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 "X86MacroFusion.h" 18 #include "X86Subtarget.h" 19 #include "X86TargetMachine.h" 20 #include "X86TargetObjectFile.h" 21 #include "X86TargetTransformInfo.h" 22 #include "llvm/ADT/Optional.h" 23 #include "llvm/ADT/SmallString.h" 24 #include "llvm/ADT/STLExtras.h" 25 #include "llvm/ADT/StringRef.h" 26 #include "llvm/ADT/Triple.h" 27 #include "llvm/Analysis/TargetTransformInfo.h" 28 #include "llvm/CodeGen/GlobalISel/CallLowering.h" 29 #include "llvm/CodeGen/GlobalISel/GISelAccessor.h" 30 #include "llvm/CodeGen/GlobalISel/IRTranslator.h" 31 #include "llvm/CodeGen/MachineScheduler.h" 32 #include "llvm/CodeGen/Passes.h" 33 #include "llvm/CodeGen/TargetPassConfig.h" 34 #include "llvm/IR/Attributes.h" 35 #include "llvm/IR/DataLayout.h" 36 #include "llvm/IR/Function.h" 37 #include "llvm/Pass.h" 38 #include "llvm/Support/CodeGen.h" 39 #include "llvm/Support/CommandLine.h" 40 #include "llvm/Support/ErrorHandling.h" 41 #include "llvm/Support/TargetRegistry.h" 42 #include "llvm/Target/TargetLoweringObjectFile.h" 43 #include "llvm/Target/TargetOptions.h" 44 #include <memory> 45 #include <string> 46 47 using namespace llvm; 48 49 static cl::opt<bool> EnableMachineCombinerPass("x86-machine-combiner", 50 cl::desc("Enable the machine combiner pass"), 51 cl::init(true), cl::Hidden); 52 53 namespace llvm { 54 55 void initializeWinEHStatePassPass(PassRegistry &); 56 57 } // end namespace llvm 58 59 extern "C" void LLVMInitializeX86Target() { 60 // Register the target. 61 RegisterTargetMachine<X86TargetMachine> X(getTheX86_32Target()); 62 RegisterTargetMachine<X86TargetMachine> Y(getTheX86_64Target()); 63 64 PassRegistry &PR = *PassRegistry::getPassRegistry(); 65 initializeGlobalISel(PR); 66 initializeWinEHStatePassPass(PR); 67 initializeFixupBWInstPassPass(PR); 68 initializeEvexToVexInstPassPass(PR); 69 } 70 71 static std::unique_ptr<TargetLoweringObjectFile> createTLOF(const Triple &TT) { 72 if (TT.isOSBinFormatMachO()) { 73 if (TT.getArch() == Triple::x86_64) 74 return llvm::make_unique<X86_64MachoTargetObjectFile>(); 75 return llvm::make_unique<TargetLoweringObjectFileMachO>(); 76 } 77 78 if (TT.isOSFreeBSD()) 79 return llvm::make_unique<X86FreeBSDTargetObjectFile>(); 80 if (TT.isOSLinux() || TT.isOSNaCl()) 81 return llvm::make_unique<X86LinuxNaClTargetObjectFile>(); 82 if (TT.isOSFuchsia()) 83 return llvm::make_unique<X86FuchsiaTargetObjectFile>(); 84 if (TT.isOSBinFormatELF()) 85 return llvm::make_unique<X86ELFTargetObjectFile>(); 86 if (TT.isKnownWindowsMSVCEnvironment() || TT.isWindowsCoreCLREnvironment()) 87 return llvm::make_unique<X86WindowsTargetObjectFile>(); 88 if (TT.isOSBinFormatCOFF()) 89 return llvm::make_unique<TargetLoweringObjectFileCOFF>(); 90 llvm_unreachable("unknown subtarget type"); 91 } 92 93 static std::string computeDataLayout(const Triple &TT) { 94 // X86 is little endian 95 std::string Ret = "e"; 96 97 Ret += DataLayout::getManglingComponent(TT); 98 // X86 and x32 have 32 bit pointers. 99 if ((TT.isArch64Bit() && 100 (TT.getEnvironment() == Triple::GNUX32 || TT.isOSNaCl())) || 101 !TT.isArch64Bit()) 102 Ret += "-p:32:32"; 103 104 // Some ABIs align 64 bit integers and doubles to 64 bits, others to 32. 105 if (TT.isArch64Bit() || TT.isOSWindows() || TT.isOSNaCl()) 106 Ret += "-i64:64"; 107 else if (TT.isOSIAMCU()) 108 Ret += "-i64:32-f64:32"; 109 else 110 Ret += "-f64:32:64"; 111 112 // Some ABIs align long double to 128 bits, others to 32. 113 if (TT.isOSNaCl() || TT.isOSIAMCU()) 114 ; // No f80 115 else if (TT.isArch64Bit() || TT.isOSDarwin()) 116 Ret += "-f80:128"; 117 else 118 Ret += "-f80:32"; 119 120 if (TT.isOSIAMCU()) 121 Ret += "-f128:32"; 122 123 // The registers can hold 8, 16, 32 or, in x86-64, 64 bits. 124 if (TT.isArch64Bit()) 125 Ret += "-n8:16:32:64"; 126 else 127 Ret += "-n8:16:32"; 128 129 // The stack is aligned to 32 bits on some ABIs and 128 bits on others. 130 if ((!TT.isArch64Bit() && TT.isOSWindows()) || TT.isOSIAMCU()) 131 Ret += "-a:0:32-S32"; 132 else 133 Ret += "-S128"; 134 135 return Ret; 136 } 137 138 static Reloc::Model getEffectiveRelocModel(const Triple &TT, 139 Optional<Reloc::Model> RM) { 140 bool is64Bit = TT.getArch() == Triple::x86_64; 141 if (!RM.hasValue()) { 142 // Darwin defaults to PIC in 64 bit mode and dynamic-no-pic in 32 bit mode. 143 // Win64 requires rip-rel addressing, thus we force it to PIC. Otherwise we 144 // use static relocation model by default. 145 if (TT.isOSDarwin()) { 146 if (is64Bit) 147 return Reloc::PIC_; 148 return Reloc::DynamicNoPIC; 149 } 150 if (TT.isOSWindows() && is64Bit) 151 return Reloc::PIC_; 152 return Reloc::Static; 153 } 154 155 // ELF and X86-64 don't have a distinct DynamicNoPIC model. DynamicNoPIC 156 // is defined as a model for code which may be used in static or dynamic 157 // executables but not necessarily a shared library. On X86-32 we just 158 // compile in -static mode, in x86-64 we use PIC. 159 if (*RM == Reloc::DynamicNoPIC) { 160 if (is64Bit) 161 return Reloc::PIC_; 162 if (!TT.isOSDarwin()) 163 return Reloc::Static; 164 } 165 166 // If we are on Darwin, disallow static relocation model in X86-64 mode, since 167 // the Mach-O file format doesn't support it. 168 if (*RM == Reloc::Static && TT.isOSDarwin() && is64Bit) 169 return Reloc::PIC_; 170 171 return *RM; 172 } 173 174 /// Create an X86 target. 175 /// 176 X86TargetMachine::X86TargetMachine(const Target &T, const Triple &TT, 177 StringRef CPU, StringRef FS, 178 const TargetOptions &Options, 179 Optional<Reloc::Model> RM, 180 CodeModel::Model CM, CodeGenOpt::Level OL) 181 : LLVMTargetMachine(T, computeDataLayout(TT), TT, CPU, FS, Options, 182 getEffectiveRelocModel(TT, RM), CM, OL), 183 TLOF(createTLOF(getTargetTriple())) { 184 // Windows stack unwinder gets confused when execution flow "falls through" 185 // after a call to 'noreturn' function. 186 // To prevent that, we emit a trap for 'unreachable' IR instructions. 187 // (which on X86, happens to be the 'ud2' instruction) 188 // On PS4, the "return address" of a 'noreturn' call must still be within 189 // the calling function, and TrapUnreachable is an easy way to get that. 190 // The check here for 64-bit windows is a bit icky, but as we're unlikely 191 // to ever want to mix 32 and 64-bit windows code in a single module 192 // this should be fine. 193 if ((TT.isOSWindows() && TT.getArch() == Triple::x86_64) || TT.isPS4()) 194 this->Options.TrapUnreachable = true; 195 196 initAsmInfo(); 197 } 198 199 X86TargetMachine::~X86TargetMachine() = default; 200 201 #ifdef LLVM_BUILD_GLOBAL_ISEL 202 namespace { 203 204 struct X86GISelActualAccessor : public GISelAccessor { 205 std::unique_ptr<CallLowering> CL; 206 207 X86GISelActualAccessor(CallLowering* CL): CL(CL) {} 208 209 const CallLowering *getCallLowering() const override { 210 return CL.get(); 211 } 212 213 const InstructionSelector *getInstructionSelector() const override { 214 //TODO: Implement 215 return nullptr; 216 } 217 218 const LegalizerInfo *getLegalizerInfo() const override { 219 //TODO: Implement 220 return nullptr; 221 } 222 223 const RegisterBankInfo *getRegBankInfo() const override { 224 //TODO: Implement 225 return nullptr; 226 } 227 }; 228 229 } // end anonymous namespace 230 #endif 231 232 const X86Subtarget * 233 X86TargetMachine::getSubtargetImpl(const Function &F) const { 234 Attribute CPUAttr = F.getFnAttribute("target-cpu"); 235 Attribute FSAttr = F.getFnAttribute("target-features"); 236 237 StringRef CPU = !CPUAttr.hasAttribute(Attribute::None) 238 ? CPUAttr.getValueAsString() 239 : (StringRef)TargetCPU; 240 StringRef FS = !FSAttr.hasAttribute(Attribute::None) 241 ? FSAttr.getValueAsString() 242 : (StringRef)TargetFS; 243 244 SmallString<512> Key; 245 Key.reserve(CPU.size() + FS.size()); 246 Key += CPU; 247 Key += FS; 248 249 // FIXME: This is related to the code below to reset the target options, 250 // we need to know whether or not the soft float flag is set on the 251 // function before we can generate a subtarget. We also need to use 252 // it as a key for the subtarget since that can be the only difference 253 // between two functions. 254 bool SoftFloat = 255 F.getFnAttribute("use-soft-float").getValueAsString() == "true"; 256 // If the soft float attribute is set on the function turn on the soft float 257 // subtarget feature. 258 if (SoftFloat) 259 Key += FS.empty() ? "+soft-float" : ",+soft-float"; 260 261 FS = Key.substr(CPU.size()); 262 263 auto &I = SubtargetMap[Key]; 264 if (!I) { 265 // This needs to be done before we create a new subtarget since any 266 // creation will depend on the TM and the code generation flags on the 267 // function that reside in TargetOptions. 268 resetTargetOptions(F); 269 I = llvm::make_unique<X86Subtarget>(TargetTriple, CPU, FS, *this, 270 Options.StackAlignmentOverride); 271 #ifndef LLVM_BUILD_GLOBAL_ISEL 272 GISelAccessor *GISel = new GISelAccessor(); 273 #else 274 X86GISelActualAccessor *GISel = new X86GISelActualAccessor( 275 new X86CallLowering(*I->getTargetLowering())); 276 #endif 277 I->setGISelAccessor(*GISel); 278 } 279 return I.get(); 280 } 281 282 //===----------------------------------------------------------------------===// 283 // Command line options for x86 284 //===----------------------------------------------------------------------===// 285 static cl::opt<bool> 286 UseVZeroUpper("x86-use-vzeroupper", cl::Hidden, 287 cl::desc("Minimize AVX to SSE transition penalty"), 288 cl::init(true)); 289 290 //===----------------------------------------------------------------------===// 291 // X86 TTI query. 292 //===----------------------------------------------------------------------===// 293 294 TargetIRAnalysis X86TargetMachine::getTargetIRAnalysis() { 295 return TargetIRAnalysis([this](const Function &F) { 296 return TargetTransformInfo(X86TTIImpl(this, F)); 297 }); 298 } 299 300 //===----------------------------------------------------------------------===// 301 // Pass Pipeline Configuration 302 //===----------------------------------------------------------------------===// 303 304 namespace { 305 306 /// X86 Code Generator Pass Configuration Options. 307 class X86PassConfig : public TargetPassConfig { 308 public: 309 X86PassConfig(X86TargetMachine *TM, PassManagerBase &PM) 310 : TargetPassConfig(TM, PM) {} 311 312 X86TargetMachine &getX86TargetMachine() const { 313 return getTM<X86TargetMachine>(); 314 } 315 316 ScheduleDAGInstrs * 317 createMachineScheduler(MachineSchedContext *C) const override { 318 ScheduleDAGMILive *DAG = createGenericSchedLive(C); 319 DAG->addMutation(createX86MacroFusionDAGMutation()); 320 return DAG; 321 } 322 323 void addIRPasses() override; 324 bool addInstSelector() override; 325 #ifdef LLVM_BUILD_GLOBAL_ISEL 326 bool addIRTranslator() override; 327 bool addLegalizeMachineIR() override; 328 bool addRegBankSelect() override; 329 bool addGlobalInstructionSelect() override; 330 #endif 331 bool addILPOpts() override; 332 bool addPreISel() override; 333 void addPreRegAlloc() override; 334 void addPostRegAlloc() override; 335 void addPreEmitPass() override; 336 void addPreSched2() override; 337 }; 338 339 } // end anonymous namespace 340 341 TargetPassConfig *X86TargetMachine::createPassConfig(PassManagerBase &PM) { 342 return new X86PassConfig(this, PM); 343 } 344 345 void X86PassConfig::addIRPasses() { 346 addPass(createAtomicExpandPass(&getX86TargetMachine())); 347 348 TargetPassConfig::addIRPasses(); 349 350 if (TM->getOptLevel() != CodeGenOpt::None) 351 addPass(createInterleavedAccessPass(TM)); 352 } 353 354 bool X86PassConfig::addInstSelector() { 355 // Install an instruction selector. 356 addPass(createX86ISelDag(getX86TargetMachine(), getOptLevel())); 357 358 // For ELF, cleanup any local-dynamic TLS accesses. 359 if (TM->getTargetTriple().isOSBinFormatELF() && 360 getOptLevel() != CodeGenOpt::None) 361 addPass(createCleanupLocalDynamicTLSPass()); 362 363 addPass(createX86GlobalBaseRegPass()); 364 return false; 365 } 366 367 #ifdef LLVM_BUILD_GLOBAL_ISEL 368 bool X86PassConfig::addIRTranslator() { 369 addPass(new IRTranslator()); 370 return false; 371 } 372 373 bool X86PassConfig::addLegalizeMachineIR() { 374 //TODO: Implement 375 return false; 376 } 377 378 bool X86PassConfig::addRegBankSelect() { 379 //TODO: Implement 380 return false; 381 } 382 383 bool X86PassConfig::addGlobalInstructionSelect() { 384 //TODO: Implement 385 return false; 386 } 387 #endif 388 389 bool X86PassConfig::addILPOpts() { 390 addPass(&EarlyIfConverterID); 391 if (EnableMachineCombinerPass) 392 addPass(&MachineCombinerID); 393 return true; 394 } 395 396 bool X86PassConfig::addPreISel() { 397 // Only add this pass for 32-bit x86 Windows. 398 const Triple &TT = TM->getTargetTriple(); 399 if (TT.isOSWindows() && TT.getArch() == Triple::x86) 400 addPass(createX86WinEHStatePass()); 401 return true; 402 } 403 404 void X86PassConfig::addPreRegAlloc() { 405 if (getOptLevel() != CodeGenOpt::None) { 406 addPass(createX86FixupSetCC()); 407 addPass(createX86OptimizeLEAs()); 408 addPass(createX86CallFrameOptimization()); 409 } 410 411 addPass(createX86WinAllocaExpander()); 412 } 413 414 void X86PassConfig::addPostRegAlloc() { 415 addPass(createX86FloatingPointStackifierPass()); 416 } 417 418 void X86PassConfig::addPreSched2() { addPass(createX86ExpandPseudoPass()); } 419 420 void X86PassConfig::addPreEmitPass() { 421 if (getOptLevel() != CodeGenOpt::None) 422 addPass(createExecutionDependencyFixPass(&X86::VR128XRegClass)); 423 424 if (UseVZeroUpper) 425 addPass(createX86IssueVZeroUpperPass()); 426 427 if (getOptLevel() != CodeGenOpt::None) { 428 addPass(createX86FixupBWInsts()); 429 addPass(createX86PadShortFunctions()); 430 addPass(createX86FixupLEAs()); 431 addPass(createX86EvexToVexInsts()); 432 } 433 } 434