| 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 | bool &IsFlat) { |
| 220 | for (Depth = 0; Input[Depth] == ' '; Depth++) |
| 221 | ; |
| 222 | if (Depth == 0) |
| 223 | return false; |
| 224 | |
| 225 | if (Input[Depth] == '!') { |
| 226 | LineTy = LineType::Metadata; |
| 227 | // This metadata is only for manual inspection only. We already created a |
| 228 | // FunctionSamples and put it in the profile map, so there is no point |
| 229 | // to skip profiles even they have no use for ThinLTO. |
| 230 | if (Input == StringRef(" !Flat" )) { |
| 231 | IsFlat = true; |
| 232 | return true; |
| 233 | } |
| 234 | return parseMetadata(Input: Input.substr(Start: Depth), FunctionHash, Attributes); |
| 235 | } |
| 236 | |
| 237 | size_t n1 = Input.find(C: ':'); |
| 238 | StringRef Loc = Input.substr(Start: Depth, N: n1 - Depth); |
| 239 | size_t n2 = Loc.find(C: '.'); |
| 240 | if (n2 == StringRef::npos) { |
| 241 | if (Loc.getAsInteger(Radix: 10, Result&: LineOffset) || !isOffsetLegal(L: LineOffset)) |
| 242 | return false; |
| 243 | Discriminator = 0; |
| 244 | } else { |
| 245 | if (Loc.substr(Start: 0, N: n2).getAsInteger(Radix: 10, Result&: LineOffset)) |
| 246 | return false; |
| 247 | if (Loc.substr(Start: n2 + 1).getAsInteger(Radix: 10, Result&: Discriminator)) |
| 248 | return false; |
| 249 | } |
| 250 | |
| 251 | StringRef Rest = Input.substr(Start: n1 + 2); |
| 252 | if (isDigit(C: Rest[0])) { |
| 253 | LineTy = LineType::BodyProfile; |
| 254 | size_t n3 = Rest.find(C: ' '); |
| 255 | if (n3 == StringRef::npos) { |
| 256 | if (Rest.getAsInteger(Radix: 10, Result&: NumSamples)) |
| 257 | return false; |
| 258 | } else { |
| 259 | if (Rest.substr(Start: 0, N: n3).getAsInteger(Radix: 10, Result&: NumSamples)) |
| 260 | return false; |
| 261 | } |
| 262 | // Find call targets and their sample counts. |
| 263 | // Note: In some cases, there are symbols in the profile which are not |
| 264 | // mangled. To accommodate such cases, use colon + integer pairs as the |
| 265 | // anchor points. |
| 266 | // An example: |
| 267 | // _M_construct<char *>:1000 string_view<std::allocator<char> >:437 |
| 268 | // ":1000" and ":437" are used as anchor points so the string above will |
| 269 | // be interpreted as |
| 270 | // target: _M_construct<char *> |
| 271 | // count: 1000 |
| 272 | // target: string_view<std::allocator<char> > |
| 273 | // count: 437 |
| 274 | while (n3 != StringRef::npos) { |
| 275 | n3 += Rest.substr(Start: n3).find_first_not_of(C: ' '); |
| 276 | Rest = Rest.substr(Start: n3); |
| 277 | n3 = Rest.find_first_of(C: ':'); |
| 278 | if (n3 == StringRef::npos || n3 == 0) |
| 279 | return false; |
| 280 | |
| 281 | StringRef Target; |
| 282 | uint64_t count, n4; |
| 283 | while (true) { |
| 284 | // Get the segment after the current colon. |
| 285 | StringRef AfterColon = Rest.substr(Start: n3 + 1); |
| 286 | // Get the target symbol before the current colon. |
| 287 | Target = Rest.substr(Start: 0, N: n3); |
| 288 | // Check if the word after the current colon is an integer. |
| 289 | n4 = AfterColon.find_first_of(C: ' '); |
| 290 | n4 = (n4 != StringRef::npos) ? n3 + n4 + 1 : Rest.size(); |
| 291 | StringRef WordAfterColon = Rest.substr(Start: n3 + 1, N: n4 - n3 - 1); |
| 292 | if (!WordAfterColon.getAsInteger(Radix: 10, Result&: count)) |
| 293 | break; |
| 294 | |
| 295 | // Try to find the next colon. |
| 296 | uint64_t n5 = AfterColon.find_first_of(C: ':'); |
| 297 | if (n5 == StringRef::npos) |
| 298 | return false; |
| 299 | n3 += n5 + 1; |
| 300 | } |
| 301 | |
| 302 | // An anchor point is found. Save the {target, count} pair |
| 303 | TargetCountMap[Target] = count; |
| 304 | if (n4 == Rest.size()) |
| 305 | break; |
| 306 | // Change n3 to the next blank space after colon + integer pair. |
| 307 | n3 = n4; |
| 308 | } |
| 309 | } else { |
| 310 | LineTy = LineType::CallSiteProfile; |
| 311 | size_t n3 = Rest.find_last_of(C: ':'); |
| 312 | CalleeName = Rest.substr(Start: 0, N: n3); |
| 313 | if (Rest.substr(Start: n3 + 1).getAsInteger(Radix: 10, Result&: NumSamples)) |
| 314 | return false; |
| 315 | } |
| 316 | return true; |
| 317 | } |
| 318 | |
| 319 | /// Load samples from a text file. |
| 320 | /// |
| 321 | /// See the documentation at the top of the file for an explanation of |
| 322 | /// the expected format. |
| 323 | /// |
| 324 | /// \returns true if the file was loaded successfully, false otherwise. |
| 325 | std::error_code SampleProfileReaderText::readImpl() { |
| 326 | line_iterator LineIt(*Buffer, /*SkipBlanks=*/true, '#'); |
| 327 | sampleprof_error Result = sampleprof_error::success; |
| 328 | |
| 329 | InlineCallStack InlineStack; |
| 330 | uint32_t TopLevelProbeProfileCount = 0; |
| 331 | |
| 332 | // DepthMetadata tracks whether we have processed metadata for the current |
| 333 | // top-level or nested function profile. |
| 334 | uint32_t DepthMetadata = 0; |
| 335 | |
| 336 | std::vector<SampleContext *> FlatSamples; |
| 337 | |
| 338 | ProfileIsFS = ProfileIsFSDisciminator; |
| 339 | FunctionSamples::ProfileIsFS = ProfileIsFS; |
| 340 | for (; !LineIt.is_at_eof(); ++LineIt) { |
| 341 | size_t pos = LineIt->find_first_not_of(C: ' '); |
| 342 | if (pos == LineIt->npos || (*LineIt)[pos] == '#') |
| 343 | continue; |
| 344 | // Read the header of each function. |
| 345 | // |
| 346 | // Note that for function identifiers we are actually expecting |
| 347 | // mangled names, but we may not always get them. This happens when |
| 348 | // the compiler decides not to emit the function (e.g., it was inlined |
| 349 | // and removed). In this case, the binary will not have the linkage |
| 350 | // name for the function, so the profiler will emit the function's |
| 351 | // unmangled name, which may contain characters like ':' and '>' in its |
| 352 | // name (member functions, templates, etc). |
| 353 | // |
| 354 | // The only requirement we place on the identifier, then, is that it |
| 355 | // should not begin with a number. |
| 356 | if ((*LineIt)[0] != ' ') { |
| 357 | uint64_t NumSamples, NumHeadSamples; |
| 358 | StringRef FName; |
| 359 | if (!ParseHead(Input: *LineIt, FName, NumSamples, NumHeadSamples)) { |
| 360 | reportError(LineNumber: LineIt.line_number(), |
| 361 | Msg: "Expected 'mangled_name:NUM:NUM', found " + *LineIt); |
| 362 | return sampleprof_error::malformed; |
| 363 | } |
| 364 | DepthMetadata = 0; |
| 365 | SampleContext FContext(FName, CSNameTable); |
| 366 | if (FContext.hasContext()) |
| 367 | ++CSProfileCount; |
| 368 | FunctionSamples &FProfile = Profiles.create(Ctx: FContext); |
| 369 | mergeSampleProfErrors(Accumulator&: Result, Result: FProfile.addTotalSamples(Num: NumSamples)); |
| 370 | mergeSampleProfErrors(Accumulator&: Result, Result: FProfile.addHeadSamples(Num: NumHeadSamples)); |
| 371 | InlineStack.clear(); |
| 372 | InlineStack.push_back(Elt: &FProfile); |
| 373 | } else { |
| 374 | uint64_t NumSamples; |
| 375 | StringRef FName; |
| 376 | DenseMap<StringRef, uint64_t> TargetCountMap; |
| 377 | uint32_t Depth, LineOffset, Discriminator; |
| 378 | LineType LineTy; |
| 379 | uint64_t FunctionHash = 0; |
| 380 | uint32_t Attributes = 0; |
| 381 | bool IsFlat = false; |
| 382 | if (!ParseLine(Input: *LineIt, LineTy, Depth, NumSamples, LineOffset, |
| 383 | Discriminator, CalleeName&: FName, TargetCountMap, FunctionHash, |
| 384 | Attributes, IsFlat)) { |
| 385 | reportError(LineNumber: LineIt.line_number(), |
| 386 | Msg: "Expected 'NUM[.NUM]: NUM[ mangled_name:NUM]*', found " + |
| 387 | *LineIt); |
| 388 | return sampleprof_error::malformed; |
| 389 | } |
| 390 | if (LineTy != LineType::Metadata && Depth == DepthMetadata) { |
| 391 | // Metadata must be put at the end of a function profile. |
| 392 | reportError(LineNumber: LineIt.line_number(), |
| 393 | Msg: "Found non-metadata after metadata: " + *LineIt); |
| 394 | return sampleprof_error::malformed; |
| 395 | } |
| 396 | |
| 397 | // Here we handle FS discriminators. |
| 398 | Discriminator &= getDiscriminatorMask(); |
| 399 | |
| 400 | while (InlineStack.size() > Depth) { |
| 401 | InlineStack.pop_back(); |
| 402 | } |
| 403 | switch (LineTy) { |
| 404 | case LineType::CallSiteProfile: { |
| 405 | FunctionSamples &FSamples = InlineStack.back()->functionSamplesAt( |
| 406 | Loc: LineLocation(LineOffset, Discriminator))[FunctionId(FName)]; |
| 407 | FSamples.setFunction(FunctionId(FName)); |
| 408 | mergeSampleProfErrors(Accumulator&: Result, Result: FSamples.addTotalSamples(Num: NumSamples)); |
| 409 | InlineStack.push_back(Elt: &FSamples); |
| 410 | DepthMetadata = 0; |
| 411 | break; |
| 412 | } |
| 413 | case LineType::BodyProfile: { |
| 414 | FunctionSamples &FProfile = *InlineStack.back(); |
| 415 | for (const auto &name_count : TargetCountMap) { |
| 416 | mergeSampleProfErrors(Accumulator&: Result, Result: FProfile.addCalledTargetSamples( |
| 417 | LineOffset, Discriminator, |
| 418 | Func: FunctionId(name_count.first), |
| 419 | Num: name_count.second)); |
| 420 | } |
| 421 | mergeSampleProfErrors( |
| 422 | Accumulator&: Result, |
| 423 | Result: FProfile.addBodySamples(LineOffset, Discriminator, Num: NumSamples)); |
| 424 | break; |
| 425 | } |
| 426 | case LineType::Metadata: { |
| 427 | FunctionSamples &FProfile = *InlineStack.back(); |
| 428 | if (FunctionHash) { |
| 429 | FProfile.setFunctionHash(FunctionHash); |
| 430 | if (Depth == 1) |
| 431 | ++TopLevelProbeProfileCount; |
| 432 | } |
| 433 | FProfile.getContext().setAllAttributes(Attributes); |
| 434 | if (Attributes & (uint32_t)ContextShouldBeInlined) |
| 435 | ProfileIsPreInlined = true; |
| 436 | DepthMetadata = Depth; |
| 437 | if (IsFlat) { |
| 438 | if (Depth == 1) |
| 439 | FlatSamples.push_back(x: &FProfile.getContext()); |
| 440 | else |
| 441 | Ctx.diagnose(DI: DiagnosticInfoSampleProfile( |
| 442 | Buffer->getBufferIdentifier(), LineIt.line_number(), |
| 443 | "!Flat may only be used at top level function." , DS_Warning)); |
| 444 | } |
| 445 | break; |
| 446 | } |
| 447 | } |
| 448 | } |
| 449 | } |
| 450 | |
| 451 | // Honor the option to skip flat functions. Since they are already added to |
| 452 | // the profile map, remove them all here. |
| 453 | if (SkipFlatProf) |
| 454 | for (SampleContext *FlatSample : FlatSamples) |
| 455 | Profiles.erase(Ctx: *FlatSample); |
| 456 | |
| 457 | assert((CSProfileCount == 0 || CSProfileCount == Profiles.size()) && |
| 458 | "Cannot have both context-sensitive and regular profile" ); |
| 459 | ProfileIsCS = (CSProfileCount > 0); |
| 460 | assert((TopLevelProbeProfileCount == 0 || |
| 461 | TopLevelProbeProfileCount == Profiles.size()) && |
| 462 | "Cannot have both probe-based profiles and regular profiles" ); |
| 463 | ProfileIsProbeBased = (TopLevelProbeProfileCount > 0); |
| 464 | FunctionSamples::ProfileIsProbeBased = ProfileIsProbeBased; |
| 465 | FunctionSamples::ProfileIsCS = ProfileIsCS; |
| 466 | FunctionSamples::ProfileIsPreInlined = ProfileIsPreInlined; |
| 467 | |
| 468 | if (Result == sampleprof_error::success) |
| 469 | computeSummary(); |
| 470 | |
| 471 | return Result; |
| 472 | } |
| 473 | |
| 474 | bool SampleProfileReaderText::hasFormat(const MemoryBuffer &Buffer) { |
| 475 | bool result = false; |
| 476 | |
| 477 | // Check that the first non-comment line is a valid function header. |
| 478 | line_iterator LineIt(Buffer, /*SkipBlanks=*/true, '#'); |
| 479 | if (!LineIt.is_at_eof()) { |
| 480 | if ((*LineIt)[0] != ' ') { |
| 481 | uint64_t NumSamples, NumHeadSamples; |
| 482 | StringRef FName; |
| 483 | result = ParseHead(Input: *LineIt, FName, NumSamples, NumHeadSamples); |
| 484 | } |
| 485 | } |
| 486 | |
| 487 | return result; |
| 488 | } |
| 489 | |
| 490 | template <typename T> ErrorOr<T> SampleProfileReaderBinary::readNumber() { |
| 491 | unsigned NumBytesRead = 0; |
| 492 | uint64_t Val = decodeULEB128(p: Data, n: &NumBytesRead); |
| 493 | |
| 494 | if (Val > std::numeric_limits<T>::max()) { |
| 495 | std::error_code EC = sampleprof_error::malformed; |
| 496 | reportError(LineNumber: 0, Msg: EC.message()); |
| 497 | return EC; |
| 498 | } else if (Data + NumBytesRead > End) { |
| 499 | std::error_code EC = sampleprof_error::truncated; |
| 500 | reportError(LineNumber: 0, Msg: EC.message()); |
| 501 | return EC; |
| 502 | } |
| 503 | |
| 504 | Data += NumBytesRead; |
| 505 | return static_cast<T>(Val); |
| 506 | } |
| 507 | |
| 508 | ErrorOr<StringRef> SampleProfileReaderBinary::readString() { |
| 509 | StringRef Str(reinterpret_cast<const char *>(Data)); |
| 510 | if (Data + Str.size() + 1 > End) { |
| 511 | std::error_code EC = sampleprof_error::truncated; |
| 512 | reportError(LineNumber: 0, Msg: EC.message()); |
| 513 | return EC; |
| 514 | } |
| 515 | |
| 516 | Data += Str.size() + 1; |
| 517 | return Str; |
| 518 | } |
| 519 | |
| 520 | template <typename T> |
| 521 | ErrorOr<T> SampleProfileReaderBinary::readUnencodedNumber() { |
| 522 | if (Data + sizeof(T) > End) { |
| 523 | std::error_code EC = sampleprof_error::truncated; |
| 524 | reportError(LineNumber: 0, Msg: EC.message()); |
| 525 | return EC; |
| 526 | } |
| 527 | |
| 528 | using namespace support; |
| 529 | T Val = endian::readNext<T, llvm::endianness::little>(Data); |
| 530 | return Val; |
| 531 | } |
| 532 | |
| 533 | template <typename T> |
| 534 | inline ErrorOr<size_t> SampleProfileReaderBinary::readStringIndex(T &Table) { |
| 535 | auto Idx = readNumber<size_t>(); |
| 536 | if (std::error_code EC = Idx.getError()) |
| 537 | return EC; |
| 538 | if (*Idx >= Table.size()) |
| 539 | return sampleprof_error::truncated_name_table; |
| 540 | return *Idx; |
| 541 | } |
| 542 | |
| 543 | ErrorOr<FunctionId> |
| 544 | SampleProfileReaderBinary::readStringFromTable(size_t *RetIdx) { |
| 545 | auto Idx = readStringIndex(Table&: NameTable); |
| 546 | if (std::error_code EC = Idx.getError()) |
| 547 | return EC; |
| 548 | if (RetIdx) |
| 549 | *RetIdx = *Idx; |
| 550 | return NameTable[*Idx]; |
| 551 | } |
| 552 | |
| 553 | ErrorOr<SampleContextFrames> |
| 554 | SampleProfileReaderBinary::readContextFromTable(size_t *RetIdx) { |
| 555 | auto ContextIdx = readNumber<size_t>(); |
| 556 | if (std::error_code EC = ContextIdx.getError()) |
| 557 | return EC; |
| 558 | if (*ContextIdx >= CSNameTable.size()) |
| 559 | return sampleprof_error::truncated_name_table; |
| 560 | if (RetIdx) |
| 561 | *RetIdx = *ContextIdx; |
| 562 | return CSNameTable[*ContextIdx]; |
| 563 | } |
| 564 | |
| 565 | ErrorOr<std::pair<SampleContext, uint64_t>> |
| 566 | SampleProfileReaderBinary::readSampleContextFromTable() { |
| 567 | SampleContext Context; |
| 568 | size_t Idx; |
| 569 | if (ProfileIsCS) { |
| 570 | auto FContext(readContextFromTable(RetIdx: &Idx)); |
| 571 | if (std::error_code EC = FContext.getError()) |
| 572 | return EC; |
| 573 | Context = SampleContext(*FContext); |
| 574 | } else { |
| 575 | auto FName(readStringFromTable(RetIdx: &Idx)); |
| 576 | if (std::error_code EC = FName.getError()) |
| 577 | return EC; |
| 578 | Context = SampleContext(*FName); |
| 579 | } |
| 580 | // Since MD5SampleContextStart may point to the profile's file data, need to |
| 581 | // make sure it is reading the same value on big endian CPU. |
| 582 | uint64_t Hash = support::endian::read64le(P: MD5SampleContextStart + Idx); |
| 583 | // Lazy computing of hash value, write back to the table to cache it. Only |
| 584 | // compute the context's hash value if it is being referenced for the first |
| 585 | // time. |
| 586 | if (Hash == 0) { |
| 587 | assert(MD5SampleContextStart == MD5SampleContextTable.data()); |
| 588 | Hash = Context.getHashCode(); |
| 589 | support::endian::write64le(P: &MD5SampleContextTable[Idx], V: Hash); |
| 590 | } |
| 591 | return std::make_pair(x&: Context, y&: Hash); |
| 592 | } |
| 593 | |
| 594 | std::error_code |
| 595 | SampleProfileReaderBinary::readProfile(FunctionSamples &FProfile) { |
| 596 | auto NumSamples = readNumber<uint64_t>(); |
| 597 | if (std::error_code EC = NumSamples.getError()) |
| 598 | return EC; |
| 599 | FProfile.addTotalSamples(Num: *NumSamples); |
| 600 | |
| 601 | // Read the samples in the body. |
| 602 | auto NumRecords = readNumber<uint32_t>(); |
| 603 | if (std::error_code EC = NumRecords.getError()) |
| 604 | return EC; |
| 605 | |
| 606 | for (uint32_t I = 0; I < *NumRecords; ++I) { |
| 607 | auto LineOffset = readNumber<uint64_t>(); |
| 608 | if (std::error_code EC = LineOffset.getError()) |
| 609 | return EC; |
| 610 | |
| 611 | if (!isOffsetLegal(L: *LineOffset)) { |
| 612 | return std::error_code(); |
| 613 | } |
| 614 | |
| 615 | auto Discriminator = readNumber<uint64_t>(); |
| 616 | if (std::error_code EC = Discriminator.getError()) |
| 617 | return EC; |
| 618 | |
| 619 | auto NumSamples = readNumber<uint64_t>(); |
| 620 | if (std::error_code EC = NumSamples.getError()) |
| 621 | return EC; |
| 622 | |
| 623 | auto NumCalls = readNumber<uint32_t>(); |
| 624 | if (std::error_code EC = NumCalls.getError()) |
| 625 | return EC; |
| 626 | |
| 627 | // Here we handle FS discriminators: |
| 628 | uint32_t DiscriminatorVal = (*Discriminator) & getDiscriminatorMask(); |
| 629 | |
| 630 | for (uint32_t J = 0; J < *NumCalls; ++J) { |
| 631 | auto CalledFunction(readStringFromTable()); |
| 632 | if (std::error_code EC = CalledFunction.getError()) |
| 633 | return EC; |
| 634 | |
| 635 | auto CalledFunctionSamples = readNumber<uint64_t>(); |
| 636 | if (std::error_code EC = CalledFunctionSamples.getError()) |
| 637 | return EC; |
| 638 | |
| 639 | FProfile.addCalledTargetSamples(LineOffset: *LineOffset, Discriminator: DiscriminatorVal, |
| 640 | Func: *CalledFunction, Num: *CalledFunctionSamples); |
| 641 | } |
| 642 | |
| 643 | FProfile.addBodySamples(LineOffset: *LineOffset, Discriminator: DiscriminatorVal, Num: *NumSamples); |
| 644 | } |
| 645 | |
| 646 | // Read all the samples for inlined function calls. |
| 647 | auto NumCallsites = readNumber<uint32_t>(); |
| 648 | if (std::error_code EC = NumCallsites.getError()) |
| 649 | return EC; |
| 650 | |
| 651 | for (uint32_t J = 0; J < *NumCallsites; ++J) { |
| 652 | auto LineOffset = readNumber<uint64_t>(); |
| 653 | if (std::error_code EC = LineOffset.getError()) |
| 654 | return EC; |
| 655 | |
| 656 | auto Discriminator = readNumber<uint64_t>(); |
| 657 | if (std::error_code EC = Discriminator.getError()) |
| 658 | return EC; |
| 659 | |
| 660 | auto FName(readStringFromTable()); |
| 661 | if (std::error_code EC = FName.getError()) |
| 662 | return EC; |
| 663 | |
| 664 | // Here we handle FS discriminators: |
| 665 | uint32_t DiscriminatorVal = (*Discriminator) & getDiscriminatorMask(); |
| 666 | |
| 667 | FunctionSamples &CalleeProfile = FProfile.functionSamplesAt( |
| 668 | Loc: LineLocation(*LineOffset, DiscriminatorVal))[*FName]; |
| 669 | CalleeProfile.setFunction(*FName); |
| 670 | if (std::error_code EC = readProfile(FProfile&: CalleeProfile)) |
| 671 | return EC; |
| 672 | } |
| 673 | |
| 674 | return sampleprof_error::success; |
| 675 | } |
| 676 | |
| 677 | std::error_code |
| 678 | SampleProfileReaderBinary::readFuncProfile(const uint8_t *Start, |
| 679 | SampleProfileMap &Profiles) { |
| 680 | Data = Start; |
| 681 | auto NumHeadSamples = readNumber<uint64_t>(); |
| 682 | if (std::error_code EC = NumHeadSamples.getError()) |
| 683 | return EC; |
| 684 | |
| 685 | auto FContextHash(readSampleContextFromTable()); |
| 686 | if (std::error_code EC = FContextHash.getError()) |
| 687 | return EC; |
| 688 | |
| 689 | auto &[FContext, Hash] = *FContextHash; |
| 690 | // Use the cached hash value for insertion instead of recalculating it. |
| 691 | auto Res = Profiles.try_emplace(Hash, Key: FContext, Args: FunctionSamples()); |
| 692 | FunctionSamples &FProfile = Res.first->second; |
| 693 | FProfile.setContext(FContext); |
| 694 | FProfile.addHeadSamples(Num: *NumHeadSamples); |
| 695 | |
| 696 | if (FContext.hasContext()) |
| 697 | CSProfileCount++; |
| 698 | |
| 699 | if (std::error_code EC = readProfile(FProfile)) |
| 700 | return EC; |
| 701 | return sampleprof_error::success; |
| 702 | } |
| 703 | |
| 704 | std::error_code |
| 705 | SampleProfileReaderBinary::readFuncProfile(const uint8_t *Start) { |
| 706 | return readFuncProfile(Start, Profiles); |
| 707 | } |
| 708 | |
| 709 | std::error_code SampleProfileReaderBinary::readImpl() { |
| 710 | ProfileIsFS = ProfileIsFSDisciminator; |
| 711 | FunctionSamples::ProfileIsFS = ProfileIsFS; |
| 712 | while (Data < End) { |
| 713 | if (std::error_code EC = readFuncProfile(Start: Data)) |
| 714 | return EC; |
| 715 | } |
| 716 | |
| 717 | return sampleprof_error::success; |
| 718 | } |
| 719 | |
| 720 | std::error_code SampleProfileReaderExtBinaryBase::readOneSection( |
| 721 | const uint8_t *Start, uint64_t Size, const SecHdrTableEntry &Entry) { |
| 722 | Data = Start; |
| 723 | End = Start + Size; |
| 724 | switch (Entry.Type) { |
| 725 | case SecProfSummary: |
| 726 | if (std::error_code EC = readSummary()) |
| 727 | return EC; |
| 728 | if (hasSecFlag(Entry, Flag: SecProfSummaryFlags::SecFlagPartial)) |
| 729 | Summary->setPartialProfile(true); |
| 730 | if (hasSecFlag(Entry, Flag: SecProfSummaryFlags::SecFlagFullContext)) |
| 731 | FunctionSamples::ProfileIsCS = ProfileIsCS = true; |
| 732 | if (hasSecFlag(Entry, Flag: SecProfSummaryFlags::SecFlagIsPreInlined)) |
| 733 | FunctionSamples::ProfileIsPreInlined = ProfileIsPreInlined = true; |
| 734 | if (hasSecFlag(Entry, Flag: SecProfSummaryFlags::SecFlagFSDiscriminator)) |
| 735 | FunctionSamples::ProfileIsFS = ProfileIsFS = true; |
| 736 | break; |
| 737 | case SecNameTable: { |
| 738 | bool FixedLengthMD5 = |
| 739 | hasSecFlag(Entry, Flag: SecNameTableFlags::SecFlagFixedLengthMD5); |
| 740 | bool UseMD5 = hasSecFlag(Entry, Flag: SecNameTableFlags::SecFlagMD5Name); |
| 741 | // UseMD5 means if THIS section uses MD5, ProfileIsMD5 means if the entire |
| 742 | // profile uses MD5 for function name matching in IPO passes. |
| 743 | ProfileIsMD5 = ProfileIsMD5 || UseMD5; |
| 744 | FunctionSamples::HasUniqSuffix = |
| 745 | hasSecFlag(Entry, Flag: SecNameTableFlags::SecFlagUniqSuffix); |
| 746 | if (std::error_code EC = readNameTableSec(IsMD5: UseMD5, FixedLengthMD5)) |
| 747 | return EC; |
| 748 | break; |
| 749 | } |
| 750 | case SecCSNameTable: { |
| 751 | if (std::error_code EC = readCSNameTableSec()) |
| 752 | return EC; |
| 753 | break; |
| 754 | } |
| 755 | case SecLBRProfile: |
| 756 | ProfileSecRange = std::make_pair(x&: Data, y&: End); |
| 757 | if (std::error_code EC = readFuncProfiles()) |
| 758 | return EC; |
| 759 | break; |
| 760 | case SecFuncOffsetTable: |
| 761 | // If module is absent, we are using LLVM tools, and need to read all |
| 762 | // profiles, so skip reading the function offset table. |
| 763 | if (!M) { |
| 764 | Data = End; |
| 765 | } else { |
| 766 | assert((!ProfileIsCS || |
| 767 | hasSecFlag(Entry, SecFuncOffsetFlags::SecFlagOrdered)) && |
| 768 | "func offset table should always be sorted in CS profile" ); |
| 769 | if (std::error_code EC = readFuncOffsetTable()) |
| 770 | return EC; |
| 771 | } |
| 772 | break; |
| 773 | case SecFuncMetadata: { |
| 774 | ProfileIsProbeBased = |
| 775 | hasSecFlag(Entry, Flag: SecFuncMetadataFlags::SecFlagIsProbeBased); |
| 776 | FunctionSamples::ProfileIsProbeBased = ProfileIsProbeBased; |
| 777 | ProfileHasAttribute = |
| 778 | hasSecFlag(Entry, Flag: SecFuncMetadataFlags::SecFlagHasAttribute); |
| 779 | if (std::error_code EC = readFuncMetadata(ProfileHasAttribute)) |
| 780 | return EC; |
| 781 | break; |
| 782 | } |
| 783 | case SecProfileSymbolList: |
| 784 | if (std::error_code EC = readProfileSymbolList()) |
| 785 | return EC; |
| 786 | break; |
| 787 | default: |
| 788 | if (std::error_code EC = readCustomSection(Entry)) |
| 789 | return EC; |
| 790 | break; |
| 791 | } |
| 792 | return sampleprof_error::success; |
| 793 | } |
| 794 | |
| 795 | bool SampleProfileReaderExtBinaryBase::useFuncOffsetList() const { |
| 796 | // If profile is CS, the function offset section is expected to consist of |
| 797 | // sequences of contexts in pre-order layout |
| 798 | // (e.g. [A, A:1 @ B, A:1 @ B:2.3 @ C] [D, D:1 @ E]), so that when a matched |
| 799 | // context in the module is found, the profiles of all its callees are |
| 800 | // recursively loaded. A list is needed since the order of profiles matters. |
| 801 | if (ProfileIsCS) |
| 802 | return true; |
| 803 | |
| 804 | // If the profile is MD5, use the map container to lookup functions in |
| 805 | // the module. A remapper has no use on MD5 names. |
| 806 | if (useMD5()) |
| 807 | return false; |
| 808 | |
| 809 | // Profile is not MD5 and if a remapper is present, the remapped name of |
| 810 | // every function needed to be matched against the module, so use the list |
| 811 | // container since each entry is accessed. |
| 812 | if (Remapper) |
| 813 | return true; |
| 814 | |
| 815 | // Otherwise use the map container for faster lookup. |
| 816 | // TODO: If the cardinality of the function offset section is much smaller |
| 817 | // than the number of functions in the module, using the list container can |
| 818 | // be always faster, but we need to figure out the constant factor to |
| 819 | // determine the cutoff. |
| 820 | return false; |
| 821 | } |
| 822 | |
| 823 | std::error_code |
| 824 | SampleProfileReaderExtBinaryBase::read(const DenseSet<StringRef> &FuncsToUse, |
| 825 | SampleProfileMap &Profiles) { |
| 826 | if (FuncsToUse.empty()) |
| 827 | return sampleprof_error::success; |
| 828 | |
| 829 | Data = ProfileSecRange.first; |
| 830 | End = ProfileSecRange.second; |
| 831 | if (std::error_code EC = readFuncProfiles(FuncsToUse, Profiles)) |
| 832 | return EC; |
| 833 | End = Data; |
| 834 | DenseSet<FunctionSamples *> ProfilesToReadMetadata; |
| 835 | for (auto FName : FuncsToUse) { |
| 836 | auto I = Profiles.find(Ctx: FName); |
| 837 | if (I != Profiles.end()) |
| 838 | ProfilesToReadMetadata.insert(V: &I->second); |
| 839 | } |
| 840 | |
| 841 | if (std::error_code EC = |
| 842 | readFuncMetadata(ProfileHasAttribute, Profiles&: ProfilesToReadMetadata)) |
| 843 | return EC; |
| 844 | return sampleprof_error::success; |
| 845 | } |
| 846 | |
| 847 | bool SampleProfileReaderExtBinaryBase::collectFuncsFromModule() { |
| 848 | if (!M) |
| 849 | return false; |
| 850 | FuncsToUse.clear(); |
| 851 | for (auto &F : *M) |
| 852 | FuncsToUse.insert(V: FunctionSamples::getCanonicalFnName(F)); |
| 853 | return true; |
| 854 | } |
| 855 | |
| 856 | std::error_code SampleProfileReaderExtBinaryBase::readFuncOffsetTable() { |
| 857 | // If there are more than one function offset section, the profile associated |
| 858 | // with the previous section has to be done reading before next one is read. |
| 859 | FuncOffsetTable.clear(); |
| 860 | FuncOffsetList.clear(); |
| 861 | |
| 862 | auto Size = readNumber<uint64_t>(); |
| 863 | if (std::error_code EC = Size.getError()) |
| 864 | return EC; |
| 865 | |
| 866 | bool UseFuncOffsetList = useFuncOffsetList(); |
| 867 | if (UseFuncOffsetList) |
| 868 | FuncOffsetList.reserve(n: *Size); |
| 869 | else |
| 870 | FuncOffsetTable.reserve(NumEntries: *Size); |
| 871 | |
| 872 | for (uint64_t I = 0; I < *Size; ++I) { |
| 873 | auto FContextHash(readSampleContextFromTable()); |
| 874 | if (std::error_code EC = FContextHash.getError()) |
| 875 | return EC; |
| 876 | |
| 877 | auto &[FContext, Hash] = *FContextHash; |
| 878 | auto Offset = readNumber<uint64_t>(); |
| 879 | if (std::error_code EC = Offset.getError()) |
| 880 | return EC; |
| 881 | |
| 882 | if (UseFuncOffsetList) |
| 883 | FuncOffsetList.emplace_back(args&: FContext, args&: *Offset); |
| 884 | else |
| 885 | // Because Porfiles replace existing value with new value if collision |
| 886 | // happens, we also use the latest offset so that they are consistent. |
| 887 | FuncOffsetTable[Hash] = *Offset; |
| 888 | } |
| 889 | |
| 890 | return sampleprof_error::success; |
| 891 | } |
| 892 | |
| 893 | std::error_code SampleProfileReaderExtBinaryBase::readFuncProfiles( |
| 894 | const DenseSet<StringRef> &FuncsToUse, SampleProfileMap &Profiles) { |
| 895 | const uint8_t *Start = Data; |
| 896 | |
| 897 | if (Remapper) { |
| 898 | for (auto Name : FuncsToUse) { |
| 899 | Remapper->insert(FunctionName: Name); |
| 900 | } |
| 901 | } |
| 902 | |
| 903 | if (ProfileIsCS) { |
| 904 | assert(useFuncOffsetList()); |
| 905 | DenseSet<uint64_t> FuncGuidsToUse; |
| 906 | if (useMD5()) { |
| 907 | for (auto Name : FuncsToUse) |
| 908 | FuncGuidsToUse.insert(V: Function::getGUIDAssumingExternalLinkage(GlobalName: Name)); |
| 909 | } |
| 910 | |
| 911 | // For each function in current module, load all context profiles for |
| 912 | // the function as well as their callee contexts which can help profile |
| 913 | // guided importing for ThinLTO. This can be achieved by walking |
| 914 | // through an ordered context container, where contexts are laid out |
| 915 | // as if they were walked in preorder of a context trie. While |
| 916 | // traversing the trie, a link to the highest common ancestor node is |
| 917 | // kept so that all of its decendants will be loaded. |
| 918 | const SampleContext *CommonContext = nullptr; |
| 919 | for (const auto &NameOffset : FuncOffsetList) { |
| 920 | const auto &FContext = NameOffset.first; |
| 921 | FunctionId FName = FContext.getFunction(); |
| 922 | StringRef FNameString; |
| 923 | if (!useMD5()) |
| 924 | FNameString = FName.stringRef(); |
| 925 | |
| 926 | // For function in the current module, keep its farthest ancestor |
| 927 | // context. This can be used to load itself and its child and |
| 928 | // sibling contexts. |
| 929 | if ((useMD5() && FuncGuidsToUse.count(V: FName.getHashCode())) || |
| 930 | (!useMD5() && (FuncsToUse.count(V: FNameString) || |
| 931 | (Remapper && Remapper->exist(FunctionName: FNameString))))) { |
| 932 | if (!CommonContext || !CommonContext->isPrefixOf(That: FContext)) |
| 933 | CommonContext = &FContext; |
| 934 | } |
| 935 | |
| 936 | if (CommonContext == &FContext || |
| 937 | (CommonContext && CommonContext->isPrefixOf(That: FContext))) { |
| 938 | // Load profile for the current context which originated from |
| 939 | // the common ancestor. |
| 940 | const uint8_t *FuncProfileAddr = Start + NameOffset.second; |
| 941 | if (std::error_code EC = readFuncProfile(Start: FuncProfileAddr)) |
| 942 | return EC; |
| 943 | } |
| 944 | } |
| 945 | } else if (useMD5()) { |
| 946 | assert(!useFuncOffsetList()); |
| 947 | for (auto Name : FuncsToUse) { |
| 948 | auto GUID = MD5Hash(Str: Name); |
| 949 | auto iter = FuncOffsetTable.find(Val: GUID); |
| 950 | if (iter == FuncOffsetTable.end()) |
| 951 | continue; |
| 952 | const uint8_t *FuncProfileAddr = Start + iter->second; |
| 953 | if (std::error_code EC = readFuncProfile(Start: FuncProfileAddr, Profiles)) |
| 954 | return EC; |
| 955 | } |
| 956 | } else if (Remapper) { |
| 957 | assert(useFuncOffsetList()); |
| 958 | for (auto NameOffset : FuncOffsetList) { |
| 959 | SampleContext FContext(NameOffset.first); |
| 960 | auto FuncName = FContext.getFunction(); |
| 961 | StringRef FuncNameStr = FuncName.stringRef(); |
| 962 | if (!FuncsToUse.count(V: FuncNameStr) && !Remapper->exist(FunctionName: FuncNameStr)) |
| 963 | continue; |
| 964 | const uint8_t *FuncProfileAddr = Start + NameOffset.second; |
| 965 | if (std::error_code EC = readFuncProfile(Start: FuncProfileAddr, Profiles)) |
| 966 | return EC; |
| 967 | } |
| 968 | } else { |
| 969 | assert(!useFuncOffsetList()); |
| 970 | for (auto Name : FuncsToUse) { |
| 971 | |
| 972 | auto iter = FuncOffsetTable.find(Val: MD5Hash(Str: Name)); |
| 973 | if (iter == FuncOffsetTable.end()) |
| 974 | continue; |
| 975 | const uint8_t *FuncProfileAddr = Start + iter->second; |
| 976 | if (std::error_code EC = readFuncProfile(Start: FuncProfileAddr, Profiles)) |
| 977 | return EC; |
| 978 | } |
| 979 | } |
| 980 | |
| 981 | return sampleprof_error::success; |
| 982 | } |
| 983 | |
| 984 | std::error_code SampleProfileReaderExtBinaryBase::readFuncProfiles() { |
| 985 | // Collect functions used by current module if the Reader has been |
| 986 | // given a module. |
| 987 | // collectFuncsFromModule uses FunctionSamples::getCanonicalFnName |
| 988 | // which will query FunctionSamples::HasUniqSuffix, so it has to be |
| 989 | // called after FunctionSamples::HasUniqSuffix is set, i.e. after |
| 990 | // NameTable section is read. |
| 991 | bool LoadFuncsToBeUsed = collectFuncsFromModule(); |
| 992 | |
| 993 | // When LoadFuncsToBeUsed is false, we are using LLVM tool, need to read all |
| 994 | // profiles. |
| 995 | if (!LoadFuncsToBeUsed) { |
| 996 | while (Data < End) { |
| 997 | if (std::error_code EC = readFuncProfile(Start: Data)) |
| 998 | return EC; |
| 999 | } |
| 1000 | assert(Data == End && "More data is read than expected" ); |
| 1001 | } else { |
| 1002 | // Load function profiles on demand. |
| 1003 | if (std::error_code EC = readFuncProfiles(FuncsToUse, Profiles)) |
| 1004 | return EC; |
| 1005 | Data = End; |
| 1006 | } |
| 1007 | assert((CSProfileCount == 0 || CSProfileCount == Profiles.size()) && |
| 1008 | "Cannot have both context-sensitive and regular profile" ); |
| 1009 | assert((!CSProfileCount || ProfileIsCS) && |
| 1010 | "Section flag should be consistent with actual profile" ); |
| 1011 | return sampleprof_error::success; |
| 1012 | } |
| 1013 | |
| 1014 | std::error_code SampleProfileReaderExtBinaryBase::readProfileSymbolList() { |
| 1015 | if (!ProfSymList) |
| 1016 | ProfSymList = std::make_unique<ProfileSymbolList>(); |
| 1017 | |
| 1018 | if (std::error_code EC = ProfSymList->read(Data, ListSize: End - Data)) |
| 1019 | return EC; |
| 1020 | |
| 1021 | Data = End; |
| 1022 | return sampleprof_error::success; |
| 1023 | } |
| 1024 | |
| 1025 | std::error_code SampleProfileReaderExtBinaryBase::decompressSection( |
| 1026 | const uint8_t *SecStart, const uint64_t SecSize, |
| 1027 | const uint8_t *&DecompressBuf, uint64_t &DecompressBufSize) { |
| 1028 | Data = SecStart; |
| 1029 | End = SecStart + SecSize; |
| 1030 | auto DecompressSize = readNumber<uint64_t>(); |
| 1031 | if (std::error_code EC = DecompressSize.getError()) |
| 1032 | return EC; |
| 1033 | DecompressBufSize = *DecompressSize; |
| 1034 | |
| 1035 | auto CompressSize = readNumber<uint64_t>(); |
| 1036 | if (std::error_code EC = CompressSize.getError()) |
| 1037 | return EC; |
| 1038 | |
| 1039 | if (!llvm::compression::zlib::isAvailable()) |
| 1040 | return sampleprof_error::zlib_unavailable; |
| 1041 | |
| 1042 | uint8_t *Buffer = Allocator.Allocate<uint8_t>(Num: DecompressBufSize); |
| 1043 | size_t UCSize = DecompressBufSize; |
| 1044 | llvm::Error E = compression::zlib::decompress(Input: ArrayRef(Data, *CompressSize), |
| 1045 | Output: Buffer, UncompressedSize&: UCSize); |
| 1046 | if (E) |
| 1047 | return sampleprof_error::uncompress_failed; |
| 1048 | DecompressBuf = reinterpret_cast<const uint8_t *>(Buffer); |
| 1049 | return sampleprof_error::success; |
| 1050 | } |
| 1051 | |
| 1052 | std::error_code SampleProfileReaderExtBinaryBase::readImpl() { |
| 1053 | const uint8_t *BufStart = |
| 1054 | reinterpret_cast<const uint8_t *>(Buffer->getBufferStart()); |
| 1055 | |
| 1056 | for (auto &Entry : SecHdrTable) { |
| 1057 | // Skip empty section. |
| 1058 | if (!Entry.Size) |
| 1059 | continue; |
| 1060 | |
| 1061 | // Skip sections without inlined functions when SkipFlatProf is true. |
| 1062 | if (SkipFlatProf && hasSecFlag(Entry, Flag: SecCommonFlags::SecFlagFlat)) |
| 1063 | continue; |
| 1064 | |
| 1065 | const uint8_t *SecStart = BufStart + Entry.Offset; |
| 1066 | uint64_t SecSize = Entry.Size; |
| 1067 | |
| 1068 | // If the section is compressed, decompress it into a buffer |
| 1069 | // DecompressBuf before reading the actual data. The pointee of |
| 1070 | // 'Data' will be changed to buffer hold by DecompressBuf |
| 1071 | // temporarily when reading the actual data. |
| 1072 | bool isCompressed = hasSecFlag(Entry, Flag: SecCommonFlags::SecFlagCompress); |
| 1073 | if (isCompressed) { |
| 1074 | const uint8_t *DecompressBuf; |
| 1075 | uint64_t DecompressBufSize; |
| 1076 | if (std::error_code EC = decompressSection( |
| 1077 | SecStart, SecSize, DecompressBuf, DecompressBufSize)) |
| 1078 | return EC; |
| 1079 | SecStart = DecompressBuf; |
| 1080 | SecSize = DecompressBufSize; |
| 1081 | } |
| 1082 | |
| 1083 | if (std::error_code EC = readOneSection(Start: SecStart, Size: SecSize, Entry)) |
| 1084 | return EC; |
| 1085 | if (Data != SecStart + SecSize) |
| 1086 | return sampleprof_error::malformed; |
| 1087 | |
| 1088 | // Change the pointee of 'Data' from DecompressBuf to original Buffer. |
| 1089 | if (isCompressed) { |
| 1090 | Data = BufStart + Entry.Offset; |
| 1091 | End = BufStart + Buffer->getBufferSize(); |
| 1092 | } |
| 1093 | } |
| 1094 | |
| 1095 | return sampleprof_error::success; |
| 1096 | } |
| 1097 | |
| 1098 | std::error_code SampleProfileReaderRawBinary::verifySPMagic(uint64_t Magic) { |
| 1099 | if (Magic == SPMagic()) |
| 1100 | return sampleprof_error::success; |
| 1101 | return sampleprof_error::bad_magic; |
| 1102 | } |
| 1103 | |
| 1104 | std::error_code SampleProfileReaderExtBinary::verifySPMagic(uint64_t Magic) { |
| 1105 | if (Magic == SPMagic(Format: SPF_Ext_Binary)) |
| 1106 | return sampleprof_error::success; |
| 1107 | return sampleprof_error::bad_magic; |
| 1108 | } |
| 1109 | |
| 1110 | std::error_code SampleProfileReaderBinary::readNameTable() { |
| 1111 | auto Size = readNumber<size_t>(); |
| 1112 | if (std::error_code EC = Size.getError()) |
| 1113 | return EC; |
| 1114 | |
| 1115 | // Normally if useMD5 is true, the name table should have MD5 values, not |
| 1116 | // strings, however in the case that ExtBinary profile has multiple name |
| 1117 | // tables mixing string and MD5, all of them have to be normalized to use MD5, |
| 1118 | // because optimization passes can only handle either type. |
| 1119 | bool UseMD5 = useMD5(); |
| 1120 | |
| 1121 | NameTable.clear(); |
| 1122 | NameTable.reserve(n: *Size); |
| 1123 | if (!ProfileIsCS) { |
| 1124 | MD5SampleContextTable.clear(); |
| 1125 | if (UseMD5) |
| 1126 | MD5SampleContextTable.reserve(n: *Size); |
| 1127 | else |
| 1128 | // If we are using strings, delay MD5 computation since only a portion of |
| 1129 | // names are used by top level functions. Use 0 to indicate MD5 value is |
| 1130 | // to be calculated as no known string has a MD5 value of 0. |
| 1131 | MD5SampleContextTable.resize(new_size: *Size); |
| 1132 | } |
| 1133 | for (size_t I = 0; I < *Size; ++I) { |
| 1134 | auto Name(readString()); |
| 1135 | if (std::error_code EC = Name.getError()) |
| 1136 | return EC; |
| 1137 | if (UseMD5) { |
| 1138 | FunctionId FID(*Name); |
| 1139 | if (!ProfileIsCS) |
| 1140 | MD5SampleContextTable.emplace_back(args: FID.getHashCode()); |
| 1141 | NameTable.emplace_back(args&: FID); |
| 1142 | } else |
| 1143 | NameTable.push_back(x: FunctionId(*Name)); |
| 1144 | } |
| 1145 | if (!ProfileIsCS) |
| 1146 | MD5SampleContextStart = MD5SampleContextTable.data(); |
| 1147 | return sampleprof_error::success; |
| 1148 | } |
| 1149 | |
| 1150 | std::error_code |
| 1151 | SampleProfileReaderExtBinaryBase::readNameTableSec(bool IsMD5, |
| 1152 | bool FixedLengthMD5) { |
| 1153 | if (FixedLengthMD5) { |
| 1154 | if (!IsMD5) |
| 1155 | errs() << "If FixedLengthMD5 is true, UseMD5 has to be true" ; |
| 1156 | auto Size = readNumber<size_t>(); |
| 1157 | if (std::error_code EC = Size.getError()) |
| 1158 | return EC; |
| 1159 | |
| 1160 | assert(Data + (*Size) * sizeof(uint64_t) == End && |
| 1161 | "Fixed length MD5 name table does not contain specified number of " |
| 1162 | "entries" ); |
| 1163 | if (Data + (*Size) * sizeof(uint64_t) > End) |
| 1164 | return sampleprof_error::truncated; |
| 1165 | |
| 1166 | NameTable.clear(); |
| 1167 | NameTable.reserve(n: *Size); |
| 1168 | for (size_t I = 0; I < *Size; ++I) { |
| 1169 | using namespace support; |
| 1170 | uint64_t FID = endian::read<uint64_t, endianness::little, unaligned>( |
| 1171 | memory: Data + I * sizeof(uint64_t)); |
| 1172 | NameTable.emplace_back(args: FunctionId(FID)); |
| 1173 | } |
| 1174 | if (!ProfileIsCS) |
| 1175 | MD5SampleContextStart = reinterpret_cast<const uint64_t *>(Data); |
| 1176 | Data = Data + (*Size) * sizeof(uint64_t); |
| 1177 | return sampleprof_error::success; |
| 1178 | } |
| 1179 | |
| 1180 | if (IsMD5) { |
| 1181 | assert(!FixedLengthMD5 && "FixedLengthMD5 should be unreachable here" ); |
| 1182 | auto Size = readNumber<size_t>(); |
| 1183 | if (std::error_code EC = Size.getError()) |
| 1184 | return EC; |
| 1185 | |
| 1186 | NameTable.clear(); |
| 1187 | NameTable.reserve(n: *Size); |
| 1188 | if (!ProfileIsCS) |
| 1189 | MD5SampleContextTable.resize(new_size: *Size); |
| 1190 | for (size_t I = 0; I < *Size; ++I) { |
| 1191 | auto FID = readNumber<uint64_t>(); |
| 1192 | if (std::error_code EC = FID.getError()) |
| 1193 | return EC; |
| 1194 | if (!ProfileIsCS) |
| 1195 | support::endian::write64le(P: &MD5SampleContextTable[I], V: *FID); |
| 1196 | NameTable.emplace_back(args: FunctionId(*FID)); |
| 1197 | } |
| 1198 | if (!ProfileIsCS) |
| 1199 | MD5SampleContextStart = MD5SampleContextTable.data(); |
| 1200 | return sampleprof_error::success; |
| 1201 | } |
| 1202 | |
| 1203 | return SampleProfileReaderBinary::readNameTable(); |
| 1204 | } |
| 1205 | |
| 1206 | // Read in the CS name table section, which basically contains a list of context |
| 1207 | // vectors. Each element of a context vector, aka a frame, refers to the |
| 1208 | // underlying raw function names that are stored in the name table, as well as |
| 1209 | // a callsite identifier that only makes sense for non-leaf frames. |
| 1210 | std::error_code SampleProfileReaderExtBinaryBase::readCSNameTableSec() { |
| 1211 | auto Size = readNumber<size_t>(); |
| 1212 | if (std::error_code EC = Size.getError()) |
| 1213 | return EC; |
| 1214 | |
| 1215 | CSNameTable.clear(); |
| 1216 | CSNameTable.reserve(n: *Size); |
| 1217 | if (ProfileIsCS) { |
| 1218 | // Delay MD5 computation of CS context until they are needed. Use 0 to |
| 1219 | // indicate MD5 value is to be calculated as no known string has a MD5 |
| 1220 | // value of 0. |
| 1221 | MD5SampleContextTable.clear(); |
| 1222 | MD5SampleContextTable.resize(new_size: *Size); |
| 1223 | MD5SampleContextStart = MD5SampleContextTable.data(); |
| 1224 | } |
| 1225 | for (size_t I = 0; I < *Size; ++I) { |
| 1226 | CSNameTable.emplace_back(args: SampleContextFrameVector()); |
| 1227 | auto ContextSize = readNumber<uint32_t>(); |
| 1228 | if (std::error_code EC = ContextSize.getError()) |
| 1229 | return EC; |
| 1230 | for (uint32_t J = 0; J < *ContextSize; ++J) { |
| 1231 | auto FName(readStringFromTable()); |
| 1232 | if (std::error_code EC = FName.getError()) |
| 1233 | return EC; |
| 1234 | auto LineOffset = readNumber<uint64_t>(); |
| 1235 | if (std::error_code EC = LineOffset.getError()) |
| 1236 | return EC; |
| 1237 | |
| 1238 | if (!isOffsetLegal(L: *LineOffset)) |
| 1239 | return std::error_code(); |
| 1240 | |
| 1241 | auto Discriminator = readNumber<uint64_t>(); |
| 1242 | if (std::error_code EC = Discriminator.getError()) |
| 1243 | return EC; |
| 1244 | |
| 1245 | CSNameTable.back().emplace_back( |
| 1246 | Args&: FName.get(), Args: LineLocation(LineOffset.get(), Discriminator.get())); |
| 1247 | } |
| 1248 | } |
| 1249 | |
| 1250 | return sampleprof_error::success; |
| 1251 | } |
| 1252 | |
| 1253 | std::error_code |
| 1254 | SampleProfileReaderExtBinaryBase::readFuncMetadata(bool ProfileHasAttribute, |
| 1255 | FunctionSamples *FProfile) { |
| 1256 | if (Data < End) { |
| 1257 | if (ProfileIsProbeBased) { |
| 1258 | auto Checksum = readNumber<uint64_t>(); |
| 1259 | if (std::error_code EC = Checksum.getError()) |
| 1260 | return EC; |
| 1261 | if (FProfile) |
| 1262 | FProfile->setFunctionHash(*Checksum); |
| 1263 | } |
| 1264 | |
| 1265 | if (ProfileHasAttribute) { |
| 1266 | auto Attributes = readNumber<uint32_t>(); |
| 1267 | if (std::error_code EC = Attributes.getError()) |
| 1268 | return EC; |
| 1269 | if (FProfile) |
| 1270 | FProfile->getContext().setAllAttributes(*Attributes); |
| 1271 | } |
| 1272 | |
| 1273 | if (!ProfileIsCS) { |
| 1274 | // Read all the attributes for inlined function calls. |
| 1275 | auto NumCallsites = readNumber<uint32_t>(); |
| 1276 | if (std::error_code EC = NumCallsites.getError()) |
| 1277 | return EC; |
| 1278 | |
| 1279 | for (uint32_t J = 0; J < *NumCallsites; ++J) { |
| 1280 | auto LineOffset = readNumber<uint64_t>(); |
| 1281 | if (std::error_code EC = LineOffset.getError()) |
| 1282 | return EC; |
| 1283 | |
| 1284 | auto Discriminator = readNumber<uint64_t>(); |
| 1285 | if (std::error_code EC = Discriminator.getError()) |
| 1286 | return EC; |
| 1287 | |
| 1288 | auto FContextHash(readSampleContextFromTable()); |
| 1289 | if (std::error_code EC = FContextHash.getError()) |
| 1290 | return EC; |
| 1291 | |
| 1292 | auto &[FContext, Hash] = *FContextHash; |
| 1293 | FunctionSamples *CalleeProfile = nullptr; |
| 1294 | if (FProfile) { |
| 1295 | CalleeProfile = const_cast<FunctionSamples *>( |
| 1296 | &FProfile->functionSamplesAt(Loc: LineLocation( |
| 1297 | *LineOffset, |
| 1298 | *Discriminator))[FContext.getFunction()]); |
| 1299 | } |
| 1300 | if (std::error_code EC = |
| 1301 | readFuncMetadata(ProfileHasAttribute, FProfile: CalleeProfile)) |
| 1302 | return EC; |
| 1303 | } |
| 1304 | } |
| 1305 | } |
| 1306 | |
| 1307 | return sampleprof_error::success; |
| 1308 | } |
| 1309 | |
| 1310 | std::error_code SampleProfileReaderExtBinaryBase::readFuncMetadata( |
| 1311 | bool ProfileHasAttribute, DenseSet<FunctionSamples *> &Profiles) { |
| 1312 | if (FuncMetadataIndex.empty()) |
| 1313 | return sampleprof_error::success; |
| 1314 | |
| 1315 | for (auto *FProfile : Profiles) { |
| 1316 | auto R = FuncMetadataIndex.find(x: FProfile->getContext().getHashCode()); |
| 1317 | if (R == FuncMetadataIndex.end()) |
| 1318 | continue; |
| 1319 | |
| 1320 | Data = R->second.first; |
| 1321 | End = R->second.second; |
| 1322 | if (std::error_code EC = readFuncMetadata(ProfileHasAttribute, FProfile)) |
| 1323 | return EC; |
| 1324 | assert(Data == End && "More data is read than expected" ); |
| 1325 | } |
| 1326 | return sampleprof_error::success; |
| 1327 | } |
| 1328 | |
| 1329 | std::error_code |
| 1330 | SampleProfileReaderExtBinaryBase::readFuncMetadata(bool ProfileHasAttribute) { |
| 1331 | while (Data < End) { |
| 1332 | auto FContextHash(readSampleContextFromTable()); |
| 1333 | if (std::error_code EC = FContextHash.getError()) |
| 1334 | return EC; |
| 1335 | auto &[FContext, Hash] = *FContextHash; |
| 1336 | FunctionSamples *FProfile = nullptr; |
| 1337 | auto It = Profiles.find(Ctx: FContext); |
| 1338 | if (It != Profiles.end()) |
| 1339 | FProfile = &It->second; |
| 1340 | |
| 1341 | const uint8_t *Start = Data; |
| 1342 | if (std::error_code EC = readFuncMetadata(ProfileHasAttribute, FProfile)) |
| 1343 | return EC; |
| 1344 | |
| 1345 | FuncMetadataIndex[FContext.getHashCode()] = {Start, Data}; |
| 1346 | } |
| 1347 | |
| 1348 | assert(Data == End && "More data is read than expected" ); |
| 1349 | return sampleprof_error::success; |
| 1350 | } |
| 1351 | |
| 1352 | std::error_code |
| 1353 | SampleProfileReaderExtBinaryBase::readSecHdrTableEntry(uint64_t Idx) { |
| 1354 | SecHdrTableEntry Entry; |
| 1355 | auto Type = readUnencodedNumber<uint64_t>(); |
| 1356 | if (std::error_code EC = Type.getError()) |
| 1357 | return EC; |
| 1358 | Entry.Type = static_cast<SecType>(*Type); |
| 1359 | |
| 1360 | auto Flags = readUnencodedNumber<uint64_t>(); |
| 1361 | if (std::error_code EC = Flags.getError()) |
| 1362 | return EC; |
| 1363 | Entry.Flags = *Flags; |
| 1364 | |
| 1365 | auto Offset = readUnencodedNumber<uint64_t>(); |
| 1366 | if (std::error_code EC = Offset.getError()) |
| 1367 | return EC; |
| 1368 | Entry.Offset = *Offset; |
| 1369 | |
| 1370 | auto Size = readUnencodedNumber<uint64_t>(); |
| 1371 | if (std::error_code EC = Size.getError()) |
| 1372 | return EC; |
| 1373 | Entry.Size = *Size; |
| 1374 | |
| 1375 | Entry.LayoutIndex = Idx; |
| 1376 | SecHdrTable.push_back(x: std::move(Entry)); |
| 1377 | return sampleprof_error::success; |
| 1378 | } |
| 1379 | |
| 1380 | std::error_code SampleProfileReaderExtBinaryBase::readSecHdrTable() { |
| 1381 | auto EntryNum = readUnencodedNumber<uint64_t>(); |
| 1382 | if (std::error_code EC = EntryNum.getError()) |
| 1383 | return EC; |
| 1384 | |
| 1385 | for (uint64_t i = 0; i < (*EntryNum); i++) |
| 1386 | if (std::error_code EC = readSecHdrTableEntry(Idx: i)) |
| 1387 | return EC; |
| 1388 | |
| 1389 | return sampleprof_error::success; |
| 1390 | } |
| 1391 | |
| 1392 | std::error_code SampleProfileReaderExtBinaryBase::() { |
| 1393 | const uint8_t *BufStart = |
| 1394 | reinterpret_cast<const uint8_t *>(Buffer->getBufferStart()); |
| 1395 | Data = BufStart; |
| 1396 | End = BufStart + Buffer->getBufferSize(); |
| 1397 | |
| 1398 | if (std::error_code EC = readMagicIdent()) |
| 1399 | return EC; |
| 1400 | |
| 1401 | if (std::error_code EC = readSecHdrTable()) |
| 1402 | return EC; |
| 1403 | |
| 1404 | return sampleprof_error::success; |
| 1405 | } |
| 1406 | |
| 1407 | uint64_t SampleProfileReaderExtBinaryBase::getSectionSize(SecType Type) { |
| 1408 | uint64_t Size = 0; |
| 1409 | for (auto &Entry : SecHdrTable) { |
| 1410 | if (Entry.Type == Type) |
| 1411 | Size += Entry.Size; |
| 1412 | } |
| 1413 | return Size; |
| 1414 | } |
| 1415 | |
| 1416 | uint64_t SampleProfileReaderExtBinaryBase::getFileSize() { |
| 1417 | // Sections in SecHdrTable is not necessarily in the same order as |
| 1418 | // sections in the profile because section like FuncOffsetTable needs |
| 1419 | // to be written after section LBRProfile but needs to be read before |
| 1420 | // section LBRProfile, so we cannot simply use the last entry in |
| 1421 | // SecHdrTable to calculate the file size. |
| 1422 | uint64_t FileSize = 0; |
| 1423 | for (auto &Entry : SecHdrTable) { |
| 1424 | FileSize = std::max(a: Entry.Offset + Entry.Size, b: FileSize); |
| 1425 | } |
| 1426 | return FileSize; |
| 1427 | } |
| 1428 | |
| 1429 | static std::string getSecFlagsStr(const SecHdrTableEntry &Entry) { |
| 1430 | std::string Flags; |
| 1431 | if (hasSecFlag(Entry, Flag: SecCommonFlags::SecFlagCompress)) |
| 1432 | Flags.append(s: "{compressed," ); |
| 1433 | else |
| 1434 | Flags.append(s: "{" ); |
| 1435 | |
| 1436 | if (hasSecFlag(Entry, Flag: SecCommonFlags::SecFlagFlat)) |
| 1437 | Flags.append(s: "flat," ); |
| 1438 | |
| 1439 | switch (Entry.Type) { |
| 1440 | case SecNameTable: |
| 1441 | if (hasSecFlag(Entry, Flag: SecNameTableFlags::SecFlagFixedLengthMD5)) |
| 1442 | Flags.append(s: "fixlenmd5," ); |
| 1443 | else if (hasSecFlag(Entry, Flag: SecNameTableFlags::SecFlagMD5Name)) |
| 1444 | Flags.append(s: "md5," ); |
| 1445 | if (hasSecFlag(Entry, Flag: SecNameTableFlags::SecFlagUniqSuffix)) |
| 1446 | Flags.append(s: "uniq," ); |
| 1447 | break; |
| 1448 | case SecProfSummary: |
| 1449 | if (hasSecFlag(Entry, Flag: SecProfSummaryFlags::SecFlagPartial)) |
| 1450 | Flags.append(s: "partial," ); |
| 1451 | if (hasSecFlag(Entry, Flag: SecProfSummaryFlags::SecFlagFullContext)) |
| 1452 | Flags.append(s: "context," ); |
| 1453 | if (hasSecFlag(Entry, Flag: SecProfSummaryFlags::SecFlagIsPreInlined)) |
| 1454 | Flags.append(s: "preInlined," ); |
| 1455 | if (hasSecFlag(Entry, Flag: SecProfSummaryFlags::SecFlagFSDiscriminator)) |
| 1456 | Flags.append(s: "fs-discriminator," ); |
| 1457 | break; |
| 1458 | case SecFuncOffsetTable: |
| 1459 | if (hasSecFlag(Entry, Flag: SecFuncOffsetFlags::SecFlagOrdered)) |
| 1460 | Flags.append(s: "ordered," ); |
| 1461 | break; |
| 1462 | case SecFuncMetadata: |
| 1463 | if (hasSecFlag(Entry, Flag: SecFuncMetadataFlags::SecFlagIsProbeBased)) |
| 1464 | Flags.append(s: "probe," ); |
| 1465 | if (hasSecFlag(Entry, Flag: SecFuncMetadataFlags::SecFlagHasAttribute)) |
| 1466 | Flags.append(s: "attr," ); |
| 1467 | break; |
| 1468 | default: |
| 1469 | break; |
| 1470 | } |
| 1471 | char &last = Flags.back(); |
| 1472 | if (last == ',') |
| 1473 | last = '}'; |
| 1474 | else |
| 1475 | Flags.append(s: "}" ); |
| 1476 | return Flags; |
| 1477 | } |
| 1478 | |
| 1479 | bool SampleProfileReaderExtBinaryBase::dumpSectionInfo(raw_ostream &OS) { |
| 1480 | uint64_t TotalSecsSize = 0; |
| 1481 | for (auto &Entry : SecHdrTable) { |
| 1482 | OS << getSecName(Type: Entry.Type) << " - Offset: " << Entry.Offset |
| 1483 | << ", Size: " << Entry.Size << ", Flags: " << getSecFlagsStr(Entry) |
| 1484 | << "\n" ; |
| 1485 | ; |
| 1486 | TotalSecsSize += Entry.Size; |
| 1487 | } |
| 1488 | uint64_t = SecHdrTable.front().Offset; |
| 1489 | assert(HeaderSize + TotalSecsSize == getFileSize() && |
| 1490 | "Size of 'header + sections' doesn't match the total size of profile" ); |
| 1491 | |
| 1492 | OS << "Header Size: " << HeaderSize << "\n" ; |
| 1493 | OS << "Total Sections Size: " << TotalSecsSize << "\n" ; |
| 1494 | OS << "File Size: " << getFileSize() << "\n" ; |
| 1495 | return true; |
| 1496 | } |
| 1497 | |
| 1498 | std::error_code SampleProfileReaderBinary::readMagicIdent() { |
| 1499 | // Read and check the magic identifier. |
| 1500 | auto Magic = readNumber<uint64_t>(); |
| 1501 | if (std::error_code EC = Magic.getError()) |
| 1502 | return EC; |
| 1503 | else if (std::error_code EC = verifySPMagic(Magic: *Magic)) |
| 1504 | return EC; |
| 1505 | |
| 1506 | // Read the version number. |
| 1507 | auto Version = readNumber<uint64_t>(); |
| 1508 | if (std::error_code EC = Version.getError()) |
| 1509 | return EC; |
| 1510 | else if (*Version != SPVersion()) |
| 1511 | return sampleprof_error::unsupported_version; |
| 1512 | |
| 1513 | return sampleprof_error::success; |
| 1514 | } |
| 1515 | |
| 1516 | std::error_code SampleProfileReaderBinary::() { |
| 1517 | Data = reinterpret_cast<const uint8_t *>(Buffer->getBufferStart()); |
| 1518 | End = Data + Buffer->getBufferSize(); |
| 1519 | |
| 1520 | if (std::error_code EC = readMagicIdent()) |
| 1521 | return EC; |
| 1522 | |
| 1523 | if (std::error_code EC = readSummary()) |
| 1524 | return EC; |
| 1525 | |
| 1526 | if (std::error_code EC = readNameTable()) |
| 1527 | return EC; |
| 1528 | return sampleprof_error::success; |
| 1529 | } |
| 1530 | |
| 1531 | std::error_code SampleProfileReaderBinary::readSummaryEntry( |
| 1532 | std::vector<ProfileSummaryEntry> &Entries) { |
| 1533 | auto Cutoff = readNumber<uint64_t>(); |
| 1534 | if (std::error_code EC = Cutoff.getError()) |
| 1535 | return EC; |
| 1536 | |
| 1537 | auto MinBlockCount = readNumber<uint64_t>(); |
| 1538 | if (std::error_code EC = MinBlockCount.getError()) |
| 1539 | return EC; |
| 1540 | |
| 1541 | auto NumBlocks = readNumber<uint64_t>(); |
| 1542 | if (std::error_code EC = NumBlocks.getError()) |
| 1543 | return EC; |
| 1544 | |
| 1545 | Entries.emplace_back(args&: *Cutoff, args&: *MinBlockCount, args&: *NumBlocks); |
| 1546 | return sampleprof_error::success; |
| 1547 | } |
| 1548 | |
| 1549 | std::error_code SampleProfileReaderBinary::readSummary() { |
| 1550 | auto TotalCount = readNumber<uint64_t>(); |
| 1551 | if (std::error_code EC = TotalCount.getError()) |
| 1552 | return EC; |
| 1553 | |
| 1554 | auto MaxBlockCount = readNumber<uint64_t>(); |
| 1555 | if (std::error_code EC = MaxBlockCount.getError()) |
| 1556 | return EC; |
| 1557 | |
| 1558 | auto MaxFunctionCount = readNumber<uint64_t>(); |
| 1559 | if (std::error_code EC = MaxFunctionCount.getError()) |
| 1560 | return EC; |
| 1561 | |
| 1562 | auto NumBlocks = readNumber<uint64_t>(); |
| 1563 | if (std::error_code EC = NumBlocks.getError()) |
| 1564 | return EC; |
| 1565 | |
| 1566 | auto NumFunctions = readNumber<uint64_t>(); |
| 1567 | if (std::error_code EC = NumFunctions.getError()) |
| 1568 | return EC; |
| 1569 | |
| 1570 | auto NumSummaryEntries = readNumber<uint64_t>(); |
| 1571 | if (std::error_code EC = NumSummaryEntries.getError()) |
| 1572 | return EC; |
| 1573 | |
| 1574 | std::vector<ProfileSummaryEntry> Entries; |
| 1575 | for (unsigned i = 0; i < *NumSummaryEntries; i++) { |
| 1576 | std::error_code EC = readSummaryEntry(Entries); |
| 1577 | if (EC != sampleprof_error::success) |
| 1578 | return EC; |
| 1579 | } |
| 1580 | Summary = std::make_unique<ProfileSummary>( |
| 1581 | args: ProfileSummary::PSK_Sample, args&: Entries, args&: *TotalCount, args&: *MaxBlockCount, args: 0, |
| 1582 | args&: *MaxFunctionCount, args&: *NumBlocks, args&: *NumFunctions); |
| 1583 | |
| 1584 | return sampleprof_error::success; |
| 1585 | } |
| 1586 | |
| 1587 | bool SampleProfileReaderRawBinary::hasFormat(const MemoryBuffer &Buffer) { |
| 1588 | const uint8_t *Data = |
| 1589 | reinterpret_cast<const uint8_t *>(Buffer.getBufferStart()); |
| 1590 | uint64_t Magic = decodeULEB128(p: Data); |
| 1591 | return Magic == SPMagic(); |
| 1592 | } |
| 1593 | |
| 1594 | bool SampleProfileReaderExtBinary::hasFormat(const MemoryBuffer &Buffer) { |
| 1595 | const uint8_t *Data = |
| 1596 | reinterpret_cast<const uint8_t *>(Buffer.getBufferStart()); |
| 1597 | uint64_t Magic = decodeULEB128(p: Data); |
| 1598 | return Magic == SPMagic(Format: SPF_Ext_Binary); |
| 1599 | } |
| 1600 | |
| 1601 | std::error_code SampleProfileReaderGCC::skipNextWord() { |
| 1602 | uint32_t dummy; |
| 1603 | if (!GcovBuffer.readInt(Val&: dummy)) |
| 1604 | return sampleprof_error::truncated; |
| 1605 | return sampleprof_error::success; |
| 1606 | } |
| 1607 | |
| 1608 | template <typename T> ErrorOr<T> SampleProfileReaderGCC::readNumber() { |
| 1609 | if (sizeof(T) <= sizeof(uint32_t)) { |
| 1610 | uint32_t Val; |
| 1611 | if (GcovBuffer.readInt(Val) && Val <= std::numeric_limits<T>::max()) |
| 1612 | return static_cast<T>(Val); |
| 1613 | } else if (sizeof(T) <= sizeof(uint64_t)) { |
| 1614 | uint64_t Val; |
| 1615 | if (GcovBuffer.readInt64(Val) && Val <= std::numeric_limits<T>::max()) |
| 1616 | return static_cast<T>(Val); |
| 1617 | } |
| 1618 | |
| 1619 | std::error_code EC = sampleprof_error::malformed; |
| 1620 | reportError(LineNumber: 0, Msg: EC.message()); |
| 1621 | return EC; |
| 1622 | } |
| 1623 | |
| 1624 | ErrorOr<StringRef> SampleProfileReaderGCC::readString() { |
| 1625 | StringRef Str; |
| 1626 | if (!GcovBuffer.readString(str&: Str)) |
| 1627 | return sampleprof_error::truncated; |
| 1628 | return Str; |
| 1629 | } |
| 1630 | |
| 1631 | std::error_code SampleProfileReaderGCC::() { |
| 1632 | // Read the magic identifier. |
| 1633 | if (!GcovBuffer.readGCDAFormat()) |
| 1634 | return sampleprof_error::unrecognized_format; |
| 1635 | |
| 1636 | // Read the version number. Note - the GCC reader does not validate this |
| 1637 | // version, but the profile creator generates v704. |
| 1638 | GCOV::GCOVVersion version; |
| 1639 | if (!GcovBuffer.readGCOVVersion(version)) |
| 1640 | return sampleprof_error::unrecognized_format; |
| 1641 | |
| 1642 | if (version != GCOV::V407) |
| 1643 | return sampleprof_error::unsupported_version; |
| 1644 | |
| 1645 | // Skip the empty integer. |
| 1646 | if (std::error_code EC = skipNextWord()) |
| 1647 | return EC; |
| 1648 | |
| 1649 | return sampleprof_error::success; |
| 1650 | } |
| 1651 | |
| 1652 | std::error_code SampleProfileReaderGCC::readSectionTag(uint32_t Expected) { |
| 1653 | uint32_t Tag; |
| 1654 | if (!GcovBuffer.readInt(Val&: Tag)) |
| 1655 | return sampleprof_error::truncated; |
| 1656 | |
| 1657 | if (Tag != Expected) |
| 1658 | return sampleprof_error::malformed; |
| 1659 | |
| 1660 | if (std::error_code EC = skipNextWord()) |
| 1661 | return EC; |
| 1662 | |
| 1663 | return sampleprof_error::success; |
| 1664 | } |
| 1665 | |
| 1666 | std::error_code SampleProfileReaderGCC::readNameTable() { |
| 1667 | if (std::error_code EC = readSectionTag(Expected: GCOVTagAFDOFileNames)) |
| 1668 | return EC; |
| 1669 | |
| 1670 | uint32_t Size; |
| 1671 | if (!GcovBuffer.readInt(Val&: Size)) |
| 1672 | return sampleprof_error::truncated; |
| 1673 | |
| 1674 | for (uint32_t I = 0; I < Size; ++I) { |
| 1675 | StringRef Str; |
| 1676 | if (!GcovBuffer.readString(str&: Str)) |
| 1677 | return sampleprof_error::truncated; |
| 1678 | Names.push_back(x: std::string(Str)); |
| 1679 | } |
| 1680 | |
| 1681 | return sampleprof_error::success; |
| 1682 | } |
| 1683 | |
| 1684 | std::error_code SampleProfileReaderGCC::readFunctionProfiles() { |
| 1685 | if (std::error_code EC = readSectionTag(Expected: GCOVTagAFDOFunction)) |
| 1686 | return EC; |
| 1687 | |
| 1688 | uint32_t NumFunctions; |
| 1689 | if (!GcovBuffer.readInt(Val&: NumFunctions)) |
| 1690 | return sampleprof_error::truncated; |
| 1691 | |
| 1692 | InlineCallStack Stack; |
| 1693 | for (uint32_t I = 0; I < NumFunctions; ++I) |
| 1694 | if (std::error_code EC = readOneFunctionProfile(InlineStack: Stack, Update: true, Offset: 0)) |
| 1695 | return EC; |
| 1696 | |
| 1697 | computeSummary(); |
| 1698 | return sampleprof_error::success; |
| 1699 | } |
| 1700 | |
| 1701 | std::error_code SampleProfileReaderGCC::readOneFunctionProfile( |
| 1702 | const InlineCallStack &InlineStack, bool Update, uint32_t Offset) { |
| 1703 | uint64_t HeadCount = 0; |
| 1704 | if (InlineStack.size() == 0) |
| 1705 | if (!GcovBuffer.readInt64(Val&: HeadCount)) |
| 1706 | return sampleprof_error::truncated; |
| 1707 | |
| 1708 | uint32_t NameIdx; |
| 1709 | if (!GcovBuffer.readInt(Val&: NameIdx)) |
| 1710 | return sampleprof_error::truncated; |
| 1711 | |
| 1712 | StringRef Name(Names[NameIdx]); |
| 1713 | |
| 1714 | uint32_t NumPosCounts; |
| 1715 | if (!GcovBuffer.readInt(Val&: NumPosCounts)) |
| 1716 | return sampleprof_error::truncated; |
| 1717 | |
| 1718 | uint32_t NumCallsites; |
| 1719 | if (!GcovBuffer.readInt(Val&: NumCallsites)) |
| 1720 | return sampleprof_error::truncated; |
| 1721 | |
| 1722 | FunctionSamples *FProfile = nullptr; |
| 1723 | if (InlineStack.size() == 0) { |
| 1724 | // If this is a top function that we have already processed, do not |
| 1725 | // update its profile again. This happens in the presence of |
| 1726 | // function aliases. Since these aliases share the same function |
| 1727 | // body, there will be identical replicated profiles for the |
| 1728 | // original function. In this case, we simply not bother updating |
| 1729 | // the profile of the original function. |
| 1730 | FProfile = &Profiles[FunctionId(Name)]; |
| 1731 | FProfile->addHeadSamples(Num: HeadCount); |
| 1732 | if (FProfile->getTotalSamples() > 0) |
| 1733 | Update = false; |
| 1734 | } else { |
| 1735 | // Otherwise, we are reading an inlined instance. The top of the |
| 1736 | // inline stack contains the profile of the caller. Insert this |
| 1737 | // callee in the caller's CallsiteMap. |
| 1738 | FunctionSamples *CallerProfile = InlineStack.front(); |
| 1739 | uint32_t LineOffset = Offset >> 16; |
| 1740 | uint32_t Discriminator = Offset & 0xffff; |
| 1741 | FProfile = &CallerProfile->functionSamplesAt( |
| 1742 | Loc: LineLocation(LineOffset, Discriminator))[FunctionId(Name)]; |
| 1743 | } |
| 1744 | FProfile->setFunction(FunctionId(Name)); |
| 1745 | |
| 1746 | for (uint32_t I = 0; I < NumPosCounts; ++I) { |
| 1747 | uint32_t Offset; |
| 1748 | if (!GcovBuffer.readInt(Val&: Offset)) |
| 1749 | return sampleprof_error::truncated; |
| 1750 | |
| 1751 | uint32_t NumTargets; |
| 1752 | if (!GcovBuffer.readInt(Val&: NumTargets)) |
| 1753 | return sampleprof_error::truncated; |
| 1754 | |
| 1755 | uint64_t Count; |
| 1756 | if (!GcovBuffer.readInt64(Val&: Count)) |
| 1757 | return sampleprof_error::truncated; |
| 1758 | |
| 1759 | // The line location is encoded in the offset as: |
| 1760 | // high 16 bits: line offset to the start of the function. |
| 1761 | // low 16 bits: discriminator. |
| 1762 | uint32_t LineOffset = Offset >> 16; |
| 1763 | uint32_t Discriminator = Offset & 0xffff; |
| 1764 | |
| 1765 | InlineCallStack NewStack; |
| 1766 | NewStack.push_back(Elt: FProfile); |
| 1767 | llvm::append_range(C&: NewStack, R: InlineStack); |
| 1768 | if (Update) { |
| 1769 | // Walk up the inline stack, adding the samples on this line to |
| 1770 | // the total sample count of the callers in the chain. |
| 1771 | for (auto *CallerProfile : NewStack) |
| 1772 | CallerProfile->addTotalSamples(Num: Count); |
| 1773 | |
| 1774 | // Update the body samples for the current profile. |
| 1775 | FProfile->addBodySamples(LineOffset, Discriminator, Num: Count); |
| 1776 | } |
| 1777 | |
| 1778 | // Process the list of functions called at an indirect call site. |
| 1779 | // These are all the targets that a function pointer (or virtual |
| 1780 | // function) resolved at runtime. |
| 1781 | for (uint32_t J = 0; J < NumTargets; J++) { |
| 1782 | uint32_t HistVal; |
| 1783 | if (!GcovBuffer.readInt(Val&: HistVal)) |
| 1784 | return sampleprof_error::truncated; |
| 1785 | |
| 1786 | if (HistVal != HIST_TYPE_INDIR_CALL_TOPN) |
| 1787 | return sampleprof_error::malformed; |
| 1788 | |
| 1789 | uint64_t TargetIdx; |
| 1790 | if (!GcovBuffer.readInt64(Val&: TargetIdx)) |
| 1791 | return sampleprof_error::truncated; |
| 1792 | StringRef TargetName(Names[TargetIdx]); |
| 1793 | |
| 1794 | uint64_t TargetCount; |
| 1795 | if (!GcovBuffer.readInt64(Val&: TargetCount)) |
| 1796 | return sampleprof_error::truncated; |
| 1797 | |
| 1798 | if (Update) |
| 1799 | FProfile->addCalledTargetSamples(LineOffset, Discriminator, |
| 1800 | Func: FunctionId(TargetName), |
| 1801 | Num: TargetCount); |
| 1802 | } |
| 1803 | } |
| 1804 | |
| 1805 | // Process all the inlined callers into the current function. These |
| 1806 | // are all the callsites that were inlined into this function. |
| 1807 | for (uint32_t I = 0; I < NumCallsites; I++) { |
| 1808 | // The offset is encoded as: |
| 1809 | // high 16 bits: line offset to the start of the function. |
| 1810 | // low 16 bits: discriminator. |
| 1811 | uint32_t Offset; |
| 1812 | if (!GcovBuffer.readInt(Val&: Offset)) |
| 1813 | return sampleprof_error::truncated; |
| 1814 | InlineCallStack NewStack; |
| 1815 | NewStack.push_back(Elt: FProfile); |
| 1816 | llvm::append_range(C&: NewStack, R: InlineStack); |
| 1817 | if (std::error_code EC = readOneFunctionProfile(InlineStack: NewStack, Update, Offset)) |
| 1818 | return EC; |
| 1819 | } |
| 1820 | |
| 1821 | return sampleprof_error::success; |
| 1822 | } |
| 1823 | |
| 1824 | /// Read a GCC AutoFDO profile. |
| 1825 | /// |
| 1826 | /// This format is generated by the Linux Perf conversion tool at |
| 1827 | /// https://github.com/google/autofdo. |
| 1828 | std::error_code SampleProfileReaderGCC::readImpl() { |
| 1829 | assert(!ProfileIsFSDisciminator && "Gcc profiles not support FSDisciminator" ); |
| 1830 | // Read the string table. |
| 1831 | if (std::error_code EC = readNameTable()) |
| 1832 | return EC; |
| 1833 | |
| 1834 | // Read the source profile. |
| 1835 | if (std::error_code EC = readFunctionProfiles()) |
| 1836 | return EC; |
| 1837 | |
| 1838 | return sampleprof_error::success; |
| 1839 | } |
| 1840 | |
| 1841 | bool SampleProfileReaderGCC::hasFormat(const MemoryBuffer &Buffer) { |
| 1842 | StringRef Magic(reinterpret_cast<const char *>(Buffer.getBufferStart())); |
| 1843 | return Magic == "adcg*704" ; |
| 1844 | } |
| 1845 | |
| 1846 | void SampleProfileReaderItaniumRemapper::applyRemapping(LLVMContext &Ctx) { |
| 1847 | // If the reader uses MD5 to represent string, we can't remap it because |
| 1848 | // we don't know what the original function names were. |
| 1849 | if (Reader.useMD5()) { |
| 1850 | Ctx.diagnose(DI: DiagnosticInfoSampleProfile( |
| 1851 | Reader.getBuffer()->getBufferIdentifier(), |
| 1852 | "Profile data remapping cannot be applied to profile data " |
| 1853 | "using MD5 names (original mangled names are not available)." , |
| 1854 | DS_Warning)); |
| 1855 | return; |
| 1856 | } |
| 1857 | |
| 1858 | // CSSPGO-TODO: Remapper is not yet supported. |
| 1859 | // We will need to remap the entire context string. |
| 1860 | assert(Remappings && "should be initialized while creating remapper" ); |
| 1861 | for (auto &Sample : Reader.getProfiles()) { |
| 1862 | DenseSet<FunctionId> NamesInSample; |
| 1863 | Sample.second.findAllNames(NameSet&: NamesInSample); |
| 1864 | for (auto &Name : NamesInSample) { |
| 1865 | StringRef NameStr = Name.stringRef(); |
| 1866 | if (auto Key = Remappings->insert(FunctionName: NameStr)) |
| 1867 | NameMap.insert(KV: {Key, NameStr}); |
| 1868 | } |
| 1869 | } |
| 1870 | |
| 1871 | RemappingApplied = true; |
| 1872 | } |
| 1873 | |
| 1874 | std::optional<StringRef> |
| 1875 | SampleProfileReaderItaniumRemapper::lookUpNameInProfile(StringRef Fname) { |
| 1876 | if (auto Key = Remappings->lookup(FunctionName: Fname)) { |
| 1877 | StringRef Result = NameMap.lookup(Val: Key); |
| 1878 | if (!Result.empty()) |
| 1879 | return Result; |
| 1880 | } |
| 1881 | return std::nullopt; |
| 1882 | } |
| 1883 | |
| 1884 | /// Prepare a memory buffer for the contents of \p Filename. |
| 1885 | /// |
| 1886 | /// \returns an error code indicating the status of the buffer. |
| 1887 | static ErrorOr<std::unique_ptr<MemoryBuffer>> |
| 1888 | setupMemoryBuffer(const Twine &Filename, vfs::FileSystem &FS) { |
| 1889 | auto BufferOrErr = Filename.str() == "-" ? MemoryBuffer::getSTDIN() |
| 1890 | : FS.getBufferForFile(Name: Filename); |
| 1891 | if (std::error_code EC = BufferOrErr.getError()) |
| 1892 | return EC; |
| 1893 | auto Buffer = std::move(BufferOrErr.get()); |
| 1894 | |
| 1895 | return std::move(Buffer); |
| 1896 | } |
| 1897 | |
| 1898 | /// Create a sample profile reader based on the format of the input file. |
| 1899 | /// |
| 1900 | /// \param Filename The file to open. |
| 1901 | /// |
| 1902 | /// \param C The LLVM context to use to emit diagnostics. |
| 1903 | /// |
| 1904 | /// \param P The FSDiscriminatorPass. |
| 1905 | /// |
| 1906 | /// \param RemapFilename The file used for profile remapping. |
| 1907 | /// |
| 1908 | /// \returns an error code indicating the status of the created reader. |
| 1909 | ErrorOr<std::unique_ptr<SampleProfileReader>> |
| 1910 | SampleProfileReader::create(StringRef Filename, LLVMContext &C, |
| 1911 | vfs::FileSystem &FS, FSDiscriminatorPass P, |
| 1912 | StringRef RemapFilename) { |
| 1913 | auto BufferOrError = setupMemoryBuffer(Filename, FS); |
| 1914 | if (std::error_code EC = BufferOrError.getError()) |
| 1915 | return EC; |
| 1916 | return create(B&: BufferOrError.get(), C, FS, P, RemapFilename); |
| 1917 | } |
| 1918 | |
| 1919 | /// Create a sample profile remapper from the given input, to remap the |
| 1920 | /// function names in the given profile data. |
| 1921 | /// |
| 1922 | /// \param Filename The file to open. |
| 1923 | /// |
| 1924 | /// \param Reader The profile reader the remapper is going to be applied to. |
| 1925 | /// |
| 1926 | /// \param C The LLVM context to use to emit diagnostics. |
| 1927 | /// |
| 1928 | /// \returns an error code indicating the status of the created reader. |
| 1929 | ErrorOr<std::unique_ptr<SampleProfileReaderItaniumRemapper>> |
| 1930 | SampleProfileReaderItaniumRemapper::create(StringRef Filename, |
| 1931 | vfs::FileSystem &FS, |
| 1932 | SampleProfileReader &Reader, |
| 1933 | LLVMContext &C) { |
| 1934 | auto BufferOrError = setupMemoryBuffer(Filename, FS); |
| 1935 | if (std::error_code EC = BufferOrError.getError()) |
| 1936 | return EC; |
| 1937 | return create(B&: BufferOrError.get(), Reader, C); |
| 1938 | } |
| 1939 | |
| 1940 | /// Create a sample profile remapper from the given input, to remap the |
| 1941 | /// function names in the given profile data. |
| 1942 | /// |
| 1943 | /// \param B The memory buffer to create the reader from (assumes ownership). |
| 1944 | /// |
| 1945 | /// \param C The LLVM context to use to emit diagnostics. |
| 1946 | /// |
| 1947 | /// \param Reader The profile reader the remapper is going to be applied to. |
| 1948 | /// |
| 1949 | /// \returns an error code indicating the status of the created reader. |
| 1950 | ErrorOr<std::unique_ptr<SampleProfileReaderItaniumRemapper>> |
| 1951 | SampleProfileReaderItaniumRemapper::create(std::unique_ptr<MemoryBuffer> &B, |
| 1952 | SampleProfileReader &Reader, |
| 1953 | LLVMContext &C) { |
| 1954 | auto Remappings = std::make_unique<SymbolRemappingReader>(); |
| 1955 | if (Error E = Remappings->read(B&: *B)) { |
| 1956 | handleAllErrors( |
| 1957 | E: std::move(E), Handlers: [&](const SymbolRemappingParseError &ParseError) { |
| 1958 | C.diagnose(DI: DiagnosticInfoSampleProfile(B->getBufferIdentifier(), |
| 1959 | ParseError.getLineNum(), |
| 1960 | ParseError.getMessage())); |
| 1961 | }); |
| 1962 | return sampleprof_error::malformed; |
| 1963 | } |
| 1964 | |
| 1965 | return std::make_unique<SampleProfileReaderItaniumRemapper>( |
| 1966 | args: std::move(B), args: std::move(Remappings), args&: Reader); |
| 1967 | } |
| 1968 | |
| 1969 | /// Create a sample profile reader based on the format of the input data. |
| 1970 | /// |
| 1971 | /// \param B The memory buffer to create the reader from (assumes ownership). |
| 1972 | /// |
| 1973 | /// \param C The LLVM context to use to emit diagnostics. |
| 1974 | /// |
| 1975 | /// \param P The FSDiscriminatorPass. |
| 1976 | /// |
| 1977 | /// \param RemapFilename The file used for profile remapping. |
| 1978 | /// |
| 1979 | /// \returns an error code indicating the status of the created reader. |
| 1980 | ErrorOr<std::unique_ptr<SampleProfileReader>> |
| 1981 | SampleProfileReader::create(std::unique_ptr<MemoryBuffer> &B, LLVMContext &C, |
| 1982 | vfs::FileSystem &FS, FSDiscriminatorPass P, |
| 1983 | StringRef RemapFilename) { |
| 1984 | std::unique_ptr<SampleProfileReader> Reader; |
| 1985 | if (SampleProfileReaderRawBinary::hasFormat(Buffer: *B)) |
| 1986 | Reader.reset(p: new SampleProfileReaderRawBinary(std::move(B), C)); |
| 1987 | else if (SampleProfileReaderExtBinary::hasFormat(Buffer: *B)) |
| 1988 | Reader.reset(p: new SampleProfileReaderExtBinary(std::move(B), C)); |
| 1989 | else if (SampleProfileReaderGCC::hasFormat(Buffer: *B)) |
| 1990 | Reader.reset(p: new SampleProfileReaderGCC(std::move(B), C)); |
| 1991 | else if (SampleProfileReaderText::hasFormat(Buffer: *B)) |
| 1992 | Reader.reset(p: new SampleProfileReaderText(std::move(B), C)); |
| 1993 | else |
| 1994 | return sampleprof_error::unrecognized_format; |
| 1995 | |
| 1996 | if (!RemapFilename.empty()) { |
| 1997 | auto ReaderOrErr = SampleProfileReaderItaniumRemapper::create( |
| 1998 | Filename: RemapFilename, FS, Reader&: *Reader, C); |
| 1999 | if (std::error_code EC = ReaderOrErr.getError()) { |
| 2000 | std::string Msg = "Could not create remapper: " + EC.message(); |
| 2001 | C.diagnose(DI: DiagnosticInfoSampleProfile(RemapFilename, Msg)); |
| 2002 | return EC; |
| 2003 | } |
| 2004 | Reader->Remapper = std::move(ReaderOrErr.get()); |
| 2005 | } |
| 2006 | |
| 2007 | if (std::error_code EC = Reader->readHeader()) { |
| 2008 | return EC; |
| 2009 | } |
| 2010 | |
| 2011 | Reader->setDiscriminatorMaskedBitFrom(P); |
| 2012 | |
| 2013 | return std::move(Reader); |
| 2014 | } |
| 2015 | |
| 2016 | // For text and GCC file formats, we compute the summary after reading the |
| 2017 | // profile. Binary format has the profile summary in its header. |
| 2018 | void SampleProfileReader::computeSummary() { |
| 2019 | SampleProfileSummaryBuilder Builder(ProfileSummaryBuilder::DefaultCutoffs); |
| 2020 | Summary = Builder.computeSummaryForProfiles(Profiles); |
| 2021 | } |
| 2022 | |