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 initializeMVEVPTOptimisationsPass(Registry); 100 initializeMVETailPredicationPass(Registry); 101 initializeARMLowOverheadLoopsPass(Registry); 102 initializeMVEGatherScatterLoweringPass(Registry); 103 } 104 105 static std::unique_ptr<TargetLoweringObjectFile> createTLOF(const Triple &TT) { 106 if (TT.isOSBinFormatMachO()) 107 return std::make_unique<TargetLoweringObjectFileMachO>(); 108 if (TT.isOSWindows()) 109 return std::make_unique<TargetLoweringObjectFileCOFF>(); 110 return std::make_unique<ARMElfTargetObjectFile>(); 111 } 112 113 static ARMBaseTargetMachine::ARMABI 114 computeTargetABI(const Triple &TT, StringRef CPU, 115 const TargetOptions &Options) { 116 StringRef ABIName = Options.MCOptions.getABIName(); 117 118 if (ABIName.empty()) 119 ABIName = ARM::computeDefaultTargetABI(TT, CPU); 120 121 if (ABIName == "aapcs16") 122 return ARMBaseTargetMachine::ARM_ABI_AAPCS16; 123 else if (ABIName.startswith("aapcs")) 124 return ARMBaseTargetMachine::ARM_ABI_AAPCS; 125 else if (ABIName.startswith("apcs")) 126 return ARMBaseTargetMachine::ARM_ABI_APCS; 127 128 llvm_unreachable("Unhandled/unknown ABI Name!"); 129 return ARMBaseTargetMachine::ARM_ABI_UNKNOWN; 130 } 131 132 static std::string computeDataLayout(const Triple &TT, StringRef CPU, 133 const TargetOptions &Options, 134 bool isLittle) { 135 auto ABI = computeTargetABI(TT, CPU, Options); 136 std::string Ret; 137 138 if (isLittle) 139 // Little endian. 140 Ret += "e"; 141 else 142 // Big endian. 143 Ret += "E"; 144 145 Ret += DataLayout::getManglingComponent(TT); 146 147 // Pointers are 32 bits and aligned to 32 bits. 148 Ret += "-p:32:32"; 149 150 // Function pointers are aligned to 8 bits (because the LSB stores the 151 // ARM/Thumb state). 152 Ret += "-Fi8"; 153 154 // ABIs other than APCS have 64 bit integers with natural alignment. 155 if (ABI != ARMBaseTargetMachine::ARM_ABI_APCS) 156 Ret += "-i64:64"; 157 158 // We have 64 bits floats. The APCS ABI requires them to be aligned to 32 159 // bits, others to 64 bits. We always try to align to 64 bits. 160 if (ABI == ARMBaseTargetMachine::ARM_ABI_APCS) 161 Ret += "-f64:32:64"; 162 163 // We have 128 and 64 bit vectors. The APCS ABI aligns them to 32 bits, others 164 // to 64. We always ty to give them natural alignment. 165 if (ABI == ARMBaseTargetMachine::ARM_ABI_APCS) 166 Ret += "-v64:32:64-v128:32:128"; 167 else if (ABI != ARMBaseTargetMachine::ARM_ABI_AAPCS16) 168 Ret += "-v128:64:128"; 169 170 // Try to align aggregates to 32 bits (the default is 64 bits, which has no 171 // particular hardware support on 32-bit ARM). 172 Ret += "-a:0:32"; 173 174 // Integer registers are 32 bits. 175 Ret += "-n32"; 176 177 // The stack is 128 bit aligned on NaCl, 64 bit aligned on AAPCS and 32 bit 178 // aligned everywhere else. 179 if (TT.isOSNaCl() || ABI == ARMBaseTargetMachine::ARM_ABI_AAPCS16) 180 Ret += "-S128"; 181 else if (ABI == ARMBaseTargetMachine::ARM_ABI_AAPCS) 182 Ret += "-S64"; 183 else 184 Ret += "-S32"; 185 186 return Ret; 187 } 188 189 static Reloc::Model getEffectiveRelocModel(const Triple &TT, 190 Optional<Reloc::Model> RM) { 191 if (!RM.hasValue()) 192 // Default relocation model on Darwin is PIC. 193 return TT.isOSBinFormatMachO() ? Reloc::PIC_ : Reloc::Static; 194 195 if (*RM == Reloc::ROPI || *RM == Reloc::RWPI || *RM == Reloc::ROPI_RWPI) 196 assert(TT.isOSBinFormatELF() && 197 "ROPI/RWPI currently only supported for ELF"); 198 199 // DynamicNoPIC is only used on darwin. 200 if (*RM == Reloc::DynamicNoPIC && !TT.isOSDarwin()) 201 return Reloc::Static; 202 203 return *RM; 204 } 205 206 /// Create an ARM architecture model. 207 /// 208 ARMBaseTargetMachine::ARMBaseTargetMachine(const Target &T, const Triple &TT, 209 StringRef CPU, StringRef FS, 210 const TargetOptions &Options, 211 Optional<Reloc::Model> RM, 212 Optional<CodeModel::Model> CM, 213 CodeGenOpt::Level OL, bool isLittle) 214 : LLVMTargetMachine(T, computeDataLayout(TT, CPU, Options, isLittle), TT, 215 CPU, FS, Options, getEffectiveRelocModel(TT, RM), 216 getEffectiveCodeModel(CM, CodeModel::Small), OL), 217 TargetABI(computeTargetABI(TT, CPU, Options)), 218 TLOF(createTLOF(getTargetTriple())), isLittle(isLittle) { 219 220 // Default to triple-appropriate float ABI 221 if (Options.FloatABIType == FloatABI::Default) { 222 if (isTargetHardFloat()) 223 this->Options.FloatABIType = FloatABI::Hard; 224 else 225 this->Options.FloatABIType = FloatABI::Soft; 226 } 227 228 // Default to triple-appropriate EABI 229 if (Options.EABIVersion == EABI::Default || 230 Options.EABIVersion == EABI::Unknown) { 231 // musl is compatible with glibc with regard to EABI version 232 if ((TargetTriple.getEnvironment() == Triple::GNUEABI || 233 TargetTriple.getEnvironment() == Triple::GNUEABIHF || 234 TargetTriple.getEnvironment() == Triple::MuslEABI || 235 TargetTriple.getEnvironment() == Triple::MuslEABIHF) && 236 !(TargetTriple.isOSWindows() || TargetTriple.isOSDarwin())) 237 this->Options.EABIVersion = EABI::GNU; 238 else 239 this->Options.EABIVersion = EABI::EABI5; 240 } 241 242 if (TT.isOSBinFormatMachO()) { 243 this->Options.TrapUnreachable = true; 244 this->Options.NoTrapAfterNoreturn = true; 245 } 246 247 // ARM supports the debug entry values. 248 setSupportsDebugEntryValues(true); 249 250 initAsmInfo(); 251 } 252 253 ARMBaseTargetMachine::~ARMBaseTargetMachine() = default; 254 255 const ARMSubtarget * 256 ARMBaseTargetMachine::getSubtargetImpl(const Function &F) const { 257 Attribute CPUAttr = F.getFnAttribute("target-cpu"); 258 Attribute FSAttr = F.getFnAttribute("target-features"); 259 260 std::string CPU = !CPUAttr.hasAttribute(Attribute::None) 261 ? CPUAttr.getValueAsString().str() 262 : TargetCPU; 263 std::string FS = !FSAttr.hasAttribute(Attribute::None) 264 ? FSAttr.getValueAsString().str() 265 : TargetFS; 266 267 // FIXME: This is related to the code below to reset the target options, 268 // we need to know whether or not the soft float flag is set on the 269 // function before we can generate a subtarget. We also need to use 270 // it as a key for the subtarget since that can be the only difference 271 // between two functions. 272 bool SoftFloat = 273 F.getFnAttribute("use-soft-float").getValueAsString() == "true"; 274 // If the soft float attribute is set on the function turn on the soft float 275 // subtarget feature. 276 if (SoftFloat) 277 FS += FS.empty() ? "+soft-float" : ",+soft-float"; 278 279 // Use the optminsize to identify the subtarget, but don't use it in the 280 // feature string. 281 std::string Key = CPU + FS; 282 if (F.hasMinSize()) 283 Key += "+minsize"; 284 285 auto &I = SubtargetMap[Key]; 286 if (!I) { 287 // This needs to be done before we create a new subtarget since any 288 // creation will depend on the TM and the code generation flags on the 289 // function that reside in TargetOptions. 290 resetTargetOptions(F); 291 I = std::make_unique<ARMSubtarget>(TargetTriple, CPU, FS, *this, isLittle, 292 F.hasMinSize()); 293 294 if (!I->isThumb() && !I->hasARMOps()) 295 F.getContext().emitError("Function '" + F.getName() + "' uses ARM " 296 "instructions, but the target does not support ARM mode execution."); 297 } 298 299 return I.get(); 300 } 301 302 TargetTransformInfo 303 ARMBaseTargetMachine::getTargetTransformInfo(const Function &F) { 304 return TargetTransformInfo(ARMTTIImpl(this, F)); 305 } 306 307 ARMLETargetMachine::ARMLETargetMachine(const Target &T, const Triple &TT, 308 StringRef CPU, StringRef FS, 309 const TargetOptions &Options, 310 Optional<Reloc::Model> RM, 311 Optional<CodeModel::Model> CM, 312 CodeGenOpt::Level OL, bool JIT) 313 : ARMBaseTargetMachine(T, TT, CPU, FS, Options, RM, CM, OL, true) {} 314 315 ARMBETargetMachine::ARMBETargetMachine(const Target &T, const Triple &TT, 316 StringRef CPU, StringRef FS, 317 const TargetOptions &Options, 318 Optional<Reloc::Model> RM, 319 Optional<CodeModel::Model> CM, 320 CodeGenOpt::Level OL, bool JIT) 321 : ARMBaseTargetMachine(T, TT, CPU, FS, Options, RM, CM, OL, false) {} 322 323 namespace { 324 325 /// ARM Code Generator Pass Configuration Options. 326 class ARMPassConfig : public TargetPassConfig { 327 public: 328 ARMPassConfig(ARMBaseTargetMachine &TM, PassManagerBase &PM) 329 : TargetPassConfig(TM, PM) {} 330 331 ARMBaseTargetMachine &getARMTargetMachine() const { 332 return getTM<ARMBaseTargetMachine>(); 333 } 334 335 ScheduleDAGInstrs * 336 createMachineScheduler(MachineSchedContext *C) const override { 337 ScheduleDAGMILive *DAG = createGenericSchedLive(C); 338 // add DAG Mutations here. 339 const ARMSubtarget &ST = C->MF->getSubtarget<ARMSubtarget>(); 340 if (ST.hasFusion()) 341 DAG->addMutation(createARMMacroFusionDAGMutation()); 342 return DAG; 343 } 344 345 ScheduleDAGInstrs * 346 createPostMachineScheduler(MachineSchedContext *C) const override { 347 ScheduleDAGMI *DAG = createGenericSchedPostRA(C); 348 // add DAG Mutations here. 349 const ARMSubtarget &ST = C->MF->getSubtarget<ARMSubtarget>(); 350 if (ST.hasFusion()) 351 DAG->addMutation(createARMMacroFusionDAGMutation()); 352 return DAG; 353 } 354 355 void addIRPasses() override; 356 void addCodeGenPrepare() override; 357 bool addPreISel() override; 358 bool addInstSelector() override; 359 bool addIRTranslator() override; 360 bool addLegalizeMachineIR() override; 361 bool addRegBankSelect() override; 362 bool addGlobalInstructionSelect() override; 363 void addPreRegAlloc() override; 364 void addPreSched2() override; 365 void addPreEmitPass() override; 366 void addPreEmitPass2() override; 367 368 std::unique_ptr<CSEConfigBase> getCSEConfig() const override; 369 }; 370 371 class ARMExecutionDomainFix : public ExecutionDomainFix { 372 public: 373 static char ID; 374 ARMExecutionDomainFix() : ExecutionDomainFix(ID, ARM::DPRRegClass) {} 375 StringRef getPassName() const override { 376 return "ARM Execution Domain Fix"; 377 } 378 }; 379 char ARMExecutionDomainFix::ID; 380 381 } // end anonymous namespace 382 383 INITIALIZE_PASS_BEGIN(ARMExecutionDomainFix, "arm-execution-domain-fix", 384 "ARM Execution Domain Fix", false, false) 385 INITIALIZE_PASS_DEPENDENCY(ReachingDefAnalysis) 386 INITIALIZE_PASS_END(ARMExecutionDomainFix, "arm-execution-domain-fix", 387 "ARM Execution Domain Fix", false, false) 388 389 TargetPassConfig *ARMBaseTargetMachine::createPassConfig(PassManagerBase &PM) { 390 return new ARMPassConfig(*this, PM); 391 } 392 393 std::unique_ptr<CSEConfigBase> ARMPassConfig::getCSEConfig() const { 394 return getStandardCSEConfigForOpt(TM->getOptLevel()); 395 } 396 397 void ARMPassConfig::addIRPasses() { 398 if (TM->Options.ThreadModel == ThreadModel::Single) 399 addPass(createLowerAtomicPass()); 400 else 401 addPass(createAtomicExpandPass()); 402 403 // Cmpxchg instructions are often used with a subsequent comparison to 404 // determine whether it succeeded. We can exploit existing control-flow in 405 // ldrex/strex loops to simplify this, but it needs tidying up. 406 if (TM->getOptLevel() != CodeGenOpt::None && EnableAtomicTidy) 407 addPass(createCFGSimplificationPass( 408 1, false, false, true, true, [this](const Function &F) { 409 const auto &ST = this->TM->getSubtarget<ARMSubtarget>(F); 410 return ST.hasAnyDataBarrier() && !ST.isThumb1Only(); 411 })); 412 413 addPass(createMVEGatherScatterLoweringPass()); 414 415 TargetPassConfig::addIRPasses(); 416 417 // Run the parallel DSP pass. 418 if (getOptLevel() == CodeGenOpt::Aggressive) 419 addPass(createARMParallelDSPPass()); 420 421 // Match interleaved memory accesses to ldN/stN intrinsics. 422 if (TM->getOptLevel() != CodeGenOpt::None) 423 addPass(createInterleavedAccessPass()); 424 425 // Add Control Flow Guard checks. 426 if (TM->getTargetTriple().isOSWindows()) 427 addPass(createCFGuardCheckPass()); 428 } 429 430 void ARMPassConfig::addCodeGenPrepare() { 431 if (getOptLevel() != CodeGenOpt::None) 432 addPass(createTypePromotionPass()); 433 TargetPassConfig::addCodeGenPrepare(); 434 } 435 436 bool ARMPassConfig::addPreISel() { 437 if ((TM->getOptLevel() != CodeGenOpt::None && 438 EnableGlobalMerge == cl::BOU_UNSET) || 439 EnableGlobalMerge == cl::BOU_TRUE) { 440 // FIXME: This is using the thumb1 only constant value for 441 // maximal global offset for merging globals. We may want 442 // to look into using the old value for non-thumb1 code of 443 // 4095 based on the TargetMachine, but this starts to become 444 // tricky when doing code gen per function. 445 bool OnlyOptimizeForSize = (TM->getOptLevel() < CodeGenOpt::Aggressive) && 446 (EnableGlobalMerge == cl::BOU_UNSET); 447 // Merging of extern globals is enabled by default on non-Mach-O as we 448 // expect it to be generally either beneficial or harmless. On Mach-O it 449 // is disabled as we emit the .subsections_via_symbols directive which 450 // means that merging extern globals is not safe. 451 bool MergeExternalByDefault = !TM->getTargetTriple().isOSBinFormatMachO(); 452 addPass(createGlobalMergePass(TM, 127, OnlyOptimizeForSize, 453 MergeExternalByDefault)); 454 } 455 456 if (TM->getOptLevel() != CodeGenOpt::None) { 457 addPass(createHardwareLoopsPass()); 458 addPass(createMVETailPredicationPass()); 459 } 460 461 return false; 462 } 463 464 bool ARMPassConfig::addInstSelector() { 465 addPass(createARMISelDag(getARMTargetMachine(), getOptLevel())); 466 return false; 467 } 468 469 bool ARMPassConfig::addIRTranslator() { 470 addPass(new IRTranslator()); 471 return false; 472 } 473 474 bool ARMPassConfig::addLegalizeMachineIR() { 475 addPass(new Legalizer()); 476 return false; 477 } 478 479 bool ARMPassConfig::addRegBankSelect() { 480 addPass(new RegBankSelect()); 481 return false; 482 } 483 484 bool ARMPassConfig::addGlobalInstructionSelect() { 485 addPass(new InstructionSelect()); 486 return false; 487 } 488 489 void ARMPassConfig::addPreRegAlloc() { 490 if (getOptLevel() != CodeGenOpt::None) { 491 addPass(createMVEVPTOptimisationsPass()); 492 493 addPass(createMLxExpansionPass()); 494 495 if (EnableARMLoadStoreOpt) 496 addPass(createARMLoadStoreOptimizationPass(/* pre-register alloc */ true)); 497 498 if (!DisableA15SDOptimization) 499 addPass(createA15SDOptimizerPass()); 500 } 501 } 502 503 void ARMPassConfig::addPreSched2() { 504 if (getOptLevel() != CodeGenOpt::None) { 505 if (EnableARMLoadStoreOpt) 506 addPass(createARMLoadStoreOptimizationPass()); 507 508 addPass(new ARMExecutionDomainFix()); 509 addPass(createBreakFalseDeps()); 510 } 511 512 // Expand some pseudo instructions into multiple instructions to allow 513 // proper scheduling. 514 addPass(createARMExpandPseudoPass()); 515 516 if (getOptLevel() != CodeGenOpt::None) { 517 // in v8, IfConversion depends on Thumb instruction widths 518 addPass(createThumb2SizeReductionPass([this](const Function &F) { 519 return this->TM->getSubtarget<ARMSubtarget>(F).restrictIT(); 520 })); 521 522 addPass(createIfConverter([](const MachineFunction &MF) { 523 return !MF.getSubtarget<ARMSubtarget>().isThumb1Only(); 524 })); 525 } 526 addPass(createMVEVPTBlockPass()); 527 addPass(createThumb2ITBlockPass()); 528 529 // Add both scheduling passes to give the subtarget an opportunity to pick 530 // between them. 531 if (getOptLevel() != CodeGenOpt::None) { 532 addPass(&PostMachineSchedulerID); 533 addPass(&PostRASchedulerID); 534 } 535 } 536 537 void ARMPassConfig::addPreEmitPass() { 538 addPass(createThumb2SizeReductionPass()); 539 540 // Constant island pass work on unbundled instructions. 541 addPass(createUnpackMachineBundles([](const MachineFunction &MF) { 542 return MF.getSubtarget<ARMSubtarget>().isThumb2(); 543 })); 544 545 // Don't optimize barriers at -O0. 546 if (getOptLevel() != CodeGenOpt::None) 547 addPass(createARMOptimizeBarriersPass()); 548 } 549 550 void ARMPassConfig::addPreEmitPass2() { 551 addPass(createARMConstantIslandPass()); 552 addPass(createARMLowOverheadLoopsPass()); 553 554 // Identify valid longjmp targets for Windows Control Flow Guard. 555 if (TM->getTargetTriple().isOSWindows()) 556 addPass(createCFGuardLongjmpPass()); 557 } 558