1//===- SampleProfile.cpp - Incorporate sample profiles into the IR --------===//
2//
3// Part of the LLVM Project, under the Apache License v2.0 with LLVM Exceptions.
4// See https://llvm.org/LICENSE.txt for license information.
5// SPDX-License-Identifier: Apache-2.0 WITH LLVM-exception
6//
7//===----------------------------------------------------------------------===//
8//
9// This file implements the SampleProfileLoader transformation. This pass
10// reads a profile file generated by a sampling profiler (e.g. Linux Perf -
11// http://perf.wiki.kernel.org/) and generates IR metadata to reflect the
12// profile information in the given profile.
13//
14// This pass generates branch weight annotations on the IR:
15//
16// - prof: Represents branch weights. This annotation is added to branches
17// to indicate the weights of each edge coming out of the branch.
18// The weight of each edge is the weight of the target block for
19// that edge. The weight of a block B is computed as the maximum
20// number of samples found in B.
21//
22//===----------------------------------------------------------------------===//
23
24#include "llvm/Transforms/IPO/SampleProfile.h"
25#include "llvm/ADT/ArrayRef.h"
26#include "llvm/ADT/DenseMap.h"
27#include "llvm/ADT/DenseSet.h"
28#include "llvm/ADT/MapVector.h"
29#include "llvm/ADT/PriorityQueue.h"
30#include "llvm/ADT/SCCIterator.h"
31#include "llvm/ADT/SmallVector.h"
32#include "llvm/ADT/Statistic.h"
33#include "llvm/ADT/StringRef.h"
34#include "llvm/ADT/Twine.h"
35#include "llvm/Analysis/AssumptionCache.h"
36#include "llvm/Analysis/BlockFrequencyInfoImpl.h"
37#include "llvm/Analysis/InlineAdvisor.h"
38#include "llvm/Analysis/InlineCost.h"
39#include "llvm/Analysis/LazyCallGraph.h"
40#include "llvm/Analysis/OptimizationRemarkEmitter.h"
41#include "llvm/Analysis/ProfileSummaryInfo.h"
42#include "llvm/Analysis/ReplayInlineAdvisor.h"
43#include "llvm/Analysis/TargetLibraryInfo.h"
44#include "llvm/Analysis/TargetTransformInfo.h"
45#include "llvm/IR/BasicBlock.h"
46#include "llvm/IR/DebugLoc.h"
47#include "llvm/IR/DiagnosticInfo.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/ProfDataUtils.h"
59#include "llvm/IR/PseudoProbe.h"
60#include "llvm/IR/ValueSymbolTable.h"
61#include "llvm/ProfileData/InstrProf.h"
62#include "llvm/ProfileData/SampleProf.h"
63#include "llvm/ProfileData/SampleProfReader.h"
64#include "llvm/Support/Casting.h"
65#include "llvm/Support/CommandLine.h"
66#include "llvm/Support/Debug.h"
67#include "llvm/Support/ErrorOr.h"
68#include "llvm/Support/VirtualFileSystem.h"
69#include "llvm/Support/raw_ostream.h"
70#include "llvm/Transforms/IPO.h"
71#include "llvm/Transforms/IPO/ProfiledCallGraph.h"
72#include "llvm/Transforms/IPO/SampleContextTracker.h"
73#include "llvm/Transforms/IPO/SampleProfileMatcher.h"
74#include "llvm/Transforms/IPO/SampleProfileProbe.h"
75#include "llvm/Transforms/Utils/CallPromotionUtils.h"
76#include "llvm/Transforms/Utils/Cloning.h"
77#include "llvm/Transforms/Utils/Instrumentation.h"
78#include "llvm/Transforms/Utils/MisExpect.h"
79#include "llvm/Transforms/Utils/SampleProfileLoaderBaseImpl.h"
80#include "llvm/Transforms/Utils/SampleProfileLoaderBaseUtil.h"
81#include <algorithm>
82#include <cassert>
83#include <cstdint>
84#include <functional>
85#include <limits>
86#include <memory>
87#include <queue>
88#include <string>
89#include <system_error>
90#include <utility>
91#include <vector>
92
93using namespace llvm;
94using namespace sampleprof;
95using namespace llvm::sampleprofutil;
96#define DEBUG_TYPE "sample-profile"
97#define CSINLINE_DEBUG DEBUG_TYPE "-inline"
98
99STATISTIC(NumCSInlined,
100 "Number of functions inlined with context sensitive profile");
101STATISTIC(NumCSNotInlined,
102 "Number of functions not inlined with context sensitive profile");
103STATISTIC(NumMismatchedProfile,
104 "Number of functions with CFG mismatched profile");
105STATISTIC(NumMatchedProfile, "Number of functions with CFG matched profile");
106STATISTIC(NumDuplicatedInlinesite,
107 "Number of inlined callsites with a partial distribution factor");
108
109STATISTIC(NumCSInlinedHitMinLimit,
110 "Number of functions with FDO inline stopped due to min size limit");
111STATISTIC(NumCSInlinedHitMaxLimit,
112 "Number of functions with FDO inline stopped due to max size limit");
113STATISTIC(
114 NumCSInlinedHitGrowthLimit,
115 "Number of functions with FDO inline stopped due to growth size limit");
116
117namespace llvm {
118
119// Command line option to specify the file to read samples from. This is
120// mainly used for debugging.
121cl::opt<std::string> SampleProfileFile(
122 "sample-profile-file", cl::init(Val: ""), cl::value_desc("filename"),
123 cl::desc("Profile file loaded by -sample-profile"), cl::Hidden);
124
125// The named file contains a set of transformations that may have been applied
126// to the symbol names between the program from which the sample data was
127// collected and the current program's symbols.
128static cl::opt<std::string> SampleProfileRemappingFile(
129 "sample-profile-remapping-file", cl::init(Val: ""), cl::value_desc("filename"),
130 cl::desc("Profile remapping file loaded by -sample-profile"), cl::Hidden);
131
132cl::opt<bool> SalvageStaleProfile(
133 "salvage-stale-profile", cl::Hidden, cl::init(Val: false),
134 cl::desc("Salvage stale profile by fuzzy matching and use the remapped "
135 "location for sample profile query."));
136cl::opt<bool>
137 SalvageUnusedProfile("salvage-unused-profile", cl::Hidden, cl::init(Val: false),
138 cl::desc("Salvage unused profile by matching with new "
139 "functions on call graph."));
140
141cl::opt<bool> ReportProfileStaleness(
142 "report-profile-staleness", cl::Hidden, cl::init(Val: false),
143 cl::desc("Compute and report stale profile statistical metrics."));
144
145cl::opt<bool> PersistProfileStaleness(
146 "persist-profile-staleness", cl::Hidden, cl::init(Val: false),
147 cl::desc("Compute stale profile statistical metrics and write it into the "
148 "native object file(.llvm_stats section)."));
149
150static cl::opt<bool> ProfileSampleAccurate(
151 "profile-sample-accurate", cl::Hidden, cl::init(Val: false),
152 cl::desc("If the sample profile is accurate, we will mark all un-sampled "
153 "callsite and function as having 0 samples. Otherwise, treat "
154 "un-sampled callsites and functions conservatively as unknown. "));
155
156static cl::opt<bool> ProfileSampleBlockAccurate(
157 "profile-sample-block-accurate", cl::Hidden, cl::init(Val: false),
158 cl::desc("If the sample profile is accurate, we will mark all un-sampled "
159 "branches and calls as having 0 samples. Otherwise, treat "
160 "them conservatively as unknown. "));
161
162static cl::opt<bool> ProfileAccurateForSymsInList(
163 "profile-accurate-for-symsinlist", cl::Hidden, cl::init(Val: true),
164 cl::desc("For symbols in profile symbol list, regard their profiles to "
165 "be accurate. It may be overridden by profile-sample-accurate. "));
166
167static cl::opt<bool> ProfileMergeInlinee(
168 "sample-profile-merge-inlinee", cl::Hidden, cl::init(Val: true),
169 cl::desc("Merge past inlinee's profile to outline version if sample "
170 "profile loader decided not to inline a call site. It will "
171 "only be enabled when top-down order of profile loading is "
172 "enabled. "));
173
174static cl::opt<bool> ProfileTopDownLoad(
175 "sample-profile-top-down-load", cl::Hidden, cl::init(Val: true),
176 cl::desc("Do profile annotation and inlining for functions in top-down "
177 "order of call graph during sample profile loading. It only "
178 "works for new pass manager. "));
179
180static cl::opt<bool>
181 UseProfiledCallGraph("use-profiled-call-graph", cl::init(Val: true), cl::Hidden,
182 cl::desc("Process functions in a top-down order "
183 "defined by the profiled call graph when "
184 "-sample-profile-top-down-load is on."));
185
186static cl::opt<bool> ProfileSizeInline(
187 "sample-profile-inline-size", cl::Hidden, cl::init(Val: false),
188 cl::desc("Inline cold call sites in profile loader if it's beneficial "
189 "for code size."));
190
191// Since profiles are consumed by many passes, turning on this option has
192// side effects. For instance, pre-link SCC inliner would see merged profiles
193// and inline the hot functions (that are skipped in this pass).
194static cl::opt<bool> DisableSampleLoaderInlining(
195 "disable-sample-loader-inlining", cl::Hidden, cl::init(Val: false),
196 cl::desc(
197 "If true, artificially skip inline transformation in sample-loader "
198 "pass, and merge (or scale) profiles (as configured by "
199 "--sample-profile-merge-inlinee)."));
200
201cl::opt<bool>
202 SortProfiledSCC("sort-profiled-scc-member", cl::init(Val: true), cl::Hidden,
203 cl::desc("Sort profiled recursion by edge weights."));
204
205cl::opt<int> ProfileInlineGrowthLimit(
206 "sample-profile-inline-growth-limit", cl::Hidden, cl::init(Val: 12),
207 cl::desc("The size growth ratio limit for proirity-based sample profile "
208 "loader inlining."));
209
210cl::opt<int> ProfileInlineLimitMin(
211 "sample-profile-inline-limit-min", cl::Hidden, cl::init(Val: 100),
212 cl::desc("The lower bound of size growth limit for "
213 "proirity-based sample profile loader inlining."));
214
215cl::opt<int> ProfileInlineLimitMax(
216 "sample-profile-inline-limit-max", cl::Hidden, cl::init(Val: 10000),
217 cl::desc("The upper bound of size growth limit for "
218 "proirity-based sample profile loader inlining."));
219
220cl::opt<int> SampleHotCallSiteThreshold(
221 "sample-profile-hot-inline-threshold", cl::Hidden, cl::init(Val: 3000),
222 cl::desc("Hot callsite threshold for proirity-based sample profile loader "
223 "inlining."));
224
225cl::opt<int> SampleColdCallSiteThreshold(
226 "sample-profile-cold-inline-threshold", cl::Hidden, cl::init(Val: 45),
227 cl::desc("Threshold for inlining cold callsites"));
228} // namespace llvm
229
230static cl::opt<unsigned> ProfileICPRelativeHotness(
231 "sample-profile-icp-relative-hotness", cl::Hidden, cl::init(Val: 25),
232 cl::desc(
233 "Relative hotness percentage threshold for indirect "
234 "call promotion in proirity-based sample profile loader inlining."));
235
236static cl::opt<unsigned> ProfileICPRelativeHotnessSkip(
237 "sample-profile-icp-relative-hotness-skip", cl::Hidden, cl::init(Val: 1),
238 cl::desc(
239 "Skip relative hotness check for ICP up to given number of targets."));
240
241static cl::opt<unsigned> HotFuncCutoffForStalenessError(
242 "hot-func-cutoff-for-staleness-error", cl::Hidden, cl::init(Val: 800000),
243 cl::desc("A function is considered hot for staleness error check if its "
244 "total sample count is above the specified percentile"));
245
246static cl::opt<unsigned> MinfuncsForStalenessError(
247 "min-functions-for-staleness-error", cl::Hidden, cl::init(Val: 50),
248 cl::desc("Skip the check if the number of hot functions is smaller than "
249 "the specified number."));
250
251static cl::opt<unsigned> PrecentMismatchForStalenessError(
252 "precent-mismatch-for-staleness-error", cl::Hidden, cl::init(Val: 80),
253 cl::desc("Reject the profile if the mismatch percent is higher than the "
254 "given number."));
255
256static cl::opt<bool> CallsitePrioritizedInline(
257 "sample-profile-prioritized-inline", cl::Hidden,
258 cl::desc("Use call site prioritized inlining for sample profile loader. "
259 "Currently only CSSPGO is supported."));
260
261static cl::opt<bool> UsePreInlinerDecision(
262 "sample-profile-use-preinliner", cl::Hidden,
263 cl::desc("Use the preinliner decisions stored in profile context."));
264
265static cl::opt<bool> AllowRecursiveInline(
266 "sample-profile-recursive-inline", cl::Hidden,
267 cl::desc("Allow sample loader inliner to inline recursive calls."));
268
269static cl::opt<bool> RemoveProbeAfterProfileAnnotation(
270 "sample-profile-remove-probe", cl::Hidden, cl::init(Val: false),
271 cl::desc("Remove pseudo-probe after sample profile annotation."));
272
273static cl::opt<std::string> ProfileInlineReplayFile(
274 "sample-profile-inline-replay", cl::init(Val: ""), cl::value_desc("filename"),
275 cl::desc(
276 "Optimization remarks file containing inline remarks to be replayed "
277 "by inlining from sample profile loader."),
278 cl::Hidden);
279
280static cl::opt<ReplayInlinerSettings::Scope> ProfileInlineReplayScope(
281 "sample-profile-inline-replay-scope",
282 cl::init(Val: ReplayInlinerSettings::Scope::Function),
283 cl::values(clEnumValN(ReplayInlinerSettings::Scope::Function, "Function",
284 "Replay on functions that have remarks associated "
285 "with them (default)"),
286 clEnumValN(ReplayInlinerSettings::Scope::Module, "Module",
287 "Replay on the entire module")),
288 cl::desc("Whether inline replay should be applied to the entire "
289 "Module or just the Functions (default) that are present as "
290 "callers in remarks during sample profile inlining."),
291 cl::Hidden);
292
293static cl::opt<ReplayInlinerSettings::Fallback> ProfileInlineReplayFallback(
294 "sample-profile-inline-replay-fallback",
295 cl::init(Val: ReplayInlinerSettings::Fallback::Original),
296 cl::values(
297 clEnumValN(
298 ReplayInlinerSettings::Fallback::Original, "Original",
299 "All decisions not in replay send to original advisor (default)"),
300 clEnumValN(ReplayInlinerSettings::Fallback::AlwaysInline,
301 "AlwaysInline", "All decisions not in replay are inlined"),
302 clEnumValN(ReplayInlinerSettings::Fallback::NeverInline, "NeverInline",
303 "All decisions not in replay are not inlined")),
304 cl::desc("How sample profile inline replay treats sites that don't come "
305 "from the replay. Original: defers to original advisor, "
306 "AlwaysInline: inline all sites not in replay, NeverInline: "
307 "inline no sites not in replay"),
308 cl::Hidden);
309
310static cl::opt<CallSiteFormat::Format> ProfileInlineReplayFormat(
311 "sample-profile-inline-replay-format",
312 cl::init(Val: CallSiteFormat::Format::LineColumnDiscriminator),
313 cl::values(
314 clEnumValN(CallSiteFormat::Format::Line, "Line", "<Line Number>"),
315 clEnumValN(CallSiteFormat::Format::LineColumn, "LineColumn",
316 "<Line Number>:<Column Number>"),
317 clEnumValN(CallSiteFormat::Format::LineDiscriminator,
318 "LineDiscriminator", "<Line Number>.<Discriminator>"),
319 clEnumValN(CallSiteFormat::Format::LineColumnDiscriminator,
320 "LineColumnDiscriminator",
321 "<Line Number>:<Column Number>.<Discriminator> (default)")),
322 cl::desc("How sample profile inline replay file is formatted"), cl::Hidden);
323
324static cl::opt<unsigned>
325 MaxNumPromotions("sample-profile-icp-max-prom", cl::init(Val: 3), cl::Hidden,
326 cl::desc("Max number of promotions for a single indirect "
327 "call callsite in sample profile loader"));
328
329static cl::opt<bool> OverwriteExistingWeights(
330 "overwrite-existing-weights", cl::Hidden, cl::init(Val: false),
331 cl::desc("Ignore existing branch weights on IR and always overwrite."));
332
333static cl::opt<bool> AnnotateSampleProfileInlinePhase(
334 "annotate-sample-profile-inline-phase", cl::Hidden, cl::init(Val: false),
335 cl::desc("Annotate LTO phase (prelink / postlink), or main (no LTO) for "
336 "sample-profile inline pass name."));
337
338namespace llvm {
339extern cl::opt<bool> EnableExtTspBlockPlacement;
340}
341
342namespace {
343
344using BlockWeightMap = DenseMap<const BasicBlock *, uint64_t>;
345using EquivalenceClassMap = DenseMap<const BasicBlock *, const BasicBlock *>;
346using Edge = std::pair<const BasicBlock *, const BasicBlock *>;
347using EdgeWeightMap = DenseMap<Edge, uint64_t>;
348using BlockEdgeMap =
349 DenseMap<const BasicBlock *, SmallVector<const BasicBlock *, 8>>;
350
351class GUIDToFuncNameMapper {
352public:
353 GUIDToFuncNameMapper(Module &M, SampleProfileReader &Reader,
354 DenseMap<uint64_t, StringRef> &GUIDToFuncNameMap)
355 : CurrentReader(Reader), CurrentModule(M),
356 CurrentGUIDToFuncNameMap(GUIDToFuncNameMap) {
357 if (!CurrentReader.useMD5())
358 return;
359
360 for (const auto &F : CurrentModule) {
361 StringRef OrigName = F.getName();
362 CurrentGUIDToFuncNameMap.insert(
363 KV: {Function::getGUIDAssumingExternalLinkage(GlobalName: OrigName), OrigName});
364
365 // Local to global var promotion used by optimization like thinlto
366 // will rename the var and add suffix like ".llvm.xxx" to the
367 // original local name. In sample profile, the suffixes of function
368 // names are all stripped. Since it is possible that the mapper is
369 // built in post-thin-link phase and var promotion has been done,
370 // we need to add the substring of function name without the suffix
371 // into the GUIDToFuncNameMap.
372 StringRef CanonName = FunctionSamples::getCanonicalFnName(F);
373 if (CanonName != OrigName)
374 CurrentGUIDToFuncNameMap.insert(
375 KV: {Function::getGUIDAssumingExternalLinkage(GlobalName: CanonName), CanonName});
376 }
377
378 // Update GUIDToFuncNameMap for each function including inlinees.
379 SetGUIDToFuncNameMapForAll(&CurrentGUIDToFuncNameMap);
380 }
381
382 ~GUIDToFuncNameMapper() {
383 if (!CurrentReader.useMD5())
384 return;
385
386 CurrentGUIDToFuncNameMap.clear();
387
388 // Reset GUIDToFuncNameMap for of each function as they're no
389 // longer valid at this point.
390 SetGUIDToFuncNameMapForAll(nullptr);
391 }
392
393private:
394 void SetGUIDToFuncNameMapForAll(DenseMap<uint64_t, StringRef> *Map) {
395 std::queue<FunctionSamples *> FSToUpdate;
396 for (auto &IFS : CurrentReader.getProfiles()) {
397 FSToUpdate.push(x: &IFS.second);
398 }
399
400 while (!FSToUpdate.empty()) {
401 FunctionSamples *FS = FSToUpdate.front();
402 FSToUpdate.pop();
403 FS->GUIDToFuncNameMap = Map;
404 for (const auto &ICS : FS->getCallsiteSamples()) {
405 const FunctionSamplesMap &FSMap = ICS.second;
406 for (const auto &IFS : FSMap) {
407 FunctionSamples &FS = const_cast<FunctionSamples &>(IFS.second);
408 FSToUpdate.push(x: &FS);
409 }
410 }
411 }
412 }
413
414 SampleProfileReader &CurrentReader;
415 Module &CurrentModule;
416 DenseMap<uint64_t, StringRef> &CurrentGUIDToFuncNameMap;
417};
418
419// Inline candidate used by iterative callsite prioritized inliner
420struct InlineCandidate {
421 CallBase *CallInstr;
422 const FunctionSamples *CalleeSamples;
423 // Prorated callsite count, which will be used to guide inlining. For example,
424 // if a callsite is duplicated in LTO prelink, then in LTO postlink the two
425 // copies will get their own distribution factors and their prorated counts
426 // will be used to decide if they should be inlined independently.
427 uint64_t CallsiteCount;
428 // Call site distribution factor to prorate the profile samples for a
429 // duplicated callsite. Default value is 1.0.
430 float CallsiteDistribution;
431};
432
433// Inline candidate comparer using call site weight
434struct CandidateComparer {
435 bool operator()(const InlineCandidate &LHS, const InlineCandidate &RHS) {
436 if (LHS.CallsiteCount != RHS.CallsiteCount)
437 return LHS.CallsiteCount < RHS.CallsiteCount;
438
439 const FunctionSamples *LCS = LHS.CalleeSamples;
440 const FunctionSamples *RCS = RHS.CalleeSamples;
441 // In inline replay mode, CalleeSamples may be null and the order doesn't
442 // matter.
443 if (!LCS || !RCS)
444 return LCS;
445
446 // Tie breaker using number of samples try to favor smaller functions first
447 if (LCS->getBodySamples().size() != RCS->getBodySamples().size())
448 return LCS->getBodySamples().size() > RCS->getBodySamples().size();
449
450 // Tie breaker using GUID so we have stable/deterministic inlining order
451 return LCS->getGUID() < RCS->getGUID();
452 }
453};
454
455using CandidateQueue =
456 PriorityQueue<InlineCandidate, std::vector<InlineCandidate>,
457 CandidateComparer>;
458
459/// Sample profile pass.
460///
461/// This pass reads profile data from the file specified by
462/// -sample-profile-file and annotates every affected function with the
463/// profile information found in that file.
464class SampleProfileLoader final : public SampleProfileLoaderBaseImpl<Function> {
465public:
466 SampleProfileLoader(
467 StringRef Name, StringRef RemapName, ThinOrFullLTOPhase LTOPhase,
468 IntrusiveRefCntPtr<vfs::FileSystem> FS,
469 std::function<AssumptionCache &(Function &)> GetAssumptionCache,
470 std::function<TargetTransformInfo &(Function &)> GetTargetTransformInfo,
471 std::function<const TargetLibraryInfo &(Function &)> GetTLI,
472 LazyCallGraph &CG, bool DisableSampleProfileInlining,
473 bool UseFlattenedProfile)
474 : SampleProfileLoaderBaseImpl(std::string(Name), std::string(RemapName),
475 std::move(FS)),
476 GetAC(std::move(GetAssumptionCache)),
477 GetTTI(std::move(GetTargetTransformInfo)), GetTLI(std::move(GetTLI)),
478 CG(CG), LTOPhase(LTOPhase),
479 AnnotatedPassName(AnnotateSampleProfileInlinePhase
480 ? llvm::AnnotateInlinePassName(IC: InlineContext{
481 .LTOPhase: LTOPhase, .Pass: InlinePass::SampleProfileInliner})
482 : CSINLINE_DEBUG),
483 DisableSampleProfileInlining(DisableSampleProfileInlining),
484 UseFlattenedProfile(UseFlattenedProfile) {}
485
486 bool doInitialization(Module &M, FunctionAnalysisManager *FAM = nullptr);
487 bool runOnModule(Module &M, ModuleAnalysisManager &AM,
488 ProfileSummaryInfo *_PSI);
489
490protected:
491 bool runOnFunction(Function &F, ModuleAnalysisManager &AM);
492 bool emitAnnotations(Function &F);
493 ErrorOr<uint64_t> getInstWeight(const Instruction &I) override;
494 const FunctionSamples *findCalleeFunctionSamples(const CallBase &I) const;
495 const FunctionSamples *
496 findFunctionSamples(const Instruction &I) const override;
497 std::vector<const FunctionSamples *>
498 findIndirectCallFunctionSamples(const Instruction &I, uint64_t &Sum) const;
499 void findExternalInlineCandidate(CallBase *CB, const FunctionSamples *Samples,
500 DenseSet<GlobalValue::GUID> &InlinedGUIDs,
501 uint64_t Threshold);
502 // Attempt to promote indirect call and also inline the promoted call
503 bool tryPromoteAndInlineCandidate(
504 Function &F, InlineCandidate &Candidate, uint64_t SumOrigin,
505 uint64_t &Sum, SmallVector<CallBase *, 8> *InlinedCallSites = nullptr);
506
507 bool inlineHotFunctions(Function &F,
508 DenseSet<GlobalValue::GUID> &InlinedGUIDs);
509 std::optional<InlineCost> getExternalInlineAdvisorCost(CallBase &CB);
510 bool getExternalInlineAdvisorShouldInline(CallBase &CB);
511 InlineCost shouldInlineCandidate(InlineCandidate &Candidate);
512 bool getInlineCandidate(InlineCandidate *NewCandidate, CallBase *CB);
513 bool
514 tryInlineCandidate(InlineCandidate &Candidate,
515 SmallVector<CallBase *, 8> *InlinedCallSites = nullptr);
516 bool
517 inlineHotFunctionsWithPriority(Function &F,
518 DenseSet<GlobalValue::GUID> &InlinedGUIDs);
519 // Inline cold/small functions in addition to hot ones
520 bool shouldInlineColdCallee(CallBase &CallInst);
521 void emitOptimizationRemarksForInlineCandidates(
522 const SmallVectorImpl<CallBase *> &Candidates, const Function &F,
523 bool Hot);
524 void promoteMergeNotInlinedContextSamples(
525 MapVector<CallBase *, const FunctionSamples *> NonInlinedCallSites,
526 const Function &F);
527 std::vector<Function *> buildFunctionOrder(Module &M, LazyCallGraph &CG);
528 std::unique_ptr<ProfiledCallGraph> buildProfiledCallGraph(Module &M);
529 void generateMDProfMetadata(Function &F);
530 bool rejectHighStalenessProfile(Module &M, ProfileSummaryInfo *PSI,
531 const SampleProfileMap &Profiles);
532 void removePseudoProbeInstsDiscriminator(Module &M);
533
534 /// Map from function name to Function *. Used to find the function from
535 /// the function name. If the function name contains suffix, additional
536 /// entry is added to map from the stripped name to the function if there
537 /// is one-to-one mapping.
538 HashKeyMap<DenseMap, FunctionId, Function *> SymbolMap;
539
540 /// Map from function name to profile name generated by call-graph based
541 /// profile fuzzy matching(--salvage-unused-profile).
542 HashKeyMap<DenseMap, FunctionId, FunctionId> FuncNameToProfNameMap;
543
544 std::function<AssumptionCache &(Function &)> GetAC;
545 std::function<TargetTransformInfo &(Function &)> GetTTI;
546 std::function<const TargetLibraryInfo &(Function &)> GetTLI;
547 LazyCallGraph &CG;
548
549 /// Profile tracker for different context.
550 std::unique_ptr<SampleContextTracker> ContextTracker;
551
552 /// Flag indicating which LTO/ThinLTO phase the pass is invoked in.
553 ///
554 /// We need to know the LTO phase because for example in ThinLTOPrelink
555 /// phase, in annotation, we should not promote indirect calls. Instead,
556 /// we will mark GUIDs that needs to be annotated to the function.
557 const ThinOrFullLTOPhase LTOPhase;
558 const std::string AnnotatedPassName;
559
560 /// Profle Symbol list tells whether a function name appears in the binary
561 /// used to generate the current profile.
562 std::shared_ptr<ProfileSymbolList> PSL;
563
564 // Information recorded when we declined to inline a call site
565 // because we have determined it is too cold is accumulated for
566 // each callee function. Initially this is just the entry count.
567 struct NotInlinedProfileInfo {
568 uint64_t entryCount;
569 };
570 DenseMap<Function *, NotInlinedProfileInfo> notInlinedCallInfo;
571
572 // GUIDToFuncNameMap saves the mapping from GUID to the symbol name, for
573 // all the function symbols defined or declared in current module.
574 DenseMap<uint64_t, StringRef> GUIDToFuncNameMap;
575
576 // All the Names used in FunctionSamples including outline function
577 // names, inline instance names and call target names.
578 StringSet<> NamesInProfile;
579 // MD5 version of NamesInProfile. Either NamesInProfile or GUIDsInProfile is
580 // populated, depends on whether the profile uses MD5. Because the name table
581 // generally contains several magnitude more entries than the number of
582 // functions, we do not want to convert all names from one form to another.
583 llvm::DenseSet<uint64_t> GUIDsInProfile;
584
585 // For symbol in profile symbol list, whether to regard their profiles
586 // to be accurate. It is mainly decided by existance of profile symbol
587 // list and -profile-accurate-for-symsinlist flag, but it can be
588 // overriden by -profile-sample-accurate or profile-sample-accurate
589 // attribute.
590 bool ProfAccForSymsInList;
591
592 bool DisableSampleProfileInlining;
593
594 bool UseFlattenedProfile;
595
596 // External inline advisor used to replay inline decision from remarks.
597 std::unique_ptr<InlineAdvisor> ExternalInlineAdvisor;
598
599 // A helper to implement the sample profile matching algorithm.
600 std::unique_ptr<SampleProfileMatcher> MatchingManager;
601
602private:
603 const char *getAnnotatedRemarkPassName() const {
604 return AnnotatedPassName.c_str();
605 }
606};
607} // end anonymous namespace
608
609namespace llvm {
610template <>
611inline bool SampleProfileInference<Function>::isExit(const BasicBlock *BB) {
612 return succ_empty(BB);
613}
614
615template <>
616inline void SampleProfileInference<Function>::findUnlikelyJumps(
617 const std::vector<const BasicBlockT *> &BasicBlocks,
618 BlockEdgeMap &Successors, FlowFunction &Func) {
619 for (auto &Jump : Func.Jumps) {
620 const auto *BB = BasicBlocks[Jump.Source];
621 const auto *Succ = BasicBlocks[Jump.Target];
622 const Instruction *TI = BB->getTerminator();
623 // Check if a block ends with InvokeInst and mark non-taken branch unlikely.
624 // In that case block Succ should be a landing pad
625 const auto &Succs = Successors[BB];
626 if (Succs.size() == 2 && Succs.back() == Succ) {
627 if (isa<InvokeInst>(Val: TI)) {
628 Jump.IsUnlikely = true;
629 }
630 }
631 const Instruction *SuccTI = Succ->getTerminator();
632 // Check if the target block contains UnreachableInst and mark it unlikely
633 if (SuccTI->getNumSuccessors() == 0) {
634 if (isa<UnreachableInst>(Val: SuccTI)) {
635 Jump.IsUnlikely = true;
636 }
637 }
638 }
639}
640
641template <>
642void SampleProfileLoaderBaseImpl<Function>::computeDominanceAndLoopInfo(
643 Function &F) {
644 DT.reset(p: new DominatorTree);
645 DT->recalculate(Func&: F);
646
647 PDT.reset(p: new PostDominatorTree(F));
648
649 LI.reset(p: new LoopInfo);
650 LI->analyze(DomTree: *DT);
651}
652} // namespace llvm
653
654ErrorOr<uint64_t> SampleProfileLoader::getInstWeight(const Instruction &Inst) {
655 if (FunctionSamples::ProfileIsProbeBased)
656 return getProbeWeight(Inst);
657
658 const DebugLoc &DLoc = Inst.getDebugLoc();
659 if (!DLoc)
660 return std::error_code();
661
662 // Ignore all intrinsics, phinodes and branch instructions.
663 // Branch and phinodes instruction usually contains debug info from sources
664 // outside of the residing basic block, thus we ignore them during annotation.
665 if (isa<UncondBrInst, CondBrInst, IntrinsicInst, PHINode>(Val: Inst))
666 return std::error_code();
667
668 // For non-CS profile, if a direct call/invoke instruction is inlined in
669 // profile (findCalleeFunctionSamples returns non-empty result), but not
670 // inlined here, it means that the inlined callsite has no sample, thus the
671 // call instruction should have 0 count.
672 // For CS profile, the callsite count of previously inlined callees is
673 // populated with the entry count of the callees.
674 if (!FunctionSamples::ProfileIsCS)
675 if (const auto *CB = dyn_cast<CallBase>(Val: &Inst))
676 if (!CB->isIndirectCall() && findCalleeFunctionSamples(I: *CB))
677 return 0;
678
679 return getInstWeightImpl(Inst);
680}
681
682/// Get the FunctionSamples for a call instruction.
683///
684/// The FunctionSamples of a call/invoke instruction \p Inst is the inlined
685/// instance in which that call instruction is calling to. It contains
686/// all samples that resides in the inlined instance. We first find the
687/// inlined instance in which the call instruction is from, then we
688/// traverse its children to find the callsite with the matching
689/// location.
690///
691/// \param Inst Call/Invoke instruction to query.
692///
693/// \returns The FunctionSamples pointer to the inlined instance.
694const FunctionSamples *
695SampleProfileLoader::findCalleeFunctionSamples(const CallBase &Inst) const {
696 const DILocation *DIL = Inst.getDebugLoc();
697 if (!DIL) {
698 return nullptr;
699 }
700
701 StringRef CalleeName;
702 if (Function *Callee = Inst.getCalledFunction())
703 CalleeName = Callee->getName();
704
705 if (FunctionSamples::ProfileIsCS)
706 return ContextTracker->getCalleeContextSamplesFor(Inst, CalleeName);
707
708 const FunctionSamples *FS = findFunctionSamples(I: Inst);
709 if (FS == nullptr)
710 return nullptr;
711
712 return FS->findFunctionSamplesAt(Loc: FunctionSamples::getCallSiteIdentifier(DIL),
713 CalleeName, Remapper: Reader->getRemapper(),
714 FuncNameToProfNameMap: &FuncNameToProfNameMap);
715}
716
717/// Returns a vector of FunctionSamples that are the indirect call targets
718/// of \p Inst. The vector is sorted by the total number of samples. Stores
719/// the total call count of the indirect call in \p Sum.
720std::vector<const FunctionSamples *>
721SampleProfileLoader::findIndirectCallFunctionSamples(
722 const Instruction &Inst, uint64_t &Sum) const {
723 const DILocation *DIL = Inst.getDebugLoc();
724 std::vector<const FunctionSamples *> R;
725
726 if (!DIL) {
727 return R;
728 }
729
730 auto FSCompare = [](const FunctionSamples *L, const FunctionSamples *R) {
731 assert(L && R && "Expect non-null FunctionSamples");
732 if (L->getHeadSamplesEstimate() != R->getHeadSamplesEstimate())
733 return L->getHeadSamplesEstimate() > R->getHeadSamplesEstimate();
734 return L->getGUID() < R->getGUID();
735 };
736
737 if (FunctionSamples::ProfileIsCS) {
738 auto CalleeSamples =
739 ContextTracker->getIndirectCalleeContextSamplesFor(DIL);
740 if (CalleeSamples.empty())
741 return R;
742
743 // For CSSPGO, we only use target context profile's entry count
744 // as that already includes both inlined callee and non-inlined ones..
745 Sum = 0;
746 for (const auto *const FS : CalleeSamples) {
747 Sum += FS->getHeadSamplesEstimate();
748 R.push_back(x: FS);
749 }
750 llvm::sort(C&: R, Comp: FSCompare);
751 return R;
752 }
753
754 const FunctionSamples *FS = findFunctionSamples(I: Inst);
755 if (FS == nullptr)
756 return R;
757
758 auto CallSite = FunctionSamples::getCallSiteIdentifier(DIL);
759 Sum = 0;
760 if (auto T = FS->findCallTargetMapAt(CallSite))
761 for (const auto &T_C : *T)
762 Sum += T_C.second;
763 if (const FunctionSamplesMap *M = FS->findFunctionSamplesMapAt(Loc: CallSite)) {
764 if (M->empty())
765 return R;
766 for (const auto &NameFS : *M) {
767 Sum += NameFS.second.getHeadSamplesEstimate();
768 R.push_back(x: &NameFS.second);
769 }
770 llvm::sort(C&: R, Comp: FSCompare);
771 }
772 return R;
773}
774
775const FunctionSamples *
776SampleProfileLoader::findFunctionSamples(const Instruction &Inst) const {
777 if (FunctionSamples::ProfileIsProbeBased) {
778 std::optional<PseudoProbe> Probe = extractProbe(Inst);
779 if (!Probe)
780 return nullptr;
781 }
782
783 const DILocation *DIL = Inst.getDebugLoc();
784 if (!DIL)
785 return Samples;
786
787 auto it = DILocation2SampleMap.try_emplace(Key: DIL,Args: nullptr);
788 if (it.second) {
789 if (FunctionSamples::ProfileIsCS)
790 it.first->second = ContextTracker->getContextSamplesFor(DIL);
791 else
792 it.first->second = Samples->findFunctionSamples(
793 DIL, Remapper: Reader->getRemapper(), FuncNameToProfNameMap: &FuncNameToProfNameMap);
794 }
795 return it.first->second;
796}
797
798/// Check whether the indirect call promotion history of \p Inst allows
799/// the promotion for \p Candidate.
800/// If the profile count for the promotion candidate \p Candidate is
801/// NOMORE_ICP_MAGICNUM, it means \p Candidate has already been promoted
802/// for \p Inst. If we already have at least MaxNumPromotions
803/// NOMORE_ICP_MAGICNUM count values in the value profile of \p Inst, we
804/// cannot promote for \p Inst anymore.
805static bool doesHistoryAllowICP(const Instruction &Inst, StringRef Candidate) {
806 uint64_t TotalCount = 0;
807 auto ValueData = getValueProfDataFromInst(Inst, ValueKind: IPVK_IndirectCallTarget,
808 MaxNumValueData: MaxNumPromotions, TotalC&: TotalCount, GetNoICPValue: true);
809 // No valid value profile so no promoted targets have been recorded
810 // before. Ok to do ICP.
811 if (ValueData.empty())
812 return true;
813
814 unsigned NumPromoted = 0;
815 for (const auto &V : ValueData) {
816 if (V.Count != NOMORE_ICP_MAGICNUM)
817 continue;
818
819 // If the promotion candidate has NOMORE_ICP_MAGICNUM count in the
820 // metadata, it means the candidate has been promoted for this
821 // indirect call.
822 if (V.Value == Function::getGUIDAssumingExternalLinkage(GlobalName: Candidate))
823 return false;
824 NumPromoted++;
825 // If already have MaxNumPromotions promotion, don't do it anymore.
826 if (NumPromoted == MaxNumPromotions)
827 return false;
828 }
829 return true;
830}
831
832/// Update indirect call target profile metadata for \p Inst.
833/// Usually \p Sum is the sum of counts of all the targets for \p Inst.
834/// If it is 0, it means updateIDTMetaData is used to mark a
835/// certain target to be promoted already. If it is not zero,
836/// we expect to use it to update the total count in the value profile.
837static void
838updateIDTMetaData(Instruction &Inst,
839 const SmallVectorImpl<InstrProfValueData> &CallTargets,
840 uint64_t Sum) {
841 // Bail out early if MaxNumPromotions is zero.
842 // This prevents allocating an array of zero length below.
843 //
844 // Note `updateIDTMetaData` is called in two places so check
845 // `MaxNumPromotions` inside it.
846 if (MaxNumPromotions == 0)
847 return;
848 // OldSum is the existing total count in the value profile data.
849 uint64_t OldSum = 0;
850 auto ValueData = getValueProfDataFromInst(Inst, ValueKind: IPVK_IndirectCallTarget,
851 MaxNumValueData: MaxNumPromotions, TotalC&: OldSum, GetNoICPValue: true);
852
853 DenseMap<uint64_t, uint64_t> ValueCountMap;
854 if (Sum == 0) {
855 assert((CallTargets.size() == 1 &&
856 CallTargets[0].Count == NOMORE_ICP_MAGICNUM) &&
857 "If sum is 0, assume only one element in CallTargets "
858 "with count being NOMORE_ICP_MAGICNUM");
859 // Initialize ValueCountMap with existing value profile data.
860 for (const auto &V : ValueData)
861 ValueCountMap[V.Value] = V.Count;
862 auto Pair =
863 ValueCountMap.try_emplace(Key: CallTargets[0].Value, Args: CallTargets[0].Count);
864 // If the target already exists in value profile, decrease the total
865 // count OldSum and reset the target's count to NOMORE_ICP_MAGICNUM.
866 if (!Pair.second) {
867 OldSum -= Pair.first->second;
868 Pair.first->second = NOMORE_ICP_MAGICNUM;
869 }
870 Sum = OldSum;
871 } else {
872 // Initialize ValueCountMap with existing NOMORE_ICP_MAGICNUM
873 // counts in the value profile.
874 for (const auto &V : ValueData) {
875 if (V.Count == NOMORE_ICP_MAGICNUM)
876 ValueCountMap[V.Value] = V.Count;
877 }
878
879 for (const auto &Data : CallTargets) {
880 auto Pair = ValueCountMap.try_emplace(Key: Data.Value, Args: Data.Count);
881 if (Pair.second)
882 continue;
883 // The target represented by Data.Value has already been promoted.
884 // Keep the count as NOMORE_ICP_MAGICNUM in the profile and decrease
885 // Sum by Data.Count.
886 assert(Sum >= Data.Count && "Sum should never be less than Data.Count");
887 Sum -= Data.Count;
888 }
889 }
890
891 SmallVector<InstrProfValueData, 8> NewCallTargets;
892 for (const auto &ValueCount : ValueCountMap) {
893 NewCallTargets.emplace_back(
894 Args: InstrProfValueData{.Value: ValueCount.first, .Count: ValueCount.second});
895 }
896
897 llvm::sort(C&: NewCallTargets,
898 Comp: [](const InstrProfValueData &L, const InstrProfValueData &R) {
899 return std::tie(args: L.Count, args: L.Value) > std::tie(args: R.Count, args: R.Value);
900 });
901
902 uint32_t MaxMDCount =
903 std::min(a: NewCallTargets.size(), b: static_cast<size_t>(MaxNumPromotions));
904 annotateValueSite(M&: *Inst.getParent()->getParent()->getParent(), Inst,
905 VDs: NewCallTargets, Sum, ValueKind: IPVK_IndirectCallTarget, MaxMDCount);
906}
907
908/// Attempt to promote indirect call and also inline the promoted call.
909///
910/// \param F Caller function.
911/// \param Candidate ICP and inline candidate.
912/// \param SumOrigin Original sum of target counts for indirect call before
913/// promoting given candidate.
914/// \param Sum Prorated sum of remaining target counts for indirect call
915/// after promoting given candidate.
916/// \param InlinedCallSite Output vector for new call sites exposed after
917/// inlining.
918bool SampleProfileLoader::tryPromoteAndInlineCandidate(
919 Function &F, InlineCandidate &Candidate, uint64_t SumOrigin, uint64_t &Sum,
920 SmallVector<CallBase *, 8> *InlinedCallSite) {
921 // Bail out early if sample-loader inliner is disabled.
922 if (DisableSampleProfileInlining)
923 return false;
924
925 // Bail out early if MaxNumPromotions is zero.
926 // This prevents allocating an array of zero length in callees below.
927 if (MaxNumPromotions == 0)
928 return false;
929 auto CalleeFunctionName = Candidate.CalleeSamples->getFunction();
930 auto R = SymbolMap.find(Key: CalleeFunctionName);
931 if (R == SymbolMap.end() || !R->second)
932 return false;
933
934 auto &CI = *Candidate.CallInstr;
935 if (!doesHistoryAllowICP(Inst: CI, Candidate: R->second->getName()))
936 return false;
937
938 const char *Reason = "Callee function not available";
939 // R->getValue() != &F is to prevent promoting a recursive call.
940 // If it is a recursive call, we do not inline it as it could bloat
941 // the code exponentially. There is way to better handle this, e.g.
942 // clone the caller first, and inline the cloned caller if it is
943 // recursive. As llvm does not inline recursive calls, we will
944 // simply ignore it instead of handling it explicitly.
945 if (!R->second->isDeclaration() && R->second->getSubprogram() &&
946 R->second->hasFnAttribute(Kind: "use-sample-profile") &&
947 R->second != &F && isLegalToPromote(CB: CI, Callee: R->second, FailureReason: &Reason)) {
948 // For promoted target, set its value with NOMORE_ICP_MAGICNUM count
949 // in the value profile metadata so the target won't be promoted again.
950 SmallVector<InstrProfValueData, 1> SortedCallTargets = {InstrProfValueData{
951 .Value: Function::getGUIDAssumingExternalLinkage(GlobalName: R->second->getName()),
952 .Count: NOMORE_ICP_MAGICNUM}};
953 updateIDTMetaData(Inst&: CI, CallTargets: SortedCallTargets, Sum: 0);
954
955 auto *DI = &pgo::promoteIndirectCall(
956 CB&: CI, F: R->second, Count: Candidate.CallsiteCount, TotalCount: Sum, AttachProfToDirectCall: false, ORE);
957 if (DI) {
958 Sum -= Candidate.CallsiteCount;
959 // Do not prorate the indirect callsite distribution since the original
960 // distribution will be used to scale down non-promoted profile target
961 // counts later. By doing this we lose track of the real callsite count
962 // for the leftover indirect callsite as a trade off for accurate call
963 // target counts.
964 // TODO: Ideally we would have two separate factors, one for call site
965 // counts and one is used to prorate call target counts.
966 // Do not update the promoted direct callsite distribution at this
967 // point since the original distribution combined with the callee profile
968 // will be used to prorate callsites from the callee if inlined. Once not
969 // inlined, the direct callsite distribution should be prorated so that
970 // the it will reflect the real callsite counts.
971 Candidate.CallInstr = DI;
972 if (isa<CallInst>(Val: DI) || isa<InvokeInst>(Val: DI)) {
973 bool Inlined = tryInlineCandidate(Candidate, InlinedCallSites: InlinedCallSite);
974 if (!Inlined) {
975 // Prorate the direct callsite distribution so that it reflects real
976 // callsite counts.
977 setProbeDistributionFactor(
978 Inst&: *DI, Factor: static_cast<float>(Candidate.CallsiteCount) / SumOrigin);
979 }
980 return Inlined;
981 }
982 }
983 } else {
984 LLVM_DEBUG(dbgs() << "\nFailed to promote indirect call to "
985 << FunctionSamples::getCanonicalFnName(
986 Candidate.CallInstr->getName())<< " because "
987 << Reason << "\n");
988 }
989 return false;
990}
991
992bool SampleProfileLoader::shouldInlineColdCallee(CallBase &CallInst) {
993 if (!ProfileSizeInline)
994 return false;
995
996 Function *Callee = CallInst.getCalledFunction();
997 if (Callee == nullptr)
998 return false;
999
1000 InlineCost Cost = getInlineCost(Call&: CallInst, Params: getInlineParams(), CalleeTTI&: GetTTI(*Callee),
1001 GetAssumptionCache: GetAC, GetTLI);
1002
1003 if (Cost.isNever())
1004 return false;
1005
1006 if (Cost.isAlways())
1007 return true;
1008
1009 return Cost.getCost() <= SampleColdCallSiteThreshold;
1010}
1011
1012void SampleProfileLoader::emitOptimizationRemarksForInlineCandidates(
1013 const SmallVectorImpl<CallBase *> &Candidates, const Function &F,
1014 bool Hot) {
1015 for (auto *I : Candidates) {
1016 Function *CalledFunction = I->getCalledFunction();
1017 if (CalledFunction) {
1018 ORE->emit(OptDiag: OptimizationRemarkAnalysis(getAnnotatedRemarkPassName(),
1019 "InlineAttempt", I->getDebugLoc(),
1020 I->getParent())
1021 << "previous inlining reattempted for "
1022 << (Hot ? "hotness: '" : "size: '")
1023 << ore::NV("Callee", CalledFunction) << "' into '"
1024 << ore::NV("Caller", &F) << "'");
1025 }
1026 }
1027}
1028
1029void SampleProfileLoader::findExternalInlineCandidate(
1030 CallBase *CB, const FunctionSamples *Samples,
1031 DenseSet<GlobalValue::GUID> &InlinedGUIDs, uint64_t Threshold) {
1032
1033 // If ExternalInlineAdvisor(ReplayInlineAdvisor) wants to inline an external
1034 // function make sure it's imported
1035 if (CB && getExternalInlineAdvisorShouldInline(CB&: *CB)) {
1036 // Samples may not exist for replayed function, if so
1037 // just add the direct GUID and move on
1038 if (!Samples) {
1039 InlinedGUIDs.insert(V: Function::getGUIDAssumingExternalLinkage(
1040 GlobalName: CB->getCalledFunction()->getName()));
1041 return;
1042 }
1043 // Otherwise, drop the threshold to import everything that we can
1044 Threshold = 0;
1045 }
1046
1047 // In some rare cases, call instruction could be changed after being pushed
1048 // into inline candidate queue, this is because earlier inlining may expose
1049 // constant propagation which can change indirect call to direct call. When
1050 // this happens, we may fail to find matching function samples for the
1051 // candidate later, even if a match was found when the candidate was enqueued.
1052 if (!Samples)
1053 return;
1054
1055 // For AutoFDO profile, retrieve candidate profiles by walking over
1056 // the nested inlinee profiles.
1057 if (!FunctionSamples::ProfileIsCS) {
1058 // Set threshold to zero to honor pre-inliner decision.
1059 if (UsePreInlinerDecision)
1060 Threshold = 0;
1061 Samples->findInlinedFunctions(S&: InlinedGUIDs, SymbolMap, Threshold);
1062 return;
1063 }
1064
1065 ContextTrieNode *Caller = ContextTracker->getContextNodeForProfile(FSamples: Samples);
1066 std::queue<ContextTrieNode *> CalleeList;
1067 CalleeList.push(x: Caller);
1068 while (!CalleeList.empty()) {
1069 ContextTrieNode *Node = CalleeList.front();
1070 CalleeList.pop();
1071 FunctionSamples *CalleeSample = Node->getFunctionSamples();
1072 // For CSSPGO profile, retrieve candidate profile by walking over the
1073 // trie built for context profile. Note that also take call targets
1074 // even if callee doesn't have a corresponding context profile.
1075 if (!CalleeSample)
1076 continue;
1077
1078 // If pre-inliner decision is used, honor that for importing as well.
1079 bool PreInline =
1080 UsePreInlinerDecision &&
1081 CalleeSample->getContext().hasAttribute(A: ContextShouldBeInlined);
1082 if (!PreInline && CalleeSample->getHeadSamplesEstimate() < Threshold)
1083 continue;
1084
1085 Function *Func = SymbolMap.lookup(Key: CalleeSample->getFunction());
1086 // Add to the import list only when it's defined out of module.
1087 if (!Func || Func->isDeclaration())
1088 InlinedGUIDs.insert(V: CalleeSample->getGUID());
1089
1090 // Import hot CallTargets, which may not be available in IR because full
1091 // profile annotation cannot be done until backend compilation in ThinLTO.
1092 for (const auto &BS : CalleeSample->getBodySamples())
1093 for (const auto &TS : BS.second.getCallTargets())
1094 if (TS.second > Threshold) {
1095 const Function *Callee = SymbolMap.lookup(Key: TS.first);
1096 if (!Callee || Callee->isDeclaration())
1097 InlinedGUIDs.insert(V: TS.first.getHashCode());
1098 }
1099
1100 // Import hot child context profile associted with callees. Note that this
1101 // may have some overlap with the call target loop above, but doing this
1102 // based child context profile again effectively allow us to use the max of
1103 // entry count and call target count to determine importing.
1104 for (auto &Child : Node->getAllChildContext()) {
1105 ContextTrieNode *CalleeNode = &Child.second;
1106 CalleeList.push(x: CalleeNode);
1107 }
1108 }
1109}
1110
1111/// Iteratively inline hot callsites of a function.
1112///
1113/// Iteratively traverse all callsites of the function \p F, so as to
1114/// find out callsites with corresponding inline instances.
1115///
1116/// For such callsites,
1117/// - If it is hot enough, inline the callsites and adds callsites of the callee
1118/// into the caller. If the call is an indirect call, first promote
1119/// it to direct call. Each indirect call is limited with a single target.
1120///
1121/// - If a callsite is not inlined, merge the its profile to the outline
1122/// version (if --sample-profile-merge-inlinee is true), or scale the
1123/// counters of standalone function based on the profile of inlined
1124/// instances (if --sample-profile-merge-inlinee is false).
1125///
1126/// Later passes may consume the updated profiles.
1127///
1128/// \param F function to perform iterative inlining.
1129/// \param InlinedGUIDs a set to be updated to include all GUIDs that are
1130/// inlined in the profiled binary.
1131///
1132/// \returns True if there is any inline happened.
1133bool SampleProfileLoader::inlineHotFunctions(
1134 Function &F, DenseSet<GlobalValue::GUID> &InlinedGUIDs) {
1135 // ProfAccForSymsInList is used in callsiteIsHot. The assertion makes sure
1136 // Profile symbol list is ignored when profile-sample-accurate is on.
1137 assert((!ProfAccForSymsInList ||
1138 (!ProfileSampleAccurate &&
1139 !F.hasFnAttribute("profile-sample-accurate"))) &&
1140 "ProfAccForSymsInList should be false when profile-sample-accurate "
1141 "is enabled");
1142
1143 MapVector<CallBase *, const FunctionSamples *> LocalNotInlinedCallSites;
1144 bool Changed = false;
1145 bool LocalChanged = true;
1146 while (LocalChanged) {
1147 LocalChanged = false;
1148 SmallVector<CallBase *, 10> CIS;
1149 for (auto &BB : F) {
1150 bool Hot = false;
1151 SmallVector<CallBase *, 10> AllCandidates;
1152 SmallVector<CallBase *, 10> ColdCandidates;
1153 for (auto &I : BB) {
1154 const FunctionSamples *FS = nullptr;
1155 if (auto *CB = dyn_cast<CallBase>(Val: &I)) {
1156 if (!isa<IntrinsicInst>(Val: I)) {
1157 if ((FS = findCalleeFunctionSamples(Inst: *CB))) {
1158 assert((!FunctionSamples::UseMD5 || FS->GUIDToFuncNameMap) &&
1159 "GUIDToFuncNameMap has to be populated");
1160 AllCandidates.push_back(Elt: CB);
1161 if (FS->getHeadSamplesEstimate() > 0 ||
1162 FunctionSamples::ProfileIsCS)
1163 LocalNotInlinedCallSites.insert(KV: {CB, FS});
1164 if (callsiteIsHot(CallsiteFS: FS, PSI, ProfAccForSymsInList))
1165 Hot = true;
1166 else if (shouldInlineColdCallee(CallInst&: *CB))
1167 ColdCandidates.push_back(Elt: CB);
1168 } else if (getExternalInlineAdvisorShouldInline(CB&: *CB)) {
1169 AllCandidates.push_back(Elt: CB);
1170 }
1171 }
1172 }
1173 }
1174 if (Hot || ExternalInlineAdvisor) {
1175 CIS.insert(I: CIS.begin(), From: AllCandidates.begin(), To: AllCandidates.end());
1176 emitOptimizationRemarksForInlineCandidates(Candidates: AllCandidates, F, Hot: true);
1177 } else {
1178 CIS.insert(I: CIS.begin(), From: ColdCandidates.begin(), To: ColdCandidates.end());
1179 emitOptimizationRemarksForInlineCandidates(Candidates: ColdCandidates, F, Hot: false);
1180 }
1181 }
1182 for (CallBase *I : CIS) {
1183 Function *CalledFunction = I->getCalledFunction();
1184 InlineCandidate Candidate = {.CallInstr: I, .CalleeSamples: LocalNotInlinedCallSites.lookup(Key: I),
1185 .CallsiteCount: 0 /* dummy count */,
1186 .CallsiteDistribution: 1.0 /* dummy distribution factor */};
1187 // Do not inline recursive calls.
1188 if (CalledFunction == &F)
1189 continue;
1190 if (I->isIndirectCall()) {
1191 uint64_t Sum;
1192 for (const auto *FS : findIndirectCallFunctionSamples(Inst: *I, Sum)) {
1193 uint64_t SumOrigin = Sum;
1194 if (LTOPhase == ThinOrFullLTOPhase::ThinLTOPreLink) {
1195 findExternalInlineCandidate(CB: I, Samples: FS, InlinedGUIDs,
1196 Threshold: PSI->getOrCompHotCountThreshold());
1197 continue;
1198 }
1199 if (!callsiteIsHot(CallsiteFS: FS, PSI, ProfAccForSymsInList))
1200 continue;
1201
1202 Candidate = {.CallInstr: I, .CalleeSamples: FS, .CallsiteCount: FS->getHeadSamplesEstimate(), .CallsiteDistribution: 1.0};
1203 if (tryPromoteAndInlineCandidate(F, Candidate, SumOrigin, Sum)) {
1204 LocalNotInlinedCallSites.erase(Key: I);
1205 LocalChanged = true;
1206 }
1207 }
1208 } else if (CalledFunction && CalledFunction->getSubprogram() &&
1209 !CalledFunction->isDeclaration()) {
1210 if (tryInlineCandidate(Candidate)) {
1211 LocalNotInlinedCallSites.erase(Key: I);
1212 LocalChanged = true;
1213 }
1214 } else if (LTOPhase == ThinOrFullLTOPhase::ThinLTOPreLink) {
1215 findExternalInlineCandidate(CB: I, Samples: findCalleeFunctionSamples(Inst: *I),
1216 InlinedGUIDs,
1217 Threshold: PSI->getOrCompHotCountThreshold());
1218 }
1219 }
1220 Changed |= LocalChanged;
1221 }
1222
1223 // For CS profile, profile for not inlined context will be merged when
1224 // base profile is being retrieved.
1225 if (!FunctionSamples::ProfileIsCS)
1226 promoteMergeNotInlinedContextSamples(NonInlinedCallSites: LocalNotInlinedCallSites, F);
1227 return Changed;
1228}
1229
1230bool SampleProfileLoader::tryInlineCandidate(
1231 InlineCandidate &Candidate, SmallVector<CallBase *, 8> *InlinedCallSites) {
1232 // Do not attempt to inline a candidate if
1233 // --disable-sample-loader-inlining is true.
1234 if (DisableSampleProfileInlining)
1235 return false;
1236
1237 CallBase &CB = *Candidate.CallInstr;
1238 Function *CalledFunction = CB.getCalledFunction();
1239 assert(CalledFunction && "Expect a callee with definition");
1240 DebugLoc DLoc = CB.getDebugLoc();
1241 BasicBlock *BB = CB.getParent();
1242
1243 InlineCost Cost = shouldInlineCandidate(Candidate);
1244 if (Cost.isNever()) {
1245 ORE->emit(OptDiag: OptimizationRemarkAnalysis(getAnnotatedRemarkPassName(),
1246 "InlineFail", DLoc, BB)
1247 << "incompatible inlining");
1248 return false;
1249 }
1250
1251 if (!Cost)
1252 return false;
1253
1254 InlineFunctionInfo IFI(GetAC);
1255 IFI.UpdateProfile = false;
1256 InlineResult IR = InlineFunction(CB, IFI,
1257 /*MergeAttributes=*/true);
1258 if (!IR.isSuccess())
1259 return false;
1260
1261 // The call to InlineFunction erases I, so we can't pass it here.
1262 emitInlinedIntoBasedOnCost(ORE&: *ORE, DLoc, Block: BB, Callee: *CalledFunction, Caller: *BB->getParent(),
1263 IC: Cost, ForProfileContext: true, PassName: getAnnotatedRemarkPassName());
1264
1265 // Now populate the list of newly exposed call sites.
1266 if (InlinedCallSites) {
1267 InlinedCallSites->clear();
1268 llvm::append_range(C&: *InlinedCallSites, R&: IFI.InlinedCallSites);
1269 }
1270
1271 if (FunctionSamples::ProfileIsCS)
1272 ContextTracker->markContextSamplesInlined(InlinedSamples: Candidate.CalleeSamples);
1273 ++NumCSInlined;
1274
1275 // Prorate inlined probes for a duplicated inlining callsite which probably
1276 // has a distribution less than 100%. Samples for an inlinee should be
1277 // distributed among the copies of the original callsite based on each
1278 // callsite's distribution factor for counts accuracy. Note that an inlined
1279 // probe may come with its own distribution factor if it has been duplicated
1280 // in the inlinee body. The two factor are multiplied to reflect the
1281 // aggregation of duplication.
1282 if (Candidate.CallsiteDistribution < 1) {
1283 for (auto &I : IFI.InlinedCallSites) {
1284 if (std::optional<PseudoProbe> Probe = extractProbe(Inst: *I))
1285 setProbeDistributionFactor(Inst&: *I, Factor: Probe->Factor *
1286 Candidate.CallsiteDistribution);
1287 }
1288 NumDuplicatedInlinesite++;
1289 }
1290
1291 return true;
1292}
1293
1294bool SampleProfileLoader::getInlineCandidate(InlineCandidate *NewCandidate,
1295 CallBase *CB) {
1296 assert(CB && "Expect non-null call instruction");
1297
1298 if (isa<IntrinsicInst>(Val: CB))
1299 return false;
1300
1301 // Find the callee's profile. For indirect call, find hottest target profile.
1302 const FunctionSamples *CalleeSamples = findCalleeFunctionSamples(Inst: *CB);
1303 // If ExternalInlineAdvisor wants to inline this site, do so even
1304 // if Samples are not present.
1305 if (!CalleeSamples && !getExternalInlineAdvisorShouldInline(CB&: *CB))
1306 return false;
1307
1308 float Factor = 1.0;
1309 if (std::optional<PseudoProbe> Probe = extractProbe(Inst: *CB))
1310 Factor = Probe->Factor;
1311
1312 uint64_t CallsiteCount =
1313 CalleeSamples ? CalleeSamples->getHeadSamplesEstimate() * Factor : 0;
1314 *NewCandidate = {.CallInstr: CB, .CalleeSamples: CalleeSamples, .CallsiteCount: CallsiteCount, .CallsiteDistribution: Factor};
1315 return true;
1316}
1317
1318std::optional<InlineCost>
1319SampleProfileLoader::getExternalInlineAdvisorCost(CallBase &CB) {
1320 std::unique_ptr<InlineAdvice> Advice = nullptr;
1321 if (ExternalInlineAdvisor) {
1322 Advice = ExternalInlineAdvisor->getAdvice(CB);
1323 if (Advice) {
1324 if (!Advice->isInliningRecommended()) {
1325 Advice->recordUnattemptedInlining();
1326 return InlineCost::getNever(Reason: "not previously inlined");
1327 }
1328 Advice->recordInlining();
1329 return InlineCost::getAlways(Reason: "previously inlined");
1330 }
1331 }
1332
1333 return {};
1334}
1335
1336bool SampleProfileLoader::getExternalInlineAdvisorShouldInline(CallBase &CB) {
1337 std::optional<InlineCost> Cost = getExternalInlineAdvisorCost(CB);
1338 return Cost ? !!*Cost : false;
1339}
1340
1341InlineCost
1342SampleProfileLoader::shouldInlineCandidate(InlineCandidate &Candidate) {
1343 if (std::optional<InlineCost> ReplayCost =
1344 getExternalInlineAdvisorCost(CB&: *Candidate.CallInstr))
1345 return *ReplayCost;
1346 // Adjust threshold based on call site hotness, only do this for callsite
1347 // prioritized inliner because otherwise cost-benefit check is done earlier.
1348 int SampleThreshold = SampleColdCallSiteThreshold;
1349 if (CallsitePrioritizedInline) {
1350 if (Candidate.CallsiteCount > PSI->getHotCountThreshold())
1351 SampleThreshold = SampleHotCallSiteThreshold;
1352 else if (!ProfileSizeInline)
1353 return InlineCost::getNever(Reason: "cold callsite");
1354 }
1355
1356 Function *Callee = Candidate.CallInstr->getCalledFunction();
1357 assert(Callee && "Expect a definition for inline candidate of direct call");
1358
1359 InlineParams Params = getInlineParams();
1360 // We will ignore the threshold from inline cost, so always get full cost.
1361 Params.ComputeFullInlineCost = true;
1362 Params.AllowRecursiveCall = AllowRecursiveInline;
1363 // Checks if there is anything in the reachable portion of the callee at
1364 // this callsite that makes this inlining potentially illegal. Need to
1365 // set ComputeFullInlineCost, otherwise getInlineCost may return early
1366 // when cost exceeds threshold without checking all IRs in the callee.
1367 // The acutal cost does not matter because we only checks isNever() to
1368 // see if it is legal to inline the callsite.
1369 InlineCost Cost = getInlineCost(Call&: *Candidate.CallInstr, Callee, Params,
1370 CalleeTTI&: GetTTI(*Callee), GetAssumptionCache: GetAC, GetTLI);
1371
1372 // Honor always inline and never inline from call analyzer
1373 if (Cost.isNever() || Cost.isAlways())
1374 return Cost;
1375
1376 // With CSSPGO, the preinliner in llvm-profgen can estimate global inline
1377 // decisions based on hotness as well as accurate function byte sizes for
1378 // given context using function/inlinee sizes from previous build. It
1379 // stores the decision in profile, and also adjust/merge context profile
1380 // aiming at better context-sensitive post-inline profile quality, assuming
1381 // all inline decision estimates are going to be honored by compiler. Here
1382 // we replay that inline decision under `sample-profile-use-preinliner`.
1383 // Note that we don't need to handle negative decision from preinliner as
1384 // context profile for not inlined calls are merged by preinliner already.
1385 if (UsePreInlinerDecision && Candidate.CalleeSamples) {
1386 // Once two node are merged due to promotion, we're losing some context
1387 // so the original context-sensitive preinliner decision should be ignored
1388 // for SyntheticContext.
1389 SampleContext &Context = Candidate.CalleeSamples->getContext();
1390 if (!Context.hasState(S: SyntheticContext) &&
1391 Context.hasAttribute(A: ContextShouldBeInlined))
1392 return InlineCost::getAlways(Reason: "preinliner");
1393 }
1394
1395 // For old FDO inliner, we inline the call site if it is below hot threshold,
1396 // even if the function is hot based on sample profile data. This is to
1397 // prevent huge functions from being inlined.
1398 if (!CallsitePrioritizedInline) {
1399 return InlineCost::get(Cost: Cost.getCost(), Threshold: SampleHotCallSiteThreshold);
1400 }
1401
1402 // Otherwise only use the cost from call analyzer, but overwite threshold with
1403 // Sample PGO threshold.
1404 return InlineCost::get(Cost: Cost.getCost(), Threshold: SampleThreshold);
1405}
1406
1407bool SampleProfileLoader::inlineHotFunctionsWithPriority(
1408 Function &F, DenseSet<GlobalValue::GUID> &InlinedGUIDs) {
1409 // ProfAccForSymsInList is used in callsiteIsHot. The assertion makes sure
1410 // Profile symbol list is ignored when profile-sample-accurate is on.
1411 assert((!ProfAccForSymsInList ||
1412 (!ProfileSampleAccurate &&
1413 !F.hasFnAttribute("profile-sample-accurate"))) &&
1414 "ProfAccForSymsInList should be false when profile-sample-accurate "
1415 "is enabled");
1416
1417 // Populating worklist with initial call sites from root inliner, along
1418 // with call site weights.
1419 CandidateQueue CQueue;
1420 InlineCandidate NewCandidate;
1421 for (auto &BB : F) {
1422 for (auto &I : BB) {
1423 auto *CB = dyn_cast<CallBase>(Val: &I);
1424 if (!CB)
1425 continue;
1426 if (getInlineCandidate(NewCandidate: &NewCandidate, CB))
1427 CQueue.push(x: NewCandidate);
1428 }
1429 }
1430
1431 // Cap the size growth from profile guided inlining. This is needed even
1432 // though cost of each inline candidate already accounts for callee size,
1433 // because with top-down inlining, we can grow inliner size significantly
1434 // with large number of smaller inlinees each pass the cost check.
1435 assert(ProfileInlineLimitMax >= ProfileInlineLimitMin &&
1436 "Max inline size limit should not be smaller than min inline size "
1437 "limit.");
1438 unsigned SizeLimit = F.getInstructionCount() * ProfileInlineGrowthLimit;
1439 SizeLimit = std::min(a: SizeLimit, b: (unsigned)ProfileInlineLimitMax);
1440 SizeLimit = std::max(a: SizeLimit, b: (unsigned)ProfileInlineLimitMin);
1441 if (ExternalInlineAdvisor)
1442 SizeLimit = std::numeric_limits<unsigned>::max();
1443
1444 MapVector<CallBase *, const FunctionSamples *> LocalNotInlinedCallSites;
1445
1446 // Perform iterative BFS call site prioritized inlining
1447 bool Changed = false;
1448 while (!CQueue.empty() && F.getInstructionCount() < SizeLimit) {
1449 InlineCandidate Candidate = CQueue.top();
1450 CQueue.pop();
1451 CallBase *I = Candidate.CallInstr;
1452 Function *CalledFunction = I->getCalledFunction();
1453
1454 if (CalledFunction == &F)
1455 continue;
1456 if (I->isIndirectCall()) {
1457 uint64_t Sum = 0;
1458 auto CalleeSamples = findIndirectCallFunctionSamples(Inst: *I, Sum);
1459 uint64_t SumOrigin = Sum;
1460 Sum *= Candidate.CallsiteDistribution;
1461 unsigned ICPCount = 0;
1462 for (const auto *FS : CalleeSamples) {
1463 // TODO: Consider disable pre-lTO ICP for MonoLTO as well
1464 if (LTOPhase == ThinOrFullLTOPhase::ThinLTOPreLink) {
1465 findExternalInlineCandidate(CB: I, Samples: FS, InlinedGUIDs,
1466 Threshold: PSI->getOrCompHotCountThreshold());
1467 continue;
1468 }
1469 uint64_t EntryCountDistributed =
1470 FS->getHeadSamplesEstimate() * Candidate.CallsiteDistribution;
1471 // In addition to regular inline cost check, we also need to make sure
1472 // ICP isn't introducing excessive speculative checks even if individual
1473 // target looks beneficial to promote and inline. That means we should
1474 // only do ICP when there's a small number dominant targets.
1475 if (ICPCount >= ProfileICPRelativeHotnessSkip &&
1476 EntryCountDistributed * 100 < SumOrigin * ProfileICPRelativeHotness)
1477 break;
1478 // TODO: Fix CallAnalyzer to handle all indirect calls.
1479 // For indirect call, we don't run CallAnalyzer to get InlineCost
1480 // before actual inlining. This is because we could see two different
1481 // types from the same definition, which makes CallAnalyzer choke as
1482 // it's expecting matching parameter type on both caller and callee
1483 // side. See example from PR18962 for the triggering cases (the bug was
1484 // fixed, but we generate different types).
1485 if (!PSI->isHotCount(C: EntryCountDistributed))
1486 break;
1487 SmallVector<CallBase *, 8> InlinedCallSites;
1488 // Attach function profile for promoted indirect callee, and update
1489 // call site count for the promoted inline candidate too.
1490 Candidate = {.CallInstr: I, .CalleeSamples: FS, .CallsiteCount: EntryCountDistributed,
1491 .CallsiteDistribution: Candidate.CallsiteDistribution};
1492 if (tryPromoteAndInlineCandidate(F, Candidate, SumOrigin, Sum,
1493 InlinedCallSite: &InlinedCallSites)) {
1494 for (auto *CB : InlinedCallSites) {
1495 if (getInlineCandidate(NewCandidate: &NewCandidate, CB))
1496 CQueue.emplace(args&: NewCandidate);
1497 }
1498 ICPCount++;
1499 Changed = true;
1500 } else if (!ContextTracker) {
1501 LocalNotInlinedCallSites.insert(KV: {I, FS});
1502 }
1503 }
1504 } else if (CalledFunction && CalledFunction->getSubprogram() &&
1505 !CalledFunction->isDeclaration()) {
1506 SmallVector<CallBase *, 8> InlinedCallSites;
1507 if (tryInlineCandidate(Candidate, InlinedCallSites: &InlinedCallSites)) {
1508 for (auto *CB : InlinedCallSites) {
1509 if (getInlineCandidate(NewCandidate: &NewCandidate, CB))
1510 CQueue.emplace(args&: NewCandidate);
1511 }
1512 Changed = true;
1513 } else if (!ContextTracker) {
1514 LocalNotInlinedCallSites.insert(KV: {I, Candidate.CalleeSamples});
1515 }
1516 } else if (LTOPhase == ThinOrFullLTOPhase::ThinLTOPreLink) {
1517 findExternalInlineCandidate(CB: I, Samples: findCalleeFunctionSamples(Inst: *I),
1518 InlinedGUIDs,
1519 Threshold: PSI->getOrCompHotCountThreshold());
1520 }
1521 }
1522
1523 if (!CQueue.empty()) {
1524 if (SizeLimit == (unsigned)ProfileInlineLimitMax)
1525 ++NumCSInlinedHitMaxLimit;
1526 else if (SizeLimit == (unsigned)ProfileInlineLimitMin)
1527 ++NumCSInlinedHitMinLimit;
1528 else
1529 ++NumCSInlinedHitGrowthLimit;
1530 }
1531
1532 // For CS profile, profile for not inlined context will be merged when
1533 // base profile is being retrieved.
1534 if (!FunctionSamples::ProfileIsCS)
1535 promoteMergeNotInlinedContextSamples(NonInlinedCallSites: LocalNotInlinedCallSites, F);
1536 return Changed;
1537}
1538
1539void SampleProfileLoader::promoteMergeNotInlinedContextSamples(
1540 MapVector<CallBase *, const FunctionSamples *> NonInlinedCallSites,
1541 const Function &F) {
1542 // Accumulate not inlined callsite information into notInlinedSamples
1543 for (const auto &Pair : NonInlinedCallSites) {
1544 CallBase *I = Pair.first;
1545 Function *Callee = I->getCalledFunction();
1546 if (!Callee || Callee->isDeclaration())
1547 continue;
1548
1549 ORE->emit(
1550 OptDiag: OptimizationRemarkAnalysis(getAnnotatedRemarkPassName(), "NotInline",
1551 I->getDebugLoc(), I->getParent())
1552 << "previous inlining not repeated: '" << ore::NV("Callee", Callee)
1553 << "' into '" << ore::NV("Caller", &F) << "'");
1554
1555 ++NumCSNotInlined;
1556 const FunctionSamples *FS = Pair.second;
1557 if (FS->getTotalSamples() == 0 && FS->getHeadSamplesEstimate() == 0) {
1558 continue;
1559 }
1560
1561 // Do not merge a context that is already duplicated into the base profile.
1562 if (FS->getContext().hasAttribute(A: sampleprof::ContextDuplicatedIntoBase))
1563 continue;
1564
1565 if (ProfileMergeInlinee) {
1566 // A function call can be replicated by optimizations like callsite
1567 // splitting or jump threading and the replicates end up sharing the
1568 // sample nested callee profile instead of slicing the original
1569 // inlinee's profile. We want to do merge exactly once by filtering out
1570 // callee profiles with a non-zero head sample count.
1571 if (FS->getHeadSamples() == 0) {
1572 // Use entry samples as head samples during the merge, as inlinees
1573 // don't have head samples.
1574 const_cast<FunctionSamples *>(FS)->addHeadSamples(
1575 Num: FS->getHeadSamplesEstimate());
1576
1577 // Note that we have to do the merge right after processing function.
1578 // This allows OutlineFS's profile to be used for annotation during
1579 // top-down processing of functions' annotation.
1580 FunctionSamples *OutlineFS = Reader->getSamplesFor(F: *Callee);
1581 // If outlined function does not exist in the profile, add it to a
1582 // separate map so that it does not rehash the original profile.
1583 if (!OutlineFS)
1584 OutlineFS = &OutlineFunctionSamples[
1585 FunctionId(FunctionSamples::getCanonicalFnName(FnName: Callee->getName()))];
1586 OutlineFS->merge(Other: *FS, Weight: 1);
1587 // Set outlined profile to be synthetic to not bias the inliner.
1588 OutlineFS->setContextSynthetic();
1589 }
1590 } else {
1591 auto pair =
1592 notInlinedCallInfo.try_emplace(Key: Callee, Args: NotInlinedProfileInfo{.entryCount: 0});
1593 pair.first->second.entryCount += FS->getHeadSamplesEstimate();
1594 }
1595 }
1596}
1597
1598/// Returns the sorted CallTargetMap \p M by count in descending order.
1599static SmallVector<InstrProfValueData, 2>
1600GetSortedValueDataFromCallTargets(const SampleRecord::CallTargetMap &M) {
1601 SmallVector<InstrProfValueData, 2> R;
1602 for (const auto &I : SampleRecord::sortCallTargets(Targets: M)) {
1603 R.emplace_back(
1604 Args: InstrProfValueData{.Value: I.first.getHashCode(), .Count: I.second});
1605 }
1606 return R;
1607}
1608
1609// Generate MD_prof metadata for every branch instruction using the
1610// edge weights computed during propagation.
1611void SampleProfileLoader::generateMDProfMetadata(Function &F) {
1612 // Generate MD_prof metadata for every branch instruction using the
1613 // edge weights computed during propagation.
1614 LLVM_DEBUG(dbgs() << "\nPropagation complete. Setting branch weights\n");
1615 LLVMContext &Ctx = F.getContext();
1616 MDBuilder MDB(Ctx);
1617 for (auto &BI : F) {
1618 BasicBlock *BB = &BI;
1619
1620 if (BlockWeights[BB]) {
1621 for (auto &I : *BB) {
1622 if (!isa<CallInst>(Val: I) && !isa<InvokeInst>(Val: I))
1623 continue;
1624 if (!cast<CallBase>(Val&: I).getCalledFunction()) {
1625 const DebugLoc &DLoc = I.getDebugLoc();
1626 if (!DLoc)
1627 continue;
1628 const DILocation *DIL = DLoc;
1629 const FunctionSamples *FS = findFunctionSamples(Inst: I);
1630 if (!FS)
1631 continue;
1632 auto CallSite = FunctionSamples::getCallSiteIdentifier(DIL);
1633 ErrorOr<SampleRecord::CallTargetMap> T =
1634 FS->findCallTargetMapAt(CallSite);
1635 if (!T || T.get().empty())
1636 continue;
1637 if (FunctionSamples::ProfileIsProbeBased) {
1638 // Prorate the callsite counts based on the pre-ICP distribution
1639 // factor to reflect what is already done to the callsite before
1640 // ICP, such as calliste cloning.
1641 if (std::optional<PseudoProbe> Probe = extractProbe(Inst: I)) {
1642 if (Probe->Factor < 1)
1643 T = SampleRecord::adjustCallTargets(Targets: T.get(), DistributionFactor: Probe->Factor);
1644 }
1645 }
1646 SmallVector<InstrProfValueData, 2> SortedCallTargets =
1647 GetSortedValueDataFromCallTargets(M: T.get());
1648 uint64_t Sum = 0;
1649 for (const auto &C : T.get())
1650 Sum += C.second;
1651 // With CSSPGO all indirect call targets are counted torwards the
1652 // original indirect call site in the profile, including both
1653 // inlined and non-inlined targets.
1654 if (!FunctionSamples::ProfileIsCS) {
1655 if (const FunctionSamplesMap *M =
1656 FS->findFunctionSamplesMapAt(Loc: CallSite)) {
1657 for (const auto &NameFS : *M)
1658 Sum += NameFS.second.getHeadSamplesEstimate();
1659 }
1660 }
1661 if (Sum)
1662 updateIDTMetaData(Inst&: I, CallTargets: SortedCallTargets, Sum);
1663 else if (OverwriteExistingWeights)
1664 I.setMetadata(KindID: LLVMContext::MD_prof, Node: nullptr);
1665 } else if (!isa<IntrinsicInst>(Val: &I)) {
1666 setBranchWeights(
1667 I, Weights: ArrayRef<uint32_t>{static_cast<uint32_t>(BlockWeights[BB])},
1668 /*IsExpected=*/false);
1669 }
1670 }
1671 } else if (OverwriteExistingWeights || ProfileSampleBlockAccurate) {
1672 // Set profile metadata (possibly annotated by LTO prelink) to zero or
1673 // clear it for cold code.
1674 for (auto &I : *BB) {
1675 if (isa<CallInst>(Val: I) || isa<InvokeInst>(Val: I)) {
1676 if (cast<CallBase>(Val&: I).isIndirectCall()) {
1677 I.setMetadata(KindID: LLVMContext::MD_prof, Node: nullptr);
1678 } else {
1679 setBranchWeights(I, Weights: ArrayRef<uint32_t>{uint32_t(0)},
1680 /*IsExpected=*/false);
1681 }
1682 }
1683 }
1684 }
1685
1686 Instruction *TI = BB->getTerminator();
1687 if (TI->getNumSuccessors() == 1)
1688 continue;
1689 if (!isa<CondBrInst>(Val: TI) && !isa<SwitchInst>(Val: TI) &&
1690 !isa<IndirectBrInst>(Val: TI))
1691 continue;
1692
1693 DebugLoc BranchLoc = TI->getDebugLoc();
1694 LLVM_DEBUG(dbgs() << "\nGetting weights for branch at line "
1695 << ((BranchLoc) ? Twine(BranchLoc.getLine())
1696 : Twine("<UNKNOWN LOCATION>"))
1697 << ".\n");
1698 SmallVector<uint32_t, 4> Weights;
1699 uint32_t MaxWeight = 0;
1700 Instruction *MaxDestInst;
1701 // Since profi treats multiple edges (multiway branches) as a single edge,
1702 // we need to distribute the computed weight among the branches. We do
1703 // this by evenly splitting the edge weight among destinations.
1704 DenseMap<const BasicBlock *, uint64_t> EdgeMultiplicity;
1705 std::vector<uint64_t> EdgeIndex;
1706 if (SampleProfileUseProfi) {
1707 EdgeIndex.resize(new_size: TI->getNumSuccessors());
1708 for (unsigned I = 0; I < TI->getNumSuccessors(); ++I) {
1709 const BasicBlock *Succ = TI->getSuccessor(Idx: I);
1710 EdgeIndex[I] = EdgeMultiplicity[Succ];
1711 EdgeMultiplicity[Succ]++;
1712 }
1713 }
1714 for (unsigned I = 0; I < TI->getNumSuccessors(); ++I) {
1715 BasicBlock *Succ = TI->getSuccessor(Idx: I);
1716 Edge E = std::make_pair(x&: BB, y&: Succ);
1717 uint64_t Weight = EdgeWeights[E];
1718 LLVM_DEBUG(dbgs() << "\t"; printEdgeWeight(dbgs(), E));
1719 // Use uint32_t saturated arithmetic to adjust the incoming weights,
1720 // if needed. Sample counts in profiles are 64-bit unsigned values,
1721 // but internally branch weights are expressed as 32-bit values.
1722 if (Weight > std::numeric_limits<uint32_t>::max()) {
1723 LLVM_DEBUG(dbgs() << " (saturated due to uint32_t overflow)\n");
1724 Weight = std::numeric_limits<uint32_t>::max();
1725 }
1726 if (!SampleProfileUseProfi) {
1727 // Weight is added by one to avoid propagation errors introduced by
1728 // 0 weights.
1729 Weights.push_back(Elt: static_cast<uint32_t>(
1730 Weight == std::numeric_limits<uint32_t>::max() ? Weight
1731 : Weight + 1));
1732 } else {
1733 // Profi creates proper weights that do not require "+1" adjustments but
1734 // we evenly split the weight among branches with the same destination.
1735 uint64_t W = Weight / EdgeMultiplicity[Succ];
1736 // Rounding up, if needed, so that first branches are hotter.
1737 if (EdgeIndex[I] < Weight % EdgeMultiplicity[Succ])
1738 W++;
1739 Weights.push_back(Elt: static_cast<uint32_t>(W));
1740 }
1741 if (Weight != 0) {
1742 if (Weight > MaxWeight) {
1743 MaxWeight = Weight;
1744 MaxDestInst = &*Succ->getFirstNonPHIOrDbgOrLifetime();
1745 }
1746 }
1747 }
1748
1749 misexpect::checkExpectAnnotations(I: *TI, ExistingWeights: Weights, /*IsFrontend=*/false);
1750
1751 uint64_t TempWeight;
1752 // Only set weights if there is at least one non-zero weight.
1753 // In any other case, let the analyzer set weights.
1754 // Do not set weights if the weights are present unless under
1755 // OverwriteExistingWeights. In ThinLTO, the profile annotation is done
1756 // twice. If the first annotation already set the weights, the second pass
1757 // does not need to set it. With OverwriteExistingWeights, Blocks with zero
1758 // weight should have their existing metadata (possibly annotated by LTO
1759 // prelink) cleared.
1760 if (MaxWeight > 0 &&
1761 (!TI->extractProfTotalWeight(TotalVal&: TempWeight) || OverwriteExistingWeights)) {
1762 LLVM_DEBUG(dbgs() << "SUCCESS. Found non-zero weights.\n");
1763 setBranchWeights(I&: *TI, Weights, /*IsExpected=*/false);
1764 ORE->emit(RemarkBuilder: [&]() {
1765 return OptimizationRemark(DEBUG_TYPE, "PopularDest", MaxDestInst)
1766 << "most popular destination for conditional branches at "
1767 << ore::NV("CondBranchesLoc", BranchLoc);
1768 });
1769 } else {
1770 if (OverwriteExistingWeights) {
1771 TI->setMetadata(KindID: LLVMContext::MD_prof, Node: nullptr);
1772 LLVM_DEBUG(dbgs() << "CLEARED. All branch weights are zero.\n");
1773 } else {
1774 LLVM_DEBUG(dbgs() << "SKIPPED. All branch weights are zero.\n");
1775 }
1776 }
1777 }
1778}
1779
1780/// Once all the branch weights are computed, we emit the MD_prof
1781/// metadata on BB using the computed values for each of its branches.
1782///
1783/// \param F The function to query.
1784///
1785/// \returns true if \p F was modified. Returns false, otherwise.
1786bool SampleProfileLoader::emitAnnotations(Function &F) {
1787 bool Changed = false;
1788
1789 if (FunctionSamples::ProfileIsProbeBased) {
1790 LLVM_DEBUG({
1791 if (!ProbeManager->getDesc(F))
1792 dbgs() << "Probe descriptor missing for Function " << F.getName()
1793 << "\n";
1794 });
1795
1796 if (ProbeManager->profileIsValid(F, Samples: *Samples)) {
1797 ++NumMatchedProfile;
1798 } else {
1799 ++NumMismatchedProfile;
1800 LLVM_DEBUG(
1801 dbgs() << "Profile is invalid due to CFG mismatch for Function "
1802 << F.getName() << "\n");
1803 if (!SalvageStaleProfile)
1804 return false;
1805 }
1806 } else {
1807 if (getFunctionLoc(F) == 0)
1808 return false;
1809
1810 LLVM_DEBUG(dbgs() << "Line number for the first instruction in "
1811 << F.getName() << ": " << getFunctionLoc(F) << "\n");
1812 }
1813
1814 DenseSet<GlobalValue::GUID> InlinedGUIDs;
1815 if (CallsitePrioritizedInline)
1816 Changed |= inlineHotFunctionsWithPriority(F, InlinedGUIDs);
1817 else
1818 Changed |= inlineHotFunctions(F, InlinedGUIDs);
1819
1820 Changed |= computeAndPropagateWeights(F, InlinedGUIDs);
1821
1822 if (Changed)
1823 generateMDProfMetadata(F);
1824
1825 emitCoverageRemarks(F);
1826 return Changed;
1827}
1828
1829std::unique_ptr<ProfiledCallGraph>
1830SampleProfileLoader::buildProfiledCallGraph(Module &M) {
1831 std::unique_ptr<ProfiledCallGraph> ProfiledCG;
1832 if (FunctionSamples::ProfileIsCS)
1833 ProfiledCG = std::make_unique<ProfiledCallGraph>(args&: *ContextTracker);
1834 else
1835 ProfiledCG = std::make_unique<ProfiledCallGraph>(args&: Reader->getProfiles());
1836
1837 // Add all functions into the profiled call graph even if they are not in
1838 // the profile. This makes sure functions missing from the profile still
1839 // gets a chance to be processed.
1840 for (Function &F : M) {
1841 if (skipProfileForFunction(F))
1842 continue;
1843 ProfiledCG->addProfiledFunction(
1844 Name: getRepInFormat(Name: FunctionSamples::getCanonicalFnName(F)));
1845 }
1846
1847 return ProfiledCG;
1848}
1849
1850std::vector<Function *>
1851SampleProfileLoader::buildFunctionOrder(Module &M, LazyCallGraph &CG) {
1852 std::vector<Function *> FunctionOrderList;
1853 FunctionOrderList.reserve(n: M.size());
1854
1855 if (!ProfileTopDownLoad && UseProfiledCallGraph)
1856 errs() << "WARNING: -use-profiled-call-graph ignored, should be used "
1857 "together with -sample-profile-top-down-load.\n";
1858
1859 if (!ProfileTopDownLoad) {
1860 if (ProfileMergeInlinee) {
1861 // Disable ProfileMergeInlinee if profile is not loaded in top down order,
1862 // because the profile for a function may be used for the profile
1863 // annotation of its outline copy before the profile merging of its
1864 // non-inlined inline instances, and that is not the way how
1865 // ProfileMergeInlinee is supposed to work.
1866 ProfileMergeInlinee = false;
1867 }
1868
1869 for (Function &F : M)
1870 if (!skipProfileForFunction(F))
1871 FunctionOrderList.push_back(x: &F);
1872 return FunctionOrderList;
1873 }
1874
1875 if (UseProfiledCallGraph || (FunctionSamples::ProfileIsCS &&
1876 !UseProfiledCallGraph.getNumOccurrences())) {
1877 // Use profiled call edges to augment the top-down order. There are cases
1878 // that the top-down order computed based on the static call graph doesn't
1879 // reflect real execution order. For example
1880 //
1881 // 1. Incomplete static call graph due to unknown indirect call targets.
1882 // Adjusting the order by considering indirect call edges from the
1883 // profile can enable the inlining of indirect call targets by allowing
1884 // the caller processed before them.
1885 // 2. Mutual call edges in an SCC. The static processing order computed for
1886 // an SCC may not reflect the call contexts in the context-sensitive
1887 // profile, thus may cause potential inlining to be overlooked. The
1888 // function order in one SCC is being adjusted to a top-down order based
1889 // on the profile to favor more inlining. This is only a problem with CS
1890 // profile.
1891 // 3. Transitive indirect call edges due to inlining. When a callee function
1892 // (say B) is inlined into a caller function (say A) in LTO prelink,
1893 // every call edge originated from the callee B will be transferred to
1894 // the caller A. If any transferred edge (say A->C) is indirect, the
1895 // original profiled indirect edge B->C, even if considered, would not
1896 // enforce a top-down order from the caller A to the potential indirect
1897 // call target C in LTO postlink since the inlined callee B is gone from
1898 // the static call graph.
1899 // 4. #3 can happen even for direct call targets, due to functions defined
1900 // in header files. A header function (say A), when included into source
1901 // files, is defined multiple times but only one definition survives due
1902 // to ODR. Therefore, the LTO prelink inlining done on those dropped
1903 // definitions can be useless based on a local file scope. More
1904 // importantly, the inlinee (say B), once fully inlined to a
1905 // to-be-dropped A, will have no profile to consume when its outlined
1906 // version is compiled. This can lead to a profile-less prelink
1907 // compilation for the outlined version of B which may be called from
1908 // external modules. while this isn't easy to fix, we rely on the
1909 // postlink AutoFDO pipeline to optimize B. Since the survived copy of
1910 // the A can be inlined in its local scope in prelink, it may not exist
1911 // in the merged IR in postlink, and we'll need the profiled call edges
1912 // to enforce a top-down order for the rest of the functions.
1913 //
1914 // Considering those cases, a profiled call graph completely independent of
1915 // the static call graph is constructed based on profile data, where
1916 // function objects are not even needed to handle case #3 and case 4.
1917 //
1918 // Note that static callgraph edges are completely ignored since they
1919 // can be conflicting with profiled edges for cyclic SCCs and may result in
1920 // an SCC order incompatible with profile-defined one. Using strictly
1921 // profile order ensures a maximum inlining experience. On the other hand,
1922 // static call edges are not so important when they don't correspond to a
1923 // context in the profile.
1924
1925 std::unique_ptr<ProfiledCallGraph> ProfiledCG = buildProfiledCallGraph(M);
1926 scc_iterator<ProfiledCallGraph *> CGI = scc_begin(G: ProfiledCG.get());
1927 while (!CGI.isAtEnd()) {
1928 auto Range = *CGI;
1929 if (SortProfiledSCC) {
1930 // Sort nodes in one SCC based on callsite hotness.
1931 scc_member_iterator<ProfiledCallGraph *> SI(*CGI);
1932 Range = *SI;
1933 }
1934 for (auto *Node : Range) {
1935 Function *F = SymbolMap.lookup(Key: Node->Name);
1936 if (F && !skipProfileForFunction(F: *F))
1937 FunctionOrderList.push_back(x: F);
1938 }
1939 ++CGI;
1940 }
1941 std::reverse(first: FunctionOrderList.begin(), last: FunctionOrderList.end());
1942 } else
1943 buildTopDownFuncOrder(CG, FunctionOrderList);
1944
1945 LLVM_DEBUG({
1946 dbgs() << "Function processing order:\n";
1947 for (auto F : FunctionOrderList) {
1948 dbgs() << F->getName() << "\n";
1949 }
1950 });
1951
1952 return FunctionOrderList;
1953}
1954
1955bool SampleProfileLoader::doInitialization(Module &M,
1956 FunctionAnalysisManager *FAM) {
1957 auto &Ctx = M.getContext();
1958
1959 auto ReaderOrErr = SampleProfileReader::create(
1960 Filename, C&: Ctx, FS&: *FS, P: FSDiscriminatorPass::Base, RemapFilename: RemappingFilename);
1961 if (std::error_code EC = ReaderOrErr.getError()) {
1962 std::string Msg = "Could not open profile: " + EC.message();
1963 Ctx.diagnose(DI: DiagnosticInfoSampleProfile(Filename, Msg));
1964 return false;
1965 }
1966 Reader = std::move(ReaderOrErr.get());
1967 Reader->setSkipFlatProf(LTOPhase == ThinOrFullLTOPhase::ThinLTOPostLink);
1968 // set module before reading the profile so reader may be able to only
1969 // read the function profiles which are used by the current module.
1970 Reader->setModule(&M);
1971 if (std::error_code EC = Reader->read()) {
1972 std::string Msg = "profile reading failed: " + EC.message();
1973 Ctx.diagnose(DI: DiagnosticInfoSampleProfile(Filename, Msg));
1974 return false;
1975 }
1976
1977 PSL = Reader->getProfileSymbolList();
1978
1979 if (DisableSampleLoaderInlining.getNumOccurrences())
1980 DisableSampleProfileInlining = DisableSampleLoaderInlining;
1981
1982 if (UseFlattenedProfile)
1983 ProfileConverter::flattenProfile(ProfileMap&: Reader->getProfiles(),
1984 ProfileIsCS: Reader->profileIsCS());
1985
1986 // While profile-sample-accurate is on, ignore symbol list.
1987 ProfAccForSymsInList =
1988 ProfileAccurateForSymsInList && PSL && !ProfileSampleAccurate;
1989 if (ProfAccForSymsInList) {
1990 NamesInProfile.clear();
1991 GUIDsInProfile.clear();
1992 auto NameTable = Reader->getNameTable();
1993 if (FunctionSamples::UseMD5) {
1994 for (FunctionId Name : NameTable)
1995 GUIDsInProfile.insert(V: Name.getHashCode());
1996 } else {
1997 for (FunctionId Name : NameTable)
1998 NamesInProfile.insert(key: Name.stringRef());
1999 }
2000 CoverageTracker.setProfAccForSymsInList(true);
2001 }
2002
2003 if (FAM && !ProfileInlineReplayFile.empty()) {
2004 ExternalInlineAdvisor = getReplayInlineAdvisor(
2005 M, FAM&: *FAM, Context&: Ctx, /*OriginalAdvisor=*/nullptr,
2006 ReplaySettings: ReplayInlinerSettings{.ReplayFile: ProfileInlineReplayFile,
2007 .ReplayScope: ProfileInlineReplayScope,
2008 .ReplayFallback: ProfileInlineReplayFallback,
2009 .ReplayFormat: {.OutputFormat: ProfileInlineReplayFormat}},
2010 /*EmitRemarks=*/false, IC: InlineContext{.LTOPhase: LTOPhase, .Pass: InlinePass::ReplaySampleProfileInliner});
2011 }
2012
2013 // Apply tweaks if context-sensitive or probe-based profile is available.
2014 if (Reader->profileIsCS() || Reader->profileIsPreInlined() ||
2015 Reader->profileIsProbeBased()) {
2016 if (!UseIterativeBFIInference.getNumOccurrences())
2017 UseIterativeBFIInference = true;
2018 if (!SampleProfileUseProfi.getNumOccurrences())
2019 SampleProfileUseProfi = true;
2020 if (!EnableExtTspBlockPlacement.getNumOccurrences())
2021 EnableExtTspBlockPlacement = true;
2022 // Enable priority-base inliner and size inline by default for CSSPGO.
2023 if (!ProfileSizeInline.getNumOccurrences())
2024 ProfileSizeInline = true;
2025 if (!CallsitePrioritizedInline.getNumOccurrences())
2026 CallsitePrioritizedInline = true;
2027 // For CSSPGO, we also allow recursive inline to best use context profile.
2028 if (!AllowRecursiveInline.getNumOccurrences())
2029 AllowRecursiveInline = true;
2030
2031 if (Reader->profileIsPreInlined()) {
2032 if (!UsePreInlinerDecision.getNumOccurrences())
2033 UsePreInlinerDecision = true;
2034 }
2035
2036 // Enable stale profile matching by default for probe-based profile.
2037 // Currently the matching relies on if the checksum mismatch is detected,
2038 // which is currently only available for pseudo-probe mode. Removing the
2039 // checksum check could cause regressions for some cases, so further tuning
2040 // might be needed if we want to enable it for all cases.
2041 if (Reader->profileIsProbeBased()) {
2042 if (!SalvageStaleProfile.getNumOccurrences())
2043 SalvageStaleProfile = true;
2044 if (!SalvageUnusedProfile.getNumOccurrences())
2045 SalvageUnusedProfile = true;
2046 }
2047
2048 if (!Reader->profileIsCS()) {
2049 // Non-CS profile should be fine without a function size budget for the
2050 // inliner since the contexts in the profile are either all from inlining
2051 // in the prevoius build or pre-computed by the preinliner with a size
2052 // cap, thus they are bounded.
2053 if (!ProfileInlineLimitMin.getNumOccurrences())
2054 ProfileInlineLimitMin = std::numeric_limits<unsigned>::max();
2055 if (!ProfileInlineLimitMax.getNumOccurrences())
2056 ProfileInlineLimitMax = std::numeric_limits<unsigned>::max();
2057 }
2058 }
2059
2060 if (Reader->profileIsCS()) {
2061 // Tracker for profiles under different context
2062 ContextTracker = std::make_unique<SampleContextTracker>(
2063 args&: Reader->getProfiles(), args: &GUIDToFuncNameMap);
2064 }
2065
2066 // Load pseudo probe descriptors for probe-based function samples.
2067 if (Reader->profileIsProbeBased()) {
2068 ProbeManager = std::make_unique<PseudoProbeManager>(args&: M);
2069 if (!ProbeManager->moduleIsProbed(M)) {
2070 const char *Msg =
2071 "Pseudo-probe-based profile requires SampleProfileProbePass";
2072 Ctx.diagnose(DI: DiagnosticInfoSampleProfile(M.getModuleIdentifier(), Msg,
2073 DS_Warning));
2074 return false;
2075 }
2076 }
2077
2078 if (ReportProfileStaleness || PersistProfileStaleness ||
2079 SalvageStaleProfile) {
2080 MatchingManager = std::make_unique<SampleProfileMatcher>(
2081 args&: M, args&: *Reader, args&: CG, args: ProbeManager.get(), args: LTOPhase, args&: SymbolMap, args&: PSL,
2082 args&: FuncNameToProfNameMap);
2083 }
2084
2085 return true;
2086}
2087
2088// Note that this is a module-level check. Even if one module is errored out,
2089// the entire build will be errored out. However, the user could make big
2090// changes to functions in single module but those changes might not be
2091// performance significant to the whole binary. Therefore, to avoid those false
2092// positives, we select a reasonable big set of hot functions that are supposed
2093// to be globally performance significant, only compute and check the mismatch
2094// within those functions. The function selection is based on two criteria:
2095// 1) The function is hot enough, which is tuned by a hotness-based
2096// flag(HotFuncCutoffForStalenessError). 2) The num of function is large enough
2097// which is tuned by the MinfuncsForStalenessError flag.
2098bool SampleProfileLoader::rejectHighStalenessProfile(
2099 Module &M, ProfileSummaryInfo *PSI, const SampleProfileMap &Profiles) {
2100 assert(FunctionSamples::ProfileIsProbeBased &&
2101 "Only support for probe-based profile");
2102 uint64_t TotalHotFunc = 0;
2103 uint64_t NumMismatchedFunc = 0;
2104 for (const auto &I : Profiles) {
2105 const auto &FS = I.second;
2106 const auto *FuncDesc = ProbeManager->getDesc(GUID: FS.getGUID());
2107 if (!FuncDesc)
2108 continue;
2109
2110 // Use a hotness-based threshold to control the function selection.
2111 if (!PSI->isHotCountNthPercentile(PercentileCutoff: HotFuncCutoffForStalenessError,
2112 C: FS.getTotalSamples()))
2113 continue;
2114
2115 TotalHotFunc++;
2116 if (ProbeManager->profileIsHashMismatched(FuncDesc: *FuncDesc, Samples: FS) &&
2117 !ProbeManager->probeFromWeakSymbol(GUID: FS.getGUID()))
2118 NumMismatchedFunc++;
2119 }
2120 // Make sure that the num of selected function is not too small to distinguish
2121 // from the user's benign changes.
2122 if (TotalHotFunc < MinfuncsForStalenessError)
2123 return false;
2124
2125 // Finally check the mismatch percentage against the threshold.
2126 if (NumMismatchedFunc * 100 >=
2127 TotalHotFunc * PrecentMismatchForStalenessError) {
2128 auto &Ctx = M.getContext();
2129 const char *Msg =
2130 "The input profile significantly mismatches current source code. "
2131 "Please recollect profile to avoid performance regression.";
2132 Ctx.diagnose(DI: DiagnosticInfoSampleProfile(M.getModuleIdentifier(), Msg));
2133 return true;
2134 }
2135 return false;
2136}
2137
2138void SampleProfileLoader::removePseudoProbeInstsDiscriminator(Module &M) {
2139 for (auto &F : M) {
2140 std::vector<Instruction *> InstsToDel;
2141 for (auto &BB : F) {
2142 for (auto &I : BB) {
2143 if (isa<PseudoProbeInst>(Val: &I))
2144 InstsToDel.push_back(x: &I);
2145 else if (isa<CallBase>(Val: &I))
2146 if (const DILocation *DIL = I.getDebugLoc().get()) {
2147 // Restore dwarf discriminator for call.
2148 unsigned Discriminator = DIL->getDiscriminator();
2149 if (DILocation::isPseudoProbeDiscriminator(Discriminator)) {
2150 std::optional<uint32_t> DwarfDiscriminator =
2151 PseudoProbeDwarfDiscriminator::extractDwarfBaseDiscriminator(
2152 Value: Discriminator);
2153 I.setDebugLoc(
2154 DIL->cloneWithDiscriminator(Discriminator: DwarfDiscriminator.value_or(u: 0)));
2155 }
2156 }
2157 }
2158 }
2159 for (auto *I : InstsToDel)
2160 I->eraseFromParent();
2161 }
2162}
2163
2164bool SampleProfileLoader::runOnModule(Module &M, ModuleAnalysisManager &AM,
2165 ProfileSummaryInfo *_PSI) {
2166 GUIDToFuncNameMapper Mapper(M, *Reader, GUIDToFuncNameMap);
2167
2168 PSI = _PSI;
2169 if (M.getProfileSummary(/* IsCS */ false) == nullptr) {
2170 M.setProfileSummary(M: Reader->getSummary().getMD(Context&: M.getContext()),
2171 Kind: ProfileSummary::PSK_Sample);
2172 PSI->refresh();
2173 }
2174
2175 if (FunctionSamples::ProfileIsProbeBased &&
2176 rejectHighStalenessProfile(M, PSI, Profiles: Reader->getProfiles()))
2177 return false;
2178
2179 auto Remapper = Reader->getRemapper();
2180 // Populate the symbol map.
2181 for (const auto &N_F : M.getValueSymbolTable()) {
2182 StringRef OrigName = N_F.getKey();
2183 Function *F = dyn_cast<Function>(Val: N_F.getValue());
2184 if (F == nullptr || OrigName.empty())
2185 continue;
2186 SymbolMap[FunctionId(OrigName)] = F;
2187 StringRef NewName = FunctionSamples::getCanonicalFnName(F: *F);
2188 if (OrigName != NewName && !NewName.empty()) {
2189 auto r = SymbolMap.emplace(Args: FunctionId(NewName), Args&: F);
2190 // Failiing to insert means there is already an entry in SymbolMap,
2191 // thus there are multiple functions that are mapped to the same
2192 // stripped name. In this case of name conflicting, set the value
2193 // to nullptr to avoid confusion.
2194 if (!r.second)
2195 r.first->second = nullptr;
2196 OrigName = NewName;
2197 }
2198 // Insert the remapped names into SymbolMap.
2199 if (Remapper) {
2200 if (auto MapName = Remapper->lookUpNameInProfile(FunctionName: OrigName)) {
2201 if (*MapName != OrigName && !MapName->empty())
2202 SymbolMap.emplace(Args: FunctionId(*MapName), Args&: F);
2203 }
2204 }
2205 }
2206
2207 // Stale profile matching.
2208 if (ReportProfileStaleness || PersistProfileStaleness ||
2209 SalvageStaleProfile) {
2210 MatchingManager->runOnModule();
2211 MatchingManager->clearMatchingData();
2212 }
2213 assert(SymbolMap.count(FunctionId()) == 0 &&
2214 "No empty StringRef should be added in SymbolMap");
2215 assert((SalvageUnusedProfile || FuncNameToProfNameMap.empty()) &&
2216 "FuncNameToProfNameMap is not empty when --salvage-unused-profile is "
2217 "not enabled");
2218
2219 bool retval = false;
2220 for (auto *F : buildFunctionOrder(M, CG)) {
2221 assert(!F->isDeclaration());
2222 clearFunctionData();
2223 retval |= runOnFunction(F&: *F, AM);
2224 }
2225
2226 // Account for cold calls not inlined....
2227 if (!FunctionSamples::ProfileIsCS)
2228 for (const std::pair<Function *, NotInlinedProfileInfo> &pair :
2229 notInlinedCallInfo)
2230 updateProfileCallee(Callee: pair.first, EntryDelta: pair.second.entryCount);
2231
2232 if (RemoveProbeAfterProfileAnnotation &&
2233 FunctionSamples::ProfileIsProbeBased) {
2234 removePseudoProbeInstsDiscriminator(M);
2235 if (auto *FuncInfo = M.getNamedMetadata(Name: PseudoProbeDescMetadataName))
2236 M.eraseNamedMetadata(NMD: FuncInfo);
2237 }
2238
2239 return retval;
2240}
2241
2242bool SampleProfileLoader::runOnFunction(Function &F,
2243 ModuleAnalysisManager &AM) {
2244 LLVM_DEBUG(dbgs() << "\n\nProcessing Function " << F.getName() << "\n");
2245 DILocation2SampleMap.clear();
2246 // By default the entry count is initialized to -1, which will be treated
2247 // conservatively by getEntryCount as the same as unknown (None). This is
2248 // to avoid newly added code to be treated as cold. If we have samples
2249 // this will be overwritten in emitAnnotations.
2250 uint64_t initialEntryCount = -1;
2251
2252 ProfAccForSymsInList = ProfileAccurateForSymsInList && PSL;
2253 if (ProfileSampleAccurate || F.hasFnAttribute(Kind: "profile-sample-accurate")) {
2254 // initialize all the function entry counts to 0. It means all the
2255 // functions without profile will be regarded as cold.
2256 initialEntryCount = 0;
2257 // profile-sample-accurate is a user assertion which has a higher precedence
2258 // than symbol list. When profile-sample-accurate is on, ignore symbol list.
2259 ProfAccForSymsInList = false;
2260 }
2261 CoverageTracker.setProfAccForSymsInList(ProfAccForSymsInList);
2262
2263 // PSL -- profile symbol list include all the symbols in sampled binary.
2264 // If ProfileAccurateForSymsInList is enabled, PSL is used to treat
2265 // old functions without samples being cold, without having to worry
2266 // about new and hot functions being mistakenly treated as cold.
2267 if (ProfAccForSymsInList) {
2268 // Initialize the entry count to 0 for functions in the list.
2269 if (PSL->contains(Name: F.getName()))
2270 initialEntryCount = 0;
2271
2272 // Function in the symbol list but without sample will be regarded as
2273 // cold. To minimize the potential negative performance impact it could
2274 // have, we want to be a little conservative here saying if a function
2275 // shows up in the profile, no matter as outline function, inline instance
2276 // or call targets, treat the function as not being cold. This will handle
2277 // the cases such as most callsites of a function are inlined in sampled
2278 // binary but not inlined in current build (because of source code drift,
2279 // imprecise debug information, or the callsites are all cold individually
2280 // but not cold accumulatively...), so the outline function showing up as
2281 // cold in sampled binary will actually not be cold after current build.
2282 StringRef CanonName = FunctionSamples::getCanonicalFnName(F);
2283 if ((FunctionSamples::UseMD5 &&
2284 GUIDsInProfile.count(
2285 V: Function::getGUIDAssumingExternalLinkage(GlobalName: CanonName))) ||
2286 (!FunctionSamples::UseMD5 && NamesInProfile.count(Key: CanonName)))
2287 initialEntryCount = -1;
2288 }
2289
2290 // Initialize entry count when the function has no existing entry
2291 // count value.
2292 if (!F.getEntryCount())
2293 F.setEntryCount(Count: initialEntryCount);
2294 auto &FAM = AM.getResult<FunctionAnalysisManagerModuleProxy>(IR&: *F.getParent())
2295 .getManager();
2296 ORE = &FAM.getResult<OptimizationRemarkEmitterAnalysis>(IR&: F);
2297
2298 if (FunctionSamples::ProfileIsCS)
2299 Samples = ContextTracker->getBaseSamplesFor(Func: F);
2300 else {
2301 Samples = Reader->getSamplesFor(F);
2302 // Try search in previously inlined functions that were split or duplicated
2303 // into base.
2304 if (!Samples) {
2305 StringRef CanonName = FunctionSamples::getCanonicalFnName(F);
2306 auto It = OutlineFunctionSamples.find(x: FunctionId(CanonName));
2307 if (It != OutlineFunctionSamples.end()) {
2308 Samples = &It->second;
2309 } else if (auto Remapper = Reader->getRemapper()) {
2310 if (auto RemppedName = Remapper->lookUpNameInProfile(FunctionName: CanonName)) {
2311 It = OutlineFunctionSamples.find(x: FunctionId(*RemppedName));
2312 if (It != OutlineFunctionSamples.end())
2313 Samples = &It->second;
2314 }
2315 }
2316 }
2317 }
2318
2319 if (Samples && !Samples->empty())
2320 return emitAnnotations(F);
2321 return false;
2322}
2323SampleProfileLoaderPass::SampleProfileLoaderPass(
2324 std::string File, std::string RemappingFile, ThinOrFullLTOPhase LTOPhase,
2325 IntrusiveRefCntPtr<vfs::FileSystem> FS, bool DisableSampleProfileInlining,
2326 bool UseFlattenedProfile)
2327 : ProfileFileName(File), ProfileRemappingFileName(RemappingFile),
2328 LTOPhase(LTOPhase), FS(std::move(FS)),
2329 DisableSampleProfileInlining(DisableSampleProfileInlining),
2330 UseFlattenedProfile(UseFlattenedProfile) {}
2331
2332PreservedAnalyses SampleProfileLoaderPass::run(Module &M,
2333 ModuleAnalysisManager &AM) {
2334 FunctionAnalysisManager &FAM =
2335 AM.getResult<FunctionAnalysisManagerModuleProxy>(IR&: M).getManager();
2336
2337 auto GetAssumptionCache = [&](Function &F) -> AssumptionCache & {
2338 return FAM.getResult<AssumptionAnalysis>(IR&: F);
2339 };
2340 auto GetTTI = [&](Function &F) -> TargetTransformInfo & {
2341 return FAM.getResult<TargetIRAnalysis>(IR&: F);
2342 };
2343 auto GetTLI = [&](Function &F) -> const TargetLibraryInfo & {
2344 return FAM.getResult<TargetLibraryAnalysis>(IR&: F);
2345 };
2346
2347 if (!FS)
2348 FS = vfs::getRealFileSystem();
2349 LazyCallGraph &CG = AM.getResult<LazyCallGraphAnalysis>(IR&: M);
2350
2351 SampleProfileLoader SampleLoader(
2352 ProfileFileName.empty() ? SampleProfileFile : ProfileFileName,
2353 ProfileRemappingFileName.empty() ? SampleProfileRemappingFile
2354 : ProfileRemappingFileName,
2355 LTOPhase, FS, GetAssumptionCache, GetTTI, GetTLI, CG,
2356 DisableSampleProfileInlining, UseFlattenedProfile);
2357 if (!SampleLoader.doInitialization(M, FAM: &FAM))
2358 return PreservedAnalyses::all();
2359
2360 ProfileSummaryInfo *PSI = &AM.getResult<ProfileSummaryAnalysis>(IR&: M);
2361 if (!SampleLoader.runOnModule(M, AM, PSI: PSI))
2362 return PreservedAnalyses::all();
2363
2364 return PreservedAnalyses::none();
2365}
2366