1 //===- ProfileSummaryInfo.cpp - Global profile summary information --------===// 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 contains a pass that provides access to the global profile summary 10 // information. 11 // 12 //===----------------------------------------------------------------------===// 13 14 #include "llvm/Analysis/ProfileSummaryInfo.h" 15 #include "llvm/Analysis/BlockFrequencyInfo.h" 16 #include "llvm/IR/BasicBlock.h" 17 #include "llvm/IR/Instructions.h" 18 #include "llvm/IR/Module.h" 19 #include "llvm/IR/ProfileSummary.h" 20 #include "llvm/InitializePasses.h" 21 #include "llvm/ProfileData/ProfileCommon.h" 22 #include "llvm/Support/CommandLine.h" 23 #include <optional> 24 using namespace llvm; 25 26 static cl::opt<bool> PartialProfile( 27 "partial-profile", cl::Hidden, cl::init(false), 28 cl::desc("Specify the current profile is used as a partial profile.")); 29 30 cl::opt<bool> ScalePartialSampleProfileWorkingSetSize( 31 "scale-partial-sample-profile-working-set-size", cl::Hidden, cl::init(true), 32 cl::desc( 33 "If true, scale the working set size of the partial sample profile " 34 "by the partial profile ratio to reflect the size of the program " 35 "being compiled.")); 36 37 static cl::opt<double> PartialSampleProfileWorkingSetSizeScaleFactor( 38 "partial-sample-profile-working-set-size-scale-factor", cl::Hidden, 39 cl::init(0.008), 40 cl::desc("The scale factor used to scale the working set size of the " 41 "partial sample profile along with the partial profile ratio. " 42 "This includes the factor of the profile counter per block " 43 "and the factor to scale the working set size to use the same " 44 "shared thresholds as PGO.")); 45 46 // The profile summary metadata may be attached either by the frontend or by 47 // any backend passes (IR level instrumentation, for example). This method 48 // checks if the Summary is null and if so checks if the summary metadata is now 49 // available in the module and parses it to get the Summary object. 50 void ProfileSummaryInfo::refresh() { 51 if (hasProfileSummary()) 52 return; 53 // First try to get context sensitive ProfileSummary. 54 auto *SummaryMD = M->getProfileSummary(/* IsCS */ true); 55 if (SummaryMD) 56 Summary.reset(ProfileSummary::getFromMD(SummaryMD)); 57 58 if (!hasProfileSummary()) { 59 // This will actually return PSK_Instr or PSK_Sample summary. 60 SummaryMD = M->getProfileSummary(/* IsCS */ false); 61 if (SummaryMD) 62 Summary.reset(ProfileSummary::getFromMD(SummaryMD)); 63 } 64 if (!hasProfileSummary()) 65 return; 66 computeThresholds(); 67 } 68 69 std::optional<uint64_t> ProfileSummaryInfo::getProfileCount( 70 const CallBase &Call, BlockFrequencyInfo *BFI, bool AllowSynthetic) const { 71 assert((isa<CallInst>(Call) || isa<InvokeInst>(Call)) && 72 "We can only get profile count for call/invoke instruction."); 73 if (hasSampleProfile()) { 74 // In sample PGO mode, check if there is a profile metadata on the 75 // instruction. If it is present, determine hotness solely based on that, 76 // since the sampled entry count may not be accurate. If there is no 77 // annotated on the instruction, return std::nullopt. 78 uint64_t TotalCount; 79 if (Call.extractProfTotalWeight(TotalCount)) 80 return TotalCount; 81 return std::nullopt; 82 } 83 if (BFI) 84 return BFI->getBlockProfileCount(Call.getParent(), AllowSynthetic); 85 return std::nullopt; 86 } 87 88 bool ProfileSummaryInfo::isFunctionHotnessUnknown(const Function &F) const { 89 assert(hasPartialSampleProfile() && "Expect partial sample profile"); 90 return !F.getEntryCount(); 91 } 92 93 /// Returns true if the function's entry is a cold. If it returns false, it 94 /// either means it is not cold or it is unknown whether it is cold or not (for 95 /// example, no profile data is available). 96 bool ProfileSummaryInfo::isFunctionEntryCold(const Function *F) const { 97 if (!F) 98 return false; 99 if (F->hasFnAttribute(Attribute::Cold)) 100 return true; 101 if (!hasProfileSummary()) 102 return false; 103 auto FunctionCount = F->getEntryCount(); 104 // FIXME: The heuristic used below for determining coldness is based on 105 // preliminary SPEC tuning for inliner. This will eventually be a 106 // convenience method that calls isHotCount. 107 return FunctionCount && isColdCount(FunctionCount->getCount()); 108 } 109 110 /// Compute the hot and cold thresholds. 111 void ProfileSummaryInfo::computeThresholds() { 112 auto &DetailedSummary = Summary->getDetailedSummary(); 113 auto &HotEntry = ProfileSummaryBuilder::getEntryForPercentile( 114 DetailedSummary, ProfileSummaryCutoffHot); 115 HotCountThreshold = 116 ProfileSummaryBuilder::getHotCountThreshold(DetailedSummary); 117 ColdCountThreshold = 118 ProfileSummaryBuilder::getColdCountThreshold(DetailedSummary); 119 assert(ColdCountThreshold <= HotCountThreshold && 120 "Cold count threshold cannot exceed hot count threshold!"); 121 if (!hasPartialSampleProfile() || !ScalePartialSampleProfileWorkingSetSize) { 122 HasHugeWorkingSetSize = 123 HotEntry.NumCounts > ProfileSummaryHugeWorkingSetSizeThreshold; 124 HasLargeWorkingSetSize = 125 HotEntry.NumCounts > ProfileSummaryLargeWorkingSetSizeThreshold; 126 } else { 127 // Scale the working set size of the partial sample profile to reflect the 128 // size of the program being compiled. 129 double PartialProfileRatio = Summary->getPartialProfileRatio(); 130 uint64_t ScaledHotEntryNumCounts = 131 static_cast<uint64_t>(HotEntry.NumCounts * PartialProfileRatio * 132 PartialSampleProfileWorkingSetSizeScaleFactor); 133 HasHugeWorkingSetSize = 134 ScaledHotEntryNumCounts > ProfileSummaryHugeWorkingSetSizeThreshold; 135 HasLargeWorkingSetSize = 136 ScaledHotEntryNumCounts > ProfileSummaryLargeWorkingSetSizeThreshold; 137 } 138 } 139 140 std::optional<uint64_t> 141 ProfileSummaryInfo::computeThreshold(int PercentileCutoff) const { 142 if (!hasProfileSummary()) 143 return std::nullopt; 144 auto iter = ThresholdCache.find(PercentileCutoff); 145 if (iter != ThresholdCache.end()) { 146 return iter->second; 147 } 148 auto &DetailedSummary = Summary->getDetailedSummary(); 149 auto &Entry = ProfileSummaryBuilder::getEntryForPercentile(DetailedSummary, 150 PercentileCutoff); 151 uint64_t CountThreshold = Entry.MinCount; 152 ThresholdCache[PercentileCutoff] = CountThreshold; 153 return CountThreshold; 154 } 155 156 bool ProfileSummaryInfo::hasHugeWorkingSetSize() const { 157 return HasHugeWorkingSetSize && *HasHugeWorkingSetSize; 158 } 159 160 bool ProfileSummaryInfo::hasLargeWorkingSetSize() const { 161 return HasLargeWorkingSetSize && *HasLargeWorkingSetSize; 162 } 163 164 bool ProfileSummaryInfo::isHotCount(uint64_t C) const { 165 return HotCountThreshold && C >= *HotCountThreshold; 166 } 167 168 bool ProfileSummaryInfo::isColdCount(uint64_t C) const { 169 return ColdCountThreshold && C <= *ColdCountThreshold; 170 } 171 172 template <bool isHot> 173 bool ProfileSummaryInfo::isHotOrColdCountNthPercentile(int PercentileCutoff, 174 uint64_t C) const { 175 auto CountThreshold = computeThreshold(PercentileCutoff); 176 if (isHot) 177 return CountThreshold && C >= *CountThreshold; 178 else 179 return CountThreshold && C <= *CountThreshold; 180 } 181 182 bool ProfileSummaryInfo::isHotCountNthPercentile(int PercentileCutoff, 183 uint64_t C) const { 184 return isHotOrColdCountNthPercentile<true>(PercentileCutoff, C); 185 } 186 187 bool ProfileSummaryInfo::isColdCountNthPercentile(int PercentileCutoff, 188 uint64_t C) const { 189 return isHotOrColdCountNthPercentile<false>(PercentileCutoff, C); 190 } 191 192 uint64_t ProfileSummaryInfo::getOrCompHotCountThreshold() const { 193 return HotCountThreshold.value_or(UINT64_MAX); 194 } 195 196 uint64_t ProfileSummaryInfo::getOrCompColdCountThreshold() const { 197 return ColdCountThreshold.value_or(0); 198 } 199 200 bool ProfileSummaryInfo::isHotCallSite(const CallBase &CB, 201 BlockFrequencyInfo *BFI) const { 202 auto C = getProfileCount(CB, BFI); 203 return C && isHotCount(*C); 204 } 205 206 bool ProfileSummaryInfo::isColdCallSite(const CallBase &CB, 207 BlockFrequencyInfo *BFI) const { 208 auto C = getProfileCount(CB, BFI); 209 if (C) 210 return isColdCount(*C); 211 212 // In SamplePGO, if the caller has been sampled, and there is no profile 213 // annotated on the callsite, we consider the callsite as cold. 214 return hasSampleProfile() && CB.getCaller()->hasProfileData(); 215 } 216 217 bool ProfileSummaryInfo::hasPartialSampleProfile() const { 218 return hasProfileSummary() && 219 Summary->getKind() == ProfileSummary::PSK_Sample && 220 (PartialProfile || Summary->isPartialProfile()); 221 } 222 223 INITIALIZE_PASS(ProfileSummaryInfoWrapperPass, "profile-summary-info", 224 "Profile summary info", false, true) 225 226 ProfileSummaryInfoWrapperPass::ProfileSummaryInfoWrapperPass() 227 : ImmutablePass(ID) { 228 initializeProfileSummaryInfoWrapperPassPass(*PassRegistry::getPassRegistry()); 229 } 230 231 bool ProfileSummaryInfoWrapperPass::doInitialization(Module &M) { 232 PSI.reset(new ProfileSummaryInfo(M)); 233 return false; 234 } 235 236 bool ProfileSummaryInfoWrapperPass::doFinalization(Module &M) { 237 PSI.reset(); 238 return false; 239 } 240 241 AnalysisKey ProfileSummaryAnalysis::Key; 242 ProfileSummaryInfo ProfileSummaryAnalysis::run(Module &M, 243 ModuleAnalysisManager &) { 244 return ProfileSummaryInfo(M); 245 } 246 247 PreservedAnalyses ProfileSummaryPrinterPass::run(Module &M, 248 ModuleAnalysisManager &AM) { 249 ProfileSummaryInfo &PSI = AM.getResult<ProfileSummaryAnalysis>(M); 250 251 OS << "Functions in " << M.getName() << " with hot/cold annotations: \n"; 252 for (auto &F : M) { 253 OS << F.getName(); 254 if (PSI.isFunctionEntryHot(&F)) 255 OS << " :hot entry "; 256 else if (PSI.isFunctionEntryCold(&F)) 257 OS << " :cold entry "; 258 OS << "\n"; 259 } 260 return PreservedAnalyses::all(); 261 } 262 263 char ProfileSummaryInfoWrapperPass::ID = 0; 264