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 "ARM.h" 14 #include "ARMSubtarget.h" 15 #include "ARMMacroFusion.h" 16 #include "ARMTargetMachine.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/ExecutionDepsFix.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 initializeARMExecutionDepsFixPass(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 initializeARMExecutionDepsFixPass(Registry); 94 initializeARMExpandPseudoPass(Registry); 95 } 96 97 static std::unique_ptr<TargetLoweringObjectFile> createTLOF(const Triple &TT) { 98 if (TT.isOSBinFormatMachO()) 99 return llvm::make_unique<TargetLoweringObjectFileMachO>(); 100 if (TT.isOSWindows()) 101 return llvm::make_unique<TargetLoweringObjectFileCOFF>(); 102 return llvm::make_unique<ARMElfTargetObjectFile>(); 103 } 104 105 static ARMBaseTargetMachine::ARMABI 106 computeTargetABI(const Triple &TT, StringRef CPU, 107 const TargetOptions &Options) { 108 StringRef ABIName = Options.MCOptions.getABIName(); 109 110 if (ABIName.empty()) 111 ABIName = ARM::computeDefaultTargetABI(TT, CPU); 112 113 if (ABIName == "aapcs16") 114 return ARMBaseTargetMachine::ARM_ABI_AAPCS16; 115 else if (ABIName.startswith("aapcs")) 116 return ARMBaseTargetMachine::ARM_ABI_AAPCS; 117 else if (ABIName.startswith("apcs")) 118 return ARMBaseTargetMachine::ARM_ABI_APCS; 119 120 llvm_unreachable("Unhandled/unknown ABI Name!"); 121 return ARMBaseTargetMachine::ARM_ABI_UNKNOWN; 122 } 123 124 static std::string computeDataLayout(const Triple &TT, StringRef CPU, 125 const TargetOptions &Options, 126 bool isLittle) { 127 auto ABI = computeTargetABI(TT, CPU, Options); 128 std::string Ret; 129 130 if (isLittle) 131 // Little endian. 132 Ret += "e"; 133 else 134 // Big endian. 135 Ret += "E"; 136 137 Ret += DataLayout::getManglingComponent(TT); 138 139 // Pointers are 32 bits and aligned to 32 bits. 140 Ret += "-p:32:32"; 141 142 // ABIs other than APCS have 64 bit integers with natural alignment. 143 if (ABI != ARMBaseTargetMachine::ARM_ABI_APCS) 144 Ret += "-i64:64"; 145 146 // We have 64 bits floats. The APCS ABI requires them to be aligned to 32 147 // bits, others to 64 bits. We always try to align to 64 bits. 148 if (ABI == ARMBaseTargetMachine::ARM_ABI_APCS) 149 Ret += "-f64:32:64"; 150 151 // We have 128 and 64 bit vectors. The APCS ABI aligns them to 32 bits, others 152 // to 64. We always ty to give them natural alignment. 153 if (ABI == ARMBaseTargetMachine::ARM_ABI_APCS) 154 Ret += "-v64:32:64-v128:32:128"; 155 else if (ABI != ARMBaseTargetMachine::ARM_ABI_AAPCS16) 156 Ret += "-v128:64:128"; 157 158 // Try to align aggregates to 32 bits (the default is 64 bits, which has no 159 // particular hardware support on 32-bit ARM). 160 Ret += "-a:0:32"; 161 162 // Integer registers are 32 bits. 163 Ret += "-n32"; 164 165 // The stack is 128 bit aligned on NaCl, 64 bit aligned on AAPCS and 32 bit 166 // aligned everywhere else. 167 if (TT.isOSNaCl() || ABI == ARMBaseTargetMachine::ARM_ABI_AAPCS16) 168 Ret += "-S128"; 169 else if (ABI == ARMBaseTargetMachine::ARM_ABI_AAPCS) 170 Ret += "-S64"; 171 else 172 Ret += "-S32"; 173 174 return Ret; 175 } 176 177 static Reloc::Model getEffectiveRelocModel(const Triple &TT, 178 Optional<Reloc::Model> RM) { 179 if (!RM.hasValue()) 180 // Default relocation model on Darwin is PIC. 181 return TT.isOSBinFormatMachO() ? Reloc::PIC_ : Reloc::Static; 182 183 if (*RM == Reloc::ROPI || *RM == Reloc::RWPI || *RM == Reloc::ROPI_RWPI) 184 assert(TT.isOSBinFormatELF() && 185 "ROPI/RWPI currently only supported for ELF"); 186 187 // DynamicNoPIC is only used on darwin. 188 if (*RM == Reloc::DynamicNoPIC && !TT.isOSDarwin()) 189 return Reloc::Static; 190 191 return *RM; 192 } 193 194 static CodeModel::Model getEffectiveCodeModel(Optional<CodeModel::Model> CM) { 195 if (CM) 196 return *CM; 197 return CodeModel::Small; 198 } 199 200 /// Create an ARM architecture model. 201 /// 202 ARMBaseTargetMachine::ARMBaseTargetMachine(const Target &T, const Triple &TT, 203 StringRef CPU, StringRef FS, 204 const TargetOptions &Options, 205 Optional<Reloc::Model> RM, 206 Optional<CodeModel::Model> CM, 207 CodeGenOpt::Level OL, bool isLittle) 208 : LLVMTargetMachine(T, computeDataLayout(TT, CPU, Options, isLittle), TT, 209 CPU, FS, Options, getEffectiveRelocModel(TT, RM), 210 getEffectiveCodeModel(CM), OL), 211 TargetABI(computeTargetABI(TT, CPU, Options)), 212 TLOF(createTLOF(getTargetTriple())), isLittle(isLittle) { 213 214 // Default to triple-appropriate float ABI 215 if (Options.FloatABIType == FloatABI::Default) { 216 if (TargetTriple.getEnvironment() == Triple::GNUEABIHF || 217 TargetTriple.getEnvironment() == Triple::MuslEABIHF || 218 TargetTriple.getEnvironment() == Triple::EABIHF || 219 TargetTriple.isOSWindows() || 220 TargetABI == ARMBaseTargetMachine::ARM_ABI_AAPCS16) 221 this->Options.FloatABIType = FloatABI::Hard; 222 else 223 this->Options.FloatABIType = FloatABI::Soft; 224 } 225 226 // Default to triple-appropriate EABI 227 if (Options.EABIVersion == EABI::Default || 228 Options.EABIVersion == EABI::Unknown) { 229 // musl is compatible with glibc with regard to EABI version 230 if ((TargetTriple.getEnvironment() == Triple::GNUEABI || 231 TargetTriple.getEnvironment() == Triple::GNUEABIHF || 232 TargetTriple.getEnvironment() == Triple::MuslEABI || 233 TargetTriple.getEnvironment() == Triple::MuslEABIHF) && 234 !(TargetTriple.isOSWindows() || TargetTriple.isOSDarwin())) 235 this->Options.EABIVersion = EABI::GNU; 236 else 237 this->Options.EABIVersion = EABI::EABI5; 238 } 239 240 initAsmInfo(); 241 } 242 243 ARMBaseTargetMachine::~ARMBaseTargetMachine() = default; 244 245 const ARMSubtarget * 246 ARMBaseTargetMachine::getSubtargetImpl(const Function &F) const { 247 Attribute CPUAttr = F.getFnAttribute("target-cpu"); 248 Attribute FSAttr = F.getFnAttribute("target-features"); 249 250 std::string CPU = !CPUAttr.hasAttribute(Attribute::None) 251 ? CPUAttr.getValueAsString().str() 252 : TargetCPU; 253 std::string FS = !FSAttr.hasAttribute(Attribute::None) 254 ? FSAttr.getValueAsString().str() 255 : TargetFS; 256 257 // FIXME: This is related to the code below to reset the target options, 258 // we need to know whether or not the soft float flag is set on the 259 // function before we can generate a subtarget. We also need to use 260 // it as a key for the subtarget since that can be the only difference 261 // between two functions. 262 bool SoftFloat = 263 F.getFnAttribute("use-soft-float").getValueAsString() == "true"; 264 // If the soft float attribute is set on the function turn on the soft float 265 // subtarget feature. 266 if (SoftFloat) 267 FS += FS.empty() ? "+soft-float" : ",+soft-float"; 268 269 auto &I = SubtargetMap[CPU + FS]; 270 if (!I) { 271 // This needs to be done before we create a new subtarget since any 272 // creation will depend on the TM and the code generation flags on the 273 // function that reside in TargetOptions. 274 resetTargetOptions(F); 275 I = llvm::make_unique<ARMSubtarget>(TargetTriple, CPU, FS, *this, isLittle); 276 277 if (!I->isThumb() && !I->hasARMOps()) 278 F.getContext().emitError("Function '" + F.getName() + "' uses ARM " 279 "instructions, but the target does not support ARM mode execution."); 280 } 281 282 return I.get(); 283 } 284 285 TargetIRAnalysis ARMBaseTargetMachine::getTargetIRAnalysis() { 286 return TargetIRAnalysis([this](const Function &F) { 287 return TargetTransformInfo(ARMTTIImpl(this, F)); 288 }); 289 } 290 291 ARMLETargetMachine::ARMLETargetMachine(const Target &T, const Triple &TT, 292 StringRef CPU, StringRef FS, 293 const TargetOptions &Options, 294 Optional<Reloc::Model> RM, 295 Optional<CodeModel::Model> CM, 296 CodeGenOpt::Level OL, bool JIT) 297 : ARMBaseTargetMachine(T, TT, CPU, FS, Options, RM, CM, OL, true) {} 298 299 ARMBETargetMachine::ARMBETargetMachine(const Target &T, const Triple &TT, 300 StringRef CPU, StringRef FS, 301 const TargetOptions &Options, 302 Optional<Reloc::Model> RM, 303 Optional<CodeModel::Model> CM, 304 CodeGenOpt::Level OL, bool JIT) 305 : ARMBaseTargetMachine(T, TT, CPU, FS, Options, RM, CM, OL, false) {} 306 307 namespace { 308 309 /// ARM Code Generator Pass Configuration Options. 310 class ARMPassConfig : public TargetPassConfig { 311 public: 312 ARMPassConfig(ARMBaseTargetMachine &TM, PassManagerBase &PM) 313 : TargetPassConfig(TM, PM) {} 314 315 ARMBaseTargetMachine &getARMTargetMachine() const { 316 return getTM<ARMBaseTargetMachine>(); 317 } 318 319 ScheduleDAGInstrs * 320 createMachineScheduler(MachineSchedContext *C) const override { 321 ScheduleDAGMILive *DAG = createGenericSchedLive(C); 322 // add DAG Mutations here. 323 const ARMSubtarget &ST = C->MF->getSubtarget<ARMSubtarget>(); 324 if (ST.hasFusion()) 325 DAG->addMutation(createARMMacroFusionDAGMutation()); 326 return DAG; 327 } 328 329 ScheduleDAGInstrs * 330 createPostMachineScheduler(MachineSchedContext *C) const override { 331 ScheduleDAGMI *DAG = createGenericSchedPostRA(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 void addIRPasses() override; 340 bool addPreISel() override; 341 bool addInstSelector() override; 342 bool addIRTranslator() override; 343 bool addLegalizeMachineIR() override; 344 bool addRegBankSelect() override; 345 bool addGlobalInstructionSelect() override; 346 void addPreRegAlloc() override; 347 void addPreSched2() override; 348 void addPreEmitPass() override; 349 }; 350 351 class ARMExecutionDepsFix : public ExecutionDepsFix { 352 public: 353 static char ID; 354 ARMExecutionDepsFix() : ExecutionDepsFix(ID, ARM::DPRRegClass) {} 355 StringRef getPassName() const override { 356 return "ARM Execution Dependency Fix"; 357 } 358 }; 359 char ARMExecutionDepsFix::ID; 360 361 } // end anonymous namespace 362 363 INITIALIZE_PASS(ARMExecutionDepsFix, "arm-execution-deps-fix", 364 "ARM Execution Dependency Fix", false, false) 365 366 TargetPassConfig *ARMBaseTargetMachine::createPassConfig(PassManagerBase &PM) { 367 return new ARMPassConfig(*this, PM); 368 } 369 370 void ARMPassConfig::addIRPasses() { 371 if (TM->Options.ThreadModel == ThreadModel::Single) 372 addPass(createLowerAtomicPass()); 373 else 374 addPass(createAtomicExpandPass()); 375 376 // Cmpxchg instructions are often used with a subsequent comparison to 377 // determine whether it succeeded. We can exploit existing control-flow in 378 // ldrex/strex loops to simplify this, but it needs tidying up. 379 if (TM->getOptLevel() != CodeGenOpt::None && EnableAtomicTidy) 380 addPass(createCFGSimplificationPass(-1, [this](const Function &F) { 381 const auto &ST = this->TM->getSubtarget<ARMSubtarget>(F); 382 return ST.hasAnyDataBarrier() && !ST.isThumb1Only(); 383 })); 384 385 TargetPassConfig::addIRPasses(); 386 387 // Match interleaved memory accesses to ldN/stN intrinsics. 388 if (TM->getOptLevel() != CodeGenOpt::None) 389 addPass(createInterleavedAccessPass()); 390 } 391 392 bool ARMPassConfig::addPreISel() { 393 if ((TM->getOptLevel() != CodeGenOpt::None && 394 EnableGlobalMerge == cl::BOU_UNSET) || 395 EnableGlobalMerge == cl::BOU_TRUE) { 396 // FIXME: This is using the thumb1 only constant value for 397 // maximal global offset for merging globals. We may want 398 // to look into using the old value for non-thumb1 code of 399 // 4095 based on the TargetMachine, but this starts to become 400 // tricky when doing code gen per function. 401 bool OnlyOptimizeForSize = (TM->getOptLevel() < CodeGenOpt::Aggressive) && 402 (EnableGlobalMerge == cl::BOU_UNSET); 403 // Merging of extern globals is enabled by default on non-Mach-O as we 404 // expect it to be generally either beneficial or harmless. On Mach-O it 405 // is disabled as we emit the .subsections_via_symbols directive which 406 // means that merging extern globals is not safe. 407 bool MergeExternalByDefault = !TM->getTargetTriple().isOSBinFormatMachO(); 408 addPass(createGlobalMergePass(TM, 127, OnlyOptimizeForSize, 409 MergeExternalByDefault)); 410 } 411 412 return false; 413 } 414 415 bool ARMPassConfig::addInstSelector() { 416 addPass(createARMISelDag(getARMTargetMachine(), getOptLevel())); 417 return false; 418 } 419 420 bool ARMPassConfig::addIRTranslator() { 421 addPass(new IRTranslator()); 422 return false; 423 } 424 425 bool ARMPassConfig::addLegalizeMachineIR() { 426 addPass(new Legalizer()); 427 return false; 428 } 429 430 bool ARMPassConfig::addRegBankSelect() { 431 addPass(new RegBankSelect()); 432 return false; 433 } 434 435 bool ARMPassConfig::addGlobalInstructionSelect() { 436 addPass(new InstructionSelect()); 437 return false; 438 } 439 440 void ARMPassConfig::addPreRegAlloc() { 441 if (getOptLevel() != CodeGenOpt::None) { 442 addPass(createMLxExpansionPass()); 443 444 if (EnableARMLoadStoreOpt) 445 addPass(createARMLoadStoreOptimizationPass(/* pre-register alloc */ true)); 446 447 if (!DisableA15SDOptimization) 448 addPass(createA15SDOptimizerPass()); 449 } 450 } 451 452 void ARMPassConfig::addPreSched2() { 453 if (getOptLevel() != CodeGenOpt::None) { 454 if (EnableARMLoadStoreOpt) 455 addPass(createARMLoadStoreOptimizationPass()); 456 457 addPass(new ARMExecutionDepsFix()); 458 } 459 460 // Expand some pseudo instructions into multiple instructions to allow 461 // proper scheduling. 462 addPass(createARMExpandPseudoPass()); 463 464 if (getOptLevel() != CodeGenOpt::None) { 465 // in v8, IfConversion depends on Thumb instruction widths 466 addPass(createThumb2SizeReductionPass([this](const Function &F) { 467 return this->TM->getSubtarget<ARMSubtarget>(F).restrictIT(); 468 })); 469 470 addPass(createIfConverter([](const MachineFunction &MF) { 471 return !MF.getSubtarget<ARMSubtarget>().isThumb1Only(); 472 })); 473 } 474 addPass(createThumb2ITBlockPass()); 475 } 476 477 void ARMPassConfig::addPreEmitPass() { 478 addPass(createThumb2SizeReductionPass()); 479 480 // Constant island pass work on unbundled instructions. 481 addPass(createUnpackMachineBundles([](const MachineFunction &MF) { 482 return MF.getSubtarget<ARMSubtarget>().isThumb2(); 483 })); 484 485 // Don't optimize barriers at -O0. 486 if (getOptLevel() != CodeGenOpt::None) 487 addPass(createARMOptimizeBarriersPass()); 488 489 addPass(createARMConstantIslandPass()); 490 } 491