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 initializeX86AvoidTrailingCallPassPass(PR); 83 initializeX86SpeculativeLoadHardeningPassPass(PR); 84 initializeX86FlagsCopyLoweringPassPass(PR); 85 initializeX86CondBrFoldingPassPass(PR); 86 initializeX86OptimizeLEAPassPass(PR); 87 initializeX86PartialReductionPass(PR); 88 } 89 90 static std::unique_ptr<TargetLoweringObjectFile> createTLOF(const Triple &TT) { 91 if (TT.isOSBinFormatMachO()) { 92 if (TT.getArch() == Triple::x86_64) 93 return std::make_unique<X86_64MachoTargetObjectFile>(); 94 return std::make_unique<TargetLoweringObjectFileMachO>(); 95 } 96 97 if (TT.isOSBinFormatCOFF()) 98 return std::make_unique<TargetLoweringObjectFileCOFF>(); 99 return std::make_unique<X86ELFTargetObjectFile>(); 100 } 101 102 static std::string computeDataLayout(const Triple &TT) { 103 // X86 is little endian 104 std::string Ret = "e"; 105 106 Ret += DataLayout::getManglingComponent(TT); 107 // X86 and x32 have 32 bit pointers. 108 if ((TT.isArch64Bit() && 109 (TT.getEnvironment() == Triple::GNUX32 || TT.isOSNaCl())) || 110 !TT.isArch64Bit()) 111 Ret += "-p:32:32"; 112 113 // Address spaces for 32 bit signed, 32 bit unsigned, and 64 bit pointers. 114 Ret += "-p270:32:32-p271:32:32-p272:64:64"; 115 116 // Some ABIs align 64 bit integers and doubles to 64 bits, others to 32. 117 if (TT.isArch64Bit() || TT.isOSWindows() || TT.isOSNaCl()) 118 Ret += "-i64:64"; 119 else if (TT.isOSIAMCU()) 120 Ret += "-i64:32-f64:32"; 121 else 122 Ret += "-f64:32:64"; 123 124 // Some ABIs align long double to 128 bits, others to 32. 125 if (TT.isOSNaCl() || TT.isOSIAMCU()) 126 ; // No f80 127 else if (TT.isArch64Bit() || TT.isOSDarwin()) 128 Ret += "-f80:128"; 129 else 130 Ret += "-f80:32"; 131 132 if (TT.isOSIAMCU()) 133 Ret += "-f128:32"; 134 135 // The registers can hold 8, 16, 32 or, in x86-64, 64 bits. 136 if (TT.isArch64Bit()) 137 Ret += "-n8:16:32:64"; 138 else 139 Ret += "-n8:16:32"; 140 141 // The stack is aligned to 32 bits on some ABIs and 128 bits on others. 142 if ((!TT.isArch64Bit() && TT.isOSWindows()) || TT.isOSIAMCU()) 143 Ret += "-a:0:32-S32"; 144 else 145 Ret += "-S128"; 146 147 return Ret; 148 } 149 150 static Reloc::Model getEffectiveRelocModel(const Triple &TT, 151 bool JIT, 152 Optional<Reloc::Model> RM) { 153 bool is64Bit = TT.getArch() == Triple::x86_64; 154 if (!RM.hasValue()) { 155 // JIT codegen should use static relocations by default, since it's 156 // typically executed in process and not relocatable. 157 if (JIT) 158 return Reloc::Static; 159 160 // Darwin defaults to PIC in 64 bit mode and dynamic-no-pic in 32 bit mode. 161 // Win64 requires rip-rel addressing, thus we force it to PIC. Otherwise we 162 // use static relocation model by default. 163 if (TT.isOSDarwin()) { 164 if (is64Bit) 165 return Reloc::PIC_; 166 return Reloc::DynamicNoPIC; 167 } 168 if (TT.isOSWindows() && is64Bit) 169 return Reloc::PIC_; 170 return Reloc::Static; 171 } 172 173 // ELF and X86-64 don't have a distinct DynamicNoPIC model. DynamicNoPIC 174 // is defined as a model for code which may be used in static or dynamic 175 // executables but not necessarily a shared library. On X86-32 we just 176 // compile in -static mode, in x86-64 we use PIC. 177 if (*RM == Reloc::DynamicNoPIC) { 178 if (is64Bit) 179 return Reloc::PIC_; 180 if (!TT.isOSDarwin()) 181 return Reloc::Static; 182 } 183 184 // If we are on Darwin, disallow static relocation model in X86-64 mode, since 185 // the Mach-O file format doesn't support it. 186 if (*RM == Reloc::Static && TT.isOSDarwin() && is64Bit) 187 return Reloc::PIC_; 188 189 return *RM; 190 } 191 192 static CodeModel::Model getEffectiveX86CodeModel(Optional<CodeModel::Model> CM, 193 bool JIT, bool Is64Bit) { 194 if (CM) { 195 if (*CM == CodeModel::Tiny) 196 report_fatal_error("Target does not support the tiny CodeModel", false); 197 return *CM; 198 } 199 if (JIT) 200 return Is64Bit ? CodeModel::Large : CodeModel::Small; 201 return CodeModel::Small; 202 } 203 204 /// Create an X86 target. 205 /// 206 X86TargetMachine::X86TargetMachine(const Target &T, const Triple &TT, 207 StringRef CPU, StringRef FS, 208 const TargetOptions &Options, 209 Optional<Reloc::Model> RM, 210 Optional<CodeModel::Model> CM, 211 CodeGenOpt::Level OL, bool JIT) 212 : LLVMTargetMachine( 213 T, computeDataLayout(TT), TT, CPU, FS, Options, 214 getEffectiveRelocModel(TT, JIT, RM), 215 getEffectiveX86CodeModel(CM, JIT, TT.getArch() == Triple::x86_64), 216 OL), 217 TLOF(createTLOF(getTargetTriple())) { 218 // On PS4, the "return address" of a 'noreturn' call must still be within 219 // the calling function, and TrapUnreachable is an easy way to get that. 220 if (TT.isPS4() || TT.isOSBinFormatMachO()) { 221 this->Options.TrapUnreachable = true; 222 this->Options.NoTrapAfterNoreturn = TT.isOSBinFormatMachO(); 223 } 224 225 setMachineOutliner(true); 226 227 // x86 supports the debug entry values. 228 setSupportsDebugEntryValues(true); 229 230 initAsmInfo(); 231 } 232 233 X86TargetMachine::~X86TargetMachine() = default; 234 235 const X86Subtarget * 236 X86TargetMachine::getSubtargetImpl(const Function &F) const { 237 Attribute CPUAttr = F.getFnAttribute("target-cpu"); 238 Attribute FSAttr = F.getFnAttribute("target-features"); 239 240 StringRef CPU = !CPUAttr.hasAttribute(Attribute::None) 241 ? CPUAttr.getValueAsString() 242 : (StringRef)TargetCPU; 243 StringRef FS = !FSAttr.hasAttribute(Attribute::None) 244 ? FSAttr.getValueAsString() 245 : (StringRef)TargetFS; 246 247 SmallString<512> Key; 248 Key.reserve(CPU.size() + FS.size()); 249 Key += CPU; 250 Key += FS; 251 252 // FIXME: This is related to the code below to reset the target options, 253 // we need to know whether or not the soft float flag is set on the 254 // function before we can generate a subtarget. We also need to use 255 // it as a key for the subtarget since that can be the only difference 256 // between two functions. 257 bool SoftFloat = 258 F.getFnAttribute("use-soft-float").getValueAsString() == "true"; 259 // If the soft float attribute is set on the function turn on the soft float 260 // subtarget feature. 261 if (SoftFloat) 262 Key += FS.empty() ? "+soft-float" : ",+soft-float"; 263 264 // Keep track of the key width after all features are added so we can extract 265 // the feature string out later. 266 unsigned CPUFSWidth = Key.size(); 267 268 // Extract prefer-vector-width attribute. 269 unsigned PreferVectorWidthOverride = 0; 270 if (F.hasFnAttribute("prefer-vector-width")) { 271 StringRef Val = F.getFnAttribute("prefer-vector-width").getValueAsString(); 272 unsigned Width; 273 if (!Val.getAsInteger(0, Width)) { 274 Key += ",prefer-vector-width="; 275 Key += Val; 276 PreferVectorWidthOverride = Width; 277 } 278 } 279 280 // Extract min-legal-vector-width attribute. 281 unsigned RequiredVectorWidth = UINT32_MAX; 282 if (F.hasFnAttribute("min-legal-vector-width")) { 283 StringRef Val = 284 F.getFnAttribute("min-legal-vector-width").getValueAsString(); 285 unsigned Width; 286 if (!Val.getAsInteger(0, Width)) { 287 Key += ",min-legal-vector-width="; 288 Key += Val; 289 RequiredVectorWidth = Width; 290 } 291 } 292 293 // Extracted here so that we make sure there is backing for the StringRef. If 294 // we assigned earlier, its possible the SmallString reallocated leaving a 295 // dangling StringRef. 296 FS = Key.slice(CPU.size(), CPUFSWidth); 297 298 auto &I = SubtargetMap[Key]; 299 if (!I) { 300 // This needs to be done before we create a new subtarget since any 301 // creation will depend on the TM and the code generation flags on the 302 // function that reside in TargetOptions. 303 resetTargetOptions(F); 304 I = std::make_unique<X86Subtarget>( 305 TargetTriple, CPU, FS, *this, 306 MaybeAlign(Options.StackAlignmentOverride), PreferVectorWidthOverride, 307 RequiredVectorWidth); 308 } 309 return I.get(); 310 } 311 312 //===----------------------------------------------------------------------===// 313 // Command line options for x86 314 //===----------------------------------------------------------------------===// 315 static cl::opt<bool> 316 UseVZeroUpper("x86-use-vzeroupper", cl::Hidden, 317 cl::desc("Minimize AVX to SSE transition penalty"), 318 cl::init(true)); 319 320 //===----------------------------------------------------------------------===// 321 // X86 TTI query. 322 //===----------------------------------------------------------------------===// 323 324 TargetTransformInfo 325 X86TargetMachine::getTargetTransformInfo(const Function &F) { 326 return TargetTransformInfo(X86TTIImpl(this, F)); 327 } 328 329 //===----------------------------------------------------------------------===// 330 // Pass Pipeline Configuration 331 //===----------------------------------------------------------------------===// 332 333 namespace { 334 335 /// X86 Code Generator Pass Configuration Options. 336 class X86PassConfig : public TargetPassConfig { 337 public: 338 X86PassConfig(X86TargetMachine &TM, PassManagerBase &PM) 339 : TargetPassConfig(TM, PM) {} 340 341 X86TargetMachine &getX86TargetMachine() const { 342 return getTM<X86TargetMachine>(); 343 } 344 345 ScheduleDAGInstrs * 346 createMachineScheduler(MachineSchedContext *C) const override { 347 ScheduleDAGMILive *DAG = createGenericSchedLive(C); 348 DAG->addMutation(createX86MacroFusionDAGMutation()); 349 return DAG; 350 } 351 352 ScheduleDAGInstrs * 353 createPostMachineScheduler(MachineSchedContext *C) const override { 354 ScheduleDAGMI *DAG = createGenericSchedPostRA(C); 355 DAG->addMutation(createX86MacroFusionDAGMutation()); 356 return DAG; 357 } 358 359 void addIRPasses() override; 360 bool addInstSelector() override; 361 bool addIRTranslator() override; 362 bool addLegalizeMachineIR() override; 363 bool addRegBankSelect() override; 364 bool addGlobalInstructionSelect() override; 365 bool addILPOpts() override; 366 bool addPreISel() override; 367 void addMachineSSAOptimization() override; 368 void addPreRegAlloc() override; 369 void addPostRegAlloc() override; 370 void addPreEmitPass() override; 371 void addPreEmitPass2() override; 372 void addPreSched2() override; 373 374 std::unique_ptr<CSEConfigBase> getCSEConfig() const override; 375 }; 376 377 class X86ExecutionDomainFix : public ExecutionDomainFix { 378 public: 379 static char ID; 380 X86ExecutionDomainFix() : ExecutionDomainFix(ID, X86::VR128XRegClass) {} 381 StringRef getPassName() const override { 382 return "X86 Execution Dependency Fix"; 383 } 384 }; 385 char X86ExecutionDomainFix::ID; 386 387 } // end anonymous namespace 388 389 INITIALIZE_PASS_BEGIN(X86ExecutionDomainFix, "x86-execution-domain-fix", 390 "X86 Execution Domain Fix", false, false) 391 INITIALIZE_PASS_DEPENDENCY(ReachingDefAnalysis) 392 INITIALIZE_PASS_END(X86ExecutionDomainFix, "x86-execution-domain-fix", 393 "X86 Execution Domain Fix", false, false) 394 395 TargetPassConfig *X86TargetMachine::createPassConfig(PassManagerBase &PM) { 396 return new X86PassConfig(*this, PM); 397 } 398 399 void X86PassConfig::addIRPasses() { 400 addPass(createAtomicExpandPass()); 401 402 TargetPassConfig::addIRPasses(); 403 404 if (TM->getOptLevel() != CodeGenOpt::None) { 405 addPass(createInterleavedAccessPass()); 406 addPass(createX86PartialReductionPass()); 407 } 408 409 // Add passes that handle indirect branch removal and insertion of a retpoline 410 // thunk. These will be a no-op unless a function subtarget has the retpoline 411 // feature enabled. 412 addPass(createIndirectBrExpandPass()); 413 414 // Add Control Flow Guard checks. 415 const Triple &TT = TM->getTargetTriple(); 416 if (TT.isOSWindows()) { 417 if (TT.getArch() == Triple::x86_64) { 418 addPass(createCFGuardDispatchPass()); 419 } else { 420 addPass(createCFGuardCheckPass()); 421 } 422 } 423 } 424 425 bool X86PassConfig::addInstSelector() { 426 // Install an instruction selector. 427 addPass(createX86ISelDag(getX86TargetMachine(), getOptLevel())); 428 429 // For ELF, cleanup any local-dynamic TLS accesses. 430 if (TM->getTargetTriple().isOSBinFormatELF() && 431 getOptLevel() != CodeGenOpt::None) 432 addPass(createCleanupLocalDynamicTLSPass()); 433 434 addPass(createX86GlobalBaseRegPass()); 435 return false; 436 } 437 438 bool X86PassConfig::addIRTranslator() { 439 addPass(new IRTranslator()); 440 return false; 441 } 442 443 bool X86PassConfig::addLegalizeMachineIR() { 444 addPass(new Legalizer()); 445 return false; 446 } 447 448 bool X86PassConfig::addRegBankSelect() { 449 addPass(new RegBankSelect()); 450 return false; 451 } 452 453 bool X86PassConfig::addGlobalInstructionSelect() { 454 addPass(new InstructionSelect()); 455 return false; 456 } 457 458 bool X86PassConfig::addILPOpts() { 459 if (EnableCondBrFoldingPass) 460 addPass(createX86CondBrFolding()); 461 addPass(&EarlyIfConverterID); 462 if (EnableMachineCombinerPass) 463 addPass(&MachineCombinerID); 464 addPass(createX86CmovConverterPass()); 465 return true; 466 } 467 468 bool X86PassConfig::addPreISel() { 469 // Only add this pass for 32-bit x86 Windows. 470 const Triple &TT = TM->getTargetTriple(); 471 if (TT.isOSWindows() && TT.getArch() == Triple::x86) 472 addPass(createX86WinEHStatePass()); 473 return true; 474 } 475 476 void X86PassConfig::addPreRegAlloc() { 477 if (getOptLevel() != CodeGenOpt::None) { 478 addPass(&LiveRangeShrinkID); 479 addPass(createX86FixupSetCC()); 480 addPass(createX86OptimizeLEAs()); 481 addPass(createX86CallFrameOptimization()); 482 addPass(createX86AvoidStoreForwardingBlocks()); 483 } 484 485 addPass(createX86SpeculativeLoadHardeningPass()); 486 addPass(createX86FlagsCopyLoweringPass()); 487 addPass(createX86WinAllocaExpander()); 488 } 489 void X86PassConfig::addMachineSSAOptimization() { 490 addPass(createX86DomainReassignmentPass()); 491 TargetPassConfig::addMachineSSAOptimization(); 492 } 493 494 void X86PassConfig::addPostRegAlloc() { 495 addPass(createX86FloatingPointStackifierPass()); 496 } 497 498 void X86PassConfig::addPreSched2() { addPass(createX86ExpandPseudoPass()); } 499 500 void X86PassConfig::addPreEmitPass() { 501 if (getOptLevel() != CodeGenOpt::None) { 502 addPass(new X86ExecutionDomainFix()); 503 addPass(createBreakFalseDeps()); 504 } 505 506 addPass(createX86IndirectBranchTrackingPass()); 507 508 if (UseVZeroUpper) 509 addPass(createX86IssueVZeroUpperPass()); 510 511 if (getOptLevel() != CodeGenOpt::None) { 512 addPass(createX86FixupBWInsts()); 513 addPass(createX86PadShortFunctions()); 514 addPass(createX86FixupLEAs()); 515 addPass(createX86EvexToVexInsts()); 516 } 517 addPass(createX86DiscriminateMemOpsPass()); 518 addPass(createX86InsertPrefetchPass()); 519 addPass(createX86InsertX87waitPass()); 520 } 521 522 void X86PassConfig::addPreEmitPass2() { 523 const Triple &TT = TM->getTargetTriple(); 524 const MCAsmInfo *MAI = TM->getMCAsmInfo(); 525 526 addPass(createX86RetpolineThunksPass()); 527 528 // Insert extra int3 instructions after trailing call instructions to avoid 529 // issues in the unwinder. 530 if (TT.isOSWindows() && TT.getArch() == Triple::x86_64) 531 addPass(createX86AvoidTrailingCallPass()); 532 533 // Verify basic block incoming and outgoing cfa offset and register values and 534 // correct CFA calculation rule where needed by inserting appropriate CFI 535 // instructions. 536 if (!TT.isOSDarwin() && 537 (!TT.isOSWindows() || 538 MAI->getExceptionHandlingType() == ExceptionHandling::DwarfCFI)) 539 addPass(createCFIInstrInserter()); 540 // Identify valid longjmp targets for Windows Control Flow Guard. 541 if (TT.isOSWindows()) 542 addPass(createCFGuardLongjmpPass()); 543 } 544 545 std::unique_ptr<CSEConfigBase> X86PassConfig::getCSEConfig() const { 546 return getStandardCSEConfigForOpt(TM->getOptLevel()); 547 } 548