| 1 | //===- BitstreamWriter.h - Low-level bitstream writer interface -*- C++ -*-===// |
| 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 header defines the BitstreamWriter class. This class can be used to |
| 10 | // write an arbitrary bitstream, regardless of its contents. |
| 11 | // |
| 12 | //===----------------------------------------------------------------------===// |
| 13 | |
| 14 | #ifndef LLVM_BITSTREAM_BITSTREAMWRITER_H |
| 15 | #define LLVM_BITSTREAM_BITSTREAMWRITER_H |
| 16 | |
| 17 | #include "llvm/ADT/ArrayRef.h" |
| 18 | #include "llvm/ADT/SmallVector.h" |
| 19 | #include "llvm/ADT/StringRef.h" |
| 20 | #include "llvm/Bitstream/BitCodes.h" |
| 21 | #include "llvm/Support/Casting.h" |
| 22 | #include "llvm/Support/Endian.h" |
| 23 | #include "llvm/Support/MathExtras.h" |
| 24 | #include "llvm/Support/raw_ostream.h" |
| 25 | #include <algorithm> |
| 26 | #include <optional> |
| 27 | #include <vector> |
| 28 | |
| 29 | namespace llvm { |
| 30 | |
| 31 | class BitstreamWriter { |
| 32 | /// Owned buffer, used to init Buffer if the provided stream doesn't happen to |
| 33 | /// be a buffer itself. |
| 34 | SmallVector<char, 0> OwnBuffer; |
| 35 | /// Internal buffer for unflushed bytes (unless there is no stream to flush |
| 36 | /// to, case in which these are "the bytes"). The writer backpatches, so it is |
| 37 | /// efficient to buffer. |
| 38 | SmallVectorImpl<char> &Buffer; |
| 39 | |
| 40 | /// FS - The file stream that Buffer flushes to. If FS is a raw_fd_stream, the |
| 41 | /// writer will incrementally flush at subblock boundaries. Otherwise flushing |
| 42 | /// will happen at the end of BitstreamWriter's lifetime. |
| 43 | raw_ostream *const FS; |
| 44 | |
| 45 | /// FlushThreshold - this is the threshold (unit B) to flush to FS, if FS is a |
| 46 | /// raw_fd_stream. |
| 47 | const uint64_t FlushThreshold; |
| 48 | |
| 49 | /// CurBit - Always between 0 and 31 inclusive, specifies the next bit to use. |
| 50 | unsigned CurBit = 0; |
| 51 | |
| 52 | /// CurValue - The current value. Only bits < CurBit are valid. |
| 53 | uint32_t CurValue = 0; |
| 54 | |
| 55 | /// CurCodeSize - This is the declared size of code values used for the |
| 56 | /// current block, in bits. |
| 57 | unsigned CurCodeSize = 2; |
| 58 | |
| 59 | /// BlockInfoCurBID - When emitting a BLOCKINFO_BLOCK, this is the currently |
| 60 | /// selected BLOCK ID. |
| 61 | unsigned BlockInfoCurBID = 0; |
| 62 | |
| 63 | /// CurAbbrevs - Abbrevs installed at in this block. |
| 64 | std::vector<std::shared_ptr<BitCodeAbbrev>> CurAbbrevs; |
| 65 | |
| 66 | // Support for retrieving a section of the output, for purposes such as |
| 67 | // checksumming. |
| 68 | std::optional<size_t> BlockFlushingStartPos; |
| 69 | |
| 70 | struct Block { |
| 71 | unsigned PrevCodeSize; |
| 72 | size_t StartSizeWord; |
| 73 | std::vector<std::shared_ptr<BitCodeAbbrev>> PrevAbbrevs; |
| 74 | Block(unsigned PCS, size_t SSW) : PrevCodeSize(PCS), StartSizeWord(SSW) {} |
| 75 | }; |
| 76 | |
| 77 | /// BlockScope - This tracks the current blocks that we have entered. |
| 78 | std::vector<Block> BlockScope; |
| 79 | |
| 80 | /// BlockInfo - This contains information emitted to BLOCKINFO_BLOCK blocks. |
| 81 | /// These describe abbreviations that all blocks of the specified ID inherit. |
| 82 | struct BlockInfo { |
| 83 | unsigned BlockID; |
| 84 | std::vector<std::shared_ptr<BitCodeAbbrev>> Abbrevs; |
| 85 | }; |
| 86 | std::vector<BlockInfo> BlockInfoRecords; |
| 87 | |
| 88 | void WriteWord(unsigned Value) { |
| 89 | Value = |
| 90 | support::endian::byte_swap<uint32_t, llvm::endianness::little>(value: Value); |
| 91 | Buffer.append(in_start: reinterpret_cast<const char *>(&Value), |
| 92 | in_end: reinterpret_cast<const char *>(&Value + 1)); |
| 93 | } |
| 94 | |
| 95 | uint64_t GetNumOfFlushedBytes() const { |
| 96 | return fdStream() ? fdStream()->tell() : 0; |
| 97 | } |
| 98 | |
| 99 | size_t GetBufferOffset() const { |
| 100 | return Buffer.size() + GetNumOfFlushedBytes(); |
| 101 | } |
| 102 | |
| 103 | size_t GetWordIndex() const { |
| 104 | size_t Offset = GetBufferOffset(); |
| 105 | assert((Offset & 3) == 0 && "Not 32-bit aligned" ); |
| 106 | return Offset / 4; |
| 107 | } |
| 108 | |
| 109 | void flushAndClear() { |
| 110 | assert(FS); |
| 111 | assert(!Buffer.empty()); |
| 112 | assert(!BlockFlushingStartPos && |
| 113 | "a call to markAndBlockFlushing should have been paired with a " |
| 114 | "call to getMarkedBufferAndResumeFlushing" ); |
| 115 | FS->write(Ptr: Buffer.data(), Size: Buffer.size()); |
| 116 | Buffer.clear(); |
| 117 | } |
| 118 | |
| 119 | /// If the related file stream is a raw_fd_stream, flush the buffer if its |
| 120 | /// size is above a threshold. If \p OnClosing is true, flushing happens |
| 121 | /// regardless of thresholds. |
| 122 | void FlushToFile(bool OnClosing = false) { |
| 123 | if (!FS || Buffer.empty()) |
| 124 | return; |
| 125 | if (OnClosing) |
| 126 | return flushAndClear(); |
| 127 | if (BlockFlushingStartPos) |
| 128 | return; |
| 129 | if (fdStream() && Buffer.size() > FlushThreshold) |
| 130 | flushAndClear(); |
| 131 | } |
| 132 | |
| 133 | raw_fd_stream *fdStream() { return dyn_cast_or_null<raw_fd_stream>(Val: FS); } |
| 134 | |
| 135 | const raw_fd_stream *fdStream() const { |
| 136 | return dyn_cast_or_null<raw_fd_stream>(Val: FS); |
| 137 | } |
| 138 | |
| 139 | SmallVectorImpl<char> &getInternalBufferFromStream(raw_ostream &OutStream) { |
| 140 | if (auto *SV = dyn_cast<raw_svector_ostream>(Val: &OutStream)) |
| 141 | return SV->buffer(); |
| 142 | return OwnBuffer; |
| 143 | } |
| 144 | |
| 145 | public: |
| 146 | /// Create a BitstreamWriter over a raw_ostream \p OutStream. |
| 147 | /// If \p OutStream is a raw_svector_ostream, the BitstreamWriter will write |
| 148 | /// directly to the latter's buffer. In all other cases, the BitstreamWriter |
| 149 | /// will use an internal buffer and flush at the end of its lifetime. |
| 150 | /// |
| 151 | /// In addition, if \p is a raw_fd_stream supporting seek, tell, and read |
| 152 | /// (besides write), the BitstreamWriter will also flush incrementally, when a |
| 153 | /// subblock is finished, and if the FlushThreshold is passed. |
| 154 | /// |
| 155 | /// NOTE: \p FlushThreshold's unit is MB. |
| 156 | BitstreamWriter(raw_ostream &OutStream, uint32_t FlushThreshold = 512) |
| 157 | : Buffer(getInternalBufferFromStream(OutStream)), |
| 158 | FS(!isa<raw_svector_ostream>(Val: OutStream) ? &OutStream : nullptr), |
| 159 | FlushThreshold(uint64_t(FlushThreshold) << 20) {} |
| 160 | |
| 161 | /// Convenience constructor for users that start with a vector - avoids |
| 162 | /// needing to wrap it in a raw_svector_ostream. |
| 163 | BitstreamWriter(SmallVectorImpl<char> &Buff) |
| 164 | : Buffer(Buff), FS(nullptr), FlushThreshold(0) {} |
| 165 | |
| 166 | ~BitstreamWriter() { |
| 167 | FlushToWord(); |
| 168 | assert(BlockScope.empty() && CurAbbrevs.empty() && "Block imbalance" ); |
| 169 | FlushToFile(/*OnClosing=*/OnClosing: true); |
| 170 | } |
| 171 | |
| 172 | /// For scenarios where the user wants to access a section of the stream to |
| 173 | /// (for example) compute some checksum, disable flushing and remember the |
| 174 | /// position in the internal buffer where that happened. Must be paired with a |
| 175 | /// call to getMarkedBufferAndResumeFlushing. |
| 176 | void markAndBlockFlushing() { |
| 177 | assert(!BlockFlushingStartPos); |
| 178 | BlockFlushingStartPos = Buffer.size(); |
| 179 | } |
| 180 | |
| 181 | /// resumes flushing, but does not flush, and returns the section in the |
| 182 | /// internal buffer starting from the position marked with |
| 183 | /// markAndBlockFlushing. The return should be processed before any additional |
| 184 | /// calls to this object, because those may cause a flush and invalidate the |
| 185 | /// return. |
| 186 | StringRef getMarkedBufferAndResumeFlushing() { |
| 187 | assert(BlockFlushingStartPos); |
| 188 | size_t Start = *BlockFlushingStartPos; |
| 189 | BlockFlushingStartPos.reset(); |
| 190 | return {&Buffer[Start], Buffer.size() - Start}; |
| 191 | } |
| 192 | |
| 193 | /// Retrieve the current position in the stream, in bits. |
| 194 | uint64_t GetCurrentBitNo() const { return GetBufferOffset() * 8 + CurBit; } |
| 195 | |
| 196 | /// Retrieve the number of bits currently used to encode an abbrev ID. |
| 197 | unsigned GetAbbrevIDWidth() const { return CurCodeSize; } |
| 198 | |
| 199 | //===--------------------------------------------------------------------===// |
| 200 | // Basic Primitives for emitting bits to the stream. |
| 201 | //===--------------------------------------------------------------------===// |
| 202 | |
| 203 | /// Backpatch a byte in the output at the given bit offset with the specified |
| 204 | /// value. |
| 205 | void BackpatchByte(uint64_t BitNo, uint8_t NewByte) { |
| 206 | using namespace llvm::support; |
| 207 | uint64_t ByteNo = BitNo / 8; |
| 208 | uint64_t StartBit = BitNo & 7; |
| 209 | uint64_t NumOfFlushedBytes = GetNumOfFlushedBytes(); |
| 210 | |
| 211 | if (ByteNo >= NumOfFlushedBytes) { |
| 212 | assert((!endian::readAtBitAlignment<uint8_t, llvm::endianness::little, |
| 213 | unaligned>( |
| 214 | &Buffer[ByteNo - NumOfFlushedBytes], StartBit)) && |
| 215 | "Expected to be patching over 0-value placeholders" ); |
| 216 | endian::writeAtBitAlignment<uint8_t, llvm::endianness::little, unaligned>( |
| 217 | memory: &Buffer[ByteNo - NumOfFlushedBytes], value: NewByte, startBit: StartBit); |
| 218 | return; |
| 219 | } |
| 220 | |
| 221 | // If we don't have a raw_fd_stream, GetNumOfFlushedBytes() should have |
| 222 | // returned 0, and we shouldn't be here. |
| 223 | assert(fdStream() != nullptr); |
| 224 | // If the byte offset to backpatch is flushed, use seek to backfill data. |
| 225 | // First, save the file position to restore later. |
| 226 | uint64_t CurPos = fdStream()->tell(); |
| 227 | |
| 228 | // Copy data to update into Bytes from the file FS and the buffer Out. |
| 229 | char Bytes[3]; // Use one more byte to silence a warning from Visual C++. |
| 230 | size_t BytesNum = StartBit ? 2 : 1; |
| 231 | size_t BytesFromDisk = std::min(a: static_cast<uint64_t>(BytesNum), b: NumOfFlushedBytes - ByteNo); |
| 232 | size_t BytesFromBuffer = BytesNum - BytesFromDisk; |
| 233 | |
| 234 | // When unaligned, copy existing data into Bytes from the file FS and the |
| 235 | // buffer Buffer so that it can be updated before writing. For debug builds |
| 236 | // read bytes unconditionally in order to check that the existing value is 0 |
| 237 | // as expected. |
| 238 | #ifdef NDEBUG |
| 239 | if (StartBit) |
| 240 | #endif |
| 241 | { |
| 242 | fdStream()->seek(off: ByteNo); |
| 243 | ssize_t BytesRead = fdStream()->read(Ptr: Bytes, Size: BytesFromDisk); |
| 244 | (void)BytesRead; // silence warning |
| 245 | assert(BytesRead >= 0 && static_cast<size_t>(BytesRead) == BytesFromDisk); |
| 246 | for (size_t i = 0; i < BytesFromBuffer; ++i) |
| 247 | Bytes[BytesFromDisk + i] = Buffer[i]; |
| 248 | assert((!endian::readAtBitAlignment<uint8_t, llvm::endianness::little, |
| 249 | unaligned>(Bytes, StartBit)) && |
| 250 | "Expected to be patching over 0-value placeholders" ); |
| 251 | } |
| 252 | |
| 253 | // Update Bytes in terms of bit offset and value. |
| 254 | endian::writeAtBitAlignment<uint8_t, llvm::endianness::little, unaligned>( |
| 255 | memory: Bytes, value: NewByte, startBit: StartBit); |
| 256 | |
| 257 | // Copy updated data back to the file FS and the buffer Out. |
| 258 | fdStream()->seek(off: ByteNo); |
| 259 | fdStream()->write(Ptr: Bytes, Size: BytesFromDisk); |
| 260 | for (size_t i = 0; i < BytesFromBuffer; ++i) |
| 261 | Buffer[i] = Bytes[BytesFromDisk + i]; |
| 262 | |
| 263 | // Restore the file position. |
| 264 | fdStream()->seek(off: CurPos); |
| 265 | } |
| 266 | |
| 267 | void BackpatchHalfWord(uint64_t BitNo, uint16_t Val) { |
| 268 | BackpatchByte(BitNo, NewByte: (uint8_t)Val); |
| 269 | BackpatchByte(BitNo: BitNo + 8, NewByte: (uint8_t)(Val >> 8)); |
| 270 | } |
| 271 | |
| 272 | void BackpatchWord(uint64_t BitNo, unsigned Val) { |
| 273 | BackpatchHalfWord(BitNo, Val: (uint16_t)Val); |
| 274 | BackpatchHalfWord(BitNo: BitNo + 16, Val: (uint16_t)(Val >> 16)); |
| 275 | } |
| 276 | |
| 277 | void BackpatchWord64(uint64_t BitNo, uint64_t Val) { |
| 278 | BackpatchWord(BitNo, Val: (uint32_t)Val); |
| 279 | BackpatchWord(BitNo: BitNo + 32, Val: (uint32_t)(Val >> 32)); |
| 280 | } |
| 281 | |
| 282 | void Emit(uint32_t Val, unsigned NumBits) { |
| 283 | assert(NumBits && NumBits <= 32 && "Invalid value size!" ); |
| 284 | assert((Val & ~(~0U >> (32-NumBits))) == 0 && "High bits set!" ); |
| 285 | CurValue |= Val << CurBit; |
| 286 | if (CurBit + NumBits < 32) { |
| 287 | CurBit += NumBits; |
| 288 | return; |
| 289 | } |
| 290 | |
| 291 | // Add the current word. |
| 292 | WriteWord(Value: CurValue); |
| 293 | |
| 294 | if (CurBit) |
| 295 | CurValue = Val >> (32-CurBit); |
| 296 | else |
| 297 | CurValue = 0; |
| 298 | CurBit = (CurBit+NumBits) & 31; |
| 299 | } |
| 300 | |
| 301 | void FlushToWord() { |
| 302 | if (CurBit) { |
| 303 | WriteWord(Value: CurValue); |
| 304 | CurBit = 0; |
| 305 | CurValue = 0; |
| 306 | } |
| 307 | } |
| 308 | |
| 309 | void EmitVBR(uint32_t Val, unsigned NumBits) { |
| 310 | assert(NumBits <= 32 && "Too many bits to emit!" ); |
| 311 | uint32_t Threshold = 1U << (NumBits-1); |
| 312 | |
| 313 | // Emit the bits with VBR encoding, NumBits-1 bits at a time. |
| 314 | while (Val >= Threshold) { |
| 315 | Emit(Val: (Val & ((1U << (NumBits - 1)) - 1)) | (1U << (NumBits - 1)), |
| 316 | NumBits); |
| 317 | Val >>= NumBits-1; |
| 318 | } |
| 319 | |
| 320 | Emit(Val, NumBits); |
| 321 | } |
| 322 | |
| 323 | void EmitVBR64(uint64_t Val, unsigned NumBits) { |
| 324 | assert(NumBits <= 32 && "Too many bits to emit!" ); |
| 325 | if ((uint32_t)Val == Val) |
| 326 | return EmitVBR(Val: (uint32_t)Val, NumBits); |
| 327 | |
| 328 | uint32_t Threshold = 1U << (NumBits-1); |
| 329 | |
| 330 | // Emit the bits with VBR encoding, NumBits-1 bits at a time. |
| 331 | while (Val >= Threshold) { |
| 332 | Emit(Val: ((uint32_t)Val & ((1U << (NumBits - 1)) - 1)) | |
| 333 | (1U << (NumBits - 1)), |
| 334 | NumBits); |
| 335 | Val >>= NumBits-1; |
| 336 | } |
| 337 | |
| 338 | Emit(Val: (uint32_t)Val, NumBits); |
| 339 | } |
| 340 | |
| 341 | /// EmitCode - Emit the specified code. |
| 342 | void EmitCode(unsigned Val) { |
| 343 | Emit(Val, NumBits: CurCodeSize); |
| 344 | } |
| 345 | |
| 346 | //===--------------------------------------------------------------------===// |
| 347 | // Block Manipulation |
| 348 | //===--------------------------------------------------------------------===// |
| 349 | |
| 350 | /// getBlockInfo - If there is block info for the specified ID, return it, |
| 351 | /// otherwise return null. |
| 352 | BlockInfo *getBlockInfo(unsigned BlockID) { |
| 353 | // Common case, the most recent entry matches BlockID. |
| 354 | if (!BlockInfoRecords.empty() && BlockInfoRecords.back().BlockID == BlockID) |
| 355 | return &BlockInfoRecords.back(); |
| 356 | |
| 357 | for (BlockInfo &BI : BlockInfoRecords) |
| 358 | if (BI.BlockID == BlockID) |
| 359 | return &BI; |
| 360 | return nullptr; |
| 361 | } |
| 362 | |
| 363 | void EnterSubblock(unsigned BlockID, unsigned CodeLen) { |
| 364 | // Block header: |
| 365 | // [ENTER_SUBBLOCK, blockid, newcodelen, <align4bytes>, blocklen] |
| 366 | EmitCode(Val: bitc::ENTER_SUBBLOCK); |
| 367 | EmitVBR(Val: BlockID, NumBits: bitc::BlockIDWidth); |
| 368 | EmitVBR(Val: CodeLen, NumBits: bitc::CodeLenWidth); |
| 369 | FlushToWord(); |
| 370 | |
| 371 | size_t BlockSizeWordIndex = GetWordIndex(); |
| 372 | unsigned OldCodeSize = CurCodeSize; |
| 373 | |
| 374 | // Emit a placeholder, which will be replaced when the block is popped. |
| 375 | Emit(Val: 0, NumBits: bitc::BlockSizeWidth); |
| 376 | |
| 377 | CurCodeSize = CodeLen; |
| 378 | |
| 379 | // Push the outer block's abbrev set onto the stack, start out with an |
| 380 | // empty abbrev set. |
| 381 | BlockScope.emplace_back(args&: OldCodeSize, args&: BlockSizeWordIndex); |
| 382 | BlockScope.back().PrevAbbrevs.swap(x&: CurAbbrevs); |
| 383 | |
| 384 | // If there is a blockinfo for this BlockID, add all the predefined abbrevs |
| 385 | // to the abbrev list. |
| 386 | if (BlockInfo *Info = getBlockInfo(BlockID)) |
| 387 | append_range(C&: CurAbbrevs, R&: Info->Abbrevs); |
| 388 | } |
| 389 | |
| 390 | void ExitBlock() { |
| 391 | assert(!BlockScope.empty() && "Block scope imbalance!" ); |
| 392 | const Block &B = BlockScope.back(); |
| 393 | |
| 394 | // Block tail: |
| 395 | // [END_BLOCK, <align4bytes>] |
| 396 | EmitCode(Val: bitc::END_BLOCK); |
| 397 | FlushToWord(); |
| 398 | |
| 399 | // Compute the size of the block, in words, not counting the size field. |
| 400 | size_t SizeInWords = GetWordIndex() - B.StartSizeWord - 1; |
| 401 | uint64_t BitNo = uint64_t(B.StartSizeWord) * 32; |
| 402 | |
| 403 | // Update the block size field in the header of this sub-block. |
| 404 | BackpatchWord(BitNo, Val: SizeInWords); |
| 405 | |
| 406 | // Restore the inner block's code size and abbrev table. |
| 407 | CurCodeSize = B.PrevCodeSize; |
| 408 | CurAbbrevs = std::move(B.PrevAbbrevs); |
| 409 | BlockScope.pop_back(); |
| 410 | FlushToFile(); |
| 411 | } |
| 412 | |
| 413 | //===--------------------------------------------------------------------===// |
| 414 | // Record Emission |
| 415 | //===--------------------------------------------------------------------===// |
| 416 | |
| 417 | private: |
| 418 | /// EmitAbbreviatedLiteral - Emit a literal value according to its abbrev |
| 419 | /// record. This is a no-op, since the abbrev specifies the literal to use. |
| 420 | template<typename uintty> |
| 421 | void EmitAbbreviatedLiteral(const BitCodeAbbrevOp &Op, uintty V) { |
| 422 | assert(Op.isLiteral() && "Not a literal" ); |
| 423 | // If the abbrev specifies the literal value to use, don't emit |
| 424 | // anything. |
| 425 | assert(V == Op.getLiteralValue() && |
| 426 | "Invalid abbrev for record!" ); |
| 427 | } |
| 428 | |
| 429 | /// EmitAbbreviatedField - Emit a single scalar field value with the specified |
| 430 | /// encoding. |
| 431 | template<typename uintty> |
| 432 | void EmitAbbreviatedField(const BitCodeAbbrevOp &Op, uintty V) { |
| 433 | assert(!Op.isLiteral() && "Literals should use EmitAbbreviatedLiteral!" ); |
| 434 | |
| 435 | // Encode the value as we are commanded. |
| 436 | switch (Op.getEncoding()) { |
| 437 | default: llvm_unreachable("Unknown encoding!" ); |
| 438 | case BitCodeAbbrevOp::Fixed: |
| 439 | if (Op.getEncodingData()) |
| 440 | Emit(Val: (unsigned)V, NumBits: (unsigned)Op.getEncodingData()); |
| 441 | break; |
| 442 | case BitCodeAbbrevOp::VBR: |
| 443 | if (Op.getEncodingData()) |
| 444 | EmitVBR64(Val: V, NumBits: (unsigned)Op.getEncodingData()); |
| 445 | break; |
| 446 | case BitCodeAbbrevOp::Char6: |
| 447 | Emit(Val: BitCodeAbbrevOp::EncodeChar6(C: (char)V), NumBits: 6); |
| 448 | break; |
| 449 | } |
| 450 | } |
| 451 | |
| 452 | /// EmitRecordWithAbbrevImpl - This is the core implementation of the record |
| 453 | /// emission code. If BlobData is non-null, then it specifies an array of |
| 454 | /// data that should be emitted as part of the Blob or Array operand that is |
| 455 | /// known to exist at the end of the record. If Code is specified, then |
| 456 | /// it is the record code to emit before the Vals, which must not contain |
| 457 | /// the code. |
| 458 | template <typename uintty> |
| 459 | void EmitRecordWithAbbrevImpl(unsigned Abbrev, ArrayRef<uintty> Vals, |
| 460 | StringRef Blob, std::optional<unsigned> Code) { |
| 461 | const char *BlobData = Blob.data(); |
| 462 | unsigned BlobLen = (unsigned) Blob.size(); |
| 463 | unsigned AbbrevNo = Abbrev-bitc::FIRST_APPLICATION_ABBREV; |
| 464 | assert(AbbrevNo < CurAbbrevs.size() && "Invalid abbrev #!" ); |
| 465 | const BitCodeAbbrev *Abbv = CurAbbrevs[AbbrevNo].get(); |
| 466 | |
| 467 | EmitCode(Val: Abbrev); |
| 468 | |
| 469 | unsigned i = 0, e = static_cast<unsigned>(Abbv->getNumOperandInfos()); |
| 470 | if (Code) { |
| 471 | assert(e && "Expected non-empty abbreviation" ); |
| 472 | const BitCodeAbbrevOp &Op = Abbv->getOperandInfo(N: i++); |
| 473 | |
| 474 | if (Op.isLiteral()) |
| 475 | EmitAbbreviatedLiteral(Op, V: *Code); |
| 476 | else { |
| 477 | assert(Op.getEncoding() != BitCodeAbbrevOp::Array && |
| 478 | Op.getEncoding() != BitCodeAbbrevOp::Blob && |
| 479 | "Expected literal or scalar" ); |
| 480 | EmitAbbreviatedField(Op, V: *Code); |
| 481 | } |
| 482 | } |
| 483 | |
| 484 | unsigned RecordIdx = 0; |
| 485 | for (; i != e; ++i) { |
| 486 | const BitCodeAbbrevOp &Op = Abbv->getOperandInfo(N: i); |
| 487 | if (Op.isLiteral()) { |
| 488 | assert(RecordIdx < Vals.size() && "Invalid abbrev/record" ); |
| 489 | EmitAbbreviatedLiteral(Op, Vals[RecordIdx]); |
| 490 | ++RecordIdx; |
| 491 | } else if (Op.getEncoding() == BitCodeAbbrevOp::Array) { |
| 492 | // Array case. |
| 493 | assert(i + 2 == e && "array op not second to last?" ); |
| 494 | const BitCodeAbbrevOp &EltEnc = Abbv->getOperandInfo(N: ++i); |
| 495 | |
| 496 | // If this record has blob data, emit it, otherwise we must have record |
| 497 | // entries to encode this way. |
| 498 | if (BlobData) { |
| 499 | assert(RecordIdx == Vals.size() && |
| 500 | "Blob data and record entries specified for array!" ); |
| 501 | // Emit a vbr6 to indicate the number of elements present. |
| 502 | EmitVBR(Val: static_cast<uint32_t>(BlobLen), NumBits: 6); |
| 503 | |
| 504 | // Emit each field. |
| 505 | for (unsigned i = 0; i != BlobLen; ++i) |
| 506 | EmitAbbreviatedField(Op: EltEnc, V: (unsigned char)BlobData[i]); |
| 507 | |
| 508 | // Know that blob data is consumed for assertion below. |
| 509 | BlobData = nullptr; |
| 510 | } else { |
| 511 | // Emit a vbr6 to indicate the number of elements present. |
| 512 | EmitVBR(Val: static_cast<uint32_t>(Vals.size()-RecordIdx), NumBits: 6); |
| 513 | |
| 514 | // Emit each field. |
| 515 | for (unsigned e = Vals.size(); RecordIdx != e; ++RecordIdx) |
| 516 | EmitAbbreviatedField(EltEnc, Vals[RecordIdx]); |
| 517 | } |
| 518 | } else if (Op.getEncoding() == BitCodeAbbrevOp::Blob) { |
| 519 | // If this record has blob data, emit it, otherwise we must have record |
| 520 | // entries to encode this way. |
| 521 | |
| 522 | if (BlobData) { |
| 523 | assert(RecordIdx == Vals.size() && |
| 524 | "Blob data and record entries specified for blob operand!" ); |
| 525 | |
| 526 | assert(Blob.data() == BlobData && "BlobData got moved" ); |
| 527 | assert(Blob.size() == BlobLen && "BlobLen got changed" ); |
| 528 | emitBlob(Bytes: Blob); |
| 529 | BlobData = nullptr; |
| 530 | } else { |
| 531 | emitBlob(Vals.slice(RecordIdx)); |
| 532 | } |
| 533 | } else { // Single scalar field. |
| 534 | assert(RecordIdx < Vals.size() && "Invalid abbrev/record" ); |
| 535 | EmitAbbreviatedField(Op, Vals[RecordIdx]); |
| 536 | ++RecordIdx; |
| 537 | } |
| 538 | } |
| 539 | assert(RecordIdx == Vals.size() && "Not all record operands emitted!" ); |
| 540 | assert(BlobData == nullptr && |
| 541 | "Blob data specified for record that doesn't use it!" ); |
| 542 | } |
| 543 | |
| 544 | public: |
| 545 | /// Emit a blob, including flushing before and tail-padding. |
| 546 | template <class UIntTy> |
| 547 | void emitBlob(ArrayRef<UIntTy> Bytes, bool ShouldEmitSize = true) { |
| 548 | // Emit a vbr6 to indicate the number of elements present. |
| 549 | if (ShouldEmitSize) |
| 550 | EmitVBR(Val: static_cast<uint32_t>(Bytes.size()), NumBits: 6); |
| 551 | |
| 552 | // Flush to a 32-bit alignment boundary. |
| 553 | FlushToWord(); |
| 554 | |
| 555 | // Emit literal bytes. |
| 556 | assert(llvm::all_of(Bytes, [](UIntTy B) { return isUInt<8>(B); })); |
| 557 | Buffer.append(Bytes.begin(), Bytes.end()); |
| 558 | |
| 559 | // Align end to 32-bits. |
| 560 | while (GetBufferOffset() & 3) |
| 561 | Buffer.push_back(Elt: 0); |
| 562 | } |
| 563 | void emitBlob(StringRef Bytes, bool ShouldEmitSize = true) { |
| 564 | emitBlob(Bytes: ArrayRef((const uint8_t *)Bytes.data(), Bytes.size()), |
| 565 | ShouldEmitSize); |
| 566 | } |
| 567 | |
| 568 | /// EmitRecord - Emit the specified record to the stream, using an abbrev if |
| 569 | /// we have one to compress the output. |
| 570 | template <typename Container> |
| 571 | void EmitRecord(unsigned Code, const Container &Vals, unsigned Abbrev = 0) { |
| 572 | if (!Abbrev) { |
| 573 | // If we don't have an abbrev to use, emit this in its fully unabbreviated |
| 574 | // form. |
| 575 | auto Count = static_cast<uint32_t>(std::size(Vals)); |
| 576 | EmitCode(Val: bitc::UNABBREV_RECORD); |
| 577 | EmitVBR(Val: Code, NumBits: 6); |
| 578 | EmitVBR(Val: Count, NumBits: 6); |
| 579 | for (unsigned i = 0, e = Count; i != e; ++i) |
| 580 | EmitVBR64(Val: Vals[i], NumBits: 6); |
| 581 | return; |
| 582 | } |
| 583 | |
| 584 | EmitRecordWithAbbrevImpl(Abbrev, ArrayRef(Vals), StringRef(), Code); |
| 585 | } |
| 586 | |
| 587 | /// EmitRecordWithAbbrev - Emit a record with the specified abbreviation. |
| 588 | /// Unlike EmitRecord, the code for the record should be included in Vals as |
| 589 | /// the first entry. |
| 590 | template <typename Container> |
| 591 | void EmitRecordWithAbbrev(unsigned Abbrev, const Container &Vals) { |
| 592 | EmitRecordWithAbbrevImpl(Abbrev, ArrayRef(Vals), StringRef(), std::nullopt); |
| 593 | } |
| 594 | |
| 595 | /// EmitRecordWithBlob - Emit the specified record to the stream, using an |
| 596 | /// abbrev that includes a blob at the end. The blob data to emit is |
| 597 | /// specified by the pointer and length specified at the end. In contrast to |
| 598 | /// EmitRecord, this routine expects that the first entry in Vals is the code |
| 599 | /// of the record. |
| 600 | template <typename Container> |
| 601 | void EmitRecordWithBlob(unsigned Abbrev, const Container &Vals, |
| 602 | StringRef Blob) { |
| 603 | EmitRecordWithAbbrevImpl(Abbrev, ArrayRef(Vals), Blob, std::nullopt); |
| 604 | } |
| 605 | template <typename Container> |
| 606 | void EmitRecordWithBlob(unsigned Abbrev, const Container &Vals, |
| 607 | const char *BlobData, unsigned BlobLen) { |
| 608 | return EmitRecordWithAbbrevImpl(Abbrev, ArrayRef(Vals), |
| 609 | StringRef(BlobData, BlobLen), std::nullopt); |
| 610 | } |
| 611 | |
| 612 | /// EmitRecordWithArray - Just like EmitRecordWithBlob, works with records |
| 613 | /// that end with an array. |
| 614 | template <typename Container> |
| 615 | void EmitRecordWithArray(unsigned Abbrev, const Container &Vals, |
| 616 | StringRef Array) { |
| 617 | EmitRecordWithAbbrevImpl(Abbrev, ArrayRef(Vals), Array, std::nullopt); |
| 618 | } |
| 619 | template <typename Container> |
| 620 | void EmitRecordWithArray(unsigned Abbrev, const Container &Vals, |
| 621 | const char *ArrayData, unsigned ArrayLen) { |
| 622 | return EmitRecordWithAbbrevImpl( |
| 623 | Abbrev, ArrayRef(Vals), StringRef(ArrayData, ArrayLen), std::nullopt); |
| 624 | } |
| 625 | |
| 626 | //===--------------------------------------------------------------------===// |
| 627 | // Abbrev Emission |
| 628 | //===--------------------------------------------------------------------===// |
| 629 | |
| 630 | private: |
| 631 | // Emit the abbreviation as a DEFINE_ABBREV record. |
| 632 | void EncodeAbbrev(const BitCodeAbbrev &Abbv) { |
| 633 | EmitCode(Val: bitc::DEFINE_ABBREV); |
| 634 | EmitVBR(Val: Abbv.getNumOperandInfos(), NumBits: 5); |
| 635 | for (unsigned i = 0, e = static_cast<unsigned>(Abbv.getNumOperandInfos()); |
| 636 | i != e; ++i) { |
| 637 | const BitCodeAbbrevOp &Op = Abbv.getOperandInfo(N: i); |
| 638 | Emit(Val: Op.isLiteral(), NumBits: 1); |
| 639 | if (Op.isLiteral()) { |
| 640 | EmitVBR64(Val: Op.getLiteralValue(), NumBits: 8); |
| 641 | } else { |
| 642 | Emit(Val: Op.getEncoding(), NumBits: 3); |
| 643 | if (Op.hasEncodingData()) |
| 644 | EmitVBR64(Val: Op.getEncodingData(), NumBits: 5); |
| 645 | } |
| 646 | } |
| 647 | } |
| 648 | public: |
| 649 | |
| 650 | /// Emits the abbreviation \p Abbv to the stream. |
| 651 | unsigned EmitAbbrev(std::shared_ptr<BitCodeAbbrev> Abbv) { |
| 652 | EncodeAbbrev(Abbv: *Abbv); |
| 653 | CurAbbrevs.push_back(x: std::move(Abbv)); |
| 654 | return static_cast<unsigned>(CurAbbrevs.size())-1 + |
| 655 | bitc::FIRST_APPLICATION_ABBREV; |
| 656 | } |
| 657 | |
| 658 | //===--------------------------------------------------------------------===// |
| 659 | // BlockInfo Block Emission |
| 660 | //===--------------------------------------------------------------------===// |
| 661 | |
| 662 | /// EnterBlockInfoBlock - Start emitting the BLOCKINFO_BLOCK. |
| 663 | void EnterBlockInfoBlock() { |
| 664 | EnterSubblock(BlockID: bitc::BLOCKINFO_BLOCK_ID, CodeLen: 2); |
| 665 | BlockInfoCurBID = ~0U; |
| 666 | BlockInfoRecords.clear(); |
| 667 | } |
| 668 | private: |
| 669 | /// SwitchToBlockID - If we aren't already talking about the specified block |
| 670 | /// ID, emit a BLOCKINFO_CODE_SETBID record. |
| 671 | void SwitchToBlockID(unsigned BlockID) { |
| 672 | if (BlockInfoCurBID == BlockID) return; |
| 673 | SmallVector<unsigned, 2> V; |
| 674 | V.push_back(Elt: BlockID); |
| 675 | EmitRecord(Code: bitc::BLOCKINFO_CODE_SETBID, Vals: V); |
| 676 | BlockInfoCurBID = BlockID; |
| 677 | } |
| 678 | |
| 679 | BlockInfo &getOrCreateBlockInfo(unsigned BlockID) { |
| 680 | if (BlockInfo *BI = getBlockInfo(BlockID)) |
| 681 | return *BI; |
| 682 | |
| 683 | // Otherwise, add a new record. |
| 684 | BlockInfoRecords.emplace_back(); |
| 685 | BlockInfoRecords.back().BlockID = BlockID; |
| 686 | return BlockInfoRecords.back(); |
| 687 | } |
| 688 | |
| 689 | public: |
| 690 | |
| 691 | /// EmitBlockInfoAbbrev - Emit a DEFINE_ABBREV record for the specified |
| 692 | /// BlockID. |
| 693 | unsigned EmitBlockInfoAbbrev(unsigned BlockID, std::shared_ptr<BitCodeAbbrev> Abbv) { |
| 694 | SwitchToBlockID(BlockID); |
| 695 | EncodeAbbrev(Abbv: *Abbv); |
| 696 | |
| 697 | // Add the abbrev to the specified block record. |
| 698 | BlockInfo &Info = getOrCreateBlockInfo(BlockID); |
| 699 | Info.Abbrevs.push_back(x: std::move(Abbv)); |
| 700 | |
| 701 | return Info.Abbrevs.size()-1+bitc::FIRST_APPLICATION_ABBREV; |
| 702 | } |
| 703 | }; |
| 704 | |
| 705 | |
| 706 | } // End llvm namespace |
| 707 | |
| 708 | #endif |
| 709 | |