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 "X86.h" 16 #include "X86CallLowering.h" 17 #include "X86TargetObjectFile.h" 18 #include "X86TargetTransformInfo.h" 19 #include "llvm/CodeGen/GlobalISel/GISelAccessor.h" 20 #include "llvm/CodeGen/GlobalISel/IRTranslator.h" 21 #include "llvm/CodeGen/MachineScheduler.h" 22 #include "llvm/CodeGen/Passes.h" 23 #include "llvm/CodeGen/TargetPassConfig.h" 24 #include "llvm/IR/Function.h" 25 #include "llvm/IR/LegacyPassManager.h" 26 #include "llvm/Support/CommandLine.h" 27 #include "llvm/Support/FormattedStream.h" 28 #include "llvm/Support/TargetRegistry.h" 29 #include "llvm/Target/TargetOptions.h" 30 using namespace llvm; 31 32 static cl::opt<bool> EnableMachineCombinerPass("x86-machine-combiner", 33 cl::desc("Enable the machine combiner pass"), 34 cl::init(true), cl::Hidden); 35 36 namespace llvm { 37 void initializeWinEHStatePassPass(PassRegistry &); 38 } 39 40 extern "C" void LLVMInitializeX86Target() { 41 // Register the target. 42 RegisterTargetMachine<X86TargetMachine> X(getTheX86_32Target()); 43 RegisterTargetMachine<X86TargetMachine> Y(getTheX86_64Target()); 44 45 PassRegistry &PR = *PassRegistry::getPassRegistry(); 46 initializeGlobalISel(PR); 47 initializeWinEHStatePassPass(PR); 48 initializeFixupBWInstPassPass(PR); 49 } 50 51 static std::unique_ptr<TargetLoweringObjectFile> createTLOF(const Triple &TT) { 52 if (TT.isOSBinFormatMachO()) { 53 if (TT.getArch() == Triple::x86_64) 54 return make_unique<X86_64MachoTargetObjectFile>(); 55 return make_unique<TargetLoweringObjectFileMachO>(); 56 } 57 58 if (TT.isOSFreeBSD()) 59 return make_unique<X86FreeBSDTargetObjectFile>(); 60 if (TT.isOSLinux() || TT.isOSNaCl()) 61 return make_unique<X86LinuxNaClTargetObjectFile>(); 62 if (TT.isOSFuchsia()) 63 return make_unique<X86FuchsiaTargetObjectFile>(); 64 if (TT.isOSBinFormatELF()) 65 return make_unique<X86ELFTargetObjectFile>(); 66 if (TT.isKnownWindowsMSVCEnvironment() || TT.isWindowsCoreCLREnvironment()) 67 return make_unique<X86WindowsTargetObjectFile>(); 68 if (TT.isOSBinFormatCOFF()) 69 return make_unique<TargetLoweringObjectFileCOFF>(); 70 llvm_unreachable("unknown subtarget type"); 71 } 72 73 static std::string computeDataLayout(const Triple &TT) { 74 // X86 is little endian 75 std::string Ret = "e"; 76 77 Ret += DataLayout::getManglingComponent(TT); 78 // X86 and x32 have 32 bit pointers. 79 if ((TT.isArch64Bit() && 80 (TT.getEnvironment() == Triple::GNUX32 || TT.isOSNaCl())) || 81 !TT.isArch64Bit()) 82 Ret += "-p:32:32"; 83 84 // Some ABIs align 64 bit integers and doubles to 64 bits, others to 32. 85 if (TT.isArch64Bit() || TT.isOSWindows() || TT.isOSNaCl()) 86 Ret += "-i64:64"; 87 else if (TT.isOSIAMCU()) 88 Ret += "-i64:32-f64:32"; 89 else 90 Ret += "-f64:32:64"; 91 92 // Some ABIs align long double to 128 bits, others to 32. 93 if (TT.isOSNaCl() || TT.isOSIAMCU()) 94 ; // No f80 95 else if (TT.isArch64Bit() || TT.isOSDarwin()) 96 Ret += "-f80:128"; 97 else 98 Ret += "-f80:32"; 99 100 if (TT.isOSIAMCU()) 101 Ret += "-f128:32"; 102 103 // The registers can hold 8, 16, 32 or, in x86-64, 64 bits. 104 if (TT.isArch64Bit()) 105 Ret += "-n8:16:32:64"; 106 else 107 Ret += "-n8:16:32"; 108 109 // The stack is aligned to 32 bits on some ABIs and 128 bits on others. 110 if ((!TT.isArch64Bit() && TT.isOSWindows()) || TT.isOSIAMCU()) 111 Ret += "-a:0:32-S32"; 112 else 113 Ret += "-S128"; 114 115 return Ret; 116 } 117 118 static Reloc::Model getEffectiveRelocModel(const Triple &TT, 119 Optional<Reloc::Model> RM) { 120 bool is64Bit = TT.getArch() == Triple::x86_64; 121 if (!RM.hasValue()) { 122 // Darwin defaults to PIC in 64 bit mode and dynamic-no-pic in 32 bit mode. 123 // Win64 requires rip-rel addressing, thus we force it to PIC. Otherwise we 124 // use static relocation model by default. 125 if (TT.isOSDarwin()) { 126 if (is64Bit) 127 return Reloc::PIC_; 128 return Reloc::DynamicNoPIC; 129 } 130 if (TT.isOSWindows() && is64Bit) 131 return Reloc::PIC_; 132 return Reloc::Static; 133 } 134 135 // ELF and X86-64 don't have a distinct DynamicNoPIC model. DynamicNoPIC 136 // is defined as a model for code which may be used in static or dynamic 137 // executables but not necessarily a shared library. On X86-32 we just 138 // compile in -static mode, in x86-64 we use PIC. 139 if (*RM == Reloc::DynamicNoPIC) { 140 if (is64Bit) 141 return Reloc::PIC_; 142 if (!TT.isOSDarwin()) 143 return Reloc::Static; 144 } 145 146 // If we are on Darwin, disallow static relocation model in X86-64 mode, since 147 // the Mach-O file format doesn't support it. 148 if (*RM == Reloc::Static && TT.isOSDarwin() && is64Bit) 149 return Reloc::PIC_; 150 151 return *RM; 152 } 153 154 /// Create an X86 target. 155 /// 156 X86TargetMachine::X86TargetMachine(const Target &T, const Triple &TT, 157 StringRef CPU, StringRef FS, 158 const TargetOptions &Options, 159 Optional<Reloc::Model> RM, 160 CodeModel::Model CM, CodeGenOpt::Level OL) 161 : LLVMTargetMachine(T, computeDataLayout(TT), TT, CPU, FS, Options, 162 getEffectiveRelocModel(TT, RM), CM, OL), 163 TLOF(createTLOF(getTargetTriple())) { 164 // Windows stack unwinder gets confused when execution flow "falls through" 165 // after a call to 'noreturn' function. 166 // To prevent that, we emit a trap for 'unreachable' IR instructions. 167 // (which on X86, happens to be the 'ud2' instruction) 168 // On PS4, the "return address" of a 'noreturn' call must still be within 169 // the calling function, and TrapUnreachable is an easy way to get that. 170 // The check here for 64-bit windows is a bit icky, but as we're unlikely 171 // to ever want to mix 32 and 64-bit windows code in a single module 172 // this should be fine. 173 if ((TT.isOSWindows() && TT.getArch() == Triple::x86_64) || TT.isPS4()) 174 this->Options.TrapUnreachable = true; 175 176 initAsmInfo(); 177 } 178 179 X86TargetMachine::~X86TargetMachine() {} 180 181 #ifdef LLVM_BUILD_GLOBAL_ISEL 182 namespace { 183 struct X86GISelActualAccessor : public GISelAccessor { 184 std::unique_ptr<CallLowering> CL; 185 X86GISelActualAccessor(CallLowering* CL): CL(CL) {} 186 const CallLowering *getCallLowering() const override { 187 return CL.get(); 188 } 189 const InstructionSelector *getInstructionSelector() const override { 190 //TODO: Implement 191 return nullptr; 192 } 193 const class LegalizerInfo *getLegalizerInfo() const override { 194 //TODO: Implement 195 return nullptr; 196 } 197 const RegisterBankInfo *getRegBankInfo() const override { 198 //TODO: Implement 199 return nullptr; 200 } 201 }; 202 } // End anonymous namespace. 203 #endif 204 const X86Subtarget * 205 X86TargetMachine::getSubtargetImpl(const Function &F) const { 206 Attribute CPUAttr = F.getFnAttribute("target-cpu"); 207 Attribute FSAttr = F.getFnAttribute("target-features"); 208 209 StringRef CPU = !CPUAttr.hasAttribute(Attribute::None) 210 ? CPUAttr.getValueAsString() 211 : (StringRef)TargetCPU; 212 StringRef FS = !FSAttr.hasAttribute(Attribute::None) 213 ? FSAttr.getValueAsString() 214 : (StringRef)TargetFS; 215 216 SmallString<512> Key; 217 Key.reserve(CPU.size() + FS.size()); 218 Key += CPU; 219 Key += FS; 220 221 // FIXME: This is related to the code below to reset the target options, 222 // we need to know whether or not the soft float flag is set on the 223 // function before we can generate a subtarget. We also need to use 224 // it as a key for the subtarget since that can be the only difference 225 // between two functions. 226 bool SoftFloat = 227 F.getFnAttribute("use-soft-float").getValueAsString() == "true"; 228 // If the soft float attribute is set on the function turn on the soft float 229 // subtarget feature. 230 if (SoftFloat) 231 Key += FS.empty() ? "+soft-float" : ",+soft-float"; 232 233 FS = Key.substr(CPU.size()); 234 235 auto &I = SubtargetMap[Key]; 236 if (!I) { 237 // This needs to be done before we create a new subtarget since any 238 // creation will depend on the TM and the code generation flags on the 239 // function that reside in TargetOptions. 240 resetTargetOptions(F); 241 I = llvm::make_unique<X86Subtarget>(TargetTriple, CPU, FS, *this, 242 Options.StackAlignmentOverride); 243 #ifndef LLVM_BUILD_GLOBAL_ISEL 244 GISelAccessor *GISel = new GISelAccessor(); 245 #else 246 X86GISelActualAccessor *GISel = new X86GISelActualAccessor( 247 new X86CallLowering(*I->getTargetLowering())); 248 #endif 249 I->setGISelAccessor(*GISel); 250 } 251 return I.get(); 252 } 253 254 //===----------------------------------------------------------------------===// 255 // Command line options for x86 256 //===----------------------------------------------------------------------===// 257 static cl::opt<bool> 258 UseVZeroUpper("x86-use-vzeroupper", cl::Hidden, 259 cl::desc("Minimize AVX to SSE transition penalty"), 260 cl::init(true)); 261 262 //===----------------------------------------------------------------------===// 263 // X86 TTI query. 264 //===----------------------------------------------------------------------===// 265 266 TargetIRAnalysis X86TargetMachine::getTargetIRAnalysis() { 267 return TargetIRAnalysis([this](const Function &F) { 268 return TargetTransformInfo(X86TTIImpl(this, F)); 269 }); 270 } 271 272 273 //===----------------------------------------------------------------------===// 274 // Pass Pipeline Configuration 275 //===----------------------------------------------------------------------===// 276 277 namespace { 278 /// X86 Code Generator Pass Configuration Options. 279 class X86PassConfig : public TargetPassConfig { 280 public: 281 X86PassConfig(X86TargetMachine *TM, PassManagerBase &PM) 282 : TargetPassConfig(TM, PM) {} 283 284 X86TargetMachine &getX86TargetMachine() const { 285 return getTM<X86TargetMachine>(); 286 } 287 288 ScheduleDAGInstrs * 289 createMachineScheduler(MachineSchedContext *C) const override { 290 ScheduleDAGMILive *DAG = createGenericSchedLive(C); 291 DAG->addMutation(createMacroFusionDAGMutation(DAG->TII)); 292 return DAG; 293 } 294 295 void addIRPasses() override; 296 bool addInstSelector() override; 297 #ifdef LLVM_BUILD_GLOBAL_ISEL 298 bool addIRTranslator() override; 299 bool addLegalizeMachineIR() override; 300 bool addRegBankSelect() override; 301 bool addGlobalInstructionSelect() override; 302 #endif 303 bool addILPOpts() override; 304 bool addPreISel() override; 305 void addPreRegAlloc() override; 306 void addPostRegAlloc() override; 307 void addPreEmitPass() override; 308 void addPreSched2() override; 309 }; 310 } // namespace 311 312 TargetPassConfig *X86TargetMachine::createPassConfig(PassManagerBase &PM) { 313 return new X86PassConfig(this, PM); 314 } 315 316 void X86PassConfig::addIRPasses() { 317 addPass(createAtomicExpandPass(&getX86TargetMachine())); 318 319 TargetPassConfig::addIRPasses(); 320 321 if (TM->getOptLevel() != CodeGenOpt::None) 322 addPass(createInterleavedAccessPass(TM)); 323 } 324 325 bool X86PassConfig::addInstSelector() { 326 // Install an instruction selector. 327 addPass(createX86ISelDag(getX86TargetMachine(), getOptLevel())); 328 329 // For ELF, cleanup any local-dynamic TLS accesses. 330 if (TM->getTargetTriple().isOSBinFormatELF() && 331 getOptLevel() != CodeGenOpt::None) 332 addPass(createCleanupLocalDynamicTLSPass()); 333 334 addPass(createX86GlobalBaseRegPass()); 335 return false; 336 } 337 338 #ifdef LLVM_BUILD_GLOBAL_ISEL 339 bool X86PassConfig::addIRTranslator() { 340 addPass(new IRTranslator()); 341 return false; 342 } 343 344 bool X86PassConfig::addLegalizeMachineIR() { 345 //TODO: Implement 346 return false; 347 } 348 349 bool X86PassConfig::addRegBankSelect() { 350 //TODO: Implement 351 return false; 352 } 353 354 bool X86PassConfig::addGlobalInstructionSelect() { 355 //TODO: Implement 356 return false; 357 } 358 #endif 359 360 bool X86PassConfig::addILPOpts() { 361 addPass(&EarlyIfConverterID); 362 if (EnableMachineCombinerPass) 363 addPass(&MachineCombinerID); 364 return true; 365 } 366 367 bool X86PassConfig::addPreISel() { 368 // Only add this pass for 32-bit x86 Windows. 369 const Triple &TT = TM->getTargetTriple(); 370 if (TT.isOSWindows() && TT.getArch() == Triple::x86) 371 addPass(createX86WinEHStatePass()); 372 return true; 373 } 374 375 void X86PassConfig::addPreRegAlloc() { 376 if (getOptLevel() != CodeGenOpt::None) { 377 addPass(createX86FixupSetCC()); 378 addPass(createX86OptimizeLEAs()); 379 addPass(createX86CallFrameOptimization()); 380 } 381 382 addPass(createX86WinAllocaExpander()); 383 } 384 385 void X86PassConfig::addPostRegAlloc() { 386 addPass(createX86FloatingPointStackifierPass()); 387 } 388 389 void X86PassConfig::addPreSched2() { addPass(createX86ExpandPseudoPass()); } 390 391 void X86PassConfig::addPreEmitPass() { 392 if (getOptLevel() != CodeGenOpt::None) 393 addPass(createExecutionDependencyFixPass(&X86::VR128XRegClass)); 394 395 if (UseVZeroUpper) 396 addPass(createX86IssueVZeroUpperPass()); 397 398 if (getOptLevel() != CodeGenOpt::None) { 399 addPass(createX86FixupBWInsts()); 400 addPass(createX86PadShortFunctions()); 401 addPass(createX86FixupLEAs()); 402 } 403 } 404