1 //===- SampleProfile.cpp - Incorporate sample profiles into the IR --------===//
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 implements the SampleProfileLoader transformation. This pass
10 // reads a profile file generated by a sampling profiler (e.g. Linux Perf -
11 // http://perf.wiki.kernel.org/) and generates IR metadata to reflect the
12 // profile information in the given profile.
13 //
14 // This pass generates branch weight annotations on the IR:
15 //
16 // - prof: Represents branch weights. This annotation is added to branches
17 // to indicate the weights of each edge coming out of the branch.
18 // The weight of each edge is the weight of the target block for
19 // that edge. The weight of a block B is computed as the maximum
20 // number of samples found in B.
21 //
22 //===----------------------------------------------------------------------===//
23
24 #include "llvm/Transforms/IPO/SampleProfile.h"
25 #include "llvm/ADT/ArrayRef.h"
26 #include "llvm/ADT/DenseMap.h"
27 #include "llvm/ADT/DenseSet.h"
28 #include "llvm/ADT/None.h"
29 #include "llvm/ADT/SCCIterator.h"
30 #include "llvm/ADT/SmallPtrSet.h"
31 #include "llvm/ADT/SmallSet.h"
32 #include "llvm/ADT/SmallVector.h"
33 #include "llvm/ADT/Statistic.h"
34 #include "llvm/ADT/StringMap.h"
35 #include "llvm/ADT/StringRef.h"
36 #include "llvm/ADT/Twine.h"
37 #include "llvm/Analysis/AssumptionCache.h"
38 #include "llvm/Analysis/CallGraph.h"
39 #include "llvm/Analysis/CallGraphSCCPass.h"
40 #include "llvm/Analysis/InlineCost.h"
41 #include "llvm/Analysis/LoopInfo.h"
42 #include "llvm/Analysis/OptimizationRemarkEmitter.h"
43 #include "llvm/Analysis/PostDominators.h"
44 #include "llvm/Analysis/ProfileSummaryInfo.h"
45 #include "llvm/Analysis/TargetTransformInfo.h"
46 #include "llvm/IR/BasicBlock.h"
47 #include "llvm/IR/CFG.h"
48 #include "llvm/IR/CallSite.h"
49 #include "llvm/IR/DebugInfoMetadata.h"
50 #include "llvm/IR/DebugLoc.h"
51 #include "llvm/IR/DiagnosticInfo.h"
52 #include "llvm/IR/Dominators.h"
53 #include "llvm/IR/Function.h"
54 #include "llvm/IR/GlobalValue.h"
55 #include "llvm/IR/InstrTypes.h"
56 #include "llvm/IR/Instruction.h"
57 #include "llvm/IR/Instructions.h"
58 #include "llvm/IR/IntrinsicInst.h"
59 #include "llvm/IR/LLVMContext.h"
60 #include "llvm/IR/MDBuilder.h"
61 #include "llvm/IR/Module.h"
62 #include "llvm/IR/PassManager.h"
63 #include "llvm/IR/ValueSymbolTable.h"
64 #include "llvm/InitializePasses.h"
65 #include "llvm/Pass.h"
66 #include "llvm/ProfileData/InstrProf.h"
67 #include "llvm/ProfileData/SampleProf.h"
68 #include "llvm/ProfileData/SampleProfReader.h"
69 #include "llvm/Support/Casting.h"
70 #include "llvm/Support/CommandLine.h"
71 #include "llvm/Support/Debug.h"
72 #include "llvm/Support/ErrorHandling.h"
73 #include "llvm/Support/ErrorOr.h"
74 #include "llvm/Support/GenericDomTree.h"
75 #include "llvm/Support/raw_ostream.h"
76 #include "llvm/Transforms/IPO.h"
77 #include "llvm/Transforms/Instrumentation.h"
78 #include "llvm/Transforms/Utils/CallPromotionUtils.h"
79 #include "llvm/Transforms/Utils/Cloning.h"
80 #include "llvm/Transforms/Utils/MisExpect.h"
81 #include <algorithm>
82 #include <cassert>
83 #include <cstdint>
84 #include <functional>
85 #include <limits>
86 #include <map>
87 #include <memory>
88 #include <queue>
89 #include <string>
90 #include <system_error>
91 #include <utility>
92 #include <vector>
93
94 using namespace llvm;
95 using namespace sampleprof;
96 using ProfileCount = Function::ProfileCount;
97 #define DEBUG_TYPE "sample-profile"
98 #define CSINLINE_DEBUG DEBUG_TYPE "-inline"
99
100 STATISTIC(NumCSInlined,
101 "Number of functions inlined with context sensitive profile");
102 STATISTIC(NumCSNotInlined,
103 "Number of functions not inlined with context sensitive profile");
104
105 // Command line option to specify the file to read samples from. This is
106 // mainly used for debugging.
107 static cl::opt<std::string> SampleProfileFile(
108 "sample-profile-file", cl::init(""), cl::value_desc("filename"),
109 cl::desc("Profile file loaded by -sample-profile"), cl::Hidden);
110
111 // The named file contains a set of transformations that may have been applied
112 // to the symbol names between the program from which the sample data was
113 // collected and the current program's symbols.
114 static cl::opt<std::string> SampleProfileRemappingFile(
115 "sample-profile-remapping-file", cl::init(""), cl::value_desc("filename"),
116 cl::desc("Profile remapping file loaded by -sample-profile"), cl::Hidden);
117
118 static cl::opt<unsigned> SampleProfileMaxPropagateIterations(
119 "sample-profile-max-propagate-iterations", cl::init(100),
120 cl::desc("Maximum number of iterations to go through when propagating "
121 "sample block/edge weights through the CFG."));
122
123 static cl::opt<unsigned> SampleProfileRecordCoverage(
124 "sample-profile-check-record-coverage", cl::init(0), cl::value_desc("N"),
125 cl::desc("Emit a warning if less than N% of records in the input profile "
126 "are matched to the IR."));
127
128 static cl::opt<unsigned> SampleProfileSampleCoverage(
129 "sample-profile-check-sample-coverage", cl::init(0), cl::value_desc("N"),
130 cl::desc("Emit a warning if less than N% of samples in the input profile "
131 "are matched to the IR."));
132
133 static cl::opt<bool> NoWarnSampleUnused(
134 "no-warn-sample-unused", cl::init(false), cl::Hidden,
135 cl::desc("Use this option to turn off/on warnings about function with "
136 "samples but without debug information to use those samples. "));
137
138 static cl::opt<bool> ProfileSampleAccurate(
139 "profile-sample-accurate", cl::Hidden, cl::init(false),
140 cl::desc("If the sample profile is accurate, we will mark all un-sampled "
141 "callsite and function as having 0 samples. Otherwise, treat "
142 "un-sampled callsites and functions conservatively as unknown. "));
143
144 static cl::opt<bool> ProfileAccurateForSymsInList(
145 "profile-accurate-for-symsinlist", cl::Hidden, cl::ZeroOrMore,
146 cl::init(true),
147 cl::desc("For symbols in profile symbol list, regard their profiles to "
148 "be accurate. It may be overriden by profile-sample-accurate. "));
149
150 static cl::opt<bool> ProfileMergeInlinee(
151 "sample-profile-merge-inlinee", cl::Hidden, cl::init(false),
152 cl::desc("Merge past inlinee's profile to outline version if sample "
153 "profile loader decided not to inline a call site."));
154
155 static cl::opt<bool> ProfileTopDownLoad(
156 "sample-profile-top-down-load", cl::Hidden, cl::init(false),
157 cl::desc("Do profile annotation and inlining for functions in top-down "
158 "order of call graph during sample profile loading."));
159
160 static cl::opt<bool> ProfileSizeInline(
161 "sample-profile-inline-size", cl::Hidden, cl::init(false),
162 cl::desc("Inline cold call sites in profile loader if it's beneficial "
163 "for code size."));
164
165 static cl::opt<int> SampleColdCallSiteThreshold(
166 "sample-profile-cold-inline-threshold", cl::Hidden, cl::init(45),
167 cl::desc("Threshold for inlining cold callsites"));
168
169 namespace {
170
171 using BlockWeightMap = DenseMap<const BasicBlock *, uint64_t>;
172 using EquivalenceClassMap = DenseMap<const BasicBlock *, const BasicBlock *>;
173 using Edge = std::pair<const BasicBlock *, const BasicBlock *>;
174 using EdgeWeightMap = DenseMap<Edge, uint64_t>;
175 using BlockEdgeMap =
176 DenseMap<const BasicBlock *, SmallVector<const BasicBlock *, 8>>;
177
178 class SampleProfileLoader;
179
180 class SampleCoverageTracker {
181 public:
SampleCoverageTracker(SampleProfileLoader & SPL)182 SampleCoverageTracker(SampleProfileLoader &SPL) : SPLoader(SPL){};
183
184 bool markSamplesUsed(const FunctionSamples *FS, uint32_t LineOffset,
185 uint32_t Discriminator, uint64_t Samples);
186 unsigned computeCoverage(unsigned Used, unsigned Total) const;
187 unsigned countUsedRecords(const FunctionSamples *FS,
188 ProfileSummaryInfo *PSI) const;
189 unsigned countBodyRecords(const FunctionSamples *FS,
190 ProfileSummaryInfo *PSI) const;
getTotalUsedSamples() const191 uint64_t getTotalUsedSamples() const { return TotalUsedSamples; }
192 uint64_t countBodySamples(const FunctionSamples *FS,
193 ProfileSummaryInfo *PSI) const;
194
clear()195 void clear() {
196 SampleCoverage.clear();
197 TotalUsedSamples = 0;
198 }
199
200 private:
201 using BodySampleCoverageMap = std::map<LineLocation, unsigned>;
202 using FunctionSamplesCoverageMap =
203 DenseMap<const FunctionSamples *, BodySampleCoverageMap>;
204
205 /// Coverage map for sampling records.
206 ///
207 /// This map keeps a record of sampling records that have been matched to
208 /// an IR instruction. This is used to detect some form of staleness in
209 /// profiles (see flag -sample-profile-check-coverage).
210 ///
211 /// Each entry in the map corresponds to a FunctionSamples instance. This is
212 /// another map that counts how many times the sample record at the
213 /// given location has been used.
214 FunctionSamplesCoverageMap SampleCoverage;
215
216 /// Number of samples used from the profile.
217 ///
218 /// When a sampling record is used for the first time, the samples from
219 /// that record are added to this accumulator. Coverage is later computed
220 /// based on the total number of samples available in this function and
221 /// its callsites.
222 ///
223 /// Note that this accumulator tracks samples used from a single function
224 /// and all the inlined callsites. Strictly, we should have a map of counters
225 /// keyed by FunctionSamples pointers, but these stats are cleared after
226 /// every function, so we just need to keep a single counter.
227 uint64_t TotalUsedSamples = 0;
228
229 SampleProfileLoader &SPLoader;
230 };
231
232 class GUIDToFuncNameMapper {
233 public:
GUIDToFuncNameMapper(Module & M,SampleProfileReader & Reader,DenseMap<uint64_t,StringRef> & GUIDToFuncNameMap)234 GUIDToFuncNameMapper(Module &M, SampleProfileReader &Reader,
235 DenseMap<uint64_t, StringRef> &GUIDToFuncNameMap)
236 : CurrentReader(Reader), CurrentModule(M),
237 CurrentGUIDToFuncNameMap(GUIDToFuncNameMap) {
238 if (CurrentReader.getFormat() != SPF_Compact_Binary)
239 return;
240
241 for (const auto &F : CurrentModule) {
242 StringRef OrigName = F.getName();
243 CurrentGUIDToFuncNameMap.insert(
244 {Function::getGUID(OrigName), OrigName});
245
246 // Local to global var promotion used by optimization like thinlto
247 // will rename the var and add suffix like ".llvm.xxx" to the
248 // original local name. In sample profile, the suffixes of function
249 // names are all stripped. Since it is possible that the mapper is
250 // built in post-thin-link phase and var promotion has been done,
251 // we need to add the substring of function name without the suffix
252 // into the GUIDToFuncNameMap.
253 StringRef CanonName = FunctionSamples::getCanonicalFnName(F);
254 if (CanonName != OrigName)
255 CurrentGUIDToFuncNameMap.insert(
256 {Function::getGUID(CanonName), CanonName});
257 }
258
259 // Update GUIDToFuncNameMap for each function including inlinees.
260 SetGUIDToFuncNameMapForAll(&CurrentGUIDToFuncNameMap);
261 }
262
~GUIDToFuncNameMapper()263 ~GUIDToFuncNameMapper() {
264 if (CurrentReader.getFormat() != SPF_Compact_Binary)
265 return;
266
267 CurrentGUIDToFuncNameMap.clear();
268
269 // Reset GUIDToFuncNameMap for of each function as they're no
270 // longer valid at this point.
271 SetGUIDToFuncNameMapForAll(nullptr);
272 }
273
274 private:
SetGUIDToFuncNameMapForAll(DenseMap<uint64_t,StringRef> * Map)275 void SetGUIDToFuncNameMapForAll(DenseMap<uint64_t, StringRef> *Map) {
276 std::queue<FunctionSamples *> FSToUpdate;
277 for (auto &IFS : CurrentReader.getProfiles()) {
278 FSToUpdate.push(&IFS.second);
279 }
280
281 while (!FSToUpdate.empty()) {
282 FunctionSamples *FS = FSToUpdate.front();
283 FSToUpdate.pop();
284 FS->GUIDToFuncNameMap = Map;
285 for (const auto &ICS : FS->getCallsiteSamples()) {
286 const FunctionSamplesMap &FSMap = ICS.second;
287 for (auto &IFS : FSMap) {
288 FunctionSamples &FS = const_cast<FunctionSamples &>(IFS.second);
289 FSToUpdate.push(&FS);
290 }
291 }
292 }
293 }
294
295 SampleProfileReader &CurrentReader;
296 Module &CurrentModule;
297 DenseMap<uint64_t, StringRef> &CurrentGUIDToFuncNameMap;
298 };
299
300 /// Sample profile pass.
301 ///
302 /// This pass reads profile data from the file specified by
303 /// -sample-profile-file and annotates every affected function with the
304 /// profile information found in that file.
305 class SampleProfileLoader {
306 public:
SampleProfileLoader(StringRef Name,StringRef RemapName,bool IsThinLTOPreLink,std::function<AssumptionCache & (Function &)> GetAssumptionCache,std::function<TargetTransformInfo & (Function &)> GetTargetTransformInfo)307 SampleProfileLoader(
308 StringRef Name, StringRef RemapName, bool IsThinLTOPreLink,
309 std::function<AssumptionCache &(Function &)> GetAssumptionCache,
310 std::function<TargetTransformInfo &(Function &)> GetTargetTransformInfo)
311 : GetAC(std::move(GetAssumptionCache)),
312 GetTTI(std::move(GetTargetTransformInfo)), CoverageTracker(*this),
313 Filename(Name), RemappingFilename(RemapName),
314 IsThinLTOPreLink(IsThinLTOPreLink) {}
315
316 bool doInitialization(Module &M);
317 bool runOnModule(Module &M, ModuleAnalysisManager *AM,
318 ProfileSummaryInfo *_PSI, CallGraph *CG);
319
dump()320 void dump() { Reader->dump(); }
321
322 protected:
323 friend class SampleCoverageTracker;
324
325 bool runOnFunction(Function &F, ModuleAnalysisManager *AM);
326 unsigned getFunctionLoc(Function &F);
327 bool emitAnnotations(Function &F);
328 ErrorOr<uint64_t> getInstWeight(const Instruction &I);
329 ErrorOr<uint64_t> getBlockWeight(const BasicBlock *BB);
330 const FunctionSamples *findCalleeFunctionSamples(const Instruction &I) const;
331 std::vector<const FunctionSamples *>
332 findIndirectCallFunctionSamples(const Instruction &I, uint64_t &Sum) const;
333 mutable DenseMap<const DILocation *, const FunctionSamples *> DILocation2SampleMap;
334 const FunctionSamples *findFunctionSamples(const Instruction &I) const;
335 bool inlineCallInstruction(Instruction *I);
336 bool inlineHotFunctions(Function &F,
337 DenseSet<GlobalValue::GUID> &InlinedGUIDs);
338 // Inline cold/small functions in addition to hot ones
339 bool shouldInlineColdCallee(Instruction &CallInst);
340 void emitOptimizationRemarksForInlineCandidates(
341 const SmallVector<Instruction *, 10> &Candidates, const Function &F, bool Hot);
342 void printEdgeWeight(raw_ostream &OS, Edge E);
343 void printBlockWeight(raw_ostream &OS, const BasicBlock *BB) const;
344 void printBlockEquivalence(raw_ostream &OS, const BasicBlock *BB);
345 bool computeBlockWeights(Function &F);
346 void findEquivalenceClasses(Function &F);
347 template <bool IsPostDom>
348 void findEquivalencesFor(BasicBlock *BB1, ArrayRef<BasicBlock *> Descendants,
349 DominatorTreeBase<BasicBlock, IsPostDom> *DomTree);
350
351 void propagateWeights(Function &F);
352 uint64_t visitEdge(Edge E, unsigned *NumUnknownEdges, Edge *UnknownEdge);
353 void buildEdges(Function &F);
354 std::vector<Function *> buildFunctionOrder(Module &M, CallGraph *CG);
355 bool propagateThroughEdges(Function &F, bool UpdateBlockCount);
356 void computeDominanceAndLoopInfo(Function &F);
357 void clearFunctionData();
358 bool callsiteIsHot(const FunctionSamples *CallsiteFS,
359 ProfileSummaryInfo *PSI);
360
361 /// Map basic blocks to their computed weights.
362 ///
363 /// The weight of a basic block is defined to be the maximum
364 /// of all the instruction weights in that block.
365 BlockWeightMap BlockWeights;
366
367 /// Map edges to their computed weights.
368 ///
369 /// Edge weights are computed by propagating basic block weights in
370 /// SampleProfile::propagateWeights.
371 EdgeWeightMap EdgeWeights;
372
373 /// Set of visited blocks during propagation.
374 SmallPtrSet<const BasicBlock *, 32> VisitedBlocks;
375
376 /// Set of visited edges during propagation.
377 SmallSet<Edge, 32> VisitedEdges;
378
379 /// Equivalence classes for block weights.
380 ///
381 /// Two blocks BB1 and BB2 are in the same equivalence class if they
382 /// dominate and post-dominate each other, and they are in the same loop
383 /// nest. When this happens, the two blocks are guaranteed to execute
384 /// the same number of times.
385 EquivalenceClassMap EquivalenceClass;
386
387 /// Map from function name to Function *. Used to find the function from
388 /// the function name. If the function name contains suffix, additional
389 /// entry is added to map from the stripped name to the function if there
390 /// is one-to-one mapping.
391 StringMap<Function *> SymbolMap;
392
393 /// Dominance, post-dominance and loop information.
394 std::unique_ptr<DominatorTree> DT;
395 std::unique_ptr<PostDominatorTree> PDT;
396 std::unique_ptr<LoopInfo> LI;
397
398 std::function<AssumptionCache &(Function &)> GetAC;
399 std::function<TargetTransformInfo &(Function &)> GetTTI;
400
401 /// Predecessors for each basic block in the CFG.
402 BlockEdgeMap Predecessors;
403
404 /// Successors for each basic block in the CFG.
405 BlockEdgeMap Successors;
406
407 SampleCoverageTracker CoverageTracker;
408
409 /// Profile reader object.
410 std::unique_ptr<SampleProfileReader> Reader;
411
412 /// Samples collected for the body of this function.
413 FunctionSamples *Samples = nullptr;
414
415 /// Name of the profile file to load.
416 std::string Filename;
417
418 /// Name of the profile remapping file to load.
419 std::string RemappingFilename;
420
421 /// Flag indicating whether the profile input loaded successfully.
422 bool ProfileIsValid = false;
423
424 /// Flag indicating if the pass is invoked in ThinLTO compile phase.
425 ///
426 /// In this phase, in annotation, we should not promote indirect calls.
427 /// Instead, we will mark GUIDs that needs to be annotated to the function.
428 bool IsThinLTOPreLink;
429
430 /// Profile Summary Info computed from sample profile.
431 ProfileSummaryInfo *PSI = nullptr;
432
433 /// Profle Symbol list tells whether a function name appears in the binary
434 /// used to generate the current profile.
435 std::unique_ptr<ProfileSymbolList> PSL;
436
437 /// Total number of samples collected in this profile.
438 ///
439 /// This is the sum of all the samples collected in all the functions executed
440 /// at runtime.
441 uint64_t TotalCollectedSamples = 0;
442
443 /// Optimization Remark Emitter used to emit diagnostic remarks.
444 OptimizationRemarkEmitter *ORE = nullptr;
445
446 // Information recorded when we declined to inline a call site
447 // because we have determined it is too cold is accumulated for
448 // each callee function. Initially this is just the entry count.
449 struct NotInlinedProfileInfo {
450 uint64_t entryCount;
451 };
452 DenseMap<Function *, NotInlinedProfileInfo> notInlinedCallInfo;
453
454 // GUIDToFuncNameMap saves the mapping from GUID to the symbol name, for
455 // all the function symbols defined or declared in current module.
456 DenseMap<uint64_t, StringRef> GUIDToFuncNameMap;
457
458 // All the Names used in FunctionSamples including outline function
459 // names, inline instance names and call target names.
460 StringSet<> NamesInProfile;
461
462 // For symbol in profile symbol list, whether to regard their profiles
463 // to be accurate. It is mainly decided by existance of profile symbol
464 // list and -profile-accurate-for-symsinlist flag, but it can be
465 // overriden by -profile-sample-accurate or profile-sample-accurate
466 // attribute.
467 bool ProfAccForSymsInList;
468 };
469
470 class SampleProfileLoaderLegacyPass : public ModulePass {
471 public:
472 // Class identification, replacement for typeinfo
473 static char ID;
474
SampleProfileLoaderLegacyPass(StringRef Name=SampleProfileFile,bool IsThinLTOPreLink=false)475 SampleProfileLoaderLegacyPass(StringRef Name = SampleProfileFile,
476 bool IsThinLTOPreLink = false)
477 : ModulePass(ID),
478 SampleLoader(Name, SampleProfileRemappingFile, IsThinLTOPreLink,
479 [&](Function &F) -> AssumptionCache & {
480 return ACT->getAssumptionCache(F);
481 },
__anon332e6c090302(Function &F) 482 [&](Function &F) -> TargetTransformInfo & {
483 return TTIWP->getTTI(F);
484 }) {
485 initializeSampleProfileLoaderLegacyPassPass(
486 *PassRegistry::getPassRegistry());
487 }
488
dump()489 void dump() { SampleLoader.dump(); }
490
doInitialization(Module & M)491 bool doInitialization(Module &M) override {
492 return SampleLoader.doInitialization(M);
493 }
494
getPassName() const495 StringRef getPassName() const override { return "Sample profile pass"; }
496 bool runOnModule(Module &M) override;
497
getAnalysisUsage(AnalysisUsage & AU) const498 void getAnalysisUsage(AnalysisUsage &AU) const override {
499 AU.addRequired<AssumptionCacheTracker>();
500 AU.addRequired<TargetTransformInfoWrapperPass>();
501 AU.addRequired<ProfileSummaryInfoWrapperPass>();
502 }
503
504 private:
505 SampleProfileLoader SampleLoader;
506 AssumptionCacheTracker *ACT = nullptr;
507 TargetTransformInfoWrapperPass *TTIWP = nullptr;
508 };
509
510 } // end anonymous namespace
511
512 /// Return true if the given callsite is hot wrt to hot cutoff threshold.
513 ///
514 /// Functions that were inlined in the original binary will be represented
515 /// in the inline stack in the sample profile. If the profile shows that
516 /// the original inline decision was "good" (i.e., the callsite is executed
517 /// frequently), then we will recreate the inline decision and apply the
518 /// profile from the inlined callsite.
519 ///
520 /// To decide whether an inlined callsite is hot, we compare the callsite
521 /// sample count with the hot cutoff computed by ProfileSummaryInfo, it is
522 /// regarded as hot if the count is above the cutoff value.
523 ///
524 /// When ProfileAccurateForSymsInList is enabled and profile symbol list
525 /// is present, functions in the profile symbol list but without profile will
526 /// be regarded as cold and much less inlining will happen in CGSCC inlining
527 /// pass, so we tend to lower the hot criteria here to allow more early
528 /// inlining to happen for warm callsites and it is helpful for performance.
callsiteIsHot(const FunctionSamples * CallsiteFS,ProfileSummaryInfo * PSI)529 bool SampleProfileLoader::callsiteIsHot(const FunctionSamples *CallsiteFS,
530 ProfileSummaryInfo *PSI) {
531 if (!CallsiteFS)
532 return false; // The callsite was not inlined in the original binary.
533
534 assert(PSI && "PSI is expected to be non null");
535 uint64_t CallsiteTotalSamples = CallsiteFS->getTotalSamples();
536 if (ProfAccForSymsInList)
537 return !PSI->isColdCount(CallsiteTotalSamples);
538 else
539 return PSI->isHotCount(CallsiteTotalSamples);
540 }
541
542 /// Mark as used the sample record for the given function samples at
543 /// (LineOffset, Discriminator).
544 ///
545 /// \returns true if this is the first time we mark the given record.
markSamplesUsed(const FunctionSamples * FS,uint32_t LineOffset,uint32_t Discriminator,uint64_t Samples)546 bool SampleCoverageTracker::markSamplesUsed(const FunctionSamples *FS,
547 uint32_t LineOffset,
548 uint32_t Discriminator,
549 uint64_t Samples) {
550 LineLocation Loc(LineOffset, Discriminator);
551 unsigned &Count = SampleCoverage[FS][Loc];
552 bool FirstTime = (++Count == 1);
553 if (FirstTime)
554 TotalUsedSamples += Samples;
555 return FirstTime;
556 }
557
558 /// Return the number of sample records that were applied from this profile.
559 ///
560 /// This count does not include records from cold inlined callsites.
561 unsigned
countUsedRecords(const FunctionSamples * FS,ProfileSummaryInfo * PSI) const562 SampleCoverageTracker::countUsedRecords(const FunctionSamples *FS,
563 ProfileSummaryInfo *PSI) const {
564 auto I = SampleCoverage.find(FS);
565
566 // The size of the coverage map for FS represents the number of records
567 // that were marked used at least once.
568 unsigned Count = (I != SampleCoverage.end()) ? I->second.size() : 0;
569
570 // If there are inlined callsites in this function, count the samples found
571 // in the respective bodies. However, do not bother counting callees with 0
572 // total samples, these are callees that were never invoked at runtime.
573 for (const auto &I : FS->getCallsiteSamples())
574 for (const auto &J : I.second) {
575 const FunctionSamples *CalleeSamples = &J.second;
576 if (SPLoader.callsiteIsHot(CalleeSamples, PSI))
577 Count += countUsedRecords(CalleeSamples, PSI);
578 }
579
580 return Count;
581 }
582
583 /// Return the number of sample records in the body of this profile.
584 ///
585 /// This count does not include records from cold inlined callsites.
586 unsigned
countBodyRecords(const FunctionSamples * FS,ProfileSummaryInfo * PSI) const587 SampleCoverageTracker::countBodyRecords(const FunctionSamples *FS,
588 ProfileSummaryInfo *PSI) const {
589 unsigned Count = FS->getBodySamples().size();
590
591 // Only count records in hot callsites.
592 for (const auto &I : FS->getCallsiteSamples())
593 for (const auto &J : I.second) {
594 const FunctionSamples *CalleeSamples = &J.second;
595 if (SPLoader.callsiteIsHot(CalleeSamples, PSI))
596 Count += countBodyRecords(CalleeSamples, PSI);
597 }
598
599 return Count;
600 }
601
602 /// Return the number of samples collected in the body of this profile.
603 ///
604 /// This count does not include samples from cold inlined callsites.
605 uint64_t
countBodySamples(const FunctionSamples * FS,ProfileSummaryInfo * PSI) const606 SampleCoverageTracker::countBodySamples(const FunctionSamples *FS,
607 ProfileSummaryInfo *PSI) const {
608 uint64_t Total = 0;
609 for (const auto &I : FS->getBodySamples())
610 Total += I.second.getSamples();
611
612 // Only count samples in hot callsites.
613 for (const auto &I : FS->getCallsiteSamples())
614 for (const auto &J : I.second) {
615 const FunctionSamples *CalleeSamples = &J.second;
616 if (SPLoader.callsiteIsHot(CalleeSamples, PSI))
617 Total += countBodySamples(CalleeSamples, PSI);
618 }
619
620 return Total;
621 }
622
623 /// Return the fraction of sample records used in this profile.
624 ///
625 /// The returned value is an unsigned integer in the range 0-100 indicating
626 /// the percentage of sample records that were used while applying this
627 /// profile to the associated function.
computeCoverage(unsigned Used,unsigned Total) const628 unsigned SampleCoverageTracker::computeCoverage(unsigned Used,
629 unsigned Total) const {
630 assert(Used <= Total &&
631 "number of used records cannot exceed the total number of records");
632 return Total > 0 ? Used * 100 / Total : 100;
633 }
634
635 /// Clear all the per-function data used to load samples and propagate weights.
clearFunctionData()636 void SampleProfileLoader::clearFunctionData() {
637 BlockWeights.clear();
638 EdgeWeights.clear();
639 VisitedBlocks.clear();
640 VisitedEdges.clear();
641 EquivalenceClass.clear();
642 DT = nullptr;
643 PDT = nullptr;
644 LI = nullptr;
645 Predecessors.clear();
646 Successors.clear();
647 CoverageTracker.clear();
648 }
649
650 #ifndef NDEBUG
651 /// Print the weight of edge \p E on stream \p OS.
652 ///
653 /// \param OS Stream to emit the output to.
654 /// \param E Edge to print.
printEdgeWeight(raw_ostream & OS,Edge E)655 void SampleProfileLoader::printEdgeWeight(raw_ostream &OS, Edge E) {
656 OS << "weight[" << E.first->getName() << "->" << E.second->getName()
657 << "]: " << EdgeWeights[E] << "\n";
658 }
659
660 /// Print the equivalence class of block \p BB on stream \p OS.
661 ///
662 /// \param OS Stream to emit the output to.
663 /// \param BB Block to print.
printBlockEquivalence(raw_ostream & OS,const BasicBlock * BB)664 void SampleProfileLoader::printBlockEquivalence(raw_ostream &OS,
665 const BasicBlock *BB) {
666 const BasicBlock *Equiv = EquivalenceClass[BB];
667 OS << "equivalence[" << BB->getName()
668 << "]: " << ((Equiv) ? EquivalenceClass[BB]->getName() : "NONE") << "\n";
669 }
670
671 /// Print the weight of block \p BB on stream \p OS.
672 ///
673 /// \param OS Stream to emit the output to.
674 /// \param BB Block to print.
printBlockWeight(raw_ostream & OS,const BasicBlock * BB) const675 void SampleProfileLoader::printBlockWeight(raw_ostream &OS,
676 const BasicBlock *BB) const {
677 const auto &I = BlockWeights.find(BB);
678 uint64_t W = (I == BlockWeights.end() ? 0 : I->second);
679 OS << "weight[" << BB->getName() << "]: " << W << "\n";
680 }
681 #endif
682
683 /// Get the weight for an instruction.
684 ///
685 /// The "weight" of an instruction \p Inst is the number of samples
686 /// collected on that instruction at runtime. To retrieve it, we
687 /// need to compute the line number of \p Inst relative to the start of its
688 /// function. We use HeaderLineno to compute the offset. We then
689 /// look up the samples collected for \p Inst using BodySamples.
690 ///
691 /// \param Inst Instruction to query.
692 ///
693 /// \returns the weight of \p Inst.
getInstWeight(const Instruction & Inst)694 ErrorOr<uint64_t> SampleProfileLoader::getInstWeight(const Instruction &Inst) {
695 const DebugLoc &DLoc = Inst.getDebugLoc();
696 if (!DLoc)
697 return std::error_code();
698
699 const FunctionSamples *FS = findFunctionSamples(Inst);
700 if (!FS)
701 return std::error_code();
702
703 // Ignore all intrinsics, phinodes and branch instructions.
704 // Branch and phinodes instruction usually contains debug info from sources outside of
705 // the residing basic block, thus we ignore them during annotation.
706 if (isa<BranchInst>(Inst) || isa<IntrinsicInst>(Inst) || isa<PHINode>(Inst))
707 return std::error_code();
708
709 // If a direct call/invoke instruction is inlined in profile
710 // (findCalleeFunctionSamples returns non-empty result), but not inlined here,
711 // it means that the inlined callsite has no sample, thus the call
712 // instruction should have 0 count.
713 if ((isa<CallInst>(Inst) || isa<InvokeInst>(Inst)) &&
714 !ImmutableCallSite(&Inst).isIndirectCall() &&
715 findCalleeFunctionSamples(Inst))
716 return 0;
717
718 const DILocation *DIL = DLoc;
719 uint32_t LineOffset = FunctionSamples::getOffset(DIL);
720 uint32_t Discriminator = DIL->getBaseDiscriminator();
721 ErrorOr<uint64_t> R = FS->findSamplesAt(LineOffset, Discriminator);
722 if (R) {
723 bool FirstMark =
724 CoverageTracker.markSamplesUsed(FS, LineOffset, Discriminator, R.get());
725 if (FirstMark) {
726 ORE->emit([&]() {
727 OptimizationRemarkAnalysis Remark(DEBUG_TYPE, "AppliedSamples", &Inst);
728 Remark << "Applied " << ore::NV("NumSamples", *R);
729 Remark << " samples from profile (offset: ";
730 Remark << ore::NV("LineOffset", LineOffset);
731 if (Discriminator) {
732 Remark << ".";
733 Remark << ore::NV("Discriminator", Discriminator);
734 }
735 Remark << ")";
736 return Remark;
737 });
738 }
739 LLVM_DEBUG(dbgs() << " " << DLoc.getLine() << "."
740 << DIL->getBaseDiscriminator() << ":" << Inst
741 << " (line offset: " << LineOffset << "."
742 << DIL->getBaseDiscriminator() << " - weight: " << R.get()
743 << ")\n");
744 }
745 return R;
746 }
747
748 /// Compute the weight of a basic block.
749 ///
750 /// The weight of basic block \p BB is the maximum weight of all the
751 /// instructions in BB.
752 ///
753 /// \param BB The basic block to query.
754 ///
755 /// \returns the weight for \p BB.
getBlockWeight(const BasicBlock * BB)756 ErrorOr<uint64_t> SampleProfileLoader::getBlockWeight(const BasicBlock *BB) {
757 uint64_t Max = 0;
758 bool HasWeight = false;
759 for (auto &I : BB->getInstList()) {
760 const ErrorOr<uint64_t> &R = getInstWeight(I);
761 if (R) {
762 Max = std::max(Max, R.get());
763 HasWeight = true;
764 }
765 }
766 return HasWeight ? ErrorOr<uint64_t>(Max) : std::error_code();
767 }
768
769 /// Compute and store the weights of every basic block.
770 ///
771 /// This populates the BlockWeights map by computing
772 /// the weights of every basic block in the CFG.
773 ///
774 /// \param F The function to query.
computeBlockWeights(Function & F)775 bool SampleProfileLoader::computeBlockWeights(Function &F) {
776 bool Changed = false;
777 LLVM_DEBUG(dbgs() << "Block weights\n");
778 for (const auto &BB : F) {
779 ErrorOr<uint64_t> Weight = getBlockWeight(&BB);
780 if (Weight) {
781 BlockWeights[&BB] = Weight.get();
782 VisitedBlocks.insert(&BB);
783 Changed = true;
784 }
785 LLVM_DEBUG(printBlockWeight(dbgs(), &BB));
786 }
787
788 return Changed;
789 }
790
791 /// Get the FunctionSamples for a call instruction.
792 ///
793 /// The FunctionSamples of a call/invoke instruction \p Inst is the inlined
794 /// instance in which that call instruction is calling to. It contains
795 /// all samples that resides in the inlined instance. We first find the
796 /// inlined instance in which the call instruction is from, then we
797 /// traverse its children to find the callsite with the matching
798 /// location.
799 ///
800 /// \param Inst Call/Invoke instruction to query.
801 ///
802 /// \returns The FunctionSamples pointer to the inlined instance.
803 const FunctionSamples *
findCalleeFunctionSamples(const Instruction & Inst) const804 SampleProfileLoader::findCalleeFunctionSamples(const Instruction &Inst) const {
805 const DILocation *DIL = Inst.getDebugLoc();
806 if (!DIL) {
807 return nullptr;
808 }
809
810 StringRef CalleeName;
811 if (const CallInst *CI = dyn_cast<CallInst>(&Inst))
812 if (Function *Callee = CI->getCalledFunction())
813 CalleeName = Callee->getName();
814
815 const FunctionSamples *FS = findFunctionSamples(Inst);
816 if (FS == nullptr)
817 return nullptr;
818
819 return FS->findFunctionSamplesAt(LineLocation(FunctionSamples::getOffset(DIL),
820 DIL->getBaseDiscriminator()),
821 CalleeName);
822 }
823
824 /// Returns a vector of FunctionSamples that are the indirect call targets
825 /// of \p Inst. The vector is sorted by the total number of samples. Stores
826 /// the total call count of the indirect call in \p Sum.
827 std::vector<const FunctionSamples *>
findIndirectCallFunctionSamples(const Instruction & Inst,uint64_t & Sum) const828 SampleProfileLoader::findIndirectCallFunctionSamples(
829 const Instruction &Inst, uint64_t &Sum) const {
830 const DILocation *DIL = Inst.getDebugLoc();
831 std::vector<const FunctionSamples *> R;
832
833 if (!DIL) {
834 return R;
835 }
836
837 const FunctionSamples *FS = findFunctionSamples(Inst);
838 if (FS == nullptr)
839 return R;
840
841 uint32_t LineOffset = FunctionSamples::getOffset(DIL);
842 uint32_t Discriminator = DIL->getBaseDiscriminator();
843
844 auto T = FS->findCallTargetMapAt(LineOffset, Discriminator);
845 Sum = 0;
846 if (T)
847 for (const auto &T_C : T.get())
848 Sum += T_C.second;
849 if (const FunctionSamplesMap *M = FS->findFunctionSamplesMapAt(LineLocation(
850 FunctionSamples::getOffset(DIL), DIL->getBaseDiscriminator()))) {
851 if (M->empty())
852 return R;
853 for (const auto &NameFS : *M) {
854 Sum += NameFS.second.getEntrySamples();
855 R.push_back(&NameFS.second);
856 }
857 llvm::sort(R, [](const FunctionSamples *L, const FunctionSamples *R) {
858 if (L->getEntrySamples() != R->getEntrySamples())
859 return L->getEntrySamples() > R->getEntrySamples();
860 return FunctionSamples::getGUID(L->getName()) <
861 FunctionSamples::getGUID(R->getName());
862 });
863 }
864 return R;
865 }
866
867 /// Get the FunctionSamples for an instruction.
868 ///
869 /// The FunctionSamples of an instruction \p Inst is the inlined instance
870 /// in which that instruction is coming from. We traverse the inline stack
871 /// of that instruction, and match it with the tree nodes in the profile.
872 ///
873 /// \param Inst Instruction to query.
874 ///
875 /// \returns the FunctionSamples pointer to the inlined instance.
876 const FunctionSamples *
findFunctionSamples(const Instruction & Inst) const877 SampleProfileLoader::findFunctionSamples(const Instruction &Inst) const {
878 const DILocation *DIL = Inst.getDebugLoc();
879 if (!DIL)
880 return Samples;
881
882 auto it = DILocation2SampleMap.try_emplace(DIL,nullptr);
883 if (it.second)
884 it.first->second = Samples->findFunctionSamples(DIL);
885 return it.first->second;
886 }
887
inlineCallInstruction(Instruction * I)888 bool SampleProfileLoader::inlineCallInstruction(Instruction *I) {
889 assert(isa<CallInst>(I) || isa<InvokeInst>(I));
890 CallSite CS(I);
891 Function *CalledFunction = CS.getCalledFunction();
892 assert(CalledFunction);
893 DebugLoc DLoc = I->getDebugLoc();
894 BasicBlock *BB = I->getParent();
895 InlineParams Params = getInlineParams();
896 Params.ComputeFullInlineCost = true;
897 // Checks if there is anything in the reachable portion of the callee at
898 // this callsite that makes this inlining potentially illegal. Need to
899 // set ComputeFullInlineCost, otherwise getInlineCost may return early
900 // when cost exceeds threshold without checking all IRs in the callee.
901 // The acutal cost does not matter because we only checks isNever() to
902 // see if it is legal to inline the callsite.
903 InlineCost Cost =
904 getInlineCost(cast<CallBase>(*I), Params, GetTTI(*CalledFunction), GetAC,
905 None, nullptr, nullptr);
906 if (Cost.isNever()) {
907 ORE->emit(OptimizationRemarkAnalysis(CSINLINE_DEBUG, "InlineFail", DLoc, BB)
908 << "incompatible inlining");
909 return false;
910 }
911 InlineFunctionInfo IFI(nullptr, &GetAC);
912 if (InlineFunction(CS, IFI)) {
913 // The call to InlineFunction erases I, so we can't pass it here.
914 ORE->emit(OptimizationRemark(CSINLINE_DEBUG, "InlineSuccess", DLoc, BB)
915 << "inlined callee '" << ore::NV("Callee", CalledFunction)
916 << "' into '" << ore::NV("Caller", BB->getParent()) << "'");
917 return true;
918 }
919 return false;
920 }
921
shouldInlineColdCallee(Instruction & CallInst)922 bool SampleProfileLoader::shouldInlineColdCallee(Instruction &CallInst) {
923 if (!ProfileSizeInline)
924 return false;
925
926 Function *Callee = CallSite(&CallInst).getCalledFunction();
927 if (Callee == nullptr)
928 return false;
929
930 InlineCost Cost =
931 getInlineCost(cast<CallBase>(CallInst), getInlineParams(),
932 GetTTI(*Callee), GetAC, None, nullptr, nullptr);
933
934 return Cost.getCost() <= SampleColdCallSiteThreshold;
935 }
936
emitOptimizationRemarksForInlineCandidates(const SmallVector<Instruction *,10> & Candidates,const Function & F,bool Hot)937 void SampleProfileLoader::emitOptimizationRemarksForInlineCandidates(
938 const SmallVector<Instruction *, 10> &Candidates, const Function &F,
939 bool Hot) {
940 for (auto I : Candidates) {
941 Function *CalledFunction = CallSite(I).getCalledFunction();
942 if (CalledFunction) {
943 ORE->emit(OptimizationRemarkAnalysis(CSINLINE_DEBUG, "InlineAttempt",
944 I->getDebugLoc(), I->getParent())
945 << "previous inlining reattempted for "
946 << (Hot ? "hotness: '" : "size: '")
947 << ore::NV("Callee", CalledFunction) << "' into '"
948 << ore::NV("Caller", &F) << "'");
949 }
950 }
951 }
952
953 /// Iteratively inline hot callsites of a function.
954 ///
955 /// Iteratively traverse all callsites of the function \p F, and find if
956 /// the corresponding inlined instance exists and is hot in profile. If
957 /// it is hot enough, inline the callsites and adds new callsites of the
958 /// callee into the caller. If the call is an indirect call, first promote
959 /// it to direct call. Each indirect call is limited with a single target.
960 ///
961 /// \param F function to perform iterative inlining.
962 /// \param InlinedGUIDs a set to be updated to include all GUIDs that are
963 /// inlined in the profiled binary.
964 ///
965 /// \returns True if there is any inline happened.
inlineHotFunctions(Function & F,DenseSet<GlobalValue::GUID> & InlinedGUIDs)966 bool SampleProfileLoader::inlineHotFunctions(
967 Function &F, DenseSet<GlobalValue::GUID> &InlinedGUIDs) {
968 DenseSet<Instruction *> PromotedInsns;
969
970 // ProfAccForSymsInList is used in callsiteIsHot. The assertion makes sure
971 // Profile symbol list is ignored when profile-sample-accurate is on.
972 assert((!ProfAccForSymsInList ||
973 (!ProfileSampleAccurate &&
974 !F.hasFnAttribute("profile-sample-accurate"))) &&
975 "ProfAccForSymsInList should be false when profile-sample-accurate "
976 "is enabled");
977
978 DenseMap<Instruction *, const FunctionSamples *> localNotInlinedCallSites;
979 bool Changed = false;
980 while (true) {
981 bool LocalChanged = false;
982 SmallVector<Instruction *, 10> CIS;
983 for (auto &BB : F) {
984 bool Hot = false;
985 SmallVector<Instruction *, 10> AllCandidates;
986 SmallVector<Instruction *, 10> ColdCandidates;
987 for (auto &I : BB.getInstList()) {
988 const FunctionSamples *FS = nullptr;
989 if ((isa<CallInst>(I) || isa<InvokeInst>(I)) &&
990 !isa<IntrinsicInst>(I) && (FS = findCalleeFunctionSamples(I))) {
991 AllCandidates.push_back(&I);
992 if (FS->getEntrySamples() > 0)
993 localNotInlinedCallSites.try_emplace(&I, FS);
994 if (callsiteIsHot(FS, PSI))
995 Hot = true;
996 else if (shouldInlineColdCallee(I))
997 ColdCandidates.push_back(&I);
998 }
999 }
1000 if (Hot) {
1001 CIS.insert(CIS.begin(), AllCandidates.begin(), AllCandidates.end());
1002 emitOptimizationRemarksForInlineCandidates(AllCandidates, F, true);
1003 }
1004 else {
1005 CIS.insert(CIS.begin(), ColdCandidates.begin(), ColdCandidates.end());
1006 emitOptimizationRemarksForInlineCandidates(ColdCandidates, F, false);
1007 }
1008 }
1009 for (auto I : CIS) {
1010 Function *CalledFunction = CallSite(I).getCalledFunction();
1011 // Do not inline recursive calls.
1012 if (CalledFunction == &F)
1013 continue;
1014 if (CallSite(I).isIndirectCall()) {
1015 if (PromotedInsns.count(I))
1016 continue;
1017 uint64_t Sum;
1018 for (const auto *FS : findIndirectCallFunctionSamples(*I, Sum)) {
1019 if (IsThinLTOPreLink) {
1020 FS->findInlinedFunctions(InlinedGUIDs, F.getParent(),
1021 PSI->getOrCompHotCountThreshold());
1022 continue;
1023 }
1024 auto CalleeFunctionName = FS->getFuncNameInModule(F.getParent());
1025 // If it is a recursive call, we do not inline it as it could bloat
1026 // the code exponentially. There is way to better handle this, e.g.
1027 // clone the caller first, and inline the cloned caller if it is
1028 // recursive. As llvm does not inline recursive calls, we will
1029 // simply ignore it instead of handling it explicitly.
1030 if (CalleeFunctionName == F.getName())
1031 continue;
1032
1033 if (!callsiteIsHot(FS, PSI))
1034 continue;
1035
1036 const char *Reason = "Callee function not available";
1037 auto R = SymbolMap.find(CalleeFunctionName);
1038 if (R != SymbolMap.end() && R->getValue() &&
1039 !R->getValue()->isDeclaration() &&
1040 R->getValue()->getSubprogram() &&
1041 isLegalToPromote(CallSite(I), R->getValue(), &Reason)) {
1042 uint64_t C = FS->getEntrySamples();
1043 Instruction *DI =
1044 pgo::promoteIndirectCall(I, R->getValue(), C, Sum, false, ORE);
1045 Sum -= C;
1046 PromotedInsns.insert(I);
1047 // If profile mismatches, we should not attempt to inline DI.
1048 if ((isa<CallInst>(DI) || isa<InvokeInst>(DI)) &&
1049 inlineCallInstruction(DI)) {
1050 localNotInlinedCallSites.erase(I);
1051 LocalChanged = true;
1052 ++NumCSInlined;
1053 }
1054 } else {
1055 LLVM_DEBUG(dbgs()
1056 << "\nFailed to promote indirect call to "
1057 << CalleeFunctionName << " because " << Reason << "\n");
1058 }
1059 }
1060 } else if (CalledFunction && CalledFunction->getSubprogram() &&
1061 !CalledFunction->isDeclaration()) {
1062 if (inlineCallInstruction(I)) {
1063 localNotInlinedCallSites.erase(I);
1064 LocalChanged = true;
1065 ++NumCSInlined;
1066 }
1067 } else if (IsThinLTOPreLink) {
1068 findCalleeFunctionSamples(*I)->findInlinedFunctions(
1069 InlinedGUIDs, F.getParent(), PSI->getOrCompHotCountThreshold());
1070 }
1071 }
1072 if (LocalChanged) {
1073 Changed = true;
1074 } else {
1075 break;
1076 }
1077 }
1078
1079 // Accumulate not inlined callsite information into notInlinedSamples
1080 for (const auto &Pair : localNotInlinedCallSites) {
1081 Instruction *I = Pair.getFirst();
1082 Function *Callee = CallSite(I).getCalledFunction();
1083 if (!Callee || Callee->isDeclaration())
1084 continue;
1085
1086 ORE->emit(OptimizationRemarkAnalysis(CSINLINE_DEBUG, "NotInline",
1087 I->getDebugLoc(), I->getParent())
1088 << "previous inlining not repeated: '"
1089 << ore::NV("Callee", Callee) << "' into '"
1090 << ore::NV("Caller", &F) << "'");
1091
1092 ++NumCSNotInlined;
1093 const FunctionSamples *FS = Pair.getSecond();
1094 if (FS->getTotalSamples() == 0 && FS->getEntrySamples() == 0) {
1095 continue;
1096 }
1097
1098 if (ProfileMergeInlinee) {
1099 // Use entry samples as head samples during the merge, as inlinees
1100 // don't have head samples.
1101 assert(FS->getHeadSamples() == 0 && "Expect 0 head sample for inlinee");
1102 const_cast<FunctionSamples *>(FS)->addHeadSamples(FS->getEntrySamples());
1103
1104 // Note that we have to do the merge right after processing function.
1105 // This allows OutlineFS's profile to be used for annotation during
1106 // top-down processing of functions' annotation.
1107 FunctionSamples *OutlineFS = Reader->getOrCreateSamplesFor(*Callee);
1108 OutlineFS->merge(*FS);
1109 } else {
1110 auto pair =
1111 notInlinedCallInfo.try_emplace(Callee, NotInlinedProfileInfo{0});
1112 pair.first->second.entryCount += FS->getEntrySamples();
1113 }
1114 }
1115 return Changed;
1116 }
1117
1118 /// Find equivalence classes for the given block.
1119 ///
1120 /// This finds all the blocks that are guaranteed to execute the same
1121 /// number of times as \p BB1. To do this, it traverses all the
1122 /// descendants of \p BB1 in the dominator or post-dominator tree.
1123 ///
1124 /// A block BB2 will be in the same equivalence class as \p BB1 if
1125 /// the following holds:
1126 ///
1127 /// 1- \p BB1 is a descendant of BB2 in the opposite tree. So, if BB2
1128 /// is a descendant of \p BB1 in the dominator tree, then BB2 should
1129 /// dominate BB1 in the post-dominator tree.
1130 ///
1131 /// 2- Both BB2 and \p BB1 must be in the same loop.
1132 ///
1133 /// For every block BB2 that meets those two requirements, we set BB2's
1134 /// equivalence class to \p BB1.
1135 ///
1136 /// \param BB1 Block to check.
1137 /// \param Descendants Descendants of \p BB1 in either the dom or pdom tree.
1138 /// \param DomTree Opposite dominator tree. If \p Descendants is filled
1139 /// with blocks from \p BB1's dominator tree, then
1140 /// this is the post-dominator tree, and vice versa.
1141 template <bool IsPostDom>
findEquivalencesFor(BasicBlock * BB1,ArrayRef<BasicBlock * > Descendants,DominatorTreeBase<BasicBlock,IsPostDom> * DomTree)1142 void SampleProfileLoader::findEquivalencesFor(
1143 BasicBlock *BB1, ArrayRef<BasicBlock *> Descendants,
1144 DominatorTreeBase<BasicBlock, IsPostDom> *DomTree) {
1145 const BasicBlock *EC = EquivalenceClass[BB1];
1146 uint64_t Weight = BlockWeights[EC];
1147 for (const auto *BB2 : Descendants) {
1148 bool IsDomParent = DomTree->dominates(BB2, BB1);
1149 bool IsInSameLoop = LI->getLoopFor(BB1) == LI->getLoopFor(BB2);
1150 if (BB1 != BB2 && IsDomParent && IsInSameLoop) {
1151 EquivalenceClass[BB2] = EC;
1152 // If BB2 is visited, then the entire EC should be marked as visited.
1153 if (VisitedBlocks.count(BB2)) {
1154 VisitedBlocks.insert(EC);
1155 }
1156
1157 // If BB2 is heavier than BB1, make BB2 have the same weight
1158 // as BB1.
1159 //
1160 // Note that we don't worry about the opposite situation here
1161 // (when BB2 is lighter than BB1). We will deal with this
1162 // during the propagation phase. Right now, we just want to
1163 // make sure that BB1 has the largest weight of all the
1164 // members of its equivalence set.
1165 Weight = std::max(Weight, BlockWeights[BB2]);
1166 }
1167 }
1168 if (EC == &EC->getParent()->getEntryBlock()) {
1169 BlockWeights[EC] = Samples->getHeadSamples() + 1;
1170 } else {
1171 BlockWeights[EC] = Weight;
1172 }
1173 }
1174
1175 /// Find equivalence classes.
1176 ///
1177 /// Since samples may be missing from blocks, we can fill in the gaps by setting
1178 /// the weights of all the blocks in the same equivalence class to the same
1179 /// weight. To compute the concept of equivalence, we use dominance and loop
1180 /// information. Two blocks B1 and B2 are in the same equivalence class if B1
1181 /// dominates B2, B2 post-dominates B1 and both are in the same loop.
1182 ///
1183 /// \param F The function to query.
findEquivalenceClasses(Function & F)1184 void SampleProfileLoader::findEquivalenceClasses(Function &F) {
1185 SmallVector<BasicBlock *, 8> DominatedBBs;
1186 LLVM_DEBUG(dbgs() << "\nBlock equivalence classes\n");
1187 // Find equivalence sets based on dominance and post-dominance information.
1188 for (auto &BB : F) {
1189 BasicBlock *BB1 = &BB;
1190
1191 // Compute BB1's equivalence class once.
1192 if (EquivalenceClass.count(BB1)) {
1193 LLVM_DEBUG(printBlockEquivalence(dbgs(), BB1));
1194 continue;
1195 }
1196
1197 // By default, blocks are in their own equivalence class.
1198 EquivalenceClass[BB1] = BB1;
1199
1200 // Traverse all the blocks dominated by BB1. We are looking for
1201 // every basic block BB2 such that:
1202 //
1203 // 1- BB1 dominates BB2.
1204 // 2- BB2 post-dominates BB1.
1205 // 3- BB1 and BB2 are in the same loop nest.
1206 //
1207 // If all those conditions hold, it means that BB2 is executed
1208 // as many times as BB1, so they are placed in the same equivalence
1209 // class by making BB2's equivalence class be BB1.
1210 DominatedBBs.clear();
1211 DT->getDescendants(BB1, DominatedBBs);
1212 findEquivalencesFor(BB1, DominatedBBs, PDT.get());
1213
1214 LLVM_DEBUG(printBlockEquivalence(dbgs(), BB1));
1215 }
1216
1217 // Assign weights to equivalence classes.
1218 //
1219 // All the basic blocks in the same equivalence class will execute
1220 // the same number of times. Since we know that the head block in
1221 // each equivalence class has the largest weight, assign that weight
1222 // to all the blocks in that equivalence class.
1223 LLVM_DEBUG(
1224 dbgs() << "\nAssign the same weight to all blocks in the same class\n");
1225 for (auto &BI : F) {
1226 const BasicBlock *BB = &BI;
1227 const BasicBlock *EquivBB = EquivalenceClass[BB];
1228 if (BB != EquivBB)
1229 BlockWeights[BB] = BlockWeights[EquivBB];
1230 LLVM_DEBUG(printBlockWeight(dbgs(), BB));
1231 }
1232 }
1233
1234 /// Visit the given edge to decide if it has a valid weight.
1235 ///
1236 /// If \p E has not been visited before, we copy to \p UnknownEdge
1237 /// and increment the count of unknown edges.
1238 ///
1239 /// \param E Edge to visit.
1240 /// \param NumUnknownEdges Current number of unknown edges.
1241 /// \param UnknownEdge Set if E has not been visited before.
1242 ///
1243 /// \returns E's weight, if known. Otherwise, return 0.
visitEdge(Edge E,unsigned * NumUnknownEdges,Edge * UnknownEdge)1244 uint64_t SampleProfileLoader::visitEdge(Edge E, unsigned *NumUnknownEdges,
1245 Edge *UnknownEdge) {
1246 if (!VisitedEdges.count(E)) {
1247 (*NumUnknownEdges)++;
1248 *UnknownEdge = E;
1249 return 0;
1250 }
1251
1252 return EdgeWeights[E];
1253 }
1254
1255 /// Propagate weights through incoming/outgoing edges.
1256 ///
1257 /// If the weight of a basic block is known, and there is only one edge
1258 /// with an unknown weight, we can calculate the weight of that edge.
1259 ///
1260 /// Similarly, if all the edges have a known count, we can calculate the
1261 /// count of the basic block, if needed.
1262 ///
1263 /// \param F Function to process.
1264 /// \param UpdateBlockCount Whether we should update basic block counts that
1265 /// has already been annotated.
1266 ///
1267 /// \returns True if new weights were assigned to edges or blocks.
propagateThroughEdges(Function & F,bool UpdateBlockCount)1268 bool SampleProfileLoader::propagateThroughEdges(Function &F,
1269 bool UpdateBlockCount) {
1270 bool Changed = false;
1271 LLVM_DEBUG(dbgs() << "\nPropagation through edges\n");
1272 for (const auto &BI : F) {
1273 const BasicBlock *BB = &BI;
1274 const BasicBlock *EC = EquivalenceClass[BB];
1275
1276 // Visit all the predecessor and successor edges to determine
1277 // which ones have a weight assigned already. Note that it doesn't
1278 // matter that we only keep track of a single unknown edge. The
1279 // only case we are interested in handling is when only a single
1280 // edge is unknown (see setEdgeOrBlockWeight).
1281 for (unsigned i = 0; i < 2; i++) {
1282 uint64_t TotalWeight = 0;
1283 unsigned NumUnknownEdges = 0, NumTotalEdges = 0;
1284 Edge UnknownEdge, SelfReferentialEdge, SingleEdge;
1285
1286 if (i == 0) {
1287 // First, visit all predecessor edges.
1288 NumTotalEdges = Predecessors[BB].size();
1289 for (auto *Pred : Predecessors[BB]) {
1290 Edge E = std::make_pair(Pred, BB);
1291 TotalWeight += visitEdge(E, &NumUnknownEdges, &UnknownEdge);
1292 if (E.first == E.second)
1293 SelfReferentialEdge = E;
1294 }
1295 if (NumTotalEdges == 1) {
1296 SingleEdge = std::make_pair(Predecessors[BB][0], BB);
1297 }
1298 } else {
1299 // On the second round, visit all successor edges.
1300 NumTotalEdges = Successors[BB].size();
1301 for (auto *Succ : Successors[BB]) {
1302 Edge E = std::make_pair(BB, Succ);
1303 TotalWeight += visitEdge(E, &NumUnknownEdges, &UnknownEdge);
1304 }
1305 if (NumTotalEdges == 1) {
1306 SingleEdge = std::make_pair(BB, Successors[BB][0]);
1307 }
1308 }
1309
1310 // After visiting all the edges, there are three cases that we
1311 // can handle immediately:
1312 //
1313 // - All the edge weights are known (i.e., NumUnknownEdges == 0).
1314 // In this case, we simply check that the sum of all the edges
1315 // is the same as BB's weight. If not, we change BB's weight
1316 // to match. Additionally, if BB had not been visited before,
1317 // we mark it visited.
1318 //
1319 // - Only one edge is unknown and BB has already been visited.
1320 // In this case, we can compute the weight of the edge by
1321 // subtracting the total block weight from all the known
1322 // edge weights. If the edges weight more than BB, then the
1323 // edge of the last remaining edge is set to zero.
1324 //
1325 // - There exists a self-referential edge and the weight of BB is
1326 // known. In this case, this edge can be based on BB's weight.
1327 // We add up all the other known edges and set the weight on
1328 // the self-referential edge as we did in the previous case.
1329 //
1330 // In any other case, we must continue iterating. Eventually,
1331 // all edges will get a weight, or iteration will stop when
1332 // it reaches SampleProfileMaxPropagateIterations.
1333 if (NumUnknownEdges <= 1) {
1334 uint64_t &BBWeight = BlockWeights[EC];
1335 if (NumUnknownEdges == 0) {
1336 if (!VisitedBlocks.count(EC)) {
1337 // If we already know the weight of all edges, the weight of the
1338 // basic block can be computed. It should be no larger than the sum
1339 // of all edge weights.
1340 if (TotalWeight > BBWeight) {
1341 BBWeight = TotalWeight;
1342 Changed = true;
1343 LLVM_DEBUG(dbgs() << "All edge weights for " << BB->getName()
1344 << " known. Set weight for block: ";
1345 printBlockWeight(dbgs(), BB););
1346 }
1347 } else if (NumTotalEdges == 1 &&
1348 EdgeWeights[SingleEdge] < BlockWeights[EC]) {
1349 // If there is only one edge for the visited basic block, use the
1350 // block weight to adjust edge weight if edge weight is smaller.
1351 EdgeWeights[SingleEdge] = BlockWeights[EC];
1352 Changed = true;
1353 }
1354 } else if (NumUnknownEdges == 1 && VisitedBlocks.count(EC)) {
1355 // If there is a single unknown edge and the block has been
1356 // visited, then we can compute E's weight.
1357 if (BBWeight >= TotalWeight)
1358 EdgeWeights[UnknownEdge] = BBWeight - TotalWeight;
1359 else
1360 EdgeWeights[UnknownEdge] = 0;
1361 const BasicBlock *OtherEC;
1362 if (i == 0)
1363 OtherEC = EquivalenceClass[UnknownEdge.first];
1364 else
1365 OtherEC = EquivalenceClass[UnknownEdge.second];
1366 // Edge weights should never exceed the BB weights it connects.
1367 if (VisitedBlocks.count(OtherEC) &&
1368 EdgeWeights[UnknownEdge] > BlockWeights[OtherEC])
1369 EdgeWeights[UnknownEdge] = BlockWeights[OtherEC];
1370 VisitedEdges.insert(UnknownEdge);
1371 Changed = true;
1372 LLVM_DEBUG(dbgs() << "Set weight for edge: ";
1373 printEdgeWeight(dbgs(), UnknownEdge));
1374 }
1375 } else if (VisitedBlocks.count(EC) && BlockWeights[EC] == 0) {
1376 // If a block Weights 0, all its in/out edges should weight 0.
1377 if (i == 0) {
1378 for (auto *Pred : Predecessors[BB]) {
1379 Edge E = std::make_pair(Pred, BB);
1380 EdgeWeights[E] = 0;
1381 VisitedEdges.insert(E);
1382 }
1383 } else {
1384 for (auto *Succ : Successors[BB]) {
1385 Edge E = std::make_pair(BB, Succ);
1386 EdgeWeights[E] = 0;
1387 VisitedEdges.insert(E);
1388 }
1389 }
1390 } else if (SelfReferentialEdge.first && VisitedBlocks.count(EC)) {
1391 uint64_t &BBWeight = BlockWeights[BB];
1392 // We have a self-referential edge and the weight of BB is known.
1393 if (BBWeight >= TotalWeight)
1394 EdgeWeights[SelfReferentialEdge] = BBWeight - TotalWeight;
1395 else
1396 EdgeWeights[SelfReferentialEdge] = 0;
1397 VisitedEdges.insert(SelfReferentialEdge);
1398 Changed = true;
1399 LLVM_DEBUG(dbgs() << "Set self-referential edge weight to: ";
1400 printEdgeWeight(dbgs(), SelfReferentialEdge));
1401 }
1402 if (UpdateBlockCount && !VisitedBlocks.count(EC) && TotalWeight > 0) {
1403 BlockWeights[EC] = TotalWeight;
1404 VisitedBlocks.insert(EC);
1405 Changed = true;
1406 }
1407 }
1408 }
1409
1410 return Changed;
1411 }
1412
1413 /// Build in/out edge lists for each basic block in the CFG.
1414 ///
1415 /// We are interested in unique edges. If a block B1 has multiple
1416 /// edges to another block B2, we only add a single B1->B2 edge.
buildEdges(Function & F)1417 void SampleProfileLoader::buildEdges(Function &F) {
1418 for (auto &BI : F) {
1419 BasicBlock *B1 = &BI;
1420
1421 // Add predecessors for B1.
1422 SmallPtrSet<BasicBlock *, 16> Visited;
1423 if (!Predecessors[B1].empty())
1424 llvm_unreachable("Found a stale predecessors list in a basic block.");
1425 for (pred_iterator PI = pred_begin(B1), PE = pred_end(B1); PI != PE; ++PI) {
1426 BasicBlock *B2 = *PI;
1427 if (Visited.insert(B2).second)
1428 Predecessors[B1].push_back(B2);
1429 }
1430
1431 // Add successors for B1.
1432 Visited.clear();
1433 if (!Successors[B1].empty())
1434 llvm_unreachable("Found a stale successors list in a basic block.");
1435 for (succ_iterator SI = succ_begin(B1), SE = succ_end(B1); SI != SE; ++SI) {
1436 BasicBlock *B2 = *SI;
1437 if (Visited.insert(B2).second)
1438 Successors[B1].push_back(B2);
1439 }
1440 }
1441 }
1442
1443 /// Returns the sorted CallTargetMap \p M by count in descending order.
GetSortedValueDataFromCallTargets(const SampleRecord::CallTargetMap & M)1444 static SmallVector<InstrProfValueData, 2> GetSortedValueDataFromCallTargets(
1445 const SampleRecord::CallTargetMap & M) {
1446 SmallVector<InstrProfValueData, 2> R;
1447 for (const auto &I : SampleRecord::SortCallTargets(M)) {
1448 R.emplace_back(InstrProfValueData{FunctionSamples::getGUID(I.first), I.second});
1449 }
1450 return R;
1451 }
1452
1453 /// Propagate weights into edges
1454 ///
1455 /// The following rules are applied to every block BB in the CFG:
1456 ///
1457 /// - If BB has a single predecessor/successor, then the weight
1458 /// of that edge is the weight of the block.
1459 ///
1460 /// - If all incoming or outgoing edges are known except one, and the
1461 /// weight of the block is already known, the weight of the unknown
1462 /// edge will be the weight of the block minus the sum of all the known
1463 /// edges. If the sum of all the known edges is larger than BB's weight,
1464 /// we set the unknown edge weight to zero.
1465 ///
1466 /// - If there is a self-referential edge, and the weight of the block is
1467 /// known, the weight for that edge is set to the weight of the block
1468 /// minus the weight of the other incoming edges to that block (if
1469 /// known).
propagateWeights(Function & F)1470 void SampleProfileLoader::propagateWeights(Function &F) {
1471 bool Changed = true;
1472 unsigned I = 0;
1473
1474 // If BB weight is larger than its corresponding loop's header BB weight,
1475 // use the BB weight to replace the loop header BB weight.
1476 for (auto &BI : F) {
1477 BasicBlock *BB = &BI;
1478 Loop *L = LI->getLoopFor(BB);
1479 if (!L) {
1480 continue;
1481 }
1482 BasicBlock *Header = L->getHeader();
1483 if (Header && BlockWeights[BB] > BlockWeights[Header]) {
1484 BlockWeights[Header] = BlockWeights[BB];
1485 }
1486 }
1487
1488 // Before propagation starts, build, for each block, a list of
1489 // unique predecessors and successors. This is necessary to handle
1490 // identical edges in multiway branches. Since we visit all blocks and all
1491 // edges of the CFG, it is cleaner to build these lists once at the start
1492 // of the pass.
1493 buildEdges(F);
1494
1495 // Propagate until we converge or we go past the iteration limit.
1496 while (Changed && I++ < SampleProfileMaxPropagateIterations) {
1497 Changed = propagateThroughEdges(F, false);
1498 }
1499
1500 // The first propagation propagates BB counts from annotated BBs to unknown
1501 // BBs. The 2nd propagation pass resets edges weights, and use all BB weights
1502 // to propagate edge weights.
1503 VisitedEdges.clear();
1504 Changed = true;
1505 while (Changed && I++ < SampleProfileMaxPropagateIterations) {
1506 Changed = propagateThroughEdges(F, false);
1507 }
1508
1509 // The 3rd propagation pass allows adjust annotated BB weights that are
1510 // obviously wrong.
1511 Changed = true;
1512 while (Changed && I++ < SampleProfileMaxPropagateIterations) {
1513 Changed = propagateThroughEdges(F, true);
1514 }
1515
1516 // Generate MD_prof metadata for every branch instruction using the
1517 // edge weights computed during propagation.
1518 LLVM_DEBUG(dbgs() << "\nPropagation complete. Setting branch weights\n");
1519 LLVMContext &Ctx = F.getContext();
1520 MDBuilder MDB(Ctx);
1521 for (auto &BI : F) {
1522 BasicBlock *BB = &BI;
1523
1524 if (BlockWeights[BB]) {
1525 for (auto &I : BB->getInstList()) {
1526 if (!isa<CallInst>(I) && !isa<InvokeInst>(I))
1527 continue;
1528 CallSite CS(&I);
1529 if (!CS.getCalledFunction()) {
1530 const DebugLoc &DLoc = I.getDebugLoc();
1531 if (!DLoc)
1532 continue;
1533 const DILocation *DIL = DLoc;
1534 uint32_t LineOffset = FunctionSamples::getOffset(DIL);
1535 uint32_t Discriminator = DIL->getBaseDiscriminator();
1536
1537 const FunctionSamples *FS = findFunctionSamples(I);
1538 if (!FS)
1539 continue;
1540 auto T = FS->findCallTargetMapAt(LineOffset, Discriminator);
1541 if (!T || T.get().empty())
1542 continue;
1543 SmallVector<InstrProfValueData, 2> SortedCallTargets =
1544 GetSortedValueDataFromCallTargets(T.get());
1545 uint64_t Sum;
1546 findIndirectCallFunctionSamples(I, Sum);
1547 annotateValueSite(*I.getParent()->getParent()->getParent(), I,
1548 SortedCallTargets, Sum, IPVK_IndirectCallTarget,
1549 SortedCallTargets.size());
1550 } else if (!isa<IntrinsicInst>(&I)) {
1551 I.setMetadata(LLVMContext::MD_prof,
1552 MDB.createBranchWeights(
1553 {static_cast<uint32_t>(BlockWeights[BB])}));
1554 }
1555 }
1556 }
1557 Instruction *TI = BB->getTerminator();
1558 if (TI->getNumSuccessors() == 1)
1559 continue;
1560 if (!isa<BranchInst>(TI) && !isa<SwitchInst>(TI))
1561 continue;
1562
1563 DebugLoc BranchLoc = TI->getDebugLoc();
1564 LLVM_DEBUG(dbgs() << "\nGetting weights for branch at line "
1565 << ((BranchLoc) ? Twine(BranchLoc.getLine())
1566 : Twine("<UNKNOWN LOCATION>"))
1567 << ".\n");
1568 SmallVector<uint32_t, 4> Weights;
1569 uint32_t MaxWeight = 0;
1570 Instruction *MaxDestInst;
1571 for (unsigned I = 0; I < TI->getNumSuccessors(); ++I) {
1572 BasicBlock *Succ = TI->getSuccessor(I);
1573 Edge E = std::make_pair(BB, Succ);
1574 uint64_t Weight = EdgeWeights[E];
1575 LLVM_DEBUG(dbgs() << "\t"; printEdgeWeight(dbgs(), E));
1576 // Use uint32_t saturated arithmetic to adjust the incoming weights,
1577 // if needed. Sample counts in profiles are 64-bit unsigned values,
1578 // but internally branch weights are expressed as 32-bit values.
1579 if (Weight > std::numeric_limits<uint32_t>::max()) {
1580 LLVM_DEBUG(dbgs() << " (saturated due to uint32_t overflow)");
1581 Weight = std::numeric_limits<uint32_t>::max();
1582 }
1583 // Weight is added by one to avoid propagation errors introduced by
1584 // 0 weights.
1585 Weights.push_back(static_cast<uint32_t>(Weight + 1));
1586 if (Weight != 0) {
1587 if (Weight > MaxWeight) {
1588 MaxWeight = Weight;
1589 MaxDestInst = Succ->getFirstNonPHIOrDbgOrLifetime();
1590 }
1591 }
1592 }
1593
1594 misexpect::verifyMisExpect(TI, Weights, TI->getContext());
1595
1596 uint64_t TempWeight;
1597 // Only set weights if there is at least one non-zero weight.
1598 // In any other case, let the analyzer set weights.
1599 // Do not set weights if the weights are present. In ThinLTO, the profile
1600 // annotation is done twice. If the first annotation already set the
1601 // weights, the second pass does not need to set it.
1602 if (MaxWeight > 0 && !TI->extractProfTotalWeight(TempWeight)) {
1603 LLVM_DEBUG(dbgs() << "SUCCESS. Found non-zero weights.\n");
1604 TI->setMetadata(LLVMContext::MD_prof,
1605 MDB.createBranchWeights(Weights));
1606 ORE->emit([&]() {
1607 return OptimizationRemark(DEBUG_TYPE, "PopularDest", MaxDestInst)
1608 << "most popular destination for conditional branches at "
1609 << ore::NV("CondBranchesLoc", BranchLoc);
1610 });
1611 } else {
1612 LLVM_DEBUG(dbgs() << "SKIPPED. All branch weights are zero.\n");
1613 }
1614 }
1615 }
1616
1617 /// Get the line number for the function header.
1618 ///
1619 /// This looks up function \p F in the current compilation unit and
1620 /// retrieves the line number where the function is defined. This is
1621 /// line 0 for all the samples read from the profile file. Every line
1622 /// number is relative to this line.
1623 ///
1624 /// \param F Function object to query.
1625 ///
1626 /// \returns the line number where \p F is defined. If it returns 0,
1627 /// it means that there is no debug information available for \p F.
getFunctionLoc(Function & F)1628 unsigned SampleProfileLoader::getFunctionLoc(Function &F) {
1629 if (DISubprogram *S = F.getSubprogram())
1630 return S->getLine();
1631
1632 if (NoWarnSampleUnused)
1633 return 0;
1634
1635 // If the start of \p F is missing, emit a diagnostic to inform the user
1636 // about the missed opportunity.
1637 F.getContext().diagnose(DiagnosticInfoSampleProfile(
1638 "No debug information found in function " + F.getName() +
1639 ": Function profile not used",
1640 DS_Warning));
1641 return 0;
1642 }
1643
computeDominanceAndLoopInfo(Function & F)1644 void SampleProfileLoader::computeDominanceAndLoopInfo(Function &F) {
1645 DT.reset(new DominatorTree);
1646 DT->recalculate(F);
1647
1648 PDT.reset(new PostDominatorTree(F));
1649
1650 LI.reset(new LoopInfo);
1651 LI->analyze(*DT);
1652 }
1653
1654 /// Generate branch weight metadata for all branches in \p F.
1655 ///
1656 /// Branch weights are computed out of instruction samples using a
1657 /// propagation heuristic. Propagation proceeds in 3 phases:
1658 ///
1659 /// 1- Assignment of block weights. All the basic blocks in the function
1660 /// are initial assigned the same weight as their most frequently
1661 /// executed instruction.
1662 ///
1663 /// 2- Creation of equivalence classes. Since samples may be missing from
1664 /// blocks, we can fill in the gaps by setting the weights of all the
1665 /// blocks in the same equivalence class to the same weight. To compute
1666 /// the concept of equivalence, we use dominance and loop information.
1667 /// Two blocks B1 and B2 are in the same equivalence class if B1
1668 /// dominates B2, B2 post-dominates B1 and both are in the same loop.
1669 ///
1670 /// 3- Propagation of block weights into edges. This uses a simple
1671 /// propagation heuristic. The following rules are applied to every
1672 /// block BB in the CFG:
1673 ///
1674 /// - If BB has a single predecessor/successor, then the weight
1675 /// of that edge is the weight of the block.
1676 ///
1677 /// - If all the edges are known except one, and the weight of the
1678 /// block is already known, the weight of the unknown edge will
1679 /// be the weight of the block minus the sum of all the known
1680 /// edges. If the sum of all the known edges is larger than BB's weight,
1681 /// we set the unknown edge weight to zero.
1682 ///
1683 /// - If there is a self-referential edge, and the weight of the block is
1684 /// known, the weight for that edge is set to the weight of the block
1685 /// minus the weight of the other incoming edges to that block (if
1686 /// known).
1687 ///
1688 /// Since this propagation is not guaranteed to finalize for every CFG, we
1689 /// only allow it to proceed for a limited number of iterations (controlled
1690 /// by -sample-profile-max-propagate-iterations).
1691 ///
1692 /// FIXME: Try to replace this propagation heuristic with a scheme
1693 /// that is guaranteed to finalize. A work-list approach similar to
1694 /// the standard value propagation algorithm used by SSA-CCP might
1695 /// work here.
1696 ///
1697 /// Once all the branch weights are computed, we emit the MD_prof
1698 /// metadata on BB using the computed values for each of its branches.
1699 ///
1700 /// \param F The function to query.
1701 ///
1702 /// \returns true if \p F was modified. Returns false, otherwise.
emitAnnotations(Function & F)1703 bool SampleProfileLoader::emitAnnotations(Function &F) {
1704 bool Changed = false;
1705
1706 if (getFunctionLoc(F) == 0)
1707 return false;
1708
1709 LLVM_DEBUG(dbgs() << "Line number for the first instruction in "
1710 << F.getName() << ": " << getFunctionLoc(F) << "\n");
1711
1712 DenseSet<GlobalValue::GUID> InlinedGUIDs;
1713 Changed |= inlineHotFunctions(F, InlinedGUIDs);
1714
1715 // Compute basic block weights.
1716 Changed |= computeBlockWeights(F);
1717
1718 if (Changed) {
1719 // Add an entry count to the function using the samples gathered at the
1720 // function entry.
1721 // Sets the GUIDs that are inlined in the profiled binary. This is used
1722 // for ThinLink to make correct liveness analysis, and also make the IR
1723 // match the profiled binary before annotation.
1724 F.setEntryCount(
1725 ProfileCount(Samples->getHeadSamples() + 1, Function::PCT_Real),
1726 &InlinedGUIDs);
1727
1728 // Compute dominance and loop info needed for propagation.
1729 computeDominanceAndLoopInfo(F);
1730
1731 // Find equivalence classes.
1732 findEquivalenceClasses(F);
1733
1734 // Propagate weights to all edges.
1735 propagateWeights(F);
1736 }
1737
1738 // If coverage checking was requested, compute it now.
1739 if (SampleProfileRecordCoverage) {
1740 unsigned Used = CoverageTracker.countUsedRecords(Samples, PSI);
1741 unsigned Total = CoverageTracker.countBodyRecords(Samples, PSI);
1742 unsigned Coverage = CoverageTracker.computeCoverage(Used, Total);
1743 if (Coverage < SampleProfileRecordCoverage) {
1744 F.getContext().diagnose(DiagnosticInfoSampleProfile(
1745 F.getSubprogram()->getFilename(), getFunctionLoc(F),
1746 Twine(Used) + " of " + Twine(Total) + " available profile records (" +
1747 Twine(Coverage) + "%) were applied",
1748 DS_Warning));
1749 }
1750 }
1751
1752 if (SampleProfileSampleCoverage) {
1753 uint64_t Used = CoverageTracker.getTotalUsedSamples();
1754 uint64_t Total = CoverageTracker.countBodySamples(Samples, PSI);
1755 unsigned Coverage = CoverageTracker.computeCoverage(Used, Total);
1756 if (Coverage < SampleProfileSampleCoverage) {
1757 F.getContext().diagnose(DiagnosticInfoSampleProfile(
1758 F.getSubprogram()->getFilename(), getFunctionLoc(F),
1759 Twine(Used) + " of " + Twine(Total) + " available profile samples (" +
1760 Twine(Coverage) + "%) were applied",
1761 DS_Warning));
1762 }
1763 }
1764 return Changed;
1765 }
1766
1767 char SampleProfileLoaderLegacyPass::ID = 0;
1768
1769 INITIALIZE_PASS_BEGIN(SampleProfileLoaderLegacyPass, "sample-profile",
1770 "Sample Profile loader", false, false)
INITIALIZE_PASS_DEPENDENCY(AssumptionCacheTracker)1771 INITIALIZE_PASS_DEPENDENCY(AssumptionCacheTracker)
1772 INITIALIZE_PASS_DEPENDENCY(TargetTransformInfoWrapperPass)
1773 INITIALIZE_PASS_DEPENDENCY(ProfileSummaryInfoWrapperPass)
1774 INITIALIZE_PASS_END(SampleProfileLoaderLegacyPass, "sample-profile",
1775 "Sample Profile loader", false, false)
1776
1777 std::vector<Function *>
1778 SampleProfileLoader::buildFunctionOrder(Module &M, CallGraph *CG) {
1779 std::vector<Function *> FunctionOrderList;
1780 FunctionOrderList.reserve(M.size());
1781
1782 if (!ProfileTopDownLoad || CG == nullptr) {
1783 for (Function &F : M)
1784 if (!F.isDeclaration())
1785 FunctionOrderList.push_back(&F);
1786 return FunctionOrderList;
1787 }
1788
1789 assert(&CG->getModule() == &M);
1790 scc_iterator<CallGraph *> CGI = scc_begin(CG);
1791 while (!CGI.isAtEnd()) {
1792 for (CallGraphNode *node : *CGI) {
1793 auto F = node->getFunction();
1794 if (F && !F->isDeclaration())
1795 FunctionOrderList.push_back(F);
1796 }
1797 ++CGI;
1798 }
1799
1800 std::reverse(FunctionOrderList.begin(), FunctionOrderList.end());
1801 return FunctionOrderList;
1802 }
1803
doInitialization(Module & M)1804 bool SampleProfileLoader::doInitialization(Module &M) {
1805 auto &Ctx = M.getContext();
1806
1807 std::unique_ptr<SampleProfileReaderItaniumRemapper> RemapReader;
1808 auto ReaderOrErr =
1809 SampleProfileReader::create(Filename, Ctx, RemappingFilename);
1810 if (std::error_code EC = ReaderOrErr.getError()) {
1811 std::string Msg = "Could not open profile: " + EC.message();
1812 Ctx.diagnose(DiagnosticInfoSampleProfile(Filename, Msg));
1813 return false;
1814 }
1815 Reader = std::move(ReaderOrErr.get());
1816 Reader->collectFuncsFrom(M);
1817 ProfileIsValid = (Reader->read() == sampleprof_error::success);
1818 PSL = Reader->getProfileSymbolList();
1819
1820 // While profile-sample-accurate is on, ignore symbol list.
1821 ProfAccForSymsInList =
1822 ProfileAccurateForSymsInList && PSL && !ProfileSampleAccurate;
1823 if (ProfAccForSymsInList) {
1824 NamesInProfile.clear();
1825 if (auto NameTable = Reader->getNameTable())
1826 NamesInProfile.insert(NameTable->begin(), NameTable->end());
1827 }
1828
1829 return true;
1830 }
1831
createSampleProfileLoaderPass()1832 ModulePass *llvm::createSampleProfileLoaderPass() {
1833 return new SampleProfileLoaderLegacyPass();
1834 }
1835
createSampleProfileLoaderPass(StringRef Name)1836 ModulePass *llvm::createSampleProfileLoaderPass(StringRef Name) {
1837 return new SampleProfileLoaderLegacyPass(Name);
1838 }
1839
runOnModule(Module & M,ModuleAnalysisManager * AM,ProfileSummaryInfo * _PSI,CallGraph * CG)1840 bool SampleProfileLoader::runOnModule(Module &M, ModuleAnalysisManager *AM,
1841 ProfileSummaryInfo *_PSI, CallGraph *CG) {
1842 GUIDToFuncNameMapper Mapper(M, *Reader, GUIDToFuncNameMap);
1843 if (!ProfileIsValid)
1844 return false;
1845
1846 PSI = _PSI;
1847 if (M.getProfileSummary(/* IsCS */ false) == nullptr)
1848 M.setProfileSummary(Reader->getSummary().getMD(M.getContext()),
1849 ProfileSummary::PSK_Sample);
1850
1851 // Compute the total number of samples collected in this profile.
1852 for (const auto &I : Reader->getProfiles())
1853 TotalCollectedSamples += I.second.getTotalSamples();
1854
1855 // Populate the symbol map.
1856 for (const auto &N_F : M.getValueSymbolTable()) {
1857 StringRef OrigName = N_F.getKey();
1858 Function *F = dyn_cast<Function>(N_F.getValue());
1859 if (F == nullptr)
1860 continue;
1861 SymbolMap[OrigName] = F;
1862 auto pos = OrigName.find('.');
1863 if (pos != StringRef::npos) {
1864 StringRef NewName = OrigName.substr(0, pos);
1865 auto r = SymbolMap.insert(std::make_pair(NewName, F));
1866 // Failiing to insert means there is already an entry in SymbolMap,
1867 // thus there are multiple functions that are mapped to the same
1868 // stripped name. In this case of name conflicting, set the value
1869 // to nullptr to avoid confusion.
1870 if (!r.second)
1871 r.first->second = nullptr;
1872 }
1873 }
1874
1875 bool retval = false;
1876 for (auto F : buildFunctionOrder(M, CG)) {
1877 assert(!F->isDeclaration());
1878 clearFunctionData();
1879 retval |= runOnFunction(*F, AM);
1880 }
1881
1882 // Account for cold calls not inlined....
1883 for (const std::pair<Function *, NotInlinedProfileInfo> &pair :
1884 notInlinedCallInfo)
1885 updateProfileCallee(pair.first, pair.second.entryCount);
1886
1887 return retval;
1888 }
1889
runOnModule(Module & M)1890 bool SampleProfileLoaderLegacyPass::runOnModule(Module &M) {
1891 ACT = &getAnalysis<AssumptionCacheTracker>();
1892 TTIWP = &getAnalysis<TargetTransformInfoWrapperPass>();
1893 ProfileSummaryInfo *PSI =
1894 &getAnalysis<ProfileSummaryInfoWrapperPass>().getPSI();
1895 return SampleLoader.runOnModule(M, nullptr, PSI, nullptr);
1896 }
1897
runOnFunction(Function & F,ModuleAnalysisManager * AM)1898 bool SampleProfileLoader::runOnFunction(Function &F, ModuleAnalysisManager *AM) {
1899
1900 DILocation2SampleMap.clear();
1901 // By default the entry count is initialized to -1, which will be treated
1902 // conservatively by getEntryCount as the same as unknown (None). This is
1903 // to avoid newly added code to be treated as cold. If we have samples
1904 // this will be overwritten in emitAnnotations.
1905 uint64_t initialEntryCount = -1;
1906
1907 ProfAccForSymsInList = ProfileAccurateForSymsInList && PSL;
1908 if (ProfileSampleAccurate || F.hasFnAttribute("profile-sample-accurate")) {
1909 // initialize all the function entry counts to 0. It means all the
1910 // functions without profile will be regarded as cold.
1911 initialEntryCount = 0;
1912 // profile-sample-accurate is a user assertion which has a higher precedence
1913 // than symbol list. When profile-sample-accurate is on, ignore symbol list.
1914 ProfAccForSymsInList = false;
1915 }
1916
1917 // PSL -- profile symbol list include all the symbols in sampled binary.
1918 // If ProfileAccurateForSymsInList is enabled, PSL is used to treat
1919 // old functions without samples being cold, without having to worry
1920 // about new and hot functions being mistakenly treated as cold.
1921 if (ProfAccForSymsInList) {
1922 // Initialize the entry count to 0 for functions in the list.
1923 if (PSL->contains(F.getName()))
1924 initialEntryCount = 0;
1925
1926 // Function in the symbol list but without sample will be regarded as
1927 // cold. To minimize the potential negative performance impact it could
1928 // have, we want to be a little conservative here saying if a function
1929 // shows up in the profile, no matter as outline function, inline instance
1930 // or call targets, treat the function as not being cold. This will handle
1931 // the cases such as most callsites of a function are inlined in sampled
1932 // binary but not inlined in current build (because of source code drift,
1933 // imprecise debug information, or the callsites are all cold individually
1934 // but not cold accumulatively...), so the outline function showing up as
1935 // cold in sampled binary will actually not be cold after current build.
1936 StringRef CanonName = FunctionSamples::getCanonicalFnName(F);
1937 if (NamesInProfile.count(CanonName))
1938 initialEntryCount = -1;
1939 }
1940
1941 F.setEntryCount(ProfileCount(initialEntryCount, Function::PCT_Real));
1942 std::unique_ptr<OptimizationRemarkEmitter> OwnedORE;
1943 if (AM) {
1944 auto &FAM =
1945 AM->getResult<FunctionAnalysisManagerModuleProxy>(*F.getParent())
1946 .getManager();
1947 ORE = &FAM.getResult<OptimizationRemarkEmitterAnalysis>(F);
1948 } else {
1949 OwnedORE = std::make_unique<OptimizationRemarkEmitter>(&F);
1950 ORE = OwnedORE.get();
1951 }
1952 Samples = Reader->getSamplesFor(F);
1953 if (Samples && !Samples->empty())
1954 return emitAnnotations(F);
1955 return false;
1956 }
1957
run(Module & M,ModuleAnalysisManager & AM)1958 PreservedAnalyses SampleProfileLoaderPass::run(Module &M,
1959 ModuleAnalysisManager &AM) {
1960 FunctionAnalysisManager &FAM =
1961 AM.getResult<FunctionAnalysisManagerModuleProxy>(M).getManager();
1962
1963 auto GetAssumptionCache = [&](Function &F) -> AssumptionCache & {
1964 return FAM.getResult<AssumptionAnalysis>(F);
1965 };
1966 auto GetTTI = [&](Function &F) -> TargetTransformInfo & {
1967 return FAM.getResult<TargetIRAnalysis>(F);
1968 };
1969
1970 SampleProfileLoader SampleLoader(
1971 ProfileFileName.empty() ? SampleProfileFile : ProfileFileName,
1972 ProfileRemappingFileName.empty() ? SampleProfileRemappingFile
1973 : ProfileRemappingFileName,
1974 IsThinLTOPreLink, GetAssumptionCache, GetTTI);
1975
1976 if (!SampleLoader.doInitialization(M))
1977 return PreservedAnalyses::all();
1978
1979 ProfileSummaryInfo *PSI = &AM.getResult<ProfileSummaryAnalysis>(M);
1980 CallGraph &CG = AM.getResult<CallGraphAnalysis>(M);
1981 if (!SampleLoader.runOnModule(M, &AM, PSI, &CG))
1982 return PreservedAnalyses::all();
1983
1984 return PreservedAnalyses::none();
1985 }
1986