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 cl::opt
<bool> EnableIPRA("enable-ipra", cl::init(false), cl::Hidden
,
51 cl::desc("Enable interprocedural register allocation "
52 "to reduce load/store at procedure calls."));
53 static cl::opt
<bool> DisablePostRASched("disable-post-ra", cl::Hidden
,
54 cl::desc("Disable Post Regalloc Scheduler"));
55 static cl::opt
<bool> DisableBranchFold("disable-branch-fold", cl::Hidden
,
56 cl::desc("Disable branch folding"));
57 static cl::opt
<bool> DisableTailDuplicate("disable-tail-duplicate", cl::Hidden
,
58 cl::desc("Disable tail duplication"));
59 static cl::opt
<bool> DisableEarlyTailDup("disable-early-taildup", cl::Hidden
,
60 cl::desc("Disable pre-register allocation tail duplication"));
61 static cl::opt
<bool> DisableBlockPlacement("disable-block-placement",
62 cl::Hidden
, cl::desc("Disable probability-driven block placement"));
63 static cl::opt
<bool> EnableBlockPlacementStats("enable-block-placement-stats",
64 cl::Hidden
, cl::desc("Collect probability-driven block placement stats"));
65 static cl::opt
<bool> DisableSSC("disable-ssc", cl::Hidden
,
66 cl::desc("Disable Stack Slot Coloring"));
67 static cl::opt
<bool> DisableMachineDCE("disable-machine-dce", cl::Hidden
,
68 cl::desc("Disable Machine Dead Code Elimination"));
69 static cl::opt
<bool> DisableEarlyIfConversion("disable-early-ifcvt", cl::Hidden
,
70 cl::desc("Disable Early If-conversion"));
71 static cl::opt
<bool> DisableMachineLICM("disable-machine-licm", cl::Hidden
,
72 cl::desc("Disable Machine LICM"));
73 static cl::opt
<bool> DisableMachineCSE("disable-machine-cse", cl::Hidden
,
74 cl::desc("Disable Machine Common Subexpression Elimination"));
75 static cl::opt
<cl::boolOrDefault
> OptimizeRegAlloc(
76 "optimize-regalloc", cl::Hidden
,
77 cl::desc("Enable optimized register allocation compilation path."));
78 static cl::opt
<bool> DisablePostRAMachineLICM("disable-postra-machine-licm",
80 cl::desc("Disable Machine LICM"));
81 static cl::opt
<bool> DisableMachineSink("disable-machine-sink", cl::Hidden
,
82 cl::desc("Disable Machine Sinking"));
83 static cl::opt
<bool> DisableLSR("disable-lsr", cl::Hidden
,
84 cl::desc("Disable Loop Strength Reduction Pass"));
85 static cl::opt
<bool> DisableConstantHoisting("disable-constant-hoisting",
86 cl::Hidden
, cl::desc("Disable ConstantHoisting"));
87 static cl::opt
<bool> DisableCGP("disable-cgp", cl::Hidden
,
88 cl::desc("Disable Codegen Prepare"));
89 static cl::opt
<bool> DisableCopyProp("disable-copyprop", cl::Hidden
,
90 cl::desc("Disable Copy Propagation pass"));
91 static cl::opt
<bool> DisablePartialLibcallInlining("disable-partial-libcall-inlining",
92 cl::Hidden
, cl::desc("Disable Partial Libcall Inlining"));
93 static cl::opt
<bool> EnableImplicitNullChecks(
94 "enable-implicit-null-checks",
95 cl::desc("Fold null checks into faulting memory operations"),
96 cl::init(false), cl::Hidden
);
98 EnableMergeICmps("enable-mergeicmps",
99 cl::desc("Merge ICmp chains into a single memcmp"),
100 cl::init(false), cl::Hidden
);
101 static cl::opt
<bool> PrintLSR("print-lsr-output", cl::Hidden
,
102 cl::desc("Print LLVM IR produced by the loop-reduce pass"));
103 static cl::opt
<bool> PrintISelInput("print-isel-input", cl::Hidden
,
104 cl::desc("Print LLVM IR input to isel pass"));
105 static cl::opt
<bool> PrintGCInfo("print-gc", cl::Hidden
,
106 cl::desc("Dump garbage collector data"));
107 static cl::opt
<bool> VerifyMachineCode("verify-machineinstrs", cl::Hidden
,
108 cl::desc("Verify generated machine code"),
111 static cl::opt
<bool> EnableMachineOutliner("enable-machine-outliner",
113 cl::desc("Enable machine outliner"));
114 static cl::opt
<bool> EnableLinkOnceODROutlining(
115 "enable-linkonceodr-outlining",
117 cl::desc("Enable the machine outliner on linkonceodr functions"),
119 // Enable or disable FastISel. Both options are needed, because
120 // FastISel is enabled by default with -fast, and we wish to be
121 // able to enable or disable fast-isel independently from -O0.
122 static cl::opt
<cl::boolOrDefault
>
123 EnableFastISelOption("fast-isel", cl::Hidden
,
124 cl::desc("Enable the \"fast\" instruction selector"));
126 static cl::opt
<cl::boolOrDefault
>
127 EnableGlobalISel("global-isel", cl::Hidden
,
128 cl::desc("Enable the \"global\" instruction selector"));
130 static cl::opt
<std::string
> PrintMachineInstrs(
131 "print-machineinstrs", cl::ValueOptional
, cl::desc("Print machine instrs"),
132 cl::value_desc("pass-name"), cl::init("option-unspecified"), cl::Hidden
);
134 static cl::opt
<int> EnableGlobalISelAbort(
135 "global-isel-abort", cl::Hidden
,
136 cl::desc("Enable abort calls when \"global\" instruction selection "
137 "fails to lower/select an instruction: 0 disable the abort, "
138 "1 enable the abort, and "
139 "2 disable the abort but emit a diagnostic on failure"),
142 // Temporary option to allow experimenting with MachineScheduler as a post-RA
143 // scheduler. Targets can "properly" enable this with
144 // substitutePass(&PostRASchedulerID, &PostMachineSchedulerID).
145 // Targets can return true in targetSchedulesPostRAScheduling() and
146 // insert a PostRA scheduling pass wherever it wants.
147 cl::opt
<bool> MISchedPostRA("misched-postra", cl::Hidden
,
148 cl::desc("Run MachineScheduler post regalloc (independent of preRA sched)"));
150 // Experimental option to run live interval analysis early.
151 static cl::opt
<bool> EarlyLiveIntervals("early-live-intervals", cl::Hidden
,
152 cl::desc("Run live interval analysis earlier in the pipeline"));
154 // Experimental option to use CFL-AA in codegen
155 enum class CFLAAType
{ None
, Steensgaard
, Andersen
, Both
};
156 static cl::opt
<CFLAAType
> UseCFLAA(
157 "use-cfl-aa-in-codegen", cl::init(CFLAAType::None
), cl::Hidden
,
158 cl::desc("Enable the new, experimental CFL alias analysis in CodeGen"),
159 cl::values(clEnumValN(CFLAAType::None
, "none", "Disable CFL-AA"),
160 clEnumValN(CFLAAType::Steensgaard
, "steens",
161 "Enable unification-based CFL-AA"),
162 clEnumValN(CFLAAType::Andersen
, "anders",
163 "Enable inclusion-based CFL-AA"),
164 clEnumValN(CFLAAType::Both
, "both",
165 "Enable both variants of CFL-AA")));
167 /// Option names for limiting the codegen pipeline.
168 /// Those are used in error reporting and we didn't want
169 /// to duplicate their names all over the place.
170 const char *StartAfterOptName
= "start-after";
171 const char *StartBeforeOptName
= "start-before";
172 const char *StopAfterOptName
= "stop-after";
173 const char *StopBeforeOptName
= "stop-before";
175 static cl::opt
<std::string
>
176 StartAfterOpt(StringRef(StartAfterOptName
),
177 cl::desc("Resume compilation after a specific pass"),
178 cl::value_desc("pass-name"), cl::init(""), cl::Hidden
);
180 static cl::opt
<std::string
>
181 StartBeforeOpt(StringRef(StartBeforeOptName
),
182 cl::desc("Resume compilation before a specific pass"),
183 cl::value_desc("pass-name"), cl::init(""), cl::Hidden
);
185 static cl::opt
<std::string
>
186 StopAfterOpt(StringRef(StopAfterOptName
),
187 cl::desc("Stop compilation after a specific pass"),
188 cl::value_desc("pass-name"), cl::init(""), cl::Hidden
);
190 static cl::opt
<std::string
>
191 StopBeforeOpt(StringRef(StopBeforeOptName
),
192 cl::desc("Stop compilation before a specific pass"),
193 cl::value_desc("pass-name"), cl::init(""), cl::Hidden
);
195 /// Allow standard passes to be disabled by command line options. This supports
196 /// simple binary flags that either suppress the pass or do nothing.
197 /// i.e. -disable-mypass=false has no effect.
198 /// These should be converted to boolOrDefault in order to use applyOverride.
199 static IdentifyingPassPtr
applyDisable(IdentifyingPassPtr PassID
,
202 return IdentifyingPassPtr();
206 /// Allow standard passes to be disabled by the command line, regardless of who
207 /// is adding the pass.
209 /// StandardID is the pass identified in the standard pass pipeline and provided
210 /// to addPass(). It may be a target-specific ID in the case that the target
211 /// directly adds its own pass, but in that case we harmlessly fall through.
213 /// TargetID is the pass that the target has configured to override StandardID.
215 /// StandardID may be a pseudo ID. In that case TargetID is the name of the real
216 /// pass to run. This allows multiple options to control a single pass depending
217 /// on where in the pipeline that pass is added.
218 static IdentifyingPassPtr
overridePass(AnalysisID StandardID
,
219 IdentifyingPassPtr TargetID
) {
220 if (StandardID
== &PostRASchedulerID
)
221 return applyDisable(TargetID
, DisablePostRASched
);
223 if (StandardID
== &BranchFolderPassID
)
224 return applyDisable(TargetID
, DisableBranchFold
);
226 if (StandardID
== &TailDuplicateID
)
227 return applyDisable(TargetID
, DisableTailDuplicate
);
229 if (StandardID
== &TargetPassConfig::EarlyTailDuplicateID
)
230 return applyDisable(TargetID
, DisableEarlyTailDup
);
232 if (StandardID
== &MachineBlockPlacementID
)
233 return applyDisable(TargetID
, DisableBlockPlacement
);
235 if (StandardID
== &StackSlotColoringID
)
236 return applyDisable(TargetID
, DisableSSC
);
238 if (StandardID
== &DeadMachineInstructionElimID
)
239 return applyDisable(TargetID
, DisableMachineDCE
);
241 if (StandardID
== &EarlyIfConverterID
)
242 return applyDisable(TargetID
, DisableEarlyIfConversion
);
244 if (StandardID
== &MachineLICMID
)
245 return applyDisable(TargetID
, DisableMachineLICM
);
247 if (StandardID
== &MachineCSEID
)
248 return applyDisable(TargetID
, DisableMachineCSE
);
250 if (StandardID
== &TargetPassConfig::PostRAMachineLICMID
)
251 return applyDisable(TargetID
, DisablePostRAMachineLICM
);
253 if (StandardID
== &MachineSinkingID
)
254 return applyDisable(TargetID
, DisableMachineSink
);
256 if (StandardID
== &MachineCopyPropagationID
)
257 return applyDisable(TargetID
, DisableCopyProp
);
262 //===---------------------------------------------------------------------===//
264 //===---------------------------------------------------------------------===//
266 INITIALIZE_PASS(TargetPassConfig
, "targetpassconfig",
267 "Target Pass Configuration", false, false)
268 char TargetPassConfig::ID
= 0;
271 char TargetPassConfig::EarlyTailDuplicateID
= 0;
272 char TargetPassConfig::PostRAMachineLICMID
= 0;
276 struct InsertedPass
{
277 AnalysisID TargetPassID
;
278 IdentifyingPassPtr InsertedPassID
;
282 InsertedPass(AnalysisID TargetPassID
, IdentifyingPassPtr InsertedPassID
,
283 bool VerifyAfter
, bool PrintAfter
)
284 : TargetPassID(TargetPassID
), InsertedPassID(InsertedPassID
),
285 VerifyAfter(VerifyAfter
), PrintAfter(PrintAfter
) {}
287 Pass
*getInsertedPass() const {
288 assert(InsertedPassID
.isValid() && "Illegal Pass ID!");
289 if (InsertedPassID
.isInstance())
290 return InsertedPassID
.getInstance();
291 Pass
*NP
= Pass::createPass(InsertedPassID
.getID());
292 assert(NP
&& "Pass ID not registered");
297 } // end anonymous namespace
301 class PassConfigImpl
{
303 // List of passes explicitly substituted by this target. Normally this is
304 // empty, but it is a convenient way to suppress or replace specific passes
305 // that are part of a standard pass pipeline without overridding the entire
306 // pipeline. This mechanism allows target options to inherit a standard pass's
307 // user interface. For example, a target may disable a standard pass by
308 // default by substituting a pass ID of zero, and the user may still enable
309 // that standard pass with an explicit command line option.
310 DenseMap
<AnalysisID
,IdentifyingPassPtr
> TargetPasses
;
312 /// Store the pairs of <AnalysisID, AnalysisID> of which the second pass
313 /// is inserted after each instance of the first one.
314 SmallVector
<InsertedPass
, 4> InsertedPasses
;
317 } // end namespace llvm
319 // Out of line virtual method.
320 TargetPassConfig::~TargetPassConfig() {
324 static const PassInfo
*getPassInfo(StringRef PassName
) {
325 if (PassName
.empty())
328 const PassRegistry
&PR
= *PassRegistry::getPassRegistry();
329 const PassInfo
*PI
= PR
.getPassInfo(PassName
);
331 report_fatal_error(Twine('\"') + Twine(PassName
) +
332 Twine("\" pass is not registered."));
336 static AnalysisID
getPassIDFromName(StringRef PassName
) {
337 const PassInfo
*PI
= getPassInfo(PassName
);
338 return PI
? PI
->getTypeInfo() : nullptr;
341 void TargetPassConfig::setStartStopPasses() {
342 StartBefore
= getPassIDFromName(StartBeforeOpt
);
343 StartAfter
= getPassIDFromName(StartAfterOpt
);
344 StopBefore
= getPassIDFromName(StopBeforeOpt
);
345 StopAfter
= getPassIDFromName(StopAfterOpt
);
346 if (StartBefore
&& StartAfter
)
347 report_fatal_error(Twine(StartBeforeOptName
) + Twine(" and ") +
348 Twine(StartAfterOptName
) + Twine(" specified!"));
349 if (StopBefore
&& StopAfter
)
350 report_fatal_error(Twine(StopBeforeOptName
) + Twine(" and ") +
351 Twine(StopAfterOptName
) + Twine(" specified!"));
352 Started
= (StartAfter
== nullptr) && (StartBefore
== nullptr);
355 // Out of line constructor provides default values for pass options and
356 // registers all common codegen passes.
357 TargetPassConfig::TargetPassConfig(LLVMTargetMachine
&TM
, PassManagerBase
&pm
)
358 : ImmutablePass(ID
), PM(&pm
), TM(&TM
) {
359 Impl
= new PassConfigImpl();
361 // Register all target independent codegen passes to activate their PassIDs,
362 // including this pass itself.
363 initializeCodeGen(*PassRegistry::getPassRegistry());
365 // Also register alias analysis passes required by codegen passes.
366 initializeBasicAAWrapperPassPass(*PassRegistry::getPassRegistry());
367 initializeAAResultsWrapperPassPass(*PassRegistry::getPassRegistry());
369 // Substitute Pseudo Pass IDs for real ones.
370 substitutePass(&EarlyTailDuplicateID
, &TailDuplicateID
);
371 substitutePass(&PostRAMachineLICMID
, &MachineLICMID
);
373 if (StringRef(PrintMachineInstrs
.getValue()).equals(""))
374 TM
.Options
.PrintMachineCode
= true;
376 if (EnableIPRA
.getNumOccurrences())
377 TM
.Options
.EnableIPRA
= EnableIPRA
;
379 // If not explicitly specified, use target default.
380 TM
.Options
.EnableIPRA
= TM
.useIPRA();
383 if (TM
.Options
.EnableIPRA
)
384 setRequiresCodeGenSCCOrder();
386 setStartStopPasses();
389 CodeGenOpt::Level
TargetPassConfig::getOptLevel() const {
390 return TM
->getOptLevel();
393 /// Insert InsertedPassID pass after TargetPassID.
394 void TargetPassConfig::insertPass(AnalysisID TargetPassID
,
395 IdentifyingPassPtr InsertedPassID
,
396 bool VerifyAfter
, bool PrintAfter
) {
397 assert(((!InsertedPassID
.isInstance() &&
398 TargetPassID
!= InsertedPassID
.getID()) ||
399 (InsertedPassID
.isInstance() &&
400 TargetPassID
!= InsertedPassID
.getInstance()->getPassID())) &&
401 "Insert a pass after itself!");
402 Impl
->InsertedPasses
.emplace_back(TargetPassID
, InsertedPassID
, VerifyAfter
,
406 /// createPassConfig - Create a pass configuration object to be used by
407 /// addPassToEmitX methods for generating a pipeline of CodeGen passes.
409 /// Targets may override this to extend TargetPassConfig.
410 TargetPassConfig
*LLVMTargetMachine::createPassConfig(PassManagerBase
&PM
) {
411 return new TargetPassConfig(*this, PM
);
414 TargetPassConfig::TargetPassConfig()
415 : ImmutablePass(ID
) {
416 report_fatal_error("Trying to construct TargetPassConfig without a target "
417 "machine. Scheduling a CodeGen pass without a target "
421 bool TargetPassConfig::hasLimitedCodeGenPipeline() const {
422 return StartBefore
|| StartAfter
|| StopBefore
|| StopAfter
;
426 TargetPassConfig::getLimitedCodeGenPipelineReason(const char *Separator
) const {
427 if (!hasLimitedCodeGenPipeline())
428 return std::string();
430 static cl::opt
<std::string
> *PassNames
[] = {&StartAfterOpt
, &StartBeforeOpt
,
431 &StopAfterOpt
, &StopBeforeOpt
};
432 static const char *OptNames
[] = {StartAfterOptName
, StartBeforeOptName
,
433 StopAfterOptName
, StopBeforeOptName
};
435 for (int Idx
= 0; Idx
< 4; ++Idx
)
436 if (!PassNames
[Idx
]->empty()) {
440 Res
+= OptNames
[Idx
];
445 // Helper to verify the analysis is really immutable.
446 void TargetPassConfig::setOpt(bool &Opt
, bool Val
) {
447 assert(!Initialized
&& "PassConfig is immutable");
451 void TargetPassConfig::substitutePass(AnalysisID StandardID
,
452 IdentifyingPassPtr TargetID
) {
453 Impl
->TargetPasses
[StandardID
] = TargetID
;
456 IdentifyingPassPtr
TargetPassConfig::getPassSubstitution(AnalysisID ID
) const {
457 DenseMap
<AnalysisID
, IdentifyingPassPtr
>::const_iterator
458 I
= Impl
->TargetPasses
.find(ID
);
459 if (I
== Impl
->TargetPasses
.end())
464 bool TargetPassConfig::isPassSubstitutedOrOverridden(AnalysisID ID
) const {
465 IdentifyingPassPtr TargetID
= getPassSubstitution(ID
);
466 IdentifyingPassPtr FinalPtr
= overridePass(ID
, TargetID
);
467 return !FinalPtr
.isValid() || FinalPtr
.isInstance() ||
468 FinalPtr
.getID() != ID
;
471 /// Add a pass to the PassManager if that pass is supposed to be run. If the
472 /// Started/Stopped flags indicate either that the compilation should start at
473 /// a later pass or that it should stop after an earlier pass, then do not add
474 /// the pass. Finally, compare the current pass against the StartAfter
475 /// and StopAfter options and change the Started/Stopped flags accordingly.
476 void TargetPassConfig::addPass(Pass
*P
, bool verifyAfter
, bool printAfter
) {
477 assert(!Initialized
&& "PassConfig is immutable");
479 // Cache the Pass ID here in case the pass manager finds this pass is
480 // redundant with ones already scheduled / available, and deletes it.
481 // Fundamentally, once we add the pass to the manager, we no longer own it
482 // and shouldn't reference it.
483 AnalysisID PassID
= P
->getPassID();
485 if (StartBefore
== PassID
)
487 if (StopBefore
== PassID
)
489 if (Started
&& !Stopped
) {
491 // Construct banner message before PM->add() as that may delete the pass.
492 if (AddingMachinePasses
&& (printAfter
|| verifyAfter
))
493 Banner
= std::string("After ") + std::string(P
->getPassName());
495 if (AddingMachinePasses
) {
497 addPrintPass(Banner
);
499 addVerifyPass(Banner
);
502 // Add the passes after the pass P if there is any.
503 for (auto IP
: Impl
->InsertedPasses
) {
504 if (IP
.TargetPassID
== PassID
)
505 addPass(IP
.getInsertedPass(), IP
.VerifyAfter
, IP
.PrintAfter
);
510 if (StopAfter
== PassID
)
512 if (StartAfter
== PassID
)
514 if (Stopped
&& !Started
)
515 report_fatal_error("Cannot stop compilation after pass that is not run");
518 /// Add a CodeGen pass at this point in the pipeline after checking for target
519 /// and command line overrides.
521 /// addPass cannot return a pointer to the pass instance because is internal the
522 /// PassManager and the instance we create here may already be freed.
523 AnalysisID
TargetPassConfig::addPass(AnalysisID PassID
, bool verifyAfter
,
525 IdentifyingPassPtr TargetID
= getPassSubstitution(PassID
);
526 IdentifyingPassPtr FinalPtr
= overridePass(PassID
, TargetID
);
527 if (!FinalPtr
.isValid())
531 if (FinalPtr
.isInstance())
532 P
= FinalPtr
.getInstance();
534 P
= Pass::createPass(FinalPtr
.getID());
536 llvm_unreachable("Pass ID not registered");
538 AnalysisID FinalID
= P
->getPassID();
539 addPass(P
, verifyAfter
, printAfter
); // Ends the lifetime of P.
544 void TargetPassConfig::printAndVerify(const std::string
&Banner
) {
545 addPrintPass(Banner
);
546 addVerifyPass(Banner
);
549 void TargetPassConfig::addPrintPass(const std::string
&Banner
) {
550 if (TM
->shouldPrintMachineCode())
551 PM
->add(createMachineFunctionPrinterPass(dbgs(), Banner
));
554 void TargetPassConfig::addVerifyPass(const std::string
&Banner
) {
555 bool Verify
= VerifyMachineCode
;
556 #ifdef EXPENSIVE_CHECKS
557 if (VerifyMachineCode
== cl::BOU_UNSET
)
558 Verify
= TM
->isMachineVerifierClean();
561 PM
->add(createMachineVerifierPass(Banner
));
564 /// Add common target configurable passes that perform LLVM IR to IR transforms
565 /// following machine independent optimization.
566 void TargetPassConfig::addIRPasses() {
568 case CFLAAType::Steensgaard
:
569 addPass(createCFLSteensAAWrapperPass());
571 case CFLAAType::Andersen
:
572 addPass(createCFLAndersAAWrapperPass());
574 case CFLAAType::Both
:
575 addPass(createCFLAndersAAWrapperPass());
576 addPass(createCFLSteensAAWrapperPass());
582 // Basic AliasAnalysis support.
583 // Add TypeBasedAliasAnalysis before BasicAliasAnalysis so that
584 // BasicAliasAnalysis wins if they disagree. This is intended to help
585 // support "obvious" type-punning idioms.
586 addPass(createTypeBasedAAWrapperPass());
587 addPass(createScopedNoAliasAAWrapperPass());
588 addPass(createBasicAAWrapperPass());
590 // Before running any passes, run the verifier to determine if the input
591 // coming from the front-end and/or optimizer is valid.
593 addPass(createVerifierPass());
595 // Run loop strength reduction before anything else.
596 if (getOptLevel() != CodeGenOpt::None
&& !DisableLSR
) {
597 addPass(createLoopStrengthReducePass());
599 addPass(createPrintFunctionPass(dbgs(), "\n\n*** Code after LSR ***\n"));
602 if (getOptLevel() != CodeGenOpt::None
) {
603 // The MergeICmpsPass tries to create memcmp calls by grouping sequences of
604 // loads and compares. ExpandMemCmpPass then tries to expand those calls
605 // into optimally-sized loads and compares. The transforms are enabled by a
606 // target lowering hook.
607 if (EnableMergeICmps
)
608 addPass(createMergeICmpsPass());
609 addPass(createExpandMemCmpPass());
612 // Run GC lowering passes for builtin collectors
613 // TODO: add a pass insertion point here
614 addPass(createGCLoweringPass());
615 addPass(createShadowStackGCLoweringPass());
617 // Make sure that no unreachable blocks are instruction selected.
618 addPass(createUnreachableBlockEliminationPass());
620 // Prepare expensive constants for SelectionDAG.
621 if (getOptLevel() != CodeGenOpt::None
&& !DisableConstantHoisting
)
622 addPass(createConstantHoistingPass());
624 if (getOptLevel() != CodeGenOpt::None
&& !DisablePartialLibcallInlining
)
625 addPass(createPartiallyInlineLibCallsPass());
627 // Instrument function entry and exit, e.g. with calls to mcount().
628 addPass(createPostInlineEntryExitInstrumenterPass());
630 // Add scalarization of target's unsupported masked memory intrinsics pass.
631 // the unsupported intrinsic will be replaced with a chain of basic blocks,
632 // that stores/loads element one-by-one if the appropriate mask bit is set.
633 addPass(createScalarizeMaskedMemIntrinPass());
635 // Expand reduction intrinsics into shuffle sequences if the target wants to.
636 addPass(createExpandReductionsPass());
639 /// Turn exception handling constructs into something the code generators can
641 void TargetPassConfig::addPassesToHandleExceptions() {
642 const MCAsmInfo
*MCAI
= TM
->getMCAsmInfo();
643 assert(MCAI
&& "No MCAsmInfo");
644 switch (MCAI
->getExceptionHandlingType()) {
645 case ExceptionHandling::SjLj
:
646 // SjLj piggy-backs on dwarf for this bit. The cleanups done apply to both
647 // Dwarf EH prepare needs to be run after SjLj prepare. Otherwise,
648 // catch info can get misplaced when a selector ends up more than one block
649 // removed from the parent invoke(s). This could happen when a landing
650 // pad is shared by multiple invokes and is also a target of a normal
651 // edge from elsewhere.
652 addPass(createSjLjEHPreparePass());
654 case ExceptionHandling::DwarfCFI
:
655 case ExceptionHandling::ARM
:
656 addPass(createDwarfEHPass());
658 case ExceptionHandling::WinEH
:
659 // We support using both GCC-style and MSVC-style exceptions on Windows, so
660 // add both preparation passes. Each pass will only actually run if it
661 // recognizes the personality function.
662 addPass(createWinEHPass());
663 addPass(createDwarfEHPass());
665 case ExceptionHandling::None
:
666 addPass(createLowerInvokePass());
668 // The lower invoke pass may create unreachable code. Remove it.
669 addPass(createUnreachableBlockEliminationPass());
674 /// Add pass to prepare the LLVM IR for code generation. This should be done
675 /// before exception handling preparation passes.
676 void TargetPassConfig::addCodeGenPrepare() {
677 if (getOptLevel() != CodeGenOpt::None
&& !DisableCGP
)
678 addPass(createCodeGenPreparePass());
679 addPass(createRewriteSymbolsPass());
682 /// Add common passes that perform LLVM IR to IR transforms in preparation for
683 /// instruction selection.
684 void TargetPassConfig::addISelPrepare() {
687 // Force codegen to run according to the callgraph.
688 if (requiresCodeGenSCCOrder())
689 addPass(new DummyCGSCCPass
);
691 // Add both the safe stack and the stack protection passes: each of them will
692 // only protect functions that have corresponding attributes.
693 addPass(createSafeStackPass());
694 addPass(createStackProtectorPass());
697 addPass(createPrintFunctionPass(
698 dbgs(), "\n\n*** Final LLVM Code input to ISel ***\n"));
700 // All passes which modify the LLVM IR are now complete; run the verifier
701 // to ensure that the IR is valid.
703 addPass(createVerifierPass());
706 bool TargetPassConfig::addCoreISelPasses() {
707 // Enable FastISel with -fast, but allow that to be overridden.
708 TM
->setO0WantsFastISel(EnableFastISelOption
!= cl::BOU_FALSE
);
709 if (EnableFastISelOption
== cl::BOU_TRUE
||
710 (TM
->getOptLevel() == CodeGenOpt::None
&& TM
->getO0WantsFastISel()))
711 TM
->setFastISel(true);
713 // Ask the target for an isel.
714 // Enable GlobalISel if the target wants to, but allow that to be overriden.
715 if (EnableGlobalISel
== cl::BOU_TRUE
||
716 (EnableGlobalISel
== cl::BOU_UNSET
&& isGlobalISelEnabled())) {
717 if (addIRTranslator())
720 addPreLegalizeMachineIR();
722 if (addLegalizeMachineIR())
725 // Before running the register bank selector, ask the target if it
726 // wants to run some passes.
727 addPreRegBankSelect();
729 if (addRegBankSelect())
732 addPreGlobalInstructionSelect();
734 if (addGlobalInstructionSelect())
737 // Pass to reset the MachineFunction if the ISel failed.
738 addPass(createResetMachineFunctionPass(
739 reportDiagnosticWhenGlobalISelFallback(), isGlobalISelAbortEnabled()));
741 // Provide a fallback path when we do not want to abort on
742 // not-yet-supported input.
743 if (!isGlobalISelAbortEnabled() && addInstSelector())
746 } else if (addInstSelector())
752 bool TargetPassConfig::addISelPasses() {
753 if (TM
->Options
.EmulatedTLS
)
754 addPass(createLowerEmuTLSPass());
756 addPass(createPreISelIntrinsicLoweringPass());
757 addPass(createTargetTransformInfoWrapperPass(TM
->getTargetIRAnalysis()));
760 addPassesToHandleExceptions();
763 return addCoreISelPasses();
766 /// -regalloc=... command line option.
767 static FunctionPass
*useDefaultRegisterAllocator() { return nullptr; }
768 static cl::opt
<RegisterRegAlloc::FunctionPassCtor
, false,
769 RegisterPassParser
<RegisterRegAlloc
>>
770 RegAlloc("regalloc", cl::Hidden
, cl::init(&useDefaultRegisterAllocator
),
771 cl::desc("Register allocator to use"));
773 /// Add the complete set of target-independent postISel code generator passes.
775 /// This can be read as the standard order of major LLVM CodeGen stages. Stages
776 /// with nontrivial configuration or multiple passes are broken out below in
777 /// add%Stage routines.
779 /// Any TargetPassConfig::addXX routine may be overriden by the Target. The
780 /// addPre/Post methods with empty header implementations allow injecting
781 /// target-specific fixups just before or after major stages. Additionally,
782 /// targets have the flexibility to change pass order within a stage by
783 /// overriding default implementation of add%Stage routines below. Each
784 /// technique has maintainability tradeoffs because alternate pass orders are
785 /// not well supported. addPre/Post works better if the target pass is easily
786 /// tied to a common pass. But if it has subtle dependencies on multiple passes,
787 /// the target should override the stage instead.
789 /// TODO: We could use a single addPre/Post(ID) hook to allow pass injection
790 /// before/after any target-independent pass. But it's currently overkill.
791 void TargetPassConfig::addMachinePasses() {
792 AddingMachinePasses
= true;
794 // Insert a machine instr printer pass after the specified pass.
795 if (!StringRef(PrintMachineInstrs
.getValue()).equals("") &&
796 !StringRef(PrintMachineInstrs
.getValue()).equals("option-unspecified")) {
797 const PassRegistry
*PR
= PassRegistry::getPassRegistry();
798 const PassInfo
*TPI
= PR
->getPassInfo(PrintMachineInstrs
.getValue());
799 const PassInfo
*IPI
= PR
->getPassInfo(StringRef("machineinstr-printer"));
800 assert (TPI
&& IPI
&& "Pass ID not registered!");
801 const char *TID
= (const char *)(TPI
->getTypeInfo());
802 const char *IID
= (const char *)(IPI
->getTypeInfo());
803 insertPass(TID
, IID
);
806 // Print the instruction selected machine code...
807 printAndVerify("After Instruction Selection");
809 // Expand pseudo-instructions emitted by ISel.
810 addPass(&ExpandISelPseudosID
);
812 // Add passes that optimize machine instructions in SSA form.
813 if (getOptLevel() != CodeGenOpt::None
) {
814 addMachineSSAOptimization();
816 // If the target requests it, assign local variables to stack slots relative
817 // to one another and simplify frame index references where possible.
818 addPass(&LocalStackSlotAllocationID
, false);
821 if (TM
->Options
.EnableIPRA
)
822 addPass(createRegUsageInfoPropPass());
824 // Run pre-ra passes.
827 // Run register allocation and passes that are tightly coupled with it,
828 // including phi elimination and scheduling.
829 if (getOptimizeRegAlloc())
830 addOptimizedRegAlloc(createRegAllocPass(true));
832 if (RegAlloc
!= &useDefaultRegisterAllocator
&&
833 RegAlloc
!= &createFastRegisterAllocator
)
834 report_fatal_error("Must use fast (default) register allocator for unoptimized regalloc.");
835 addFastRegAlloc(createRegAllocPass(false));
838 // Run post-ra passes.
841 // Insert prolog/epilog code. Eliminate abstract frame index references...
842 if (getOptLevel() != CodeGenOpt::None
)
843 addPass(&ShrinkWrapID
);
845 // Prolog/Epilog inserter needs a TargetMachine to instantiate. But only
846 // do so if it hasn't been disabled, substituted, or overridden.
847 if (!isPassSubstitutedOrOverridden(&PrologEpilogCodeInserterID
))
848 addPass(createPrologEpilogInserterPass());
850 /// Add passes that optimize machine instructions after register allocation.
851 if (getOptLevel() != CodeGenOpt::None
)
852 addMachineLateOptimization();
854 // Expand pseudo instructions before second scheduling pass.
855 addPass(&ExpandPostRAPseudosID
);
857 // Run pre-sched2 passes.
860 if (EnableImplicitNullChecks
)
861 addPass(&ImplicitNullChecksID
);
863 // Second pass scheduler.
864 // Let Target optionally insert this pass by itself at some other
866 if (getOptLevel() != CodeGenOpt::None
&&
867 !TM
->targetSchedulesPostRAScheduling()) {
869 addPass(&PostMachineSchedulerID
);
871 addPass(&PostRASchedulerID
);
877 addPass(createGCInfoPrinter(dbgs()), false, false);
880 // Basic block placement.
881 if (getOptLevel() != CodeGenOpt::None
)
886 if (TM
->Options
.EnableIPRA
)
887 // Collect register usage information and produce a register mask of
888 // clobbered registers, to be used to optimize call sites.
889 addPass(createRegUsageInfoCollector());
891 addPass(&FuncletLayoutID
, false);
893 addPass(&StackMapLivenessID
, false);
894 addPass(&LiveDebugValuesID
, false);
896 // Insert before XRay Instrumentation.
897 addPass(&FEntryInserterID
, false);
899 addPass(&XRayInstrumentationID
, false);
900 addPass(&PatchableFunctionID
, false);
902 if (EnableMachineOutliner
)
903 PM
->add(createMachineOutlinerPass(EnableLinkOnceODROutlining
));
905 AddingMachinePasses
= false;
908 /// Add passes that optimize machine instructions in SSA form.
909 void TargetPassConfig::addMachineSSAOptimization() {
910 // Pre-ra tail duplication.
911 addPass(&EarlyTailDuplicateID
);
913 // Optimize PHIs before DCE: removing dead PHI cycles may make more
914 // instructions dead.
915 addPass(&OptimizePHIsID
, false);
917 // This pass merges large allocas. StackSlotColoring is a different pass
918 // which merges spill slots.
919 addPass(&StackColoringID
, false);
921 // If the target requests it, assign local variables to stack slots relative
922 // to one another and simplify frame index references where possible.
923 addPass(&LocalStackSlotAllocationID
, false);
925 // With optimization, dead code should already be eliminated. However
926 // there is one known exception: lowered code for arguments that are only
927 // used by tail calls, where the tail calls reuse the incoming stack
928 // arguments directly (see t11 in test/CodeGen/X86/sibcall.ll).
929 addPass(&DeadMachineInstructionElimID
);
931 // Allow targets to insert passes that improve instruction level parallelism,
932 // like if-conversion. Such passes will typically need dominator trees and
933 // loop info, just like LICM and CSE below.
936 addPass(&MachineLICMID
, false);
937 addPass(&MachineCSEID
, false);
939 addPass(&MachineSinkingID
);
941 addPass(&PeepholeOptimizerID
);
942 // Clean-up the dead code that may have been generated by peephole
944 addPass(&DeadMachineInstructionElimID
);
947 //===---------------------------------------------------------------------===//
948 /// Register Allocation Pass Configuration
949 //===---------------------------------------------------------------------===//
951 bool TargetPassConfig::getOptimizeRegAlloc() const {
952 switch (OptimizeRegAlloc
) {
953 case cl::BOU_UNSET
: return getOptLevel() != CodeGenOpt::None
;
954 case cl::BOU_TRUE
: return true;
955 case cl::BOU_FALSE
: return false;
957 llvm_unreachable("Invalid optimize-regalloc state");
960 /// RegisterRegAlloc's global Registry tracks allocator registration.
961 MachinePassRegistry
RegisterRegAlloc::Registry
;
963 /// A dummy default pass factory indicates whether the register allocator is
964 /// overridden on the command line.
965 static llvm::once_flag InitializeDefaultRegisterAllocatorFlag
;
967 static RegisterRegAlloc
968 defaultRegAlloc("default",
969 "pick register allocator based on -O option",
970 useDefaultRegisterAllocator
);
972 static void initializeDefaultRegisterAllocatorOnce() {
973 RegisterRegAlloc::FunctionPassCtor Ctor
= RegisterRegAlloc::getDefault();
977 RegisterRegAlloc::setDefault(RegAlloc
);
981 /// Instantiate the default register allocator pass for this target for either
982 /// the optimized or unoptimized allocation path. This will be added to the pass
983 /// manager by addFastRegAlloc in the unoptimized case or addOptimizedRegAlloc
984 /// in the optimized case.
986 /// A target that uses the standard regalloc pass order for fast or optimized
987 /// allocation may still override this for per-target regalloc
988 /// selection. But -regalloc=... always takes precedence.
989 FunctionPass
*TargetPassConfig::createTargetRegisterAllocator(bool Optimized
) {
991 return createGreedyRegisterAllocator();
993 return createFastRegisterAllocator();
996 /// Find and instantiate the register allocation pass requested by this target
997 /// at the current optimization level. Different register allocators are
998 /// defined as separate passes because they may require different analysis.
1000 /// This helper ensures that the regalloc= option is always available,
1001 /// even for targets that override the default allocator.
1003 /// FIXME: When MachinePassRegistry register pass IDs instead of function ptrs,
1004 /// this can be folded into addPass.
1005 FunctionPass
*TargetPassConfig::createRegAllocPass(bool Optimized
) {
1006 // Initialize the global default.
1007 llvm::call_once(InitializeDefaultRegisterAllocatorFlag
,
1008 initializeDefaultRegisterAllocatorOnce
);
1010 RegisterRegAlloc::FunctionPassCtor Ctor
= RegisterRegAlloc::getDefault();
1011 if (Ctor
!= useDefaultRegisterAllocator
)
1014 // With no -regalloc= override, ask the target for a regalloc pass.
1015 return createTargetRegisterAllocator(Optimized
);
1018 /// Return true if the default global register allocator is in use and
1019 /// has not be overriden on the command line with '-regalloc=...'
1020 bool TargetPassConfig::usingDefaultRegAlloc() const {
1021 return RegAlloc
.getNumOccurrences() == 0;
1024 /// Add the minimum set of target-independent passes that are required for
1025 /// register allocation. No coalescing or scheduling.
1026 void TargetPassConfig::addFastRegAlloc(FunctionPass
*RegAllocPass
) {
1027 addPass(&PHIEliminationID
, false);
1028 addPass(&TwoAddressInstructionPassID
, false);
1031 addPass(RegAllocPass
);
1034 /// Add standard target-independent passes that are tightly coupled with
1035 /// optimized register allocation, including coalescing, machine instruction
1036 /// scheduling, and register allocation itself.
1037 void TargetPassConfig::addOptimizedRegAlloc(FunctionPass
*RegAllocPass
) {
1038 addPass(&DetectDeadLanesID
, false);
1040 addPass(&ProcessImplicitDefsID
, false);
1042 // LiveVariables currently requires pure SSA form.
1044 // FIXME: Once TwoAddressInstruction pass no longer uses kill flags,
1045 // LiveVariables can be removed completely, and LiveIntervals can be directly
1046 // computed. (We still either need to regenerate kill flags after regalloc, or
1047 // preferably fix the scavenger to not depend on them).
1048 addPass(&LiveVariablesID
, false);
1050 // Edge splitting is smarter with machine loop info.
1051 addPass(&MachineLoopInfoID
, false);
1052 addPass(&PHIEliminationID
, false);
1054 // Eventually, we want to run LiveIntervals before PHI elimination.
1055 if (EarlyLiveIntervals
)
1056 addPass(&LiveIntervalsID
, false);
1058 addPass(&TwoAddressInstructionPassID
, false);
1059 addPass(&RegisterCoalescerID
);
1061 // The machine scheduler may accidentally create disconnected components
1062 // when moving subregister definitions around, avoid this by splitting them to
1063 // separate vregs before. Splitting can also improve reg. allocation quality.
1064 addPass(&RenameIndependentSubregsID
);
1066 // PreRA instruction scheduling.
1067 addPass(&MachineSchedulerID
);
1070 // Add the selected register allocation pass.
1071 addPass(RegAllocPass
);
1073 // Allow targets to change the register assignments before rewriting.
1076 // Finally rewrite virtual registers.
1077 addPass(&VirtRegRewriterID
);
1079 // Perform stack slot coloring and post-ra machine LICM.
1081 // FIXME: Re-enable coloring with register when it's capable of adding
1083 addPass(&StackSlotColoringID
);
1085 // Run post-ra machine LICM to hoist reloads / remats.
1087 // FIXME: can this move into MachineLateOptimization?
1088 addPass(&PostRAMachineLICMID
);
1092 //===---------------------------------------------------------------------===//
1093 /// Post RegAlloc Pass Configuration
1094 //===---------------------------------------------------------------------===//
1096 /// Add passes that optimize machine instructions after register allocation.
1097 void TargetPassConfig::addMachineLateOptimization() {
1098 // Branch folding must be run after regalloc and prolog/epilog insertion.
1099 addPass(&BranchFolderPassID
);
1101 // Tail duplication.
1102 // Note that duplicating tail just increases code size and degrades
1103 // performance for targets that require Structured Control Flow.
1104 // In addition it can also make CFG irreducible. Thus we disable it.
1105 if (!TM
->requiresStructuredCFG())
1106 addPass(&TailDuplicateID
);
1108 // Copy propagation.
1109 addPass(&MachineCopyPropagationID
);
1112 /// Add standard GC passes.
1113 bool TargetPassConfig::addGCPasses() {
1114 addPass(&GCMachineCodeAnalysisID
, false);
1118 /// Add standard basic block placement passes.
1119 void TargetPassConfig::addBlockPlacement() {
1120 if (addPass(&MachineBlockPlacementID
)) {
1121 // Run a separate pass to collect block placement statistics.
1122 if (EnableBlockPlacementStats
)
1123 addPass(&MachineBlockPlacementStatsID
);
1127 //===---------------------------------------------------------------------===//
1128 /// GlobalISel Configuration
1129 //===---------------------------------------------------------------------===//
1131 bool TargetPassConfig::isGlobalISelEnabled() const {
1135 bool TargetPassConfig::isGlobalISelAbortEnabled() const {
1136 return EnableGlobalISelAbort
== 1;
1139 bool TargetPassConfig::reportDiagnosticWhenGlobalISelFallback() const {
1140 return EnableGlobalISelAbort
== 2;