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 "AArch64MacroFusion.h" 15 #include "AArch64Subtarget.h" 16 #include "AArch64TargetMachine.h" 17 #include "AArch64TargetObjectFile.h" 18 #include "AArch64TargetTransformInfo.h" 19 #include "MCTargetDesc/AArch64MCTargetDesc.h" 20 #include "llvm/ADT/STLExtras.h" 21 #include "llvm/ADT/Triple.h" 22 #include "llvm/Analysis/TargetTransformInfo.h" 23 #include "llvm/CodeGen/GlobalISel/IRTranslator.h" 24 #include "llvm/CodeGen/GlobalISel/InstructionSelect.h" 25 #include "llvm/CodeGen/GlobalISel/Legalizer.h" 26 #include "llvm/CodeGen/GlobalISel/RegBankSelect.h" 27 #include "llvm/CodeGen/MachineScheduler.h" 28 #include "llvm/CodeGen/Passes.h" 29 #include "llvm/CodeGen/TargetPassConfig.h" 30 #include "llvm/IR/Attributes.h" 31 #include "llvm/IR/Function.h" 32 #include "llvm/MC/MCTargetOptions.h" 33 #include "llvm/Pass.h" 34 #include "llvm/Support/CodeGen.h" 35 #include "llvm/Support/CommandLine.h" 36 #include "llvm/Support/TargetRegistry.h" 37 #include "llvm/Target/TargetLoweringObjectFile.h" 38 #include "llvm/Target/TargetOptions.h" 39 #include "llvm/Transforms/Scalar.h" 40 #include <memory> 41 #include <string> 42 43 using namespace llvm; 44 45 static cl::opt<bool> EnableCCMP("aarch64-enable-ccmp", 46 cl::desc("Enable the CCMP formation pass"), 47 cl::init(true), cl::Hidden); 48 49 static cl::opt<bool> EnableMCR("aarch64-enable-mcr", 50 cl::desc("Enable the machine combiner pass"), 51 cl::init(true), cl::Hidden); 52 53 static cl::opt<bool> EnableStPairSuppress("aarch64-enable-stp-suppress", 54 cl::desc("Suppress STP for AArch64"), 55 cl::init(true), cl::Hidden); 56 57 static cl::opt<bool> EnableAdvSIMDScalar( 58 "aarch64-enable-simd-scalar", 59 cl::desc("Enable use of AdvSIMD scalar integer instructions"), 60 cl::init(false), cl::Hidden); 61 62 static cl::opt<bool> 63 EnablePromoteConstant("aarch64-enable-promote-const", 64 cl::desc("Enable the promote constant pass"), 65 cl::init(true), cl::Hidden); 66 67 static cl::opt<bool> EnableCollectLOH( 68 "aarch64-enable-collect-loh", 69 cl::desc("Enable the pass that emits the linker optimization hints (LOH)"), 70 cl::init(true), cl::Hidden); 71 72 static cl::opt<bool> 73 EnableDeadRegisterElimination("aarch64-enable-dead-defs", cl::Hidden, 74 cl::desc("Enable the pass that removes dead" 75 " definitons and replaces stores to" 76 " them with stores to the zero" 77 " register"), 78 cl::init(true)); 79 80 static cl::opt<bool> EnableRedundantCopyElimination( 81 "aarch64-enable-copyelim", 82 cl::desc("Enable the redundant copy elimination pass"), cl::init(true), 83 cl::Hidden); 84 85 static cl::opt<bool> EnableLoadStoreOpt("aarch64-enable-ldst-opt", 86 cl::desc("Enable the load/store pair" 87 " optimization pass"), 88 cl::init(true), cl::Hidden); 89 90 static cl::opt<bool> EnableAtomicTidy( 91 "aarch64-enable-atomic-cfg-tidy", cl::Hidden, 92 cl::desc("Run SimplifyCFG after expanding atomic operations" 93 " to make use of cmpxchg flow-based information"), 94 cl::init(true)); 95 96 static cl::opt<bool> 97 EnableEarlyIfConversion("aarch64-enable-early-ifcvt", cl::Hidden, 98 cl::desc("Run early if-conversion"), 99 cl::init(true)); 100 101 static cl::opt<bool> 102 EnableCondOpt("aarch64-enable-condopt", 103 cl::desc("Enable the condition optimizer pass"), 104 cl::init(true), cl::Hidden); 105 106 static cl::opt<bool> 107 EnableA53Fix835769("aarch64-fix-cortex-a53-835769", cl::Hidden, 108 cl::desc("Work around Cortex-A53 erratum 835769"), 109 cl::init(false)); 110 111 static cl::opt<bool> 112 EnableGEPOpt("aarch64-enable-gep-opt", cl::Hidden, 113 cl::desc("Enable optimizations on complex GEPs"), 114 cl::init(false)); 115 116 static cl::opt<bool> 117 BranchRelaxation("aarch64-enable-branch-relax", cl::Hidden, cl::init(true), 118 cl::desc("Relax out of range conditional branches")); 119 120 // FIXME: Unify control over GlobalMerge. 121 static cl::opt<cl::boolOrDefault> 122 EnableGlobalMerge("aarch64-enable-global-merge", cl::Hidden, 123 cl::desc("Enable the global merge pass")); 124 125 static cl::opt<bool> 126 EnableLoopDataPrefetch("aarch64-enable-loop-data-prefetch", cl::Hidden, 127 cl::desc("Enable the loop data prefetch pass"), 128 cl::init(true)); 129 130 static cl::opt<int> EnableGlobalISelAtO( 131 "aarch64-enable-global-isel-at-O", cl::Hidden, 132 cl::desc("Enable GlobalISel at or below an opt level (-1 to disable)"), 133 cl::init(-1)); 134 135 extern "C" void LLVMInitializeAArch64Target() { 136 // Register the target. 137 RegisterTargetMachine<AArch64leTargetMachine> X(getTheAArch64leTarget()); 138 RegisterTargetMachine<AArch64beTargetMachine> Y(getTheAArch64beTarget()); 139 RegisterTargetMachine<AArch64leTargetMachine> Z(getTheARM64Target()); 140 auto PR = PassRegistry::getPassRegistry(); 141 initializeGlobalISel(*PR); 142 initializeAArch64A53Fix835769Pass(*PR); 143 initializeAArch64A57FPLoadBalancingPass(*PR); 144 initializeAArch64AdvSIMDScalarPass(*PR); 145 initializeAArch64CollectLOHPass(*PR); 146 initializeAArch64ConditionalComparesPass(*PR); 147 initializeAArch64ConditionOptimizerPass(*PR); 148 initializeAArch64DeadRegisterDefinitionsPass(*PR); 149 initializeAArch64ExpandPseudoPass(*PR); 150 initializeAArch64LoadStoreOptPass(*PR); 151 initializeAArch64VectorByElementOptPass(*PR); 152 initializeAArch64PromoteConstantPass(*PR); 153 initializeAArch64RedundantCopyEliminationPass(*PR); 154 initializeAArch64StorePairSuppressPass(*PR); 155 initializeLDTLSCleanupPass(*PR); 156 } 157 158 //===----------------------------------------------------------------------===// 159 // AArch64 Lowering public interface. 160 //===----------------------------------------------------------------------===// 161 static std::unique_ptr<TargetLoweringObjectFile> createTLOF(const Triple &TT) { 162 if (TT.isOSBinFormatMachO()) 163 return llvm::make_unique<AArch64_MachoTargetObjectFile>(); 164 165 return llvm::make_unique<AArch64_ELFTargetObjectFile>(); 166 } 167 168 // Helper function to build a DataLayout string 169 static std::string computeDataLayout(const Triple &TT, 170 const MCTargetOptions &Options, 171 bool LittleEndian) { 172 if (Options.getABIName() == "ilp32") 173 return "e-m:e-p:32:32-i8:8-i16:16-i64:64-S128"; 174 if (TT.isOSBinFormatMachO()) 175 return "e-m:o-i64:64-i128:128-n32:64-S128"; 176 if (LittleEndian) 177 return "e-m:e-i8:8:32-i16:16:32-i64:64-i128:128-n32:64-S128"; 178 return "E-m:e-i8:8:32-i16:16:32-i64:64-i128:128-n32:64-S128"; 179 } 180 181 static Reloc::Model getEffectiveRelocModel(const Triple &TT, 182 Optional<Reloc::Model> RM) { 183 // AArch64 Darwin is always PIC. 184 if (TT.isOSDarwin()) 185 return Reloc::PIC_; 186 // On ELF platforms the default static relocation model has a smart enough 187 // linker to cope with referencing external symbols defined in a shared 188 // library. Hence DynamicNoPIC doesn't need to be promoted to PIC. 189 if (!RM.hasValue() || *RM == Reloc::DynamicNoPIC) 190 return Reloc::Static; 191 return *RM; 192 } 193 194 /// Create an AArch64 architecture model. 195 /// 196 AArch64TargetMachine::AArch64TargetMachine( 197 const Target &T, const Triple &TT, StringRef CPU, StringRef FS, 198 const TargetOptions &Options, Optional<Reloc::Model> RM, 199 CodeModel::Model CM, CodeGenOpt::Level OL, bool LittleEndian) 200 // This nested ternary is horrible, but DL needs to be properly 201 // initialized before TLInfo is constructed. 202 : LLVMTargetMachine(T, computeDataLayout(TT, Options.MCOptions, 203 LittleEndian), 204 TT, CPU, FS, Options, 205 getEffectiveRelocModel(TT, RM), CM, OL), 206 TLOF(createTLOF(getTargetTriple())), 207 isLittle(LittleEndian) { 208 initAsmInfo(); 209 } 210 211 AArch64TargetMachine::~AArch64TargetMachine() = default; 212 213 const AArch64Subtarget * 214 AArch64TargetMachine::getSubtargetImpl(const Function &F) const { 215 Attribute CPUAttr = F.getFnAttribute("target-cpu"); 216 Attribute FSAttr = F.getFnAttribute("target-features"); 217 218 std::string CPU = !CPUAttr.hasAttribute(Attribute::None) 219 ? CPUAttr.getValueAsString().str() 220 : TargetCPU; 221 std::string FS = !FSAttr.hasAttribute(Attribute::None) 222 ? FSAttr.getValueAsString().str() 223 : TargetFS; 224 225 auto &I = SubtargetMap[CPU + FS]; 226 if (!I) { 227 // This needs to be done before we create a new subtarget since any 228 // creation will depend on the TM and the code generation flags on the 229 // function that reside in TargetOptions. 230 resetTargetOptions(F); 231 I = llvm::make_unique<AArch64Subtarget>(TargetTriple, CPU, FS, *this, 232 isLittle); 233 } 234 return I.get(); 235 } 236 237 void AArch64leTargetMachine::anchor() { } 238 239 AArch64leTargetMachine::AArch64leTargetMachine( 240 const Target &T, const Triple &TT, StringRef CPU, StringRef FS, 241 const TargetOptions &Options, Optional<Reloc::Model> RM, 242 CodeModel::Model CM, CodeGenOpt::Level OL) 243 : AArch64TargetMachine(T, TT, CPU, FS, Options, RM, CM, OL, true) {} 244 245 void AArch64beTargetMachine::anchor() { } 246 247 AArch64beTargetMachine::AArch64beTargetMachine( 248 const Target &T, const Triple &TT, StringRef CPU, StringRef FS, 249 const TargetOptions &Options, Optional<Reloc::Model> RM, 250 CodeModel::Model CM, CodeGenOpt::Level OL) 251 : AArch64TargetMachine(T, TT, CPU, FS, Options, RM, CM, OL, false) {} 252 253 namespace { 254 255 /// AArch64 Code Generator Pass Configuration Options. 256 class AArch64PassConfig : public TargetPassConfig { 257 public: 258 AArch64PassConfig(AArch64TargetMachine *TM, PassManagerBase &PM) 259 : TargetPassConfig(TM, PM) { 260 if (TM->getOptLevel() != CodeGenOpt::None) 261 substitutePass(&PostRASchedulerID, &PostMachineSchedulerID); 262 } 263 264 AArch64TargetMachine &getAArch64TargetMachine() const { 265 return getTM<AArch64TargetMachine>(); 266 } 267 268 ScheduleDAGInstrs * 269 createMachineScheduler(MachineSchedContext *C) const override { 270 ScheduleDAGMILive *DAG = createGenericSchedLive(C); 271 DAG->addMutation(createLoadClusterDAGMutation(DAG->TII, DAG->TRI)); 272 DAG->addMutation(createStoreClusterDAGMutation(DAG->TII, DAG->TRI)); 273 DAG->addMutation(createAArch64MacroFusionDAGMutation()); 274 return DAG; 275 } 276 277 ScheduleDAGInstrs * 278 createPostMachineScheduler(MachineSchedContext *C) const override { 279 const AArch64Subtarget &ST = C->MF->getSubtarget<AArch64Subtarget>(); 280 if (ST.hasFuseAES() || ST.hasFuseLiterals()) { 281 // Run the Macro Fusion after RA again since literals are expanded from 282 // pseudos then (v. addPreSched2()). 283 ScheduleDAGMI *DAG = createGenericSchedPostRA(C); 284 DAG->addMutation(createAArch64MacroFusionDAGMutation()); 285 return DAG; 286 } 287 288 return nullptr; 289 } 290 291 void addIRPasses() override; 292 bool addPreISel() override; 293 bool addInstSelector() override; 294 #ifdef LLVM_BUILD_GLOBAL_ISEL 295 bool addIRTranslator() override; 296 bool addLegalizeMachineIR() override; 297 bool addRegBankSelect() override; 298 bool addGlobalInstructionSelect() override; 299 #endif 300 bool addILPOpts() override; 301 void addPreRegAlloc() override; 302 void addPostRegAlloc() override; 303 void addPreSched2() override; 304 void addPreEmitPass() override; 305 306 bool isGlobalISelEnabled() const override; 307 }; 308 309 } // end anonymous namespace 310 311 TargetIRAnalysis AArch64TargetMachine::getTargetIRAnalysis() { 312 return TargetIRAnalysis([this](const Function &F) { 313 return TargetTransformInfo(AArch64TTIImpl(this, F)); 314 }); 315 } 316 317 TargetPassConfig *AArch64TargetMachine::createPassConfig(PassManagerBase &PM) { 318 return new AArch64PassConfig(this, PM); 319 } 320 321 void AArch64PassConfig::addIRPasses() { 322 // Always expand atomic operations, we don't deal with atomicrmw or cmpxchg 323 // ourselves. 324 addPass(createAtomicExpandPass()); 325 326 // Cmpxchg instructions are often used with a subsequent comparison to 327 // determine whether it succeeded. We can exploit existing control-flow in 328 // ldrex/strex loops to simplify this, but it needs tidying up. 329 if (TM->getOptLevel() != CodeGenOpt::None && EnableAtomicTidy) 330 addPass(createCFGSimplificationPass()); 331 332 // Run LoopDataPrefetch 333 // 334 // Run this before LSR to remove the multiplies involved in computing the 335 // pointer values N iterations ahead. 336 if (TM->getOptLevel() != CodeGenOpt::None && EnableLoopDataPrefetch) 337 addPass(createLoopDataPrefetchPass()); 338 339 TargetPassConfig::addIRPasses(); 340 341 // Match interleaved memory accesses to ldN/stN intrinsics. 342 if (TM->getOptLevel() != CodeGenOpt::None) 343 addPass(createInterleavedAccessPass()); 344 345 if (TM->getOptLevel() == CodeGenOpt::Aggressive && EnableGEPOpt) { 346 // Call SeparateConstOffsetFromGEP pass to extract constants within indices 347 // and lower a GEP with multiple indices to either arithmetic operations or 348 // multiple GEPs with single index. 349 addPass(createSeparateConstOffsetFromGEPPass(TM, true)); 350 // Call EarlyCSE pass to find and remove subexpressions in the lowered 351 // result. 352 addPass(createEarlyCSEPass()); 353 // Do loop invariant code motion in case part of the lowered result is 354 // invariant. 355 addPass(createLICMPass()); 356 } 357 } 358 359 // Pass Pipeline Configuration 360 bool AArch64PassConfig::addPreISel() { 361 // Run promote constant before global merge, so that the promoted constants 362 // get a chance to be merged 363 if (TM->getOptLevel() != CodeGenOpt::None && EnablePromoteConstant) 364 addPass(createAArch64PromoteConstantPass()); 365 // FIXME: On AArch64, this depends on the type. 366 // Basically, the addressable offsets are up to 4095 * Ty.getSizeInBytes(). 367 // and the offset has to be a multiple of the related size in bytes. 368 if ((TM->getOptLevel() != CodeGenOpt::None && 369 EnableGlobalMerge == cl::BOU_UNSET) || 370 EnableGlobalMerge == cl::BOU_TRUE) { 371 bool OnlyOptimizeForSize = (TM->getOptLevel() < CodeGenOpt::Aggressive) && 372 (EnableGlobalMerge == cl::BOU_UNSET); 373 addPass(createGlobalMergePass(TM, 4095, OnlyOptimizeForSize)); 374 } 375 376 return false; 377 } 378 379 bool AArch64PassConfig::addInstSelector() { 380 addPass(createAArch64ISelDag(getAArch64TargetMachine(), getOptLevel())); 381 382 // For ELF, cleanup any local-dynamic TLS accesses (i.e. combine as many 383 // references to _TLS_MODULE_BASE_ as possible. 384 if (TM->getTargetTriple().isOSBinFormatELF() && 385 getOptLevel() != CodeGenOpt::None) 386 addPass(createAArch64CleanupLocalDynamicTLSPass()); 387 388 return false; 389 } 390 391 #ifdef LLVM_BUILD_GLOBAL_ISEL 392 bool AArch64PassConfig::addIRTranslator() { 393 addPass(new IRTranslator()); 394 return false; 395 } 396 397 bool AArch64PassConfig::addLegalizeMachineIR() { 398 addPass(new Legalizer()); 399 return false; 400 } 401 402 bool AArch64PassConfig::addRegBankSelect() { 403 addPass(new RegBankSelect()); 404 return false; 405 } 406 407 bool AArch64PassConfig::addGlobalInstructionSelect() { 408 addPass(new InstructionSelect()); 409 return false; 410 } 411 #endif 412 413 bool AArch64PassConfig::isGlobalISelEnabled() const { 414 return TM->getOptLevel() <= EnableGlobalISelAtO; 415 } 416 417 bool AArch64PassConfig::addILPOpts() { 418 if (EnableCondOpt) 419 addPass(createAArch64ConditionOptimizerPass()); 420 if (EnableCCMP) 421 addPass(createAArch64ConditionalCompares()); 422 if (EnableMCR) 423 addPass(&MachineCombinerID); 424 if (EnableEarlyIfConversion) 425 addPass(&EarlyIfConverterID); 426 if (EnableStPairSuppress) 427 addPass(createAArch64StorePairSuppressPass()); 428 addPass(createAArch64VectorByElementOptPass()); 429 return true; 430 } 431 432 void AArch64PassConfig::addPreRegAlloc() { 433 // Change dead register definitions to refer to the zero register. 434 if (TM->getOptLevel() != CodeGenOpt::None && EnableDeadRegisterElimination) 435 addPass(createAArch64DeadRegisterDefinitions()); 436 437 // Use AdvSIMD scalar instructions whenever profitable. 438 if (TM->getOptLevel() != CodeGenOpt::None && EnableAdvSIMDScalar) { 439 addPass(createAArch64AdvSIMDScalar()); 440 // The AdvSIMD pass may produce copies that can be rewritten to 441 // be register coaleascer friendly. 442 addPass(&PeepholeOptimizerID); 443 } 444 } 445 446 void AArch64PassConfig::addPostRegAlloc() { 447 // Remove redundant copy instructions. 448 if (TM->getOptLevel() != CodeGenOpt::None && EnableRedundantCopyElimination) 449 addPass(createAArch64RedundantCopyEliminationPass()); 450 451 if (TM->getOptLevel() != CodeGenOpt::None && usingDefaultRegAlloc()) 452 // Improve performance for some FP/SIMD code for A57. 453 addPass(createAArch64A57FPLoadBalancing()); 454 } 455 456 void AArch64PassConfig::addPreSched2() { 457 // Expand some pseudo instructions to allow proper scheduling. 458 addPass(createAArch64ExpandPseudoPass()); 459 // Use load/store pair instructions when possible. 460 if (TM->getOptLevel() != CodeGenOpt::None && EnableLoadStoreOpt) 461 addPass(createAArch64LoadStoreOptimizationPass()); 462 } 463 464 void AArch64PassConfig::addPreEmitPass() { 465 if (EnableA53Fix835769) 466 addPass(createAArch64A53Fix835769()); 467 // Relax conditional branch instructions if they're otherwise out of 468 // range of their destination. 469 if (BranchRelaxation) 470 addPass(&BranchRelaxationPassID); 471 472 if (TM->getOptLevel() != CodeGenOpt::None && EnableCollectLOH && 473 TM->getTargetTriple().isOSBinFormatMachO()) 474 addPass(createAArch64CollectLOHPass()); 475 } 476