Bug Summary

File:lib/Transforms/IPO/SampleProfile.cpp
Warning:line 706, column 10
Called C++ object pointer is null

Annotated Source Code

Press '?' to see keyboard shortcuts

clang -cc1 -triple x86_64-pc-linux-gnu -analyze -disable-free -disable-llvm-verifier -discard-value-names -main-file-name SampleProfile.cpp -analyzer-store=region -analyzer-opt-analyze-nested-blocks -analyzer-eagerly-assume -analyzer-checker=core -analyzer-checker=apiModeling -analyzer-checker=unix -analyzer-checker=deadcode -analyzer-checker=cplusplus -analyzer-checker=security.insecureAPI.UncheckedReturn -analyzer-checker=security.insecureAPI.getpw -analyzer-checker=security.insecureAPI.gets -analyzer-checker=security.insecureAPI.mktemp -analyzer-checker=security.insecureAPI.mkstemp -analyzer-checker=security.insecureAPI.vfork -analyzer-checker=nullability.NullPassedToNonnull -analyzer-checker=nullability.NullReturnedFromNonnull -analyzer-output plist -w -mrelocation-model pic -pic-level 2 -mthread-model posix -fmath-errno -masm-verbose -mconstructor-aliases -munwind-tables -fuse-init-array -target-cpu x86-64 -dwarf-column-info -debugger-tuning=gdb -momit-leaf-frame-pointer -ffunction-sections -fdata-sections -resource-dir /usr/lib/llvm-7/lib/clang/7.0.0 -D _DEBUG -D _GNU_SOURCE -D __STDC_CONSTANT_MACROS -D __STDC_FORMAT_MACROS -D __STDC_LIMIT_MACROS -I /build/llvm-toolchain-snapshot-7~svn329677/build-llvm/lib/Transforms/IPO -I /build/llvm-toolchain-snapshot-7~svn329677/lib/Transforms/IPO -I /build/llvm-toolchain-snapshot-7~svn329677/build-llvm/include -I /build/llvm-toolchain-snapshot-7~svn329677/include -U NDEBUG -internal-isystem /usr/lib/gcc/x86_64-linux-gnu/7.3.0/../../../../include/c++/7.3.0 -internal-isystem /usr/lib/gcc/x86_64-linux-gnu/7.3.0/../../../../include/x86_64-linux-gnu/c++/7.3.0 -internal-isystem /usr/lib/gcc/x86_64-linux-gnu/7.3.0/../../../../include/x86_64-linux-gnu/c++/7.3.0 -internal-isystem /usr/lib/gcc/x86_64-linux-gnu/7.3.0/../../../../include/c++/7.3.0/backward -internal-isystem /usr/include/clang/7.0.0/include/ -internal-isystem /usr/local/include -internal-isystem /usr/lib/llvm-7/lib/clang/7.0.0/include -internal-externc-isystem /usr/include/x86_64-linux-gnu -internal-externc-isystem /include -internal-externc-isystem /usr/include -O2 -Wno-unused-parameter -Wwrite-strings -Wno-missing-field-initializers -Wno-long-long -Wno-maybe-uninitialized -Wno-comment -std=c++11 -fdeprecated-macro -fdebug-compilation-dir /build/llvm-toolchain-snapshot-7~svn329677/build-llvm/lib/Transforms/IPO -ferror-limit 19 -fmessage-length 0 -fvisibility-inlines-hidden -fobjc-runtime=gcc -fdiagnostics-show-option -vectorize-loops -vectorize-slp -analyzer-checker optin.performance.Padding -analyzer-output=html -analyzer-config stable-report-filename=true -o /tmp/scan-build-2018-04-11-031539-24776-1 -x c++ /build/llvm-toolchain-snapshot-7~svn329677/lib/Transforms/IPO/SampleProfile.cpp
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
25#include "llvm/Transforms/IPO/SampleProfile.h"
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"
36#include "llvm/Analysis/AssumptionCache.h"
37#include "llvm/Analysis/InlineCost.h"
38#include "llvm/Analysis/LoopInfo.h"
39#include "llvm/Analysis/OptimizationRemarkEmitter.h"
40#include "llvm/Analysis/TargetTransformInfo.h"
41#include "llvm/IR/BasicBlock.h"
42#include "llvm/IR/CFG.h"
43#include "llvm/IR/CallSite.h"
44#include "llvm/IR/DebugInfoMetadata.h"
45#include "llvm/IR/DebugLoc.h"
46#include "llvm/IR/DiagnosticInfo.h"
47#include "llvm/IR/Dominators.h"
48#include "llvm/IR/Function.h"
49#include "llvm/IR/GlobalValue.h"
50#include "llvm/IR/InstrTypes.h"
51#include "llvm/IR/Instruction.h"
52#include "llvm/IR/Instructions.h"
53#include "llvm/IR/IntrinsicInst.h"
54#include "llvm/IR/LLVMContext.h"
55#include "llvm/IR/MDBuilder.h"
56#include "llvm/IR/Module.h"
57#include "llvm/IR/PassManager.h"
58#include "llvm/IR/ValueSymbolTable.h"
59#include "llvm/Pass.h"
60#include "llvm/ProfileData/InstrProf.h"
61#include "llvm/ProfileData/SampleProf.h"
62#include "llvm/ProfileData/SampleProfReader.h"
63#include "llvm/Support/Casting.h"
64#include "llvm/Support/CommandLine.h"
65#include "llvm/Support/Debug.h"
66#include "llvm/Support/ErrorHandling.h"
67#include "llvm/Support/ErrorOr.h"
68#include "llvm/Support/GenericDomTree.h"
69#include "llvm/Support/raw_ostream.h"
70#include "llvm/Transforms/IPO.h"
71#include "llvm/Transforms/Instrumentation.h"
72#include "llvm/Transforms/Utils/CallPromotionUtils.h"
73#include "llvm/Transforms/Utils/Cloning.h"
74#include <algorithm>
75#include <cassert>
76#include <cstdint>
77#include <functional>
78#include <limits>
79#include <map>
80#include <memory>
81#include <string>
82#include <system_error>
83#include <utility>
84#include <vector>
85
86using namespace llvm;
87using namespace sampleprof;
88using ProfileCount = Function::ProfileCount;
89#define DEBUG_TYPE"sample-profile" "sample-profile"
90
91// Command line option to specify the file to read samples from. This is
92// mainly used for debugging.
93static cl::opt<std::string> SampleProfileFile(
94 "sample-profile-file", cl::init(""), cl::value_desc("filename"),
95 cl::desc("Profile file loaded by -sample-profile"), cl::Hidden);
96
97static cl::opt<unsigned> SampleProfileMaxPropagateIterations(
98 "sample-profile-max-propagate-iterations", cl::init(100),
99 cl::desc("Maximum number of iterations to go through when propagating "
100 "sample block/edge weights through the CFG."));
101
102static cl::opt<unsigned> SampleProfileRecordCoverage(
103 "sample-profile-check-record-coverage", cl::init(0), cl::value_desc("N"),
104 cl::desc("Emit a warning if less than N% of records in the input profile "
105 "are matched to the IR."));
106
107static cl::opt<unsigned> SampleProfileSampleCoverage(
108 "sample-profile-check-sample-coverage", cl::init(0), cl::value_desc("N"),
109 cl::desc("Emit a warning if less than N% of samples in the input profile "
110 "are matched to the IR."));
111
112static cl::opt<double> SampleProfileHotThreshold(
113 "sample-profile-inline-hot-threshold", cl::init(0.1), cl::value_desc("N"),
114 cl::desc("Inlined functions that account for more than N% of all samples "
115 "collected in the parent function, will be inlined again."));
116
117namespace {
118
119using BlockWeightMap = DenseMap<const BasicBlock *, uint64_t>;
120using EquivalenceClassMap = DenseMap<const BasicBlock *, const BasicBlock *>;
121using Edge = std::pair<const BasicBlock *, const BasicBlock *>;
122using EdgeWeightMap = DenseMap<Edge, uint64_t>;
123using BlockEdgeMap =
124 DenseMap<const BasicBlock *, SmallVector<const BasicBlock *, 8>>;
125
126class SampleCoverageTracker {
127public:
128 SampleCoverageTracker() = default;
129
130 bool markSamplesUsed(const FunctionSamples *FS, uint32_t LineOffset,
131 uint32_t Discriminator, uint64_t Samples);
132 unsigned computeCoverage(unsigned Used, unsigned Total) const;
133 unsigned countUsedRecords(const FunctionSamples *FS) const;
134 unsigned countBodyRecords(const FunctionSamples *FS) const;
135 uint64_t getTotalUsedSamples() const { return TotalUsedSamples; }
136 uint64_t countBodySamples(const FunctionSamples *FS) const;
137
138 void clear() {
139 SampleCoverage.clear();
140 TotalUsedSamples = 0;
141 }
142
143private:
144 using BodySampleCoverageMap = std::map<LineLocation, unsigned>;
145 using FunctionSamplesCoverageMap =
146 DenseMap<const FunctionSamples *, BodySampleCoverageMap>;
147
148 /// Coverage map for sampling records.
149 ///
150 /// This map keeps a record of sampling records that have been matched to
151 /// an IR instruction. This is used to detect some form of staleness in
152 /// profiles (see flag -sample-profile-check-coverage).
153 ///
154 /// Each entry in the map corresponds to a FunctionSamples instance. This is
155 /// another map that counts how many times the sample record at the
156 /// given location has been used.
157 FunctionSamplesCoverageMap SampleCoverage;
158
159 /// Number of samples used from the profile.
160 ///
161 /// When a sampling record is used for the first time, the samples from
162 /// that record are added to this accumulator. Coverage is later computed
163 /// based on the total number of samples available in this function and
164 /// its callsites.
165 ///
166 /// Note that this accumulator tracks samples used from a single function
167 /// and all the inlined callsites. Strictly, we should have a map of counters
168 /// keyed by FunctionSamples pointers, but these stats are cleared after
169 /// every function, so we just need to keep a single counter.
170 uint64_t TotalUsedSamples = 0;
171};
172
173/// \brief Sample profile pass.
174///
175/// This pass reads profile data from the file specified by
176/// -sample-profile-file and annotates every affected function with the
177/// profile information found in that file.
178class SampleProfileLoader {
179public:
180 SampleProfileLoader(
181 StringRef Name, bool IsThinLTOPreLink,
182 std::function<AssumptionCache &(Function &)> GetAssumptionCache,
183 std::function<TargetTransformInfo &(Function &)> GetTargetTransformInfo)
184 : GetAC(std::move(GetAssumptionCache)),
185 GetTTI(std::move(GetTargetTransformInfo)), Filename(Name),
186 IsThinLTOPreLink(IsThinLTOPreLink) {}
187
188 bool doInitialization(Module &M);
189 bool runOnModule(Module &M, ModuleAnalysisManager *AM);
190
191 void dump() { Reader->dump(); }
192
193protected:
194 bool runOnFunction(Function &F, ModuleAnalysisManager *AM);
195 unsigned getFunctionLoc(Function &F);
196 bool emitAnnotations(Function &F);
197 ErrorOr<uint64_t> getInstWeight(const Instruction &I);
198 ErrorOr<uint64_t> getBlockWeight(const BasicBlock *BB);
199 const FunctionSamples *findCalleeFunctionSamples(const Instruction &I) const;
200 std::vector<const FunctionSamples *>
201 findIndirectCallFunctionSamples(const Instruction &I, uint64_t &Sum) const;
202 const FunctionSamples *findFunctionSamples(const Instruction &I) const;
203 bool inlineCallInstruction(Instruction *I);
204 bool inlineHotFunctions(Function &F,
205 DenseSet<GlobalValue::GUID> &InlinedGUIDs);
206 void printEdgeWeight(raw_ostream &OS, Edge E);
207 void printBlockWeight(raw_ostream &OS, const BasicBlock *BB) const;
208 void printBlockEquivalence(raw_ostream &OS, const BasicBlock *BB);
209 bool computeBlockWeights(Function &F);
210 void findEquivalenceClasses(Function &F);
211 template <bool IsPostDom>
212 void findEquivalencesFor(BasicBlock *BB1, ArrayRef<BasicBlock *> Descendants,
213 DominatorTreeBase<BasicBlock, IsPostDom> *DomTree);
214
215 void propagateWeights(Function &F);
216 uint64_t visitEdge(Edge E, unsigned *NumUnknownEdges, Edge *UnknownEdge);
217 void buildEdges(Function &F);
218 bool propagateThroughEdges(Function &F, bool UpdateBlockCount);
219 void computeDominanceAndLoopInfo(Function &F);
220 void clearFunctionData();
221
222 /// \brief Map basic blocks to their computed weights.
223 ///
224 /// The weight of a basic block is defined to be the maximum
225 /// of all the instruction weights in that block.
226 BlockWeightMap BlockWeights;
227
228 /// \brief Map edges to their computed weights.
229 ///
230 /// Edge weights are computed by propagating basic block weights in
231 /// SampleProfile::propagateWeights.
232 EdgeWeightMap EdgeWeights;
233
234 /// \brief Set of visited blocks during propagation.
235 SmallPtrSet<const BasicBlock *, 32> VisitedBlocks;
236
237 /// \brief Set of visited edges during propagation.
238 SmallSet<Edge, 32> VisitedEdges;
239
240 /// \brief Equivalence classes for block weights.
241 ///
242 /// Two blocks BB1 and BB2 are in the same equivalence class if they
243 /// dominate and post-dominate each other, and they are in the same loop
244 /// nest. When this happens, the two blocks are guaranteed to execute
245 /// the same number of times.
246 EquivalenceClassMap EquivalenceClass;
247
248 /// Map from function name to Function *. Used to find the function from
249 /// the function name. If the function name contains suffix, additional
250 /// entry is added to map from the stripped name to the function if there
251 /// is one-to-one mapping.
252 StringMap<Function *> SymbolMap;
253
254 /// \brief Dominance, post-dominance and loop information.
255 std::unique_ptr<DominatorTree> DT;
256 std::unique_ptr<PostDomTreeBase<BasicBlock>> PDT;
257 std::unique_ptr<LoopInfo> LI;
258
259 std::function<AssumptionCache &(Function &)> GetAC;
260 std::function<TargetTransformInfo &(Function &)> GetTTI;
261
262 /// \brief Predecessors for each basic block in the CFG.
263 BlockEdgeMap Predecessors;
264
265 /// \brief Successors for each basic block in the CFG.
266 BlockEdgeMap Successors;
267
268 SampleCoverageTracker CoverageTracker;
269
270 /// \brief Profile reader object.
271 std::unique_ptr<SampleProfileReader> Reader;
272
273 /// \brief Samples collected for the body of this function.
274 FunctionSamples *Samples = nullptr;
275
276 /// \brief Name of the profile file to load.
277 std::string Filename;
278
279 /// \brief Flag indicating whether the profile input loaded successfully.
280 bool ProfileIsValid = false;
281
282 /// \brief Flag indicating if the pass is invoked in ThinLTO compile phase.
283 ///
284 /// In this phase, in annotation, we should not promote indirect calls.
285 /// Instead, we will mark GUIDs that needs to be annotated to the function.
286 bool IsThinLTOPreLink;
287
288 /// \brief Total number of samples collected in this profile.
289 ///
290 /// This is the sum of all the samples collected in all the functions executed
291 /// at runtime.
292 uint64_t TotalCollectedSamples = 0;
293
294 /// \brief Optimization Remark Emitter used to emit diagnostic remarks.
295 OptimizationRemarkEmitter *ORE = nullptr;
296};
297
298class SampleProfileLoaderLegacyPass : public ModulePass {
299public:
300 // Class identification, replacement for typeinfo
301 static char ID;
302
303 SampleProfileLoaderLegacyPass(StringRef Name = SampleProfileFile,
304 bool IsThinLTOPreLink = false)
305 : ModulePass(ID), SampleLoader(Name, IsThinLTOPreLink,
306 [&](Function &F) -> AssumptionCache & {
307 return ACT->getAssumptionCache(F);
308 },
309 [&](Function &F) -> TargetTransformInfo & {
310 return TTIWP->getTTI(F);
311 }) {
312 initializeSampleProfileLoaderLegacyPassPass(
313 *PassRegistry::getPassRegistry());
314 }
315
316 void dump() { SampleLoader.dump(); }
317
318 bool doInitialization(Module &M) override {
319 return SampleLoader.doInitialization(M);
320 }
321
322 StringRef getPassName() const override { return "Sample profile pass"; }
323 bool runOnModule(Module &M) override;
324
325 void getAnalysisUsage(AnalysisUsage &AU) const override {
326 AU.addRequired<AssumptionCacheTracker>();
327 AU.addRequired<TargetTransformInfoWrapperPass>();
328 }
329
330private:
331 SampleProfileLoader SampleLoader;
332 AssumptionCacheTracker *ACT = nullptr;
333 TargetTransformInfoWrapperPass *TTIWP = nullptr;
334};
335
336} // end anonymous namespace
337
338/// Return true if the given callsite is hot wrt to its caller.
339///
340/// Functions that were inlined in the original binary will be represented
341/// in the inline stack in the sample profile. If the profile shows that
342/// the original inline decision was "good" (i.e., the callsite is executed
343/// frequently), then we will recreate the inline decision and apply the
344/// profile from the inlined callsite.
345///
346/// To decide whether an inlined callsite is hot, we compute the fraction
347/// of samples used by the callsite with respect to the total number of samples
348/// collected in the caller.
349///
350/// If that fraction is larger than the default given by
351/// SampleProfileHotThreshold, the callsite will be inlined again.
352static bool callsiteIsHot(const FunctionSamples *CallerFS,
353 const FunctionSamples *CallsiteFS) {
354 if (!CallsiteFS)
355 return false; // The callsite was not inlined in the original binary.
356
357 uint64_t ParentTotalSamples = CallerFS->getTotalSamples();
358 if (ParentTotalSamples == 0)
359 return false; // Avoid division by zero.
360
361 uint64_t CallsiteTotalSamples = CallsiteFS->getTotalSamples();
362 if (CallsiteTotalSamples == 0)
363 return false; // Callsite is trivially cold.
364
365 double PercentSamples =
366 (double)CallsiteTotalSamples / (double)ParentTotalSamples * 100.0;
367 return PercentSamples >= SampleProfileHotThreshold;
368}
369
370/// Mark as used the sample record for the given function samples at
371/// (LineOffset, Discriminator).
372///
373/// \returns true if this is the first time we mark the given record.
374bool SampleCoverageTracker::markSamplesUsed(const FunctionSamples *FS,
375 uint32_t LineOffset,
376 uint32_t Discriminator,
377 uint64_t Samples) {
378 LineLocation Loc(LineOffset, Discriminator);
379 unsigned &Count = SampleCoverage[FS][Loc];
380 bool FirstTime = (++Count == 1);
381 if (FirstTime)
382 TotalUsedSamples += Samples;
383 return FirstTime;
384}
385
386/// Return the number of sample records that were applied from this profile.
387///
388/// This count does not include records from cold inlined callsites.
389unsigned
390SampleCoverageTracker::countUsedRecords(const FunctionSamples *FS) const {
391 auto I = SampleCoverage.find(FS);
392
393 // The size of the coverage map for FS represents the number of records
394 // that were marked used at least once.
395 unsigned Count = (I != SampleCoverage.end()) ? I->second.size() : 0;
396
397 // If there are inlined callsites in this function, count the samples found
398 // in the respective bodies. However, do not bother counting callees with 0
399 // total samples, these are callees that were never invoked at runtime.
400 for (const auto &I : FS->getCallsiteSamples())
401 for (const auto &J : I.second) {
402 const FunctionSamples *CalleeSamples = &J.second;
403 if (callsiteIsHot(FS, CalleeSamples))
404 Count += countUsedRecords(CalleeSamples);
405 }
406
407 return Count;
408}
409
410/// Return the number of sample records in the body of this profile.
411///
412/// This count does not include records from cold inlined callsites.
413unsigned
414SampleCoverageTracker::countBodyRecords(const FunctionSamples *FS) const {
415 unsigned Count = FS->getBodySamples().size();
416
417 // Only count records in hot callsites.
418 for (const auto &I : FS->getCallsiteSamples())
419 for (const auto &J : I.second) {
420 const FunctionSamples *CalleeSamples = &J.second;
421 if (callsiteIsHot(FS, CalleeSamples))
422 Count += countBodyRecords(CalleeSamples);
423 }
424
425 return Count;
426}
427
428/// Return the number of samples collected in the body of this profile.
429///
430/// This count does not include samples from cold inlined callsites.
431uint64_t
432SampleCoverageTracker::countBodySamples(const FunctionSamples *FS) const {
433 uint64_t Total = 0;
434 for (const auto &I : FS->getBodySamples())
435 Total += I.second.getSamples();
436
437 // Only count samples in hot callsites.
438 for (const auto &I : FS->getCallsiteSamples())
439 for (const auto &J : I.second) {
440 const FunctionSamples *CalleeSamples = &J.second;
441 if (callsiteIsHot(FS, CalleeSamples))
442 Total += countBodySamples(CalleeSamples);
443 }
444
445 return Total;
446}
447
448/// Return the fraction of sample records used in this profile.
449///
450/// The returned value is an unsigned integer in the range 0-100 indicating
451/// the percentage of sample records that were used while applying this
452/// profile to the associated function.
453unsigned SampleCoverageTracker::computeCoverage(unsigned Used,
454 unsigned Total) const {
455 assert(Used <= Total &&(static_cast <bool> (Used <= Total && "number of used records cannot exceed the total number of records"
) ? void (0) : __assert_fail ("Used <= Total && \"number of used records cannot exceed the total number of records\""
, "/build/llvm-toolchain-snapshot-7~svn329677/lib/Transforms/IPO/SampleProfile.cpp"
, 456, __extension__ __PRETTY_FUNCTION__))
456 "number of used records cannot exceed the total number of records")(static_cast <bool> (Used <= Total && "number of used records cannot exceed the total number of records"
) ? void (0) : __assert_fail ("Used <= Total && \"number of used records cannot exceed the total number of records\""
, "/build/llvm-toolchain-snapshot-7~svn329677/lib/Transforms/IPO/SampleProfile.cpp"
, 456, __extension__ __PRETTY_FUNCTION__))
;
457 return Total > 0 ? Used * 100 / Total : 100;
458}
459
460/// Clear all the per-function data used to load samples and propagate weights.
461void SampleProfileLoader::clearFunctionData() {
462 BlockWeights.clear();
463 EdgeWeights.clear();
464 VisitedBlocks.clear();
465 VisitedEdges.clear();
466 EquivalenceClass.clear();
467 DT = nullptr;
468 PDT = nullptr;
469 LI = nullptr;
470 Predecessors.clear();
471 Successors.clear();
472 CoverageTracker.clear();
473}
474
475#ifndef NDEBUG
476/// \brief Print the weight of edge \p E on stream \p OS.
477///
478/// \param OS Stream to emit the output to.
479/// \param E Edge to print.
480void SampleProfileLoader::printEdgeWeight(raw_ostream &OS, Edge E) {
481 OS << "weight[" << E.first->getName() << "->" << E.second->getName()
482 << "]: " << EdgeWeights[E] << "\n";
483}
484
485/// \brief Print the equivalence class of block \p BB on stream \p OS.
486///
487/// \param OS Stream to emit the output to.
488/// \param BB Block to print.
489void SampleProfileLoader::printBlockEquivalence(raw_ostream &OS,
490 const BasicBlock *BB) {
491 const BasicBlock *Equiv = EquivalenceClass[BB];
492 OS << "equivalence[" << BB->getName()
493 << "]: " << ((Equiv) ? EquivalenceClass[BB]->getName() : "NONE") << "\n";
494}
495
496/// \brief Print the weight of block \p BB on stream \p OS.
497///
498/// \param OS Stream to emit the output to.
499/// \param BB Block to print.
500void SampleProfileLoader::printBlockWeight(raw_ostream &OS,
501 const BasicBlock *BB) const {
502 const auto &I = BlockWeights.find(BB);
503 uint64_t W = (I == BlockWeights.end() ? 0 : I->second);
504 OS << "weight[" << BB->getName() << "]: " << W << "\n";
505}
506#endif
507
508/// \brief Get the weight for an instruction.
509///
510/// The "weight" of an instruction \p Inst is the number of samples
511/// collected on that instruction at runtime. To retrieve it, we
512/// need to compute the line number of \p Inst relative to the start of its
513/// function. We use HeaderLineno to compute the offset. We then
514/// look up the samples collected for \p Inst using BodySamples.
515///
516/// \param Inst Instruction to query.
517///
518/// \returns the weight of \p Inst.
519ErrorOr<uint64_t> SampleProfileLoader::getInstWeight(const Instruction &Inst) {
520 const DebugLoc &DLoc = Inst.getDebugLoc();
521 if (!DLoc)
2
Assuming the condition is false
3
Taking false branch
9
Assuming the condition is false
10
Taking false branch
522 return std::error_code();
523
524 const FunctionSamples *FS = findFunctionSamples(Inst);
11
Calling 'SampleProfileLoader::findFunctionSamples'
525 if (!FS)
4
Assuming 'FS' is null
5
Taking true branch
526 return std::error_code();
527
528 // Ignore all intrinsics and branch instructions.
529 // Branch instruction usually contains debug info from sources outside of
530 // the residing basic block, thus we ignore them during annotation.
531 if (isa<BranchInst>(Inst) || isa<IntrinsicInst>(Inst))
532 return std::error_code();
533
534 // If a direct call/invoke instruction is inlined in profile
535 // (findCalleeFunctionSamples returns non-empty result), but not inlined here,
536 // it means that the inlined callsite has no sample, thus the call
537 // instruction should have 0 count.
538 if ((isa<CallInst>(Inst) || isa<InvokeInst>(Inst)) &&
539 !ImmutableCallSite(&Inst).isIndirectCall() &&
540 findCalleeFunctionSamples(Inst))
541 return 0;
542
543 const DILocation *DIL = DLoc;
544 uint32_t LineOffset = FunctionSamples::getOffset(DIL);
545 uint32_t Discriminator = DIL->getBaseDiscriminator();
546 ErrorOr<uint64_t> R = FS->findSamplesAt(LineOffset, Discriminator);
547 if (R) {
548 bool FirstMark =
549 CoverageTracker.markSamplesUsed(FS, LineOffset, Discriminator, R.get());
550 if (FirstMark) {
551 ORE->emit([&]() {
552 OptimizationRemarkAnalysis Remark(DEBUG_TYPE"sample-profile", "AppliedSamples", &Inst);
553 Remark << "Applied " << ore::NV("NumSamples", *R);
554 Remark << " samples from profile (offset: ";
555 Remark << ore::NV("LineOffset", LineOffset);
556 if (Discriminator) {
557 Remark << ".";
558 Remark << ore::NV("Discriminator", Discriminator);
559 }
560 Remark << ")";
561 return Remark;
562 });
563 }
564 DEBUG(dbgs() << " " << DLoc.getLine() << "."do { if (::llvm::DebugFlag && ::llvm::isCurrentDebugType
("sample-profile")) { dbgs() << " " << DLoc.getLine
() << "." << DIL->getBaseDiscriminator() <<
":" << Inst << " (line offset: " << LineOffset
<< "." << DIL->getBaseDiscriminator() <<
" - weight: " << R.get() << ")\n"; } } while (false
)
565 << DIL->getBaseDiscriminator() << ":" << Instdo { if (::llvm::DebugFlag && ::llvm::isCurrentDebugType
("sample-profile")) { dbgs() << " " << DLoc.getLine
() << "." << DIL->getBaseDiscriminator() <<
":" << Inst << " (line offset: " << LineOffset
<< "." << DIL->getBaseDiscriminator() <<
" - weight: " << R.get() << ")\n"; } } while (false
)
566 << " (line offset: " << LineOffset << "."do { if (::llvm::DebugFlag && ::llvm::isCurrentDebugType
("sample-profile")) { dbgs() << " " << DLoc.getLine
() << "." << DIL->getBaseDiscriminator() <<
":" << Inst << " (line offset: " << LineOffset
<< "." << DIL->getBaseDiscriminator() <<
" - weight: " << R.get() << ")\n"; } } while (false
)
567 << DIL->getBaseDiscriminator() << " - weight: " << R.get()do { if (::llvm::DebugFlag && ::llvm::isCurrentDebugType
("sample-profile")) { dbgs() << " " << DLoc.getLine
() << "." << DIL->getBaseDiscriminator() <<
":" << Inst << " (line offset: " << LineOffset
<< "." << DIL->getBaseDiscriminator() <<
" - weight: " << R.get() << ")\n"; } } while (false
)
568 << ")\n")do { if (::llvm::DebugFlag && ::llvm::isCurrentDebugType
("sample-profile")) { dbgs() << " " << DLoc.getLine
() << "." << DIL->getBaseDiscriminator() <<
":" << Inst << " (line offset: " << LineOffset
<< "." << DIL->getBaseDiscriminator() <<
" - weight: " << R.get() << ")\n"; } } while (false
)
;
569 }
570 return R;
571}
572
573/// \brief Compute the weight of a basic block.
574///
575/// The weight of basic block \p BB is the maximum weight of all the
576/// instructions in BB.
577///
578/// \param BB The basic block to query.
579///
580/// \returns the weight for \p BB.
581ErrorOr<uint64_t> SampleProfileLoader::getBlockWeight(const BasicBlock *BB) {
582 uint64_t Max = 0;
583 bool HasWeight = false;
584 for (auto &I : BB->getInstList()) {
585 const ErrorOr<uint64_t> &R = getInstWeight(I);
1
Calling 'SampleProfileLoader::getInstWeight'
6
Returning from 'SampleProfileLoader::getInstWeight'
8
Calling 'SampleProfileLoader::getInstWeight'
586 if (R) {
7
Taking false branch
587 Max = std::max(Max, R.get());
588 HasWeight = true;
589 }
590 }
591 return HasWeight ? ErrorOr<uint64_t>(Max) : std::error_code();
592}
593
594/// \brief Compute and store the weights of every basic block.
595///
596/// This populates the BlockWeights map by computing
597/// the weights of every basic block in the CFG.
598///
599/// \param F The function to query.
600bool SampleProfileLoader::computeBlockWeights(Function &F) {
601 bool Changed = false;
602 DEBUG(dbgs() << "Block weights\n")do { if (::llvm::DebugFlag && ::llvm::isCurrentDebugType
("sample-profile")) { dbgs() << "Block weights\n"; } } while
(false)
;
603 for (const auto &BB : F) {
604 ErrorOr<uint64_t> Weight = getBlockWeight(&BB);
605 if (Weight) {
606 BlockWeights[&BB] = Weight.get();
607 VisitedBlocks.insert(&BB);
608 Changed = true;
609 }
610 DEBUG(printBlockWeight(dbgs(), &BB))do { if (::llvm::DebugFlag && ::llvm::isCurrentDebugType
("sample-profile")) { printBlockWeight(dbgs(), &BB); } } while
(false)
;
611 }
612
613 return Changed;
614}
615
616/// \brief Get the FunctionSamples for a call instruction.
617///
618/// The FunctionSamples of a call/invoke instruction \p Inst is the inlined
619/// instance in which that call instruction is calling to. It contains
620/// all samples that resides in the inlined instance. We first find the
621/// inlined instance in which the call instruction is from, then we
622/// traverse its children to find the callsite with the matching
623/// location.
624///
625/// \param Inst Call/Invoke instruction to query.
626///
627/// \returns The FunctionSamples pointer to the inlined instance.
628const FunctionSamples *
629SampleProfileLoader::findCalleeFunctionSamples(const Instruction &Inst) const {
630 const DILocation *DIL = Inst.getDebugLoc();
631 if (!DIL) {
632 return nullptr;
633 }
634
635 StringRef CalleeName;
636 if (const CallInst *CI = dyn_cast<CallInst>(&Inst))
637 if (Function *Callee = CI->getCalledFunction())
638 CalleeName = Callee->getName();
639
640 const FunctionSamples *FS = findFunctionSamples(Inst);
641 if (FS == nullptr)
642 return nullptr;
643
644 return FS->findFunctionSamplesAt(LineLocation(FunctionSamples::getOffset(DIL),
645 DIL->getBaseDiscriminator()),
646 CalleeName);
647}
648
649/// Returns a vector of FunctionSamples that are the indirect call targets
650/// of \p Inst. The vector is sorted by the total number of samples. Stores
651/// the total call count of the indirect call in \p Sum.
652std::vector<const FunctionSamples *>
653SampleProfileLoader::findIndirectCallFunctionSamples(
654 const Instruction &Inst, uint64_t &Sum) const {
655 const DILocation *DIL = Inst.getDebugLoc();
656 std::vector<const FunctionSamples *> R;
657
658 if (!DIL) {
659 return R;
660 }
661
662 const FunctionSamples *FS = findFunctionSamples(Inst);
663 if (FS == nullptr)
664 return R;
665
666 uint32_t LineOffset = FunctionSamples::getOffset(DIL);
667 uint32_t Discriminator = DIL->getBaseDiscriminator();
668
669 auto T = FS->findCallTargetMapAt(LineOffset, Discriminator);
670 Sum = 0;
671 if (T)
672 for (const auto &T_C : T.get())
673 Sum += T_C.second;
674 if (const FunctionSamplesMap *M = FS->findFunctionSamplesMapAt(LineLocation(
675 FunctionSamples::getOffset(DIL), DIL->getBaseDiscriminator()))) {
676 if (M->empty())
677 return R;
678 for (const auto &NameFS : *M) {
679 Sum += NameFS.second.getEntrySamples();
680 R.push_back(&NameFS.second);
681 }
682 std::sort(R.begin(), R.end(),
683 [](const FunctionSamples *L, const FunctionSamples *R) {
684 return L->getEntrySamples() > R->getEntrySamples();
685 });
686 }
687 return R;
688}
689
690/// \brief Get the FunctionSamples for an instruction.
691///
692/// The FunctionSamples of an instruction \p Inst is the inlined instance
693/// in which that instruction is coming from. We traverse the inline stack
694/// of that instruction, and match it with the tree nodes in the profile.
695///
696/// \param Inst Instruction to query.
697///
698/// \returns the FunctionSamples pointer to the inlined instance.
699const FunctionSamples *
700SampleProfileLoader::findFunctionSamples(const Instruction &Inst) const {
701 SmallVector<std::pair<LineLocation, StringRef>, 10> S;
702 const DILocation *DIL = Inst.getDebugLoc();
703 if (!DIL)
12
Assuming 'DIL' is non-null
13
Taking false branch
704 return Samples;
705
706 return Samples->findFunctionSamples(DIL);
14
Called C++ object pointer is null
707}
708
709bool SampleProfileLoader::inlineCallInstruction(Instruction *I) {
710 assert(isa<CallInst>(I) || isa<InvokeInst>(I))(static_cast <bool> (isa<CallInst>(I) || isa<InvokeInst
>(I)) ? void (0) : __assert_fail ("isa<CallInst>(I) || isa<InvokeInst>(I)"
, "/build/llvm-toolchain-snapshot-7~svn329677/lib/Transforms/IPO/SampleProfile.cpp"
, 710, __extension__ __PRETTY_FUNCTION__))
;
711 CallSite CS(I);
712 Function *CalledFunction = CS.getCalledFunction();
713 assert(CalledFunction)(static_cast <bool> (CalledFunction) ? void (0) : __assert_fail
("CalledFunction", "/build/llvm-toolchain-snapshot-7~svn329677/lib/Transforms/IPO/SampleProfile.cpp"
, 713, __extension__ __PRETTY_FUNCTION__))
;
714 DebugLoc DLoc = I->getDebugLoc();
715 BasicBlock *BB = I->getParent();
716 InlineParams Params = getInlineParams();
717 Params.ComputeFullInlineCost = true;
718 // Checks if there is anything in the reachable portion of the callee at
719 // this callsite that makes this inlining potentially illegal. Need to
720 // set ComputeFullInlineCost, otherwise getInlineCost may return early
721 // when cost exceeds threshold without checking all IRs in the callee.
722 // The acutal cost does not matter because we only checks isNever() to
723 // see if it is legal to inline the callsite.
724 InlineCost Cost = getInlineCost(CS, Params, GetTTI(*CalledFunction), GetAC,
725 None, nullptr, nullptr);
726 if (Cost.isNever()) {
727 ORE->emit(OptimizationRemark(DEBUG_TYPE"sample-profile", "Not inline", DLoc, BB)
728 << "incompatible inlining");
729 return false;
730 }
731 InlineFunctionInfo IFI(nullptr, &GetAC);
732 if (InlineFunction(CS, IFI)) {
733 // The call to InlineFunction erases I, so we can't pass it here.
734 ORE->emit(OptimizationRemark(DEBUG_TYPE"sample-profile", "HotInline", DLoc, BB)
735 << "inlined hot callee '" << ore::NV("Callee", CalledFunction)
736 << "' into '" << ore::NV("Caller", BB->getParent()) << "'");
737 return true;
738 }
739 return false;
740}
741
742/// \brief Iteratively inline hot callsites of a function.
743///
744/// Iteratively traverse all callsites of the function \p F, and find if
745/// the corresponding inlined instance exists and is hot in profile. If
746/// it is hot enough, inline the callsites and adds new callsites of the
747/// callee into the caller. If the call is an indirect call, first promote
748/// it to direct call. Each indirect call is limited with a single target.
749///
750/// \param F function to perform iterative inlining.
751/// \param InlinedGUIDs a set to be updated to include all GUIDs that are
752/// inlined in the profiled binary.
753///
754/// \returns True if there is any inline happened.
755bool SampleProfileLoader::inlineHotFunctions(
756 Function &F, DenseSet<GlobalValue::GUID> &InlinedGUIDs) {
757 DenseSet<Instruction *> PromotedInsns;
758 bool Changed = false;
759 while (true) {
760 bool LocalChanged = false;
761 SmallVector<Instruction *, 10> CIS;
762 for (auto &BB : F) {
763 bool Hot = false;
764 SmallVector<Instruction *, 10> Candidates;
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(Samples, FS))
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(),
790 Samples->getTotalSamples() *
791 SampleProfileHotThreshold / 100);
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 if (CalleeFunctionName == F.getName())
801 continue;
802
803 const char *Reason = "Callee function not available";
804 auto R = SymbolMap.find(CalleeFunctionName);
805 if (R != SymbolMap.end() && R->getValue() &&
806 !R->getValue()->isDeclaration() &&
807 R->getValue()->getSubprogram() &&
808 isLegalToPromote(CallSite(I), R->getValue(), &Reason)) {
809 uint64_t C = FS->getEntrySamples();
810 Instruction *DI =
811 pgo::promoteIndirectCall(I, R->getValue(), C, Sum, false, ORE);
812 Sum -= C;
813 PromotedInsns.insert(I);
814 // If profile mismatches, we should not attempt to inline DI.
815 if ((isa<CallInst>(DI) || isa<InvokeInst>(DI)) &&
816 inlineCallInstruction(DI))
817 LocalChanged = true;
818 } else {
819 DEBUG(dbgs()do { if (::llvm::DebugFlag && ::llvm::isCurrentDebugType
("sample-profile")) { dbgs() << "\nFailed to promote indirect call to "
<< CalleeFunctionName << " because " << Reason
<< "\n"; } } while (false)
820 << "\nFailed to promote indirect call to "do { if (::llvm::DebugFlag && ::llvm::isCurrentDebugType
("sample-profile")) { dbgs() << "\nFailed to promote indirect call to "
<< CalleeFunctionName << " because " << Reason
<< "\n"; } } while (false)
821 << CalleeFunctionName << " because " << Reason << "\n")do { if (::llvm::DebugFlag && ::llvm::isCurrentDebugType
("sample-profile")) { dbgs() << "\nFailed to promote indirect call to "
<< CalleeFunctionName << " because " << Reason
<< "\n"; } } while (false)
;
822 }
823 }
824 } else if (CalledFunction && CalledFunction->getSubprogram() &&
825 !CalledFunction->isDeclaration()) {
826 if (inlineCallInstruction(I))
827 LocalChanged = true;
828 } else if (IsThinLTOPreLink) {
829 findCalleeFunctionSamples(*I)->findInlinedFunctions(
830 InlinedGUIDs, F.getParent(),
831 Samples->getTotalSamples() * SampleProfileHotThreshold / 100);
832 }
833 }
834 if (LocalChanged) {
835 Changed = true;
836 } else {
837 break;
838 }
839 }
840 return Changed;
841}
842
843/// \brief Find equivalence classes for the given block.
844///
845/// This finds all the blocks that are guaranteed to execute the same
846/// number of times as \p BB1. To do this, it traverses all the
847/// descendants of \p BB1 in the dominator or post-dominator tree.
848///
849/// A block BB2 will be in the same equivalence class as \p BB1 if
850/// the following holds:
851///
852/// 1- \p BB1 is a descendant of BB2 in the opposite tree. So, if BB2
853/// is a descendant of \p BB1 in the dominator tree, then BB2 should
854/// dominate BB1 in the post-dominator tree.
855///
856/// 2- Both BB2 and \p BB1 must be in the same loop.
857///
858/// For every block BB2 that meets those two requirements, we set BB2's
859/// equivalence class to \p BB1.
860///
861/// \param BB1 Block to check.
862/// \param Descendants Descendants of \p BB1 in either the dom or pdom tree.
863/// \param DomTree Opposite dominator tree. If \p Descendants is filled
864/// with blocks from \p BB1's dominator tree, then
865/// this is the post-dominator tree, and vice versa.
866template <bool IsPostDom>
867void SampleProfileLoader::findEquivalencesFor(
868 BasicBlock *BB1, ArrayRef<BasicBlock *> Descendants,
869 DominatorTreeBase<BasicBlock, IsPostDom> *DomTree) {
870 const BasicBlock *EC = EquivalenceClass[BB1];
871 uint64_t Weight = BlockWeights[EC];
872 for (const auto *BB2 : Descendants) {
873 bool IsDomParent = DomTree->dominates(BB2, BB1);
874 bool IsInSameLoop = LI->getLoopFor(BB1) == LI->getLoopFor(BB2);
875 if (BB1 != BB2 && IsDomParent && IsInSameLoop) {
876 EquivalenceClass[BB2] = EC;
877 // If BB2 is visited, then the entire EC should be marked as visited.
878 if (VisitedBlocks.count(BB2)) {
879 VisitedBlocks.insert(EC);
880 }
881
882 // If BB2 is heavier than BB1, make BB2 have the same weight
883 // as BB1.
884 //
885 // Note that we don't worry about the opposite situation here
886 // (when BB2 is lighter than BB1). We will deal with this
887 // during the propagation phase. Right now, we just want to
888 // make sure that BB1 has the largest weight of all the
889 // members of its equivalence set.
890 Weight = std::max(Weight, BlockWeights[BB2]);
891 }
892 }
893 if (EC == &EC->getParent()->getEntryBlock()) {
894 BlockWeights[EC] = Samples->getHeadSamples() + 1;
895 } else {
896 BlockWeights[EC] = Weight;
897 }
898}
899
900/// \brief Find equivalence classes.
901///
902/// Since samples may be missing from blocks, we can fill in the gaps by setting
903/// the weights of all the blocks in the same equivalence class to the same
904/// weight. To compute the concept of equivalence, we use dominance and loop
905/// information. Two blocks B1 and B2 are in the same equivalence class if B1
906/// dominates B2, B2 post-dominates B1 and both are in the same loop.
907///
908/// \param F The function to query.
909void SampleProfileLoader::findEquivalenceClasses(Function &F) {
910 SmallVector<BasicBlock *, 8> DominatedBBs;
911 DEBUG(dbgs() << "\nBlock equivalence classes\n")do { if (::llvm::DebugFlag && ::llvm::isCurrentDebugType
("sample-profile")) { dbgs() << "\nBlock equivalence classes\n"
; } } while (false)
;
912 // Find equivalence sets based on dominance and post-dominance information.
913 for (auto &BB : F) {
914 BasicBlock *BB1 = &BB;
915
916 // Compute BB1's equivalence class once.
917 if (EquivalenceClass.count(BB1)) {
918 DEBUG(printBlockEquivalence(dbgs(), BB1))do { if (::llvm::DebugFlag && ::llvm::isCurrentDebugType
("sample-profile")) { printBlockEquivalence(dbgs(), BB1); } }
while (false)
;
919 continue;
920 }
921
922 // By default, blocks are in their own equivalence class.
923 EquivalenceClass[BB1] = BB1;
924
925 // Traverse all the blocks dominated by BB1. We are looking for
926 // every basic block BB2 such that:
927 //
928 // 1- BB1 dominates BB2.
929 // 2- BB2 post-dominates BB1.
930 // 3- BB1 and BB2 are in the same loop nest.
931 //
932 // If all those conditions hold, it means that BB2 is executed
933 // as many times as BB1, so they are placed in the same equivalence
934 // class by making BB2's equivalence class be BB1.
935 DominatedBBs.clear();
936 DT->getDescendants(BB1, DominatedBBs);
937 findEquivalencesFor(BB1, DominatedBBs, PDT.get());
938
939 DEBUG(printBlockEquivalence(dbgs(), BB1))do { if (::llvm::DebugFlag && ::llvm::isCurrentDebugType
("sample-profile")) { printBlockEquivalence(dbgs(), BB1); } }
while (false)
;
940 }
941
942 // Assign weights to equivalence classes.
943 //
944 // All the basic blocks in the same equivalence class will execute
945 // the same number of times. Since we know that the head block in
946 // each equivalence class has the largest weight, assign that weight
947 // to all the blocks in that equivalence class.
948 DEBUG(dbgs() << "\nAssign the same weight to all blocks in the same class\n")do { if (::llvm::DebugFlag && ::llvm::isCurrentDebugType
("sample-profile")) { dbgs() << "\nAssign the same weight to all blocks in the same class\n"
; } } while (false)
;
949 for (auto &BI : F) {
950 const BasicBlock *BB = &BI;
951 const BasicBlock *EquivBB = EquivalenceClass[BB];
952 if (BB != EquivBB)
953 BlockWeights[BB] = BlockWeights[EquivBB];
954 DEBUG(printBlockWeight(dbgs(), BB))do { if (::llvm::DebugFlag && ::llvm::isCurrentDebugType
("sample-profile")) { printBlockWeight(dbgs(), BB); } } while
(false)
;
955 }
956}
957
958/// \brief Visit the given edge to decide if it has a valid weight.
959///
960/// If \p E has not been visited before, we copy to \p UnknownEdge
961/// and increment the count of unknown edges.
962///
963/// \param E Edge to visit.
964/// \param NumUnknownEdges Current number of unknown edges.
965/// \param UnknownEdge Set if E has not been visited before.
966///
967/// \returns E's weight, if known. Otherwise, return 0.
968uint64_t SampleProfileLoader::visitEdge(Edge E, unsigned *NumUnknownEdges,
969 Edge *UnknownEdge) {
970 if (!VisitedEdges.count(E)) {
971 (*NumUnknownEdges)++;
972 *UnknownEdge = E;
973 return 0;
974 }
975
976 return EdgeWeights[E];
977}
978
979/// \brief Propagate weights through incoming/outgoing edges.
980///
981/// If the weight of a basic block is known, and there is only one edge
982/// with an unknown weight, we can calculate the weight of that edge.
983///
984/// Similarly, if all the edges have a known count, we can calculate the
985/// count of the basic block, if needed.
986///
987/// \param F Function to process.
988/// \param UpdateBlockCount Whether we should update basic block counts that
989/// has already been annotated.
990///
991/// \returns True if new weights were assigned to edges or blocks.
992bool SampleProfileLoader::propagateThroughEdges(Function &F,
993 bool UpdateBlockCount) {
994 bool Changed = false;
995 DEBUG(dbgs() << "\nPropagation through edges\n")do { if (::llvm::DebugFlag && ::llvm::isCurrentDebugType
("sample-profile")) { dbgs() << "\nPropagation through edges\n"
; } } while (false)
;
996 for (const auto &BI : F) {
997 const BasicBlock *BB = &BI;
998 const BasicBlock *EC = EquivalenceClass[BB];
999
1000 // Visit all the predecessor and successor edges to determine
1001 // which ones have a weight assigned already. Note that it doesn't
1002 // matter that we only keep track of a single unknown edge. The
1003 // only case we are interested in handling is when only a single
1004 // edge is unknown (see setEdgeOrBlockWeight).
1005 for (unsigned i = 0; i < 2; i++) {
1006 uint64_t TotalWeight = 0;
1007 unsigned NumUnknownEdges = 0, NumTotalEdges = 0;
1008 Edge UnknownEdge, SelfReferentialEdge, SingleEdge;
1009
1010 if (i == 0) {
1011 // First, visit all predecessor edges.
1012 NumTotalEdges = Predecessors[BB].size();
1013 for (auto *Pred : Predecessors[BB]) {
1014 Edge E = std::make_pair(Pred, BB);
1015 TotalWeight += visitEdge(E, &NumUnknownEdges, &UnknownEdge);
1016 if (E.first == E.second)
1017 SelfReferentialEdge = E;
1018 }
1019 if (NumTotalEdges == 1) {
1020 SingleEdge = std::make_pair(Predecessors[BB][0], BB);
1021 }
1022 } else {
1023 // On the second round, visit all successor edges.
1024 NumTotalEdges = Successors[BB].size();
1025 for (auto *Succ : Successors[BB]) {
1026 Edge E = std::make_pair(BB, Succ);
1027 TotalWeight += visitEdge(E, &NumUnknownEdges, &UnknownEdge);
1028 }
1029 if (NumTotalEdges == 1) {
1030 SingleEdge = std::make_pair(BB, Successors[BB][0]);
1031 }
1032 }
1033
1034 // After visiting all the edges, there are three cases that we
1035 // can handle immediately:
1036 //
1037 // - All the edge weights are known (i.e., NumUnknownEdges == 0).
1038 // In this case, we simply check that the sum of all the edges
1039 // is the same as BB's weight. If not, we change BB's weight
1040 // to match. Additionally, if BB had not been visited before,
1041 // we mark it visited.
1042 //
1043 // - Only one edge is unknown and BB has already been visited.
1044 // In this case, we can compute the weight of the edge by
1045 // subtracting the total block weight from all the known
1046 // edge weights. If the edges weight more than BB, then the
1047 // edge of the last remaining edge is set to zero.
1048 //
1049 // - There exists a self-referential edge and the weight of BB is
1050 // known. In this case, this edge can be based on BB's weight.
1051 // We add up all the other known edges and set the weight on
1052 // the self-referential edge as we did in the previous case.
1053 //
1054 // In any other case, we must continue iterating. Eventually,
1055 // all edges will get a weight, or iteration will stop when
1056 // it reaches SampleProfileMaxPropagateIterations.
1057 if (NumUnknownEdges <= 1) {
1058 uint64_t &BBWeight = BlockWeights[EC];
1059 if (NumUnknownEdges == 0) {
1060 if (!VisitedBlocks.count(EC)) {
1061 // If we already know the weight of all edges, the weight of the
1062 // basic block can be computed. It should be no larger than the sum
1063 // of all edge weights.
1064 if (TotalWeight > BBWeight) {
1065 BBWeight = TotalWeight;
1066 Changed = true;
1067 DEBUG(dbgs() << "All edge weights for " << BB->getName()do { if (::llvm::DebugFlag && ::llvm::isCurrentDebugType
("sample-profile")) { dbgs() << "All edge weights for "
<< BB->getName() << " known. Set weight for block: "
; printBlockWeight(dbgs(), BB);; } } while (false)
1068 << " known. Set weight for block: ";do { if (::llvm::DebugFlag && ::llvm::isCurrentDebugType
("sample-profile")) { dbgs() << "All edge weights for "
<< BB->getName() << " known. Set weight for block: "
; printBlockWeight(dbgs(), BB);; } } while (false)
1069 printBlockWeight(dbgs(), BB);)do { if (::llvm::DebugFlag && ::llvm::isCurrentDebugType
("sample-profile")) { dbgs() << "All edge weights for "
<< BB->getName() << " known. Set weight for block: "
; printBlockWeight(dbgs(), BB);; } } while (false)
;
1070 }
1071 } else if (NumTotalEdges == 1 &&
1072 EdgeWeights[SingleEdge] < BlockWeights[EC]) {
1073 // If there is only one edge for the visited basic block, use the
1074 // block weight to adjust edge weight if edge weight is smaller.
1075 EdgeWeights[SingleEdge] = BlockWeights[EC];
1076 Changed = true;
1077 }
1078 } else if (NumUnknownEdges == 1 && VisitedBlocks.count(EC)) {
1079 // If there is a single unknown edge and the block has been
1080 // visited, then we can compute E's weight.
1081 if (BBWeight >= TotalWeight)
1082 EdgeWeights[UnknownEdge] = BBWeight - TotalWeight;
1083 else
1084 EdgeWeights[UnknownEdge] = 0;
1085 const BasicBlock *OtherEC;
1086 if (i == 0)
1087 OtherEC = EquivalenceClass[UnknownEdge.first];
1088 else
1089 OtherEC = EquivalenceClass[UnknownEdge.second];
1090 // Edge weights should never exceed the BB weights it connects.
1091 if (VisitedBlocks.count(OtherEC) &&
1092 EdgeWeights[UnknownEdge] > BlockWeights[OtherEC])
1093 EdgeWeights[UnknownEdge] = BlockWeights[OtherEC];
1094 VisitedEdges.insert(UnknownEdge);
1095 Changed = true;
1096 DEBUG(dbgs() << "Set weight for edge: ";do { if (::llvm::DebugFlag && ::llvm::isCurrentDebugType
("sample-profile")) { dbgs() << "Set weight for edge: "
; printEdgeWeight(dbgs(), UnknownEdge); } } while (false)
1097 printEdgeWeight(dbgs(), UnknownEdge))do { if (::llvm::DebugFlag && ::llvm::isCurrentDebugType
("sample-profile")) { dbgs() << "Set weight for edge: "
; printEdgeWeight(dbgs(), UnknownEdge); } } while (false)
;
1098 }
1099 } else if (VisitedBlocks.count(EC) && BlockWeights[EC] == 0) {
1100 // If a block Weights 0, all its in/out edges should weight 0.
1101 if (i == 0) {
1102 for (auto *Pred : Predecessors[BB]) {
1103 Edge E = std::make_pair(Pred, BB);
1104 EdgeWeights[E] = 0;
1105 VisitedEdges.insert(E);
1106 }
1107 } else {
1108 for (auto *Succ : Successors[BB]) {
1109 Edge E = std::make_pair(BB, Succ);
1110 EdgeWeights[E] = 0;
1111 VisitedEdges.insert(E);
1112 }
1113 }
1114 } else if (SelfReferentialEdge.first && VisitedBlocks.count(EC)) {
1115 uint64_t &BBWeight = BlockWeights[BB];
1116 // We have a self-referential edge and the weight of BB is known.
1117 if (BBWeight >= TotalWeight)
1118 EdgeWeights[SelfReferentialEdge] = BBWeight - TotalWeight;
1119 else
1120 EdgeWeights[SelfReferentialEdge] = 0;
1121 VisitedEdges.insert(SelfReferentialEdge);
1122 Changed = true;
1123 DEBUG(dbgs() << "Set self-referential edge weight to: ";do { if (::llvm::DebugFlag && ::llvm::isCurrentDebugType
("sample-profile")) { dbgs() << "Set self-referential edge weight to: "
; printEdgeWeight(dbgs(), SelfReferentialEdge); } } while (false
)
1124 printEdgeWeight(dbgs(), SelfReferentialEdge))do { if (::llvm::DebugFlag && ::llvm::isCurrentDebugType
("sample-profile")) { dbgs() << "Set self-referential edge weight to: "
; printEdgeWeight(dbgs(), SelfReferentialEdge); } } while (false
)
;
1125 }
1126 if (UpdateBlockCount && !VisitedBlocks.count(EC) && TotalWeight > 0) {
1127 BlockWeights[EC] = TotalWeight;
1128 VisitedBlocks.insert(EC);
1129 Changed = true;
1130 }
1131 }
1132 }
1133
1134 return Changed;
1135}
1136
1137/// \brief Build in/out edge lists for each basic block in the CFG.
1138///
1139/// We are interested in unique edges. If a block B1 has multiple
1140/// edges to another block B2, we only add a single B1->B2 edge.
1141void SampleProfileLoader::buildEdges(Function &F) {
1142 for (auto &BI : F) {
1143 BasicBlock *B1 = &BI;
1144
1145 // Add predecessors for B1.
1146 SmallPtrSet<BasicBlock *, 16> Visited;
1147 if (!Predecessors[B1].empty())
1148 llvm_unreachable("Found a stale predecessors list in a basic block.")::llvm::llvm_unreachable_internal("Found a stale predecessors list in a basic block."
, "/build/llvm-toolchain-snapshot-7~svn329677/lib/Transforms/IPO/SampleProfile.cpp"
, 1148)
;
1149 for (pred_iterator PI = pred_begin(B1), PE = pred_end(B1); PI != PE; ++PI) {
1150 BasicBlock *B2 = *PI;
1151 if (Visited.insert(B2).second)
1152 Predecessors[B1].push_back(B2);
1153 }
1154
1155 // Add successors for B1.
1156 Visited.clear();
1157 if (!Successors[B1].empty())
1158 llvm_unreachable("Found a stale successors list in a basic block.")::llvm::llvm_unreachable_internal("Found a stale successors list in a basic block."
, "/build/llvm-toolchain-snapshot-7~svn329677/lib/Transforms/IPO/SampleProfile.cpp"
, 1158)
;
1159 for (succ_iterator SI = succ_begin(B1), SE = succ_end(B1); SI != SE; ++SI) {
1160 BasicBlock *B2 = *SI;
1161 if (Visited.insert(B2).second)
1162 Successors[B1].push_back(B2);
1163 }
1164 }
1165}
1166
1167/// Returns the sorted CallTargetMap \p M by count in descending order.
1168static SmallVector<InstrProfValueData, 2> SortCallTargets(
1169 const SampleRecord::CallTargetMap &M) {
1170 SmallVector<InstrProfValueData, 2> R;
1171 for (auto I = M.begin(); I != M.end(); ++I)
1172 R.push_back({Function::getGUID(I->getKey()), I->getValue()});
1173 std::sort(R.begin(), R.end(),
1174 [](const InstrProfValueData &L, const InstrProfValueData &R) {
1175 if (L.Count == R.Count)
1176 return L.Value > R.Value;
1177 else
1178 return L.Count > R.Count;
1179 });
1180 return R;
1181}
1182
1183/// \brief Propagate weights into edges
1184///
1185/// The following rules are applied to every block BB in the CFG:
1186///
1187/// - If BB has a single predecessor/successor, then the weight
1188/// of that edge is the weight of the block.
1189///
1190/// - If all incoming or outgoing edges are known except one, and the
1191/// weight of the block is already known, the weight of the unknown
1192/// edge will be the weight of the block minus the sum of all the known
1193/// edges. If the sum of all the known edges is larger than BB's weight,
1194/// we set the unknown edge weight to zero.
1195///
1196/// - If there is a self-referential edge, and the weight of the block is
1197/// known, the weight for that edge is set to the weight of the block
1198/// minus the weight of the other incoming edges to that block (if
1199/// known).
1200void SampleProfileLoader::propagateWeights(Function &F) {
1201 bool Changed = true;
1202 unsigned I = 0;
1203
1204 // If BB weight is larger than its corresponding loop's header BB weight,
1205 // use the BB weight to replace the loop header BB weight.
1206 for (auto &BI : F) {
1207 BasicBlock *BB = &BI;
1208 Loop *L = LI->getLoopFor(BB);
1209 if (!L) {
1210 continue;
1211 }
1212 BasicBlock *Header = L->getHeader();
1213 if (Header && BlockWeights[BB] > BlockWeights[Header]) {
1214 BlockWeights[Header] = BlockWeights[BB];
1215 }
1216 }
1217
1218 // Before propagation starts, build, for each block, a list of
1219 // unique predecessors and successors. This is necessary to handle
1220 // identical edges in multiway branches. Since we visit all blocks and all
1221 // edges of the CFG, it is cleaner to build these lists once at the start
1222 // of the pass.
1223 buildEdges(F);
1224
1225 // Propagate until we converge or we go past the iteration limit.
1226 while (Changed && I++ < SampleProfileMaxPropagateIterations) {
1227 Changed = propagateThroughEdges(F, false);
1228 }
1229
1230 // The first propagation propagates BB counts from annotated BBs to unknown
1231 // BBs. The 2nd propagation pass resets edges weights, and use all BB weights
1232 // to propagate edge weights.
1233 VisitedEdges.clear();
1234 Changed = true;
1235 while (Changed && I++ < SampleProfileMaxPropagateIterations) {
1236 Changed = propagateThroughEdges(F, false);
1237 }
1238
1239 // The 3rd propagation pass allows adjust annotated BB weights that are
1240 // obviously wrong.
1241 Changed = true;
1242 while (Changed && I++ < SampleProfileMaxPropagateIterations) {
1243 Changed = propagateThroughEdges(F, true);
1244 }
1245
1246 // Generate MD_prof metadata for every branch instruction using the
1247 // edge weights computed during propagation.
1248 DEBUG(dbgs() << "\nPropagation complete. Setting branch weights\n")do { if (::llvm::DebugFlag && ::llvm::isCurrentDebugType
("sample-profile")) { dbgs() << "\nPropagation complete. Setting branch weights\n"
; } } while (false)
;
1249 LLVMContext &Ctx = F.getContext();
1250 MDBuilder MDB(Ctx);
1251 for (auto &BI : F) {
1252 BasicBlock *BB = &BI;
1253
1254 if (BlockWeights[BB]) {
1255 for (auto &I : BB->getInstList()) {
1256 if (!isa<CallInst>(I) && !isa<InvokeInst>(I))
1257 continue;
1258 CallSite CS(&I);
1259 if (!CS.getCalledFunction()) {
1260 const DebugLoc &DLoc = I.getDebugLoc();
1261 if (!DLoc)
1262 continue;
1263 const DILocation *DIL = DLoc;
1264 uint32_t LineOffset = FunctionSamples::getOffset(DIL);
1265 uint32_t Discriminator = DIL->getBaseDiscriminator();
1266
1267 const FunctionSamples *FS = findFunctionSamples(I);
1268 if (!FS)
1269 continue;
1270 auto T = FS->findCallTargetMapAt(LineOffset, Discriminator);
1271 if (!T || T.get().empty())
1272 continue;
1273 SmallVector<InstrProfValueData, 2> SortedCallTargets =
1274 SortCallTargets(T.get());
1275 uint64_t Sum;
1276 findIndirectCallFunctionSamples(I, Sum);
1277 annotateValueSite(*I.getParent()->getParent()->getParent(), I,
1278 SortedCallTargets, Sum, IPVK_IndirectCallTarget,
1279 SortedCallTargets.size());
1280 } else if (!dyn_cast<IntrinsicInst>(&I)) {
1281 SmallVector<uint32_t, 1> Weights;
1282 Weights.push_back(BlockWeights[BB]);
1283 I.setMetadata(LLVMContext::MD_prof, MDB.createBranchWeights(Weights));
1284 }
1285 }
1286 }
1287 TerminatorInst *TI = BB->getTerminator();
1288 if (TI->getNumSuccessors() == 1)
1289 continue;
1290 if (!isa<BranchInst>(TI) && !isa<SwitchInst>(TI))
1291 continue;
1292
1293 DebugLoc BranchLoc = TI->getDebugLoc();
1294 DEBUG(dbgs() << "\nGetting weights for branch at line "do { if (::llvm::DebugFlag && ::llvm::isCurrentDebugType
("sample-profile")) { dbgs() << "\nGetting weights for branch at line "
<< ((BranchLoc) ? Twine(BranchLoc.getLine()) : Twine("<UNKNOWN LOCATION>"
)) << ".\n"; } } while (false)
1295 << ((BranchLoc) ? Twine(BranchLoc.getLine())do { if (::llvm::DebugFlag && ::llvm::isCurrentDebugType
("sample-profile")) { dbgs() << "\nGetting weights for branch at line "
<< ((BranchLoc) ? Twine(BranchLoc.getLine()) : Twine("<UNKNOWN LOCATION>"
)) << ".\n"; } } while (false)
1296 : Twine("<UNKNOWN LOCATION>"))do { if (::llvm::DebugFlag && ::llvm::isCurrentDebugType
("sample-profile")) { dbgs() << "\nGetting weights for branch at line "
<< ((BranchLoc) ? Twine(BranchLoc.getLine()) : Twine("<UNKNOWN LOCATION>"
)) << ".\n"; } } while (false)
1297 << ".\n")do { if (::llvm::DebugFlag && ::llvm::isCurrentDebugType
("sample-profile")) { dbgs() << "\nGetting weights for branch at line "
<< ((BranchLoc) ? Twine(BranchLoc.getLine()) : Twine("<UNKNOWN LOCATION>"
)) << ".\n"; } } while (false)
;
1298 SmallVector<uint32_t, 4> Weights;
1299 uint32_t MaxWeight = 0;
1300 Instruction *MaxDestInst;
1301 for (unsigned I = 0; I < TI->getNumSuccessors(); ++I) {
1302 BasicBlock *Succ = TI->getSuccessor(I);
1303 Edge E = std::make_pair(BB, Succ);
1304 uint64_t Weight = EdgeWeights[E];
1305 DEBUG(dbgs() << "\t"; printEdgeWeight(dbgs(), E))do { if (::llvm::DebugFlag && ::llvm::isCurrentDebugType
("sample-profile")) { dbgs() << "\t"; printEdgeWeight(dbgs
(), E); } } while (false)
;
1306 // Use uint32_t saturated arithmetic to adjust the incoming weights,
1307 // if needed. Sample counts in profiles are 64-bit unsigned values,
1308 // but internally branch weights are expressed as 32-bit values.
1309 if (Weight > std::numeric_limits<uint32_t>::max()) {
1310 DEBUG(dbgs() << " (saturated due to uint32_t overflow)")do { if (::llvm::DebugFlag && ::llvm::isCurrentDebugType
("sample-profile")) { dbgs() << " (saturated due to uint32_t overflow)"
; } } while (false)
;
1311 Weight = std::numeric_limits<uint32_t>::max();
1312 }
1313 // Weight is added by one to avoid propagation errors introduced by
1314 // 0 weights.
1315 Weights.push_back(static_cast<uint32_t>(Weight + 1));
1316 if (Weight != 0) {
1317 if (Weight > MaxWeight) {
1318 MaxWeight = Weight;
1319 MaxDestInst = Succ->getFirstNonPHIOrDbgOrLifetime();
1320 }
1321 }
1322 }
1323
1324 uint64_t TempWeight;
1325 // Only set weights if there is at least one non-zero weight.
1326 // In any other case, let the analyzer set weights.
1327 // Do not set weights if the weights are present. In ThinLTO, the profile
1328 // annotation is done twice. If the first annotation already set the
1329 // weights, the second pass does not need to set it.
1330 if (MaxWeight > 0 && !TI->extractProfTotalWeight(TempWeight)) {
1331 DEBUG(dbgs() << "SUCCESS. Found non-zero weights.\n")do { if (::llvm::DebugFlag && ::llvm::isCurrentDebugType
("sample-profile")) { dbgs() << "SUCCESS. Found non-zero weights.\n"
; } } while (false)
;
1332 TI->setMetadata(LLVMContext::MD_prof,
1333 MDB.createBranchWeights(Weights));
1334 ORE->emit([&]() {
1335 return OptimizationRemark(DEBUG_TYPE"sample-profile", "PopularDest", MaxDestInst)
1336 << "most popular destination for conditional branches at "
1337 << ore::NV("CondBranchesLoc", BranchLoc);
1338 });
1339 } else {
1340 DEBUG(dbgs() << "SKIPPED. All branch weights are zero.\n")do { if (::llvm::DebugFlag && ::llvm::isCurrentDebugType
("sample-profile")) { dbgs() << "SKIPPED. All branch weights are zero.\n"
; } } while (false)
;
1341 }
1342 }
1343}
1344
1345/// \brief Get the line number for the function header.
1346///
1347/// This looks up function \p F in the current compilation unit and
1348/// retrieves the line number where the function is defined. This is
1349/// line 0 for all the samples read from the profile file. Every line
1350/// number is relative to this line.
1351///
1352/// \param F Function object to query.
1353///
1354/// \returns the line number where \p F is defined. If it returns 0,
1355/// it means that there is no debug information available for \p F.
1356unsigned SampleProfileLoader::getFunctionLoc(Function &F) {
1357 if (DISubprogram *S = F.getSubprogram())
1358 return S->getLine();
1359
1360 // If the start of \p F is missing, emit a diagnostic to inform the user
1361 // about the missed opportunity.
1362 F.getContext().diagnose(DiagnosticInfoSampleProfile(
1363 "No debug information found in function " + F.getName() +
1364 ": Function profile not used",
1365 DS_Warning));
1366 return 0;
1367}
1368
1369void SampleProfileLoader::computeDominanceAndLoopInfo(Function &F) {
1370 DT.reset(new DominatorTree);
1371 DT->recalculate(F);
1372
1373 PDT.reset(new PostDomTreeBase<BasicBlock>());
1374 PDT->recalculate(F);
1375
1376 LI.reset(new LoopInfo);
1377 LI->analyze(*DT);
1378}
1379
1380/// \brief Generate branch weight metadata for all branches in \p F.
1381///
1382/// Branch weights are computed out of instruction samples using a
1383/// propagation heuristic. Propagation proceeds in 3 phases:
1384///
1385/// 1- Assignment of block weights. All the basic blocks in the function
1386/// are initial assigned the same weight as their most frequently
1387/// executed instruction.
1388///
1389/// 2- Creation of equivalence classes. Since samples may be missing from
1390/// blocks, we can fill in the gaps by setting the weights of all the
1391/// blocks in the same equivalence class to the same weight. To compute
1392/// the concept of equivalence, we use dominance and loop information.
1393/// Two blocks B1 and B2 are in the same equivalence class if B1
1394/// dominates B2, B2 post-dominates B1 and both are in the same loop.
1395///
1396/// 3- Propagation of block weights into edges. This uses a simple
1397/// propagation heuristic. The following rules are applied to every
1398/// block BB in the CFG:
1399///
1400/// - If BB has a single predecessor/successor, then the weight
1401/// of that edge is the weight of the block.
1402///
1403/// - If all the edges are known except one, and the weight of the
1404/// block is already known, the weight of the unknown edge will
1405/// be the weight of the block minus the sum of all the known
1406/// edges. If the sum of all the known edges is larger than BB's weight,
1407/// we set the unknown edge weight to zero.
1408///
1409/// - If there is a self-referential edge, and the weight of the block is
1410/// known, the weight for that edge is set to the weight of the block
1411/// minus the weight of the other incoming edges to that block (if
1412/// known).
1413///
1414/// Since this propagation is not guaranteed to finalize for every CFG, we
1415/// only allow it to proceed for a limited number of iterations (controlled
1416/// by -sample-profile-max-propagate-iterations).
1417///
1418/// FIXME: Try to replace this propagation heuristic with a scheme
1419/// that is guaranteed to finalize. A work-list approach similar to
1420/// the standard value propagation algorithm used by SSA-CCP might
1421/// work here.
1422///
1423/// Once all the branch weights are computed, we emit the MD_prof
1424/// metadata on BB using the computed values for each of its branches.
1425///
1426/// \param F The function to query.
1427///
1428/// \returns true if \p F was modified. Returns false, otherwise.
1429bool SampleProfileLoader::emitAnnotations(Function &F) {
1430 bool Changed = false;
1431
1432 if (getFunctionLoc(F) == 0)
1433 return false;
1434
1435 DEBUG(dbgs() << "Line number for the first instruction in " << F.getName()do { if (::llvm::DebugFlag && ::llvm::isCurrentDebugType
("sample-profile")) { dbgs() << "Line number for the first instruction in "
<< F.getName() << ": " << getFunctionLoc(F
) << "\n"; } } while (false)
1436 << ": " << getFunctionLoc(F) << "\n")do { if (::llvm::DebugFlag && ::llvm::isCurrentDebugType
("sample-profile")) { dbgs() << "Line number for the first instruction in "
<< F.getName() << ": " << getFunctionLoc(F
) << "\n"; } } while (false)
;
1437
1438 DenseSet<GlobalValue::GUID> InlinedGUIDs;
1439 Changed |= inlineHotFunctions(F, InlinedGUIDs);
1440
1441 // Compute basic block weights.
1442 Changed |= computeBlockWeights(F);
1443
1444 if (Changed) {
1445 // Add an entry count to the function using the samples gathered at the
1446 // function entry.
1447 // Sets the GUIDs that are inlined in the profiled binary. This is used
1448 // for ThinLink to make correct liveness analysis, and also make the IR
1449 // match the profiled binary before annotation.
1450 F.setEntryCount(
1451 ProfileCount(Samples->getHeadSamples() + 1, Function::PCT_Real),
1452 &InlinedGUIDs);
1453
1454 // Compute dominance and loop info needed for propagation.
1455 computeDominanceAndLoopInfo(F);
1456
1457 // Find equivalence classes.
1458 findEquivalenceClasses(F);
1459
1460 // Propagate weights to all edges.
1461 propagateWeights(F);
1462 }
1463
1464 // If coverage checking was requested, compute it now.
1465 if (SampleProfileRecordCoverage) {
1466 unsigned Used = CoverageTracker.countUsedRecords(Samples);
1467 unsigned Total = CoverageTracker.countBodyRecords(Samples);
1468 unsigned Coverage = CoverageTracker.computeCoverage(Used, Total);
1469 if (Coverage < SampleProfileRecordCoverage) {
1470 F.getContext().diagnose(DiagnosticInfoSampleProfile(
1471 F.getSubprogram()->getFilename(), getFunctionLoc(F),
1472 Twine(Used) + " of " + Twine(Total) + " available profile records (" +
1473 Twine(Coverage) + "%) were applied",
1474 DS_Warning));
1475 }
1476 }
1477
1478 if (SampleProfileSampleCoverage) {
1479 uint64_t Used = CoverageTracker.getTotalUsedSamples();
1480 uint64_t Total = CoverageTracker.countBodySamples(Samples);
1481 unsigned Coverage = CoverageTracker.computeCoverage(Used, Total);
1482 if (Coverage < SampleProfileSampleCoverage) {
1483 F.getContext().diagnose(DiagnosticInfoSampleProfile(
1484 F.getSubprogram()->getFilename(), getFunctionLoc(F),
1485 Twine(Used) + " of " + Twine(Total) + " available profile samples (" +
1486 Twine(Coverage) + "%) were applied",
1487 DS_Warning));
1488 }
1489 }
1490 return Changed;
1491}
1492
1493char SampleProfileLoaderLegacyPass::ID = 0;
1494
1495INITIALIZE_PASS_BEGIN(SampleProfileLoaderLegacyPass, "sample-profile",static void *initializeSampleProfileLoaderLegacyPassPassOnce(
PassRegistry &Registry) {
1496 "Sample Profile loader", false, false)static void *initializeSampleProfileLoaderLegacyPassPassOnce(
PassRegistry &Registry) {
1497INITIALIZE_PASS_DEPENDENCY(AssumptionCacheTracker)initializeAssumptionCacheTrackerPass(Registry);
1498INITIALIZE_PASS_DEPENDENCY(TargetTransformInfoWrapperPass)initializeTargetTransformInfoWrapperPassPass(Registry);
1499INITIALIZE_PASS_END(SampleProfileLoaderLegacyPass, "sample-profile",PassInfo *PI = new PassInfo( "Sample Profile loader", "sample-profile"
, &SampleProfileLoaderLegacyPass::ID, PassInfo::NormalCtor_t
(callDefaultCtor<SampleProfileLoaderLegacyPass>), false
, false); Registry.registerPass(*PI, true); return PI; } static
llvm::once_flag InitializeSampleProfileLoaderLegacyPassPassFlag
; void llvm::initializeSampleProfileLoaderLegacyPassPass(PassRegistry
&Registry) { llvm::call_once(InitializeSampleProfileLoaderLegacyPassPassFlag
, initializeSampleProfileLoaderLegacyPassPassOnce, std::ref(Registry
)); }
1500 "Sample Profile loader", false, false)PassInfo *PI = new PassInfo( "Sample Profile loader", "sample-profile"
, &SampleProfileLoaderLegacyPass::ID, PassInfo::NormalCtor_t
(callDefaultCtor<SampleProfileLoaderLegacyPass>), false
, false); Registry.registerPass(*PI, true); return PI; } static
llvm::once_flag InitializeSampleProfileLoaderLegacyPassPassFlag
; void llvm::initializeSampleProfileLoaderLegacyPassPass(PassRegistry
&Registry) { llvm::call_once(InitializeSampleProfileLoaderLegacyPassPassFlag
, initializeSampleProfileLoaderLegacyPassPassOnce, std::ref(Registry
)); }
1501
1502bool SampleProfileLoader::doInitialization(Module &M) {
1503 auto &Ctx = M.getContext();
1504 auto ReaderOrErr = SampleProfileReader::create(Filename, Ctx);
1505 if (std::error_code EC = ReaderOrErr.getError()) {
1506 std::string Msg = "Could not open profile: " + EC.message();
1507 Ctx.diagnose(DiagnosticInfoSampleProfile(Filename, Msg));
1508 return false;
1509 }
1510 Reader = std::move(ReaderOrErr.get());
1511 ProfileIsValid = (Reader->read() == sampleprof_error::success);
1512 return true;
1513}
1514
1515ModulePass *llvm::createSampleProfileLoaderPass() {
1516 return new SampleProfileLoaderLegacyPass(SampleProfileFile);
1517}
1518
1519ModulePass *llvm::createSampleProfileLoaderPass(StringRef Name) {
1520 return new SampleProfileLoaderLegacyPass(Name);
1521}
1522
1523bool SampleProfileLoader::runOnModule(Module &M, ModuleAnalysisManager *AM) {
1524 if (!ProfileIsValid)
1525 return false;
1526
1527 // Compute the total number of samples collected in this profile.
1528 for (const auto &I : Reader->getProfiles())
1529 TotalCollectedSamples += I.second.getTotalSamples();
1530
1531 // Populate the symbol map.
1532 for (const auto &N_F : M.getValueSymbolTable()) {
1533 StringRef OrigName = N_F.getKey();
1534 Function *F = dyn_cast<Function>(N_F.getValue());
1535 if (F == nullptr)
1536 continue;
1537 SymbolMap[OrigName] = F;
1538 auto pos = OrigName.find('.');
1539 if (pos != StringRef::npos) {
1540 StringRef NewName = OrigName.substr(0, pos);
1541 auto r = SymbolMap.insert(std::make_pair(NewName, F));
1542 // Failiing to insert means there is already an entry in SymbolMap,
1543 // thus there are multiple functions that are mapped to the same
1544 // stripped name. In this case of name conflicting, set the value
1545 // to nullptr to avoid confusion.
1546 if (!r.second)
1547 r.first->second = nullptr;
1548 }
1549 }
1550
1551 bool retval = false;
1552 for (auto &F : M)
1553 if (!F.isDeclaration()) {
1554 clearFunctionData();
1555 retval |= runOnFunction(F, AM);
1556 }
1557 if (M.getProfileSummary() == nullptr)
1558 M.setProfileSummary(Reader->getSummary().getMD(M.getContext()));
1559 return retval;
1560}
1561
1562bool SampleProfileLoaderLegacyPass::runOnModule(Module &M) {
1563 ACT = &getAnalysis<AssumptionCacheTracker>();
1564 TTIWP = &getAnalysis<TargetTransformInfoWrapperPass>();
1565 return SampleLoader.runOnModule(M, nullptr);
1566}
1567
1568bool SampleProfileLoader::runOnFunction(Function &F, ModuleAnalysisManager *AM) {
1569 // Initialize the entry count to -1, which will be treated conservatively
1570 // by getEntryCount as the same as unknown (None). If we have samples this
1571 // will be overwritten in emitAnnotations.
1572 F.setEntryCount(ProfileCount(-1, Function::PCT_Real));
1573 std::unique_ptr<OptimizationRemarkEmitter> OwnedORE;
1574 if (AM) {
1575 auto &FAM =
1576 AM->getResult<FunctionAnalysisManagerModuleProxy>(*F.getParent())
1577 .getManager();
1578 ORE = &FAM.getResult<OptimizationRemarkEmitterAnalysis>(F);
1579 } else {
1580 OwnedORE = make_unique<OptimizationRemarkEmitter>(&F);
1581 ORE = OwnedORE.get();
1582 }
1583 Samples = Reader->getSamplesFor(F);
1584 if (Samples && !Samples->empty())
1585 return emitAnnotations(F);
1586 return false;
1587}
1588
1589PreservedAnalyses SampleProfileLoaderPass::run(Module &M,
1590 ModuleAnalysisManager &AM) {
1591 FunctionAnalysisManager &FAM =
1592 AM.getResult<FunctionAnalysisManagerModuleProxy>(M).getManager();
1593
1594 auto GetAssumptionCache = [&](Function &F) -> AssumptionCache & {
1595 return FAM.getResult<AssumptionAnalysis>(F);
1596 };
1597 auto GetTTI = [&](Function &F) -> TargetTransformInfo & {
1598 return FAM.getResult<TargetIRAnalysis>(F);
1599 };
1600
1601 SampleProfileLoader SampleLoader(
1602 ProfileFileName.empty() ? SampleProfileFile : ProfileFileName,
1603 IsThinLTOPreLink, GetAssumptionCache, GetTTI);
1604
1605 SampleLoader.doInitialization(M);
1606
1607 if (!SampleLoader.runOnModule(M, &AM))
1608 return PreservedAnalyses::all();
1609
1610 return PreservedAnalyses::none();
1611}