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 32768 bytes.  This is quite similar to the cabinet (.cab)
  42  * file format, but they are not the same.  According to the cabinet format
  43  * documentation, the LZX block size is independent from the CFDATA blocks, and
  44  * a LZX block may span several CFDATA blocks.  However, in WIMs, LZX blocks do
  45  * not appear to ever span multiple WIM chunks.  Note that this means any WIM
  46  * chunk may be decompressed or compressed independently from any other chunk,
  47  * 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_MAINCODE_TABLEBITS) +
 124                                         (LZX_MAINCODE_MAX_NUM_SYMBOLS * 2)]
 125                                         _aligned_attribute(DECODE_TABLE_ALIGNMENT);
 126         u8 maintree_lens[LZX_MAINCODE_MAX_NUM_SYMBOLS];
 127
 128
 129         u16 lentree_decode_table[(1 << LZX_LENCODE_TABLEBITS) +
 130                                         (LZX_LENCODE_NUM_SYMBOLS * 2)]
 131                                         _aligned_attribute(DECODE_TABLE_ALIGNMENT);
 132         u8 lentree_lens[LZX_LENCODE_NUM_SYMBOLS];
 133
 134
 135         u16 alignedtree_decode_table[(1 << LZX_ALIGNEDCODE_TABLEBITS) +
 136                                         (LZX_ALIGNEDCODE_NUM_SYMBOLS * 2)]
 137                                         _aligned_attribute(DECODE_TABLE_ALIGNMENT);
 138         u8 alignedtree_lens[LZX_ALIGNEDCODE_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_PRECODE_NUM_SYMBOLS, LZX_PRECODE_TABLEBITS, n,
 152                             LZX_MAX_PRE_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                             unsigned num_main_syms)
 161 {
 162         return read_huffsym(istream, tables->maintree_decode_table,
 163                             tables->maintree_lens, num_main_syms,
 164                             LZX_MAINCODE_TABLEBITS, n, LZX_MAX_MAIN_CODEWORD_LEN);
 165 }
 166
 167 /* Reads a Huffman-encoded symbol using the length tree. */
 168 static inline int
 169 read_huffsym_using_lentree(struct input_bitstream *istream,
 170                            const struct lzx_tables *tables,
 171                            unsigned *n)
 172 {
 173         return read_huffsym(istream, tables->lentree_decode_table,
 174                             tables->lentree_lens, LZX_LENCODE_NUM_SYMBOLS,
 175                             LZX_LENCODE_TABLEBITS, n, LZX_MAX_LEN_CODEWORD_LEN);
 176 }
 177
 178 /* Reads a Huffman-encoded symbol using the aligned offset tree. */
 179 static inline int
 180 read_huffsym_using_alignedtree(struct input_bitstream *istream,
 181                                const struct lzx_tables *tables,
 182                                unsigned *n)
 183 {
 184         return read_huffsym(istream, tables->alignedtree_decode_table,
 185                             tables->alignedtree_lens,
 186                             LZX_ALIGNEDCODE_NUM_SYMBOLS,
 187                             LZX_ALIGNEDCODE_TABLEBITS, n,
 188                             LZX_MAX_ALIGNED_CODEWORD_LEN);
 189 }
 190
 191 /*
 192  * Reads the pretree from the input, then uses the pretree to decode @num_lens
 193  * code length values from the input.
 194  *
 195  * @istream:    The bit stream for the input.  It is positioned on the beginning
 196  *                      of the pretree for the code length values.
 197  * @lens:       An array that contains the length values from the previous time
 198  *                      the code lengths for this Huffman tree were read, or all
 199  *                      0's if this is the first time.
 200  * @num_lens:   Number of length values to decode and return.
 201  *
 202  */
 203 static int
 204 lzx_read_code_lens(struct input_bitstream *istream, u8 lens[],
 205                    unsigned num_lens)
 206 {
 207         /* Declare the decoding table and length table for the pretree. */
 208         u16 pretree_decode_table[(1 << LZX_PRECODE_TABLEBITS) +
 209                                         (LZX_PRECODE_NUM_SYMBOLS * 2)]
 210                                         _aligned_attribute(DECODE_TABLE_ALIGNMENT);
 211         u8 pretree_lens[LZX_PRECODE_NUM_SYMBOLS];
 212         unsigned i;
 213         u32 len;
 214         int ret;
 215
 216         /* Read the code lengths of the pretree codes.  There are 20 lengths of
 217          * 4 bits each. */
 218         for (i = 0; i < LZX_PRECODE_NUM_SYMBOLS; i++) {
 219                 ret = bitstream_read_bits(istream, LZX_PRECODE_ELEMENT_SIZE,
 220                                           &len);
 221                 if (ret)
 222                         return ret;
 223                 pretree_lens[i] = len;
 224         }
 225
 226         /* Make the decoding table for the pretree. */
 227         ret = make_huffman_decode_table(pretree_decode_table,
 228                                         LZX_PRECODE_NUM_SYMBOLS,
 229                                         LZX_PRECODE_TABLEBITS,
 230                                         pretree_lens,
 231                                         LZX_MAX_PRE_CODEWORD_LEN);
 232         if (ret)
 233                 return ret;
 234
 235         /* Pointer past the last length value that needs to be filled in. */
 236         u8 *lens_end = lens + num_lens;
 237
 238         while (1) {
 239
 240                 /* Decode a symbol from the input.  If the symbol is between 0
 241                  * and 16, it is the difference from the old length.  If it is
 242                  * between 17 and 19, it is a special code that indicates that
 243                  * some number of the next lengths are all 0, or some number of
 244                  * the next lengths are all equal to the next symbol in the
 245                  * input. */
 246                 unsigned tree_code;
 247                 u32 num_zeroes;
 248                 unsigned code;
 249                 u32 num_same;
 250                 signed char value;
 251
 252                 ret = read_huffsym_using_pretree(istream, pretree_decode_table,
 253                                                  pretree_lens, &tree_code);
 254                 if (ret)
 255                         return ret;
 256                 switch (tree_code) {
 257                 case 17: /* Run of 0's */
 258                         ret = bitstream_read_bits(istream, 4, &num_zeroes);
 259                         if (ret)
 260                                 return ret;
 261                         num_zeroes += 4;
 262                         while (num_zeroes--) {
 263                                 *lens = 0;
 264                                 if (++lens == lens_end)
 265                                         return 0;
 266                         }
 267                         break;
 268                 case 18: /* Longer run of 0's */
 269                         ret = bitstream_read_bits(istream, 5, &num_zeroes);
 270                         if (ret)
 271                                 return ret;
 272                         num_zeroes += 20;
 273                         while (num_zeroes--) {
 274                                 *lens = 0;
 275                                 if (++lens == lens_end)
 276                                         return 0;
 277                         }
 278                         break;
 279                 case 19: /* Run of identical lengths */
 280                         ret = bitstream_read_bits(istream, 1, &num_same);
 281                         if (ret)
 282                                 return ret;
 283                         num_same += 4;
 284                         ret = read_huffsym_using_pretree(istream,
 285                                                          pretree_decode_table,
 286                                                          pretree_lens,
 287                                                          &code);
 288                         if (ret)
 289                                 return ret;
 290                         value = (signed char)*lens - (signed char)code;
 291                         if (value < 0)
 292                                 value += 17;
 293                         while (num_same--) {
 294                                 *lens = value;
 295                                 if (++lens == lens_end)
 296                                         return 0;
 297                         }
 298                         break;
 299                 default: /* Difference from old length. */
 300                         value = (signed char)*lens - (signed char)tree_code;
 301                         if (value < 0)
 302                                 value += 17;
 303                         *lens = value;
 304                         if (++lens == lens_end)
 305                                 return 0;
 306                         break;
 307                 }
 308         }
 309 }
 310
 311 /*
 312  * Reads the header for an LZX-compressed block.
 313  *
 314  * @istream:            The input bitstream.
 315  * @block_size_ret:     A pointer to an int into which the size of the block,
 316  *                              in bytes, will be returned.
 317  * @block_type_ret:     A pointer to an int into which the type of the block
 318  *                              (LZX_BLOCKTYPE_*) will be returned.
 319  * @tables:             A pointer to a lzx_tables structure in which the
 320  *                              main tree, the length tree, and possibly the
 321  *                              aligned offset tree will be constructed.
 322  * @queue:      A pointer to the least-recently-used queue into which
 323  *                      R0, R1, and R2 will be written (only for uncompressed
 324  *                      blocks, which contain this information in the header)
 325  */
 326 static int
 327 lzx_read_block_header(struct input_bitstream *istream,
 328                       unsigned num_main_syms,
 329                       unsigned max_window_size,
 330                       unsigned *block_size_ret,
 331                       unsigned *block_type_ret,
 332                       struct lzx_tables *tables,
 333                       struct lzx_lru_queue *queue)
 334 {
 335         int ret;
 336         unsigned block_type;
 337         unsigned block_size;
 338
 339         ret = bitstream_ensure_bits(istream, 4);
 340         if (ret)
 341                 return ret;
 342
 343         /* The first three bits tell us what kind of block it is, and are one
 344          * of the LZX_BLOCKTYPE_* values.  */
 345         block_type = bitstream_read_bits_nocheck(istream, 3);
 346
 347         /* Read the block size.  This mirrors the behavior
 348          * lzx_write_compressed_block() in lzx-compress.c; see that for more
 349          * details.  */
 350         if (bitstream_read_bits_nocheck(istream, 1)) {
 351                 block_size = LZX_DEFAULT_BLOCK_SIZE;
 352         } else {
 353                 u32 tmp;
 354                 block_size = 0;
 355
 356                 ret = bitstream_read_bits(istream, 8, &tmp);
 357                 if (ret)
 358                         return ret;
 359                 block_size |= tmp;
 360
 361                 ret = bitstream_read_bits(istream, 8, &tmp);
 362                 if (ret)
 363                         return ret;
 364                 block_size <<= 8;
 365                 block_size |= tmp;
 366
 367                 if (max_window_size >= 65536) {
 368                         ret = bitstream_read_bits(istream, 8, &tmp);
 369                         if (ret)
 370                                 return ret;
 371                         block_size <<= 8;
 372                         block_size |= tmp;
 373                 }
 374         }
 375
 376         switch (block_type) {
 377         case LZX_BLOCKTYPE_ALIGNED:
 378                 /* Read the path lengths for the elements of the aligned tree,
 379                  * then build it. */
 380
 381                 for (unsigned i = 0; i < LZX_ALIGNEDCODE_NUM_SYMBOLS; i++) {
 382                         u32 len;
 383
 384                         ret = bitstream_read_bits(istream,
 385                                                   LZX_ALIGNEDCODE_ELEMENT_SIZE,
 386                                                   &len);
 387                         if (ret)
 388                                 return ret;
 389                         tables->alignedtree_lens[i] = len;
 390                 }
 391
 392                 LZX_DEBUG("Building the aligned tree.");
 393                 ret = make_huffman_decode_table(tables->alignedtree_decode_table,
 394                                                 LZX_ALIGNEDCODE_NUM_SYMBOLS,
 395                                                 LZX_ALIGNEDCODE_TABLEBITS,
 396                                                 tables->alignedtree_lens,
 397                                                 LZX_MAX_ALIGNED_CODEWORD_LEN);
 398                 if (ret) {
 399                         LZX_DEBUG("Failed to make the decode table for the "
 400                                   "aligned offset tree");
 401                         return ret;
 402                 }
 403
 404                 /* Fall though, since the rest of the header for aligned offset
 405                  * blocks is the same as that for verbatim blocks */
 406
 407         case LZX_BLOCKTYPE_VERBATIM:
 408                 if (block_type == LZX_BLOCKTYPE_VERBATIM)
 409                         LZX_DEBUG("Found verbatim block.");
 410
 411                 LZX_DEBUG("Reading path lengths for main tree.");
 412                 /* Read the path lengths for the first 256 elements of the main
 413                  * tree. */
 414                 ret = lzx_read_code_lens(istream, tables->maintree_lens,
 415                                          LZX_NUM_CHARS);
 416                 if (ret) {
 417                         LZX_DEBUG("Failed to read the code lengths for the "
 418                                   "first 256 elements of the main tree");
 419                         return ret;
 420                 }
 421
 422                 /* Read the path lengths for the remaining elements of the main
 423                  * tree. */
 424                 LZX_DEBUG("Reading path lengths for remaining elements of "
 425                           "main tree (%d elements).",
 426                           num_main_syms - LZX_NUM_CHARS);
 427                 ret = lzx_read_code_lens(istream,
 428                                          tables->maintree_lens + LZX_NUM_CHARS,
 429                                          num_main_syms - LZX_NUM_CHARS);
 430                 if (ret) {
 431                         LZX_DEBUG("Failed to read the path lengths for the "
 432                                   "remaining elements of the main tree");
 433                         return ret;
 434                 }
 435
 436                 LZX_DEBUG("Building the Huffman decoding "
 437                           "table for the main tree.");
 438
 439                 ret = make_huffman_decode_table(tables->maintree_decode_table,
 440                                                 num_main_syms,
 441                                                 LZX_MAINCODE_TABLEBITS,
 442                                                 tables->maintree_lens,
 443                                                 LZX_MAX_MAIN_CODEWORD_LEN);
 444                 if (ret) {
 445                         LZX_DEBUG("Failed to make the decode "
 446                                   "table for the main tree");
 447                         return ret;
 448                 }
 449
 450                 LZX_DEBUG("Reading path lengths for the length tree.");
 451                 ret = lzx_read_code_lens(istream, tables->lentree_lens,
 452                                          LZX_LENCODE_NUM_SYMBOLS);
 453                 if (ret) {
 454                         LZX_DEBUG("Failed to read the path "
 455                                   "lengths for the length tree");
 456                         return ret;
 457                 }
 458
 459                 LZX_DEBUG("Building the length tree.");
 460                 ret = make_huffman_decode_table(tables->lentree_decode_table,
 461                                                 LZX_LENCODE_NUM_SYMBOLS,
 462                                                 LZX_LENCODE_TABLEBITS,
 463                                                 tables->lentree_lens,
 464                                                 LZX_MAX_LEN_CODEWORD_LEN);
 465                 if (ret) {
 466                         LZX_DEBUG("Failed to build the length Huffman tree");
 467                         return ret;
 468                 }
 469                 /* The bitstream of compressed literals and matches for this
 470                  * block directly follows and will be read in
 471                  * lzx_decompress_block(). */
 472                 break;
 473         case LZX_BLOCKTYPE_UNCOMPRESSED:
 474                 LZX_DEBUG("Found uncompressed block.");
 475                 /* Before reading the three LRU match offsets from the
 476                  * uncompressed block header, the stream needs to be aligned on
 477                  * a 16-bit boundary.  But, unexpectedly, if the stream is
 478                  * *already* aligned, the correct thing to do is to throw away
 479                  * the next 16 bits. */
 480                 if (istream->bitsleft == 0) {
 481                         if (istream->data_bytes_left < 14) {
 482                                 LZX_DEBUG("Insufficient length in "
 483                                           "uncompressed block");
 484                                 return -1;
 485                         }
 486                         istream->data += 2;
 487                         istream->data_bytes_left -= 2;
 488                 } else {
 489                         if (istream->data_bytes_left < 12) {
 490                                 LZX_DEBUG("Insufficient length in "
 491                                           "uncompressed block");
 492                                 return -1;
 493                         }
 494                         istream->bitsleft = 0;
 495                         istream->bitbuf = 0;
 496                 }
 497                 queue->R[0] = le32_to_cpu(*(u32*)(istream->data + 0));
 498                 queue->R[1] = le32_to_cpu(*(u32*)(istream->data + 4));
 499                 queue->R[2] = le32_to_cpu(*(u32*)(istream->data + 8));
 500                 istream->data += 12;
 501                 istream->data_bytes_left -= 12;
 502                 /* The uncompressed data of this block directly follows and will
 503                  * be read in lzx_decompress(). */
 504                 break;
 505         default:
 506                 LZX_DEBUG("Found invalid block");
 507                 return -1;
 508         }
 509         *block_type_ret = block_type;
 510         *block_size_ret = block_size;
 511         return 0;
 512 }
 513
 514 /*
 515  * Decodes a compressed match from a block of LZX-compressed data.  A match
 516  * refers to some match_offset to a point earlier in the window as well as some
 517  * match_len, for which the data is to be copied to the current position in the
 518  * window.
 519  *
 520  * @main_element:       The start of the match data, as decoded using the main
 521  *                      tree.
 522  *
 523  * @block_type:         The type of the block (LZX_BLOCKTYPE_ALIGNED or
 524  *                      LZX_BLOCKTYPE_VERBATIM)
 525  *
 526  * @bytes_remaining:    The amount of uncompressed data remaining to be
 527  *                      uncompressed in this block.  It is an error if the match
 528  *                      is longer than this number.
 529  *
 530  * @window:             A pointer to the window into which the uncompressed
 531  *                      data is being written.
 532  *
 533  * @window_pos:         The current byte offset in the window.
 534  *
 535  * @tables:             The Huffman decoding tables for this LZX block (main
 536  *                      code, length code, and for LZX_BLOCKTYPE_ALIGNED blocks,
 537  *                      also the aligned offset code).
 538  *
 539  * @queue:              The least-recently used queue for match offsets.
 540  *
 541  * @istream:            The input bitstream.
 542  *
 543  * Returns the length of the match, or a negative number on error.  The possible
 544  * error cases are:
 545  *      - Match would exceed the amount of data remaining to be uncompressed.
 546  *      - Match refers to data before the window.
 547  *      - The input bitstream ended unexpectedly.
 548  */
 549 static int
 550 lzx_decode_match(unsigned main_element, int block_type,
 551                  unsigned bytes_remaining, u8 *window,
 552                  unsigned window_pos,
 553                  const struct lzx_tables *tables,
 554                  struct lzx_lru_queue *queue,
 555                  struct input_bitstream *istream)
 556 {
 557         unsigned length_header;
 558         unsigned position_slot;
 559         unsigned match_len;
 560         unsigned match_offset;
 561         unsigned additional_len;
 562         unsigned num_extra_bits;
 563         u32 verbatim_bits;
 564         u32 aligned_bits;
 565         unsigned i;
 566         int ret;
 567         u8 *match_dest;
 568         u8 *match_src;
 569
 570         /* The main element is offset by 256 because values under 256 indicate a
 571          * literal value. */
 572         main_element -= LZX_NUM_CHARS;
 573
 574         /* The length header consists of the lower 3 bits of the main element.
 575          * The position slot is the rest of it. */
 576         length_header = main_element & LZX_NUM_PRIMARY_LENS;
 577         position_slot = main_element >> 3;
 578
 579         /* If the length_header is less than LZX_NUM_PRIMARY_LENS (= 7), it
 580          * gives the match length as the offset from LZX_MIN_MATCH_LEN.
 581          * Otherwise, the length is given by an additional symbol encoded using
 582          * the length tree, offset by 9 (LZX_MIN_MATCH_LEN +
 583          * LZX_NUM_PRIMARY_LENS) */
 584         match_len = LZX_MIN_MATCH_LEN + length_header;
 585         if (length_header == LZX_NUM_PRIMARY_LENS) {
 586                 ret = read_huffsym_using_lentree(istream, tables,
 587                                                  &additional_len);
 588                 if (ret)
 589                         return ret;
 590                 match_len += additional_len;
 591         }
 592
 593
 594         /* If the position_slot is 0, 1, or 2, the match offset is retrieved
 595          * from the LRU queue.  Otherwise, the match offset is not in the LRU
 596          * queue. */
 597         switch (position_slot) {
 598         case 0:
 599                 match_offset = queue->R[0];
 600                 break;
 601         case 1:
 602                 match_offset = queue->R[1];
 603                 swap(queue->R[0], queue->R[1]);
 604                 break;
 605         case 2:
 606                 /* The queue doesn't work quite the same as a real LRU queue,
 607                  * since using the R2 offset doesn't bump the R1 offset down to
 608                  * R2. */
 609                 match_offset = queue->R[2];
 610                 swap(queue->R[0], queue->R[2]);
 611                 break;
 612         default:
 613                 /* Otherwise, the offset was not encoded as one the offsets in
 614                  * the queue.  Depending on the position slot, there is a
 615                  * certain number of extra bits that need to be read to fully
 616                  * decode the match offset. */
 617
 618                 /* Look up the number of extra bits that need to be read. */
 619                 num_extra_bits = lzx_get_num_extra_bits(position_slot);
 620
 621                 /* For aligned blocks, if there are at least 3 extra bits, the
 622                  * actual number of extra bits is 3 less, and they encode a
 623                  * number of 8-byte words that are added to the offset; there
 624                  * is then an additional symbol read using the aligned tree that
 625                  * specifies the actual byte alignment. */
 626                 if (block_type == LZX_BLOCKTYPE_ALIGNED && num_extra_bits >= 3) {
 627
 628                         /* There is an error in the LZX "specification" at this
 629                          * point; it indicates that a Huffman symbol is to be
 630                          * read only if num_extra_bits is greater than 3, but
 631                          * actually it is if num_extra_bits is greater than or
 632                          * equal to 3.  (Note that in the case with
 633                          * num_extra_bits == 3, the assignment to verbatim_bits
 634                          * will just set it to 0. ) */
 635                         ret = bitstream_read_bits(istream, num_extra_bits - 3,
 636                                                   &verbatim_bits);
 637                         if (ret)
 638                                 return ret;
 639
 640                         verbatim_bits <<= 3;
 641
 642                         ret = read_huffsym_using_alignedtree(istream, tables,
 643                                                              &aligned_bits);
 644                         if (ret)
 645                                 return ret;
 646                 } else {
 647                         /* For non-aligned blocks, or for aligned blocks with
 648                          * less than 3 extra bits, the extra bits are added
 649                          * directly to the match offset, and the correction for
 650                          * the alignment is taken to be 0. */
 651                         ret = bitstream_read_bits(istream, num_extra_bits,
 652                                                   &verbatim_bits);
 653                         if (ret)
 654                                 return ret;
 655
 656                         aligned_bits = 0;
 657                 }
 658
 659                 /* Calculate the match offset. */
 660                 match_offset = lzx_position_base[position_slot] +
 661                                verbatim_bits + aligned_bits - LZX_OFFSET_OFFSET;
 662
 663                 /* Update the LRU queue. */
 664                 queue->R[2] = queue->R[1];
 665                 queue->R[1] = queue->R[0];
 666                 queue->R[0] = match_offset;
 667                 break;
 668         }
 669
 670         /* Verify that the match is in the bounds of the part of the window
 671          * currently in use, then copy the source of the match to the current
 672          * position. */
 673
 674         if (match_len > bytes_remaining) {
 675                 LZX_DEBUG("Match of length %u bytes overflows "
 676                           "uncompressed block size", match_len);
 677                 return -1;
 678         }
 679
 680         if (match_offset > window_pos) {
 681                 LZX_DEBUG("Match of length %u bytes references "
 682                           "data before window (match_offset = %u, "
 683                           "window_pos = %u)",
 684                           match_len, match_offset, window_pos);
 685                 return -1;
 686         }
 687
 688         match_dest = window + window_pos;
 689         match_src = match_dest - match_offset;
 690
 691 #if 0
 692         printf("Match: src %u, dst %u, len %u\n", match_src - window,
 693                                                 match_dest - window,
 694                                                 match_len);
 695         putchar('|');
 696         for (i = 0; i < match_len; i++) {
 697                 match_dest[i] = match_src[i];
 698                 putchar(match_src[i]);
 699         }
 700         putchar('|');
 701         putchar('\n');
 702 #else
 703         for (i = 0; i < match_len; i++)
 704                 match_dest[i] = match_src[i];
 705 #endif
 706
 707         return match_len;
 708 }
 709
 710 static void
 711 undo_call_insn_translation(u32 *call_insn_target, int input_pos,
 712                            s32 file_size)
 713 {
 714         s32 abs_offset;
 715         s32 rel_offset;
 716
 717         abs_offset = le32_to_cpu(*call_insn_target);
 718         if (abs_offset >= -input_pos && abs_offset < file_size) {
 719                 if (abs_offset >= 0) {
 720                         /* "good translation" */
 721                         rel_offset = abs_offset - input_pos;
 722                 } else {
 723                         /* "compensating translation" */
 724                         rel_offset = abs_offset + file_size;
 725                 }
 726                 *call_insn_target = cpu_to_le32(rel_offset);
 727         }
 728 }
 729
 730 /* Undo the 'E8' preprocessing, where the targets of x86 CALL instructions were
 731  * changed from relative offsets to absolute offsets.
 732  *
 733  * Note that this call instruction preprocessing can and will be used on any
 734  * data even if it is not actually x86 machine code.  In fact, this type of
 735  * preprocessing appears to always be used in LZX-compressed resources in WIM
 736  * files; there is no bit to indicate whether it is used or not, unlike in the
 737  * LZX compressed format as used in cabinet files, where a bit is reserved for
 738  * that purpose.
 739  *
 740  * Call instruction preprocessing is disabled in the last 6 bytes of the
 741  * uncompressed data, which really means the 5-byte call instruction cannot
 742  * start in the last 10 bytes of the uncompressed data.  This is one of the
 743  * errors in the LZX documentation.
 744  *
 745  * Call instruction preprocessing does not appear to be disabled after the
 746  * 32768th chunk of a WIM stream, which is apparently is yet another difference
 747  * from the LZX compression used in cabinet files.
 748  *
 749  * Call instruction processing is supposed to take the file size as a parameter,
 750  * as it is used in calculating the translated jump targets.  But in WIM files,
 751  * this file size is always the same (LZX_WIM_MAGIC_FILESIZE == 12000000).*/
 752 static void
 753 undo_call_insn_preprocessing(u8 uncompressed_data[], int uncompressed_data_len)
 754 {
 755         for (int i = 0; i < uncompressed_data_len - 10; i++) {
 756                 if (uncompressed_data[i] == 0xe8) {
 757                         undo_call_insn_translation((u32*)&uncompressed_data[i + 1],
 758                                                    i,
 759                                                    LZX_WIM_MAGIC_FILESIZE);
 760                         i += 4;
 761                 }
 762         }
 763 }
 764
 765 /*
 766  * Decompresses a LZX-compressed block of data from which the header has already
 767  * been read.
 768  *
 769  * @block_type: The type of the block (LZX_BLOCKTYPE_VERBATIM or
 770  *              LZX_BLOCKTYPE_ALIGNED)
 771  * @block_size: The size of the block, in bytes.
 772  * @num_main_syms:      Number of symbols in the main alphabet.
 773  * @window:     Pointer to the decompression window.
 774  * @window_pos: The current position in the window.  Will be 0 for the first
 775  *                      block.
 776  * @tables:     The Huffman decoding tables for the block (main, length, and
 777  *                      aligned offset, the latter only for LZX_BLOCKTYPE_ALIGNED)
 778  * @queue:      The least-recently-used queue for match offsets.
 779  * @istream:    The input bitstream for the compressed literals.
 780  */
 781 static int
 782 lzx_decompress_block(int block_type, unsigned block_size,
 783                      unsigned num_main_syms,
 784                      u8 *window,
 785                      unsigned window_pos,
 786                      const struct lzx_tables *tables,
 787                      struct lzx_lru_queue *queue,
 788                      struct input_bitstream *istream)
 789 {
 790         unsigned main_element;
 791         unsigned end;
 792         int ret;
 793         int match_len;
 794
 795         end = window_pos + block_size;
 796         while (window_pos < end) {
 797                 ret = read_huffsym_using_maintree(istream, tables,
 798                                                   &main_element,
 799                                                   num_main_syms);
 800                 if (ret)
 801                         return ret;
 802
 803                 if (main_element < LZX_NUM_CHARS) {
 804                         /* literal: 0 to LZX_NUM_CHARS - 1 */
 805                         window[window_pos++] = main_element;
 806                 } else {
 807                         /* match: LZX_NUM_CHARS to num_main_syms - 1 */
 808                         match_len = lzx_decode_match(main_element,
 809                                                      block_type,
 810                                                      end - window_pos,
 811                                                      window,
 812                                                      window_pos,
 813                                                      tables,
 814                                                      queue,
 815                                                      istream);
 816                         if (match_len < 0)
 817                                 return match_len;
 818                         window_pos += match_len;
 819                 }
 820         }
 821         return 0;
 822 }
 823
 824 WIMLIBAPI int
 825 wimlib_lzx_decompress2(const void *compressed_data, unsigned compressed_len,
 826                        void *uncompressed_data, unsigned uncompressed_len,
 827                        u32 max_window_size)
 828 {
 829         struct lzx_tables tables;
 830         struct input_bitstream istream;
 831         struct lzx_lru_queue queue;
 832         unsigned window_pos;
 833         unsigned block_size;
 834         unsigned block_type;
 835         unsigned num_main_syms;
 836         int ret;
 837         bool e8_preprocessing_done;
 838
 839         LZX_DEBUG("compressed_data = %p, compressed_len = %u, "
 840                   "uncompressed_data = %p, uncompressed_len = %u, "
 841                   "max_window_size=%u).",
 842                   compressed_data, compressed_len,
 843                   uncompressed_data, uncompressed_len, max_window_size);
 844
 845         if (!lzx_window_size_valid(max_window_size)) {
 846                 LZX_DEBUG("Window size of %u is invalid!",
 847                           max_window_size);
 848                 return -1;
 849         }
 850
 851         num_main_syms = lzx_get_num_main_syms(max_window_size);
 852
 853         if (uncompressed_len > max_window_size) {
 854                 LZX_DEBUG("Uncompressed chunk size of %u exceeds "
 855                           "window size of %u!",
 856                           uncompressed_len, max_window_size);
 857                 return -1;
 858         }
 859
 860         memset(tables.maintree_lens, 0, sizeof(tables.maintree_lens));
 861         memset(tables.lentree_lens, 0, sizeof(tables.lentree_lens));
 862         lzx_lru_queue_init(&queue);
 863         init_input_bitstream(&istream, compressed_data, compressed_len);
 864
 865         e8_preprocessing_done = false; /* Set to true if there may be 0xe8 bytes
 866                                           in the uncompressed data. */
 867
 868         /* The compressed data will consist of one or more blocks.  The
 869          * following loop decompresses one block, and it runs until there all
 870          * the compressed data has been decompressed, so there are no more
 871          * blocks.  */
 872
 873         for (window_pos = 0;
 874              window_pos < uncompressed_len;
 875              window_pos += block_size)
 876         {
 877                 LZX_DEBUG("Reading block header.");
 878                 ret = lzx_read_block_header(&istream, num_main_syms,
 879                                             max_window_size, &block_size,
 880                                             &block_type, &tables, &queue);
 881                 if (ret)
 882                         return ret;
 883
 884                 LZX_DEBUG("block_size = %u, window_pos = %u",
 885                           block_size, window_pos);
 886
 887                 if (block_size > uncompressed_len - window_pos) {
 888                         LZX_DEBUG("Expected a block size of at "
 889                                   "most %u bytes (found %u bytes)",
 890                                   uncompressed_len - window_pos, block_size);
 891                         return -1;
 892                 }
 893
 894                 switch (block_type) {
 895                 case LZX_BLOCKTYPE_VERBATIM:
 896                 case LZX_BLOCKTYPE_ALIGNED:
 897                         if (block_type == LZX_BLOCKTYPE_VERBATIM)
 898                                 LZX_DEBUG("LZX_BLOCKTYPE_VERBATIM");
 899                         else
 900                                 LZX_DEBUG("LZX_BLOCKTYPE_ALIGNED");
 901                         ret = lzx_decompress_block(block_type,
 902                                                    block_size,
 903                                                    num_main_syms,
 904                                                    uncompressed_data,
 905                                                    window_pos,
 906                                                    &tables,
 907                                                    &queue,
 908                                                    &istream);
 909                         if (ret)
 910                                 return ret;
 911
 912                         if (tables.maintree_lens[0xe8] != 0)
 913                                 e8_preprocessing_done = true;
 914                         break;
 915                 case LZX_BLOCKTYPE_UNCOMPRESSED:
 916                         LZX_DEBUG("LZX_BLOCKTYPE_UNCOMPRESSED");
 917                         if (istream.data_bytes_left < block_size) {
 918                                 LZX_DEBUG("Unexpected end of input when "
 919                                           "reading %u bytes from LZX bitstream "
 920                                           "(only have %u bytes left)",
 921                                           block_size, istream.data_bytes_left);
 922                                 return -1;
 923                         }
 924                         memcpy(&((u8*)uncompressed_data)[window_pos], istream.data,
 925                                block_size);
 926                         istream.data += block_size;
 927                         istream.data_bytes_left -= block_size;
 928                         /* Re-align bitstream if an odd number of bytes were
 929                          * read.  */
 930                         if (istream.data_bytes_left && (block_size & 1)) {
 931                                 istream.data_bytes_left--;
 932                                 istream.data++;
 933                         }
 934                         e8_preprocessing_done = true;
 935                         break;
 936                 }
 937         }
 938         if (e8_preprocessing_done)
 939                 undo_call_insn_preprocessing(uncompressed_data, uncompressed_len);
 940         return 0;
 941 }
 942
 943 /* API function documented in wimlib.h  */
 944 WIMLIBAPI int
 945 wimlib_lzx_decompress(const void *compressed_data, unsigned compressed_len,
 946                       void *uncompressed_data, unsigned uncompressed_len)
 947 {
 948         return wimlib_lzx_decompress2(compressed_data, compressed_len,
 949                                       uncompressed_data, uncompressed_len,
 950                                       32768);
 951 }