1 //===-- X86TargetMachine.cpp - Define TargetMachine for the X86 -----------===// 2 // 3 // Part of the LLVM Project, under the Apache License v2.0 with LLVM Exceptions. 4 // See https://llvm.org/LICENSE.txt for license information. 5 // SPDX-License-Identifier: Apache-2.0 WITH LLVM-exception 6 // 7 //===----------------------------------------------------------------------===// 8 // 9 // This file defines the X86 specific subclass of TargetMachine. 10 // 11 //===----------------------------------------------------------------------===// 12 13 #include "X86TargetMachine.h" 14 #include "MCTargetDesc/X86MCTargetDesc.h" 15 #include "TargetInfo/X86TargetInfo.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/TargetPassConfig.h" 38 #include "llvm/IR/Attributes.h" 39 #include "llvm/IR/DataLayout.h" 40 #include "llvm/IR/Function.h" 41 #include "llvm/MC/MCAsmInfo.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/TargetLoweringObjectFile.h" 48 #include "llvm/Target/TargetOptions.h" 49 #include "llvm/Transforms/CFGuard.h" 50 #include <memory> 51 #include <string> 52 53 using namespace llvm; 54 55 static cl::opt<bool> EnableMachineCombinerPass("x86-machine-combiner", 56 cl::desc("Enable the machine combiner pass"), 57 cl::init(true), cl::Hidden); 58 59 static cl::opt<bool> EnableCondBrFoldingPass("x86-condbr-folding", 60 cl::desc("Enable the conditional branch " 61 "folding pass"), 62 cl::init(false), cl::Hidden); 63 64 extern "C" LLVM_EXTERNAL_VISIBILITY void LLVMInitializeX86Target() { 65 // Register the target. 66 RegisterTargetMachine<X86TargetMachine> X(getTheX86_32Target()); 67 RegisterTargetMachine<X86TargetMachine> Y(getTheX86_64Target()); 68 69 PassRegistry &PR = *PassRegistry::getPassRegistry(); 70 initializeGlobalISel(PR); 71 initializeWinEHStatePassPass(PR); 72 initializeFixupBWInstPassPass(PR); 73 initializeEvexToVexInstPassPass(PR); 74 initializeFixupLEAPassPass(PR); 75 initializeFPSPass(PR); 76 initializeX86CallFrameOptimizationPass(PR); 77 initializeX86CmovConverterPassPass(PR); 78 initializeX86ExpandPseudoPass(PR); 79 initializeX86ExecutionDomainFixPass(PR); 80 initializeX86DomainReassignmentPass(PR); 81 initializeX86AvoidSFBPassPass(PR); 82 initializeX86SpeculativeLoadHardeningPassPass(PR); 83 initializeX86FlagsCopyLoweringPassPass(PR); 84 initializeX86CondBrFoldingPassPass(PR); 85 initializeX86OptimizeLEAPassPass(PR); 86 } 87 88 static std::unique_ptr<TargetLoweringObjectFile> createTLOF(const Triple &TT) { 89 if (TT.isOSBinFormatMachO()) { 90 if (TT.getArch() == Triple::x86_64) 91 return std::make_unique<X86_64MachoTargetObjectFile>(); 92 return std::make_unique<TargetLoweringObjectFileMachO>(); 93 } 94 95 if (TT.isOSFreeBSD()) 96 return std::make_unique<X86FreeBSDTargetObjectFile>(); 97 if (TT.isOSLinux() || TT.isOSNaCl() || TT.isOSIAMCU()) 98 return std::make_unique<X86LinuxNaClTargetObjectFile>(); 99 if (TT.isOSSolaris()) 100 return std::make_unique<X86SolarisTargetObjectFile>(); 101 if (TT.isOSFuchsia()) 102 return std::make_unique<X86FuchsiaTargetObjectFile>(); 103 if (TT.isOSBinFormatELF()) 104 return std::make_unique<X86ELFTargetObjectFile>(); 105 if (TT.isOSBinFormatCOFF()) 106 return std::make_unique<TargetLoweringObjectFileCOFF>(); 107 llvm_unreachable("unknown subtarget type"); 108 } 109 110 static std::string computeDataLayout(const Triple &TT) { 111 // X86 is little endian 112 std::string Ret = "e"; 113 114 Ret += DataLayout::getManglingComponent(TT); 115 // X86 and x32 have 32 bit pointers. 116 if ((TT.isArch64Bit() && 117 (TT.getEnvironment() == Triple::GNUX32 || TT.isOSNaCl())) || 118 !TT.isArch64Bit()) 119 Ret += "-p:32:32"; 120 121 // Address spaces for 32 bit signed, 32 bit unsigned, and 64 bit pointers. 122 Ret += "-p270:32:32-p271:32:32-p272:64:64"; 123 124 // Some ABIs align 64 bit integers and doubles to 64 bits, others to 32. 125 if (TT.isArch64Bit() || TT.isOSWindows() || TT.isOSNaCl()) 126 Ret += "-i64:64"; 127 else if (TT.isOSIAMCU()) 128 Ret += "-i64:32-f64:32"; 129 else 130 Ret += "-f64:32:64"; 131 132 // Some ABIs align long double to 128 bits, others to 32. 133 if (TT.isOSNaCl() || TT.isOSIAMCU()) 134 ; // No f80 135 else if (TT.isArch64Bit() || TT.isOSDarwin()) 136 Ret += "-f80:128"; 137 else 138 Ret += "-f80:32"; 139 140 if (TT.isOSIAMCU()) 141 Ret += "-f128:32"; 142 143 // The registers can hold 8, 16, 32 or, in x86-64, 64 bits. 144 if (TT.isArch64Bit()) 145 Ret += "-n8:16:32:64"; 146 else 147 Ret += "-n8:16:32"; 148 149 // The stack is aligned to 32 bits on some ABIs and 128 bits on others. 150 if ((!TT.isArch64Bit() && TT.isOSWindows()) || TT.isOSIAMCU()) 151 Ret += "-a:0:32-S32"; 152 else 153 Ret += "-S128"; 154 155 return Ret; 156 } 157 158 static Reloc::Model getEffectiveRelocModel(const Triple &TT, 159 bool JIT, 160 Optional<Reloc::Model> RM) { 161 bool is64Bit = TT.getArch() == Triple::x86_64; 162 if (!RM.hasValue()) { 163 // JIT codegen should use static relocations by default, since it's 164 // typically executed in process and not relocatable. 165 if (JIT) 166 return Reloc::Static; 167 168 // Darwin defaults to PIC in 64 bit mode and dynamic-no-pic in 32 bit mode. 169 // Win64 requires rip-rel addressing, thus we force it to PIC. Otherwise we 170 // use static relocation model by default. 171 if (TT.isOSDarwin()) { 172 if (is64Bit) 173 return Reloc::PIC_; 174 return Reloc::DynamicNoPIC; 175 } 176 if (TT.isOSWindows() && is64Bit) 177 return Reloc::PIC_; 178 return Reloc::Static; 179 } 180 181 // ELF and X86-64 don't have a distinct DynamicNoPIC model. DynamicNoPIC 182 // is defined as a model for code which may be used in static or dynamic 183 // executables but not necessarily a shared library. On X86-32 we just 184 // compile in -static mode, in x86-64 we use PIC. 185 if (*RM == Reloc::DynamicNoPIC) { 186 if (is64Bit) 187 return Reloc::PIC_; 188 if (!TT.isOSDarwin()) 189 return Reloc::Static; 190 } 191 192 // If we are on Darwin, disallow static relocation model in X86-64 mode, since 193 // the Mach-O file format doesn't support it. 194 if (*RM == Reloc::Static && TT.isOSDarwin() && is64Bit) 195 return Reloc::PIC_; 196 197 return *RM; 198 } 199 200 static CodeModel::Model getEffectiveX86CodeModel(Optional<CodeModel::Model> CM, 201 bool JIT, bool Is64Bit) { 202 if (CM) { 203 if (*CM == CodeModel::Tiny) 204 report_fatal_error("Target does not support the tiny CodeModel", false); 205 return *CM; 206 } 207 if (JIT) 208 return Is64Bit ? CodeModel::Large : CodeModel::Small; 209 return CodeModel::Small; 210 } 211 212 /// Create an X86 target. 213 /// 214 X86TargetMachine::X86TargetMachine(const Target &T, const Triple &TT, 215 StringRef CPU, StringRef FS, 216 const TargetOptions &Options, 217 Optional<Reloc::Model> RM, 218 Optional<CodeModel::Model> CM, 219 CodeGenOpt::Level OL, bool JIT) 220 : LLVMTargetMachine( 221 T, computeDataLayout(TT), TT, CPU, FS, Options, 222 getEffectiveRelocModel(TT, JIT, RM), 223 getEffectiveX86CodeModel(CM, JIT, TT.getArch() == Triple::x86_64), 224 OL), 225 TLOF(createTLOF(getTargetTriple())) { 226 // On PS4, the "return address" of a 'noreturn' call must still be within 227 // the calling function, and TrapUnreachable is an easy way to get that. 228 if (TT.isPS4() || TT.isOSBinFormatMachO()) { 229 this->Options.TrapUnreachable = true; 230 this->Options.NoTrapAfterNoreturn = TT.isOSBinFormatMachO(); 231 } 232 233 setMachineOutliner(true); 234 235 // x86 supports the debug entry values. 236 setSupportsDebugEntryValues(true); 237 238 initAsmInfo(); 239 } 240 241 X86TargetMachine::~X86TargetMachine() = default; 242 243 const X86Subtarget * 244 X86TargetMachine::getSubtargetImpl(const Function &F) const { 245 Attribute CPUAttr = F.getFnAttribute("target-cpu"); 246 Attribute FSAttr = F.getFnAttribute("target-features"); 247 248 StringRef CPU = !CPUAttr.hasAttribute(Attribute::None) 249 ? CPUAttr.getValueAsString() 250 : (StringRef)TargetCPU; 251 StringRef FS = !FSAttr.hasAttribute(Attribute::None) 252 ? FSAttr.getValueAsString() 253 : (StringRef)TargetFS; 254 255 SmallString<512> Key; 256 Key.reserve(CPU.size() + FS.size()); 257 Key += CPU; 258 Key += FS; 259 260 // FIXME: This is related to the code below to reset the target options, 261 // we need to know whether or not the soft float flag is set on the 262 // function before we can generate a subtarget. We also need to use 263 // it as a key for the subtarget since that can be the only difference 264 // between two functions. 265 bool SoftFloat = 266 F.getFnAttribute("use-soft-float").getValueAsString() == "true"; 267 // If the soft float attribute is set on the function turn on the soft float 268 // subtarget feature. 269 if (SoftFloat) 270 Key += FS.empty() ? "+soft-float" : ",+soft-float"; 271 272 // Keep track of the key width after all features are added so we can extract 273 // the feature string out later. 274 unsigned CPUFSWidth = Key.size(); 275 276 // Extract prefer-vector-width attribute. 277 unsigned PreferVectorWidthOverride = 0; 278 if (F.hasFnAttribute("prefer-vector-width")) { 279 StringRef Val = F.getFnAttribute("prefer-vector-width").getValueAsString(); 280 unsigned Width; 281 if (!Val.getAsInteger(0, Width)) { 282 Key += ",prefer-vector-width="; 283 Key += Val; 284 PreferVectorWidthOverride = Width; 285 } 286 } 287 288 // Extract min-legal-vector-width attribute. 289 unsigned RequiredVectorWidth = UINT32_MAX; 290 if (F.hasFnAttribute("min-legal-vector-width")) { 291 StringRef Val = 292 F.getFnAttribute("min-legal-vector-width").getValueAsString(); 293 unsigned Width; 294 if (!Val.getAsInteger(0, Width)) { 295 Key += ",min-legal-vector-width="; 296 Key += Val; 297 RequiredVectorWidth = Width; 298 } 299 } 300 301 // Extracted here so that we make sure there is backing for the StringRef. If 302 // we assigned earlier, its possible the SmallString reallocated leaving a 303 // dangling StringRef. 304 FS = Key.slice(CPU.size(), CPUFSWidth); 305 306 auto &I = SubtargetMap[Key]; 307 if (!I) { 308 // This needs to be done before we create a new subtarget since any 309 // creation will depend on the TM and the code generation flags on the 310 // function that reside in TargetOptions. 311 resetTargetOptions(F); 312 I = std::make_unique<X86Subtarget>( 313 TargetTriple, CPU, FS, *this, 314 MaybeAlign(Options.StackAlignmentOverride), PreferVectorWidthOverride, 315 RequiredVectorWidth); 316 } 317 return I.get(); 318 } 319 320 //===----------------------------------------------------------------------===// 321 // Command line options for x86 322 //===----------------------------------------------------------------------===// 323 static cl::opt<bool> 324 UseVZeroUpper("x86-use-vzeroupper", cl::Hidden, 325 cl::desc("Minimize AVX to SSE transition penalty"), 326 cl::init(true)); 327 328 //===----------------------------------------------------------------------===// 329 // X86 TTI query. 330 //===----------------------------------------------------------------------===// 331 332 TargetTransformInfo 333 X86TargetMachine::getTargetTransformInfo(const Function &F) { 334 return TargetTransformInfo(X86TTIImpl(this, F)); 335 } 336 337 //===----------------------------------------------------------------------===// 338 // Pass Pipeline Configuration 339 //===----------------------------------------------------------------------===// 340 341 namespace { 342 343 /// X86 Code Generator Pass Configuration Options. 344 class X86PassConfig : public TargetPassConfig { 345 public: 346 X86PassConfig(X86TargetMachine &TM, PassManagerBase &PM) 347 : TargetPassConfig(TM, PM) {} 348 349 X86TargetMachine &getX86TargetMachine() const { 350 return getTM<X86TargetMachine>(); 351 } 352 353 ScheduleDAGInstrs * 354 createMachineScheduler(MachineSchedContext *C) const override { 355 ScheduleDAGMILive *DAG = createGenericSchedLive(C); 356 DAG->addMutation(createX86MacroFusionDAGMutation()); 357 return DAG; 358 } 359 360 ScheduleDAGInstrs * 361 createPostMachineScheduler(MachineSchedContext *C) const override { 362 ScheduleDAGMI *DAG = createGenericSchedPostRA(C); 363 DAG->addMutation(createX86MacroFusionDAGMutation()); 364 return DAG; 365 } 366 367 void addIRPasses() override; 368 bool addInstSelector() override; 369 bool addIRTranslator() override; 370 bool addLegalizeMachineIR() override; 371 bool addRegBankSelect() override; 372 bool addGlobalInstructionSelect() override; 373 bool addILPOpts() override; 374 bool addPreISel() override; 375 void addMachineSSAOptimization() override; 376 void addPreRegAlloc() override; 377 void addPostRegAlloc() override; 378 void addPreEmitPass() override; 379 void addPreEmitPass2() override; 380 void addPreSched2() override; 381 382 std::unique_ptr<CSEConfigBase> getCSEConfig() const override; 383 }; 384 385 class X86ExecutionDomainFix : public ExecutionDomainFix { 386 public: 387 static char ID; 388 X86ExecutionDomainFix() : ExecutionDomainFix(ID, X86::VR128XRegClass) {} 389 StringRef getPassName() const override { 390 return "X86 Execution Dependency Fix"; 391 } 392 }; 393 char X86ExecutionDomainFix::ID; 394 395 } // end anonymous namespace 396 397 INITIALIZE_PASS_BEGIN(X86ExecutionDomainFix, "x86-execution-domain-fix", 398 "X86 Execution Domain Fix", false, false) 399 INITIALIZE_PASS_DEPENDENCY(ReachingDefAnalysis) 400 INITIALIZE_PASS_END(X86ExecutionDomainFix, "x86-execution-domain-fix", 401 "X86 Execution Domain Fix", false, false) 402 403 TargetPassConfig *X86TargetMachine::createPassConfig(PassManagerBase &PM) { 404 return new X86PassConfig(*this, PM); 405 } 406 407 void X86PassConfig::addIRPasses() { 408 addPass(createAtomicExpandPass()); 409 410 TargetPassConfig::addIRPasses(); 411 412 if (TM->getOptLevel() != CodeGenOpt::None) 413 addPass(createInterleavedAccessPass()); 414 415 // Add passes that handle indirect branch removal and insertion of a retpoline 416 // thunk. These will be a no-op unless a function subtarget has the retpoline 417 // feature enabled. 418 addPass(createIndirectBrExpandPass()); 419 420 // Add Control Flow Guard checks. 421 const Triple &TT = TM->getTargetTriple(); 422 if (TT.isOSWindows()) { 423 if (TT.getArch() == Triple::x86_64) { 424 addPass(createCFGuardDispatchPass()); 425 } else { 426 addPass(createCFGuardCheckPass()); 427 } 428 } 429 } 430 431 bool X86PassConfig::addInstSelector() { 432 // Install an instruction selector. 433 addPass(createX86ISelDag(getX86TargetMachine(), getOptLevel())); 434 435 // For ELF, cleanup any local-dynamic TLS accesses. 436 if (TM->getTargetTriple().isOSBinFormatELF() && 437 getOptLevel() != CodeGenOpt::None) 438 addPass(createCleanupLocalDynamicTLSPass()); 439 440 addPass(createX86GlobalBaseRegPass()); 441 return false; 442 } 443 444 bool X86PassConfig::addIRTranslator() { 445 addPass(new IRTranslator()); 446 return false; 447 } 448 449 bool X86PassConfig::addLegalizeMachineIR() { 450 addPass(new Legalizer()); 451 return false; 452 } 453 454 bool X86PassConfig::addRegBankSelect() { 455 addPass(new RegBankSelect()); 456 return false; 457 } 458 459 bool X86PassConfig::addGlobalInstructionSelect() { 460 addPass(new InstructionSelect()); 461 return false; 462 } 463 464 bool X86PassConfig::addILPOpts() { 465 if (EnableCondBrFoldingPass) 466 addPass(createX86CondBrFolding()); 467 addPass(&EarlyIfConverterID); 468 if (EnableMachineCombinerPass) 469 addPass(&MachineCombinerID); 470 addPass(createX86CmovConverterPass()); 471 return true; 472 } 473 474 bool X86PassConfig::addPreISel() { 475 // Only add this pass for 32-bit x86 Windows. 476 const Triple &TT = TM->getTargetTriple(); 477 if (TT.isOSWindows() && TT.getArch() == Triple::x86) 478 addPass(createX86WinEHStatePass()); 479 return true; 480 } 481 482 void X86PassConfig::addPreRegAlloc() { 483 if (getOptLevel() != CodeGenOpt::None) { 484 addPass(&LiveRangeShrinkID); 485 addPass(createX86FixupSetCC()); 486 addPass(createX86OptimizeLEAs()); 487 addPass(createX86CallFrameOptimization()); 488 addPass(createX86AvoidStoreForwardingBlocks()); 489 } 490 491 addPass(createX86SpeculativeLoadHardeningPass()); 492 addPass(createX86FlagsCopyLoweringPass()); 493 addPass(createX86WinAllocaExpander()); 494 } 495 void X86PassConfig::addMachineSSAOptimization() { 496 addPass(createX86DomainReassignmentPass()); 497 TargetPassConfig::addMachineSSAOptimization(); 498 } 499 500 void X86PassConfig::addPostRegAlloc() { 501 addPass(createX86FloatingPointStackifierPass()); 502 } 503 504 void X86PassConfig::addPreSched2() { addPass(createX86ExpandPseudoPass()); } 505 506 void X86PassConfig::addPreEmitPass() { 507 if (getOptLevel() != CodeGenOpt::None) { 508 addPass(new X86ExecutionDomainFix()); 509 addPass(createBreakFalseDeps()); 510 } 511 512 addPass(createX86IndirectBranchTrackingPass()); 513 514 if (UseVZeroUpper) 515 addPass(createX86IssueVZeroUpperPass()); 516 517 if (getOptLevel() != CodeGenOpt::None) { 518 addPass(createX86FixupBWInsts()); 519 addPass(createX86PadShortFunctions()); 520 addPass(createX86FixupLEAs()); 521 addPass(createX86EvexToVexInsts()); 522 } 523 addPass(createX86DiscriminateMemOpsPass()); 524 addPass(createX86InsertPrefetchPass()); 525 addPass(createX86InsertX87waitPass()); 526 } 527 528 void X86PassConfig::addPreEmitPass2() { 529 const Triple &TT = TM->getTargetTriple(); 530 const MCAsmInfo *MAI = TM->getMCAsmInfo(); 531 532 addPass(createX86RetpolineThunksPass()); 533 534 // Insert extra int3 instructions after trailing call instructions to avoid 535 // issues in the unwinder. 536 if (TT.isOSWindows() && TT.getArch() == Triple::x86_64) 537 addPass(createX86AvoidTrailingCallPass()); 538 539 // Verify basic block incoming and outgoing cfa offset and register values and 540 // correct CFA calculation rule where needed by inserting appropriate CFI 541 // instructions. 542 if (!TT.isOSDarwin() && 543 (!TT.isOSWindows() || 544 MAI->getExceptionHandlingType() == ExceptionHandling::DwarfCFI)) 545 addPass(createCFIInstrInserter()); 546 // Identify valid longjmp targets for Windows Control Flow Guard. 547 if (TT.isOSWindows()) 548 addPass(createCFGuardLongjmpPass()); 549 } 550 551 std::unique_ptr<CSEConfigBase> X86PassConfig::getCSEConfig() const { 552 return getStandardCSEConfigForOpt(TM->getOptLevel()); 553 } 554