1 //===- llvm/CodeGen/GlobalISel/Utils.cpp -------------------------*- C++ -*-==// 2 // 3 // Part of the LLVM Project, under the Apache License v2.0 with LLVM Exceptions. 4 // See https://llvm.org/LICENSE.txt for license information. 5 // SPDX-License-Identifier: Apache-2.0 WITH LLVM-exception 6 // 7 //===----------------------------------------------------------------------===// 8 /// \file This file implements the utility functions used by the GlobalISel 9 /// pipeline. 10 //===----------------------------------------------------------------------===// 11 12 #include "llvm/CodeGen/GlobalISel/Utils.h" 13 #include "llvm/ADT/APFloat.h" 14 #include "llvm/ADT/APInt.h" 15 #include "llvm/CodeGen/CodeGenCommonISel.h" 16 #include "llvm/CodeGen/GlobalISel/GISelChangeObserver.h" 17 #include "llvm/CodeGen/GlobalISel/GISelKnownBits.h" 18 #include "llvm/CodeGen/GlobalISel/GenericMachineInstrs.h" 19 #include "llvm/CodeGen/GlobalISel/LostDebugLocObserver.h" 20 #include "llvm/CodeGen/GlobalISel/MIPatternMatch.h" 21 #include "llvm/CodeGen/GlobalISel/MachineIRBuilder.h" 22 #include "llvm/CodeGen/MachineInstr.h" 23 #include "llvm/CodeGen/MachineInstrBuilder.h" 24 #include "llvm/CodeGen/MachineOptimizationRemarkEmitter.h" 25 #include "llvm/CodeGen/MachineRegisterInfo.h" 26 #include "llvm/CodeGen/MachineSizeOpts.h" 27 #include "llvm/CodeGen/RegisterBankInfo.h" 28 #include "llvm/CodeGen/StackProtector.h" 29 #include "llvm/CodeGen/TargetInstrInfo.h" 30 #include "llvm/CodeGen/TargetLowering.h" 31 #include "llvm/CodeGen/TargetPassConfig.h" 32 #include "llvm/CodeGen/TargetRegisterInfo.h" 33 #include "llvm/IR/Constants.h" 34 #include "llvm/Target/TargetMachine.h" 35 #include "llvm/Transforms/Utils/SizeOpts.h" 36 #include <numeric> 37 #include <optional> 38 39 #define DEBUG_TYPE "globalisel-utils" 40 41 using namespace llvm; 42 using namespace MIPatternMatch; 43 44 Register llvm::constrainRegToClass(MachineRegisterInfo &MRI, 45 const TargetInstrInfo &TII, 46 const RegisterBankInfo &RBI, Register Reg, 47 const TargetRegisterClass &RegClass) { 48 if (!RBI.constrainGenericRegister(Reg, RegClass, MRI)) 49 return MRI.createVirtualRegister(&RegClass); 50 51 return Reg; 52 } 53 54 Register llvm::constrainOperandRegClass( 55 const MachineFunction &MF, const TargetRegisterInfo &TRI, 56 MachineRegisterInfo &MRI, const TargetInstrInfo &TII, 57 const RegisterBankInfo &RBI, MachineInstr &InsertPt, 58 const TargetRegisterClass &RegClass, MachineOperand &RegMO) { 59 Register Reg = RegMO.getReg(); 60 // Assume physical registers are properly constrained. 61 assert(Reg.isVirtual() && "PhysReg not implemented"); 62 63 // Save the old register class to check whether 64 // the change notifications will be required. 65 // TODO: A better approach would be to pass 66 // the observers to constrainRegToClass(). 67 auto *OldRegClass = MRI.getRegClassOrNull(Reg); 68 Register ConstrainedReg = constrainRegToClass(MRI, TII, RBI, Reg, RegClass); 69 // If we created a new virtual register because the class is not compatible 70 // then create a copy between the new and the old register. 71 if (ConstrainedReg != Reg) { 72 MachineBasicBlock::iterator InsertIt(&InsertPt); 73 MachineBasicBlock &MBB = *InsertPt.getParent(); 74 // FIXME: The copy needs to have the classes constrained for its operands. 75 // Use operand's regbank to get the class for old register (Reg). 76 if (RegMO.isUse()) { 77 BuildMI(MBB, InsertIt, InsertPt.getDebugLoc(), 78 TII.get(TargetOpcode::COPY), ConstrainedReg) 79 .addReg(Reg); 80 } else { 81 assert(RegMO.isDef() && "Must be a definition"); 82 BuildMI(MBB, std::next(InsertIt), InsertPt.getDebugLoc(), 83 TII.get(TargetOpcode::COPY), Reg) 84 .addReg(ConstrainedReg); 85 } 86 if (GISelChangeObserver *Observer = MF.getObserver()) { 87 Observer->changingInstr(*RegMO.getParent()); 88 } 89 RegMO.setReg(ConstrainedReg); 90 if (GISelChangeObserver *Observer = MF.getObserver()) { 91 Observer->changedInstr(*RegMO.getParent()); 92 } 93 } else if (OldRegClass != MRI.getRegClassOrNull(Reg)) { 94 if (GISelChangeObserver *Observer = MF.getObserver()) { 95 if (!RegMO.isDef()) { 96 MachineInstr *RegDef = MRI.getVRegDef(Reg); 97 Observer->changedInstr(*RegDef); 98 } 99 Observer->changingAllUsesOfReg(MRI, Reg); 100 Observer->finishedChangingAllUsesOfReg(); 101 } 102 } 103 return ConstrainedReg; 104 } 105 106 Register llvm::constrainOperandRegClass( 107 const MachineFunction &MF, const TargetRegisterInfo &TRI, 108 MachineRegisterInfo &MRI, const TargetInstrInfo &TII, 109 const RegisterBankInfo &RBI, MachineInstr &InsertPt, const MCInstrDesc &II, 110 MachineOperand &RegMO, unsigned OpIdx) { 111 Register Reg = RegMO.getReg(); 112 // Assume physical registers are properly constrained. 113 assert(Reg.isVirtual() && "PhysReg not implemented"); 114 115 const TargetRegisterClass *OpRC = TII.getRegClass(II, OpIdx, &TRI, MF); 116 // Some of the target independent instructions, like COPY, may not impose any 117 // register class constraints on some of their operands: If it's a use, we can 118 // skip constraining as the instruction defining the register would constrain 119 // it. 120 121 if (OpRC) { 122 // Obtain the RC from incoming regbank if it is a proper sub-class. Operands 123 // can have multiple regbanks for a superclass that combine different 124 // register types (E.g., AMDGPU's VGPR and AGPR). The regbank ambiguity 125 // resolved by targets during regbankselect should not be overridden. 126 if (const auto *SubRC = TRI.getCommonSubClass( 127 OpRC, TRI.getConstrainedRegClassForOperand(RegMO, MRI))) 128 OpRC = SubRC; 129 130 OpRC = TRI.getAllocatableClass(OpRC); 131 } 132 133 if (!OpRC) { 134 assert((!isTargetSpecificOpcode(II.getOpcode()) || RegMO.isUse()) && 135 "Register class constraint is required unless either the " 136 "instruction is target independent or the operand is a use"); 137 // FIXME: Just bailing out like this here could be not enough, unless we 138 // expect the users of this function to do the right thing for PHIs and 139 // COPY: 140 // v1 = COPY v0 141 // v2 = COPY v1 142 // v1 here may end up not being constrained at all. Please notice that to 143 // reproduce the issue we likely need a destination pattern of a selection 144 // rule producing such extra copies, not just an input GMIR with them as 145 // every existing target using selectImpl handles copies before calling it 146 // and they never reach this function. 147 return Reg; 148 } 149 return constrainOperandRegClass(MF, TRI, MRI, TII, RBI, InsertPt, *OpRC, 150 RegMO); 151 } 152 153 bool llvm::constrainSelectedInstRegOperands(MachineInstr &I, 154 const TargetInstrInfo &TII, 155 const TargetRegisterInfo &TRI, 156 const RegisterBankInfo &RBI) { 157 assert(!isPreISelGenericOpcode(I.getOpcode()) && 158 "A selected instruction is expected"); 159 MachineBasicBlock &MBB = *I.getParent(); 160 MachineFunction &MF = *MBB.getParent(); 161 MachineRegisterInfo &MRI = MF.getRegInfo(); 162 163 for (unsigned OpI = 0, OpE = I.getNumExplicitOperands(); OpI != OpE; ++OpI) { 164 MachineOperand &MO = I.getOperand(OpI); 165 166 // There's nothing to be done on non-register operands. 167 if (!MO.isReg()) 168 continue; 169 170 LLVM_DEBUG(dbgs() << "Converting operand: " << MO << '\n'); 171 assert(MO.isReg() && "Unsupported non-reg operand"); 172 173 Register Reg = MO.getReg(); 174 // Physical registers don't need to be constrained. 175 if (Reg.isPhysical()) 176 continue; 177 178 // Register operands with a value of 0 (e.g. predicate operands) don't need 179 // to be constrained. 180 if (Reg == 0) 181 continue; 182 183 // If the operand is a vreg, we should constrain its regclass, and only 184 // insert COPYs if that's impossible. 185 // constrainOperandRegClass does that for us. 186 constrainOperandRegClass(MF, TRI, MRI, TII, RBI, I, I.getDesc(), MO, OpI); 187 188 // Tie uses to defs as indicated in MCInstrDesc if this hasn't already been 189 // done. 190 if (MO.isUse()) { 191 int DefIdx = I.getDesc().getOperandConstraint(OpI, MCOI::TIED_TO); 192 if (DefIdx != -1 && !I.isRegTiedToUseOperand(DefIdx)) 193 I.tieOperands(DefIdx, OpI); 194 } 195 } 196 return true; 197 } 198 199 bool llvm::canReplaceReg(Register DstReg, Register SrcReg, 200 MachineRegisterInfo &MRI) { 201 // Give up if either DstReg or SrcReg is a physical register. 202 if (DstReg.isPhysical() || SrcReg.isPhysical()) 203 return false; 204 // Give up if the types don't match. 205 if (MRI.getType(DstReg) != MRI.getType(SrcReg)) 206 return false; 207 // Replace if either DstReg has no constraints or the register 208 // constraints match. 209 const auto &DstRBC = MRI.getRegClassOrRegBank(DstReg); 210 if (!DstRBC || DstRBC == MRI.getRegClassOrRegBank(SrcReg)) 211 return true; 212 213 // Otherwise match if the Src is already a regclass that is covered by the Dst 214 // RegBank. 215 return DstRBC.is<const RegisterBank *>() && MRI.getRegClassOrNull(SrcReg) && 216 DstRBC.get<const RegisterBank *>()->covers( 217 *MRI.getRegClassOrNull(SrcReg)); 218 } 219 220 bool llvm::isTriviallyDead(const MachineInstr &MI, 221 const MachineRegisterInfo &MRI) { 222 // FIXME: This logical is mostly duplicated with 223 // DeadMachineInstructionElim::isDead. Why is LOCAL_ESCAPE not considered in 224 // MachineInstr::isLabel? 225 226 // Don't delete frame allocation labels. 227 if (MI.getOpcode() == TargetOpcode::LOCAL_ESCAPE) 228 return false; 229 // LIFETIME markers should be preserved even if they seem dead. 230 if (MI.getOpcode() == TargetOpcode::LIFETIME_START || 231 MI.getOpcode() == TargetOpcode::LIFETIME_END) 232 return false; 233 234 // If we can move an instruction, we can remove it. Otherwise, it has 235 // a side-effect of some sort. 236 bool SawStore = false; 237 if (!MI.isSafeToMove(/*AA=*/nullptr, SawStore) && !MI.isPHI()) 238 return false; 239 240 // Instructions without side-effects are dead iff they only define dead vregs. 241 for (const auto &MO : MI.all_defs()) { 242 Register Reg = MO.getReg(); 243 if (Reg.isPhysical() || !MRI.use_nodbg_empty(Reg)) 244 return false; 245 } 246 return true; 247 } 248 249 static void reportGISelDiagnostic(DiagnosticSeverity Severity, 250 MachineFunction &MF, 251 const TargetPassConfig &TPC, 252 MachineOptimizationRemarkEmitter &MORE, 253 MachineOptimizationRemarkMissed &R) { 254 bool IsFatal = Severity == DS_Error && 255 TPC.isGlobalISelAbortEnabled(); 256 // Print the function name explicitly if we don't have a debug location (which 257 // makes the diagnostic less useful) or if we're going to emit a raw error. 258 if (!R.getLocation().isValid() || IsFatal) 259 R << (" (in function: " + MF.getName() + ")").str(); 260 261 if (IsFatal) 262 report_fatal_error(Twine(R.getMsg())); 263 else 264 MORE.emit(R); 265 } 266 267 void llvm::reportGISelWarning(MachineFunction &MF, const TargetPassConfig &TPC, 268 MachineOptimizationRemarkEmitter &MORE, 269 MachineOptimizationRemarkMissed &R) { 270 reportGISelDiagnostic(DS_Warning, MF, TPC, MORE, R); 271 } 272 273 void llvm::reportGISelFailure(MachineFunction &MF, const TargetPassConfig &TPC, 274 MachineOptimizationRemarkEmitter &MORE, 275 MachineOptimizationRemarkMissed &R) { 276 MF.getProperties().set(MachineFunctionProperties::Property::FailedISel); 277 reportGISelDiagnostic(DS_Error, MF, TPC, MORE, R); 278 } 279 280 void llvm::reportGISelFailure(MachineFunction &MF, const TargetPassConfig &TPC, 281 MachineOptimizationRemarkEmitter &MORE, 282 const char *PassName, StringRef Msg, 283 const MachineInstr &MI) { 284 MachineOptimizationRemarkMissed R(PassName, "GISelFailure: ", 285 MI.getDebugLoc(), MI.getParent()); 286 R << Msg; 287 // Printing MI is expensive; only do it if expensive remarks are enabled. 288 if (TPC.isGlobalISelAbortEnabled() || MORE.allowExtraAnalysis(PassName)) 289 R << ": " << ore::MNV("Inst", MI); 290 reportGISelFailure(MF, TPC, MORE, R); 291 } 292 293 std::optional<APInt> llvm::getIConstantVRegVal(Register VReg, 294 const MachineRegisterInfo &MRI) { 295 std::optional<ValueAndVReg> ValAndVReg = getIConstantVRegValWithLookThrough( 296 VReg, MRI, /*LookThroughInstrs*/ false); 297 assert((!ValAndVReg || ValAndVReg->VReg == VReg) && 298 "Value found while looking through instrs"); 299 if (!ValAndVReg) 300 return std::nullopt; 301 return ValAndVReg->Value; 302 } 303 304 std::optional<int64_t> 305 llvm::getIConstantVRegSExtVal(Register VReg, const MachineRegisterInfo &MRI) { 306 std::optional<APInt> Val = getIConstantVRegVal(VReg, MRI); 307 if (Val && Val->getBitWidth() <= 64) 308 return Val->getSExtValue(); 309 return std::nullopt; 310 } 311 312 namespace { 313 314 typedef std::function<bool(const MachineInstr *)> IsOpcodeFn; 315 typedef std::function<std::optional<APInt>(const MachineInstr *MI)> GetAPCstFn; 316 317 std::optional<ValueAndVReg> getConstantVRegValWithLookThrough( 318 Register VReg, const MachineRegisterInfo &MRI, IsOpcodeFn IsConstantOpcode, 319 GetAPCstFn getAPCstValue, bool LookThroughInstrs = true, 320 bool LookThroughAnyExt = false) { 321 SmallVector<std::pair<unsigned, unsigned>, 4> SeenOpcodes; 322 MachineInstr *MI; 323 324 while ((MI = MRI.getVRegDef(VReg)) && !IsConstantOpcode(MI) && 325 LookThroughInstrs) { 326 switch (MI->getOpcode()) { 327 case TargetOpcode::G_ANYEXT: 328 if (!LookThroughAnyExt) 329 return std::nullopt; 330 [[fallthrough]]; 331 case TargetOpcode::G_TRUNC: 332 case TargetOpcode::G_SEXT: 333 case TargetOpcode::G_ZEXT: 334 SeenOpcodes.push_back(std::make_pair( 335 MI->getOpcode(), 336 MRI.getType(MI->getOperand(0).getReg()).getSizeInBits())); 337 VReg = MI->getOperand(1).getReg(); 338 break; 339 case TargetOpcode::COPY: 340 VReg = MI->getOperand(1).getReg(); 341 if (VReg.isPhysical()) 342 return std::nullopt; 343 break; 344 case TargetOpcode::G_INTTOPTR: 345 VReg = MI->getOperand(1).getReg(); 346 break; 347 default: 348 return std::nullopt; 349 } 350 } 351 if (!MI || !IsConstantOpcode(MI)) 352 return std::nullopt; 353 354 std::optional<APInt> MaybeVal = getAPCstValue(MI); 355 if (!MaybeVal) 356 return std::nullopt; 357 APInt &Val = *MaybeVal; 358 while (!SeenOpcodes.empty()) { 359 std::pair<unsigned, unsigned> OpcodeAndSize = SeenOpcodes.pop_back_val(); 360 switch (OpcodeAndSize.first) { 361 case TargetOpcode::G_TRUNC: 362 Val = Val.trunc(OpcodeAndSize.second); 363 break; 364 case TargetOpcode::G_ANYEXT: 365 case TargetOpcode::G_SEXT: 366 Val = Val.sext(OpcodeAndSize.second); 367 break; 368 case TargetOpcode::G_ZEXT: 369 Val = Val.zext(OpcodeAndSize.second); 370 break; 371 } 372 } 373 374 return ValueAndVReg{Val, VReg}; 375 } 376 377 bool isIConstant(const MachineInstr *MI) { 378 if (!MI) 379 return false; 380 return MI->getOpcode() == TargetOpcode::G_CONSTANT; 381 } 382 383 bool isFConstant(const MachineInstr *MI) { 384 if (!MI) 385 return false; 386 return MI->getOpcode() == TargetOpcode::G_FCONSTANT; 387 } 388 389 bool isAnyConstant(const MachineInstr *MI) { 390 if (!MI) 391 return false; 392 unsigned Opc = MI->getOpcode(); 393 return Opc == TargetOpcode::G_CONSTANT || Opc == TargetOpcode::G_FCONSTANT; 394 } 395 396 std::optional<APInt> getCImmAsAPInt(const MachineInstr *MI) { 397 const MachineOperand &CstVal = MI->getOperand(1); 398 if (CstVal.isCImm()) 399 return CstVal.getCImm()->getValue(); 400 return std::nullopt; 401 } 402 403 std::optional<APInt> getCImmOrFPImmAsAPInt(const MachineInstr *MI) { 404 const MachineOperand &CstVal = MI->getOperand(1); 405 if (CstVal.isCImm()) 406 return CstVal.getCImm()->getValue(); 407 if (CstVal.isFPImm()) 408 return CstVal.getFPImm()->getValueAPF().bitcastToAPInt(); 409 return std::nullopt; 410 } 411 412 } // end anonymous namespace 413 414 std::optional<ValueAndVReg> llvm::getIConstantVRegValWithLookThrough( 415 Register VReg, const MachineRegisterInfo &MRI, bool LookThroughInstrs) { 416 return getConstantVRegValWithLookThrough(VReg, MRI, isIConstant, 417 getCImmAsAPInt, LookThroughInstrs); 418 } 419 420 std::optional<ValueAndVReg> llvm::getAnyConstantVRegValWithLookThrough( 421 Register VReg, const MachineRegisterInfo &MRI, bool LookThroughInstrs, 422 bool LookThroughAnyExt) { 423 return getConstantVRegValWithLookThrough( 424 VReg, MRI, isAnyConstant, getCImmOrFPImmAsAPInt, LookThroughInstrs, 425 LookThroughAnyExt); 426 } 427 428 std::optional<FPValueAndVReg> llvm::getFConstantVRegValWithLookThrough( 429 Register VReg, const MachineRegisterInfo &MRI, bool LookThroughInstrs) { 430 auto Reg = getConstantVRegValWithLookThrough( 431 VReg, MRI, isFConstant, getCImmOrFPImmAsAPInt, LookThroughInstrs); 432 if (!Reg) 433 return std::nullopt; 434 return FPValueAndVReg{getConstantFPVRegVal(Reg->VReg, MRI)->getValueAPF(), 435 Reg->VReg}; 436 } 437 438 const ConstantFP * 439 llvm::getConstantFPVRegVal(Register VReg, const MachineRegisterInfo &MRI) { 440 MachineInstr *MI = MRI.getVRegDef(VReg); 441 if (TargetOpcode::G_FCONSTANT != MI->getOpcode()) 442 return nullptr; 443 return MI->getOperand(1).getFPImm(); 444 } 445 446 std::optional<DefinitionAndSourceRegister> 447 llvm::getDefSrcRegIgnoringCopies(Register Reg, const MachineRegisterInfo &MRI) { 448 Register DefSrcReg = Reg; 449 auto *DefMI = MRI.getVRegDef(Reg); 450 auto DstTy = MRI.getType(DefMI->getOperand(0).getReg()); 451 if (!DstTy.isValid()) 452 return std::nullopt; 453 unsigned Opc = DefMI->getOpcode(); 454 while (Opc == TargetOpcode::COPY || isPreISelGenericOptimizationHint(Opc)) { 455 Register SrcReg = DefMI->getOperand(1).getReg(); 456 auto SrcTy = MRI.getType(SrcReg); 457 if (!SrcTy.isValid()) 458 break; 459 DefMI = MRI.getVRegDef(SrcReg); 460 DefSrcReg = SrcReg; 461 Opc = DefMI->getOpcode(); 462 } 463 return DefinitionAndSourceRegister{DefMI, DefSrcReg}; 464 } 465 466 MachineInstr *llvm::getDefIgnoringCopies(Register Reg, 467 const MachineRegisterInfo &MRI) { 468 std::optional<DefinitionAndSourceRegister> DefSrcReg = 469 getDefSrcRegIgnoringCopies(Reg, MRI); 470 return DefSrcReg ? DefSrcReg->MI : nullptr; 471 } 472 473 Register llvm::getSrcRegIgnoringCopies(Register Reg, 474 const MachineRegisterInfo &MRI) { 475 std::optional<DefinitionAndSourceRegister> DefSrcReg = 476 getDefSrcRegIgnoringCopies(Reg, MRI); 477 return DefSrcReg ? DefSrcReg->Reg : Register(); 478 } 479 480 void llvm::extractParts(Register Reg, LLT Ty, int NumParts, 481 SmallVectorImpl<Register> &VRegs, 482 MachineIRBuilder &MIRBuilder, 483 MachineRegisterInfo &MRI) { 484 for (int i = 0; i < NumParts; ++i) 485 VRegs.push_back(MRI.createGenericVirtualRegister(Ty)); 486 MIRBuilder.buildUnmerge(VRegs, Reg); 487 } 488 489 bool llvm::extractParts(Register Reg, LLT RegTy, LLT MainTy, LLT &LeftoverTy, 490 SmallVectorImpl<Register> &VRegs, 491 SmallVectorImpl<Register> &LeftoverRegs, 492 MachineIRBuilder &MIRBuilder, 493 MachineRegisterInfo &MRI) { 494 assert(!LeftoverTy.isValid() && "this is an out argument"); 495 496 unsigned RegSize = RegTy.getSizeInBits(); 497 unsigned MainSize = MainTy.getSizeInBits(); 498 unsigned NumParts = RegSize / MainSize; 499 unsigned LeftoverSize = RegSize - NumParts * MainSize; 500 501 // Use an unmerge when possible. 502 if (LeftoverSize == 0) { 503 for (unsigned I = 0; I < NumParts; ++I) 504 VRegs.push_back(MRI.createGenericVirtualRegister(MainTy)); 505 MIRBuilder.buildUnmerge(VRegs, Reg); 506 return true; 507 } 508 509 // Try to use unmerge for irregular vector split where possible 510 // For example when splitting a <6 x i32> into <4 x i32> with <2 x i32> 511 // leftover, it becomes: 512 // <2 x i32> %2, <2 x i32>%3, <2 x i32> %4 = G_UNMERGE_VALUE <6 x i32> %1 513 // <4 x i32> %5 = G_CONCAT_VECTOR <2 x i32> %2, <2 x i32> %3 514 if (RegTy.isVector() && MainTy.isVector()) { 515 unsigned RegNumElts = RegTy.getNumElements(); 516 unsigned MainNumElts = MainTy.getNumElements(); 517 unsigned LeftoverNumElts = RegNumElts % MainNumElts; 518 // If can unmerge to LeftoverTy, do it 519 if (MainNumElts % LeftoverNumElts == 0 && 520 RegNumElts % LeftoverNumElts == 0 && 521 RegTy.getScalarSizeInBits() == MainTy.getScalarSizeInBits() && 522 LeftoverNumElts > 1) { 523 LeftoverTy = 524 LLT::fixed_vector(LeftoverNumElts, RegTy.getScalarSizeInBits()); 525 526 // Unmerge the SrcReg to LeftoverTy vectors 527 SmallVector<Register, 4> UnmergeValues; 528 extractParts(Reg, LeftoverTy, RegNumElts / LeftoverNumElts, UnmergeValues, 529 MIRBuilder, MRI); 530 531 // Find how many LeftoverTy makes one MainTy 532 unsigned LeftoverPerMain = MainNumElts / LeftoverNumElts; 533 unsigned NumOfLeftoverVal = 534 ((RegNumElts % MainNumElts) / LeftoverNumElts); 535 536 // Create as many MainTy as possible using unmerged value 537 SmallVector<Register, 4> MergeValues; 538 for (unsigned I = 0; I < UnmergeValues.size() - NumOfLeftoverVal; I++) { 539 MergeValues.push_back(UnmergeValues[I]); 540 if (MergeValues.size() == LeftoverPerMain) { 541 VRegs.push_back( 542 MIRBuilder.buildMergeLikeInstr(MainTy, MergeValues).getReg(0)); 543 MergeValues.clear(); 544 } 545 } 546 // Populate LeftoverRegs with the leftovers 547 for (unsigned I = UnmergeValues.size() - NumOfLeftoverVal; 548 I < UnmergeValues.size(); I++) { 549 LeftoverRegs.push_back(UnmergeValues[I]); 550 } 551 return true; 552 } 553 } 554 // Perform irregular split. Leftover is last element of RegPieces. 555 if (MainTy.isVector()) { 556 SmallVector<Register, 8> RegPieces; 557 extractVectorParts(Reg, MainTy.getNumElements(), RegPieces, MIRBuilder, 558 MRI); 559 for (unsigned i = 0; i < RegPieces.size() - 1; ++i) 560 VRegs.push_back(RegPieces[i]); 561 LeftoverRegs.push_back(RegPieces[RegPieces.size() - 1]); 562 LeftoverTy = MRI.getType(LeftoverRegs[0]); 563 return true; 564 } 565 566 LeftoverTy = LLT::scalar(LeftoverSize); 567 // For irregular sizes, extract the individual parts. 568 for (unsigned I = 0; I != NumParts; ++I) { 569 Register NewReg = MRI.createGenericVirtualRegister(MainTy); 570 VRegs.push_back(NewReg); 571 MIRBuilder.buildExtract(NewReg, Reg, MainSize * I); 572 } 573 574 for (unsigned Offset = MainSize * NumParts; Offset < RegSize; 575 Offset += LeftoverSize) { 576 Register NewReg = MRI.createGenericVirtualRegister(LeftoverTy); 577 LeftoverRegs.push_back(NewReg); 578 MIRBuilder.buildExtract(NewReg, Reg, Offset); 579 } 580 581 return true; 582 } 583 584 void llvm::extractVectorParts(Register Reg, unsigned NumElts, 585 SmallVectorImpl<Register> &VRegs, 586 MachineIRBuilder &MIRBuilder, 587 MachineRegisterInfo &MRI) { 588 LLT RegTy = MRI.getType(Reg); 589 assert(RegTy.isVector() && "Expected a vector type"); 590 591 LLT EltTy = RegTy.getElementType(); 592 LLT NarrowTy = (NumElts == 1) ? EltTy : LLT::fixed_vector(NumElts, EltTy); 593 unsigned RegNumElts = RegTy.getNumElements(); 594 unsigned LeftoverNumElts = RegNumElts % NumElts; 595 unsigned NumNarrowTyPieces = RegNumElts / NumElts; 596 597 // Perfect split without leftover 598 if (LeftoverNumElts == 0) 599 return extractParts(Reg, NarrowTy, NumNarrowTyPieces, VRegs, MIRBuilder, 600 MRI); 601 602 // Irregular split. Provide direct access to all elements for artifact 603 // combiner using unmerge to elements. Then build vectors with NumElts 604 // elements. Remaining element(s) will be (used to build vector) Leftover. 605 SmallVector<Register, 8> Elts; 606 extractParts(Reg, EltTy, RegNumElts, Elts, MIRBuilder, MRI); 607 608 unsigned Offset = 0; 609 // Requested sub-vectors of NarrowTy. 610 for (unsigned i = 0; i < NumNarrowTyPieces; ++i, Offset += NumElts) { 611 ArrayRef<Register> Pieces(&Elts[Offset], NumElts); 612 VRegs.push_back(MIRBuilder.buildMergeLikeInstr(NarrowTy, Pieces).getReg(0)); 613 } 614 615 // Leftover element(s). 616 if (LeftoverNumElts == 1) { 617 VRegs.push_back(Elts[Offset]); 618 } else { 619 LLT LeftoverTy = LLT::fixed_vector(LeftoverNumElts, EltTy); 620 ArrayRef<Register> Pieces(&Elts[Offset], LeftoverNumElts); 621 VRegs.push_back( 622 MIRBuilder.buildMergeLikeInstr(LeftoverTy, Pieces).getReg(0)); 623 } 624 } 625 626 MachineInstr *llvm::getOpcodeDef(unsigned Opcode, Register Reg, 627 const MachineRegisterInfo &MRI) { 628 MachineInstr *DefMI = getDefIgnoringCopies(Reg, MRI); 629 return DefMI && DefMI->getOpcode() == Opcode ? DefMI : nullptr; 630 } 631 632 APFloat llvm::getAPFloatFromSize(double Val, unsigned Size) { 633 if (Size == 32) 634 return APFloat(float(Val)); 635 if (Size == 64) 636 return APFloat(Val); 637 if (Size != 16) 638 llvm_unreachable("Unsupported FPConstant size"); 639 bool Ignored; 640 APFloat APF(Val); 641 APF.convert(APFloat::IEEEhalf(), APFloat::rmNearestTiesToEven, &Ignored); 642 return APF; 643 } 644 645 std::optional<APInt> llvm::ConstantFoldBinOp(unsigned Opcode, 646 const Register Op1, 647 const Register Op2, 648 const MachineRegisterInfo &MRI) { 649 auto MaybeOp2Cst = getAnyConstantVRegValWithLookThrough(Op2, MRI, false); 650 if (!MaybeOp2Cst) 651 return std::nullopt; 652 653 auto MaybeOp1Cst = getAnyConstantVRegValWithLookThrough(Op1, MRI, false); 654 if (!MaybeOp1Cst) 655 return std::nullopt; 656 657 const APInt &C1 = MaybeOp1Cst->Value; 658 const APInt &C2 = MaybeOp2Cst->Value; 659 switch (Opcode) { 660 default: 661 break; 662 case TargetOpcode::G_ADD: 663 case TargetOpcode::G_PTR_ADD: 664 return C1 + C2; 665 case TargetOpcode::G_AND: 666 return C1 & C2; 667 case TargetOpcode::G_ASHR: 668 return C1.ashr(C2); 669 case TargetOpcode::G_LSHR: 670 return C1.lshr(C2); 671 case TargetOpcode::G_MUL: 672 return C1 * C2; 673 case TargetOpcode::G_OR: 674 return C1 | C2; 675 case TargetOpcode::G_SHL: 676 return C1 << C2; 677 case TargetOpcode::G_SUB: 678 return C1 - C2; 679 case TargetOpcode::G_XOR: 680 return C1 ^ C2; 681 case TargetOpcode::G_UDIV: 682 if (!C2.getBoolValue()) 683 break; 684 return C1.udiv(C2); 685 case TargetOpcode::G_SDIV: 686 if (!C2.getBoolValue()) 687 break; 688 return C1.sdiv(C2); 689 case TargetOpcode::G_UREM: 690 if (!C2.getBoolValue()) 691 break; 692 return C1.urem(C2); 693 case TargetOpcode::G_SREM: 694 if (!C2.getBoolValue()) 695 break; 696 return C1.srem(C2); 697 case TargetOpcode::G_SMIN: 698 return APIntOps::smin(C1, C2); 699 case TargetOpcode::G_SMAX: 700 return APIntOps::smax(C1, C2); 701 case TargetOpcode::G_UMIN: 702 return APIntOps::umin(C1, C2); 703 case TargetOpcode::G_UMAX: 704 return APIntOps::umax(C1, C2); 705 } 706 707 return std::nullopt; 708 } 709 710 std::optional<APFloat> 711 llvm::ConstantFoldFPBinOp(unsigned Opcode, const Register Op1, 712 const Register Op2, const MachineRegisterInfo &MRI) { 713 const ConstantFP *Op2Cst = getConstantFPVRegVal(Op2, MRI); 714 if (!Op2Cst) 715 return std::nullopt; 716 717 const ConstantFP *Op1Cst = getConstantFPVRegVal(Op1, MRI); 718 if (!Op1Cst) 719 return std::nullopt; 720 721 APFloat C1 = Op1Cst->getValueAPF(); 722 const APFloat &C2 = Op2Cst->getValueAPF(); 723 switch (Opcode) { 724 case TargetOpcode::G_FADD: 725 C1.add(C2, APFloat::rmNearestTiesToEven); 726 return C1; 727 case TargetOpcode::G_FSUB: 728 C1.subtract(C2, APFloat::rmNearestTiesToEven); 729 return C1; 730 case TargetOpcode::G_FMUL: 731 C1.multiply(C2, APFloat::rmNearestTiesToEven); 732 return C1; 733 case TargetOpcode::G_FDIV: 734 C1.divide(C2, APFloat::rmNearestTiesToEven); 735 return C1; 736 case TargetOpcode::G_FREM: 737 C1.mod(C2); 738 return C1; 739 case TargetOpcode::G_FCOPYSIGN: 740 C1.copySign(C2); 741 return C1; 742 case TargetOpcode::G_FMINNUM: 743 return minnum(C1, C2); 744 case TargetOpcode::G_FMAXNUM: 745 return maxnum(C1, C2); 746 case TargetOpcode::G_FMINIMUM: 747 return minimum(C1, C2); 748 case TargetOpcode::G_FMAXIMUM: 749 return maximum(C1, C2); 750 case TargetOpcode::G_FMINNUM_IEEE: 751 case TargetOpcode::G_FMAXNUM_IEEE: 752 // FIXME: These operations were unfortunately named. fminnum/fmaxnum do not 753 // follow the IEEE behavior for signaling nans and follow libm's fmin/fmax, 754 // and currently there isn't a nice wrapper in APFloat for the version with 755 // correct snan handling. 756 break; 757 default: 758 break; 759 } 760 761 return std::nullopt; 762 } 763 764 SmallVector<APInt> 765 llvm::ConstantFoldVectorBinop(unsigned Opcode, const Register Op1, 766 const Register Op2, 767 const MachineRegisterInfo &MRI) { 768 auto *SrcVec2 = getOpcodeDef<GBuildVector>(Op2, MRI); 769 if (!SrcVec2) 770 return SmallVector<APInt>(); 771 772 auto *SrcVec1 = getOpcodeDef<GBuildVector>(Op1, MRI); 773 if (!SrcVec1) 774 return SmallVector<APInt>(); 775 776 SmallVector<APInt> FoldedElements; 777 for (unsigned Idx = 0, E = SrcVec1->getNumSources(); Idx < E; ++Idx) { 778 auto MaybeCst = ConstantFoldBinOp(Opcode, SrcVec1->getSourceReg(Idx), 779 SrcVec2->getSourceReg(Idx), MRI); 780 if (!MaybeCst) 781 return SmallVector<APInt>(); 782 FoldedElements.push_back(*MaybeCst); 783 } 784 return FoldedElements; 785 } 786 787 bool llvm::isKnownNeverNaN(Register Val, const MachineRegisterInfo &MRI, 788 bool SNaN) { 789 const MachineInstr *DefMI = MRI.getVRegDef(Val); 790 if (!DefMI) 791 return false; 792 793 const TargetMachine& TM = DefMI->getMF()->getTarget(); 794 if (DefMI->getFlag(MachineInstr::FmNoNans) || TM.Options.NoNaNsFPMath) 795 return true; 796 797 // If the value is a constant, we can obviously see if it is a NaN or not. 798 if (const ConstantFP *FPVal = getConstantFPVRegVal(Val, MRI)) { 799 return !FPVal->getValueAPF().isNaN() || 800 (SNaN && !FPVal->getValueAPF().isSignaling()); 801 } 802 803 if (DefMI->getOpcode() == TargetOpcode::G_BUILD_VECTOR) { 804 for (const auto &Op : DefMI->uses()) 805 if (!isKnownNeverNaN(Op.getReg(), MRI, SNaN)) 806 return false; 807 return true; 808 } 809 810 switch (DefMI->getOpcode()) { 811 default: 812 break; 813 case TargetOpcode::G_FADD: 814 case TargetOpcode::G_FSUB: 815 case TargetOpcode::G_FMUL: 816 case TargetOpcode::G_FDIV: 817 case TargetOpcode::G_FREM: 818 case TargetOpcode::G_FSIN: 819 case TargetOpcode::G_FCOS: 820 case TargetOpcode::G_FMA: 821 case TargetOpcode::G_FMAD: 822 if (SNaN) 823 return true; 824 825 // TODO: Need isKnownNeverInfinity 826 return false; 827 case TargetOpcode::G_FMINNUM_IEEE: 828 case TargetOpcode::G_FMAXNUM_IEEE: { 829 if (SNaN) 830 return true; 831 // This can return a NaN if either operand is an sNaN, or if both operands 832 // are NaN. 833 return (isKnownNeverNaN(DefMI->getOperand(1).getReg(), MRI) && 834 isKnownNeverSNaN(DefMI->getOperand(2).getReg(), MRI)) || 835 (isKnownNeverSNaN(DefMI->getOperand(1).getReg(), MRI) && 836 isKnownNeverNaN(DefMI->getOperand(2).getReg(), MRI)); 837 } 838 case TargetOpcode::G_FMINNUM: 839 case TargetOpcode::G_FMAXNUM: { 840 // Only one needs to be known not-nan, since it will be returned if the 841 // other ends up being one. 842 return isKnownNeverNaN(DefMI->getOperand(1).getReg(), MRI, SNaN) || 843 isKnownNeverNaN(DefMI->getOperand(2).getReg(), MRI, SNaN); 844 } 845 } 846 847 if (SNaN) { 848 // FP operations quiet. For now, just handle the ones inserted during 849 // legalization. 850 switch (DefMI->getOpcode()) { 851 case TargetOpcode::G_FPEXT: 852 case TargetOpcode::G_FPTRUNC: 853 case TargetOpcode::G_FCANONICALIZE: 854 return true; 855 default: 856 return false; 857 } 858 } 859 860 return false; 861 } 862 863 Align llvm::inferAlignFromPtrInfo(MachineFunction &MF, 864 const MachinePointerInfo &MPO) { 865 auto PSV = dyn_cast_if_present<const PseudoSourceValue *>(MPO.V); 866 if (auto FSPV = dyn_cast_or_null<FixedStackPseudoSourceValue>(PSV)) { 867 MachineFrameInfo &MFI = MF.getFrameInfo(); 868 return commonAlignment(MFI.getObjectAlign(FSPV->getFrameIndex()), 869 MPO.Offset); 870 } 871 872 if (const Value *V = dyn_cast_if_present<const Value *>(MPO.V)) { 873 const Module *M = MF.getFunction().getParent(); 874 return V->getPointerAlignment(M->getDataLayout()); 875 } 876 877 return Align(1); 878 } 879 880 Register llvm::getFunctionLiveInPhysReg(MachineFunction &MF, 881 const TargetInstrInfo &TII, 882 MCRegister PhysReg, 883 const TargetRegisterClass &RC, 884 const DebugLoc &DL, LLT RegTy) { 885 MachineBasicBlock &EntryMBB = MF.front(); 886 MachineRegisterInfo &MRI = MF.getRegInfo(); 887 Register LiveIn = MRI.getLiveInVirtReg(PhysReg); 888 if (LiveIn) { 889 MachineInstr *Def = MRI.getVRegDef(LiveIn); 890 if (Def) { 891 // FIXME: Should the verifier check this is in the entry block? 892 assert(Def->getParent() == &EntryMBB && "live-in copy not in entry block"); 893 return LiveIn; 894 } 895 896 // It's possible the incoming argument register and copy was added during 897 // lowering, but later deleted due to being/becoming dead. If this happens, 898 // re-insert the copy. 899 } else { 900 // The live in register was not present, so add it. 901 LiveIn = MF.addLiveIn(PhysReg, &RC); 902 if (RegTy.isValid()) 903 MRI.setType(LiveIn, RegTy); 904 } 905 906 BuildMI(EntryMBB, EntryMBB.begin(), DL, TII.get(TargetOpcode::COPY), LiveIn) 907 .addReg(PhysReg); 908 if (!EntryMBB.isLiveIn(PhysReg)) 909 EntryMBB.addLiveIn(PhysReg); 910 return LiveIn; 911 } 912 913 std::optional<APInt> llvm::ConstantFoldExtOp(unsigned Opcode, 914 const Register Op1, uint64_t Imm, 915 const MachineRegisterInfo &MRI) { 916 auto MaybeOp1Cst = getIConstantVRegVal(Op1, MRI); 917 if (MaybeOp1Cst) { 918 switch (Opcode) { 919 default: 920 break; 921 case TargetOpcode::G_SEXT_INREG: { 922 LLT Ty = MRI.getType(Op1); 923 return MaybeOp1Cst->trunc(Imm).sext(Ty.getScalarSizeInBits()); 924 } 925 } 926 } 927 return std::nullopt; 928 } 929 930 std::optional<APInt> llvm::ConstantFoldCastOp(unsigned Opcode, LLT DstTy, 931 const Register Op0, 932 const MachineRegisterInfo &MRI) { 933 std::optional<APInt> Val = getIConstantVRegVal(Op0, MRI); 934 if (!Val) 935 return Val; 936 937 const unsigned DstSize = DstTy.getScalarSizeInBits(); 938 939 switch (Opcode) { 940 case TargetOpcode::G_SEXT: 941 return Val->sext(DstSize); 942 case TargetOpcode::G_ZEXT: 943 case TargetOpcode::G_ANYEXT: 944 // TODO: DAG considers target preference when constant folding any_extend. 945 return Val->zext(DstSize); 946 default: 947 break; 948 } 949 950 llvm_unreachable("unexpected cast opcode to constant fold"); 951 } 952 953 std::optional<APFloat> 954 llvm::ConstantFoldIntToFloat(unsigned Opcode, LLT DstTy, Register Src, 955 const MachineRegisterInfo &MRI) { 956 assert(Opcode == TargetOpcode::G_SITOFP || Opcode == TargetOpcode::G_UITOFP); 957 if (auto MaybeSrcVal = getIConstantVRegVal(Src, MRI)) { 958 APFloat DstVal(getFltSemanticForLLT(DstTy)); 959 DstVal.convertFromAPInt(*MaybeSrcVal, Opcode == TargetOpcode::G_SITOFP, 960 APFloat::rmNearestTiesToEven); 961 return DstVal; 962 } 963 return std::nullopt; 964 } 965 966 std::optional<SmallVector<unsigned>> 967 llvm::ConstantFoldCTLZ(Register Src, const MachineRegisterInfo &MRI) { 968 LLT Ty = MRI.getType(Src); 969 SmallVector<unsigned> FoldedCTLZs; 970 auto tryFoldScalar = [&](Register R) -> std::optional<unsigned> { 971 auto MaybeCst = getIConstantVRegVal(R, MRI); 972 if (!MaybeCst) 973 return std::nullopt; 974 return MaybeCst->countl_zero(); 975 }; 976 if (Ty.isVector()) { 977 // Try to constant fold each element. 978 auto *BV = getOpcodeDef<GBuildVector>(Src, MRI); 979 if (!BV) 980 return std::nullopt; 981 for (unsigned SrcIdx = 0; SrcIdx < BV->getNumSources(); ++SrcIdx) { 982 if (auto MaybeFold = tryFoldScalar(BV->getSourceReg(SrcIdx))) { 983 FoldedCTLZs.emplace_back(*MaybeFold); 984 continue; 985 } 986 return std::nullopt; 987 } 988 return FoldedCTLZs; 989 } 990 if (auto MaybeCst = tryFoldScalar(Src)) { 991 FoldedCTLZs.emplace_back(*MaybeCst); 992 return FoldedCTLZs; 993 } 994 return std::nullopt; 995 } 996 997 bool llvm::isKnownToBeAPowerOfTwo(Register Reg, const MachineRegisterInfo &MRI, 998 GISelKnownBits *KB) { 999 std::optional<DefinitionAndSourceRegister> DefSrcReg = 1000 getDefSrcRegIgnoringCopies(Reg, MRI); 1001 if (!DefSrcReg) 1002 return false; 1003 1004 const MachineInstr &MI = *DefSrcReg->MI; 1005 const LLT Ty = MRI.getType(Reg); 1006 1007 switch (MI.getOpcode()) { 1008 case TargetOpcode::G_CONSTANT: { 1009 unsigned BitWidth = Ty.getScalarSizeInBits(); 1010 const ConstantInt *CI = MI.getOperand(1).getCImm(); 1011 return CI->getValue().zextOrTrunc(BitWidth).isPowerOf2(); 1012 } 1013 case TargetOpcode::G_SHL: { 1014 // A left-shift of a constant one will have exactly one bit set because 1015 // shifting the bit off the end is undefined. 1016 1017 // TODO: Constant splat 1018 if (auto ConstLHS = getIConstantVRegVal(MI.getOperand(1).getReg(), MRI)) { 1019 if (*ConstLHS == 1) 1020 return true; 1021 } 1022 1023 break; 1024 } 1025 case TargetOpcode::G_LSHR: { 1026 if (auto ConstLHS = getIConstantVRegVal(MI.getOperand(1).getReg(), MRI)) { 1027 if (ConstLHS->isSignMask()) 1028 return true; 1029 } 1030 1031 break; 1032 } 1033 case TargetOpcode::G_BUILD_VECTOR: { 1034 // TODO: Probably should have a recursion depth guard since you could have 1035 // bitcasted vector elements. 1036 for (const MachineOperand &MO : llvm::drop_begin(MI.operands())) 1037 if (!isKnownToBeAPowerOfTwo(MO.getReg(), MRI, KB)) 1038 return false; 1039 1040 return true; 1041 } 1042 case TargetOpcode::G_BUILD_VECTOR_TRUNC: { 1043 // Only handle constants since we would need to know if number of leading 1044 // zeros is greater than the truncation amount. 1045 const unsigned BitWidth = Ty.getScalarSizeInBits(); 1046 for (const MachineOperand &MO : llvm::drop_begin(MI.operands())) { 1047 auto Const = getIConstantVRegVal(MO.getReg(), MRI); 1048 if (!Const || !Const->zextOrTrunc(BitWidth).isPowerOf2()) 1049 return false; 1050 } 1051 1052 return true; 1053 } 1054 default: 1055 break; 1056 } 1057 1058 if (!KB) 1059 return false; 1060 1061 // More could be done here, though the above checks are enough 1062 // to handle some common cases. 1063 1064 // Fall back to computeKnownBits to catch other known cases. 1065 KnownBits Known = KB->getKnownBits(Reg); 1066 return (Known.countMaxPopulation() == 1) && (Known.countMinPopulation() == 1); 1067 } 1068 1069 void llvm::getSelectionDAGFallbackAnalysisUsage(AnalysisUsage &AU) { 1070 AU.addPreserved<StackProtector>(); 1071 } 1072 1073 LLT llvm::getLCMType(LLT OrigTy, LLT TargetTy) { 1074 const unsigned OrigSize = OrigTy.getSizeInBits(); 1075 const unsigned TargetSize = TargetTy.getSizeInBits(); 1076 1077 if (OrigSize == TargetSize) 1078 return OrigTy; 1079 1080 if (OrigTy.isVector()) { 1081 const LLT OrigElt = OrigTy.getElementType(); 1082 1083 if (TargetTy.isVector()) { 1084 const LLT TargetElt = TargetTy.getElementType(); 1085 1086 if (OrigElt.getSizeInBits() == TargetElt.getSizeInBits()) { 1087 int GCDElts = 1088 std::gcd(OrigTy.getNumElements(), TargetTy.getNumElements()); 1089 // Prefer the original element type. 1090 ElementCount Mul = OrigTy.getElementCount() * TargetTy.getNumElements(); 1091 return LLT::vector(Mul.divideCoefficientBy(GCDElts), 1092 OrigTy.getElementType()); 1093 } 1094 } else { 1095 if (OrigElt.getSizeInBits() == TargetSize) 1096 return OrigTy; 1097 } 1098 1099 unsigned LCMSize = std::lcm(OrigSize, TargetSize); 1100 return LLT::fixed_vector(LCMSize / OrigElt.getSizeInBits(), OrigElt); 1101 } 1102 1103 if (TargetTy.isVector()) { 1104 unsigned LCMSize = std::lcm(OrigSize, TargetSize); 1105 return LLT::fixed_vector(LCMSize / OrigSize, OrigTy); 1106 } 1107 1108 unsigned LCMSize = std::lcm(OrigSize, TargetSize); 1109 1110 // Preserve pointer types. 1111 if (LCMSize == OrigSize) 1112 return OrigTy; 1113 if (LCMSize == TargetSize) 1114 return TargetTy; 1115 1116 return LLT::scalar(LCMSize); 1117 } 1118 1119 LLT llvm::getCoverTy(LLT OrigTy, LLT TargetTy) { 1120 if (!OrigTy.isVector() || !TargetTy.isVector() || OrigTy == TargetTy || 1121 (OrigTy.getScalarSizeInBits() != TargetTy.getScalarSizeInBits())) 1122 return getLCMType(OrigTy, TargetTy); 1123 1124 unsigned OrigTyNumElts = OrigTy.getNumElements(); 1125 unsigned TargetTyNumElts = TargetTy.getNumElements(); 1126 if (OrigTyNumElts % TargetTyNumElts == 0) 1127 return OrigTy; 1128 1129 unsigned NumElts = alignTo(OrigTyNumElts, TargetTyNumElts); 1130 return LLT::scalarOrVector(ElementCount::getFixed(NumElts), 1131 OrigTy.getElementType()); 1132 } 1133 1134 LLT llvm::getGCDType(LLT OrigTy, LLT TargetTy) { 1135 const unsigned OrigSize = OrigTy.getSizeInBits(); 1136 const unsigned TargetSize = TargetTy.getSizeInBits(); 1137 1138 if (OrigSize == TargetSize) 1139 return OrigTy; 1140 1141 if (OrigTy.isVector()) { 1142 LLT OrigElt = OrigTy.getElementType(); 1143 if (TargetTy.isVector()) { 1144 LLT TargetElt = TargetTy.getElementType(); 1145 if (OrigElt.getSizeInBits() == TargetElt.getSizeInBits()) { 1146 int GCD = std::gcd(OrigTy.getNumElements(), TargetTy.getNumElements()); 1147 return LLT::scalarOrVector(ElementCount::getFixed(GCD), OrigElt); 1148 } 1149 } else { 1150 // If the source is a vector of pointers, return a pointer element. 1151 if (OrigElt.getSizeInBits() == TargetSize) 1152 return OrigElt; 1153 } 1154 1155 unsigned GCD = std::gcd(OrigSize, TargetSize); 1156 if (GCD == OrigElt.getSizeInBits()) 1157 return OrigElt; 1158 1159 // If we can't produce the original element type, we have to use a smaller 1160 // scalar. 1161 if (GCD < OrigElt.getSizeInBits()) 1162 return LLT::scalar(GCD); 1163 return LLT::fixed_vector(GCD / OrigElt.getSizeInBits(), OrigElt); 1164 } 1165 1166 if (TargetTy.isVector()) { 1167 // Try to preserve the original element type. 1168 LLT TargetElt = TargetTy.getElementType(); 1169 if (TargetElt.getSizeInBits() == OrigSize) 1170 return OrigTy; 1171 } 1172 1173 unsigned GCD = std::gcd(OrigSize, TargetSize); 1174 return LLT::scalar(GCD); 1175 } 1176 1177 std::optional<int> llvm::getSplatIndex(MachineInstr &MI) { 1178 assert(MI.getOpcode() == TargetOpcode::G_SHUFFLE_VECTOR && 1179 "Only G_SHUFFLE_VECTOR can have a splat index!"); 1180 ArrayRef<int> Mask = MI.getOperand(3).getShuffleMask(); 1181 auto FirstDefinedIdx = find_if(Mask, [](int Elt) { return Elt >= 0; }); 1182 1183 // If all elements are undefined, this shuffle can be considered a splat. 1184 // Return 0 for better potential for callers to simplify. 1185 if (FirstDefinedIdx == Mask.end()) 1186 return 0; 1187 1188 // Make sure all remaining elements are either undef or the same 1189 // as the first non-undef value. 1190 int SplatValue = *FirstDefinedIdx; 1191 if (any_of(make_range(std::next(FirstDefinedIdx), Mask.end()), 1192 [&SplatValue](int Elt) { return Elt >= 0 && Elt != SplatValue; })) 1193 return std::nullopt; 1194 1195 return SplatValue; 1196 } 1197 1198 static bool isBuildVectorOp(unsigned Opcode) { 1199 return Opcode == TargetOpcode::G_BUILD_VECTOR || 1200 Opcode == TargetOpcode::G_BUILD_VECTOR_TRUNC; 1201 } 1202 1203 namespace { 1204 1205 std::optional<ValueAndVReg> getAnyConstantSplat(Register VReg, 1206 const MachineRegisterInfo &MRI, 1207 bool AllowUndef) { 1208 MachineInstr *MI = getDefIgnoringCopies(VReg, MRI); 1209 if (!MI) 1210 return std::nullopt; 1211 1212 bool isConcatVectorsOp = MI->getOpcode() == TargetOpcode::G_CONCAT_VECTORS; 1213 if (!isBuildVectorOp(MI->getOpcode()) && !isConcatVectorsOp) 1214 return std::nullopt; 1215 1216 std::optional<ValueAndVReg> SplatValAndReg; 1217 for (MachineOperand &Op : MI->uses()) { 1218 Register Element = Op.getReg(); 1219 // If we have a G_CONCAT_VECTOR, we recursively look into the 1220 // vectors that we're concatenating to see if they're splats. 1221 auto ElementValAndReg = 1222 isConcatVectorsOp 1223 ? getAnyConstantSplat(Element, MRI, AllowUndef) 1224 : getAnyConstantVRegValWithLookThrough(Element, MRI, true, true); 1225 1226 // If AllowUndef, treat undef as value that will result in a constant splat. 1227 if (!ElementValAndReg) { 1228 if (AllowUndef && isa<GImplicitDef>(MRI.getVRegDef(Element))) 1229 continue; 1230 return std::nullopt; 1231 } 1232 1233 // Record splat value 1234 if (!SplatValAndReg) 1235 SplatValAndReg = ElementValAndReg; 1236 1237 // Different constant than the one already recorded, not a constant splat. 1238 if (SplatValAndReg->Value != ElementValAndReg->Value) 1239 return std::nullopt; 1240 } 1241 1242 return SplatValAndReg; 1243 } 1244 1245 } // end anonymous namespace 1246 1247 bool llvm::isBuildVectorConstantSplat(const Register Reg, 1248 const MachineRegisterInfo &MRI, 1249 int64_t SplatValue, bool AllowUndef) { 1250 if (auto SplatValAndReg = getAnyConstantSplat(Reg, MRI, AllowUndef)) 1251 return mi_match(SplatValAndReg->VReg, MRI, m_SpecificICst(SplatValue)); 1252 return false; 1253 } 1254 1255 bool llvm::isBuildVectorConstantSplat(const MachineInstr &MI, 1256 const MachineRegisterInfo &MRI, 1257 int64_t SplatValue, bool AllowUndef) { 1258 return isBuildVectorConstantSplat(MI.getOperand(0).getReg(), MRI, SplatValue, 1259 AllowUndef); 1260 } 1261 1262 std::optional<APInt> 1263 llvm::getIConstantSplatVal(const Register Reg, const MachineRegisterInfo &MRI) { 1264 if (auto SplatValAndReg = 1265 getAnyConstantSplat(Reg, MRI, /* AllowUndef */ false)) { 1266 if (std::optional<ValueAndVReg> ValAndVReg = 1267 getIConstantVRegValWithLookThrough(SplatValAndReg->VReg, MRI)) 1268 return ValAndVReg->Value; 1269 } 1270 1271 return std::nullopt; 1272 } 1273 1274 std::optional<APInt> 1275 llvm::getIConstantSplatVal(const MachineInstr &MI, 1276 const MachineRegisterInfo &MRI) { 1277 return getIConstantSplatVal(MI.getOperand(0).getReg(), MRI); 1278 } 1279 1280 std::optional<int64_t> 1281 llvm::getIConstantSplatSExtVal(const Register Reg, 1282 const MachineRegisterInfo &MRI) { 1283 if (auto SplatValAndReg = 1284 getAnyConstantSplat(Reg, MRI, /* AllowUndef */ false)) 1285 return getIConstantVRegSExtVal(SplatValAndReg->VReg, MRI); 1286 return std::nullopt; 1287 } 1288 1289 std::optional<int64_t> 1290 llvm::getIConstantSplatSExtVal(const MachineInstr &MI, 1291 const MachineRegisterInfo &MRI) { 1292 return getIConstantSplatSExtVal(MI.getOperand(0).getReg(), MRI); 1293 } 1294 1295 std::optional<FPValueAndVReg> 1296 llvm::getFConstantSplat(Register VReg, const MachineRegisterInfo &MRI, 1297 bool AllowUndef) { 1298 if (auto SplatValAndReg = getAnyConstantSplat(VReg, MRI, AllowUndef)) 1299 return getFConstantVRegValWithLookThrough(SplatValAndReg->VReg, MRI); 1300 return std::nullopt; 1301 } 1302 1303 bool llvm::isBuildVectorAllZeros(const MachineInstr &MI, 1304 const MachineRegisterInfo &MRI, 1305 bool AllowUndef) { 1306 return isBuildVectorConstantSplat(MI, MRI, 0, AllowUndef); 1307 } 1308 1309 bool llvm::isBuildVectorAllOnes(const MachineInstr &MI, 1310 const MachineRegisterInfo &MRI, 1311 bool AllowUndef) { 1312 return isBuildVectorConstantSplat(MI, MRI, -1, AllowUndef); 1313 } 1314 1315 std::optional<RegOrConstant> 1316 llvm::getVectorSplat(const MachineInstr &MI, const MachineRegisterInfo &MRI) { 1317 unsigned Opc = MI.getOpcode(); 1318 if (!isBuildVectorOp(Opc)) 1319 return std::nullopt; 1320 if (auto Splat = getIConstantSplatSExtVal(MI, MRI)) 1321 return RegOrConstant(*Splat); 1322 auto Reg = MI.getOperand(1).getReg(); 1323 if (any_of(drop_begin(MI.operands(), 2), 1324 [&Reg](const MachineOperand &Op) { return Op.getReg() != Reg; })) 1325 return std::nullopt; 1326 return RegOrConstant(Reg); 1327 } 1328 1329 static bool isConstantScalar(const MachineInstr &MI, 1330 const MachineRegisterInfo &MRI, 1331 bool AllowFP = true, 1332 bool AllowOpaqueConstants = true) { 1333 switch (MI.getOpcode()) { 1334 case TargetOpcode::G_CONSTANT: 1335 case TargetOpcode::G_IMPLICIT_DEF: 1336 return true; 1337 case TargetOpcode::G_FCONSTANT: 1338 return AllowFP; 1339 case TargetOpcode::G_GLOBAL_VALUE: 1340 case TargetOpcode::G_FRAME_INDEX: 1341 case TargetOpcode::G_BLOCK_ADDR: 1342 case TargetOpcode::G_JUMP_TABLE: 1343 return AllowOpaqueConstants; 1344 default: 1345 return false; 1346 } 1347 } 1348 1349 bool llvm::isConstantOrConstantVector(MachineInstr &MI, 1350 const MachineRegisterInfo &MRI) { 1351 Register Def = MI.getOperand(0).getReg(); 1352 if (auto C = getIConstantVRegValWithLookThrough(Def, MRI)) 1353 return true; 1354 GBuildVector *BV = dyn_cast<GBuildVector>(&MI); 1355 if (!BV) 1356 return false; 1357 for (unsigned SrcIdx = 0; SrcIdx < BV->getNumSources(); ++SrcIdx) { 1358 if (getIConstantVRegValWithLookThrough(BV->getSourceReg(SrcIdx), MRI) || 1359 getOpcodeDef<GImplicitDef>(BV->getSourceReg(SrcIdx), MRI)) 1360 continue; 1361 return false; 1362 } 1363 return true; 1364 } 1365 1366 bool llvm::isConstantOrConstantVector(const MachineInstr &MI, 1367 const MachineRegisterInfo &MRI, 1368 bool AllowFP, bool AllowOpaqueConstants) { 1369 if (isConstantScalar(MI, MRI, AllowFP, AllowOpaqueConstants)) 1370 return true; 1371 1372 if (!isBuildVectorOp(MI.getOpcode())) 1373 return false; 1374 1375 const unsigned NumOps = MI.getNumOperands(); 1376 for (unsigned I = 1; I != NumOps; ++I) { 1377 const MachineInstr *ElementDef = MRI.getVRegDef(MI.getOperand(I).getReg()); 1378 if (!isConstantScalar(*ElementDef, MRI, AllowFP, AllowOpaqueConstants)) 1379 return false; 1380 } 1381 1382 return true; 1383 } 1384 1385 std::optional<APInt> 1386 llvm::isConstantOrConstantSplatVector(MachineInstr &MI, 1387 const MachineRegisterInfo &MRI) { 1388 Register Def = MI.getOperand(0).getReg(); 1389 if (auto C = getIConstantVRegValWithLookThrough(Def, MRI)) 1390 return C->Value; 1391 auto MaybeCst = getIConstantSplatSExtVal(MI, MRI); 1392 if (!MaybeCst) 1393 return std::nullopt; 1394 const unsigned ScalarSize = MRI.getType(Def).getScalarSizeInBits(); 1395 return APInt(ScalarSize, *MaybeCst, true); 1396 } 1397 1398 bool llvm::isNullOrNullSplat(const MachineInstr &MI, 1399 const MachineRegisterInfo &MRI, bool AllowUndefs) { 1400 switch (MI.getOpcode()) { 1401 case TargetOpcode::G_IMPLICIT_DEF: 1402 return AllowUndefs; 1403 case TargetOpcode::G_CONSTANT: 1404 return MI.getOperand(1).getCImm()->isNullValue(); 1405 case TargetOpcode::G_FCONSTANT: { 1406 const ConstantFP *FPImm = MI.getOperand(1).getFPImm(); 1407 return FPImm->isZero() && !FPImm->isNegative(); 1408 } 1409 default: 1410 if (!AllowUndefs) // TODO: isBuildVectorAllZeros assumes undef is OK already 1411 return false; 1412 return isBuildVectorAllZeros(MI, MRI); 1413 } 1414 } 1415 1416 bool llvm::isAllOnesOrAllOnesSplat(const MachineInstr &MI, 1417 const MachineRegisterInfo &MRI, 1418 bool AllowUndefs) { 1419 switch (MI.getOpcode()) { 1420 case TargetOpcode::G_IMPLICIT_DEF: 1421 return AllowUndefs; 1422 case TargetOpcode::G_CONSTANT: 1423 return MI.getOperand(1).getCImm()->isAllOnesValue(); 1424 default: 1425 if (!AllowUndefs) // TODO: isBuildVectorAllOnes assumes undef is OK already 1426 return false; 1427 return isBuildVectorAllOnes(MI, MRI); 1428 } 1429 } 1430 1431 bool llvm::matchUnaryPredicate( 1432 const MachineRegisterInfo &MRI, Register Reg, 1433 std::function<bool(const Constant *ConstVal)> Match, bool AllowUndefs) { 1434 1435 const MachineInstr *Def = getDefIgnoringCopies(Reg, MRI); 1436 if (AllowUndefs && Def->getOpcode() == TargetOpcode::G_IMPLICIT_DEF) 1437 return Match(nullptr); 1438 1439 // TODO: Also handle fconstant 1440 if (Def->getOpcode() == TargetOpcode::G_CONSTANT) 1441 return Match(Def->getOperand(1).getCImm()); 1442 1443 if (Def->getOpcode() != TargetOpcode::G_BUILD_VECTOR) 1444 return false; 1445 1446 for (unsigned I = 1, E = Def->getNumOperands(); I != E; ++I) { 1447 Register SrcElt = Def->getOperand(I).getReg(); 1448 const MachineInstr *SrcDef = getDefIgnoringCopies(SrcElt, MRI); 1449 if (AllowUndefs && SrcDef->getOpcode() == TargetOpcode::G_IMPLICIT_DEF) { 1450 if (!Match(nullptr)) 1451 return false; 1452 continue; 1453 } 1454 1455 if (SrcDef->getOpcode() != TargetOpcode::G_CONSTANT || 1456 !Match(SrcDef->getOperand(1).getCImm())) 1457 return false; 1458 } 1459 1460 return true; 1461 } 1462 1463 bool llvm::isConstTrueVal(const TargetLowering &TLI, int64_t Val, bool IsVector, 1464 bool IsFP) { 1465 switch (TLI.getBooleanContents(IsVector, IsFP)) { 1466 case TargetLowering::UndefinedBooleanContent: 1467 return Val & 0x1; 1468 case TargetLowering::ZeroOrOneBooleanContent: 1469 return Val == 1; 1470 case TargetLowering::ZeroOrNegativeOneBooleanContent: 1471 return Val == -1; 1472 } 1473 llvm_unreachable("Invalid boolean contents"); 1474 } 1475 1476 bool llvm::isConstFalseVal(const TargetLowering &TLI, int64_t Val, 1477 bool IsVector, bool IsFP) { 1478 switch (TLI.getBooleanContents(IsVector, IsFP)) { 1479 case TargetLowering::UndefinedBooleanContent: 1480 return ~Val & 0x1; 1481 case TargetLowering::ZeroOrOneBooleanContent: 1482 case TargetLowering::ZeroOrNegativeOneBooleanContent: 1483 return Val == 0; 1484 } 1485 llvm_unreachable("Invalid boolean contents"); 1486 } 1487 1488 int64_t llvm::getICmpTrueVal(const TargetLowering &TLI, bool IsVector, 1489 bool IsFP) { 1490 switch (TLI.getBooleanContents(IsVector, IsFP)) { 1491 case TargetLowering::UndefinedBooleanContent: 1492 case TargetLowering::ZeroOrOneBooleanContent: 1493 return 1; 1494 case TargetLowering::ZeroOrNegativeOneBooleanContent: 1495 return -1; 1496 } 1497 llvm_unreachable("Invalid boolean contents"); 1498 } 1499 1500 bool llvm::shouldOptForSize(const MachineBasicBlock &MBB, 1501 ProfileSummaryInfo *PSI, BlockFrequencyInfo *BFI) { 1502 const auto &F = MBB.getParent()->getFunction(); 1503 return F.hasOptSize() || F.hasMinSize() || 1504 llvm::shouldOptimizeForSize(MBB.getBasicBlock(), PSI, BFI); 1505 } 1506 1507 void llvm::saveUsesAndErase(MachineInstr &MI, MachineRegisterInfo &MRI, 1508 LostDebugLocObserver *LocObserver, 1509 SmallInstListTy &DeadInstChain) { 1510 for (MachineOperand &Op : MI.uses()) { 1511 if (Op.isReg() && Op.getReg().isVirtual()) 1512 DeadInstChain.insert(MRI.getVRegDef(Op.getReg())); 1513 } 1514 LLVM_DEBUG(dbgs() << MI << "Is dead; erasing.\n"); 1515 DeadInstChain.remove(&MI); 1516 MI.eraseFromParent(); 1517 if (LocObserver) 1518 LocObserver->checkpoint(false); 1519 } 1520 1521 void llvm::eraseInstrs(ArrayRef<MachineInstr *> DeadInstrs, 1522 MachineRegisterInfo &MRI, 1523 LostDebugLocObserver *LocObserver) { 1524 SmallInstListTy DeadInstChain; 1525 for (MachineInstr *MI : DeadInstrs) 1526 saveUsesAndErase(*MI, MRI, LocObserver, DeadInstChain); 1527 1528 while (!DeadInstChain.empty()) { 1529 MachineInstr *Inst = DeadInstChain.pop_back_val(); 1530 if (!isTriviallyDead(*Inst, MRI)) 1531 continue; 1532 saveUsesAndErase(*Inst, MRI, LocObserver, DeadInstChain); 1533 } 1534 } 1535 1536 void llvm::eraseInstr(MachineInstr &MI, MachineRegisterInfo &MRI, 1537 LostDebugLocObserver *LocObserver) { 1538 return eraseInstrs({&MI}, MRI, LocObserver); 1539 } 1540 1541 void llvm::salvageDebugInfo(const MachineRegisterInfo &MRI, MachineInstr &MI) { 1542 for (auto &Def : MI.defs()) { 1543 assert(Def.isReg() && "Must be a reg"); 1544 1545 SmallVector<MachineOperand *, 16> DbgUsers; 1546 for (auto &MOUse : MRI.use_operands(Def.getReg())) { 1547 MachineInstr *DbgValue = MOUse.getParent(); 1548 // Ignore partially formed DBG_VALUEs. 1549 if (DbgValue->isNonListDebugValue() && DbgValue->getNumOperands() == 4) { 1550 DbgUsers.push_back(&MOUse); 1551 } 1552 } 1553 1554 if (!DbgUsers.empty()) { 1555 salvageDebugInfoForDbgValue(MRI, MI, DbgUsers); 1556 } 1557 } 1558 } 1559