1 //===- TargetPassConfig.cpp - Target independent code generation passes ---===//
3 // The LLVM Compiler Infrastructure
5 // This file is distributed under the University of Illinois Open Source
6 // License. See LICENSE.TXT for details.
8 //===----------------------------------------------------------------------===//
10 // This file defines interfaces to access the target independent code
11 // generation passes provided by the LLVM backend.
13 //===---------------------------------------------------------------------===//
15 #include "llvm/CodeGen/TargetPassConfig.h"
16 #include "llvm/ADT/DenseMap.h"
17 #include "llvm/ADT/SmallVector.h"
18 #include "llvm/ADT/StringRef.h"
19 #include "llvm/Analysis/BasicAliasAnalysis.h"
20 #include "llvm/Analysis/CFLAndersAliasAnalysis.h"
21 #include "llvm/Analysis/CFLSteensAliasAnalysis.h"
22 #include "llvm/Analysis/CallGraphSCCPass.h"
23 #include "llvm/Analysis/ScopedNoAliasAA.h"
24 #include "llvm/Analysis/TargetTransformInfo.h"
25 #include "llvm/Analysis/TypeBasedAliasAnalysis.h"
26 #include "llvm/CodeGen/MachineFunctionPass.h"
27 #include "llvm/CodeGen/MachinePassRegistry.h"
28 #include "llvm/CodeGen/Passes.h"
29 #include "llvm/CodeGen/RegAllocRegistry.h"
30 #include "llvm/IR/IRPrintingPasses.h"
31 #include "llvm/IR/LegacyPassManager.h"
32 #include "llvm/IR/Verifier.h"
33 #include "llvm/MC/MCAsmInfo.h"
34 #include "llvm/MC/MCTargetOptions.h"
35 #include "llvm/Pass.h"
36 #include "llvm/Support/CodeGen.h"
37 #include "llvm/Support/CommandLine.h"
38 #include "llvm/Support/Compiler.h"
39 #include "llvm/Support/Debug.h"
40 #include "llvm/Support/ErrorHandling.h"
41 #include "llvm/Support/Threading.h"
42 #include "llvm/Target/TargetMachine.h"
43 #include "llvm/Transforms/Scalar.h"
44 #include "llvm/Transforms/Utils/SymbolRewriter.h"
50 static cl::opt<bool> DisablePostRASched("disable-post-ra", cl::Hidden,
51 cl::desc("Disable Post Regalloc Scheduler"));
52 static cl::opt<bool> DisableBranchFold("disable-branch-fold", cl::Hidden,
53 cl::desc("Disable branch folding"));
54 static cl::opt<bool> DisableTailDuplicate("disable-tail-duplicate", cl::Hidden,
55 cl::desc("Disable tail duplication"));
56 static cl::opt<bool> DisableEarlyTailDup("disable-early-taildup", cl::Hidden,
57 cl::desc("Disable pre-register allocation tail duplication"));
58 static cl::opt<bool> DisableBlockPlacement("disable-block-placement",
59 cl::Hidden, cl::desc("Disable probability-driven block placement"));
60 static cl::opt<bool> EnableBlockPlacementStats("enable-block-placement-stats",
61 cl::Hidden, cl::desc("Collect probability-driven block placement stats"));
62 static cl::opt<bool> DisableSSC("disable-ssc", cl::Hidden,
63 cl::desc("Disable Stack Slot Coloring"));
64 static cl::opt<bool> DisableMachineDCE("disable-machine-dce", cl::Hidden,
65 cl::desc("Disable Machine Dead Code Elimination"));
66 static cl::opt<bool> DisableEarlyIfConversion("disable-early-ifcvt", cl::Hidden,
67 cl::desc("Disable Early If-conversion"));
68 static cl::opt<bool> DisableMachineLICM("disable-machine-licm", cl::Hidden,
69 cl::desc("Disable Machine LICM"));
70 static cl::opt<bool> DisableMachineCSE("disable-machine-cse", cl::Hidden,
71 cl::desc("Disable Machine Common Subexpression Elimination"));
72 static cl::opt<cl::boolOrDefault> OptimizeRegAlloc(
73 "optimize-regalloc", cl::Hidden,
74 cl::desc("Enable optimized register allocation compilation path."));
75 static cl::opt<bool> DisablePostRAMachineLICM("disable-postra-machine-licm",
77 cl::desc("Disable Machine LICM"));
78 static cl::opt<bool> DisableMachineSink("disable-machine-sink", cl::Hidden,
79 cl::desc("Disable Machine Sinking"));
80 static cl::opt<bool> DisableLSR("disable-lsr", cl::Hidden,
81 cl::desc("Disable Loop Strength Reduction Pass"));
82 static cl::opt<bool> DisableConstantHoisting("disable-constant-hoisting",
83 cl::Hidden, cl::desc("Disable ConstantHoisting"));
84 static cl::opt<bool> DisableCGP("disable-cgp", cl::Hidden,
85 cl::desc("Disable Codegen Prepare"));
86 static cl::opt<bool> DisableCopyProp("disable-copyprop", cl::Hidden,
87 cl::desc("Disable Copy Propagation pass"));
88 static cl::opt<bool> DisablePartialLibcallInlining("disable-partial-libcall-inlining",
89 cl::Hidden, cl::desc("Disable Partial Libcall Inlining"));
90 static cl::opt<bool> EnableImplicitNullChecks(
91 "enable-implicit-null-checks",
92 cl::desc("Fold null checks into faulting memory operations"),
94 static cl::opt<bool> PrintLSR("print-lsr-output", cl::Hidden,
95 cl::desc("Print LLVM IR produced by the loop-reduce pass"));
96 static cl::opt<bool> PrintISelInput("print-isel-input", cl::Hidden,
97 cl::desc("Print LLVM IR input to isel pass"));
98 static cl::opt<bool> PrintGCInfo("print-gc", cl::Hidden,
99 cl::desc("Dump garbage collector data"));
100 static cl::opt<bool> VerifyMachineCode("verify-machineinstrs", cl::Hidden,
101 cl::desc("Verify generated machine code"),
104 static cl::opt<bool> EnableMachineOutliner("enable-machine-outliner",
106 cl::desc("Enable machine outliner"));
107 // Enable or disable FastISel. Both options are needed, because
108 // FastISel is enabled by default with -fast, and we wish to be
109 // able to enable or disable fast-isel independently from -O0.
110 static cl::opt<cl::boolOrDefault>
111 EnableFastISelOption("fast-isel", cl::Hidden,
112 cl::desc("Enable the \"fast\" instruction selector"));
114 static cl::opt<cl::boolOrDefault>
115 EnableGlobalISel("global-isel", cl::Hidden,
116 cl::desc("Enable the \"global\" instruction selector"));
118 static cl::opt<std::string>
119 PrintMachineInstrs("print-machineinstrs", cl::ValueOptional,
120 cl::desc("Print machine instrs"),
121 cl::value_desc("pass-name"), cl::init("option-unspecified"));
123 static cl::opt<int> EnableGlobalISelAbort(
124 "global-isel-abort", cl::Hidden,
125 cl::desc("Enable abort calls when \"global\" instruction selection "
126 "fails to lower/select an instruction: 0 disable the abort, "
127 "1 enable the abort, and "
128 "2 disable the abort but emit a diagnostic on failure"),
131 // Temporary option to allow experimenting with MachineScheduler as a post-RA
132 // scheduler. Targets can "properly" enable this with
133 // substitutePass(&PostRASchedulerID, &PostMachineSchedulerID).
134 // Targets can return true in targetSchedulesPostRAScheduling() and
135 // insert a PostRA scheduling pass wherever it wants.
136 cl::opt<bool> MISchedPostRA("misched-postra", cl::Hidden,
137 cl::desc("Run MachineScheduler post regalloc (independent of preRA sched)"));
139 // Experimental option to run live interval analysis early.
140 static cl::opt<bool> EarlyLiveIntervals("early-live-intervals", cl::Hidden,
141 cl::desc("Run live interval analysis earlier in the pipeline"));
143 // Experimental option to use CFL-AA in codegen
144 enum class CFLAAType { None, Steensgaard, Andersen, Both };
145 static cl::opt<CFLAAType> UseCFLAA(
146 "use-cfl-aa-in-codegen", cl::init(CFLAAType::None), cl::Hidden,
147 cl::desc("Enable the new, experimental CFL alias analysis in CodeGen"),
148 cl::values(clEnumValN(CFLAAType::None, "none", "Disable CFL-AA"),
149 clEnumValN(CFLAAType::Steensgaard, "steens",
150 "Enable unification-based CFL-AA"),
151 clEnumValN(CFLAAType::Andersen, "anders",
152 "Enable inclusion-based CFL-AA"),
153 clEnumValN(CFLAAType::Both, "both",
154 "Enable both variants of CFL-AA")));
156 /// Allow standard passes to be disabled by command line options. This supports
157 /// simple binary flags that either suppress the pass or do nothing.
158 /// i.e. -disable-mypass=false has no effect.
159 /// These should be converted to boolOrDefault in order to use applyOverride.
160 static IdentifyingPassPtr applyDisable(IdentifyingPassPtr PassID,
163 return IdentifyingPassPtr();
167 /// Allow standard passes to be disabled by the command line, regardless of who
168 /// is adding the pass.
170 /// StandardID is the pass identified in the standard pass pipeline and provided
171 /// to addPass(). It may be a target-specific ID in the case that the target
172 /// directly adds its own pass, but in that case we harmlessly fall through.
174 /// TargetID is the pass that the target has configured to override StandardID.
176 /// StandardID may be a pseudo ID. In that case TargetID is the name of the real
177 /// pass to run. This allows multiple options to control a single pass depending
178 /// on where in the pipeline that pass is added.
179 static IdentifyingPassPtr overridePass(AnalysisID StandardID,
180 IdentifyingPassPtr TargetID) {
181 if (StandardID == &PostRASchedulerID)
182 return applyDisable(TargetID, DisablePostRASched);
184 if (StandardID == &BranchFolderPassID)
185 return applyDisable(TargetID, DisableBranchFold);
187 if (StandardID == &TailDuplicateID)
188 return applyDisable(TargetID, DisableTailDuplicate);
190 if (StandardID == &TargetPassConfig::EarlyTailDuplicateID)
191 return applyDisable(TargetID, DisableEarlyTailDup);
193 if (StandardID == &MachineBlockPlacementID)
194 return applyDisable(TargetID, DisableBlockPlacement);
196 if (StandardID == &StackSlotColoringID)
197 return applyDisable(TargetID, DisableSSC);
199 if (StandardID == &DeadMachineInstructionElimID)
200 return applyDisable(TargetID, DisableMachineDCE);
202 if (StandardID == &EarlyIfConverterID)
203 return applyDisable(TargetID, DisableEarlyIfConversion);
205 if (StandardID == &MachineLICMID)
206 return applyDisable(TargetID, DisableMachineLICM);
208 if (StandardID == &MachineCSEID)
209 return applyDisable(TargetID, DisableMachineCSE);
211 if (StandardID == &TargetPassConfig::PostRAMachineLICMID)
212 return applyDisable(TargetID, DisablePostRAMachineLICM);
214 if (StandardID == &MachineSinkingID)
215 return applyDisable(TargetID, DisableMachineSink);
217 if (StandardID == &MachineCopyPropagationID)
218 return applyDisable(TargetID, DisableCopyProp);
223 //===---------------------------------------------------------------------===//
225 //===---------------------------------------------------------------------===//
227 INITIALIZE_PASS(TargetPassConfig, "targetpassconfig",
228 "Target Pass Configuration", false, false)
229 char TargetPassConfig::ID = 0;
232 char TargetPassConfig::EarlyTailDuplicateID = 0;
233 char TargetPassConfig::PostRAMachineLICMID = 0;
237 struct InsertedPass {
238 AnalysisID TargetPassID;
239 IdentifyingPassPtr InsertedPassID;
243 InsertedPass(AnalysisID TargetPassID, IdentifyingPassPtr InsertedPassID,
244 bool VerifyAfter, bool PrintAfter)
245 : TargetPassID(TargetPassID), InsertedPassID(InsertedPassID),
246 VerifyAfter(VerifyAfter), PrintAfter(PrintAfter) {}
248 Pass *getInsertedPass() const {
249 assert(InsertedPassID.isValid() && "Illegal Pass ID!");
250 if (InsertedPassID.isInstance())
251 return InsertedPassID.getInstance();
252 Pass *NP = Pass::createPass(InsertedPassID.getID());
253 assert(NP && "Pass ID not registered");
258 } // end anonymous namespace
262 class PassConfigImpl {
264 // List of passes explicitly substituted by this target. Normally this is
265 // empty, but it is a convenient way to suppress or replace specific passes
266 // that are part of a standard pass pipeline without overridding the entire
267 // pipeline. This mechanism allows target options to inherit a standard pass's
268 // user interface. For example, a target may disable a standard pass by
269 // default by substituting a pass ID of zero, and the user may still enable
270 // that standard pass with an explicit command line option.
271 DenseMap<AnalysisID,IdentifyingPassPtr> TargetPasses;
273 /// Store the pairs of <AnalysisID, AnalysisID> of which the second pass
274 /// is inserted after each instance of the first one.
275 SmallVector<InsertedPass, 4> InsertedPasses;
278 } // end namespace llvm
280 // Out of line virtual method.
281 TargetPassConfig::~TargetPassConfig() {
285 // Out of line constructor provides default values for pass options and
286 // registers all common codegen passes.
287 TargetPassConfig::TargetPassConfig(LLVMTargetMachine &TM, PassManagerBase &pm)
288 : ImmutablePass(ID), PM(&pm), TM(&TM) {
289 Impl = new PassConfigImpl();
291 // Register all target independent codegen passes to activate their PassIDs,
292 // including this pass itself.
293 initializeCodeGen(*PassRegistry::getPassRegistry());
295 // Also register alias analysis passes required by codegen passes.
296 initializeBasicAAWrapperPassPass(*PassRegistry::getPassRegistry());
297 initializeAAResultsWrapperPassPass(*PassRegistry::getPassRegistry());
299 // Substitute Pseudo Pass IDs for real ones.
300 substitutePass(&EarlyTailDuplicateID, &TailDuplicateID);
301 substitutePass(&PostRAMachineLICMID, &MachineLICMID);
303 if (StringRef(PrintMachineInstrs.getValue()).equals(""))
304 TM.Options.PrintMachineCode = true;
306 if (TM.Options.EnableIPRA)
307 setRequiresCodeGenSCCOrder();
310 CodeGenOpt::Level TargetPassConfig::getOptLevel() const {
311 return TM->getOptLevel();
314 /// Insert InsertedPassID pass after TargetPassID.
315 void TargetPassConfig::insertPass(AnalysisID TargetPassID,
316 IdentifyingPassPtr InsertedPassID,
317 bool VerifyAfter, bool PrintAfter) {
318 assert(((!InsertedPassID.isInstance() &&
319 TargetPassID != InsertedPassID.getID()) ||
320 (InsertedPassID.isInstance() &&
321 TargetPassID != InsertedPassID.getInstance()->getPassID())) &&
322 "Insert a pass after itself!");
323 Impl->InsertedPasses.emplace_back(TargetPassID, InsertedPassID, VerifyAfter,
327 /// createPassConfig - Create a pass configuration object to be used by
328 /// addPassToEmitX methods for generating a pipeline of CodeGen passes.
330 /// Targets may override this to extend TargetPassConfig.
331 TargetPassConfig *LLVMTargetMachine::createPassConfig(PassManagerBase &PM) {
332 return new TargetPassConfig(*this, PM);
335 TargetPassConfig::TargetPassConfig()
336 : ImmutablePass(ID) {
337 report_fatal_error("Trying to construct TargetPassConfig without a target "
338 "machine. Scheduling a CodeGen pass without a target "
342 // Helper to verify the analysis is really immutable.
343 void TargetPassConfig::setOpt(bool &Opt, bool Val) {
344 assert(!Initialized && "PassConfig is immutable");
348 void TargetPassConfig::substitutePass(AnalysisID StandardID,
349 IdentifyingPassPtr TargetID) {
350 Impl->TargetPasses[StandardID] = TargetID;
353 IdentifyingPassPtr TargetPassConfig::getPassSubstitution(AnalysisID ID) const {
354 DenseMap<AnalysisID, IdentifyingPassPtr>::const_iterator
355 I = Impl->TargetPasses.find(ID);
356 if (I == Impl->TargetPasses.end())
361 bool TargetPassConfig::isPassSubstitutedOrOverridden(AnalysisID ID) const {
362 IdentifyingPassPtr TargetID = getPassSubstitution(ID);
363 IdentifyingPassPtr FinalPtr = overridePass(ID, TargetID);
364 return !FinalPtr.isValid() || FinalPtr.isInstance() ||
365 FinalPtr.getID() != ID;
368 /// Add a pass to the PassManager if that pass is supposed to be run. If the
369 /// Started/Stopped flags indicate either that the compilation should start at
370 /// a later pass or that it should stop after an earlier pass, then do not add
371 /// the pass. Finally, compare the current pass against the StartAfter
372 /// and StopAfter options and change the Started/Stopped flags accordingly.
373 void TargetPassConfig::addPass(Pass *P, bool verifyAfter, bool printAfter) {
374 assert(!Initialized && "PassConfig is immutable");
376 // Cache the Pass ID here in case the pass manager finds this pass is
377 // redundant with ones already scheduled / available, and deletes it.
378 // Fundamentally, once we add the pass to the manager, we no longer own it
379 // and shouldn't reference it.
380 AnalysisID PassID = P->getPassID();
382 if (StartBefore == PassID)
384 if (StopBefore == PassID)
386 if (Started && !Stopped) {
388 // Construct banner message before PM->add() as that may delete the pass.
389 if (AddingMachinePasses && (printAfter || verifyAfter))
390 Banner = std::string("After ") + std::string(P->getPassName());
392 if (AddingMachinePasses) {
394 addPrintPass(Banner);
396 addVerifyPass(Banner);
399 // Add the passes after the pass P if there is any.
400 for (auto IP : Impl->InsertedPasses) {
401 if (IP.TargetPassID == PassID)
402 addPass(IP.getInsertedPass(), IP.VerifyAfter, IP.PrintAfter);
407 if (StopAfter == PassID)
409 if (StartAfter == PassID)
411 if (Stopped && !Started)
412 report_fatal_error("Cannot stop compilation after pass that is not run");
415 /// Add a CodeGen pass at this point in the pipeline after checking for target
416 /// and command line overrides.
418 /// addPass cannot return a pointer to the pass instance because is internal the
419 /// PassManager and the instance we create here may already be freed.
420 AnalysisID TargetPassConfig::addPass(AnalysisID PassID, bool verifyAfter,
422 IdentifyingPassPtr TargetID = getPassSubstitution(PassID);
423 IdentifyingPassPtr FinalPtr = overridePass(PassID, TargetID);
424 if (!FinalPtr.isValid())
428 if (FinalPtr.isInstance())
429 P = FinalPtr.getInstance();
431 P = Pass::createPass(FinalPtr.getID());
433 llvm_unreachable("Pass ID not registered");
435 AnalysisID FinalID = P->getPassID();
436 addPass(P, verifyAfter, printAfter); // Ends the lifetime of P.
441 void TargetPassConfig::printAndVerify(const std::string &Banner) {
442 addPrintPass(Banner);
443 addVerifyPass(Banner);
446 void TargetPassConfig::addPrintPass(const std::string &Banner) {
447 if (TM->shouldPrintMachineCode())
448 PM->add(createMachineFunctionPrinterPass(dbgs(), Banner));
451 void TargetPassConfig::addVerifyPass(const std::string &Banner) {
452 bool Verify = VerifyMachineCode;
453 #ifdef EXPENSIVE_CHECKS
454 if (VerifyMachineCode == cl::BOU_UNSET)
455 Verify = TM->isMachineVerifierClean();
458 PM->add(createMachineVerifierPass(Banner));
461 /// Add common target configurable passes that perform LLVM IR to IR transforms
462 /// following machine independent optimization.
463 void TargetPassConfig::addIRPasses() {
465 case CFLAAType::Steensgaard:
466 addPass(createCFLSteensAAWrapperPass());
468 case CFLAAType::Andersen:
469 addPass(createCFLAndersAAWrapperPass());
471 case CFLAAType::Both:
472 addPass(createCFLAndersAAWrapperPass());
473 addPass(createCFLSteensAAWrapperPass());
479 // Basic AliasAnalysis support.
480 // Add TypeBasedAliasAnalysis before BasicAliasAnalysis so that
481 // BasicAliasAnalysis wins if they disagree. This is intended to help
482 // support "obvious" type-punning idioms.
483 addPass(createTypeBasedAAWrapperPass());
484 addPass(createScopedNoAliasAAWrapperPass());
485 addPass(createBasicAAWrapperPass());
487 // Before running any passes, run the verifier to determine if the input
488 // coming from the front-end and/or optimizer is valid.
490 addPass(createVerifierPass());
492 // Run loop strength reduction before anything else.
493 if (getOptLevel() != CodeGenOpt::None && !DisableLSR) {
494 addPass(createLoopStrengthReducePass());
496 addPass(createPrintFunctionPass(dbgs(), "\n\n*** Code after LSR ***\n"));
499 // Run GC lowering passes for builtin collectors
500 // TODO: add a pass insertion point here
501 addPass(createGCLoweringPass());
502 addPass(createShadowStackGCLoweringPass());
504 // Make sure that no unreachable blocks are instruction selected.
505 addPass(createUnreachableBlockEliminationPass());
507 // Prepare expensive constants for SelectionDAG.
508 if (getOptLevel() != CodeGenOpt::None && !DisableConstantHoisting)
509 addPass(createConstantHoistingPass());
511 if (getOptLevel() != CodeGenOpt::None && !DisablePartialLibcallInlining)
512 addPass(createPartiallyInlineLibCallsPass());
514 // Insert calls to mcount-like functions.
515 addPass(createCountingFunctionInserterPass());
517 // Add scalarization of target's unsupported masked memory intrinsics pass.
518 // the unsupported intrinsic will be replaced with a chain of basic blocks,
519 // that stores/loads element one-by-one if the appropriate mask bit is set.
520 addPass(createScalarizeMaskedMemIntrinPass());
522 // Expand reduction intrinsics into shuffle sequences if the target wants to.
523 addPass(createExpandReductionsPass());
526 /// Turn exception handling constructs into something the code generators can
528 void TargetPassConfig::addPassesToHandleExceptions() {
529 const MCAsmInfo *MCAI = TM->getMCAsmInfo();
530 assert(MCAI && "No MCAsmInfo");
531 switch (MCAI->getExceptionHandlingType()) {
532 case ExceptionHandling::SjLj:
533 // SjLj piggy-backs on dwarf for this bit. The cleanups done apply to both
534 // Dwarf EH prepare needs to be run after SjLj prepare. Otherwise,
535 // catch info can get misplaced when a selector ends up more than one block
536 // removed from the parent invoke(s). This could happen when a landing
537 // pad is shared by multiple invokes and is also a target of a normal
538 // edge from elsewhere.
539 addPass(createSjLjEHPreparePass());
541 case ExceptionHandling::DwarfCFI:
542 case ExceptionHandling::ARM:
543 addPass(createDwarfEHPass());
545 case ExceptionHandling::WinEH:
546 // We support using both GCC-style and MSVC-style exceptions on Windows, so
547 // add both preparation passes. Each pass will only actually run if it
548 // recognizes the personality function.
549 addPass(createWinEHPass());
550 addPass(createDwarfEHPass());
552 case ExceptionHandling::None:
553 addPass(createLowerInvokePass());
555 // The lower invoke pass may create unreachable code. Remove it.
556 addPass(createUnreachableBlockEliminationPass());
561 /// Add pass to prepare the LLVM IR for code generation. This should be done
562 /// before exception handling preparation passes.
563 void TargetPassConfig::addCodeGenPrepare() {
564 if (getOptLevel() != CodeGenOpt::None && !DisableCGP)
565 addPass(createCodeGenPreparePass());
566 addPass(createRewriteSymbolsPass());
569 /// Add common passes that perform LLVM IR to IR transforms in preparation for
570 /// instruction selection.
571 void TargetPassConfig::addISelPrepare() {
574 // Force codegen to run according to the callgraph.
575 if (requiresCodeGenSCCOrder())
576 addPass(new DummyCGSCCPass);
578 // Add both the safe stack and the stack protection passes: each of them will
579 // only protect functions that have corresponding attributes.
580 addPass(createSafeStackPass());
581 addPass(createStackProtectorPass());
584 addPass(createPrintFunctionPass(
585 dbgs(), "\n\n*** Final LLVM Code input to ISel ***\n"));
587 // All passes which modify the LLVM IR are now complete; run the verifier
588 // to ensure that the IR is valid.
590 addPass(createVerifierPass());
593 bool TargetPassConfig::addCoreISelPasses() {
594 // Enable FastISel with -fast, but allow that to be overridden.
595 TM->setO0WantsFastISel(EnableFastISelOption != cl::BOU_FALSE);
596 if (EnableFastISelOption == cl::BOU_TRUE ||
597 (TM->getOptLevel() == CodeGenOpt::None && TM->getO0WantsFastISel()))
598 TM->setFastISel(true);
600 // Ask the target for an isel.
601 // Enable GlobalISel if the target wants to, but allow that to be overriden.
602 if (EnableGlobalISel == cl::BOU_TRUE ||
603 (EnableGlobalISel == cl::BOU_UNSET && isGlobalISelEnabled())) {
604 if (addIRTranslator())
607 addPreLegalizeMachineIR();
609 if (addLegalizeMachineIR())
612 // Before running the register bank selector, ask the target if it
613 // wants to run some passes.
614 addPreRegBankSelect();
616 if (addRegBankSelect())
619 addPreGlobalInstructionSelect();
621 if (addGlobalInstructionSelect())
624 // Pass to reset the MachineFunction if the ISel failed.
625 addPass(createResetMachineFunctionPass(
626 reportDiagnosticWhenGlobalISelFallback(), isGlobalISelAbortEnabled()));
628 // Provide a fallback path when we do not want to abort on
629 // not-yet-supported input.
630 if (!isGlobalISelAbortEnabled() && addInstSelector())
633 } else if (addInstSelector())
639 bool TargetPassConfig::addISelPasses() {
640 if (TM->Options.EmulatedTLS)
641 addPass(createLowerEmuTLSPass());
643 addPass(createPreISelIntrinsicLoweringPass());
644 addPass(createTargetTransformInfoWrapperPass(TM->getTargetIRAnalysis()));
647 addPassesToHandleExceptions();
650 return addCoreISelPasses();
653 /// -regalloc=... command line option.
654 static FunctionPass *useDefaultRegisterAllocator() { return nullptr; }
655 static cl::opt<RegisterRegAlloc::FunctionPassCtor, false,
656 RegisterPassParser<RegisterRegAlloc> >
658 cl::init(&useDefaultRegisterAllocator),
659 cl::desc("Register allocator to use"));
661 /// Add the complete set of target-independent postISel code generator passes.
663 /// This can be read as the standard order of major LLVM CodeGen stages. Stages
664 /// with nontrivial configuration or multiple passes are broken out below in
665 /// add%Stage routines.
667 /// Any TargetPassConfig::addXX routine may be overriden by the Target. The
668 /// addPre/Post methods with empty header implementations allow injecting
669 /// target-specific fixups just before or after major stages. Additionally,
670 /// targets have the flexibility to change pass order within a stage by
671 /// overriding default implementation of add%Stage routines below. Each
672 /// technique has maintainability tradeoffs because alternate pass orders are
673 /// not well supported. addPre/Post works better if the target pass is easily
674 /// tied to a common pass. But if it has subtle dependencies on multiple passes,
675 /// the target should override the stage instead.
677 /// TODO: We could use a single addPre/Post(ID) hook to allow pass injection
678 /// before/after any target-independent pass. But it's currently overkill.
679 void TargetPassConfig::addMachinePasses() {
680 AddingMachinePasses = true;
682 // Insert a machine instr printer pass after the specified pass.
683 if (!StringRef(PrintMachineInstrs.getValue()).equals("") &&
684 !StringRef(PrintMachineInstrs.getValue()).equals("option-unspecified")) {
685 const PassRegistry *PR = PassRegistry::getPassRegistry();
686 const PassInfo *TPI = PR->getPassInfo(PrintMachineInstrs.getValue());
687 const PassInfo *IPI = PR->getPassInfo(StringRef("machineinstr-printer"));
688 assert (TPI && IPI && "Pass ID not registered!");
689 const char *TID = (const char *)(TPI->getTypeInfo());
690 const char *IID = (const char *)(IPI->getTypeInfo());
691 insertPass(TID, IID);
694 // Print the instruction selected machine code...
695 printAndVerify("After Instruction Selection");
697 if (TM->Options.EnableIPRA)
698 addPass(createRegUsageInfoPropPass());
700 // Expand pseudo-instructions emitted by ISel.
701 addPass(&ExpandISelPseudosID);
703 // Add passes that optimize machine instructions in SSA form.
704 if (getOptLevel() != CodeGenOpt::None) {
705 addMachineSSAOptimization();
707 // If the target requests it, assign local variables to stack slots relative
708 // to one another and simplify frame index references where possible.
709 addPass(&LocalStackSlotAllocationID, false);
712 // Run pre-ra passes.
715 // Run register allocation and passes that are tightly coupled with it,
716 // including phi elimination and scheduling.
717 if (getOptimizeRegAlloc())
718 addOptimizedRegAlloc(createRegAllocPass(true));
720 if (RegAlloc != &useDefaultRegisterAllocator &&
721 RegAlloc != &createFastRegisterAllocator)
722 report_fatal_error("Must use fast (default) register allocator for unoptimized regalloc.");
723 addFastRegAlloc(createRegAllocPass(false));
726 // Run post-ra passes.
729 // Insert prolog/epilog code. Eliminate abstract frame index references...
730 if (getOptLevel() != CodeGenOpt::None)
731 addPass(&ShrinkWrapID);
733 // Prolog/Epilog inserter needs a TargetMachine to instantiate. But only
734 // do so if it hasn't been disabled, substituted, or overridden.
735 if (!isPassSubstitutedOrOverridden(&PrologEpilogCodeInserterID))
736 addPass(createPrologEpilogInserterPass());
738 /// Add passes that optimize machine instructions after register allocation.
739 if (getOptLevel() != CodeGenOpt::None)
740 addMachineLateOptimization();
742 // Expand pseudo instructions before second scheduling pass.
743 addPass(&ExpandPostRAPseudosID);
745 // Run pre-sched2 passes.
748 if (EnableImplicitNullChecks)
749 addPass(&ImplicitNullChecksID);
751 // Second pass scheduler.
752 // Let Target optionally insert this pass by itself at some other
754 if (getOptLevel() != CodeGenOpt::None &&
755 !TM->targetSchedulesPostRAScheduling()) {
757 addPass(&PostMachineSchedulerID);
759 addPass(&PostRASchedulerID);
765 addPass(createGCInfoPrinter(dbgs()), false, false);
768 // Basic block placement.
769 if (getOptLevel() != CodeGenOpt::None)
774 if (TM->Options.EnableIPRA)
775 // Collect register usage information and produce a register mask of
776 // clobbered registers, to be used to optimize call sites.
777 addPass(createRegUsageInfoCollector());
779 addPass(&FuncletLayoutID, false);
781 addPass(&StackMapLivenessID, false);
782 addPass(&LiveDebugValuesID, false);
784 // Insert before XRay Instrumentation.
785 addPass(&FEntryInserterID, false);
787 addPass(&XRayInstrumentationID, false);
788 addPass(&PatchableFunctionID, false);
790 if (EnableMachineOutliner)
791 PM->add(createMachineOutlinerPass());
793 AddingMachinePasses = false;
796 /// Add passes that optimize machine instructions in SSA form.
797 void TargetPassConfig::addMachineSSAOptimization() {
798 // Pre-ra tail duplication.
799 addPass(&EarlyTailDuplicateID);
801 // Optimize PHIs before DCE: removing dead PHI cycles may make more
802 // instructions dead.
803 addPass(&OptimizePHIsID, false);
805 // This pass merges large allocas. StackSlotColoring is a different pass
806 // which merges spill slots.
807 addPass(&StackColoringID, false);
809 // If the target requests it, assign local variables to stack slots relative
810 // to one another and simplify frame index references where possible.
811 addPass(&LocalStackSlotAllocationID, false);
813 // With optimization, dead code should already be eliminated. However
814 // there is one known exception: lowered code for arguments that are only
815 // used by tail calls, where the tail calls reuse the incoming stack
816 // arguments directly (see t11 in test/CodeGen/X86/sibcall.ll).
817 addPass(&DeadMachineInstructionElimID);
819 // Allow targets to insert passes that improve instruction level parallelism,
820 // like if-conversion. Such passes will typically need dominator trees and
821 // loop info, just like LICM and CSE below.
824 addPass(&MachineLICMID, false);
825 addPass(&MachineCSEID, false);
827 // Coalesce basic blocks with the same branch condition
828 addPass(&BranchCoalescingID);
830 addPass(&MachineSinkingID);
832 addPass(&PeepholeOptimizerID);
833 // Clean-up the dead code that may have been generated by peephole
835 addPass(&DeadMachineInstructionElimID);
838 //===---------------------------------------------------------------------===//
839 /// Register Allocation Pass Configuration
840 //===---------------------------------------------------------------------===//
842 bool TargetPassConfig::getOptimizeRegAlloc() const {
843 switch (OptimizeRegAlloc) {
844 case cl::BOU_UNSET: return getOptLevel() != CodeGenOpt::None;
845 case cl::BOU_TRUE: return true;
846 case cl::BOU_FALSE: return false;
848 llvm_unreachable("Invalid optimize-regalloc state");
851 /// RegisterRegAlloc's global Registry tracks allocator registration.
852 MachinePassRegistry RegisterRegAlloc::Registry;
854 /// A dummy default pass factory indicates whether the register allocator is
855 /// overridden on the command line.
856 static llvm::once_flag InitializeDefaultRegisterAllocatorFlag;
858 static RegisterRegAlloc
859 defaultRegAlloc("default",
860 "pick register allocator based on -O option",
861 useDefaultRegisterAllocator);
863 static void initializeDefaultRegisterAllocatorOnce() {
864 RegisterRegAlloc::FunctionPassCtor Ctor = RegisterRegAlloc::getDefault();
868 RegisterRegAlloc::setDefault(RegAlloc);
872 /// Instantiate the default register allocator pass for this target for either
873 /// the optimized or unoptimized allocation path. This will be added to the pass
874 /// manager by addFastRegAlloc in the unoptimized case or addOptimizedRegAlloc
875 /// in the optimized case.
877 /// A target that uses the standard regalloc pass order for fast or optimized
878 /// allocation may still override this for per-target regalloc
879 /// selection. But -regalloc=... always takes precedence.
880 FunctionPass *TargetPassConfig::createTargetRegisterAllocator(bool Optimized) {
882 return createGreedyRegisterAllocator();
884 return createFastRegisterAllocator();
887 /// Find and instantiate the register allocation pass requested by this target
888 /// at the current optimization level. Different register allocators are
889 /// defined as separate passes because they may require different analysis.
891 /// This helper ensures that the regalloc= option is always available,
892 /// even for targets that override the default allocator.
894 /// FIXME: When MachinePassRegistry register pass IDs instead of function ptrs,
895 /// this can be folded into addPass.
896 FunctionPass *TargetPassConfig::createRegAllocPass(bool Optimized) {
897 // Initialize the global default.
898 llvm::call_once(InitializeDefaultRegisterAllocatorFlag,
899 initializeDefaultRegisterAllocatorOnce);
901 RegisterRegAlloc::FunctionPassCtor Ctor = RegisterRegAlloc::getDefault();
902 if (Ctor != useDefaultRegisterAllocator)
905 // With no -regalloc= override, ask the target for a regalloc pass.
906 return createTargetRegisterAllocator(Optimized);
909 /// Return true if the default global register allocator is in use and
910 /// has not be overriden on the command line with '-regalloc=...'
911 bool TargetPassConfig::usingDefaultRegAlloc() const {
912 return RegAlloc.getNumOccurrences() == 0;
915 /// Add the minimum set of target-independent passes that are required for
916 /// register allocation. No coalescing or scheduling.
917 void TargetPassConfig::addFastRegAlloc(FunctionPass *RegAllocPass) {
918 addPass(&PHIEliminationID, false);
919 addPass(&TwoAddressInstructionPassID, false);
922 addPass(RegAllocPass);
925 /// Add standard target-independent passes that are tightly coupled with
926 /// optimized register allocation, including coalescing, machine instruction
927 /// scheduling, and register allocation itself.
928 void TargetPassConfig::addOptimizedRegAlloc(FunctionPass *RegAllocPass) {
929 addPass(&DetectDeadLanesID, false);
931 addPass(&ProcessImplicitDefsID, false);
933 // LiveVariables currently requires pure SSA form.
935 // FIXME: Once TwoAddressInstruction pass no longer uses kill flags,
936 // LiveVariables can be removed completely, and LiveIntervals can be directly
937 // computed. (We still either need to regenerate kill flags after regalloc, or
938 // preferably fix the scavenger to not depend on them).
939 addPass(&LiveVariablesID, false);
941 // Edge splitting is smarter with machine loop info.
942 addPass(&MachineLoopInfoID, false);
943 addPass(&PHIEliminationID, false);
945 // Eventually, we want to run LiveIntervals before PHI elimination.
946 if (EarlyLiveIntervals)
947 addPass(&LiveIntervalsID, false);
949 addPass(&TwoAddressInstructionPassID, false);
950 addPass(&RegisterCoalescerID);
952 // The machine scheduler may accidentally create disconnected components
953 // when moving subregister definitions around, avoid this by splitting them to
954 // separate vregs before. Splitting can also improve reg. allocation quality.
955 addPass(&RenameIndependentSubregsID);
957 // PreRA instruction scheduling.
958 addPass(&MachineSchedulerID);
961 // Add the selected register allocation pass.
962 addPass(RegAllocPass);
964 // Allow targets to change the register assignments before rewriting.
967 // Finally rewrite virtual registers.
968 addPass(&VirtRegRewriterID);
970 // Perform stack slot coloring and post-ra machine LICM.
972 // FIXME: Re-enable coloring with register when it's capable of adding
974 addPass(&StackSlotColoringID);
976 // Run post-ra machine LICM to hoist reloads / remats.
978 // FIXME: can this move into MachineLateOptimization?
979 addPass(&PostRAMachineLICMID);
983 //===---------------------------------------------------------------------===//
984 /// Post RegAlloc Pass Configuration
985 //===---------------------------------------------------------------------===//
987 /// Add passes that optimize machine instructions after register allocation.
988 void TargetPassConfig::addMachineLateOptimization() {
989 // Branch folding must be run after regalloc and prolog/epilog insertion.
990 addPass(&BranchFolderPassID);
993 // Note that duplicating tail just increases code size and degrades
994 // performance for targets that require Structured Control Flow.
995 // In addition it can also make CFG irreducible. Thus we disable it.
996 if (!TM->requiresStructuredCFG())
997 addPass(&TailDuplicateID);
1000 addPass(&MachineCopyPropagationID);
1003 /// Add standard GC passes.
1004 bool TargetPassConfig::addGCPasses() {
1005 addPass(&GCMachineCodeAnalysisID, false);
1009 /// Add standard basic block placement passes.
1010 void TargetPassConfig::addBlockPlacement() {
1011 if (addPass(&MachineBlockPlacementID)) {
1012 // Run a separate pass to collect block placement statistics.
1013 if (EnableBlockPlacementStats)
1014 addPass(&MachineBlockPlacementStatsID);
1018 //===---------------------------------------------------------------------===//
1019 /// GlobalISel Configuration
1020 //===---------------------------------------------------------------------===//
1022 bool TargetPassConfig::isGlobalISelEnabled() const {
1026 bool TargetPassConfig::isGlobalISelAbortEnabled() const {
1027 return EnableGlobalISelAbort == 1;
1030 bool TargetPassConfig::reportDiagnosticWhenGlobalISelFallback() const {
1031 return EnableGlobalISelAbort == 2;