1 //===- SampleProf.h - Sampling profiling format support ---------*- C++ -*-===//
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
7 //===----------------------------------------------------------------------===//
9 // This file contains common definitions used in the reading and writing of
10 // sample profile data.
12 //===----------------------------------------------------------------------===//
14 #ifndef LLVM_PROFILEDATA_SAMPLEPROF_H
15 #define LLVM_PROFILEDATA_SAMPLEPROF_H
17 #include "llvm/ADT/DenseSet.h"
18 #include "llvm/ADT/SmallVector.h"
19 #include "llvm/ADT/StringMap.h"
20 #include "llvm/ADT/StringRef.h"
21 #include "llvm/ADT/StringSet.h"
22 #include "llvm/IR/Function.h"
23 #include "llvm/IR/GlobalValue.h"
24 #include "llvm/IR/Module.h"
25 #include "llvm/Support/Debug.h"
26 #include "llvm/Support/ErrorOr.h"
27 #include "llvm/Support/MathExtras.h"
28 #include "llvm/Support/raw_ostream.h"
34 #include <system_error>
41 const std::error_category &sampleprof_category();
43 enum class sampleprof_error {
51 unsupported_writing_format,
55 ostream_seek_unsupported,
61 inline std::error_code make_error_code(sampleprof_error E) {
62 return std::error_code(static_cast<int>(E), sampleprof_category());
65 inline sampleprof_error MergeResult(sampleprof_error &Accumulator,
66 sampleprof_error Result) {
67 // Prefer first error encountered as later errors may be secondary effects of
68 // the initial problem.
69 if (Accumulator == sampleprof_error::success &&
70 Result != sampleprof_error::success)
75 } // end namespace llvm
80 struct is_error_code_enum<llvm::sampleprof_error> : std::true_type {};
82 } // end namespace std
85 namespace sampleprof {
87 enum SampleProfileFormat {
90 SPF_Compact_Binary = 0x2,
96 static inline uint64_t SPMagic(SampleProfileFormat Format = SPF_Binary) {
97 return uint64_t('S') << (64 - 8) | uint64_t('P') << (64 - 16) |
98 uint64_t('R') << (64 - 24) | uint64_t('O') << (64 - 32) |
99 uint64_t('F') << (64 - 40) | uint64_t('4') << (64 - 48) |
100 uint64_t('2') << (64 - 56) | uint64_t(Format);
103 // Get the proper representation of a string in the input Format.
104 static inline StringRef getRepInFormat(StringRef Name,
105 SampleProfileFormat Format,
106 std::string &GUIDBuf) {
109 GUIDBuf = std::to_string(Function::getGUID(Name));
110 return (Format == SPF_Compact_Binary) ? StringRef(GUIDBuf) : Name;
113 static inline uint64_t SPVersion() { return 103; }
115 // Section Type used by SampleProfileExtBinaryBaseReader and
116 // SampleProfileExtBinaryBaseWriter. Never change the existing
117 // value of enum. Only append new ones.
122 SecProfileSymbolList = 3,
123 SecFuncOffsetTable = 4,
124 // marker for the first type of profile.
125 SecFuncProfileFirst = 32,
126 SecLBRProfile = SecFuncProfileFirst
129 static inline std::string getSecName(SecType Type) {
132 return "InvalidSection";
134 return "ProfileSummarySection";
136 return "NameTableSection";
137 case SecProfileSymbolList:
138 return "ProfileSymbolListSection";
139 case SecFuncOffsetTable:
140 return "FuncOffsetTableSection";
142 return "LBRProfileSection";
144 llvm_unreachable("A SecType has no name for output");
147 // Entry type of section header table used by SampleProfileExtBinaryBaseReader
148 // and SampleProfileExtBinaryBaseWriter.
149 struct SecHdrTableEntry {
156 enum SecFlags { SecFlagInValid = 0, SecFlagCompress = (1 << 0) };
158 static inline void addSecFlags(SecHdrTableEntry &Entry, uint64_t Flags) {
159 Entry.Flags |= Flags;
162 static inline void removeSecFlags(SecHdrTableEntry &Entry, uint64_t Flags) {
163 Entry.Flags &= ~Flags;
166 static inline bool hasSecFlag(SecHdrTableEntry &Entry, SecFlags Flag) {
167 return Entry.Flags & Flag;
170 /// Represents the relative location of an instruction.
172 /// Instruction locations are specified by the line offset from the
173 /// beginning of the function (marked by the line where the function
174 /// header is) and the discriminator value within that line.
176 /// The discriminator value is useful to distinguish instructions
177 /// that are on the same line but belong to different basic blocks
178 /// (e.g., the two post-increment instructions in "if (p) x++; else y++;").
179 struct LineLocation {
180 LineLocation(uint32_t L, uint32_t D) : LineOffset(L), Discriminator(D) {}
182 void print(raw_ostream &OS) const;
185 bool operator<(const LineLocation &O) const {
186 return LineOffset < O.LineOffset ||
187 (LineOffset == O.LineOffset && Discriminator < O.Discriminator);
191 uint32_t Discriminator;
194 raw_ostream &operator<<(raw_ostream &OS, const LineLocation &Loc);
196 /// Representation of a single sample record.
198 /// A sample record is represented by a positive integer value, which
199 /// indicates how frequently was the associated line location executed.
201 /// Additionally, if the associated location contains a function call,
202 /// the record will hold a list of all the possible called targets. For
203 /// direct calls, this will be the exact function being invoked. For
204 /// indirect calls (function pointers, virtual table dispatch), this
205 /// will be a list of one or more functions.
208 using CallTarget = std::pair<StringRef, uint64_t>;
209 struct CallTargetComparator {
210 bool operator()(const CallTarget &LHS, const CallTarget &RHS) const {
211 if (LHS.second != RHS.second)
212 return LHS.second > RHS.second;
214 return LHS.first < RHS.first;
218 using SortedCallTargetSet = std::set<CallTarget, CallTargetComparator>;
219 using CallTargetMap = StringMap<uint64_t>;
220 SampleRecord() = default;
222 /// Increment the number of samples for this record by \p S.
223 /// Optionally scale sample count \p S by \p Weight.
225 /// Sample counts accumulate using saturating arithmetic, to avoid wrapping
226 /// around unsigned integers.
227 sampleprof_error addSamples(uint64_t S, uint64_t Weight = 1) {
229 NumSamples = SaturatingMultiplyAdd(S, Weight, NumSamples, &Overflowed);
230 return Overflowed ? sampleprof_error::counter_overflow
231 : sampleprof_error::success;
234 /// Add called function \p F with samples \p S.
235 /// Optionally scale sample count \p S by \p Weight.
237 /// Sample counts accumulate using saturating arithmetic, to avoid wrapping
238 /// around unsigned integers.
239 sampleprof_error addCalledTarget(StringRef F, uint64_t S,
240 uint64_t Weight = 1) {
241 uint64_t &TargetSamples = CallTargets[F];
244 SaturatingMultiplyAdd(S, Weight, TargetSamples, &Overflowed);
245 return Overflowed ? sampleprof_error::counter_overflow
246 : sampleprof_error::success;
249 /// Return true if this sample record contains function calls.
250 bool hasCalls() const { return !CallTargets.empty(); }
252 uint64_t getSamples() const { return NumSamples; }
253 const CallTargetMap &getCallTargets() const { return CallTargets; }
254 const SortedCallTargetSet getSortedCallTargets() const {
255 return SortCallTargets(CallTargets);
258 /// Sort call targets in descending order of call frequency.
259 static const SortedCallTargetSet SortCallTargets(const CallTargetMap &Targets) {
260 SortedCallTargetSet SortedTargets;
261 for (const auto &I : Targets) {
262 SortedTargets.emplace(I.first(), I.second);
264 return SortedTargets;
267 /// Merge the samples in \p Other into this record.
268 /// Optionally scale sample counts by \p Weight.
269 sampleprof_error merge(const SampleRecord &Other, uint64_t Weight = 1) {
270 sampleprof_error Result = addSamples(Other.getSamples(), Weight);
271 for (const auto &I : Other.getCallTargets()) {
272 MergeResult(Result, addCalledTarget(I.first(), I.second, Weight));
277 void print(raw_ostream &OS, unsigned Indent) const;
281 uint64_t NumSamples = 0;
282 CallTargetMap CallTargets;
285 raw_ostream &operator<<(raw_ostream &OS, const SampleRecord &Sample);
287 class FunctionSamples;
289 using BodySampleMap = std::map<LineLocation, SampleRecord>;
290 // NOTE: Using a StringMap here makes parsed profiles consume around 17% more
291 // memory, which is *very* significant for large profiles.
292 using FunctionSamplesMap = std::map<std::string, FunctionSamples, std::less<>>;
293 using CallsiteSampleMap = std::map<LineLocation, FunctionSamplesMap>;
295 /// Representation of the samples collected for a function.
297 /// This data structure contains all the collected samples for the body
298 /// of a function. Each sample corresponds to a LineLocation instance
299 /// within the body of the function.
300 class FunctionSamples {
302 FunctionSamples() = default;
304 void print(raw_ostream &OS = dbgs(), unsigned Indent = 0) const;
307 sampleprof_error addTotalSamples(uint64_t Num, uint64_t Weight = 1) {
310 SaturatingMultiplyAdd(Num, Weight, TotalSamples, &Overflowed);
311 return Overflowed ? sampleprof_error::counter_overflow
312 : sampleprof_error::success;
315 sampleprof_error addHeadSamples(uint64_t Num, uint64_t Weight = 1) {
318 SaturatingMultiplyAdd(Num, Weight, TotalHeadSamples, &Overflowed);
319 return Overflowed ? sampleprof_error::counter_overflow
320 : sampleprof_error::success;
323 sampleprof_error addBodySamples(uint32_t LineOffset, uint32_t Discriminator,
324 uint64_t Num, uint64_t Weight = 1) {
325 return BodySamples[LineLocation(LineOffset, Discriminator)].addSamples(
329 sampleprof_error addCalledTargetSamples(uint32_t LineOffset,
330 uint32_t Discriminator,
331 StringRef FName, uint64_t Num,
332 uint64_t Weight = 1) {
333 return BodySamples[LineLocation(LineOffset, Discriminator)].addCalledTarget(
337 /// Return the number of samples collected at the given location.
338 /// Each location is specified by \p LineOffset and \p Discriminator.
339 /// If the location is not found in profile, return error.
340 ErrorOr<uint64_t> findSamplesAt(uint32_t LineOffset,
341 uint32_t Discriminator) const {
342 const auto &ret = BodySamples.find(LineLocation(LineOffset, Discriminator));
343 if (ret == BodySamples.end())
344 return std::error_code();
346 return ret->second.getSamples();
349 /// Returns the call target map collected at a given location.
350 /// Each location is specified by \p LineOffset and \p Discriminator.
351 /// If the location is not found in profile, return error.
352 ErrorOr<SampleRecord::CallTargetMap>
353 findCallTargetMapAt(uint32_t LineOffset, uint32_t Discriminator) const {
354 const auto &ret = BodySamples.find(LineLocation(LineOffset, Discriminator));
355 if (ret == BodySamples.end())
356 return std::error_code();
357 return ret->second.getCallTargets();
360 /// Return the function samples at the given callsite location.
361 FunctionSamplesMap &functionSamplesAt(const LineLocation &Loc) {
362 return CallsiteSamples[Loc];
365 /// Returns the FunctionSamplesMap at the given \p Loc.
366 const FunctionSamplesMap *
367 findFunctionSamplesMapAt(const LineLocation &Loc) const {
368 auto iter = CallsiteSamples.find(Loc);
369 if (iter == CallsiteSamples.end())
371 return &iter->second;
374 /// Returns a pointer to FunctionSamples at the given callsite location \p Loc
375 /// with callee \p CalleeName. If no callsite can be found, relax the
376 /// restriction to return the FunctionSamples at callsite location \p Loc
377 /// with the maximum total sample count.
378 const FunctionSamples *findFunctionSamplesAt(const LineLocation &Loc,
379 StringRef CalleeName) const {
380 std::string CalleeGUID;
381 CalleeName = getRepInFormat(CalleeName, Format, CalleeGUID);
383 auto iter = CallsiteSamples.find(Loc);
384 if (iter == CallsiteSamples.end())
386 auto FS = iter->second.find(CalleeName);
387 if (FS != iter->second.end())
389 // If we cannot find exact match of the callee name, return the FS with
390 // the max total count. Only do this when CalleeName is not provided,
391 // i.e., only for indirect calls.
392 if (!CalleeName.empty())
394 uint64_t MaxTotalSamples = 0;
395 const FunctionSamples *R = nullptr;
396 for (const auto &NameFS : iter->second)
397 if (NameFS.second.getTotalSamples() >= MaxTotalSamples) {
398 MaxTotalSamples = NameFS.second.getTotalSamples();
404 bool empty() const { return TotalSamples == 0; }
406 /// Return the total number of samples collected inside the function.
407 uint64_t getTotalSamples() const { return TotalSamples; }
409 /// Return the total number of branch samples that have the function as the
410 /// branch target. This should be equivalent to the sample of the first
411 /// instruction of the symbol. But as we directly get this info for raw
412 /// profile without referring to potentially inaccurate debug info, this
413 /// gives more accurate profile data and is preferred for standalone symbols.
414 uint64_t getHeadSamples() const { return TotalHeadSamples; }
416 /// Return the sample count of the first instruction of the function.
417 /// The function can be either a standalone symbol or an inlined function.
418 uint64_t getEntrySamples() const {
419 // Use either BodySamples or CallsiteSamples which ever has the smaller
421 if (!BodySamples.empty() &&
422 (CallsiteSamples.empty() ||
423 BodySamples.begin()->first < CallsiteSamples.begin()->first))
424 return BodySamples.begin()->second.getSamples();
425 if (!CallsiteSamples.empty()) {
427 // An indirect callsite may be promoted to several inlined direct calls.
428 // We need to get the sum of them.
429 for (const auto &N_FS : CallsiteSamples.begin()->second)
430 T += N_FS.second.getEntrySamples();
436 /// Return all the samples collected in the body of the function.
437 const BodySampleMap &getBodySamples() const { return BodySamples; }
439 /// Return all the callsite samples collected in the body of the function.
440 const CallsiteSampleMap &getCallsiteSamples() const {
441 return CallsiteSamples;
444 /// Merge the samples in \p Other into this one.
445 /// Optionally scale samples by \p Weight.
446 sampleprof_error merge(const FunctionSamples &Other, uint64_t Weight = 1) {
447 sampleprof_error Result = sampleprof_error::success;
448 Name = Other.getName();
449 MergeResult(Result, addTotalSamples(Other.getTotalSamples(), Weight));
450 MergeResult(Result, addHeadSamples(Other.getHeadSamples(), Weight));
451 for (const auto &I : Other.getBodySamples()) {
452 const LineLocation &Loc = I.first;
453 const SampleRecord &Rec = I.second;
454 MergeResult(Result, BodySamples[Loc].merge(Rec, Weight));
456 for (const auto &I : Other.getCallsiteSamples()) {
457 const LineLocation &Loc = I.first;
458 FunctionSamplesMap &FSMap = functionSamplesAt(Loc);
459 for (const auto &Rec : I.second)
460 MergeResult(Result, FSMap[Rec.first].merge(Rec.second, Weight));
465 /// Recursively traverses all children, if the total sample count of the
466 /// corresponding function is no less than \p Threshold, add its corresponding
467 /// GUID to \p S. Also traverse the BodySamples to add hot CallTarget's GUID
469 void findInlinedFunctions(DenseSet<GlobalValue::GUID> &S, const Module *M,
470 uint64_t Threshold) const {
471 if (TotalSamples <= Threshold)
473 S.insert(getGUID(Name));
474 // Import hot CallTargets, which may not be available in IR because full
475 // profile annotation cannot be done until backend compilation in ThinLTO.
476 for (const auto &BS : BodySamples)
477 for (const auto &TS : BS.second.getCallTargets())
478 if (TS.getValue() > Threshold) {
479 const Function *Callee =
480 M->getFunction(getNameInModule(TS.getKey(), M));
481 if (!Callee || !Callee->getSubprogram())
482 S.insert(getGUID(TS.getKey()));
484 for (const auto &CS : CallsiteSamples)
485 for (const auto &NameFS : CS.second)
486 NameFS.second.findInlinedFunctions(S, M, Threshold);
489 /// Set the name of the function.
490 void setName(StringRef FunctionName) { Name = FunctionName; }
492 /// Return the function name.
493 StringRef getName() const { return Name; }
495 /// Return the original function name if it exists in Module \p M.
496 StringRef getFuncNameInModule(const Module *M) const {
497 return getNameInModule(Name, M);
500 /// Return the canonical name for a function, taking into account
501 /// suffix elision policy attributes.
502 static StringRef getCanonicalFnName(const Function &F) {
503 static const char *knownSuffixes[] = { ".llvm.", ".part." };
504 auto AttrName = "sample-profile-suffix-elision-policy";
505 auto Attr = F.getFnAttribute(AttrName).getValueAsString();
506 if (Attr == "" || Attr == "all") {
507 return F.getName().split('.').first;
508 } else if (Attr == "selected") {
509 StringRef Cand(F.getName());
510 for (const auto &Suf : knownSuffixes) {
511 StringRef Suffix(Suf);
512 auto It = Cand.rfind(Suffix);
513 if (It == StringRef::npos)
515 auto Dit = Cand.rfind('.');
516 if (Dit == It + Suffix.size() - 1)
517 Cand = Cand.substr(0, It);
520 } else if (Attr == "none") {
523 assert(false && "internal error: unknown suffix elision policy");
528 /// Translate \p Name into its original name in Module.
529 /// When the Format is not SPF_Compact_Binary, \p Name needs no translation.
530 /// When the Format is SPF_Compact_Binary, \p Name in current FunctionSamples
531 /// is actually GUID of the original function name. getNameInModule will
532 /// translate \p Name in current FunctionSamples into its original name.
533 /// If the original name doesn't exist in \p M, return empty StringRef.
534 StringRef getNameInModule(StringRef Name, const Module *M) const {
535 if (Format != SPF_Compact_Binary)
538 assert(GUIDToFuncNameMap && "GUIDToFuncNameMap needs to be popluated first");
539 auto iter = GUIDToFuncNameMap->find(std::stoull(Name.data()));
540 if (iter == GUIDToFuncNameMap->end())
545 /// Returns the line offset to the start line of the subprogram.
546 /// We assume that a single function will not exceed 65535 LOC.
547 static unsigned getOffset(const DILocation *DIL);
549 /// Get the FunctionSamples of the inline instance where DIL originates
552 /// The FunctionSamples of the instruction (Machine or IR) associated to
553 /// \p DIL is the inlined instance in which that instruction is coming from.
554 /// We traverse the inline stack of that instruction, and match it with the
555 /// tree nodes in the profile.
557 /// \returns the FunctionSamples pointer to the inlined instance.
558 const FunctionSamples *findFunctionSamples(const DILocation *DIL) const;
560 static SampleProfileFormat Format;
562 /// GUIDToFuncNameMap saves the mapping from GUID to the symbol name, for
563 /// all the function symbols defined or declared in current module.
564 DenseMap<uint64_t, StringRef> *GUIDToFuncNameMap = nullptr;
566 // Assume the input \p Name is a name coming from FunctionSamples itself.
567 // If the format is SPF_Compact_Binary, the name is already a GUID and we
568 // don't want to return the GUID of GUID.
569 static uint64_t getGUID(StringRef Name) {
570 return (Format == SPF_Compact_Binary) ? std::stoull(Name.data())
571 : Function::getGUID(Name);
575 /// Mangled name of the function.
578 /// Total number of samples collected inside this function.
580 /// Samples are cumulative, they include all the samples collected
581 /// inside this function and all its inlined callees.
582 uint64_t TotalSamples = 0;
584 /// Total number of samples collected at the head of the function.
585 /// This is an approximation of the number of calls made to this function
587 uint64_t TotalHeadSamples = 0;
589 /// Map instruction locations to collected samples.
591 /// Each entry in this map contains the number of samples
592 /// collected at the corresponding line offset. All line locations
593 /// are an offset from the start of the function.
594 BodySampleMap BodySamples;
596 /// Map call sites to collected samples for the called function.
598 /// Each entry in this map corresponds to all the samples
599 /// collected for the inlined function call at the given
600 /// location. For example, given:
608 /// If the bar() and baz() calls were inlined inside foo(), this
609 /// map will contain two entries. One for all the samples collected
610 /// in the call to bar() at line offset 1, the other for all the samples
611 /// collected in the call to baz() at line offset 8.
612 CallsiteSampleMap CallsiteSamples;
615 raw_ostream &operator<<(raw_ostream &OS, const FunctionSamples &FS);
617 /// Sort a LocationT->SampleT map by LocationT.
619 /// It produces a sorted list of <LocationT, SampleT> records by ascending
620 /// order of LocationT.
621 template <class LocationT, class SampleT> class SampleSorter {
623 using SamplesWithLoc = std::pair<const LocationT, SampleT>;
624 using SamplesWithLocList = SmallVector<const SamplesWithLoc *, 20>;
626 SampleSorter(const std::map<LocationT, SampleT> &Samples) {
627 for (const auto &I : Samples)
629 llvm::stable_sort(V, [](const SamplesWithLoc *A, const SamplesWithLoc *B) {
630 return A->first < B->first;
634 const SamplesWithLocList &get() const { return V; }
637 SamplesWithLocList V;
640 /// ProfileSymbolList records the list of function symbols shown up
641 /// in the binary used to generate the profile. It is useful to
642 /// to discriminate a function being so cold as not to shown up
643 /// in the profile and a function newly added.
644 class ProfileSymbolList {
646 /// copy indicates whether we need to copy the underlying memory
647 /// for the input Name.
648 void add(StringRef Name, bool copy = false) {
653 Syms.insert(Name.copy(Allocator));
656 bool contains(StringRef Name) { return Syms.count(Name); }
658 void merge(const ProfileSymbolList &List) {
659 for (auto Sym : List.Syms)
663 unsigned size() { return Syms.size(); }
665 void setToCompress(bool TC) { ToCompress = TC; }
666 bool toCompress() { return ToCompress; }
668 std::error_code read(const uint8_t *Data, uint64_t ListSize);
669 std::error_code write(raw_ostream &OS);
670 void dump(raw_ostream &OS = dbgs()) const;
673 // Determine whether or not to compress the symbol list when
674 // writing it into profile. The variable is unused when the symbol
675 // list is read from an existing profile.
676 bool ToCompress = false;
677 DenseSet<StringRef> Syms;
678 BumpPtrAllocator Allocator;
681 } // end namespace sampleprof
682 } // end namespace llvm
684 #endif // LLVM_PROFILEDATA_SAMPLEPROF_H