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

   1 /*
   2  * lzx-decompress.c
   3  *
   4  * LZX decompression routines, originally based on code taken from cabextract
   5  * v0.5, which was, itself, a modified version of the lzx decompression code
   6  * from unlzx.
   7  */
   8
   9 /*
  10  * Copyright (C) 2012, 2013 Eric Biggers
  11  *
  12  * This file is part of wimlib, a library for working with WIM files.
  13  *
  14  * wimlib is free software; you can redistribute it and/or modify it under the
  15  * terms of the GNU General Public License as published by the Free
  16  * Software Foundation; either version 3 of the License, or (at your option)
  17  * any later version.
  18  *
  19  * wimlib is distributed in the hope that it will be useful, but WITHOUT ANY
  20  * WARRANTY; without even the implied warranty of MERCHANTABILITY or FITNESS FOR
  21  * A PARTICULAR PURPOSE. See the GNU General Public License for more
  22  * details.
  23  *
  24  * You should have received a copy of the GNU General Public License
  25  * along with wimlib; if not, see http://www.gnu.org/licenses/.
  26  */
  27
  28 /*
  29  * LZX is a LZ77 and Huffman-code based compression format that has many
  30  * similarities to the DEFLATE format used in zlib.  The compression ratio is as
  31  * good or better than DEFLATE.  However, in WIM files only up to 32768 bytes of
  32  * data can ever compressed be in the same LZX block, so a .tar.gz file could
  33  * potentially be smaller than a WIM file that uses LZX compression because it
  34  * can use a larger LZ77 window size.
  35  *
  36  * Some notes on the LZX compression format as used in Windows Imaging (WIM)
  37  * files:
  38  *
  39  * A compressed WIM resource consists of a table of chunk offsets followed by
  40  * the compressed chunks themselves.  All compressed chunks except possibly the
  41  * last decompress to WIM_CHUNK_SIZE (= 32768) bytes.  This is quite similar to
  42  * the cabinet (.cab) file format, but they are not the same.  According to the
  43  * cabinet format documentation, the LZX block size is independent from the
  44  * CFDATA blocks, and a LZX block may span several CFDATA blocks.  However, in
  45  * WIMs, LZX blocks do not appear to ever span multiple WIM chunks.  Note that
  46  * this means any WIM chunk may be decompressed or compressed independently from
  47  * any other chunk, which is convenient.
  48  *
  49  * A LZX compressed WIM chunk contains one or more LZX blocks of the aligned,
  50  * verbatim, or uncompressed block types.  For aligned and verbatim blocks, the
  51  * size of the block in uncompressed bytes is specified by a bit following the 3
  52  * bits that specify the block type, possibly followed by an additional 16 bits.
  53  * '1' means to use the default block size (equal to 32768, the size of a WIM
  54  * chunk--- and this seems to only be valid for the first LZX block in a WIM
  55  * chunk), while '0' means that the block size is provided by the next 16 bits.
  56  *
  57  * The cabinet format, as documented, allows for the possibility that a
  58  * compressed CFDATA chunk is up to 6144 bytes larger than the data it
  59  * uncompresses to.  However, in the WIM format it appears that every chunk that
  60  * would be 32768 bytes or more when compressed is actually stored fully
  61  * uncompressed.
  62  *
  63  * The 'e8' preprocessing step that changes x86 call instructions to use
  64  * absolute offsets instead of relative offsets relies on a filesize parameter.
  65  * There is no such parameter for this in the WIM files (even though the size of
  66  * the file resource could be used for this purpose), and instead a magic file
  67  * size of 12000000 is used.  The 'e8' preprocessing is always done, and there
  68  * is no bit to indicate whether it is done or not.
  69  */
  70
  71 /*
  72  * Some more notes about errors in Microsoft's LZX documentation:
  73  *
  74  * Microsoft's LZX document and their implementation of the com.ms.util.cab Java
  75  * package do not concur.
  76  *
  77  * In the LZX document, there is a table showing the correlation between window
  78  * size and the number of position slots. It states that the 1MB window = 40
  79  * slots and the 2MB window = 42 slots. In the implementation, 1MB = 42 slots,
  80  * 2MB = 50 slots. The actual calculation is 'find the first slot whose position
  81  * base is equal to or more than the required window size'. This would explain
  82  * why other tables in the document refer to 50 slots rather than 42.
  83  *
  84  * The constant NUM_PRIMARY_LENS used in the decompression pseudocode is not
  85  * defined in the specification.
  86  *
  87  * The LZX document states that aligned offset blocks have their aligned offset
  88  * huffman tree AFTER the main and length trees. The implementation suggests
  89  * that the aligned offset tree is BEFORE the main and length trees.
  90  *
  91  * The LZX document decoding algorithm states that, in an aligned offset block,
  92  * if an extra_bits value is 1, 2 or 3, then that number of bits should be read
  93  * and the result added to the match offset. This is correct for 1 and 2, but
  94  * not 3, where just a huffman symbol (using the aligned tree) should be read.
  95  *
  96  * Regarding the E8 preprocessing, the LZX document states 'No translation may
  97  * be performed on the last 6 bytes of the input block'. This is correct.
  98  * However, the pseudocode provided checks for the *E8 leader* up to the last 6
  99  * bytes. If the leader appears between -10 and -7 bytes from the end, this
 100  * would cause the next four bytes to be modified, at least one of which would
 101  * be in the last 6 bytes, which is not allowed according to the spec.
 102  *
 103  * The specification states that the huffman trees must always contain at least
 104  * one element. However, many CAB files contain blocks where the length tree is
 105  * completely empty (because there are no matches), and this is expected to
 106  * succeed.
 107  */
 108
 109 #ifdef HAVE_CONFIG_H
 110 #  include "config.h"
 111 #endif
 112
 113 #include "wimlib.h"
 114 #include "wimlib/decompress.h"
 115 #include "wimlib/lzx.h"
 116 #include "wimlib/util.h"
 117
 118 #include <string.h>
 119
 120 /* Huffman decoding tables and maps from symbols to code lengths. */
 121 struct lzx_tables {
 122
 123         u16 maintree_decode_table[(1 << LZX_MAINTREE_TABLEBITS) +
 124                                         (LZX_MAINTREE_NUM_SYMBOLS * 2)]
 125                                         _aligned_attribute(DECODE_TABLE_ALIGNMENT);
 126         u8 maintree_lens[LZX_MAINTREE_NUM_SYMBOLS];
 127
 128
 129         u16 lentree_decode_table[(1 << LZX_LENTREE_TABLEBITS) +
 130                                         (LZX_LENTREE_NUM_SYMBOLS * 2)]
 131                                         _aligned_attribute(DECODE_TABLE_ALIGNMENT);
 132         u8 lentree_lens[LZX_LENTREE_NUM_SYMBOLS];
 133
 134
 135         u16 alignedtree_decode_table[(1 << LZX_ALIGNEDTREE_TABLEBITS) +
 136                                         (LZX_ALIGNEDTREE_NUM_SYMBOLS * 2)]
 137                                         _aligned_attribute(DECODE_TABLE_ALIGNMENT);
 138         u8 alignedtree_lens[LZX_ALIGNEDTREE_NUM_SYMBOLS];
 139 } _aligned_attribute(DECODE_TABLE_ALIGNMENT);
 140
 141
 142 /*
 143  * Reads a Huffman-encoded symbol using the pre-tree.
 144  */
 145 static inline int
 146 read_huffsym_using_pretree(struct input_bitstream *istream,
 147                            const u16 pretree_decode_table[],
 148                            const u8 pretree_lens[], unsigned *n)
 149 {
 150         return read_huffsym(istream, pretree_decode_table, pretree_lens,
 151                             LZX_PRETREE_NUM_SYMBOLS, LZX_PRETREE_TABLEBITS, n,
 152                             LZX_MAX_CODEWORD_LEN);
 153 }
 154
 155 /* Reads a Huffman-encoded symbol using the main tree. */
 156 static inline int
 157 read_huffsym_using_maintree(struct input_bitstream *istream,
 158                             const struct lzx_tables *tables,
 159                             unsigned *n)
 160 {
 161         return read_huffsym(istream, tables->maintree_decode_table,
 162                             tables->maintree_lens, LZX_MAINTREE_NUM_SYMBOLS,
 163                             LZX_MAINTREE_TABLEBITS, n, LZX_MAX_CODEWORD_LEN);
 164 }
 165
 166 /* Reads a Huffman-encoded symbol using the length tree. */
 167 static inline int
 168 read_huffsym_using_lentree(struct input_bitstream *istream,
 169                            const struct lzx_tables *tables,
 170                            unsigned *n)
 171 {
 172         return read_huffsym(istream, tables->lentree_decode_table,
 173                             tables->lentree_lens, LZX_LENTREE_NUM_SYMBOLS,
 174                             LZX_LENTREE_TABLEBITS, n, LZX_MAX_CODEWORD_LEN);
 175 }
 176
 177 /* Reads a Huffman-encoded symbol using the aligned offset tree. */
 178 static inline int
 179 read_huffsym_using_alignedtree(struct input_bitstream *istream,
 180                                const struct lzx_tables *tables,
 181                                unsigned *n)
 182 {
 183         return read_huffsym(istream, tables->alignedtree_decode_table,
 184                             tables->alignedtree_lens,
 185                             LZX_ALIGNEDTREE_NUM_SYMBOLS,
 186                             LZX_ALIGNEDTREE_TABLEBITS, n, 8);
 187 }
 188
 189 /*
 190  * Reads the pretree from the input, then uses the pretree to decode @num_lens
 191  * code length values from the input.
 192  *
 193  * @istream:    The bit stream for the input.  It is positioned on the beginning
 194  *                      of the pretree for the code length values.
 195  * @lens:       An array that contains the length values from the previous time
 196  *                      the code lengths for this Huffman tree were read, or all
 197  *                      0's if this is the first time.
 198  * @num_lens:   Number of length values to decode and return.
 199  *
 200  */
 201 static int
 202 lzx_read_code_lens(struct input_bitstream *istream, u8 lens[],
 203                    unsigned num_lens)
 204 {
 205         /* Declare the decoding table and length table for the pretree. */
 206         u16 pretree_decode_table[(1 << LZX_PRETREE_TABLEBITS) +
 207                                         (LZX_PRETREE_NUM_SYMBOLS * 2)]
 208                                         _aligned_attribute(DECODE_TABLE_ALIGNMENT);
 209         u8 pretree_lens[LZX_PRETREE_NUM_SYMBOLS];
 210         unsigned i;
 211         unsigned len;
 212         int ret;
 213
 214         /* Read the code lengths of the pretree codes.  There are 20 lengths of
 215          * 4 bits each. */
 216         for (i = 0; i < LZX_PRETREE_NUM_SYMBOLS; i++) {
 217                 ret = bitstream_read_bits(istream, LZX_PRETREE_ELEMENT_SIZE,
 218                                           &len);
 219                 if (ret != 0)
 220                         return ret;
 221                 pretree_lens[i] = len;
 222         }
 223
 224         /* Make the decoding table for the pretree. */
 225         ret = make_huffman_decode_table(pretree_decode_table,
 226                                         LZX_PRETREE_NUM_SYMBOLS,
 227                                         LZX_PRETREE_TABLEBITS,
 228                                         pretree_lens,
 229                                         LZX_MAX_CODEWORD_LEN);
 230         if (ret != 0)
 231                 return ret;
 232
 233         /* Pointer past the last length value that needs to be filled in. */
 234         u8 *lens_end = lens + num_lens;
 235
 236         while (1) {
 237
 238                 /* Decode a symbol from the input.  If the symbol is between 0
 239                  * and 16, it is the difference from the old length.  If it is
 240                  * between 17 and 19, it is a special code that indicates that
 241                  * some number of the next lengths are all 0, or some number of
 242                  * the next lengths are all equal to the next symbol in the
 243                  * input. */
 244                 unsigned tree_code;
 245                 unsigned num_zeroes;
 246                 unsigned code;
 247                 unsigned num_same;
 248                 signed char value;
 249
 250                 ret = read_huffsym_using_pretree(istream, pretree_decode_table,
 251                                                  pretree_lens, &tree_code);
 252                 if (ret != 0)
 253                         return ret;
 254                 switch (tree_code) {
 255                 case 17: /* Run of 0's */
 256                         ret = bitstream_read_bits(istream, 4, &num_zeroes);
 257                         if (ret != 0)
 258                                 return ret;
 259                         num_zeroes += 4;
 260                         while (num_zeroes--) {
 261                                 *lens = 0;
 262                                 if (++lens == lens_end)
 263                                         return 0;
 264                         }
 265                         break;
 266                 case 18: /* Longer run of 0's */
 267                         ret = bitstream_read_bits(istream, 5, &num_zeroes);
 268                         if (ret != 0)
 269                                 return ret;
 270                         num_zeroes += 20;
 271                         while (num_zeroes--) {
 272                                 *lens = 0;
 273                                 if (++lens == lens_end)
 274                                         return 0;
 275                         }
 276                         break;
 277                 case 19: /* Run of identical lengths */
 278                         ret = bitstream_read_bits(istream, 1, &num_same);
 279                         if (ret != 0)
 280                                 return ret;
 281                         num_same += 4;
 282                         ret = read_huffsym_using_pretree(istream,
 283                                                          pretree_decode_table,
 284                                                          pretree_lens,
 285                                                          &code);
 286                         if (ret != 0)
 287                                 return ret;
 288                         value = (signed char)*lens - (signed char)code;
 289                         if (value < 0)
 290                                 value += 17;
 291                         while (num_same--) {
 292                                 *lens = value;
 293                                 if (++lens == lens_end)
 294                                         return 0;
 295                         }
 296                         break;
 297                 default: /* Difference from old length. */
 298                         value = (signed char)*lens - (signed char)tree_code;
 299                         if (value < 0)
 300                                 value += 17;
 301                         *lens = value;
 302                         if (++lens == lens_end)
 303                                 return 0;
 304                         break;
 305                 }
 306         }
 307 }
 308
 309 /*
 310  * Reads the header for an LZX-compressed block.
 311  *
 312  * @istream:            The input bitstream.
 313  * @block_size_ret:     A pointer to an int into which the size of the block,
 314  *                              in bytes, will be returned.
 315  * @block_type_ret:     A pointer to an int into which the type of the block
 316  *                              (LZX_BLOCKTYPE_*) will be returned.
 317  * @tables:             A pointer to a lzx_tables structure in which the
 318  *                              main tree, the length tree, and possibly the
 319  *                              aligned offset tree will be constructed.
 320  * @queue:      A pointer to the least-recently-used queue into which
 321  *                      R0, R1, and R2 will be written (only for uncompressed
 322  *                      blocks, which contain this information in the header)
 323  */
 324 static int
 325 lzx_read_block_header(struct input_bitstream *istream,
 326                       unsigned *block_size_ret,
 327                       unsigned *block_type_ret,
 328                       struct lzx_tables *tables,
 329                       struct lru_queue *queue)
 330 {
 331         int ret;
 332         unsigned block_type;
 333         unsigned block_size;
 334         unsigned s;
 335         unsigned i;
 336         unsigned len;
 337
 338         ret = bitstream_ensure_bits(istream, 4);
 339         if (ret) {
 340                 DEBUG("LZX input stream overrun");
 341                 return ret;
 342         }
 343
 344         /* The first three bits tell us what kind of block it is, and are one
 345          * of the LZX_BLOCKTYPE_* values.  */
 346         block_type = bitstream_read_bits_nocheck(istream, 3);
 347
 348         /* The next bit indicates whether the block size is the default (32768),
 349          * indicated by a 1 bit, or whether the block size is given by the next
 350          * 16 bits, indicated by a 0 bit. */
 351         s = bitstream_read_bits_nocheck(istream, 1);
 352
 353         if (s) {
 354                 block_size = 32768;
 355         } else {
 356                 ret = bitstream_read_bits(istream, 16, &block_size);
 357                 if (ret)
 358                         return ret;
 359                 block_size = le16_to_cpu(block_size);
 360         }
 361
 362         switch (block_type) {
 363         case LZX_BLOCKTYPE_ALIGNED:
 364                 /* Read the path lengths for the elements of the aligned tree,
 365                  * then build it. */
 366
 367                 for (i = 0; i < LZX_ALIGNEDTREE_NUM_SYMBOLS; i++) {
 368                         ret = bitstream_read_bits(istream,
 369                                                   LZX_ALIGNEDTREE_ELEMENT_SIZE,
 370                                                   &len);
 371                         if (ret)
 372                                 return ret;
 373                         tables->alignedtree_lens[i] = len;
 374                 }
 375
 376                 LZX_DEBUG("Building the aligned tree.");
 377                 ret = make_huffman_decode_table(tables->alignedtree_decode_table,
 378                                                 LZX_ALIGNEDTREE_NUM_SYMBOLS,
 379                                                 LZX_ALIGNEDTREE_TABLEBITS,
 380                                                 tables->alignedtree_lens,
 381                                                 8);
 382                 if (ret) {
 383                         DEBUG("lzx_decompress(): Failed to make the decode "
 384                               "table for the aligned offset tree");
 385                         return ret;
 386                 }
 387
 388                 /* Fall though, since the rest of the header for aligned offset
 389                  * blocks is the same as that for verbatim blocks */
 390
 391         case LZX_BLOCKTYPE_VERBATIM:
 392                 if (block_type == LZX_BLOCKTYPE_VERBATIM)
 393                         LZX_DEBUG("Found verbatim block.");
 394
 395                 LZX_DEBUG("Reading path lengths for main tree.");
 396                 /* Read the path lengths for the first 256 elements of the main
 397                  * tree. */
 398                 ret = lzx_read_code_lens(istream, tables->maintree_lens,
 399                                          LZX_NUM_CHARS);
 400                 if (ret) {
 401                         DEBUG("lzx_decompress(): Failed to read the code "
 402                               "lengths for the first 256 elements of the "
 403                               "main tree");
 404                         return ret;
 405                 }
 406
 407                 /* Read the path lengths for the remaining elements of the main
 408                  * tree. */
 409                 LZX_DEBUG("Reading path lengths for remaining elements of "
 410                           "main tree (%d elements).",
 411                           LZX_MAINTREE_NUM_SYMBOLS - LZX_NUM_CHARS);
 412                 ret = lzx_read_code_lens(istream,
 413                                          tables->maintree_lens + LZX_NUM_CHARS,
 414                                          LZX_MAINTREE_NUM_SYMBOLS - LZX_NUM_CHARS);
 415                 if (ret) {
 416                         DEBUG("lzx_decompress(): Failed to read the path "
 417                               "lengths for the remaining elements of the main "
 418                               "tree");
 419                         return ret;
 420                 }
 421
 422                 LZX_DEBUG("Building the Huffman decoding "
 423                           "table for the main tree.");
 424
 425                 ret = make_huffman_decode_table(tables->maintree_decode_table,
 426                                                 LZX_MAINTREE_NUM_SYMBOLS,
 427                                                 LZX_MAINTREE_TABLEBITS,
 428                                                 tables->maintree_lens,
 429                                                 LZX_MAX_CODEWORD_LEN);
 430                 if (ret) {
 431                         DEBUG("lzx_decompress(): Failed to make the decode "
 432                               "table for the main tree");
 433                         return ret;
 434                 }
 435
 436                 LZX_DEBUG("Reading path lengths for the length tree.");
 437                 ret = lzx_read_code_lens(istream, tables->lentree_lens,
 438                                          LZX_LENTREE_NUM_SYMBOLS);
 439                 if (ret) {
 440                         DEBUG("lzx_decompress(): Failed to read the path "
 441                               "lengths for the length tree");
 442                         return ret;
 443                 }
 444
 445                 LZX_DEBUG("Building the length tree.");
 446                 ret = make_huffman_decode_table(tables->lentree_decode_table,
 447                                                 LZX_LENTREE_NUM_SYMBOLS,
 448                                                 LZX_LENTREE_TABLEBITS,
 449                                                 tables->lentree_lens,
 450                                                 LZX_MAX_CODEWORD_LEN);
 451                 if (ret) {
 452                         DEBUG("lzx_decompress(): Failed to build the length "
 453                               "Huffman tree");
 454                         return ret;
 455                 }
 456                 /* The bitstream of compressed literals and matches for this
 457                  * block directly follows and will be read in
 458                  * lzx_decompress_block(). */
 459                 break;
 460         case LZX_BLOCKTYPE_UNCOMPRESSED:
 461                 LZX_DEBUG("Found uncompressed block.");
 462                 /* Before reading the three LRU match offsets from the
 463                  * uncompressed block header, the stream needs to be aligned on
 464                  * a 16-bit boundary.  But, unexpectedly, if the stream is
 465                  * *already* aligned, the correct thing to do is to throw away
 466                  * the next 16 bits. */
 467                 if (istream->bitsleft == 0) {
 468                         if (istream->data_bytes_left < 14) {
 469                                 DEBUG("lzx_decompress(): Insufficient length in "
 470                                       "uncompressed block");
 471                                 return -1;
 472                         }
 473                         istream->data += 2;
 474                         istream->data_bytes_left -= 2;
 475                 } else {
 476                         if (istream->data_bytes_left < 12) {
 477                                 DEBUG("lzx_decompress(): Insufficient length in "
 478                                       "uncompressed block");
 479                                 return -1;
 480                         }
 481                         istream->bitsleft = 0;
 482                         istream->bitbuf = 0;
 483                 }
 484                 queue->R0 = le32_to_cpu(*(u32*)(istream->data + 0));
 485                 queue->R1 = le32_to_cpu(*(u32*)(istream->data + 4));
 486                 queue->R2 = le32_to_cpu(*(u32*)(istream->data + 8));
 487                 istream->data += 12;
 488                 istream->data_bytes_left -= 12;
 489                 /* The uncompressed data of this block directly follows and will
 490                  * be read in lzx_decompress(). */
 491                 break;
 492         default:
 493                 DEBUG("lzx_decompress(): Found invalid block");
 494                 return -1;
 495         }
 496         *block_type_ret = block_type;
 497         *block_size_ret = block_size;
 498         return 0;
 499 }
 500
 501 /*
 502  * Decodes a compressed match from a block of LZX-compressed data.  A match
 503  * refers to some match_offset to a point earlier in the window as well as some
 504  * match_len, for which the data is to be copied to the current position in the
 505  * window.
 506  *
 507  * @main_element:       The start of the match data, as decoded using the main
 508  *                      tree.
 509  *
 510  * @block_type:         The type of the block (LZX_BLOCKTYPE_ALIGNED or
 511  *                      LZX_BLOCKTYPE_VERBATIM)
 512  *
 513  * @bytes_remaining:    The amount of uncompressed data remaining to be
 514  *                      uncompressed in this block.  It is an error if the match
 515  *                      is longer than this number.
 516  *
 517  * @window:             A pointer to the window into which the uncompressed
 518  *                      data is being written.
 519  *
 520  * @window_pos:         The current byte offset in the window.
 521  *
 522  * @tables:             The Huffman decoding tables for this LZX block (main
 523  *                      code, length code, and for LZX_BLOCKTYPE_ALIGNED blocks,
 524  *                      also the aligned offset code).
 525  *
 526  * @queue:              The least-recently used queue for match offsets.
 527  *
 528  * @istream:            The input bitstream.
 529  *
 530  * Returns the length of the match, or a negative number on error.  The possible
 531  * error cases are:
 532  *      - Match would exceed the amount of data remaining to be uncompressed.
 533  *      - Match refers to data before the window.
 534  *      - The input bitstream ended unexpectedly.
 535  */
 536 static int
 537 lzx_decode_match(unsigned main_element, int block_type,
 538                  unsigned bytes_remaining, u8 *window,
 539                  unsigned window_pos,
 540                  const struct lzx_tables *tables,
 541                  struct lru_queue *queue,
 542                  struct input_bitstream *istream)
 543 {
 544         unsigned length_header;
 545         unsigned position_slot;
 546         unsigned match_len;
 547         unsigned match_offset;
 548         unsigned additional_len;
 549         unsigned num_extra_bits;
 550         unsigned verbatim_bits;
 551         unsigned aligned_bits;
 552         unsigned i;
 553         int ret;
 554         u8 *match_dest;
 555         u8 *match_src;
 556
 557         /* The main element is offset by 256 because values under 256 indicate a
 558          * literal value. */
 559         main_element -= LZX_NUM_CHARS;
 560
 561         /* The length header consists of the lower 3 bits of the main element.
 562          * The position slot is the rest of it. */
 563         length_header = main_element & LZX_NUM_PRIMARY_LENS;
 564         position_slot = main_element >> 3;
 565
 566         /* If the length_header is less than LZX_NUM_PRIMARY_LENS (= 7), it
 567          * gives the match length as the offset from LZX_MIN_MATCH.  Otherwise,
 568          * the length is given by an additional symbol encoded using the length
 569          * tree, offset by 9 (LZX_MIN_MATCH + LZX_NUM_PRIMARY_LENS) */
 570         match_len = LZX_MIN_MATCH + length_header;
 571         if (length_header == LZX_NUM_PRIMARY_LENS) {
 572                 ret = read_huffsym_using_lentree(istream, tables,
 573                                                  &additional_len);
 574                 if (ret != 0)
 575                         return ret;
 576                 match_len += additional_len;
 577         }
 578
 579
 580         /* If the position_slot is 0, 1, or 2, the match offset is retrieved
 581          * from the LRU queue.  Otherwise, the match offset is not in the LRU
 582          * queue. */
 583         switch (position_slot) {
 584         case 0:
 585                 match_offset = queue->R0;
 586                 break;
 587         case 1:
 588                 match_offset = queue->R1;
 589                 swap(queue->R0, queue->R1);
 590                 break;
 591         case 2:
 592                 /* The queue doesn't work quite the same as a real LRU queue,
 593                  * since using the R2 offset doesn't bump the R1 offset down to
 594                  * R2. */
 595                 match_offset = queue->R2;
 596                 swap(queue->R0, queue->R2);
 597                 break;
 598         default:
 599                 /* Otherwise, the offset was not encoded as one the offsets in
 600                  * the queue.  Depending on the position slot, there is a
 601                  * certain number of extra bits that need to be read to fully
 602                  * decode the match offset. */
 603
 604                 /* Look up the number of extra bits that need to be read. */
 605                 num_extra_bits = lzx_get_num_extra_bits(position_slot);
 606
 607                 /* For aligned blocks, if there are at least 3 extra bits, the
 608                  * actual number of extra bits is 3 less, and they encode a
 609                  * number of 8-byte words that are added to the offset; there
 610                  * is then an additional symbol read using the aligned tree that
 611                  * specifies the actual byte alignment. */
 612                 if (block_type == LZX_BLOCKTYPE_ALIGNED && num_extra_bits >= 3) {
 613
 614                         /* There is an error in the LZX "specification" at this
 615                          * point; it indicates that a Huffman symbol is to be
 616                          * read only if num_extra_bits is greater than 3, but
 617                          * actually it is if num_extra_bits is greater than or
 618                          * equal to 3.  (Note that in the case with
 619                          * num_extra_bits == 3, the assignment to verbatim_bits
 620                          * will just set it to 0. ) */
 621                         ret = bitstream_read_bits(istream, num_extra_bits - 3,
 622                                                   &verbatim_bits);
 623                         if (ret != 0)
 624                                 return ret;
 625
 626                         verbatim_bits <<= 3;
 627
 628                         ret = read_huffsym_using_alignedtree(istream, tables,
 629                                                              &aligned_bits);
 630                         if (ret != 0)
 631                                 return ret;
 632                 } else {
 633                         /* For non-aligned blocks, or for aligned blocks with
 634                          * less than 3 extra bits, the extra bits are added
 635                          * directly to the match offset, and the correction for
 636                          * the alignment is taken to be 0. */
 637                         ret = bitstream_read_bits(istream, num_extra_bits,
 638                                                   &verbatim_bits);
 639                         if (ret != 0)
 640                                 return ret;
 641
 642                         aligned_bits = 0;
 643                 }
 644
 645                 /* Calculate the match offset. */
 646                 match_offset = lzx_position_base[position_slot] +
 647                                verbatim_bits + aligned_bits - 2;
 648
 649                 /* Update the LRU queue. */
 650                 queue->R2 = queue->R1;
 651                 queue->R1 = queue->R0;
 652                 queue->R0 = match_offset;
 653                 break;
 654         }
 655
 656         /* Verify that the match is in the bounds of the part of the window
 657          * currently in use, then copy the source of the match to the current
 658          * position. */
 659
 660         if (match_len > bytes_remaining) {
 661                 DEBUG("lzx_decode_match(): Match of length %u bytes overflows "
 662                       "uncompressed block size", match_len);
 663                 return -1;
 664         }
 665
 666         if (match_offset > window_pos) {
 667                 DEBUG("lzx_decode_match(): Match of length %u bytes references "
 668                       "data before window (match_offset = %u, window_pos = %u)",
 669                       match_len, match_offset, window_pos);
 670                 return -1;
 671         }
 672
 673         match_dest = window + window_pos;
 674         match_src = match_dest - match_offset;
 675
 676 #if 0
 677         printf("Match: src %u, dst %u, len %u\n", match_src - window,
 678                                                 match_dest - window,
 679                                                 match_len);
 680         putchar('|');
 681         for (i = 0; i < match_len; i++) {
 682                 match_dest[i] = match_src[i];
 683                 putchar(match_src[i]);
 684         }
 685         putchar('|');
 686         putchar('\n');
 687 #else
 688         for (i = 0; i < match_len; i++)
 689                 match_dest[i] = match_src[i];
 690 #endif
 691
 692         return match_len;
 693 }
 694
 695 static void
 696 undo_call_insn_translation(u32 *call_insn_target, int input_pos,
 697                            s32 file_size)
 698 {
 699         s32 abs_offset;
 700         s32 rel_offset;
 701
 702         abs_offset = le32_to_cpu(*call_insn_target);
 703         if (abs_offset >= -input_pos && abs_offset < file_size) {
 704                 if (abs_offset >= 0) {
 705                         /* "good translation" */
 706                         rel_offset = abs_offset - input_pos;
 707                 } else {
 708                         /* "compensating translation" */
 709                         rel_offset = abs_offset + file_size;
 710                 }
 711                 *call_insn_target = cpu_to_le32(rel_offset);
 712         }
 713 }
 714
 715 /* Undo the 'E8' preprocessing, where the targets of x86 CALL instructions were
 716  * changed from relative offsets to absolute offsets.
 717  *
 718  * Note that this call instruction preprocessing can and will be used on any
 719  * data even if it is not actually x86 machine code.  In fact, this type of
 720  * preprocessing appears to always be used in LZX-compressed resources in WIM
 721  * files; there is no bit to indicate whether it is used or not, unlike in the
 722  * LZX compressed format as used in cabinet files, where a bit is reserved for
 723  * that purpose.
 724  *
 725  * Call instruction preprocessing is disabled in the last 6 bytes of the
 726  * uncompressed data, which really means the 5-byte call instruction cannot
 727  * start in the last 10 bytes of the uncompressed data.  This is one of the
 728  * errors in the LZX documentation.
 729  *
 730  * Call instruction preprocessing does not appear to be disabled after the
 731  * 32768th chunk of a WIM stream, which is apparently is yet another difference
 732  * from the LZX compression used in cabinet files.
 733  *
 734  * Call instruction processing is supposed to take the file size as a parameter,
 735  * as it is used in calculating the translated jump targets.  But in WIM files,
 736  * this file size is always the same (LZX_WIM_MAGIC_FILESIZE == 12000000).*/
 737 static void
 738 undo_call_insn_preprocessing(u8 uncompressed_data[], int uncompressed_data_len)
 739 {
 740         for (int i = 0; i < uncompressed_data_len - 10; i++) {
 741                 if (uncompressed_data[i] == 0xe8) {
 742                         undo_call_insn_translation((u32*)&uncompressed_data[i + 1],
 743                                                    i,
 744                                                    LZX_WIM_MAGIC_FILESIZE);
 745                         i += 4;
 746                 }
 747         }
 748 }
 749
 750 /*
 751  * Decompresses a LZX-compressed block of data from which the header has already
 752  * been read.
 753  *
 754  * @block_type: The type of the block (LZX_BLOCKTYPE_VERBATIM or
 755  *              LZX_BLOCKTYPE_ALIGNED)
 756  * @block_size: The size of the block, in bytes.
 757  * @window:     Pointer to the decompression window.
 758  * @window_pos: The current position in the window.  Will be 0 for the first
 759  *                      block.
 760  * @tables:     The Huffman decoding tables for the block (main, length, and
 761  *                      aligned offset, the latter only for LZX_BLOCKTYPE_ALIGNED)
 762  * @queue:      The least-recently-used queue for match offsets.
 763  * @istream:    The input bitstream for the compressed literals.
 764  */
 765 static int
 766 lzx_decompress_block(int block_type, unsigned block_size,
 767                      u8 *window,
 768                      unsigned window_pos,
 769                      const struct lzx_tables *tables,
 770                      struct lru_queue *queue,
 771                      struct input_bitstream *istream)
 772 {
 773         unsigned main_element;
 774         unsigned end;
 775         int ret;
 776         int match_len;
 777
 778         end = window_pos + block_size;
 779         while (window_pos < end) {
 780                 ret = read_huffsym_using_maintree(istream, tables,
 781                                                   &main_element);
 782                 if (ret)
 783                         return ret;
 784
 785                 if (main_element < LZX_NUM_CHARS) {
 786                         /* literal: 0 to LZX_NUM_CHARS - 1 */
 787                         window[window_pos++] = main_element;
 788                 } else {
 789                         /* match: LZX_NUM_CHARS to LZX_MAINTREE_NUM_SYMBOLS - 1 */
 790                         match_len = lzx_decode_match(main_element,
 791                                                      block_type,
 792                                                      end - window_pos,
 793                                                      window,
 794                                                      window_pos,
 795                                                      tables,
 796                                                      queue,
 797                                                      istream);
 798                         if (match_len < 0)
 799                                 return match_len;
 800                         window_pos += match_len;
 801                 }
 802         }
 803         return 0;
 804 }
 805
 806 /* API function documented in wimlib.h  */
 807 WIMLIBAPI int
 808 wimlib_lzx_decompress(const void *compressed_data, unsigned compressed_len,
 809                       void *uncompressed_data, unsigned uncompressed_len)
 810 {
 811         struct lzx_tables tables;
 812         struct input_bitstream istream;
 813         struct lru_queue queue;
 814         unsigned window_pos;
 815         unsigned block_size;
 816         unsigned block_type;
 817         int ret;
 818         bool e8_preprocessing_done;
 819
 820         LZX_DEBUG("lzx_decompress (compressed_data = %p, compressed_len = %d, "
 821                   "uncompressed_data = %p, uncompressed_len = %d).",
 822                   compressed_data, compressed_len,
 823                   uncompressed_data, uncompressed_len);
 824
 825         wimlib_assert(uncompressed_len <= 32768);
 826
 827         memset(tables.maintree_lens, 0, sizeof(tables.maintree_lens));
 828         memset(tables.lentree_lens, 0, sizeof(tables.lentree_lens));
 829         queue.R0 = 1;
 830         queue.R1 = 1;
 831         queue.R2 = 1;
 832         init_input_bitstream(&istream, compressed_data, compressed_len);
 833
 834         e8_preprocessing_done = false; /* Set to true if there may be 0xe8 bytes
 835                                           in the uncompressed data. */
 836
 837         /* The compressed data will consist of one or more blocks.  The
 838          * following loop decompresses one block, and it runs until there all
 839          * the compressed data has been decompressed, so there are no more
 840          * blocks.  */
 841
 842         for (window_pos = 0;
 843              window_pos < uncompressed_len;
 844              window_pos += block_size)
 845         {
 846                 LZX_DEBUG("Reading block header.");
 847                 ret = lzx_read_block_header(&istream, &block_size,
 848                                             &block_type, &tables, &queue);
 849                 if (ret)
 850                         return ret;
 851
 852                 LZX_DEBUG("block_size = %u, window_pos = %u",
 853                           block_size, window_pos);
 854
 855                 if (block_size > uncompressed_len - window_pos) {
 856                         DEBUG("lzx_decompress(): Expected a block size of at "
 857                               "most %u bytes (found %u bytes)",
 858                               uncompressed_len - window_pos, block_size);
 859                         return -1;
 860                 }
 861
 862                 switch (block_type) {
 863                 case LZX_BLOCKTYPE_VERBATIM:
 864                 case LZX_BLOCKTYPE_ALIGNED:
 865                         if (block_type == LZX_BLOCKTYPE_VERBATIM)
 866                                 LZX_DEBUG("LZX_BLOCKTYPE_VERBATIM");
 867                         else
 868                                 LZX_DEBUG("LZX_BLOCKTYPE_ALIGNED");
 869                         ret = lzx_decompress_block(block_type,
 870                                                    block_size,
 871                                                    uncompressed_data,
 872                                                    window_pos,
 873                                                    &tables,
 874                                                    &queue,
 875                                                    &istream);
 876                         if (ret)
 877                                 return ret;
 878                         if (tables.maintree_lens[0xe8] != 0)
 879                                 e8_preprocessing_done = true;
 880                         break;
 881                 case LZX_BLOCKTYPE_UNCOMPRESSED:
 882                         LZX_DEBUG("LZX_BLOCKTYPE_UNCOMPRESSED");
 883                         if (istream.data_bytes_left < block_size) {
 884                                 DEBUG("Unexpected end of input when "
 885                                       "reading %u bytes from LZX bitstream "
 886                                       "(only have %u bytes left)",
 887                                       block_size, istream.data_bytes_left);
 888                                 return -1;
 889                         }
 890                         memcpy(&((u8*)uncompressed_data)[window_pos], istream.data,
 891                                block_size);
 892                         istream.data += block_size;
 893                         istream.data_bytes_left -= block_size;
 894                         /* Re-align bitstream if an odd number of bytes were
 895                          * read.  */
 896                         if (istream.data_bytes_left && (block_size & 1)) {
 897                                 istream.data_bytes_left--;
 898                                 istream.data++;
 899                         }
 900                         e8_preprocessing_done = true;
 901                         break;
 902                 }
 903         }
 904         if (e8_preprocessing_done)
 905                 undo_call_insn_preprocessing(uncompressed_data, uncompressed_len);
 906         return 0;
 907 }