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/Support/SaveAndRestore.h"
43 #include "llvm/Target/TargetMachine.h"
44 #include "llvm/Transforms/Scalar.h"
45 #include "llvm/Transforms/Utils.h"
46 #include "llvm/Transforms/Utils/SymbolRewriter.h"
52 cl::opt
<bool> EnableIPRA("enable-ipra", cl::init(false), cl::Hidden
,
53 cl::desc("Enable interprocedural register allocation "
54 "to reduce load/store at procedure calls."));
55 static cl::opt
<bool> DisablePostRASched("disable-post-ra", cl::Hidden
,
56 cl::desc("Disable Post Regalloc Scheduler"));
57 static cl::opt
<bool> DisableBranchFold("disable-branch-fold", cl::Hidden
,
58 cl::desc("Disable branch folding"));
59 static cl::opt
<bool> DisableTailDuplicate("disable-tail-duplicate", cl::Hidden
,
60 cl::desc("Disable tail duplication"));
61 static cl::opt
<bool> DisableEarlyTailDup("disable-early-taildup", cl::Hidden
,
62 cl::desc("Disable pre-register allocation tail duplication"));
63 static cl::opt
<bool> DisableBlockPlacement("disable-block-placement",
64 cl::Hidden
, cl::desc("Disable probability-driven block placement"));
65 static cl::opt
<bool> EnableBlockPlacementStats("enable-block-placement-stats",
66 cl::Hidden
, cl::desc("Collect probability-driven block placement stats"));
67 static cl::opt
<bool> DisableSSC("disable-ssc", cl::Hidden
,
68 cl::desc("Disable Stack Slot Coloring"));
69 static cl::opt
<bool> DisableMachineDCE("disable-machine-dce", cl::Hidden
,
70 cl::desc("Disable Machine Dead Code Elimination"));
71 static cl::opt
<bool> DisableEarlyIfConversion("disable-early-ifcvt", cl::Hidden
,
72 cl::desc("Disable Early If-conversion"));
73 static cl::opt
<bool> DisableMachineLICM("disable-machine-licm", cl::Hidden
,
74 cl::desc("Disable Machine LICM"));
75 static cl::opt
<bool> DisableMachineCSE("disable-machine-cse", cl::Hidden
,
76 cl::desc("Disable Machine Common Subexpression Elimination"));
77 static cl::opt
<cl::boolOrDefault
> OptimizeRegAlloc(
78 "optimize-regalloc", cl::Hidden
,
79 cl::desc("Enable optimized register allocation compilation path."));
80 static cl::opt
<bool> DisablePostRAMachineLICM("disable-postra-machine-licm",
82 cl::desc("Disable Machine LICM"));
83 static cl::opt
<bool> DisableMachineSink("disable-machine-sink", cl::Hidden
,
84 cl::desc("Disable Machine Sinking"));
85 static cl::opt
<bool> DisablePostRAMachineSink("disable-postra-machine-sink",
87 cl::desc("Disable PostRA Machine Sinking"));
88 static cl::opt
<bool> DisableLSR("disable-lsr", cl::Hidden
,
89 cl::desc("Disable Loop Strength Reduction Pass"));
90 static cl::opt
<bool> DisableConstantHoisting("disable-constant-hoisting",
91 cl::Hidden
, cl::desc("Disable ConstantHoisting"));
92 static cl::opt
<bool> DisableCGP("disable-cgp", cl::Hidden
,
93 cl::desc("Disable Codegen Prepare"));
94 static cl::opt
<bool> DisableCopyProp("disable-copyprop", cl::Hidden
,
95 cl::desc("Disable Copy Propagation pass"));
96 static cl::opt
<bool> DisablePartialLibcallInlining("disable-partial-libcall-inlining",
97 cl::Hidden
, cl::desc("Disable Partial Libcall Inlining"));
98 static cl::opt
<bool> EnableImplicitNullChecks(
99 "enable-implicit-null-checks",
100 cl::desc("Fold null checks into faulting memory operations"),
101 cl::init(false), cl::Hidden
);
102 static cl::opt
<bool> DisableMergeICmps("disable-mergeicmps",
103 cl::desc("Disable MergeICmps Pass"),
104 cl::init(false), cl::Hidden
);
105 static cl::opt
<bool> PrintLSR("print-lsr-output", cl::Hidden
,
106 cl::desc("Print LLVM IR produced by the loop-reduce pass"));
107 static cl::opt
<bool> PrintISelInput("print-isel-input", cl::Hidden
,
108 cl::desc("Print LLVM IR input to isel pass"));
109 static cl::opt
<bool> PrintGCInfo("print-gc", cl::Hidden
,
110 cl::desc("Dump garbage collector data"));
111 static cl::opt
<cl::boolOrDefault
>
112 VerifyMachineCode("verify-machineinstrs", cl::Hidden
,
113 cl::desc("Verify generated machine code"),
115 enum RunOutliner
{ AlwaysOutline
, NeverOutline
, TargetDefault
};
116 // Enable or disable the MachineOutliner.
117 static cl::opt
<RunOutliner
> EnableMachineOutliner(
118 "enable-machine-outliner", cl::desc("Enable the machine outliner"),
119 cl::Hidden
, cl::ValueOptional
, cl::init(TargetDefault
),
120 cl::values(clEnumValN(AlwaysOutline
, "always",
121 "Run on all functions guaranteed to be beneficial"),
122 clEnumValN(NeverOutline
, "never", "Disable all outlining"),
123 // Sentinel value for unspecified option.
124 clEnumValN(AlwaysOutline
, "", "")));
125 // Enable or disable FastISel. Both options are needed, because
126 // FastISel is enabled by default with -fast, and we wish to be
127 // able to enable or disable fast-isel independently from -O0.
128 static cl::opt
<cl::boolOrDefault
>
129 EnableFastISelOption("fast-isel", cl::Hidden
,
130 cl::desc("Enable the \"fast\" instruction selector"));
132 static cl::opt
<cl::boolOrDefault
> EnableGlobalISelOption(
133 "global-isel", cl::Hidden
,
134 cl::desc("Enable the \"global\" instruction selector"));
136 static cl::opt
<std::string
> PrintMachineInstrs(
137 "print-machineinstrs", cl::ValueOptional
, cl::desc("Print machine instrs"),
138 cl::value_desc("pass-name"), cl::init("option-unspecified"), cl::Hidden
);
140 static cl::opt
<int> EnableGlobalISelAbort(
141 "global-isel-abort", cl::Hidden
,
142 cl::desc("Enable abort calls when \"global\" instruction selection "
143 "fails to lower/select an instruction: 0 disable the abort, "
144 "1 enable the abort, and "
145 "2 disable the abort but emit a diagnostic on failure"),
148 // Temporary option to allow experimenting with MachineScheduler as a post-RA
149 // scheduler. Targets can "properly" enable this with
150 // substitutePass(&PostRASchedulerID, &PostMachineSchedulerID).
151 // Targets can return true in targetSchedulesPostRAScheduling() and
152 // insert a PostRA scheduling pass wherever it wants.
153 cl::opt
<bool> MISchedPostRA("misched-postra", cl::Hidden
,
154 cl::desc("Run MachineScheduler post regalloc (independent of preRA sched)"));
156 // Experimental option to run live interval analysis early.
157 static cl::opt
<bool> EarlyLiveIntervals("early-live-intervals", cl::Hidden
,
158 cl::desc("Run live interval analysis earlier in the pipeline"));
160 // Experimental option to use CFL-AA in codegen
161 enum class CFLAAType
{ None
, Steensgaard
, Andersen
, Both
};
162 static cl::opt
<CFLAAType
> UseCFLAA(
163 "use-cfl-aa-in-codegen", cl::init(CFLAAType::None
), cl::Hidden
,
164 cl::desc("Enable the new, experimental CFL alias analysis in CodeGen"),
165 cl::values(clEnumValN(CFLAAType::None
, "none", "Disable CFL-AA"),
166 clEnumValN(CFLAAType::Steensgaard
, "steens",
167 "Enable unification-based CFL-AA"),
168 clEnumValN(CFLAAType::Andersen
, "anders",
169 "Enable inclusion-based CFL-AA"),
170 clEnumValN(CFLAAType::Both
, "both",
171 "Enable both variants of CFL-AA")));
173 /// Option names for limiting the codegen pipeline.
174 /// Those are used in error reporting and we didn't want
175 /// to duplicate their names all over the place.
176 const char *StartAfterOptName
= "start-after";
177 const char *StartBeforeOptName
= "start-before";
178 const char *StopAfterOptName
= "stop-after";
179 const char *StopBeforeOptName
= "stop-before";
181 static cl::opt
<std::string
>
182 StartAfterOpt(StringRef(StartAfterOptName
),
183 cl::desc("Resume compilation after a specific pass"),
184 cl::value_desc("pass-name"), cl::init(""), cl::Hidden
);
186 static cl::opt
<std::string
>
187 StartBeforeOpt(StringRef(StartBeforeOptName
),
188 cl::desc("Resume compilation before a specific pass"),
189 cl::value_desc("pass-name"), cl::init(""), cl::Hidden
);
191 static cl::opt
<std::string
>
192 StopAfterOpt(StringRef(StopAfterOptName
),
193 cl::desc("Stop compilation after a specific pass"),
194 cl::value_desc("pass-name"), cl::init(""), cl::Hidden
);
196 static cl::opt
<std::string
>
197 StopBeforeOpt(StringRef(StopBeforeOptName
),
198 cl::desc("Stop compilation before a specific pass"),
199 cl::value_desc("pass-name"), cl::init(""), cl::Hidden
);
201 /// Allow standard passes to be disabled by command line options. This supports
202 /// simple binary flags that either suppress the pass or do nothing.
203 /// i.e. -disable-mypass=false has no effect.
204 /// These should be converted to boolOrDefault in order to use applyOverride.
205 static IdentifyingPassPtr
applyDisable(IdentifyingPassPtr PassID
,
208 return IdentifyingPassPtr();
212 /// Allow standard passes to be disabled by the command line, regardless of who
213 /// is adding the pass.
215 /// StandardID is the pass identified in the standard pass pipeline and provided
216 /// to addPass(). It may be a target-specific ID in the case that the target
217 /// directly adds its own pass, but in that case we harmlessly fall through.
219 /// TargetID is the pass that the target has configured to override StandardID.
221 /// StandardID may be a pseudo ID. In that case TargetID is the name of the real
222 /// pass to run. This allows multiple options to control a single pass depending
223 /// on where in the pipeline that pass is added.
224 static IdentifyingPassPtr
overridePass(AnalysisID StandardID
,
225 IdentifyingPassPtr TargetID
) {
226 if (StandardID
== &PostRASchedulerID
)
227 return applyDisable(TargetID
, DisablePostRASched
);
229 if (StandardID
== &BranchFolderPassID
)
230 return applyDisable(TargetID
, DisableBranchFold
);
232 if (StandardID
== &TailDuplicateID
)
233 return applyDisable(TargetID
, DisableTailDuplicate
);
235 if (StandardID
== &EarlyTailDuplicateID
)
236 return applyDisable(TargetID
, DisableEarlyTailDup
);
238 if (StandardID
== &MachineBlockPlacementID
)
239 return applyDisable(TargetID
, DisableBlockPlacement
);
241 if (StandardID
== &StackSlotColoringID
)
242 return applyDisable(TargetID
, DisableSSC
);
244 if (StandardID
== &DeadMachineInstructionElimID
)
245 return applyDisable(TargetID
, DisableMachineDCE
);
247 if (StandardID
== &EarlyIfConverterID
)
248 return applyDisable(TargetID
, DisableEarlyIfConversion
);
250 if (StandardID
== &EarlyMachineLICMID
)
251 return applyDisable(TargetID
, DisableMachineLICM
);
253 if (StandardID
== &MachineCSEID
)
254 return applyDisable(TargetID
, DisableMachineCSE
);
256 if (StandardID
== &MachineLICMID
)
257 return applyDisable(TargetID
, DisablePostRAMachineLICM
);
259 if (StandardID
== &MachineSinkingID
)
260 return applyDisable(TargetID
, DisableMachineSink
);
262 if (StandardID
== &PostRAMachineSinkingID
)
263 return applyDisable(TargetID
, DisablePostRAMachineSink
);
265 if (StandardID
== &MachineCopyPropagationID
)
266 return applyDisable(TargetID
, DisableCopyProp
);
271 //===---------------------------------------------------------------------===//
273 //===---------------------------------------------------------------------===//
275 INITIALIZE_PASS(TargetPassConfig
, "targetpassconfig",
276 "Target Pass Configuration", false, false)
277 char TargetPassConfig::ID
= 0;
281 struct InsertedPass
{
282 AnalysisID TargetPassID
;
283 IdentifyingPassPtr InsertedPassID
;
287 InsertedPass(AnalysisID TargetPassID
, IdentifyingPassPtr InsertedPassID
,
288 bool VerifyAfter
, bool PrintAfter
)
289 : TargetPassID(TargetPassID
), InsertedPassID(InsertedPassID
),
290 VerifyAfter(VerifyAfter
), PrintAfter(PrintAfter
) {}
292 Pass
*getInsertedPass() const {
293 assert(InsertedPassID
.isValid() && "Illegal Pass ID!");
294 if (InsertedPassID
.isInstance())
295 return InsertedPassID
.getInstance();
296 Pass
*NP
= Pass::createPass(InsertedPassID
.getID());
297 assert(NP
&& "Pass ID not registered");
302 } // end anonymous namespace
306 class PassConfigImpl
{
308 // List of passes explicitly substituted by this target. Normally this is
309 // empty, but it is a convenient way to suppress or replace specific passes
310 // that are part of a standard pass pipeline without overridding the entire
311 // pipeline. This mechanism allows target options to inherit a standard pass's
312 // user interface. For example, a target may disable a standard pass by
313 // default by substituting a pass ID of zero, and the user may still enable
314 // that standard pass with an explicit command line option.
315 DenseMap
<AnalysisID
,IdentifyingPassPtr
> TargetPasses
;
317 /// Store the pairs of <AnalysisID, AnalysisID> of which the second pass
318 /// is inserted after each instance of the first one.
319 SmallVector
<InsertedPass
, 4> InsertedPasses
;
322 } // end namespace llvm
324 // Out of line virtual method.
325 TargetPassConfig::~TargetPassConfig() {
329 static const PassInfo
*getPassInfo(StringRef PassName
) {
330 if (PassName
.empty())
333 const PassRegistry
&PR
= *PassRegistry::getPassRegistry();
334 const PassInfo
*PI
= PR
.getPassInfo(PassName
);
336 report_fatal_error(Twine('\"') + Twine(PassName
) +
337 Twine("\" pass is not registered."));
341 static AnalysisID
getPassIDFromName(StringRef PassName
) {
342 const PassInfo
*PI
= getPassInfo(PassName
);
343 return PI
? PI
->getTypeInfo() : nullptr;
346 void TargetPassConfig::setStartStopPasses() {
347 StartBefore
= getPassIDFromName(StartBeforeOpt
);
348 StartAfter
= getPassIDFromName(StartAfterOpt
);
349 StopBefore
= getPassIDFromName(StopBeforeOpt
);
350 StopAfter
= getPassIDFromName(StopAfterOpt
);
351 if (StartBefore
&& StartAfter
)
352 report_fatal_error(Twine(StartBeforeOptName
) + Twine(" and ") +
353 Twine(StartAfterOptName
) + Twine(" specified!"));
354 if (StopBefore
&& StopAfter
)
355 report_fatal_error(Twine(StopBeforeOptName
) + Twine(" and ") +
356 Twine(StopAfterOptName
) + Twine(" specified!"));
357 Started
= (StartAfter
== nullptr) && (StartBefore
== nullptr);
360 // Out of line constructor provides default values for pass options and
361 // registers all common codegen passes.
362 TargetPassConfig::TargetPassConfig(LLVMTargetMachine
&TM
, PassManagerBase
&pm
)
363 : ImmutablePass(ID
), PM(&pm
), TM(&TM
) {
364 Impl
= new PassConfigImpl();
366 // Register all target independent codegen passes to activate their PassIDs,
367 // including this pass itself.
368 initializeCodeGen(*PassRegistry::getPassRegistry());
370 // Also register alias analysis passes required by codegen passes.
371 initializeBasicAAWrapperPassPass(*PassRegistry::getPassRegistry());
372 initializeAAResultsWrapperPassPass(*PassRegistry::getPassRegistry());
374 if (StringRef(PrintMachineInstrs
.getValue()).equals(""))
375 TM
.Options
.PrintMachineCode
= true;
377 if (EnableIPRA
.getNumOccurrences())
378 TM
.Options
.EnableIPRA
= EnableIPRA
;
380 // If not explicitly specified, use target default.
381 TM
.Options
.EnableIPRA
= TM
.useIPRA();
384 if (TM
.Options
.EnableIPRA
)
385 setRequiresCodeGenSCCOrder();
387 setStartStopPasses();
390 CodeGenOpt::Level
TargetPassConfig::getOptLevel() const {
391 return TM
->getOptLevel();
394 /// Insert InsertedPassID pass after TargetPassID.
395 void TargetPassConfig::insertPass(AnalysisID TargetPassID
,
396 IdentifyingPassPtr InsertedPassID
,
397 bool VerifyAfter
, bool PrintAfter
) {
398 assert(((!InsertedPassID
.isInstance() &&
399 TargetPassID
!= InsertedPassID
.getID()) ||
400 (InsertedPassID
.isInstance() &&
401 TargetPassID
!= InsertedPassID
.getInstance()->getPassID())) &&
402 "Insert a pass after itself!");
403 Impl
->InsertedPasses
.emplace_back(TargetPassID
, InsertedPassID
, VerifyAfter
,
407 /// createPassConfig - Create a pass configuration object to be used by
408 /// addPassToEmitX methods for generating a pipeline of CodeGen passes.
410 /// Targets may override this to extend TargetPassConfig.
411 TargetPassConfig
*LLVMTargetMachine::createPassConfig(PassManagerBase
&PM
) {
412 return new TargetPassConfig(*this, PM
);
415 TargetPassConfig::TargetPassConfig()
416 : ImmutablePass(ID
) {
417 report_fatal_error("Trying to construct TargetPassConfig without a target "
418 "machine. Scheduling a CodeGen pass without a target "
422 bool TargetPassConfig::hasLimitedCodeGenPipeline() const {
423 return StartBefore
|| StartAfter
|| StopBefore
|| StopAfter
;
427 TargetPassConfig::getLimitedCodeGenPipelineReason(const char *Separator
) const {
428 if (!hasLimitedCodeGenPipeline())
429 return std::string();
431 static cl::opt
<std::string
> *PassNames
[] = {&StartAfterOpt
, &StartBeforeOpt
,
432 &StopAfterOpt
, &StopBeforeOpt
};
433 static const char *OptNames
[] = {StartAfterOptName
, StartBeforeOptName
,
434 StopAfterOptName
, StopBeforeOptName
};
436 for (int Idx
= 0; Idx
< 4; ++Idx
)
437 if (!PassNames
[Idx
]->empty()) {
441 Res
+= OptNames
[Idx
];
446 // Helper to verify the analysis is really immutable.
447 void TargetPassConfig::setOpt(bool &Opt
, bool Val
) {
448 assert(!Initialized
&& "PassConfig is immutable");
452 void TargetPassConfig::substitutePass(AnalysisID StandardID
,
453 IdentifyingPassPtr TargetID
) {
454 Impl
->TargetPasses
[StandardID
] = TargetID
;
457 IdentifyingPassPtr
TargetPassConfig::getPassSubstitution(AnalysisID ID
) const {
458 DenseMap
<AnalysisID
, IdentifyingPassPtr
>::const_iterator
459 I
= Impl
->TargetPasses
.find(ID
);
460 if (I
== Impl
->TargetPasses
.end())
465 bool TargetPassConfig::isPassSubstitutedOrOverridden(AnalysisID ID
) const {
466 IdentifyingPassPtr TargetID
= getPassSubstitution(ID
);
467 IdentifyingPassPtr FinalPtr
= overridePass(ID
, TargetID
);
468 return !FinalPtr
.isValid() || FinalPtr
.isInstance() ||
469 FinalPtr
.getID() != ID
;
472 /// Add a pass to the PassManager if that pass is supposed to be run. If the
473 /// Started/Stopped flags indicate either that the compilation should start at
474 /// a later pass or that it should stop after an earlier pass, then do not add
475 /// the pass. Finally, compare the current pass against the StartAfter
476 /// and StopAfter options and change the Started/Stopped flags accordingly.
477 void TargetPassConfig::addPass(Pass
*P
, bool verifyAfter
, bool printAfter
) {
478 assert(!Initialized
&& "PassConfig is immutable");
480 // Cache the Pass ID here in case the pass manager finds this pass is
481 // redundant with ones already scheduled / available, and deletes it.
482 // Fundamentally, once we add the pass to the manager, we no longer own it
483 // and shouldn't reference it.
484 AnalysisID PassID
= P
->getPassID();
486 if (StartBefore
== PassID
)
488 if (StopBefore
== PassID
)
490 if (Started
&& !Stopped
) {
492 // Construct banner message before PM->add() as that may delete the pass.
493 if (AddingMachinePasses
&& (printAfter
|| verifyAfter
))
494 Banner
= std::string("After ") + std::string(P
->getPassName());
496 if (AddingMachinePasses
) {
498 addPrintPass(Banner
);
500 addVerifyPass(Banner
);
503 // Add the passes after the pass P if there is any.
504 for (auto IP
: Impl
->InsertedPasses
) {
505 if (IP
.TargetPassID
== PassID
)
506 addPass(IP
.getInsertedPass(), IP
.VerifyAfter
, IP
.PrintAfter
);
511 if (StopAfter
== PassID
)
513 if (StartAfter
== PassID
)
515 if (Stopped
&& !Started
)
516 report_fatal_error("Cannot stop compilation after pass that is not run");
519 /// Add a CodeGen pass at this point in the pipeline after checking for target
520 /// and command line overrides.
522 /// addPass cannot return a pointer to the pass instance because is internal the
523 /// PassManager and the instance we create here may already be freed.
524 AnalysisID
TargetPassConfig::addPass(AnalysisID PassID
, bool verifyAfter
,
526 IdentifyingPassPtr TargetID
= getPassSubstitution(PassID
);
527 IdentifyingPassPtr FinalPtr
= overridePass(PassID
, TargetID
);
528 if (!FinalPtr
.isValid())
532 if (FinalPtr
.isInstance())
533 P
= FinalPtr
.getInstance();
535 P
= Pass::createPass(FinalPtr
.getID());
537 llvm_unreachable("Pass ID not registered");
539 AnalysisID FinalID
= P
->getPassID();
540 addPass(P
, verifyAfter
, printAfter
); // Ends the lifetime of P.
545 void TargetPassConfig::printAndVerify(const std::string
&Banner
) {
546 addPrintPass(Banner
);
547 addVerifyPass(Banner
);
550 void TargetPassConfig::addPrintPass(const std::string
&Banner
) {
551 if (TM
->shouldPrintMachineCode())
552 PM
->add(createMachineFunctionPrinterPass(dbgs(), Banner
));
555 void TargetPassConfig::addVerifyPass(const std::string
&Banner
) {
556 bool Verify
= VerifyMachineCode
== cl::BOU_TRUE
;
557 #ifdef EXPENSIVE_CHECKS
558 if (VerifyMachineCode
== cl::BOU_UNSET
)
559 Verify
= TM
->isMachineVerifierClean();
562 PM
->add(createMachineVerifierPass(Banner
));
565 /// Add common target configurable passes that perform LLVM IR to IR transforms
566 /// following machine independent optimization.
567 void TargetPassConfig::addIRPasses() {
569 case CFLAAType::Steensgaard
:
570 addPass(createCFLSteensAAWrapperPass());
572 case CFLAAType::Andersen
:
573 addPass(createCFLAndersAAWrapperPass());
575 case CFLAAType::Both
:
576 addPass(createCFLAndersAAWrapperPass());
577 addPass(createCFLSteensAAWrapperPass());
583 // Basic AliasAnalysis support.
584 // Add TypeBasedAliasAnalysis before BasicAliasAnalysis so that
585 // BasicAliasAnalysis wins if they disagree. This is intended to help
586 // support "obvious" type-punning idioms.
587 addPass(createTypeBasedAAWrapperPass());
588 addPass(createScopedNoAliasAAWrapperPass());
589 addPass(createBasicAAWrapperPass());
591 // Before running any passes, run the verifier to determine if the input
592 // coming from the front-end and/or optimizer is valid.
594 addPass(createVerifierPass());
596 // Run loop strength reduction before anything else.
597 if (getOptLevel() != CodeGenOpt::None
&& !DisableLSR
) {
598 addPass(createLoopStrengthReducePass());
600 addPass(createPrintFunctionPass(dbgs(), "\n\n*** Code after LSR ***\n"));
603 if (getOptLevel() != CodeGenOpt::None
) {
604 // The MergeICmpsPass tries to create memcmp calls by grouping sequences of
605 // loads and compares. ExpandMemCmpPass then tries to expand those calls
606 // into optimally-sized loads and compares. The transforms are enabled by a
607 // target lowering hook.
608 if (!DisableMergeICmps
)
609 addPass(createMergeICmpsPass());
610 addPass(createExpandMemCmpPass());
613 // Run GC lowering passes for builtin collectors
614 // TODO: add a pass insertion point here
615 addPass(createGCLoweringPass());
616 addPass(createShadowStackGCLoweringPass());
618 // Make sure that no unreachable blocks are instruction selected.
619 addPass(createUnreachableBlockEliminationPass());
621 // Prepare expensive constants for SelectionDAG.
622 if (getOptLevel() != CodeGenOpt::None
&& !DisableConstantHoisting
)
623 addPass(createConstantHoistingPass());
625 if (getOptLevel() != CodeGenOpt::None
&& !DisablePartialLibcallInlining
)
626 addPass(createPartiallyInlineLibCallsPass());
628 // Instrument function entry and exit, e.g. with calls to mcount().
629 addPass(createPostInlineEntryExitInstrumenterPass());
631 // Add scalarization of target's unsupported masked memory intrinsics pass.
632 // the unsupported intrinsic will be replaced with a chain of basic blocks,
633 // that stores/loads element one-by-one if the appropriate mask bit is set.
634 addPass(createScalarizeMaskedMemIntrinPass());
636 // Expand reduction intrinsics into shuffle sequences if the target wants to.
637 addPass(createExpandReductionsPass());
640 /// Turn exception handling constructs into something the code generators can
642 void TargetPassConfig::addPassesToHandleExceptions() {
643 const MCAsmInfo
*MCAI
= TM
->getMCAsmInfo();
644 assert(MCAI
&& "No MCAsmInfo");
645 switch (MCAI
->getExceptionHandlingType()) {
646 case ExceptionHandling::SjLj
:
647 // SjLj piggy-backs on dwarf for this bit. The cleanups done apply to both
648 // Dwarf EH prepare needs to be run after SjLj prepare. Otherwise,
649 // catch info can get misplaced when a selector ends up more than one block
650 // removed from the parent invoke(s). This could happen when a landing
651 // pad is shared by multiple invokes and is also a target of a normal
652 // edge from elsewhere.
653 addPass(createSjLjEHPreparePass());
655 case ExceptionHandling::DwarfCFI
:
656 case ExceptionHandling::ARM
:
657 addPass(createDwarfEHPass());
659 case ExceptionHandling::WinEH
:
660 // We support using both GCC-style and MSVC-style exceptions on Windows, so
661 // add both preparation passes. Each pass will only actually run if it
662 // recognizes the personality function.
663 addPass(createWinEHPass());
664 addPass(createDwarfEHPass());
666 case ExceptionHandling::Wasm
:
667 // Wasm EH uses Windows EH instructions, but it does not need to demote PHIs
668 // on catchpads and cleanuppads because it does not outline them into
669 // funclets. Catchswitch blocks are not lowered in SelectionDAG, so we
670 // should remove PHIs there.
671 addPass(createWinEHPass(/*DemoteCatchSwitchPHIOnly=*/false));
672 addPass(createWasmEHPass());
674 case ExceptionHandling::None
:
675 addPass(createLowerInvokePass());
677 // The lower invoke pass may create unreachable code. Remove it.
678 addPass(createUnreachableBlockEliminationPass());
683 /// Add pass to prepare the LLVM IR for code generation. This should be done
684 /// before exception handling preparation passes.
685 void TargetPassConfig::addCodeGenPrepare() {
686 if (getOptLevel() != CodeGenOpt::None
&& !DisableCGP
)
687 addPass(createCodeGenPreparePass());
688 addPass(createRewriteSymbolsPass());
691 /// Add common passes that perform LLVM IR to IR transforms in preparation for
692 /// instruction selection.
693 void TargetPassConfig::addISelPrepare() {
696 // Force codegen to run according to the callgraph.
697 if (requiresCodeGenSCCOrder())
698 addPass(new DummyCGSCCPass
);
700 // Add both the safe stack and the stack protection passes: each of them will
701 // only protect functions that have corresponding attributes.
702 addPass(createSafeStackPass());
703 addPass(createStackProtectorPass());
706 addPass(createPrintFunctionPass(
707 dbgs(), "\n\n*** Final LLVM Code input to ISel ***\n"));
709 // All passes which modify the LLVM IR are now complete; run the verifier
710 // to ensure that the IR is valid.
712 addPass(createVerifierPass());
715 bool TargetPassConfig::addCoreISelPasses() {
716 // Enable FastISel with -fast-isel, but allow that to be overridden.
717 TM
->setO0WantsFastISel(EnableFastISelOption
!= cl::BOU_FALSE
);
718 if (EnableFastISelOption
== cl::BOU_TRUE
||
719 (TM
->getOptLevel() == CodeGenOpt::None
&& TM
->getO0WantsFastISel()))
720 TM
->setFastISel(true);
722 // Ask the target for an instruction selector.
723 // Explicitly enabling fast-isel should override implicitly enabled
725 if (EnableGlobalISelOption
== cl::BOU_TRUE
||
726 (EnableGlobalISelOption
== cl::BOU_UNSET
&&
727 TM
->Options
.EnableGlobalISel
&& EnableFastISelOption
!= cl::BOU_TRUE
)) {
728 TM
->setFastISel(false);
730 SaveAndRestore
<bool> SavedAddingMachinePasses(AddingMachinePasses
, true);
731 if (addIRTranslator())
734 addPreLegalizeMachineIR();
736 if (addLegalizeMachineIR())
739 // Before running the register bank selector, ask the target if it
740 // wants to run some passes.
741 addPreRegBankSelect();
743 if (addRegBankSelect())
746 addPreGlobalInstructionSelect();
748 if (addGlobalInstructionSelect())
751 // Pass to reset the MachineFunction if the ISel failed.
752 addPass(createResetMachineFunctionPass(
753 reportDiagnosticWhenGlobalISelFallback(), isGlobalISelAbortEnabled()));
755 // Provide a fallback path when we do not want to abort on
756 // not-yet-supported input.
757 if (!isGlobalISelAbortEnabled() && addInstSelector())
760 } else if (addInstSelector())
766 bool TargetPassConfig::addISelPasses() {
767 if (TM
->useEmulatedTLS())
768 addPass(createLowerEmuTLSPass());
770 addPass(createPreISelIntrinsicLoweringPass());
771 addPass(createTargetTransformInfoWrapperPass(TM
->getTargetIRAnalysis()));
774 addPassesToHandleExceptions();
777 return addCoreISelPasses();
780 /// -regalloc=... command line option.
781 static FunctionPass
*useDefaultRegisterAllocator() { return nullptr; }
782 static cl::opt
<RegisterRegAlloc::FunctionPassCtor
, false,
783 RegisterPassParser
<RegisterRegAlloc
>>
784 RegAlloc("regalloc", cl::Hidden
, cl::init(&useDefaultRegisterAllocator
),
785 cl::desc("Register allocator to use"));
787 /// Add the complete set of target-independent postISel code generator passes.
789 /// This can be read as the standard order of major LLVM CodeGen stages. Stages
790 /// with nontrivial configuration or multiple passes are broken out below in
791 /// add%Stage routines.
793 /// Any TargetPassConfig::addXX routine may be overriden by the Target. The
794 /// addPre/Post methods with empty header implementations allow injecting
795 /// target-specific fixups just before or after major stages. Additionally,
796 /// targets have the flexibility to change pass order within a stage by
797 /// overriding default implementation of add%Stage routines below. Each
798 /// technique has maintainability tradeoffs because alternate pass orders are
799 /// not well supported. addPre/Post works better if the target pass is easily
800 /// tied to a common pass. But if it has subtle dependencies on multiple passes,
801 /// the target should override the stage instead.
803 /// TODO: We could use a single addPre/Post(ID) hook to allow pass injection
804 /// before/after any target-independent pass. But it's currently overkill.
805 void TargetPassConfig::addMachinePasses() {
806 AddingMachinePasses
= true;
808 // Insert a machine instr printer pass after the specified pass.
809 if (!StringRef(PrintMachineInstrs
.getValue()).equals("") &&
810 !StringRef(PrintMachineInstrs
.getValue()).equals("option-unspecified")) {
811 const PassRegistry
*PR
= PassRegistry::getPassRegistry();
812 const PassInfo
*TPI
= PR
->getPassInfo(PrintMachineInstrs
.getValue());
813 const PassInfo
*IPI
= PR
->getPassInfo(StringRef("machineinstr-printer"));
814 assert (TPI
&& IPI
&& "Pass ID not registered!");
815 const char *TID
= (const char *)(TPI
->getTypeInfo());
816 const char *IID
= (const char *)(IPI
->getTypeInfo());
817 insertPass(TID
, IID
);
820 // Print the instruction selected machine code...
821 printAndVerify("After Instruction Selection");
823 // Expand pseudo-instructions emitted by ISel.
824 addPass(&ExpandISelPseudosID
);
826 // Add passes that optimize machine instructions in SSA form.
827 if (getOptLevel() != CodeGenOpt::None
) {
828 addMachineSSAOptimization();
830 // If the target requests it, assign local variables to stack slots relative
831 // to one another and simplify frame index references where possible.
832 addPass(&LocalStackSlotAllocationID
, false);
835 if (TM
->Options
.EnableIPRA
)
836 addPass(createRegUsageInfoPropPass());
838 // Run pre-ra passes.
841 // Run register allocation and passes that are tightly coupled with it,
842 // including phi elimination and scheduling.
843 if (getOptimizeRegAlloc())
844 addOptimizedRegAlloc(createRegAllocPass(true));
846 if (RegAlloc
!= &useDefaultRegisterAllocator
&&
847 RegAlloc
!= &createFastRegisterAllocator
)
848 report_fatal_error("Must use fast (default) register allocator for unoptimized regalloc.");
849 addFastRegAlloc(createRegAllocPass(false));
852 // Run post-ra passes.
855 // Insert prolog/epilog code. Eliminate abstract frame index references...
856 if (getOptLevel() != CodeGenOpt::None
) {
857 addPass(&PostRAMachineSinkingID
);
858 addPass(&ShrinkWrapID
);
861 // Prolog/Epilog inserter needs a TargetMachine to instantiate. But only
862 // do so if it hasn't been disabled, substituted, or overridden.
863 if (!isPassSubstitutedOrOverridden(&PrologEpilogCodeInserterID
))
864 addPass(createPrologEpilogInserterPass());
866 /// Add passes that optimize machine instructions after register allocation.
867 if (getOptLevel() != CodeGenOpt::None
)
868 addMachineLateOptimization();
870 // Expand pseudo instructions before second scheduling pass.
871 addPass(&ExpandPostRAPseudosID
);
873 // Run pre-sched2 passes.
876 if (EnableImplicitNullChecks
)
877 addPass(&ImplicitNullChecksID
);
879 // Second pass scheduler.
880 // Let Target optionally insert this pass by itself at some other
882 if (getOptLevel() != CodeGenOpt::None
&&
883 !TM
->targetSchedulesPostRAScheduling()) {
885 addPass(&PostMachineSchedulerID
);
887 addPass(&PostRASchedulerID
);
893 addPass(createGCInfoPrinter(dbgs()), false, false);
896 // Basic block placement.
897 if (getOptLevel() != CodeGenOpt::None
)
902 if (TM
->Options
.EnableIPRA
)
903 // Collect register usage information and produce a register mask of
904 // clobbered registers, to be used to optimize call sites.
905 addPass(createRegUsageInfoCollector());
907 addPass(&FuncletLayoutID
, false);
909 addPass(&StackMapLivenessID
, false);
910 addPass(&LiveDebugValuesID
, false);
912 // Insert before XRay Instrumentation.
913 addPass(&FEntryInserterID
, false);
915 addPass(&XRayInstrumentationID
, false);
916 addPass(&PatchableFunctionID
, false);
918 if (TM
->Options
.EnableMachineOutliner
&& getOptLevel() != CodeGenOpt::None
&&
919 EnableMachineOutliner
!= NeverOutline
) {
920 bool RunOnAllFunctions
= (EnableMachineOutliner
== AlwaysOutline
);
921 bool AddOutliner
= RunOnAllFunctions
||
922 TM
->Options
.SupportsDefaultOutlining
;
924 addPass(createMachineOutlinerPass(RunOnAllFunctions
));
927 // Add passes that directly emit MI after all other MI passes.
930 AddingMachinePasses
= false;
933 /// Add passes that optimize machine instructions in SSA form.
934 void TargetPassConfig::addMachineSSAOptimization() {
935 // Pre-ra tail duplication.
936 addPass(&EarlyTailDuplicateID
);
938 // Optimize PHIs before DCE: removing dead PHI cycles may make more
939 // instructions dead.
940 addPass(&OptimizePHIsID
, false);
942 // This pass merges large allocas. StackSlotColoring is a different pass
943 // which merges spill slots.
944 addPass(&StackColoringID
, false);
946 // If the target requests it, assign local variables to stack slots relative
947 // to one another and simplify frame index references where possible.
948 addPass(&LocalStackSlotAllocationID
, false);
950 // With optimization, dead code should already be eliminated. However
951 // there is one known exception: lowered code for arguments that are only
952 // used by tail calls, where the tail calls reuse the incoming stack
953 // arguments directly (see t11 in test/CodeGen/X86/sibcall.ll).
954 addPass(&DeadMachineInstructionElimID
);
956 // Allow targets to insert passes that improve instruction level parallelism,
957 // like if-conversion. Such passes will typically need dominator trees and
958 // loop info, just like LICM and CSE below.
961 addPass(&EarlyMachineLICMID
, false);
962 addPass(&MachineCSEID
, false);
964 addPass(&MachineSinkingID
);
966 addPass(&PeepholeOptimizerID
);
967 // Clean-up the dead code that may have been generated by peephole
969 addPass(&DeadMachineInstructionElimID
);
972 //===---------------------------------------------------------------------===//
973 /// Register Allocation Pass Configuration
974 //===---------------------------------------------------------------------===//
976 bool TargetPassConfig::getOptimizeRegAlloc() const {
977 switch (OptimizeRegAlloc
) {
978 case cl::BOU_UNSET
: return getOptLevel() != CodeGenOpt::None
;
979 case cl::BOU_TRUE
: return true;
980 case cl::BOU_FALSE
: return false;
982 llvm_unreachable("Invalid optimize-regalloc state");
985 /// RegisterRegAlloc's global Registry tracks allocator registration.
986 MachinePassRegistry
RegisterRegAlloc::Registry
;
988 /// A dummy default pass factory indicates whether the register allocator is
989 /// overridden on the command line.
990 static llvm::once_flag InitializeDefaultRegisterAllocatorFlag
;
992 static RegisterRegAlloc
993 defaultRegAlloc("default",
994 "pick register allocator based on -O option",
995 useDefaultRegisterAllocator
);
997 static void initializeDefaultRegisterAllocatorOnce() {
998 RegisterRegAlloc::FunctionPassCtor Ctor
= RegisterRegAlloc::getDefault();
1002 RegisterRegAlloc::setDefault(RegAlloc
);
1006 /// Instantiate the default register allocator pass for this target for either
1007 /// the optimized or unoptimized allocation path. This will be added to the pass
1008 /// manager by addFastRegAlloc in the unoptimized case or addOptimizedRegAlloc
1009 /// in the optimized case.
1011 /// A target that uses the standard regalloc pass order for fast or optimized
1012 /// allocation may still override this for per-target regalloc
1013 /// selection. But -regalloc=... always takes precedence.
1014 FunctionPass
*TargetPassConfig::createTargetRegisterAllocator(bool Optimized
) {
1016 return createGreedyRegisterAllocator();
1018 return createFastRegisterAllocator();
1021 /// Find and instantiate the register allocation pass requested by this target
1022 /// at the current optimization level. Different register allocators are
1023 /// defined as separate passes because they may require different analysis.
1025 /// This helper ensures that the regalloc= option is always available,
1026 /// even for targets that override the default allocator.
1028 /// FIXME: When MachinePassRegistry register pass IDs instead of function ptrs,
1029 /// this can be folded into addPass.
1030 FunctionPass
*TargetPassConfig::createRegAllocPass(bool Optimized
) {
1031 // Initialize the global default.
1032 llvm::call_once(InitializeDefaultRegisterAllocatorFlag
,
1033 initializeDefaultRegisterAllocatorOnce
);
1035 RegisterRegAlloc::FunctionPassCtor Ctor
= RegisterRegAlloc::getDefault();
1036 if (Ctor
!= useDefaultRegisterAllocator
)
1039 // With no -regalloc= override, ask the target for a regalloc pass.
1040 return createTargetRegisterAllocator(Optimized
);
1043 /// Return true if the default global register allocator is in use and
1044 /// has not be overriden on the command line with '-regalloc=...'
1045 bool TargetPassConfig::usingDefaultRegAlloc() const {
1046 return RegAlloc
.getNumOccurrences() == 0;
1049 /// Add the minimum set of target-independent passes that are required for
1050 /// register allocation. No coalescing or scheduling.
1051 void TargetPassConfig::addFastRegAlloc(FunctionPass
*RegAllocPass
) {
1052 addPass(&PHIEliminationID
, false);
1053 addPass(&TwoAddressInstructionPassID
, false);
1056 addPass(RegAllocPass
);
1059 /// Add standard target-independent passes that are tightly coupled with
1060 /// optimized register allocation, including coalescing, machine instruction
1061 /// scheduling, and register allocation itself.
1062 void TargetPassConfig::addOptimizedRegAlloc(FunctionPass
*RegAllocPass
) {
1063 addPass(&DetectDeadLanesID
, false);
1065 addPass(&ProcessImplicitDefsID
, false);
1067 // LiveVariables currently requires pure SSA form.
1069 // FIXME: Once TwoAddressInstruction pass no longer uses kill flags,
1070 // LiveVariables can be removed completely, and LiveIntervals can be directly
1071 // computed. (We still either need to regenerate kill flags after regalloc, or
1072 // preferably fix the scavenger to not depend on them).
1073 addPass(&LiveVariablesID
, false);
1075 // Edge splitting is smarter with machine loop info.
1076 addPass(&MachineLoopInfoID
, false);
1077 addPass(&PHIEliminationID
, false);
1079 // Eventually, we want to run LiveIntervals before PHI elimination.
1080 if (EarlyLiveIntervals
)
1081 addPass(&LiveIntervalsID
, false);
1083 addPass(&TwoAddressInstructionPassID
, false);
1084 addPass(&RegisterCoalescerID
);
1086 // The machine scheduler may accidentally create disconnected components
1087 // when moving subregister definitions around, avoid this by splitting them to
1088 // separate vregs before. Splitting can also improve reg. allocation quality.
1089 addPass(&RenameIndependentSubregsID
);
1091 // PreRA instruction scheduling.
1092 addPass(&MachineSchedulerID
);
1095 // Add the selected register allocation pass.
1096 addPass(RegAllocPass
);
1098 // Allow targets to change the register assignments before rewriting.
1101 // Finally rewrite virtual registers.
1102 addPass(&VirtRegRewriterID
);
1104 // Perform stack slot coloring and post-ra machine LICM.
1106 // FIXME: Re-enable coloring with register when it's capable of adding
1108 addPass(&StackSlotColoringID
);
1110 // Copy propagate to forward register uses and try to eliminate COPYs that
1111 // were not coalesced.
1112 addPass(&MachineCopyPropagationID
);
1114 // Run post-ra machine LICM to hoist reloads / remats.
1116 // FIXME: can this move into MachineLateOptimization?
1117 addPass(&MachineLICMID
);
1121 //===---------------------------------------------------------------------===//
1122 /// Post RegAlloc Pass Configuration
1123 //===---------------------------------------------------------------------===//
1125 /// Add passes that optimize machine instructions after register allocation.
1126 void TargetPassConfig::addMachineLateOptimization() {
1127 // Branch folding must be run after regalloc and prolog/epilog insertion.
1128 addPass(&BranchFolderPassID
);
1130 // Tail duplication.
1131 // Note that duplicating tail just increases code size and degrades
1132 // performance for targets that require Structured Control Flow.
1133 // In addition it can also make CFG irreducible. Thus we disable it.
1134 if (!TM
->requiresStructuredCFG())
1135 addPass(&TailDuplicateID
);
1137 // Copy propagation.
1138 addPass(&MachineCopyPropagationID
);
1141 /// Add standard GC passes.
1142 bool TargetPassConfig::addGCPasses() {
1143 addPass(&GCMachineCodeAnalysisID
, false);
1147 /// Add standard basic block placement passes.
1148 void TargetPassConfig::addBlockPlacement() {
1149 if (addPass(&MachineBlockPlacementID
)) {
1150 // Run a separate pass to collect block placement statistics.
1151 if (EnableBlockPlacementStats
)
1152 addPass(&MachineBlockPlacementStatsID
);
1156 //===---------------------------------------------------------------------===//
1157 /// GlobalISel Configuration
1158 //===---------------------------------------------------------------------===//
1159 bool TargetPassConfig::isGlobalISelAbortEnabled() const {
1160 if (EnableGlobalISelAbort
.getNumOccurrences() > 0)
1161 return EnableGlobalISelAbort
== 1;
1163 // When no abort behaviour is specified, we don't abort if the target says
1164 // that GISel is enabled.
1165 return !TM
->Options
.EnableGlobalISel
;
1168 bool TargetPassConfig::reportDiagnosticWhenGlobalISelFallback() const {
1169 return EnableGlobalISelAbort
== 2;