| 1 | //===-- ARMTargetMachine.cpp - Define TargetMachine for ARM ---------------===// |
| 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 | // |
| 10 | //===----------------------------------------------------------------------===// |
| 11 | |
| 12 | #include "ARMTargetMachine.h" |
| 13 | #include "ARM.h" |
| 14 | #include "ARMLatencyMutations.h" |
| 15 | #include "ARMMachineFunctionInfo.h" |
| 16 | #include "ARMMacroFusion.h" |
| 17 | #include "ARMSubtarget.h" |
| 18 | #include "ARMTargetObjectFile.h" |
| 19 | #include "ARMTargetTransformInfo.h" |
| 20 | #include "MCTargetDesc/ARMMCTargetDesc.h" |
| 21 | #include "TargetInfo/ARMTargetInfo.h" |
| 22 | #include "llvm/ADT/StringRef.h" |
| 23 | #include "llvm/Analysis/TargetTransformInfo.h" |
| 24 | #include "llvm/CodeGen/ExecutionDomainFix.h" |
| 25 | #include "llvm/CodeGen/GlobalISel/CSEInfo.h" |
| 26 | #include "llvm/CodeGen/GlobalISel/CallLowering.h" |
| 27 | #include "llvm/CodeGen/GlobalISel/IRTranslator.h" |
| 28 | #include "llvm/CodeGen/GlobalISel/InstructionSelect.h" |
| 29 | #include "llvm/CodeGen/GlobalISel/Legalizer.h" |
| 30 | #include "llvm/CodeGen/GlobalISel/LegalizerInfo.h" |
| 31 | #include "llvm/CodeGen/GlobalISel/RegBankSelect.h" |
| 32 | #include "llvm/CodeGen/MIRParser/MIParser.h" |
| 33 | #include "llvm/CodeGen/MachineFunction.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/CallingConv.h" |
| 39 | #include "llvm/IR/DataLayout.h" |
| 40 | #include "llvm/IR/DiagnosticInfo.h" |
| 41 | #include "llvm/IR/Function.h" |
| 42 | #include "llvm/IR/InstIterator.h" |
| 43 | #include "llvm/IR/InstrTypes.h" |
| 44 | #include "llvm/IR/Module.h" |
| 45 | #include "llvm/MC/TargetRegistry.h" |
| 46 | #include "llvm/Pass.h" |
| 47 | #include "llvm/Passes/PassBuilder.h" |
| 48 | #include "llvm/Support/CodeGen.h" |
| 49 | #include "llvm/Support/CommandLine.h" |
| 50 | #include "llvm/Support/Compiler.h" |
| 51 | #include "llvm/Support/ErrorHandling.h" |
| 52 | #include "llvm/Target/TargetLoweringObjectFile.h" |
| 53 | #include "llvm/Target/TargetOptions.h" |
| 54 | #include "llvm/TargetParser/ARMTargetParser.h" |
| 55 | #include "llvm/TargetParser/Triple.h" |
| 56 | #include "llvm/Transforms/CFGuard.h" |
| 57 | #include "llvm/Transforms/IPO.h" |
| 58 | #include "llvm/Transforms/Scalar.h" |
| 59 | #include <cassert> |
| 60 | #include <memory> |
| 61 | #include <optional> |
| 62 | #include <string> |
| 63 | |
| 64 | using namespace llvm; |
| 65 | |
| 66 | static cl::opt<bool> |
| 67 | DisableA15SDOptimization("disable-a15-sd-optimization" , cl::Hidden, |
| 68 | cl::desc("Inhibit optimization of S->D register accesses on A15" ), |
| 69 | cl::init(Val: false)); |
| 70 | |
| 71 | static cl::opt<bool> |
| 72 | EnableAtomicTidy("arm-atomic-cfg-tidy" , cl::Hidden, |
| 73 | cl::desc("Run SimplifyCFG after expanding atomic operations" |
| 74 | " to make use of cmpxchg flow-based information" ), |
| 75 | cl::init(Val: true)); |
| 76 | |
| 77 | static cl::opt<bool> |
| 78 | EnableARMLoadStoreOpt("arm-load-store-opt" , cl::Hidden, |
| 79 | cl::desc("Enable ARM load/store optimization pass" ), |
| 80 | cl::init(Val: true)); |
| 81 | |
| 82 | // FIXME: Unify control over GlobalMerge. |
| 83 | static cl::opt<cl::boolOrDefault> |
| 84 | EnableGlobalMerge("arm-global-merge" , cl::Hidden, |
| 85 | cl::desc("Enable the global merge pass" )); |
| 86 | |
| 87 | namespace llvm { |
| 88 | void initializeARMExecutionDomainFixPass(PassRegistry&); |
| 89 | } |
| 90 | |
| 91 | extern "C" LLVM_ABI LLVM_EXTERNAL_VISIBILITY void LLVMInitializeARMTarget() { |
| 92 | // Register the target. |
| 93 | RegisterTargetMachine<ARMLETargetMachine> X(getTheARMLETarget()); |
| 94 | RegisterTargetMachine<ARMLETargetMachine> A(getTheThumbLETarget()); |
| 95 | RegisterTargetMachine<ARMBETargetMachine> Y(getTheARMBETarget()); |
| 96 | RegisterTargetMachine<ARMBETargetMachine> B(getTheThumbBETarget()); |
| 97 | |
| 98 | PassRegistry &Registry = *PassRegistry::getPassRegistry(); |
| 99 | initializeGlobalISel(Registry); |
| 100 | initializeARMAsmPrinterPass(Registry); |
| 101 | initializeARMLoadStoreOptLegacyPass(Registry); |
| 102 | initializeARMPreAllocLoadStoreOptLegacyPass(Registry); |
| 103 | initializeARMParallelDSPPass(Registry); |
| 104 | initializeARMBranchTargetsPass(Registry); |
| 105 | initializeARMConstantIslandsPass(Registry); |
| 106 | initializeARMExecutionDomainFixPass(Registry); |
| 107 | initializeARMExpandPseudoPass(Registry); |
| 108 | initializeThumb2SizeReducePass(Registry); |
| 109 | initializeMVEVPTBlockPass(Registry); |
| 110 | initializeMVETPAndVPTOptimisationsPass(Registry); |
| 111 | initializeMVETailPredicationPass(Registry); |
| 112 | initializeARMLowOverheadLoopsPass(Registry); |
| 113 | initializeARMBlockPlacementPass(Registry); |
| 114 | initializeMVEGatherScatterLoweringPass(Registry); |
| 115 | initializeARMSLSHardeningPass(Registry); |
| 116 | initializeMVELaneInterleavingPass(Registry); |
| 117 | initializeARMFixCortexA57AES1742098Pass(Registry); |
| 118 | initializeARMDAGToDAGISelLegacyPass(Registry); |
| 119 | initializeMachineKCFILegacyPass(Registry); |
| 120 | } |
| 121 | |
| 122 | static std::unique_ptr<TargetLoweringObjectFile> createTLOF(const Triple &TT) { |
| 123 | if (TT.isOSBinFormatMachO()) |
| 124 | return std::make_unique<TargetLoweringObjectFileMachO>(); |
| 125 | if (TT.isOSWindows()) |
| 126 | return std::make_unique<TargetLoweringObjectFileCOFF>(); |
| 127 | return std::make_unique<ARMElfTargetObjectFile>(); |
| 128 | } |
| 129 | |
| 130 | static Reloc::Model getEffectiveRelocModel(const Triple &TT, |
| 131 | std::optional<Reloc::Model> RM) { |
| 132 | if (!RM) |
| 133 | // Default relocation model on Darwin is PIC. |
| 134 | return TT.isOSBinFormatMachO() ? Reloc::PIC_ : Reloc::Static; |
| 135 | |
| 136 | if (*RM == Reloc::ROPI || *RM == Reloc::RWPI || *RM == Reloc::ROPI_RWPI) |
| 137 | assert(TT.isOSBinFormatELF() && |
| 138 | "ROPI/RWPI currently only supported for ELF" ); |
| 139 | |
| 140 | // DynamicNoPIC is only used on darwin. |
| 141 | if (*RM == Reloc::DynamicNoPIC && !TT.isOSDarwin()) |
| 142 | return Reloc::Static; |
| 143 | |
| 144 | return *RM; |
| 145 | } |
| 146 | |
| 147 | /// Create an ARM architecture model. |
| 148 | /// |
| 149 | ARMBaseTargetMachine::ARMBaseTargetMachine(const Target &T, const Triple &TT, |
| 150 | StringRef CPU, StringRef FS, |
| 151 | const TargetOptions &Options, |
| 152 | std::optional<Reloc::Model> RM, |
| 153 | std::optional<CodeModel::Model> CM, |
| 154 | CodeGenOptLevel OL) |
| 155 | : CodeGenTargetMachineImpl( |
| 156 | T, TT.computeDataLayout(ABIName: Options.MCOptions.ABIName), TT, CPU, FS, |
| 157 | Options, getEffectiveRelocModel(TT, RM), |
| 158 | getEffectiveCodeModel(CM, Default: CodeModel::Small), OL), |
| 159 | TargetABI(ARM::computeTargetABI(TT, ABIName: Options.MCOptions.ABIName)), |
| 160 | TLOF(createTLOF(TT: getTargetTriple())), isLittle(TT.isLittleEndian()) { |
| 161 | |
| 162 | if (TT.isOSBinFormatMachO()) { |
| 163 | this->Options.TrapUnreachable = true; |
| 164 | this->Options.NoTrapAfterNoreturn = true; |
| 165 | } |
| 166 | |
| 167 | // ARM supports the debug entry values. |
| 168 | setSupportsDebugEntryValues(true); |
| 169 | |
| 170 | initAsmInfo(); |
| 171 | |
| 172 | // ARM supports the MachineOutliner. |
| 173 | setMachineOutliner(true); |
| 174 | setSupportsDefaultOutlining(true); |
| 175 | } |
| 176 | |
| 177 | ARMBaseTargetMachine::~ARMBaseTargetMachine() = default; |
| 178 | |
| 179 | MachineFunctionInfo *ARMBaseTargetMachine::createMachineFunctionInfo( |
| 180 | BumpPtrAllocator &Allocator, const Function &F, |
| 181 | const TargetSubtargetInfo *STI) const { |
| 182 | const auto *ARMSTI = static_cast<const ARMSubtarget *>(STI); |
| 183 | if (!ARMSTI->hasFPRegs() || ARMSTI->isThumb1Only() || |
| 184 | ARMSTI->useSoftFloat()) { |
| 185 | const StringRef FPRegsUnavailableMsg = |
| 186 | ", but floating-point registers are unavailable" ; |
| 187 | const ARMTargetLowering *TLI = ARMSTI->getTargetLowering(); |
| 188 | |
| 189 | if (TLI->getEffectiveCallingConv(CC: F.getCallingConv(), isVarArg: F.isVarArg()) == |
| 190 | CallingConv::ARM_AAPCS_VFP) { |
| 191 | F.getContext().diagnose(DI: DiagnosticInfoUnsupported( |
| 192 | F, Twine("calling convention is hard-float" ) + FPRegsUnavailableMsg, |
| 193 | DiagnosticLocation(F.getSubprogram()))); |
| 194 | } else { |
| 195 | for (const Instruction &I : instructions(F)) { |
| 196 | const auto *CB = dyn_cast<CallBase>(Val: &I); |
| 197 | if (!CB || CB->isInlineAsm() || |
| 198 | (CB->getCalledFunction() && CB->getCalledFunction()->isIntrinsic())) |
| 199 | continue; |
| 200 | if (TLI->getEffectiveCallingConv(CC: CB->getCallingConv(), |
| 201 | isVarArg: CB->getFunctionType()->isVarArg()) == |
| 202 | CallingConv::ARM_AAPCS_VFP) { |
| 203 | const Function *Callee = CB->getCalledFunction(); |
| 204 | F.getContext().diagnose(DI: DiagnosticInfoUnsupported( |
| 205 | F, |
| 206 | (Callee ? Twine("'" ) + F.getName() + "' calls '" + |
| 207 | Callee->getName() + "', which" |
| 208 | : Twine("'" ) + F.getName() + |
| 209 | "' makes an indirect call that" ) + |
| 210 | " expects a hard-float calling convention" + |
| 211 | FPRegsUnavailableMsg, |
| 212 | CB->getDebugLoc())); |
| 213 | } |
| 214 | } |
| 215 | } |
| 216 | } |
| 217 | return ARMFunctionInfo::create<ARMFunctionInfo>(Allocator, F, STI: ARMSTI); |
| 218 | } |
| 219 | |
| 220 | FloatABI::ABIType ARMBaseTargetMachine::getFloatABI(const Module &M) const { |
| 221 | // An explicit "float-abi" module flag always wins, even for AAPCS16. |
| 222 | if (auto *Val = dyn_cast_or_null<MDString>(Val: M.getModuleFlag(Key: "float-abi" ))) |
| 223 | return *FloatABI::parseABIType(S: Val->getString()); |
| 224 | |
| 225 | // With no explicit ABI, an explicit -target-abi=aapcs16 forces hard float |
| 226 | // even on triples whose default float ABI is soft (the triple default only |
| 227 | // detects AAPCS16 when it is the triple's own default ABI). |
| 228 | if (TargetABI == ARM::ARM_ABI_AAPCS16) |
| 229 | return FloatABI::Hard; |
| 230 | // Otherwise fall back to the ABI implied by the target triple. |
| 231 | return M.getTargetTriple().getDefaultFloatABI(); |
| 232 | } |
| 233 | |
| 234 | ARM::ARMABI ARMBaseTargetMachine::getEffectiveABI(const Module &M) const { |
| 235 | // Consistency of "target-abi" and -target-abi is validated elsewhere. |
| 236 | if (const auto *MD = cast_or_null<MDString>(Val: M.getModuleFlag(Key: "target-abi" ))) |
| 237 | return ARM::computeTargetABI(TT: TargetTriple, ABIName: MD->getString()); |
| 238 | return TargetABI; |
| 239 | } |
| 240 | |
| 241 | const ARMSubtarget * |
| 242 | ARMBaseTargetMachine::getSubtargetImpl(const Function &F) const { |
| 243 | Attribute CPUAttr = F.getFnAttribute(Kind: "target-cpu" ); |
| 244 | Attribute FSAttr = F.getFnAttribute(Kind: "target-features" ); |
| 245 | |
| 246 | std::string CPU = |
| 247 | CPUAttr.isValid() ? CPUAttr.getValueAsString().str() : TargetCPU; |
| 248 | std::string FS = |
| 249 | FSAttr.isValid() ? FSAttr.getValueAsString().str() : TargetFS; |
| 250 | |
| 251 | // FIXME: This is related to the code below to reset the target options, |
| 252 | // we need to know whether or not the soft float flag is set on the |
| 253 | // function before we can generate a subtarget. We also need to use |
| 254 | // it as a key for the subtarget since that can be the only difference |
| 255 | // between two functions. |
| 256 | bool SoftFloat = F.getFnAttribute(Kind: "use-soft-float" ).getValueAsBool(); |
| 257 | // If the soft float attribute is set on the function turn on the soft float |
| 258 | // subtarget feature. |
| 259 | if (SoftFloat) |
| 260 | FS += FS.empty() ? "+soft-float" : ",+soft-float" ; |
| 261 | |
| 262 | // Use the optminsize to identify the subtarget, but don't use it in the |
| 263 | // feature string. |
| 264 | std::string Key = CPU + FS; |
| 265 | if (F.hasMinSize()) |
| 266 | Key += "+minsize" ; |
| 267 | |
| 268 | DenormalMode DM = F.getDenormalFPEnv().DefaultMode; |
| 269 | if (DM != DenormalMode::getIEEE()) |
| 270 | Key += "denormal-fp-math=" + DM.str(); |
| 271 | |
| 272 | FloatABI::ABIType FloatABI = getFloatABI(M: *F.getParent()); |
| 273 | // It is legal to have FloatABI::Hard for targets with SIMD registers |
| 274 | // but no floating-point hardware (mve+nofp). |
| 275 | Key += FloatABI == FloatABI::Hard ? "+hard-float-abi" : "+soft-float-abi" ; |
| 276 | |
| 277 | ARM::ARMABI ABI = getEffectiveABI(M: *F.getParent()); |
| 278 | Key += "+abi=" + std::to_string(val: (int)ABI); |
| 279 | |
| 280 | auto &I = SubtargetMap[Key]; |
| 281 | if (!I) { |
| 282 | I = std::make_unique<ARMSubtarget>(args: TargetTriple, args&: CPU, args&: FS, args: *this, args: isLittle, |
| 283 | args&: FloatABI, args&: ABI, args: F.hasMinSize(), args&: DM); |
| 284 | |
| 285 | if (!I->isThumb() && !I->hasARMOps()) |
| 286 | F.getContext().emitError(ErrorStr: "Function '" + F.getName() + "' uses ARM " |
| 287 | "instructions, but the target does not support ARM mode execution." ); |
| 288 | } |
| 289 | |
| 290 | return I.get(); |
| 291 | } |
| 292 | |
| 293 | TargetTransformInfo |
| 294 | ARMBaseTargetMachine::getTargetTransformInfo(const Function &F) const { |
| 295 | return TargetTransformInfo(std::make_unique<ARMTTIImpl>(args: this, args: F)); |
| 296 | } |
| 297 | |
| 298 | ScheduleDAGInstrs * |
| 299 | ARMBaseTargetMachine::createMachineScheduler(MachineSchedContext *C) const { |
| 300 | ScheduleDAGMILive *DAG = createSchedLive(C); |
| 301 | // add DAG Mutations here. |
| 302 | const ARMSubtarget &ST = C->MF->getSubtarget<ARMSubtarget>(); |
| 303 | if (ST.hasFusion()) |
| 304 | DAG->addMutation(Mutation: createARMMacroFusionDAGMutation()); |
| 305 | return DAG; |
| 306 | } |
| 307 | |
| 308 | ScheduleDAGInstrs * |
| 309 | ARMBaseTargetMachine::createPostMachineScheduler(MachineSchedContext *C) const { |
| 310 | ScheduleDAGMI *DAG = createSchedPostRA(C); |
| 311 | // add DAG Mutations here. |
| 312 | const ARMSubtarget &ST = C->MF->getSubtarget<ARMSubtarget>(); |
| 313 | if (ST.hasFusion()) |
| 314 | DAG->addMutation(Mutation: createARMMacroFusionDAGMutation()); |
| 315 | if (auto Mutation = createARMLatencyMutations(ST, AA: C->AA)) |
| 316 | DAG->addMutation(Mutation: std::move(Mutation)); |
| 317 | return DAG; |
| 318 | } |
| 319 | |
| 320 | ARMLETargetMachine::ARMLETargetMachine(const Target &T, const Triple &TT, |
| 321 | StringRef CPU, StringRef FS, |
| 322 | const TargetOptions &Options, |
| 323 | std::optional<Reloc::Model> RM, |
| 324 | std::optional<CodeModel::Model> CM, |
| 325 | CodeGenOptLevel OL, bool JIT) |
| 326 | : ARMBaseTargetMachine(T, TT, CPU, FS, Options, RM, CM, OL) {} |
| 327 | |
| 328 | ARMBETargetMachine::ARMBETargetMachine(const Target &T, const Triple &TT, |
| 329 | StringRef CPU, StringRef FS, |
| 330 | const TargetOptions &Options, |
| 331 | std::optional<Reloc::Model> RM, |
| 332 | std::optional<CodeModel::Model> CM, |
| 333 | CodeGenOptLevel OL, bool JIT) |
| 334 | : ARMBaseTargetMachine(T, TT, CPU, FS, Options, RM, CM, OL) {} |
| 335 | |
| 336 | namespace { |
| 337 | |
| 338 | /// ARM Code Generator Pass Configuration Options. |
| 339 | class ARMPassConfig : public TargetPassConfig { |
| 340 | public: |
| 341 | ARMPassConfig(ARMBaseTargetMachine &TM, PassManagerBase &PM) |
| 342 | : TargetPassConfig(TM, PM) {} |
| 343 | |
| 344 | ARMBaseTargetMachine &getARMTargetMachine() const { |
| 345 | return getTM<ARMBaseTargetMachine>(); |
| 346 | } |
| 347 | |
| 348 | void addIRPasses() override; |
| 349 | void addCodeGenPrepare() override; |
| 350 | bool addPreISel() override; |
| 351 | bool addInstSelector() override; |
| 352 | bool addIRTranslator() override; |
| 353 | bool addLegalizeMachineIR() override; |
| 354 | bool addRegBankSelect() override; |
| 355 | bool addGlobalInstructionSelect() override; |
| 356 | void addPreRegAlloc() override; |
| 357 | void addPreSched2() override; |
| 358 | void addPreEmitPass() override; |
| 359 | void addPreEmitPass2() override; |
| 360 | |
| 361 | std::unique_ptr<CSEConfigBase> getCSEConfig() const override; |
| 362 | }; |
| 363 | |
| 364 | class ARMExecutionDomainFix : public ExecutionDomainFix { |
| 365 | public: |
| 366 | static char ID; |
| 367 | ARMExecutionDomainFix() : ExecutionDomainFix(ID, ARM::DPRRegClass) {} |
| 368 | StringRef getPassName() const override { |
| 369 | return "ARM Execution Domain Fix" ; |
| 370 | } |
| 371 | }; |
| 372 | char ARMExecutionDomainFix::ID; |
| 373 | |
| 374 | } // end anonymous namespace |
| 375 | |
| 376 | INITIALIZE_PASS_BEGIN(ARMExecutionDomainFix, "arm-execution-domain-fix" , |
| 377 | "ARM Execution Domain Fix" , false, false) |
| 378 | INITIALIZE_PASS_DEPENDENCY(ReachingDefInfoWrapperPass) |
| 379 | INITIALIZE_PASS_END(ARMExecutionDomainFix, "arm-execution-domain-fix" , |
| 380 | "ARM Execution Domain Fix" , false, false) |
| 381 | |
| 382 | void ARMBaseTargetMachine::registerPassBuilderCallbacks(PassBuilder &PB) { |
| 383 | #define GET_PASS_REGISTRY "ARMPassRegistry.def" |
| 384 | #include "llvm/Passes/TargetPassRegistry.inc" |
| 385 | } |
| 386 | |
| 387 | TargetPassConfig *ARMBaseTargetMachine::createPassConfig(PassManagerBase &PM) { |
| 388 | return new ARMPassConfig(*this, PM); |
| 389 | } |
| 390 | |
| 391 | std::unique_ptr<CSEConfigBase> ARMPassConfig::getCSEConfig() const { |
| 392 | return getStandardCSEConfigForOpt(Level: TM->getOptLevel()); |
| 393 | } |
| 394 | |
| 395 | void ARMPassConfig::addIRPasses() { |
| 396 | addPass(P: createAtomicExpandLegacyPass()); |
| 397 | |
| 398 | // Cmpxchg instructions are often used with a subsequent comparison to |
| 399 | // determine whether it succeeded. We can exploit existing control-flow in |
| 400 | // ldrex/strex loops to simplify this, but it needs tidying up. |
| 401 | if (TM->getOptLevel() != CodeGenOptLevel::None && EnableAtomicTidy) |
| 402 | addPass(P: createCFGSimplificationPass( |
| 403 | Options: SimplifyCFGOptions().hoistCommonInsts(B: true).sinkCommonInsts(B: true), |
| 404 | Ftor: [this](const Function &F) { |
| 405 | const auto &ST = this->TM->getSubtarget<ARMSubtarget>(F); |
| 406 | return ST.hasAnyDataBarrier() && !ST.isThumb1Only(); |
| 407 | })); |
| 408 | |
| 409 | addPass(P: createMVEGatherScatterLoweringPass()); |
| 410 | addPass(P: createMVELaneInterleavingPass()); |
| 411 | |
| 412 | TargetPassConfig::addIRPasses(); |
| 413 | |
| 414 | // Run the parallel DSP pass. |
| 415 | if (getOptLevel() == CodeGenOptLevel::Aggressive) |
| 416 | addPass(P: createARMParallelDSPPass()); |
| 417 | |
| 418 | // Match complex arithmetic patterns |
| 419 | if (TM->getOptLevel() >= CodeGenOptLevel::Default) |
| 420 | addPass(P: createComplexDeinterleavingPass(TM)); |
| 421 | |
| 422 | // Match interleaved memory accesses to ldN/stN intrinsics. |
| 423 | if (TM->getOptLevel() != CodeGenOptLevel::None) |
| 424 | addPass(P: createInterleavedAccessPass()); |
| 425 | |
| 426 | // Add Control Flow Guard checks. |
| 427 | if (TM->getTargetTriple().isOSWindows()) |
| 428 | addPass(P: createCFGuardPass()); |
| 429 | |
| 430 | if (TM->Options.JMCInstrument) |
| 431 | addPass(P: createJMCInstrumenterPass()); |
| 432 | } |
| 433 | |
| 434 | void ARMPassConfig::addCodeGenPrepare() { |
| 435 | if (getOptLevel() != CodeGenOptLevel::None) |
| 436 | addPass(P: createTypePromotionLegacyPass()); |
| 437 | TargetPassConfig::addCodeGenPrepare(); |
| 438 | } |
| 439 | |
| 440 | bool ARMPassConfig::addPreISel() { |
| 441 | if ((TM->getOptLevel() != CodeGenOptLevel::None && |
| 442 | EnableGlobalMerge == cl::boolOrDefault::BOU_UNSET) || |
| 443 | EnableGlobalMerge == cl::boolOrDefault::BOU_TRUE) { |
| 444 | // FIXME: This is using the thumb1 only constant value for |
| 445 | // maximal global offset for merging globals. We may want |
| 446 | // to look into using the old value for non-thumb1 code of |
| 447 | // 4095 based on the TargetMachine, but this starts to become |
| 448 | // tricky when doing code gen per function. |
| 449 | bool OnlyOptimizeForSize = |
| 450 | (TM->getOptLevel() < CodeGenOptLevel::Aggressive) && |
| 451 | (EnableGlobalMerge == cl::boolOrDefault::BOU_UNSET); |
| 452 | // Merging of extern globals is enabled by default on non-Mach-O as we |
| 453 | // expect it to be generally either beneficial or harmless. On Mach-O it |
| 454 | // is disabled as we emit the .subsections_via_symbols directive which |
| 455 | // means that merging extern globals is not safe. |
| 456 | bool MergeExternalByDefault = !TM->getTargetTriple().isOSBinFormatMachO(); |
| 457 | addPass(P: createGlobalMergePass(TM, MaximalOffset: 127, OnlyOptimizeForSize, |
| 458 | MergeExternalByDefault)); |
| 459 | } |
| 460 | |
| 461 | if (TM->getOptLevel() != CodeGenOptLevel::None) { |
| 462 | addPass(P: createHardwareLoopsLegacyPass()); |
| 463 | addPass(P: createMVETailPredicationPass()); |
| 464 | // FIXME: IR passes can delete address-taken basic blocks, deleting |
| 465 | // corresponding blockaddresses. ARMConstantPoolConstant holds references to |
| 466 | // address-taken basic blocks which can be invalidated if the function |
| 467 | // containing the blockaddress has already been codegen'd and the basic |
| 468 | // block is removed. Work around this by forcing all IR passes to run before |
| 469 | // any ISel takes place. We should have a more principled way of handling |
| 470 | // this. See D99707 for more details. |
| 471 | addPass(P: createBarrierNoopPass()); |
| 472 | } |
| 473 | |
| 474 | return false; |
| 475 | } |
| 476 | |
| 477 | bool ARMPassConfig::addInstSelector() { |
| 478 | addPass(P: createARMISelDag(TM&: getARMTargetMachine(), OptLevel: getOptLevel())); |
| 479 | return false; |
| 480 | } |
| 481 | |
| 482 | bool ARMPassConfig::addIRTranslator() { |
| 483 | addPass(P: new IRTranslatorLegacy(getOptLevel())); |
| 484 | return false; |
| 485 | } |
| 486 | |
| 487 | bool ARMPassConfig::addLegalizeMachineIR() { |
| 488 | addPass(P: new LegalizerLegacy()); |
| 489 | return false; |
| 490 | } |
| 491 | |
| 492 | bool ARMPassConfig::addRegBankSelect() { |
| 493 | addPass(P: new RegBankSelectLegacy()); |
| 494 | return false; |
| 495 | } |
| 496 | |
| 497 | bool ARMPassConfig::addGlobalInstructionSelect() { |
| 498 | addPass(P: new InstructionSelectLegacy(getOptLevel())); |
| 499 | return false; |
| 500 | } |
| 501 | |
| 502 | void ARMPassConfig::addPreRegAlloc() { |
| 503 | if (getOptLevel() != CodeGenOptLevel::None) { |
| 504 | if (getOptLevel() == CodeGenOptLevel::Aggressive) |
| 505 | addPass(PassID: &MachinePipelinerID); |
| 506 | |
| 507 | addPass(P: createMVETPAndVPTOptimisationsPass()); |
| 508 | |
| 509 | addPass(P: createMLxExpansionPass()); |
| 510 | |
| 511 | if (EnableARMLoadStoreOpt) |
| 512 | addPass(P: createARMLoadStoreOptLegacyPass(/* pre-register alloc */ PreAlloc: true)); |
| 513 | |
| 514 | if (!DisableA15SDOptimization) |
| 515 | addPass(P: createA15SDOptimizerPass()); |
| 516 | } |
| 517 | } |
| 518 | |
| 519 | void ARMPassConfig::addPreSched2() { |
| 520 | if (getOptLevel() != CodeGenOptLevel::None) { |
| 521 | if (EnableARMLoadStoreOpt) |
| 522 | addPass(P: createARMLoadStoreOptLegacyPass()); |
| 523 | |
| 524 | addPass(P: new ARMExecutionDomainFix()); |
| 525 | addPass(P: createBreakFalseDepsLegacyPass()); |
| 526 | } |
| 527 | |
| 528 | // Expand some pseudo instructions into multiple instructions to allow |
| 529 | // proper scheduling. |
| 530 | addPass(P: createARMExpandPseudoPass()); |
| 531 | |
| 532 | // Emit KCFI checks for indirect calls. |
| 533 | addPass(P: createKCFIPass()); |
| 534 | |
| 535 | if (getOptLevel() != CodeGenOptLevel::None) { |
| 536 | // When optimising for size, always run the Thumb2SizeReduction pass before |
| 537 | // IfConversion. Otherwise, check whether IT blocks are restricted |
| 538 | // (e.g. in v8, IfConversion depends on Thumb instruction widths) |
| 539 | addPass(P: createThumb2SizeReductionPass(Ftor: [this](const Function &F) { |
| 540 | return this->TM->getSubtarget<ARMSubtarget>(F).hasMinSize() || |
| 541 | this->TM->getSubtarget<ARMSubtarget>(F).restrictIT(); |
| 542 | })); |
| 543 | |
| 544 | addPass(P: createIfConverter(Ftor: [](const MachineFunction &MF) { |
| 545 | return !MF.getSubtarget<ARMSubtarget>().isThumb1Only(); |
| 546 | })); |
| 547 | } |
| 548 | addPass(P: createThumb2ITBlockPass()); |
| 549 | |
| 550 | // Add both scheduling passes to give the subtarget an opportunity to pick |
| 551 | // between them. |
| 552 | if (getOptLevel() != CodeGenOptLevel::None) { |
| 553 | addPass(PassID: &PostMachineSchedulerID); |
| 554 | addPass(PassID: &PostRASchedulerID); |
| 555 | } |
| 556 | |
| 557 | addPass(P: createMVEVPTBlockPass()); |
| 558 | addPass(P: createARMIndirectThunks()); |
| 559 | addPass(P: createARMSLSHardeningPass()); |
| 560 | } |
| 561 | |
| 562 | void ARMPassConfig::addPreEmitPass() { |
| 563 | addPass(P: createThumb2SizeReductionPass()); |
| 564 | |
| 565 | // Unpack bundles for: |
| 566 | // - Thumb2: Constant island pass requires unbundled instructions |
| 567 | // - KCFI: KCFI_CHECK pseudo instructions need to be unbundled for AsmPrinter |
| 568 | addPass(P: createUnpackMachineBundlesLegacy(Ftor: [](const MachineFunction &MF) { |
| 569 | return MF.getSubtarget<ARMSubtarget>().isThumb2() || |
| 570 | MF.getFunction().getParent()->getModuleFlag(Key: "kcfi" ); |
| 571 | })); |
| 572 | |
| 573 | // Don't optimize barriers or block placement at -O0. |
| 574 | if (getOptLevel() != CodeGenOptLevel::None) { |
| 575 | addPass(P: createARMBlockPlacementPass()); |
| 576 | addPass(P: createARMOptimizeBarriersPass()); |
| 577 | } |
| 578 | } |
| 579 | |
| 580 | void ARMPassConfig::addPreEmitPass2() { |
| 581 | |
| 582 | // Inserts fixup instructions before unsafe AES operations. Instructions may |
| 583 | // be inserted at the start of blocks and at within blocks so this pass has to |
| 584 | // come before those below. |
| 585 | addPass(P: createARMFixCortexA57AES1742098Pass()); |
| 586 | // Inserts BTIs at the start of functions and indirectly-called basic blocks, |
| 587 | // so passes cannot add to the start of basic blocks once this has run. |
| 588 | addPass(P: createARMBranchTargetsPass()); |
| 589 | // Inserts Constant Islands. Block sizes cannot be increased after this point, |
| 590 | // as this may push the branch ranges and load offsets of accessing constant |
| 591 | // pools out of range.. |
| 592 | addPass(P: createARMConstantIslandPass()); |
| 593 | // Finalises Low-Overhead Loops. This replaces pseudo instructions with real |
| 594 | // instructions, but the pseudos all have conservative sizes so that block |
| 595 | // sizes will only be decreased by this pass. |
| 596 | addPass(P: createARMLowOverheadLoopsPass()); |
| 597 | |
| 598 | if (TM->getTargetTriple().isOSWindows()) { |
| 599 | // Identify valid longjmp targets for Windows Control Flow Guard. |
| 600 | addPass(P: createCFGuardLongjmpPass()); |
| 601 | // Identify valid eh continuation targets for Windows EHCont Guard. |
| 602 | addPass(P: createEHContGuardTargetsLegacy()); |
| 603 | } |
| 604 | } |
| 605 | |
| 606 | yaml::MachineFunctionInfo * |
| 607 | ARMBaseTargetMachine::createDefaultFuncInfoYAML() const { |
| 608 | return new yaml::ARMFunctionInfo(); |
| 609 | } |
| 610 | |
| 611 | yaml::MachineFunctionInfo * |
| 612 | ARMBaseTargetMachine::convertFuncInfoToYAML(const MachineFunction &MF) const { |
| 613 | const auto *MFI = MF.getInfo<ARMFunctionInfo>(); |
| 614 | return new yaml::ARMFunctionInfo(*MFI); |
| 615 | } |
| 616 | |
| 617 | bool ARMBaseTargetMachine::parseMachineFunctionInfo( |
| 618 | const yaml::MachineFunctionInfo &MFI, PerFunctionMIParsingState &PFS, |
| 619 | SMDiagnostic &Error, SMRange &SourceRange) const { |
| 620 | const auto &YamlMFI = static_cast<const yaml::ARMFunctionInfo &>(MFI); |
| 621 | MachineFunction &MF = PFS.MF; |
| 622 | MF.getInfo<ARMFunctionInfo>()->initializeBaseYamlFields(YamlMFI); |
| 623 | return false; |
| 624 | } |
| 625 | |
| 626 | void ARMBaseTargetMachine::reset() { SubtargetMap.clear(); } |
| 627 | |