1 //===-- AArch64TargetMachine.cpp - Define TargetMachine for AArch64 -------===// 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 "AArch64TargetMachine.h" 13 #include "AArch64.h" 14 #include "AArch64MachineFunctionInfo.h" 15 #include "AArch64MacroFusion.h" 16 #include "AArch64Subtarget.h" 17 #include "AArch64TargetObjectFile.h" 18 #include "AArch64TargetTransformInfo.h" 19 #include "MCTargetDesc/AArch64MCTargetDesc.h" 20 #include "TargetInfo/AArch64TargetInfo.h" 21 #include "llvm/ADT/STLExtras.h" 22 #include "llvm/ADT/Triple.h" 23 #include "llvm/Analysis/TargetTransformInfo.h" 24 #include "llvm/CodeGen/CSEConfigBase.h" 25 #include "llvm/CodeGen/GlobalISel/IRTranslator.h" 26 #include "llvm/CodeGen/GlobalISel/InstructionSelect.h" 27 #include "llvm/CodeGen/GlobalISel/Legalizer.h" 28 #include "llvm/CodeGen/GlobalISel/Localizer.h" 29 #include "llvm/CodeGen/GlobalISel/RegBankSelect.h" 30 #include "llvm/CodeGen/MIRParser/MIParser.h" 31 #include "llvm/CodeGen/MachineScheduler.h" 32 #include "llvm/CodeGen/Passes.h" 33 #include "llvm/CodeGen/TargetPassConfig.h" 34 #include "llvm/IR/Attributes.h" 35 #include "llvm/IR/Function.h" 36 #include "llvm/InitializePasses.h" 37 #include "llvm/MC/MCAsmInfo.h" 38 #include "llvm/MC/MCTargetOptions.h" 39 #include "llvm/MC/TargetRegistry.h" 40 #include "llvm/Pass.h" 41 #include "llvm/Support/CodeGen.h" 42 #include "llvm/Support/CommandLine.h" 43 #include "llvm/Target/TargetLoweringObjectFile.h" 44 #include "llvm/Target/TargetOptions.h" 45 #include "llvm/Transforms/CFGuard.h" 46 #include "llvm/Transforms/Scalar.h" 47 #include <memory> 48 #include <string> 49 50 using namespace llvm; 51 52 static cl::opt<bool> EnableCCMP("aarch64-enable-ccmp", 53 cl::desc("Enable the CCMP formation pass"), 54 cl::init(true), cl::Hidden); 55 56 static cl::opt<bool> 57 EnableCondBrTuning("aarch64-enable-cond-br-tune", 58 cl::desc("Enable the conditional branch tuning pass"), 59 cl::init(true), cl::Hidden); 60 61 static cl::opt<bool> EnableMCR("aarch64-enable-mcr", 62 cl::desc("Enable the machine combiner pass"), 63 cl::init(true), cl::Hidden); 64 65 static cl::opt<bool> EnableStPairSuppress("aarch64-enable-stp-suppress", 66 cl::desc("Suppress STP for AArch64"), 67 cl::init(true), cl::Hidden); 68 69 static cl::opt<bool> EnableAdvSIMDScalar( 70 "aarch64-enable-simd-scalar", 71 cl::desc("Enable use of AdvSIMD scalar integer instructions"), 72 cl::init(false), cl::Hidden); 73 74 static cl::opt<bool> 75 EnablePromoteConstant("aarch64-enable-promote-const", 76 cl::desc("Enable the promote constant pass"), 77 cl::init(true), cl::Hidden); 78 79 static cl::opt<bool> EnableCollectLOH( 80 "aarch64-enable-collect-loh", 81 cl::desc("Enable the pass that emits the linker optimization hints (LOH)"), 82 cl::init(true), cl::Hidden); 83 84 static cl::opt<bool> 85 EnableDeadRegisterElimination("aarch64-enable-dead-defs", cl::Hidden, 86 cl::desc("Enable the pass that removes dead" 87 " definitons and replaces stores to" 88 " them with stores to the zero" 89 " register"), 90 cl::init(true)); 91 92 static cl::opt<bool> EnableRedundantCopyElimination( 93 "aarch64-enable-copyelim", 94 cl::desc("Enable the redundant copy elimination pass"), cl::init(true), 95 cl::Hidden); 96 97 static cl::opt<bool> EnableLoadStoreOpt("aarch64-enable-ldst-opt", 98 cl::desc("Enable the load/store pair" 99 " optimization pass"), 100 cl::init(true), cl::Hidden); 101 102 static cl::opt<bool> EnableAtomicTidy( 103 "aarch64-enable-atomic-cfg-tidy", cl::Hidden, 104 cl::desc("Run SimplifyCFG after expanding atomic operations" 105 " to make use of cmpxchg flow-based information"), 106 cl::init(true)); 107 108 static cl::opt<bool> 109 EnableEarlyIfConversion("aarch64-enable-early-ifcvt", cl::Hidden, 110 cl::desc("Run early if-conversion"), 111 cl::init(true)); 112 113 static cl::opt<bool> 114 EnableCondOpt("aarch64-enable-condopt", 115 cl::desc("Enable the condition optimizer pass"), 116 cl::init(true), cl::Hidden); 117 118 static cl::opt<bool> 119 EnableA53Fix835769("aarch64-fix-cortex-a53-835769", cl::Hidden, 120 cl::desc("Work around Cortex-A53 erratum 835769"), 121 cl::init(false)); 122 123 static cl::opt<bool> 124 EnableGEPOpt("aarch64-enable-gep-opt", cl::Hidden, 125 cl::desc("Enable optimizations on complex GEPs"), 126 cl::init(false)); 127 128 static cl::opt<bool> 129 BranchRelaxation("aarch64-enable-branch-relax", cl::Hidden, cl::init(true), 130 cl::desc("Relax out of range conditional branches")); 131 132 static cl::opt<bool> EnableCompressJumpTables( 133 "aarch64-enable-compress-jump-tables", cl::Hidden, cl::init(true), 134 cl::desc("Use smallest entry possible for jump tables")); 135 136 // FIXME: Unify control over GlobalMerge. 137 static cl::opt<cl::boolOrDefault> 138 EnableGlobalMerge("aarch64-enable-global-merge", cl::Hidden, 139 cl::desc("Enable the global merge pass")); 140 141 static cl::opt<bool> 142 EnableLoopDataPrefetch("aarch64-enable-loop-data-prefetch", cl::Hidden, 143 cl::desc("Enable the loop data prefetch pass"), 144 cl::init(true)); 145 146 static cl::opt<int> EnableGlobalISelAtO( 147 "aarch64-enable-global-isel-at-O", cl::Hidden, 148 cl::desc("Enable GlobalISel at or below an opt level (-1 to disable)"), 149 cl::init(0)); 150 151 static cl::opt<bool> 152 EnableSVEIntrinsicOpts("aarch64-enable-sve-intrinsic-opts", cl::Hidden, 153 cl::desc("Enable SVE intrinsic opts"), 154 cl::init(true)); 155 156 static cl::opt<bool> EnableFalkorHWPFFix("aarch64-enable-falkor-hwpf-fix", 157 cl::init(true), cl::Hidden); 158 159 static cl::opt<bool> 160 EnableBranchTargets("aarch64-enable-branch-targets", cl::Hidden, 161 cl::desc("Enable the AArch64 branch target pass"), 162 cl::init(true)); 163 164 static cl::opt<unsigned> SVEVectorBitsMaxOpt( 165 "aarch64-sve-vector-bits-max", 166 cl::desc("Assume SVE vector registers are at most this big, " 167 "with zero meaning no maximum size is assumed."), 168 cl::init(0), cl::Hidden); 169 170 static cl::opt<unsigned> SVEVectorBitsMinOpt( 171 "aarch64-sve-vector-bits-min", 172 cl::desc("Assume SVE vector registers are at least this big, " 173 "with zero meaning no minimum size is assumed."), 174 cl::init(0), cl::Hidden); 175 176 extern cl::opt<bool> EnableHomogeneousPrologEpilog; 177 178 extern "C" LLVM_EXTERNAL_VISIBILITY void LLVMInitializeAArch64Target() { 179 // Register the target. 180 RegisterTargetMachine<AArch64leTargetMachine> X(getTheAArch64leTarget()); 181 RegisterTargetMachine<AArch64beTargetMachine> Y(getTheAArch64beTarget()); 182 RegisterTargetMachine<AArch64leTargetMachine> Z(getTheARM64Target()); 183 RegisterTargetMachine<AArch64leTargetMachine> W(getTheARM64_32Target()); 184 RegisterTargetMachine<AArch64leTargetMachine> V(getTheAArch64_32Target()); 185 auto PR = PassRegistry::getPassRegistry(); 186 initializeGlobalISel(*PR); 187 initializeAArch64A53Fix835769Pass(*PR); 188 initializeAArch64A57FPLoadBalancingPass(*PR); 189 initializeAArch64AdvSIMDScalarPass(*PR); 190 initializeAArch64BranchTargetsPass(*PR); 191 initializeAArch64CollectLOHPass(*PR); 192 initializeAArch64CompressJumpTablesPass(*PR); 193 initializeAArch64ConditionalComparesPass(*PR); 194 initializeAArch64ConditionOptimizerPass(*PR); 195 initializeAArch64DeadRegisterDefinitionsPass(*PR); 196 initializeAArch64ExpandPseudoPass(*PR); 197 initializeAArch64LoadStoreOptPass(*PR); 198 initializeAArch64MIPeepholeOptPass(*PR); 199 initializeAArch64SIMDInstrOptPass(*PR); 200 initializeAArch64O0PreLegalizerCombinerPass(*PR); 201 initializeAArch64PreLegalizerCombinerPass(*PR); 202 initializeAArch64PostLegalizerCombinerPass(*PR); 203 initializeAArch64PostLegalizerLoweringPass(*PR); 204 initializeAArch64PostSelectOptimizePass(*PR); 205 initializeAArch64PromoteConstantPass(*PR); 206 initializeAArch64RedundantCopyEliminationPass(*PR); 207 initializeAArch64StorePairSuppressPass(*PR); 208 initializeFalkorHWPFFixPass(*PR); 209 initializeFalkorMarkStridedAccessesLegacyPass(*PR); 210 initializeLDTLSCleanupPass(*PR); 211 initializeSVEIntrinsicOptsPass(*PR); 212 initializeAArch64SpeculationHardeningPass(*PR); 213 initializeAArch64SLSHardeningPass(*PR); 214 initializeAArch64StackTaggingPass(*PR); 215 initializeAArch64StackTaggingPreRAPass(*PR); 216 initializeAArch64LowerHomogeneousPrologEpilogPass(*PR); 217 } 218 219 //===----------------------------------------------------------------------===// 220 // AArch64 Lowering public interface. 221 //===----------------------------------------------------------------------===// 222 static std::unique_ptr<TargetLoweringObjectFile> createTLOF(const Triple &TT) { 223 if (TT.isOSBinFormatMachO()) 224 return std::make_unique<AArch64_MachoTargetObjectFile>(); 225 if (TT.isOSBinFormatCOFF()) 226 return std::make_unique<AArch64_COFFTargetObjectFile>(); 227 228 return std::make_unique<AArch64_ELFTargetObjectFile>(); 229 } 230 231 // Helper function to build a DataLayout string 232 static std::string computeDataLayout(const Triple &TT, 233 const MCTargetOptions &Options, 234 bool LittleEndian) { 235 if (TT.isOSBinFormatMachO()) { 236 if (TT.getArch() == Triple::aarch64_32) 237 return "e-m:o-p:32:32-i64:64-i128:128-n32:64-S128"; 238 return "e-m:o-i64:64-i128:128-n32:64-S128"; 239 } 240 if (TT.isOSBinFormatCOFF()) 241 return "e-m:w-p:64:64-i32:32-i64:64-i128:128-n32:64-S128"; 242 std::string Endian = LittleEndian ? "e" : "E"; 243 std::string Ptr32 = TT.getEnvironment() == Triple::GNUILP32 ? "-p:32:32" : ""; 244 return Endian + "-m:e" + Ptr32 + 245 "-i8:8:32-i16:16:32-i64:64-i128:128-n32:64-S128"; 246 } 247 248 static StringRef computeDefaultCPU(const Triple &TT, StringRef CPU) { 249 if (CPU.empty() && TT.isArm64e()) 250 return "apple-a12"; 251 return CPU; 252 } 253 254 static Reloc::Model getEffectiveRelocModel(const Triple &TT, 255 Optional<Reloc::Model> RM) { 256 // AArch64 Darwin and Windows are always PIC. 257 if (TT.isOSDarwin() || TT.isOSWindows()) 258 return Reloc::PIC_; 259 // On ELF platforms the default static relocation model has a smart enough 260 // linker to cope with referencing external symbols defined in a shared 261 // library. Hence DynamicNoPIC doesn't need to be promoted to PIC. 262 if (!RM.hasValue() || *RM == Reloc::DynamicNoPIC) 263 return Reloc::Static; 264 return *RM; 265 } 266 267 static CodeModel::Model 268 getEffectiveAArch64CodeModel(const Triple &TT, Optional<CodeModel::Model> CM, 269 bool JIT) { 270 if (CM) { 271 if (*CM != CodeModel::Small && *CM != CodeModel::Tiny && 272 *CM != CodeModel::Large) { 273 report_fatal_error( 274 "Only small, tiny and large code models are allowed on AArch64"); 275 } else if (*CM == CodeModel::Tiny && !TT.isOSBinFormatELF()) 276 report_fatal_error("tiny code model is only supported on ELF"); 277 return *CM; 278 } 279 // The default MCJIT memory managers make no guarantees about where they can 280 // find an executable page; JITed code needs to be able to refer to globals 281 // no matter how far away they are. 282 // We should set the CodeModel::Small for Windows ARM64 in JIT mode, 283 // since with large code model LLVM generating 4 MOV instructions, and 284 // Windows doesn't support relocating these long branch (4 MOVs). 285 if (JIT && !TT.isOSWindows()) 286 return CodeModel::Large; 287 return CodeModel::Small; 288 } 289 290 /// Create an AArch64 architecture model. 291 /// 292 AArch64TargetMachine::AArch64TargetMachine(const Target &T, const Triple &TT, 293 StringRef CPU, StringRef FS, 294 const TargetOptions &Options, 295 Optional<Reloc::Model> RM, 296 Optional<CodeModel::Model> CM, 297 CodeGenOpt::Level OL, bool JIT, 298 bool LittleEndian) 299 : LLVMTargetMachine(T, 300 computeDataLayout(TT, Options.MCOptions, LittleEndian), 301 TT, computeDefaultCPU(TT, CPU), FS, Options, 302 getEffectiveRelocModel(TT, RM), 303 getEffectiveAArch64CodeModel(TT, CM, JIT), OL), 304 TLOF(createTLOF(getTargetTriple())), isLittle(LittleEndian) { 305 initAsmInfo(); 306 307 if (TT.isOSBinFormatMachO()) { 308 this->Options.TrapUnreachable = true; 309 this->Options.NoTrapAfterNoreturn = true; 310 } 311 312 if (getMCAsmInfo()->usesWindowsCFI()) { 313 // Unwinding can get confused if the last instruction in an 314 // exception-handling region (function, funclet, try block, etc.) 315 // is a call. 316 // 317 // FIXME: We could elide the trap if the next instruction would be in 318 // the same region anyway. 319 this->Options.TrapUnreachable = true; 320 } 321 322 if (this->Options.TLSSize == 0) // default 323 this->Options.TLSSize = 24; 324 if ((getCodeModel() == CodeModel::Small || 325 getCodeModel() == CodeModel::Kernel) && 326 this->Options.TLSSize > 32) 327 // for the small (and kernel) code model, the maximum TLS size is 4GiB 328 this->Options.TLSSize = 32; 329 else if (getCodeModel() == CodeModel::Tiny && this->Options.TLSSize > 24) 330 // for the tiny code model, the maximum TLS size is 1MiB (< 16MiB) 331 this->Options.TLSSize = 24; 332 333 // Enable GlobalISel at or below EnableGlobalISelAt0, unless this is 334 // MachO/CodeModel::Large, which GlobalISel does not support. 335 if (getOptLevel() <= EnableGlobalISelAtO && 336 TT.getArch() != Triple::aarch64_32 && 337 TT.getEnvironment() != Triple::GNUILP32 && 338 !(getCodeModel() == CodeModel::Large && TT.isOSBinFormatMachO())) { 339 setGlobalISel(true); 340 setGlobalISelAbort(GlobalISelAbortMode::Disable); 341 } 342 343 // AArch64 supports the MachineOutliner. 344 setMachineOutliner(true); 345 346 // AArch64 supports default outlining behaviour. 347 setSupportsDefaultOutlining(true); 348 349 // AArch64 supports the debug entry values. 350 setSupportsDebugEntryValues(true); 351 } 352 353 AArch64TargetMachine::~AArch64TargetMachine() = default; 354 355 const AArch64Subtarget * 356 AArch64TargetMachine::getSubtargetImpl(const Function &F) const { 357 Attribute CPUAttr = F.getFnAttribute("target-cpu"); 358 Attribute TuneAttr = F.getFnAttribute("tune-cpu"); 359 Attribute FSAttr = F.getFnAttribute("target-features"); 360 361 std::string CPU = 362 CPUAttr.isValid() ? CPUAttr.getValueAsString().str() : TargetCPU; 363 std::string TuneCPU = 364 TuneAttr.isValid() ? TuneAttr.getValueAsString().str() : CPU; 365 std::string FS = 366 FSAttr.isValid() ? FSAttr.getValueAsString().str() : TargetFS; 367 368 SmallString<512> Key; 369 370 unsigned MinSVEVectorSize = 0; 371 unsigned MaxSVEVectorSize = 0; 372 Attribute VScaleRangeAttr = F.getFnAttribute(Attribute::VScaleRange); 373 if (VScaleRangeAttr.isValid()) { 374 std::tie(MinSVEVectorSize, MaxSVEVectorSize) = 375 VScaleRangeAttr.getVScaleRangeArgs(); 376 MinSVEVectorSize *= 128; 377 MaxSVEVectorSize *= 128; 378 } else { 379 MinSVEVectorSize = SVEVectorBitsMinOpt; 380 MaxSVEVectorSize = SVEVectorBitsMaxOpt; 381 } 382 383 assert(MinSVEVectorSize % 128 == 0 && 384 "SVE requires vector length in multiples of 128!"); 385 assert(MaxSVEVectorSize % 128 == 0 && 386 "SVE requires vector length in multiples of 128!"); 387 assert((MaxSVEVectorSize >= MinSVEVectorSize || MaxSVEVectorSize == 0) && 388 "Minimum SVE vector size should not be larger than its maximum!"); 389 390 // Sanitize user input in case of no asserts 391 if (MaxSVEVectorSize == 0) 392 MinSVEVectorSize = (MinSVEVectorSize / 128) * 128; 393 else { 394 MinSVEVectorSize = 395 (std::min(MinSVEVectorSize, MaxSVEVectorSize) / 128) * 128; 396 MaxSVEVectorSize = 397 (std::max(MinSVEVectorSize, MaxSVEVectorSize) / 128) * 128; 398 } 399 400 Key += "SVEMin"; 401 Key += std::to_string(MinSVEVectorSize); 402 Key += "SVEMax"; 403 Key += std::to_string(MaxSVEVectorSize); 404 Key += CPU; 405 Key += TuneCPU; 406 Key += FS; 407 408 auto &I = SubtargetMap[Key]; 409 if (!I) { 410 // This needs to be done before we create a new subtarget since any 411 // creation will depend on the TM and the code generation flags on the 412 // function that reside in TargetOptions. 413 resetTargetOptions(F); 414 I = std::make_unique<AArch64Subtarget>(TargetTriple, CPU, TuneCPU, FS, 415 *this, isLittle, MinSVEVectorSize, 416 MaxSVEVectorSize); 417 } 418 return I.get(); 419 } 420 421 void AArch64leTargetMachine::anchor() { } 422 423 AArch64leTargetMachine::AArch64leTargetMachine( 424 const Target &T, const Triple &TT, StringRef CPU, StringRef FS, 425 const TargetOptions &Options, Optional<Reloc::Model> RM, 426 Optional<CodeModel::Model> CM, CodeGenOpt::Level OL, bool JIT) 427 : AArch64TargetMachine(T, TT, CPU, FS, Options, RM, CM, OL, JIT, true) {} 428 429 void AArch64beTargetMachine::anchor() { } 430 431 AArch64beTargetMachine::AArch64beTargetMachine( 432 const Target &T, const Triple &TT, StringRef CPU, StringRef FS, 433 const TargetOptions &Options, Optional<Reloc::Model> RM, 434 Optional<CodeModel::Model> CM, CodeGenOpt::Level OL, bool JIT) 435 : AArch64TargetMachine(T, TT, CPU, FS, Options, RM, CM, OL, JIT, false) {} 436 437 namespace { 438 439 /// AArch64 Code Generator Pass Configuration Options. 440 class AArch64PassConfig : public TargetPassConfig { 441 public: 442 AArch64PassConfig(AArch64TargetMachine &TM, PassManagerBase &PM) 443 : TargetPassConfig(TM, PM) { 444 if (TM.getOptLevel() != CodeGenOpt::None) 445 substitutePass(&PostRASchedulerID, &PostMachineSchedulerID); 446 } 447 448 AArch64TargetMachine &getAArch64TargetMachine() const { 449 return getTM<AArch64TargetMachine>(); 450 } 451 452 ScheduleDAGInstrs * 453 createMachineScheduler(MachineSchedContext *C) const override { 454 const AArch64Subtarget &ST = C->MF->getSubtarget<AArch64Subtarget>(); 455 ScheduleDAGMILive *DAG = createGenericSchedLive(C); 456 DAG->addMutation(createLoadClusterDAGMutation(DAG->TII, DAG->TRI)); 457 DAG->addMutation(createStoreClusterDAGMutation(DAG->TII, DAG->TRI)); 458 if (ST.hasFusion()) 459 DAG->addMutation(createAArch64MacroFusionDAGMutation()); 460 return DAG; 461 } 462 463 ScheduleDAGInstrs * 464 createPostMachineScheduler(MachineSchedContext *C) const override { 465 const AArch64Subtarget &ST = C->MF->getSubtarget<AArch64Subtarget>(); 466 if (ST.hasFusion()) { 467 // Run the Macro Fusion after RA again since literals are expanded from 468 // pseudos then (v. addPreSched2()). 469 ScheduleDAGMI *DAG = createGenericSchedPostRA(C); 470 DAG->addMutation(createAArch64MacroFusionDAGMutation()); 471 return DAG; 472 } 473 474 return nullptr; 475 } 476 477 void addIRPasses() override; 478 bool addPreISel() override; 479 void addCodeGenPrepare() override; 480 bool addInstSelector() override; 481 bool addIRTranslator() override; 482 void addPreLegalizeMachineIR() override; 483 bool addLegalizeMachineIR() override; 484 void addPreRegBankSelect() override; 485 bool addRegBankSelect() override; 486 void addPreGlobalInstructionSelect() override; 487 bool addGlobalInstructionSelect() override; 488 void addMachineSSAOptimization() override; 489 bool addILPOpts() override; 490 void addPreRegAlloc() override; 491 void addPostRegAlloc() override; 492 void addPreSched2() override; 493 void addPreEmitPass() override; 494 void addPreEmitPass2() override; 495 496 std::unique_ptr<CSEConfigBase> getCSEConfig() const override; 497 }; 498 499 } // end anonymous namespace 500 501 TargetTransformInfo 502 AArch64TargetMachine::getTargetTransformInfo(const Function &F) { 503 return TargetTransformInfo(AArch64TTIImpl(this, F)); 504 } 505 506 TargetPassConfig *AArch64TargetMachine::createPassConfig(PassManagerBase &PM) { 507 return new AArch64PassConfig(*this, PM); 508 } 509 510 std::unique_ptr<CSEConfigBase> AArch64PassConfig::getCSEConfig() const { 511 return getStandardCSEConfigForOpt(TM->getOptLevel()); 512 } 513 514 void AArch64PassConfig::addIRPasses() { 515 // Always expand atomic operations, we don't deal with atomicrmw or cmpxchg 516 // ourselves. 517 addPass(createAtomicExpandPass()); 518 519 // Expand any SVE vector library calls that we can't code generate directly. 520 if (EnableSVEIntrinsicOpts && TM->getOptLevel() == CodeGenOpt::Aggressive) 521 addPass(createSVEIntrinsicOptsPass()); 522 523 // Cmpxchg instructions are often used with a subsequent comparison to 524 // determine whether it succeeded. We can exploit existing control-flow in 525 // ldrex/strex loops to simplify this, but it needs tidying up. 526 if (TM->getOptLevel() != CodeGenOpt::None && EnableAtomicTidy) 527 addPass(createCFGSimplificationPass(SimplifyCFGOptions() 528 .forwardSwitchCondToPhi(true) 529 .convertSwitchToLookupTable(true) 530 .needCanonicalLoops(false) 531 .hoistCommonInsts(true) 532 .sinkCommonInsts(true))); 533 534 // Run LoopDataPrefetch 535 // 536 // Run this before LSR to remove the multiplies involved in computing the 537 // pointer values N iterations ahead. 538 if (TM->getOptLevel() != CodeGenOpt::None) { 539 if (EnableLoopDataPrefetch) 540 addPass(createLoopDataPrefetchPass()); 541 if (EnableFalkorHWPFFix) 542 addPass(createFalkorMarkStridedAccessesPass()); 543 } 544 545 TargetPassConfig::addIRPasses(); 546 547 addPass(createAArch64StackTaggingPass( 548 /*IsOptNone=*/TM->getOptLevel() == CodeGenOpt::None)); 549 550 // Match interleaved memory accesses to ldN/stN intrinsics. 551 if (TM->getOptLevel() != CodeGenOpt::None) { 552 addPass(createInterleavedLoadCombinePass()); 553 addPass(createInterleavedAccessPass()); 554 } 555 556 if (TM->getOptLevel() == CodeGenOpt::Aggressive && EnableGEPOpt) { 557 // Call SeparateConstOffsetFromGEP pass to extract constants within indices 558 // and lower a GEP with multiple indices to either arithmetic operations or 559 // multiple GEPs with single index. 560 addPass(createSeparateConstOffsetFromGEPPass(true)); 561 // Call EarlyCSE pass to find and remove subexpressions in the lowered 562 // result. 563 addPass(createEarlyCSEPass()); 564 // Do loop invariant code motion in case part of the lowered result is 565 // invariant. 566 addPass(createLICMPass()); 567 } 568 569 // Add Control Flow Guard checks. 570 if (TM->getTargetTriple().isOSWindows()) 571 addPass(createCFGuardCheckPass()); 572 } 573 574 // Pass Pipeline Configuration 575 bool AArch64PassConfig::addPreISel() { 576 // Run promote constant before global merge, so that the promoted constants 577 // get a chance to be merged 578 if (TM->getOptLevel() != CodeGenOpt::None && EnablePromoteConstant) 579 addPass(createAArch64PromoteConstantPass()); 580 // FIXME: On AArch64, this depends on the type. 581 // Basically, the addressable offsets are up to 4095 * Ty.getSizeInBytes(). 582 // and the offset has to be a multiple of the related size in bytes. 583 if ((TM->getOptLevel() != CodeGenOpt::None && 584 EnableGlobalMerge == cl::BOU_UNSET) || 585 EnableGlobalMerge == cl::BOU_TRUE) { 586 bool OnlyOptimizeForSize = (TM->getOptLevel() < CodeGenOpt::Aggressive) && 587 (EnableGlobalMerge == cl::BOU_UNSET); 588 589 // Merging of extern globals is enabled by default on non-Mach-O as we 590 // expect it to be generally either beneficial or harmless. On Mach-O it 591 // is disabled as we emit the .subsections_via_symbols directive which 592 // means that merging extern globals is not safe. 593 bool MergeExternalByDefault = !TM->getTargetTriple().isOSBinFormatMachO(); 594 595 // FIXME: extern global merging is only enabled when we optimise for size 596 // because there are some regressions with it also enabled for performance. 597 if (!OnlyOptimizeForSize) 598 MergeExternalByDefault = false; 599 600 addPass(createGlobalMergePass(TM, 4095, OnlyOptimizeForSize, 601 MergeExternalByDefault)); 602 } 603 604 return false; 605 } 606 607 void AArch64PassConfig::addCodeGenPrepare() { 608 if (getOptLevel() != CodeGenOpt::None) 609 addPass(createTypePromotionPass()); 610 TargetPassConfig::addCodeGenPrepare(); 611 } 612 613 bool AArch64PassConfig::addInstSelector() { 614 addPass(createAArch64ISelDag(getAArch64TargetMachine(), getOptLevel())); 615 616 // For ELF, cleanup any local-dynamic TLS accesses (i.e. combine as many 617 // references to _TLS_MODULE_BASE_ as possible. 618 if (TM->getTargetTriple().isOSBinFormatELF() && 619 getOptLevel() != CodeGenOpt::None) 620 addPass(createAArch64CleanupLocalDynamicTLSPass()); 621 622 return false; 623 } 624 625 bool AArch64PassConfig::addIRTranslator() { 626 addPass(new IRTranslator(getOptLevel())); 627 return false; 628 } 629 630 void AArch64PassConfig::addPreLegalizeMachineIR() { 631 if (getOptLevel() == CodeGenOpt::None) 632 addPass(createAArch64O0PreLegalizerCombiner()); 633 else 634 addPass(createAArch64PreLegalizerCombiner()); 635 } 636 637 bool AArch64PassConfig::addLegalizeMachineIR() { 638 addPass(new Legalizer()); 639 return false; 640 } 641 642 void AArch64PassConfig::addPreRegBankSelect() { 643 bool IsOptNone = getOptLevel() == CodeGenOpt::None; 644 if (!IsOptNone) 645 addPass(createAArch64PostLegalizerCombiner(IsOptNone)); 646 addPass(createAArch64PostLegalizerLowering()); 647 } 648 649 bool AArch64PassConfig::addRegBankSelect() { 650 addPass(new RegBankSelect()); 651 return false; 652 } 653 654 void AArch64PassConfig::addPreGlobalInstructionSelect() { 655 addPass(new Localizer()); 656 } 657 658 bool AArch64PassConfig::addGlobalInstructionSelect() { 659 addPass(new InstructionSelect(getOptLevel())); 660 if (getOptLevel() != CodeGenOpt::None) 661 addPass(createAArch64PostSelectOptimize()); 662 return false; 663 } 664 665 void AArch64PassConfig::addMachineSSAOptimization() { 666 // Run default MachineSSAOptimization first. 667 TargetPassConfig::addMachineSSAOptimization(); 668 669 if (TM->getOptLevel() != CodeGenOpt::None) 670 addPass(createAArch64MIPeepholeOptPass()); 671 } 672 673 bool AArch64PassConfig::addILPOpts() { 674 if (EnableCondOpt) 675 addPass(createAArch64ConditionOptimizerPass()); 676 if (EnableCCMP) 677 addPass(createAArch64ConditionalCompares()); 678 if (EnableMCR) 679 addPass(&MachineCombinerID); 680 if (EnableCondBrTuning) 681 addPass(createAArch64CondBrTuning()); 682 if (EnableEarlyIfConversion) 683 addPass(&EarlyIfConverterID); 684 if (EnableStPairSuppress) 685 addPass(createAArch64StorePairSuppressPass()); 686 addPass(createAArch64SIMDInstrOptPass()); 687 if (TM->getOptLevel() != CodeGenOpt::None) 688 addPass(createAArch64StackTaggingPreRAPass()); 689 return true; 690 } 691 692 void AArch64PassConfig::addPreRegAlloc() { 693 // Change dead register definitions to refer to the zero register. 694 if (TM->getOptLevel() != CodeGenOpt::None && EnableDeadRegisterElimination) 695 addPass(createAArch64DeadRegisterDefinitions()); 696 697 // Use AdvSIMD scalar instructions whenever profitable. 698 if (TM->getOptLevel() != CodeGenOpt::None && EnableAdvSIMDScalar) { 699 addPass(createAArch64AdvSIMDScalar()); 700 // The AdvSIMD pass may produce copies that can be rewritten to 701 // be register coalescer friendly. 702 addPass(&PeepholeOptimizerID); 703 } 704 } 705 706 void AArch64PassConfig::addPostRegAlloc() { 707 // Remove redundant copy instructions. 708 if (TM->getOptLevel() != CodeGenOpt::None && EnableRedundantCopyElimination) 709 addPass(createAArch64RedundantCopyEliminationPass()); 710 711 if (TM->getOptLevel() != CodeGenOpt::None && usingDefaultRegAlloc()) 712 // Improve performance for some FP/SIMD code for A57. 713 addPass(createAArch64A57FPLoadBalancing()); 714 } 715 716 void AArch64PassConfig::addPreSched2() { 717 // Lower homogeneous frame instructions 718 if (EnableHomogeneousPrologEpilog) 719 addPass(createAArch64LowerHomogeneousPrologEpilogPass()); 720 // Expand some pseudo instructions to allow proper scheduling. 721 addPass(createAArch64ExpandPseudoPass()); 722 // Use load/store pair instructions when possible. 723 if (TM->getOptLevel() != CodeGenOpt::None) { 724 if (EnableLoadStoreOpt) 725 addPass(createAArch64LoadStoreOptimizationPass()); 726 } 727 728 // The AArch64SpeculationHardeningPass destroys dominator tree and natural 729 // loop info, which is needed for the FalkorHWPFFixPass and also later on. 730 // Therefore, run the AArch64SpeculationHardeningPass before the 731 // FalkorHWPFFixPass to avoid recomputing dominator tree and natural loop 732 // info. 733 addPass(createAArch64SpeculationHardeningPass()); 734 735 addPass(createAArch64IndirectThunks()); 736 addPass(createAArch64SLSHardeningPass()); 737 738 if (TM->getOptLevel() != CodeGenOpt::None) { 739 if (EnableFalkorHWPFFix) 740 addPass(createFalkorHWPFFixPass()); 741 } 742 } 743 744 void AArch64PassConfig::addPreEmitPass() { 745 // Machine Block Placement might have created new opportunities when run 746 // at O3, where the Tail Duplication Threshold is set to 4 instructions. 747 // Run the load/store optimizer once more. 748 if (TM->getOptLevel() >= CodeGenOpt::Aggressive && EnableLoadStoreOpt) 749 addPass(createAArch64LoadStoreOptimizationPass()); 750 751 if (EnableA53Fix835769) 752 addPass(createAArch64A53Fix835769()); 753 754 if (EnableBranchTargets) 755 addPass(createAArch64BranchTargetsPass()); 756 757 // Relax conditional branch instructions if they're otherwise out of 758 // range of their destination. 759 if (BranchRelaxation) 760 addPass(&BranchRelaxationPassID); 761 762 if (TM->getTargetTriple().isOSWindows()) { 763 // Identify valid longjmp targets for Windows Control Flow Guard. 764 addPass(createCFGuardLongjmpPass()); 765 // Identify valid eh continuation targets for Windows EHCont Guard. 766 addPass(createEHContGuardCatchretPass()); 767 } 768 769 if (TM->getOptLevel() != CodeGenOpt::None && EnableCompressJumpTables) 770 addPass(createAArch64CompressJumpTablesPass()); 771 772 if (TM->getOptLevel() != CodeGenOpt::None && EnableCollectLOH && 773 TM->getTargetTriple().isOSBinFormatMachO()) 774 addPass(createAArch64CollectLOHPass()); 775 } 776 777 void AArch64PassConfig::addPreEmitPass2() { 778 // SVE bundles move prefixes with destructive operations. BLR_RVMARKER pseudo 779 // instructions are lowered to bundles as well. 780 addPass(createUnpackMachineBundles(nullptr)); 781 } 782 783 yaml::MachineFunctionInfo * 784 AArch64TargetMachine::createDefaultFuncInfoYAML() const { 785 return new yaml::AArch64FunctionInfo(); 786 } 787 788 yaml::MachineFunctionInfo * 789 AArch64TargetMachine::convertFuncInfoToYAML(const MachineFunction &MF) const { 790 const auto *MFI = MF.getInfo<AArch64FunctionInfo>(); 791 return new yaml::AArch64FunctionInfo(*MFI); 792 } 793 794 bool AArch64TargetMachine::parseMachineFunctionInfo( 795 const yaml::MachineFunctionInfo &MFI, PerFunctionMIParsingState &PFS, 796 SMDiagnostic &Error, SMRange &SourceRange) const { 797 const auto &YamlMFI = 798 reinterpret_cast<const yaml::AArch64FunctionInfo &>(MFI); 799 MachineFunction &MF = PFS.MF; 800 MF.getInfo<AArch64FunctionInfo>()->initializeBaseYamlFields(YamlMFI); 801 return false; 802 } 803