1 //===-- X86TargetMachine.cpp - Define TargetMachine for the X86 -----------===// 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 // This file defines the X86 specific subclass of TargetMachine. 11 // 12 //===----------------------------------------------------------------------===// 13 14 #include "MCTargetDesc/X86MCTargetDesc.h" 15 #include "X86.h" 16 #include "X86CallLowering.h" 17 #include "X86LegalizerInfo.h" 18 #ifdef LLVM_BUILD_GLOBAL_ISEL 19 #include "X86RegisterBankInfo.h" 20 #endif 21 #include "X86MacroFusion.h" 22 #include "X86Subtarget.h" 23 #include "X86TargetMachine.h" 24 #include "X86TargetObjectFile.h" 25 #include "X86TargetTransformInfo.h" 26 #include "llvm/ADT/Optional.h" 27 #include "llvm/ADT/STLExtras.h" 28 #include "llvm/ADT/SmallString.h" 29 #include "llvm/ADT/StringRef.h" 30 #include "llvm/ADT/Triple.h" 31 #include "llvm/Analysis/TargetTransformInfo.h" 32 #include "llvm/CodeGen/ExecutionDepsFix.h" 33 #include "llvm/CodeGen/GlobalISel/CallLowering.h" 34 #include "llvm/CodeGen/GlobalISel/GISelAccessor.h" 35 #include "llvm/CodeGen/GlobalISel/IRTranslator.h" 36 #include "llvm/CodeGen/GlobalISel/InstructionSelect.h" 37 #include "llvm/CodeGen/GlobalISel/Legalizer.h" 38 #include "llvm/CodeGen/GlobalISel/RegBankSelect.h" 39 #include "llvm/CodeGen/MachineScheduler.h" 40 #include "llvm/CodeGen/Passes.h" 41 #include "llvm/CodeGen/TargetPassConfig.h" 42 #include "llvm/IR/Attributes.h" 43 #include "llvm/IR/DataLayout.h" 44 #include "llvm/IR/Function.h" 45 #include "llvm/Pass.h" 46 #include "llvm/Support/CodeGen.h" 47 #include "llvm/Support/CommandLine.h" 48 #include "llvm/Support/ErrorHandling.h" 49 #include "llvm/Support/TargetRegistry.h" 50 #include "llvm/Target/TargetLoweringObjectFile.h" 51 #include "llvm/Target/TargetOptions.h" 52 #include <memory> 53 #include <string> 54 55 using namespace llvm; 56 57 static cl::opt<bool> EnableMachineCombinerPass("x86-machine-combiner", 58 cl::desc("Enable the machine combiner pass"), 59 cl::init(true), cl::Hidden); 60 61 namespace llvm { 62 63 void initializeWinEHStatePassPass(PassRegistry &); 64 void initializeFixupLEAPassPass(PassRegistry &); 65 void initializeX86ExecutionDepsFixPass(PassRegistry &); 66 67 } // end namespace llvm 68 69 extern "C" void LLVMInitializeX86Target() { 70 // Register the target. 71 RegisterTargetMachine<X86TargetMachine> X(getTheX86_32Target()); 72 RegisterTargetMachine<X86TargetMachine> Y(getTheX86_64Target()); 73 74 PassRegistry &PR = *PassRegistry::getPassRegistry(); 75 initializeGlobalISel(PR); 76 initializeWinEHStatePassPass(PR); 77 initializeFixupBWInstPassPass(PR); 78 initializeEvexToVexInstPassPass(PR); 79 initializeFixupLEAPassPass(PR); 80 initializeX86ExecutionDepsFixPass(PR); 81 } 82 83 static std::unique_ptr<TargetLoweringObjectFile> createTLOF(const Triple &TT) { 84 if (TT.isOSBinFormatMachO()) { 85 if (TT.getArch() == Triple::x86_64) 86 return llvm::make_unique<X86_64MachoTargetObjectFile>(); 87 return llvm::make_unique<TargetLoweringObjectFileMachO>(); 88 } 89 90 if (TT.isOSFreeBSD()) 91 return llvm::make_unique<X86FreeBSDTargetObjectFile>(); 92 if (TT.isOSLinux() || TT.isOSNaCl() || TT.isOSIAMCU()) 93 return llvm::make_unique<X86LinuxNaClTargetObjectFile>(); 94 if (TT.isOSSolaris()) 95 return llvm::make_unique<X86SolarisTargetObjectFile>(); 96 if (TT.isOSFuchsia()) 97 return llvm::make_unique<X86FuchsiaTargetObjectFile>(); 98 if (TT.isOSBinFormatELF()) 99 return llvm::make_unique<X86ELFTargetObjectFile>(); 100 if (TT.isKnownWindowsMSVCEnvironment() || TT.isWindowsCoreCLREnvironment()) 101 return llvm::make_unique<X86WindowsTargetObjectFile>(); 102 if (TT.isOSBinFormatCOFF()) 103 return llvm::make_unique<TargetLoweringObjectFileCOFF>(); 104 llvm_unreachable("unknown subtarget type"); 105 } 106 107 static std::string computeDataLayout(const Triple &TT) { 108 // X86 is little endian 109 std::string Ret = "e"; 110 111 Ret += DataLayout::getManglingComponent(TT); 112 // X86 and x32 have 32 bit pointers. 113 if ((TT.isArch64Bit() && 114 (TT.getEnvironment() == Triple::GNUX32 || TT.isOSNaCl())) || 115 !TT.isArch64Bit()) 116 Ret += "-p:32:32"; 117 118 // Some ABIs align 64 bit integers and doubles to 64 bits, others to 32. 119 if (TT.isArch64Bit() || TT.isOSWindows() || TT.isOSNaCl()) 120 Ret += "-i64:64"; 121 else if (TT.isOSIAMCU()) 122 Ret += "-i64:32-f64:32"; 123 else 124 Ret += "-f64:32:64"; 125 126 // Some ABIs align long double to 128 bits, others to 32. 127 if (TT.isOSNaCl() || TT.isOSIAMCU()) 128 ; // No f80 129 else if (TT.isArch64Bit() || TT.isOSDarwin()) 130 Ret += "-f80:128"; 131 else 132 Ret += "-f80:32"; 133 134 if (TT.isOSIAMCU()) 135 Ret += "-f128:32"; 136 137 // The registers can hold 8, 16, 32 or, in x86-64, 64 bits. 138 if (TT.isArch64Bit()) 139 Ret += "-n8:16:32:64"; 140 else 141 Ret += "-n8:16:32"; 142 143 // The stack is aligned to 32 bits on some ABIs and 128 bits on others. 144 if ((!TT.isArch64Bit() && TT.isOSWindows()) || TT.isOSIAMCU()) 145 Ret += "-a:0:32-S32"; 146 else 147 Ret += "-S128"; 148 149 return Ret; 150 } 151 152 static Reloc::Model getEffectiveRelocModel(const Triple &TT, 153 Optional<Reloc::Model> RM) { 154 bool is64Bit = TT.getArch() == Triple::x86_64; 155 if (!RM.hasValue()) { 156 // Darwin defaults to PIC in 64 bit mode and dynamic-no-pic in 32 bit mode. 157 // Win64 requires rip-rel addressing, thus we force it to PIC. Otherwise we 158 // use static relocation model by default. 159 if (TT.isOSDarwin()) { 160 if (is64Bit) 161 return Reloc::PIC_; 162 return Reloc::DynamicNoPIC; 163 } 164 if (TT.isOSWindows() && is64Bit) 165 return Reloc::PIC_; 166 return Reloc::Static; 167 } 168 169 // ELF and X86-64 don't have a distinct DynamicNoPIC model. DynamicNoPIC 170 // is defined as a model for code which may be used in static or dynamic 171 // executables but not necessarily a shared library. On X86-32 we just 172 // compile in -static mode, in x86-64 we use PIC. 173 if (*RM == Reloc::DynamicNoPIC) { 174 if (is64Bit) 175 return Reloc::PIC_; 176 if (!TT.isOSDarwin()) 177 return Reloc::Static; 178 } 179 180 // If we are on Darwin, disallow static relocation model in X86-64 mode, since 181 // the Mach-O file format doesn't support it. 182 if (*RM == Reloc::Static && TT.isOSDarwin() && is64Bit) 183 return Reloc::PIC_; 184 185 return *RM; 186 } 187 188 /// Create an X86 target. 189 /// 190 X86TargetMachine::X86TargetMachine(const Target &T, const Triple &TT, 191 StringRef CPU, StringRef FS, 192 const TargetOptions &Options, 193 Optional<Reloc::Model> RM, 194 CodeModel::Model CM, CodeGenOpt::Level OL) 195 : LLVMTargetMachine(T, computeDataLayout(TT), TT, CPU, FS, Options, 196 getEffectiveRelocModel(TT, RM), CM, OL), 197 TLOF(createTLOF(getTargetTriple())) { 198 // Windows stack unwinder gets confused when execution flow "falls through" 199 // after a call to 'noreturn' function. 200 // To prevent that, we emit a trap for 'unreachable' IR instructions. 201 // (which on X86, happens to be the 'ud2' instruction) 202 // On PS4, the "return address" of a 'noreturn' call must still be within 203 // the calling function, and TrapUnreachable is an easy way to get that. 204 // The check here for 64-bit windows is a bit icky, but as we're unlikely 205 // to ever want to mix 32 and 64-bit windows code in a single module 206 // this should be fine. 207 if ((TT.isOSWindows() && TT.getArch() == Triple::x86_64) || TT.isPS4()) 208 this->Options.TrapUnreachable = true; 209 210 initAsmInfo(); 211 } 212 213 X86TargetMachine::~X86TargetMachine() = default; 214 215 #ifdef LLVM_BUILD_GLOBAL_ISEL 216 namespace { 217 218 struct X86GISelActualAccessor : public GISelAccessor { 219 std::unique_ptr<CallLowering> CallLoweringInfo; 220 std::unique_ptr<LegalizerInfo> Legalizer; 221 std::unique_ptr<RegisterBankInfo> RegBankInfo; 222 std::unique_ptr<InstructionSelector> InstSelector; 223 224 const CallLowering *getCallLowering() const override { 225 return CallLoweringInfo.get(); 226 } 227 228 const InstructionSelector *getInstructionSelector() const override { 229 return InstSelector.get(); 230 } 231 232 const LegalizerInfo *getLegalizerInfo() const override { 233 return Legalizer.get(); 234 } 235 236 const RegisterBankInfo *getRegBankInfo() const override { 237 return RegBankInfo.get(); 238 } 239 }; 240 241 } // end anonymous namespace 242 #endif 243 244 const X86Subtarget * 245 X86TargetMachine::getSubtargetImpl(const Function &F) const { 246 Attribute CPUAttr = F.getFnAttribute("target-cpu"); 247 Attribute FSAttr = F.getFnAttribute("target-features"); 248 249 StringRef CPU = !CPUAttr.hasAttribute(Attribute::None) 250 ? CPUAttr.getValueAsString() 251 : (StringRef)TargetCPU; 252 StringRef FS = !FSAttr.hasAttribute(Attribute::None) 253 ? FSAttr.getValueAsString() 254 : (StringRef)TargetFS; 255 256 SmallString<512> Key; 257 Key.reserve(CPU.size() + FS.size()); 258 Key += CPU; 259 Key += FS; 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 Key += FS.empty() ? "+soft-float" : ",+soft-float"; 272 273 FS = Key.substr(CPU.size()); 274 275 auto &I = SubtargetMap[Key]; 276 if (!I) { 277 // This needs to be done before we create a new subtarget since any 278 // creation will depend on the TM and the code generation flags on the 279 // function that reside in TargetOptions. 280 resetTargetOptions(F); 281 I = llvm::make_unique<X86Subtarget>(TargetTriple, CPU, FS, *this, 282 Options.StackAlignmentOverride); 283 #ifndef LLVM_BUILD_GLOBAL_ISEL 284 GISelAccessor *GISel = new GISelAccessor(); 285 #else 286 X86GISelActualAccessor *GISel = new X86GISelActualAccessor(); 287 288 GISel->CallLoweringInfo.reset(new X86CallLowering(*I->getTargetLowering())); 289 GISel->Legalizer.reset(new X86LegalizerInfo(*I, *this)); 290 291 auto *RBI = new X86RegisterBankInfo(*I->getRegisterInfo()); 292 GISel->RegBankInfo.reset(RBI); 293 GISel->InstSelector.reset(createX86InstructionSelector( 294 *this, *I, *RBI)); 295 #endif 296 I->setGISelAccessor(*GISel); 297 } 298 return I.get(); 299 } 300 301 //===----------------------------------------------------------------------===// 302 // Command line options for x86 303 //===----------------------------------------------------------------------===// 304 static cl::opt<bool> 305 UseVZeroUpper("x86-use-vzeroupper", cl::Hidden, 306 cl::desc("Minimize AVX to SSE transition penalty"), 307 cl::init(true)); 308 309 //===----------------------------------------------------------------------===// 310 // X86 TTI query. 311 //===----------------------------------------------------------------------===// 312 313 TargetIRAnalysis X86TargetMachine::getTargetIRAnalysis() { 314 return TargetIRAnalysis([this](const Function &F) { 315 return TargetTransformInfo(X86TTIImpl(this, F)); 316 }); 317 } 318 319 //===----------------------------------------------------------------------===// 320 // Pass Pipeline Configuration 321 //===----------------------------------------------------------------------===// 322 323 namespace { 324 325 /// X86 Code Generator Pass Configuration Options. 326 class X86PassConfig : public TargetPassConfig { 327 public: 328 X86PassConfig(X86TargetMachine &TM, PassManagerBase &PM) 329 : TargetPassConfig(TM, PM) {} 330 331 X86TargetMachine &getX86TargetMachine() const { 332 return getTM<X86TargetMachine>(); 333 } 334 335 ScheduleDAGInstrs * 336 createMachineScheduler(MachineSchedContext *C) const override { 337 ScheduleDAGMILive *DAG = createGenericSchedLive(C); 338 DAG->addMutation(createX86MacroFusionDAGMutation()); 339 return DAG; 340 } 341 342 void addIRPasses() override; 343 bool addInstSelector() override; 344 #ifdef LLVM_BUILD_GLOBAL_ISEL 345 bool addIRTranslator() override; 346 bool addLegalizeMachineIR() override; 347 bool addRegBankSelect() override; 348 bool addGlobalInstructionSelect() override; 349 #endif 350 bool addILPOpts() override; 351 bool addPreISel() override; 352 void addPreRegAlloc() override; 353 void addPostRegAlloc() override; 354 void addPreEmitPass() override; 355 void addPreSched2() override; 356 }; 357 358 class X86ExecutionDepsFix : public ExecutionDepsFix { 359 public: 360 static char ID; 361 X86ExecutionDepsFix() : ExecutionDepsFix(ID, X86::VR128XRegClass) {} 362 StringRef getPassName() const override { 363 return "X86 Execution Dependency Fix"; 364 } 365 }; 366 char X86ExecutionDepsFix::ID; 367 368 } // end anonymous namespace 369 370 INITIALIZE_PASS(X86ExecutionDepsFix, "x86-execution-deps-fix", 371 "X86 Execution Dependency Fix", false, false) 372 373 TargetPassConfig *X86TargetMachine::createPassConfig(PassManagerBase &PM) { 374 return new X86PassConfig(*this, PM); 375 } 376 377 void X86PassConfig::addIRPasses() { 378 addPass(createAtomicExpandPass()); 379 380 TargetPassConfig::addIRPasses(); 381 382 if (TM->getOptLevel() != CodeGenOpt::None) 383 addPass(createInterleavedAccessPass()); 384 } 385 386 bool X86PassConfig::addInstSelector() { 387 // Install an instruction selector. 388 addPass(createX86ISelDag(getX86TargetMachine(), getOptLevel())); 389 390 // For ELF, cleanup any local-dynamic TLS accesses. 391 if (TM->getTargetTriple().isOSBinFormatELF() && 392 getOptLevel() != CodeGenOpt::None) 393 addPass(createCleanupLocalDynamicTLSPass()); 394 395 addPass(createX86GlobalBaseRegPass()); 396 return false; 397 } 398 399 #ifdef LLVM_BUILD_GLOBAL_ISEL 400 bool X86PassConfig::addIRTranslator() { 401 addPass(new IRTranslator()); 402 return false; 403 } 404 405 bool X86PassConfig::addLegalizeMachineIR() { 406 addPass(new Legalizer()); 407 return false; 408 } 409 410 bool X86PassConfig::addRegBankSelect() { 411 addPass(new RegBankSelect()); 412 return false; 413 } 414 415 bool X86PassConfig::addGlobalInstructionSelect() { 416 addPass(new InstructionSelect()); 417 return false; 418 } 419 #endif 420 421 bool X86PassConfig::addILPOpts() { 422 addPass(&EarlyIfConverterID); 423 if (EnableMachineCombinerPass) 424 addPass(&MachineCombinerID); 425 return true; 426 } 427 428 bool X86PassConfig::addPreISel() { 429 // Only add this pass for 32-bit x86 Windows. 430 const Triple &TT = TM->getTargetTriple(); 431 if (TT.isOSWindows() && TT.getArch() == Triple::x86) 432 addPass(createX86WinEHStatePass()); 433 return true; 434 } 435 436 void X86PassConfig::addPreRegAlloc() { 437 if (getOptLevel() != CodeGenOpt::None) { 438 addPass(&LiveRangeShrinkID); 439 addPass(createX86FixupSetCC()); 440 addPass(createX86OptimizeLEAs()); 441 addPass(createX86CallFrameOptimization()); 442 } 443 444 addPass(createX86WinAllocaExpander()); 445 } 446 447 void X86PassConfig::addPostRegAlloc() { 448 addPass(createX86FloatingPointStackifierPass()); 449 } 450 451 void X86PassConfig::addPreSched2() { addPass(createX86ExpandPseudoPass()); } 452 453 void X86PassConfig::addPreEmitPass() { 454 if (getOptLevel() != CodeGenOpt::None) 455 addPass(new X86ExecutionDepsFix()); 456 457 if (UseVZeroUpper) 458 addPass(createX86IssueVZeroUpperPass()); 459 460 if (getOptLevel() != CodeGenOpt::None) { 461 addPass(createX86FixupBWInsts()); 462 addPass(createX86PadShortFunctions()); 463 addPass(createX86FixupLEAs()); 464 addPass(createX86EvexToVexInsts()); 465 } 466 } 467