1 //===-- llvm/CodeGen/GlobalISel/LegalizerHelper.cpp -----------------------===// 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 // 9 /// \file This file implements the LegalizerHelper class to legalize 10 /// individual instructions and the LegalizeMachineIR wrapper pass for the 11 /// primary legalization. 12 // 13 //===----------------------------------------------------------------------===// 14 15 #include "llvm/CodeGen/GlobalISel/LegalizerHelper.h" 16 #include "llvm/CodeGen/GlobalISel/CallLowering.h" 17 #include "llvm/CodeGen/GlobalISel/GISelChangeObserver.h" 18 #include "llvm/CodeGen/GlobalISel/LegalizerInfo.h" 19 #include "llvm/CodeGen/MachineRegisterInfo.h" 20 #include "llvm/CodeGen/TargetFrameLowering.h" 21 #include "llvm/CodeGen/TargetInstrInfo.h" 22 #include "llvm/CodeGen/TargetLowering.h" 23 #include "llvm/CodeGen/TargetSubtargetInfo.h" 24 #include "llvm/Support/Debug.h" 25 #include "llvm/Support/MathExtras.h" 26 #include "llvm/Support/raw_ostream.h" 27 28 #define DEBUG_TYPE "legalizer" 29 30 using namespace llvm; 31 using namespace LegalizeActions; 32 33 /// Try to break down \p OrigTy into \p NarrowTy sized pieces. 34 /// 35 /// Returns the number of \p NarrowTy elements needed to reconstruct \p OrigTy, 36 /// with any leftover piece as type \p LeftoverTy 37 /// 38 /// Returns -1 in the first element of the pair if the breakdown is not 39 /// satisfiable. 40 static std::pair<int, int> 41 getNarrowTypeBreakDown(LLT OrigTy, LLT NarrowTy, LLT &LeftoverTy) { 42 assert(!LeftoverTy.isValid() && "this is an out argument"); 43 44 unsigned Size = OrigTy.getSizeInBits(); 45 unsigned NarrowSize = NarrowTy.getSizeInBits(); 46 unsigned NumParts = Size / NarrowSize; 47 unsigned LeftoverSize = Size - NumParts * NarrowSize; 48 assert(Size > NarrowSize); 49 50 if (LeftoverSize == 0) 51 return {NumParts, 0}; 52 53 if (NarrowTy.isVector()) { 54 unsigned EltSize = OrigTy.getScalarSizeInBits(); 55 if (LeftoverSize % EltSize != 0) 56 return {-1, -1}; 57 LeftoverTy = LLT::scalarOrVector(LeftoverSize / EltSize, EltSize); 58 } else { 59 LeftoverTy = LLT::scalar(LeftoverSize); 60 } 61 62 int NumLeftover = LeftoverSize / LeftoverTy.getSizeInBits(); 63 return std::make_pair(NumParts, NumLeftover); 64 } 65 66 LegalizerHelper::LegalizerHelper(MachineFunction &MF, 67 GISelChangeObserver &Observer, 68 MachineIRBuilder &Builder) 69 : MIRBuilder(Builder), MRI(MF.getRegInfo()), 70 LI(*MF.getSubtarget().getLegalizerInfo()), Observer(Observer) { 71 MIRBuilder.setMF(MF); 72 MIRBuilder.setChangeObserver(Observer); 73 } 74 75 LegalizerHelper::LegalizerHelper(MachineFunction &MF, const LegalizerInfo &LI, 76 GISelChangeObserver &Observer, 77 MachineIRBuilder &B) 78 : MIRBuilder(B), MRI(MF.getRegInfo()), LI(LI), Observer(Observer) { 79 MIRBuilder.setMF(MF); 80 MIRBuilder.setChangeObserver(Observer); 81 } 82 LegalizerHelper::LegalizeResult 83 LegalizerHelper::legalizeInstrStep(MachineInstr &MI) { 84 LLVM_DEBUG(dbgs() << "Legalizing: "; MI.print(dbgs())); 85 86 if (MI.getOpcode() == TargetOpcode::G_INTRINSIC || 87 MI.getOpcode() == TargetOpcode::G_INTRINSIC_W_SIDE_EFFECTS) 88 return LI.legalizeIntrinsic(MI, MRI, MIRBuilder) ? Legalized 89 : UnableToLegalize; 90 auto Step = LI.getAction(MI, MRI); 91 switch (Step.Action) { 92 case Legal: 93 LLVM_DEBUG(dbgs() << ".. Already legal\n"); 94 return AlreadyLegal; 95 case Libcall: 96 LLVM_DEBUG(dbgs() << ".. Convert to libcall\n"); 97 return libcall(MI); 98 case NarrowScalar: 99 LLVM_DEBUG(dbgs() << ".. Narrow scalar\n"); 100 return narrowScalar(MI, Step.TypeIdx, Step.NewType); 101 case WidenScalar: 102 LLVM_DEBUG(dbgs() << ".. Widen scalar\n"); 103 return widenScalar(MI, Step.TypeIdx, Step.NewType); 104 case Lower: 105 LLVM_DEBUG(dbgs() << ".. Lower\n"); 106 return lower(MI, Step.TypeIdx, Step.NewType); 107 case FewerElements: 108 LLVM_DEBUG(dbgs() << ".. Reduce number of elements\n"); 109 return fewerElementsVector(MI, Step.TypeIdx, Step.NewType); 110 case MoreElements: 111 LLVM_DEBUG(dbgs() << ".. Increase number of elements\n"); 112 return moreElementsVector(MI, Step.TypeIdx, Step.NewType); 113 case Custom: 114 LLVM_DEBUG(dbgs() << ".. Custom legalization\n"); 115 return LI.legalizeCustom(MI, MRI, MIRBuilder, Observer) ? Legalized 116 : UnableToLegalize; 117 default: 118 LLVM_DEBUG(dbgs() << ".. Unable to legalize\n"); 119 return UnableToLegalize; 120 } 121 } 122 123 void LegalizerHelper::extractParts(Register Reg, LLT Ty, int NumParts, 124 SmallVectorImpl<Register> &VRegs) { 125 for (int i = 0; i < NumParts; ++i) 126 VRegs.push_back(MRI.createGenericVirtualRegister(Ty)); 127 MIRBuilder.buildUnmerge(VRegs, Reg); 128 } 129 130 bool LegalizerHelper::extractParts(Register Reg, LLT RegTy, 131 LLT MainTy, LLT &LeftoverTy, 132 SmallVectorImpl<Register> &VRegs, 133 SmallVectorImpl<Register> &LeftoverRegs) { 134 assert(!LeftoverTy.isValid() && "this is an out argument"); 135 136 unsigned RegSize = RegTy.getSizeInBits(); 137 unsigned MainSize = MainTy.getSizeInBits(); 138 unsigned NumParts = RegSize / MainSize; 139 unsigned LeftoverSize = RegSize - NumParts * MainSize; 140 141 // Use an unmerge when possible. 142 if (LeftoverSize == 0) { 143 for (unsigned I = 0; I < NumParts; ++I) 144 VRegs.push_back(MRI.createGenericVirtualRegister(MainTy)); 145 MIRBuilder.buildUnmerge(VRegs, Reg); 146 return true; 147 } 148 149 if (MainTy.isVector()) { 150 unsigned EltSize = MainTy.getScalarSizeInBits(); 151 if (LeftoverSize % EltSize != 0) 152 return false; 153 LeftoverTy = LLT::scalarOrVector(LeftoverSize / EltSize, EltSize); 154 } else { 155 LeftoverTy = LLT::scalar(LeftoverSize); 156 } 157 158 // For irregular sizes, extract the individual parts. 159 for (unsigned I = 0; I != NumParts; ++I) { 160 Register NewReg = MRI.createGenericVirtualRegister(MainTy); 161 VRegs.push_back(NewReg); 162 MIRBuilder.buildExtract(NewReg, Reg, MainSize * I); 163 } 164 165 for (unsigned Offset = MainSize * NumParts; Offset < RegSize; 166 Offset += LeftoverSize) { 167 Register NewReg = MRI.createGenericVirtualRegister(LeftoverTy); 168 LeftoverRegs.push_back(NewReg); 169 MIRBuilder.buildExtract(NewReg, Reg, Offset); 170 } 171 172 return true; 173 } 174 175 static LLT getGCDType(LLT OrigTy, LLT TargetTy) { 176 if (OrigTy.isVector() && TargetTy.isVector()) { 177 assert(OrigTy.getElementType() == TargetTy.getElementType()); 178 int GCD = greatestCommonDivisor(OrigTy.getNumElements(), 179 TargetTy.getNumElements()); 180 return LLT::scalarOrVector(GCD, OrigTy.getElementType()); 181 } 182 183 if (OrigTy.isVector() && !TargetTy.isVector()) { 184 assert(OrigTy.getElementType() == TargetTy); 185 return TargetTy; 186 } 187 188 assert(!OrigTy.isVector() && !TargetTy.isVector()); 189 190 int GCD = greatestCommonDivisor(OrigTy.getSizeInBits(), 191 TargetTy.getSizeInBits()); 192 return LLT::scalar(GCD); 193 } 194 195 void LegalizerHelper::insertParts(Register DstReg, 196 LLT ResultTy, LLT PartTy, 197 ArrayRef<Register> PartRegs, 198 LLT LeftoverTy, 199 ArrayRef<Register> LeftoverRegs) { 200 if (!LeftoverTy.isValid()) { 201 assert(LeftoverRegs.empty()); 202 203 if (!ResultTy.isVector()) { 204 MIRBuilder.buildMerge(DstReg, PartRegs); 205 return; 206 } 207 208 if (PartTy.isVector()) 209 MIRBuilder.buildConcatVectors(DstReg, PartRegs); 210 else 211 MIRBuilder.buildBuildVector(DstReg, PartRegs); 212 return; 213 } 214 215 unsigned PartSize = PartTy.getSizeInBits(); 216 unsigned LeftoverPartSize = LeftoverTy.getSizeInBits(); 217 218 Register CurResultReg = MRI.createGenericVirtualRegister(ResultTy); 219 MIRBuilder.buildUndef(CurResultReg); 220 221 unsigned Offset = 0; 222 for (Register PartReg : PartRegs) { 223 Register NewResultReg = MRI.createGenericVirtualRegister(ResultTy); 224 MIRBuilder.buildInsert(NewResultReg, CurResultReg, PartReg, Offset); 225 CurResultReg = NewResultReg; 226 Offset += PartSize; 227 } 228 229 for (unsigned I = 0, E = LeftoverRegs.size(); I != E; ++I) { 230 // Use the original output register for the final insert to avoid a copy. 231 Register NewResultReg = (I + 1 == E) ? 232 DstReg : MRI.createGenericVirtualRegister(ResultTy); 233 234 MIRBuilder.buildInsert(NewResultReg, CurResultReg, LeftoverRegs[I], Offset); 235 CurResultReg = NewResultReg; 236 Offset += LeftoverPartSize; 237 } 238 } 239 240 static RTLIB::Libcall getRTLibDesc(unsigned Opcode, unsigned Size) { 241 switch (Opcode) { 242 case TargetOpcode::G_SDIV: 243 assert((Size == 32 || Size == 64 || Size == 128) && "Unsupported size"); 244 switch (Size) { 245 case 32: 246 return RTLIB::SDIV_I32; 247 case 64: 248 return RTLIB::SDIV_I64; 249 case 128: 250 return RTLIB::SDIV_I128; 251 default: 252 llvm_unreachable("unexpected size"); 253 } 254 case TargetOpcode::G_UDIV: 255 assert((Size == 32 || Size == 64 || Size == 128) && "Unsupported size"); 256 switch (Size) { 257 case 32: 258 return RTLIB::UDIV_I32; 259 case 64: 260 return RTLIB::UDIV_I64; 261 case 128: 262 return RTLIB::UDIV_I128; 263 default: 264 llvm_unreachable("unexpected size"); 265 } 266 case TargetOpcode::G_SREM: 267 assert((Size == 32 || Size == 64) && "Unsupported size"); 268 return Size == 64 ? RTLIB::SREM_I64 : RTLIB::SREM_I32; 269 case TargetOpcode::G_UREM: 270 assert((Size == 32 || Size == 64) && "Unsupported size"); 271 return Size == 64 ? RTLIB::UREM_I64 : RTLIB::UREM_I32; 272 case TargetOpcode::G_CTLZ_ZERO_UNDEF: 273 assert(Size == 32 && "Unsupported size"); 274 return RTLIB::CTLZ_I32; 275 case TargetOpcode::G_FADD: 276 assert((Size == 32 || Size == 64) && "Unsupported size"); 277 return Size == 64 ? RTLIB::ADD_F64 : RTLIB::ADD_F32; 278 case TargetOpcode::G_FSUB: 279 assert((Size == 32 || Size == 64) && "Unsupported size"); 280 return Size == 64 ? RTLIB::SUB_F64 : RTLIB::SUB_F32; 281 case TargetOpcode::G_FMUL: 282 assert((Size == 32 || Size == 64) && "Unsupported size"); 283 return Size == 64 ? RTLIB::MUL_F64 : RTLIB::MUL_F32; 284 case TargetOpcode::G_FDIV: 285 assert((Size == 32 || Size == 64) && "Unsupported size"); 286 return Size == 64 ? RTLIB::DIV_F64 : RTLIB::DIV_F32; 287 case TargetOpcode::G_FEXP: 288 assert((Size == 32 || Size == 64) && "Unsupported size"); 289 return Size == 64 ? RTLIB::EXP_F64 : RTLIB::EXP_F32; 290 case TargetOpcode::G_FEXP2: 291 assert((Size == 32 || Size == 64) && "Unsupported size"); 292 return Size == 64 ? RTLIB::EXP2_F64 : RTLIB::EXP2_F32; 293 case TargetOpcode::G_FREM: 294 return Size == 64 ? RTLIB::REM_F64 : RTLIB::REM_F32; 295 case TargetOpcode::G_FPOW: 296 return Size == 64 ? RTLIB::POW_F64 : RTLIB::POW_F32; 297 case TargetOpcode::G_FMA: 298 assert((Size == 32 || Size == 64) && "Unsupported size"); 299 return Size == 64 ? RTLIB::FMA_F64 : RTLIB::FMA_F32; 300 case TargetOpcode::G_FSIN: 301 assert((Size == 32 || Size == 64 || Size == 128) && "Unsupported size"); 302 return Size == 128 ? RTLIB::SIN_F128 303 : Size == 64 ? RTLIB::SIN_F64 : RTLIB::SIN_F32; 304 case TargetOpcode::G_FCOS: 305 assert((Size == 32 || Size == 64 || Size == 128) && "Unsupported size"); 306 return Size == 128 ? RTLIB::COS_F128 307 : Size == 64 ? RTLIB::COS_F64 : RTLIB::COS_F32; 308 case TargetOpcode::G_FLOG10: 309 assert((Size == 32 || Size == 64 || Size == 128) && "Unsupported size"); 310 return Size == 128 ? RTLIB::LOG10_F128 311 : Size == 64 ? RTLIB::LOG10_F64 : RTLIB::LOG10_F32; 312 case TargetOpcode::G_FLOG: 313 assert((Size == 32 || Size == 64 || Size == 128) && "Unsupported size"); 314 return Size == 128 ? RTLIB::LOG_F128 315 : Size == 64 ? RTLIB::LOG_F64 : RTLIB::LOG_F32; 316 case TargetOpcode::G_FLOG2: 317 assert((Size == 32 || Size == 64 || Size == 128) && "Unsupported size"); 318 return Size == 128 ? RTLIB::LOG2_F128 319 : Size == 64 ? RTLIB::LOG2_F64 : RTLIB::LOG2_F32; 320 case TargetOpcode::G_FCEIL: 321 assert((Size == 32 || Size == 64) && "Unsupported size"); 322 return Size == 64 ? RTLIB::CEIL_F64 : RTLIB::CEIL_F32; 323 case TargetOpcode::G_FFLOOR: 324 assert((Size == 32 || Size == 64) && "Unsupported size"); 325 return Size == 64 ? RTLIB::FLOOR_F64 : RTLIB::FLOOR_F32; 326 } 327 llvm_unreachable("Unknown libcall function"); 328 } 329 330 /// True if an instruction is in tail position in its caller. Intended for 331 /// legalizing libcalls as tail calls when possible. 332 static bool isLibCallInTailPosition(MachineInstr &MI) { 333 const Function &F = MI.getParent()->getParent()->getFunction(); 334 335 // Conservatively require the attributes of the call to match those of 336 // the return. Ignore NoAlias and NonNull because they don't affect the 337 // call sequence. 338 AttributeList CallerAttrs = F.getAttributes(); 339 if (AttrBuilder(CallerAttrs, AttributeList::ReturnIndex) 340 .removeAttribute(Attribute::NoAlias) 341 .removeAttribute(Attribute::NonNull) 342 .hasAttributes()) 343 return false; 344 345 // It's not safe to eliminate the sign / zero extension of the return value. 346 if (CallerAttrs.hasAttribute(AttributeList::ReturnIndex, Attribute::ZExt) || 347 CallerAttrs.hasAttribute(AttributeList::ReturnIndex, Attribute::SExt)) 348 return false; 349 350 // Only tail call if the following instruction is a standard return. 351 auto &TII = *MI.getMF()->getSubtarget().getInstrInfo(); 352 MachineInstr *Next = MI.getNextNode(); 353 if (!Next || TII.isTailCall(*Next) || !Next->isReturn()) 354 return false; 355 356 return true; 357 } 358 359 LegalizerHelper::LegalizeResult 360 llvm::createLibcall(MachineIRBuilder &MIRBuilder, RTLIB::Libcall Libcall, 361 const CallLowering::ArgInfo &Result, 362 ArrayRef<CallLowering::ArgInfo> Args) { 363 auto &CLI = *MIRBuilder.getMF().getSubtarget().getCallLowering(); 364 auto &TLI = *MIRBuilder.getMF().getSubtarget().getTargetLowering(); 365 const char *Name = TLI.getLibcallName(Libcall); 366 367 CallLowering::CallLoweringInfo Info; 368 Info.CallConv = TLI.getLibcallCallingConv(Libcall); 369 Info.Callee = MachineOperand::CreateES(Name); 370 Info.OrigRet = Result; 371 std::copy(Args.begin(), Args.end(), std::back_inserter(Info.OrigArgs)); 372 if (!CLI.lowerCall(MIRBuilder, Info)) 373 return LegalizerHelper::UnableToLegalize; 374 375 return LegalizerHelper::Legalized; 376 } 377 378 // Useful for libcalls where all operands have the same type. 379 static LegalizerHelper::LegalizeResult 380 simpleLibcall(MachineInstr &MI, MachineIRBuilder &MIRBuilder, unsigned Size, 381 Type *OpType) { 382 auto Libcall = getRTLibDesc(MI.getOpcode(), Size); 383 384 SmallVector<CallLowering::ArgInfo, 3> Args; 385 for (unsigned i = 1; i < MI.getNumOperands(); i++) 386 Args.push_back({MI.getOperand(i).getReg(), OpType}); 387 return createLibcall(MIRBuilder, Libcall, {MI.getOperand(0).getReg(), OpType}, 388 Args); 389 } 390 391 LegalizerHelper::LegalizeResult 392 llvm::createMemLibcall(MachineIRBuilder &MIRBuilder, MachineRegisterInfo &MRI, 393 MachineInstr &MI) { 394 assert(MI.getOpcode() == TargetOpcode::G_INTRINSIC_W_SIDE_EFFECTS); 395 auto &Ctx = MIRBuilder.getMF().getFunction().getContext(); 396 397 SmallVector<CallLowering::ArgInfo, 3> Args; 398 // Add all the args, except for the last which is an imm denoting 'tail'. 399 for (unsigned i = 1; i < MI.getNumOperands() - 1; i++) { 400 Register Reg = MI.getOperand(i).getReg(); 401 402 // Need derive an IR type for call lowering. 403 LLT OpLLT = MRI.getType(Reg); 404 Type *OpTy = nullptr; 405 if (OpLLT.isPointer()) 406 OpTy = Type::getInt8PtrTy(Ctx, OpLLT.getAddressSpace()); 407 else 408 OpTy = IntegerType::get(Ctx, OpLLT.getSizeInBits()); 409 Args.push_back({Reg, OpTy}); 410 } 411 412 auto &CLI = *MIRBuilder.getMF().getSubtarget().getCallLowering(); 413 auto &TLI = *MIRBuilder.getMF().getSubtarget().getTargetLowering(); 414 Intrinsic::ID ID = MI.getOperand(0).getIntrinsicID(); 415 RTLIB::Libcall RTLibcall; 416 switch (ID) { 417 case Intrinsic::memcpy: 418 RTLibcall = RTLIB::MEMCPY; 419 break; 420 case Intrinsic::memset: 421 RTLibcall = RTLIB::MEMSET; 422 break; 423 case Intrinsic::memmove: 424 RTLibcall = RTLIB::MEMMOVE; 425 break; 426 default: 427 return LegalizerHelper::UnableToLegalize; 428 } 429 const char *Name = TLI.getLibcallName(RTLibcall); 430 431 MIRBuilder.setInstr(MI); 432 433 CallLowering::CallLoweringInfo Info; 434 Info.CallConv = TLI.getLibcallCallingConv(RTLibcall); 435 Info.Callee = MachineOperand::CreateES(Name); 436 Info.OrigRet = CallLowering::ArgInfo({0}, Type::getVoidTy(Ctx)); 437 Info.IsTailCall = MI.getOperand(MI.getNumOperands() - 1).getImm() == 1 && 438 isLibCallInTailPosition(MI); 439 440 std::copy(Args.begin(), Args.end(), std::back_inserter(Info.OrigArgs)); 441 if (!CLI.lowerCall(MIRBuilder, Info)) 442 return LegalizerHelper::UnableToLegalize; 443 444 if (Info.LoweredTailCall) { 445 assert(Info.IsTailCall && "Lowered tail call when it wasn't a tail call?"); 446 // We must have a return following the call to get past 447 // isLibCallInTailPosition. 448 assert(MI.getNextNode() && MI.getNextNode()->isReturn() && 449 "Expected instr following MI to be a return?"); 450 451 // We lowered a tail call, so the call is now the return from the block. 452 // Delete the old return. 453 MI.getNextNode()->eraseFromParent(); 454 } 455 456 return LegalizerHelper::Legalized; 457 } 458 459 static RTLIB::Libcall getConvRTLibDesc(unsigned Opcode, Type *ToType, 460 Type *FromType) { 461 auto ToMVT = MVT::getVT(ToType); 462 auto FromMVT = MVT::getVT(FromType); 463 464 switch (Opcode) { 465 case TargetOpcode::G_FPEXT: 466 return RTLIB::getFPEXT(FromMVT, ToMVT); 467 case TargetOpcode::G_FPTRUNC: 468 return RTLIB::getFPROUND(FromMVT, ToMVT); 469 case TargetOpcode::G_FPTOSI: 470 return RTLIB::getFPTOSINT(FromMVT, ToMVT); 471 case TargetOpcode::G_FPTOUI: 472 return RTLIB::getFPTOUINT(FromMVT, ToMVT); 473 case TargetOpcode::G_SITOFP: 474 return RTLIB::getSINTTOFP(FromMVT, ToMVT); 475 case TargetOpcode::G_UITOFP: 476 return RTLIB::getUINTTOFP(FromMVT, ToMVT); 477 } 478 llvm_unreachable("Unsupported libcall function"); 479 } 480 481 static LegalizerHelper::LegalizeResult 482 conversionLibcall(MachineInstr &MI, MachineIRBuilder &MIRBuilder, Type *ToType, 483 Type *FromType) { 484 RTLIB::Libcall Libcall = getConvRTLibDesc(MI.getOpcode(), ToType, FromType); 485 return createLibcall(MIRBuilder, Libcall, {MI.getOperand(0).getReg(), ToType}, 486 {{MI.getOperand(1).getReg(), FromType}}); 487 } 488 489 LegalizerHelper::LegalizeResult 490 LegalizerHelper::libcall(MachineInstr &MI) { 491 LLT LLTy = MRI.getType(MI.getOperand(0).getReg()); 492 unsigned Size = LLTy.getSizeInBits(); 493 auto &Ctx = MIRBuilder.getMF().getFunction().getContext(); 494 495 MIRBuilder.setInstr(MI); 496 497 switch (MI.getOpcode()) { 498 default: 499 return UnableToLegalize; 500 case TargetOpcode::G_SDIV: 501 case TargetOpcode::G_UDIV: 502 case TargetOpcode::G_SREM: 503 case TargetOpcode::G_UREM: 504 case TargetOpcode::G_CTLZ_ZERO_UNDEF: { 505 Type *HLTy = IntegerType::get(Ctx, Size); 506 auto Status = simpleLibcall(MI, MIRBuilder, Size, HLTy); 507 if (Status != Legalized) 508 return Status; 509 break; 510 } 511 case TargetOpcode::G_FADD: 512 case TargetOpcode::G_FSUB: 513 case TargetOpcode::G_FMUL: 514 case TargetOpcode::G_FDIV: 515 case TargetOpcode::G_FMA: 516 case TargetOpcode::G_FPOW: 517 case TargetOpcode::G_FREM: 518 case TargetOpcode::G_FCOS: 519 case TargetOpcode::G_FSIN: 520 case TargetOpcode::G_FLOG10: 521 case TargetOpcode::G_FLOG: 522 case TargetOpcode::G_FLOG2: 523 case TargetOpcode::G_FEXP: 524 case TargetOpcode::G_FEXP2: 525 case TargetOpcode::G_FCEIL: 526 case TargetOpcode::G_FFLOOR: { 527 if (Size > 64) { 528 LLVM_DEBUG(dbgs() << "Size " << Size << " too large to legalize.\n"); 529 return UnableToLegalize; 530 } 531 Type *HLTy = Size == 64 ? Type::getDoubleTy(Ctx) : Type::getFloatTy(Ctx); 532 auto Status = simpleLibcall(MI, MIRBuilder, Size, HLTy); 533 if (Status != Legalized) 534 return Status; 535 break; 536 } 537 case TargetOpcode::G_FPEXT: { 538 // FIXME: Support other floating point types (half, fp128 etc) 539 unsigned FromSize = MRI.getType(MI.getOperand(1).getReg()).getSizeInBits(); 540 unsigned ToSize = MRI.getType(MI.getOperand(0).getReg()).getSizeInBits(); 541 if (ToSize != 64 || FromSize != 32) 542 return UnableToLegalize; 543 LegalizeResult Status = conversionLibcall( 544 MI, MIRBuilder, Type::getDoubleTy(Ctx), Type::getFloatTy(Ctx)); 545 if (Status != Legalized) 546 return Status; 547 break; 548 } 549 case TargetOpcode::G_FPTRUNC: { 550 // FIXME: Support other floating point types (half, fp128 etc) 551 unsigned FromSize = MRI.getType(MI.getOperand(1).getReg()).getSizeInBits(); 552 unsigned ToSize = MRI.getType(MI.getOperand(0).getReg()).getSizeInBits(); 553 if (ToSize != 32 || FromSize != 64) 554 return UnableToLegalize; 555 LegalizeResult Status = conversionLibcall( 556 MI, MIRBuilder, Type::getFloatTy(Ctx), Type::getDoubleTy(Ctx)); 557 if (Status != Legalized) 558 return Status; 559 break; 560 } 561 case TargetOpcode::G_FPTOSI: 562 case TargetOpcode::G_FPTOUI: { 563 // FIXME: Support other types 564 unsigned FromSize = MRI.getType(MI.getOperand(1).getReg()).getSizeInBits(); 565 unsigned ToSize = MRI.getType(MI.getOperand(0).getReg()).getSizeInBits(); 566 if ((ToSize != 32 && ToSize != 64) || (FromSize != 32 && FromSize != 64)) 567 return UnableToLegalize; 568 LegalizeResult Status = conversionLibcall( 569 MI, MIRBuilder, 570 ToSize == 32 ? Type::getInt32Ty(Ctx) : Type::getInt64Ty(Ctx), 571 FromSize == 64 ? Type::getDoubleTy(Ctx) : Type::getFloatTy(Ctx)); 572 if (Status != Legalized) 573 return Status; 574 break; 575 } 576 case TargetOpcode::G_SITOFP: 577 case TargetOpcode::G_UITOFP: { 578 // FIXME: Support other types 579 unsigned FromSize = MRI.getType(MI.getOperand(1).getReg()).getSizeInBits(); 580 unsigned ToSize = MRI.getType(MI.getOperand(0).getReg()).getSizeInBits(); 581 if ((FromSize != 32 && FromSize != 64) || (ToSize != 32 && ToSize != 64)) 582 return UnableToLegalize; 583 LegalizeResult Status = conversionLibcall( 584 MI, MIRBuilder, 585 ToSize == 64 ? Type::getDoubleTy(Ctx) : Type::getFloatTy(Ctx), 586 FromSize == 32 ? Type::getInt32Ty(Ctx) : Type::getInt64Ty(Ctx)); 587 if (Status != Legalized) 588 return Status; 589 break; 590 } 591 } 592 593 MI.eraseFromParent(); 594 return Legalized; 595 } 596 597 LegalizerHelper::LegalizeResult LegalizerHelper::narrowScalar(MachineInstr &MI, 598 unsigned TypeIdx, 599 LLT NarrowTy) { 600 MIRBuilder.setInstr(MI); 601 602 uint64_t SizeOp0 = MRI.getType(MI.getOperand(0).getReg()).getSizeInBits(); 603 uint64_t NarrowSize = NarrowTy.getSizeInBits(); 604 605 switch (MI.getOpcode()) { 606 default: 607 return UnableToLegalize; 608 case TargetOpcode::G_IMPLICIT_DEF: { 609 // FIXME: add support for when SizeOp0 isn't an exact multiple of 610 // NarrowSize. 611 if (SizeOp0 % NarrowSize != 0) 612 return UnableToLegalize; 613 int NumParts = SizeOp0 / NarrowSize; 614 615 SmallVector<Register, 2> DstRegs; 616 for (int i = 0; i < NumParts; ++i) 617 DstRegs.push_back( 618 MIRBuilder.buildUndef(NarrowTy)->getOperand(0).getReg()); 619 620 Register DstReg = MI.getOperand(0).getReg(); 621 if(MRI.getType(DstReg).isVector()) 622 MIRBuilder.buildBuildVector(DstReg, DstRegs); 623 else 624 MIRBuilder.buildMerge(DstReg, DstRegs); 625 MI.eraseFromParent(); 626 return Legalized; 627 } 628 case TargetOpcode::G_CONSTANT: { 629 LLT Ty = MRI.getType(MI.getOperand(0).getReg()); 630 const APInt &Val = MI.getOperand(1).getCImm()->getValue(); 631 unsigned TotalSize = Ty.getSizeInBits(); 632 unsigned NarrowSize = NarrowTy.getSizeInBits(); 633 int NumParts = TotalSize / NarrowSize; 634 635 SmallVector<Register, 4> PartRegs; 636 for (int I = 0; I != NumParts; ++I) { 637 unsigned Offset = I * NarrowSize; 638 auto K = MIRBuilder.buildConstant(NarrowTy, 639 Val.lshr(Offset).trunc(NarrowSize)); 640 PartRegs.push_back(K.getReg(0)); 641 } 642 643 LLT LeftoverTy; 644 unsigned LeftoverBits = TotalSize - NumParts * NarrowSize; 645 SmallVector<Register, 1> LeftoverRegs; 646 if (LeftoverBits != 0) { 647 LeftoverTy = LLT::scalar(LeftoverBits); 648 auto K = MIRBuilder.buildConstant( 649 LeftoverTy, 650 Val.lshr(NumParts * NarrowSize).trunc(LeftoverBits)); 651 LeftoverRegs.push_back(K.getReg(0)); 652 } 653 654 insertParts(MI.getOperand(0).getReg(), 655 Ty, NarrowTy, PartRegs, LeftoverTy, LeftoverRegs); 656 657 MI.eraseFromParent(); 658 return Legalized; 659 } 660 case TargetOpcode::G_SEXT: { 661 if (TypeIdx != 0) 662 return UnableToLegalize; 663 664 Register SrcReg = MI.getOperand(1).getReg(); 665 LLT SrcTy = MRI.getType(SrcReg); 666 667 // FIXME: support the general case where the requested NarrowTy may not be 668 // the same as the source type. E.g. s128 = sext(s32) 669 if ((SrcTy.getSizeInBits() != SizeOp0 / 2) || 670 SrcTy.getSizeInBits() != NarrowTy.getSizeInBits()) { 671 LLVM_DEBUG(dbgs() << "Can't narrow sext to type " << NarrowTy << "\n"); 672 return UnableToLegalize; 673 } 674 675 // Shift the sign bit of the low register through the high register. 676 auto ShiftAmt = 677 MIRBuilder.buildConstant(LLT::scalar(64), NarrowTy.getSizeInBits() - 1); 678 auto Shift = MIRBuilder.buildAShr(NarrowTy, SrcReg, ShiftAmt); 679 MIRBuilder.buildMerge(MI.getOperand(0).getReg(), {SrcReg, Shift.getReg(0)}); 680 MI.eraseFromParent(); 681 return Legalized; 682 } 683 case TargetOpcode::G_ZEXT: { 684 if (TypeIdx != 0) 685 return UnableToLegalize; 686 687 LLT SrcTy = MRI.getType(MI.getOperand(1).getReg()); 688 uint64_t SizeOp1 = SrcTy.getSizeInBits(); 689 if (SizeOp0 % SizeOp1 != 0) 690 return UnableToLegalize; 691 692 // Generate a merge where the bottom bits are taken from the source, and 693 // zero everything else. 694 Register ZeroReg = MIRBuilder.buildConstant(SrcTy, 0).getReg(0); 695 unsigned NumParts = SizeOp0 / SizeOp1; 696 SmallVector<Register, 4> Srcs = {MI.getOperand(1).getReg()}; 697 for (unsigned Part = 1; Part < NumParts; ++Part) 698 Srcs.push_back(ZeroReg); 699 MIRBuilder.buildMerge(MI.getOperand(0).getReg(), Srcs); 700 MI.eraseFromParent(); 701 return Legalized; 702 } 703 case TargetOpcode::G_TRUNC: { 704 if (TypeIdx != 1) 705 return UnableToLegalize; 706 707 uint64_t SizeOp1 = MRI.getType(MI.getOperand(1).getReg()).getSizeInBits(); 708 if (NarrowTy.getSizeInBits() * 2 != SizeOp1) { 709 LLVM_DEBUG(dbgs() << "Can't narrow trunc to type " << NarrowTy << "\n"); 710 return UnableToLegalize; 711 } 712 713 auto Unmerge = MIRBuilder.buildUnmerge(NarrowTy, MI.getOperand(1).getReg()); 714 MIRBuilder.buildCopy(MI.getOperand(0).getReg(), Unmerge.getReg(0)); 715 MI.eraseFromParent(); 716 return Legalized; 717 } 718 719 case TargetOpcode::G_ADD: { 720 // FIXME: add support for when SizeOp0 isn't an exact multiple of 721 // NarrowSize. 722 if (SizeOp0 % NarrowSize != 0) 723 return UnableToLegalize; 724 // Expand in terms of carry-setting/consuming G_ADDE instructions. 725 int NumParts = SizeOp0 / NarrowTy.getSizeInBits(); 726 727 SmallVector<Register, 2> Src1Regs, Src2Regs, DstRegs; 728 extractParts(MI.getOperand(1).getReg(), NarrowTy, NumParts, Src1Regs); 729 extractParts(MI.getOperand(2).getReg(), NarrowTy, NumParts, Src2Regs); 730 731 Register CarryIn; 732 for (int i = 0; i < NumParts; ++i) { 733 Register DstReg = MRI.createGenericVirtualRegister(NarrowTy); 734 Register CarryOut = MRI.createGenericVirtualRegister(LLT::scalar(1)); 735 736 if (i == 0) 737 MIRBuilder.buildUAddo(DstReg, CarryOut, Src1Regs[i], Src2Regs[i]); 738 else { 739 MIRBuilder.buildUAdde(DstReg, CarryOut, Src1Regs[i], 740 Src2Regs[i], CarryIn); 741 } 742 743 DstRegs.push_back(DstReg); 744 CarryIn = CarryOut; 745 } 746 Register DstReg = MI.getOperand(0).getReg(); 747 if(MRI.getType(DstReg).isVector()) 748 MIRBuilder.buildBuildVector(DstReg, DstRegs); 749 else 750 MIRBuilder.buildMerge(DstReg, DstRegs); 751 MI.eraseFromParent(); 752 return Legalized; 753 } 754 case TargetOpcode::G_SUB: { 755 // FIXME: add support for when SizeOp0 isn't an exact multiple of 756 // NarrowSize. 757 if (SizeOp0 % NarrowSize != 0) 758 return UnableToLegalize; 759 760 int NumParts = SizeOp0 / NarrowTy.getSizeInBits(); 761 762 SmallVector<Register, 2> Src1Regs, Src2Regs, DstRegs; 763 extractParts(MI.getOperand(1).getReg(), NarrowTy, NumParts, Src1Regs); 764 extractParts(MI.getOperand(2).getReg(), NarrowTy, NumParts, Src2Regs); 765 766 Register DstReg = MRI.createGenericVirtualRegister(NarrowTy); 767 Register BorrowOut = MRI.createGenericVirtualRegister(LLT::scalar(1)); 768 MIRBuilder.buildInstr(TargetOpcode::G_USUBO, {DstReg, BorrowOut}, 769 {Src1Regs[0], Src2Regs[0]}); 770 DstRegs.push_back(DstReg); 771 Register BorrowIn = BorrowOut; 772 for (int i = 1; i < NumParts; ++i) { 773 DstReg = MRI.createGenericVirtualRegister(NarrowTy); 774 BorrowOut = MRI.createGenericVirtualRegister(LLT::scalar(1)); 775 776 MIRBuilder.buildInstr(TargetOpcode::G_USUBE, {DstReg, BorrowOut}, 777 {Src1Regs[i], Src2Regs[i], BorrowIn}); 778 779 DstRegs.push_back(DstReg); 780 BorrowIn = BorrowOut; 781 } 782 MIRBuilder.buildMerge(MI.getOperand(0).getReg(), DstRegs); 783 MI.eraseFromParent(); 784 return Legalized; 785 } 786 case TargetOpcode::G_MUL: 787 case TargetOpcode::G_UMULH: 788 return narrowScalarMul(MI, NarrowTy); 789 case TargetOpcode::G_EXTRACT: 790 return narrowScalarExtract(MI, TypeIdx, NarrowTy); 791 case TargetOpcode::G_INSERT: 792 return narrowScalarInsert(MI, TypeIdx, NarrowTy); 793 case TargetOpcode::G_LOAD: { 794 const auto &MMO = **MI.memoperands_begin(); 795 Register DstReg = MI.getOperand(0).getReg(); 796 LLT DstTy = MRI.getType(DstReg); 797 if (DstTy.isVector()) 798 return UnableToLegalize; 799 800 if (8 * MMO.getSize() != DstTy.getSizeInBits()) { 801 Register TmpReg = MRI.createGenericVirtualRegister(NarrowTy); 802 auto &MMO = **MI.memoperands_begin(); 803 MIRBuilder.buildLoad(TmpReg, MI.getOperand(1).getReg(), MMO); 804 MIRBuilder.buildAnyExt(DstReg, TmpReg); 805 MI.eraseFromParent(); 806 return Legalized; 807 } 808 809 return reduceLoadStoreWidth(MI, TypeIdx, NarrowTy); 810 } 811 case TargetOpcode::G_ZEXTLOAD: 812 case TargetOpcode::G_SEXTLOAD: { 813 bool ZExt = MI.getOpcode() == TargetOpcode::G_ZEXTLOAD; 814 Register DstReg = MI.getOperand(0).getReg(); 815 Register PtrReg = MI.getOperand(1).getReg(); 816 817 Register TmpReg = MRI.createGenericVirtualRegister(NarrowTy); 818 auto &MMO = **MI.memoperands_begin(); 819 if (MMO.getSizeInBits() == NarrowSize) { 820 MIRBuilder.buildLoad(TmpReg, PtrReg, MMO); 821 } else { 822 unsigned ExtLoad = ZExt ? TargetOpcode::G_ZEXTLOAD 823 : TargetOpcode::G_SEXTLOAD; 824 MIRBuilder.buildInstr(ExtLoad) 825 .addDef(TmpReg) 826 .addUse(PtrReg) 827 .addMemOperand(&MMO); 828 } 829 830 if (ZExt) 831 MIRBuilder.buildZExt(DstReg, TmpReg); 832 else 833 MIRBuilder.buildSExt(DstReg, TmpReg); 834 835 MI.eraseFromParent(); 836 return Legalized; 837 } 838 case TargetOpcode::G_STORE: { 839 const auto &MMO = **MI.memoperands_begin(); 840 841 Register SrcReg = MI.getOperand(0).getReg(); 842 LLT SrcTy = MRI.getType(SrcReg); 843 if (SrcTy.isVector()) 844 return UnableToLegalize; 845 846 int NumParts = SizeOp0 / NarrowSize; 847 unsigned HandledSize = NumParts * NarrowTy.getSizeInBits(); 848 unsigned LeftoverBits = SrcTy.getSizeInBits() - HandledSize; 849 if (SrcTy.isVector() && LeftoverBits != 0) 850 return UnableToLegalize; 851 852 if (8 * MMO.getSize() != SrcTy.getSizeInBits()) { 853 Register TmpReg = MRI.createGenericVirtualRegister(NarrowTy); 854 auto &MMO = **MI.memoperands_begin(); 855 MIRBuilder.buildTrunc(TmpReg, SrcReg); 856 MIRBuilder.buildStore(TmpReg, MI.getOperand(1).getReg(), MMO); 857 MI.eraseFromParent(); 858 return Legalized; 859 } 860 861 return reduceLoadStoreWidth(MI, 0, NarrowTy); 862 } 863 case TargetOpcode::G_SELECT: 864 return narrowScalarSelect(MI, TypeIdx, NarrowTy); 865 case TargetOpcode::G_AND: 866 case TargetOpcode::G_OR: 867 case TargetOpcode::G_XOR: { 868 // Legalize bitwise operation: 869 // A = BinOp<Ty> B, C 870 // into: 871 // B1, ..., BN = G_UNMERGE_VALUES B 872 // C1, ..., CN = G_UNMERGE_VALUES C 873 // A1 = BinOp<Ty/N> B1, C2 874 // ... 875 // AN = BinOp<Ty/N> BN, CN 876 // A = G_MERGE_VALUES A1, ..., AN 877 return narrowScalarBasic(MI, TypeIdx, NarrowTy); 878 } 879 case TargetOpcode::G_SHL: 880 case TargetOpcode::G_LSHR: 881 case TargetOpcode::G_ASHR: 882 return narrowScalarShift(MI, TypeIdx, NarrowTy); 883 case TargetOpcode::G_CTLZ: 884 case TargetOpcode::G_CTLZ_ZERO_UNDEF: 885 case TargetOpcode::G_CTTZ: 886 case TargetOpcode::G_CTTZ_ZERO_UNDEF: 887 case TargetOpcode::G_CTPOP: 888 if (TypeIdx != 0) 889 return UnableToLegalize; // TODO 890 891 Observer.changingInstr(MI); 892 narrowScalarDst(MI, NarrowTy, 0, TargetOpcode::G_ZEXT); 893 Observer.changedInstr(MI); 894 return Legalized; 895 case TargetOpcode::G_INTTOPTR: 896 if (TypeIdx != 1) 897 return UnableToLegalize; 898 899 Observer.changingInstr(MI); 900 narrowScalarSrc(MI, NarrowTy, 1); 901 Observer.changedInstr(MI); 902 return Legalized; 903 case TargetOpcode::G_PTRTOINT: 904 if (TypeIdx != 0) 905 return UnableToLegalize; 906 907 Observer.changingInstr(MI); 908 narrowScalarDst(MI, NarrowTy, 0, TargetOpcode::G_ZEXT); 909 Observer.changedInstr(MI); 910 return Legalized; 911 case TargetOpcode::G_PHI: { 912 unsigned NumParts = SizeOp0 / NarrowSize; 913 SmallVector<Register, 2> DstRegs; 914 SmallVector<SmallVector<Register, 2>, 2> SrcRegs; 915 DstRegs.resize(NumParts); 916 SrcRegs.resize(MI.getNumOperands() / 2); 917 Observer.changingInstr(MI); 918 for (unsigned i = 1; i < MI.getNumOperands(); i += 2) { 919 MachineBasicBlock &OpMBB = *MI.getOperand(i + 1).getMBB(); 920 MIRBuilder.setInsertPt(OpMBB, OpMBB.getFirstTerminator()); 921 extractParts(MI.getOperand(i).getReg(), NarrowTy, NumParts, 922 SrcRegs[i / 2]); 923 } 924 MachineBasicBlock &MBB = *MI.getParent(); 925 MIRBuilder.setInsertPt(MBB, MI); 926 for (unsigned i = 0; i < NumParts; ++i) { 927 DstRegs[i] = MRI.createGenericVirtualRegister(NarrowTy); 928 MachineInstrBuilder MIB = 929 MIRBuilder.buildInstr(TargetOpcode::G_PHI).addDef(DstRegs[i]); 930 for (unsigned j = 1; j < MI.getNumOperands(); j += 2) 931 MIB.addUse(SrcRegs[j / 2][i]).add(MI.getOperand(j + 1)); 932 } 933 MIRBuilder.setInsertPt(MBB, MBB.getFirstNonPHI()); 934 MIRBuilder.buildMerge(MI.getOperand(0).getReg(), DstRegs); 935 Observer.changedInstr(MI); 936 MI.eraseFromParent(); 937 return Legalized; 938 } 939 case TargetOpcode::G_EXTRACT_VECTOR_ELT: 940 case TargetOpcode::G_INSERT_VECTOR_ELT: { 941 if (TypeIdx != 2) 942 return UnableToLegalize; 943 944 int OpIdx = MI.getOpcode() == TargetOpcode::G_EXTRACT_VECTOR_ELT ? 2 : 3; 945 Observer.changingInstr(MI); 946 narrowScalarSrc(MI, NarrowTy, OpIdx); 947 Observer.changedInstr(MI); 948 return Legalized; 949 } 950 case TargetOpcode::G_ICMP: { 951 uint64_t SrcSize = MRI.getType(MI.getOperand(2).getReg()).getSizeInBits(); 952 if (NarrowSize * 2 != SrcSize) 953 return UnableToLegalize; 954 955 Observer.changingInstr(MI); 956 Register LHSL = MRI.createGenericVirtualRegister(NarrowTy); 957 Register LHSH = MRI.createGenericVirtualRegister(NarrowTy); 958 MIRBuilder.buildUnmerge({LHSL, LHSH}, MI.getOperand(2).getReg()); 959 960 Register RHSL = MRI.createGenericVirtualRegister(NarrowTy); 961 Register RHSH = MRI.createGenericVirtualRegister(NarrowTy); 962 MIRBuilder.buildUnmerge({RHSL, RHSH}, MI.getOperand(3).getReg()); 963 964 CmpInst::Predicate Pred = 965 static_cast<CmpInst::Predicate>(MI.getOperand(1).getPredicate()); 966 LLT ResTy = MRI.getType(MI.getOperand(0).getReg()); 967 968 if (Pred == CmpInst::ICMP_EQ || Pred == CmpInst::ICMP_NE) { 969 MachineInstrBuilder XorL = MIRBuilder.buildXor(NarrowTy, LHSL, RHSL); 970 MachineInstrBuilder XorH = MIRBuilder.buildXor(NarrowTy, LHSH, RHSH); 971 MachineInstrBuilder Or = MIRBuilder.buildOr(NarrowTy, XorL, XorH); 972 MachineInstrBuilder Zero = MIRBuilder.buildConstant(NarrowTy, 0); 973 MIRBuilder.buildICmp(Pred, MI.getOperand(0).getReg(), Or, Zero); 974 } else { 975 MachineInstrBuilder CmpH = MIRBuilder.buildICmp(Pred, ResTy, LHSH, RHSH); 976 MachineInstrBuilder CmpHEQ = 977 MIRBuilder.buildICmp(CmpInst::Predicate::ICMP_EQ, ResTy, LHSH, RHSH); 978 MachineInstrBuilder CmpLU = MIRBuilder.buildICmp( 979 ICmpInst::getUnsignedPredicate(Pred), ResTy, LHSL, RHSL); 980 MIRBuilder.buildSelect(MI.getOperand(0).getReg(), CmpHEQ, CmpLU, CmpH); 981 } 982 Observer.changedInstr(MI); 983 MI.eraseFromParent(); 984 return Legalized; 985 } 986 case TargetOpcode::G_SEXT_INREG: { 987 if (TypeIdx != 0) 988 return UnableToLegalize; 989 990 if (!MI.getOperand(2).isImm()) 991 return UnableToLegalize; 992 int64_t SizeInBits = MI.getOperand(2).getImm(); 993 994 // So long as the new type has more bits than the bits we're extending we 995 // don't need to break it apart. 996 if (NarrowTy.getScalarSizeInBits() >= SizeInBits) { 997 Observer.changingInstr(MI); 998 // We don't lose any non-extension bits by truncating the src and 999 // sign-extending the dst. 1000 MachineOperand &MO1 = MI.getOperand(1); 1001 auto TruncMIB = MIRBuilder.buildTrunc(NarrowTy, MO1.getReg()); 1002 MO1.setReg(TruncMIB->getOperand(0).getReg()); 1003 1004 MachineOperand &MO2 = MI.getOperand(0); 1005 Register DstExt = MRI.createGenericVirtualRegister(NarrowTy); 1006 MIRBuilder.setInsertPt(MIRBuilder.getMBB(), ++MIRBuilder.getInsertPt()); 1007 MIRBuilder.buildInstr(TargetOpcode::G_SEXT, {MO2.getReg()}, {DstExt}); 1008 MO2.setReg(DstExt); 1009 Observer.changedInstr(MI); 1010 return Legalized; 1011 } 1012 1013 // Break it apart. Components below the extension point are unmodified. The 1014 // component containing the extension point becomes a narrower SEXT_INREG. 1015 // Components above it are ashr'd from the component containing the 1016 // extension point. 1017 if (SizeOp0 % NarrowSize != 0) 1018 return UnableToLegalize; 1019 int NumParts = SizeOp0 / NarrowSize; 1020 1021 // List the registers where the destination will be scattered. 1022 SmallVector<Register, 2> DstRegs; 1023 // List the registers where the source will be split. 1024 SmallVector<Register, 2> SrcRegs; 1025 1026 // Create all the temporary registers. 1027 for (int i = 0; i < NumParts; ++i) { 1028 Register SrcReg = MRI.createGenericVirtualRegister(NarrowTy); 1029 1030 SrcRegs.push_back(SrcReg); 1031 } 1032 1033 // Explode the big arguments into smaller chunks. 1034 MIRBuilder.buildUnmerge(SrcRegs, MI.getOperand(1).getReg()); 1035 1036 Register AshrCstReg = 1037 MIRBuilder.buildConstant(NarrowTy, NarrowTy.getScalarSizeInBits() - 1) 1038 ->getOperand(0) 1039 .getReg(); 1040 Register FullExtensionReg = 0; 1041 Register PartialExtensionReg = 0; 1042 1043 // Do the operation on each small part. 1044 for (int i = 0; i < NumParts; ++i) { 1045 if ((i + 1) * NarrowTy.getScalarSizeInBits() < SizeInBits) 1046 DstRegs.push_back(SrcRegs[i]); 1047 else if (i * NarrowTy.getScalarSizeInBits() > SizeInBits) { 1048 assert(PartialExtensionReg && 1049 "Expected to visit partial extension before full"); 1050 if (FullExtensionReg) { 1051 DstRegs.push_back(FullExtensionReg); 1052 continue; 1053 } 1054 DstRegs.push_back(MIRBuilder 1055 .buildInstr(TargetOpcode::G_ASHR, {NarrowTy}, 1056 {PartialExtensionReg, AshrCstReg}) 1057 ->getOperand(0) 1058 .getReg()); 1059 FullExtensionReg = DstRegs.back(); 1060 } else { 1061 DstRegs.push_back( 1062 MIRBuilder 1063 .buildInstr( 1064 TargetOpcode::G_SEXT_INREG, {NarrowTy}, 1065 {SrcRegs[i], SizeInBits % NarrowTy.getScalarSizeInBits()}) 1066 ->getOperand(0) 1067 .getReg()); 1068 PartialExtensionReg = DstRegs.back(); 1069 } 1070 } 1071 1072 // Gather the destination registers into the final destination. 1073 Register DstReg = MI.getOperand(0).getReg(); 1074 MIRBuilder.buildMerge(DstReg, DstRegs); 1075 MI.eraseFromParent(); 1076 return Legalized; 1077 } 1078 case TargetOpcode::G_BSWAP: 1079 case TargetOpcode::G_BITREVERSE: { 1080 if (SizeOp0 % NarrowSize != 0) 1081 return UnableToLegalize; 1082 1083 Observer.changingInstr(MI); 1084 SmallVector<Register, 2> SrcRegs, DstRegs; 1085 unsigned NumParts = SizeOp0 / NarrowSize; 1086 extractParts(MI.getOperand(1).getReg(), NarrowTy, NumParts, SrcRegs); 1087 1088 for (unsigned i = 0; i < NumParts; ++i) { 1089 auto DstPart = MIRBuilder.buildInstr(MI.getOpcode(), {NarrowTy}, 1090 {SrcRegs[NumParts - 1 - i]}); 1091 DstRegs.push_back(DstPart.getReg(0)); 1092 } 1093 1094 MIRBuilder.buildMerge(MI.getOperand(0).getReg(), DstRegs); 1095 1096 Observer.changedInstr(MI); 1097 MI.eraseFromParent(); 1098 return Legalized; 1099 } 1100 } 1101 } 1102 1103 void LegalizerHelper::widenScalarSrc(MachineInstr &MI, LLT WideTy, 1104 unsigned OpIdx, unsigned ExtOpcode) { 1105 MachineOperand &MO = MI.getOperand(OpIdx); 1106 auto ExtB = MIRBuilder.buildInstr(ExtOpcode, {WideTy}, {MO.getReg()}); 1107 MO.setReg(ExtB->getOperand(0).getReg()); 1108 } 1109 1110 void LegalizerHelper::narrowScalarSrc(MachineInstr &MI, LLT NarrowTy, 1111 unsigned OpIdx) { 1112 MachineOperand &MO = MI.getOperand(OpIdx); 1113 auto ExtB = MIRBuilder.buildInstr(TargetOpcode::G_TRUNC, {NarrowTy}, 1114 {MO.getReg()}); 1115 MO.setReg(ExtB->getOperand(0).getReg()); 1116 } 1117 1118 void LegalizerHelper::widenScalarDst(MachineInstr &MI, LLT WideTy, 1119 unsigned OpIdx, unsigned TruncOpcode) { 1120 MachineOperand &MO = MI.getOperand(OpIdx); 1121 Register DstExt = MRI.createGenericVirtualRegister(WideTy); 1122 MIRBuilder.setInsertPt(MIRBuilder.getMBB(), ++MIRBuilder.getInsertPt()); 1123 MIRBuilder.buildInstr(TruncOpcode, {MO.getReg()}, {DstExt}); 1124 MO.setReg(DstExt); 1125 } 1126 1127 void LegalizerHelper::narrowScalarDst(MachineInstr &MI, LLT NarrowTy, 1128 unsigned OpIdx, unsigned ExtOpcode) { 1129 MachineOperand &MO = MI.getOperand(OpIdx); 1130 Register DstTrunc = MRI.createGenericVirtualRegister(NarrowTy); 1131 MIRBuilder.setInsertPt(MIRBuilder.getMBB(), ++MIRBuilder.getInsertPt()); 1132 MIRBuilder.buildInstr(ExtOpcode, {MO.getReg()}, {DstTrunc}); 1133 MO.setReg(DstTrunc); 1134 } 1135 1136 void LegalizerHelper::moreElementsVectorDst(MachineInstr &MI, LLT WideTy, 1137 unsigned OpIdx) { 1138 MachineOperand &MO = MI.getOperand(OpIdx); 1139 Register DstExt = MRI.createGenericVirtualRegister(WideTy); 1140 MIRBuilder.setInsertPt(MIRBuilder.getMBB(), ++MIRBuilder.getInsertPt()); 1141 MIRBuilder.buildExtract(MO.getReg(), DstExt, 0); 1142 MO.setReg(DstExt); 1143 } 1144 1145 void LegalizerHelper::moreElementsVectorSrc(MachineInstr &MI, LLT MoreTy, 1146 unsigned OpIdx) { 1147 MachineOperand &MO = MI.getOperand(OpIdx); 1148 1149 LLT OldTy = MRI.getType(MO.getReg()); 1150 unsigned OldElts = OldTy.getNumElements(); 1151 unsigned NewElts = MoreTy.getNumElements(); 1152 1153 unsigned NumParts = NewElts / OldElts; 1154 1155 // Use concat_vectors if the result is a multiple of the number of elements. 1156 if (NumParts * OldElts == NewElts) { 1157 SmallVector<Register, 8> Parts; 1158 Parts.push_back(MO.getReg()); 1159 1160 Register ImpDef = MIRBuilder.buildUndef(OldTy).getReg(0); 1161 for (unsigned I = 1; I != NumParts; ++I) 1162 Parts.push_back(ImpDef); 1163 1164 auto Concat = MIRBuilder.buildConcatVectors(MoreTy, Parts); 1165 MO.setReg(Concat.getReg(0)); 1166 return; 1167 } 1168 1169 Register MoreReg = MRI.createGenericVirtualRegister(MoreTy); 1170 Register ImpDef = MIRBuilder.buildUndef(MoreTy).getReg(0); 1171 MIRBuilder.buildInsert(MoreReg, ImpDef, MO.getReg(), 0); 1172 MO.setReg(MoreReg); 1173 } 1174 1175 LegalizerHelper::LegalizeResult 1176 LegalizerHelper::widenScalarMergeValues(MachineInstr &MI, unsigned TypeIdx, 1177 LLT WideTy) { 1178 if (TypeIdx != 1) 1179 return UnableToLegalize; 1180 1181 Register DstReg = MI.getOperand(0).getReg(); 1182 LLT DstTy = MRI.getType(DstReg); 1183 if (DstTy.isVector()) 1184 return UnableToLegalize; 1185 1186 Register Src1 = MI.getOperand(1).getReg(); 1187 LLT SrcTy = MRI.getType(Src1); 1188 const int DstSize = DstTy.getSizeInBits(); 1189 const int SrcSize = SrcTy.getSizeInBits(); 1190 const int WideSize = WideTy.getSizeInBits(); 1191 const int NumMerge = (DstSize + WideSize - 1) / WideSize; 1192 1193 unsigned NumOps = MI.getNumOperands(); 1194 unsigned NumSrc = MI.getNumOperands() - 1; 1195 unsigned PartSize = DstTy.getSizeInBits() / NumSrc; 1196 1197 if (WideSize >= DstSize) { 1198 // Directly pack the bits in the target type. 1199 Register ResultReg = MIRBuilder.buildZExt(WideTy, Src1).getReg(0); 1200 1201 for (unsigned I = 2; I != NumOps; ++I) { 1202 const unsigned Offset = (I - 1) * PartSize; 1203 1204 Register SrcReg = MI.getOperand(I).getReg(); 1205 assert(MRI.getType(SrcReg) == LLT::scalar(PartSize)); 1206 1207 auto ZextInput = MIRBuilder.buildZExt(WideTy, SrcReg); 1208 1209 Register NextResult = I + 1 == NumOps && WideTy == DstTy ? DstReg : 1210 MRI.createGenericVirtualRegister(WideTy); 1211 1212 auto ShiftAmt = MIRBuilder.buildConstant(WideTy, Offset); 1213 auto Shl = MIRBuilder.buildShl(WideTy, ZextInput, ShiftAmt); 1214 MIRBuilder.buildOr(NextResult, ResultReg, Shl); 1215 ResultReg = NextResult; 1216 } 1217 1218 if (WideSize > DstSize) 1219 MIRBuilder.buildTrunc(DstReg, ResultReg); 1220 else if (DstTy.isPointer()) 1221 MIRBuilder.buildIntToPtr(DstReg, ResultReg); 1222 1223 MI.eraseFromParent(); 1224 return Legalized; 1225 } 1226 1227 // Unmerge the original values to the GCD type, and recombine to the next 1228 // multiple greater than the original type. 1229 // 1230 // %3:_(s12) = G_MERGE_VALUES %0:_(s4), %1:_(s4), %2:_(s4) -> s6 1231 // %4:_(s2), %5:_(s2) = G_UNMERGE_VALUES %0 1232 // %6:_(s2), %7:_(s2) = G_UNMERGE_VALUES %1 1233 // %8:_(s2), %9:_(s2) = G_UNMERGE_VALUES %2 1234 // %10:_(s6) = G_MERGE_VALUES %4, %5, %6 1235 // %11:_(s6) = G_MERGE_VALUES %7, %8, %9 1236 // %12:_(s12) = G_MERGE_VALUES %10, %11 1237 // 1238 // Padding with undef if necessary: 1239 // 1240 // %2:_(s8) = G_MERGE_VALUES %0:_(s4), %1:_(s4) -> s6 1241 // %3:_(s2), %4:_(s2) = G_UNMERGE_VALUES %0 1242 // %5:_(s2), %6:_(s2) = G_UNMERGE_VALUES %1 1243 // %7:_(s2) = G_IMPLICIT_DEF 1244 // %8:_(s6) = G_MERGE_VALUES %3, %4, %5 1245 // %9:_(s6) = G_MERGE_VALUES %6, %7, %7 1246 // %10:_(s12) = G_MERGE_VALUES %8, %9 1247 1248 const int GCD = greatestCommonDivisor(SrcSize, WideSize); 1249 LLT GCDTy = LLT::scalar(GCD); 1250 1251 SmallVector<Register, 8> Parts; 1252 SmallVector<Register, 8> NewMergeRegs; 1253 SmallVector<Register, 8> Unmerges; 1254 LLT WideDstTy = LLT::scalar(NumMerge * WideSize); 1255 1256 // Decompose the original operands if they don't evenly divide. 1257 for (int I = 1, E = MI.getNumOperands(); I != E; ++I) { 1258 Register SrcReg = MI.getOperand(I).getReg(); 1259 if (GCD == SrcSize) { 1260 Unmerges.push_back(SrcReg); 1261 } else { 1262 auto Unmerge = MIRBuilder.buildUnmerge(GCDTy, SrcReg); 1263 for (int J = 0, JE = Unmerge->getNumOperands() - 1; J != JE; ++J) 1264 Unmerges.push_back(Unmerge.getReg(J)); 1265 } 1266 } 1267 1268 // Pad with undef to the next size that is a multiple of the requested size. 1269 if (static_cast<int>(Unmerges.size()) != NumMerge * WideSize) { 1270 Register UndefReg = MIRBuilder.buildUndef(GCDTy).getReg(0); 1271 for (int I = Unmerges.size(); I != NumMerge * WideSize; ++I) 1272 Unmerges.push_back(UndefReg); 1273 } 1274 1275 const int PartsPerGCD = WideSize / GCD; 1276 1277 // Build merges of each piece. 1278 ArrayRef<Register> Slicer(Unmerges); 1279 for (int I = 0; I != NumMerge; ++I, Slicer = Slicer.drop_front(PartsPerGCD)) { 1280 auto Merge = MIRBuilder.buildMerge(WideTy, Slicer.take_front(PartsPerGCD)); 1281 NewMergeRegs.push_back(Merge.getReg(0)); 1282 } 1283 1284 // A truncate may be necessary if the requested type doesn't evenly divide the 1285 // original result type. 1286 if (DstTy.getSizeInBits() == WideDstTy.getSizeInBits()) { 1287 MIRBuilder.buildMerge(DstReg, NewMergeRegs); 1288 } else { 1289 auto FinalMerge = MIRBuilder.buildMerge(WideDstTy, NewMergeRegs); 1290 MIRBuilder.buildTrunc(DstReg, FinalMerge.getReg(0)); 1291 } 1292 1293 MI.eraseFromParent(); 1294 return Legalized; 1295 } 1296 1297 LegalizerHelper::LegalizeResult 1298 LegalizerHelper::widenScalarUnmergeValues(MachineInstr &MI, unsigned TypeIdx, 1299 LLT WideTy) { 1300 if (TypeIdx != 0) 1301 return UnableToLegalize; 1302 1303 unsigned NumDst = MI.getNumOperands() - 1; 1304 Register SrcReg = MI.getOperand(NumDst).getReg(); 1305 LLT SrcTy = MRI.getType(SrcReg); 1306 if (!SrcTy.isScalar()) 1307 return UnableToLegalize; 1308 1309 Register Dst0Reg = MI.getOperand(0).getReg(); 1310 LLT DstTy = MRI.getType(Dst0Reg); 1311 if (!DstTy.isScalar()) 1312 return UnableToLegalize; 1313 1314 unsigned NewSrcSize = NumDst * WideTy.getSizeInBits(); 1315 LLT NewSrcTy = LLT::scalar(NewSrcSize); 1316 unsigned SizeDiff = WideTy.getSizeInBits() - DstTy.getSizeInBits(); 1317 1318 auto WideSrc = MIRBuilder.buildZExt(NewSrcTy, SrcReg); 1319 1320 for (unsigned I = 1; I != NumDst; ++I) { 1321 auto ShiftAmt = MIRBuilder.buildConstant(NewSrcTy, SizeDiff * I); 1322 auto Shl = MIRBuilder.buildShl(NewSrcTy, WideSrc, ShiftAmt); 1323 WideSrc = MIRBuilder.buildOr(NewSrcTy, WideSrc, Shl); 1324 } 1325 1326 Observer.changingInstr(MI); 1327 1328 MI.getOperand(NumDst).setReg(WideSrc->getOperand(0).getReg()); 1329 for (unsigned I = 0; I != NumDst; ++I) 1330 widenScalarDst(MI, WideTy, I); 1331 1332 Observer.changedInstr(MI); 1333 1334 return Legalized; 1335 } 1336 1337 LegalizerHelper::LegalizeResult 1338 LegalizerHelper::widenScalarExtract(MachineInstr &MI, unsigned TypeIdx, 1339 LLT WideTy) { 1340 Register DstReg = MI.getOperand(0).getReg(); 1341 Register SrcReg = MI.getOperand(1).getReg(); 1342 LLT SrcTy = MRI.getType(SrcReg); 1343 1344 LLT DstTy = MRI.getType(DstReg); 1345 unsigned Offset = MI.getOperand(2).getImm(); 1346 1347 if (TypeIdx == 0) { 1348 if (SrcTy.isVector() || DstTy.isVector()) 1349 return UnableToLegalize; 1350 1351 SrcOp Src(SrcReg); 1352 if (SrcTy.isPointer()) { 1353 // Extracts from pointers can be handled only if they are really just 1354 // simple integers. 1355 const DataLayout &DL = MIRBuilder.getDataLayout(); 1356 if (DL.isNonIntegralAddressSpace(SrcTy.getAddressSpace())) 1357 return UnableToLegalize; 1358 1359 LLT SrcAsIntTy = LLT::scalar(SrcTy.getSizeInBits()); 1360 Src = MIRBuilder.buildPtrToInt(SrcAsIntTy, Src); 1361 SrcTy = SrcAsIntTy; 1362 } 1363 1364 if (DstTy.isPointer()) 1365 return UnableToLegalize; 1366 1367 if (Offset == 0) { 1368 // Avoid a shift in the degenerate case. 1369 MIRBuilder.buildTrunc(DstReg, 1370 MIRBuilder.buildAnyExtOrTrunc(WideTy, Src)); 1371 MI.eraseFromParent(); 1372 return Legalized; 1373 } 1374 1375 // Do a shift in the source type. 1376 LLT ShiftTy = SrcTy; 1377 if (WideTy.getSizeInBits() > SrcTy.getSizeInBits()) { 1378 Src = MIRBuilder.buildAnyExt(WideTy, Src); 1379 ShiftTy = WideTy; 1380 } else if (WideTy.getSizeInBits() > SrcTy.getSizeInBits()) 1381 return UnableToLegalize; 1382 1383 auto LShr = MIRBuilder.buildLShr( 1384 ShiftTy, Src, MIRBuilder.buildConstant(ShiftTy, Offset)); 1385 MIRBuilder.buildTrunc(DstReg, LShr); 1386 MI.eraseFromParent(); 1387 return Legalized; 1388 } 1389 1390 if (SrcTy.isScalar()) { 1391 Observer.changingInstr(MI); 1392 widenScalarSrc(MI, WideTy, 1, TargetOpcode::G_ANYEXT); 1393 Observer.changedInstr(MI); 1394 return Legalized; 1395 } 1396 1397 if (!SrcTy.isVector()) 1398 return UnableToLegalize; 1399 1400 if (DstTy != SrcTy.getElementType()) 1401 return UnableToLegalize; 1402 1403 if (Offset % SrcTy.getScalarSizeInBits() != 0) 1404 return UnableToLegalize; 1405 1406 Observer.changingInstr(MI); 1407 widenScalarSrc(MI, WideTy, 1, TargetOpcode::G_ANYEXT); 1408 1409 MI.getOperand(2).setImm((WideTy.getSizeInBits() / SrcTy.getSizeInBits()) * 1410 Offset); 1411 widenScalarDst(MI, WideTy.getScalarType(), 0); 1412 Observer.changedInstr(MI); 1413 return Legalized; 1414 } 1415 1416 LegalizerHelper::LegalizeResult 1417 LegalizerHelper::widenScalarInsert(MachineInstr &MI, unsigned TypeIdx, 1418 LLT WideTy) { 1419 if (TypeIdx != 0) 1420 return UnableToLegalize; 1421 Observer.changingInstr(MI); 1422 widenScalarSrc(MI, WideTy, 1, TargetOpcode::G_ANYEXT); 1423 widenScalarDst(MI, WideTy); 1424 Observer.changedInstr(MI); 1425 return Legalized; 1426 } 1427 1428 LegalizerHelper::LegalizeResult 1429 LegalizerHelper::widenScalar(MachineInstr &MI, unsigned TypeIdx, LLT WideTy) { 1430 MIRBuilder.setInstr(MI); 1431 1432 switch (MI.getOpcode()) { 1433 default: 1434 return UnableToLegalize; 1435 case TargetOpcode::G_EXTRACT: 1436 return widenScalarExtract(MI, TypeIdx, WideTy); 1437 case TargetOpcode::G_INSERT: 1438 return widenScalarInsert(MI, TypeIdx, WideTy); 1439 case TargetOpcode::G_MERGE_VALUES: 1440 return widenScalarMergeValues(MI, TypeIdx, WideTy); 1441 case TargetOpcode::G_UNMERGE_VALUES: 1442 return widenScalarUnmergeValues(MI, TypeIdx, WideTy); 1443 case TargetOpcode::G_UADDO: 1444 case TargetOpcode::G_USUBO: { 1445 if (TypeIdx == 1) 1446 return UnableToLegalize; // TODO 1447 auto LHSZext = MIRBuilder.buildInstr(TargetOpcode::G_ZEXT, {WideTy}, 1448 {MI.getOperand(2).getReg()}); 1449 auto RHSZext = MIRBuilder.buildInstr(TargetOpcode::G_ZEXT, {WideTy}, 1450 {MI.getOperand(3).getReg()}); 1451 unsigned Opcode = MI.getOpcode() == TargetOpcode::G_UADDO 1452 ? TargetOpcode::G_ADD 1453 : TargetOpcode::G_SUB; 1454 // Do the arithmetic in the larger type. 1455 auto NewOp = MIRBuilder.buildInstr(Opcode, {WideTy}, {LHSZext, RHSZext}); 1456 LLT OrigTy = MRI.getType(MI.getOperand(0).getReg()); 1457 APInt Mask = APInt::getAllOnesValue(OrigTy.getSizeInBits()); 1458 auto AndOp = MIRBuilder.buildInstr( 1459 TargetOpcode::G_AND, {WideTy}, 1460 {NewOp, MIRBuilder.buildConstant(WideTy, Mask.getZExtValue())}); 1461 // There is no overflow if the AndOp is the same as NewOp. 1462 MIRBuilder.buildICmp(CmpInst::ICMP_NE, MI.getOperand(1).getReg(), NewOp, 1463 AndOp); 1464 // Now trunc the NewOp to the original result. 1465 MIRBuilder.buildTrunc(MI.getOperand(0).getReg(), NewOp); 1466 MI.eraseFromParent(); 1467 return Legalized; 1468 } 1469 case TargetOpcode::G_CTTZ: 1470 case TargetOpcode::G_CTTZ_ZERO_UNDEF: 1471 case TargetOpcode::G_CTLZ: 1472 case TargetOpcode::G_CTLZ_ZERO_UNDEF: 1473 case TargetOpcode::G_CTPOP: { 1474 if (TypeIdx == 0) { 1475 Observer.changingInstr(MI); 1476 widenScalarDst(MI, WideTy, 0); 1477 Observer.changedInstr(MI); 1478 return Legalized; 1479 } 1480 1481 Register SrcReg = MI.getOperand(1).getReg(); 1482 1483 // First ZEXT the input. 1484 auto MIBSrc = MIRBuilder.buildZExt(WideTy, SrcReg); 1485 LLT CurTy = MRI.getType(SrcReg); 1486 if (MI.getOpcode() == TargetOpcode::G_CTTZ) { 1487 // The count is the same in the larger type except if the original 1488 // value was zero. This can be handled by setting the bit just off 1489 // the top of the original type. 1490 auto TopBit = 1491 APInt::getOneBitSet(WideTy.getSizeInBits(), CurTy.getSizeInBits()); 1492 MIBSrc = MIRBuilder.buildOr( 1493 WideTy, MIBSrc, MIRBuilder.buildConstant(WideTy, TopBit)); 1494 } 1495 1496 // Perform the operation at the larger size. 1497 auto MIBNewOp = MIRBuilder.buildInstr(MI.getOpcode(), {WideTy}, {MIBSrc}); 1498 // This is already the correct result for CTPOP and CTTZs 1499 if (MI.getOpcode() == TargetOpcode::G_CTLZ || 1500 MI.getOpcode() == TargetOpcode::G_CTLZ_ZERO_UNDEF) { 1501 // The correct result is NewOp - (Difference in widety and current ty). 1502 unsigned SizeDiff = WideTy.getSizeInBits() - CurTy.getSizeInBits(); 1503 MIBNewOp = MIRBuilder.buildInstr( 1504 TargetOpcode::G_SUB, {WideTy}, 1505 {MIBNewOp, MIRBuilder.buildConstant(WideTy, SizeDiff)}); 1506 } 1507 1508 MIRBuilder.buildZExtOrTrunc(MI.getOperand(0), MIBNewOp); 1509 MI.eraseFromParent(); 1510 return Legalized; 1511 } 1512 case TargetOpcode::G_BSWAP: { 1513 Observer.changingInstr(MI); 1514 Register DstReg = MI.getOperand(0).getReg(); 1515 1516 Register ShrReg = MRI.createGenericVirtualRegister(WideTy); 1517 Register DstExt = MRI.createGenericVirtualRegister(WideTy); 1518 Register ShiftAmtReg = MRI.createGenericVirtualRegister(WideTy); 1519 widenScalarSrc(MI, WideTy, 1, TargetOpcode::G_ANYEXT); 1520 1521 MI.getOperand(0).setReg(DstExt); 1522 1523 MIRBuilder.setInsertPt(MIRBuilder.getMBB(), ++MIRBuilder.getInsertPt()); 1524 1525 LLT Ty = MRI.getType(DstReg); 1526 unsigned DiffBits = WideTy.getScalarSizeInBits() - Ty.getScalarSizeInBits(); 1527 MIRBuilder.buildConstant(ShiftAmtReg, DiffBits); 1528 MIRBuilder.buildInstr(TargetOpcode::G_LSHR) 1529 .addDef(ShrReg) 1530 .addUse(DstExt) 1531 .addUse(ShiftAmtReg); 1532 1533 MIRBuilder.buildTrunc(DstReg, ShrReg); 1534 Observer.changedInstr(MI); 1535 return Legalized; 1536 } 1537 case TargetOpcode::G_BITREVERSE: { 1538 Observer.changingInstr(MI); 1539 1540 Register DstReg = MI.getOperand(0).getReg(); 1541 LLT Ty = MRI.getType(DstReg); 1542 unsigned DiffBits = WideTy.getScalarSizeInBits() - Ty.getScalarSizeInBits(); 1543 1544 Register DstExt = MRI.createGenericVirtualRegister(WideTy); 1545 widenScalarSrc(MI, WideTy, 1, TargetOpcode::G_ANYEXT); 1546 MI.getOperand(0).setReg(DstExt); 1547 MIRBuilder.setInsertPt(MIRBuilder.getMBB(), ++MIRBuilder.getInsertPt()); 1548 1549 auto ShiftAmt = MIRBuilder.buildConstant(WideTy, DiffBits); 1550 auto Shift = MIRBuilder.buildLShr(WideTy, DstExt, ShiftAmt); 1551 MIRBuilder.buildTrunc(DstReg, Shift); 1552 Observer.changedInstr(MI); 1553 return Legalized; 1554 } 1555 case TargetOpcode::G_ADD: 1556 case TargetOpcode::G_AND: 1557 case TargetOpcode::G_MUL: 1558 case TargetOpcode::G_OR: 1559 case TargetOpcode::G_XOR: 1560 case TargetOpcode::G_SUB: 1561 // Perform operation at larger width (any extension is fines here, high bits 1562 // don't affect the result) and then truncate the result back to the 1563 // original type. 1564 Observer.changingInstr(MI); 1565 widenScalarSrc(MI, WideTy, 1, TargetOpcode::G_ANYEXT); 1566 widenScalarSrc(MI, WideTy, 2, TargetOpcode::G_ANYEXT); 1567 widenScalarDst(MI, WideTy); 1568 Observer.changedInstr(MI); 1569 return Legalized; 1570 1571 case TargetOpcode::G_SHL: 1572 Observer.changingInstr(MI); 1573 1574 if (TypeIdx == 0) { 1575 widenScalarSrc(MI, WideTy, 1, TargetOpcode::G_ANYEXT); 1576 widenScalarDst(MI, WideTy); 1577 } else { 1578 assert(TypeIdx == 1); 1579 // The "number of bits to shift" operand must preserve its value as an 1580 // unsigned integer: 1581 widenScalarSrc(MI, WideTy, 2, TargetOpcode::G_ZEXT); 1582 } 1583 1584 Observer.changedInstr(MI); 1585 return Legalized; 1586 1587 case TargetOpcode::G_SDIV: 1588 case TargetOpcode::G_SREM: 1589 case TargetOpcode::G_SMIN: 1590 case TargetOpcode::G_SMAX: 1591 Observer.changingInstr(MI); 1592 widenScalarSrc(MI, WideTy, 1, TargetOpcode::G_SEXT); 1593 widenScalarSrc(MI, WideTy, 2, TargetOpcode::G_SEXT); 1594 widenScalarDst(MI, WideTy); 1595 Observer.changedInstr(MI); 1596 return Legalized; 1597 1598 case TargetOpcode::G_ASHR: 1599 case TargetOpcode::G_LSHR: 1600 Observer.changingInstr(MI); 1601 1602 if (TypeIdx == 0) { 1603 unsigned CvtOp = MI.getOpcode() == TargetOpcode::G_ASHR ? 1604 TargetOpcode::G_SEXT : TargetOpcode::G_ZEXT; 1605 1606 widenScalarSrc(MI, WideTy, 1, CvtOp); 1607 widenScalarDst(MI, WideTy); 1608 } else { 1609 assert(TypeIdx == 1); 1610 // The "number of bits to shift" operand must preserve its value as an 1611 // unsigned integer: 1612 widenScalarSrc(MI, WideTy, 2, TargetOpcode::G_ZEXT); 1613 } 1614 1615 Observer.changedInstr(MI); 1616 return Legalized; 1617 case TargetOpcode::G_UDIV: 1618 case TargetOpcode::G_UREM: 1619 case TargetOpcode::G_UMIN: 1620 case TargetOpcode::G_UMAX: 1621 Observer.changingInstr(MI); 1622 widenScalarSrc(MI, WideTy, 1, TargetOpcode::G_ZEXT); 1623 widenScalarSrc(MI, WideTy, 2, TargetOpcode::G_ZEXT); 1624 widenScalarDst(MI, WideTy); 1625 Observer.changedInstr(MI); 1626 return Legalized; 1627 1628 case TargetOpcode::G_SELECT: 1629 Observer.changingInstr(MI); 1630 if (TypeIdx == 0) { 1631 // Perform operation at larger width (any extension is fine here, high 1632 // bits don't affect the result) and then truncate the result back to the 1633 // original type. 1634 widenScalarSrc(MI, WideTy, 2, TargetOpcode::G_ANYEXT); 1635 widenScalarSrc(MI, WideTy, 3, TargetOpcode::G_ANYEXT); 1636 widenScalarDst(MI, WideTy); 1637 } else { 1638 bool IsVec = MRI.getType(MI.getOperand(1).getReg()).isVector(); 1639 // Explicit extension is required here since high bits affect the result. 1640 widenScalarSrc(MI, WideTy, 1, MIRBuilder.getBoolExtOp(IsVec, false)); 1641 } 1642 Observer.changedInstr(MI); 1643 return Legalized; 1644 1645 case TargetOpcode::G_FPTOSI: 1646 case TargetOpcode::G_FPTOUI: 1647 Observer.changingInstr(MI); 1648 1649 if (TypeIdx == 0) 1650 widenScalarDst(MI, WideTy); 1651 else 1652 widenScalarSrc(MI, WideTy, 1, TargetOpcode::G_FPEXT); 1653 1654 Observer.changedInstr(MI); 1655 return Legalized; 1656 case TargetOpcode::G_SITOFP: 1657 if (TypeIdx != 1) 1658 return UnableToLegalize; 1659 Observer.changingInstr(MI); 1660 widenScalarSrc(MI, WideTy, 1, TargetOpcode::G_SEXT); 1661 Observer.changedInstr(MI); 1662 return Legalized; 1663 1664 case TargetOpcode::G_UITOFP: 1665 if (TypeIdx != 1) 1666 return UnableToLegalize; 1667 Observer.changingInstr(MI); 1668 widenScalarSrc(MI, WideTy, 1, TargetOpcode::G_ZEXT); 1669 Observer.changedInstr(MI); 1670 return Legalized; 1671 1672 case TargetOpcode::G_LOAD: 1673 case TargetOpcode::G_SEXTLOAD: 1674 case TargetOpcode::G_ZEXTLOAD: 1675 Observer.changingInstr(MI); 1676 widenScalarDst(MI, WideTy); 1677 Observer.changedInstr(MI); 1678 return Legalized; 1679 1680 case TargetOpcode::G_STORE: { 1681 if (TypeIdx != 0) 1682 return UnableToLegalize; 1683 1684 LLT Ty = MRI.getType(MI.getOperand(0).getReg()); 1685 if (!isPowerOf2_32(Ty.getSizeInBits())) 1686 return UnableToLegalize; 1687 1688 Observer.changingInstr(MI); 1689 1690 unsigned ExtType = Ty.getScalarSizeInBits() == 1 ? 1691 TargetOpcode::G_ZEXT : TargetOpcode::G_ANYEXT; 1692 widenScalarSrc(MI, WideTy, 0, ExtType); 1693 1694 Observer.changedInstr(MI); 1695 return Legalized; 1696 } 1697 case TargetOpcode::G_CONSTANT: { 1698 MachineOperand &SrcMO = MI.getOperand(1); 1699 LLVMContext &Ctx = MIRBuilder.getMF().getFunction().getContext(); 1700 unsigned ExtOpc = LI.getExtOpcodeForWideningConstant( 1701 MRI.getType(MI.getOperand(0).getReg())); 1702 assert((ExtOpc == TargetOpcode::G_ZEXT || ExtOpc == TargetOpcode::G_SEXT || 1703 ExtOpc == TargetOpcode::G_ANYEXT) && 1704 "Illegal Extend"); 1705 const APInt &SrcVal = SrcMO.getCImm()->getValue(); 1706 const APInt &Val = (ExtOpc == TargetOpcode::G_SEXT) 1707 ? SrcVal.sext(WideTy.getSizeInBits()) 1708 : SrcVal.zext(WideTy.getSizeInBits()); 1709 Observer.changingInstr(MI); 1710 SrcMO.setCImm(ConstantInt::get(Ctx, Val)); 1711 1712 widenScalarDst(MI, WideTy); 1713 Observer.changedInstr(MI); 1714 return Legalized; 1715 } 1716 case TargetOpcode::G_FCONSTANT: { 1717 MachineOperand &SrcMO = MI.getOperand(1); 1718 LLVMContext &Ctx = MIRBuilder.getMF().getFunction().getContext(); 1719 APFloat Val = SrcMO.getFPImm()->getValueAPF(); 1720 bool LosesInfo; 1721 switch (WideTy.getSizeInBits()) { 1722 case 32: 1723 Val.convert(APFloat::IEEEsingle(), APFloat::rmNearestTiesToEven, 1724 &LosesInfo); 1725 break; 1726 case 64: 1727 Val.convert(APFloat::IEEEdouble(), APFloat::rmNearestTiesToEven, 1728 &LosesInfo); 1729 break; 1730 default: 1731 return UnableToLegalize; 1732 } 1733 1734 assert(!LosesInfo && "extend should always be lossless"); 1735 1736 Observer.changingInstr(MI); 1737 SrcMO.setFPImm(ConstantFP::get(Ctx, Val)); 1738 1739 widenScalarDst(MI, WideTy, 0, TargetOpcode::G_FPTRUNC); 1740 Observer.changedInstr(MI); 1741 return Legalized; 1742 } 1743 case TargetOpcode::G_IMPLICIT_DEF: { 1744 Observer.changingInstr(MI); 1745 widenScalarDst(MI, WideTy); 1746 Observer.changedInstr(MI); 1747 return Legalized; 1748 } 1749 case TargetOpcode::G_BRCOND: 1750 Observer.changingInstr(MI); 1751 widenScalarSrc(MI, WideTy, 0, MIRBuilder.getBoolExtOp(false, false)); 1752 Observer.changedInstr(MI); 1753 return Legalized; 1754 1755 case TargetOpcode::G_FCMP: 1756 Observer.changingInstr(MI); 1757 if (TypeIdx == 0) 1758 widenScalarDst(MI, WideTy); 1759 else { 1760 widenScalarSrc(MI, WideTy, 2, TargetOpcode::G_FPEXT); 1761 widenScalarSrc(MI, WideTy, 3, TargetOpcode::G_FPEXT); 1762 } 1763 Observer.changedInstr(MI); 1764 return Legalized; 1765 1766 case TargetOpcode::G_ICMP: 1767 Observer.changingInstr(MI); 1768 if (TypeIdx == 0) 1769 widenScalarDst(MI, WideTy); 1770 else { 1771 unsigned ExtOpcode = CmpInst::isSigned(static_cast<CmpInst::Predicate>( 1772 MI.getOperand(1).getPredicate())) 1773 ? TargetOpcode::G_SEXT 1774 : TargetOpcode::G_ZEXT; 1775 widenScalarSrc(MI, WideTy, 2, ExtOpcode); 1776 widenScalarSrc(MI, WideTy, 3, ExtOpcode); 1777 } 1778 Observer.changedInstr(MI); 1779 return Legalized; 1780 1781 case TargetOpcode::G_PTR_ADD: 1782 assert(TypeIdx == 1 && "unable to legalize pointer of G_PTR_ADD"); 1783 Observer.changingInstr(MI); 1784 widenScalarSrc(MI, WideTy, 2, TargetOpcode::G_SEXT); 1785 Observer.changedInstr(MI); 1786 return Legalized; 1787 1788 case TargetOpcode::G_PHI: { 1789 assert(TypeIdx == 0 && "Expecting only Idx 0"); 1790 1791 Observer.changingInstr(MI); 1792 for (unsigned I = 1; I < MI.getNumOperands(); I += 2) { 1793 MachineBasicBlock &OpMBB = *MI.getOperand(I + 1).getMBB(); 1794 MIRBuilder.setInsertPt(OpMBB, OpMBB.getFirstTerminator()); 1795 widenScalarSrc(MI, WideTy, I, TargetOpcode::G_ANYEXT); 1796 } 1797 1798 MachineBasicBlock &MBB = *MI.getParent(); 1799 MIRBuilder.setInsertPt(MBB, --MBB.getFirstNonPHI()); 1800 widenScalarDst(MI, WideTy); 1801 Observer.changedInstr(MI); 1802 return Legalized; 1803 } 1804 case TargetOpcode::G_EXTRACT_VECTOR_ELT: { 1805 if (TypeIdx == 0) { 1806 Register VecReg = MI.getOperand(1).getReg(); 1807 LLT VecTy = MRI.getType(VecReg); 1808 Observer.changingInstr(MI); 1809 1810 widenScalarSrc(MI, LLT::vector(VecTy.getNumElements(), 1811 WideTy.getSizeInBits()), 1812 1, TargetOpcode::G_SEXT); 1813 1814 widenScalarDst(MI, WideTy, 0); 1815 Observer.changedInstr(MI); 1816 return Legalized; 1817 } 1818 1819 if (TypeIdx != 2) 1820 return UnableToLegalize; 1821 Observer.changingInstr(MI); 1822 // TODO: Probably should be zext 1823 widenScalarSrc(MI, WideTy, 2, TargetOpcode::G_SEXT); 1824 Observer.changedInstr(MI); 1825 return Legalized; 1826 } 1827 case TargetOpcode::G_INSERT_VECTOR_ELT: { 1828 if (TypeIdx == 1) { 1829 Observer.changingInstr(MI); 1830 1831 Register VecReg = MI.getOperand(1).getReg(); 1832 LLT VecTy = MRI.getType(VecReg); 1833 LLT WideVecTy = LLT::vector(VecTy.getNumElements(), WideTy); 1834 1835 widenScalarSrc(MI, WideVecTy, 1, TargetOpcode::G_ANYEXT); 1836 widenScalarSrc(MI, WideTy, 2, TargetOpcode::G_ANYEXT); 1837 widenScalarDst(MI, WideVecTy, 0); 1838 Observer.changedInstr(MI); 1839 return Legalized; 1840 } 1841 1842 if (TypeIdx == 2) { 1843 Observer.changingInstr(MI); 1844 // TODO: Probably should be zext 1845 widenScalarSrc(MI, WideTy, 3, TargetOpcode::G_SEXT); 1846 Observer.changedInstr(MI); 1847 } 1848 1849 return Legalized; 1850 } 1851 case TargetOpcode::G_FADD: 1852 case TargetOpcode::G_FMUL: 1853 case TargetOpcode::G_FSUB: 1854 case TargetOpcode::G_FMA: 1855 case TargetOpcode::G_FMAD: 1856 case TargetOpcode::G_FNEG: 1857 case TargetOpcode::G_FABS: 1858 case TargetOpcode::G_FCANONICALIZE: 1859 case TargetOpcode::G_FMINNUM: 1860 case TargetOpcode::G_FMAXNUM: 1861 case TargetOpcode::G_FMINNUM_IEEE: 1862 case TargetOpcode::G_FMAXNUM_IEEE: 1863 case TargetOpcode::G_FMINIMUM: 1864 case TargetOpcode::G_FMAXIMUM: 1865 case TargetOpcode::G_FDIV: 1866 case TargetOpcode::G_FREM: 1867 case TargetOpcode::G_FCEIL: 1868 case TargetOpcode::G_FFLOOR: 1869 case TargetOpcode::G_FCOS: 1870 case TargetOpcode::G_FSIN: 1871 case TargetOpcode::G_FLOG10: 1872 case TargetOpcode::G_FLOG: 1873 case TargetOpcode::G_FLOG2: 1874 case TargetOpcode::G_FRINT: 1875 case TargetOpcode::G_FNEARBYINT: 1876 case TargetOpcode::G_FSQRT: 1877 case TargetOpcode::G_FEXP: 1878 case TargetOpcode::G_FEXP2: 1879 case TargetOpcode::G_FPOW: 1880 case TargetOpcode::G_INTRINSIC_TRUNC: 1881 case TargetOpcode::G_INTRINSIC_ROUND: 1882 assert(TypeIdx == 0); 1883 Observer.changingInstr(MI); 1884 1885 for (unsigned I = 1, E = MI.getNumOperands(); I != E; ++I) 1886 widenScalarSrc(MI, WideTy, I, TargetOpcode::G_FPEXT); 1887 1888 widenScalarDst(MI, WideTy, 0, TargetOpcode::G_FPTRUNC); 1889 Observer.changedInstr(MI); 1890 return Legalized; 1891 case TargetOpcode::G_INTTOPTR: 1892 if (TypeIdx != 1) 1893 return UnableToLegalize; 1894 1895 Observer.changingInstr(MI); 1896 widenScalarSrc(MI, WideTy, 1, TargetOpcode::G_ZEXT); 1897 Observer.changedInstr(MI); 1898 return Legalized; 1899 case TargetOpcode::G_PTRTOINT: 1900 if (TypeIdx != 0) 1901 return UnableToLegalize; 1902 1903 Observer.changingInstr(MI); 1904 widenScalarDst(MI, WideTy, 0); 1905 Observer.changedInstr(MI); 1906 return Legalized; 1907 case TargetOpcode::G_BUILD_VECTOR: { 1908 Observer.changingInstr(MI); 1909 1910 const LLT WideEltTy = TypeIdx == 1 ? WideTy : WideTy.getElementType(); 1911 for (int I = 1, E = MI.getNumOperands(); I != E; ++I) 1912 widenScalarSrc(MI, WideEltTy, I, TargetOpcode::G_ANYEXT); 1913 1914 // Avoid changing the result vector type if the source element type was 1915 // requested. 1916 if (TypeIdx == 1) { 1917 auto &TII = *MI.getMF()->getSubtarget().getInstrInfo(); 1918 MI.setDesc(TII.get(TargetOpcode::G_BUILD_VECTOR_TRUNC)); 1919 } else { 1920 widenScalarDst(MI, WideTy, 0); 1921 } 1922 1923 Observer.changedInstr(MI); 1924 return Legalized; 1925 } 1926 case TargetOpcode::G_SEXT_INREG: 1927 if (TypeIdx != 0) 1928 return UnableToLegalize; 1929 1930 Observer.changingInstr(MI); 1931 widenScalarSrc(MI, WideTy, 1, TargetOpcode::G_ANYEXT); 1932 widenScalarDst(MI, WideTy, 0, TargetOpcode::G_TRUNC); 1933 Observer.changedInstr(MI); 1934 return Legalized; 1935 } 1936 } 1937 1938 LegalizerHelper::LegalizeResult 1939 LegalizerHelper::lower(MachineInstr &MI, unsigned TypeIdx, LLT Ty) { 1940 using namespace TargetOpcode; 1941 MIRBuilder.setInstr(MI); 1942 1943 switch(MI.getOpcode()) { 1944 default: 1945 return UnableToLegalize; 1946 case TargetOpcode::G_SREM: 1947 case TargetOpcode::G_UREM: { 1948 Register QuotReg = MRI.createGenericVirtualRegister(Ty); 1949 MIRBuilder.buildInstr(MI.getOpcode() == G_SREM ? G_SDIV : G_UDIV) 1950 .addDef(QuotReg) 1951 .addUse(MI.getOperand(1).getReg()) 1952 .addUse(MI.getOperand(2).getReg()); 1953 1954 Register ProdReg = MRI.createGenericVirtualRegister(Ty); 1955 MIRBuilder.buildMul(ProdReg, QuotReg, MI.getOperand(2).getReg()); 1956 MIRBuilder.buildSub(MI.getOperand(0).getReg(), MI.getOperand(1).getReg(), 1957 ProdReg); 1958 MI.eraseFromParent(); 1959 return Legalized; 1960 } 1961 case TargetOpcode::G_SADDO: 1962 case TargetOpcode::G_SSUBO: 1963 return lowerSADDO_SSUBO(MI); 1964 case TargetOpcode::G_SMULO: 1965 case TargetOpcode::G_UMULO: { 1966 // Generate G_UMULH/G_SMULH to check for overflow and a normal G_MUL for the 1967 // result. 1968 Register Res = MI.getOperand(0).getReg(); 1969 Register Overflow = MI.getOperand(1).getReg(); 1970 Register LHS = MI.getOperand(2).getReg(); 1971 Register RHS = MI.getOperand(3).getReg(); 1972 1973 MIRBuilder.buildMul(Res, LHS, RHS); 1974 1975 unsigned Opcode = MI.getOpcode() == TargetOpcode::G_SMULO 1976 ? TargetOpcode::G_SMULH 1977 : TargetOpcode::G_UMULH; 1978 1979 Register HiPart = MRI.createGenericVirtualRegister(Ty); 1980 MIRBuilder.buildInstr(Opcode) 1981 .addDef(HiPart) 1982 .addUse(LHS) 1983 .addUse(RHS); 1984 1985 Register Zero = MRI.createGenericVirtualRegister(Ty); 1986 MIRBuilder.buildConstant(Zero, 0); 1987 1988 // For *signed* multiply, overflow is detected by checking: 1989 // (hi != (lo >> bitwidth-1)) 1990 if (Opcode == TargetOpcode::G_SMULH) { 1991 Register Shifted = MRI.createGenericVirtualRegister(Ty); 1992 Register ShiftAmt = MRI.createGenericVirtualRegister(Ty); 1993 MIRBuilder.buildConstant(ShiftAmt, Ty.getSizeInBits() - 1); 1994 MIRBuilder.buildInstr(TargetOpcode::G_ASHR) 1995 .addDef(Shifted) 1996 .addUse(Res) 1997 .addUse(ShiftAmt); 1998 MIRBuilder.buildICmp(CmpInst::ICMP_NE, Overflow, HiPart, Shifted); 1999 } else { 2000 MIRBuilder.buildICmp(CmpInst::ICMP_NE, Overflow, HiPart, Zero); 2001 } 2002 MI.eraseFromParent(); 2003 return Legalized; 2004 } 2005 case TargetOpcode::G_FNEG: { 2006 // TODO: Handle vector types once we are able to 2007 // represent them. 2008 if (Ty.isVector()) 2009 return UnableToLegalize; 2010 Register Res = MI.getOperand(0).getReg(); 2011 Type *ZeroTy; 2012 LLVMContext &Ctx = MIRBuilder.getMF().getFunction().getContext(); 2013 switch (Ty.getSizeInBits()) { 2014 case 16: 2015 ZeroTy = Type::getHalfTy(Ctx); 2016 break; 2017 case 32: 2018 ZeroTy = Type::getFloatTy(Ctx); 2019 break; 2020 case 64: 2021 ZeroTy = Type::getDoubleTy(Ctx); 2022 break; 2023 case 128: 2024 ZeroTy = Type::getFP128Ty(Ctx); 2025 break; 2026 default: 2027 llvm_unreachable("unexpected floating-point type"); 2028 } 2029 ConstantFP &ZeroForNegation = 2030 *cast<ConstantFP>(ConstantFP::getZeroValueForNegation(ZeroTy)); 2031 auto Zero = MIRBuilder.buildFConstant(Ty, ZeroForNegation); 2032 Register SubByReg = MI.getOperand(1).getReg(); 2033 Register ZeroReg = Zero->getOperand(0).getReg(); 2034 MIRBuilder.buildInstr(TargetOpcode::G_FSUB, {Res}, {ZeroReg, SubByReg}, 2035 MI.getFlags()); 2036 MI.eraseFromParent(); 2037 return Legalized; 2038 } 2039 case TargetOpcode::G_FSUB: { 2040 // Lower (G_FSUB LHS, RHS) to (G_FADD LHS, (G_FNEG RHS)). 2041 // First, check if G_FNEG is marked as Lower. If so, we may 2042 // end up with an infinite loop as G_FSUB is used to legalize G_FNEG. 2043 if (LI.getAction({G_FNEG, {Ty}}).Action == Lower) 2044 return UnableToLegalize; 2045 Register Res = MI.getOperand(0).getReg(); 2046 Register LHS = MI.getOperand(1).getReg(); 2047 Register RHS = MI.getOperand(2).getReg(); 2048 Register Neg = MRI.createGenericVirtualRegister(Ty); 2049 MIRBuilder.buildInstr(TargetOpcode::G_FNEG).addDef(Neg).addUse(RHS); 2050 MIRBuilder.buildInstr(TargetOpcode::G_FADD, {Res}, {LHS, Neg}, MI.getFlags()); 2051 MI.eraseFromParent(); 2052 return Legalized; 2053 } 2054 case TargetOpcode::G_FMAD: 2055 return lowerFMad(MI); 2056 case TargetOpcode::G_INTRINSIC_ROUND: 2057 return lowerIntrinsicRound(MI); 2058 case TargetOpcode::G_ATOMIC_CMPXCHG_WITH_SUCCESS: { 2059 Register OldValRes = MI.getOperand(0).getReg(); 2060 Register SuccessRes = MI.getOperand(1).getReg(); 2061 Register Addr = MI.getOperand(2).getReg(); 2062 Register CmpVal = MI.getOperand(3).getReg(); 2063 Register NewVal = MI.getOperand(4).getReg(); 2064 MIRBuilder.buildAtomicCmpXchg(OldValRes, Addr, CmpVal, NewVal, 2065 **MI.memoperands_begin()); 2066 MIRBuilder.buildICmp(CmpInst::ICMP_EQ, SuccessRes, OldValRes, CmpVal); 2067 MI.eraseFromParent(); 2068 return Legalized; 2069 } 2070 case TargetOpcode::G_LOAD: 2071 case TargetOpcode::G_SEXTLOAD: 2072 case TargetOpcode::G_ZEXTLOAD: { 2073 // Lower to a memory-width G_LOAD and a G_SEXT/G_ZEXT/G_ANYEXT 2074 Register DstReg = MI.getOperand(0).getReg(); 2075 Register PtrReg = MI.getOperand(1).getReg(); 2076 LLT DstTy = MRI.getType(DstReg); 2077 auto &MMO = **MI.memoperands_begin(); 2078 2079 if (DstTy.getSizeInBits() == MMO.getSizeInBits()) { 2080 if (MI.getOpcode() == TargetOpcode::G_LOAD) { 2081 // This load needs splitting into power of 2 sized loads. 2082 if (DstTy.isVector()) 2083 return UnableToLegalize; 2084 if (isPowerOf2_32(DstTy.getSizeInBits())) 2085 return UnableToLegalize; // Don't know what we're being asked to do. 2086 2087 // Our strategy here is to generate anyextending loads for the smaller 2088 // types up to next power-2 result type, and then combine the two larger 2089 // result values together, before truncating back down to the non-pow-2 2090 // type. 2091 // E.g. v1 = i24 load => 2092 // v2 = i32 load (2 byte) 2093 // v3 = i32 load (1 byte) 2094 // v4 = i32 shl v3, 16 2095 // v5 = i32 or v4, v2 2096 // v1 = i24 trunc v5 2097 // By doing this we generate the correct truncate which should get 2098 // combined away as an artifact with a matching extend. 2099 uint64_t LargeSplitSize = PowerOf2Floor(DstTy.getSizeInBits()); 2100 uint64_t SmallSplitSize = DstTy.getSizeInBits() - LargeSplitSize; 2101 2102 MachineFunction &MF = MIRBuilder.getMF(); 2103 MachineMemOperand *LargeMMO = 2104 MF.getMachineMemOperand(&MMO, 0, LargeSplitSize / 8); 2105 MachineMemOperand *SmallMMO = MF.getMachineMemOperand( 2106 &MMO, LargeSplitSize / 8, SmallSplitSize / 8); 2107 2108 LLT PtrTy = MRI.getType(PtrReg); 2109 unsigned AnyExtSize = NextPowerOf2(DstTy.getSizeInBits()); 2110 LLT AnyExtTy = LLT::scalar(AnyExtSize); 2111 Register LargeLdReg = MRI.createGenericVirtualRegister(AnyExtTy); 2112 Register SmallLdReg = MRI.createGenericVirtualRegister(AnyExtTy); 2113 auto LargeLoad = 2114 MIRBuilder.buildLoad(LargeLdReg, PtrReg, *LargeMMO); 2115 2116 auto OffsetCst = 2117 MIRBuilder.buildConstant(LLT::scalar(64), LargeSplitSize / 8); 2118 Register PtrAddReg = MRI.createGenericVirtualRegister(PtrTy); 2119 auto SmallPtr = 2120 MIRBuilder.buildPtrAdd(PtrAddReg, PtrReg, OffsetCst.getReg(0)); 2121 auto SmallLoad = MIRBuilder.buildLoad(SmallLdReg, SmallPtr.getReg(0), 2122 *SmallMMO); 2123 2124 auto ShiftAmt = MIRBuilder.buildConstant(AnyExtTy, LargeSplitSize); 2125 auto Shift = MIRBuilder.buildShl(AnyExtTy, SmallLoad, ShiftAmt); 2126 auto Or = MIRBuilder.buildOr(AnyExtTy, Shift, LargeLoad); 2127 MIRBuilder.buildTrunc(DstReg, {Or.getReg(0)}); 2128 MI.eraseFromParent(); 2129 return Legalized; 2130 } 2131 MIRBuilder.buildLoad(DstReg, PtrReg, MMO); 2132 MI.eraseFromParent(); 2133 return Legalized; 2134 } 2135 2136 if (DstTy.isScalar()) { 2137 Register TmpReg = 2138 MRI.createGenericVirtualRegister(LLT::scalar(MMO.getSizeInBits())); 2139 MIRBuilder.buildLoad(TmpReg, PtrReg, MMO); 2140 switch (MI.getOpcode()) { 2141 default: 2142 llvm_unreachable("Unexpected opcode"); 2143 case TargetOpcode::G_LOAD: 2144 MIRBuilder.buildExtOrTrunc(TargetOpcode::G_ANYEXT, DstReg, TmpReg); 2145 break; 2146 case TargetOpcode::G_SEXTLOAD: 2147 MIRBuilder.buildSExt(DstReg, TmpReg); 2148 break; 2149 case TargetOpcode::G_ZEXTLOAD: 2150 MIRBuilder.buildZExt(DstReg, TmpReg); 2151 break; 2152 } 2153 MI.eraseFromParent(); 2154 return Legalized; 2155 } 2156 2157 return UnableToLegalize; 2158 } 2159 case TargetOpcode::G_STORE: { 2160 // Lower a non-power of 2 store into multiple pow-2 stores. 2161 // E.g. split an i24 store into an i16 store + i8 store. 2162 // We do this by first extending the stored value to the next largest power 2163 // of 2 type, and then using truncating stores to store the components. 2164 // By doing this, likewise with G_LOAD, generate an extend that can be 2165 // artifact-combined away instead of leaving behind extracts. 2166 Register SrcReg = MI.getOperand(0).getReg(); 2167 Register PtrReg = MI.getOperand(1).getReg(); 2168 LLT SrcTy = MRI.getType(SrcReg); 2169 MachineMemOperand &MMO = **MI.memoperands_begin(); 2170 if (SrcTy.getSizeInBits() != MMO.getSizeInBits()) 2171 return UnableToLegalize; 2172 if (SrcTy.isVector()) 2173 return UnableToLegalize; 2174 if (isPowerOf2_32(SrcTy.getSizeInBits())) 2175 return UnableToLegalize; // Don't know what we're being asked to do. 2176 2177 // Extend to the next pow-2. 2178 const LLT ExtendTy = LLT::scalar(NextPowerOf2(SrcTy.getSizeInBits())); 2179 auto ExtVal = MIRBuilder.buildAnyExt(ExtendTy, SrcReg); 2180 2181 // Obtain the smaller value by shifting away the larger value. 2182 uint64_t LargeSplitSize = PowerOf2Floor(SrcTy.getSizeInBits()); 2183 uint64_t SmallSplitSize = SrcTy.getSizeInBits() - LargeSplitSize; 2184 auto ShiftAmt = MIRBuilder.buildConstant(ExtendTy, LargeSplitSize); 2185 auto SmallVal = MIRBuilder.buildLShr(ExtendTy, ExtVal, ShiftAmt); 2186 2187 // Generate the PtrAdd and truncating stores. 2188 LLT PtrTy = MRI.getType(PtrReg); 2189 auto OffsetCst = 2190 MIRBuilder.buildConstant(LLT::scalar(64), LargeSplitSize / 8); 2191 Register PtrAddReg = MRI.createGenericVirtualRegister(PtrTy); 2192 auto SmallPtr = 2193 MIRBuilder.buildPtrAdd(PtrAddReg, PtrReg, OffsetCst.getReg(0)); 2194 2195 MachineFunction &MF = MIRBuilder.getMF(); 2196 MachineMemOperand *LargeMMO = 2197 MF.getMachineMemOperand(&MMO, 0, LargeSplitSize / 8); 2198 MachineMemOperand *SmallMMO = 2199 MF.getMachineMemOperand(&MMO, LargeSplitSize / 8, SmallSplitSize / 8); 2200 MIRBuilder.buildStore(ExtVal.getReg(0), PtrReg, *LargeMMO); 2201 MIRBuilder.buildStore(SmallVal.getReg(0), SmallPtr.getReg(0), *SmallMMO); 2202 MI.eraseFromParent(); 2203 return Legalized; 2204 } 2205 case TargetOpcode::G_CTLZ_ZERO_UNDEF: 2206 case TargetOpcode::G_CTTZ_ZERO_UNDEF: 2207 case TargetOpcode::G_CTLZ: 2208 case TargetOpcode::G_CTTZ: 2209 case TargetOpcode::G_CTPOP: 2210 return lowerBitCount(MI, TypeIdx, Ty); 2211 case G_UADDO: { 2212 Register Res = MI.getOperand(0).getReg(); 2213 Register CarryOut = MI.getOperand(1).getReg(); 2214 Register LHS = MI.getOperand(2).getReg(); 2215 Register RHS = MI.getOperand(3).getReg(); 2216 2217 MIRBuilder.buildAdd(Res, LHS, RHS); 2218 MIRBuilder.buildICmp(CmpInst::ICMP_ULT, CarryOut, Res, RHS); 2219 2220 MI.eraseFromParent(); 2221 return Legalized; 2222 } 2223 case G_UADDE: { 2224 Register Res = MI.getOperand(0).getReg(); 2225 Register CarryOut = MI.getOperand(1).getReg(); 2226 Register LHS = MI.getOperand(2).getReg(); 2227 Register RHS = MI.getOperand(3).getReg(); 2228 Register CarryIn = MI.getOperand(4).getReg(); 2229 2230 Register TmpRes = MRI.createGenericVirtualRegister(Ty); 2231 Register ZExtCarryIn = MRI.createGenericVirtualRegister(Ty); 2232 2233 MIRBuilder.buildAdd(TmpRes, LHS, RHS); 2234 MIRBuilder.buildZExt(ZExtCarryIn, CarryIn); 2235 MIRBuilder.buildAdd(Res, TmpRes, ZExtCarryIn); 2236 MIRBuilder.buildICmp(CmpInst::ICMP_ULT, CarryOut, Res, LHS); 2237 2238 MI.eraseFromParent(); 2239 return Legalized; 2240 } 2241 case G_USUBO: { 2242 Register Res = MI.getOperand(0).getReg(); 2243 Register BorrowOut = MI.getOperand(1).getReg(); 2244 Register LHS = MI.getOperand(2).getReg(); 2245 Register RHS = MI.getOperand(3).getReg(); 2246 2247 MIRBuilder.buildSub(Res, LHS, RHS); 2248 MIRBuilder.buildICmp(CmpInst::ICMP_ULT, BorrowOut, LHS, RHS); 2249 2250 MI.eraseFromParent(); 2251 return Legalized; 2252 } 2253 case G_USUBE: { 2254 Register Res = MI.getOperand(0).getReg(); 2255 Register BorrowOut = MI.getOperand(1).getReg(); 2256 Register LHS = MI.getOperand(2).getReg(); 2257 Register RHS = MI.getOperand(3).getReg(); 2258 Register BorrowIn = MI.getOperand(4).getReg(); 2259 2260 Register TmpRes = MRI.createGenericVirtualRegister(Ty); 2261 Register ZExtBorrowIn = MRI.createGenericVirtualRegister(Ty); 2262 Register LHS_EQ_RHS = MRI.createGenericVirtualRegister(LLT::scalar(1)); 2263 Register LHS_ULT_RHS = MRI.createGenericVirtualRegister(LLT::scalar(1)); 2264 2265 MIRBuilder.buildSub(TmpRes, LHS, RHS); 2266 MIRBuilder.buildZExt(ZExtBorrowIn, BorrowIn); 2267 MIRBuilder.buildSub(Res, TmpRes, ZExtBorrowIn); 2268 MIRBuilder.buildICmp(CmpInst::ICMP_EQ, LHS_EQ_RHS, LHS, RHS); 2269 MIRBuilder.buildICmp(CmpInst::ICMP_ULT, LHS_ULT_RHS, LHS, RHS); 2270 MIRBuilder.buildSelect(BorrowOut, LHS_EQ_RHS, BorrowIn, LHS_ULT_RHS); 2271 2272 MI.eraseFromParent(); 2273 return Legalized; 2274 } 2275 case G_UITOFP: 2276 return lowerUITOFP(MI, TypeIdx, Ty); 2277 case G_SITOFP: 2278 return lowerSITOFP(MI, TypeIdx, Ty); 2279 case G_FPTOUI: 2280 return lowerFPTOUI(MI, TypeIdx, Ty); 2281 case G_SMIN: 2282 case G_SMAX: 2283 case G_UMIN: 2284 case G_UMAX: 2285 return lowerMinMax(MI, TypeIdx, Ty); 2286 case G_FCOPYSIGN: 2287 return lowerFCopySign(MI, TypeIdx, Ty); 2288 case G_FMINNUM: 2289 case G_FMAXNUM: 2290 return lowerFMinNumMaxNum(MI); 2291 case G_UNMERGE_VALUES: 2292 return lowerUnmergeValues(MI); 2293 case TargetOpcode::G_SEXT_INREG: { 2294 assert(MI.getOperand(2).isImm() && "Expected immediate"); 2295 int64_t SizeInBits = MI.getOperand(2).getImm(); 2296 2297 Register DstReg = MI.getOperand(0).getReg(); 2298 Register SrcReg = MI.getOperand(1).getReg(); 2299 LLT DstTy = MRI.getType(DstReg); 2300 Register TmpRes = MRI.createGenericVirtualRegister(DstTy); 2301 2302 auto MIBSz = MIRBuilder.buildConstant(DstTy, DstTy.getScalarSizeInBits() - SizeInBits); 2303 MIRBuilder.buildInstr(TargetOpcode::G_SHL, {TmpRes}, {SrcReg, MIBSz->getOperand(0).getReg()}); 2304 MIRBuilder.buildInstr(TargetOpcode::G_ASHR, {DstReg}, {TmpRes, MIBSz->getOperand(0).getReg()}); 2305 MI.eraseFromParent(); 2306 return Legalized; 2307 } 2308 case G_SHUFFLE_VECTOR: 2309 return lowerShuffleVector(MI); 2310 case G_DYN_STACKALLOC: 2311 return lowerDynStackAlloc(MI); 2312 case G_EXTRACT: 2313 return lowerExtract(MI); 2314 case G_INSERT: 2315 return lowerInsert(MI); 2316 case G_BSWAP: 2317 return lowerBswap(MI); 2318 case G_BITREVERSE: 2319 return lowerBitreverse(MI); 2320 case G_READ_REGISTER: 2321 return lowerReadRegister(MI); 2322 } 2323 } 2324 2325 LegalizerHelper::LegalizeResult LegalizerHelper::fewerElementsVectorImplicitDef( 2326 MachineInstr &MI, unsigned TypeIdx, LLT NarrowTy) { 2327 SmallVector<Register, 2> DstRegs; 2328 2329 unsigned NarrowSize = NarrowTy.getSizeInBits(); 2330 Register DstReg = MI.getOperand(0).getReg(); 2331 unsigned Size = MRI.getType(DstReg).getSizeInBits(); 2332 int NumParts = Size / NarrowSize; 2333 // FIXME: Don't know how to handle the situation where the small vectors 2334 // aren't all the same size yet. 2335 if (Size % NarrowSize != 0) 2336 return UnableToLegalize; 2337 2338 for (int i = 0; i < NumParts; ++i) { 2339 Register TmpReg = MRI.createGenericVirtualRegister(NarrowTy); 2340 MIRBuilder.buildUndef(TmpReg); 2341 DstRegs.push_back(TmpReg); 2342 } 2343 2344 if (NarrowTy.isVector()) 2345 MIRBuilder.buildConcatVectors(DstReg, DstRegs); 2346 else 2347 MIRBuilder.buildBuildVector(DstReg, DstRegs); 2348 2349 MI.eraseFromParent(); 2350 return Legalized; 2351 } 2352 2353 LegalizerHelper::LegalizeResult 2354 LegalizerHelper::fewerElementsVectorBasic(MachineInstr &MI, unsigned TypeIdx, 2355 LLT NarrowTy) { 2356 const unsigned Opc = MI.getOpcode(); 2357 const unsigned NumOps = MI.getNumOperands() - 1; 2358 const unsigned NarrowSize = NarrowTy.getSizeInBits(); 2359 const Register DstReg = MI.getOperand(0).getReg(); 2360 const unsigned Flags = MI.getFlags(); 2361 const LLT DstTy = MRI.getType(DstReg); 2362 const unsigned Size = DstTy.getSizeInBits(); 2363 const int NumParts = Size / NarrowSize; 2364 const LLT EltTy = DstTy.getElementType(); 2365 const unsigned EltSize = EltTy.getSizeInBits(); 2366 const unsigned BitsForNumParts = NarrowSize * NumParts; 2367 2368 // Check if we have any leftovers. If we do, then only handle the case where 2369 // the leftover is one element. 2370 if (BitsForNumParts != Size && BitsForNumParts + EltSize != Size) 2371 return UnableToLegalize; 2372 2373 if (BitsForNumParts != Size) { 2374 Register AccumDstReg = MRI.createGenericVirtualRegister(DstTy); 2375 MIRBuilder.buildUndef(AccumDstReg); 2376 2377 // Handle the pieces which evenly divide into the requested type with 2378 // extract/op/insert sequence. 2379 for (unsigned Offset = 0; Offset < BitsForNumParts; Offset += NarrowSize) { 2380 SmallVector<SrcOp, 4> SrcOps; 2381 for (unsigned I = 1, E = MI.getNumOperands(); I != E; ++I) { 2382 Register PartOpReg = MRI.createGenericVirtualRegister(NarrowTy); 2383 MIRBuilder.buildExtract(PartOpReg, MI.getOperand(I).getReg(), Offset); 2384 SrcOps.push_back(PartOpReg); 2385 } 2386 2387 Register PartDstReg = MRI.createGenericVirtualRegister(NarrowTy); 2388 MIRBuilder.buildInstr(Opc, {PartDstReg}, SrcOps, Flags); 2389 2390 Register PartInsertReg = MRI.createGenericVirtualRegister(DstTy); 2391 MIRBuilder.buildInsert(PartInsertReg, AccumDstReg, PartDstReg, Offset); 2392 AccumDstReg = PartInsertReg; 2393 } 2394 2395 // Handle the remaining element sized leftover piece. 2396 SmallVector<SrcOp, 4> SrcOps; 2397 for (unsigned I = 1, E = MI.getNumOperands(); I != E; ++I) { 2398 Register PartOpReg = MRI.createGenericVirtualRegister(EltTy); 2399 MIRBuilder.buildExtract(PartOpReg, MI.getOperand(I).getReg(), 2400 BitsForNumParts); 2401 SrcOps.push_back(PartOpReg); 2402 } 2403 2404 Register PartDstReg = MRI.createGenericVirtualRegister(EltTy); 2405 MIRBuilder.buildInstr(Opc, {PartDstReg}, SrcOps, Flags); 2406 MIRBuilder.buildInsert(DstReg, AccumDstReg, PartDstReg, BitsForNumParts); 2407 MI.eraseFromParent(); 2408 2409 return Legalized; 2410 } 2411 2412 SmallVector<Register, 2> DstRegs, Src0Regs, Src1Regs, Src2Regs; 2413 2414 extractParts(MI.getOperand(1).getReg(), NarrowTy, NumParts, Src0Regs); 2415 2416 if (NumOps >= 2) 2417 extractParts(MI.getOperand(2).getReg(), NarrowTy, NumParts, Src1Regs); 2418 2419 if (NumOps >= 3) 2420 extractParts(MI.getOperand(3).getReg(), NarrowTy, NumParts, Src2Regs); 2421 2422 for (int i = 0; i < NumParts; ++i) { 2423 Register DstReg = MRI.createGenericVirtualRegister(NarrowTy); 2424 2425 if (NumOps == 1) 2426 MIRBuilder.buildInstr(Opc, {DstReg}, {Src0Regs[i]}, Flags); 2427 else if (NumOps == 2) { 2428 MIRBuilder.buildInstr(Opc, {DstReg}, {Src0Regs[i], Src1Regs[i]}, Flags); 2429 } else if (NumOps == 3) { 2430 MIRBuilder.buildInstr(Opc, {DstReg}, 2431 {Src0Regs[i], Src1Regs[i], Src2Regs[i]}, Flags); 2432 } 2433 2434 DstRegs.push_back(DstReg); 2435 } 2436 2437 if (NarrowTy.isVector()) 2438 MIRBuilder.buildConcatVectors(DstReg, DstRegs); 2439 else 2440 MIRBuilder.buildBuildVector(DstReg, DstRegs); 2441 2442 MI.eraseFromParent(); 2443 return Legalized; 2444 } 2445 2446 // Handle splitting vector operations which need to have the same number of 2447 // elements in each type index, but each type index may have a different element 2448 // type. 2449 // 2450 // e.g. <4 x s64> = G_SHL <4 x s64>, <4 x s32> -> 2451 // <2 x s64> = G_SHL <2 x s64>, <2 x s32> 2452 // <2 x s64> = G_SHL <2 x s64>, <2 x s32> 2453 // 2454 // Also handles some irregular breakdown cases, e.g. 2455 // e.g. <3 x s64> = G_SHL <3 x s64>, <3 x s32> -> 2456 // <2 x s64> = G_SHL <2 x s64>, <2 x s32> 2457 // s64 = G_SHL s64, s32 2458 LegalizerHelper::LegalizeResult 2459 LegalizerHelper::fewerElementsVectorMultiEltType( 2460 MachineInstr &MI, unsigned TypeIdx, LLT NarrowTyArg) { 2461 if (TypeIdx != 0) 2462 return UnableToLegalize; 2463 2464 const LLT NarrowTy0 = NarrowTyArg; 2465 const unsigned NewNumElts = 2466 NarrowTy0.isVector() ? NarrowTy0.getNumElements() : 1; 2467 2468 const Register DstReg = MI.getOperand(0).getReg(); 2469 LLT DstTy = MRI.getType(DstReg); 2470 LLT LeftoverTy0; 2471 2472 // All of the operands need to have the same number of elements, so if we can 2473 // determine a type breakdown for the result type, we can for all of the 2474 // source types. 2475 int NumParts = getNarrowTypeBreakDown(DstTy, NarrowTy0, LeftoverTy0).first; 2476 if (NumParts < 0) 2477 return UnableToLegalize; 2478 2479 SmallVector<MachineInstrBuilder, 4> NewInsts; 2480 2481 SmallVector<Register, 4> DstRegs, LeftoverDstRegs; 2482 SmallVector<Register, 4> PartRegs, LeftoverRegs; 2483 2484 for (unsigned I = 1, E = MI.getNumOperands(); I != E; ++I) { 2485 LLT LeftoverTy; 2486 Register SrcReg = MI.getOperand(I).getReg(); 2487 LLT SrcTyI = MRI.getType(SrcReg); 2488 LLT NarrowTyI = LLT::scalarOrVector(NewNumElts, SrcTyI.getScalarType()); 2489 LLT LeftoverTyI; 2490 2491 // Split this operand into the requested typed registers, and any leftover 2492 // required to reproduce the original type. 2493 if (!extractParts(SrcReg, SrcTyI, NarrowTyI, LeftoverTyI, PartRegs, 2494 LeftoverRegs)) 2495 return UnableToLegalize; 2496 2497 if (I == 1) { 2498 // For the first operand, create an instruction for each part and setup 2499 // the result. 2500 for (Register PartReg : PartRegs) { 2501 Register PartDstReg = MRI.createGenericVirtualRegister(NarrowTy0); 2502 NewInsts.push_back(MIRBuilder.buildInstrNoInsert(MI.getOpcode()) 2503 .addDef(PartDstReg) 2504 .addUse(PartReg)); 2505 DstRegs.push_back(PartDstReg); 2506 } 2507 2508 for (Register LeftoverReg : LeftoverRegs) { 2509 Register PartDstReg = MRI.createGenericVirtualRegister(LeftoverTy0); 2510 NewInsts.push_back(MIRBuilder.buildInstrNoInsert(MI.getOpcode()) 2511 .addDef(PartDstReg) 2512 .addUse(LeftoverReg)); 2513 LeftoverDstRegs.push_back(PartDstReg); 2514 } 2515 } else { 2516 assert(NewInsts.size() == PartRegs.size() + LeftoverRegs.size()); 2517 2518 // Add the newly created operand splits to the existing instructions. The 2519 // odd-sized pieces are ordered after the requested NarrowTyArg sized 2520 // pieces. 2521 unsigned InstCount = 0; 2522 for (unsigned J = 0, JE = PartRegs.size(); J != JE; ++J) 2523 NewInsts[InstCount++].addUse(PartRegs[J]); 2524 for (unsigned J = 0, JE = LeftoverRegs.size(); J != JE; ++J) 2525 NewInsts[InstCount++].addUse(LeftoverRegs[J]); 2526 } 2527 2528 PartRegs.clear(); 2529 LeftoverRegs.clear(); 2530 } 2531 2532 // Insert the newly built operations and rebuild the result register. 2533 for (auto &MIB : NewInsts) 2534 MIRBuilder.insertInstr(MIB); 2535 2536 insertParts(DstReg, DstTy, NarrowTy0, DstRegs, LeftoverTy0, LeftoverDstRegs); 2537 2538 MI.eraseFromParent(); 2539 return Legalized; 2540 } 2541 2542 LegalizerHelper::LegalizeResult 2543 LegalizerHelper::fewerElementsVectorCasts(MachineInstr &MI, unsigned TypeIdx, 2544 LLT NarrowTy) { 2545 if (TypeIdx != 0) 2546 return UnableToLegalize; 2547 2548 Register DstReg = MI.getOperand(0).getReg(); 2549 Register SrcReg = MI.getOperand(1).getReg(); 2550 LLT DstTy = MRI.getType(DstReg); 2551 LLT SrcTy = MRI.getType(SrcReg); 2552 2553 LLT NarrowTy0 = NarrowTy; 2554 LLT NarrowTy1; 2555 unsigned NumParts; 2556 2557 if (NarrowTy.isVector()) { 2558 // Uneven breakdown not handled. 2559 NumParts = DstTy.getNumElements() / NarrowTy.getNumElements(); 2560 if (NumParts * NarrowTy.getNumElements() != DstTy.getNumElements()) 2561 return UnableToLegalize; 2562 2563 NarrowTy1 = LLT::vector(NumParts, SrcTy.getElementType().getSizeInBits()); 2564 } else { 2565 NumParts = DstTy.getNumElements(); 2566 NarrowTy1 = SrcTy.getElementType(); 2567 } 2568 2569 SmallVector<Register, 4> SrcRegs, DstRegs; 2570 extractParts(SrcReg, NarrowTy1, NumParts, SrcRegs); 2571 2572 for (unsigned I = 0; I < NumParts; ++I) { 2573 Register DstReg = MRI.createGenericVirtualRegister(NarrowTy0); 2574 MachineInstr *NewInst = MIRBuilder.buildInstr(MI.getOpcode()) 2575 .addDef(DstReg) 2576 .addUse(SrcRegs[I]); 2577 2578 NewInst->setFlags(MI.getFlags()); 2579 DstRegs.push_back(DstReg); 2580 } 2581 2582 if (NarrowTy.isVector()) 2583 MIRBuilder.buildConcatVectors(DstReg, DstRegs); 2584 else 2585 MIRBuilder.buildBuildVector(DstReg, DstRegs); 2586 2587 MI.eraseFromParent(); 2588 return Legalized; 2589 } 2590 2591 LegalizerHelper::LegalizeResult 2592 LegalizerHelper::fewerElementsVectorCmp(MachineInstr &MI, unsigned TypeIdx, 2593 LLT NarrowTy) { 2594 Register DstReg = MI.getOperand(0).getReg(); 2595 Register Src0Reg = MI.getOperand(2).getReg(); 2596 LLT DstTy = MRI.getType(DstReg); 2597 LLT SrcTy = MRI.getType(Src0Reg); 2598 2599 unsigned NumParts; 2600 LLT NarrowTy0, NarrowTy1; 2601 2602 if (TypeIdx == 0) { 2603 unsigned NewElts = NarrowTy.isVector() ? NarrowTy.getNumElements() : 1; 2604 unsigned OldElts = DstTy.getNumElements(); 2605 2606 NarrowTy0 = NarrowTy; 2607 NumParts = NarrowTy.isVector() ? (OldElts / NewElts) : DstTy.getNumElements(); 2608 NarrowTy1 = NarrowTy.isVector() ? 2609 LLT::vector(NarrowTy.getNumElements(), SrcTy.getScalarSizeInBits()) : 2610 SrcTy.getElementType(); 2611 2612 } else { 2613 unsigned NewElts = NarrowTy.isVector() ? NarrowTy.getNumElements() : 1; 2614 unsigned OldElts = SrcTy.getNumElements(); 2615 2616 NumParts = NarrowTy.isVector() ? (OldElts / NewElts) : 2617 NarrowTy.getNumElements(); 2618 NarrowTy0 = LLT::vector(NarrowTy.getNumElements(), 2619 DstTy.getScalarSizeInBits()); 2620 NarrowTy1 = NarrowTy; 2621 } 2622 2623 // FIXME: Don't know how to handle the situation where the small vectors 2624 // aren't all the same size yet. 2625 if (NarrowTy1.isVector() && 2626 NarrowTy1.getNumElements() * NumParts != DstTy.getNumElements()) 2627 return UnableToLegalize; 2628 2629 CmpInst::Predicate Pred 2630 = static_cast<CmpInst::Predicate>(MI.getOperand(1).getPredicate()); 2631 2632 SmallVector<Register, 2> Src1Regs, Src2Regs, DstRegs; 2633 extractParts(MI.getOperand(2).getReg(), NarrowTy1, NumParts, Src1Regs); 2634 extractParts(MI.getOperand(3).getReg(), NarrowTy1, NumParts, Src2Regs); 2635 2636 for (unsigned I = 0; I < NumParts; ++I) { 2637 Register DstReg = MRI.createGenericVirtualRegister(NarrowTy0); 2638 DstRegs.push_back(DstReg); 2639 2640 if (MI.getOpcode() == TargetOpcode::G_ICMP) 2641 MIRBuilder.buildICmp(Pred, DstReg, Src1Regs[I], Src2Regs[I]); 2642 else { 2643 MachineInstr *NewCmp 2644 = MIRBuilder.buildFCmp(Pred, DstReg, Src1Regs[I], Src2Regs[I]); 2645 NewCmp->setFlags(MI.getFlags()); 2646 } 2647 } 2648 2649 if (NarrowTy1.isVector()) 2650 MIRBuilder.buildConcatVectors(DstReg, DstRegs); 2651 else 2652 MIRBuilder.buildBuildVector(DstReg, DstRegs); 2653 2654 MI.eraseFromParent(); 2655 return Legalized; 2656 } 2657 2658 LegalizerHelper::LegalizeResult 2659 LegalizerHelper::fewerElementsVectorSelect(MachineInstr &MI, unsigned TypeIdx, 2660 LLT NarrowTy) { 2661 Register DstReg = MI.getOperand(0).getReg(); 2662 Register CondReg = MI.getOperand(1).getReg(); 2663 2664 unsigned NumParts = 0; 2665 LLT NarrowTy0, NarrowTy1; 2666 2667 LLT DstTy = MRI.getType(DstReg); 2668 LLT CondTy = MRI.getType(CondReg); 2669 unsigned Size = DstTy.getSizeInBits(); 2670 2671 assert(TypeIdx == 0 || CondTy.isVector()); 2672 2673 if (TypeIdx == 0) { 2674 NarrowTy0 = NarrowTy; 2675 NarrowTy1 = CondTy; 2676 2677 unsigned NarrowSize = NarrowTy0.getSizeInBits(); 2678 // FIXME: Don't know how to handle the situation where the small vectors 2679 // aren't all the same size yet. 2680 if (Size % NarrowSize != 0) 2681 return UnableToLegalize; 2682 2683 NumParts = Size / NarrowSize; 2684 2685 // Need to break down the condition type 2686 if (CondTy.isVector()) { 2687 if (CondTy.getNumElements() == NumParts) 2688 NarrowTy1 = CondTy.getElementType(); 2689 else 2690 NarrowTy1 = LLT::vector(CondTy.getNumElements() / NumParts, 2691 CondTy.getScalarSizeInBits()); 2692 } 2693 } else { 2694 NumParts = CondTy.getNumElements(); 2695 if (NarrowTy.isVector()) { 2696 // TODO: Handle uneven breakdown. 2697 if (NumParts * NarrowTy.getNumElements() != CondTy.getNumElements()) 2698 return UnableToLegalize; 2699 2700 return UnableToLegalize; 2701 } else { 2702 NarrowTy0 = DstTy.getElementType(); 2703 NarrowTy1 = NarrowTy; 2704 } 2705 } 2706 2707 SmallVector<Register, 2> DstRegs, Src0Regs, Src1Regs, Src2Regs; 2708 if (CondTy.isVector()) 2709 extractParts(MI.getOperand(1).getReg(), NarrowTy1, NumParts, Src0Regs); 2710 2711 extractParts(MI.getOperand(2).getReg(), NarrowTy0, NumParts, Src1Regs); 2712 extractParts(MI.getOperand(3).getReg(), NarrowTy0, NumParts, Src2Regs); 2713 2714 for (unsigned i = 0; i < NumParts; ++i) { 2715 Register DstReg = MRI.createGenericVirtualRegister(NarrowTy0); 2716 MIRBuilder.buildSelect(DstReg, CondTy.isVector() ? Src0Regs[i] : CondReg, 2717 Src1Regs[i], Src2Regs[i]); 2718 DstRegs.push_back(DstReg); 2719 } 2720 2721 if (NarrowTy0.isVector()) 2722 MIRBuilder.buildConcatVectors(DstReg, DstRegs); 2723 else 2724 MIRBuilder.buildBuildVector(DstReg, DstRegs); 2725 2726 MI.eraseFromParent(); 2727 return Legalized; 2728 } 2729 2730 LegalizerHelper::LegalizeResult 2731 LegalizerHelper::fewerElementsVectorPhi(MachineInstr &MI, unsigned TypeIdx, 2732 LLT NarrowTy) { 2733 const Register DstReg = MI.getOperand(0).getReg(); 2734 LLT PhiTy = MRI.getType(DstReg); 2735 LLT LeftoverTy; 2736 2737 // All of the operands need to have the same number of elements, so if we can 2738 // determine a type breakdown for the result type, we can for all of the 2739 // source types. 2740 int NumParts, NumLeftover; 2741 std::tie(NumParts, NumLeftover) 2742 = getNarrowTypeBreakDown(PhiTy, NarrowTy, LeftoverTy); 2743 if (NumParts < 0) 2744 return UnableToLegalize; 2745 2746 SmallVector<Register, 4> DstRegs, LeftoverDstRegs; 2747 SmallVector<MachineInstrBuilder, 4> NewInsts; 2748 2749 const int TotalNumParts = NumParts + NumLeftover; 2750 2751 // Insert the new phis in the result block first. 2752 for (int I = 0; I != TotalNumParts; ++I) { 2753 LLT Ty = I < NumParts ? NarrowTy : LeftoverTy; 2754 Register PartDstReg = MRI.createGenericVirtualRegister(Ty); 2755 NewInsts.push_back(MIRBuilder.buildInstr(TargetOpcode::G_PHI) 2756 .addDef(PartDstReg)); 2757 if (I < NumParts) 2758 DstRegs.push_back(PartDstReg); 2759 else 2760 LeftoverDstRegs.push_back(PartDstReg); 2761 } 2762 2763 MachineBasicBlock *MBB = MI.getParent(); 2764 MIRBuilder.setInsertPt(*MBB, MBB->getFirstNonPHI()); 2765 insertParts(DstReg, PhiTy, NarrowTy, DstRegs, LeftoverTy, LeftoverDstRegs); 2766 2767 SmallVector<Register, 4> PartRegs, LeftoverRegs; 2768 2769 // Insert code to extract the incoming values in each predecessor block. 2770 for (unsigned I = 1, E = MI.getNumOperands(); I != E; I += 2) { 2771 PartRegs.clear(); 2772 LeftoverRegs.clear(); 2773 2774 Register SrcReg = MI.getOperand(I).getReg(); 2775 MachineBasicBlock &OpMBB = *MI.getOperand(I + 1).getMBB(); 2776 MIRBuilder.setInsertPt(OpMBB, OpMBB.getFirstTerminator()); 2777 2778 LLT Unused; 2779 if (!extractParts(SrcReg, PhiTy, NarrowTy, Unused, PartRegs, 2780 LeftoverRegs)) 2781 return UnableToLegalize; 2782 2783 // Add the newly created operand splits to the existing instructions. The 2784 // odd-sized pieces are ordered after the requested NarrowTyArg sized 2785 // pieces. 2786 for (int J = 0; J != TotalNumParts; ++J) { 2787 MachineInstrBuilder MIB = NewInsts[J]; 2788 MIB.addUse(J < NumParts ? PartRegs[J] : LeftoverRegs[J - NumParts]); 2789 MIB.addMBB(&OpMBB); 2790 } 2791 } 2792 2793 MI.eraseFromParent(); 2794 return Legalized; 2795 } 2796 2797 LegalizerHelper::LegalizeResult 2798 LegalizerHelper::fewerElementsVectorUnmergeValues(MachineInstr &MI, 2799 unsigned TypeIdx, 2800 LLT NarrowTy) { 2801 if (TypeIdx != 1) 2802 return UnableToLegalize; 2803 2804 const int NumDst = MI.getNumOperands() - 1; 2805 const Register SrcReg = MI.getOperand(NumDst).getReg(); 2806 LLT SrcTy = MRI.getType(SrcReg); 2807 2808 LLT DstTy = MRI.getType(MI.getOperand(0).getReg()); 2809 2810 // TODO: Create sequence of extracts. 2811 if (DstTy == NarrowTy) 2812 return UnableToLegalize; 2813 2814 LLT GCDTy = getGCDType(SrcTy, NarrowTy); 2815 if (DstTy == GCDTy) { 2816 // This would just be a copy of the same unmerge. 2817 // TODO: Create extracts, pad with undef and create intermediate merges. 2818 return UnableToLegalize; 2819 } 2820 2821 auto Unmerge = MIRBuilder.buildUnmerge(GCDTy, SrcReg); 2822 const int NumUnmerge = Unmerge->getNumOperands() - 1; 2823 const int PartsPerUnmerge = NumDst / NumUnmerge; 2824 2825 for (int I = 0; I != NumUnmerge; ++I) { 2826 auto MIB = MIRBuilder.buildInstr(TargetOpcode::G_UNMERGE_VALUES); 2827 2828 for (int J = 0; J != PartsPerUnmerge; ++J) 2829 MIB.addDef(MI.getOperand(I * PartsPerUnmerge + J).getReg()); 2830 MIB.addUse(Unmerge.getReg(I)); 2831 } 2832 2833 MI.eraseFromParent(); 2834 return Legalized; 2835 } 2836 2837 LegalizerHelper::LegalizeResult 2838 LegalizerHelper::fewerElementsVectorBuildVector(MachineInstr &MI, 2839 unsigned TypeIdx, 2840 LLT NarrowTy) { 2841 assert(TypeIdx == 0 && "not a vector type index"); 2842 Register DstReg = MI.getOperand(0).getReg(); 2843 LLT DstTy = MRI.getType(DstReg); 2844 LLT SrcTy = DstTy.getElementType(); 2845 2846 int DstNumElts = DstTy.getNumElements(); 2847 int NarrowNumElts = NarrowTy.getNumElements(); 2848 int NumConcat = (DstNumElts + NarrowNumElts - 1) / NarrowNumElts; 2849 LLT WidenedDstTy = LLT::vector(NarrowNumElts * NumConcat, SrcTy); 2850 2851 SmallVector<Register, 8> ConcatOps; 2852 SmallVector<Register, 8> SubBuildVector; 2853 2854 Register UndefReg; 2855 if (WidenedDstTy != DstTy) 2856 UndefReg = MIRBuilder.buildUndef(SrcTy).getReg(0); 2857 2858 // Create a G_CONCAT_VECTORS of NarrowTy pieces, padding with undef as 2859 // necessary. 2860 // 2861 // %3:_(<3 x s16>) = G_BUILD_VECTOR %0, %1, %2 2862 // -> <2 x s16> 2863 // 2864 // %4:_(s16) = G_IMPLICIT_DEF 2865 // %5:_(<2 x s16>) = G_BUILD_VECTOR %0, %1 2866 // %6:_(<2 x s16>) = G_BUILD_VECTOR %2, %4 2867 // %7:_(<4 x s16>) = G_CONCAT_VECTORS %5, %6 2868 // %3:_(<3 x s16>) = G_EXTRACT %7, 0 2869 for (int I = 0; I != NumConcat; ++I) { 2870 for (int J = 0; J != NarrowNumElts; ++J) { 2871 int SrcIdx = NarrowNumElts * I + J; 2872 2873 if (SrcIdx < DstNumElts) { 2874 Register SrcReg = MI.getOperand(SrcIdx + 1).getReg(); 2875 SubBuildVector.push_back(SrcReg); 2876 } else 2877 SubBuildVector.push_back(UndefReg); 2878 } 2879 2880 auto BuildVec = MIRBuilder.buildBuildVector(NarrowTy, SubBuildVector); 2881 ConcatOps.push_back(BuildVec.getReg(0)); 2882 SubBuildVector.clear(); 2883 } 2884 2885 if (DstTy == WidenedDstTy) 2886 MIRBuilder.buildConcatVectors(DstReg, ConcatOps); 2887 else { 2888 auto Concat = MIRBuilder.buildConcatVectors(WidenedDstTy, ConcatOps); 2889 MIRBuilder.buildExtract(DstReg, Concat, 0); 2890 } 2891 2892 MI.eraseFromParent(); 2893 return Legalized; 2894 } 2895 2896 LegalizerHelper::LegalizeResult 2897 LegalizerHelper::reduceLoadStoreWidth(MachineInstr &MI, unsigned TypeIdx, 2898 LLT NarrowTy) { 2899 // FIXME: Don't know how to handle secondary types yet. 2900 if (TypeIdx != 0) 2901 return UnableToLegalize; 2902 2903 MachineMemOperand *MMO = *MI.memoperands_begin(); 2904 2905 // This implementation doesn't work for atomics. Give up instead of doing 2906 // something invalid. 2907 if (MMO->getOrdering() != AtomicOrdering::NotAtomic || 2908 MMO->getFailureOrdering() != AtomicOrdering::NotAtomic) 2909 return UnableToLegalize; 2910 2911 bool IsLoad = MI.getOpcode() == TargetOpcode::G_LOAD; 2912 Register ValReg = MI.getOperand(0).getReg(); 2913 Register AddrReg = MI.getOperand(1).getReg(); 2914 LLT ValTy = MRI.getType(ValReg); 2915 2916 int NumParts = -1; 2917 int NumLeftover = -1; 2918 LLT LeftoverTy; 2919 SmallVector<Register, 8> NarrowRegs, NarrowLeftoverRegs; 2920 if (IsLoad) { 2921 std::tie(NumParts, NumLeftover) = getNarrowTypeBreakDown(ValTy, NarrowTy, LeftoverTy); 2922 } else { 2923 if (extractParts(ValReg, ValTy, NarrowTy, LeftoverTy, NarrowRegs, 2924 NarrowLeftoverRegs)) { 2925 NumParts = NarrowRegs.size(); 2926 NumLeftover = NarrowLeftoverRegs.size(); 2927 } 2928 } 2929 2930 if (NumParts == -1) 2931 return UnableToLegalize; 2932 2933 const LLT OffsetTy = LLT::scalar(MRI.getType(AddrReg).getScalarSizeInBits()); 2934 2935 unsigned TotalSize = ValTy.getSizeInBits(); 2936 2937 // Split the load/store into PartTy sized pieces starting at Offset. If this 2938 // is a load, return the new registers in ValRegs. For a store, each elements 2939 // of ValRegs should be PartTy. Returns the next offset that needs to be 2940 // handled. 2941 auto splitTypePieces = [=](LLT PartTy, SmallVectorImpl<Register> &ValRegs, 2942 unsigned Offset) -> unsigned { 2943 MachineFunction &MF = MIRBuilder.getMF(); 2944 unsigned PartSize = PartTy.getSizeInBits(); 2945 for (unsigned Idx = 0, E = NumParts; Idx != E && Offset < TotalSize; 2946 Offset += PartSize, ++Idx) { 2947 unsigned ByteSize = PartSize / 8; 2948 unsigned ByteOffset = Offset / 8; 2949 Register NewAddrReg; 2950 2951 MIRBuilder.materializePtrAdd(NewAddrReg, AddrReg, OffsetTy, ByteOffset); 2952 2953 MachineMemOperand *NewMMO = 2954 MF.getMachineMemOperand(MMO, ByteOffset, ByteSize); 2955 2956 if (IsLoad) { 2957 Register Dst = MRI.createGenericVirtualRegister(PartTy); 2958 ValRegs.push_back(Dst); 2959 MIRBuilder.buildLoad(Dst, NewAddrReg, *NewMMO); 2960 } else { 2961 MIRBuilder.buildStore(ValRegs[Idx], NewAddrReg, *NewMMO); 2962 } 2963 } 2964 2965 return Offset; 2966 }; 2967 2968 unsigned HandledOffset = splitTypePieces(NarrowTy, NarrowRegs, 0); 2969 2970 // Handle the rest of the register if this isn't an even type breakdown. 2971 if (LeftoverTy.isValid()) 2972 splitTypePieces(LeftoverTy, NarrowLeftoverRegs, HandledOffset); 2973 2974 if (IsLoad) { 2975 insertParts(ValReg, ValTy, NarrowTy, NarrowRegs, 2976 LeftoverTy, NarrowLeftoverRegs); 2977 } 2978 2979 MI.eraseFromParent(); 2980 return Legalized; 2981 } 2982 2983 LegalizerHelper::LegalizeResult 2984 LegalizerHelper::fewerElementsVector(MachineInstr &MI, unsigned TypeIdx, 2985 LLT NarrowTy) { 2986 using namespace TargetOpcode; 2987 2988 MIRBuilder.setInstr(MI); 2989 switch (MI.getOpcode()) { 2990 case G_IMPLICIT_DEF: 2991 return fewerElementsVectorImplicitDef(MI, TypeIdx, NarrowTy); 2992 case G_AND: 2993 case G_OR: 2994 case G_XOR: 2995 case G_ADD: 2996 case G_SUB: 2997 case G_MUL: 2998 case G_SMULH: 2999 case G_UMULH: 3000 case G_FADD: 3001 case G_FMUL: 3002 case G_FSUB: 3003 case G_FNEG: 3004 case G_FABS: 3005 case G_FCANONICALIZE: 3006 case G_FDIV: 3007 case G_FREM: 3008 case G_FMA: 3009 case G_FMAD: 3010 case G_FPOW: 3011 case G_FEXP: 3012 case G_FEXP2: 3013 case G_FLOG: 3014 case G_FLOG2: 3015 case G_FLOG10: 3016 case G_FNEARBYINT: 3017 case G_FCEIL: 3018 case G_FFLOOR: 3019 case G_FRINT: 3020 case G_INTRINSIC_ROUND: 3021 case G_INTRINSIC_TRUNC: 3022 case G_FCOS: 3023 case G_FSIN: 3024 case G_FSQRT: 3025 case G_BSWAP: 3026 case G_BITREVERSE: 3027 case G_SDIV: 3028 case G_UDIV: 3029 case G_SREM: 3030 case G_UREM: 3031 case G_SMIN: 3032 case G_SMAX: 3033 case G_UMIN: 3034 case G_UMAX: 3035 case G_FMINNUM: 3036 case G_FMAXNUM: 3037 case G_FMINNUM_IEEE: 3038 case G_FMAXNUM_IEEE: 3039 case G_FMINIMUM: 3040 case G_FMAXIMUM: 3041 return fewerElementsVectorBasic(MI, TypeIdx, NarrowTy); 3042 case G_SHL: 3043 case G_LSHR: 3044 case G_ASHR: 3045 case G_CTLZ: 3046 case G_CTLZ_ZERO_UNDEF: 3047 case G_CTTZ: 3048 case G_CTTZ_ZERO_UNDEF: 3049 case G_CTPOP: 3050 case G_FCOPYSIGN: 3051 return fewerElementsVectorMultiEltType(MI, TypeIdx, NarrowTy); 3052 case G_ZEXT: 3053 case G_SEXT: 3054 case G_ANYEXT: 3055 case G_FPEXT: 3056 case G_FPTRUNC: 3057 case G_SITOFP: 3058 case G_UITOFP: 3059 case G_FPTOSI: 3060 case G_FPTOUI: 3061 case G_INTTOPTR: 3062 case G_PTRTOINT: 3063 case G_ADDRSPACE_CAST: 3064 return fewerElementsVectorCasts(MI, TypeIdx, NarrowTy); 3065 case G_ICMP: 3066 case G_FCMP: 3067 return fewerElementsVectorCmp(MI, TypeIdx, NarrowTy); 3068 case G_SELECT: 3069 return fewerElementsVectorSelect(MI, TypeIdx, NarrowTy); 3070 case G_PHI: 3071 return fewerElementsVectorPhi(MI, TypeIdx, NarrowTy); 3072 case G_UNMERGE_VALUES: 3073 return fewerElementsVectorUnmergeValues(MI, TypeIdx, NarrowTy); 3074 case G_BUILD_VECTOR: 3075 return fewerElementsVectorBuildVector(MI, TypeIdx, NarrowTy); 3076 case G_LOAD: 3077 case G_STORE: 3078 return reduceLoadStoreWidth(MI, TypeIdx, NarrowTy); 3079 default: 3080 return UnableToLegalize; 3081 } 3082 } 3083 3084 LegalizerHelper::LegalizeResult 3085 LegalizerHelper::narrowScalarShiftByConstant(MachineInstr &MI, const APInt &Amt, 3086 const LLT HalfTy, const LLT AmtTy) { 3087 3088 Register InL = MRI.createGenericVirtualRegister(HalfTy); 3089 Register InH = MRI.createGenericVirtualRegister(HalfTy); 3090 MIRBuilder.buildUnmerge({InL, InH}, MI.getOperand(1).getReg()); 3091 3092 if (Amt.isNullValue()) { 3093 MIRBuilder.buildMerge(MI.getOperand(0).getReg(), {InL, InH}); 3094 MI.eraseFromParent(); 3095 return Legalized; 3096 } 3097 3098 LLT NVT = HalfTy; 3099 unsigned NVTBits = HalfTy.getSizeInBits(); 3100 unsigned VTBits = 2 * NVTBits; 3101 3102 SrcOp Lo(Register(0)), Hi(Register(0)); 3103 if (MI.getOpcode() == TargetOpcode::G_SHL) { 3104 if (Amt.ugt(VTBits)) { 3105 Lo = Hi = MIRBuilder.buildConstant(NVT, 0); 3106 } else if (Amt.ugt(NVTBits)) { 3107 Lo = MIRBuilder.buildConstant(NVT, 0); 3108 Hi = MIRBuilder.buildShl(NVT, InL, 3109 MIRBuilder.buildConstant(AmtTy, Amt - NVTBits)); 3110 } else if (Amt == NVTBits) { 3111 Lo = MIRBuilder.buildConstant(NVT, 0); 3112 Hi = InL; 3113 } else { 3114 Lo = MIRBuilder.buildShl(NVT, InL, MIRBuilder.buildConstant(AmtTy, Amt)); 3115 auto OrLHS = 3116 MIRBuilder.buildShl(NVT, InH, MIRBuilder.buildConstant(AmtTy, Amt)); 3117 auto OrRHS = MIRBuilder.buildLShr( 3118 NVT, InL, MIRBuilder.buildConstant(AmtTy, -Amt + NVTBits)); 3119 Hi = MIRBuilder.buildOr(NVT, OrLHS, OrRHS); 3120 } 3121 } else if (MI.getOpcode() == TargetOpcode::G_LSHR) { 3122 if (Amt.ugt(VTBits)) { 3123 Lo = Hi = MIRBuilder.buildConstant(NVT, 0); 3124 } else if (Amt.ugt(NVTBits)) { 3125 Lo = MIRBuilder.buildLShr(NVT, InH, 3126 MIRBuilder.buildConstant(AmtTy, Amt - NVTBits)); 3127 Hi = MIRBuilder.buildConstant(NVT, 0); 3128 } else if (Amt == NVTBits) { 3129 Lo = InH; 3130 Hi = MIRBuilder.buildConstant(NVT, 0); 3131 } else { 3132 auto ShiftAmtConst = MIRBuilder.buildConstant(AmtTy, Amt); 3133 3134 auto OrLHS = MIRBuilder.buildLShr(NVT, InL, ShiftAmtConst); 3135 auto OrRHS = MIRBuilder.buildShl( 3136 NVT, InH, MIRBuilder.buildConstant(AmtTy, -Amt + NVTBits)); 3137 3138 Lo = MIRBuilder.buildOr(NVT, OrLHS, OrRHS); 3139 Hi = MIRBuilder.buildLShr(NVT, InH, ShiftAmtConst); 3140 } 3141 } else { 3142 if (Amt.ugt(VTBits)) { 3143 Hi = Lo = MIRBuilder.buildAShr( 3144 NVT, InH, MIRBuilder.buildConstant(AmtTy, NVTBits - 1)); 3145 } else if (Amt.ugt(NVTBits)) { 3146 Lo = MIRBuilder.buildAShr(NVT, InH, 3147 MIRBuilder.buildConstant(AmtTy, Amt - NVTBits)); 3148 Hi = MIRBuilder.buildAShr(NVT, InH, 3149 MIRBuilder.buildConstant(AmtTy, NVTBits - 1)); 3150 } else if (Amt == NVTBits) { 3151 Lo = InH; 3152 Hi = MIRBuilder.buildAShr(NVT, InH, 3153 MIRBuilder.buildConstant(AmtTy, NVTBits - 1)); 3154 } else { 3155 auto ShiftAmtConst = MIRBuilder.buildConstant(AmtTy, Amt); 3156 3157 auto OrLHS = MIRBuilder.buildLShr(NVT, InL, ShiftAmtConst); 3158 auto OrRHS = MIRBuilder.buildShl( 3159 NVT, InH, MIRBuilder.buildConstant(AmtTy, -Amt + NVTBits)); 3160 3161 Lo = MIRBuilder.buildOr(NVT, OrLHS, OrRHS); 3162 Hi = MIRBuilder.buildAShr(NVT, InH, ShiftAmtConst); 3163 } 3164 } 3165 3166 MIRBuilder.buildMerge(MI.getOperand(0).getReg(), {Lo.getReg(), Hi.getReg()}); 3167 MI.eraseFromParent(); 3168 3169 return Legalized; 3170 } 3171 3172 // TODO: Optimize if constant shift amount. 3173 LegalizerHelper::LegalizeResult 3174 LegalizerHelper::narrowScalarShift(MachineInstr &MI, unsigned TypeIdx, 3175 LLT RequestedTy) { 3176 if (TypeIdx == 1) { 3177 Observer.changingInstr(MI); 3178 narrowScalarSrc(MI, RequestedTy, 2); 3179 Observer.changedInstr(MI); 3180 return Legalized; 3181 } 3182 3183 Register DstReg = MI.getOperand(0).getReg(); 3184 LLT DstTy = MRI.getType(DstReg); 3185 if (DstTy.isVector()) 3186 return UnableToLegalize; 3187 3188 Register Amt = MI.getOperand(2).getReg(); 3189 LLT ShiftAmtTy = MRI.getType(Amt); 3190 const unsigned DstEltSize = DstTy.getScalarSizeInBits(); 3191 if (DstEltSize % 2 != 0) 3192 return UnableToLegalize; 3193 3194 // Ignore the input type. We can only go to exactly half the size of the 3195 // input. If that isn't small enough, the resulting pieces will be further 3196 // legalized. 3197 const unsigned NewBitSize = DstEltSize / 2; 3198 const LLT HalfTy = LLT::scalar(NewBitSize); 3199 const LLT CondTy = LLT::scalar(1); 3200 3201 if (const MachineInstr *KShiftAmt = 3202 getOpcodeDef(TargetOpcode::G_CONSTANT, Amt, MRI)) { 3203 return narrowScalarShiftByConstant( 3204 MI, KShiftAmt->getOperand(1).getCImm()->getValue(), HalfTy, ShiftAmtTy); 3205 } 3206 3207 // TODO: Expand with known bits. 3208 3209 // Handle the fully general expansion by an unknown amount. 3210 auto NewBits = MIRBuilder.buildConstant(ShiftAmtTy, NewBitSize); 3211 3212 Register InL = MRI.createGenericVirtualRegister(HalfTy); 3213 Register InH = MRI.createGenericVirtualRegister(HalfTy); 3214 MIRBuilder.buildUnmerge({InL, InH}, MI.getOperand(1).getReg()); 3215 3216 auto AmtExcess = MIRBuilder.buildSub(ShiftAmtTy, Amt, NewBits); 3217 auto AmtLack = MIRBuilder.buildSub(ShiftAmtTy, NewBits, Amt); 3218 3219 auto Zero = MIRBuilder.buildConstant(ShiftAmtTy, 0); 3220 auto IsShort = MIRBuilder.buildICmp(ICmpInst::ICMP_ULT, CondTy, Amt, NewBits); 3221 auto IsZero = MIRBuilder.buildICmp(ICmpInst::ICMP_EQ, CondTy, Amt, Zero); 3222 3223 Register ResultRegs[2]; 3224 switch (MI.getOpcode()) { 3225 case TargetOpcode::G_SHL: { 3226 // Short: ShAmt < NewBitSize 3227 auto LoS = MIRBuilder.buildShl(HalfTy, InL, Amt); 3228 3229 auto LoOr = MIRBuilder.buildLShr(HalfTy, InL, AmtLack); 3230 auto HiOr = MIRBuilder.buildShl(HalfTy, InH, Amt); 3231 auto HiS = MIRBuilder.buildOr(HalfTy, LoOr, HiOr); 3232 3233 // Long: ShAmt >= NewBitSize 3234 auto LoL = MIRBuilder.buildConstant(HalfTy, 0); // Lo part is zero. 3235 auto HiL = MIRBuilder.buildShl(HalfTy, InL, AmtExcess); // Hi from Lo part. 3236 3237 auto Lo = MIRBuilder.buildSelect(HalfTy, IsShort, LoS, LoL); 3238 auto Hi = MIRBuilder.buildSelect( 3239 HalfTy, IsZero, InH, MIRBuilder.buildSelect(HalfTy, IsShort, HiS, HiL)); 3240 3241 ResultRegs[0] = Lo.getReg(0); 3242 ResultRegs[1] = Hi.getReg(0); 3243 break; 3244 } 3245 case TargetOpcode::G_LSHR: 3246 case TargetOpcode::G_ASHR: { 3247 // Short: ShAmt < NewBitSize 3248 auto HiS = MIRBuilder.buildInstr(MI.getOpcode(), {HalfTy}, {InH, Amt}); 3249 3250 auto LoOr = MIRBuilder.buildLShr(HalfTy, InL, Amt); 3251 auto HiOr = MIRBuilder.buildShl(HalfTy, InH, AmtLack); 3252 auto LoS = MIRBuilder.buildOr(HalfTy, LoOr, HiOr); 3253 3254 // Long: ShAmt >= NewBitSize 3255 MachineInstrBuilder HiL; 3256 if (MI.getOpcode() == TargetOpcode::G_LSHR) { 3257 HiL = MIRBuilder.buildConstant(HalfTy, 0); // Hi part is zero. 3258 } else { 3259 auto ShiftAmt = MIRBuilder.buildConstant(ShiftAmtTy, NewBitSize - 1); 3260 HiL = MIRBuilder.buildAShr(HalfTy, InH, ShiftAmt); // Sign of Hi part. 3261 } 3262 auto LoL = MIRBuilder.buildInstr(MI.getOpcode(), {HalfTy}, 3263 {InH, AmtExcess}); // Lo from Hi part. 3264 3265 auto Lo = MIRBuilder.buildSelect( 3266 HalfTy, IsZero, InL, MIRBuilder.buildSelect(HalfTy, IsShort, LoS, LoL)); 3267 3268 auto Hi = MIRBuilder.buildSelect(HalfTy, IsShort, HiS, HiL); 3269 3270 ResultRegs[0] = Lo.getReg(0); 3271 ResultRegs[1] = Hi.getReg(0); 3272 break; 3273 } 3274 default: 3275 llvm_unreachable("not a shift"); 3276 } 3277 3278 MIRBuilder.buildMerge(DstReg, ResultRegs); 3279 MI.eraseFromParent(); 3280 return Legalized; 3281 } 3282 3283 LegalizerHelper::LegalizeResult 3284 LegalizerHelper::moreElementsVectorPhi(MachineInstr &MI, unsigned TypeIdx, 3285 LLT MoreTy) { 3286 assert(TypeIdx == 0 && "Expecting only Idx 0"); 3287 3288 Observer.changingInstr(MI); 3289 for (unsigned I = 1, E = MI.getNumOperands(); I != E; I += 2) { 3290 MachineBasicBlock &OpMBB = *MI.getOperand(I + 1).getMBB(); 3291 MIRBuilder.setInsertPt(OpMBB, OpMBB.getFirstTerminator()); 3292 moreElementsVectorSrc(MI, MoreTy, I); 3293 } 3294 3295 MachineBasicBlock &MBB = *MI.getParent(); 3296 MIRBuilder.setInsertPt(MBB, --MBB.getFirstNonPHI()); 3297 moreElementsVectorDst(MI, MoreTy, 0); 3298 Observer.changedInstr(MI); 3299 return Legalized; 3300 } 3301 3302 LegalizerHelper::LegalizeResult 3303 LegalizerHelper::moreElementsVector(MachineInstr &MI, unsigned TypeIdx, 3304 LLT MoreTy) { 3305 MIRBuilder.setInstr(MI); 3306 unsigned Opc = MI.getOpcode(); 3307 switch (Opc) { 3308 case TargetOpcode::G_IMPLICIT_DEF: 3309 case TargetOpcode::G_LOAD: { 3310 if (TypeIdx != 0) 3311 return UnableToLegalize; 3312 Observer.changingInstr(MI); 3313 moreElementsVectorDst(MI, MoreTy, 0); 3314 Observer.changedInstr(MI); 3315 return Legalized; 3316 } 3317 case TargetOpcode::G_STORE: 3318 if (TypeIdx != 0) 3319 return UnableToLegalize; 3320 Observer.changingInstr(MI); 3321 moreElementsVectorSrc(MI, MoreTy, 0); 3322 Observer.changedInstr(MI); 3323 return Legalized; 3324 case TargetOpcode::G_AND: 3325 case TargetOpcode::G_OR: 3326 case TargetOpcode::G_XOR: 3327 case TargetOpcode::G_SMIN: 3328 case TargetOpcode::G_SMAX: 3329 case TargetOpcode::G_UMIN: 3330 case TargetOpcode::G_UMAX: 3331 case TargetOpcode::G_FMINNUM: 3332 case TargetOpcode::G_FMAXNUM: 3333 case TargetOpcode::G_FMINNUM_IEEE: 3334 case TargetOpcode::G_FMAXNUM_IEEE: 3335 case TargetOpcode::G_FMINIMUM: 3336 case TargetOpcode::G_FMAXIMUM: { 3337 Observer.changingInstr(MI); 3338 moreElementsVectorSrc(MI, MoreTy, 1); 3339 moreElementsVectorSrc(MI, MoreTy, 2); 3340 moreElementsVectorDst(MI, MoreTy, 0); 3341 Observer.changedInstr(MI); 3342 return Legalized; 3343 } 3344 case TargetOpcode::G_EXTRACT: 3345 if (TypeIdx != 1) 3346 return UnableToLegalize; 3347 Observer.changingInstr(MI); 3348 moreElementsVectorSrc(MI, MoreTy, 1); 3349 Observer.changedInstr(MI); 3350 return Legalized; 3351 case TargetOpcode::G_INSERT: 3352 if (TypeIdx != 0) 3353 return UnableToLegalize; 3354 Observer.changingInstr(MI); 3355 moreElementsVectorSrc(MI, MoreTy, 1); 3356 moreElementsVectorDst(MI, MoreTy, 0); 3357 Observer.changedInstr(MI); 3358 return Legalized; 3359 case TargetOpcode::G_SELECT: 3360 if (TypeIdx != 0) 3361 return UnableToLegalize; 3362 if (MRI.getType(MI.getOperand(1).getReg()).isVector()) 3363 return UnableToLegalize; 3364 3365 Observer.changingInstr(MI); 3366 moreElementsVectorSrc(MI, MoreTy, 2); 3367 moreElementsVectorSrc(MI, MoreTy, 3); 3368 moreElementsVectorDst(MI, MoreTy, 0); 3369 Observer.changedInstr(MI); 3370 return Legalized; 3371 case TargetOpcode::G_UNMERGE_VALUES: { 3372 if (TypeIdx != 1) 3373 return UnableToLegalize; 3374 3375 LLT DstTy = MRI.getType(MI.getOperand(0).getReg()); 3376 int NumDst = MI.getNumOperands() - 1; 3377 moreElementsVectorSrc(MI, MoreTy, NumDst); 3378 3379 auto MIB = MIRBuilder.buildInstr(TargetOpcode::G_UNMERGE_VALUES); 3380 for (int I = 0; I != NumDst; ++I) 3381 MIB.addDef(MI.getOperand(I).getReg()); 3382 3383 int NewNumDst = MoreTy.getSizeInBits() / DstTy.getSizeInBits(); 3384 for (int I = NumDst; I != NewNumDst; ++I) 3385 MIB.addDef(MRI.createGenericVirtualRegister(DstTy)); 3386 3387 MIB.addUse(MI.getOperand(NumDst).getReg()); 3388 MI.eraseFromParent(); 3389 return Legalized; 3390 } 3391 case TargetOpcode::G_PHI: 3392 return moreElementsVectorPhi(MI, TypeIdx, MoreTy); 3393 default: 3394 return UnableToLegalize; 3395 } 3396 } 3397 3398 void LegalizerHelper::multiplyRegisters(SmallVectorImpl<Register> &DstRegs, 3399 ArrayRef<Register> Src1Regs, 3400 ArrayRef<Register> Src2Regs, 3401 LLT NarrowTy) { 3402 MachineIRBuilder &B = MIRBuilder; 3403 unsigned SrcParts = Src1Regs.size(); 3404 unsigned DstParts = DstRegs.size(); 3405 3406 unsigned DstIdx = 0; // Low bits of the result. 3407 Register FactorSum = 3408 B.buildMul(NarrowTy, Src1Regs[DstIdx], Src2Regs[DstIdx]).getReg(0); 3409 DstRegs[DstIdx] = FactorSum; 3410 3411 unsigned CarrySumPrevDstIdx; 3412 SmallVector<Register, 4> Factors; 3413 3414 for (DstIdx = 1; DstIdx < DstParts; DstIdx++) { 3415 // Collect low parts of muls for DstIdx. 3416 for (unsigned i = DstIdx + 1 < SrcParts ? 0 : DstIdx - SrcParts + 1; 3417 i <= std::min(DstIdx, SrcParts - 1); ++i) { 3418 MachineInstrBuilder Mul = 3419 B.buildMul(NarrowTy, Src1Regs[DstIdx - i], Src2Regs[i]); 3420 Factors.push_back(Mul.getReg(0)); 3421 } 3422 // Collect high parts of muls from previous DstIdx. 3423 for (unsigned i = DstIdx < SrcParts ? 0 : DstIdx - SrcParts; 3424 i <= std::min(DstIdx - 1, SrcParts - 1); ++i) { 3425 MachineInstrBuilder Umulh = 3426 B.buildUMulH(NarrowTy, Src1Regs[DstIdx - 1 - i], Src2Regs[i]); 3427 Factors.push_back(Umulh.getReg(0)); 3428 } 3429 // Add CarrySum from additions calculated for previous DstIdx. 3430 if (DstIdx != 1) { 3431 Factors.push_back(CarrySumPrevDstIdx); 3432 } 3433 3434 Register CarrySum; 3435 // Add all factors and accumulate all carries into CarrySum. 3436 if (DstIdx != DstParts - 1) { 3437 MachineInstrBuilder Uaddo = 3438 B.buildUAddo(NarrowTy, LLT::scalar(1), Factors[0], Factors[1]); 3439 FactorSum = Uaddo.getReg(0); 3440 CarrySum = B.buildZExt(NarrowTy, Uaddo.getReg(1)).getReg(0); 3441 for (unsigned i = 2; i < Factors.size(); ++i) { 3442 MachineInstrBuilder Uaddo = 3443 B.buildUAddo(NarrowTy, LLT::scalar(1), FactorSum, Factors[i]); 3444 FactorSum = Uaddo.getReg(0); 3445 MachineInstrBuilder Carry = B.buildZExt(NarrowTy, Uaddo.getReg(1)); 3446 CarrySum = B.buildAdd(NarrowTy, CarrySum, Carry).getReg(0); 3447 } 3448 } else { 3449 // Since value for the next index is not calculated, neither is CarrySum. 3450 FactorSum = B.buildAdd(NarrowTy, Factors[0], Factors[1]).getReg(0); 3451 for (unsigned i = 2; i < Factors.size(); ++i) 3452 FactorSum = B.buildAdd(NarrowTy, FactorSum, Factors[i]).getReg(0); 3453 } 3454 3455 CarrySumPrevDstIdx = CarrySum; 3456 DstRegs[DstIdx] = FactorSum; 3457 Factors.clear(); 3458 } 3459 } 3460 3461 LegalizerHelper::LegalizeResult 3462 LegalizerHelper::narrowScalarMul(MachineInstr &MI, LLT NarrowTy) { 3463 Register DstReg = MI.getOperand(0).getReg(); 3464 Register Src1 = MI.getOperand(1).getReg(); 3465 Register Src2 = MI.getOperand(2).getReg(); 3466 3467 LLT Ty = MRI.getType(DstReg); 3468 if (Ty.isVector()) 3469 return UnableToLegalize; 3470 3471 unsigned SrcSize = MRI.getType(Src1).getSizeInBits(); 3472 unsigned DstSize = Ty.getSizeInBits(); 3473 unsigned NarrowSize = NarrowTy.getSizeInBits(); 3474 if (DstSize % NarrowSize != 0 || SrcSize % NarrowSize != 0) 3475 return UnableToLegalize; 3476 3477 unsigned NumDstParts = DstSize / NarrowSize; 3478 unsigned NumSrcParts = SrcSize / NarrowSize; 3479 bool IsMulHigh = MI.getOpcode() == TargetOpcode::G_UMULH; 3480 unsigned DstTmpParts = NumDstParts * (IsMulHigh ? 2 : 1); 3481 3482 SmallVector<Register, 2> Src1Parts, Src2Parts, DstTmpRegs; 3483 extractParts(Src1, NarrowTy, NumSrcParts, Src1Parts); 3484 extractParts(Src2, NarrowTy, NumSrcParts, Src2Parts); 3485 DstTmpRegs.resize(DstTmpParts); 3486 multiplyRegisters(DstTmpRegs, Src1Parts, Src2Parts, NarrowTy); 3487 3488 // Take only high half of registers if this is high mul. 3489 ArrayRef<Register> DstRegs( 3490 IsMulHigh ? &DstTmpRegs[DstTmpParts / 2] : &DstTmpRegs[0], NumDstParts); 3491 MIRBuilder.buildMerge(DstReg, DstRegs); 3492 MI.eraseFromParent(); 3493 return Legalized; 3494 } 3495 3496 LegalizerHelper::LegalizeResult 3497 LegalizerHelper::narrowScalarExtract(MachineInstr &MI, unsigned TypeIdx, 3498 LLT NarrowTy) { 3499 if (TypeIdx != 1) 3500 return UnableToLegalize; 3501 3502 uint64_t NarrowSize = NarrowTy.getSizeInBits(); 3503 3504 int64_t SizeOp1 = MRI.getType(MI.getOperand(1).getReg()).getSizeInBits(); 3505 // FIXME: add support for when SizeOp1 isn't an exact multiple of 3506 // NarrowSize. 3507 if (SizeOp1 % NarrowSize != 0) 3508 return UnableToLegalize; 3509 int NumParts = SizeOp1 / NarrowSize; 3510 3511 SmallVector<Register, 2> SrcRegs, DstRegs; 3512 SmallVector<uint64_t, 2> Indexes; 3513 extractParts(MI.getOperand(1).getReg(), NarrowTy, NumParts, SrcRegs); 3514 3515 Register OpReg = MI.getOperand(0).getReg(); 3516 uint64_t OpStart = MI.getOperand(2).getImm(); 3517 uint64_t OpSize = MRI.getType(OpReg).getSizeInBits(); 3518 for (int i = 0; i < NumParts; ++i) { 3519 unsigned SrcStart = i * NarrowSize; 3520 3521 if (SrcStart + NarrowSize <= OpStart || SrcStart >= OpStart + OpSize) { 3522 // No part of the extract uses this subregister, ignore it. 3523 continue; 3524 } else if (SrcStart == OpStart && NarrowTy == MRI.getType(OpReg)) { 3525 // The entire subregister is extracted, forward the value. 3526 DstRegs.push_back(SrcRegs[i]); 3527 continue; 3528 } 3529 3530 // OpSegStart is where this destination segment would start in OpReg if it 3531 // extended infinitely in both directions. 3532 int64_t ExtractOffset; 3533 uint64_t SegSize; 3534 if (OpStart < SrcStart) { 3535 ExtractOffset = 0; 3536 SegSize = std::min(NarrowSize, OpStart + OpSize - SrcStart); 3537 } else { 3538 ExtractOffset = OpStart - SrcStart; 3539 SegSize = std::min(SrcStart + NarrowSize - OpStart, OpSize); 3540 } 3541 3542 Register SegReg = SrcRegs[i]; 3543 if (ExtractOffset != 0 || SegSize != NarrowSize) { 3544 // A genuine extract is needed. 3545 SegReg = MRI.createGenericVirtualRegister(LLT::scalar(SegSize)); 3546 MIRBuilder.buildExtract(SegReg, SrcRegs[i], ExtractOffset); 3547 } 3548 3549 DstRegs.push_back(SegReg); 3550 } 3551 3552 Register DstReg = MI.getOperand(0).getReg(); 3553 if(MRI.getType(DstReg).isVector()) 3554 MIRBuilder.buildBuildVector(DstReg, DstRegs); 3555 else 3556 MIRBuilder.buildMerge(DstReg, DstRegs); 3557 MI.eraseFromParent(); 3558 return Legalized; 3559 } 3560 3561 LegalizerHelper::LegalizeResult 3562 LegalizerHelper::narrowScalarInsert(MachineInstr &MI, unsigned TypeIdx, 3563 LLT NarrowTy) { 3564 // FIXME: Don't know how to handle secondary types yet. 3565 if (TypeIdx != 0) 3566 return UnableToLegalize; 3567 3568 uint64_t SizeOp0 = MRI.getType(MI.getOperand(0).getReg()).getSizeInBits(); 3569 uint64_t NarrowSize = NarrowTy.getSizeInBits(); 3570 3571 // FIXME: add support for when SizeOp0 isn't an exact multiple of 3572 // NarrowSize. 3573 if (SizeOp0 % NarrowSize != 0) 3574 return UnableToLegalize; 3575 3576 int NumParts = SizeOp0 / NarrowSize; 3577 3578 SmallVector<Register, 2> SrcRegs, DstRegs; 3579 SmallVector<uint64_t, 2> Indexes; 3580 extractParts(MI.getOperand(1).getReg(), NarrowTy, NumParts, SrcRegs); 3581 3582 Register OpReg = MI.getOperand(2).getReg(); 3583 uint64_t OpStart = MI.getOperand(3).getImm(); 3584 uint64_t OpSize = MRI.getType(OpReg).getSizeInBits(); 3585 for (int i = 0; i < NumParts; ++i) { 3586 unsigned DstStart = i * NarrowSize; 3587 3588 if (DstStart + NarrowSize <= OpStart || DstStart >= OpStart + OpSize) { 3589 // No part of the insert affects this subregister, forward the original. 3590 DstRegs.push_back(SrcRegs[i]); 3591 continue; 3592 } else if (DstStart == OpStart && NarrowTy == MRI.getType(OpReg)) { 3593 // The entire subregister is defined by this insert, forward the new 3594 // value. 3595 DstRegs.push_back(OpReg); 3596 continue; 3597 } 3598 3599 // OpSegStart is where this destination segment would start in OpReg if it 3600 // extended infinitely in both directions. 3601 int64_t ExtractOffset, InsertOffset; 3602 uint64_t SegSize; 3603 if (OpStart < DstStart) { 3604 InsertOffset = 0; 3605 ExtractOffset = DstStart - OpStart; 3606 SegSize = std::min(NarrowSize, OpStart + OpSize - DstStart); 3607 } else { 3608 InsertOffset = OpStart - DstStart; 3609 ExtractOffset = 0; 3610 SegSize = 3611 std::min(NarrowSize - InsertOffset, OpStart + OpSize - DstStart); 3612 } 3613 3614 Register SegReg = OpReg; 3615 if (ExtractOffset != 0 || SegSize != OpSize) { 3616 // A genuine extract is needed. 3617 SegReg = MRI.createGenericVirtualRegister(LLT::scalar(SegSize)); 3618 MIRBuilder.buildExtract(SegReg, OpReg, ExtractOffset); 3619 } 3620 3621 Register DstReg = MRI.createGenericVirtualRegister(NarrowTy); 3622 MIRBuilder.buildInsert(DstReg, SrcRegs[i], SegReg, InsertOffset); 3623 DstRegs.push_back(DstReg); 3624 } 3625 3626 assert(DstRegs.size() == (unsigned)NumParts && "not all parts covered"); 3627 Register DstReg = MI.getOperand(0).getReg(); 3628 if(MRI.getType(DstReg).isVector()) 3629 MIRBuilder.buildBuildVector(DstReg, DstRegs); 3630 else 3631 MIRBuilder.buildMerge(DstReg, DstRegs); 3632 MI.eraseFromParent(); 3633 return Legalized; 3634 } 3635 3636 LegalizerHelper::LegalizeResult 3637 LegalizerHelper::narrowScalarBasic(MachineInstr &MI, unsigned TypeIdx, 3638 LLT NarrowTy) { 3639 Register DstReg = MI.getOperand(0).getReg(); 3640 LLT DstTy = MRI.getType(DstReg); 3641 3642 assert(MI.getNumOperands() == 3 && TypeIdx == 0); 3643 3644 SmallVector<Register, 4> DstRegs, DstLeftoverRegs; 3645 SmallVector<Register, 4> Src0Regs, Src0LeftoverRegs; 3646 SmallVector<Register, 4> Src1Regs, Src1LeftoverRegs; 3647 LLT LeftoverTy; 3648 if (!extractParts(MI.getOperand(1).getReg(), DstTy, NarrowTy, LeftoverTy, 3649 Src0Regs, Src0LeftoverRegs)) 3650 return UnableToLegalize; 3651 3652 LLT Unused; 3653 if (!extractParts(MI.getOperand(2).getReg(), DstTy, NarrowTy, Unused, 3654 Src1Regs, Src1LeftoverRegs)) 3655 llvm_unreachable("inconsistent extractParts result"); 3656 3657 for (unsigned I = 0, E = Src1Regs.size(); I != E; ++I) { 3658 auto Inst = MIRBuilder.buildInstr(MI.getOpcode(), {NarrowTy}, 3659 {Src0Regs[I], Src1Regs[I]}); 3660 DstRegs.push_back(Inst->getOperand(0).getReg()); 3661 } 3662 3663 for (unsigned I = 0, E = Src1LeftoverRegs.size(); I != E; ++I) { 3664 auto Inst = MIRBuilder.buildInstr( 3665 MI.getOpcode(), 3666 {LeftoverTy}, {Src0LeftoverRegs[I], Src1LeftoverRegs[I]}); 3667 DstLeftoverRegs.push_back(Inst->getOperand(0).getReg()); 3668 } 3669 3670 insertParts(DstReg, DstTy, NarrowTy, DstRegs, 3671 LeftoverTy, DstLeftoverRegs); 3672 3673 MI.eraseFromParent(); 3674 return Legalized; 3675 } 3676 3677 LegalizerHelper::LegalizeResult 3678 LegalizerHelper::narrowScalarSelect(MachineInstr &MI, unsigned TypeIdx, 3679 LLT NarrowTy) { 3680 if (TypeIdx != 0) 3681 return UnableToLegalize; 3682 3683 Register CondReg = MI.getOperand(1).getReg(); 3684 LLT CondTy = MRI.getType(CondReg); 3685 if (CondTy.isVector()) // TODO: Handle vselect 3686 return UnableToLegalize; 3687 3688 Register DstReg = MI.getOperand(0).getReg(); 3689 LLT DstTy = MRI.getType(DstReg); 3690 3691 SmallVector<Register, 4> DstRegs, DstLeftoverRegs; 3692 SmallVector<Register, 4> Src1Regs, Src1LeftoverRegs; 3693 SmallVector<Register, 4> Src2Regs, Src2LeftoverRegs; 3694 LLT LeftoverTy; 3695 if (!extractParts(MI.getOperand(2).getReg(), DstTy, NarrowTy, LeftoverTy, 3696 Src1Regs, Src1LeftoverRegs)) 3697 return UnableToLegalize; 3698 3699 LLT Unused; 3700 if (!extractParts(MI.getOperand(3).getReg(), DstTy, NarrowTy, Unused, 3701 Src2Regs, Src2LeftoverRegs)) 3702 llvm_unreachable("inconsistent extractParts result"); 3703 3704 for (unsigned I = 0, E = Src1Regs.size(); I != E; ++I) { 3705 auto Select = MIRBuilder.buildSelect(NarrowTy, 3706 CondReg, Src1Regs[I], Src2Regs[I]); 3707 DstRegs.push_back(Select->getOperand(0).getReg()); 3708 } 3709 3710 for (unsigned I = 0, E = Src1LeftoverRegs.size(); I != E; ++I) { 3711 auto Select = MIRBuilder.buildSelect( 3712 LeftoverTy, CondReg, Src1LeftoverRegs[I], Src2LeftoverRegs[I]); 3713 DstLeftoverRegs.push_back(Select->getOperand(0).getReg()); 3714 } 3715 3716 insertParts(DstReg, DstTy, NarrowTy, DstRegs, 3717 LeftoverTy, DstLeftoverRegs); 3718 3719 MI.eraseFromParent(); 3720 return Legalized; 3721 } 3722 3723 LegalizerHelper::LegalizeResult 3724 LegalizerHelper::lowerBitCount(MachineInstr &MI, unsigned TypeIdx, LLT Ty) { 3725 unsigned Opc = MI.getOpcode(); 3726 auto &TII = *MI.getMF()->getSubtarget().getInstrInfo(); 3727 auto isSupported = [this](const LegalityQuery &Q) { 3728 auto QAction = LI.getAction(Q).Action; 3729 return QAction == Legal || QAction == Libcall || QAction == Custom; 3730 }; 3731 switch (Opc) { 3732 default: 3733 return UnableToLegalize; 3734 case TargetOpcode::G_CTLZ_ZERO_UNDEF: { 3735 // This trivially expands to CTLZ. 3736 Observer.changingInstr(MI); 3737 MI.setDesc(TII.get(TargetOpcode::G_CTLZ)); 3738 Observer.changedInstr(MI); 3739 return Legalized; 3740 } 3741 case TargetOpcode::G_CTLZ: { 3742 Register SrcReg = MI.getOperand(1).getReg(); 3743 unsigned Len = Ty.getSizeInBits(); 3744 if (isSupported({TargetOpcode::G_CTLZ_ZERO_UNDEF, {Ty, Ty}})) { 3745 // If CTLZ_ZERO_UNDEF is supported, emit that and a select for zero. 3746 auto MIBCtlzZU = MIRBuilder.buildInstr(TargetOpcode::G_CTLZ_ZERO_UNDEF, 3747 {Ty}, {SrcReg}); 3748 auto MIBZero = MIRBuilder.buildConstant(Ty, 0); 3749 auto MIBLen = MIRBuilder.buildConstant(Ty, Len); 3750 auto MIBICmp = MIRBuilder.buildICmp(CmpInst::ICMP_EQ, LLT::scalar(1), 3751 SrcReg, MIBZero); 3752 MIRBuilder.buildSelect(MI.getOperand(0).getReg(), MIBICmp, MIBLen, 3753 MIBCtlzZU); 3754 MI.eraseFromParent(); 3755 return Legalized; 3756 } 3757 // for now, we do this: 3758 // NewLen = NextPowerOf2(Len); 3759 // x = x | (x >> 1); 3760 // x = x | (x >> 2); 3761 // ... 3762 // x = x | (x >>16); 3763 // x = x | (x >>32); // for 64-bit input 3764 // Upto NewLen/2 3765 // return Len - popcount(x); 3766 // 3767 // Ref: "Hacker's Delight" by Henry Warren 3768 Register Op = SrcReg; 3769 unsigned NewLen = PowerOf2Ceil(Len); 3770 for (unsigned i = 0; (1U << i) <= (NewLen / 2); ++i) { 3771 auto MIBShiftAmt = MIRBuilder.buildConstant(Ty, 1ULL << i); 3772 auto MIBOp = MIRBuilder.buildInstr( 3773 TargetOpcode::G_OR, {Ty}, 3774 {Op, MIRBuilder.buildInstr(TargetOpcode::G_LSHR, {Ty}, 3775 {Op, MIBShiftAmt})}); 3776 Op = MIBOp->getOperand(0).getReg(); 3777 } 3778 auto MIBPop = MIRBuilder.buildInstr(TargetOpcode::G_CTPOP, {Ty}, {Op}); 3779 MIRBuilder.buildInstr(TargetOpcode::G_SUB, {MI.getOperand(0).getReg()}, 3780 {MIRBuilder.buildConstant(Ty, Len), MIBPop}); 3781 MI.eraseFromParent(); 3782 return Legalized; 3783 } 3784 case TargetOpcode::G_CTTZ_ZERO_UNDEF: { 3785 // This trivially expands to CTTZ. 3786 Observer.changingInstr(MI); 3787 MI.setDesc(TII.get(TargetOpcode::G_CTTZ)); 3788 Observer.changedInstr(MI); 3789 return Legalized; 3790 } 3791 case TargetOpcode::G_CTTZ: { 3792 Register SrcReg = MI.getOperand(1).getReg(); 3793 unsigned Len = Ty.getSizeInBits(); 3794 if (isSupported({TargetOpcode::G_CTTZ_ZERO_UNDEF, {Ty, Ty}})) { 3795 // If CTTZ_ZERO_UNDEF is legal or custom, emit that and a select with 3796 // zero. 3797 auto MIBCttzZU = MIRBuilder.buildInstr(TargetOpcode::G_CTTZ_ZERO_UNDEF, 3798 {Ty}, {SrcReg}); 3799 auto MIBZero = MIRBuilder.buildConstant(Ty, 0); 3800 auto MIBLen = MIRBuilder.buildConstant(Ty, Len); 3801 auto MIBICmp = MIRBuilder.buildICmp(CmpInst::ICMP_EQ, LLT::scalar(1), 3802 SrcReg, MIBZero); 3803 MIRBuilder.buildSelect(MI.getOperand(0).getReg(), MIBICmp, MIBLen, 3804 MIBCttzZU); 3805 MI.eraseFromParent(); 3806 return Legalized; 3807 } 3808 // for now, we use: { return popcount(~x & (x - 1)); } 3809 // unless the target has ctlz but not ctpop, in which case we use: 3810 // { return 32 - nlz(~x & (x-1)); } 3811 // Ref: "Hacker's Delight" by Henry Warren 3812 auto MIBCstNeg1 = MIRBuilder.buildConstant(Ty, -1); 3813 auto MIBNot = 3814 MIRBuilder.buildInstr(TargetOpcode::G_XOR, {Ty}, {SrcReg, MIBCstNeg1}); 3815 auto MIBTmp = MIRBuilder.buildInstr( 3816 TargetOpcode::G_AND, {Ty}, 3817 {MIBNot, MIRBuilder.buildInstr(TargetOpcode::G_ADD, {Ty}, 3818 {SrcReg, MIBCstNeg1})}); 3819 if (!isSupported({TargetOpcode::G_CTPOP, {Ty, Ty}}) && 3820 isSupported({TargetOpcode::G_CTLZ, {Ty, Ty}})) { 3821 auto MIBCstLen = MIRBuilder.buildConstant(Ty, Len); 3822 MIRBuilder.buildInstr( 3823 TargetOpcode::G_SUB, {MI.getOperand(0).getReg()}, 3824 {MIBCstLen, 3825 MIRBuilder.buildInstr(TargetOpcode::G_CTLZ, {Ty}, {MIBTmp})}); 3826 MI.eraseFromParent(); 3827 return Legalized; 3828 } 3829 MI.setDesc(TII.get(TargetOpcode::G_CTPOP)); 3830 MI.getOperand(1).setReg(MIBTmp->getOperand(0).getReg()); 3831 return Legalized; 3832 } 3833 } 3834 } 3835 3836 // Expand s32 = G_UITOFP s64 using bit operations to an IEEE float 3837 // representation. 3838 LegalizerHelper::LegalizeResult 3839 LegalizerHelper::lowerU64ToF32BitOps(MachineInstr &MI) { 3840 Register Dst = MI.getOperand(0).getReg(); 3841 Register Src = MI.getOperand(1).getReg(); 3842 const LLT S64 = LLT::scalar(64); 3843 const LLT S32 = LLT::scalar(32); 3844 const LLT S1 = LLT::scalar(1); 3845 3846 assert(MRI.getType(Src) == S64 && MRI.getType(Dst) == S32); 3847 3848 // unsigned cul2f(ulong u) { 3849 // uint lz = clz(u); 3850 // uint e = (u != 0) ? 127U + 63U - lz : 0; 3851 // u = (u << lz) & 0x7fffffffffffffffUL; 3852 // ulong t = u & 0xffffffffffUL; 3853 // uint v = (e << 23) | (uint)(u >> 40); 3854 // uint r = t > 0x8000000000UL ? 1U : (t == 0x8000000000UL ? v & 1U : 0U); 3855 // return as_float(v + r); 3856 // } 3857 3858 auto Zero32 = MIRBuilder.buildConstant(S32, 0); 3859 auto Zero64 = MIRBuilder.buildConstant(S64, 0); 3860 3861 auto LZ = MIRBuilder.buildCTLZ_ZERO_UNDEF(S32, Src); 3862 3863 auto K = MIRBuilder.buildConstant(S32, 127U + 63U); 3864 auto Sub = MIRBuilder.buildSub(S32, K, LZ); 3865 3866 auto NotZero = MIRBuilder.buildICmp(CmpInst::ICMP_NE, S1, Src, Zero64); 3867 auto E = MIRBuilder.buildSelect(S32, NotZero, Sub, Zero32); 3868 3869 auto Mask0 = MIRBuilder.buildConstant(S64, (-1ULL) >> 1); 3870 auto ShlLZ = MIRBuilder.buildShl(S64, Src, LZ); 3871 3872 auto U = MIRBuilder.buildAnd(S64, ShlLZ, Mask0); 3873 3874 auto Mask1 = MIRBuilder.buildConstant(S64, 0xffffffffffULL); 3875 auto T = MIRBuilder.buildAnd(S64, U, Mask1); 3876 3877 auto UShl = MIRBuilder.buildLShr(S64, U, MIRBuilder.buildConstant(S64, 40)); 3878 auto ShlE = MIRBuilder.buildShl(S32, E, MIRBuilder.buildConstant(S32, 23)); 3879 auto V = MIRBuilder.buildOr(S32, ShlE, MIRBuilder.buildTrunc(S32, UShl)); 3880 3881 auto C = MIRBuilder.buildConstant(S64, 0x8000000000ULL); 3882 auto RCmp = MIRBuilder.buildICmp(CmpInst::ICMP_UGT, S1, T, C); 3883 auto TCmp = MIRBuilder.buildICmp(CmpInst::ICMP_EQ, S1, T, C); 3884 auto One = MIRBuilder.buildConstant(S32, 1); 3885 3886 auto VTrunc1 = MIRBuilder.buildAnd(S32, V, One); 3887 auto Select0 = MIRBuilder.buildSelect(S32, TCmp, VTrunc1, Zero32); 3888 auto R = MIRBuilder.buildSelect(S32, RCmp, One, Select0); 3889 MIRBuilder.buildAdd(Dst, V, R); 3890 3891 return Legalized; 3892 } 3893 3894 LegalizerHelper::LegalizeResult 3895 LegalizerHelper::lowerUITOFP(MachineInstr &MI, unsigned TypeIdx, LLT Ty) { 3896 Register Dst = MI.getOperand(0).getReg(); 3897 Register Src = MI.getOperand(1).getReg(); 3898 LLT DstTy = MRI.getType(Dst); 3899 LLT SrcTy = MRI.getType(Src); 3900 3901 if (SrcTy == LLT::scalar(1)) { 3902 auto True = MIRBuilder.buildFConstant(DstTy, 1.0); 3903 auto False = MIRBuilder.buildFConstant(DstTy, 0.0); 3904 MIRBuilder.buildSelect(Dst, Src, True, False); 3905 MI.eraseFromParent(); 3906 return Legalized; 3907 } 3908 3909 if (SrcTy != LLT::scalar(64)) 3910 return UnableToLegalize; 3911 3912 if (DstTy == LLT::scalar(32)) { 3913 // TODO: SelectionDAG has several alternative expansions to port which may 3914 // be more reasonble depending on the available instructions. If a target 3915 // has sitofp, does not have CTLZ, or can efficiently use f64 as an 3916 // intermediate type, this is probably worse. 3917 return lowerU64ToF32BitOps(MI); 3918 } 3919 3920 return UnableToLegalize; 3921 } 3922 3923 LegalizerHelper::LegalizeResult 3924 LegalizerHelper::lowerSITOFP(MachineInstr &MI, unsigned TypeIdx, LLT Ty) { 3925 Register Dst = MI.getOperand(0).getReg(); 3926 Register Src = MI.getOperand(1).getReg(); 3927 LLT DstTy = MRI.getType(Dst); 3928 LLT SrcTy = MRI.getType(Src); 3929 3930 const LLT S64 = LLT::scalar(64); 3931 const LLT S32 = LLT::scalar(32); 3932 const LLT S1 = LLT::scalar(1); 3933 3934 if (SrcTy == S1) { 3935 auto True = MIRBuilder.buildFConstant(DstTy, -1.0); 3936 auto False = MIRBuilder.buildFConstant(DstTy, 0.0); 3937 MIRBuilder.buildSelect(Dst, Src, True, False); 3938 MI.eraseFromParent(); 3939 return Legalized; 3940 } 3941 3942 if (SrcTy != S64) 3943 return UnableToLegalize; 3944 3945 if (DstTy == S32) { 3946 // signed cl2f(long l) { 3947 // long s = l >> 63; 3948 // float r = cul2f((l + s) ^ s); 3949 // return s ? -r : r; 3950 // } 3951 Register L = Src; 3952 auto SignBit = MIRBuilder.buildConstant(S64, 63); 3953 auto S = MIRBuilder.buildAShr(S64, L, SignBit); 3954 3955 auto LPlusS = MIRBuilder.buildAdd(S64, L, S); 3956 auto Xor = MIRBuilder.buildXor(S64, LPlusS, S); 3957 auto R = MIRBuilder.buildUITOFP(S32, Xor); 3958 3959 auto RNeg = MIRBuilder.buildFNeg(S32, R); 3960 auto SignNotZero = MIRBuilder.buildICmp(CmpInst::ICMP_NE, S1, S, 3961 MIRBuilder.buildConstant(S64, 0)); 3962 MIRBuilder.buildSelect(Dst, SignNotZero, RNeg, R); 3963 return Legalized; 3964 } 3965 3966 return UnableToLegalize; 3967 } 3968 3969 LegalizerHelper::LegalizeResult 3970 LegalizerHelper::lowerFPTOUI(MachineInstr &MI, unsigned TypeIdx, LLT Ty) { 3971 Register Dst = MI.getOperand(0).getReg(); 3972 Register Src = MI.getOperand(1).getReg(); 3973 LLT DstTy = MRI.getType(Dst); 3974 LLT SrcTy = MRI.getType(Src); 3975 const LLT S64 = LLT::scalar(64); 3976 const LLT S32 = LLT::scalar(32); 3977 3978 if (SrcTy != S64 && SrcTy != S32) 3979 return UnableToLegalize; 3980 if (DstTy != S32 && DstTy != S64) 3981 return UnableToLegalize; 3982 3983 // FPTOSI gives same result as FPTOUI for positive signed integers. 3984 // FPTOUI needs to deal with fp values that convert to unsigned integers 3985 // greater or equal to 2^31 for float or 2^63 for double. For brevity 2^Exp. 3986 3987 APInt TwoPExpInt = APInt::getSignMask(DstTy.getSizeInBits()); 3988 APFloat TwoPExpFP(SrcTy.getSizeInBits() == 32 ? APFloat::IEEEsingle() 3989 : APFloat::IEEEdouble(), 3990 APInt::getNullValue(SrcTy.getSizeInBits())); 3991 TwoPExpFP.convertFromAPInt(TwoPExpInt, false, APFloat::rmNearestTiesToEven); 3992 3993 MachineInstrBuilder FPTOSI = MIRBuilder.buildFPTOSI(DstTy, Src); 3994 3995 MachineInstrBuilder Threshold = MIRBuilder.buildFConstant(SrcTy, TwoPExpFP); 3996 // For fp Value greater or equal to Threshold(2^Exp), we use FPTOSI on 3997 // (Value - 2^Exp) and add 2^Exp by setting highest bit in result to 1. 3998 MachineInstrBuilder FSub = MIRBuilder.buildFSub(SrcTy, Src, Threshold); 3999 MachineInstrBuilder ResLowBits = MIRBuilder.buildFPTOSI(DstTy, FSub); 4000 MachineInstrBuilder ResHighBit = MIRBuilder.buildConstant(DstTy, TwoPExpInt); 4001 MachineInstrBuilder Res = MIRBuilder.buildXor(DstTy, ResLowBits, ResHighBit); 4002 4003 const LLT S1 = LLT::scalar(1); 4004 4005 MachineInstrBuilder FCMP = 4006 MIRBuilder.buildFCmp(CmpInst::FCMP_ULT, S1, Src, Threshold); 4007 MIRBuilder.buildSelect(Dst, FCMP, FPTOSI, Res); 4008 4009 MI.eraseFromParent(); 4010 return Legalized; 4011 } 4012 4013 static CmpInst::Predicate minMaxToCompare(unsigned Opc) { 4014 switch (Opc) { 4015 case TargetOpcode::G_SMIN: 4016 return CmpInst::ICMP_SLT; 4017 case TargetOpcode::G_SMAX: 4018 return CmpInst::ICMP_SGT; 4019 case TargetOpcode::G_UMIN: 4020 return CmpInst::ICMP_ULT; 4021 case TargetOpcode::G_UMAX: 4022 return CmpInst::ICMP_UGT; 4023 default: 4024 llvm_unreachable("not in integer min/max"); 4025 } 4026 } 4027 4028 LegalizerHelper::LegalizeResult 4029 LegalizerHelper::lowerMinMax(MachineInstr &MI, unsigned TypeIdx, LLT Ty) { 4030 Register Dst = MI.getOperand(0).getReg(); 4031 Register Src0 = MI.getOperand(1).getReg(); 4032 Register Src1 = MI.getOperand(2).getReg(); 4033 4034 const CmpInst::Predicate Pred = minMaxToCompare(MI.getOpcode()); 4035 LLT CmpType = MRI.getType(Dst).changeElementSize(1); 4036 4037 auto Cmp = MIRBuilder.buildICmp(Pred, CmpType, Src0, Src1); 4038 MIRBuilder.buildSelect(Dst, Cmp, Src0, Src1); 4039 4040 MI.eraseFromParent(); 4041 return Legalized; 4042 } 4043 4044 LegalizerHelper::LegalizeResult 4045 LegalizerHelper::lowerFCopySign(MachineInstr &MI, unsigned TypeIdx, LLT Ty) { 4046 Register Dst = MI.getOperand(0).getReg(); 4047 Register Src0 = MI.getOperand(1).getReg(); 4048 Register Src1 = MI.getOperand(2).getReg(); 4049 4050 const LLT Src0Ty = MRI.getType(Src0); 4051 const LLT Src1Ty = MRI.getType(Src1); 4052 4053 const int Src0Size = Src0Ty.getScalarSizeInBits(); 4054 const int Src1Size = Src1Ty.getScalarSizeInBits(); 4055 4056 auto SignBitMask = MIRBuilder.buildConstant( 4057 Src0Ty, APInt::getSignMask(Src0Size)); 4058 4059 auto NotSignBitMask = MIRBuilder.buildConstant( 4060 Src0Ty, APInt::getLowBitsSet(Src0Size, Src0Size - 1)); 4061 4062 auto And0 = MIRBuilder.buildAnd(Src0Ty, Src0, NotSignBitMask); 4063 MachineInstr *Or; 4064 4065 if (Src0Ty == Src1Ty) { 4066 auto And1 = MIRBuilder.buildAnd(Src1Ty, Src0, SignBitMask); 4067 Or = MIRBuilder.buildOr(Dst, And0, And1); 4068 } else if (Src0Size > Src1Size) { 4069 auto ShiftAmt = MIRBuilder.buildConstant(Src0Ty, Src0Size - Src1Size); 4070 auto Zext = MIRBuilder.buildZExt(Src0Ty, Src1); 4071 auto Shift = MIRBuilder.buildShl(Src0Ty, Zext, ShiftAmt); 4072 auto And1 = MIRBuilder.buildAnd(Src0Ty, Shift, SignBitMask); 4073 Or = MIRBuilder.buildOr(Dst, And0, And1); 4074 } else { 4075 auto ShiftAmt = MIRBuilder.buildConstant(Src1Ty, Src1Size - Src0Size); 4076 auto Shift = MIRBuilder.buildLShr(Src1Ty, Src1, ShiftAmt); 4077 auto Trunc = MIRBuilder.buildTrunc(Src0Ty, Shift); 4078 auto And1 = MIRBuilder.buildAnd(Src0Ty, Trunc, SignBitMask); 4079 Or = MIRBuilder.buildOr(Dst, And0, And1); 4080 } 4081 4082 // Be careful about setting nsz/nnan/ninf on every instruction, since the 4083 // constants are a nan and -0.0, but the final result should preserve 4084 // everything. 4085 if (unsigned Flags = MI.getFlags()) 4086 Or->setFlags(Flags); 4087 4088 MI.eraseFromParent(); 4089 return Legalized; 4090 } 4091 4092 LegalizerHelper::LegalizeResult 4093 LegalizerHelper::lowerFMinNumMaxNum(MachineInstr &MI) { 4094 unsigned NewOp = MI.getOpcode() == TargetOpcode::G_FMINNUM ? 4095 TargetOpcode::G_FMINNUM_IEEE : TargetOpcode::G_FMAXNUM_IEEE; 4096 4097 Register Dst = MI.getOperand(0).getReg(); 4098 Register Src0 = MI.getOperand(1).getReg(); 4099 Register Src1 = MI.getOperand(2).getReg(); 4100 LLT Ty = MRI.getType(Dst); 4101 4102 if (!MI.getFlag(MachineInstr::FmNoNans)) { 4103 // Insert canonicalizes if it's possible we need to quiet to get correct 4104 // sNaN behavior. 4105 4106 // Note this must be done here, and not as an optimization combine in the 4107 // absence of a dedicate quiet-snan instruction as we're using an 4108 // omni-purpose G_FCANONICALIZE. 4109 if (!isKnownNeverSNaN(Src0, MRI)) 4110 Src0 = MIRBuilder.buildFCanonicalize(Ty, Src0, MI.getFlags()).getReg(0); 4111 4112 if (!isKnownNeverSNaN(Src1, MRI)) 4113 Src1 = MIRBuilder.buildFCanonicalize(Ty, Src1, MI.getFlags()).getReg(0); 4114 } 4115 4116 // If there are no nans, it's safe to simply replace this with the non-IEEE 4117 // version. 4118 MIRBuilder.buildInstr(NewOp, {Dst}, {Src0, Src1}, MI.getFlags()); 4119 MI.eraseFromParent(); 4120 return Legalized; 4121 } 4122 4123 LegalizerHelper::LegalizeResult LegalizerHelper::lowerFMad(MachineInstr &MI) { 4124 // Expand G_FMAD a, b, c -> G_FADD (G_FMUL a, b), c 4125 Register DstReg = MI.getOperand(0).getReg(); 4126 LLT Ty = MRI.getType(DstReg); 4127 unsigned Flags = MI.getFlags(); 4128 4129 auto Mul = MIRBuilder.buildFMul(Ty, MI.getOperand(1), MI.getOperand(2), 4130 Flags); 4131 MIRBuilder.buildFAdd(DstReg, Mul, MI.getOperand(3), Flags); 4132 MI.eraseFromParent(); 4133 return Legalized; 4134 } 4135 4136 LegalizerHelper::LegalizeResult 4137 LegalizerHelper::lowerIntrinsicRound(MachineInstr &MI) { 4138 Register DstReg = MI.getOperand(0).getReg(); 4139 Register SrcReg = MI.getOperand(1).getReg(); 4140 unsigned Flags = MI.getFlags(); 4141 LLT Ty = MRI.getType(DstReg); 4142 const LLT CondTy = Ty.changeElementSize(1); 4143 4144 // result = trunc(src); 4145 // if (src < 0.0 && src != result) 4146 // result += -1.0. 4147 4148 auto Zero = MIRBuilder.buildFConstant(Ty, 0.0); 4149 auto Trunc = MIRBuilder.buildIntrinsicTrunc(Ty, SrcReg, Flags); 4150 4151 auto Lt0 = MIRBuilder.buildFCmp(CmpInst::FCMP_OLT, CondTy, 4152 SrcReg, Zero, Flags); 4153 auto NeTrunc = MIRBuilder.buildFCmp(CmpInst::FCMP_ONE, CondTy, 4154 SrcReg, Trunc, Flags); 4155 auto And = MIRBuilder.buildAnd(CondTy, Lt0, NeTrunc); 4156 auto AddVal = MIRBuilder.buildSITOFP(Ty, And); 4157 4158 MIRBuilder.buildFAdd(DstReg, Trunc, AddVal); 4159 MI.eraseFromParent(); 4160 return Legalized; 4161 } 4162 4163 LegalizerHelper::LegalizeResult 4164 LegalizerHelper::lowerUnmergeValues(MachineInstr &MI) { 4165 const unsigned NumDst = MI.getNumOperands() - 1; 4166 const Register SrcReg = MI.getOperand(NumDst).getReg(); 4167 LLT SrcTy = MRI.getType(SrcReg); 4168 4169 Register Dst0Reg = MI.getOperand(0).getReg(); 4170 LLT DstTy = MRI.getType(Dst0Reg); 4171 4172 4173 // Expand scalarizing unmerge as bitcast to integer and shift. 4174 if (!DstTy.isVector() && SrcTy.isVector() && 4175 SrcTy.getElementType() == DstTy) { 4176 LLT IntTy = LLT::scalar(SrcTy.getSizeInBits()); 4177 Register Cast = MIRBuilder.buildBitcast(IntTy, SrcReg).getReg(0); 4178 4179 MIRBuilder.buildTrunc(Dst0Reg, Cast); 4180 4181 const unsigned DstSize = DstTy.getSizeInBits(); 4182 unsigned Offset = DstSize; 4183 for (unsigned I = 1; I != NumDst; ++I, Offset += DstSize) { 4184 auto ShiftAmt = MIRBuilder.buildConstant(IntTy, Offset); 4185 auto Shift = MIRBuilder.buildLShr(IntTy, Cast, ShiftAmt); 4186 MIRBuilder.buildTrunc(MI.getOperand(I), Shift); 4187 } 4188 4189 MI.eraseFromParent(); 4190 return Legalized; 4191 } 4192 4193 return UnableToLegalize; 4194 } 4195 4196 LegalizerHelper::LegalizeResult 4197 LegalizerHelper::lowerShuffleVector(MachineInstr &MI) { 4198 Register DstReg = MI.getOperand(0).getReg(); 4199 Register Src0Reg = MI.getOperand(1).getReg(); 4200 Register Src1Reg = MI.getOperand(2).getReg(); 4201 LLT Src0Ty = MRI.getType(Src0Reg); 4202 LLT DstTy = MRI.getType(DstReg); 4203 LLT IdxTy = LLT::scalar(32); 4204 4205 ArrayRef<int> Mask = MI.getOperand(3).getShuffleMask(); 4206 4207 if (DstTy.isScalar()) { 4208 if (Src0Ty.isVector()) 4209 return UnableToLegalize; 4210 4211 // This is just a SELECT. 4212 assert(Mask.size() == 1 && "Expected a single mask element"); 4213 Register Val; 4214 if (Mask[0] < 0 || Mask[0] > 1) 4215 Val = MIRBuilder.buildUndef(DstTy).getReg(0); 4216 else 4217 Val = Mask[0] == 0 ? Src0Reg : Src1Reg; 4218 MIRBuilder.buildCopy(DstReg, Val); 4219 MI.eraseFromParent(); 4220 return Legalized; 4221 } 4222 4223 Register Undef; 4224 SmallVector<Register, 32> BuildVec; 4225 LLT EltTy = DstTy.getElementType(); 4226 4227 for (int Idx : Mask) { 4228 if (Idx < 0) { 4229 if (!Undef.isValid()) 4230 Undef = MIRBuilder.buildUndef(EltTy).getReg(0); 4231 BuildVec.push_back(Undef); 4232 continue; 4233 } 4234 4235 if (Src0Ty.isScalar()) { 4236 BuildVec.push_back(Idx == 0 ? Src0Reg : Src1Reg); 4237 } else { 4238 int NumElts = Src0Ty.getNumElements(); 4239 Register SrcVec = Idx < NumElts ? Src0Reg : Src1Reg; 4240 int ExtractIdx = Idx < NumElts ? Idx : Idx - NumElts; 4241 auto IdxK = MIRBuilder.buildConstant(IdxTy, ExtractIdx); 4242 auto Extract = MIRBuilder.buildExtractVectorElement(EltTy, SrcVec, IdxK); 4243 BuildVec.push_back(Extract.getReg(0)); 4244 } 4245 } 4246 4247 MIRBuilder.buildBuildVector(DstReg, BuildVec); 4248 MI.eraseFromParent(); 4249 return Legalized; 4250 } 4251 4252 LegalizerHelper::LegalizeResult 4253 LegalizerHelper::lowerDynStackAlloc(MachineInstr &MI) { 4254 Register Dst = MI.getOperand(0).getReg(); 4255 Register AllocSize = MI.getOperand(1).getReg(); 4256 unsigned Align = MI.getOperand(2).getImm(); 4257 4258 const auto &MF = *MI.getMF(); 4259 const auto &TLI = *MF.getSubtarget().getTargetLowering(); 4260 4261 LLT PtrTy = MRI.getType(Dst); 4262 LLT IntPtrTy = LLT::scalar(PtrTy.getSizeInBits()); 4263 4264 Register SPReg = TLI.getStackPointerRegisterToSaveRestore(); 4265 auto SPTmp = MIRBuilder.buildCopy(PtrTy, SPReg); 4266 SPTmp = MIRBuilder.buildCast(IntPtrTy, SPTmp); 4267 4268 // Subtract the final alloc from the SP. We use G_PTRTOINT here so we don't 4269 // have to generate an extra instruction to negate the alloc and then use 4270 // G_PTR_ADD to add the negative offset. 4271 auto Alloc = MIRBuilder.buildSub(IntPtrTy, SPTmp, AllocSize); 4272 if (Align) { 4273 APInt AlignMask(IntPtrTy.getSizeInBits(), Align, true); 4274 AlignMask.negate(); 4275 auto AlignCst = MIRBuilder.buildConstant(IntPtrTy, AlignMask); 4276 Alloc = MIRBuilder.buildAnd(IntPtrTy, Alloc, AlignCst); 4277 } 4278 4279 SPTmp = MIRBuilder.buildCast(PtrTy, Alloc); 4280 MIRBuilder.buildCopy(SPReg, SPTmp); 4281 MIRBuilder.buildCopy(Dst, SPTmp); 4282 4283 MI.eraseFromParent(); 4284 return Legalized; 4285 } 4286 4287 LegalizerHelper::LegalizeResult 4288 LegalizerHelper::lowerExtract(MachineInstr &MI) { 4289 Register Dst = MI.getOperand(0).getReg(); 4290 Register Src = MI.getOperand(1).getReg(); 4291 unsigned Offset = MI.getOperand(2).getImm(); 4292 4293 LLT DstTy = MRI.getType(Dst); 4294 LLT SrcTy = MRI.getType(Src); 4295 4296 if (DstTy.isScalar() && 4297 (SrcTy.isScalar() || 4298 (SrcTy.isVector() && DstTy == SrcTy.getElementType()))) { 4299 LLT SrcIntTy = SrcTy; 4300 if (!SrcTy.isScalar()) { 4301 SrcIntTy = LLT::scalar(SrcTy.getSizeInBits()); 4302 Src = MIRBuilder.buildBitcast(SrcIntTy, Src).getReg(0); 4303 } 4304 4305 if (Offset == 0) 4306 MIRBuilder.buildTrunc(Dst, Src); 4307 else { 4308 auto ShiftAmt = MIRBuilder.buildConstant(SrcIntTy, Offset); 4309 auto Shr = MIRBuilder.buildLShr(SrcIntTy, Src, ShiftAmt); 4310 MIRBuilder.buildTrunc(Dst, Shr); 4311 } 4312 4313 MI.eraseFromParent(); 4314 return Legalized; 4315 } 4316 4317 return UnableToLegalize; 4318 } 4319 4320 LegalizerHelper::LegalizeResult LegalizerHelper::lowerInsert(MachineInstr &MI) { 4321 Register Dst = MI.getOperand(0).getReg(); 4322 Register Src = MI.getOperand(1).getReg(); 4323 Register InsertSrc = MI.getOperand(2).getReg(); 4324 uint64_t Offset = MI.getOperand(3).getImm(); 4325 4326 LLT DstTy = MRI.getType(Src); 4327 LLT InsertTy = MRI.getType(InsertSrc); 4328 4329 if (InsertTy.isScalar() && 4330 (DstTy.isScalar() || 4331 (DstTy.isVector() && DstTy.getElementType() == InsertTy))) { 4332 LLT IntDstTy = DstTy; 4333 if (!DstTy.isScalar()) { 4334 IntDstTy = LLT::scalar(DstTy.getSizeInBits()); 4335 Src = MIRBuilder.buildBitcast(IntDstTy, Src).getReg(0); 4336 } 4337 4338 Register ExtInsSrc = MIRBuilder.buildZExt(IntDstTy, InsertSrc).getReg(0); 4339 if (Offset != 0) { 4340 auto ShiftAmt = MIRBuilder.buildConstant(IntDstTy, Offset); 4341 ExtInsSrc = MIRBuilder.buildShl(IntDstTy, ExtInsSrc, ShiftAmt).getReg(0); 4342 } 4343 4344 APInt MaskVal = ~APInt::getBitsSet(DstTy.getSizeInBits(), Offset, 4345 InsertTy.getSizeInBits()); 4346 4347 auto Mask = MIRBuilder.buildConstant(IntDstTy, MaskVal); 4348 auto MaskedSrc = MIRBuilder.buildAnd(IntDstTy, Src, Mask); 4349 auto Or = MIRBuilder.buildOr(IntDstTy, MaskedSrc, ExtInsSrc); 4350 4351 MIRBuilder.buildBitcast(Dst, Or); 4352 MI.eraseFromParent(); 4353 return Legalized; 4354 } 4355 4356 return UnableToLegalize; 4357 } 4358 4359 LegalizerHelper::LegalizeResult 4360 LegalizerHelper::lowerSADDO_SSUBO(MachineInstr &MI) { 4361 Register Dst0 = MI.getOperand(0).getReg(); 4362 Register Dst1 = MI.getOperand(1).getReg(); 4363 Register LHS = MI.getOperand(2).getReg(); 4364 Register RHS = MI.getOperand(3).getReg(); 4365 const bool IsAdd = MI.getOpcode() == TargetOpcode::G_SADDO; 4366 4367 LLT Ty = MRI.getType(Dst0); 4368 LLT BoolTy = MRI.getType(Dst1); 4369 4370 if (IsAdd) 4371 MIRBuilder.buildAdd(Dst0, LHS, RHS); 4372 else 4373 MIRBuilder.buildSub(Dst0, LHS, RHS); 4374 4375 // TODO: If SADDSAT/SSUBSAT is legal, compare results to detect overflow. 4376 4377 auto Zero = MIRBuilder.buildConstant(Ty, 0); 4378 4379 // For an addition, the result should be less than one of the operands (LHS) 4380 // if and only if the other operand (RHS) is negative, otherwise there will 4381 // be overflow. 4382 // For a subtraction, the result should be less than one of the operands 4383 // (LHS) if and only if the other operand (RHS) is (non-zero) positive, 4384 // otherwise there will be overflow. 4385 auto ResultLowerThanLHS = 4386 MIRBuilder.buildICmp(CmpInst::ICMP_SLT, BoolTy, Dst0, LHS); 4387 auto ConditionRHS = MIRBuilder.buildICmp( 4388 IsAdd ? CmpInst::ICMP_SLT : CmpInst::ICMP_SGT, BoolTy, RHS, Zero); 4389 4390 MIRBuilder.buildXor(Dst1, ConditionRHS, ResultLowerThanLHS); 4391 MI.eraseFromParent(); 4392 return Legalized; 4393 } 4394 4395 LegalizerHelper::LegalizeResult 4396 LegalizerHelper::lowerBswap(MachineInstr &MI) { 4397 Register Dst = MI.getOperand(0).getReg(); 4398 Register Src = MI.getOperand(1).getReg(); 4399 const LLT Ty = MRI.getType(Src); 4400 unsigned SizeInBytes = Ty.getSizeInBytes(); 4401 unsigned BaseShiftAmt = (SizeInBytes - 1) * 8; 4402 4403 // Swap most and least significant byte, set remaining bytes in Res to zero. 4404 auto ShiftAmt = MIRBuilder.buildConstant(Ty, BaseShiftAmt); 4405 auto LSByteShiftedLeft = MIRBuilder.buildShl(Ty, Src, ShiftAmt); 4406 auto MSByteShiftedRight = MIRBuilder.buildLShr(Ty, Src, ShiftAmt); 4407 auto Res = MIRBuilder.buildOr(Ty, MSByteShiftedRight, LSByteShiftedLeft); 4408 4409 // Set i-th high/low byte in Res to i-th low/high byte from Src. 4410 for (unsigned i = 1; i < SizeInBytes / 2; ++i) { 4411 // AND with Mask leaves byte i unchanged and sets remaining bytes to 0. 4412 APInt APMask(SizeInBytes * 8, 0xFF << (i * 8)); 4413 auto Mask = MIRBuilder.buildConstant(Ty, APMask); 4414 auto ShiftAmt = MIRBuilder.buildConstant(Ty, BaseShiftAmt - 16 * i); 4415 // Low byte shifted left to place of high byte: (Src & Mask) << ShiftAmt. 4416 auto LoByte = MIRBuilder.buildAnd(Ty, Src, Mask); 4417 auto LoShiftedLeft = MIRBuilder.buildShl(Ty, LoByte, ShiftAmt); 4418 Res = MIRBuilder.buildOr(Ty, Res, LoShiftedLeft); 4419 // High byte shifted right to place of low byte: (Src >> ShiftAmt) & Mask. 4420 auto SrcShiftedRight = MIRBuilder.buildLShr(Ty, Src, ShiftAmt); 4421 auto HiShiftedRight = MIRBuilder.buildAnd(Ty, SrcShiftedRight, Mask); 4422 Res = MIRBuilder.buildOr(Ty, Res, HiShiftedRight); 4423 } 4424 Res.getInstr()->getOperand(0).setReg(Dst); 4425 4426 MI.eraseFromParent(); 4427 return Legalized; 4428 } 4429 4430 //{ (Src & Mask) >> N } | { (Src << N) & Mask } 4431 static MachineInstrBuilder SwapN(unsigned N, DstOp Dst, MachineIRBuilder &B, 4432 MachineInstrBuilder Src, APInt Mask) { 4433 const LLT Ty = Dst.getLLTTy(*B.getMRI()); 4434 MachineInstrBuilder C_N = B.buildConstant(Ty, N); 4435 MachineInstrBuilder MaskLoNTo0 = B.buildConstant(Ty, Mask); 4436 auto LHS = B.buildLShr(Ty, B.buildAnd(Ty, Src, MaskLoNTo0), C_N); 4437 auto RHS = B.buildAnd(Ty, B.buildShl(Ty, Src, C_N), MaskLoNTo0); 4438 return B.buildOr(Dst, LHS, RHS); 4439 } 4440 4441 LegalizerHelper::LegalizeResult 4442 LegalizerHelper::lowerBitreverse(MachineInstr &MI) { 4443 Register Dst = MI.getOperand(0).getReg(); 4444 Register Src = MI.getOperand(1).getReg(); 4445 const LLT Ty = MRI.getType(Src); 4446 unsigned Size = Ty.getSizeInBits(); 4447 4448 MachineInstrBuilder BSWAP = 4449 MIRBuilder.buildInstr(TargetOpcode::G_BSWAP, {Ty}, {Src}); 4450 4451 // swap high and low 4 bits in 8 bit blocks 7654|3210 -> 3210|7654 4452 // [(val & 0xF0F0F0F0) >> 4] | [(val & 0x0F0F0F0F) << 4] 4453 // -> [(val & 0xF0F0F0F0) >> 4] | [(val << 4) & 0xF0F0F0F0] 4454 MachineInstrBuilder Swap4 = 4455 SwapN(4, Ty, MIRBuilder, BSWAP, APInt::getSplat(Size, APInt(8, 0xF0))); 4456 4457 // swap high and low 2 bits in 4 bit blocks 32|10 76|54 -> 10|32 54|76 4458 // [(val & 0xCCCCCCCC) >> 2] & [(val & 0x33333333) << 2] 4459 // -> [(val & 0xCCCCCCCC) >> 2] & [(val << 2) & 0xCCCCCCCC] 4460 MachineInstrBuilder Swap2 = 4461 SwapN(2, Ty, MIRBuilder, Swap4, APInt::getSplat(Size, APInt(8, 0xCC))); 4462 4463 // swap high and low 1 bit in 2 bit blocks 1|0 3|2 5|4 7|6 -> 0|1 2|3 4|5 6|7 4464 // [(val & 0xAAAAAAAA) >> 1] & [(val & 0x55555555) << 1] 4465 // -> [(val & 0xAAAAAAAA) >> 1] & [(val << 1) & 0xAAAAAAAA] 4466 SwapN(1, Dst, MIRBuilder, Swap2, APInt::getSplat(Size, APInt(8, 0xAA))); 4467 4468 MI.eraseFromParent(); 4469 return Legalized; 4470 } 4471 4472 LegalizerHelper::LegalizeResult 4473 LegalizerHelper::lowerReadRegister(MachineInstr &MI) { 4474 Register Dst = MI.getOperand(0).getReg(); 4475 const LLT Ty = MRI.getType(Dst); 4476 const MDString *RegStr = cast<MDString>( 4477 cast<MDNode>(MI.getOperand(1).getMetadata())->getOperand(0)); 4478 4479 MachineFunction &MF = MIRBuilder.getMF(); 4480 const TargetSubtargetInfo &STI = MF.getSubtarget(); 4481 const TargetLowering *TLI = STI.getTargetLowering(); 4482 Register Reg = TLI->getRegisterByName(RegStr->getString().data(), Ty, MF); 4483 if (!Reg.isValid()) 4484 return UnableToLegalize; 4485 4486 MIRBuilder.buildCopy(Dst, Reg); 4487 MI.eraseFromParent(); 4488 return Legalized; 4489 } 4490