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/GenericMachineInstrs.h" 19 #include "llvm/CodeGen/GlobalISel/LegalizerInfo.h" 20 #include "llvm/CodeGen/GlobalISel/LostDebugLocObserver.h" 21 #include "llvm/CodeGen/GlobalISel/MIPatternMatch.h" 22 #include "llvm/CodeGen/GlobalISel/MachineIRBuilder.h" 23 #include "llvm/CodeGen/GlobalISel/Utils.h" 24 #include "llvm/CodeGen/MachineFrameInfo.h" 25 #include "llvm/CodeGen/MachineRegisterInfo.h" 26 #include "llvm/CodeGen/TargetFrameLowering.h" 27 #include "llvm/CodeGen/TargetInstrInfo.h" 28 #include "llvm/CodeGen/TargetLowering.h" 29 #include "llvm/CodeGen/TargetOpcodes.h" 30 #include "llvm/CodeGen/TargetSubtargetInfo.h" 31 #include "llvm/IR/Instructions.h" 32 #include "llvm/Support/Debug.h" 33 #include "llvm/Support/MathExtras.h" 34 #include "llvm/Support/raw_ostream.h" 35 #include "llvm/Target/TargetMachine.h" 36 #include <numeric> 37 #include <optional> 38 39 #define DEBUG_TYPE "legalizer" 40 41 using namespace llvm; 42 using namespace LegalizeActions; 43 using namespace MIPatternMatch; 44 45 /// Try to break down \p OrigTy into \p NarrowTy sized pieces. 46 /// 47 /// Returns the number of \p NarrowTy elements needed to reconstruct \p OrigTy, 48 /// with any leftover piece as type \p LeftoverTy 49 /// 50 /// Returns -1 in the first element of the pair if the breakdown is not 51 /// satisfiable. 52 static std::pair<int, int> 53 getNarrowTypeBreakDown(LLT OrigTy, LLT NarrowTy, LLT &LeftoverTy) { 54 assert(!LeftoverTy.isValid() && "this is an out argument"); 55 56 unsigned Size = OrigTy.getSizeInBits(); 57 unsigned NarrowSize = NarrowTy.getSizeInBits(); 58 unsigned NumParts = Size / NarrowSize; 59 unsigned LeftoverSize = Size - NumParts * NarrowSize; 60 assert(Size > NarrowSize); 61 62 if (LeftoverSize == 0) 63 return {NumParts, 0}; 64 65 if (NarrowTy.isVector()) { 66 unsigned EltSize = OrigTy.getScalarSizeInBits(); 67 if (LeftoverSize % EltSize != 0) 68 return {-1, -1}; 69 LeftoverTy = LLT::scalarOrVector( 70 ElementCount::getFixed(LeftoverSize / EltSize), EltSize); 71 } else { 72 LeftoverTy = LLT::scalar(LeftoverSize); 73 } 74 75 int NumLeftover = LeftoverSize / LeftoverTy.getSizeInBits(); 76 return std::make_pair(NumParts, NumLeftover); 77 } 78 79 static Type *getFloatTypeForLLT(LLVMContext &Ctx, LLT Ty) { 80 81 if (!Ty.isScalar()) 82 return nullptr; 83 84 switch (Ty.getSizeInBits()) { 85 case 16: 86 return Type::getHalfTy(Ctx); 87 case 32: 88 return Type::getFloatTy(Ctx); 89 case 64: 90 return Type::getDoubleTy(Ctx); 91 case 80: 92 return Type::getX86_FP80Ty(Ctx); 93 case 128: 94 return Type::getFP128Ty(Ctx); 95 default: 96 return nullptr; 97 } 98 } 99 100 LegalizerHelper::LegalizerHelper(MachineFunction &MF, 101 GISelChangeObserver &Observer, 102 MachineIRBuilder &Builder) 103 : MIRBuilder(Builder), Observer(Observer), MRI(MF.getRegInfo()), 104 LI(*MF.getSubtarget().getLegalizerInfo()), 105 TLI(*MF.getSubtarget().getTargetLowering()) { } 106 107 LegalizerHelper::LegalizerHelper(MachineFunction &MF, const LegalizerInfo &LI, 108 GISelChangeObserver &Observer, 109 MachineIRBuilder &B) 110 : MIRBuilder(B), Observer(Observer), MRI(MF.getRegInfo()), LI(LI), 111 TLI(*MF.getSubtarget().getTargetLowering()) { } 112 113 LegalizerHelper::LegalizeResult 114 LegalizerHelper::legalizeInstrStep(MachineInstr &MI, 115 LostDebugLocObserver &LocObserver) { 116 LLVM_DEBUG(dbgs() << "Legalizing: " << MI); 117 118 MIRBuilder.setInstrAndDebugLoc(MI); 119 120 if (MI.getOpcode() == TargetOpcode::G_INTRINSIC || 121 MI.getOpcode() == TargetOpcode::G_INTRINSIC_W_SIDE_EFFECTS) 122 return LI.legalizeIntrinsic(*this, MI) ? Legalized : UnableToLegalize; 123 auto Step = LI.getAction(MI, MRI); 124 switch (Step.Action) { 125 case Legal: 126 LLVM_DEBUG(dbgs() << ".. Already legal\n"); 127 return AlreadyLegal; 128 case Libcall: 129 LLVM_DEBUG(dbgs() << ".. Convert to libcall\n"); 130 return libcall(MI, LocObserver); 131 case NarrowScalar: 132 LLVM_DEBUG(dbgs() << ".. Narrow scalar\n"); 133 return narrowScalar(MI, Step.TypeIdx, Step.NewType); 134 case WidenScalar: 135 LLVM_DEBUG(dbgs() << ".. Widen scalar\n"); 136 return widenScalar(MI, Step.TypeIdx, Step.NewType); 137 case Bitcast: 138 LLVM_DEBUG(dbgs() << ".. Bitcast type\n"); 139 return bitcast(MI, Step.TypeIdx, Step.NewType); 140 case Lower: 141 LLVM_DEBUG(dbgs() << ".. Lower\n"); 142 return lower(MI, Step.TypeIdx, Step.NewType); 143 case FewerElements: 144 LLVM_DEBUG(dbgs() << ".. Reduce number of elements\n"); 145 return fewerElementsVector(MI, Step.TypeIdx, Step.NewType); 146 case MoreElements: 147 LLVM_DEBUG(dbgs() << ".. Increase number of elements\n"); 148 return moreElementsVector(MI, Step.TypeIdx, Step.NewType); 149 case Custom: 150 LLVM_DEBUG(dbgs() << ".. Custom legalization\n"); 151 return LI.legalizeCustom(*this, MI) ? Legalized : UnableToLegalize; 152 default: 153 LLVM_DEBUG(dbgs() << ".. Unable to legalize\n"); 154 return UnableToLegalize; 155 } 156 } 157 158 void LegalizerHelper::extractParts(Register Reg, LLT Ty, int NumParts, 159 SmallVectorImpl<Register> &VRegs) { 160 for (int i = 0; i < NumParts; ++i) 161 VRegs.push_back(MRI.createGenericVirtualRegister(Ty)); 162 MIRBuilder.buildUnmerge(VRegs, Reg); 163 } 164 165 bool LegalizerHelper::extractParts(Register Reg, LLT RegTy, 166 LLT MainTy, LLT &LeftoverTy, 167 SmallVectorImpl<Register> &VRegs, 168 SmallVectorImpl<Register> &LeftoverRegs) { 169 assert(!LeftoverTy.isValid() && "this is an out argument"); 170 171 unsigned RegSize = RegTy.getSizeInBits(); 172 unsigned MainSize = MainTy.getSizeInBits(); 173 unsigned NumParts = RegSize / MainSize; 174 unsigned LeftoverSize = RegSize - NumParts * MainSize; 175 176 // Use an unmerge when possible. 177 if (LeftoverSize == 0) { 178 for (unsigned I = 0; I < NumParts; ++I) 179 VRegs.push_back(MRI.createGenericVirtualRegister(MainTy)); 180 MIRBuilder.buildUnmerge(VRegs, Reg); 181 return true; 182 } 183 184 // Perform irregular split. Leftover is last element of RegPieces. 185 if (MainTy.isVector()) { 186 SmallVector<Register, 8> RegPieces; 187 extractVectorParts(Reg, MainTy.getNumElements(), RegPieces); 188 for (unsigned i = 0; i < RegPieces.size() - 1; ++i) 189 VRegs.push_back(RegPieces[i]); 190 LeftoverRegs.push_back(RegPieces[RegPieces.size() - 1]); 191 LeftoverTy = MRI.getType(LeftoverRegs[0]); 192 return true; 193 } 194 195 LeftoverTy = LLT::scalar(LeftoverSize); 196 // For irregular sizes, extract the individual parts. 197 for (unsigned I = 0; I != NumParts; ++I) { 198 Register NewReg = MRI.createGenericVirtualRegister(MainTy); 199 VRegs.push_back(NewReg); 200 MIRBuilder.buildExtract(NewReg, Reg, MainSize * I); 201 } 202 203 for (unsigned Offset = MainSize * NumParts; Offset < RegSize; 204 Offset += LeftoverSize) { 205 Register NewReg = MRI.createGenericVirtualRegister(LeftoverTy); 206 LeftoverRegs.push_back(NewReg); 207 MIRBuilder.buildExtract(NewReg, Reg, Offset); 208 } 209 210 return true; 211 } 212 213 void LegalizerHelper::extractVectorParts(Register Reg, unsigned NumElts, 214 SmallVectorImpl<Register> &VRegs) { 215 LLT RegTy = MRI.getType(Reg); 216 assert(RegTy.isVector() && "Expected a vector type"); 217 218 LLT EltTy = RegTy.getElementType(); 219 LLT NarrowTy = (NumElts == 1) ? EltTy : LLT::fixed_vector(NumElts, EltTy); 220 unsigned RegNumElts = RegTy.getNumElements(); 221 unsigned LeftoverNumElts = RegNumElts % NumElts; 222 unsigned NumNarrowTyPieces = RegNumElts / NumElts; 223 224 // Perfect split without leftover 225 if (LeftoverNumElts == 0) 226 return extractParts(Reg, NarrowTy, NumNarrowTyPieces, VRegs); 227 228 // Irregular split. Provide direct access to all elements for artifact 229 // combiner using unmerge to elements. Then build vectors with NumElts 230 // elements. Remaining element(s) will be (used to build vector) Leftover. 231 SmallVector<Register, 8> Elts; 232 extractParts(Reg, EltTy, RegNumElts, Elts); 233 234 unsigned Offset = 0; 235 // Requested sub-vectors of NarrowTy. 236 for (unsigned i = 0; i < NumNarrowTyPieces; ++i, Offset += NumElts) { 237 ArrayRef<Register> Pieces(&Elts[Offset], NumElts); 238 VRegs.push_back(MIRBuilder.buildMergeLikeInstr(NarrowTy, Pieces).getReg(0)); 239 } 240 241 // Leftover element(s). 242 if (LeftoverNumElts == 1) { 243 VRegs.push_back(Elts[Offset]); 244 } else { 245 LLT LeftoverTy = LLT::fixed_vector(LeftoverNumElts, EltTy); 246 ArrayRef<Register> Pieces(&Elts[Offset], LeftoverNumElts); 247 VRegs.push_back( 248 MIRBuilder.buildMergeLikeInstr(LeftoverTy, Pieces).getReg(0)); 249 } 250 } 251 252 void LegalizerHelper::insertParts(Register DstReg, 253 LLT ResultTy, LLT PartTy, 254 ArrayRef<Register> PartRegs, 255 LLT LeftoverTy, 256 ArrayRef<Register> LeftoverRegs) { 257 if (!LeftoverTy.isValid()) { 258 assert(LeftoverRegs.empty()); 259 260 if (!ResultTy.isVector()) { 261 MIRBuilder.buildMergeLikeInstr(DstReg, PartRegs); 262 return; 263 } 264 265 if (PartTy.isVector()) 266 MIRBuilder.buildConcatVectors(DstReg, PartRegs); 267 else 268 MIRBuilder.buildBuildVector(DstReg, PartRegs); 269 return; 270 } 271 272 // Merge sub-vectors with different number of elements and insert into DstReg. 273 if (ResultTy.isVector()) { 274 assert(LeftoverRegs.size() == 1 && "Expected one leftover register"); 275 SmallVector<Register, 8> AllRegs; 276 for (auto Reg : concat<const Register>(PartRegs, LeftoverRegs)) 277 AllRegs.push_back(Reg); 278 return mergeMixedSubvectors(DstReg, AllRegs); 279 } 280 281 SmallVector<Register> GCDRegs; 282 LLT GCDTy = getGCDType(getGCDType(ResultTy, LeftoverTy), PartTy); 283 for (auto PartReg : concat<const Register>(PartRegs, LeftoverRegs)) 284 extractGCDType(GCDRegs, GCDTy, PartReg); 285 LLT ResultLCMTy = buildLCMMergePieces(ResultTy, LeftoverTy, GCDTy, GCDRegs); 286 buildWidenedRemergeToDst(DstReg, ResultLCMTy, GCDRegs); 287 } 288 289 void LegalizerHelper::appendVectorElts(SmallVectorImpl<Register> &Elts, 290 Register Reg) { 291 LLT Ty = MRI.getType(Reg); 292 SmallVector<Register, 8> RegElts; 293 extractParts(Reg, Ty.getScalarType(), Ty.getNumElements(), RegElts); 294 Elts.append(RegElts); 295 } 296 297 /// Merge \p PartRegs with different types into \p DstReg. 298 void LegalizerHelper::mergeMixedSubvectors(Register DstReg, 299 ArrayRef<Register> PartRegs) { 300 SmallVector<Register, 8> AllElts; 301 for (unsigned i = 0; i < PartRegs.size() - 1; ++i) 302 appendVectorElts(AllElts, PartRegs[i]); 303 304 Register Leftover = PartRegs[PartRegs.size() - 1]; 305 if (MRI.getType(Leftover).isScalar()) 306 AllElts.push_back(Leftover); 307 else 308 appendVectorElts(AllElts, Leftover); 309 310 MIRBuilder.buildMergeLikeInstr(DstReg, AllElts); 311 } 312 313 /// Append the result registers of G_UNMERGE_VALUES \p MI to \p Regs. 314 static void getUnmergeResults(SmallVectorImpl<Register> &Regs, 315 const MachineInstr &MI) { 316 assert(MI.getOpcode() == TargetOpcode::G_UNMERGE_VALUES); 317 318 const int StartIdx = Regs.size(); 319 const int NumResults = MI.getNumOperands() - 1; 320 Regs.resize(Regs.size() + NumResults); 321 for (int I = 0; I != NumResults; ++I) 322 Regs[StartIdx + I] = MI.getOperand(I).getReg(); 323 } 324 325 void LegalizerHelper::extractGCDType(SmallVectorImpl<Register> &Parts, 326 LLT GCDTy, Register SrcReg) { 327 LLT SrcTy = MRI.getType(SrcReg); 328 if (SrcTy == GCDTy) { 329 // If the source already evenly divides the result type, we don't need to do 330 // anything. 331 Parts.push_back(SrcReg); 332 } else { 333 // Need to split into common type sized pieces. 334 auto Unmerge = MIRBuilder.buildUnmerge(GCDTy, SrcReg); 335 getUnmergeResults(Parts, *Unmerge); 336 } 337 } 338 339 LLT LegalizerHelper::extractGCDType(SmallVectorImpl<Register> &Parts, LLT DstTy, 340 LLT NarrowTy, Register SrcReg) { 341 LLT SrcTy = MRI.getType(SrcReg); 342 LLT GCDTy = getGCDType(getGCDType(SrcTy, NarrowTy), DstTy); 343 extractGCDType(Parts, GCDTy, SrcReg); 344 return GCDTy; 345 } 346 347 LLT LegalizerHelper::buildLCMMergePieces(LLT DstTy, LLT NarrowTy, LLT GCDTy, 348 SmallVectorImpl<Register> &VRegs, 349 unsigned PadStrategy) { 350 LLT LCMTy = getLCMType(DstTy, NarrowTy); 351 352 int NumParts = LCMTy.getSizeInBits() / NarrowTy.getSizeInBits(); 353 int NumSubParts = NarrowTy.getSizeInBits() / GCDTy.getSizeInBits(); 354 int NumOrigSrc = VRegs.size(); 355 356 Register PadReg; 357 358 // Get a value we can use to pad the source value if the sources won't evenly 359 // cover the result type. 360 if (NumOrigSrc < NumParts * NumSubParts) { 361 if (PadStrategy == TargetOpcode::G_ZEXT) 362 PadReg = MIRBuilder.buildConstant(GCDTy, 0).getReg(0); 363 else if (PadStrategy == TargetOpcode::G_ANYEXT) 364 PadReg = MIRBuilder.buildUndef(GCDTy).getReg(0); 365 else { 366 assert(PadStrategy == TargetOpcode::G_SEXT); 367 368 // Shift the sign bit of the low register through the high register. 369 auto ShiftAmt = 370 MIRBuilder.buildConstant(LLT::scalar(64), GCDTy.getSizeInBits() - 1); 371 PadReg = MIRBuilder.buildAShr(GCDTy, VRegs.back(), ShiftAmt).getReg(0); 372 } 373 } 374 375 // Registers for the final merge to be produced. 376 SmallVector<Register, 4> Remerge(NumParts); 377 378 // Registers needed for intermediate merges, which will be merged into a 379 // source for Remerge. 380 SmallVector<Register, 4> SubMerge(NumSubParts); 381 382 // Once we've fully read off the end of the original source bits, we can reuse 383 // the same high bits for remaining padding elements. 384 Register AllPadReg; 385 386 // Build merges to the LCM type to cover the original result type. 387 for (int I = 0; I != NumParts; ++I) { 388 bool AllMergePartsArePadding = true; 389 390 // Build the requested merges to the requested type. 391 for (int J = 0; J != NumSubParts; ++J) { 392 int Idx = I * NumSubParts + J; 393 if (Idx >= NumOrigSrc) { 394 SubMerge[J] = PadReg; 395 continue; 396 } 397 398 SubMerge[J] = VRegs[Idx]; 399 400 // There are meaningful bits here we can't reuse later. 401 AllMergePartsArePadding = false; 402 } 403 404 // If we've filled up a complete piece with padding bits, we can directly 405 // emit the natural sized constant if applicable, rather than a merge of 406 // smaller constants. 407 if (AllMergePartsArePadding && !AllPadReg) { 408 if (PadStrategy == TargetOpcode::G_ANYEXT) 409 AllPadReg = MIRBuilder.buildUndef(NarrowTy).getReg(0); 410 else if (PadStrategy == TargetOpcode::G_ZEXT) 411 AllPadReg = MIRBuilder.buildConstant(NarrowTy, 0).getReg(0); 412 413 // If this is a sign extension, we can't materialize a trivial constant 414 // with the right type and have to produce a merge. 415 } 416 417 if (AllPadReg) { 418 // Avoid creating additional instructions if we're just adding additional 419 // copies of padding bits. 420 Remerge[I] = AllPadReg; 421 continue; 422 } 423 424 if (NumSubParts == 1) 425 Remerge[I] = SubMerge[0]; 426 else 427 Remerge[I] = MIRBuilder.buildMergeLikeInstr(NarrowTy, SubMerge).getReg(0); 428 429 // In the sign extend padding case, re-use the first all-signbit merge. 430 if (AllMergePartsArePadding && !AllPadReg) 431 AllPadReg = Remerge[I]; 432 } 433 434 VRegs = std::move(Remerge); 435 return LCMTy; 436 } 437 438 void LegalizerHelper::buildWidenedRemergeToDst(Register DstReg, LLT LCMTy, 439 ArrayRef<Register> RemergeRegs) { 440 LLT DstTy = MRI.getType(DstReg); 441 442 // Create the merge to the widened source, and extract the relevant bits into 443 // the result. 444 445 if (DstTy == LCMTy) { 446 MIRBuilder.buildMergeLikeInstr(DstReg, RemergeRegs); 447 return; 448 } 449 450 auto Remerge = MIRBuilder.buildMergeLikeInstr(LCMTy, RemergeRegs); 451 if (DstTy.isScalar() && LCMTy.isScalar()) { 452 MIRBuilder.buildTrunc(DstReg, Remerge); 453 return; 454 } 455 456 if (LCMTy.isVector()) { 457 unsigned NumDefs = LCMTy.getSizeInBits() / DstTy.getSizeInBits(); 458 SmallVector<Register, 8> UnmergeDefs(NumDefs); 459 UnmergeDefs[0] = DstReg; 460 for (unsigned I = 1; I != NumDefs; ++I) 461 UnmergeDefs[I] = MRI.createGenericVirtualRegister(DstTy); 462 463 MIRBuilder.buildUnmerge(UnmergeDefs, 464 MIRBuilder.buildMergeLikeInstr(LCMTy, RemergeRegs)); 465 return; 466 } 467 468 llvm_unreachable("unhandled case"); 469 } 470 471 static RTLIB::Libcall getRTLibDesc(unsigned Opcode, unsigned Size) { 472 #define RTLIBCASE_INT(LibcallPrefix) \ 473 do { \ 474 switch (Size) { \ 475 case 32: \ 476 return RTLIB::LibcallPrefix##32; \ 477 case 64: \ 478 return RTLIB::LibcallPrefix##64; \ 479 case 128: \ 480 return RTLIB::LibcallPrefix##128; \ 481 default: \ 482 llvm_unreachable("unexpected size"); \ 483 } \ 484 } while (0) 485 486 #define RTLIBCASE(LibcallPrefix) \ 487 do { \ 488 switch (Size) { \ 489 case 32: \ 490 return RTLIB::LibcallPrefix##32; \ 491 case 64: \ 492 return RTLIB::LibcallPrefix##64; \ 493 case 80: \ 494 return RTLIB::LibcallPrefix##80; \ 495 case 128: \ 496 return RTLIB::LibcallPrefix##128; \ 497 default: \ 498 llvm_unreachable("unexpected size"); \ 499 } \ 500 } while (0) 501 502 switch (Opcode) { 503 case TargetOpcode::G_MUL: 504 RTLIBCASE_INT(MUL_I); 505 case TargetOpcode::G_SDIV: 506 RTLIBCASE_INT(SDIV_I); 507 case TargetOpcode::G_UDIV: 508 RTLIBCASE_INT(UDIV_I); 509 case TargetOpcode::G_SREM: 510 RTLIBCASE_INT(SREM_I); 511 case TargetOpcode::G_UREM: 512 RTLIBCASE_INT(UREM_I); 513 case TargetOpcode::G_CTLZ_ZERO_UNDEF: 514 RTLIBCASE_INT(CTLZ_I); 515 case TargetOpcode::G_FADD: 516 RTLIBCASE(ADD_F); 517 case TargetOpcode::G_FSUB: 518 RTLIBCASE(SUB_F); 519 case TargetOpcode::G_FMUL: 520 RTLIBCASE(MUL_F); 521 case TargetOpcode::G_FDIV: 522 RTLIBCASE(DIV_F); 523 case TargetOpcode::G_FEXP: 524 RTLIBCASE(EXP_F); 525 case TargetOpcode::G_FEXP2: 526 RTLIBCASE(EXP2_F); 527 case TargetOpcode::G_FREM: 528 RTLIBCASE(REM_F); 529 case TargetOpcode::G_FPOW: 530 RTLIBCASE(POW_F); 531 case TargetOpcode::G_FMA: 532 RTLIBCASE(FMA_F); 533 case TargetOpcode::G_FSIN: 534 RTLIBCASE(SIN_F); 535 case TargetOpcode::G_FCOS: 536 RTLIBCASE(COS_F); 537 case TargetOpcode::G_FLOG10: 538 RTLIBCASE(LOG10_F); 539 case TargetOpcode::G_FLOG: 540 RTLIBCASE(LOG_F); 541 case TargetOpcode::G_FLOG2: 542 RTLIBCASE(LOG2_F); 543 case TargetOpcode::G_FCEIL: 544 RTLIBCASE(CEIL_F); 545 case TargetOpcode::G_FFLOOR: 546 RTLIBCASE(FLOOR_F); 547 case TargetOpcode::G_FMINNUM: 548 RTLIBCASE(FMIN_F); 549 case TargetOpcode::G_FMAXNUM: 550 RTLIBCASE(FMAX_F); 551 case TargetOpcode::G_FSQRT: 552 RTLIBCASE(SQRT_F); 553 case TargetOpcode::G_FRINT: 554 RTLIBCASE(RINT_F); 555 case TargetOpcode::G_FNEARBYINT: 556 RTLIBCASE(NEARBYINT_F); 557 case TargetOpcode::G_INTRINSIC_ROUNDEVEN: 558 RTLIBCASE(ROUNDEVEN_F); 559 } 560 llvm_unreachable("Unknown libcall function"); 561 } 562 563 /// True if an instruction is in tail position in its caller. Intended for 564 /// legalizing libcalls as tail calls when possible. 565 static bool isLibCallInTailPosition(MachineInstr &MI, 566 const TargetInstrInfo &TII, 567 MachineRegisterInfo &MRI) { 568 MachineBasicBlock &MBB = *MI.getParent(); 569 const Function &F = MBB.getParent()->getFunction(); 570 571 // Conservatively require the attributes of the call to match those of 572 // the return. Ignore NoAlias and NonNull because they don't affect the 573 // call sequence. 574 AttributeList CallerAttrs = F.getAttributes(); 575 if (AttrBuilder(F.getContext(), CallerAttrs.getRetAttrs()) 576 .removeAttribute(Attribute::NoAlias) 577 .removeAttribute(Attribute::NonNull) 578 .hasAttributes()) 579 return false; 580 581 // It's not safe to eliminate the sign / zero extension of the return value. 582 if (CallerAttrs.hasRetAttr(Attribute::ZExt) || 583 CallerAttrs.hasRetAttr(Attribute::SExt)) 584 return false; 585 586 // Only tail call if the following instruction is a standard return or if we 587 // have a `thisreturn` callee, and a sequence like: 588 // 589 // G_MEMCPY %0, %1, %2 590 // $x0 = COPY %0 591 // RET_ReallyLR implicit $x0 592 auto Next = next_nodbg(MI.getIterator(), MBB.instr_end()); 593 if (Next != MBB.instr_end() && Next->isCopy()) { 594 switch (MI.getOpcode()) { 595 default: 596 llvm_unreachable("unsupported opcode"); 597 case TargetOpcode::G_BZERO: 598 return false; 599 case TargetOpcode::G_MEMCPY: 600 case TargetOpcode::G_MEMMOVE: 601 case TargetOpcode::G_MEMSET: 602 break; 603 } 604 605 Register VReg = MI.getOperand(0).getReg(); 606 if (!VReg.isVirtual() || VReg != Next->getOperand(1).getReg()) 607 return false; 608 609 Register PReg = Next->getOperand(0).getReg(); 610 if (!PReg.isPhysical()) 611 return false; 612 613 auto Ret = next_nodbg(Next, MBB.instr_end()); 614 if (Ret == MBB.instr_end() || !Ret->isReturn()) 615 return false; 616 617 if (Ret->getNumImplicitOperands() != 1) 618 return false; 619 620 if (PReg != Ret->getOperand(0).getReg()) 621 return false; 622 623 // Skip over the COPY that we just validated. 624 Next = Ret; 625 } 626 627 if (Next == MBB.instr_end() || TII.isTailCall(*Next) || !Next->isReturn()) 628 return false; 629 630 return true; 631 } 632 633 LegalizerHelper::LegalizeResult 634 llvm::createLibcall(MachineIRBuilder &MIRBuilder, const char *Name, 635 const CallLowering::ArgInfo &Result, 636 ArrayRef<CallLowering::ArgInfo> Args, 637 const CallingConv::ID CC) { 638 auto &CLI = *MIRBuilder.getMF().getSubtarget().getCallLowering(); 639 640 CallLowering::CallLoweringInfo Info; 641 Info.CallConv = CC; 642 Info.Callee = MachineOperand::CreateES(Name); 643 Info.OrigRet = Result; 644 std::copy(Args.begin(), Args.end(), std::back_inserter(Info.OrigArgs)); 645 if (!CLI.lowerCall(MIRBuilder, Info)) 646 return LegalizerHelper::UnableToLegalize; 647 648 return LegalizerHelper::Legalized; 649 } 650 651 LegalizerHelper::LegalizeResult 652 llvm::createLibcall(MachineIRBuilder &MIRBuilder, RTLIB::Libcall Libcall, 653 const CallLowering::ArgInfo &Result, 654 ArrayRef<CallLowering::ArgInfo> Args) { 655 auto &TLI = *MIRBuilder.getMF().getSubtarget().getTargetLowering(); 656 const char *Name = TLI.getLibcallName(Libcall); 657 const CallingConv::ID CC = TLI.getLibcallCallingConv(Libcall); 658 return createLibcall(MIRBuilder, Name, Result, Args, CC); 659 } 660 661 // Useful for libcalls where all operands have the same type. 662 static LegalizerHelper::LegalizeResult 663 simpleLibcall(MachineInstr &MI, MachineIRBuilder &MIRBuilder, unsigned Size, 664 Type *OpType) { 665 auto Libcall = getRTLibDesc(MI.getOpcode(), Size); 666 667 // FIXME: What does the original arg index mean here? 668 SmallVector<CallLowering::ArgInfo, 3> Args; 669 for (const MachineOperand &MO : llvm::drop_begin(MI.operands())) 670 Args.push_back({MO.getReg(), OpType, 0}); 671 return createLibcall(MIRBuilder, Libcall, 672 {MI.getOperand(0).getReg(), OpType, 0}, Args); 673 } 674 675 LegalizerHelper::LegalizeResult 676 llvm::createMemLibcall(MachineIRBuilder &MIRBuilder, MachineRegisterInfo &MRI, 677 MachineInstr &MI, LostDebugLocObserver &LocObserver) { 678 auto &Ctx = MIRBuilder.getMF().getFunction().getContext(); 679 680 SmallVector<CallLowering::ArgInfo, 3> Args; 681 // Add all the args, except for the last which is an imm denoting 'tail'. 682 for (unsigned i = 0; i < MI.getNumOperands() - 1; ++i) { 683 Register Reg = MI.getOperand(i).getReg(); 684 685 // Need derive an IR type for call lowering. 686 LLT OpLLT = MRI.getType(Reg); 687 Type *OpTy = nullptr; 688 if (OpLLT.isPointer()) 689 OpTy = Type::getInt8PtrTy(Ctx, OpLLT.getAddressSpace()); 690 else 691 OpTy = IntegerType::get(Ctx, OpLLT.getSizeInBits()); 692 Args.push_back({Reg, OpTy, 0}); 693 } 694 695 auto &CLI = *MIRBuilder.getMF().getSubtarget().getCallLowering(); 696 auto &TLI = *MIRBuilder.getMF().getSubtarget().getTargetLowering(); 697 RTLIB::Libcall RTLibcall; 698 unsigned Opc = MI.getOpcode(); 699 switch (Opc) { 700 case TargetOpcode::G_BZERO: 701 RTLibcall = RTLIB::BZERO; 702 break; 703 case TargetOpcode::G_MEMCPY: 704 RTLibcall = RTLIB::MEMCPY; 705 Args[0].Flags[0].setReturned(); 706 break; 707 case TargetOpcode::G_MEMMOVE: 708 RTLibcall = RTLIB::MEMMOVE; 709 Args[0].Flags[0].setReturned(); 710 break; 711 case TargetOpcode::G_MEMSET: 712 RTLibcall = RTLIB::MEMSET; 713 Args[0].Flags[0].setReturned(); 714 break; 715 default: 716 llvm_unreachable("unsupported opcode"); 717 } 718 const char *Name = TLI.getLibcallName(RTLibcall); 719 720 // Unsupported libcall on the target. 721 if (!Name) { 722 LLVM_DEBUG(dbgs() << ".. .. Could not find libcall name for " 723 << MIRBuilder.getTII().getName(Opc) << "\n"); 724 return LegalizerHelper::UnableToLegalize; 725 } 726 727 CallLowering::CallLoweringInfo Info; 728 Info.CallConv = TLI.getLibcallCallingConv(RTLibcall); 729 Info.Callee = MachineOperand::CreateES(Name); 730 Info.OrigRet = CallLowering::ArgInfo({0}, Type::getVoidTy(Ctx), 0); 731 Info.IsTailCall = MI.getOperand(MI.getNumOperands() - 1).getImm() && 732 isLibCallInTailPosition(MI, MIRBuilder.getTII(), MRI); 733 734 std::copy(Args.begin(), Args.end(), std::back_inserter(Info.OrigArgs)); 735 if (!CLI.lowerCall(MIRBuilder, Info)) 736 return LegalizerHelper::UnableToLegalize; 737 738 if (Info.LoweredTailCall) { 739 assert(Info.IsTailCall && "Lowered tail call when it wasn't a tail call?"); 740 741 // Check debug locations before removing the return. 742 LocObserver.checkpoint(true); 743 744 // We must have a return following the call (or debug insts) to get past 745 // isLibCallInTailPosition. 746 do { 747 MachineInstr *Next = MI.getNextNode(); 748 assert(Next && 749 (Next->isCopy() || Next->isReturn() || Next->isDebugInstr()) && 750 "Expected instr following MI to be return or debug inst?"); 751 // We lowered a tail call, so the call is now the return from the block. 752 // Delete the old return. 753 Next->eraseFromParent(); 754 } while (MI.getNextNode()); 755 756 // We expect to lose the debug location from the return. 757 LocObserver.checkpoint(false); 758 } 759 760 return LegalizerHelper::Legalized; 761 } 762 763 static RTLIB::Libcall getConvRTLibDesc(unsigned Opcode, Type *ToType, 764 Type *FromType) { 765 auto ToMVT = MVT::getVT(ToType); 766 auto FromMVT = MVT::getVT(FromType); 767 768 switch (Opcode) { 769 case TargetOpcode::G_FPEXT: 770 return RTLIB::getFPEXT(FromMVT, ToMVT); 771 case TargetOpcode::G_FPTRUNC: 772 return RTLIB::getFPROUND(FromMVT, ToMVT); 773 case TargetOpcode::G_FPTOSI: 774 return RTLIB::getFPTOSINT(FromMVT, ToMVT); 775 case TargetOpcode::G_FPTOUI: 776 return RTLIB::getFPTOUINT(FromMVT, ToMVT); 777 case TargetOpcode::G_SITOFP: 778 return RTLIB::getSINTTOFP(FromMVT, ToMVT); 779 case TargetOpcode::G_UITOFP: 780 return RTLIB::getUINTTOFP(FromMVT, ToMVT); 781 } 782 llvm_unreachable("Unsupported libcall function"); 783 } 784 785 static LegalizerHelper::LegalizeResult 786 conversionLibcall(MachineInstr &MI, MachineIRBuilder &MIRBuilder, Type *ToType, 787 Type *FromType) { 788 RTLIB::Libcall Libcall = getConvRTLibDesc(MI.getOpcode(), ToType, FromType); 789 return createLibcall(MIRBuilder, Libcall, 790 {MI.getOperand(0).getReg(), ToType, 0}, 791 {{MI.getOperand(1).getReg(), FromType, 0}}); 792 } 793 794 LegalizerHelper::LegalizeResult 795 LegalizerHelper::libcall(MachineInstr &MI, LostDebugLocObserver &LocObserver) { 796 LLT LLTy = MRI.getType(MI.getOperand(0).getReg()); 797 unsigned Size = LLTy.getSizeInBits(); 798 auto &Ctx = MIRBuilder.getMF().getFunction().getContext(); 799 800 switch (MI.getOpcode()) { 801 default: 802 return UnableToLegalize; 803 case TargetOpcode::G_MUL: 804 case TargetOpcode::G_SDIV: 805 case TargetOpcode::G_UDIV: 806 case TargetOpcode::G_SREM: 807 case TargetOpcode::G_UREM: 808 case TargetOpcode::G_CTLZ_ZERO_UNDEF: { 809 Type *HLTy = IntegerType::get(Ctx, Size); 810 auto Status = simpleLibcall(MI, MIRBuilder, Size, HLTy); 811 if (Status != Legalized) 812 return Status; 813 break; 814 } 815 case TargetOpcode::G_FADD: 816 case TargetOpcode::G_FSUB: 817 case TargetOpcode::G_FMUL: 818 case TargetOpcode::G_FDIV: 819 case TargetOpcode::G_FMA: 820 case TargetOpcode::G_FPOW: 821 case TargetOpcode::G_FREM: 822 case TargetOpcode::G_FCOS: 823 case TargetOpcode::G_FSIN: 824 case TargetOpcode::G_FLOG10: 825 case TargetOpcode::G_FLOG: 826 case TargetOpcode::G_FLOG2: 827 case TargetOpcode::G_FEXP: 828 case TargetOpcode::G_FEXP2: 829 case TargetOpcode::G_FCEIL: 830 case TargetOpcode::G_FFLOOR: 831 case TargetOpcode::G_FMINNUM: 832 case TargetOpcode::G_FMAXNUM: 833 case TargetOpcode::G_FSQRT: 834 case TargetOpcode::G_FRINT: 835 case TargetOpcode::G_FNEARBYINT: 836 case TargetOpcode::G_INTRINSIC_ROUNDEVEN: { 837 Type *HLTy = getFloatTypeForLLT(Ctx, LLTy); 838 if (!HLTy || (Size != 32 && Size != 64 && Size != 80 && Size != 128)) { 839 LLVM_DEBUG(dbgs() << "No libcall available for type " << LLTy << ".\n"); 840 return UnableToLegalize; 841 } 842 auto Status = simpleLibcall(MI, MIRBuilder, Size, HLTy); 843 if (Status != Legalized) 844 return Status; 845 break; 846 } 847 case TargetOpcode::G_FPEXT: 848 case TargetOpcode::G_FPTRUNC: { 849 Type *FromTy = getFloatTypeForLLT(Ctx, MRI.getType(MI.getOperand(1).getReg())); 850 Type *ToTy = getFloatTypeForLLT(Ctx, MRI.getType(MI.getOperand(0).getReg())); 851 if (!FromTy || !ToTy) 852 return UnableToLegalize; 853 LegalizeResult Status = conversionLibcall(MI, MIRBuilder, ToTy, FromTy ); 854 if (Status != Legalized) 855 return Status; 856 break; 857 } 858 case TargetOpcode::G_FPTOSI: 859 case TargetOpcode::G_FPTOUI: { 860 // FIXME: Support other types 861 unsigned FromSize = MRI.getType(MI.getOperand(1).getReg()).getSizeInBits(); 862 unsigned ToSize = MRI.getType(MI.getOperand(0).getReg()).getSizeInBits(); 863 if ((ToSize != 32 && ToSize != 64) || (FromSize != 32 && FromSize != 64)) 864 return UnableToLegalize; 865 LegalizeResult Status = conversionLibcall( 866 MI, MIRBuilder, 867 ToSize == 32 ? Type::getInt32Ty(Ctx) : Type::getInt64Ty(Ctx), 868 FromSize == 64 ? Type::getDoubleTy(Ctx) : Type::getFloatTy(Ctx)); 869 if (Status != Legalized) 870 return Status; 871 break; 872 } 873 case TargetOpcode::G_SITOFP: 874 case TargetOpcode::G_UITOFP: { 875 // FIXME: Support other types 876 unsigned FromSize = MRI.getType(MI.getOperand(1).getReg()).getSizeInBits(); 877 unsigned ToSize = MRI.getType(MI.getOperand(0).getReg()).getSizeInBits(); 878 if ((FromSize != 32 && FromSize != 64) || (ToSize != 32 && ToSize != 64)) 879 return UnableToLegalize; 880 LegalizeResult Status = conversionLibcall( 881 MI, MIRBuilder, 882 ToSize == 64 ? Type::getDoubleTy(Ctx) : Type::getFloatTy(Ctx), 883 FromSize == 32 ? Type::getInt32Ty(Ctx) : Type::getInt64Ty(Ctx)); 884 if (Status != Legalized) 885 return Status; 886 break; 887 } 888 case TargetOpcode::G_BZERO: 889 case TargetOpcode::G_MEMCPY: 890 case TargetOpcode::G_MEMMOVE: 891 case TargetOpcode::G_MEMSET: { 892 LegalizeResult Result = 893 createMemLibcall(MIRBuilder, *MIRBuilder.getMRI(), MI, LocObserver); 894 if (Result != Legalized) 895 return Result; 896 MI.eraseFromParent(); 897 return Result; 898 } 899 } 900 901 MI.eraseFromParent(); 902 return Legalized; 903 } 904 905 LegalizerHelper::LegalizeResult LegalizerHelper::narrowScalar(MachineInstr &MI, 906 unsigned TypeIdx, 907 LLT NarrowTy) { 908 uint64_t SizeOp0 = MRI.getType(MI.getOperand(0).getReg()).getSizeInBits(); 909 uint64_t NarrowSize = NarrowTy.getSizeInBits(); 910 911 switch (MI.getOpcode()) { 912 default: 913 return UnableToLegalize; 914 case TargetOpcode::G_IMPLICIT_DEF: { 915 Register DstReg = MI.getOperand(0).getReg(); 916 LLT DstTy = MRI.getType(DstReg); 917 918 // If SizeOp0 is not an exact multiple of NarrowSize, emit 919 // G_ANYEXT(G_IMPLICIT_DEF). Cast result to vector if needed. 920 // FIXME: Although this would also be legal for the general case, it causes 921 // a lot of regressions in the emitted code (superfluous COPYs, artifact 922 // combines not being hit). This seems to be a problem related to the 923 // artifact combiner. 924 if (SizeOp0 % NarrowSize != 0) { 925 LLT ImplicitTy = NarrowTy; 926 if (DstTy.isVector()) 927 ImplicitTy = LLT::vector(DstTy.getElementCount(), ImplicitTy); 928 929 Register ImplicitReg = MIRBuilder.buildUndef(ImplicitTy).getReg(0); 930 MIRBuilder.buildAnyExt(DstReg, ImplicitReg); 931 932 MI.eraseFromParent(); 933 return Legalized; 934 } 935 936 int NumParts = SizeOp0 / NarrowSize; 937 938 SmallVector<Register, 2> DstRegs; 939 for (int i = 0; i < NumParts; ++i) 940 DstRegs.push_back(MIRBuilder.buildUndef(NarrowTy).getReg(0)); 941 942 if (DstTy.isVector()) 943 MIRBuilder.buildBuildVector(DstReg, DstRegs); 944 else 945 MIRBuilder.buildMergeLikeInstr(DstReg, DstRegs); 946 MI.eraseFromParent(); 947 return Legalized; 948 } 949 case TargetOpcode::G_CONSTANT: { 950 LLT Ty = MRI.getType(MI.getOperand(0).getReg()); 951 const APInt &Val = MI.getOperand(1).getCImm()->getValue(); 952 unsigned TotalSize = Ty.getSizeInBits(); 953 unsigned NarrowSize = NarrowTy.getSizeInBits(); 954 int NumParts = TotalSize / NarrowSize; 955 956 SmallVector<Register, 4> PartRegs; 957 for (int I = 0; I != NumParts; ++I) { 958 unsigned Offset = I * NarrowSize; 959 auto K = MIRBuilder.buildConstant(NarrowTy, 960 Val.lshr(Offset).trunc(NarrowSize)); 961 PartRegs.push_back(K.getReg(0)); 962 } 963 964 LLT LeftoverTy; 965 unsigned LeftoverBits = TotalSize - NumParts * NarrowSize; 966 SmallVector<Register, 1> LeftoverRegs; 967 if (LeftoverBits != 0) { 968 LeftoverTy = LLT::scalar(LeftoverBits); 969 auto K = MIRBuilder.buildConstant( 970 LeftoverTy, 971 Val.lshr(NumParts * NarrowSize).trunc(LeftoverBits)); 972 LeftoverRegs.push_back(K.getReg(0)); 973 } 974 975 insertParts(MI.getOperand(0).getReg(), 976 Ty, NarrowTy, PartRegs, LeftoverTy, LeftoverRegs); 977 978 MI.eraseFromParent(); 979 return Legalized; 980 } 981 case TargetOpcode::G_SEXT: 982 case TargetOpcode::G_ZEXT: 983 case TargetOpcode::G_ANYEXT: 984 return narrowScalarExt(MI, TypeIdx, NarrowTy); 985 case TargetOpcode::G_TRUNC: { 986 if (TypeIdx != 1) 987 return UnableToLegalize; 988 989 uint64_t SizeOp1 = MRI.getType(MI.getOperand(1).getReg()).getSizeInBits(); 990 if (NarrowTy.getSizeInBits() * 2 != SizeOp1) { 991 LLVM_DEBUG(dbgs() << "Can't narrow trunc to type " << NarrowTy << "\n"); 992 return UnableToLegalize; 993 } 994 995 auto Unmerge = MIRBuilder.buildUnmerge(NarrowTy, MI.getOperand(1)); 996 MIRBuilder.buildCopy(MI.getOperand(0), Unmerge.getReg(0)); 997 MI.eraseFromParent(); 998 return Legalized; 999 } 1000 1001 case TargetOpcode::G_FREEZE: { 1002 if (TypeIdx != 0) 1003 return UnableToLegalize; 1004 1005 LLT Ty = MRI.getType(MI.getOperand(0).getReg()); 1006 // Should widen scalar first 1007 if (Ty.getSizeInBits() % NarrowTy.getSizeInBits() != 0) 1008 return UnableToLegalize; 1009 1010 auto Unmerge = MIRBuilder.buildUnmerge(NarrowTy, MI.getOperand(1).getReg()); 1011 SmallVector<Register, 8> Parts; 1012 for (unsigned i = 0; i < Unmerge->getNumDefs(); ++i) { 1013 Parts.push_back( 1014 MIRBuilder.buildFreeze(NarrowTy, Unmerge.getReg(i)).getReg(0)); 1015 } 1016 1017 MIRBuilder.buildMergeLikeInstr(MI.getOperand(0).getReg(), Parts); 1018 MI.eraseFromParent(); 1019 return Legalized; 1020 } 1021 case TargetOpcode::G_ADD: 1022 case TargetOpcode::G_SUB: 1023 case TargetOpcode::G_SADDO: 1024 case TargetOpcode::G_SSUBO: 1025 case TargetOpcode::G_SADDE: 1026 case TargetOpcode::G_SSUBE: 1027 case TargetOpcode::G_UADDO: 1028 case TargetOpcode::G_USUBO: 1029 case TargetOpcode::G_UADDE: 1030 case TargetOpcode::G_USUBE: 1031 return narrowScalarAddSub(MI, TypeIdx, NarrowTy); 1032 case TargetOpcode::G_MUL: 1033 case TargetOpcode::G_UMULH: 1034 return narrowScalarMul(MI, NarrowTy); 1035 case TargetOpcode::G_EXTRACT: 1036 return narrowScalarExtract(MI, TypeIdx, NarrowTy); 1037 case TargetOpcode::G_INSERT: 1038 return narrowScalarInsert(MI, TypeIdx, NarrowTy); 1039 case TargetOpcode::G_LOAD: { 1040 auto &LoadMI = cast<GLoad>(MI); 1041 Register DstReg = LoadMI.getDstReg(); 1042 LLT DstTy = MRI.getType(DstReg); 1043 if (DstTy.isVector()) 1044 return UnableToLegalize; 1045 1046 if (8 * LoadMI.getMemSize() != DstTy.getSizeInBits()) { 1047 Register TmpReg = MRI.createGenericVirtualRegister(NarrowTy); 1048 MIRBuilder.buildLoad(TmpReg, LoadMI.getPointerReg(), LoadMI.getMMO()); 1049 MIRBuilder.buildAnyExt(DstReg, TmpReg); 1050 LoadMI.eraseFromParent(); 1051 return Legalized; 1052 } 1053 1054 return reduceLoadStoreWidth(LoadMI, TypeIdx, NarrowTy); 1055 } 1056 case TargetOpcode::G_ZEXTLOAD: 1057 case TargetOpcode::G_SEXTLOAD: { 1058 auto &LoadMI = cast<GExtLoad>(MI); 1059 Register DstReg = LoadMI.getDstReg(); 1060 Register PtrReg = LoadMI.getPointerReg(); 1061 1062 Register TmpReg = MRI.createGenericVirtualRegister(NarrowTy); 1063 auto &MMO = LoadMI.getMMO(); 1064 unsigned MemSize = MMO.getSizeInBits(); 1065 1066 if (MemSize == NarrowSize) { 1067 MIRBuilder.buildLoad(TmpReg, PtrReg, MMO); 1068 } else if (MemSize < NarrowSize) { 1069 MIRBuilder.buildLoadInstr(LoadMI.getOpcode(), TmpReg, PtrReg, MMO); 1070 } else if (MemSize > NarrowSize) { 1071 // FIXME: Need to split the load. 1072 return UnableToLegalize; 1073 } 1074 1075 if (isa<GZExtLoad>(LoadMI)) 1076 MIRBuilder.buildZExt(DstReg, TmpReg); 1077 else 1078 MIRBuilder.buildSExt(DstReg, TmpReg); 1079 1080 LoadMI.eraseFromParent(); 1081 return Legalized; 1082 } 1083 case TargetOpcode::G_STORE: { 1084 auto &StoreMI = cast<GStore>(MI); 1085 1086 Register SrcReg = StoreMI.getValueReg(); 1087 LLT SrcTy = MRI.getType(SrcReg); 1088 if (SrcTy.isVector()) 1089 return UnableToLegalize; 1090 1091 int NumParts = SizeOp0 / NarrowSize; 1092 unsigned HandledSize = NumParts * NarrowTy.getSizeInBits(); 1093 unsigned LeftoverBits = SrcTy.getSizeInBits() - HandledSize; 1094 if (SrcTy.isVector() && LeftoverBits != 0) 1095 return UnableToLegalize; 1096 1097 if (8 * StoreMI.getMemSize() != SrcTy.getSizeInBits()) { 1098 Register TmpReg = MRI.createGenericVirtualRegister(NarrowTy); 1099 MIRBuilder.buildTrunc(TmpReg, SrcReg); 1100 MIRBuilder.buildStore(TmpReg, StoreMI.getPointerReg(), StoreMI.getMMO()); 1101 StoreMI.eraseFromParent(); 1102 return Legalized; 1103 } 1104 1105 return reduceLoadStoreWidth(StoreMI, 0, NarrowTy); 1106 } 1107 case TargetOpcode::G_SELECT: 1108 return narrowScalarSelect(MI, TypeIdx, NarrowTy); 1109 case TargetOpcode::G_AND: 1110 case TargetOpcode::G_OR: 1111 case TargetOpcode::G_XOR: { 1112 // Legalize bitwise operation: 1113 // A = BinOp<Ty> B, C 1114 // into: 1115 // B1, ..., BN = G_UNMERGE_VALUES B 1116 // C1, ..., CN = G_UNMERGE_VALUES C 1117 // A1 = BinOp<Ty/N> B1, C2 1118 // ... 1119 // AN = BinOp<Ty/N> BN, CN 1120 // A = G_MERGE_VALUES A1, ..., AN 1121 return narrowScalarBasic(MI, TypeIdx, NarrowTy); 1122 } 1123 case TargetOpcode::G_SHL: 1124 case TargetOpcode::G_LSHR: 1125 case TargetOpcode::G_ASHR: 1126 return narrowScalarShift(MI, TypeIdx, NarrowTy); 1127 case TargetOpcode::G_CTLZ: 1128 case TargetOpcode::G_CTLZ_ZERO_UNDEF: 1129 case TargetOpcode::G_CTTZ: 1130 case TargetOpcode::G_CTTZ_ZERO_UNDEF: 1131 case TargetOpcode::G_CTPOP: 1132 if (TypeIdx == 1) 1133 switch (MI.getOpcode()) { 1134 case TargetOpcode::G_CTLZ: 1135 case TargetOpcode::G_CTLZ_ZERO_UNDEF: 1136 return narrowScalarCTLZ(MI, TypeIdx, NarrowTy); 1137 case TargetOpcode::G_CTTZ: 1138 case TargetOpcode::G_CTTZ_ZERO_UNDEF: 1139 return narrowScalarCTTZ(MI, TypeIdx, NarrowTy); 1140 case TargetOpcode::G_CTPOP: 1141 return narrowScalarCTPOP(MI, TypeIdx, NarrowTy); 1142 default: 1143 return UnableToLegalize; 1144 } 1145 1146 Observer.changingInstr(MI); 1147 narrowScalarDst(MI, NarrowTy, 0, TargetOpcode::G_ZEXT); 1148 Observer.changedInstr(MI); 1149 return Legalized; 1150 case TargetOpcode::G_INTTOPTR: 1151 if (TypeIdx != 1) 1152 return UnableToLegalize; 1153 1154 Observer.changingInstr(MI); 1155 narrowScalarSrc(MI, NarrowTy, 1); 1156 Observer.changedInstr(MI); 1157 return Legalized; 1158 case TargetOpcode::G_PTRTOINT: 1159 if (TypeIdx != 0) 1160 return UnableToLegalize; 1161 1162 Observer.changingInstr(MI); 1163 narrowScalarDst(MI, NarrowTy, 0, TargetOpcode::G_ZEXT); 1164 Observer.changedInstr(MI); 1165 return Legalized; 1166 case TargetOpcode::G_PHI: { 1167 // FIXME: add support for when SizeOp0 isn't an exact multiple of 1168 // NarrowSize. 1169 if (SizeOp0 % NarrowSize != 0) 1170 return UnableToLegalize; 1171 1172 unsigned NumParts = SizeOp0 / NarrowSize; 1173 SmallVector<Register, 2> DstRegs(NumParts); 1174 SmallVector<SmallVector<Register, 2>, 2> SrcRegs(MI.getNumOperands() / 2); 1175 Observer.changingInstr(MI); 1176 for (unsigned i = 1; i < MI.getNumOperands(); i += 2) { 1177 MachineBasicBlock &OpMBB = *MI.getOperand(i + 1).getMBB(); 1178 MIRBuilder.setInsertPt(OpMBB, OpMBB.getFirstTerminatorForward()); 1179 extractParts(MI.getOperand(i).getReg(), NarrowTy, NumParts, 1180 SrcRegs[i / 2]); 1181 } 1182 MachineBasicBlock &MBB = *MI.getParent(); 1183 MIRBuilder.setInsertPt(MBB, MI); 1184 for (unsigned i = 0; i < NumParts; ++i) { 1185 DstRegs[i] = MRI.createGenericVirtualRegister(NarrowTy); 1186 MachineInstrBuilder MIB = 1187 MIRBuilder.buildInstr(TargetOpcode::G_PHI).addDef(DstRegs[i]); 1188 for (unsigned j = 1; j < MI.getNumOperands(); j += 2) 1189 MIB.addUse(SrcRegs[j / 2][i]).add(MI.getOperand(j + 1)); 1190 } 1191 MIRBuilder.setInsertPt(MBB, MBB.getFirstNonPHI()); 1192 MIRBuilder.buildMergeLikeInstr(MI.getOperand(0), DstRegs); 1193 Observer.changedInstr(MI); 1194 MI.eraseFromParent(); 1195 return Legalized; 1196 } 1197 case TargetOpcode::G_EXTRACT_VECTOR_ELT: 1198 case TargetOpcode::G_INSERT_VECTOR_ELT: { 1199 if (TypeIdx != 2) 1200 return UnableToLegalize; 1201 1202 int OpIdx = MI.getOpcode() == TargetOpcode::G_EXTRACT_VECTOR_ELT ? 2 : 3; 1203 Observer.changingInstr(MI); 1204 narrowScalarSrc(MI, NarrowTy, OpIdx); 1205 Observer.changedInstr(MI); 1206 return Legalized; 1207 } 1208 case TargetOpcode::G_ICMP: { 1209 Register LHS = MI.getOperand(2).getReg(); 1210 LLT SrcTy = MRI.getType(LHS); 1211 uint64_t SrcSize = SrcTy.getSizeInBits(); 1212 CmpInst::Predicate Pred = 1213 static_cast<CmpInst::Predicate>(MI.getOperand(1).getPredicate()); 1214 1215 // TODO: Handle the non-equality case for weird sizes. 1216 if (NarrowSize * 2 != SrcSize && !ICmpInst::isEquality(Pred)) 1217 return UnableToLegalize; 1218 1219 LLT LeftoverTy; // Example: s88 -> s64 (NarrowTy) + s24 (leftover) 1220 SmallVector<Register, 4> LHSPartRegs, LHSLeftoverRegs; 1221 if (!extractParts(LHS, SrcTy, NarrowTy, LeftoverTy, LHSPartRegs, 1222 LHSLeftoverRegs)) 1223 return UnableToLegalize; 1224 1225 LLT Unused; // Matches LeftoverTy; G_ICMP LHS and RHS are the same type. 1226 SmallVector<Register, 4> RHSPartRegs, RHSLeftoverRegs; 1227 if (!extractParts(MI.getOperand(3).getReg(), SrcTy, NarrowTy, Unused, 1228 RHSPartRegs, RHSLeftoverRegs)) 1229 return UnableToLegalize; 1230 1231 // We now have the LHS and RHS of the compare split into narrow-type 1232 // registers, plus potentially some leftover type. 1233 Register Dst = MI.getOperand(0).getReg(); 1234 LLT ResTy = MRI.getType(Dst); 1235 if (ICmpInst::isEquality(Pred)) { 1236 // For each part on the LHS and RHS, keep track of the result of XOR-ing 1237 // them together. For each equal part, the result should be all 0s. For 1238 // each non-equal part, we'll get at least one 1. 1239 auto Zero = MIRBuilder.buildConstant(NarrowTy, 0); 1240 SmallVector<Register, 4> Xors; 1241 for (auto LHSAndRHS : zip(LHSPartRegs, RHSPartRegs)) { 1242 auto LHS = std::get<0>(LHSAndRHS); 1243 auto RHS = std::get<1>(LHSAndRHS); 1244 auto Xor = MIRBuilder.buildXor(NarrowTy, LHS, RHS).getReg(0); 1245 Xors.push_back(Xor); 1246 } 1247 1248 // Build a G_XOR for each leftover register. Each G_XOR must be widened 1249 // to the desired narrow type so that we can OR them together later. 1250 SmallVector<Register, 4> WidenedXors; 1251 for (auto LHSAndRHS : zip(LHSLeftoverRegs, RHSLeftoverRegs)) { 1252 auto LHS = std::get<0>(LHSAndRHS); 1253 auto RHS = std::get<1>(LHSAndRHS); 1254 auto Xor = MIRBuilder.buildXor(LeftoverTy, LHS, RHS).getReg(0); 1255 LLT GCDTy = extractGCDType(WidenedXors, NarrowTy, LeftoverTy, Xor); 1256 buildLCMMergePieces(LeftoverTy, NarrowTy, GCDTy, WidenedXors, 1257 /* PadStrategy = */ TargetOpcode::G_ZEXT); 1258 Xors.insert(Xors.end(), WidenedXors.begin(), WidenedXors.end()); 1259 } 1260 1261 // Now, for each part we broke up, we know if they are equal/not equal 1262 // based off the G_XOR. We can OR these all together and compare against 1263 // 0 to get the result. 1264 assert(Xors.size() >= 2 && "Should have gotten at least two Xors?"); 1265 auto Or = MIRBuilder.buildOr(NarrowTy, Xors[0], Xors[1]); 1266 for (unsigned I = 2, E = Xors.size(); I < E; ++I) 1267 Or = MIRBuilder.buildOr(NarrowTy, Or, Xors[I]); 1268 MIRBuilder.buildICmp(Pred, Dst, Or, Zero); 1269 } else { 1270 // TODO: Handle non-power-of-two types. 1271 assert(LHSPartRegs.size() == 2 && "Expected exactly 2 LHS part regs?"); 1272 assert(RHSPartRegs.size() == 2 && "Expected exactly 2 RHS part regs?"); 1273 Register LHSL = LHSPartRegs[0]; 1274 Register LHSH = LHSPartRegs[1]; 1275 Register RHSL = RHSPartRegs[0]; 1276 Register RHSH = RHSPartRegs[1]; 1277 MachineInstrBuilder CmpH = MIRBuilder.buildICmp(Pred, ResTy, LHSH, RHSH); 1278 MachineInstrBuilder CmpHEQ = 1279 MIRBuilder.buildICmp(CmpInst::Predicate::ICMP_EQ, ResTy, LHSH, RHSH); 1280 MachineInstrBuilder CmpLU = MIRBuilder.buildICmp( 1281 ICmpInst::getUnsignedPredicate(Pred), ResTy, LHSL, RHSL); 1282 MIRBuilder.buildSelect(Dst, CmpHEQ, CmpLU, CmpH); 1283 } 1284 MI.eraseFromParent(); 1285 return Legalized; 1286 } 1287 case TargetOpcode::G_SEXT_INREG: { 1288 if (TypeIdx != 0) 1289 return UnableToLegalize; 1290 1291 int64_t SizeInBits = MI.getOperand(2).getImm(); 1292 1293 // So long as the new type has more bits than the bits we're extending we 1294 // don't need to break it apart. 1295 if (NarrowTy.getScalarSizeInBits() >= SizeInBits) { 1296 Observer.changingInstr(MI); 1297 // We don't lose any non-extension bits by truncating the src and 1298 // sign-extending the dst. 1299 MachineOperand &MO1 = MI.getOperand(1); 1300 auto TruncMIB = MIRBuilder.buildTrunc(NarrowTy, MO1); 1301 MO1.setReg(TruncMIB.getReg(0)); 1302 1303 MachineOperand &MO2 = MI.getOperand(0); 1304 Register DstExt = MRI.createGenericVirtualRegister(NarrowTy); 1305 MIRBuilder.setInsertPt(MIRBuilder.getMBB(), ++MIRBuilder.getInsertPt()); 1306 MIRBuilder.buildSExt(MO2, DstExt); 1307 MO2.setReg(DstExt); 1308 Observer.changedInstr(MI); 1309 return Legalized; 1310 } 1311 1312 // Break it apart. Components below the extension point are unmodified. The 1313 // component containing the extension point becomes a narrower SEXT_INREG. 1314 // Components above it are ashr'd from the component containing the 1315 // extension point. 1316 if (SizeOp0 % NarrowSize != 0) 1317 return UnableToLegalize; 1318 int NumParts = SizeOp0 / NarrowSize; 1319 1320 // List the registers where the destination will be scattered. 1321 SmallVector<Register, 2> DstRegs; 1322 // List the registers where the source will be split. 1323 SmallVector<Register, 2> SrcRegs; 1324 1325 // Create all the temporary registers. 1326 for (int i = 0; i < NumParts; ++i) { 1327 Register SrcReg = MRI.createGenericVirtualRegister(NarrowTy); 1328 1329 SrcRegs.push_back(SrcReg); 1330 } 1331 1332 // Explode the big arguments into smaller chunks. 1333 MIRBuilder.buildUnmerge(SrcRegs, MI.getOperand(1)); 1334 1335 Register AshrCstReg = 1336 MIRBuilder.buildConstant(NarrowTy, NarrowTy.getScalarSizeInBits() - 1) 1337 .getReg(0); 1338 Register FullExtensionReg = 0; 1339 Register PartialExtensionReg = 0; 1340 1341 // Do the operation on each small part. 1342 for (int i = 0; i < NumParts; ++i) { 1343 if ((i + 1) * NarrowTy.getScalarSizeInBits() < SizeInBits) 1344 DstRegs.push_back(SrcRegs[i]); 1345 else if (i * NarrowTy.getScalarSizeInBits() > SizeInBits) { 1346 assert(PartialExtensionReg && 1347 "Expected to visit partial extension before full"); 1348 if (FullExtensionReg) { 1349 DstRegs.push_back(FullExtensionReg); 1350 continue; 1351 } 1352 DstRegs.push_back( 1353 MIRBuilder.buildAShr(NarrowTy, PartialExtensionReg, AshrCstReg) 1354 .getReg(0)); 1355 FullExtensionReg = DstRegs.back(); 1356 } else { 1357 DstRegs.push_back( 1358 MIRBuilder 1359 .buildInstr( 1360 TargetOpcode::G_SEXT_INREG, {NarrowTy}, 1361 {SrcRegs[i], SizeInBits % NarrowTy.getScalarSizeInBits()}) 1362 .getReg(0)); 1363 PartialExtensionReg = DstRegs.back(); 1364 } 1365 } 1366 1367 // Gather the destination registers into the final destination. 1368 Register DstReg = MI.getOperand(0).getReg(); 1369 MIRBuilder.buildMergeLikeInstr(DstReg, DstRegs); 1370 MI.eraseFromParent(); 1371 return Legalized; 1372 } 1373 case TargetOpcode::G_BSWAP: 1374 case TargetOpcode::G_BITREVERSE: { 1375 if (SizeOp0 % NarrowSize != 0) 1376 return UnableToLegalize; 1377 1378 Observer.changingInstr(MI); 1379 SmallVector<Register, 2> SrcRegs, DstRegs; 1380 unsigned NumParts = SizeOp0 / NarrowSize; 1381 extractParts(MI.getOperand(1).getReg(), NarrowTy, NumParts, SrcRegs); 1382 1383 for (unsigned i = 0; i < NumParts; ++i) { 1384 auto DstPart = MIRBuilder.buildInstr(MI.getOpcode(), {NarrowTy}, 1385 {SrcRegs[NumParts - 1 - i]}); 1386 DstRegs.push_back(DstPart.getReg(0)); 1387 } 1388 1389 MIRBuilder.buildMergeLikeInstr(MI.getOperand(0), DstRegs); 1390 1391 Observer.changedInstr(MI); 1392 MI.eraseFromParent(); 1393 return Legalized; 1394 } 1395 case TargetOpcode::G_PTR_ADD: 1396 case TargetOpcode::G_PTRMASK: { 1397 if (TypeIdx != 1) 1398 return UnableToLegalize; 1399 Observer.changingInstr(MI); 1400 narrowScalarSrc(MI, NarrowTy, 2); 1401 Observer.changedInstr(MI); 1402 return Legalized; 1403 } 1404 case TargetOpcode::G_FPTOUI: 1405 case TargetOpcode::G_FPTOSI: 1406 return narrowScalarFPTOI(MI, TypeIdx, NarrowTy); 1407 case TargetOpcode::G_FPEXT: 1408 if (TypeIdx != 0) 1409 return UnableToLegalize; 1410 Observer.changingInstr(MI); 1411 narrowScalarDst(MI, NarrowTy, 0, TargetOpcode::G_FPEXT); 1412 Observer.changedInstr(MI); 1413 return Legalized; 1414 } 1415 } 1416 1417 Register LegalizerHelper::coerceToScalar(Register Val) { 1418 LLT Ty = MRI.getType(Val); 1419 if (Ty.isScalar()) 1420 return Val; 1421 1422 const DataLayout &DL = MIRBuilder.getDataLayout(); 1423 LLT NewTy = LLT::scalar(Ty.getSizeInBits()); 1424 if (Ty.isPointer()) { 1425 if (DL.isNonIntegralAddressSpace(Ty.getAddressSpace())) 1426 return Register(); 1427 return MIRBuilder.buildPtrToInt(NewTy, Val).getReg(0); 1428 } 1429 1430 Register NewVal = Val; 1431 1432 assert(Ty.isVector()); 1433 LLT EltTy = Ty.getElementType(); 1434 if (EltTy.isPointer()) 1435 NewVal = MIRBuilder.buildPtrToInt(NewTy, NewVal).getReg(0); 1436 return MIRBuilder.buildBitcast(NewTy, NewVal).getReg(0); 1437 } 1438 1439 void LegalizerHelper::widenScalarSrc(MachineInstr &MI, LLT WideTy, 1440 unsigned OpIdx, unsigned ExtOpcode) { 1441 MachineOperand &MO = MI.getOperand(OpIdx); 1442 auto ExtB = MIRBuilder.buildInstr(ExtOpcode, {WideTy}, {MO}); 1443 MO.setReg(ExtB.getReg(0)); 1444 } 1445 1446 void LegalizerHelper::narrowScalarSrc(MachineInstr &MI, LLT NarrowTy, 1447 unsigned OpIdx) { 1448 MachineOperand &MO = MI.getOperand(OpIdx); 1449 auto ExtB = MIRBuilder.buildTrunc(NarrowTy, MO); 1450 MO.setReg(ExtB.getReg(0)); 1451 } 1452 1453 void LegalizerHelper::widenScalarDst(MachineInstr &MI, LLT WideTy, 1454 unsigned OpIdx, unsigned TruncOpcode) { 1455 MachineOperand &MO = MI.getOperand(OpIdx); 1456 Register DstExt = MRI.createGenericVirtualRegister(WideTy); 1457 MIRBuilder.setInsertPt(MIRBuilder.getMBB(), ++MIRBuilder.getInsertPt()); 1458 MIRBuilder.buildInstr(TruncOpcode, {MO}, {DstExt}); 1459 MO.setReg(DstExt); 1460 } 1461 1462 void LegalizerHelper::narrowScalarDst(MachineInstr &MI, LLT NarrowTy, 1463 unsigned OpIdx, unsigned ExtOpcode) { 1464 MachineOperand &MO = MI.getOperand(OpIdx); 1465 Register DstTrunc = MRI.createGenericVirtualRegister(NarrowTy); 1466 MIRBuilder.setInsertPt(MIRBuilder.getMBB(), ++MIRBuilder.getInsertPt()); 1467 MIRBuilder.buildInstr(ExtOpcode, {MO}, {DstTrunc}); 1468 MO.setReg(DstTrunc); 1469 } 1470 1471 void LegalizerHelper::moreElementsVectorDst(MachineInstr &MI, LLT WideTy, 1472 unsigned OpIdx) { 1473 MachineOperand &MO = MI.getOperand(OpIdx); 1474 MIRBuilder.setInsertPt(MIRBuilder.getMBB(), ++MIRBuilder.getInsertPt()); 1475 Register Dst = MO.getReg(); 1476 Register DstExt = MRI.createGenericVirtualRegister(WideTy); 1477 MO.setReg(DstExt); 1478 MIRBuilder.buildDeleteTrailingVectorElements(Dst, DstExt); 1479 } 1480 1481 void LegalizerHelper::moreElementsVectorSrc(MachineInstr &MI, LLT MoreTy, 1482 unsigned OpIdx) { 1483 MachineOperand &MO = MI.getOperand(OpIdx); 1484 SmallVector<Register, 8> Regs; 1485 MO.setReg(MIRBuilder.buildPadVectorWithUndefElements(MoreTy, MO).getReg(0)); 1486 } 1487 1488 void LegalizerHelper::bitcastSrc(MachineInstr &MI, LLT CastTy, unsigned OpIdx) { 1489 MachineOperand &Op = MI.getOperand(OpIdx); 1490 Op.setReg(MIRBuilder.buildBitcast(CastTy, Op).getReg(0)); 1491 } 1492 1493 void LegalizerHelper::bitcastDst(MachineInstr &MI, LLT CastTy, unsigned OpIdx) { 1494 MachineOperand &MO = MI.getOperand(OpIdx); 1495 Register CastDst = MRI.createGenericVirtualRegister(CastTy); 1496 MIRBuilder.setInsertPt(MIRBuilder.getMBB(), ++MIRBuilder.getInsertPt()); 1497 MIRBuilder.buildBitcast(MO, CastDst); 1498 MO.setReg(CastDst); 1499 } 1500 1501 LegalizerHelper::LegalizeResult 1502 LegalizerHelper::widenScalarMergeValues(MachineInstr &MI, unsigned TypeIdx, 1503 LLT WideTy) { 1504 if (TypeIdx != 1) 1505 return UnableToLegalize; 1506 1507 Register DstReg = MI.getOperand(0).getReg(); 1508 LLT DstTy = MRI.getType(DstReg); 1509 if (DstTy.isVector()) 1510 return UnableToLegalize; 1511 1512 Register Src1 = MI.getOperand(1).getReg(); 1513 LLT SrcTy = MRI.getType(Src1); 1514 const int DstSize = DstTy.getSizeInBits(); 1515 const int SrcSize = SrcTy.getSizeInBits(); 1516 const int WideSize = WideTy.getSizeInBits(); 1517 const int NumMerge = (DstSize + WideSize - 1) / WideSize; 1518 1519 unsigned NumOps = MI.getNumOperands(); 1520 unsigned NumSrc = MI.getNumOperands() - 1; 1521 unsigned PartSize = DstTy.getSizeInBits() / NumSrc; 1522 1523 if (WideSize >= DstSize) { 1524 // Directly pack the bits in the target type. 1525 Register ResultReg = MIRBuilder.buildZExt(WideTy, Src1).getReg(0); 1526 1527 for (unsigned I = 2; I != NumOps; ++I) { 1528 const unsigned Offset = (I - 1) * PartSize; 1529 1530 Register SrcReg = MI.getOperand(I).getReg(); 1531 assert(MRI.getType(SrcReg) == LLT::scalar(PartSize)); 1532 1533 auto ZextInput = MIRBuilder.buildZExt(WideTy, SrcReg); 1534 1535 Register NextResult = I + 1 == NumOps && WideTy == DstTy ? DstReg : 1536 MRI.createGenericVirtualRegister(WideTy); 1537 1538 auto ShiftAmt = MIRBuilder.buildConstant(WideTy, Offset); 1539 auto Shl = MIRBuilder.buildShl(WideTy, ZextInput, ShiftAmt); 1540 MIRBuilder.buildOr(NextResult, ResultReg, Shl); 1541 ResultReg = NextResult; 1542 } 1543 1544 if (WideSize > DstSize) 1545 MIRBuilder.buildTrunc(DstReg, ResultReg); 1546 else if (DstTy.isPointer()) 1547 MIRBuilder.buildIntToPtr(DstReg, ResultReg); 1548 1549 MI.eraseFromParent(); 1550 return Legalized; 1551 } 1552 1553 // Unmerge the original values to the GCD type, and recombine to the next 1554 // multiple greater than the original type. 1555 // 1556 // %3:_(s12) = G_MERGE_VALUES %0:_(s4), %1:_(s4), %2:_(s4) -> s6 1557 // %4:_(s2), %5:_(s2) = G_UNMERGE_VALUES %0 1558 // %6:_(s2), %7:_(s2) = G_UNMERGE_VALUES %1 1559 // %8:_(s2), %9:_(s2) = G_UNMERGE_VALUES %2 1560 // %10:_(s6) = G_MERGE_VALUES %4, %5, %6 1561 // %11:_(s6) = G_MERGE_VALUES %7, %8, %9 1562 // %12:_(s12) = G_MERGE_VALUES %10, %11 1563 // 1564 // Padding with undef if necessary: 1565 // 1566 // %2:_(s8) = G_MERGE_VALUES %0:_(s4), %1:_(s4) -> s6 1567 // %3:_(s2), %4:_(s2) = G_UNMERGE_VALUES %0 1568 // %5:_(s2), %6:_(s2) = G_UNMERGE_VALUES %1 1569 // %7:_(s2) = G_IMPLICIT_DEF 1570 // %8:_(s6) = G_MERGE_VALUES %3, %4, %5 1571 // %9:_(s6) = G_MERGE_VALUES %6, %7, %7 1572 // %10:_(s12) = G_MERGE_VALUES %8, %9 1573 1574 const int GCD = std::gcd(SrcSize, WideSize); 1575 LLT GCDTy = LLT::scalar(GCD); 1576 1577 SmallVector<Register, 8> Parts; 1578 SmallVector<Register, 8> NewMergeRegs; 1579 SmallVector<Register, 8> Unmerges; 1580 LLT WideDstTy = LLT::scalar(NumMerge * WideSize); 1581 1582 // Decompose the original operands if they don't evenly divide. 1583 for (const MachineOperand &MO : llvm::drop_begin(MI.operands())) { 1584 Register SrcReg = MO.getReg(); 1585 if (GCD == SrcSize) { 1586 Unmerges.push_back(SrcReg); 1587 } else { 1588 auto Unmerge = MIRBuilder.buildUnmerge(GCDTy, SrcReg); 1589 for (int J = 0, JE = Unmerge->getNumOperands() - 1; J != JE; ++J) 1590 Unmerges.push_back(Unmerge.getReg(J)); 1591 } 1592 } 1593 1594 // Pad with undef to the next size that is a multiple of the requested size. 1595 if (static_cast<int>(Unmerges.size()) != NumMerge * WideSize) { 1596 Register UndefReg = MIRBuilder.buildUndef(GCDTy).getReg(0); 1597 for (int I = Unmerges.size(); I != NumMerge * WideSize; ++I) 1598 Unmerges.push_back(UndefReg); 1599 } 1600 1601 const int PartsPerGCD = WideSize / GCD; 1602 1603 // Build merges of each piece. 1604 ArrayRef<Register> Slicer(Unmerges); 1605 for (int I = 0; I != NumMerge; ++I, Slicer = Slicer.drop_front(PartsPerGCD)) { 1606 auto Merge = 1607 MIRBuilder.buildMergeLikeInstr(WideTy, Slicer.take_front(PartsPerGCD)); 1608 NewMergeRegs.push_back(Merge.getReg(0)); 1609 } 1610 1611 // A truncate may be necessary if the requested type doesn't evenly divide the 1612 // original result type. 1613 if (DstTy.getSizeInBits() == WideDstTy.getSizeInBits()) { 1614 MIRBuilder.buildMergeLikeInstr(DstReg, NewMergeRegs); 1615 } else { 1616 auto FinalMerge = MIRBuilder.buildMergeLikeInstr(WideDstTy, NewMergeRegs); 1617 MIRBuilder.buildTrunc(DstReg, FinalMerge.getReg(0)); 1618 } 1619 1620 MI.eraseFromParent(); 1621 return Legalized; 1622 } 1623 1624 LegalizerHelper::LegalizeResult 1625 LegalizerHelper::widenScalarUnmergeValues(MachineInstr &MI, unsigned TypeIdx, 1626 LLT WideTy) { 1627 if (TypeIdx != 0) 1628 return UnableToLegalize; 1629 1630 int NumDst = MI.getNumOperands() - 1; 1631 Register SrcReg = MI.getOperand(NumDst).getReg(); 1632 LLT SrcTy = MRI.getType(SrcReg); 1633 if (SrcTy.isVector()) 1634 return UnableToLegalize; 1635 1636 Register Dst0Reg = MI.getOperand(0).getReg(); 1637 LLT DstTy = MRI.getType(Dst0Reg); 1638 if (!DstTy.isScalar()) 1639 return UnableToLegalize; 1640 1641 if (WideTy.getSizeInBits() >= SrcTy.getSizeInBits()) { 1642 if (SrcTy.isPointer()) { 1643 const DataLayout &DL = MIRBuilder.getDataLayout(); 1644 if (DL.isNonIntegralAddressSpace(SrcTy.getAddressSpace())) { 1645 LLVM_DEBUG( 1646 dbgs() << "Not casting non-integral address space integer\n"); 1647 return UnableToLegalize; 1648 } 1649 1650 SrcTy = LLT::scalar(SrcTy.getSizeInBits()); 1651 SrcReg = MIRBuilder.buildPtrToInt(SrcTy, SrcReg).getReg(0); 1652 } 1653 1654 // Widen SrcTy to WideTy. This does not affect the result, but since the 1655 // user requested this size, it is probably better handled than SrcTy and 1656 // should reduce the total number of legalization artifacts. 1657 if (WideTy.getSizeInBits() > SrcTy.getSizeInBits()) { 1658 SrcTy = WideTy; 1659 SrcReg = MIRBuilder.buildAnyExt(WideTy, SrcReg).getReg(0); 1660 } 1661 1662 // Theres no unmerge type to target. Directly extract the bits from the 1663 // source type 1664 unsigned DstSize = DstTy.getSizeInBits(); 1665 1666 MIRBuilder.buildTrunc(Dst0Reg, SrcReg); 1667 for (int I = 1; I != NumDst; ++I) { 1668 auto ShiftAmt = MIRBuilder.buildConstant(SrcTy, DstSize * I); 1669 auto Shr = MIRBuilder.buildLShr(SrcTy, SrcReg, ShiftAmt); 1670 MIRBuilder.buildTrunc(MI.getOperand(I), Shr); 1671 } 1672 1673 MI.eraseFromParent(); 1674 return Legalized; 1675 } 1676 1677 // Extend the source to a wider type. 1678 LLT LCMTy = getLCMType(SrcTy, WideTy); 1679 1680 Register WideSrc = SrcReg; 1681 if (LCMTy.getSizeInBits() != SrcTy.getSizeInBits()) { 1682 // TODO: If this is an integral address space, cast to integer and anyext. 1683 if (SrcTy.isPointer()) { 1684 LLVM_DEBUG(dbgs() << "Widening pointer source types not implemented\n"); 1685 return UnableToLegalize; 1686 } 1687 1688 WideSrc = MIRBuilder.buildAnyExt(LCMTy, WideSrc).getReg(0); 1689 } 1690 1691 auto Unmerge = MIRBuilder.buildUnmerge(WideTy, WideSrc); 1692 1693 // Create a sequence of unmerges and merges to the original results. Since we 1694 // may have widened the source, we will need to pad the results with dead defs 1695 // to cover the source register. 1696 // e.g. widen s48 to s64: 1697 // %1:_(s48), %2:_(s48) = G_UNMERGE_VALUES %0:_(s96) 1698 // 1699 // => 1700 // %4:_(s192) = G_ANYEXT %0:_(s96) 1701 // %5:_(s64), %6, %7 = G_UNMERGE_VALUES %4 ; Requested unmerge 1702 // ; unpack to GCD type, with extra dead defs 1703 // %8:_(s16), %9, %10, %11 = G_UNMERGE_VALUES %5:_(s64) 1704 // %12:_(s16), %13, dead %14, dead %15 = G_UNMERGE_VALUES %6:_(s64) 1705 // dead %16:_(s16), dead %17, dead %18, dead %18 = G_UNMERGE_VALUES %7:_(s64) 1706 // %1:_(s48) = G_MERGE_VALUES %8:_(s16), %9, %10 ; Remerge to destination 1707 // %2:_(s48) = G_MERGE_VALUES %11:_(s16), %12, %13 ; Remerge to destination 1708 const LLT GCDTy = getGCDType(WideTy, DstTy); 1709 const int NumUnmerge = Unmerge->getNumOperands() - 1; 1710 const int PartsPerRemerge = DstTy.getSizeInBits() / GCDTy.getSizeInBits(); 1711 1712 // Directly unmerge to the destination without going through a GCD type 1713 // if possible 1714 if (PartsPerRemerge == 1) { 1715 const int PartsPerUnmerge = WideTy.getSizeInBits() / DstTy.getSizeInBits(); 1716 1717 for (int I = 0; I != NumUnmerge; ++I) { 1718 auto MIB = MIRBuilder.buildInstr(TargetOpcode::G_UNMERGE_VALUES); 1719 1720 for (int J = 0; J != PartsPerUnmerge; ++J) { 1721 int Idx = I * PartsPerUnmerge + J; 1722 if (Idx < NumDst) 1723 MIB.addDef(MI.getOperand(Idx).getReg()); 1724 else { 1725 // Create dead def for excess components. 1726 MIB.addDef(MRI.createGenericVirtualRegister(DstTy)); 1727 } 1728 } 1729 1730 MIB.addUse(Unmerge.getReg(I)); 1731 } 1732 } else { 1733 SmallVector<Register, 16> Parts; 1734 for (int J = 0; J != NumUnmerge; ++J) 1735 extractGCDType(Parts, GCDTy, Unmerge.getReg(J)); 1736 1737 SmallVector<Register, 8> RemergeParts; 1738 for (int I = 0; I != NumDst; ++I) { 1739 for (int J = 0; J < PartsPerRemerge; ++J) { 1740 const int Idx = I * PartsPerRemerge + J; 1741 RemergeParts.emplace_back(Parts[Idx]); 1742 } 1743 1744 MIRBuilder.buildMergeLikeInstr(MI.getOperand(I).getReg(), RemergeParts); 1745 RemergeParts.clear(); 1746 } 1747 } 1748 1749 MI.eraseFromParent(); 1750 return Legalized; 1751 } 1752 1753 LegalizerHelper::LegalizeResult 1754 LegalizerHelper::widenScalarExtract(MachineInstr &MI, unsigned TypeIdx, 1755 LLT WideTy) { 1756 Register DstReg = MI.getOperand(0).getReg(); 1757 Register SrcReg = MI.getOperand(1).getReg(); 1758 LLT SrcTy = MRI.getType(SrcReg); 1759 1760 LLT DstTy = MRI.getType(DstReg); 1761 unsigned Offset = MI.getOperand(2).getImm(); 1762 1763 if (TypeIdx == 0) { 1764 if (SrcTy.isVector() || DstTy.isVector()) 1765 return UnableToLegalize; 1766 1767 SrcOp Src(SrcReg); 1768 if (SrcTy.isPointer()) { 1769 // Extracts from pointers can be handled only if they are really just 1770 // simple integers. 1771 const DataLayout &DL = MIRBuilder.getDataLayout(); 1772 if (DL.isNonIntegralAddressSpace(SrcTy.getAddressSpace())) 1773 return UnableToLegalize; 1774 1775 LLT SrcAsIntTy = LLT::scalar(SrcTy.getSizeInBits()); 1776 Src = MIRBuilder.buildPtrToInt(SrcAsIntTy, Src); 1777 SrcTy = SrcAsIntTy; 1778 } 1779 1780 if (DstTy.isPointer()) 1781 return UnableToLegalize; 1782 1783 if (Offset == 0) { 1784 // Avoid a shift in the degenerate case. 1785 MIRBuilder.buildTrunc(DstReg, 1786 MIRBuilder.buildAnyExtOrTrunc(WideTy, Src)); 1787 MI.eraseFromParent(); 1788 return Legalized; 1789 } 1790 1791 // Do a shift in the source type. 1792 LLT ShiftTy = SrcTy; 1793 if (WideTy.getSizeInBits() > SrcTy.getSizeInBits()) { 1794 Src = MIRBuilder.buildAnyExt(WideTy, Src); 1795 ShiftTy = WideTy; 1796 } 1797 1798 auto LShr = MIRBuilder.buildLShr( 1799 ShiftTy, Src, MIRBuilder.buildConstant(ShiftTy, Offset)); 1800 MIRBuilder.buildTrunc(DstReg, LShr); 1801 MI.eraseFromParent(); 1802 return Legalized; 1803 } 1804 1805 if (SrcTy.isScalar()) { 1806 Observer.changingInstr(MI); 1807 widenScalarSrc(MI, WideTy, 1, TargetOpcode::G_ANYEXT); 1808 Observer.changedInstr(MI); 1809 return Legalized; 1810 } 1811 1812 if (!SrcTy.isVector()) 1813 return UnableToLegalize; 1814 1815 if (DstTy != SrcTy.getElementType()) 1816 return UnableToLegalize; 1817 1818 if (Offset % SrcTy.getScalarSizeInBits() != 0) 1819 return UnableToLegalize; 1820 1821 Observer.changingInstr(MI); 1822 widenScalarSrc(MI, WideTy, 1, TargetOpcode::G_ANYEXT); 1823 1824 MI.getOperand(2).setImm((WideTy.getSizeInBits() / SrcTy.getSizeInBits()) * 1825 Offset); 1826 widenScalarDst(MI, WideTy.getScalarType(), 0); 1827 Observer.changedInstr(MI); 1828 return Legalized; 1829 } 1830 1831 LegalizerHelper::LegalizeResult 1832 LegalizerHelper::widenScalarInsert(MachineInstr &MI, unsigned TypeIdx, 1833 LLT WideTy) { 1834 if (TypeIdx != 0 || WideTy.isVector()) 1835 return UnableToLegalize; 1836 Observer.changingInstr(MI); 1837 widenScalarSrc(MI, WideTy, 1, TargetOpcode::G_ANYEXT); 1838 widenScalarDst(MI, WideTy); 1839 Observer.changedInstr(MI); 1840 return Legalized; 1841 } 1842 1843 LegalizerHelper::LegalizeResult 1844 LegalizerHelper::widenScalarAddSubOverflow(MachineInstr &MI, unsigned TypeIdx, 1845 LLT WideTy) { 1846 unsigned Opcode; 1847 unsigned ExtOpcode; 1848 std::optional<Register> CarryIn; 1849 switch (MI.getOpcode()) { 1850 default: 1851 llvm_unreachable("Unexpected opcode!"); 1852 case TargetOpcode::G_SADDO: 1853 Opcode = TargetOpcode::G_ADD; 1854 ExtOpcode = TargetOpcode::G_SEXT; 1855 break; 1856 case TargetOpcode::G_SSUBO: 1857 Opcode = TargetOpcode::G_SUB; 1858 ExtOpcode = TargetOpcode::G_SEXT; 1859 break; 1860 case TargetOpcode::G_UADDO: 1861 Opcode = TargetOpcode::G_ADD; 1862 ExtOpcode = TargetOpcode::G_ZEXT; 1863 break; 1864 case TargetOpcode::G_USUBO: 1865 Opcode = TargetOpcode::G_SUB; 1866 ExtOpcode = TargetOpcode::G_ZEXT; 1867 break; 1868 case TargetOpcode::G_SADDE: 1869 Opcode = TargetOpcode::G_UADDE; 1870 ExtOpcode = TargetOpcode::G_SEXT; 1871 CarryIn = MI.getOperand(4).getReg(); 1872 break; 1873 case TargetOpcode::G_SSUBE: 1874 Opcode = TargetOpcode::G_USUBE; 1875 ExtOpcode = TargetOpcode::G_SEXT; 1876 CarryIn = MI.getOperand(4).getReg(); 1877 break; 1878 case TargetOpcode::G_UADDE: 1879 Opcode = TargetOpcode::G_UADDE; 1880 ExtOpcode = TargetOpcode::G_ZEXT; 1881 CarryIn = MI.getOperand(4).getReg(); 1882 break; 1883 case TargetOpcode::G_USUBE: 1884 Opcode = TargetOpcode::G_USUBE; 1885 ExtOpcode = TargetOpcode::G_ZEXT; 1886 CarryIn = MI.getOperand(4).getReg(); 1887 break; 1888 } 1889 1890 if (TypeIdx == 1) { 1891 unsigned BoolExtOp = MIRBuilder.getBoolExtOp(WideTy.isVector(), false); 1892 1893 Observer.changingInstr(MI); 1894 if (CarryIn) 1895 widenScalarSrc(MI, WideTy, 4, BoolExtOp); 1896 widenScalarDst(MI, WideTy, 1); 1897 1898 Observer.changedInstr(MI); 1899 return Legalized; 1900 } 1901 1902 auto LHSExt = MIRBuilder.buildInstr(ExtOpcode, {WideTy}, {MI.getOperand(2)}); 1903 auto RHSExt = MIRBuilder.buildInstr(ExtOpcode, {WideTy}, {MI.getOperand(3)}); 1904 // Do the arithmetic in the larger type. 1905 Register NewOp; 1906 if (CarryIn) { 1907 LLT CarryOutTy = MRI.getType(MI.getOperand(1).getReg()); 1908 NewOp = MIRBuilder 1909 .buildInstr(Opcode, {WideTy, CarryOutTy}, 1910 {LHSExt, RHSExt, *CarryIn}) 1911 .getReg(0); 1912 } else { 1913 NewOp = MIRBuilder.buildInstr(Opcode, {WideTy}, {LHSExt, RHSExt}).getReg(0); 1914 } 1915 LLT OrigTy = MRI.getType(MI.getOperand(0).getReg()); 1916 auto TruncOp = MIRBuilder.buildTrunc(OrigTy, NewOp); 1917 auto ExtOp = MIRBuilder.buildInstr(ExtOpcode, {WideTy}, {TruncOp}); 1918 // There is no overflow if the ExtOp is the same as NewOp. 1919 MIRBuilder.buildICmp(CmpInst::ICMP_NE, MI.getOperand(1), NewOp, ExtOp); 1920 // Now trunc the NewOp to the original result. 1921 MIRBuilder.buildTrunc(MI.getOperand(0), NewOp); 1922 MI.eraseFromParent(); 1923 return Legalized; 1924 } 1925 1926 LegalizerHelper::LegalizeResult 1927 LegalizerHelper::widenScalarAddSubShlSat(MachineInstr &MI, unsigned TypeIdx, 1928 LLT WideTy) { 1929 bool IsSigned = MI.getOpcode() == TargetOpcode::G_SADDSAT || 1930 MI.getOpcode() == TargetOpcode::G_SSUBSAT || 1931 MI.getOpcode() == TargetOpcode::G_SSHLSAT; 1932 bool IsShift = MI.getOpcode() == TargetOpcode::G_SSHLSAT || 1933 MI.getOpcode() == TargetOpcode::G_USHLSAT; 1934 // We can convert this to: 1935 // 1. Any extend iN to iM 1936 // 2. SHL by M-N 1937 // 3. [US][ADD|SUB|SHL]SAT 1938 // 4. L/ASHR by M-N 1939 // 1940 // It may be more efficient to lower this to a min and a max operation in 1941 // the higher precision arithmetic if the promoted operation isn't legal, 1942 // but this decision is up to the target's lowering request. 1943 Register DstReg = MI.getOperand(0).getReg(); 1944 1945 unsigned NewBits = WideTy.getScalarSizeInBits(); 1946 unsigned SHLAmount = NewBits - MRI.getType(DstReg).getScalarSizeInBits(); 1947 1948 // Shifts must zero-extend the RHS to preserve the unsigned quantity, and 1949 // must not left shift the RHS to preserve the shift amount. 1950 auto LHS = MIRBuilder.buildAnyExt(WideTy, MI.getOperand(1)); 1951 auto RHS = IsShift ? MIRBuilder.buildZExt(WideTy, MI.getOperand(2)) 1952 : MIRBuilder.buildAnyExt(WideTy, MI.getOperand(2)); 1953 auto ShiftK = MIRBuilder.buildConstant(WideTy, SHLAmount); 1954 auto ShiftL = MIRBuilder.buildShl(WideTy, LHS, ShiftK); 1955 auto ShiftR = IsShift ? RHS : MIRBuilder.buildShl(WideTy, RHS, ShiftK); 1956 1957 auto WideInst = MIRBuilder.buildInstr(MI.getOpcode(), {WideTy}, 1958 {ShiftL, ShiftR}, MI.getFlags()); 1959 1960 // Use a shift that will preserve the number of sign bits when the trunc is 1961 // folded away. 1962 auto Result = IsSigned ? MIRBuilder.buildAShr(WideTy, WideInst, ShiftK) 1963 : MIRBuilder.buildLShr(WideTy, WideInst, ShiftK); 1964 1965 MIRBuilder.buildTrunc(DstReg, Result); 1966 MI.eraseFromParent(); 1967 return Legalized; 1968 } 1969 1970 LegalizerHelper::LegalizeResult 1971 LegalizerHelper::widenScalarMulo(MachineInstr &MI, unsigned TypeIdx, 1972 LLT WideTy) { 1973 if (TypeIdx == 1) { 1974 Observer.changingInstr(MI); 1975 widenScalarDst(MI, WideTy, 1); 1976 Observer.changedInstr(MI); 1977 return Legalized; 1978 } 1979 1980 bool IsSigned = MI.getOpcode() == TargetOpcode::G_SMULO; 1981 Register Result = MI.getOperand(0).getReg(); 1982 Register OriginalOverflow = MI.getOperand(1).getReg(); 1983 Register LHS = MI.getOperand(2).getReg(); 1984 Register RHS = MI.getOperand(3).getReg(); 1985 LLT SrcTy = MRI.getType(LHS); 1986 LLT OverflowTy = MRI.getType(OriginalOverflow); 1987 unsigned SrcBitWidth = SrcTy.getScalarSizeInBits(); 1988 1989 // To determine if the result overflowed in the larger type, we extend the 1990 // input to the larger type, do the multiply (checking if it overflows), 1991 // then also check the high bits of the result to see if overflow happened 1992 // there. 1993 unsigned ExtOp = IsSigned ? TargetOpcode::G_SEXT : TargetOpcode::G_ZEXT; 1994 auto LeftOperand = MIRBuilder.buildInstr(ExtOp, {WideTy}, {LHS}); 1995 auto RightOperand = MIRBuilder.buildInstr(ExtOp, {WideTy}, {RHS}); 1996 1997 auto Mulo = MIRBuilder.buildInstr(MI.getOpcode(), {WideTy, OverflowTy}, 1998 {LeftOperand, RightOperand}); 1999 auto Mul = Mulo->getOperand(0); 2000 MIRBuilder.buildTrunc(Result, Mul); 2001 2002 MachineInstrBuilder ExtResult; 2003 // Overflow occurred if it occurred in the larger type, or if the high part 2004 // of the result does not zero/sign-extend the low part. Check this second 2005 // possibility first. 2006 if (IsSigned) { 2007 // For signed, overflow occurred when the high part does not sign-extend 2008 // the low part. 2009 ExtResult = MIRBuilder.buildSExtInReg(WideTy, Mul, SrcBitWidth); 2010 } else { 2011 // Unsigned overflow occurred when the high part does not zero-extend the 2012 // low part. 2013 ExtResult = MIRBuilder.buildZExtInReg(WideTy, Mul, SrcBitWidth); 2014 } 2015 2016 // Multiplication cannot overflow if the WideTy is >= 2 * original width, 2017 // so we don't need to check the overflow result of larger type Mulo. 2018 if (WideTy.getScalarSizeInBits() < 2 * SrcBitWidth) { 2019 auto Overflow = 2020 MIRBuilder.buildICmp(CmpInst::ICMP_NE, OverflowTy, Mul, ExtResult); 2021 // Finally check if the multiplication in the larger type itself overflowed. 2022 MIRBuilder.buildOr(OriginalOverflow, Mulo->getOperand(1), Overflow); 2023 } else { 2024 MIRBuilder.buildICmp(CmpInst::ICMP_NE, OriginalOverflow, Mul, ExtResult); 2025 } 2026 MI.eraseFromParent(); 2027 return Legalized; 2028 } 2029 2030 LegalizerHelper::LegalizeResult 2031 LegalizerHelper::widenScalar(MachineInstr &MI, unsigned TypeIdx, LLT WideTy) { 2032 switch (MI.getOpcode()) { 2033 default: 2034 return UnableToLegalize; 2035 case TargetOpcode::G_ATOMICRMW_XCHG: 2036 case TargetOpcode::G_ATOMICRMW_ADD: 2037 case TargetOpcode::G_ATOMICRMW_SUB: 2038 case TargetOpcode::G_ATOMICRMW_AND: 2039 case TargetOpcode::G_ATOMICRMW_OR: 2040 case TargetOpcode::G_ATOMICRMW_XOR: 2041 case TargetOpcode::G_ATOMICRMW_MIN: 2042 case TargetOpcode::G_ATOMICRMW_MAX: 2043 case TargetOpcode::G_ATOMICRMW_UMIN: 2044 case TargetOpcode::G_ATOMICRMW_UMAX: 2045 assert(TypeIdx == 0 && "atomicrmw with second scalar type"); 2046 Observer.changingInstr(MI); 2047 widenScalarSrc(MI, WideTy, 2, TargetOpcode::G_ANYEXT); 2048 widenScalarDst(MI, WideTy, 0); 2049 Observer.changedInstr(MI); 2050 return Legalized; 2051 case TargetOpcode::G_ATOMIC_CMPXCHG: 2052 assert(TypeIdx == 0 && "G_ATOMIC_CMPXCHG with second scalar type"); 2053 Observer.changingInstr(MI); 2054 widenScalarSrc(MI, WideTy, 2, TargetOpcode::G_ANYEXT); 2055 widenScalarSrc(MI, WideTy, 3, TargetOpcode::G_ANYEXT); 2056 widenScalarDst(MI, WideTy, 0); 2057 Observer.changedInstr(MI); 2058 return Legalized; 2059 case TargetOpcode::G_ATOMIC_CMPXCHG_WITH_SUCCESS: 2060 if (TypeIdx == 0) { 2061 Observer.changingInstr(MI); 2062 widenScalarSrc(MI, WideTy, 3, TargetOpcode::G_ANYEXT); 2063 widenScalarSrc(MI, WideTy, 4, TargetOpcode::G_ANYEXT); 2064 widenScalarDst(MI, WideTy, 0); 2065 Observer.changedInstr(MI); 2066 return Legalized; 2067 } 2068 assert(TypeIdx == 1 && 2069 "G_ATOMIC_CMPXCHG_WITH_SUCCESS with third scalar type"); 2070 Observer.changingInstr(MI); 2071 widenScalarDst(MI, WideTy, 1); 2072 Observer.changedInstr(MI); 2073 return Legalized; 2074 case TargetOpcode::G_EXTRACT: 2075 return widenScalarExtract(MI, TypeIdx, WideTy); 2076 case TargetOpcode::G_INSERT: 2077 return widenScalarInsert(MI, TypeIdx, WideTy); 2078 case TargetOpcode::G_MERGE_VALUES: 2079 return widenScalarMergeValues(MI, TypeIdx, WideTy); 2080 case TargetOpcode::G_UNMERGE_VALUES: 2081 return widenScalarUnmergeValues(MI, TypeIdx, WideTy); 2082 case TargetOpcode::G_SADDO: 2083 case TargetOpcode::G_SSUBO: 2084 case TargetOpcode::G_UADDO: 2085 case TargetOpcode::G_USUBO: 2086 case TargetOpcode::G_SADDE: 2087 case TargetOpcode::G_SSUBE: 2088 case TargetOpcode::G_UADDE: 2089 case TargetOpcode::G_USUBE: 2090 return widenScalarAddSubOverflow(MI, TypeIdx, WideTy); 2091 case TargetOpcode::G_UMULO: 2092 case TargetOpcode::G_SMULO: 2093 return widenScalarMulo(MI, TypeIdx, WideTy); 2094 case TargetOpcode::G_SADDSAT: 2095 case TargetOpcode::G_SSUBSAT: 2096 case TargetOpcode::G_SSHLSAT: 2097 case TargetOpcode::G_UADDSAT: 2098 case TargetOpcode::G_USUBSAT: 2099 case TargetOpcode::G_USHLSAT: 2100 return widenScalarAddSubShlSat(MI, TypeIdx, WideTy); 2101 case TargetOpcode::G_CTTZ: 2102 case TargetOpcode::G_CTTZ_ZERO_UNDEF: 2103 case TargetOpcode::G_CTLZ: 2104 case TargetOpcode::G_CTLZ_ZERO_UNDEF: 2105 case TargetOpcode::G_CTPOP: { 2106 if (TypeIdx == 0) { 2107 Observer.changingInstr(MI); 2108 widenScalarDst(MI, WideTy, 0); 2109 Observer.changedInstr(MI); 2110 return Legalized; 2111 } 2112 2113 Register SrcReg = MI.getOperand(1).getReg(); 2114 2115 // First extend the input. 2116 unsigned ExtOpc = MI.getOpcode() == TargetOpcode::G_CTTZ || 2117 MI.getOpcode() == TargetOpcode::G_CTTZ_ZERO_UNDEF 2118 ? TargetOpcode::G_ANYEXT 2119 : TargetOpcode::G_ZEXT; 2120 auto MIBSrc = MIRBuilder.buildInstr(ExtOpc, {WideTy}, {SrcReg}); 2121 LLT CurTy = MRI.getType(SrcReg); 2122 unsigned NewOpc = MI.getOpcode(); 2123 if (NewOpc == TargetOpcode::G_CTTZ) { 2124 // The count is the same in the larger type except if the original 2125 // value was zero. This can be handled by setting the bit just off 2126 // the top of the original type. 2127 auto TopBit = 2128 APInt::getOneBitSet(WideTy.getSizeInBits(), CurTy.getSizeInBits()); 2129 MIBSrc = MIRBuilder.buildOr( 2130 WideTy, MIBSrc, MIRBuilder.buildConstant(WideTy, TopBit)); 2131 // Now we know the operand is non-zero, use the more relaxed opcode. 2132 NewOpc = TargetOpcode::G_CTTZ_ZERO_UNDEF; 2133 } 2134 2135 // Perform the operation at the larger size. 2136 auto MIBNewOp = MIRBuilder.buildInstr(NewOpc, {WideTy}, {MIBSrc}); 2137 // This is already the correct result for CTPOP and CTTZs 2138 if (MI.getOpcode() == TargetOpcode::G_CTLZ || 2139 MI.getOpcode() == TargetOpcode::G_CTLZ_ZERO_UNDEF) { 2140 // The correct result is NewOp - (Difference in widety and current ty). 2141 unsigned SizeDiff = WideTy.getSizeInBits() - CurTy.getSizeInBits(); 2142 MIBNewOp = MIRBuilder.buildSub( 2143 WideTy, MIBNewOp, MIRBuilder.buildConstant(WideTy, SizeDiff)); 2144 } 2145 2146 MIRBuilder.buildZExtOrTrunc(MI.getOperand(0), MIBNewOp); 2147 MI.eraseFromParent(); 2148 return Legalized; 2149 } 2150 case TargetOpcode::G_BSWAP: { 2151 Observer.changingInstr(MI); 2152 Register DstReg = MI.getOperand(0).getReg(); 2153 2154 Register ShrReg = MRI.createGenericVirtualRegister(WideTy); 2155 Register DstExt = MRI.createGenericVirtualRegister(WideTy); 2156 Register ShiftAmtReg = MRI.createGenericVirtualRegister(WideTy); 2157 widenScalarSrc(MI, WideTy, 1, TargetOpcode::G_ANYEXT); 2158 2159 MI.getOperand(0).setReg(DstExt); 2160 2161 MIRBuilder.setInsertPt(MIRBuilder.getMBB(), ++MIRBuilder.getInsertPt()); 2162 2163 LLT Ty = MRI.getType(DstReg); 2164 unsigned DiffBits = WideTy.getScalarSizeInBits() - Ty.getScalarSizeInBits(); 2165 MIRBuilder.buildConstant(ShiftAmtReg, DiffBits); 2166 MIRBuilder.buildLShr(ShrReg, DstExt, ShiftAmtReg); 2167 2168 MIRBuilder.buildTrunc(DstReg, ShrReg); 2169 Observer.changedInstr(MI); 2170 return Legalized; 2171 } 2172 case TargetOpcode::G_BITREVERSE: { 2173 Observer.changingInstr(MI); 2174 2175 Register DstReg = MI.getOperand(0).getReg(); 2176 LLT Ty = MRI.getType(DstReg); 2177 unsigned DiffBits = WideTy.getScalarSizeInBits() - Ty.getScalarSizeInBits(); 2178 2179 Register DstExt = MRI.createGenericVirtualRegister(WideTy); 2180 widenScalarSrc(MI, WideTy, 1, TargetOpcode::G_ANYEXT); 2181 MI.getOperand(0).setReg(DstExt); 2182 MIRBuilder.setInsertPt(MIRBuilder.getMBB(), ++MIRBuilder.getInsertPt()); 2183 2184 auto ShiftAmt = MIRBuilder.buildConstant(WideTy, DiffBits); 2185 auto Shift = MIRBuilder.buildLShr(WideTy, DstExt, ShiftAmt); 2186 MIRBuilder.buildTrunc(DstReg, Shift); 2187 Observer.changedInstr(MI); 2188 return Legalized; 2189 } 2190 case TargetOpcode::G_FREEZE: 2191 Observer.changingInstr(MI); 2192 widenScalarSrc(MI, WideTy, 1, TargetOpcode::G_ANYEXT); 2193 widenScalarDst(MI, WideTy); 2194 Observer.changedInstr(MI); 2195 return Legalized; 2196 2197 case TargetOpcode::G_ABS: 2198 Observer.changingInstr(MI); 2199 widenScalarSrc(MI, WideTy, 1, TargetOpcode::G_SEXT); 2200 widenScalarDst(MI, WideTy); 2201 Observer.changedInstr(MI); 2202 return Legalized; 2203 2204 case TargetOpcode::G_ADD: 2205 case TargetOpcode::G_AND: 2206 case TargetOpcode::G_MUL: 2207 case TargetOpcode::G_OR: 2208 case TargetOpcode::G_XOR: 2209 case TargetOpcode::G_SUB: 2210 // Perform operation at larger width (any extension is fines here, high bits 2211 // don't affect the result) and then truncate the result back to the 2212 // original type. 2213 Observer.changingInstr(MI); 2214 widenScalarSrc(MI, WideTy, 1, TargetOpcode::G_ANYEXT); 2215 widenScalarSrc(MI, WideTy, 2, TargetOpcode::G_ANYEXT); 2216 widenScalarDst(MI, WideTy); 2217 Observer.changedInstr(MI); 2218 return Legalized; 2219 2220 case TargetOpcode::G_SBFX: 2221 case TargetOpcode::G_UBFX: 2222 Observer.changingInstr(MI); 2223 2224 if (TypeIdx == 0) { 2225 widenScalarSrc(MI, WideTy, 1, TargetOpcode::G_ANYEXT); 2226 widenScalarDst(MI, WideTy); 2227 } else { 2228 widenScalarSrc(MI, WideTy, 2, TargetOpcode::G_ZEXT); 2229 widenScalarSrc(MI, WideTy, 3, TargetOpcode::G_ZEXT); 2230 } 2231 2232 Observer.changedInstr(MI); 2233 return Legalized; 2234 2235 case TargetOpcode::G_SHL: 2236 Observer.changingInstr(MI); 2237 2238 if (TypeIdx == 0) { 2239 widenScalarSrc(MI, WideTy, 1, TargetOpcode::G_ANYEXT); 2240 widenScalarDst(MI, WideTy); 2241 } else { 2242 assert(TypeIdx == 1); 2243 // The "number of bits to shift" operand must preserve its value as an 2244 // unsigned integer: 2245 widenScalarSrc(MI, WideTy, 2, TargetOpcode::G_ZEXT); 2246 } 2247 2248 Observer.changedInstr(MI); 2249 return Legalized; 2250 2251 case TargetOpcode::G_SDIV: 2252 case TargetOpcode::G_SREM: 2253 case TargetOpcode::G_SMIN: 2254 case TargetOpcode::G_SMAX: 2255 Observer.changingInstr(MI); 2256 widenScalarSrc(MI, WideTy, 1, TargetOpcode::G_SEXT); 2257 widenScalarSrc(MI, WideTy, 2, TargetOpcode::G_SEXT); 2258 widenScalarDst(MI, WideTy); 2259 Observer.changedInstr(MI); 2260 return Legalized; 2261 2262 case TargetOpcode::G_SDIVREM: 2263 Observer.changingInstr(MI); 2264 widenScalarSrc(MI, WideTy, 2, TargetOpcode::G_SEXT); 2265 widenScalarSrc(MI, WideTy, 3, TargetOpcode::G_SEXT); 2266 widenScalarDst(MI, WideTy); 2267 widenScalarDst(MI, WideTy, 1); 2268 Observer.changedInstr(MI); 2269 return Legalized; 2270 2271 case TargetOpcode::G_ASHR: 2272 case TargetOpcode::G_LSHR: 2273 Observer.changingInstr(MI); 2274 2275 if (TypeIdx == 0) { 2276 unsigned CvtOp = MI.getOpcode() == TargetOpcode::G_ASHR ? 2277 TargetOpcode::G_SEXT : TargetOpcode::G_ZEXT; 2278 2279 widenScalarSrc(MI, WideTy, 1, CvtOp); 2280 widenScalarDst(MI, WideTy); 2281 } else { 2282 assert(TypeIdx == 1); 2283 // The "number of bits to shift" operand must preserve its value as an 2284 // unsigned integer: 2285 widenScalarSrc(MI, WideTy, 2, TargetOpcode::G_ZEXT); 2286 } 2287 2288 Observer.changedInstr(MI); 2289 return Legalized; 2290 case TargetOpcode::G_UDIV: 2291 case TargetOpcode::G_UREM: 2292 case TargetOpcode::G_UMIN: 2293 case TargetOpcode::G_UMAX: 2294 Observer.changingInstr(MI); 2295 widenScalarSrc(MI, WideTy, 1, TargetOpcode::G_ZEXT); 2296 widenScalarSrc(MI, WideTy, 2, TargetOpcode::G_ZEXT); 2297 widenScalarDst(MI, WideTy); 2298 Observer.changedInstr(MI); 2299 return Legalized; 2300 2301 case TargetOpcode::G_UDIVREM: 2302 Observer.changingInstr(MI); 2303 widenScalarSrc(MI, WideTy, 2, TargetOpcode::G_ZEXT); 2304 widenScalarSrc(MI, WideTy, 3, TargetOpcode::G_ZEXT); 2305 widenScalarDst(MI, WideTy); 2306 widenScalarDst(MI, WideTy, 1); 2307 Observer.changedInstr(MI); 2308 return Legalized; 2309 2310 case TargetOpcode::G_SELECT: 2311 Observer.changingInstr(MI); 2312 if (TypeIdx == 0) { 2313 // Perform operation at larger width (any extension is fine here, high 2314 // bits don't affect the result) and then truncate the result back to the 2315 // original type. 2316 widenScalarSrc(MI, WideTy, 2, TargetOpcode::G_ANYEXT); 2317 widenScalarSrc(MI, WideTy, 3, TargetOpcode::G_ANYEXT); 2318 widenScalarDst(MI, WideTy); 2319 } else { 2320 bool IsVec = MRI.getType(MI.getOperand(1).getReg()).isVector(); 2321 // Explicit extension is required here since high bits affect the result. 2322 widenScalarSrc(MI, WideTy, 1, MIRBuilder.getBoolExtOp(IsVec, false)); 2323 } 2324 Observer.changedInstr(MI); 2325 return Legalized; 2326 2327 case TargetOpcode::G_FPTOSI: 2328 case TargetOpcode::G_FPTOUI: 2329 Observer.changingInstr(MI); 2330 2331 if (TypeIdx == 0) 2332 widenScalarDst(MI, WideTy); 2333 else 2334 widenScalarSrc(MI, WideTy, 1, TargetOpcode::G_FPEXT); 2335 2336 Observer.changedInstr(MI); 2337 return Legalized; 2338 case TargetOpcode::G_SITOFP: 2339 Observer.changingInstr(MI); 2340 2341 if (TypeIdx == 0) 2342 widenScalarDst(MI, WideTy, 0, TargetOpcode::G_FPTRUNC); 2343 else 2344 widenScalarSrc(MI, WideTy, 1, TargetOpcode::G_SEXT); 2345 2346 Observer.changedInstr(MI); 2347 return Legalized; 2348 case TargetOpcode::G_UITOFP: 2349 Observer.changingInstr(MI); 2350 2351 if (TypeIdx == 0) 2352 widenScalarDst(MI, WideTy, 0, TargetOpcode::G_FPTRUNC); 2353 else 2354 widenScalarSrc(MI, WideTy, 1, TargetOpcode::G_ZEXT); 2355 2356 Observer.changedInstr(MI); 2357 return Legalized; 2358 case TargetOpcode::G_LOAD: 2359 case TargetOpcode::G_SEXTLOAD: 2360 case TargetOpcode::G_ZEXTLOAD: 2361 Observer.changingInstr(MI); 2362 widenScalarDst(MI, WideTy); 2363 Observer.changedInstr(MI); 2364 return Legalized; 2365 2366 case TargetOpcode::G_STORE: { 2367 if (TypeIdx != 0) 2368 return UnableToLegalize; 2369 2370 LLT Ty = MRI.getType(MI.getOperand(0).getReg()); 2371 if (!Ty.isScalar()) 2372 return UnableToLegalize; 2373 2374 Observer.changingInstr(MI); 2375 2376 unsigned ExtType = Ty.getScalarSizeInBits() == 1 ? 2377 TargetOpcode::G_ZEXT : TargetOpcode::G_ANYEXT; 2378 widenScalarSrc(MI, WideTy, 0, ExtType); 2379 2380 Observer.changedInstr(MI); 2381 return Legalized; 2382 } 2383 case TargetOpcode::G_CONSTANT: { 2384 MachineOperand &SrcMO = MI.getOperand(1); 2385 LLVMContext &Ctx = MIRBuilder.getMF().getFunction().getContext(); 2386 unsigned ExtOpc = LI.getExtOpcodeForWideningConstant( 2387 MRI.getType(MI.getOperand(0).getReg())); 2388 assert((ExtOpc == TargetOpcode::G_ZEXT || ExtOpc == TargetOpcode::G_SEXT || 2389 ExtOpc == TargetOpcode::G_ANYEXT) && 2390 "Illegal Extend"); 2391 const APInt &SrcVal = SrcMO.getCImm()->getValue(); 2392 const APInt &Val = (ExtOpc == TargetOpcode::G_SEXT) 2393 ? SrcVal.sext(WideTy.getSizeInBits()) 2394 : SrcVal.zext(WideTy.getSizeInBits()); 2395 Observer.changingInstr(MI); 2396 SrcMO.setCImm(ConstantInt::get(Ctx, Val)); 2397 2398 widenScalarDst(MI, WideTy); 2399 Observer.changedInstr(MI); 2400 return Legalized; 2401 } 2402 case TargetOpcode::G_FCONSTANT: { 2403 // To avoid changing the bits of the constant due to extension to a larger 2404 // type and then using G_FPTRUNC, we simply convert to a G_CONSTANT. 2405 MachineOperand &SrcMO = MI.getOperand(1); 2406 APInt Val = SrcMO.getFPImm()->getValueAPF().bitcastToAPInt(); 2407 MIRBuilder.setInstrAndDebugLoc(MI); 2408 auto IntCst = MIRBuilder.buildConstant(MI.getOperand(0).getReg(), Val); 2409 widenScalarDst(*IntCst, WideTy, 0, TargetOpcode::G_TRUNC); 2410 MI.eraseFromParent(); 2411 return Legalized; 2412 } 2413 case TargetOpcode::G_IMPLICIT_DEF: { 2414 Observer.changingInstr(MI); 2415 widenScalarDst(MI, WideTy); 2416 Observer.changedInstr(MI); 2417 return Legalized; 2418 } 2419 case TargetOpcode::G_BRCOND: 2420 Observer.changingInstr(MI); 2421 widenScalarSrc(MI, WideTy, 0, MIRBuilder.getBoolExtOp(false, false)); 2422 Observer.changedInstr(MI); 2423 return Legalized; 2424 2425 case TargetOpcode::G_FCMP: 2426 Observer.changingInstr(MI); 2427 if (TypeIdx == 0) 2428 widenScalarDst(MI, WideTy); 2429 else { 2430 widenScalarSrc(MI, WideTy, 2, TargetOpcode::G_FPEXT); 2431 widenScalarSrc(MI, WideTy, 3, TargetOpcode::G_FPEXT); 2432 } 2433 Observer.changedInstr(MI); 2434 return Legalized; 2435 2436 case TargetOpcode::G_ICMP: 2437 Observer.changingInstr(MI); 2438 if (TypeIdx == 0) 2439 widenScalarDst(MI, WideTy); 2440 else { 2441 unsigned ExtOpcode = CmpInst::isSigned(static_cast<CmpInst::Predicate>( 2442 MI.getOperand(1).getPredicate())) 2443 ? TargetOpcode::G_SEXT 2444 : TargetOpcode::G_ZEXT; 2445 widenScalarSrc(MI, WideTy, 2, ExtOpcode); 2446 widenScalarSrc(MI, WideTy, 3, ExtOpcode); 2447 } 2448 Observer.changedInstr(MI); 2449 return Legalized; 2450 2451 case TargetOpcode::G_PTR_ADD: 2452 assert(TypeIdx == 1 && "unable to legalize pointer of G_PTR_ADD"); 2453 Observer.changingInstr(MI); 2454 widenScalarSrc(MI, WideTy, 2, TargetOpcode::G_SEXT); 2455 Observer.changedInstr(MI); 2456 return Legalized; 2457 2458 case TargetOpcode::G_PHI: { 2459 assert(TypeIdx == 0 && "Expecting only Idx 0"); 2460 2461 Observer.changingInstr(MI); 2462 for (unsigned I = 1; I < MI.getNumOperands(); I += 2) { 2463 MachineBasicBlock &OpMBB = *MI.getOperand(I + 1).getMBB(); 2464 MIRBuilder.setInsertPt(OpMBB, OpMBB.getFirstTerminatorForward()); 2465 widenScalarSrc(MI, WideTy, I, TargetOpcode::G_ANYEXT); 2466 } 2467 2468 MachineBasicBlock &MBB = *MI.getParent(); 2469 MIRBuilder.setInsertPt(MBB, --MBB.getFirstNonPHI()); 2470 widenScalarDst(MI, WideTy); 2471 Observer.changedInstr(MI); 2472 return Legalized; 2473 } 2474 case TargetOpcode::G_EXTRACT_VECTOR_ELT: { 2475 if (TypeIdx == 0) { 2476 Register VecReg = MI.getOperand(1).getReg(); 2477 LLT VecTy = MRI.getType(VecReg); 2478 Observer.changingInstr(MI); 2479 2480 widenScalarSrc( 2481 MI, LLT::vector(VecTy.getElementCount(), WideTy.getSizeInBits()), 1, 2482 TargetOpcode::G_ANYEXT); 2483 2484 widenScalarDst(MI, WideTy, 0); 2485 Observer.changedInstr(MI); 2486 return Legalized; 2487 } 2488 2489 if (TypeIdx != 2) 2490 return UnableToLegalize; 2491 Observer.changingInstr(MI); 2492 // TODO: Probably should be zext 2493 widenScalarSrc(MI, WideTy, 2, TargetOpcode::G_SEXT); 2494 Observer.changedInstr(MI); 2495 return Legalized; 2496 } 2497 case TargetOpcode::G_INSERT_VECTOR_ELT: { 2498 if (TypeIdx == 1) { 2499 Observer.changingInstr(MI); 2500 2501 Register VecReg = MI.getOperand(1).getReg(); 2502 LLT VecTy = MRI.getType(VecReg); 2503 LLT WideVecTy = LLT::vector(VecTy.getElementCount(), WideTy); 2504 2505 widenScalarSrc(MI, WideVecTy, 1, TargetOpcode::G_ANYEXT); 2506 widenScalarSrc(MI, WideTy, 2, TargetOpcode::G_ANYEXT); 2507 widenScalarDst(MI, WideVecTy, 0); 2508 Observer.changedInstr(MI); 2509 return Legalized; 2510 } 2511 2512 if (TypeIdx == 2) { 2513 Observer.changingInstr(MI); 2514 // TODO: Probably should be zext 2515 widenScalarSrc(MI, WideTy, 3, TargetOpcode::G_SEXT); 2516 Observer.changedInstr(MI); 2517 return Legalized; 2518 } 2519 2520 return UnableToLegalize; 2521 } 2522 case TargetOpcode::G_FADD: 2523 case TargetOpcode::G_FMUL: 2524 case TargetOpcode::G_FSUB: 2525 case TargetOpcode::G_FMA: 2526 case TargetOpcode::G_FMAD: 2527 case TargetOpcode::G_FNEG: 2528 case TargetOpcode::G_FABS: 2529 case TargetOpcode::G_FCANONICALIZE: 2530 case TargetOpcode::G_FMINNUM: 2531 case TargetOpcode::G_FMAXNUM: 2532 case TargetOpcode::G_FMINNUM_IEEE: 2533 case TargetOpcode::G_FMAXNUM_IEEE: 2534 case TargetOpcode::G_FMINIMUM: 2535 case TargetOpcode::G_FMAXIMUM: 2536 case TargetOpcode::G_FDIV: 2537 case TargetOpcode::G_FREM: 2538 case TargetOpcode::G_FCEIL: 2539 case TargetOpcode::G_FFLOOR: 2540 case TargetOpcode::G_FCOS: 2541 case TargetOpcode::G_FSIN: 2542 case TargetOpcode::G_FLOG10: 2543 case TargetOpcode::G_FLOG: 2544 case TargetOpcode::G_FLOG2: 2545 case TargetOpcode::G_FRINT: 2546 case TargetOpcode::G_FNEARBYINT: 2547 case TargetOpcode::G_FSQRT: 2548 case TargetOpcode::G_FEXP: 2549 case TargetOpcode::G_FEXP2: 2550 case TargetOpcode::G_FPOW: 2551 case TargetOpcode::G_INTRINSIC_TRUNC: 2552 case TargetOpcode::G_INTRINSIC_ROUND: 2553 case TargetOpcode::G_INTRINSIC_ROUNDEVEN: 2554 assert(TypeIdx == 0); 2555 Observer.changingInstr(MI); 2556 2557 for (unsigned I = 1, E = MI.getNumOperands(); I != E; ++I) 2558 widenScalarSrc(MI, WideTy, I, TargetOpcode::G_FPEXT); 2559 2560 widenScalarDst(MI, WideTy, 0, TargetOpcode::G_FPTRUNC); 2561 Observer.changedInstr(MI); 2562 return Legalized; 2563 case TargetOpcode::G_FPOWI: { 2564 if (TypeIdx != 0) 2565 return UnableToLegalize; 2566 Observer.changingInstr(MI); 2567 widenScalarSrc(MI, WideTy, 1, TargetOpcode::G_FPEXT); 2568 widenScalarDst(MI, WideTy, 0, TargetOpcode::G_FPTRUNC); 2569 Observer.changedInstr(MI); 2570 return Legalized; 2571 } 2572 case TargetOpcode::G_INTTOPTR: 2573 if (TypeIdx != 1) 2574 return UnableToLegalize; 2575 2576 Observer.changingInstr(MI); 2577 widenScalarSrc(MI, WideTy, 1, TargetOpcode::G_ZEXT); 2578 Observer.changedInstr(MI); 2579 return Legalized; 2580 case TargetOpcode::G_PTRTOINT: 2581 if (TypeIdx != 0) 2582 return UnableToLegalize; 2583 2584 Observer.changingInstr(MI); 2585 widenScalarDst(MI, WideTy, 0); 2586 Observer.changedInstr(MI); 2587 return Legalized; 2588 case TargetOpcode::G_BUILD_VECTOR: { 2589 Observer.changingInstr(MI); 2590 2591 const LLT WideEltTy = TypeIdx == 1 ? WideTy : WideTy.getElementType(); 2592 for (int I = 1, E = MI.getNumOperands(); I != E; ++I) 2593 widenScalarSrc(MI, WideEltTy, I, TargetOpcode::G_ANYEXT); 2594 2595 // Avoid changing the result vector type if the source element type was 2596 // requested. 2597 if (TypeIdx == 1) { 2598 MI.setDesc(MIRBuilder.getTII().get(TargetOpcode::G_BUILD_VECTOR_TRUNC)); 2599 } else { 2600 widenScalarDst(MI, WideTy, 0); 2601 } 2602 2603 Observer.changedInstr(MI); 2604 return Legalized; 2605 } 2606 case TargetOpcode::G_SEXT_INREG: 2607 if (TypeIdx != 0) 2608 return UnableToLegalize; 2609 2610 Observer.changingInstr(MI); 2611 widenScalarSrc(MI, WideTy, 1, TargetOpcode::G_ANYEXT); 2612 widenScalarDst(MI, WideTy, 0, TargetOpcode::G_TRUNC); 2613 Observer.changedInstr(MI); 2614 return Legalized; 2615 case TargetOpcode::G_PTRMASK: { 2616 if (TypeIdx != 1) 2617 return UnableToLegalize; 2618 Observer.changingInstr(MI); 2619 widenScalarSrc(MI, WideTy, 2, TargetOpcode::G_ZEXT); 2620 Observer.changedInstr(MI); 2621 return Legalized; 2622 } 2623 } 2624 } 2625 2626 static void getUnmergePieces(SmallVectorImpl<Register> &Pieces, 2627 MachineIRBuilder &B, Register Src, LLT Ty) { 2628 auto Unmerge = B.buildUnmerge(Ty, Src); 2629 for (int I = 0, E = Unmerge->getNumOperands() - 1; I != E; ++I) 2630 Pieces.push_back(Unmerge.getReg(I)); 2631 } 2632 2633 LegalizerHelper::LegalizeResult 2634 LegalizerHelper::lowerBitcast(MachineInstr &MI) { 2635 Register Dst = MI.getOperand(0).getReg(); 2636 Register Src = MI.getOperand(1).getReg(); 2637 LLT DstTy = MRI.getType(Dst); 2638 LLT SrcTy = MRI.getType(Src); 2639 2640 if (SrcTy.isVector()) { 2641 LLT SrcEltTy = SrcTy.getElementType(); 2642 SmallVector<Register, 8> SrcRegs; 2643 2644 if (DstTy.isVector()) { 2645 int NumDstElt = DstTy.getNumElements(); 2646 int NumSrcElt = SrcTy.getNumElements(); 2647 2648 LLT DstEltTy = DstTy.getElementType(); 2649 LLT DstCastTy = DstEltTy; // Intermediate bitcast result type 2650 LLT SrcPartTy = SrcEltTy; // Original unmerge result type. 2651 2652 // If there's an element size mismatch, insert intermediate casts to match 2653 // the result element type. 2654 if (NumSrcElt < NumDstElt) { // Source element type is larger. 2655 // %1:_(<4 x s8>) = G_BITCAST %0:_(<2 x s16>) 2656 // 2657 // => 2658 // 2659 // %2:_(s16), %3:_(s16) = G_UNMERGE_VALUES %0 2660 // %3:_(<2 x s8>) = G_BITCAST %2 2661 // %4:_(<2 x s8>) = G_BITCAST %3 2662 // %1:_(<4 x s16>) = G_CONCAT_VECTORS %3, %4 2663 DstCastTy = LLT::fixed_vector(NumDstElt / NumSrcElt, DstEltTy); 2664 SrcPartTy = SrcEltTy; 2665 } else if (NumSrcElt > NumDstElt) { // Source element type is smaller. 2666 // 2667 // %1:_(<2 x s16>) = G_BITCAST %0:_(<4 x s8>) 2668 // 2669 // => 2670 // 2671 // %2:_(<2 x s8>), %3:_(<2 x s8>) = G_UNMERGE_VALUES %0 2672 // %3:_(s16) = G_BITCAST %2 2673 // %4:_(s16) = G_BITCAST %3 2674 // %1:_(<2 x s16>) = G_BUILD_VECTOR %3, %4 2675 SrcPartTy = LLT::fixed_vector(NumSrcElt / NumDstElt, SrcEltTy); 2676 DstCastTy = DstEltTy; 2677 } 2678 2679 getUnmergePieces(SrcRegs, MIRBuilder, Src, SrcPartTy); 2680 for (Register &SrcReg : SrcRegs) 2681 SrcReg = MIRBuilder.buildBitcast(DstCastTy, SrcReg).getReg(0); 2682 } else 2683 getUnmergePieces(SrcRegs, MIRBuilder, Src, SrcEltTy); 2684 2685 MIRBuilder.buildMergeLikeInstr(Dst, SrcRegs); 2686 MI.eraseFromParent(); 2687 return Legalized; 2688 } 2689 2690 if (DstTy.isVector()) { 2691 SmallVector<Register, 8> SrcRegs; 2692 getUnmergePieces(SrcRegs, MIRBuilder, Src, DstTy.getElementType()); 2693 MIRBuilder.buildMergeLikeInstr(Dst, SrcRegs); 2694 MI.eraseFromParent(); 2695 return Legalized; 2696 } 2697 2698 return UnableToLegalize; 2699 } 2700 2701 /// Figure out the bit offset into a register when coercing a vector index for 2702 /// the wide element type. This is only for the case when promoting vector to 2703 /// one with larger elements. 2704 // 2705 /// 2706 /// %offset_idx = G_AND %idx, ~(-1 << Log2(DstEltSize / SrcEltSize)) 2707 /// %offset_bits = G_SHL %offset_idx, Log2(SrcEltSize) 2708 static Register getBitcastWiderVectorElementOffset(MachineIRBuilder &B, 2709 Register Idx, 2710 unsigned NewEltSize, 2711 unsigned OldEltSize) { 2712 const unsigned Log2EltRatio = Log2_32(NewEltSize / OldEltSize); 2713 LLT IdxTy = B.getMRI()->getType(Idx); 2714 2715 // Now figure out the amount we need to shift to get the target bits. 2716 auto OffsetMask = B.buildConstant( 2717 IdxTy, ~(APInt::getAllOnes(IdxTy.getSizeInBits()) << Log2EltRatio)); 2718 auto OffsetIdx = B.buildAnd(IdxTy, Idx, OffsetMask); 2719 return B.buildShl(IdxTy, OffsetIdx, 2720 B.buildConstant(IdxTy, Log2_32(OldEltSize))).getReg(0); 2721 } 2722 2723 /// Perform a G_EXTRACT_VECTOR_ELT in a different sized vector element. If this 2724 /// is casting to a vector with a smaller element size, perform multiple element 2725 /// extracts and merge the results. If this is coercing to a vector with larger 2726 /// elements, index the bitcasted vector and extract the target element with bit 2727 /// operations. This is intended to force the indexing in the native register 2728 /// size for architectures that can dynamically index the register file. 2729 LegalizerHelper::LegalizeResult 2730 LegalizerHelper::bitcastExtractVectorElt(MachineInstr &MI, unsigned TypeIdx, 2731 LLT CastTy) { 2732 if (TypeIdx != 1) 2733 return UnableToLegalize; 2734 2735 Register Dst = MI.getOperand(0).getReg(); 2736 Register SrcVec = MI.getOperand(1).getReg(); 2737 Register Idx = MI.getOperand(2).getReg(); 2738 LLT SrcVecTy = MRI.getType(SrcVec); 2739 LLT IdxTy = MRI.getType(Idx); 2740 2741 LLT SrcEltTy = SrcVecTy.getElementType(); 2742 unsigned NewNumElts = CastTy.isVector() ? CastTy.getNumElements() : 1; 2743 unsigned OldNumElts = SrcVecTy.getNumElements(); 2744 2745 LLT NewEltTy = CastTy.isVector() ? CastTy.getElementType() : CastTy; 2746 Register CastVec = MIRBuilder.buildBitcast(CastTy, SrcVec).getReg(0); 2747 2748 const unsigned NewEltSize = NewEltTy.getSizeInBits(); 2749 const unsigned OldEltSize = SrcEltTy.getSizeInBits(); 2750 if (NewNumElts > OldNumElts) { 2751 // Decreasing the vector element size 2752 // 2753 // e.g. i64 = extract_vector_elt x:v2i64, y:i32 2754 // => 2755 // v4i32:castx = bitcast x:v2i64 2756 // 2757 // i64 = bitcast 2758 // (v2i32 build_vector (i32 (extract_vector_elt castx, (2 * y))), 2759 // (i32 (extract_vector_elt castx, (2 * y + 1))) 2760 // 2761 if (NewNumElts % OldNumElts != 0) 2762 return UnableToLegalize; 2763 2764 // Type of the intermediate result vector. 2765 const unsigned NewEltsPerOldElt = NewNumElts / OldNumElts; 2766 LLT MidTy = 2767 LLT::scalarOrVector(ElementCount::getFixed(NewEltsPerOldElt), NewEltTy); 2768 2769 auto NewEltsPerOldEltK = MIRBuilder.buildConstant(IdxTy, NewEltsPerOldElt); 2770 2771 SmallVector<Register, 8> NewOps(NewEltsPerOldElt); 2772 auto NewBaseIdx = MIRBuilder.buildMul(IdxTy, Idx, NewEltsPerOldEltK); 2773 2774 for (unsigned I = 0; I < NewEltsPerOldElt; ++I) { 2775 auto IdxOffset = MIRBuilder.buildConstant(IdxTy, I); 2776 auto TmpIdx = MIRBuilder.buildAdd(IdxTy, NewBaseIdx, IdxOffset); 2777 auto Elt = MIRBuilder.buildExtractVectorElement(NewEltTy, CastVec, TmpIdx); 2778 NewOps[I] = Elt.getReg(0); 2779 } 2780 2781 auto NewVec = MIRBuilder.buildBuildVector(MidTy, NewOps); 2782 MIRBuilder.buildBitcast(Dst, NewVec); 2783 MI.eraseFromParent(); 2784 return Legalized; 2785 } 2786 2787 if (NewNumElts < OldNumElts) { 2788 if (NewEltSize % OldEltSize != 0) 2789 return UnableToLegalize; 2790 2791 // This only depends on powers of 2 because we use bit tricks to figure out 2792 // the bit offset we need to shift to get the target element. A general 2793 // expansion could emit division/multiply. 2794 if (!isPowerOf2_32(NewEltSize / OldEltSize)) 2795 return UnableToLegalize; 2796 2797 // Increasing the vector element size. 2798 // %elt:_(small_elt) = G_EXTRACT_VECTOR_ELT %vec:_(<N x small_elt>), %idx 2799 // 2800 // => 2801 // 2802 // %cast = G_BITCAST %vec 2803 // %scaled_idx = G_LSHR %idx, Log2(DstEltSize / SrcEltSize) 2804 // %wide_elt = G_EXTRACT_VECTOR_ELT %cast, %scaled_idx 2805 // %offset_idx = G_AND %idx, ~(-1 << Log2(DstEltSize / SrcEltSize)) 2806 // %offset_bits = G_SHL %offset_idx, Log2(SrcEltSize) 2807 // %elt_bits = G_LSHR %wide_elt, %offset_bits 2808 // %elt = G_TRUNC %elt_bits 2809 2810 const unsigned Log2EltRatio = Log2_32(NewEltSize / OldEltSize); 2811 auto Log2Ratio = MIRBuilder.buildConstant(IdxTy, Log2EltRatio); 2812 2813 // Divide to get the index in the wider element type. 2814 auto ScaledIdx = MIRBuilder.buildLShr(IdxTy, Idx, Log2Ratio); 2815 2816 Register WideElt = CastVec; 2817 if (CastTy.isVector()) { 2818 WideElt = MIRBuilder.buildExtractVectorElement(NewEltTy, CastVec, 2819 ScaledIdx).getReg(0); 2820 } 2821 2822 // Compute the bit offset into the register of the target element. 2823 Register OffsetBits = getBitcastWiderVectorElementOffset( 2824 MIRBuilder, Idx, NewEltSize, OldEltSize); 2825 2826 // Shift the wide element to get the target element. 2827 auto ExtractedBits = MIRBuilder.buildLShr(NewEltTy, WideElt, OffsetBits); 2828 MIRBuilder.buildTrunc(Dst, ExtractedBits); 2829 MI.eraseFromParent(); 2830 return Legalized; 2831 } 2832 2833 return UnableToLegalize; 2834 } 2835 2836 /// Emit code to insert \p InsertReg into \p TargetRet at \p OffsetBits in \p 2837 /// TargetReg, while preserving other bits in \p TargetReg. 2838 /// 2839 /// (InsertReg << Offset) | (TargetReg & ~(-1 >> InsertReg.size()) << Offset) 2840 static Register buildBitFieldInsert(MachineIRBuilder &B, 2841 Register TargetReg, Register InsertReg, 2842 Register OffsetBits) { 2843 LLT TargetTy = B.getMRI()->getType(TargetReg); 2844 LLT InsertTy = B.getMRI()->getType(InsertReg); 2845 auto ZextVal = B.buildZExt(TargetTy, InsertReg); 2846 auto ShiftedInsertVal = B.buildShl(TargetTy, ZextVal, OffsetBits); 2847 2848 // Produce a bitmask of the value to insert 2849 auto EltMask = B.buildConstant( 2850 TargetTy, APInt::getLowBitsSet(TargetTy.getSizeInBits(), 2851 InsertTy.getSizeInBits())); 2852 // Shift it into position 2853 auto ShiftedMask = B.buildShl(TargetTy, EltMask, OffsetBits); 2854 auto InvShiftedMask = B.buildNot(TargetTy, ShiftedMask); 2855 2856 // Clear out the bits in the wide element 2857 auto MaskedOldElt = B.buildAnd(TargetTy, TargetReg, InvShiftedMask); 2858 2859 // The value to insert has all zeros already, so stick it into the masked 2860 // wide element. 2861 return B.buildOr(TargetTy, MaskedOldElt, ShiftedInsertVal).getReg(0); 2862 } 2863 2864 /// Perform a G_INSERT_VECTOR_ELT in a different sized vector element. If this 2865 /// is increasing the element size, perform the indexing in the target element 2866 /// type, and use bit operations to insert at the element position. This is 2867 /// intended for architectures that can dynamically index the register file and 2868 /// want to force indexing in the native register size. 2869 LegalizerHelper::LegalizeResult 2870 LegalizerHelper::bitcastInsertVectorElt(MachineInstr &MI, unsigned TypeIdx, 2871 LLT CastTy) { 2872 if (TypeIdx != 0) 2873 return UnableToLegalize; 2874 2875 Register Dst = MI.getOperand(0).getReg(); 2876 Register SrcVec = MI.getOperand(1).getReg(); 2877 Register Val = MI.getOperand(2).getReg(); 2878 Register Idx = MI.getOperand(3).getReg(); 2879 2880 LLT VecTy = MRI.getType(Dst); 2881 LLT IdxTy = MRI.getType(Idx); 2882 2883 LLT VecEltTy = VecTy.getElementType(); 2884 LLT NewEltTy = CastTy.isVector() ? CastTy.getElementType() : CastTy; 2885 const unsigned NewEltSize = NewEltTy.getSizeInBits(); 2886 const unsigned OldEltSize = VecEltTy.getSizeInBits(); 2887 2888 unsigned NewNumElts = CastTy.isVector() ? CastTy.getNumElements() : 1; 2889 unsigned OldNumElts = VecTy.getNumElements(); 2890 2891 Register CastVec = MIRBuilder.buildBitcast(CastTy, SrcVec).getReg(0); 2892 if (NewNumElts < OldNumElts) { 2893 if (NewEltSize % OldEltSize != 0) 2894 return UnableToLegalize; 2895 2896 // This only depends on powers of 2 because we use bit tricks to figure out 2897 // the bit offset we need to shift to get the target element. A general 2898 // expansion could emit division/multiply. 2899 if (!isPowerOf2_32(NewEltSize / OldEltSize)) 2900 return UnableToLegalize; 2901 2902 const unsigned Log2EltRatio = Log2_32(NewEltSize / OldEltSize); 2903 auto Log2Ratio = MIRBuilder.buildConstant(IdxTy, Log2EltRatio); 2904 2905 // Divide to get the index in the wider element type. 2906 auto ScaledIdx = MIRBuilder.buildLShr(IdxTy, Idx, Log2Ratio); 2907 2908 Register ExtractedElt = CastVec; 2909 if (CastTy.isVector()) { 2910 ExtractedElt = MIRBuilder.buildExtractVectorElement(NewEltTy, CastVec, 2911 ScaledIdx).getReg(0); 2912 } 2913 2914 // Compute the bit offset into the register of the target element. 2915 Register OffsetBits = getBitcastWiderVectorElementOffset( 2916 MIRBuilder, Idx, NewEltSize, OldEltSize); 2917 2918 Register InsertedElt = buildBitFieldInsert(MIRBuilder, ExtractedElt, 2919 Val, OffsetBits); 2920 if (CastTy.isVector()) { 2921 InsertedElt = MIRBuilder.buildInsertVectorElement( 2922 CastTy, CastVec, InsertedElt, ScaledIdx).getReg(0); 2923 } 2924 2925 MIRBuilder.buildBitcast(Dst, InsertedElt); 2926 MI.eraseFromParent(); 2927 return Legalized; 2928 } 2929 2930 return UnableToLegalize; 2931 } 2932 2933 LegalizerHelper::LegalizeResult LegalizerHelper::lowerLoad(GAnyLoad &LoadMI) { 2934 // Lower to a memory-width G_LOAD and a G_SEXT/G_ZEXT/G_ANYEXT 2935 Register DstReg = LoadMI.getDstReg(); 2936 Register PtrReg = LoadMI.getPointerReg(); 2937 LLT DstTy = MRI.getType(DstReg); 2938 MachineMemOperand &MMO = LoadMI.getMMO(); 2939 LLT MemTy = MMO.getMemoryType(); 2940 MachineFunction &MF = MIRBuilder.getMF(); 2941 2942 unsigned MemSizeInBits = MemTy.getSizeInBits(); 2943 unsigned MemStoreSizeInBits = 8 * MemTy.getSizeInBytes(); 2944 2945 if (MemSizeInBits != MemStoreSizeInBits) { 2946 if (MemTy.isVector()) 2947 return UnableToLegalize; 2948 2949 // Promote to a byte-sized load if not loading an integral number of 2950 // bytes. For example, promote EXTLOAD:i20 -> EXTLOAD:i24. 2951 LLT WideMemTy = LLT::scalar(MemStoreSizeInBits); 2952 MachineMemOperand *NewMMO = 2953 MF.getMachineMemOperand(&MMO, MMO.getPointerInfo(), WideMemTy); 2954 2955 Register LoadReg = DstReg; 2956 LLT LoadTy = DstTy; 2957 2958 // If this wasn't already an extending load, we need to widen the result 2959 // register to avoid creating a load with a narrower result than the source. 2960 if (MemStoreSizeInBits > DstTy.getSizeInBits()) { 2961 LoadTy = WideMemTy; 2962 LoadReg = MRI.createGenericVirtualRegister(WideMemTy); 2963 } 2964 2965 if (isa<GSExtLoad>(LoadMI)) { 2966 auto NewLoad = MIRBuilder.buildLoad(LoadTy, PtrReg, *NewMMO); 2967 MIRBuilder.buildSExtInReg(LoadReg, NewLoad, MemSizeInBits); 2968 } else if (isa<GZExtLoad>(LoadMI) || WideMemTy == LoadTy) { 2969 auto NewLoad = MIRBuilder.buildLoad(LoadTy, PtrReg, *NewMMO); 2970 // The extra bits are guaranteed to be zero, since we stored them that 2971 // way. A zext load from Wide thus automatically gives zext from MemVT. 2972 MIRBuilder.buildAssertZExt(LoadReg, NewLoad, MemSizeInBits); 2973 } else { 2974 MIRBuilder.buildLoad(LoadReg, PtrReg, *NewMMO); 2975 } 2976 2977 if (DstTy != LoadTy) 2978 MIRBuilder.buildTrunc(DstReg, LoadReg); 2979 2980 LoadMI.eraseFromParent(); 2981 return Legalized; 2982 } 2983 2984 // Big endian lowering not implemented. 2985 if (MIRBuilder.getDataLayout().isBigEndian()) 2986 return UnableToLegalize; 2987 2988 // This load needs splitting into power of 2 sized loads. 2989 // 2990 // Our strategy here is to generate anyextending loads for the smaller 2991 // types up to next power-2 result type, and then combine the two larger 2992 // result values together, before truncating back down to the non-pow-2 2993 // type. 2994 // E.g. v1 = i24 load => 2995 // v2 = i32 zextload (2 byte) 2996 // v3 = i32 load (1 byte) 2997 // v4 = i32 shl v3, 16 2998 // v5 = i32 or v4, v2 2999 // v1 = i24 trunc v5 3000 // By doing this we generate the correct truncate which should get 3001 // combined away as an artifact with a matching extend. 3002 3003 uint64_t LargeSplitSize, SmallSplitSize; 3004 3005 if (!isPowerOf2_32(MemSizeInBits)) { 3006 // This load needs splitting into power of 2 sized loads. 3007 LargeSplitSize = PowerOf2Floor(MemSizeInBits); 3008 SmallSplitSize = MemSizeInBits - LargeSplitSize; 3009 } else { 3010 // This is already a power of 2, but we still need to split this in half. 3011 // 3012 // Assume we're being asked to decompose an unaligned load. 3013 // TODO: If this requires multiple splits, handle them all at once. 3014 auto &Ctx = MF.getFunction().getContext(); 3015 if (TLI.allowsMemoryAccess(Ctx, MIRBuilder.getDataLayout(), MemTy, MMO)) 3016 return UnableToLegalize; 3017 3018 SmallSplitSize = LargeSplitSize = MemSizeInBits / 2; 3019 } 3020 3021 if (MemTy.isVector()) { 3022 // TODO: Handle vector extloads 3023 if (MemTy != DstTy) 3024 return UnableToLegalize; 3025 3026 // TODO: We can do better than scalarizing the vector and at least split it 3027 // in half. 3028 return reduceLoadStoreWidth(LoadMI, 0, DstTy.getElementType()); 3029 } 3030 3031 MachineMemOperand *LargeMMO = 3032 MF.getMachineMemOperand(&MMO, 0, LargeSplitSize / 8); 3033 MachineMemOperand *SmallMMO = 3034 MF.getMachineMemOperand(&MMO, LargeSplitSize / 8, SmallSplitSize / 8); 3035 3036 LLT PtrTy = MRI.getType(PtrReg); 3037 unsigned AnyExtSize = PowerOf2Ceil(DstTy.getSizeInBits()); 3038 LLT AnyExtTy = LLT::scalar(AnyExtSize); 3039 auto LargeLoad = MIRBuilder.buildLoadInstr(TargetOpcode::G_ZEXTLOAD, AnyExtTy, 3040 PtrReg, *LargeMMO); 3041 3042 auto OffsetCst = MIRBuilder.buildConstant(LLT::scalar(PtrTy.getSizeInBits()), 3043 LargeSplitSize / 8); 3044 Register PtrAddReg = MRI.createGenericVirtualRegister(PtrTy); 3045 auto SmallPtr = MIRBuilder.buildPtrAdd(PtrAddReg, PtrReg, OffsetCst); 3046 auto SmallLoad = MIRBuilder.buildLoadInstr(LoadMI.getOpcode(), AnyExtTy, 3047 SmallPtr, *SmallMMO); 3048 3049 auto ShiftAmt = MIRBuilder.buildConstant(AnyExtTy, LargeSplitSize); 3050 auto Shift = MIRBuilder.buildShl(AnyExtTy, SmallLoad, ShiftAmt); 3051 3052 if (AnyExtTy == DstTy) 3053 MIRBuilder.buildOr(DstReg, Shift, LargeLoad); 3054 else if (AnyExtTy.getSizeInBits() != DstTy.getSizeInBits()) { 3055 auto Or = MIRBuilder.buildOr(AnyExtTy, Shift, LargeLoad); 3056 MIRBuilder.buildTrunc(DstReg, {Or}); 3057 } else { 3058 assert(DstTy.isPointer() && "expected pointer"); 3059 auto Or = MIRBuilder.buildOr(AnyExtTy, Shift, LargeLoad); 3060 3061 // FIXME: We currently consider this to be illegal for non-integral address 3062 // spaces, but we need still need a way to reinterpret the bits. 3063 MIRBuilder.buildIntToPtr(DstReg, Or); 3064 } 3065 3066 LoadMI.eraseFromParent(); 3067 return Legalized; 3068 } 3069 3070 LegalizerHelper::LegalizeResult LegalizerHelper::lowerStore(GStore &StoreMI) { 3071 // Lower a non-power of 2 store into multiple pow-2 stores. 3072 // E.g. split an i24 store into an i16 store + i8 store. 3073 // We do this by first extending the stored value to the next largest power 3074 // of 2 type, and then using truncating stores to store the components. 3075 // By doing this, likewise with G_LOAD, generate an extend that can be 3076 // artifact-combined away instead of leaving behind extracts. 3077 Register SrcReg = StoreMI.getValueReg(); 3078 Register PtrReg = StoreMI.getPointerReg(); 3079 LLT SrcTy = MRI.getType(SrcReg); 3080 MachineFunction &MF = MIRBuilder.getMF(); 3081 MachineMemOperand &MMO = **StoreMI.memoperands_begin(); 3082 LLT MemTy = MMO.getMemoryType(); 3083 3084 unsigned StoreWidth = MemTy.getSizeInBits(); 3085 unsigned StoreSizeInBits = 8 * MemTy.getSizeInBytes(); 3086 3087 if (StoreWidth != StoreSizeInBits) { 3088 if (SrcTy.isVector()) 3089 return UnableToLegalize; 3090 3091 // Promote to a byte-sized store with upper bits zero if not 3092 // storing an integral number of bytes. For example, promote 3093 // TRUNCSTORE:i1 X -> TRUNCSTORE:i8 (and X, 1) 3094 LLT WideTy = LLT::scalar(StoreSizeInBits); 3095 3096 if (StoreSizeInBits > SrcTy.getSizeInBits()) { 3097 // Avoid creating a store with a narrower source than result. 3098 SrcReg = MIRBuilder.buildAnyExt(WideTy, SrcReg).getReg(0); 3099 SrcTy = WideTy; 3100 } 3101 3102 auto ZextInReg = MIRBuilder.buildZExtInReg(SrcTy, SrcReg, StoreWidth); 3103 3104 MachineMemOperand *NewMMO = 3105 MF.getMachineMemOperand(&MMO, MMO.getPointerInfo(), WideTy); 3106 MIRBuilder.buildStore(ZextInReg, PtrReg, *NewMMO); 3107 StoreMI.eraseFromParent(); 3108 return Legalized; 3109 } 3110 3111 if (MemTy.isVector()) { 3112 // TODO: Handle vector trunc stores 3113 if (MemTy != SrcTy) 3114 return UnableToLegalize; 3115 3116 // TODO: We can do better than scalarizing the vector and at least split it 3117 // in half. 3118 return reduceLoadStoreWidth(StoreMI, 0, SrcTy.getElementType()); 3119 } 3120 3121 unsigned MemSizeInBits = MemTy.getSizeInBits(); 3122 uint64_t LargeSplitSize, SmallSplitSize; 3123 3124 if (!isPowerOf2_32(MemSizeInBits)) { 3125 LargeSplitSize = PowerOf2Floor(MemTy.getSizeInBits()); 3126 SmallSplitSize = MemTy.getSizeInBits() - LargeSplitSize; 3127 } else { 3128 auto &Ctx = MF.getFunction().getContext(); 3129 if (TLI.allowsMemoryAccess(Ctx, MIRBuilder.getDataLayout(), MemTy, MMO)) 3130 return UnableToLegalize; // Don't know what we're being asked to do. 3131 3132 SmallSplitSize = LargeSplitSize = MemSizeInBits / 2; 3133 } 3134 3135 // Extend to the next pow-2. If this store was itself the result of lowering, 3136 // e.g. an s56 store being broken into s32 + s24, we might have a stored type 3137 // that's wider than the stored size. 3138 unsigned AnyExtSize = PowerOf2Ceil(MemTy.getSizeInBits()); 3139 const LLT NewSrcTy = LLT::scalar(AnyExtSize); 3140 3141 if (SrcTy.isPointer()) { 3142 const LLT IntPtrTy = LLT::scalar(SrcTy.getSizeInBits()); 3143 SrcReg = MIRBuilder.buildPtrToInt(IntPtrTy, SrcReg).getReg(0); 3144 } 3145 3146 auto ExtVal = MIRBuilder.buildAnyExtOrTrunc(NewSrcTy, SrcReg); 3147 3148 // Obtain the smaller value by shifting away the larger value. 3149 auto ShiftAmt = MIRBuilder.buildConstant(NewSrcTy, LargeSplitSize); 3150 auto SmallVal = MIRBuilder.buildLShr(NewSrcTy, ExtVal, ShiftAmt); 3151 3152 // Generate the PtrAdd and truncating stores. 3153 LLT PtrTy = MRI.getType(PtrReg); 3154 auto OffsetCst = MIRBuilder.buildConstant( 3155 LLT::scalar(PtrTy.getSizeInBits()), LargeSplitSize / 8); 3156 auto SmallPtr = 3157 MIRBuilder.buildPtrAdd(PtrTy, PtrReg, OffsetCst); 3158 3159 MachineMemOperand *LargeMMO = 3160 MF.getMachineMemOperand(&MMO, 0, LargeSplitSize / 8); 3161 MachineMemOperand *SmallMMO = 3162 MF.getMachineMemOperand(&MMO, LargeSplitSize / 8, SmallSplitSize / 8); 3163 MIRBuilder.buildStore(ExtVal, PtrReg, *LargeMMO); 3164 MIRBuilder.buildStore(SmallVal, SmallPtr, *SmallMMO); 3165 StoreMI.eraseFromParent(); 3166 return Legalized; 3167 } 3168 3169 LegalizerHelper::LegalizeResult 3170 LegalizerHelper::bitcast(MachineInstr &MI, unsigned TypeIdx, LLT CastTy) { 3171 switch (MI.getOpcode()) { 3172 case TargetOpcode::G_LOAD: { 3173 if (TypeIdx != 0) 3174 return UnableToLegalize; 3175 MachineMemOperand &MMO = **MI.memoperands_begin(); 3176 3177 // Not sure how to interpret a bitcast of an extending load. 3178 if (MMO.getMemoryType().getSizeInBits() != CastTy.getSizeInBits()) 3179 return UnableToLegalize; 3180 3181 Observer.changingInstr(MI); 3182 bitcastDst(MI, CastTy, 0); 3183 MMO.setType(CastTy); 3184 Observer.changedInstr(MI); 3185 return Legalized; 3186 } 3187 case TargetOpcode::G_STORE: { 3188 if (TypeIdx != 0) 3189 return UnableToLegalize; 3190 3191 MachineMemOperand &MMO = **MI.memoperands_begin(); 3192 3193 // Not sure how to interpret a bitcast of a truncating store. 3194 if (MMO.getMemoryType().getSizeInBits() != CastTy.getSizeInBits()) 3195 return UnableToLegalize; 3196 3197 Observer.changingInstr(MI); 3198 bitcastSrc(MI, CastTy, 0); 3199 MMO.setType(CastTy); 3200 Observer.changedInstr(MI); 3201 return Legalized; 3202 } 3203 case TargetOpcode::G_SELECT: { 3204 if (TypeIdx != 0) 3205 return UnableToLegalize; 3206 3207 if (MRI.getType(MI.getOperand(1).getReg()).isVector()) { 3208 LLVM_DEBUG( 3209 dbgs() << "bitcast action not implemented for vector select\n"); 3210 return UnableToLegalize; 3211 } 3212 3213 Observer.changingInstr(MI); 3214 bitcastSrc(MI, CastTy, 2); 3215 bitcastSrc(MI, CastTy, 3); 3216 bitcastDst(MI, CastTy, 0); 3217 Observer.changedInstr(MI); 3218 return Legalized; 3219 } 3220 case TargetOpcode::G_AND: 3221 case TargetOpcode::G_OR: 3222 case TargetOpcode::G_XOR: { 3223 Observer.changingInstr(MI); 3224 bitcastSrc(MI, CastTy, 1); 3225 bitcastSrc(MI, CastTy, 2); 3226 bitcastDst(MI, CastTy, 0); 3227 Observer.changedInstr(MI); 3228 return Legalized; 3229 } 3230 case TargetOpcode::G_EXTRACT_VECTOR_ELT: 3231 return bitcastExtractVectorElt(MI, TypeIdx, CastTy); 3232 case TargetOpcode::G_INSERT_VECTOR_ELT: 3233 return bitcastInsertVectorElt(MI, TypeIdx, CastTy); 3234 default: 3235 return UnableToLegalize; 3236 } 3237 } 3238 3239 // Legalize an instruction by changing the opcode in place. 3240 void LegalizerHelper::changeOpcode(MachineInstr &MI, unsigned NewOpcode) { 3241 Observer.changingInstr(MI); 3242 MI.setDesc(MIRBuilder.getTII().get(NewOpcode)); 3243 Observer.changedInstr(MI); 3244 } 3245 3246 LegalizerHelper::LegalizeResult 3247 LegalizerHelper::lower(MachineInstr &MI, unsigned TypeIdx, LLT LowerHintTy) { 3248 using namespace TargetOpcode; 3249 3250 switch(MI.getOpcode()) { 3251 default: 3252 return UnableToLegalize; 3253 case TargetOpcode::G_BITCAST: 3254 return lowerBitcast(MI); 3255 case TargetOpcode::G_SREM: 3256 case TargetOpcode::G_UREM: { 3257 LLT Ty = MRI.getType(MI.getOperand(0).getReg()); 3258 auto Quot = 3259 MIRBuilder.buildInstr(MI.getOpcode() == G_SREM ? G_SDIV : G_UDIV, {Ty}, 3260 {MI.getOperand(1), MI.getOperand(2)}); 3261 3262 auto Prod = MIRBuilder.buildMul(Ty, Quot, MI.getOperand(2)); 3263 MIRBuilder.buildSub(MI.getOperand(0), MI.getOperand(1), Prod); 3264 MI.eraseFromParent(); 3265 return Legalized; 3266 } 3267 case TargetOpcode::G_SADDO: 3268 case TargetOpcode::G_SSUBO: 3269 return lowerSADDO_SSUBO(MI); 3270 case TargetOpcode::G_UMULH: 3271 case TargetOpcode::G_SMULH: 3272 return lowerSMULH_UMULH(MI); 3273 case TargetOpcode::G_SMULO: 3274 case TargetOpcode::G_UMULO: { 3275 // Generate G_UMULH/G_SMULH to check for overflow and a normal G_MUL for the 3276 // result. 3277 Register Res = MI.getOperand(0).getReg(); 3278 Register Overflow = MI.getOperand(1).getReg(); 3279 Register LHS = MI.getOperand(2).getReg(); 3280 Register RHS = MI.getOperand(3).getReg(); 3281 LLT Ty = MRI.getType(Res); 3282 3283 unsigned Opcode = MI.getOpcode() == TargetOpcode::G_SMULO 3284 ? TargetOpcode::G_SMULH 3285 : TargetOpcode::G_UMULH; 3286 3287 Observer.changingInstr(MI); 3288 const auto &TII = MIRBuilder.getTII(); 3289 MI.setDesc(TII.get(TargetOpcode::G_MUL)); 3290 MI.removeOperand(1); 3291 Observer.changedInstr(MI); 3292 3293 auto HiPart = MIRBuilder.buildInstr(Opcode, {Ty}, {LHS, RHS}); 3294 auto Zero = MIRBuilder.buildConstant(Ty, 0); 3295 3296 // Move insert point forward so we can use the Res register if needed. 3297 MIRBuilder.setInsertPt(MIRBuilder.getMBB(), ++MIRBuilder.getInsertPt()); 3298 3299 // For *signed* multiply, overflow is detected by checking: 3300 // (hi != (lo >> bitwidth-1)) 3301 if (Opcode == TargetOpcode::G_SMULH) { 3302 auto ShiftAmt = MIRBuilder.buildConstant(Ty, Ty.getSizeInBits() - 1); 3303 auto Shifted = MIRBuilder.buildAShr(Ty, Res, ShiftAmt); 3304 MIRBuilder.buildICmp(CmpInst::ICMP_NE, Overflow, HiPart, Shifted); 3305 } else { 3306 MIRBuilder.buildICmp(CmpInst::ICMP_NE, Overflow, HiPart, Zero); 3307 } 3308 return Legalized; 3309 } 3310 case TargetOpcode::G_FNEG: { 3311 Register Res = MI.getOperand(0).getReg(); 3312 LLT Ty = MRI.getType(Res); 3313 3314 // TODO: Handle vector types once we are able to 3315 // represent them. 3316 if (Ty.isVector()) 3317 return UnableToLegalize; 3318 auto SignMask = 3319 MIRBuilder.buildConstant(Ty, APInt::getSignMask(Ty.getSizeInBits())); 3320 Register SubByReg = MI.getOperand(1).getReg(); 3321 MIRBuilder.buildXor(Res, SubByReg, SignMask); 3322 MI.eraseFromParent(); 3323 return Legalized; 3324 } 3325 case TargetOpcode::G_FSUB: 3326 case TargetOpcode::G_STRICT_FSUB: { 3327 Register Res = MI.getOperand(0).getReg(); 3328 LLT Ty = MRI.getType(Res); 3329 3330 // Lower (G_FSUB LHS, RHS) to (G_FADD LHS, (G_FNEG RHS)). 3331 // First, check if G_FNEG is marked as Lower. If so, we may 3332 // end up with an infinite loop as G_FSUB is used to legalize G_FNEG. 3333 if (LI.getAction({G_FNEG, {Ty}}).Action == Lower) 3334 return UnableToLegalize; 3335 Register LHS = MI.getOperand(1).getReg(); 3336 Register RHS = MI.getOperand(2).getReg(); 3337 auto Neg = MIRBuilder.buildFNeg(Ty, RHS); 3338 3339 if (MI.getOpcode() == TargetOpcode::G_STRICT_FSUB) 3340 MIRBuilder.buildStrictFAdd(Res, LHS, Neg, MI.getFlags()); 3341 else 3342 MIRBuilder.buildFAdd(Res, LHS, Neg, MI.getFlags()); 3343 3344 MI.eraseFromParent(); 3345 return Legalized; 3346 } 3347 case TargetOpcode::G_FMAD: 3348 return lowerFMad(MI); 3349 case TargetOpcode::G_FFLOOR: 3350 return lowerFFloor(MI); 3351 case TargetOpcode::G_INTRINSIC_ROUND: 3352 return lowerIntrinsicRound(MI); 3353 case TargetOpcode::G_INTRINSIC_ROUNDEVEN: { 3354 // Since round even is the assumed rounding mode for unconstrained FP 3355 // operations, rint and roundeven are the same operation. 3356 changeOpcode(MI, TargetOpcode::G_FRINT); 3357 return Legalized; 3358 } 3359 case TargetOpcode::G_ATOMIC_CMPXCHG_WITH_SUCCESS: { 3360 Register OldValRes = MI.getOperand(0).getReg(); 3361 Register SuccessRes = MI.getOperand(1).getReg(); 3362 Register Addr = MI.getOperand(2).getReg(); 3363 Register CmpVal = MI.getOperand(3).getReg(); 3364 Register NewVal = MI.getOperand(4).getReg(); 3365 MIRBuilder.buildAtomicCmpXchg(OldValRes, Addr, CmpVal, NewVal, 3366 **MI.memoperands_begin()); 3367 MIRBuilder.buildICmp(CmpInst::ICMP_EQ, SuccessRes, OldValRes, CmpVal); 3368 MI.eraseFromParent(); 3369 return Legalized; 3370 } 3371 case TargetOpcode::G_LOAD: 3372 case TargetOpcode::G_SEXTLOAD: 3373 case TargetOpcode::G_ZEXTLOAD: 3374 return lowerLoad(cast<GAnyLoad>(MI)); 3375 case TargetOpcode::G_STORE: 3376 return lowerStore(cast<GStore>(MI)); 3377 case TargetOpcode::G_CTLZ_ZERO_UNDEF: 3378 case TargetOpcode::G_CTTZ_ZERO_UNDEF: 3379 case TargetOpcode::G_CTLZ: 3380 case TargetOpcode::G_CTTZ: 3381 case TargetOpcode::G_CTPOP: 3382 return lowerBitCount(MI); 3383 case G_UADDO: { 3384 Register Res = MI.getOperand(0).getReg(); 3385 Register CarryOut = MI.getOperand(1).getReg(); 3386 Register LHS = MI.getOperand(2).getReg(); 3387 Register RHS = MI.getOperand(3).getReg(); 3388 3389 MIRBuilder.buildAdd(Res, LHS, RHS); 3390 MIRBuilder.buildICmp(CmpInst::ICMP_ULT, CarryOut, Res, RHS); 3391 3392 MI.eraseFromParent(); 3393 return Legalized; 3394 } 3395 case G_UADDE: { 3396 Register Res = MI.getOperand(0).getReg(); 3397 Register CarryOut = MI.getOperand(1).getReg(); 3398 Register LHS = MI.getOperand(2).getReg(); 3399 Register RHS = MI.getOperand(3).getReg(); 3400 Register CarryIn = MI.getOperand(4).getReg(); 3401 LLT Ty = MRI.getType(Res); 3402 3403 auto TmpRes = MIRBuilder.buildAdd(Ty, LHS, RHS); 3404 auto ZExtCarryIn = MIRBuilder.buildZExt(Ty, CarryIn); 3405 MIRBuilder.buildAdd(Res, TmpRes, ZExtCarryIn); 3406 MIRBuilder.buildICmp(CmpInst::ICMP_ULT, CarryOut, Res, LHS); 3407 3408 MI.eraseFromParent(); 3409 return Legalized; 3410 } 3411 case G_USUBO: { 3412 Register Res = MI.getOperand(0).getReg(); 3413 Register BorrowOut = MI.getOperand(1).getReg(); 3414 Register LHS = MI.getOperand(2).getReg(); 3415 Register RHS = MI.getOperand(3).getReg(); 3416 3417 MIRBuilder.buildSub(Res, LHS, RHS); 3418 MIRBuilder.buildICmp(CmpInst::ICMP_ULT, BorrowOut, LHS, RHS); 3419 3420 MI.eraseFromParent(); 3421 return Legalized; 3422 } 3423 case G_USUBE: { 3424 Register Res = MI.getOperand(0).getReg(); 3425 Register BorrowOut = MI.getOperand(1).getReg(); 3426 Register LHS = MI.getOperand(2).getReg(); 3427 Register RHS = MI.getOperand(3).getReg(); 3428 Register BorrowIn = MI.getOperand(4).getReg(); 3429 const LLT CondTy = MRI.getType(BorrowOut); 3430 const LLT Ty = MRI.getType(Res); 3431 3432 auto TmpRes = MIRBuilder.buildSub(Ty, LHS, RHS); 3433 auto ZExtBorrowIn = MIRBuilder.buildZExt(Ty, BorrowIn); 3434 MIRBuilder.buildSub(Res, TmpRes, ZExtBorrowIn); 3435 3436 auto LHS_EQ_RHS = MIRBuilder.buildICmp(CmpInst::ICMP_EQ, CondTy, LHS, RHS); 3437 auto LHS_ULT_RHS = MIRBuilder.buildICmp(CmpInst::ICMP_ULT, CondTy, LHS, RHS); 3438 MIRBuilder.buildSelect(BorrowOut, LHS_EQ_RHS, BorrowIn, LHS_ULT_RHS); 3439 3440 MI.eraseFromParent(); 3441 return Legalized; 3442 } 3443 case G_UITOFP: 3444 return lowerUITOFP(MI); 3445 case G_SITOFP: 3446 return lowerSITOFP(MI); 3447 case G_FPTOUI: 3448 return lowerFPTOUI(MI); 3449 case G_FPTOSI: 3450 return lowerFPTOSI(MI); 3451 case G_FPTRUNC: 3452 return lowerFPTRUNC(MI); 3453 case G_FPOWI: 3454 return lowerFPOWI(MI); 3455 case G_SMIN: 3456 case G_SMAX: 3457 case G_UMIN: 3458 case G_UMAX: 3459 return lowerMinMax(MI); 3460 case G_FCOPYSIGN: 3461 return lowerFCopySign(MI); 3462 case G_FMINNUM: 3463 case G_FMAXNUM: 3464 return lowerFMinNumMaxNum(MI); 3465 case G_MERGE_VALUES: 3466 return lowerMergeValues(MI); 3467 case G_UNMERGE_VALUES: 3468 return lowerUnmergeValues(MI); 3469 case TargetOpcode::G_SEXT_INREG: { 3470 assert(MI.getOperand(2).isImm() && "Expected immediate"); 3471 int64_t SizeInBits = MI.getOperand(2).getImm(); 3472 3473 Register DstReg = MI.getOperand(0).getReg(); 3474 Register SrcReg = MI.getOperand(1).getReg(); 3475 LLT DstTy = MRI.getType(DstReg); 3476 Register TmpRes = MRI.createGenericVirtualRegister(DstTy); 3477 3478 auto MIBSz = MIRBuilder.buildConstant(DstTy, DstTy.getScalarSizeInBits() - SizeInBits); 3479 MIRBuilder.buildShl(TmpRes, SrcReg, MIBSz->getOperand(0)); 3480 MIRBuilder.buildAShr(DstReg, TmpRes, MIBSz->getOperand(0)); 3481 MI.eraseFromParent(); 3482 return Legalized; 3483 } 3484 case G_EXTRACT_VECTOR_ELT: 3485 case G_INSERT_VECTOR_ELT: 3486 return lowerExtractInsertVectorElt(MI); 3487 case G_SHUFFLE_VECTOR: 3488 return lowerShuffleVector(MI); 3489 case G_DYN_STACKALLOC: 3490 return lowerDynStackAlloc(MI); 3491 case G_EXTRACT: 3492 return lowerExtract(MI); 3493 case G_INSERT: 3494 return lowerInsert(MI); 3495 case G_BSWAP: 3496 return lowerBswap(MI); 3497 case G_BITREVERSE: 3498 return lowerBitreverse(MI); 3499 case G_READ_REGISTER: 3500 case G_WRITE_REGISTER: 3501 return lowerReadWriteRegister(MI); 3502 case G_UADDSAT: 3503 case G_USUBSAT: { 3504 // Try to make a reasonable guess about which lowering strategy to use. The 3505 // target can override this with custom lowering and calling the 3506 // implementation functions. 3507 LLT Ty = MRI.getType(MI.getOperand(0).getReg()); 3508 if (LI.isLegalOrCustom({G_UMIN, Ty})) 3509 return lowerAddSubSatToMinMax(MI); 3510 return lowerAddSubSatToAddoSubo(MI); 3511 } 3512 case G_SADDSAT: 3513 case G_SSUBSAT: { 3514 LLT Ty = MRI.getType(MI.getOperand(0).getReg()); 3515 3516 // FIXME: It would probably make more sense to see if G_SADDO is preferred, 3517 // since it's a shorter expansion. However, we would need to figure out the 3518 // preferred boolean type for the carry out for the query. 3519 if (LI.isLegalOrCustom({G_SMIN, Ty}) && LI.isLegalOrCustom({G_SMAX, Ty})) 3520 return lowerAddSubSatToMinMax(MI); 3521 return lowerAddSubSatToAddoSubo(MI); 3522 } 3523 case G_SSHLSAT: 3524 case G_USHLSAT: 3525 return lowerShlSat(MI); 3526 case G_ABS: 3527 return lowerAbsToAddXor(MI); 3528 case G_SELECT: 3529 return lowerSelect(MI); 3530 case G_IS_FPCLASS: 3531 return lowerISFPCLASS(MI); 3532 case G_SDIVREM: 3533 case G_UDIVREM: 3534 return lowerDIVREM(MI); 3535 case G_FSHL: 3536 case G_FSHR: 3537 return lowerFunnelShift(MI); 3538 case G_ROTL: 3539 case G_ROTR: 3540 return lowerRotate(MI); 3541 case G_MEMSET: 3542 case G_MEMCPY: 3543 case G_MEMMOVE: 3544 return lowerMemCpyFamily(MI); 3545 case G_MEMCPY_INLINE: 3546 return lowerMemcpyInline(MI); 3547 GISEL_VECREDUCE_CASES_NONSEQ 3548 return lowerVectorReduction(MI); 3549 } 3550 } 3551 3552 Align LegalizerHelper::getStackTemporaryAlignment(LLT Ty, 3553 Align MinAlign) const { 3554 // FIXME: We're missing a way to go back from LLT to llvm::Type to query the 3555 // datalayout for the preferred alignment. Also there should be a target hook 3556 // for this to allow targets to reduce the alignment and ignore the 3557 // datalayout. e.g. AMDGPU should always use a 4-byte alignment, regardless of 3558 // the type. 3559 return std::max(Align(PowerOf2Ceil(Ty.getSizeInBytes())), MinAlign); 3560 } 3561 3562 MachineInstrBuilder 3563 LegalizerHelper::createStackTemporary(TypeSize Bytes, Align Alignment, 3564 MachinePointerInfo &PtrInfo) { 3565 MachineFunction &MF = MIRBuilder.getMF(); 3566 const DataLayout &DL = MIRBuilder.getDataLayout(); 3567 int FrameIdx = MF.getFrameInfo().CreateStackObject(Bytes, Alignment, false); 3568 3569 unsigned AddrSpace = DL.getAllocaAddrSpace(); 3570 LLT FramePtrTy = LLT::pointer(AddrSpace, DL.getPointerSizeInBits(AddrSpace)); 3571 3572 PtrInfo = MachinePointerInfo::getFixedStack(MF, FrameIdx); 3573 return MIRBuilder.buildFrameIndex(FramePtrTy, FrameIdx); 3574 } 3575 3576 static Register clampDynamicVectorIndex(MachineIRBuilder &B, Register IdxReg, 3577 LLT VecTy) { 3578 int64_t IdxVal; 3579 if (mi_match(IdxReg, *B.getMRI(), m_ICst(IdxVal))) 3580 return IdxReg; 3581 3582 LLT IdxTy = B.getMRI()->getType(IdxReg); 3583 unsigned NElts = VecTy.getNumElements(); 3584 if (isPowerOf2_32(NElts)) { 3585 APInt Imm = APInt::getLowBitsSet(IdxTy.getSizeInBits(), Log2_32(NElts)); 3586 return B.buildAnd(IdxTy, IdxReg, B.buildConstant(IdxTy, Imm)).getReg(0); 3587 } 3588 3589 return B.buildUMin(IdxTy, IdxReg, B.buildConstant(IdxTy, NElts - 1)) 3590 .getReg(0); 3591 } 3592 3593 Register LegalizerHelper::getVectorElementPointer(Register VecPtr, LLT VecTy, 3594 Register Index) { 3595 LLT EltTy = VecTy.getElementType(); 3596 3597 // Calculate the element offset and add it to the pointer. 3598 unsigned EltSize = EltTy.getSizeInBits() / 8; // FIXME: should be ABI size. 3599 assert(EltSize * 8 == EltTy.getSizeInBits() && 3600 "Converting bits to bytes lost precision"); 3601 3602 Index = clampDynamicVectorIndex(MIRBuilder, Index, VecTy); 3603 3604 LLT IdxTy = MRI.getType(Index); 3605 auto Mul = MIRBuilder.buildMul(IdxTy, Index, 3606 MIRBuilder.buildConstant(IdxTy, EltSize)); 3607 3608 LLT PtrTy = MRI.getType(VecPtr); 3609 return MIRBuilder.buildPtrAdd(PtrTy, VecPtr, Mul).getReg(0); 3610 } 3611 3612 #ifndef NDEBUG 3613 /// Check that all vector operands have same number of elements. Other operands 3614 /// should be listed in NonVecOp. 3615 static bool hasSameNumEltsOnAllVectorOperands( 3616 GenericMachineInstr &MI, MachineRegisterInfo &MRI, 3617 std::initializer_list<unsigned> NonVecOpIndices) { 3618 if (MI.getNumMemOperands() != 0) 3619 return false; 3620 3621 LLT VecTy = MRI.getType(MI.getReg(0)); 3622 if (!VecTy.isVector()) 3623 return false; 3624 unsigned NumElts = VecTy.getNumElements(); 3625 3626 for (unsigned OpIdx = 1; OpIdx < MI.getNumOperands(); ++OpIdx) { 3627 MachineOperand &Op = MI.getOperand(OpIdx); 3628 if (!Op.isReg()) { 3629 if (!is_contained(NonVecOpIndices, OpIdx)) 3630 return false; 3631 continue; 3632 } 3633 3634 LLT Ty = MRI.getType(Op.getReg()); 3635 if (!Ty.isVector()) { 3636 if (!is_contained(NonVecOpIndices, OpIdx)) 3637 return false; 3638 continue; 3639 } 3640 3641 if (Ty.getNumElements() != NumElts) 3642 return false; 3643 } 3644 3645 return true; 3646 } 3647 #endif 3648 3649 /// Fill \p DstOps with DstOps that have same number of elements combined as 3650 /// the Ty. These DstOps have either scalar type when \p NumElts = 1 or are 3651 /// vectors with \p NumElts elements. When Ty.getNumElements() is not multiple 3652 /// of \p NumElts last DstOp (leftover) has fewer then \p NumElts elements. 3653 static void makeDstOps(SmallVectorImpl<DstOp> &DstOps, LLT Ty, 3654 unsigned NumElts) { 3655 LLT LeftoverTy; 3656 assert(Ty.isVector() && "Expected vector type"); 3657 LLT EltTy = Ty.getElementType(); 3658 LLT NarrowTy = (NumElts == 1) ? EltTy : LLT::fixed_vector(NumElts, EltTy); 3659 int NumParts, NumLeftover; 3660 std::tie(NumParts, NumLeftover) = 3661 getNarrowTypeBreakDown(Ty, NarrowTy, LeftoverTy); 3662 3663 assert(NumParts > 0 && "Error in getNarrowTypeBreakDown"); 3664 for (int i = 0; i < NumParts; ++i) { 3665 DstOps.push_back(NarrowTy); 3666 } 3667 3668 if (LeftoverTy.isValid()) { 3669 assert(NumLeftover == 1 && "expected exactly one leftover"); 3670 DstOps.push_back(LeftoverTy); 3671 } 3672 } 3673 3674 /// Operand \p Op is used on \p N sub-instructions. Fill \p Ops with \p N SrcOps 3675 /// made from \p Op depending on operand type. 3676 static void broadcastSrcOp(SmallVectorImpl<SrcOp> &Ops, unsigned N, 3677 MachineOperand &Op) { 3678 for (unsigned i = 0; i < N; ++i) { 3679 if (Op.isReg()) 3680 Ops.push_back(Op.getReg()); 3681 else if (Op.isImm()) 3682 Ops.push_back(Op.getImm()); 3683 else if (Op.isPredicate()) 3684 Ops.push_back(static_cast<CmpInst::Predicate>(Op.getPredicate())); 3685 else 3686 llvm_unreachable("Unsupported type"); 3687 } 3688 } 3689 3690 // Handle splitting vector operations which need to have the same number of 3691 // elements in each type index, but each type index may have a different element 3692 // type. 3693 // 3694 // e.g. <4 x s64> = G_SHL <4 x s64>, <4 x s32> -> 3695 // <2 x s64> = G_SHL <2 x s64>, <2 x s32> 3696 // <2 x s64> = G_SHL <2 x s64>, <2 x s32> 3697 // 3698 // Also handles some irregular breakdown cases, e.g. 3699 // e.g. <3 x s64> = G_SHL <3 x s64>, <3 x s32> -> 3700 // <2 x s64> = G_SHL <2 x s64>, <2 x s32> 3701 // s64 = G_SHL s64, s32 3702 LegalizerHelper::LegalizeResult 3703 LegalizerHelper::fewerElementsVectorMultiEltType( 3704 GenericMachineInstr &MI, unsigned NumElts, 3705 std::initializer_list<unsigned> NonVecOpIndices) { 3706 assert(hasSameNumEltsOnAllVectorOperands(MI, MRI, NonVecOpIndices) && 3707 "Non-compatible opcode or not specified non-vector operands"); 3708 unsigned OrigNumElts = MRI.getType(MI.getReg(0)).getNumElements(); 3709 3710 unsigned NumInputs = MI.getNumOperands() - MI.getNumDefs(); 3711 unsigned NumDefs = MI.getNumDefs(); 3712 3713 // Create DstOps (sub-vectors with NumElts elts + Leftover) for each output. 3714 // Build instructions with DstOps to use instruction found by CSE directly. 3715 // CSE copies found instruction into given vreg when building with vreg dest. 3716 SmallVector<SmallVector<DstOp, 8>, 2> OutputOpsPieces(NumDefs); 3717 // Output registers will be taken from created instructions. 3718 SmallVector<SmallVector<Register, 8>, 2> OutputRegs(NumDefs); 3719 for (unsigned i = 0; i < NumDefs; ++i) { 3720 makeDstOps(OutputOpsPieces[i], MRI.getType(MI.getReg(i)), NumElts); 3721 } 3722 3723 // Split vector input operands into sub-vectors with NumElts elts + Leftover. 3724 // Operands listed in NonVecOpIndices will be used as is without splitting; 3725 // examples: compare predicate in icmp and fcmp (op 1), vector select with i1 3726 // scalar condition (op 1), immediate in sext_inreg (op 2). 3727 SmallVector<SmallVector<SrcOp, 8>, 3> InputOpsPieces(NumInputs); 3728 for (unsigned UseIdx = NumDefs, UseNo = 0; UseIdx < MI.getNumOperands(); 3729 ++UseIdx, ++UseNo) { 3730 if (is_contained(NonVecOpIndices, UseIdx)) { 3731 broadcastSrcOp(InputOpsPieces[UseNo], OutputOpsPieces[0].size(), 3732 MI.getOperand(UseIdx)); 3733 } else { 3734 SmallVector<Register, 8> SplitPieces; 3735 extractVectorParts(MI.getReg(UseIdx), NumElts, SplitPieces); 3736 for (auto Reg : SplitPieces) 3737 InputOpsPieces[UseNo].push_back(Reg); 3738 } 3739 } 3740 3741 unsigned NumLeftovers = OrigNumElts % NumElts ? 1 : 0; 3742 3743 // Take i-th piece of each input operand split and build sub-vector/scalar 3744 // instruction. Set i-th DstOp(s) from OutputOpsPieces as destination(s). 3745 for (unsigned i = 0; i < OrigNumElts / NumElts + NumLeftovers; ++i) { 3746 SmallVector<DstOp, 2> Defs; 3747 for (unsigned DstNo = 0; DstNo < NumDefs; ++DstNo) 3748 Defs.push_back(OutputOpsPieces[DstNo][i]); 3749 3750 SmallVector<SrcOp, 3> Uses; 3751 for (unsigned InputNo = 0; InputNo < NumInputs; ++InputNo) 3752 Uses.push_back(InputOpsPieces[InputNo][i]); 3753 3754 auto I = MIRBuilder.buildInstr(MI.getOpcode(), Defs, Uses, MI.getFlags()); 3755 for (unsigned DstNo = 0; DstNo < NumDefs; ++DstNo) 3756 OutputRegs[DstNo].push_back(I.getReg(DstNo)); 3757 } 3758 3759 // Merge small outputs into MI's output for each def operand. 3760 if (NumLeftovers) { 3761 for (unsigned i = 0; i < NumDefs; ++i) 3762 mergeMixedSubvectors(MI.getReg(i), OutputRegs[i]); 3763 } else { 3764 for (unsigned i = 0; i < NumDefs; ++i) 3765 MIRBuilder.buildMergeLikeInstr(MI.getReg(i), OutputRegs[i]); 3766 } 3767 3768 MI.eraseFromParent(); 3769 return Legalized; 3770 } 3771 3772 LegalizerHelper::LegalizeResult 3773 LegalizerHelper::fewerElementsVectorPhi(GenericMachineInstr &MI, 3774 unsigned NumElts) { 3775 unsigned OrigNumElts = MRI.getType(MI.getReg(0)).getNumElements(); 3776 3777 unsigned NumInputs = MI.getNumOperands() - MI.getNumDefs(); 3778 unsigned NumDefs = MI.getNumDefs(); 3779 3780 SmallVector<DstOp, 8> OutputOpsPieces; 3781 SmallVector<Register, 8> OutputRegs; 3782 makeDstOps(OutputOpsPieces, MRI.getType(MI.getReg(0)), NumElts); 3783 3784 // Instructions that perform register split will be inserted in basic block 3785 // where register is defined (basic block is in the next operand). 3786 SmallVector<SmallVector<Register, 8>, 3> InputOpsPieces(NumInputs / 2); 3787 for (unsigned UseIdx = NumDefs, UseNo = 0; UseIdx < MI.getNumOperands(); 3788 UseIdx += 2, ++UseNo) { 3789 MachineBasicBlock &OpMBB = *MI.getOperand(UseIdx + 1).getMBB(); 3790 MIRBuilder.setInsertPt(OpMBB, OpMBB.getFirstTerminatorForward()); 3791 extractVectorParts(MI.getReg(UseIdx), NumElts, InputOpsPieces[UseNo]); 3792 } 3793 3794 // Build PHIs with fewer elements. 3795 unsigned NumLeftovers = OrigNumElts % NumElts ? 1 : 0; 3796 MIRBuilder.setInsertPt(*MI.getParent(), MI); 3797 for (unsigned i = 0; i < OrigNumElts / NumElts + NumLeftovers; ++i) { 3798 auto Phi = MIRBuilder.buildInstr(TargetOpcode::G_PHI); 3799 Phi.addDef( 3800 MRI.createGenericVirtualRegister(OutputOpsPieces[i].getLLTTy(MRI))); 3801 OutputRegs.push_back(Phi.getReg(0)); 3802 3803 for (unsigned j = 0; j < NumInputs / 2; ++j) { 3804 Phi.addUse(InputOpsPieces[j][i]); 3805 Phi.add(MI.getOperand(1 + j * 2 + 1)); 3806 } 3807 } 3808 3809 // Merge small outputs into MI's def. 3810 if (NumLeftovers) { 3811 mergeMixedSubvectors(MI.getReg(0), OutputRegs); 3812 } else { 3813 MIRBuilder.buildMergeLikeInstr(MI.getReg(0), OutputRegs); 3814 } 3815 3816 MI.eraseFromParent(); 3817 return Legalized; 3818 } 3819 3820 LegalizerHelper::LegalizeResult 3821 LegalizerHelper::fewerElementsVectorUnmergeValues(MachineInstr &MI, 3822 unsigned TypeIdx, 3823 LLT NarrowTy) { 3824 const int NumDst = MI.getNumOperands() - 1; 3825 const Register SrcReg = MI.getOperand(NumDst).getReg(); 3826 LLT DstTy = MRI.getType(MI.getOperand(0).getReg()); 3827 LLT SrcTy = MRI.getType(SrcReg); 3828 3829 if (TypeIdx != 1 || NarrowTy == DstTy) 3830 return UnableToLegalize; 3831 3832 // Requires compatible types. Otherwise SrcReg should have been defined by 3833 // merge-like instruction that would get artifact combined. Most likely 3834 // instruction that defines SrcReg has to perform more/fewer elements 3835 // legalization compatible with NarrowTy. 3836 assert(SrcTy.isVector() && NarrowTy.isVector() && "Expected vector types"); 3837 assert((SrcTy.getScalarType() == NarrowTy.getScalarType()) && "bad type"); 3838 3839 if ((SrcTy.getSizeInBits() % NarrowTy.getSizeInBits() != 0) || 3840 (NarrowTy.getSizeInBits() % DstTy.getSizeInBits() != 0)) 3841 return UnableToLegalize; 3842 3843 // This is most likely DstTy (smaller then register size) packed in SrcTy 3844 // (larger then register size) and since unmerge was not combined it will be 3845 // lowered to bit sequence extracts from register. Unpack SrcTy to NarrowTy 3846 // (register size) pieces first. Then unpack each of NarrowTy pieces to DstTy. 3847 3848 // %1:_(DstTy), %2, %3, %4 = G_UNMERGE_VALUES %0:_(SrcTy) 3849 // 3850 // %5:_(NarrowTy), %6 = G_UNMERGE_VALUES %0:_(SrcTy) - reg sequence 3851 // %1:_(DstTy), %2 = G_UNMERGE_VALUES %5:_(NarrowTy) - sequence of bits in reg 3852 // %3:_(DstTy), %4 = G_UNMERGE_VALUES %6:_(NarrowTy) 3853 auto Unmerge = MIRBuilder.buildUnmerge(NarrowTy, SrcReg); 3854 const int NumUnmerge = Unmerge->getNumOperands() - 1; 3855 const int PartsPerUnmerge = NumDst / NumUnmerge; 3856 3857 for (int I = 0; I != NumUnmerge; ++I) { 3858 auto MIB = MIRBuilder.buildInstr(TargetOpcode::G_UNMERGE_VALUES); 3859 3860 for (int J = 0; J != PartsPerUnmerge; ++J) 3861 MIB.addDef(MI.getOperand(I * PartsPerUnmerge + J).getReg()); 3862 MIB.addUse(Unmerge.getReg(I)); 3863 } 3864 3865 MI.eraseFromParent(); 3866 return Legalized; 3867 } 3868 3869 LegalizerHelper::LegalizeResult 3870 LegalizerHelper::fewerElementsVectorMerge(MachineInstr &MI, unsigned TypeIdx, 3871 LLT NarrowTy) { 3872 Register DstReg = MI.getOperand(0).getReg(); 3873 LLT DstTy = MRI.getType(DstReg); 3874 LLT SrcTy = MRI.getType(MI.getOperand(1).getReg()); 3875 // Requires compatible types. Otherwise user of DstReg did not perform unmerge 3876 // that should have been artifact combined. Most likely instruction that uses 3877 // DstReg has to do more/fewer elements legalization compatible with NarrowTy. 3878 assert(DstTy.isVector() && NarrowTy.isVector() && "Expected vector types"); 3879 assert((DstTy.getScalarType() == NarrowTy.getScalarType()) && "bad type"); 3880 if (NarrowTy == SrcTy) 3881 return UnableToLegalize; 3882 3883 // This attempts to lower part of LCMTy merge/unmerge sequence. Intended use 3884 // is for old mir tests. Since the changes to more/fewer elements it should no 3885 // longer be possible to generate MIR like this when starting from llvm-ir 3886 // because LCMTy approach was replaced with merge/unmerge to vector elements. 3887 if (TypeIdx == 1) { 3888 assert(SrcTy.isVector() && "Expected vector types"); 3889 assert((SrcTy.getScalarType() == NarrowTy.getScalarType()) && "bad type"); 3890 if ((DstTy.getSizeInBits() % NarrowTy.getSizeInBits() != 0) || 3891 (NarrowTy.getNumElements() >= SrcTy.getNumElements())) 3892 return UnableToLegalize; 3893 // %2:_(DstTy) = G_CONCAT_VECTORS %0:_(SrcTy), %1:_(SrcTy) 3894 // 3895 // %3:_(EltTy), %4, %5 = G_UNMERGE_VALUES %0:_(SrcTy) 3896 // %6:_(EltTy), %7, %8 = G_UNMERGE_VALUES %1:_(SrcTy) 3897 // %9:_(NarrowTy) = G_BUILD_VECTOR %3:_(EltTy), %4 3898 // %10:_(NarrowTy) = G_BUILD_VECTOR %5:_(EltTy), %6 3899 // %11:_(NarrowTy) = G_BUILD_VECTOR %7:_(EltTy), %8 3900 // %2:_(DstTy) = G_CONCAT_VECTORS %9:_(NarrowTy), %10, %11 3901 3902 SmallVector<Register, 8> Elts; 3903 LLT EltTy = MRI.getType(MI.getOperand(1).getReg()).getScalarType(); 3904 for (unsigned i = 1; i < MI.getNumOperands(); ++i) { 3905 auto Unmerge = MIRBuilder.buildUnmerge(EltTy, MI.getOperand(i).getReg()); 3906 for (unsigned j = 0; j < Unmerge->getNumDefs(); ++j) 3907 Elts.push_back(Unmerge.getReg(j)); 3908 } 3909 3910 SmallVector<Register, 8> NarrowTyElts; 3911 unsigned NumNarrowTyElts = NarrowTy.getNumElements(); 3912 unsigned NumNarrowTyPieces = DstTy.getNumElements() / NumNarrowTyElts; 3913 for (unsigned i = 0, Offset = 0; i < NumNarrowTyPieces; 3914 ++i, Offset += NumNarrowTyElts) { 3915 ArrayRef<Register> Pieces(&Elts[Offset], NumNarrowTyElts); 3916 NarrowTyElts.push_back( 3917 MIRBuilder.buildMergeLikeInstr(NarrowTy, Pieces).getReg(0)); 3918 } 3919 3920 MIRBuilder.buildMergeLikeInstr(DstReg, NarrowTyElts); 3921 MI.eraseFromParent(); 3922 return Legalized; 3923 } 3924 3925 assert(TypeIdx == 0 && "Bad type index"); 3926 if ((NarrowTy.getSizeInBits() % SrcTy.getSizeInBits() != 0) || 3927 (DstTy.getSizeInBits() % NarrowTy.getSizeInBits() != 0)) 3928 return UnableToLegalize; 3929 3930 // This is most likely SrcTy (smaller then register size) packed in DstTy 3931 // (larger then register size) and since merge was not combined it will be 3932 // lowered to bit sequence packing into register. Merge SrcTy to NarrowTy 3933 // (register size) pieces first. Then merge each of NarrowTy pieces to DstTy. 3934 3935 // %0:_(DstTy) = G_MERGE_VALUES %1:_(SrcTy), %2, %3, %4 3936 // 3937 // %5:_(NarrowTy) = G_MERGE_VALUES %1:_(SrcTy), %2 - sequence of bits in reg 3938 // %6:_(NarrowTy) = G_MERGE_VALUES %3:_(SrcTy), %4 3939 // %0:_(DstTy) = G_MERGE_VALUES %5:_(NarrowTy), %6 - reg sequence 3940 SmallVector<Register, 8> NarrowTyElts; 3941 unsigned NumParts = DstTy.getNumElements() / NarrowTy.getNumElements(); 3942 unsigned NumSrcElts = SrcTy.isVector() ? SrcTy.getNumElements() : 1; 3943 unsigned NumElts = NarrowTy.getNumElements() / NumSrcElts; 3944 for (unsigned i = 0; i < NumParts; ++i) { 3945 SmallVector<Register, 8> Sources; 3946 for (unsigned j = 0; j < NumElts; ++j) 3947 Sources.push_back(MI.getOperand(1 + i * NumElts + j).getReg()); 3948 NarrowTyElts.push_back( 3949 MIRBuilder.buildMergeLikeInstr(NarrowTy, Sources).getReg(0)); 3950 } 3951 3952 MIRBuilder.buildMergeLikeInstr(DstReg, NarrowTyElts); 3953 MI.eraseFromParent(); 3954 return Legalized; 3955 } 3956 3957 LegalizerHelper::LegalizeResult 3958 LegalizerHelper::fewerElementsVectorExtractInsertVectorElt(MachineInstr &MI, 3959 unsigned TypeIdx, 3960 LLT NarrowVecTy) { 3961 Register DstReg = MI.getOperand(0).getReg(); 3962 Register SrcVec = MI.getOperand(1).getReg(); 3963 Register InsertVal; 3964 bool IsInsert = MI.getOpcode() == TargetOpcode::G_INSERT_VECTOR_ELT; 3965 3966 assert((IsInsert ? TypeIdx == 0 : TypeIdx == 1) && "not a vector type index"); 3967 if (IsInsert) 3968 InsertVal = MI.getOperand(2).getReg(); 3969 3970 Register Idx = MI.getOperand(MI.getNumOperands() - 1).getReg(); 3971 3972 // TODO: Handle total scalarization case. 3973 if (!NarrowVecTy.isVector()) 3974 return UnableToLegalize; 3975 3976 LLT VecTy = MRI.getType(SrcVec); 3977 3978 // If the index is a constant, we can really break this down as you would 3979 // expect, and index into the target size pieces. 3980 int64_t IdxVal; 3981 auto MaybeCst = getIConstantVRegValWithLookThrough(Idx, MRI); 3982 if (MaybeCst) { 3983 IdxVal = MaybeCst->Value.getSExtValue(); 3984 // Avoid out of bounds indexing the pieces. 3985 if (IdxVal >= VecTy.getNumElements()) { 3986 MIRBuilder.buildUndef(DstReg); 3987 MI.eraseFromParent(); 3988 return Legalized; 3989 } 3990 3991 SmallVector<Register, 8> VecParts; 3992 LLT GCDTy = extractGCDType(VecParts, VecTy, NarrowVecTy, SrcVec); 3993 3994 // Build a sequence of NarrowTy pieces in VecParts for this operand. 3995 LLT LCMTy = buildLCMMergePieces(VecTy, NarrowVecTy, GCDTy, VecParts, 3996 TargetOpcode::G_ANYEXT); 3997 3998 unsigned NewNumElts = NarrowVecTy.getNumElements(); 3999 4000 LLT IdxTy = MRI.getType(Idx); 4001 int64_t PartIdx = IdxVal / NewNumElts; 4002 auto NewIdx = 4003 MIRBuilder.buildConstant(IdxTy, IdxVal - NewNumElts * PartIdx); 4004 4005 if (IsInsert) { 4006 LLT PartTy = MRI.getType(VecParts[PartIdx]); 4007 4008 // Use the adjusted index to insert into one of the subvectors. 4009 auto InsertPart = MIRBuilder.buildInsertVectorElement( 4010 PartTy, VecParts[PartIdx], InsertVal, NewIdx); 4011 VecParts[PartIdx] = InsertPart.getReg(0); 4012 4013 // Recombine the inserted subvector with the others to reform the result 4014 // vector. 4015 buildWidenedRemergeToDst(DstReg, LCMTy, VecParts); 4016 } else { 4017 MIRBuilder.buildExtractVectorElement(DstReg, VecParts[PartIdx], NewIdx); 4018 } 4019 4020 MI.eraseFromParent(); 4021 return Legalized; 4022 } 4023 4024 // With a variable index, we can't perform the operation in a smaller type, so 4025 // we're forced to expand this. 4026 // 4027 // TODO: We could emit a chain of compare/select to figure out which piece to 4028 // index. 4029 return lowerExtractInsertVectorElt(MI); 4030 } 4031 4032 LegalizerHelper::LegalizeResult 4033 LegalizerHelper::reduceLoadStoreWidth(GLoadStore &LdStMI, unsigned TypeIdx, 4034 LLT NarrowTy) { 4035 // FIXME: Don't know how to handle secondary types yet. 4036 if (TypeIdx != 0) 4037 return UnableToLegalize; 4038 4039 // This implementation doesn't work for atomics. Give up instead of doing 4040 // something invalid. 4041 if (LdStMI.isAtomic()) 4042 return UnableToLegalize; 4043 4044 bool IsLoad = isa<GLoad>(LdStMI); 4045 Register ValReg = LdStMI.getReg(0); 4046 Register AddrReg = LdStMI.getPointerReg(); 4047 LLT ValTy = MRI.getType(ValReg); 4048 4049 // FIXME: Do we need a distinct NarrowMemory legalize action? 4050 if (ValTy.getSizeInBits() != 8 * LdStMI.getMemSize()) { 4051 LLVM_DEBUG(dbgs() << "Can't narrow extload/truncstore\n"); 4052 return UnableToLegalize; 4053 } 4054 4055 int NumParts = -1; 4056 int NumLeftover = -1; 4057 LLT LeftoverTy; 4058 SmallVector<Register, 8> NarrowRegs, NarrowLeftoverRegs; 4059 if (IsLoad) { 4060 std::tie(NumParts, NumLeftover) = getNarrowTypeBreakDown(ValTy, NarrowTy, LeftoverTy); 4061 } else { 4062 if (extractParts(ValReg, ValTy, NarrowTy, LeftoverTy, NarrowRegs, 4063 NarrowLeftoverRegs)) { 4064 NumParts = NarrowRegs.size(); 4065 NumLeftover = NarrowLeftoverRegs.size(); 4066 } 4067 } 4068 4069 if (NumParts == -1) 4070 return UnableToLegalize; 4071 4072 LLT PtrTy = MRI.getType(AddrReg); 4073 const LLT OffsetTy = LLT::scalar(PtrTy.getSizeInBits()); 4074 4075 unsigned TotalSize = ValTy.getSizeInBits(); 4076 4077 // Split the load/store into PartTy sized pieces starting at Offset. If this 4078 // is a load, return the new registers in ValRegs. For a store, each elements 4079 // of ValRegs should be PartTy. Returns the next offset that needs to be 4080 // handled. 4081 bool isBigEndian = MIRBuilder.getDataLayout().isBigEndian(); 4082 auto MMO = LdStMI.getMMO(); 4083 auto splitTypePieces = [=](LLT PartTy, SmallVectorImpl<Register> &ValRegs, 4084 unsigned NumParts, unsigned Offset) -> unsigned { 4085 MachineFunction &MF = MIRBuilder.getMF(); 4086 unsigned PartSize = PartTy.getSizeInBits(); 4087 for (unsigned Idx = 0, E = NumParts; Idx != E && Offset < TotalSize; 4088 ++Idx) { 4089 unsigned ByteOffset = Offset / 8; 4090 Register NewAddrReg; 4091 4092 MIRBuilder.materializePtrAdd(NewAddrReg, AddrReg, OffsetTy, ByteOffset); 4093 4094 MachineMemOperand *NewMMO = 4095 MF.getMachineMemOperand(&MMO, ByteOffset, PartTy); 4096 4097 if (IsLoad) { 4098 Register Dst = MRI.createGenericVirtualRegister(PartTy); 4099 ValRegs.push_back(Dst); 4100 MIRBuilder.buildLoad(Dst, NewAddrReg, *NewMMO); 4101 } else { 4102 MIRBuilder.buildStore(ValRegs[Idx], NewAddrReg, *NewMMO); 4103 } 4104 Offset = isBigEndian ? Offset - PartSize : Offset + PartSize; 4105 } 4106 4107 return Offset; 4108 }; 4109 4110 unsigned Offset = isBigEndian ? TotalSize - NarrowTy.getSizeInBits() : 0; 4111 unsigned HandledOffset = 4112 splitTypePieces(NarrowTy, NarrowRegs, NumParts, Offset); 4113 4114 // Handle the rest of the register if this isn't an even type breakdown. 4115 if (LeftoverTy.isValid()) 4116 splitTypePieces(LeftoverTy, NarrowLeftoverRegs, NumLeftover, HandledOffset); 4117 4118 if (IsLoad) { 4119 insertParts(ValReg, ValTy, NarrowTy, NarrowRegs, 4120 LeftoverTy, NarrowLeftoverRegs); 4121 } 4122 4123 LdStMI.eraseFromParent(); 4124 return Legalized; 4125 } 4126 4127 LegalizerHelper::LegalizeResult 4128 LegalizerHelper::fewerElementsVector(MachineInstr &MI, unsigned TypeIdx, 4129 LLT NarrowTy) { 4130 using namespace TargetOpcode; 4131 GenericMachineInstr &GMI = cast<GenericMachineInstr>(MI); 4132 unsigned NumElts = NarrowTy.isVector() ? NarrowTy.getNumElements() : 1; 4133 4134 switch (MI.getOpcode()) { 4135 case G_IMPLICIT_DEF: 4136 case G_TRUNC: 4137 case G_AND: 4138 case G_OR: 4139 case G_XOR: 4140 case G_ADD: 4141 case G_SUB: 4142 case G_MUL: 4143 case G_PTR_ADD: 4144 case G_SMULH: 4145 case G_UMULH: 4146 case G_FADD: 4147 case G_FMUL: 4148 case G_FSUB: 4149 case G_FNEG: 4150 case G_FABS: 4151 case G_FCANONICALIZE: 4152 case G_FDIV: 4153 case G_FREM: 4154 case G_FMA: 4155 case G_FMAD: 4156 case G_FPOW: 4157 case G_FEXP: 4158 case G_FEXP2: 4159 case G_FLOG: 4160 case G_FLOG2: 4161 case G_FLOG10: 4162 case G_FNEARBYINT: 4163 case G_FCEIL: 4164 case G_FFLOOR: 4165 case G_FRINT: 4166 case G_INTRINSIC_ROUND: 4167 case G_INTRINSIC_ROUNDEVEN: 4168 case G_INTRINSIC_TRUNC: 4169 case G_FCOS: 4170 case G_FSIN: 4171 case G_FSQRT: 4172 case G_BSWAP: 4173 case G_BITREVERSE: 4174 case G_SDIV: 4175 case G_UDIV: 4176 case G_SREM: 4177 case G_UREM: 4178 case G_SDIVREM: 4179 case G_UDIVREM: 4180 case G_SMIN: 4181 case G_SMAX: 4182 case G_UMIN: 4183 case G_UMAX: 4184 case G_ABS: 4185 case G_FMINNUM: 4186 case G_FMAXNUM: 4187 case G_FMINNUM_IEEE: 4188 case G_FMAXNUM_IEEE: 4189 case G_FMINIMUM: 4190 case G_FMAXIMUM: 4191 case G_FSHL: 4192 case G_FSHR: 4193 case G_ROTL: 4194 case G_ROTR: 4195 case G_FREEZE: 4196 case G_SADDSAT: 4197 case G_SSUBSAT: 4198 case G_UADDSAT: 4199 case G_USUBSAT: 4200 case G_UMULO: 4201 case G_SMULO: 4202 case G_SHL: 4203 case G_LSHR: 4204 case G_ASHR: 4205 case G_SSHLSAT: 4206 case G_USHLSAT: 4207 case G_CTLZ: 4208 case G_CTLZ_ZERO_UNDEF: 4209 case G_CTTZ: 4210 case G_CTTZ_ZERO_UNDEF: 4211 case G_CTPOP: 4212 case G_FCOPYSIGN: 4213 case G_ZEXT: 4214 case G_SEXT: 4215 case G_ANYEXT: 4216 case G_FPEXT: 4217 case G_FPTRUNC: 4218 case G_SITOFP: 4219 case G_UITOFP: 4220 case G_FPTOSI: 4221 case G_FPTOUI: 4222 case G_INTTOPTR: 4223 case G_PTRTOINT: 4224 case G_ADDRSPACE_CAST: 4225 case G_UADDO: 4226 case G_USUBO: 4227 case G_UADDE: 4228 case G_USUBE: 4229 case G_SADDO: 4230 case G_SSUBO: 4231 case G_SADDE: 4232 case G_SSUBE: 4233 case G_STRICT_FADD: 4234 case G_STRICT_FSUB: 4235 case G_STRICT_FMUL: 4236 case G_STRICT_FMA: 4237 return fewerElementsVectorMultiEltType(GMI, NumElts); 4238 case G_ICMP: 4239 case G_FCMP: 4240 return fewerElementsVectorMultiEltType(GMI, NumElts, {1 /*cpm predicate*/}); 4241 case G_IS_FPCLASS: 4242 return fewerElementsVectorMultiEltType(GMI, NumElts, {2, 3 /*mask,fpsem*/}); 4243 case G_SELECT: 4244 if (MRI.getType(MI.getOperand(1).getReg()).isVector()) 4245 return fewerElementsVectorMultiEltType(GMI, NumElts); 4246 return fewerElementsVectorMultiEltType(GMI, NumElts, {1 /*scalar cond*/}); 4247 case G_PHI: 4248 return fewerElementsVectorPhi(GMI, NumElts); 4249 case G_UNMERGE_VALUES: 4250 return fewerElementsVectorUnmergeValues(MI, TypeIdx, NarrowTy); 4251 case G_BUILD_VECTOR: 4252 assert(TypeIdx == 0 && "not a vector type index"); 4253 return fewerElementsVectorMerge(MI, TypeIdx, NarrowTy); 4254 case G_CONCAT_VECTORS: 4255 if (TypeIdx != 1) // TODO: This probably does work as expected already. 4256 return UnableToLegalize; 4257 return fewerElementsVectorMerge(MI, TypeIdx, NarrowTy); 4258 case G_EXTRACT_VECTOR_ELT: 4259 case G_INSERT_VECTOR_ELT: 4260 return fewerElementsVectorExtractInsertVectorElt(MI, TypeIdx, NarrowTy); 4261 case G_LOAD: 4262 case G_STORE: 4263 return reduceLoadStoreWidth(cast<GLoadStore>(MI), TypeIdx, NarrowTy); 4264 case G_SEXT_INREG: 4265 return fewerElementsVectorMultiEltType(GMI, NumElts, {2 /*imm*/}); 4266 GISEL_VECREDUCE_CASES_NONSEQ 4267 return fewerElementsVectorReductions(MI, TypeIdx, NarrowTy); 4268 case G_SHUFFLE_VECTOR: 4269 return fewerElementsVectorShuffle(MI, TypeIdx, NarrowTy); 4270 default: 4271 return UnableToLegalize; 4272 } 4273 } 4274 4275 LegalizerHelper::LegalizeResult LegalizerHelper::fewerElementsVectorShuffle( 4276 MachineInstr &MI, unsigned int TypeIdx, LLT NarrowTy) { 4277 assert(MI.getOpcode() == TargetOpcode::G_SHUFFLE_VECTOR); 4278 if (TypeIdx != 0) 4279 return UnableToLegalize; 4280 4281 Register DstReg = MI.getOperand(0).getReg(); 4282 Register Src1Reg = MI.getOperand(1).getReg(); 4283 Register Src2Reg = MI.getOperand(2).getReg(); 4284 ArrayRef<int> Mask = MI.getOperand(3).getShuffleMask(); 4285 LLT DstTy = MRI.getType(DstReg); 4286 LLT Src1Ty = MRI.getType(Src1Reg); 4287 LLT Src2Ty = MRI.getType(Src2Reg); 4288 // The shuffle should be canonicalized by now. 4289 if (DstTy != Src1Ty) 4290 return UnableToLegalize; 4291 if (DstTy != Src2Ty) 4292 return UnableToLegalize; 4293 4294 if (!isPowerOf2_32(DstTy.getNumElements())) 4295 return UnableToLegalize; 4296 4297 // We only support splitting a shuffle into 2, so adjust NarrowTy accordingly. 4298 // Further legalization attempts will be needed to do split further. 4299 NarrowTy = 4300 DstTy.changeElementCount(DstTy.getElementCount().divideCoefficientBy(2)); 4301 unsigned NewElts = NarrowTy.getNumElements(); 4302 4303 SmallVector<Register> SplitSrc1Regs, SplitSrc2Regs; 4304 extractParts(Src1Reg, NarrowTy, 2, SplitSrc1Regs); 4305 extractParts(Src2Reg, NarrowTy, 2, SplitSrc2Regs); 4306 Register Inputs[4] = {SplitSrc1Regs[0], SplitSrc1Regs[1], SplitSrc2Regs[0], 4307 SplitSrc2Regs[1]}; 4308 4309 Register Hi, Lo; 4310 4311 // If Lo or Hi uses elements from at most two of the four input vectors, then 4312 // express it as a vector shuffle of those two inputs. Otherwise extract the 4313 // input elements by hand and construct the Lo/Hi output using a BUILD_VECTOR. 4314 SmallVector<int, 16> Ops; 4315 for (unsigned High = 0; High < 2; ++High) { 4316 Register &Output = High ? Hi : Lo; 4317 4318 // Build a shuffle mask for the output, discovering on the fly which 4319 // input vectors to use as shuffle operands (recorded in InputUsed). 4320 // If building a suitable shuffle vector proves too hard, then bail 4321 // out with useBuildVector set. 4322 unsigned InputUsed[2] = {-1U, -1U}; // Not yet discovered. 4323 unsigned FirstMaskIdx = High * NewElts; 4324 bool UseBuildVector = false; 4325 for (unsigned MaskOffset = 0; MaskOffset < NewElts; ++MaskOffset) { 4326 // The mask element. This indexes into the input. 4327 int Idx = Mask[FirstMaskIdx + MaskOffset]; 4328 4329 // The input vector this mask element indexes into. 4330 unsigned Input = (unsigned)Idx / NewElts; 4331 4332 if (Input >= std::size(Inputs)) { 4333 // The mask element does not index into any input vector. 4334 Ops.push_back(-1); 4335 continue; 4336 } 4337 4338 // Turn the index into an offset from the start of the input vector. 4339 Idx -= Input * NewElts; 4340 4341 // Find or create a shuffle vector operand to hold this input. 4342 unsigned OpNo; 4343 for (OpNo = 0; OpNo < std::size(InputUsed); ++OpNo) { 4344 if (InputUsed[OpNo] == Input) { 4345 // This input vector is already an operand. 4346 break; 4347 } else if (InputUsed[OpNo] == -1U) { 4348 // Create a new operand for this input vector. 4349 InputUsed[OpNo] = Input; 4350 break; 4351 } 4352 } 4353 4354 if (OpNo >= std::size(InputUsed)) { 4355 // More than two input vectors used! Give up on trying to create a 4356 // shuffle vector. Insert all elements into a BUILD_VECTOR instead. 4357 UseBuildVector = true; 4358 break; 4359 } 4360 4361 // Add the mask index for the new shuffle vector. 4362 Ops.push_back(Idx + OpNo * NewElts); 4363 } 4364 4365 if (UseBuildVector) { 4366 LLT EltTy = NarrowTy.getElementType(); 4367 SmallVector<Register, 16> SVOps; 4368 4369 // Extract the input elements by hand. 4370 for (unsigned MaskOffset = 0; MaskOffset < NewElts; ++MaskOffset) { 4371 // The mask element. This indexes into the input. 4372 int Idx = Mask[FirstMaskIdx + MaskOffset]; 4373 4374 // The input vector this mask element indexes into. 4375 unsigned Input = (unsigned)Idx / NewElts; 4376 4377 if (Input >= std::size(Inputs)) { 4378 // The mask element is "undef" or indexes off the end of the input. 4379 SVOps.push_back(MIRBuilder.buildUndef(EltTy).getReg(0)); 4380 continue; 4381 } 4382 4383 // Turn the index into an offset from the start of the input vector. 4384 Idx -= Input * NewElts; 4385 4386 // Extract the vector element by hand. 4387 SVOps.push_back(MIRBuilder 4388 .buildExtractVectorElement( 4389 EltTy, Inputs[Input], 4390 MIRBuilder.buildConstant(LLT::scalar(32), Idx)) 4391 .getReg(0)); 4392 } 4393 4394 // Construct the Lo/Hi output using a G_BUILD_VECTOR. 4395 Output = MIRBuilder.buildBuildVector(NarrowTy, SVOps).getReg(0); 4396 } else if (InputUsed[0] == -1U) { 4397 // No input vectors were used! The result is undefined. 4398 Output = MIRBuilder.buildUndef(NarrowTy).getReg(0); 4399 } else { 4400 Register Op0 = Inputs[InputUsed[0]]; 4401 // If only one input was used, use an undefined vector for the other. 4402 Register Op1 = InputUsed[1] == -1U 4403 ? MIRBuilder.buildUndef(NarrowTy).getReg(0) 4404 : Inputs[InputUsed[1]]; 4405 // At least one input vector was used. Create a new shuffle vector. 4406 Output = MIRBuilder.buildShuffleVector(NarrowTy, Op0, Op1, Ops).getReg(0); 4407 } 4408 4409 Ops.clear(); 4410 } 4411 4412 MIRBuilder.buildConcatVectors(DstReg, {Lo, Hi}); 4413 MI.eraseFromParent(); 4414 return Legalized; 4415 } 4416 4417 static unsigned getScalarOpcForReduction(unsigned Opc) { 4418 unsigned ScalarOpc; 4419 switch (Opc) { 4420 case TargetOpcode::G_VECREDUCE_FADD: 4421 ScalarOpc = TargetOpcode::G_FADD; 4422 break; 4423 case TargetOpcode::G_VECREDUCE_FMUL: 4424 ScalarOpc = TargetOpcode::G_FMUL; 4425 break; 4426 case TargetOpcode::G_VECREDUCE_FMAX: 4427 ScalarOpc = TargetOpcode::G_FMAXNUM; 4428 break; 4429 case TargetOpcode::G_VECREDUCE_FMIN: 4430 ScalarOpc = TargetOpcode::G_FMINNUM; 4431 break; 4432 case TargetOpcode::G_VECREDUCE_ADD: 4433 ScalarOpc = TargetOpcode::G_ADD; 4434 break; 4435 case TargetOpcode::G_VECREDUCE_MUL: 4436 ScalarOpc = TargetOpcode::G_MUL; 4437 break; 4438 case TargetOpcode::G_VECREDUCE_AND: 4439 ScalarOpc = TargetOpcode::G_AND; 4440 break; 4441 case TargetOpcode::G_VECREDUCE_OR: 4442 ScalarOpc = TargetOpcode::G_OR; 4443 break; 4444 case TargetOpcode::G_VECREDUCE_XOR: 4445 ScalarOpc = TargetOpcode::G_XOR; 4446 break; 4447 case TargetOpcode::G_VECREDUCE_SMAX: 4448 ScalarOpc = TargetOpcode::G_SMAX; 4449 break; 4450 case TargetOpcode::G_VECREDUCE_SMIN: 4451 ScalarOpc = TargetOpcode::G_SMIN; 4452 break; 4453 case TargetOpcode::G_VECREDUCE_UMAX: 4454 ScalarOpc = TargetOpcode::G_UMAX; 4455 break; 4456 case TargetOpcode::G_VECREDUCE_UMIN: 4457 ScalarOpc = TargetOpcode::G_UMIN; 4458 break; 4459 default: 4460 llvm_unreachable("Unhandled reduction"); 4461 } 4462 return ScalarOpc; 4463 } 4464 4465 LegalizerHelper::LegalizeResult LegalizerHelper::fewerElementsVectorReductions( 4466 MachineInstr &MI, unsigned int TypeIdx, LLT NarrowTy) { 4467 unsigned Opc = MI.getOpcode(); 4468 assert(Opc != TargetOpcode::G_VECREDUCE_SEQ_FADD && 4469 Opc != TargetOpcode::G_VECREDUCE_SEQ_FMUL && 4470 "Sequential reductions not expected"); 4471 4472 if (TypeIdx != 1) 4473 return UnableToLegalize; 4474 4475 // The semantics of the normal non-sequential reductions allow us to freely 4476 // re-associate the operation. 4477 Register SrcReg = MI.getOperand(1).getReg(); 4478 LLT SrcTy = MRI.getType(SrcReg); 4479 Register DstReg = MI.getOperand(0).getReg(); 4480 LLT DstTy = MRI.getType(DstReg); 4481 4482 if (NarrowTy.isVector() && 4483 (SrcTy.getNumElements() % NarrowTy.getNumElements() != 0)) 4484 return UnableToLegalize; 4485 4486 unsigned ScalarOpc = getScalarOpcForReduction(Opc); 4487 SmallVector<Register> SplitSrcs; 4488 // If NarrowTy is a scalar then we're being asked to scalarize. 4489 const unsigned NumParts = 4490 NarrowTy.isVector() ? SrcTy.getNumElements() / NarrowTy.getNumElements() 4491 : SrcTy.getNumElements(); 4492 4493 extractParts(SrcReg, NarrowTy, NumParts, SplitSrcs); 4494 if (NarrowTy.isScalar()) { 4495 if (DstTy != NarrowTy) 4496 return UnableToLegalize; // FIXME: handle implicit extensions. 4497 4498 if (isPowerOf2_32(NumParts)) { 4499 // Generate a tree of scalar operations to reduce the critical path. 4500 SmallVector<Register> PartialResults; 4501 unsigned NumPartsLeft = NumParts; 4502 while (NumPartsLeft > 1) { 4503 for (unsigned Idx = 0; Idx < NumPartsLeft - 1; Idx += 2) { 4504 PartialResults.emplace_back( 4505 MIRBuilder 4506 .buildInstr(ScalarOpc, {NarrowTy}, 4507 {SplitSrcs[Idx], SplitSrcs[Idx + 1]}) 4508 .getReg(0)); 4509 } 4510 SplitSrcs = PartialResults; 4511 PartialResults.clear(); 4512 NumPartsLeft = SplitSrcs.size(); 4513 } 4514 assert(SplitSrcs.size() == 1); 4515 MIRBuilder.buildCopy(DstReg, SplitSrcs[0]); 4516 MI.eraseFromParent(); 4517 return Legalized; 4518 } 4519 // If we can't generate a tree, then just do sequential operations. 4520 Register Acc = SplitSrcs[0]; 4521 for (unsigned Idx = 1; Idx < NumParts; ++Idx) 4522 Acc = MIRBuilder.buildInstr(ScalarOpc, {NarrowTy}, {Acc, SplitSrcs[Idx]}) 4523 .getReg(0); 4524 MIRBuilder.buildCopy(DstReg, Acc); 4525 MI.eraseFromParent(); 4526 return Legalized; 4527 } 4528 SmallVector<Register> PartialReductions; 4529 for (unsigned Part = 0; Part < NumParts; ++Part) { 4530 PartialReductions.push_back( 4531 MIRBuilder.buildInstr(Opc, {DstTy}, {SplitSrcs[Part]}).getReg(0)); 4532 } 4533 4534 4535 // If the types involved are powers of 2, we can generate intermediate vector 4536 // ops, before generating a final reduction operation. 4537 if (isPowerOf2_32(SrcTy.getNumElements()) && 4538 isPowerOf2_32(NarrowTy.getNumElements())) { 4539 return tryNarrowPow2Reduction(MI, SrcReg, SrcTy, NarrowTy, ScalarOpc); 4540 } 4541 4542 Register Acc = PartialReductions[0]; 4543 for (unsigned Part = 1; Part < NumParts; ++Part) { 4544 if (Part == NumParts - 1) { 4545 MIRBuilder.buildInstr(ScalarOpc, {DstReg}, 4546 {Acc, PartialReductions[Part]}); 4547 } else { 4548 Acc = MIRBuilder 4549 .buildInstr(ScalarOpc, {DstTy}, {Acc, PartialReductions[Part]}) 4550 .getReg(0); 4551 } 4552 } 4553 MI.eraseFromParent(); 4554 return Legalized; 4555 } 4556 4557 LegalizerHelper::LegalizeResult 4558 LegalizerHelper::tryNarrowPow2Reduction(MachineInstr &MI, Register SrcReg, 4559 LLT SrcTy, LLT NarrowTy, 4560 unsigned ScalarOpc) { 4561 SmallVector<Register> SplitSrcs; 4562 // Split the sources into NarrowTy size pieces. 4563 extractParts(SrcReg, NarrowTy, 4564 SrcTy.getNumElements() / NarrowTy.getNumElements(), SplitSrcs); 4565 // We're going to do a tree reduction using vector operations until we have 4566 // one NarrowTy size value left. 4567 while (SplitSrcs.size() > 1) { 4568 SmallVector<Register> PartialRdxs; 4569 for (unsigned Idx = 0; Idx < SplitSrcs.size()-1; Idx += 2) { 4570 Register LHS = SplitSrcs[Idx]; 4571 Register RHS = SplitSrcs[Idx + 1]; 4572 // Create the intermediate vector op. 4573 Register Res = 4574 MIRBuilder.buildInstr(ScalarOpc, {NarrowTy}, {LHS, RHS}).getReg(0); 4575 PartialRdxs.push_back(Res); 4576 } 4577 SplitSrcs = std::move(PartialRdxs); 4578 } 4579 // Finally generate the requested NarrowTy based reduction. 4580 Observer.changingInstr(MI); 4581 MI.getOperand(1).setReg(SplitSrcs[0]); 4582 Observer.changedInstr(MI); 4583 return Legalized; 4584 } 4585 4586 LegalizerHelper::LegalizeResult 4587 LegalizerHelper::narrowScalarShiftByConstant(MachineInstr &MI, const APInt &Amt, 4588 const LLT HalfTy, const LLT AmtTy) { 4589 4590 Register InL = MRI.createGenericVirtualRegister(HalfTy); 4591 Register InH = MRI.createGenericVirtualRegister(HalfTy); 4592 MIRBuilder.buildUnmerge({InL, InH}, MI.getOperand(1)); 4593 4594 if (Amt.isZero()) { 4595 MIRBuilder.buildMergeLikeInstr(MI.getOperand(0), {InL, InH}); 4596 MI.eraseFromParent(); 4597 return Legalized; 4598 } 4599 4600 LLT NVT = HalfTy; 4601 unsigned NVTBits = HalfTy.getSizeInBits(); 4602 unsigned VTBits = 2 * NVTBits; 4603 4604 SrcOp Lo(Register(0)), Hi(Register(0)); 4605 if (MI.getOpcode() == TargetOpcode::G_SHL) { 4606 if (Amt.ugt(VTBits)) { 4607 Lo = Hi = MIRBuilder.buildConstant(NVT, 0); 4608 } else if (Amt.ugt(NVTBits)) { 4609 Lo = MIRBuilder.buildConstant(NVT, 0); 4610 Hi = MIRBuilder.buildShl(NVT, InL, 4611 MIRBuilder.buildConstant(AmtTy, Amt - NVTBits)); 4612 } else if (Amt == NVTBits) { 4613 Lo = MIRBuilder.buildConstant(NVT, 0); 4614 Hi = InL; 4615 } else { 4616 Lo = MIRBuilder.buildShl(NVT, InL, MIRBuilder.buildConstant(AmtTy, Amt)); 4617 auto OrLHS = 4618 MIRBuilder.buildShl(NVT, InH, MIRBuilder.buildConstant(AmtTy, Amt)); 4619 auto OrRHS = MIRBuilder.buildLShr( 4620 NVT, InL, MIRBuilder.buildConstant(AmtTy, -Amt + NVTBits)); 4621 Hi = MIRBuilder.buildOr(NVT, OrLHS, OrRHS); 4622 } 4623 } else if (MI.getOpcode() == TargetOpcode::G_LSHR) { 4624 if (Amt.ugt(VTBits)) { 4625 Lo = Hi = MIRBuilder.buildConstant(NVT, 0); 4626 } else if (Amt.ugt(NVTBits)) { 4627 Lo = MIRBuilder.buildLShr(NVT, InH, 4628 MIRBuilder.buildConstant(AmtTy, Amt - NVTBits)); 4629 Hi = MIRBuilder.buildConstant(NVT, 0); 4630 } else if (Amt == NVTBits) { 4631 Lo = InH; 4632 Hi = MIRBuilder.buildConstant(NVT, 0); 4633 } else { 4634 auto ShiftAmtConst = MIRBuilder.buildConstant(AmtTy, Amt); 4635 4636 auto OrLHS = MIRBuilder.buildLShr(NVT, InL, ShiftAmtConst); 4637 auto OrRHS = MIRBuilder.buildShl( 4638 NVT, InH, MIRBuilder.buildConstant(AmtTy, -Amt + NVTBits)); 4639 4640 Lo = MIRBuilder.buildOr(NVT, OrLHS, OrRHS); 4641 Hi = MIRBuilder.buildLShr(NVT, InH, ShiftAmtConst); 4642 } 4643 } else { 4644 if (Amt.ugt(VTBits)) { 4645 Hi = Lo = MIRBuilder.buildAShr( 4646 NVT, InH, MIRBuilder.buildConstant(AmtTy, NVTBits - 1)); 4647 } else if (Amt.ugt(NVTBits)) { 4648 Lo = MIRBuilder.buildAShr(NVT, InH, 4649 MIRBuilder.buildConstant(AmtTy, Amt - NVTBits)); 4650 Hi = MIRBuilder.buildAShr(NVT, InH, 4651 MIRBuilder.buildConstant(AmtTy, NVTBits - 1)); 4652 } else if (Amt == NVTBits) { 4653 Lo = InH; 4654 Hi = MIRBuilder.buildAShr(NVT, InH, 4655 MIRBuilder.buildConstant(AmtTy, NVTBits - 1)); 4656 } else { 4657 auto ShiftAmtConst = MIRBuilder.buildConstant(AmtTy, Amt); 4658 4659 auto OrLHS = MIRBuilder.buildLShr(NVT, InL, ShiftAmtConst); 4660 auto OrRHS = MIRBuilder.buildShl( 4661 NVT, InH, MIRBuilder.buildConstant(AmtTy, -Amt + NVTBits)); 4662 4663 Lo = MIRBuilder.buildOr(NVT, OrLHS, OrRHS); 4664 Hi = MIRBuilder.buildAShr(NVT, InH, ShiftAmtConst); 4665 } 4666 } 4667 4668 MIRBuilder.buildMergeLikeInstr(MI.getOperand(0), {Lo, Hi}); 4669 MI.eraseFromParent(); 4670 4671 return Legalized; 4672 } 4673 4674 // TODO: Optimize if constant shift amount. 4675 LegalizerHelper::LegalizeResult 4676 LegalizerHelper::narrowScalarShift(MachineInstr &MI, unsigned TypeIdx, 4677 LLT RequestedTy) { 4678 if (TypeIdx == 1) { 4679 Observer.changingInstr(MI); 4680 narrowScalarSrc(MI, RequestedTy, 2); 4681 Observer.changedInstr(MI); 4682 return Legalized; 4683 } 4684 4685 Register DstReg = MI.getOperand(0).getReg(); 4686 LLT DstTy = MRI.getType(DstReg); 4687 if (DstTy.isVector()) 4688 return UnableToLegalize; 4689 4690 Register Amt = MI.getOperand(2).getReg(); 4691 LLT ShiftAmtTy = MRI.getType(Amt); 4692 const unsigned DstEltSize = DstTy.getScalarSizeInBits(); 4693 if (DstEltSize % 2 != 0) 4694 return UnableToLegalize; 4695 4696 // Ignore the input type. We can only go to exactly half the size of the 4697 // input. If that isn't small enough, the resulting pieces will be further 4698 // legalized. 4699 const unsigned NewBitSize = DstEltSize / 2; 4700 const LLT HalfTy = LLT::scalar(NewBitSize); 4701 const LLT CondTy = LLT::scalar(1); 4702 4703 if (auto VRegAndVal = getIConstantVRegValWithLookThrough(Amt, MRI)) { 4704 return narrowScalarShiftByConstant(MI, VRegAndVal->Value, HalfTy, 4705 ShiftAmtTy); 4706 } 4707 4708 // TODO: Expand with known bits. 4709 4710 // Handle the fully general expansion by an unknown amount. 4711 auto NewBits = MIRBuilder.buildConstant(ShiftAmtTy, NewBitSize); 4712 4713 Register InL = MRI.createGenericVirtualRegister(HalfTy); 4714 Register InH = MRI.createGenericVirtualRegister(HalfTy); 4715 MIRBuilder.buildUnmerge({InL, InH}, MI.getOperand(1)); 4716 4717 auto AmtExcess = MIRBuilder.buildSub(ShiftAmtTy, Amt, NewBits); 4718 auto AmtLack = MIRBuilder.buildSub(ShiftAmtTy, NewBits, Amt); 4719 4720 auto Zero = MIRBuilder.buildConstant(ShiftAmtTy, 0); 4721 auto IsShort = MIRBuilder.buildICmp(ICmpInst::ICMP_ULT, CondTy, Amt, NewBits); 4722 auto IsZero = MIRBuilder.buildICmp(ICmpInst::ICMP_EQ, CondTy, Amt, Zero); 4723 4724 Register ResultRegs[2]; 4725 switch (MI.getOpcode()) { 4726 case TargetOpcode::G_SHL: { 4727 // Short: ShAmt < NewBitSize 4728 auto LoS = MIRBuilder.buildShl(HalfTy, InL, Amt); 4729 4730 auto LoOr = MIRBuilder.buildLShr(HalfTy, InL, AmtLack); 4731 auto HiOr = MIRBuilder.buildShl(HalfTy, InH, Amt); 4732 auto HiS = MIRBuilder.buildOr(HalfTy, LoOr, HiOr); 4733 4734 // Long: ShAmt >= NewBitSize 4735 auto LoL = MIRBuilder.buildConstant(HalfTy, 0); // Lo part is zero. 4736 auto HiL = MIRBuilder.buildShl(HalfTy, InL, AmtExcess); // Hi from Lo part. 4737 4738 auto Lo = MIRBuilder.buildSelect(HalfTy, IsShort, LoS, LoL); 4739 auto Hi = MIRBuilder.buildSelect( 4740 HalfTy, IsZero, InH, MIRBuilder.buildSelect(HalfTy, IsShort, HiS, HiL)); 4741 4742 ResultRegs[0] = Lo.getReg(0); 4743 ResultRegs[1] = Hi.getReg(0); 4744 break; 4745 } 4746 case TargetOpcode::G_LSHR: 4747 case TargetOpcode::G_ASHR: { 4748 // Short: ShAmt < NewBitSize 4749 auto HiS = MIRBuilder.buildInstr(MI.getOpcode(), {HalfTy}, {InH, Amt}); 4750 4751 auto LoOr = MIRBuilder.buildLShr(HalfTy, InL, Amt); 4752 auto HiOr = MIRBuilder.buildShl(HalfTy, InH, AmtLack); 4753 auto LoS = MIRBuilder.buildOr(HalfTy, LoOr, HiOr); 4754 4755 // Long: ShAmt >= NewBitSize 4756 MachineInstrBuilder HiL; 4757 if (MI.getOpcode() == TargetOpcode::G_LSHR) { 4758 HiL = MIRBuilder.buildConstant(HalfTy, 0); // Hi part is zero. 4759 } else { 4760 auto ShiftAmt = MIRBuilder.buildConstant(ShiftAmtTy, NewBitSize - 1); 4761 HiL = MIRBuilder.buildAShr(HalfTy, InH, ShiftAmt); // Sign of Hi part. 4762 } 4763 auto LoL = MIRBuilder.buildInstr(MI.getOpcode(), {HalfTy}, 4764 {InH, AmtExcess}); // Lo from Hi part. 4765 4766 auto Lo = MIRBuilder.buildSelect( 4767 HalfTy, IsZero, InL, MIRBuilder.buildSelect(HalfTy, IsShort, LoS, LoL)); 4768 4769 auto Hi = MIRBuilder.buildSelect(HalfTy, IsShort, HiS, HiL); 4770 4771 ResultRegs[0] = Lo.getReg(0); 4772 ResultRegs[1] = Hi.getReg(0); 4773 break; 4774 } 4775 default: 4776 llvm_unreachable("not a shift"); 4777 } 4778 4779 MIRBuilder.buildMergeLikeInstr(DstReg, ResultRegs); 4780 MI.eraseFromParent(); 4781 return Legalized; 4782 } 4783 4784 LegalizerHelper::LegalizeResult 4785 LegalizerHelper::moreElementsVectorPhi(MachineInstr &MI, unsigned TypeIdx, 4786 LLT MoreTy) { 4787 assert(TypeIdx == 0 && "Expecting only Idx 0"); 4788 4789 Observer.changingInstr(MI); 4790 for (unsigned I = 1, E = MI.getNumOperands(); I != E; I += 2) { 4791 MachineBasicBlock &OpMBB = *MI.getOperand(I + 1).getMBB(); 4792 MIRBuilder.setInsertPt(OpMBB, OpMBB.getFirstTerminator()); 4793 moreElementsVectorSrc(MI, MoreTy, I); 4794 } 4795 4796 MachineBasicBlock &MBB = *MI.getParent(); 4797 MIRBuilder.setInsertPt(MBB, --MBB.getFirstNonPHI()); 4798 moreElementsVectorDst(MI, MoreTy, 0); 4799 Observer.changedInstr(MI); 4800 return Legalized; 4801 } 4802 4803 LegalizerHelper::LegalizeResult 4804 LegalizerHelper::moreElementsVector(MachineInstr &MI, unsigned TypeIdx, 4805 LLT MoreTy) { 4806 unsigned Opc = MI.getOpcode(); 4807 switch (Opc) { 4808 case TargetOpcode::G_IMPLICIT_DEF: 4809 case TargetOpcode::G_LOAD: { 4810 if (TypeIdx != 0) 4811 return UnableToLegalize; 4812 Observer.changingInstr(MI); 4813 moreElementsVectorDst(MI, MoreTy, 0); 4814 Observer.changedInstr(MI); 4815 return Legalized; 4816 } 4817 case TargetOpcode::G_STORE: 4818 if (TypeIdx != 0) 4819 return UnableToLegalize; 4820 Observer.changingInstr(MI); 4821 moreElementsVectorSrc(MI, MoreTy, 0); 4822 Observer.changedInstr(MI); 4823 return Legalized; 4824 case TargetOpcode::G_AND: 4825 case TargetOpcode::G_OR: 4826 case TargetOpcode::G_XOR: 4827 case TargetOpcode::G_ADD: 4828 case TargetOpcode::G_SUB: 4829 case TargetOpcode::G_MUL: 4830 case TargetOpcode::G_FADD: 4831 case TargetOpcode::G_FMUL: 4832 case TargetOpcode::G_UADDSAT: 4833 case TargetOpcode::G_USUBSAT: 4834 case TargetOpcode::G_SADDSAT: 4835 case TargetOpcode::G_SSUBSAT: 4836 case TargetOpcode::G_SMIN: 4837 case TargetOpcode::G_SMAX: 4838 case TargetOpcode::G_UMIN: 4839 case TargetOpcode::G_UMAX: 4840 case TargetOpcode::G_FMINNUM: 4841 case TargetOpcode::G_FMAXNUM: 4842 case TargetOpcode::G_FMINNUM_IEEE: 4843 case TargetOpcode::G_FMAXNUM_IEEE: 4844 case TargetOpcode::G_FMINIMUM: 4845 case TargetOpcode::G_FMAXIMUM: 4846 case TargetOpcode::G_STRICT_FADD: 4847 case TargetOpcode::G_STRICT_FSUB: 4848 case TargetOpcode::G_STRICT_FMUL: { 4849 Observer.changingInstr(MI); 4850 moreElementsVectorSrc(MI, MoreTy, 1); 4851 moreElementsVectorSrc(MI, MoreTy, 2); 4852 moreElementsVectorDst(MI, MoreTy, 0); 4853 Observer.changedInstr(MI); 4854 return Legalized; 4855 } 4856 case TargetOpcode::G_FMA: 4857 case TargetOpcode::G_STRICT_FMA: 4858 case TargetOpcode::G_FSHR: 4859 case TargetOpcode::G_FSHL: { 4860 Observer.changingInstr(MI); 4861 moreElementsVectorSrc(MI, MoreTy, 1); 4862 moreElementsVectorSrc(MI, MoreTy, 2); 4863 moreElementsVectorSrc(MI, MoreTy, 3); 4864 moreElementsVectorDst(MI, MoreTy, 0); 4865 Observer.changedInstr(MI); 4866 return Legalized; 4867 } 4868 case TargetOpcode::G_EXTRACT: 4869 if (TypeIdx != 1) 4870 return UnableToLegalize; 4871 Observer.changingInstr(MI); 4872 moreElementsVectorSrc(MI, MoreTy, 1); 4873 Observer.changedInstr(MI); 4874 return Legalized; 4875 case TargetOpcode::G_INSERT: 4876 case TargetOpcode::G_FREEZE: 4877 case TargetOpcode::G_FNEG: 4878 case TargetOpcode::G_FABS: 4879 case TargetOpcode::G_BSWAP: 4880 case TargetOpcode::G_FCANONICALIZE: 4881 case TargetOpcode::G_SEXT_INREG: 4882 if (TypeIdx != 0) 4883 return UnableToLegalize; 4884 Observer.changingInstr(MI); 4885 moreElementsVectorSrc(MI, MoreTy, 1); 4886 moreElementsVectorDst(MI, MoreTy, 0); 4887 Observer.changedInstr(MI); 4888 return Legalized; 4889 case TargetOpcode::G_SELECT: { 4890 Register DstReg = MI.getOperand(0).getReg(); 4891 Register CondReg = MI.getOperand(1).getReg(); 4892 LLT DstTy = MRI.getType(DstReg); 4893 LLT CondTy = MRI.getType(CondReg); 4894 if (TypeIdx == 1) { 4895 if (!CondTy.isScalar() || 4896 DstTy.getElementCount() != MoreTy.getElementCount()) 4897 return UnableToLegalize; 4898 4899 // This is turning a scalar select of vectors into a vector 4900 // select. Broadcast the select condition. 4901 auto ShufSplat = MIRBuilder.buildShuffleSplat(MoreTy, CondReg); 4902 Observer.changingInstr(MI); 4903 MI.getOperand(1).setReg(ShufSplat.getReg(0)); 4904 Observer.changedInstr(MI); 4905 return Legalized; 4906 } 4907 4908 if (CondTy.isVector()) 4909 return UnableToLegalize; 4910 4911 Observer.changingInstr(MI); 4912 moreElementsVectorSrc(MI, MoreTy, 2); 4913 moreElementsVectorSrc(MI, MoreTy, 3); 4914 moreElementsVectorDst(MI, MoreTy, 0); 4915 Observer.changedInstr(MI); 4916 return Legalized; 4917 } 4918 case TargetOpcode::G_UNMERGE_VALUES: 4919 return UnableToLegalize; 4920 case TargetOpcode::G_PHI: 4921 return moreElementsVectorPhi(MI, TypeIdx, MoreTy); 4922 case TargetOpcode::G_SHUFFLE_VECTOR: 4923 return moreElementsVectorShuffle(MI, TypeIdx, MoreTy); 4924 case TargetOpcode::G_BUILD_VECTOR: { 4925 SmallVector<SrcOp, 8> Elts; 4926 for (auto Op : MI.uses()) { 4927 Elts.push_back(Op.getReg()); 4928 } 4929 4930 for (unsigned i = Elts.size(); i < MoreTy.getNumElements(); ++i) { 4931 Elts.push_back(MIRBuilder.buildUndef(MoreTy.getScalarType())); 4932 } 4933 4934 MIRBuilder.buildDeleteTrailingVectorElements( 4935 MI.getOperand(0).getReg(), MIRBuilder.buildInstr(Opc, {MoreTy}, Elts)); 4936 MI.eraseFromParent(); 4937 return Legalized; 4938 } 4939 case TargetOpcode::G_TRUNC: { 4940 Observer.changingInstr(MI); 4941 moreElementsVectorSrc(MI, MoreTy, 1); 4942 moreElementsVectorDst(MI, MoreTy, 0); 4943 Observer.changedInstr(MI); 4944 return Legalized; 4945 } 4946 default: 4947 return UnableToLegalize; 4948 } 4949 } 4950 4951 /// Expand source vectors to the size of destination vector. 4952 static LegalizerHelper::LegalizeResult 4953 equalizeVectorShuffleLengths(MachineInstr &MI, MachineIRBuilder &MIRBuilder) { 4954 MachineRegisterInfo &MRI = *MIRBuilder.getMRI(); 4955 4956 LLT DstTy = MRI.getType(MI.getOperand(0).getReg()); 4957 LLT SrcTy = MRI.getType(MI.getOperand(1).getReg()); 4958 ArrayRef<int> Mask = MI.getOperand(3).getShuffleMask(); 4959 unsigned MaskNumElts = Mask.size(); 4960 unsigned SrcNumElts = SrcTy.getNumElements(); 4961 Register DstReg = MI.getOperand(0).getReg(); 4962 LLT DestEltTy = DstTy.getElementType(); 4963 4964 // TODO: Normalize the shuffle vector since mask and vector length don't 4965 // match. 4966 if (MaskNumElts <= SrcNumElts) { 4967 return LegalizerHelper::LegalizeResult::UnableToLegalize; 4968 } 4969 4970 unsigned PaddedMaskNumElts = alignTo(MaskNumElts, SrcNumElts); 4971 unsigned NumConcat = PaddedMaskNumElts / SrcNumElts; 4972 LLT PaddedTy = LLT::fixed_vector(PaddedMaskNumElts, DestEltTy); 4973 4974 // Create new source vectors by concatenating the initial 4975 // source vectors with undefined vectors of the same size. 4976 auto Undef = MIRBuilder.buildUndef(SrcTy); 4977 SmallVector<Register, 8> MOps1(NumConcat, Undef.getReg(0)); 4978 SmallVector<Register, 8> MOps2(NumConcat, Undef.getReg(0)); 4979 MOps1[0] = MI.getOperand(1).getReg(); 4980 MOps2[0] = MI.getOperand(2).getReg(); 4981 4982 auto Src1 = MIRBuilder.buildConcatVectors(PaddedTy, MOps1); 4983 auto Src2 = MIRBuilder.buildConcatVectors(PaddedTy, MOps2); 4984 4985 // Readjust mask for new input vector length. 4986 SmallVector<int, 8> MappedOps(PaddedMaskNumElts, -1); 4987 for (unsigned I = 0; I != MaskNumElts; ++I) { 4988 int Idx = Mask[I]; 4989 if (Idx >= static_cast<int>(SrcNumElts)) 4990 Idx += PaddedMaskNumElts - SrcNumElts; 4991 MappedOps[I] = Idx; 4992 } 4993 4994 // If we got more elements than required, extract subvector. 4995 if (MaskNumElts != PaddedMaskNumElts) { 4996 auto Shuffle = 4997 MIRBuilder.buildShuffleVector(PaddedTy, Src1, Src2, MappedOps); 4998 4999 SmallVector<Register, 16> Elts(MaskNumElts); 5000 for (unsigned I = 0; I < MaskNumElts; ++I) { 5001 Elts[I] = 5002 MIRBuilder.buildExtractVectorElementConstant(DestEltTy, Shuffle, I) 5003 .getReg(0); 5004 } 5005 MIRBuilder.buildBuildVector(DstReg, Elts); 5006 } else { 5007 MIRBuilder.buildShuffleVector(DstReg, Src1, Src2, MappedOps); 5008 } 5009 5010 MI.eraseFromParent(); 5011 return LegalizerHelper::LegalizeResult::Legalized; 5012 } 5013 5014 LegalizerHelper::LegalizeResult 5015 LegalizerHelper::moreElementsVectorShuffle(MachineInstr &MI, 5016 unsigned int TypeIdx, LLT MoreTy) { 5017 Register DstReg = MI.getOperand(0).getReg(); 5018 Register Src1Reg = MI.getOperand(1).getReg(); 5019 Register Src2Reg = MI.getOperand(2).getReg(); 5020 ArrayRef<int> Mask = MI.getOperand(3).getShuffleMask(); 5021 LLT DstTy = MRI.getType(DstReg); 5022 LLT Src1Ty = MRI.getType(Src1Reg); 5023 LLT Src2Ty = MRI.getType(Src2Reg); 5024 unsigned NumElts = DstTy.getNumElements(); 5025 unsigned WidenNumElts = MoreTy.getNumElements(); 5026 5027 if (DstTy.isVector() && Src1Ty.isVector() && 5028 DstTy.getNumElements() > Src1Ty.getNumElements()) { 5029 return equalizeVectorShuffleLengths(MI, MIRBuilder); 5030 } 5031 5032 if (TypeIdx != 0) 5033 return UnableToLegalize; 5034 5035 // Expect a canonicalized shuffle. 5036 if (DstTy != Src1Ty || DstTy != Src2Ty) 5037 return UnableToLegalize; 5038 5039 moreElementsVectorSrc(MI, MoreTy, 1); 5040 moreElementsVectorSrc(MI, MoreTy, 2); 5041 5042 // Adjust mask based on new input vector length. 5043 SmallVector<int, 16> NewMask; 5044 for (unsigned I = 0; I != NumElts; ++I) { 5045 int Idx = Mask[I]; 5046 if (Idx < static_cast<int>(NumElts)) 5047 NewMask.push_back(Idx); 5048 else 5049 NewMask.push_back(Idx - NumElts + WidenNumElts); 5050 } 5051 for (unsigned I = NumElts; I != WidenNumElts; ++I) 5052 NewMask.push_back(-1); 5053 moreElementsVectorDst(MI, MoreTy, 0); 5054 MIRBuilder.setInstrAndDebugLoc(MI); 5055 MIRBuilder.buildShuffleVector(MI.getOperand(0).getReg(), 5056 MI.getOperand(1).getReg(), 5057 MI.getOperand(2).getReg(), NewMask); 5058 MI.eraseFromParent(); 5059 return Legalized; 5060 } 5061 5062 void LegalizerHelper::multiplyRegisters(SmallVectorImpl<Register> &DstRegs, 5063 ArrayRef<Register> Src1Regs, 5064 ArrayRef<Register> Src2Regs, 5065 LLT NarrowTy) { 5066 MachineIRBuilder &B = MIRBuilder; 5067 unsigned SrcParts = Src1Regs.size(); 5068 unsigned DstParts = DstRegs.size(); 5069 5070 unsigned DstIdx = 0; // Low bits of the result. 5071 Register FactorSum = 5072 B.buildMul(NarrowTy, Src1Regs[DstIdx], Src2Regs[DstIdx]).getReg(0); 5073 DstRegs[DstIdx] = FactorSum; 5074 5075 unsigned CarrySumPrevDstIdx; 5076 SmallVector<Register, 4> Factors; 5077 5078 for (DstIdx = 1; DstIdx < DstParts; DstIdx++) { 5079 // Collect low parts of muls for DstIdx. 5080 for (unsigned i = DstIdx + 1 < SrcParts ? 0 : DstIdx - SrcParts + 1; 5081 i <= std::min(DstIdx, SrcParts - 1); ++i) { 5082 MachineInstrBuilder Mul = 5083 B.buildMul(NarrowTy, Src1Regs[DstIdx - i], Src2Regs[i]); 5084 Factors.push_back(Mul.getReg(0)); 5085 } 5086 // Collect high parts of muls from previous DstIdx. 5087 for (unsigned i = DstIdx < SrcParts ? 0 : DstIdx - SrcParts; 5088 i <= std::min(DstIdx - 1, SrcParts - 1); ++i) { 5089 MachineInstrBuilder Umulh = 5090 B.buildUMulH(NarrowTy, Src1Regs[DstIdx - 1 - i], Src2Regs[i]); 5091 Factors.push_back(Umulh.getReg(0)); 5092 } 5093 // Add CarrySum from additions calculated for previous DstIdx. 5094 if (DstIdx != 1) { 5095 Factors.push_back(CarrySumPrevDstIdx); 5096 } 5097 5098 Register CarrySum; 5099 // Add all factors and accumulate all carries into CarrySum. 5100 if (DstIdx != DstParts - 1) { 5101 MachineInstrBuilder Uaddo = 5102 B.buildUAddo(NarrowTy, LLT::scalar(1), Factors[0], Factors[1]); 5103 FactorSum = Uaddo.getReg(0); 5104 CarrySum = B.buildZExt(NarrowTy, Uaddo.getReg(1)).getReg(0); 5105 for (unsigned i = 2; i < Factors.size(); ++i) { 5106 MachineInstrBuilder Uaddo = 5107 B.buildUAddo(NarrowTy, LLT::scalar(1), FactorSum, Factors[i]); 5108 FactorSum = Uaddo.getReg(0); 5109 MachineInstrBuilder Carry = B.buildZExt(NarrowTy, Uaddo.getReg(1)); 5110 CarrySum = B.buildAdd(NarrowTy, CarrySum, Carry).getReg(0); 5111 } 5112 } else { 5113 // Since value for the next index is not calculated, neither is CarrySum. 5114 FactorSum = B.buildAdd(NarrowTy, Factors[0], Factors[1]).getReg(0); 5115 for (unsigned i = 2; i < Factors.size(); ++i) 5116 FactorSum = B.buildAdd(NarrowTy, FactorSum, Factors[i]).getReg(0); 5117 } 5118 5119 CarrySumPrevDstIdx = CarrySum; 5120 DstRegs[DstIdx] = FactorSum; 5121 Factors.clear(); 5122 } 5123 } 5124 5125 LegalizerHelper::LegalizeResult 5126 LegalizerHelper::narrowScalarAddSub(MachineInstr &MI, unsigned TypeIdx, 5127 LLT NarrowTy) { 5128 if (TypeIdx != 0) 5129 return UnableToLegalize; 5130 5131 Register DstReg = MI.getOperand(0).getReg(); 5132 LLT DstType = MRI.getType(DstReg); 5133 // FIXME: add support for vector types 5134 if (DstType.isVector()) 5135 return UnableToLegalize; 5136 5137 unsigned Opcode = MI.getOpcode(); 5138 unsigned OpO, OpE, OpF; 5139 switch (Opcode) { 5140 case TargetOpcode::G_SADDO: 5141 case TargetOpcode::G_SADDE: 5142 case TargetOpcode::G_UADDO: 5143 case TargetOpcode::G_UADDE: 5144 case TargetOpcode::G_ADD: 5145 OpO = TargetOpcode::G_UADDO; 5146 OpE = TargetOpcode::G_UADDE; 5147 OpF = TargetOpcode::G_UADDE; 5148 if (Opcode == TargetOpcode::G_SADDO || Opcode == TargetOpcode::G_SADDE) 5149 OpF = TargetOpcode::G_SADDE; 5150 break; 5151 case TargetOpcode::G_SSUBO: 5152 case TargetOpcode::G_SSUBE: 5153 case TargetOpcode::G_USUBO: 5154 case TargetOpcode::G_USUBE: 5155 case TargetOpcode::G_SUB: 5156 OpO = TargetOpcode::G_USUBO; 5157 OpE = TargetOpcode::G_USUBE; 5158 OpF = TargetOpcode::G_USUBE; 5159 if (Opcode == TargetOpcode::G_SSUBO || Opcode == TargetOpcode::G_SSUBE) 5160 OpF = TargetOpcode::G_SSUBE; 5161 break; 5162 default: 5163 llvm_unreachable("Unexpected add/sub opcode!"); 5164 } 5165 5166 // 1 for a plain add/sub, 2 if this is an operation with a carry-out. 5167 unsigned NumDefs = MI.getNumExplicitDefs(); 5168 Register Src1 = MI.getOperand(NumDefs).getReg(); 5169 Register Src2 = MI.getOperand(NumDefs + 1).getReg(); 5170 Register CarryDst, CarryIn; 5171 if (NumDefs == 2) 5172 CarryDst = MI.getOperand(1).getReg(); 5173 if (MI.getNumOperands() == NumDefs + 3) 5174 CarryIn = MI.getOperand(NumDefs + 2).getReg(); 5175 5176 LLT RegTy = MRI.getType(MI.getOperand(0).getReg()); 5177 LLT LeftoverTy, DummyTy; 5178 SmallVector<Register, 2> Src1Regs, Src2Regs, Src1Left, Src2Left, DstRegs; 5179 extractParts(Src1, RegTy, NarrowTy, LeftoverTy, Src1Regs, Src1Left); 5180 extractParts(Src2, RegTy, NarrowTy, DummyTy, Src2Regs, Src2Left); 5181 5182 int NarrowParts = Src1Regs.size(); 5183 for (int I = 0, E = Src1Left.size(); I != E; ++I) { 5184 Src1Regs.push_back(Src1Left[I]); 5185 Src2Regs.push_back(Src2Left[I]); 5186 } 5187 DstRegs.reserve(Src1Regs.size()); 5188 5189 for (int i = 0, e = Src1Regs.size(); i != e; ++i) { 5190 Register DstReg = 5191 MRI.createGenericVirtualRegister(MRI.getType(Src1Regs[i])); 5192 Register CarryOut = MRI.createGenericVirtualRegister(LLT::scalar(1)); 5193 // Forward the final carry-out to the destination register 5194 if (i == e - 1 && CarryDst) 5195 CarryOut = CarryDst; 5196 5197 if (!CarryIn) { 5198 MIRBuilder.buildInstr(OpO, {DstReg, CarryOut}, 5199 {Src1Regs[i], Src2Regs[i]}); 5200 } else if (i == e - 1) { 5201 MIRBuilder.buildInstr(OpF, {DstReg, CarryOut}, 5202 {Src1Regs[i], Src2Regs[i], CarryIn}); 5203 } else { 5204 MIRBuilder.buildInstr(OpE, {DstReg, CarryOut}, 5205 {Src1Regs[i], Src2Regs[i], CarryIn}); 5206 } 5207 5208 DstRegs.push_back(DstReg); 5209 CarryIn = CarryOut; 5210 } 5211 insertParts(MI.getOperand(0).getReg(), RegTy, NarrowTy, 5212 ArrayRef(DstRegs).take_front(NarrowParts), LeftoverTy, 5213 ArrayRef(DstRegs).drop_front(NarrowParts)); 5214 5215 MI.eraseFromParent(); 5216 return Legalized; 5217 } 5218 5219 LegalizerHelper::LegalizeResult 5220 LegalizerHelper::narrowScalarMul(MachineInstr &MI, LLT NarrowTy) { 5221 Register DstReg = MI.getOperand(0).getReg(); 5222 Register Src1 = MI.getOperand(1).getReg(); 5223 Register Src2 = MI.getOperand(2).getReg(); 5224 5225 LLT Ty = MRI.getType(DstReg); 5226 if (Ty.isVector()) 5227 return UnableToLegalize; 5228 5229 unsigned Size = Ty.getSizeInBits(); 5230 unsigned NarrowSize = NarrowTy.getSizeInBits(); 5231 if (Size % NarrowSize != 0) 5232 return UnableToLegalize; 5233 5234 unsigned NumParts = Size / NarrowSize; 5235 bool IsMulHigh = MI.getOpcode() == TargetOpcode::G_UMULH; 5236 unsigned DstTmpParts = NumParts * (IsMulHigh ? 2 : 1); 5237 5238 SmallVector<Register, 2> Src1Parts, Src2Parts; 5239 SmallVector<Register, 2> DstTmpRegs(DstTmpParts); 5240 extractParts(Src1, NarrowTy, NumParts, Src1Parts); 5241 extractParts(Src2, NarrowTy, NumParts, Src2Parts); 5242 multiplyRegisters(DstTmpRegs, Src1Parts, Src2Parts, NarrowTy); 5243 5244 // Take only high half of registers if this is high mul. 5245 ArrayRef<Register> DstRegs(&DstTmpRegs[DstTmpParts - NumParts], NumParts); 5246 MIRBuilder.buildMergeLikeInstr(DstReg, DstRegs); 5247 MI.eraseFromParent(); 5248 return Legalized; 5249 } 5250 5251 LegalizerHelper::LegalizeResult 5252 LegalizerHelper::narrowScalarFPTOI(MachineInstr &MI, unsigned TypeIdx, 5253 LLT NarrowTy) { 5254 if (TypeIdx != 0) 5255 return UnableToLegalize; 5256 5257 bool IsSigned = MI.getOpcode() == TargetOpcode::G_FPTOSI; 5258 5259 Register Src = MI.getOperand(1).getReg(); 5260 LLT SrcTy = MRI.getType(Src); 5261 5262 // If all finite floats fit into the narrowed integer type, we can just swap 5263 // out the result type. This is practically only useful for conversions from 5264 // half to at least 16-bits, so just handle the one case. 5265 if (SrcTy.getScalarType() != LLT::scalar(16) || 5266 NarrowTy.getScalarSizeInBits() < (IsSigned ? 17u : 16u)) 5267 return UnableToLegalize; 5268 5269 Observer.changingInstr(MI); 5270 narrowScalarDst(MI, NarrowTy, 0, 5271 IsSigned ? TargetOpcode::G_SEXT : TargetOpcode::G_ZEXT); 5272 Observer.changedInstr(MI); 5273 return Legalized; 5274 } 5275 5276 LegalizerHelper::LegalizeResult 5277 LegalizerHelper::narrowScalarExtract(MachineInstr &MI, unsigned TypeIdx, 5278 LLT NarrowTy) { 5279 if (TypeIdx != 1) 5280 return UnableToLegalize; 5281 5282 uint64_t NarrowSize = NarrowTy.getSizeInBits(); 5283 5284 int64_t SizeOp1 = MRI.getType(MI.getOperand(1).getReg()).getSizeInBits(); 5285 // FIXME: add support for when SizeOp1 isn't an exact multiple of 5286 // NarrowSize. 5287 if (SizeOp1 % NarrowSize != 0) 5288 return UnableToLegalize; 5289 int NumParts = SizeOp1 / NarrowSize; 5290 5291 SmallVector<Register, 2> SrcRegs, DstRegs; 5292 SmallVector<uint64_t, 2> Indexes; 5293 extractParts(MI.getOperand(1).getReg(), NarrowTy, NumParts, SrcRegs); 5294 5295 Register OpReg = MI.getOperand(0).getReg(); 5296 uint64_t OpStart = MI.getOperand(2).getImm(); 5297 uint64_t OpSize = MRI.getType(OpReg).getSizeInBits(); 5298 for (int i = 0; i < NumParts; ++i) { 5299 unsigned SrcStart = i * NarrowSize; 5300 5301 if (SrcStart + NarrowSize <= OpStart || SrcStart >= OpStart + OpSize) { 5302 // No part of the extract uses this subregister, ignore it. 5303 continue; 5304 } else if (SrcStart == OpStart && NarrowTy == MRI.getType(OpReg)) { 5305 // The entire subregister is extracted, forward the value. 5306 DstRegs.push_back(SrcRegs[i]); 5307 continue; 5308 } 5309 5310 // OpSegStart is where this destination segment would start in OpReg if it 5311 // extended infinitely in both directions. 5312 int64_t ExtractOffset; 5313 uint64_t SegSize; 5314 if (OpStart < SrcStart) { 5315 ExtractOffset = 0; 5316 SegSize = std::min(NarrowSize, OpStart + OpSize - SrcStart); 5317 } else { 5318 ExtractOffset = OpStart - SrcStart; 5319 SegSize = std::min(SrcStart + NarrowSize - OpStart, OpSize); 5320 } 5321 5322 Register SegReg = SrcRegs[i]; 5323 if (ExtractOffset != 0 || SegSize != NarrowSize) { 5324 // A genuine extract is needed. 5325 SegReg = MRI.createGenericVirtualRegister(LLT::scalar(SegSize)); 5326 MIRBuilder.buildExtract(SegReg, SrcRegs[i], ExtractOffset); 5327 } 5328 5329 DstRegs.push_back(SegReg); 5330 } 5331 5332 Register DstReg = MI.getOperand(0).getReg(); 5333 if (MRI.getType(DstReg).isVector()) 5334 MIRBuilder.buildBuildVector(DstReg, DstRegs); 5335 else if (DstRegs.size() > 1) 5336 MIRBuilder.buildMergeLikeInstr(DstReg, DstRegs); 5337 else 5338 MIRBuilder.buildCopy(DstReg, DstRegs[0]); 5339 MI.eraseFromParent(); 5340 return Legalized; 5341 } 5342 5343 LegalizerHelper::LegalizeResult 5344 LegalizerHelper::narrowScalarInsert(MachineInstr &MI, unsigned TypeIdx, 5345 LLT NarrowTy) { 5346 // FIXME: Don't know how to handle secondary types yet. 5347 if (TypeIdx != 0) 5348 return UnableToLegalize; 5349 5350 SmallVector<Register, 2> SrcRegs, LeftoverRegs, DstRegs; 5351 SmallVector<uint64_t, 2> Indexes; 5352 LLT RegTy = MRI.getType(MI.getOperand(0).getReg()); 5353 LLT LeftoverTy; 5354 extractParts(MI.getOperand(1).getReg(), RegTy, NarrowTy, LeftoverTy, SrcRegs, 5355 LeftoverRegs); 5356 5357 for (Register Reg : LeftoverRegs) 5358 SrcRegs.push_back(Reg); 5359 5360 uint64_t NarrowSize = NarrowTy.getSizeInBits(); 5361 Register OpReg = MI.getOperand(2).getReg(); 5362 uint64_t OpStart = MI.getOperand(3).getImm(); 5363 uint64_t OpSize = MRI.getType(OpReg).getSizeInBits(); 5364 for (int I = 0, E = SrcRegs.size(); I != E; ++I) { 5365 unsigned DstStart = I * NarrowSize; 5366 5367 if (DstStart == OpStart && NarrowTy == MRI.getType(OpReg)) { 5368 // The entire subregister is defined by this insert, forward the new 5369 // value. 5370 DstRegs.push_back(OpReg); 5371 continue; 5372 } 5373 5374 Register SrcReg = SrcRegs[I]; 5375 if (MRI.getType(SrcRegs[I]) == LeftoverTy) { 5376 // The leftover reg is smaller than NarrowTy, so we need to extend it. 5377 SrcReg = MRI.createGenericVirtualRegister(NarrowTy); 5378 MIRBuilder.buildAnyExt(SrcReg, SrcRegs[I]); 5379 } 5380 5381 if (DstStart + NarrowSize <= OpStart || DstStart >= OpStart + OpSize) { 5382 // No part of the insert affects this subregister, forward the original. 5383 DstRegs.push_back(SrcReg); 5384 continue; 5385 } 5386 5387 // OpSegStart is where this destination segment would start in OpReg if it 5388 // extended infinitely in both directions. 5389 int64_t ExtractOffset, InsertOffset; 5390 uint64_t SegSize; 5391 if (OpStart < DstStart) { 5392 InsertOffset = 0; 5393 ExtractOffset = DstStart - OpStart; 5394 SegSize = std::min(NarrowSize, OpStart + OpSize - DstStart); 5395 } else { 5396 InsertOffset = OpStart - DstStart; 5397 ExtractOffset = 0; 5398 SegSize = 5399 std::min(NarrowSize - InsertOffset, OpStart + OpSize - DstStart); 5400 } 5401 5402 Register SegReg = OpReg; 5403 if (ExtractOffset != 0 || SegSize != OpSize) { 5404 // A genuine extract is needed. 5405 SegReg = MRI.createGenericVirtualRegister(LLT::scalar(SegSize)); 5406 MIRBuilder.buildExtract(SegReg, OpReg, ExtractOffset); 5407 } 5408 5409 Register DstReg = MRI.createGenericVirtualRegister(NarrowTy); 5410 MIRBuilder.buildInsert(DstReg, SrcReg, SegReg, InsertOffset); 5411 DstRegs.push_back(DstReg); 5412 } 5413 5414 uint64_t WideSize = DstRegs.size() * NarrowSize; 5415 Register DstReg = MI.getOperand(0).getReg(); 5416 if (WideSize > RegTy.getSizeInBits()) { 5417 Register MergeReg = MRI.createGenericVirtualRegister(LLT::scalar(WideSize)); 5418 MIRBuilder.buildMergeLikeInstr(MergeReg, DstRegs); 5419 MIRBuilder.buildTrunc(DstReg, MergeReg); 5420 } else 5421 MIRBuilder.buildMergeLikeInstr(DstReg, DstRegs); 5422 5423 MI.eraseFromParent(); 5424 return Legalized; 5425 } 5426 5427 LegalizerHelper::LegalizeResult 5428 LegalizerHelper::narrowScalarBasic(MachineInstr &MI, unsigned TypeIdx, 5429 LLT NarrowTy) { 5430 Register DstReg = MI.getOperand(0).getReg(); 5431 LLT DstTy = MRI.getType(DstReg); 5432 5433 assert(MI.getNumOperands() == 3 && TypeIdx == 0); 5434 5435 SmallVector<Register, 4> DstRegs, DstLeftoverRegs; 5436 SmallVector<Register, 4> Src0Regs, Src0LeftoverRegs; 5437 SmallVector<Register, 4> Src1Regs, Src1LeftoverRegs; 5438 LLT LeftoverTy; 5439 if (!extractParts(MI.getOperand(1).getReg(), DstTy, NarrowTy, LeftoverTy, 5440 Src0Regs, Src0LeftoverRegs)) 5441 return UnableToLegalize; 5442 5443 LLT Unused; 5444 if (!extractParts(MI.getOperand(2).getReg(), DstTy, NarrowTy, Unused, 5445 Src1Regs, Src1LeftoverRegs)) 5446 llvm_unreachable("inconsistent extractParts result"); 5447 5448 for (unsigned I = 0, E = Src1Regs.size(); I != E; ++I) { 5449 auto Inst = MIRBuilder.buildInstr(MI.getOpcode(), {NarrowTy}, 5450 {Src0Regs[I], Src1Regs[I]}); 5451 DstRegs.push_back(Inst.getReg(0)); 5452 } 5453 5454 for (unsigned I = 0, E = Src1LeftoverRegs.size(); I != E; ++I) { 5455 auto Inst = MIRBuilder.buildInstr( 5456 MI.getOpcode(), 5457 {LeftoverTy}, {Src0LeftoverRegs[I], Src1LeftoverRegs[I]}); 5458 DstLeftoverRegs.push_back(Inst.getReg(0)); 5459 } 5460 5461 insertParts(DstReg, DstTy, NarrowTy, DstRegs, 5462 LeftoverTy, DstLeftoverRegs); 5463 5464 MI.eraseFromParent(); 5465 return Legalized; 5466 } 5467 5468 LegalizerHelper::LegalizeResult 5469 LegalizerHelper::narrowScalarExt(MachineInstr &MI, unsigned TypeIdx, 5470 LLT NarrowTy) { 5471 if (TypeIdx != 0) 5472 return UnableToLegalize; 5473 5474 Register DstReg = MI.getOperand(0).getReg(); 5475 Register SrcReg = MI.getOperand(1).getReg(); 5476 5477 LLT DstTy = MRI.getType(DstReg); 5478 if (DstTy.isVector()) 5479 return UnableToLegalize; 5480 5481 SmallVector<Register, 8> Parts; 5482 LLT GCDTy = extractGCDType(Parts, DstTy, NarrowTy, SrcReg); 5483 LLT LCMTy = buildLCMMergePieces(DstTy, NarrowTy, GCDTy, Parts, MI.getOpcode()); 5484 buildWidenedRemergeToDst(DstReg, LCMTy, Parts); 5485 5486 MI.eraseFromParent(); 5487 return Legalized; 5488 } 5489 5490 LegalizerHelper::LegalizeResult 5491 LegalizerHelper::narrowScalarSelect(MachineInstr &MI, unsigned TypeIdx, 5492 LLT NarrowTy) { 5493 if (TypeIdx != 0) 5494 return UnableToLegalize; 5495 5496 Register CondReg = MI.getOperand(1).getReg(); 5497 LLT CondTy = MRI.getType(CondReg); 5498 if (CondTy.isVector()) // TODO: Handle vselect 5499 return UnableToLegalize; 5500 5501 Register DstReg = MI.getOperand(0).getReg(); 5502 LLT DstTy = MRI.getType(DstReg); 5503 5504 SmallVector<Register, 4> DstRegs, DstLeftoverRegs; 5505 SmallVector<Register, 4> Src1Regs, Src1LeftoverRegs; 5506 SmallVector<Register, 4> Src2Regs, Src2LeftoverRegs; 5507 LLT LeftoverTy; 5508 if (!extractParts(MI.getOperand(2).getReg(), DstTy, NarrowTy, LeftoverTy, 5509 Src1Regs, Src1LeftoverRegs)) 5510 return UnableToLegalize; 5511 5512 LLT Unused; 5513 if (!extractParts(MI.getOperand(3).getReg(), DstTy, NarrowTy, Unused, 5514 Src2Regs, Src2LeftoverRegs)) 5515 llvm_unreachable("inconsistent extractParts result"); 5516 5517 for (unsigned I = 0, E = Src1Regs.size(); I != E; ++I) { 5518 auto Select = MIRBuilder.buildSelect(NarrowTy, 5519 CondReg, Src1Regs[I], Src2Regs[I]); 5520 DstRegs.push_back(Select.getReg(0)); 5521 } 5522 5523 for (unsigned I = 0, E = Src1LeftoverRegs.size(); I != E; ++I) { 5524 auto Select = MIRBuilder.buildSelect( 5525 LeftoverTy, CondReg, Src1LeftoverRegs[I], Src2LeftoverRegs[I]); 5526 DstLeftoverRegs.push_back(Select.getReg(0)); 5527 } 5528 5529 insertParts(DstReg, DstTy, NarrowTy, DstRegs, 5530 LeftoverTy, DstLeftoverRegs); 5531 5532 MI.eraseFromParent(); 5533 return Legalized; 5534 } 5535 5536 LegalizerHelper::LegalizeResult 5537 LegalizerHelper::narrowScalarCTLZ(MachineInstr &MI, unsigned TypeIdx, 5538 LLT NarrowTy) { 5539 if (TypeIdx != 1) 5540 return UnableToLegalize; 5541 5542 Register DstReg = MI.getOperand(0).getReg(); 5543 Register SrcReg = MI.getOperand(1).getReg(); 5544 LLT DstTy = MRI.getType(DstReg); 5545 LLT SrcTy = MRI.getType(SrcReg); 5546 unsigned NarrowSize = NarrowTy.getSizeInBits(); 5547 5548 if (SrcTy.isScalar() && SrcTy.getSizeInBits() == 2 * NarrowSize) { 5549 const bool IsUndef = MI.getOpcode() == TargetOpcode::G_CTLZ_ZERO_UNDEF; 5550 5551 MachineIRBuilder &B = MIRBuilder; 5552 auto UnmergeSrc = B.buildUnmerge(NarrowTy, SrcReg); 5553 // ctlz(Hi:Lo) -> Hi == 0 ? (NarrowSize + ctlz(Lo)) : ctlz(Hi) 5554 auto C_0 = B.buildConstant(NarrowTy, 0); 5555 auto HiIsZero = B.buildICmp(CmpInst::ICMP_EQ, LLT::scalar(1), 5556 UnmergeSrc.getReg(1), C_0); 5557 auto LoCTLZ = IsUndef ? 5558 B.buildCTLZ_ZERO_UNDEF(DstTy, UnmergeSrc.getReg(0)) : 5559 B.buildCTLZ(DstTy, UnmergeSrc.getReg(0)); 5560 auto C_NarrowSize = B.buildConstant(DstTy, NarrowSize); 5561 auto HiIsZeroCTLZ = B.buildAdd(DstTy, LoCTLZ, C_NarrowSize); 5562 auto HiCTLZ = B.buildCTLZ_ZERO_UNDEF(DstTy, UnmergeSrc.getReg(1)); 5563 B.buildSelect(DstReg, HiIsZero, HiIsZeroCTLZ, HiCTLZ); 5564 5565 MI.eraseFromParent(); 5566 return Legalized; 5567 } 5568 5569 return UnableToLegalize; 5570 } 5571 5572 LegalizerHelper::LegalizeResult 5573 LegalizerHelper::narrowScalarCTTZ(MachineInstr &MI, unsigned TypeIdx, 5574 LLT NarrowTy) { 5575 if (TypeIdx != 1) 5576 return UnableToLegalize; 5577 5578 Register DstReg = MI.getOperand(0).getReg(); 5579 Register SrcReg = MI.getOperand(1).getReg(); 5580 LLT DstTy = MRI.getType(DstReg); 5581 LLT SrcTy = MRI.getType(SrcReg); 5582 unsigned NarrowSize = NarrowTy.getSizeInBits(); 5583 5584 if (SrcTy.isScalar() && SrcTy.getSizeInBits() == 2 * NarrowSize) { 5585 const bool IsUndef = MI.getOpcode() == TargetOpcode::G_CTTZ_ZERO_UNDEF; 5586 5587 MachineIRBuilder &B = MIRBuilder; 5588 auto UnmergeSrc = B.buildUnmerge(NarrowTy, SrcReg); 5589 // cttz(Hi:Lo) -> Lo == 0 ? (cttz(Hi) + NarrowSize) : cttz(Lo) 5590 auto C_0 = B.buildConstant(NarrowTy, 0); 5591 auto LoIsZero = B.buildICmp(CmpInst::ICMP_EQ, LLT::scalar(1), 5592 UnmergeSrc.getReg(0), C_0); 5593 auto HiCTTZ = IsUndef ? 5594 B.buildCTTZ_ZERO_UNDEF(DstTy, UnmergeSrc.getReg(1)) : 5595 B.buildCTTZ(DstTy, UnmergeSrc.getReg(1)); 5596 auto C_NarrowSize = B.buildConstant(DstTy, NarrowSize); 5597 auto LoIsZeroCTTZ = B.buildAdd(DstTy, HiCTTZ, C_NarrowSize); 5598 auto LoCTTZ = B.buildCTTZ_ZERO_UNDEF(DstTy, UnmergeSrc.getReg(0)); 5599 B.buildSelect(DstReg, LoIsZero, LoIsZeroCTTZ, LoCTTZ); 5600 5601 MI.eraseFromParent(); 5602 return Legalized; 5603 } 5604 5605 return UnableToLegalize; 5606 } 5607 5608 LegalizerHelper::LegalizeResult 5609 LegalizerHelper::narrowScalarCTPOP(MachineInstr &MI, unsigned TypeIdx, 5610 LLT NarrowTy) { 5611 if (TypeIdx != 1) 5612 return UnableToLegalize; 5613 5614 Register DstReg = MI.getOperand(0).getReg(); 5615 LLT DstTy = MRI.getType(DstReg); 5616 LLT SrcTy = MRI.getType(MI.getOperand(1).getReg()); 5617 unsigned NarrowSize = NarrowTy.getSizeInBits(); 5618 5619 if (SrcTy.isScalar() && SrcTy.getSizeInBits() == 2 * NarrowSize) { 5620 auto UnmergeSrc = MIRBuilder.buildUnmerge(NarrowTy, MI.getOperand(1)); 5621 5622 auto LoCTPOP = MIRBuilder.buildCTPOP(DstTy, UnmergeSrc.getReg(0)); 5623 auto HiCTPOP = MIRBuilder.buildCTPOP(DstTy, UnmergeSrc.getReg(1)); 5624 MIRBuilder.buildAdd(DstReg, HiCTPOP, LoCTPOP); 5625 5626 MI.eraseFromParent(); 5627 return Legalized; 5628 } 5629 5630 return UnableToLegalize; 5631 } 5632 5633 LegalizerHelper::LegalizeResult 5634 LegalizerHelper::lowerBitCount(MachineInstr &MI) { 5635 unsigned Opc = MI.getOpcode(); 5636 const auto &TII = MIRBuilder.getTII(); 5637 auto isSupported = [this](const LegalityQuery &Q) { 5638 auto QAction = LI.getAction(Q).Action; 5639 return QAction == Legal || QAction == Libcall || QAction == Custom; 5640 }; 5641 switch (Opc) { 5642 default: 5643 return UnableToLegalize; 5644 case TargetOpcode::G_CTLZ_ZERO_UNDEF: { 5645 // This trivially expands to CTLZ. 5646 Observer.changingInstr(MI); 5647 MI.setDesc(TII.get(TargetOpcode::G_CTLZ)); 5648 Observer.changedInstr(MI); 5649 return Legalized; 5650 } 5651 case TargetOpcode::G_CTLZ: { 5652 Register DstReg = MI.getOperand(0).getReg(); 5653 Register SrcReg = MI.getOperand(1).getReg(); 5654 LLT DstTy = MRI.getType(DstReg); 5655 LLT SrcTy = MRI.getType(SrcReg); 5656 unsigned Len = SrcTy.getSizeInBits(); 5657 5658 if (isSupported({TargetOpcode::G_CTLZ_ZERO_UNDEF, {DstTy, SrcTy}})) { 5659 // If CTLZ_ZERO_UNDEF is supported, emit that and a select for zero. 5660 auto CtlzZU = MIRBuilder.buildCTLZ_ZERO_UNDEF(DstTy, SrcReg); 5661 auto ZeroSrc = MIRBuilder.buildConstant(SrcTy, 0); 5662 auto ICmp = MIRBuilder.buildICmp( 5663 CmpInst::ICMP_EQ, SrcTy.changeElementSize(1), SrcReg, ZeroSrc); 5664 auto LenConst = MIRBuilder.buildConstant(DstTy, Len); 5665 MIRBuilder.buildSelect(DstReg, ICmp, LenConst, CtlzZU); 5666 MI.eraseFromParent(); 5667 return Legalized; 5668 } 5669 // for now, we do this: 5670 // NewLen = NextPowerOf2(Len); 5671 // x = x | (x >> 1); 5672 // x = x | (x >> 2); 5673 // ... 5674 // x = x | (x >>16); 5675 // x = x | (x >>32); // for 64-bit input 5676 // Upto NewLen/2 5677 // return Len - popcount(x); 5678 // 5679 // Ref: "Hacker's Delight" by Henry Warren 5680 Register Op = SrcReg; 5681 unsigned NewLen = PowerOf2Ceil(Len); 5682 for (unsigned i = 0; (1U << i) <= (NewLen / 2); ++i) { 5683 auto MIBShiftAmt = MIRBuilder.buildConstant(SrcTy, 1ULL << i); 5684 auto MIBOp = MIRBuilder.buildOr( 5685 SrcTy, Op, MIRBuilder.buildLShr(SrcTy, Op, MIBShiftAmt)); 5686 Op = MIBOp.getReg(0); 5687 } 5688 auto MIBPop = MIRBuilder.buildCTPOP(DstTy, Op); 5689 MIRBuilder.buildSub(MI.getOperand(0), MIRBuilder.buildConstant(DstTy, Len), 5690 MIBPop); 5691 MI.eraseFromParent(); 5692 return Legalized; 5693 } 5694 case TargetOpcode::G_CTTZ_ZERO_UNDEF: { 5695 // This trivially expands to CTTZ. 5696 Observer.changingInstr(MI); 5697 MI.setDesc(TII.get(TargetOpcode::G_CTTZ)); 5698 Observer.changedInstr(MI); 5699 return Legalized; 5700 } 5701 case TargetOpcode::G_CTTZ: { 5702 Register DstReg = MI.getOperand(0).getReg(); 5703 Register SrcReg = MI.getOperand(1).getReg(); 5704 LLT DstTy = MRI.getType(DstReg); 5705 LLT SrcTy = MRI.getType(SrcReg); 5706 5707 unsigned Len = SrcTy.getSizeInBits(); 5708 if (isSupported({TargetOpcode::G_CTTZ_ZERO_UNDEF, {DstTy, SrcTy}})) { 5709 // If CTTZ_ZERO_UNDEF is legal or custom, emit that and a select with 5710 // zero. 5711 auto CttzZU = MIRBuilder.buildCTTZ_ZERO_UNDEF(DstTy, SrcReg); 5712 auto Zero = MIRBuilder.buildConstant(SrcTy, 0); 5713 auto ICmp = MIRBuilder.buildICmp( 5714 CmpInst::ICMP_EQ, DstTy.changeElementSize(1), SrcReg, Zero); 5715 auto LenConst = MIRBuilder.buildConstant(DstTy, Len); 5716 MIRBuilder.buildSelect(DstReg, ICmp, LenConst, CttzZU); 5717 MI.eraseFromParent(); 5718 return Legalized; 5719 } 5720 // for now, we use: { return popcount(~x & (x - 1)); } 5721 // unless the target has ctlz but not ctpop, in which case we use: 5722 // { return 32 - nlz(~x & (x-1)); } 5723 // Ref: "Hacker's Delight" by Henry Warren 5724 auto MIBCstNeg1 = MIRBuilder.buildConstant(SrcTy, -1); 5725 auto MIBNot = MIRBuilder.buildXor(SrcTy, SrcReg, MIBCstNeg1); 5726 auto MIBTmp = MIRBuilder.buildAnd( 5727 SrcTy, MIBNot, MIRBuilder.buildAdd(SrcTy, SrcReg, MIBCstNeg1)); 5728 if (!isSupported({TargetOpcode::G_CTPOP, {SrcTy, SrcTy}}) && 5729 isSupported({TargetOpcode::G_CTLZ, {SrcTy, SrcTy}})) { 5730 auto MIBCstLen = MIRBuilder.buildConstant(SrcTy, Len); 5731 MIRBuilder.buildSub(MI.getOperand(0), MIBCstLen, 5732 MIRBuilder.buildCTLZ(SrcTy, MIBTmp)); 5733 MI.eraseFromParent(); 5734 return Legalized; 5735 } 5736 MI.setDesc(TII.get(TargetOpcode::G_CTPOP)); 5737 MI.getOperand(1).setReg(MIBTmp.getReg(0)); 5738 return Legalized; 5739 } 5740 case TargetOpcode::G_CTPOP: { 5741 Register SrcReg = MI.getOperand(1).getReg(); 5742 LLT Ty = MRI.getType(SrcReg); 5743 unsigned Size = Ty.getSizeInBits(); 5744 MachineIRBuilder &B = MIRBuilder; 5745 5746 // Count set bits in blocks of 2 bits. Default approach would be 5747 // B2Count = { val & 0x55555555 } + { (val >> 1) & 0x55555555 } 5748 // We use following formula instead: 5749 // B2Count = val - { (val >> 1) & 0x55555555 } 5750 // since it gives same result in blocks of 2 with one instruction less. 5751 auto C_1 = B.buildConstant(Ty, 1); 5752 auto B2Set1LoTo1Hi = B.buildLShr(Ty, SrcReg, C_1); 5753 APInt B2Mask1HiTo0 = APInt::getSplat(Size, APInt(8, 0x55)); 5754 auto C_B2Mask1HiTo0 = B.buildConstant(Ty, B2Mask1HiTo0); 5755 auto B2Count1Hi = B.buildAnd(Ty, B2Set1LoTo1Hi, C_B2Mask1HiTo0); 5756 auto B2Count = B.buildSub(Ty, SrcReg, B2Count1Hi); 5757 5758 // In order to get count in blocks of 4 add values from adjacent block of 2. 5759 // B4Count = { B2Count & 0x33333333 } + { (B2Count >> 2) & 0x33333333 } 5760 auto C_2 = B.buildConstant(Ty, 2); 5761 auto B4Set2LoTo2Hi = B.buildLShr(Ty, B2Count, C_2); 5762 APInt B4Mask2HiTo0 = APInt::getSplat(Size, APInt(8, 0x33)); 5763 auto C_B4Mask2HiTo0 = B.buildConstant(Ty, B4Mask2HiTo0); 5764 auto B4HiB2Count = B.buildAnd(Ty, B4Set2LoTo2Hi, C_B4Mask2HiTo0); 5765 auto B4LoB2Count = B.buildAnd(Ty, B2Count, C_B4Mask2HiTo0); 5766 auto B4Count = B.buildAdd(Ty, B4HiB2Count, B4LoB2Count); 5767 5768 // For count in blocks of 8 bits we don't have to mask high 4 bits before 5769 // addition since count value sits in range {0,...,8} and 4 bits are enough 5770 // to hold such binary values. After addition high 4 bits still hold count 5771 // of set bits in high 4 bit block, set them to zero and get 8 bit result. 5772 // B8Count = { B4Count + (B4Count >> 4) } & 0x0F0F0F0F 5773 auto C_4 = B.buildConstant(Ty, 4); 5774 auto B8HiB4Count = B.buildLShr(Ty, B4Count, C_4); 5775 auto B8CountDirty4Hi = B.buildAdd(Ty, B8HiB4Count, B4Count); 5776 APInt B8Mask4HiTo0 = APInt::getSplat(Size, APInt(8, 0x0F)); 5777 auto C_B8Mask4HiTo0 = B.buildConstant(Ty, B8Mask4HiTo0); 5778 auto B8Count = B.buildAnd(Ty, B8CountDirty4Hi, C_B8Mask4HiTo0); 5779 5780 assert(Size<=128 && "Scalar size is too large for CTPOP lower algorithm"); 5781 // 8 bits can hold CTPOP result of 128 bit int or smaller. Mul with this 5782 // bitmask will set 8 msb in ResTmp to sum of all B8Counts in 8 bit blocks. 5783 auto MulMask = B.buildConstant(Ty, APInt::getSplat(Size, APInt(8, 0x01))); 5784 auto ResTmp = B.buildMul(Ty, B8Count, MulMask); 5785 5786 // Shift count result from 8 high bits to low bits. 5787 auto C_SizeM8 = B.buildConstant(Ty, Size - 8); 5788 B.buildLShr(MI.getOperand(0).getReg(), ResTmp, C_SizeM8); 5789 5790 MI.eraseFromParent(); 5791 return Legalized; 5792 } 5793 } 5794 } 5795 5796 // Check that (every element of) Reg is undef or not an exact multiple of BW. 5797 static bool isNonZeroModBitWidthOrUndef(const MachineRegisterInfo &MRI, 5798 Register Reg, unsigned BW) { 5799 return matchUnaryPredicate( 5800 MRI, Reg, 5801 [=](const Constant *C) { 5802 // Null constant here means an undef. 5803 const ConstantInt *CI = dyn_cast_or_null<ConstantInt>(C); 5804 return !CI || CI->getValue().urem(BW) != 0; 5805 }, 5806 /*AllowUndefs*/ true); 5807 } 5808 5809 LegalizerHelper::LegalizeResult 5810 LegalizerHelper::lowerFunnelShiftWithInverse(MachineInstr &MI) { 5811 Register Dst = MI.getOperand(0).getReg(); 5812 Register X = MI.getOperand(1).getReg(); 5813 Register Y = MI.getOperand(2).getReg(); 5814 Register Z = MI.getOperand(3).getReg(); 5815 LLT Ty = MRI.getType(Dst); 5816 LLT ShTy = MRI.getType(Z); 5817 5818 unsigned BW = Ty.getScalarSizeInBits(); 5819 5820 if (!isPowerOf2_32(BW)) 5821 return UnableToLegalize; 5822 5823 const bool IsFSHL = MI.getOpcode() == TargetOpcode::G_FSHL; 5824 unsigned RevOpcode = IsFSHL ? TargetOpcode::G_FSHR : TargetOpcode::G_FSHL; 5825 5826 if (isNonZeroModBitWidthOrUndef(MRI, Z, BW)) { 5827 // fshl X, Y, Z -> fshr X, Y, -Z 5828 // fshr X, Y, Z -> fshl X, Y, -Z 5829 auto Zero = MIRBuilder.buildConstant(ShTy, 0); 5830 Z = MIRBuilder.buildSub(Ty, Zero, Z).getReg(0); 5831 } else { 5832 // fshl X, Y, Z -> fshr (srl X, 1), (fshr X, Y, 1), ~Z 5833 // fshr X, Y, Z -> fshl (fshl X, Y, 1), (shl Y, 1), ~Z 5834 auto One = MIRBuilder.buildConstant(ShTy, 1); 5835 if (IsFSHL) { 5836 Y = MIRBuilder.buildInstr(RevOpcode, {Ty}, {X, Y, One}).getReg(0); 5837 X = MIRBuilder.buildLShr(Ty, X, One).getReg(0); 5838 } else { 5839 X = MIRBuilder.buildInstr(RevOpcode, {Ty}, {X, Y, One}).getReg(0); 5840 Y = MIRBuilder.buildShl(Ty, Y, One).getReg(0); 5841 } 5842 5843 Z = MIRBuilder.buildNot(ShTy, Z).getReg(0); 5844 } 5845 5846 MIRBuilder.buildInstr(RevOpcode, {Dst}, {X, Y, Z}); 5847 MI.eraseFromParent(); 5848 return Legalized; 5849 } 5850 5851 LegalizerHelper::LegalizeResult 5852 LegalizerHelper::lowerFunnelShiftAsShifts(MachineInstr &MI) { 5853 Register Dst = MI.getOperand(0).getReg(); 5854 Register X = MI.getOperand(1).getReg(); 5855 Register Y = MI.getOperand(2).getReg(); 5856 Register Z = MI.getOperand(3).getReg(); 5857 LLT Ty = MRI.getType(Dst); 5858 LLT ShTy = MRI.getType(Z); 5859 5860 const unsigned BW = Ty.getScalarSizeInBits(); 5861 const bool IsFSHL = MI.getOpcode() == TargetOpcode::G_FSHL; 5862 5863 Register ShX, ShY; 5864 Register ShAmt, InvShAmt; 5865 5866 // FIXME: Emit optimized urem by constant instead of letting it expand later. 5867 if (isNonZeroModBitWidthOrUndef(MRI, Z, BW)) { 5868 // fshl: X << C | Y >> (BW - C) 5869 // fshr: X << (BW - C) | Y >> C 5870 // where C = Z % BW is not zero 5871 auto BitWidthC = MIRBuilder.buildConstant(ShTy, BW); 5872 ShAmt = MIRBuilder.buildURem(ShTy, Z, BitWidthC).getReg(0); 5873 InvShAmt = MIRBuilder.buildSub(ShTy, BitWidthC, ShAmt).getReg(0); 5874 ShX = MIRBuilder.buildShl(Ty, X, IsFSHL ? ShAmt : InvShAmt).getReg(0); 5875 ShY = MIRBuilder.buildLShr(Ty, Y, IsFSHL ? InvShAmt : ShAmt).getReg(0); 5876 } else { 5877 // fshl: X << (Z % BW) | Y >> 1 >> (BW - 1 - (Z % BW)) 5878 // fshr: X << 1 << (BW - 1 - (Z % BW)) | Y >> (Z % BW) 5879 auto Mask = MIRBuilder.buildConstant(ShTy, BW - 1); 5880 if (isPowerOf2_32(BW)) { 5881 // Z % BW -> Z & (BW - 1) 5882 ShAmt = MIRBuilder.buildAnd(ShTy, Z, Mask).getReg(0); 5883 // (BW - 1) - (Z % BW) -> ~Z & (BW - 1) 5884 auto NotZ = MIRBuilder.buildNot(ShTy, Z); 5885 InvShAmt = MIRBuilder.buildAnd(ShTy, NotZ, Mask).getReg(0); 5886 } else { 5887 auto BitWidthC = MIRBuilder.buildConstant(ShTy, BW); 5888 ShAmt = MIRBuilder.buildURem(ShTy, Z, BitWidthC).getReg(0); 5889 InvShAmt = MIRBuilder.buildSub(ShTy, Mask, ShAmt).getReg(0); 5890 } 5891 5892 auto One = MIRBuilder.buildConstant(ShTy, 1); 5893 if (IsFSHL) { 5894 ShX = MIRBuilder.buildShl(Ty, X, ShAmt).getReg(0); 5895 auto ShY1 = MIRBuilder.buildLShr(Ty, Y, One); 5896 ShY = MIRBuilder.buildLShr(Ty, ShY1, InvShAmt).getReg(0); 5897 } else { 5898 auto ShX1 = MIRBuilder.buildShl(Ty, X, One); 5899 ShX = MIRBuilder.buildShl(Ty, ShX1, InvShAmt).getReg(0); 5900 ShY = MIRBuilder.buildLShr(Ty, Y, ShAmt).getReg(0); 5901 } 5902 } 5903 5904 MIRBuilder.buildOr(Dst, ShX, ShY); 5905 MI.eraseFromParent(); 5906 return Legalized; 5907 } 5908 5909 LegalizerHelper::LegalizeResult 5910 LegalizerHelper::lowerFunnelShift(MachineInstr &MI) { 5911 // These operations approximately do the following (while avoiding undefined 5912 // shifts by BW): 5913 // G_FSHL: (X << (Z % BW)) | (Y >> (BW - (Z % BW))) 5914 // G_FSHR: (X << (BW - (Z % BW))) | (Y >> (Z % BW)) 5915 Register Dst = MI.getOperand(0).getReg(); 5916 LLT Ty = MRI.getType(Dst); 5917 LLT ShTy = MRI.getType(MI.getOperand(3).getReg()); 5918 5919 bool IsFSHL = MI.getOpcode() == TargetOpcode::G_FSHL; 5920 unsigned RevOpcode = IsFSHL ? TargetOpcode::G_FSHR : TargetOpcode::G_FSHL; 5921 5922 // TODO: Use smarter heuristic that accounts for vector legalization. 5923 if (LI.getAction({RevOpcode, {Ty, ShTy}}).Action == Lower) 5924 return lowerFunnelShiftAsShifts(MI); 5925 5926 // This only works for powers of 2, fallback to shifts if it fails. 5927 LegalizerHelper::LegalizeResult Result = lowerFunnelShiftWithInverse(MI); 5928 if (Result == UnableToLegalize) 5929 return lowerFunnelShiftAsShifts(MI); 5930 return Result; 5931 } 5932 5933 LegalizerHelper::LegalizeResult 5934 LegalizerHelper::lowerRotateWithReverseRotate(MachineInstr &MI) { 5935 Register Dst = MI.getOperand(0).getReg(); 5936 Register Src = MI.getOperand(1).getReg(); 5937 Register Amt = MI.getOperand(2).getReg(); 5938 LLT AmtTy = MRI.getType(Amt); 5939 auto Zero = MIRBuilder.buildConstant(AmtTy, 0); 5940 bool IsLeft = MI.getOpcode() == TargetOpcode::G_ROTL; 5941 unsigned RevRot = IsLeft ? TargetOpcode::G_ROTR : TargetOpcode::G_ROTL; 5942 auto Neg = MIRBuilder.buildSub(AmtTy, Zero, Amt); 5943 MIRBuilder.buildInstr(RevRot, {Dst}, {Src, Neg}); 5944 MI.eraseFromParent(); 5945 return Legalized; 5946 } 5947 5948 LegalizerHelper::LegalizeResult LegalizerHelper::lowerRotate(MachineInstr &MI) { 5949 Register Dst = MI.getOperand(0).getReg(); 5950 Register Src = MI.getOperand(1).getReg(); 5951 Register Amt = MI.getOperand(2).getReg(); 5952 LLT DstTy = MRI.getType(Dst); 5953 LLT SrcTy = MRI.getType(Src); 5954 LLT AmtTy = MRI.getType(Amt); 5955 5956 unsigned EltSizeInBits = DstTy.getScalarSizeInBits(); 5957 bool IsLeft = MI.getOpcode() == TargetOpcode::G_ROTL; 5958 5959 MIRBuilder.setInstrAndDebugLoc(MI); 5960 5961 // If a rotate in the other direction is supported, use it. 5962 unsigned RevRot = IsLeft ? TargetOpcode::G_ROTR : TargetOpcode::G_ROTL; 5963 if (LI.isLegalOrCustom({RevRot, {DstTy, SrcTy}}) && 5964 isPowerOf2_32(EltSizeInBits)) 5965 return lowerRotateWithReverseRotate(MI); 5966 5967 // If a funnel shift is supported, use it. 5968 unsigned FShOpc = IsLeft ? TargetOpcode::G_FSHL : TargetOpcode::G_FSHR; 5969 unsigned RevFsh = !IsLeft ? TargetOpcode::G_FSHL : TargetOpcode::G_FSHR; 5970 bool IsFShLegal = false; 5971 if ((IsFShLegal = LI.isLegalOrCustom({FShOpc, {DstTy, AmtTy}})) || 5972 LI.isLegalOrCustom({RevFsh, {DstTy, AmtTy}})) { 5973 auto buildFunnelShift = [&](unsigned Opc, Register R1, Register R2, 5974 Register R3) { 5975 MIRBuilder.buildInstr(Opc, {R1}, {R2, R2, R3}); 5976 MI.eraseFromParent(); 5977 return Legalized; 5978 }; 5979 // If a funnel shift in the other direction is supported, use it. 5980 if (IsFShLegal) { 5981 return buildFunnelShift(FShOpc, Dst, Src, Amt); 5982 } else if (isPowerOf2_32(EltSizeInBits)) { 5983 Amt = MIRBuilder.buildNeg(DstTy, Amt).getReg(0); 5984 return buildFunnelShift(RevFsh, Dst, Src, Amt); 5985 } 5986 } 5987 5988 auto Zero = MIRBuilder.buildConstant(AmtTy, 0); 5989 unsigned ShOpc = IsLeft ? TargetOpcode::G_SHL : TargetOpcode::G_LSHR; 5990 unsigned RevShiftOpc = IsLeft ? TargetOpcode::G_LSHR : TargetOpcode::G_SHL; 5991 auto BitWidthMinusOneC = MIRBuilder.buildConstant(AmtTy, EltSizeInBits - 1); 5992 Register ShVal; 5993 Register RevShiftVal; 5994 if (isPowerOf2_32(EltSizeInBits)) { 5995 // (rotl x, c) -> x << (c & (w - 1)) | x >> (-c & (w - 1)) 5996 // (rotr x, c) -> x >> (c & (w - 1)) | x << (-c & (w - 1)) 5997 auto NegAmt = MIRBuilder.buildSub(AmtTy, Zero, Amt); 5998 auto ShAmt = MIRBuilder.buildAnd(AmtTy, Amt, BitWidthMinusOneC); 5999 ShVal = MIRBuilder.buildInstr(ShOpc, {DstTy}, {Src, ShAmt}).getReg(0); 6000 auto RevAmt = MIRBuilder.buildAnd(AmtTy, NegAmt, BitWidthMinusOneC); 6001 RevShiftVal = 6002 MIRBuilder.buildInstr(RevShiftOpc, {DstTy}, {Src, RevAmt}).getReg(0); 6003 } else { 6004 // (rotl x, c) -> x << (c % w) | x >> 1 >> (w - 1 - (c % w)) 6005 // (rotr x, c) -> x >> (c % w) | x << 1 << (w - 1 - (c % w)) 6006 auto BitWidthC = MIRBuilder.buildConstant(AmtTy, EltSizeInBits); 6007 auto ShAmt = MIRBuilder.buildURem(AmtTy, Amt, BitWidthC); 6008 ShVal = MIRBuilder.buildInstr(ShOpc, {DstTy}, {Src, ShAmt}).getReg(0); 6009 auto RevAmt = MIRBuilder.buildSub(AmtTy, BitWidthMinusOneC, ShAmt); 6010 auto One = MIRBuilder.buildConstant(AmtTy, 1); 6011 auto Inner = MIRBuilder.buildInstr(RevShiftOpc, {DstTy}, {Src, One}); 6012 RevShiftVal = 6013 MIRBuilder.buildInstr(RevShiftOpc, {DstTy}, {Inner, RevAmt}).getReg(0); 6014 } 6015 MIRBuilder.buildOr(Dst, ShVal, RevShiftVal); 6016 MI.eraseFromParent(); 6017 return Legalized; 6018 } 6019 6020 // Expand s32 = G_UITOFP s64 using bit operations to an IEEE float 6021 // representation. 6022 LegalizerHelper::LegalizeResult 6023 LegalizerHelper::lowerU64ToF32BitOps(MachineInstr &MI) { 6024 Register Dst = MI.getOperand(0).getReg(); 6025 Register Src = MI.getOperand(1).getReg(); 6026 const LLT S64 = LLT::scalar(64); 6027 const LLT S32 = LLT::scalar(32); 6028 const LLT S1 = LLT::scalar(1); 6029 6030 assert(MRI.getType(Src) == S64 && MRI.getType(Dst) == S32); 6031 6032 // unsigned cul2f(ulong u) { 6033 // uint lz = clz(u); 6034 // uint e = (u != 0) ? 127U + 63U - lz : 0; 6035 // u = (u << lz) & 0x7fffffffffffffffUL; 6036 // ulong t = u & 0xffffffffffUL; 6037 // uint v = (e << 23) | (uint)(u >> 40); 6038 // uint r = t > 0x8000000000UL ? 1U : (t == 0x8000000000UL ? v & 1U : 0U); 6039 // return as_float(v + r); 6040 // } 6041 6042 auto Zero32 = MIRBuilder.buildConstant(S32, 0); 6043 auto Zero64 = MIRBuilder.buildConstant(S64, 0); 6044 6045 auto LZ = MIRBuilder.buildCTLZ_ZERO_UNDEF(S32, Src); 6046 6047 auto K = MIRBuilder.buildConstant(S32, 127U + 63U); 6048 auto Sub = MIRBuilder.buildSub(S32, K, LZ); 6049 6050 auto NotZero = MIRBuilder.buildICmp(CmpInst::ICMP_NE, S1, Src, Zero64); 6051 auto E = MIRBuilder.buildSelect(S32, NotZero, Sub, Zero32); 6052 6053 auto Mask0 = MIRBuilder.buildConstant(S64, (-1ULL) >> 1); 6054 auto ShlLZ = MIRBuilder.buildShl(S64, Src, LZ); 6055 6056 auto U = MIRBuilder.buildAnd(S64, ShlLZ, Mask0); 6057 6058 auto Mask1 = MIRBuilder.buildConstant(S64, 0xffffffffffULL); 6059 auto T = MIRBuilder.buildAnd(S64, U, Mask1); 6060 6061 auto UShl = MIRBuilder.buildLShr(S64, U, MIRBuilder.buildConstant(S64, 40)); 6062 auto ShlE = MIRBuilder.buildShl(S32, E, MIRBuilder.buildConstant(S32, 23)); 6063 auto V = MIRBuilder.buildOr(S32, ShlE, MIRBuilder.buildTrunc(S32, UShl)); 6064 6065 auto C = MIRBuilder.buildConstant(S64, 0x8000000000ULL); 6066 auto RCmp = MIRBuilder.buildICmp(CmpInst::ICMP_UGT, S1, T, C); 6067 auto TCmp = MIRBuilder.buildICmp(CmpInst::ICMP_EQ, S1, T, C); 6068 auto One = MIRBuilder.buildConstant(S32, 1); 6069 6070 auto VTrunc1 = MIRBuilder.buildAnd(S32, V, One); 6071 auto Select0 = MIRBuilder.buildSelect(S32, TCmp, VTrunc1, Zero32); 6072 auto R = MIRBuilder.buildSelect(S32, RCmp, One, Select0); 6073 MIRBuilder.buildAdd(Dst, V, R); 6074 6075 MI.eraseFromParent(); 6076 return Legalized; 6077 } 6078 6079 LegalizerHelper::LegalizeResult LegalizerHelper::lowerUITOFP(MachineInstr &MI) { 6080 Register Dst = MI.getOperand(0).getReg(); 6081 Register Src = MI.getOperand(1).getReg(); 6082 LLT DstTy = MRI.getType(Dst); 6083 LLT SrcTy = MRI.getType(Src); 6084 6085 if (SrcTy == LLT::scalar(1)) { 6086 auto True = MIRBuilder.buildFConstant(DstTy, 1.0); 6087 auto False = MIRBuilder.buildFConstant(DstTy, 0.0); 6088 MIRBuilder.buildSelect(Dst, Src, True, False); 6089 MI.eraseFromParent(); 6090 return Legalized; 6091 } 6092 6093 if (SrcTy != LLT::scalar(64)) 6094 return UnableToLegalize; 6095 6096 if (DstTy == LLT::scalar(32)) { 6097 // TODO: SelectionDAG has several alternative expansions to port which may 6098 // be more reasonble depending on the available instructions. If a target 6099 // has sitofp, does not have CTLZ, or can efficiently use f64 as an 6100 // intermediate type, this is probably worse. 6101 return lowerU64ToF32BitOps(MI); 6102 } 6103 6104 return UnableToLegalize; 6105 } 6106 6107 LegalizerHelper::LegalizeResult LegalizerHelper::lowerSITOFP(MachineInstr &MI) { 6108 Register Dst = MI.getOperand(0).getReg(); 6109 Register Src = MI.getOperand(1).getReg(); 6110 LLT DstTy = MRI.getType(Dst); 6111 LLT SrcTy = MRI.getType(Src); 6112 6113 const LLT S64 = LLT::scalar(64); 6114 const LLT S32 = LLT::scalar(32); 6115 const LLT S1 = LLT::scalar(1); 6116 6117 if (SrcTy == S1) { 6118 auto True = MIRBuilder.buildFConstant(DstTy, -1.0); 6119 auto False = MIRBuilder.buildFConstant(DstTy, 0.0); 6120 MIRBuilder.buildSelect(Dst, Src, True, False); 6121 MI.eraseFromParent(); 6122 return Legalized; 6123 } 6124 6125 if (SrcTy != S64) 6126 return UnableToLegalize; 6127 6128 if (DstTy == S32) { 6129 // signed cl2f(long l) { 6130 // long s = l >> 63; 6131 // float r = cul2f((l + s) ^ s); 6132 // return s ? -r : r; 6133 // } 6134 Register L = Src; 6135 auto SignBit = MIRBuilder.buildConstant(S64, 63); 6136 auto S = MIRBuilder.buildAShr(S64, L, SignBit); 6137 6138 auto LPlusS = MIRBuilder.buildAdd(S64, L, S); 6139 auto Xor = MIRBuilder.buildXor(S64, LPlusS, S); 6140 auto R = MIRBuilder.buildUITOFP(S32, Xor); 6141 6142 auto RNeg = MIRBuilder.buildFNeg(S32, R); 6143 auto SignNotZero = MIRBuilder.buildICmp(CmpInst::ICMP_NE, S1, S, 6144 MIRBuilder.buildConstant(S64, 0)); 6145 MIRBuilder.buildSelect(Dst, SignNotZero, RNeg, R); 6146 MI.eraseFromParent(); 6147 return Legalized; 6148 } 6149 6150 return UnableToLegalize; 6151 } 6152 6153 LegalizerHelper::LegalizeResult LegalizerHelper::lowerFPTOUI(MachineInstr &MI) { 6154 Register Dst = MI.getOperand(0).getReg(); 6155 Register Src = MI.getOperand(1).getReg(); 6156 LLT DstTy = MRI.getType(Dst); 6157 LLT SrcTy = MRI.getType(Src); 6158 const LLT S64 = LLT::scalar(64); 6159 const LLT S32 = LLT::scalar(32); 6160 6161 if (SrcTy != S64 && SrcTy != S32) 6162 return UnableToLegalize; 6163 if (DstTy != S32 && DstTy != S64) 6164 return UnableToLegalize; 6165 6166 // FPTOSI gives same result as FPTOUI for positive signed integers. 6167 // FPTOUI needs to deal with fp values that convert to unsigned integers 6168 // greater or equal to 2^31 for float or 2^63 for double. For brevity 2^Exp. 6169 6170 APInt TwoPExpInt = APInt::getSignMask(DstTy.getSizeInBits()); 6171 APFloat TwoPExpFP(SrcTy.getSizeInBits() == 32 ? APFloat::IEEEsingle() 6172 : APFloat::IEEEdouble(), 6173 APInt::getZero(SrcTy.getSizeInBits())); 6174 TwoPExpFP.convertFromAPInt(TwoPExpInt, false, APFloat::rmNearestTiesToEven); 6175 6176 MachineInstrBuilder FPTOSI = MIRBuilder.buildFPTOSI(DstTy, Src); 6177 6178 MachineInstrBuilder Threshold = MIRBuilder.buildFConstant(SrcTy, TwoPExpFP); 6179 // For fp Value greater or equal to Threshold(2^Exp), we use FPTOSI on 6180 // (Value - 2^Exp) and add 2^Exp by setting highest bit in result to 1. 6181 MachineInstrBuilder FSub = MIRBuilder.buildFSub(SrcTy, Src, Threshold); 6182 MachineInstrBuilder ResLowBits = MIRBuilder.buildFPTOSI(DstTy, FSub); 6183 MachineInstrBuilder ResHighBit = MIRBuilder.buildConstant(DstTy, TwoPExpInt); 6184 MachineInstrBuilder Res = MIRBuilder.buildXor(DstTy, ResLowBits, ResHighBit); 6185 6186 const LLT S1 = LLT::scalar(1); 6187 6188 MachineInstrBuilder FCMP = 6189 MIRBuilder.buildFCmp(CmpInst::FCMP_ULT, S1, Src, Threshold); 6190 MIRBuilder.buildSelect(Dst, FCMP, FPTOSI, Res); 6191 6192 MI.eraseFromParent(); 6193 return Legalized; 6194 } 6195 6196 LegalizerHelper::LegalizeResult LegalizerHelper::lowerFPTOSI(MachineInstr &MI) { 6197 Register Dst = MI.getOperand(0).getReg(); 6198 Register Src = MI.getOperand(1).getReg(); 6199 LLT DstTy = MRI.getType(Dst); 6200 LLT SrcTy = MRI.getType(Src); 6201 const LLT S64 = LLT::scalar(64); 6202 const LLT S32 = LLT::scalar(32); 6203 6204 // FIXME: Only f32 to i64 conversions are supported. 6205 if (SrcTy.getScalarType() != S32 || DstTy.getScalarType() != S64) 6206 return UnableToLegalize; 6207 6208 // Expand f32 -> i64 conversion 6209 // This algorithm comes from compiler-rt's implementation of fixsfdi: 6210 // https://github.com/llvm/llvm-project/blob/main/compiler-rt/lib/builtins/fixsfdi.c 6211 6212 unsigned SrcEltBits = SrcTy.getScalarSizeInBits(); 6213 6214 auto ExponentMask = MIRBuilder.buildConstant(SrcTy, 0x7F800000); 6215 auto ExponentLoBit = MIRBuilder.buildConstant(SrcTy, 23); 6216 6217 auto AndExpMask = MIRBuilder.buildAnd(SrcTy, Src, ExponentMask); 6218 auto ExponentBits = MIRBuilder.buildLShr(SrcTy, AndExpMask, ExponentLoBit); 6219 6220 auto SignMask = MIRBuilder.buildConstant(SrcTy, 6221 APInt::getSignMask(SrcEltBits)); 6222 auto AndSignMask = MIRBuilder.buildAnd(SrcTy, Src, SignMask); 6223 auto SignLowBit = MIRBuilder.buildConstant(SrcTy, SrcEltBits - 1); 6224 auto Sign = MIRBuilder.buildAShr(SrcTy, AndSignMask, SignLowBit); 6225 Sign = MIRBuilder.buildSExt(DstTy, Sign); 6226 6227 auto MantissaMask = MIRBuilder.buildConstant(SrcTy, 0x007FFFFF); 6228 auto AndMantissaMask = MIRBuilder.buildAnd(SrcTy, Src, MantissaMask); 6229 auto K = MIRBuilder.buildConstant(SrcTy, 0x00800000); 6230 6231 auto R = MIRBuilder.buildOr(SrcTy, AndMantissaMask, K); 6232 R = MIRBuilder.buildZExt(DstTy, R); 6233 6234 auto Bias = MIRBuilder.buildConstant(SrcTy, 127); 6235 auto Exponent = MIRBuilder.buildSub(SrcTy, ExponentBits, Bias); 6236 auto SubExponent = MIRBuilder.buildSub(SrcTy, Exponent, ExponentLoBit); 6237 auto ExponentSub = MIRBuilder.buildSub(SrcTy, ExponentLoBit, Exponent); 6238 6239 auto Shl = MIRBuilder.buildShl(DstTy, R, SubExponent); 6240 auto Srl = MIRBuilder.buildLShr(DstTy, R, ExponentSub); 6241 6242 const LLT S1 = LLT::scalar(1); 6243 auto CmpGt = MIRBuilder.buildICmp(CmpInst::ICMP_SGT, 6244 S1, Exponent, ExponentLoBit); 6245 6246 R = MIRBuilder.buildSelect(DstTy, CmpGt, Shl, Srl); 6247 6248 auto XorSign = MIRBuilder.buildXor(DstTy, R, Sign); 6249 auto Ret = MIRBuilder.buildSub(DstTy, XorSign, Sign); 6250 6251 auto ZeroSrcTy = MIRBuilder.buildConstant(SrcTy, 0); 6252 6253 auto ExponentLt0 = MIRBuilder.buildICmp(CmpInst::ICMP_SLT, 6254 S1, Exponent, ZeroSrcTy); 6255 6256 auto ZeroDstTy = MIRBuilder.buildConstant(DstTy, 0); 6257 MIRBuilder.buildSelect(Dst, ExponentLt0, ZeroDstTy, Ret); 6258 6259 MI.eraseFromParent(); 6260 return Legalized; 6261 } 6262 6263 // f64 -> f16 conversion using round-to-nearest-even rounding mode. 6264 LegalizerHelper::LegalizeResult 6265 LegalizerHelper::lowerFPTRUNC_F64_TO_F16(MachineInstr &MI) { 6266 Register Dst = MI.getOperand(0).getReg(); 6267 Register Src = MI.getOperand(1).getReg(); 6268 6269 if (MRI.getType(Src).isVector()) // TODO: Handle vectors directly. 6270 return UnableToLegalize; 6271 6272 const unsigned ExpMask = 0x7ff; 6273 const unsigned ExpBiasf64 = 1023; 6274 const unsigned ExpBiasf16 = 15; 6275 const LLT S32 = LLT::scalar(32); 6276 const LLT S1 = LLT::scalar(1); 6277 6278 auto Unmerge = MIRBuilder.buildUnmerge(S32, Src); 6279 Register U = Unmerge.getReg(0); 6280 Register UH = Unmerge.getReg(1); 6281 6282 auto E = MIRBuilder.buildLShr(S32, UH, MIRBuilder.buildConstant(S32, 20)); 6283 E = MIRBuilder.buildAnd(S32, E, MIRBuilder.buildConstant(S32, ExpMask)); 6284 6285 // Subtract the fp64 exponent bias (1023) to get the real exponent and 6286 // add the f16 bias (15) to get the biased exponent for the f16 format. 6287 E = MIRBuilder.buildAdd( 6288 S32, E, MIRBuilder.buildConstant(S32, -ExpBiasf64 + ExpBiasf16)); 6289 6290 auto M = MIRBuilder.buildLShr(S32, UH, MIRBuilder.buildConstant(S32, 8)); 6291 M = MIRBuilder.buildAnd(S32, M, MIRBuilder.buildConstant(S32, 0xffe)); 6292 6293 auto MaskedSig = MIRBuilder.buildAnd(S32, UH, 6294 MIRBuilder.buildConstant(S32, 0x1ff)); 6295 MaskedSig = MIRBuilder.buildOr(S32, MaskedSig, U); 6296 6297 auto Zero = MIRBuilder.buildConstant(S32, 0); 6298 auto SigCmpNE0 = MIRBuilder.buildICmp(CmpInst::ICMP_NE, S1, MaskedSig, Zero); 6299 auto Lo40Set = MIRBuilder.buildZExt(S32, SigCmpNE0); 6300 M = MIRBuilder.buildOr(S32, M, Lo40Set); 6301 6302 // (M != 0 ? 0x0200 : 0) | 0x7c00; 6303 auto Bits0x200 = MIRBuilder.buildConstant(S32, 0x0200); 6304 auto CmpM_NE0 = MIRBuilder.buildICmp(CmpInst::ICMP_NE, S1, M, Zero); 6305 auto SelectCC = MIRBuilder.buildSelect(S32, CmpM_NE0, Bits0x200, Zero); 6306 6307 auto Bits0x7c00 = MIRBuilder.buildConstant(S32, 0x7c00); 6308 auto I = MIRBuilder.buildOr(S32, SelectCC, Bits0x7c00); 6309 6310 // N = M | (E << 12); 6311 auto EShl12 = MIRBuilder.buildShl(S32, E, MIRBuilder.buildConstant(S32, 12)); 6312 auto N = MIRBuilder.buildOr(S32, M, EShl12); 6313 6314 // B = clamp(1-E, 0, 13); 6315 auto One = MIRBuilder.buildConstant(S32, 1); 6316 auto OneSubExp = MIRBuilder.buildSub(S32, One, E); 6317 auto B = MIRBuilder.buildSMax(S32, OneSubExp, Zero); 6318 B = MIRBuilder.buildSMin(S32, B, MIRBuilder.buildConstant(S32, 13)); 6319 6320 auto SigSetHigh = MIRBuilder.buildOr(S32, M, 6321 MIRBuilder.buildConstant(S32, 0x1000)); 6322 6323 auto D = MIRBuilder.buildLShr(S32, SigSetHigh, B); 6324 auto D0 = MIRBuilder.buildShl(S32, D, B); 6325 6326 auto D0_NE_SigSetHigh = MIRBuilder.buildICmp(CmpInst::ICMP_NE, S1, 6327 D0, SigSetHigh); 6328 auto D1 = MIRBuilder.buildZExt(S32, D0_NE_SigSetHigh); 6329 D = MIRBuilder.buildOr(S32, D, D1); 6330 6331 auto CmpELtOne = MIRBuilder.buildICmp(CmpInst::ICMP_SLT, S1, E, One); 6332 auto V = MIRBuilder.buildSelect(S32, CmpELtOne, D, N); 6333 6334 auto VLow3 = MIRBuilder.buildAnd(S32, V, MIRBuilder.buildConstant(S32, 7)); 6335 V = MIRBuilder.buildLShr(S32, V, MIRBuilder.buildConstant(S32, 2)); 6336 6337 auto VLow3Eq3 = MIRBuilder.buildICmp(CmpInst::ICMP_EQ, S1, VLow3, 6338 MIRBuilder.buildConstant(S32, 3)); 6339 auto V0 = MIRBuilder.buildZExt(S32, VLow3Eq3); 6340 6341 auto VLow3Gt5 = MIRBuilder.buildICmp(CmpInst::ICMP_SGT, S1, VLow3, 6342 MIRBuilder.buildConstant(S32, 5)); 6343 auto V1 = MIRBuilder.buildZExt(S32, VLow3Gt5); 6344 6345 V1 = MIRBuilder.buildOr(S32, V0, V1); 6346 V = MIRBuilder.buildAdd(S32, V, V1); 6347 6348 auto CmpEGt30 = MIRBuilder.buildICmp(CmpInst::ICMP_SGT, S1, 6349 E, MIRBuilder.buildConstant(S32, 30)); 6350 V = MIRBuilder.buildSelect(S32, CmpEGt30, 6351 MIRBuilder.buildConstant(S32, 0x7c00), V); 6352 6353 auto CmpEGt1039 = MIRBuilder.buildICmp(CmpInst::ICMP_EQ, S1, 6354 E, MIRBuilder.buildConstant(S32, 1039)); 6355 V = MIRBuilder.buildSelect(S32, CmpEGt1039, I, V); 6356 6357 // Extract the sign bit. 6358 auto Sign = MIRBuilder.buildLShr(S32, UH, MIRBuilder.buildConstant(S32, 16)); 6359 Sign = MIRBuilder.buildAnd(S32, Sign, MIRBuilder.buildConstant(S32, 0x8000)); 6360 6361 // Insert the sign bit 6362 V = MIRBuilder.buildOr(S32, Sign, V); 6363 6364 MIRBuilder.buildTrunc(Dst, V); 6365 MI.eraseFromParent(); 6366 return Legalized; 6367 } 6368 6369 LegalizerHelper::LegalizeResult 6370 LegalizerHelper::lowerFPTRUNC(MachineInstr &MI) { 6371 Register Dst = MI.getOperand(0).getReg(); 6372 Register Src = MI.getOperand(1).getReg(); 6373 6374 LLT DstTy = MRI.getType(Dst); 6375 LLT SrcTy = MRI.getType(Src); 6376 const LLT S64 = LLT::scalar(64); 6377 const LLT S16 = LLT::scalar(16); 6378 6379 if (DstTy.getScalarType() == S16 && SrcTy.getScalarType() == S64) 6380 return lowerFPTRUNC_F64_TO_F16(MI); 6381 6382 return UnableToLegalize; 6383 } 6384 6385 // TODO: If RHS is a constant SelectionDAGBuilder expands this into a 6386 // multiplication tree. 6387 LegalizerHelper::LegalizeResult LegalizerHelper::lowerFPOWI(MachineInstr &MI) { 6388 Register Dst = MI.getOperand(0).getReg(); 6389 Register Src0 = MI.getOperand(1).getReg(); 6390 Register Src1 = MI.getOperand(2).getReg(); 6391 LLT Ty = MRI.getType(Dst); 6392 6393 auto CvtSrc1 = MIRBuilder.buildSITOFP(Ty, Src1); 6394 MIRBuilder.buildFPow(Dst, Src0, CvtSrc1, MI.getFlags()); 6395 MI.eraseFromParent(); 6396 return Legalized; 6397 } 6398 6399 static CmpInst::Predicate minMaxToCompare(unsigned Opc) { 6400 switch (Opc) { 6401 case TargetOpcode::G_SMIN: 6402 return CmpInst::ICMP_SLT; 6403 case TargetOpcode::G_SMAX: 6404 return CmpInst::ICMP_SGT; 6405 case TargetOpcode::G_UMIN: 6406 return CmpInst::ICMP_ULT; 6407 case TargetOpcode::G_UMAX: 6408 return CmpInst::ICMP_UGT; 6409 default: 6410 llvm_unreachable("not in integer min/max"); 6411 } 6412 } 6413 6414 LegalizerHelper::LegalizeResult LegalizerHelper::lowerMinMax(MachineInstr &MI) { 6415 Register Dst = MI.getOperand(0).getReg(); 6416 Register Src0 = MI.getOperand(1).getReg(); 6417 Register Src1 = MI.getOperand(2).getReg(); 6418 6419 const CmpInst::Predicate Pred = minMaxToCompare(MI.getOpcode()); 6420 LLT CmpType = MRI.getType(Dst).changeElementSize(1); 6421 6422 auto Cmp = MIRBuilder.buildICmp(Pred, CmpType, Src0, Src1); 6423 MIRBuilder.buildSelect(Dst, Cmp, Src0, Src1); 6424 6425 MI.eraseFromParent(); 6426 return Legalized; 6427 } 6428 6429 LegalizerHelper::LegalizeResult 6430 LegalizerHelper::lowerFCopySign(MachineInstr &MI) { 6431 Register Dst = MI.getOperand(0).getReg(); 6432 Register Src0 = MI.getOperand(1).getReg(); 6433 Register Src1 = MI.getOperand(2).getReg(); 6434 6435 const LLT Src0Ty = MRI.getType(Src0); 6436 const LLT Src1Ty = MRI.getType(Src1); 6437 6438 const int Src0Size = Src0Ty.getScalarSizeInBits(); 6439 const int Src1Size = Src1Ty.getScalarSizeInBits(); 6440 6441 auto SignBitMask = MIRBuilder.buildConstant( 6442 Src0Ty, APInt::getSignMask(Src0Size)); 6443 6444 auto NotSignBitMask = MIRBuilder.buildConstant( 6445 Src0Ty, APInt::getLowBitsSet(Src0Size, Src0Size - 1)); 6446 6447 Register And0 = MIRBuilder.buildAnd(Src0Ty, Src0, NotSignBitMask).getReg(0); 6448 Register And1; 6449 if (Src0Ty == Src1Ty) { 6450 And1 = MIRBuilder.buildAnd(Src1Ty, Src1, SignBitMask).getReg(0); 6451 } else if (Src0Size > Src1Size) { 6452 auto ShiftAmt = MIRBuilder.buildConstant(Src0Ty, Src0Size - Src1Size); 6453 auto Zext = MIRBuilder.buildZExt(Src0Ty, Src1); 6454 auto Shift = MIRBuilder.buildShl(Src0Ty, Zext, ShiftAmt); 6455 And1 = MIRBuilder.buildAnd(Src0Ty, Shift, SignBitMask).getReg(0); 6456 } else { 6457 auto ShiftAmt = MIRBuilder.buildConstant(Src1Ty, Src1Size - Src0Size); 6458 auto Shift = MIRBuilder.buildLShr(Src1Ty, Src1, ShiftAmt); 6459 auto Trunc = MIRBuilder.buildTrunc(Src0Ty, Shift); 6460 And1 = MIRBuilder.buildAnd(Src0Ty, Trunc, SignBitMask).getReg(0); 6461 } 6462 6463 // Be careful about setting nsz/nnan/ninf on every instruction, since the 6464 // constants are a nan and -0.0, but the final result should preserve 6465 // everything. 6466 unsigned Flags = MI.getFlags(); 6467 MIRBuilder.buildOr(Dst, And0, And1, Flags); 6468 6469 MI.eraseFromParent(); 6470 return Legalized; 6471 } 6472 6473 LegalizerHelper::LegalizeResult 6474 LegalizerHelper::lowerFMinNumMaxNum(MachineInstr &MI) { 6475 unsigned NewOp = MI.getOpcode() == TargetOpcode::G_FMINNUM ? 6476 TargetOpcode::G_FMINNUM_IEEE : TargetOpcode::G_FMAXNUM_IEEE; 6477 6478 Register Dst = MI.getOperand(0).getReg(); 6479 Register Src0 = MI.getOperand(1).getReg(); 6480 Register Src1 = MI.getOperand(2).getReg(); 6481 LLT Ty = MRI.getType(Dst); 6482 6483 if (!MI.getFlag(MachineInstr::FmNoNans)) { 6484 // Insert canonicalizes if it's possible we need to quiet to get correct 6485 // sNaN behavior. 6486 6487 // Note this must be done here, and not as an optimization combine in the 6488 // absence of a dedicate quiet-snan instruction as we're using an 6489 // omni-purpose G_FCANONICALIZE. 6490 if (!isKnownNeverSNaN(Src0, MRI)) 6491 Src0 = MIRBuilder.buildFCanonicalize(Ty, Src0, MI.getFlags()).getReg(0); 6492 6493 if (!isKnownNeverSNaN(Src1, MRI)) 6494 Src1 = MIRBuilder.buildFCanonicalize(Ty, Src1, MI.getFlags()).getReg(0); 6495 } 6496 6497 // If there are no nans, it's safe to simply replace this with the non-IEEE 6498 // version. 6499 MIRBuilder.buildInstr(NewOp, {Dst}, {Src0, Src1}, MI.getFlags()); 6500 MI.eraseFromParent(); 6501 return Legalized; 6502 } 6503 6504 LegalizerHelper::LegalizeResult LegalizerHelper::lowerFMad(MachineInstr &MI) { 6505 // Expand G_FMAD a, b, c -> G_FADD (G_FMUL a, b), c 6506 Register DstReg = MI.getOperand(0).getReg(); 6507 LLT Ty = MRI.getType(DstReg); 6508 unsigned Flags = MI.getFlags(); 6509 6510 auto Mul = MIRBuilder.buildFMul(Ty, MI.getOperand(1), MI.getOperand(2), 6511 Flags); 6512 MIRBuilder.buildFAdd(DstReg, Mul, MI.getOperand(3), Flags); 6513 MI.eraseFromParent(); 6514 return Legalized; 6515 } 6516 6517 LegalizerHelper::LegalizeResult 6518 LegalizerHelper::lowerIntrinsicRound(MachineInstr &MI) { 6519 Register DstReg = MI.getOperand(0).getReg(); 6520 Register X = MI.getOperand(1).getReg(); 6521 const unsigned Flags = MI.getFlags(); 6522 const LLT Ty = MRI.getType(DstReg); 6523 const LLT CondTy = Ty.changeElementSize(1); 6524 6525 // round(x) => 6526 // t = trunc(x); 6527 // d = fabs(x - t); 6528 // o = copysign(1.0f, x); 6529 // return t + (d >= 0.5 ? o : 0.0); 6530 6531 auto T = MIRBuilder.buildIntrinsicTrunc(Ty, X, Flags); 6532 6533 auto Diff = MIRBuilder.buildFSub(Ty, X, T, Flags); 6534 auto AbsDiff = MIRBuilder.buildFAbs(Ty, Diff, Flags); 6535 auto Zero = MIRBuilder.buildFConstant(Ty, 0.0); 6536 auto One = MIRBuilder.buildFConstant(Ty, 1.0); 6537 auto Half = MIRBuilder.buildFConstant(Ty, 0.5); 6538 auto SignOne = MIRBuilder.buildFCopysign(Ty, One, X); 6539 6540 auto Cmp = MIRBuilder.buildFCmp(CmpInst::FCMP_OGE, CondTy, AbsDiff, Half, 6541 Flags); 6542 auto Sel = MIRBuilder.buildSelect(Ty, Cmp, SignOne, Zero, Flags); 6543 6544 MIRBuilder.buildFAdd(DstReg, T, Sel, Flags); 6545 6546 MI.eraseFromParent(); 6547 return Legalized; 6548 } 6549 6550 LegalizerHelper::LegalizeResult 6551 LegalizerHelper::lowerFFloor(MachineInstr &MI) { 6552 Register DstReg = MI.getOperand(0).getReg(); 6553 Register SrcReg = MI.getOperand(1).getReg(); 6554 unsigned Flags = MI.getFlags(); 6555 LLT Ty = MRI.getType(DstReg); 6556 const LLT CondTy = Ty.changeElementSize(1); 6557 6558 // result = trunc(src); 6559 // if (src < 0.0 && src != result) 6560 // result += -1.0. 6561 6562 auto Trunc = MIRBuilder.buildIntrinsicTrunc(Ty, SrcReg, Flags); 6563 auto Zero = MIRBuilder.buildFConstant(Ty, 0.0); 6564 6565 auto Lt0 = MIRBuilder.buildFCmp(CmpInst::FCMP_OLT, CondTy, 6566 SrcReg, Zero, Flags); 6567 auto NeTrunc = MIRBuilder.buildFCmp(CmpInst::FCMP_ONE, CondTy, 6568 SrcReg, Trunc, Flags); 6569 auto And = MIRBuilder.buildAnd(CondTy, Lt0, NeTrunc); 6570 auto AddVal = MIRBuilder.buildSITOFP(Ty, And); 6571 6572 MIRBuilder.buildFAdd(DstReg, Trunc, AddVal, Flags); 6573 MI.eraseFromParent(); 6574 return Legalized; 6575 } 6576 6577 LegalizerHelper::LegalizeResult 6578 LegalizerHelper::lowerMergeValues(MachineInstr &MI) { 6579 const unsigned NumOps = MI.getNumOperands(); 6580 Register DstReg = MI.getOperand(0).getReg(); 6581 Register Src0Reg = MI.getOperand(1).getReg(); 6582 LLT DstTy = MRI.getType(DstReg); 6583 LLT SrcTy = MRI.getType(Src0Reg); 6584 unsigned PartSize = SrcTy.getSizeInBits(); 6585 6586 LLT WideTy = LLT::scalar(DstTy.getSizeInBits()); 6587 Register ResultReg = MIRBuilder.buildZExt(WideTy, Src0Reg).getReg(0); 6588 6589 for (unsigned I = 2; I != NumOps; ++I) { 6590 const unsigned Offset = (I - 1) * PartSize; 6591 6592 Register SrcReg = MI.getOperand(I).getReg(); 6593 auto ZextInput = MIRBuilder.buildZExt(WideTy, SrcReg); 6594 6595 Register NextResult = I + 1 == NumOps && WideTy == DstTy ? DstReg : 6596 MRI.createGenericVirtualRegister(WideTy); 6597 6598 auto ShiftAmt = MIRBuilder.buildConstant(WideTy, Offset); 6599 auto Shl = MIRBuilder.buildShl(WideTy, ZextInput, ShiftAmt); 6600 MIRBuilder.buildOr(NextResult, ResultReg, Shl); 6601 ResultReg = NextResult; 6602 } 6603 6604 if (DstTy.isPointer()) { 6605 if (MIRBuilder.getDataLayout().isNonIntegralAddressSpace( 6606 DstTy.getAddressSpace())) { 6607 LLVM_DEBUG(dbgs() << "Not casting nonintegral address space\n"); 6608 return UnableToLegalize; 6609 } 6610 6611 MIRBuilder.buildIntToPtr(DstReg, ResultReg); 6612 } 6613 6614 MI.eraseFromParent(); 6615 return Legalized; 6616 } 6617 6618 LegalizerHelper::LegalizeResult 6619 LegalizerHelper::lowerUnmergeValues(MachineInstr &MI) { 6620 const unsigned NumDst = MI.getNumOperands() - 1; 6621 Register SrcReg = MI.getOperand(NumDst).getReg(); 6622 Register Dst0Reg = MI.getOperand(0).getReg(); 6623 LLT DstTy = MRI.getType(Dst0Reg); 6624 if (DstTy.isPointer()) 6625 return UnableToLegalize; // TODO 6626 6627 SrcReg = coerceToScalar(SrcReg); 6628 if (!SrcReg) 6629 return UnableToLegalize; 6630 6631 // Expand scalarizing unmerge as bitcast to integer and shift. 6632 LLT IntTy = MRI.getType(SrcReg); 6633 6634 MIRBuilder.buildTrunc(Dst0Reg, SrcReg); 6635 6636 const unsigned DstSize = DstTy.getSizeInBits(); 6637 unsigned Offset = DstSize; 6638 for (unsigned I = 1; I != NumDst; ++I, Offset += DstSize) { 6639 auto ShiftAmt = MIRBuilder.buildConstant(IntTy, Offset); 6640 auto Shift = MIRBuilder.buildLShr(IntTy, SrcReg, ShiftAmt); 6641 MIRBuilder.buildTrunc(MI.getOperand(I), Shift); 6642 } 6643 6644 MI.eraseFromParent(); 6645 return Legalized; 6646 } 6647 6648 /// Lower a vector extract or insert by writing the vector to a stack temporary 6649 /// and reloading the element or vector. 6650 /// 6651 /// %dst = G_EXTRACT_VECTOR_ELT %vec, %idx 6652 /// => 6653 /// %stack_temp = G_FRAME_INDEX 6654 /// G_STORE %vec, %stack_temp 6655 /// %idx = clamp(%idx, %vec.getNumElements()) 6656 /// %element_ptr = G_PTR_ADD %stack_temp, %idx 6657 /// %dst = G_LOAD %element_ptr 6658 LegalizerHelper::LegalizeResult 6659 LegalizerHelper::lowerExtractInsertVectorElt(MachineInstr &MI) { 6660 Register DstReg = MI.getOperand(0).getReg(); 6661 Register SrcVec = MI.getOperand(1).getReg(); 6662 Register InsertVal; 6663 if (MI.getOpcode() == TargetOpcode::G_INSERT_VECTOR_ELT) 6664 InsertVal = MI.getOperand(2).getReg(); 6665 6666 Register Idx = MI.getOperand(MI.getNumOperands() - 1).getReg(); 6667 6668 LLT VecTy = MRI.getType(SrcVec); 6669 LLT EltTy = VecTy.getElementType(); 6670 unsigned NumElts = VecTy.getNumElements(); 6671 6672 int64_t IdxVal; 6673 if (mi_match(Idx, MRI, m_ICst(IdxVal)) && IdxVal <= NumElts) { 6674 SmallVector<Register, 8> SrcRegs; 6675 extractParts(SrcVec, EltTy, NumElts, SrcRegs); 6676 6677 if (InsertVal) { 6678 SrcRegs[IdxVal] = MI.getOperand(2).getReg(); 6679 MIRBuilder.buildMergeLikeInstr(DstReg, SrcRegs); 6680 } else { 6681 MIRBuilder.buildCopy(DstReg, SrcRegs[IdxVal]); 6682 } 6683 6684 MI.eraseFromParent(); 6685 return Legalized; 6686 } 6687 6688 if (!EltTy.isByteSized()) { // Not implemented. 6689 LLVM_DEBUG(dbgs() << "Can't handle non-byte element vectors yet\n"); 6690 return UnableToLegalize; 6691 } 6692 6693 unsigned EltBytes = EltTy.getSizeInBytes(); 6694 Align VecAlign = getStackTemporaryAlignment(VecTy); 6695 Align EltAlign; 6696 6697 MachinePointerInfo PtrInfo; 6698 auto StackTemp = createStackTemporary(TypeSize::Fixed(VecTy.getSizeInBytes()), 6699 VecAlign, PtrInfo); 6700 MIRBuilder.buildStore(SrcVec, StackTemp, PtrInfo, VecAlign); 6701 6702 // Get the pointer to the element, and be sure not to hit undefined behavior 6703 // if the index is out of bounds. 6704 Register EltPtr = getVectorElementPointer(StackTemp.getReg(0), VecTy, Idx); 6705 6706 if (mi_match(Idx, MRI, m_ICst(IdxVal))) { 6707 int64_t Offset = IdxVal * EltBytes; 6708 PtrInfo = PtrInfo.getWithOffset(Offset); 6709 EltAlign = commonAlignment(VecAlign, Offset); 6710 } else { 6711 // We lose information with a variable offset. 6712 EltAlign = getStackTemporaryAlignment(EltTy); 6713 PtrInfo = MachinePointerInfo(MRI.getType(EltPtr).getAddressSpace()); 6714 } 6715 6716 if (InsertVal) { 6717 // Write the inserted element 6718 MIRBuilder.buildStore(InsertVal, EltPtr, PtrInfo, EltAlign); 6719 6720 // Reload the whole vector. 6721 MIRBuilder.buildLoad(DstReg, StackTemp, PtrInfo, VecAlign); 6722 } else { 6723 MIRBuilder.buildLoad(DstReg, EltPtr, PtrInfo, EltAlign); 6724 } 6725 6726 MI.eraseFromParent(); 6727 return Legalized; 6728 } 6729 6730 LegalizerHelper::LegalizeResult 6731 LegalizerHelper::lowerShuffleVector(MachineInstr &MI) { 6732 Register DstReg = MI.getOperand(0).getReg(); 6733 Register Src0Reg = MI.getOperand(1).getReg(); 6734 Register Src1Reg = MI.getOperand(2).getReg(); 6735 LLT Src0Ty = MRI.getType(Src0Reg); 6736 LLT DstTy = MRI.getType(DstReg); 6737 LLT IdxTy = LLT::scalar(32); 6738 6739 ArrayRef<int> Mask = MI.getOperand(3).getShuffleMask(); 6740 6741 if (DstTy.isScalar()) { 6742 if (Src0Ty.isVector()) 6743 return UnableToLegalize; 6744 6745 // This is just a SELECT. 6746 assert(Mask.size() == 1 && "Expected a single mask element"); 6747 Register Val; 6748 if (Mask[0] < 0 || Mask[0] > 1) 6749 Val = MIRBuilder.buildUndef(DstTy).getReg(0); 6750 else 6751 Val = Mask[0] == 0 ? Src0Reg : Src1Reg; 6752 MIRBuilder.buildCopy(DstReg, Val); 6753 MI.eraseFromParent(); 6754 return Legalized; 6755 } 6756 6757 Register Undef; 6758 SmallVector<Register, 32> BuildVec; 6759 LLT EltTy = DstTy.getElementType(); 6760 6761 for (int Idx : Mask) { 6762 if (Idx < 0) { 6763 if (!Undef.isValid()) 6764 Undef = MIRBuilder.buildUndef(EltTy).getReg(0); 6765 BuildVec.push_back(Undef); 6766 continue; 6767 } 6768 6769 if (Src0Ty.isScalar()) { 6770 BuildVec.push_back(Idx == 0 ? Src0Reg : Src1Reg); 6771 } else { 6772 int NumElts = Src0Ty.getNumElements(); 6773 Register SrcVec = Idx < NumElts ? Src0Reg : Src1Reg; 6774 int ExtractIdx = Idx < NumElts ? Idx : Idx - NumElts; 6775 auto IdxK = MIRBuilder.buildConstant(IdxTy, ExtractIdx); 6776 auto Extract = MIRBuilder.buildExtractVectorElement(EltTy, SrcVec, IdxK); 6777 BuildVec.push_back(Extract.getReg(0)); 6778 } 6779 } 6780 6781 MIRBuilder.buildBuildVector(DstReg, BuildVec); 6782 MI.eraseFromParent(); 6783 return Legalized; 6784 } 6785 6786 LegalizerHelper::LegalizeResult 6787 LegalizerHelper::lowerDynStackAlloc(MachineInstr &MI) { 6788 const auto &MF = *MI.getMF(); 6789 const auto &TFI = *MF.getSubtarget().getFrameLowering(); 6790 if (TFI.getStackGrowthDirection() == TargetFrameLowering::StackGrowsUp) 6791 return UnableToLegalize; 6792 6793 Register Dst = MI.getOperand(0).getReg(); 6794 Register AllocSize = MI.getOperand(1).getReg(); 6795 Align Alignment = assumeAligned(MI.getOperand(2).getImm()); 6796 6797 LLT PtrTy = MRI.getType(Dst); 6798 LLT IntPtrTy = LLT::scalar(PtrTy.getSizeInBits()); 6799 6800 Register SPReg = TLI.getStackPointerRegisterToSaveRestore(); 6801 auto SPTmp = MIRBuilder.buildCopy(PtrTy, SPReg); 6802 SPTmp = MIRBuilder.buildCast(IntPtrTy, SPTmp); 6803 6804 // Subtract the final alloc from the SP. We use G_PTRTOINT here so we don't 6805 // have to generate an extra instruction to negate the alloc and then use 6806 // G_PTR_ADD to add the negative offset. 6807 auto Alloc = MIRBuilder.buildSub(IntPtrTy, SPTmp, AllocSize); 6808 if (Alignment > Align(1)) { 6809 APInt AlignMask(IntPtrTy.getSizeInBits(), Alignment.value(), true); 6810 AlignMask.negate(); 6811 auto AlignCst = MIRBuilder.buildConstant(IntPtrTy, AlignMask); 6812 Alloc = MIRBuilder.buildAnd(IntPtrTy, Alloc, AlignCst); 6813 } 6814 6815 SPTmp = MIRBuilder.buildCast(PtrTy, Alloc); 6816 MIRBuilder.buildCopy(SPReg, SPTmp); 6817 MIRBuilder.buildCopy(Dst, SPTmp); 6818 6819 MI.eraseFromParent(); 6820 return Legalized; 6821 } 6822 6823 LegalizerHelper::LegalizeResult 6824 LegalizerHelper::lowerExtract(MachineInstr &MI) { 6825 Register Dst = MI.getOperand(0).getReg(); 6826 Register Src = MI.getOperand(1).getReg(); 6827 unsigned Offset = MI.getOperand(2).getImm(); 6828 6829 LLT DstTy = MRI.getType(Dst); 6830 LLT SrcTy = MRI.getType(Src); 6831 6832 // Extract sub-vector or one element 6833 if (SrcTy.isVector()) { 6834 unsigned SrcEltSize = SrcTy.getElementType().getSizeInBits(); 6835 unsigned DstSize = DstTy.getSizeInBits(); 6836 6837 if ((Offset % SrcEltSize == 0) && (DstSize % SrcEltSize == 0) && 6838 (Offset + DstSize <= SrcTy.getSizeInBits())) { 6839 // Unmerge and allow access to each Src element for the artifact combiner. 6840 auto Unmerge = MIRBuilder.buildUnmerge(SrcTy.getElementType(), Src); 6841 6842 // Take element(s) we need to extract and copy it (merge them). 6843 SmallVector<Register, 8> SubVectorElts; 6844 for (unsigned Idx = Offset / SrcEltSize; 6845 Idx < (Offset + DstSize) / SrcEltSize; ++Idx) { 6846 SubVectorElts.push_back(Unmerge.getReg(Idx)); 6847 } 6848 if (SubVectorElts.size() == 1) 6849 MIRBuilder.buildCopy(Dst, SubVectorElts[0]); 6850 else 6851 MIRBuilder.buildMergeLikeInstr(Dst, SubVectorElts); 6852 6853 MI.eraseFromParent(); 6854 return Legalized; 6855 } 6856 } 6857 6858 if (DstTy.isScalar() && 6859 (SrcTy.isScalar() || 6860 (SrcTy.isVector() && DstTy == SrcTy.getElementType()))) { 6861 LLT SrcIntTy = SrcTy; 6862 if (!SrcTy.isScalar()) { 6863 SrcIntTy = LLT::scalar(SrcTy.getSizeInBits()); 6864 Src = MIRBuilder.buildBitcast(SrcIntTy, Src).getReg(0); 6865 } 6866 6867 if (Offset == 0) 6868 MIRBuilder.buildTrunc(Dst, Src); 6869 else { 6870 auto ShiftAmt = MIRBuilder.buildConstant(SrcIntTy, Offset); 6871 auto Shr = MIRBuilder.buildLShr(SrcIntTy, Src, ShiftAmt); 6872 MIRBuilder.buildTrunc(Dst, Shr); 6873 } 6874 6875 MI.eraseFromParent(); 6876 return Legalized; 6877 } 6878 6879 return UnableToLegalize; 6880 } 6881 6882 LegalizerHelper::LegalizeResult LegalizerHelper::lowerInsert(MachineInstr &MI) { 6883 Register Dst = MI.getOperand(0).getReg(); 6884 Register Src = MI.getOperand(1).getReg(); 6885 Register InsertSrc = MI.getOperand(2).getReg(); 6886 uint64_t Offset = MI.getOperand(3).getImm(); 6887 6888 LLT DstTy = MRI.getType(Src); 6889 LLT InsertTy = MRI.getType(InsertSrc); 6890 6891 // Insert sub-vector or one element 6892 if (DstTy.isVector() && !InsertTy.isPointer()) { 6893 LLT EltTy = DstTy.getElementType(); 6894 unsigned EltSize = EltTy.getSizeInBits(); 6895 unsigned InsertSize = InsertTy.getSizeInBits(); 6896 6897 if ((Offset % EltSize == 0) && (InsertSize % EltSize == 0) && 6898 (Offset + InsertSize <= DstTy.getSizeInBits())) { 6899 auto UnmergeSrc = MIRBuilder.buildUnmerge(EltTy, Src); 6900 SmallVector<Register, 8> DstElts; 6901 unsigned Idx = 0; 6902 // Elements from Src before insert start Offset 6903 for (; Idx < Offset / EltSize; ++Idx) { 6904 DstElts.push_back(UnmergeSrc.getReg(Idx)); 6905 } 6906 6907 // Replace elements in Src with elements from InsertSrc 6908 if (InsertTy.getSizeInBits() > EltSize) { 6909 auto UnmergeInsertSrc = MIRBuilder.buildUnmerge(EltTy, InsertSrc); 6910 for (unsigned i = 0; Idx < (Offset + InsertSize) / EltSize; 6911 ++Idx, ++i) { 6912 DstElts.push_back(UnmergeInsertSrc.getReg(i)); 6913 } 6914 } else { 6915 DstElts.push_back(InsertSrc); 6916 ++Idx; 6917 } 6918 6919 // Remaining elements from Src after insert 6920 for (; Idx < DstTy.getNumElements(); ++Idx) { 6921 DstElts.push_back(UnmergeSrc.getReg(Idx)); 6922 } 6923 6924 MIRBuilder.buildMergeLikeInstr(Dst, DstElts); 6925 MI.eraseFromParent(); 6926 return Legalized; 6927 } 6928 } 6929 6930 if (InsertTy.isVector() || 6931 (DstTy.isVector() && DstTy.getElementType() != InsertTy)) 6932 return UnableToLegalize; 6933 6934 const DataLayout &DL = MIRBuilder.getDataLayout(); 6935 if ((DstTy.isPointer() && 6936 DL.isNonIntegralAddressSpace(DstTy.getAddressSpace())) || 6937 (InsertTy.isPointer() && 6938 DL.isNonIntegralAddressSpace(InsertTy.getAddressSpace()))) { 6939 LLVM_DEBUG(dbgs() << "Not casting non-integral address space integer\n"); 6940 return UnableToLegalize; 6941 } 6942 6943 LLT IntDstTy = DstTy; 6944 6945 if (!DstTy.isScalar()) { 6946 IntDstTy = LLT::scalar(DstTy.getSizeInBits()); 6947 Src = MIRBuilder.buildCast(IntDstTy, Src).getReg(0); 6948 } 6949 6950 if (!InsertTy.isScalar()) { 6951 const LLT IntInsertTy = LLT::scalar(InsertTy.getSizeInBits()); 6952 InsertSrc = MIRBuilder.buildPtrToInt(IntInsertTy, InsertSrc).getReg(0); 6953 } 6954 6955 Register ExtInsSrc = MIRBuilder.buildZExt(IntDstTy, InsertSrc).getReg(0); 6956 if (Offset != 0) { 6957 auto ShiftAmt = MIRBuilder.buildConstant(IntDstTy, Offset); 6958 ExtInsSrc = MIRBuilder.buildShl(IntDstTy, ExtInsSrc, ShiftAmt).getReg(0); 6959 } 6960 6961 APInt MaskVal = APInt::getBitsSetWithWrap( 6962 DstTy.getSizeInBits(), Offset + InsertTy.getSizeInBits(), Offset); 6963 6964 auto Mask = MIRBuilder.buildConstant(IntDstTy, MaskVal); 6965 auto MaskedSrc = MIRBuilder.buildAnd(IntDstTy, Src, Mask); 6966 auto Or = MIRBuilder.buildOr(IntDstTy, MaskedSrc, ExtInsSrc); 6967 6968 MIRBuilder.buildCast(Dst, Or); 6969 MI.eraseFromParent(); 6970 return Legalized; 6971 } 6972 6973 LegalizerHelper::LegalizeResult 6974 LegalizerHelper::lowerSADDO_SSUBO(MachineInstr &MI) { 6975 Register Dst0 = MI.getOperand(0).getReg(); 6976 Register Dst1 = MI.getOperand(1).getReg(); 6977 Register LHS = MI.getOperand(2).getReg(); 6978 Register RHS = MI.getOperand(3).getReg(); 6979 const bool IsAdd = MI.getOpcode() == TargetOpcode::G_SADDO; 6980 6981 LLT Ty = MRI.getType(Dst0); 6982 LLT BoolTy = MRI.getType(Dst1); 6983 6984 if (IsAdd) 6985 MIRBuilder.buildAdd(Dst0, LHS, RHS); 6986 else 6987 MIRBuilder.buildSub(Dst0, LHS, RHS); 6988 6989 // TODO: If SADDSAT/SSUBSAT is legal, compare results to detect overflow. 6990 6991 auto Zero = MIRBuilder.buildConstant(Ty, 0); 6992 6993 // For an addition, the result should be less than one of the operands (LHS) 6994 // if and only if the other operand (RHS) is negative, otherwise there will 6995 // be overflow. 6996 // For a subtraction, the result should be less than one of the operands 6997 // (LHS) if and only if the other operand (RHS) is (non-zero) positive, 6998 // otherwise there will be overflow. 6999 auto ResultLowerThanLHS = 7000 MIRBuilder.buildICmp(CmpInst::ICMP_SLT, BoolTy, Dst0, LHS); 7001 auto ConditionRHS = MIRBuilder.buildICmp( 7002 IsAdd ? CmpInst::ICMP_SLT : CmpInst::ICMP_SGT, BoolTy, RHS, Zero); 7003 7004 MIRBuilder.buildXor(Dst1, ConditionRHS, ResultLowerThanLHS); 7005 MI.eraseFromParent(); 7006 return Legalized; 7007 } 7008 7009 LegalizerHelper::LegalizeResult 7010 LegalizerHelper::lowerAddSubSatToMinMax(MachineInstr &MI) { 7011 Register Res = MI.getOperand(0).getReg(); 7012 Register LHS = MI.getOperand(1).getReg(); 7013 Register RHS = MI.getOperand(2).getReg(); 7014 LLT Ty = MRI.getType(Res); 7015 bool IsSigned; 7016 bool IsAdd; 7017 unsigned BaseOp; 7018 switch (MI.getOpcode()) { 7019 default: 7020 llvm_unreachable("unexpected addsat/subsat opcode"); 7021 case TargetOpcode::G_UADDSAT: 7022 IsSigned = false; 7023 IsAdd = true; 7024 BaseOp = TargetOpcode::G_ADD; 7025 break; 7026 case TargetOpcode::G_SADDSAT: 7027 IsSigned = true; 7028 IsAdd = true; 7029 BaseOp = TargetOpcode::G_ADD; 7030 break; 7031 case TargetOpcode::G_USUBSAT: 7032 IsSigned = false; 7033 IsAdd = false; 7034 BaseOp = TargetOpcode::G_SUB; 7035 break; 7036 case TargetOpcode::G_SSUBSAT: 7037 IsSigned = true; 7038 IsAdd = false; 7039 BaseOp = TargetOpcode::G_SUB; 7040 break; 7041 } 7042 7043 if (IsSigned) { 7044 // sadd.sat(a, b) -> 7045 // hi = 0x7fffffff - smax(a, 0) 7046 // lo = 0x80000000 - smin(a, 0) 7047 // a + smin(smax(lo, b), hi) 7048 // ssub.sat(a, b) -> 7049 // lo = smax(a, -1) - 0x7fffffff 7050 // hi = smin(a, -1) - 0x80000000 7051 // a - smin(smax(lo, b), hi) 7052 // TODO: AMDGPU can use a "median of 3" instruction here: 7053 // a +/- med3(lo, b, hi) 7054 uint64_t NumBits = Ty.getScalarSizeInBits(); 7055 auto MaxVal = 7056 MIRBuilder.buildConstant(Ty, APInt::getSignedMaxValue(NumBits)); 7057 auto MinVal = 7058 MIRBuilder.buildConstant(Ty, APInt::getSignedMinValue(NumBits)); 7059 MachineInstrBuilder Hi, Lo; 7060 if (IsAdd) { 7061 auto Zero = MIRBuilder.buildConstant(Ty, 0); 7062 Hi = MIRBuilder.buildSub(Ty, MaxVal, MIRBuilder.buildSMax(Ty, LHS, Zero)); 7063 Lo = MIRBuilder.buildSub(Ty, MinVal, MIRBuilder.buildSMin(Ty, LHS, Zero)); 7064 } else { 7065 auto NegOne = MIRBuilder.buildConstant(Ty, -1); 7066 Lo = MIRBuilder.buildSub(Ty, MIRBuilder.buildSMax(Ty, LHS, NegOne), 7067 MaxVal); 7068 Hi = MIRBuilder.buildSub(Ty, MIRBuilder.buildSMin(Ty, LHS, NegOne), 7069 MinVal); 7070 } 7071 auto RHSClamped = 7072 MIRBuilder.buildSMin(Ty, MIRBuilder.buildSMax(Ty, Lo, RHS), Hi); 7073 MIRBuilder.buildInstr(BaseOp, {Res}, {LHS, RHSClamped}); 7074 } else { 7075 // uadd.sat(a, b) -> a + umin(~a, b) 7076 // usub.sat(a, b) -> a - umin(a, b) 7077 Register Not = IsAdd ? MIRBuilder.buildNot(Ty, LHS).getReg(0) : LHS; 7078 auto Min = MIRBuilder.buildUMin(Ty, Not, RHS); 7079 MIRBuilder.buildInstr(BaseOp, {Res}, {LHS, Min}); 7080 } 7081 7082 MI.eraseFromParent(); 7083 return Legalized; 7084 } 7085 7086 LegalizerHelper::LegalizeResult 7087 LegalizerHelper::lowerAddSubSatToAddoSubo(MachineInstr &MI) { 7088 Register Res = MI.getOperand(0).getReg(); 7089 Register LHS = MI.getOperand(1).getReg(); 7090 Register RHS = MI.getOperand(2).getReg(); 7091 LLT Ty = MRI.getType(Res); 7092 LLT BoolTy = Ty.changeElementSize(1); 7093 bool IsSigned; 7094 bool IsAdd; 7095 unsigned OverflowOp; 7096 switch (MI.getOpcode()) { 7097 default: 7098 llvm_unreachable("unexpected addsat/subsat opcode"); 7099 case TargetOpcode::G_UADDSAT: 7100 IsSigned = false; 7101 IsAdd = true; 7102 OverflowOp = TargetOpcode::G_UADDO; 7103 break; 7104 case TargetOpcode::G_SADDSAT: 7105 IsSigned = true; 7106 IsAdd = true; 7107 OverflowOp = TargetOpcode::G_SADDO; 7108 break; 7109 case TargetOpcode::G_USUBSAT: 7110 IsSigned = false; 7111 IsAdd = false; 7112 OverflowOp = TargetOpcode::G_USUBO; 7113 break; 7114 case TargetOpcode::G_SSUBSAT: 7115 IsSigned = true; 7116 IsAdd = false; 7117 OverflowOp = TargetOpcode::G_SSUBO; 7118 break; 7119 } 7120 7121 auto OverflowRes = 7122 MIRBuilder.buildInstr(OverflowOp, {Ty, BoolTy}, {LHS, RHS}); 7123 Register Tmp = OverflowRes.getReg(0); 7124 Register Ov = OverflowRes.getReg(1); 7125 MachineInstrBuilder Clamp; 7126 if (IsSigned) { 7127 // sadd.sat(a, b) -> 7128 // {tmp, ov} = saddo(a, b) 7129 // ov ? (tmp >>s 31) + 0x80000000 : r 7130 // ssub.sat(a, b) -> 7131 // {tmp, ov} = ssubo(a, b) 7132 // ov ? (tmp >>s 31) + 0x80000000 : r 7133 uint64_t NumBits = Ty.getScalarSizeInBits(); 7134 auto ShiftAmount = MIRBuilder.buildConstant(Ty, NumBits - 1); 7135 auto Sign = MIRBuilder.buildAShr(Ty, Tmp, ShiftAmount); 7136 auto MinVal = 7137 MIRBuilder.buildConstant(Ty, APInt::getSignedMinValue(NumBits)); 7138 Clamp = MIRBuilder.buildAdd(Ty, Sign, MinVal); 7139 } else { 7140 // uadd.sat(a, b) -> 7141 // {tmp, ov} = uaddo(a, b) 7142 // ov ? 0xffffffff : tmp 7143 // usub.sat(a, b) -> 7144 // {tmp, ov} = usubo(a, b) 7145 // ov ? 0 : tmp 7146 Clamp = MIRBuilder.buildConstant(Ty, IsAdd ? -1 : 0); 7147 } 7148 MIRBuilder.buildSelect(Res, Ov, Clamp, Tmp); 7149 7150 MI.eraseFromParent(); 7151 return Legalized; 7152 } 7153 7154 LegalizerHelper::LegalizeResult 7155 LegalizerHelper::lowerShlSat(MachineInstr &MI) { 7156 assert((MI.getOpcode() == TargetOpcode::G_SSHLSAT || 7157 MI.getOpcode() == TargetOpcode::G_USHLSAT) && 7158 "Expected shlsat opcode!"); 7159 bool IsSigned = MI.getOpcode() == TargetOpcode::G_SSHLSAT; 7160 Register Res = MI.getOperand(0).getReg(); 7161 Register LHS = MI.getOperand(1).getReg(); 7162 Register RHS = MI.getOperand(2).getReg(); 7163 LLT Ty = MRI.getType(Res); 7164 LLT BoolTy = Ty.changeElementSize(1); 7165 7166 unsigned BW = Ty.getScalarSizeInBits(); 7167 auto Result = MIRBuilder.buildShl(Ty, LHS, RHS); 7168 auto Orig = IsSigned ? MIRBuilder.buildAShr(Ty, Result, RHS) 7169 : MIRBuilder.buildLShr(Ty, Result, RHS); 7170 7171 MachineInstrBuilder SatVal; 7172 if (IsSigned) { 7173 auto SatMin = MIRBuilder.buildConstant(Ty, APInt::getSignedMinValue(BW)); 7174 auto SatMax = MIRBuilder.buildConstant(Ty, APInt::getSignedMaxValue(BW)); 7175 auto Cmp = MIRBuilder.buildICmp(CmpInst::ICMP_SLT, BoolTy, LHS, 7176 MIRBuilder.buildConstant(Ty, 0)); 7177 SatVal = MIRBuilder.buildSelect(Ty, Cmp, SatMin, SatMax); 7178 } else { 7179 SatVal = MIRBuilder.buildConstant(Ty, APInt::getMaxValue(BW)); 7180 } 7181 auto Ov = MIRBuilder.buildICmp(CmpInst::ICMP_NE, BoolTy, LHS, Orig); 7182 MIRBuilder.buildSelect(Res, Ov, SatVal, Result); 7183 7184 MI.eraseFromParent(); 7185 return Legalized; 7186 } 7187 7188 LegalizerHelper::LegalizeResult 7189 LegalizerHelper::lowerBswap(MachineInstr &MI) { 7190 Register Dst = MI.getOperand(0).getReg(); 7191 Register Src = MI.getOperand(1).getReg(); 7192 const LLT Ty = MRI.getType(Src); 7193 unsigned SizeInBytes = (Ty.getScalarSizeInBits() + 7) / 8; 7194 unsigned BaseShiftAmt = (SizeInBytes - 1) * 8; 7195 7196 // Swap most and least significant byte, set remaining bytes in Res to zero. 7197 auto ShiftAmt = MIRBuilder.buildConstant(Ty, BaseShiftAmt); 7198 auto LSByteShiftedLeft = MIRBuilder.buildShl(Ty, Src, ShiftAmt); 7199 auto MSByteShiftedRight = MIRBuilder.buildLShr(Ty, Src, ShiftAmt); 7200 auto Res = MIRBuilder.buildOr(Ty, MSByteShiftedRight, LSByteShiftedLeft); 7201 7202 // Set i-th high/low byte in Res to i-th low/high byte from Src. 7203 for (unsigned i = 1; i < SizeInBytes / 2; ++i) { 7204 // AND with Mask leaves byte i unchanged and sets remaining bytes to 0. 7205 APInt APMask(SizeInBytes * 8, 0xFF << (i * 8)); 7206 auto Mask = MIRBuilder.buildConstant(Ty, APMask); 7207 auto ShiftAmt = MIRBuilder.buildConstant(Ty, BaseShiftAmt - 16 * i); 7208 // Low byte shifted left to place of high byte: (Src & Mask) << ShiftAmt. 7209 auto LoByte = MIRBuilder.buildAnd(Ty, Src, Mask); 7210 auto LoShiftedLeft = MIRBuilder.buildShl(Ty, LoByte, ShiftAmt); 7211 Res = MIRBuilder.buildOr(Ty, Res, LoShiftedLeft); 7212 // High byte shifted right to place of low byte: (Src >> ShiftAmt) & Mask. 7213 auto SrcShiftedRight = MIRBuilder.buildLShr(Ty, Src, ShiftAmt); 7214 auto HiShiftedRight = MIRBuilder.buildAnd(Ty, SrcShiftedRight, Mask); 7215 Res = MIRBuilder.buildOr(Ty, Res, HiShiftedRight); 7216 } 7217 Res.getInstr()->getOperand(0).setReg(Dst); 7218 7219 MI.eraseFromParent(); 7220 return Legalized; 7221 } 7222 7223 //{ (Src & Mask) >> N } | { (Src << N) & Mask } 7224 static MachineInstrBuilder SwapN(unsigned N, DstOp Dst, MachineIRBuilder &B, 7225 MachineInstrBuilder Src, APInt Mask) { 7226 const LLT Ty = Dst.getLLTTy(*B.getMRI()); 7227 MachineInstrBuilder C_N = B.buildConstant(Ty, N); 7228 MachineInstrBuilder MaskLoNTo0 = B.buildConstant(Ty, Mask); 7229 auto LHS = B.buildLShr(Ty, B.buildAnd(Ty, Src, MaskLoNTo0), C_N); 7230 auto RHS = B.buildAnd(Ty, B.buildShl(Ty, Src, C_N), MaskLoNTo0); 7231 return B.buildOr(Dst, LHS, RHS); 7232 } 7233 7234 LegalizerHelper::LegalizeResult 7235 LegalizerHelper::lowerBitreverse(MachineInstr &MI) { 7236 Register Dst = MI.getOperand(0).getReg(); 7237 Register Src = MI.getOperand(1).getReg(); 7238 const LLT Ty = MRI.getType(Src); 7239 unsigned Size = Ty.getSizeInBits(); 7240 7241 MachineInstrBuilder BSWAP = 7242 MIRBuilder.buildInstr(TargetOpcode::G_BSWAP, {Ty}, {Src}); 7243 7244 // swap high and low 4 bits in 8 bit blocks 7654|3210 -> 3210|7654 7245 // [(val & 0xF0F0F0F0) >> 4] | [(val & 0x0F0F0F0F) << 4] 7246 // -> [(val & 0xF0F0F0F0) >> 4] | [(val << 4) & 0xF0F0F0F0] 7247 MachineInstrBuilder Swap4 = 7248 SwapN(4, Ty, MIRBuilder, BSWAP, APInt::getSplat(Size, APInt(8, 0xF0))); 7249 7250 // swap high and low 2 bits in 4 bit blocks 32|10 76|54 -> 10|32 54|76 7251 // [(val & 0xCCCCCCCC) >> 2] & [(val & 0x33333333) << 2] 7252 // -> [(val & 0xCCCCCCCC) >> 2] & [(val << 2) & 0xCCCCCCCC] 7253 MachineInstrBuilder Swap2 = 7254 SwapN(2, Ty, MIRBuilder, Swap4, APInt::getSplat(Size, APInt(8, 0xCC))); 7255 7256 // 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 7257 // [(val & 0xAAAAAAAA) >> 1] & [(val & 0x55555555) << 1] 7258 // -> [(val & 0xAAAAAAAA) >> 1] & [(val << 1) & 0xAAAAAAAA] 7259 SwapN(1, Dst, MIRBuilder, Swap2, APInt::getSplat(Size, APInt(8, 0xAA))); 7260 7261 MI.eraseFromParent(); 7262 return Legalized; 7263 } 7264 7265 LegalizerHelper::LegalizeResult 7266 LegalizerHelper::lowerReadWriteRegister(MachineInstr &MI) { 7267 MachineFunction &MF = MIRBuilder.getMF(); 7268 7269 bool IsRead = MI.getOpcode() == TargetOpcode::G_READ_REGISTER; 7270 int NameOpIdx = IsRead ? 1 : 0; 7271 int ValRegIndex = IsRead ? 0 : 1; 7272 7273 Register ValReg = MI.getOperand(ValRegIndex).getReg(); 7274 const LLT Ty = MRI.getType(ValReg); 7275 const MDString *RegStr = cast<MDString>( 7276 cast<MDNode>(MI.getOperand(NameOpIdx).getMetadata())->getOperand(0)); 7277 7278 Register PhysReg = TLI.getRegisterByName(RegStr->getString().data(), Ty, MF); 7279 if (!PhysReg.isValid()) 7280 return UnableToLegalize; 7281 7282 if (IsRead) 7283 MIRBuilder.buildCopy(ValReg, PhysReg); 7284 else 7285 MIRBuilder.buildCopy(PhysReg, ValReg); 7286 7287 MI.eraseFromParent(); 7288 return Legalized; 7289 } 7290 7291 LegalizerHelper::LegalizeResult 7292 LegalizerHelper::lowerSMULH_UMULH(MachineInstr &MI) { 7293 bool IsSigned = MI.getOpcode() == TargetOpcode::G_SMULH; 7294 unsigned ExtOp = IsSigned ? TargetOpcode::G_SEXT : TargetOpcode::G_ZEXT; 7295 Register Result = MI.getOperand(0).getReg(); 7296 LLT OrigTy = MRI.getType(Result); 7297 auto SizeInBits = OrigTy.getScalarSizeInBits(); 7298 LLT WideTy = OrigTy.changeElementSize(SizeInBits * 2); 7299 7300 auto LHS = MIRBuilder.buildInstr(ExtOp, {WideTy}, {MI.getOperand(1)}); 7301 auto RHS = MIRBuilder.buildInstr(ExtOp, {WideTy}, {MI.getOperand(2)}); 7302 auto Mul = MIRBuilder.buildMul(WideTy, LHS, RHS); 7303 unsigned ShiftOp = IsSigned ? TargetOpcode::G_ASHR : TargetOpcode::G_LSHR; 7304 7305 auto ShiftAmt = MIRBuilder.buildConstant(WideTy, SizeInBits); 7306 auto Shifted = MIRBuilder.buildInstr(ShiftOp, {WideTy}, {Mul, ShiftAmt}); 7307 MIRBuilder.buildTrunc(Result, Shifted); 7308 7309 MI.eraseFromParent(); 7310 return Legalized; 7311 } 7312 7313 LegalizerHelper::LegalizeResult 7314 LegalizerHelper::lowerISFPCLASS(MachineInstr &MI) { 7315 Register DstReg = MI.getOperand(0).getReg(); 7316 Register SrcReg = MI.getOperand(1).getReg(); 7317 LLT DstTy = MRI.getType(DstReg); 7318 LLT SrcTy = MRI.getType(SrcReg); 7319 uint64_t Mask = MI.getOperand(2).getImm(); 7320 7321 if (Mask == 0) { 7322 MIRBuilder.buildConstant(DstReg, 0); 7323 MI.eraseFromParent(); 7324 return Legalized; 7325 } 7326 if ((Mask & fcAllFlags) == fcAllFlags) { 7327 MIRBuilder.buildConstant(DstReg, 1); 7328 MI.eraseFromParent(); 7329 return Legalized; 7330 } 7331 7332 unsigned BitSize = SrcTy.getScalarSizeInBits(); 7333 const fltSemantics &Semantics = getFltSemanticForLLT(SrcTy.getScalarType()); 7334 7335 LLT IntTy = LLT::scalar(BitSize); 7336 if (SrcTy.isVector()) 7337 IntTy = LLT::vector(SrcTy.getElementCount(), IntTy); 7338 auto AsInt = MIRBuilder.buildCopy(IntTy, SrcReg); 7339 7340 // Various masks. 7341 APInt SignBit = APInt::getSignMask(BitSize); 7342 APInt ValueMask = APInt::getSignedMaxValue(BitSize); // All bits but sign. 7343 APInt Inf = APFloat::getInf(Semantics).bitcastToAPInt(); // Exp and int bit. 7344 APInt ExpMask = Inf; 7345 APInt AllOneMantissa = APFloat::getLargest(Semantics).bitcastToAPInt() & ~Inf; 7346 APInt QNaNBitMask = 7347 APInt::getOneBitSet(BitSize, AllOneMantissa.getActiveBits() - 1); 7348 APInt InvertionMask = APInt::getAllOnesValue(DstTy.getScalarSizeInBits()); 7349 7350 auto SignBitC = MIRBuilder.buildConstant(IntTy, SignBit); 7351 auto ValueMaskC = MIRBuilder.buildConstant(IntTy, ValueMask); 7352 auto InfC = MIRBuilder.buildConstant(IntTy, Inf); 7353 auto ExpMaskC = MIRBuilder.buildConstant(IntTy, ExpMask); 7354 auto ZeroC = MIRBuilder.buildConstant(IntTy, 0); 7355 7356 auto Abs = MIRBuilder.buildAnd(IntTy, AsInt, ValueMaskC); 7357 auto Sign = 7358 MIRBuilder.buildICmp(CmpInst::Predicate::ICMP_NE, DstTy, AsInt, Abs); 7359 7360 auto Res = MIRBuilder.buildConstant(DstTy, 0); 7361 const auto appendToRes = [&](MachineInstrBuilder ToAppend) { 7362 Res = MIRBuilder.buildOr(DstTy, Res, ToAppend); 7363 }; 7364 7365 // Tests that involve more than one class should be processed first. 7366 if ((Mask & fcFinite) == fcFinite) { 7367 // finite(V) ==> abs(V) u< exp_mask 7368 appendToRes(MIRBuilder.buildICmp(CmpInst::Predicate::ICMP_ULT, DstTy, Abs, 7369 ExpMaskC)); 7370 Mask &= ~fcFinite; 7371 } else if ((Mask & fcFinite) == fcPosFinite) { 7372 // finite(V) && V > 0 ==> V u< exp_mask 7373 appendToRes(MIRBuilder.buildICmp(CmpInst::Predicate::ICMP_ULT, DstTy, AsInt, 7374 ExpMaskC)); 7375 Mask &= ~fcPosFinite; 7376 } else if ((Mask & fcFinite) == fcNegFinite) { 7377 // finite(V) && V < 0 ==> abs(V) u< exp_mask && signbit == 1 7378 auto Cmp = MIRBuilder.buildICmp(CmpInst::Predicate::ICMP_ULT, DstTy, Abs, 7379 ExpMaskC); 7380 auto And = MIRBuilder.buildAnd(DstTy, Cmp, Sign); 7381 appendToRes(And); 7382 Mask &= ~fcNegFinite; 7383 } 7384 7385 // Check for individual classes. 7386 if (unsigned PartialCheck = Mask & fcZero) { 7387 if (PartialCheck == fcPosZero) 7388 appendToRes(MIRBuilder.buildICmp(CmpInst::Predicate::ICMP_EQ, DstTy, 7389 AsInt, ZeroC)); 7390 else if (PartialCheck == fcZero) 7391 appendToRes( 7392 MIRBuilder.buildICmp(CmpInst::Predicate::ICMP_EQ, DstTy, Abs, ZeroC)); 7393 else // fcNegZero 7394 appendToRes(MIRBuilder.buildICmp(CmpInst::Predicate::ICMP_EQ, DstTy, 7395 AsInt, SignBitC)); 7396 } 7397 7398 if (unsigned PartialCheck = Mask & fcInf) { 7399 if (PartialCheck == fcPosInf) 7400 appendToRes(MIRBuilder.buildICmp(CmpInst::Predicate::ICMP_EQ, DstTy, 7401 AsInt, InfC)); 7402 else if (PartialCheck == fcInf) 7403 appendToRes( 7404 MIRBuilder.buildICmp(CmpInst::Predicate::ICMP_EQ, DstTy, Abs, InfC)); 7405 else { // fcNegInf 7406 APInt NegInf = APFloat::getInf(Semantics, true).bitcastToAPInt(); 7407 auto NegInfC = MIRBuilder.buildConstant(IntTy, NegInf); 7408 appendToRes(MIRBuilder.buildICmp(CmpInst::Predicate::ICMP_EQ, DstTy, 7409 AsInt, NegInfC)); 7410 } 7411 } 7412 7413 if (unsigned PartialCheck = Mask & fcNan) { 7414 auto InfWithQnanBitC = MIRBuilder.buildConstant(IntTy, Inf | QNaNBitMask); 7415 if (PartialCheck == fcNan) { 7416 // isnan(V) ==> abs(V) u> int(inf) 7417 appendToRes( 7418 MIRBuilder.buildICmp(CmpInst::Predicate::ICMP_UGT, DstTy, Abs, InfC)); 7419 } else if (PartialCheck == fcQNan) { 7420 // isquiet(V) ==> abs(V) u>= (unsigned(Inf) | quiet_bit) 7421 appendToRes(MIRBuilder.buildICmp(CmpInst::Predicate::ICMP_UGE, DstTy, Abs, 7422 InfWithQnanBitC)); 7423 } else { // fcSNan 7424 // issignaling(V) ==> abs(V) u> unsigned(Inf) && 7425 // abs(V) u< (unsigned(Inf) | quiet_bit) 7426 auto IsNan = 7427 MIRBuilder.buildICmp(CmpInst::Predicate::ICMP_UGT, DstTy, Abs, InfC); 7428 auto IsNotQnan = MIRBuilder.buildICmp(CmpInst::Predicate::ICMP_ULT, DstTy, 7429 Abs, InfWithQnanBitC); 7430 appendToRes(MIRBuilder.buildAnd(DstTy, IsNan, IsNotQnan)); 7431 } 7432 } 7433 7434 if (unsigned PartialCheck = Mask & fcSubnormal) { 7435 // issubnormal(V) ==> unsigned(abs(V) - 1) u< (all mantissa bits set) 7436 // issubnormal(V) && V>0 ==> unsigned(V - 1) u< (all mantissa bits set) 7437 auto V = (PartialCheck == fcPosSubnormal) ? AsInt : Abs; 7438 auto OneC = MIRBuilder.buildConstant(IntTy, 1); 7439 auto VMinusOne = MIRBuilder.buildSub(IntTy, V, OneC); 7440 auto SubnormalRes = 7441 MIRBuilder.buildICmp(CmpInst::Predicate::ICMP_ULT, DstTy, VMinusOne, 7442 MIRBuilder.buildConstant(IntTy, AllOneMantissa)); 7443 if (PartialCheck == fcNegSubnormal) 7444 SubnormalRes = MIRBuilder.buildAnd(DstTy, SubnormalRes, Sign); 7445 appendToRes(SubnormalRes); 7446 } 7447 7448 if (unsigned PartialCheck = Mask & fcNormal) { 7449 // isnormal(V) ==> (0 u< exp u< max_exp) ==> (unsigned(exp-1) u< 7450 // (max_exp-1)) 7451 APInt ExpLSB = ExpMask & ~(ExpMask.shl(1)); 7452 auto ExpMinusOne = MIRBuilder.buildSub( 7453 IntTy, Abs, MIRBuilder.buildConstant(IntTy, ExpLSB)); 7454 APInt MaxExpMinusOne = ExpMask - ExpLSB; 7455 auto NormalRes = 7456 MIRBuilder.buildICmp(CmpInst::Predicate::ICMP_ULT, DstTy, ExpMinusOne, 7457 MIRBuilder.buildConstant(IntTy, MaxExpMinusOne)); 7458 if (PartialCheck == fcNegNormal) 7459 NormalRes = MIRBuilder.buildAnd(DstTy, NormalRes, Sign); 7460 else if (PartialCheck == fcPosNormal) { 7461 auto PosSign = MIRBuilder.buildXor( 7462 DstTy, Sign, MIRBuilder.buildConstant(DstTy, InvertionMask)); 7463 NormalRes = MIRBuilder.buildAnd(DstTy, NormalRes, PosSign); 7464 } 7465 appendToRes(NormalRes); 7466 } 7467 7468 MIRBuilder.buildCopy(DstReg, Res); 7469 MI.eraseFromParent(); 7470 return Legalized; 7471 } 7472 7473 LegalizerHelper::LegalizeResult LegalizerHelper::lowerSelect(MachineInstr &MI) { 7474 // Implement vector G_SELECT in terms of XOR, AND, OR. 7475 Register DstReg = MI.getOperand(0).getReg(); 7476 Register MaskReg = MI.getOperand(1).getReg(); 7477 Register Op1Reg = MI.getOperand(2).getReg(); 7478 Register Op2Reg = MI.getOperand(3).getReg(); 7479 LLT DstTy = MRI.getType(DstReg); 7480 LLT MaskTy = MRI.getType(MaskReg); 7481 if (!DstTy.isVector()) 7482 return UnableToLegalize; 7483 7484 bool IsEltPtr = DstTy.getElementType().isPointer(); 7485 if (IsEltPtr) { 7486 LLT ScalarPtrTy = LLT::scalar(DstTy.getScalarSizeInBits()); 7487 LLT NewTy = DstTy.changeElementType(ScalarPtrTy); 7488 Op1Reg = MIRBuilder.buildPtrToInt(NewTy, Op1Reg).getReg(0); 7489 Op2Reg = MIRBuilder.buildPtrToInt(NewTy, Op2Reg).getReg(0); 7490 DstTy = NewTy; 7491 } 7492 7493 if (MaskTy.isScalar()) { 7494 // Turn the scalar condition into a vector condition mask. 7495 7496 Register MaskElt = MaskReg; 7497 7498 // The condition was potentially zero extended before, but we want a sign 7499 // extended boolean. 7500 if (MaskTy != LLT::scalar(1)) 7501 MaskElt = MIRBuilder.buildSExtInReg(MaskTy, MaskElt, 1).getReg(0); 7502 7503 // Continue the sign extension (or truncate) to match the data type. 7504 MaskElt = MIRBuilder.buildSExtOrTrunc(DstTy.getElementType(), 7505 MaskElt).getReg(0); 7506 7507 // Generate a vector splat idiom. 7508 auto ShufSplat = MIRBuilder.buildShuffleSplat(DstTy, MaskElt); 7509 MaskReg = ShufSplat.getReg(0); 7510 MaskTy = DstTy; 7511 } 7512 7513 if (MaskTy.getSizeInBits() != DstTy.getSizeInBits()) { 7514 return UnableToLegalize; 7515 } 7516 7517 auto NotMask = MIRBuilder.buildNot(MaskTy, MaskReg); 7518 auto NewOp1 = MIRBuilder.buildAnd(MaskTy, Op1Reg, MaskReg); 7519 auto NewOp2 = MIRBuilder.buildAnd(MaskTy, Op2Reg, NotMask); 7520 if (IsEltPtr) { 7521 auto Or = MIRBuilder.buildOr(DstTy, NewOp1, NewOp2); 7522 MIRBuilder.buildIntToPtr(DstReg, Or); 7523 } else { 7524 MIRBuilder.buildOr(DstReg, NewOp1, NewOp2); 7525 } 7526 MI.eraseFromParent(); 7527 return Legalized; 7528 } 7529 7530 LegalizerHelper::LegalizeResult LegalizerHelper::lowerDIVREM(MachineInstr &MI) { 7531 // Split DIVREM into individual instructions. 7532 unsigned Opcode = MI.getOpcode(); 7533 7534 MIRBuilder.buildInstr( 7535 Opcode == TargetOpcode::G_SDIVREM ? TargetOpcode::G_SDIV 7536 : TargetOpcode::G_UDIV, 7537 {MI.getOperand(0).getReg()}, {MI.getOperand(2), MI.getOperand(3)}); 7538 MIRBuilder.buildInstr( 7539 Opcode == TargetOpcode::G_SDIVREM ? TargetOpcode::G_SREM 7540 : TargetOpcode::G_UREM, 7541 {MI.getOperand(1).getReg()}, {MI.getOperand(2), MI.getOperand(3)}); 7542 MI.eraseFromParent(); 7543 return Legalized; 7544 } 7545 7546 LegalizerHelper::LegalizeResult 7547 LegalizerHelper::lowerAbsToAddXor(MachineInstr &MI) { 7548 // Expand %res = G_ABS %a into: 7549 // %v1 = G_ASHR %a, scalar_size-1 7550 // %v2 = G_ADD %a, %v1 7551 // %res = G_XOR %v2, %v1 7552 LLT DstTy = MRI.getType(MI.getOperand(0).getReg()); 7553 Register OpReg = MI.getOperand(1).getReg(); 7554 auto ShiftAmt = 7555 MIRBuilder.buildConstant(DstTy, DstTy.getScalarSizeInBits() - 1); 7556 auto Shift = MIRBuilder.buildAShr(DstTy, OpReg, ShiftAmt); 7557 auto Add = MIRBuilder.buildAdd(DstTy, OpReg, Shift); 7558 MIRBuilder.buildXor(MI.getOperand(0).getReg(), Add, Shift); 7559 MI.eraseFromParent(); 7560 return Legalized; 7561 } 7562 7563 LegalizerHelper::LegalizeResult 7564 LegalizerHelper::lowerAbsToMaxNeg(MachineInstr &MI) { 7565 // Expand %res = G_ABS %a into: 7566 // %v1 = G_CONSTANT 0 7567 // %v2 = G_SUB %v1, %a 7568 // %res = G_SMAX %a, %v2 7569 Register SrcReg = MI.getOperand(1).getReg(); 7570 LLT Ty = MRI.getType(SrcReg); 7571 auto Zero = MIRBuilder.buildConstant(Ty, 0).getReg(0); 7572 auto Sub = MIRBuilder.buildSub(Ty, Zero, SrcReg).getReg(0); 7573 MIRBuilder.buildSMax(MI.getOperand(0), SrcReg, Sub); 7574 MI.eraseFromParent(); 7575 return Legalized; 7576 } 7577 7578 LegalizerHelper::LegalizeResult 7579 LegalizerHelper::lowerVectorReduction(MachineInstr &MI) { 7580 Register SrcReg = MI.getOperand(1).getReg(); 7581 LLT SrcTy = MRI.getType(SrcReg); 7582 LLT DstTy = MRI.getType(SrcReg); 7583 7584 // The source could be a scalar if the IR type was <1 x sN>. 7585 if (SrcTy.isScalar()) { 7586 if (DstTy.getSizeInBits() > SrcTy.getSizeInBits()) 7587 return UnableToLegalize; // FIXME: handle extension. 7588 // This can be just a plain copy. 7589 Observer.changingInstr(MI); 7590 MI.setDesc(MIRBuilder.getTII().get(TargetOpcode::COPY)); 7591 Observer.changedInstr(MI); 7592 return Legalized; 7593 } 7594 return UnableToLegalize;; 7595 } 7596 7597 static bool shouldLowerMemFuncForSize(const MachineFunction &MF) { 7598 // On Darwin, -Os means optimize for size without hurting performance, so 7599 // only really optimize for size when -Oz (MinSize) is used. 7600 if (MF.getTarget().getTargetTriple().isOSDarwin()) 7601 return MF.getFunction().hasMinSize(); 7602 return MF.getFunction().hasOptSize(); 7603 } 7604 7605 // Returns a list of types to use for memory op lowering in MemOps. A partial 7606 // port of findOptimalMemOpLowering in TargetLowering. 7607 static bool findGISelOptimalMemOpLowering(std::vector<LLT> &MemOps, 7608 unsigned Limit, const MemOp &Op, 7609 unsigned DstAS, unsigned SrcAS, 7610 const AttributeList &FuncAttributes, 7611 const TargetLowering &TLI) { 7612 if (Op.isMemcpyWithFixedDstAlign() && Op.getSrcAlign() < Op.getDstAlign()) 7613 return false; 7614 7615 LLT Ty = TLI.getOptimalMemOpLLT(Op, FuncAttributes); 7616 7617 if (Ty == LLT()) { 7618 // Use the largest scalar type whose alignment constraints are satisfied. 7619 // We only need to check DstAlign here as SrcAlign is always greater or 7620 // equal to DstAlign (or zero). 7621 Ty = LLT::scalar(64); 7622 if (Op.isFixedDstAlign()) 7623 while (Op.getDstAlign() < Ty.getSizeInBytes() && 7624 !TLI.allowsMisalignedMemoryAccesses(Ty, DstAS, Op.getDstAlign())) 7625 Ty = LLT::scalar(Ty.getSizeInBytes()); 7626 assert(Ty.getSizeInBits() > 0 && "Could not find valid type"); 7627 // FIXME: check for the largest legal type we can load/store to. 7628 } 7629 7630 unsigned NumMemOps = 0; 7631 uint64_t Size = Op.size(); 7632 while (Size) { 7633 unsigned TySize = Ty.getSizeInBytes(); 7634 while (TySize > Size) { 7635 // For now, only use non-vector load / store's for the left-over pieces. 7636 LLT NewTy = Ty; 7637 // FIXME: check for mem op safety and legality of the types. Not all of 7638 // SDAGisms map cleanly to GISel concepts. 7639 if (NewTy.isVector()) 7640 NewTy = NewTy.getSizeInBits() > 64 ? LLT::scalar(64) : LLT::scalar(32); 7641 NewTy = LLT::scalar(PowerOf2Floor(NewTy.getSizeInBits() - 1)); 7642 unsigned NewTySize = NewTy.getSizeInBytes(); 7643 assert(NewTySize > 0 && "Could not find appropriate type"); 7644 7645 // If the new LLT cannot cover all of the remaining bits, then consider 7646 // issuing a (or a pair of) unaligned and overlapping load / store. 7647 unsigned Fast; 7648 // Need to get a VT equivalent for allowMisalignedMemoryAccesses(). 7649 MVT VT = getMVTForLLT(Ty); 7650 if (NumMemOps && Op.allowOverlap() && NewTySize < Size && 7651 TLI.allowsMisalignedMemoryAccesses( 7652 VT, DstAS, Op.isFixedDstAlign() ? Op.getDstAlign() : Align(1), 7653 MachineMemOperand::MONone, &Fast) && 7654 Fast) 7655 TySize = Size; 7656 else { 7657 Ty = NewTy; 7658 TySize = NewTySize; 7659 } 7660 } 7661 7662 if (++NumMemOps > Limit) 7663 return false; 7664 7665 MemOps.push_back(Ty); 7666 Size -= TySize; 7667 } 7668 7669 return true; 7670 } 7671 7672 static Type *getTypeForLLT(LLT Ty, LLVMContext &C) { 7673 if (Ty.isVector()) 7674 return FixedVectorType::get(IntegerType::get(C, Ty.getScalarSizeInBits()), 7675 Ty.getNumElements()); 7676 return IntegerType::get(C, Ty.getSizeInBits()); 7677 } 7678 7679 // Get a vectorized representation of the memset value operand, GISel edition. 7680 static Register getMemsetValue(Register Val, LLT Ty, MachineIRBuilder &MIB) { 7681 MachineRegisterInfo &MRI = *MIB.getMRI(); 7682 unsigned NumBits = Ty.getScalarSizeInBits(); 7683 auto ValVRegAndVal = getIConstantVRegValWithLookThrough(Val, MRI); 7684 if (!Ty.isVector() && ValVRegAndVal) { 7685 APInt Scalar = ValVRegAndVal->Value.trunc(8); 7686 APInt SplatVal = APInt::getSplat(NumBits, Scalar); 7687 return MIB.buildConstant(Ty, SplatVal).getReg(0); 7688 } 7689 7690 // Extend the byte value to the larger type, and then multiply by a magic 7691 // value 0x010101... in order to replicate it across every byte. 7692 // Unless it's zero, in which case just emit a larger G_CONSTANT 0. 7693 if (ValVRegAndVal && ValVRegAndVal->Value == 0) { 7694 return MIB.buildConstant(Ty, 0).getReg(0); 7695 } 7696 7697 LLT ExtType = Ty.getScalarType(); 7698 auto ZExt = MIB.buildZExtOrTrunc(ExtType, Val); 7699 if (NumBits > 8) { 7700 APInt Magic = APInt::getSplat(NumBits, APInt(8, 0x01)); 7701 auto MagicMI = MIB.buildConstant(ExtType, Magic); 7702 Val = MIB.buildMul(ExtType, ZExt, MagicMI).getReg(0); 7703 } 7704 7705 // For vector types create a G_BUILD_VECTOR. 7706 if (Ty.isVector()) 7707 Val = MIB.buildSplatVector(Ty, Val).getReg(0); 7708 7709 return Val; 7710 } 7711 7712 LegalizerHelper::LegalizeResult 7713 LegalizerHelper::lowerMemset(MachineInstr &MI, Register Dst, Register Val, 7714 uint64_t KnownLen, Align Alignment, 7715 bool IsVolatile) { 7716 auto &MF = *MI.getParent()->getParent(); 7717 const auto &TLI = *MF.getSubtarget().getTargetLowering(); 7718 auto &DL = MF.getDataLayout(); 7719 LLVMContext &C = MF.getFunction().getContext(); 7720 7721 assert(KnownLen != 0 && "Have a zero length memset length!"); 7722 7723 bool DstAlignCanChange = false; 7724 MachineFrameInfo &MFI = MF.getFrameInfo(); 7725 bool OptSize = shouldLowerMemFuncForSize(MF); 7726 7727 MachineInstr *FIDef = getOpcodeDef(TargetOpcode::G_FRAME_INDEX, Dst, MRI); 7728 if (FIDef && !MFI.isFixedObjectIndex(FIDef->getOperand(1).getIndex())) 7729 DstAlignCanChange = true; 7730 7731 unsigned Limit = TLI.getMaxStoresPerMemset(OptSize); 7732 std::vector<LLT> MemOps; 7733 7734 const auto &DstMMO = **MI.memoperands_begin(); 7735 MachinePointerInfo DstPtrInfo = DstMMO.getPointerInfo(); 7736 7737 auto ValVRegAndVal = getIConstantVRegValWithLookThrough(Val, MRI); 7738 bool IsZeroVal = ValVRegAndVal && ValVRegAndVal->Value == 0; 7739 7740 if (!findGISelOptimalMemOpLowering(MemOps, Limit, 7741 MemOp::Set(KnownLen, DstAlignCanChange, 7742 Alignment, 7743 /*IsZeroMemset=*/IsZeroVal, 7744 /*IsVolatile=*/IsVolatile), 7745 DstPtrInfo.getAddrSpace(), ~0u, 7746 MF.getFunction().getAttributes(), TLI)) 7747 return UnableToLegalize; 7748 7749 if (DstAlignCanChange) { 7750 // Get an estimate of the type from the LLT. 7751 Type *IRTy = getTypeForLLT(MemOps[0], C); 7752 Align NewAlign = DL.getABITypeAlign(IRTy); 7753 if (NewAlign > Alignment) { 7754 Alignment = NewAlign; 7755 unsigned FI = FIDef->getOperand(1).getIndex(); 7756 // Give the stack frame object a larger alignment if needed. 7757 if (MFI.getObjectAlign(FI) < Alignment) 7758 MFI.setObjectAlignment(FI, Alignment); 7759 } 7760 } 7761 7762 MachineIRBuilder MIB(MI); 7763 // Find the largest store and generate the bit pattern for it. 7764 LLT LargestTy = MemOps[0]; 7765 for (unsigned i = 1; i < MemOps.size(); i++) 7766 if (MemOps[i].getSizeInBits() > LargestTy.getSizeInBits()) 7767 LargestTy = MemOps[i]; 7768 7769 // The memset stored value is always defined as an s8, so in order to make it 7770 // work with larger store types we need to repeat the bit pattern across the 7771 // wider type. 7772 Register MemSetValue = getMemsetValue(Val, LargestTy, MIB); 7773 7774 if (!MemSetValue) 7775 return UnableToLegalize; 7776 7777 // Generate the stores. For each store type in the list, we generate the 7778 // matching store of that type to the destination address. 7779 LLT PtrTy = MRI.getType(Dst); 7780 unsigned DstOff = 0; 7781 unsigned Size = KnownLen; 7782 for (unsigned I = 0; I < MemOps.size(); I++) { 7783 LLT Ty = MemOps[I]; 7784 unsigned TySize = Ty.getSizeInBytes(); 7785 if (TySize > Size) { 7786 // Issuing an unaligned load / store pair that overlaps with the previous 7787 // pair. Adjust the offset accordingly. 7788 assert(I == MemOps.size() - 1 && I != 0); 7789 DstOff -= TySize - Size; 7790 } 7791 7792 // If this store is smaller than the largest store see whether we can get 7793 // the smaller value for free with a truncate. 7794 Register Value = MemSetValue; 7795 if (Ty.getSizeInBits() < LargestTy.getSizeInBits()) { 7796 MVT VT = getMVTForLLT(Ty); 7797 MVT LargestVT = getMVTForLLT(LargestTy); 7798 if (!LargestTy.isVector() && !Ty.isVector() && 7799 TLI.isTruncateFree(LargestVT, VT)) 7800 Value = MIB.buildTrunc(Ty, MemSetValue).getReg(0); 7801 else 7802 Value = getMemsetValue(Val, Ty, MIB); 7803 if (!Value) 7804 return UnableToLegalize; 7805 } 7806 7807 auto *StoreMMO = MF.getMachineMemOperand(&DstMMO, DstOff, Ty); 7808 7809 Register Ptr = Dst; 7810 if (DstOff != 0) { 7811 auto Offset = 7812 MIB.buildConstant(LLT::scalar(PtrTy.getSizeInBits()), DstOff); 7813 Ptr = MIB.buildPtrAdd(PtrTy, Dst, Offset).getReg(0); 7814 } 7815 7816 MIB.buildStore(Value, Ptr, *StoreMMO); 7817 DstOff += Ty.getSizeInBytes(); 7818 Size -= TySize; 7819 } 7820 7821 MI.eraseFromParent(); 7822 return Legalized; 7823 } 7824 7825 LegalizerHelper::LegalizeResult 7826 LegalizerHelper::lowerMemcpyInline(MachineInstr &MI) { 7827 assert(MI.getOpcode() == TargetOpcode::G_MEMCPY_INLINE); 7828 7829 Register Dst = MI.getOperand(0).getReg(); 7830 Register Src = MI.getOperand(1).getReg(); 7831 Register Len = MI.getOperand(2).getReg(); 7832 7833 const auto *MMOIt = MI.memoperands_begin(); 7834 const MachineMemOperand *MemOp = *MMOIt; 7835 bool IsVolatile = MemOp->isVolatile(); 7836 7837 // See if this is a constant length copy 7838 auto LenVRegAndVal = getIConstantVRegValWithLookThrough(Len, MRI); 7839 // FIXME: support dynamically sized G_MEMCPY_INLINE 7840 assert(LenVRegAndVal && 7841 "inline memcpy with dynamic size is not yet supported"); 7842 uint64_t KnownLen = LenVRegAndVal->Value.getZExtValue(); 7843 if (KnownLen == 0) { 7844 MI.eraseFromParent(); 7845 return Legalized; 7846 } 7847 7848 const auto &DstMMO = **MI.memoperands_begin(); 7849 const auto &SrcMMO = **std::next(MI.memoperands_begin()); 7850 Align DstAlign = DstMMO.getBaseAlign(); 7851 Align SrcAlign = SrcMMO.getBaseAlign(); 7852 7853 return lowerMemcpyInline(MI, Dst, Src, KnownLen, DstAlign, SrcAlign, 7854 IsVolatile); 7855 } 7856 7857 LegalizerHelper::LegalizeResult 7858 LegalizerHelper::lowerMemcpyInline(MachineInstr &MI, Register Dst, Register Src, 7859 uint64_t KnownLen, Align DstAlign, 7860 Align SrcAlign, bool IsVolatile) { 7861 assert(MI.getOpcode() == TargetOpcode::G_MEMCPY_INLINE); 7862 return lowerMemcpy(MI, Dst, Src, KnownLen, 7863 std::numeric_limits<uint64_t>::max(), DstAlign, SrcAlign, 7864 IsVolatile); 7865 } 7866 7867 LegalizerHelper::LegalizeResult 7868 LegalizerHelper::lowerMemcpy(MachineInstr &MI, Register Dst, Register Src, 7869 uint64_t KnownLen, uint64_t Limit, Align DstAlign, 7870 Align SrcAlign, bool IsVolatile) { 7871 auto &MF = *MI.getParent()->getParent(); 7872 const auto &TLI = *MF.getSubtarget().getTargetLowering(); 7873 auto &DL = MF.getDataLayout(); 7874 LLVMContext &C = MF.getFunction().getContext(); 7875 7876 assert(KnownLen != 0 && "Have a zero length memcpy length!"); 7877 7878 bool DstAlignCanChange = false; 7879 MachineFrameInfo &MFI = MF.getFrameInfo(); 7880 Align Alignment = std::min(DstAlign, SrcAlign); 7881 7882 MachineInstr *FIDef = getOpcodeDef(TargetOpcode::G_FRAME_INDEX, Dst, MRI); 7883 if (FIDef && !MFI.isFixedObjectIndex(FIDef->getOperand(1).getIndex())) 7884 DstAlignCanChange = true; 7885 7886 // FIXME: infer better src pointer alignment like SelectionDAG does here. 7887 // FIXME: also use the equivalent of isMemSrcFromConstant and alwaysinlining 7888 // if the memcpy is in a tail call position. 7889 7890 std::vector<LLT> MemOps; 7891 7892 const auto &DstMMO = **MI.memoperands_begin(); 7893 const auto &SrcMMO = **std::next(MI.memoperands_begin()); 7894 MachinePointerInfo DstPtrInfo = DstMMO.getPointerInfo(); 7895 MachinePointerInfo SrcPtrInfo = SrcMMO.getPointerInfo(); 7896 7897 if (!findGISelOptimalMemOpLowering( 7898 MemOps, Limit, 7899 MemOp::Copy(KnownLen, DstAlignCanChange, Alignment, SrcAlign, 7900 IsVolatile), 7901 DstPtrInfo.getAddrSpace(), SrcPtrInfo.getAddrSpace(), 7902 MF.getFunction().getAttributes(), TLI)) 7903 return UnableToLegalize; 7904 7905 if (DstAlignCanChange) { 7906 // Get an estimate of the type from the LLT. 7907 Type *IRTy = getTypeForLLT(MemOps[0], C); 7908 Align NewAlign = DL.getABITypeAlign(IRTy); 7909 7910 // Don't promote to an alignment that would require dynamic stack 7911 // realignment. 7912 const TargetRegisterInfo *TRI = MF.getSubtarget().getRegisterInfo(); 7913 if (!TRI->hasStackRealignment(MF)) 7914 while (NewAlign > Alignment && DL.exceedsNaturalStackAlignment(NewAlign)) 7915 NewAlign = NewAlign.previous(); 7916 7917 if (NewAlign > Alignment) { 7918 Alignment = NewAlign; 7919 unsigned FI = FIDef->getOperand(1).getIndex(); 7920 // Give the stack frame object a larger alignment if needed. 7921 if (MFI.getObjectAlign(FI) < Alignment) 7922 MFI.setObjectAlignment(FI, Alignment); 7923 } 7924 } 7925 7926 LLVM_DEBUG(dbgs() << "Inlining memcpy: " << MI << " into loads & stores\n"); 7927 7928 MachineIRBuilder MIB(MI); 7929 // Now we need to emit a pair of load and stores for each of the types we've 7930 // collected. I.e. for each type, generate a load from the source pointer of 7931 // that type width, and then generate a corresponding store to the dest buffer 7932 // of that value loaded. This can result in a sequence of loads and stores 7933 // mixed types, depending on what the target specifies as good types to use. 7934 unsigned CurrOffset = 0; 7935 unsigned Size = KnownLen; 7936 for (auto CopyTy : MemOps) { 7937 // Issuing an unaligned load / store pair that overlaps with the previous 7938 // pair. Adjust the offset accordingly. 7939 if (CopyTy.getSizeInBytes() > Size) 7940 CurrOffset -= CopyTy.getSizeInBytes() - Size; 7941 7942 // Construct MMOs for the accesses. 7943 auto *LoadMMO = 7944 MF.getMachineMemOperand(&SrcMMO, CurrOffset, CopyTy.getSizeInBytes()); 7945 auto *StoreMMO = 7946 MF.getMachineMemOperand(&DstMMO, CurrOffset, CopyTy.getSizeInBytes()); 7947 7948 // Create the load. 7949 Register LoadPtr = Src; 7950 Register Offset; 7951 if (CurrOffset != 0) { 7952 LLT SrcTy = MRI.getType(Src); 7953 Offset = MIB.buildConstant(LLT::scalar(SrcTy.getSizeInBits()), CurrOffset) 7954 .getReg(0); 7955 LoadPtr = MIB.buildPtrAdd(SrcTy, Src, Offset).getReg(0); 7956 } 7957 auto LdVal = MIB.buildLoad(CopyTy, LoadPtr, *LoadMMO); 7958 7959 // Create the store. 7960 Register StorePtr = Dst; 7961 if (CurrOffset != 0) { 7962 LLT DstTy = MRI.getType(Dst); 7963 StorePtr = MIB.buildPtrAdd(DstTy, Dst, Offset).getReg(0); 7964 } 7965 MIB.buildStore(LdVal, StorePtr, *StoreMMO); 7966 CurrOffset += CopyTy.getSizeInBytes(); 7967 Size -= CopyTy.getSizeInBytes(); 7968 } 7969 7970 MI.eraseFromParent(); 7971 return Legalized; 7972 } 7973 7974 LegalizerHelper::LegalizeResult 7975 LegalizerHelper::lowerMemmove(MachineInstr &MI, Register Dst, Register Src, 7976 uint64_t KnownLen, Align DstAlign, Align SrcAlign, 7977 bool IsVolatile) { 7978 auto &MF = *MI.getParent()->getParent(); 7979 const auto &TLI = *MF.getSubtarget().getTargetLowering(); 7980 auto &DL = MF.getDataLayout(); 7981 LLVMContext &C = MF.getFunction().getContext(); 7982 7983 assert(KnownLen != 0 && "Have a zero length memmove length!"); 7984 7985 bool DstAlignCanChange = false; 7986 MachineFrameInfo &MFI = MF.getFrameInfo(); 7987 bool OptSize = shouldLowerMemFuncForSize(MF); 7988 Align Alignment = std::min(DstAlign, SrcAlign); 7989 7990 MachineInstr *FIDef = getOpcodeDef(TargetOpcode::G_FRAME_INDEX, Dst, MRI); 7991 if (FIDef && !MFI.isFixedObjectIndex(FIDef->getOperand(1).getIndex())) 7992 DstAlignCanChange = true; 7993 7994 unsigned Limit = TLI.getMaxStoresPerMemmove(OptSize); 7995 std::vector<LLT> MemOps; 7996 7997 const auto &DstMMO = **MI.memoperands_begin(); 7998 const auto &SrcMMO = **std::next(MI.memoperands_begin()); 7999 MachinePointerInfo DstPtrInfo = DstMMO.getPointerInfo(); 8000 MachinePointerInfo SrcPtrInfo = SrcMMO.getPointerInfo(); 8001 8002 // FIXME: SelectionDAG always passes false for 'AllowOverlap', apparently due 8003 // to a bug in it's findOptimalMemOpLowering implementation. For now do the 8004 // same thing here. 8005 if (!findGISelOptimalMemOpLowering( 8006 MemOps, Limit, 8007 MemOp::Copy(KnownLen, DstAlignCanChange, Alignment, SrcAlign, 8008 /*IsVolatile*/ true), 8009 DstPtrInfo.getAddrSpace(), SrcPtrInfo.getAddrSpace(), 8010 MF.getFunction().getAttributes(), TLI)) 8011 return UnableToLegalize; 8012 8013 if (DstAlignCanChange) { 8014 // Get an estimate of the type from the LLT. 8015 Type *IRTy = getTypeForLLT(MemOps[0], C); 8016 Align NewAlign = DL.getABITypeAlign(IRTy); 8017 8018 // Don't promote to an alignment that would require dynamic stack 8019 // realignment. 8020 const TargetRegisterInfo *TRI = MF.getSubtarget().getRegisterInfo(); 8021 if (!TRI->hasStackRealignment(MF)) 8022 while (NewAlign > Alignment && DL.exceedsNaturalStackAlignment(NewAlign)) 8023 NewAlign = NewAlign.previous(); 8024 8025 if (NewAlign > Alignment) { 8026 Alignment = NewAlign; 8027 unsigned FI = FIDef->getOperand(1).getIndex(); 8028 // Give the stack frame object a larger alignment if needed. 8029 if (MFI.getObjectAlign(FI) < Alignment) 8030 MFI.setObjectAlignment(FI, Alignment); 8031 } 8032 } 8033 8034 LLVM_DEBUG(dbgs() << "Inlining memmove: " << MI << " into loads & stores\n"); 8035 8036 MachineIRBuilder MIB(MI); 8037 // Memmove requires that we perform the loads first before issuing the stores. 8038 // Apart from that, this loop is pretty much doing the same thing as the 8039 // memcpy codegen function. 8040 unsigned CurrOffset = 0; 8041 SmallVector<Register, 16> LoadVals; 8042 for (auto CopyTy : MemOps) { 8043 // Construct MMO for the load. 8044 auto *LoadMMO = 8045 MF.getMachineMemOperand(&SrcMMO, CurrOffset, CopyTy.getSizeInBytes()); 8046 8047 // Create the load. 8048 Register LoadPtr = Src; 8049 if (CurrOffset != 0) { 8050 LLT SrcTy = MRI.getType(Src); 8051 auto Offset = 8052 MIB.buildConstant(LLT::scalar(SrcTy.getSizeInBits()), CurrOffset); 8053 LoadPtr = MIB.buildPtrAdd(SrcTy, Src, Offset).getReg(0); 8054 } 8055 LoadVals.push_back(MIB.buildLoad(CopyTy, LoadPtr, *LoadMMO).getReg(0)); 8056 CurrOffset += CopyTy.getSizeInBytes(); 8057 } 8058 8059 CurrOffset = 0; 8060 for (unsigned I = 0; I < MemOps.size(); ++I) { 8061 LLT CopyTy = MemOps[I]; 8062 // Now store the values loaded. 8063 auto *StoreMMO = 8064 MF.getMachineMemOperand(&DstMMO, CurrOffset, CopyTy.getSizeInBytes()); 8065 8066 Register StorePtr = Dst; 8067 if (CurrOffset != 0) { 8068 LLT DstTy = MRI.getType(Dst); 8069 auto Offset = 8070 MIB.buildConstant(LLT::scalar(DstTy.getSizeInBits()), CurrOffset); 8071 StorePtr = MIB.buildPtrAdd(DstTy, Dst, Offset).getReg(0); 8072 } 8073 MIB.buildStore(LoadVals[I], StorePtr, *StoreMMO); 8074 CurrOffset += CopyTy.getSizeInBytes(); 8075 } 8076 MI.eraseFromParent(); 8077 return Legalized; 8078 } 8079 8080 LegalizerHelper::LegalizeResult 8081 LegalizerHelper::lowerMemCpyFamily(MachineInstr &MI, unsigned MaxLen) { 8082 const unsigned Opc = MI.getOpcode(); 8083 // This combine is fairly complex so it's not written with a separate 8084 // matcher function. 8085 assert((Opc == TargetOpcode::G_MEMCPY || Opc == TargetOpcode::G_MEMMOVE || 8086 Opc == TargetOpcode::G_MEMSET) && 8087 "Expected memcpy like instruction"); 8088 8089 auto MMOIt = MI.memoperands_begin(); 8090 const MachineMemOperand *MemOp = *MMOIt; 8091 8092 Align DstAlign = MemOp->getBaseAlign(); 8093 Align SrcAlign; 8094 Register Dst = MI.getOperand(0).getReg(); 8095 Register Src = MI.getOperand(1).getReg(); 8096 Register Len = MI.getOperand(2).getReg(); 8097 8098 if (Opc != TargetOpcode::G_MEMSET) { 8099 assert(MMOIt != MI.memoperands_end() && "Expected a second MMO on MI"); 8100 MemOp = *(++MMOIt); 8101 SrcAlign = MemOp->getBaseAlign(); 8102 } 8103 8104 // See if this is a constant length copy 8105 auto LenVRegAndVal = getIConstantVRegValWithLookThrough(Len, MRI); 8106 if (!LenVRegAndVal) 8107 return UnableToLegalize; 8108 uint64_t KnownLen = LenVRegAndVal->Value.getZExtValue(); 8109 8110 if (KnownLen == 0) { 8111 MI.eraseFromParent(); 8112 return Legalized; 8113 } 8114 8115 bool IsVolatile = MemOp->isVolatile(); 8116 if (Opc == TargetOpcode::G_MEMCPY_INLINE) 8117 return lowerMemcpyInline(MI, Dst, Src, KnownLen, DstAlign, SrcAlign, 8118 IsVolatile); 8119 8120 // Don't try to optimize volatile. 8121 if (IsVolatile) 8122 return UnableToLegalize; 8123 8124 if (MaxLen && KnownLen > MaxLen) 8125 return UnableToLegalize; 8126 8127 if (Opc == TargetOpcode::G_MEMCPY) { 8128 auto &MF = *MI.getParent()->getParent(); 8129 const auto &TLI = *MF.getSubtarget().getTargetLowering(); 8130 bool OptSize = shouldLowerMemFuncForSize(MF); 8131 uint64_t Limit = TLI.getMaxStoresPerMemcpy(OptSize); 8132 return lowerMemcpy(MI, Dst, Src, KnownLen, Limit, DstAlign, SrcAlign, 8133 IsVolatile); 8134 } 8135 if (Opc == TargetOpcode::G_MEMMOVE) 8136 return lowerMemmove(MI, Dst, Src, KnownLen, DstAlign, SrcAlign, IsVolatile); 8137 if (Opc == TargetOpcode::G_MEMSET) 8138 return lowerMemset(MI, Dst, Src, KnownLen, DstAlign, IsVolatile); 8139 return UnableToLegalize; 8140 } 8141