[ORC] Add std::tuple support to SimplePackedSerialization.
[llvm-project.git] / llvm / lib / CodeGen / TargetPassConfig.cpp
blob2a90c3154bb41e6cbe5111cf95b3b17cd2b393aa
1 //===- TargetPassConfig.cpp - Target independent code generation passes ---===//
2 //
3 // Part of the LLVM Project, under the Apache License v2.0 with LLVM Exceptions.
4 // See https://llvm.org/LICENSE.txt for license information.
5 // SPDX-License-Identifier: Apache-2.0 WITH LLVM-exception
6 //
7 //===----------------------------------------------------------------------===//
8 //
9 // This file defines interfaces to access the target independent code
10 // generation passes provided by the LLVM backend.
12 //===---------------------------------------------------------------------===//
14 #include "llvm/CodeGen/TargetPassConfig.h"
15 #include "llvm/ADT/DenseMap.h"
16 #include "llvm/ADT/SmallVector.h"
17 #include "llvm/ADT/StringRef.h"
18 #include "llvm/Analysis/BasicAliasAnalysis.h"
19 #include "llvm/Analysis/CFLAndersAliasAnalysis.h"
20 #include "llvm/Analysis/CFLSteensAliasAnalysis.h"
21 #include "llvm/Analysis/CallGraphSCCPass.h"
22 #include "llvm/Analysis/ScopedNoAliasAA.h"
23 #include "llvm/Analysis/TargetTransformInfo.h"
24 #include "llvm/Analysis/TypeBasedAliasAnalysis.h"
25 #include "llvm/CodeGen/CSEConfigBase.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/PassInstrumentation.h"
33 #include "llvm/IR/Verifier.h"
34 #include "llvm/InitializePasses.h"
35 #include "llvm/MC/MCAsmInfo.h"
36 #include "llvm/MC/MCTargetOptions.h"
37 #include "llvm/Pass.h"
38 #include "llvm/Support/CodeGen.h"
39 #include "llvm/Support/CommandLine.h"
40 #include "llvm/Support/Compiler.h"
41 #include "llvm/Support/Debug.h"
42 #include "llvm/Support/Discriminator.h"
43 #include "llvm/Support/ErrorHandling.h"
44 #include "llvm/Support/SaveAndRestore.h"
45 #include "llvm/Support/Threading.h"
46 #include "llvm/Target/CGPassBuilderOption.h"
47 #include "llvm/Target/TargetMachine.h"
48 #include "llvm/Transforms/Scalar.h"
49 #include "llvm/Transforms/Utils.h"
50 #include "llvm/Transforms/Utils/SymbolRewriter.h"
51 #include <cassert>
52 #include <string>
54 using namespace llvm;
56 static cl::opt<bool>
57 EnableIPRA("enable-ipra", cl::init(false), cl::Hidden,
58 cl::desc("Enable interprocedural register allocation "
59 "to reduce load/store at procedure calls."));
60 static cl::opt<bool> DisablePostRASched("disable-post-ra", cl::Hidden,
61 cl::desc("Disable Post Regalloc Scheduler"));
62 static cl::opt<bool> DisableBranchFold("disable-branch-fold", cl::Hidden,
63 cl::desc("Disable branch folding"));
64 static cl::opt<bool> DisableTailDuplicate("disable-tail-duplicate", cl::Hidden,
65 cl::desc("Disable tail duplication"));
66 static cl::opt<bool> DisableEarlyTailDup("disable-early-taildup", cl::Hidden,
67 cl::desc("Disable pre-register allocation tail duplication"));
68 static cl::opt<bool> DisableBlockPlacement("disable-block-placement",
69 cl::Hidden, cl::desc("Disable probability-driven block placement"));
70 static cl::opt<bool> EnableBlockPlacementStats("enable-block-placement-stats",
71 cl::Hidden, cl::desc("Collect probability-driven block placement stats"));
72 static cl::opt<bool> DisableSSC("disable-ssc", cl::Hidden,
73 cl::desc("Disable Stack Slot Coloring"));
74 static cl::opt<bool> DisableMachineDCE("disable-machine-dce", cl::Hidden,
75 cl::desc("Disable Machine Dead Code Elimination"));
76 static cl::opt<bool> DisableEarlyIfConversion("disable-early-ifcvt", cl::Hidden,
77 cl::desc("Disable Early If-conversion"));
78 static cl::opt<bool> DisableMachineLICM("disable-machine-licm", cl::Hidden,
79 cl::desc("Disable Machine LICM"));
80 static cl::opt<bool> DisableMachineCSE("disable-machine-cse", cl::Hidden,
81 cl::desc("Disable Machine Common Subexpression Elimination"));
82 static cl::opt<cl::boolOrDefault> OptimizeRegAlloc(
83 "optimize-regalloc", cl::Hidden,
84 cl::desc("Enable optimized register allocation compilation path."));
85 static cl::opt<bool> DisablePostRAMachineLICM("disable-postra-machine-licm",
86 cl::Hidden,
87 cl::desc("Disable Machine LICM"));
88 static cl::opt<bool> DisableMachineSink("disable-machine-sink", cl::Hidden,
89 cl::desc("Disable Machine Sinking"));
90 static cl::opt<bool> DisablePostRAMachineSink("disable-postra-machine-sink",
91 cl::Hidden,
92 cl::desc("Disable PostRA Machine Sinking"));
93 static cl::opt<bool> DisableLSR("disable-lsr", cl::Hidden,
94 cl::desc("Disable Loop Strength Reduction Pass"));
95 static cl::opt<bool> DisableConstantHoisting("disable-constant-hoisting",
96 cl::Hidden, cl::desc("Disable ConstantHoisting"));
97 static cl::opt<bool> DisableCGP("disable-cgp", cl::Hidden,
98 cl::desc("Disable Codegen Prepare"));
99 static cl::opt<bool> DisableCopyProp("disable-copyprop", cl::Hidden,
100 cl::desc("Disable Copy Propagation pass"));
101 static cl::opt<bool> DisablePartialLibcallInlining("disable-partial-libcall-inlining",
102 cl::Hidden, cl::desc("Disable Partial Libcall Inlining"));
103 static cl::opt<bool> EnableImplicitNullChecks(
104 "enable-implicit-null-checks",
105 cl::desc("Fold null checks into faulting memory operations"),
106 cl::init(false), cl::Hidden);
107 static cl::opt<bool> DisableMergeICmps("disable-mergeicmps",
108 cl::desc("Disable MergeICmps Pass"),
109 cl::init(false), cl::Hidden);
110 static cl::opt<bool> PrintLSR("print-lsr-output", cl::Hidden,
111 cl::desc("Print LLVM IR produced by the loop-reduce pass"));
112 static cl::opt<bool> PrintISelInput("print-isel-input", cl::Hidden,
113 cl::desc("Print LLVM IR input to isel pass"));
114 static cl::opt<bool> PrintGCInfo("print-gc", cl::Hidden,
115 cl::desc("Dump garbage collector data"));
116 static cl::opt<cl::boolOrDefault>
117 VerifyMachineCode("verify-machineinstrs", cl::Hidden,
118 cl::desc("Verify generated machine code"),
119 cl::ZeroOrMore);
120 static cl::opt<cl::boolOrDefault> DebugifyAndStripAll(
121 "debugify-and-strip-all-safe", cl::Hidden,
122 cl::desc(
123 "Debugify MIR before and Strip debug after "
124 "each pass except those known to be unsafe when debug info is present"),
125 cl::ZeroOrMore);
126 static cl::opt<cl::boolOrDefault> DebugifyCheckAndStripAll(
127 "debugify-check-and-strip-all-safe", cl::Hidden,
128 cl::desc(
129 "Debugify MIR before, by checking and stripping the debug info after, "
130 "each pass except those known to be unsafe when debug info is present"),
131 cl::ZeroOrMore);
132 // Enable or disable the MachineOutliner.
133 static cl::opt<RunOutliner> EnableMachineOutliner(
134 "enable-machine-outliner", cl::desc("Enable the machine outliner"),
135 cl::Hidden, cl::ValueOptional, cl::init(RunOutliner::TargetDefault),
136 cl::values(clEnumValN(RunOutliner::AlwaysOutline, "always",
137 "Run on all functions guaranteed to be beneficial"),
138 clEnumValN(RunOutliner::NeverOutline, "never",
139 "Disable all outlining"),
140 // Sentinel value for unspecified option.
141 clEnumValN(RunOutliner::AlwaysOutline, "", "")));
142 // Enable or disable FastISel. Both options are needed, because
143 // FastISel is enabled by default with -fast, and we wish to be
144 // able to enable or disable fast-isel independently from -O0.
145 static cl::opt<cl::boolOrDefault>
146 EnableFastISelOption("fast-isel", cl::Hidden,
147 cl::desc("Enable the \"fast\" instruction selector"));
149 static cl::opt<cl::boolOrDefault> EnableGlobalISelOption(
150 "global-isel", cl::Hidden,
151 cl::desc("Enable the \"global\" instruction selector"));
153 // FIXME: remove this after switching to NPM or GlobalISel, whichever gets there
154 // first...
155 static cl::opt<bool>
156 PrintAfterISel("print-after-isel", cl::init(false), cl::Hidden,
157 cl::desc("Print machine instrs after ISel"));
159 static cl::opt<GlobalISelAbortMode> EnableGlobalISelAbort(
160 "global-isel-abort", cl::Hidden,
161 cl::desc("Enable abort calls when \"global\" instruction selection "
162 "fails to lower/select an instruction"),
163 cl::values(
164 clEnumValN(GlobalISelAbortMode::Disable, "0", "Disable the abort"),
165 clEnumValN(GlobalISelAbortMode::Enable, "1", "Enable the abort"),
166 clEnumValN(GlobalISelAbortMode::DisableWithDiag, "2",
167 "Disable the abort but emit a diagnostic on failure")));
169 // An option that disables inserting FS-AFDO discriminators before emit.
170 // This is mainly for debugging and tuning purpose.
171 static cl::opt<bool>
172 FSNoFinalDiscrim("fs-no-final-discrim", cl::init(false), cl::Hidden,
173 cl::desc("Do not insert FS-AFDO discriminators before "
174 "emit."));
175 // Disable MIRProfileLoader before RegAlloc. This is for for debugging and
176 // tuning purpose.
177 static cl::opt<bool> DisableRAFSProfileLoader(
178 "disable-ra-fsprofile-loader", cl::init(true), cl::Hidden,
179 cl::desc("Disable MIRProfileLoader before RegAlloc"));
180 // Disable MIRProfileLoader before BloackPlacement. This is for for debugging
181 // and tuning purpose.
182 static cl::opt<bool> DisableLayoutFSProfileLoader(
183 "disable-layout-fsprofile-loader", cl::init(true), cl::Hidden,
184 cl::desc("Disable MIRProfileLoader before BlockPlacement"));
185 // Specify FSProfile file name.
186 static cl::opt<std::string>
187 FSProfileFile("fs-profile-file", cl::init(""), cl::value_desc("filename"),
188 cl::desc("Flow Sensitive profile file name."), cl::Hidden);
189 // Specify Remapping file for FSProfile.
190 static cl::opt<std::string> FSRemappingFile(
191 "fs-remapping-file", cl::init(""), cl::value_desc("filename"),
192 cl::desc("Flow Sensitive profile remapping file name."), cl::Hidden);
194 // Temporary option to allow experimenting with MachineScheduler as a post-RA
195 // scheduler. Targets can "properly" enable this with
196 // substitutePass(&PostRASchedulerID, &PostMachineSchedulerID).
197 // Targets can return true in targetSchedulesPostRAScheduling() and
198 // insert a PostRA scheduling pass wherever it wants.
199 static cl::opt<bool> MISchedPostRA(
200 "misched-postra", cl::Hidden,
201 cl::desc(
202 "Run MachineScheduler post regalloc (independent of preRA sched)"));
204 // Experimental option to run live interval analysis early.
205 static cl::opt<bool> EarlyLiveIntervals("early-live-intervals", cl::Hidden,
206 cl::desc("Run live interval analysis earlier in the pipeline"));
208 // Experimental option to use CFL-AA in codegen
209 static cl::opt<CFLAAType> UseCFLAA(
210 "use-cfl-aa-in-codegen", cl::init(CFLAAType::None), cl::Hidden,
211 cl::desc("Enable the new, experimental CFL alias analysis in CodeGen"),
212 cl::values(clEnumValN(CFLAAType::None, "none", "Disable CFL-AA"),
213 clEnumValN(CFLAAType::Steensgaard, "steens",
214 "Enable unification-based CFL-AA"),
215 clEnumValN(CFLAAType::Andersen, "anders",
216 "Enable inclusion-based CFL-AA"),
217 clEnumValN(CFLAAType::Both, "both",
218 "Enable both variants of CFL-AA")));
220 /// Option names for limiting the codegen pipeline.
221 /// Those are used in error reporting and we didn't want
222 /// to duplicate their names all over the place.
223 static const char StartAfterOptName[] = "start-after";
224 static const char StartBeforeOptName[] = "start-before";
225 static const char StopAfterOptName[] = "stop-after";
226 static const char StopBeforeOptName[] = "stop-before";
228 static cl::opt<std::string>
229 StartAfterOpt(StringRef(StartAfterOptName),
230 cl::desc("Resume compilation after a specific pass"),
231 cl::value_desc("pass-name"), cl::init(""), cl::Hidden);
233 static cl::opt<std::string>
234 StartBeforeOpt(StringRef(StartBeforeOptName),
235 cl::desc("Resume compilation before a specific pass"),
236 cl::value_desc("pass-name"), cl::init(""), cl::Hidden);
238 static cl::opt<std::string>
239 StopAfterOpt(StringRef(StopAfterOptName),
240 cl::desc("Stop compilation after a specific pass"),
241 cl::value_desc("pass-name"), cl::init(""), cl::Hidden);
243 static cl::opt<std::string>
244 StopBeforeOpt(StringRef(StopBeforeOptName),
245 cl::desc("Stop compilation before a specific pass"),
246 cl::value_desc("pass-name"), cl::init(""), cl::Hidden);
248 /// Enable the machine function splitter pass.
249 static cl::opt<bool> EnableMachineFunctionSplitter(
250 "enable-split-machine-functions", cl::Hidden,
251 cl::desc("Split out cold blocks from machine functions based on profile "
252 "information."));
254 /// Disable the expand reductions pass for testing.
255 static cl::opt<bool> DisableExpandReductions(
256 "disable-expand-reductions", cl::init(false), cl::Hidden,
257 cl::desc("Disable the expand reduction intrinsics pass from running"));
259 /// Allow standard passes to be disabled by command line options. This supports
260 /// simple binary flags that either suppress the pass or do nothing.
261 /// i.e. -disable-mypass=false has no effect.
262 /// These should be converted to boolOrDefault in order to use applyOverride.
263 static IdentifyingPassPtr applyDisable(IdentifyingPassPtr PassID,
264 bool Override) {
265 if (Override)
266 return IdentifyingPassPtr();
267 return PassID;
270 /// Allow standard passes to be disabled by the command line, regardless of who
271 /// is adding the pass.
273 /// StandardID is the pass identified in the standard pass pipeline and provided
274 /// to addPass(). It may be a target-specific ID in the case that the target
275 /// directly adds its own pass, but in that case we harmlessly fall through.
277 /// TargetID is the pass that the target has configured to override StandardID.
279 /// StandardID may be a pseudo ID. In that case TargetID is the name of the real
280 /// pass to run. This allows multiple options to control a single pass depending
281 /// on where in the pipeline that pass is added.
282 static IdentifyingPassPtr overridePass(AnalysisID StandardID,
283 IdentifyingPassPtr TargetID) {
284 if (StandardID == &PostRASchedulerID)
285 return applyDisable(TargetID, DisablePostRASched);
287 if (StandardID == &BranchFolderPassID)
288 return applyDisable(TargetID, DisableBranchFold);
290 if (StandardID == &TailDuplicateID)
291 return applyDisable(TargetID, DisableTailDuplicate);
293 if (StandardID == &EarlyTailDuplicateID)
294 return applyDisable(TargetID, DisableEarlyTailDup);
296 if (StandardID == &MachineBlockPlacementID)
297 return applyDisable(TargetID, DisableBlockPlacement);
299 if (StandardID == &StackSlotColoringID)
300 return applyDisable(TargetID, DisableSSC);
302 if (StandardID == &DeadMachineInstructionElimID)
303 return applyDisable(TargetID, DisableMachineDCE);
305 if (StandardID == &EarlyIfConverterID)
306 return applyDisable(TargetID, DisableEarlyIfConversion);
308 if (StandardID == &EarlyMachineLICMID)
309 return applyDisable(TargetID, DisableMachineLICM);
311 if (StandardID == &MachineCSEID)
312 return applyDisable(TargetID, DisableMachineCSE);
314 if (StandardID == &MachineLICMID)
315 return applyDisable(TargetID, DisablePostRAMachineLICM);
317 if (StandardID == &MachineSinkingID)
318 return applyDisable(TargetID, DisableMachineSink);
320 if (StandardID == &PostRAMachineSinkingID)
321 return applyDisable(TargetID, DisablePostRAMachineSink);
323 if (StandardID == &MachineCopyPropagationID)
324 return applyDisable(TargetID, DisableCopyProp);
326 return TargetID;
329 // Find the FSProfile file name. The internal option takes the precedence
330 // before getting from TargetMachine.
331 static const std::string getFSProfileFile(const TargetMachine *TM) {
332 if (!FSProfileFile.empty())
333 return FSProfileFile.getValue();
334 const Optional<PGOOptions> &PGOOpt = TM->getPGOOption();
335 if (PGOOpt == None || PGOOpt->Action != PGOOptions::SampleUse)
336 return std::string();
337 return PGOOpt->ProfileFile;
340 // Find the Profile remapping file name. The internal option takes the
341 // precedence before getting from TargetMachine.
342 static const std::string getFSRemappingFile(const TargetMachine *TM) {
343 if (!FSRemappingFile.empty())
344 return FSRemappingFile.getValue();
345 const Optional<PGOOptions> &PGOOpt = TM->getPGOOption();
346 if (PGOOpt == None || PGOOpt->Action != PGOOptions::SampleUse)
347 return std::string();
348 return PGOOpt->ProfileRemappingFile;
351 //===---------------------------------------------------------------------===//
352 /// TargetPassConfig
353 //===---------------------------------------------------------------------===//
355 INITIALIZE_PASS(TargetPassConfig, "targetpassconfig",
356 "Target Pass Configuration", false, false)
357 char TargetPassConfig::ID = 0;
359 namespace {
361 struct InsertedPass {
362 AnalysisID TargetPassID;
363 IdentifyingPassPtr InsertedPassID;
364 bool VerifyAfter;
366 InsertedPass(AnalysisID TargetPassID, IdentifyingPassPtr InsertedPassID,
367 bool VerifyAfter)
368 : TargetPassID(TargetPassID), InsertedPassID(InsertedPassID),
369 VerifyAfter(VerifyAfter) {}
371 Pass *getInsertedPass() const {
372 assert(InsertedPassID.isValid() && "Illegal Pass ID!");
373 if (InsertedPassID.isInstance())
374 return InsertedPassID.getInstance();
375 Pass *NP = Pass::createPass(InsertedPassID.getID());
376 assert(NP && "Pass ID not registered");
377 return NP;
381 } // end anonymous namespace
383 namespace llvm {
385 extern cl::opt<bool> EnableFSDiscriminator;
387 class PassConfigImpl {
388 public:
389 // List of passes explicitly substituted by this target. Normally this is
390 // empty, but it is a convenient way to suppress or replace specific passes
391 // that are part of a standard pass pipeline without overridding the entire
392 // pipeline. This mechanism allows target options to inherit a standard pass's
393 // user interface. For example, a target may disable a standard pass by
394 // default by substituting a pass ID of zero, and the user may still enable
395 // that standard pass with an explicit command line option.
396 DenseMap<AnalysisID,IdentifyingPassPtr> TargetPasses;
398 /// Store the pairs of <AnalysisID, AnalysisID> of which the second pass
399 /// is inserted after each instance of the first one.
400 SmallVector<InsertedPass, 4> InsertedPasses;
403 } // end namespace llvm
405 // Out of line virtual method.
406 TargetPassConfig::~TargetPassConfig() {
407 delete Impl;
410 static const PassInfo *getPassInfo(StringRef PassName) {
411 if (PassName.empty())
412 return nullptr;
414 const PassRegistry &PR = *PassRegistry::getPassRegistry();
415 const PassInfo *PI = PR.getPassInfo(PassName);
416 if (!PI)
417 report_fatal_error(Twine('\"') + Twine(PassName) +
418 Twine("\" pass is not registered."));
419 return PI;
422 static AnalysisID getPassIDFromName(StringRef PassName) {
423 const PassInfo *PI = getPassInfo(PassName);
424 return PI ? PI->getTypeInfo() : nullptr;
427 static std::pair<StringRef, unsigned>
428 getPassNameAndInstanceNum(StringRef PassName) {
429 StringRef Name, InstanceNumStr;
430 std::tie(Name, InstanceNumStr) = PassName.split(',');
432 unsigned InstanceNum = 0;
433 if (!InstanceNumStr.empty() && InstanceNumStr.getAsInteger(10, InstanceNum))
434 report_fatal_error("invalid pass instance specifier " + PassName);
436 return std::make_pair(Name, InstanceNum);
439 void TargetPassConfig::setStartStopPasses() {
440 StringRef StartBeforeName;
441 std::tie(StartBeforeName, StartBeforeInstanceNum) =
442 getPassNameAndInstanceNum(StartBeforeOpt);
444 StringRef StartAfterName;
445 std::tie(StartAfterName, StartAfterInstanceNum) =
446 getPassNameAndInstanceNum(StartAfterOpt);
448 StringRef StopBeforeName;
449 std::tie(StopBeforeName, StopBeforeInstanceNum)
450 = getPassNameAndInstanceNum(StopBeforeOpt);
452 StringRef StopAfterName;
453 std::tie(StopAfterName, StopAfterInstanceNum)
454 = getPassNameAndInstanceNum(StopAfterOpt);
456 StartBefore = getPassIDFromName(StartBeforeName);
457 StartAfter = getPassIDFromName(StartAfterName);
458 StopBefore = getPassIDFromName(StopBeforeName);
459 StopAfter = getPassIDFromName(StopAfterName);
460 if (StartBefore && StartAfter)
461 report_fatal_error(Twine(StartBeforeOptName) + Twine(" and ") +
462 Twine(StartAfterOptName) + Twine(" specified!"));
463 if (StopBefore && StopAfter)
464 report_fatal_error(Twine(StopBeforeOptName) + Twine(" and ") +
465 Twine(StopAfterOptName) + Twine(" specified!"));
466 Started = (StartAfter == nullptr) && (StartBefore == nullptr);
469 CGPassBuilderOption llvm::getCGPassBuilderOption() {
470 CGPassBuilderOption Opt;
472 #define SET_OPTION(Option) \
473 if (Option.getNumOccurrences()) \
474 Opt.Option = Option;
476 SET_OPTION(EnableFastISelOption)
477 SET_OPTION(EnableGlobalISelAbort)
478 SET_OPTION(EnableGlobalISelOption)
479 SET_OPTION(EnableIPRA)
480 SET_OPTION(OptimizeRegAlloc)
481 SET_OPTION(VerifyMachineCode)
483 #define SET_BOOLEAN_OPTION(Option) Opt.Option = Option;
485 SET_BOOLEAN_OPTION(EarlyLiveIntervals)
486 SET_BOOLEAN_OPTION(EnableBlockPlacementStats)
487 SET_BOOLEAN_OPTION(EnableImplicitNullChecks)
488 SET_BOOLEAN_OPTION(EnableMachineOutliner)
489 SET_BOOLEAN_OPTION(MISchedPostRA)
490 SET_BOOLEAN_OPTION(UseCFLAA)
491 SET_BOOLEAN_OPTION(DisableMergeICmps)
492 SET_BOOLEAN_OPTION(DisableLSR)
493 SET_BOOLEAN_OPTION(DisableConstantHoisting)
494 SET_BOOLEAN_OPTION(DisableCGP)
495 SET_BOOLEAN_OPTION(DisablePartialLibcallInlining)
496 SET_BOOLEAN_OPTION(PrintLSR)
497 SET_BOOLEAN_OPTION(PrintISelInput)
498 SET_BOOLEAN_OPTION(PrintGCInfo)
500 return Opt;
503 static void registerPartialPipelineCallback(PassInstrumentationCallbacks &PIC,
504 LLVMTargetMachine &LLVMTM) {
505 StringRef StartBefore;
506 StringRef StartAfter;
507 StringRef StopBefore;
508 StringRef StopAfter;
510 unsigned StartBeforeInstanceNum = 0;
511 unsigned StartAfterInstanceNum = 0;
512 unsigned StopBeforeInstanceNum = 0;
513 unsigned StopAfterInstanceNum = 0;
515 std::tie(StartBefore, StartBeforeInstanceNum) =
516 getPassNameAndInstanceNum(StartBeforeOpt);
517 std::tie(StartAfter, StartAfterInstanceNum) =
518 getPassNameAndInstanceNum(StartAfterOpt);
519 std::tie(StopBefore, StopBeforeInstanceNum) =
520 getPassNameAndInstanceNum(StopBeforeOpt);
521 std::tie(StopAfter, StopAfterInstanceNum) =
522 getPassNameAndInstanceNum(StopAfterOpt);
524 if (StartBefore.empty() && StartAfter.empty() && StopBefore.empty() &&
525 StopAfter.empty())
526 return;
528 std::tie(StartBefore, std::ignore) =
529 LLVMTM.getPassNameFromLegacyName(StartBefore);
530 std::tie(StartAfter, std::ignore) =
531 LLVMTM.getPassNameFromLegacyName(StartAfter);
532 std::tie(StopBefore, std::ignore) =
533 LLVMTM.getPassNameFromLegacyName(StopBefore);
534 std::tie(StopAfter, std::ignore) =
535 LLVMTM.getPassNameFromLegacyName(StopAfter);
536 if (!StartBefore.empty() && !StartAfter.empty())
537 report_fatal_error(Twine(StartBeforeOptName) + Twine(" and ") +
538 Twine(StartAfterOptName) + Twine(" specified!"));
539 if (!StopBefore.empty() && !StopAfter.empty())
540 report_fatal_error(Twine(StopBeforeOptName) + Twine(" and ") +
541 Twine(StopAfterOptName) + Twine(" specified!"));
543 PIC.registerShouldRunOptionalPassCallback(
544 [=, EnableCurrent = StartBefore.empty() && StartAfter.empty(),
545 EnableNext = Optional<bool>(), StartBeforeCount = 0u,
546 StartAfterCount = 0u, StopBeforeCount = 0u,
547 StopAfterCount = 0u](StringRef P, Any) mutable {
548 bool StartBeforePass = !StartBefore.empty() && P.contains(StartBefore);
549 bool StartAfterPass = !StartAfter.empty() && P.contains(StartAfter);
550 bool StopBeforePass = !StopBefore.empty() && P.contains(StopBefore);
551 bool StopAfterPass = !StopAfter.empty() && P.contains(StopAfter);
553 // Implement -start-after/-stop-after
554 if (EnableNext) {
555 EnableCurrent = *EnableNext;
556 EnableNext.reset();
559 // Using PIC.registerAfterPassCallback won't work because if this
560 // callback returns false, AfterPassCallback is also skipped.
561 if (StartAfterPass && StartAfterCount++ == StartAfterInstanceNum) {
562 assert(!EnableNext && "Error: assign to EnableNext more than once");
563 EnableNext = true;
565 if (StopAfterPass && StopAfterCount++ == StopAfterInstanceNum) {
566 assert(!EnableNext && "Error: assign to EnableNext more than once");
567 EnableNext = false;
570 if (StartBeforePass && StartBeforeCount++ == StartBeforeInstanceNum)
571 EnableCurrent = true;
572 if (StopBeforePass && StopBeforeCount++ == StopBeforeInstanceNum)
573 EnableCurrent = false;
574 return EnableCurrent;
578 void llvm::registerCodeGenCallback(PassInstrumentationCallbacks &PIC,
579 LLVMTargetMachine &LLVMTM) {
581 // Register a callback for disabling passes.
582 PIC.registerShouldRunOptionalPassCallback([](StringRef P, Any) {
584 #define DISABLE_PASS(Option, Name) \
585 if (Option && P.contains(#Name)) \
586 return false;
587 DISABLE_PASS(DisableBlockPlacement, MachineBlockPlacementPass)
588 DISABLE_PASS(DisableBranchFold, BranchFolderPass)
589 DISABLE_PASS(DisableCopyProp, MachineCopyPropagationPass)
590 DISABLE_PASS(DisableEarlyIfConversion, EarlyIfConverterPass)
591 DISABLE_PASS(DisableEarlyTailDup, EarlyTailDuplicatePass)
592 DISABLE_PASS(DisableMachineCSE, MachineCSEPass)
593 DISABLE_PASS(DisableMachineDCE, DeadMachineInstructionElimPass)
594 DISABLE_PASS(DisableMachineLICM, EarlyMachineLICMPass)
595 DISABLE_PASS(DisableMachineSink, MachineSinkingPass)
596 DISABLE_PASS(DisablePostRAMachineLICM, MachineLICMPass)
597 DISABLE_PASS(DisablePostRAMachineSink, PostRAMachineSinkingPass)
598 DISABLE_PASS(DisablePostRASched, PostRASchedulerPass)
599 DISABLE_PASS(DisableSSC, StackSlotColoringPass)
600 DISABLE_PASS(DisableTailDuplicate, TailDuplicatePass)
602 return true;
605 registerPartialPipelineCallback(PIC, LLVMTM);
608 // Out of line constructor provides default values for pass options and
609 // registers all common codegen passes.
610 TargetPassConfig::TargetPassConfig(LLVMTargetMachine &TM, PassManagerBase &pm)
611 : ImmutablePass(ID), PM(&pm), TM(&TM) {
612 Impl = new PassConfigImpl();
614 // Register all target independent codegen passes to activate their PassIDs,
615 // including this pass itself.
616 initializeCodeGen(*PassRegistry::getPassRegistry());
618 // Also register alias analysis passes required by codegen passes.
619 initializeBasicAAWrapperPassPass(*PassRegistry::getPassRegistry());
620 initializeAAResultsWrapperPassPass(*PassRegistry::getPassRegistry());
622 if (EnableIPRA.getNumOccurrences())
623 TM.Options.EnableIPRA = EnableIPRA;
624 else {
625 // If not explicitly specified, use target default.
626 TM.Options.EnableIPRA |= TM.useIPRA();
629 if (TM.Options.EnableIPRA)
630 setRequiresCodeGenSCCOrder();
632 if (EnableGlobalISelAbort.getNumOccurrences())
633 TM.Options.GlobalISelAbort = EnableGlobalISelAbort;
635 setStartStopPasses();
638 CodeGenOpt::Level TargetPassConfig::getOptLevel() const {
639 return TM->getOptLevel();
642 /// Insert InsertedPassID pass after TargetPassID.
643 void TargetPassConfig::insertPass(AnalysisID TargetPassID,
644 IdentifyingPassPtr InsertedPassID,
645 bool VerifyAfter) {
646 assert(((!InsertedPassID.isInstance() &&
647 TargetPassID != InsertedPassID.getID()) ||
648 (InsertedPassID.isInstance() &&
649 TargetPassID != InsertedPassID.getInstance()->getPassID())) &&
650 "Insert a pass after itself!");
651 Impl->InsertedPasses.emplace_back(TargetPassID, InsertedPassID, VerifyAfter);
654 /// createPassConfig - Create a pass configuration object to be used by
655 /// addPassToEmitX methods for generating a pipeline of CodeGen passes.
657 /// Targets may override this to extend TargetPassConfig.
658 TargetPassConfig *LLVMTargetMachine::createPassConfig(PassManagerBase &PM) {
659 return new TargetPassConfig(*this, PM);
662 TargetPassConfig::TargetPassConfig()
663 : ImmutablePass(ID) {
664 report_fatal_error("Trying to construct TargetPassConfig without a target "
665 "machine. Scheduling a CodeGen pass without a target "
666 "triple set?");
669 bool TargetPassConfig::willCompleteCodeGenPipeline() {
670 return StopBeforeOpt.empty() && StopAfterOpt.empty();
673 bool TargetPassConfig::hasLimitedCodeGenPipeline() {
674 return !StartBeforeOpt.empty() || !StartAfterOpt.empty() ||
675 !willCompleteCodeGenPipeline();
678 std::string
679 TargetPassConfig::getLimitedCodeGenPipelineReason(const char *Separator) {
680 if (!hasLimitedCodeGenPipeline())
681 return std::string();
682 std::string Res;
683 static cl::opt<std::string> *PassNames[] = {&StartAfterOpt, &StartBeforeOpt,
684 &StopAfterOpt, &StopBeforeOpt};
685 static const char *OptNames[] = {StartAfterOptName, StartBeforeOptName,
686 StopAfterOptName, StopBeforeOptName};
687 bool IsFirst = true;
688 for (int Idx = 0; Idx < 4; ++Idx)
689 if (!PassNames[Idx]->empty()) {
690 if (!IsFirst)
691 Res += Separator;
692 IsFirst = false;
693 Res += OptNames[Idx];
695 return Res;
698 // Helper to verify the analysis is really immutable.
699 void TargetPassConfig::setOpt(bool &Opt, bool Val) {
700 assert(!Initialized && "PassConfig is immutable");
701 Opt = Val;
704 void TargetPassConfig::substitutePass(AnalysisID StandardID,
705 IdentifyingPassPtr TargetID) {
706 Impl->TargetPasses[StandardID] = TargetID;
709 IdentifyingPassPtr TargetPassConfig::getPassSubstitution(AnalysisID ID) const {
710 DenseMap<AnalysisID, IdentifyingPassPtr>::const_iterator
711 I = Impl->TargetPasses.find(ID);
712 if (I == Impl->TargetPasses.end())
713 return ID;
714 return I->second;
717 bool TargetPassConfig::isPassSubstitutedOrOverridden(AnalysisID ID) const {
718 IdentifyingPassPtr TargetID = getPassSubstitution(ID);
719 IdentifyingPassPtr FinalPtr = overridePass(ID, TargetID);
720 return !FinalPtr.isValid() || FinalPtr.isInstance() ||
721 FinalPtr.getID() != ID;
724 /// Add a pass to the PassManager if that pass is supposed to be run. If the
725 /// Started/Stopped flags indicate either that the compilation should start at
726 /// a later pass or that it should stop after an earlier pass, then do not add
727 /// the pass. Finally, compare the current pass against the StartAfter
728 /// and StopAfter options and change the Started/Stopped flags accordingly.
729 void TargetPassConfig::addPass(Pass *P, bool verifyAfter) {
730 assert(!Initialized && "PassConfig is immutable");
732 // Cache the Pass ID here in case the pass manager finds this pass is
733 // redundant with ones already scheduled / available, and deletes it.
734 // Fundamentally, once we add the pass to the manager, we no longer own it
735 // and shouldn't reference it.
736 AnalysisID PassID = P->getPassID();
738 if (StartBefore == PassID && StartBeforeCount++ == StartBeforeInstanceNum)
739 Started = true;
740 if (StopBefore == PassID && StopBeforeCount++ == StopBeforeInstanceNum)
741 Stopped = true;
742 if (Started && !Stopped) {
743 if (AddingMachinePasses)
744 addMachinePrePasses();
745 std::string Banner;
746 // Construct banner message before PM->add() as that may delete the pass.
747 if (AddingMachinePasses && verifyAfter)
748 Banner = std::string("After ") + std::string(P->getPassName());
749 PM->add(P);
750 if (AddingMachinePasses)
751 addMachinePostPasses(Banner, /*AllowVerify*/ verifyAfter);
753 // Add the passes after the pass P if there is any.
754 for (const auto &IP : Impl->InsertedPasses) {
755 if (IP.TargetPassID == PassID)
756 addPass(IP.getInsertedPass(), IP.VerifyAfter);
758 } else {
759 delete P;
762 if (StopAfter == PassID && StopAfterCount++ == StopAfterInstanceNum)
763 Stopped = true;
765 if (StartAfter == PassID && StartAfterCount++ == StartAfterInstanceNum)
766 Started = true;
767 if (Stopped && !Started)
768 report_fatal_error("Cannot stop compilation after pass that is not run");
771 /// Add a CodeGen pass at this point in the pipeline after checking for target
772 /// and command line overrides.
774 /// addPass cannot return a pointer to the pass instance because is internal the
775 /// PassManager and the instance we create here may already be freed.
776 AnalysisID TargetPassConfig::addPass(AnalysisID PassID, bool verifyAfter) {
777 IdentifyingPassPtr TargetID = getPassSubstitution(PassID);
778 IdentifyingPassPtr FinalPtr = overridePass(PassID, TargetID);
779 if (!FinalPtr.isValid())
780 return nullptr;
782 Pass *P;
783 if (FinalPtr.isInstance())
784 P = FinalPtr.getInstance();
785 else {
786 P = Pass::createPass(FinalPtr.getID());
787 if (!P)
788 llvm_unreachable("Pass ID not registered");
790 AnalysisID FinalID = P->getPassID();
791 addPass(P, verifyAfter); // Ends the lifetime of P.
793 return FinalID;
796 void TargetPassConfig::printAndVerify(const std::string &Banner) {
797 addPrintPass(Banner);
798 addVerifyPass(Banner);
801 void TargetPassConfig::addPrintPass(const std::string &Banner) {
802 if (PrintAfterISel)
803 PM->add(createMachineFunctionPrinterPass(dbgs(), Banner));
806 void TargetPassConfig::addVerifyPass(const std::string &Banner) {
807 bool Verify = VerifyMachineCode == cl::BOU_TRUE;
808 #ifdef EXPENSIVE_CHECKS
809 if (VerifyMachineCode == cl::BOU_UNSET)
810 Verify = TM->isMachineVerifierClean();
811 #endif
812 if (Verify)
813 PM->add(createMachineVerifierPass(Banner));
816 void TargetPassConfig::addDebugifyPass() {
817 PM->add(createDebugifyMachineModulePass());
820 void TargetPassConfig::addStripDebugPass() {
821 PM->add(createStripDebugMachineModulePass(/*OnlyDebugified=*/true));
824 void TargetPassConfig::addCheckDebugPass() {
825 PM->add(createCheckDebugMachineModulePass());
828 void TargetPassConfig::addMachinePrePasses(bool AllowDebugify) {
829 if (AllowDebugify && DebugifyIsSafe &&
830 (DebugifyAndStripAll == cl::BOU_TRUE ||
831 DebugifyCheckAndStripAll == cl::BOU_TRUE))
832 addDebugifyPass();
835 void TargetPassConfig::addMachinePostPasses(const std::string &Banner,
836 bool AllowVerify, bool AllowStrip) {
837 if (DebugifyIsSafe) {
838 if (DebugifyCheckAndStripAll == cl::BOU_TRUE) {
839 addCheckDebugPass();
840 addStripDebugPass();
841 } else if (DebugifyAndStripAll == cl::BOU_TRUE)
842 addStripDebugPass();
844 if (AllowVerify)
845 addVerifyPass(Banner);
848 /// Add common target configurable passes that perform LLVM IR to IR transforms
849 /// following machine independent optimization.
850 void TargetPassConfig::addIRPasses() {
851 // Before running any passes, run the verifier to determine if the input
852 // coming from the front-end and/or optimizer is valid.
853 if (!DisableVerify)
854 addPass(createVerifierPass());
856 if (getOptLevel() != CodeGenOpt::None) {
857 switch (UseCFLAA) {
858 case CFLAAType::Steensgaard:
859 addPass(createCFLSteensAAWrapperPass());
860 break;
861 case CFLAAType::Andersen:
862 addPass(createCFLAndersAAWrapperPass());
863 break;
864 case CFLAAType::Both:
865 addPass(createCFLAndersAAWrapperPass());
866 addPass(createCFLSteensAAWrapperPass());
867 break;
868 default:
869 break;
872 // Basic AliasAnalysis support.
873 // Add TypeBasedAliasAnalysis before BasicAliasAnalysis so that
874 // BasicAliasAnalysis wins if they disagree. This is intended to help
875 // support "obvious" type-punning idioms.
876 addPass(createTypeBasedAAWrapperPass());
877 addPass(createScopedNoAliasAAWrapperPass());
878 addPass(createBasicAAWrapperPass());
880 // Run loop strength reduction before anything else.
881 if (!DisableLSR) {
882 addPass(createCanonicalizeFreezeInLoopsPass());
883 addPass(createLoopStrengthReducePass());
884 if (PrintLSR)
885 addPass(createPrintFunctionPass(dbgs(),
886 "\n\n*** Code after LSR ***\n"));
889 // The MergeICmpsPass tries to create memcmp calls by grouping sequences of
890 // loads and compares. ExpandMemCmpPass then tries to expand those calls
891 // into optimally-sized loads and compares. The transforms are enabled by a
892 // target lowering hook.
893 if (!DisableMergeICmps)
894 addPass(createMergeICmpsLegacyPass());
895 addPass(createExpandMemCmpPass());
898 // Run GC lowering passes for builtin collectors
899 // TODO: add a pass insertion point here
900 addPass(&GCLoweringID);
901 addPass(&ShadowStackGCLoweringID);
902 addPass(createLowerConstantIntrinsicsPass());
904 // Make sure that no unreachable blocks are instruction selected.
905 addPass(createUnreachableBlockEliminationPass());
907 // Prepare expensive constants for SelectionDAG.
908 if (getOptLevel() != CodeGenOpt::None && !DisableConstantHoisting)
909 addPass(createConstantHoistingPass());
911 if (getOptLevel() != CodeGenOpt::None)
912 addPass(createReplaceWithVeclibLegacyPass());
914 if (getOptLevel() != CodeGenOpt::None && !DisablePartialLibcallInlining)
915 addPass(createPartiallyInlineLibCallsPass());
917 // Expand vector predication intrinsics into standard IR instructions.
918 // This pass has to run before ScalarizeMaskedMemIntrin and ExpandReduction
919 // passes since it emits those kinds of intrinsics.
920 addPass(createExpandVectorPredicationPass());
922 // Add scalarization of target's unsupported masked memory intrinsics pass.
923 // the unsupported intrinsic will be replaced with a chain of basic blocks,
924 // that stores/loads element one-by-one if the appropriate mask bit is set.
925 addPass(createScalarizeMaskedMemIntrinLegacyPass());
927 // Expand reduction intrinsics into shuffle sequences if the target wants to.
928 // Allow disabling it for testing purposes.
929 if (!DisableExpandReductions)
930 addPass(createExpandReductionsPass());
933 /// Turn exception handling constructs into something the code generators can
934 /// handle.
935 void TargetPassConfig::addPassesToHandleExceptions() {
936 const MCAsmInfo *MCAI = TM->getMCAsmInfo();
937 assert(MCAI && "No MCAsmInfo");
938 switch (MCAI->getExceptionHandlingType()) {
939 case ExceptionHandling::SjLj:
940 // SjLj piggy-backs on dwarf for this bit. The cleanups done apply to both
941 // Dwarf EH prepare needs to be run after SjLj prepare. Otherwise,
942 // catch info can get misplaced when a selector ends up more than one block
943 // removed from the parent invoke(s). This could happen when a landing
944 // pad is shared by multiple invokes and is also a target of a normal
945 // edge from elsewhere.
946 addPass(createSjLjEHPreparePass(TM));
947 LLVM_FALLTHROUGH;
948 case ExceptionHandling::DwarfCFI:
949 case ExceptionHandling::ARM:
950 case ExceptionHandling::AIX:
951 addPass(createDwarfEHPass(getOptLevel()));
952 break;
953 case ExceptionHandling::WinEH:
954 // We support using both GCC-style and MSVC-style exceptions on Windows, so
955 // add both preparation passes. Each pass will only actually run if it
956 // recognizes the personality function.
957 addPass(createWinEHPass());
958 addPass(createDwarfEHPass(getOptLevel()));
959 break;
960 case ExceptionHandling::Wasm:
961 // Wasm EH uses Windows EH instructions, but it does not need to demote PHIs
962 // on catchpads and cleanuppads because it does not outline them into
963 // funclets. Catchswitch blocks are not lowered in SelectionDAG, so we
964 // should remove PHIs there.
965 addPass(createWinEHPass(/*DemoteCatchSwitchPHIOnly=*/false));
966 addPass(createWasmEHPass());
967 break;
968 case ExceptionHandling::None:
969 addPass(createLowerInvokePass());
971 // The lower invoke pass may create unreachable code. Remove it.
972 addPass(createUnreachableBlockEliminationPass());
973 break;
977 /// Add pass to prepare the LLVM IR for code generation. This should be done
978 /// before exception handling preparation passes.
979 void TargetPassConfig::addCodeGenPrepare() {
980 if (getOptLevel() != CodeGenOpt::None && !DisableCGP)
981 addPass(createCodeGenPreparePass());
984 /// Add common passes that perform LLVM IR to IR transforms in preparation for
985 /// instruction selection.
986 void TargetPassConfig::addISelPrepare() {
987 addPreISel();
989 // Force codegen to run according to the callgraph.
990 if (requiresCodeGenSCCOrder())
991 addPass(new DummyCGSCCPass);
993 // Add both the safe stack and the stack protection passes: each of them will
994 // only protect functions that have corresponding attributes.
995 addPass(createSafeStackPass());
996 addPass(createStackProtectorPass());
998 if (PrintISelInput)
999 addPass(createPrintFunctionPass(
1000 dbgs(), "\n\n*** Final LLVM Code input to ISel ***\n"));
1002 // All passes which modify the LLVM IR are now complete; run the verifier
1003 // to ensure that the IR is valid.
1004 if (!DisableVerify)
1005 addPass(createVerifierPass());
1008 bool TargetPassConfig::addCoreISelPasses() {
1009 // Enable FastISel with -fast-isel, but allow that to be overridden.
1010 TM->setO0WantsFastISel(EnableFastISelOption != cl::BOU_FALSE);
1012 // Determine an instruction selector.
1013 enum class SelectorType { SelectionDAG, FastISel, GlobalISel };
1014 SelectorType Selector;
1016 if (EnableFastISelOption == cl::BOU_TRUE)
1017 Selector = SelectorType::FastISel;
1018 else if (EnableGlobalISelOption == cl::BOU_TRUE ||
1019 (TM->Options.EnableGlobalISel &&
1020 EnableGlobalISelOption != cl::BOU_FALSE))
1021 Selector = SelectorType::GlobalISel;
1022 else if (TM->getOptLevel() == CodeGenOpt::None && TM->getO0WantsFastISel())
1023 Selector = SelectorType::FastISel;
1024 else
1025 Selector = SelectorType::SelectionDAG;
1027 // Set consistently TM->Options.EnableFastISel and EnableGlobalISel.
1028 if (Selector == SelectorType::FastISel) {
1029 TM->setFastISel(true);
1030 TM->setGlobalISel(false);
1031 } else if (Selector == SelectorType::GlobalISel) {
1032 TM->setFastISel(false);
1033 TM->setGlobalISel(true);
1036 // FIXME: Injecting into the DAGISel pipeline seems to cause issues with
1037 // analyses needing to be re-run. This can result in being unable to
1038 // schedule passes (particularly with 'Function Alias Analysis
1039 // Results'). It's not entirely clear why but AFAICT this seems to be
1040 // due to one FunctionPassManager not being able to use analyses from a
1041 // previous one. As we're injecting a ModulePass we break the usual
1042 // pass manager into two. GlobalISel with the fallback path disabled
1043 // and -run-pass seem to be unaffected. The majority of GlobalISel
1044 // testing uses -run-pass so this probably isn't too bad.
1045 SaveAndRestore<bool> SavedDebugifyIsSafe(DebugifyIsSafe);
1046 if (Selector != SelectorType::GlobalISel || !isGlobalISelAbortEnabled())
1047 DebugifyIsSafe = false;
1049 // Add instruction selector passes.
1050 if (Selector == SelectorType::GlobalISel) {
1051 SaveAndRestore<bool> SavedAddingMachinePasses(AddingMachinePasses, true);
1052 if (addIRTranslator())
1053 return true;
1055 addPreLegalizeMachineIR();
1057 if (addLegalizeMachineIR())
1058 return true;
1060 // Before running the register bank selector, ask the target if it
1061 // wants to run some passes.
1062 addPreRegBankSelect();
1064 if (addRegBankSelect())
1065 return true;
1067 addPreGlobalInstructionSelect();
1069 if (addGlobalInstructionSelect())
1070 return true;
1072 // Pass to reset the MachineFunction if the ISel failed.
1073 addPass(createResetMachineFunctionPass(
1074 reportDiagnosticWhenGlobalISelFallback(), isGlobalISelAbortEnabled()));
1076 // Provide a fallback path when we do not want to abort on
1077 // not-yet-supported input.
1078 if (!isGlobalISelAbortEnabled() && addInstSelector())
1079 return true;
1081 } else if (addInstSelector())
1082 return true;
1084 // Expand pseudo-instructions emitted by ISel. Don't run the verifier before
1085 // FinalizeISel.
1086 addPass(&FinalizeISelID);
1088 // Print the instruction selected machine code...
1089 printAndVerify("After Instruction Selection");
1091 return false;
1094 bool TargetPassConfig::addISelPasses() {
1095 if (TM->useEmulatedTLS())
1096 addPass(createLowerEmuTLSPass());
1098 addPass(createPreISelIntrinsicLoweringPass());
1099 PM->add(createTargetTransformInfoWrapperPass(TM->getTargetIRAnalysis()));
1100 addIRPasses();
1101 addCodeGenPrepare();
1102 addPassesToHandleExceptions();
1103 addISelPrepare();
1105 return addCoreISelPasses();
1108 /// -regalloc=... command line option.
1109 static FunctionPass *useDefaultRegisterAllocator() { return nullptr; }
1110 static cl::opt<RegisterRegAlloc::FunctionPassCtor, false,
1111 RegisterPassParser<RegisterRegAlloc>>
1112 RegAlloc("regalloc", cl::Hidden, cl::init(&useDefaultRegisterAllocator),
1113 cl::desc("Register allocator to use"));
1115 /// Add the complete set of target-independent postISel code generator passes.
1117 /// This can be read as the standard order of major LLVM CodeGen stages. Stages
1118 /// with nontrivial configuration or multiple passes are broken out below in
1119 /// add%Stage routines.
1121 /// Any TargetPassConfig::addXX routine may be overriden by the Target. The
1122 /// addPre/Post methods with empty header implementations allow injecting
1123 /// target-specific fixups just before or after major stages. Additionally,
1124 /// targets have the flexibility to change pass order within a stage by
1125 /// overriding default implementation of add%Stage routines below. Each
1126 /// technique has maintainability tradeoffs because alternate pass orders are
1127 /// not well supported. addPre/Post works better if the target pass is easily
1128 /// tied to a common pass. But if it has subtle dependencies on multiple passes,
1129 /// the target should override the stage instead.
1131 /// TODO: We could use a single addPre/Post(ID) hook to allow pass injection
1132 /// before/after any target-independent pass. But it's currently overkill.
1133 void TargetPassConfig::addMachinePasses() {
1134 AddingMachinePasses = true;
1136 // Add passes that optimize machine instructions in SSA form.
1137 if (getOptLevel() != CodeGenOpt::None) {
1138 addMachineSSAOptimization();
1139 } else {
1140 // If the target requests it, assign local variables to stack slots relative
1141 // to one another and simplify frame index references where possible.
1142 addPass(&LocalStackSlotAllocationID);
1145 if (TM->Options.EnableIPRA)
1146 addPass(createRegUsageInfoPropPass());
1148 // Run pre-ra passes.
1149 addPreRegAlloc();
1151 // Debugifying the register allocator passes seems to provoke some
1152 // non-determinism that affects CodeGen and there doesn't seem to be a point
1153 // where it becomes safe again so stop debugifying here.
1154 DebugifyIsSafe = false;
1156 // Add a FSDiscriminator pass right before RA, so that we could get
1157 // more precise SampleFDO profile for RA.
1158 if (EnableFSDiscriminator) {
1159 addPass(createMIRAddFSDiscriminatorsPass(
1160 sampleprof::FSDiscriminatorPass::Pass1));
1161 const std::string ProfileFile = getFSProfileFile(TM);
1162 if (!ProfileFile.empty() && !DisableRAFSProfileLoader)
1163 addPass(
1164 createMIRProfileLoaderPass(ProfileFile, getFSRemappingFile(TM),
1165 sampleprof::FSDiscriminatorPass::Pass1));
1168 // Run register allocation and passes that are tightly coupled with it,
1169 // including phi elimination and scheduling.
1170 if (getOptimizeRegAlloc())
1171 addOptimizedRegAlloc();
1172 else
1173 addFastRegAlloc();
1175 // Run post-ra passes.
1176 addPostRegAlloc();
1178 addPass(&RemoveRedundantDebugValuesID, false);
1180 addPass(&FixupStatepointCallerSavedID);
1182 // Insert prolog/epilog code. Eliminate abstract frame index references...
1183 if (getOptLevel() != CodeGenOpt::None) {
1184 addPass(&PostRAMachineSinkingID);
1185 addPass(&ShrinkWrapID);
1188 // Prolog/Epilog inserter needs a TargetMachine to instantiate. But only
1189 // do so if it hasn't been disabled, substituted, or overridden.
1190 if (!isPassSubstitutedOrOverridden(&PrologEpilogCodeInserterID))
1191 addPass(createPrologEpilogInserterPass());
1193 /// Add passes that optimize machine instructions after register allocation.
1194 if (getOptLevel() != CodeGenOpt::None)
1195 addMachineLateOptimization();
1197 // Expand pseudo instructions before second scheduling pass.
1198 addPass(&ExpandPostRAPseudosID);
1200 // Run pre-sched2 passes.
1201 addPreSched2();
1203 if (EnableImplicitNullChecks)
1204 addPass(&ImplicitNullChecksID);
1206 // Second pass scheduler.
1207 // Let Target optionally insert this pass by itself at some other
1208 // point.
1209 if (getOptLevel() != CodeGenOpt::None &&
1210 !TM->targetSchedulesPostRAScheduling()) {
1211 if (MISchedPostRA)
1212 addPass(&PostMachineSchedulerID);
1213 else
1214 addPass(&PostRASchedulerID);
1217 // GC
1218 if (addGCPasses()) {
1219 if (PrintGCInfo)
1220 addPass(createGCInfoPrinter(dbgs()), false);
1223 // Basic block placement.
1224 if (getOptLevel() != CodeGenOpt::None)
1225 addBlockPlacement();
1227 // Insert before XRay Instrumentation.
1228 addPass(&FEntryInserterID);
1230 addPass(&XRayInstrumentationID);
1231 addPass(&PatchableFunctionID);
1233 if (EnableFSDiscriminator && !FSNoFinalDiscrim)
1234 // Add FS discriminators here so that all the instruction duplicates
1235 // in different BBs get their own discriminators. With this, we can "sum"
1236 // the SampleFDO counters instead of using MAX. This will improve the
1237 // SampleFDO profile quality.
1238 addPass(createMIRAddFSDiscriminatorsPass(
1239 sampleprof::FSDiscriminatorPass::PassLast));
1241 addPreEmitPass();
1243 if (TM->Options.EnableIPRA)
1244 // Collect register usage information and produce a register mask of
1245 // clobbered registers, to be used to optimize call sites.
1246 addPass(createRegUsageInfoCollector());
1248 // FIXME: Some backends are incompatible with running the verifier after
1249 // addPreEmitPass. Maybe only pass "false" here for those targets?
1250 addPass(&FuncletLayoutID, false);
1252 addPass(&StackMapLivenessID, false);
1253 addPass(&LiveDebugValuesID, false);
1255 if (TM->Options.EnableMachineOutliner && getOptLevel() != CodeGenOpt::None &&
1256 EnableMachineOutliner != RunOutliner::NeverOutline) {
1257 bool RunOnAllFunctions =
1258 (EnableMachineOutliner == RunOutliner::AlwaysOutline);
1259 bool AddOutliner =
1260 RunOnAllFunctions || TM->Options.SupportsDefaultOutlining;
1261 if (AddOutliner)
1262 addPass(createMachineOutlinerPass(RunOnAllFunctions));
1265 // Machine function splitter uses the basic block sections feature. Both
1266 // cannot be enabled at the same time. Basic block sections takes precedence.
1267 // FIXME: In principle, BasicBlockSection::Labels and splitting can used
1268 // together. Update this check once we have addressed any issues.
1269 if (TM->getBBSectionsType() != llvm::BasicBlockSection::None) {
1270 addPass(llvm::createBasicBlockSectionsPass(TM->getBBSectionsFuncListBuf()));
1271 } else if (TM->Options.EnableMachineFunctionSplitter ||
1272 EnableMachineFunctionSplitter) {
1273 addPass(createMachineFunctionSplitterPass());
1276 // Add passes that directly emit MI after all other MI passes.
1277 addPreEmitPass2();
1279 // Insert pseudo probe annotation for callsite profiling
1280 if (TM->Options.PseudoProbeForProfiling)
1281 addPass(createPseudoProbeInserter());
1283 AddingMachinePasses = false;
1286 /// Add passes that optimize machine instructions in SSA form.
1287 void TargetPassConfig::addMachineSSAOptimization() {
1288 // Pre-ra tail duplication.
1289 addPass(&EarlyTailDuplicateID);
1291 // Optimize PHIs before DCE: removing dead PHI cycles may make more
1292 // instructions dead.
1293 addPass(&OptimizePHIsID);
1295 // This pass merges large allocas. StackSlotColoring is a different pass
1296 // which merges spill slots.
1297 addPass(&StackColoringID);
1299 // If the target requests it, assign local variables to stack slots relative
1300 // to one another and simplify frame index references where possible.
1301 addPass(&LocalStackSlotAllocationID);
1303 // With optimization, dead code should already be eliminated. However
1304 // there is one known exception: lowered code for arguments that are only
1305 // used by tail calls, where the tail calls reuse the incoming stack
1306 // arguments directly (see t11 in test/CodeGen/X86/sibcall.ll).
1307 addPass(&DeadMachineInstructionElimID);
1309 // Allow targets to insert passes that improve instruction level parallelism,
1310 // like if-conversion. Such passes will typically need dominator trees and
1311 // loop info, just like LICM and CSE below.
1312 addILPOpts();
1314 addPass(&EarlyMachineLICMID);
1315 addPass(&MachineCSEID);
1317 addPass(&MachineSinkingID);
1319 addPass(&PeepholeOptimizerID);
1320 // Clean-up the dead code that may have been generated by peephole
1321 // rewriting.
1322 addPass(&DeadMachineInstructionElimID);
1325 //===---------------------------------------------------------------------===//
1326 /// Register Allocation Pass Configuration
1327 //===---------------------------------------------------------------------===//
1329 bool TargetPassConfig::getOptimizeRegAlloc() const {
1330 switch (OptimizeRegAlloc) {
1331 case cl::BOU_UNSET: return getOptLevel() != CodeGenOpt::None;
1332 case cl::BOU_TRUE: return true;
1333 case cl::BOU_FALSE: return false;
1335 llvm_unreachable("Invalid optimize-regalloc state");
1338 /// A dummy default pass factory indicates whether the register allocator is
1339 /// overridden on the command line.
1340 static llvm::once_flag InitializeDefaultRegisterAllocatorFlag;
1342 static RegisterRegAlloc
1343 defaultRegAlloc("default",
1344 "pick register allocator based on -O option",
1345 useDefaultRegisterAllocator);
1347 static void initializeDefaultRegisterAllocatorOnce() {
1348 if (!RegisterRegAlloc::getDefault())
1349 RegisterRegAlloc::setDefault(RegAlloc);
1352 /// Instantiate the default register allocator pass for this target for either
1353 /// the optimized or unoptimized allocation path. This will be added to the pass
1354 /// manager by addFastRegAlloc in the unoptimized case or addOptimizedRegAlloc
1355 /// in the optimized case.
1357 /// A target that uses the standard regalloc pass order for fast or optimized
1358 /// allocation may still override this for per-target regalloc
1359 /// selection. But -regalloc=... always takes precedence.
1360 FunctionPass *TargetPassConfig::createTargetRegisterAllocator(bool Optimized) {
1361 if (Optimized)
1362 return createGreedyRegisterAllocator();
1363 else
1364 return createFastRegisterAllocator();
1367 /// Find and instantiate the register allocation pass requested by this target
1368 /// at the current optimization level. Different register allocators are
1369 /// defined as separate passes because they may require different analysis.
1371 /// This helper ensures that the regalloc= option is always available,
1372 /// even for targets that override the default allocator.
1374 /// FIXME: When MachinePassRegistry register pass IDs instead of function ptrs,
1375 /// this can be folded into addPass.
1376 FunctionPass *TargetPassConfig::createRegAllocPass(bool Optimized) {
1377 // Initialize the global default.
1378 llvm::call_once(InitializeDefaultRegisterAllocatorFlag,
1379 initializeDefaultRegisterAllocatorOnce);
1381 RegisterRegAlloc::FunctionPassCtor Ctor = RegisterRegAlloc::getDefault();
1382 if (Ctor != useDefaultRegisterAllocator)
1383 return Ctor();
1385 // With no -regalloc= override, ask the target for a regalloc pass.
1386 return createTargetRegisterAllocator(Optimized);
1389 bool TargetPassConfig::addRegAssignAndRewriteFast() {
1390 if (RegAlloc != (RegisterRegAlloc::FunctionPassCtor)&useDefaultRegisterAllocator &&
1391 RegAlloc != (RegisterRegAlloc::FunctionPassCtor)&createFastRegisterAllocator)
1392 report_fatal_error("Must use fast (default) register allocator for unoptimized regalloc.");
1394 addPass(createRegAllocPass(false));
1396 // Allow targets to change the register assignments after
1397 // fast register allocation.
1398 addPostFastRegAllocRewrite();
1399 return true;
1402 bool TargetPassConfig::addRegAssignAndRewriteOptimized() {
1403 // Add the selected register allocation pass.
1404 addPass(createRegAllocPass(true));
1406 // Allow targets to change the register assignments before rewriting.
1407 addPreRewrite();
1409 // Finally rewrite virtual registers.
1410 addPass(&VirtRegRewriterID);
1412 return true;
1415 /// Return true if the default global register allocator is in use and
1416 /// has not be overriden on the command line with '-regalloc=...'
1417 bool TargetPassConfig::usingDefaultRegAlloc() const {
1418 return RegAlloc.getNumOccurrences() == 0;
1421 /// Add the minimum set of target-independent passes that are required for
1422 /// register allocation. No coalescing or scheduling.
1423 void TargetPassConfig::addFastRegAlloc() {
1424 addPass(&PHIEliminationID, false);
1425 addPass(&TwoAddressInstructionPassID, false);
1427 addRegAssignAndRewriteFast();
1430 /// Add standard target-independent passes that are tightly coupled with
1431 /// optimized register allocation, including coalescing, machine instruction
1432 /// scheduling, and register allocation itself.
1433 void TargetPassConfig::addOptimizedRegAlloc() {
1434 addPass(&DetectDeadLanesID, false);
1436 addPass(&ProcessImplicitDefsID, false);
1438 // LiveVariables currently requires pure SSA form.
1440 // FIXME: Once TwoAddressInstruction pass no longer uses kill flags,
1441 // LiveVariables can be removed completely, and LiveIntervals can be directly
1442 // computed. (We still either need to regenerate kill flags after regalloc, or
1443 // preferably fix the scavenger to not depend on them).
1444 // FIXME: UnreachableMachineBlockElim is a dependant pass of LiveVariables.
1445 // When LiveVariables is removed this has to be removed/moved either.
1446 // Explicit addition of UnreachableMachineBlockElim allows stopping before or
1447 // after it with -stop-before/-stop-after.
1448 addPass(&UnreachableMachineBlockElimID, false);
1449 addPass(&LiveVariablesID, false);
1451 // Edge splitting is smarter with machine loop info.
1452 addPass(&MachineLoopInfoID, false);
1453 addPass(&PHIEliminationID, false);
1455 // Eventually, we want to run LiveIntervals before PHI elimination.
1456 if (EarlyLiveIntervals)
1457 addPass(&LiveIntervalsID, false);
1459 addPass(&TwoAddressInstructionPassID, false);
1460 addPass(&RegisterCoalescerID);
1462 // The machine scheduler may accidentally create disconnected components
1463 // when moving subregister definitions around, avoid this by splitting them to
1464 // separate vregs before. Splitting can also improve reg. allocation quality.
1465 addPass(&RenameIndependentSubregsID);
1467 // PreRA instruction scheduling.
1468 addPass(&MachineSchedulerID);
1470 if (addRegAssignAndRewriteOptimized()) {
1471 // Perform stack slot coloring and post-ra machine LICM.
1473 // FIXME: Re-enable coloring with register when it's capable of adding
1474 // kill markers.
1475 addPass(&StackSlotColoringID);
1477 // Allow targets to expand pseudo instructions depending on the choice of
1478 // registers before MachineCopyPropagation.
1479 addPostRewrite();
1481 // Copy propagate to forward register uses and try to eliminate COPYs that
1482 // were not coalesced.
1483 addPass(&MachineCopyPropagationID);
1485 // Run post-ra machine LICM to hoist reloads / remats.
1487 // FIXME: can this move into MachineLateOptimization?
1488 addPass(&MachineLICMID);
1492 //===---------------------------------------------------------------------===//
1493 /// Post RegAlloc Pass Configuration
1494 //===---------------------------------------------------------------------===//
1496 /// Add passes that optimize machine instructions after register allocation.
1497 void TargetPassConfig::addMachineLateOptimization() {
1498 // Branch folding must be run after regalloc and prolog/epilog insertion.
1499 addPass(&BranchFolderPassID);
1501 // Tail duplication.
1502 // Note that duplicating tail just increases code size and degrades
1503 // performance for targets that require Structured Control Flow.
1504 // In addition it can also make CFG irreducible. Thus we disable it.
1505 if (!TM->requiresStructuredCFG())
1506 addPass(&TailDuplicateID);
1508 // Copy propagation.
1509 addPass(&MachineCopyPropagationID);
1512 /// Add standard GC passes.
1513 bool TargetPassConfig::addGCPasses() {
1514 addPass(&GCMachineCodeAnalysisID, false);
1515 return true;
1518 /// Add standard basic block placement passes.
1519 void TargetPassConfig::addBlockPlacement() {
1520 if (EnableFSDiscriminator) {
1521 addPass(createMIRAddFSDiscriminatorsPass(
1522 sampleprof::FSDiscriminatorPass::Pass2));
1523 const std::string ProfileFile = getFSProfileFile(TM);
1524 if (!ProfileFile.empty() && !DisableLayoutFSProfileLoader)
1525 addPass(
1526 createMIRProfileLoaderPass(ProfileFile, getFSRemappingFile(TM),
1527 sampleprof::FSDiscriminatorPass::Pass2));
1529 if (addPass(&MachineBlockPlacementID)) {
1530 // Run a separate pass to collect block placement statistics.
1531 if (EnableBlockPlacementStats)
1532 addPass(&MachineBlockPlacementStatsID);
1536 //===---------------------------------------------------------------------===//
1537 /// GlobalISel Configuration
1538 //===---------------------------------------------------------------------===//
1539 bool TargetPassConfig::isGlobalISelAbortEnabled() const {
1540 return TM->Options.GlobalISelAbort == GlobalISelAbortMode::Enable;
1543 bool TargetPassConfig::reportDiagnosticWhenGlobalISelFallback() const {
1544 return TM->Options.GlobalISelAbort == GlobalISelAbortMode::DisableWithDiag;
1547 bool TargetPassConfig::isGISelCSEEnabled() const {
1548 return true;
1551 std::unique_ptr<CSEConfigBase> TargetPassConfig::getCSEConfig() const {
1552 return std::make_unique<CSEConfigBase>();