1 //===- SampleProfile.cpp - Incorporate sample profiles into the IR --------===//
3 // The LLVM Compiler Infrastructure
5 // This file is distributed under the University of Illinois Open Source
6 // License. See LICENSE.TXT for details.
8 //===----------------------------------------------------------------------===//
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.
15 // This pass generates branch weight annotations on the IR:
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.
23 //===----------------------------------------------------------------------===//
25 #include "llvm/ADT/DenseMap.h"
26 #include "llvm/ADT/SmallPtrSet.h"
27 #include "llvm/ADT/SmallSet.h"
28 #include "llvm/ADT/StringRef.h"
29 #include "llvm/Analysis/LoopInfo.h"
30 #include "llvm/Analysis/PostDominators.h"
31 #include "llvm/IR/Constants.h"
32 #include "llvm/IR/DebugInfo.h"
33 #include "llvm/IR/DiagnosticInfo.h"
34 #include "llvm/IR/Dominators.h"
35 #include "llvm/IR/Function.h"
36 #include "llvm/IR/InstIterator.h"
37 #include "llvm/IR/Instructions.h"
38 #include "llvm/IR/LLVMContext.h"
39 #include "llvm/IR/MDBuilder.h"
40 #include "llvm/IR/Metadata.h"
41 #include "llvm/IR/Module.h"
42 #include "llvm/Pass.h"
43 #include "llvm/ProfileData/SampleProfReader.h"
44 #include "llvm/Support/CommandLine.h"
45 #include "llvm/Support/Debug.h"
46 #include "llvm/Support/raw_ostream.h"
47 #include "llvm/Transforms/IPO.h"
51 using namespace sampleprof;
53 #define DEBUG_TYPE "sample-profile"
55 // Command line option to specify the file to read samples from. This is
56 // mainly used for debugging.
57 static cl::opt<std::string> SampleProfileFile(
58 "sample-profile-file", cl::init(""), cl::value_desc("filename"),
59 cl::desc("Profile file loaded by -sample-profile"), cl::Hidden);
60 static cl::opt<unsigned> SampleProfileMaxPropagateIterations(
61 "sample-profile-max-propagate-iterations", cl::init(100),
62 cl::desc("Maximum number of iterations to go through when propagating "
63 "sample block/edge weights through the CFG."));
66 typedef DenseMap<BasicBlock *, unsigned> BlockWeightMap;
67 typedef DenseMap<BasicBlock *, BasicBlock *> EquivalenceClassMap;
68 typedef std::pair<BasicBlock *, BasicBlock *> Edge;
69 typedef DenseMap<Edge, unsigned> EdgeWeightMap;
70 typedef DenseMap<BasicBlock *, SmallVector<BasicBlock *, 8>> BlockEdgeMap;
72 /// \brief Sample profile pass.
74 /// This pass reads profile data from the file specified by
75 /// -sample-profile-file and annotates every affected function with the
76 /// profile information found in that file.
77 class SampleProfileLoader : public ModulePass {
79 // Class identification, replacement for typeinfo
82 SampleProfileLoader(StringRef Name = SampleProfileFile)
83 : ModulePass(ID), DT(nullptr), PDT(nullptr), LI(nullptr), Reader(),
84 Samples(nullptr), Filename(Name), ProfileIsValid(false) {
85 initializeSampleProfileLoaderPass(*PassRegistry::getPassRegistry());
88 bool doInitialization(Module &M) override;
90 void dump() { Reader->dump(); }
92 const char *getPassName() const override { return "Sample profile pass"; }
94 bool runOnModule(Module &M) override;
96 void getAnalysisUsage(AnalysisUsage &AU) const override {
101 bool runOnFunction(Function &F);
102 unsigned getFunctionLoc(Function &F);
103 bool emitAnnotations(Function &F);
104 unsigned getInstWeight(Instruction &I);
105 unsigned getBlockWeight(BasicBlock *BB);
106 void printEdgeWeight(raw_ostream &OS, Edge E);
107 void printBlockWeight(raw_ostream &OS, BasicBlock *BB);
108 void printBlockEquivalence(raw_ostream &OS, BasicBlock *BB);
109 bool computeBlockWeights(Function &F);
110 void findEquivalenceClasses(Function &F);
111 void findEquivalencesFor(BasicBlock *BB1,
112 SmallVector<BasicBlock *, 8> Descendants,
113 DominatorTreeBase<BasicBlock> *DomTree);
114 void propagateWeights(Function &F);
115 unsigned visitEdge(Edge E, unsigned *NumUnknownEdges, Edge *UnknownEdge);
116 void buildEdges(Function &F);
117 bool propagateThroughEdges(Function &F);
118 void computeDominanceAndLoopInfo(Function &F);
120 /// \brief Line number for the function header. Used to compute absolute
121 /// line numbers from the relative line numbers found in the profile.
122 unsigned HeaderLineno;
124 /// \brief Map basic blocks to their computed weights.
126 /// The weight of a basic block is defined to be the maximum
127 /// of all the instruction weights in that block.
128 BlockWeightMap BlockWeights;
130 /// \brief Map edges to their computed weights.
132 /// Edge weights are computed by propagating basic block weights in
133 /// SampleProfile::propagateWeights.
134 EdgeWeightMap EdgeWeights;
136 /// \brief Set of visited blocks during propagation.
137 SmallPtrSet<BasicBlock *, 128> VisitedBlocks;
139 /// \brief Set of visited edges during propagation.
140 SmallSet<Edge, 128> VisitedEdges;
142 /// \brief Equivalence classes for block weights.
144 /// Two blocks BB1 and BB2 are in the same equivalence class if they
145 /// dominate and post-dominate each other, and they are in the same loop
146 /// nest. When this happens, the two blocks are guaranteed to execute
147 /// the same number of times.
148 EquivalenceClassMap EquivalenceClass;
150 /// \brief Dominance, post-dominance and loop information.
151 std::unique_ptr<DominatorTree> DT;
152 std::unique_ptr<DominatorTreeBase<BasicBlock>> PDT;
153 std::unique_ptr<LoopInfo> LI;
155 /// \brief Predecessors for each basic block in the CFG.
156 BlockEdgeMap Predecessors;
158 /// \brief Successors for each basic block in the CFG.
159 BlockEdgeMap Successors;
161 /// \brief Profile reader object.
162 std::unique_ptr<SampleProfileReader> Reader;
164 /// \brief Samples collected for the body of this function.
165 FunctionSamples *Samples;
167 /// \brief Name of the profile file to load.
170 /// \brief Flag indicating whether the profile input loaded successfully.
175 /// \brief Print the weight of edge \p E on stream \p OS.
177 /// \param OS Stream to emit the output to.
178 /// \param E Edge to print.
179 void SampleProfileLoader::printEdgeWeight(raw_ostream &OS, Edge E) {
180 OS << "weight[" << E.first->getName() << "->" << E.second->getName()
181 << "]: " << EdgeWeights[E] << "\n";
184 /// \brief Print the equivalence class of block \p BB on stream \p OS.
186 /// \param OS Stream to emit the output to.
187 /// \param BB Block to print.
188 void SampleProfileLoader::printBlockEquivalence(raw_ostream &OS,
190 BasicBlock *Equiv = EquivalenceClass[BB];
191 OS << "equivalence[" << BB->getName()
192 << "]: " << ((Equiv) ? EquivalenceClass[BB]->getName() : "NONE") << "\n";
195 /// \brief Print the weight of block \p BB on stream \p OS.
197 /// \param OS Stream to emit the output to.
198 /// \param BB Block to print.
199 void SampleProfileLoader::printBlockWeight(raw_ostream &OS, BasicBlock *BB) {
200 OS << "weight[" << BB->getName() << "]: " << BlockWeights[BB] << "\n";
203 /// \brief Get the weight for an instruction.
205 /// The "weight" of an instruction \p Inst is the number of samples
206 /// collected on that instruction at runtime. To retrieve it, we
207 /// need to compute the line number of \p Inst relative to the start of its
208 /// function. We use HeaderLineno to compute the offset. We then
209 /// look up the samples collected for \p Inst using BodySamples.
211 /// \param Inst Instruction to query.
213 /// \returns The profiled weight of I.
214 unsigned SampleProfileLoader::getInstWeight(Instruction &Inst) {
215 DebugLoc DLoc = Inst.getDebugLoc();
219 unsigned Lineno = DLoc.getLine();
220 if (Lineno < HeaderLineno)
223 const DILocation *DIL = DLoc;
224 int LOffset = Lineno - HeaderLineno;
225 unsigned Discriminator = DIL->getDiscriminator();
226 unsigned Weight = Samples->samplesAt(LOffset, Discriminator);
227 DEBUG(dbgs() << " " << Lineno << "." << Discriminator << ":" << Inst
228 << " (line offset: " << LOffset << "." << Discriminator
229 << " - weight: " << Weight << ")\n");
233 /// \brief Compute the weight of a basic block.
235 /// The weight of basic block \p BB is the maximum weight of all the
236 /// instructions in BB. The weight of \p BB is computed and cached in
237 /// the BlockWeights map.
239 /// \param BB The basic block to query.
241 /// \returns The computed weight of BB.
242 unsigned SampleProfileLoader::getBlockWeight(BasicBlock *BB) {
243 // If we've computed BB's weight before, return it.
244 std::pair<BlockWeightMap::iterator, bool> Entry =
245 BlockWeights.insert(std::make_pair(BB, 0));
247 return Entry.first->second;
249 // Otherwise, compute and cache BB's weight.
251 for (auto &I : BB->getInstList()) {
252 unsigned InstWeight = getInstWeight(I);
253 if (InstWeight > Weight)
256 Entry.first->second = Weight;
260 /// \brief Compute and store the weights of every basic block.
262 /// This populates the BlockWeights map by computing
263 /// the weights of every basic block in the CFG.
265 /// \param F The function to query.
266 bool SampleProfileLoader::computeBlockWeights(Function &F) {
267 bool Changed = false;
268 DEBUG(dbgs() << "Block weights\n");
270 unsigned Weight = getBlockWeight(&BB);
271 Changed |= (Weight > 0);
272 DEBUG(printBlockWeight(dbgs(), &BB));
278 /// \brief Find equivalence classes for the given block.
280 /// This finds all the blocks that are guaranteed to execute the same
281 /// number of times as \p BB1. To do this, it traverses all the
282 /// descendants of \p BB1 in the dominator or post-dominator tree.
284 /// A block BB2 will be in the same equivalence class as \p BB1 if
285 /// the following holds:
287 /// 1- \p BB1 is a descendant of BB2 in the opposite tree. So, if BB2
288 /// is a descendant of \p BB1 in the dominator tree, then BB2 should
289 /// dominate BB1 in the post-dominator tree.
291 /// 2- Both BB2 and \p BB1 must be in the same loop.
293 /// For every block BB2 that meets those two requirements, we set BB2's
294 /// equivalence class to \p BB1.
296 /// \param BB1 Block to check.
297 /// \param Descendants Descendants of \p BB1 in either the dom or pdom tree.
298 /// \param DomTree Opposite dominator tree. If \p Descendants is filled
299 /// with blocks from \p BB1's dominator tree, then
300 /// this is the post-dominator tree, and vice versa.
301 void SampleProfileLoader::findEquivalencesFor(
302 BasicBlock *BB1, SmallVector<BasicBlock *, 8> Descendants,
303 DominatorTreeBase<BasicBlock> *DomTree) {
304 for (auto *BB2 : Descendants) {
305 bool IsDomParent = DomTree->dominates(BB2, BB1);
306 bool IsInSameLoop = LI->getLoopFor(BB1) == LI->getLoopFor(BB2);
307 if (BB1 != BB2 && VisitedBlocks.insert(BB2).second && IsDomParent &&
309 EquivalenceClass[BB2] = BB1;
311 // If BB2 is heavier than BB1, make BB2 have the same weight
314 // Note that we don't worry about the opposite situation here
315 // (when BB2 is lighter than BB1). We will deal with this
316 // during the propagation phase. Right now, we just want to
317 // make sure that BB1 has the largest weight of all the
318 // members of its equivalence set.
319 unsigned &BB1Weight = BlockWeights[BB1];
320 unsigned &BB2Weight = BlockWeights[BB2];
321 BB1Weight = std::max(BB1Weight, BB2Weight);
326 /// \brief Find equivalence classes.
328 /// Since samples may be missing from blocks, we can fill in the gaps by setting
329 /// the weights of all the blocks in the same equivalence class to the same
330 /// weight. To compute the concept of equivalence, we use dominance and loop
331 /// information. Two blocks B1 and B2 are in the same equivalence class if B1
332 /// dominates B2, B2 post-dominates B1 and both are in the same loop.
334 /// \param F The function to query.
335 void SampleProfileLoader::findEquivalenceClasses(Function &F) {
336 SmallVector<BasicBlock *, 8> DominatedBBs;
337 DEBUG(dbgs() << "\nBlock equivalence classes\n");
338 // Find equivalence sets based on dominance and post-dominance information.
340 BasicBlock *BB1 = &BB;
342 // Compute BB1's equivalence class once.
343 if (EquivalenceClass.count(BB1)) {
344 DEBUG(printBlockEquivalence(dbgs(), BB1));
348 // By default, blocks are in their own equivalence class.
349 EquivalenceClass[BB1] = BB1;
351 // Traverse all the blocks dominated by BB1. We are looking for
352 // every basic block BB2 such that:
354 // 1- BB1 dominates BB2.
355 // 2- BB2 post-dominates BB1.
356 // 3- BB1 and BB2 are in the same loop nest.
358 // If all those conditions hold, it means that BB2 is executed
359 // as many times as BB1, so they are placed in the same equivalence
360 // class by making BB2's equivalence class be BB1.
361 DominatedBBs.clear();
362 DT->getDescendants(BB1, DominatedBBs);
363 findEquivalencesFor(BB1, DominatedBBs, PDT.get());
365 // Repeat the same logic for all the blocks post-dominated by BB1.
366 // We are looking for every basic block BB2 such that:
368 // 1- BB1 post-dominates BB2.
369 // 2- BB2 dominates BB1.
370 // 3- BB1 and BB2 are in the same loop nest.
372 // If all those conditions hold, BB2's equivalence class is BB1.
373 DominatedBBs.clear();
374 PDT->getDescendants(BB1, DominatedBBs);
375 findEquivalencesFor(BB1, DominatedBBs, DT.get());
377 DEBUG(printBlockEquivalence(dbgs(), BB1));
380 // Assign weights to equivalence classes.
382 // All the basic blocks in the same equivalence class will execute
383 // the same number of times. Since we know that the head block in
384 // each equivalence class has the largest weight, assign that weight
385 // to all the blocks in that equivalence class.
386 DEBUG(dbgs() << "\nAssign the same weight to all blocks in the same class\n");
388 BasicBlock *BB = &BI;
389 BasicBlock *EquivBB = EquivalenceClass[BB];
391 BlockWeights[BB] = BlockWeights[EquivBB];
392 DEBUG(printBlockWeight(dbgs(), BB));
396 /// \brief Visit the given edge to decide if it has a valid weight.
398 /// If \p E has not been visited before, we copy to \p UnknownEdge
399 /// and increment the count of unknown edges.
401 /// \param E Edge to visit.
402 /// \param NumUnknownEdges Current number of unknown edges.
403 /// \param UnknownEdge Set if E has not been visited before.
405 /// \returns E's weight, if known. Otherwise, return 0.
406 unsigned SampleProfileLoader::visitEdge(Edge E, unsigned *NumUnknownEdges,
408 if (!VisitedEdges.count(E)) {
409 (*NumUnknownEdges)++;
414 return EdgeWeights[E];
417 /// \brief Propagate weights through incoming/outgoing edges.
419 /// If the weight of a basic block is known, and there is only one edge
420 /// with an unknown weight, we can calculate the weight of that edge.
422 /// Similarly, if all the edges have a known count, we can calculate the
423 /// count of the basic block, if needed.
425 /// \param F Function to process.
427 /// \returns True if new weights were assigned to edges or blocks.
428 bool SampleProfileLoader::propagateThroughEdges(Function &F) {
429 bool Changed = false;
430 DEBUG(dbgs() << "\nPropagation through edges\n");
432 BasicBlock *BB = &BI;
434 // Visit all the predecessor and successor edges to determine
435 // which ones have a weight assigned already. Note that it doesn't
436 // matter that we only keep track of a single unknown edge. The
437 // only case we are interested in handling is when only a single
438 // edge is unknown (see setEdgeOrBlockWeight).
439 for (unsigned i = 0; i < 2; i++) {
440 unsigned TotalWeight = 0;
441 unsigned NumUnknownEdges = 0;
442 Edge UnknownEdge, SelfReferentialEdge;
445 // First, visit all predecessor edges.
446 for (auto *Pred : Predecessors[BB]) {
447 Edge E = std::make_pair(Pred, BB);
448 TotalWeight += visitEdge(E, &NumUnknownEdges, &UnknownEdge);
449 if (E.first == E.second)
450 SelfReferentialEdge = E;
453 // On the second round, visit all successor edges.
454 for (auto *Succ : Successors[BB]) {
455 Edge E = std::make_pair(BB, Succ);
456 TotalWeight += visitEdge(E, &NumUnknownEdges, &UnknownEdge);
460 // After visiting all the edges, there are three cases that we
461 // can handle immediately:
463 // - All the edge weights are known (i.e., NumUnknownEdges == 0).
464 // In this case, we simply check that the sum of all the edges
465 // is the same as BB's weight. If not, we change BB's weight
466 // to match. Additionally, if BB had not been visited before,
467 // we mark it visited.
469 // - Only one edge is unknown and BB has already been visited.
470 // In this case, we can compute the weight of the edge by
471 // subtracting the total block weight from all the known
472 // edge weights. If the edges weight more than BB, then the
473 // edge of the last remaining edge is set to zero.
475 // - There exists a self-referential edge and the weight of BB is
476 // known. In this case, this edge can be based on BB's weight.
477 // We add up all the other known edges and set the weight on
478 // the self-referential edge as we did in the previous case.
480 // In any other case, we must continue iterating. Eventually,
481 // all edges will get a weight, or iteration will stop when
482 // it reaches SampleProfileMaxPropagateIterations.
483 if (NumUnknownEdges <= 1) {
484 unsigned &BBWeight = BlockWeights[BB];
485 if (NumUnknownEdges == 0) {
486 // If we already know the weight of all edges, the weight of the
487 // basic block can be computed. It should be no larger than the sum
488 // of all edge weights.
489 if (TotalWeight > BBWeight) {
490 BBWeight = TotalWeight;
492 DEBUG(dbgs() << "All edge weights for " << BB->getName()
493 << " known. Set weight for block: ";
494 printBlockWeight(dbgs(), BB););
496 if (VisitedBlocks.insert(BB).second)
498 } else if (NumUnknownEdges == 1 && VisitedBlocks.count(BB)) {
499 // If there is a single unknown edge and the block has been
500 // visited, then we can compute E's weight.
501 if (BBWeight >= TotalWeight)
502 EdgeWeights[UnknownEdge] = BBWeight - TotalWeight;
504 EdgeWeights[UnknownEdge] = 0;
505 VisitedEdges.insert(UnknownEdge);
507 DEBUG(dbgs() << "Set weight for edge: ";
508 printEdgeWeight(dbgs(), UnknownEdge));
510 } else if (SelfReferentialEdge.first && VisitedBlocks.count(BB)) {
511 unsigned &BBWeight = BlockWeights[BB];
512 // We have a self-referential edge and the weight of BB is known.
513 if (BBWeight >= TotalWeight)
514 EdgeWeights[SelfReferentialEdge] = BBWeight - TotalWeight;
516 EdgeWeights[SelfReferentialEdge] = 0;
517 VisitedEdges.insert(SelfReferentialEdge);
519 DEBUG(dbgs() << "Set self-referential edge weight to: ";
520 printEdgeWeight(dbgs(), SelfReferentialEdge));
528 /// \brief Build in/out edge lists for each basic block in the CFG.
530 /// We are interested in unique edges. If a block B1 has multiple
531 /// edges to another block B2, we only add a single B1->B2 edge.
532 void SampleProfileLoader::buildEdges(Function &F) {
534 BasicBlock *B1 = &BI;
536 // Add predecessors for B1.
537 SmallPtrSet<BasicBlock *, 16> Visited;
538 if (!Predecessors[B1].empty())
539 llvm_unreachable("Found a stale predecessors list in a basic block.");
540 for (pred_iterator PI = pred_begin(B1), PE = pred_end(B1); PI != PE; ++PI) {
541 BasicBlock *B2 = *PI;
542 if (Visited.insert(B2).second)
543 Predecessors[B1].push_back(B2);
546 // Add successors for B1.
548 if (!Successors[B1].empty())
549 llvm_unreachable("Found a stale successors list in a basic block.");
550 for (succ_iterator SI = succ_begin(B1), SE = succ_end(B1); SI != SE; ++SI) {
551 BasicBlock *B2 = *SI;
552 if (Visited.insert(B2).second)
553 Successors[B1].push_back(B2);
558 /// \brief Propagate weights into edges
560 /// The following rules are applied to every block BB in the CFG:
562 /// - If BB has a single predecessor/successor, then the weight
563 /// of that edge is the weight of the block.
565 /// - If all incoming or outgoing edges are known except one, and the
566 /// weight of the block is already known, the weight of the unknown
567 /// edge will be the weight of the block minus the sum of all the known
568 /// edges. If the sum of all the known edges is larger than BB's weight,
569 /// we set the unknown edge weight to zero.
571 /// - If there is a self-referential edge, and the weight of the block is
572 /// known, the weight for that edge is set to the weight of the block
573 /// minus the weight of the other incoming edges to that block (if
575 void SampleProfileLoader::propagateWeights(Function &F) {
579 // Add an entry count to the function using the samples gathered
580 // at the function entry.
581 F.setEntryCount(Samples->getHeadSamples());
583 // Before propagation starts, build, for each block, a list of
584 // unique predecessors and successors. This is necessary to handle
585 // identical edges in multiway branches. Since we visit all blocks and all
586 // edges of the CFG, it is cleaner to build these lists once at the start
590 // Propagate until we converge or we go past the iteration limit.
591 while (Changed && i++ < SampleProfileMaxPropagateIterations) {
592 Changed = propagateThroughEdges(F);
595 // Generate MD_prof metadata for every branch instruction using the
596 // edge weights computed during propagation.
597 DEBUG(dbgs() << "\nPropagation complete. Setting branch weights\n");
598 MDBuilder MDB(F.getContext());
600 BasicBlock *BB = &BI;
601 TerminatorInst *TI = BB->getTerminator();
602 if (TI->getNumSuccessors() == 1)
604 if (!isa<BranchInst>(TI) && !isa<SwitchInst>(TI))
607 DEBUG(dbgs() << "\nGetting weights for branch at line "
608 << TI->getDebugLoc().getLine() << ".\n");
609 SmallVector<unsigned, 4> Weights;
610 bool AllWeightsZero = true;
611 for (unsigned I = 0; I < TI->getNumSuccessors(); ++I) {
612 BasicBlock *Succ = TI->getSuccessor(I);
613 Edge E = std::make_pair(BB, Succ);
614 unsigned Weight = EdgeWeights[E];
615 DEBUG(dbgs() << "\t"; printEdgeWeight(dbgs(), E));
616 Weights.push_back(Weight);
618 AllWeightsZero = false;
621 // Only set weights if there is at least one non-zero weight.
622 // In any other case, let the analyzer set weights.
623 if (!AllWeightsZero) {
624 DEBUG(dbgs() << "SUCCESS. Found non-zero weights.\n");
625 TI->setMetadata(llvm::LLVMContext::MD_prof,
626 MDB.createBranchWeights(Weights));
628 DEBUG(dbgs() << "SKIPPED. All branch weights are zero.\n");
633 /// \brief Get the line number for the function header.
635 /// This looks up function \p F in the current compilation unit and
636 /// retrieves the line number where the function is defined. This is
637 /// line 0 for all the samples read from the profile file. Every line
638 /// number is relative to this line.
640 /// \param F Function object to query.
642 /// \returns the line number where \p F is defined. If it returns 0,
643 /// it means that there is no debug information available for \p F.
644 unsigned SampleProfileLoader::getFunctionLoc(Function &F) {
645 if (DISubprogram *S = getDISubprogram(&F))
648 // If could not find the start of \p F, emit a diagnostic to inform the user
649 // about the missed opportunity.
650 F.getContext().diagnose(DiagnosticInfoSampleProfile(
651 "No debug information found in function " + F.getName() +
652 ": Function profile not used",
657 void SampleProfileLoader::computeDominanceAndLoopInfo(Function &F) {
658 DT.reset(new DominatorTree);
661 PDT.reset(new DominatorTreeBase<BasicBlock>(true));
664 LI.reset(new LoopInfo);
668 /// \brief Generate branch weight metadata for all branches in \p F.
670 /// Branch weights are computed out of instruction samples using a
671 /// propagation heuristic. Propagation proceeds in 3 phases:
673 /// 1- Assignment of block weights. All the basic blocks in the function
674 /// are initial assigned the same weight as their most frequently
675 /// executed instruction.
677 /// 2- Creation of equivalence classes. Since samples may be missing from
678 /// blocks, we can fill in the gaps by setting the weights of all the
679 /// blocks in the same equivalence class to the same weight. To compute
680 /// the concept of equivalence, we use dominance and loop information.
681 /// Two blocks B1 and B2 are in the same equivalence class if B1
682 /// dominates B2, B2 post-dominates B1 and both are in the same loop.
684 /// 3- Propagation of block weights into edges. This uses a simple
685 /// propagation heuristic. The following rules are applied to every
686 /// block BB in the CFG:
688 /// - If BB has a single predecessor/successor, then the weight
689 /// of that edge is the weight of the block.
691 /// - If all the edges are known except one, and the weight of the
692 /// block is already known, the weight of the unknown edge will
693 /// be the weight of the block minus the sum of all the known
694 /// edges. If the sum of all the known edges is larger than BB's weight,
695 /// we set the unknown edge weight to zero.
697 /// - If there is a self-referential edge, and the weight of the block is
698 /// known, the weight for that edge is set to the weight of the block
699 /// minus the weight of the other incoming edges to that block (if
702 /// Since this propagation is not guaranteed to finalize for every CFG, we
703 /// only allow it to proceed for a limited number of iterations (controlled
704 /// by -sample-profile-max-propagate-iterations).
706 /// FIXME: Try to replace this propagation heuristic with a scheme
707 /// that is guaranteed to finalize. A work-list approach similar to
708 /// the standard value propagation algorithm used by SSA-CCP might
711 /// Once all the branch weights are computed, we emit the MD_prof
712 /// metadata on BB using the computed values for each of its branches.
714 /// \param F The function to query.
716 /// \returns true if \p F was modified. Returns false, otherwise.
717 bool SampleProfileLoader::emitAnnotations(Function &F) {
718 bool Changed = false;
720 // Initialize invariants used during computation and propagation.
721 HeaderLineno = getFunctionLoc(F);
722 if (HeaderLineno == 0)
725 DEBUG(dbgs() << "Line number for the first instruction in " << F.getName()
726 << ": " << HeaderLineno << "\n");
728 // Compute basic block weights.
729 Changed |= computeBlockWeights(F);
732 // Compute dominance and loop info needed for propagation.
733 computeDominanceAndLoopInfo(F);
735 // Find equivalence classes.
736 findEquivalenceClasses(F);
738 // Propagate weights to all edges.
745 char SampleProfileLoader::ID = 0;
746 INITIALIZE_PASS_BEGIN(SampleProfileLoader, "sample-profile",
747 "Sample Profile loader", false, false)
748 INITIALIZE_PASS_DEPENDENCY(AddDiscriminators)
749 INITIALIZE_PASS_END(SampleProfileLoader, "sample-profile",
750 "Sample Profile loader", false, false)
752 bool SampleProfileLoader::doInitialization(Module &M) {
753 auto &Ctx = M.getContext();
754 auto ReaderOrErr = SampleProfileReader::create(Filename, Ctx);
755 if (std::error_code EC = ReaderOrErr.getError()) {
756 std::string Msg = "Could not open profile: " + EC.message();
757 Ctx.diagnose(DiagnosticInfoSampleProfile(Filename.data(), Msg));
760 Reader = std::move(ReaderOrErr.get());
761 ProfileIsValid = (Reader->read() == sampleprof_error::success);
765 ModulePass *llvm::createSampleProfileLoaderPass() {
766 return new SampleProfileLoader(SampleProfileFile);
769 ModulePass *llvm::createSampleProfileLoaderPass(StringRef Name) {
770 return new SampleProfileLoader(Name);
773 bool SampleProfileLoader::runOnModule(Module &M) {
776 if (!F.isDeclaration())
777 retval |= runOnFunction(F);
781 bool SampleProfileLoader::runOnFunction(Function &F) {
785 Samples = Reader->getSamplesFor(F);
786 if (!Samples->empty())
787 return emitAnnotations(F);