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/InitializePasses.h" 35 #include "llvm/MC/MCAsmInfo.h" 36 #include "llvm/MC/MCTargetOptions.h" 37 #include "llvm/Pass.h" 38 #include "llvm/Support/CodeGen.h" 39 #include "llvm/Support/CommandLine.h" 40 #include "llvm/Support/TargetRegistry.h" 41 #include "llvm/Target/TargetLoweringObjectFile.h" 42 #include "llvm/Target/TargetOptions.h" 43 #include "llvm/Transforms/CFGuard.h" 44 #include "llvm/Transforms/Scalar.h" 45 #include <memory> 46 #include <string> 47 48 using namespace llvm; 49 50 static cl::opt<bool> EnableCCMP("aarch64-enable-ccmp", 51 cl::desc("Enable the CCMP formation pass"), 52 cl::init(true), cl::Hidden); 53 54 static cl::opt<bool> 55 EnableCondBrTuning("aarch64-enable-cond-br-tune", 56 cl::desc("Enable the conditional branch tuning pass"), 57 cl::init(true), cl::Hidden); 58 59 static cl::opt<bool> EnableMCR("aarch64-enable-mcr", 60 cl::desc("Enable the machine combiner pass"), 61 cl::init(true), cl::Hidden); 62 63 static cl::opt<bool> EnableStPairSuppress("aarch64-enable-stp-suppress", 64 cl::desc("Suppress STP for AArch64"), 65 cl::init(true), cl::Hidden); 66 67 static cl::opt<bool> EnableAdvSIMDScalar( 68 "aarch64-enable-simd-scalar", 69 cl::desc("Enable use of AdvSIMD scalar integer instructions"), 70 cl::init(false), cl::Hidden); 71 72 static cl::opt<bool> 73 EnablePromoteConstant("aarch64-enable-promote-const", 74 cl::desc("Enable the promote constant pass"), 75 cl::init(true), cl::Hidden); 76 77 static cl::opt<bool> EnableCollectLOH( 78 "aarch64-enable-collect-loh", 79 cl::desc("Enable the pass that emits the linker optimization hints (LOH)"), 80 cl::init(true), cl::Hidden); 81 82 static cl::opt<bool> 83 EnableDeadRegisterElimination("aarch64-enable-dead-defs", cl::Hidden, 84 cl::desc("Enable the pass that removes dead" 85 " definitons and replaces stores to" 86 " them with stores to the zero" 87 " register"), 88 cl::init(true)); 89 90 static cl::opt<bool> EnableRedundantCopyElimination( 91 "aarch64-enable-copyelim", 92 cl::desc("Enable the redundant copy elimination pass"), cl::init(true), 93 cl::Hidden); 94 95 static cl::opt<bool> EnableLoadStoreOpt("aarch64-enable-ldst-opt", 96 cl::desc("Enable the load/store pair" 97 " optimization pass"), 98 cl::init(true), cl::Hidden); 99 100 static cl::opt<bool> EnableAtomicTidy( 101 "aarch64-enable-atomic-cfg-tidy", cl::Hidden, 102 cl::desc("Run SimplifyCFG after expanding atomic operations" 103 " to make use of cmpxchg flow-based information"), 104 cl::init(true)); 105 106 static cl::opt<bool> 107 EnableEarlyIfConversion("aarch64-enable-early-ifcvt", cl::Hidden, 108 cl::desc("Run early if-conversion"), 109 cl::init(true)); 110 111 static cl::opt<bool> 112 EnableCondOpt("aarch64-enable-condopt", 113 cl::desc("Enable the condition optimizer pass"), 114 cl::init(true), cl::Hidden); 115 116 static cl::opt<bool> 117 EnableA53Fix835769("aarch64-fix-cortex-a53-835769", cl::Hidden, 118 cl::desc("Work around Cortex-A53 erratum 835769"), 119 cl::init(false)); 120 121 static cl::opt<bool> 122 EnableGEPOpt("aarch64-enable-gep-opt", cl::Hidden, 123 cl::desc("Enable optimizations on complex GEPs"), 124 cl::init(false)); 125 126 static cl::opt<bool> 127 BranchRelaxation("aarch64-enable-branch-relax", cl::Hidden, cl::init(true), 128 cl::desc("Relax out of range conditional branches")); 129 130 static cl::opt<bool> EnableCompressJumpTables( 131 "aarch64-enable-compress-jump-tables", cl::Hidden, cl::init(true), 132 cl::desc("Use smallest entry possible for jump tables")); 133 134 // FIXME: Unify control over GlobalMerge. 135 static cl::opt<cl::boolOrDefault> 136 EnableGlobalMerge("aarch64-enable-global-merge", cl::Hidden, 137 cl::desc("Enable the global merge pass")); 138 139 static cl::opt<bool> 140 EnableLoopDataPrefetch("aarch64-enable-loop-data-prefetch", cl::Hidden, 141 cl::desc("Enable the loop data prefetch pass"), 142 cl::init(true)); 143 144 static cl::opt<int> EnableGlobalISelAtO( 145 "aarch64-enable-global-isel-at-O", cl::Hidden, 146 cl::desc("Enable GlobalISel at or below an opt level (-1 to disable)"), 147 cl::init(0)); 148 149 static cl::opt<bool> EnableFalkorHWPFFix("aarch64-enable-falkor-hwpf-fix", 150 cl::init(true), cl::Hidden); 151 152 static cl::opt<bool> 153 EnableBranchTargets("aarch64-enable-branch-targets", cl::Hidden, 154 cl::desc("Enable the AAcrh64 branch target pass"), 155 cl::init(true)); 156 157 extern "C" void LLVMInitializeAArch64Target() { 158 // Register the target. 159 RegisterTargetMachine<AArch64leTargetMachine> X(getTheAArch64leTarget()); 160 RegisterTargetMachine<AArch64beTargetMachine> Y(getTheAArch64beTarget()); 161 RegisterTargetMachine<AArch64leTargetMachine> Z(getTheARM64Target()); 162 RegisterTargetMachine<AArch64leTargetMachine> W(getTheARM64_32Target()); 163 RegisterTargetMachine<AArch64leTargetMachine> V(getTheAArch64_32Target()); 164 auto PR = PassRegistry::getPassRegistry(); 165 initializeGlobalISel(*PR); 166 initializeAArch64A53Fix835769Pass(*PR); 167 initializeAArch64A57FPLoadBalancingPass(*PR); 168 initializeAArch64AdvSIMDScalarPass(*PR); 169 initializeAArch64BranchTargetsPass(*PR); 170 initializeAArch64CollectLOHPass(*PR); 171 initializeAArch64CompressJumpTablesPass(*PR); 172 initializeAArch64ConditionalComparesPass(*PR); 173 initializeAArch64ConditionOptimizerPass(*PR); 174 initializeAArch64DeadRegisterDefinitionsPass(*PR); 175 initializeAArch64ExpandPseudoPass(*PR); 176 initializeAArch64LoadStoreOptPass(*PR); 177 initializeAArch64SIMDInstrOptPass(*PR); 178 initializeAArch64PreLegalizerCombinerPass(*PR); 179 initializeAArch64PromoteConstantPass(*PR); 180 initializeAArch64RedundantCopyEliminationPass(*PR); 181 initializeAArch64StorePairSuppressPass(*PR); 182 initializeFalkorHWPFFixPass(*PR); 183 initializeFalkorMarkStridedAccessesLegacyPass(*PR); 184 initializeLDTLSCleanupPass(*PR); 185 initializeAArch64SpeculationHardeningPass(*PR); 186 initializeAArch64StackTaggingPass(*PR); 187 initializeAArch64StackTaggingPreRAPass(*PR); 188 } 189 190 //===----------------------------------------------------------------------===// 191 // AArch64 Lowering public interface. 192 //===----------------------------------------------------------------------===// 193 static std::unique_ptr<TargetLoweringObjectFile> createTLOF(const Triple &TT) { 194 if (TT.isOSBinFormatMachO()) 195 return std::make_unique<AArch64_MachoTargetObjectFile>(); 196 if (TT.isOSBinFormatCOFF()) 197 return std::make_unique<AArch64_COFFTargetObjectFile>(); 198 199 return std::make_unique<AArch64_ELFTargetObjectFile>(); 200 } 201 202 // Helper function to build a DataLayout string 203 static std::string computeDataLayout(const Triple &TT, 204 const MCTargetOptions &Options, 205 bool LittleEndian) { 206 if (Options.getABIName() == "ilp32") 207 return "e-m:e-p:32:32-i8:8-i16:16-i64:64-S128"; 208 if (TT.isOSBinFormatMachO()) { 209 if (TT.getArch() == Triple::aarch64_32) 210 return "e-m:o-p:32:32-i64:64-i128:128-n32:64-S128"; 211 return "e-m:o-i64:64-i128:128-n32:64-S128"; 212 } 213 if (TT.isOSBinFormatCOFF()) 214 return "e-m:w-p:64:64-i32:32-i64:64-i128:128-n32:64-S128"; 215 if (LittleEndian) 216 return "e-m:e-i8:8:32-i16:16:32-i64:64-i128:128-n32:64-S128"; 217 return "E-m:e-i8:8:32-i16:16:32-i64:64-i128:128-n32:64-S128"; 218 } 219 220 static Reloc::Model getEffectiveRelocModel(const Triple &TT, 221 Optional<Reloc::Model> RM) { 222 // AArch64 Darwin and Windows are always PIC. 223 if (TT.isOSDarwin() || TT.isOSWindows()) 224 return Reloc::PIC_; 225 // On ELF platforms the default static relocation model has a smart enough 226 // linker to cope with referencing external symbols defined in a shared 227 // library. Hence DynamicNoPIC doesn't need to be promoted to PIC. 228 if (!RM.hasValue() || *RM == Reloc::DynamicNoPIC) 229 return Reloc::Static; 230 return *RM; 231 } 232 233 static CodeModel::Model 234 getEffectiveAArch64CodeModel(const Triple &TT, Optional<CodeModel::Model> CM, 235 bool JIT) { 236 if (CM) { 237 if (*CM != CodeModel::Small && *CM != CodeModel::Tiny && 238 *CM != CodeModel::Large) { 239 if (!TT.isOSFuchsia()) 240 report_fatal_error( 241 "Only small, tiny and large code models are allowed on AArch64"); 242 else if (*CM != CodeModel::Kernel) 243 report_fatal_error("Only small, tiny, kernel, and large code models " 244 "are allowed on AArch64"); 245 } else if (*CM == CodeModel::Tiny && !TT.isOSBinFormatELF()) 246 report_fatal_error("tiny code model is only supported on ELF"); 247 return *CM; 248 } 249 // The default MCJIT memory managers make no guarantees about where they can 250 // find an executable page; JITed code needs to be able to refer to globals 251 // no matter how far away they are. 252 if (JIT) 253 return CodeModel::Large; 254 return CodeModel::Small; 255 } 256 257 /// Create an AArch64 architecture model. 258 /// 259 AArch64TargetMachine::AArch64TargetMachine(const Target &T, const Triple &TT, 260 StringRef CPU, StringRef FS, 261 const TargetOptions &Options, 262 Optional<Reloc::Model> RM, 263 Optional<CodeModel::Model> CM, 264 CodeGenOpt::Level OL, bool JIT, 265 bool LittleEndian) 266 : LLVMTargetMachine(T, 267 computeDataLayout(TT, Options.MCOptions, LittleEndian), 268 TT, CPU, FS, Options, getEffectiveRelocModel(TT, RM), 269 getEffectiveAArch64CodeModel(TT, CM, JIT), OL), 270 TLOF(createTLOF(getTargetTriple())), isLittle(LittleEndian) { 271 initAsmInfo(); 272 273 if (TT.isOSBinFormatMachO()) { 274 this->Options.TrapUnreachable = true; 275 this->Options.NoTrapAfterNoreturn = true; 276 } 277 278 if (getMCAsmInfo()->usesWindowsCFI()) { 279 // Unwinding can get confused if the last instruction in an 280 // exception-handling region (function, funclet, try block, etc.) 281 // is a call. 282 // 283 // FIXME: We could elide the trap if the next instruction would be in 284 // the same region anyway. 285 this->Options.TrapUnreachable = true; 286 } 287 288 // Enable GlobalISel at or below EnableGlobalISelAt0, unless this is 289 // MachO/CodeModel::Large, which GlobalISel does not support. 290 if (getOptLevel() <= EnableGlobalISelAtO && 291 TT.getArch() != Triple::aarch64_32 && 292 !(getCodeModel() == CodeModel::Large && TT.isOSBinFormatMachO())) { 293 setGlobalISel(true); 294 setGlobalISelAbort(GlobalISelAbortMode::Disable); 295 } 296 297 // AArch64 supports the MachineOutliner. 298 setMachineOutliner(true); 299 300 // AArch64 supports default outlining behaviour. 301 setSupportsDefaultOutlining(true); 302 } 303 304 AArch64TargetMachine::~AArch64TargetMachine() = default; 305 306 const AArch64Subtarget * 307 AArch64TargetMachine::getSubtargetImpl(const Function &F) const { 308 Attribute CPUAttr = F.getFnAttribute("target-cpu"); 309 Attribute FSAttr = F.getFnAttribute("target-features"); 310 311 std::string CPU = !CPUAttr.hasAttribute(Attribute::None) 312 ? CPUAttr.getValueAsString().str() 313 : TargetCPU; 314 std::string FS = !FSAttr.hasAttribute(Attribute::None) 315 ? FSAttr.getValueAsString().str() 316 : TargetFS; 317 318 auto &I = SubtargetMap[CPU + FS]; 319 if (!I) { 320 // This needs to be done before we create a new subtarget since any 321 // creation will depend on the TM and the code generation flags on the 322 // function that reside in TargetOptions. 323 resetTargetOptions(F); 324 I = std::make_unique<AArch64Subtarget>(TargetTriple, CPU, FS, *this, 325 isLittle); 326 } 327 return I.get(); 328 } 329 330 void AArch64leTargetMachine::anchor() { } 331 332 AArch64leTargetMachine::AArch64leTargetMachine( 333 const Target &T, const Triple &TT, StringRef CPU, StringRef FS, 334 const TargetOptions &Options, Optional<Reloc::Model> RM, 335 Optional<CodeModel::Model> CM, CodeGenOpt::Level OL, bool JIT) 336 : AArch64TargetMachine(T, TT, CPU, FS, Options, RM, CM, OL, JIT, true) {} 337 338 void AArch64beTargetMachine::anchor() { } 339 340 AArch64beTargetMachine::AArch64beTargetMachine( 341 const Target &T, const Triple &TT, StringRef CPU, StringRef FS, 342 const TargetOptions &Options, Optional<Reloc::Model> RM, 343 Optional<CodeModel::Model> CM, CodeGenOpt::Level OL, bool JIT) 344 : AArch64TargetMachine(T, TT, CPU, FS, Options, RM, CM, OL, JIT, false) {} 345 346 namespace { 347 348 /// AArch64 Code Generator Pass Configuration Options. 349 class AArch64PassConfig : public TargetPassConfig { 350 public: 351 AArch64PassConfig(AArch64TargetMachine &TM, PassManagerBase &PM) 352 : TargetPassConfig(TM, PM) { 353 if (TM.getOptLevel() != CodeGenOpt::None) 354 substitutePass(&PostRASchedulerID, &PostMachineSchedulerID); 355 } 356 357 AArch64TargetMachine &getAArch64TargetMachine() const { 358 return getTM<AArch64TargetMachine>(); 359 } 360 361 ScheduleDAGInstrs * 362 createMachineScheduler(MachineSchedContext *C) const override { 363 const AArch64Subtarget &ST = C->MF->getSubtarget<AArch64Subtarget>(); 364 ScheduleDAGMILive *DAG = createGenericSchedLive(C); 365 DAG->addMutation(createLoadClusterDAGMutation(DAG->TII, DAG->TRI)); 366 DAG->addMutation(createStoreClusterDAGMutation(DAG->TII, DAG->TRI)); 367 if (ST.hasFusion()) 368 DAG->addMutation(createAArch64MacroFusionDAGMutation()); 369 return DAG; 370 } 371 372 ScheduleDAGInstrs * 373 createPostMachineScheduler(MachineSchedContext *C) const override { 374 const AArch64Subtarget &ST = C->MF->getSubtarget<AArch64Subtarget>(); 375 if (ST.hasFusion()) { 376 // Run the Macro Fusion after RA again since literals are expanded from 377 // pseudos then (v. addPreSched2()). 378 ScheduleDAGMI *DAG = createGenericSchedPostRA(C); 379 DAG->addMutation(createAArch64MacroFusionDAGMutation()); 380 return DAG; 381 } 382 383 return nullptr; 384 } 385 386 void addIRPasses() override; 387 bool addPreISel() override; 388 bool addInstSelector() override; 389 bool addIRTranslator() override; 390 void addPreLegalizeMachineIR() override; 391 bool addLegalizeMachineIR() override; 392 bool addRegBankSelect() override; 393 void addPreGlobalInstructionSelect() override; 394 bool addGlobalInstructionSelect() override; 395 bool addILPOpts() override; 396 void addPreRegAlloc() override; 397 void addPostRegAlloc() override; 398 void addPreSched2() override; 399 void addPreEmitPass() override; 400 401 std::unique_ptr<CSEConfigBase> getCSEConfig() const override; 402 }; 403 404 } // end anonymous namespace 405 406 TargetTransformInfo 407 AArch64TargetMachine::getTargetTransformInfo(const Function &F) { 408 return TargetTransformInfo(AArch64TTIImpl(this, F)); 409 } 410 411 TargetPassConfig *AArch64TargetMachine::createPassConfig(PassManagerBase &PM) { 412 return new AArch64PassConfig(*this, PM); 413 } 414 415 std::unique_ptr<CSEConfigBase> AArch64PassConfig::getCSEConfig() const { 416 return getStandardCSEConfigForOpt(TM->getOptLevel()); 417 } 418 419 void AArch64PassConfig::addIRPasses() { 420 // Always expand atomic operations, we don't deal with atomicrmw or cmpxchg 421 // ourselves. 422 addPass(createAtomicExpandPass()); 423 424 // Cmpxchg instructions are often used with a subsequent comparison to 425 // determine whether it succeeded. We can exploit existing control-flow in 426 // ldrex/strex loops to simplify this, but it needs tidying up. 427 if (TM->getOptLevel() != CodeGenOpt::None && EnableAtomicTidy) 428 addPass(createCFGSimplificationPass(1, true, true, false, true)); 429 430 // Run LoopDataPrefetch 431 // 432 // Run this before LSR to remove the multiplies involved in computing the 433 // pointer values N iterations ahead. 434 if (TM->getOptLevel() != CodeGenOpt::None) { 435 if (EnableLoopDataPrefetch) 436 addPass(createLoopDataPrefetchPass()); 437 if (EnableFalkorHWPFFix) 438 addPass(createFalkorMarkStridedAccessesPass()); 439 } 440 441 TargetPassConfig::addIRPasses(); 442 443 // Match interleaved memory accesses to ldN/stN intrinsics. 444 if (TM->getOptLevel() != CodeGenOpt::None) { 445 addPass(createInterleavedLoadCombinePass()); 446 addPass(createInterleavedAccessPass()); 447 } 448 449 if (TM->getOptLevel() == CodeGenOpt::Aggressive && EnableGEPOpt) { 450 // Call SeparateConstOffsetFromGEP pass to extract constants within indices 451 // and lower a GEP with multiple indices to either arithmetic operations or 452 // multiple GEPs with single index. 453 addPass(createSeparateConstOffsetFromGEPPass(true)); 454 // Call EarlyCSE pass to find and remove subexpressions in the lowered 455 // result. 456 addPass(createEarlyCSEPass()); 457 // Do loop invariant code motion in case part of the lowered result is 458 // invariant. 459 addPass(createLICMPass()); 460 } 461 462 addPass(createAArch64StackTaggingPass(/* MergeInit = */ TM->getOptLevel() != 463 CodeGenOpt::None)); 464 465 // Add Control Flow Guard checks. 466 if (TM->getTargetTriple().isOSWindows()) 467 addPass(createCFGuardCheckPass()); 468 } 469 470 // Pass Pipeline Configuration 471 bool AArch64PassConfig::addPreISel() { 472 // Run promote constant before global merge, so that the promoted constants 473 // get a chance to be merged 474 if (TM->getOptLevel() != CodeGenOpt::None && EnablePromoteConstant) 475 addPass(createAArch64PromoteConstantPass()); 476 // FIXME: On AArch64, this depends on the type. 477 // Basically, the addressable offsets are up to 4095 * Ty.getSizeInBytes(). 478 // and the offset has to be a multiple of the related size in bytes. 479 if ((TM->getOptLevel() != CodeGenOpt::None && 480 EnableGlobalMerge == cl::BOU_UNSET) || 481 EnableGlobalMerge == cl::BOU_TRUE) { 482 bool OnlyOptimizeForSize = (TM->getOptLevel() < CodeGenOpt::Aggressive) && 483 (EnableGlobalMerge == cl::BOU_UNSET); 484 485 // Merging of extern globals is enabled by default on non-Mach-O as we 486 // expect it to be generally either beneficial or harmless. On Mach-O it 487 // is disabled as we emit the .subsections_via_symbols directive which 488 // means that merging extern globals is not safe. 489 bool MergeExternalByDefault = !TM->getTargetTriple().isOSBinFormatMachO(); 490 491 // FIXME: extern global merging is only enabled when we optimise for size 492 // because there are some regressions with it also enabled for performance. 493 if (!OnlyOptimizeForSize) 494 MergeExternalByDefault = false; 495 496 addPass(createGlobalMergePass(TM, 4095, OnlyOptimizeForSize, 497 MergeExternalByDefault)); 498 } 499 500 return false; 501 } 502 503 bool AArch64PassConfig::addInstSelector() { 504 addPass(createAArch64ISelDag(getAArch64TargetMachine(), getOptLevel())); 505 506 // For ELF, cleanup any local-dynamic TLS accesses (i.e. combine as many 507 // references to _TLS_MODULE_BASE_ as possible. 508 if (TM->getTargetTriple().isOSBinFormatELF() && 509 getOptLevel() != CodeGenOpt::None) 510 addPass(createAArch64CleanupLocalDynamicTLSPass()); 511 512 return false; 513 } 514 515 bool AArch64PassConfig::addIRTranslator() { 516 addPass(new IRTranslator()); 517 return false; 518 } 519 520 void AArch64PassConfig::addPreLegalizeMachineIR() { 521 bool IsOptNone = getOptLevel() == CodeGenOpt::None; 522 addPass(createAArch64PreLegalizeCombiner(IsOptNone)); 523 } 524 525 bool AArch64PassConfig::addLegalizeMachineIR() { 526 addPass(new Legalizer()); 527 return false; 528 } 529 530 bool AArch64PassConfig::addRegBankSelect() { 531 addPass(new RegBankSelect()); 532 return false; 533 } 534 535 void AArch64PassConfig::addPreGlobalInstructionSelect() { 536 addPass(new Localizer()); 537 } 538 539 bool AArch64PassConfig::addGlobalInstructionSelect() { 540 addPass(new InstructionSelect()); 541 return false; 542 } 543 544 bool AArch64PassConfig::addILPOpts() { 545 if (EnableCondOpt) 546 addPass(createAArch64ConditionOptimizerPass()); 547 if (EnableCCMP) 548 addPass(createAArch64ConditionalCompares()); 549 if (EnableMCR) 550 addPass(&MachineCombinerID); 551 if (EnableCondBrTuning) 552 addPass(createAArch64CondBrTuning()); 553 if (EnableEarlyIfConversion) 554 addPass(&EarlyIfConverterID); 555 if (EnableStPairSuppress) 556 addPass(createAArch64StorePairSuppressPass()); 557 addPass(createAArch64SIMDInstrOptPass()); 558 if (TM->getOptLevel() != CodeGenOpt::None) 559 addPass(createAArch64StackTaggingPreRAPass()); 560 return true; 561 } 562 563 void AArch64PassConfig::addPreRegAlloc() { 564 // Change dead register definitions to refer to the zero register. 565 if (TM->getOptLevel() != CodeGenOpt::None && EnableDeadRegisterElimination) 566 addPass(createAArch64DeadRegisterDefinitions()); 567 568 // Use AdvSIMD scalar instructions whenever profitable. 569 if (TM->getOptLevel() != CodeGenOpt::None && EnableAdvSIMDScalar) { 570 addPass(createAArch64AdvSIMDScalar()); 571 // The AdvSIMD pass may produce copies that can be rewritten to 572 // be register coaleascer friendly. 573 addPass(&PeepholeOptimizerID); 574 } 575 } 576 577 void AArch64PassConfig::addPostRegAlloc() { 578 // Remove redundant copy instructions. 579 if (TM->getOptLevel() != CodeGenOpt::None && EnableRedundantCopyElimination) 580 addPass(createAArch64RedundantCopyEliminationPass()); 581 582 if (TM->getOptLevel() != CodeGenOpt::None && usingDefaultRegAlloc()) 583 // Improve performance for some FP/SIMD code for A57. 584 addPass(createAArch64A57FPLoadBalancing()); 585 } 586 587 void AArch64PassConfig::addPreSched2() { 588 // Expand some pseudo instructions to allow proper scheduling. 589 addPass(createAArch64ExpandPseudoPass()); 590 // Use load/store pair instructions when possible. 591 if (TM->getOptLevel() != CodeGenOpt::None) { 592 if (EnableLoadStoreOpt) 593 addPass(createAArch64LoadStoreOptimizationPass()); 594 } 595 596 // The AArch64SpeculationHardeningPass destroys dominator tree and natural 597 // loop info, which is needed for the FalkorHWPFFixPass and also later on. 598 // Therefore, run the AArch64SpeculationHardeningPass before the 599 // FalkorHWPFFixPass to avoid recomputing dominator tree and natural loop 600 // info. 601 addPass(createAArch64SpeculationHardeningPass()); 602 603 if (TM->getOptLevel() != CodeGenOpt::None) { 604 if (EnableFalkorHWPFFix) 605 addPass(createFalkorHWPFFixPass()); 606 } 607 } 608 609 void AArch64PassConfig::addPreEmitPass() { 610 // Machine Block Placement might have created new opportunities when run 611 // at O3, where the Tail Duplication Threshold is set to 4 instructions. 612 // Run the load/store optimizer once more. 613 if (TM->getOptLevel() >= CodeGenOpt::Aggressive && EnableLoadStoreOpt) 614 addPass(createAArch64LoadStoreOptimizationPass()); 615 616 if (EnableA53Fix835769) 617 addPass(createAArch64A53Fix835769()); 618 619 if (EnableBranchTargets) 620 addPass(createAArch64BranchTargetsPass()); 621 622 // Relax conditional branch instructions if they're otherwise out of 623 // range of their destination. 624 if (BranchRelaxation) 625 addPass(&BranchRelaxationPassID); 626 627 // Identify valid longjmp targets for Windows Control Flow Guard. 628 if (TM->getTargetTriple().isOSWindows()) 629 addPass(createCFGuardLongjmpPass()); 630 631 if (TM->getOptLevel() != CodeGenOpt::None && EnableCompressJumpTables) 632 addPass(createAArch64CompressJumpTablesPass()); 633 634 if (TM->getOptLevel() != CodeGenOpt::None && EnableCollectLOH && 635 TM->getTargetTriple().isOSBinFormatMachO()) 636 addPass(createAArch64CollectLOHPass()); 637 } 638