| 1 | //===-- X86TargetMachine.cpp - Define TargetMachine for the X86 -----------===// |
| 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 defines the X86 specific subclass of TargetMachine. |
| 10 | // |
| 11 | //===----------------------------------------------------------------------===// |
| 12 | |
| 13 | #include "X86TargetMachine.h" |
| 14 | #include "MCTargetDesc/X86MCTargetDesc.h" |
| 15 | #include "TargetInfo/X86TargetInfo.h" |
| 16 | #include "X86.h" |
| 17 | #include "X86MachineFunctionInfo.h" |
| 18 | #include "X86MacroFusion.h" |
| 19 | #include "X86Subtarget.h" |
| 20 | #include "X86TargetObjectFile.h" |
| 21 | #include "X86TargetTransformInfo.h" |
| 22 | #include "llvm-c/Visibility.h" |
| 23 | #include "llvm/ADT/SmallString.h" |
| 24 | #include "llvm/ADT/StringRef.h" |
| 25 | #include "llvm/Analysis/TargetTransformInfo.h" |
| 26 | #include "llvm/CodeGen/GlobalISel/CSEInfo.h" |
| 27 | #include "llvm/CodeGen/GlobalISel/CallLowering.h" |
| 28 | #include "llvm/CodeGen/GlobalISel/IRTranslator.h" |
| 29 | #include "llvm/CodeGen/GlobalISel/InstructionSelect.h" |
| 30 | #include "llvm/CodeGen/GlobalISel/Legalizer.h" |
| 31 | #include "llvm/CodeGen/GlobalISel/RegBankSelect.h" |
| 32 | #include "llvm/CodeGen/MIRParser/MIParser.h" |
| 33 | #include "llvm/CodeGen/MIRYamlMapping.h" |
| 34 | #include "llvm/CodeGen/MachineScheduler.h" |
| 35 | #include "llvm/CodeGen/Passes.h" |
| 36 | #include "llvm/CodeGen/TargetPassConfig.h" |
| 37 | #include "llvm/IR/Attributes.h" |
| 38 | #include "llvm/IR/DataLayout.h" |
| 39 | #include "llvm/IR/Function.h" |
| 40 | #include "llvm/MC/MCAsmInfo.h" |
| 41 | #include "llvm/MC/TargetRegistry.h" |
| 42 | #include "llvm/Option/LibraryOptions.h" |
| 43 | #include "llvm/Pass.h" |
| 44 | #include "llvm/Support/CodeGen.h" |
| 45 | #include "llvm/Support/ErrorHandling.h" |
| 46 | #include "llvm/Target/TargetLoweringObjectFile.h" |
| 47 | #include "llvm/Target/TargetOptions.h" |
| 48 | #include "llvm/TargetParser/Triple.h" |
| 49 | #include "llvm/Transforms/CFGuard.h" |
| 50 | #include <memory> |
| 51 | #include <optional> |
| 52 | |
| 53 | using namespace llvm; |
| 54 | |
| 55 | extern "C" LLVM_C_ABI void LLVMInitializeX86Target() { |
| 56 | // Register the target. |
| 57 | RegisterTargetMachine<X86TargetMachine> X(getTheX86_32Target()); |
| 58 | RegisterTargetMachine<X86TargetMachine> Y(getTheX86_64Target()); |
| 59 | static opt::RegisterLibraryOptions<X86Options> O; |
| 60 | |
| 61 | PassRegistry &PR = *PassRegistry::getPassRegistry(); |
| 62 | initializeX86LowerAMXIntrinsicsLegacyPassPass(PR); |
| 63 | initializeX86LowerAMXTypeLegacyPassPass(PR); |
| 64 | initializeX86PreTileConfigLegacyPass(PR); |
| 65 | initializeGlobalISel(PR); |
| 66 | initializeWinEHStateLegacyPass(PR); |
| 67 | initializeX86FixupBWInstLegacyPass(PR); |
| 68 | initializeCompressEVEXLegacyPass(PR); |
| 69 | initializeFixupLEAsLegacyPass(PR); |
| 70 | initializeX86FPStackifierLegacyPass(PR); |
| 71 | initializeX86FenceNonTemporalStoresLegacyPass(PR); |
| 72 | initializeX86FixupSetCCLegacyPass(PR); |
| 73 | initializeX86CallFrameOptimizationLegacyPass(PR); |
| 74 | initializeX86CmovConversionLegacyPass(PR); |
| 75 | initializeX86TileConfigLegacyPass(PR); |
| 76 | initializeX86FastPreTileConfigLegacyPass(PR); |
| 77 | initializeX86FastTileConfigLegacyPass(PR); |
| 78 | initializeMachineKCFILegacyPass(PR); |
| 79 | initializeX86LowerTileCopyLegacyPass(PR); |
| 80 | initializeX86ExpandPseudoLegacyPass(PR); |
| 81 | initializeX86DomainReassignmentLegacyPass(PR); |
| 82 | initializeX86AvoidSFBLegacyPass(PR); |
| 83 | initializeX86AvoidTrailingCallLegacyPassPass(PR); |
| 84 | initializeX86SpeculativeLoadHardeningLegacyPass(PR); |
| 85 | initializeX86SpeculativeExecutionSideEffectSuppressionLegacyPass(PR); |
| 86 | initializeX86FlagsCopyLoweringLegacyPass(PR); |
| 87 | initializeX86LoadValueInjectionLoadHardeningLegacyPass(PR); |
| 88 | initializeX86LoadValueInjectionRetHardeningLegacyPass(PR); |
| 89 | initializeX86OptimizeLEAsLegacyPass(PR); |
| 90 | initializeX86PartialReductionLegacyPass(PR); |
| 91 | initializeX86ReturnThunksLegacyPass(PR); |
| 92 | initializeX86DAGToDAGISelLegacyPass(PR); |
| 93 | initializeX86ArgumentStackSlotLegacyPass(PR); |
| 94 | initializeX86AsmPrinterPass(PR); |
| 95 | initializeX86FixupInstTuningLegacyPass(PR); |
| 96 | initializeX86FixupVectorConstantsLegacyPass(PR); |
| 97 | initializeX86DynAllocaExpanderLegacyPass(PR); |
| 98 | initializeX86SuppressAPXForRelocationLegacyPass(PR); |
| 99 | initializeX86WinEHUnwindV2LegacyPass(PR); |
| 100 | initializeX86PreLegalizerCombinerLegacyPass(PR); |
| 101 | initializeX86PostLegalizerCombinerLegacyPass(PR); |
| 102 | initializeX86WinEHUnwindV3Pass(PR); |
| 103 | } |
| 104 | |
| 105 | static std::unique_ptr<TargetLoweringObjectFile> createTLOF(const Triple &TT) { |
| 106 | if (TT.isOSBinFormatMachO()) { |
| 107 | if (TT.isX86_64()) |
| 108 | return std::make_unique<X86_64MachoTargetObjectFile>(); |
| 109 | return std::make_unique<TargetLoweringObjectFileMachO>(); |
| 110 | } |
| 111 | |
| 112 | if (TT.isOSBinFormatCOFF()) |
| 113 | return std::make_unique<TargetLoweringObjectFileCOFF>(); |
| 114 | |
| 115 | if (TT.isX86_64()) |
| 116 | return std::make_unique<X86_64ELFTargetObjectFile>(); |
| 117 | return std::make_unique<X86ELFTargetObjectFile>(); |
| 118 | } |
| 119 | |
| 120 | static Reloc::Model getEffectiveRelocModel(const Triple &TT, bool JIT, |
| 121 | std::optional<Reloc::Model> RM) { |
| 122 | bool is64Bit = TT.isX86_64(); |
| 123 | if (!RM) { |
| 124 | // JIT codegen should use static relocations by default, since it's |
| 125 | // typically executed in process and not relocatable. |
| 126 | if (JIT) |
| 127 | return Reloc::Static; |
| 128 | |
| 129 | // Darwin defaults to PIC in 64 bit mode and dynamic-no-pic in 32 bit mode. |
| 130 | // Win64 requires rip-rel addressing, thus we force it to PIC. Otherwise we |
| 131 | // use static relocation model by default. |
| 132 | if (TT.isOSDarwin()) { |
| 133 | if (is64Bit) |
| 134 | return Reloc::PIC_; |
| 135 | return Reloc::DynamicNoPIC; |
| 136 | } |
| 137 | if (TT.isOSWindows() && is64Bit) |
| 138 | return Reloc::PIC_; |
| 139 | return Reloc::Static; |
| 140 | } |
| 141 | |
| 142 | // ELF and X86-64 don't have a distinct DynamicNoPIC model. DynamicNoPIC |
| 143 | // is defined as a model for code which may be used in static or dynamic |
| 144 | // executables but not necessarily a shared library. On X86-32 we just |
| 145 | // compile in -static mode, in x86-64 we use PIC. |
| 146 | if (*RM == Reloc::DynamicNoPIC) { |
| 147 | if (is64Bit) |
| 148 | return Reloc::PIC_; |
| 149 | if (!TT.isOSDarwin()) |
| 150 | return Reloc::Static; |
| 151 | } |
| 152 | |
| 153 | // If we are on Darwin, disallow static relocation model in X86-64 mode, since |
| 154 | // the Mach-O file format doesn't support it. |
| 155 | if (*RM == Reloc::Static && TT.isOSDarwin() && is64Bit) |
| 156 | return Reloc::PIC_; |
| 157 | |
| 158 | return *RM; |
| 159 | } |
| 160 | |
| 161 | static CodeModel::Model |
| 162 | getEffectiveX86CodeModel(const Triple &TT, std::optional<CodeModel::Model> CM, |
| 163 | bool JIT) { |
| 164 | bool Is64Bit = TT.isX86_64(); |
| 165 | if (CM) { |
| 166 | if (*CM == CodeModel::Tiny) |
| 167 | reportFatalUsageError(reason: "target does not support the tiny CodeModel" ); |
| 168 | return *CM; |
| 169 | } |
| 170 | if (JIT) |
| 171 | return Is64Bit ? CodeModel::Large : CodeModel::Small; |
| 172 | return CodeModel::Small; |
| 173 | } |
| 174 | |
| 175 | /// Create an X86 target. |
| 176 | /// |
| 177 | X86TargetMachine::X86TargetMachine(const Target &T, const Triple &TT, |
| 178 | StringRef CPU, StringRef FS, |
| 179 | const TargetOptions &Options, |
| 180 | std::optional<Reloc::Model> RM, |
| 181 | std::optional<CodeModel::Model> CM, |
| 182 | CodeGenOptLevel OL, bool JIT) |
| 183 | : CodeGenTargetMachineImpl(T, TT, CPU, FS, Options, |
| 184 | getEffectiveRelocModel(TT, JIT, RM), |
| 185 | getEffectiveX86CodeModel(TT, CM, JIT), OL), |
| 186 | CLOpts(cl::LibraryOptionsContext::get<X86Options>( |
| 187 | C: Options.MCOptions.LibraryOptions)), |
| 188 | TLOF(createTLOF(TT: getTargetTriple())), IsJIT(JIT) { |
| 189 | // On PS4/PS5, the "return address" of a 'noreturn' call must still be within |
| 190 | // the calling function. Note that this also includes __stack_chk_fail, |
| 191 | // so there was some target-specific logic in the instruction selectors |
| 192 | // to handle that. That code has since been generalized, so the only thing |
| 193 | // needed is to set TrapUnreachable here. |
| 194 | if (TT.isPS() || TT.isOSBinFormatMachO()) { |
| 195 | this->Options.TrapUnreachable = true; |
| 196 | this->Options.NoTrapAfterNoreturn = TT.isOSBinFormatMachO(); |
| 197 | } |
| 198 | |
| 199 | setMachineOutliner(true); |
| 200 | |
| 201 | // x86 supports the debug entry values. |
| 202 | setSupportsDebugEntryValues(true); |
| 203 | |
| 204 | initAsmInfo(); |
| 205 | } |
| 206 | |
| 207 | X86TargetMachine::~X86TargetMachine() = default; |
| 208 | |
| 209 | const X86Subtarget * |
| 210 | X86TargetMachine::getSubtargetImpl(const Function &F) const { |
| 211 | AttributeSet FnAttrs = F.getAttributes().getFnAttrs(); |
| 212 | const Module *M = F.getParent(); |
| 213 | if (const X86Subtarget *ST = LastSubtarget.lookup(NewFnAttrs: FnAttrs, NewM: M)) |
| 214 | return ST; |
| 215 | |
| 216 | SubtargetAttrs Attrs; |
| 217 | Attrs.CPU = FnAttrs.getAttribute(Kind: "target-cpu" ); |
| 218 | Attrs.TuneCPU = FnAttrs.getAttribute(Kind: "tune-cpu" ); |
| 219 | Attrs.FS = FnAttrs.getAttribute(Kind: "target-features" ); |
| 220 | Attrs.PreferVecWidth = FnAttrs.getAttribute(Kind: "prefer-vector-width" ); |
| 221 | Attrs.MinLegalVecWidth = FnAttrs.getAttribute(Kind: "min-legal-vector-width" ); |
| 222 | Attrs.SoftFloat = FnAttrs.getAttribute(Kind: "use-soft-float" ); |
| 223 | |
| 224 | if (const X86Subtarget *ST = LastSubtarget.lookup(NewFnAttrs: FnAttrs, NewM: M, NewAttrs: Attrs)) |
| 225 | return ST; |
| 226 | |
| 227 | StringRef CPU = |
| 228 | Attrs.CPU.isValid() ? Attrs.CPU.getValueAsString() : (StringRef)TargetCPU; |
| 229 | // "x86-64" is a default target setting for many front ends. In these cases, |
| 230 | // they actually request for "generic" tuning unless the "tune-cpu" was |
| 231 | // specified. |
| 232 | StringRef TuneCPU = Attrs.TuneCPU.isValid() ? Attrs.TuneCPU.getValueAsString() |
| 233 | : CPU == "x86-64" ? "generic" |
| 234 | : (StringRef)CPU; |
| 235 | StringRef FS = |
| 236 | Attrs.FS.isValid() ? Attrs.FS.getValueAsString() : (StringRef)TargetFS; |
| 237 | |
| 238 | SmallString<512> Key; |
| 239 | // The additions here are ordered so that the definitely short strings are |
| 240 | // added first so we won't exceed the small size. We append the |
| 241 | // much longer FS string at the end so that we only heap allocate at most |
| 242 | // one time. |
| 243 | |
| 244 | // Extract prefer-vector-width attribute. |
| 245 | unsigned PreferVectorWidthOverride = 0; |
| 246 | if (Attrs.PreferVecWidth.isValid()) { |
| 247 | StringRef Val = Attrs.PreferVecWidth.getValueAsString(); |
| 248 | unsigned Width; |
| 249 | if (!Val.getAsInteger(Radix: 0, Result&: Width)) { |
| 250 | Key += 'p'; |
| 251 | Key += Val; |
| 252 | PreferVectorWidthOverride = Width; |
| 253 | } |
| 254 | } |
| 255 | |
| 256 | // Extract min-legal-vector-width attribute. |
| 257 | unsigned RequiredVectorWidth = UINT32_MAX; |
| 258 | if (Attrs.MinLegalVecWidth.isValid()) { |
| 259 | StringRef Val = Attrs.MinLegalVecWidth.getValueAsString(); |
| 260 | unsigned Width; |
| 261 | if (!Val.getAsInteger(Radix: 0, Result&: Width)) { |
| 262 | Key += 'm'; |
| 263 | Key += Val; |
| 264 | RequiredVectorWidth = Width; |
| 265 | } |
| 266 | } |
| 267 | |
| 268 | // Add CPU to the Key. |
| 269 | Key += CPU; |
| 270 | |
| 271 | // Add tune CPU to the Key. |
| 272 | Key += TuneCPU; |
| 273 | |
| 274 | // Keep track of the start of the feature portion of the string. |
| 275 | unsigned FSStart = Key.size(); |
| 276 | |
| 277 | // FIXME: This is related to the code below to reset the target options, |
| 278 | // we need to know whether or not the soft float flag is set on the |
| 279 | // function before we can generate a subtarget. We also need to use |
| 280 | // it as a key for the subtarget since that can be the only difference |
| 281 | // between two functions. |
| 282 | bool SoftFloat = Attrs.SoftFloat.getValueAsBool(); |
| 283 | // If the soft float attribute is set on the function turn on the soft float |
| 284 | // subtarget feature. |
| 285 | if (SoftFloat) |
| 286 | Key += FS.empty() ? "+soft-float" : "+soft-float," ; |
| 287 | |
| 288 | Key += FS; |
| 289 | |
| 290 | // We may have added +soft-float to the features so move the StringRef to |
| 291 | // point to the full string in the Key. |
| 292 | FS = Key.substr(Start: FSStart); |
| 293 | |
| 294 | auto &I = SubtargetMap[Key]; |
| 295 | if (!I) { |
| 296 | I = std::make_unique<X86Subtarget>( |
| 297 | args: TargetTriple, args&: CPU, args&: TuneCPU, args&: FS, args: *this, |
| 298 | args: MaybeAlign(F.getParent()->getOverrideStackAlignment()), |
| 299 | args&: PreferVectorWidthOverride, args&: RequiredVectorWidth); |
| 300 | } |
| 301 | LastSubtarget.set(NewFnAttrs: FnAttrs, NewM: M, NewAttrs: Attrs, NewSubtarget: I.get()); |
| 302 | return I.get(); |
| 303 | } |
| 304 | |
| 305 | yaml::MachineFunctionInfo *X86TargetMachine::createDefaultFuncInfoYAML() const { |
| 306 | return new yaml::X86MachineFunctionInfo(); |
| 307 | } |
| 308 | |
| 309 | yaml::MachineFunctionInfo * |
| 310 | X86TargetMachine::convertFuncInfoToYAML(const MachineFunction &MF) const { |
| 311 | const auto *MFI = MF.getInfo<X86MachineFunctionInfo>(); |
| 312 | return new yaml::X86MachineFunctionInfo(*MFI); |
| 313 | } |
| 314 | |
| 315 | bool X86TargetMachine::parseMachineFunctionInfo( |
| 316 | const yaml::MachineFunctionInfo &MFI, PerFunctionMIParsingState &PFS, |
| 317 | SMDiagnostic &Error, SMRange &SourceRange) const { |
| 318 | const auto &YamlMFI = static_cast<const yaml::X86MachineFunctionInfo &>(MFI); |
| 319 | PFS.MF.getInfo<X86MachineFunctionInfo>()->initializeBaseYamlFields(YamlMFI); |
| 320 | return false; |
| 321 | } |
| 322 | |
| 323 | bool X86TargetMachine::isNoopAddrSpaceCast(const DataLayout &DL, unsigned SrcAS, |
| 324 | unsigned DestAS) const { |
| 325 | assert(SrcAS != DestAS && "Expected different address spaces!" ); |
| 326 | if (DL.getPointerSize(AS: SrcAS) != DL.getPointerSize(AS: DestAS)) |
| 327 | return false; |
| 328 | return SrcAS < 256 && DestAS < 256; |
| 329 | } |
| 330 | |
| 331 | void X86TargetMachine::reset() { |
| 332 | SubtargetMap.clear(); |
| 333 | LastSubtarget.clear(); |
| 334 | } |
| 335 | |
| 336 | ScheduleDAGInstrs * |
| 337 | X86TargetMachine::createMachineScheduler(MachineSchedContext *C) const { |
| 338 | ScheduleDAGMILive *DAG = createSchedLive(C); |
| 339 | DAG->addMutation(Mutation: createX86MacroFusionDAGMutation()); |
| 340 | return DAG; |
| 341 | } |
| 342 | |
| 343 | ScheduleDAGInstrs * |
| 344 | X86TargetMachine::createPostMachineScheduler(MachineSchedContext *C) const { |
| 345 | ScheduleDAGMI *DAG = createSchedPostRA(C); |
| 346 | DAG->addMutation(Mutation: createX86MacroFusionDAGMutation()); |
| 347 | return DAG; |
| 348 | } |
| 349 | |
| 350 | //===----------------------------------------------------------------------===// |
| 351 | // X86 TTI query. |
| 352 | //===----------------------------------------------------------------------===// |
| 353 | |
| 354 | TargetTransformInfo |
| 355 | X86TargetMachine::getTargetTransformInfo(const Function &F) const { |
| 356 | return TargetTransformInfo(std::make_unique<X86TTIImpl>(args: this, args: F)); |
| 357 | } |
| 358 | |
| 359 | //===----------------------------------------------------------------------===// |
| 360 | // Pass Pipeline Configuration |
| 361 | //===----------------------------------------------------------------------===// |
| 362 | |
| 363 | namespace { |
| 364 | |
| 365 | /// X86 Code Generator Pass Configuration Options. |
| 366 | class X86PassConfig : public TargetPassConfig { |
| 367 | public: |
| 368 | X86PassConfig(X86TargetMachine &TM, PassManagerBase &PM) |
| 369 | : TargetPassConfig(TM, PM) {} |
| 370 | |
| 371 | X86TargetMachine &getX86TargetMachine() const { |
| 372 | return getTM<X86TargetMachine>(); |
| 373 | } |
| 374 | |
| 375 | void addIRPasses() override; |
| 376 | bool addInstSelector() override; |
| 377 | bool addIRTranslator() override; |
| 378 | bool addLegalizeMachineIR() override; |
| 379 | void addPreRegBankSelect() override; |
| 380 | bool addRegBankSelect() override; |
| 381 | bool addGlobalInstructionSelect() override; |
| 382 | void addPreLegalizeMachineIR() override; |
| 383 | bool addILPOpts() override; |
| 384 | bool addPreISel() override; |
| 385 | void addMachineSSAOptimization() override; |
| 386 | void addPreRegAlloc() override; |
| 387 | bool addRegAssignAndRewriteFast() override; |
| 388 | void addPostRegAlloc() override; |
| 389 | void addPreEmitPass() override; |
| 390 | void addPreEmitPass2() override; |
| 391 | void addPreSched2() override; |
| 392 | bool addRegAssignAndRewriteOptimized() override; |
| 393 | |
| 394 | std::unique_ptr<CSEConfigBase> getCSEConfig() const override; |
| 395 | }; |
| 396 | |
| 397 | } // end anonymous namespace |
| 398 | |
| 399 | TargetPassConfig *X86TargetMachine::createPassConfig(PassManagerBase &PM) { |
| 400 | return new X86PassConfig(*this, PM); |
| 401 | } |
| 402 | |
| 403 | MachineFunctionInfo *X86TargetMachine::createMachineFunctionInfo( |
| 404 | BumpPtrAllocator &Allocator, const Function &F, |
| 405 | const TargetSubtargetInfo *STI) const { |
| 406 | return X86MachineFunctionInfo::create<X86MachineFunctionInfo>(Allocator, F, |
| 407 | STI); |
| 408 | } |
| 409 | |
| 410 | void X86PassConfig::addIRPasses() { |
| 411 | addPass(P: createAtomicExpandLegacyPass()); |
| 412 | |
| 413 | // We add both pass anyway and when these two passes run, we skip the pass |
| 414 | // based on the option level and option attribute. |
| 415 | addPass(P: createX86LowerAMXIntrinsicsLegacyPass()); |
| 416 | addPass(P: createX86LowerAMXTypeLegacyPass()); |
| 417 | |
| 418 | TargetPassConfig::addIRPasses(); |
| 419 | |
| 420 | if (TM->getOptLevel() != CodeGenOptLevel::None) { |
| 421 | addPass(P: createInterleavedAccessPass()); |
| 422 | addPass(P: createX86PartialReductionLegacyPass()); |
| 423 | } |
| 424 | |
| 425 | // Add passes that handle indirect branch removal and insertion of a retpoline |
| 426 | // thunk. These will be a no-op unless a function subtarget has the retpoline |
| 427 | // feature enabled. |
| 428 | addPass(P: createIndirectBrExpandPass(OptLevel: getOptLevel())); |
| 429 | |
| 430 | // Add Control Flow Guard checks. |
| 431 | const Triple &TT = TM->getTargetTriple(); |
| 432 | if (TT.isOSWindows()) { |
| 433 | addPass(P: createCFGuardPass()); |
| 434 | } |
| 435 | |
| 436 | if (TM->Options.JMCInstrument) |
| 437 | addPass(P: createJMCInstrumenterPass()); |
| 438 | |
| 439 | addPass(P: createX86FenceNonTemporalStoresLegacyPass()); |
| 440 | } |
| 441 | |
| 442 | bool X86PassConfig::addInstSelector() { |
| 443 | // Install an instruction selector. |
| 444 | addPass(P: createX86ISelDag(TM&: getX86TargetMachine(), OptLevel: getOptLevel())); |
| 445 | |
| 446 | // For ELF, cleanup any local-dynamic TLS accesses. |
| 447 | if (TM->getTargetTriple().isOSBinFormatELF() && |
| 448 | getOptLevel() != CodeGenOptLevel::None) |
| 449 | addPass(P: createCleanupLocalDynamicTLSLegacyPass()); |
| 450 | |
| 451 | addPass(P: createX86GlobalBaseRegLegacyPass()); |
| 452 | addPass(P: createX86ArgumentStackSlotLegacyPass()); |
| 453 | return false; |
| 454 | } |
| 455 | |
| 456 | bool X86PassConfig::addIRTranslator() { |
| 457 | addPass(P: new IRTranslatorLegacy(getOptLevel())); |
| 458 | return false; |
| 459 | } |
| 460 | |
| 461 | void X86PassConfig::addPreRegBankSelect() { |
| 462 | bool IsOptNone = getOptLevel() == CodeGenOptLevel::None; |
| 463 | if (!IsOptNone) { |
| 464 | addPass(P: createX86PostLegalizerCombinerLegacy()); |
| 465 | } |
| 466 | } |
| 467 | bool X86PassConfig::addLegalizeMachineIR() { |
| 468 | addPass(P: new LegalizerLegacy()); |
| 469 | return false; |
| 470 | } |
| 471 | |
| 472 | bool X86PassConfig::addRegBankSelect() { |
| 473 | addPass(P: new RegBankSelectLegacy()); |
| 474 | return false; |
| 475 | } |
| 476 | |
| 477 | bool X86PassConfig::addGlobalInstructionSelect() { |
| 478 | addPass(P: new InstructionSelectLegacy(getOptLevel())); |
| 479 | // Add GlobalBaseReg in case there is no SelectionDAG passes afterwards |
| 480 | if (isGlobalISelAbortEnabled()) |
| 481 | addPass(P: createX86GlobalBaseRegLegacyPass()); |
| 482 | return false; |
| 483 | } |
| 484 | |
| 485 | void X86PassConfig::addPreLegalizeMachineIR() { |
| 486 | if (getOptLevel() != CodeGenOptLevel::None) { |
| 487 | addPass(P: createX86PreLegalizerCombinerLegacy()); |
| 488 | } |
| 489 | } |
| 490 | |
| 491 | bool X86PassConfig::addILPOpts() { |
| 492 | addPass(PassID: &EarlyIfConverterLegacyID); |
| 493 | if (getX86TargetMachine().getCLOpts().machine_combiner) |
| 494 | addPass(PassID: &MachineCombinerID); |
| 495 | addPass(P: createX86CmovConversionLegacyPass()); |
| 496 | return true; |
| 497 | } |
| 498 | |
| 499 | bool X86PassConfig::addPreISel() { |
| 500 | // Only add this pass for 32-bit x86 Windows. |
| 501 | const Triple &TT = TM->getTargetTriple(); |
| 502 | if (TT.isOSWindows() && TT.isX86_32()) |
| 503 | addPass(P: createX86WinEHStateLegacyPass()); |
| 504 | return true; |
| 505 | } |
| 506 | |
| 507 | void X86PassConfig::addPreRegAlloc() { |
| 508 | if (getOptLevel() != CodeGenOptLevel::None) { |
| 509 | addPass(PassID: &LiveRangeShrinkID); |
| 510 | addPass(P: createX86FixupSetCCLegacyPass()); |
| 511 | addPass(P: createX86OptimizeLEAsLegacyPass()); |
| 512 | addPass(P: createX86CallFrameOptimizationLegacyPass()); |
| 513 | addPass(P: createX86AvoidStoreForwardingBlocksLegacyPass()); |
| 514 | } |
| 515 | |
| 516 | addPass(P: createX86SuppressAPXForRelocationLegacyPass()); |
| 517 | |
| 518 | addPass(P: createX86SpeculativeLoadHardeningLegacyPass()); |
| 519 | addPass(P: createX86FlagsCopyLoweringLegacyPass()); |
| 520 | addPass(P: createX86DynAllocaExpanderLegacyPass()); |
| 521 | |
| 522 | if (getOptLevel() != CodeGenOptLevel::None) |
| 523 | addPass(P: createX86PreTileConfigLegacyPass()); |
| 524 | else |
| 525 | addPass(P: createX86FastPreTileConfigLegacyPass()); |
| 526 | } |
| 527 | |
| 528 | bool X86PassConfig::addRegAssignAndRewriteFast() { |
| 529 | bool ToReturn = TargetPassConfig::addRegAssignAndRewriteFast(); |
| 530 | addPass(P: createX86FastTileConfigLegacyPass()); |
| 531 | return ToReturn; |
| 532 | } |
| 533 | |
| 534 | void X86PassConfig::addMachineSSAOptimization() { |
| 535 | addPass(P: createX86DomainReassignmentLegacyPass()); |
| 536 | TargetPassConfig::addMachineSSAOptimization(); |
| 537 | } |
| 538 | |
| 539 | void X86PassConfig::addPostRegAlloc() { |
| 540 | addPass(P: createX86LowerTileCopyLegacyPass()); |
| 541 | addPass(P: createX86FPStackifierLegacyPass()); |
| 542 | // When -O0 is enabled, the Load Value Injection Hardening pass will fall back |
| 543 | // to using the Speculative Execution Side Effect Suppression pass for |
| 544 | // mitigation. This is to prevent slow downs due to |
| 545 | // analyses needed by the LVIHardening pass when compiling at -O0. |
| 546 | if (getOptLevel() != CodeGenOptLevel::None) |
| 547 | addPass(P: createX86LoadValueInjectionLoadHardeningLegacyPass()); |
| 548 | } |
| 549 | |
| 550 | void X86PassConfig::addPreSched2() { |
| 551 | addPass(P: createX86ExpandPseudoLegacyPass()); |
| 552 | addPass(P: createKCFIPass()); |
| 553 | } |
| 554 | |
| 555 | void X86PassConfig::addPreEmitPass() { |
| 556 | if (getOptLevel() != CodeGenOptLevel::None) { |
| 557 | addPass(P: createExecutionDomainFixLegacyPass()); |
| 558 | addPass(P: createBreakFalseDepsLegacyPass()); |
| 559 | } |
| 560 | |
| 561 | addPass(P: createX86IndirectBranchTrackingLegacyPass()); |
| 562 | |
| 563 | addPass(P: createX86InsertVZeroUpperLegacyPass()); |
| 564 | |
| 565 | if (getOptLevel() != CodeGenOptLevel::None) { |
| 566 | addPass(P: createX86FixupBWInstsLegacyPass()); |
| 567 | addPass(P: createX86PadShortFunctionsLegacyPass()); |
| 568 | addPass(P: createX86FixupLEAsLegacyPass()); |
| 569 | addPass(P: createX86FixupInstTuningLegacyPass()); |
| 570 | addPass(P: createX86FixupVectorConstantsLegacyPass()); |
| 571 | } |
| 572 | addPass(P: createX86CompressEVEXLegacyPass()); |
| 573 | addPass(P: createX86InsertX87WaitLegacyPass()); |
| 574 | |
| 575 | if (TM->getTargetTriple().isLFI()) |
| 576 | addPass(P: createX86LFIRewritePass()); |
| 577 | } |
| 578 | |
| 579 | void X86PassConfig::addPreEmitPass2() { |
| 580 | const Triple &TT = TM->getTargetTriple(); |
| 581 | |
| 582 | // The X86 Speculative Execution Pass must run after all control |
| 583 | // flow graph modifying passes. As a result it was listed to run right before |
| 584 | // the X86 Retpoline Thunks pass. The reason it must run after control flow |
| 585 | // graph modifications is that the model of LFENCE in LLVM has to be updated |
| 586 | // (FIXME: https://bugs.llvm.org/show_bug.cgi?id=45167). Currently the |
| 587 | // placement of this pass was hand checked to ensure that the subsequent |
| 588 | // passes don't move the code around the LFENCEs in a way that will hurt the |
| 589 | // correctness of this pass. This placement has been shown to work based on |
| 590 | // hand inspection of the codegen output. |
| 591 | addPass(P: createX86SpeculativeExecutionSideEffectSuppressionLegacyPass()); |
| 592 | addPass(P: createX86IndirectThunksLegacyPass()); |
| 593 | addPass(P: createX86ReturnThunksLegacyPass()); |
| 594 | |
| 595 | // Insert extra int3 instructions after trailing call instructions to avoid |
| 596 | // issues in the unwinder. |
| 597 | if (TT.isOSWindows() && TT.isX86_64()) |
| 598 | addPass(P: createX86AvoidTrailingCallLegacyPass()); |
| 599 | |
| 600 | // Verify basic block incoming and outgoing cfa offset and register values and |
| 601 | // correct CFA calculation rule where needed by inserting appropriate CFI |
| 602 | // instructions. CFIInstrInserter self-skips targets that do not use DWARF |
| 603 | // CFI. |
| 604 | if (!TT.isOSDarwin()) |
| 605 | addPass(P: createCFIInstrInserterLegacy()); |
| 606 | |
| 607 | if (TT.isOSWindows()) { |
| 608 | // Identify valid longjmp targets for Windows Control Flow Guard. |
| 609 | addPass(P: createCFGuardLongjmpLegacyPass()); |
| 610 | // Identify valid eh continuation targets for Windows EHCont Guard. |
| 611 | addPass(P: createEHContGuardTargetsLegacy()); |
| 612 | } |
| 613 | addPass(P: createX86LoadValueInjectionRetHardeningLegacyPass()); |
| 614 | |
| 615 | // Insert pseudo probe annotation for callsite profiling |
| 616 | addPass(P: createPseudoProbeInserterLegacyPass()); |
| 617 | |
| 618 | // KCFI indirect call checks are lowered to a bundle, and on Darwin platforms, |
| 619 | // also CALL_RVMARKER. |
| 620 | addPass(P: createUnpackMachineBundlesLegacy(Ftor: [&TT](const MachineFunction &MF) { |
| 621 | // Only run bundle expansion if the module uses kcfi, or there are relevant |
| 622 | // ObjC runtime functions present in the module. |
| 623 | const Function &F = MF.getFunction(); |
| 624 | const Module *M = F.getParent(); |
| 625 | return M->getModuleFlag(Key: "kcfi" ) || |
| 626 | (TT.isOSDarwin() && |
| 627 | (M->getFunction(Name: "objc_retainAutoreleasedReturnValue" ) || |
| 628 | M->getFunction(Name: "objc_unsafeClaimAutoreleasedReturnValue" ))); |
| 629 | })); |
| 630 | |
| 631 | // Analyzes and emits pseudos to support Win x64 Unwind V2. This pass must run |
| 632 | // after all real instructions have been added to the epilog. |
| 633 | if (TT.isOSWindows() && TT.isX86_64()) { |
| 634 | addPass(P: createX86WinEHUnwindV2LegacyPass()); |
| 635 | addPass(P: createX86WinEHUnwindV3Pass()); |
| 636 | } |
| 637 | } |
| 638 | |
| 639 | std::unique_ptr<CSEConfigBase> X86PassConfig::getCSEConfig() const { |
| 640 | return getStandardCSEConfigForOpt(Level: TM->getOptLevel()); |
| 641 | } |
| 642 | |
| 643 | static bool onlyAllocateTileRegisters(const TargetRegisterInfo &TRI, |
| 644 | const MachineRegisterInfo &MRI, |
| 645 | const Register Reg) { |
| 646 | const TargetRegisterClass *RC = MRI.getRegClass(Reg); |
| 647 | return static_cast<const X86RegisterInfo &>(TRI).isTileRegisterClass(RC); |
| 648 | } |
| 649 | |
| 650 | bool X86PassConfig::addRegAssignAndRewriteOptimized() { |
| 651 | // Don't support tile RA when RA is specified by command line "-regalloc". |
| 652 | if (!isCustomizedRegAlloc() && getX86TargetMachine().getCLOpts().tile_ra) { |
| 653 | // Allocate tile register first. |
| 654 | addPass(P: createGreedyRegisterAllocator(F: onlyAllocateTileRegisters)); |
| 655 | addPass(P: createX86TileConfigLegacyPass()); |
| 656 | } |
| 657 | return TargetPassConfig::addRegAssignAndRewriteOptimized(); |
| 658 | } |
| 659 | |