1 //===- SPIRVInstructionSelector.cpp ------------------------------*- C++ -*-==// 2 // 3 // Part of the LLVM Project, under the Apache License v2.0 with LLVM Exceptions. 4 // See https://llvm.org/LICENSE.txt for license information. 5 // SPDX-License-Identifier: Apache-2.0 WITH LLVM-exception 6 // 7 //===----------------------------------------------------------------------===// 8 // 9 // This file implements the targeting of the InstructionSelector class for 10 // SPIRV. 11 // TODO: This should be generated by TableGen. 12 // 13 //===----------------------------------------------------------------------===// 14 15 #include "MCTargetDesc/SPIRVBaseInfo.h" 16 #include "MCTargetDesc/SPIRVMCTargetDesc.h" 17 #include "SPIRV.h" 18 #include "SPIRVGlobalRegistry.h" 19 #include "SPIRVInstrInfo.h" 20 #include "SPIRVRegisterBankInfo.h" 21 #include "SPIRVRegisterInfo.h" 22 #include "SPIRVTargetMachine.h" 23 #include "SPIRVUtils.h" 24 #include "llvm/ADT/APFloat.h" 25 #include "llvm/CodeGen/GlobalISel/GIMatchTableExecutorImpl.h" 26 #include "llvm/CodeGen/GlobalISel/GenericMachineInstrs.h" 27 #include "llvm/CodeGen/GlobalISel/InstructionSelector.h" 28 #include "llvm/CodeGen/MachineInstrBuilder.h" 29 #include "llvm/CodeGen/MachineModuleInfoImpls.h" 30 #include "llvm/CodeGen/MachineRegisterInfo.h" 31 #include "llvm/CodeGen/TargetOpcodes.h" 32 #include "llvm/IR/IntrinsicsSPIRV.h" 33 #include "llvm/Support/Debug.h" 34 35 namespace llvm { 36 37 class SPIRVMachineModuleInfo : public MachineModuleInfoImpl { 38 public: 39 SyncScope::ID Work_ItemSSID; 40 SyncScope::ID WorkGroupSSID; 41 SyncScope::ID DeviceSSID; 42 SyncScope::ID AllSVMDevicesSSID; 43 SyncScope::ID SubGroupSSID; 44 45 SPIRVMachineModuleInfo(const MachineModuleInfo &MMI) { 46 LLVMContext &CTX = MMI.getModule()->getContext(); 47 Work_ItemSSID = CTX.getOrInsertSyncScopeID("work_item"); 48 WorkGroupSSID = CTX.getOrInsertSyncScopeID("workgroup"); 49 DeviceSSID = CTX.getOrInsertSyncScopeID("device"); 50 AllSVMDevicesSSID = CTX.getOrInsertSyncScopeID("all_svm_devices"); 51 SubGroupSSID = CTX.getOrInsertSyncScopeID("sub_group"); 52 } 53 }; 54 55 } // end namespace llvm 56 57 #define DEBUG_TYPE "spirv-isel" 58 59 using namespace llvm; 60 namespace CL = SPIRV::OpenCLExtInst; 61 namespace GL = SPIRV::GLSLExtInst; 62 63 using ExtInstList = 64 std::vector<std::pair<SPIRV::InstructionSet::InstructionSet, uint32_t>>; 65 66 namespace { 67 68 #define GET_GLOBALISEL_PREDICATE_BITSET 69 #include "SPIRVGenGlobalISel.inc" 70 #undef GET_GLOBALISEL_PREDICATE_BITSET 71 72 class SPIRVInstructionSelector : public InstructionSelector { 73 const SPIRVSubtarget &STI; 74 const SPIRVInstrInfo &TII; 75 const SPIRVRegisterInfo &TRI; 76 const RegisterBankInfo &RBI; 77 SPIRVGlobalRegistry &GR; 78 MachineRegisterInfo *MRI; 79 SPIRVMachineModuleInfo *MMI = nullptr; 80 81 /// We need to keep track of the number we give to anonymous global values to 82 /// generate the same name every time when this is needed. 83 mutable DenseMap<const GlobalValue *, unsigned> UnnamedGlobalIDs; 84 85 public: 86 SPIRVInstructionSelector(const SPIRVTargetMachine &TM, 87 const SPIRVSubtarget &ST, 88 const RegisterBankInfo &RBI); 89 void setupMF(MachineFunction &MF, GISelKnownBits *KB, 90 CodeGenCoverage *CoverageInfo, ProfileSummaryInfo *PSI, 91 BlockFrequencyInfo *BFI) override; 92 // Common selection code. Instruction-specific selection occurs in spvSelect. 93 bool select(MachineInstr &I) override; 94 static const char *getName() { return DEBUG_TYPE; } 95 96 #define GET_GLOBALISEL_PREDICATES_DECL 97 #include "SPIRVGenGlobalISel.inc" 98 #undef GET_GLOBALISEL_PREDICATES_DECL 99 100 #define GET_GLOBALISEL_TEMPORARIES_DECL 101 #include "SPIRVGenGlobalISel.inc" 102 #undef GET_GLOBALISEL_TEMPORARIES_DECL 103 104 private: 105 // tblgen-erated 'select' implementation, used as the initial selector for 106 // the patterns that don't require complex C++. 107 bool selectImpl(MachineInstr &I, CodeGenCoverage &CoverageInfo) const; 108 109 // All instruction-specific selection that didn't happen in "select()". 110 // Is basically a large Switch/Case delegating to all other select method. 111 bool spvSelect(Register ResVReg, const SPIRVType *ResType, 112 MachineInstr &I) const; 113 114 bool selectGlobalValue(Register ResVReg, MachineInstr &I, 115 const MachineInstr *Init = nullptr) const; 116 117 bool selectUnOpWithSrc(Register ResVReg, const SPIRVType *ResType, 118 MachineInstr &I, Register SrcReg, 119 unsigned Opcode) const; 120 bool selectUnOp(Register ResVReg, const SPIRVType *ResType, MachineInstr &I, 121 unsigned Opcode) const; 122 123 bool selectBitcast(Register ResVReg, const SPIRVType *ResType, 124 MachineInstr &I) const; 125 126 bool selectLoad(Register ResVReg, const SPIRVType *ResType, 127 MachineInstr &I) const; 128 bool selectStore(MachineInstr &I) const; 129 130 bool selectStackSave(Register ResVReg, const SPIRVType *ResType, 131 MachineInstr &I) const; 132 bool selectStackRestore(MachineInstr &I) const; 133 134 bool selectMemOperation(Register ResVReg, MachineInstr &I) const; 135 136 bool selectAtomicRMW(Register ResVReg, const SPIRVType *ResType, 137 MachineInstr &I, unsigned NewOpcode, 138 unsigned NegateOpcode = 0) const; 139 140 bool selectAtomicCmpXchg(Register ResVReg, const SPIRVType *ResType, 141 MachineInstr &I) const; 142 143 bool selectFence(MachineInstr &I) const; 144 145 bool selectAddrSpaceCast(Register ResVReg, const SPIRVType *ResType, 146 MachineInstr &I) const; 147 148 bool selectAnyOrAll(Register ResVReg, const SPIRVType *ResType, 149 MachineInstr &I, unsigned OpType) const; 150 151 bool selectAll(Register ResVReg, const SPIRVType *ResType, 152 MachineInstr &I) const; 153 154 bool selectAny(Register ResVReg, const SPIRVType *ResType, 155 MachineInstr &I) const; 156 157 bool selectBitreverse(Register ResVReg, const SPIRVType *ResType, 158 MachineInstr &I) const; 159 160 bool selectConstVector(Register ResVReg, const SPIRVType *ResType, 161 MachineInstr &I) const; 162 bool selectSplatVector(Register ResVReg, const SPIRVType *ResType, 163 MachineInstr &I) const; 164 165 bool selectCmp(Register ResVReg, const SPIRVType *ResType, 166 unsigned comparisonOpcode, MachineInstr &I) const; 167 168 bool selectICmp(Register ResVReg, const SPIRVType *ResType, 169 MachineInstr &I) const; 170 bool selectFCmp(Register ResVReg, const SPIRVType *ResType, 171 MachineInstr &I) const; 172 173 bool selectFmix(Register ResVReg, const SPIRVType *ResType, 174 MachineInstr &I) const; 175 176 void renderImm32(MachineInstrBuilder &MIB, const MachineInstr &I, 177 int OpIdx) const; 178 void renderFImm32(MachineInstrBuilder &MIB, const MachineInstr &I, 179 int OpIdx) const; 180 181 bool selectConst(Register ResVReg, const SPIRVType *ResType, const APInt &Imm, 182 MachineInstr &I) const; 183 184 bool selectSelect(Register ResVReg, const SPIRVType *ResType, MachineInstr &I, 185 bool IsSigned) const; 186 bool selectIToF(Register ResVReg, const SPIRVType *ResType, MachineInstr &I, 187 bool IsSigned, unsigned Opcode) const; 188 bool selectExt(Register ResVReg, const SPIRVType *ResType, MachineInstr &I, 189 bool IsSigned) const; 190 191 bool selectTrunc(Register ResVReg, const SPIRVType *ResType, 192 MachineInstr &I) const; 193 194 bool selectIntToBool(Register IntReg, Register ResVReg, MachineInstr &I, 195 const SPIRVType *intTy, const SPIRVType *boolTy) const; 196 197 bool selectOpUndef(Register ResVReg, const SPIRVType *ResType, 198 MachineInstr &I) const; 199 bool selectFreeze(Register ResVReg, const SPIRVType *ResType, 200 MachineInstr &I) const; 201 bool selectIntrinsic(Register ResVReg, const SPIRVType *ResType, 202 MachineInstr &I) const; 203 bool selectExtractVal(Register ResVReg, const SPIRVType *ResType, 204 MachineInstr &I) const; 205 bool selectInsertVal(Register ResVReg, const SPIRVType *ResType, 206 MachineInstr &I) const; 207 bool selectExtractElt(Register ResVReg, const SPIRVType *ResType, 208 MachineInstr &I) const; 209 bool selectInsertElt(Register ResVReg, const SPIRVType *ResType, 210 MachineInstr &I) const; 211 bool selectGEP(Register ResVReg, const SPIRVType *ResType, 212 MachineInstr &I) const; 213 214 bool selectFrameIndex(Register ResVReg, const SPIRVType *ResType, 215 MachineInstr &I) const; 216 bool selectAllocaArray(Register ResVReg, const SPIRVType *ResType, 217 MachineInstr &I) const; 218 219 bool selectBranch(MachineInstr &I) const; 220 bool selectBranchCond(MachineInstr &I) const; 221 222 bool selectPhi(Register ResVReg, const SPIRVType *ResType, 223 MachineInstr &I) const; 224 225 bool selectExtInst(Register ResVReg, const SPIRVType *ResType, 226 MachineInstr &I, CL::OpenCLExtInst CLInst) const; 227 bool selectExtInst(Register ResVReg, const SPIRVType *ResType, 228 MachineInstr &I, CL::OpenCLExtInst CLInst, 229 GL::GLSLExtInst GLInst) const; 230 bool selectExtInst(Register ResVReg, const SPIRVType *ResType, 231 MachineInstr &I, const ExtInstList &ExtInsts) const; 232 233 bool selectLog10(Register ResVReg, const SPIRVType *ResType, 234 MachineInstr &I) const; 235 236 bool selectSpvThreadId(Register ResVReg, const SPIRVType *ResType, 237 MachineInstr &I) const; 238 239 bool selectUnmergeValues(MachineInstr &I) const; 240 241 Register buildI32Constant(uint32_t Val, MachineInstr &I, 242 const SPIRVType *ResType = nullptr) const; 243 244 Register buildZerosVal(const SPIRVType *ResType, MachineInstr &I) const; 245 Register buildZerosValF(const SPIRVType *ResType, MachineInstr &I) const; 246 Register buildOnesVal(bool AllOnes, const SPIRVType *ResType, 247 MachineInstr &I) const; 248 249 bool wrapIntoSpecConstantOp(MachineInstr &I, 250 SmallVector<Register> &CompositeArgs) const; 251 }; 252 253 } // end anonymous namespace 254 255 #define GET_GLOBALISEL_IMPL 256 #include "SPIRVGenGlobalISel.inc" 257 #undef GET_GLOBALISEL_IMPL 258 259 SPIRVInstructionSelector::SPIRVInstructionSelector(const SPIRVTargetMachine &TM, 260 const SPIRVSubtarget &ST, 261 const RegisterBankInfo &RBI) 262 : InstructionSelector(), STI(ST), TII(*ST.getInstrInfo()), 263 TRI(*ST.getRegisterInfo()), RBI(RBI), GR(*ST.getSPIRVGlobalRegistry()), 264 #define GET_GLOBALISEL_PREDICATES_INIT 265 #include "SPIRVGenGlobalISel.inc" 266 #undef GET_GLOBALISEL_PREDICATES_INIT 267 #define GET_GLOBALISEL_TEMPORARIES_INIT 268 #include "SPIRVGenGlobalISel.inc" 269 #undef GET_GLOBALISEL_TEMPORARIES_INIT 270 { 271 } 272 273 void SPIRVInstructionSelector::setupMF(MachineFunction &MF, GISelKnownBits *KB, 274 CodeGenCoverage *CoverageInfo, 275 ProfileSummaryInfo *PSI, 276 BlockFrequencyInfo *BFI) { 277 MMI = &MF.getMMI().getObjFileInfo<SPIRVMachineModuleInfo>(); 278 MRI = &MF.getRegInfo(); 279 GR.setCurrentFunc(MF); 280 InstructionSelector::setupMF(MF, KB, CoverageInfo, PSI, BFI); 281 } 282 283 static bool isImm(const MachineOperand &MO, MachineRegisterInfo *MRI); 284 285 // Defined in SPIRVLegalizerInfo.cpp. 286 extern bool isTypeFoldingSupported(unsigned Opcode); 287 288 bool SPIRVInstructionSelector::select(MachineInstr &I) { 289 assert(I.getParent() && "Instruction should be in a basic block!"); 290 assert(I.getParent()->getParent() && "Instruction should be in a function!"); 291 292 Register Opcode = I.getOpcode(); 293 // If it's not a GMIR instruction, we've selected it already. 294 if (!isPreISelGenericOpcode(Opcode)) { 295 if (Opcode == SPIRV::ASSIGN_TYPE) { // These pseudos aren't needed any more. 296 Register DstReg = I.getOperand(0).getReg(); 297 Register SrcReg = I.getOperand(1).getReg(); 298 auto *Def = MRI->getVRegDef(SrcReg); 299 if (isTypeFoldingSupported(Def->getOpcode())) { 300 if (MRI->getType(DstReg).isPointer()) 301 MRI->setType(DstReg, LLT::scalar(32)); 302 bool Res = selectImpl(I, *CoverageInfo); 303 assert(Res || Def->getOpcode() == TargetOpcode::G_CONSTANT); 304 if (Res) 305 return Res; 306 } 307 MRI->replaceRegWith(SrcReg, DstReg); 308 I.removeFromParent(); 309 return true; 310 } else if (I.getNumDefs() == 1) { 311 // Make all vregs 32 bits (for SPIR-V IDs). 312 MRI->setType(I.getOperand(0).getReg(), LLT::scalar(32)); 313 } 314 return constrainSelectedInstRegOperands(I, TII, TRI, RBI); 315 } 316 317 if (I.getNumOperands() != I.getNumExplicitOperands()) { 318 LLVM_DEBUG(errs() << "Generic instr has unexpected implicit operands\n"); 319 return false; 320 } 321 322 // Common code for getting return reg+type, and removing selected instr 323 // from parent occurs here. Instr-specific selection happens in spvSelect(). 324 bool HasDefs = I.getNumDefs() > 0; 325 Register ResVReg = HasDefs ? I.getOperand(0).getReg() : Register(0); 326 SPIRVType *ResType = HasDefs ? GR.getSPIRVTypeForVReg(ResVReg) : nullptr; 327 assert(!HasDefs || ResType || I.getOpcode() == TargetOpcode::G_GLOBAL_VALUE); 328 if (spvSelect(ResVReg, ResType, I)) { 329 if (HasDefs) // Make all vregs 32 bits (for SPIR-V IDs). 330 for (unsigned i = 0; i < I.getNumDefs(); ++i) 331 MRI->setType(I.getOperand(i).getReg(), LLT::scalar(32)); 332 I.removeFromParent(); 333 return true; 334 } 335 return false; 336 } 337 338 bool SPIRVInstructionSelector::spvSelect(Register ResVReg, 339 const SPIRVType *ResType, 340 MachineInstr &I) const { 341 const unsigned Opcode = I.getOpcode(); 342 if (isTypeFoldingSupported(Opcode) && Opcode != TargetOpcode::G_CONSTANT) 343 return selectImpl(I, *CoverageInfo); 344 switch (Opcode) { 345 case TargetOpcode::G_CONSTANT: 346 return selectConst(ResVReg, ResType, I.getOperand(1).getCImm()->getValue(), 347 I); 348 case TargetOpcode::G_GLOBAL_VALUE: 349 return selectGlobalValue(ResVReg, I); 350 case TargetOpcode::G_IMPLICIT_DEF: 351 return selectOpUndef(ResVReg, ResType, I); 352 case TargetOpcode::G_FREEZE: 353 return selectFreeze(ResVReg, ResType, I); 354 355 case TargetOpcode::G_INTRINSIC: 356 case TargetOpcode::G_INTRINSIC_W_SIDE_EFFECTS: 357 case TargetOpcode::G_INTRINSIC_CONVERGENT_W_SIDE_EFFECTS: 358 return selectIntrinsic(ResVReg, ResType, I); 359 case TargetOpcode::G_BITREVERSE: 360 return selectBitreverse(ResVReg, ResType, I); 361 362 case TargetOpcode::G_BUILD_VECTOR: 363 return selectConstVector(ResVReg, ResType, I); 364 case TargetOpcode::G_SPLAT_VECTOR: 365 return selectSplatVector(ResVReg, ResType, I); 366 367 case TargetOpcode::G_SHUFFLE_VECTOR: { 368 MachineBasicBlock &BB = *I.getParent(); 369 auto MIB = BuildMI(BB, I, I.getDebugLoc(), TII.get(SPIRV::OpVectorShuffle)) 370 .addDef(ResVReg) 371 .addUse(GR.getSPIRVTypeID(ResType)) 372 .addUse(I.getOperand(1).getReg()) 373 .addUse(I.getOperand(2).getReg()); 374 for (auto V : I.getOperand(3).getShuffleMask()) 375 MIB.addImm(V); 376 return MIB.constrainAllUses(TII, TRI, RBI); 377 } 378 case TargetOpcode::G_MEMMOVE: 379 case TargetOpcode::G_MEMCPY: 380 case TargetOpcode::G_MEMSET: 381 return selectMemOperation(ResVReg, I); 382 383 case TargetOpcode::G_ICMP: 384 return selectICmp(ResVReg, ResType, I); 385 case TargetOpcode::G_FCMP: 386 return selectFCmp(ResVReg, ResType, I); 387 388 case TargetOpcode::G_FRAME_INDEX: 389 return selectFrameIndex(ResVReg, ResType, I); 390 391 case TargetOpcode::G_LOAD: 392 return selectLoad(ResVReg, ResType, I); 393 case TargetOpcode::G_STORE: 394 return selectStore(I); 395 396 case TargetOpcode::G_BR: 397 return selectBranch(I); 398 case TargetOpcode::G_BRCOND: 399 return selectBranchCond(I); 400 401 case TargetOpcode::G_PHI: 402 return selectPhi(ResVReg, ResType, I); 403 404 case TargetOpcode::G_FPTOSI: 405 return selectUnOp(ResVReg, ResType, I, SPIRV::OpConvertFToS); 406 case TargetOpcode::G_FPTOUI: 407 return selectUnOp(ResVReg, ResType, I, SPIRV::OpConvertFToU); 408 409 case TargetOpcode::G_SITOFP: 410 return selectIToF(ResVReg, ResType, I, true, SPIRV::OpConvertSToF); 411 case TargetOpcode::G_UITOFP: 412 return selectIToF(ResVReg, ResType, I, false, SPIRV::OpConvertUToF); 413 414 case TargetOpcode::G_CTPOP: 415 return selectUnOp(ResVReg, ResType, I, SPIRV::OpBitCount); 416 case TargetOpcode::G_SMIN: 417 return selectExtInst(ResVReg, ResType, I, CL::s_min, GL::SMin); 418 case TargetOpcode::G_UMIN: 419 return selectExtInst(ResVReg, ResType, I, CL::u_min, GL::UMin); 420 421 case TargetOpcode::G_SMAX: 422 return selectExtInst(ResVReg, ResType, I, CL::s_max, GL::SMax); 423 case TargetOpcode::G_UMAX: 424 return selectExtInst(ResVReg, ResType, I, CL::u_max, GL::UMax); 425 426 case TargetOpcode::G_FMA: 427 return selectExtInst(ResVReg, ResType, I, CL::fma, GL::Fma); 428 429 case TargetOpcode::G_FPOW: 430 return selectExtInst(ResVReg, ResType, I, CL::pow, GL::Pow); 431 case TargetOpcode::G_FPOWI: 432 return selectExtInst(ResVReg, ResType, I, CL::pown); 433 434 case TargetOpcode::G_FEXP: 435 return selectExtInst(ResVReg, ResType, I, CL::exp, GL::Exp); 436 case TargetOpcode::G_FEXP2: 437 return selectExtInst(ResVReg, ResType, I, CL::exp2, GL::Exp2); 438 439 case TargetOpcode::G_FLOG: 440 return selectExtInst(ResVReg, ResType, I, CL::log, GL::Log); 441 case TargetOpcode::G_FLOG2: 442 return selectExtInst(ResVReg, ResType, I, CL::log2, GL::Log2); 443 case TargetOpcode::G_FLOG10: 444 return selectLog10(ResVReg, ResType, I); 445 446 case TargetOpcode::G_FABS: 447 return selectExtInst(ResVReg, ResType, I, CL::fabs, GL::FAbs); 448 case TargetOpcode::G_ABS: 449 return selectExtInst(ResVReg, ResType, I, CL::s_abs, GL::SAbs); 450 451 case TargetOpcode::G_FMINNUM: 452 case TargetOpcode::G_FMINIMUM: 453 return selectExtInst(ResVReg, ResType, I, CL::fmin, GL::NMin); 454 case TargetOpcode::G_FMAXNUM: 455 case TargetOpcode::G_FMAXIMUM: 456 return selectExtInst(ResVReg, ResType, I, CL::fmax, GL::NMax); 457 458 case TargetOpcode::G_FCOPYSIGN: 459 return selectExtInst(ResVReg, ResType, I, CL::copysign); 460 461 case TargetOpcode::G_FCEIL: 462 return selectExtInst(ResVReg, ResType, I, CL::ceil, GL::Ceil); 463 case TargetOpcode::G_FFLOOR: 464 return selectExtInst(ResVReg, ResType, I, CL::floor, GL::Floor); 465 466 case TargetOpcode::G_FCOS: 467 return selectExtInst(ResVReg, ResType, I, CL::cos, GL::Cos); 468 case TargetOpcode::G_FSIN: 469 return selectExtInst(ResVReg, ResType, I, CL::sin, GL::Sin); 470 471 case TargetOpcode::G_FSQRT: 472 return selectExtInst(ResVReg, ResType, I, CL::sqrt, GL::Sqrt); 473 474 case TargetOpcode::G_CTTZ: 475 case TargetOpcode::G_CTTZ_ZERO_UNDEF: 476 return selectExtInst(ResVReg, ResType, I, CL::ctz); 477 case TargetOpcode::G_CTLZ: 478 case TargetOpcode::G_CTLZ_ZERO_UNDEF: 479 return selectExtInst(ResVReg, ResType, I, CL::clz); 480 481 case TargetOpcode::G_INTRINSIC_ROUND: 482 return selectExtInst(ResVReg, ResType, I, CL::round, GL::Round); 483 case TargetOpcode::G_INTRINSIC_ROUNDEVEN: 484 return selectExtInst(ResVReg, ResType, I, CL::rint, GL::RoundEven); 485 case TargetOpcode::G_INTRINSIC_TRUNC: 486 return selectExtInst(ResVReg, ResType, I, CL::trunc, GL::Trunc); 487 case TargetOpcode::G_FRINT: 488 case TargetOpcode::G_FNEARBYINT: 489 return selectExtInst(ResVReg, ResType, I, CL::rint, GL::RoundEven); 490 491 case TargetOpcode::G_SMULH: 492 return selectExtInst(ResVReg, ResType, I, CL::s_mul_hi); 493 case TargetOpcode::G_UMULH: 494 return selectExtInst(ResVReg, ResType, I, CL::u_mul_hi); 495 496 case TargetOpcode::G_SEXT: 497 return selectExt(ResVReg, ResType, I, true); 498 case TargetOpcode::G_ANYEXT: 499 case TargetOpcode::G_ZEXT: 500 return selectExt(ResVReg, ResType, I, false); 501 case TargetOpcode::G_TRUNC: 502 return selectTrunc(ResVReg, ResType, I); 503 case TargetOpcode::G_FPTRUNC: 504 case TargetOpcode::G_FPEXT: 505 return selectUnOp(ResVReg, ResType, I, SPIRV::OpFConvert); 506 507 case TargetOpcode::G_PTRTOINT: 508 return selectUnOp(ResVReg, ResType, I, SPIRV::OpConvertPtrToU); 509 case TargetOpcode::G_INTTOPTR: 510 return selectUnOp(ResVReg, ResType, I, SPIRV::OpConvertUToPtr); 511 case TargetOpcode::G_BITCAST: 512 return selectBitcast(ResVReg, ResType, I); 513 case TargetOpcode::G_ADDRSPACE_CAST: 514 return selectAddrSpaceCast(ResVReg, ResType, I); 515 case TargetOpcode::G_PTR_ADD: { 516 // Currently, we get G_PTR_ADD only as a result of translating 517 // global variables, initialized with constant expressions like GV + Const 518 // (see test opencl/basic/progvar_prog_scope_init.ll). 519 // TODO: extend the handler once we have other cases. 520 assert(I.getOperand(1).isReg() && I.getOperand(2).isReg()); 521 Register GV = I.getOperand(1).getReg(); 522 MachineRegisterInfo::def_instr_iterator II = MRI->def_instr_begin(GV); 523 (void)II; 524 assert(((*II).getOpcode() == TargetOpcode::G_GLOBAL_VALUE || 525 (*II).getOpcode() == TargetOpcode::COPY || 526 (*II).getOpcode() == SPIRV::OpVariable) && 527 isImm(I.getOperand(2), MRI)); 528 Register Idx = buildZerosVal(GR.getOrCreateSPIRVIntegerType(32, I, TII), I); 529 MachineBasicBlock &BB = *I.getParent(); 530 auto MIB = BuildMI(BB, I, I.getDebugLoc(), TII.get(SPIRV::OpSpecConstantOp)) 531 .addDef(ResVReg) 532 .addUse(GR.getSPIRVTypeID(ResType)) 533 .addImm(static_cast<uint32_t>( 534 SPIRV::Opcode::InBoundsPtrAccessChain)) 535 .addUse(GV) 536 .addUse(Idx) 537 .addUse(I.getOperand(2).getReg()); 538 return MIB.constrainAllUses(TII, TRI, RBI); 539 } 540 541 case TargetOpcode::G_ATOMICRMW_OR: 542 return selectAtomicRMW(ResVReg, ResType, I, SPIRV::OpAtomicOr); 543 case TargetOpcode::G_ATOMICRMW_ADD: 544 return selectAtomicRMW(ResVReg, ResType, I, SPIRV::OpAtomicIAdd); 545 case TargetOpcode::G_ATOMICRMW_AND: 546 return selectAtomicRMW(ResVReg, ResType, I, SPIRV::OpAtomicAnd); 547 case TargetOpcode::G_ATOMICRMW_MAX: 548 return selectAtomicRMW(ResVReg, ResType, I, SPIRV::OpAtomicSMax); 549 case TargetOpcode::G_ATOMICRMW_MIN: 550 return selectAtomicRMW(ResVReg, ResType, I, SPIRV::OpAtomicSMin); 551 case TargetOpcode::G_ATOMICRMW_SUB: 552 return selectAtomicRMW(ResVReg, ResType, I, SPIRV::OpAtomicISub); 553 case TargetOpcode::G_ATOMICRMW_XOR: 554 return selectAtomicRMW(ResVReg, ResType, I, SPIRV::OpAtomicXor); 555 case TargetOpcode::G_ATOMICRMW_UMAX: 556 return selectAtomicRMW(ResVReg, ResType, I, SPIRV::OpAtomicUMax); 557 case TargetOpcode::G_ATOMICRMW_UMIN: 558 return selectAtomicRMW(ResVReg, ResType, I, SPIRV::OpAtomicUMin); 559 case TargetOpcode::G_ATOMICRMW_XCHG: 560 return selectAtomicRMW(ResVReg, ResType, I, SPIRV::OpAtomicExchange); 561 case TargetOpcode::G_ATOMIC_CMPXCHG: 562 return selectAtomicCmpXchg(ResVReg, ResType, I); 563 564 case TargetOpcode::G_ATOMICRMW_FADD: 565 return selectAtomicRMW(ResVReg, ResType, I, SPIRV::OpAtomicFAddEXT); 566 case TargetOpcode::G_ATOMICRMW_FSUB: 567 // Translate G_ATOMICRMW_FSUB to OpAtomicFAddEXT with negative value operand 568 return selectAtomicRMW(ResVReg, ResType, I, SPIRV::OpAtomicFAddEXT, 569 SPIRV::OpFNegate); 570 case TargetOpcode::G_ATOMICRMW_FMIN: 571 return selectAtomicRMW(ResVReg, ResType, I, SPIRV::OpAtomicFMinEXT); 572 case TargetOpcode::G_ATOMICRMW_FMAX: 573 return selectAtomicRMW(ResVReg, ResType, I, SPIRV::OpAtomicFMaxEXT); 574 575 case TargetOpcode::G_FENCE: 576 return selectFence(I); 577 578 case TargetOpcode::G_STACKSAVE: 579 return selectStackSave(ResVReg, ResType, I); 580 case TargetOpcode::G_STACKRESTORE: 581 return selectStackRestore(I); 582 583 case TargetOpcode::G_UNMERGE_VALUES: 584 return selectUnmergeValues(I); 585 586 default: 587 return false; 588 } 589 } 590 591 bool SPIRVInstructionSelector::selectExtInst(Register ResVReg, 592 const SPIRVType *ResType, 593 MachineInstr &I, 594 CL::OpenCLExtInst CLInst) const { 595 return selectExtInst(ResVReg, ResType, I, 596 {{SPIRV::InstructionSet::OpenCL_std, CLInst}}); 597 } 598 599 bool SPIRVInstructionSelector::selectExtInst(Register ResVReg, 600 const SPIRVType *ResType, 601 MachineInstr &I, 602 CL::OpenCLExtInst CLInst, 603 GL::GLSLExtInst GLInst) const { 604 ExtInstList ExtInsts = {{SPIRV::InstructionSet::OpenCL_std, CLInst}, 605 {SPIRV::InstructionSet::GLSL_std_450, GLInst}}; 606 return selectExtInst(ResVReg, ResType, I, ExtInsts); 607 } 608 609 bool SPIRVInstructionSelector::selectExtInst(Register ResVReg, 610 const SPIRVType *ResType, 611 MachineInstr &I, 612 const ExtInstList &Insts) const { 613 614 for (const auto &Ex : Insts) { 615 SPIRV::InstructionSet::InstructionSet Set = Ex.first; 616 uint32_t Opcode = Ex.second; 617 if (STI.canUseExtInstSet(Set)) { 618 MachineBasicBlock &BB = *I.getParent(); 619 auto MIB = BuildMI(BB, I, I.getDebugLoc(), TII.get(SPIRV::OpExtInst)) 620 .addDef(ResVReg) 621 .addUse(GR.getSPIRVTypeID(ResType)) 622 .addImm(static_cast<uint32_t>(Set)) 623 .addImm(Opcode); 624 const unsigned NumOps = I.getNumOperands(); 625 for (unsigned i = 1; i < NumOps; ++i) 626 MIB.add(I.getOperand(i)); 627 return MIB.constrainAllUses(TII, TRI, RBI); 628 } 629 } 630 return false; 631 } 632 633 bool SPIRVInstructionSelector::selectUnOpWithSrc(Register ResVReg, 634 const SPIRVType *ResType, 635 MachineInstr &I, 636 Register SrcReg, 637 unsigned Opcode) const { 638 return BuildMI(*I.getParent(), I, I.getDebugLoc(), TII.get(Opcode)) 639 .addDef(ResVReg) 640 .addUse(GR.getSPIRVTypeID(ResType)) 641 .addUse(SrcReg) 642 .constrainAllUses(TII, TRI, RBI); 643 } 644 645 bool SPIRVInstructionSelector::selectUnOp(Register ResVReg, 646 const SPIRVType *ResType, 647 MachineInstr &I, 648 unsigned Opcode) const { 649 return selectUnOpWithSrc(ResVReg, ResType, I, I.getOperand(1).getReg(), 650 Opcode); 651 } 652 653 bool SPIRVInstructionSelector::selectBitcast(Register ResVReg, 654 const SPIRVType *ResType, 655 MachineInstr &I) const { 656 Register OpReg = I.getOperand(1).getReg(); 657 SPIRVType *OpType = OpReg.isValid() ? GR.getSPIRVTypeForVReg(OpReg) : nullptr; 658 if (!GR.isBitcastCompatible(ResType, OpType)) 659 report_fatal_error("incompatible result and operand types in a bitcast"); 660 return selectUnOp(ResVReg, ResType, I, SPIRV::OpBitcast); 661 } 662 663 static SPIRV::Scope::Scope getScope(SyncScope::ID Ord, 664 SPIRVMachineModuleInfo *MMI) { 665 if (Ord == SyncScope::SingleThread || Ord == MMI->Work_ItemSSID) 666 return SPIRV::Scope::Invocation; 667 else if (Ord == SyncScope::System || Ord == MMI->DeviceSSID) 668 return SPIRV::Scope::Device; 669 else if (Ord == MMI->WorkGroupSSID) 670 return SPIRV::Scope::Workgroup; 671 else if (Ord == MMI->AllSVMDevicesSSID) 672 return SPIRV::Scope::CrossDevice; 673 else if (Ord == MMI->SubGroupSSID) 674 return SPIRV::Scope::Subgroup; 675 else 676 // OpenCL approach is: "The functions that do not have memory_scope argument 677 // have the same semantics as the corresponding functions with the 678 // memory_scope argument set to memory_scope_device." See ref.: // 679 // https://registry.khronos.org/OpenCL/specs/3.0-unified/html/OpenCL_C.html#atomic-functions 680 // In our case if the scope is unknown, assuming that SPIR-V code is to be 681 // consumed in an OpenCL environment, we use the same approach and set the 682 // scope to memory_scope_device. 683 return SPIRV::Scope::Device; 684 } 685 686 static void addMemoryOperands(MachineMemOperand *MemOp, 687 MachineInstrBuilder &MIB) { 688 uint32_t SpvMemOp = static_cast<uint32_t>(SPIRV::MemoryOperand::None); 689 if (MemOp->isVolatile()) 690 SpvMemOp |= static_cast<uint32_t>(SPIRV::MemoryOperand::Volatile); 691 if (MemOp->isNonTemporal()) 692 SpvMemOp |= static_cast<uint32_t>(SPIRV::MemoryOperand::Nontemporal); 693 if (MemOp->getAlign().value()) 694 SpvMemOp |= static_cast<uint32_t>(SPIRV::MemoryOperand::Aligned); 695 696 if (SpvMemOp != static_cast<uint32_t>(SPIRV::MemoryOperand::None)) { 697 MIB.addImm(SpvMemOp); 698 if (SpvMemOp & static_cast<uint32_t>(SPIRV::MemoryOperand::Aligned)) 699 MIB.addImm(MemOp->getAlign().value()); 700 } 701 } 702 703 static void addMemoryOperands(uint64_t Flags, MachineInstrBuilder &MIB) { 704 uint32_t SpvMemOp = static_cast<uint32_t>(SPIRV::MemoryOperand::None); 705 if (Flags & MachineMemOperand::Flags::MOVolatile) 706 SpvMemOp |= static_cast<uint32_t>(SPIRV::MemoryOperand::Volatile); 707 if (Flags & MachineMemOperand::Flags::MONonTemporal) 708 SpvMemOp |= static_cast<uint32_t>(SPIRV::MemoryOperand::Nontemporal); 709 710 if (SpvMemOp != static_cast<uint32_t>(SPIRV::MemoryOperand::None)) 711 MIB.addImm(SpvMemOp); 712 } 713 714 bool SPIRVInstructionSelector::selectLoad(Register ResVReg, 715 const SPIRVType *ResType, 716 MachineInstr &I) const { 717 unsigned OpOffset = isa<GIntrinsic>(I) ? 1 : 0; 718 Register Ptr = I.getOperand(1 + OpOffset).getReg(); 719 auto MIB = BuildMI(*I.getParent(), I, I.getDebugLoc(), TII.get(SPIRV::OpLoad)) 720 .addDef(ResVReg) 721 .addUse(GR.getSPIRVTypeID(ResType)) 722 .addUse(Ptr); 723 if (!I.getNumMemOperands()) { 724 assert(I.getOpcode() == TargetOpcode::G_INTRINSIC_W_SIDE_EFFECTS || 725 I.getOpcode() == 726 TargetOpcode::G_INTRINSIC_CONVERGENT_W_SIDE_EFFECTS); 727 addMemoryOperands(I.getOperand(2 + OpOffset).getImm(), MIB); 728 } else { 729 addMemoryOperands(*I.memoperands_begin(), MIB); 730 } 731 return MIB.constrainAllUses(TII, TRI, RBI); 732 } 733 734 bool SPIRVInstructionSelector::selectStore(MachineInstr &I) const { 735 unsigned OpOffset = isa<GIntrinsic>(I) ? 1 : 0; 736 Register StoreVal = I.getOperand(0 + OpOffset).getReg(); 737 Register Ptr = I.getOperand(1 + OpOffset).getReg(); 738 MachineBasicBlock &BB = *I.getParent(); 739 auto MIB = BuildMI(BB, I, I.getDebugLoc(), TII.get(SPIRV::OpStore)) 740 .addUse(Ptr) 741 .addUse(StoreVal); 742 if (!I.getNumMemOperands()) { 743 assert(I.getOpcode() == TargetOpcode::G_INTRINSIC_W_SIDE_EFFECTS || 744 I.getOpcode() == 745 TargetOpcode::G_INTRINSIC_CONVERGENT_W_SIDE_EFFECTS); 746 addMemoryOperands(I.getOperand(2 + OpOffset).getImm(), MIB); 747 } else { 748 addMemoryOperands(*I.memoperands_begin(), MIB); 749 } 750 return MIB.constrainAllUses(TII, TRI, RBI); 751 } 752 753 bool SPIRVInstructionSelector::selectStackSave(Register ResVReg, 754 const SPIRVType *ResType, 755 MachineInstr &I) const { 756 if (!STI.canUseExtension(SPIRV::Extension::SPV_INTEL_variable_length_array)) 757 report_fatal_error( 758 "llvm.stacksave intrinsic: this instruction requires the following " 759 "SPIR-V extension: SPV_INTEL_variable_length_array", 760 false); 761 MachineBasicBlock &BB = *I.getParent(); 762 return BuildMI(BB, I, I.getDebugLoc(), TII.get(SPIRV::OpSaveMemoryINTEL)) 763 .addDef(ResVReg) 764 .addUse(GR.getSPIRVTypeID(ResType)) 765 .constrainAllUses(TII, TRI, RBI); 766 } 767 768 bool SPIRVInstructionSelector::selectStackRestore(MachineInstr &I) const { 769 if (!STI.canUseExtension(SPIRV::Extension::SPV_INTEL_variable_length_array)) 770 report_fatal_error( 771 "llvm.stackrestore intrinsic: this instruction requires the following " 772 "SPIR-V extension: SPV_INTEL_variable_length_array", 773 false); 774 if (!I.getOperand(0).isReg()) 775 return false; 776 MachineBasicBlock &BB = *I.getParent(); 777 return BuildMI(BB, I, I.getDebugLoc(), TII.get(SPIRV::OpRestoreMemoryINTEL)) 778 .addUse(I.getOperand(0).getReg()) 779 .constrainAllUses(TII, TRI, RBI); 780 } 781 782 bool SPIRVInstructionSelector::selectMemOperation(Register ResVReg, 783 MachineInstr &I) const { 784 MachineBasicBlock &BB = *I.getParent(); 785 Register SrcReg = I.getOperand(1).getReg(); 786 if (I.getOpcode() == TargetOpcode::G_MEMSET) { 787 assert(I.getOperand(1).isReg() && I.getOperand(2).isReg()); 788 unsigned Val = getIConstVal(I.getOperand(1).getReg(), MRI); 789 unsigned Num = getIConstVal(I.getOperand(2).getReg(), MRI); 790 SPIRVType *ValTy = GR.getOrCreateSPIRVIntegerType(8, I, TII); 791 SPIRVType *ArrTy = GR.getOrCreateSPIRVArrayType(ValTy, Num, I, TII); 792 Register Const = GR.getOrCreateConsIntArray(Val, I, ArrTy, TII); 793 SPIRVType *VarTy = GR.getOrCreateSPIRVPointerType( 794 ArrTy, I, TII, SPIRV::StorageClass::UniformConstant); 795 // TODO: check if we have such GV, add init, use buildGlobalVariable. 796 Function &CurFunction = GR.CurMF->getFunction(); 797 Type *LLVMArrTy = 798 ArrayType::get(IntegerType::get(CurFunction.getContext(), 8), Num); 799 // Module takes ownership of the global var. 800 GlobalVariable *GV = new GlobalVariable(*CurFunction.getParent(), LLVMArrTy, 801 true, GlobalValue::InternalLinkage, 802 Constant::getNullValue(LLVMArrTy)); 803 Register VarReg = MRI->createGenericVirtualRegister(LLT::scalar(32)); 804 GR.add(GV, GR.CurMF, VarReg); 805 806 buildOpDecorate(VarReg, I, TII, SPIRV::Decoration::Constant, {}); 807 BuildMI(*I.getParent(), I, I.getDebugLoc(), TII.get(SPIRV::OpVariable)) 808 .addDef(VarReg) 809 .addUse(GR.getSPIRVTypeID(VarTy)) 810 .addImm(SPIRV::StorageClass::UniformConstant) 811 .addUse(Const) 812 .constrainAllUses(TII, TRI, RBI); 813 SPIRVType *SourceTy = GR.getOrCreateSPIRVPointerType( 814 ValTy, I, TII, SPIRV::StorageClass::UniformConstant); 815 SrcReg = MRI->createGenericVirtualRegister(LLT::scalar(32)); 816 selectUnOpWithSrc(SrcReg, SourceTy, I, VarReg, SPIRV::OpBitcast); 817 } 818 auto MIB = BuildMI(BB, I, I.getDebugLoc(), TII.get(SPIRV::OpCopyMemorySized)) 819 .addUse(I.getOperand(0).getReg()) 820 .addUse(SrcReg) 821 .addUse(I.getOperand(2).getReg()); 822 if (I.getNumMemOperands()) 823 addMemoryOperands(*I.memoperands_begin(), MIB); 824 bool Result = MIB.constrainAllUses(TII, TRI, RBI); 825 if (ResVReg.isValid() && ResVReg != MIB->getOperand(0).getReg()) 826 BuildMI(BB, I, I.getDebugLoc(), TII.get(TargetOpcode::COPY), ResVReg) 827 .addUse(MIB->getOperand(0).getReg()); 828 return Result; 829 } 830 831 bool SPIRVInstructionSelector::selectAtomicRMW(Register ResVReg, 832 const SPIRVType *ResType, 833 MachineInstr &I, 834 unsigned NewOpcode, 835 unsigned NegateOpcode) const { 836 assert(I.hasOneMemOperand()); 837 const MachineMemOperand *MemOp = *I.memoperands_begin(); 838 uint32_t Scope = 839 static_cast<uint32_t>(getScope(MemOp->getSyncScopeID(), MMI)); 840 Register ScopeReg = buildI32Constant(Scope, I); 841 842 Register Ptr = I.getOperand(1).getReg(); 843 // TODO: Changed as it's implemented in the translator. See test/atomicrmw.ll 844 // auto ScSem = 845 // getMemSemanticsForStorageClass(GR.getPointerStorageClass(Ptr)); 846 AtomicOrdering AO = MemOp->getSuccessOrdering(); 847 uint32_t MemSem = static_cast<uint32_t>(getMemSemantics(AO)); 848 Register MemSemReg = buildI32Constant(MemSem /*| ScSem*/, I); 849 850 bool Result = false; 851 Register ValueReg = I.getOperand(2).getReg(); 852 if (NegateOpcode != 0) { 853 // Translation with negative value operand is requested 854 Register TmpReg = MRI->createVirtualRegister(&SPIRV::IDRegClass); 855 Result |= selectUnOpWithSrc(TmpReg, ResType, I, ValueReg, NegateOpcode); 856 ValueReg = TmpReg; 857 } 858 859 Result |= BuildMI(*I.getParent(), I, I.getDebugLoc(), TII.get(NewOpcode)) 860 .addDef(ResVReg) 861 .addUse(GR.getSPIRVTypeID(ResType)) 862 .addUse(Ptr) 863 .addUse(ScopeReg) 864 .addUse(MemSemReg) 865 .addUse(ValueReg) 866 .constrainAllUses(TII, TRI, RBI); 867 return Result; 868 } 869 870 bool SPIRVInstructionSelector::selectUnmergeValues(MachineInstr &I) const { 871 unsigned ArgI = I.getNumOperands() - 1; 872 Register SrcReg = 873 I.getOperand(ArgI).isReg() ? I.getOperand(ArgI).getReg() : Register(0); 874 SPIRVType *DefType = 875 SrcReg.isValid() ? GR.getSPIRVTypeForVReg(SrcReg) : nullptr; 876 if (!DefType || DefType->getOpcode() != SPIRV::OpTypeVector) 877 report_fatal_error( 878 "cannot select G_UNMERGE_VALUES with a non-vector argument"); 879 880 SPIRVType *ScalarType = 881 GR.getSPIRVTypeForVReg(DefType->getOperand(1).getReg()); 882 MachineBasicBlock &BB = *I.getParent(); 883 bool Res = false; 884 for (unsigned i = 0; i < I.getNumDefs(); ++i) { 885 Register ResVReg = I.getOperand(i).getReg(); 886 SPIRVType *ResType = GR.getSPIRVTypeForVReg(ResVReg); 887 if (!ResType) { 888 // There was no "assign type" actions, let's fix this now 889 ResType = ScalarType; 890 MRI->setRegClass(ResVReg, &SPIRV::IDRegClass); 891 MRI->setType(ResVReg, LLT::scalar(GR.getScalarOrVectorBitWidth(ResType))); 892 GR.assignSPIRVTypeToVReg(ResType, ResVReg, *GR.CurMF); 893 } 894 auto MIB = 895 BuildMI(BB, I, I.getDebugLoc(), TII.get(SPIRV::OpCompositeExtract)) 896 .addDef(ResVReg) 897 .addUse(GR.getSPIRVTypeID(ResType)) 898 .addUse(SrcReg) 899 .addImm(static_cast<int64_t>(i)); 900 Res |= MIB.constrainAllUses(TII, TRI, RBI); 901 } 902 return Res; 903 } 904 905 bool SPIRVInstructionSelector::selectFence(MachineInstr &I) const { 906 AtomicOrdering AO = AtomicOrdering(I.getOperand(0).getImm()); 907 uint32_t MemSem = static_cast<uint32_t>(getMemSemantics(AO)); 908 Register MemSemReg = buildI32Constant(MemSem, I); 909 SyncScope::ID Ord = SyncScope::ID(I.getOperand(1).getImm()); 910 uint32_t Scope = static_cast<uint32_t>(getScope(Ord, MMI)); 911 Register ScopeReg = buildI32Constant(Scope, I); 912 MachineBasicBlock &BB = *I.getParent(); 913 return BuildMI(BB, I, I.getDebugLoc(), TII.get(SPIRV::OpMemoryBarrier)) 914 .addUse(ScopeReg) 915 .addUse(MemSemReg) 916 .constrainAllUses(TII, TRI, RBI); 917 } 918 919 bool SPIRVInstructionSelector::selectAtomicCmpXchg(Register ResVReg, 920 const SPIRVType *ResType, 921 MachineInstr &I) const { 922 Register ScopeReg; 923 Register MemSemEqReg; 924 Register MemSemNeqReg; 925 Register Ptr = I.getOperand(2).getReg(); 926 if (!isa<GIntrinsic>(I)) { 927 assert(I.hasOneMemOperand()); 928 const MachineMemOperand *MemOp = *I.memoperands_begin(); 929 unsigned Scope = 930 static_cast<uint32_t>(getScope(MemOp->getSyncScopeID(), MMI)); 931 ScopeReg = buildI32Constant(Scope, I); 932 933 unsigned ScSem = static_cast<uint32_t>( 934 getMemSemanticsForStorageClass(GR.getPointerStorageClass(Ptr))); 935 AtomicOrdering AO = MemOp->getSuccessOrdering(); 936 unsigned MemSemEq = static_cast<uint32_t>(getMemSemantics(AO)) | ScSem; 937 MemSemEqReg = buildI32Constant(MemSemEq, I); 938 AtomicOrdering FO = MemOp->getFailureOrdering(); 939 unsigned MemSemNeq = static_cast<uint32_t>(getMemSemantics(FO)) | ScSem; 940 MemSemNeqReg = 941 MemSemEq == MemSemNeq ? MemSemEqReg : buildI32Constant(MemSemNeq, I); 942 } else { 943 ScopeReg = I.getOperand(5).getReg(); 944 MemSemEqReg = I.getOperand(6).getReg(); 945 MemSemNeqReg = I.getOperand(7).getReg(); 946 } 947 948 Register Cmp = I.getOperand(3).getReg(); 949 Register Val = I.getOperand(4).getReg(); 950 SPIRVType *SpvValTy = GR.getSPIRVTypeForVReg(Val); 951 Register ACmpRes = MRI->createVirtualRegister(&SPIRV::IDRegClass); 952 const DebugLoc &DL = I.getDebugLoc(); 953 bool Result = 954 BuildMI(*I.getParent(), I, DL, TII.get(SPIRV::OpAtomicCompareExchange)) 955 .addDef(ACmpRes) 956 .addUse(GR.getSPIRVTypeID(SpvValTy)) 957 .addUse(Ptr) 958 .addUse(ScopeReg) 959 .addUse(MemSemEqReg) 960 .addUse(MemSemNeqReg) 961 .addUse(Val) 962 .addUse(Cmp) 963 .constrainAllUses(TII, TRI, RBI); 964 Register CmpSuccReg = MRI->createVirtualRegister(&SPIRV::IDRegClass); 965 SPIRVType *BoolTy = GR.getOrCreateSPIRVBoolType(I, TII); 966 Result |= BuildMI(*I.getParent(), I, DL, TII.get(SPIRV::OpIEqual)) 967 .addDef(CmpSuccReg) 968 .addUse(GR.getSPIRVTypeID(BoolTy)) 969 .addUse(ACmpRes) 970 .addUse(Cmp) 971 .constrainAllUses(TII, TRI, RBI); 972 Register TmpReg = MRI->createVirtualRegister(&SPIRV::IDRegClass); 973 Result |= BuildMI(*I.getParent(), I, DL, TII.get(SPIRV::OpCompositeInsert)) 974 .addDef(TmpReg) 975 .addUse(GR.getSPIRVTypeID(ResType)) 976 .addUse(ACmpRes) 977 .addUse(GR.getOrCreateUndef(I, ResType, TII)) 978 .addImm(0) 979 .constrainAllUses(TII, TRI, RBI); 980 Result |= BuildMI(*I.getParent(), I, DL, TII.get(SPIRV::OpCompositeInsert)) 981 .addDef(ResVReg) 982 .addUse(GR.getSPIRVTypeID(ResType)) 983 .addUse(CmpSuccReg) 984 .addUse(TmpReg) 985 .addImm(1) 986 .constrainAllUses(TII, TRI, RBI); 987 return Result; 988 } 989 990 static bool isGenericCastablePtr(SPIRV::StorageClass::StorageClass SC) { 991 switch (SC) { 992 case SPIRV::StorageClass::Workgroup: 993 case SPIRV::StorageClass::CrossWorkgroup: 994 case SPIRV::StorageClass::Function: 995 return true; 996 default: 997 return false; 998 } 999 } 1000 1001 static bool isUSMStorageClass(SPIRV::StorageClass::StorageClass SC) { 1002 switch (SC) { 1003 case SPIRV::StorageClass::DeviceOnlyINTEL: 1004 case SPIRV::StorageClass::HostOnlyINTEL: 1005 return true; 1006 default: 1007 return false; 1008 } 1009 } 1010 1011 // In SPIR-V address space casting can only happen to and from the Generic 1012 // storage class. We can also only cast Workgroup, CrossWorkgroup, or Function 1013 // pointers to and from Generic pointers. As such, we can convert e.g. from 1014 // Workgroup to Function by going via a Generic pointer as an intermediary. All 1015 // other combinations can only be done by a bitcast, and are probably not safe. 1016 bool SPIRVInstructionSelector::selectAddrSpaceCast(Register ResVReg, 1017 const SPIRVType *ResType, 1018 MachineInstr &I) const { 1019 // If the AddrSpaceCast user is single and in OpConstantComposite or 1020 // OpVariable, we should select OpSpecConstantOp. 1021 auto UIs = MRI->use_instructions(ResVReg); 1022 if (!UIs.empty() && ++UIs.begin() == UIs.end() && 1023 (UIs.begin()->getOpcode() == SPIRV::OpConstantComposite || 1024 UIs.begin()->getOpcode() == SPIRV::OpVariable || 1025 isSpvIntrinsic(*UIs.begin(), Intrinsic::spv_init_global))) { 1026 Register NewReg = I.getOperand(1).getReg(); 1027 MachineBasicBlock &BB = *I.getParent(); 1028 SPIRVType *SpvBaseTy = GR.getOrCreateSPIRVIntegerType(8, I, TII); 1029 ResType = GR.getOrCreateSPIRVPointerType(SpvBaseTy, I, TII, 1030 SPIRV::StorageClass::Generic); 1031 bool Result = 1032 BuildMI(BB, I, I.getDebugLoc(), TII.get(SPIRV::OpSpecConstantOp)) 1033 .addDef(ResVReg) 1034 .addUse(GR.getSPIRVTypeID(ResType)) 1035 .addImm(static_cast<uint32_t>(SPIRV::Opcode::PtrCastToGeneric)) 1036 .addUse(NewReg) 1037 .constrainAllUses(TII, TRI, RBI); 1038 return Result; 1039 } 1040 Register SrcPtr = I.getOperand(1).getReg(); 1041 SPIRVType *SrcPtrTy = GR.getSPIRVTypeForVReg(SrcPtr); 1042 SPIRV::StorageClass::StorageClass SrcSC = GR.getPointerStorageClass(SrcPtr); 1043 SPIRV::StorageClass::StorageClass DstSC = GR.getPointerStorageClass(ResVReg); 1044 1045 // don't generate a cast between identical storage classes 1046 if (SrcSC == DstSC) 1047 return true; 1048 1049 // Casting from an eligible pointer to Generic. 1050 if (DstSC == SPIRV::StorageClass::Generic && isGenericCastablePtr(SrcSC)) 1051 return selectUnOp(ResVReg, ResType, I, SPIRV::OpPtrCastToGeneric); 1052 // Casting from Generic to an eligible pointer. 1053 if (SrcSC == SPIRV::StorageClass::Generic && isGenericCastablePtr(DstSC)) 1054 return selectUnOp(ResVReg, ResType, I, SPIRV::OpGenericCastToPtr); 1055 // Casting between 2 eligible pointers using Generic as an intermediary. 1056 if (isGenericCastablePtr(SrcSC) && isGenericCastablePtr(DstSC)) { 1057 Register Tmp = MRI->createVirtualRegister(&SPIRV::IDRegClass); 1058 SPIRVType *GenericPtrTy = GR.getOrCreateSPIRVPointerType( 1059 SrcPtrTy, I, TII, SPIRV::StorageClass::Generic); 1060 MachineBasicBlock &BB = *I.getParent(); 1061 const DebugLoc &DL = I.getDebugLoc(); 1062 bool Success = BuildMI(BB, I, DL, TII.get(SPIRV::OpPtrCastToGeneric)) 1063 .addDef(Tmp) 1064 .addUse(GR.getSPIRVTypeID(GenericPtrTy)) 1065 .addUse(SrcPtr) 1066 .constrainAllUses(TII, TRI, RBI); 1067 return Success && BuildMI(BB, I, DL, TII.get(SPIRV::OpGenericCastToPtr)) 1068 .addDef(ResVReg) 1069 .addUse(GR.getSPIRVTypeID(ResType)) 1070 .addUse(Tmp) 1071 .constrainAllUses(TII, TRI, RBI); 1072 } 1073 1074 // Check if instructions from the SPV_INTEL_usm_storage_classes extension may 1075 // be applied 1076 if (isUSMStorageClass(SrcSC) && DstSC == SPIRV::StorageClass::CrossWorkgroup) 1077 return selectUnOp(ResVReg, ResType, I, 1078 SPIRV::OpPtrCastToCrossWorkgroupINTEL); 1079 if (SrcSC == SPIRV::StorageClass::CrossWorkgroup && isUSMStorageClass(DstSC)) 1080 return selectUnOp(ResVReg, ResType, I, 1081 SPIRV::OpCrossWorkgroupCastToPtrINTEL); 1082 1083 // TODO Should this case just be disallowed completely? 1084 // We're casting 2 other arbitrary address spaces, so have to bitcast. 1085 return selectUnOp(ResVReg, ResType, I, SPIRV::OpBitcast); 1086 } 1087 1088 static unsigned getFCmpOpcode(unsigned PredNum) { 1089 auto Pred = static_cast<CmpInst::Predicate>(PredNum); 1090 switch (Pred) { 1091 case CmpInst::FCMP_OEQ: 1092 return SPIRV::OpFOrdEqual; 1093 case CmpInst::FCMP_OGE: 1094 return SPIRV::OpFOrdGreaterThanEqual; 1095 case CmpInst::FCMP_OGT: 1096 return SPIRV::OpFOrdGreaterThan; 1097 case CmpInst::FCMP_OLE: 1098 return SPIRV::OpFOrdLessThanEqual; 1099 case CmpInst::FCMP_OLT: 1100 return SPIRV::OpFOrdLessThan; 1101 case CmpInst::FCMP_ONE: 1102 return SPIRV::OpFOrdNotEqual; 1103 case CmpInst::FCMP_ORD: 1104 return SPIRV::OpOrdered; 1105 case CmpInst::FCMP_UEQ: 1106 return SPIRV::OpFUnordEqual; 1107 case CmpInst::FCMP_UGE: 1108 return SPIRV::OpFUnordGreaterThanEqual; 1109 case CmpInst::FCMP_UGT: 1110 return SPIRV::OpFUnordGreaterThan; 1111 case CmpInst::FCMP_ULE: 1112 return SPIRV::OpFUnordLessThanEqual; 1113 case CmpInst::FCMP_ULT: 1114 return SPIRV::OpFUnordLessThan; 1115 case CmpInst::FCMP_UNE: 1116 return SPIRV::OpFUnordNotEqual; 1117 case CmpInst::FCMP_UNO: 1118 return SPIRV::OpUnordered; 1119 default: 1120 llvm_unreachable("Unknown predicate type for FCmp"); 1121 } 1122 } 1123 1124 static unsigned getICmpOpcode(unsigned PredNum) { 1125 auto Pred = static_cast<CmpInst::Predicate>(PredNum); 1126 switch (Pred) { 1127 case CmpInst::ICMP_EQ: 1128 return SPIRV::OpIEqual; 1129 case CmpInst::ICMP_NE: 1130 return SPIRV::OpINotEqual; 1131 case CmpInst::ICMP_SGE: 1132 return SPIRV::OpSGreaterThanEqual; 1133 case CmpInst::ICMP_SGT: 1134 return SPIRV::OpSGreaterThan; 1135 case CmpInst::ICMP_SLE: 1136 return SPIRV::OpSLessThanEqual; 1137 case CmpInst::ICMP_SLT: 1138 return SPIRV::OpSLessThan; 1139 case CmpInst::ICMP_UGE: 1140 return SPIRV::OpUGreaterThanEqual; 1141 case CmpInst::ICMP_UGT: 1142 return SPIRV::OpUGreaterThan; 1143 case CmpInst::ICMP_ULE: 1144 return SPIRV::OpULessThanEqual; 1145 case CmpInst::ICMP_ULT: 1146 return SPIRV::OpULessThan; 1147 default: 1148 llvm_unreachable("Unknown predicate type for ICmp"); 1149 } 1150 } 1151 1152 static unsigned getPtrCmpOpcode(unsigned Pred) { 1153 switch (static_cast<CmpInst::Predicate>(Pred)) { 1154 case CmpInst::ICMP_EQ: 1155 return SPIRV::OpPtrEqual; 1156 case CmpInst::ICMP_NE: 1157 return SPIRV::OpPtrNotEqual; 1158 default: 1159 llvm_unreachable("Unknown predicate type for pointer comparison"); 1160 } 1161 } 1162 1163 // Return the logical operation, or abort if none exists. 1164 static unsigned getBoolCmpOpcode(unsigned PredNum) { 1165 auto Pred = static_cast<CmpInst::Predicate>(PredNum); 1166 switch (Pred) { 1167 case CmpInst::ICMP_EQ: 1168 return SPIRV::OpLogicalEqual; 1169 case CmpInst::ICMP_NE: 1170 return SPIRV::OpLogicalNotEqual; 1171 default: 1172 llvm_unreachable("Unknown predicate type for Bool comparison"); 1173 } 1174 } 1175 1176 bool SPIRVInstructionSelector::selectAnyOrAll(Register ResVReg, 1177 const SPIRVType *ResType, 1178 MachineInstr &I, 1179 unsigned OpAnyOrAll) const { 1180 assert(I.getNumOperands() == 3); 1181 assert(I.getOperand(2).isReg()); 1182 MachineBasicBlock &BB = *I.getParent(); 1183 Register InputRegister = I.getOperand(2).getReg(); 1184 SPIRVType *InputType = GR.getSPIRVTypeForVReg(InputRegister); 1185 1186 if (!InputType) 1187 report_fatal_error("Input Type could not be determined."); 1188 1189 bool IsBoolTy = GR.isScalarOrVectorOfType(InputRegister, SPIRV::OpTypeBool); 1190 bool IsVectorTy = InputType->getOpcode() == SPIRV::OpTypeVector; 1191 if (IsBoolTy && !IsVectorTy) { 1192 assert(ResVReg == I.getOperand(0).getReg()); 1193 return BuildMI(*I.getParent(), I, I.getDebugLoc(), 1194 TII.get(TargetOpcode::COPY)) 1195 .addDef(ResVReg) 1196 .addUse(InputRegister) 1197 .constrainAllUses(TII, TRI, RBI); 1198 } 1199 1200 bool IsFloatTy = GR.isScalarOrVectorOfType(InputRegister, SPIRV::OpTypeFloat); 1201 unsigned SpirvNotEqualId = 1202 IsFloatTy ? SPIRV::OpFOrdNotEqual : SPIRV::OpINotEqual; 1203 SPIRVType *SpvBoolScalarTy = GR.getOrCreateSPIRVBoolType(I, TII); 1204 SPIRVType *SpvBoolTy = SpvBoolScalarTy; 1205 Register NotEqualReg = ResVReg; 1206 1207 if (IsVectorTy) { 1208 NotEqualReg = IsBoolTy ? InputRegister 1209 : MRI->createVirtualRegister(&SPIRV::IDRegClass); 1210 const unsigned NumElts = InputType->getOperand(2).getImm(); 1211 SpvBoolTy = GR.getOrCreateSPIRVVectorType(SpvBoolTy, NumElts, I, TII); 1212 } 1213 1214 if (!IsBoolTy) { 1215 Register ConstZeroReg = 1216 IsFloatTy ? buildZerosValF(InputType, I) : buildZerosVal(InputType, I); 1217 1218 BuildMI(BB, I, I.getDebugLoc(), TII.get(SpirvNotEqualId)) 1219 .addDef(NotEqualReg) 1220 .addUse(GR.getSPIRVTypeID(SpvBoolTy)) 1221 .addUse(InputRegister) 1222 .addUse(ConstZeroReg) 1223 .constrainAllUses(TII, TRI, RBI); 1224 } 1225 1226 if (!IsVectorTy) 1227 return true; 1228 1229 return BuildMI(BB, I, I.getDebugLoc(), TII.get(OpAnyOrAll)) 1230 .addDef(ResVReg) 1231 .addUse(GR.getSPIRVTypeID(SpvBoolScalarTy)) 1232 .addUse(NotEqualReg) 1233 .constrainAllUses(TII, TRI, RBI); 1234 } 1235 1236 bool SPIRVInstructionSelector::selectAll(Register ResVReg, 1237 const SPIRVType *ResType, 1238 MachineInstr &I) const { 1239 return selectAnyOrAll(ResVReg, ResType, I, SPIRV::OpAll); 1240 } 1241 1242 bool SPIRVInstructionSelector::selectAny(Register ResVReg, 1243 const SPIRVType *ResType, 1244 MachineInstr &I) const { 1245 return selectAnyOrAll(ResVReg, ResType, I, SPIRV::OpAny); 1246 } 1247 1248 bool SPIRVInstructionSelector::selectFmix(Register ResVReg, 1249 const SPIRVType *ResType, 1250 MachineInstr &I) const { 1251 1252 assert(I.getNumOperands() == 5); 1253 assert(I.getOperand(2).isReg()); 1254 assert(I.getOperand(3).isReg()); 1255 assert(I.getOperand(4).isReg()); 1256 MachineBasicBlock &BB = *I.getParent(); 1257 1258 return BuildMI(BB, I, I.getDebugLoc(), TII.get(SPIRV::OpExtInst)) 1259 .addDef(ResVReg) 1260 .addUse(GR.getSPIRVTypeID(ResType)) 1261 .addImm(static_cast<uint32_t>(SPIRV::InstructionSet::GLSL_std_450)) 1262 .addImm(GL::FMix) 1263 .addUse(I.getOperand(2).getReg()) 1264 .addUse(I.getOperand(3).getReg()) 1265 .addUse(I.getOperand(4).getReg()) 1266 .constrainAllUses(TII, TRI, RBI); 1267 } 1268 1269 bool SPIRVInstructionSelector::selectBitreverse(Register ResVReg, 1270 const SPIRVType *ResType, 1271 MachineInstr &I) const { 1272 MachineBasicBlock &BB = *I.getParent(); 1273 return BuildMI(BB, I, I.getDebugLoc(), TII.get(SPIRV::OpBitReverse)) 1274 .addDef(ResVReg) 1275 .addUse(GR.getSPIRVTypeID(ResType)) 1276 .addUse(I.getOperand(1).getReg()) 1277 .constrainAllUses(TII, TRI, RBI); 1278 } 1279 1280 bool SPIRVInstructionSelector::selectFreeze(Register ResVReg, 1281 const SPIRVType *ResType, 1282 MachineInstr &I) const { 1283 // There is no way to implement `freeze` correctly without support on SPIR-V 1284 // standard side, but we may at least address a simple (static) case when 1285 // undef/poison value presence is obvious. The main benefit of even 1286 // incomplete `freeze` support is preventing of translation from crashing due 1287 // to lack of support on legalization and instruction selection steps. 1288 if (!I.getOperand(0).isReg() || !I.getOperand(1).isReg()) 1289 return false; 1290 Register OpReg = I.getOperand(1).getReg(); 1291 if (MachineInstr *Def = MRI->getVRegDef(OpReg)) { 1292 Register Reg; 1293 switch (Def->getOpcode()) { 1294 case SPIRV::ASSIGN_TYPE: 1295 if (MachineInstr *AssignToDef = 1296 MRI->getVRegDef(Def->getOperand(1).getReg())) { 1297 if (AssignToDef->getOpcode() == TargetOpcode::G_IMPLICIT_DEF) 1298 Reg = Def->getOperand(2).getReg(); 1299 } 1300 break; 1301 case SPIRV::OpUndef: 1302 Reg = Def->getOperand(1).getReg(); 1303 break; 1304 } 1305 unsigned DestOpCode; 1306 if (Reg.isValid()) { 1307 DestOpCode = SPIRV::OpConstantNull; 1308 } else { 1309 DestOpCode = TargetOpcode::COPY; 1310 Reg = OpReg; 1311 } 1312 return BuildMI(*I.getParent(), I, I.getDebugLoc(), TII.get(DestOpCode)) 1313 .addDef(I.getOperand(0).getReg()) 1314 .addUse(Reg) 1315 .constrainAllUses(TII, TRI, RBI); 1316 } 1317 return false; 1318 } 1319 1320 bool SPIRVInstructionSelector::selectConstVector(Register ResVReg, 1321 const SPIRVType *ResType, 1322 MachineInstr &I) const { 1323 // TODO: only const case is supported for now. 1324 assert(std::all_of( 1325 I.operands_begin(), I.operands_end(), [this](const MachineOperand &MO) { 1326 if (MO.isDef()) 1327 return true; 1328 if (!MO.isReg()) 1329 return false; 1330 SPIRVType *ConstTy = this->MRI->getVRegDef(MO.getReg()); 1331 assert(ConstTy && ConstTy->getOpcode() == SPIRV::ASSIGN_TYPE && 1332 ConstTy->getOperand(1).isReg()); 1333 Register ConstReg = ConstTy->getOperand(1).getReg(); 1334 const MachineInstr *Const = this->MRI->getVRegDef(ConstReg); 1335 assert(Const); 1336 return (Const->getOpcode() == TargetOpcode::G_CONSTANT || 1337 Const->getOpcode() == TargetOpcode::G_FCONSTANT); 1338 })); 1339 1340 auto MIB = BuildMI(*I.getParent(), I, I.getDebugLoc(), 1341 TII.get(SPIRV::OpConstantComposite)) 1342 .addDef(ResVReg) 1343 .addUse(GR.getSPIRVTypeID(ResType)); 1344 for (unsigned i = I.getNumExplicitDefs(); i < I.getNumExplicitOperands(); ++i) 1345 MIB.addUse(I.getOperand(i).getReg()); 1346 return MIB.constrainAllUses(TII, TRI, RBI); 1347 } 1348 1349 static unsigned getArrayComponentCount(MachineRegisterInfo *MRI, 1350 const SPIRVType *ResType) { 1351 Register OpReg = ResType->getOperand(2).getReg(); 1352 SPIRVType *OpDef = MRI->getVRegDef(OpReg); 1353 if (!OpDef) 1354 return 0; 1355 if (OpDef->getOpcode() == SPIRV::ASSIGN_TYPE && 1356 OpDef->getOperand(1).isReg()) { 1357 if (SPIRVType *RefDef = MRI->getVRegDef(OpDef->getOperand(1).getReg())) 1358 OpDef = RefDef; 1359 } 1360 unsigned N = OpDef->getOpcode() == TargetOpcode::G_CONSTANT 1361 ? OpDef->getOperand(1).getCImm()->getValue().getZExtValue() 1362 : 0; 1363 return N; 1364 } 1365 1366 // Return true if the type represents a constant register 1367 static bool isConstReg(MachineRegisterInfo *MRI, SPIRVType *OpDef) { 1368 if (OpDef->getOpcode() == SPIRV::ASSIGN_TYPE && 1369 OpDef->getOperand(1).isReg()) { 1370 if (SPIRVType *RefDef = MRI->getVRegDef(OpDef->getOperand(1).getReg())) 1371 OpDef = RefDef; 1372 } 1373 return OpDef->getOpcode() == TargetOpcode::G_CONSTANT || 1374 OpDef->getOpcode() == TargetOpcode::G_FCONSTANT; 1375 } 1376 1377 // Return true if the virtual register represents a constant 1378 static bool isConstReg(MachineRegisterInfo *MRI, Register OpReg) { 1379 if (SPIRVType *OpDef = MRI->getVRegDef(OpReg)) 1380 return isConstReg(MRI, OpDef); 1381 return false; 1382 } 1383 1384 bool SPIRVInstructionSelector::selectSplatVector(Register ResVReg, 1385 const SPIRVType *ResType, 1386 MachineInstr &I) const { 1387 unsigned N = 0; 1388 if (ResType->getOpcode() == SPIRV::OpTypeVector) 1389 N = GR.getScalarOrVectorComponentCount(ResType); 1390 else if (ResType->getOpcode() == SPIRV::OpTypeArray) 1391 N = getArrayComponentCount(MRI, ResType); 1392 else 1393 report_fatal_error("Cannot select G_SPLAT_VECTOR with a non-vector result"); 1394 1395 unsigned OpIdx = I.getNumExplicitDefs(); 1396 if (!I.getOperand(OpIdx).isReg()) 1397 report_fatal_error("Unexpected argument in G_SPLAT_VECTOR"); 1398 1399 // check if we may construct a constant vector 1400 Register OpReg = I.getOperand(OpIdx).getReg(); 1401 bool IsConst = isConstReg(MRI, OpReg); 1402 1403 if (!IsConst && N < 2) 1404 report_fatal_error( 1405 "There must be at least two constituent operands in a vector"); 1406 1407 auto MIB = BuildMI(*I.getParent(), I, I.getDebugLoc(), 1408 TII.get(IsConst ? SPIRV::OpConstantComposite 1409 : SPIRV::OpCompositeConstruct)) 1410 .addDef(ResVReg) 1411 .addUse(GR.getSPIRVTypeID(ResType)); 1412 for (unsigned i = 0; i < N; ++i) 1413 MIB.addUse(OpReg); 1414 return MIB.constrainAllUses(TII, TRI, RBI); 1415 } 1416 1417 bool SPIRVInstructionSelector::selectCmp(Register ResVReg, 1418 const SPIRVType *ResType, 1419 unsigned CmpOpc, 1420 MachineInstr &I) const { 1421 Register Cmp0 = I.getOperand(2).getReg(); 1422 Register Cmp1 = I.getOperand(3).getReg(); 1423 assert(GR.getSPIRVTypeForVReg(Cmp0)->getOpcode() == 1424 GR.getSPIRVTypeForVReg(Cmp1)->getOpcode() && 1425 "CMP operands should have the same type"); 1426 return BuildMI(*I.getParent(), I, I.getDebugLoc(), TII.get(CmpOpc)) 1427 .addDef(ResVReg) 1428 .addUse(GR.getSPIRVTypeID(ResType)) 1429 .addUse(Cmp0) 1430 .addUse(Cmp1) 1431 .constrainAllUses(TII, TRI, RBI); 1432 } 1433 1434 bool SPIRVInstructionSelector::selectICmp(Register ResVReg, 1435 const SPIRVType *ResType, 1436 MachineInstr &I) const { 1437 auto Pred = I.getOperand(1).getPredicate(); 1438 unsigned CmpOpc; 1439 1440 Register CmpOperand = I.getOperand(2).getReg(); 1441 if (GR.isScalarOfType(CmpOperand, SPIRV::OpTypePointer)) 1442 CmpOpc = getPtrCmpOpcode(Pred); 1443 else if (GR.isScalarOrVectorOfType(CmpOperand, SPIRV::OpTypeBool)) 1444 CmpOpc = getBoolCmpOpcode(Pred); 1445 else 1446 CmpOpc = getICmpOpcode(Pred); 1447 return selectCmp(ResVReg, ResType, CmpOpc, I); 1448 } 1449 1450 void SPIRVInstructionSelector::renderFImm32(MachineInstrBuilder &MIB, 1451 const MachineInstr &I, 1452 int OpIdx) const { 1453 assert(I.getOpcode() == TargetOpcode::G_FCONSTANT && OpIdx == -1 && 1454 "Expected G_FCONSTANT"); 1455 const ConstantFP *FPImm = I.getOperand(1).getFPImm(); 1456 addNumImm(FPImm->getValueAPF().bitcastToAPInt(), MIB); 1457 } 1458 1459 void SPIRVInstructionSelector::renderImm32(MachineInstrBuilder &MIB, 1460 const MachineInstr &I, 1461 int OpIdx) const { 1462 assert(I.getOpcode() == TargetOpcode::G_CONSTANT && OpIdx == -1 && 1463 "Expected G_CONSTANT"); 1464 addNumImm(I.getOperand(1).getCImm()->getValue(), MIB); 1465 } 1466 1467 Register 1468 SPIRVInstructionSelector::buildI32Constant(uint32_t Val, MachineInstr &I, 1469 const SPIRVType *ResType) const { 1470 Type *LLVMTy = IntegerType::get(GR.CurMF->getFunction().getContext(), 32); 1471 const SPIRVType *SpvI32Ty = 1472 ResType ? ResType : GR.getOrCreateSPIRVIntegerType(32, I, TII); 1473 // Find a constant in DT or build a new one. 1474 auto ConstInt = ConstantInt::get(LLVMTy, Val); 1475 Register NewReg = GR.find(ConstInt, GR.CurMF); 1476 if (!NewReg.isValid()) { 1477 NewReg = MRI->createGenericVirtualRegister(LLT::scalar(32)); 1478 GR.add(ConstInt, GR.CurMF, NewReg); 1479 MachineInstr *MI; 1480 MachineBasicBlock &BB = *I.getParent(); 1481 if (Val == 0) { 1482 MI = BuildMI(BB, I, I.getDebugLoc(), TII.get(SPIRV::OpConstantNull)) 1483 .addDef(NewReg) 1484 .addUse(GR.getSPIRVTypeID(SpvI32Ty)); 1485 } else { 1486 MI = BuildMI(BB, I, I.getDebugLoc(), TII.get(SPIRV::OpConstantI)) 1487 .addDef(NewReg) 1488 .addUse(GR.getSPIRVTypeID(SpvI32Ty)) 1489 .addImm(APInt(32, Val).getZExtValue()); 1490 } 1491 constrainSelectedInstRegOperands(*MI, TII, TRI, RBI); 1492 } 1493 return NewReg; 1494 } 1495 1496 bool SPIRVInstructionSelector::selectFCmp(Register ResVReg, 1497 const SPIRVType *ResType, 1498 MachineInstr &I) const { 1499 unsigned CmpOp = getFCmpOpcode(I.getOperand(1).getPredicate()); 1500 return selectCmp(ResVReg, ResType, CmpOp, I); 1501 } 1502 1503 Register SPIRVInstructionSelector::buildZerosVal(const SPIRVType *ResType, 1504 MachineInstr &I) const { 1505 // OpenCL uses nulls for Zero. In HLSL we don't use null constants. 1506 bool ZeroAsNull = STI.isOpenCLEnv(); 1507 if (ResType->getOpcode() == SPIRV::OpTypeVector) 1508 return GR.getOrCreateConstVector(0UL, I, ResType, TII, ZeroAsNull); 1509 return GR.getOrCreateConstInt(0, I, ResType, TII, ZeroAsNull); 1510 } 1511 1512 static APFloat getZeroFP(const Type *LLVMFloatTy) { 1513 if (!LLVMFloatTy) 1514 return APFloat::getZero(APFloat::IEEEsingle()); 1515 switch (LLVMFloatTy->getScalarType()->getTypeID()) { 1516 case Type::HalfTyID: 1517 return APFloat::getZero(APFloat::IEEEhalf()); 1518 default: 1519 case Type::FloatTyID: 1520 return APFloat::getZero(APFloat::IEEEsingle()); 1521 case Type::DoubleTyID: 1522 return APFloat::getZero(APFloat::IEEEdouble()); 1523 } 1524 } 1525 1526 Register SPIRVInstructionSelector::buildZerosValF(const SPIRVType *ResType, 1527 MachineInstr &I) const { 1528 // OpenCL uses nulls for Zero. In HLSL we don't use null constants. 1529 bool ZeroAsNull = STI.isOpenCLEnv(); 1530 APFloat VZero = getZeroFP(GR.getTypeForSPIRVType(ResType)); 1531 if (ResType->getOpcode() == SPIRV::OpTypeVector) 1532 return GR.getOrCreateConstVector(VZero, I, ResType, TII, ZeroAsNull); 1533 return GR.getOrCreateConstFP(VZero, I, ResType, TII, ZeroAsNull); 1534 } 1535 1536 Register SPIRVInstructionSelector::buildOnesVal(bool AllOnes, 1537 const SPIRVType *ResType, 1538 MachineInstr &I) const { 1539 unsigned BitWidth = GR.getScalarOrVectorBitWidth(ResType); 1540 APInt One = 1541 AllOnes ? APInt::getAllOnes(BitWidth) : APInt::getOneBitSet(BitWidth, 0); 1542 if (ResType->getOpcode() == SPIRV::OpTypeVector) 1543 return GR.getOrCreateConstVector(One.getZExtValue(), I, ResType, TII); 1544 return GR.getOrCreateConstInt(One.getZExtValue(), I, ResType, TII); 1545 } 1546 1547 bool SPIRVInstructionSelector::selectSelect(Register ResVReg, 1548 const SPIRVType *ResType, 1549 MachineInstr &I, 1550 bool IsSigned) const { 1551 // To extend a bool, we need to use OpSelect between constants. 1552 Register ZeroReg = buildZerosVal(ResType, I); 1553 Register OneReg = buildOnesVal(IsSigned, ResType, I); 1554 bool IsScalarBool = 1555 GR.isScalarOfType(I.getOperand(1).getReg(), SPIRV::OpTypeBool); 1556 unsigned Opcode = 1557 IsScalarBool ? SPIRV::OpSelectSISCond : SPIRV::OpSelectSIVCond; 1558 return BuildMI(*I.getParent(), I, I.getDebugLoc(), TII.get(Opcode)) 1559 .addDef(ResVReg) 1560 .addUse(GR.getSPIRVTypeID(ResType)) 1561 .addUse(I.getOperand(1).getReg()) 1562 .addUse(OneReg) 1563 .addUse(ZeroReg) 1564 .constrainAllUses(TII, TRI, RBI); 1565 } 1566 1567 bool SPIRVInstructionSelector::selectIToF(Register ResVReg, 1568 const SPIRVType *ResType, 1569 MachineInstr &I, bool IsSigned, 1570 unsigned Opcode) const { 1571 Register SrcReg = I.getOperand(1).getReg(); 1572 // We can convert bool value directly to float type without OpConvert*ToF, 1573 // however the translator generates OpSelect+OpConvert*ToF, so we do the same. 1574 if (GR.isScalarOrVectorOfType(I.getOperand(1).getReg(), SPIRV::OpTypeBool)) { 1575 unsigned BitWidth = GR.getScalarOrVectorBitWidth(ResType); 1576 SPIRVType *TmpType = GR.getOrCreateSPIRVIntegerType(BitWidth, I, TII); 1577 if (ResType->getOpcode() == SPIRV::OpTypeVector) { 1578 const unsigned NumElts = ResType->getOperand(2).getImm(); 1579 TmpType = GR.getOrCreateSPIRVVectorType(TmpType, NumElts, I, TII); 1580 } 1581 SrcReg = MRI->createVirtualRegister(&SPIRV::IDRegClass); 1582 selectSelect(SrcReg, TmpType, I, false); 1583 } 1584 return selectUnOpWithSrc(ResVReg, ResType, I, SrcReg, Opcode); 1585 } 1586 1587 bool SPIRVInstructionSelector::selectExt(Register ResVReg, 1588 const SPIRVType *ResType, 1589 MachineInstr &I, bool IsSigned) const { 1590 Register SrcReg = I.getOperand(1).getReg(); 1591 if (GR.isScalarOrVectorOfType(SrcReg, SPIRV::OpTypeBool)) 1592 return selectSelect(ResVReg, ResType, I, IsSigned); 1593 1594 SPIRVType *SrcType = GR.getSPIRVTypeForVReg(SrcReg); 1595 if (SrcType == ResType) 1596 return BuildMI(*I.getParent(), I, I.getDebugLoc(), 1597 TII.get(TargetOpcode::COPY)) 1598 .addDef(ResVReg) 1599 .addUse(SrcReg) 1600 .constrainAllUses(TII, TRI, RBI); 1601 1602 unsigned Opcode = IsSigned ? SPIRV::OpSConvert : SPIRV::OpUConvert; 1603 return selectUnOp(ResVReg, ResType, I, Opcode); 1604 } 1605 1606 bool SPIRVInstructionSelector::selectIntToBool(Register IntReg, 1607 Register ResVReg, 1608 MachineInstr &I, 1609 const SPIRVType *IntTy, 1610 const SPIRVType *BoolTy) const { 1611 // To truncate to a bool, we use OpBitwiseAnd 1 and OpINotEqual to zero. 1612 Register BitIntReg = MRI->createVirtualRegister(&SPIRV::IDRegClass); 1613 bool IsVectorTy = IntTy->getOpcode() == SPIRV::OpTypeVector; 1614 unsigned Opcode = IsVectorTy ? SPIRV::OpBitwiseAndV : SPIRV::OpBitwiseAndS; 1615 Register Zero = buildZerosVal(IntTy, I); 1616 Register One = buildOnesVal(false, IntTy, I); 1617 MachineBasicBlock &BB = *I.getParent(); 1618 BuildMI(BB, I, I.getDebugLoc(), TII.get(Opcode)) 1619 .addDef(BitIntReg) 1620 .addUse(GR.getSPIRVTypeID(IntTy)) 1621 .addUse(IntReg) 1622 .addUse(One) 1623 .constrainAllUses(TII, TRI, RBI); 1624 return BuildMI(BB, I, I.getDebugLoc(), TII.get(SPIRV::OpINotEqual)) 1625 .addDef(ResVReg) 1626 .addUse(GR.getSPIRVTypeID(BoolTy)) 1627 .addUse(BitIntReg) 1628 .addUse(Zero) 1629 .constrainAllUses(TII, TRI, RBI); 1630 } 1631 1632 bool SPIRVInstructionSelector::selectTrunc(Register ResVReg, 1633 const SPIRVType *ResType, 1634 MachineInstr &I) const { 1635 Register IntReg = I.getOperand(1).getReg(); 1636 const SPIRVType *ArgType = GR.getSPIRVTypeForVReg(IntReg); 1637 if (GR.isScalarOrVectorOfType(ResVReg, SPIRV::OpTypeBool)) 1638 return selectIntToBool(IntReg, ResVReg, I, ArgType, ResType); 1639 if (ArgType == ResType) 1640 return BuildMI(*I.getParent(), I, I.getDebugLoc(), 1641 TII.get(TargetOpcode::COPY)) 1642 .addDef(ResVReg) 1643 .addUse(IntReg) 1644 .constrainAllUses(TII, TRI, RBI); 1645 bool IsSigned = GR.isScalarOrVectorSigned(ResType); 1646 unsigned Opcode = IsSigned ? SPIRV::OpSConvert : SPIRV::OpUConvert; 1647 return selectUnOp(ResVReg, ResType, I, Opcode); 1648 } 1649 1650 bool SPIRVInstructionSelector::selectConst(Register ResVReg, 1651 const SPIRVType *ResType, 1652 const APInt &Imm, 1653 MachineInstr &I) const { 1654 unsigned TyOpcode = ResType->getOpcode(); 1655 assert(TyOpcode != SPIRV::OpTypePointer || Imm.isZero()); 1656 MachineBasicBlock &BB = *I.getParent(); 1657 if ((TyOpcode == SPIRV::OpTypePointer || TyOpcode == SPIRV::OpTypeEvent) && 1658 Imm.isZero()) 1659 return BuildMI(BB, I, I.getDebugLoc(), TII.get(SPIRV::OpConstantNull)) 1660 .addDef(ResVReg) 1661 .addUse(GR.getSPIRVTypeID(ResType)) 1662 .constrainAllUses(TII, TRI, RBI); 1663 if (TyOpcode == SPIRV::OpTypeInt) { 1664 assert(Imm.getBitWidth() <= 64 && "Unsupported integer width!"); 1665 Register Reg = GR.getOrCreateConstInt(Imm.getZExtValue(), I, ResType, TII); 1666 if (Reg == ResVReg) 1667 return true; 1668 return BuildMI(BB, I, I.getDebugLoc(), TII.get(TargetOpcode::COPY)) 1669 .addDef(ResVReg) 1670 .addUse(Reg) 1671 .constrainAllUses(TII, TRI, RBI); 1672 } 1673 auto MIB = BuildMI(BB, I, I.getDebugLoc(), TII.get(SPIRV::OpConstantI)) 1674 .addDef(ResVReg) 1675 .addUse(GR.getSPIRVTypeID(ResType)); 1676 // <=32-bit integers should be caught by the sdag pattern. 1677 assert(Imm.getBitWidth() > 32); 1678 addNumImm(Imm, MIB); 1679 return MIB.constrainAllUses(TII, TRI, RBI); 1680 } 1681 1682 bool SPIRVInstructionSelector::selectOpUndef(Register ResVReg, 1683 const SPIRVType *ResType, 1684 MachineInstr &I) const { 1685 return BuildMI(*I.getParent(), I, I.getDebugLoc(), TII.get(SPIRV::OpUndef)) 1686 .addDef(ResVReg) 1687 .addUse(GR.getSPIRVTypeID(ResType)) 1688 .constrainAllUses(TII, TRI, RBI); 1689 } 1690 1691 static bool isImm(const MachineOperand &MO, MachineRegisterInfo *MRI) { 1692 assert(MO.isReg()); 1693 const SPIRVType *TypeInst = MRI->getVRegDef(MO.getReg()); 1694 if (TypeInst->getOpcode() != SPIRV::ASSIGN_TYPE) 1695 return false; 1696 assert(TypeInst->getOperand(1).isReg()); 1697 MachineInstr *ImmInst = MRI->getVRegDef(TypeInst->getOperand(1).getReg()); 1698 return ImmInst->getOpcode() == TargetOpcode::G_CONSTANT; 1699 } 1700 1701 static int64_t foldImm(const MachineOperand &MO, MachineRegisterInfo *MRI) { 1702 const SPIRVType *TypeInst = MRI->getVRegDef(MO.getReg()); 1703 MachineInstr *ImmInst = MRI->getVRegDef(TypeInst->getOperand(1).getReg()); 1704 assert(ImmInst->getOpcode() == TargetOpcode::G_CONSTANT); 1705 return ImmInst->getOperand(1).getCImm()->getZExtValue(); 1706 } 1707 1708 bool SPIRVInstructionSelector::selectInsertVal(Register ResVReg, 1709 const SPIRVType *ResType, 1710 MachineInstr &I) const { 1711 MachineBasicBlock &BB = *I.getParent(); 1712 auto MIB = BuildMI(BB, I, I.getDebugLoc(), TII.get(SPIRV::OpCompositeInsert)) 1713 .addDef(ResVReg) 1714 .addUse(GR.getSPIRVTypeID(ResType)) 1715 // object to insert 1716 .addUse(I.getOperand(3).getReg()) 1717 // composite to insert into 1718 .addUse(I.getOperand(2).getReg()); 1719 for (unsigned i = 4; i < I.getNumOperands(); i++) 1720 MIB.addImm(foldImm(I.getOperand(i), MRI)); 1721 return MIB.constrainAllUses(TII, TRI, RBI); 1722 } 1723 1724 bool SPIRVInstructionSelector::selectExtractVal(Register ResVReg, 1725 const SPIRVType *ResType, 1726 MachineInstr &I) const { 1727 MachineBasicBlock &BB = *I.getParent(); 1728 auto MIB = BuildMI(BB, I, I.getDebugLoc(), TII.get(SPIRV::OpCompositeExtract)) 1729 .addDef(ResVReg) 1730 .addUse(GR.getSPIRVTypeID(ResType)) 1731 .addUse(I.getOperand(2).getReg()); 1732 for (unsigned i = 3; i < I.getNumOperands(); i++) 1733 MIB.addImm(foldImm(I.getOperand(i), MRI)); 1734 return MIB.constrainAllUses(TII, TRI, RBI); 1735 } 1736 1737 bool SPIRVInstructionSelector::selectInsertElt(Register ResVReg, 1738 const SPIRVType *ResType, 1739 MachineInstr &I) const { 1740 if (isImm(I.getOperand(4), MRI)) 1741 return selectInsertVal(ResVReg, ResType, I); 1742 MachineBasicBlock &BB = *I.getParent(); 1743 return BuildMI(BB, I, I.getDebugLoc(), TII.get(SPIRV::OpVectorInsertDynamic)) 1744 .addDef(ResVReg) 1745 .addUse(GR.getSPIRVTypeID(ResType)) 1746 .addUse(I.getOperand(2).getReg()) 1747 .addUse(I.getOperand(3).getReg()) 1748 .addUse(I.getOperand(4).getReg()) 1749 .constrainAllUses(TII, TRI, RBI); 1750 } 1751 1752 bool SPIRVInstructionSelector::selectExtractElt(Register ResVReg, 1753 const SPIRVType *ResType, 1754 MachineInstr &I) const { 1755 if (isImm(I.getOperand(3), MRI)) 1756 return selectExtractVal(ResVReg, ResType, I); 1757 MachineBasicBlock &BB = *I.getParent(); 1758 return BuildMI(BB, I, I.getDebugLoc(), TII.get(SPIRV::OpVectorExtractDynamic)) 1759 .addDef(ResVReg) 1760 .addUse(GR.getSPIRVTypeID(ResType)) 1761 .addUse(I.getOperand(2).getReg()) 1762 .addUse(I.getOperand(3).getReg()) 1763 .constrainAllUses(TII, TRI, RBI); 1764 } 1765 1766 bool SPIRVInstructionSelector::selectGEP(Register ResVReg, 1767 const SPIRVType *ResType, 1768 MachineInstr &I) const { 1769 const bool IsGEPInBounds = I.getOperand(2).getImm(); 1770 1771 // OpAccessChain could be used for OpenCL, but the SPIRV-LLVM Translator only 1772 // relies on PtrAccessChain, so we'll try not to deviate. For Vulkan however, 1773 // we have to use Op[InBounds]AccessChain. 1774 const unsigned Opcode = STI.isVulkanEnv() 1775 ? (IsGEPInBounds ? SPIRV::OpInBoundsAccessChain 1776 : SPIRV::OpAccessChain) 1777 : (IsGEPInBounds ? SPIRV::OpInBoundsPtrAccessChain 1778 : SPIRV::OpPtrAccessChain); 1779 1780 auto Res = BuildMI(*I.getParent(), I, I.getDebugLoc(), TII.get(Opcode)) 1781 .addDef(ResVReg) 1782 .addUse(GR.getSPIRVTypeID(ResType)) 1783 // Object to get a pointer to. 1784 .addUse(I.getOperand(3).getReg()); 1785 // Adding indices. 1786 const unsigned StartingIndex = 1787 (Opcode == SPIRV::OpAccessChain || Opcode == SPIRV::OpInBoundsAccessChain) 1788 ? 5 1789 : 4; 1790 for (unsigned i = StartingIndex; i < I.getNumExplicitOperands(); ++i) 1791 Res.addUse(I.getOperand(i).getReg()); 1792 return Res.constrainAllUses(TII, TRI, RBI); 1793 } 1794 1795 // Maybe wrap a value into OpSpecConstantOp 1796 bool SPIRVInstructionSelector::wrapIntoSpecConstantOp( 1797 MachineInstr &I, SmallVector<Register> &CompositeArgs) const { 1798 bool Result = true; 1799 unsigned Lim = I.getNumExplicitOperands(); 1800 for (unsigned i = I.getNumExplicitDefs() + 1; i < Lim; ++i) { 1801 Register OpReg = I.getOperand(i).getReg(); 1802 SPIRVType *OpDefine = MRI->getVRegDef(OpReg); 1803 SPIRVType *OpType = GR.getSPIRVTypeForVReg(OpReg); 1804 if (!OpDefine || !OpType || isConstReg(MRI, OpDefine) || 1805 OpDefine->getOpcode() == TargetOpcode::G_ADDRSPACE_CAST) { 1806 // The case of G_ADDRSPACE_CAST inside spv_const_composite() is processed 1807 // by selectAddrSpaceCast() 1808 CompositeArgs.push_back(OpReg); 1809 continue; 1810 } 1811 MachineFunction *MF = I.getMF(); 1812 Register WrapReg = GR.find(OpDefine, MF); 1813 if (WrapReg.isValid()) { 1814 CompositeArgs.push_back(WrapReg); 1815 continue; 1816 } 1817 // Create a new register for the wrapper 1818 WrapReg = MRI->createVirtualRegister(&SPIRV::IDRegClass); 1819 GR.add(OpDefine, MF, WrapReg); 1820 CompositeArgs.push_back(WrapReg); 1821 // Decorate the wrapper register and generate a new instruction 1822 MRI->setType(WrapReg, LLT::pointer(0, 32)); 1823 GR.assignSPIRVTypeToVReg(OpType, WrapReg, *MF); 1824 MachineBasicBlock &BB = *I.getParent(); 1825 Result = BuildMI(BB, I, I.getDebugLoc(), TII.get(SPIRV::OpSpecConstantOp)) 1826 .addDef(WrapReg) 1827 .addUse(GR.getSPIRVTypeID(OpType)) 1828 .addImm(static_cast<uint32_t>(SPIRV::Opcode::Bitcast)) 1829 .addUse(OpReg) 1830 .constrainAllUses(TII, TRI, RBI); 1831 if (!Result) 1832 break; 1833 } 1834 return Result; 1835 } 1836 1837 bool SPIRVInstructionSelector::selectIntrinsic(Register ResVReg, 1838 const SPIRVType *ResType, 1839 MachineInstr &I) const { 1840 MachineBasicBlock &BB = *I.getParent(); 1841 Intrinsic::ID IID = cast<GIntrinsic>(I).getIntrinsicID(); 1842 switch (IID) { 1843 case Intrinsic::spv_load: 1844 return selectLoad(ResVReg, ResType, I); 1845 case Intrinsic::spv_store: 1846 return selectStore(I); 1847 case Intrinsic::spv_extractv: 1848 return selectExtractVal(ResVReg, ResType, I); 1849 case Intrinsic::spv_insertv: 1850 return selectInsertVal(ResVReg, ResType, I); 1851 case Intrinsic::spv_extractelt: 1852 return selectExtractElt(ResVReg, ResType, I); 1853 case Intrinsic::spv_insertelt: 1854 return selectInsertElt(ResVReg, ResType, I); 1855 case Intrinsic::spv_gep: 1856 return selectGEP(ResVReg, ResType, I); 1857 case Intrinsic::spv_unref_global: 1858 case Intrinsic::spv_init_global: { 1859 MachineInstr *MI = MRI->getVRegDef(I.getOperand(1).getReg()); 1860 MachineInstr *Init = I.getNumExplicitOperands() > 2 1861 ? MRI->getVRegDef(I.getOperand(2).getReg()) 1862 : nullptr; 1863 assert(MI); 1864 return selectGlobalValue(MI->getOperand(0).getReg(), *MI, Init); 1865 } 1866 case Intrinsic::spv_undef: { 1867 auto MIB = BuildMI(BB, I, I.getDebugLoc(), TII.get(SPIRV::OpUndef)) 1868 .addDef(ResVReg) 1869 .addUse(GR.getSPIRVTypeID(ResType)); 1870 return MIB.constrainAllUses(TII, TRI, RBI); 1871 } 1872 case Intrinsic::spv_const_composite: { 1873 // If no values are attached, the composite is null constant. 1874 bool IsNull = I.getNumExplicitDefs() + 1 == I.getNumExplicitOperands(); 1875 // Select a proper instruction. 1876 unsigned Opcode = SPIRV::OpConstantNull; 1877 SmallVector<Register> CompositeArgs; 1878 if (!IsNull) { 1879 Opcode = SPIRV::OpConstantComposite; 1880 if (!wrapIntoSpecConstantOp(I, CompositeArgs)) 1881 return false; 1882 } 1883 auto MIB = BuildMI(BB, I, I.getDebugLoc(), TII.get(Opcode)) 1884 .addDef(ResVReg) 1885 .addUse(GR.getSPIRVTypeID(ResType)); 1886 // skip type MD node we already used when generated assign.type for this 1887 if (!IsNull) { 1888 for (Register OpReg : CompositeArgs) 1889 MIB.addUse(OpReg); 1890 } 1891 return MIB.constrainAllUses(TII, TRI, RBI); 1892 } 1893 case Intrinsic::spv_assign_name: { 1894 auto MIB = BuildMI(BB, I, I.getDebugLoc(), TII.get(SPIRV::OpName)); 1895 MIB.addUse(I.getOperand(I.getNumExplicitDefs() + 1).getReg()); 1896 for (unsigned i = I.getNumExplicitDefs() + 2; 1897 i < I.getNumExplicitOperands(); ++i) { 1898 MIB.addImm(I.getOperand(i).getImm()); 1899 } 1900 return MIB.constrainAllUses(TII, TRI, RBI); 1901 } 1902 case Intrinsic::spv_switch: { 1903 auto MIB = BuildMI(BB, I, I.getDebugLoc(), TII.get(SPIRV::OpSwitch)); 1904 for (unsigned i = 1; i < I.getNumExplicitOperands(); ++i) { 1905 if (I.getOperand(i).isReg()) 1906 MIB.addReg(I.getOperand(i).getReg()); 1907 else if (I.getOperand(i).isCImm()) 1908 addNumImm(I.getOperand(i).getCImm()->getValue(), MIB); 1909 else if (I.getOperand(i).isMBB()) 1910 MIB.addMBB(I.getOperand(i).getMBB()); 1911 else 1912 llvm_unreachable("Unexpected OpSwitch operand"); 1913 } 1914 return MIB.constrainAllUses(TII, TRI, RBI); 1915 } 1916 case Intrinsic::spv_cmpxchg: 1917 return selectAtomicCmpXchg(ResVReg, ResType, I); 1918 case Intrinsic::spv_unreachable: 1919 BuildMI(BB, I, I.getDebugLoc(), TII.get(SPIRV::OpUnreachable)); 1920 break; 1921 case Intrinsic::spv_alloca: 1922 return selectFrameIndex(ResVReg, ResType, I); 1923 case Intrinsic::spv_alloca_array: 1924 return selectAllocaArray(ResVReg, ResType, I); 1925 case Intrinsic::spv_assume: 1926 if (STI.canUseExtension(SPIRV::Extension::SPV_KHR_expect_assume)) 1927 BuildMI(BB, I, I.getDebugLoc(), TII.get(SPIRV::OpAssumeTrueKHR)) 1928 .addUse(I.getOperand(1).getReg()); 1929 break; 1930 case Intrinsic::spv_expect: 1931 if (STI.canUseExtension(SPIRV::Extension::SPV_KHR_expect_assume)) 1932 BuildMI(BB, I, I.getDebugLoc(), TII.get(SPIRV::OpExpectKHR)) 1933 .addDef(ResVReg) 1934 .addUse(GR.getSPIRVTypeID(ResType)) 1935 .addUse(I.getOperand(2).getReg()) 1936 .addUse(I.getOperand(3).getReg()); 1937 break; 1938 case Intrinsic::spv_thread_id: 1939 return selectSpvThreadId(ResVReg, ResType, I); 1940 case Intrinsic::spv_all: 1941 return selectAll(ResVReg, ResType, I); 1942 case Intrinsic::spv_any: 1943 return selectAny(ResVReg, ResType, I); 1944 case Intrinsic::spv_lerp: 1945 return selectFmix(ResVReg, ResType, I); 1946 case Intrinsic::spv_lifetime_start: 1947 case Intrinsic::spv_lifetime_end: { 1948 unsigned Op = IID == Intrinsic::spv_lifetime_start ? SPIRV::OpLifetimeStart 1949 : SPIRV::OpLifetimeStop; 1950 int64_t Size = I.getOperand(I.getNumExplicitDefs() + 1).getImm(); 1951 Register PtrReg = I.getOperand(I.getNumExplicitDefs() + 2).getReg(); 1952 unsigned PonteeOpType = GR.getPointeeTypeOp(PtrReg); 1953 bool IsNonvoidPtr = PonteeOpType != 0 && PonteeOpType != SPIRV::OpTypeVoid; 1954 if (Size == -1 || IsNonvoidPtr) 1955 Size = 0; 1956 BuildMI(BB, I, I.getDebugLoc(), TII.get(Op)).addUse(PtrReg).addImm(Size); 1957 } break; 1958 default: { 1959 std::string DiagMsg; 1960 raw_string_ostream OS(DiagMsg); 1961 I.print(OS); 1962 DiagMsg = "Intrinsic selection not implemented: " + DiagMsg; 1963 report_fatal_error(DiagMsg.c_str(), false); 1964 } 1965 } 1966 return true; 1967 } 1968 1969 bool SPIRVInstructionSelector::selectAllocaArray(Register ResVReg, 1970 const SPIRVType *ResType, 1971 MachineInstr &I) const { 1972 // there was an allocation size parameter to the allocation instruction 1973 // that is not 1 1974 MachineBasicBlock &BB = *I.getParent(); 1975 return BuildMI(BB, I, I.getDebugLoc(), 1976 TII.get(SPIRV::OpVariableLengthArrayINTEL)) 1977 .addDef(ResVReg) 1978 .addUse(GR.getSPIRVTypeID(ResType)) 1979 .addUse(I.getOperand(2).getReg()) 1980 .constrainAllUses(TII, TRI, RBI); 1981 } 1982 1983 bool SPIRVInstructionSelector::selectFrameIndex(Register ResVReg, 1984 const SPIRVType *ResType, 1985 MachineInstr &I) const { 1986 // Change order of instructions if needed: all OpVariable instructions in a 1987 // function must be the first instructions in the first block 1988 MachineFunction *MF = I.getParent()->getParent(); 1989 MachineBasicBlock *MBB = &MF->front(); 1990 auto It = MBB->SkipPHIsAndLabels(MBB->begin()), E = MBB->end(); 1991 bool IsHeader = false; 1992 unsigned Opcode; 1993 for (; It != E && It != I; ++It) { 1994 Opcode = It->getOpcode(); 1995 if (Opcode == SPIRV::OpFunction || Opcode == SPIRV::OpFunctionParameter) { 1996 IsHeader = true; 1997 } else if (IsHeader && 1998 !(Opcode == SPIRV::ASSIGN_TYPE || Opcode == SPIRV::OpLabel)) { 1999 ++It; 2000 break; 2001 } 2002 } 2003 return BuildMI(*MBB, It, It->getDebugLoc(), TII.get(SPIRV::OpVariable)) 2004 .addDef(ResVReg) 2005 .addUse(GR.getSPIRVTypeID(ResType)) 2006 .addImm(static_cast<uint32_t>(SPIRV::StorageClass::Function)) 2007 .constrainAllUses(TII, TRI, RBI); 2008 } 2009 2010 bool SPIRVInstructionSelector::selectBranch(MachineInstr &I) const { 2011 // InstructionSelector walks backwards through the instructions. We can use 2012 // both a G_BR and a G_BRCOND to create an OpBranchConditional. We hit G_BR 2013 // first, so can generate an OpBranchConditional here. If there is no 2014 // G_BRCOND, we just use OpBranch for a regular unconditional branch. 2015 const MachineInstr *PrevI = I.getPrevNode(); 2016 MachineBasicBlock &MBB = *I.getParent(); 2017 if (PrevI != nullptr && PrevI->getOpcode() == TargetOpcode::G_BRCOND) { 2018 return BuildMI(MBB, I, I.getDebugLoc(), TII.get(SPIRV::OpBranchConditional)) 2019 .addUse(PrevI->getOperand(0).getReg()) 2020 .addMBB(PrevI->getOperand(1).getMBB()) 2021 .addMBB(I.getOperand(0).getMBB()) 2022 .constrainAllUses(TII, TRI, RBI); 2023 } 2024 return BuildMI(MBB, I, I.getDebugLoc(), TII.get(SPIRV::OpBranch)) 2025 .addMBB(I.getOperand(0).getMBB()) 2026 .constrainAllUses(TII, TRI, RBI); 2027 } 2028 2029 bool SPIRVInstructionSelector::selectBranchCond(MachineInstr &I) const { 2030 // InstructionSelector walks backwards through the instructions. For an 2031 // explicit conditional branch with no fallthrough, we use both a G_BR and a 2032 // G_BRCOND to create an OpBranchConditional. We should hit G_BR first, and 2033 // generate the OpBranchConditional in selectBranch above. 2034 // 2035 // If an OpBranchConditional has been generated, we simply return, as the work 2036 // is alread done. If there is no OpBranchConditional, LLVM must be relying on 2037 // implicit fallthrough to the next basic block, so we need to create an 2038 // OpBranchConditional with an explicit "false" argument pointing to the next 2039 // basic block that LLVM would fall through to. 2040 const MachineInstr *NextI = I.getNextNode(); 2041 // Check if this has already been successfully selected. 2042 if (NextI != nullptr && NextI->getOpcode() == SPIRV::OpBranchConditional) 2043 return true; 2044 // Must be relying on implicit block fallthrough, so generate an 2045 // OpBranchConditional with the "next" basic block as the "false" target. 2046 MachineBasicBlock &MBB = *I.getParent(); 2047 unsigned NextMBBNum = MBB.getNextNode()->getNumber(); 2048 MachineBasicBlock *NextMBB = I.getMF()->getBlockNumbered(NextMBBNum); 2049 return BuildMI(MBB, I, I.getDebugLoc(), TII.get(SPIRV::OpBranchConditional)) 2050 .addUse(I.getOperand(0).getReg()) 2051 .addMBB(I.getOperand(1).getMBB()) 2052 .addMBB(NextMBB) 2053 .constrainAllUses(TII, TRI, RBI); 2054 } 2055 2056 bool SPIRVInstructionSelector::selectPhi(Register ResVReg, 2057 const SPIRVType *ResType, 2058 MachineInstr &I) const { 2059 auto MIB = BuildMI(*I.getParent(), I, I.getDebugLoc(), TII.get(SPIRV::OpPhi)) 2060 .addDef(ResVReg) 2061 .addUse(GR.getSPIRVTypeID(ResType)); 2062 const unsigned NumOps = I.getNumOperands(); 2063 for (unsigned i = 1; i < NumOps; i += 2) { 2064 MIB.addUse(I.getOperand(i + 0).getReg()); 2065 MIB.addMBB(I.getOperand(i + 1).getMBB()); 2066 } 2067 return MIB.constrainAllUses(TII, TRI, RBI); 2068 } 2069 2070 bool SPIRVInstructionSelector::selectGlobalValue( 2071 Register ResVReg, MachineInstr &I, const MachineInstr *Init) const { 2072 // FIXME: don't use MachineIRBuilder here, replace it with BuildMI. 2073 MachineIRBuilder MIRBuilder(I); 2074 const GlobalValue *GV = I.getOperand(1).getGlobal(); 2075 Type *GVType = GR.getDeducedGlobalValueType(GV); 2076 SPIRVType *PointerBaseType; 2077 if (GVType->isArrayTy()) { 2078 SPIRVType *ArrayElementType = 2079 GR.getOrCreateSPIRVType(GVType->getArrayElementType(), MIRBuilder, 2080 SPIRV::AccessQualifier::ReadWrite, false); 2081 PointerBaseType = GR.getOrCreateSPIRVArrayType( 2082 ArrayElementType, GVType->getArrayNumElements(), I, TII); 2083 } else { 2084 PointerBaseType = GR.getOrCreateSPIRVType( 2085 GVType, MIRBuilder, SPIRV::AccessQualifier::ReadWrite, false); 2086 } 2087 SPIRVType *ResType = GR.getOrCreateSPIRVPointerType( 2088 PointerBaseType, I, TII, 2089 addressSpaceToStorageClass(GV->getAddressSpace(), STI)); 2090 2091 std::string GlobalIdent; 2092 if (!GV->hasName()) { 2093 unsigned &ID = UnnamedGlobalIDs[GV]; 2094 if (ID == 0) 2095 ID = UnnamedGlobalIDs.size(); 2096 GlobalIdent = "__unnamed_" + Twine(ID).str(); 2097 } else { 2098 GlobalIdent = GV->getGlobalIdentifier(); 2099 } 2100 2101 // Behaviour of functions as operands depends on availability of the 2102 // corresponding extension (SPV_INTEL_function_pointers): 2103 // - If there is an extension to operate with functions as operands: 2104 // We create a proper constant operand and evaluate a correct type for a 2105 // function pointer. 2106 // - Without the required extension: 2107 // We have functions as operands in tests with blocks of instruction e.g. in 2108 // transcoding/global_block.ll. These operands are not used and should be 2109 // substituted by zero constants. Their type is expected to be always 2110 // OpTypePointer Function %uchar. 2111 if (isa<Function>(GV)) { 2112 const Constant *ConstVal = GV; 2113 MachineBasicBlock &BB = *I.getParent(); 2114 Register NewReg = GR.find(ConstVal, GR.CurMF); 2115 if (!NewReg.isValid()) { 2116 Register NewReg = ResVReg; 2117 GR.add(ConstVal, GR.CurMF, NewReg); 2118 const Function *GVFun = 2119 STI.canUseExtension(SPIRV::Extension::SPV_INTEL_function_pointers) 2120 ? dyn_cast<Function>(GV) 2121 : nullptr; 2122 if (GVFun) { 2123 // References to a function via function pointers generate virtual 2124 // registers without a definition. We will resolve it later, during 2125 // module analysis stage. 2126 MachineRegisterInfo *MRI = MIRBuilder.getMRI(); 2127 Register FuncVReg = MRI->createGenericVirtualRegister(LLT::scalar(32)); 2128 MRI->setRegClass(FuncVReg, &SPIRV::IDRegClass); 2129 MachineInstrBuilder MB = 2130 BuildMI(BB, I, I.getDebugLoc(), 2131 TII.get(SPIRV::OpConstantFunctionPointerINTEL)) 2132 .addDef(NewReg) 2133 .addUse(GR.getSPIRVTypeID(ResType)) 2134 .addUse(FuncVReg); 2135 // mapping the function pointer to the used Function 2136 GR.recordFunctionPointer(&MB.getInstr()->getOperand(2), GVFun); 2137 return MB.constrainAllUses(TII, TRI, RBI); 2138 } 2139 return BuildMI(BB, I, I.getDebugLoc(), TII.get(SPIRV::OpConstantNull)) 2140 .addDef(NewReg) 2141 .addUse(GR.getSPIRVTypeID(ResType)) 2142 .constrainAllUses(TII, TRI, RBI); 2143 } 2144 assert(NewReg != ResVReg); 2145 return BuildMI(BB, I, I.getDebugLoc(), TII.get(TargetOpcode::COPY)) 2146 .addDef(ResVReg) 2147 .addUse(NewReg) 2148 .constrainAllUses(TII, TRI, RBI); 2149 } 2150 auto GlobalVar = cast<GlobalVariable>(GV); 2151 assert(GlobalVar->getName() != "llvm.global.annotations"); 2152 2153 bool HasInit = GlobalVar->hasInitializer() && 2154 !isa<UndefValue>(GlobalVar->getInitializer()); 2155 // Skip empty declaration for GVs with initilaizers till we get the decl with 2156 // passed initializer. 2157 if (HasInit && !Init) 2158 return true; 2159 2160 unsigned AddrSpace = GV->getAddressSpace(); 2161 SPIRV::StorageClass::StorageClass Storage = 2162 addressSpaceToStorageClass(AddrSpace, STI); 2163 bool HasLnkTy = GV->getLinkage() != GlobalValue::InternalLinkage && 2164 Storage != SPIRV::StorageClass::Function; 2165 SPIRV::LinkageType::LinkageType LnkType = 2166 (GV->isDeclaration() || GV->hasAvailableExternallyLinkage()) 2167 ? SPIRV::LinkageType::Import 2168 : (GV->getLinkage() == GlobalValue::LinkOnceODRLinkage && 2169 STI.canUseExtension(SPIRV::Extension::SPV_KHR_linkonce_odr) 2170 ? SPIRV::LinkageType::LinkOnceODR 2171 : SPIRV::LinkageType::Export); 2172 2173 Register Reg = GR.buildGlobalVariable(ResVReg, ResType, GlobalIdent, GV, 2174 Storage, Init, GlobalVar->isConstant(), 2175 HasLnkTy, LnkType, MIRBuilder, true); 2176 return Reg.isValid(); 2177 } 2178 2179 bool SPIRVInstructionSelector::selectLog10(Register ResVReg, 2180 const SPIRVType *ResType, 2181 MachineInstr &I) const { 2182 if (STI.canUseExtInstSet(SPIRV::InstructionSet::OpenCL_std)) { 2183 return selectExtInst(ResVReg, ResType, I, CL::log10); 2184 } 2185 2186 // There is no log10 instruction in the GLSL Extended Instruction set, so it 2187 // is implemented as: 2188 // log10(x) = log2(x) * (1 / log2(10)) 2189 // = log2(x) * 0.30103 2190 2191 MachineIRBuilder MIRBuilder(I); 2192 MachineBasicBlock &BB = *I.getParent(); 2193 2194 // Build log2(x). 2195 Register VarReg = MRI->createVirtualRegister(&SPIRV::IDRegClass); 2196 bool Result = 2197 BuildMI(BB, I, I.getDebugLoc(), TII.get(SPIRV::OpExtInst)) 2198 .addDef(VarReg) 2199 .addUse(GR.getSPIRVTypeID(ResType)) 2200 .addImm(static_cast<uint32_t>(SPIRV::InstructionSet::GLSL_std_450)) 2201 .addImm(GL::Log2) 2202 .add(I.getOperand(1)) 2203 .constrainAllUses(TII, TRI, RBI); 2204 2205 // Build 0.30103. 2206 assert(ResType->getOpcode() == SPIRV::OpTypeVector || 2207 ResType->getOpcode() == SPIRV::OpTypeFloat); 2208 // TODO: Add matrix implementation once supported by the HLSL frontend. 2209 const SPIRVType *SpirvScalarType = 2210 ResType->getOpcode() == SPIRV::OpTypeVector 2211 ? GR.getSPIRVTypeForVReg(ResType->getOperand(1).getReg()) 2212 : ResType; 2213 Register ScaleReg = 2214 GR.buildConstantFP(APFloat(0.30103f), MIRBuilder, SpirvScalarType); 2215 2216 // Multiply log2(x) by 0.30103 to get log10(x) result. 2217 auto Opcode = ResType->getOpcode() == SPIRV::OpTypeVector 2218 ? SPIRV::OpVectorTimesScalar 2219 : SPIRV::OpFMulS; 2220 Result &= BuildMI(BB, I, I.getDebugLoc(), TII.get(Opcode)) 2221 .addDef(ResVReg) 2222 .addUse(GR.getSPIRVTypeID(ResType)) 2223 .addUse(VarReg) 2224 .addUse(ScaleReg) 2225 .constrainAllUses(TII, TRI, RBI); 2226 2227 return Result; 2228 } 2229 2230 bool SPIRVInstructionSelector::selectSpvThreadId(Register ResVReg, 2231 const SPIRVType *ResType, 2232 MachineInstr &I) const { 2233 // DX intrinsic: @llvm.dx.thread.id(i32) 2234 // ID Name Description 2235 // 93 ThreadId reads the thread ID 2236 2237 MachineIRBuilder MIRBuilder(I); 2238 const SPIRVType *U32Type = GR.getOrCreateSPIRVIntegerType(32, MIRBuilder); 2239 const SPIRVType *Vec3Ty = 2240 GR.getOrCreateSPIRVVectorType(U32Type, 3, MIRBuilder); 2241 const SPIRVType *PtrType = GR.getOrCreateSPIRVPointerType( 2242 Vec3Ty, MIRBuilder, SPIRV::StorageClass::Input); 2243 2244 // Create new register for GlobalInvocationID builtin variable. 2245 Register NewRegister = 2246 MIRBuilder.getMRI()->createVirtualRegister(&SPIRV::IDRegClass); 2247 MIRBuilder.getMRI()->setType(NewRegister, LLT::pointer(0, 32)); 2248 GR.assignSPIRVTypeToVReg(PtrType, NewRegister, MIRBuilder.getMF()); 2249 2250 // Build GlobalInvocationID global variable with the necessary decorations. 2251 Register Variable = GR.buildGlobalVariable( 2252 NewRegister, PtrType, 2253 getLinkStringForBuiltIn(SPIRV::BuiltIn::GlobalInvocationId), nullptr, 2254 SPIRV::StorageClass::Input, nullptr, true, true, 2255 SPIRV::LinkageType::Import, MIRBuilder, false); 2256 2257 // Create new register for loading value. 2258 MachineRegisterInfo *MRI = MIRBuilder.getMRI(); 2259 Register LoadedRegister = MRI->createVirtualRegister(&SPIRV::IDRegClass); 2260 MIRBuilder.getMRI()->setType(LoadedRegister, LLT::pointer(0, 32)); 2261 GR.assignSPIRVTypeToVReg(Vec3Ty, LoadedRegister, MIRBuilder.getMF()); 2262 2263 // Load v3uint value from the global variable. 2264 BuildMI(*I.getParent(), I, I.getDebugLoc(), TII.get(SPIRV::OpLoad)) 2265 .addDef(LoadedRegister) 2266 .addUse(GR.getSPIRVTypeID(Vec3Ty)) 2267 .addUse(Variable); 2268 2269 // Get Thread ID index. Expecting operand is a constant immediate value, 2270 // wrapped in a type assignment. 2271 assert(I.getOperand(2).isReg()); 2272 Register ThreadIdReg = I.getOperand(2).getReg(); 2273 SPIRVType *ConstTy = this->MRI->getVRegDef(ThreadIdReg); 2274 assert(ConstTy && ConstTy->getOpcode() == SPIRV::ASSIGN_TYPE && 2275 ConstTy->getOperand(1).isReg()); 2276 Register ConstReg = ConstTy->getOperand(1).getReg(); 2277 const MachineInstr *Const = this->MRI->getVRegDef(ConstReg); 2278 assert(Const && Const->getOpcode() == TargetOpcode::G_CONSTANT); 2279 const llvm::APInt &Val = Const->getOperand(1).getCImm()->getValue(); 2280 const uint32_t ThreadId = Val.getZExtValue(); 2281 2282 // Extract the thread ID from the loaded vector value. 2283 MachineBasicBlock &BB = *I.getParent(); 2284 auto MIB = BuildMI(BB, I, I.getDebugLoc(), TII.get(SPIRV::OpCompositeExtract)) 2285 .addDef(ResVReg) 2286 .addUse(GR.getSPIRVTypeID(ResType)) 2287 .addUse(LoadedRegister) 2288 .addImm(ThreadId); 2289 return MIB.constrainAllUses(TII, TRI, RBI); 2290 } 2291 2292 namespace llvm { 2293 InstructionSelector * 2294 createSPIRVInstructionSelector(const SPIRVTargetMachine &TM, 2295 const SPIRVSubtarget &Subtarget, 2296 const RegisterBankInfo &RBI) { 2297 return new SPIRVInstructionSelector(TM, Subtarget, RBI); 2298 } 2299 } // namespace llvm 2300