1 //===- SampleProfReader.cpp - Read LLVM sample profile data ---------------===//
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 implements the class that reads LLVM sample profiles. It
10 // supports three file formats: text, binary and gcov.
12 // The textual representation is useful for debugging and testing purposes. The
13 // binary representation is more compact, resulting in smaller file sizes.
15 // The gcov encoding is the one generated by GCC's AutoFDO profile creation
16 // tool (https://github.com/google/autofdo)
18 // All three encodings can be used interchangeably as an input sample profile.
20 //===----------------------------------------------------------------------===//
22 #include "llvm/ProfileData/SampleProfReader.h"
23 #include "llvm/ADT/DenseMap.h"
24 #include "llvm/ADT/STLExtras.h"
25 #include "llvm/ADT/StringRef.h"
26 #include "llvm/IR/Module.h"
27 #include "llvm/IR/ProfileSummary.h"
28 #include "llvm/ProfileData/ProfileCommon.h"
29 #include "llvm/ProfileData/SampleProf.h"
30 #include "llvm/Support/CommandLine.h"
31 #include "llvm/Support/Compression.h"
32 #include "llvm/Support/ErrorOr.h"
33 #include "llvm/Support/JSON.h"
34 #include "llvm/Support/LEB128.h"
35 #include "llvm/Support/LineIterator.h"
36 #include "llvm/Support/MD5.h"
37 #include "llvm/Support/MemoryBuffer.h"
38 #include "llvm/Support/VirtualFileSystem.h"
39 #include "llvm/Support/raw_ostream.h"
45 #include <system_error>
49 using namespace sampleprof
;
51 #define DEBUG_TYPE "samplepgo-reader"
53 // This internal option specifies if the profile uses FS discriminators.
54 // It only applies to text, and binary format profiles.
55 // For ext-binary format profiles, the flag is set in the summary.
56 static cl::opt
<bool> ProfileIsFSDisciminator(
57 "profile-isfs", cl::Hidden
, cl::init(false),
58 cl::desc("Profile uses flow sensitive discriminators"));
60 /// Dump the function profile for \p FName.
62 /// \param FContext Name + context of the function to print.
63 /// \param OS Stream to emit the output to.
64 void SampleProfileReader::dumpFunctionProfile(const FunctionSamples
&FS
,
66 OS
<< "Function: " << FS
.getContext().toString() << ": " << FS
;
69 /// Dump all the function profiles found on stream \p OS.
70 void SampleProfileReader::dump(raw_ostream
&OS
) {
71 std::vector
<NameFunctionSamples
> V
;
72 sortFuncProfiles(Profiles
, V
);
73 for (const auto &I
: V
)
74 dumpFunctionProfile(*I
.second
, OS
);
77 static void dumpFunctionProfileJson(const FunctionSamples
&S
,
78 json::OStream
&JOS
, bool TopLevel
= false) {
79 auto DumpBody
= [&](const BodySampleMap
&BodySamples
) {
80 for (const auto &I
: BodySamples
) {
81 const LineLocation
&Loc
= I
.first
;
82 const SampleRecord
&Sample
= I
.second
;
84 JOS
.attribute("line", Loc
.LineOffset
);
85 if (Loc
.Discriminator
)
86 JOS
.attribute("discriminator", Loc
.Discriminator
);
87 JOS
.attribute("samples", Sample
.getSamples());
89 auto CallTargets
= Sample
.getSortedCallTargets();
90 if (!CallTargets
.empty()) {
91 JOS
.attributeArray("calls", [&] {
92 for (const auto &J
: CallTargets
) {
94 JOS
.attribute("function", J
.first
.str());
95 JOS
.attribute("samples", J
.second
);
104 auto DumpCallsiteSamples
= [&](const CallsiteSampleMap
&CallsiteSamples
) {
105 for (const auto &I
: CallsiteSamples
)
106 for (const auto &FS
: I
.second
) {
107 const LineLocation
&Loc
= I
.first
;
108 const FunctionSamples
&CalleeSamples
= FS
.second
;
110 JOS
.attribute("line", Loc
.LineOffset
);
111 if (Loc
.Discriminator
)
112 JOS
.attribute("discriminator", Loc
.Discriminator
);
114 "samples", [&] { dumpFunctionProfileJson(CalleeSamples
, JOS
); });
120 JOS
.attribute("name", S
.getFunction().str());
121 JOS
.attribute("total", S
.getTotalSamples());
123 JOS
.attribute("head", S
.getHeadSamples());
125 const auto &BodySamples
= S
.getBodySamples();
126 if (!BodySamples
.empty())
127 JOS
.attributeArray("body", [&] { DumpBody(BodySamples
); });
129 const auto &CallsiteSamples
= S
.getCallsiteSamples();
130 if (!CallsiteSamples
.empty())
131 JOS
.attributeArray("callsites",
132 [&] { DumpCallsiteSamples(CallsiteSamples
); });
136 /// Dump all the function profiles found on stream \p OS in the JSON format.
137 void SampleProfileReader::dumpJson(raw_ostream
&OS
) {
138 std::vector
<NameFunctionSamples
> V
;
139 sortFuncProfiles(Profiles
, V
);
140 json::OStream
JOS(OS
, 2);
142 for (const auto &F
: V
)
143 dumpFunctionProfileJson(*F
.second
, JOS
, true);
146 // Emit a newline character at the end as json::OStream doesn't emit one.
150 /// Parse \p Input as function head.
152 /// Parse one line of \p Input, and update function name in \p FName,
153 /// function's total sample count in \p NumSamples, function's entry
154 /// count in \p NumHeadSamples.
156 /// \returns true if parsing is successful.
157 static bool ParseHead(const StringRef
&Input
, StringRef
&FName
,
158 uint64_t &NumSamples
, uint64_t &NumHeadSamples
) {
161 size_t n2
= Input
.rfind(':');
162 size_t n1
= Input
.rfind(':', n2
- 1);
163 FName
= Input
.substr(0, n1
);
164 if (Input
.substr(n1
+ 1, n2
- n1
- 1).getAsInteger(10, NumSamples
))
166 if (Input
.substr(n2
+ 1).getAsInteger(10, NumHeadSamples
))
171 /// Returns true if line offset \p L is legal (only has 16 bits).
172 static bool isOffsetLegal(unsigned L
) { return (L
& 0xffff) == L
; }
174 /// Parse \p Input that contains metadata.
175 /// Possible metadata:
176 /// - CFG Checksum information:
177 /// !CFGChecksum: 12345
178 /// - CFG Checksum information:
180 /// Stores the FunctionHash (a.k.a. CFG Checksum) into \p FunctionHash.
181 static bool parseMetadata(const StringRef
&Input
, uint64_t &FunctionHash
,
182 uint32_t &Attributes
) {
183 if (Input
.starts_with("!CFGChecksum:")) {
184 StringRef CFGInfo
= Input
.substr(strlen("!CFGChecksum:")).trim();
185 return !CFGInfo
.getAsInteger(10, FunctionHash
);
188 if (Input
.starts_with("!Attributes:")) {
189 StringRef Attrib
= Input
.substr(strlen("!Attributes:")).trim();
190 return !Attrib
.getAsInteger(10, Attributes
);
196 enum class LineType
{
202 /// Parse \p Input as line sample.
204 /// \param Input input line.
205 /// \param LineTy Type of this line.
206 /// \param Depth the depth of the inline stack.
207 /// \param NumSamples total samples of the line/inlined callsite.
208 /// \param LineOffset line offset to the start of the function.
209 /// \param Discriminator discriminator of the line.
210 /// \param TargetCountMap map from indirect call target to count.
211 /// \param FunctionHash the function's CFG hash, used by pseudo probe.
213 /// returns true if parsing is successful.
214 static bool ParseLine(const StringRef
&Input
, LineType
&LineTy
, uint32_t &Depth
,
215 uint64_t &NumSamples
, uint32_t &LineOffset
,
216 uint32_t &Discriminator
, StringRef
&CalleeName
,
217 DenseMap
<StringRef
, uint64_t> &TargetCountMap
,
218 uint64_t &FunctionHash
, uint32_t &Attributes
) {
219 for (Depth
= 0; Input
[Depth
] == ' '; Depth
++)
224 if (Input
[Depth
] == '!') {
225 LineTy
= LineType::Metadata
;
226 return parseMetadata(Input
.substr(Depth
), FunctionHash
, Attributes
);
229 size_t n1
= Input
.find(':');
230 StringRef Loc
= Input
.substr(Depth
, n1
- Depth
);
231 size_t n2
= Loc
.find('.');
232 if (n2
== StringRef::npos
) {
233 if (Loc
.getAsInteger(10, LineOffset
) || !isOffsetLegal(LineOffset
))
237 if (Loc
.substr(0, n2
).getAsInteger(10, LineOffset
))
239 if (Loc
.substr(n2
+ 1).getAsInteger(10, Discriminator
))
243 StringRef Rest
= Input
.substr(n1
+ 2);
244 if (isDigit(Rest
[0])) {
245 LineTy
= LineType::BodyProfile
;
246 size_t n3
= Rest
.find(' ');
247 if (n3
== StringRef::npos
) {
248 if (Rest
.getAsInteger(10, NumSamples
))
251 if (Rest
.substr(0, n3
).getAsInteger(10, NumSamples
))
254 // Find call targets and their sample counts.
255 // Note: In some cases, there are symbols in the profile which are not
256 // mangled. To accommodate such cases, use colon + integer pairs as the
259 // _M_construct<char *>:1000 string_view<std::allocator<char> >:437
260 // ":1000" and ":437" are used as anchor points so the string above will
262 // target: _M_construct<char *>
264 // target: string_view<std::allocator<char> >
266 while (n3
!= StringRef::npos
) {
267 n3
+= Rest
.substr(n3
).find_first_not_of(' ');
268 Rest
= Rest
.substr(n3
);
269 n3
= Rest
.find_first_of(':');
270 if (n3
== StringRef::npos
|| n3
== 0)
276 // Get the segment after the current colon.
277 StringRef AfterColon
= Rest
.substr(n3
+ 1);
278 // Get the target symbol before the current colon.
279 Target
= Rest
.substr(0, n3
);
280 // Check if the word after the current colon is an integer.
281 n4
= AfterColon
.find_first_of(' ');
282 n4
= (n4
!= StringRef::npos
) ? n3
+ n4
+ 1 : Rest
.size();
283 StringRef WordAfterColon
= Rest
.substr(n3
+ 1, n4
- n3
- 1);
284 if (!WordAfterColon
.getAsInteger(10, count
))
287 // Try to find the next colon.
288 uint64_t n5
= AfterColon
.find_first_of(':');
289 if (n5
== StringRef::npos
)
294 // An anchor point is found. Save the {target, count} pair
295 TargetCountMap
[Target
] = count
;
296 if (n4
== Rest
.size())
298 // Change n3 to the next blank space after colon + integer pair.
302 LineTy
= LineType::CallSiteProfile
;
303 size_t n3
= Rest
.find_last_of(':');
304 CalleeName
= Rest
.substr(0, n3
);
305 if (Rest
.substr(n3
+ 1).getAsInteger(10, NumSamples
))
311 /// Load samples from a text file.
313 /// See the documentation at the top of the file for an explanation of
314 /// the expected format.
316 /// \returns true if the file was loaded successfully, false otherwise.
317 std::error_code
SampleProfileReaderText::readImpl() {
318 line_iterator
LineIt(*Buffer
, /*SkipBlanks=*/true, '#');
319 sampleprof_error Result
= sampleprof_error::success
;
321 InlineCallStack InlineStack
;
322 uint32_t TopLevelProbeProfileCount
= 0;
324 // DepthMetadata tracks whether we have processed metadata for the current
325 // top-level or nested function profile.
326 uint32_t DepthMetadata
= 0;
328 ProfileIsFS
= ProfileIsFSDisciminator
;
329 FunctionSamples::ProfileIsFS
= ProfileIsFS
;
330 for (; !LineIt
.is_at_eof(); ++LineIt
) {
331 size_t pos
= LineIt
->find_first_not_of(' ');
332 if (pos
== LineIt
->npos
|| (*LineIt
)[pos
] == '#')
334 // Read the header of each function.
336 // Note that for function identifiers we are actually expecting
337 // mangled names, but we may not always get them. This happens when
338 // the compiler decides not to emit the function (e.g., it was inlined
339 // and removed). In this case, the binary will not have the linkage
340 // name for the function, so the profiler will emit the function's
341 // unmangled name, which may contain characters like ':' and '>' in its
342 // name (member functions, templates, etc).
344 // The only requirement we place on the identifier, then, is that it
345 // should not begin with a number.
346 if ((*LineIt
)[0] != ' ') {
347 uint64_t NumSamples
, NumHeadSamples
;
349 if (!ParseHead(*LineIt
, FName
, NumSamples
, NumHeadSamples
)) {
350 reportError(LineIt
.line_number(),
351 "Expected 'mangled_name:NUM:NUM', found " + *LineIt
);
352 return sampleprof_error::malformed
;
355 SampleContext
FContext(FName
, CSNameTable
);
356 if (FContext
.hasContext())
358 FunctionSamples
&FProfile
= Profiles
.Create(FContext
);
359 MergeResult(Result
, FProfile
.addTotalSamples(NumSamples
));
360 MergeResult(Result
, FProfile
.addHeadSamples(NumHeadSamples
));
362 InlineStack
.push_back(&FProfile
);
366 DenseMap
<StringRef
, uint64_t> TargetCountMap
;
367 uint32_t Depth
, LineOffset
, Discriminator
;
369 uint64_t FunctionHash
= 0;
370 uint32_t Attributes
= 0;
371 if (!ParseLine(*LineIt
, LineTy
, Depth
, NumSamples
, LineOffset
,
372 Discriminator
, FName
, TargetCountMap
, FunctionHash
,
374 reportError(LineIt
.line_number(),
375 "Expected 'NUM[.NUM]: NUM[ mangled_name:NUM]*', found " +
377 return sampleprof_error::malformed
;
379 if (LineTy
!= LineType::Metadata
&& Depth
== DepthMetadata
) {
380 // Metadata must be put at the end of a function profile.
381 reportError(LineIt
.line_number(),
382 "Found non-metadata after metadata: " + *LineIt
);
383 return sampleprof_error::malformed
;
386 // Here we handle FS discriminators.
387 Discriminator
&= getDiscriminatorMask();
389 while (InlineStack
.size() > Depth
) {
390 InlineStack
.pop_back();
393 case LineType::CallSiteProfile
: {
394 FunctionSamples
&FSamples
= InlineStack
.back()->functionSamplesAt(
395 LineLocation(LineOffset
, Discriminator
))[FunctionId(FName
)];
396 FSamples
.setFunction(FunctionId(FName
));
397 MergeResult(Result
, FSamples
.addTotalSamples(NumSamples
));
398 InlineStack
.push_back(&FSamples
);
402 case LineType::BodyProfile
: {
403 while (InlineStack
.size() > Depth
) {
404 InlineStack
.pop_back();
406 FunctionSamples
&FProfile
= *InlineStack
.back();
407 for (const auto &name_count
: TargetCountMap
) {
408 MergeResult(Result
, FProfile
.addCalledTargetSamples(
409 LineOffset
, Discriminator
,
410 FunctionId(name_count
.first
),
413 MergeResult(Result
, FProfile
.addBodySamples(LineOffset
, Discriminator
,
417 case LineType::Metadata
: {
418 FunctionSamples
&FProfile
= *InlineStack
.back();
420 FProfile
.setFunctionHash(FunctionHash
);
422 ++TopLevelProbeProfileCount
;
424 FProfile
.getContext().setAllAttributes(Attributes
);
425 if (Attributes
& (uint32_t)ContextShouldBeInlined
)
426 ProfileIsPreInlined
= true;
427 DepthMetadata
= Depth
;
434 assert((CSProfileCount
== 0 || CSProfileCount
== Profiles
.size()) &&
435 "Cannot have both context-sensitive and regular profile");
436 ProfileIsCS
= (CSProfileCount
> 0);
437 assert((TopLevelProbeProfileCount
== 0 ||
438 TopLevelProbeProfileCount
== Profiles
.size()) &&
439 "Cannot have both probe-based profiles and regular profiles");
440 ProfileIsProbeBased
= (TopLevelProbeProfileCount
> 0);
441 FunctionSamples::ProfileIsProbeBased
= ProfileIsProbeBased
;
442 FunctionSamples::ProfileIsCS
= ProfileIsCS
;
443 FunctionSamples::ProfileIsPreInlined
= ProfileIsPreInlined
;
445 if (Result
== sampleprof_error::success
)
451 bool SampleProfileReaderText::hasFormat(const MemoryBuffer
&Buffer
) {
454 // Check that the first non-comment line is a valid function header.
455 line_iterator
LineIt(Buffer
, /*SkipBlanks=*/true, '#');
456 if (!LineIt
.is_at_eof()) {
457 if ((*LineIt
)[0] != ' ') {
458 uint64_t NumSamples
, NumHeadSamples
;
460 result
= ParseHead(*LineIt
, FName
, NumSamples
, NumHeadSamples
);
467 template <typename T
> ErrorOr
<T
> SampleProfileReaderBinary::readNumber() {
468 unsigned NumBytesRead
= 0;
469 uint64_t Val
= decodeULEB128(Data
, &NumBytesRead
);
471 if (Val
> std::numeric_limits
<T
>::max()) {
472 std::error_code EC
= sampleprof_error::malformed
;
473 reportError(0, EC
.message());
475 } else if (Data
+ NumBytesRead
> End
) {
476 std::error_code EC
= sampleprof_error::truncated
;
477 reportError(0, EC
.message());
481 Data
+= NumBytesRead
;
482 return static_cast<T
>(Val
);
485 ErrorOr
<StringRef
> SampleProfileReaderBinary::readString() {
486 StringRef
Str(reinterpret_cast<const char *>(Data
));
487 if (Data
+ Str
.size() + 1 > End
) {
488 std::error_code EC
= sampleprof_error::truncated
;
489 reportError(0, EC
.message());
493 Data
+= Str
.size() + 1;
497 template <typename T
>
498 ErrorOr
<T
> SampleProfileReaderBinary::readUnencodedNumber() {
499 if (Data
+ sizeof(T
) > End
) {
500 std::error_code EC
= sampleprof_error::truncated
;
501 reportError(0, EC
.message());
505 using namespace support
;
506 T Val
= endian::readNext
<T
, llvm::endianness::little
, unaligned
>(Data
);
510 template <typename T
>
511 inline ErrorOr
<size_t> SampleProfileReaderBinary::readStringIndex(T
&Table
) {
512 auto Idx
= readNumber
<size_t>();
513 if (std::error_code EC
= Idx
.getError())
515 if (*Idx
>= Table
.size())
516 return sampleprof_error::truncated_name_table
;
521 SampleProfileReaderBinary::readStringFromTable(size_t *RetIdx
) {
522 auto Idx
= readStringIndex(NameTable
);
523 if (std::error_code EC
= Idx
.getError())
527 return NameTable
[*Idx
];
530 ErrorOr
<SampleContextFrames
>
531 SampleProfileReaderBinary::readContextFromTable(size_t *RetIdx
) {
532 auto ContextIdx
= readNumber
<size_t>();
533 if (std::error_code EC
= ContextIdx
.getError())
535 if (*ContextIdx
>= CSNameTable
.size())
536 return sampleprof_error::truncated_name_table
;
538 *RetIdx
= *ContextIdx
;
539 return CSNameTable
[*ContextIdx
];
542 ErrorOr
<std::pair
<SampleContext
, uint64_t>>
543 SampleProfileReaderBinary::readSampleContextFromTable() {
544 SampleContext Context
;
547 auto FContext(readContextFromTable(&Idx
));
548 if (std::error_code EC
= FContext
.getError())
550 Context
= SampleContext(*FContext
);
552 auto FName(readStringFromTable(&Idx
));
553 if (std::error_code EC
= FName
.getError())
555 Context
= SampleContext(*FName
);
557 // Since MD5SampleContextStart may point to the profile's file data, need to
558 // make sure it is reading the same value on big endian CPU.
559 uint64_t Hash
= support::endian::read64le(MD5SampleContextStart
+ Idx
);
560 // Lazy computing of hash value, write back to the table to cache it. Only
561 // compute the context's hash value if it is being referenced for the first
564 assert(MD5SampleContextStart
== MD5SampleContextTable
.data());
565 Hash
= Context
.getHashCode();
566 support::endian::write64le(&MD5SampleContextTable
[Idx
], Hash
);
568 return std::make_pair(Context
, Hash
);
572 SampleProfileReaderBinary::readProfile(FunctionSamples
&FProfile
) {
573 auto NumSamples
= readNumber
<uint64_t>();
574 if (std::error_code EC
= NumSamples
.getError())
576 FProfile
.addTotalSamples(*NumSamples
);
578 // Read the samples in the body.
579 auto NumRecords
= readNumber
<uint32_t>();
580 if (std::error_code EC
= NumRecords
.getError())
583 for (uint32_t I
= 0; I
< *NumRecords
; ++I
) {
584 auto LineOffset
= readNumber
<uint64_t>();
585 if (std::error_code EC
= LineOffset
.getError())
588 if (!isOffsetLegal(*LineOffset
)) {
589 return std::error_code();
592 auto Discriminator
= readNumber
<uint64_t>();
593 if (std::error_code EC
= Discriminator
.getError())
596 auto NumSamples
= readNumber
<uint64_t>();
597 if (std::error_code EC
= NumSamples
.getError())
600 auto NumCalls
= readNumber
<uint32_t>();
601 if (std::error_code EC
= NumCalls
.getError())
604 // Here we handle FS discriminators:
605 uint32_t DiscriminatorVal
= (*Discriminator
) & getDiscriminatorMask();
607 for (uint32_t J
= 0; J
< *NumCalls
; ++J
) {
608 auto CalledFunction(readStringFromTable());
609 if (std::error_code EC
= CalledFunction
.getError())
612 auto CalledFunctionSamples
= readNumber
<uint64_t>();
613 if (std::error_code EC
= CalledFunctionSamples
.getError())
616 FProfile
.addCalledTargetSamples(*LineOffset
, DiscriminatorVal
,
617 *CalledFunction
, *CalledFunctionSamples
);
620 FProfile
.addBodySamples(*LineOffset
, DiscriminatorVal
, *NumSamples
);
623 // Read all the samples for inlined function calls.
624 auto NumCallsites
= readNumber
<uint32_t>();
625 if (std::error_code EC
= NumCallsites
.getError())
628 for (uint32_t J
= 0; J
< *NumCallsites
; ++J
) {
629 auto LineOffset
= readNumber
<uint64_t>();
630 if (std::error_code EC
= LineOffset
.getError())
633 auto Discriminator
= readNumber
<uint64_t>();
634 if (std::error_code EC
= Discriminator
.getError())
637 auto FName(readStringFromTable());
638 if (std::error_code EC
= FName
.getError())
641 // Here we handle FS discriminators:
642 uint32_t DiscriminatorVal
= (*Discriminator
) & getDiscriminatorMask();
644 FunctionSamples
&CalleeProfile
= FProfile
.functionSamplesAt(
645 LineLocation(*LineOffset
, DiscriminatorVal
))[*FName
];
646 CalleeProfile
.setFunction(*FName
);
647 if (std::error_code EC
= readProfile(CalleeProfile
))
651 return sampleprof_error::success
;
655 SampleProfileReaderBinary::readFuncProfile(const uint8_t *Start
) {
657 auto NumHeadSamples
= readNumber
<uint64_t>();
658 if (std::error_code EC
= NumHeadSamples
.getError())
661 auto FContextHash(readSampleContextFromTable());
662 if (std::error_code EC
= FContextHash
.getError())
665 auto &[FContext
, Hash
] = *FContextHash
;
666 // Use the cached hash value for insertion instead of recalculating it.
667 auto Res
= Profiles
.try_emplace(Hash
, FContext
, FunctionSamples());
668 FunctionSamples
&FProfile
= Res
.first
->second
;
669 FProfile
.setContext(FContext
);
670 FProfile
.addHeadSamples(*NumHeadSamples
);
672 if (FContext
.hasContext())
675 if (std::error_code EC
= readProfile(FProfile
))
677 return sampleprof_error::success
;
680 std::error_code
SampleProfileReaderBinary::readImpl() {
681 ProfileIsFS
= ProfileIsFSDisciminator
;
682 FunctionSamples::ProfileIsFS
= ProfileIsFS
;
684 if (std::error_code EC
= readFuncProfile(Data
))
688 return sampleprof_error::success
;
691 std::error_code
SampleProfileReaderExtBinaryBase::readOneSection(
692 const uint8_t *Start
, uint64_t Size
, const SecHdrTableEntry
&Entry
) {
695 switch (Entry
.Type
) {
697 if (std::error_code EC
= readSummary())
699 if (hasSecFlag(Entry
, SecProfSummaryFlags::SecFlagPartial
))
700 Summary
->setPartialProfile(true);
701 if (hasSecFlag(Entry
, SecProfSummaryFlags::SecFlagFullContext
))
702 FunctionSamples::ProfileIsCS
= ProfileIsCS
= true;
703 if (hasSecFlag(Entry
, SecProfSummaryFlags::SecFlagIsPreInlined
))
704 FunctionSamples::ProfileIsPreInlined
= ProfileIsPreInlined
= true;
705 if (hasSecFlag(Entry
, SecProfSummaryFlags::SecFlagFSDiscriminator
))
706 FunctionSamples::ProfileIsFS
= ProfileIsFS
= true;
709 bool FixedLengthMD5
=
710 hasSecFlag(Entry
, SecNameTableFlags::SecFlagFixedLengthMD5
);
711 bool UseMD5
= hasSecFlag(Entry
, SecNameTableFlags::SecFlagMD5Name
);
712 // UseMD5 means if THIS section uses MD5, ProfileIsMD5 means if the entire
713 // profile uses MD5 for function name matching in IPO passes.
714 ProfileIsMD5
= ProfileIsMD5
|| UseMD5
;
715 FunctionSamples::HasUniqSuffix
=
716 hasSecFlag(Entry
, SecNameTableFlags::SecFlagUniqSuffix
);
717 if (std::error_code EC
= readNameTableSec(UseMD5
, FixedLengthMD5
))
721 case SecCSNameTable
: {
722 if (std::error_code EC
= readCSNameTableSec())
727 if (std::error_code EC
= readFuncProfiles())
730 case SecFuncOffsetTable
:
731 // If module is absent, we are using LLVM tools, and need to read all
732 // profiles, so skip reading the function offset table.
736 assert((!ProfileIsCS
||
737 hasSecFlag(Entry
, SecFuncOffsetFlags::SecFlagOrdered
)) &&
738 "func offset table should always be sorted in CS profile");
739 if (std::error_code EC
= readFuncOffsetTable())
743 case SecFuncMetadata
: {
744 ProfileIsProbeBased
=
745 hasSecFlag(Entry
, SecFuncMetadataFlags::SecFlagIsProbeBased
);
746 FunctionSamples::ProfileIsProbeBased
= ProfileIsProbeBased
;
748 hasSecFlag(Entry
, SecFuncMetadataFlags::SecFlagHasAttribute
);
749 if (std::error_code EC
= readFuncMetadata(HasAttribute
))
753 case SecProfileSymbolList
:
754 if (std::error_code EC
= readProfileSymbolList())
758 if (std::error_code EC
= readCustomSection(Entry
))
762 return sampleprof_error::success
;
765 bool SampleProfileReaderExtBinaryBase::useFuncOffsetList() const {
766 // If profile is CS, the function offset section is expected to consist of
767 // sequences of contexts in pre-order layout
768 // (e.g. [A, A:1 @ B, A:1 @ B:2.3 @ C] [D, D:1 @ E]), so that when a matched
769 // context in the module is found, the profiles of all its callees are
770 // recursively loaded. A list is needed since the order of profiles matters.
774 // If the profile is MD5, use the map container to lookup functions in
775 // the module. A remapper has no use on MD5 names.
779 // Profile is not MD5 and if a remapper is present, the remapped name of
780 // every function needed to be matched against the module, so use the list
781 // container since each entry is accessed.
785 // Otherwise use the map container for faster lookup.
786 // TODO: If the cardinality of the function offset section is much smaller
787 // than the number of functions in the module, using the list container can
788 // be always faster, but we need to figure out the constant factor to
789 // determine the cutoff.
794 bool SampleProfileReaderExtBinaryBase::collectFuncsFromModule() {
799 FuncsToUse
.insert(FunctionSamples::getCanonicalFnName(F
));
803 std::error_code
SampleProfileReaderExtBinaryBase::readFuncOffsetTable() {
804 // If there are more than one function offset section, the profile associated
805 // with the previous section has to be done reading before next one is read.
806 FuncOffsetTable
.clear();
807 FuncOffsetList
.clear();
809 auto Size
= readNumber
<uint64_t>();
810 if (std::error_code EC
= Size
.getError())
813 bool UseFuncOffsetList
= useFuncOffsetList();
814 if (UseFuncOffsetList
)
815 FuncOffsetList
.reserve(*Size
);
817 FuncOffsetTable
.reserve(*Size
);
819 for (uint64_t I
= 0; I
< *Size
; ++I
) {
820 auto FContextHash(readSampleContextFromTable());
821 if (std::error_code EC
= FContextHash
.getError())
824 auto &[FContext
, Hash
] = *FContextHash
;
825 auto Offset
= readNumber
<uint64_t>();
826 if (std::error_code EC
= Offset
.getError())
829 if (UseFuncOffsetList
)
830 FuncOffsetList
.emplace_back(FContext
, *Offset
);
832 // Because Porfiles replace existing value with new value if collision
833 // happens, we also use the latest offset so that they are consistent.
834 FuncOffsetTable
[Hash
] = *Offset
;
837 return sampleprof_error::success
;
840 std::error_code
SampleProfileReaderExtBinaryBase::readFuncProfiles() {
841 // Collect functions used by current module if the Reader has been
843 // collectFuncsFromModule uses FunctionSamples::getCanonicalFnName
844 // which will query FunctionSamples::HasUniqSuffix, so it has to be
845 // called after FunctionSamples::HasUniqSuffix is set, i.e. after
846 // NameTable section is read.
847 bool LoadFuncsToBeUsed
= collectFuncsFromModule();
849 // When LoadFuncsToBeUsed is false, we are using LLVM tool, need to read all
851 const uint8_t *Start
= Data
;
852 if (!LoadFuncsToBeUsed
) {
854 if (std::error_code EC
= readFuncProfile(Data
))
857 assert(Data
== End
&& "More data is read than expected");
859 // Load function profiles on demand.
861 for (auto Name
: FuncsToUse
) {
862 Remapper
->insert(Name
);
867 assert(useFuncOffsetList());
868 DenseSet
<uint64_t> FuncGuidsToUse
;
870 for (auto Name
: FuncsToUse
)
871 FuncGuidsToUse
.insert(Function::getGUID(Name
));
874 // For each function in current module, load all context profiles for
875 // the function as well as their callee contexts which can help profile
876 // guided importing for ThinLTO. This can be achieved by walking
877 // through an ordered context container, where contexts are laid out
878 // as if they were walked in preorder of a context trie. While
879 // traversing the trie, a link to the highest common ancestor node is
880 // kept so that all of its decendants will be loaded.
881 const SampleContext
*CommonContext
= nullptr;
882 for (const auto &NameOffset
: FuncOffsetList
) {
883 const auto &FContext
= NameOffset
.first
;
884 FunctionId FName
= FContext
.getFunction();
885 StringRef FNameString
;
887 FNameString
= FName
.stringRef();
889 // For function in the current module, keep its farthest ancestor
890 // context. This can be used to load itself and its child and
892 if ((useMD5() && FuncGuidsToUse
.count(FName
.getHashCode())) ||
893 (!useMD5() && (FuncsToUse
.count(FNameString
) ||
894 (Remapper
&& Remapper
->exist(FNameString
))))) {
895 if (!CommonContext
|| !CommonContext
->IsPrefixOf(FContext
))
896 CommonContext
= &FContext
;
899 if (CommonContext
== &FContext
||
900 (CommonContext
&& CommonContext
->IsPrefixOf(FContext
))) {
901 // Load profile for the current context which originated from
902 // the common ancestor.
903 const uint8_t *FuncProfileAddr
= Start
+ NameOffset
.second
;
904 if (std::error_code EC
= readFuncProfile(FuncProfileAddr
))
908 } else if (useMD5()) {
909 assert(!useFuncOffsetList());
910 for (auto Name
: FuncsToUse
) {
911 auto GUID
= MD5Hash(Name
);
912 auto iter
= FuncOffsetTable
.find(GUID
);
913 if (iter
== FuncOffsetTable
.end())
915 const uint8_t *FuncProfileAddr
= Start
+ iter
->second
;
916 if (std::error_code EC
= readFuncProfile(FuncProfileAddr
))
919 } else if (Remapper
) {
920 assert(useFuncOffsetList());
921 for (auto NameOffset
: FuncOffsetList
) {
922 SampleContext
FContext(NameOffset
.first
);
923 auto FuncName
= FContext
.getFunction();
924 StringRef FuncNameStr
= FuncName
.stringRef();
925 if (!FuncsToUse
.count(FuncNameStr
) && !Remapper
->exist(FuncNameStr
))
927 const uint8_t *FuncProfileAddr
= Start
+ NameOffset
.second
;
928 if (std::error_code EC
= readFuncProfile(FuncProfileAddr
))
932 assert(!useFuncOffsetList());
933 for (auto Name
: FuncsToUse
) {
934 auto iter
= FuncOffsetTable
.find(MD5Hash(Name
));
935 if (iter
== FuncOffsetTable
.end())
937 const uint8_t *FuncProfileAddr
= Start
+ iter
->second
;
938 if (std::error_code EC
= readFuncProfile(FuncProfileAddr
))
944 assert((CSProfileCount
== 0 || CSProfileCount
== Profiles
.size()) &&
945 "Cannot have both context-sensitive and regular profile");
946 assert((!CSProfileCount
|| ProfileIsCS
) &&
947 "Section flag should be consistent with actual profile");
948 return sampleprof_error::success
;
951 std::error_code
SampleProfileReaderExtBinaryBase::readProfileSymbolList() {
953 ProfSymList
= std::make_unique
<ProfileSymbolList
>();
955 if (std::error_code EC
= ProfSymList
->read(Data
, End
- Data
))
959 return sampleprof_error::success
;
962 std::error_code
SampleProfileReaderExtBinaryBase::decompressSection(
963 const uint8_t *SecStart
, const uint64_t SecSize
,
964 const uint8_t *&DecompressBuf
, uint64_t &DecompressBufSize
) {
966 End
= SecStart
+ SecSize
;
967 auto DecompressSize
= readNumber
<uint64_t>();
968 if (std::error_code EC
= DecompressSize
.getError())
970 DecompressBufSize
= *DecompressSize
;
972 auto CompressSize
= readNumber
<uint64_t>();
973 if (std::error_code EC
= CompressSize
.getError())
976 if (!llvm::compression::zlib::isAvailable())
977 return sampleprof_error::zlib_unavailable
;
979 uint8_t *Buffer
= Allocator
.Allocate
<uint8_t>(DecompressBufSize
);
980 size_t UCSize
= DecompressBufSize
;
981 llvm::Error E
= compression::zlib::decompress(ArrayRef(Data
, *CompressSize
),
984 return sampleprof_error::uncompress_failed
;
985 DecompressBuf
= reinterpret_cast<const uint8_t *>(Buffer
);
986 return sampleprof_error::success
;
989 std::error_code
SampleProfileReaderExtBinaryBase::readImpl() {
990 const uint8_t *BufStart
=
991 reinterpret_cast<const uint8_t *>(Buffer
->getBufferStart());
993 for (auto &Entry
: SecHdrTable
) {
994 // Skip empty section.
998 // Skip sections without context when SkipFlatProf is true.
999 if (SkipFlatProf
&& hasSecFlag(Entry
, SecCommonFlags::SecFlagFlat
))
1002 const uint8_t *SecStart
= BufStart
+ Entry
.Offset
;
1003 uint64_t SecSize
= Entry
.Size
;
1005 // If the section is compressed, decompress it into a buffer
1006 // DecompressBuf before reading the actual data. The pointee of
1007 // 'Data' will be changed to buffer hold by DecompressBuf
1008 // temporarily when reading the actual data.
1009 bool isCompressed
= hasSecFlag(Entry
, SecCommonFlags::SecFlagCompress
);
1011 const uint8_t *DecompressBuf
;
1012 uint64_t DecompressBufSize
;
1013 if (std::error_code EC
= decompressSection(
1014 SecStart
, SecSize
, DecompressBuf
, DecompressBufSize
))
1016 SecStart
= DecompressBuf
;
1017 SecSize
= DecompressBufSize
;
1020 if (std::error_code EC
= readOneSection(SecStart
, SecSize
, Entry
))
1022 if (Data
!= SecStart
+ SecSize
)
1023 return sampleprof_error::malformed
;
1025 // Change the pointee of 'Data' from DecompressBuf to original Buffer.
1027 Data
= BufStart
+ Entry
.Offset
;
1028 End
= BufStart
+ Buffer
->getBufferSize();
1032 return sampleprof_error::success
;
1035 std::error_code
SampleProfileReaderRawBinary::verifySPMagic(uint64_t Magic
) {
1036 if (Magic
== SPMagic())
1037 return sampleprof_error::success
;
1038 return sampleprof_error::bad_magic
;
1041 std::error_code
SampleProfileReaderExtBinary::verifySPMagic(uint64_t Magic
) {
1042 if (Magic
== SPMagic(SPF_Ext_Binary
))
1043 return sampleprof_error::success
;
1044 return sampleprof_error::bad_magic
;
1047 std::error_code
SampleProfileReaderBinary::readNameTable() {
1048 auto Size
= readNumber
<size_t>();
1049 if (std::error_code EC
= Size
.getError())
1052 // Normally if useMD5 is true, the name table should have MD5 values, not
1053 // strings, however in the case that ExtBinary profile has multiple name
1054 // tables mixing string and MD5, all of them have to be normalized to use MD5,
1055 // because optimization passes can only handle either type.
1056 bool UseMD5
= useMD5();
1059 NameTable
.reserve(*Size
);
1061 MD5SampleContextTable
.clear();
1063 MD5SampleContextTable
.reserve(*Size
);
1065 // If we are using strings, delay MD5 computation since only a portion of
1066 // names are used by top level functions. Use 0 to indicate MD5 value is
1067 // to be calculated as no known string has a MD5 value of 0.
1068 MD5SampleContextTable
.resize(*Size
);
1070 for (size_t I
= 0; I
< *Size
; ++I
) {
1071 auto Name(readString());
1072 if (std::error_code EC
= Name
.getError())
1075 FunctionId
FID(*Name
);
1077 MD5SampleContextTable
.emplace_back(FID
.getHashCode());
1078 NameTable
.emplace_back(FID
);
1080 NameTable
.push_back(FunctionId(*Name
));
1083 MD5SampleContextStart
= MD5SampleContextTable
.data();
1084 return sampleprof_error::success
;
1088 SampleProfileReaderExtBinaryBase::readNameTableSec(bool IsMD5
,
1089 bool FixedLengthMD5
) {
1090 if (FixedLengthMD5
) {
1092 errs() << "If FixedLengthMD5 is true, UseMD5 has to be true";
1093 auto Size
= readNumber
<size_t>();
1094 if (std::error_code EC
= Size
.getError())
1097 assert(Data
+ (*Size
) * sizeof(uint64_t) == End
&&
1098 "Fixed length MD5 name table does not contain specified number of "
1100 if (Data
+ (*Size
) * sizeof(uint64_t) > End
)
1101 return sampleprof_error::truncated
;
1104 NameTable
.reserve(*Size
);
1105 for (size_t I
= 0; I
< *Size
; ++I
) {
1106 using namespace support
;
1107 uint64_t FID
= endian::read
<uint64_t, endianness::little
, unaligned
>(
1108 Data
+ I
* sizeof(uint64_t));
1109 NameTable
.emplace_back(FunctionId(FID
));
1112 MD5SampleContextStart
= reinterpret_cast<const uint64_t *>(Data
);
1113 Data
= Data
+ (*Size
) * sizeof(uint64_t);
1114 return sampleprof_error::success
;
1118 assert(!FixedLengthMD5
&& "FixedLengthMD5 should be unreachable here");
1119 auto Size
= readNumber
<size_t>();
1120 if (std::error_code EC
= Size
.getError())
1124 NameTable
.reserve(*Size
);
1126 MD5SampleContextTable
.resize(*Size
);
1127 for (size_t I
= 0; I
< *Size
; ++I
) {
1128 auto FID
= readNumber
<uint64_t>();
1129 if (std::error_code EC
= FID
.getError())
1132 support::endian::write64le(&MD5SampleContextTable
[I
], *FID
);
1133 NameTable
.emplace_back(FunctionId(*FID
));
1136 MD5SampleContextStart
= MD5SampleContextTable
.data();
1137 return sampleprof_error::success
;
1140 return SampleProfileReaderBinary::readNameTable();
1143 // Read in the CS name table section, which basically contains a list of context
1144 // vectors. Each element of a context vector, aka a frame, refers to the
1145 // underlying raw function names that are stored in the name table, as well as
1146 // a callsite identifier that only makes sense for non-leaf frames.
1147 std::error_code
SampleProfileReaderExtBinaryBase::readCSNameTableSec() {
1148 auto Size
= readNumber
<size_t>();
1149 if (std::error_code EC
= Size
.getError())
1152 CSNameTable
.clear();
1153 CSNameTable
.reserve(*Size
);
1155 // Delay MD5 computation of CS context until they are needed. Use 0 to
1156 // indicate MD5 value is to be calculated as no known string has a MD5
1158 MD5SampleContextTable
.clear();
1159 MD5SampleContextTable
.resize(*Size
);
1160 MD5SampleContextStart
= MD5SampleContextTable
.data();
1162 for (size_t I
= 0; I
< *Size
; ++I
) {
1163 CSNameTable
.emplace_back(SampleContextFrameVector());
1164 auto ContextSize
= readNumber
<uint32_t>();
1165 if (std::error_code EC
= ContextSize
.getError())
1167 for (uint32_t J
= 0; J
< *ContextSize
; ++J
) {
1168 auto FName(readStringFromTable());
1169 if (std::error_code EC
= FName
.getError())
1171 auto LineOffset
= readNumber
<uint64_t>();
1172 if (std::error_code EC
= LineOffset
.getError())
1175 if (!isOffsetLegal(*LineOffset
))
1176 return std::error_code();
1178 auto Discriminator
= readNumber
<uint64_t>();
1179 if (std::error_code EC
= Discriminator
.getError())
1182 CSNameTable
.back().emplace_back(
1183 FName
.get(), LineLocation(LineOffset
.get(), Discriminator
.get()));
1187 return sampleprof_error::success
;
1191 SampleProfileReaderExtBinaryBase::readFuncMetadata(bool ProfileHasAttribute
,
1192 FunctionSamples
*FProfile
) {
1194 if (ProfileIsProbeBased
) {
1195 auto Checksum
= readNumber
<uint64_t>();
1196 if (std::error_code EC
= Checksum
.getError())
1199 FProfile
->setFunctionHash(*Checksum
);
1202 if (ProfileHasAttribute
) {
1203 auto Attributes
= readNumber
<uint32_t>();
1204 if (std::error_code EC
= Attributes
.getError())
1207 FProfile
->getContext().setAllAttributes(*Attributes
);
1211 // Read all the attributes for inlined function calls.
1212 auto NumCallsites
= readNumber
<uint32_t>();
1213 if (std::error_code EC
= NumCallsites
.getError())
1216 for (uint32_t J
= 0; J
< *NumCallsites
; ++J
) {
1217 auto LineOffset
= readNumber
<uint64_t>();
1218 if (std::error_code EC
= LineOffset
.getError())
1221 auto Discriminator
= readNumber
<uint64_t>();
1222 if (std::error_code EC
= Discriminator
.getError())
1225 auto FContextHash(readSampleContextFromTable());
1226 if (std::error_code EC
= FContextHash
.getError())
1229 auto &[FContext
, Hash
] = *FContextHash
;
1230 FunctionSamples
*CalleeProfile
= nullptr;
1232 CalleeProfile
= const_cast<FunctionSamples
*>(
1233 &FProfile
->functionSamplesAt(LineLocation(
1235 *Discriminator
))[FContext
.getFunction()]);
1237 if (std::error_code EC
=
1238 readFuncMetadata(ProfileHasAttribute
, CalleeProfile
))
1244 return sampleprof_error::success
;
1248 SampleProfileReaderExtBinaryBase::readFuncMetadata(bool ProfileHasAttribute
) {
1249 while (Data
< End
) {
1250 auto FContextHash(readSampleContextFromTable());
1251 if (std::error_code EC
= FContextHash
.getError())
1253 auto &[FContext
, Hash
] = *FContextHash
;
1254 FunctionSamples
*FProfile
= nullptr;
1255 auto It
= Profiles
.find(FContext
);
1256 if (It
!= Profiles
.end())
1257 FProfile
= &It
->second
;
1259 if (std::error_code EC
= readFuncMetadata(ProfileHasAttribute
, FProfile
))
1263 assert(Data
== End
&& "More data is read than expected");
1264 return sampleprof_error::success
;
1268 SampleProfileReaderExtBinaryBase::readSecHdrTableEntry(uint64_t Idx
) {
1269 SecHdrTableEntry Entry
;
1270 auto Type
= readUnencodedNumber
<uint64_t>();
1271 if (std::error_code EC
= Type
.getError())
1273 Entry
.Type
= static_cast<SecType
>(*Type
);
1275 auto Flags
= readUnencodedNumber
<uint64_t>();
1276 if (std::error_code EC
= Flags
.getError())
1278 Entry
.Flags
= *Flags
;
1280 auto Offset
= readUnencodedNumber
<uint64_t>();
1281 if (std::error_code EC
= Offset
.getError())
1283 Entry
.Offset
= *Offset
;
1285 auto Size
= readUnencodedNumber
<uint64_t>();
1286 if (std::error_code EC
= Size
.getError())
1290 Entry
.LayoutIndex
= Idx
;
1291 SecHdrTable
.push_back(std::move(Entry
));
1292 return sampleprof_error::success
;
1295 std::error_code
SampleProfileReaderExtBinaryBase::readSecHdrTable() {
1296 auto EntryNum
= readUnencodedNumber
<uint64_t>();
1297 if (std::error_code EC
= EntryNum
.getError())
1300 for (uint64_t i
= 0; i
< (*EntryNum
); i
++)
1301 if (std::error_code EC
= readSecHdrTableEntry(i
))
1304 return sampleprof_error::success
;
1307 std::error_code
SampleProfileReaderExtBinaryBase::readHeader() {
1308 const uint8_t *BufStart
=
1309 reinterpret_cast<const uint8_t *>(Buffer
->getBufferStart());
1311 End
= BufStart
+ Buffer
->getBufferSize();
1313 if (std::error_code EC
= readMagicIdent())
1316 if (std::error_code EC
= readSecHdrTable())
1319 return sampleprof_error::success
;
1322 uint64_t SampleProfileReaderExtBinaryBase::getSectionSize(SecType Type
) {
1324 for (auto &Entry
: SecHdrTable
) {
1325 if (Entry
.Type
== Type
)
1331 uint64_t SampleProfileReaderExtBinaryBase::getFileSize() {
1332 // Sections in SecHdrTable is not necessarily in the same order as
1333 // sections in the profile because section like FuncOffsetTable needs
1334 // to be written after section LBRProfile but needs to be read before
1335 // section LBRProfile, so we cannot simply use the last entry in
1336 // SecHdrTable to calculate the file size.
1337 uint64_t FileSize
= 0;
1338 for (auto &Entry
: SecHdrTable
) {
1339 FileSize
= std::max(Entry
.Offset
+ Entry
.Size
, FileSize
);
1344 static std::string
getSecFlagsStr(const SecHdrTableEntry
&Entry
) {
1346 if (hasSecFlag(Entry
, SecCommonFlags::SecFlagCompress
))
1347 Flags
.append("{compressed,");
1351 if (hasSecFlag(Entry
, SecCommonFlags::SecFlagFlat
))
1352 Flags
.append("flat,");
1354 switch (Entry
.Type
) {
1356 if (hasSecFlag(Entry
, SecNameTableFlags::SecFlagFixedLengthMD5
))
1357 Flags
.append("fixlenmd5,");
1358 else if (hasSecFlag(Entry
, SecNameTableFlags::SecFlagMD5Name
))
1359 Flags
.append("md5,");
1360 if (hasSecFlag(Entry
, SecNameTableFlags::SecFlagUniqSuffix
))
1361 Flags
.append("uniq,");
1363 case SecProfSummary
:
1364 if (hasSecFlag(Entry
, SecProfSummaryFlags::SecFlagPartial
))
1365 Flags
.append("partial,");
1366 if (hasSecFlag(Entry
, SecProfSummaryFlags::SecFlagFullContext
))
1367 Flags
.append("context,");
1368 if (hasSecFlag(Entry
, SecProfSummaryFlags::SecFlagIsPreInlined
))
1369 Flags
.append("preInlined,");
1370 if (hasSecFlag(Entry
, SecProfSummaryFlags::SecFlagFSDiscriminator
))
1371 Flags
.append("fs-discriminator,");
1373 case SecFuncOffsetTable
:
1374 if (hasSecFlag(Entry
, SecFuncOffsetFlags::SecFlagOrdered
))
1375 Flags
.append("ordered,");
1377 case SecFuncMetadata
:
1378 if (hasSecFlag(Entry
, SecFuncMetadataFlags::SecFlagIsProbeBased
))
1379 Flags
.append("probe,");
1380 if (hasSecFlag(Entry
, SecFuncMetadataFlags::SecFlagHasAttribute
))
1381 Flags
.append("attr,");
1386 char &last
= Flags
.back();
1394 bool SampleProfileReaderExtBinaryBase::dumpSectionInfo(raw_ostream
&OS
) {
1395 uint64_t TotalSecsSize
= 0;
1396 for (auto &Entry
: SecHdrTable
) {
1397 OS
<< getSecName(Entry
.Type
) << " - Offset: " << Entry
.Offset
1398 << ", Size: " << Entry
.Size
<< ", Flags: " << getSecFlagsStr(Entry
)
1401 TotalSecsSize
+= Entry
.Size
;
1403 uint64_t HeaderSize
= SecHdrTable
.front().Offset
;
1404 assert(HeaderSize
+ TotalSecsSize
== getFileSize() &&
1405 "Size of 'header + sections' doesn't match the total size of profile");
1407 OS
<< "Header Size: " << HeaderSize
<< "\n";
1408 OS
<< "Total Sections Size: " << TotalSecsSize
<< "\n";
1409 OS
<< "File Size: " << getFileSize() << "\n";
1413 std::error_code
SampleProfileReaderBinary::readMagicIdent() {
1414 // Read and check the magic identifier.
1415 auto Magic
= readNumber
<uint64_t>();
1416 if (std::error_code EC
= Magic
.getError())
1418 else if (std::error_code EC
= verifySPMagic(*Magic
))
1421 // Read the version number.
1422 auto Version
= readNumber
<uint64_t>();
1423 if (std::error_code EC
= Version
.getError())
1425 else if (*Version
!= SPVersion())
1426 return sampleprof_error::unsupported_version
;
1428 return sampleprof_error::success
;
1431 std::error_code
SampleProfileReaderBinary::readHeader() {
1432 Data
= reinterpret_cast<const uint8_t *>(Buffer
->getBufferStart());
1433 End
= Data
+ Buffer
->getBufferSize();
1435 if (std::error_code EC
= readMagicIdent())
1438 if (std::error_code EC
= readSummary())
1441 if (std::error_code EC
= readNameTable())
1443 return sampleprof_error::success
;
1446 std::error_code
SampleProfileReaderBinary::readSummaryEntry(
1447 std::vector
<ProfileSummaryEntry
> &Entries
) {
1448 auto Cutoff
= readNumber
<uint64_t>();
1449 if (std::error_code EC
= Cutoff
.getError())
1452 auto MinBlockCount
= readNumber
<uint64_t>();
1453 if (std::error_code EC
= MinBlockCount
.getError())
1456 auto NumBlocks
= readNumber
<uint64_t>();
1457 if (std::error_code EC
= NumBlocks
.getError())
1460 Entries
.emplace_back(*Cutoff
, *MinBlockCount
, *NumBlocks
);
1461 return sampleprof_error::success
;
1464 std::error_code
SampleProfileReaderBinary::readSummary() {
1465 auto TotalCount
= readNumber
<uint64_t>();
1466 if (std::error_code EC
= TotalCount
.getError())
1469 auto MaxBlockCount
= readNumber
<uint64_t>();
1470 if (std::error_code EC
= MaxBlockCount
.getError())
1473 auto MaxFunctionCount
= readNumber
<uint64_t>();
1474 if (std::error_code EC
= MaxFunctionCount
.getError())
1477 auto NumBlocks
= readNumber
<uint64_t>();
1478 if (std::error_code EC
= NumBlocks
.getError())
1481 auto NumFunctions
= readNumber
<uint64_t>();
1482 if (std::error_code EC
= NumFunctions
.getError())
1485 auto NumSummaryEntries
= readNumber
<uint64_t>();
1486 if (std::error_code EC
= NumSummaryEntries
.getError())
1489 std::vector
<ProfileSummaryEntry
> Entries
;
1490 for (unsigned i
= 0; i
< *NumSummaryEntries
; i
++) {
1491 std::error_code EC
= readSummaryEntry(Entries
);
1492 if (EC
!= sampleprof_error::success
)
1495 Summary
= std::make_unique
<ProfileSummary
>(
1496 ProfileSummary::PSK_Sample
, Entries
, *TotalCount
, *MaxBlockCount
, 0,
1497 *MaxFunctionCount
, *NumBlocks
, *NumFunctions
);
1499 return sampleprof_error::success
;
1502 bool SampleProfileReaderRawBinary::hasFormat(const MemoryBuffer
&Buffer
) {
1503 const uint8_t *Data
=
1504 reinterpret_cast<const uint8_t *>(Buffer
.getBufferStart());
1505 uint64_t Magic
= decodeULEB128(Data
);
1506 return Magic
== SPMagic();
1509 bool SampleProfileReaderExtBinary::hasFormat(const MemoryBuffer
&Buffer
) {
1510 const uint8_t *Data
=
1511 reinterpret_cast<const uint8_t *>(Buffer
.getBufferStart());
1512 uint64_t Magic
= decodeULEB128(Data
);
1513 return Magic
== SPMagic(SPF_Ext_Binary
);
1516 std::error_code
SampleProfileReaderGCC::skipNextWord() {
1518 if (!GcovBuffer
.readInt(dummy
))
1519 return sampleprof_error::truncated
;
1520 return sampleprof_error::success
;
1523 template <typename T
> ErrorOr
<T
> SampleProfileReaderGCC::readNumber() {
1524 if (sizeof(T
) <= sizeof(uint32_t)) {
1526 if (GcovBuffer
.readInt(Val
) && Val
<= std::numeric_limits
<T
>::max())
1527 return static_cast<T
>(Val
);
1528 } else if (sizeof(T
) <= sizeof(uint64_t)) {
1530 if (GcovBuffer
.readInt64(Val
) && Val
<= std::numeric_limits
<T
>::max())
1531 return static_cast<T
>(Val
);
1534 std::error_code EC
= sampleprof_error::malformed
;
1535 reportError(0, EC
.message());
1539 ErrorOr
<StringRef
> SampleProfileReaderGCC::readString() {
1541 if (!GcovBuffer
.readString(Str
))
1542 return sampleprof_error::truncated
;
1546 std::error_code
SampleProfileReaderGCC::readHeader() {
1547 // Read the magic identifier.
1548 if (!GcovBuffer
.readGCDAFormat())
1549 return sampleprof_error::unrecognized_format
;
1551 // Read the version number. Note - the GCC reader does not validate this
1552 // version, but the profile creator generates v704.
1553 GCOV::GCOVVersion version
;
1554 if (!GcovBuffer
.readGCOVVersion(version
))
1555 return sampleprof_error::unrecognized_format
;
1557 if (version
!= GCOV::V407
)
1558 return sampleprof_error::unsupported_version
;
1560 // Skip the empty integer.
1561 if (std::error_code EC
= skipNextWord())
1564 return sampleprof_error::success
;
1567 std::error_code
SampleProfileReaderGCC::readSectionTag(uint32_t Expected
) {
1569 if (!GcovBuffer
.readInt(Tag
))
1570 return sampleprof_error::truncated
;
1572 if (Tag
!= Expected
)
1573 return sampleprof_error::malformed
;
1575 if (std::error_code EC
= skipNextWord())
1578 return sampleprof_error::success
;
1581 std::error_code
SampleProfileReaderGCC::readNameTable() {
1582 if (std::error_code EC
= readSectionTag(GCOVTagAFDOFileNames
))
1586 if (!GcovBuffer
.readInt(Size
))
1587 return sampleprof_error::truncated
;
1589 for (uint32_t I
= 0; I
< Size
; ++I
) {
1591 if (!GcovBuffer
.readString(Str
))
1592 return sampleprof_error::truncated
;
1593 Names
.push_back(std::string(Str
));
1596 return sampleprof_error::success
;
1599 std::error_code
SampleProfileReaderGCC::readFunctionProfiles() {
1600 if (std::error_code EC
= readSectionTag(GCOVTagAFDOFunction
))
1603 uint32_t NumFunctions
;
1604 if (!GcovBuffer
.readInt(NumFunctions
))
1605 return sampleprof_error::truncated
;
1607 InlineCallStack Stack
;
1608 for (uint32_t I
= 0; I
< NumFunctions
; ++I
)
1609 if (std::error_code EC
= readOneFunctionProfile(Stack
, true, 0))
1613 return sampleprof_error::success
;
1616 std::error_code
SampleProfileReaderGCC::readOneFunctionProfile(
1617 const InlineCallStack
&InlineStack
, bool Update
, uint32_t Offset
) {
1618 uint64_t HeadCount
= 0;
1619 if (InlineStack
.size() == 0)
1620 if (!GcovBuffer
.readInt64(HeadCount
))
1621 return sampleprof_error::truncated
;
1624 if (!GcovBuffer
.readInt(NameIdx
))
1625 return sampleprof_error::truncated
;
1627 StringRef
Name(Names
[NameIdx
]);
1629 uint32_t NumPosCounts
;
1630 if (!GcovBuffer
.readInt(NumPosCounts
))
1631 return sampleprof_error::truncated
;
1633 uint32_t NumCallsites
;
1634 if (!GcovBuffer
.readInt(NumCallsites
))
1635 return sampleprof_error::truncated
;
1637 FunctionSamples
*FProfile
= nullptr;
1638 if (InlineStack
.size() == 0) {
1639 // If this is a top function that we have already processed, do not
1640 // update its profile again. This happens in the presence of
1641 // function aliases. Since these aliases share the same function
1642 // body, there will be identical replicated profiles for the
1643 // original function. In this case, we simply not bother updating
1644 // the profile of the original function.
1645 FProfile
= &Profiles
[FunctionId(Name
)];
1646 FProfile
->addHeadSamples(HeadCount
);
1647 if (FProfile
->getTotalSamples() > 0)
1650 // Otherwise, we are reading an inlined instance. The top of the
1651 // inline stack contains the profile of the caller. Insert this
1652 // callee in the caller's CallsiteMap.
1653 FunctionSamples
*CallerProfile
= InlineStack
.front();
1654 uint32_t LineOffset
= Offset
>> 16;
1655 uint32_t Discriminator
= Offset
& 0xffff;
1656 FProfile
= &CallerProfile
->functionSamplesAt(
1657 LineLocation(LineOffset
, Discriminator
))[FunctionId(Name
)];
1659 FProfile
->setFunction(FunctionId(Name
));
1661 for (uint32_t I
= 0; I
< NumPosCounts
; ++I
) {
1663 if (!GcovBuffer
.readInt(Offset
))
1664 return sampleprof_error::truncated
;
1666 uint32_t NumTargets
;
1667 if (!GcovBuffer
.readInt(NumTargets
))
1668 return sampleprof_error::truncated
;
1671 if (!GcovBuffer
.readInt64(Count
))
1672 return sampleprof_error::truncated
;
1674 // The line location is encoded in the offset as:
1675 // high 16 bits: line offset to the start of the function.
1676 // low 16 bits: discriminator.
1677 uint32_t LineOffset
= Offset
>> 16;
1678 uint32_t Discriminator
= Offset
& 0xffff;
1680 InlineCallStack NewStack
;
1681 NewStack
.push_back(FProfile
);
1682 llvm::append_range(NewStack
, InlineStack
);
1684 // Walk up the inline stack, adding the samples on this line to
1685 // the total sample count of the callers in the chain.
1686 for (auto *CallerProfile
: NewStack
)
1687 CallerProfile
->addTotalSamples(Count
);
1689 // Update the body samples for the current profile.
1690 FProfile
->addBodySamples(LineOffset
, Discriminator
, Count
);
1693 // Process the list of functions called at an indirect call site.
1694 // These are all the targets that a function pointer (or virtual
1695 // function) resolved at runtime.
1696 for (uint32_t J
= 0; J
< NumTargets
; J
++) {
1698 if (!GcovBuffer
.readInt(HistVal
))
1699 return sampleprof_error::truncated
;
1701 if (HistVal
!= HIST_TYPE_INDIR_CALL_TOPN
)
1702 return sampleprof_error::malformed
;
1705 if (!GcovBuffer
.readInt64(TargetIdx
))
1706 return sampleprof_error::truncated
;
1707 StringRef
TargetName(Names
[TargetIdx
]);
1709 uint64_t TargetCount
;
1710 if (!GcovBuffer
.readInt64(TargetCount
))
1711 return sampleprof_error::truncated
;
1714 FProfile
->addCalledTargetSamples(LineOffset
, Discriminator
,
1715 FunctionId(TargetName
),
1720 // Process all the inlined callers into the current function. These
1721 // are all the callsites that were inlined into this function.
1722 for (uint32_t I
= 0; I
< NumCallsites
; I
++) {
1723 // The offset is encoded as:
1724 // high 16 bits: line offset to the start of the function.
1725 // low 16 bits: discriminator.
1727 if (!GcovBuffer
.readInt(Offset
))
1728 return sampleprof_error::truncated
;
1729 InlineCallStack NewStack
;
1730 NewStack
.push_back(FProfile
);
1731 llvm::append_range(NewStack
, InlineStack
);
1732 if (std::error_code EC
= readOneFunctionProfile(NewStack
, Update
, Offset
))
1736 return sampleprof_error::success
;
1739 /// Read a GCC AutoFDO profile.
1741 /// This format is generated by the Linux Perf conversion tool at
1742 /// https://github.com/google/autofdo.
1743 std::error_code
SampleProfileReaderGCC::readImpl() {
1744 assert(!ProfileIsFSDisciminator
&& "Gcc profiles not support FSDisciminator");
1745 // Read the string table.
1746 if (std::error_code EC
= readNameTable())
1749 // Read the source profile.
1750 if (std::error_code EC
= readFunctionProfiles())
1753 return sampleprof_error::success
;
1756 bool SampleProfileReaderGCC::hasFormat(const MemoryBuffer
&Buffer
) {
1757 StringRef
Magic(reinterpret_cast<const char *>(Buffer
.getBufferStart()));
1758 return Magic
== "adcg*704";
1761 void SampleProfileReaderItaniumRemapper::applyRemapping(LLVMContext
&Ctx
) {
1762 // If the reader uses MD5 to represent string, we can't remap it because
1763 // we don't know what the original function names were.
1764 if (Reader
.useMD5()) {
1765 Ctx
.diagnose(DiagnosticInfoSampleProfile(
1766 Reader
.getBuffer()->getBufferIdentifier(),
1767 "Profile data remapping cannot be applied to profile data "
1768 "using MD5 names (original mangled names are not available).",
1773 // CSSPGO-TODO: Remapper is not yet supported.
1774 // We will need to remap the entire context string.
1775 assert(Remappings
&& "should be initialized while creating remapper");
1776 for (auto &Sample
: Reader
.getProfiles()) {
1777 DenseSet
<FunctionId
> NamesInSample
;
1778 Sample
.second
.findAllNames(NamesInSample
);
1779 for (auto &Name
: NamesInSample
) {
1780 StringRef NameStr
= Name
.stringRef();
1781 if (auto Key
= Remappings
->insert(NameStr
))
1782 NameMap
.insert({Key
, NameStr
});
1786 RemappingApplied
= true;
1789 std::optional
<StringRef
>
1790 SampleProfileReaderItaniumRemapper::lookUpNameInProfile(StringRef Fname
) {
1791 if (auto Key
= Remappings
->lookup(Fname
)) {
1792 StringRef Result
= NameMap
.lookup(Key
);
1793 if (!Result
.empty())
1796 return std::nullopt
;
1799 /// Prepare a memory buffer for the contents of \p Filename.
1801 /// \returns an error code indicating the status of the buffer.
1802 static ErrorOr
<std::unique_ptr
<MemoryBuffer
>>
1803 setupMemoryBuffer(const Twine
&Filename
, vfs::FileSystem
&FS
) {
1804 auto BufferOrErr
= Filename
.str() == "-" ? MemoryBuffer::getSTDIN()
1805 : FS
.getBufferForFile(Filename
);
1806 if (std::error_code EC
= BufferOrErr
.getError())
1808 auto Buffer
= std::move(BufferOrErr
.get());
1810 return std::move(Buffer
);
1813 /// Create a sample profile reader based on the format of the input file.
1815 /// \param Filename The file to open.
1817 /// \param C The LLVM context to use to emit diagnostics.
1819 /// \param P The FSDiscriminatorPass.
1821 /// \param RemapFilename The file used for profile remapping.
1823 /// \returns an error code indicating the status of the created reader.
1824 ErrorOr
<std::unique_ptr
<SampleProfileReader
>>
1825 SampleProfileReader::create(const std::string Filename
, LLVMContext
&C
,
1826 vfs::FileSystem
&FS
, FSDiscriminatorPass P
,
1827 const std::string RemapFilename
) {
1828 auto BufferOrError
= setupMemoryBuffer(Filename
, FS
);
1829 if (std::error_code EC
= BufferOrError
.getError())
1831 return create(BufferOrError
.get(), C
, FS
, P
, RemapFilename
);
1834 /// Create a sample profile remapper from the given input, to remap the
1835 /// function names in the given profile data.
1837 /// \param Filename The file to open.
1839 /// \param Reader The profile reader the remapper is going to be applied to.
1841 /// \param C The LLVM context to use to emit diagnostics.
1843 /// \returns an error code indicating the status of the created reader.
1844 ErrorOr
<std::unique_ptr
<SampleProfileReaderItaniumRemapper
>>
1845 SampleProfileReaderItaniumRemapper::create(const std::string Filename
,
1846 vfs::FileSystem
&FS
,
1847 SampleProfileReader
&Reader
,
1849 auto BufferOrError
= setupMemoryBuffer(Filename
, FS
);
1850 if (std::error_code EC
= BufferOrError
.getError())
1852 return create(BufferOrError
.get(), Reader
, C
);
1855 /// Create a sample profile remapper from the given input, to remap the
1856 /// function names in the given profile data.
1858 /// \param B The memory buffer to create the reader from (assumes ownership).
1860 /// \param C The LLVM context to use to emit diagnostics.
1862 /// \param Reader The profile reader the remapper is going to be applied to.
1864 /// \returns an error code indicating the status of the created reader.
1865 ErrorOr
<std::unique_ptr
<SampleProfileReaderItaniumRemapper
>>
1866 SampleProfileReaderItaniumRemapper::create(std::unique_ptr
<MemoryBuffer
> &B
,
1867 SampleProfileReader
&Reader
,
1869 auto Remappings
= std::make_unique
<SymbolRemappingReader
>();
1870 if (Error E
= Remappings
->read(*B
)) {
1872 std::move(E
), [&](const SymbolRemappingParseError
&ParseError
) {
1873 C
.diagnose(DiagnosticInfoSampleProfile(B
->getBufferIdentifier(),
1874 ParseError
.getLineNum(),
1875 ParseError
.getMessage()));
1877 return sampleprof_error::malformed
;
1880 return std::make_unique
<SampleProfileReaderItaniumRemapper
>(
1881 std::move(B
), std::move(Remappings
), Reader
);
1884 /// Create a sample profile reader based on the format of the input data.
1886 /// \param B The memory buffer to create the reader from (assumes ownership).
1888 /// \param C The LLVM context to use to emit diagnostics.
1890 /// \param P The FSDiscriminatorPass.
1892 /// \param RemapFilename The file used for profile remapping.
1894 /// \returns an error code indicating the status of the created reader.
1895 ErrorOr
<std::unique_ptr
<SampleProfileReader
>>
1896 SampleProfileReader::create(std::unique_ptr
<MemoryBuffer
> &B
, LLVMContext
&C
,
1897 vfs::FileSystem
&FS
, FSDiscriminatorPass P
,
1898 const std::string RemapFilename
) {
1899 std::unique_ptr
<SampleProfileReader
> Reader
;
1900 if (SampleProfileReaderRawBinary::hasFormat(*B
))
1901 Reader
.reset(new SampleProfileReaderRawBinary(std::move(B
), C
));
1902 else if (SampleProfileReaderExtBinary::hasFormat(*B
))
1903 Reader
.reset(new SampleProfileReaderExtBinary(std::move(B
), C
));
1904 else if (SampleProfileReaderGCC::hasFormat(*B
))
1905 Reader
.reset(new SampleProfileReaderGCC(std::move(B
), C
));
1906 else if (SampleProfileReaderText::hasFormat(*B
))
1907 Reader
.reset(new SampleProfileReaderText(std::move(B
), C
));
1909 return sampleprof_error::unrecognized_format
;
1911 if (!RemapFilename
.empty()) {
1912 auto ReaderOrErr
= SampleProfileReaderItaniumRemapper::create(
1913 RemapFilename
, FS
, *Reader
, C
);
1914 if (std::error_code EC
= ReaderOrErr
.getError()) {
1915 std::string Msg
= "Could not create remapper: " + EC
.message();
1916 C
.diagnose(DiagnosticInfoSampleProfile(RemapFilename
, Msg
));
1919 Reader
->Remapper
= std::move(ReaderOrErr
.get());
1922 if (std::error_code EC
= Reader
->readHeader()) {
1926 Reader
->setDiscriminatorMaskedBitFrom(P
);
1928 return std::move(Reader
);
1931 // For text and GCC file formats, we compute the summary after reading the
1932 // profile. Binary format has the profile summary in its header.
1933 void SampleProfileReader::computeSummary() {
1934 SampleProfileSummaryBuilder
Builder(ProfileSummaryBuilder::DefaultCutoffs
);
1935 Summary
= Builder
.computeSummaryForProfiles(Profiles
);