wimlib.net Git - wimlib/blob - src/lzx-decompress.c

   1 /*
   2  * lzx-decompress.c
   3  *
   4  * A very fast decompressor for LZX, as used in WIM files.
   5  */
   6
   7 /*
   8  * Copyright (C) 2012, 2013, 2014 Eric Biggers
   9  *
  10  * This file is part of wimlib, a library for working with WIM files.
  11  *
  12  * wimlib is free software; you can redistribute it and/or modify it under the
  13  * terms of the GNU General Public License as published by the Free
  14  * Software Foundation; either version 3 of the License, or (at your option)
  15  * any later version.
  16  *
  17  * wimlib is distributed in the hope that it will be useful, but WITHOUT ANY
  18  * WARRANTY; without even the implied warranty of MERCHANTABILITY or FITNESS FOR
  19  * A PARTICULAR PURPOSE. See the GNU General Public License for more
  20  * details.
  21  *
  22  * You should have received a copy of the GNU General Public License
  23  * along with wimlib; if not, see http://www.gnu.org/licenses/.
  24  */
  25
  26 /*
  27  * LZX is an LZ77 and Huffman-code based compression format that has many
  28  * similarities to DEFLATE (the format used by zlib/gzip).  The compression
  29  * ratio is as good or better than DEFLATE.  See lzx-compress.c for a format
  30  * overview, and see https://en.wikipedia.org/wiki/LZX_(algorithm) for a
  31  * historical overview.  Here I make some pragmatic notes.
  32  *
  33  * The old specification for LZX is the document "Microsoft LZX Data Compression
  34  * Format" (1997).  It defines the LZX format as used in cabinet files.  Allowed
  35  * window sizes are 2^n where 15 <= n <= 21.  However, this document contains
  36  * several errors, so don't read too much into it...
  37  *
  38  * The new specification for LZX is the document "[MS-PATCH]: LZX DELTA
  39  * Compression and Decompression" (2014).  It defines the LZX format as used by
  40  * Microsoft's binary patcher.  It corrects several errors in the 1997 document
  41  * and extends the format in several ways --- namely, optional reference data,
  42  * up to 2^25 byte windows, and longer match lengths.
  43  *
  44  * WIM files use a more restricted form of LZX.  No LZX DELTA extensions are
  45  * present, the window is not "sliding", E8 preprocessing is done
  46  * unconditionally with a fixed file size, and the maximum window size is always
  47  * 2^15 bytes (equal to the size of each "chunk" in a compressed WIM resource).
  48  * This code is primarily intended to implement this form of LZX.  But although
  49  * not compatible with WIMGAPI, this code also supports maximum window sizes up
  50  * to 2^21 bytes.
  51  *
  52  * TODO: Add support for window sizes up to 2^25 bytes.
  53  */
  54
  55 #ifdef HAVE_CONFIG_H
  56 #  include "config.h"
  57 #endif
  58
  59 #include "wimlib/decompressor_ops.h"
  60 #include "wimlib/decompress_common.h"
  61 #include "wimlib/error.h"
  62 #include "wimlib/lzx.h"
  63 #include "wimlib/util.h"
  64
  65 #include <string.h>
  66
  67 /* These values are chosen for fast decompression.  */
  68 #define LZX_MAINCODE_TABLEBITS          11
  69 #define LZX_LENCODE_TABLEBITS           10
  70 #define LZX_PRECODE_TABLEBITS           6
  71 #define LZX_ALIGNEDCODE_TABLEBITS       7
  72
  73 /* Huffman decoding tables, and arrays that map symbols to codeword lengths.  */
  74 struct lzx_tables {
  75
  76         u16 maincode_decode_table[(1 << LZX_MAINCODE_TABLEBITS) +
  77                                         (LZX_MAINCODE_MAX_NUM_SYMBOLS * 2)]
  78                                         _aligned_attribute(DECODE_TABLE_ALIGNMENT);
  79         u8 maincode_lens[LZX_MAINCODE_MAX_NUM_SYMBOLS];
  80
  81
  82         u16 lencode_decode_table[(1 << LZX_LENCODE_TABLEBITS) +
  83                                         (LZX_LENCODE_NUM_SYMBOLS * 2)]
  84                                         _aligned_attribute(DECODE_TABLE_ALIGNMENT);
  85         u8 lencode_lens[LZX_LENCODE_NUM_SYMBOLS];
  86
  87
  88         u16 alignedcode_decode_table[(1 << LZX_ALIGNEDCODE_TABLEBITS) +
  89                                         (LZX_ALIGNEDCODE_NUM_SYMBOLS * 2)]
  90                                         _aligned_attribute(DECODE_TABLE_ALIGNMENT);
  91         u8 alignedcode_lens[LZX_ALIGNEDCODE_NUM_SYMBOLS];
  92 } _aligned_attribute(DECODE_TABLE_ALIGNMENT);
  93
  94 /* The main LZX decompressor structure.
  95  *
  96  * Note: we keep track of most of the decompression state outside this
  97  * structure.  This structure only exists so that (1) we can store @window_order
  98  * and @num_main_syms for multiple calls to lzx_decompress(); and (2) so that we
  99  * don't have to allocate the large 'struct lzx_tables' on the stack.  */
 100 struct lzx_decompressor {
 101         unsigned window_order;
 102         unsigned num_main_syms;
 103         struct lzx_tables tables;
 104 };
 105
 106 /* Read a Huffman-encoded symbol using the precode.  */
 107 static inline u16
 108 read_huffsym_using_precode(struct input_bitstream *istream,
 109                            const u16 precode_decode_table[])
 110 {
 111         return read_huffsym(istream, precode_decode_table,
 112                             LZX_PRECODE_TABLEBITS, LZX_MAX_PRE_CODEWORD_LEN);
 113 }
 114
 115 /* Read a Huffman-encoded symbol using the main code.  */
 116 static inline u16
 117 read_huffsym_using_maincode(struct input_bitstream *istream,
 118                             const struct lzx_tables *tables)
 119 {
 120         return read_huffsym(istream, tables->maincode_decode_table,
 121                             LZX_MAINCODE_TABLEBITS, LZX_MAX_MAIN_CODEWORD_LEN);
 122 }
 123
 124 /* Read a Huffman-encoded symbol using the length code.  */
 125 static inline u16
 126 read_huffsym_using_lencode(struct input_bitstream *istream,
 127                            const struct lzx_tables *tables)
 128 {
 129         return read_huffsym(istream, tables->lencode_decode_table,
 130                             LZX_LENCODE_TABLEBITS, LZX_MAX_LEN_CODEWORD_LEN);
 131 }
 132
 133 /* Read a Huffman-encoded symbol using the aligned offset code.  */
 134 static inline u16
 135 read_huffsym_using_alignedcode(struct input_bitstream *istream,
 136                                const struct lzx_tables *tables)
 137 {
 138         return read_huffsym(istream, tables->alignedcode_decode_table,
 139                             LZX_ALIGNEDCODE_TABLEBITS, LZX_MAX_ALIGNED_CODEWORD_LEN);
 140 }
 141
 142 /*
 143  * Read the precode from the compressed input bitstream, then use it to decode
 144  * @num_lens codeword length values.
 145  *
 146  * @istream:
 147  *      The input bitstream.
 148  *
 149  * @lens:
 150  *      An array that contains the length values from the previous time the
 151  *      codeword lengths for this Huffman code were read, or all 0's if this is
 152  *      the first time.
 153  *
 154  * @num_lens:
 155  *      Number of length values to decode.
 156  *
 157  * Returns 0 on success, or -1 if the data was invalid.
 158  */
 159 static int
 160 lzx_read_codeword_lens(struct input_bitstream *istream, u8 lens[], unsigned num_lens)
 161 {
 162         /* Declare the decoding table and length table for the precode.  */
 163         u16 precode_decode_table[(1 << LZX_PRECODE_TABLEBITS) +
 164                                         (LZX_PRECODE_NUM_SYMBOLS * 2)]
 165                                         _aligned_attribute(DECODE_TABLE_ALIGNMENT);
 166         u8 precode_lens[LZX_PRECODE_NUM_SYMBOLS];
 167         int ret;
 168
 169         /* Read the lengths of the precode codewords.  These are given
 170          * explicitly.  */
 171         for (int i = 0; i < LZX_PRECODE_NUM_SYMBOLS; i++) {
 172                 precode_lens[i] = bitstream_read_bits(istream,
 173                                                       LZX_PRECODE_ELEMENT_SIZE);
 174         }
 175
 176         /* Make the decoding table for the precode.  */
 177         ret = make_huffman_decode_table(precode_decode_table,
 178                                         LZX_PRECODE_NUM_SYMBOLS,
 179                                         LZX_PRECODE_TABLEBITS,
 180                                         precode_lens,
 181                                         LZX_MAX_PRE_CODEWORD_LEN);
 182         if (ret)
 183                 return ret;
 184
 185         /* Pointer past the last length value that needs to be filled in.  */
 186         u8 *lens_end = lens + num_lens;
 187
 188         for (;;) {
 189
 190                 unsigned presym;
 191                 unsigned run_len;
 192                 signed char value;
 193
 194                 /* Decode a symbol from the input.
 195                  *
 196                  * If the symbol is between 0 and 16, it is the difference from
 197                  * the old length, modulo 17.
 198                  *
 199                  * If the symbol is between 17 and 19, it is a special symbol
 200                  * that indicates that some number of the next lengths are all
 201                  * 0, or that some number of the next lengths are all equal to
 202                  * the next symbol.  */
 203
 204                 presym = read_huffsym_using_precode(istream,
 205                                                     precode_decode_table);
 206                 switch (presym) {
 207
 208                 case 17: /* Run of 0's  */
 209                         run_len = 4 + bitstream_read_bits(istream, 4);
 210                         do {
 211                                 *lens = 0;
 212                                 if (++lens == lens_end)
 213                                         return 0;
 214                         } while (--run_len);
 215                         break;
 216
 217                 case 18: /* Longer run of 0's  */
 218                         run_len = 20 + bitstream_read_bits(istream, 5);
 219                         do {
 220                                 *lens = 0;
 221                                 if (++lens == lens_end)
 222                                         return 0;
 223                         } while (--run_len);
 224                         break;
 225
 226                 case 19: /* Run of identical lengths  */
 227                         run_len = 4 + bitstream_read_bits(istream, 1);
 228                         presym = read_huffsym_using_precode(istream,
 229                                                             precode_decode_table);
 230                         value = (signed char)*lens - (signed char)presym;
 231                         if (value < 0)
 232                                 value += 17;
 233                         do {
 234                                 *lens = value;
 235                                 if (++lens == lens_end)
 236                                         return 0;
 237                         } while (--run_len);
 238                         break;
 239
 240                 default: /* Difference from old length  */
 241                         value = (signed char)*lens - (signed char)presym;
 242                         if (value < 0)
 243                                 value += 17;
 244                         *lens = value;
 245                         if (++lens == lens_end)
 246                                 return 0;
 247                         break;
 248                 }
 249         }
 250 }
 251
 252 /*
 253  * Read the header of an LZX block and save the block type and size in
 254  * *block_type_ret and *block_size_ret, respectively.
 255  *
 256  * If the block is compressed, also update the Huffman decode @tables with the
 257  * new Huffman codes.
 258  *
 259  * If the block is uncompressed, also update the match offset @queue with the
 260  * new match offsets.
 261  *
 262  * Return 0 on success, or -1 if the data was invalid.
 263  */
 264 static int
 265 lzx_read_block_header(struct input_bitstream *istream,
 266                       unsigned num_main_syms,
 267                       unsigned window_order,
 268                       int *block_type_ret,
 269                       u32 *block_size_ret,
 270                       struct lzx_tables *tables,
 271                       struct lzx_lru_queue *queue)
 272 {
 273         int block_type;
 274         u32 block_size;
 275         int ret;
 276
 277         bitstream_ensure_bits(istream, 4);
 278
 279         /* The first three bits tell us what kind of block it is, and should be
 280          * one of the LZX_BLOCKTYPE_* values.  */
 281         block_type = bitstream_pop_bits(istream, 3);
 282
 283         /* Read the block size.  This mirrors the behavior of
 284          * lzx_write_compressed_block() in lzx-compress.c; see that for more
 285          * details.  */
 286         if (bitstream_pop_bits(istream, 1)) {
 287                 block_size = LZX_DEFAULT_BLOCK_SIZE;
 288         } else {
 289                 u32 tmp;
 290                 block_size = 0;
 291
 292                 tmp = bitstream_read_bits(istream, 8);
 293                 block_size |= tmp;
 294                 tmp = bitstream_read_bits(istream, 8);
 295                 block_size <<= 8;
 296                 block_size |= tmp;
 297
 298                 if (window_order >= 16) {
 299                         tmp = bitstream_read_bits(istream, 8);
 300                         block_size <<= 8;
 301                         block_size |= tmp;
 302                 }
 303         }
 304
 305         switch (block_type) {
 306
 307         case LZX_BLOCKTYPE_ALIGNED:
 308
 309                 /* Read the aligned offset code and prepare its decode table.
 310                  */
 311
 312                 for (int i = 0; i < LZX_ALIGNEDCODE_NUM_SYMBOLS; i++) {
 313                         tables->alignedcode_lens[i] =
 314                                 bitstream_read_bits(istream,
 315                                                     LZX_ALIGNEDCODE_ELEMENT_SIZE);
 316                 }
 317
 318                 ret = make_huffman_decode_table(tables->alignedcode_decode_table,
 319                                                 LZX_ALIGNEDCODE_NUM_SYMBOLS,
 320                                                 LZX_ALIGNEDCODE_TABLEBITS,
 321                                                 tables->alignedcode_lens,
 322                                                 LZX_MAX_ALIGNED_CODEWORD_LEN);
 323                 if (ret)
 324                         return ret;
 325
 326                 /* Fall though, since the rest of the header for aligned offset
 327                  * blocks is the same as that for verbatim blocks.  */
 328
 329         case LZX_BLOCKTYPE_VERBATIM:
 330
 331                 /* Read the main code and prepare its decode table.
 332                  *
 333                  * Note that the codeword lengths in the main code are encoded
 334                  * in two parts: one part for literal symbols, and one part for
 335                  * match symbols.  */
 336
 337                 ret = lzx_read_codeword_lens(istream, tables->maincode_lens,
 338                                              LZX_NUM_CHARS);
 339                 if (ret)
 340                         return ret;
 341
 342                 ret = lzx_read_codeword_lens(istream,
 343                                              tables->maincode_lens + LZX_NUM_CHARS,
 344                                              num_main_syms - LZX_NUM_CHARS);
 345                 if (ret)
 346                         return ret;
 347
 348                 ret = make_huffman_decode_table(tables->maincode_decode_table,
 349                                                 num_main_syms,
 350                                                 LZX_MAINCODE_TABLEBITS,
 351                                                 tables->maincode_lens,
 352                                                 LZX_MAX_MAIN_CODEWORD_LEN);
 353                 if (ret)
 354                         return ret;
 355
 356                 /* Read the length code and prepare its decode table.  */
 357
 358                 ret = lzx_read_codeword_lens(istream, tables->lencode_lens,
 359                                              LZX_LENCODE_NUM_SYMBOLS);
 360                 if (ret)
 361                         return ret;
 362
 363                 ret = make_huffman_decode_table(tables->lencode_decode_table,
 364                                                 LZX_LENCODE_NUM_SYMBOLS,
 365                                                 LZX_LENCODE_TABLEBITS,
 366                                                 tables->lencode_lens,
 367                                                 LZX_MAX_LEN_CODEWORD_LEN);
 368                 if (ret)
 369                         return ret;
 370
 371                 break;
 372
 373         case LZX_BLOCKTYPE_UNCOMPRESSED:
 374
 375                 /* Before reading the three LRU match offsets from the
 376                  * uncompressed block header, the stream must be aligned on a
 377                  * 16-bit boundary.  But, unexpectedly, if the stream is
 378                  * *already* aligned, the correct thing to do is to throw away
 379                  * the next 16 bits.  */
 380
 381                 bitstream_ensure_bits(istream, 1);
 382                 bitstream_align(istream);
 383                 queue->R[0] = bitstream_read_u32(istream);
 384                 queue->R[1] = bitstream_read_u32(istream);
 385                 queue->R[2] = bitstream_read_u32(istream);
 386
 387                 /* Offsets of 0 are invalid.  */
 388                 if (queue->R[0] == 0 || queue->R[1] == 0 || queue->R[2] == 0)
 389                         return -1;
 390                 break;
 391
 392         default:
 393                 /* Unrecognized block type.  */
 394                 return -1;
 395         }
 396
 397         *block_type_ret = block_type;
 398         *block_size_ret = block_size;
 399         return 0;
 400 }
 401
 402 /*
 403  * Decode a match and copy its bytes into the decompression window.
 404  *
 405  * Return the length of the match in bytes, or 0 if the match underflowed the
 406  * window or overflowed the current block.
 407  */
 408 static u32
 409 lzx_decode_match(unsigned main_symbol, int block_type,
 410                  u32 bytes_remaining, u8 *window, u32 window_pos,
 411                  const struct lzx_tables *tables,
 412                  struct lzx_lru_queue *queue,
 413                  struct input_bitstream *istream)
 414 {
 415         unsigned length_header;
 416         unsigned position_slot;
 417         u32 match_len;
 418         u32 match_offset;
 419         unsigned num_extra_bits;
 420         u32 verbatim_bits;
 421         u32 aligned_bits;
 422
 423         /* The main symbol is offset by 256 because values under 256 indicate a
 424          * literal value.  */
 425         main_symbol -= LZX_NUM_CHARS;
 426
 427         /* The length header consists of the lower 3 bits of the main element.
 428          * The position slot is the rest of it. */
 429         length_header = main_symbol & LZX_NUM_PRIMARY_LENS;
 430         position_slot = main_symbol >> 3;
 431
 432         /* If the length_header is less than LZX_NUM_PRIMARY_LENS (= 7), it
 433          * gives the match length as the offset from LZX_MIN_MATCH_LEN.
 434          * Otherwise, the length is given by an additional symbol encoded using
 435          * the length code, offset by 9 (LZX_MIN_MATCH_LEN +
 436          * LZX_NUM_PRIMARY_LENS) */
 437         match_len = LZX_MIN_MATCH_LEN + length_header;
 438         if (length_header == LZX_NUM_PRIMARY_LENS)
 439                 match_len += read_huffsym_using_lencode(istream, tables);
 440
 441         /* If the position_slot is 0, 1, or 2, the match offset is retrieved
 442          * from the LRU queue.  Otherwise, the match offset is not in the LRU
 443          * queue. */
 444         if (position_slot <= 2) {
 445                 /* Note: This isn't a real LRU queue, since using the R2 offset
 446                  * doesn't bump the R1 offset down to R2.  This quirk allows all
 447                  * 3 recent offsets to be handled by the same code.  (For R0,
 448                  * the swap is a no-op.)  */
 449                 match_offset = queue->R[position_slot];
 450                 queue->R[position_slot] = queue->R[0];
 451                 queue->R[0] = match_offset;
 452         } else {
 453                 /* Otherwise, the offset was not encoded as one the offsets in
 454                  * the queue.  Depending on the position slot, there is a
 455                  * certain number of extra bits that need to be read to fully
 456                  * decode the match offset. */
 457
 458                 /* Look up the number of extra bits that need to be read. */
 459                 num_extra_bits = lzx_get_num_extra_bits(position_slot);
 460
 461                 /* For aligned blocks, if there are at least 3 extra bits, the
 462                  * actual number of extra bits is 3 less, and they encode a
 463                  * number of 8-byte words that are added to the offset; there
 464                  * is then an additional symbol read using the aligned offset
 465                  * code that specifies the actual byte alignment. */
 466                 if (block_type == LZX_BLOCKTYPE_ALIGNED && num_extra_bits >= 3) {
 467
 468                         /* There is an error in the LZX "specification" at this
 469                          * point; it indicates that a Huffman symbol is to be
 470                          * read only if num_extra_bits is greater than 3, but
 471                          * actually it is if num_extra_bits is greater than or
 472                          * equal to 3.  (Note that in the case with
 473                          * num_extra_bits == 3, the assignment to verbatim_bits
 474                          * will just set it to 0. ) */
 475                         verbatim_bits = bitstream_read_bits(istream,
 476                                                             num_extra_bits - 3);
 477                         verbatim_bits <<= 3;
 478                         aligned_bits = read_huffsym_using_alignedcode(istream,
 479                                                                       tables);
 480                 } else {
 481                         /* For non-aligned blocks, or for aligned blocks with
 482                          * less than 3 extra bits, the extra bits are added
 483                          * directly to the match offset, and the correction for
 484                          * the alignment is taken to be 0. */
 485                         verbatim_bits = bitstream_read_bits(istream, num_extra_bits);
 486                         aligned_bits = 0;
 487                 }
 488
 489                 /* Calculate the match offset. */
 490                 match_offset = lzx_position_base[position_slot] +
 491                                verbatim_bits + aligned_bits - LZX_OFFSET_OFFSET;
 492
 493                 /* Update the LRU queue. */
 494                 queue->R[2] = queue->R[1];
 495                 queue->R[1] = queue->R[0];
 496                 queue->R[0] = match_offset;
 497         }
 498
 499         /* Validate the match, then copy it to the current position.  */
 500
 501         if (unlikely(match_len > bytes_remaining))
 502                 return 0;
 503
 504         if (unlikely(match_offset > window_pos))
 505                 return 0;
 506
 507         lz_copy(&window[window_pos], match_len, match_offset,
 508                 &window[window_pos + bytes_remaining]);
 509
 510         return match_len;
 511 }
 512
 513 /*
 514  * Decompress an LZX-compressed block of data.
 515  *
 516  * @block_type:
 517  *      The type of the block (LZX_BLOCKTYPE_VERBATIM or LZX_BLOCKTYPE_ALIGNED).
 518  *
 519  * @block_size:
 520  *      The size of the block, in bytes.
 521  *
 522  * @window:
 523  *      Pointer to the beginning of the decompression window.
 524  *
 525  * @window_pos:
 526  *      The position in the window at which the block starts.
 527  *
 528  * @tables:
 529  *      The Huffman decoding tables for the block.
 530  *
 531  * @queue:
 532  *      The least-recently-used queue for match offsets.
 533  *
 534  * @istream:
 535  *      The input bitstream, positioned at the start of the block data.
 536  *
 537  * Returns 0 on success, or -1 if the data was invalid.
 538  */
 539 static int
 540 lzx_decompress_block(int block_type, u32 block_size,
 541                      u8 *window, u32 window_pos,
 542                      const struct lzx_tables *tables,
 543                      struct lzx_lru_queue *queue,
 544                      struct input_bitstream *istream)
 545 {
 546         u32 block_end;
 547         unsigned main_symbol;
 548         u32 match_len;
 549
 550         block_end = window_pos + block_size;
 551         while (window_pos < block_end) {
 552                 main_symbol = read_huffsym_using_maincode(istream, tables);
 553                 if (main_symbol < LZX_NUM_CHARS) {
 554                         /* Literal  */
 555                         window[window_pos++] = main_symbol;
 556                 } else {
 557                         /* Match  */
 558                         match_len = lzx_decode_match(main_symbol,
 559                                                      block_type,
 560                                                      block_end - window_pos,
 561                                                      window,
 562                                                      window_pos,
 563                                                      tables,
 564                                                      queue,
 565                                                      istream);
 566                         if (unlikely(match_len == 0))
 567                                 return -1;
 568                         window_pos += match_len;
 569                 }
 570         }
 571         return 0;
 572 }
 573
 574 static int
 575 lzx_decompress(const void *compressed_data, size_t compressed_size,
 576                void *uncompressed_data, size_t uncompressed_size,
 577                void *_dec)
 578 {
 579         struct lzx_decompressor *dec = _dec;
 580         struct input_bitstream istream;
 581         struct lzx_lru_queue queue;
 582         u32 window_pos;
 583         int block_type;
 584         u32 block_size;
 585         bool may_have_e8_byte;
 586         int ret;
 587
 588         init_input_bitstream(&istream, compressed_data, compressed_size);
 589
 590         /* Initialize the recent offsets queue.  */
 591         lzx_lru_queue_init(&queue);
 592
 593         /* Codeword lengths begin as all 0's for delta encoding purposes.  */
 594         memset(dec->tables.maincode_lens, 0, sizeof(dec->tables.maincode_lens));
 595         memset(dec->tables.lencode_lens, 0, sizeof(dec->tables.lencode_lens));
 596
 597         /* Set this to true if there may be 0xe8 bytes in the uncompressed data.
 598          */
 599         may_have_e8_byte = false;
 600
 601         /* The compressed data will consist of one or more blocks.  The
 602          * following loop decompresses one block, and it runs until there all
 603          * the compressed data has been decompressed, so there are no more
 604          * blocks.  */
 605
 606         for (window_pos = 0;
 607              window_pos < uncompressed_size;
 608              window_pos += block_size)
 609         {
 610                 ret = lzx_read_block_header(&istream, dec->num_main_syms,
 611                                             dec->window_order, &block_type,
 612                                             &block_size, &dec->tables, &queue);
 613                 if (ret)
 614                         return ret;
 615
 616                 if (block_size > uncompressed_size - window_pos)
 617                         return -1;
 618
 619                 if (block_type != LZX_BLOCKTYPE_UNCOMPRESSED) {
 620
 621                         /* Compressed block.  */
 622
 623                         ret = lzx_decompress_block(block_type,
 624                                                    block_size,
 625                                                    uncompressed_data,
 626                                                    window_pos,
 627                                                    &dec->tables,
 628                                                    &queue,
 629                                                    &istream);
 630                         if (ret)
 631                                 return ret;
 632
 633                         /* If the first 0xe8 byte was in this block, it must
 634                          * have been encoded as a literal using mainsym 0xe8. */
 635                         if (dec->tables.maincode_lens[0xe8] != 0)
 636                                 may_have_e8_byte = true;
 637                 } else {
 638
 639                         /* Uncompressed block.  */
 640                         const u8 *p;
 641
 642                         p = bitstream_read_bytes(&istream, block_size);
 643                         if (!p)
 644                                 return -1;
 645
 646                         memcpy(&((u8*)uncompressed_data)[window_pos], p, block_size);
 647
 648                         /* Re-align the bitstream if an odd number of bytes was
 649                          * read.  */
 650                         if (block_size & 1)
 651                                 bitstream_read_byte(&istream);
 652
 653                         may_have_e8_byte = true;
 654                 }
 655         }
 656
 657         /* Postprocess the data unless it cannot possibly contain 0xe8 bytes  */
 658         if (may_have_e8_byte)
 659                 lzx_undo_e8_preprocessing(uncompressed_data, uncompressed_size);
 660
 661         return 0;
 662 }
 663
 664 static void
 665 lzx_free_decompressor(void *_dec)
 666 {
 667         struct lzx_decompressor *dec = _dec;
 668
 669         ALIGNED_FREE(dec);
 670 }
 671
 672 static int
 673 lzx_create_decompressor(size_t max_block_size, void **dec_ret)
 674 {
 675         struct lzx_decompressor *dec;
 676         unsigned window_order;
 677
 678         window_order = lzx_get_window_order(max_block_size);
 679         if (window_order == 0)
 680                 return WIMLIB_ERR_INVALID_PARAM;
 681
 682         /* The aligned allocation is needed to ensure that the lzx_tables are
 683          * aligned properly.  */
 684         dec = ALIGNED_MALLOC(sizeof(struct lzx_decompressor),
 685                              DECODE_TABLE_ALIGNMENT);
 686         if (!dec)
 687                 return WIMLIB_ERR_NOMEM;
 688
 689         dec->window_order = window_order;
 690         dec->num_main_syms = lzx_get_num_main_syms(window_order);
 691
 692         *dec_ret = dec;
 693         return 0;
 694 }
 695
 696 const struct decompressor_ops lzx_decompressor_ops = {
 697         .create_decompressor = lzx_create_decompressor,
 698         .decompress          = lzx_decompress,
 699         .free_decompressor   = lzx_free_decompressor,
 700 };