1 //===-- TargetPassConfig.cpp - Target independent code generation passes --===// 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 interfaces to access the target independent code 11 // generation passes provided by the LLVM backend. 12 // 13 //===---------------------------------------------------------------------===// 14 15 #include "llvm/CodeGen/TargetPassConfig.h" 16 17 #include "llvm/Analysis/BasicAliasAnalysis.h" 18 #include "llvm/Analysis/CFLAliasAnalysis.h" 19 #include "llvm/Analysis/CallGraphSCCPass.h" 20 #include "llvm/Analysis/Passes.h" 21 #include "llvm/Analysis/ScopedNoAliasAA.h" 22 #include "llvm/Analysis/TypeBasedAliasAnalysis.h" 23 #include "llvm/CodeGen/MachineFunctionPass.h" 24 #include "llvm/CodeGen/RegAllocRegistry.h" 25 #include "llvm/CodeGen/RegisterUsageInfo.h" 26 #include "llvm/IR/IRPrintingPasses.h" 27 #include "llvm/IR/LegacyPassManager.h" 28 #include "llvm/IR/Verifier.h" 29 #include "llvm/MC/MCAsmInfo.h" 30 #include "llvm/Support/CommandLine.h" 31 #include "llvm/Support/Debug.h" 32 #include "llvm/Support/ErrorHandling.h" 33 #include "llvm/Support/raw_ostream.h" 34 #include "llvm/Target/TargetMachine.h" 35 #include "llvm/Transforms/Instrumentation.h" 36 #include "llvm/Transforms/Scalar.h" 37 #include "llvm/Transforms/Utils/SymbolRewriter.h" 38 39 using namespace llvm; 40 41 static cl::opt<bool> DisablePostRA("disable-post-ra", cl::Hidden, 42 cl::desc("Disable Post Regalloc")); 43 static cl::opt<bool> DisableBranchFold("disable-branch-fold", cl::Hidden, 44 cl::desc("Disable branch folding")); 45 static cl::opt<bool> DisableTailDuplicate("disable-tail-duplicate", cl::Hidden, 46 cl::desc("Disable tail duplication")); 47 static cl::opt<bool> DisableEarlyTailDup("disable-early-taildup", cl::Hidden, 48 cl::desc("Disable pre-register allocation tail duplication")); 49 static cl::opt<bool> DisableBlockPlacement("disable-block-placement", 50 cl::Hidden, cl::desc("Disable probability-driven block placement")); 51 static cl::opt<bool> EnableBlockPlacementStats("enable-block-placement-stats", 52 cl::Hidden, cl::desc("Collect probability-driven block placement stats")); 53 static cl::opt<bool> DisableSSC("disable-ssc", cl::Hidden, 54 cl::desc("Disable Stack Slot Coloring")); 55 static cl::opt<bool> DisableMachineDCE("disable-machine-dce", cl::Hidden, 56 cl::desc("Disable Machine Dead Code Elimination")); 57 static cl::opt<bool> DisableEarlyIfConversion("disable-early-ifcvt", cl::Hidden, 58 cl::desc("Disable Early If-conversion")); 59 static cl::opt<bool> DisableMachineLICM("disable-machine-licm", cl::Hidden, 60 cl::desc("Disable Machine LICM")); 61 static cl::opt<bool> DisableMachineCSE("disable-machine-cse", cl::Hidden, 62 cl::desc("Disable Machine Common Subexpression Elimination")); 63 static cl::opt<cl::boolOrDefault> OptimizeRegAlloc( 64 "optimize-regalloc", cl::Hidden, 65 cl::desc("Enable optimized register allocation compilation path.")); 66 static cl::opt<bool> DisablePostRAMachineLICM("disable-postra-machine-licm", 67 cl::Hidden, 68 cl::desc("Disable Machine LICM")); 69 static cl::opt<bool> DisableMachineSink("disable-machine-sink", cl::Hidden, 70 cl::desc("Disable Machine Sinking")); 71 static cl::opt<bool> DisableLSR("disable-lsr", cl::Hidden, 72 cl::desc("Disable Loop Strength Reduction Pass")); 73 static cl::opt<bool> DisableConstantHoisting("disable-constant-hoisting", 74 cl::Hidden, cl::desc("Disable ConstantHoisting")); 75 static cl::opt<bool> DisableCGP("disable-cgp", cl::Hidden, 76 cl::desc("Disable Codegen Prepare")); 77 static cl::opt<bool> DisableCopyProp("disable-copyprop", cl::Hidden, 78 cl::desc("Disable Copy Propagation pass")); 79 static cl::opt<bool> DisablePartialLibcallInlining("disable-partial-libcall-inlining", 80 cl::Hidden, cl::desc("Disable Partial Libcall Inlining")); 81 static cl::opt<bool> EnableImplicitNullChecks( 82 "enable-implicit-null-checks", 83 cl::desc("Fold null checks into faulting memory operations"), 84 cl::init(false)); 85 static cl::opt<bool> PrintLSR("print-lsr-output", cl::Hidden, 86 cl::desc("Print LLVM IR produced by the loop-reduce pass")); 87 static cl::opt<bool> PrintISelInput("print-isel-input", cl::Hidden, 88 cl::desc("Print LLVM IR input to isel pass")); 89 static cl::opt<bool> PrintGCInfo("print-gc", cl::Hidden, 90 cl::desc("Dump garbage collector data")); 91 static cl::opt<bool> VerifyMachineCode("verify-machineinstrs", cl::Hidden, 92 cl::desc("Verify generated machine code"), 93 cl::init(false), 94 cl::ZeroOrMore); 95 96 static cl::opt<std::string> 97 PrintMachineInstrs("print-machineinstrs", cl::ValueOptional, 98 cl::desc("Print machine instrs"), 99 cl::value_desc("pass-name"), cl::init("option-unspecified")); 100 101 // Temporary option to allow experimenting with MachineScheduler as a post-RA 102 // scheduler. Targets can "properly" enable this with 103 // substitutePass(&PostRASchedulerID, &PostMachineSchedulerID). 104 // Targets can return true in targetSchedulesPostRAScheduling() and 105 // insert a PostRA scheduling pass wherever it wants. 106 cl::opt<bool> MISchedPostRA("misched-postra", cl::Hidden, 107 cl::desc("Run MachineScheduler post regalloc (independent of preRA sched)")); 108 109 // Experimental option to run live interval analysis early. 110 static cl::opt<bool> EarlyLiveIntervals("early-live-intervals", cl::Hidden, 111 cl::desc("Run live interval analysis earlier in the pipeline")); 112 113 static cl::opt<bool> UseCFLAA("use-cfl-aa-in-codegen", 114 cl::init(false), cl::Hidden, 115 cl::desc("Enable the new, experimental CFL alias analysis in CodeGen")); 116 117 cl::opt<bool> UseIPRA("enable-ipra", cl::init(false), cl::Hidden, 118 cl::desc("Enable interprocedural register allocation " 119 "to reduce load/store at procedure calls.")); 120 121 /// Allow standard passes to be disabled by command line options. This supports 122 /// simple binary flags that either suppress the pass or do nothing. 123 /// i.e. -disable-mypass=false has no effect. 124 /// These should be converted to boolOrDefault in order to use applyOverride. 125 static IdentifyingPassPtr applyDisable(IdentifyingPassPtr PassID, 126 bool Override) { 127 if (Override) 128 return IdentifyingPassPtr(); 129 return PassID; 130 } 131 132 /// Allow standard passes to be disabled by the command line, regardless of who 133 /// is adding the pass. 134 /// 135 /// StandardID is the pass identified in the standard pass pipeline and provided 136 /// to addPass(). It may be a target-specific ID in the case that the target 137 /// directly adds its own pass, but in that case we harmlessly fall through. 138 /// 139 /// TargetID is the pass that the target has configured to override StandardID. 140 /// 141 /// StandardID may be a pseudo ID. In that case TargetID is the name of the real 142 /// pass to run. This allows multiple options to control a single pass depending 143 /// on where in the pipeline that pass is added. 144 static IdentifyingPassPtr overridePass(AnalysisID StandardID, 145 IdentifyingPassPtr TargetID) { 146 if (StandardID == &PostRASchedulerID) 147 return applyDisable(TargetID, DisablePostRA); 148 149 if (StandardID == &BranchFolderPassID) 150 return applyDisable(TargetID, DisableBranchFold); 151 152 if (StandardID == &TailDuplicateID) 153 return applyDisable(TargetID, DisableTailDuplicate); 154 155 if (StandardID == &TargetPassConfig::EarlyTailDuplicateID) 156 return applyDisable(TargetID, DisableEarlyTailDup); 157 158 if (StandardID == &MachineBlockPlacementID) 159 return applyDisable(TargetID, DisableBlockPlacement); 160 161 if (StandardID == &StackSlotColoringID) 162 return applyDisable(TargetID, DisableSSC); 163 164 if (StandardID == &DeadMachineInstructionElimID) 165 return applyDisable(TargetID, DisableMachineDCE); 166 167 if (StandardID == &EarlyIfConverterID) 168 return applyDisable(TargetID, DisableEarlyIfConversion); 169 170 if (StandardID == &MachineLICMID) 171 return applyDisable(TargetID, DisableMachineLICM); 172 173 if (StandardID == &MachineCSEID) 174 return applyDisable(TargetID, DisableMachineCSE); 175 176 if (StandardID == &TargetPassConfig::PostRAMachineLICMID) 177 return applyDisable(TargetID, DisablePostRAMachineLICM); 178 179 if (StandardID == &MachineSinkingID) 180 return applyDisable(TargetID, DisableMachineSink); 181 182 if (StandardID == &MachineCopyPropagationID) 183 return applyDisable(TargetID, DisableCopyProp); 184 185 return TargetID; 186 } 187 188 //===---------------------------------------------------------------------===// 189 /// TargetPassConfig 190 //===---------------------------------------------------------------------===// 191 192 INITIALIZE_PASS(TargetPassConfig, "targetpassconfig", 193 "Target Pass Configuration", false, false) 194 char TargetPassConfig::ID = 0; 195 196 // Pseudo Pass IDs. 197 char TargetPassConfig::EarlyTailDuplicateID = 0; 198 char TargetPassConfig::PostRAMachineLICMID = 0; 199 200 namespace { 201 struct InsertedPass { 202 AnalysisID TargetPassID; 203 IdentifyingPassPtr InsertedPassID; 204 bool VerifyAfter; 205 bool PrintAfter; 206 207 InsertedPass(AnalysisID TargetPassID, IdentifyingPassPtr InsertedPassID, 208 bool VerifyAfter, bool PrintAfter) 209 : TargetPassID(TargetPassID), InsertedPassID(InsertedPassID), 210 VerifyAfter(VerifyAfter), PrintAfter(PrintAfter) {} 211 212 Pass *getInsertedPass() const { 213 assert(InsertedPassID.isValid() && "Illegal Pass ID!"); 214 if (InsertedPassID.isInstance()) 215 return InsertedPassID.getInstance(); 216 Pass *NP = Pass::createPass(InsertedPassID.getID()); 217 assert(NP && "Pass ID not registered"); 218 return NP; 219 } 220 }; 221 } 222 223 namespace llvm { 224 class PassConfigImpl { 225 public: 226 // List of passes explicitly substituted by this target. Normally this is 227 // empty, but it is a convenient way to suppress or replace specific passes 228 // that are part of a standard pass pipeline without overridding the entire 229 // pipeline. This mechanism allows target options to inherit a standard pass's 230 // user interface. For example, a target may disable a standard pass by 231 // default by substituting a pass ID of zero, and the user may still enable 232 // that standard pass with an explicit command line option. 233 DenseMap<AnalysisID,IdentifyingPassPtr> TargetPasses; 234 235 /// Store the pairs of <AnalysisID, AnalysisID> of which the second pass 236 /// is inserted after each instance of the first one. 237 SmallVector<InsertedPass, 4> InsertedPasses; 238 }; 239 } // namespace llvm 240 241 // Out of line virtual method. 242 TargetPassConfig::~TargetPassConfig() { 243 delete Impl; 244 } 245 246 // Out of line constructor provides default values for pass options and 247 // registers all common codegen passes. 248 TargetPassConfig::TargetPassConfig(TargetMachine *tm, PassManagerBase &pm) 249 : ImmutablePass(ID), PM(&pm), StartBefore(nullptr), StartAfter(nullptr), 250 StopAfter(nullptr), Started(true), Stopped(false), 251 AddingMachinePasses(false), TM(tm), Impl(nullptr), Initialized(false), 252 DisableVerify(false), EnableTailMerge(true) { 253 254 Impl = new PassConfigImpl(); 255 256 // Register all target independent codegen passes to activate their PassIDs, 257 // including this pass itself. 258 initializeCodeGen(*PassRegistry::getPassRegistry()); 259 260 // Also register alias analysis passes required by codegen passes. 261 initializeBasicAAWrapperPassPass(*PassRegistry::getPassRegistry()); 262 initializeAAResultsWrapperPassPass(*PassRegistry::getPassRegistry()); 263 264 // Substitute Pseudo Pass IDs for real ones. 265 substitutePass(&EarlyTailDuplicateID, &TailDuplicateID); 266 substitutePass(&PostRAMachineLICMID, &MachineLICMID); 267 268 if (StringRef(PrintMachineInstrs.getValue()).equals("")) 269 TM->Options.PrintMachineCode = true; 270 } 271 272 CodeGenOpt::Level TargetPassConfig::getOptLevel() const { 273 return TM->getOptLevel(); 274 } 275 276 /// Insert InsertedPassID pass after TargetPassID. 277 void TargetPassConfig::insertPass(AnalysisID TargetPassID, 278 IdentifyingPassPtr InsertedPassID, 279 bool VerifyAfter, bool PrintAfter) { 280 assert(((!InsertedPassID.isInstance() && 281 TargetPassID != InsertedPassID.getID()) || 282 (InsertedPassID.isInstance() && 283 TargetPassID != InsertedPassID.getInstance()->getPassID())) && 284 "Insert a pass after itself!"); 285 Impl->InsertedPasses.emplace_back(TargetPassID, InsertedPassID, VerifyAfter, 286 PrintAfter); 287 } 288 289 /// createPassConfig - Create a pass configuration object to be used by 290 /// addPassToEmitX methods for generating a pipeline of CodeGen passes. 291 /// 292 /// Targets may override this to extend TargetPassConfig. 293 TargetPassConfig *LLVMTargetMachine::createPassConfig(PassManagerBase &PM) { 294 return new TargetPassConfig(this, PM); 295 } 296 297 TargetPassConfig::TargetPassConfig() 298 : ImmutablePass(ID), PM(nullptr) { 299 llvm_unreachable("TargetPassConfig should not be constructed on-the-fly"); 300 } 301 302 // Helper to verify the analysis is really immutable. 303 void TargetPassConfig::setOpt(bool &Opt, bool Val) { 304 assert(!Initialized && "PassConfig is immutable"); 305 Opt = Val; 306 } 307 308 void TargetPassConfig::substitutePass(AnalysisID StandardID, 309 IdentifyingPassPtr TargetID) { 310 Impl->TargetPasses[StandardID] = TargetID; 311 } 312 313 IdentifyingPassPtr TargetPassConfig::getPassSubstitution(AnalysisID ID) const { 314 DenseMap<AnalysisID, IdentifyingPassPtr>::const_iterator 315 I = Impl->TargetPasses.find(ID); 316 if (I == Impl->TargetPasses.end()) 317 return ID; 318 return I->second; 319 } 320 321 bool TargetPassConfig::isPassSubstitutedOrOverridden(AnalysisID ID) const { 322 IdentifyingPassPtr TargetID = getPassSubstitution(ID); 323 IdentifyingPassPtr FinalPtr = overridePass(ID, TargetID); 324 return !FinalPtr.isValid() || FinalPtr.isInstance() || 325 FinalPtr.getID() != ID; 326 } 327 328 /// Add a pass to the PassManager if that pass is supposed to be run. If the 329 /// Started/Stopped flags indicate either that the compilation should start at 330 /// a later pass or that it should stop after an earlier pass, then do not add 331 /// the pass. Finally, compare the current pass against the StartAfter 332 /// and StopAfter options and change the Started/Stopped flags accordingly. 333 void TargetPassConfig::addPass(Pass *P, bool verifyAfter, bool printAfter) { 334 assert(!Initialized && "PassConfig is immutable"); 335 336 // Cache the Pass ID here in case the pass manager finds this pass is 337 // redundant with ones already scheduled / available, and deletes it. 338 // Fundamentally, once we add the pass to the manager, we no longer own it 339 // and shouldn't reference it. 340 AnalysisID PassID = P->getPassID(); 341 342 if (StartBefore == PassID) 343 Started = true; 344 if (Started && !Stopped) { 345 std::string Banner; 346 // Construct banner message before PM->add() as that may delete the pass. 347 if (AddingMachinePasses && (printAfter || verifyAfter)) 348 Banner = std::string("After ") + std::string(P->getPassName()); 349 PM->add(P); 350 if (AddingMachinePasses) { 351 if (printAfter) 352 addPrintPass(Banner); 353 if (verifyAfter) 354 addVerifyPass(Banner); 355 } 356 357 // Add the passes after the pass P if there is any. 358 for (auto IP : Impl->InsertedPasses) { 359 if (IP.TargetPassID == PassID) 360 addPass(IP.getInsertedPass(), IP.VerifyAfter, IP.PrintAfter); 361 } 362 } else { 363 delete P; 364 } 365 if (StopAfter == PassID) 366 Stopped = true; 367 if (StartAfter == PassID) 368 Started = true; 369 if (Stopped && !Started) 370 report_fatal_error("Cannot stop compilation after pass that is not run"); 371 } 372 373 /// Add a CodeGen pass at this point in the pipeline after checking for target 374 /// and command line overrides. 375 /// 376 /// addPass cannot return a pointer to the pass instance because is internal the 377 /// PassManager and the instance we create here may already be freed. 378 AnalysisID TargetPassConfig::addPass(AnalysisID PassID, bool verifyAfter, 379 bool printAfter) { 380 IdentifyingPassPtr TargetID = getPassSubstitution(PassID); 381 IdentifyingPassPtr FinalPtr = overridePass(PassID, TargetID); 382 if (!FinalPtr.isValid()) 383 return nullptr; 384 385 Pass *P; 386 if (FinalPtr.isInstance()) 387 P = FinalPtr.getInstance(); 388 else { 389 P = Pass::createPass(FinalPtr.getID()); 390 if (!P) 391 llvm_unreachable("Pass ID not registered"); 392 } 393 AnalysisID FinalID = P->getPassID(); 394 addPass(P, verifyAfter, printAfter); // Ends the lifetime of P. 395 396 return FinalID; 397 } 398 399 void TargetPassConfig::printAndVerify(const std::string &Banner) { 400 addPrintPass(Banner); 401 addVerifyPass(Banner); 402 } 403 404 void TargetPassConfig::addPrintPass(const std::string &Banner) { 405 if (TM->shouldPrintMachineCode()) 406 PM->add(createMachineFunctionPrinterPass(dbgs(), Banner)); 407 } 408 409 void TargetPassConfig::addVerifyPass(const std::string &Banner) { 410 if (VerifyMachineCode) 411 PM->add(createMachineVerifierPass(Banner)); 412 } 413 414 /// Add common target configurable passes that perform LLVM IR to IR transforms 415 /// following machine independent optimization. 416 void TargetPassConfig::addIRPasses() { 417 // Basic AliasAnalysis support. 418 // Add TypeBasedAliasAnalysis before BasicAliasAnalysis so that 419 // BasicAliasAnalysis wins if they disagree. This is intended to help 420 // support "obvious" type-punning idioms. 421 if (UseCFLAA) 422 addPass(createCFLAAWrapperPass()); 423 addPass(createTypeBasedAAWrapperPass()); 424 addPass(createScopedNoAliasAAWrapperPass()); 425 addPass(createBasicAAWrapperPass()); 426 427 // Before running any passes, run the verifier to determine if the input 428 // coming from the front-end and/or optimizer is valid. 429 if (!DisableVerify) 430 addPass(createVerifierPass()); 431 432 // Run loop strength reduction before anything else. 433 if (getOptLevel() != CodeGenOpt::None && !DisableLSR) { 434 addPass(createLoopStrengthReducePass()); 435 if (PrintLSR) 436 addPass(createPrintFunctionPass(dbgs(), "\n\n*** Code after LSR ***\n")); 437 } 438 439 // Run GC lowering passes for builtin collectors 440 // TODO: add a pass insertion point here 441 addPass(createGCLoweringPass()); 442 addPass(createShadowStackGCLoweringPass()); 443 444 // Make sure that no unreachable blocks are instruction selected. 445 addPass(createUnreachableBlockEliminationPass()); 446 447 // Prepare expensive constants for SelectionDAG. 448 if (getOptLevel() != CodeGenOpt::None && !DisableConstantHoisting) 449 addPass(createConstantHoistingPass()); 450 451 if (getOptLevel() != CodeGenOpt::None && !DisablePartialLibcallInlining) 452 addPass(createPartiallyInlineLibCallsPass()); 453 } 454 455 /// Turn exception handling constructs into something the code generators can 456 /// handle. 457 void TargetPassConfig::addPassesToHandleExceptions() { 458 switch (TM->getMCAsmInfo()->getExceptionHandlingType()) { 459 case ExceptionHandling::SjLj: 460 // SjLj piggy-backs on dwarf for this bit. The cleanups done apply to both 461 // Dwarf EH prepare needs to be run after SjLj prepare. Otherwise, 462 // catch info can get misplaced when a selector ends up more than one block 463 // removed from the parent invoke(s). This could happen when a landing 464 // pad is shared by multiple invokes and is also a target of a normal 465 // edge from elsewhere. 466 addPass(createSjLjEHPreparePass()); 467 // FALLTHROUGH 468 case ExceptionHandling::DwarfCFI: 469 case ExceptionHandling::ARM: 470 addPass(createDwarfEHPass(TM)); 471 break; 472 case ExceptionHandling::WinEH: 473 // We support using both GCC-style and MSVC-style exceptions on Windows, so 474 // add both preparation passes. Each pass will only actually run if it 475 // recognizes the personality function. 476 addPass(createWinEHPass(TM)); 477 addPass(createDwarfEHPass(TM)); 478 break; 479 case ExceptionHandling::None: 480 addPass(createLowerInvokePass()); 481 482 // The lower invoke pass may create unreachable code. Remove it. 483 addPass(createUnreachableBlockEliminationPass()); 484 break; 485 } 486 } 487 488 /// Add pass to prepare the LLVM IR for code generation. This should be done 489 /// before exception handling preparation passes. 490 void TargetPassConfig::addCodeGenPrepare() { 491 if (getOptLevel() != CodeGenOpt::None && !DisableCGP) 492 addPass(createCodeGenPreparePass(TM)); 493 addPass(createRewriteSymbolsPass()); 494 } 495 496 /// Add common passes that perform LLVM IR to IR transforms in preparation for 497 /// instruction selection. 498 void TargetPassConfig::addISelPrepare() { 499 addPreISel(); 500 501 // Force codegen to run according to the callgraph. 502 if (UseIPRA) 503 addPass(new DummyCGSCCPass); 504 505 // Add both the safe stack and the stack protection passes: each of them will 506 // only protect functions that have corresponding attributes. 507 addPass(createSafeStackPass(TM)); 508 addPass(createStackProtectorPass(TM)); 509 510 if (PrintISelInput) 511 addPass(createPrintFunctionPass( 512 dbgs(), "\n\n*** Final LLVM Code input to ISel ***\n")); 513 514 // All passes which modify the LLVM IR are now complete; run the verifier 515 // to ensure that the IR is valid. 516 if (!DisableVerify) 517 addPass(createVerifierPass()); 518 } 519 520 /// Add the complete set of target-independent postISel code generator passes. 521 /// 522 /// This can be read as the standard order of major LLVM CodeGen stages. Stages 523 /// with nontrivial configuration or multiple passes are broken out below in 524 /// add%Stage routines. 525 /// 526 /// Any TargetPassConfig::addXX routine may be overriden by the Target. The 527 /// addPre/Post methods with empty header implementations allow injecting 528 /// target-specific fixups just before or after major stages. Additionally, 529 /// targets have the flexibility to change pass order within a stage by 530 /// overriding default implementation of add%Stage routines below. Each 531 /// technique has maintainability tradeoffs because alternate pass orders are 532 /// not well supported. addPre/Post works better if the target pass is easily 533 /// tied to a common pass. But if it has subtle dependencies on multiple passes, 534 /// the target should override the stage instead. 535 /// 536 /// TODO: We could use a single addPre/Post(ID) hook to allow pass injection 537 /// before/after any target-independent pass. But it's currently overkill. 538 void TargetPassConfig::addMachinePasses() { 539 AddingMachinePasses = true; 540 541 if (UseIPRA) 542 addPass(createRegUsageInfoPropPass()); 543 544 // Insert a machine instr printer pass after the specified pass. 545 if (!StringRef(PrintMachineInstrs.getValue()).equals("") && 546 !StringRef(PrintMachineInstrs.getValue()).equals("option-unspecified")) { 547 const PassRegistry *PR = PassRegistry::getPassRegistry(); 548 const PassInfo *TPI = PR->getPassInfo(PrintMachineInstrs.getValue()); 549 const PassInfo *IPI = PR->getPassInfo(StringRef("machineinstr-printer")); 550 assert (TPI && IPI && "Pass ID not registered!"); 551 const char *TID = (const char *)(TPI->getTypeInfo()); 552 const char *IID = (const char *)(IPI->getTypeInfo()); 553 insertPass(TID, IID); 554 } 555 556 // Print the instruction selected machine code... 557 printAndVerify("After Instruction Selection"); 558 559 // Expand pseudo-instructions emitted by ISel. 560 addPass(&ExpandISelPseudosID); 561 562 // Add passes that optimize machine instructions in SSA form. 563 if (getOptLevel() != CodeGenOpt::None) { 564 addMachineSSAOptimization(); 565 } else { 566 // If the target requests it, assign local variables to stack slots relative 567 // to one another and simplify frame index references where possible. 568 addPass(&LocalStackSlotAllocationID, false); 569 } 570 571 // Run pre-ra passes. 572 addPreRegAlloc(); 573 574 // Run register allocation and passes that are tightly coupled with it, 575 // including phi elimination and scheduling. 576 if (getOptimizeRegAlloc()) 577 addOptimizedRegAlloc(createRegAllocPass(true)); 578 else 579 addFastRegAlloc(createRegAllocPass(false)); 580 581 // Run post-ra passes. 582 addPostRegAlloc(); 583 584 // Insert prolog/epilog code. Eliminate abstract frame index references... 585 if (getOptLevel() != CodeGenOpt::None) 586 addPass(&ShrinkWrapID); 587 588 // Prolog/Epilog inserter needs a TargetMachine to instantiate. But only 589 // do so if it hasn't been disabled, substituted, or overridden. 590 if (!isPassSubstitutedOrOverridden(&PrologEpilogCodeInserterID)) 591 addPass(createPrologEpilogInserterPass(TM)); 592 593 /// Add passes that optimize machine instructions after register allocation. 594 if (getOptLevel() != CodeGenOpt::None) 595 addMachineLateOptimization(); 596 597 // Expand pseudo instructions before second scheduling pass. 598 addPass(&ExpandPostRAPseudosID); 599 600 // Run pre-sched2 passes. 601 addPreSched2(); 602 603 if (EnableImplicitNullChecks) 604 addPass(&ImplicitNullChecksID); 605 606 // Second pass scheduler. 607 // Let Target optionally insert this pass by itself at some other 608 // point. 609 if (getOptLevel() != CodeGenOpt::None && 610 !TM->targetSchedulesPostRAScheduling()) { 611 if (MISchedPostRA) 612 addPass(&PostMachineSchedulerID); 613 else 614 addPass(&PostRASchedulerID); 615 } 616 617 // GC 618 if (addGCPasses()) { 619 if (PrintGCInfo) 620 addPass(createGCInfoPrinter(dbgs()), false, false); 621 } 622 623 // Basic block placement. 624 if (getOptLevel() != CodeGenOpt::None) 625 addBlockPlacement(); 626 627 addPreEmitPass(); 628 629 if (UseIPRA) 630 // Collect register usage information and produce a register mask of 631 // clobbered registers, to be used to optimize call sites. 632 addPass(createRegUsageInfoCollector()); 633 634 addPass(&FuncletLayoutID, false); 635 636 addPass(&StackMapLivenessID, false); 637 addPass(&LiveDebugValuesID, false); 638 639 addPass(&PatchableFunctionID, false); 640 641 AddingMachinePasses = false; 642 } 643 644 /// Add passes that optimize machine instructions in SSA form. 645 void TargetPassConfig::addMachineSSAOptimization() { 646 // Pre-ra tail duplication. 647 addPass(&EarlyTailDuplicateID); 648 649 // Optimize PHIs before DCE: removing dead PHI cycles may make more 650 // instructions dead. 651 addPass(&OptimizePHIsID, false); 652 653 // This pass merges large allocas. StackSlotColoring is a different pass 654 // which merges spill slots. 655 addPass(&StackColoringID, false); 656 657 // If the target requests it, assign local variables to stack slots relative 658 // to one another and simplify frame index references where possible. 659 addPass(&LocalStackSlotAllocationID, false); 660 661 // With optimization, dead code should already be eliminated. However 662 // there is one known exception: lowered code for arguments that are only 663 // used by tail calls, where the tail calls reuse the incoming stack 664 // arguments directly (see t11 in test/CodeGen/X86/sibcall.ll). 665 addPass(&DeadMachineInstructionElimID); 666 667 // Allow targets to insert passes that improve instruction level parallelism, 668 // like if-conversion. Such passes will typically need dominator trees and 669 // loop info, just like LICM and CSE below. 670 addILPOpts(); 671 672 addPass(&MachineLICMID, false); 673 addPass(&MachineCSEID, false); 674 addPass(&MachineSinkingID); 675 676 addPass(&PeepholeOptimizerID); 677 // Clean-up the dead code that may have been generated by peephole 678 // rewriting. 679 addPass(&DeadMachineInstructionElimID); 680 } 681 682 //===---------------------------------------------------------------------===// 683 /// Register Allocation Pass Configuration 684 //===---------------------------------------------------------------------===// 685 686 bool TargetPassConfig::getOptimizeRegAlloc() const { 687 switch (OptimizeRegAlloc) { 688 case cl::BOU_UNSET: return getOptLevel() != CodeGenOpt::None; 689 case cl::BOU_TRUE: return true; 690 case cl::BOU_FALSE: return false; 691 } 692 llvm_unreachable("Invalid optimize-regalloc state"); 693 } 694 695 /// RegisterRegAlloc's global Registry tracks allocator registration. 696 MachinePassRegistry RegisterRegAlloc::Registry; 697 698 /// A dummy default pass factory indicates whether the register allocator is 699 /// overridden on the command line. 700 static FunctionPass *useDefaultRegisterAllocator() { return nullptr; } 701 static RegisterRegAlloc 702 defaultRegAlloc("default", 703 "pick register allocator based on -O option", 704 useDefaultRegisterAllocator); 705 706 /// -regalloc=... command line option. 707 static cl::opt<RegisterRegAlloc::FunctionPassCtor, false, 708 RegisterPassParser<RegisterRegAlloc> > 709 RegAlloc("regalloc", 710 cl::init(&useDefaultRegisterAllocator), 711 cl::desc("Register allocator to use")); 712 713 714 /// Instantiate the default register allocator pass for this target for either 715 /// the optimized or unoptimized allocation path. This will be added to the pass 716 /// manager by addFastRegAlloc in the unoptimized case or addOptimizedRegAlloc 717 /// in the optimized case. 718 /// 719 /// A target that uses the standard regalloc pass order for fast or optimized 720 /// allocation may still override this for per-target regalloc 721 /// selection. But -regalloc=... always takes precedence. 722 FunctionPass *TargetPassConfig::createTargetRegisterAllocator(bool Optimized) { 723 if (Optimized) 724 return createGreedyRegisterAllocator(); 725 else 726 return createFastRegisterAllocator(); 727 } 728 729 /// Find and instantiate the register allocation pass requested by this target 730 /// at the current optimization level. Different register allocators are 731 /// defined as separate passes because they may require different analysis. 732 /// 733 /// This helper ensures that the regalloc= option is always available, 734 /// even for targets that override the default allocator. 735 /// 736 /// FIXME: When MachinePassRegistry register pass IDs instead of function ptrs, 737 /// this can be folded into addPass. 738 FunctionPass *TargetPassConfig::createRegAllocPass(bool Optimized) { 739 RegisterRegAlloc::FunctionPassCtor Ctor = RegisterRegAlloc::getDefault(); 740 741 // Initialize the global default. 742 if (!Ctor) { 743 Ctor = RegAlloc; 744 RegisterRegAlloc::setDefault(RegAlloc); 745 } 746 if (Ctor != useDefaultRegisterAllocator) 747 return Ctor(); 748 749 // With no -regalloc= override, ask the target for a regalloc pass. 750 return createTargetRegisterAllocator(Optimized); 751 } 752 753 /// Return true if the default global register allocator is in use and 754 /// has not be overriden on the command line with '-regalloc=...' 755 bool TargetPassConfig::usingDefaultRegAlloc() const { 756 return RegAlloc.getNumOccurrences() == 0; 757 } 758 759 /// Add the minimum set of target-independent passes that are required for 760 /// register allocation. No coalescing or scheduling. 761 void TargetPassConfig::addFastRegAlloc(FunctionPass *RegAllocPass) { 762 addPass(&PHIEliminationID, false); 763 addPass(&TwoAddressInstructionPassID, false); 764 765 if (RegAllocPass) 766 addPass(RegAllocPass); 767 } 768 769 /// Add standard target-independent passes that are tightly coupled with 770 /// optimized register allocation, including coalescing, machine instruction 771 /// scheduling, and register allocation itself. 772 void TargetPassConfig::addOptimizedRegAlloc(FunctionPass *RegAllocPass) { 773 addPass(&DetectDeadLanesID, false); 774 775 addPass(&ProcessImplicitDefsID, false); 776 777 // LiveVariables currently requires pure SSA form. 778 // 779 // FIXME: Once TwoAddressInstruction pass no longer uses kill flags, 780 // LiveVariables can be removed completely, and LiveIntervals can be directly 781 // computed. (We still either need to regenerate kill flags after regalloc, or 782 // preferably fix the scavenger to not depend on them). 783 addPass(&LiveVariablesID, false); 784 785 // Edge splitting is smarter with machine loop info. 786 addPass(&MachineLoopInfoID, false); 787 addPass(&PHIEliminationID, false); 788 789 // Eventually, we want to run LiveIntervals before PHI elimination. 790 if (EarlyLiveIntervals) 791 addPass(&LiveIntervalsID, false); 792 793 addPass(&TwoAddressInstructionPassID, false); 794 addPass(&RegisterCoalescerID); 795 796 // The machine scheduler may accidentally create disconnected components 797 // when moving subregister definitions around, avoid this by splitting them to 798 // separate vregs before. Splitting can also improve reg. allocation quality. 799 addPass(&RenameIndependentSubregsID); 800 801 // PreRA instruction scheduling. 802 addPass(&MachineSchedulerID); 803 804 if (RegAllocPass) { 805 // Add the selected register allocation pass. 806 addPass(RegAllocPass); 807 808 // Allow targets to change the register assignments before rewriting. 809 addPreRewrite(); 810 811 // Finally rewrite virtual registers. 812 addPass(&VirtRegRewriterID); 813 814 // Perform stack slot coloring and post-ra machine LICM. 815 // 816 // FIXME: Re-enable coloring with register when it's capable of adding 817 // kill markers. 818 addPass(&StackSlotColoringID); 819 820 // Run post-ra machine LICM to hoist reloads / remats. 821 // 822 // FIXME: can this move into MachineLateOptimization? 823 addPass(&PostRAMachineLICMID); 824 } 825 } 826 827 //===---------------------------------------------------------------------===// 828 /// Post RegAlloc Pass Configuration 829 //===---------------------------------------------------------------------===// 830 831 /// Add passes that optimize machine instructions after register allocation. 832 void TargetPassConfig::addMachineLateOptimization() { 833 // Branch folding must be run after regalloc and prolog/epilog insertion. 834 addPass(&BranchFolderPassID); 835 836 // Tail duplication. 837 // Note that duplicating tail just increases code size and degrades 838 // performance for targets that require Structured Control Flow. 839 // In addition it can also make CFG irreducible. Thus we disable it. 840 if (!TM->requiresStructuredCFG()) 841 addPass(&TailDuplicateID); 842 843 // Copy propagation. 844 addPass(&MachineCopyPropagationID); 845 } 846 847 /// Add standard GC passes. 848 bool TargetPassConfig::addGCPasses() { 849 addPass(&GCMachineCodeAnalysisID, false); 850 return true; 851 } 852 853 /// Add standard basic block placement passes. 854 void TargetPassConfig::addBlockPlacement() { 855 if (addPass(&MachineBlockPlacementID)) { 856 // Run a separate pass to collect block placement statistics. 857 if (EnableBlockPlacementStats) 858 addPass(&MachineBlockPlacementStatsID); 859 } 860 } 861