LLVM  7.0.0svn
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"
41 #include "llvm/IR/BasicBlock.h"
42 #include "llvm/IR/CFG.h"
43 #include "llvm/IR/CallSite.h"
45 #include "llvm/IR/DebugLoc.h"
46 #include "llvm/IR/DiagnosticInfo.h"
47 #include "llvm/IR/Dominators.h"
48 #include "llvm/IR/Function.h"
49 #include "llvm/IR/GlobalValue.h"
50 #include "llvm/IR/InstrTypes.h"
51 #include "llvm/IR/Instruction.h"
52 #include "llvm/IR/Instructions.h"
53 #include "llvm/IR/IntrinsicInst.h"
54 #include "llvm/IR/LLVMContext.h"
55 #include "llvm/IR/MDBuilder.h"
56 #include "llvm/IR/Module.h"
57 #include "llvm/IR/PassManager.h"
59 #include "llvm/Pass.h"
63 #include "llvm/Support/Casting.h"
65 #include "llvm/Support/Debug.h"
67 #include "llvm/Support/ErrorOr.h"
70 #include "llvm/Transforms/IPO.h"
74 #include <algorithm>
75 #include <cassert>
76 #include <cstdint>
77 #include <functional>
78 #include <limits>
79 #include <map>
80 #include <memory>
81 #include <string>
82 #include <system_error>
83 #include <utility>
84 #include <vector>
85 
86 using namespace llvm;
87 using namespace sampleprof;
89 #define DEBUG_TYPE "sample-profile"
90 
91 // Command line option to specify the file to read samples from. This is
92 // mainly used for debugging.
94  "sample-profile-file", cl::init(""), cl::value_desc("filename"),
95  cl::desc("Profile file loaded by -sample-profile"), cl::Hidden);
96 
98  "sample-profile-max-propagate-iterations", cl::init(100),
99  cl::desc("Maximum number of iterations to go through when propagating "
100  "sample block/edge weights through the CFG."));
101 
103  "sample-profile-check-record-coverage", cl::init(0), cl::value_desc("N"),
104  cl::desc("Emit a warning if less than N% of records in the input profile "
105  "are matched to the IR."));
106 
108  "sample-profile-check-sample-coverage", cl::init(0), cl::value_desc("N"),
109  cl::desc("Emit a warning if less than N% of samples in the input profile "
110  "are matched to the IR."));
111 
113  "sample-profile-inline-hot-threshold", cl::init(0.1), cl::value_desc("N"),
114  cl::desc("Inlined functions that account for more than N% of all samples "
115  "collected in the parent function, will be inlined again."));
116 
117 namespace {
118 
119 using BlockWeightMap = DenseMap<const BasicBlock *, uint64_t>;
120 using EquivalenceClassMap = DenseMap<const BasicBlock *, const BasicBlock *>;
121 using Edge = std::pair<const BasicBlock *, const BasicBlock *>;
122 using EdgeWeightMap = DenseMap<Edge, uint64_t>;
123 using BlockEdgeMap =
125 
126 class SampleCoverageTracker {
127 public:
128  SampleCoverageTracker() = default;
129 
130  bool markSamplesUsed(const FunctionSamples *FS, uint32_t LineOffset,
131  uint32_t Discriminator, uint64_t Samples);
132  unsigned computeCoverage(unsigned Used, unsigned Total) const;
133  unsigned countUsedRecords(const FunctionSamples *FS) const;
134  unsigned countBodyRecords(const FunctionSamples *FS) const;
135  uint64_t getTotalUsedSamples() const { return TotalUsedSamples; }
136  uint64_t countBodySamples(const FunctionSamples *FS) const;
137 
138  void clear() {
139  SampleCoverage.clear();
140  TotalUsedSamples = 0;
141  }
142 
143 private:
144  using BodySampleCoverageMap = std::map<LineLocation, unsigned>;
145  using FunctionSamplesCoverageMap =
147 
148  /// Coverage map for sampling records.
149  ///
150  /// This map keeps a record of sampling records that have been matched to
151  /// an IR instruction. This is used to detect some form of staleness in
152  /// profiles (see flag -sample-profile-check-coverage).
153  ///
154  /// Each entry in the map corresponds to a FunctionSamples instance. This is
155  /// another map that counts how many times the sample record at the
156  /// given location has been used.
157  FunctionSamplesCoverageMap SampleCoverage;
158 
159  /// Number of samples used from the profile.
160  ///
161  /// When a sampling record is used for the first time, the samples from
162  /// that record are added to this accumulator. Coverage is later computed
163  /// based on the total number of samples available in this function and
164  /// its callsites.
165  ///
166  /// Note that this accumulator tracks samples used from a single function
167  /// and all the inlined callsites. Strictly, we should have a map of counters
168  /// keyed by FunctionSamples pointers, but these stats are cleared after
169  /// every function, so we just need to keep a single counter.
170  uint64_t TotalUsedSamples = 0;
171 };
172 
173 /// \brief Sample profile pass.
174 ///
175 /// This pass reads profile data from the file specified by
176 /// -sample-profile-file and annotates every affected function with the
177 /// profile information found in that file.
178 class SampleProfileLoader {
179 public:
180  SampleProfileLoader(
181  StringRef Name, bool IsThinLTOPreLink,
182  std::function<AssumptionCache &(Function &)> GetAssumptionCache,
183  std::function<TargetTransformInfo &(Function &)> GetTargetTransformInfo)
184  : GetAC(std::move(GetAssumptionCache)),
185  GetTTI(std::move(GetTargetTransformInfo)), Filename(Name),
186  IsThinLTOPreLink(IsThinLTOPreLink) {}
187 
188  bool doInitialization(Module &M);
189  bool runOnModule(Module &M, ModuleAnalysisManager *AM);
190 
191  void dump() { Reader->dump(); }
192 
193 protected:
195  unsigned getFunctionLoc(Function &F);
196  bool emitAnnotations(Function &F);
197  ErrorOr<uint64_t> getInstWeight(const Instruction &I);
198  ErrorOr<uint64_t> getBlockWeight(const BasicBlock *BB);
199  const FunctionSamples *findCalleeFunctionSamples(const Instruction &I) const;
200  std::vector<const FunctionSamples *>
201  findIndirectCallFunctionSamples(const Instruction &I, uint64_t &Sum) const;
202  const FunctionSamples *findFunctionSamples(const Instruction &I) const;
203  bool inlineCallInstruction(Instruction *I);
204  bool inlineHotFunctions(Function &F,
205  DenseSet<GlobalValue::GUID> &InlinedGUIDs);
206  void printEdgeWeight(raw_ostream &OS, Edge E);
207  void printBlockWeight(raw_ostream &OS, const BasicBlock *BB) const;
208  void printBlockEquivalence(raw_ostream &OS, const BasicBlock *BB);
209  bool computeBlockWeights(Function &F);
210  void findEquivalenceClasses(Function &F);
211  template <bool IsPostDom>
212  void findEquivalencesFor(BasicBlock *BB1, ArrayRef<BasicBlock *> Descendants,
214 
215  void propagateWeights(Function &F);
216  uint64_t visitEdge(Edge E, unsigned *NumUnknownEdges, Edge *UnknownEdge);
217  void buildEdges(Function &F);
218  bool propagateThroughEdges(Function &F, bool UpdateBlockCount);
219  void computeDominanceAndLoopInfo(Function &F);
220  unsigned getOffset(const DILocation *DIL) const;
221  void clearFunctionData();
222 
223  /// \brief Map basic blocks to their computed weights.
224  ///
225  /// The weight of a basic block is defined to be the maximum
226  /// of all the instruction weights in that block.
227  BlockWeightMap BlockWeights;
228 
229  /// \brief Map edges to their computed weights.
230  ///
231  /// Edge weights are computed by propagating basic block weights in
232  /// SampleProfile::propagateWeights.
233  EdgeWeightMap EdgeWeights;
234 
235  /// \brief Set of visited blocks during propagation.
237 
238  /// \brief Set of visited edges during propagation.
239  SmallSet<Edge, 32> VisitedEdges;
240 
241  /// \brief Equivalence classes for block weights.
242  ///
243  /// Two blocks BB1 and BB2 are in the same equivalence class if they
244  /// dominate and post-dominate each other, and they are in the same loop
245  /// nest. When this happens, the two blocks are guaranteed to execute
246  /// the same number of times.
247  EquivalenceClassMap EquivalenceClass;
248 
249  /// Map from function name to Function *. Used to find the function from
250  /// the function name. If the function name contains suffix, additional
251  /// entry is added to map from the stripped name to the function if there
252  /// is one-to-one mapping.
254 
255  /// \brief Dominance, post-dominance and loop information.
256  std::unique_ptr<DominatorTree> DT;
257  std::unique_ptr<PostDomTreeBase<BasicBlock>> PDT;
258  std::unique_ptr<LoopInfo> LI;
259 
260  std::function<AssumptionCache &(Function &)> GetAC;
261  std::function<TargetTransformInfo &(Function &)> GetTTI;
262 
263  /// \brief Predecessors for each basic block in the CFG.
264  BlockEdgeMap Predecessors;
265 
266  /// \brief Successors for each basic block in the CFG.
267  BlockEdgeMap Successors;
268 
269  SampleCoverageTracker CoverageTracker;
270 
271  /// \brief Profile reader object.
272  std::unique_ptr<SampleProfileReader> Reader;
273 
274  /// \brief Samples collected for the body of this function.
275  FunctionSamples *Samples = nullptr;
276 
277  /// \brief Name of the profile file to load.
278  std::string Filename;
279 
280  /// \brief Flag indicating whether the profile input loaded successfully.
281  bool ProfileIsValid = false;
282 
283  /// \brief Flag indicating if the pass is invoked in ThinLTO compile phase.
284  ///
285  /// In this phase, in annotation, we should not promote indirect calls.
286  /// Instead, we will mark GUIDs that needs to be annotated to the function.
287  bool IsThinLTOPreLink;
288 
289  /// \brief Total number of samples collected in this profile.
290  ///
291  /// This is the sum of all the samples collected in all the functions executed
292  /// at runtime.
293  uint64_t TotalCollectedSamples = 0;
294 
295  /// \brief Optimization Remark Emitter used to emit diagnostic remarks.
296  OptimizationRemarkEmitter *ORE = nullptr;
297 };
298 
299 class SampleProfileLoaderLegacyPass : public ModulePass {
300 public:
301  // Class identification, replacement for typeinfo
302  static char ID;
303 
304  SampleProfileLoaderLegacyPass(StringRef Name = SampleProfileFile,
305  bool IsThinLTOPreLink = false)
306  : ModulePass(ID), SampleLoader(Name, IsThinLTOPreLink,
307  [&](Function &F) -> AssumptionCache & {
308  return ACT->getAssumptionCache(F);
309  },
310  [&](Function &F) -> TargetTransformInfo & {
311  return TTIWP->getTTI(F);
312  }) {
315  }
316 
317  void dump() { SampleLoader.dump(); }
318 
319  bool doInitialization(Module &M) override {
320  return SampleLoader.doInitialization(M);
321  }
322 
323  StringRef getPassName() const override { return "Sample profile pass"; }
324  bool runOnModule(Module &M) override;
325 
326  void getAnalysisUsage(AnalysisUsage &AU) const override {
329  }
330 
331 private:
332  SampleProfileLoader SampleLoader;
333  AssumptionCacheTracker *ACT = nullptr;
334  TargetTransformInfoWrapperPass *TTIWP = nullptr;
335 };
336 
337 } // end anonymous namespace
338 
339 /// Return true if the given callsite is hot wrt to its caller.
340 ///
341 /// Functions that were inlined in the original binary will be represented
342 /// in the inline stack in the sample profile. If the profile shows that
343 /// the original inline decision was "good" (i.e., the callsite is executed
344 /// frequently), then we will recreate the inline decision and apply the
345 /// profile from the inlined callsite.
346 ///
347 /// To decide whether an inlined callsite is hot, we compute the fraction
348 /// of samples used by the callsite with respect to the total number of samples
349 /// collected in the caller.
350 ///
351 /// If that fraction is larger than the default given by
352 /// SampleProfileHotThreshold, the callsite will be inlined again.
353 static bool callsiteIsHot(const FunctionSamples *CallerFS,
354  const FunctionSamples *CallsiteFS) {
355  if (!CallsiteFS)
356  return false; // The callsite was not inlined in the original binary.
357 
358  uint64_t ParentTotalSamples = CallerFS->getTotalSamples();
359  if (ParentTotalSamples == 0)
360  return false; // Avoid division by zero.
361 
362  uint64_t CallsiteTotalSamples = CallsiteFS->getTotalSamples();
363  if (CallsiteTotalSamples == 0)
364  return false; // Callsite is trivially cold.
365 
366  double PercentSamples =
367  (double)CallsiteTotalSamples / (double)ParentTotalSamples * 100.0;
368  return PercentSamples >= SampleProfileHotThreshold;
369 }
370 
371 /// Mark as used the sample record for the given function samples at
372 /// (LineOffset, Discriminator).
373 ///
374 /// \returns true if this is the first time we mark the given record.
375 bool SampleCoverageTracker::markSamplesUsed(const FunctionSamples *FS,
376  uint32_t LineOffset,
377  uint32_t Discriminator,
378  uint64_t Samples) {
379  LineLocation Loc(LineOffset, Discriminator);
380  unsigned &Count = SampleCoverage[FS][Loc];
381  bool FirstTime = (++Count == 1);
382  if (FirstTime)
383  TotalUsedSamples += Samples;
384  return FirstTime;
385 }
386 
387 /// Return the number of sample records that were applied from this profile.
388 ///
389 /// This count does not include records from cold inlined callsites.
390 unsigned
391 SampleCoverageTracker::countUsedRecords(const FunctionSamples *FS) const {
392  auto I = SampleCoverage.find(FS);
393 
394  // The size of the coverage map for FS represents the number of records
395  // that were marked used at least once.
396  unsigned Count = (I != SampleCoverage.end()) ? I->second.size() : 0;
397 
398  // If there are inlined callsites in this function, count the samples found
399  // in the respective bodies. However, do not bother counting callees with 0
400  // total samples, these are callees that were never invoked at runtime.
401  for (const auto &I : FS->getCallsiteSamples())
402  for (const auto &J : I.second) {
403  const FunctionSamples *CalleeSamples = &J.second;
404  if (callsiteIsHot(FS, CalleeSamples))
405  Count += countUsedRecords(CalleeSamples);
406  }
407 
408  return Count;
409 }
410 
411 /// Return the number of sample records in the body of this profile.
412 ///
413 /// This count does not include records from cold inlined callsites.
414 unsigned
415 SampleCoverageTracker::countBodyRecords(const FunctionSamples *FS) 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(FS, CalleeSamples))
423  Count += countBodyRecords(CalleeSamples);
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) const {
434  uint64_t Total = 0;
435  for (const auto &I : FS->getBodySamples())
436  Total += I.second.getSamples();
437 
438  // Only count samples in hot callsites.
439  for (const auto &I : FS->getCallsiteSamples())
440  for (const auto &J : I.second) {
441  const FunctionSamples *CalleeSamples = &J.second;
442  if (callsiteIsHot(FS, CalleeSamples))
443  Total += countBodySamples(CalleeSamples);
444  }
445 
446  return Total;
447 }
448 
449 /// Return the fraction of sample records used in this profile.
450 ///
451 /// The returned value is an unsigned integer in the range 0-100 indicating
452 /// the percentage of sample records that were used while applying this
453 /// profile to the associated function.
454 unsigned SampleCoverageTracker::computeCoverage(unsigned Used,
455  unsigned Total) const {
456  assert(Used <= Total &&
457  "number of used records cannot exceed the total number of records");
458  return Total > 0 ? Used * 100 / Total : 100;
459 }
460 
461 /// Clear all the per-function data used to load samples and propagate weights.
462 void SampleProfileLoader::clearFunctionData() {
463  BlockWeights.clear();
464  EdgeWeights.clear();
465  VisitedBlocks.clear();
466  VisitedEdges.clear();
467  EquivalenceClass.clear();
468  DT = nullptr;
469  PDT = nullptr;
470  LI = nullptr;
471  Predecessors.clear();
472  Successors.clear();
473  CoverageTracker.clear();
474 }
475 
476 /// Returns the line offset to the start line of the subprogram.
477 /// We assume that a single function will not exceed 65535 LOC.
478 unsigned SampleProfileLoader::getOffset(const DILocation *DIL) const {
479  return (DIL->getLine() - DIL->getScope()->getSubprogram()->getLine()) &
480  0xffff;
481 }
482 
483 #ifndef NDEBUG
484 /// \brief Print the weight of edge \p E on stream \p OS.
485 ///
486 /// \param OS Stream to emit the output to.
487 /// \param E Edge to print.
488 void SampleProfileLoader::printEdgeWeight(raw_ostream &OS, Edge E) {
489  OS << "weight[" << E.first->getName() << "->" << E.second->getName()
490  << "]: " << EdgeWeights[E] << "\n";
491 }
492 
493 /// \brief Print the equivalence class of block \p BB on stream \p OS.
494 ///
495 /// \param OS Stream to emit the output to.
496 /// \param BB Block to print.
497 void SampleProfileLoader::printBlockEquivalence(raw_ostream &OS,
498  const BasicBlock *BB) {
499  const BasicBlock *Equiv = EquivalenceClass[BB];
500  OS << "equivalence[" << BB->getName()
501  << "]: " << ((Equiv) ? EquivalenceClass[BB]->getName() : "NONE") << "\n";
502 }
503 
504 /// \brief Print the weight of block \p BB on stream \p OS.
505 ///
506 /// \param OS Stream to emit the output to.
507 /// \param BB Block to print.
508 void SampleProfileLoader::printBlockWeight(raw_ostream &OS,
509  const BasicBlock *BB) const {
510  const auto &I = BlockWeights.find(BB);
511  uint64_t W = (I == BlockWeights.end() ? 0 : I->second);
512  OS << "weight[" << BB->getName() << "]: " << W << "\n";
513 }
514 #endif
515 
516 /// \brief Get the weight for an instruction.
517 ///
518 /// The "weight" of an instruction \p Inst is the number of samples
519 /// collected on that instruction at runtime. To retrieve it, we
520 /// need to compute the line number of \p Inst relative to the start of its
521 /// function. We use HeaderLineno to compute the offset. We then
522 /// look up the samples collected for \p Inst using BodySamples.
523 ///
524 /// \param Inst Instruction to query.
525 ///
526 /// \returns the weight of \p Inst.
527 ErrorOr<uint64_t> SampleProfileLoader::getInstWeight(const Instruction &Inst) {
528  const DebugLoc &DLoc = Inst.getDebugLoc();
529  if (!DLoc)
530  return std::error_code();
531 
532  const FunctionSamples *FS = findFunctionSamples(Inst);
533  if (!FS)
534  return std::error_code();
535 
536  // Ignore all intrinsics and branch instructions.
537  // Branch instruction usually contains debug info from sources outside of
538  // the residing basic block, thus we ignore them during annotation.
539  if (isa<BranchInst>(Inst) || isa<IntrinsicInst>(Inst))
540  return std::error_code();
541 
542  // If a direct call/invoke instruction is inlined in profile
543  // (findCalleeFunctionSamples returns non-empty result), but not inlined here,
544  // it means that the inlined callsite has no sample, thus the call
545  // instruction should have 0 count.
546  if ((isa<CallInst>(Inst) || isa<InvokeInst>(Inst)) &&
547  !ImmutableCallSite(&Inst).isIndirectCall() &&
548  findCalleeFunctionSamples(Inst))
549  return 0;
550 
551  const DILocation *DIL = DLoc;
552  uint32_t LineOffset = getOffset(DIL);
553  uint32_t Discriminator = DIL->getBaseDiscriminator();
554  ErrorOr<uint64_t> R = FS->findSamplesAt(LineOffset, Discriminator);
555  if (R) {
556  bool FirstMark =
557  CoverageTracker.markSamplesUsed(FS, LineOffset, Discriminator, R.get());
558  if (FirstMark) {
559  ORE->emit([&]() {
560  OptimizationRemarkAnalysis Remark(DEBUG_TYPE, "AppliedSamples", &Inst);
561  Remark << "Applied " << ore::NV("NumSamples", *R);
562  Remark << " samples from profile (offset: ";
563  Remark << ore::NV("LineOffset", LineOffset);
564  if (Discriminator) {
565  Remark << ".";
566  Remark << ore::NV("Discriminator", Discriminator);
567  }
568  Remark << ")";
569  return Remark;
570  });
571  }
572  DEBUG(dbgs() << " " << DLoc.getLine() << "."
573  << DIL->getBaseDiscriminator() << ":" << Inst
574  << " (line offset: " << LineOffset << "."
575  << DIL->getBaseDiscriminator() << " - weight: " << R.get()
576  << ")\n");
577  }
578  return R;
579 }
580 
581 /// \brief Compute the weight of a basic block.
582 ///
583 /// The weight of basic block \p BB is the maximum weight of all the
584 /// instructions in BB.
585 ///
586 /// \param BB The basic block to query.
587 ///
588 /// \returns the weight for \p BB.
589 ErrorOr<uint64_t> SampleProfileLoader::getBlockWeight(const BasicBlock *BB) {
590  uint64_t Max = 0;
591  bool HasWeight = false;
592  for (auto &I : BB->getInstList()) {
593  const ErrorOr<uint64_t> &R = getInstWeight(I);
594  if (R) {
595  Max = std::max(Max, R.get());
596  HasWeight = true;
597  }
598  }
599  return HasWeight ? ErrorOr<uint64_t>(Max) : std::error_code();
600 }
601 
602 /// \brief Compute and store the weights of every basic block.
603 ///
604 /// This populates the BlockWeights map by computing
605 /// the weights of every basic block in the CFG.
606 ///
607 /// \param F The function to query.
608 bool SampleProfileLoader::computeBlockWeights(Function &F) {
609  bool Changed = false;
610  DEBUG(dbgs() << "Block weights\n");
611  for (const auto &BB : F) {
612  ErrorOr<uint64_t> Weight = getBlockWeight(&BB);
613  if (Weight) {
614  BlockWeights[&BB] = Weight.get();
615  VisitedBlocks.insert(&BB);
616  Changed = true;
617  }
618  DEBUG(printBlockWeight(dbgs(), &BB));
619  }
620 
621  return Changed;
622 }
623 
624 /// \brief Get the FunctionSamples for a call instruction.
625 ///
626 /// The FunctionSamples of a call/invoke instruction \p Inst is the inlined
627 /// instance in which that call instruction is calling to. It contains
628 /// all samples that resides in the inlined instance. We first find the
629 /// inlined instance in which the call instruction is from, then we
630 /// traverse its children to find the callsite with the matching
631 /// location.
632 ///
633 /// \param Inst Call/Invoke instruction to query.
634 ///
635 /// \returns The FunctionSamples pointer to the inlined instance.
636 const FunctionSamples *
637 SampleProfileLoader::findCalleeFunctionSamples(const Instruction &Inst) const {
638  const DILocation *DIL = Inst.getDebugLoc();
639  if (!DIL) {
640  return nullptr;
641  }
642 
643  StringRef CalleeName;
644  if (const CallInst *CI = dyn_cast<CallInst>(&Inst))
645  if (Function *Callee = CI->getCalledFunction())
646  CalleeName = Callee->getName();
647 
648  const FunctionSamples *FS = findFunctionSamples(Inst);
649  if (FS == nullptr)
650  return nullptr;
651 
652  return FS->findFunctionSamplesAt(
653  LineLocation(getOffset(DIL), DIL->getBaseDiscriminator()), CalleeName);
654 }
655 
656 /// Returns a vector of FunctionSamples that are the indirect call targets
657 /// of \p Inst. The vector is sorted by the total number of samples. Stores
658 /// the total call count of the indirect call in \p Sum.
659 std::vector<const FunctionSamples *>
660 SampleProfileLoader::findIndirectCallFunctionSamples(
661  const Instruction &Inst, uint64_t &Sum) const {
662  const DILocation *DIL = Inst.getDebugLoc();
663  std::vector<const FunctionSamples *> R;
664 
665  if (!DIL) {
666  return R;
667  }
668 
669  const FunctionSamples *FS = findFunctionSamples(Inst);
670  if (FS == nullptr)
671  return R;
672 
673  uint32_t LineOffset = getOffset(DIL);
674  uint32_t Discriminator = DIL->getBaseDiscriminator();
675 
676  auto T = FS->findCallTargetMapAt(LineOffset, Discriminator);
677  Sum = 0;
678  if (T)
679  for (const auto &T_C : T.get())
680  Sum += T_C.second;
683  if (M->empty())
684  return R;
685  for (const auto &NameFS : *M) {
686  Sum += NameFS.second.getEntrySamples();
687  R.push_back(&NameFS.second);
688  }
689  std::sort(R.begin(), R.end(),
690  [](const FunctionSamples *L, const FunctionSamples *R) {
691  return L->getEntrySamples() > R->getEntrySamples();
692  });
693  }
694  return R;
695 }
696 
697 /// \brief Get the FunctionSamples for an instruction.
698 ///
699 /// The FunctionSamples of an instruction \p Inst is the inlined instance
700 /// in which that instruction is coming from. We traverse the inline stack
701 /// of that instruction, and match it with the tree nodes in the profile.
702 ///
703 /// \param Inst Instruction to query.
704 ///
705 /// \returns the FunctionSamples pointer to the inlined instance.
706 const FunctionSamples *
707 SampleProfileLoader::findFunctionSamples(const Instruction &Inst) const {
709  const DILocation *DIL = Inst.getDebugLoc();
710  if (!DIL)
711  return Samples;
712 
713  const DILocation *PrevDIL = DIL;
714  for (DIL = DIL->getInlinedAt(); DIL; DIL = DIL->getInlinedAt()) {
715  S.push_back(std::make_pair(
717  PrevDIL->getScope()->getSubprogram()->getLinkageName()));
718  PrevDIL = DIL;
719  }
720  if (S.size() == 0)
721  return Samples;
722  const FunctionSamples *FS = Samples;
723  for (int i = S.size() - 1; i >= 0 && FS != nullptr; i--) {
724  FS = FS->findFunctionSamplesAt(S[i].first, S[i].second);
725  }
726  return FS;
727 }
728 
729 bool SampleProfileLoader::inlineCallInstruction(Instruction *I) {
730  assert(isa<CallInst>(I) || isa<InvokeInst>(I));
731  CallSite CS(I);
732  Function *CalledFunction = CS.getCalledFunction();
733  assert(CalledFunction);
734  DebugLoc DLoc = I->getDebugLoc();
735  BasicBlock *BB = I->getParent();
736  InlineParams Params = getInlineParams();
737  Params.ComputeFullInlineCost = true;
738  // Checks if there is anything in the reachable portion of the callee at
739  // this callsite that makes this inlining potentially illegal. Need to
740  // set ComputeFullInlineCost, otherwise getInlineCost may return early
741  // when cost exceeds threshold without checking all IRs in the callee.
742  // The acutal cost does not matter because we only checks isNever() to
743  // see if it is legal to inline the callsite.
744  InlineCost Cost = getInlineCost(CS, Params, GetTTI(*CalledFunction), GetAC,
745  None, nullptr, nullptr);
746  if (Cost.isNever()) {
747  ORE->emit(OptimizationRemark(DEBUG_TYPE, "Not inline", DLoc, BB)
748  << "incompatible inlining");
749  return false;
750  }
751  InlineFunctionInfo IFI(nullptr, &GetAC);
752  if (InlineFunction(CS, IFI)) {
753  // The call to InlineFunction erases I, so we can't pass it here.
754  ORE->emit(OptimizationRemark(DEBUG_TYPE, "HotInline", DLoc, BB)
755  << "inlined hot callee '" << ore::NV("Callee", CalledFunction)
756  << "' into '" << ore::NV("Caller", BB->getParent()) << "'");
757  return true;
758  }
759  return false;
760 }
761 
762 /// \brief Iteratively inline hot callsites of a function.
763 ///
764 /// Iteratively traverse all callsites of the function \p F, and find if
765 /// the corresponding inlined instance exists and is hot in profile. If
766 /// it is hot enough, inline the callsites and adds new callsites of the
767 /// callee into the caller. If the call is an indirect call, first promote
768 /// it to direct call. Each indirect call is limited with a single target.
769 ///
770 /// \param F function to perform iterative inlining.
771 /// \param InlinedGUIDs a set to be updated to include all GUIDs that are
772 /// inlined in the profiled binary.
773 ///
774 /// \returns True if there is any inline happened.
775 bool SampleProfileLoader::inlineHotFunctions(
776  Function &F, DenseSet<GlobalValue::GUID> &InlinedGUIDs) {
777  DenseSet<Instruction *> PromotedInsns;
778  bool Changed = false;
779  while (true) {
780  bool LocalChanged = false;
782  for (auto &BB : F) {
783  bool Hot = false;
785  for (auto &I : BB.getInstList()) {
786  const FunctionSamples *FS = nullptr;
787  if ((isa<CallInst>(I) || isa<InvokeInst>(I)) &&
788  !isa<IntrinsicInst>(I) && (FS = findCalleeFunctionSamples(I))) {
789  Candidates.push_back(&I);
790  if (callsiteIsHot(Samples, FS))
791  Hot = true;
792  }
793  }
794  if (Hot) {
795  CIS.insert(CIS.begin(), Candidates.begin(), Candidates.end());
796  }
797  }
798  for (auto I : CIS) {
799  Function *CalledFunction = CallSite(I).getCalledFunction();
800  // Do not inline recursive calls.
801  if (CalledFunction == &F)
802  continue;
803  if (CallSite(I).isIndirectCall()) {
804  if (PromotedInsns.count(I))
805  continue;
806  uint64_t Sum;
807  for (const auto *FS : findIndirectCallFunctionSamples(*I, Sum)) {
808  if (IsThinLTOPreLink) {
809  FS->findInlinedFunctions(InlinedGUIDs, F.getParent(),
810  Samples->getTotalSamples() *
812  continue;
813  }
814  auto CalleeFunctionName = FS->getName();
815  // If it is a recursive call, we do not inline it as it could bloat
816  // the code exponentially. There is way to better handle this, e.g.
817  // clone the caller first, and inline the cloned caller if it is
818  // recursive. As llvm does not inline recursive calls, we will
819  // simply ignore it instead of handling it explicitly.
820  if (CalleeFunctionName == F.getName())
821  continue;
822 
823  const char *Reason = "Callee function not available";
824  auto R = SymbolMap.find(CalleeFunctionName);
825  if (R != SymbolMap.end() && R->getValue() &&
826  !R->getValue()->isDeclaration() &&
827  R->getValue()->getSubprogram() &&
828  isLegalToPromote(CallSite(I), R->getValue(), &Reason)) {
829  uint64_t C = FS->getEntrySamples();
830  Instruction *DI =
831  pgo::promoteIndirectCall(I, R->getValue(), C, Sum, false, ORE);
832  Sum -= C;
833  PromotedInsns.insert(I);
834  // If profile mismatches, we should not attempt to inline DI.
835  if ((isa<CallInst>(DI) || isa<InvokeInst>(DI)) &&
836  inlineCallInstruction(DI))
837  LocalChanged = true;
838  } else {
839  DEBUG(dbgs()
840  << "\nFailed to promote indirect call to "
841  << CalleeFunctionName << " because " << Reason << "\n");
842  }
843  }
844  } else if (CalledFunction && CalledFunction->getSubprogram() &&
845  !CalledFunction->isDeclaration()) {
846  if (inlineCallInstruction(I))
847  LocalChanged = true;
848  } else if (IsThinLTOPreLink) {
849  findCalleeFunctionSamples(*I)->findInlinedFunctions(
850  InlinedGUIDs, F.getParent(),
851  Samples->getTotalSamples() * SampleProfileHotThreshold / 100);
852  }
853  }
854  if (LocalChanged) {
855  Changed = true;
856  } else {
857  break;
858  }
859  }
860  return Changed;
861 }
862 
863 /// \brief Find equivalence classes for the given block.
864 ///
865 /// This finds all the blocks that are guaranteed to execute the same
866 /// number of times as \p BB1. To do this, it traverses all the
867 /// descendants of \p BB1 in the dominator or post-dominator tree.
868 ///
869 /// A block BB2 will be in the same equivalence class as \p BB1 if
870 /// the following holds:
871 ///
872 /// 1- \p BB1 is a descendant of BB2 in the opposite tree. So, if BB2
873 /// is a descendant of \p BB1 in the dominator tree, then BB2 should
874 /// dominate BB1 in the post-dominator tree.
875 ///
876 /// 2- Both BB2 and \p BB1 must be in the same loop.
877 ///
878 /// For every block BB2 that meets those two requirements, we set BB2's
879 /// equivalence class to \p BB1.
880 ///
881 /// \param BB1 Block to check.
882 /// \param Descendants Descendants of \p BB1 in either the dom or pdom tree.
883 /// \param DomTree Opposite dominator tree. If \p Descendants is filled
884 /// with blocks from \p BB1's dominator tree, then
885 /// this is the post-dominator tree, and vice versa.
886 template <bool IsPostDom>
887 void SampleProfileLoader::findEquivalencesFor(
888  BasicBlock *BB1, ArrayRef<BasicBlock *> Descendants,
890  const BasicBlock *EC = EquivalenceClass[BB1];
891  uint64_t Weight = BlockWeights[EC];
892  for (const auto *BB2 : Descendants) {
893  bool IsDomParent = DomTree->dominates(BB2, BB1);
894  bool IsInSameLoop = LI->getLoopFor(BB1) == LI->getLoopFor(BB2);
895  if (BB1 != BB2 && IsDomParent && IsInSameLoop) {
896  EquivalenceClass[BB2] = EC;
897  // If BB2 is visited, then the entire EC should be marked as visited.
898  if (VisitedBlocks.count(BB2)) {
899  VisitedBlocks.insert(EC);
900  }
901 
902  // If BB2 is heavier than BB1, make BB2 have the same weight
903  // as BB1.
904  //
905  // Note that we don't worry about the opposite situation here
906  // (when BB2 is lighter than BB1). We will deal with this
907  // during the propagation phase. Right now, we just want to
908  // make sure that BB1 has the largest weight of all the
909  // members of its equivalence set.
910  Weight = std::max(Weight, BlockWeights[BB2]);
911  }
912  }
913  if (EC == &EC->getParent()->getEntryBlock()) {
914  BlockWeights[EC] = Samples->getHeadSamples() + 1;
915  } else {
916  BlockWeights[EC] = Weight;
917  }
918 }
919 
920 /// \brief Find equivalence classes.
921 ///
922 /// Since samples may be missing from blocks, we can fill in the gaps by setting
923 /// the weights of all the blocks in the same equivalence class to the same
924 /// weight. To compute the concept of equivalence, we use dominance and loop
925 /// information. Two blocks B1 and B2 are in the same equivalence class if B1
926 /// dominates B2, B2 post-dominates B1 and both are in the same loop.
927 ///
928 /// \param F The function to query.
929 void SampleProfileLoader::findEquivalenceClasses(Function &F) {
930  SmallVector<BasicBlock *, 8> DominatedBBs;
931  DEBUG(dbgs() << "\nBlock equivalence classes\n");
932  // Find equivalence sets based on dominance and post-dominance information.
933  for (auto &BB : F) {
934  BasicBlock *BB1 = &BB;
935 
936  // Compute BB1's equivalence class once.
937  if (EquivalenceClass.count(BB1)) {
938  DEBUG(printBlockEquivalence(dbgs(), BB1));
939  continue;
940  }
941 
942  // By default, blocks are in their own equivalence class.
943  EquivalenceClass[BB1] = BB1;
944 
945  // Traverse all the blocks dominated by BB1. We are looking for
946  // every basic block BB2 such that:
947  //
948  // 1- BB1 dominates BB2.
949  // 2- BB2 post-dominates BB1.
950  // 3- BB1 and BB2 are in the same loop nest.
951  //
952  // If all those conditions hold, it means that BB2 is executed
953  // as many times as BB1, so they are placed in the same equivalence
954  // class by making BB2's equivalence class be BB1.
955  DominatedBBs.clear();
956  DT->getDescendants(BB1, DominatedBBs);
957  findEquivalencesFor(BB1, DominatedBBs, PDT.get());
958 
959  DEBUG(printBlockEquivalence(dbgs(), BB1));
960  }
961 
962  // Assign weights to equivalence classes.
963  //
964  // All the basic blocks in the same equivalence class will execute
965  // the same number of times. Since we know that the head block in
966  // each equivalence class has the largest weight, assign that weight
967  // to all the blocks in that equivalence class.
968  DEBUG(dbgs() << "\nAssign the same weight to all blocks in the same class\n");
969  for (auto &BI : F) {
970  const BasicBlock *BB = &BI;
971  const BasicBlock *EquivBB = EquivalenceClass[BB];
972  if (BB != EquivBB)
973  BlockWeights[BB] = BlockWeights[EquivBB];
974  DEBUG(printBlockWeight(dbgs(), BB));
975  }
976 }
977 
978 /// \brief Visit the given edge to decide if it has a valid weight.
979 ///
980 /// If \p E has not been visited before, we copy to \p UnknownEdge
981 /// and increment the count of unknown edges.
982 ///
983 /// \param E Edge to visit.
984 /// \param NumUnknownEdges Current number of unknown edges.
985 /// \param UnknownEdge Set if E has not been visited before.
986 ///
987 /// \returns E's weight, if known. Otherwise, return 0.
988 uint64_t SampleProfileLoader::visitEdge(Edge E, unsigned *NumUnknownEdges,
989  Edge *UnknownEdge) {
990  if (!VisitedEdges.count(E)) {
991  (*NumUnknownEdges)++;
992  *UnknownEdge = E;
993  return 0;
994  }
995 
996  return EdgeWeights[E];
997 }
998 
999 /// \brief Propagate weights through incoming/outgoing edges.
1000 ///
1001 /// If the weight of a basic block is known, and there is only one edge
1002 /// with an unknown weight, we can calculate the weight of that edge.
1003 ///
1004 /// Similarly, if all the edges have a known count, we can calculate the
1005 /// count of the basic block, if needed.
1006 ///
1007 /// \param F Function to process.
1008 /// \param UpdateBlockCount Whether we should update basic block counts that
1009 /// has already been annotated.
1010 ///
1011 /// \returns True if new weights were assigned to edges or blocks.
1012 bool SampleProfileLoader::propagateThroughEdges(Function &F,
1013  bool UpdateBlockCount) {
1014  bool Changed = false;
1015  DEBUG(dbgs() << "\nPropagation through edges\n");
1016  for (const auto &BI : F) {
1017  const BasicBlock *BB = &BI;
1018  const BasicBlock *EC = EquivalenceClass[BB];
1019 
1020  // Visit all the predecessor and successor edges to determine
1021  // which ones have a weight assigned already. Note that it doesn't
1022  // matter that we only keep track of a single unknown edge. The
1023  // only case we are interested in handling is when only a single
1024  // edge is unknown (see setEdgeOrBlockWeight).
1025  for (unsigned i = 0; i < 2; i++) {
1026  uint64_t TotalWeight = 0;
1027  unsigned NumUnknownEdges = 0, NumTotalEdges = 0;
1028  Edge UnknownEdge, SelfReferentialEdge, SingleEdge;
1029 
1030  if (i == 0) {
1031  // First, visit all predecessor edges.
1032  NumTotalEdges = Predecessors[BB].size();
1033  for (auto *Pred : Predecessors[BB]) {
1034  Edge E = std::make_pair(Pred, BB);
1035  TotalWeight += visitEdge(E, &NumUnknownEdges, &UnknownEdge);
1036  if (E.first == E.second)
1037  SelfReferentialEdge = E;
1038  }
1039  if (NumTotalEdges == 1) {
1040  SingleEdge = std::make_pair(Predecessors[BB][0], BB);
1041  }
1042  } else {
1043  // On the second round, visit all successor edges.
1044  NumTotalEdges = Successors[BB].size();
1045  for (auto *Succ : Successors[BB]) {
1046  Edge E = std::make_pair(BB, Succ);
1047  TotalWeight += visitEdge(E, &NumUnknownEdges, &UnknownEdge);
1048  }
1049  if (NumTotalEdges == 1) {
1050  SingleEdge = std::make_pair(BB, Successors[BB][0]);
1051  }
1052  }
1053 
1054  // After visiting all the edges, there are three cases that we
1055  // can handle immediately:
1056  //
1057  // - All the edge weights are known (i.e., NumUnknownEdges == 0).
1058  // In this case, we simply check that the sum of all the edges
1059  // is the same as BB's weight. If not, we change BB's weight
1060  // to match. Additionally, if BB had not been visited before,
1061  // we mark it visited.
1062  //
1063  // - Only one edge is unknown and BB has already been visited.
1064  // In this case, we can compute the weight of the edge by
1065  // subtracting the total block weight from all the known
1066  // edge weights. If the edges weight more than BB, then the
1067  // edge of the last remaining edge is set to zero.
1068  //
1069  // - There exists a self-referential edge and the weight of BB is
1070  // known. In this case, this edge can be based on BB's weight.
1071  // We add up all the other known edges and set the weight on
1072  // the self-referential edge as we did in the previous case.
1073  //
1074  // In any other case, we must continue iterating. Eventually,
1075  // all edges will get a weight, or iteration will stop when
1076  // it reaches SampleProfileMaxPropagateIterations.
1077  if (NumUnknownEdges <= 1) {
1078  uint64_t &BBWeight = BlockWeights[EC];
1079  if (NumUnknownEdges == 0) {
1080  if (!VisitedBlocks.count(EC)) {
1081  // If we already know the weight of all edges, the weight of the
1082  // basic block can be computed. It should be no larger than the sum
1083  // of all edge weights.
1084  if (TotalWeight > BBWeight) {
1085  BBWeight = TotalWeight;
1086  Changed = true;
1087  DEBUG(dbgs() << "All edge weights for " << BB->getName()
1088  << " known. Set weight for block: ";
1089  printBlockWeight(dbgs(), BB););
1090  }
1091  } else if (NumTotalEdges == 1 &&
1092  EdgeWeights[SingleEdge] < BlockWeights[EC]) {
1093  // If there is only one edge for the visited basic block, use the
1094  // block weight to adjust edge weight if edge weight is smaller.
1095  EdgeWeights[SingleEdge] = BlockWeights[EC];
1096  Changed = true;
1097  }
1098  } else if (NumUnknownEdges == 1 && VisitedBlocks.count(EC)) {
1099  // If there is a single unknown edge and the block has been
1100  // visited, then we can compute E's weight.
1101  if (BBWeight >= TotalWeight)
1102  EdgeWeights[UnknownEdge] = BBWeight - TotalWeight;
1103  else
1104  EdgeWeights[UnknownEdge] = 0;
1105  const BasicBlock *OtherEC;
1106  if (i == 0)
1107  OtherEC = EquivalenceClass[UnknownEdge.first];
1108  else
1109  OtherEC = EquivalenceClass[UnknownEdge.second];
1110  // Edge weights should never exceed the BB weights it connects.
1111  if (VisitedBlocks.count(OtherEC) &&
1112  EdgeWeights[UnknownEdge] > BlockWeights[OtherEC])
1113  EdgeWeights[UnknownEdge] = BlockWeights[OtherEC];
1114  VisitedEdges.insert(UnknownEdge);
1115  Changed = true;
1116  DEBUG(dbgs() << "Set weight for edge: ";
1117  printEdgeWeight(dbgs(), UnknownEdge));
1118  }
1119  } else if (VisitedBlocks.count(EC) && BlockWeights[EC] == 0) {
1120  // If a block Weights 0, all its in/out edges should weight 0.
1121  if (i == 0) {
1122  for (auto *Pred : Predecessors[BB]) {
1123  Edge E = std::make_pair(Pred, BB);
1124  EdgeWeights[E] = 0;
1125  VisitedEdges.insert(E);
1126  }
1127  } else {
1128  for (auto *Succ : Successors[BB]) {
1129  Edge E = std::make_pair(BB, Succ);
1130  EdgeWeights[E] = 0;
1131  VisitedEdges.insert(E);
1132  }
1133  }
1134  } else if (SelfReferentialEdge.first && VisitedBlocks.count(EC)) {
1135  uint64_t &BBWeight = BlockWeights[BB];
1136  // We have a self-referential edge and the weight of BB is known.
1137  if (BBWeight >= TotalWeight)
1138  EdgeWeights[SelfReferentialEdge] = BBWeight - TotalWeight;
1139  else
1140  EdgeWeights[SelfReferentialEdge] = 0;
1141  VisitedEdges.insert(SelfReferentialEdge);
1142  Changed = true;
1143  DEBUG(dbgs() << "Set self-referential edge weight to: ";
1144  printEdgeWeight(dbgs(), SelfReferentialEdge));
1145  }
1146  if (UpdateBlockCount && !VisitedBlocks.count(EC) && TotalWeight > 0) {
1147  BlockWeights[EC] = TotalWeight;
1148  VisitedBlocks.insert(EC);
1149  Changed = true;
1150  }
1151  }
1152  }
1153 
1154  return Changed;
1155 }
1156 
1157 /// \brief Build in/out edge lists for each basic block in the CFG.
1158 ///
1159 /// We are interested in unique edges. If a block B1 has multiple
1160 /// edges to another block B2, we only add a single B1->B2 edge.
1161 void SampleProfileLoader::buildEdges(Function &F) {
1162  for (auto &BI : F) {
1163  BasicBlock *B1 = &BI;
1164 
1165  // Add predecessors for B1.
1167  if (!Predecessors[B1].empty())
1168  llvm_unreachable("Found a stale predecessors list in a basic block.");
1169  for (pred_iterator PI = pred_begin(B1), PE = pred_end(B1); PI != PE; ++PI) {
1170  BasicBlock *B2 = *PI;
1171  if (Visited.insert(B2).second)
1172  Predecessors[B1].push_back(B2);
1173  }
1174 
1175  // Add successors for B1.
1176  Visited.clear();
1177  if (!Successors[B1].empty())
1178  llvm_unreachable("Found a stale successors list in a basic block.");
1179  for (succ_iterator SI = succ_begin(B1), SE = succ_end(B1); SI != SE; ++SI) {
1180  BasicBlock *B2 = *SI;
1181  if (Visited.insert(B2).second)
1182  Successors[B1].push_back(B2);
1183  }
1184  }
1185 }
1186 
1187 /// Returns the sorted CallTargetMap \p M by count in descending order.
1189  const SampleRecord::CallTargetMap &M) {
1191  for (auto I = M.begin(); I != M.end(); ++I)
1192  R.push_back({Function::getGUID(I->getKey()), I->getValue()});
1193  std::sort(R.begin(), R.end(),
1194  [](const InstrProfValueData &L, const InstrProfValueData &R) {
1195  if (L.Count == R.Count)
1196  return L.Value > R.Value;
1197  else
1198  return L.Count > R.Count;
1199  });
1200  return R;
1201 }
1202 
1203 /// \brief Propagate weights into edges
1204 ///
1205 /// The following rules are applied to every block BB in the CFG:
1206 ///
1207 /// - If BB has a single predecessor/successor, then the weight
1208 /// of that edge is the weight of the block.
1209 ///
1210 /// - If all incoming or outgoing edges are known except one, and the
1211 /// weight of the block is already known, the weight of the unknown
1212 /// edge will be the weight of the block minus the sum of all the known
1213 /// edges. If the sum of all the known edges is larger than BB's weight,
1214 /// we set the unknown edge weight to zero.
1215 ///
1216 /// - If there is a self-referential edge, and the weight of the block is
1217 /// known, the weight for that edge is set to the weight of the block
1218 /// minus the weight of the other incoming edges to that block (if
1219 /// known).
1220 void SampleProfileLoader::propagateWeights(Function &F) {
1221  bool Changed = true;
1222  unsigned I = 0;
1223 
1224  // If BB weight is larger than its corresponding loop's header BB weight,
1225  // use the BB weight to replace the loop header BB weight.
1226  for (auto &BI : F) {
1227  BasicBlock *BB = &BI;
1228  Loop *L = LI->getLoopFor(BB);
1229  if (!L) {
1230  continue;
1231  }
1232  BasicBlock *Header = L->getHeader();
1233  if (Header && BlockWeights[BB] > BlockWeights[Header]) {
1234  BlockWeights[Header] = BlockWeights[BB];
1235  }
1236  }
1237 
1238  // Before propagation starts, build, for each block, a list of
1239  // unique predecessors and successors. This is necessary to handle
1240  // identical edges in multiway branches. Since we visit all blocks and all
1241  // edges of the CFG, it is cleaner to build these lists once at the start
1242  // of the pass.
1243  buildEdges(F);
1244 
1245  // Propagate until we converge or we go past the iteration limit.
1246  while (Changed && I++ < SampleProfileMaxPropagateIterations) {
1247  Changed = propagateThroughEdges(F, false);
1248  }
1249 
1250  // The first propagation propagates BB counts from annotated BBs to unknown
1251  // BBs. The 2nd propagation pass resets edges weights, and use all BB weights
1252  // to propagate edge weights.
1253  VisitedEdges.clear();
1254  Changed = true;
1255  while (Changed && I++ < SampleProfileMaxPropagateIterations) {
1256  Changed = propagateThroughEdges(F, false);
1257  }
1258 
1259  // The 3rd propagation pass allows adjust annotated BB weights that are
1260  // obviously wrong.
1261  Changed = true;
1262  while (Changed && I++ < SampleProfileMaxPropagateIterations) {
1263  Changed = propagateThroughEdges(F, true);
1264  }
1265 
1266  // Generate MD_prof metadata for every branch instruction using the
1267  // edge weights computed during propagation.
1268  DEBUG(dbgs() << "\nPropagation complete. Setting branch weights\n");
1269  LLVMContext &Ctx = F.getContext();
1270  MDBuilder MDB(Ctx);
1271  for (auto &BI : F) {
1272  BasicBlock *BB = &BI;
1273 
1274  if (BlockWeights[BB]) {
1275  for (auto &I : BB->getInstList()) {
1276  if (!isa<CallInst>(I) && !isa<InvokeInst>(I))
1277  continue;
1278  CallSite CS(&I);
1279  if (!CS.getCalledFunction()) {
1280  const DebugLoc &DLoc = I.getDebugLoc();
1281  if (!DLoc)
1282  continue;
1283  const DILocation *DIL = DLoc;
1284  uint32_t LineOffset = getOffset(DIL);
1285  uint32_t Discriminator = DIL->getBaseDiscriminator();
1286 
1287  const FunctionSamples *FS = findFunctionSamples(I);
1288  if (!FS)
1289  continue;
1290  auto T = FS->findCallTargetMapAt(LineOffset, Discriminator);
1291  if (!T || T.get().empty())
1292  continue;
1293  SmallVector<InstrProfValueData, 2> SortedCallTargets =
1294  SortCallTargets(T.get());
1295  uint64_t Sum;
1296  findIndirectCallFunctionSamples(I, Sum);
1297  annotateValueSite(*I.getParent()->getParent()->getParent(), I,
1298  SortedCallTargets, Sum, IPVK_IndirectCallTarget,
1299  SortedCallTargets.size());
1300  } else if (!dyn_cast<IntrinsicInst>(&I)) {
1301  SmallVector<uint32_t, 1> Weights;
1302  Weights.push_back(BlockWeights[BB]);
1303  I.setMetadata(LLVMContext::MD_prof, MDB.createBranchWeights(Weights));
1304  }
1305  }
1306  }
1307  TerminatorInst *TI = BB->getTerminator();
1308  if (TI->getNumSuccessors() == 1)
1309  continue;
1310  if (!isa<BranchInst>(TI) && !isa<SwitchInst>(TI))
1311  continue;
1312 
1313  DebugLoc BranchLoc = TI->getDebugLoc();
1314  DEBUG(dbgs() << "\nGetting weights for branch at line "
1315  << ((BranchLoc) ? Twine(BranchLoc.getLine())
1316  : Twine("<UNKNOWN LOCATION>"))
1317  << ".\n");
1318  SmallVector<uint32_t, 4> Weights;
1319  uint32_t MaxWeight = 0;
1320  Instruction *MaxDestInst;
1321  for (unsigned I = 0; I < TI->getNumSuccessors(); ++I) {
1322  BasicBlock *Succ = TI->getSuccessor(I);
1323  Edge E = std::make_pair(BB, Succ);
1324  uint64_t Weight = EdgeWeights[E];
1325  DEBUG(dbgs() << "\t"; printEdgeWeight(dbgs(), E));
1326  // Use uint32_t saturated arithmetic to adjust the incoming weights,
1327  // if needed. Sample counts in profiles are 64-bit unsigned values,
1328  // but internally branch weights are expressed as 32-bit values.
1329  if (Weight > std::numeric_limits<uint32_t>::max()) {
1330  DEBUG(dbgs() << " (saturated due to uint32_t overflow)");
1332  }
1333  // Weight is added by one to avoid propagation errors introduced by
1334  // 0 weights.
1335  Weights.push_back(static_cast<uint32_t>(Weight + 1));
1336  if (Weight != 0) {
1337  if (Weight > MaxWeight) {
1338  MaxWeight = Weight;
1339  MaxDestInst = Succ->getFirstNonPHIOrDbgOrLifetime();
1340  }
1341  }
1342  }
1343 
1344  uint64_t TempWeight;
1345  // Only set weights if there is at least one non-zero weight.
1346  // In any other case, let the analyzer set weights.
1347  // Do not set weights if the weights are present. In ThinLTO, the profile
1348  // annotation is done twice. If the first annotation already set the
1349  // weights, the second pass does not need to set it.
1350  if (MaxWeight > 0 && !TI->extractProfTotalWeight(TempWeight)) {
1351  DEBUG(dbgs() << "SUCCESS. Found non-zero weights.\n");
1353  MDB.createBranchWeights(Weights));
1354  ORE->emit([&]() {
1355  return OptimizationRemark(DEBUG_TYPE, "PopularDest", MaxDestInst)
1356  << "most popular destination for conditional branches at "
1357  << ore::NV("CondBranchesLoc", BranchLoc);
1358  });
1359  } else {
1360  DEBUG(dbgs() << "SKIPPED. All branch weights are zero.\n");
1361  }
1362  }
1363 }
1364 
1365 /// \brief Get the line number for the function header.
1366 ///
1367 /// This looks up function \p F in the current compilation unit and
1368 /// retrieves the line number where the function is defined. This is
1369 /// line 0 for all the samples read from the profile file. Every line
1370 /// number is relative to this line.
1371 ///
1372 /// \param F Function object to query.
1373 ///
1374 /// \returns the line number where \p F is defined. If it returns 0,
1375 /// it means that there is no debug information available for \p F.
1376 unsigned SampleProfileLoader::getFunctionLoc(Function &F) {
1377  if (DISubprogram *S = F.getSubprogram())
1378  return S->getLine();
1379 
1380  // If the start of \p F is missing, emit a diagnostic to inform the user
1381  // about the missed opportunity.
1383  "No debug information found in function " + F.getName() +
1384  ": Function profile not used",
1385  DS_Warning));
1386  return 0;
1387 }
1388 
1389 void SampleProfileLoader::computeDominanceAndLoopInfo(Function &F) {
1390  DT.reset(new DominatorTree);
1391  DT->recalculate(F);
1392 
1393  PDT.reset(new PostDomTreeBase<BasicBlock>());
1394  PDT->recalculate(F);
1395 
1396  LI.reset(new LoopInfo);
1397  LI->analyze(*DT);
1398 }
1399 
1400 /// \brief Generate branch weight metadata for all branches in \p F.
1401 ///
1402 /// Branch weights are computed out of instruction samples using a
1403 /// propagation heuristic. Propagation proceeds in 3 phases:
1404 ///
1405 /// 1- Assignment of block weights. All the basic blocks in the function
1406 /// are initial assigned the same weight as their most frequently
1407 /// executed instruction.
1408 ///
1409 /// 2- Creation of equivalence classes. Since samples may be missing from
1410 /// blocks, we can fill in the gaps by setting the weights of all the
1411 /// blocks in the same equivalence class to the same weight. To compute
1412 /// the concept of equivalence, we use dominance and loop information.
1413 /// Two blocks B1 and B2 are in the same equivalence class if B1
1414 /// dominates B2, B2 post-dominates B1 and both are in the same loop.
1415 ///
1416 /// 3- Propagation of block weights into edges. This uses a simple
1417 /// propagation heuristic. The following rules are applied to every
1418 /// block BB in the CFG:
1419 ///
1420 /// - If BB has a single predecessor/successor, then the weight
1421 /// of that edge is the weight of the block.
1422 ///
1423 /// - If all the edges are known except one, and the weight of the
1424 /// block is already known, the weight of the unknown edge will
1425 /// be the weight of the block minus the sum of all the known
1426 /// edges. If the sum of all the known edges is larger than BB's weight,
1427 /// we set the unknown edge weight to zero.
1428 ///
1429 /// - If there is a self-referential edge, and the weight of the block is
1430 /// known, the weight for that edge is set to the weight of the block
1431 /// minus the weight of the other incoming edges to that block (if
1432 /// known).
1433 ///
1434 /// Since this propagation is not guaranteed to finalize for every CFG, we
1435 /// only allow it to proceed for a limited number of iterations (controlled
1436 /// by -sample-profile-max-propagate-iterations).
1437 ///
1438 /// FIXME: Try to replace this propagation heuristic with a scheme
1439 /// that is guaranteed to finalize. A work-list approach similar to
1440 /// the standard value propagation algorithm used by SSA-CCP might
1441 /// work here.
1442 ///
1443 /// Once all the branch weights are computed, we emit the MD_prof
1444 /// metadata on BB using the computed values for each of its branches.
1445 ///
1446 /// \param F The function to query.
1447 ///
1448 /// \returns true if \p F was modified. Returns false, otherwise.
1449 bool SampleProfileLoader::emitAnnotations(Function &F) {
1450  bool Changed = false;
1451 
1452  if (getFunctionLoc(F) == 0)
1453  return false;
1454 
1455  DEBUG(dbgs() << "Line number for the first instruction in " << F.getName()
1456  << ": " << getFunctionLoc(F) << "\n");
1457 
1458  DenseSet<GlobalValue::GUID> InlinedGUIDs;
1459  Changed |= inlineHotFunctions(F, InlinedGUIDs);
1460 
1461  // Compute basic block weights.
1462  Changed |= computeBlockWeights(F);
1463 
1464  if (Changed) {
1465  // Add an entry count to the function using the samples gathered at the
1466  // function entry.
1467  // Sets the GUIDs that are inlined in the profiled binary. This is used
1468  // for ThinLink to make correct liveness analysis, and also make the IR
1469  // match the profiled binary before annotation.
1470  F.setEntryCount(
1471  ProfileCount(Samples->getHeadSamples() + 1, Function::PCT_Real),
1472  &InlinedGUIDs);
1473 
1474  // Compute dominance and loop info needed for propagation.
1475  computeDominanceAndLoopInfo(F);
1476 
1477  // Find equivalence classes.
1478  findEquivalenceClasses(F);
1479 
1480  // Propagate weights to all edges.
1481  propagateWeights(F);
1482  }
1483 
1484  // If coverage checking was requested, compute it now.
1486  unsigned Used = CoverageTracker.countUsedRecords(Samples);
1487  unsigned Total = CoverageTracker.countBodyRecords(Samples);
1488  unsigned Coverage = CoverageTracker.computeCoverage(Used, Total);
1489  if (Coverage < SampleProfileRecordCoverage) {
1491  F.getSubprogram()->getFilename(), getFunctionLoc(F),
1492  Twine(Used) + " of " + Twine(Total) + " available profile records (" +
1493  Twine(Coverage) + "%) were applied",
1494  DS_Warning));
1495  }
1496  }
1497 
1499  uint64_t Used = CoverageTracker.getTotalUsedSamples();
1500  uint64_t Total = CoverageTracker.countBodySamples(Samples);
1501  unsigned Coverage = CoverageTracker.computeCoverage(Used, Total);
1502  if (Coverage < SampleProfileSampleCoverage) {
1504  F.getSubprogram()->getFilename(), getFunctionLoc(F),
1505  Twine(Used) + " of " + Twine(Total) + " available profile samples (" +
1506  Twine(Coverage) + "%) were applied",
1507  DS_Warning));
1508  }
1509  }
1510  return Changed;
1511 }
1512 
1514 
1515 INITIALIZE_PASS_BEGIN(SampleProfileLoaderLegacyPass, "sample-profile",
1516  "Sample Profile loader", false, false)
1519 INITIALIZE_PASS_END(SampleProfileLoaderLegacyPass, "sample-profile",
1520  "Sample Profile loader", false, false)
1521 
1522 bool SampleProfileLoader::doInitialization(Module &M) {
1523  auto &Ctx = M.getContext();
1524  auto ReaderOrErr = SampleProfileReader::create(Filename, Ctx);
1525  if (std::error_code EC = ReaderOrErr.getError()) {
1526  std::string Msg = "Could not open profile: " + EC.message();
1527  Ctx.diagnose(DiagnosticInfoSampleProfile(Filename, Msg));
1528  return false;
1529  }
1530  Reader = std::move(ReaderOrErr.get());
1531  ProfileIsValid = (Reader->read() == sampleprof_error::success);
1532  return true;
1533 }
1534 
1536  return new SampleProfileLoaderLegacyPass(SampleProfileFile);
1537 }
1538 
1540  return new SampleProfileLoaderLegacyPass(Name);
1541 }
1542 
1543 bool SampleProfileLoader::runOnModule(Module &M, ModuleAnalysisManager *AM) {
1544  if (!ProfileIsValid)
1545  return false;
1546 
1547  // Compute the total number of samples collected in this profile.
1548  for (const auto &I : Reader->getProfiles())
1549  TotalCollectedSamples += I.second.getTotalSamples();
1550 
1551  // Populate the symbol map.
1552  for (const auto &N_F : M.getValueSymbolTable()) {
1553  StringRef OrigName = N_F.getKey();
1554  Function *F = dyn_cast<Function>(N_F.getValue());
1555  if (F == nullptr)
1556  continue;
1557  SymbolMap[OrigName] = F;
1558  auto pos = OrigName.find('.');
1559  if (pos != StringRef::npos) {
1560  StringRef NewName = OrigName.substr(0, pos);
1561  auto r = SymbolMap.insert(std::make_pair(NewName, F));
1562  // Failiing to insert means there is already an entry in SymbolMap,
1563  // thus there are multiple functions that are mapped to the same
1564  // stripped name. In this case of name conflicting, set the value
1565  // to nullptr to avoid confusion.
1566  if (!r.second)
1567  r.first->second = nullptr;
1568  }
1569  }
1570 
1571  bool retval = false;
1572  for (auto &F : M)
1573  if (!F.isDeclaration()) {
1574  clearFunctionData();
1575  retval |= runOnFunction(F, AM);
1576  }
1577  if (M.getProfileSummary() == nullptr)
1578  M.setProfileSummary(Reader->getSummary().getMD(M.getContext()));
1579  return retval;
1580 }
1581 
1582 bool SampleProfileLoaderLegacyPass::runOnModule(Module &M) {
1583  ACT = &getAnalysis<AssumptionCacheTracker>();
1584  TTIWP = &getAnalysis<TargetTransformInfoWrapperPass>();
1585  return SampleLoader.runOnModule(M, nullptr);
1586 }
1587 
1589  // Initialize the entry count to -1, which will be treated conservatively
1590  // by getEntryCount as the same as unknown (None). If we have samples this
1591  // will be overwritten in emitAnnotations.
1593  std::unique_ptr<OptimizationRemarkEmitter> OwnedORE;
1594  if (AM) {
1595  auto &FAM =
1597  .getManager();
1598  ORE = &FAM.getResult<OptimizationRemarkEmitterAnalysis>(F);
1599  } else {
1600  OwnedORE = make_unique<OptimizationRemarkEmitter>(&F);
1601  ORE = OwnedORE.get();
1602  }
1603  Samples = Reader->getSamplesFor(F);
1604  if (Samples && !Samples->empty())
1605  return emitAnnotations(F);
1606  return false;
1607 }
1608 
1610  ModuleAnalysisManager &AM) {
1612  AM.getResult<FunctionAnalysisManagerModuleProxy>(M).getManager();
1613 
1614  auto GetAssumptionCache = [&](Function &F) -> AssumptionCache & {
1615  return FAM.getResult<AssumptionAnalysis>(F);
1616  };
1617  auto GetTTI = [&](Function &F) -> TargetTransformInfo & {
1618  return FAM.getResult<TargetIRAnalysis>(F);
1619  };
1620 
1621  SampleProfileLoader SampleLoader(
1622  ProfileFileName.empty() ? SampleProfileFile : ProfileFileName,
1623  IsThinLTOPreLink, GetAssumptionCache, GetTTI);
1624 
1625  SampleLoader.doInitialization(M);
1626 
1627  if (!SampleLoader.runOnModule(M, &AM))
1628  return PreservedAnalyses::all();
1629 
1630  return PreservedAnalyses::none();
1631 }
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:265
Thresholds to tune inline cost analysis.
Definition: InlineCost.h:130
Represents either an error or a value T.
Definition: ErrorOr.h:69
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:99
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
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.
LLVM_ATTRIBUTE_ALWAYS_INLINE size_type size() const
Definition: SmallVector.h:136
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:542
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 .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 second
unsigned getLine() const
Definition: DebugLoc.cpp:25
ErrorOr< uint64_t > findSamplesAt(uint32_t LineOffset, uint32_t Discriminator) const
Return the number of samples collected at the given location.
Definition: SampleProf.h:238
std::map< SymbolStringPtr, JITEvaluatedSymbol > SymbolMap
A map from symbol names (as SymbolStringPtrs) to JITSymbols (address/flags pairs).
Definition: Core.h:40
A debug info location.
Definition: DebugLoc.h:34
F(f)
void findInlinedFunctions(DenseSet< GlobalValue::GUID > &S, const Module *M, uint64_t Threshold) const
Recursively traverses all children, if the total sample count of the corresponding function is no les...
Definition: SampleProf.h:361
const StringRef & getName() const
Return the function name.
Definition: SampleProf.h:384
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:176
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:198
bool 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...
AnalysisUsage & addRequired()
#define INITIALIZE_PASS_DEPENDENCY(depName)
Definition: PassSupport.h:51
Definition: BitVector.h:920
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:1324
Diagnostic information for optimization analysis remarks.
static cl::opt< double > SampleProfileHotThreshold("sample-profile-inline-hot-threshold", cl::init(0.1), cl::value_desc("N"), cl::desc("Inlined functions that account for more than N% of all samples " "collected in the parent function, will be inlined again."))
static StringRef getName(Value *V)
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:185
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:1336
static Error getOffset(const SymbolRef &Sym, SectionRef Sec, uint64_t &Result)
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:618
static bool runOnFunction(Function &F, bool PostInlining)
initializer< Ty > init(const Ty &Val)
Definition: CommandLine.h:406
void dump(const SparseBitVector< ElementSize > &LHS, raw_ostream &out)
Subclasses of this class are all able to terminate a basic block.
Definition: InstrTypes.h:54
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.
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:1505
const BodySampleMap & getBodySamples() const
Return all the samples collected in the body of the function.
Definition: SampleProf.h:329
static GCRegistry::Add< CoreCLRGC > E("coreclr", "CoreCLR-compatible GC")
LLVM_ATTRIBUTE_ALWAYS_INLINE iterator begin()
Definition: SmallVector.h:116
SmallSet - This maintains a set of unique values, optimizing for the case when the set is small (less...
Definition: SmallSet.h:36
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:811
static PreservedAnalyses all()
Construct a special preserved set that preserves all passes.
Definition: PassManager.h:159
Class to represent profile counts.
Definition: Function.h:242
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:1222
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:157
unsigned first
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:276
A function analysis which provides an AssumptionCache.
const InstListType & getInstList() const
Return the underlying instruction list container.
Definition: BasicBlock.h:317
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:862
Module.h This file contains the declarations for the Module class.
static bool callsiteIsHot(const FunctionSamples *CallerFS, const FunctionSamples *CallsiteFS)
Return true if the given callsite is hot wrt to its caller.
ErrorOr< SampleRecord::CallTargetMap > findCallTargetMapAt(uint32_t LineOffset, uint32_t Discriminator) const
Returns the call target map collected at a given location.
Definition: SampleProf.h:251
raw_ostream & dbgs()
dbgs() - This returns a reference to a raw_ostream for debugging messages.
Definition: Debug.cpp:132
StringMap - This is an unconventional map that is specialized for handling keys that are "strings"...
Definition: StringMap.h:222
static void clear(coro::Shape &Shape)
Definition: Coroutines.cpp:210
iterator insert(iterator I, T &&Elt)
Definition: SmallVector.h:480
bool dominates(const DomTreeNodeBase< NodeT > *A, const DomTreeNodeBase< NodeT > *B) const
dominates - Returns true iff A dominates B.
LLVM_ATTRIBUTE_ALWAYS_INLINE iterator end()
Definition: SmallVector.h:120
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:285
#define DEBUG_TYPE
static const size_t npos
Definition: StringRef.h:51
iterator begin()
Definition: StringMap.h:317
Represents the relative location of an instruction.
Definition: SampleProf.h:99
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."))
Represents a single loop in the control flow graph.
Definition: LoopInfo.h:439
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:713
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:190
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:299
sample Sample Profile loader
const CallsiteSampleMap & getCallsiteSamples() const
Return all the callsite samples collected in the body of the function.
Definition: SampleProf.h:332
This class implements an extremely fast bulk output stream that can only output to a stream...
Definition: raw_ostream.h:44
#define DEBUG(X)
Definition: Debug.h:118
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.
void sort(Policy policy, RandomAccessIterator Start, RandomAccessIterator End, const Comparator &Comp=Comparator())
Definition: Parallel.h:199
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:310
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
The optimization diagnostic interface.
iterator end()
Definition: StringMap.h:320
This file provides the interface for the sampled PGO loader pass.
reference get()
Definition: ErrorOr.h:169
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