LLVM  8.0.0svn
SampleProfile.cpp
Go to the documentation of this file.
1 //===- SampleProfile.cpp - Incorporate sample profiles into the IR --------===//
2 //
3 // The LLVM Compiler Infrastructure
4 //
5 // This file is distributed under the University of Illinois Open Source
6 // License. See LICENSE.TXT for details.
7 //
8 //===----------------------------------------------------------------------===//
9 //
10 // This file implements the SampleProfileLoader transformation. This pass
11 // reads a profile file generated by a sampling profiler (e.g. Linux Perf -
12 // http://perf.wiki.kernel.org/) and generates IR metadata to reflect the
13 // profile information in the given profile.
14 //
15 // This pass generates branch weight annotations on the IR:
16 //
17 // - prof: Represents branch weights. This annotation is added to branches
18 // to indicate the weights of each edge coming out of the branch.
19 // The weight of each edge is the weight of the target block for
20 // that edge. The weight of a block B is computed as the maximum
21 // number of samples found in B.
22 //
23 //===----------------------------------------------------------------------===//
24 
26 #include "llvm/ADT/ArrayRef.h"
27 #include "llvm/ADT/DenseMap.h"
28 #include "llvm/ADT/DenseSet.h"
29 #include "llvm/ADT/None.h"
30 #include "llvm/ADT/SmallPtrSet.h"
31 #include "llvm/ADT/SmallSet.h"
32 #include "llvm/ADT/SmallVector.h"
33 #include "llvm/ADT/StringMap.h"
34 #include "llvm/ADT/StringRef.h"
35 #include "llvm/ADT/Twine.h"
38 #include "llvm/Analysis/LoopInfo.h"
43 #include "llvm/IR/BasicBlock.h"
44 #include "llvm/IR/CFG.h"
45 #include "llvm/IR/CallSite.h"
47 #include "llvm/IR/DebugLoc.h"
48 #include "llvm/IR/DiagnosticInfo.h"
49 #include "llvm/IR/Dominators.h"
50 #include "llvm/IR/Function.h"
51 #include "llvm/IR/GlobalValue.h"
52 #include "llvm/IR/InstrTypes.h"
53 #include "llvm/IR/Instruction.h"
54 #include "llvm/IR/Instructions.h"
55 #include "llvm/IR/IntrinsicInst.h"
56 #include "llvm/IR/LLVMContext.h"
57 #include "llvm/IR/MDBuilder.h"
58 #include "llvm/IR/Module.h"
59 #include "llvm/IR/PassManager.h"
61 #include "llvm/Pass.h"
65 #include "llvm/Support/Casting.h"
67 #include "llvm/Support/Debug.h"
69 #include "llvm/Support/ErrorOr.h"
72 #include "llvm/Transforms/IPO.h"
76 #include <algorithm>
77 #include <cassert>
78 #include <cstdint>
79 #include <functional>
80 #include <limits>
81 #include <map>
82 #include <memory>
83 #include <string>
84 #include <system_error>
85 #include <utility>
86 #include <vector>
87 
88 using namespace llvm;
89 using namespace sampleprof;
91 #define DEBUG_TYPE "sample-profile"
92 
93 // Command line option to specify the file to read samples from. This is
94 // mainly used for debugging.
96  "sample-profile-file", cl::init(""), cl::value_desc("filename"),
97  cl::desc("Profile file loaded by -sample-profile"), cl::Hidden);
98 
100  "sample-profile-max-propagate-iterations", cl::init(100),
101  cl::desc("Maximum number of iterations to go through when propagating "
102  "sample block/edge weights through the CFG."));
103 
105  "sample-profile-check-record-coverage", cl::init(0), cl::value_desc("N"),
106  cl::desc("Emit a warning if less than N% of records in the input profile "
107  "are matched to the IR."));
108 
110  "sample-profile-check-sample-coverage", cl::init(0), cl::value_desc("N"),
111  cl::desc("Emit a warning if less than N% of samples in the input profile "
112  "are matched to the IR."));
113 
115  "no-warn-sample-unused", cl::init(false), cl::Hidden,
116  cl::desc("Use this option to turn off/on warnings about function with "
117  "samples but without debug information to use those samples. "));
118 
119 namespace {
120 
121 using BlockWeightMap = DenseMap<const BasicBlock *, uint64_t>;
122 using EquivalenceClassMap = DenseMap<const BasicBlock *, const BasicBlock *>;
123 using Edge = std::pair<const BasicBlock *, const BasicBlock *>;
124 using EdgeWeightMap = DenseMap<Edge, uint64_t>;
125 using BlockEdgeMap =
127 
128 class SampleCoverageTracker {
129 public:
130  SampleCoverageTracker() = default;
131 
132  bool markSamplesUsed(const FunctionSamples *FS, uint32_t LineOffset,
133  uint32_t Discriminator, uint64_t Samples);
134  unsigned computeCoverage(unsigned Used, unsigned Total) const;
135  unsigned countUsedRecords(const FunctionSamples *FS,
136  ProfileSummaryInfo *PSI) const;
137  unsigned countBodyRecords(const FunctionSamples *FS,
138  ProfileSummaryInfo *PSI) const;
139  uint64_t getTotalUsedSamples() const { return TotalUsedSamples; }
140  uint64_t countBodySamples(const FunctionSamples *FS,
141  ProfileSummaryInfo *PSI) const;
142 
143  void clear() {
144  SampleCoverage.clear();
145  TotalUsedSamples = 0;
146  }
147 
148 private:
149  using BodySampleCoverageMap = std::map<LineLocation, unsigned>;
150  using FunctionSamplesCoverageMap =
152 
153  /// Coverage map for sampling records.
154  ///
155  /// This map keeps a record of sampling records that have been matched to
156  /// an IR instruction. This is used to detect some form of staleness in
157  /// profiles (see flag -sample-profile-check-coverage).
158  ///
159  /// Each entry in the map corresponds to a FunctionSamples instance. This is
160  /// another map that counts how many times the sample record at the
161  /// given location has been used.
162  FunctionSamplesCoverageMap SampleCoverage;
163 
164  /// Number of samples used from the profile.
165  ///
166  /// When a sampling record is used for the first time, the samples from
167  /// that record are added to this accumulator. Coverage is later computed
168  /// based on the total number of samples available in this function and
169  /// its callsites.
170  ///
171  /// Note that this accumulator tracks samples used from a single function
172  /// and all the inlined callsites. Strictly, we should have a map of counters
173  /// keyed by FunctionSamples pointers, but these stats are cleared after
174  /// every function, so we just need to keep a single counter.
175  uint64_t TotalUsedSamples = 0;
176 };
177 
178 /// Sample profile pass.
179 ///
180 /// This pass reads profile data from the file specified by
181 /// -sample-profile-file and annotates every affected function with the
182 /// profile information found in that file.
183 class SampleProfileLoader {
184 public:
185  SampleProfileLoader(
186  StringRef Name, bool IsThinLTOPreLink,
187  std::function<AssumptionCache &(Function &)> GetAssumptionCache,
188  std::function<TargetTransformInfo &(Function &)> GetTargetTransformInfo)
189  : GetAC(std::move(GetAssumptionCache)),
190  GetTTI(std::move(GetTargetTransformInfo)), Filename(Name),
191  IsThinLTOPreLink(IsThinLTOPreLink) {}
192 
193  bool doInitialization(Module &M);
194  bool runOnModule(Module &M, ModuleAnalysisManager *AM,
195  ProfileSummaryInfo *_PSI);
196 
197  void dump() { Reader->dump(); }
198 
199 protected:
201  unsigned getFunctionLoc(Function &F);
202  bool emitAnnotations(Function &F);
203  ErrorOr<uint64_t> getInstWeight(const Instruction &I);
204  ErrorOr<uint64_t> getBlockWeight(const BasicBlock *BB);
205  const FunctionSamples *findCalleeFunctionSamples(const Instruction &I) const;
206  std::vector<const FunctionSamples *>
207  findIndirectCallFunctionSamples(const Instruction &I, uint64_t &Sum) const;
208  const FunctionSamples *findFunctionSamples(const Instruction &I) const;
209  bool inlineCallInstruction(Instruction *I);
210  bool inlineHotFunctions(Function &F,
211  DenseSet<GlobalValue::GUID> &InlinedGUIDs);
212  void printEdgeWeight(raw_ostream &OS, Edge E);
213  void printBlockWeight(raw_ostream &OS, const BasicBlock *BB) const;
214  void printBlockEquivalence(raw_ostream &OS, const BasicBlock *BB);
215  bool computeBlockWeights(Function &F);
216  void findEquivalenceClasses(Function &F);
217  template <bool IsPostDom>
218  void findEquivalencesFor(BasicBlock *BB1, ArrayRef<BasicBlock *> Descendants,
220 
221  void propagateWeights(Function &F);
222  uint64_t visitEdge(Edge E, unsigned *NumUnknownEdges, Edge *UnknownEdge);
223  void buildEdges(Function &F);
224  bool propagateThroughEdges(Function &F, bool UpdateBlockCount);
225  void computeDominanceAndLoopInfo(Function &F);
226  void clearFunctionData();
227 
228  /// Map basic blocks to their computed weights.
229  ///
230  /// The weight of a basic block is defined to be the maximum
231  /// of all the instruction weights in that block.
232  BlockWeightMap BlockWeights;
233 
234  /// Map edges to their computed weights.
235  ///
236  /// Edge weights are computed by propagating basic block weights in
237  /// SampleProfile::propagateWeights.
238  EdgeWeightMap EdgeWeights;
239 
240  /// Set of visited blocks during propagation.
242 
243  /// Set of visited edges during propagation.
244  SmallSet<Edge, 32> VisitedEdges;
245 
246  /// Equivalence classes for block weights.
247  ///
248  /// Two blocks BB1 and BB2 are in the same equivalence class if they
249  /// dominate and post-dominate each other, and they are in the same loop
250  /// nest. When this happens, the two blocks are guaranteed to execute
251  /// the same number of times.
252  EquivalenceClassMap EquivalenceClass;
253 
254  /// Map from function name to Function *. Used to find the function from
255  /// the function name. If the function name contains suffix, additional
256  /// entry is added to map from the stripped name to the function if there
257  /// is one-to-one mapping.
259 
260  /// Dominance, post-dominance and loop information.
261  std::unique_ptr<DominatorTree> DT;
262  std::unique_ptr<PostDominatorTree> PDT;
263  std::unique_ptr<LoopInfo> LI;
264 
265  std::function<AssumptionCache &(Function &)> GetAC;
266  std::function<TargetTransformInfo &(Function &)> GetTTI;
267 
268  /// Predecessors for each basic block in the CFG.
269  BlockEdgeMap Predecessors;
270 
271  /// Successors for each basic block in the CFG.
272  BlockEdgeMap Successors;
273 
274  SampleCoverageTracker CoverageTracker;
275 
276  /// Profile reader object.
277  std::unique_ptr<SampleProfileReader> Reader;
278 
279  /// Samples collected for the body of this function.
280  FunctionSamples *Samples = nullptr;
281 
282  /// Name of the profile file to load.
283  std::string Filename;
284 
285  /// Flag indicating whether the profile input loaded successfully.
286  bool ProfileIsValid = false;
287 
288  /// Flag indicating if the pass is invoked in ThinLTO compile phase.
289  ///
290  /// In this phase, in annotation, we should not promote indirect calls.
291  /// Instead, we will mark GUIDs that needs to be annotated to the function.
292  bool IsThinLTOPreLink;
293 
294  /// Profile Summary Info computed from sample profile.
295  ProfileSummaryInfo *PSI = nullptr;
296 
297  /// Total number of samples collected in this profile.
298  ///
299  /// This is the sum of all the samples collected in all the functions executed
300  /// at runtime.
301  uint64_t TotalCollectedSamples = 0;
302 
303  /// Optimization Remark Emitter used to emit diagnostic remarks.
304  OptimizationRemarkEmitter *ORE = nullptr;
305 };
306 
307 class SampleProfileLoaderLegacyPass : public ModulePass {
308 public:
309  // Class identification, replacement for typeinfo
310  static char ID;
311 
312  SampleProfileLoaderLegacyPass(StringRef Name = SampleProfileFile,
313  bool IsThinLTOPreLink = false)
314  : ModulePass(ID), SampleLoader(Name, IsThinLTOPreLink,
315  [&](Function &F) -> AssumptionCache & {
316  return ACT->getAssumptionCache(F);
317  },
318  [&](Function &F) -> TargetTransformInfo & {
319  return TTIWP->getTTI(F);
320  }) {
323  }
324 
325  void dump() { SampleLoader.dump(); }
326 
327  bool doInitialization(Module &M) override {
328  return SampleLoader.doInitialization(M);
329  }
330 
331  StringRef getPassName() const override { return "Sample profile pass"; }
332  bool runOnModule(Module &M) override;
333 
334  void getAnalysisUsage(AnalysisUsage &AU) const override {
338  }
339 
340 private:
341  SampleProfileLoader SampleLoader;
342  AssumptionCacheTracker *ACT = nullptr;
343  TargetTransformInfoWrapperPass *TTIWP = nullptr;
344 };
345 
346 } // end anonymous namespace
347 
348 /// Return true if the given callsite is hot wrt to hot cutoff threshold.
349 ///
350 /// Functions that were inlined in the original binary will be represented
351 /// in the inline stack in the sample profile. If the profile shows that
352 /// the original inline decision was "good" (i.e., the callsite is executed
353 /// frequently), then we will recreate the inline decision and apply the
354 /// profile from the inlined callsite.
355 ///
356 /// To decide whether an inlined callsite is hot, we compare the callsite
357 /// sample count with the hot cutoff computed by ProfileSummaryInfo, it is
358 /// regarded as hot if the count is above the cutoff value.
359 static bool callsiteIsHot(const FunctionSamples *CallsiteFS,
360  ProfileSummaryInfo *PSI) {
361  if (!CallsiteFS)
362  return false; // The callsite was not inlined in the original binary.
363 
364  assert(PSI && "PSI is expected to be non null");
365  uint64_t CallsiteTotalSamples = CallsiteFS->getTotalSamples();
366  return PSI->isHotCount(CallsiteTotalSamples);
367 }
368 
369 /// Mark as used the sample record for the given function samples at
370 /// (LineOffset, Discriminator).
371 ///
372 /// \returns true if this is the first time we mark the given record.
373 bool SampleCoverageTracker::markSamplesUsed(const FunctionSamples *FS,
374  uint32_t LineOffset,
375  uint32_t Discriminator,
376  uint64_t Samples) {
377  LineLocation Loc(LineOffset, Discriminator);
378  unsigned &Count = SampleCoverage[FS][Loc];
379  bool FirstTime = (++Count == 1);
380  if (FirstTime)
381  TotalUsedSamples += Samples;
382  return FirstTime;
383 }
384 
385 /// Return the number of sample records that were applied from this profile.
386 ///
387 /// This count does not include records from cold inlined callsites.
388 unsigned
389 SampleCoverageTracker::countUsedRecords(const FunctionSamples *FS,
390  ProfileSummaryInfo *PSI) const {
391  auto I = SampleCoverage.find(FS);
392 
393  // The size of the coverage map for FS represents the number of records
394  // that were marked used at least once.
395  unsigned Count = (I != SampleCoverage.end()) ? I->second.size() : 0;
396 
397  // If there are inlined callsites in this function, count the samples found
398  // in the respective bodies. However, do not bother counting callees with 0
399  // total samples, these are callees that were never invoked at runtime.
400  for (const auto &I : FS->getCallsiteSamples())
401  for (const auto &J : I.second) {
402  const FunctionSamples *CalleeSamples = &J.second;
403  if (callsiteIsHot(CalleeSamples, PSI))
404  Count += countUsedRecords(CalleeSamples, PSI);
405  }
406 
407  return Count;
408 }
409 
410 /// Return the number of sample records in the body of this profile.
411 ///
412 /// This count does not include records from cold inlined callsites.
413 unsigned
414 SampleCoverageTracker::countBodyRecords(const FunctionSamples *FS,
415  ProfileSummaryInfo *PSI) const {
416  unsigned Count = FS->getBodySamples().size();
417 
418  // Only count records in hot callsites.
419  for (const auto &I : FS->getCallsiteSamples())
420  for (const auto &J : I.second) {
421  const FunctionSamples *CalleeSamples = &J.second;
422  if (callsiteIsHot(CalleeSamples, PSI))
423  Count += countBodyRecords(CalleeSamples, PSI);
424  }
425 
426  return Count;
427 }
428 
429 /// Return the number of samples collected in the body of this profile.
430 ///
431 /// This count does not include samples from cold inlined callsites.
432 uint64_t
433 SampleCoverageTracker::countBodySamples(const FunctionSamples *FS,
434  ProfileSummaryInfo *PSI) const {
435  uint64_t Total = 0;
436  for (const auto &I : FS->getBodySamples())
437  Total += I.second.getSamples();
438 
439  // Only count samples in hot callsites.
440  for (const auto &I : FS->getCallsiteSamples())
441  for (const auto &J : I.second) {
442  const FunctionSamples *CalleeSamples = &J.second;
443  if (callsiteIsHot(CalleeSamples, PSI))
444  Total += countBodySamples(CalleeSamples, PSI);
445  }
446 
447  return Total;
448 }
449 
450 /// Return the fraction of sample records used in this profile.
451 ///
452 /// The returned value is an unsigned integer in the range 0-100 indicating
453 /// the percentage of sample records that were used while applying this
454 /// profile to the associated function.
455 unsigned SampleCoverageTracker::computeCoverage(unsigned Used,
456  unsigned Total) const {
457  assert(Used <= Total &&
458  "number of used records cannot exceed the total number of records");
459  return Total > 0 ? Used * 100 / Total : 100;
460 }
461 
462 /// Clear all the per-function data used to load samples and propagate weights.
463 void SampleProfileLoader::clearFunctionData() {
464  BlockWeights.clear();
465  EdgeWeights.clear();
466  VisitedBlocks.clear();
467  VisitedEdges.clear();
468  EquivalenceClass.clear();
469  DT = nullptr;
470  PDT = nullptr;
471  LI = nullptr;
472  Predecessors.clear();
473  Successors.clear();
474  CoverageTracker.clear();
475 }
476 
477 #ifndef NDEBUG
478 /// Print the weight of edge \p E on stream \p OS.
479 ///
480 /// \param OS Stream to emit the output to.
481 /// \param E Edge to print.
482 void SampleProfileLoader::printEdgeWeight(raw_ostream &OS, Edge E) {
483  OS << "weight[" << E.first->getName() << "->" << E.second->getName()
484  << "]: " << EdgeWeights[E] << "\n";
485 }
486 
487 /// Print the equivalence class of block \p BB on stream \p OS.
488 ///
489 /// \param OS Stream to emit the output to.
490 /// \param BB Block to print.
491 void SampleProfileLoader::printBlockEquivalence(raw_ostream &OS,
492  const BasicBlock *BB) {
493  const BasicBlock *Equiv = EquivalenceClass[BB];
494  OS << "equivalence[" << BB->getName()
495  << "]: " << ((Equiv) ? EquivalenceClass[BB]->getName() : "NONE") << "\n";
496 }
497 
498 /// Print the weight of block \p BB on stream \p OS.
499 ///
500 /// \param OS Stream to emit the output to.
501 /// \param BB Block to print.
502 void SampleProfileLoader::printBlockWeight(raw_ostream &OS,
503  const BasicBlock *BB) const {
504  const auto &I = BlockWeights.find(BB);
505  uint64_t W = (I == BlockWeights.end() ? 0 : I->second);
506  OS << "weight[" << BB->getName() << "]: " << W << "\n";
507 }
508 #endif
509 
510 /// Get the weight for an instruction.
511 ///
512 /// The "weight" of an instruction \p Inst is the number of samples
513 /// collected on that instruction at runtime. To retrieve it, we
514 /// need to compute the line number of \p Inst relative to the start of its
515 /// function. We use HeaderLineno to compute the offset. We then
516 /// look up the samples collected for \p Inst using BodySamples.
517 ///
518 /// \param Inst Instruction to query.
519 ///
520 /// \returns the weight of \p Inst.
521 ErrorOr<uint64_t> SampleProfileLoader::getInstWeight(const Instruction &Inst) {
522  const DebugLoc &DLoc = Inst.getDebugLoc();
523  if (!DLoc)
524  return std::error_code();
525 
526  const FunctionSamples *FS = findFunctionSamples(Inst);
527  if (!FS)
528  return std::error_code();
529 
530  // Ignore all intrinsics and branch instructions.
531  // Branch instruction usually contains debug info from sources outside of
532  // the residing basic block, thus we ignore them during annotation.
533  if (isa<BranchInst>(Inst) || isa<IntrinsicInst>(Inst))
534  return std::error_code();
535 
536  // If a direct call/invoke instruction is inlined in profile
537  // (findCalleeFunctionSamples returns non-empty result), but not inlined here,
538  // it means that the inlined callsite has no sample, thus the call
539  // instruction should have 0 count.
540  if ((isa<CallInst>(Inst) || isa<InvokeInst>(Inst)) &&
541  !ImmutableCallSite(&Inst).isIndirectCall() &&
542  findCalleeFunctionSamples(Inst))
543  return 0;
544 
545  const DILocation *DIL = DLoc;
546  uint32_t LineOffset = FunctionSamples::getOffset(DIL);
547  uint32_t Discriminator = DIL->getBaseDiscriminator();
548  ErrorOr<uint64_t> R = FS->findSamplesAt(LineOffset, Discriminator);
549  if (R) {
550  bool FirstMark =
551  CoverageTracker.markSamplesUsed(FS, LineOffset, Discriminator, R.get());
552  if (FirstMark) {
553  ORE->emit([&]() {
554  OptimizationRemarkAnalysis Remark(DEBUG_TYPE, "AppliedSamples", &Inst);
555  Remark << "Applied " << ore::NV("NumSamples", *R);
556  Remark << " samples from profile (offset: ";
557  Remark << ore::NV("LineOffset", LineOffset);
558  if (Discriminator) {
559  Remark << ".";
560  Remark << ore::NV("Discriminator", Discriminator);
561  }
562  Remark << ")";
563  return Remark;
564  });
565  }
566  LLVM_DEBUG(dbgs() << " " << DLoc.getLine() << "."
567  << DIL->getBaseDiscriminator() << ":" << Inst
568  << " (line offset: " << LineOffset << "."
569  << DIL->getBaseDiscriminator() << " - weight: " << R.get()
570  << ")\n");
571  }
572  return R;
573 }
574 
575 /// Compute the weight of a basic block.
576 ///
577 /// The weight of basic block \p BB is the maximum weight of all the
578 /// instructions in BB.
579 ///
580 /// \param BB The basic block to query.
581 ///
582 /// \returns the weight for \p BB.
583 ErrorOr<uint64_t> SampleProfileLoader::getBlockWeight(const BasicBlock *BB) {
584  uint64_t Max = 0;
585  bool HasWeight = false;
586  for (auto &I : BB->getInstList()) {
587  const ErrorOr<uint64_t> &R = getInstWeight(I);
588  if (R) {
589  Max = std::max(Max, R.get());
590  HasWeight = true;
591  }
592  }
593  return HasWeight ? ErrorOr<uint64_t>(Max) : std::error_code();
594 }
595 
596 /// Compute and store the weights of every basic block.
597 ///
598 /// This populates the BlockWeights map by computing
599 /// the weights of every basic block in the CFG.
600 ///
601 /// \param F The function to query.
602 bool SampleProfileLoader::computeBlockWeights(Function &F) {
603  bool Changed = false;
604  LLVM_DEBUG(dbgs() << "Block weights\n");
605  for (const auto &BB : F) {
606  ErrorOr<uint64_t> Weight = getBlockWeight(&BB);
607  if (Weight) {
608  BlockWeights[&BB] = Weight.get();
609  VisitedBlocks.insert(&BB);
610  Changed = true;
611  }
612  LLVM_DEBUG(printBlockWeight(dbgs(), &BB));
613  }
614 
615  return Changed;
616 }
617 
618 /// Get the FunctionSamples for a call instruction.
619 ///
620 /// The FunctionSamples of a call/invoke instruction \p Inst is the inlined
621 /// instance in which that call instruction is calling to. It contains
622 /// all samples that resides in the inlined instance. We first find the
623 /// inlined instance in which the call instruction is from, then we
624 /// traverse its children to find the callsite with the matching
625 /// location.
626 ///
627 /// \param Inst Call/Invoke instruction to query.
628 ///
629 /// \returns The FunctionSamples pointer to the inlined instance.
630 const FunctionSamples *
631 SampleProfileLoader::findCalleeFunctionSamples(const Instruction &Inst) const {
632  const DILocation *DIL = Inst.getDebugLoc();
633  if (!DIL) {
634  return nullptr;
635  }
636 
637  StringRef CalleeName;
638  if (const CallInst *CI = dyn_cast<CallInst>(&Inst))
639  if (Function *Callee = CI->getCalledFunction())
640  CalleeName = Callee->getName();
641 
642  const FunctionSamples *FS = findFunctionSamples(Inst);
643  if (FS == nullptr)
644  return nullptr;
645 
646  std::string CalleeGUID;
647  CalleeName = getRepInFormat(CalleeName, Reader->getFormat(), CalleeGUID);
649  DIL->getBaseDiscriminator()),
650  CalleeName);
651 }
652 
653 /// Returns a vector of FunctionSamples that are the indirect call targets
654 /// of \p Inst. The vector is sorted by the total number of samples. Stores
655 /// the total call count of the indirect call in \p Sum.
656 std::vector<const FunctionSamples *>
657 SampleProfileLoader::findIndirectCallFunctionSamples(
658  const Instruction &Inst, uint64_t &Sum) const {
659  const DILocation *DIL = Inst.getDebugLoc();
660  std::vector<const FunctionSamples *> R;
661 
662  if (!DIL) {
663  return R;
664  }
665 
666  const FunctionSamples *FS = findFunctionSamples(Inst);
667  if (FS == nullptr)
668  return R;
669 
670  uint32_t LineOffset = FunctionSamples::getOffset(DIL);
671  uint32_t Discriminator = DIL->getBaseDiscriminator();
672 
673  auto T = FS->findCallTargetMapAt(LineOffset, Discriminator);
674  Sum = 0;
675  if (T)
676  for (const auto &T_C : T.get())
677  Sum += T_C.second;
680  if (M->empty())
681  return R;
682  for (const auto &NameFS : *M) {
683  Sum += NameFS.second.getEntrySamples();
684  R.push_back(&NameFS.second);
685  }
686  llvm::sort(R.begin(), R.end(),
687  [](const FunctionSamples *L, const FunctionSamples *R) {
688  return L->getEntrySamples() > R->getEntrySamples();
689  });
690  }
691  return R;
692 }
693 
694 /// Get the FunctionSamples for an instruction.
695 ///
696 /// The FunctionSamples of an instruction \p Inst is the inlined instance
697 /// in which that instruction is coming from. We traverse the inline stack
698 /// of that instruction, and match it with the tree nodes in the profile.
699 ///
700 /// \param Inst Instruction to query.
701 ///
702 /// \returns the FunctionSamples pointer to the inlined instance.
703 const FunctionSamples *
704 SampleProfileLoader::findFunctionSamples(const Instruction &Inst) const {
706  const DILocation *DIL = Inst.getDebugLoc();
707  if (!DIL)
708  return Samples;
709 
710  return Samples->findFunctionSamples(DIL);
711 }
712 
713 bool SampleProfileLoader::inlineCallInstruction(Instruction *I) {
714  assert(isa<CallInst>(I) || isa<InvokeInst>(I));
715  CallSite CS(I);
716  Function *CalledFunction = CS.getCalledFunction();
717  assert(CalledFunction);
718  DebugLoc DLoc = I->getDebugLoc();
719  BasicBlock *BB = I->getParent();
720  InlineParams Params = getInlineParams();
721  Params.ComputeFullInlineCost = true;
722  // Checks if there is anything in the reachable portion of the callee at
723  // this callsite that makes this inlining potentially illegal. Need to
724  // set ComputeFullInlineCost, otherwise getInlineCost may return early
725  // when cost exceeds threshold without checking all IRs in the callee.
726  // The acutal cost does not matter because we only checks isNever() to
727  // see if it is legal to inline the callsite.
728  InlineCost Cost = getInlineCost(CS, Params, GetTTI(*CalledFunction), GetAC,
729  None, nullptr, nullptr);
730  if (Cost.isNever()) {
731  ORE->emit(OptimizationRemark(DEBUG_TYPE, "Not inline", DLoc, BB)
732  << "incompatible inlining");
733  return false;
734  }
735  InlineFunctionInfo IFI(nullptr, &GetAC);
736  if (InlineFunction(CS, IFI)) {
737  // The call to InlineFunction erases I, so we can't pass it here.
738  ORE->emit(OptimizationRemark(DEBUG_TYPE, "HotInline", DLoc, BB)
739  << "inlined hot callee '" << ore::NV("Callee", CalledFunction)
740  << "' into '" << ore::NV("Caller", BB->getParent()) << "'");
741  return true;
742  }
743  return false;
744 }
745 
746 /// Iteratively inline hot callsites of a function.
747 ///
748 /// Iteratively traverse all callsites of the function \p F, and find if
749 /// the corresponding inlined instance exists and is hot in profile. If
750 /// it is hot enough, inline the callsites and adds new callsites of the
751 /// callee into the caller. If the call is an indirect call, first promote
752 /// it to direct call. Each indirect call is limited with a single target.
753 ///
754 /// \param F function to perform iterative inlining.
755 /// \param InlinedGUIDs a set to be updated to include all GUIDs that are
756 /// inlined in the profiled binary.
757 ///
758 /// \returns True if there is any inline happened.
759 bool SampleProfileLoader::inlineHotFunctions(
760  Function &F, DenseSet<GlobalValue::GUID> &InlinedGUIDs) {
761  DenseSet<Instruction *> PromotedInsns;
762  bool Changed = false;
763  bool isCompact = (Reader->getFormat() == SPF_Compact_Binary);
764  while (true) {
765  bool LocalChanged = false;
767  for (auto &BB : F) {
768  bool Hot = false;
770  for (auto &I : BB.getInstList()) {
771  const FunctionSamples *FS = nullptr;
772  if ((isa<CallInst>(I) || isa<InvokeInst>(I)) &&
773  !isa<IntrinsicInst>(I) && (FS = findCalleeFunctionSamples(I))) {
774  Candidates.push_back(&I);
775  if (callsiteIsHot(FS, PSI))
776  Hot = true;
777  }
778  }
779  if (Hot) {
780  CIS.insert(CIS.begin(), Candidates.begin(), Candidates.end());
781  }
782  }
783  for (auto I : CIS) {
784  Function *CalledFunction = CallSite(I).getCalledFunction();
785  // Do not inline recursive calls.
786  if (CalledFunction == &F)
787  continue;
788  if (CallSite(I).isIndirectCall()) {
789  if (PromotedInsns.count(I))
790  continue;
791  uint64_t Sum;
792  for (const auto *FS : findIndirectCallFunctionSamples(*I, Sum)) {
793  if (IsThinLTOPreLink) {
794  FS->findInlinedFunctions(InlinedGUIDs, F.getParent(),
796  isCompact);
797  continue;
798  }
799  auto CalleeFunctionName = FS->getName();
800  // If it is a recursive call, we do not inline it as it could bloat
801  // the code exponentially. There is way to better handle this, e.g.
802  // clone the caller first, and inline the cloned caller if it is
803  // recursive. As llvm does not inline recursive calls, we will
804  // simply ignore it instead of handling it explicitly.
805  std::string FGUID;
806  auto Fname = getRepInFormat(F.getName(), Reader->getFormat(), FGUID);
807  if (CalleeFunctionName == Fname)
808  continue;
809 
810  const char *Reason = "Callee function not available";
811  auto R = SymbolMap.find(CalleeFunctionName);
812  if (R != SymbolMap.end() && R->getValue() &&
813  !R->getValue()->isDeclaration() &&
814  R->getValue()->getSubprogram() &&
815  isLegalToPromote(CallSite(I), R->getValue(), &Reason)) {
816  uint64_t C = FS->getEntrySamples();
817  Instruction *DI =
818  pgo::promoteIndirectCall(I, R->getValue(), C, Sum, false, ORE);
819  Sum -= C;
820  PromotedInsns.insert(I);
821  // If profile mismatches, we should not attempt to inline DI.
822  if ((isa<CallInst>(DI) || isa<InvokeInst>(DI)) &&
823  inlineCallInstruction(DI))
824  LocalChanged = true;
825  } else {
826  LLVM_DEBUG(dbgs()
827  << "\nFailed to promote indirect call to "
828  << CalleeFunctionName << " because " << Reason << "\n");
829  }
830  }
831  } else if (CalledFunction && CalledFunction->getSubprogram() &&
832  !CalledFunction->isDeclaration()) {
833  if (inlineCallInstruction(I))
834  LocalChanged = true;
835  } else if (IsThinLTOPreLink) {
836  findCalleeFunctionSamples(*I)->findInlinedFunctions(
837  InlinedGUIDs, F.getParent(), PSI->getOrCompHotCountThreshold(),
838  isCompact);
839  }
840  }
841  if (LocalChanged) {
842  Changed = true;
843  } else {
844  break;
845  }
846  }
847  return Changed;
848 }
849 
850 /// Find equivalence classes for the given block.
851 ///
852 /// This finds all the blocks that are guaranteed to execute the same
853 /// number of times as \p BB1. To do this, it traverses all the
854 /// descendants of \p BB1 in the dominator or post-dominator tree.
855 ///
856 /// A block BB2 will be in the same equivalence class as \p BB1 if
857 /// the following holds:
858 ///
859 /// 1- \p BB1 is a descendant of BB2 in the opposite tree. So, if BB2
860 /// is a descendant of \p BB1 in the dominator tree, then BB2 should
861 /// dominate BB1 in the post-dominator tree.
862 ///
863 /// 2- Both BB2 and \p BB1 must be in the same loop.
864 ///
865 /// For every block BB2 that meets those two requirements, we set BB2's
866 /// equivalence class to \p BB1.
867 ///
868 /// \param BB1 Block to check.
869 /// \param Descendants Descendants of \p BB1 in either the dom or pdom tree.
870 /// \param DomTree Opposite dominator tree. If \p Descendants is filled
871 /// with blocks from \p BB1's dominator tree, then
872 /// this is the post-dominator tree, and vice versa.
873 template <bool IsPostDom>
874 void SampleProfileLoader::findEquivalencesFor(
875  BasicBlock *BB1, ArrayRef<BasicBlock *> Descendants,
877  const BasicBlock *EC = EquivalenceClass[BB1];
878  uint64_t Weight = BlockWeights[EC];
879  for (const auto *BB2 : Descendants) {
880  bool IsDomParent = DomTree->dominates(BB2, BB1);
881  bool IsInSameLoop = LI->getLoopFor(BB1) == LI->getLoopFor(BB2);
882  if (BB1 != BB2 && IsDomParent && IsInSameLoop) {
883  EquivalenceClass[BB2] = EC;
884  // If BB2 is visited, then the entire EC should be marked as visited.
885  if (VisitedBlocks.count(BB2)) {
886  VisitedBlocks.insert(EC);
887  }
888 
889  // If BB2 is heavier than BB1, make BB2 have the same weight
890  // as BB1.
891  //
892  // Note that we don't worry about the opposite situation here
893  // (when BB2 is lighter than BB1). We will deal with this
894  // during the propagation phase. Right now, we just want to
895  // make sure that BB1 has the largest weight of all the
896  // members of its equivalence set.
897  Weight = std::max(Weight, BlockWeights[BB2]);
898  }
899  }
900  if (EC == &EC->getParent()->getEntryBlock()) {
901  BlockWeights[EC] = Samples->getHeadSamples() + 1;
902  } else {
903  BlockWeights[EC] = Weight;
904  }
905 }
906 
907 /// Find equivalence classes.
908 ///
909 /// Since samples may be missing from blocks, we can fill in the gaps by setting
910 /// the weights of all the blocks in the same equivalence class to the same
911 /// weight. To compute the concept of equivalence, we use dominance and loop
912 /// information. Two blocks B1 and B2 are in the same equivalence class if B1
913 /// dominates B2, B2 post-dominates B1 and both are in the same loop.
914 ///
915 /// \param F The function to query.
916 void SampleProfileLoader::findEquivalenceClasses(Function &F) {
917  SmallVector<BasicBlock *, 8> DominatedBBs;
918  LLVM_DEBUG(dbgs() << "\nBlock equivalence classes\n");
919  // Find equivalence sets based on dominance and post-dominance information.
920  for (auto &BB : F) {
921  BasicBlock *BB1 = &BB;
922 
923  // Compute BB1's equivalence class once.
924  if (EquivalenceClass.count(BB1)) {
925  LLVM_DEBUG(printBlockEquivalence(dbgs(), BB1));
926  continue;
927  }
928 
929  // By default, blocks are in their own equivalence class.
930  EquivalenceClass[BB1] = BB1;
931 
932  // Traverse all the blocks dominated by BB1. We are looking for
933  // every basic block BB2 such that:
934  //
935  // 1- BB1 dominates BB2.
936  // 2- BB2 post-dominates BB1.
937  // 3- BB1 and BB2 are in the same loop nest.
938  //
939  // If all those conditions hold, it means that BB2 is executed
940  // as many times as BB1, so they are placed in the same equivalence
941  // class by making BB2's equivalence class be BB1.
942  DominatedBBs.clear();
943  DT->getDescendants(BB1, DominatedBBs);
944  findEquivalencesFor(BB1, DominatedBBs, PDT.get());
945 
946  LLVM_DEBUG(printBlockEquivalence(dbgs(), BB1));
947  }
948 
949  // Assign weights to equivalence classes.
950  //
951  // All the basic blocks in the same equivalence class will execute
952  // the same number of times. Since we know that the head block in
953  // each equivalence class has the largest weight, assign that weight
954  // to all the blocks in that equivalence class.
955  LLVM_DEBUG(
956  dbgs() << "\nAssign the same weight to all blocks in the same class\n");
957  for (auto &BI : F) {
958  const BasicBlock *BB = &BI;
959  const BasicBlock *EquivBB = EquivalenceClass[BB];
960  if (BB != EquivBB)
961  BlockWeights[BB] = BlockWeights[EquivBB];
962  LLVM_DEBUG(printBlockWeight(dbgs(), BB));
963  }
964 }
965 
966 /// Visit the given edge to decide if it has a valid weight.
967 ///
968 /// If \p E has not been visited before, we copy to \p UnknownEdge
969 /// and increment the count of unknown edges.
970 ///
971 /// \param E Edge to visit.
972 /// \param NumUnknownEdges Current number of unknown edges.
973 /// \param UnknownEdge Set if E has not been visited before.
974 ///
975 /// \returns E's weight, if known. Otherwise, return 0.
976 uint64_t SampleProfileLoader::visitEdge(Edge E, unsigned *NumUnknownEdges,
977  Edge *UnknownEdge) {
978  if (!VisitedEdges.count(E)) {
979  (*NumUnknownEdges)++;
980  *UnknownEdge = E;
981  return 0;
982  }
983 
984  return EdgeWeights[E];
985 }
986 
987 /// Propagate weights through incoming/outgoing edges.
988 ///
989 /// If the weight of a basic block is known, and there is only one edge
990 /// with an unknown weight, we can calculate the weight of that edge.
991 ///
992 /// Similarly, if all the edges have a known count, we can calculate the
993 /// count of the basic block, if needed.
994 ///
995 /// \param F Function to process.
996 /// \param UpdateBlockCount Whether we should update basic block counts that
997 /// has already been annotated.
998 ///
999 /// \returns True if new weights were assigned to edges or blocks.
1000 bool SampleProfileLoader::propagateThroughEdges(Function &F,
1001  bool UpdateBlockCount) {
1002  bool Changed = false;
1003  LLVM_DEBUG(dbgs() << "\nPropagation through edges\n");
1004  for (const auto &BI : F) {
1005  const BasicBlock *BB = &BI;
1006  const BasicBlock *EC = EquivalenceClass[BB];
1007 
1008  // Visit all the predecessor and successor edges to determine
1009  // which ones have a weight assigned already. Note that it doesn't
1010  // matter that we only keep track of a single unknown edge. The
1011  // only case we are interested in handling is when only a single
1012  // edge is unknown (see setEdgeOrBlockWeight).
1013  for (unsigned i = 0; i < 2; i++) {
1014  uint64_t TotalWeight = 0;
1015  unsigned NumUnknownEdges = 0, NumTotalEdges = 0;
1016  Edge UnknownEdge, SelfReferentialEdge, SingleEdge;
1017 
1018  if (i == 0) {
1019  // First, visit all predecessor edges.
1020  NumTotalEdges = Predecessors[BB].size();
1021  for (auto *Pred : Predecessors[BB]) {
1022  Edge E = std::make_pair(Pred, BB);
1023  TotalWeight += visitEdge(E, &NumUnknownEdges, &UnknownEdge);
1024  if (E.first == E.second)
1025  SelfReferentialEdge = E;
1026  }
1027  if (NumTotalEdges == 1) {
1028  SingleEdge = std::make_pair(Predecessors[BB][0], BB);
1029  }
1030  } else {
1031  // On the second round, visit all successor edges.
1032  NumTotalEdges = Successors[BB].size();
1033  for (auto *Succ : Successors[BB]) {
1034  Edge E = std::make_pair(BB, Succ);
1035  TotalWeight += visitEdge(E, &NumUnknownEdges, &UnknownEdge);
1036  }
1037  if (NumTotalEdges == 1) {
1038  SingleEdge = std::make_pair(BB, Successors[BB][0]);
1039  }
1040  }
1041 
1042  // After visiting all the edges, there are three cases that we
1043  // can handle immediately:
1044  //
1045  // - All the edge weights are known (i.e., NumUnknownEdges == 0).
1046  // In this case, we simply check that the sum of all the edges
1047  // is the same as BB's weight. If not, we change BB's weight
1048  // to match. Additionally, if BB had not been visited before,
1049  // we mark it visited.
1050  //
1051  // - Only one edge is unknown and BB has already been visited.
1052  // In this case, we can compute the weight of the edge by
1053  // subtracting the total block weight from all the known
1054  // edge weights. If the edges weight more than BB, then the
1055  // edge of the last remaining edge is set to zero.
1056  //
1057  // - There exists a self-referential edge and the weight of BB is
1058  // known. In this case, this edge can be based on BB's weight.
1059  // We add up all the other known edges and set the weight on
1060  // the self-referential edge as we did in the previous case.
1061  //
1062  // In any other case, we must continue iterating. Eventually,
1063  // all edges will get a weight, or iteration will stop when
1064  // it reaches SampleProfileMaxPropagateIterations.
1065  if (NumUnknownEdges <= 1) {
1066  uint64_t &BBWeight = BlockWeights[EC];
1067  if (NumUnknownEdges == 0) {
1068  if (!VisitedBlocks.count(EC)) {
1069  // If we already know the weight of all edges, the weight of the
1070  // basic block can be computed. It should be no larger than the sum
1071  // of all edge weights.
1072  if (TotalWeight > BBWeight) {
1073  BBWeight = TotalWeight;
1074  Changed = true;
1075  LLVM_DEBUG(dbgs() << "All edge weights for " << BB->getName()
1076  << " known. Set weight for block: ";
1077  printBlockWeight(dbgs(), BB););
1078  }
1079  } else if (NumTotalEdges == 1 &&
1080  EdgeWeights[SingleEdge] < BlockWeights[EC]) {
1081  // If there is only one edge for the visited basic block, use the
1082  // block weight to adjust edge weight if edge weight is smaller.
1083  EdgeWeights[SingleEdge] = BlockWeights[EC];
1084  Changed = true;
1085  }
1086  } else if (NumUnknownEdges == 1 && VisitedBlocks.count(EC)) {
1087  // If there is a single unknown edge and the block has been
1088  // visited, then we can compute E's weight.
1089  if (BBWeight >= TotalWeight)
1090  EdgeWeights[UnknownEdge] = BBWeight - TotalWeight;
1091  else
1092  EdgeWeights[UnknownEdge] = 0;
1093  const BasicBlock *OtherEC;
1094  if (i == 0)
1095  OtherEC = EquivalenceClass[UnknownEdge.first];
1096  else
1097  OtherEC = EquivalenceClass[UnknownEdge.second];
1098  // Edge weights should never exceed the BB weights it connects.
1099  if (VisitedBlocks.count(OtherEC) &&
1100  EdgeWeights[UnknownEdge] > BlockWeights[OtherEC])
1101  EdgeWeights[UnknownEdge] = BlockWeights[OtherEC];
1102  VisitedEdges.insert(UnknownEdge);
1103  Changed = true;
1104  LLVM_DEBUG(dbgs() << "Set weight for edge: ";
1105  printEdgeWeight(dbgs(), UnknownEdge));
1106  }
1107  } else if (VisitedBlocks.count(EC) && BlockWeights[EC] == 0) {
1108  // If a block Weights 0, all its in/out edges should weight 0.
1109  if (i == 0) {
1110  for (auto *Pred : Predecessors[BB]) {
1111  Edge E = std::make_pair(Pred, BB);
1112  EdgeWeights[E] = 0;
1113  VisitedEdges.insert(E);
1114  }
1115  } else {
1116  for (auto *Succ : Successors[BB]) {
1117  Edge E = std::make_pair(BB, Succ);
1118  EdgeWeights[E] = 0;
1119  VisitedEdges.insert(E);
1120  }
1121  }
1122  } else if (SelfReferentialEdge.first && VisitedBlocks.count(EC)) {
1123  uint64_t &BBWeight = BlockWeights[BB];
1124  // We have a self-referential edge and the weight of BB is known.
1125  if (BBWeight >= TotalWeight)
1126  EdgeWeights[SelfReferentialEdge] = BBWeight - TotalWeight;
1127  else
1128  EdgeWeights[SelfReferentialEdge] = 0;
1129  VisitedEdges.insert(SelfReferentialEdge);
1130  Changed = true;
1131  LLVM_DEBUG(dbgs() << "Set self-referential edge weight to: ";
1132  printEdgeWeight(dbgs(), SelfReferentialEdge));
1133  }
1134  if (UpdateBlockCount && !VisitedBlocks.count(EC) && TotalWeight > 0) {
1135  BlockWeights[EC] = TotalWeight;
1136  VisitedBlocks.insert(EC);
1137  Changed = true;
1138  }
1139  }
1140  }
1141 
1142  return Changed;
1143 }
1144 
1145 /// Build in/out edge lists for each basic block in the CFG.
1146 ///
1147 /// We are interested in unique edges. If a block B1 has multiple
1148 /// edges to another block B2, we only add a single B1->B2 edge.
1149 void SampleProfileLoader::buildEdges(Function &F) {
1150  for (auto &BI : F) {
1151  BasicBlock *B1 = &BI;
1152 
1153  // Add predecessors for B1.
1155  if (!Predecessors[B1].empty())
1156  llvm_unreachable("Found a stale predecessors list in a basic block.");
1157  for (pred_iterator PI = pred_begin(B1), PE = pred_end(B1); PI != PE; ++PI) {
1158  BasicBlock *B2 = *PI;
1159  if (Visited.insert(B2).second)
1160  Predecessors[B1].push_back(B2);
1161  }
1162 
1163  // Add successors for B1.
1164  Visited.clear();
1165  if (!Successors[B1].empty())
1166  llvm_unreachable("Found a stale successors list in a basic block.");
1167  for (succ_iterator SI = succ_begin(B1), SE = succ_end(B1); SI != SE; ++SI) {
1168  BasicBlock *B2 = *SI;
1169  if (Visited.insert(B2).second)
1170  Successors[B1].push_back(B2);
1171  }
1172  }
1173 }
1174 
1175 /// Returns the sorted CallTargetMap \p M by count in descending order.
1177  const SampleRecord::CallTargetMap &M) {
1179  for (auto I = M.begin(); I != M.end(); ++I)
1180  R.push_back({Function::getGUID(I->getKey()), I->getValue()});
1181  llvm::sort(R.begin(), R.end(),
1182  [](const InstrProfValueData &L, const InstrProfValueData &R) {
1183  if (L.Count == R.Count)
1184  return L.Value > R.Value;
1185  else
1186  return L.Count > R.Count;
1187  });
1188  return R;
1189 }
1190 
1191 /// Propagate weights into edges
1192 ///
1193 /// The following rules are applied to every block BB in the CFG:
1194 ///
1195 /// - If BB has a single predecessor/successor, then the weight
1196 /// of that edge is the weight of the block.
1197 ///
1198 /// - If all incoming or outgoing edges are known except one, and the
1199 /// weight of the block is already known, the weight of the unknown
1200 /// edge will be the weight of the block minus the sum of all the known
1201 /// edges. If the sum of all the known edges is larger than BB's weight,
1202 /// we set the unknown edge weight to zero.
1203 ///
1204 /// - If there is a self-referential edge, and the weight of the block is
1205 /// known, the weight for that edge is set to the weight of the block
1206 /// minus the weight of the other incoming edges to that block (if
1207 /// known).
1208 void SampleProfileLoader::propagateWeights(Function &F) {
1209  bool Changed = true;
1210  unsigned I = 0;
1211 
1212  // If BB weight is larger than its corresponding loop's header BB weight,
1213  // use the BB weight to replace the loop header BB weight.
1214  for (auto &BI : F) {
1215  BasicBlock *BB = &BI;
1216  Loop *L = LI->getLoopFor(BB);
1217  if (!L) {
1218  continue;
1219  }
1220  BasicBlock *Header = L->getHeader();
1221  if (Header && BlockWeights[BB] > BlockWeights[Header]) {
1222  BlockWeights[Header] = BlockWeights[BB];
1223  }
1224  }
1225 
1226  // Before propagation starts, build, for each block, a list of
1227  // unique predecessors and successors. This is necessary to handle
1228  // identical edges in multiway branches. Since we visit all blocks and all
1229  // edges of the CFG, it is cleaner to build these lists once at the start
1230  // of the pass.
1231  buildEdges(F);
1232 
1233  // Propagate until we converge or we go past the iteration limit.
1234  while (Changed && I++ < SampleProfileMaxPropagateIterations) {
1235  Changed = propagateThroughEdges(F, false);
1236  }
1237 
1238  // The first propagation propagates BB counts from annotated BBs to unknown
1239  // BBs. The 2nd propagation pass resets edges weights, and use all BB weights
1240  // to propagate edge weights.
1241  VisitedEdges.clear();
1242  Changed = true;
1243  while (Changed && I++ < SampleProfileMaxPropagateIterations) {
1244  Changed = propagateThroughEdges(F, false);
1245  }
1246 
1247  // The 3rd propagation pass allows adjust annotated BB weights that are
1248  // obviously wrong.
1249  Changed = true;
1250  while (Changed && I++ < SampleProfileMaxPropagateIterations) {
1251  Changed = propagateThroughEdges(F, true);
1252  }
1253 
1254  // Generate MD_prof metadata for every branch instruction using the
1255  // edge weights computed during propagation.
1256  LLVM_DEBUG(dbgs() << "\nPropagation complete. Setting branch weights\n");
1257  LLVMContext &Ctx = F.getContext();
1258  MDBuilder MDB(Ctx);
1259  for (auto &BI : F) {
1260  BasicBlock *BB = &BI;
1261 
1262  if (BlockWeights[BB]) {
1263  for (auto &I : BB->getInstList()) {
1264  if (!isa<CallInst>(I) && !isa<InvokeInst>(I))
1265  continue;
1266  CallSite CS(&I);
1267  if (!CS.getCalledFunction()) {
1268  const DebugLoc &DLoc = I.getDebugLoc();
1269  if (!DLoc)
1270  continue;
1271  const DILocation *DIL = DLoc;
1272  uint32_t LineOffset = FunctionSamples::getOffset(DIL);
1273  uint32_t Discriminator = DIL->getBaseDiscriminator();
1274 
1275  const FunctionSamples *FS = findFunctionSamples(I);
1276  if (!FS)
1277  continue;
1278  auto T = FS->findCallTargetMapAt(LineOffset, Discriminator);
1279  if (!T || T.get().empty())
1280  continue;
1281  SmallVector<InstrProfValueData, 2> SortedCallTargets =
1282  SortCallTargets(T.get());
1283  uint64_t Sum;
1284  findIndirectCallFunctionSamples(I, Sum);
1285  annotateValueSite(*I.getParent()->getParent()->getParent(), I,
1286  SortedCallTargets, Sum, IPVK_IndirectCallTarget,
1287  SortedCallTargets.size());
1288  } else if (!dyn_cast<IntrinsicInst>(&I)) {
1289  SmallVector<uint32_t, 1> Weights;
1290  Weights.push_back(BlockWeights[BB]);
1291  I.setMetadata(LLVMContext::MD_prof, MDB.createBranchWeights(Weights));
1292  }
1293  }
1294  }
1295  TerminatorInst *TI = BB->getTerminator();
1296  if (TI->getNumSuccessors() == 1)
1297  continue;
1298  if (!isa<BranchInst>(TI) && !isa<SwitchInst>(TI))
1299  continue;
1300 
1301  DebugLoc BranchLoc = TI->getDebugLoc();
1302  LLVM_DEBUG(dbgs() << "\nGetting weights for branch at line "
1303  << ((BranchLoc) ? Twine(BranchLoc.getLine())
1304  : Twine("<UNKNOWN LOCATION>"))
1305  << ".\n");
1306  SmallVector<uint32_t, 4> Weights;
1307  uint32_t MaxWeight = 0;
1308  Instruction *MaxDestInst;
1309  for (unsigned I = 0; I < TI->getNumSuccessors(); ++I) {
1310  BasicBlock *Succ = TI->getSuccessor(I);
1311  Edge E = std::make_pair(BB, Succ);
1312  uint64_t Weight = EdgeWeights[E];
1313  LLVM_DEBUG(dbgs() << "\t"; printEdgeWeight(dbgs(), E));
1314  // Use uint32_t saturated arithmetic to adjust the incoming weights,
1315  // if needed. Sample counts in profiles are 64-bit unsigned values,
1316  // but internally branch weights are expressed as 32-bit values.
1317  if (Weight > std::numeric_limits<uint32_t>::max()) {
1318  LLVM_DEBUG(dbgs() << " (saturated due to uint32_t overflow)");
1320  }
1321  // Weight is added by one to avoid propagation errors introduced by
1322  // 0 weights.
1323  Weights.push_back(static_cast<uint32_t>(Weight + 1));
1324  if (Weight != 0) {
1325  if (Weight > MaxWeight) {
1326  MaxWeight = Weight;
1327  MaxDestInst = Succ->getFirstNonPHIOrDbgOrLifetime();
1328  }
1329  }
1330  }
1331 
1332  uint64_t TempWeight;
1333  // Only set weights if there is at least one non-zero weight.
1334  // In any other case, let the analyzer set weights.
1335  // Do not set weights if the weights are present. In ThinLTO, the profile
1336  // annotation is done twice. If the first annotation already set the
1337  // weights, the second pass does not need to set it.
1338  if (MaxWeight > 0 && !TI->extractProfTotalWeight(TempWeight)) {
1339  LLVM_DEBUG(dbgs() << "SUCCESS. Found non-zero weights.\n");
1341  MDB.createBranchWeights(Weights));
1342  ORE->emit([&]() {
1343  return OptimizationRemark(DEBUG_TYPE, "PopularDest", MaxDestInst)
1344  << "most popular destination for conditional branches at "
1345  << ore::NV("CondBranchesLoc", BranchLoc);
1346  });
1347  } else {
1348  LLVM_DEBUG(dbgs() << "SKIPPED. All branch weights are zero.\n");
1349  }
1350  }
1351 }
1352 
1353 /// Get the line number for the function header.
1354 ///
1355 /// This looks up function \p F in the current compilation unit and
1356 /// retrieves the line number where the function is defined. This is
1357 /// line 0 for all the samples read from the profile file. Every line
1358 /// number is relative to this line.
1359 ///
1360 /// \param F Function object to query.
1361 ///
1362 /// \returns the line number where \p F is defined. If it returns 0,
1363 /// it means that there is no debug information available for \p F.
1364 unsigned SampleProfileLoader::getFunctionLoc(Function &F) {
1365  if (DISubprogram *S = F.getSubprogram())
1366  return S->getLine();
1367 
1368  if (NoWarnSampleUnused)
1369  return 0;
1370 
1371  // If the start of \p F is missing, emit a diagnostic to inform the user
1372  // about the missed opportunity.
1374  "No debug information found in function " + F.getName() +
1375  ": Function profile not used",
1376  DS_Warning));
1377  return 0;
1378 }
1379 
1380 void SampleProfileLoader::computeDominanceAndLoopInfo(Function &F) {
1381  DT.reset(new DominatorTree);
1382  DT->recalculate(F);
1383 
1384  PDT.reset(new PostDominatorTree(F));
1385 
1386  LI.reset(new LoopInfo);
1387  LI->analyze(*DT);
1388 }
1389 
1390 /// Generate branch weight metadata for all branches in \p F.
1391 ///
1392 /// Branch weights are computed out of instruction samples using a
1393 /// propagation heuristic. Propagation proceeds in 3 phases:
1394 ///
1395 /// 1- Assignment of block weights. All the basic blocks in the function
1396 /// are initial assigned the same weight as their most frequently
1397 /// executed instruction.
1398 ///
1399 /// 2- Creation of equivalence classes. Since samples may be missing from
1400 /// blocks, we can fill in the gaps by setting the weights of all the
1401 /// blocks in the same equivalence class to the same weight. To compute
1402 /// the concept of equivalence, we use dominance and loop information.
1403 /// Two blocks B1 and B2 are in the same equivalence class if B1
1404 /// dominates B2, B2 post-dominates B1 and both are in the same loop.
1405 ///
1406 /// 3- Propagation of block weights into edges. This uses a simple
1407 /// propagation heuristic. The following rules are applied to every
1408 /// block BB in the CFG:
1409 ///
1410 /// - If BB has a single predecessor/successor, then the weight
1411 /// of that edge is the weight of the block.
1412 ///
1413 /// - If all the edges are known except one, and the weight of the
1414 /// block is already known, the weight of the unknown edge will
1415 /// be the weight of the block minus the sum of all the known
1416 /// edges. If the sum of all the known edges is larger than BB's weight,
1417 /// we set the unknown edge weight to zero.
1418 ///
1419 /// - If there is a self-referential edge, and the weight of the block is
1420 /// known, the weight for that edge is set to the weight of the block
1421 /// minus the weight of the other incoming edges to that block (if
1422 /// known).
1423 ///
1424 /// Since this propagation is not guaranteed to finalize for every CFG, we
1425 /// only allow it to proceed for a limited number of iterations (controlled
1426 /// by -sample-profile-max-propagate-iterations).
1427 ///
1428 /// FIXME: Try to replace this propagation heuristic with a scheme
1429 /// that is guaranteed to finalize. A work-list approach similar to
1430 /// the standard value propagation algorithm used by SSA-CCP might
1431 /// work here.
1432 ///
1433 /// Once all the branch weights are computed, we emit the MD_prof
1434 /// metadata on BB using the computed values for each of its branches.
1435 ///
1436 /// \param F The function to query.
1437 ///
1438 /// \returns true if \p F was modified. Returns false, otherwise.
1439 bool SampleProfileLoader::emitAnnotations(Function &F) {
1440  bool Changed = false;
1441 
1442  if (getFunctionLoc(F) == 0)
1443  return false;
1444 
1445  LLVM_DEBUG(dbgs() << "Line number for the first instruction in "
1446  << F.getName() << ": " << getFunctionLoc(F) << "\n");
1447 
1448  DenseSet<GlobalValue::GUID> InlinedGUIDs;
1449  Changed |= inlineHotFunctions(F, InlinedGUIDs);
1450 
1451  // Compute basic block weights.
1452  Changed |= computeBlockWeights(F);
1453 
1454  if (Changed) {
1455  // Add an entry count to the function using the samples gathered at the
1456  // function entry.
1457  // Sets the GUIDs that are inlined in the profiled binary. This is used
1458  // for ThinLink to make correct liveness analysis, and also make the IR
1459  // match the profiled binary before annotation.
1460  F.setEntryCount(
1461  ProfileCount(Samples->getHeadSamples() + 1, Function::PCT_Real),
1462  &InlinedGUIDs);
1463 
1464  // Compute dominance and loop info needed for propagation.
1465  computeDominanceAndLoopInfo(F);
1466 
1467  // Find equivalence classes.
1468  findEquivalenceClasses(F);
1469 
1470  // Propagate weights to all edges.
1471  propagateWeights(F);
1472  }
1473 
1474  // If coverage checking was requested, compute it now.
1476  unsigned Used = CoverageTracker.countUsedRecords(Samples, PSI);
1477  unsigned Total = CoverageTracker.countBodyRecords(Samples, PSI);
1478  unsigned Coverage = CoverageTracker.computeCoverage(Used, Total);
1479  if (Coverage < SampleProfileRecordCoverage) {
1481  F.getSubprogram()->getFilename(), getFunctionLoc(F),
1482  Twine(Used) + " of " + Twine(Total) + " available profile records (" +
1483  Twine(Coverage) + "%) were applied",
1484  DS_Warning));
1485  }
1486  }
1487 
1489  uint64_t Used = CoverageTracker.getTotalUsedSamples();
1490  uint64_t Total = CoverageTracker.countBodySamples(Samples, PSI);
1491  unsigned Coverage = CoverageTracker.computeCoverage(Used, Total);
1492  if (Coverage < SampleProfileSampleCoverage) {
1494  F.getSubprogram()->getFilename(), getFunctionLoc(F),
1495  Twine(Used) + " of " + Twine(Total) + " available profile samples (" +
1496  Twine(Coverage) + "%) were applied",
1497  DS_Warning));
1498  }
1499  }
1500  return Changed;
1501 }
1502 
1504 
1505 INITIALIZE_PASS_BEGIN(SampleProfileLoaderLegacyPass, "sample-profile",
1506  "Sample Profile loader", false, false)
1510 INITIALIZE_PASS_END(SampleProfileLoaderLegacyPass, "sample-profile",
1511  "Sample Profile loader", false, false)
1512 
1513 bool SampleProfileLoader::doInitialization(Module &M) {
1514  auto &Ctx = M.getContext();
1515  auto ReaderOrErr = SampleProfileReader::create(Filename, Ctx);
1516  if (std::error_code EC = ReaderOrErr.getError()) {
1517  std::string Msg = "Could not open profile: " + EC.message();
1518  Ctx.diagnose(DiagnosticInfoSampleProfile(Filename, Msg));
1519  return false;
1520  }
1521  Reader = std::move(ReaderOrErr.get());
1522  ProfileIsValid = (Reader->read() == sampleprof_error::success);
1523  return true;
1524 }
1525 
1527  return new SampleProfileLoaderLegacyPass(SampleProfileFile);
1528 }
1529 
1531  return new SampleProfileLoaderLegacyPass(Name);
1532 }
1533 
1534 bool SampleProfileLoader::runOnModule(Module &M, ModuleAnalysisManager *AM,
1535  ProfileSummaryInfo *_PSI) {
1536  if (!ProfileIsValid)
1537  return false;
1538 
1539  PSI = _PSI;
1540  if (M.getProfileSummary() == nullptr)
1541  M.setProfileSummary(Reader->getSummary().getMD(M.getContext()));
1542 
1543  // Compute the total number of samples collected in this profile.
1544  for (const auto &I : Reader->getProfiles())
1545  TotalCollectedSamples += I.second.getTotalSamples();
1546 
1547  // Populate the symbol map.
1548  for (const auto &N_F : M.getValueSymbolTable()) {
1549  StringRef OrigName = N_F.getKey();
1550  Function *F = dyn_cast<Function>(N_F.getValue());
1551  if (F == nullptr)
1552  continue;
1553  SymbolMap[OrigName] = F;
1554  auto pos = OrigName.find('.');
1555  if (pos != StringRef::npos) {
1556  StringRef NewName = OrigName.substr(0, pos);
1557  auto r = SymbolMap.insert(std::make_pair(NewName, F));
1558  // Failiing to insert means there is already an entry in SymbolMap,
1559  // thus there are multiple functions that are mapped to the same
1560  // stripped name. In this case of name conflicting, set the value
1561  // to nullptr to avoid confusion.
1562  if (!r.second)
1563  r.first->second = nullptr;
1564  }
1565  }
1566 
1567  bool retval = false;
1568  for (auto &F : M)
1569  if (!F.isDeclaration()) {
1570  clearFunctionData();
1571  retval |= runOnFunction(F, AM);
1572  }
1573  return retval;
1574 }
1575 
1576 bool SampleProfileLoaderLegacyPass::runOnModule(Module &M) {
1577  ACT = &getAnalysis<AssumptionCacheTracker>();
1578  TTIWP = &getAnalysis<TargetTransformInfoWrapperPass>();
1579  ProfileSummaryInfo *PSI =
1580  getAnalysis<ProfileSummaryInfoWrapperPass>().getPSI();
1581  return SampleLoader.runOnModule(M, nullptr, PSI);
1582 }
1583 
1585  // Initialize the entry count to -1, which will be treated conservatively
1586  // by getEntryCount as the same as unknown (None). If we have samples this
1587  // will be overwritten in emitAnnotations.
1589  std::unique_ptr<OptimizationRemarkEmitter> OwnedORE;
1590  if (AM) {
1591  auto &FAM =
1593  .getManager();
1594  ORE = &FAM.getResult<OptimizationRemarkEmitterAnalysis>(F);
1595  } else {
1596  OwnedORE = make_unique<OptimizationRemarkEmitter>(&F);
1597  ORE = OwnedORE.get();
1598  }
1599  Samples = Reader->getSamplesFor(F);
1600  if (Samples && !Samples->empty())
1601  return emitAnnotations(F);
1602  return false;
1603 }
1604 
1606  ModuleAnalysisManager &AM) {
1608  AM.getResult<FunctionAnalysisManagerModuleProxy>(M).getManager();
1609 
1610  auto GetAssumptionCache = [&](Function &F) -> AssumptionCache & {
1611  return FAM.getResult<AssumptionAnalysis>(F);
1612  };
1613  auto GetTTI = [&](Function &F) -> TargetTransformInfo & {
1614  return FAM.getResult<TargetIRAnalysis>(F);
1615  };
1616 
1617  SampleProfileLoader SampleLoader(
1618  ProfileFileName.empty() ? SampleProfileFile : ProfileFileName,
1619  IsThinLTOPreLink, GetAssumptionCache, GetTTI);
1620 
1621  SampleLoader.doInitialization(M);
1622 
1624  if (!SampleLoader.runOnModule(M, &AM, PSI))
1625  return PreservedAnalyses::all();
1626 
1627  return PreservedAnalyses::none();
1628 }
const NoneType None
Definition: None.h:24
uint64_t CallInst * C
const FunctionSamplesMap * findFunctionSamplesMapAt(const LineLocation &Loc) const
Returns the FunctionSamplesMap at the given Loc.
Definition: SampleProf.h:283
Thresholds to tune inline cost analysis.
Definition: InlineCost.h:155
Represents either an error or a value T.
Definition: ErrorOr.h:57
static PassRegistry * getPassRegistry()
getPassRegistry - Access the global registry object, which is automatically initialized at applicatio...
GCNRegPressure max(const GCNRegPressure &P1, const GCNRegPressure &P2)
DiagnosticInfoOptimizationBase::Argument NV
bool isNever() const
Definition: InlineCost.h:106
PassT::Result & getResult(IRUnitT &IR, ExtraArgTs... ExtraArgs)
Get the result of an analysis pass for a given IR unit.
Definition: PassManager.h:687
Compute iterated dominance frontiers using a linear time algorithm.
Definition: AllocatorList.h:24
sample profile
static StringRef getRepInFormat(StringRef Name, SampleProfileFormat Format, std::string &GUIDBuf)
Definition: SampleProf.h:97
A Module instance is used to store all the information related to an LLVM module. ...
Definition: Module.h:63
BasicBlock * getSuccessor(unsigned idx) const
Return the specified successor.
Implements a dense probed hash-table based set.
Definition: DenseSet.h:221
const ValueSymbolTable & getValueSymbolTable() const
Get the symbol table of global variable and function identifiers.
Definition: Module.h:554
Analysis providing profile information.
This class represents a function call, abstracting a target machine&#39;s calling convention.
An immutable pass that tracks lazily created AssumptionCache objects.
A cache of @llvm.assume calls within a function.
Analysis pass providing the TargetTransformInfo.
bool isLegalToPromote(CallSite CS, Function *Callee, const char **FailureReason=nullptr)
Return true if the given indirect call site can be made to call Callee.
static ErrorOr< std::unique_ptr< SampleProfileReader > > create(const Twine &Filename, LLVMContext &C)
Create a sample profile reader appropriate to the file format.
unsigned getLine() const
Definition: DebugLoc.cpp:26
ErrorOr< uint64_t > findSamplesAt(uint32_t LineOffset, uint32_t Discriminator) const
Return the number of samples collected at the given location.
Definition: SampleProf.h:256
std::map< SymbolStringPtr, JITEvaluatedSymbol > SymbolMap
A map from symbol names (as SymbolStringPtrs) to JITSymbols (address/flags pairs).
Definition: Core.h:51
A debug info location.
Definition: DebugLoc.h:34
F(f)
uint64_t getOrCompHotCountThreshold()
Returns HotCountThreshold if set.
InlineResult InlineFunction(CallInst *C, InlineFunctionInfo &IFI, AAResults *CalleeAAR=nullptr, bool InsertLifetime=true)
InlineFunction - This function inlines the called function into the basic block of the caller...
const StringRef & getName() const
Return the function name.
Definition: SampleProf.h:403
static SmallVector< InstrProfValueData, 2 > SortCallTargets(const SampleRecord::CallTargetMap &M)
Returns the sorted CallTargetMap M by count in descending order.
InlineFunctionInfo - This class captures the data input to the InlineFunction call, and records the auxiliary results produced by it.
Definition: Cloning.h:177
bool isHotCount(uint64_t C)
Returns true if F is a hot function.
Represents the cost of inlining a function.
Definition: InlineCost.h:65
PreservedAnalyses run(Module &M, ModuleAnalysisManager &AM)
Representation of the samples collected for a function.
Definition: SampleProf.h:216
AnalysisUsage & addRequired()
#define INITIALIZE_PASS_DEPENDENCY(depName)
Definition: PassSupport.h:51
Definition: BitVector.h:921
Twine - A lightweight data structure for efficiently representing the concatenation of temporary valu...
Definition: Twine.h:81
void setEntryCount(ProfileCount Count, const DenseSet< GlobalValue::GUID > *Imports=nullptr)
Set the entry count for this function.
Definition: Function.cpp:1352
An analysis pass based on legacy pass manager to deliver ProfileSummaryInfo.
LLVMContext & getContext() const
Get the global data context.
Definition: Module.h:242
Diagnostic information for optimization analysis remarks.
static StringRef getName(Value *V)
Metadata * getProfileSummary()
Returns profile summary metadata.
Definition: Module.cpp:512
StringRef getFilename() const
BlockT * getHeader() const
Definition: LoopInfo.h:100
Interval::succ_iterator succ_begin(Interval *I)
succ_begin/succ_end - define methods so that Intervals may be used just like BasicBlocks can with the...
Definition: Interval.h:103
Subprogram description.
const Instruction * getFirstNonPHIOrDbgOrLifetime() const
Returns a pointer to the first instruction in this block that is not a PHINode, a debug intrinsic...
Definition: BasicBlock.cpp:203
ArrayRef - Represent a constant reference to an array (0 or more elements consecutively in memory)...
Definition: APInt.h:33
bool extractProfTotalWeight(uint64_t &TotalVal) const
Retrieve total raw weight values of a branch.
Definition: Metadata.cpp:1340
static cl::opt< std::string > SampleProfileFile("sample-profile-file", cl::init(""), cl::value_desc("filename"), cl::desc("Profile file loaded by -sample-profile"), cl::Hidden)
LLVM_NODISCARD LLVM_ATTRIBUTE_ALWAYS_INLINE StringRef substr(size_t Start, size_t N=npos) const
Return a reference to the substring from [Start, Start + N).
Definition: StringRef.h:598
Debug location.
ModulePass * createSampleProfileLoaderPass()
Instruction * promoteIndirectCall(Instruction *Inst, Function *F, uint64_t Count, uint64_t TotalCount, bool AttachProfToDirectCall, OptimizationRemarkEmitter *ORE)
Concrete subclass of DominatorTreeBase that is used to compute a normal dominator tree...
Definition: Dominators.h:142
amdgpu Simplify well known AMD library false Value * Callee
static cl::opt< unsigned > SampleProfileRecordCoverage("sample-profile-check-record-coverage", cl::init(0), cl::value_desc("N"), cl::desc("Emit a warning if less than N% of records in the input profile " "are matched to the IR."))
static PreservedAnalyses none()
Convenience factory function for the empty preserved set.
Definition: PassManager.h:156
Interval::succ_iterator succ_end(Interval *I)
Definition: Interval.h:106
Core dominator tree base class.
Definition: LoopInfo.h:61
const BasicBlock & getEntryBlock() const
Definition: Function.h:626
static bool runOnFunction(Function &F, bool PostInlining)
initializer< Ty > init(const Ty &Val)
Definition: CommandLine.h:410
void dump(const SparseBitVector< ElementSize > &LHS, raw_ostream &out)
Subclasses of this class are all able to terminate a basic block.
Definition: InstrTypes.h:55
Wrapper pass for TargetTransformInfo.
A set of analyses that are preserved following a run of a transformation pass.
Definition: PassManager.h:153
* if(!EatIfPresent(lltok::kw_thread_local)) return false
ParseOptionalThreadLocal := /*empty.
void findInlinedFunctions(DenseSet< GlobalValue::GUID > &S, const Module *M, uint64_t Threshold, bool isCompact) const
Recursively traverses all children, if the total sample count of the corresponding function is no les...
Definition: SampleProf.h:379
LLVM Basic Block Representation.
Definition: BasicBlock.h:59
This is an important class for using LLVM in a threaded context.
Definition: LLVMContext.h:69
static cl::opt< unsigned > SampleProfileMaxPropagateIterations("sample-profile-max-propagate-iterations", cl::init(100), cl::desc("Maximum number of iterations to go through when propagating " "sample block/edge weights through the CFG."))
DISubprogram * getSubprogram() const
Get the attached subprogram.
Definition: Metadata.cpp:1508
An analysis pass based on the new PM to deliver ProfileSummaryInfo.
const BodySampleMap & getBodySamples() const
Return all the samples collected in the body of the function.
Definition: SampleProf.h:347
static GCRegistry::Add< CoreCLRGC > E("coreclr", "CoreCLR-compatible GC")
LLVM_ATTRIBUTE_ALWAYS_INLINE iterator begin()
Definition: SmallVector.h:129
SmallSet - This maintains a set of unique values, optimizing for the case when the set is small (less...
Definition: SmallSet.h:135
std::pair< iterator, bool > insert(PtrType Ptr)
Inserts Ptr if and only if there is no element in the container equal to Ptr.
Definition: SmallPtrSet.h:371
Diagnostic information for applied optimization remarks.
Interval::pred_iterator pred_begin(Interval *I)
pred_begin/pred_end - define methods so that Intervals may be used just like BasicBlocks can with the...
Definition: Interval.h:113
Function::ProfileCount ProfileCount
Represent the analysis usage information of a pass.
Interval::pred_iterator pred_end(Interval *I)
Definition: Interval.h:116
void initializeSampleProfileLoaderLegacyPassPass(PassRegistry &)
LLVMContext & getContext() const
getContext - Return a reference to the LLVMContext associated with this function. ...
Definition: Function.cpp:194
Used in the streaming interface as the general argument type.
void annotateValueSite(Module &M, Instruction &Inst, const InstrProfRecord &InstrProfR, InstrProfValueKind ValueKind, uint32_t SiteIndx, uint32_t MaxMDCount=3)
Get the value profile data for value site SiteIdx from InstrProfR and annotate the instruction Inst w...
Definition: InstrProf.cpp:812
static PreservedAnalyses all()
Construct a special preserved set that preserves all passes.
Definition: PassManager.h:159
Class to represent profile counts.
Definition: Function.h:247
size_t size() const
Definition: SmallVector.h:53
MDNode * createBranchWeights(uint32_t TrueWeight, uint32_t FalseWeight)
Return metadata containing two branch weights.
Definition: MDBuilder.cpp:38
INITIALIZE_PASS_END(RegBankSelect, DEBUG_TYPE, "Assign register bank of generic virtual registers", false, false) RegBankSelect
#define llvm_unreachable(msg)
Marks that the current location is not supposed to be reachable.
void setMetadata(unsigned KindID, MDNode *Node)
Set the metadata of the specified kind to the specified node.
Definition: Metadata.cpp:1226
InlineCost getInlineCost(CallSite CS, const InlineParams &Params, TargetTransformInfo &CalleeTTI, std::function< AssumptionCache &(Function &)> &GetAssumptionCache, Optional< function_ref< BlockFrequencyInfo &(Function &)>> GetBFI, ProfileSummaryInfo *PSI, OptimizationRemarkEmitter *ORE=nullptr)
Get an InlineCost object representing the cost of inlining this callsite.
Optional< bool > ComputeFullInlineCost
Compute inline cost even when the cost has exceeded the threshold.
Definition: InlineCost.h:182
void sort(IteratorTy Start, IteratorTy End)
Definition: STLExtras.h:969
const FunctionSamples * findFunctionSamplesAt(const LineLocation &Loc, StringRef CalleeName) const
Returns a pointer to FunctionSamples at the given callsite location Loc with callee CalleeName...
Definition: SampleProf.h:294
A function analysis which provides an AssumptionCache.
const InstListType & getInstList() const
Return the underlying instruction list container.
Definition: BasicBlock.h:329
SmallPtrSet - This class implements a set which is optimized for holding SmallSize or less elements...
Definition: SmallPtrSet.h:418
This pass provides access to the codegen interfaces that are needed for IR-level transformations.
InlineParams getInlineParams()
Generate the parameters to tune the inline cost analysis based only on the commandline options...
This is a &#39;vector&#39; (really, a variable-sized array), optimized for the case when the array is small...
Definition: SmallVector.h:847
Module.h This file contains the declarations for the Module class.
static unsigned getOffset(const DILocation *DIL)
Returns the line offset to the start line of the subprogram.
Definition: SampleProf.cpp:155
ErrorOr< SampleRecord::CallTargetMap > findCallTargetMapAt(uint32_t LineOffset, uint32_t Discriminator) const
Returns the call target map collected at a given location.
Definition: SampleProf.h:269
raw_ostream & dbgs()
dbgs() - This returns a reference to a raw_ostream for debugging messages.
Definition: Debug.cpp:133
StringMap - This is an unconventional map that is specialized for handling keys that are "strings"...
Definition: StringMap.h:220
PostDominatorTree Class - Concrete subclass of DominatorTree that is used to compute the post-dominat...
static void clear(coro::Shape &Shape)
Definition: Coroutines.cpp:211
iterator insert(iterator I, T &&Elt)
Definition: SmallVector.h:478
bool dominates(const DomTreeNodeBase< NodeT > *A, const DomTreeNodeBase< NodeT > *B) const
dominates - Returns true iff A dominates B.
void setProfileSummary(Metadata *M)
Attach profile summary metadata to this module.
Definition: Module.cpp:508
LLVM_ATTRIBUTE_ALWAYS_INLINE iterator end()
Definition: SmallVector.h:133
unsigned getBaseDiscriminator() const
Returns the base discriminator stored in the discriminator.
const DebugLoc & getDebugLoc() const
Return the debug location for this node as a DebugLoc.
Definition: Instruction.h:290
#define DEBUG_TYPE
static const size_t npos
Definition: StringRef.h:51
iterator begin()
Definition: StringMap.h:315
Represents the relative location of an instruction.
Definition: SampleProf.h:117
static cl::opt< unsigned > SampleProfileSampleCoverage("sample-profile-check-sample-coverage", cl::init(0), cl::value_desc("N"), cl::desc("Emit a warning if less than N% of samples in the input profile " "are matched to the IR."))
static bool callsiteIsHot(const FunctionSamples *CallsiteFS, ProfileSummaryInfo *PSI)
Return true if the given callsite is hot wrt to hot cutoff threshold.
Represents a single loop in the control flow graph.
Definition: LoopInfo.h:459
StringRef getName() const
Return a constant reference to the value&#39;s name.
Definition: Value.cpp:224
Establish a view to a call site for examination.
Definition: CallSite.h:714
const Function * getParent() const
Return the enclosing method, or null if none.
Definition: BasicBlock.h:108
#define I(x, y, z)
Definition: MD5.cpp:58
ModulePass class - This class is used to implement unstructured interprocedural optimizations and ana...
Definition: Pass.h:225
LLVM_NODISCARD std::enable_if<!is_simple_type< Y >::value, typename cast_retty< X, const Y >::ret_type >::type dyn_cast(const Y &Val)
Definition: Casting.h:323
std::map< std::string, FunctionSamples > FunctionSamplesMap
Definition: SampleProf.h:208
size_type count(const_arg_type_t< ValueT > V) const
Return 1 if the specified key is in the set, 0 otherwise.
Definition: DenseSet.h:91
void diagnose(const DiagnosticInfo &DI)
Report a message to the currently installed diagnostic handler.
INITIALIZE_PASS_BEGIN(SampleProfileLoaderLegacyPass, "sample-profile", "Sample Profile loader", false, false) INITIALIZE_PASS_END(SampleProfileLoaderLegacyPass
bool isDeclaration() const
Return true if the primary definition of this global value is outside of the current translation unit...
Definition: Globals.cpp:201
Provides ErrorOr<T> smart pointer.
FunTy * getCalledFunction() const
Return the function being called if this is a direct call, otherwise return null (if it&#39;s an indirect...
Definition: CallSite.h:107
assert(ImpDefSCC.getReg()==AMDGPU::SCC &&ImpDefSCC.isDef())
unsigned getNumSuccessors() const
Return the number of successors that this terminator has.
Module * getParent()
Get the module that this global value is contained inside of...
Definition: GlobalValue.h:565
uint64_t getTotalSamples() const
Return the total number of samples collected inside the function.
Definition: SampleProf.h:317
sample Sample Profile loader
const CallsiteSampleMap & getCallsiteSamples() const
Return all the callsite samples collected in the body of the function.
Definition: SampleProf.h:350
This class implements an extremely fast bulk output stream that can only output to a stream...
Definition: raw_ostream.h:46
print Print MemDeps of function
This file defines a set of templates that efficiently compute a dominator tree over a generic graph...
StringRef - Represent a constant reference to a string, i.e.
Definition: StringRef.h:49
A container for analyses that lazily runs them and caches their results.
This pass exposes codegen information to IR-level passes.
uint64_t getEntrySamples() const
Return the sample count of the first instruction of the function.
Definition: SampleProf.h:328
static cl::opt< bool > NoWarnSampleUnused("no-warn-sample-unused", cl::init(false), cl::Hidden, cl::desc("Use this option to turn off/on warnings about function with " "samples but without debug information to use those samples. "))
This header defines various interfaces for pass management in LLVM.
Diagnostic information for the sample profiler.
LLVM_NODISCARD LLVM_ATTRIBUTE_ALWAYS_INLINE size_t find(char C, size_t From=0) const
Search for the first character C in the string.
Definition: StringRef.h:298
#define LLVM_DEBUG(X)
Definition: Debug.h:119
The optimization diagnostic interface.
iterator end()
Definition: StringMap.h:318
This file provides the interface for the sampled PGO loader pass.
reference get()
Definition: ErrorOr.h:157
const BasicBlock * getParent() const
Definition: Instruction.h:67
An analysis over an "outer" IR unit that provides access to an analysis manager over an "inner" IR un...
Definition: PassManager.h:946