LLVM  9.0.0svn
ProfileSummaryInfo.cpp
Go to the documentation of this file.
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 
16 #include "llvm/IR/BasicBlock.h"
17 #include "llvm/IR/CallSite.h"
18 #include "llvm/IR/Metadata.h"
19 #include "llvm/IR/Module.h"
20 #include "llvm/IR/ProfileSummary.h"
21 using namespace llvm;
22 
23 // The following two parameters determine the threshold for a count to be
24 // considered hot/cold. These two parameters are percentile values (multiplied
25 // by 10000). If the counts are sorted in descending order, the minimum count to
26 // reach ProfileSummaryCutoffHot gives the threshold to determine a hot count.
27 // Similarly, the minimum count to reach ProfileSummaryCutoffCold gives the
28 // threshold for determining cold count (everything <= this threshold is
29 // considered cold).
30 
32  "profile-summary-cutoff-hot", cl::Hidden, cl::init(990000), cl::ZeroOrMore,
33  cl::desc("A count is hot if it exceeds the minimum count to"
34  " reach this percentile of total counts."));
35 
37  "profile-summary-cutoff-cold", cl::Hidden, cl::init(999999), cl::ZeroOrMore,
38  cl::desc("A count is cold if it is below the minimum count"
39  " to reach this percentile of total counts."));
40 
42  "profile-summary-huge-working-set-size-threshold", cl::Hidden,
43  cl::init(15000), cl::ZeroOrMore,
44  cl::desc("The code working set size is considered huge if the number of"
45  " blocks required to reach the -profile-summary-cutoff-hot"
46  " percentile exceeds this count."));
47 
48 // The next two options override the counts derived from summary computation and
49 // are useful for debugging purposes.
51  "profile-summary-hot-count", cl::ReallyHidden, cl::ZeroOrMore,
52  cl::desc("A fixed hot count that overrides the count derived from"
53  " profile-summary-cutoff-hot"));
54 
56  "profile-summary-cold-count", cl::ReallyHidden, cl::ZeroOrMore,
57  cl::desc("A fixed cold count that overrides the count derived from"
58  " profile-summary-cutoff-cold"));
59 
60 // Find the summary entry for a desired percentile of counts.
62  uint64_t Percentile) {
63  auto It = llvm::bsearch(DS, [=](const ProfileSummaryEntry &Entry) {
64  return Percentile <= Entry.Cutoff;
65  });
66  // The required percentile has to be <= one of the percentiles in the
67  // detailed summary.
68  if (It == DS.end())
69  report_fatal_error("Desired percentile exceeds the maximum cutoff");
70  return *It;
71 }
72 
73 // The profile summary metadata may be attached either by the frontend or by
74 // any backend passes (IR level instrumentation, for example). This method
75 // checks if the Summary is null and if so checks if the summary metadata is now
76 // available in the module and parses it to get the Summary object. Returns true
77 // if a valid Summary is available.
78 bool ProfileSummaryInfo::computeSummary() {
79  if (Summary)
80  return true;
81  // First try to get context sensitive ProfileSummary.
82  auto *SummaryMD = M.getProfileSummary(/* IsCS */ true);
83  if (SummaryMD) {
84  Summary.reset(ProfileSummary::getFromMD(SummaryMD));
85  return true;
86  }
87  // This will actually return PSK_Instr or PSK_Sample summary.
88  SummaryMD = M.getProfileSummary(/* IsCS */ false);
89  if (!SummaryMD)
90  return false;
91  Summary.reset(ProfileSummary::getFromMD(SummaryMD));
92  return true;
93 }
94 
98  bool AllowSynthetic) {
99  if (!Inst)
100  return None;
101  assert((isa<CallInst>(Inst) || isa<InvokeInst>(Inst)) &&
102  "We can only get profile count for call/invoke instruction.");
103  if (hasSampleProfile()) {
104  // In sample PGO mode, check if there is a profile metadata on the
105  // instruction. If it is present, determine hotness solely based on that,
106  // since the sampled entry count may not be accurate. If there is no
107  // annotated on the instruction, return None.
108  uint64_t TotalCount;
109  if (Inst->extractProfTotalWeight(TotalCount))
110  return TotalCount;
111  return None;
112  }
113  if (BFI)
114  return BFI->getBlockProfileCount(Inst->getParent(), AllowSynthetic);
115  return None;
116 }
117 
118 /// Returns true if the function's entry is hot. If it returns false, it
119 /// either means it is not hot or it is unknown whether it is hot or not (for
120 /// example, no profile data is available).
122  if (!F || !computeSummary())
123  return false;
124  auto FunctionCount = F->getEntryCount();
125  // FIXME: The heuristic used below for determining hotness is based on
126  // preliminary SPEC tuning for inliner. This will eventually be a
127  // convenience method that calls isHotCount.
128  return FunctionCount && isHotCount(FunctionCount.getCount());
129 }
130 
131 /// Returns true if the function contains hot code. This can include a hot
132 /// function entry count, hot basic block, or (in the case of Sample PGO)
133 /// hot total call edge count.
134 /// If it returns false, it either means it is not hot or it is unknown
135 /// (for example, no profile data is available).
138  if (!F || !computeSummary())
139  return false;
140  if (auto FunctionCount = F->getEntryCount())
141  if (isHotCount(FunctionCount.getCount()))
142  return true;
143 
144  if (hasSampleProfile()) {
145  uint64_t TotalCallCount = 0;
146  for (const auto &BB : *F)
147  for (const auto &I : BB)
148  if (isa<CallInst>(I) || isa<InvokeInst>(I))
149  if (auto CallCount = getProfileCount(&I, nullptr))
150  TotalCallCount += CallCount.getValue();
151  if (isHotCount(TotalCallCount))
152  return true;
153  }
154  for (const auto &BB : *F)
155  if (isHotBlock(&BB, &BFI))
156  return true;
157  return false;
158 }
159 
160 /// Returns true if the function only contains cold code. This means that
161 /// the function entry and blocks are all cold, and (in the case of Sample PGO)
162 /// the total call edge count is cold.
163 /// If it returns false, it either means it is not cold or it is unknown
164 /// (for example, no profile data is available).
167  if (!F || !computeSummary())
168  return false;
169  if (auto FunctionCount = F->getEntryCount())
170  if (!isColdCount(FunctionCount.getCount()))
171  return false;
172 
173  if (hasSampleProfile()) {
174  uint64_t TotalCallCount = 0;
175  for (const auto &BB : *F)
176  for (const auto &I : BB)
177  if (isa<CallInst>(I) || isa<InvokeInst>(I))
178  if (auto CallCount = getProfileCount(&I, nullptr))
179  TotalCallCount += CallCount.getValue();
180  if (!isColdCount(TotalCallCount))
181  return false;
182  }
183  for (const auto &BB : *F)
184  if (!isColdBlock(&BB, &BFI))
185  return false;
186  return true;
187 }
188 
189 /// Returns true if the function's entry is a cold. If it returns false, it
190 /// either means it is not cold or it is unknown whether it is cold or not (for
191 /// example, no profile data is available).
193  if (!F)
194  return false;
196  return true;
197  if (!computeSummary())
198  return false;
199  auto FunctionCount = F->getEntryCount();
200  // FIXME: The heuristic used below for determining coldness is based on
201  // preliminary SPEC tuning for inliner. This will eventually be a
202  // convenience method that calls isHotCount.
203  return FunctionCount && isColdCount(FunctionCount.getCount());
204 }
205 
206 /// Compute the hot and cold thresholds.
207 void ProfileSummaryInfo::computeThresholds() {
208  if (!computeSummary())
209  return;
210  auto &DetailedSummary = Summary->getDetailedSummary();
211  auto &HotEntry =
213  HotCountThreshold = HotEntry.MinCount;
214  if (ProfileSummaryHotCount.getNumOccurrences() > 0)
215  HotCountThreshold = ProfileSummaryHotCount;
216  auto &ColdEntry =
218  ColdCountThreshold = ColdEntry.MinCount;
219  if (ProfileSummaryColdCount.getNumOccurrences() > 0)
220  ColdCountThreshold = ProfileSummaryColdCount;
221  assert(ColdCountThreshold <= HotCountThreshold &&
222  "Cold count threshold cannot exceed hot count threshold!");
223  HasHugeWorkingSetSize =
224  HotEntry.NumCounts > ProfileSummaryHugeWorkingSetSizeThreshold;
225 }
226 
228  if (!HasHugeWorkingSetSize)
229  computeThresholds();
230  return HasHugeWorkingSetSize && HasHugeWorkingSetSize.getValue();
231 }
232 
234  if (!HotCountThreshold)
235  computeThresholds();
236  return HotCountThreshold && C >= HotCountThreshold.getValue();
237 }
238 
240  if (!ColdCountThreshold)
241  computeThresholds();
242  return ColdCountThreshold && C <= ColdCountThreshold.getValue();
243 }
244 
246  if (!HotCountThreshold)
247  computeThresholds();
248  return HotCountThreshold ? HotCountThreshold.getValue() : UINT64_MAX;
249 }
250 
252  if (!ColdCountThreshold)
253  computeThresholds();
254  return ColdCountThreshold ? ColdCountThreshold.getValue() : 0;
255 }
256 
258  auto Count = BFI->getBlockProfileCount(BB);
259  return Count && isHotCount(*Count);
260 }
261 
264  auto Count = BFI->getBlockProfileCount(BB);
265  return Count && isColdCount(*Count);
266 }
267 
270  auto C = getProfileCount(CS.getInstruction(), BFI);
271  return C && isHotCount(*C);
272 }
273 
276  auto C = getProfileCount(CS.getInstruction(), BFI);
277  if (C)
278  return isColdCount(*C);
279 
280  // In SamplePGO, if the caller has been sampled, and there is no profile
281  // annotated on the callsite, we consider the callsite as cold.
282  return hasSampleProfile() && CS.getCaller()->hasProfileData();
283 }
284 
286  "Profile summary info", false, true)
287 
289  : ImmutablePass(ID) {
291 }
292 
294  PSI.reset(new ProfileSummaryInfo(M));
295  return false;
296 }
297 
299  PSI.reset();
300  return false;
301 }
302 
303 AnalysisKey ProfileSummaryAnalysis::Key;
306  return ProfileSummaryInfo(M);
307 }
308 
310  ModuleAnalysisManager &AM) {
312 
313  OS << "Functions in " << M.getName() << " with hot/cold annotations: \n";
314  for (auto &F : M) {
315  OS << F.getName();
316  if (PSI.isFunctionEntryHot(&F))
317  OS << " :hot entry ";
318  else if (PSI.isFunctionEntryCold(&F))
319  OS << " :cold entry ";
320  OS << "\n";
321  }
322  return PreservedAnalyses::all();
323 }
324 
uint64_t CallInst * C
Optional< uint64_t > getBlockProfileCount(const BasicBlock *BB, bool AllowSynthetic=false) const
Returns the estimated profile count of BB.
static PassRegistry * getPassRegistry()
getPassRegistry - Access the global registry object, which is automatically initialized at applicatio...
uint32_t Cutoff
The required percentile of counts.
PassT::Result & getResult(IRUnitT &IR, ExtraArgTs... ExtraArgs)
Get the result of an analysis pass for a given IR unit.
Definition: PassManager.h:776
LLVM_ATTRIBUTE_NORETURN void report_fatal_error(Error Err, bool gen_crash_diag=true)
Report a serious error, calling any installed error handler.
Definition: Error.cpp:139
This class represents lattice values for constants.
Definition: AllocatorList.h:23
A Module instance is used to store all the information related to an LLVM module. ...
Definition: Module.h:65
bool isColdCount(uint64_t C)
Returns true if count C is considered cold.
Analysis providing profile information.
This file contains the declarations for metadata subclasses.
Metadata * getProfileSummary(bool IsCS)
Returns profile summary metadata.
Definition: Module.cpp:541
bool hasFnAttribute(Attribute::AttrKind Kind) const
Return true if the function has the attribute.
Definition: Function.h:323
StringRef getName() const
Get a short "name" for the module.
Definition: Module.h:227
F(f)
uint64_t getOrCompHotCountThreshold()
Returns HotCountThreshold if set.
static cl::opt< unsigned > ProfileSummaryHugeWorkingSetSizeThreshold("profile-summary-huge-working-set-size-threshold", cl::Hidden, cl::init(15000), cl::ZeroOrMore, cl::desc("The code working set size is considered huge if the number of" " blocks required to reach the -profile-summary-cutoff-hot" " percentile exceeds this count."))
bool isHotCount(uint64_t C)
Returns true if count C is considered hot.
bool isFunctionEntryCold(const Function *F)
Returns true if F has cold function entry.
bool isHotCallSite(const CallSite &CS, BlockFrequencyInfo *BFI)
Returns true if CallSite CS is considered hot.
static cl::opt< int > ProfileSummaryHotCount("profile-summary-hot-count", cl::ReallyHidden, cl::ZeroOrMore, cl::desc("A fixed hot count that overrides the count derived from" " profile-summary-cutoff-hot"))
uint64_t getOrCompColdCountThreshold()
Returns ColdCountThreshold if set.
Result run(Module &M, ModuleAnalysisManager &)
bool isFunctionHotInCallGraph(const Function *F, BlockFrequencyInfo &BFI)
Returns true if F contains hot code.
An analysis pass based on legacy pass manager to deliver ProfileSummaryInfo.
InstrTy * getInstruction() const
Definition: CallSite.h:96
bool isColdBlock(const BasicBlock *BB, BlockFrequencyInfo *BFI)
Returns true if BasicBlock BB is considered cold.
#define UINT64_MAX
Definition: DataTypes.h:83
const T & getValue() const LLVM_LVALUE_FUNCTION
Definition: Optional.h:255
bool extractProfTotalWeight(uint64_t &TotalVal) const
Retrieve total raw weight values of a branch.
Definition: Metadata.cpp:1336
size_t bsearch(size_t Lo, size_t Hi, Predicate P)
Binary search for the first index where a predicate is true.
Definition: STLExtras.h:1336
bool doInitialization(Module &M) override
doInitialization - Virtual method overridden by subclasses to do any necessary initialization before ...
BlockFrequencyInfo pass uses BlockFrequencyInfoImpl implementation to estimate IR basic block frequen...
initializer< Ty > init(const Ty &Val)
Definition: CommandLine.h:432
A set of analyses that are preserved following a run of a transformation pass.
Definition: PassManager.h:153
LLVM Basic Block Representation.
Definition: BasicBlock.h:57
bool hasSampleProfile()
Returns true if module M has sample profile.
An analysis pass based on the new PM to deliver ProfileSummaryInfo.
ProfileCount getEntryCount(bool AllowSynthetic=false) const
Get the entry count for this function.
Definition: Function.cpp:1467
INITIALIZE_PASS(ProfileSummaryInfoWrapperPass, "profile-summary-info", "Profile summary info", false, true) ProfileSummaryInfoWrapperPass
bool hasProfileData(bool IncludeSynthetic=false) const
Return true if the function is annotated with profile data.
Definition: Function.h:308
bool doFinalization(Module &M) override
doFinalization - Virtual method overriden by subclasses to do any necessary clean up after all passes...
static PreservedAnalyses all()
Construct a special preserved set that preserves all passes.
Definition: PassManager.h:159
Optional< uint64_t > getProfileCount(const Instruction *CallInst, BlockFrequencyInfo *BFI, bool AllowSynthetic=false)
Returns the profile count for CallInst.
PreservedAnalyses run(Module &M, ModuleAnalysisManager &AM)
ImmutablePass class - This class is used to provide information that does not need to be run...
Definition: Pass.h:255
Module.h This file contains the declarations for the Module class.
void initializeProfileSummaryInfoWrapperPassPass(PassRegistry &)
bool isFunctionColdInCallGraph(const Function *F, BlockFrequencyInfo &BFI)
Returns true if F contains only cold code.
static ProfileSummary * getFromMD(Metadata *MD)
Construct profile summary from metdata.
static cl::opt< int > ProfileSummaryCutoffHot("profile-summary-cutoff-hot", cl::Hidden, cl::init(990000), cl::ZeroOrMore, cl::desc("A count is hot if it exceeds the minimum count to" " reach this percentile of total counts."))
bool isFunctionEntryHot(const Function *F)
Returns true if F has hot function entry.
#define I(x, y, z)
Definition: MD5.cpp:58
bool isHotBlock(const BasicBlock *BB, BlockFrequencyInfo *BFI)
Returns true if BasicBlock BB is considered hot.
bool isColdCallSite(const CallSite &CS, BlockFrequencyInfo *BFI)
Returns true if Callsite CS is considered cold.
assert(ImpDefSCC.getReg()==AMDGPU::SCC &&ImpDefSCC.isDef())
bool hasHugeWorkingSetSize()
Returns true if the working set size of the code is considered huge.
static cl::opt< int > ProfileSummaryCutoffCold("profile-summary-cutoff-cold", cl::Hidden, cl::init(999999), cl::ZeroOrMore, cl::desc("A count is cold if it is below the minimum count" " to reach this percentile of total counts."))
A container for analyses that lazily runs them and caches their results.
std::vector< ProfileSummaryEntry > SummaryEntryVector
static cl::opt< int > ProfileSummaryColdCount("profile-summary-cold-count", cl::ReallyHidden, cl::ZeroOrMore, cl::desc("A fixed cold count that overrides the count derived from" " profile-summary-cutoff-cold"))
A special type used by analysis passes to provide an address that identifies that particular analysis...
Definition: PassManager.h:70
static const ProfileSummaryEntry & getEntryForPercentile(SummaryEntryVector &DS, uint64_t Percentile)
const BasicBlock * getParent() const
Definition: Instruction.h:66
FunTy * getCaller() const
Return the caller function for this call site.
Definition: CallSite.h:275