| 1 | //===- CodeGenDataReader.cpp ----------------------------------------------===// |
| 2 | // |
| 3 | // Part of the LLVM Project, under the Apache License v2.0 with LLVM Exceptions. |
| 4 | // See https://llvm.org/LICENSE.txt for license information. |
| 5 | // SPDX-License-Identifier: Apache-2.0 WITH LLVM-exception |
| 6 | // |
| 7 | //===----------------------------------------------------------------------===// |
| 8 | // |
| 9 | // This file contains support for reading codegen data. |
| 10 | // |
| 11 | //===----------------------------------------------------------------------===// |
| 12 | |
| 13 | #include "llvm/CGData/CodeGenDataReader.h" |
| 14 | #include "llvm/CGData/OutlinedHashTreeRecord.h" |
| 15 | #include "llvm/Object/ObjectFile.h" |
| 16 | #include "llvm/Support/CommandLine.h" |
| 17 | #include "llvm/Support/MemoryBuffer.h" |
| 18 | |
| 19 | #define DEBUG_TYPE "cg-data-reader" |
| 20 | |
| 21 | using namespace llvm; |
| 22 | |
| 23 | static cl::opt<bool> IndexedCodeGenDataReadFunctionMapNames( |
| 24 | "indexed-codegen-data-read-function-map-names" , cl::init(Val: true), cl::Hidden, |
| 25 | cl::desc("Read function map names in indexed CodeGenData. Can be " |
| 26 | "disabled to save memory and time for final consumption of the " |
| 27 | "indexed CodeGenData in production." )); |
| 28 | |
| 29 | namespace llvm { |
| 30 | |
| 31 | cl::opt<bool> IndexedCodeGenDataLazyLoading( |
| 32 | "indexed-codegen-data-lazy-loading" , cl::init(Val: false), cl::Hidden, |
| 33 | cl::desc( |
| 34 | "Lazily load indexed CodeGenData. Enable to save memory and time " |
| 35 | "for final consumption of the indexed CodeGenData in production." )); |
| 36 | |
| 37 | static Expected<std::unique_ptr<MemoryBuffer>> |
| 38 | setupMemoryBuffer(const Twine &Filename, vfs::FileSystem &FS) { |
| 39 | auto BufferOrErr = Filename.str() == "-" ? MemoryBuffer::getSTDIN() |
| 40 | : FS.getBufferForFile(Name: Filename); |
| 41 | if (std::error_code EC = BufferOrErr.getError()) |
| 42 | return errorCodeToError(EC); |
| 43 | return std::move(BufferOrErr.get()); |
| 44 | } |
| 45 | |
| 46 | Error CodeGenDataReader::mergeFromObjectFile( |
| 47 | const object::ObjectFile *Obj, OutlinedHashTreeRecord &GlobalOutlineRecord, |
| 48 | StableFunctionMapRecord &GlobalFunctionMapRecord, |
| 49 | stable_hash *CombinedHash) { |
| 50 | Triple TT = Obj->makeTriple(); |
| 51 | auto CGOutlineName = |
| 52 | getCodeGenDataSectionName(CGSK: CG_outline, OF: TT.getObjectFormat(), AddSegmentInfo: false); |
| 53 | auto CGMergeName = |
| 54 | getCodeGenDataSectionName(CGSK: CG_merge, OF: TT.getObjectFormat(), AddSegmentInfo: false); |
| 55 | |
| 56 | auto processSectionContents = [&](const StringRef &Name, |
| 57 | const StringRef &Contents) { |
| 58 | if (Name != CGOutlineName && Name != CGMergeName) |
| 59 | return; |
| 60 | if (CombinedHash) |
| 61 | *CombinedHash = stable_hash_combine(A: *CombinedHash, B: xxh3_64bits(data: Contents)); |
| 62 | auto *Data = reinterpret_cast<const unsigned char *>(Contents.data()); |
| 63 | auto *EndData = Data + Contents.size(); |
| 64 | // In case dealing with an executable that has concatenated cgdata, |
| 65 | // we want to merge them into a single cgdata. |
| 66 | // Although it's not a typical workflow, we support this scenario |
| 67 | // by looping over all data in the sections. |
| 68 | if (Name == CGOutlineName) { |
| 69 | while (Data != EndData) { |
| 70 | OutlinedHashTreeRecord LocalOutlineRecord; |
| 71 | LocalOutlineRecord.deserialize(Ptr&: Data); |
| 72 | GlobalOutlineRecord.merge(Other: LocalOutlineRecord); |
| 73 | } |
| 74 | } else if (Name == CGMergeName) { |
| 75 | while (Data != EndData) { |
| 76 | StableFunctionMapRecord LocalFunctionMapRecord; |
| 77 | LocalFunctionMapRecord.deserialize(Ptr&: Data); |
| 78 | GlobalFunctionMapRecord.merge(Other: LocalFunctionMapRecord); |
| 79 | } |
| 80 | } |
| 81 | }; |
| 82 | |
| 83 | for (auto &Section : Obj->sections()) { |
| 84 | Expected<StringRef> NameOrErr = Section.getName(); |
| 85 | if (!NameOrErr) |
| 86 | return NameOrErr.takeError(); |
| 87 | Expected<StringRef> ContentsOrErr = Section.getContents(); |
| 88 | if (!ContentsOrErr) |
| 89 | return ContentsOrErr.takeError(); |
| 90 | processSectionContents(*NameOrErr, *ContentsOrErr); |
| 91 | } |
| 92 | |
| 93 | return Error::success(); |
| 94 | } |
| 95 | |
| 96 | Error IndexedCodeGenDataReader::read() { |
| 97 | using namespace support; |
| 98 | |
| 99 | // The smallest header with the version 1 is 24 bytes. |
| 100 | // Do not update this value even with the new version of the header. |
| 101 | const unsigned = 24; |
| 102 | if (DataBuffer->getBufferSize() < MinHeaderSize) |
| 103 | return error(Err: cgdata_error::bad_header); |
| 104 | |
| 105 | auto *Start = |
| 106 | reinterpret_cast<const unsigned char *>(DataBuffer->getBufferStart()); |
| 107 | auto *End = |
| 108 | reinterpret_cast<const unsigned char *>(DataBuffer->getBufferEnd()); |
| 109 | if (auto E = IndexedCGData::Header::readFromBuffer(Curr: Start).moveInto(Value&: Header)) |
| 110 | return E; |
| 111 | |
| 112 | if (hasOutlinedHashTree()) { |
| 113 | const unsigned char *Ptr = Start + Header.OutlinedHashTreeOffset; |
| 114 | if (Ptr >= End) |
| 115 | return error(Err: cgdata_error::eof); |
| 116 | HashTreeRecord.deserialize(Ptr); |
| 117 | } |
| 118 | |
| 119 | // TODO: lazy loading support for outlined hash tree. |
| 120 | std::shared_ptr<MemoryBuffer> SharedDataBuffer = std::move(DataBuffer); |
| 121 | if (hasStableFunctionMap()) { |
| 122 | const unsigned char *Ptr = Start + Header.StableFunctionMapOffset; |
| 123 | if (Ptr >= End) |
| 124 | return error(Err: cgdata_error::eof); |
| 125 | FunctionMapRecord.setReadStableFunctionMapNames( |
| 126 | IndexedCodeGenDataReadFunctionMapNames); |
| 127 | if (IndexedCodeGenDataLazyLoading) |
| 128 | FunctionMapRecord.lazyDeserialize(Buffer: std::move(SharedDataBuffer), |
| 129 | Offset: Header.StableFunctionMapOffset); |
| 130 | else |
| 131 | FunctionMapRecord.deserialize(Ptr); |
| 132 | } |
| 133 | |
| 134 | return success(); |
| 135 | } |
| 136 | |
| 137 | Expected<std::unique_ptr<CodeGenDataReader>> |
| 138 | CodeGenDataReader::create(const Twine &Path, vfs::FileSystem &FS) { |
| 139 | // Set up the buffer to read. |
| 140 | auto BufferOrError = setupMemoryBuffer(Filename: Path, FS); |
| 141 | if (Error E = BufferOrError.takeError()) |
| 142 | return std::move(E); |
| 143 | return CodeGenDataReader::create(Buffer: std::move(BufferOrError.get())); |
| 144 | } |
| 145 | |
| 146 | Expected<std::unique_ptr<CodeGenDataReader>> |
| 147 | CodeGenDataReader::create(std::unique_ptr<MemoryBuffer> Buffer) { |
| 148 | if (Buffer->getBufferSize() == 0) |
| 149 | return make_error<CGDataError>(Args: cgdata_error::empty_cgdata); |
| 150 | |
| 151 | std::unique_ptr<CodeGenDataReader> Reader; |
| 152 | // Create the reader. |
| 153 | if (IndexedCodeGenDataReader::hasFormat(Buffer: *Buffer)) |
| 154 | Reader = std::make_unique<IndexedCodeGenDataReader>(args: std::move(Buffer)); |
| 155 | else if (TextCodeGenDataReader::hasFormat(Buffer: *Buffer)) |
| 156 | Reader = std::make_unique<TextCodeGenDataReader>(args: std::move(Buffer)); |
| 157 | else |
| 158 | return make_error<CGDataError>(Args: cgdata_error::malformed); |
| 159 | |
| 160 | // Initialize the reader and return the result. |
| 161 | if (Error E = Reader->read()) |
| 162 | return std::move(E); |
| 163 | |
| 164 | return std::move(Reader); |
| 165 | } |
| 166 | |
| 167 | bool IndexedCodeGenDataReader::hasFormat(const MemoryBuffer &DataBuffer) { |
| 168 | using namespace support; |
| 169 | if (DataBuffer.getBufferSize() < sizeof(IndexedCGData::Magic)) |
| 170 | return false; |
| 171 | |
| 172 | uint64_t Magic = endian::read<uint64_t, aligned>(memory: DataBuffer.getBufferStart(), |
| 173 | endian: llvm::endianness::little); |
| 174 | // Verify that it's magical. |
| 175 | return Magic == IndexedCGData::Magic; |
| 176 | } |
| 177 | |
| 178 | bool TextCodeGenDataReader::hasFormat(const MemoryBuffer &Buffer) { |
| 179 | // Verify that this really looks like plain ASCII text by checking a |
| 180 | // 'reasonable' number of characters (up to the magic size). |
| 181 | StringRef Prefix = Buffer.getBuffer().take_front(N: sizeof(uint64_t)); |
| 182 | return llvm::all_of(Range&: Prefix, P: [](char c) { return isPrint(C: c) || isSpace(C: c); }); |
| 183 | } |
| 184 | Error TextCodeGenDataReader::read() { |
| 185 | using namespace support; |
| 186 | |
| 187 | // Parse the custom header line by line. |
| 188 | for (; !Line.is_at_eof(); ++Line) { |
| 189 | // Skip empty or whitespace-only lines |
| 190 | if (Line->trim().empty()) |
| 191 | continue; |
| 192 | |
| 193 | if (!Line->starts_with(Prefix: ":" )) |
| 194 | break; |
| 195 | StringRef Str = Line->drop_front().rtrim(); |
| 196 | if (Str.equals_insensitive(RHS: "outlined_hash_tree" )) |
| 197 | DataKind |= CGDataKind::FunctionOutlinedHashTree; |
| 198 | else if (Str.equals_insensitive(RHS: "stable_function_map" )) |
| 199 | DataKind |= CGDataKind::StableFunctionMergingMap; |
| 200 | else |
| 201 | return error(Err: cgdata_error::bad_header); |
| 202 | } |
| 203 | |
| 204 | // We treat an empty header (that is a comment # only) as a valid header. |
| 205 | if (Line.is_at_eof()) { |
| 206 | if (DataKind == CGDataKind::Unknown) |
| 207 | return Error::success(); |
| 208 | return error(Err: cgdata_error::bad_header); |
| 209 | } |
| 210 | |
| 211 | // The YAML docs follow after the header. |
| 212 | const char *Pos = Line->data(); |
| 213 | size_t Size = reinterpret_cast<size_t>(DataBuffer->getBufferEnd()) - |
| 214 | reinterpret_cast<size_t>(Pos); |
| 215 | yaml::Input YOS(StringRef(Pos, Size)); |
| 216 | if (hasOutlinedHashTree()) |
| 217 | HashTreeRecord.deserializeYAML(YIS&: YOS); |
| 218 | if (hasStableFunctionMap()) |
| 219 | FunctionMapRecord.deserializeYAML(YIS&: YOS); |
| 220 | |
| 221 | return Error::success(); |
| 222 | } |
| 223 | } // end namespace llvm |
| 224 | |