Andrew Scull | 5e1ddfa | 2018-08-14 10:06:54 +0100 | [diff] [blame] | 1 | //===- BitstreamWriter.h - Low-level bitstream writer interface -*- C++ -*-===// |
| 2 | // |
Andrew Walbran | 16937d0 | 2019-10-22 13:54:20 +0100 | [diff] [blame] | 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 |
Andrew Scull | 5e1ddfa | 2018-08-14 10:06:54 +0100 | [diff] [blame] | 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 | |
Andrew Walbran | 3d2c197 | 2020-04-07 12:24:26 +0100 | [diff] [blame] | 14 | #ifndef LLVM_BITSTREAM_BITSTREAMWRITER_H |
| 15 | #define LLVM_BITSTREAM_BITSTREAMWRITER_H |
Andrew Scull | 5e1ddfa | 2018-08-14 10:06:54 +0100 | [diff] [blame] | 16 | |
| 17 | #include "llvm/ADT/ArrayRef.h" |
| 18 | #include "llvm/ADT/Optional.h" |
| 19 | #include "llvm/ADT/SmallVector.h" |
| 20 | #include "llvm/ADT/StringRef.h" |
Andrew Walbran | 3d2c197 | 2020-04-07 12:24:26 +0100 | [diff] [blame] | 21 | #include "llvm/Bitstream/BitCodes.h" |
Andrew Scull | 5e1ddfa | 2018-08-14 10:06:54 +0100 | [diff] [blame] | 22 | #include "llvm/Support/Endian.h" |
Olivier Deprez | f4ef2d0 | 2021-04-20 13:36:24 +0200 | [diff] [blame^] | 23 | #include "llvm/Support/raw_ostream.h" |
| 24 | #include <algorithm> |
Andrew Scull | 5e1ddfa | 2018-08-14 10:06:54 +0100 | [diff] [blame] | 25 | #include <vector> |
| 26 | |
| 27 | namespace llvm { |
| 28 | |
| 29 | class BitstreamWriter { |
Olivier Deprez | f4ef2d0 | 2021-04-20 13:36:24 +0200 | [diff] [blame^] | 30 | /// Out - The buffer that keeps unflushed bytes. |
Andrew Scull | 5e1ddfa | 2018-08-14 10:06:54 +0100 | [diff] [blame] | 31 | SmallVectorImpl<char> &Out; |
| 32 | |
Olivier Deprez | f4ef2d0 | 2021-04-20 13:36:24 +0200 | [diff] [blame^] | 33 | /// FS - The file stream that Out flushes to. If FS is nullptr, it does not |
| 34 | /// support read or seek, Out cannot be flushed until all data are written. |
| 35 | raw_fd_stream *FS; |
| 36 | |
| 37 | /// FlushThreshold - If FS is valid, this is the threshold (unit B) to flush |
| 38 | /// FS. |
| 39 | const uint64_t FlushThreshold; |
| 40 | |
Andrew Scull | 5e1ddfa | 2018-08-14 10:06:54 +0100 | [diff] [blame] | 41 | /// CurBit - Always between 0 and 31 inclusive, specifies the next bit to use. |
| 42 | unsigned CurBit; |
| 43 | |
Olivier Deprez | f4ef2d0 | 2021-04-20 13:36:24 +0200 | [diff] [blame^] | 44 | /// CurValue - The current value. Only bits < CurBit are valid. |
Andrew Scull | 5e1ddfa | 2018-08-14 10:06:54 +0100 | [diff] [blame] | 45 | uint32_t CurValue; |
| 46 | |
| 47 | /// CurCodeSize - This is the declared size of code values used for the |
| 48 | /// current block, in bits. |
| 49 | unsigned CurCodeSize; |
| 50 | |
| 51 | /// BlockInfoCurBID - When emitting a BLOCKINFO_BLOCK, this is the currently |
| 52 | /// selected BLOCK ID. |
| 53 | unsigned BlockInfoCurBID; |
| 54 | |
| 55 | /// CurAbbrevs - Abbrevs installed at in this block. |
| 56 | std::vector<std::shared_ptr<BitCodeAbbrev>> CurAbbrevs; |
| 57 | |
| 58 | struct Block { |
| 59 | unsigned PrevCodeSize; |
| 60 | size_t StartSizeWord; |
| 61 | std::vector<std::shared_ptr<BitCodeAbbrev>> PrevAbbrevs; |
| 62 | Block(unsigned PCS, size_t SSW) : PrevCodeSize(PCS), StartSizeWord(SSW) {} |
| 63 | }; |
| 64 | |
| 65 | /// BlockScope - This tracks the current blocks that we have entered. |
| 66 | std::vector<Block> BlockScope; |
| 67 | |
| 68 | /// BlockInfo - This contains information emitted to BLOCKINFO_BLOCK blocks. |
| 69 | /// These describe abbreviations that all blocks of the specified ID inherit. |
| 70 | struct BlockInfo { |
| 71 | unsigned BlockID; |
| 72 | std::vector<std::shared_ptr<BitCodeAbbrev>> Abbrevs; |
| 73 | }; |
| 74 | std::vector<BlockInfo> BlockInfoRecords; |
| 75 | |
| 76 | void WriteByte(unsigned char Value) { |
| 77 | Out.push_back(Value); |
Olivier Deprez | f4ef2d0 | 2021-04-20 13:36:24 +0200 | [diff] [blame^] | 78 | FlushToFile(); |
Andrew Scull | 5e1ddfa | 2018-08-14 10:06:54 +0100 | [diff] [blame] | 79 | } |
| 80 | |
| 81 | void WriteWord(unsigned Value) { |
| 82 | Value = support::endian::byte_swap<uint32_t, support::little>(Value); |
| 83 | Out.append(reinterpret_cast<const char *>(&Value), |
| 84 | reinterpret_cast<const char *>(&Value + 1)); |
Olivier Deprez | f4ef2d0 | 2021-04-20 13:36:24 +0200 | [diff] [blame^] | 85 | FlushToFile(); |
Andrew Scull | 5e1ddfa | 2018-08-14 10:06:54 +0100 | [diff] [blame] | 86 | } |
| 87 | |
Olivier Deprez | f4ef2d0 | 2021-04-20 13:36:24 +0200 | [diff] [blame^] | 88 | uint64_t GetNumOfFlushedBytes() const { return FS ? FS->tell() : 0; } |
| 89 | |
| 90 | size_t GetBufferOffset() const { return Out.size() + GetNumOfFlushedBytes(); } |
Andrew Scull | 5e1ddfa | 2018-08-14 10:06:54 +0100 | [diff] [blame] | 91 | |
| 92 | size_t GetWordIndex() const { |
| 93 | size_t Offset = GetBufferOffset(); |
| 94 | assert((Offset & 3) == 0 && "Not 32-bit aligned"); |
| 95 | return Offset / 4; |
| 96 | } |
| 97 | |
Olivier Deprez | f4ef2d0 | 2021-04-20 13:36:24 +0200 | [diff] [blame^] | 98 | /// If the related file stream supports reading, seeking and writing, flush |
| 99 | /// the buffer if its size is above a threshold. |
| 100 | void FlushToFile() { |
| 101 | if (!FS) |
| 102 | return; |
| 103 | if (Out.size() < FlushThreshold) |
| 104 | return; |
| 105 | FS->write((char *)&Out.front(), Out.size()); |
| 106 | Out.clear(); |
| 107 | } |
| 108 | |
Andrew Scull | 5e1ddfa | 2018-08-14 10:06:54 +0100 | [diff] [blame] | 109 | public: |
Olivier Deprez | f4ef2d0 | 2021-04-20 13:36:24 +0200 | [diff] [blame^] | 110 | /// Create a BitstreamWriter that writes to Buffer \p O. |
| 111 | /// |
| 112 | /// \p FS is the file stream that \p O flushes to incrementally. If \p FS is |
| 113 | /// null, \p O does not flush incrementially, but writes to disk at the end. |
| 114 | /// |
| 115 | /// \p FlushThreshold is the threshold (unit M) to flush \p O if \p FS is |
| 116 | /// valid. |
| 117 | BitstreamWriter(SmallVectorImpl<char> &O, raw_fd_stream *FS = nullptr, |
| 118 | uint32_t FlushThreshold = 512) |
| 119 | : Out(O), FS(FS), FlushThreshold(FlushThreshold << 20), CurBit(0), |
| 120 | CurValue(0), CurCodeSize(2) {} |
Andrew Scull | 5e1ddfa | 2018-08-14 10:06:54 +0100 | [diff] [blame] | 121 | |
| 122 | ~BitstreamWriter() { |
| 123 | assert(CurBit == 0 && "Unflushed data remaining"); |
| 124 | assert(BlockScope.empty() && CurAbbrevs.empty() && "Block imbalance"); |
| 125 | } |
| 126 | |
Andrew Scull | cdfcccc | 2018-10-05 20:58:37 +0100 | [diff] [blame] | 127 | /// Retrieve the current position in the stream, in bits. |
Andrew Scull | 5e1ddfa | 2018-08-14 10:06:54 +0100 | [diff] [blame] | 128 | uint64_t GetCurrentBitNo() const { return GetBufferOffset() * 8 + CurBit; } |
| 129 | |
Andrew Scull | cdfcccc | 2018-10-05 20:58:37 +0100 | [diff] [blame] | 130 | /// Retrieve the number of bits currently used to encode an abbrev ID. |
Andrew Scull | 5e1ddfa | 2018-08-14 10:06:54 +0100 | [diff] [blame] | 131 | unsigned GetAbbrevIDWidth() const { return CurCodeSize; } |
| 132 | |
| 133 | //===--------------------------------------------------------------------===// |
| 134 | // Basic Primitives for emitting bits to the stream. |
| 135 | //===--------------------------------------------------------------------===// |
| 136 | |
| 137 | /// Backpatch a 32-bit word in the output at the given bit offset |
| 138 | /// with the specified value. |
| 139 | void BackpatchWord(uint64_t BitNo, unsigned NewWord) { |
| 140 | using namespace llvm::support; |
Olivier Deprez | f4ef2d0 | 2021-04-20 13:36:24 +0200 | [diff] [blame^] | 141 | uint64_t ByteNo = BitNo / 8; |
| 142 | uint64_t StartBit = BitNo & 7; |
| 143 | uint64_t NumOfFlushedBytes = GetNumOfFlushedBytes(); |
| 144 | |
| 145 | if (ByteNo >= NumOfFlushedBytes) { |
| 146 | assert((!endian::readAtBitAlignment<uint32_t, little, unaligned>( |
| 147 | &Out[ByteNo - NumOfFlushedBytes], StartBit)) && |
| 148 | "Expected to be patching over 0-value placeholders"); |
| 149 | endian::writeAtBitAlignment<uint32_t, little, unaligned>( |
| 150 | &Out[ByteNo - NumOfFlushedBytes], NewWord, StartBit); |
| 151 | return; |
| 152 | } |
| 153 | |
| 154 | // If the byte offset to backpatch is flushed, use seek to backfill data. |
| 155 | // First, save the file position to restore later. |
| 156 | uint64_t CurPos = FS->tell(); |
| 157 | |
| 158 | // Copy data to update into Bytes from the file FS and the buffer Out. |
| 159 | char Bytes[9]; // Use one more byte to silence a warning from Visual C++. |
| 160 | size_t BytesNum = StartBit ? 8 : 4; |
| 161 | size_t BytesFromDisk = std::min(static_cast<uint64_t>(BytesNum), NumOfFlushedBytes - ByteNo); |
| 162 | size_t BytesFromBuffer = BytesNum - BytesFromDisk; |
| 163 | |
| 164 | // When unaligned, copy existing data into Bytes from the file FS and the |
| 165 | // buffer Out so that it can be updated before writing. For debug builds |
| 166 | // read bytes unconditionally in order to check that the existing value is 0 |
| 167 | // as expected. |
| 168 | #ifdef NDEBUG |
| 169 | if (StartBit) |
| 170 | #endif |
| 171 | { |
| 172 | FS->seek(ByteNo); |
| 173 | ssize_t BytesRead = FS->read(Bytes, BytesFromDisk); |
| 174 | (void)BytesRead; // silence warning |
| 175 | assert(BytesRead >= 0 && static_cast<size_t>(BytesRead) == BytesFromDisk); |
| 176 | for (size_t i = 0; i < BytesFromBuffer; ++i) |
| 177 | Bytes[BytesFromDisk + i] = Out[i]; |
| 178 | assert((!endian::readAtBitAlignment<uint32_t, little, unaligned>( |
| 179 | Bytes, StartBit)) && |
| 180 | "Expected to be patching over 0-value placeholders"); |
| 181 | } |
| 182 | |
| 183 | // Update Bytes in terms of bit offset and value. |
| 184 | endian::writeAtBitAlignment<uint32_t, little, unaligned>(Bytes, NewWord, |
| 185 | StartBit); |
| 186 | |
| 187 | // Copy updated data back to the file FS and the buffer Out. |
| 188 | FS->seek(ByteNo); |
| 189 | FS->write(Bytes, BytesFromDisk); |
| 190 | for (size_t i = 0; i < BytesFromBuffer; ++i) |
| 191 | Out[i] = Bytes[BytesFromDisk + i]; |
| 192 | |
| 193 | // Restore the file position. |
| 194 | FS->seek(CurPos); |
Andrew Scull | 5e1ddfa | 2018-08-14 10:06:54 +0100 | [diff] [blame] | 195 | } |
| 196 | |
| 197 | void BackpatchWord64(uint64_t BitNo, uint64_t Val) { |
| 198 | BackpatchWord(BitNo, (uint32_t)Val); |
| 199 | BackpatchWord(BitNo + 32, (uint32_t)(Val >> 32)); |
| 200 | } |
| 201 | |
| 202 | void Emit(uint32_t Val, unsigned NumBits) { |
| 203 | assert(NumBits && NumBits <= 32 && "Invalid value size!"); |
| 204 | assert((Val & ~(~0U >> (32-NumBits))) == 0 && "High bits set!"); |
| 205 | CurValue |= Val << CurBit; |
| 206 | if (CurBit + NumBits < 32) { |
| 207 | CurBit += NumBits; |
| 208 | return; |
| 209 | } |
| 210 | |
| 211 | // Add the current word. |
| 212 | WriteWord(CurValue); |
| 213 | |
| 214 | if (CurBit) |
| 215 | CurValue = Val >> (32-CurBit); |
| 216 | else |
| 217 | CurValue = 0; |
| 218 | CurBit = (CurBit+NumBits) & 31; |
| 219 | } |
| 220 | |
| 221 | void FlushToWord() { |
| 222 | if (CurBit) { |
| 223 | WriteWord(CurValue); |
| 224 | CurBit = 0; |
| 225 | CurValue = 0; |
| 226 | } |
| 227 | } |
| 228 | |
| 229 | void EmitVBR(uint32_t Val, unsigned NumBits) { |
| 230 | assert(NumBits <= 32 && "Too many bits to emit!"); |
| 231 | uint32_t Threshold = 1U << (NumBits-1); |
| 232 | |
| 233 | // Emit the bits with VBR encoding, NumBits-1 bits at a time. |
| 234 | while (Val >= Threshold) { |
| 235 | Emit((Val & ((1 << (NumBits-1))-1)) | (1 << (NumBits-1)), NumBits); |
| 236 | Val >>= NumBits-1; |
| 237 | } |
| 238 | |
| 239 | Emit(Val, NumBits); |
| 240 | } |
| 241 | |
| 242 | void EmitVBR64(uint64_t Val, unsigned NumBits) { |
| 243 | assert(NumBits <= 32 && "Too many bits to emit!"); |
| 244 | if ((uint32_t)Val == Val) |
| 245 | return EmitVBR((uint32_t)Val, NumBits); |
| 246 | |
| 247 | uint32_t Threshold = 1U << (NumBits-1); |
| 248 | |
| 249 | // Emit the bits with VBR encoding, NumBits-1 bits at a time. |
| 250 | while (Val >= Threshold) { |
| 251 | Emit(((uint32_t)Val & ((1 << (NumBits-1))-1)) | |
| 252 | (1 << (NumBits-1)), NumBits); |
| 253 | Val >>= NumBits-1; |
| 254 | } |
| 255 | |
| 256 | Emit((uint32_t)Val, NumBits); |
| 257 | } |
| 258 | |
| 259 | /// EmitCode - Emit the specified code. |
| 260 | void EmitCode(unsigned Val) { |
| 261 | Emit(Val, CurCodeSize); |
| 262 | } |
| 263 | |
| 264 | //===--------------------------------------------------------------------===// |
| 265 | // Block Manipulation |
| 266 | //===--------------------------------------------------------------------===// |
| 267 | |
| 268 | /// getBlockInfo - If there is block info for the specified ID, return it, |
| 269 | /// otherwise return null. |
| 270 | BlockInfo *getBlockInfo(unsigned BlockID) { |
| 271 | // Common case, the most recent entry matches BlockID. |
| 272 | if (!BlockInfoRecords.empty() && BlockInfoRecords.back().BlockID == BlockID) |
| 273 | return &BlockInfoRecords.back(); |
| 274 | |
| 275 | for (unsigned i = 0, e = static_cast<unsigned>(BlockInfoRecords.size()); |
| 276 | i != e; ++i) |
| 277 | if (BlockInfoRecords[i].BlockID == BlockID) |
| 278 | return &BlockInfoRecords[i]; |
| 279 | return nullptr; |
| 280 | } |
| 281 | |
| 282 | void EnterSubblock(unsigned BlockID, unsigned CodeLen) { |
| 283 | // Block header: |
| 284 | // [ENTER_SUBBLOCK, blockid, newcodelen, <align4bytes>, blocklen] |
| 285 | EmitCode(bitc::ENTER_SUBBLOCK); |
| 286 | EmitVBR(BlockID, bitc::BlockIDWidth); |
| 287 | EmitVBR(CodeLen, bitc::CodeLenWidth); |
| 288 | FlushToWord(); |
| 289 | |
| 290 | size_t BlockSizeWordIndex = GetWordIndex(); |
| 291 | unsigned OldCodeSize = CurCodeSize; |
| 292 | |
| 293 | // Emit a placeholder, which will be replaced when the block is popped. |
| 294 | Emit(0, bitc::BlockSizeWidth); |
| 295 | |
| 296 | CurCodeSize = CodeLen; |
| 297 | |
| 298 | // Push the outer block's abbrev set onto the stack, start out with an |
| 299 | // empty abbrev set. |
| 300 | BlockScope.emplace_back(OldCodeSize, BlockSizeWordIndex); |
| 301 | BlockScope.back().PrevAbbrevs.swap(CurAbbrevs); |
| 302 | |
| 303 | // If there is a blockinfo for this BlockID, add all the predefined abbrevs |
| 304 | // to the abbrev list. |
| 305 | if (BlockInfo *Info = getBlockInfo(BlockID)) { |
| 306 | CurAbbrevs.insert(CurAbbrevs.end(), Info->Abbrevs.begin(), |
| 307 | Info->Abbrevs.end()); |
| 308 | } |
| 309 | } |
| 310 | |
| 311 | void ExitBlock() { |
| 312 | assert(!BlockScope.empty() && "Block scope imbalance!"); |
| 313 | const Block &B = BlockScope.back(); |
| 314 | |
| 315 | // Block tail: |
| 316 | // [END_BLOCK, <align4bytes>] |
| 317 | EmitCode(bitc::END_BLOCK); |
| 318 | FlushToWord(); |
| 319 | |
| 320 | // Compute the size of the block, in words, not counting the size field. |
| 321 | size_t SizeInWords = GetWordIndex() - B.StartSizeWord - 1; |
| 322 | uint64_t BitNo = uint64_t(B.StartSizeWord) * 32; |
| 323 | |
| 324 | // Update the block size field in the header of this sub-block. |
| 325 | BackpatchWord(BitNo, SizeInWords); |
| 326 | |
| 327 | // Restore the inner block's code size and abbrev table. |
| 328 | CurCodeSize = B.PrevCodeSize; |
| 329 | CurAbbrevs = std::move(B.PrevAbbrevs); |
| 330 | BlockScope.pop_back(); |
| 331 | } |
| 332 | |
| 333 | //===--------------------------------------------------------------------===// |
| 334 | // Record Emission |
| 335 | //===--------------------------------------------------------------------===// |
| 336 | |
| 337 | private: |
| 338 | /// EmitAbbreviatedLiteral - Emit a literal value according to its abbrev |
| 339 | /// record. This is a no-op, since the abbrev specifies the literal to use. |
| 340 | template<typename uintty> |
| 341 | void EmitAbbreviatedLiteral(const BitCodeAbbrevOp &Op, uintty V) { |
| 342 | assert(Op.isLiteral() && "Not a literal"); |
| 343 | // If the abbrev specifies the literal value to use, don't emit |
| 344 | // anything. |
| 345 | assert(V == Op.getLiteralValue() && |
| 346 | "Invalid abbrev for record!"); |
| 347 | } |
| 348 | |
| 349 | /// EmitAbbreviatedField - Emit a single scalar field value with the specified |
| 350 | /// encoding. |
| 351 | template<typename uintty> |
| 352 | void EmitAbbreviatedField(const BitCodeAbbrevOp &Op, uintty V) { |
| 353 | assert(!Op.isLiteral() && "Literals should use EmitAbbreviatedLiteral!"); |
| 354 | |
| 355 | // Encode the value as we are commanded. |
| 356 | switch (Op.getEncoding()) { |
| 357 | default: llvm_unreachable("Unknown encoding!"); |
| 358 | case BitCodeAbbrevOp::Fixed: |
| 359 | if (Op.getEncodingData()) |
| 360 | Emit((unsigned)V, (unsigned)Op.getEncodingData()); |
| 361 | break; |
| 362 | case BitCodeAbbrevOp::VBR: |
| 363 | if (Op.getEncodingData()) |
| 364 | EmitVBR64(V, (unsigned)Op.getEncodingData()); |
| 365 | break; |
| 366 | case BitCodeAbbrevOp::Char6: |
| 367 | Emit(BitCodeAbbrevOp::EncodeChar6((char)V), 6); |
| 368 | break; |
| 369 | } |
| 370 | } |
| 371 | |
| 372 | /// EmitRecordWithAbbrevImpl - This is the core implementation of the record |
| 373 | /// emission code. If BlobData is non-null, then it specifies an array of |
| 374 | /// data that should be emitted as part of the Blob or Array operand that is |
| 375 | /// known to exist at the end of the record. If Code is specified, then |
| 376 | /// it is the record code to emit before the Vals, which must not contain |
| 377 | /// the code. |
| 378 | template <typename uintty> |
| 379 | void EmitRecordWithAbbrevImpl(unsigned Abbrev, ArrayRef<uintty> Vals, |
| 380 | StringRef Blob, Optional<unsigned> Code) { |
| 381 | const char *BlobData = Blob.data(); |
| 382 | unsigned BlobLen = (unsigned) Blob.size(); |
| 383 | unsigned AbbrevNo = Abbrev-bitc::FIRST_APPLICATION_ABBREV; |
| 384 | assert(AbbrevNo < CurAbbrevs.size() && "Invalid abbrev #!"); |
| 385 | const BitCodeAbbrev *Abbv = CurAbbrevs[AbbrevNo].get(); |
| 386 | |
| 387 | EmitCode(Abbrev); |
| 388 | |
| 389 | unsigned i = 0, e = static_cast<unsigned>(Abbv->getNumOperandInfos()); |
| 390 | if (Code) { |
| 391 | assert(e && "Expected non-empty abbreviation"); |
| 392 | const BitCodeAbbrevOp &Op = Abbv->getOperandInfo(i++); |
| 393 | |
| 394 | if (Op.isLiteral()) |
| 395 | EmitAbbreviatedLiteral(Op, Code.getValue()); |
| 396 | else { |
| 397 | assert(Op.getEncoding() != BitCodeAbbrevOp::Array && |
| 398 | Op.getEncoding() != BitCodeAbbrevOp::Blob && |
| 399 | "Expected literal or scalar"); |
| 400 | EmitAbbreviatedField(Op, Code.getValue()); |
| 401 | } |
| 402 | } |
| 403 | |
| 404 | unsigned RecordIdx = 0; |
| 405 | for (; i != e; ++i) { |
| 406 | const BitCodeAbbrevOp &Op = Abbv->getOperandInfo(i); |
| 407 | if (Op.isLiteral()) { |
| 408 | assert(RecordIdx < Vals.size() && "Invalid abbrev/record"); |
| 409 | EmitAbbreviatedLiteral(Op, Vals[RecordIdx]); |
| 410 | ++RecordIdx; |
| 411 | } else if (Op.getEncoding() == BitCodeAbbrevOp::Array) { |
| 412 | // Array case. |
| 413 | assert(i + 2 == e && "array op not second to last?"); |
| 414 | const BitCodeAbbrevOp &EltEnc = Abbv->getOperandInfo(++i); |
| 415 | |
| 416 | // If this record has blob data, emit it, otherwise we must have record |
| 417 | // entries to encode this way. |
| 418 | if (BlobData) { |
| 419 | assert(RecordIdx == Vals.size() && |
| 420 | "Blob data and record entries specified for array!"); |
| 421 | // Emit a vbr6 to indicate the number of elements present. |
| 422 | EmitVBR(static_cast<uint32_t>(BlobLen), 6); |
| 423 | |
| 424 | // Emit each field. |
| 425 | for (unsigned i = 0; i != BlobLen; ++i) |
| 426 | EmitAbbreviatedField(EltEnc, (unsigned char)BlobData[i]); |
| 427 | |
| 428 | // Know that blob data is consumed for assertion below. |
| 429 | BlobData = nullptr; |
| 430 | } else { |
| 431 | // Emit a vbr6 to indicate the number of elements present. |
| 432 | EmitVBR(static_cast<uint32_t>(Vals.size()-RecordIdx), 6); |
| 433 | |
| 434 | // Emit each field. |
| 435 | for (unsigned e = Vals.size(); RecordIdx != e; ++RecordIdx) |
| 436 | EmitAbbreviatedField(EltEnc, Vals[RecordIdx]); |
| 437 | } |
| 438 | } else if (Op.getEncoding() == BitCodeAbbrevOp::Blob) { |
| 439 | // If this record has blob data, emit it, otherwise we must have record |
| 440 | // entries to encode this way. |
| 441 | |
| 442 | if (BlobData) { |
| 443 | assert(RecordIdx == Vals.size() && |
| 444 | "Blob data and record entries specified for blob operand!"); |
| 445 | |
| 446 | assert(Blob.data() == BlobData && "BlobData got moved"); |
| 447 | assert(Blob.size() == BlobLen && "BlobLen got changed"); |
| 448 | emitBlob(Blob); |
| 449 | BlobData = nullptr; |
| 450 | } else { |
| 451 | emitBlob(Vals.slice(RecordIdx)); |
| 452 | } |
| 453 | } else { // Single scalar field. |
| 454 | assert(RecordIdx < Vals.size() && "Invalid abbrev/record"); |
| 455 | EmitAbbreviatedField(Op, Vals[RecordIdx]); |
| 456 | ++RecordIdx; |
| 457 | } |
| 458 | } |
| 459 | assert(RecordIdx == Vals.size() && "Not all record operands emitted!"); |
| 460 | assert(BlobData == nullptr && |
| 461 | "Blob data specified for record that doesn't use it!"); |
| 462 | } |
| 463 | |
| 464 | public: |
| 465 | /// Emit a blob, including flushing before and tail-padding. |
| 466 | template <class UIntTy> |
| 467 | void emitBlob(ArrayRef<UIntTy> Bytes, bool ShouldEmitSize = true) { |
| 468 | // Emit a vbr6 to indicate the number of elements present. |
| 469 | if (ShouldEmitSize) |
| 470 | EmitVBR(static_cast<uint32_t>(Bytes.size()), 6); |
| 471 | |
| 472 | // Flush to a 32-bit alignment boundary. |
| 473 | FlushToWord(); |
| 474 | |
| 475 | // Emit literal bytes. |
| 476 | for (const auto &B : Bytes) { |
| 477 | assert(isUInt<8>(B) && "Value too large to emit as byte"); |
| 478 | WriteByte((unsigned char)B); |
| 479 | } |
| 480 | |
| 481 | // Align end to 32-bits. |
| 482 | while (GetBufferOffset() & 3) |
| 483 | WriteByte(0); |
| 484 | } |
| 485 | void emitBlob(StringRef Bytes, bool ShouldEmitSize = true) { |
| 486 | emitBlob(makeArrayRef((const uint8_t *)Bytes.data(), Bytes.size()), |
| 487 | ShouldEmitSize); |
| 488 | } |
| 489 | |
| 490 | /// EmitRecord - Emit the specified record to the stream, using an abbrev if |
| 491 | /// we have one to compress the output. |
| 492 | template <typename Container> |
| 493 | void EmitRecord(unsigned Code, const Container &Vals, unsigned Abbrev = 0) { |
| 494 | if (!Abbrev) { |
| 495 | // If we don't have an abbrev to use, emit this in its fully unabbreviated |
| 496 | // form. |
| 497 | auto Count = static_cast<uint32_t>(makeArrayRef(Vals).size()); |
| 498 | EmitCode(bitc::UNABBREV_RECORD); |
| 499 | EmitVBR(Code, 6); |
| 500 | EmitVBR(Count, 6); |
| 501 | for (unsigned i = 0, e = Count; i != e; ++i) |
| 502 | EmitVBR64(Vals[i], 6); |
| 503 | return; |
| 504 | } |
| 505 | |
| 506 | EmitRecordWithAbbrevImpl(Abbrev, makeArrayRef(Vals), StringRef(), Code); |
| 507 | } |
| 508 | |
| 509 | /// EmitRecordWithAbbrev - Emit a record with the specified abbreviation. |
| 510 | /// Unlike EmitRecord, the code for the record should be included in Vals as |
| 511 | /// the first entry. |
| 512 | template <typename Container> |
| 513 | void EmitRecordWithAbbrev(unsigned Abbrev, const Container &Vals) { |
| 514 | EmitRecordWithAbbrevImpl(Abbrev, makeArrayRef(Vals), StringRef(), None); |
| 515 | } |
| 516 | |
| 517 | /// EmitRecordWithBlob - Emit the specified record to the stream, using an |
| 518 | /// abbrev that includes a blob at the end. The blob data to emit is |
| 519 | /// specified by the pointer and length specified at the end. In contrast to |
| 520 | /// EmitRecord, this routine expects that the first entry in Vals is the code |
| 521 | /// of the record. |
| 522 | template <typename Container> |
| 523 | void EmitRecordWithBlob(unsigned Abbrev, const Container &Vals, |
| 524 | StringRef Blob) { |
| 525 | EmitRecordWithAbbrevImpl(Abbrev, makeArrayRef(Vals), Blob, None); |
| 526 | } |
| 527 | template <typename Container> |
| 528 | void EmitRecordWithBlob(unsigned Abbrev, const Container &Vals, |
| 529 | const char *BlobData, unsigned BlobLen) { |
| 530 | return EmitRecordWithAbbrevImpl(Abbrev, makeArrayRef(Vals), |
| 531 | StringRef(BlobData, BlobLen), None); |
| 532 | } |
| 533 | |
| 534 | /// EmitRecordWithArray - Just like EmitRecordWithBlob, works with records |
| 535 | /// that end with an array. |
| 536 | template <typename Container> |
| 537 | void EmitRecordWithArray(unsigned Abbrev, const Container &Vals, |
| 538 | StringRef Array) { |
| 539 | EmitRecordWithAbbrevImpl(Abbrev, makeArrayRef(Vals), Array, None); |
| 540 | } |
| 541 | template <typename Container> |
| 542 | void EmitRecordWithArray(unsigned Abbrev, const Container &Vals, |
| 543 | const char *ArrayData, unsigned ArrayLen) { |
| 544 | return EmitRecordWithAbbrevImpl(Abbrev, makeArrayRef(Vals), |
| 545 | StringRef(ArrayData, ArrayLen), None); |
| 546 | } |
| 547 | |
| 548 | //===--------------------------------------------------------------------===// |
| 549 | // Abbrev Emission |
| 550 | //===--------------------------------------------------------------------===// |
| 551 | |
| 552 | private: |
| 553 | // Emit the abbreviation as a DEFINE_ABBREV record. |
| 554 | void EncodeAbbrev(const BitCodeAbbrev &Abbv) { |
| 555 | EmitCode(bitc::DEFINE_ABBREV); |
| 556 | EmitVBR(Abbv.getNumOperandInfos(), 5); |
| 557 | for (unsigned i = 0, e = static_cast<unsigned>(Abbv.getNumOperandInfos()); |
| 558 | i != e; ++i) { |
| 559 | const BitCodeAbbrevOp &Op = Abbv.getOperandInfo(i); |
| 560 | Emit(Op.isLiteral(), 1); |
| 561 | if (Op.isLiteral()) { |
| 562 | EmitVBR64(Op.getLiteralValue(), 8); |
| 563 | } else { |
| 564 | Emit(Op.getEncoding(), 3); |
| 565 | if (Op.hasEncodingData()) |
| 566 | EmitVBR64(Op.getEncodingData(), 5); |
| 567 | } |
| 568 | } |
| 569 | } |
| 570 | public: |
| 571 | |
Andrew Walbran | 3d2c197 | 2020-04-07 12:24:26 +0100 | [diff] [blame] | 572 | /// Emits the abbreviation \p Abbv to the stream. |
Andrew Scull | 5e1ddfa | 2018-08-14 10:06:54 +0100 | [diff] [blame] | 573 | unsigned EmitAbbrev(std::shared_ptr<BitCodeAbbrev> Abbv) { |
Andrew Scull | 5e1ddfa | 2018-08-14 10:06:54 +0100 | [diff] [blame] | 574 | EncodeAbbrev(*Abbv); |
| 575 | CurAbbrevs.push_back(std::move(Abbv)); |
| 576 | return static_cast<unsigned>(CurAbbrevs.size())-1 + |
| 577 | bitc::FIRST_APPLICATION_ABBREV; |
| 578 | } |
| 579 | |
| 580 | //===--------------------------------------------------------------------===// |
| 581 | // BlockInfo Block Emission |
| 582 | //===--------------------------------------------------------------------===// |
| 583 | |
| 584 | /// EnterBlockInfoBlock - Start emitting the BLOCKINFO_BLOCK. |
| 585 | void EnterBlockInfoBlock() { |
| 586 | EnterSubblock(bitc::BLOCKINFO_BLOCK_ID, 2); |
| 587 | BlockInfoCurBID = ~0U; |
| 588 | BlockInfoRecords.clear(); |
| 589 | } |
| 590 | private: |
| 591 | /// SwitchToBlockID - If we aren't already talking about the specified block |
| 592 | /// ID, emit a BLOCKINFO_CODE_SETBID record. |
| 593 | void SwitchToBlockID(unsigned BlockID) { |
| 594 | if (BlockInfoCurBID == BlockID) return; |
| 595 | SmallVector<unsigned, 2> V; |
| 596 | V.push_back(BlockID); |
| 597 | EmitRecord(bitc::BLOCKINFO_CODE_SETBID, V); |
| 598 | BlockInfoCurBID = BlockID; |
| 599 | } |
| 600 | |
| 601 | BlockInfo &getOrCreateBlockInfo(unsigned BlockID) { |
| 602 | if (BlockInfo *BI = getBlockInfo(BlockID)) |
| 603 | return *BI; |
| 604 | |
| 605 | // Otherwise, add a new record. |
| 606 | BlockInfoRecords.emplace_back(); |
| 607 | BlockInfoRecords.back().BlockID = BlockID; |
| 608 | return BlockInfoRecords.back(); |
| 609 | } |
| 610 | |
| 611 | public: |
| 612 | |
| 613 | /// EmitBlockInfoAbbrev - Emit a DEFINE_ABBREV record for the specified |
| 614 | /// BlockID. |
| 615 | unsigned EmitBlockInfoAbbrev(unsigned BlockID, std::shared_ptr<BitCodeAbbrev> Abbv) { |
| 616 | SwitchToBlockID(BlockID); |
| 617 | EncodeAbbrev(*Abbv); |
| 618 | |
| 619 | // Add the abbrev to the specified block record. |
| 620 | BlockInfo &Info = getOrCreateBlockInfo(BlockID); |
| 621 | Info.Abbrevs.push_back(std::move(Abbv)); |
| 622 | |
| 623 | return Info.Abbrevs.size()-1+bitc::FIRST_APPLICATION_ABBREV; |
| 624 | } |
| 625 | }; |
| 626 | |
| 627 | |
| 628 | } // End llvm namespace |
| 629 | |
| 630 | #endif |