1 //===-- ARMTargetMachine.cpp - Define TargetMachine for ARM ---------------===// 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 "ARMTargetMachine.h" 13 #include "ARM.h" 14 #include "ARMMacroFusion.h" 15 #include "ARMSubtarget.h" 16 #include "ARMTargetObjectFile.h" 17 #include "ARMTargetTransformInfo.h" 18 #include "MCTargetDesc/ARMMCTargetDesc.h" 19 #include "TargetInfo/ARMTargetInfo.h" 20 #include "llvm/ADT/Optional.h" 21 #include "llvm/ADT/STLExtras.h" 22 #include "llvm/ADT/StringRef.h" 23 #include "llvm/ADT/Triple.h" 24 #include "llvm/Analysis/TargetTransformInfo.h" 25 #include "llvm/CodeGen/ExecutionDomainFix.h" 26 #include "llvm/CodeGen/GlobalISel/CallLowering.h" 27 #include "llvm/CodeGen/GlobalISel/IRTranslator.h" 28 #include "llvm/CodeGen/GlobalISel/InstructionSelect.h" 29 #include "llvm/CodeGen/GlobalISel/InstructionSelector.h" 30 #include "llvm/CodeGen/GlobalISel/Legalizer.h" 31 #include "llvm/CodeGen/GlobalISel/LegalizerInfo.h" 32 #include "llvm/CodeGen/GlobalISel/RegBankSelect.h" 33 #include "llvm/CodeGen/GlobalISel/RegisterBankInfo.h" 34 #include "llvm/CodeGen/MachineFunction.h" 35 #include "llvm/CodeGen/MachineScheduler.h" 36 #include "llvm/CodeGen/Passes.h" 37 #include "llvm/CodeGen/TargetPassConfig.h" 38 #include "llvm/IR/Attributes.h" 39 #include "llvm/IR/DataLayout.h" 40 #include "llvm/IR/Function.h" 41 #include "llvm/Pass.h" 42 #include "llvm/Support/CodeGen.h" 43 #include "llvm/Support/CommandLine.h" 44 #include "llvm/Support/ErrorHandling.h" 45 #include "llvm/Support/TargetParser.h" 46 #include "llvm/Support/TargetRegistry.h" 47 #include "llvm/Target/TargetLoweringObjectFile.h" 48 #include "llvm/Target/TargetOptions.h" 49 #include "llvm/Transforms/CFGuard.h" 50 #include "llvm/Transforms/Scalar.h" 51 #include <cassert> 52 #include <memory> 53 #include <string> 54 55 using namespace llvm; 56 57 static cl::opt<bool> 58 DisableA15SDOptimization("disable-a15-sd-optimization", cl::Hidden, 59 cl::desc("Inhibit optimization of S->D register accesses on A15"), 60 cl::init(false)); 61 62 static cl::opt<bool> 63 EnableAtomicTidy("arm-atomic-cfg-tidy", cl::Hidden, 64 cl::desc("Run SimplifyCFG after expanding atomic operations" 65 " to make use of cmpxchg flow-based information"), 66 cl::init(true)); 67 68 static cl::opt<bool> 69 EnableARMLoadStoreOpt("arm-load-store-opt", cl::Hidden, 70 cl::desc("Enable ARM load/store optimization pass"), 71 cl::init(true)); 72 73 // FIXME: Unify control over GlobalMerge. 74 static cl::opt<cl::boolOrDefault> 75 EnableGlobalMerge("arm-global-merge", cl::Hidden, 76 cl::desc("Enable the global merge pass")); 77 78 namespace llvm { 79 void initializeARMExecutionDomainFixPass(PassRegistry&); 80 } 81 82 extern "C" LLVM_EXTERNAL_VISIBILITY void LLVMInitializeARMTarget() { 83 // Register the target. 84 RegisterTargetMachine<ARMLETargetMachine> X(getTheARMLETarget()); 85 RegisterTargetMachine<ARMLETargetMachine> A(getTheThumbLETarget()); 86 RegisterTargetMachine<ARMBETargetMachine> Y(getTheARMBETarget()); 87 RegisterTargetMachine<ARMBETargetMachine> B(getTheThumbBETarget()); 88 89 PassRegistry &Registry = *PassRegistry::getPassRegistry(); 90 initializeGlobalISel(Registry); 91 initializeARMLoadStoreOptPass(Registry); 92 initializeARMPreAllocLoadStoreOptPass(Registry); 93 initializeARMParallelDSPPass(Registry); 94 initializeARMConstantIslandsPass(Registry); 95 initializeARMExecutionDomainFixPass(Registry); 96 initializeARMExpandPseudoPass(Registry); 97 initializeThumb2SizeReducePass(Registry); 98 initializeMVEVPTBlockPass(Registry); 99 initializeMVETailPredicationPass(Registry); 100 initializeARMLowOverheadLoopsPass(Registry); 101 initializeMVEGatherScatterLoweringPass(Registry); 102 } 103 104 static std::unique_ptr<TargetLoweringObjectFile> createTLOF(const Triple &TT) { 105 if (TT.isOSBinFormatMachO()) 106 return std::make_unique<TargetLoweringObjectFileMachO>(); 107 if (TT.isOSWindows()) 108 return std::make_unique<TargetLoweringObjectFileCOFF>(); 109 return std::make_unique<ARMElfTargetObjectFile>(); 110 } 111 112 static ARMBaseTargetMachine::ARMABI 113 computeTargetABI(const Triple &TT, StringRef CPU, 114 const TargetOptions &Options) { 115 StringRef ABIName = Options.MCOptions.getABIName(); 116 117 if (ABIName.empty()) 118 ABIName = ARM::computeDefaultTargetABI(TT, CPU); 119 120 if (ABIName == "aapcs16") 121 return ARMBaseTargetMachine::ARM_ABI_AAPCS16; 122 else if (ABIName.startswith("aapcs")) 123 return ARMBaseTargetMachine::ARM_ABI_AAPCS; 124 else if (ABIName.startswith("apcs")) 125 return ARMBaseTargetMachine::ARM_ABI_APCS; 126 127 llvm_unreachable("Unhandled/unknown ABI Name!"); 128 return ARMBaseTargetMachine::ARM_ABI_UNKNOWN; 129 } 130 131 static std::string computeDataLayout(const Triple &TT, StringRef CPU, 132 const TargetOptions &Options, 133 bool isLittle) { 134 auto ABI = computeTargetABI(TT, CPU, Options); 135 std::string Ret; 136 137 if (isLittle) 138 // Little endian. 139 Ret += "e"; 140 else 141 // Big endian. 142 Ret += "E"; 143 144 Ret += DataLayout::getManglingComponent(TT); 145 146 // Pointers are 32 bits and aligned to 32 bits. 147 Ret += "-p:32:32"; 148 149 // Function pointers are aligned to 8 bits (because the LSB stores the 150 // ARM/Thumb state). 151 Ret += "-Fi8"; 152 153 // ABIs other than APCS have 64 bit integers with natural alignment. 154 if (ABI != ARMBaseTargetMachine::ARM_ABI_APCS) 155 Ret += "-i64:64"; 156 157 // We have 64 bits floats. The APCS ABI requires them to be aligned to 32 158 // bits, others to 64 bits. We always try to align to 64 bits. 159 if (ABI == ARMBaseTargetMachine::ARM_ABI_APCS) 160 Ret += "-f64:32:64"; 161 162 // We have 128 and 64 bit vectors. The APCS ABI aligns them to 32 bits, others 163 // to 64. We always ty to give them natural alignment. 164 if (ABI == ARMBaseTargetMachine::ARM_ABI_APCS) 165 Ret += "-v64:32:64-v128:32:128"; 166 else if (ABI != ARMBaseTargetMachine::ARM_ABI_AAPCS16) 167 Ret += "-v128:64:128"; 168 169 // Try to align aggregates to 32 bits (the default is 64 bits, which has no 170 // particular hardware support on 32-bit ARM). 171 Ret += "-a:0:32"; 172 173 // Integer registers are 32 bits. 174 Ret += "-n32"; 175 176 // The stack is 128 bit aligned on NaCl, 64 bit aligned on AAPCS and 32 bit 177 // aligned everywhere else. 178 if (TT.isOSNaCl() || ABI == ARMBaseTargetMachine::ARM_ABI_AAPCS16) 179 Ret += "-S128"; 180 else if (ABI == ARMBaseTargetMachine::ARM_ABI_AAPCS) 181 Ret += "-S64"; 182 else 183 Ret += "-S32"; 184 185 return Ret; 186 } 187 188 static Reloc::Model getEffectiveRelocModel(const Triple &TT, 189 Optional<Reloc::Model> RM) { 190 if (!RM.hasValue()) 191 // Default relocation model on Darwin is PIC. 192 return TT.isOSBinFormatMachO() ? Reloc::PIC_ : Reloc::Static; 193 194 if (*RM == Reloc::ROPI || *RM == Reloc::RWPI || *RM == Reloc::ROPI_RWPI) 195 assert(TT.isOSBinFormatELF() && 196 "ROPI/RWPI currently only supported for ELF"); 197 198 // DynamicNoPIC is only used on darwin. 199 if (*RM == Reloc::DynamicNoPIC && !TT.isOSDarwin()) 200 return Reloc::Static; 201 202 return *RM; 203 } 204 205 /// Create an ARM architecture model. 206 /// 207 ARMBaseTargetMachine::ARMBaseTargetMachine(const Target &T, const Triple &TT, 208 StringRef CPU, StringRef FS, 209 const TargetOptions &Options, 210 Optional<Reloc::Model> RM, 211 Optional<CodeModel::Model> CM, 212 CodeGenOpt::Level OL, bool isLittle) 213 : LLVMTargetMachine(T, computeDataLayout(TT, CPU, Options, isLittle), TT, 214 CPU, FS, Options, getEffectiveRelocModel(TT, RM), 215 getEffectiveCodeModel(CM, CodeModel::Small), OL), 216 TargetABI(computeTargetABI(TT, CPU, Options)), 217 TLOF(createTLOF(getTargetTriple())), isLittle(isLittle) { 218 219 // Default to triple-appropriate float ABI 220 if (Options.FloatABIType == FloatABI::Default) { 221 if (isTargetHardFloat()) 222 this->Options.FloatABIType = FloatABI::Hard; 223 else 224 this->Options.FloatABIType = FloatABI::Soft; 225 } 226 227 // Default to triple-appropriate EABI 228 if (Options.EABIVersion == EABI::Default || 229 Options.EABIVersion == EABI::Unknown) { 230 // musl is compatible with glibc with regard to EABI version 231 if ((TargetTriple.getEnvironment() == Triple::GNUEABI || 232 TargetTriple.getEnvironment() == Triple::GNUEABIHF || 233 TargetTriple.getEnvironment() == Triple::MuslEABI || 234 TargetTriple.getEnvironment() == Triple::MuslEABIHF) && 235 !(TargetTriple.isOSWindows() || TargetTriple.isOSDarwin())) 236 this->Options.EABIVersion = EABI::GNU; 237 else 238 this->Options.EABIVersion = EABI::EABI5; 239 } 240 241 if (TT.isOSBinFormatMachO()) { 242 this->Options.TrapUnreachable = true; 243 this->Options.NoTrapAfterNoreturn = true; 244 } 245 246 // ARM supports the debug entry values. 247 setSupportsDebugEntryValues(true); 248 249 initAsmInfo(); 250 } 251 252 ARMBaseTargetMachine::~ARMBaseTargetMachine() = default; 253 254 const ARMSubtarget * 255 ARMBaseTargetMachine::getSubtargetImpl(const Function &F) const { 256 Attribute CPUAttr = F.getFnAttribute("target-cpu"); 257 Attribute FSAttr = F.getFnAttribute("target-features"); 258 259 std::string CPU = !CPUAttr.hasAttribute(Attribute::None) 260 ? CPUAttr.getValueAsString().str() 261 : TargetCPU; 262 std::string FS = !FSAttr.hasAttribute(Attribute::None) 263 ? FSAttr.getValueAsString().str() 264 : TargetFS; 265 266 // FIXME: This is related to the code below to reset the target options, 267 // we need to know whether or not the soft float flag is set on the 268 // function before we can generate a subtarget. We also need to use 269 // it as a key for the subtarget since that can be the only difference 270 // between two functions. 271 bool SoftFloat = 272 F.getFnAttribute("use-soft-float").getValueAsString() == "true"; 273 // If the soft float attribute is set on the function turn on the soft float 274 // subtarget feature. 275 if (SoftFloat) 276 FS += FS.empty() ? "+soft-float" : ",+soft-float"; 277 278 // Use the optminsize to identify the subtarget, but don't use it in the 279 // feature string. 280 std::string Key = CPU + FS; 281 if (F.hasMinSize()) 282 Key += "+minsize"; 283 284 auto &I = SubtargetMap[Key]; 285 if (!I) { 286 // This needs to be done before we create a new subtarget since any 287 // creation will depend on the TM and the code generation flags on the 288 // function that reside in TargetOptions. 289 resetTargetOptions(F); 290 I = std::make_unique<ARMSubtarget>(TargetTriple, CPU, FS, *this, isLittle, 291 F.hasMinSize()); 292 293 if (!I->isThumb() && !I->hasARMOps()) 294 F.getContext().emitError("Function '" + F.getName() + "' uses ARM " 295 "instructions, but the target does not support ARM mode execution."); 296 } 297 298 return I.get(); 299 } 300 301 TargetTransformInfo 302 ARMBaseTargetMachine::getTargetTransformInfo(const Function &F) { 303 return TargetTransformInfo(ARMTTIImpl(this, F)); 304 } 305 306 ARMLETargetMachine::ARMLETargetMachine(const Target &T, const Triple &TT, 307 StringRef CPU, StringRef FS, 308 const TargetOptions &Options, 309 Optional<Reloc::Model> RM, 310 Optional<CodeModel::Model> CM, 311 CodeGenOpt::Level OL, bool JIT) 312 : ARMBaseTargetMachine(T, TT, CPU, FS, Options, RM, CM, OL, true) {} 313 314 ARMBETargetMachine::ARMBETargetMachine(const Target &T, const Triple &TT, 315 StringRef CPU, StringRef FS, 316 const TargetOptions &Options, 317 Optional<Reloc::Model> RM, 318 Optional<CodeModel::Model> CM, 319 CodeGenOpt::Level OL, bool JIT) 320 : ARMBaseTargetMachine(T, TT, CPU, FS, Options, RM, CM, OL, false) {} 321 322 namespace { 323 324 /// ARM Code Generator Pass Configuration Options. 325 class ARMPassConfig : public TargetPassConfig { 326 public: 327 ARMPassConfig(ARMBaseTargetMachine &TM, PassManagerBase &PM) 328 : TargetPassConfig(TM, PM) {} 329 330 ARMBaseTargetMachine &getARMTargetMachine() const { 331 return getTM<ARMBaseTargetMachine>(); 332 } 333 334 ScheduleDAGInstrs * 335 createMachineScheduler(MachineSchedContext *C) const override { 336 ScheduleDAGMILive *DAG = createGenericSchedLive(C); 337 // add DAG Mutations here. 338 const ARMSubtarget &ST = C->MF->getSubtarget<ARMSubtarget>(); 339 if (ST.hasFusion()) 340 DAG->addMutation(createARMMacroFusionDAGMutation()); 341 return DAG; 342 } 343 344 ScheduleDAGInstrs * 345 createPostMachineScheduler(MachineSchedContext *C) const override { 346 ScheduleDAGMI *DAG = createGenericSchedPostRA(C); 347 // add DAG Mutations here. 348 const ARMSubtarget &ST = C->MF->getSubtarget<ARMSubtarget>(); 349 if (ST.hasFusion()) 350 DAG->addMutation(createARMMacroFusionDAGMutation()); 351 return DAG; 352 } 353 354 void addIRPasses() override; 355 void addCodeGenPrepare() override; 356 bool addPreISel() override; 357 bool addInstSelector() override; 358 bool addIRTranslator() override; 359 bool addLegalizeMachineIR() override; 360 bool addRegBankSelect() override; 361 bool addGlobalInstructionSelect() override; 362 void addPreRegAlloc() override; 363 void addPreSched2() override; 364 void addPreEmitPass() override; 365 366 std::unique_ptr<CSEConfigBase> getCSEConfig() const override; 367 }; 368 369 class ARMExecutionDomainFix : public ExecutionDomainFix { 370 public: 371 static char ID; 372 ARMExecutionDomainFix() : ExecutionDomainFix(ID, ARM::DPRRegClass) {} 373 StringRef getPassName() const override { 374 return "ARM Execution Domain Fix"; 375 } 376 }; 377 char ARMExecutionDomainFix::ID; 378 379 } // end anonymous namespace 380 381 INITIALIZE_PASS_BEGIN(ARMExecutionDomainFix, "arm-execution-domain-fix", 382 "ARM Execution Domain Fix", false, false) 383 INITIALIZE_PASS_DEPENDENCY(ReachingDefAnalysis) 384 INITIALIZE_PASS_END(ARMExecutionDomainFix, "arm-execution-domain-fix", 385 "ARM Execution Domain Fix", false, false) 386 387 TargetPassConfig *ARMBaseTargetMachine::createPassConfig(PassManagerBase &PM) { 388 return new ARMPassConfig(*this, PM); 389 } 390 391 std::unique_ptr<CSEConfigBase> ARMPassConfig::getCSEConfig() const { 392 return getStandardCSEConfigForOpt(TM->getOptLevel()); 393 } 394 395 void ARMPassConfig::addIRPasses() { 396 if (TM->Options.ThreadModel == ThreadModel::Single) 397 addPass(createLowerAtomicPass()); 398 else 399 addPass(createAtomicExpandPass()); 400 401 // Cmpxchg instructions are often used with a subsequent comparison to 402 // determine whether it succeeded. We can exploit existing control-flow in 403 // ldrex/strex loops to simplify this, but it needs tidying up. 404 if (TM->getOptLevel() != CodeGenOpt::None && EnableAtomicTidy) 405 addPass(createCFGSimplificationPass( 406 1, false, false, true, true, [this](const Function &F) { 407 const auto &ST = this->TM->getSubtarget<ARMSubtarget>(F); 408 return ST.hasAnyDataBarrier() && !ST.isThumb1Only(); 409 })); 410 411 addPass(createMVEGatherScatterLoweringPass()); 412 413 TargetPassConfig::addIRPasses(); 414 415 // Run the parallel DSP pass. 416 if (getOptLevel() == CodeGenOpt::Aggressive) 417 addPass(createARMParallelDSPPass()); 418 419 // Match interleaved memory accesses to ldN/stN intrinsics. 420 if (TM->getOptLevel() != CodeGenOpt::None) 421 addPass(createInterleavedAccessPass()); 422 423 // Add Control Flow Guard checks. 424 if (TM->getTargetTriple().isOSWindows()) 425 addPass(createCFGuardCheckPass()); 426 } 427 428 void ARMPassConfig::addCodeGenPrepare() { 429 if (getOptLevel() != CodeGenOpt::None) 430 addPass(createTypePromotionPass()); 431 TargetPassConfig::addCodeGenPrepare(); 432 } 433 434 bool ARMPassConfig::addPreISel() { 435 if ((TM->getOptLevel() != CodeGenOpt::None && 436 EnableGlobalMerge == cl::BOU_UNSET) || 437 EnableGlobalMerge == cl::BOU_TRUE) { 438 // FIXME: This is using the thumb1 only constant value for 439 // maximal global offset for merging globals. We may want 440 // to look into using the old value for non-thumb1 code of 441 // 4095 based on the TargetMachine, but this starts to become 442 // tricky when doing code gen per function. 443 bool OnlyOptimizeForSize = (TM->getOptLevel() < CodeGenOpt::Aggressive) && 444 (EnableGlobalMerge == cl::BOU_UNSET); 445 // Merging of extern globals is enabled by default on non-Mach-O as we 446 // expect it to be generally either beneficial or harmless. On Mach-O it 447 // is disabled as we emit the .subsections_via_symbols directive which 448 // means that merging extern globals is not safe. 449 bool MergeExternalByDefault = !TM->getTargetTriple().isOSBinFormatMachO(); 450 addPass(createGlobalMergePass(TM, 127, OnlyOptimizeForSize, 451 MergeExternalByDefault)); 452 } 453 454 if (TM->getOptLevel() != CodeGenOpt::None) { 455 addPass(createHardwareLoopsPass()); 456 addPass(createMVETailPredicationPass()); 457 } 458 459 return false; 460 } 461 462 bool ARMPassConfig::addInstSelector() { 463 addPass(createARMISelDag(getARMTargetMachine(), getOptLevel())); 464 return false; 465 } 466 467 bool ARMPassConfig::addIRTranslator() { 468 addPass(new IRTranslator()); 469 return false; 470 } 471 472 bool ARMPassConfig::addLegalizeMachineIR() { 473 addPass(new Legalizer()); 474 return false; 475 } 476 477 bool ARMPassConfig::addRegBankSelect() { 478 addPass(new RegBankSelect()); 479 return false; 480 } 481 482 bool ARMPassConfig::addGlobalInstructionSelect() { 483 addPass(new InstructionSelect()); 484 return false; 485 } 486 487 void ARMPassConfig::addPreRegAlloc() { 488 if (getOptLevel() != CodeGenOpt::None) { 489 addPass(createMLxExpansionPass()); 490 491 if (EnableARMLoadStoreOpt) 492 addPass(createARMLoadStoreOptimizationPass(/* pre-register alloc */ true)); 493 494 if (!DisableA15SDOptimization) 495 addPass(createA15SDOptimizerPass()); 496 } 497 } 498 499 void ARMPassConfig::addPreSched2() { 500 if (getOptLevel() != CodeGenOpt::None) { 501 if (EnableARMLoadStoreOpt) 502 addPass(createARMLoadStoreOptimizationPass()); 503 504 addPass(new ARMExecutionDomainFix()); 505 addPass(createBreakFalseDeps()); 506 } 507 508 // Expand some pseudo instructions into multiple instructions to allow 509 // proper scheduling. 510 addPass(createARMExpandPseudoPass()); 511 512 if (getOptLevel() != CodeGenOpt::None) { 513 // in v8, IfConversion depends on Thumb instruction widths 514 addPass(createThumb2SizeReductionPass([this](const Function &F) { 515 return this->TM->getSubtarget<ARMSubtarget>(F).restrictIT(); 516 })); 517 518 addPass(createIfConverter([](const MachineFunction &MF) { 519 return !MF.getSubtarget<ARMSubtarget>().isThumb1Only(); 520 })); 521 } 522 addPass(createMVEVPTBlockPass()); 523 addPass(createThumb2ITBlockPass()); 524 525 // Add both scheduling passes to give the subtarget an opportunity to pick 526 // between them. 527 if (getOptLevel() != CodeGenOpt::None) { 528 addPass(&PostMachineSchedulerID); 529 addPass(&PostRASchedulerID); 530 } 531 } 532 533 void ARMPassConfig::addPreEmitPass() { 534 addPass(createThumb2SizeReductionPass()); 535 536 // Constant island pass work on unbundled instructions. 537 addPass(createUnpackMachineBundles([](const MachineFunction &MF) { 538 return MF.getSubtarget<ARMSubtarget>().isThumb2(); 539 })); 540 541 // Don't optimize barriers at -O0. 542 if (getOptLevel() != CodeGenOpt::None) 543 addPass(createARMOptimizeBarriersPass()); 544 545 addPass(createARMConstantIslandPass()); 546 addPass(createARMLowOverheadLoopsPass()); 547 548 // Identify valid longjmp targets for Windows Control Flow Guard. 549 if (TM->getTargetTriple().isOSWindows()) 550 addPass(createCFGuardLongjmpPass()); 551 } 552