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