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"
43 #include "llvm/IR/BasicBlock.h"
44 #include "llvm/IR/CFG.h"
45 #include "llvm/IR/CallSite.h"
47 #include "llvm/IR/DebugLoc.h"
48 #include "llvm/IR/DiagnosticInfo.h"
49 #include "llvm/IR/Dominators.h"
50 #include "llvm/IR/Function.h"
51 #include "llvm/IR/GlobalValue.h"
52 #include "llvm/IR/InstrTypes.h"
53 #include "llvm/IR/Instruction.h"
54 #include "llvm/IR/Instructions.h"
55 #include "llvm/IR/IntrinsicInst.h"
56 #include "llvm/IR/LLVMContext.h"
57 #include "llvm/IR/MDBuilder.h"
58 #include "llvm/IR/Module.h"
59 #include "llvm/IR/PassManager.h"
61 #include "llvm/Pass.h"
65 #include "llvm/Support/Casting.h"
67 #include "llvm/Support/Debug.h"
69 #include "llvm/Support/ErrorOr.h"
72 #include "llvm/Transforms/IPO.h"
76 #include <algorithm>
77 #include <cassert>
78 #include <cstdint>
79 #include <functional>
80 #include <limits>
81 #include <map>
82 #include <memory>
83 #include <string>
84 #include <system_error>
85 #include <utility>
86 #include <vector>
87 
88 using namespace llvm;
89 using namespace sampleprof;
91 #define DEBUG_TYPE "sample-profile"
92 
93 // Command line option to specify the file to read samples from. This is
94 // mainly used for debugging.
96  "sample-profile-file", cl::init(""), cl::value_desc("filename"),
97  cl::desc("Profile file loaded by -sample-profile"), cl::Hidden);
98 
100  "sample-profile-max-propagate-iterations", cl::init(100),
101  cl::desc("Maximum number of iterations to go through when propagating "
102  "sample block/edge weights through the CFG."));
103 
105  "sample-profile-check-record-coverage", cl::init(0), cl::value_desc("N"),
106  cl::desc("Emit a warning if less than N% of records in the input profile "
107  "are matched to the IR."));
108 
110  "sample-profile-check-sample-coverage", cl::init(0), cl::value_desc("N"),
111  cl::desc("Emit a warning if less than N% of samples in the input profile "
112  "are matched to the IR."));
113 
114 namespace {
115 
116 using BlockWeightMap = DenseMap<const BasicBlock *, uint64_t>;
117 using EquivalenceClassMap = DenseMap<const BasicBlock *, const BasicBlock *>;
118 using Edge = std::pair<const BasicBlock *, const BasicBlock *>;
119 using EdgeWeightMap = DenseMap<Edge, uint64_t>;
120 using BlockEdgeMap =
122 
123 class SampleCoverageTracker {
124 public:
125  SampleCoverageTracker() = default;
126 
127  bool markSamplesUsed(const FunctionSamples *FS, uint32_t LineOffset,
128  uint32_t Discriminator, uint64_t Samples);
129  unsigned computeCoverage(unsigned Used, unsigned Total) const;
130  unsigned countUsedRecords(const FunctionSamples *FS,
131  ProfileSummaryInfo *PSI) const;
132  unsigned countBodyRecords(const FunctionSamples *FS,
133  ProfileSummaryInfo *PSI) const;
134  uint64_t getTotalUsedSamples() const { return TotalUsedSamples; }
135  uint64_t countBodySamples(const FunctionSamples *FS,
136  ProfileSummaryInfo *PSI) 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 /// 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  ProfileSummaryInfo *_PSI);
191 
192  void dump() { Reader->dump(); }
193 
194 protected:
196  unsigned getFunctionLoc(Function &F);
197  bool emitAnnotations(Function &F);
198  ErrorOr<uint64_t> getInstWeight(const Instruction &I);
199  ErrorOr<uint64_t> getBlockWeight(const BasicBlock *BB);
200  const FunctionSamples *findCalleeFunctionSamples(const Instruction &I) const;
201  std::vector<const FunctionSamples *>
202  findIndirectCallFunctionSamples(const Instruction &I, uint64_t &Sum) const;
203  const FunctionSamples *findFunctionSamples(const Instruction &I) const;
204  bool inlineCallInstruction(Instruction *I);
205  bool inlineHotFunctions(Function &F,
206  DenseSet<GlobalValue::GUID> &InlinedGUIDs);
207  void printEdgeWeight(raw_ostream &OS, Edge E);
208  void printBlockWeight(raw_ostream &OS, const BasicBlock *BB) const;
209  void printBlockEquivalence(raw_ostream &OS, const BasicBlock *BB);
210  bool computeBlockWeights(Function &F);
211  void findEquivalenceClasses(Function &F);
212  template <bool IsPostDom>
213  void findEquivalencesFor(BasicBlock *BB1, ArrayRef<BasicBlock *> Descendants,
215 
216  void propagateWeights(Function &F);
217  uint64_t visitEdge(Edge E, unsigned *NumUnknownEdges, Edge *UnknownEdge);
218  void buildEdges(Function &F);
219  bool propagateThroughEdges(Function &F, bool UpdateBlockCount);
220  void computeDominanceAndLoopInfo(Function &F);
221  void clearFunctionData();
222 
223  /// 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  /// 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  /// Set of visited blocks during propagation.
237 
238  /// Set of visited edges during propagation.
239  SmallSet<Edge, 32> VisitedEdges;
240 
241  /// 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  /// Dominance, post-dominance and loop information.
256  std::unique_ptr<DominatorTree> DT;
257  std::unique_ptr<PostDominatorTree> PDT;
258  std::unique_ptr<LoopInfo> LI;
259 
260  std::function<AssumptionCache &(Function &)> GetAC;
261  std::function<TargetTransformInfo &(Function &)> GetTTI;
262 
263  /// Predecessors for each basic block in the CFG.
264  BlockEdgeMap Predecessors;
265 
266  /// Successors for each basic block in the CFG.
267  BlockEdgeMap Successors;
268 
269  SampleCoverageTracker CoverageTracker;
270 
271  /// Profile reader object.
272  std::unique_ptr<SampleProfileReader> Reader;
273 
274  /// Samples collected for the body of this function.
275  FunctionSamples *Samples = nullptr;
276 
277  /// Name of the profile file to load.
278  std::string Filename;
279 
280  /// Flag indicating whether the profile input loaded successfully.
281  bool ProfileIsValid = false;
282 
283  /// 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  /// Profile Summary Info computed from sample profile.
290  ProfileSummaryInfo *PSI = nullptr;
291 
292  /// Total number of samples collected in this profile.
293  ///
294  /// This is the sum of all the samples collected in all the functions executed
295  /// at runtime.
296  uint64_t TotalCollectedSamples = 0;
297 
298  /// Optimization Remark Emitter used to emit diagnostic remarks.
299  OptimizationRemarkEmitter *ORE = nullptr;
300 };
301 
302 class SampleProfileLoaderLegacyPass : public ModulePass {
303 public:
304  // Class identification, replacement for typeinfo
305  static char ID;
306 
307  SampleProfileLoaderLegacyPass(StringRef Name = SampleProfileFile,
308  bool IsThinLTOPreLink = false)
309  : ModulePass(ID), SampleLoader(Name, IsThinLTOPreLink,
310  [&](Function &F) -> AssumptionCache & {
311  return ACT->getAssumptionCache(F);
312  },
313  [&](Function &F) -> TargetTransformInfo & {
314  return TTIWP->getTTI(F);
315  }) {
318  }
319 
320  void dump() { SampleLoader.dump(); }
321 
322  bool doInitialization(Module &M) override {
323  return SampleLoader.doInitialization(M);
324  }
325 
326  StringRef getPassName() const override { return "Sample profile pass"; }
327  bool runOnModule(Module &M) override;
328 
329  void getAnalysisUsage(AnalysisUsage &AU) const override {
333  }
334 
335 private:
336  SampleProfileLoader SampleLoader;
337  AssumptionCacheTracker *ACT = nullptr;
338  TargetTransformInfoWrapperPass *TTIWP = nullptr;
339 };
340 
341 } // end anonymous namespace
342 
343 /// Return true if the given callsite is hot wrt to hot cutoff threshold.
344 ///
345 /// Functions that were inlined in the original binary will be represented
346 /// in the inline stack in the sample profile. If the profile shows that
347 /// the original inline decision was "good" (i.e., the callsite is executed
348 /// frequently), then we will recreate the inline decision and apply the
349 /// profile from the inlined callsite.
350 ///
351 /// To decide whether an inlined callsite is hot, we compare the callsite
352 /// sample count with the hot cutoff computed by ProfileSummaryInfo, it is
353 /// regarded as hot if the count is above the cutoff value.
354 static bool callsiteIsHot(const FunctionSamples *CallsiteFS,
355  ProfileSummaryInfo *PSI) {
356  if (!CallsiteFS)
357  return false; // The callsite was not inlined in the original binary.
358 
359  assert(PSI && "PSI is expected to be non null");
360  uint64_t CallsiteTotalSamples = CallsiteFS->getTotalSamples();
361  return PSI->isHotCount(CallsiteTotalSamples);
362 }
363 
364 /// Mark as used the sample record for the given function samples at
365 /// (LineOffset, Discriminator).
366 ///
367 /// \returns true if this is the first time we mark the given record.
368 bool SampleCoverageTracker::markSamplesUsed(const FunctionSamples *FS,
369  uint32_t LineOffset,
370  uint32_t Discriminator,
371  uint64_t Samples) {
372  LineLocation Loc(LineOffset, Discriminator);
373  unsigned &Count = SampleCoverage[FS][Loc];
374  bool FirstTime = (++Count == 1);
375  if (FirstTime)
376  TotalUsedSamples += Samples;
377  return FirstTime;
378 }
379 
380 /// Return the number of sample records that were applied from this profile.
381 ///
382 /// This count does not include records from cold inlined callsites.
383 unsigned
384 SampleCoverageTracker::countUsedRecords(const FunctionSamples *FS,
385  ProfileSummaryInfo *PSI) const {
386  auto I = SampleCoverage.find(FS);
387 
388  // The size of the coverage map for FS represents the number of records
389  // that were marked used at least once.
390  unsigned Count = (I != SampleCoverage.end()) ? I->second.size() : 0;
391 
392  // If there are inlined callsites in this function, count the samples found
393  // in the respective bodies. However, do not bother counting callees with 0
394  // total samples, these are callees that were never invoked at runtime.
395  for (const auto &I : FS->getCallsiteSamples())
396  for (const auto &J : I.second) {
397  const FunctionSamples *CalleeSamples = &J.second;
398  if (callsiteIsHot(CalleeSamples, PSI))
399  Count += countUsedRecords(CalleeSamples, PSI);
400  }
401 
402  return Count;
403 }
404 
405 /// Return the number of sample records in the body of this profile.
406 ///
407 /// This count does not include records from cold inlined callsites.
408 unsigned
409 SampleCoverageTracker::countBodyRecords(const FunctionSamples *FS,
410  ProfileSummaryInfo *PSI) const {
411  unsigned Count = FS->getBodySamples().size();
412 
413  // Only count records in hot callsites.
414  for (const auto &I : FS->getCallsiteSamples())
415  for (const auto &J : I.second) {
416  const FunctionSamples *CalleeSamples = &J.second;
417  if (callsiteIsHot(CalleeSamples, PSI))
418  Count += countBodyRecords(CalleeSamples, PSI);
419  }
420 
421  return Count;
422 }
423 
424 /// Return the number of samples collected in the body of this profile.
425 ///
426 /// This count does not include samples from cold inlined callsites.
427 uint64_t
428 SampleCoverageTracker::countBodySamples(const FunctionSamples *FS,
429  ProfileSummaryInfo *PSI) const {
430  uint64_t Total = 0;
431  for (const auto &I : FS->getBodySamples())
432  Total += I.second.getSamples();
433 
434  // Only count samples in hot callsites.
435  for (const auto &I : FS->getCallsiteSamples())
436  for (const auto &J : I.second) {
437  const FunctionSamples *CalleeSamples = &J.second;
438  if (callsiteIsHot(CalleeSamples, PSI))
439  Total += countBodySamples(CalleeSamples, PSI);
440  }
441 
442  return Total;
443 }
444 
445 /// Return the fraction of sample records used in this profile.
446 ///
447 /// The returned value is an unsigned integer in the range 0-100 indicating
448 /// the percentage of sample records that were used while applying this
449 /// profile to the associated function.
450 unsigned SampleCoverageTracker::computeCoverage(unsigned Used,
451  unsigned Total) const {
452  assert(Used <= Total &&
453  "number of used records cannot exceed the total number of records");
454  return Total > 0 ? Used * 100 / Total : 100;
455 }
456 
457 /// Clear all the per-function data used to load samples and propagate weights.
458 void SampleProfileLoader::clearFunctionData() {
459  BlockWeights.clear();
460  EdgeWeights.clear();
461  VisitedBlocks.clear();
462  VisitedEdges.clear();
463  EquivalenceClass.clear();
464  DT = nullptr;
465  PDT = nullptr;
466  LI = nullptr;
467  Predecessors.clear();
468  Successors.clear();
469  CoverageTracker.clear();
470 }
471 
472 #ifndef NDEBUG
473 /// Print the weight of edge \p E on stream \p OS.
474 ///
475 /// \param OS Stream to emit the output to.
476 /// \param E Edge to print.
477 void SampleProfileLoader::printEdgeWeight(raw_ostream &OS, Edge E) {
478  OS << "weight[" << E.first->getName() << "->" << E.second->getName()
479  << "]: " << EdgeWeights[E] << "\n";
480 }
481 
482 /// Print the equivalence class of block \p BB on stream \p OS.
483 ///
484 /// \param OS Stream to emit the output to.
485 /// \param BB Block to print.
486 void SampleProfileLoader::printBlockEquivalence(raw_ostream &OS,
487  const BasicBlock *BB) {
488  const BasicBlock *Equiv = EquivalenceClass[BB];
489  OS << "equivalence[" << BB->getName()
490  << "]: " << ((Equiv) ? EquivalenceClass[BB]->getName() : "NONE") << "\n";
491 }
492 
493 /// Print the weight 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::printBlockWeight(raw_ostream &OS,
498  const BasicBlock *BB) const {
499  const auto &I = BlockWeights.find(BB);
500  uint64_t W = (I == BlockWeights.end() ? 0 : I->second);
501  OS << "weight[" << BB->getName() << "]: " << W << "\n";
502 }
503 #endif
504 
505 /// Get the weight for an instruction.
506 ///
507 /// The "weight" of an instruction \p Inst is the number of samples
508 /// collected on that instruction at runtime. To retrieve it, we
509 /// need to compute the line number of \p Inst relative to the start of its
510 /// function. We use HeaderLineno to compute the offset. We then
511 /// look up the samples collected for \p Inst using BodySamples.
512 ///
513 /// \param Inst Instruction to query.
514 ///
515 /// \returns the weight of \p Inst.
516 ErrorOr<uint64_t> SampleProfileLoader::getInstWeight(const Instruction &Inst) {
517  const DebugLoc &DLoc = Inst.getDebugLoc();
518  if (!DLoc)
519  return std::error_code();
520 
521  const FunctionSamples *FS = findFunctionSamples(Inst);
522  if (!FS)
523  return std::error_code();
524 
525  // Ignore all intrinsics and branch instructions.
526  // Branch instruction usually contains debug info from sources outside of
527  // the residing basic block, thus we ignore them during annotation.
528  if (isa<BranchInst>(Inst) || isa<IntrinsicInst>(Inst))
529  return std::error_code();
530 
531  // If a direct call/invoke instruction is inlined in profile
532  // (findCalleeFunctionSamples returns non-empty result), but not inlined here,
533  // it means that the inlined callsite has no sample, thus the call
534  // instruction should have 0 count.
535  if ((isa<CallInst>(Inst) || isa<InvokeInst>(Inst)) &&
536  !ImmutableCallSite(&Inst).isIndirectCall() &&
537  findCalleeFunctionSamples(Inst))
538  return 0;
539 
540  const DILocation *DIL = DLoc;
541  uint32_t LineOffset = FunctionSamples::getOffset(DIL);
542  uint32_t Discriminator = DIL->getBaseDiscriminator();
543  ErrorOr<uint64_t> R = FS->findSamplesAt(LineOffset, Discriminator);
544  if (R) {
545  bool FirstMark =
546  CoverageTracker.markSamplesUsed(FS, LineOffset, Discriminator, R.get());
547  if (FirstMark) {
548  ORE->emit([&]() {
549  OptimizationRemarkAnalysis Remark(DEBUG_TYPE, "AppliedSamples", &Inst);
550  Remark << "Applied " << ore::NV("NumSamples", *R);
551  Remark << " samples from profile (offset: ";
552  Remark << ore::NV("LineOffset", LineOffset);
553  if (Discriminator) {
554  Remark << ".";
555  Remark << ore::NV("Discriminator", Discriminator);
556  }
557  Remark << ")";
558  return Remark;
559  });
560  }
561  LLVM_DEBUG(dbgs() << " " << DLoc.getLine() << "."
562  << DIL->getBaseDiscriminator() << ":" << Inst
563  << " (line offset: " << LineOffset << "."
564  << DIL->getBaseDiscriminator() << " - weight: " << R.get()
565  << ")\n");
566  }
567  return R;
568 }
569 
570 /// Compute the weight of a basic block.
571 ///
572 /// The weight of basic block \p BB is the maximum weight of all the
573 /// instructions in BB.
574 ///
575 /// \param BB The basic block to query.
576 ///
577 /// \returns the weight for \p BB.
578 ErrorOr<uint64_t> SampleProfileLoader::getBlockWeight(const BasicBlock *BB) {
579  uint64_t Max = 0;
580  bool HasWeight = false;
581  for (auto &I : BB->getInstList()) {
582  const ErrorOr<uint64_t> &R = getInstWeight(I);
583  if (R) {
584  Max = std::max(Max, R.get());
585  HasWeight = true;
586  }
587  }
588  return HasWeight ? ErrorOr<uint64_t>(Max) : std::error_code();
589 }
590 
591 /// Compute and store the weights of every basic block.
592 ///
593 /// This populates the BlockWeights map by computing
594 /// the weights of every basic block in the CFG.
595 ///
596 /// \param F The function to query.
597 bool SampleProfileLoader::computeBlockWeights(Function &F) {
598  bool Changed = false;
599  LLVM_DEBUG(dbgs() << "Block weights\n");
600  for (const auto &BB : F) {
601  ErrorOr<uint64_t> Weight = getBlockWeight(&BB);
602  if (Weight) {
603  BlockWeights[&BB] = Weight.get();
604  VisitedBlocks.insert(&BB);
605  Changed = true;
606  }
607  LLVM_DEBUG(printBlockWeight(dbgs(), &BB));
608  }
609 
610  return Changed;
611 }
612 
613 /// Get the FunctionSamples for a call instruction.
614 ///
615 /// The FunctionSamples of a call/invoke instruction \p Inst is the inlined
616 /// instance in which that call instruction is calling to. It contains
617 /// all samples that resides in the inlined instance. We first find the
618 /// inlined instance in which the call instruction is from, then we
619 /// traverse its children to find the callsite with the matching
620 /// location.
621 ///
622 /// \param Inst Call/Invoke instruction to query.
623 ///
624 /// \returns The FunctionSamples pointer to the inlined instance.
625 const FunctionSamples *
626 SampleProfileLoader::findCalleeFunctionSamples(const Instruction &Inst) const {
627  const DILocation *DIL = Inst.getDebugLoc();
628  if (!DIL) {
629  return nullptr;
630  }
631 
632  StringRef CalleeName;
633  if (const CallInst *CI = dyn_cast<CallInst>(&Inst))
634  if (Function *Callee = CI->getCalledFunction())
635  CalleeName = Callee->getName();
636 
637  const FunctionSamples *FS = findFunctionSamples(Inst);
638  if (FS == nullptr)
639  return nullptr;
640 
641  std::string CalleeGUID;
642  CalleeName = getRepInFormat(CalleeName, Reader->getFormat(), CalleeGUID);
644  DIL->getBaseDiscriminator()),
645  CalleeName);
646 }
647 
648 /// Returns a vector of FunctionSamples that are the indirect call targets
649 /// of \p Inst. The vector is sorted by the total number of samples. Stores
650 /// the total call count of the indirect call in \p Sum.
651 std::vector<const FunctionSamples *>
652 SampleProfileLoader::findIndirectCallFunctionSamples(
653  const Instruction &Inst, uint64_t &Sum) const {
654  const DILocation *DIL = Inst.getDebugLoc();
655  std::vector<const FunctionSamples *> R;
656 
657  if (!DIL) {
658  return R;
659  }
660 
661  const FunctionSamples *FS = findFunctionSamples(Inst);
662  if (FS == nullptr)
663  return R;
664 
665  uint32_t LineOffset = FunctionSamples::getOffset(DIL);
666  uint32_t Discriminator = DIL->getBaseDiscriminator();
667 
668  auto T = FS->findCallTargetMapAt(LineOffset, Discriminator);
669  Sum = 0;
670  if (T)
671  for (const auto &T_C : T.get())
672  Sum += T_C.second;
675  if (M->empty())
676  return R;
677  for (const auto &NameFS : *M) {
678  Sum += NameFS.second.getEntrySamples();
679  R.push_back(&NameFS.second);
680  }
681  llvm::sort(R.begin(), R.end(),
682  [](const FunctionSamples *L, const FunctionSamples *R) {
683  return L->getEntrySamples() > R->getEntrySamples();
684  });
685  }
686  return R;
687 }
688 
689 /// Get the FunctionSamples for an instruction.
690 ///
691 /// The FunctionSamples of an instruction \p Inst is the inlined instance
692 /// in which that instruction is coming from. We traverse the inline stack
693 /// of that instruction, and match it with the tree nodes in the profile.
694 ///
695 /// \param Inst Instruction to query.
696 ///
697 /// \returns the FunctionSamples pointer to the inlined instance.
698 const FunctionSamples *
699 SampleProfileLoader::findFunctionSamples(const Instruction &Inst) const {
701  const DILocation *DIL = Inst.getDebugLoc();
702  if (!DIL)
703  return Samples;
704 
705  return Samples->findFunctionSamples(DIL);
706 }
707 
708 bool SampleProfileLoader::inlineCallInstruction(Instruction *I) {
709  assert(isa<CallInst>(I) || isa<InvokeInst>(I));
710  CallSite CS(I);
711  Function *CalledFunction = CS.getCalledFunction();
712  assert(CalledFunction);
713  DebugLoc DLoc = I->getDebugLoc();
714  BasicBlock *BB = I->getParent();
715  InlineParams Params = getInlineParams();
716  Params.ComputeFullInlineCost = true;
717  // Checks if there is anything in the reachable portion of the callee at
718  // this callsite that makes this inlining potentially illegal. Need to
719  // set ComputeFullInlineCost, otherwise getInlineCost may return early
720  // when cost exceeds threshold without checking all IRs in the callee.
721  // The acutal cost does not matter because we only checks isNever() to
722  // see if it is legal to inline the callsite.
723  InlineCost Cost = getInlineCost(CS, Params, GetTTI(*CalledFunction), GetAC,
724  None, nullptr, nullptr);
725  if (Cost.isNever()) {
726  ORE->emit(OptimizationRemark(DEBUG_TYPE, "Not inline", DLoc, BB)
727  << "incompatible inlining");
728  return false;
729  }
730  InlineFunctionInfo IFI(nullptr, &GetAC);
731  if (InlineFunction(CS, IFI)) {
732  // The call to InlineFunction erases I, so we can't pass it here.
733  ORE->emit(OptimizationRemark(DEBUG_TYPE, "HotInline", DLoc, BB)
734  << "inlined hot callee '" << ore::NV("Callee", CalledFunction)
735  << "' into '" << ore::NV("Caller", BB->getParent()) << "'");
736  return true;
737  }
738  return false;
739 }
740 
741 /// Iteratively inline hot callsites of a function.
742 ///
743 /// Iteratively traverse all callsites of the function \p F, and find if
744 /// the corresponding inlined instance exists and is hot in profile. If
745 /// it is hot enough, inline the callsites and adds new callsites of the
746 /// callee into the caller. If the call is an indirect call, first promote
747 /// it to direct call. Each indirect call is limited with a single target.
748 ///
749 /// \param F function to perform iterative inlining.
750 /// \param InlinedGUIDs a set to be updated to include all GUIDs that are
751 /// inlined in the profiled binary.
752 ///
753 /// \returns True if there is any inline happened.
754 bool SampleProfileLoader::inlineHotFunctions(
755  Function &F, DenseSet<GlobalValue::GUID> &InlinedGUIDs) {
756  DenseSet<Instruction *> PromotedInsns;
757  bool Changed = false;
758  bool isCompact = (Reader->getFormat() == SPF_Compact_Binary);
759  while (true) {
760  bool LocalChanged = false;
762  for (auto &BB : F) {
763  bool Hot = false;
765  for (auto &I : BB.getInstList()) {
766  const FunctionSamples *FS = nullptr;
767  if ((isa<CallInst>(I) || isa<InvokeInst>(I)) &&
768  !isa<IntrinsicInst>(I) && (FS = findCalleeFunctionSamples(I))) {
769  Candidates.push_back(&I);
770  if (callsiteIsHot(FS, PSI))
771  Hot = true;
772  }
773  }
774  if (Hot) {
775  CIS.insert(CIS.begin(), Candidates.begin(), Candidates.end());
776  }
777  }
778  for (auto I : CIS) {
779  Function *CalledFunction = CallSite(I).getCalledFunction();
780  // Do not inline recursive calls.
781  if (CalledFunction == &F)
782  continue;
783  if (CallSite(I).isIndirectCall()) {
784  if (PromotedInsns.count(I))
785  continue;
786  uint64_t Sum;
787  for (const auto *FS : findIndirectCallFunctionSamples(*I, Sum)) {
788  if (IsThinLTOPreLink) {
789  FS->findInlinedFunctions(InlinedGUIDs, F.getParent(),
791  isCompact);
792  continue;
793  }
794  auto CalleeFunctionName = FS->getName();
795  // If it is a recursive call, we do not inline it as it could bloat
796  // the code exponentially. There is way to better handle this, e.g.
797  // clone the caller first, and inline the cloned caller if it is
798  // recursive. As llvm does not inline recursive calls, we will
799  // simply ignore it instead of handling it explicitly.
800  std::string FGUID;
801  auto Fname = getRepInFormat(F.getName(), Reader->getFormat(), FGUID);
802  if (CalleeFunctionName == Fname)
803  continue;
804 
805  const char *Reason = "Callee function not available";
806  auto R = SymbolMap.find(CalleeFunctionName);
807  if (R != SymbolMap.end() && R->getValue() &&
808  !R->getValue()->isDeclaration() &&
809  R->getValue()->getSubprogram() &&
810  isLegalToPromote(CallSite(I), R->getValue(), &Reason)) {
811  uint64_t C = FS->getEntrySamples();
812  Instruction *DI =
813  pgo::promoteIndirectCall(I, R->getValue(), C, Sum, false, ORE);
814  Sum -= C;
815  PromotedInsns.insert(I);
816  // If profile mismatches, we should not attempt to inline DI.
817  if ((isa<CallInst>(DI) || isa<InvokeInst>(DI)) &&
818  inlineCallInstruction(DI))
819  LocalChanged = true;
820  } else {
821  LLVM_DEBUG(dbgs()
822  << "\nFailed to promote indirect call to "
823  << CalleeFunctionName << " because " << Reason << "\n");
824  }
825  }
826  } else if (CalledFunction && CalledFunction->getSubprogram() &&
827  !CalledFunction->isDeclaration()) {
828  if (inlineCallInstruction(I))
829  LocalChanged = true;
830  } else if (IsThinLTOPreLink) {
831  findCalleeFunctionSamples(*I)->findInlinedFunctions(
832  InlinedGUIDs, F.getParent(), PSI->getOrCompHotCountThreshold(),
833  isCompact);
834  }
835  }
836  if (LocalChanged) {
837  Changed = true;
838  } else {
839  break;
840  }
841  }
842  return Changed;
843 }
844 
845 /// Find equivalence classes for the given block.
846 ///
847 /// This finds all the blocks that are guaranteed to execute the same
848 /// number of times as \p BB1. To do this, it traverses all the
849 /// descendants of \p BB1 in the dominator or post-dominator tree.
850 ///
851 /// A block BB2 will be in the same equivalence class as \p BB1 if
852 /// the following holds:
853 ///
854 /// 1- \p BB1 is a descendant of BB2 in the opposite tree. So, if BB2
855 /// is a descendant of \p BB1 in the dominator tree, then BB2 should
856 /// dominate BB1 in the post-dominator tree.
857 ///
858 /// 2- Both BB2 and \p BB1 must be in the same loop.
859 ///
860 /// For every block BB2 that meets those two requirements, we set BB2's
861 /// equivalence class to \p BB1.
862 ///
863 /// \param BB1 Block to check.
864 /// \param Descendants Descendants of \p BB1 in either the dom or pdom tree.
865 /// \param DomTree Opposite dominator tree. If \p Descendants is filled
866 /// with blocks from \p BB1's dominator tree, then
867 /// this is the post-dominator tree, and vice versa.
868 template <bool IsPostDom>
869 void SampleProfileLoader::findEquivalencesFor(
870  BasicBlock *BB1, ArrayRef<BasicBlock *> Descendants,
872  const BasicBlock *EC = EquivalenceClass[BB1];
873  uint64_t Weight = BlockWeights[EC];
874  for (const auto *BB2 : Descendants) {
875  bool IsDomParent = DomTree->dominates(BB2, BB1);
876  bool IsInSameLoop = LI->getLoopFor(BB1) == LI->getLoopFor(BB2);
877  if (BB1 != BB2 && IsDomParent && IsInSameLoop) {
878  EquivalenceClass[BB2] = EC;
879  // If BB2 is visited, then the entire EC should be marked as visited.
880  if (VisitedBlocks.count(BB2)) {
881  VisitedBlocks.insert(EC);
882  }
883 
884  // If BB2 is heavier than BB1, make BB2 have the same weight
885  // as BB1.
886  //
887  // Note that we don't worry about the opposite situation here
888  // (when BB2 is lighter than BB1). We will deal with this
889  // during the propagation phase. Right now, we just want to
890  // make sure that BB1 has the largest weight of all the
891  // members of its equivalence set.
892  Weight = std::max(Weight, BlockWeights[BB2]);
893  }
894  }
895  if (EC == &EC->getParent()->getEntryBlock()) {
896  BlockWeights[EC] = Samples->getHeadSamples() + 1;
897  } else {
898  BlockWeights[EC] = Weight;
899  }
900 }
901 
902 /// Find equivalence classes.
903 ///
904 /// Since samples may be missing from blocks, we can fill in the gaps by setting
905 /// the weights of all the blocks in the same equivalence class to the same
906 /// weight. To compute the concept of equivalence, we use dominance and loop
907 /// information. Two blocks B1 and B2 are in the same equivalence class if B1
908 /// dominates B2, B2 post-dominates B1 and both are in the same loop.
909 ///
910 /// \param F The function to query.
911 void SampleProfileLoader::findEquivalenceClasses(Function &F) {
912  SmallVector<BasicBlock *, 8> DominatedBBs;
913  LLVM_DEBUG(dbgs() << "\nBlock equivalence classes\n");
914  // Find equivalence sets based on dominance and post-dominance information.
915  for (auto &BB : F) {
916  BasicBlock *BB1 = &BB;
917 
918  // Compute BB1's equivalence class once.
919  if (EquivalenceClass.count(BB1)) {
920  LLVM_DEBUG(printBlockEquivalence(dbgs(), BB1));
921  continue;
922  }
923 
924  // By default, blocks are in their own equivalence class.
925  EquivalenceClass[BB1] = BB1;
926 
927  // Traverse all the blocks dominated by BB1. We are looking for
928  // every basic block BB2 such that:
929  //
930  // 1- BB1 dominates BB2.
931  // 2- BB2 post-dominates BB1.
932  // 3- BB1 and BB2 are in the same loop nest.
933  //
934  // If all those conditions hold, it means that BB2 is executed
935  // as many times as BB1, so they are placed in the same equivalence
936  // class by making BB2's equivalence class be BB1.
937  DominatedBBs.clear();
938  DT->getDescendants(BB1, DominatedBBs);
939  findEquivalencesFor(BB1, DominatedBBs, PDT.get());
940 
941  LLVM_DEBUG(printBlockEquivalence(dbgs(), BB1));
942  }
943 
944  // Assign weights to equivalence classes.
945  //
946  // All the basic blocks in the same equivalence class will execute
947  // the same number of times. Since we know that the head block in
948  // each equivalence class has the largest weight, assign that weight
949  // to all the blocks in that equivalence class.
950  LLVM_DEBUG(
951  dbgs() << "\nAssign the same weight to all blocks in the same class\n");
952  for (auto &BI : F) {
953  const BasicBlock *BB = &BI;
954  const BasicBlock *EquivBB = EquivalenceClass[BB];
955  if (BB != EquivBB)
956  BlockWeights[BB] = BlockWeights[EquivBB];
957  LLVM_DEBUG(printBlockWeight(dbgs(), BB));
958  }
959 }
960 
961 /// Visit the given edge to decide if it has a valid weight.
962 ///
963 /// If \p E has not been visited before, we copy to \p UnknownEdge
964 /// and increment the count of unknown edges.
965 ///
966 /// \param E Edge to visit.
967 /// \param NumUnknownEdges Current number of unknown edges.
968 /// \param UnknownEdge Set if E has not been visited before.
969 ///
970 /// \returns E's weight, if known. Otherwise, return 0.
971 uint64_t SampleProfileLoader::visitEdge(Edge E, unsigned *NumUnknownEdges,
972  Edge *UnknownEdge) {
973  if (!VisitedEdges.count(E)) {
974  (*NumUnknownEdges)++;
975  *UnknownEdge = E;
976  return 0;
977  }
978 
979  return EdgeWeights[E];
980 }
981 
982 /// Propagate weights through incoming/outgoing edges.
983 ///
984 /// If the weight of a basic block is known, and there is only one edge
985 /// with an unknown weight, we can calculate the weight of that edge.
986 ///
987 /// Similarly, if all the edges have a known count, we can calculate the
988 /// count of the basic block, if needed.
989 ///
990 /// \param F Function to process.
991 /// \param UpdateBlockCount Whether we should update basic block counts that
992 /// has already been annotated.
993 ///
994 /// \returns True if new weights were assigned to edges or blocks.
995 bool SampleProfileLoader::propagateThroughEdges(Function &F,
996  bool UpdateBlockCount) {
997  bool Changed = false;
998  LLVM_DEBUG(dbgs() << "\nPropagation through edges\n");
999  for (const auto &BI : F) {
1000  const BasicBlock *BB = &BI;
1001  const BasicBlock *EC = EquivalenceClass[BB];
1002 
1003  // Visit all the predecessor and successor edges to determine
1004  // which ones have a weight assigned already. Note that it doesn't
1005  // matter that we only keep track of a single unknown edge. The
1006  // only case we are interested in handling is when only a single
1007  // edge is unknown (see setEdgeOrBlockWeight).
1008  for (unsigned i = 0; i < 2; i++) {
1009  uint64_t TotalWeight = 0;
1010  unsigned NumUnknownEdges = 0, NumTotalEdges = 0;
1011  Edge UnknownEdge, SelfReferentialEdge, SingleEdge;
1012 
1013  if (i == 0) {
1014  // First, visit all predecessor edges.
1015  NumTotalEdges = Predecessors[BB].size();
1016  for (auto *Pred : Predecessors[BB]) {
1017  Edge E = std::make_pair(Pred, BB);
1018  TotalWeight += visitEdge(E, &NumUnknownEdges, &UnknownEdge);
1019  if (E.first == E.second)
1020  SelfReferentialEdge = E;
1021  }
1022  if (NumTotalEdges == 1) {
1023  SingleEdge = std::make_pair(Predecessors[BB][0], BB);
1024  }
1025  } else {
1026  // On the second round, visit all successor edges.
1027  NumTotalEdges = Successors[BB].size();
1028  for (auto *Succ : Successors[BB]) {
1029  Edge E = std::make_pair(BB, Succ);
1030  TotalWeight += visitEdge(E, &NumUnknownEdges, &UnknownEdge);
1031  }
1032  if (NumTotalEdges == 1) {
1033  SingleEdge = std::make_pair(BB, Successors[BB][0]);
1034  }
1035  }
1036 
1037  // After visiting all the edges, there are three cases that we
1038  // can handle immediately:
1039  //
1040  // - All the edge weights are known (i.e., NumUnknownEdges == 0).
1041  // In this case, we simply check that the sum of all the edges
1042  // is the same as BB's weight. If not, we change BB's weight
1043  // to match. Additionally, if BB had not been visited before,
1044  // we mark it visited.
1045  //
1046  // - Only one edge is unknown and BB has already been visited.
1047  // In this case, we can compute the weight of the edge by
1048  // subtracting the total block weight from all the known
1049  // edge weights. If the edges weight more than BB, then the
1050  // edge of the last remaining edge is set to zero.
1051  //
1052  // - There exists a self-referential edge and the weight of BB is
1053  // known. In this case, this edge can be based on BB's weight.
1054  // We add up all the other known edges and set the weight on
1055  // the self-referential edge as we did in the previous case.
1056  //
1057  // In any other case, we must continue iterating. Eventually,
1058  // all edges will get a weight, or iteration will stop when
1059  // it reaches SampleProfileMaxPropagateIterations.
1060  if (NumUnknownEdges <= 1) {
1061  uint64_t &BBWeight = BlockWeights[EC];
1062  if (NumUnknownEdges == 0) {
1063  if (!VisitedBlocks.count(EC)) {
1064  // If we already know the weight of all edges, the weight of the
1065  // basic block can be computed. It should be no larger than the sum
1066  // of all edge weights.
1067  if (TotalWeight > BBWeight) {
1068  BBWeight = TotalWeight;
1069  Changed = true;
1070  LLVM_DEBUG(dbgs() << "All edge weights for " << BB->getName()
1071  << " known. Set weight for block: ";
1072  printBlockWeight(dbgs(), BB););
1073  }
1074  } else if (NumTotalEdges == 1 &&
1075  EdgeWeights[SingleEdge] < BlockWeights[EC]) {
1076  // If there is only one edge for the visited basic block, use the
1077  // block weight to adjust edge weight if edge weight is smaller.
1078  EdgeWeights[SingleEdge] = BlockWeights[EC];
1079  Changed = true;
1080  }
1081  } else if (NumUnknownEdges == 1 && VisitedBlocks.count(EC)) {
1082  // If there is a single unknown edge and the block has been
1083  // visited, then we can compute E's weight.
1084  if (BBWeight >= TotalWeight)
1085  EdgeWeights[UnknownEdge] = BBWeight - TotalWeight;
1086  else
1087  EdgeWeights[UnknownEdge] = 0;
1088  const BasicBlock *OtherEC;
1089  if (i == 0)
1090  OtherEC = EquivalenceClass[UnknownEdge.first];
1091  else
1092  OtherEC = EquivalenceClass[UnknownEdge.second];
1093  // Edge weights should never exceed the BB weights it connects.
1094  if (VisitedBlocks.count(OtherEC) &&
1095  EdgeWeights[UnknownEdge] > BlockWeights[OtherEC])
1096  EdgeWeights[UnknownEdge] = BlockWeights[OtherEC];
1097  VisitedEdges.insert(UnknownEdge);
1098  Changed = true;
1099  LLVM_DEBUG(dbgs() << "Set weight for edge: ";
1100  printEdgeWeight(dbgs(), UnknownEdge));
1101  }
1102  } else if (VisitedBlocks.count(EC) && BlockWeights[EC] == 0) {
1103  // If a block Weights 0, all its in/out edges should weight 0.
1104  if (i == 0) {
1105  for (auto *Pred : Predecessors[BB]) {
1106  Edge E = std::make_pair(Pred, BB);
1107  EdgeWeights[E] = 0;
1108  VisitedEdges.insert(E);
1109  }
1110  } else {
1111  for (auto *Succ : Successors[BB]) {
1112  Edge E = std::make_pair(BB, Succ);
1113  EdgeWeights[E] = 0;
1114  VisitedEdges.insert(E);
1115  }
1116  }
1117  } else if (SelfReferentialEdge.first && VisitedBlocks.count(EC)) {
1118  uint64_t &BBWeight = BlockWeights[BB];
1119  // We have a self-referential edge and the weight of BB is known.
1120  if (BBWeight >= TotalWeight)
1121  EdgeWeights[SelfReferentialEdge] = BBWeight - TotalWeight;
1122  else
1123  EdgeWeights[SelfReferentialEdge] = 0;
1124  VisitedEdges.insert(SelfReferentialEdge);
1125  Changed = true;
1126  LLVM_DEBUG(dbgs() << "Set self-referential edge weight to: ";
1127  printEdgeWeight(dbgs(), SelfReferentialEdge));
1128  }
1129  if (UpdateBlockCount && !VisitedBlocks.count(EC) && TotalWeight > 0) {
1130  BlockWeights[EC] = TotalWeight;
1131  VisitedBlocks.insert(EC);
1132  Changed = true;
1133  }
1134  }
1135  }
1136 
1137  return Changed;
1138 }
1139 
1140 /// Build in/out edge lists for each basic block in the CFG.
1141 ///
1142 /// We are interested in unique edges. If a block B1 has multiple
1143 /// edges to another block B2, we only add a single B1->B2 edge.
1144 void SampleProfileLoader::buildEdges(Function &F) {
1145  for (auto &BI : F) {
1146  BasicBlock *B1 = &BI;
1147 
1148  // Add predecessors for B1.
1150  if (!Predecessors[B1].empty())
1151  llvm_unreachable("Found a stale predecessors list in a basic block.");
1152  for (pred_iterator PI = pred_begin(B1), PE = pred_end(B1); PI != PE; ++PI) {
1153  BasicBlock *B2 = *PI;
1154  if (Visited.insert(B2).second)
1155  Predecessors[B1].push_back(B2);
1156  }
1157 
1158  // Add successors for B1.
1159  Visited.clear();
1160  if (!Successors[B1].empty())
1161  llvm_unreachable("Found a stale successors list in a basic block.");
1162  for (succ_iterator SI = succ_begin(B1), SE = succ_end(B1); SI != SE; ++SI) {
1163  BasicBlock *B2 = *SI;
1164  if (Visited.insert(B2).second)
1165  Successors[B1].push_back(B2);
1166  }
1167  }
1168 }
1169 
1170 /// Returns the sorted CallTargetMap \p M by count in descending order.
1172  const SampleRecord::CallTargetMap &M) {
1174  for (auto I = M.begin(); I != M.end(); ++I)
1175  R.push_back({Function::getGUID(I->getKey()), I->getValue()});
1176  llvm::sort(R.begin(), R.end(),
1177  [](const InstrProfValueData &L, const InstrProfValueData &R) {
1178  if (L.Count == R.Count)
1179  return L.Value > R.Value;
1180  else
1181  return L.Count > R.Count;
1182  });
1183  return R;
1184 }
1185 
1186 /// Propagate weights into edges
1187 ///
1188 /// The following rules are applied to every block BB in the CFG:
1189 ///
1190 /// - If BB has a single predecessor/successor, then the weight
1191 /// of that edge is the weight of the block.
1192 ///
1193 /// - If all incoming or outgoing edges are known except one, and the
1194 /// weight of the block is already known, the weight of the unknown
1195 /// edge will be the weight of the block minus the sum of all the known
1196 /// edges. If the sum of all the known edges is larger than BB's weight,
1197 /// we set the unknown edge weight to zero.
1198 ///
1199 /// - If there is a self-referential edge, and the weight of the block is
1200 /// known, the weight for that edge is set to the weight of the block
1201 /// minus the weight of the other incoming edges to that block (if
1202 /// known).
1203 void SampleProfileLoader::propagateWeights(Function &F) {
1204  bool Changed = true;
1205  unsigned I = 0;
1206 
1207  // If BB weight is larger than its corresponding loop's header BB weight,
1208  // use the BB weight to replace the loop header BB weight.
1209  for (auto &BI : F) {
1210  BasicBlock *BB = &BI;
1211  Loop *L = LI->getLoopFor(BB);
1212  if (!L) {
1213  continue;
1214  }
1215  BasicBlock *Header = L->getHeader();
1216  if (Header && BlockWeights[BB] > BlockWeights[Header]) {
1217  BlockWeights[Header] = BlockWeights[BB];
1218  }
1219  }
1220 
1221  // Before propagation starts, build, for each block, a list of
1222  // unique predecessors and successors. This is necessary to handle
1223  // identical edges in multiway branches. Since we visit all blocks and all
1224  // edges of the CFG, it is cleaner to build these lists once at the start
1225  // of the pass.
1226  buildEdges(F);
1227 
1228  // Propagate until we converge or we go past the iteration limit.
1229  while (Changed && I++ < SampleProfileMaxPropagateIterations) {
1230  Changed = propagateThroughEdges(F, false);
1231  }
1232 
1233  // The first propagation propagates BB counts from annotated BBs to unknown
1234  // BBs. The 2nd propagation pass resets edges weights, and use all BB weights
1235  // to propagate edge weights.
1236  VisitedEdges.clear();
1237  Changed = true;
1238  while (Changed && I++ < SampleProfileMaxPropagateIterations) {
1239  Changed = propagateThroughEdges(F, false);
1240  }
1241 
1242  // The 3rd propagation pass allows adjust annotated BB weights that are
1243  // obviously wrong.
1244  Changed = true;
1245  while (Changed && I++ < SampleProfileMaxPropagateIterations) {
1246  Changed = propagateThroughEdges(F, true);
1247  }
1248 
1249  // Generate MD_prof metadata for every branch instruction using the
1250  // edge weights computed during propagation.
1251  LLVM_DEBUG(dbgs() << "\nPropagation complete. Setting branch weights\n");
1252  LLVMContext &Ctx = F.getContext();
1253  MDBuilder MDB(Ctx);
1254  for (auto &BI : F) {
1255  BasicBlock *BB = &BI;
1256 
1257  if (BlockWeights[BB]) {
1258  for (auto &I : BB->getInstList()) {
1259  if (!isa<CallInst>(I) && !isa<InvokeInst>(I))
1260  continue;
1261  CallSite CS(&I);
1262  if (!CS.getCalledFunction()) {
1263  const DebugLoc &DLoc = I.getDebugLoc();
1264  if (!DLoc)
1265  continue;
1266  const DILocation *DIL = DLoc;
1267  uint32_t LineOffset = FunctionSamples::getOffset(DIL);
1268  uint32_t Discriminator = DIL->getBaseDiscriminator();
1269 
1270  const FunctionSamples *FS = findFunctionSamples(I);
1271  if (!FS)
1272  continue;
1273  auto T = FS->findCallTargetMapAt(LineOffset, Discriminator);
1274  if (!T || T.get().empty())
1275  continue;
1276  SmallVector<InstrProfValueData, 2> SortedCallTargets =
1277  SortCallTargets(T.get());
1278  uint64_t Sum;
1279  findIndirectCallFunctionSamples(I, Sum);
1280  annotateValueSite(*I.getParent()->getParent()->getParent(), I,
1281  SortedCallTargets, Sum, IPVK_IndirectCallTarget,
1282  SortedCallTargets.size());
1283  } else if (!dyn_cast<IntrinsicInst>(&I)) {
1284  SmallVector<uint32_t, 1> Weights;
1285  Weights.push_back(BlockWeights[BB]);
1286  I.setMetadata(LLVMContext::MD_prof, MDB.createBranchWeights(Weights));
1287  }
1288  }
1289  }
1290  TerminatorInst *TI = BB->getTerminator();
1291  if (TI->getNumSuccessors() == 1)
1292  continue;
1293  if (!isa<BranchInst>(TI) && !isa<SwitchInst>(TI))
1294  continue;
1295 
1296  DebugLoc BranchLoc = TI->getDebugLoc();
1297  LLVM_DEBUG(dbgs() << "\nGetting weights for branch at line "
1298  << ((BranchLoc) ? Twine(BranchLoc.getLine())
1299  : Twine("<UNKNOWN LOCATION>"))
1300  << ".\n");
1301  SmallVector<uint32_t, 4> Weights;
1302  uint32_t MaxWeight = 0;
1303  Instruction *MaxDestInst;
1304  for (unsigned I = 0; I < TI->getNumSuccessors(); ++I) {
1305  BasicBlock *Succ = TI->getSuccessor(I);
1306  Edge E = std::make_pair(BB, Succ);
1307  uint64_t Weight = EdgeWeights[E];
1308  LLVM_DEBUG(dbgs() << "\t"; printEdgeWeight(dbgs(), E));
1309  // Use uint32_t saturated arithmetic to adjust the incoming weights,
1310  // if needed. Sample counts in profiles are 64-bit unsigned values,
1311  // but internally branch weights are expressed as 32-bit values.
1312  if (Weight > std::numeric_limits<uint32_t>::max()) {
1313  LLVM_DEBUG(dbgs() << " (saturated due to uint32_t overflow)");
1315  }
1316  // Weight is added by one to avoid propagation errors introduced by
1317  // 0 weights.
1318  Weights.push_back(static_cast<uint32_t>(Weight + 1));
1319  if (Weight != 0) {
1320  if (Weight > MaxWeight) {
1321  MaxWeight = Weight;
1322  MaxDestInst = Succ->getFirstNonPHIOrDbgOrLifetime();
1323  }
1324  }
1325  }
1326 
1327  uint64_t TempWeight;
1328  // Only set weights if there is at least one non-zero weight.
1329  // In any other case, let the analyzer set weights.
1330  // Do not set weights if the weights are present. In ThinLTO, the profile
1331  // annotation is done twice. If the first annotation already set the
1332  // weights, the second pass does not need to set it.
1333  if (MaxWeight > 0 && !TI->extractProfTotalWeight(TempWeight)) {
1334  LLVM_DEBUG(dbgs() << "SUCCESS. Found non-zero weights.\n");
1336  MDB.createBranchWeights(Weights));
1337  ORE->emit([&]() {
1338  return OptimizationRemark(DEBUG_TYPE, "PopularDest", MaxDestInst)
1339  << "most popular destination for conditional branches at "
1340  << ore::NV("CondBranchesLoc", BranchLoc);
1341  });
1342  } else {
1343  LLVM_DEBUG(dbgs() << "SKIPPED. All branch weights are zero.\n");
1344  }
1345  }
1346 }
1347 
1348 /// Get the line number for the function header.
1349 ///
1350 /// This looks up function \p F in the current compilation unit and
1351 /// retrieves the line number where the function is defined. This is
1352 /// line 0 for all the samples read from the profile file. Every line
1353 /// number is relative to this line.
1354 ///
1355 /// \param F Function object to query.
1356 ///
1357 /// \returns the line number where \p F is defined. If it returns 0,
1358 /// it means that there is no debug information available for \p F.
1359 unsigned SampleProfileLoader::getFunctionLoc(Function &F) {
1360  if (DISubprogram *S = F.getSubprogram())
1361  return S->getLine();
1362 
1363  // If the start of \p F is missing, emit a diagnostic to inform the user
1364  // about the missed opportunity.
1366  "No debug information found in function " + F.getName() +
1367  ": Function profile not used",
1368  DS_Warning));
1369  return 0;
1370 }
1371 
1372 void SampleProfileLoader::computeDominanceAndLoopInfo(Function &F) {
1373  DT.reset(new DominatorTree);
1374  DT->recalculate(F);
1375 
1376  PDT.reset(new PostDominatorTree(F));
1377 
1378  LI.reset(new LoopInfo);
1379  LI->analyze(*DT);
1380 }
1381 
1382 /// Generate branch weight metadata for all branches in \p F.
1383 ///
1384 /// Branch weights are computed out of instruction samples using a
1385 /// propagation heuristic. Propagation proceeds in 3 phases:
1386 ///
1387 /// 1- Assignment of block weights. All the basic blocks in the function
1388 /// are initial assigned the same weight as their most frequently
1389 /// executed instruction.
1390 ///
1391 /// 2- Creation of equivalence classes. Since samples may be missing from
1392 /// blocks, we can fill in the gaps by setting the weights of all the
1393 /// blocks in the same equivalence class to the same weight. To compute
1394 /// the concept of equivalence, we use dominance and loop information.
1395 /// Two blocks B1 and B2 are in the same equivalence class if B1
1396 /// dominates B2, B2 post-dominates B1 and both are in the same loop.
1397 ///
1398 /// 3- Propagation of block weights into edges. This uses a simple
1399 /// propagation heuristic. The following rules are applied to every
1400 /// block BB in the CFG:
1401 ///
1402 /// - If BB has a single predecessor/successor, then the weight
1403 /// of that edge is the weight of the block.
1404 ///
1405 /// - If all the edges are known except one, and the weight of the
1406 /// block is already known, the weight of the unknown edge will
1407 /// be the weight of the block minus the sum of all the known
1408 /// edges. If the sum of all the known edges is larger than BB's weight,
1409 /// we set the unknown edge weight to zero.
1410 ///
1411 /// - If there is a self-referential edge, and the weight of the block is
1412 /// known, the weight for that edge is set to the weight of the block
1413 /// minus the weight of the other incoming edges to that block (if
1414 /// known).
1415 ///
1416 /// Since this propagation is not guaranteed to finalize for every CFG, we
1417 /// only allow it to proceed for a limited number of iterations (controlled
1418 /// by -sample-profile-max-propagate-iterations).
1419 ///
1420 /// FIXME: Try to replace this propagation heuristic with a scheme
1421 /// that is guaranteed to finalize. A work-list approach similar to
1422 /// the standard value propagation algorithm used by SSA-CCP might
1423 /// work here.
1424 ///
1425 /// Once all the branch weights are computed, we emit the MD_prof
1426 /// metadata on BB using the computed values for each of its branches.
1427 ///
1428 /// \param F The function to query.
1429 ///
1430 /// \returns true if \p F was modified. Returns false, otherwise.
1431 bool SampleProfileLoader::emitAnnotations(Function &F) {
1432  bool Changed = false;
1433 
1434  if (getFunctionLoc(F) == 0)
1435  return false;
1436 
1437  LLVM_DEBUG(dbgs() << "Line number for the first instruction in "
1438  << F.getName() << ": " << getFunctionLoc(F) << "\n");
1439 
1440  DenseSet<GlobalValue::GUID> InlinedGUIDs;
1441  Changed |= inlineHotFunctions(F, InlinedGUIDs);
1442 
1443  // Compute basic block weights.
1444  Changed |= computeBlockWeights(F);
1445 
1446  if (Changed) {
1447  // Add an entry count to the function using the samples gathered at the
1448  // function entry.
1449  // Sets the GUIDs that are inlined in the profiled binary. This is used
1450  // for ThinLink to make correct liveness analysis, and also make the IR
1451  // match the profiled binary before annotation.
1452  F.setEntryCount(
1453  ProfileCount(Samples->getHeadSamples() + 1, Function::PCT_Real),
1454  &InlinedGUIDs);
1455 
1456  // Compute dominance and loop info needed for propagation.
1457  computeDominanceAndLoopInfo(F);
1458 
1459  // Find equivalence classes.
1460  findEquivalenceClasses(F);
1461 
1462  // Propagate weights to all edges.
1463  propagateWeights(F);
1464  }
1465 
1466  // If coverage checking was requested, compute it now.
1468  unsigned Used = CoverageTracker.countUsedRecords(Samples, PSI);
1469  unsigned Total = CoverageTracker.countBodyRecords(Samples, PSI);
1470  unsigned Coverage = CoverageTracker.computeCoverage(Used, Total);
1471  if (Coverage < SampleProfileRecordCoverage) {
1473  F.getSubprogram()->getFilename(), getFunctionLoc(F),
1474  Twine(Used) + " of " + Twine(Total) + " available profile records (" +
1475  Twine(Coverage) + "%) were applied",
1476  DS_Warning));
1477  }
1478  }
1479 
1481  uint64_t Used = CoverageTracker.getTotalUsedSamples();
1482  uint64_t Total = CoverageTracker.countBodySamples(Samples, PSI);
1483  unsigned Coverage = CoverageTracker.computeCoverage(Used, Total);
1484  if (Coverage < SampleProfileSampleCoverage) {
1486  F.getSubprogram()->getFilename(), getFunctionLoc(F),
1487  Twine(Used) + " of " + Twine(Total) + " available profile samples (" +
1488  Twine(Coverage) + "%) were applied",
1489  DS_Warning));
1490  }
1491  }
1492  return Changed;
1493 }
1494 
1496 
1497 INITIALIZE_PASS_BEGIN(SampleProfileLoaderLegacyPass, "sample-profile",
1498  "Sample Profile loader", false, false)
1502 INITIALIZE_PASS_END(SampleProfileLoaderLegacyPass, "sample-profile",
1503  "Sample Profile loader", false, false)
1504 
1505 bool SampleProfileLoader::doInitialization(Module &M) {
1506  auto &Ctx = M.getContext();
1507  auto ReaderOrErr = SampleProfileReader::create(Filename, Ctx);
1508  if (std::error_code EC = ReaderOrErr.getError()) {
1509  std::string Msg = "Could not open profile: " + EC.message();
1510  Ctx.diagnose(DiagnosticInfoSampleProfile(Filename, Msg));
1511  return false;
1512  }
1513  Reader = std::move(ReaderOrErr.get());
1514  ProfileIsValid = (Reader->read() == sampleprof_error::success);
1515  return true;
1516 }
1517 
1519  return new SampleProfileLoaderLegacyPass(SampleProfileFile);
1520 }
1521 
1523  return new SampleProfileLoaderLegacyPass(Name);
1524 }
1525 
1526 bool SampleProfileLoader::runOnModule(Module &M, ModuleAnalysisManager *AM,
1527  ProfileSummaryInfo *_PSI) {
1528  if (!ProfileIsValid)
1529  return false;
1530 
1531  PSI = _PSI;
1532  if (M.getProfileSummary() == nullptr)
1533  M.setProfileSummary(Reader->getSummary().getMD(M.getContext()));
1534 
1535  // Compute the total number of samples collected in this profile.
1536  for (const auto &I : Reader->getProfiles())
1537  TotalCollectedSamples += I.second.getTotalSamples();
1538 
1539  // Populate the symbol map.
1540  for (const auto &N_F : M.getValueSymbolTable()) {
1541  StringRef OrigName = N_F.getKey();
1542  Function *F = dyn_cast<Function>(N_F.getValue());
1543  if (F == nullptr)
1544  continue;
1545  SymbolMap[OrigName] = F;
1546  auto pos = OrigName.find('.');
1547  if (pos != StringRef::npos) {
1548  StringRef NewName = OrigName.substr(0, pos);
1549  auto r = SymbolMap.insert(std::make_pair(NewName, F));
1550  // Failiing to insert means there is already an entry in SymbolMap,
1551  // thus there are multiple functions that are mapped to the same
1552  // stripped name. In this case of name conflicting, set the value
1553  // to nullptr to avoid confusion.
1554  if (!r.second)
1555  r.first->second = nullptr;
1556  }
1557  }
1558 
1559  bool retval = false;
1560  for (auto &F : M)
1561  if (!F.isDeclaration()) {
1562  clearFunctionData();
1563  retval |= runOnFunction(F, AM);
1564  }
1565  return retval;
1566 }
1567 
1568 bool SampleProfileLoaderLegacyPass::runOnModule(Module &M) {
1569  ACT = &getAnalysis<AssumptionCacheTracker>();
1570  TTIWP = &getAnalysis<TargetTransformInfoWrapperPass>();
1571  ProfileSummaryInfo *PSI =
1572  getAnalysis<ProfileSummaryInfoWrapperPass>().getPSI();
1573  return SampleLoader.runOnModule(M, nullptr, PSI);
1574 }
1575 
1577  // Initialize the entry count to -1, which will be treated conservatively
1578  // by getEntryCount as the same as unknown (None). If we have samples this
1579  // will be overwritten in emitAnnotations.
1581  std::unique_ptr<OptimizationRemarkEmitter> OwnedORE;
1582  if (AM) {
1583  auto &FAM =
1585  .getManager();
1586  ORE = &FAM.getResult<OptimizationRemarkEmitterAnalysis>(F);
1587  } else {
1588  OwnedORE = make_unique<OptimizationRemarkEmitter>(&F);
1589  ORE = OwnedORE.get();
1590  }
1591  Samples = Reader->getSamplesFor(F);
1592  if (Samples && !Samples->empty())
1593  return emitAnnotations(F);
1594  return false;
1595 }
1596 
1598  ModuleAnalysisManager &AM) {
1600  AM.getResult<FunctionAnalysisManagerModuleProxy>(M).getManager();
1601 
1602  auto GetAssumptionCache = [&](Function &F) -> AssumptionCache & {
1603  return FAM.getResult<AssumptionAnalysis>(F);
1604  };
1605  auto GetTTI = [&](Function &F) -> TargetTransformInfo & {
1606  return FAM.getResult<TargetIRAnalysis>(F);
1607  };
1608 
1609  SampleProfileLoader SampleLoader(
1610  ProfileFileName.empty() ? SampleProfileFile : ProfileFileName,
1611  IsThinLTOPreLink, GetAssumptionCache, GetTTI);
1612 
1613  SampleLoader.doInitialization(M);
1614 
1616  if (!SampleLoader.runOnModule(M, &AM, PSI))
1617  return PreservedAnalyses::all();
1618 
1619  return PreservedAnalyses::none();
1620 }
const NoneType None
Definition: None.h:24
uint64_t CallInst * C
const FunctionSamplesMap * findFunctionSamplesMapAt(const LineLocation &Loc) const
Returns the FunctionSamplesMap at the given Loc.
Definition: SampleProf.h:283
Thresholds to tune inline cost analysis.
Definition: InlineCost.h: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
static StringRef getRepInFormat(StringRef Name, SampleProfileFormat Format, std::string &GUIDBuf)
Definition: SampleProf.h:97
A Module instance is used to store all the information related to an LLVM module. ...
Definition: Module.h:63
BasicBlock * getSuccessor(unsigned idx) const
Return the specified successor.
LLVM_ATTRIBUTE_ALWAYS_INLINE size_type size() const
Definition: SmallVector.h:137
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:547
Analysis providing profile information.
This class represents a function call, abstracting a target machine&#39;s calling convention.
An immutable pass that tracks lazily created AssumptionCache objects.
A cache of @llvm.assume calls within a function.
Analysis pass providing the TargetTransformInfo.
bool isLegalToPromote(CallSite CS, Function *Callee, const char **FailureReason=nullptr)
Return true if the given indirect call site can be made to call Callee.
static ErrorOr< std::unique_ptr< SampleProfileReader > > create(const Twine &Filename, LLVMContext &C)
Create a sample profile reader appropriate to the file format.
unsigned getLine() const
Definition: DebugLoc.cpp:26
ErrorOr< uint64_t > findSamplesAt(uint32_t LineOffset, uint32_t Discriminator) const
Return the number of samples collected at the given location.
Definition: SampleProf.h:256
std::map< SymbolStringPtr, JITEvaluatedSymbol > SymbolMap
A map from symbol names (as SymbolStringPtrs) to JITSymbols (address/flags pairs).
Definition: Core.h:51
A debug info location.
Definition: DebugLoc.h:34
F(f)
uint64_t getOrCompHotCountThreshold()
Returns HotCountThreshold if set.
const StringRef & getName() const
Return the function name.
Definition: SampleProf.h:403
static SmallVector< InstrProfValueData, 2 > SortCallTargets(const SampleRecord::CallTargetMap &M)
Returns the sorted CallTargetMap M by count in descending order.
InlineFunctionInfo - This class captures the data input to the InlineFunction call, and records the auxiliary results produced by it.
Definition: Cloning.h:176
bool isHotCount(uint64_t C)
Returns true if F is a hot function.
Represents the cost of inlining a function.
Definition: InlineCost.h:65
PreservedAnalyses run(Module &M, ModuleAnalysisManager &AM)
Representation of the samples collected for a function.
Definition: SampleProf.h:216
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:921
Twine - A lightweight data structure for efficiently representing the concatenation of temporary valu...
Definition: Twine.h:81
void setEntryCount(ProfileCount Count, const DenseSet< GlobalValue::GUID > *Imports=nullptr)
Set the entry count for this function.
Definition: Function.cpp:1345
An analysis pass based on legacy pass manager to deliver ProfileSummaryInfo.
LLVMContext & getContext() const
Get the global data context.
Definition: Module.h:242
Diagnostic information for optimization analysis remarks.
static StringRef getName(Value *V)
Metadata * getProfileSummary()
Returns profile summary metadata.
Definition: Module.cpp:512
StringRef getFilename() const
BlockT * getHeader() const
Definition: LoopInfo.h:100
Interval::succ_iterator succ_begin(Interval *I)
succ_begin/succ_end - define methods so that Intervals may be used just like BasicBlocks can with the...
Definition: Interval.h:103
Subprogram description.
const Instruction * getFirstNonPHIOrDbgOrLifetime() const
Returns a pointer to the first instruction in this block that is not a PHINode, a debug intrinsic...
Definition: BasicBlock.cpp:203
ArrayRef - Represent a constant reference to an array (0 or more elements consecutively in memory)...
Definition: APInt.h:33
bool extractProfTotalWeight(uint64_t &TotalVal) const
Retrieve total raw weight values of a branch.
Definition: Metadata.cpp:1340
static cl::opt< std::string > SampleProfileFile("sample-profile-file", cl::init(""), cl::value_desc("filename"), cl::desc("Profile file loaded by -sample-profile"), cl::Hidden)
LLVM_NODISCARD LLVM_ATTRIBUTE_ALWAYS_INLINE StringRef substr(size_t Start, size_t N=npos) const
Return a reference to the substring from [Start, Start + N).
Definition: StringRef.h:598
Debug location.
ModulePass * createSampleProfileLoaderPass()
Instruction * promoteIndirectCall(Instruction *Inst, Function *F, uint64_t Count, uint64_t TotalCount, bool AttachProfToDirectCall, OptimizationRemarkEmitter *ORE)
Concrete subclass of DominatorTreeBase that is used to compute a normal dominator tree...
Definition: Dominators.h:142
amdgpu Simplify well known AMD library false Value * Callee
static cl::opt< unsigned > SampleProfileRecordCoverage("sample-profile-check-record-coverage", cl::init(0), cl::value_desc("N"), cl::desc("Emit a warning if less than N% of records in the input profile " "are matched to the IR."))
static PreservedAnalyses none()
Convenience factory function for the empty preserved set.
Definition: PassManager.h:156
Interval::succ_iterator succ_end(Interval *I)
Definition: Interval.h:106
Core dominator tree base class.
Definition: LoopInfo.h:61
const BasicBlock & getEntryBlock() const
Definition: Function.h:626
static bool runOnFunction(Function &F, bool PostInlining)
initializer< Ty > init(const Ty &Val)
Definition: CommandLine.h:410
void dump(const SparseBitVector< ElementSize > &LHS, raw_ostream &out)
Subclasses of this class are all able to terminate a basic block.
Definition: InstrTypes.h:55
Wrapper pass for TargetTransformInfo.
A set of analyses that are preserved following a run of a transformation pass.
Definition: PassManager.h:153
* if(!EatIfPresent(lltok::kw_thread_local)) return false
ParseOptionalThreadLocal := /*empty.
void findInlinedFunctions(DenseSet< GlobalValue::GUID > &S, const Module *M, uint64_t Threshold, bool isCompact) const
Recursively traverses all children, if the total sample count of the corresponding function is no les...
Definition: SampleProf.h:379
LLVM Basic Block Representation.
Definition: BasicBlock.h:59
This is an important class for using LLVM in a threaded context.
Definition: LLVMContext.h:69
static cl::opt< unsigned > SampleProfileMaxPropagateIterations("sample-profile-max-propagate-iterations", cl::init(100), cl::desc("Maximum number of iterations to go through when propagating " "sample block/edge weights through the CFG."))
DISubprogram * getSubprogram() const
Get the attached subprogram.
Definition: Metadata.cpp:1508
An analysis pass based on the new PM to deliver ProfileSummaryInfo.
const BodySampleMap & getBodySamples() const
Return all the samples collected in the body of the function.
Definition: SampleProf.h:347
static GCRegistry::Add< CoreCLRGC > E("coreclr", "CoreCLR-compatible GC")
LLVM_ATTRIBUTE_ALWAYS_INLINE iterator begin()
Definition: SmallVector.h:117
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:193
Used in the streaming interface as the general argument type.
void annotateValueSite(Module &M, Instruction &Inst, const InstrProfRecord &InstrProfR, InstrProfValueKind ValueKind, uint32_t SiteIndx, uint32_t MaxMDCount=3)
Get the value profile data for value site SiteIdx from InstrProfR and annotate the instruction Inst w...
Definition: InstrProf.cpp:812
static PreservedAnalyses all()
Construct a special preserved set that preserves all passes.
Definition: PassManager.h:159
Class to represent profile counts.
Definition: Function.h:247
MDNode * createBranchWeights(uint32_t TrueWeight, uint32_t FalseWeight)
Return metadata containing two branch weights.
Definition: MDBuilder.cpp:38
INITIALIZE_PASS_END(RegBankSelect, DEBUG_TYPE, "Assign register bank of generic virtual registers", false, false) RegBankSelect
#define llvm_unreachable(msg)
Marks that the current location is not supposed to be reachable.
void setMetadata(unsigned KindID, MDNode *Node)
Set the metadata of the specified kind to the specified node.
Definition: Metadata.cpp:1226
InlineCost getInlineCost(CallSite CS, const InlineParams &Params, TargetTransformInfo &CalleeTTI, std::function< AssumptionCache &(Function &)> &GetAssumptionCache, Optional< function_ref< BlockFrequencyInfo &(Function &)>> GetBFI, ProfileSummaryInfo *PSI, OptimizationRemarkEmitter *ORE=nullptr)
Get an InlineCost object representing the cost of inlining this callsite.
Optional< bool > ComputeFullInlineCost
Compute inline cost even when the cost has exceeded the threshold.
Definition: InlineCost.h:157
void sort(IteratorTy Start, IteratorTy End)
Definition: STLExtras.h:859
const FunctionSamples * findFunctionSamplesAt(const LineLocation &Loc, StringRef CalleeName) const
Returns a pointer to FunctionSamples at the given callsite location Loc with callee CalleeName...
Definition: SampleProf.h:294
A function analysis which provides an AssumptionCache.
const InstListType & getInstList() const
Return the underlying instruction list container.
Definition: BasicBlock.h:329
SmallPtrSet - This class implements a set which is optimized for holding SmallSize or less elements...
Definition: SmallPtrSet.h:418
This pass provides access to the codegen interfaces that are needed for IR-level transformations.
InlineParams getInlineParams()
Generate the parameters to tune the inline cost analysis based only on the commandline options...
This is a &#39;vector&#39; (really, a variable-sized array), optimized for the case when the array is small...
Definition: SmallVector.h:861
Module.h This file contains the declarations for the Module class.
static unsigned getOffset(const DILocation *DIL)
Returns the line offset to the start line of the subprogram.
Definition: SampleProf.cpp:155
ErrorOr< SampleRecord::CallTargetMap > findCallTargetMapAt(uint32_t LineOffset, uint32_t Discriminator) const
Returns the call target map collected at a given location.
Definition: SampleProf.h:269
raw_ostream & dbgs()
dbgs() - This returns a reference to a raw_ostream for debugging messages.
Definition: Debug.cpp:133
StringMap - This is an unconventional map that is specialized for handling keys that are "strings"...
Definition: StringMap.h:220
PostDominatorTree Class - Concrete subclass of DominatorTree that is used to compute the post-dominat...
static void clear(coro::Shape &Shape)
Definition: Coroutines.cpp:211
iterator insert(iterator I, T &&Elt)
Definition: SmallVector.h:479
bool dominates(const DomTreeNodeBase< NodeT > *A, const DomTreeNodeBase< NodeT > *B) const
dominates - Returns true iff A dominates B.
void setProfileSummary(Metadata *M)
Attach profile summary metadata to this module.
Definition: Module.cpp:508
LLVM_ATTRIBUTE_ALWAYS_INLINE iterator end()
Definition: SmallVector.h:121
unsigned getBaseDiscriminator() const
Returns the base discriminator stored in the discriminator.
const DebugLoc & getDebugLoc() const
Return the debug location for this node as a DebugLoc.
Definition: Instruction.h:290
#define DEBUG_TYPE
static const size_t npos
Definition: StringRef.h:51
iterator begin()
Definition: StringMap.h:315
Represents the relative location of an instruction.
Definition: SampleProf.h:117
static cl::opt< unsigned > SampleProfileSampleCoverage("sample-profile-check-sample-coverage", cl::init(0), cl::value_desc("N"), cl::desc("Emit a warning if less than N% of samples in the input profile " "are matched to the IR."))
static bool callsiteIsHot(const FunctionSamples *CallsiteFS, ProfileSummaryInfo *PSI)
Return true if the given callsite is hot wrt to hot cutoff threshold.
Represents a single loop in the control flow graph.
Definition: LoopInfo.h:445
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:208
size_type count(const_arg_type_t< ValueT > V) const
Return 1 if the specified key is in the set, 0 otherwise.
Definition: DenseSet.h:91
void diagnose(const DiagnosticInfo &DI)
Report a message to the currently installed diagnostic handler.
INITIALIZE_PASS_BEGIN(SampleProfileLoaderLegacyPass, "sample-profile", "Sample Profile loader", false, false) INITIALIZE_PASS_END(SampleProfileLoaderLegacyPass
bool isDeclaration() const
Return true if the primary definition of this global value is outside of the current translation unit...
Definition: Globals.cpp:201
Provides ErrorOr<T> smart pointer.
FunTy * getCalledFunction() const
Return the function being called if this is a direct call, otherwise return null (if it&#39;s an indirect...
Definition: CallSite.h:107
assert(ImpDefSCC.getReg()==AMDGPU::SCC &&ImpDefSCC.isDef())
unsigned getNumSuccessors() const
Return the number of successors that this terminator has.
Module * getParent()
Get the module that this global value is contained inside of...
Definition: GlobalValue.h:565
uint64_t getTotalSamples() const
Return the total number of samples collected inside the function.
Definition: SampleProf.h:317
sample Sample Profile loader
const CallsiteSampleMap & getCallsiteSamples() const
Return all the callsite samples collected in the body of the function.
Definition: SampleProf.h:350
This class implements an extremely fast bulk output stream that can only output to a stream...
Definition: raw_ostream.h:46
print Print MemDeps of function
This file defines a set of templates that efficiently compute a dominator tree over a generic graph...
StringRef - Represent a constant reference to a string, i.e.
Definition: StringRef.h:49
A container for analyses that lazily runs them and caches their results.
This pass exposes codegen information to IR-level passes.
uint64_t getEntrySamples() const
Return the sample count of the first instruction of the function.
Definition: SampleProf.h:328
This header defines various interfaces for pass management in LLVM.
Diagnostic information for the sample profiler.
LLVM_NODISCARD LLVM_ATTRIBUTE_ALWAYS_INLINE size_t find(char C, size_t From=0) const
Search for the first character C in the string.
Definition: StringRef.h:298
#define LLVM_DEBUG(X)
Definition: Debug.h:119
The optimization diagnostic interface.
iterator end()
Definition: StringMap.h:318
This file provides the interface for the sampled PGO loader pass.
reference get()
Definition: ErrorOr.h: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