[ORC] Add std::tuple support to SimplePackedSerialization.
[llvm-project.git] / llvm / lib / ProfileData / InstrProfReader.cpp
blobd7b8844ff6fb1a54a0f1feea6d5693cbec643308
1 //===- InstrProfReader.cpp - Instrumented profiling reader ----------------===//
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 support for reading profiling data for clang's
10 // instrumentation based PGO and coverage.
12 //===----------------------------------------------------------------------===//
14 #include "llvm/ProfileData/InstrProfReader.h"
15 #include "llvm/ADT/ArrayRef.h"
16 #include "llvm/ADT/DenseMap.h"
17 #include "llvm/ADT/STLExtras.h"
18 #include "llvm/ADT/StringExtras.h"
19 #include "llvm/ADT/StringRef.h"
20 #include "llvm/IR/ProfileSummary.h"
21 #include "llvm/ProfileData/InstrProf.h"
22 #include "llvm/ProfileData/ProfileCommon.h"
23 #include "llvm/Support/Endian.h"
24 #include "llvm/Support/Error.h"
25 #include "llvm/Support/ErrorOr.h"
26 #include "llvm/Support/MemoryBuffer.h"
27 #include "llvm/Support/SwapByteOrder.h"
28 #include "llvm/Support/SymbolRemappingReader.h"
29 #include <algorithm>
30 #include <cctype>
31 #include <cstddef>
32 #include <cstdint>
33 #include <limits>
34 #include <memory>
35 #include <system_error>
36 #include <utility>
37 #include <vector>
39 using namespace llvm;
41 static Expected<std::unique_ptr<MemoryBuffer>>
42 setupMemoryBuffer(const Twine &Path) {
43 ErrorOr<std::unique_ptr<MemoryBuffer>> BufferOrErr =
44 MemoryBuffer::getFileOrSTDIN(Path, /*IsText=*/true);
45 if (std::error_code EC = BufferOrErr.getError())
46 return errorCodeToError(EC);
47 return std::move(BufferOrErr.get());
50 static Error initializeReader(InstrProfReader &Reader) {
51 return Reader.readHeader();
54 Expected<std::unique_ptr<InstrProfReader>>
55 InstrProfReader::create(const Twine &Path) {
56 // Set up the buffer to read.
57 auto BufferOrError = setupMemoryBuffer(Path);
58 if (Error E = BufferOrError.takeError())
59 return std::move(E);
60 return InstrProfReader::create(std::move(BufferOrError.get()));
63 Expected<std::unique_ptr<InstrProfReader>>
64 InstrProfReader::create(std::unique_ptr<MemoryBuffer> Buffer) {
65 // Sanity check the buffer.
66 if (uint64_t(Buffer->getBufferSize()) > std::numeric_limits<uint64_t>::max())
67 return make_error<InstrProfError>(instrprof_error::too_large);
69 if (Buffer->getBufferSize() == 0)
70 return make_error<InstrProfError>(instrprof_error::empty_raw_profile);
72 std::unique_ptr<InstrProfReader> Result;
73 // Create the reader.
74 if (IndexedInstrProfReader::hasFormat(*Buffer))
75 Result.reset(new IndexedInstrProfReader(std::move(Buffer)));
76 else if (RawInstrProfReader64::hasFormat(*Buffer))
77 Result.reset(new RawInstrProfReader64(std::move(Buffer)));
78 else if (RawInstrProfReader32::hasFormat(*Buffer))
79 Result.reset(new RawInstrProfReader32(std::move(Buffer)));
80 else if (TextInstrProfReader::hasFormat(*Buffer))
81 Result.reset(new TextInstrProfReader(std::move(Buffer)));
82 else
83 return make_error<InstrProfError>(instrprof_error::unrecognized_format);
85 // Initialize the reader and return the result.
86 if (Error E = initializeReader(*Result))
87 return std::move(E);
89 return std::move(Result);
92 Expected<std::unique_ptr<IndexedInstrProfReader>>
93 IndexedInstrProfReader::create(const Twine &Path, const Twine &RemappingPath) {
94 // Set up the buffer to read.
95 auto BufferOrError = setupMemoryBuffer(Path);
96 if (Error E = BufferOrError.takeError())
97 return std::move(E);
99 // Set up the remapping buffer if requested.
100 std::unique_ptr<MemoryBuffer> RemappingBuffer;
101 std::string RemappingPathStr = RemappingPath.str();
102 if (!RemappingPathStr.empty()) {
103 auto RemappingBufferOrError = setupMemoryBuffer(RemappingPathStr);
104 if (Error E = RemappingBufferOrError.takeError())
105 return std::move(E);
106 RemappingBuffer = std::move(RemappingBufferOrError.get());
109 return IndexedInstrProfReader::create(std::move(BufferOrError.get()),
110 std::move(RemappingBuffer));
113 Expected<std::unique_ptr<IndexedInstrProfReader>>
114 IndexedInstrProfReader::create(std::unique_ptr<MemoryBuffer> Buffer,
115 std::unique_ptr<MemoryBuffer> RemappingBuffer) {
116 // Sanity check the buffer.
117 if (uint64_t(Buffer->getBufferSize()) > std::numeric_limits<uint64_t>::max())
118 return make_error<InstrProfError>(instrprof_error::too_large);
120 // Create the reader.
121 if (!IndexedInstrProfReader::hasFormat(*Buffer))
122 return make_error<InstrProfError>(instrprof_error::bad_magic);
123 auto Result = std::make_unique<IndexedInstrProfReader>(
124 std::move(Buffer), std::move(RemappingBuffer));
126 // Initialize the reader and return the result.
127 if (Error E = initializeReader(*Result))
128 return std::move(E);
130 return std::move(Result);
133 void InstrProfIterator::Increment() {
134 if (auto E = Reader->readNextRecord(Record)) {
135 // Handle errors in the reader.
136 InstrProfError::take(std::move(E));
137 *this = InstrProfIterator();
141 bool TextInstrProfReader::hasFormat(const MemoryBuffer &Buffer) {
142 // Verify that this really looks like plain ASCII text by checking a
143 // 'reasonable' number of characters (up to profile magic size).
144 size_t count = std::min(Buffer.getBufferSize(), sizeof(uint64_t));
145 StringRef buffer = Buffer.getBufferStart();
146 return count == 0 ||
147 std::all_of(buffer.begin(), buffer.begin() + count,
148 [](char c) { return isPrint(c) || isSpace(c); });
151 // Read the profile variant flag from the header: ":FE" means this is a FE
152 // generated profile. ":IR" means this is an IR level profile. Other strings
153 // with a leading ':' will be reported an error format.
154 Error TextInstrProfReader::readHeader() {
155 Symtab.reset(new InstrProfSymtab());
156 bool IsIRInstr = false;
157 bool IsEntryFirst = false;
158 bool IsCS = false;
160 while (Line->startswith(":")) {
161 StringRef Str = Line->substr(1);
162 if (Str.equals_insensitive("ir"))
163 IsIRInstr = true;
164 else if (Str.equals_insensitive("fe"))
165 IsIRInstr = false;
166 else if (Str.equals_insensitive("csir")) {
167 IsIRInstr = true;
168 IsCS = true;
169 } else if (Str.equals_insensitive("entry_first"))
170 IsEntryFirst = true;
171 else if (Str.equals_insensitive("not_entry_first"))
172 IsEntryFirst = false;
173 else
174 return error(instrprof_error::bad_header);
175 ++Line;
177 IsIRLevelProfile = IsIRInstr;
178 InstrEntryBBEnabled = IsEntryFirst;
179 HasCSIRLevelProfile = IsCS;
180 return success();
183 Error
184 TextInstrProfReader::readValueProfileData(InstrProfRecord &Record) {
186 #define CHECK_LINE_END(Line) \
187 if (Line.is_at_end()) \
188 return error(instrprof_error::truncated);
189 #define READ_NUM(Str, Dst) \
190 if ((Str).getAsInteger(10, (Dst))) \
191 return error(instrprof_error::malformed);
192 #define VP_READ_ADVANCE(Val) \
193 CHECK_LINE_END(Line); \
194 uint32_t Val; \
195 READ_NUM((*Line), (Val)); \
196 Line++;
198 if (Line.is_at_end())
199 return success();
201 uint32_t NumValueKinds;
202 if (Line->getAsInteger(10, NumValueKinds)) {
203 // No value profile data
204 return success();
206 if (NumValueKinds == 0 || NumValueKinds > IPVK_Last + 1)
207 return error(instrprof_error::malformed);
208 Line++;
210 for (uint32_t VK = 0; VK < NumValueKinds; VK++) {
211 VP_READ_ADVANCE(ValueKind);
212 if (ValueKind > IPVK_Last)
213 return error(instrprof_error::malformed);
214 VP_READ_ADVANCE(NumValueSites);
215 if (!NumValueSites)
216 continue;
218 Record.reserveSites(VK, NumValueSites);
219 for (uint32_t S = 0; S < NumValueSites; S++) {
220 VP_READ_ADVANCE(NumValueData);
222 std::vector<InstrProfValueData> CurrentValues;
223 for (uint32_t V = 0; V < NumValueData; V++) {
224 CHECK_LINE_END(Line);
225 std::pair<StringRef, StringRef> VD = Line->rsplit(':');
226 uint64_t TakenCount, Value;
227 if (ValueKind == IPVK_IndirectCallTarget) {
228 if (InstrProfSymtab::isExternalSymbol(VD.first)) {
229 Value = 0;
230 } else {
231 if (Error E = Symtab->addFuncName(VD.first))
232 return E;
233 Value = IndexedInstrProf::ComputeHash(VD.first);
235 } else {
236 READ_NUM(VD.first, Value);
238 READ_NUM(VD.second, TakenCount);
239 CurrentValues.push_back({Value, TakenCount});
240 Line++;
242 Record.addValueData(ValueKind, S, CurrentValues.data(), NumValueData,
243 nullptr);
246 return success();
248 #undef CHECK_LINE_END
249 #undef READ_NUM
250 #undef VP_READ_ADVANCE
253 Error TextInstrProfReader::readNextRecord(NamedInstrProfRecord &Record) {
254 // Skip empty lines and comments.
255 while (!Line.is_at_end() && (Line->empty() || Line->startswith("#")))
256 ++Line;
257 // If we hit EOF while looking for a name, we're done.
258 if (Line.is_at_end()) {
259 return error(instrprof_error::eof);
262 // Read the function name.
263 Record.Name = *Line++;
264 if (Error E = Symtab->addFuncName(Record.Name))
265 return error(std::move(E));
267 // Read the function hash.
268 if (Line.is_at_end())
269 return error(instrprof_error::truncated);
270 if ((Line++)->getAsInteger(0, Record.Hash))
271 return error(instrprof_error::malformed);
273 // Read the number of counters.
274 uint64_t NumCounters;
275 if (Line.is_at_end())
276 return error(instrprof_error::truncated);
277 if ((Line++)->getAsInteger(10, NumCounters))
278 return error(instrprof_error::malformed);
279 if (NumCounters == 0)
280 return error(instrprof_error::malformed);
282 // Read each counter and fill our internal storage with the values.
283 Record.Clear();
284 Record.Counts.reserve(NumCounters);
285 for (uint64_t I = 0; I < NumCounters; ++I) {
286 if (Line.is_at_end())
287 return error(instrprof_error::truncated);
288 uint64_t Count;
289 if ((Line++)->getAsInteger(10, Count))
290 return error(instrprof_error::malformed);
291 Record.Counts.push_back(Count);
294 // Check if value profile data exists and read it if so.
295 if (Error E = readValueProfileData(Record))
296 return error(std::move(E));
298 return success();
301 template <class IntPtrT>
302 bool RawInstrProfReader<IntPtrT>::hasFormat(const MemoryBuffer &DataBuffer) {
303 if (DataBuffer.getBufferSize() < sizeof(uint64_t))
304 return false;
305 uint64_t Magic =
306 *reinterpret_cast<const uint64_t *>(DataBuffer.getBufferStart());
307 return RawInstrProf::getMagic<IntPtrT>() == Magic ||
308 sys::getSwappedBytes(RawInstrProf::getMagic<IntPtrT>()) == Magic;
311 template <class IntPtrT>
312 Error RawInstrProfReader<IntPtrT>::readHeader() {
313 if (!hasFormat(*DataBuffer))
314 return error(instrprof_error::bad_magic);
315 if (DataBuffer->getBufferSize() < sizeof(RawInstrProf::Header))
316 return error(instrprof_error::bad_header);
317 auto *Header = reinterpret_cast<const RawInstrProf::Header *>(
318 DataBuffer->getBufferStart());
319 ShouldSwapBytes = Header->Magic != RawInstrProf::getMagic<IntPtrT>();
320 return readHeader(*Header);
323 template <class IntPtrT>
324 Error RawInstrProfReader<IntPtrT>::readNextHeader(const char *CurrentPos) {
325 const char *End = DataBuffer->getBufferEnd();
326 // Skip zero padding between profiles.
327 while (CurrentPos != End && *CurrentPos == 0)
328 ++CurrentPos;
329 // If there's nothing left, we're done.
330 if (CurrentPos == End)
331 return make_error<InstrProfError>(instrprof_error::eof);
332 // If there isn't enough space for another header, this is probably just
333 // garbage at the end of the file.
334 if (CurrentPos + sizeof(RawInstrProf::Header) > End)
335 return make_error<InstrProfError>(instrprof_error::malformed);
336 // The writer ensures each profile is padded to start at an aligned address.
337 if (reinterpret_cast<size_t>(CurrentPos) % alignof(uint64_t))
338 return make_error<InstrProfError>(instrprof_error::malformed);
339 // The magic should have the same byte order as in the previous header.
340 uint64_t Magic = *reinterpret_cast<const uint64_t *>(CurrentPos);
341 if (Magic != swap(RawInstrProf::getMagic<IntPtrT>()))
342 return make_error<InstrProfError>(instrprof_error::bad_magic);
344 // There's another profile to read, so we need to process the header.
345 auto *Header = reinterpret_cast<const RawInstrProf::Header *>(CurrentPos);
346 return readHeader(*Header);
349 template <class IntPtrT>
350 Error RawInstrProfReader<IntPtrT>::createSymtab(InstrProfSymtab &Symtab) {
351 if (Error E = Symtab.create(StringRef(NamesStart, NamesSize)))
352 return error(std::move(E));
353 for (const RawInstrProf::ProfileData<IntPtrT> *I = Data; I != DataEnd; ++I) {
354 const IntPtrT FPtr = swap(I->FunctionPointer);
355 if (!FPtr)
356 continue;
357 Symtab.mapAddress(FPtr, I->NameRef);
359 return success();
362 template <class IntPtrT>
363 Error RawInstrProfReader<IntPtrT>::readHeader(
364 const RawInstrProf::Header &Header) {
365 Version = swap(Header.Version);
366 if (GET_VERSION(Version) != RawInstrProf::Version)
367 return error(instrprof_error::unsupported_version);
369 BinaryIdsSize = swap(Header.BinaryIdsSize);
370 CountersDelta = swap(Header.CountersDelta);
371 NamesDelta = swap(Header.NamesDelta);
372 auto DataSize = swap(Header.DataSize);
373 auto PaddingBytesBeforeCounters = swap(Header.PaddingBytesBeforeCounters);
374 auto CountersSize = swap(Header.CountersSize);
375 auto PaddingBytesAfterCounters = swap(Header.PaddingBytesAfterCounters);
376 NamesSize = swap(Header.NamesSize);
377 ValueKindLast = swap(Header.ValueKindLast);
379 auto DataSizeInBytes = DataSize * sizeof(RawInstrProf::ProfileData<IntPtrT>);
380 auto PaddingSize = getNumPaddingBytes(NamesSize);
382 // Profile data starts after profile header and binary ids if exist.
383 ptrdiff_t DataOffset = sizeof(RawInstrProf::Header) + BinaryIdsSize;
384 ptrdiff_t CountersOffset =
385 DataOffset + DataSizeInBytes + PaddingBytesBeforeCounters;
386 ptrdiff_t NamesOffset = CountersOffset + (sizeof(uint64_t) * CountersSize) +
387 PaddingBytesAfterCounters;
388 ptrdiff_t ValueDataOffset = NamesOffset + NamesSize + PaddingSize;
390 auto *Start = reinterpret_cast<const char *>(&Header);
391 if (Start + ValueDataOffset > DataBuffer->getBufferEnd())
392 return error(instrprof_error::bad_header);
394 Data = reinterpret_cast<const RawInstrProf::ProfileData<IntPtrT> *>(
395 Start + DataOffset);
396 DataEnd = Data + DataSize;
398 // Binary ids start just after the header.
399 BinaryIdsStart =
400 reinterpret_cast<const uint8_t *>(&Header) + sizeof(RawInstrProf::Header);
401 CountersStart = reinterpret_cast<const uint64_t *>(Start + CountersOffset);
402 NamesStart = Start + NamesOffset;
403 ValueDataStart = reinterpret_cast<const uint8_t *>(Start + ValueDataOffset);
405 std::unique_ptr<InstrProfSymtab> NewSymtab = std::make_unique<InstrProfSymtab>();
406 if (Error E = createSymtab(*NewSymtab.get()))
407 return E;
409 Symtab = std::move(NewSymtab);
410 return success();
413 template <class IntPtrT>
414 Error RawInstrProfReader<IntPtrT>::readName(NamedInstrProfRecord &Record) {
415 Record.Name = getName(Data->NameRef);
416 return success();
419 template <class IntPtrT>
420 Error RawInstrProfReader<IntPtrT>::readFuncHash(NamedInstrProfRecord &Record) {
421 Record.Hash = swap(Data->FuncHash);
422 return success();
425 template <class IntPtrT>
426 Error RawInstrProfReader<IntPtrT>::readRawCounts(
427 InstrProfRecord &Record) {
428 uint32_t NumCounters = swap(Data->NumCounters);
429 IntPtrT CounterPtr = Data->CounterPtr;
430 if (NumCounters == 0)
431 return error(instrprof_error::malformed);
433 auto *NamesStartAsCounter = reinterpret_cast<const uint64_t *>(NamesStart);
434 ptrdiff_t MaxNumCounters = NamesStartAsCounter - CountersStart;
436 // Check bounds. Note that the counter pointer embedded in the data record
437 // may itself be corrupt.
438 if (MaxNumCounters < 0 || NumCounters > (uint32_t)MaxNumCounters)
439 return error(instrprof_error::malformed);
441 // We need to compute the in-buffer counter offset from the in-memory address
442 // distance. The initial CountersDelta is the in-memory address difference
443 // start(__llvm_prf_cnts)-start(__llvm_prf_data), so SrcData->CounterPtr -
444 // CountersDelta computes the offset into the in-buffer counter section.
446 // CountersDelta decreases as we advance to the next data record.
447 ptrdiff_t CounterOffset = getCounterOffset(CounterPtr);
448 CountersDelta -= sizeof(*Data);
449 if (CounterOffset < 0 || CounterOffset > MaxNumCounters ||
450 ((uint32_t)CounterOffset + NumCounters) > (uint32_t)MaxNumCounters)
451 return error(instrprof_error::malformed);
453 auto RawCounts = makeArrayRef(getCounter(CounterOffset), NumCounters);
455 if (ShouldSwapBytes) {
456 Record.Counts.clear();
457 Record.Counts.reserve(RawCounts.size());
458 for (uint64_t Count : RawCounts)
459 Record.Counts.push_back(swap(Count));
460 } else
461 Record.Counts = RawCounts;
463 return success();
466 template <class IntPtrT>
467 Error RawInstrProfReader<IntPtrT>::readValueProfilingData(
468 InstrProfRecord &Record) {
469 Record.clearValueData();
470 CurValueDataSize = 0;
471 // Need to match the logic in value profile dumper code in compiler-rt:
472 uint32_t NumValueKinds = 0;
473 for (uint32_t I = 0; I < IPVK_Last + 1; I++)
474 NumValueKinds += (Data->NumValueSites[I] != 0);
476 if (!NumValueKinds)
477 return success();
479 Expected<std::unique_ptr<ValueProfData>> VDataPtrOrErr =
480 ValueProfData::getValueProfData(
481 ValueDataStart, (const unsigned char *)DataBuffer->getBufferEnd(),
482 getDataEndianness());
484 if (Error E = VDataPtrOrErr.takeError())
485 return E;
487 // Note that besides deserialization, this also performs the conversion for
488 // indirect call targets. The function pointers from the raw profile are
489 // remapped into function name hashes.
490 VDataPtrOrErr.get()->deserializeTo(Record, Symtab.get());
491 CurValueDataSize = VDataPtrOrErr.get()->getSize();
492 return success();
495 template <class IntPtrT>
496 Error RawInstrProfReader<IntPtrT>::readNextRecord(NamedInstrProfRecord &Record) {
497 if (atEnd())
498 // At this point, ValueDataStart field points to the next header.
499 if (Error E = readNextHeader(getNextHeaderPos()))
500 return error(std::move(E));
502 // Read name ad set it in Record.
503 if (Error E = readName(Record))
504 return error(std::move(E));
506 // Read FuncHash and set it in Record.
507 if (Error E = readFuncHash(Record))
508 return error(std::move(E));
510 // Read raw counts and set Record.
511 if (Error E = readRawCounts(Record))
512 return error(std::move(E));
514 // Read value data and set Record.
515 if (Error E = readValueProfilingData(Record))
516 return error(std::move(E));
518 // Iterate.
519 advanceData();
520 return success();
523 template <class IntPtrT>
524 Error RawInstrProfReader<IntPtrT>::printBinaryIds(raw_ostream &OS) {
525 if (BinaryIdsSize == 0)
526 return success();
528 OS << "Binary IDs: \n";
529 const uint8_t *BI = BinaryIdsStart;
530 while (BI < BinaryIdsStart + BinaryIdsSize) {
531 uint64_t BinaryIdLen = swap(*reinterpret_cast<const uint64_t *>(BI));
532 // Increment by binary id length data type size.
533 BI += sizeof(BinaryIdLen);
534 if (BI > (const uint8_t *)DataBuffer->getBufferEnd())
535 return make_error<InstrProfError>(instrprof_error::malformed);
537 for (uint64_t I = 0; I < BinaryIdLen; I++)
538 OS << format("%02x", BI[I]);
539 OS << "\n";
541 // Increment by binary id data length.
542 BI += BinaryIdLen;
543 if (BI > (const uint8_t *)DataBuffer->getBufferEnd())
544 return make_error<InstrProfError>(instrprof_error::malformed);
547 return success();
550 namespace llvm {
552 template class RawInstrProfReader<uint32_t>;
553 template class RawInstrProfReader<uint64_t>;
555 } // end namespace llvm
557 InstrProfLookupTrait::hash_value_type
558 InstrProfLookupTrait::ComputeHash(StringRef K) {
559 return IndexedInstrProf::ComputeHash(HashType, K);
562 using data_type = InstrProfLookupTrait::data_type;
563 using offset_type = InstrProfLookupTrait::offset_type;
565 bool InstrProfLookupTrait::readValueProfilingData(
566 const unsigned char *&D, const unsigned char *const End) {
567 Expected<std::unique_ptr<ValueProfData>> VDataPtrOrErr =
568 ValueProfData::getValueProfData(D, End, ValueProfDataEndianness);
570 if (VDataPtrOrErr.takeError())
571 return false;
573 VDataPtrOrErr.get()->deserializeTo(DataBuffer.back(), nullptr);
574 D += VDataPtrOrErr.get()->TotalSize;
576 return true;
579 data_type InstrProfLookupTrait::ReadData(StringRef K, const unsigned char *D,
580 offset_type N) {
581 using namespace support;
583 // Check if the data is corrupt. If so, don't try to read it.
584 if (N % sizeof(uint64_t))
585 return data_type();
587 DataBuffer.clear();
588 std::vector<uint64_t> CounterBuffer;
590 const unsigned char *End = D + N;
591 while (D < End) {
592 // Read hash.
593 if (D + sizeof(uint64_t) >= End)
594 return data_type();
595 uint64_t Hash = endian::readNext<uint64_t, little, unaligned>(D);
597 // Initialize number of counters for GET_VERSION(FormatVersion) == 1.
598 uint64_t CountsSize = N / sizeof(uint64_t) - 1;
599 // If format version is different then read the number of counters.
600 if (GET_VERSION(FormatVersion) != IndexedInstrProf::ProfVersion::Version1) {
601 if (D + sizeof(uint64_t) > End)
602 return data_type();
603 CountsSize = endian::readNext<uint64_t, little, unaligned>(D);
605 // Read counter values.
606 if (D + CountsSize * sizeof(uint64_t) > End)
607 return data_type();
609 CounterBuffer.clear();
610 CounterBuffer.reserve(CountsSize);
611 for (uint64_t J = 0; J < CountsSize; ++J)
612 CounterBuffer.push_back(endian::readNext<uint64_t, little, unaligned>(D));
614 DataBuffer.emplace_back(K, Hash, std::move(CounterBuffer));
616 // Read value profiling data.
617 if (GET_VERSION(FormatVersion) > IndexedInstrProf::ProfVersion::Version2 &&
618 !readValueProfilingData(D, End)) {
619 DataBuffer.clear();
620 return data_type();
623 return DataBuffer;
626 template <typename HashTableImpl>
627 Error InstrProfReaderIndex<HashTableImpl>::getRecords(
628 StringRef FuncName, ArrayRef<NamedInstrProfRecord> &Data) {
629 auto Iter = HashTable->find(FuncName);
630 if (Iter == HashTable->end())
631 return make_error<InstrProfError>(instrprof_error::unknown_function);
633 Data = (*Iter);
634 if (Data.empty())
635 return make_error<InstrProfError>(instrprof_error::malformed);
637 return Error::success();
640 template <typename HashTableImpl>
641 Error InstrProfReaderIndex<HashTableImpl>::getRecords(
642 ArrayRef<NamedInstrProfRecord> &Data) {
643 if (atEnd())
644 return make_error<InstrProfError>(instrprof_error::eof);
646 Data = *RecordIterator;
648 if (Data.empty())
649 return make_error<InstrProfError>(instrprof_error::malformed);
651 return Error::success();
654 template <typename HashTableImpl>
655 InstrProfReaderIndex<HashTableImpl>::InstrProfReaderIndex(
656 const unsigned char *Buckets, const unsigned char *const Payload,
657 const unsigned char *const Base, IndexedInstrProf::HashT HashType,
658 uint64_t Version) {
659 FormatVersion = Version;
660 HashTable.reset(HashTableImpl::Create(
661 Buckets, Payload, Base,
662 typename HashTableImpl::InfoType(HashType, Version)));
663 RecordIterator = HashTable->data_begin();
666 namespace {
667 /// A remapper that does not apply any remappings.
668 class InstrProfReaderNullRemapper : public InstrProfReaderRemapper {
669 InstrProfReaderIndexBase &Underlying;
671 public:
672 InstrProfReaderNullRemapper(InstrProfReaderIndexBase &Underlying)
673 : Underlying(Underlying) {}
675 Error getRecords(StringRef FuncName,
676 ArrayRef<NamedInstrProfRecord> &Data) override {
677 return Underlying.getRecords(FuncName, Data);
682 /// A remapper that applies remappings based on a symbol remapping file.
683 template <typename HashTableImpl>
684 class llvm::InstrProfReaderItaniumRemapper
685 : public InstrProfReaderRemapper {
686 public:
687 InstrProfReaderItaniumRemapper(
688 std::unique_ptr<MemoryBuffer> RemapBuffer,
689 InstrProfReaderIndex<HashTableImpl> &Underlying)
690 : RemapBuffer(std::move(RemapBuffer)), Underlying(Underlying) {
693 /// Extract the original function name from a PGO function name.
694 static StringRef extractName(StringRef Name) {
695 // We can have multiple :-separated pieces; there can be pieces both
696 // before and after the mangled name. Find the first part that starts
697 // with '_Z'; we'll assume that's the mangled name we want.
698 std::pair<StringRef, StringRef> Parts = {StringRef(), Name};
699 while (true) {
700 Parts = Parts.second.split(':');
701 if (Parts.first.startswith("_Z"))
702 return Parts.first;
703 if (Parts.second.empty())
704 return Name;
708 /// Given a mangled name extracted from a PGO function name, and a new
709 /// form for that mangled name, reconstitute the name.
710 static void reconstituteName(StringRef OrigName, StringRef ExtractedName,
711 StringRef Replacement,
712 SmallVectorImpl<char> &Out) {
713 Out.reserve(OrigName.size() + Replacement.size() - ExtractedName.size());
714 Out.insert(Out.end(), OrigName.begin(), ExtractedName.begin());
715 Out.insert(Out.end(), Replacement.begin(), Replacement.end());
716 Out.insert(Out.end(), ExtractedName.end(), OrigName.end());
719 Error populateRemappings() override {
720 if (Error E = Remappings.read(*RemapBuffer))
721 return E;
722 for (StringRef Name : Underlying.HashTable->keys()) {
723 StringRef RealName = extractName(Name);
724 if (auto Key = Remappings.insert(RealName)) {
725 // FIXME: We could theoretically map the same equivalence class to
726 // multiple names in the profile data. If that happens, we should
727 // return NamedInstrProfRecords from all of them.
728 MappedNames.insert({Key, RealName});
731 return Error::success();
734 Error getRecords(StringRef FuncName,
735 ArrayRef<NamedInstrProfRecord> &Data) override {
736 StringRef RealName = extractName(FuncName);
737 if (auto Key = Remappings.lookup(RealName)) {
738 StringRef Remapped = MappedNames.lookup(Key);
739 if (!Remapped.empty()) {
740 if (RealName.begin() == FuncName.begin() &&
741 RealName.end() == FuncName.end())
742 FuncName = Remapped;
743 else {
744 // Try rebuilding the name from the given remapping.
745 SmallString<256> Reconstituted;
746 reconstituteName(FuncName, RealName, Remapped, Reconstituted);
747 Error E = Underlying.getRecords(Reconstituted, Data);
748 if (!E)
749 return E;
751 // If we failed because the name doesn't exist, fall back to asking
752 // about the original name.
753 if (Error Unhandled = handleErrors(
754 std::move(E), [](std::unique_ptr<InstrProfError> Err) {
755 return Err->get() == instrprof_error::unknown_function
756 ? Error::success()
757 : Error(std::move(Err));
759 return Unhandled;
763 return Underlying.getRecords(FuncName, Data);
766 private:
767 /// The memory buffer containing the remapping configuration. Remappings
768 /// holds pointers into this buffer.
769 std::unique_ptr<MemoryBuffer> RemapBuffer;
771 /// The mangling remapper.
772 SymbolRemappingReader Remappings;
774 /// Mapping from mangled name keys to the name used for the key in the
775 /// profile data.
776 /// FIXME: Can we store a location within the on-disk hash table instead of
777 /// redoing lookup?
778 DenseMap<SymbolRemappingReader::Key, StringRef> MappedNames;
780 /// The real profile data reader.
781 InstrProfReaderIndex<HashTableImpl> &Underlying;
784 bool IndexedInstrProfReader::hasFormat(const MemoryBuffer &DataBuffer) {
785 using namespace support;
787 if (DataBuffer.getBufferSize() < 8)
788 return false;
789 uint64_t Magic =
790 endian::read<uint64_t, little, aligned>(DataBuffer.getBufferStart());
791 // Verify that it's magical.
792 return Magic == IndexedInstrProf::Magic;
795 const unsigned char *
796 IndexedInstrProfReader::readSummary(IndexedInstrProf::ProfVersion Version,
797 const unsigned char *Cur, bool UseCS) {
798 using namespace IndexedInstrProf;
799 using namespace support;
801 if (Version >= IndexedInstrProf::Version4) {
802 const IndexedInstrProf::Summary *SummaryInLE =
803 reinterpret_cast<const IndexedInstrProf::Summary *>(Cur);
804 uint64_t NFields =
805 endian::byte_swap<uint64_t, little>(SummaryInLE->NumSummaryFields);
806 uint64_t NEntries =
807 endian::byte_swap<uint64_t, little>(SummaryInLE->NumCutoffEntries);
808 uint32_t SummarySize =
809 IndexedInstrProf::Summary::getSize(NFields, NEntries);
810 std::unique_ptr<IndexedInstrProf::Summary> SummaryData =
811 IndexedInstrProf::allocSummary(SummarySize);
813 const uint64_t *Src = reinterpret_cast<const uint64_t *>(SummaryInLE);
814 uint64_t *Dst = reinterpret_cast<uint64_t *>(SummaryData.get());
815 for (unsigned I = 0; I < SummarySize / sizeof(uint64_t); I++)
816 Dst[I] = endian::byte_swap<uint64_t, little>(Src[I]);
818 SummaryEntryVector DetailedSummary;
819 for (unsigned I = 0; I < SummaryData->NumCutoffEntries; I++) {
820 const IndexedInstrProf::Summary::Entry &Ent = SummaryData->getEntry(I);
821 DetailedSummary.emplace_back((uint32_t)Ent.Cutoff, Ent.MinBlockCount,
822 Ent.NumBlocks);
824 std::unique_ptr<llvm::ProfileSummary> &Summary =
825 UseCS ? this->CS_Summary : this->Summary;
827 // initialize InstrProfSummary using the SummaryData from disk.
828 Summary = std::make_unique<ProfileSummary>(
829 UseCS ? ProfileSummary::PSK_CSInstr : ProfileSummary::PSK_Instr,
830 DetailedSummary, SummaryData->get(Summary::TotalBlockCount),
831 SummaryData->get(Summary::MaxBlockCount),
832 SummaryData->get(Summary::MaxInternalBlockCount),
833 SummaryData->get(Summary::MaxFunctionCount),
834 SummaryData->get(Summary::TotalNumBlocks),
835 SummaryData->get(Summary::TotalNumFunctions));
836 return Cur + SummarySize;
837 } else {
838 // The older versions do not support a profile summary. This just computes
839 // an empty summary, which will not result in accurate hot/cold detection.
840 // We would need to call addRecord for all NamedInstrProfRecords to get the
841 // correct summary. However, this version is old (prior to early 2016) and
842 // has not been supporting an accurate summary for several years.
843 InstrProfSummaryBuilder Builder(ProfileSummaryBuilder::DefaultCutoffs);
844 Summary = Builder.getSummary();
845 return Cur;
849 Error IndexedInstrProfReader::readHeader() {
850 using namespace support;
852 const unsigned char *Start =
853 (const unsigned char *)DataBuffer->getBufferStart();
854 const unsigned char *Cur = Start;
855 if ((const unsigned char *)DataBuffer->getBufferEnd() - Cur < 24)
856 return error(instrprof_error::truncated);
858 auto *Header = reinterpret_cast<const IndexedInstrProf::Header *>(Cur);
859 Cur += sizeof(IndexedInstrProf::Header);
861 // Check the magic number.
862 uint64_t Magic = endian::byte_swap<uint64_t, little>(Header->Magic);
863 if (Magic != IndexedInstrProf::Magic)
864 return error(instrprof_error::bad_magic);
866 // Read the version.
867 uint64_t FormatVersion = endian::byte_swap<uint64_t, little>(Header->Version);
868 if (GET_VERSION(FormatVersion) >
869 IndexedInstrProf::ProfVersion::CurrentVersion)
870 return error(instrprof_error::unsupported_version);
872 Cur = readSummary((IndexedInstrProf::ProfVersion)FormatVersion, Cur,
873 /* UseCS */ false);
874 if (FormatVersion & VARIANT_MASK_CSIR_PROF)
875 Cur = readSummary((IndexedInstrProf::ProfVersion)FormatVersion, Cur,
876 /* UseCS */ true);
878 // Read the hash type and start offset.
879 IndexedInstrProf::HashT HashType = static_cast<IndexedInstrProf::HashT>(
880 endian::byte_swap<uint64_t, little>(Header->HashType));
881 if (HashType > IndexedInstrProf::HashT::Last)
882 return error(instrprof_error::unsupported_hash_type);
884 uint64_t HashOffset = endian::byte_swap<uint64_t, little>(Header->HashOffset);
886 // The rest of the file is an on disk hash table.
887 auto IndexPtr =
888 std::make_unique<InstrProfReaderIndex<OnDiskHashTableImplV3>>(
889 Start + HashOffset, Cur, Start, HashType, FormatVersion);
891 // Load the remapping table now if requested.
892 if (RemappingBuffer) {
893 Remapper = std::make_unique<
894 InstrProfReaderItaniumRemapper<OnDiskHashTableImplV3>>(
895 std::move(RemappingBuffer), *IndexPtr);
896 if (Error E = Remapper->populateRemappings())
897 return E;
898 } else {
899 Remapper = std::make_unique<InstrProfReaderNullRemapper>(*IndexPtr);
901 Index = std::move(IndexPtr);
903 return success();
906 InstrProfSymtab &IndexedInstrProfReader::getSymtab() {
907 if (Symtab.get())
908 return *Symtab.get();
910 std::unique_ptr<InstrProfSymtab> NewSymtab = std::make_unique<InstrProfSymtab>();
911 if (Error E = Index->populateSymtab(*NewSymtab.get())) {
912 consumeError(error(InstrProfError::take(std::move(E))));
915 Symtab = std::move(NewSymtab);
916 return *Symtab.get();
919 Expected<InstrProfRecord>
920 IndexedInstrProfReader::getInstrProfRecord(StringRef FuncName,
921 uint64_t FuncHash) {
922 ArrayRef<NamedInstrProfRecord> Data;
923 Error Err = Remapper->getRecords(FuncName, Data);
924 if (Err)
925 return std::move(Err);
926 // Found it. Look for counters with the right hash.
927 for (unsigned I = 0, E = Data.size(); I < E; ++I) {
928 // Check for a match and fill the vector if there is one.
929 if (Data[I].Hash == FuncHash) {
930 return std::move(Data[I]);
933 return error(instrprof_error::hash_mismatch);
936 Error IndexedInstrProfReader::getFunctionCounts(StringRef FuncName,
937 uint64_t FuncHash,
938 std::vector<uint64_t> &Counts) {
939 Expected<InstrProfRecord> Record = getInstrProfRecord(FuncName, FuncHash);
940 if (Error E = Record.takeError())
941 return error(std::move(E));
943 Counts = Record.get().Counts;
944 return success();
947 Error IndexedInstrProfReader::readNextRecord(NamedInstrProfRecord &Record) {
948 ArrayRef<NamedInstrProfRecord> Data;
950 Error E = Index->getRecords(Data);
951 if (E)
952 return error(std::move(E));
954 Record = Data[RecordIndex++];
955 if (RecordIndex >= Data.size()) {
956 Index->advanceToNextKey();
957 RecordIndex = 0;
959 return success();
962 void InstrProfReader::accumulateCounts(CountSumOrPercent &Sum, bool IsCS) {
963 uint64_t NumFuncs = 0;
964 for (const auto &Func : *this) {
965 if (isIRLevelProfile()) {
966 bool FuncIsCS = NamedInstrProfRecord::hasCSFlagInHash(Func.Hash);
967 if (FuncIsCS != IsCS)
968 continue;
970 Func.accumulateCounts(Sum);
971 ++NumFuncs;
973 Sum.NumEntries = NumFuncs;