1 | //===- SampleProfReader.cpp - Read LLVM sample profile data ---------------===// |
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 implements the class that reads LLVM sample profiles. It |
10 | // supports three file formats: text, binary and gcov. |
11 | // |
12 | // The textual representation is useful for debugging and testing purposes. The |
13 | // binary representation is more compact, resulting in smaller file sizes. |
14 | // |
15 | // The gcov encoding is the one generated by GCC's AutoFDO profile creation |
16 | // tool (https://github.com/google/autofdo) |
17 | // |
18 | // All three encodings can be used interchangeably as an input sample profile. |
19 | // |
20 | //===----------------------------------------------------------------------===// |
21 | |
22 | #include "llvm/ProfileData/SampleProfReader.h" |
23 | #include "llvm/ADT/DenseMap.h" |
24 | #include "llvm/ADT/STLExtras.h" |
25 | #include "llvm/ADT/StringRef.h" |
26 | #include "llvm/IR/Module.h" |
27 | #include "llvm/IR/ProfileSummary.h" |
28 | #include "llvm/ProfileData/ProfileCommon.h" |
29 | #include "llvm/ProfileData/SampleProf.h" |
30 | #include "llvm/Support/CommandLine.h" |
31 | #include "llvm/Support/Compression.h" |
32 | #include "llvm/Support/ErrorOr.h" |
33 | #include "llvm/Support/JSON.h" |
34 | #include "llvm/Support/LEB128.h" |
35 | #include "llvm/Support/LineIterator.h" |
36 | #include "llvm/Support/MD5.h" |
37 | #include "llvm/Support/MemoryBuffer.h" |
38 | #include "llvm/Support/VirtualFileSystem.h" |
39 | #include "llvm/Support/raw_ostream.h" |
40 | #include <algorithm> |
41 | #include <cstddef> |
42 | #include <cstdint> |
43 | #include <limits> |
44 | #include <memory> |
45 | #include <system_error> |
46 | #include <vector> |
47 | |
48 | using namespace llvm; |
49 | using namespace sampleprof; |
50 | |
51 | #define DEBUG_TYPE "samplepgo-reader" |
52 | |
53 | // This internal option specifies if the profile uses FS discriminators. |
54 | // It only applies to text, and binary format profiles. |
55 | // For ext-binary format profiles, the flag is set in the summary. |
56 | static cl::opt<bool> ProfileIsFSDisciminator( |
57 | "profile-isfs" , cl::Hidden, cl::init(Val: false), |
58 | cl::desc("Profile uses flow sensitive discriminators" )); |
59 | |
60 | /// Dump the function profile for \p FName. |
61 | /// |
62 | /// \param FContext Name + context of the function to print. |
63 | /// \param OS Stream to emit the output to. |
64 | void SampleProfileReader::dumpFunctionProfile(const FunctionSamples &FS, |
65 | raw_ostream &OS) { |
66 | OS << "Function: " << FS.getContext().toString() << ": " << FS; |
67 | } |
68 | |
69 | /// Dump all the function profiles found on stream \p OS. |
70 | void SampleProfileReader::dump(raw_ostream &OS) { |
71 | std::vector<NameFunctionSamples> V; |
72 | sortFuncProfiles(ProfileMap: Profiles, SortedProfiles&: V); |
73 | for (const auto &I : V) |
74 | dumpFunctionProfile(FS: *I.second, OS); |
75 | } |
76 | |
77 | static void dumpFunctionProfileJson(const FunctionSamples &S, |
78 | json::OStream &JOS, bool TopLevel = false) { |
79 | auto DumpBody = [&](const BodySampleMap &BodySamples) { |
80 | for (const auto &I : BodySamples) { |
81 | const LineLocation &Loc = I.first; |
82 | const SampleRecord &Sample = I.second; |
83 | JOS.object(Contents: [&] { |
84 | JOS.attribute(Key: "line" , Contents: Loc.LineOffset); |
85 | if (Loc.Discriminator) |
86 | JOS.attribute(Key: "discriminator" , Contents: Loc.Discriminator); |
87 | JOS.attribute(Key: "samples" , Contents: Sample.getSamples()); |
88 | |
89 | auto CallTargets = Sample.getSortedCallTargets(); |
90 | if (!CallTargets.empty()) { |
91 | JOS.attributeArray(Key: "calls" , Contents: [&] { |
92 | for (const auto &J : CallTargets) { |
93 | JOS.object(Contents: [&] { |
94 | JOS.attribute(Key: "function" , Contents: J.first.str()); |
95 | JOS.attribute(Key: "samples" , Contents: J.second); |
96 | }); |
97 | } |
98 | }); |
99 | } |
100 | }); |
101 | } |
102 | }; |
103 | |
104 | auto DumpCallsiteSamples = [&](const CallsiteSampleMap &CallsiteSamples) { |
105 | for (const auto &I : CallsiteSamples) |
106 | for (const auto &FS : I.second) { |
107 | const LineLocation &Loc = I.first; |
108 | const FunctionSamples &CalleeSamples = FS.second; |
109 | JOS.object(Contents: [&] { |
110 | JOS.attribute(Key: "line" , Contents: Loc.LineOffset); |
111 | if (Loc.Discriminator) |
112 | JOS.attribute(Key: "discriminator" , Contents: Loc.Discriminator); |
113 | JOS.attributeArray( |
114 | Key: "samples" , Contents: [&] { dumpFunctionProfileJson(S: CalleeSamples, JOS); }); |
115 | }); |
116 | } |
117 | }; |
118 | |
119 | JOS.object(Contents: [&] { |
120 | JOS.attribute(Key: "name" , Contents: S.getFunction().str()); |
121 | JOS.attribute(Key: "total" , Contents: S.getTotalSamples()); |
122 | if (TopLevel) |
123 | JOS.attribute(Key: "head" , Contents: S.getHeadSamples()); |
124 | |
125 | const auto &BodySamples = S.getBodySamples(); |
126 | if (!BodySamples.empty()) |
127 | JOS.attributeArray(Key: "body" , Contents: [&] { DumpBody(BodySamples); }); |
128 | |
129 | const auto &CallsiteSamples = S.getCallsiteSamples(); |
130 | if (!CallsiteSamples.empty()) |
131 | JOS.attributeArray(Key: "callsites" , |
132 | Contents: [&] { DumpCallsiteSamples(CallsiteSamples); }); |
133 | }); |
134 | } |
135 | |
136 | /// Dump all the function profiles found on stream \p OS in the JSON format. |
137 | void SampleProfileReader::dumpJson(raw_ostream &OS) { |
138 | std::vector<NameFunctionSamples> V; |
139 | sortFuncProfiles(ProfileMap: Profiles, SortedProfiles&: V); |
140 | json::OStream JOS(OS, 2); |
141 | JOS.arrayBegin(); |
142 | for (const auto &F : V) |
143 | dumpFunctionProfileJson(S: *F.second, JOS, TopLevel: true); |
144 | JOS.arrayEnd(); |
145 | |
146 | // Emit a newline character at the end as json::OStream doesn't emit one. |
147 | OS << "\n" ; |
148 | } |
149 | |
150 | /// Parse \p Input as function head. |
151 | /// |
152 | /// Parse one line of \p Input, and update function name in \p FName, |
153 | /// function's total sample count in \p NumSamples, function's entry |
154 | /// count in \p NumHeadSamples. |
155 | /// |
156 | /// \returns true if parsing is successful. |
157 | static bool ParseHead(const StringRef &Input, StringRef &FName, |
158 | uint64_t &NumSamples, uint64_t &NumHeadSamples) { |
159 | if (Input[0] == ' ') |
160 | return false; |
161 | size_t n2 = Input.rfind(C: ':'); |
162 | size_t n1 = Input.rfind(C: ':', From: n2 - 1); |
163 | FName = Input.substr(Start: 0, N: n1); |
164 | if (Input.substr(Start: n1 + 1, N: n2 - n1 - 1).getAsInteger(Radix: 10, Result&: NumSamples)) |
165 | return false; |
166 | if (Input.substr(Start: n2 + 1).getAsInteger(Radix: 10, Result&: NumHeadSamples)) |
167 | return false; |
168 | return true; |
169 | } |
170 | |
171 | /// Returns true if line offset \p L is legal (only has 16 bits). |
172 | static bool isOffsetLegal(unsigned L) { return (L & 0xffff) == L; } |
173 | |
174 | /// Parse \p Input that contains metadata. |
175 | /// Possible metadata: |
176 | /// - CFG Checksum information: |
177 | /// !CFGChecksum: 12345 |
178 | /// - CFG Checksum information: |
179 | /// !Attributes: 1 |
180 | /// Stores the FunctionHash (a.k.a. CFG Checksum) into \p FunctionHash. |
181 | static bool parseMetadata(const StringRef &Input, uint64_t &FunctionHash, |
182 | uint32_t &Attributes) { |
183 | if (Input.starts_with(Prefix: "!CFGChecksum:" )) { |
184 | StringRef CFGInfo = Input.substr(Start: strlen(s: "!CFGChecksum:" )).trim(); |
185 | return !CFGInfo.getAsInteger(Radix: 10, Result&: FunctionHash); |
186 | } |
187 | |
188 | if (Input.starts_with(Prefix: "!Attributes:" )) { |
189 | StringRef Attrib = Input.substr(Start: strlen(s: "!Attributes:" )).trim(); |
190 | return !Attrib.getAsInteger(Radix: 10, Result&: Attributes); |
191 | } |
192 | |
193 | return false; |
194 | } |
195 | |
196 | enum class LineType { |
197 | CallSiteProfile, |
198 | BodyProfile, |
199 | Metadata, |
200 | }; |
201 | |
202 | /// Parse \p Input as line sample. |
203 | /// |
204 | /// \param Input input line. |
205 | /// \param LineTy Type of this line. |
206 | /// \param Depth the depth of the inline stack. |
207 | /// \param NumSamples total samples of the line/inlined callsite. |
208 | /// \param LineOffset line offset to the start of the function. |
209 | /// \param Discriminator discriminator of the line. |
210 | /// \param TargetCountMap map from indirect call target to count. |
211 | /// \param FunctionHash the function's CFG hash, used by pseudo probe. |
212 | /// |
213 | /// returns true if parsing is successful. |
214 | static bool ParseLine(const StringRef &Input, LineType &LineTy, uint32_t &Depth, |
215 | uint64_t &NumSamples, uint32_t &LineOffset, |
216 | uint32_t &Discriminator, StringRef &CalleeName, |
217 | DenseMap<StringRef, uint64_t> &TargetCountMap, |
218 | uint64_t &FunctionHash, uint32_t &Attributes) { |
219 | for (Depth = 0; Input[Depth] == ' '; Depth++) |
220 | ; |
221 | if (Depth == 0) |
222 | return false; |
223 | |
224 | if (Input[Depth] == '!') { |
225 | LineTy = LineType::Metadata; |
226 | return parseMetadata(Input: Input.substr(Start: Depth), FunctionHash, Attributes); |
227 | } |
228 | |
229 | size_t n1 = Input.find(C: ':'); |
230 | StringRef Loc = Input.substr(Start: Depth, N: n1 - Depth); |
231 | size_t n2 = Loc.find(C: '.'); |
232 | if (n2 == StringRef::npos) { |
233 | if (Loc.getAsInteger(Radix: 10, Result&: LineOffset) || !isOffsetLegal(L: LineOffset)) |
234 | return false; |
235 | Discriminator = 0; |
236 | } else { |
237 | if (Loc.substr(Start: 0, N: n2).getAsInteger(Radix: 10, Result&: LineOffset)) |
238 | return false; |
239 | if (Loc.substr(Start: n2 + 1).getAsInteger(Radix: 10, Result&: Discriminator)) |
240 | return false; |
241 | } |
242 | |
243 | StringRef Rest = Input.substr(Start: n1 + 2); |
244 | if (isDigit(C: Rest[0])) { |
245 | LineTy = LineType::BodyProfile; |
246 | size_t n3 = Rest.find(C: ' '); |
247 | if (n3 == StringRef::npos) { |
248 | if (Rest.getAsInteger(Radix: 10, Result&: NumSamples)) |
249 | return false; |
250 | } else { |
251 | if (Rest.substr(Start: 0, N: n3).getAsInteger(Radix: 10, Result&: NumSamples)) |
252 | return false; |
253 | } |
254 | // Find call targets and their sample counts. |
255 | // Note: In some cases, there are symbols in the profile which are not |
256 | // mangled. To accommodate such cases, use colon + integer pairs as the |
257 | // anchor points. |
258 | // An example: |
259 | // _M_construct<char *>:1000 string_view<std::allocator<char> >:437 |
260 | // ":1000" and ":437" are used as anchor points so the string above will |
261 | // be interpreted as |
262 | // target: _M_construct<char *> |
263 | // count: 1000 |
264 | // target: string_view<std::allocator<char> > |
265 | // count: 437 |
266 | while (n3 != StringRef::npos) { |
267 | n3 += Rest.substr(Start: n3).find_first_not_of(C: ' '); |
268 | Rest = Rest.substr(Start: n3); |
269 | n3 = Rest.find_first_of(C: ':'); |
270 | if (n3 == StringRef::npos || n3 == 0) |
271 | return false; |
272 | |
273 | StringRef Target; |
274 | uint64_t count, n4; |
275 | while (true) { |
276 | // Get the segment after the current colon. |
277 | StringRef AfterColon = Rest.substr(Start: n3 + 1); |
278 | // Get the target symbol before the current colon. |
279 | Target = Rest.substr(Start: 0, N: n3); |
280 | // Check if the word after the current colon is an integer. |
281 | n4 = AfterColon.find_first_of(C: ' '); |
282 | n4 = (n4 != StringRef::npos) ? n3 + n4 + 1 : Rest.size(); |
283 | StringRef WordAfterColon = Rest.substr(Start: n3 + 1, N: n4 - n3 - 1); |
284 | if (!WordAfterColon.getAsInteger(Radix: 10, Result&: count)) |
285 | break; |
286 | |
287 | // Try to find the next colon. |
288 | uint64_t n5 = AfterColon.find_first_of(C: ':'); |
289 | if (n5 == StringRef::npos) |
290 | return false; |
291 | n3 += n5 + 1; |
292 | } |
293 | |
294 | // An anchor point is found. Save the {target, count} pair |
295 | TargetCountMap[Target] = count; |
296 | if (n4 == Rest.size()) |
297 | break; |
298 | // Change n3 to the next blank space after colon + integer pair. |
299 | n3 = n4; |
300 | } |
301 | } else { |
302 | LineTy = LineType::CallSiteProfile; |
303 | size_t n3 = Rest.find_last_of(C: ':'); |
304 | CalleeName = Rest.substr(Start: 0, N: n3); |
305 | if (Rest.substr(Start: n3 + 1).getAsInteger(Radix: 10, Result&: NumSamples)) |
306 | return false; |
307 | } |
308 | return true; |
309 | } |
310 | |
311 | /// Load samples from a text file. |
312 | /// |
313 | /// See the documentation at the top of the file for an explanation of |
314 | /// the expected format. |
315 | /// |
316 | /// \returns true if the file was loaded successfully, false otherwise. |
317 | std::error_code SampleProfileReaderText::readImpl() { |
318 | line_iterator LineIt(*Buffer, /*SkipBlanks=*/true, '#'); |
319 | sampleprof_error Result = sampleprof_error::success; |
320 | |
321 | InlineCallStack InlineStack; |
322 | uint32_t TopLevelProbeProfileCount = 0; |
323 | |
324 | // DepthMetadata tracks whether we have processed metadata for the current |
325 | // top-level or nested function profile. |
326 | uint32_t DepthMetadata = 0; |
327 | |
328 | ProfileIsFS = ProfileIsFSDisciminator; |
329 | FunctionSamples::ProfileIsFS = ProfileIsFS; |
330 | for (; !LineIt.is_at_eof(); ++LineIt) { |
331 | size_t pos = LineIt->find_first_not_of(C: ' '); |
332 | if (pos == LineIt->npos || (*LineIt)[pos] == '#') |
333 | continue; |
334 | // Read the header of each function. |
335 | // |
336 | // Note that for function identifiers we are actually expecting |
337 | // mangled names, but we may not always get them. This happens when |
338 | // the compiler decides not to emit the function (e.g., it was inlined |
339 | // and removed). In this case, the binary will not have the linkage |
340 | // name for the function, so the profiler will emit the function's |
341 | // unmangled name, which may contain characters like ':' and '>' in its |
342 | // name (member functions, templates, etc). |
343 | // |
344 | // The only requirement we place on the identifier, then, is that it |
345 | // should not begin with a number. |
346 | if ((*LineIt)[0] != ' ') { |
347 | uint64_t NumSamples, NumHeadSamples; |
348 | StringRef FName; |
349 | if (!ParseHead(Input: *LineIt, FName, NumSamples, NumHeadSamples)) { |
350 | reportError(LineNumber: LineIt.line_number(), |
351 | Msg: "Expected 'mangled_name:NUM:NUM', found " + *LineIt); |
352 | return sampleprof_error::malformed; |
353 | } |
354 | DepthMetadata = 0; |
355 | SampleContext FContext(FName, CSNameTable); |
356 | if (FContext.hasContext()) |
357 | ++CSProfileCount; |
358 | FunctionSamples &FProfile = Profiles.create(Ctx: FContext); |
359 | mergeSampleProfErrors(Accumulator&: Result, Result: FProfile.addTotalSamples(Num: NumSamples)); |
360 | mergeSampleProfErrors(Accumulator&: Result, Result: FProfile.addHeadSamples(Num: NumHeadSamples)); |
361 | InlineStack.clear(); |
362 | InlineStack.push_back(Elt: &FProfile); |
363 | } else { |
364 | uint64_t NumSamples; |
365 | StringRef FName; |
366 | DenseMap<StringRef, uint64_t> TargetCountMap; |
367 | uint32_t Depth, LineOffset, Discriminator; |
368 | LineType LineTy; |
369 | uint64_t FunctionHash = 0; |
370 | uint32_t Attributes = 0; |
371 | if (!ParseLine(Input: *LineIt, LineTy, Depth, NumSamples, LineOffset, |
372 | Discriminator, CalleeName&: FName, TargetCountMap, FunctionHash, |
373 | Attributes)) { |
374 | reportError(LineNumber: LineIt.line_number(), |
375 | Msg: "Expected 'NUM[.NUM]: NUM[ mangled_name:NUM]*', found " + |
376 | *LineIt); |
377 | return sampleprof_error::malformed; |
378 | } |
379 | if (LineTy != LineType::Metadata && Depth == DepthMetadata) { |
380 | // Metadata must be put at the end of a function profile. |
381 | reportError(LineNumber: LineIt.line_number(), |
382 | Msg: "Found non-metadata after metadata: " + *LineIt); |
383 | return sampleprof_error::malformed; |
384 | } |
385 | |
386 | // Here we handle FS discriminators. |
387 | Discriminator &= getDiscriminatorMask(); |
388 | |
389 | while (InlineStack.size() > Depth) { |
390 | InlineStack.pop_back(); |
391 | } |
392 | switch (LineTy) { |
393 | case LineType::CallSiteProfile: { |
394 | FunctionSamples &FSamples = InlineStack.back()->functionSamplesAt( |
395 | Loc: LineLocation(LineOffset, Discriminator))[FunctionId(FName)]; |
396 | FSamples.setFunction(FunctionId(FName)); |
397 | mergeSampleProfErrors(Accumulator&: Result, Result: FSamples.addTotalSamples(Num: NumSamples)); |
398 | InlineStack.push_back(Elt: &FSamples); |
399 | DepthMetadata = 0; |
400 | break; |
401 | } |
402 | case LineType::BodyProfile: { |
403 | while (InlineStack.size() > Depth) { |
404 | InlineStack.pop_back(); |
405 | } |
406 | FunctionSamples &FProfile = *InlineStack.back(); |
407 | for (const auto &name_count : TargetCountMap) { |
408 | mergeSampleProfErrors(Accumulator&: Result, Result: FProfile.addCalledTargetSamples( |
409 | LineOffset, Discriminator, |
410 | Func: FunctionId(name_count.first), |
411 | Num: name_count.second)); |
412 | } |
413 | mergeSampleProfErrors( |
414 | Accumulator&: Result, |
415 | Result: FProfile.addBodySamples(LineOffset, Discriminator, Num: NumSamples)); |
416 | break; |
417 | } |
418 | case LineType::Metadata: { |
419 | FunctionSamples &FProfile = *InlineStack.back(); |
420 | if (FunctionHash) { |
421 | FProfile.setFunctionHash(FunctionHash); |
422 | if (Depth == 1) |
423 | ++TopLevelProbeProfileCount; |
424 | } |
425 | FProfile.getContext().setAllAttributes(Attributes); |
426 | if (Attributes & (uint32_t)ContextShouldBeInlined) |
427 | ProfileIsPreInlined = true; |
428 | DepthMetadata = Depth; |
429 | break; |
430 | } |
431 | } |
432 | } |
433 | } |
434 | |
435 | assert((CSProfileCount == 0 || CSProfileCount == Profiles.size()) && |
436 | "Cannot have both context-sensitive and regular profile" ); |
437 | ProfileIsCS = (CSProfileCount > 0); |
438 | assert((TopLevelProbeProfileCount == 0 || |
439 | TopLevelProbeProfileCount == Profiles.size()) && |
440 | "Cannot have both probe-based profiles and regular profiles" ); |
441 | ProfileIsProbeBased = (TopLevelProbeProfileCount > 0); |
442 | FunctionSamples::ProfileIsProbeBased = ProfileIsProbeBased; |
443 | FunctionSamples::ProfileIsCS = ProfileIsCS; |
444 | FunctionSamples::ProfileIsPreInlined = ProfileIsPreInlined; |
445 | |
446 | if (Result == sampleprof_error::success) |
447 | computeSummary(); |
448 | |
449 | return Result; |
450 | } |
451 | |
452 | bool SampleProfileReaderText::hasFormat(const MemoryBuffer &Buffer) { |
453 | bool result = false; |
454 | |
455 | // Check that the first non-comment line is a valid function header. |
456 | line_iterator LineIt(Buffer, /*SkipBlanks=*/true, '#'); |
457 | if (!LineIt.is_at_eof()) { |
458 | if ((*LineIt)[0] != ' ') { |
459 | uint64_t NumSamples, NumHeadSamples; |
460 | StringRef FName; |
461 | result = ParseHead(Input: *LineIt, FName, NumSamples, NumHeadSamples); |
462 | } |
463 | } |
464 | |
465 | return result; |
466 | } |
467 | |
468 | template <typename T> ErrorOr<T> SampleProfileReaderBinary::readNumber() { |
469 | unsigned NumBytesRead = 0; |
470 | uint64_t Val = decodeULEB128(p: Data, n: &NumBytesRead); |
471 | |
472 | if (Val > std::numeric_limits<T>::max()) { |
473 | std::error_code EC = sampleprof_error::malformed; |
474 | reportError(LineNumber: 0, Msg: EC.message()); |
475 | return EC; |
476 | } else if (Data + NumBytesRead > End) { |
477 | std::error_code EC = sampleprof_error::truncated; |
478 | reportError(LineNumber: 0, Msg: EC.message()); |
479 | return EC; |
480 | } |
481 | |
482 | Data += NumBytesRead; |
483 | return static_cast<T>(Val); |
484 | } |
485 | |
486 | ErrorOr<StringRef> SampleProfileReaderBinary::readString() { |
487 | StringRef Str(reinterpret_cast<const char *>(Data)); |
488 | if (Data + Str.size() + 1 > End) { |
489 | std::error_code EC = sampleprof_error::truncated; |
490 | reportError(LineNumber: 0, Msg: EC.message()); |
491 | return EC; |
492 | } |
493 | |
494 | Data += Str.size() + 1; |
495 | return Str; |
496 | } |
497 | |
498 | template <typename T> |
499 | ErrorOr<T> SampleProfileReaderBinary::readUnencodedNumber() { |
500 | if (Data + sizeof(T) > End) { |
501 | std::error_code EC = sampleprof_error::truncated; |
502 | reportError(LineNumber: 0, Msg: EC.message()); |
503 | return EC; |
504 | } |
505 | |
506 | using namespace support; |
507 | T Val = endian::readNext<T, llvm::endianness::little>(Data); |
508 | return Val; |
509 | } |
510 | |
511 | template <typename T> |
512 | inline ErrorOr<size_t> SampleProfileReaderBinary::readStringIndex(T &Table) { |
513 | auto Idx = readNumber<size_t>(); |
514 | if (std::error_code EC = Idx.getError()) |
515 | return EC; |
516 | if (*Idx >= Table.size()) |
517 | return sampleprof_error::truncated_name_table; |
518 | return *Idx; |
519 | } |
520 | |
521 | ErrorOr<FunctionId> |
522 | SampleProfileReaderBinary::readStringFromTable(size_t *RetIdx) { |
523 | auto Idx = readStringIndex(Table&: NameTable); |
524 | if (std::error_code EC = Idx.getError()) |
525 | return EC; |
526 | if (RetIdx) |
527 | *RetIdx = *Idx; |
528 | return NameTable[*Idx]; |
529 | } |
530 | |
531 | ErrorOr<SampleContextFrames> |
532 | SampleProfileReaderBinary::readContextFromTable(size_t *RetIdx) { |
533 | auto ContextIdx = readNumber<size_t>(); |
534 | if (std::error_code EC = ContextIdx.getError()) |
535 | return EC; |
536 | if (*ContextIdx >= CSNameTable.size()) |
537 | return sampleprof_error::truncated_name_table; |
538 | if (RetIdx) |
539 | *RetIdx = *ContextIdx; |
540 | return CSNameTable[*ContextIdx]; |
541 | } |
542 | |
543 | ErrorOr<std::pair<SampleContext, uint64_t>> |
544 | SampleProfileReaderBinary::readSampleContextFromTable() { |
545 | SampleContext Context; |
546 | size_t Idx; |
547 | if (ProfileIsCS) { |
548 | auto FContext(readContextFromTable(RetIdx: &Idx)); |
549 | if (std::error_code EC = FContext.getError()) |
550 | return EC; |
551 | Context = SampleContext(*FContext); |
552 | } else { |
553 | auto FName(readStringFromTable(RetIdx: &Idx)); |
554 | if (std::error_code EC = FName.getError()) |
555 | return EC; |
556 | Context = SampleContext(*FName); |
557 | } |
558 | // Since MD5SampleContextStart may point to the profile's file data, need to |
559 | // make sure it is reading the same value on big endian CPU. |
560 | uint64_t Hash = support::endian::read64le(P: MD5SampleContextStart + Idx); |
561 | // Lazy computing of hash value, write back to the table to cache it. Only |
562 | // compute the context's hash value if it is being referenced for the first |
563 | // time. |
564 | if (Hash == 0) { |
565 | assert(MD5SampleContextStart == MD5SampleContextTable.data()); |
566 | Hash = Context.getHashCode(); |
567 | support::endian::write64le(P: &MD5SampleContextTable[Idx], V: Hash); |
568 | } |
569 | return std::make_pair(x&: Context, y&: Hash); |
570 | } |
571 | |
572 | std::error_code |
573 | SampleProfileReaderBinary::readProfile(FunctionSamples &FProfile) { |
574 | auto NumSamples = readNumber<uint64_t>(); |
575 | if (std::error_code EC = NumSamples.getError()) |
576 | return EC; |
577 | FProfile.addTotalSamples(Num: *NumSamples); |
578 | |
579 | // Read the samples in the body. |
580 | auto NumRecords = readNumber<uint32_t>(); |
581 | if (std::error_code EC = NumRecords.getError()) |
582 | return EC; |
583 | |
584 | for (uint32_t I = 0; I < *NumRecords; ++I) { |
585 | auto LineOffset = readNumber<uint64_t>(); |
586 | if (std::error_code EC = LineOffset.getError()) |
587 | return EC; |
588 | |
589 | if (!isOffsetLegal(L: *LineOffset)) { |
590 | return std::error_code(); |
591 | } |
592 | |
593 | auto Discriminator = readNumber<uint64_t>(); |
594 | if (std::error_code EC = Discriminator.getError()) |
595 | return EC; |
596 | |
597 | auto NumSamples = readNumber<uint64_t>(); |
598 | if (std::error_code EC = NumSamples.getError()) |
599 | return EC; |
600 | |
601 | auto NumCalls = readNumber<uint32_t>(); |
602 | if (std::error_code EC = NumCalls.getError()) |
603 | return EC; |
604 | |
605 | // Here we handle FS discriminators: |
606 | uint32_t DiscriminatorVal = (*Discriminator) & getDiscriminatorMask(); |
607 | |
608 | for (uint32_t J = 0; J < *NumCalls; ++J) { |
609 | auto CalledFunction(readStringFromTable()); |
610 | if (std::error_code EC = CalledFunction.getError()) |
611 | return EC; |
612 | |
613 | auto CalledFunctionSamples = readNumber<uint64_t>(); |
614 | if (std::error_code EC = CalledFunctionSamples.getError()) |
615 | return EC; |
616 | |
617 | FProfile.addCalledTargetSamples(LineOffset: *LineOffset, Discriminator: DiscriminatorVal, |
618 | Func: *CalledFunction, Num: *CalledFunctionSamples); |
619 | } |
620 | |
621 | FProfile.addBodySamples(LineOffset: *LineOffset, Discriminator: DiscriminatorVal, Num: *NumSamples); |
622 | } |
623 | |
624 | // Read all the samples for inlined function calls. |
625 | auto NumCallsites = readNumber<uint32_t>(); |
626 | if (std::error_code EC = NumCallsites.getError()) |
627 | return EC; |
628 | |
629 | for (uint32_t J = 0; J < *NumCallsites; ++J) { |
630 | auto LineOffset = readNumber<uint64_t>(); |
631 | if (std::error_code EC = LineOffset.getError()) |
632 | return EC; |
633 | |
634 | auto Discriminator = readNumber<uint64_t>(); |
635 | if (std::error_code EC = Discriminator.getError()) |
636 | return EC; |
637 | |
638 | auto FName(readStringFromTable()); |
639 | if (std::error_code EC = FName.getError()) |
640 | return EC; |
641 | |
642 | // Here we handle FS discriminators: |
643 | uint32_t DiscriminatorVal = (*Discriminator) & getDiscriminatorMask(); |
644 | |
645 | FunctionSamples &CalleeProfile = FProfile.functionSamplesAt( |
646 | Loc: LineLocation(*LineOffset, DiscriminatorVal))[*FName]; |
647 | CalleeProfile.setFunction(*FName); |
648 | if (std::error_code EC = readProfile(FProfile&: CalleeProfile)) |
649 | return EC; |
650 | } |
651 | |
652 | return sampleprof_error::success; |
653 | } |
654 | |
655 | std::error_code |
656 | SampleProfileReaderBinary::readFuncProfile(const uint8_t *Start) { |
657 | Data = Start; |
658 | auto NumHeadSamples = readNumber<uint64_t>(); |
659 | if (std::error_code EC = NumHeadSamples.getError()) |
660 | return EC; |
661 | |
662 | auto FContextHash(readSampleContextFromTable()); |
663 | if (std::error_code EC = FContextHash.getError()) |
664 | return EC; |
665 | |
666 | auto &[FContext, Hash] = *FContextHash; |
667 | // Use the cached hash value for insertion instead of recalculating it. |
668 | auto Res = Profiles.try_emplace(Hash, Key: FContext, Args: FunctionSamples()); |
669 | FunctionSamples &FProfile = Res.first->second; |
670 | FProfile.setContext(FContext); |
671 | FProfile.addHeadSamples(Num: *NumHeadSamples); |
672 | |
673 | if (FContext.hasContext()) |
674 | CSProfileCount++; |
675 | |
676 | if (std::error_code EC = readProfile(FProfile)) |
677 | return EC; |
678 | return sampleprof_error::success; |
679 | } |
680 | |
681 | std::error_code SampleProfileReaderBinary::readImpl() { |
682 | ProfileIsFS = ProfileIsFSDisciminator; |
683 | FunctionSamples::ProfileIsFS = ProfileIsFS; |
684 | while (Data < End) { |
685 | if (std::error_code EC = readFuncProfile(Start: Data)) |
686 | return EC; |
687 | } |
688 | |
689 | return sampleprof_error::success; |
690 | } |
691 | |
692 | std::error_code SampleProfileReaderExtBinaryBase::readOneSection( |
693 | const uint8_t *Start, uint64_t Size, const SecHdrTableEntry &Entry) { |
694 | Data = Start; |
695 | End = Start + Size; |
696 | switch (Entry.Type) { |
697 | case SecProfSummary: |
698 | if (std::error_code EC = readSummary()) |
699 | return EC; |
700 | if (hasSecFlag(Entry, Flag: SecProfSummaryFlags::SecFlagPartial)) |
701 | Summary->setPartialProfile(true); |
702 | if (hasSecFlag(Entry, Flag: SecProfSummaryFlags::SecFlagFullContext)) |
703 | FunctionSamples::ProfileIsCS = ProfileIsCS = true; |
704 | if (hasSecFlag(Entry, Flag: SecProfSummaryFlags::SecFlagIsPreInlined)) |
705 | FunctionSamples::ProfileIsPreInlined = ProfileIsPreInlined = true; |
706 | if (hasSecFlag(Entry, Flag: SecProfSummaryFlags::SecFlagFSDiscriminator)) |
707 | FunctionSamples::ProfileIsFS = ProfileIsFS = true; |
708 | break; |
709 | case SecNameTable: { |
710 | bool FixedLengthMD5 = |
711 | hasSecFlag(Entry, Flag: SecNameTableFlags::SecFlagFixedLengthMD5); |
712 | bool UseMD5 = hasSecFlag(Entry, Flag: SecNameTableFlags::SecFlagMD5Name); |
713 | // UseMD5 means if THIS section uses MD5, ProfileIsMD5 means if the entire |
714 | // profile uses MD5 for function name matching in IPO passes. |
715 | ProfileIsMD5 = ProfileIsMD5 || UseMD5; |
716 | FunctionSamples::HasUniqSuffix = |
717 | hasSecFlag(Entry, Flag: SecNameTableFlags::SecFlagUniqSuffix); |
718 | if (std::error_code EC = readNameTableSec(IsMD5: UseMD5, FixedLengthMD5)) |
719 | return EC; |
720 | break; |
721 | } |
722 | case SecCSNameTable: { |
723 | if (std::error_code EC = readCSNameTableSec()) |
724 | return EC; |
725 | break; |
726 | } |
727 | case SecLBRProfile: |
728 | if (std::error_code EC = readFuncProfiles()) |
729 | return EC; |
730 | break; |
731 | case SecFuncOffsetTable: |
732 | // If module is absent, we are using LLVM tools, and need to read all |
733 | // profiles, so skip reading the function offset table. |
734 | if (!M) { |
735 | Data = End; |
736 | } else { |
737 | assert((!ProfileIsCS || |
738 | hasSecFlag(Entry, SecFuncOffsetFlags::SecFlagOrdered)) && |
739 | "func offset table should always be sorted in CS profile" ); |
740 | if (std::error_code EC = readFuncOffsetTable()) |
741 | return EC; |
742 | } |
743 | break; |
744 | case SecFuncMetadata: { |
745 | ProfileIsProbeBased = |
746 | hasSecFlag(Entry, Flag: SecFuncMetadataFlags::SecFlagIsProbeBased); |
747 | FunctionSamples::ProfileIsProbeBased = ProfileIsProbeBased; |
748 | bool HasAttribute = |
749 | hasSecFlag(Entry, Flag: SecFuncMetadataFlags::SecFlagHasAttribute); |
750 | if (std::error_code EC = readFuncMetadata(ProfileHasAttribute: HasAttribute)) |
751 | return EC; |
752 | break; |
753 | } |
754 | case SecProfileSymbolList: |
755 | if (std::error_code EC = readProfileSymbolList()) |
756 | return EC; |
757 | break; |
758 | default: |
759 | if (std::error_code EC = readCustomSection(Entry)) |
760 | return EC; |
761 | break; |
762 | } |
763 | return sampleprof_error::success; |
764 | } |
765 | |
766 | bool SampleProfileReaderExtBinaryBase::useFuncOffsetList() const { |
767 | // If profile is CS, the function offset section is expected to consist of |
768 | // sequences of contexts in pre-order layout |
769 | // (e.g. [A, A:1 @ B, A:1 @ B:2.3 @ C] [D, D:1 @ E]), so that when a matched |
770 | // context in the module is found, the profiles of all its callees are |
771 | // recursively loaded. A list is needed since the order of profiles matters. |
772 | if (ProfileIsCS) |
773 | return true; |
774 | |
775 | // If the profile is MD5, use the map container to lookup functions in |
776 | // the module. A remapper has no use on MD5 names. |
777 | if (useMD5()) |
778 | return false; |
779 | |
780 | // Profile is not MD5 and if a remapper is present, the remapped name of |
781 | // every function needed to be matched against the module, so use the list |
782 | // container since each entry is accessed. |
783 | if (Remapper) |
784 | return true; |
785 | |
786 | // Otherwise use the map container for faster lookup. |
787 | // TODO: If the cardinality of the function offset section is much smaller |
788 | // than the number of functions in the module, using the list container can |
789 | // be always faster, but we need to figure out the constant factor to |
790 | // determine the cutoff. |
791 | return false; |
792 | } |
793 | |
794 | |
795 | bool SampleProfileReaderExtBinaryBase::collectFuncsFromModule() { |
796 | if (!M) |
797 | return false; |
798 | FuncsToUse.clear(); |
799 | for (auto &F : *M) |
800 | FuncsToUse.insert(V: FunctionSamples::getCanonicalFnName(F)); |
801 | return true; |
802 | } |
803 | |
804 | std::error_code SampleProfileReaderExtBinaryBase::readFuncOffsetTable() { |
805 | // If there are more than one function offset section, the profile associated |
806 | // with the previous section has to be done reading before next one is read. |
807 | FuncOffsetTable.clear(); |
808 | FuncOffsetList.clear(); |
809 | |
810 | auto Size = readNumber<uint64_t>(); |
811 | if (std::error_code EC = Size.getError()) |
812 | return EC; |
813 | |
814 | bool UseFuncOffsetList = useFuncOffsetList(); |
815 | if (UseFuncOffsetList) |
816 | FuncOffsetList.reserve(n: *Size); |
817 | else |
818 | FuncOffsetTable.reserve(NumEntries: *Size); |
819 | |
820 | for (uint64_t I = 0; I < *Size; ++I) { |
821 | auto FContextHash(readSampleContextFromTable()); |
822 | if (std::error_code EC = FContextHash.getError()) |
823 | return EC; |
824 | |
825 | auto &[FContext, Hash] = *FContextHash; |
826 | auto Offset = readNumber<uint64_t>(); |
827 | if (std::error_code EC = Offset.getError()) |
828 | return EC; |
829 | |
830 | if (UseFuncOffsetList) |
831 | FuncOffsetList.emplace_back(args&: FContext, args&: *Offset); |
832 | else |
833 | // Because Porfiles replace existing value with new value if collision |
834 | // happens, we also use the latest offset so that they are consistent. |
835 | FuncOffsetTable[Hash] = *Offset; |
836 | } |
837 | |
838 | return sampleprof_error::success; |
839 | } |
840 | |
841 | std::error_code SampleProfileReaderExtBinaryBase::readFuncProfiles() { |
842 | // Collect functions used by current module if the Reader has been |
843 | // given a module. |
844 | // collectFuncsFromModule uses FunctionSamples::getCanonicalFnName |
845 | // which will query FunctionSamples::HasUniqSuffix, so it has to be |
846 | // called after FunctionSamples::HasUniqSuffix is set, i.e. after |
847 | // NameTable section is read. |
848 | bool LoadFuncsToBeUsed = collectFuncsFromModule(); |
849 | |
850 | // When LoadFuncsToBeUsed is false, we are using LLVM tool, need to read all |
851 | // profiles. |
852 | const uint8_t *Start = Data; |
853 | if (!LoadFuncsToBeUsed) { |
854 | while (Data < End) { |
855 | if (std::error_code EC = readFuncProfile(Start: Data)) |
856 | return EC; |
857 | } |
858 | assert(Data == End && "More data is read than expected" ); |
859 | } else { |
860 | // Load function profiles on demand. |
861 | if (Remapper) { |
862 | for (auto Name : FuncsToUse) { |
863 | Remapper->insert(FunctionName: Name); |
864 | } |
865 | } |
866 | |
867 | if (ProfileIsCS) { |
868 | assert(useFuncOffsetList()); |
869 | DenseSet<uint64_t> FuncGuidsToUse; |
870 | if (useMD5()) { |
871 | for (auto Name : FuncsToUse) |
872 | FuncGuidsToUse.insert(V: Function::getGUID(GlobalName: Name)); |
873 | } |
874 | |
875 | // For each function in current module, load all context profiles for |
876 | // the function as well as their callee contexts which can help profile |
877 | // guided importing for ThinLTO. This can be achieved by walking |
878 | // through an ordered context container, where contexts are laid out |
879 | // as if they were walked in preorder of a context trie. While |
880 | // traversing the trie, a link to the highest common ancestor node is |
881 | // kept so that all of its decendants will be loaded. |
882 | const SampleContext *CommonContext = nullptr; |
883 | for (const auto &NameOffset : FuncOffsetList) { |
884 | const auto &FContext = NameOffset.first; |
885 | FunctionId FName = FContext.getFunction(); |
886 | StringRef FNameString; |
887 | if (!useMD5()) |
888 | FNameString = FName.stringRef(); |
889 | |
890 | // For function in the current module, keep its farthest ancestor |
891 | // context. This can be used to load itself and its child and |
892 | // sibling contexts. |
893 | if ((useMD5() && FuncGuidsToUse.count(V: FName.getHashCode())) || |
894 | (!useMD5() && (FuncsToUse.count(V: FNameString) || |
895 | (Remapper && Remapper->exist(FunctionName: FNameString))))) { |
896 | if (!CommonContext || !CommonContext->isPrefixOf(That: FContext)) |
897 | CommonContext = &FContext; |
898 | } |
899 | |
900 | if (CommonContext == &FContext || |
901 | (CommonContext && CommonContext->isPrefixOf(That: FContext))) { |
902 | // Load profile for the current context which originated from |
903 | // the common ancestor. |
904 | const uint8_t *FuncProfileAddr = Start + NameOffset.second; |
905 | if (std::error_code EC = readFuncProfile(Start: FuncProfileAddr)) |
906 | return EC; |
907 | } |
908 | } |
909 | } else if (useMD5()) { |
910 | assert(!useFuncOffsetList()); |
911 | for (auto Name : FuncsToUse) { |
912 | auto GUID = MD5Hash(Str: Name); |
913 | auto iter = FuncOffsetTable.find(Val: GUID); |
914 | if (iter == FuncOffsetTable.end()) |
915 | continue; |
916 | const uint8_t *FuncProfileAddr = Start + iter->second; |
917 | if (std::error_code EC = readFuncProfile(Start: FuncProfileAddr)) |
918 | return EC; |
919 | } |
920 | } else if (Remapper) { |
921 | assert(useFuncOffsetList()); |
922 | for (auto NameOffset : FuncOffsetList) { |
923 | SampleContext FContext(NameOffset.first); |
924 | auto FuncName = FContext.getFunction(); |
925 | StringRef FuncNameStr = FuncName.stringRef(); |
926 | if (!FuncsToUse.count(V: FuncNameStr) && !Remapper->exist(FunctionName: FuncNameStr)) |
927 | continue; |
928 | const uint8_t *FuncProfileAddr = Start + NameOffset.second; |
929 | if (std::error_code EC = readFuncProfile(Start: FuncProfileAddr)) |
930 | return EC; |
931 | } |
932 | } else { |
933 | assert(!useFuncOffsetList()); |
934 | for (auto Name : FuncsToUse) { |
935 | auto iter = FuncOffsetTable.find(Val: MD5Hash(Str: Name)); |
936 | if (iter == FuncOffsetTable.end()) |
937 | continue; |
938 | const uint8_t *FuncProfileAddr = Start + iter->second; |
939 | if (std::error_code EC = readFuncProfile(Start: FuncProfileAddr)) |
940 | return EC; |
941 | } |
942 | } |
943 | Data = End; |
944 | } |
945 | assert((CSProfileCount == 0 || CSProfileCount == Profiles.size()) && |
946 | "Cannot have both context-sensitive and regular profile" ); |
947 | assert((!CSProfileCount || ProfileIsCS) && |
948 | "Section flag should be consistent with actual profile" ); |
949 | return sampleprof_error::success; |
950 | } |
951 | |
952 | std::error_code SampleProfileReaderExtBinaryBase::readProfileSymbolList() { |
953 | if (!ProfSymList) |
954 | ProfSymList = std::make_unique<ProfileSymbolList>(); |
955 | |
956 | if (std::error_code EC = ProfSymList->read(Data, ListSize: End - Data)) |
957 | return EC; |
958 | |
959 | Data = End; |
960 | return sampleprof_error::success; |
961 | } |
962 | |
963 | std::error_code SampleProfileReaderExtBinaryBase::decompressSection( |
964 | const uint8_t *SecStart, const uint64_t SecSize, |
965 | const uint8_t *&DecompressBuf, uint64_t &DecompressBufSize) { |
966 | Data = SecStart; |
967 | End = SecStart + SecSize; |
968 | auto DecompressSize = readNumber<uint64_t>(); |
969 | if (std::error_code EC = DecompressSize.getError()) |
970 | return EC; |
971 | DecompressBufSize = *DecompressSize; |
972 | |
973 | auto CompressSize = readNumber<uint64_t>(); |
974 | if (std::error_code EC = CompressSize.getError()) |
975 | return EC; |
976 | |
977 | if (!llvm::compression::zlib::isAvailable()) |
978 | return sampleprof_error::zlib_unavailable; |
979 | |
980 | uint8_t *Buffer = Allocator.Allocate<uint8_t>(Num: DecompressBufSize); |
981 | size_t UCSize = DecompressBufSize; |
982 | llvm::Error E = compression::zlib::decompress(Input: ArrayRef(Data, *CompressSize), |
983 | Output: Buffer, UncompressedSize&: UCSize); |
984 | if (E) |
985 | return sampleprof_error::uncompress_failed; |
986 | DecompressBuf = reinterpret_cast<const uint8_t *>(Buffer); |
987 | return sampleprof_error::success; |
988 | } |
989 | |
990 | std::error_code SampleProfileReaderExtBinaryBase::readImpl() { |
991 | const uint8_t *BufStart = |
992 | reinterpret_cast<const uint8_t *>(Buffer->getBufferStart()); |
993 | |
994 | for (auto &Entry : SecHdrTable) { |
995 | // Skip empty section. |
996 | if (!Entry.Size) |
997 | continue; |
998 | |
999 | // Skip sections without context when SkipFlatProf is true. |
1000 | if (SkipFlatProf && hasSecFlag(Entry, Flag: SecCommonFlags::SecFlagFlat)) |
1001 | continue; |
1002 | |
1003 | const uint8_t *SecStart = BufStart + Entry.Offset; |
1004 | uint64_t SecSize = Entry.Size; |
1005 | |
1006 | // If the section is compressed, decompress it into a buffer |
1007 | // DecompressBuf before reading the actual data. The pointee of |
1008 | // 'Data' will be changed to buffer hold by DecompressBuf |
1009 | // temporarily when reading the actual data. |
1010 | bool isCompressed = hasSecFlag(Entry, Flag: SecCommonFlags::SecFlagCompress); |
1011 | if (isCompressed) { |
1012 | const uint8_t *DecompressBuf; |
1013 | uint64_t DecompressBufSize; |
1014 | if (std::error_code EC = decompressSection( |
1015 | SecStart, SecSize, DecompressBuf, DecompressBufSize)) |
1016 | return EC; |
1017 | SecStart = DecompressBuf; |
1018 | SecSize = DecompressBufSize; |
1019 | } |
1020 | |
1021 | if (std::error_code EC = readOneSection(Start: SecStart, Size: SecSize, Entry)) |
1022 | return EC; |
1023 | if (Data != SecStart + SecSize) |
1024 | return sampleprof_error::malformed; |
1025 | |
1026 | // Change the pointee of 'Data' from DecompressBuf to original Buffer. |
1027 | if (isCompressed) { |
1028 | Data = BufStart + Entry.Offset; |
1029 | End = BufStart + Buffer->getBufferSize(); |
1030 | } |
1031 | } |
1032 | |
1033 | return sampleprof_error::success; |
1034 | } |
1035 | |
1036 | std::error_code SampleProfileReaderRawBinary::verifySPMagic(uint64_t Magic) { |
1037 | if (Magic == SPMagic()) |
1038 | return sampleprof_error::success; |
1039 | return sampleprof_error::bad_magic; |
1040 | } |
1041 | |
1042 | std::error_code SampleProfileReaderExtBinary::verifySPMagic(uint64_t Magic) { |
1043 | if (Magic == SPMagic(Format: SPF_Ext_Binary)) |
1044 | return sampleprof_error::success; |
1045 | return sampleprof_error::bad_magic; |
1046 | } |
1047 | |
1048 | std::error_code SampleProfileReaderBinary::readNameTable() { |
1049 | auto Size = readNumber<size_t>(); |
1050 | if (std::error_code EC = Size.getError()) |
1051 | return EC; |
1052 | |
1053 | // Normally if useMD5 is true, the name table should have MD5 values, not |
1054 | // strings, however in the case that ExtBinary profile has multiple name |
1055 | // tables mixing string and MD5, all of them have to be normalized to use MD5, |
1056 | // because optimization passes can only handle either type. |
1057 | bool UseMD5 = useMD5(); |
1058 | |
1059 | NameTable.clear(); |
1060 | NameTable.reserve(n: *Size); |
1061 | if (!ProfileIsCS) { |
1062 | MD5SampleContextTable.clear(); |
1063 | if (UseMD5) |
1064 | MD5SampleContextTable.reserve(n: *Size); |
1065 | else |
1066 | // If we are using strings, delay MD5 computation since only a portion of |
1067 | // names are used by top level functions. Use 0 to indicate MD5 value is |
1068 | // to be calculated as no known string has a MD5 value of 0. |
1069 | MD5SampleContextTable.resize(new_size: *Size); |
1070 | } |
1071 | for (size_t I = 0; I < *Size; ++I) { |
1072 | auto Name(readString()); |
1073 | if (std::error_code EC = Name.getError()) |
1074 | return EC; |
1075 | if (UseMD5) { |
1076 | FunctionId FID(*Name); |
1077 | if (!ProfileIsCS) |
1078 | MD5SampleContextTable.emplace_back(args: FID.getHashCode()); |
1079 | NameTable.emplace_back(args&: FID); |
1080 | } else |
1081 | NameTable.push_back(x: FunctionId(*Name)); |
1082 | } |
1083 | if (!ProfileIsCS) |
1084 | MD5SampleContextStart = MD5SampleContextTable.data(); |
1085 | return sampleprof_error::success; |
1086 | } |
1087 | |
1088 | std::error_code |
1089 | SampleProfileReaderExtBinaryBase::readNameTableSec(bool IsMD5, |
1090 | bool FixedLengthMD5) { |
1091 | if (FixedLengthMD5) { |
1092 | if (!IsMD5) |
1093 | errs() << "If FixedLengthMD5 is true, UseMD5 has to be true" ; |
1094 | auto Size = readNumber<size_t>(); |
1095 | if (std::error_code EC = Size.getError()) |
1096 | return EC; |
1097 | |
1098 | assert(Data + (*Size) * sizeof(uint64_t) == End && |
1099 | "Fixed length MD5 name table does not contain specified number of " |
1100 | "entries" ); |
1101 | if (Data + (*Size) * sizeof(uint64_t) > End) |
1102 | return sampleprof_error::truncated; |
1103 | |
1104 | NameTable.clear(); |
1105 | NameTable.reserve(n: *Size); |
1106 | for (size_t I = 0; I < *Size; ++I) { |
1107 | using namespace support; |
1108 | uint64_t FID = endian::read<uint64_t, endianness::little, unaligned>( |
1109 | memory: Data + I * sizeof(uint64_t)); |
1110 | NameTable.emplace_back(args: FunctionId(FID)); |
1111 | } |
1112 | if (!ProfileIsCS) |
1113 | MD5SampleContextStart = reinterpret_cast<const uint64_t *>(Data); |
1114 | Data = Data + (*Size) * sizeof(uint64_t); |
1115 | return sampleprof_error::success; |
1116 | } |
1117 | |
1118 | if (IsMD5) { |
1119 | assert(!FixedLengthMD5 && "FixedLengthMD5 should be unreachable here" ); |
1120 | auto Size = readNumber<size_t>(); |
1121 | if (std::error_code EC = Size.getError()) |
1122 | return EC; |
1123 | |
1124 | NameTable.clear(); |
1125 | NameTable.reserve(n: *Size); |
1126 | if (!ProfileIsCS) |
1127 | MD5SampleContextTable.resize(new_size: *Size); |
1128 | for (size_t I = 0; I < *Size; ++I) { |
1129 | auto FID = readNumber<uint64_t>(); |
1130 | if (std::error_code EC = FID.getError()) |
1131 | return EC; |
1132 | if (!ProfileIsCS) |
1133 | support::endian::write64le(P: &MD5SampleContextTable[I], V: *FID); |
1134 | NameTable.emplace_back(args: FunctionId(*FID)); |
1135 | } |
1136 | if (!ProfileIsCS) |
1137 | MD5SampleContextStart = MD5SampleContextTable.data(); |
1138 | return sampleprof_error::success; |
1139 | } |
1140 | |
1141 | return SampleProfileReaderBinary::readNameTable(); |
1142 | } |
1143 | |
1144 | // Read in the CS name table section, which basically contains a list of context |
1145 | // vectors. Each element of a context vector, aka a frame, refers to the |
1146 | // underlying raw function names that are stored in the name table, as well as |
1147 | // a callsite identifier that only makes sense for non-leaf frames. |
1148 | std::error_code SampleProfileReaderExtBinaryBase::readCSNameTableSec() { |
1149 | auto Size = readNumber<size_t>(); |
1150 | if (std::error_code EC = Size.getError()) |
1151 | return EC; |
1152 | |
1153 | CSNameTable.clear(); |
1154 | CSNameTable.reserve(n: *Size); |
1155 | if (ProfileIsCS) { |
1156 | // Delay MD5 computation of CS context until they are needed. Use 0 to |
1157 | // indicate MD5 value is to be calculated as no known string has a MD5 |
1158 | // value of 0. |
1159 | MD5SampleContextTable.clear(); |
1160 | MD5SampleContextTable.resize(new_size: *Size); |
1161 | MD5SampleContextStart = MD5SampleContextTable.data(); |
1162 | } |
1163 | for (size_t I = 0; I < *Size; ++I) { |
1164 | CSNameTable.emplace_back(args: SampleContextFrameVector()); |
1165 | auto ContextSize = readNumber<uint32_t>(); |
1166 | if (std::error_code EC = ContextSize.getError()) |
1167 | return EC; |
1168 | for (uint32_t J = 0; J < *ContextSize; ++J) { |
1169 | auto FName(readStringFromTable()); |
1170 | if (std::error_code EC = FName.getError()) |
1171 | return EC; |
1172 | auto LineOffset = readNumber<uint64_t>(); |
1173 | if (std::error_code EC = LineOffset.getError()) |
1174 | return EC; |
1175 | |
1176 | if (!isOffsetLegal(L: *LineOffset)) |
1177 | return std::error_code(); |
1178 | |
1179 | auto Discriminator = readNumber<uint64_t>(); |
1180 | if (std::error_code EC = Discriminator.getError()) |
1181 | return EC; |
1182 | |
1183 | CSNameTable.back().emplace_back( |
1184 | Args&: FName.get(), Args: LineLocation(LineOffset.get(), Discriminator.get())); |
1185 | } |
1186 | } |
1187 | |
1188 | return sampleprof_error::success; |
1189 | } |
1190 | |
1191 | std::error_code |
1192 | SampleProfileReaderExtBinaryBase::readFuncMetadata(bool ProfileHasAttribute, |
1193 | FunctionSamples *FProfile) { |
1194 | if (Data < End) { |
1195 | if (ProfileIsProbeBased) { |
1196 | auto Checksum = readNumber<uint64_t>(); |
1197 | if (std::error_code EC = Checksum.getError()) |
1198 | return EC; |
1199 | if (FProfile) |
1200 | FProfile->setFunctionHash(*Checksum); |
1201 | } |
1202 | |
1203 | if (ProfileHasAttribute) { |
1204 | auto Attributes = readNumber<uint32_t>(); |
1205 | if (std::error_code EC = Attributes.getError()) |
1206 | return EC; |
1207 | if (FProfile) |
1208 | FProfile->getContext().setAllAttributes(*Attributes); |
1209 | } |
1210 | |
1211 | if (!ProfileIsCS) { |
1212 | // Read all the attributes for inlined function calls. |
1213 | auto NumCallsites = readNumber<uint32_t>(); |
1214 | if (std::error_code EC = NumCallsites.getError()) |
1215 | return EC; |
1216 | |
1217 | for (uint32_t J = 0; J < *NumCallsites; ++J) { |
1218 | auto LineOffset = readNumber<uint64_t>(); |
1219 | if (std::error_code EC = LineOffset.getError()) |
1220 | return EC; |
1221 | |
1222 | auto Discriminator = readNumber<uint64_t>(); |
1223 | if (std::error_code EC = Discriminator.getError()) |
1224 | return EC; |
1225 | |
1226 | auto FContextHash(readSampleContextFromTable()); |
1227 | if (std::error_code EC = FContextHash.getError()) |
1228 | return EC; |
1229 | |
1230 | auto &[FContext, Hash] = *FContextHash; |
1231 | FunctionSamples *CalleeProfile = nullptr; |
1232 | if (FProfile) { |
1233 | CalleeProfile = const_cast<FunctionSamples *>( |
1234 | &FProfile->functionSamplesAt(Loc: LineLocation( |
1235 | *LineOffset, |
1236 | *Discriminator))[FContext.getFunction()]); |
1237 | } |
1238 | if (std::error_code EC = |
1239 | readFuncMetadata(ProfileHasAttribute, FProfile: CalleeProfile)) |
1240 | return EC; |
1241 | } |
1242 | } |
1243 | } |
1244 | |
1245 | return sampleprof_error::success; |
1246 | } |
1247 | |
1248 | std::error_code |
1249 | SampleProfileReaderExtBinaryBase::readFuncMetadata(bool ProfileHasAttribute) { |
1250 | while (Data < End) { |
1251 | auto FContextHash(readSampleContextFromTable()); |
1252 | if (std::error_code EC = FContextHash.getError()) |
1253 | return EC; |
1254 | auto &[FContext, Hash] = *FContextHash; |
1255 | FunctionSamples *FProfile = nullptr; |
1256 | auto It = Profiles.find(Ctx: FContext); |
1257 | if (It != Profiles.end()) |
1258 | FProfile = &It->second; |
1259 | |
1260 | if (std::error_code EC = readFuncMetadata(ProfileHasAttribute, FProfile)) |
1261 | return EC; |
1262 | } |
1263 | |
1264 | assert(Data == End && "More data is read than expected" ); |
1265 | return sampleprof_error::success; |
1266 | } |
1267 | |
1268 | std::error_code |
1269 | SampleProfileReaderExtBinaryBase::readSecHdrTableEntry(uint64_t Idx) { |
1270 | SecHdrTableEntry Entry; |
1271 | auto Type = readUnencodedNumber<uint64_t>(); |
1272 | if (std::error_code EC = Type.getError()) |
1273 | return EC; |
1274 | Entry.Type = static_cast<SecType>(*Type); |
1275 | |
1276 | auto Flags = readUnencodedNumber<uint64_t>(); |
1277 | if (std::error_code EC = Flags.getError()) |
1278 | return EC; |
1279 | Entry.Flags = *Flags; |
1280 | |
1281 | auto Offset = readUnencodedNumber<uint64_t>(); |
1282 | if (std::error_code EC = Offset.getError()) |
1283 | return EC; |
1284 | Entry.Offset = *Offset; |
1285 | |
1286 | auto Size = readUnencodedNumber<uint64_t>(); |
1287 | if (std::error_code EC = Size.getError()) |
1288 | return EC; |
1289 | Entry.Size = *Size; |
1290 | |
1291 | Entry.LayoutIndex = Idx; |
1292 | SecHdrTable.push_back(x: std::move(Entry)); |
1293 | return sampleprof_error::success; |
1294 | } |
1295 | |
1296 | std::error_code SampleProfileReaderExtBinaryBase::readSecHdrTable() { |
1297 | auto EntryNum = readUnencodedNumber<uint64_t>(); |
1298 | if (std::error_code EC = EntryNum.getError()) |
1299 | return EC; |
1300 | |
1301 | for (uint64_t i = 0; i < (*EntryNum); i++) |
1302 | if (std::error_code EC = readSecHdrTableEntry(Idx: i)) |
1303 | return EC; |
1304 | |
1305 | return sampleprof_error::success; |
1306 | } |
1307 | |
1308 | std::error_code SampleProfileReaderExtBinaryBase::() { |
1309 | const uint8_t *BufStart = |
1310 | reinterpret_cast<const uint8_t *>(Buffer->getBufferStart()); |
1311 | Data = BufStart; |
1312 | End = BufStart + Buffer->getBufferSize(); |
1313 | |
1314 | if (std::error_code EC = readMagicIdent()) |
1315 | return EC; |
1316 | |
1317 | if (std::error_code EC = readSecHdrTable()) |
1318 | return EC; |
1319 | |
1320 | return sampleprof_error::success; |
1321 | } |
1322 | |
1323 | uint64_t SampleProfileReaderExtBinaryBase::getSectionSize(SecType Type) { |
1324 | uint64_t Size = 0; |
1325 | for (auto &Entry : SecHdrTable) { |
1326 | if (Entry.Type == Type) |
1327 | Size += Entry.Size; |
1328 | } |
1329 | return Size; |
1330 | } |
1331 | |
1332 | uint64_t SampleProfileReaderExtBinaryBase::getFileSize() { |
1333 | // Sections in SecHdrTable is not necessarily in the same order as |
1334 | // sections in the profile because section like FuncOffsetTable needs |
1335 | // to be written after section LBRProfile but needs to be read before |
1336 | // section LBRProfile, so we cannot simply use the last entry in |
1337 | // SecHdrTable to calculate the file size. |
1338 | uint64_t FileSize = 0; |
1339 | for (auto &Entry : SecHdrTable) { |
1340 | FileSize = std::max(a: Entry.Offset + Entry.Size, b: FileSize); |
1341 | } |
1342 | return FileSize; |
1343 | } |
1344 | |
1345 | static std::string getSecFlagsStr(const SecHdrTableEntry &Entry) { |
1346 | std::string Flags; |
1347 | if (hasSecFlag(Entry, Flag: SecCommonFlags::SecFlagCompress)) |
1348 | Flags.append(s: "{compressed," ); |
1349 | else |
1350 | Flags.append(s: "{" ); |
1351 | |
1352 | if (hasSecFlag(Entry, Flag: SecCommonFlags::SecFlagFlat)) |
1353 | Flags.append(s: "flat," ); |
1354 | |
1355 | switch (Entry.Type) { |
1356 | case SecNameTable: |
1357 | if (hasSecFlag(Entry, Flag: SecNameTableFlags::SecFlagFixedLengthMD5)) |
1358 | Flags.append(s: "fixlenmd5," ); |
1359 | else if (hasSecFlag(Entry, Flag: SecNameTableFlags::SecFlagMD5Name)) |
1360 | Flags.append(s: "md5," ); |
1361 | if (hasSecFlag(Entry, Flag: SecNameTableFlags::SecFlagUniqSuffix)) |
1362 | Flags.append(s: "uniq," ); |
1363 | break; |
1364 | case SecProfSummary: |
1365 | if (hasSecFlag(Entry, Flag: SecProfSummaryFlags::SecFlagPartial)) |
1366 | Flags.append(s: "partial," ); |
1367 | if (hasSecFlag(Entry, Flag: SecProfSummaryFlags::SecFlagFullContext)) |
1368 | Flags.append(s: "context," ); |
1369 | if (hasSecFlag(Entry, Flag: SecProfSummaryFlags::SecFlagIsPreInlined)) |
1370 | Flags.append(s: "preInlined," ); |
1371 | if (hasSecFlag(Entry, Flag: SecProfSummaryFlags::SecFlagFSDiscriminator)) |
1372 | Flags.append(s: "fs-discriminator," ); |
1373 | break; |
1374 | case SecFuncOffsetTable: |
1375 | if (hasSecFlag(Entry, Flag: SecFuncOffsetFlags::SecFlagOrdered)) |
1376 | Flags.append(s: "ordered," ); |
1377 | break; |
1378 | case SecFuncMetadata: |
1379 | if (hasSecFlag(Entry, Flag: SecFuncMetadataFlags::SecFlagIsProbeBased)) |
1380 | Flags.append(s: "probe," ); |
1381 | if (hasSecFlag(Entry, Flag: SecFuncMetadataFlags::SecFlagHasAttribute)) |
1382 | Flags.append(s: "attr," ); |
1383 | break; |
1384 | default: |
1385 | break; |
1386 | } |
1387 | char &last = Flags.back(); |
1388 | if (last == ',') |
1389 | last = '}'; |
1390 | else |
1391 | Flags.append(s: "}" ); |
1392 | return Flags; |
1393 | } |
1394 | |
1395 | bool SampleProfileReaderExtBinaryBase::dumpSectionInfo(raw_ostream &OS) { |
1396 | uint64_t TotalSecsSize = 0; |
1397 | for (auto &Entry : SecHdrTable) { |
1398 | OS << getSecName(Type: Entry.Type) << " - Offset: " << Entry.Offset |
1399 | << ", Size: " << Entry.Size << ", Flags: " << getSecFlagsStr(Entry) |
1400 | << "\n" ; |
1401 | ; |
1402 | TotalSecsSize += Entry.Size; |
1403 | } |
1404 | uint64_t = SecHdrTable.front().Offset; |
1405 | assert(HeaderSize + TotalSecsSize == getFileSize() && |
1406 | "Size of 'header + sections' doesn't match the total size of profile" ); |
1407 | |
1408 | OS << "Header Size: " << HeaderSize << "\n" ; |
1409 | OS << "Total Sections Size: " << TotalSecsSize << "\n" ; |
1410 | OS << "File Size: " << getFileSize() << "\n" ; |
1411 | return true; |
1412 | } |
1413 | |
1414 | std::error_code SampleProfileReaderBinary::readMagicIdent() { |
1415 | // Read and check the magic identifier. |
1416 | auto Magic = readNumber<uint64_t>(); |
1417 | if (std::error_code EC = Magic.getError()) |
1418 | return EC; |
1419 | else if (std::error_code EC = verifySPMagic(Magic: *Magic)) |
1420 | return EC; |
1421 | |
1422 | // Read the version number. |
1423 | auto Version = readNumber<uint64_t>(); |
1424 | if (std::error_code EC = Version.getError()) |
1425 | return EC; |
1426 | else if (*Version != SPVersion()) |
1427 | return sampleprof_error::unsupported_version; |
1428 | |
1429 | return sampleprof_error::success; |
1430 | } |
1431 | |
1432 | std::error_code SampleProfileReaderBinary::() { |
1433 | Data = reinterpret_cast<const uint8_t *>(Buffer->getBufferStart()); |
1434 | End = Data + Buffer->getBufferSize(); |
1435 | |
1436 | if (std::error_code EC = readMagicIdent()) |
1437 | return EC; |
1438 | |
1439 | if (std::error_code EC = readSummary()) |
1440 | return EC; |
1441 | |
1442 | if (std::error_code EC = readNameTable()) |
1443 | return EC; |
1444 | return sampleprof_error::success; |
1445 | } |
1446 | |
1447 | std::error_code SampleProfileReaderBinary::readSummaryEntry( |
1448 | std::vector<ProfileSummaryEntry> &Entries) { |
1449 | auto Cutoff = readNumber<uint64_t>(); |
1450 | if (std::error_code EC = Cutoff.getError()) |
1451 | return EC; |
1452 | |
1453 | auto MinBlockCount = readNumber<uint64_t>(); |
1454 | if (std::error_code EC = MinBlockCount.getError()) |
1455 | return EC; |
1456 | |
1457 | auto NumBlocks = readNumber<uint64_t>(); |
1458 | if (std::error_code EC = NumBlocks.getError()) |
1459 | return EC; |
1460 | |
1461 | Entries.emplace_back(args&: *Cutoff, args&: *MinBlockCount, args&: *NumBlocks); |
1462 | return sampleprof_error::success; |
1463 | } |
1464 | |
1465 | std::error_code SampleProfileReaderBinary::readSummary() { |
1466 | auto TotalCount = readNumber<uint64_t>(); |
1467 | if (std::error_code EC = TotalCount.getError()) |
1468 | return EC; |
1469 | |
1470 | auto MaxBlockCount = readNumber<uint64_t>(); |
1471 | if (std::error_code EC = MaxBlockCount.getError()) |
1472 | return EC; |
1473 | |
1474 | auto MaxFunctionCount = readNumber<uint64_t>(); |
1475 | if (std::error_code EC = MaxFunctionCount.getError()) |
1476 | return EC; |
1477 | |
1478 | auto NumBlocks = readNumber<uint64_t>(); |
1479 | if (std::error_code EC = NumBlocks.getError()) |
1480 | return EC; |
1481 | |
1482 | auto NumFunctions = readNumber<uint64_t>(); |
1483 | if (std::error_code EC = NumFunctions.getError()) |
1484 | return EC; |
1485 | |
1486 | auto NumSummaryEntries = readNumber<uint64_t>(); |
1487 | if (std::error_code EC = NumSummaryEntries.getError()) |
1488 | return EC; |
1489 | |
1490 | std::vector<ProfileSummaryEntry> Entries; |
1491 | for (unsigned i = 0; i < *NumSummaryEntries; i++) { |
1492 | std::error_code EC = readSummaryEntry(Entries); |
1493 | if (EC != sampleprof_error::success) |
1494 | return EC; |
1495 | } |
1496 | Summary = std::make_unique<ProfileSummary>( |
1497 | args: ProfileSummary::PSK_Sample, args&: Entries, args&: *TotalCount, args&: *MaxBlockCount, args: 0, |
1498 | args&: *MaxFunctionCount, args&: *NumBlocks, args&: *NumFunctions); |
1499 | |
1500 | return sampleprof_error::success; |
1501 | } |
1502 | |
1503 | bool SampleProfileReaderRawBinary::hasFormat(const MemoryBuffer &Buffer) { |
1504 | const uint8_t *Data = |
1505 | reinterpret_cast<const uint8_t *>(Buffer.getBufferStart()); |
1506 | uint64_t Magic = decodeULEB128(p: Data); |
1507 | return Magic == SPMagic(); |
1508 | } |
1509 | |
1510 | bool SampleProfileReaderExtBinary::hasFormat(const MemoryBuffer &Buffer) { |
1511 | const uint8_t *Data = |
1512 | reinterpret_cast<const uint8_t *>(Buffer.getBufferStart()); |
1513 | uint64_t Magic = decodeULEB128(p: Data); |
1514 | return Magic == SPMagic(Format: SPF_Ext_Binary); |
1515 | } |
1516 | |
1517 | std::error_code SampleProfileReaderGCC::skipNextWord() { |
1518 | uint32_t dummy; |
1519 | if (!GcovBuffer.readInt(Val&: dummy)) |
1520 | return sampleprof_error::truncated; |
1521 | return sampleprof_error::success; |
1522 | } |
1523 | |
1524 | template <typename T> ErrorOr<T> SampleProfileReaderGCC::readNumber() { |
1525 | if (sizeof(T) <= sizeof(uint32_t)) { |
1526 | uint32_t Val; |
1527 | if (GcovBuffer.readInt(Val) && Val <= std::numeric_limits<T>::max()) |
1528 | return static_cast<T>(Val); |
1529 | } else if (sizeof(T) <= sizeof(uint64_t)) { |
1530 | uint64_t Val; |
1531 | if (GcovBuffer.readInt64(Val) && Val <= std::numeric_limits<T>::max()) |
1532 | return static_cast<T>(Val); |
1533 | } |
1534 | |
1535 | std::error_code EC = sampleprof_error::malformed; |
1536 | reportError(LineNumber: 0, Msg: EC.message()); |
1537 | return EC; |
1538 | } |
1539 | |
1540 | ErrorOr<StringRef> SampleProfileReaderGCC::readString() { |
1541 | StringRef Str; |
1542 | if (!GcovBuffer.readString(str&: Str)) |
1543 | return sampleprof_error::truncated; |
1544 | return Str; |
1545 | } |
1546 | |
1547 | std::error_code SampleProfileReaderGCC::() { |
1548 | // Read the magic identifier. |
1549 | if (!GcovBuffer.readGCDAFormat()) |
1550 | return sampleprof_error::unrecognized_format; |
1551 | |
1552 | // Read the version number. Note - the GCC reader does not validate this |
1553 | // version, but the profile creator generates v704. |
1554 | GCOV::GCOVVersion version; |
1555 | if (!GcovBuffer.readGCOVVersion(version)) |
1556 | return sampleprof_error::unrecognized_format; |
1557 | |
1558 | if (version != GCOV::V407) |
1559 | return sampleprof_error::unsupported_version; |
1560 | |
1561 | // Skip the empty integer. |
1562 | if (std::error_code EC = skipNextWord()) |
1563 | return EC; |
1564 | |
1565 | return sampleprof_error::success; |
1566 | } |
1567 | |
1568 | std::error_code SampleProfileReaderGCC::readSectionTag(uint32_t Expected) { |
1569 | uint32_t Tag; |
1570 | if (!GcovBuffer.readInt(Val&: Tag)) |
1571 | return sampleprof_error::truncated; |
1572 | |
1573 | if (Tag != Expected) |
1574 | return sampleprof_error::malformed; |
1575 | |
1576 | if (std::error_code EC = skipNextWord()) |
1577 | return EC; |
1578 | |
1579 | return sampleprof_error::success; |
1580 | } |
1581 | |
1582 | std::error_code SampleProfileReaderGCC::readNameTable() { |
1583 | if (std::error_code EC = readSectionTag(Expected: GCOVTagAFDOFileNames)) |
1584 | return EC; |
1585 | |
1586 | uint32_t Size; |
1587 | if (!GcovBuffer.readInt(Val&: Size)) |
1588 | return sampleprof_error::truncated; |
1589 | |
1590 | for (uint32_t I = 0; I < Size; ++I) { |
1591 | StringRef Str; |
1592 | if (!GcovBuffer.readString(str&: Str)) |
1593 | return sampleprof_error::truncated; |
1594 | Names.push_back(x: std::string(Str)); |
1595 | } |
1596 | |
1597 | return sampleprof_error::success; |
1598 | } |
1599 | |
1600 | std::error_code SampleProfileReaderGCC::readFunctionProfiles() { |
1601 | if (std::error_code EC = readSectionTag(Expected: GCOVTagAFDOFunction)) |
1602 | return EC; |
1603 | |
1604 | uint32_t NumFunctions; |
1605 | if (!GcovBuffer.readInt(Val&: NumFunctions)) |
1606 | return sampleprof_error::truncated; |
1607 | |
1608 | InlineCallStack Stack; |
1609 | for (uint32_t I = 0; I < NumFunctions; ++I) |
1610 | if (std::error_code EC = readOneFunctionProfile(InlineStack: Stack, Update: true, Offset: 0)) |
1611 | return EC; |
1612 | |
1613 | computeSummary(); |
1614 | return sampleprof_error::success; |
1615 | } |
1616 | |
1617 | std::error_code SampleProfileReaderGCC::readOneFunctionProfile( |
1618 | const InlineCallStack &InlineStack, bool Update, uint32_t Offset) { |
1619 | uint64_t HeadCount = 0; |
1620 | if (InlineStack.size() == 0) |
1621 | if (!GcovBuffer.readInt64(Val&: HeadCount)) |
1622 | return sampleprof_error::truncated; |
1623 | |
1624 | uint32_t NameIdx; |
1625 | if (!GcovBuffer.readInt(Val&: NameIdx)) |
1626 | return sampleprof_error::truncated; |
1627 | |
1628 | StringRef Name(Names[NameIdx]); |
1629 | |
1630 | uint32_t NumPosCounts; |
1631 | if (!GcovBuffer.readInt(Val&: NumPosCounts)) |
1632 | return sampleprof_error::truncated; |
1633 | |
1634 | uint32_t NumCallsites; |
1635 | if (!GcovBuffer.readInt(Val&: NumCallsites)) |
1636 | return sampleprof_error::truncated; |
1637 | |
1638 | FunctionSamples *FProfile = nullptr; |
1639 | if (InlineStack.size() == 0) { |
1640 | // If this is a top function that we have already processed, do not |
1641 | // update its profile again. This happens in the presence of |
1642 | // function aliases. Since these aliases share the same function |
1643 | // body, there will be identical replicated profiles for the |
1644 | // original function. In this case, we simply not bother updating |
1645 | // the profile of the original function. |
1646 | FProfile = &Profiles[FunctionId(Name)]; |
1647 | FProfile->addHeadSamples(Num: HeadCount); |
1648 | if (FProfile->getTotalSamples() > 0) |
1649 | Update = false; |
1650 | } else { |
1651 | // Otherwise, we are reading an inlined instance. The top of the |
1652 | // inline stack contains the profile of the caller. Insert this |
1653 | // callee in the caller's CallsiteMap. |
1654 | FunctionSamples *CallerProfile = InlineStack.front(); |
1655 | uint32_t LineOffset = Offset >> 16; |
1656 | uint32_t Discriminator = Offset & 0xffff; |
1657 | FProfile = &CallerProfile->functionSamplesAt( |
1658 | Loc: LineLocation(LineOffset, Discriminator))[FunctionId(Name)]; |
1659 | } |
1660 | FProfile->setFunction(FunctionId(Name)); |
1661 | |
1662 | for (uint32_t I = 0; I < NumPosCounts; ++I) { |
1663 | uint32_t Offset; |
1664 | if (!GcovBuffer.readInt(Val&: Offset)) |
1665 | return sampleprof_error::truncated; |
1666 | |
1667 | uint32_t NumTargets; |
1668 | if (!GcovBuffer.readInt(Val&: NumTargets)) |
1669 | return sampleprof_error::truncated; |
1670 | |
1671 | uint64_t Count; |
1672 | if (!GcovBuffer.readInt64(Val&: Count)) |
1673 | return sampleprof_error::truncated; |
1674 | |
1675 | // The line location is encoded in the offset as: |
1676 | // high 16 bits: line offset to the start of the function. |
1677 | // low 16 bits: discriminator. |
1678 | uint32_t LineOffset = Offset >> 16; |
1679 | uint32_t Discriminator = Offset & 0xffff; |
1680 | |
1681 | InlineCallStack NewStack; |
1682 | NewStack.push_back(Elt: FProfile); |
1683 | llvm::append_range(C&: NewStack, R: InlineStack); |
1684 | if (Update) { |
1685 | // Walk up the inline stack, adding the samples on this line to |
1686 | // the total sample count of the callers in the chain. |
1687 | for (auto *CallerProfile : NewStack) |
1688 | CallerProfile->addTotalSamples(Num: Count); |
1689 | |
1690 | // Update the body samples for the current profile. |
1691 | FProfile->addBodySamples(LineOffset, Discriminator, Num: Count); |
1692 | } |
1693 | |
1694 | // Process the list of functions called at an indirect call site. |
1695 | // These are all the targets that a function pointer (or virtual |
1696 | // function) resolved at runtime. |
1697 | for (uint32_t J = 0; J < NumTargets; J++) { |
1698 | uint32_t HistVal; |
1699 | if (!GcovBuffer.readInt(Val&: HistVal)) |
1700 | return sampleprof_error::truncated; |
1701 | |
1702 | if (HistVal != HIST_TYPE_INDIR_CALL_TOPN) |
1703 | return sampleprof_error::malformed; |
1704 | |
1705 | uint64_t TargetIdx; |
1706 | if (!GcovBuffer.readInt64(Val&: TargetIdx)) |
1707 | return sampleprof_error::truncated; |
1708 | StringRef TargetName(Names[TargetIdx]); |
1709 | |
1710 | uint64_t TargetCount; |
1711 | if (!GcovBuffer.readInt64(Val&: TargetCount)) |
1712 | return sampleprof_error::truncated; |
1713 | |
1714 | if (Update) |
1715 | FProfile->addCalledTargetSamples(LineOffset, Discriminator, |
1716 | Func: FunctionId(TargetName), |
1717 | Num: TargetCount); |
1718 | } |
1719 | } |
1720 | |
1721 | // Process all the inlined callers into the current function. These |
1722 | // are all the callsites that were inlined into this function. |
1723 | for (uint32_t I = 0; I < NumCallsites; I++) { |
1724 | // The offset is encoded as: |
1725 | // high 16 bits: line offset to the start of the function. |
1726 | // low 16 bits: discriminator. |
1727 | uint32_t Offset; |
1728 | if (!GcovBuffer.readInt(Val&: Offset)) |
1729 | return sampleprof_error::truncated; |
1730 | InlineCallStack NewStack; |
1731 | NewStack.push_back(Elt: FProfile); |
1732 | llvm::append_range(C&: NewStack, R: InlineStack); |
1733 | if (std::error_code EC = readOneFunctionProfile(InlineStack: NewStack, Update, Offset)) |
1734 | return EC; |
1735 | } |
1736 | |
1737 | return sampleprof_error::success; |
1738 | } |
1739 | |
1740 | /// Read a GCC AutoFDO profile. |
1741 | /// |
1742 | /// This format is generated by the Linux Perf conversion tool at |
1743 | /// https://github.com/google/autofdo. |
1744 | std::error_code SampleProfileReaderGCC::readImpl() { |
1745 | assert(!ProfileIsFSDisciminator && "Gcc profiles not support FSDisciminator" ); |
1746 | // Read the string table. |
1747 | if (std::error_code EC = readNameTable()) |
1748 | return EC; |
1749 | |
1750 | // Read the source profile. |
1751 | if (std::error_code EC = readFunctionProfiles()) |
1752 | return EC; |
1753 | |
1754 | return sampleprof_error::success; |
1755 | } |
1756 | |
1757 | bool SampleProfileReaderGCC::hasFormat(const MemoryBuffer &Buffer) { |
1758 | StringRef Magic(reinterpret_cast<const char *>(Buffer.getBufferStart())); |
1759 | return Magic == "adcg*704" ; |
1760 | } |
1761 | |
1762 | void SampleProfileReaderItaniumRemapper::applyRemapping(LLVMContext &Ctx) { |
1763 | // If the reader uses MD5 to represent string, we can't remap it because |
1764 | // we don't know what the original function names were. |
1765 | if (Reader.useMD5()) { |
1766 | Ctx.diagnose(DI: DiagnosticInfoSampleProfile( |
1767 | Reader.getBuffer()->getBufferIdentifier(), |
1768 | "Profile data remapping cannot be applied to profile data " |
1769 | "using MD5 names (original mangled names are not available)." , |
1770 | DS_Warning)); |
1771 | return; |
1772 | } |
1773 | |
1774 | // CSSPGO-TODO: Remapper is not yet supported. |
1775 | // We will need to remap the entire context string. |
1776 | assert(Remappings && "should be initialized while creating remapper" ); |
1777 | for (auto &Sample : Reader.getProfiles()) { |
1778 | DenseSet<FunctionId> NamesInSample; |
1779 | Sample.second.findAllNames(NameSet&: NamesInSample); |
1780 | for (auto &Name : NamesInSample) { |
1781 | StringRef NameStr = Name.stringRef(); |
1782 | if (auto Key = Remappings->insert(FunctionName: NameStr)) |
1783 | NameMap.insert(KV: {Key, NameStr}); |
1784 | } |
1785 | } |
1786 | |
1787 | RemappingApplied = true; |
1788 | } |
1789 | |
1790 | std::optional<StringRef> |
1791 | SampleProfileReaderItaniumRemapper::lookUpNameInProfile(StringRef Fname) { |
1792 | if (auto Key = Remappings->lookup(FunctionName: Fname)) { |
1793 | StringRef Result = NameMap.lookup(Val: Key); |
1794 | if (!Result.empty()) |
1795 | return Result; |
1796 | } |
1797 | return std::nullopt; |
1798 | } |
1799 | |
1800 | /// Prepare a memory buffer for the contents of \p Filename. |
1801 | /// |
1802 | /// \returns an error code indicating the status of the buffer. |
1803 | static ErrorOr<std::unique_ptr<MemoryBuffer>> |
1804 | setupMemoryBuffer(const Twine &Filename, vfs::FileSystem &FS) { |
1805 | auto BufferOrErr = Filename.str() == "-" ? MemoryBuffer::getSTDIN() |
1806 | : FS.getBufferForFile(Name: Filename); |
1807 | if (std::error_code EC = BufferOrErr.getError()) |
1808 | return EC; |
1809 | auto Buffer = std::move(BufferOrErr.get()); |
1810 | |
1811 | return std::move(Buffer); |
1812 | } |
1813 | |
1814 | /// Create a sample profile reader based on the format of the input file. |
1815 | /// |
1816 | /// \param Filename The file to open. |
1817 | /// |
1818 | /// \param C The LLVM context to use to emit diagnostics. |
1819 | /// |
1820 | /// \param P The FSDiscriminatorPass. |
1821 | /// |
1822 | /// \param RemapFilename The file used for profile remapping. |
1823 | /// |
1824 | /// \returns an error code indicating the status of the created reader. |
1825 | ErrorOr<std::unique_ptr<SampleProfileReader>> |
1826 | SampleProfileReader::create(StringRef Filename, LLVMContext &C, |
1827 | vfs::FileSystem &FS, FSDiscriminatorPass P, |
1828 | StringRef RemapFilename) { |
1829 | auto BufferOrError = setupMemoryBuffer(Filename, FS); |
1830 | if (std::error_code EC = BufferOrError.getError()) |
1831 | return EC; |
1832 | return create(B&: BufferOrError.get(), C, FS, P, RemapFilename); |
1833 | } |
1834 | |
1835 | /// Create a sample profile remapper from the given input, to remap the |
1836 | /// function names in the given profile data. |
1837 | /// |
1838 | /// \param Filename The file to open. |
1839 | /// |
1840 | /// \param Reader The profile reader the remapper is going to be applied to. |
1841 | /// |
1842 | /// \param C The LLVM context to use to emit diagnostics. |
1843 | /// |
1844 | /// \returns an error code indicating the status of the created reader. |
1845 | ErrorOr<std::unique_ptr<SampleProfileReaderItaniumRemapper>> |
1846 | SampleProfileReaderItaniumRemapper::create(StringRef Filename, |
1847 | vfs::FileSystem &FS, |
1848 | SampleProfileReader &Reader, |
1849 | LLVMContext &C) { |
1850 | auto BufferOrError = setupMemoryBuffer(Filename, FS); |
1851 | if (std::error_code EC = BufferOrError.getError()) |
1852 | return EC; |
1853 | return create(B&: BufferOrError.get(), Reader, C); |
1854 | } |
1855 | |
1856 | /// Create a sample profile remapper from the given input, to remap the |
1857 | /// function names in the given profile data. |
1858 | /// |
1859 | /// \param B The memory buffer to create the reader from (assumes ownership). |
1860 | /// |
1861 | /// \param C The LLVM context to use to emit diagnostics. |
1862 | /// |
1863 | /// \param Reader The profile reader the remapper is going to be applied to. |
1864 | /// |
1865 | /// \returns an error code indicating the status of the created reader. |
1866 | ErrorOr<std::unique_ptr<SampleProfileReaderItaniumRemapper>> |
1867 | SampleProfileReaderItaniumRemapper::create(std::unique_ptr<MemoryBuffer> &B, |
1868 | SampleProfileReader &Reader, |
1869 | LLVMContext &C) { |
1870 | auto Remappings = std::make_unique<SymbolRemappingReader>(); |
1871 | if (Error E = Remappings->read(B&: *B)) { |
1872 | handleAllErrors( |
1873 | E: std::move(E), Handlers: [&](const SymbolRemappingParseError &ParseError) { |
1874 | C.diagnose(DI: DiagnosticInfoSampleProfile(B->getBufferIdentifier(), |
1875 | ParseError.getLineNum(), |
1876 | ParseError.getMessage())); |
1877 | }); |
1878 | return sampleprof_error::malformed; |
1879 | } |
1880 | |
1881 | return std::make_unique<SampleProfileReaderItaniumRemapper>( |
1882 | args: std::move(B), args: std::move(Remappings), args&: Reader); |
1883 | } |
1884 | |
1885 | /// Create a sample profile reader based on the format of the input data. |
1886 | /// |
1887 | /// \param B The memory buffer to create the reader from (assumes ownership). |
1888 | /// |
1889 | /// \param C The LLVM context to use to emit diagnostics. |
1890 | /// |
1891 | /// \param P The FSDiscriminatorPass. |
1892 | /// |
1893 | /// \param RemapFilename The file used for profile remapping. |
1894 | /// |
1895 | /// \returns an error code indicating the status of the created reader. |
1896 | ErrorOr<std::unique_ptr<SampleProfileReader>> |
1897 | SampleProfileReader::create(std::unique_ptr<MemoryBuffer> &B, LLVMContext &C, |
1898 | vfs::FileSystem &FS, FSDiscriminatorPass P, |
1899 | StringRef RemapFilename) { |
1900 | std::unique_ptr<SampleProfileReader> Reader; |
1901 | if (SampleProfileReaderRawBinary::hasFormat(Buffer: *B)) |
1902 | Reader.reset(p: new SampleProfileReaderRawBinary(std::move(B), C)); |
1903 | else if (SampleProfileReaderExtBinary::hasFormat(Buffer: *B)) |
1904 | Reader.reset(p: new SampleProfileReaderExtBinary(std::move(B), C)); |
1905 | else if (SampleProfileReaderGCC::hasFormat(Buffer: *B)) |
1906 | Reader.reset(p: new SampleProfileReaderGCC(std::move(B), C)); |
1907 | else if (SampleProfileReaderText::hasFormat(Buffer: *B)) |
1908 | Reader.reset(p: new SampleProfileReaderText(std::move(B), C)); |
1909 | else |
1910 | return sampleprof_error::unrecognized_format; |
1911 | |
1912 | if (!RemapFilename.empty()) { |
1913 | auto ReaderOrErr = SampleProfileReaderItaniumRemapper::create( |
1914 | Filename: RemapFilename, FS, Reader&: *Reader, C); |
1915 | if (std::error_code EC = ReaderOrErr.getError()) { |
1916 | std::string Msg = "Could not create remapper: " + EC.message(); |
1917 | C.diagnose(DI: DiagnosticInfoSampleProfile(RemapFilename, Msg)); |
1918 | return EC; |
1919 | } |
1920 | Reader->Remapper = std::move(ReaderOrErr.get()); |
1921 | } |
1922 | |
1923 | if (std::error_code EC = Reader->readHeader()) { |
1924 | return EC; |
1925 | } |
1926 | |
1927 | Reader->setDiscriminatorMaskedBitFrom(P); |
1928 | |
1929 | return std::move(Reader); |
1930 | } |
1931 | |
1932 | // For text and GCC file formats, we compute the summary after reading the |
1933 | // profile. Binary format has the profile summary in its header. |
1934 | void SampleProfileReader::computeSummary() { |
1935 | SampleProfileSummaryBuilder Builder(ProfileSummaryBuilder::DefaultCutoffs); |
1936 | Summary = Builder.computeSummaryForProfiles(Profiles); |
1937 | } |
1938 | |