1 //===- PassManagerBuilder.cpp - Build Standard Pass -----------------------===//
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 the PassManagerBuilder class, which is used to set up a
10 // "standard" optimization sequence suitable for languages like C and C++.
11 //
12 //===----------------------------------------------------------------------===//
13
14 #include "llvm/Transforms/IPO/PassManagerBuilder.h"
15 #include "llvm-c/Transforms/PassManagerBuilder.h"
16 #include "llvm/ADT/STLExtras.h"
17 #include "llvm/ADT/SmallVector.h"
18 #include "llvm/Analysis/BasicAliasAnalysis.h"
19 #include "llvm/Analysis/CFLAndersAliasAnalysis.h"
20 #include "llvm/Analysis/CFLSteensAliasAnalysis.h"
21 #include "llvm/Analysis/GlobalsModRef.h"
22 #include "llvm/Analysis/InlineCost.h"
23 #include "llvm/Analysis/Passes.h"
24 #include "llvm/Analysis/ScopedNoAliasAA.h"
25 #include "llvm/Analysis/TargetLibraryInfo.h"
26 #include "llvm/Analysis/TypeBasedAliasAnalysis.h"
27 #include "llvm/IR/DataLayout.h"
28 #include "llvm/IR/LegacyPassManager.h"
29 #include "llvm/IR/Verifier.h"
30 #include "llvm/Support/CommandLine.h"
31 #include "llvm/Support/ManagedStatic.h"
32 #include "llvm/Transforms/AggressiveInstCombine/AggressiveInstCombine.h"
33 #include "llvm/Transforms/IPO.h"
34 #include "llvm/Transforms/IPO/Attributor.h"
35 #include "llvm/Transforms/IPO/ForceFunctionAttrs.h"
36 #include "llvm/Transforms/IPO/FunctionAttrs.h"
37 #include "llvm/Transforms/IPO/InferFunctionAttrs.h"
38 #include "llvm/Transforms/InstCombine/InstCombine.h"
39 #include "llvm/Transforms/Instrumentation.h"
40 #include "llvm/Transforms/Scalar.h"
41 #include "llvm/Transforms/Scalar/GVN.h"
42 #include "llvm/Transforms/Scalar/InstSimplifyPass.h"
43 #include "llvm/Transforms/Scalar/LICM.h"
44 #include "llvm/Transforms/Scalar/LoopUnrollPass.h"
45 #include "llvm/Transforms/Scalar/SimpleLoopUnswitch.h"
46 #include "llvm/Transforms/Utils.h"
47 #include "llvm/Transforms/Vectorize.h"
48 #include "llvm/Transforms/Vectorize/LoopVectorize.h"
49 #include "llvm/Transforms/Vectorize/SLPVectorizer.h"
50
51 using namespace llvm;
52
53 static cl::opt<bool>
54 RunPartialInlining("enable-partial-inlining", cl::init(false), cl::Hidden,
55 cl::ZeroOrMore, cl::desc("Run Partial inlinining pass"));
56
57 static cl::opt<bool>
58 UseGVNAfterVectorization("use-gvn-after-vectorization",
59 cl::init(false), cl::Hidden,
60 cl::desc("Run GVN instead of Early CSE after vectorization passes"));
61
62 static cl::opt<bool> ExtraVectorizerPasses(
63 "extra-vectorizer-passes", cl::init(false), cl::Hidden,
64 cl::desc("Run cleanup optimization passes after vectorization."));
65
66 static cl::opt<bool>
67 RunLoopRerolling("reroll-loops", cl::Hidden,
68 cl::desc("Run the loop rerolling pass"));
69
70 static cl::opt<bool> RunNewGVN("enable-newgvn", cl::init(false), cl::Hidden,
71 cl::desc("Run the NewGVN pass"));
72
73 // Experimental option to use CFL-AA
74 enum class CFLAAType { None, Steensgaard, Andersen, Both };
75 static cl::opt<CFLAAType>
76 UseCFLAA("use-cfl-aa", cl::init(CFLAAType::None), cl::Hidden,
77 cl::desc("Enable the new, experimental CFL alias analysis"),
78 cl::values(clEnumValN(CFLAAType::None, "none", "Disable CFL-AA"),
79 clEnumValN(CFLAAType::Steensgaard, "steens",
80 "Enable unification-based CFL-AA"),
81 clEnumValN(CFLAAType::Andersen, "anders",
82 "Enable inclusion-based CFL-AA"),
83 clEnumValN(CFLAAType::Both, "both",
84 "Enable both variants of CFL-AA")));
85
86 static cl::opt<bool> EnableLoopInterchange(
87 "enable-loopinterchange", cl::init(false), cl::Hidden,
88 cl::desc("Enable the new, experimental LoopInterchange Pass"));
89
90 static cl::opt<bool> EnableUnrollAndJam("enable-unroll-and-jam",
91 cl::init(false), cl::Hidden,
92 cl::desc("Enable Unroll And Jam Pass"));
93
94 static cl::opt<bool>
95 EnablePrepareForThinLTO("prepare-for-thinlto", cl::init(false), cl::Hidden,
96 cl::desc("Enable preparation for ThinLTO."));
97
98 static cl::opt<bool>
99 EnablePerformThinLTO("perform-thinlto", cl::init(false), cl::Hidden,
100 cl::desc("Enable performing ThinLTO."));
101
102 cl::opt<bool> EnableHotColdSplit("hot-cold-split", cl::init(false), cl::Hidden,
103 cl::desc("Enable hot-cold splitting pass"));
104
105 static cl::opt<bool> UseLoopVersioningLICM(
106 "enable-loop-versioning-licm", cl::init(false), cl::Hidden,
107 cl::desc("Enable the experimental Loop Versioning LICM pass"));
108
109 static cl::opt<bool>
110 DisablePreInliner("disable-preinline", cl::init(false), cl::Hidden,
111 cl::desc("Disable pre-instrumentation inliner"));
112
113 static cl::opt<int> PreInlineThreshold(
114 "preinline-threshold", cl::Hidden, cl::init(75), cl::ZeroOrMore,
115 cl::desc("Control the amount of inlining in pre-instrumentation inliner "
116 "(default = 75)"));
117
118 static cl::opt<bool> EnableGVNHoist(
119 "enable-gvn-hoist", cl::init(false), cl::Hidden,
120 cl::desc("Enable the GVN hoisting pass (default = off)"));
121
122 static cl::opt<bool>
123 DisableLibCallsShrinkWrap("disable-libcalls-shrinkwrap", cl::init(false),
124 cl::Hidden,
125 cl::desc("Disable shrink-wrap library calls"));
126
127 static cl::opt<bool> EnableSimpleLoopUnswitch(
128 "enable-simple-loop-unswitch", cl::init(false), cl::Hidden,
129 cl::desc("Enable the simple loop unswitch pass. Also enables independent "
130 "cleanup passes integrated into the loop pass manager pipeline."));
131
132 static cl::opt<bool> EnableGVNSink(
133 "enable-gvn-sink", cl::init(false), cl::Hidden,
134 cl::desc("Enable the GVN sinking pass (default = off)"));
135
136 // This option is used in simplifying testing SampleFDO optimizations for
137 // profile loading.
138 static cl::opt<bool>
139 EnableCHR("enable-chr", cl::init(true), cl::Hidden,
140 cl::desc("Enable control height reduction optimization (CHR)"));
141
142 cl::opt<bool> FlattenedProfileUsed(
143 "flattened-profile-used", cl::init(false), cl::Hidden,
144 cl::desc("Indicate the sample profile being used is flattened, i.e., "
145 "no inline hierachy exists in the profile. "));
146
147 cl::opt<bool> EnableOrderFileInstrumentation(
148 "enable-order-file-instrumentation", cl::init(false), cl::Hidden,
149 cl::desc("Enable order file instrumentation (default = off)"));
150
151 static cl::opt<bool>
152 EnableMatrix("enable-matrix", cl::init(false), cl::Hidden,
153 cl::desc("Enable lowering of the matrix intrinsics"));
154
PassManagerBuilder()155 PassManagerBuilder::PassManagerBuilder() {
156 OptLevel = 2;
157 SizeLevel = 0;
158 LibraryInfo = nullptr;
159 Inliner = nullptr;
160 DisableUnrollLoops = false;
161 SLPVectorize = RunSLPVectorization;
162 LoopVectorize = EnableLoopVectorization;
163 LoopsInterleaved = EnableLoopInterleaving;
164 RerollLoops = RunLoopRerolling;
165 NewGVN = RunNewGVN;
166 LicmMssaOptCap = SetLicmMssaOptCap;
167 LicmMssaNoAccForPromotionCap = SetLicmMssaNoAccForPromotionCap;
168 DisableGVNLoadPRE = false;
169 ForgetAllSCEVInLoopUnroll = ForgetSCEVInLoopUnroll;
170 VerifyInput = false;
171 VerifyOutput = false;
172 MergeFunctions = false;
173 PrepareForLTO = false;
174 EnablePGOInstrGen = false;
175 EnablePGOCSInstrGen = false;
176 EnablePGOCSInstrUse = false;
177 PGOInstrGen = "";
178 PGOInstrUse = "";
179 PGOSampleUse = "";
180 PrepareForThinLTO = EnablePrepareForThinLTO;
181 PerformThinLTO = EnablePerformThinLTO;
182 DivergentTarget = false;
183 }
184
~PassManagerBuilder()185 PassManagerBuilder::~PassManagerBuilder() {
186 delete LibraryInfo;
187 delete Inliner;
188 }
189
190 /// Set of global extensions, automatically added as part of the standard set.
191 static ManagedStatic<
192 SmallVector<std::tuple<PassManagerBuilder::ExtensionPointTy,
193 PassManagerBuilder::ExtensionFn,
194 PassManagerBuilder::GlobalExtensionID>,
195 8>>
196 GlobalExtensions;
197 static PassManagerBuilder::GlobalExtensionID GlobalExtensionsCounter;
198
199 /// Check if GlobalExtensions is constructed and not empty.
200 /// Since GlobalExtensions is a managed static, calling 'empty()' will trigger
201 /// the construction of the object.
GlobalExtensionsNotEmpty()202 static bool GlobalExtensionsNotEmpty() {
203 return GlobalExtensions.isConstructed() && !GlobalExtensions->empty();
204 }
205
206 PassManagerBuilder::GlobalExtensionID
addGlobalExtension(PassManagerBuilder::ExtensionPointTy Ty,PassManagerBuilder::ExtensionFn Fn)207 PassManagerBuilder::addGlobalExtension(PassManagerBuilder::ExtensionPointTy Ty,
208 PassManagerBuilder::ExtensionFn Fn) {
209 auto ExtensionID = GlobalExtensionsCounter++;
210 GlobalExtensions->push_back(std::make_tuple(Ty, std::move(Fn), ExtensionID));
211 return ExtensionID;
212 }
213
removeGlobalExtension(PassManagerBuilder::GlobalExtensionID ExtensionID)214 void PassManagerBuilder::removeGlobalExtension(
215 PassManagerBuilder::GlobalExtensionID ExtensionID) {
216 // RegisterStandardPasses may try to call this function after GlobalExtensions
217 // has already been destroyed; doing so should not generate an error.
218 if (!GlobalExtensions.isConstructed())
219 return;
220
221 auto GlobalExtension =
222 llvm::find_if(*GlobalExtensions, [ExtensionID](const auto &elem) {
223 return std::get<2>(elem) == ExtensionID;
224 });
225 assert(GlobalExtension != GlobalExtensions->end() &&
226 "The extension ID to be removed should always be valid.");
227
228 GlobalExtensions->erase(GlobalExtension);
229 }
230
addExtension(ExtensionPointTy Ty,ExtensionFn Fn)231 void PassManagerBuilder::addExtension(ExtensionPointTy Ty, ExtensionFn Fn) {
232 Extensions.push_back(std::make_pair(Ty, std::move(Fn)));
233 }
234
addExtensionsToPM(ExtensionPointTy ETy,legacy::PassManagerBase & PM) const235 void PassManagerBuilder::addExtensionsToPM(ExtensionPointTy ETy,
236 legacy::PassManagerBase &PM) const {
237 if (GlobalExtensionsNotEmpty()) {
238 for (auto &Ext : *GlobalExtensions) {
239 if (std::get<0>(Ext) == ETy)
240 std::get<1>(Ext)(*this, PM);
241 }
242 }
243 for (unsigned i = 0, e = Extensions.size(); i != e; ++i)
244 if (Extensions[i].first == ETy)
245 Extensions[i].second(*this, PM);
246 }
247
addInitialAliasAnalysisPasses(legacy::PassManagerBase & PM) const248 void PassManagerBuilder::addInitialAliasAnalysisPasses(
249 legacy::PassManagerBase &PM) const {
250 switch (UseCFLAA) {
251 case CFLAAType::Steensgaard:
252 PM.add(createCFLSteensAAWrapperPass());
253 break;
254 case CFLAAType::Andersen:
255 PM.add(createCFLAndersAAWrapperPass());
256 break;
257 case CFLAAType::Both:
258 PM.add(createCFLSteensAAWrapperPass());
259 PM.add(createCFLAndersAAWrapperPass());
260 break;
261 default:
262 break;
263 }
264
265 // Add TypeBasedAliasAnalysis before BasicAliasAnalysis so that
266 // BasicAliasAnalysis wins if they disagree. This is intended to help
267 // support "obvious" type-punning idioms.
268 PM.add(createTypeBasedAAWrapperPass());
269 PM.add(createScopedNoAliasAAWrapperPass());
270 }
271
addInstructionCombiningPass(legacy::PassManagerBase & PM) const272 void PassManagerBuilder::addInstructionCombiningPass(
273 legacy::PassManagerBase &PM) const {
274 bool ExpensiveCombines = OptLevel > 2;
275 PM.add(createInstructionCombiningPass(ExpensiveCombines));
276 }
277
populateFunctionPassManager(legacy::FunctionPassManager & FPM)278 void PassManagerBuilder::populateFunctionPassManager(
279 legacy::FunctionPassManager &FPM) {
280 addExtensionsToPM(EP_EarlyAsPossible, FPM);
281 FPM.add(createEntryExitInstrumenterPass());
282
283 // Add LibraryInfo if we have some.
284 if (LibraryInfo)
285 FPM.add(new TargetLibraryInfoWrapperPass(*LibraryInfo));
286
287 if (OptLevel == 0) return;
288
289 addInitialAliasAnalysisPasses(FPM);
290
291 FPM.add(createCFGSimplificationPass());
292 FPM.add(createSROAPass());
293 FPM.add(createEarlyCSEPass());
294 FPM.add(createLowerExpectIntrinsicPass());
295 }
296
297 // Do PGO instrumentation generation or use pass as the option specified.
addPGOInstrPasses(legacy::PassManagerBase & MPM,bool IsCS=false)298 void PassManagerBuilder::addPGOInstrPasses(legacy::PassManagerBase &MPM,
299 bool IsCS = false) {
300 if (IsCS) {
301 if (!EnablePGOCSInstrGen && !EnablePGOCSInstrUse)
302 return;
303 } else if (!EnablePGOInstrGen && PGOInstrUse.empty() && PGOSampleUse.empty())
304 return;
305
306 // Perform the preinline and cleanup passes for O1 and above.
307 // And avoid doing them if optimizing for size.
308 // We will not do this inline for context sensitive PGO (when IsCS is true).
309 if (OptLevel > 0 && SizeLevel == 0 && !DisablePreInliner &&
310 PGOSampleUse.empty() && !IsCS) {
311 // Create preinline pass. We construct an InlineParams object and specify
312 // the threshold here to avoid the command line options of the regular
313 // inliner to influence pre-inlining. The only fields of InlineParams we
314 // care about are DefaultThreshold and HintThreshold.
315 InlineParams IP;
316 IP.DefaultThreshold = PreInlineThreshold;
317 // FIXME: The hint threshold has the same value used by the regular inliner.
318 // This should probably be lowered after performance testing.
319 IP.HintThreshold = 325;
320
321 MPM.add(createFunctionInliningPass(IP));
322 MPM.add(createSROAPass());
323 MPM.add(createEarlyCSEPass()); // Catch trivial redundancies
324 MPM.add(createCFGSimplificationPass()); // Merge & remove BBs
325 MPM.add(createInstructionCombiningPass()); // Combine silly seq's
326 addExtensionsToPM(EP_Peephole, MPM);
327 }
328 if ((EnablePGOInstrGen && !IsCS) || (EnablePGOCSInstrGen && IsCS)) {
329 MPM.add(createPGOInstrumentationGenLegacyPass(IsCS));
330 // Add the profile lowering pass.
331 InstrProfOptions Options;
332 if (!PGOInstrGen.empty())
333 Options.InstrProfileOutput = PGOInstrGen;
334 Options.DoCounterPromotion = true;
335 Options.UseBFIInPromotion = IsCS;
336 MPM.add(createLoopRotatePass());
337 MPM.add(createInstrProfilingLegacyPass(Options, IsCS));
338 }
339 if (!PGOInstrUse.empty())
340 MPM.add(createPGOInstrumentationUseLegacyPass(PGOInstrUse, IsCS));
341 // Indirect call promotion that promotes intra-module targets only.
342 // For ThinLTO this is done earlier due to interactions with globalopt
343 // for imported functions. We don't run this at -O0.
344 if (OptLevel > 0 && !IsCS)
345 MPM.add(
346 createPGOIndirectCallPromotionLegacyPass(false, !PGOSampleUse.empty()));
347 }
addFunctionSimplificationPasses(legacy::PassManagerBase & MPM)348 void PassManagerBuilder::addFunctionSimplificationPasses(
349 legacy::PassManagerBase &MPM) {
350 // Start of function pass.
351 // Break up aggregate allocas, using SSAUpdater.
352 assert(OptLevel >= 1 && "Calling function optimizer with no optimization level!");
353 MPM.add(createSROAPass());
354 MPM.add(createEarlyCSEPass(true /* Enable mem-ssa. */)); // Catch trivial redundancies
355
356 if (OptLevel > 1) {
357 if (EnableGVNHoist)
358 MPM.add(createGVNHoistPass());
359 if (EnableGVNSink) {
360 MPM.add(createGVNSinkPass());
361 MPM.add(createCFGSimplificationPass());
362 }
363 }
364
365 if (OptLevel > 1) {
366 // Speculative execution if the target has divergent branches; otherwise nop.
367 MPM.add(createSpeculativeExecutionIfHasBranchDivergencePass());
368
369 MPM.add(createJumpThreadingPass()); // Thread jumps.
370 MPM.add(createCorrelatedValuePropagationPass()); // Propagate conditionals
371 }
372 MPM.add(createCFGSimplificationPass()); // Merge & remove BBs
373 // Combine silly seq's
374 if (OptLevel > 2)
375 MPM.add(createAggressiveInstCombinerPass());
376 addInstructionCombiningPass(MPM);
377 if (SizeLevel == 0 && !DisableLibCallsShrinkWrap)
378 MPM.add(createLibCallsShrinkWrapPass());
379 addExtensionsToPM(EP_Peephole, MPM);
380
381 // Optimize memory intrinsic calls based on the profiled size information.
382 if (SizeLevel == 0)
383 MPM.add(createPGOMemOPSizeOptLegacyPass());
384
385 // TODO: Investigate the cost/benefit of tail call elimination on debugging.
386 if (OptLevel > 1)
387 MPM.add(createTailCallEliminationPass()); // Eliminate tail calls
388 MPM.add(createCFGSimplificationPass()); // Merge & remove BBs
389 MPM.add(createReassociatePass()); // Reassociate expressions
390
391 // Begin the loop pass pipeline.
392 if (EnableSimpleLoopUnswitch) {
393 // The simple loop unswitch pass relies on separate cleanup passes. Schedule
394 // them first so when we re-process a loop they run before other loop
395 // passes.
396 MPM.add(createLoopInstSimplifyPass());
397 MPM.add(createLoopSimplifyCFGPass());
398 }
399 // Rotate Loop - disable header duplication at -Oz
400 MPM.add(createLoopRotatePass(SizeLevel == 2 ? 0 : -1));
401 // TODO: Investigate promotion cap for O1.
402 MPM.add(createLICMPass(LicmMssaOptCap, LicmMssaNoAccForPromotionCap));
403 if (EnableSimpleLoopUnswitch)
404 MPM.add(createSimpleLoopUnswitchLegacyPass());
405 else
406 MPM.add(createLoopUnswitchPass(SizeLevel || OptLevel < 3, DivergentTarget));
407 // FIXME: We break the loop pass pipeline here in order to do full
408 // simplify-cfg. Eventually loop-simplifycfg should be enhanced to replace the
409 // need for this.
410 MPM.add(createCFGSimplificationPass());
411 addInstructionCombiningPass(MPM);
412 // We resume loop passes creating a second loop pipeline here.
413 MPM.add(createIndVarSimplifyPass()); // Canonicalize indvars
414 MPM.add(createLoopIdiomPass()); // Recognize idioms like memset.
415 addExtensionsToPM(EP_LateLoopOptimizations, MPM);
416 MPM.add(createLoopDeletionPass()); // Delete dead loops
417
418 if (EnableLoopInterchange)
419 MPM.add(createLoopInterchangePass()); // Interchange loops
420
421 // Unroll small loops
422 MPM.add(createSimpleLoopUnrollPass(OptLevel, DisableUnrollLoops,
423 ForgetAllSCEVInLoopUnroll));
424 addExtensionsToPM(EP_LoopOptimizerEnd, MPM);
425 // This ends the loop pass pipelines.
426
427 if (OptLevel > 1) {
428 MPM.add(createMergedLoadStoreMotionPass()); // Merge ld/st in diamonds
429 MPM.add(NewGVN ? createNewGVNPass()
430 : createGVNPass(DisableGVNLoadPRE)); // Remove redundancies
431 }
432 MPM.add(createMemCpyOptPass()); // Remove memcpy / form memset
433 MPM.add(createSCCPPass()); // Constant prop with SCCP
434
435 // Delete dead bit computations (instcombine runs after to fold away the dead
436 // computations, and then ADCE will run later to exploit any new DCE
437 // opportunities that creates).
438 MPM.add(createBitTrackingDCEPass()); // Delete dead bit computations
439
440 // Run instcombine after redundancy elimination to exploit opportunities
441 // opened up by them.
442 addInstructionCombiningPass(MPM);
443 addExtensionsToPM(EP_Peephole, MPM);
444 if (OptLevel > 1) {
445 MPM.add(createJumpThreadingPass()); // Thread jumps
446 MPM.add(createCorrelatedValuePropagationPass());
447 MPM.add(createDeadStoreEliminationPass()); // Delete dead stores
448 MPM.add(createLICMPass(LicmMssaOptCap, LicmMssaNoAccForPromotionCap));
449 }
450
451 addExtensionsToPM(EP_ScalarOptimizerLate, MPM);
452
453 if (RerollLoops)
454 MPM.add(createLoopRerollPass());
455
456 // TODO: Investigate if this is too expensive at O1.
457 MPM.add(createAggressiveDCEPass()); // Delete dead instructions
458 MPM.add(createCFGSimplificationPass()); // Merge & remove BBs
459 // Clean up after everything.
460 addInstructionCombiningPass(MPM);
461 addExtensionsToPM(EP_Peephole, MPM);
462
463 if (EnableCHR && OptLevel >= 3 &&
464 (!PGOInstrUse.empty() || !PGOSampleUse.empty() || EnablePGOCSInstrGen))
465 MPM.add(createControlHeightReductionLegacyPass());
466 }
467
populateModulePassManager(legacy::PassManagerBase & MPM)468 void PassManagerBuilder::populateModulePassManager(
469 legacy::PassManagerBase &MPM) {
470 // Whether this is a default or *LTO pre-link pipeline. The FullLTO post-link
471 // is handled separately, so just check this is not the ThinLTO post-link.
472 bool DefaultOrPreLinkPipeline = !PerformThinLTO;
473
474 if (!PGOSampleUse.empty()) {
475 MPM.add(createPruneEHPass());
476 // In ThinLTO mode, when flattened profile is used, all the available
477 // profile information will be annotated in PreLink phase so there is
478 // no need to load the profile again in PostLink.
479 if (!(FlattenedProfileUsed && PerformThinLTO))
480 MPM.add(createSampleProfileLoaderPass(PGOSampleUse));
481 }
482
483 // Allow forcing function attributes as a debugging and tuning aid.
484 MPM.add(createForceFunctionAttrsLegacyPass());
485
486 // If all optimizations are disabled, just run the always-inline pass and,
487 // if enabled, the function merging pass.
488 if (OptLevel == 0) {
489 addPGOInstrPasses(MPM);
490 if (Inliner) {
491 MPM.add(Inliner);
492 Inliner = nullptr;
493 }
494
495 // FIXME: The BarrierNoopPass is a HACK! The inliner pass above implicitly
496 // creates a CGSCC pass manager, but we don't want to add extensions into
497 // that pass manager. To prevent this we insert a no-op module pass to reset
498 // the pass manager to get the same behavior as EP_OptimizerLast in non-O0
499 // builds. The function merging pass is
500 if (MergeFunctions)
501 MPM.add(createMergeFunctionsPass());
502 else if (GlobalExtensionsNotEmpty() || !Extensions.empty())
503 MPM.add(createBarrierNoopPass());
504
505 if (PerformThinLTO) {
506 // Drop available_externally and unreferenced globals. This is necessary
507 // with ThinLTO in order to avoid leaving undefined references to dead
508 // globals in the object file.
509 MPM.add(createEliminateAvailableExternallyPass());
510 MPM.add(createGlobalDCEPass());
511 }
512
513 addExtensionsToPM(EP_EnabledOnOptLevel0, MPM);
514
515 if (PrepareForLTO || PrepareForThinLTO) {
516 MPM.add(createCanonicalizeAliasesPass());
517 // Rename anon globals to be able to export them in the summary.
518 // This has to be done after we add the extensions to the pass manager
519 // as there could be passes (e.g. Adddress sanitizer) which introduce
520 // new unnamed globals.
521 MPM.add(createNameAnonGlobalPass());
522 }
523 return;
524 }
525
526 // Add LibraryInfo if we have some.
527 if (LibraryInfo)
528 MPM.add(new TargetLibraryInfoWrapperPass(*LibraryInfo));
529
530 addInitialAliasAnalysisPasses(MPM);
531
532 // For ThinLTO there are two passes of indirect call promotion. The
533 // first is during the compile phase when PerformThinLTO=false and
534 // intra-module indirect call targets are promoted. The second is during
535 // the ThinLTO backend when PerformThinLTO=true, when we promote imported
536 // inter-module indirect calls. For that we perform indirect call promotion
537 // earlier in the pass pipeline, here before globalopt. Otherwise imported
538 // available_externally functions look unreferenced and are removed.
539 if (PerformThinLTO)
540 MPM.add(createPGOIndirectCallPromotionLegacyPass(/*InLTO = */ true,
541 !PGOSampleUse.empty()));
542
543 // For SamplePGO in ThinLTO compile phase, we do not want to unroll loops
544 // as it will change the CFG too much to make the 2nd profile annotation
545 // in backend more difficult.
546 bool PrepareForThinLTOUsingPGOSampleProfile =
547 PrepareForThinLTO && !PGOSampleUse.empty();
548 if (PrepareForThinLTOUsingPGOSampleProfile)
549 DisableUnrollLoops = true;
550
551 // Infer attributes about declarations if possible.
552 MPM.add(createInferFunctionAttrsLegacyPass());
553
554 addExtensionsToPM(EP_ModuleOptimizerEarly, MPM);
555
556 if (OptLevel > 2)
557 MPM.add(createCallSiteSplittingPass());
558
559 MPM.add(createIPSCCPPass()); // IP SCCP
560 MPM.add(createCalledValuePropagationPass());
561
562 // Infer attributes on declarations, call sites, arguments, etc.
563 MPM.add(createAttributorLegacyPass());
564
565 MPM.add(createGlobalOptimizerPass()); // Optimize out global vars
566 // Promote any localized global vars.
567 MPM.add(createPromoteMemoryToRegisterPass());
568
569 MPM.add(createDeadArgEliminationPass()); // Dead argument elimination
570
571 addInstructionCombiningPass(MPM); // Clean up after IPCP & DAE
572 addExtensionsToPM(EP_Peephole, MPM);
573 MPM.add(createCFGSimplificationPass()); // Clean up after IPCP & DAE
574
575 // For SamplePGO in ThinLTO compile phase, we do not want to do indirect
576 // call promotion as it will change the CFG too much to make the 2nd
577 // profile annotation in backend more difficult.
578 // PGO instrumentation is added during the compile phase for ThinLTO, do
579 // not run it a second time
580 if (DefaultOrPreLinkPipeline && !PrepareForThinLTOUsingPGOSampleProfile)
581 addPGOInstrPasses(MPM);
582
583 // Create profile COMDAT variables. Lld linker wants to see all variables
584 // before the LTO/ThinLTO link since it needs to resolve symbols/comdats.
585 if (!PerformThinLTO && EnablePGOCSInstrGen)
586 MPM.add(createPGOInstrumentationGenCreateVarLegacyPass(PGOInstrGen));
587
588 // We add a module alias analysis pass here. In part due to bugs in the
589 // analysis infrastructure this "works" in that the analysis stays alive
590 // for the entire SCC pass run below.
591 MPM.add(createGlobalsAAWrapperPass());
592
593 // Start of CallGraph SCC passes.
594 MPM.add(createPruneEHPass()); // Remove dead EH info
595 bool RunInliner = false;
596 if (Inliner) {
597 MPM.add(Inliner);
598 Inliner = nullptr;
599 RunInliner = true;
600 }
601
602 MPM.add(createPostOrderFunctionAttrsLegacyPass());
603 if (OptLevel > 2)
604 MPM.add(createArgumentPromotionPass()); // Scalarize uninlined fn args
605
606 addExtensionsToPM(EP_CGSCCOptimizerLate, MPM);
607 addFunctionSimplificationPasses(MPM);
608
609 // FIXME: This is a HACK! The inliner pass above implicitly creates a CGSCC
610 // pass manager that we are specifically trying to avoid. To prevent this
611 // we must insert a no-op module pass to reset the pass manager.
612 MPM.add(createBarrierNoopPass());
613
614 if (RunPartialInlining)
615 MPM.add(createPartialInliningPass());
616
617 if (OptLevel > 1 && !PrepareForLTO && !PrepareForThinLTO)
618 // Remove avail extern fns and globals definitions if we aren't
619 // compiling an object file for later LTO. For LTO we want to preserve
620 // these so they are eligible for inlining at link-time. Note if they
621 // are unreferenced they will be removed by GlobalDCE later, so
622 // this only impacts referenced available externally globals.
623 // Eventually they will be suppressed during codegen, but eliminating
624 // here enables more opportunity for GlobalDCE as it may make
625 // globals referenced by available external functions dead
626 // and saves running remaining passes on the eliminated functions.
627 MPM.add(createEliminateAvailableExternallyPass());
628
629 // CSFDO instrumentation and use pass. Don't invoke this for Prepare pass
630 // for LTO and ThinLTO -- The actual pass will be called after all inlines
631 // are performed.
632 // Need to do this after COMDAT variables have been eliminated,
633 // (i.e. after EliminateAvailableExternallyPass).
634 if (!(PrepareForLTO || PrepareForThinLTO))
635 addPGOInstrPasses(MPM, /* IsCS */ true);
636
637 if (EnableOrderFileInstrumentation)
638 MPM.add(createInstrOrderFilePass());
639
640 MPM.add(createReversePostOrderFunctionAttrsPass());
641
642 // The inliner performs some kind of dead code elimination as it goes,
643 // but there are cases that are not really caught by it. We might
644 // at some point consider teaching the inliner about them, but it
645 // is OK for now to run GlobalOpt + GlobalDCE in tandem as their
646 // benefits generally outweight the cost, making the whole pipeline
647 // faster.
648 if (RunInliner) {
649 MPM.add(createGlobalOptimizerPass());
650 MPM.add(createGlobalDCEPass());
651 }
652
653 // If we are planning to perform ThinLTO later, let's not bloat the code with
654 // unrolling/vectorization/... now. We'll first run the inliner + CGSCC passes
655 // during ThinLTO and perform the rest of the optimizations afterward.
656 if (PrepareForThinLTO) {
657 // Ensure we perform any last passes, but do so before renaming anonymous
658 // globals in case the passes add any.
659 addExtensionsToPM(EP_OptimizerLast, MPM);
660 MPM.add(createCanonicalizeAliasesPass());
661 // Rename anon globals to be able to export them in the summary.
662 MPM.add(createNameAnonGlobalPass());
663 return;
664 }
665
666 if (PerformThinLTO)
667 // Optimize globals now when performing ThinLTO, this enables more
668 // optimizations later.
669 MPM.add(createGlobalOptimizerPass());
670
671 // Scheduling LoopVersioningLICM when inlining is over, because after that
672 // we may see more accurate aliasing. Reason to run this late is that too
673 // early versioning may prevent further inlining due to increase of code
674 // size. By placing it just after inlining other optimizations which runs
675 // later might get benefit of no-alias assumption in clone loop.
676 if (UseLoopVersioningLICM) {
677 MPM.add(createLoopVersioningLICMPass()); // Do LoopVersioningLICM
678 MPM.add(createLICMPass(LicmMssaOptCap, LicmMssaNoAccForPromotionCap));
679 }
680
681 // We add a fresh GlobalsModRef run at this point. This is particularly
682 // useful as the above will have inlined, DCE'ed, and function-attr
683 // propagated everything. We should at this point have a reasonably minimal
684 // and richly annotated call graph. By computing aliasing and mod/ref
685 // information for all local globals here, the late loop passes and notably
686 // the vectorizer will be able to use them to help recognize vectorizable
687 // memory operations.
688 //
689 // Note that this relies on a bug in the pass manager which preserves
690 // a module analysis into a function pass pipeline (and throughout it) so
691 // long as the first function pass doesn't invalidate the module analysis.
692 // Thus both Float2Int and LoopRotate have to preserve AliasAnalysis for
693 // this to work. Fortunately, it is trivial to preserve AliasAnalysis
694 // (doing nothing preserves it as it is required to be conservatively
695 // correct in the face of IR changes).
696 MPM.add(createGlobalsAAWrapperPass());
697
698 MPM.add(createFloat2IntPass());
699 MPM.add(createLowerConstantIntrinsicsPass());
700
701 if (EnableMatrix) {
702 MPM.add(createLowerMatrixIntrinsicsPass());
703 // CSE the pointer arithmetic of the column vectors. This allows alias
704 // analysis to establish no-aliasing between loads and stores of different
705 // columns of the same matrix.
706 MPM.add(createEarlyCSEPass(false));
707 }
708
709 addExtensionsToPM(EP_VectorizerStart, MPM);
710
711 // Re-rotate loops in all our loop nests. These may have fallout out of
712 // rotated form due to GVN or other transformations, and the vectorizer relies
713 // on the rotated form. Disable header duplication at -Oz.
714 MPM.add(createLoopRotatePass(SizeLevel == 2 ? 0 : -1));
715
716 // Distribute loops to allow partial vectorization. I.e. isolate dependences
717 // into separate loop that would otherwise inhibit vectorization. This is
718 // currently only performed for loops marked with the metadata
719 // llvm.loop.distribute=true or when -enable-loop-distribute is specified.
720 MPM.add(createLoopDistributePass());
721
722 MPM.add(createLoopVectorizePass(!LoopsInterleaved, !LoopVectorize));
723
724 // Eliminate loads by forwarding stores from the previous iteration to loads
725 // of the current iteration.
726 MPM.add(createLoopLoadEliminationPass());
727
728 // FIXME: Because of #pragma vectorize enable, the passes below are always
729 // inserted in the pipeline, even when the vectorizer doesn't run (ex. when
730 // on -O1 and no #pragma is found). Would be good to have these two passes
731 // as function calls, so that we can only pass them when the vectorizer
732 // changed the code.
733 addInstructionCombiningPass(MPM);
734 if (OptLevel > 1 && ExtraVectorizerPasses) {
735 // At higher optimization levels, try to clean up any runtime overlap and
736 // alignment checks inserted by the vectorizer. We want to track correllated
737 // runtime checks for two inner loops in the same outer loop, fold any
738 // common computations, hoist loop-invariant aspects out of any outer loop,
739 // and unswitch the runtime checks if possible. Once hoisted, we may have
740 // dead (or speculatable) control flows or more combining opportunities.
741 MPM.add(createEarlyCSEPass());
742 MPM.add(createCorrelatedValuePropagationPass());
743 addInstructionCombiningPass(MPM);
744 MPM.add(createLICMPass(LicmMssaOptCap, LicmMssaNoAccForPromotionCap));
745 MPM.add(createLoopUnswitchPass(SizeLevel || OptLevel < 3, DivergentTarget));
746 MPM.add(createCFGSimplificationPass());
747 addInstructionCombiningPass(MPM);
748 }
749
750 // Cleanup after loop vectorization, etc. Simplification passes like CVP and
751 // GVN, loop transforms, and others have already run, so it's now better to
752 // convert to more optimized IR using more aggressive simplify CFG options.
753 // The extra sinking transform can create larger basic blocks, so do this
754 // before SLP vectorization.
755 MPM.add(createCFGSimplificationPass(1, true, true, false, true));
756
757 if (SLPVectorize) {
758 MPM.add(createSLPVectorizerPass()); // Vectorize parallel scalar chains.
759 if (OptLevel > 1 && ExtraVectorizerPasses) {
760 MPM.add(createEarlyCSEPass());
761 }
762 }
763
764 addExtensionsToPM(EP_Peephole, MPM);
765 addInstructionCombiningPass(MPM);
766
767 if (EnableUnrollAndJam && !DisableUnrollLoops) {
768 // Unroll and Jam. We do this before unroll but need to be in a separate
769 // loop pass manager in order for the outer loop to be processed by
770 // unroll and jam before the inner loop is unrolled.
771 MPM.add(createLoopUnrollAndJamPass(OptLevel));
772 }
773
774 // Unroll small loops
775 MPM.add(createLoopUnrollPass(OptLevel, DisableUnrollLoops,
776 ForgetAllSCEVInLoopUnroll));
777
778 if (!DisableUnrollLoops) {
779 // LoopUnroll may generate some redundency to cleanup.
780 addInstructionCombiningPass(MPM);
781
782 // Runtime unrolling will introduce runtime check in loop prologue. If the
783 // unrolled loop is a inner loop, then the prologue will be inside the
784 // outer loop. LICM pass can help to promote the runtime check out if the
785 // checked value is loop invariant.
786 MPM.add(createLICMPass(LicmMssaOptCap, LicmMssaNoAccForPromotionCap));
787 }
788
789 MPM.add(createWarnMissedTransformationsPass());
790
791 // After vectorization and unrolling, assume intrinsics may tell us more
792 // about pointer alignments.
793 MPM.add(createAlignmentFromAssumptionsPass());
794
795 // FIXME: We shouldn't bother with this anymore.
796 MPM.add(createStripDeadPrototypesPass()); // Get rid of dead prototypes
797
798 // GlobalOpt already deletes dead functions and globals, at -O2 try a
799 // late pass of GlobalDCE. It is capable of deleting dead cycles.
800 if (OptLevel > 1) {
801 MPM.add(createGlobalDCEPass()); // Remove dead fns and globals.
802 MPM.add(createConstantMergePass()); // Merge dup global constants
803 }
804
805 // See comment in the new PM for justification of scheduling splitting at
806 // this stage (\ref buildModuleSimplificationPipeline).
807 if (EnableHotColdSplit && !(PrepareForLTO || PrepareForThinLTO))
808 MPM.add(createHotColdSplittingPass());
809
810 if (MergeFunctions)
811 MPM.add(createMergeFunctionsPass());
812
813 // LoopSink pass sinks instructions hoisted by LICM, which serves as a
814 // canonicalization pass that enables other optimizations. As a result,
815 // LoopSink pass needs to be a very late IR pass to avoid undoing LICM
816 // result too early.
817 MPM.add(createLoopSinkPass());
818 // Get rid of LCSSA nodes.
819 MPM.add(createInstSimplifyLegacyPass());
820
821 // This hoists/decomposes div/rem ops. It should run after other sink/hoist
822 // passes to avoid re-sinking, but before SimplifyCFG because it can allow
823 // flattening of blocks.
824 MPM.add(createDivRemPairsPass());
825
826 // LoopSink (and other loop passes since the last simplifyCFG) might have
827 // resulted in single-entry-single-exit or empty blocks. Clean up the CFG.
828 MPM.add(createCFGSimplificationPass());
829
830 addExtensionsToPM(EP_OptimizerLast, MPM);
831
832 if (PrepareForLTO) {
833 MPM.add(createCanonicalizeAliasesPass());
834 // Rename anon globals to be able to handle them in the summary
835 MPM.add(createNameAnonGlobalPass());
836 }
837 }
838
addLTOOptimizationPasses(legacy::PassManagerBase & PM)839 void PassManagerBuilder::addLTOOptimizationPasses(legacy::PassManagerBase &PM) {
840 // Load sample profile before running the LTO optimization pipeline.
841 if (!PGOSampleUse.empty()) {
842 PM.add(createPruneEHPass());
843 PM.add(createSampleProfileLoaderPass(PGOSampleUse));
844 }
845
846 // Remove unused virtual tables to improve the quality of code generated by
847 // whole-program devirtualization and bitset lowering.
848 PM.add(createGlobalDCEPass());
849
850 // Provide AliasAnalysis services for optimizations.
851 addInitialAliasAnalysisPasses(PM);
852
853 // Allow forcing function attributes as a debugging and tuning aid.
854 PM.add(createForceFunctionAttrsLegacyPass());
855
856 // Infer attributes about declarations if possible.
857 PM.add(createInferFunctionAttrsLegacyPass());
858
859 if (OptLevel > 1) {
860 // Split call-site with more constrained arguments.
861 PM.add(createCallSiteSplittingPass());
862
863 // Indirect call promotion. This should promote all the targets that are
864 // left by the earlier promotion pass that promotes intra-module targets.
865 // This two-step promotion is to save the compile time. For LTO, it should
866 // produce the same result as if we only do promotion here.
867 PM.add(
868 createPGOIndirectCallPromotionLegacyPass(true, !PGOSampleUse.empty()));
869
870 // Propagate constants at call sites into the functions they call. This
871 // opens opportunities for globalopt (and inlining) by substituting function
872 // pointers passed as arguments to direct uses of functions.
873 PM.add(createIPSCCPPass());
874
875 // Attach metadata to indirect call sites indicating the set of functions
876 // they may target at run-time. This should follow IPSCCP.
877 PM.add(createCalledValuePropagationPass());
878
879 // Infer attributes on declarations, call sites, arguments, etc.
880 PM.add(createAttributorLegacyPass());
881 }
882
883 // Infer attributes about definitions. The readnone attribute in particular is
884 // required for virtual constant propagation.
885 PM.add(createPostOrderFunctionAttrsLegacyPass());
886 PM.add(createReversePostOrderFunctionAttrsPass());
887
888 // Split globals using inrange annotations on GEP indices. This can help
889 // improve the quality of generated code when virtual constant propagation or
890 // control flow integrity are enabled.
891 PM.add(createGlobalSplitPass());
892
893 // Apply whole-program devirtualization and virtual constant propagation.
894 PM.add(createWholeProgramDevirtPass(ExportSummary, nullptr));
895
896 // That's all we need at opt level 1.
897 if (OptLevel == 1)
898 return;
899
900 // Now that we internalized some globals, see if we can hack on them!
901 PM.add(createGlobalOptimizerPass());
902 // Promote any localized global vars.
903 PM.add(createPromoteMemoryToRegisterPass());
904
905 // Linking modules together can lead to duplicated global constants, only
906 // keep one copy of each constant.
907 PM.add(createConstantMergePass());
908
909 // Remove unused arguments from functions.
910 PM.add(createDeadArgEliminationPass());
911
912 // Reduce the code after globalopt and ipsccp. Both can open up significant
913 // simplification opportunities, and both can propagate functions through
914 // function pointers. When this happens, we often have to resolve varargs
915 // calls, etc, so let instcombine do this.
916 if (OptLevel > 2)
917 PM.add(createAggressiveInstCombinerPass());
918 addInstructionCombiningPass(PM);
919 addExtensionsToPM(EP_Peephole, PM);
920
921 // Inline small functions
922 bool RunInliner = Inliner;
923 if (RunInliner) {
924 PM.add(Inliner);
925 Inliner = nullptr;
926 }
927
928 PM.add(createPruneEHPass()); // Remove dead EH info.
929
930 // CSFDO instrumentation and use pass.
931 addPGOInstrPasses(PM, /* IsCS */ true);
932
933 // Optimize globals again if we ran the inliner.
934 if (RunInliner)
935 PM.add(createGlobalOptimizerPass());
936 PM.add(createGlobalDCEPass()); // Remove dead functions.
937
938 // If we didn't decide to inline a function, check to see if we can
939 // transform it to pass arguments by value instead of by reference.
940 PM.add(createArgumentPromotionPass());
941
942 // The IPO passes may leave cruft around. Clean up after them.
943 addInstructionCombiningPass(PM);
944 addExtensionsToPM(EP_Peephole, PM);
945 PM.add(createJumpThreadingPass());
946
947 // Break up allocas
948 PM.add(createSROAPass());
949
950 // LTO provides additional opportunities for tailcall elimination due to
951 // link-time inlining, and visibility of nocapture attribute.
952 if (OptLevel > 1)
953 PM.add(createTailCallEliminationPass());
954
955 // Infer attributes on declarations, call sites, arguments, etc.
956 PM.add(createPostOrderFunctionAttrsLegacyPass()); // Add nocapture.
957 // Run a few AA driven optimizations here and now, to cleanup the code.
958 PM.add(createGlobalsAAWrapperPass()); // IP alias analysis.
959
960 PM.add(createLICMPass(LicmMssaOptCap, LicmMssaNoAccForPromotionCap));
961 PM.add(createMergedLoadStoreMotionPass()); // Merge ld/st in diamonds.
962 PM.add(NewGVN ? createNewGVNPass()
963 : createGVNPass(DisableGVNLoadPRE)); // Remove redundancies.
964 PM.add(createMemCpyOptPass()); // Remove dead memcpys.
965
966 // Nuke dead stores.
967 PM.add(createDeadStoreEliminationPass());
968
969 // More loops are countable; try to optimize them.
970 PM.add(createIndVarSimplifyPass());
971 PM.add(createLoopDeletionPass());
972 if (EnableLoopInterchange)
973 PM.add(createLoopInterchangePass());
974
975 // Unroll small loops
976 PM.add(createSimpleLoopUnrollPass(OptLevel, DisableUnrollLoops,
977 ForgetAllSCEVInLoopUnroll));
978 PM.add(createLoopVectorizePass(true, !LoopVectorize));
979 // The vectorizer may have significantly shortened a loop body; unroll again.
980 PM.add(createLoopUnrollPass(OptLevel, DisableUnrollLoops,
981 ForgetAllSCEVInLoopUnroll));
982
983 PM.add(createWarnMissedTransformationsPass());
984
985 // Now that we've optimized loops (in particular loop induction variables),
986 // we may have exposed more scalar opportunities. Run parts of the scalar
987 // optimizer again at this point.
988 addInstructionCombiningPass(PM); // Initial cleanup
989 PM.add(createCFGSimplificationPass()); // if-convert
990 PM.add(createSCCPPass()); // Propagate exposed constants
991 addInstructionCombiningPass(PM); // Clean up again
992 PM.add(createBitTrackingDCEPass());
993
994 // More scalar chains could be vectorized due to more alias information
995 if (SLPVectorize)
996 PM.add(createSLPVectorizerPass()); // Vectorize parallel scalar chains.
997
998 // After vectorization, assume intrinsics may tell us more about pointer
999 // alignments.
1000 PM.add(createAlignmentFromAssumptionsPass());
1001
1002 // Cleanup and simplify the code after the scalar optimizations.
1003 addInstructionCombiningPass(PM);
1004 addExtensionsToPM(EP_Peephole, PM);
1005
1006 PM.add(createJumpThreadingPass());
1007 }
1008
addLateLTOOptimizationPasses(legacy::PassManagerBase & PM)1009 void PassManagerBuilder::addLateLTOOptimizationPasses(
1010 legacy::PassManagerBase &PM) {
1011 // See comment in the new PM for justification of scheduling splitting at
1012 // this stage (\ref buildLTODefaultPipeline).
1013 if (EnableHotColdSplit)
1014 PM.add(createHotColdSplittingPass());
1015
1016 // Delete basic blocks, which optimization passes may have killed.
1017 PM.add(createCFGSimplificationPass());
1018
1019 // Drop bodies of available externally objects to improve GlobalDCE.
1020 PM.add(createEliminateAvailableExternallyPass());
1021
1022 // Now that we have optimized the program, discard unreachable functions.
1023 PM.add(createGlobalDCEPass());
1024
1025 // FIXME: this is profitable (for compiler time) to do at -O0 too, but
1026 // currently it damages debug info.
1027 if (MergeFunctions)
1028 PM.add(createMergeFunctionsPass());
1029 }
1030
populateThinLTOPassManager(legacy::PassManagerBase & PM)1031 void PassManagerBuilder::populateThinLTOPassManager(
1032 legacy::PassManagerBase &PM) {
1033 PerformThinLTO = true;
1034 if (LibraryInfo)
1035 PM.add(new TargetLibraryInfoWrapperPass(*LibraryInfo));
1036
1037 if (VerifyInput)
1038 PM.add(createVerifierPass());
1039
1040 if (ImportSummary) {
1041 // These passes import type identifier resolutions for whole-program
1042 // devirtualization and CFI. They must run early because other passes may
1043 // disturb the specific instruction patterns that these passes look for,
1044 // creating dependencies on resolutions that may not appear in the summary.
1045 //
1046 // For example, GVN may transform the pattern assume(type.test) appearing in
1047 // two basic blocks into assume(phi(type.test, type.test)), which would
1048 // transform a dependency on a WPD resolution into a dependency on a type
1049 // identifier resolution for CFI.
1050 //
1051 // Also, WPD has access to more precise information than ICP and can
1052 // devirtualize more effectively, so it should operate on the IR first.
1053 PM.add(createWholeProgramDevirtPass(nullptr, ImportSummary));
1054 PM.add(createLowerTypeTestsPass(nullptr, ImportSummary));
1055 }
1056
1057 populateModulePassManager(PM);
1058
1059 if (VerifyOutput)
1060 PM.add(createVerifierPass());
1061 PerformThinLTO = false;
1062 }
1063
populateLTOPassManager(legacy::PassManagerBase & PM)1064 void PassManagerBuilder::populateLTOPassManager(legacy::PassManagerBase &PM) {
1065 if (LibraryInfo)
1066 PM.add(new TargetLibraryInfoWrapperPass(*LibraryInfo));
1067
1068 if (VerifyInput)
1069 PM.add(createVerifierPass());
1070
1071 addExtensionsToPM(EP_FullLinkTimeOptimizationEarly, PM);
1072
1073 if (OptLevel != 0)
1074 addLTOOptimizationPasses(PM);
1075 else {
1076 // The whole-program-devirt pass needs to run at -O0 because only it knows
1077 // about the llvm.type.checked.load intrinsic: it needs to both lower the
1078 // intrinsic itself and handle it in the summary.
1079 PM.add(createWholeProgramDevirtPass(ExportSummary, nullptr));
1080 }
1081
1082 // Create a function that performs CFI checks for cross-DSO calls with targets
1083 // in the current module.
1084 PM.add(createCrossDSOCFIPass());
1085
1086 // Lower type metadata and the type.test intrinsic. This pass supports Clang's
1087 // control flow integrity mechanisms (-fsanitize=cfi*) and needs to run at
1088 // link time if CFI is enabled. The pass does nothing if CFI is disabled.
1089 PM.add(createLowerTypeTestsPass(ExportSummary, nullptr));
1090
1091 if (OptLevel != 0)
1092 addLateLTOOptimizationPasses(PM);
1093
1094 addExtensionsToPM(EP_FullLinkTimeOptimizationLast, PM);
1095
1096 if (VerifyOutput)
1097 PM.add(createVerifierPass());
1098 }
1099
unwrap(LLVMPassManagerBuilderRef P)1100 inline PassManagerBuilder *unwrap(LLVMPassManagerBuilderRef P) {
1101 return reinterpret_cast<PassManagerBuilder*>(P);
1102 }
1103
wrap(PassManagerBuilder * P)1104 inline LLVMPassManagerBuilderRef wrap(PassManagerBuilder *P) {
1105 return reinterpret_cast<LLVMPassManagerBuilderRef>(P);
1106 }
1107
LLVMPassManagerBuilderCreate()1108 LLVMPassManagerBuilderRef LLVMPassManagerBuilderCreate() {
1109 PassManagerBuilder *PMB = new PassManagerBuilder();
1110 return wrap(PMB);
1111 }
1112
LLVMPassManagerBuilderDispose(LLVMPassManagerBuilderRef PMB)1113 void LLVMPassManagerBuilderDispose(LLVMPassManagerBuilderRef PMB) {
1114 PassManagerBuilder *Builder = unwrap(PMB);
1115 delete Builder;
1116 }
1117
1118 void
LLVMPassManagerBuilderSetOptLevel(LLVMPassManagerBuilderRef PMB,unsigned OptLevel)1119 LLVMPassManagerBuilderSetOptLevel(LLVMPassManagerBuilderRef PMB,
1120 unsigned OptLevel) {
1121 PassManagerBuilder *Builder = unwrap(PMB);
1122 Builder->OptLevel = OptLevel;
1123 }
1124
1125 void
LLVMPassManagerBuilderSetSizeLevel(LLVMPassManagerBuilderRef PMB,unsigned SizeLevel)1126 LLVMPassManagerBuilderSetSizeLevel(LLVMPassManagerBuilderRef PMB,
1127 unsigned SizeLevel) {
1128 PassManagerBuilder *Builder = unwrap(PMB);
1129 Builder->SizeLevel = SizeLevel;
1130 }
1131
1132 void
LLVMPassManagerBuilderSetDisableUnitAtATime(LLVMPassManagerBuilderRef PMB,LLVMBool Value)1133 LLVMPassManagerBuilderSetDisableUnitAtATime(LLVMPassManagerBuilderRef PMB,
1134 LLVMBool Value) {
1135 // NOTE: The DisableUnitAtATime switch has been removed.
1136 }
1137
1138 void
LLVMPassManagerBuilderSetDisableUnrollLoops(LLVMPassManagerBuilderRef PMB,LLVMBool Value)1139 LLVMPassManagerBuilderSetDisableUnrollLoops(LLVMPassManagerBuilderRef PMB,
1140 LLVMBool Value) {
1141 PassManagerBuilder *Builder = unwrap(PMB);
1142 Builder->DisableUnrollLoops = Value;
1143 }
1144
1145 void
LLVMPassManagerBuilderSetDisableSimplifyLibCalls(LLVMPassManagerBuilderRef PMB,LLVMBool Value)1146 LLVMPassManagerBuilderSetDisableSimplifyLibCalls(LLVMPassManagerBuilderRef PMB,
1147 LLVMBool Value) {
1148 // NOTE: The simplify-libcalls pass has been removed.
1149 }
1150
1151 void
LLVMPassManagerBuilderUseInlinerWithThreshold(LLVMPassManagerBuilderRef PMB,unsigned Threshold)1152 LLVMPassManagerBuilderUseInlinerWithThreshold(LLVMPassManagerBuilderRef PMB,
1153 unsigned Threshold) {
1154 PassManagerBuilder *Builder = unwrap(PMB);
1155 Builder->Inliner = createFunctionInliningPass(Threshold);
1156 }
1157
1158 void
LLVMPassManagerBuilderPopulateFunctionPassManager(LLVMPassManagerBuilderRef PMB,LLVMPassManagerRef PM)1159 LLVMPassManagerBuilderPopulateFunctionPassManager(LLVMPassManagerBuilderRef PMB,
1160 LLVMPassManagerRef PM) {
1161 PassManagerBuilder *Builder = unwrap(PMB);
1162 legacy::FunctionPassManager *FPM = unwrap<legacy::FunctionPassManager>(PM);
1163 Builder->populateFunctionPassManager(*FPM);
1164 }
1165
1166 void
LLVMPassManagerBuilderPopulateModulePassManager(LLVMPassManagerBuilderRef PMB,LLVMPassManagerRef PM)1167 LLVMPassManagerBuilderPopulateModulePassManager(LLVMPassManagerBuilderRef PMB,
1168 LLVMPassManagerRef PM) {
1169 PassManagerBuilder *Builder = unwrap(PMB);
1170 legacy::PassManagerBase *MPM = unwrap(PM);
1171 Builder->populateModulePassManager(*MPM);
1172 }
1173
LLVMPassManagerBuilderPopulateLTOPassManager(LLVMPassManagerBuilderRef PMB,LLVMPassManagerRef PM,LLVMBool Internalize,LLVMBool RunInliner)1174 void LLVMPassManagerBuilderPopulateLTOPassManager(LLVMPassManagerBuilderRef PMB,
1175 LLVMPassManagerRef PM,
1176 LLVMBool Internalize,
1177 LLVMBool RunInliner) {
1178 PassManagerBuilder *Builder = unwrap(PMB);
1179 legacy::PassManagerBase *LPM = unwrap(PM);
1180
1181 // A small backwards compatibility hack. populateLTOPassManager used to take
1182 // an RunInliner option.
1183 if (RunInliner && !Builder->Inliner)
1184 Builder->Inliner = createFunctionInliningPass();
1185
1186 Builder->populateLTOPassManager(*LPM);
1187 }
1188