1 //===- ProfileSummaryInfo.cpp - Global profile summary information --------===//
3 // The LLVM Compiler Infrastructure
5 // This file is distributed under the University of Illinois Open Source
6 // License. See LICENSE.TXT for details.
8 //===----------------------------------------------------------------------===//
10 // This file contains a pass that provides access to the global profile summary
13 //===----------------------------------------------------------------------===//
15 #include "llvm/Analysis/ProfileSummaryInfo.h"
16 #include "llvm/Analysis/BlockFrequencyInfo.h"
17 #include "llvm/IR/BasicBlock.h"
18 #include "llvm/IR/CallSite.h"
19 #include "llvm/IR/Metadata.h"
20 #include "llvm/IR/Module.h"
21 #include "llvm/IR/ProfileSummary.h"
24 // The following two parameters determine the threshold for a count to be
25 // considered hot/cold. These two parameters are percentile values (multiplied
26 // by 10000). If the counts are sorted in descending order, the minimum count to
27 // reach ProfileSummaryCutoffHot gives the threshold to determine a hot count.
28 // Similarly, the minimum count to reach ProfileSummaryCutoffCold gives the
29 // threshold for determining cold count (everything <= this threshold is
32 static cl::opt<int> ProfileSummaryCutoffHot(
33 "profile-summary-cutoff-hot", cl::Hidden, cl::init(990000), cl::ZeroOrMore,
34 cl::desc("A count is hot if it exceeds the minimum count to"
35 " reach this percentile of total counts."));
37 static cl::opt<int> ProfileSummaryCutoffCold(
38 "profile-summary-cutoff-cold", cl::Hidden, cl::init(999999), cl::ZeroOrMore,
39 cl::desc("A count is cold if it is below the minimum count"
40 " to reach this percentile of total counts."));
42 static cl::opt<bool> ProfileSampleAccurate(
43 "profile-sample-accurate", cl::Hidden, cl::init(false),
44 cl::desc("If the sample profile is accurate, we will mark all un-sampled "
45 "callsite as cold. Otherwise, treat un-sampled callsites as if "
46 "we have no profile."));
47 static cl::opt<unsigned> ProfileSummaryHugeWorkingSetSizeThreshold(
48 "profile-summary-huge-working-set-size-threshold", cl::Hidden,
49 cl::init(15000), cl::ZeroOrMore,
50 cl::desc("The code working set size is considered huge if the number of"
51 " blocks required to reach the -profile-summary-cutoff-hot"
52 " percentile exceeds this count."));
54 // Find the summary entry for a desired percentile of counts.
55 static const ProfileSummaryEntry &getEntryForPercentile(SummaryEntryVector &DS,
56 uint64_t Percentile) {
57 auto Compare = [](const ProfileSummaryEntry &Entry, uint64_t Percentile) {
58 return Entry.Cutoff < Percentile;
60 auto It = std::lower_bound(DS.begin(), DS.end(), Percentile, Compare);
61 // The required percentile has to be <= one of the percentiles in the
64 report_fatal_error("Desired percentile exceeds the maximum cutoff");
68 // The profile summary metadata may be attached either by the frontend or by
69 // any backend passes (IR level instrumentation, for example). This method
70 // checks if the Summary is null and if so checks if the summary metadata is now
71 // available in the module and parses it to get the Summary object. Returns true
72 // if a valid Summary is available.
73 bool ProfileSummaryInfo::computeSummary() {
76 auto *SummaryMD = M.getProfileSummary();
79 Summary.reset(ProfileSummary::getFromMD(SummaryMD));
84 ProfileSummaryInfo::getProfileCount(const Instruction *Inst,
85 BlockFrequencyInfo *BFI) {
88 assert((isa<CallInst>(Inst) || isa<InvokeInst>(Inst)) &&
89 "We can only get profile count for call/invoke instruction.");
90 if (hasSampleProfile()) {
91 // In sample PGO mode, check if there is a profile metadata on the
92 // instruction. If it is present, determine hotness solely based on that,
93 // since the sampled entry count may not be accurate. If there is no
94 // annotated on the instruction, return None.
96 if (Inst->extractProfTotalWeight(TotalCount))
101 return BFI->getBlockProfileCount(Inst->getParent());
105 /// Returns true if the function's entry is hot. If it returns false, it
106 /// either means it is not hot or it is unknown whether it is hot or not (for
107 /// example, no profile data is available).
108 bool ProfileSummaryInfo::isFunctionEntryHot(const Function *F) {
109 if (!F || !computeSummary())
111 auto FunctionCount = F->getEntryCount();
112 // FIXME: The heuristic used below for determining hotness is based on
113 // preliminary SPEC tuning for inliner. This will eventually be a
114 // convenience method that calls isHotCount.
115 return FunctionCount && isHotCount(FunctionCount.getValue());
118 /// Returns true if the function contains hot code. This can include a hot
119 /// function entry count, hot basic block, or (in the case of Sample PGO)
120 /// hot total call edge count.
121 /// If it returns false, it either means it is not hot or it is unknown
122 /// (for example, no profile data is available).
123 bool ProfileSummaryInfo::isFunctionHotInCallGraph(const Function *F,
124 BlockFrequencyInfo &BFI) {
125 if (!F || !computeSummary())
127 if (auto FunctionCount = F->getEntryCount())
128 if (isHotCount(FunctionCount.getValue()))
131 if (hasSampleProfile()) {
132 uint64_t TotalCallCount = 0;
133 for (const auto &BB : *F)
134 for (const auto &I : BB)
135 if (isa<CallInst>(I) || isa<InvokeInst>(I))
136 if (auto CallCount = getProfileCount(&I, nullptr))
137 TotalCallCount += CallCount.getValue();
138 if (isHotCount(TotalCallCount))
141 for (const auto &BB : *F)
142 if (isHotBB(&BB, &BFI))
147 /// Returns true if the function only contains cold code. This means that
148 /// the function entry and blocks are all cold, and (in the case of Sample PGO)
149 /// the total call edge count is cold.
150 /// If it returns false, it either means it is not cold or it is unknown
151 /// (for example, no profile data is available).
152 bool ProfileSummaryInfo::isFunctionColdInCallGraph(const Function *F,
153 BlockFrequencyInfo &BFI) {
154 if (!F || !computeSummary())
156 if (auto FunctionCount = F->getEntryCount())
157 if (!isColdCount(FunctionCount.getValue()))
160 if (hasSampleProfile()) {
161 uint64_t TotalCallCount = 0;
162 for (const auto &BB : *F)
163 for (const auto &I : BB)
164 if (isa<CallInst>(I) || isa<InvokeInst>(I))
165 if (auto CallCount = getProfileCount(&I, nullptr))
166 TotalCallCount += CallCount.getValue();
167 if (!isColdCount(TotalCallCount))
170 for (const auto &BB : *F)
171 if (!isColdBB(&BB, &BFI))
176 /// Returns true if the function's entry is a cold. If it returns false, it
177 /// either means it is not cold or it is unknown whether it is cold or not (for
178 /// example, no profile data is available).
179 bool ProfileSummaryInfo::isFunctionEntryCold(const Function *F) {
182 if (F->hasFnAttribute(Attribute::Cold))
184 if (!computeSummary())
186 auto FunctionCount = F->getEntryCount();
187 // FIXME: The heuristic used below for determining coldness is based on
188 // preliminary SPEC tuning for inliner. This will eventually be a
189 // convenience method that calls isHotCount.
190 return FunctionCount && isColdCount(FunctionCount.getValue());
193 /// Compute the hot and cold thresholds.
194 void ProfileSummaryInfo::computeThresholds() {
195 if (!computeSummary())
197 auto &DetailedSummary = Summary->getDetailedSummary();
199 getEntryForPercentile(DetailedSummary, ProfileSummaryCutoffHot);
200 HotCountThreshold = HotEntry.MinCount;
202 getEntryForPercentile(DetailedSummary, ProfileSummaryCutoffCold);
203 ColdCountThreshold = ColdEntry.MinCount;
204 HasHugeWorkingSetSize =
205 HotEntry.NumCounts > ProfileSummaryHugeWorkingSetSizeThreshold;
208 bool ProfileSummaryInfo::hasHugeWorkingSetSize() {
209 if (!HasHugeWorkingSetSize)
211 return HasHugeWorkingSetSize && HasHugeWorkingSetSize.getValue();
214 bool ProfileSummaryInfo::isHotCount(uint64_t C) {
215 if (!HotCountThreshold)
217 return HotCountThreshold && C >= HotCountThreshold.getValue();
220 bool ProfileSummaryInfo::isColdCount(uint64_t C) {
221 if (!ColdCountThreshold)
223 return ColdCountThreshold && C <= ColdCountThreshold.getValue();
226 bool ProfileSummaryInfo::isHotBB(const BasicBlock *B, BlockFrequencyInfo *BFI) {
227 auto Count = BFI->getBlockProfileCount(B);
228 return Count && isHotCount(*Count);
231 bool ProfileSummaryInfo::isColdBB(const BasicBlock *B,
232 BlockFrequencyInfo *BFI) {
233 auto Count = BFI->getBlockProfileCount(B);
234 return Count && isColdCount(*Count);
237 bool ProfileSummaryInfo::isHotCallSite(const CallSite &CS,
238 BlockFrequencyInfo *BFI) {
239 auto C = getProfileCount(CS.getInstruction(), BFI);
240 return C && isHotCount(*C);
243 bool ProfileSummaryInfo::isColdCallSite(const CallSite &CS,
244 BlockFrequencyInfo *BFI) {
245 auto C = getProfileCount(CS.getInstruction(), BFI);
247 return isColdCount(*C);
249 // In SamplePGO, if the caller has been sampled, and there is no profile
250 // annotatedon the callsite, we consider the callsite as cold.
251 // If there is no profile for the caller, and we know the profile is
252 // accurate, we consider the callsite as cold.
253 return (hasSampleProfile() &&
254 (CS.getCaller()->hasProfileData() || ProfileSampleAccurate ||
255 CS.getCaller()->hasFnAttribute("profile-sample-accurate")));
258 INITIALIZE_PASS(ProfileSummaryInfoWrapperPass, "profile-summary-info",
259 "Profile summary info", false, true)
261 ProfileSummaryInfoWrapperPass::ProfileSummaryInfoWrapperPass()
262 : ImmutablePass(ID) {
263 initializeProfileSummaryInfoWrapperPassPass(*PassRegistry::getPassRegistry());
266 bool ProfileSummaryInfoWrapperPass::doInitialization(Module &M) {
267 PSI.reset(new ProfileSummaryInfo(M));
271 bool ProfileSummaryInfoWrapperPass::doFinalization(Module &M) {
276 AnalysisKey ProfileSummaryAnalysis::Key;
277 ProfileSummaryInfo ProfileSummaryAnalysis::run(Module &M,
278 ModuleAnalysisManager &) {
279 return ProfileSummaryInfo(M);
282 PreservedAnalyses ProfileSummaryPrinterPass::run(Module &M,
283 ModuleAnalysisManager &AM) {
284 ProfileSummaryInfo &PSI = AM.getResult<ProfileSummaryAnalysis>(M);
286 OS << "Functions in " << M.getName() << " with hot/cold annotations: \n";
289 if (PSI.isFunctionEntryHot(&F))
290 OS << " :hot entry ";
291 else if (PSI.isFunctionEntryCold(&F))
292 OS << " :cold entry ";
295 return PreservedAnalyses::all();
298 char ProfileSummaryInfoWrapperPass::ID = 0;