), MRI)) { 1104 if (*ConstLHS == 1) 1105 return true; 1106 } 1107 1108 break; 1109 } 1110 case TargetOpcode::G_LSHR: { 1111 if (auto ConstLHS = getIConstantVRegVal(MI.getOperand(1).getReg(), MRI)) { 1112 if (ConstLHS->isSignMask()) 1113 return true; 1114 } 1115 1116 break; 1117 } 1118 case TargetOpcode::G_BUILD_VECTOR: { 1119 // TODO: Probably should have a recursion depth guard since you could have 1120 // bitcasted vector elements. 1121 for (const MachineOperand &MO : llvm::drop_begin(MI.operands())) 1122 if (!isKnownToBeAPowerOfTwo(MO.getReg(), MRI, KB)) 1123 return false; 1124 1125 return true; 1126 } 1127 case TargetOpcode::G_BUILD_VECTOR_TRUNC: { 1128 // Only handle constants since we would need to know if number of leading 1129 // zeros is greater than the truncation amount. 1130 const unsigned BitWidth = Ty.getScalarSizeInBits(); 1131 for (const MachineOperand &MO : llvm::drop_begin(MI.operands())) { 1132 auto Const = getIConstantVRegVal(MO.getReg(), MRI); 1133 if (!Const || !Const->zextOrTrunc(BitWidth).isPowerOf2()) 1134 return false; 1135 } 1136 1137 return true; 1138 } 1139 default: 1140 break; 1141 } 1142 1143 if (!KB) 1144 return false; 1145 1146 // More could be done here, though the above checks are enough 1147 // to handle some common cases. 1148 1149 // Fall back to computeKnownBits to catch other known cases. 1150 KnownBits Known = KB->getKnownBits(Reg); 1151 return (Known.countMaxPopulation() == 1) && (Known.countMinPopulation() == 1); 1152 } 1153 1154 void llvm::getSelectionDAGFallbackAnalysisUsage(AnalysisUsage &AU) { 1155 AU.addPreserved<StackProtector>(); 1156 } 1157 1158 LLT llvm::getLCMType(LLT OrigTy, LLT TargetTy) { 1159 if (OrigTy.getSizeInBits() == TargetTy.getSizeInBits()) 1160 return OrigTy; 1161 1162 if (OrigTy.isVector() && TargetTy.isVector()) { 1163 LLT OrigElt = OrigTy.getElementType(); 1164 LLT TargetElt = TargetTy.getElementType(); 1165 1166 // TODO: The docstring for this function says the intention is to use this 1167 // function to build MERGE/UNMERGE instructions. It won't be the case that 1168 // we generate a MERGE/UNMERGE between fixed and scalable vector types. We 1169 // could implement getLCMType between the two in the future if there was a 1170 // need, but it is not worth it now as this function should not be used in 1171 // that way. 1172 assert(((OrigTy.isScalableVector() && !TargetTy.isFixedVector()) || 1173 (OrigTy.isFixedVector() && !TargetTy.isScalableVector())) && 1174 "getLCMType not implemented between fixed and scalable vectors."); 1175 1176 if (OrigElt.getSizeInBits() == TargetElt.getSizeInBits()) { 1177 int GCDMinElts = std::gcd(OrigTy.getElementCount().getKnownMinValue(), 1178 TargetTy.getElementCount().getKnownMinValue()); 1179 // Prefer the original element type. 1180 ElementCount Mul = OrigTy.getElementCount().multiplyCoefficientBy( 1181 TargetTy.getElementCount().getKnownMinValue()); 1182 return LLT::vector(Mul.divideCoefficientBy(GCDMinElts), 1183 OrigTy.getElementType()); 1184 } 1185 unsigned LCM = std::lcm(OrigTy.getSizeInBits().getKnownMinValue(), 1186 TargetTy.getSizeInBits().getKnownMinValue()); 1187 return LLT::vector( 1188 ElementCount::get(LCM / OrigElt.getSizeInBits(), OrigTy.isScalable()), 1189 OrigElt); 1190 } 1191 1192 // One type is scalar, one type is vector 1193 if (OrigTy.isVector() || TargetTy.isVector()) { 1194 LLT VecTy = OrigTy.isVector() ? OrigTy : TargetTy; 1195 LLT ScalarTy = OrigTy.isVector() ? TargetTy : OrigTy; 1196 LLT EltTy = VecTy.getElementType(); 1197 LLT OrigEltTy = OrigTy.isVector() ? OrigTy.getElementType() : OrigTy; 1198 1199 // Prefer scalar type from OrigTy. 1200 if (EltTy.getSizeInBits() == ScalarTy.getSizeInBits()) 1201 return LLT::vector(VecTy.getElementCount(), OrigEltTy); 1202 1203 // Different size scalars. Create vector with the same total size. 1204 // LCM will take fixed/scalable from VecTy. 1205 unsigned LCM = std::lcm(EltTy.getSizeInBits().getFixedValue() * 1206 VecTy.getElementCount().getKnownMinValue(), 1207 ScalarTy.getSizeInBits().getFixedValue()); 1208 // Prefer type from OrigTy 1209 return LLT::vector(ElementCount::get(LCM / OrigEltTy.getSizeInBits(), 1210 VecTy.getElementCount().isScalable()), 1211 OrigEltTy); 1212 } 1213 1214 // At this point, both types are scalars of different size 1215 unsigned LCM = std::lcm(OrigTy.getSizeInBits().getFixedValue(), 1216 TargetTy.getSizeInBits().getFixedValue()); 1217 // Preserve pointer types. 1218 if (LCM == OrigTy.getSizeInBits()) 1219 return OrigTy; 1220 if (LCM == TargetTy.getSizeInBits()) 1221 return TargetTy; 1222 return LLT::scalar(LCM); 1223 } 1224 1225 LLT llvm::getCoverTy(LLT OrigTy, LLT TargetTy) { 1226 1227 if ((OrigTy.isScalableVector() && TargetTy.isFixedVector()) || 1228 (OrigTy.isFixedVector() && TargetTy.isScalableVector())) 1229 llvm_unreachable( 1230 "getCoverTy not implemented between fixed and scalable vectors."); 1231 1232 if (!OrigTy.isVector() || !TargetTy.isVector() || OrigTy == TargetTy || 1233 (OrigTy.getScalarSizeInBits() != TargetTy.getScalarSizeInBits())) 1234 return getLCMType(OrigTy, TargetTy); 1235 1236 unsigned OrigTyNumElts = OrigTy.getElementCount().getKnownMinValue(); 1237 unsigned TargetTyNumElts = TargetTy.getElementCount().getKnownMinValue(); 1238 if (OrigTyNumElts % TargetTyNumElts == 0) 1239 return OrigTy; 1240 1241 unsigned NumElts = alignTo(OrigTyNumElts, TargetTyNumElts); 1242 return LLT::scalarOrVector(ElementCount::getFixed(NumElts), 1243 OrigTy.getElementType()); 1244 } 1245 1246 LLT llvm::getGCDType(LLT OrigTy, LLT TargetTy) { 1247 if (OrigTy.getSizeInBits() == TargetTy.getSizeInBits()) 1248 return OrigTy; 1249 1250 if (OrigTy.isVector() && TargetTy.isVector()) { 1251 LLT OrigElt = OrigTy.getElementType(); 1252 1253 // TODO: The docstring for this function says the intention is to use this 1254 // function to build MERGE/UNMERGE instructions. It won't be the case that 1255 // we generate a MERGE/UNMERGE between fixed and scalable vector types. We 1256 // could implement getGCDType between the two in the future if there was a 1257 // need, but it is not worth it now as this function should not be used in 1258 // that way. 1259 assert(((OrigTy.isScalableVector() && !TargetTy.isFixedVector()) || 1260 (OrigTy.isFixedVector() && !TargetTy.isScalableVector())) && 1261 "getGCDType not implemented between fixed and scalable vectors."); 1262 1263 unsigned GCD = std::gcd(OrigTy.getSizeInBits().getKnownMinValue(), 1264 TargetTy.getSizeInBits().getKnownMinValue()); 1265 if (GCD == OrigElt.getSizeInBits()) 1266 return LLT::scalarOrVector(ElementCount::get(1, OrigTy.isScalable()), 1267 OrigElt); 1268 1269 // Cannot produce original element type, but both have vscale in common. 1270 if (GCD < OrigElt.getSizeInBits()) 1271 return LLT::scalarOrVector(ElementCount::get(1, OrigTy.isScalable()), 1272 GCD); 1273 1274 return LLT::vector( 1275 ElementCount::get(GCD / OrigElt.getSizeInBits().getFixedValue(), 1276 OrigTy.isScalable()), 1277 OrigElt); 1278 } 1279 1280 // If one type is vector and the element size matches the scalar size, then 1281 // the gcd is the scalar type. 1282 if (OrigTy.isVector() && 1283 OrigTy.getElementType().getSizeInBits() == TargetTy.getSizeInBits()) 1284 return OrigTy.getElementType(); 1285 if (TargetTy.isVector() && 1286 TargetTy.getElementType().getSizeInBits() == OrigTy.getSizeInBits()) 1287 return OrigTy; 1288 1289 // At this point, both types are either scalars of different type or one is a 1290 // vector and one is a scalar. If both types are scalars, the GCD type is the 1291 // GCD between the two scalar sizes. If one is vector and one is scalar, then 1292 // the GCD type is the GCD between the scalar and the vector element size. 1293 LLT OrigScalar = OrigTy.getScalarType(); 1294 LLT TargetScalar = TargetTy.getScalarType(); 1295 unsigned GCD = std::gcd(OrigScalar.getSizeInBits().getFixedValue(), 1296 TargetScalar.getSizeInBits().getFixedValue()); 1297 return LLT::scalar(GCD); 1298 } 1299 1300 std::optional<int> llvm::getSplatIndex(MachineInstr &MI) { 1301 assert(MI.getOpcode() == TargetOpcode::G_SHUFFLE_VECTOR && 1302 "Only G_SHUFFLE_VECTOR can have a splat index!"); 1303 ArrayRef<int> Mask = MI.getOperand(3).getShuffleMask(); 1304 auto FirstDefinedIdx = find_if(Mask, [](int Elt) { return Elt >= 0; }); 1305 1306 // If all elements are undefined, this shuffle can be considered a splat. 1307 // Return 0 for better potential for callers to simplify. 1308 if (FirstDefinedIdx == Mask.end()) 1309 return 0; 1310 1311 // Make sure all remaining elements are either undef or the same 1312 // as the first non-undef value. 1313 int SplatValue = *FirstDefinedIdx; 1314 if (any_of(make_range(std::next(FirstDefinedIdx), Mask.end()), 1315 [&SplatValue](int Elt) { return Elt >= 0 && Elt != SplatValue; })) 1316 return std::nullopt; 1317 1318 return SplatValue; 1319 } 1320 1321 static bool isBuildVectorOp(unsigned Opcode) { 1322 return Opcode == TargetOpcode::G_BUILD_VECTOR || 1323 Opcode == TargetOpcode::G_BUILD_VECTOR_TRUNC; 1324 } 1325 1326 namespace { 1327 1328 std::optional<ValueAndVReg> getAnyConstantSplat(Register VReg, 1329 const MachineRegisterInfo &MRI, 1330 bool AllowUndef) { 1331 MachineInstr *MI = getDefIgnoringCopies(VReg, MRI); 1332 if (!MI) 1333 return std::nullopt; 1334 1335 bool isConcatVectorsOp = MI->getOpcode() == TargetOpcode::G_CONCAT_VECTORS; 1336 if (!isBuildVectorOp(MI->getOpcode()) && !isConcatVectorsOp) 1337 return std::nullopt; 1338 1339 std::optional<ValueAndVReg> SplatValAndReg; 1340 for (MachineOperand &Op : MI->uses()) { 1341 Register Element = Op.getReg(); 1342 // If we have a G_CONCAT_VECTOR, we recursively look into the 1343 // vectors that we're concatenating to see if they're splats. 1344 auto ElementValAndReg = 1345 isConcatVectorsOp 1346 ? getAnyConstantSplat(Element, MRI, AllowUndef) 1347 : getAnyConstantVRegValWithLookThrough(Element, MRI, true, true); 1348 1349 // If AllowUndef, treat undef as value that will result in a constant splat. 1350 if (!ElementValAndReg) { 1351 if (AllowUndef && isa<GImplicitDef>(MRI.getVRegDef(Element))) 1352 continue; 1353 return std::nullopt; 1354 } 1355 1356 // Record splat value 1357 if (!SplatValAndReg) 1358 SplatValAndReg = ElementValAndReg; 1359 1360 // Different constant than the one already recorded, not a constant splat. 1361 if (SplatValAndReg->Value != ElementValAndReg->Value) 1362 return std::nullopt; 1363 } 1364 1365 return SplatValAndReg; 1366 } 1367 1368 } // end anonymous namespace 1369 1370 bool llvm::isBuildVectorConstantSplat(const Register Reg, 1371 const MachineRegisterInfo &MRI, 1372 int64_t SplatValue, bool AllowUndef) { 1373 if (auto SplatValAndReg = getAnyConstantSplat(Reg, MRI, AllowUndef)) 1374 return mi_match(SplatValAndReg->VReg, MRI, m_SpecificICst(SplatValue)); 1375 return false; 1376 } 1377 1378 bool llvm::isBuildVectorConstantSplat(const MachineInstr &MI, 1379 const MachineRegisterInfo &MRI, 1380 int64_t SplatValue, bool AllowUndef) { 1381 return isBuildVectorConstantSplat(MI.getOperand(0).getReg(), MRI, SplatValue, 1382 AllowUndef); 1383 } 1384 1385 std::optional<APInt> 1386 llvm::getIConstantSplatVal(const Register Reg, const MachineRegisterInfo &MRI) { 1387 if (auto SplatValAndReg = 1388 getAnyConstantSplat(Reg, MRI, /* AllowUndef */ false)) { 1389 if (std::optional<ValueAndVReg> ValAndVReg = 1390 getIConstantVRegValWithLookThrough(SplatValAndReg->VReg, MRI)) 1391 return ValAndVReg->Value; 1392 } 1393 1394 return std::nullopt; 1395 } 1396 1397 std::optional<APInt> 1398 llvm::getIConstantSplatVal(const MachineInstr &MI, 1399 const MachineRegisterInfo &MRI) { 1400 return getIConstantSplatVal(MI.getOperand(0).getReg(), MRI); 1401 } 1402 1403 std::optional<int64_t> 1404 llvm::getIConstantSplatSExtVal(const Register Reg, 1405 const MachineRegisterInfo &MRI) { 1406 if (auto SplatValAndReg = 1407 getAnyConstantSplat(Reg, MRI, /* AllowUndef */ false)) 1408 return getIConstantVRegSExtVal(SplatValAndReg->VReg, MRI); 1409 return std::nullopt; 1410 } 1411 1412 std::optional<int64_t> 1413 llvm::getIConstantSplatSExtVal(const MachineInstr &MI, 1414 const MachineRegisterInfo &MRI) { 1415 return getIConstantSplatSExtVal(MI.getOperand(0).getReg(), MRI); 1416 } 1417 1418 std::optional<FPValueAndVReg> 1419 llvm::getFConstantSplat(Register VReg, const MachineRegisterInfo &MRI, 1420 bool AllowUndef) { 1421 if (auto SplatValAndReg = getAnyConstantSplat(VReg, MRI, AllowUndef)) 1422 return getFConstantVRegValWithLookThrough(SplatValAndReg->VReg, MRI); 1423 return std::nullopt; 1424 } 1425 1426 bool llvm::isBuildVectorAllZeros(const MachineInstr &MI, 1427 const MachineRegisterInfo &MRI, 1428 bool AllowUndef) { 1429 return isBuildVectorConstantSplat(MI, MRI, 0, AllowUndef); 1430 } 1431 1432 bool llvm::isBuildVectorAllOnes(const MachineInstr &MI, 1433 const MachineRegisterInfo &MRI, 1434 bool AllowUndef) { 1435 return isBuildVectorConstantSplat(MI, MRI, -1, AllowUndef); 1436 } 1437 1438 std::optional<RegOrConstant> 1439 llvm::getVectorSplat(const MachineInstr &MI, const MachineRegisterInfo &MRI) { 1440 unsigned Opc = MI.getOpcode(); 1441 if (!isBuildVectorOp(Opc)) 1442 return std::nullopt; 1443 if (auto Splat = getIConstantSplatSExtVal(MI, MRI)) 1444 return RegOrConstant(*Splat); 1445 auto Reg = MI.getOperand(1).getReg(); 1446 if (any_of(drop_begin(MI.operands(), 2), 1447 [&Reg](const MachineOperand &Op) { return Op.getReg() != Reg; })) 1448 return std::nullopt; 1449 return RegOrConstant(Reg); 1450 } 1451 1452 static bool isConstantScalar(const MachineInstr &MI, 1453 const MachineRegisterInfo &MRI, 1454 bool AllowFP = true, 1455 bool AllowOpaqueConstants = true) { 1456 switch (MI.getOpcode()) { 1457 case TargetOpcode::G_CONSTANT: 1458 case TargetOpcode::G_IMPLICIT_DEF: 1459 return true; 1460 case TargetOpcode::G_FCONSTANT: 1461 return AllowFP; 1462 case TargetOpcode::G_GLOBAL_VALUE: 1463 case TargetOpcode::G_FRAME_INDEX: 1464 case TargetOpcode::G_BLOCK_ADDR: 1465 case TargetOpcode::G_JUMP_TABLE: 1466 return AllowOpaqueConstants; 1467 default: 1468 return false; 1469 } 1470 } 1471 1472 bool llvm::isConstantOrConstantVector(MachineInstr &MI, 1473 const MachineRegisterInfo &MRI) { 1474 Register Def = MI.getOperand(0).getReg(); 1475 if (auto C = getIConstantVRegValWithLookThrough(Def, MRI)) 1476 return true; 1477 GBuildVector *BV = dyn_cast<GBuildVector>(&MI); 1478 if (!BV) 1479 return false; 1480 for (unsigned SrcIdx = 0; SrcIdx < BV->getNumSources(); ++SrcIdx) { 1481 if (getIConstantVRegValWithLookThrough(BV->getSourceReg(SrcIdx), MRI) || 1482 getOpcodeDef<GImplicitDef>(BV->getSourceReg(SrcIdx), MRI)) 1483 continue; 1484 return false; 1485 } 1486 return true; 1487 } 1488 1489 bool llvm::isConstantOrConstantVector(const MachineInstr &MI, 1490 const MachineRegisterInfo &MRI, 1491 bool AllowFP, bool AllowOpaqueConstants) { 1492 if (isConstantScalar(MI, MRI, AllowFP, AllowOpaqueConstants)) 1493 return true; 1494 1495 if (!isBuildVectorOp(MI.getOpcode())) 1496 return false; 1497 1498 const unsigned NumOps = MI.getNumOperands(); 1499 for (unsigned I = 1; I != NumOps; ++I) { 1500 const MachineInstr *ElementDef = MRI.getVRegDef(MI.getOperand(I).getReg()); 1501 if (!isConstantScalar(*ElementDef, MRI, AllowFP, AllowOpaqueConstants)) 1502 return false; 1503 } 1504 1505 return true; 1506 } 1507 1508 std::optional<APInt> 1509 llvm::isConstantOrConstantSplatVector(MachineInstr &MI, 1510 const MachineRegisterInfo &MRI) { 1511 Register Def = MI.getOperand(0).getReg(); 1512 if (auto C = getIConstantVRegValWithLookThrough(Def, MRI)) 1513 return C->Value; 1514 auto MaybeCst = getIConstantSplatSExtVal(MI, MRI); 1515 if (!MaybeCst) 1516 return std::nullopt; 1517 const unsigned ScalarSize = MRI.getType(Def).getScalarSizeInBits(); 1518 return APInt(ScalarSize, *MaybeCst, true); 1519 } 1520 1521 bool llvm::isNullOrNullSplat(const MachineInstr &MI, 1522 const MachineRegisterInfo &MRI, bool AllowUndefs) { 1523 switch (MI.getOpcode()) { 1524 case TargetOpcode::G_IMPLICIT_DEF: 1525 return AllowUndefs; 1526 case TargetOpcode::G_CONSTANT: 1527 return MI.getOperand(1).getCImm()->isNullValue(); 1528 case TargetOpcode::G_FCONSTANT: { 1529 const ConstantFP *FPImm = MI.getOperand(1).getFPImm(); 1530 return FPImm->isZero() && !FPImm->isNegative(); 1531 } 1532 default: 1533 if (!AllowUndefs) // TODO: isBuildVectorAllZeros assumes undef is OK already 1534 return false; 1535 return isBuildVectorAllZeros(MI, MRI); 1536 } 1537 } 1538 1539 bool llvm::isAllOnesOrAllOnesSplat(const MachineInstr &MI, 1540 const MachineRegisterInfo &MRI, 1541 bool AllowUndefs) { 1542 switch (MI.getOpcode()) { 1543 case TargetOpcode::G_IMPLICIT_DEF: 1544 return AllowUndefs; 1545 case TargetOpcode::G_CONSTANT: 1546 return MI.getOperand(1).getCImm()->isAllOnesValue(); 1547 default: 1548 if (!AllowUndefs) // TODO: isBuildVectorAllOnes assumes undef is OK already 1549 return false; 1550 return isBuildVectorAllOnes(MI, MRI); 1551 } 1552 } 1553 1554 bool llvm::matchUnaryPredicate( 1555 const MachineRegisterInfo &MRI, Register Reg, 1556 std::function<bool(const Constant *ConstVal)> Match, bool AllowUndefs) { 1557 1558 const MachineInstr *Def = getDefIgnoringCopies(Reg, MRI); 1559 if (AllowUndefs && Def->getOpcode() == TargetOpcode::G_IMPLICIT_DEF) 1560 return Match(nullptr); 1561 1562 // TODO: Also handle fconstant 1563 if (Def->getOpcode() == TargetOpcode::G_CONSTANT) 1564 return Match(Def->getOperand(1).getCImm()); 1565 1566 if (Def->getOpcode() != TargetOpcode::G_BUILD_VECTOR) 1567 return false; 1568 1569 for (unsigned I = 1, E = Def->getNumOperands(); I != E; ++I) { 1570 Register SrcElt = Def->getOperand(I).getReg(); 1571 const MachineInstr *SrcDef = getDefIgnoringCopies(SrcElt, MRI); 1572 if (AllowUndefs && SrcDef->getOpcode() == TargetOpcode::G_IMPLICIT_DEF) { 1573 if (!Match(nullptr)) 1574 return false; 1575 continue; 1576 } 1577 1578 if (SrcDef->getOpcode() != TargetOpcode::G_CONSTANT || 1579 !Match(SrcDef->getOperand(1).getCImm())) 1580 return false; 1581 } 1582 1583 return true; 1584 } 1585 1586 bool llvm::isConstTrueVal(const TargetLowering &TLI, int64_t Val, bool IsVector, 1587 bool IsFP) { 1588 switch (TLI.getBooleanContents(IsVector, IsFP)) { 1589 case TargetLowering::UndefinedBooleanContent: 1590 return Val & 0x1; 1591 case TargetLowering::ZeroOrOneBooleanContent: 1592 return Val == 1; 1593 case TargetLowering::ZeroOrNegativeOneBooleanContent: 1594 return Val == -1; 1595 } 1596 llvm_unreachable("Invalid boolean contents"); 1597 } 1598 1599 bool llvm::isConstFalseVal(const TargetLowering &TLI, int64_t Val, 1600 bool IsVector, bool IsFP) { 1601 switch (TLI.getBooleanContents(IsVector, IsFP)) { 1602 case TargetLowering::UndefinedBooleanContent: 1603 return ~Val & 0x1; 1604 case TargetLowering::ZeroOrOneBooleanContent: 1605 case TargetLowering::ZeroOrNegativeOneBooleanContent: 1606 return Val == 0; 1607 } 1608 llvm_unreachable("Invalid boolean contents"); 1609 } 1610 1611 int64_t llvm::getICmpTrueVal(const TargetLowering &TLI, bool IsVector, 1612 bool IsFP) { 1613 switch (TLI.getBooleanContents(IsVector, IsFP)) { 1614 case TargetLowering::UndefinedBooleanContent: 1615 case TargetLowering::ZeroOrOneBooleanContent: 1616 return 1; 1617 case TargetLowering::ZeroOrNegativeOneBooleanContent: 1618 return -1; 1619 } 1620 llvm_unreachable("Invalid boolean contents"); 1621 } 1622 1623 bool llvm::shouldOptForSize(const MachineBasicBlock &MBB, 1624 ProfileSummaryInfo *PSI, BlockFrequencyInfo *BFI) { 1625 const auto &F = MBB.getParent()->getFunction(); 1626 return F.hasOptSize() || F.hasMinSize() || 1627 llvm::shouldOptimizeForSize(MBB.getBasicBlock(), PSI, BFI); 1628 } 1629 1630 void llvm::saveUsesAndErase(MachineInstr &MI, MachineRegisterInfo &MRI, 1631 LostDebugLocObserver *LocObserver, 1632 SmallInstListTy &DeadInstChain) { 1633 for (MachineOperand &Op : MI.uses()) { 1634 if (Op.isReg() && Op.getReg().isVirtual()) 1635 DeadInstChain.insert(MRI.getVRegDef(Op.getReg())); 1636 } 1637 LLVM_DEBUG(dbgs() << MI << "Is dead; erasing.\n"); 1638 DeadInstChain.remove(&MI); 1639 MI.eraseFromParent(); 1640 if (LocObserver) 1641 LocObserver->checkpoint(false); 1642 } 1643 1644 void llvm::eraseInstrs(ArrayRef<MachineInstr *> DeadInstrs, 1645 MachineRegisterInfo &MRI, 1646 LostDebugLocObserver *LocObserver) { 1647 SmallInstListTy DeadInstChain; 1648 for (MachineInstr *MI : DeadInstrs) 1649 saveUsesAndErase(*MI, MRI, LocObserver, DeadInstChain); 1650 1651 while (!DeadInstChain.empty()) { 1652 MachineInstr *Inst = DeadInstChain.pop_back_val(); 1653 if (!isTriviallyDead(*Inst, MRI)) 1654 continue; 1655 saveUsesAndErase(*Inst, MRI, LocObserver, DeadInstChain); 1656 } 1657 } 1658 1659 void llvm::eraseInstr(MachineInstr &MI, MachineRegisterInfo &MRI, 1660 LostDebugLocObserver *LocObserver) { 1661 return eraseInstrs({&MI}, MRI, LocObserver); 1662 } 1663 1664 void llvm::salvageDebugInfo(const MachineRegisterInfo &MRI, MachineInstr &MI) { 1665 for (auto &Def : MI.defs()) { 1666 assert(Def.isReg() && "Must be a reg"); 1667 1668 SmallVector<MachineOperand *, 16> DbgUsers; 1669 for (auto &MOUse : MRI.use_operands(Def.getReg())) { 1670 MachineInstr *DbgValue = MOUse.getParent(); 1671 // Ignore partially formed DBG_VALUEs. 1672 if (DbgValue->isNonListDebugValue() && DbgValue->getNumOperands() == 4) { 1673 DbgUsers.push_back(&MOUse); 1674 } 1675 } 1676 1677 if (!DbgUsers.empty()) { 1678 salvageDebugInfoForDbgValue(MRI, MI, DbgUsers); 1679 } 1680 } 1681 } 1682 1683 bool llvm::isPreISelGenericFloatingPointOpcode(unsigned Opc) { 1684 switch (Opc) { 1685 case TargetOpcode::G_FABS: 1686 case TargetOpcode::G_FADD: 1687 case TargetOpcode::G_FCANONICALIZE: 1688 case TargetOpcode::G_FCEIL: 1689 case TargetOpcode::G_FCONSTANT: 1690 case TargetOpcode::G_FCOPYSIGN: 1691 case TargetOpcode::G_FCOS: 1692 case TargetOpcode::G_FDIV: 1693 case TargetOpcode::G_FEXP2: 1694 case TargetOpcode::G_FEXP: 1695 case TargetOpcode::G_FFLOOR: 1696 case TargetOpcode::G_FLOG10: 1697 case TargetOpcode::G_FLOG2: 1698 case TargetOpcode::G_FLOG: 1699 case TargetOpcode::G_FMA: 1700 case TargetOpcode::G_FMAD: 1701 case TargetOpcode::G_FMAXIMUM: 1702 case TargetOpcode::G_FMAXNUM: 1703 case TargetOpcode::G_FMAXNUM_IEEE: 1704 case TargetOpcode::G_FMINIMUM: 1705 case TargetOpcode::G_FMINNUM: 1706 case TargetOpcode::G_FMINNUM_IEEE: 1707 case TargetOpcode::G_FMUL: 1708 case TargetOpcode::G_FNEARBYINT: 1709 case TargetOpcode::G_FNEG: 1710 case TargetOpcode::G_FPEXT: 1711 case TargetOpcode::G_FPOW: 1712 case TargetOpcode::G_FPTRUNC: 1713 case TargetOpcode::G_FREM: 1714 case TargetOpcode::G_FRINT: 1715 case TargetOpcode::G_FSIN: 1716 case TargetOpcode::G_FSQRT: 1717 case TargetOpcode::G_FSUB: 1718 case TargetOpcode::G_INTRINSIC_ROUND: 1719 case TargetOpcode::G_INTRINSIC_ROUNDEVEN: 1720 case TargetOpcode::G_INTRINSIC_TRUNC: 1721 return true; 1722 default: 1723 return false; 1724 } 1725 } 1726 1727 namespace { 1728 enum class UndefPoisonKind { 1729 PoisonOnly = (1 << 0), 1730 UndefOnly = (1 << 1), 1731 UndefOrPoison = PoisonOnly | UndefOnly, 1732 }; 1733 } 1734 1735 [[maybe_unused]] static bool includesPoison(UndefPoisonKind Kind) { 1736 return (unsigned(Kind) & unsigned(UndefPoisonKind::PoisonOnly)) != 0; 1737 } 1738 1739 [[maybe_unused]] static bool includesUndef(UndefPoisonKind Kind) { 1740 return (unsigned(Kind) & unsigned(UndefPoisonKind::UndefOnly)) != 0; 1741 } 1742 1743 static bool canCreateUndefOrPoison(Register Reg, const MachineRegisterInfo &MRI, 1744 bool ConsiderFlagsAndMetadata, 1745 UndefPoisonKind Kind) { 1746 MachineInstr *RegDef = MRI.getVRegDef(Reg); 1747 1748 switch (RegDef->getOpcode()) { 1749 case TargetOpcode::G_FREEZE: 1750 return false; 1751 default: 1752 return true; 1753 } 1754 } 1755 1756 static bool isGuaranteedNotToBeUndefOrPoison(Register Reg, 1757 const MachineRegisterInfo &MRI, 1758 unsigned Depth, 1759 UndefPoisonKind Kind) { 1760 if (Depth >= MaxAnalysisRecursionDepth) 1761 return false; 1762 1763 MachineInstr *RegDef = MRI.getVRegDef(Reg); 1764 1765 switch (RegDef->getOpcode()) { 1766 case TargetOpcode::G_FREEZE: 1767 return true; 1768 case TargetOpcode::G_IMPLICIT_DEF: 1769 return !includesUndef(Kind); 1770 default: 1771 return false; 1772 } 1773 } 1774 1775 bool llvm::canCreateUndefOrPoison(Register Reg, const MachineRegisterInfo &MRI, 1776 bool ConsiderFlagsAndMetadata) { 1777 return ::canCreateUndefOrPoison(Reg, MRI, ConsiderFlagsAndMetadata, 1778 UndefPoisonKind::UndefOrPoison); 1779 } 1780 1781 bool canCreatePoison(Register Reg, const MachineRegisterInfo &MRI, 1782 bool ConsiderFlagsAndMetadata = true) { 1783 return ::canCreateUndefOrPoison(Reg, MRI, ConsiderFlagsAndMetadata, 1784 UndefPoisonKind::PoisonOnly); 1785 } 1786 1787 bool llvm::isGuaranteedNotToBeUndefOrPoison(Register Reg, 1788 const MachineRegisterInfo &MRI, 1789 unsigned Depth) { 1790 return ::isGuaranteedNotToBeUndefOrPoison(Reg, MRI, Depth, 1791 UndefPoisonKind::UndefOrPoison); 1792 } 1793 1794 bool llvm::isGuaranteedNotToBePoison(Register Reg, 1795 const MachineRegisterInfo &MRI, 1796 unsigned Depth) { 1797 return ::isGuaranteedNotToBeUndefOrPoison(Reg, MRI, Depth, 1798 UndefPoisonKind::PoisonOnly); 1799 } 1800 1801 bool llvm::isGuaranteedNotToBeUndef(Register Reg, 1802 const MachineRegisterInfo &MRI, 1803 unsigned Depth) { 1804 return ::isGuaranteedNotToBeUndefOrPoison(Reg, MRI, Depth, 1805 UndefPoisonKind::UndefOnly); 1806 } 1807