|
120
|
1 //===- ProfileSummaryInfo.cpp - Global profile summary information --------===//
|
|
|
2 //
|
|
|
3 // The LLVM Compiler Infrastructure
|
|
|
4 //
|
|
|
5 // This file is distributed under the University of Illinois Open Source
|
|
|
6 // License. See LICENSE.TXT for details.
|
|
|
7 //
|
|
|
8 //===----------------------------------------------------------------------===//
|
|
|
9 //
|
|
|
10 // This file contains a pass that provides access to the global profile summary
|
|
|
11 // information.
|
|
|
12 //
|
|
|
13 //===----------------------------------------------------------------------===//
|
|
|
14
|
|
121
|
15 #include "llvm/Analysis/ProfileSummaryInfo.h"
|
|
120
|
16 #include "llvm/Analysis/BlockFrequencyInfo.h"
|
|
|
17 #include "llvm/IR/BasicBlock.h"
|
|
121
|
18 #include "llvm/IR/CallSite.h"
|
|
120
|
19 #include "llvm/IR/Metadata.h"
|
|
|
20 #include "llvm/IR/Module.h"
|
|
|
21 #include "llvm/IR/ProfileSummary.h"
|
|
|
22 using namespace llvm;
|
|
|
23
|
|
|
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
|
|
|
30 // considered cold).
|
|
|
31
|
|
|
32 static cl::opt<int> ProfileSummaryCutoffHot(
|
|
121
|
33 "profile-summary-cutoff-hot", cl::Hidden, cl::init(990000), cl::ZeroOrMore,
|
|
120
|
34 cl::desc("A count is hot if it exceeds the minimum count to"
|
|
|
35 " reach this percentile of total counts."));
|
|
|
36
|
|
|
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."));
|
|
|
41
|
|
121
|
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."));
|
|
|
53
|
|
|
54 // Find the summary entry for a desired percentile of counts.
|
|
|
55 static const ProfileSummaryEntry &getEntryForPercentile(SummaryEntryVector &DS,
|
|
|
56 uint64_t Percentile) {
|
|
120
|
57 auto Compare = [](const ProfileSummaryEntry &Entry, uint64_t Percentile) {
|
|
|
58 return Entry.Cutoff < Percentile;
|
|
|
59 };
|
|
|
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
|
|
|
62 // detailed summary.
|
|
|
63 if (It == DS.end())
|
|
|
64 report_fatal_error("Desired percentile exceeds the maximum cutoff");
|
|
121
|
65 return *It;
|
|
120
|
66 }
|
|
|
67
|
|
|
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
|
|
121
|
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() {
|
|
120
|
74 if (Summary)
|
|
121
|
75 return true;
|
|
120
|
76 auto *SummaryMD = M.getProfileSummary();
|
|
|
77 if (!SummaryMD)
|
|
121
|
78 return false;
|
|
120
|
79 Summary.reset(ProfileSummary::getFromMD(SummaryMD));
|
|
121
|
80 return true;
|
|
|
81 }
|
|
|
82
|
|
|
83 Optional<uint64_t>
|
|
|
84 ProfileSummaryInfo::getProfileCount(const Instruction *Inst,
|
|
|
85 BlockFrequencyInfo *BFI) {
|
|
|
86 if (!Inst)
|
|
|
87 return None;
|
|
|
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.
|
|
|
95 uint64_t TotalCount;
|
|
|
96 if (Inst->extractProfTotalWeight(TotalCount))
|
|
|
97 return TotalCount;
|
|
|
98 return None;
|
|
|
99 }
|
|
|
100 if (BFI)
|
|
|
101 return BFI->getBlockProfileCount(Inst->getParent());
|
|
|
102 return None;
|
|
120
|
103 }
|
|
|
104
|
|
|
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) {
|
|
121
|
109 if (!F || !computeSummary())
|
|
120
|
110 return false;
|
|
|
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());
|
|
|
116 }
|
|
|
117
|
|
121
|
118 /// Returns true if the function's entry or total call edge count is hot.
|
|
|
119 /// If it returns false, it either means it is not hot or it is unknown
|
|
|
120 /// whether it is hot or not (for example, no profile data is available).
|
|
|
121 bool ProfileSummaryInfo::isFunctionHotInCallGraph(const Function *F) {
|
|
|
122 if (!F || !computeSummary())
|
|
|
123 return false;
|
|
|
124 if (auto FunctionCount = F->getEntryCount())
|
|
|
125 if (isHotCount(FunctionCount.getValue()))
|
|
|
126 return true;
|
|
|
127
|
|
|
128 uint64_t TotalCallCount = 0;
|
|
|
129 for (const auto &BB : *F)
|
|
|
130 for (const auto &I : BB)
|
|
|
131 if (isa<CallInst>(I) || isa<InvokeInst>(I))
|
|
|
132 if (auto CallCount = getProfileCount(&I, nullptr))
|
|
|
133 TotalCallCount += CallCount.getValue();
|
|
|
134 return isHotCount(TotalCallCount);
|
|
|
135 }
|
|
|
136
|
|
|
137 /// Returns true if the function's entry and total call edge count is cold.
|
|
|
138 /// If it returns false, it either means it is not cold or it is unknown
|
|
|
139 /// whether it is cold or not (for example, no profile data is available).
|
|
|
140 bool ProfileSummaryInfo::isFunctionColdInCallGraph(const Function *F) {
|
|
|
141 if (!F || !computeSummary())
|
|
|
142 return false;
|
|
|
143 if (auto FunctionCount = F->getEntryCount())
|
|
|
144 if (!isColdCount(FunctionCount.getValue()))
|
|
|
145 return false;
|
|
|
146
|
|
|
147 uint64_t TotalCallCount = 0;
|
|
|
148 for (const auto &BB : *F)
|
|
|
149 for (const auto &I : BB)
|
|
|
150 if (isa<CallInst>(I) || isa<InvokeInst>(I))
|
|
|
151 if (auto CallCount = getProfileCount(&I, nullptr))
|
|
|
152 TotalCallCount += CallCount.getValue();
|
|
|
153 return isColdCount(TotalCallCount);
|
|
|
154 }
|
|
|
155
|
|
120
|
156 /// Returns true if the function's entry is a cold. If it returns false, it
|
|
|
157 /// either means it is not cold or it is unknown whether it is cold or not (for
|
|
|
158 /// example, no profile data is available).
|
|
|
159 bool ProfileSummaryInfo::isFunctionEntryCold(const Function *F) {
|
|
|
160 if (!F)
|
|
|
161 return false;
|
|
121
|
162 if (F->hasFnAttribute(Attribute::Cold))
|
|
120
|
163 return true;
|
|
121
|
164 if (!computeSummary())
|
|
120
|
165 return false;
|
|
|
166 auto FunctionCount = F->getEntryCount();
|
|
|
167 // FIXME: The heuristic used below for determining coldness is based on
|
|
|
168 // preliminary SPEC tuning for inliner. This will eventually be a
|
|
|
169 // convenience method that calls isHotCount.
|
|
|
170 return FunctionCount && isColdCount(FunctionCount.getValue());
|
|
|
171 }
|
|
|
172
|
|
|
173 /// Compute the hot and cold thresholds.
|
|
|
174 void ProfileSummaryInfo::computeThresholds() {
|
|
121
|
175 if (!computeSummary())
|
|
120
|
176 return;
|
|
|
177 auto &DetailedSummary = Summary->getDetailedSummary();
|
|
121
|
178 auto &HotEntry =
|
|
|
179 getEntryForPercentile(DetailedSummary, ProfileSummaryCutoffHot);
|
|
|
180 HotCountThreshold = HotEntry.MinCount;
|
|
|
181 auto &ColdEntry =
|
|
|
182 getEntryForPercentile(DetailedSummary, ProfileSummaryCutoffCold);
|
|
|
183 ColdCountThreshold = ColdEntry.MinCount;
|
|
|
184 HasHugeWorkingSetSize =
|
|
|
185 HotEntry.NumCounts > ProfileSummaryHugeWorkingSetSizeThreshold;
|
|
|
186 }
|
|
|
187
|
|
|
188 bool ProfileSummaryInfo::hasHugeWorkingSetSize() {
|
|
|
189 if (!HasHugeWorkingSetSize)
|
|
|
190 computeThresholds();
|
|
|
191 return HasHugeWorkingSetSize && HasHugeWorkingSetSize.getValue();
|
|
120
|
192 }
|
|
|
193
|
|
|
194 bool ProfileSummaryInfo::isHotCount(uint64_t C) {
|
|
|
195 if (!HotCountThreshold)
|
|
|
196 computeThresholds();
|
|
|
197 return HotCountThreshold && C >= HotCountThreshold.getValue();
|
|
|
198 }
|
|
|
199
|
|
|
200 bool ProfileSummaryInfo::isColdCount(uint64_t C) {
|
|
|
201 if (!ColdCountThreshold)
|
|
|
202 computeThresholds();
|
|
|
203 return ColdCountThreshold && C <= ColdCountThreshold.getValue();
|
|
|
204 }
|
|
|
205
|
|
|
206 bool ProfileSummaryInfo::isHotBB(const BasicBlock *B, BlockFrequencyInfo *BFI) {
|
|
|
207 auto Count = BFI->getBlockProfileCount(B);
|
|
121
|
208 return Count && isHotCount(*Count);
|
|
|
209 }
|
|
|
210
|
|
|
211 bool ProfileSummaryInfo::isColdBB(const BasicBlock *B,
|
|
|
212 BlockFrequencyInfo *BFI) {
|
|
|
213 auto Count = BFI->getBlockProfileCount(B);
|
|
|
214 return Count && isColdCount(*Count);
|
|
|
215 }
|
|
|
216
|
|
|
217 bool ProfileSummaryInfo::isHotCallSite(const CallSite &CS,
|
|
|
218 BlockFrequencyInfo *BFI) {
|
|
|
219 auto C = getProfileCount(CS.getInstruction(), BFI);
|
|
|
220 return C && isHotCount(*C);
|
|
|
221 }
|
|
|
222
|
|
|
223 bool ProfileSummaryInfo::isColdCallSite(const CallSite &CS,
|
|
|
224 BlockFrequencyInfo *BFI) {
|
|
|
225 auto C = getProfileCount(CS.getInstruction(), BFI);
|
|
|
226 if (C)
|
|
|
227 return isColdCount(*C);
|
|
|
228
|
|
|
229 // In SamplePGO, if the caller has been sampled, and there is no profile
|
|
|
230 // annotatedon the callsite, we consider the callsite as cold.
|
|
|
231 // If there is no profile for the caller, and we know the profile is
|
|
|
232 // accurate, we consider the callsite as cold.
|
|
|
233 return (hasSampleProfile() &&
|
|
|
234 (CS.getCaller()->getEntryCount() || ProfileSampleAccurate ||
|
|
|
235 CS.getCaller()->hasFnAttribute("profile-sample-accurate")));
|
|
120
|
236 }
|
|
|
237
|
|
|
238 INITIALIZE_PASS(ProfileSummaryInfoWrapperPass, "profile-summary-info",
|
|
|
239 "Profile summary info", false, true)
|
|
|
240
|
|
|
241 ProfileSummaryInfoWrapperPass::ProfileSummaryInfoWrapperPass()
|
|
|
242 : ImmutablePass(ID) {
|
|
|
243 initializeProfileSummaryInfoWrapperPassPass(*PassRegistry::getPassRegistry());
|
|
|
244 }
|
|
|
245
|
|
|
246 bool ProfileSummaryInfoWrapperPass::doInitialization(Module &M) {
|
|
|
247 PSI.reset(new ProfileSummaryInfo(M));
|
|
|
248 return false;
|
|
|
249 }
|
|
|
250
|
|
|
251 bool ProfileSummaryInfoWrapperPass::doFinalization(Module &M) {
|
|
|
252 PSI.reset();
|
|
|
253 return false;
|
|
|
254 }
|
|
|
255
|
|
|
256 AnalysisKey ProfileSummaryAnalysis::Key;
|
|
|
257 ProfileSummaryInfo ProfileSummaryAnalysis::run(Module &M,
|
|
|
258 ModuleAnalysisManager &) {
|
|
|
259 return ProfileSummaryInfo(M);
|
|
|
260 }
|
|
|
261
|
|
|
262 PreservedAnalyses ProfileSummaryPrinterPass::run(Module &M,
|
|
|
263 ModuleAnalysisManager &AM) {
|
|
|
264 ProfileSummaryInfo &PSI = AM.getResult<ProfileSummaryAnalysis>(M);
|
|
|
265
|
|
|
266 OS << "Functions in " << M.getName() << " with hot/cold annotations: \n";
|
|
|
267 for (auto &F : M) {
|
|
|
268 OS << F.getName();
|
|
|
269 if (PSI.isFunctionEntryHot(&F))
|
|
|
270 OS << " :hot entry ";
|
|
|
271 else if (PSI.isFunctionEntryCold(&F))
|
|
|
272 OS << " :cold entry ";
|
|
|
273 OS << "\n";
|
|
|
274 }
|
|
|
275 return PreservedAnalyses::all();
|
|
|
276 }
|
|
|
277
|
|
|
278 char ProfileSummaryInfoWrapperPass::ID = 0;
|
