| 1 | //===-- SPIRVPreLegalizer.cpp - prepare IR for legalization -----*- 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 | // The pass prepares IR for legalization: it assigns SPIR-V types to registers |
| 10 | // and removes intrinsics which holded these types during IR translation. |
| 11 | // Also it processes constants and registers them in GR to avoid duplication. |
| 12 | // |
| 13 | //===----------------------------------------------------------------------===// |
| 14 | |
| 15 | #include "SPIRV.h" |
| 16 | #include "SPIRVSubtarget.h" |
| 17 | #include "SPIRVUtils.h" |
| 18 | #include "llvm/ADT/PostOrderIterator.h" |
| 19 | #include "llvm/CodeGen/GlobalISel/CSEInfo.h" |
| 20 | #include "llvm/CodeGen/GlobalISel/GISelValueTracking.h" |
| 21 | #include "llvm/CodeGen/GlobalISel/MIPatternMatch.h" |
| 22 | #include "llvm/CodeGen/MachineFunctionAnalysisManager.h" |
| 23 | #include "llvm/CodeGen/MachinePassManager.h" |
| 24 | #include "llvm/IR/Analysis.h" |
| 25 | #include "llvm/IR/Attributes.h" |
| 26 | #include "llvm/IR/Constants.h" |
| 27 | #include "llvm/IR/InstrTypes.h" |
| 28 | #include "llvm/IR/IntrinsicsSPIRV.h" |
| 29 | #include "llvm/Support/MathExtras.h" |
| 30 | |
| 31 | #define DEBUG_TYPE "spirv-prelegalizer" |
| 32 | |
| 33 | using namespace llvm; |
| 34 | using namespace llvm::MIPatternMatch; |
| 35 | |
| 36 | namespace { |
| 37 | class SPIRVPreLegalizerLegacy : public MachineFunctionPass { |
| 38 | public: |
| 39 | static char ID; |
| 40 | SPIRVPreLegalizerLegacy() : MachineFunctionPass(ID) {} |
| 41 | bool runOnMachineFunction(MachineFunction &MF) override; |
| 42 | void getAnalysisUsage(AnalysisUsage &AU) const override; |
| 43 | }; |
| 44 | } // namespace |
| 45 | |
| 46 | void SPIRVPreLegalizerLegacy::getAnalysisUsage(AnalysisUsage &AU) const { |
| 47 | AU.addPreserved<GISelValueTrackingAnalysisLegacy>(); |
| 48 | MachineFunctionPass::getAnalysisUsage(AU); |
| 49 | } |
| 50 | |
| 51 | static inline void invalidateAndEraseMI(SPIRVGlobalRegistry *GR, |
| 52 | MachineInstr *MI) { |
| 53 | GR->invalidateMachineInstr(MI); |
| 54 | MI->eraseFromParent(); |
| 55 | } |
| 56 | |
| 57 | static void |
| 58 | addConstantsToTrack(MachineFunction &MF, SPIRVGlobalRegistry *GR, |
| 59 | const SPIRVSubtarget &STI, |
| 60 | DenseMap<MachineInstr *, Type *> &TargetExtConstTypes) { |
| 61 | MachineRegisterInfo &MRI = MF.getRegInfo(); |
| 62 | DenseMap<MachineInstr *, Register> RegsAlreadyAddedToDT; |
| 63 | SmallVector<MachineInstr *, 10> ToErase, ToEraseComposites; |
| 64 | for (MachineBasicBlock &MBB : MF) { |
| 65 | for (MachineInstr &MI : MBB) { |
| 66 | if (!isSpvIntrinsic(MI, IntrinsicID: Intrinsic::spv_track_constant)) |
| 67 | continue; |
| 68 | ToErase.push_back(Elt: &MI); |
| 69 | Register SrcReg = MI.getOperand(i: 2).getReg(); |
| 70 | auto *Const = |
| 71 | cast<Constant>(Val: cast<ConstantAsMetadata>( |
| 72 | Val: MI.getOperand(i: 3).getMetadata()->getOperand(I: 0)) |
| 73 | ->getValue()); |
| 74 | if (auto *GV = dyn_cast<GlobalValue>(Val: Const)) { |
| 75 | Register Reg = GR->find(V: GV, MF: &MF); |
| 76 | if (!Reg.isValid()) { |
| 77 | GR->add(V: GV, MI: MRI.getVRegDef(Reg: SrcReg)); |
| 78 | GR->addGlobalObject(V: GV, MF: &MF, R: SrcReg); |
| 79 | } else |
| 80 | RegsAlreadyAddedToDT[&MI] = Reg; |
| 81 | } else { |
| 82 | Register Reg = GR->find(V: Const, MF: &MF); |
| 83 | if (!Reg.isValid()) { |
| 84 | if (auto *ConstVec = dyn_cast<ConstantDataVector>(Val: Const)) { |
| 85 | auto *BuildVec = MRI.getVRegDef(Reg: SrcReg); |
| 86 | assert(BuildVec && |
| 87 | BuildVec->getOpcode() == TargetOpcode::G_BUILD_VECTOR); |
| 88 | GR->add(V: Const, MI: BuildVec); |
| 89 | for (unsigned i = 0; i < ConstVec->getNumElements(); ++i) { |
| 90 | // Ensure that OpConstantComposite reuses a constant when it's |
| 91 | // already created and available in the same machine function. |
| 92 | Constant *ElemConst = ConstVec->getElementAsConstant(i); |
| 93 | Register ElemReg = GR->find(V: ElemConst, MF: &MF); |
| 94 | if (!ElemReg.isValid()) |
| 95 | GR->add(V: ElemConst, |
| 96 | MI: MRI.getVRegDef(Reg: BuildVec->getOperand(i: 1 + i).getReg())); |
| 97 | else |
| 98 | BuildVec->getOperand(i: 1 + i).setReg(ElemReg); |
| 99 | } |
| 100 | } |
| 101 | if (Const->getType()->isTargetExtTy()) { |
| 102 | // remember association so that we can restore it when assign types |
| 103 | MachineInstr *SrcMI = MRI.getVRegDef(Reg: SrcReg); |
| 104 | if (SrcMI) |
| 105 | GR->add(V: Const, MI: SrcMI); |
| 106 | if (SrcMI && (SrcMI->getOpcode() == TargetOpcode::G_CONSTANT || |
| 107 | SrcMI->getOpcode() == TargetOpcode::G_IMPLICIT_DEF)) |
| 108 | TargetExtConstTypes[SrcMI] = Const->getType(); |
| 109 | if (Const->isNullValue()) { |
| 110 | MachineBasicBlock &DepMBB = MF.front(); |
| 111 | MachineIRBuilder MIB(DepMBB, DepMBB.getFirstNonPHI()); |
| 112 | SPIRVTypeInst ExtType = GR->getOrCreateSPIRVType( |
| 113 | Type: Const->getType(), MIRBuilder&: MIB, AQ: SPIRV::AccessQualifier::ReadWrite, |
| 114 | EmitIR: true); |
| 115 | assert(SrcMI && "Expected source instruction to be valid" ); |
| 116 | SrcMI->setDesc(STI.getInstrInfo()->get(Opcode: SPIRV::OpConstantNull)); |
| 117 | SrcMI->addOperand(Op: MachineOperand::CreateReg( |
| 118 | Reg: GR->getSPIRVTypeID(SpirvType: ExtType), isDef: false)); |
| 119 | } |
| 120 | } |
| 121 | } else { |
| 122 | RegsAlreadyAddedToDT[&MI] = Reg; |
| 123 | // This MI is unused and will be removed. If the MI uses |
| 124 | // const_composite, it will be unused and should be removed too. |
| 125 | assert(MI.getOperand(2).isReg() && "Reg operand is expected" ); |
| 126 | MachineInstr *SrcMI = MRI.getVRegDef(Reg: MI.getOperand(i: 2).getReg()); |
| 127 | if (SrcMI && isSpvIntrinsic(MI: *SrcMI, IntrinsicID: Intrinsic::spv_const_composite)) |
| 128 | ToEraseComposites.push_back(Elt: SrcMI); |
| 129 | } |
| 130 | } |
| 131 | } |
| 132 | } |
| 133 | for (MachineInstr *MI : ToErase) { |
| 134 | Register Reg = MI->getOperand(i: 2).getReg(); |
| 135 | auto It = RegsAlreadyAddedToDT.find(Val: MI); |
| 136 | if (It != RegsAlreadyAddedToDT.end()) |
| 137 | Reg = It->second; |
| 138 | auto *RC = MRI.getRegClassOrNull(Reg: MI->getOperand(i: 0).getReg()); |
| 139 | if (!MRI.getRegClassOrNull(Reg) && RC) |
| 140 | MRI.setRegClass(Reg, RC); |
| 141 | MRI.replaceRegWith(FromReg: MI->getOperand(i: 0).getReg(), ToReg: Reg); |
| 142 | invalidateAndEraseMI(GR, MI); |
| 143 | } |
| 144 | for (MachineInstr *MI : ToEraseComposites) |
| 145 | invalidateAndEraseMI(GR, MI); |
| 146 | } |
| 147 | |
| 148 | static void foldConstantsIntoIntrinsics(MachineFunction &MF, |
| 149 | SPIRVGlobalRegistry *GR, |
| 150 | MachineIRBuilder MIB) { |
| 151 | SmallVector<MachineInstr *, 64> ToErase; |
| 152 | for (MachineBasicBlock &MBB : MF) { |
| 153 | for (MachineInstr &MI : MBB) { |
| 154 | if (!isSpvIntrinsic(MI, IntrinsicID: Intrinsic::spv_assign_name)) |
| 155 | continue; |
| 156 | const MDNode *MD = MI.getOperand(i: 2).getMetadata(); |
| 157 | StringRef ValueName = cast<MDString>(Val: MD->getOperand(I: 0))->getString(); |
| 158 | if (ValueName.size() > 0) { |
| 159 | MIB.setInsertPt(MBB&: *MI.getParent(), II: MI); |
| 160 | buildOpName(Target: MI.getOperand(i: 1).getReg(), Name: ValueName, MIRBuilder&: MIB); |
| 161 | } |
| 162 | ToErase.push_back(Elt: &MI); |
| 163 | } |
| 164 | for (MachineInstr *MI : ToErase) |
| 165 | invalidateAndEraseMI(GR, MI); |
| 166 | ToErase.clear(); |
| 167 | } |
| 168 | } |
| 169 | |
| 170 | static MachineInstr *findAssignTypeInstr(Register Reg, |
| 171 | MachineRegisterInfo *MRI) { |
| 172 | for (MachineRegisterInfo::use_instr_iterator I = MRI->use_instr_begin(RegNo: Reg), |
| 173 | IE = MRI->use_instr_end(); |
| 174 | I != IE; ++I) { |
| 175 | MachineInstr *UseMI = &*I; |
| 176 | if ((isSpvIntrinsic(MI: *UseMI, IntrinsicID: Intrinsic::spv_assign_ptr_type) || |
| 177 | isSpvIntrinsic(MI: *UseMI, IntrinsicID: Intrinsic::spv_assign_type)) && |
| 178 | UseMI->getOperand(i: 1).getReg() == Reg) |
| 179 | return UseMI; |
| 180 | } |
| 181 | return nullptr; |
| 182 | } |
| 183 | |
| 184 | static void buildOpBitcast(SPIRVGlobalRegistry *GR, MachineIRBuilder &MIB, |
| 185 | Register ResVReg, Register OpReg) { |
| 186 | SPIRVTypeInst ResType = GR->getSPIRVTypeForVReg(VReg: ResVReg); |
| 187 | SPIRVTypeInst OpType = GR->getSPIRVTypeForVReg(VReg: OpReg); |
| 188 | assert(ResType && OpType && "Operand types are expected" ); |
| 189 | if (!GR->isBitcastCompatible(Type1: ResType, Type2: OpType)) |
| 190 | report_fatal_error(reason: "incompatible result and operand types in a bitcast" ); |
| 191 | MachineRegisterInfo *MRI = MIB.getMRI(); |
| 192 | if (!MRI->getRegClassOrNull(Reg: ResVReg)) |
| 193 | MRI->setRegClass(Reg: ResVReg, RC: GR->getRegClass(SpvType: ResType)); |
| 194 | if (ResType == OpType) |
| 195 | MIB.buildInstr(Opcode: TargetOpcode::COPY).addDef(RegNo: ResVReg).addUse(RegNo: OpReg); |
| 196 | else |
| 197 | MIB.buildInstr(Opcode: SPIRV::OpBitcast) |
| 198 | .addDef(RegNo: ResVReg) |
| 199 | .addUse(RegNo: GR->getSPIRVTypeID(SpirvType: ResType)) |
| 200 | .addUse(RegNo: OpReg); |
| 201 | } |
| 202 | |
| 203 | // We lower G_BITCAST to OpBitcast here to avoid a MachineVerifier error. |
| 204 | // The verifier checks if the source and destination LLTs of a G_BITCAST are |
| 205 | // different, but this check is too strict for SPIR-V's typed pointers, which |
| 206 | // may have the same LLT but different SPIRV type (e.g. pointers to different |
| 207 | // pointee types). By lowering to OpBitcast here, we bypass the verifier's |
| 208 | // check. See discussion in https://github.com/llvm/llvm-project/pull/110270 |
| 209 | // for more context. |
| 210 | // |
| 211 | // We also handle the llvm.spv.bitcast intrinsic here. If the source and |
| 212 | // destination SPIR-V types are the same, we lower it to a COPY to enable |
| 213 | // further optimizations like copy propagation. |
| 214 | static void lowerBitcasts(MachineFunction &MF, SPIRVGlobalRegistry *GR, |
| 215 | MachineIRBuilder MIB) { |
| 216 | SmallVector<MachineInstr *, 16> ToErase; |
| 217 | for (MachineBasicBlock &MBB : MF) { |
| 218 | for (MachineInstr &MI : MBB) { |
| 219 | if (isSpvIntrinsic(MI, IntrinsicID: Intrinsic::spv_bitcast)) { |
| 220 | Register DstReg = MI.getOperand(i: 0).getReg(); |
| 221 | Register SrcReg = MI.getOperand(i: 2).getReg(); |
| 222 | SPIRVTypeInst DstType = GR->getSPIRVTypeForVReg(VReg: DstReg); |
| 223 | assert( |
| 224 | DstType && |
| 225 | "Expected destination SPIR-V type to have been assigned already." ); |
| 226 | SPIRVTypeInst SrcType = GR->getSPIRVTypeForVReg(VReg: SrcReg); |
| 227 | assert(SrcType && |
| 228 | "Expected source SPIR-V type to have been assigned already." ); |
| 229 | if (DstType == SrcType) { |
| 230 | MIB.setInsertPt(MBB&: *MI.getParent(), II: MI); |
| 231 | MIB.buildCopy(Res: DstReg, Op: SrcReg); |
| 232 | ToErase.push_back(Elt: &MI); |
| 233 | continue; |
| 234 | } |
| 235 | } |
| 236 | |
| 237 | if (MI.getOpcode() != TargetOpcode::G_BITCAST) |
| 238 | continue; |
| 239 | |
| 240 | MIB.setInsertPt(MBB&: *MI.getParent(), II: MI); |
| 241 | buildOpBitcast(GR, MIB, ResVReg: MI.getOperand(i: 0).getReg(), |
| 242 | OpReg: MI.getOperand(i: 1).getReg()); |
| 243 | ToErase.push_back(Elt: &MI); |
| 244 | } |
| 245 | } |
| 246 | for (MachineInstr *MI : ToErase) |
| 247 | invalidateAndEraseMI(GR, MI); |
| 248 | } |
| 249 | |
| 250 | static void insertBitcasts(MachineFunction &MF, SPIRVGlobalRegistry *GR, |
| 251 | MachineIRBuilder MIB) { |
| 252 | // Get access to information about available extensions |
| 253 | const SPIRVSubtarget *ST = |
| 254 | static_cast<const SPIRVSubtarget *>(&MIB.getMF().getSubtarget()); |
| 255 | SmallVector<MachineInstr *, 10> ToErase; |
| 256 | for (MachineBasicBlock &MBB : MF) { |
| 257 | for (MachineInstr &MI : MBB) { |
| 258 | if (!isSpvIntrinsic(MI, IntrinsicID: Intrinsic::spv_ptrcast)) |
| 259 | continue; |
| 260 | assert(MI.getOperand(2).isReg()); |
| 261 | MIB.setInsertPt(MBB&: *MI.getParent(), II: MI); |
| 262 | ToErase.push_back(Elt: &MI); |
| 263 | Register Def = MI.getOperand(i: 0).getReg(); |
| 264 | Register Source = MI.getOperand(i: 2).getReg(); |
| 265 | Type *ElemTy = getMDOperandAsType(N: MI.getOperand(i: 3).getMetadata(), I: 0); |
| 266 | auto SC = |
| 267 | isa<FunctionType>(Val: ElemTy) && |
| 268 | ST->canUseExtension( |
| 269 | E: SPIRV::Extension::SPV_INTEL_function_pointers) |
| 270 | ? SPIRV::StorageClass::CodeSectionINTEL |
| 271 | : addressSpaceToStorageClass(AddrSpace: MI.getOperand(i: 4).getImm(), STI: *ST); |
| 272 | SPIRVTypeInst AssignedPtrType = |
| 273 | GR->getOrCreateSPIRVPointerType(BaseType: ElemTy, I&: MI, SC); |
| 274 | |
| 275 | // If the ptrcast would be redundant, replace all uses with the source |
| 276 | // register. |
| 277 | MachineRegisterInfo *MRI = MIB.getMRI(); |
| 278 | // For untyped pointers the SPIR-V pointer type does not encode the |
| 279 | // pointee, so two pointers with different element types share the same |
| 280 | // pointer type. The element type still matters because it selects the |
| 281 | // Base Type operand of OpUntyped*AccessChainKHR. Treat the cast as |
| 282 | // redundant only when the source already carries the same element type. |
| 283 | // Otherwise keep a distinct register so the element type is preserved. |
| 284 | bool Redundant = |
| 285 | AssignedPtrType->getOpcode() == SPIRV::OpTypeUntypedPointerKHR |
| 286 | ? GR->getUntypedPtrElementType(Reg: Source) == |
| 287 | GR->getOrCreateSPIRVType(Type: ElemTy, MIRBuilder&: MIB, |
| 288 | AQ: SPIRV::AccessQualifier::ReadWrite, |
| 289 | /*EmitIR=*/true) |
| 290 | : GR->getSPIRVTypeForVReg(VReg: Source) == AssignedPtrType; |
| 291 | if (Redundant) { |
| 292 | // Erase Def's assign type instruction if we are going to replace Def. |
| 293 | if (MachineInstr *AssignMI = findAssignTypeInstr(Reg: Def, MRI)) |
| 294 | ToErase.push_back(Elt: AssignMI); |
| 295 | MRI->replaceRegWith(FromReg: Def, ToReg: Source); |
| 296 | } else { |
| 297 | if (!GR->getSPIRVTypeForVReg(VReg: Def, MF: &MF)) |
| 298 | GR->assignSPIRVTypeToVReg(Type: AssignedPtrType, VReg: Def, MF); |
| 299 | MIB.buildBitcast(Dst: Def, Src: Source); |
| 300 | } |
| 301 | } |
| 302 | } |
| 303 | for (MachineInstr *MI : ToErase) |
| 304 | invalidateAndEraseMI(GR, MI); |
| 305 | } |
| 306 | |
| 307 | // Translating GV, IRTranslator sometimes generates following IR: |
| 308 | // %1 = G_GLOBAL_VALUE |
| 309 | // %2 = COPY %1 |
| 310 | // %3 = G_ADDRSPACE_CAST %2 |
| 311 | // |
| 312 | // or |
| 313 | // |
| 314 | // %1 = G_ZEXT %2 |
| 315 | // G_MEMCPY ... %2 ... |
| 316 | // |
| 317 | // New registers have no SPIRV type and no register class info. |
| 318 | // |
| 319 | // Set SPIRV type for GV, propagate it from GV to other instructions, |
| 320 | // also set register classes. |
| 321 | static SPIRVTypeInst propagateSPIRVType(MachineInstr *MI, |
| 322 | SPIRVGlobalRegistry *GR, |
| 323 | MachineRegisterInfo &MRI, |
| 324 | MachineIRBuilder &MIB) { |
| 325 | SPIRVTypeInst SpvType = nullptr; |
| 326 | assert(MI && "Machine instr is expected" ); |
| 327 | if (MI->getOperand(i: 0).isReg()) { |
| 328 | Register Reg = MI->getOperand(i: 0).getReg(); |
| 329 | SpvType = GR->getSPIRVTypeForVReg(VReg: Reg); |
| 330 | if (!SpvType) { |
| 331 | switch (MI->getOpcode()) { |
| 332 | case TargetOpcode::G_FCONSTANT: |
| 333 | case TargetOpcode::G_CONSTANT: { |
| 334 | MIB.setInsertPt(MBB&: *MI->getParent(), II: MI); |
| 335 | Type *Ty = MI->getOperand(i: 1).getCImm()->getType(); |
| 336 | SpvType = GR->getOrCreateSPIRVType( |
| 337 | Type: Ty, MIRBuilder&: MIB, AQ: SPIRV::AccessQualifier::ReadWrite, EmitIR: true); |
| 338 | break; |
| 339 | } |
| 340 | case TargetOpcode::G_GLOBAL_VALUE: { |
| 341 | MIB.setInsertPt(MBB&: *MI->getParent(), II: MI); |
| 342 | const GlobalValue *Global = MI->getOperand(i: 1).getGlobal(); |
| 343 | Type *ElementTy = toTypedPointer(Ty: GR->getDeducedGlobalValueType(Global)); |
| 344 | unsigned AddrSpace = Global->getType()->getAddressSpace(); |
| 345 | // Function pointers use CodeSectionINTEL storage class in SPIR-V when |
| 346 | // the SPV_INTEL_function_pointers extension is enabled. |
| 347 | const SPIRVSubtarget &ST = MIB.getMF().getSubtarget<SPIRVSubtarget>(); |
| 348 | if (isa<Function>(Val: Global) && |
| 349 | ST.canUseExtension(E: SPIRV::Extension::SPV_INTEL_function_pointers)) |
| 350 | AddrSpace = |
| 351 | storageClassToAddressSpace(SC: SPIRV::StorageClass::CodeSectionINTEL); |
| 352 | auto *Ty = TypedPointerType::get(ElementType: ElementTy, AddressSpace: AddrSpace); |
| 353 | SpvType = GR->getOrCreateSPIRVType( |
| 354 | Type: Ty, MIRBuilder&: MIB, AQ: SPIRV::AccessQualifier::ReadWrite, EmitIR: true); |
| 355 | break; |
| 356 | } |
| 357 | case TargetOpcode::G_ANYEXT: |
| 358 | case TargetOpcode::G_SEXT: |
| 359 | case TargetOpcode::G_ZEXT: { |
| 360 | if (MI->getOperand(i: 1).isReg()) { |
| 361 | if (MachineInstr *DefInstr = |
| 362 | MRI.getVRegDef(Reg: MI->getOperand(i: 1).getReg())) { |
| 363 | if (SPIRVTypeInst Def = |
| 364 | propagateSPIRVType(MI: DefInstr, GR, MRI, MIB)) { |
| 365 | unsigned CurrentBW = GR->getScalarOrVectorBitWidth(Type: Def); |
| 366 | unsigned ExpectedBW = |
| 367 | std::max(a: MRI.getType(Reg).getScalarSizeInBits(), b: CurrentBW); |
| 368 | unsigned NumElements = GR->getScalarOrVectorComponentCount(Type: Def); |
| 369 | SpvType = GR->getOrCreateSPIRVIntegerType(BitWidth: ExpectedBW, MIRBuilder&: MIB); |
| 370 | if (NumElements > 1) |
| 371 | SpvType = GR->getOrCreateSPIRVVectorType(BaseType: SpvType, NumElements, |
| 372 | MIRBuilder&: MIB, EmitIR: true); |
| 373 | } |
| 374 | } |
| 375 | } |
| 376 | break; |
| 377 | } |
| 378 | case TargetOpcode::G_PTRTOINT: |
| 379 | SpvType = GR->getOrCreateSPIRVIntegerType( |
| 380 | BitWidth: MRI.getType(Reg).getScalarSizeInBits(), MIRBuilder&: MIB); |
| 381 | break; |
| 382 | case TargetOpcode::G_TRUNC: |
| 383 | case TargetOpcode::G_ADDRSPACE_CAST: |
| 384 | case TargetOpcode::G_PTR_ADD: |
| 385 | case TargetOpcode::COPY: { |
| 386 | MachineOperand &Op = MI->getOperand(i: 1); |
| 387 | MachineInstr *Def = Op.isReg() ? MRI.getVRegDef(Reg: Op.getReg()) : nullptr; |
| 388 | if (Def) |
| 389 | SpvType = propagateSPIRVType(MI: Def, GR, MRI, MIB); |
| 390 | break; |
| 391 | } |
| 392 | default: |
| 393 | break; |
| 394 | } |
| 395 | if (SpvType) { |
| 396 | // check if the address space needs correction |
| 397 | LLT RegType = MRI.getType(Reg); |
| 398 | if (SpvType.isPointer() && RegType.isPointer() && |
| 399 | storageClassToAddressSpace(SC: GR->getPointerStorageClass(Type: SpvType)) != |
| 400 | RegType.getAddressSpace()) { |
| 401 | // Don't correct CodeSectionINTEL back to Function for function |
| 402 | // pointer G_GLOBAL_VALUE - the LLVM register has address space 0 |
| 403 | // but the SPIR-V type was intentionally set to CodeSectionINTEL. |
| 404 | bool SkipCorrection = |
| 405 | MI->getOpcode() == TargetOpcode::G_GLOBAL_VALUE && |
| 406 | GR->getPointerStorageClass(Type: SpvType) == |
| 407 | SPIRV::StorageClass::CodeSectionINTEL; |
| 408 | if (!SkipCorrection) { |
| 409 | const SPIRVSubtarget &ST = |
| 410 | MI->getParent()->getParent()->getSubtarget<SPIRVSubtarget>(); |
| 411 | auto TSC = |
| 412 | addressSpaceToStorageClass(AddrSpace: RegType.getAddressSpace(), STI: ST); |
| 413 | SpvType = GR->changePointerStorageClass(PtrType: SpvType, SC: TSC, I&: *MI); |
| 414 | } |
| 415 | } |
| 416 | GR->assignSPIRVTypeToVReg(Type: SpvType, VReg: Reg, MF: MIB.getMF()); |
| 417 | } |
| 418 | if (!MRI.getRegClassOrNull(Reg)) |
| 419 | MRI.setRegClass(Reg, RC: SpvType ? GR->getRegClass(SpvType) |
| 420 | : &SPIRV::iIDRegClass); |
| 421 | } |
| 422 | } |
| 423 | return SpvType; |
| 424 | } |
| 425 | |
| 426 | // To support current approach and limitations wrt. bit width here we widen a |
| 427 | // scalar register with a bit width greater than 1 to valid sizes and cap it to |
| 428 | // 128 width. |
| 429 | static unsigned widenBitWidthToNextPow2(unsigned BitWidth) { |
| 430 | if (BitWidth == 1) |
| 431 | return 1; // No need to widen 1-bit values |
| 432 | return std::min(a: std::max<unsigned>(a: PowerOf2Ceil(A: BitWidth), b: 8u), b: 128u); |
| 433 | } |
| 434 | |
| 435 | static std::optional<unsigned> |
| 436 | getNarrowScalarWidth(Register Reg, const MachineRegisterInfo &MRI) { |
| 437 | LLT Ty = MRI.getType(Reg); |
| 438 | if (!Ty.isScalar()) |
| 439 | return std::nullopt; |
| 440 | unsigned W = Ty.getScalarSizeInBits(); |
| 441 | // <= and not == because widenBitWidthToNextPow2 caps at 128. |
| 442 | if (widenBitWidthToNextPow2(BitWidth: W) <= W) |
| 443 | return std::nullopt; |
| 444 | return W; |
| 445 | } |
| 446 | |
| 447 | static void widenScalarType(Register Reg, MachineRegisterInfo &MRI) { |
| 448 | LLT RegType = MRI.getType(Reg); |
| 449 | if (!RegType.isScalar()) |
| 450 | return; |
| 451 | unsigned CurrentWidth = RegType.getScalarSizeInBits(); |
| 452 | unsigned NewWidth = widenBitWidthToNextPow2(BitWidth: CurrentWidth); |
| 453 | if (NewWidth != CurrentWidth) |
| 454 | MRI.setType(VReg: Reg, Ty: LLT::scalar(SizeInBits: NewWidth)); |
| 455 | } |
| 456 | |
| 457 | static void widenCImmType(MachineOperand &MOP) { |
| 458 | const ConstantInt *CImmVal = MOP.getCImm(); |
| 459 | unsigned CurrentWidth = CImmVal->getBitWidth(); |
| 460 | unsigned NewWidth = widenBitWidthToNextPow2(BitWidth: CurrentWidth); |
| 461 | if (NewWidth != CurrentWidth) { |
| 462 | // Replace the immediate value with the widened version |
| 463 | MOP.setCImm(ConstantInt::get(Context&: CImmVal->getType()->getContext(), |
| 464 | V: CImmVal->getValue().zextOrTrunc(width: NewWidth))); |
| 465 | } |
| 466 | } |
| 467 | |
| 468 | static void setInsertPtAfterDef(MachineIRBuilder &MIB, MachineInstr *Def) { |
| 469 | MachineBasicBlock &MBB = *Def->getParent(); |
| 470 | MachineBasicBlock::iterator DefIt = |
| 471 | Def->getNextNode() ? Def->getNextNode()->getIterator() : MBB.end(); |
| 472 | // Skip all the PHI and debug instructions. |
| 473 | while (DefIt != MBB.end() && |
| 474 | (DefIt->isPHI() || DefIt->isDebugOrPseudoInstr())) |
| 475 | DefIt = std::next(x: DefIt); |
| 476 | MIB.setInsertPt(MBB, II: DefIt); |
| 477 | } |
| 478 | |
| 479 | namespace llvm { |
| 480 | void updateRegType(Register Reg, Type *Ty, SPIRVTypeInst SpvType, |
| 481 | SPIRVGlobalRegistry *GR, MachineIRBuilder &MIB, |
| 482 | MachineRegisterInfo &MRI) { |
| 483 | assert((Ty || SpvType) && "Either LLVM or SPIRV type is expected." ); |
| 484 | MachineInstr *Def = MRI.getVRegDef(Reg); |
| 485 | setInsertPtAfterDef(MIB, Def); |
| 486 | if (!SpvType) |
| 487 | SpvType = GR->getOrCreateSPIRVType(Type: Ty, MIRBuilder&: MIB, |
| 488 | AQ: SPIRV::AccessQualifier::ReadWrite, EmitIR: true); |
| 489 | if (!MRI.getRegClassOrNull(Reg)) |
| 490 | MRI.setRegClass(Reg, RC: GR->getRegClass(SpvType)); |
| 491 | if (!MRI.getType(Reg).isValid()) |
| 492 | MRI.setType(VReg: Reg, Ty: GR->getRegType(SpvType)); |
| 493 | GR->assignSPIRVTypeToVReg(Type: SpvType, VReg: Reg, MF: MIB.getMF()); |
| 494 | } |
| 495 | |
| 496 | void processInstr(MachineInstr &MI, MachineIRBuilder &MIB, |
| 497 | MachineRegisterInfo &MRI, SPIRVGlobalRegistry *GR, |
| 498 | SPIRVTypeInst KnownResType) { |
| 499 | MIB.setInsertPt(MBB&: *MI.getParent(), II: MI.getIterator()); |
| 500 | for (auto &Op : MI.operands()) { |
| 501 | if (!Op.isReg() || Op.isDef()) |
| 502 | continue; |
| 503 | Register OpReg = Op.getReg(); |
| 504 | SPIRVTypeInst SpvType = GR->getSPIRVTypeForVReg(VReg: OpReg); |
| 505 | if (!SpvType && KnownResType) { |
| 506 | SpvType = KnownResType; |
| 507 | GR->assignSPIRVTypeToVReg(Type: KnownResType, VReg: OpReg, MF: *MI.getMF()); |
| 508 | } |
| 509 | assert(SpvType); |
| 510 | if (!MRI.getRegClassOrNull(Reg: OpReg)) |
| 511 | MRI.setRegClass(Reg: OpReg, RC: GR->getRegClass(SpvType)); |
| 512 | if (!MRI.getType(Reg: OpReg).isValid()) |
| 513 | MRI.setType(VReg: OpReg, Ty: GR->getRegType(SpvType)); |
| 514 | } |
| 515 | } |
| 516 | } // namespace llvm |
| 517 | |
| 518 | // Sign-sensitive integer ops: their result depends on the value of the input |
| 519 | // sign bit at position (width-1). On sub-pow2 widths the general widening |
| 520 | // loop is a pure LLT relabel, which leaves the sign bit at the *original* |
| 521 | // position instead of the widened MSB. These ops therefore need an explicit |
| 522 | // G_SEXT_INREG on each value operand to move the sign bit up. |
| 523 | // |
| 524 | // Signed-vs-unsigned G_ICMP is distinguished by its predicate operand. |
| 525 | // |
| 526 | // TODO: follow-up PRs will add the remaining sign-sensitive opcodes |
| 527 | // (e.g. G_SMIN/G_SMAX, G_SADDSAT/G_SSUBSAT, signed overflow ops). |
| 528 | static bool isSignSensitiveOp(const MachineInstr &MI) { |
| 529 | switch (MI.getOpcode()) { |
| 530 | case TargetOpcode::G_ASHR: |
| 531 | case TargetOpcode::G_SDIV: |
| 532 | case TargetOpcode::G_SREM: |
| 533 | return true; |
| 534 | case TargetOpcode::G_ICMP: |
| 535 | return CmpInst::isSigned( |
| 536 | Pred: static_cast<CmpInst::Predicate>(MI.getOperand(i: 1).getPredicate())); |
| 537 | default: |
| 538 | return false; |
| 539 | } |
| 540 | } |
| 541 | |
| 542 | struct NarrowWideningInfo { |
| 543 | // Width before widening of each sign-sensitive value-operand vreg (one entry |
| 544 | // per vreg). |
| 545 | DenseMap<Register, unsigned> OrigWidth; |
| 546 | // Sign-sensitive ops whose value operand(s) need replacing, ordered for |
| 547 | // reproducible vreg numbering. |
| 548 | SmallVector<MachineInstr *> SignSensitiveWorklist; |
| 549 | // Keyed by instruction, not vreg: G_TRUNC handling can replace the source. |
| 550 | SmallVector<std::pair<MachineInstr *, unsigned>> BitCountWorklist; |
| 551 | }; |
| 552 | |
| 553 | // G_CTTZ_ZERO_POISON is absent because its low bits are known non-zero, G_CTLS |
| 554 | // because the backend does not select it. |
| 555 | static bool isWidthSensitiveBitCountOp(unsigned Opcode) { |
| 556 | switch (Opcode) { |
| 557 | case TargetOpcode::G_CTLZ: |
| 558 | case TargetOpcode::G_CTLZ_ZERO_POISON: |
| 559 | case TargetOpcode::G_CTTZ: |
| 560 | case TargetOpcode::G_CTPOP: |
| 561 | return true; |
| 562 | default: |
| 563 | return false; |
| 564 | } |
| 565 | } |
| 566 | |
| 567 | // Collect ops whose semantics depend on the operand width along with their |
| 568 | // pre-widening widths, before later passes retype those vregs to pow2 LLTs |
| 569 | // and the original width is no longer recoverable. |
| 570 | static NarrowWideningInfo |
| 571 | recordNarrowOperandWidths(MachineFunction &MF, const MachineRegisterInfo &MRI) { |
| 572 | NarrowWideningInfo Info; |
| 573 | auto RecordIfNarrow = [&](Register Reg) { |
| 574 | std::optional<unsigned> W = getNarrowScalarWidth(Reg, MRI); |
| 575 | if (!W) |
| 576 | return false; |
| 577 | Info.OrigWidth.try_emplace(Key: Reg, Args&: *W); |
| 578 | return true; |
| 579 | }; |
| 580 | for (MachineBasicBlock &MBB : MF) { |
| 581 | for (MachineInstr &MI : MBB) { |
| 582 | if (isWidthSensitiveBitCountOp(Opcode: MI.getOpcode())) { |
| 583 | if (std::optional<unsigned> W = |
| 584 | getNarrowScalarWidth(Reg: MI.getOperand(i: 1).getReg(), MRI)) |
| 585 | Info.BitCountWorklist.emplace_back(Args: &MI, Args&: *W); |
| 586 | continue; |
| 587 | } |
| 588 | if (!isSignSensitiveOp(MI)) |
| 589 | continue; |
| 590 | // Value operands are the trailing two, past any def or predicate. |
| 591 | unsigned N = MI.getNumOperands(); |
| 592 | const MachineOperand &LHS = MI.getOperand(i: N - 2); |
| 593 | const MachineOperand &RHS = MI.getOperand(i: N - 1); |
| 594 | // Sign-sensitive opcodes carry register operands only. |
| 595 | assert(LHS.isReg() && RHS.isReg()); |
| 596 | bool NeedsRewrite = RecordIfNarrow(LHS.getReg()); |
| 597 | NeedsRewrite = RecordIfNarrow(RHS.getReg()) || NeedsRewrite; |
| 598 | if (NeedsRewrite) |
| 599 | Info.SignSensitiveWorklist.push_back(Elt: &MI); |
| 600 | } |
| 601 | } |
| 602 | return Info; |
| 603 | } |
| 604 | |
| 605 | // For every recorded sign-sensitive op, insert G_SEXT_INREG on each value |
| 606 | // operand whose original width was narrower than the widened pow2 width and |
| 607 | // retype the operand's vreg LLT in place to the widened width. |
| 608 | // |
| 609 | // Info must have been populated by recordNarrowOperandWidths before |
| 610 | // other passes retyped the vregs; otherwise the narrow widths needed here |
| 611 | // are lost. |
| 612 | // |
| 613 | // TODO: handle vector operands. |
| 614 | static void widenSignSensitiveOps(MachineFunction &MF, SPIRVGlobalRegistry *GR, |
| 615 | MachineIRBuilder &MIB, |
| 616 | MachineRegisterInfo &MRI, |
| 617 | const NarrowWideningInfo &Info) { |
| 618 | // Emit G_SEXT_INREG from Reg's recorded narrow width; retypes Reg to the |
| 619 | // widened width and returns the sign-extended vreg. |
| 620 | auto SignExtendReg = [&](Register Reg, unsigned OldW, |
| 621 | MachineInstr &MI) -> Register { |
| 622 | unsigned NewW = widenBitWidthToNextPow2(BitWidth: OldW); |
| 623 | LLT NewLLT = LLT::scalar(SizeInBits: NewW); |
| 624 | MIB.setInsertPt(MBB&: *MI.getParent(), II: MI.getIterator()); |
| 625 | SPIRVTypeInst SpvTy = GR->getOrCreateSPIRVIntegerType(BitWidth: NewW, MIRBuilder&: MIB); |
| 626 | Register SExted = MRI.createGenericVirtualRegister(Ty: NewLLT); |
| 627 | GR->assignSPIRVTypeToVReg(Type: SpvTy, VReg: SExted, MF); |
| 628 | MRI.setRegClass(Reg: SExted, RC: GR->getRegClass(SpvType: SpvTy)); |
| 629 | MRI.setType(VReg: Reg, Ty: NewLLT); |
| 630 | MIB.buildSExtInReg(Res: SExted, Op: Reg, ImmOp: OldW); |
| 631 | return SExted; |
| 632 | }; |
| 633 | |
| 634 | // TODO: when the same narrow vreg feeds multiple sign-sensitive ops (e.g. |
| 635 | // sdiv %x, %y and srem %x, %y), emit one shared G_SEXT_INREG instead of one |
| 636 | // per use. |
| 637 | for (MachineInstr *MI : Info.SignSensitiveWorklist) { |
| 638 | unsigned N = MI->getNumOperands(); |
| 639 | MachineOperand &LHS = MI->getOperand(i: N - 2); |
| 640 | MachineOperand &RHS = MI->getOperand(i: N - 1); |
| 641 | Register LHSReg = LHS.getReg(); |
| 642 | Register RHSReg = RHS.getReg(); |
| 643 | if (auto It = Info.OrigWidth.find(Val: LHSReg); It != Info.OrigWidth.end()) |
| 644 | LHS.setReg(SignExtendReg(LHSReg, It->second, *MI)); |
| 645 | // Same vreg on both sides (e.g. G_ICMP slt %x, %x): reuse the sext just |
| 646 | // emitted for LHS instead of emitting a second one. |
| 647 | if (RHSReg == LHSReg) { |
| 648 | RHS.setReg(LHS.getReg()); |
| 649 | continue; |
| 650 | } |
| 651 | if (auto It = Info.OrigWidth.find(Val: RHSReg); It != Info.OrigWidth.end()) |
| 652 | RHS.setReg(SignExtendReg(RHSReg, It->second, *MI)); |
| 653 | } |
| 654 | } |
| 655 | |
| 656 | // LegalizerHelper::widenScalar has the same cases but cannot be reached: the |
| 657 | // relabel retypes every narrow scalar to a pow2 LLT, so no illegal narrow type |
| 658 | // ever reaches the legalizer. |
| 659 | // |
| 660 | // TODO: handle vector operands. |
| 661 | static void widenBitCountOps(SPIRVGlobalRegistry *GR, MachineIRBuilder &MIB, |
| 662 | MachineRegisterInfo &MRI, |
| 663 | const NarrowWideningInfo &Info) { |
| 664 | for (auto [MI, OldWidth] : Info.BitCountWorklist) { |
| 665 | Register SrcReg = MI->getOperand(i: 1).getReg(); |
| 666 | unsigned NewWidth = widenBitWidthToNextPow2(BitWidth: OldWidth); |
| 667 | LLT NewTy = LLT::scalar(SizeInBits: NewWidth); |
| 668 | widenScalarType(Reg: SrcReg, MRI); |
| 669 | MIB.setInstrAndDebugLoc(*MI); |
| 670 | SPIRVTypeInst SpvTy = GR->getOrCreateSPIRVIntegerType(BitWidth: NewWidth, MIRBuilder&: MIB); |
| 671 | |
| 672 | // The G_TRUNC lowering masks its result to the narrow width, so a source |
| 673 | // coming from it needs no second mask. |
| 674 | APInt Cst; |
| 675 | bool HighBitsAlreadyZero = |
| 676 | mi_match(R: SrcReg, MRI, P: m_GAnd(L: m_Reg(), R: m_ICst(Cst))) && |
| 677 | Cst.isSubsetOf(RHS: APInt::getLowBitsSet(numBits: Cst.getBitWidth(), loBitsSet: OldWidth)); |
| 678 | auto ClearHighBits = [&](unsigned Width) -> Register { |
| 679 | if (HighBitsAlreadyZero) |
| 680 | return SrcReg; |
| 681 | Register Masked = createVirtualRegister(SpvType: SpvTy, GR, MIRBuilder&: MIB); |
| 682 | MIB.buildZExtInReg(Res: Masked, Op: SrcReg, ImmOp: Width); |
| 683 | return Masked; |
| 684 | }; |
| 685 | |
| 686 | Register Input; |
| 687 | switch (MI->getOpcode()) { |
| 688 | case TargetOpcode::G_CTLZ_ZERO_POISON: { |
| 689 | // Shifting up to the widened MSB moves the poison out too, so no |
| 690 | // adjustment. |
| 691 | Input = createVirtualRegister(SpvType: SpvTy, GR, MIRBuilder&: MIB); |
| 692 | auto Diff = MIB.buildConstant(Res: NewTy, Val: NewWidth - OldWidth); |
| 693 | MIB.buildShl(Dst: Input, Src0: SrcReg, Src1: Diff); |
| 694 | break; |
| 695 | } |
| 696 | case TargetOpcode::G_CTTZ: { |
| 697 | // Keeps an all-zero narrow value counting exactly OldWidth zeros. |
| 698 | Input = createVirtualRegister(SpvType: SpvTy, GR, MIRBuilder&: MIB); |
| 699 | auto TopBit = |
| 700 | MIB.buildConstant(Res: NewTy, Val: APInt::getOneBitSet(numBits: NewWidth, BitNo: OldWidth)); |
| 701 | MIB.buildOr(Dst: Input, Src0: SrcReg, Src1: TopBit); |
| 702 | break; |
| 703 | } |
| 704 | case TargetOpcode::G_CTPOP: |
| 705 | Input = ClearHighBits(OldWidth); |
| 706 | break; |
| 707 | case TargetOpcode::G_CTLZ: { |
| 708 | // Clearing the extra bits adds leading zeros the count has to drop. |
| 709 | Input = ClearHighBits(OldWidth); |
| 710 | Register DstReg = MI->getOperand(i: 0).getReg(); |
| 711 | widenScalarType(Reg: DstReg, MRI); |
| 712 | Register Count = createVirtualRegister(SpvType: SpvTy, GR, MIRBuilder&: MIB); |
| 713 | MI->getOperand(i: 0).setReg(Count); |
| 714 | setInsertPtAfterDef(MIB, Def: MI); |
| 715 | auto Diff = MIB.buildConstant(Res: NewTy, Val: NewWidth - OldWidth); |
| 716 | MIB.buildSub(Dst: DstReg, Src0: Count, Src1: Diff); |
| 717 | break; |
| 718 | } |
| 719 | default: |
| 720 | llvm_unreachable("unexpected width-sensitive bit-count opcode" ); |
| 721 | } |
| 722 | MI->getOperand(i: 1).setReg(Input); |
| 723 | } |
| 724 | } |
| 725 | |
| 726 | static void |
| 727 | generateAssignInstrs(MachineFunction &MF, SPIRVGlobalRegistry *GR, |
| 728 | MachineIRBuilder MIB, |
| 729 | DenseMap<MachineInstr *, Type *> &TargetExtConstTypes) { |
| 730 | // Get access to information about available extensions |
| 731 | const SPIRVSubtarget *ST = |
| 732 | static_cast<const SPIRVSubtarget *>(&MIB.getMF().getSubtarget()); |
| 733 | |
| 734 | MachineRegisterInfo &MRI = MF.getRegInfo(); |
| 735 | SmallVector<MachineInstr *, 10> ToErase; |
| 736 | DenseMap<MachineInstr *, Register> RegsAlreadyAddedToDT; |
| 737 | |
| 738 | bool IsExtendedInts = |
| 739 | ST->canUseExtension( |
| 740 | E: SPIRV::Extension::SPV_ALTERA_arbitrary_precision_integers) || |
| 741 | ST->canUseExtension(E: SPIRV::Extension::SPV_KHR_bit_instructions) || |
| 742 | ST->canUseExtension(E: SPIRV::Extension::SPV_INTEL_int4); |
| 743 | |
| 744 | if (!IsExtendedInts) { |
| 745 | // Without arbitrary precision integer extensions, SPIR-V only supports |
| 746 | // integer widths of 8, 16, 32, 64. Non-standard widths (e.g., i24, i40) |
| 747 | // must be widened to the next power of two. |
| 748 | // |
| 749 | // Record the original widths of width-sensitive operands before either |
| 750 | // the G_TRUNC handling or the general widening loop retypes vregs, then |
| 751 | // rewrite those ops after G_TRUNC processing using the recorded widths. |
| 752 | NarrowWideningInfo WideningInfo = recordNarrowOperandWidths(MF, MRI); |
| 753 | |
| 754 | // G_TRUNC requires special handling because its semantics depend on the |
| 755 | // original destination width. For example: |
| 756 | // %dst:s24 = G_TRUNC %src:s64 |
| 757 | // After widening s24 to s32, we cannot simply do: |
| 758 | // %dst:s32 = G_TRUNC %src:s64 |
| 759 | // because this would keep 32 bits instead of 24. Instead, we insert a |
| 760 | // G_AND to mask the value to the original width: |
| 761 | // %mask:s64 = G_CONSTANT 0xFFFFFF ; 24-bit mask |
| 762 | // %masked:s64 = G_AND %src:s64, %mask |
| 763 | // %dst:s32 = G_TRUNC %masked:s64 |
| 764 | // If src and dst widen to the same size, G_TRUNC is replaced entirely: |
| 765 | // %mask:s64 = G_CONSTANT 0xFFFFFFFFFF ; 40-bit mask |
| 766 | // %dst:s64 = G_AND %src:s64, %mask |
| 767 | SmallVector<MachineInstr *, 8> TruncToRemove; |
| 768 | for (MachineBasicBlock &MBB : MF) { |
| 769 | for (MachineInstr &MI : MBB) { |
| 770 | unsigned MIOp = MI.getOpcode(); |
| 771 | if (MIOp != TargetOpcode::G_TRUNC) |
| 772 | continue; |
| 773 | assert(MI.getNumOperands() == 2); |
| 774 | assert(MI.getOperand(0).isReg()); |
| 775 | assert(MI.getOperand(1).isReg()); |
| 776 | |
| 777 | Register DstReg = MI.getOperand(i: 0).getReg(); |
| 778 | Register SrcReg = MI.getOperand(i: 1).getReg(); |
| 779 | |
| 780 | LLT DstTy = MRI.getType(Reg: DstReg); |
| 781 | LLT SrcTy = MRI.getType(Reg: SrcReg); |
| 782 | assert((DstTy.isScalar() || DstTy.isVector()) && |
| 783 | (SrcTy.isScalar() || SrcTy.isVector()) && |
| 784 | "Expected scalar or vector G_TRUNC types" ); |
| 785 | assert(DstTy.isVector() == SrcTy.isVector() && |
| 786 | "Expected matching scalar/vector G_TRUNC types" ); |
| 787 | assert((!DstTy.isVector() || |
| 788 | DstTy.getElementCount() == SrcTy.getElementCount()) && |
| 789 | "Expected equal vector element counts" ); |
| 790 | |
| 791 | unsigned OriginalDstWidth = DstTy.getScalarSizeInBits(); |
| 792 | unsigned OriginalSrcWidth = SrcTy.getScalarSizeInBits(); |
| 793 | |
| 794 | unsigned NewDstWidth = widenBitWidthToNextPow2(BitWidth: OriginalDstWidth); |
| 795 | unsigned NewSrcWidth = widenBitWidthToNextPow2(BitWidth: OriginalSrcWidth); |
| 796 | LLT NewDstTy = DstTy.changeElementSize(NewEltSize: NewDstWidth); |
| 797 | LLT NewSrcTy = SrcTy.changeElementSize(NewEltSize: NewSrcWidth); |
| 798 | |
| 799 | // No Dst width change means no truncation semantics change, but the |
| 800 | // source still needs a legal type. |
| 801 | if (OriginalDstWidth == NewDstWidth) { |
| 802 | MRI.setType(VReg: SrcReg, Ty: NewSrcTy); |
| 803 | continue; |
| 804 | } |
| 805 | |
| 806 | MRI.setType(VReg: SrcReg, Ty: NewSrcTy); |
| 807 | MRI.setType(VReg: DstReg, Ty: NewDstTy); |
| 808 | |
| 809 | MIB.setInsertPt(MBB, II: MI.getIterator()); |
| 810 | APInt Mask = APInt::getLowBitsSet(numBits: NewSrcWidth, loBitsSet: OriginalDstWidth); |
| 811 | MachineInstrBuilder MaskReg = |
| 812 | DstTy.isVector() |
| 813 | ? MIB.buildBuildVectorConstant( |
| 814 | Res: NewSrcTy, |
| 815 | Ops: SmallVector<APInt, 4>(DstTy.getNumElements(), Mask)) |
| 816 | : MIB.buildConstant(Res: NewSrcTy, Val: Mask); |
| 817 | Register MaskedReg = MRI.createGenericVirtualRegister(Ty: NewSrcTy); |
| 818 | MIB.buildAnd(Dst: MaskedReg, Src0: SrcReg, Src1: MaskReg); |
| 819 | |
| 820 | if (NewSrcWidth == NewDstWidth) { |
| 821 | // Rekey OrigWidth from DstReg to MaskedReg so widenSignSensitiveOps |
| 822 | // still sees the narrow original width after replaceRegWith. |
| 823 | if (auto It = WideningInfo.OrigWidth.find(Val: DstReg); |
| 824 | It != WideningInfo.OrigWidth.end()) { |
| 825 | unsigned W = It->second; |
| 826 | WideningInfo.OrigWidth.erase(I: It); |
| 827 | WideningInfo.OrigWidth.try_emplace(Key: MaskedReg, Args&: W); |
| 828 | } |
| 829 | MRI.replaceRegWith(FromReg: DstReg, ToReg: MaskedReg); |
| 830 | TruncToRemove.push_back(Elt: &MI); |
| 831 | } else { |
| 832 | MI.getOperand(i: 1).setReg(MaskedReg); |
| 833 | } |
| 834 | } |
| 835 | } |
| 836 | for (MachineInstr *MI : TruncToRemove) |
| 837 | MI->eraseFromParent(); |
| 838 | |
| 839 | widenSignSensitiveOps(MF, GR, MIB, MRI, Info: WideningInfo); |
| 840 | widenBitCountOps(GR, MIB, MRI, Info: WideningInfo); |
| 841 | } |
| 842 | |
| 843 | for (MachineBasicBlock *MBB : post_order(G: &MF)) { |
| 844 | if (MBB->empty()) |
| 845 | continue; |
| 846 | |
| 847 | bool ReachedBegin = false; |
| 848 | for (auto MII = std::prev(x: MBB->end()), Begin = MBB->begin(); |
| 849 | !ReachedBegin;) { |
| 850 | MachineInstr &MI = *MII; |
| 851 | unsigned MIOp = MI.getOpcode(); |
| 852 | |
| 853 | if (!IsExtendedInts) { |
| 854 | // validate bit width of scalar registers and constant immediates |
| 855 | for (auto &MOP : MI.operands()) { |
| 856 | if (MOP.isReg()) |
| 857 | widenScalarType(Reg: MOP.getReg(), MRI); |
| 858 | else if (MOP.isCImm()) |
| 859 | widenCImmType(MOP); |
| 860 | } |
| 861 | } |
| 862 | |
| 863 | if (isSpvIntrinsic(MI, IntrinsicID: Intrinsic::spv_assign_ptr_type)) { |
| 864 | Register Reg = MI.getOperand(i: 1).getReg(); |
| 865 | MIB.setInsertPt(MBB&: *MI.getParent(), II: MI.getIterator()); |
| 866 | Type *ElementTy = getMDOperandAsType(N: MI.getOperand(i: 2).getMetadata(), I: 0); |
| 867 | auto SC = addressSpaceToStorageClass(AddrSpace: MI.getOperand(i: 3).getImm(), STI: *ST); |
| 868 | if (SC == SPIRV::StorageClass::Function && |
| 869 | isa<FunctionType>(Val: ElementTy) && |
| 870 | ST->canUseExtension(E: SPIRV::Extension::SPV_INTEL_function_pointers)) |
| 871 | SC = SPIRV::StorageClass::CodeSectionINTEL; |
| 872 | SPIRVTypeInst AssignedPtrType = |
| 873 | GR->getOrCreateSPIRVPointerType(BaseType: ElementTy, I&: MI, SC); |
| 874 | |
| 875 | // For untyped pointers, store the element type for later use. |
| 876 | if (ST->canUseExtension(E: SPIRV::Extension::SPV_KHR_untyped_pointers) && |
| 877 | !ST->isShader()) { |
| 878 | SPIRVTypeInst ElemSpvType = GR->getOrCreateSPIRVType( |
| 879 | Type: ElementTy, MIRBuilder&: MIB, AQ: SPIRV::AccessQualifier::ReadWrite, |
| 880 | /*EmitIR=*/true); |
| 881 | GR->setUntypedPtrElementType(Reg, ElemType: ElemSpvType); |
| 882 | } |
| 883 | |
| 884 | // The intrinsic also carries vector-of-pointer values produced by |
| 885 | // scalarized vector GEPs; wrap the pointer in OpTypeVector to match |
| 886 | // the vreg's LLT. |
| 887 | LLT RegTy = MRI.getType(Reg); |
| 888 | if (RegTy.isValid() && RegTy.isVector()) |
| 889 | AssignedPtrType = GR->getOrCreateSPIRVVectorType( |
| 890 | BaseType: AssignedPtrType, NumElements: RegTy.getNumElements(), MIRBuilder&: MIB, |
| 891 | /*EmitIR=*/true); |
| 892 | MachineInstr *Def = MRI.getVRegDef(Reg); |
| 893 | assert(Def && "Expecting an instruction that defines the register" ); |
| 894 | // G_GLOBAL_VALUE already has type info. |
| 895 | if (Def->getOpcode() != TargetOpcode::G_GLOBAL_VALUE) |
| 896 | updateRegType(Reg, Ty: nullptr, SpvType: AssignedPtrType, GR, MIB, |
| 897 | MRI&: MF.getRegInfo()); |
| 898 | ToErase.push_back(Elt: &MI); |
| 899 | } else if (isSpvIntrinsic(MI, IntrinsicID: Intrinsic::spv_assign_type)) { |
| 900 | Register Reg = MI.getOperand(i: 1).getReg(); |
| 901 | Type *Ty = getMDOperandAsType(N: MI.getOperand(i: 2).getMetadata(), I: 0); |
| 902 | MachineInstr *Def = MRI.getVRegDef(Reg); |
| 903 | assert(Def && "Expecting an instruction that defines the register" ); |
| 904 | // G_GLOBAL_VALUE already has type info. |
| 905 | if (Def->getOpcode() != TargetOpcode::G_GLOBAL_VALUE) |
| 906 | updateRegType(Reg, Ty, SpvType: nullptr, GR, MIB, MRI&: MF.getRegInfo()); |
| 907 | if (Def->getOpcode() == TargetOpcode::COPY && isVector1(Ty)) |
| 908 | updateRegType(Reg: passCopy(Def, MRI: &MF.getRegInfo())->getOperand(i: 0).getReg(), |
| 909 | Ty, SpvType: nullptr, GR, MIB, MRI&: MF.getRegInfo()); |
| 910 | ToErase.push_back(Elt: &MI); |
| 911 | } else if (MIOp == TargetOpcode::FAKE_USE && MI.getNumOperands() > 0) { |
| 912 | MachineInstr *MdMI = MI.getPrevNode(); |
| 913 | if (MdMI && isSpvIntrinsic(MI: *MdMI, IntrinsicID: Intrinsic::spv_value_md)) { |
| 914 | // It's an internal service info from before IRTranslator passes. |
| 915 | MachineInstr *Def = getVRegDef(MRI, Reg: MI.getOperand(i: 0).getReg()); |
| 916 | for (unsigned I = 1, E = MI.getNumOperands(); I != E && Def; ++I) |
| 917 | if (getVRegDef(MRI, Reg: MI.getOperand(i: I).getReg()) != Def) |
| 918 | Def = nullptr; |
| 919 | if (Def) { |
| 920 | const MDNode *MD = MdMI->getOperand(i: 1).getMetadata(); |
| 921 | StringRef ValueName = |
| 922 | cast<MDString>(Val: MD->getOperand(I: 1))->getString(); |
| 923 | const MDNode *TypeMD = cast<MDNode>(Val: MD->getOperand(I: 0)); |
| 924 | Type *ValueTy = getMDOperandAsType(N: TypeMD, I: 0); |
| 925 | GR->addValueAttrs(Key: Def, Val: std::make_pair(x&: ValueTy, y: ValueName.str())); |
| 926 | } |
| 927 | ToErase.push_back(Elt: MdMI); |
| 928 | } |
| 929 | ToErase.push_back(Elt: &MI); |
| 930 | } else if (MIOp == TargetOpcode::G_CONSTANT || |
| 931 | MIOp == TargetOpcode::G_FCONSTANT || |
| 932 | MIOp == TargetOpcode::G_BUILD_VECTOR) { |
| 933 | // %rc = G_CONSTANT ty Val |
| 934 | // Ensure %rc has a valid SPIR-V type assigned in the Global Registry. |
| 935 | Register Reg = MI.getOperand(i: 0).getReg(); |
| 936 | bool NeedAssignType = !GR->getSPIRVTypeForVReg(VReg: Reg); |
| 937 | Type *Ty = nullptr; |
| 938 | if (MIOp == TargetOpcode::G_CONSTANT) { |
| 939 | auto TargetExtIt = TargetExtConstTypes.find(Val: &MI); |
| 940 | Ty = TargetExtIt == TargetExtConstTypes.end() |
| 941 | ? MI.getOperand(i: 1).getCImm()->getType() |
| 942 | : TargetExtIt->second; |
| 943 | const ConstantInt *OpCI = MI.getOperand(i: 1).getCImm(); |
| 944 | // TODO: we may wish to analyze here if OpCI is zero and LLT RegType = |
| 945 | // MRI.getType(Reg); RegType.isPointer() is true, so that we observe |
| 946 | // at this point not i64/i32 constant but null pointer in the |
| 947 | // corresponding address space of RegType.getAddressSpace(). This may |
| 948 | // help to successfully validate the case when a OpConstantComposite's |
| 949 | // constituent has type that does not match Result Type of |
| 950 | // OpConstantComposite (see, for example, |
| 951 | // pointers/PtrCast-null-in-OpSpecConstantOp.ll). |
| 952 | Register PrimaryReg = GR->find(V: OpCI, MF: &MF); |
| 953 | if (!PrimaryReg.isValid()) { |
| 954 | GR->add(V: OpCI, MI: &MI); |
| 955 | } else if (PrimaryReg != Reg && |
| 956 | MRI.getType(Reg) == MRI.getType(Reg: PrimaryReg)) { |
| 957 | auto *RCReg = MRI.getRegClassOrNull(Reg); |
| 958 | auto *RCPrimary = MRI.getRegClassOrNull(Reg: PrimaryReg); |
| 959 | if (!RCReg || RCPrimary == RCReg) { |
| 960 | RegsAlreadyAddedToDT[&MI] = PrimaryReg; |
| 961 | ToErase.push_back(Elt: &MI); |
| 962 | NeedAssignType = false; |
| 963 | } |
| 964 | } |
| 965 | } else if (MIOp == TargetOpcode::G_FCONSTANT) { |
| 966 | Ty = MI.getOperand(i: 1).getFPImm()->getType(); |
| 967 | } else { |
| 968 | assert(MIOp == TargetOpcode::G_BUILD_VECTOR); |
| 969 | Type *ElemTy = nullptr; |
| 970 | MachineInstr *ElemMI = MRI.getVRegDef(Reg: MI.getOperand(i: 1).getReg()); |
| 971 | assert(ElemMI); |
| 972 | |
| 973 | if (ElemMI->getOpcode() == TargetOpcode::G_CONSTANT) { |
| 974 | ElemTy = ElemMI->getOperand(i: 1).getCImm()->getType(); |
| 975 | } else if (ElemMI->getOpcode() == TargetOpcode::G_FCONSTANT) { |
| 976 | ElemTy = ElemMI->getOperand(i: 1).getFPImm()->getType(); |
| 977 | } else { |
| 978 | if (SPIRVTypeInst ElemSpvType = |
| 979 | GR->getSPIRVTypeForVReg(VReg: MI.getOperand(i: 1).getReg(), MF: &MF)) |
| 980 | ElemTy = const_cast<Type *>(GR->getTypeForSPIRVType(Ty: ElemSpvType)); |
| 981 | } |
| 982 | if (ElemTy) |
| 983 | Ty = VectorType::get( |
| 984 | ElementType: ElemTy, NumElements: MI.getNumExplicitOperands() - MI.getNumExplicitDefs(), |
| 985 | Scalable: false); |
| 986 | else |
| 987 | NeedAssignType = false; |
| 988 | } |
| 989 | if (NeedAssignType) |
| 990 | updateRegType(Reg, Ty, SpvType: nullptr, GR, MIB, MRI); |
| 991 | } else if (MIOp == TargetOpcode::G_GLOBAL_VALUE) { |
| 992 | propagateSPIRVType(MI: &MI, GR, MRI, MIB); |
| 993 | } |
| 994 | |
| 995 | if (MII == Begin) |
| 996 | ReachedBegin = true; |
| 997 | else |
| 998 | --MII; |
| 999 | } |
| 1000 | } |
| 1001 | for (MachineInstr *MI : ToErase) { |
| 1002 | auto It = RegsAlreadyAddedToDT.find(Val: MI); |
| 1003 | if (It != RegsAlreadyAddedToDT.end()) |
| 1004 | MRI.replaceRegWith(FromReg: MI->getOperand(i: 0).getReg(), ToReg: It->second); |
| 1005 | invalidateAndEraseMI(GR, MI); |
| 1006 | } |
| 1007 | |
| 1008 | // Address the case when IRTranslator introduces instructions with new |
| 1009 | // registers without associated SPIRV type. |
| 1010 | for (MachineBasicBlock &MBB : MF) { |
| 1011 | for (MachineInstr &MI : MBB) { |
| 1012 | switch (MI.getOpcode()) { |
| 1013 | case TargetOpcode::G_TRUNC: |
| 1014 | case TargetOpcode::G_ANYEXT: |
| 1015 | case TargetOpcode::G_SEXT: |
| 1016 | case TargetOpcode::G_ZEXT: |
| 1017 | case TargetOpcode::G_PTRTOINT: |
| 1018 | case TargetOpcode::COPY: |
| 1019 | case TargetOpcode::G_ADDRSPACE_CAST: |
| 1020 | propagateSPIRVType(MI: &MI, GR, MRI, MIB); |
| 1021 | break; |
| 1022 | } |
| 1023 | } |
| 1024 | } |
| 1025 | } |
| 1026 | |
| 1027 | static void processInstrsWithTypeFolding(MachineFunction &MF, |
| 1028 | SPIRVGlobalRegistry *GR, |
| 1029 | MachineIRBuilder MIB) { |
| 1030 | MachineRegisterInfo &MRI = MF.getRegInfo(); |
| 1031 | for (MachineBasicBlock &MBB : MF) |
| 1032 | for (MachineInstr &MI : MBB) |
| 1033 | if (isTypeFoldingSupported(Opcode: MI.getOpcode())) |
| 1034 | processInstr(MI, MIB, MRI, GR, KnownResType: nullptr); |
| 1035 | } |
| 1036 | |
| 1037 | static Register |
| 1038 | collectInlineAsmInstrOperands(MachineInstr *MI, |
| 1039 | SmallVector<unsigned, 4> *Ops = nullptr) { |
| 1040 | Register DefReg; |
| 1041 | unsigned StartOp = InlineAsm::MIOp_FirstOperand, |
| 1042 | AsmDescOp = InlineAsm::MIOp_FirstOperand; |
| 1043 | for (unsigned Idx = StartOp, MISz = MI->getNumOperands(); Idx != MISz; |
| 1044 | ++Idx) { |
| 1045 | const MachineOperand &MO = MI->getOperand(i: Idx); |
| 1046 | if (MO.isMetadata()) |
| 1047 | continue; |
| 1048 | if (Idx == AsmDescOp && MO.isImm()) { |
| 1049 | // compute the index of the next operand descriptor |
| 1050 | const InlineAsm::Flag F(MO.getImm()); |
| 1051 | AsmDescOp += 1 + F.getNumOperandRegisters(); |
| 1052 | continue; |
| 1053 | } |
| 1054 | if (MO.isReg() && MO.isDef()) { |
| 1055 | if (!Ops) |
| 1056 | return MO.getReg(); |
| 1057 | DefReg = MO.getReg(); |
| 1058 | } else if (Ops) { |
| 1059 | Ops->push_back(Elt: Idx); |
| 1060 | } |
| 1061 | } |
| 1062 | return DefReg; |
| 1063 | } |
| 1064 | |
| 1065 | static void |
| 1066 | insertInlineAsmProcess(MachineFunction &MF, SPIRVGlobalRegistry *GR, |
| 1067 | const SPIRVSubtarget &ST, MachineIRBuilder MIRBuilder, |
| 1068 | const SmallVector<MachineInstr *> &ToProcess) { |
| 1069 | MachineRegisterInfo &MRI = MF.getRegInfo(); |
| 1070 | Register AsmTargetReg; |
| 1071 | for (unsigned i = 0, Sz = ToProcess.size(); i + 1 < Sz; i += 2) { |
| 1072 | MachineInstr *I1 = ToProcess[i], *I2 = ToProcess[i + 1]; |
| 1073 | assert(isSpvIntrinsic(*I1, Intrinsic::spv_inline_asm) && I2->isInlineAsm()); |
| 1074 | MIRBuilder.setInsertPt(MBB&: *I2->getParent(), II: *I2); |
| 1075 | |
| 1076 | if (!AsmTargetReg.isValid()) { |
| 1077 | // define vendor specific assembly target or dialect |
| 1078 | AsmTargetReg = MRI.createGenericVirtualRegister(Ty: LLT::scalar(SizeInBits: 32)); |
| 1079 | MRI.setRegClass(Reg: AsmTargetReg, RC: &SPIRV::iIDRegClass); |
| 1080 | auto AsmTargetMIB = |
| 1081 | MIRBuilder.buildInstr(Opcode: SPIRV::OpAsmTargetINTEL).addDef(RegNo: AsmTargetReg); |
| 1082 | addStringImm(Str: ST.getTargetTripleAsStr(), MIB&: AsmTargetMIB); |
| 1083 | GR->add(Obj: AsmTargetMIB.getInstr(), MI: AsmTargetMIB); |
| 1084 | } |
| 1085 | |
| 1086 | // create types |
| 1087 | const MDNode *IAMD = I1->getOperand(i: 1).getMetadata(); |
| 1088 | FunctionType *FTy = cast<FunctionType>(Val: getMDOperandAsType(N: IAMD, I: 0)); |
| 1089 | SmallVector<SPIRVTypeInst, 4> ArgTypes; |
| 1090 | for (const auto &ArgTy : FTy->params()) |
| 1091 | ArgTypes.push_back(Elt: GR->getOrCreateSPIRVType( |
| 1092 | Type: ArgTy, MIRBuilder, AQ: SPIRV::AccessQualifier::ReadWrite, EmitIR: true)); |
| 1093 | SPIRVTypeInst RetType = |
| 1094 | GR->getOrCreateSPIRVType(Type: FTy->getReturnType(), MIRBuilder, |
| 1095 | AQ: SPIRV::AccessQualifier::ReadWrite, EmitIR: true); |
| 1096 | SPIRVTypeInst FuncType = GR->getOrCreateOpTypeFunctionWithArgs( |
| 1097 | Ty: FTy, RetType, ArgTypes, MIRBuilder); |
| 1098 | |
| 1099 | // define vendor specific assembly instructions string |
| 1100 | Register AsmReg = MRI.createGenericVirtualRegister(Ty: LLT::scalar(SizeInBits: 32)); |
| 1101 | MRI.setRegClass(Reg: AsmReg, RC: &SPIRV::iIDRegClass); |
| 1102 | auto AsmMIB = MIRBuilder.buildInstr(Opcode: SPIRV::OpAsmINTEL) |
| 1103 | .addDef(RegNo: AsmReg) |
| 1104 | .addUse(RegNo: GR->getSPIRVTypeID(SpirvType: RetType)) |
| 1105 | .addUse(RegNo: GR->getSPIRVTypeID(SpirvType: FuncType)) |
| 1106 | .addUse(RegNo: AsmTargetReg); |
| 1107 | // inline asm string: |
| 1108 | addStringImm(Str: I2->getOperand(i: InlineAsm::MIOp_AsmString).getSymbolName(), |
| 1109 | MIB&: AsmMIB); |
| 1110 | // inline asm constraint string: |
| 1111 | addStringImm(Str: cast<MDString>(Val: I1->getOperand(i: 2).getMetadata()->getOperand(I: 0)) |
| 1112 | ->getString(), |
| 1113 | MIB&: AsmMIB); |
| 1114 | GR->add(Obj: AsmMIB.getInstr(), MI: AsmMIB); |
| 1115 | |
| 1116 | // calls the inline assembly instruction |
| 1117 | unsigned = I2->getOperand(i: InlineAsm::MIOp_ExtraInfo).getImm(); |
| 1118 | if (ExtraInfo & InlineAsm::Extra_HasSideEffects) |
| 1119 | MIRBuilder.buildInstr(Opcode: SPIRV::OpDecorate) |
| 1120 | .addUse(RegNo: AsmReg) |
| 1121 | .addImm(Val: static_cast<uint32_t>(SPIRV::Decoration::SideEffectsINTEL)); |
| 1122 | |
| 1123 | Register DefReg = collectInlineAsmInstrOperands(MI: I2); |
| 1124 | if (!DefReg.isValid()) { |
| 1125 | DefReg = MRI.createGenericVirtualRegister(Ty: LLT::scalar(SizeInBits: 32)); |
| 1126 | MRI.setRegClass(Reg: DefReg, RC: &SPIRV::iIDRegClass); |
| 1127 | SPIRVTypeInst VoidType = GR->getOrCreateSPIRVType( |
| 1128 | Type: Type::getVoidTy(C&: MF.getFunction().getContext()), MIRBuilder, |
| 1129 | AQ: SPIRV::AccessQualifier::ReadWrite, EmitIR: true); |
| 1130 | GR->assignSPIRVTypeToVReg(Type: VoidType, VReg: DefReg, MF); |
| 1131 | } |
| 1132 | |
| 1133 | auto AsmCall = MIRBuilder.buildInstr(Opcode: SPIRV::OpAsmCallINTEL) |
| 1134 | .addDef(RegNo: DefReg) |
| 1135 | .addUse(RegNo: GR->getSPIRVTypeID(SpirvType: RetType)) |
| 1136 | .addUse(RegNo: AsmReg); |
| 1137 | for (unsigned IntrIdx = 3; IntrIdx < I1->getNumOperands(); ++IntrIdx) |
| 1138 | AsmCall.addUse(RegNo: I1->getOperand(i: IntrIdx).getReg()); |
| 1139 | |
| 1140 | // IRTranslator gets a bit confused when lowering inline ASM with outputs |
| 1141 | // and inserts a spurious COPY & TRUNC as registers are assumed to be i64; |
| 1142 | // we have to clean that up here to prevent erroneous trunc casts either on |
| 1143 | // a struct (for multiple outputs) or same width integers to get lowered |
| 1144 | // into SPIR-V |
| 1145 | if (MRI.hasOneUse(RegNo: DefReg)) { |
| 1146 | MachineInstr &CopyMI = *MRI.use_instr_begin(RegNo: DefReg); |
| 1147 | if (CopyMI.getOpcode() == TargetOpcode::COPY) { |
| 1148 | Register CopyDst = CopyMI.getOperand(i: 0).getReg(); |
| 1149 | if (MRI.hasOneUse(RegNo: CopyDst)) { |
| 1150 | MachineInstr &TruncMI = *MRI.use_instr_begin(RegNo: CopyDst); |
| 1151 | if (TruncMI.getOpcode() == TargetOpcode::G_TRUNC) { |
| 1152 | MRI.setType(VReg: DefReg, Ty: GR->getRegType(SpvType: RetType)); |
| 1153 | Register TruncReg = TruncMI.defs().begin()->getReg(); |
| 1154 | MRI.replaceRegWith(FromReg: TruncReg, ToReg: DefReg); |
| 1155 | invalidateAndEraseMI(GR, MI: &TruncMI); |
| 1156 | invalidateAndEraseMI(GR, MI: &CopyMI); |
| 1157 | } |
| 1158 | } |
| 1159 | } |
| 1160 | } |
| 1161 | } |
| 1162 | for (MachineInstr *MI : ToProcess) |
| 1163 | invalidateAndEraseMI(GR, MI); |
| 1164 | } |
| 1165 | |
| 1166 | static void insertInlineAsm(MachineFunction &MF, SPIRVGlobalRegistry *GR, |
| 1167 | const SPIRVSubtarget &ST, |
| 1168 | MachineIRBuilder MIRBuilder) { |
| 1169 | SmallVector<MachineInstr *> ToProcess; |
| 1170 | for (MachineBasicBlock &MBB : MF) { |
| 1171 | for (MachineInstr &MI : MBB) { |
| 1172 | if (isSpvIntrinsic(MI, IntrinsicID: Intrinsic::spv_inline_asm) || |
| 1173 | MI.getOpcode() == TargetOpcode::INLINEASM) |
| 1174 | ToProcess.push_back(Elt: &MI); |
| 1175 | } |
| 1176 | } |
| 1177 | if (ToProcess.size() == 0) |
| 1178 | return; |
| 1179 | |
| 1180 | if (!ST.canUseExtension(E: SPIRV::Extension::SPV_INTEL_inline_assembly)) |
| 1181 | report_fatal_error(reason: "Inline assembly instructions require the " |
| 1182 | "following SPIR-V extension: SPV_INTEL_inline_assembly" , |
| 1183 | gen_crash_diag: false); |
| 1184 | |
| 1185 | insertInlineAsmProcess(MF, GR, ST, MIRBuilder, ToProcess); |
| 1186 | } |
| 1187 | |
| 1188 | static void insertSpirvDecorations(MachineFunction &MF, SPIRVGlobalRegistry *GR, |
| 1189 | MachineIRBuilder MIB) { |
| 1190 | const SPIRVSubtarget &ST = cast<SPIRVSubtarget>(Val: MIB.getMF().getSubtarget()); |
| 1191 | SmallVector<MachineInstr *, 10> ToErase; |
| 1192 | for (MachineBasicBlock &MBB : MF) { |
| 1193 | for (MachineInstr &MI : MBB) { |
| 1194 | if (!isSpvIntrinsic(MI, IntrinsicID: Intrinsic::spv_assign_decoration) && |
| 1195 | !isSpvIntrinsic(MI, IntrinsicID: Intrinsic::spv_assign_aliasing_decoration) && |
| 1196 | !isSpvIntrinsic(MI, IntrinsicID: Intrinsic::spv_assign_fpmaxerror_decoration)) |
| 1197 | continue; |
| 1198 | MIB.setInsertPt(MBB&: *MI.getParent(), II: MI.getNextNode()); |
| 1199 | if (isSpvIntrinsic(MI, IntrinsicID: Intrinsic::spv_assign_decoration)) { |
| 1200 | buildOpSpirvDecorations(Reg: MI.getOperand(i: 1).getReg(), MIRBuilder&: MIB, |
| 1201 | GVarMD: MI.getOperand(i: 2).getMetadata(), ST); |
| 1202 | } else if (isSpvIntrinsic(MI, |
| 1203 | IntrinsicID: Intrinsic::spv_assign_fpmaxerror_decoration)) { |
| 1204 | ConstantFP *OpV = mdconst::dyn_extract<ConstantFP>( |
| 1205 | MD: MI.getOperand(i: 2).getMetadata()->getOperand(I: 0)); |
| 1206 | uint32_t OpValue = OpV->getValueAPF().bitcastToAPInt().getZExtValue(); |
| 1207 | |
| 1208 | buildOpDecorate(Reg: MI.getOperand(i: 1).getReg(), MIRBuilder&: MIB, |
| 1209 | Dec: SPIRV::Decoration::FPMaxErrorDecorationINTEL, |
| 1210 | DecArgs: {OpValue}); |
| 1211 | } else { |
| 1212 | GR->buildMemAliasingOpDecorate(Reg: MI.getOperand(i: 1).getReg(), MIRBuilder&: MIB, |
| 1213 | Dec: MI.getOperand(i: 2).getImm(), |
| 1214 | GVarMD: MI.getOperand(i: 3).getMetadata()); |
| 1215 | } |
| 1216 | |
| 1217 | ToErase.push_back(Elt: &MI); |
| 1218 | } |
| 1219 | } |
| 1220 | for (MachineInstr *MI : ToErase) |
| 1221 | invalidateAndEraseMI(GR, MI); |
| 1222 | } |
| 1223 | |
| 1224 | // Returns the value of the switch case operand in Reg. The case value stays a |
| 1225 | // G_CONSTANT until the module emits a SPIR-V constant for the same value, at |
| 1226 | // which point the case register is replaced with the one defining that |
| 1227 | // constant, which keeps its value in literal operands rather than in a CImm. |
| 1228 | static const ConstantInt *getSwitchCaseValue(Register Reg, |
| 1229 | const MachineRegisterInfo &MRI, |
| 1230 | LLVMContext &Ctx) { |
| 1231 | APInt Val; |
| 1232 | if (mi_match(R: Reg, MRI, P: m_ICst(Cst&: Val))) |
| 1233 | return ConstantInt::get(Context&: Ctx, V: Val); |
| 1234 | |
| 1235 | const MachineInstr *Def = nullptr; |
| 1236 | if (!mi_match(R: Reg, MRI, P: m_MInstr(MI&: Def))) |
| 1237 | llvm_unreachable("Switch case operand has no definition" ); |
| 1238 | |
| 1239 | LLT Ty = MRI.getType(Reg); |
| 1240 | assert(Ty.isValid() && "Expected a typed switch case value" ); |
| 1241 | Val = APInt(Ty.getScalarSizeInBits(), 0); |
| 1242 | |
| 1243 | switch (Def->getOpcode()) { |
| 1244 | case SPIRV::OpConstantNull: |
| 1245 | case SPIRV::OpConstantI: |
| 1246 | // The operands after the type are 32-bit literal words, least significant |
| 1247 | // first, as written by addNumImm(). OpConstantNull carries none, so it |
| 1248 | // decodes to zero without a case of its own. |
| 1249 | for (unsigned I = 2, E = Def->getNumExplicitOperands(); I != E; ++I) { |
| 1250 | uint32_t Word = static_cast<uint32_t>(Def->getOperand(i: I).getImm()); |
| 1251 | Val |= APInt(Val.getBitWidth(), Word).shl(shiftAmt: (I - 2) * 32); |
| 1252 | } |
| 1253 | break; |
| 1254 | default: |
| 1255 | llvm_unreachable("Unexpected definition of a switch case value" ); |
| 1256 | } |
| 1257 | return ConstantInt::get(Context&: Ctx, V: Val); |
| 1258 | } |
| 1259 | |
| 1260 | // LLVM allows the switches to use registers as cases, while SPIR-V required |
| 1261 | // those to be immediate values. This function replaces such operands with the |
| 1262 | // equivalent immediate constant. |
| 1263 | static void processSwitchesConstants(MachineFunction &MF, |
| 1264 | SPIRVGlobalRegistry *GR, |
| 1265 | MachineIRBuilder MIB) { |
| 1266 | MachineRegisterInfo &MRI = MF.getRegInfo(); |
| 1267 | LLVMContext &Ctx = MF.getFunction().getContext(); |
| 1268 | for (MachineBasicBlock &MBB : MF) { |
| 1269 | for (MachineInstr &MI : MBB) { |
| 1270 | if (!isSpvIntrinsic(MI, IntrinsicID: Intrinsic::spv_switch)) |
| 1271 | continue; |
| 1272 | |
| 1273 | SmallVector<MachineOperand, 8> NewOperands; |
| 1274 | NewOperands.push_back(Elt: MI.getOperand(i: 0)); // Opcode |
| 1275 | NewOperands.push_back(Elt: MI.getOperand(i: 1)); // Condition |
| 1276 | NewOperands.push_back(Elt: MI.getOperand(i: 2)); // Default |
| 1277 | for (unsigned i = 3; i < MI.getNumOperands(); i += 2) { |
| 1278 | Register Reg = MI.getOperand(i).getReg(); |
| 1279 | NewOperands.push_back( |
| 1280 | Elt: MachineOperand::CreateCImm(CI: getSwitchCaseValue(Reg, MRI, Ctx))); |
| 1281 | |
| 1282 | NewOperands.push_back(Elt: MI.getOperand(i: i + 1)); |
| 1283 | } |
| 1284 | |
| 1285 | assert(MI.getNumOperands() == NewOperands.size()); |
| 1286 | while (MI.getNumOperands() > 0) |
| 1287 | MI.removeOperand(OpNo: 0); |
| 1288 | for (auto &MO : NewOperands) |
| 1289 | MI.addOperand(Op: MO); |
| 1290 | } |
| 1291 | } |
| 1292 | } |
| 1293 | |
| 1294 | // Some instructions are used during CodeGen but should never be emitted. |
| 1295 | // Cleaning up those. |
| 1296 | static void cleanupHelperInstructions(MachineFunction &MF, |
| 1297 | SPIRVGlobalRegistry *GR) { |
| 1298 | SmallVector<MachineInstr *, 8> ToEraseMI; |
| 1299 | for (MachineBasicBlock &MBB : MF) { |
| 1300 | for (MachineInstr &MI : MBB) { |
| 1301 | if (isSpvIntrinsic(MI, IntrinsicID: Intrinsic::spv_track_constant) || |
| 1302 | MI.getOpcode() == TargetOpcode::G_BRINDIRECT) |
| 1303 | ToEraseMI.push_back(Elt: &MI); |
| 1304 | } |
| 1305 | } |
| 1306 | |
| 1307 | for (MachineInstr *MI : ToEraseMI) |
| 1308 | invalidateAndEraseMI(GR, MI); |
| 1309 | } |
| 1310 | |
| 1311 | // Find all usages of G_BLOCK_ADDR in our intrinsics and replace those |
| 1312 | // operands/registers by the actual MBB it references. |
| 1313 | static void processBlockAddr(MachineFunction &MF, SPIRVGlobalRegistry *GR, |
| 1314 | MachineIRBuilder MIB) { |
| 1315 | // Gather the reverse-mapping BB -> MBB. |
| 1316 | DenseMap<const BasicBlock *, MachineBasicBlock *> BB2MBB; |
| 1317 | for (MachineBasicBlock &MBB : MF) |
| 1318 | BB2MBB[MBB.getBasicBlock()] = &MBB; |
| 1319 | |
| 1320 | // Gather instructions requiring patching. For now, only those can use |
| 1321 | // G_BLOCK_ADDR. |
| 1322 | SmallVector<MachineInstr *, 8> InstructionsToPatch; |
| 1323 | for (MachineBasicBlock &MBB : MF) { |
| 1324 | for (MachineInstr &MI : MBB) { |
| 1325 | if (isSpvIntrinsic(MI, IntrinsicID: Intrinsic::spv_switch) || |
| 1326 | isSpvIntrinsic(MI, IntrinsicID: Intrinsic::spv_loop_merge) || |
| 1327 | isSpvIntrinsic(MI, IntrinsicID: Intrinsic::spv_selection_merge)) |
| 1328 | InstructionsToPatch.push_back(Elt: &MI); |
| 1329 | } |
| 1330 | } |
| 1331 | |
| 1332 | // For each instruction to fix, we replace all the G_BLOCK_ADDR operands by |
| 1333 | // the actual MBB it references. Once those references have been updated, we |
| 1334 | // can cleanup remaining G_BLOCK_ADDR references. |
| 1335 | SmallPtrSet<MachineBasicBlock *, 8> ClearAddressTaken; |
| 1336 | SmallPtrSet<MachineInstr *, 8> ToEraseMI; |
| 1337 | MachineRegisterInfo &MRI = MF.getRegInfo(); |
| 1338 | for (MachineInstr *MI : InstructionsToPatch) { |
| 1339 | SmallVector<MachineOperand, 8> NewOps; |
| 1340 | for (unsigned i = 0; i < MI->getNumOperands(); ++i) { |
| 1341 | // The operand is not a register, keep as-is. |
| 1342 | if (!MI->getOperand(i).isReg()) { |
| 1343 | NewOps.push_back(Elt: MI->getOperand(i)); |
| 1344 | continue; |
| 1345 | } |
| 1346 | |
| 1347 | Register Reg = MI->getOperand(i).getReg(); |
| 1348 | MachineInstr *BuildMBB = MRI.getVRegDef(Reg); |
| 1349 | // The register is not the result of G_BLOCK_ADDR, keep as-is. |
| 1350 | if (!BuildMBB || BuildMBB->getOpcode() != TargetOpcode::G_BLOCK_ADDR) { |
| 1351 | NewOps.push_back(Elt: MI->getOperand(i)); |
| 1352 | continue; |
| 1353 | } |
| 1354 | |
| 1355 | assert(BuildMBB && BuildMBB->getOpcode() == TargetOpcode::G_BLOCK_ADDR && |
| 1356 | BuildMBB->getOperand(1).isBlockAddress() && |
| 1357 | BuildMBB->getOperand(1).getBlockAddress()); |
| 1358 | BasicBlock *BB = |
| 1359 | BuildMBB->getOperand(i: 1).getBlockAddress()->getBasicBlock(); |
| 1360 | auto It = BB2MBB.find(Val: BB); |
| 1361 | if (It == BB2MBB.end()) |
| 1362 | report_fatal_error(reason: "cannot find a machine basic block by a basic block " |
| 1363 | "in a switch statement" ); |
| 1364 | MachineBasicBlock *ReferencedBlock = It->second; |
| 1365 | NewOps.push_back(Elt: MachineOperand::CreateMBB(MBB: ReferencedBlock)); |
| 1366 | |
| 1367 | ClearAddressTaken.insert(Ptr: ReferencedBlock); |
| 1368 | ToEraseMI.insert(Ptr: BuildMBB); |
| 1369 | } |
| 1370 | |
| 1371 | // Replace the operands. |
| 1372 | assert(MI->getNumOperands() == NewOps.size()); |
| 1373 | while (MI->getNumOperands() > 0) |
| 1374 | MI->removeOperand(OpNo: 0); |
| 1375 | for (auto &MO : NewOps) |
| 1376 | MI->addOperand(Op: MO); |
| 1377 | |
| 1378 | if (MachineInstr *Next = MI->getNextNode()) { |
| 1379 | if (isSpvIntrinsic(MI: *Next, IntrinsicID: Intrinsic::spv_track_constant)) { |
| 1380 | ToEraseMI.insert(Ptr: Next); |
| 1381 | Next = MI->getNextNode(); |
| 1382 | } |
| 1383 | if (Next && Next->getOpcode() == TargetOpcode::G_BRINDIRECT) |
| 1384 | ToEraseMI.insert(Ptr: Next); |
| 1385 | } |
| 1386 | } |
| 1387 | |
| 1388 | // BlockAddress operands were used to keep information between passes, |
| 1389 | // let's undo the "address taken" status to reflect that Succ doesn't |
| 1390 | // actually correspond to an IR-level basic block. |
| 1391 | for (MachineBasicBlock *Succ : ClearAddressTaken) |
| 1392 | Succ->setAddressTakenIRBlock(nullptr); |
| 1393 | |
| 1394 | // If we just delete G_BLOCK_ADDR instructions with BlockAddress operands, |
| 1395 | // this leaves their BasicBlock counterparts in a "address taken" status. This |
| 1396 | // would make AsmPrinter to generate a series of unneeded labels of a "Address |
| 1397 | // of block that was removed by CodeGen" kind. Let's first ensure that we |
| 1398 | // don't have a dangling BlockAddress constants by zapping the BlockAddress |
| 1399 | // nodes, and only after that proceed with erasing G_BLOCK_ADDR instructions. |
| 1400 | Constant *Replacement = |
| 1401 | ConstantInt::get(Ty: Type::getInt32Ty(C&: MF.getFunction().getContext()), V: 1); |
| 1402 | for (MachineInstr *BlockAddrI : ToEraseMI) { |
| 1403 | if (BlockAddrI->getOpcode() == TargetOpcode::G_BLOCK_ADDR) { |
| 1404 | BlockAddress *BA = const_cast<BlockAddress *>( |
| 1405 | BlockAddrI->getOperand(i: 1).getBlockAddress()); |
| 1406 | BA->replaceAllUsesWith( |
| 1407 | V: ConstantExpr::getIntToPtr(C: Replacement, Ty: BA->getType())); |
| 1408 | BA->destroyConstant(); |
| 1409 | } |
| 1410 | invalidateAndEraseMI(GR, MI: BlockAddrI); |
| 1411 | } |
| 1412 | } |
| 1413 | |
| 1414 | static bool isImplicitFallthrough(MachineBasicBlock &MBB) { |
| 1415 | if (MBB.empty()) |
| 1416 | return MBB.getNextNode() != nullptr; |
| 1417 | |
| 1418 | // Branching SPIR-V intrinsics are not detected by this generic method. |
| 1419 | // Thus, we can only trust negative result. |
| 1420 | if (!MBB.canFallThrough()) |
| 1421 | return false; |
| 1422 | |
| 1423 | // Otherwise, we must manually check if we have a SPIR-V intrinsic which |
| 1424 | // prevent an implicit fallthrough. |
| 1425 | for (MachineBasicBlock::reverse_iterator It = MBB.rbegin(), E = MBB.rend(); |
| 1426 | It != E; ++It) { |
| 1427 | if (isSpvIntrinsic(MI: *It, IntrinsicID: Intrinsic::spv_switch)) |
| 1428 | return false; |
| 1429 | } |
| 1430 | return true; |
| 1431 | } |
| 1432 | |
| 1433 | static void removeImplicitFallthroughs(MachineFunction &MF, |
| 1434 | MachineIRBuilder MIB) { |
| 1435 | // It is valid for MachineBasicBlocks to not finish with a branch instruction. |
| 1436 | // In such cases, they will simply fallthrough their immediate successor. |
| 1437 | for (MachineBasicBlock &MBB : MF) { |
| 1438 | if (!isImplicitFallthrough(MBB)) |
| 1439 | continue; |
| 1440 | |
| 1441 | assert(MBB.succ_size() == 1); |
| 1442 | MIB.setInsertPt(MBB, II: MBB.end()); |
| 1443 | MIB.buildBr(Dest&: **MBB.successors().begin()); |
| 1444 | } |
| 1445 | } |
| 1446 | |
| 1447 | static bool runPreLegalizer(MachineFunction &MF) { |
| 1448 | // Initialize the type registry. |
| 1449 | const SPIRVSubtarget &ST = MF.getSubtarget<SPIRVSubtarget>(); |
| 1450 | SPIRVGlobalRegistry *GR = ST.getSPIRVGlobalRegistry(); |
| 1451 | GR->setCurrentFunc(MF); |
| 1452 | MachineIRBuilder MIB(MF); |
| 1453 | // a registry of target extension constants |
| 1454 | DenseMap<MachineInstr *, Type *> TargetExtConstTypes; |
| 1455 | // to keep record of tracked constants |
| 1456 | addConstantsToTrack(MF, GR, STI: ST, TargetExtConstTypes); |
| 1457 | foldConstantsIntoIntrinsics(MF, GR, MIB); |
| 1458 | insertBitcasts(MF, GR, MIB); |
| 1459 | generateAssignInstrs(MF, GR, MIB, TargetExtConstTypes); |
| 1460 | |
| 1461 | processSwitchesConstants(MF, GR, MIB); |
| 1462 | processBlockAddr(MF, GR, MIB); |
| 1463 | cleanupHelperInstructions(MF, GR); |
| 1464 | |
| 1465 | processInstrsWithTypeFolding(MF, GR, MIB); |
| 1466 | removeImplicitFallthroughs(MF, MIB); |
| 1467 | insertSpirvDecorations(MF, GR, MIB); |
| 1468 | insertInlineAsm(MF, GR, ST, MIRBuilder: MIB); |
| 1469 | lowerBitcasts(MF, GR, MIB); |
| 1470 | |
| 1471 | return true; |
| 1472 | } |
| 1473 | |
| 1474 | INITIALIZE_PASS(SPIRVPreLegalizerLegacy, DEBUG_TYPE, "SPIRV pre legalizer" , |
| 1475 | false, false) |
| 1476 | |
| 1477 | char SPIRVPreLegalizerLegacy::ID = 0; |
| 1478 | |
| 1479 | FunctionPass *llvm::createSPIRVPreLegalizerLegacyPass() { |
| 1480 | return new SPIRVPreLegalizerLegacy(); |
| 1481 | } |
| 1482 | |
| 1483 | bool SPIRVPreLegalizerLegacy::runOnMachineFunction(MachineFunction &MF) { |
| 1484 | return runPreLegalizer(MF); |
| 1485 | } |
| 1486 | |
| 1487 | PreservedAnalyses |
| 1488 | SPIRVPreLegalizerPass::run(MachineFunction &MF, |
| 1489 | MachineFunctionAnalysisManager &MFAM) { |
| 1490 | bool Changed = runPreLegalizer(MF); |
| 1491 | if (!Changed) |
| 1492 | return PreservedAnalyses::all(); |
| 1493 | |
| 1494 | return getMachineFunctionPassPreservedAnalyses() |
| 1495 | .preserve<GISelValueTrackingAnalysis>(); |
| 1496 | } |
| 1497 | |