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 "AArch64MacroFusion.h" 15 #include "AArch64Subtarget.h" 16 #include "AArch64TargetObjectFile.h" 17 #include "AArch64TargetTransformInfo.h" 18 #include "MCTargetDesc/AArch64MCTargetDesc.h" 19 #include "TargetInfo/AArch64TargetInfo.h" 20 #include "llvm/ADT/STLExtras.h" 21 #include "llvm/ADT/Triple.h" 22 #include "llvm/Analysis/TargetTransformInfo.h" 23 #include "llvm/CodeGen/CSEConfigBase.h" 24 #include "llvm/CodeGen/GlobalISel/IRTranslator.h" 25 #include "llvm/CodeGen/GlobalISel/InstructionSelect.h" 26 #include "llvm/CodeGen/GlobalISel/Legalizer.h" 27 #include "llvm/CodeGen/GlobalISel/Localizer.h" 28 #include "llvm/CodeGen/GlobalISel/RegBankSelect.h" 29 #include "llvm/CodeGen/MachineScheduler.h" 30 #include "llvm/CodeGen/Passes.h" 31 #include "llvm/CodeGen/TargetPassConfig.h" 32 #include "llvm/IR/Attributes.h" 33 #include "llvm/IR/Function.h" 34 #include "llvm/MC/MCAsmInfo.h" 35 #include "llvm/MC/MCTargetOptions.h" 36 #include "llvm/Pass.h" 37 #include "llvm/Support/CodeGen.h" 38 #include "llvm/Support/CommandLine.h" 39 #include "llvm/Support/TargetRegistry.h" 40 #include "llvm/Target/TargetLoweringObjectFile.h" 41 #include "llvm/Target/TargetOptions.h" 42 #include "llvm/Transforms/CFGuard.h" 43 #include "llvm/Transforms/Scalar.h" 44 #include <memory> 45 #include <string> 46 47 using namespace llvm; 48 49 static cl::opt<bool> EnableCCMP("aarch64-enable-ccmp", 50 cl::desc("Enable the CCMP formation pass"), 51 cl::init(true), cl::Hidden); 52 53 static cl::opt<bool> 54 EnableCondBrTuning("aarch64-enable-cond-br-tune", 55 cl::desc("Enable the conditional branch tuning pass"), 56 cl::init(true), cl::Hidden); 57 58 static cl::opt<bool> EnableMCR("aarch64-enable-mcr", 59 cl::desc("Enable the machine combiner pass"), 60 cl::init(true), cl::Hidden); 61 62 static cl::opt<bool> EnableStPairSuppress("aarch64-enable-stp-suppress", 63 cl::desc("Suppress STP for AArch64"), 64 cl::init(true), cl::Hidden); 65 66 static cl::opt<bool> EnableAdvSIMDScalar( 67 "aarch64-enable-simd-scalar", 68 cl::desc("Enable use of AdvSIMD scalar integer instructions"), 69 cl::init(false), cl::Hidden); 70 71 static cl::opt<bool> 72 EnablePromoteConstant("aarch64-enable-promote-const", 73 cl::desc("Enable the promote constant pass"), 74 cl::init(true), cl::Hidden); 75 76 static cl::opt<bool> EnableCollectLOH( 77 "aarch64-enable-collect-loh", 78 cl::desc("Enable the pass that emits the linker optimization hints (LOH)"), 79 cl::init(true), cl::Hidden); 80 81 static cl::opt<bool> 82 EnableDeadRegisterElimination("aarch64-enable-dead-defs", cl::Hidden, 83 cl::desc("Enable the pass that removes dead" 84 " definitons and replaces stores to" 85 " them with stores to the zero" 86 " register"), 87 cl::init(true)); 88 89 static cl::opt<bool> EnableRedundantCopyElimination( 90 "aarch64-enable-copyelim", 91 cl::desc("Enable the redundant copy elimination pass"), cl::init(true), 92 cl::Hidden); 93 94 static cl::opt<bool> EnableLoadStoreOpt("aarch64-enable-ldst-opt", 95 cl::desc("Enable the load/store pair" 96 " optimization pass"), 97 cl::init(true), cl::Hidden); 98 99 static cl::opt<bool> EnableAtomicTidy( 100 "aarch64-enable-atomic-cfg-tidy", cl::Hidden, 101 cl::desc("Run SimplifyCFG after expanding atomic operations" 102 " to make use of cmpxchg flow-based information"), 103 cl::init(true)); 104 105 static cl::opt<bool> 106 EnableEarlyIfConversion("aarch64-enable-early-ifcvt", cl::Hidden, 107 cl::desc("Run early if-conversion"), 108 cl::init(true)); 109 110 static cl::opt<bool> 111 EnableCondOpt("aarch64-enable-condopt", 112 cl::desc("Enable the condition optimizer pass"), 113 cl::init(true), cl::Hidden); 114 115 static cl::opt<bool> 116 EnableA53Fix835769("aarch64-fix-cortex-a53-835769", cl::Hidden, 117 cl::desc("Work around Cortex-A53 erratum 835769"), 118 cl::init(false)); 119 120 static cl::opt<bool> 121 EnableGEPOpt("aarch64-enable-gep-opt", cl::Hidden, 122 cl::desc("Enable optimizations on complex GEPs"), 123 cl::init(false)); 124 125 static cl::opt<bool> 126 BranchRelaxation("aarch64-enable-branch-relax", cl::Hidden, cl::init(true), 127 cl::desc("Relax out of range conditional branches")); 128 129 static cl::opt<bool> EnableCompressJumpTables( 130 "aarch64-enable-compress-jump-tables", cl::Hidden, cl::init(true), 131 cl::desc("Use smallest entry possible for jump tables")); 132 133 // FIXME: Unify control over GlobalMerge. 134 static cl::opt<cl::boolOrDefault> 135 EnableGlobalMerge("aarch64-enable-global-merge", cl::Hidden, 136 cl::desc("Enable the global merge pass")); 137 138 static cl::opt<bool> 139 EnableLoopDataPrefetch("aarch64-enable-loop-data-prefetch", cl::Hidden, 140 cl::desc("Enable the loop data prefetch pass"), 141 cl::init(true)); 142 143 static cl::opt<int> EnableGlobalISelAtO( 144 "aarch64-enable-global-isel-at-O", cl::Hidden, 145 cl::desc("Enable GlobalISel at or below an opt level (-1 to disable)"), 146 cl::init(0)); 147 148 static cl::opt<bool> EnableFalkorHWPFFix("aarch64-enable-falkor-hwpf-fix", 149 cl::init(true), cl::Hidden); 150 151 static cl::opt<bool> 152 EnableBranchTargets("aarch64-enable-branch-targets", cl::Hidden, 153 cl::desc("Enable the AAcrh64 branch target pass"), 154 cl::init(true)); 155 156 extern "C" void LLVMInitializeAArch64Target() { 157 // Register the target. 158 RegisterTargetMachine<AArch64leTargetMachine> X(getTheAArch64leTarget()); 159 RegisterTargetMachine<AArch64beTargetMachine> Y(getTheAArch64beTarget()); 160 RegisterTargetMachine<AArch64leTargetMachine> Z(getTheARM64Target()); 161 RegisterTargetMachine<AArch64leTargetMachine> W(getTheARM64_32Target()); 162 RegisterTargetMachine<AArch64leTargetMachine> V(getTheAArch64_32Target()); 163 auto PR = PassRegistry::getPassRegistry(); 164 initializeGlobalISel(*PR); 165 initializeAArch64A53Fix835769Pass(*PR); 166 initializeAArch64A57FPLoadBalancingPass(*PR); 167 initializeAArch64AdvSIMDScalarPass(*PR); 168 initializeAArch64BranchTargetsPass(*PR); 169 initializeAArch64CollectLOHPass(*PR); 170 initializeAArch64CompressJumpTablesPass(*PR); 171 initializeAArch64ConditionalComparesPass(*PR); 172 initializeAArch64ConditionOptimizerPass(*PR); 173 initializeAArch64DeadRegisterDefinitionsPass(*PR); 174 initializeAArch64ExpandPseudoPass(*PR); 175 initializeAArch64LoadStoreOptPass(*PR); 176 initializeAArch64SIMDInstrOptPass(*PR); 177 initializeAArch64PreLegalizerCombinerPass(*PR); 178 initializeAArch64PromoteConstantPass(*PR); 179 initializeAArch64RedundantCopyEliminationPass(*PR); 180 initializeAArch64StorePairSuppressPass(*PR); 181 initializeFalkorHWPFFixPass(*PR); 182 initializeFalkorMarkStridedAccessesLegacyPass(*PR); 183 initializeLDTLSCleanupPass(*PR); 184 initializeAArch64SpeculationHardeningPass(*PR); 185 initializeAArch64StackTaggingPass(*PR); 186 initializeAArch64StackTaggingPreRAPass(*PR); 187 } 188 189 //===----------------------------------------------------------------------===// 190 // AArch64 Lowering public interface. 191 //===----------------------------------------------------------------------===// 192 static std::unique_ptr<TargetLoweringObjectFile> createTLOF(const Triple &TT) { 193 if (TT.isOSBinFormatMachO()) 194 return std::make_unique<AArch64_MachoTargetObjectFile>(); 195 if (TT.isOSBinFormatCOFF()) 196 return std::make_unique<AArch64_COFFTargetObjectFile>(); 197 198 return std::make_unique<AArch64_ELFTargetObjectFile>(); 199 } 200 201 // Helper function to build a DataLayout string 202 static std::string computeDataLayout(const Triple &TT, 203 const MCTargetOptions &Options, 204 bool LittleEndian) { 205 if (Options.getABIName() == "ilp32") 206 return "e-m:e-p:32:32-i8:8-i16:16-i64:64-S128"; 207 if (TT.isOSBinFormatMachO()) { 208 if (TT.getArch() == Triple::aarch64_32) 209 return "e-m:o-p:32:32-i64:64-i128:128-n32:64-S128"; 210 return "e-m:o-i64:64-i128:128-n32:64-S128"; 211 } 212 if (TT.isOSBinFormatCOFF()) 213 return "e-m:w-p:64:64-i32:32-i64:64-i128:128-n32:64-S128"; 214 if (LittleEndian) 215 return "e-m:e-i8:8:32-i16:16:32-i64:64-i128:128-n32:64-S128"; 216 return "E-m:e-i8:8:32-i16:16:32-i64:64-i128:128-n32:64-S128"; 217 } 218 219 static Reloc::Model getEffectiveRelocModel(const Triple &TT, 220 Optional<Reloc::Model> RM) { 221 // AArch64 Darwin and Windows are always PIC. 222 if (TT.isOSDarwin() || TT.isOSWindows()) 223 return Reloc::PIC_; 224 // On ELF platforms the default static relocation model has a smart enough 225 // linker to cope with referencing external symbols defined in a shared 226 // library. Hence DynamicNoPIC doesn't need to be promoted to PIC. 227 if (!RM.hasValue() || *RM == Reloc::DynamicNoPIC) 228 return Reloc::Static; 229 return *RM; 230 } 231 232 static CodeModel::Model 233 getEffectiveAArch64CodeModel(const Triple &TT, Optional<CodeModel::Model> CM, 234 bool JIT) { 235 if (CM) { 236 if (*CM != CodeModel::Small && *CM != CodeModel::Tiny && 237 *CM != CodeModel::Large) { 238 if (!TT.isOSFuchsia()) 239 report_fatal_error( 240 "Only small, tiny and large code models are allowed on AArch64"); 241 else if (*CM != CodeModel::Kernel) 242 report_fatal_error("Only small, tiny, kernel, and large code models " 243 "are allowed on AArch64"); 244 } else if (*CM == CodeModel::Tiny && !TT.isOSBinFormatELF()) 245 report_fatal_error("tiny code model is only supported on ELF"); 246 return *CM; 247 } 248 // The default MCJIT memory managers make no guarantees about where they can 249 // find an executable page; JITed code needs to be able to refer to globals 250 // no matter how far away they are. 251 if (JIT) 252 return CodeModel::Large; 253 return CodeModel::Small; 254 } 255 256 /// Create an AArch64 architecture model. 257 /// 258 AArch64TargetMachine::AArch64TargetMachine(const Target &T, const Triple &TT, 259 StringRef CPU, StringRef FS, 260 const TargetOptions &Options, 261 Optional<Reloc::Model> RM, 262 Optional<CodeModel::Model> CM, 263 CodeGenOpt::Level OL, bool JIT, 264 bool LittleEndian) 265 : LLVMTargetMachine(T, 266 computeDataLayout(TT, Options.MCOptions, LittleEndian), 267 TT, CPU, FS, Options, getEffectiveRelocModel(TT, RM), 268 getEffectiveAArch64CodeModel(TT, CM, JIT), OL), 269 TLOF(createTLOF(getTargetTriple())), isLittle(LittleEndian) { 270 initAsmInfo(); 271 272 if (TT.isOSBinFormatMachO()) { 273 this->Options.TrapUnreachable = true; 274 this->Options.NoTrapAfterNoreturn = true; 275 } 276 277 if (getMCAsmInfo()->usesWindowsCFI()) { 278 // Unwinding can get confused if the last instruction in an 279 // exception-handling region (function, funclet, try block, etc.) 280 // is a call. 281 // 282 // FIXME: We could elide the trap if the next instruction would be in 283 // the same region anyway. 284 this->Options.TrapUnreachable = true; 285 } 286 287 // Enable GlobalISel at or below EnableGlobalISelAt0, unless this is 288 // MachO/CodeModel::Large, which GlobalISel does not support. 289 if (getOptLevel() <= EnableGlobalISelAtO && 290 TT.getArch() != Triple::aarch64_32 && 291 !(getCodeModel() == CodeModel::Large && TT.isOSBinFormatMachO())) { 292 setGlobalISel(true); 293 setGlobalISelAbort(GlobalISelAbortMode::Disable); 294 } 295 296 // AArch64 supports the MachineOutliner. 297 setMachineOutliner(true); 298 299 // AArch64 supports default outlining behaviour. 300 setSupportsDefaultOutlining(true); 301 } 302 303 AArch64TargetMachine::~AArch64TargetMachine() = default; 304 305 const AArch64Subtarget * 306 AArch64TargetMachine::getSubtargetImpl(const Function &F) const { 307 Attribute CPUAttr = F.getFnAttribute("target-cpu"); 308 Attribute FSAttr = F.getFnAttribute("target-features"); 309 310 std::string CPU = !CPUAttr.hasAttribute(Attribute::None) 311 ? CPUAttr.getValueAsString().str() 312 : TargetCPU; 313 std::string FS = !FSAttr.hasAttribute(Attribute::None) 314 ? FSAttr.getValueAsString().str() 315 : TargetFS; 316 317 auto &I = SubtargetMap[CPU + FS]; 318 if (!I) { 319 // This needs to be done before we create a new subtarget since any 320 // creation will depend on the TM and the code generation flags on the 321 // function that reside in TargetOptions. 322 resetTargetOptions(F); 323 I = std::make_unique<AArch64Subtarget>(TargetTriple, CPU, FS, *this, 324 isLittle); 325 } 326 return I.get(); 327 } 328 329 void AArch64leTargetMachine::anchor() { } 330 331 AArch64leTargetMachine::AArch64leTargetMachine( 332 const Target &T, const Triple &TT, StringRef CPU, StringRef FS, 333 const TargetOptions &Options, Optional<Reloc::Model> RM, 334 Optional<CodeModel::Model> CM, CodeGenOpt::Level OL, bool JIT) 335 : AArch64TargetMachine(T, TT, CPU, FS, Options, RM, CM, OL, JIT, true) {} 336 337 void AArch64beTargetMachine::anchor() { } 338 339 AArch64beTargetMachine::AArch64beTargetMachine( 340 const Target &T, const Triple &TT, StringRef CPU, StringRef FS, 341 const TargetOptions &Options, Optional<Reloc::Model> RM, 342 Optional<CodeModel::Model> CM, CodeGenOpt::Level OL, bool JIT) 343 : AArch64TargetMachine(T, TT, CPU, FS, Options, RM, CM, OL, JIT, false) {} 344 345 namespace { 346 347 /// AArch64 Code Generator Pass Configuration Options. 348 class AArch64PassConfig : public TargetPassConfig { 349 public: 350 AArch64PassConfig(AArch64TargetMachine &TM, PassManagerBase &PM) 351 : TargetPassConfig(TM, PM) { 352 if (TM.getOptLevel() != CodeGenOpt::None) 353 substitutePass(&PostRASchedulerID, &PostMachineSchedulerID); 354 } 355 356 AArch64TargetMachine &getAArch64TargetMachine() const { 357 return getTM<AArch64TargetMachine>(); 358 } 359 360 ScheduleDAGInstrs * 361 createMachineScheduler(MachineSchedContext *C) const override { 362 const AArch64Subtarget &ST = C->MF->getSubtarget<AArch64Subtarget>(); 363 ScheduleDAGMILive *DAG = createGenericSchedLive(C); 364 DAG->addMutation(createLoadClusterDAGMutation(DAG->TII, DAG->TRI)); 365 DAG->addMutation(createStoreClusterDAGMutation(DAG->TII, DAG->TRI)); 366 if (ST.hasFusion()) 367 DAG->addMutation(createAArch64MacroFusionDAGMutation()); 368 return DAG; 369 } 370 371 ScheduleDAGInstrs * 372 createPostMachineScheduler(MachineSchedContext *C) const override { 373 const AArch64Subtarget &ST = C->MF->getSubtarget<AArch64Subtarget>(); 374 if (ST.hasFusion()) { 375 // Run the Macro Fusion after RA again since literals are expanded from 376 // pseudos then (v. addPreSched2()). 377 ScheduleDAGMI *DAG = createGenericSchedPostRA(C); 378 DAG->addMutation(createAArch64MacroFusionDAGMutation()); 379 return DAG; 380 } 381 382 return nullptr; 383 } 384 385 void addIRPasses() override; 386 bool addPreISel() override; 387 bool addInstSelector() override; 388 bool addIRTranslator() override; 389 void addPreLegalizeMachineIR() override; 390 bool addLegalizeMachineIR() override; 391 bool addRegBankSelect() override; 392 void addPreGlobalInstructionSelect() override; 393 bool addGlobalInstructionSelect() override; 394 bool addILPOpts() override; 395 void addPreRegAlloc() override; 396 void addPostRegAlloc() override; 397 void addPreSched2() override; 398 void addPreEmitPass() override; 399 400 std::unique_ptr<CSEConfigBase> getCSEConfig() const override; 401 }; 402 403 } // end anonymous namespace 404 405 TargetTransformInfo 406 AArch64TargetMachine::getTargetTransformInfo(const Function &F) { 407 return TargetTransformInfo(AArch64TTIImpl(this, F)); 408 } 409 410 TargetPassConfig *AArch64TargetMachine::createPassConfig(PassManagerBase &PM) { 411 return new AArch64PassConfig(*this, PM); 412 } 413 414 std::unique_ptr<CSEConfigBase> AArch64PassConfig::getCSEConfig() const { 415 return getStandardCSEConfigForOpt(TM->getOptLevel()); 416 } 417 418 void AArch64PassConfig::addIRPasses() { 419 // Always expand atomic operations, we don't deal with atomicrmw or cmpxchg 420 // ourselves. 421 addPass(createAtomicExpandPass()); 422 423 // Cmpxchg instructions are often used with a subsequent comparison to 424 // determine whether it succeeded. We can exploit existing control-flow in 425 // ldrex/strex loops to simplify this, but it needs tidying up. 426 if (TM->getOptLevel() != CodeGenOpt::None && EnableAtomicTidy) 427 addPass(createCFGSimplificationPass(1, true, true, false, true)); 428 429 // Run LoopDataPrefetch 430 // 431 // Run this before LSR to remove the multiplies involved in computing the 432 // pointer values N iterations ahead. 433 if (TM->getOptLevel() != CodeGenOpt::None) { 434 if (EnableLoopDataPrefetch) 435 addPass(createLoopDataPrefetchPass()); 436 if (EnableFalkorHWPFFix) 437 addPass(createFalkorMarkStridedAccessesPass()); 438 } 439 440 TargetPassConfig::addIRPasses(); 441 442 // Match interleaved memory accesses to ldN/stN intrinsics. 443 if (TM->getOptLevel() != CodeGenOpt::None) { 444 addPass(createInterleavedLoadCombinePass()); 445 addPass(createInterleavedAccessPass()); 446 } 447 448 if (TM->getOptLevel() == CodeGenOpt::Aggressive && EnableGEPOpt) { 449 // Call SeparateConstOffsetFromGEP pass to extract constants within indices 450 // and lower a GEP with multiple indices to either arithmetic operations or 451 // multiple GEPs with single index. 452 addPass(createSeparateConstOffsetFromGEPPass(true)); 453 // Call EarlyCSE pass to find and remove subexpressions in the lowered 454 // result. 455 addPass(createEarlyCSEPass()); 456 // Do loop invariant code motion in case part of the lowered result is 457 // invariant. 458 addPass(createLICMPass()); 459 } 460 461 addPass(createAArch64StackTaggingPass(/* MergeInit = */ TM->getOptLevel() != 462 CodeGenOpt::None)); 463 464 // Add Control Flow Guard checks. 465 if (TM->getTargetTriple().isOSWindows()) 466 addPass(createCFGuardCheckPass()); 467 } 468 469 // Pass Pipeline Configuration 470 bool AArch64PassConfig::addPreISel() { 471 // Run promote constant before global merge, so that the promoted constants 472 // get a chance to be merged 473 if (TM->getOptLevel() != CodeGenOpt::None && EnablePromoteConstant) 474 addPass(createAArch64PromoteConstantPass()); 475 // FIXME: On AArch64, this depends on the type. 476 // Basically, the addressable offsets are up to 4095 * Ty.getSizeInBytes(). 477 // and the offset has to be a multiple of the related size in bytes. 478 if ((TM->getOptLevel() != CodeGenOpt::None && 479 EnableGlobalMerge == cl::BOU_UNSET) || 480 EnableGlobalMerge == cl::BOU_TRUE) { 481 bool OnlyOptimizeForSize = (TM->getOptLevel() < CodeGenOpt::Aggressive) && 482 (EnableGlobalMerge == cl::BOU_UNSET); 483 484 // Merging of extern globals is enabled by default on non-Mach-O as we 485 // expect it to be generally either beneficial or harmless. On Mach-O it 486 // is disabled as we emit the .subsections_via_symbols directive which 487 // means that merging extern globals is not safe. 488 bool MergeExternalByDefault = !TM->getTargetTriple().isOSBinFormatMachO(); 489 490 // FIXME: extern global merging is only enabled when we optimise for size 491 // because there are some regressions with it also enabled for performance. 492 if (!OnlyOptimizeForSize) 493 MergeExternalByDefault = false; 494 495 addPass(createGlobalMergePass(TM, 4095, OnlyOptimizeForSize, 496 MergeExternalByDefault)); 497 } 498 499 return false; 500 } 501 502 bool AArch64PassConfig::addInstSelector() { 503 addPass(createAArch64ISelDag(getAArch64TargetMachine(), getOptLevel())); 504 505 // For ELF, cleanup any local-dynamic TLS accesses (i.e. combine as many 506 // references to _TLS_MODULE_BASE_ as possible. 507 if (TM->getTargetTriple().isOSBinFormatELF() && 508 getOptLevel() != CodeGenOpt::None) 509 addPass(createAArch64CleanupLocalDynamicTLSPass()); 510 511 return false; 512 } 513 514 bool AArch64PassConfig::addIRTranslator() { 515 addPass(new IRTranslator()); 516 return false; 517 } 518 519 void AArch64PassConfig::addPreLegalizeMachineIR() { 520 bool IsOptNone = getOptLevel() == CodeGenOpt::None; 521 addPass(createAArch64PreLegalizeCombiner(IsOptNone)); 522 } 523 524 bool AArch64PassConfig::addLegalizeMachineIR() { 525 addPass(new Legalizer()); 526 return false; 527 } 528 529 bool AArch64PassConfig::addRegBankSelect() { 530 addPass(new RegBankSelect()); 531 return false; 532 } 533 534 void AArch64PassConfig::addPreGlobalInstructionSelect() { 535 addPass(new Localizer()); 536 } 537 538 bool AArch64PassConfig::addGlobalInstructionSelect() { 539 addPass(new InstructionSelect()); 540 return false; 541 } 542 543 bool AArch64PassConfig::addILPOpts() { 544 if (EnableCondOpt) 545 addPass(createAArch64ConditionOptimizerPass()); 546 if (EnableCCMP) 547 addPass(createAArch64ConditionalCompares()); 548 if (EnableMCR) 549 addPass(&MachineCombinerID); 550 if (EnableCondBrTuning) 551 addPass(createAArch64CondBrTuning()); 552 if (EnableEarlyIfConversion) 553 addPass(&EarlyIfConverterID); 554 if (EnableStPairSuppress) 555 addPass(createAArch64StorePairSuppressPass()); 556 addPass(createAArch64SIMDInstrOptPass()); 557 if (TM->getOptLevel() != CodeGenOpt::None) 558 addPass(createAArch64StackTaggingPreRAPass()); 559 return true; 560 } 561 562 void AArch64PassConfig::addPreRegAlloc() { 563 // Change dead register definitions to refer to the zero register. 564 if (TM->getOptLevel() != CodeGenOpt::None && EnableDeadRegisterElimination) 565 addPass(createAArch64DeadRegisterDefinitions()); 566 567 // Use AdvSIMD scalar instructions whenever profitable. 568 if (TM->getOptLevel() != CodeGenOpt::None && EnableAdvSIMDScalar) { 569 addPass(createAArch64AdvSIMDScalar()); 570 // The AdvSIMD pass may produce copies that can be rewritten to 571 // be register coaleascer friendly. 572 addPass(&PeepholeOptimizerID); 573 } 574 } 575 576 void AArch64PassConfig::addPostRegAlloc() { 577 // Remove redundant copy instructions. 578 if (TM->getOptLevel() != CodeGenOpt::None && EnableRedundantCopyElimination) 579 addPass(createAArch64RedundantCopyEliminationPass()); 580 581 if (TM->getOptLevel() != CodeGenOpt::None && usingDefaultRegAlloc()) 582 // Improve performance for some FP/SIMD code for A57. 583 addPass(createAArch64A57FPLoadBalancing()); 584 } 585 586 void AArch64PassConfig::addPreSched2() { 587 // Expand some pseudo instructions to allow proper scheduling. 588 addPass(createAArch64ExpandPseudoPass()); 589 // Use load/store pair instructions when possible. 590 if (TM->getOptLevel() != CodeGenOpt::None) { 591 if (EnableLoadStoreOpt) 592 addPass(createAArch64LoadStoreOptimizationPass()); 593 } 594 595 // The AArch64SpeculationHardeningPass destroys dominator tree and natural 596 // loop info, which is needed for the FalkorHWPFFixPass and also later on. 597 // Therefore, run the AArch64SpeculationHardeningPass before the 598 // FalkorHWPFFixPass to avoid recomputing dominator tree and natural loop 599 // info. 600 addPass(createAArch64SpeculationHardeningPass()); 601 602 if (TM->getOptLevel() != CodeGenOpt::None) { 603 if (EnableFalkorHWPFFix) 604 addPass(createFalkorHWPFFixPass()); 605 } 606 } 607 608 void AArch64PassConfig::addPreEmitPass() { 609 // Machine Block Placement might have created new opportunities when run 610 // at O3, where the Tail Duplication Threshold is set to 4 instructions. 611 // Run the load/store optimizer once more. 612 if (TM->getOptLevel() >= CodeGenOpt::Aggressive && EnableLoadStoreOpt) 613 addPass(createAArch64LoadStoreOptimizationPass()); 614 615 if (EnableA53Fix835769) 616 addPass(createAArch64A53Fix835769()); 617 // Relax conditional branch instructions if they're otherwise out of 618 // range of their destination. 619 if (BranchRelaxation) 620 addPass(&BranchRelaxationPassID); 621 622 if (EnableBranchTargets) 623 addPass(createAArch64BranchTargetsPass()); 624 625 // Identify valid longjmp targets for Windows Control Flow Guard. 626 if (TM->getTargetTriple().isOSWindows()) 627 addPass(createCFGuardLongjmpPass()); 628 629 if (TM->getOptLevel() != CodeGenOpt::None && EnableCompressJumpTables) 630 addPass(createAArch64CompressJumpTablesPass()); 631 632 if (TM->getOptLevel() != CodeGenOpt::None && EnableCollectLOH && 633 TM->getTargetTriple().isOSBinFormatMachO()) 634 addPass(createAArch64CollectLOHPass()); 635 } 636