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. |
11 | // |
12 | //===----------------------------------------------------------------------===// |
13 | |
14 | #include "llvm/ProfileData/InstrProfReader.h" |
15 | #include "llvm/ADT/ArrayRef.h" |
16 | #include "llvm/ADT/DenseMap.h" |
17 | #include "llvm/ADT/StringExtras.h" |
18 | #include "llvm/ADT/StringRef.h" |
19 | #include "llvm/IR/ProfileSummary.h" |
20 | #include "llvm/ProfileData/InstrProf.h" |
21 | #include "llvm/ProfileData/MemProf.h" |
22 | #include "llvm/ProfileData/ProfileCommon.h" |
23 | #include "llvm/ProfileData/SymbolRemappingReader.h" |
24 | #include "llvm/Support/Endian.h" |
25 | #include "llvm/Support/Error.h" |
26 | #include "llvm/Support/ErrorOr.h" |
27 | #include "llvm/Support/FormatVariadic.h" |
28 | #include "llvm/Support/MemoryBuffer.h" |
29 | #include "llvm/Support/SwapByteOrder.h" |
30 | #include "llvm/Support/VirtualFileSystem.h" |
31 | #include <algorithm> |
32 | #include <cstddef> |
33 | #include <cstdint> |
34 | #include <limits> |
35 | #include <memory> |
36 | #include <optional> |
37 | #include <system_error> |
38 | #include <utility> |
39 | #include <vector> |
40 | |
41 | using namespace llvm; |
42 | |
43 | // Extracts the variant information from the top 32 bits in the version and |
44 | // returns an enum specifying the variants present. |
45 | static InstrProfKind getProfileKindFromVersion(uint64_t Version) { |
46 | InstrProfKind ProfileKind = InstrProfKind::Unknown; |
47 | if (Version & VARIANT_MASK_IR_PROF) { |
48 | ProfileKind |= InstrProfKind::IRInstrumentation; |
49 | } |
50 | if (Version & VARIANT_MASK_CSIR_PROF) { |
51 | ProfileKind |= InstrProfKind::ContextSensitive; |
52 | } |
53 | if (Version & VARIANT_MASK_INSTR_ENTRY) { |
54 | ProfileKind |= InstrProfKind::FunctionEntryInstrumentation; |
55 | } |
56 | if (Version & VARIANT_MASK_BYTE_COVERAGE) { |
57 | ProfileKind |= InstrProfKind::SingleByteCoverage; |
58 | } |
59 | if (Version & VARIANT_MASK_FUNCTION_ENTRY_ONLY) { |
60 | ProfileKind |= InstrProfKind::FunctionEntryOnly; |
61 | } |
62 | if (Version & VARIANT_MASK_MEMPROF) { |
63 | ProfileKind |= InstrProfKind::MemProf; |
64 | } |
65 | if (Version & VARIANT_MASK_TEMPORAL_PROF) { |
66 | ProfileKind |= InstrProfKind::TemporalProfile; |
67 | } |
68 | return ProfileKind; |
69 | } |
70 | |
71 | static Expected<std::unique_ptr<MemoryBuffer>> |
72 | setupMemoryBuffer(const Twine &Filename, vfs::FileSystem &FS) { |
73 | auto BufferOrErr = Filename.str() == "-" ? MemoryBuffer::getSTDIN() |
74 | : FS.getBufferForFile(Name: Filename); |
75 | if (std::error_code EC = BufferOrErr.getError()) |
76 | return errorCodeToError(EC); |
77 | return std::move(BufferOrErr.get()); |
78 | } |
79 | |
80 | static Error initializeReader(InstrProfReader &Reader) { |
81 | return Reader.readHeader(); |
82 | } |
83 | |
84 | /// Read a list of binary ids from a profile that consist of |
85 | /// a. uint64_t binary id length |
86 | /// b. uint8_t binary id data |
87 | /// c. uint8_t padding (if necessary) |
88 | /// This function is shared between raw and indexed profiles. |
89 | /// Raw profiles are in host-endian format, and indexed profiles are in |
90 | /// little-endian format. So, this function takes an argument indicating the |
91 | /// associated endian format to read the binary ids correctly. |
92 | static Error |
93 | readBinaryIdsInternal(const MemoryBuffer &DataBuffer, |
94 | ArrayRef<uint8_t> BinaryIdsBuffer, |
95 | std::vector<llvm::object::BuildID> &BinaryIds, |
96 | const llvm::endianness Endian) { |
97 | using namespace support; |
98 | |
99 | const uint64_t BinaryIdsSize = BinaryIdsBuffer.size(); |
100 | const uint8_t *BinaryIdsStart = BinaryIdsBuffer.data(); |
101 | |
102 | if (BinaryIdsSize == 0) |
103 | return Error::success(); |
104 | |
105 | const uint8_t *BI = BinaryIdsStart; |
106 | const uint8_t *BIEnd = BinaryIdsStart + BinaryIdsSize; |
107 | const uint8_t *End = |
108 | reinterpret_cast<const uint8_t *>(DataBuffer.getBufferEnd()); |
109 | |
110 | while (BI < BIEnd) { |
111 | size_t Remaining = BIEnd - BI; |
112 | // There should be enough left to read the binary id length. |
113 | if (Remaining < sizeof(uint64_t)) |
114 | return make_error<InstrProfError>( |
115 | Args: instrprof_error::malformed, |
116 | Args: "not enough data to read binary id length" ); |
117 | |
118 | uint64_t BILen = endian::readNext<uint64_t>(memory&: BI, endian: Endian); |
119 | if (BILen == 0) |
120 | return make_error<InstrProfError>(Args: instrprof_error::malformed, |
121 | Args: "binary id length is 0" ); |
122 | |
123 | Remaining = BIEnd - BI; |
124 | // There should be enough left to read the binary id data. |
125 | if (Remaining < alignToPowerOf2(Value: BILen, Align: sizeof(uint64_t))) |
126 | return make_error<InstrProfError>( |
127 | Args: instrprof_error::malformed, Args: "not enough data to read binary id data" ); |
128 | |
129 | // Add binary id to the binary ids list. |
130 | BinaryIds.push_back(x: object::BuildID(BI, BI + BILen)); |
131 | |
132 | // Increment by binary id data length, which aligned to the size of uint64. |
133 | BI += alignToPowerOf2(Value: BILen, Align: sizeof(uint64_t)); |
134 | if (BI > End) |
135 | return make_error<InstrProfError>( |
136 | Args: instrprof_error::malformed, |
137 | Args: "binary id section is greater than buffer size" ); |
138 | } |
139 | |
140 | return Error::success(); |
141 | } |
142 | |
143 | static void printBinaryIdsInternal(raw_ostream &OS, |
144 | ArrayRef<llvm::object::BuildID> BinaryIds) { |
145 | OS << "Binary IDs: \n" ; |
146 | for (const auto &BI : BinaryIds) { |
147 | for (auto I : BI) |
148 | OS << format(Fmt: "%02x" , Vals: I); |
149 | OS << "\n" ; |
150 | } |
151 | } |
152 | |
153 | Expected<std::unique_ptr<InstrProfReader>> |
154 | InstrProfReader::create(const Twine &Path, vfs::FileSystem &FS, |
155 | const InstrProfCorrelator *Correlator, |
156 | std::function<void(Error)> Warn) { |
157 | // Set up the buffer to read. |
158 | auto BufferOrError = setupMemoryBuffer(Filename: Path, FS); |
159 | if (Error E = BufferOrError.takeError()) |
160 | return std::move(E); |
161 | return InstrProfReader::create(Buffer: std::move(BufferOrError.get()), Correlator, |
162 | Warn); |
163 | } |
164 | |
165 | Expected<std::unique_ptr<InstrProfReader>> |
166 | InstrProfReader::create(std::unique_ptr<MemoryBuffer> Buffer, |
167 | const InstrProfCorrelator *Correlator, |
168 | std::function<void(Error)> Warn) { |
169 | if (Buffer->getBufferSize() == 0) |
170 | return make_error<InstrProfError>(Args: instrprof_error::empty_raw_profile); |
171 | |
172 | std::unique_ptr<InstrProfReader> Result; |
173 | // Create the reader. |
174 | if (IndexedInstrProfReader::hasFormat(DataBuffer: *Buffer)) |
175 | Result.reset(p: new IndexedInstrProfReader(std::move(Buffer))); |
176 | else if (RawInstrProfReader64::hasFormat(DataBuffer: *Buffer)) |
177 | Result.reset(p: new RawInstrProfReader64(std::move(Buffer), Correlator, Warn)); |
178 | else if (RawInstrProfReader32::hasFormat(DataBuffer: *Buffer)) |
179 | Result.reset(p: new RawInstrProfReader32(std::move(Buffer), Correlator, Warn)); |
180 | else if (TextInstrProfReader::hasFormat(Buffer: *Buffer)) |
181 | Result.reset(p: new TextInstrProfReader(std::move(Buffer))); |
182 | else |
183 | return make_error<InstrProfError>(Args: instrprof_error::unrecognized_format); |
184 | |
185 | // Initialize the reader and return the result. |
186 | if (Error E = initializeReader(Reader&: *Result)) |
187 | return std::move(E); |
188 | |
189 | return std::move(Result); |
190 | } |
191 | |
192 | Expected<std::unique_ptr<IndexedInstrProfReader>> |
193 | IndexedInstrProfReader::create(const Twine &Path, vfs::FileSystem &FS, |
194 | const Twine &RemappingPath) { |
195 | // Set up the buffer to read. |
196 | auto BufferOrError = setupMemoryBuffer(Filename: Path, FS); |
197 | if (Error E = BufferOrError.takeError()) |
198 | return std::move(E); |
199 | |
200 | // Set up the remapping buffer if requested. |
201 | std::unique_ptr<MemoryBuffer> RemappingBuffer; |
202 | std::string RemappingPathStr = RemappingPath.str(); |
203 | if (!RemappingPathStr.empty()) { |
204 | auto RemappingBufferOrError = setupMemoryBuffer(Filename: RemappingPathStr, FS); |
205 | if (Error E = RemappingBufferOrError.takeError()) |
206 | return std::move(E); |
207 | RemappingBuffer = std::move(RemappingBufferOrError.get()); |
208 | } |
209 | |
210 | return IndexedInstrProfReader::create(Buffer: std::move(BufferOrError.get()), |
211 | RemappingBuffer: std::move(RemappingBuffer)); |
212 | } |
213 | |
214 | Expected<std::unique_ptr<IndexedInstrProfReader>> |
215 | IndexedInstrProfReader::create(std::unique_ptr<MemoryBuffer> Buffer, |
216 | std::unique_ptr<MemoryBuffer> RemappingBuffer) { |
217 | // Create the reader. |
218 | if (!IndexedInstrProfReader::hasFormat(DataBuffer: *Buffer)) |
219 | return make_error<InstrProfError>(Args: instrprof_error::bad_magic); |
220 | auto Result = std::make_unique<IndexedInstrProfReader>( |
221 | args: std::move(Buffer), args: std::move(RemappingBuffer)); |
222 | |
223 | // Initialize the reader and return the result. |
224 | if (Error E = initializeReader(Reader&: *Result)) |
225 | return std::move(E); |
226 | |
227 | return std::move(Result); |
228 | } |
229 | |
230 | bool TextInstrProfReader::hasFormat(const MemoryBuffer &Buffer) { |
231 | // Verify that this really looks like plain ASCII text by checking a |
232 | // 'reasonable' number of characters (up to profile magic size). |
233 | size_t count = std::min(a: Buffer.getBufferSize(), b: sizeof(uint64_t)); |
234 | StringRef buffer = Buffer.getBufferStart(); |
235 | return count == 0 || |
236 | std::all_of(first: buffer.begin(), last: buffer.begin() + count, |
237 | pred: [](char c) { return isPrint(C: c) || isSpace(C: c); }); |
238 | } |
239 | |
240 | // Read the profile variant flag from the header: ":FE" means this is a FE |
241 | // generated profile. ":IR" means this is an IR level profile. Other strings |
242 | // with a leading ':' will be reported an error format. |
243 | Error TextInstrProfReader::() { |
244 | Symtab.reset(p: new InstrProfSymtab()); |
245 | |
246 | while (Line->starts_with(Prefix: ":" )) { |
247 | StringRef Str = Line->substr(Start: 1); |
248 | if (Str.equals_insensitive(RHS: "ir" )) |
249 | ProfileKind |= InstrProfKind::IRInstrumentation; |
250 | else if (Str.equals_insensitive(RHS: "fe" )) |
251 | ProfileKind |= InstrProfKind::FrontendInstrumentation; |
252 | else if (Str.equals_insensitive(RHS: "csir" )) { |
253 | ProfileKind |= InstrProfKind::IRInstrumentation; |
254 | ProfileKind |= InstrProfKind::ContextSensitive; |
255 | } else if (Str.equals_insensitive(RHS: "entry_first" )) |
256 | ProfileKind |= InstrProfKind::FunctionEntryInstrumentation; |
257 | else if (Str.equals_insensitive(RHS: "not_entry_first" )) |
258 | ProfileKind &= ~InstrProfKind::FunctionEntryInstrumentation; |
259 | else if (Str.equals_insensitive(RHS: "single_byte_coverage" )) |
260 | ProfileKind |= InstrProfKind::SingleByteCoverage; |
261 | else if (Str.equals_insensitive(RHS: "temporal_prof_traces" )) { |
262 | ProfileKind |= InstrProfKind::TemporalProfile; |
263 | if (auto Err = readTemporalProfTraceData()) |
264 | return error(E: std::move(Err)); |
265 | } else |
266 | return error(Err: instrprof_error::bad_header); |
267 | ++Line; |
268 | } |
269 | return success(); |
270 | } |
271 | |
272 | /// Temporal profile trace data is stored in the header immediately after |
273 | /// ":temporal_prof_traces". The first integer is the number of traces, the |
274 | /// second integer is the stream size, then the following lines are the actual |
275 | /// traces which consist of a weight and a comma separated list of function |
276 | /// names. |
277 | Error TextInstrProfReader::readTemporalProfTraceData() { |
278 | if ((++Line).is_at_end()) |
279 | return error(Err: instrprof_error::eof); |
280 | |
281 | uint32_t NumTraces; |
282 | if (Line->getAsInteger(Radix: 0, Result&: NumTraces)) |
283 | return error(Err: instrprof_error::malformed); |
284 | |
285 | if ((++Line).is_at_end()) |
286 | return error(Err: instrprof_error::eof); |
287 | |
288 | if (Line->getAsInteger(Radix: 0, Result&: TemporalProfTraceStreamSize)) |
289 | return error(Err: instrprof_error::malformed); |
290 | |
291 | for (uint32_t i = 0; i < NumTraces; i++) { |
292 | if ((++Line).is_at_end()) |
293 | return error(Err: instrprof_error::eof); |
294 | |
295 | TemporalProfTraceTy Trace; |
296 | if (Line->getAsInteger(Radix: 0, Result&: Trace.Weight)) |
297 | return error(Err: instrprof_error::malformed); |
298 | |
299 | if ((++Line).is_at_end()) |
300 | return error(Err: instrprof_error::eof); |
301 | |
302 | SmallVector<StringRef> FuncNames; |
303 | Line->split(A&: FuncNames, Separator: "," , /*MaxSplit=*/-1, /*KeepEmpty=*/false); |
304 | for (auto &FuncName : FuncNames) |
305 | Trace.FunctionNameRefs.push_back( |
306 | x: IndexedInstrProf::ComputeHash(K: FuncName.trim())); |
307 | TemporalProfTraces.push_back(Elt: std::move(Trace)); |
308 | } |
309 | return success(); |
310 | } |
311 | |
312 | Error |
313 | TextInstrProfReader::readValueProfileData(InstrProfRecord &Record) { |
314 | |
315 | #define CHECK_LINE_END(Line) \ |
316 | if (Line.is_at_end()) \ |
317 | return error(instrprof_error::truncated); |
318 | #define READ_NUM(Str, Dst) \ |
319 | if ((Str).getAsInteger(10, (Dst))) \ |
320 | return error(instrprof_error::malformed); |
321 | #define VP_READ_ADVANCE(Val) \ |
322 | CHECK_LINE_END(Line); \ |
323 | uint32_t Val; \ |
324 | READ_NUM((*Line), (Val)); \ |
325 | Line++; |
326 | |
327 | if (Line.is_at_end()) |
328 | return success(); |
329 | |
330 | uint32_t NumValueKinds; |
331 | if (Line->getAsInteger(Radix: 10, Result&: NumValueKinds)) { |
332 | // No value profile data |
333 | return success(); |
334 | } |
335 | if (NumValueKinds == 0 || NumValueKinds > IPVK_Last + 1) |
336 | return error(Err: instrprof_error::malformed, |
337 | ErrMsg: "number of value kinds is invalid" ); |
338 | Line++; |
339 | |
340 | for (uint32_t VK = 0; VK < NumValueKinds; VK++) { |
341 | VP_READ_ADVANCE(ValueKind); |
342 | if (ValueKind > IPVK_Last) |
343 | return error(Err: instrprof_error::malformed, ErrMsg: "value kind is invalid" ); |
344 | ; |
345 | VP_READ_ADVANCE(NumValueSites); |
346 | if (!NumValueSites) |
347 | continue; |
348 | |
349 | Record.reserveSites(ValueKind: VK, NumValueSites); |
350 | for (uint32_t S = 0; S < NumValueSites; S++) { |
351 | VP_READ_ADVANCE(NumValueData); |
352 | |
353 | std::vector<InstrProfValueData> CurrentValues; |
354 | for (uint32_t V = 0; V < NumValueData; V++) { |
355 | CHECK_LINE_END(Line); |
356 | std::pair<StringRef, StringRef> VD = Line->rsplit(Separator: ':'); |
357 | uint64_t TakenCount, Value; |
358 | if (ValueKind == IPVK_IndirectCallTarget) { |
359 | if (InstrProfSymtab::isExternalSymbol(Symbol: VD.first)) { |
360 | Value = 0; |
361 | } else { |
362 | if (Error E = Symtab->addFuncName(FuncName: VD.first)) |
363 | return E; |
364 | Value = IndexedInstrProf::ComputeHash(K: VD.first); |
365 | } |
366 | } else if (ValueKind == IPVK_VTableTarget) { |
367 | if (InstrProfSymtab::isExternalSymbol(Symbol: VD.first)) |
368 | Value = 0; |
369 | else { |
370 | if (Error E = Symtab->addVTableName(VTableName: VD.first)) |
371 | return E; |
372 | Value = IndexedInstrProf::ComputeHash(K: VD.first); |
373 | } |
374 | } else { |
375 | READ_NUM(VD.first, Value); |
376 | } |
377 | READ_NUM(VD.second, TakenCount); |
378 | CurrentValues.push_back(x: {.Value: Value, .Count: TakenCount}); |
379 | Line++; |
380 | } |
381 | assert(CurrentValues.size() == NumValueData); |
382 | Record.addValueData(ValueKind, Site: S, VData: CurrentValues, SymTab: nullptr); |
383 | } |
384 | } |
385 | return success(); |
386 | |
387 | #undef CHECK_LINE_END |
388 | #undef READ_NUM |
389 | #undef VP_READ_ADVANCE |
390 | } |
391 | |
392 | Error TextInstrProfReader::readNextRecord(NamedInstrProfRecord &Record) { |
393 | // Skip empty lines and comments. |
394 | while (!Line.is_at_end() && (Line->empty() || Line->starts_with(Prefix: "#" ))) |
395 | ++Line; |
396 | // If we hit EOF while looking for a name, we're done. |
397 | if (Line.is_at_end()) { |
398 | return error(Err: instrprof_error::eof); |
399 | } |
400 | |
401 | // Read the function name. |
402 | Record.Name = *Line++; |
403 | if (Error E = Symtab->addFuncName(FuncName: Record.Name)) |
404 | return error(E: std::move(E)); |
405 | |
406 | // Read the function hash. |
407 | if (Line.is_at_end()) |
408 | return error(Err: instrprof_error::truncated); |
409 | if ((Line++)->getAsInteger(Radix: 0, Result&: Record.Hash)) |
410 | return error(Err: instrprof_error::malformed, |
411 | ErrMsg: "function hash is not a valid integer" ); |
412 | |
413 | // Read the number of counters. |
414 | uint64_t NumCounters; |
415 | if (Line.is_at_end()) |
416 | return error(Err: instrprof_error::truncated); |
417 | if ((Line++)->getAsInteger(Radix: 10, Result&: NumCounters)) |
418 | return error(Err: instrprof_error::malformed, |
419 | ErrMsg: "number of counters is not a valid integer" ); |
420 | if (NumCounters == 0) |
421 | return error(Err: instrprof_error::malformed, ErrMsg: "number of counters is zero" ); |
422 | |
423 | // Read each counter and fill our internal storage with the values. |
424 | Record.Clear(); |
425 | Record.Counts.reserve(n: NumCounters); |
426 | for (uint64_t I = 0; I < NumCounters; ++I) { |
427 | if (Line.is_at_end()) |
428 | return error(Err: instrprof_error::truncated); |
429 | uint64_t Count; |
430 | if ((Line++)->getAsInteger(Radix: 10, Result&: Count)) |
431 | return error(Err: instrprof_error::malformed, ErrMsg: "count is invalid" ); |
432 | Record.Counts.push_back(x: Count); |
433 | } |
434 | |
435 | // Bitmap byte information is indicated with special character. |
436 | if (Line->starts_with(Prefix: "$" )) { |
437 | Record.BitmapBytes.clear(); |
438 | // Read the number of bitmap bytes. |
439 | uint64_t NumBitmapBytes; |
440 | if ((Line++)->drop_front(N: 1).trim().getAsInteger(Radix: 0, Result&: NumBitmapBytes)) |
441 | return error(Err: instrprof_error::malformed, |
442 | ErrMsg: "number of bitmap bytes is not a valid integer" ); |
443 | if (NumBitmapBytes != 0) { |
444 | // Read each bitmap and fill our internal storage with the values. |
445 | Record.BitmapBytes.reserve(n: NumBitmapBytes); |
446 | for (uint8_t I = 0; I < NumBitmapBytes; ++I) { |
447 | if (Line.is_at_end()) |
448 | return error(Err: instrprof_error::truncated); |
449 | uint8_t BitmapByte; |
450 | if ((Line++)->getAsInteger(Radix: 0, Result&: BitmapByte)) |
451 | return error(Err: instrprof_error::malformed, |
452 | ErrMsg: "bitmap byte is not a valid integer" ); |
453 | Record.BitmapBytes.push_back(x: BitmapByte); |
454 | } |
455 | } |
456 | } |
457 | |
458 | // Check if value profile data exists and read it if so. |
459 | if (Error E = readValueProfileData(Record)) |
460 | return error(E: std::move(E)); |
461 | |
462 | return success(); |
463 | } |
464 | |
465 | template <class IntPtrT> |
466 | InstrProfKind RawInstrProfReader<IntPtrT>::getProfileKind() const { |
467 | return getProfileKindFromVersion(Version); |
468 | } |
469 | |
470 | template <class IntPtrT> |
471 | SmallVector<TemporalProfTraceTy> & |
472 | RawInstrProfReader<IntPtrT>::getTemporalProfTraces( |
473 | std::optional<uint64_t> Weight) { |
474 | if (TemporalProfTimestamps.empty()) { |
475 | assert(TemporalProfTraces.empty()); |
476 | return TemporalProfTraces; |
477 | } |
478 | // Sort functions by their timestamps to build the trace. |
479 | std::sort(first: TemporalProfTimestamps.begin(), last: TemporalProfTimestamps.end()); |
480 | TemporalProfTraceTy Trace; |
481 | if (Weight) |
482 | Trace.Weight = *Weight; |
483 | for (auto &[TimestampValue, NameRef] : TemporalProfTimestamps) |
484 | Trace.FunctionNameRefs.push_back(x: NameRef); |
485 | TemporalProfTraces = {std::move(Trace)}; |
486 | return TemporalProfTraces; |
487 | } |
488 | |
489 | template <class IntPtrT> |
490 | bool RawInstrProfReader<IntPtrT>::hasFormat(const MemoryBuffer &DataBuffer) { |
491 | if (DataBuffer.getBufferSize() < sizeof(uint64_t)) |
492 | return false; |
493 | uint64_t Magic = |
494 | *reinterpret_cast<const uint64_t *>(DataBuffer.getBufferStart()); |
495 | return RawInstrProf::getMagic<IntPtrT>() == Magic || |
496 | llvm::byteswap(RawInstrProf::getMagic<IntPtrT>()) == Magic; |
497 | } |
498 | |
499 | template <class IntPtrT> |
500 | Error RawInstrProfReader<IntPtrT>::() { |
501 | if (!hasFormat(DataBuffer: *DataBuffer)) |
502 | return error(instrprof_error::bad_magic); |
503 | if (DataBuffer->getBufferSize() < sizeof(RawInstrProf::Header)) |
504 | return error(instrprof_error::bad_header); |
505 | auto * = reinterpret_cast<const RawInstrProf::Header *>( |
506 | DataBuffer->getBufferStart()); |
507 | ShouldSwapBytes = Header->Magic != RawInstrProf::getMagic<IntPtrT>(); |
508 | return readHeader(*Header); |
509 | } |
510 | |
511 | template <class IntPtrT> |
512 | Error RawInstrProfReader<IntPtrT>::(const char *CurrentPos) { |
513 | const char *End = DataBuffer->getBufferEnd(); |
514 | // Skip zero padding between profiles. |
515 | while (CurrentPos != End && *CurrentPos == 0) |
516 | ++CurrentPos; |
517 | // If there's nothing left, we're done. |
518 | if (CurrentPos == End) |
519 | return make_error<InstrProfError>(Args: instrprof_error::eof); |
520 | // If there isn't enough space for another header, this is probably just |
521 | // garbage at the end of the file. |
522 | if (CurrentPos + sizeof(RawInstrProf::Header) > End) |
523 | return make_error<InstrProfError>(Args: instrprof_error::malformed, |
524 | Args: "not enough space for another header" ); |
525 | // The writer ensures each profile is padded to start at an aligned address. |
526 | if (reinterpret_cast<size_t>(CurrentPos) % alignof(uint64_t)) |
527 | return make_error<InstrProfError>(Args: instrprof_error::malformed, |
528 | Args: "insufficient padding" ); |
529 | // The magic should have the same byte order as in the previous header. |
530 | uint64_t Magic = *reinterpret_cast<const uint64_t *>(CurrentPos); |
531 | if (Magic != swap(RawInstrProf::getMagic<IntPtrT>())) |
532 | return make_error<InstrProfError>(Args: instrprof_error::bad_magic); |
533 | |
534 | // There's another profile to read, so we need to process the header. |
535 | auto * = reinterpret_cast<const RawInstrProf::Header *>(CurrentPos); |
536 | return readHeader(*Header); |
537 | } |
538 | |
539 | template <class IntPtrT> |
540 | Error RawInstrProfReader<IntPtrT>::createSymtab(InstrProfSymtab &Symtab) { |
541 | if (Error E = Symtab.create(FuncNameStrings: StringRef(NamesStart, NamesEnd - NamesStart), |
542 | VTableNameStrings: StringRef(VNamesStart, VNamesEnd - VNamesStart))) |
543 | return error(std::move(E)); |
544 | for (const RawInstrProf::ProfileData<IntPtrT> *I = Data; I != DataEnd; ++I) { |
545 | const IntPtrT FPtr = swap(I->FunctionPointer); |
546 | if (!FPtr) |
547 | continue; |
548 | Symtab.mapAddress(Addr: FPtr, MD5Val: swap(I->NameRef)); |
549 | } |
550 | |
551 | if (VTableBegin != nullptr && VTableEnd != nullptr) { |
552 | for (const RawInstrProf::VTableProfileData<IntPtrT> *I = VTableBegin; |
553 | I != VTableEnd; ++I) { |
554 | const IntPtrT VPtr = swap(I->VTablePointer); |
555 | if (!VPtr) |
556 | continue; |
557 | // Map both begin and end address to the name hash, since the instrumented |
558 | // address could be somewhere in the middle. |
559 | // VPtr is of type uint32_t or uint64_t so 'VPtr + I->VTableSize' marks |
560 | // the end of vtable address. |
561 | Symtab.mapVTableAddress(StartAddr: VPtr, EndAddr: VPtr + swap(I->VTableSize), |
562 | MD5Val: swap(I->VTableNameHash)); |
563 | } |
564 | } |
565 | return success(); |
566 | } |
567 | |
568 | template <class IntPtrT> |
569 | Error RawInstrProfReader<IntPtrT>::( |
570 | const RawInstrProf::Header &) { |
571 | Version = swap(Header.Version); |
572 | if (GET_VERSION(Version) != RawInstrProf::Version) |
573 | return error(instrprof_error::raw_profile_version_mismatch, |
574 | ("Profile uses raw profile format version = " + |
575 | Twine(GET_VERSION(Version)) + |
576 | "; expected version = " + Twine(RawInstrProf::Version) + |
577 | "\nPLEASE update this tool to version in the raw profile, or " |
578 | "regenerate raw profile with expected version." ) |
579 | .str()); |
580 | |
581 | uint64_t BinaryIdSize = swap(Header.BinaryIdsSize); |
582 | // Binary id start just after the header if exists. |
583 | const uint8_t *BinaryIdStart = |
584 | reinterpret_cast<const uint8_t *>(&Header) + sizeof(RawInstrProf::Header); |
585 | const uint8_t *BinaryIdEnd = BinaryIdStart + BinaryIdSize; |
586 | const uint8_t *BufferEnd = (const uint8_t *)DataBuffer->getBufferEnd(); |
587 | if (BinaryIdSize % sizeof(uint64_t) || BinaryIdEnd > BufferEnd) |
588 | return error(instrprof_error::bad_header); |
589 | ArrayRef<uint8_t> BinaryIdsBuffer(BinaryIdStart, BinaryIdSize); |
590 | if (!BinaryIdsBuffer.empty()) { |
591 | if (Error Err = readBinaryIdsInternal(*DataBuffer, BinaryIdsBuffer, |
592 | BinaryIds, getDataEndianness())) |
593 | return Err; |
594 | } |
595 | |
596 | CountersDelta = swap(Header.CountersDelta); |
597 | BitmapDelta = swap(Header.BitmapDelta); |
598 | NamesDelta = swap(Header.NamesDelta); |
599 | auto NumData = swap(Header.NumData); |
600 | auto PaddingBytesBeforeCounters = swap(Header.PaddingBytesBeforeCounters); |
601 | auto = swap(Header.NumCounters) * getCounterTypeSize(); |
602 | auto PaddingBytesAfterCounters = swap(Header.PaddingBytesAfterCounters); |
603 | auto NumBitmapBytes = swap(Header.NumBitmapBytes); |
604 | auto PaddingBytesAfterBitmapBytes = swap(Header.PaddingBytesAfterBitmapBytes); |
605 | auto NamesSize = swap(Header.NamesSize); |
606 | auto VTableNameSize = swap(Header.VNamesSize); |
607 | auto NumVTables = swap(Header.NumVTables); |
608 | ValueKindLast = swap(Header.ValueKindLast); |
609 | |
610 | auto DataSize = NumData * sizeof(RawInstrProf::ProfileData<IntPtrT>); |
611 | auto PaddingBytesAfterNames = getNumPaddingBytes(SizeInBytes: NamesSize); |
612 | auto PaddingBytesAfterVTableNames = getNumPaddingBytes(SizeInBytes: VTableNameSize); |
613 | |
614 | auto VTableSectionSize = |
615 | NumVTables * sizeof(RawInstrProf::VTableProfileData<IntPtrT>); |
616 | auto PaddingBytesAfterVTableProfData = getNumPaddingBytes(SizeInBytes: VTableSectionSize); |
617 | |
618 | // Profile data starts after profile header and binary ids if exist. |
619 | ptrdiff_t DataOffset = sizeof(RawInstrProf::Header) + BinaryIdSize; |
620 | ptrdiff_t CountersOffset = DataOffset + DataSize + PaddingBytesBeforeCounters; |
621 | ptrdiff_t BitmapOffset = |
622 | CountersOffset + CountersSize + PaddingBytesAfterCounters; |
623 | ptrdiff_t NamesOffset = |
624 | BitmapOffset + NumBitmapBytes + PaddingBytesAfterBitmapBytes; |
625 | ptrdiff_t VTableProfDataOffset = |
626 | NamesOffset + NamesSize + PaddingBytesAfterNames; |
627 | ptrdiff_t VTableNameOffset = VTableProfDataOffset + VTableSectionSize + |
628 | PaddingBytesAfterVTableProfData; |
629 | ptrdiff_t ValueDataOffset = |
630 | VTableNameOffset + VTableNameSize + PaddingBytesAfterVTableNames; |
631 | |
632 | auto *Start = reinterpret_cast<const char *>(&Header); |
633 | if (Start + ValueDataOffset > DataBuffer->getBufferEnd()) |
634 | return error(instrprof_error::bad_header); |
635 | |
636 | if (Correlator) { |
637 | // These sizes in the raw file are zero because we constructed them in the |
638 | // Correlator. |
639 | if (!(DataSize == 0 && NamesSize == 0 && CountersDelta == 0 && |
640 | NamesDelta == 0)) |
641 | return error(instrprof_error::unexpected_correlation_info); |
642 | Data = Correlator->getDataPointer(); |
643 | DataEnd = Data + Correlator->getDataSize(); |
644 | NamesStart = Correlator->getNamesPointer(); |
645 | NamesEnd = NamesStart + Correlator->getNamesSize(); |
646 | } else { |
647 | Data = reinterpret_cast<const RawInstrProf::ProfileData<IntPtrT> *>( |
648 | Start + DataOffset); |
649 | DataEnd = Data + NumData; |
650 | VTableBegin = |
651 | reinterpret_cast<const RawInstrProf::VTableProfileData<IntPtrT> *>( |
652 | Start + VTableProfDataOffset); |
653 | VTableEnd = VTableBegin + NumVTables; |
654 | NamesStart = Start + NamesOffset; |
655 | NamesEnd = NamesStart + NamesSize; |
656 | VNamesStart = Start + VTableNameOffset; |
657 | VNamesEnd = VNamesStart + VTableNameSize; |
658 | } |
659 | |
660 | CountersStart = Start + CountersOffset; |
661 | CountersEnd = CountersStart + CountersSize; |
662 | BitmapStart = Start + BitmapOffset; |
663 | BitmapEnd = BitmapStart + NumBitmapBytes; |
664 | ValueDataStart = reinterpret_cast<const uint8_t *>(Start + ValueDataOffset); |
665 | |
666 | std::unique_ptr<InstrProfSymtab> NewSymtab = std::make_unique<InstrProfSymtab>(); |
667 | if (Error E = createSymtab(Symtab&: *NewSymtab)) |
668 | return E; |
669 | |
670 | Symtab = std::move(NewSymtab); |
671 | return success(); |
672 | } |
673 | |
674 | template <class IntPtrT> |
675 | Error RawInstrProfReader<IntPtrT>::readName(NamedInstrProfRecord &Record) { |
676 | Record.Name = getName(NameRef: Data->NameRef); |
677 | return success(); |
678 | } |
679 | |
680 | template <class IntPtrT> |
681 | Error RawInstrProfReader<IntPtrT>::readFuncHash(NamedInstrProfRecord &Record) { |
682 | Record.Hash = swap(Data->FuncHash); |
683 | return success(); |
684 | } |
685 | |
686 | template <class IntPtrT> |
687 | Error RawInstrProfReader<IntPtrT>::readRawCounts( |
688 | InstrProfRecord &Record) { |
689 | uint32_t NumCounters = swap(Data->NumCounters); |
690 | if (NumCounters == 0) |
691 | return error(instrprof_error::malformed, "number of counters is zero" ); |
692 | |
693 | ptrdiff_t CounterBaseOffset = swap(Data->CounterPtr) - CountersDelta; |
694 | if (CounterBaseOffset < 0) |
695 | return error( |
696 | instrprof_error::malformed, |
697 | ("counter offset " + Twine(CounterBaseOffset) + " is negative" ).str()); |
698 | |
699 | if (CounterBaseOffset >= CountersEnd - CountersStart) |
700 | return error(instrprof_error::malformed, |
701 | ("counter offset " + Twine(CounterBaseOffset) + |
702 | " is greater than the maximum counter offset " + |
703 | Twine(CountersEnd - CountersStart - 1)) |
704 | .str()); |
705 | |
706 | uint64_t MaxNumCounters = |
707 | (CountersEnd - (CountersStart + CounterBaseOffset)) / |
708 | getCounterTypeSize(); |
709 | if (NumCounters > MaxNumCounters) |
710 | return error(instrprof_error::malformed, |
711 | ("number of counters " + Twine(NumCounters) + |
712 | " is greater than the maximum number of counters " + |
713 | Twine(MaxNumCounters)) |
714 | .str()); |
715 | |
716 | Record.Counts.clear(); |
717 | Record.Counts.reserve(n: NumCounters); |
718 | for (uint32_t I = 0; I < NumCounters; I++) { |
719 | const char *Ptr = |
720 | CountersStart + CounterBaseOffset + I * getCounterTypeSize(); |
721 | if (I == 0 && hasTemporalProfile()) { |
722 | uint64_t TimestampValue = swap(*reinterpret_cast<const uint64_t *>(Ptr)); |
723 | if (TimestampValue != 0 && |
724 | TimestampValue != std::numeric_limits<uint64_t>::max()) { |
725 | TemporalProfTimestamps.emplace_back(TimestampValue, |
726 | swap(Data->NameRef)); |
727 | TemporalProfTraceStreamSize = 1; |
728 | } |
729 | if (hasSingleByteCoverage()) { |
730 | // In coverage mode, getCounterTypeSize() returns 1 byte but our |
731 | // timestamp field has size uint64_t. Increment I so that the next |
732 | // iteration of this for loop points to the byte after the timestamp |
733 | // field, i.e., I += 8. |
734 | I += 7; |
735 | } |
736 | continue; |
737 | } |
738 | if (hasSingleByteCoverage()) { |
739 | // A value of zero signifies the block is covered. |
740 | Record.Counts.push_back(x: *Ptr == 0 ? 1 : 0); |
741 | } else { |
742 | uint64_t CounterValue = swap(*reinterpret_cast<const uint64_t *>(Ptr)); |
743 | if (CounterValue > MaxCounterValue && Warn) |
744 | Warn(make_error<InstrProfError>( |
745 | Args: instrprof_error::counter_value_too_large, Args: Twine(CounterValue))); |
746 | |
747 | Record.Counts.push_back(x: CounterValue); |
748 | } |
749 | } |
750 | |
751 | return success(); |
752 | } |
753 | |
754 | template <class IntPtrT> |
755 | Error RawInstrProfReader<IntPtrT>::readRawBitmapBytes(InstrProfRecord &Record) { |
756 | uint32_t NumBitmapBytes = swap(Data->NumBitmapBytes); |
757 | |
758 | Record.BitmapBytes.clear(); |
759 | Record.BitmapBytes.reserve(n: NumBitmapBytes); |
760 | |
761 | // It's possible MCDC is either not enabled or only used for some functions |
762 | // and not others. So if we record 0 bytes, just move on. |
763 | if (NumBitmapBytes == 0) |
764 | return success(); |
765 | |
766 | // BitmapDelta decreases as we advance to the next data record. |
767 | ptrdiff_t BitmapOffset = swap(Data->BitmapPtr) - BitmapDelta; |
768 | if (BitmapOffset < 0) |
769 | return error( |
770 | instrprof_error::malformed, |
771 | ("bitmap offset " + Twine(BitmapOffset) + " is negative" ).str()); |
772 | |
773 | if (BitmapOffset >= BitmapEnd - BitmapStart) |
774 | return error(instrprof_error::malformed, |
775 | ("bitmap offset " + Twine(BitmapOffset) + |
776 | " is greater than the maximum bitmap offset " + |
777 | Twine(BitmapEnd - BitmapStart - 1)) |
778 | .str()); |
779 | |
780 | uint64_t MaxNumBitmapBytes = |
781 | (BitmapEnd - (BitmapStart + BitmapOffset)) / sizeof(uint8_t); |
782 | if (NumBitmapBytes > MaxNumBitmapBytes) |
783 | return error(instrprof_error::malformed, |
784 | ("number of bitmap bytes " + Twine(NumBitmapBytes) + |
785 | " is greater than the maximum number of bitmap bytes " + |
786 | Twine(MaxNumBitmapBytes)) |
787 | .str()); |
788 | |
789 | for (uint32_t I = 0; I < NumBitmapBytes; I++) { |
790 | const char *Ptr = BitmapStart + BitmapOffset + I; |
791 | Record.BitmapBytes.push_back(swap(*Ptr)); |
792 | } |
793 | |
794 | return success(); |
795 | } |
796 | |
797 | template <class IntPtrT> |
798 | Error RawInstrProfReader<IntPtrT>::readValueProfilingData( |
799 | InstrProfRecord &Record) { |
800 | Record.clearValueData(); |
801 | CurValueDataSize = 0; |
802 | // Need to match the logic in value profile dumper code in compiler-rt: |
803 | uint32_t NumValueKinds = 0; |
804 | for (uint32_t I = 0; I < IPVK_Last + 1; I++) |
805 | NumValueKinds += (Data->NumValueSites[I] != 0); |
806 | |
807 | if (!NumValueKinds) |
808 | return success(); |
809 | |
810 | Expected<std::unique_ptr<ValueProfData>> VDataPtrOrErr = |
811 | ValueProfData::getValueProfData( |
812 | SrcBuffer: ValueDataStart, SrcBufferEnd: (const unsigned char *)DataBuffer->getBufferEnd(), |
813 | SrcDataEndianness: getDataEndianness()); |
814 | |
815 | if (Error E = VDataPtrOrErr.takeError()) |
816 | return E; |
817 | |
818 | // Note that besides deserialization, this also performs the conversion for |
819 | // indirect call targets. The function pointers from the raw profile are |
820 | // remapped into function name hashes. |
821 | VDataPtrOrErr.get()->deserializeTo(Record, SymTab: Symtab.get()); |
822 | CurValueDataSize = VDataPtrOrErr.get()->getSize(); |
823 | return success(); |
824 | } |
825 | |
826 | template <class IntPtrT> |
827 | Error RawInstrProfReader<IntPtrT>::readNextRecord(NamedInstrProfRecord &Record) { |
828 | // Keep reading profiles that consist of only headers and no profile data and |
829 | // counters. |
830 | while (atEnd()) |
831 | // At this point, ValueDataStart field points to the next header. |
832 | if (Error E = readNextHeader(CurrentPos: getNextHeaderPos())) |
833 | return error(std::move(E)); |
834 | |
835 | // Read name and set it in Record. |
836 | if (Error E = readName(Record)) |
837 | return error(std::move(E)); |
838 | |
839 | // Read FuncHash and set it in Record. |
840 | if (Error E = readFuncHash(Record)) |
841 | return error(std::move(E)); |
842 | |
843 | // Read raw counts and set Record. |
844 | if (Error E = readRawCounts(Record)) |
845 | return error(std::move(E)); |
846 | |
847 | // Read raw bitmap bytes and set Record. |
848 | if (Error E = readRawBitmapBytes(Record)) |
849 | return error(std::move(E)); |
850 | |
851 | // Read value data and set Record. |
852 | if (Error E = readValueProfilingData(Record)) |
853 | return error(std::move(E)); |
854 | |
855 | // Iterate. |
856 | advanceData(); |
857 | return success(); |
858 | } |
859 | |
860 | template <class IntPtrT> |
861 | Error RawInstrProfReader<IntPtrT>::readBinaryIds( |
862 | std::vector<llvm::object::BuildID> &BinaryIds) { |
863 | BinaryIds.insert(BinaryIds.begin(), this->BinaryIds.begin(), |
864 | this->BinaryIds.end()); |
865 | return Error::success(); |
866 | } |
867 | |
868 | template <class IntPtrT> |
869 | Error RawInstrProfReader<IntPtrT>::printBinaryIds(raw_ostream &OS) { |
870 | if (!BinaryIds.empty()) |
871 | printBinaryIdsInternal(OS, BinaryIds); |
872 | return Error::success(); |
873 | } |
874 | |
875 | namespace llvm { |
876 | |
877 | template class RawInstrProfReader<uint32_t>; |
878 | template class RawInstrProfReader<uint64_t>; |
879 | |
880 | } // end namespace llvm |
881 | |
882 | InstrProfLookupTrait::hash_value_type |
883 | InstrProfLookupTrait::ComputeHash(StringRef K) { |
884 | return IndexedInstrProf::ComputeHash(Type: HashType, K); |
885 | } |
886 | |
887 | using data_type = InstrProfLookupTrait::data_type; |
888 | using offset_type = InstrProfLookupTrait::offset_type; |
889 | |
890 | bool InstrProfLookupTrait::readValueProfilingData( |
891 | const unsigned char *&D, const unsigned char *const End) { |
892 | Expected<std::unique_ptr<ValueProfData>> VDataPtrOrErr = |
893 | ValueProfData::getValueProfData(SrcBuffer: D, SrcBufferEnd: End, SrcDataEndianness: ValueProfDataEndianness); |
894 | |
895 | if (VDataPtrOrErr.takeError()) |
896 | return false; |
897 | |
898 | VDataPtrOrErr.get()->deserializeTo(Record&: DataBuffer.back(), SymTab: nullptr); |
899 | D += VDataPtrOrErr.get()->TotalSize; |
900 | |
901 | return true; |
902 | } |
903 | |
904 | data_type InstrProfLookupTrait::ReadData(StringRef K, const unsigned char *D, |
905 | offset_type N) { |
906 | using namespace support; |
907 | |
908 | // Check if the data is corrupt. If so, don't try to read it. |
909 | if (N % sizeof(uint64_t)) |
910 | return data_type(); |
911 | |
912 | DataBuffer.clear(); |
913 | std::vector<uint64_t> CounterBuffer; |
914 | std::vector<uint8_t> BitmapByteBuffer; |
915 | |
916 | const unsigned char *End = D + N; |
917 | while (D < End) { |
918 | // Read hash. |
919 | if (D + sizeof(uint64_t) >= End) |
920 | return data_type(); |
921 | uint64_t Hash = endian::readNext<uint64_t, llvm::endianness::little>(memory&: D); |
922 | |
923 | // Initialize number of counters for GET_VERSION(FormatVersion) == 1. |
924 | uint64_t CountsSize = N / sizeof(uint64_t) - 1; |
925 | // If format version is different then read the number of counters. |
926 | if (GET_VERSION(FormatVersion) != IndexedInstrProf::ProfVersion::Version1) { |
927 | if (D + sizeof(uint64_t) > End) |
928 | return data_type(); |
929 | CountsSize = endian::readNext<uint64_t, llvm::endianness::little>(memory&: D); |
930 | } |
931 | // Read counter values. |
932 | if (D + CountsSize * sizeof(uint64_t) > End) |
933 | return data_type(); |
934 | |
935 | CounterBuffer.clear(); |
936 | CounterBuffer.reserve(n: CountsSize); |
937 | for (uint64_t J = 0; J < CountsSize; ++J) |
938 | CounterBuffer.push_back( |
939 | x: endian::readNext<uint64_t, llvm::endianness::little>(memory&: D)); |
940 | |
941 | // Read bitmap bytes for GET_VERSION(FormatVersion) > 10. |
942 | if (GET_VERSION(FormatVersion) > IndexedInstrProf::ProfVersion::Version10) { |
943 | uint64_t BitmapBytes = 0; |
944 | if (D + sizeof(uint64_t) > End) |
945 | return data_type(); |
946 | BitmapBytes = endian::readNext<uint64_t, llvm::endianness::little>(memory&: D); |
947 | // Read bitmap byte values. |
948 | if (D + BitmapBytes * sizeof(uint8_t) > End) |
949 | return data_type(); |
950 | BitmapByteBuffer.clear(); |
951 | BitmapByteBuffer.reserve(n: BitmapBytes); |
952 | for (uint64_t J = 0; J < BitmapBytes; ++J) |
953 | BitmapByteBuffer.push_back(x: static_cast<uint8_t>( |
954 | endian::readNext<uint64_t, llvm::endianness::little>(memory&: D))); |
955 | } |
956 | |
957 | DataBuffer.emplace_back(args&: K, args&: Hash, args: std::move(CounterBuffer), |
958 | args: std::move(BitmapByteBuffer)); |
959 | |
960 | // Read value profiling data. |
961 | if (GET_VERSION(FormatVersion) > IndexedInstrProf::ProfVersion::Version2 && |
962 | !readValueProfilingData(D, End)) { |
963 | DataBuffer.clear(); |
964 | return data_type(); |
965 | } |
966 | } |
967 | return DataBuffer; |
968 | } |
969 | |
970 | template <typename HashTableImpl> |
971 | Error InstrProfReaderIndex<HashTableImpl>::getRecords( |
972 | StringRef FuncName, ArrayRef<NamedInstrProfRecord> &Data) { |
973 | auto Iter = HashTable->find(FuncName); |
974 | if (Iter == HashTable->end()) |
975 | return make_error<InstrProfError>(Args: instrprof_error::unknown_function); |
976 | |
977 | Data = (*Iter); |
978 | if (Data.empty()) |
979 | return make_error<InstrProfError>(Args: instrprof_error::malformed, |
980 | Args: "profile data is empty" ); |
981 | |
982 | return Error::success(); |
983 | } |
984 | |
985 | template <typename HashTableImpl> |
986 | Error InstrProfReaderIndex<HashTableImpl>::getRecords( |
987 | ArrayRef<NamedInstrProfRecord> &Data) { |
988 | if (atEnd()) |
989 | return make_error<InstrProfError>(Args: instrprof_error::eof); |
990 | |
991 | Data = *RecordIterator; |
992 | |
993 | if (Data.empty()) |
994 | return make_error<InstrProfError>(Args: instrprof_error::malformed, |
995 | Args: "profile data is empty" ); |
996 | |
997 | return Error::success(); |
998 | } |
999 | |
1000 | template <typename HashTableImpl> |
1001 | InstrProfReaderIndex<HashTableImpl>::InstrProfReaderIndex( |
1002 | const unsigned char *Buckets, const unsigned char *const Payload, |
1003 | const unsigned char *const Base, IndexedInstrProf::HashT HashType, |
1004 | uint64_t Version) { |
1005 | FormatVersion = Version; |
1006 | HashTable.reset(HashTableImpl::Create( |
1007 | Buckets, Payload, Base, |
1008 | typename HashTableImpl::InfoType(HashType, Version))); |
1009 | RecordIterator = HashTable->data_begin(); |
1010 | } |
1011 | |
1012 | template <typename HashTableImpl> |
1013 | InstrProfKind InstrProfReaderIndex<HashTableImpl>::getProfileKind() const { |
1014 | return getProfileKindFromVersion(Version: FormatVersion); |
1015 | } |
1016 | |
1017 | namespace { |
1018 | /// A remapper that does not apply any remappings. |
1019 | class InstrProfReaderNullRemapper : public InstrProfReaderRemapper { |
1020 | InstrProfReaderIndexBase &Underlying; |
1021 | |
1022 | public: |
1023 | InstrProfReaderNullRemapper(InstrProfReaderIndexBase &Underlying) |
1024 | : Underlying(Underlying) {} |
1025 | |
1026 | Error getRecords(StringRef FuncName, |
1027 | ArrayRef<NamedInstrProfRecord> &Data) override { |
1028 | return Underlying.getRecords(FuncName, Data); |
1029 | } |
1030 | }; |
1031 | } // namespace |
1032 | |
1033 | /// A remapper that applies remappings based on a symbol remapping file. |
1034 | template <typename HashTableImpl> |
1035 | class llvm::InstrProfReaderItaniumRemapper |
1036 | : public InstrProfReaderRemapper { |
1037 | public: |
1038 | InstrProfReaderItaniumRemapper( |
1039 | std::unique_ptr<MemoryBuffer> RemapBuffer, |
1040 | InstrProfReaderIndex<HashTableImpl> &Underlying) |
1041 | : RemapBuffer(std::move(RemapBuffer)), Underlying(Underlying) { |
1042 | } |
1043 | |
1044 | /// Extract the original function name from a PGO function name. |
1045 | static StringRef (StringRef Name) { |
1046 | // We can have multiple pieces separated by kGlobalIdentifierDelimiter ( |
1047 | // semicolon now and colon in older profiles); there can be pieces both |
1048 | // before and after the mangled name. Find the first part that starts with |
1049 | // '_Z'; we'll assume that's the mangled name we want. |
1050 | std::pair<StringRef, StringRef> Parts = {StringRef(), Name}; |
1051 | while (true) { |
1052 | Parts = Parts.second.split(Separator: GlobalIdentifierDelimiter); |
1053 | if (Parts.first.starts_with(Prefix: "_Z" )) |
1054 | return Parts.first; |
1055 | if (Parts.second.empty()) |
1056 | return Name; |
1057 | } |
1058 | } |
1059 | |
1060 | /// Given a mangled name extracted from a PGO function name, and a new |
1061 | /// form for that mangled name, reconstitute the name. |
1062 | static void reconstituteName(StringRef OrigName, StringRef , |
1063 | StringRef Replacement, |
1064 | SmallVectorImpl<char> &Out) { |
1065 | Out.reserve(N: OrigName.size() + Replacement.size() - ExtractedName.size()); |
1066 | Out.insert(I: Out.end(), From: OrigName.begin(), To: ExtractedName.begin()); |
1067 | Out.insert(I: Out.end(), From: Replacement.begin(), To: Replacement.end()); |
1068 | Out.insert(I: Out.end(), From: ExtractedName.end(), To: OrigName.end()); |
1069 | } |
1070 | |
1071 | Error populateRemappings() override { |
1072 | if (Error E = Remappings.read(B&: *RemapBuffer)) |
1073 | return E; |
1074 | for (StringRef Name : Underlying.HashTable->keys()) { |
1075 | StringRef RealName = extractName(Name); |
1076 | if (auto Key = Remappings.insert(FunctionName: RealName)) { |
1077 | // FIXME: We could theoretically map the same equivalence class to |
1078 | // multiple names in the profile data. If that happens, we should |
1079 | // return NamedInstrProfRecords from all of them. |
1080 | MappedNames.insert(KV: {Key, RealName}); |
1081 | } |
1082 | } |
1083 | return Error::success(); |
1084 | } |
1085 | |
1086 | Error getRecords(StringRef FuncName, |
1087 | ArrayRef<NamedInstrProfRecord> &Data) override { |
1088 | StringRef RealName = extractName(Name: FuncName); |
1089 | if (auto Key = Remappings.lookup(FunctionName: RealName)) { |
1090 | StringRef Remapped = MappedNames.lookup(Val: Key); |
1091 | if (!Remapped.empty()) { |
1092 | if (RealName.begin() == FuncName.begin() && |
1093 | RealName.end() == FuncName.end()) |
1094 | FuncName = Remapped; |
1095 | else { |
1096 | // Try rebuilding the name from the given remapping. |
1097 | SmallString<256> Reconstituted; |
1098 | reconstituteName(OrigName: FuncName, ExtractedName: RealName, Replacement: Remapped, Out&: Reconstituted); |
1099 | Error E = Underlying.getRecords(Reconstituted, Data); |
1100 | if (!E) |
1101 | return E; |
1102 | |
1103 | // If we failed because the name doesn't exist, fall back to asking |
1104 | // about the original name. |
1105 | if (Error Unhandled = handleErrors( |
1106 | std::move(E), [](std::unique_ptr<InstrProfError> Err) { |
1107 | return Err->get() == instrprof_error::unknown_function |
1108 | ? Error::success() |
1109 | : Error(std::move(Err)); |
1110 | })) |
1111 | return Unhandled; |
1112 | } |
1113 | } |
1114 | } |
1115 | return Underlying.getRecords(FuncName, Data); |
1116 | } |
1117 | |
1118 | private: |
1119 | /// The memory buffer containing the remapping configuration. Remappings |
1120 | /// holds pointers into this buffer. |
1121 | std::unique_ptr<MemoryBuffer> RemapBuffer; |
1122 | |
1123 | /// The mangling remapper. |
1124 | SymbolRemappingReader Remappings; |
1125 | |
1126 | /// Mapping from mangled name keys to the name used for the key in the |
1127 | /// profile data. |
1128 | /// FIXME: Can we store a location within the on-disk hash table instead of |
1129 | /// redoing lookup? |
1130 | DenseMap<SymbolRemappingReader::Key, StringRef> MappedNames; |
1131 | |
1132 | /// The real profile data reader. |
1133 | InstrProfReaderIndex<HashTableImpl> &Underlying; |
1134 | }; |
1135 | |
1136 | bool IndexedInstrProfReader::hasFormat(const MemoryBuffer &DataBuffer) { |
1137 | using namespace support; |
1138 | |
1139 | if (DataBuffer.getBufferSize() < 8) |
1140 | return false; |
1141 | uint64_t Magic = endian::read<uint64_t, llvm::endianness::little, aligned>( |
1142 | memory: DataBuffer.getBufferStart()); |
1143 | // Verify that it's magical. |
1144 | return Magic == IndexedInstrProf::Magic; |
1145 | } |
1146 | |
1147 | const unsigned char * |
1148 | IndexedInstrProfReader::readSummary(IndexedInstrProf::ProfVersion Version, |
1149 | const unsigned char *Cur, bool UseCS) { |
1150 | using namespace IndexedInstrProf; |
1151 | using namespace support; |
1152 | |
1153 | if (Version >= IndexedInstrProf::Version4) { |
1154 | const IndexedInstrProf::Summary *SummaryInLE = |
1155 | reinterpret_cast<const IndexedInstrProf::Summary *>(Cur); |
1156 | uint64_t NFields = endian::byte_swap<uint64_t, llvm::endianness::little>( |
1157 | value: SummaryInLE->NumSummaryFields); |
1158 | uint64_t NEntries = endian::byte_swap<uint64_t, llvm::endianness::little>( |
1159 | value: SummaryInLE->NumCutoffEntries); |
1160 | uint32_t SummarySize = |
1161 | IndexedInstrProf::Summary::getSize(NumSumFields: NFields, NumCutoffEntries: NEntries); |
1162 | std::unique_ptr<IndexedInstrProf::Summary> SummaryData = |
1163 | IndexedInstrProf::allocSummary(TotalSize: SummarySize); |
1164 | |
1165 | const uint64_t *Src = reinterpret_cast<const uint64_t *>(SummaryInLE); |
1166 | uint64_t *Dst = reinterpret_cast<uint64_t *>(SummaryData.get()); |
1167 | for (unsigned I = 0; I < SummarySize / sizeof(uint64_t); I++) |
1168 | Dst[I] = endian::byte_swap<uint64_t, llvm::endianness::little>(value: Src[I]); |
1169 | |
1170 | SummaryEntryVector DetailedSummary; |
1171 | for (unsigned I = 0; I < SummaryData->NumCutoffEntries; I++) { |
1172 | const IndexedInstrProf::Summary::Entry &Ent = SummaryData->getEntry(I); |
1173 | DetailedSummary.emplace_back(args: (uint32_t)Ent.Cutoff, args: Ent.MinBlockCount, |
1174 | args: Ent.NumBlocks); |
1175 | } |
1176 | std::unique_ptr<llvm::ProfileSummary> &Summary = |
1177 | UseCS ? this->CS_Summary : this->Summary; |
1178 | |
1179 | // initialize InstrProfSummary using the SummaryData from disk. |
1180 | Summary = std::make_unique<ProfileSummary>( |
1181 | args: UseCS ? ProfileSummary::PSK_CSInstr : ProfileSummary::PSK_Instr, |
1182 | args&: DetailedSummary, args: SummaryData->get(K: Summary::TotalBlockCount), |
1183 | args: SummaryData->get(K: Summary::MaxBlockCount), |
1184 | args: SummaryData->get(K: Summary::MaxInternalBlockCount), |
1185 | args: SummaryData->get(K: Summary::MaxFunctionCount), |
1186 | args: SummaryData->get(K: Summary::TotalNumBlocks), |
1187 | args: SummaryData->get(K: Summary::TotalNumFunctions)); |
1188 | return Cur + SummarySize; |
1189 | } else { |
1190 | // The older versions do not support a profile summary. This just computes |
1191 | // an empty summary, which will not result in accurate hot/cold detection. |
1192 | // We would need to call addRecord for all NamedInstrProfRecords to get the |
1193 | // correct summary. However, this version is old (prior to early 2016) and |
1194 | // has not been supporting an accurate summary for several years. |
1195 | InstrProfSummaryBuilder Builder(ProfileSummaryBuilder::DefaultCutoffs); |
1196 | Summary = Builder.getSummary(); |
1197 | return Cur; |
1198 | } |
1199 | } |
1200 | |
1201 | Error IndexedMemProfReader::deserializeV012(const unsigned char *Start, |
1202 | const unsigned char *Ptr, |
1203 | uint64_t FirstWord) { |
1204 | // The value returned from RecordTableGenerator.Emit. |
1205 | const uint64_t RecordTableOffset = |
1206 | Version == memprof::Version0 |
1207 | ? FirstWord |
1208 | : support::endian::readNext<uint64_t, llvm::endianness::little>(memory&: Ptr); |
1209 | // The offset in the stream right before invoking |
1210 | // FrameTableGenerator.Emit. |
1211 | const uint64_t FramePayloadOffset = |
1212 | support::endian::readNext<uint64_t, llvm::endianness::little>(memory&: Ptr); |
1213 | // The value returned from FrameTableGenerator.Emit. |
1214 | const uint64_t FrameTableOffset = |
1215 | support::endian::readNext<uint64_t, llvm::endianness::little>(memory&: Ptr); |
1216 | |
1217 | // The offset in the stream right before invoking |
1218 | // CallStackTableGenerator.Emit. |
1219 | uint64_t CallStackPayloadOffset = 0; |
1220 | // The value returned from CallStackTableGenerator.Emit. |
1221 | uint64_t CallStackTableOffset = 0; |
1222 | if (Version >= memprof::Version2) { |
1223 | CallStackPayloadOffset = |
1224 | support::endian::readNext<uint64_t, llvm::endianness::little>(memory&: Ptr); |
1225 | CallStackTableOffset = |
1226 | support::endian::readNext<uint64_t, llvm::endianness::little>(memory&: Ptr); |
1227 | } |
1228 | |
1229 | // Read the schema. |
1230 | auto SchemaOr = memprof::readMemProfSchema(Buffer&: Ptr); |
1231 | if (!SchemaOr) |
1232 | return SchemaOr.takeError(); |
1233 | Schema = SchemaOr.get(); |
1234 | |
1235 | // Now initialize the table reader with a pointer into data buffer. |
1236 | MemProfRecordTable.reset(p: MemProfRecordHashTable::Create( |
1237 | /*Buckets=*/Start + RecordTableOffset, |
1238 | /*Payload=*/Ptr, |
1239 | /*Base=*/Start, InfoObj: memprof::RecordLookupTrait(Version, Schema))); |
1240 | |
1241 | // Initialize the frame table reader with the payload and bucket offsets. |
1242 | MemProfFrameTable.reset(p: MemProfFrameHashTable::Create( |
1243 | /*Buckets=*/Start + FrameTableOffset, |
1244 | /*Payload=*/Start + FramePayloadOffset, |
1245 | /*Base=*/Start)); |
1246 | |
1247 | if (Version >= memprof::Version2) |
1248 | MemProfCallStackTable.reset(p: MemProfCallStackHashTable::Create( |
1249 | /*Buckets=*/Start + CallStackTableOffset, |
1250 | /*Payload=*/Start + CallStackPayloadOffset, |
1251 | /*Base=*/Start)); |
1252 | |
1253 | return Error::success(); |
1254 | } |
1255 | |
1256 | Error IndexedMemProfReader::deserializeV3(const unsigned char *Start, |
1257 | const unsigned char *Ptr) { |
1258 | // The offset in the stream right before invoking |
1259 | // CallStackTableGenerator.Emit. |
1260 | const uint64_t CallStackPayloadOffset = |
1261 | support::endian::readNext<uint64_t, llvm::endianness::little>(memory&: Ptr); |
1262 | // The offset in the stream right before invoking RecordTableGenerator.Emit. |
1263 | const uint64_t RecordPayloadOffset = |
1264 | support::endian::readNext<uint64_t, llvm::endianness::little>(memory&: Ptr); |
1265 | // The value returned from RecordTableGenerator.Emit. |
1266 | const uint64_t RecordTableOffset = |
1267 | support::endian::readNext<uint64_t, llvm::endianness::little>(memory&: Ptr); |
1268 | |
1269 | // Read the schema. |
1270 | auto SchemaOr = memprof::readMemProfSchema(Buffer&: Ptr); |
1271 | if (!SchemaOr) |
1272 | return SchemaOr.takeError(); |
1273 | Schema = SchemaOr.get(); |
1274 | |
1275 | FrameBase = Ptr; |
1276 | CallStackBase = Start + CallStackPayloadOffset; |
1277 | |
1278 | // Now initialize the table reader with a pointer into data buffer. |
1279 | MemProfRecordTable.reset(p: MemProfRecordHashTable::Create( |
1280 | /*Buckets=*/Start + RecordTableOffset, |
1281 | /*Payload=*/Start + RecordPayloadOffset, |
1282 | /*Base=*/Start, InfoObj: memprof::RecordLookupTrait(memprof::Version3, Schema))); |
1283 | |
1284 | return Error::success(); |
1285 | } |
1286 | |
1287 | Error IndexedMemProfReader::deserialize(const unsigned char *Start, |
1288 | uint64_t MemProfOffset) { |
1289 | const unsigned char *Ptr = Start + MemProfOffset; |
1290 | |
1291 | // Read the first 64-bit word, which may be RecordTableOffset in |
1292 | // memprof::MemProfVersion0 or the MemProf version number in |
1293 | // memprof::MemProfVersion1 and above. |
1294 | const uint64_t FirstWord = |
1295 | support::endian::readNext<uint64_t, llvm::endianness::little>(memory&: Ptr); |
1296 | |
1297 | if (FirstWord == memprof::Version1 || FirstWord == memprof::Version2 || |
1298 | FirstWord == memprof::Version3) { |
1299 | // Everything is good. We can proceed to deserialize the rest. |
1300 | Version = static_cast<memprof::IndexedVersion>(FirstWord); |
1301 | } else if (FirstWord >= 24) { |
1302 | // This is a heuristic/hack to detect memprof::MemProfVersion0, |
1303 | // which does not have a version field in the header. |
1304 | // In memprof::MemProfVersion0, FirstWord will be RecordTableOffset, |
1305 | // which should be at least 24 because of the MemProf header size. |
1306 | Version = memprof::Version0; |
1307 | } else { |
1308 | return make_error<InstrProfError>( |
1309 | Args: instrprof_error::unsupported_version, |
1310 | Args: formatv(Fmt: "MemProf version {} not supported; " |
1311 | "requires version between {} and {}, inclusive" , |
1312 | Vals: FirstWord, Vals: memprof::MinimumSupportedVersion, |
1313 | Vals: memprof::MaximumSupportedVersion)); |
1314 | } |
1315 | |
1316 | switch (Version) { |
1317 | case memprof::Version0: |
1318 | case memprof::Version1: |
1319 | case memprof::Version2: |
1320 | if (Error E = deserializeV012(Start, Ptr, FirstWord)) |
1321 | return E; |
1322 | break; |
1323 | case memprof::Version3: |
1324 | if (Error E = deserializeV3(Start, Ptr)) |
1325 | return E; |
1326 | break; |
1327 | } |
1328 | |
1329 | #ifdef EXPENSIVE_CHECKS |
1330 | // Go through all the records and verify that CSId has been correctly |
1331 | // populated. Do this only under EXPENSIVE_CHECKS. Otherwise, we |
1332 | // would defeat the purpose of OnDiskIterableChainedHashTable. |
1333 | // Note that we can compare CSId against actual call stacks only for |
1334 | // Version0 and Version1 because IndexedAllocationInfo::CallStack and |
1335 | // IndexedMemProfRecord::CallSites are not populated in Version2. |
1336 | if (Version <= memprof::Version1) |
1337 | for (const auto &Record : MemProfRecordTable->data()) |
1338 | verifyIndexedMemProfRecord(Record); |
1339 | #endif |
1340 | |
1341 | return Error::success(); |
1342 | } |
1343 | |
1344 | Error IndexedInstrProfReader::() { |
1345 | using namespace support; |
1346 | |
1347 | const unsigned char *Start = |
1348 | (const unsigned char *)DataBuffer->getBufferStart(); |
1349 | const unsigned char *Cur = Start; |
1350 | if ((const unsigned char *)DataBuffer->getBufferEnd() - Cur < 24) |
1351 | return error(Err: instrprof_error::truncated); |
1352 | |
1353 | auto = IndexedInstrProf::Header::readFromBuffer(Buffer: Start); |
1354 | if (!HeaderOr) |
1355 | return HeaderOr.takeError(); |
1356 | |
1357 | const IndexedInstrProf::Header * = &HeaderOr.get(); |
1358 | Cur += Header->size(); |
1359 | |
1360 | Cur = readSummary(Version: (IndexedInstrProf::ProfVersion)Header->Version, Cur, |
1361 | /* UseCS */ false); |
1362 | if (Header->Version & VARIANT_MASK_CSIR_PROF) |
1363 | Cur = readSummary(Version: (IndexedInstrProf::ProfVersion)Header->Version, Cur, |
1364 | /* UseCS */ true); |
1365 | // Read the hash type and start offset. |
1366 | IndexedInstrProf::HashT HashType = |
1367 | static_cast<IndexedInstrProf::HashT>(Header->HashType); |
1368 | if (HashType > IndexedInstrProf::HashT::Last) |
1369 | return error(Err: instrprof_error::unsupported_hash_type); |
1370 | |
1371 | // The hash table with profile counts comes next. |
1372 | auto IndexPtr = std::make_unique<InstrProfReaderIndex<OnDiskHashTableImplV3>>( |
1373 | args: Start + Header->HashOffset, args&: Cur, args&: Start, args&: HashType, args: Header->Version); |
1374 | |
1375 | // The MemProfOffset field in the header is only valid when the format |
1376 | // version is higher than 8 (when it was introduced). |
1377 | if (Header->getIndexedProfileVersion() >= 8 && |
1378 | Header->Version & VARIANT_MASK_MEMPROF) { |
1379 | if (Error E = MemProfReader.deserialize(Start, MemProfOffset: Header->MemProfOffset)) |
1380 | return E; |
1381 | } |
1382 | |
1383 | // BinaryIdOffset field in the header is only valid when the format version |
1384 | // is higher than 9 (when it was introduced). |
1385 | if (Header->getIndexedProfileVersion() >= 9) { |
1386 | const unsigned char *Ptr = Start + Header->BinaryIdOffset; |
1387 | // Read binary ids size. |
1388 | uint64_t BinaryIdsSize = |
1389 | support::endian::readNext<uint64_t, llvm::endianness::little>(memory&: Ptr); |
1390 | if (BinaryIdsSize % sizeof(uint64_t)) |
1391 | return error(Err: instrprof_error::bad_header); |
1392 | // Set the binary ids start. |
1393 | BinaryIdsBuffer = ArrayRef<uint8_t>(Ptr, BinaryIdsSize); |
1394 | if (Ptr > (const unsigned char *)DataBuffer->getBufferEnd()) |
1395 | return make_error<InstrProfError>(Args: instrprof_error::malformed, |
1396 | Args: "corrupted binary ids" ); |
1397 | } |
1398 | |
1399 | if (Header->getIndexedProfileVersion() >= 12) { |
1400 | const unsigned char *Ptr = Start + Header->VTableNamesOffset; |
1401 | |
1402 | uint64_t CompressedVTableNamesLen = |
1403 | support::endian::readNext<uint64_t, llvm::endianness::little>(memory&: Ptr); |
1404 | |
1405 | // Writer first writes the length of compressed string, and then the actual |
1406 | // content. |
1407 | const char *VTableNamePtr = (const char *)Ptr; |
1408 | if (VTableNamePtr > (const char *)DataBuffer->getBufferEnd()) |
1409 | return make_error<InstrProfError>(Args: instrprof_error::truncated); |
1410 | |
1411 | VTableName = StringRef(VTableNamePtr, CompressedVTableNamesLen); |
1412 | } |
1413 | |
1414 | if (Header->getIndexedProfileVersion() >= 10 && |
1415 | Header->Version & VARIANT_MASK_TEMPORAL_PROF) { |
1416 | const unsigned char *Ptr = Start + Header->TemporalProfTracesOffset; |
1417 | const auto *PtrEnd = (const unsigned char *)DataBuffer->getBufferEnd(); |
1418 | // Expect at least two 64 bit fields: NumTraces, and TraceStreamSize |
1419 | if (Ptr + 2 * sizeof(uint64_t) > PtrEnd) |
1420 | return error(Err: instrprof_error::truncated); |
1421 | const uint64_t NumTraces = |
1422 | support::endian::readNext<uint64_t, llvm::endianness::little>(memory&: Ptr); |
1423 | TemporalProfTraceStreamSize = |
1424 | support::endian::readNext<uint64_t, llvm::endianness::little>(memory&: Ptr); |
1425 | for (unsigned i = 0; i < NumTraces; i++) { |
1426 | // Expect at least two 64 bit fields: Weight and NumFunctions |
1427 | if (Ptr + 2 * sizeof(uint64_t) > PtrEnd) |
1428 | return error(Err: instrprof_error::truncated); |
1429 | TemporalProfTraceTy Trace; |
1430 | Trace.Weight = |
1431 | support::endian::readNext<uint64_t, llvm::endianness::little>(memory&: Ptr); |
1432 | const uint64_t NumFunctions = |
1433 | support::endian::readNext<uint64_t, llvm::endianness::little>(memory&: Ptr); |
1434 | // Expect at least NumFunctions 64 bit fields |
1435 | if (Ptr + NumFunctions * sizeof(uint64_t) > PtrEnd) |
1436 | return error(Err: instrprof_error::truncated); |
1437 | for (unsigned j = 0; j < NumFunctions; j++) { |
1438 | const uint64_t NameRef = |
1439 | support::endian::readNext<uint64_t, llvm::endianness::little>(memory&: Ptr); |
1440 | Trace.FunctionNameRefs.push_back(x: NameRef); |
1441 | } |
1442 | TemporalProfTraces.push_back(Elt: std::move(Trace)); |
1443 | } |
1444 | } |
1445 | |
1446 | // Load the remapping table now if requested. |
1447 | if (RemappingBuffer) { |
1448 | Remapper = |
1449 | std::make_unique<InstrProfReaderItaniumRemapper<OnDiskHashTableImplV3>>( |
1450 | args: std::move(RemappingBuffer), args&: *IndexPtr); |
1451 | if (Error E = Remapper->populateRemappings()) |
1452 | return E; |
1453 | } else { |
1454 | Remapper = std::make_unique<InstrProfReaderNullRemapper>(args&: *IndexPtr); |
1455 | } |
1456 | Index = std::move(IndexPtr); |
1457 | |
1458 | return success(); |
1459 | } |
1460 | |
1461 | InstrProfSymtab &IndexedInstrProfReader::getSymtab() { |
1462 | if (Symtab) |
1463 | return *Symtab; |
1464 | |
1465 | auto NewSymtab = std::make_unique<InstrProfSymtab>(); |
1466 | |
1467 | if (Error E = NewSymtab->initVTableNamesFromCompressedStrings(CompressedVTableNames: VTableName)) { |
1468 | auto [ErrCode, Msg] = InstrProfError::take(E: std::move(E)); |
1469 | consumeError(Err: error(Err: ErrCode, ErrMsg: Msg)); |
1470 | } |
1471 | |
1472 | // finalizeSymtab is called inside populateSymtab. |
1473 | if (Error E = Index->populateSymtab(*NewSymtab)) { |
1474 | auto [ErrCode, Msg] = InstrProfError::take(E: std::move(E)); |
1475 | consumeError(Err: error(Err: ErrCode, ErrMsg: Msg)); |
1476 | } |
1477 | |
1478 | Symtab = std::move(NewSymtab); |
1479 | return *Symtab; |
1480 | } |
1481 | |
1482 | Expected<InstrProfRecord> IndexedInstrProfReader::getInstrProfRecord( |
1483 | StringRef FuncName, uint64_t FuncHash, StringRef DeprecatedFuncName, |
1484 | uint64_t *MismatchedFuncSum) { |
1485 | ArrayRef<NamedInstrProfRecord> Data; |
1486 | uint64_t FuncSum = 0; |
1487 | auto Err = Remapper->getRecords(FuncName, Data); |
1488 | if (Err) { |
1489 | // If we don't find FuncName, try DeprecatedFuncName to handle profiles |
1490 | // built by older compilers. |
1491 | auto Err2 = |
1492 | handleErrors(E: std::move(Err), Hs: [&](const InstrProfError &IE) -> Error { |
1493 | if (IE.get() != instrprof_error::unknown_function) |
1494 | return make_error<InstrProfError>(Args: IE); |
1495 | if (auto Err = Remapper->getRecords(FuncName: DeprecatedFuncName, Data)) |
1496 | return Err; |
1497 | return Error::success(); |
1498 | }); |
1499 | if (Err2) |
1500 | return std::move(Err2); |
1501 | } |
1502 | // Found it. Look for counters with the right hash. |
1503 | |
1504 | // A flag to indicate if the records are from the same type |
1505 | // of profile (i.e cs vs nocs). |
1506 | bool CSBitMatch = false; |
1507 | auto getFuncSum = [](ArrayRef<uint64_t> Counts) { |
1508 | uint64_t ValueSum = 0; |
1509 | for (uint64_t CountValue : Counts) { |
1510 | if (CountValue == (uint64_t)-1) |
1511 | continue; |
1512 | // Handle overflow -- if that happens, return max. |
1513 | if (std::numeric_limits<uint64_t>::max() - CountValue <= ValueSum) |
1514 | return std::numeric_limits<uint64_t>::max(); |
1515 | ValueSum += CountValue; |
1516 | } |
1517 | return ValueSum; |
1518 | }; |
1519 | |
1520 | for (const NamedInstrProfRecord &I : Data) { |
1521 | // Check for a match and fill the vector if there is one. |
1522 | if (I.Hash == FuncHash) |
1523 | return std::move(I); |
1524 | if (NamedInstrProfRecord::hasCSFlagInHash(FuncHash: I.Hash) == |
1525 | NamedInstrProfRecord::hasCSFlagInHash(FuncHash)) { |
1526 | CSBitMatch = true; |
1527 | if (MismatchedFuncSum == nullptr) |
1528 | continue; |
1529 | FuncSum = std::max(a: FuncSum, b: getFuncSum(I.Counts)); |
1530 | } |
1531 | } |
1532 | if (CSBitMatch) { |
1533 | if (MismatchedFuncSum != nullptr) |
1534 | *MismatchedFuncSum = FuncSum; |
1535 | return error(Err: instrprof_error::hash_mismatch); |
1536 | } |
1537 | return error(Err: instrprof_error::unknown_function); |
1538 | } |
1539 | |
1540 | static Expected<memprof::MemProfRecord> |
1541 | getMemProfRecordV0(const memprof::IndexedMemProfRecord &IndexedRecord, |
1542 | MemProfFrameHashTable &MemProfFrameTable) { |
1543 | memprof::FrameIdConverter<MemProfFrameHashTable> FrameIdConv( |
1544 | MemProfFrameTable); |
1545 | |
1546 | memprof::MemProfRecord Record = |
1547 | memprof::MemProfRecord(IndexedRecord, FrameIdConv); |
1548 | |
1549 | // Check that all frame ids were successfully converted to frames. |
1550 | if (FrameIdConv.LastUnmappedId) { |
1551 | return make_error<InstrProfError>(Args: instrprof_error::hash_mismatch, |
1552 | Args: "memprof frame not found for frame id " + |
1553 | Twine(*FrameIdConv.LastUnmappedId)); |
1554 | } |
1555 | |
1556 | return Record; |
1557 | } |
1558 | |
1559 | static Expected<memprof::MemProfRecord> |
1560 | getMemProfRecordV2(const memprof::IndexedMemProfRecord &IndexedRecord, |
1561 | MemProfFrameHashTable &MemProfFrameTable, |
1562 | MemProfCallStackHashTable &MemProfCallStackTable) { |
1563 | memprof::FrameIdConverter<MemProfFrameHashTable> FrameIdConv( |
1564 | MemProfFrameTable); |
1565 | |
1566 | memprof::CallStackIdConverter<MemProfCallStackHashTable> CSIdConv( |
1567 | MemProfCallStackTable, FrameIdConv); |
1568 | |
1569 | memprof::MemProfRecord Record = IndexedRecord.toMemProfRecord(Callback: CSIdConv); |
1570 | |
1571 | // Check that all call stack ids were successfully converted to call stacks. |
1572 | if (CSIdConv.LastUnmappedId) { |
1573 | return make_error<InstrProfError>( |
1574 | Args: instrprof_error::hash_mismatch, |
1575 | Args: "memprof call stack not found for call stack id " + |
1576 | Twine(*CSIdConv.LastUnmappedId)); |
1577 | } |
1578 | |
1579 | // Check that all frame ids were successfully converted to frames. |
1580 | if (FrameIdConv.LastUnmappedId) { |
1581 | return make_error<InstrProfError>(Args: instrprof_error::hash_mismatch, |
1582 | Args: "memprof frame not found for frame id " + |
1583 | Twine(*FrameIdConv.LastUnmappedId)); |
1584 | } |
1585 | |
1586 | return Record; |
1587 | } |
1588 | |
1589 | static Expected<memprof::MemProfRecord> |
1590 | getMemProfRecordV3(const memprof::IndexedMemProfRecord &IndexedRecord, |
1591 | const unsigned char *FrameBase, |
1592 | const unsigned char *CallStackBase) { |
1593 | memprof::LinearFrameIdConverter FrameIdConv(FrameBase); |
1594 | memprof::LinearCallStackIdConverter CSIdConv(CallStackBase, FrameIdConv); |
1595 | memprof::MemProfRecord Record = IndexedRecord.toMemProfRecord(Callback: CSIdConv); |
1596 | return Record; |
1597 | } |
1598 | |
1599 | Expected<memprof::MemProfRecord> |
1600 | IndexedMemProfReader::getMemProfRecord(const uint64_t FuncNameHash) const { |
1601 | // TODO: Add memprof specific errors. |
1602 | if (MemProfRecordTable == nullptr) |
1603 | return make_error<InstrProfError>(Args: instrprof_error::invalid_prof, |
1604 | Args: "no memprof data available in profile" ); |
1605 | auto Iter = MemProfRecordTable->find(EKey: FuncNameHash); |
1606 | if (Iter == MemProfRecordTable->end()) |
1607 | return make_error<InstrProfError>( |
1608 | Args: instrprof_error::unknown_function, |
1609 | Args: "memprof record not found for function hash " + Twine(FuncNameHash)); |
1610 | |
1611 | const memprof::IndexedMemProfRecord &IndexedRecord = *Iter; |
1612 | switch (Version) { |
1613 | case memprof::Version0: |
1614 | case memprof::Version1: |
1615 | assert(MemProfFrameTable && "MemProfFrameTable must be available" ); |
1616 | assert(!MemProfCallStackTable && |
1617 | "MemProfCallStackTable must not be available" ); |
1618 | return getMemProfRecordV0(IndexedRecord, MemProfFrameTable&: *MemProfFrameTable); |
1619 | case memprof::Version2: |
1620 | assert(MemProfFrameTable && "MemProfFrameTable must be available" ); |
1621 | assert(MemProfCallStackTable && "MemProfCallStackTable must be available" ); |
1622 | return getMemProfRecordV2(IndexedRecord, MemProfFrameTable&: *MemProfFrameTable, |
1623 | MemProfCallStackTable&: *MemProfCallStackTable); |
1624 | case memprof::Version3: |
1625 | assert(!MemProfFrameTable && "MemProfFrameTable must not be available" ); |
1626 | assert(!MemProfCallStackTable && |
1627 | "MemProfCallStackTable must not be available" ); |
1628 | assert(FrameBase && "FrameBase must be available" ); |
1629 | assert(CallStackBase && "CallStackBase must be available" ); |
1630 | return getMemProfRecordV3(IndexedRecord, FrameBase, CallStackBase); |
1631 | } |
1632 | |
1633 | return make_error<InstrProfError>( |
1634 | Args: instrprof_error::unsupported_version, |
1635 | Args: formatv(Fmt: "MemProf version {} not supported; " |
1636 | "requires version between {} and {}, inclusive" , |
1637 | Vals: Version, Vals: memprof::MinimumSupportedVersion, |
1638 | Vals: memprof::MaximumSupportedVersion)); |
1639 | } |
1640 | |
1641 | Error IndexedInstrProfReader::getFunctionCounts(StringRef FuncName, |
1642 | uint64_t FuncHash, |
1643 | std::vector<uint64_t> &Counts) { |
1644 | Expected<InstrProfRecord> Record = getInstrProfRecord(FuncName, FuncHash); |
1645 | if (Error E = Record.takeError()) |
1646 | return error(E: std::move(E)); |
1647 | |
1648 | Counts = Record.get().Counts; |
1649 | return success(); |
1650 | } |
1651 | |
1652 | Error IndexedInstrProfReader::getFunctionBitmap(StringRef FuncName, |
1653 | uint64_t FuncHash, |
1654 | BitVector &Bitmap) { |
1655 | Expected<InstrProfRecord> Record = getInstrProfRecord(FuncName, FuncHash); |
1656 | if (Error E = Record.takeError()) |
1657 | return error(E: std::move(E)); |
1658 | |
1659 | const auto &BitmapBytes = Record.get().BitmapBytes; |
1660 | size_t I = 0, E = BitmapBytes.size(); |
1661 | Bitmap.resize(N: E * CHAR_BIT); |
1662 | BitVector::apply( |
1663 | f: [&](auto X) { |
1664 | using XTy = decltype(X); |
1665 | alignas(XTy) uint8_t W[sizeof(X)]; |
1666 | size_t N = std::min(a: E - I, b: sizeof(W)); |
1667 | std::memset(s: W, c: 0, n: sizeof(W)); |
1668 | std::memcpy(dest: W, src: &BitmapBytes[I], n: N); |
1669 | I += N; |
1670 | return support::endian::read<XTy, llvm::endianness::little, |
1671 | support::aligned>(W); |
1672 | }, |
1673 | Out&: Bitmap, Arg: Bitmap); |
1674 | assert(I == E); |
1675 | |
1676 | return success(); |
1677 | } |
1678 | |
1679 | Error IndexedInstrProfReader::readNextRecord(NamedInstrProfRecord &Record) { |
1680 | ArrayRef<NamedInstrProfRecord> Data; |
1681 | |
1682 | Error E = Index->getRecords(Data); |
1683 | if (E) |
1684 | return error(E: std::move(E)); |
1685 | |
1686 | Record = Data[RecordIndex++]; |
1687 | if (RecordIndex >= Data.size()) { |
1688 | Index->advanceToNextKey(); |
1689 | RecordIndex = 0; |
1690 | } |
1691 | return success(); |
1692 | } |
1693 | |
1694 | Error IndexedInstrProfReader::readBinaryIds( |
1695 | std::vector<llvm::object::BuildID> &BinaryIds) { |
1696 | return readBinaryIdsInternal(DataBuffer: *DataBuffer, BinaryIdsBuffer, BinaryIds, |
1697 | Endian: llvm::endianness::little); |
1698 | } |
1699 | |
1700 | Error IndexedInstrProfReader::printBinaryIds(raw_ostream &OS) { |
1701 | std::vector<llvm::object::BuildID> BinaryIds; |
1702 | if (Error E = readBinaryIds(BinaryIds)) |
1703 | return E; |
1704 | printBinaryIdsInternal(OS, BinaryIds); |
1705 | return Error::success(); |
1706 | } |
1707 | |
1708 | void InstrProfReader::accumulateCounts(CountSumOrPercent &Sum, bool IsCS) { |
1709 | uint64_t NumFuncs = 0; |
1710 | for (const auto &Func : *this) { |
1711 | if (isIRLevelProfile()) { |
1712 | bool FuncIsCS = NamedInstrProfRecord::hasCSFlagInHash(FuncHash: Func.Hash); |
1713 | if (FuncIsCS != IsCS) |
1714 | continue; |
1715 | } |
1716 | Func.accumulateCounts(Sum); |
1717 | ++NumFuncs; |
1718 | } |
1719 | Sum.NumEntries = NumFuncs; |
1720 | } |
1721 | |