1 //===-- AArch64TargetMachine.cpp - Define TargetMachine for AArch64 -------===// 2 // 3 // The LLVM Compiler Infrastructure 4 // 5 // This file is distributed under the University of Illinois Open Source 6 // License. See LICENSE.TXT for details. 7 // 8 //===----------------------------------------------------------------------===// 9 // 10 // 11 //===----------------------------------------------------------------------===// 12 13 #include "AArch64.h" 14 #include "AArch64CallLowering.h" 15 #include "AArch64InstructionSelector.h" 16 #include "AArch64LegalizerInfo.h" 17 #ifdef LLVM_BUILD_GLOBAL_ISEL 18 #include "AArch64RegisterBankInfo.h" 19 #endif 20 #include "AArch64Subtarget.h" 21 #include "AArch64TargetMachine.h" 22 #include "AArch64TargetObjectFile.h" 23 #include "AArch64TargetTransformInfo.h" 24 #include "MCTargetDesc/AArch64MCTargetDesc.h" 25 #include "llvm/ADT/STLExtras.h" 26 #include "llvm/ADT/Triple.h" 27 #include "llvm/Analysis/TargetTransformInfo.h" 28 #include "llvm/CodeGen/GlobalISel/GISelAccessor.h" 29 #include "llvm/CodeGen/GlobalISel/IRTranslator.h" 30 #include "llvm/CodeGen/GlobalISel/InstructionSelect.h" 31 #include "llvm/CodeGen/GlobalISel/Legalizer.h" 32 #include "llvm/CodeGen/GlobalISel/RegBankSelect.h" 33 #include "llvm/CodeGen/MachineScheduler.h" 34 #include "llvm/CodeGen/Passes.h" 35 #include "llvm/CodeGen/TargetPassConfig.h" 36 #include "llvm/IR/Attributes.h" 37 #include "llvm/IR/Function.h" 38 #include "llvm/MC/MCTargetOptions.h" 39 #include "llvm/Pass.h" 40 #include "llvm/Support/CodeGen.h" 41 #include "llvm/Support/CommandLine.h" 42 #include "llvm/Support/TargetRegistry.h" 43 #include "llvm/Target/TargetLoweringObjectFile.h" 44 #include "llvm/Target/TargetOptions.h" 45 #include "llvm/Transforms/Scalar.h" 46 #include <memory> 47 #include <string> 48 49 using namespace llvm; 50 51 static cl::opt<bool> EnableCCMP("aarch64-enable-ccmp", 52 cl::desc("Enable the CCMP formation pass"), 53 cl::init(true), cl::Hidden); 54 55 static cl::opt<bool> EnableMCR("aarch64-enable-mcr", 56 cl::desc("Enable the machine combiner pass"), 57 cl::init(true), cl::Hidden); 58 59 static cl::opt<bool> EnableStPairSuppress("aarch64-enable-stp-suppress", 60 cl::desc("Suppress STP for AArch64"), 61 cl::init(true), cl::Hidden); 62 63 static cl::opt<bool> EnableAdvSIMDScalar( 64 "aarch64-enable-simd-scalar", 65 cl::desc("Enable use of AdvSIMD scalar integer instructions"), 66 cl::init(false), cl::Hidden); 67 68 static cl::opt<bool> 69 EnablePromoteConstant("aarch64-enable-promote-const", 70 cl::desc("Enable the promote constant pass"), 71 cl::init(true), cl::Hidden); 72 73 static cl::opt<bool> EnableCollectLOH( 74 "aarch64-enable-collect-loh", 75 cl::desc("Enable the pass that emits the linker optimization hints (LOH)"), 76 cl::init(true), cl::Hidden); 77 78 static cl::opt<bool> 79 EnableDeadRegisterElimination("aarch64-enable-dead-defs", cl::Hidden, 80 cl::desc("Enable the pass that removes dead" 81 " definitons and replaces stores to" 82 " them with stores to the zero" 83 " register"), 84 cl::init(true)); 85 86 static cl::opt<bool> EnableRedundantCopyElimination( 87 "aarch64-enable-copyelim", 88 cl::desc("Enable the redundant copy elimination pass"), cl::init(true), 89 cl::Hidden); 90 91 static cl::opt<bool> EnableLoadStoreOpt("aarch64-enable-ldst-opt", 92 cl::desc("Enable the load/store pair" 93 " optimization pass"), 94 cl::init(true), cl::Hidden); 95 96 static cl::opt<bool> EnableAtomicTidy( 97 "aarch64-enable-atomic-cfg-tidy", cl::Hidden, 98 cl::desc("Run SimplifyCFG after expanding atomic operations" 99 " to make use of cmpxchg flow-based information"), 100 cl::init(true)); 101 102 static cl::opt<bool> 103 EnableEarlyIfConversion("aarch64-enable-early-ifcvt", cl::Hidden, 104 cl::desc("Run early if-conversion"), 105 cl::init(true)); 106 107 static cl::opt<bool> 108 EnableCondOpt("aarch64-enable-condopt", 109 cl::desc("Enable the condition optimizer pass"), 110 cl::init(true), cl::Hidden); 111 112 static cl::opt<bool> 113 EnableA53Fix835769("aarch64-fix-cortex-a53-835769", cl::Hidden, 114 cl::desc("Work around Cortex-A53 erratum 835769"), 115 cl::init(false)); 116 117 static cl::opt<bool> 118 EnableAddressTypePromotion("aarch64-enable-type-promotion", cl::Hidden, 119 cl::desc("Enable the type promotion pass"), 120 cl::init(true)); 121 122 static cl::opt<bool> 123 EnableGEPOpt("aarch64-enable-gep-opt", cl::Hidden, 124 cl::desc("Enable optimizations on complex GEPs"), 125 cl::init(false)); 126 127 static cl::opt<bool> 128 BranchRelaxation("aarch64-enable-branch-relax", cl::Hidden, cl::init(true), 129 cl::desc("Relax out of range conditional branches")); 130 131 // FIXME: Unify control over GlobalMerge. 132 static cl::opt<cl::boolOrDefault> 133 EnableGlobalMerge("aarch64-enable-global-merge", cl::Hidden, 134 cl::desc("Enable the global merge pass")); 135 136 static cl::opt<bool> 137 EnableLoopDataPrefetch("aarch64-enable-loop-data-prefetch", cl::Hidden, 138 cl::desc("Enable the loop data prefetch pass"), 139 cl::init(true)); 140 141 extern "C" void LLVMInitializeAArch64Target() { 142 // Register the target. 143 RegisterTargetMachine<AArch64leTargetMachine> X(getTheAArch64leTarget()); 144 RegisterTargetMachine<AArch64beTargetMachine> Y(getTheAArch64beTarget()); 145 RegisterTargetMachine<AArch64leTargetMachine> Z(getTheARM64Target()); 146 auto PR = PassRegistry::getPassRegistry(); 147 initializeGlobalISel(*PR); 148 initializeAArch64A53Fix835769Pass(*PR); 149 initializeAArch64A57FPLoadBalancingPass(*PR); 150 initializeAArch64AddressTypePromotionPass(*PR); 151 initializeAArch64AdvSIMDScalarPass(*PR); 152 initializeAArch64CollectLOHPass(*PR); 153 initializeAArch64ConditionalComparesPass(*PR); 154 initializeAArch64ConditionOptimizerPass(*PR); 155 initializeAArch64DeadRegisterDefinitionsPass(*PR); 156 initializeAArch64ExpandPseudoPass(*PR); 157 initializeAArch64LoadStoreOptPass(*PR); 158 initializeAArch64VectorByElementOptPass(*PR); 159 initializeAArch64PromoteConstantPass(*PR); 160 initializeAArch64RedundantCopyEliminationPass(*PR); 161 initializeAArch64StorePairSuppressPass(*PR); 162 initializeLDTLSCleanupPass(*PR); 163 } 164 165 //===----------------------------------------------------------------------===// 166 // AArch64 Lowering public interface. 167 //===----------------------------------------------------------------------===// 168 static std::unique_ptr<TargetLoweringObjectFile> createTLOF(const Triple &TT) { 169 if (TT.isOSBinFormatMachO()) 170 return llvm::make_unique<AArch64_MachoTargetObjectFile>(); 171 172 return llvm::make_unique<AArch64_ELFTargetObjectFile>(); 173 } 174 175 // Helper function to build a DataLayout string 176 static std::string computeDataLayout(const Triple &TT, 177 const MCTargetOptions &Options, 178 bool LittleEndian) { 179 if (Options.getABIName() == "ilp32") 180 return "e-m:e-p:32:32-i8:8-i16:16-i64:64-S128"; 181 if (TT.isOSBinFormatMachO()) 182 return "e-m:o-i64:64-i128:128-n32:64-S128"; 183 if (LittleEndian) 184 return "e-m:e-i8:8:32-i16:16:32-i64:64-i128:128-n32:64-S128"; 185 return "E-m:e-i8:8:32-i16:16:32-i64:64-i128:128-n32:64-S128"; 186 } 187 188 static Reloc::Model getEffectiveRelocModel(const Triple &TT, 189 Optional<Reloc::Model> RM) { 190 // AArch64 Darwin is always PIC. 191 if (TT.isOSDarwin()) 192 return Reloc::PIC_; 193 // On ELF platforms the default static relocation model has a smart enough 194 // linker to cope with referencing external symbols defined in a shared 195 // library. Hence DynamicNoPIC doesn't need to be promoted to PIC. 196 if (!RM.hasValue() || *RM == Reloc::DynamicNoPIC) 197 return Reloc::Static; 198 return *RM; 199 } 200 201 /// Create an AArch64 architecture model. 202 /// 203 AArch64TargetMachine::AArch64TargetMachine( 204 const Target &T, const Triple &TT, StringRef CPU, StringRef FS, 205 const TargetOptions &Options, Optional<Reloc::Model> RM, 206 CodeModel::Model CM, CodeGenOpt::Level OL, bool LittleEndian) 207 // This nested ternary is horrible, but DL needs to be properly 208 // initialized before TLInfo is constructed. 209 : LLVMTargetMachine(T, computeDataLayout(TT, Options.MCOptions, 210 LittleEndian), 211 TT, CPU, FS, Options, 212 getEffectiveRelocModel(TT, RM), CM, OL), 213 TLOF(createTLOF(getTargetTriple())), 214 isLittle(LittleEndian) { 215 initAsmInfo(); 216 } 217 218 AArch64TargetMachine::~AArch64TargetMachine() = default; 219 220 #ifdef LLVM_BUILD_GLOBAL_ISEL 221 namespace { 222 223 struct AArch64GISelActualAccessor : public GISelAccessor { 224 std::unique_ptr<CallLowering> CallLoweringInfo; 225 std::unique_ptr<InstructionSelector> InstSelector; 226 std::unique_ptr<LegalizerInfo> Legalizer; 227 std::unique_ptr<RegisterBankInfo> RegBankInfo; 228 229 const CallLowering *getCallLowering() const override { 230 return CallLoweringInfo.get(); 231 } 232 233 const InstructionSelector *getInstructionSelector() const override { 234 return InstSelector.get(); 235 } 236 237 const LegalizerInfo *getLegalizerInfo() const override { 238 return Legalizer.get(); 239 } 240 241 const RegisterBankInfo *getRegBankInfo() const override { 242 return RegBankInfo.get(); 243 } 244 }; 245 246 } // end anonymous namespace 247 #endif 248 249 const AArch64Subtarget * 250 AArch64TargetMachine::getSubtargetImpl(const Function &F) const { 251 Attribute CPUAttr = F.getFnAttribute("target-cpu"); 252 Attribute FSAttr = F.getFnAttribute("target-features"); 253 254 std::string CPU = !CPUAttr.hasAttribute(Attribute::None) 255 ? CPUAttr.getValueAsString().str() 256 : TargetCPU; 257 std::string FS = !FSAttr.hasAttribute(Attribute::None) 258 ? FSAttr.getValueAsString().str() 259 : TargetFS; 260 261 auto &I = SubtargetMap[CPU + FS]; 262 if (!I) { 263 // This needs to be done before we create a new subtarget since any 264 // creation will depend on the TM and the code generation flags on the 265 // function that reside in TargetOptions. 266 resetTargetOptions(F); 267 I = llvm::make_unique<AArch64Subtarget>(TargetTriple, CPU, FS, *this, 268 isLittle); 269 #ifndef LLVM_BUILD_GLOBAL_ISEL 270 GISelAccessor *GISel = new GISelAccessor(); 271 #else 272 AArch64GISelActualAccessor *GISel = 273 new AArch64GISelActualAccessor(); 274 GISel->CallLoweringInfo.reset( 275 new AArch64CallLowering(*I->getTargetLowering())); 276 GISel->Legalizer.reset(new AArch64LegalizerInfo()); 277 278 auto *RBI = new AArch64RegisterBankInfo(*I->getRegisterInfo()); 279 280 // FIXME: At this point, we can't rely on Subtarget having RBI. 281 // It's awkward to mix passing RBI and the Subtarget; should we pass 282 // TII/TRI as well? 283 GISel->InstSelector.reset(new AArch64InstructionSelector(*this, *I, *RBI)); 284 285 GISel->RegBankInfo.reset(RBI); 286 #endif 287 I->setGISelAccessor(*GISel); 288 } 289 return I.get(); 290 } 291 292 void AArch64leTargetMachine::anchor() { } 293 294 AArch64leTargetMachine::AArch64leTargetMachine( 295 const Target &T, const Triple &TT, StringRef CPU, StringRef FS, 296 const TargetOptions &Options, Optional<Reloc::Model> RM, 297 CodeModel::Model CM, CodeGenOpt::Level OL) 298 : AArch64TargetMachine(T, TT, CPU, FS, Options, RM, CM, OL, true) {} 299 300 void AArch64beTargetMachine::anchor() { } 301 302 AArch64beTargetMachine::AArch64beTargetMachine( 303 const Target &T, const Triple &TT, StringRef CPU, StringRef FS, 304 const TargetOptions &Options, Optional<Reloc::Model> RM, 305 CodeModel::Model CM, CodeGenOpt::Level OL) 306 : AArch64TargetMachine(T, TT, CPU, FS, Options, RM, CM, OL, false) {} 307 308 namespace { 309 310 /// AArch64 Code Generator Pass Configuration Options. 311 class AArch64PassConfig : public TargetPassConfig { 312 public: 313 AArch64PassConfig(AArch64TargetMachine *TM, PassManagerBase &PM) 314 : TargetPassConfig(TM, PM) { 315 if (TM->getOptLevel() != CodeGenOpt::None) 316 substitutePass(&PostRASchedulerID, &PostMachineSchedulerID); 317 } 318 319 AArch64TargetMachine &getAArch64TargetMachine() const { 320 return getTM<AArch64TargetMachine>(); 321 } 322 323 ScheduleDAGInstrs * 324 createMachineScheduler(MachineSchedContext *C) const override { 325 ScheduleDAGMILive *DAG = createGenericSchedLive(C); 326 DAG->addMutation(createLoadClusterDAGMutation(DAG->TII, DAG->TRI)); 327 DAG->addMutation(createStoreClusterDAGMutation(DAG->TII, DAG->TRI)); 328 DAG->addMutation(createMacroFusionDAGMutation(DAG->TII)); 329 return DAG; 330 } 331 332 void addIRPasses() override; 333 bool addPreISel() override; 334 bool addInstSelector() override; 335 #ifdef LLVM_BUILD_GLOBAL_ISEL 336 bool addIRTranslator() override; 337 bool addLegalizeMachineIR() override; 338 bool addRegBankSelect() override; 339 bool addGlobalInstructionSelect() override; 340 #endif 341 bool addILPOpts() override; 342 void addPreRegAlloc() override; 343 void addPostRegAlloc() override; 344 void addPreSched2() override; 345 void addPreEmitPass() override; 346 }; 347 348 } // end anonymous namespace 349 350 TargetIRAnalysis AArch64TargetMachine::getTargetIRAnalysis() { 351 return TargetIRAnalysis([this](const Function &F) { 352 return TargetTransformInfo(AArch64TTIImpl(this, F)); 353 }); 354 } 355 356 TargetPassConfig *AArch64TargetMachine::createPassConfig(PassManagerBase &PM) { 357 return new AArch64PassConfig(this, PM); 358 } 359 360 void AArch64PassConfig::addIRPasses() { 361 // Always expand atomic operations, we don't deal with atomicrmw or cmpxchg 362 // ourselves. 363 addPass(createAtomicExpandPass(TM)); 364 365 // Cmpxchg instructions are often used with a subsequent comparison to 366 // determine whether it succeeded. We can exploit existing control-flow in 367 // ldrex/strex loops to simplify this, but it needs tidying up. 368 if (TM->getOptLevel() != CodeGenOpt::None && EnableAtomicTidy) 369 addPass(createCFGSimplificationPass()); 370 371 // Run LoopDataPrefetch 372 // 373 // Run this before LSR to remove the multiplies involved in computing the 374 // pointer values N iterations ahead. 375 if (TM->getOptLevel() != CodeGenOpt::None && EnableLoopDataPrefetch) 376 addPass(createLoopDataPrefetchPass()); 377 378 TargetPassConfig::addIRPasses(); 379 380 // Match interleaved memory accesses to ldN/stN intrinsics. 381 if (TM->getOptLevel() != CodeGenOpt::None) 382 addPass(createInterleavedAccessPass(TM)); 383 384 if (TM->getOptLevel() == CodeGenOpt::Aggressive && EnableGEPOpt) { 385 // Call SeparateConstOffsetFromGEP pass to extract constants within indices 386 // and lower a GEP with multiple indices to either arithmetic operations or 387 // multiple GEPs with single index. 388 addPass(createSeparateConstOffsetFromGEPPass(TM, true)); 389 // Call EarlyCSE pass to find and remove subexpressions in the lowered 390 // result. 391 addPass(createEarlyCSEPass()); 392 // Do loop invariant code motion in case part of the lowered result is 393 // invariant. 394 addPass(createLICMPass()); 395 } 396 } 397 398 // Pass Pipeline Configuration 399 bool AArch64PassConfig::addPreISel() { 400 // Run promote constant before global merge, so that the promoted constants 401 // get a chance to be merged 402 if (TM->getOptLevel() != CodeGenOpt::None && EnablePromoteConstant) 403 addPass(createAArch64PromoteConstantPass()); 404 // FIXME: On AArch64, this depends on the type. 405 // Basically, the addressable offsets are up to 4095 * Ty.getSizeInBytes(). 406 // and the offset has to be a multiple of the related size in bytes. 407 if ((TM->getOptLevel() != CodeGenOpt::None && 408 EnableGlobalMerge == cl::BOU_UNSET) || 409 EnableGlobalMerge == cl::BOU_TRUE) { 410 bool OnlyOptimizeForSize = (TM->getOptLevel() < CodeGenOpt::Aggressive) && 411 (EnableGlobalMerge == cl::BOU_UNSET); 412 addPass(createGlobalMergePass(TM, 4095, OnlyOptimizeForSize)); 413 } 414 415 if (TM->getOptLevel() != CodeGenOpt::None && EnableAddressTypePromotion) 416 addPass(createAArch64AddressTypePromotionPass()); 417 418 return false; 419 } 420 421 bool AArch64PassConfig::addInstSelector() { 422 addPass(createAArch64ISelDag(getAArch64TargetMachine(), getOptLevel())); 423 424 // For ELF, cleanup any local-dynamic TLS accesses (i.e. combine as many 425 // references to _TLS_MODULE_BASE_ as possible. 426 if (TM->getTargetTriple().isOSBinFormatELF() && 427 getOptLevel() != CodeGenOpt::None) 428 addPass(createAArch64CleanupLocalDynamicTLSPass()); 429 430 return false; 431 } 432 433 #ifdef LLVM_BUILD_GLOBAL_ISEL 434 bool AArch64PassConfig::addIRTranslator() { 435 addPass(new IRTranslator()); 436 return false; 437 } 438 439 bool AArch64PassConfig::addLegalizeMachineIR() { 440 addPass(new Legalizer()); 441 return false; 442 } 443 444 bool AArch64PassConfig::addRegBankSelect() { 445 addPass(new RegBankSelect()); 446 return false; 447 } 448 449 bool AArch64PassConfig::addGlobalInstructionSelect() { 450 addPass(new InstructionSelect()); 451 return false; 452 } 453 #endif 454 455 bool AArch64PassConfig::addILPOpts() { 456 if (EnableCondOpt) 457 addPass(createAArch64ConditionOptimizerPass()); 458 if (EnableCCMP) 459 addPass(createAArch64ConditionalCompares()); 460 if (EnableMCR) 461 addPass(&MachineCombinerID); 462 if (EnableEarlyIfConversion) 463 addPass(&EarlyIfConverterID); 464 if (EnableStPairSuppress) 465 addPass(createAArch64StorePairSuppressPass()); 466 addPass(createAArch64VectorByElementOptPass()); 467 return true; 468 } 469 470 void AArch64PassConfig::addPreRegAlloc() { 471 // Change dead register definitions to refer to the zero register. 472 if (TM->getOptLevel() != CodeGenOpt::None && EnableDeadRegisterElimination) 473 addPass(createAArch64DeadRegisterDefinitions()); 474 475 // Use AdvSIMD scalar instructions whenever profitable. 476 if (TM->getOptLevel() != CodeGenOpt::None && EnableAdvSIMDScalar) { 477 addPass(createAArch64AdvSIMDScalar()); 478 // The AdvSIMD pass may produce copies that can be rewritten to 479 // be register coaleascer friendly. 480 addPass(&PeepholeOptimizerID); 481 } 482 } 483 484 void AArch64PassConfig::addPostRegAlloc() { 485 // Remove redundant copy instructions. 486 if (TM->getOptLevel() != CodeGenOpt::None && EnableRedundantCopyElimination) 487 addPass(createAArch64RedundantCopyEliminationPass()); 488 489 if (TM->getOptLevel() != CodeGenOpt::None && usingDefaultRegAlloc()) 490 // Improve performance for some FP/SIMD code for A57. 491 addPass(createAArch64A57FPLoadBalancing()); 492 } 493 494 void AArch64PassConfig::addPreSched2() { 495 // Expand some pseudo instructions to allow proper scheduling. 496 addPass(createAArch64ExpandPseudoPass()); 497 // Use load/store pair instructions when possible. 498 if (TM->getOptLevel() != CodeGenOpt::None && EnableLoadStoreOpt) 499 addPass(createAArch64LoadStoreOptimizationPass()); 500 } 501 502 void AArch64PassConfig::addPreEmitPass() { 503 if (EnableA53Fix835769) 504 addPass(createAArch64A53Fix835769()); 505 // Relax conditional branch instructions if they're otherwise out of 506 // range of their destination. 507 if (BranchRelaxation) 508 addPass(&BranchRelaxationPassID); 509 510 if (TM->getOptLevel() != CodeGenOpt::None && EnableCollectLOH && 511 TM->getTargetTriple().isOSBinFormatMachO()) 512 addPass(createAArch64CollectLOHPass()); 513 } 514