LLVM  6.0.0svn
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"
73 #include <algorithm>
74 #include <cassert>
75 #include <cstdint>
76 #include <functional>
77 #include <limits>
78 #include <map>
79 #include <memory>
80 #include <string>
81 #include <system_error>
82 #include <utility>
83 #include <vector>
84 
85 using namespace llvm;
86 using namespace sampleprof;
87 
88 #define DEBUG_TYPE "sample-profile"
89 
90 // Command line option to specify the file to read samples from. This is
91 // mainly used for debugging.
93  "sample-profile-file", cl::init(""), cl::value_desc("filename"),
94  cl::desc("Profile file loaded by -sample-profile"), cl::Hidden);
95 
97  "sample-profile-max-propagate-iterations", cl::init(100),
98  cl::desc("Maximum number of iterations to go through when propagating "
99  "sample block/edge weights through the CFG."));
100 
102  "sample-profile-check-record-coverage", cl::init(0), cl::value_desc("N"),
103  cl::desc("Emit a warning if less than N% of records in the input profile "
104  "are matched to the IR."));
105 
107  "sample-profile-check-sample-coverage", cl::init(0), cl::value_desc("N"),
108  cl::desc("Emit a warning if less than N% of samples in the input profile "
109  "are matched to the IR."));
110 
112  "sample-profile-inline-hot-threshold", cl::init(0.1), cl::value_desc("N"),
113  cl::desc("Inlined functions that account for more than N% of all samples "
114  "collected in the parent function, will be inlined again."));
115 
116 namespace {
117 
118 using BlockWeightMap = DenseMap<const BasicBlock *, uint64_t>;
119 using EquivalenceClassMap = DenseMap<const BasicBlock *, const BasicBlock *>;
120 using Edge = std::pair<const BasicBlock *, const BasicBlock *>;
121 using EdgeWeightMap = DenseMap<Edge, uint64_t>;
122 using BlockEdgeMap =
124 
125 class SampleCoverageTracker {
126 public:
127  SampleCoverageTracker() = default;
128 
129  bool markSamplesUsed(const FunctionSamples *FS, uint32_t LineOffset,
130  uint32_t Discriminator, uint64_t Samples);
131  unsigned computeCoverage(unsigned Used, unsigned Total) const;
132  unsigned countUsedRecords(const FunctionSamples *FS) const;
133  unsigned countBodyRecords(const FunctionSamples *FS) const;
134  uint64_t getTotalUsedSamples() const { return TotalUsedSamples; }
135  uint64_t countBodySamples(const FunctionSamples *FS) const;
136 
137  void clear() {
138  SampleCoverage.clear();
139  TotalUsedSamples = 0;
140  }
141 
142 private:
143  using BodySampleCoverageMap = std::map<LineLocation, unsigned>;
144  using FunctionSamplesCoverageMap =
146 
147  /// Coverage map for sampling records.
148  ///
149  /// This map keeps a record of sampling records that have been matched to
150  /// an IR instruction. This is used to detect some form of staleness in
151  /// profiles (see flag -sample-profile-check-coverage).
152  ///
153  /// Each entry in the map corresponds to a FunctionSamples instance. This is
154  /// another map that counts how many times the sample record at the
155  /// given location has been used.
156  FunctionSamplesCoverageMap SampleCoverage;
157 
158  /// Number of samples used from the profile.
159  ///
160  /// When a sampling record is used for the first time, the samples from
161  /// that record are added to this accumulator. Coverage is later computed
162  /// based on the total number of samples available in this function and
163  /// its callsites.
164  ///
165  /// Note that this accumulator tracks samples used from a single function
166  /// and all the inlined callsites. Strictly, we should have a map of counters
167  /// keyed by FunctionSamples pointers, but these stats are cleared after
168  /// every function, so we just need to keep a single counter.
169  uint64_t TotalUsedSamples = 0;
170 };
171 
172 /// \brief Sample profile pass.
173 ///
174 /// This pass reads profile data from the file specified by
175 /// -sample-profile-file and annotates every affected function with the
176 /// profile information found in that file.
177 class SampleProfileLoader {
178 public:
179  SampleProfileLoader(
180  StringRef Name, bool IsThinLTOPreLink,
181  std::function<AssumptionCache &(Function &)> GetAssumptionCache,
182  std::function<TargetTransformInfo &(Function &)> GetTargetTransformInfo)
183  : GetAC(GetAssumptionCache), GetTTI(GetTargetTransformInfo),
184  Filename(Name), IsThinLTOPreLink(IsThinLTOPreLink) {}
185 
186  bool doInitialization(Module &M);
187  bool runOnModule(Module &M, ModuleAnalysisManager *AM);
188 
189  void dump() { Reader->dump(); }
190 
191 protected:
193  unsigned getFunctionLoc(Function &F);
194  bool emitAnnotations(Function &F);
195  ErrorOr<uint64_t> getInstWeight(const Instruction &I);
196  ErrorOr<uint64_t> getBlockWeight(const BasicBlock *BB);
197  const FunctionSamples *findCalleeFunctionSamples(const Instruction &I) const;
198  std::vector<const FunctionSamples *>
199  findIndirectCallFunctionSamples(const Instruction &I, uint64_t &Sum) const;
200  const FunctionSamples *findFunctionSamples(const Instruction &I) const;
201  bool inlineCallInstruction(Instruction *I);
202  bool inlineHotFunctions(Function &F,
203  DenseSet<GlobalValue::GUID> &InlinedGUIDs);
204  void printEdgeWeight(raw_ostream &OS, Edge E);
205  void printBlockWeight(raw_ostream &OS, const BasicBlock *BB) const;
206  void printBlockEquivalence(raw_ostream &OS, const BasicBlock *BB);
207  bool computeBlockWeights(Function &F);
208  void findEquivalenceClasses(Function &F);
209  template <bool IsPostDom>
210  void findEquivalencesFor(BasicBlock *BB1, ArrayRef<BasicBlock *> Descendants,
212 
213  void propagateWeights(Function &F);
214  uint64_t visitEdge(Edge E, unsigned *NumUnknownEdges, Edge *UnknownEdge);
215  void buildEdges(Function &F);
216  bool propagateThroughEdges(Function &F, bool UpdateBlockCount);
217  void computeDominanceAndLoopInfo(Function &F);
218  unsigned getOffset(const DILocation *DIL) const;
219  void clearFunctionData();
220 
221  /// \brief Map basic blocks to their computed weights.
222  ///
223  /// The weight of a basic block is defined to be the maximum
224  /// of all the instruction weights in that block.
225  BlockWeightMap BlockWeights;
226 
227  /// \brief Map edges to their computed weights.
228  ///
229  /// Edge weights are computed by propagating basic block weights in
230  /// SampleProfile::propagateWeights.
231  EdgeWeightMap EdgeWeights;
232 
233  /// \brief Set of visited blocks during propagation.
235 
236  /// \brief Set of visited edges during propagation.
237  SmallSet<Edge, 32> VisitedEdges;
238 
239  /// \brief Equivalence classes for block weights.
240  ///
241  /// Two blocks BB1 and BB2 are in the same equivalence class if they
242  /// dominate and post-dominate each other, and they are in the same loop
243  /// nest. When this happens, the two blocks are guaranteed to execute
244  /// the same number of times.
245  EquivalenceClassMap EquivalenceClass;
246 
247  /// Map from function name to Function *. Used to find the function from
248  /// the function name. If the function name contains suffix, additional
249  /// entry is added to map from the stripped name to the function if there
250  /// is one-to-one mapping.
251  StringMap<Function *> SymbolMap;
252 
253  /// \brief Dominance, post-dominance and loop information.
254  std::unique_ptr<DominatorTree> DT;
255  std::unique_ptr<PostDomTreeBase<BasicBlock>> PDT;
256  std::unique_ptr<LoopInfo> LI;
257 
258  std::function<AssumptionCache &(Function &)> GetAC;
259  std::function<TargetTransformInfo &(Function &)> GetTTI;
260 
261  /// \brief Predecessors for each basic block in the CFG.
262  BlockEdgeMap Predecessors;
263 
264  /// \brief Successors for each basic block in the CFG.
265  BlockEdgeMap Successors;
266 
267  SampleCoverageTracker CoverageTracker;
268 
269  /// \brief Profile reader object.
270  std::unique_ptr<SampleProfileReader> Reader;
271 
272  /// \brief Samples collected for the body of this function.
273  FunctionSamples *Samples = nullptr;
274 
275  /// \brief Name of the profile file to load.
276  std::string Filename;
277 
278  /// \brief Flag indicating whether the profile input loaded successfully.
279  bool ProfileIsValid = false;
280 
281  /// \brief Flag indicating if the pass is invoked in ThinLTO compile phase.
282  ///
283  /// In this phase, in annotation, we should not promote indirect calls.
284  /// Instead, we will mark GUIDs that needs to be annotated to the function.
285  bool IsThinLTOPreLink;
286 
287  /// \brief Total number of samples collected in this profile.
288  ///
289  /// This is the sum of all the samples collected in all the functions executed
290  /// at runtime.
291  uint64_t TotalCollectedSamples = 0;
292 
293  /// \brief Optimization Remark Emitter used to emit diagnostic remarks.
294  OptimizationRemarkEmitter *ORE = nullptr;
295 };
296 
297 class SampleProfileLoaderLegacyPass : public ModulePass {
298 public:
299  // Class identification, replacement for typeinfo
300  static char ID;
301 
302  SampleProfileLoaderLegacyPass(StringRef Name = SampleProfileFile,
303  bool IsThinLTOPreLink = false)
304  : ModulePass(ID), SampleLoader(Name, IsThinLTOPreLink,
305  [&](Function &F) -> AssumptionCache & {
306  return ACT->getAssumptionCache(F);
307  },
308  [&](Function &F) -> TargetTransformInfo & {
309  return TTIWP->getTTI(F);
310  }) {
313  }
314 
315  void dump() { SampleLoader.dump(); }
316 
317  bool doInitialization(Module &M) override {
318  return SampleLoader.doInitialization(M);
319  }
320 
321  StringRef getPassName() const override { return "Sample profile pass"; }
322  bool runOnModule(Module &M) override;
323 
324  void getAnalysisUsage(AnalysisUsage &AU) const override {
327  }
328 
329 private:
330  SampleProfileLoader SampleLoader;
331  AssumptionCacheTracker *ACT = nullptr;
332  TargetTransformInfoWrapperPass *TTIWP = nullptr;
333 };
334 
335 } // end anonymous namespace
336 
337 /// Return true if the given callsite is hot wrt to its caller.
338 ///
339 /// Functions that were inlined in the original binary will be represented
340 /// in the inline stack in the sample profile. If the profile shows that
341 /// the original inline decision was "good" (i.e., the callsite is executed
342 /// frequently), then we will recreate the inline decision and apply the
343 /// profile from the inlined callsite.
344 ///
345 /// To decide whether an inlined callsite is hot, we compute the fraction
346 /// of samples used by the callsite with respect to the total number of samples
347 /// collected in the caller.
348 ///
349 /// If that fraction is larger than the default given by
350 /// SampleProfileHotThreshold, the callsite will be inlined again.
351 static bool callsiteIsHot(const FunctionSamples *CallerFS,
352  const FunctionSamples *CallsiteFS) {
353  if (!CallsiteFS)
354  return false; // The callsite was not inlined in the original binary.
355 
356  uint64_t ParentTotalSamples = CallerFS->getTotalSamples();
357  if (ParentTotalSamples == 0)
358  return false; // Avoid division by zero.
359 
360  uint64_t CallsiteTotalSamples = CallsiteFS->getTotalSamples();
361  if (CallsiteTotalSamples == 0)
362  return false; // Callsite is trivially cold.
363 
364  double PercentSamples =
365  (double)CallsiteTotalSamples / (double)ParentTotalSamples * 100.0;
366  return PercentSamples >= SampleProfileHotThreshold;
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) const {
390  auto I = SampleCoverage.find(FS);
391 
392  // The size of the coverage map for FS represents the number of records
393  // that were marked used at least once.
394  unsigned Count = (I != SampleCoverage.end()) ? I->second.size() : 0;
395 
396  // If there are inlined callsites in this function, count the samples found
397  // in the respective bodies. However, do not bother counting callees with 0
398  // total samples, these are callees that were never invoked at runtime.
399  for (const auto &I : FS->getCallsiteSamples())
400  for (const auto &J : I.second) {
401  const FunctionSamples *CalleeSamples = &J.second;
402  if (callsiteIsHot(FS, CalleeSamples))
403  Count += countUsedRecords(CalleeSamples);
404  }
405 
406  return Count;
407 }
408 
409 /// Return the number of sample records in the body of this profile.
410 ///
411 /// This count does not include records from cold inlined callsites.
412 unsigned
413 SampleCoverageTracker::countBodyRecords(const FunctionSamples *FS) const {
414  unsigned Count = FS->getBodySamples().size();
415 
416  // Only count records in hot callsites.
417  for (const auto &I : FS->getCallsiteSamples())
418  for (const auto &J : I.second) {
419  const FunctionSamples *CalleeSamples = &J.second;
420  if (callsiteIsHot(FS, CalleeSamples))
421  Count += countBodyRecords(CalleeSamples);
422  }
423 
424  return Count;
425 }
426 
427 /// Return the number of samples collected in the body of this profile.
428 ///
429 /// This count does not include samples from cold inlined callsites.
430 uint64_t
431 SampleCoverageTracker::countBodySamples(const FunctionSamples *FS) const {
432  uint64_t Total = 0;
433  for (const auto &I : FS->getBodySamples())
434  Total += I.second.getSamples();
435 
436  // Only count samples in hot callsites.
437  for (const auto &I : FS->getCallsiteSamples())
438  for (const auto &J : I.second) {
439  const FunctionSamples *CalleeSamples = &J.second;
440  if (callsiteIsHot(FS, CalleeSamples))
441  Total += countBodySamples(CalleeSamples);
442  }
443 
444  return Total;
445 }
446 
447 /// Return the fraction of sample records used in this profile.
448 ///
449 /// The returned value is an unsigned integer in the range 0-100 indicating
450 /// the percentage of sample records that were used while applying this
451 /// profile to the associated function.
452 unsigned SampleCoverageTracker::computeCoverage(unsigned Used,
453  unsigned Total) const {
454  assert(Used <= Total &&
455  "number of used records cannot exceed the total number of records");
456  return Total > 0 ? Used * 100 / Total : 100;
457 }
458 
459 /// Clear all the per-function data used to load samples and propagate weights.
460 void SampleProfileLoader::clearFunctionData() {
461  BlockWeights.clear();
462  EdgeWeights.clear();
463  VisitedBlocks.clear();
464  VisitedEdges.clear();
465  EquivalenceClass.clear();
466  DT = nullptr;
467  PDT = nullptr;
468  LI = nullptr;
469  Predecessors.clear();
470  Successors.clear();
471  CoverageTracker.clear();
472 }
473 
474 /// Returns the line offset to the start line of the subprogram.
475 /// We assume that a single function will not exceed 65535 LOC.
476 unsigned SampleProfileLoader::getOffset(const DILocation *DIL) const {
477  return (DIL->getLine() - DIL->getScope()->getSubprogram()->getLine()) &
478  0xffff;
479 }
480 
481 #ifndef NDEBUG
482 /// \brief Print the weight of edge \p E on stream \p OS.
483 ///
484 /// \param OS Stream to emit the output to.
485 /// \param E Edge to print.
486 void SampleProfileLoader::printEdgeWeight(raw_ostream &OS, Edge E) {
487  OS << "weight[" << E.first->getName() << "->" << E.second->getName()
488  << "]: " << EdgeWeights[E] << "\n";
489 }
490 
491 /// \brief Print the equivalence class of block \p BB on stream \p OS.
492 ///
493 /// \param OS Stream to emit the output to.
494 /// \param BB Block to print.
495 void SampleProfileLoader::printBlockEquivalence(raw_ostream &OS,
496  const BasicBlock *BB) {
497  const BasicBlock *Equiv = EquivalenceClass[BB];
498  OS << "equivalence[" << BB->getName()
499  << "]: " << ((Equiv) ? EquivalenceClass[BB]->getName() : "NONE") << "\n";
500 }
501 
502 /// \brief Print the weight of block \p BB on stream \p OS.
503 ///
504 /// \param OS Stream to emit the output to.
505 /// \param BB Block to print.
506 void SampleProfileLoader::printBlockWeight(raw_ostream &OS,
507  const BasicBlock *BB) const {
508  const auto &I = BlockWeights.find(BB);
509  uint64_t W = (I == BlockWeights.end() ? 0 : I->second);
510  OS << "weight[" << BB->getName() << "]: " << W << "\n";
511 }
512 #endif
513 
514 /// \brief Get the weight for an instruction.
515 ///
516 /// The "weight" of an instruction \p Inst is the number of samples
517 /// collected on that instruction at runtime. To retrieve it, we
518 /// need to compute the line number of \p Inst relative to the start of its
519 /// function. We use HeaderLineno to compute the offset. We then
520 /// look up the samples collected for \p Inst using BodySamples.
521 ///
522 /// \param Inst Instruction to query.
523 ///
524 /// \returns the weight of \p Inst.
525 ErrorOr<uint64_t> SampleProfileLoader::getInstWeight(const Instruction &Inst) {
526  const DebugLoc &DLoc = Inst.getDebugLoc();
527  if (!DLoc)
528  return std::error_code();
529 
530  const FunctionSamples *FS = findFunctionSamples(Inst);
531  if (!FS)
532  return std::error_code();
533 
534  // Ignore all intrinsics and branch instructions.
535  // Branch instruction usually contains debug info from sources outside of
536  // the residing basic block, thus we ignore them during annotation.
537  if (isa<BranchInst>(Inst) || isa<IntrinsicInst>(Inst))
538  return std::error_code();
539 
540  // If a direct call/invoke instruction is inlined in profile
541  // (findCalleeFunctionSamples returns non-empty result), but not inlined here,
542  // it means that the inlined callsite has no sample, thus the call
543  // instruction should have 0 count.
544  if ((isa<CallInst>(Inst) || isa<InvokeInst>(Inst)) &&
545  !ImmutableCallSite(&Inst).isIndirectCall() &&
546  findCalleeFunctionSamples(Inst))
547  return 0;
548 
549  const DILocation *DIL = DLoc;
550  uint32_t LineOffset = getOffset(DIL);
551  uint32_t Discriminator = DIL->getBaseDiscriminator();
552  ErrorOr<uint64_t> R = FS->findSamplesAt(LineOffset, Discriminator);
553  if (R) {
554  bool FirstMark =
555  CoverageTracker.markSamplesUsed(FS, LineOffset, Discriminator, R.get());
556  if (FirstMark) {
557  ORE->emit([&]() {
558  OptimizationRemarkAnalysis Remark(DEBUG_TYPE, "AppliedSamples", &Inst);
559  Remark << "Applied " << ore::NV("NumSamples", *R);
560  Remark << " samples from profile (offset: ";
561  Remark << ore::NV("LineOffset", LineOffset);
562  if (Discriminator) {
563  Remark << ".";
564  Remark << ore::NV("Discriminator", Discriminator);
565  }
566  Remark << ")";
567  return Remark;
568  });
569  }
570  DEBUG(dbgs() << " " << DLoc.getLine() << "."
571  << DIL->getBaseDiscriminator() << ":" << Inst
572  << " (line offset: " << LineOffset << "."
573  << DIL->getBaseDiscriminator() << " - weight: " << R.get()
574  << ")\n");
575  }
576  return R;
577 }
578 
579 /// \brief Compute the weight of a basic block.
580 ///
581 /// The weight of basic block \p BB is the maximum weight of all the
582 /// instructions in BB.
583 ///
584 /// \param BB The basic block to query.
585 ///
586 /// \returns the weight for \p BB.
587 ErrorOr<uint64_t> SampleProfileLoader::getBlockWeight(const BasicBlock *BB) {
588  uint64_t Max = 0;
589  bool HasWeight = false;
590  for (auto &I : BB->getInstList()) {
591  const ErrorOr<uint64_t> &R = getInstWeight(I);
592  if (R) {
593  Max = std::max(Max, R.get());
594  HasWeight = true;
595  }
596  }
597  return HasWeight ? ErrorOr<uint64_t>(Max) : std::error_code();
598 }
599 
600 /// \brief Compute and store the weights of every basic block.
601 ///
602 /// This populates the BlockWeights map by computing
603 /// the weights of every basic block in the CFG.
604 ///
605 /// \param F The function to query.
606 bool SampleProfileLoader::computeBlockWeights(Function &F) {
607  bool Changed = false;
608  DEBUG(dbgs() << "Block weights\n");
609  for (const auto &BB : F) {
610  ErrorOr<uint64_t> Weight = getBlockWeight(&BB);
611  if (Weight) {
612  BlockWeights[&BB] = Weight.get();
613  VisitedBlocks.insert(&BB);
614  Changed = true;
615  }
616  DEBUG(printBlockWeight(dbgs(), &BB));
617  }
618 
619  return Changed;
620 }
621 
622 /// \brief Get the FunctionSamples for a call instruction.
623 ///
624 /// The FunctionSamples of a call/invoke instruction \p Inst is the inlined
625 /// instance in which that call instruction is calling to. It contains
626 /// all samples that resides in the inlined instance. We first find the
627 /// inlined instance in which the call instruction is from, then we
628 /// traverse its children to find the callsite with the matching
629 /// location.
630 ///
631 /// \param Inst Call/Invoke instruction to query.
632 ///
633 /// \returns The FunctionSamples pointer to the inlined instance.
634 const FunctionSamples *
635 SampleProfileLoader::findCalleeFunctionSamples(const Instruction &Inst) const {
636  const DILocation *DIL = Inst.getDebugLoc();
637  if (!DIL) {
638  return nullptr;
639  }
640 
641  StringRef CalleeName;
642  if (const CallInst *CI = dyn_cast<CallInst>(&Inst))
643  if (Function *Callee = CI->getCalledFunction())
644  CalleeName = Callee->getName();
645 
646  const FunctionSamples *FS = findFunctionSamples(Inst);
647  if (FS == nullptr)
648  return nullptr;
649 
650  return FS->findFunctionSamplesAt(
651  LineLocation(getOffset(DIL), DIL->getBaseDiscriminator()), CalleeName);
652 }
653 
654 /// Returns a vector of FunctionSamples that are the indirect call targets
655 /// of \p Inst. The vector is sorted by the total number of samples. Stores
656 /// the total call count of the indirect call in \p Sum.
657 std::vector<const FunctionSamples *>
658 SampleProfileLoader::findIndirectCallFunctionSamples(
659  const Instruction &Inst, uint64_t &Sum) const {
660  const DILocation *DIL = Inst.getDebugLoc();
661  std::vector<const FunctionSamples *> R;
662 
663  if (!DIL) {
664  return R;
665  }
666 
667  const FunctionSamples *FS = findFunctionSamples(Inst);
668  if (FS == nullptr)
669  return R;
670 
671  uint32_t LineOffset = getOffset(DIL);
672  uint32_t Discriminator = DIL->getBaseDiscriminator();
673 
674  auto T = FS->findCallTargetMapAt(LineOffset, Discriminator);
675  Sum = 0;
676  if (T)
677  for (const auto &T_C : T.get())
678  Sum += T_C.second;
681  if (M->empty())
682  return R;
683  for (const auto &NameFS : *M) {
684  Sum += NameFS.second.getEntrySamples();
685  R.push_back(&NameFS.second);
686  }
687  std::sort(R.begin(), R.end(),
688  [](const FunctionSamples *L, const FunctionSamples *R) {
689  return L->getEntrySamples() > R->getEntrySamples();
690  });
691  }
692  return R;
693 }
694 
695 /// \brief Get the FunctionSamples for an instruction.
696 ///
697 /// The FunctionSamples of an instruction \p Inst is the inlined instance
698 /// in which that instruction is coming from. We traverse the inline stack
699 /// of that instruction, and match it with the tree nodes in the profile.
700 ///
701 /// \param Inst Instruction to query.
702 ///
703 /// \returns the FunctionSamples pointer to the inlined instance.
704 const FunctionSamples *
705 SampleProfileLoader::findFunctionSamples(const Instruction &Inst) const {
707  const DILocation *DIL = Inst.getDebugLoc();
708  if (!DIL)
709  return Samples;
710 
711  const DILocation *PrevDIL = DIL;
712  for (DIL = DIL->getInlinedAt(); DIL; DIL = DIL->getInlinedAt()) {
713  S.push_back(std::make_pair(
715  PrevDIL->getScope()->getSubprogram()->getLinkageName()));
716  PrevDIL = DIL;
717  }
718  if (S.size() == 0)
719  return Samples;
720  const FunctionSamples *FS = Samples;
721  for (int i = S.size() - 1; i >= 0 && FS != nullptr; i--) {
722  FS = FS->findFunctionSamplesAt(S[i].first, S[i].second);
723  }
724  return FS;
725 }
726 
727 bool SampleProfileLoader::inlineCallInstruction(Instruction *I) {
728  assert(isa<CallInst>(I) || isa<InvokeInst>(I));
729  CallSite CS(I);
730  Function *CalledFunction = CS.getCalledFunction();
731  assert(CalledFunction);
732  DebugLoc DLoc = I->getDebugLoc();
733  BasicBlock *BB = I->getParent();
734  InlineParams Params = getInlineParams();
735  Params.ComputeFullInlineCost = true;
736  // Checks if there is anything in the reachable portion of the callee at
737  // this callsite that makes this inlining potentially illegal. Need to
738  // set ComputeFullInlineCost, otherwise getInlineCost may return early
739  // when cost exceeds threshold without checking all IRs in the callee.
740  // The acutal cost does not matter because we only checks isNever() to
741  // see if it is legal to inline the callsite.
742  InlineCost Cost = getInlineCost(CS, Params, GetTTI(*CalledFunction), GetAC,
743  None, nullptr, nullptr);
744  if (Cost.isNever()) {
745  ORE->emit(OptimizationRemark(DEBUG_TYPE, "Not inline", DLoc, BB)
746  << "incompatible inlining");
747  return false;
748  }
749  InlineFunctionInfo IFI(nullptr, &GetAC);
750  if (InlineFunction(CS, IFI)) {
751  // The call to InlineFunction erases I, so we can't pass it here.
752  ORE->emit(OptimizationRemark(DEBUG_TYPE, "HotInline", DLoc, BB)
753  << "inlined hot callee '" << ore::NV("Callee", CalledFunction)
754  << "' into '" << ore::NV("Caller", BB->getParent()) << "'");
755  return true;
756  }
757  return false;
758 }
759 
760 /// \brief Iteratively inline hot callsites of a function.
761 ///
762 /// Iteratively traverse all callsites of the function \p F, and find if
763 /// the corresponding inlined instance exists and is hot in profile. If
764 /// it is hot enough, inline the callsites and adds new callsites of the
765 /// callee into the caller. If the call is an indirect call, first promote
766 /// it to direct call. Each indirect call is limited with a single target.
767 ///
768 /// \param F function to perform iterative inlining.
769 /// \param InlinedGUIDs a set to be updated to include all GUIDs that are
770 /// inlined in the profiled binary.
771 ///
772 /// \returns True if there is any inline happened.
773 bool SampleProfileLoader::inlineHotFunctions(
774  Function &F, DenseSet<GlobalValue::GUID> &InlinedGUIDs) {
775  DenseSet<Instruction *> PromotedInsns;
776  bool Changed = false;
777  while (true) {
778  bool LocalChanged = false;
780  for (auto &BB : F) {
781  bool Hot = false;
783  for (auto &I : BB.getInstList()) {
784  const FunctionSamples *FS = nullptr;
785  if ((isa<CallInst>(I) || isa<InvokeInst>(I)) &&
786  !isa<IntrinsicInst>(I) && (FS = findCalleeFunctionSamples(I))) {
787  Candidates.push_back(&I);
788  if (callsiteIsHot(Samples, FS))
789  Hot = true;
790  }
791  }
792  if (Hot) {
793  CIS.insert(CIS.begin(), Candidates.begin(), Candidates.end());
794  }
795  }
796  for (auto I : CIS) {
797  Function *CalledFunction = CallSite(I).getCalledFunction();
798  // Do not inline recursive calls.
799  if (CalledFunction == &F)
800  continue;
801  if (CallSite(I).isIndirectCall()) {
802  if (PromotedInsns.count(I))
803  continue;
804  uint64_t Sum;
805  for (const auto *FS : findIndirectCallFunctionSamples(*I, Sum)) {
806  if (IsThinLTOPreLink) {
807  FS->findInlinedFunctions(InlinedGUIDs, F.getParent(),
808  Samples->getTotalSamples() *
810  continue;
811  }
812  auto CalleeFunctionName = FS->getName();
813  // If it is a recursive call, we do not inline it as it could bloat
814  // the code exponentially. There is way to better handle this, e.g.
815  // clone the caller first, and inline the cloned caller if it is
816  // recursive. As llvm does not inline recursive calls, we will
817  // simply ignore it instead of handling it explicitly.
818  if (CalleeFunctionName == F.getName())
819  continue;
820 
821  const char *Reason = "Callee function not available";
822  auto R = SymbolMap.find(CalleeFunctionName);
823  if (R != SymbolMap.end() && R->getValue() &&
824  !R->getValue()->isDeclaration() &&
825  R->getValue()->getSubprogram() &&
826  isLegalToPromote(I, R->getValue(), &Reason)) {
827  uint64_t C = FS->getEntrySamples();
829  I, R->getValue(), C, Sum, false, ORE);
830  Sum -= C;
831  PromotedInsns.insert(I);
832  // If profile mismatches, we should not attempt to inline DI.
833  if ((isa<CallInst>(DI) || isa<InvokeInst>(DI)) &&
834  inlineCallInstruction(DI))
835  LocalChanged = true;
836  } else {
837  DEBUG(dbgs()
838  << "\nFailed to promote indirect call to "
839  << CalleeFunctionName << " because " << Reason << "\n");
840  }
841  }
842  } else if (CalledFunction && CalledFunction->getSubprogram() &&
843  !CalledFunction->isDeclaration()) {
844  if (inlineCallInstruction(I))
845  LocalChanged = true;
846  } else if (IsThinLTOPreLink) {
847  findCalleeFunctionSamples(*I)->findInlinedFunctions(
848  InlinedGUIDs, F.getParent(),
849  Samples->getTotalSamples() * SampleProfileHotThreshold / 100);
850  }
851  }
852  if (LocalChanged) {
853  Changed = true;
854  } else {
855  break;
856  }
857  }
858  return Changed;
859 }
860 
861 /// \brief Find equivalence classes for the given block.
862 ///
863 /// This finds all the blocks that are guaranteed to execute the same
864 /// number of times as \p BB1. To do this, it traverses all the
865 /// descendants of \p BB1 in the dominator or post-dominator tree.
866 ///
867 /// A block BB2 will be in the same equivalence class as \p BB1 if
868 /// the following holds:
869 ///
870 /// 1- \p BB1 is a descendant of BB2 in the opposite tree. So, if BB2
871 /// is a descendant of \p BB1 in the dominator tree, then BB2 should
872 /// dominate BB1 in the post-dominator tree.
873 ///
874 /// 2- Both BB2 and \p BB1 must be in the same loop.
875 ///
876 /// For every block BB2 that meets those two requirements, we set BB2's
877 /// equivalence class to \p BB1.
878 ///
879 /// \param BB1 Block to check.
880 /// \param Descendants Descendants of \p BB1 in either the dom or pdom tree.
881 /// \param DomTree Opposite dominator tree. If \p Descendants is filled
882 /// with blocks from \p BB1's dominator tree, then
883 /// this is the post-dominator tree, and vice versa.
884 template <bool IsPostDom>
885 void SampleProfileLoader::findEquivalencesFor(
886  BasicBlock *BB1, ArrayRef<BasicBlock *> Descendants,
888  const BasicBlock *EC = EquivalenceClass[BB1];
889  uint64_t Weight = BlockWeights[EC];
890  for (const auto *BB2 : Descendants) {
891  bool IsDomParent = DomTree->dominates(BB2, BB1);
892  bool IsInSameLoop = LI->getLoopFor(BB1) == LI->getLoopFor(BB2);
893  if (BB1 != BB2 && IsDomParent && IsInSameLoop) {
894  EquivalenceClass[BB2] = EC;
895  // If BB2 is visited, then the entire EC should be marked as visited.
896  if (VisitedBlocks.count(BB2)) {
897  VisitedBlocks.insert(EC);
898  }
899 
900  // If BB2 is heavier than BB1, make BB2 have the same weight
901  // as BB1.
902  //
903  // Note that we don't worry about the opposite situation here
904  // (when BB2 is lighter than BB1). We will deal with this
905  // during the propagation phase. Right now, we just want to
906  // make sure that BB1 has the largest weight of all the
907  // members of its equivalence set.
908  Weight = std::max(Weight, BlockWeights[BB2]);
909  }
910  }
911  if (EC == &EC->getParent()->getEntryBlock()) {
912  BlockWeights[EC] = Samples->getHeadSamples() + 1;
913  } else {
914  BlockWeights[EC] = Weight;
915  }
916 }
917 
918 /// \brief Find equivalence classes.
919 ///
920 /// Since samples may be missing from blocks, we can fill in the gaps by setting
921 /// the weights of all the blocks in the same equivalence class to the same
922 /// weight. To compute the concept of equivalence, we use dominance and loop
923 /// information. Two blocks B1 and B2 are in the same equivalence class if B1
924 /// dominates B2, B2 post-dominates B1 and both are in the same loop.
925 ///
926 /// \param F The function to query.
927 void SampleProfileLoader::findEquivalenceClasses(Function &F) {
928  SmallVector<BasicBlock *, 8> DominatedBBs;
929  DEBUG(dbgs() << "\nBlock equivalence classes\n");
930  // Find equivalence sets based on dominance and post-dominance information.
931  for (auto &BB : F) {
932  BasicBlock *BB1 = &BB;
933 
934  // Compute BB1's equivalence class once.
935  if (EquivalenceClass.count(BB1)) {
936  DEBUG(printBlockEquivalence(dbgs(), BB1));
937  continue;
938  }
939 
940  // By default, blocks are in their own equivalence class.
941  EquivalenceClass[BB1] = BB1;
942 
943  // Traverse all the blocks dominated by BB1. We are looking for
944  // every basic block BB2 such that:
945  //
946  // 1- BB1 dominates BB2.
947  // 2- BB2 post-dominates BB1.
948  // 3- BB1 and BB2 are in the same loop nest.
949  //
950  // If all those conditions hold, it means that BB2 is executed
951  // as many times as BB1, so they are placed in the same equivalence
952  // class by making BB2's equivalence class be BB1.
953  DominatedBBs.clear();
954  DT->getDescendants(BB1, DominatedBBs);
955  findEquivalencesFor(BB1, DominatedBBs, PDT.get());
956 
957  DEBUG(printBlockEquivalence(dbgs(), BB1));
958  }
959 
960  // Assign weights to equivalence classes.
961  //
962  // All the basic blocks in the same equivalence class will execute
963  // the same number of times. Since we know that the head block in
964  // each equivalence class has the largest weight, assign that weight
965  // to all the blocks in that equivalence class.
966  DEBUG(dbgs() << "\nAssign the same weight to all blocks in the same class\n");
967  for (auto &BI : F) {
968  const BasicBlock *BB = &BI;
969  const BasicBlock *EquivBB = EquivalenceClass[BB];
970  if (BB != EquivBB)
971  BlockWeights[BB] = BlockWeights[EquivBB];
972  DEBUG(printBlockWeight(dbgs(), BB));
973  }
974 }
975 
976 /// \brief Visit the given edge to decide if it has a valid weight.
977 ///
978 /// If \p E has not been visited before, we copy to \p UnknownEdge
979 /// and increment the count of unknown edges.
980 ///
981 /// \param E Edge to visit.
982 /// \param NumUnknownEdges Current number of unknown edges.
983 /// \param UnknownEdge Set if E has not been visited before.
984 ///
985 /// \returns E's weight, if known. Otherwise, return 0.
986 uint64_t SampleProfileLoader::visitEdge(Edge E, unsigned *NumUnknownEdges,
987  Edge *UnknownEdge) {
988  if (!VisitedEdges.count(E)) {
989  (*NumUnknownEdges)++;
990  *UnknownEdge = E;
991  return 0;
992  }
993 
994  return EdgeWeights[E];
995 }
996 
997 /// \brief Propagate weights through incoming/outgoing edges.
998 ///
999 /// If the weight of a basic block is known, and there is only one edge
1000 /// with an unknown weight, we can calculate the weight of that edge.
1001 ///
1002 /// Similarly, if all the edges have a known count, we can calculate the
1003 /// count of the basic block, if needed.
1004 ///
1005 /// \param F Function to process.
1006 /// \param UpdateBlockCount Whether we should update basic block counts that
1007 /// has already been annotated.
1008 ///
1009 /// \returns True if new weights were assigned to edges or blocks.
1010 bool SampleProfileLoader::propagateThroughEdges(Function &F,
1011  bool UpdateBlockCount) {
1012  bool Changed = false;
1013  DEBUG(dbgs() << "\nPropagation through edges\n");
1014  for (const auto &BI : F) {
1015  const BasicBlock *BB = &BI;
1016  const BasicBlock *EC = EquivalenceClass[BB];
1017 
1018  // Visit all the predecessor and successor edges to determine
1019  // which ones have a weight assigned already. Note that it doesn't
1020  // matter that we only keep track of a single unknown edge. The
1021  // only case we are interested in handling is when only a single
1022  // edge is unknown (see setEdgeOrBlockWeight).
1023  for (unsigned i = 0; i < 2; i++) {
1024  uint64_t TotalWeight = 0;
1025  unsigned NumUnknownEdges = 0, NumTotalEdges = 0;
1026  Edge UnknownEdge, SelfReferentialEdge, SingleEdge;
1027 
1028  if (i == 0) {
1029  // First, visit all predecessor edges.
1030  NumTotalEdges = Predecessors[BB].size();
1031  for (auto *Pred : Predecessors[BB]) {
1032  Edge E = std::make_pair(Pred, BB);
1033  TotalWeight += visitEdge(E, &NumUnknownEdges, &UnknownEdge);
1034  if (E.first == E.second)
1035  SelfReferentialEdge = E;
1036  }
1037  if (NumTotalEdges == 1) {
1038  SingleEdge = std::make_pair(Predecessors[BB][0], BB);
1039  }
1040  } else {
1041  // On the second round, visit all successor edges.
1042  NumTotalEdges = Successors[BB].size();
1043  for (auto *Succ : Successors[BB]) {
1044  Edge E = std::make_pair(BB, Succ);
1045  TotalWeight += visitEdge(E, &NumUnknownEdges, &UnknownEdge);
1046  }
1047  if (NumTotalEdges == 1) {
1048  SingleEdge = std::make_pair(BB, Successors[BB][0]);
1049  }
1050  }
1051 
1052  // After visiting all the edges, there are three cases that we
1053  // can handle immediately:
1054  //
1055  // - All the edge weights are known (i.e., NumUnknownEdges == 0).
1056  // In this case, we simply check that the sum of all the edges
1057  // is the same as BB's weight. If not, we change BB's weight
1058  // to match. Additionally, if BB had not been visited before,
1059  // we mark it visited.
1060  //
1061  // - Only one edge is unknown and BB has already been visited.
1062  // In this case, we can compute the weight of the edge by
1063  // subtracting the total block weight from all the known
1064  // edge weights. If the edges weight more than BB, then the
1065  // edge of the last remaining edge is set to zero.
1066  //
1067  // - There exists a self-referential edge and the weight of BB is
1068  // known. In this case, this edge can be based on BB's weight.
1069  // We add up all the other known edges and set the weight on
1070  // the self-referential edge as we did in the previous case.
1071  //
1072  // In any other case, we must continue iterating. Eventually,
1073  // all edges will get a weight, or iteration will stop when
1074  // it reaches SampleProfileMaxPropagateIterations.
1075  if (NumUnknownEdges <= 1) {
1076  uint64_t &BBWeight = BlockWeights[EC];
1077  if (NumUnknownEdges == 0) {
1078  if (!VisitedBlocks.count(EC)) {
1079  // If we already know the weight of all edges, the weight of the
1080  // basic block can be computed. It should be no larger than the sum
1081  // of all edge weights.
1082  if (TotalWeight > BBWeight) {
1083  BBWeight = TotalWeight;
1084  Changed = true;
1085  DEBUG(dbgs() << "All edge weights for " << BB->getName()
1086  << " known. Set weight for block: ";
1087  printBlockWeight(dbgs(), BB););
1088  }
1089  } else if (NumTotalEdges == 1 &&
1090  EdgeWeights[SingleEdge] < BlockWeights[EC]) {
1091  // If there is only one edge for the visited basic block, use the
1092  // block weight to adjust edge weight if edge weight is smaller.
1093  EdgeWeights[SingleEdge] = BlockWeights[EC];
1094  Changed = true;
1095  }
1096  } else if (NumUnknownEdges == 1 && VisitedBlocks.count(EC)) {
1097  // If there is a single unknown edge and the block has been
1098  // visited, then we can compute E's weight.
1099  if (BBWeight >= TotalWeight)
1100  EdgeWeights[UnknownEdge] = BBWeight - TotalWeight;
1101  else
1102  EdgeWeights[UnknownEdge] = 0;
1103  const BasicBlock *OtherEC;
1104  if (i == 0)
1105  OtherEC = EquivalenceClass[UnknownEdge.first];
1106  else
1107  OtherEC = EquivalenceClass[UnknownEdge.second];
1108  // Edge weights should never exceed the BB weights it connects.
1109  if (VisitedBlocks.count(OtherEC) &&
1110  EdgeWeights[UnknownEdge] > BlockWeights[OtherEC])
1111  EdgeWeights[UnknownEdge] = BlockWeights[OtherEC];
1112  VisitedEdges.insert(UnknownEdge);
1113  Changed = true;
1114  DEBUG(dbgs() << "Set weight for edge: ";
1115  printEdgeWeight(dbgs(), UnknownEdge));
1116  }
1117  } else if (VisitedBlocks.count(EC) && BlockWeights[EC] == 0) {
1118  // If a block Weights 0, all its in/out edges should weight 0.
1119  if (i == 0) {
1120  for (auto *Pred : Predecessors[BB]) {
1121  Edge E = std::make_pair(Pred, BB);
1122  EdgeWeights[E] = 0;
1123  VisitedEdges.insert(E);
1124  }
1125  } else {
1126  for (auto *Succ : Successors[BB]) {
1127  Edge E = std::make_pair(BB, Succ);
1128  EdgeWeights[E] = 0;
1129  VisitedEdges.insert(E);
1130  }
1131  }
1132  } else if (SelfReferentialEdge.first && VisitedBlocks.count(EC)) {
1133  uint64_t &BBWeight = BlockWeights[BB];
1134  // We have a self-referential edge and the weight of BB is known.
1135  if (BBWeight >= TotalWeight)
1136  EdgeWeights[SelfReferentialEdge] = BBWeight - TotalWeight;
1137  else
1138  EdgeWeights[SelfReferentialEdge] = 0;
1139  VisitedEdges.insert(SelfReferentialEdge);
1140  Changed = true;
1141  DEBUG(dbgs() << "Set self-referential edge weight to: ";
1142  printEdgeWeight(dbgs(), SelfReferentialEdge));
1143  }
1144  if (UpdateBlockCount && !VisitedBlocks.count(EC) && TotalWeight > 0) {
1145  BlockWeights[EC] = TotalWeight;
1146  VisitedBlocks.insert(EC);
1147  Changed = true;
1148  }
1149  }
1150  }
1151 
1152  return Changed;
1153 }
1154 
1155 /// \brief Build in/out edge lists for each basic block in the CFG.
1156 ///
1157 /// We are interested in unique edges. If a block B1 has multiple
1158 /// edges to another block B2, we only add a single B1->B2 edge.
1159 void SampleProfileLoader::buildEdges(Function &F) {
1160  for (auto &BI : F) {
1161  BasicBlock *B1 = &BI;
1162 
1163  // Add predecessors for B1.
1165  if (!Predecessors[B1].empty())
1166  llvm_unreachable("Found a stale predecessors list in a basic block.");
1167  for (pred_iterator PI = pred_begin(B1), PE = pred_end(B1); PI != PE; ++PI) {
1168  BasicBlock *B2 = *PI;
1169  if (Visited.insert(B2).second)
1170  Predecessors[B1].push_back(B2);
1171  }
1172 
1173  // Add successors for B1.
1174  Visited.clear();
1175  if (!Successors[B1].empty())
1176  llvm_unreachable("Found a stale successors list in a basic block.");
1177  for (succ_iterator SI = succ_begin(B1), SE = succ_end(B1); SI != SE; ++SI) {
1178  BasicBlock *B2 = *SI;
1179  if (Visited.insert(B2).second)
1180  Successors[B1].push_back(B2);
1181  }
1182  }
1183 }
1184 
1185 /// Returns the sorted CallTargetMap \p M by count in descending order.
1187  const SampleRecord::CallTargetMap &M) {
1189  for (auto I = M.begin(); I != M.end(); ++I)
1190  R.push_back({Function::getGUID(I->getKey()), I->getValue()});
1191  std::sort(R.begin(), R.end(),
1192  [](const InstrProfValueData &L, const InstrProfValueData &R) {
1193  if (L.Count == R.Count)
1194  return L.Value > R.Value;
1195  else
1196  return L.Count > R.Count;
1197  });
1198  return R;
1199 }
1200 
1201 /// \brief Propagate weights into edges
1202 ///
1203 /// The following rules are applied to every block BB in the CFG:
1204 ///
1205 /// - If BB has a single predecessor/successor, then the weight
1206 /// of that edge is the weight of the block.
1207 ///
1208 /// - If all incoming or outgoing edges are known except one, and the
1209 /// weight of the block is already known, the weight of the unknown
1210 /// edge will be the weight of the block minus the sum of all the known
1211 /// edges. If the sum of all the known edges is larger than BB's weight,
1212 /// we set the unknown edge weight to zero.
1213 ///
1214 /// - If there is a self-referential edge, and the weight of the block is
1215 /// known, the weight for that edge is set to the weight of the block
1216 /// minus the weight of the other incoming edges to that block (if
1217 /// known).
1218 void SampleProfileLoader::propagateWeights(Function &F) {
1219  bool Changed = true;
1220  unsigned I = 0;
1221 
1222  // If BB weight is larger than its corresponding loop's header BB weight,
1223  // use the BB weight to replace the loop header BB weight.
1224  for (auto &BI : F) {
1225  BasicBlock *BB = &BI;
1226  Loop *L = LI->getLoopFor(BB);
1227  if (!L) {
1228  continue;
1229  }
1230  BasicBlock *Header = L->getHeader();
1231  if (Header && BlockWeights[BB] > BlockWeights[Header]) {
1232  BlockWeights[Header] = BlockWeights[BB];
1233  }
1234  }
1235 
1236  // Before propagation starts, build, for each block, a list of
1237  // unique predecessors and successors. This is necessary to handle
1238  // identical edges in multiway branches. Since we visit all blocks and all
1239  // edges of the CFG, it is cleaner to build these lists once at the start
1240  // of the pass.
1241  buildEdges(F);
1242 
1243  // Propagate until we converge or we go past the iteration limit.
1244  while (Changed && I++ < SampleProfileMaxPropagateIterations) {
1245  Changed = propagateThroughEdges(F, false);
1246  }
1247 
1248  // The first propagation propagates BB counts from annotated BBs to unknown
1249  // BBs. The 2nd propagation pass resets edges weights, and use all BB weights
1250  // to propagate edge weights.
1251  VisitedEdges.clear();
1252  Changed = true;
1253  while (Changed && I++ < SampleProfileMaxPropagateIterations) {
1254  Changed = propagateThroughEdges(F, false);
1255  }
1256 
1257  // The 3rd propagation pass allows adjust annotated BB weights that are
1258  // obviously wrong.
1259  Changed = true;
1260  while (Changed && I++ < SampleProfileMaxPropagateIterations) {
1261  Changed = propagateThroughEdges(F, true);
1262  }
1263 
1264  // Generate MD_prof metadata for every branch instruction using the
1265  // edge weights computed during propagation.
1266  DEBUG(dbgs() << "\nPropagation complete. Setting branch weights\n");
1267  LLVMContext &Ctx = F.getContext();
1268  MDBuilder MDB(Ctx);
1269  for (auto &BI : F) {
1270  BasicBlock *BB = &BI;
1271 
1272  if (BlockWeights[BB]) {
1273  for (auto &I : BB->getInstList()) {
1274  if (!isa<CallInst>(I) && !isa<InvokeInst>(I))
1275  continue;
1276  CallSite CS(&I);
1277  if (!CS.getCalledFunction()) {
1278  const DebugLoc &DLoc = I.getDebugLoc();
1279  if (!DLoc)
1280  continue;
1281  const DILocation *DIL = DLoc;
1282  uint32_t LineOffset = getOffset(DIL);
1283  uint32_t Discriminator = DIL->getBaseDiscriminator();
1284 
1285  const FunctionSamples *FS = findFunctionSamples(I);
1286  if (!FS)
1287  continue;
1288  auto T = FS->findCallTargetMapAt(LineOffset, Discriminator);
1289  if (!T || T.get().empty())
1290  continue;
1291  SmallVector<InstrProfValueData, 2> SortedCallTargets =
1292  SortCallTargets(T.get());
1293  uint64_t Sum;
1294  findIndirectCallFunctionSamples(I, Sum);
1295  annotateValueSite(*I.getParent()->getParent()->getParent(), I,
1296  SortedCallTargets, Sum, IPVK_IndirectCallTarget,
1297  SortedCallTargets.size());
1298  } else if (!dyn_cast<IntrinsicInst>(&I)) {
1299  SmallVector<uint32_t, 1> Weights;
1300  Weights.push_back(BlockWeights[BB]);
1301  I.setMetadata(LLVMContext::MD_prof, MDB.createBranchWeights(Weights));
1302  }
1303  }
1304  }
1305  TerminatorInst *TI = BB->getTerminator();
1306  if (TI->getNumSuccessors() == 1)
1307  continue;
1308  if (!isa<BranchInst>(TI) && !isa<SwitchInst>(TI))
1309  continue;
1310 
1311  DebugLoc BranchLoc = TI->getDebugLoc();
1312  DEBUG(dbgs() << "\nGetting weights for branch at line "
1313  << ((BranchLoc) ? Twine(BranchLoc.getLine())
1314  : Twine("<UNKNOWN LOCATION>"))
1315  << ".\n");
1316  SmallVector<uint32_t, 4> Weights;
1317  uint32_t MaxWeight = 0;
1318  Instruction *MaxDestInst;
1319  for (unsigned I = 0; I < TI->getNumSuccessors(); ++I) {
1320  BasicBlock *Succ = TI->getSuccessor(I);
1321  Edge E = std::make_pair(BB, Succ);
1322  uint64_t Weight = EdgeWeights[E];
1323  DEBUG(dbgs() << "\t"; printEdgeWeight(dbgs(), E));
1324  // Use uint32_t saturated arithmetic to adjust the incoming weights,
1325  // if needed. Sample counts in profiles are 64-bit unsigned values,
1326  // but internally branch weights are expressed as 32-bit values.
1327  if (Weight > std::numeric_limits<uint32_t>::max()) {
1328  DEBUG(dbgs() << " (saturated due to uint32_t overflow)");
1330  }
1331  // Weight is added by one to avoid propagation errors introduced by
1332  // 0 weights.
1333  Weights.push_back(static_cast<uint32_t>(Weight + 1));
1334  if (Weight != 0) {
1335  if (Weight > MaxWeight) {
1336  MaxWeight = Weight;
1337  MaxDestInst = Succ->getFirstNonPHIOrDbgOrLifetime();
1338  }
1339  }
1340  }
1341 
1342  uint64_t TempWeight;
1343  // Only set weights if there is at least one non-zero weight.
1344  // In any other case, let the analyzer set weights.
1345  // Do not set weights if the weights are present. In ThinLTO, the profile
1346  // annotation is done twice. If the first annotation already set the
1347  // weights, the second pass does not need to set it.
1348  if (MaxWeight > 0 && !TI->extractProfTotalWeight(TempWeight)) {
1349  DEBUG(dbgs() << "SUCCESS. Found non-zero weights.\n");
1351  MDB.createBranchWeights(Weights));
1352  ORE->emit([&]() {
1353  return OptimizationRemark(DEBUG_TYPE, "PopularDest", MaxDestInst)
1354  << "most popular destination for conditional branches at "
1355  << ore::NV("CondBranchesLoc", BranchLoc);
1356  });
1357  } else {
1358  DEBUG(dbgs() << "SKIPPED. All branch weights are zero.\n");
1359  }
1360  }
1361 }
1362 
1363 /// \brief Get the line number for the function header.
1364 ///
1365 /// This looks up function \p F in the current compilation unit and
1366 /// retrieves the line number where the function is defined. This is
1367 /// line 0 for all the samples read from the profile file. Every line
1368 /// number is relative to this line.
1369 ///
1370 /// \param F Function object to query.
1371 ///
1372 /// \returns the line number where \p F is defined. If it returns 0,
1373 /// it means that there is no debug information available for \p F.
1374 unsigned SampleProfileLoader::getFunctionLoc(Function &F) {
1375  if (DISubprogram *S = F.getSubprogram())
1376  return S->getLine();
1377 
1378  // If the start of \p F is missing, emit a diagnostic to inform the user
1379  // about the missed opportunity.
1381  "No debug information found in function " + F.getName() +
1382  ": Function profile not used",
1383  DS_Warning));
1384  return 0;
1385 }
1386 
1387 void SampleProfileLoader::computeDominanceAndLoopInfo(Function &F) {
1388  DT.reset(new DominatorTree);
1389  DT->recalculate(F);
1390 
1391  PDT.reset(new PostDomTreeBase<BasicBlock>());
1392  PDT->recalculate(F);
1393 
1394  LI.reset(new LoopInfo);
1395  LI->analyze(*DT);
1396 }
1397 
1398 /// \brief Generate branch weight metadata for all branches in \p F.
1399 ///
1400 /// Branch weights are computed out of instruction samples using a
1401 /// propagation heuristic. Propagation proceeds in 3 phases:
1402 ///
1403 /// 1- Assignment of block weights. All the basic blocks in the function
1404 /// are initial assigned the same weight as their most frequently
1405 /// executed instruction.
1406 ///
1407 /// 2- Creation of equivalence classes. Since samples may be missing from
1408 /// blocks, we can fill in the gaps by setting the weights of all the
1409 /// blocks in the same equivalence class to the same weight. To compute
1410 /// the concept of equivalence, we use dominance and loop information.
1411 /// Two blocks B1 and B2 are in the same equivalence class if B1
1412 /// dominates B2, B2 post-dominates B1 and both are in the same loop.
1413 ///
1414 /// 3- Propagation of block weights into edges. This uses a simple
1415 /// propagation heuristic. The following rules are applied to every
1416 /// block BB in the CFG:
1417 ///
1418 /// - If BB has a single predecessor/successor, then the weight
1419 /// of that edge is the weight of the block.
1420 ///
1421 /// - If all the edges are known except one, and the weight of the
1422 /// block is already known, the weight of the unknown edge will
1423 /// be the weight of the block minus the sum of all the known
1424 /// edges. If the sum of all the known edges is larger than BB's weight,
1425 /// we set the unknown edge weight to zero.
1426 ///
1427 /// - If there is a self-referential edge, and the weight of the block is
1428 /// known, the weight for that edge is set to the weight of the block
1429 /// minus the weight of the other incoming edges to that block (if
1430 /// known).
1431 ///
1432 /// Since this propagation is not guaranteed to finalize for every CFG, we
1433 /// only allow it to proceed for a limited number of iterations (controlled
1434 /// by -sample-profile-max-propagate-iterations).
1435 ///
1436 /// FIXME: Try to replace this propagation heuristic with a scheme
1437 /// that is guaranteed to finalize. A work-list approach similar to
1438 /// the standard value propagation algorithm used by SSA-CCP might
1439 /// work here.
1440 ///
1441 /// Once all the branch weights are computed, we emit the MD_prof
1442 /// metadata on BB using the computed values for each of its branches.
1443 ///
1444 /// \param F The function to query.
1445 ///
1446 /// \returns true if \p F was modified. Returns false, otherwise.
1447 bool SampleProfileLoader::emitAnnotations(Function &F) {
1448  bool Changed = false;
1449 
1450  if (getFunctionLoc(F) == 0)
1451  return false;
1452 
1453  DEBUG(dbgs() << "Line number for the first instruction in " << F.getName()
1454  << ": " << getFunctionLoc(F) << "\n");
1455 
1456  DenseSet<GlobalValue::GUID> InlinedGUIDs;
1457  Changed |= inlineHotFunctions(F, InlinedGUIDs);
1458 
1459  // Compute basic block weights.
1460  Changed |= computeBlockWeights(F);
1461 
1462  if (Changed) {
1463  // Add an entry count to the function using the samples gathered at the
1464  // function entry.
1465  // Sets the GUIDs that are inlined in the profiled binary. This is used
1466  // for ThinLink to make correct liveness analysis, and also make the IR
1467  // match the profiled binary before annotation.
1468  F.setEntryCount(Samples->getHeadSamples() + 1, &InlinedGUIDs);
1469 
1470  // Compute dominance and loop info needed for propagation.
1471  computeDominanceAndLoopInfo(F);
1472 
1473  // Find equivalence classes.
1474  findEquivalenceClasses(F);
1475 
1476  // Propagate weights to all edges.
1477  propagateWeights(F);
1478  }
1479 
1480  // If coverage checking was requested, compute it now.
1482  unsigned Used = CoverageTracker.countUsedRecords(Samples);
1483  unsigned Total = CoverageTracker.countBodyRecords(Samples);
1484  unsigned Coverage = CoverageTracker.computeCoverage(Used, Total);
1485  if (Coverage < SampleProfileRecordCoverage) {
1487  F.getSubprogram()->getFilename(), getFunctionLoc(F),
1488  Twine(Used) + " of " + Twine(Total) + " available profile records (" +
1489  Twine(Coverage) + "%) were applied",
1490  DS_Warning));
1491  }
1492  }
1493 
1495  uint64_t Used = CoverageTracker.getTotalUsedSamples();
1496  uint64_t Total = CoverageTracker.countBodySamples(Samples);
1497  unsigned Coverage = CoverageTracker.computeCoverage(Used, Total);
1498  if (Coverage < SampleProfileSampleCoverage) {
1500  F.getSubprogram()->getFilename(), getFunctionLoc(F),
1501  Twine(Used) + " of " + Twine(Total) + " available profile samples (" +
1502  Twine(Coverage) + "%) were applied",
1503  DS_Warning));
1504  }
1505  }
1506  return Changed;
1507 }
1508 
1510 
1511 INITIALIZE_PASS_BEGIN(SampleProfileLoaderLegacyPass, "sample-profile",
1512  "Sample Profile loader", false, false)
1515 INITIALIZE_PASS_END(SampleProfileLoaderLegacyPass, "sample-profile",
1516  "Sample Profile loader", false, false)
1517 
1518 bool SampleProfileLoader::doInitialization(Module &M) {
1519  auto &Ctx = M.getContext();
1520  auto ReaderOrErr = SampleProfileReader::create(Filename, Ctx);
1521  if (std::error_code EC = ReaderOrErr.getError()) {
1522  std::string Msg = "Could not open profile: " + EC.message();
1523  Ctx.diagnose(DiagnosticInfoSampleProfile(Filename, Msg));
1524  return false;
1525  }
1526  Reader = std::move(ReaderOrErr.get());
1527  ProfileIsValid = (Reader->read() == sampleprof_error::success);
1528  return true;
1529 }
1530 
1532  return new SampleProfileLoaderLegacyPass(SampleProfileFile);
1533 }
1534 
1536  return new SampleProfileLoaderLegacyPass(Name);
1537 }
1538 
1539 bool SampleProfileLoader::runOnModule(Module &M, ModuleAnalysisManager *AM) {
1540  if (!ProfileIsValid)
1541  return false;
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  std::string 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 != std::string::npos) {
1556  std::string 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  if (M.getProfileSummary() == nullptr)
1574  M.setProfileSummary(Reader->getSummary().getMD(M.getContext()));
1575  return retval;
1576 }
1577 
1578 bool SampleProfileLoaderLegacyPass::runOnModule(Module &M) {
1579  ACT = &getAnalysis<AssumptionCacheTracker>();
1580  TTIWP = &getAnalysis<TargetTransformInfoWrapperPass>();
1581  return SampleLoader.runOnModule(M, nullptr);
1582 }
1583 
1585  F.setEntryCount(0);
1586  std::unique_ptr<OptimizationRemarkEmitter> OwnedORE;
1587  if (AM) {
1588  auto &FAM =
1590  .getManager();
1591  ORE = &FAM.getResult<OptimizationRemarkEmitterAnalysis>(F);
1592  } else {
1593  OwnedORE = make_unique<OptimizationRemarkEmitter>(&F);
1594  ORE = OwnedORE.get();
1595  }
1596  Samples = Reader->getSamplesFor(F);
1597  if (Samples && !Samples->empty())
1598  return emitAnnotations(F);
1599  return false;
1600 }
1601 
1603  ModuleAnalysisManager &AM) {
1605  AM.getResult<FunctionAnalysisManagerModuleProxy>(M).getManager();
1606 
1607  auto GetAssumptionCache = [&](Function &F) -> AssumptionCache & {
1608  return FAM.getResult<AssumptionAnalysis>(F);
1609  };
1610  auto GetTTI = [&](Function &F) -> TargetTransformInfo & {
1611  return FAM.getResult<TargetIRAnalysis>(F);
1612  };
1613 
1614  SampleProfileLoader SampleLoader(
1615  ProfileFileName.empty() ? SampleProfileFile : ProfileFileName,
1616  IsThinLTOPreLink, GetAssumptionCache, GetTTI);
1617 
1618  SampleLoader.doInitialization(M);
1619 
1620  if (!SampleLoader.runOnModule(M, &AM))
1621  return PreservedAnalyses::all();
1622 
1623  return PreservedAnalyses::none();
1624 }
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:263
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.
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: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:236
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:359
const StringRef & getName() const
Return the function name.
Definition: SampleProf.h:382
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:196
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
Twine - A lightweight data structure for efficiently representing the concatenation of temporary valu...
Definition: Twine.h:81
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:1328
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)
Debug location.
ModulePass * createSampleProfileLoaderPass()
Concrete subclass of DominatorTreeBase that is used to compute a normal dominator tree...
Definition: Dominators.h:140
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:572
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:1497
const BodySampleMap & getBodySamples() const
Return all the samples collected in the body of the function.
Definition: SampleProf.h:327
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
Represent the analysis usage information of a pass.
bool isLegalToPromote(Instruction *Inst, Function *F, const char **Reason)
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
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:1214
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:274
A function analysis which provides an AssumptionCache.
void setEntryCount(uint64_t Count, const DenseSet< GlobalValue::GUID > *Imports=nullptr)
Set the entry count for this function.
Definition: Function.cpp:1324
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:864
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:249
raw_ostream & dbgs()
dbgs() - This returns a reference to a raw_ostream for debugging messages.
Definition: Debug.cpp:132
Instruction * promoteIndirectCall(Instruction *Inst, Function *F, uint64_t Count, uint64_t TotalCount, bool AttachProfToDirectCall, OptimizationRemarkEmitter *ORE)
StringMap - This is an unconventional map that is specialized for handling keys that are "strings"...
Definition: StringMap.h:224
static void clear(coro::Shape &Shape)
Definition: Coroutines.cpp:210
iterator insert(iterator I, T &&Elt)
Definition: SmallVector.h:482
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:284
#define DEBUG_TYPE
iterator begin()
Definition: StringMap.h:319
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:220
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
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:556
uint64_t getTotalSamples() const
Return the total number of samples collected inside the function.
Definition: SampleProf.h:297
sample Sample Profile loader
const CallsiteSampleMap & getCallsiteSamples() const
Return all the callsite samples collected in the body of the function.
Definition: SampleProf.h:330
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:308
This header defines various interfaces for pass management in LLVM.
Diagnostic information for the sample profiler.
The optimization diagnostic interface.
iterator end()
Definition: StringMap.h:322
This file provides the interface for the sampled PGO loader pass.
reference get()
Definition: ErrorOr.h:169
const BasicBlock * getParent() const
Definition: Instruction.h:66
An analysis over an "outer" IR unit that provides access to an analysis manager over an "inner" IR un...
Definition: PassManager.h:946