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 "X86TargetMachine.h" 15 #include "MCTargetDesc/X86MCTargetDesc.h" 16 #include "X86.h" 17 #include "X86CallLowering.h" 18 #include "X86LegalizerInfo.h" 19 #include "X86MacroFusion.h" 20 #include "X86Subtarget.h" 21 #include "X86TargetObjectFile.h" 22 #include "X86TargetTransformInfo.h" 23 #include "llvm/ADT/Optional.h" 24 #include "llvm/ADT/STLExtras.h" 25 #include "llvm/ADT/SmallString.h" 26 #include "llvm/ADT/StringRef.h" 27 #include "llvm/ADT/Triple.h" 28 #include "llvm/Analysis/TargetTransformInfo.h" 29 #include "llvm/CodeGen/ExecutionDomainFix.h" 30 #include "llvm/CodeGen/GlobalISel/CallLowering.h" 31 #include "llvm/CodeGen/GlobalISel/IRTranslator.h" 32 #include "llvm/CodeGen/GlobalISel/InstructionSelect.h" 33 #include "llvm/CodeGen/GlobalISel/Legalizer.h" 34 #include "llvm/CodeGen/GlobalISel/RegBankSelect.h" 35 #include "llvm/CodeGen/MachineScheduler.h" 36 #include "llvm/CodeGen/Passes.h" 37 #include "llvm/CodeGen/TargetLoweringObjectFile.h" 38 #include "llvm/CodeGen/TargetPassConfig.h" 39 #include "llvm/IR/Attributes.h" 40 #include "llvm/IR/DataLayout.h" 41 #include "llvm/IR/Function.h" 42 #include "llvm/Pass.h" 43 #include "llvm/Support/CodeGen.h" 44 #include "llvm/Support/CommandLine.h" 45 #include "llvm/Support/ErrorHandling.h" 46 #include "llvm/Support/TargetRegistry.h" 47 #include "llvm/Target/TargetOptions.h" 48 #include <memory> 49 #include <string> 50 51 using namespace llvm; 52 53 static cl::opt<bool> EnableMachineCombinerPass("x86-machine-combiner", 54 cl::desc("Enable the machine combiner pass"), 55 cl::init(true), cl::Hidden); 56 57 namespace llvm { 58 59 void initializeWinEHStatePassPass(PassRegistry &); 60 void initializeFixupLEAPassPass(PassRegistry &); 61 void initializeX86CallFrameOptimizationPass(PassRegistry &); 62 void initializeX86CmovConverterPassPass(PassRegistry &); 63 void initializeX86ExecutionDomainFixPass(PassRegistry &); 64 void initializeX86DomainReassignmentPass(PassRegistry &); 65 66 } // end namespace llvm 67 68 extern "C" void LLVMInitializeX86Target() { 69 // Register the target. 70 RegisterTargetMachine<X86TargetMachine> X(getTheX86_32Target()); 71 RegisterTargetMachine<X86TargetMachine> Y(getTheX86_64Target()); 72 73 PassRegistry &PR = *PassRegistry::getPassRegistry(); 74 initializeGlobalISel(PR); 75 initializeWinEHStatePassPass(PR); 76 initializeFixupBWInstPassPass(PR); 77 initializeEvexToVexInstPassPass(PR); 78 initializeFixupLEAPassPass(PR); 79 initializeX86CallFrameOptimizationPass(PR); 80 initializeX86CmovConverterPassPass(PR); 81 initializeX86ExecutionDomainFixPass(PR); 82 initializeX86DomainReassignmentPass(PR); 83 } 84 85 static std::unique_ptr<TargetLoweringObjectFile> createTLOF(const Triple &TT) { 86 if (TT.isOSBinFormatMachO()) { 87 if (TT.getArch() == Triple::x86_64) 88 return llvm::make_unique<X86_64MachoTargetObjectFile>(); 89 return llvm::make_unique<TargetLoweringObjectFileMachO>(); 90 } 91 92 if (TT.isOSFreeBSD()) 93 return llvm::make_unique<X86FreeBSDTargetObjectFile>(); 94 if (TT.isOSLinux() || TT.isOSNaCl() || TT.isOSIAMCU()) 95 return llvm::make_unique<X86LinuxNaClTargetObjectFile>(); 96 if (TT.isOSSolaris()) 97 return llvm::make_unique<X86SolarisTargetObjectFile>(); 98 if (TT.isOSFuchsia()) 99 return llvm::make_unique<X86FuchsiaTargetObjectFile>(); 100 if (TT.isOSBinFormatELF()) 101 return llvm::make_unique<X86ELFTargetObjectFile>(); 102 if (TT.isKnownWindowsMSVCEnvironment() || TT.isWindowsCoreCLREnvironment()) 103 return llvm::make_unique<X86WindowsTargetObjectFile>(); 104 if (TT.isOSBinFormatCOFF()) 105 return llvm::make_unique<TargetLoweringObjectFileCOFF>(); 106 llvm_unreachable("unknown subtarget type"); 107 } 108 109 static std::string computeDataLayout(const Triple &TT) { 110 // X86 is little endian 111 std::string Ret = "e"; 112 113 Ret += DataLayout::getManglingComponent(TT); 114 // X86 and x32 have 32 bit pointers. 115 if ((TT.isArch64Bit() && 116 (TT.getEnvironment() == Triple::GNUX32 || TT.isOSNaCl())) || 117 !TT.isArch64Bit()) 118 Ret += "-p:32:32"; 119 120 // Some ABIs align 64 bit integers and doubles to 64 bits, others to 32. 121 if (TT.isArch64Bit() || TT.isOSWindows() || TT.isOSNaCl()) 122 Ret += "-i64:64"; 123 else if (TT.isOSIAMCU()) 124 Ret += "-i64:32-f64:32"; 125 else 126 Ret += "-f64:32:64"; 127 128 // Some ABIs align long double to 128 bits, others to 32. 129 if (TT.isOSNaCl() || TT.isOSIAMCU()) 130 ; // No f80 131 else if (TT.isArch64Bit() || TT.isOSDarwin()) 132 Ret += "-f80:128"; 133 else 134 Ret += "-f80:32"; 135 136 if (TT.isOSIAMCU()) 137 Ret += "-f128:32"; 138 139 // The registers can hold 8, 16, 32 or, in x86-64, 64 bits. 140 if (TT.isArch64Bit()) 141 Ret += "-n8:16:32:64"; 142 else 143 Ret += "-n8:16:32"; 144 145 // The stack is aligned to 32 bits on some ABIs and 128 bits on others. 146 if ((!TT.isArch64Bit() && TT.isOSWindows()) || TT.isOSIAMCU()) 147 Ret += "-a:0:32-S32"; 148 else 149 Ret += "-S128"; 150 151 return Ret; 152 } 153 154 static Reloc::Model getEffectiveRelocModel(const Triple &TT, 155 Optional<Reloc::Model> RM) { 156 bool is64Bit = TT.getArch() == Triple::x86_64; 157 if (!RM.hasValue()) { 158 // Darwin defaults to PIC in 64 bit mode and dynamic-no-pic in 32 bit mode. 159 // Win64 requires rip-rel addressing, thus we force it to PIC. Otherwise we 160 // use static relocation model by default. 161 if (TT.isOSDarwin()) { 162 if (is64Bit) 163 return Reloc::PIC_; 164 return Reloc::DynamicNoPIC; 165 } 166 if (TT.isOSWindows() && is64Bit) 167 return Reloc::PIC_; 168 return Reloc::Static; 169 } 170 171 // ELF and X86-64 don't have a distinct DynamicNoPIC model. DynamicNoPIC 172 // is defined as a model for code which may be used in static or dynamic 173 // executables but not necessarily a shared library. On X86-32 we just 174 // compile in -static mode, in x86-64 we use PIC. 175 if (*RM == Reloc::DynamicNoPIC) { 176 if (is64Bit) 177 return Reloc::PIC_; 178 if (!TT.isOSDarwin()) 179 return Reloc::Static; 180 } 181 182 // If we are on Darwin, disallow static relocation model in X86-64 mode, since 183 // the Mach-O file format doesn't support it. 184 if (*RM == Reloc::Static && TT.isOSDarwin() && is64Bit) 185 return Reloc::PIC_; 186 187 return *RM; 188 } 189 190 static CodeModel::Model getEffectiveCodeModel(Optional<CodeModel::Model> CM, 191 bool JIT, bool Is64Bit) { 192 if (CM) 193 return *CM; 194 if (JIT) 195 return Is64Bit ? CodeModel::Large : CodeModel::Small; 196 return CodeModel::Small; 197 } 198 199 /// Create an X86 target. 200 /// 201 X86TargetMachine::X86TargetMachine(const Target &T, const Triple &TT, 202 StringRef CPU, StringRef FS, 203 const TargetOptions &Options, 204 Optional<Reloc::Model> RM, 205 Optional<CodeModel::Model> CM, 206 CodeGenOpt::Level OL, bool JIT) 207 : LLVMTargetMachine( 208 T, computeDataLayout(TT), TT, CPU, FS, Options, 209 getEffectiveRelocModel(TT, RM), 210 getEffectiveCodeModel(CM, JIT, TT.getArch() == Triple::x86_64), OL), 211 TLOF(createTLOF(getTargetTriple())) { 212 // Windows stack unwinder gets confused when execution flow "falls through" 213 // after a call to 'noreturn' function. 214 // To prevent that, we emit a trap for 'unreachable' IR instructions. 215 // (which on X86, happens to be the 'ud2' instruction) 216 // On PS4, the "return address" of a 'noreturn' call must still be within 217 // the calling function, and TrapUnreachable is an easy way to get that. 218 // The check here for 64-bit windows is a bit icky, but as we're unlikely 219 // to ever want to mix 32 and 64-bit windows code in a single module 220 // this should be fine. 221 if ((TT.isOSWindows() && TT.getArch() == Triple::x86_64) || TT.isPS4()) 222 this->Options.TrapUnreachable = true; 223 224 initAsmInfo(); 225 } 226 227 X86TargetMachine::~X86TargetMachine() = default; 228 229 const X86Subtarget * 230 X86TargetMachine::getSubtargetImpl(const Function &F) const { 231 Attribute CPUAttr = F.getFnAttribute("target-cpu"); 232 Attribute FSAttr = F.getFnAttribute("target-features"); 233 234 StringRef CPU = !CPUAttr.hasAttribute(Attribute::None) 235 ? CPUAttr.getValueAsString() 236 : (StringRef)TargetCPU; 237 StringRef FS = !FSAttr.hasAttribute(Attribute::None) 238 ? FSAttr.getValueAsString() 239 : (StringRef)TargetFS; 240 241 SmallString<512> Key; 242 Key.reserve(CPU.size() + FS.size()); 243 Key += CPU; 244 Key += FS; 245 246 // FIXME: This is related to the code below to reset the target options, 247 // we need to know whether or not the soft float flag is set on the 248 // function before we can generate a subtarget. We also need to use 249 // it as a key for the subtarget since that can be the only difference 250 // between two functions. 251 bool SoftFloat = 252 F.getFnAttribute("use-soft-float").getValueAsString() == "true"; 253 // If the soft float attribute is set on the function turn on the soft float 254 // subtarget feature. 255 if (SoftFloat) 256 Key += FS.empty() ? "+soft-float" : ",+soft-float"; 257 258 // Keep track of the key width after all features are added so we can extract 259 // the feature string out later. 260 unsigned CPUFSWidth = Key.size(); 261 262 // Translate vector width function attribute into subtarget features. This 263 // overrides any CPU specific turning parameter 264 unsigned PreferVectorWidthOverride = 0; 265 if (F.hasFnAttribute("prefer-vector-width")) { 266 StringRef Val = F.getFnAttribute("prefer-vector-width").getValueAsString(); 267 unsigned Width; 268 if (!Val.getAsInteger(0, Width)) { 269 Key += ",prefer-vector-width="; 270 Key += Val; 271 PreferVectorWidthOverride = Width; 272 } 273 } 274 275 FS = Key.slice(CPU.size(), CPUFSWidth); 276 277 auto &I = SubtargetMap[Key]; 278 if (!I) { 279 // This needs to be done before we create a new subtarget since any 280 // creation will depend on the TM and the code generation flags on the 281 // function that reside in TargetOptions. 282 resetTargetOptions(F); 283 I = llvm::make_unique<X86Subtarget>(TargetTriple, CPU, FS, *this, 284 Options.StackAlignmentOverride, 285 PreferVectorWidthOverride); 286 } 287 return I.get(); 288 } 289 290 //===----------------------------------------------------------------------===// 291 // Command line options for x86 292 //===----------------------------------------------------------------------===// 293 static cl::opt<bool> 294 UseVZeroUpper("x86-use-vzeroupper", cl::Hidden, 295 cl::desc("Minimize AVX to SSE transition penalty"), 296 cl::init(true)); 297 298 //===----------------------------------------------------------------------===// 299 // X86 TTI query. 300 //===----------------------------------------------------------------------===// 301 302 TargetTransformInfo 303 X86TargetMachine::getTargetTransformInfo(const Function &F) { 304 return TargetTransformInfo(X86TTIImpl(this, F)); 305 } 306 307 //===----------------------------------------------------------------------===// 308 // Pass Pipeline Configuration 309 //===----------------------------------------------------------------------===// 310 311 namespace { 312 313 /// X86 Code Generator Pass Configuration Options. 314 class X86PassConfig : public TargetPassConfig { 315 public: 316 X86PassConfig(X86TargetMachine &TM, PassManagerBase &PM) 317 : TargetPassConfig(TM, PM) {} 318 319 X86TargetMachine &getX86TargetMachine() const { 320 return getTM<X86TargetMachine>(); 321 } 322 323 ScheduleDAGInstrs * 324 createMachineScheduler(MachineSchedContext *C) const override { 325 ScheduleDAGMILive *DAG = createGenericSchedLive(C); 326 DAG->addMutation(createX86MacroFusionDAGMutation()); 327 return DAG; 328 } 329 330 void addIRPasses() override; 331 bool addInstSelector() override; 332 bool addIRTranslator() override; 333 bool addLegalizeMachineIR() override; 334 bool addRegBankSelect() override; 335 bool addGlobalInstructionSelect() override; 336 bool addILPOpts() override; 337 bool addPreISel() override; 338 void addMachineSSAOptimization() override; 339 void addPreRegAlloc() override; 340 void addPostRegAlloc() override; 341 void addPreEmitPass() override; 342 void addPreSched2() override; 343 }; 344 345 class X86ExecutionDomainFix : public ExecutionDomainFix { 346 public: 347 static char ID; 348 X86ExecutionDomainFix() : ExecutionDomainFix(ID, X86::VR128XRegClass) {} 349 StringRef getPassName() const override { 350 return "X86 Execution Dependency Fix"; 351 } 352 }; 353 char X86ExecutionDomainFix::ID; 354 355 } // end anonymous namespace 356 357 INITIALIZE_PASS_BEGIN(X86ExecutionDomainFix, "x86-execution-domain-fix", 358 "X86 Execution Domain Fix", false, false) 359 INITIALIZE_PASS_DEPENDENCY(ReachingDefAnalysis) 360 INITIALIZE_PASS_END(X86ExecutionDomainFix, "x86-execution-domain-fix", 361 "X86 Execution Domain Fix", false, false) 362 363 TargetPassConfig *X86TargetMachine::createPassConfig(PassManagerBase &PM) { 364 return new X86PassConfig(*this, PM); 365 } 366 367 void X86PassConfig::addIRPasses() { 368 addPass(createAtomicExpandPass()); 369 370 TargetPassConfig::addIRPasses(); 371 372 if (TM->getOptLevel() != CodeGenOpt::None) 373 addPass(createInterleavedAccessPass()); 374 } 375 376 bool X86PassConfig::addInstSelector() { 377 // Install an instruction selector. 378 addPass(createX86ISelDag(getX86TargetMachine(), getOptLevel())); 379 380 // For ELF, cleanup any local-dynamic TLS accesses. 381 if (TM->getTargetTriple().isOSBinFormatELF() && 382 getOptLevel() != CodeGenOpt::None) 383 addPass(createCleanupLocalDynamicTLSPass()); 384 385 addPass(createX86GlobalBaseRegPass()); 386 return false; 387 } 388 389 bool X86PassConfig::addIRTranslator() { 390 addPass(new IRTranslator()); 391 return false; 392 } 393 394 bool X86PassConfig::addLegalizeMachineIR() { 395 addPass(new Legalizer()); 396 return false; 397 } 398 399 bool X86PassConfig::addRegBankSelect() { 400 addPass(new RegBankSelect()); 401 return false; 402 } 403 404 bool X86PassConfig::addGlobalInstructionSelect() { 405 addPass(new InstructionSelect()); 406 return false; 407 } 408 409 bool X86PassConfig::addILPOpts() { 410 addPass(&EarlyIfConverterID); 411 if (EnableMachineCombinerPass) 412 addPass(&MachineCombinerID); 413 addPass(createX86CmovConverterPass()); 414 return true; 415 } 416 417 bool X86PassConfig::addPreISel() { 418 // Only add this pass for 32-bit x86 Windows. 419 const Triple &TT = TM->getTargetTriple(); 420 if (TT.isOSWindows() && TT.getArch() == Triple::x86) 421 addPass(createX86WinEHStatePass()); 422 return true; 423 } 424 425 void X86PassConfig::addPreRegAlloc() { 426 if (getOptLevel() != CodeGenOpt::None) { 427 addPass(&LiveRangeShrinkID); 428 addPass(createX86FixupSetCC()); 429 addPass(createX86OptimizeLEAs()); 430 addPass(createX86CallFrameOptimization()); 431 } 432 433 addPass(createX86WinAllocaExpander()); 434 } 435 void X86PassConfig::addMachineSSAOptimization() { 436 addPass(createX86DomainReassignmentPass()); 437 TargetPassConfig::addMachineSSAOptimization(); 438 } 439 440 void X86PassConfig::addPostRegAlloc() { 441 addPass(createX86FloatingPointStackifierPass()); 442 } 443 444 void X86PassConfig::addPreSched2() { addPass(createX86ExpandPseudoPass()); } 445 446 void X86PassConfig::addPreEmitPass() { 447 if (getOptLevel() != CodeGenOpt::None) { 448 addPass(new X86ExecutionDomainFix()); 449 addPass(createBreakFalseDeps()); 450 } 451 452 addPass(createX86IndirectBranchTrackingPass()); 453 454 if (UseVZeroUpper) 455 addPass(createX86IssueVZeroUpperPass()); 456 457 if (getOptLevel() != CodeGenOpt::None) { 458 addPass(createX86FixupBWInsts()); 459 addPass(createX86PadShortFunctions()); 460 addPass(createX86FixupLEAs()); 461 addPass(createX86EvexToVexInsts()); 462 } 463 } 464