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