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, 2014 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 an 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.
  32  *
  33  * Some notes on the LZX compression format as used in Windows Imaging (WIM)
  34  * files:
  35  *
  36  * A compressed WIM resource consists of a table of chunk offsets followed by
  37  * the compressed chunks themselves.  All compressed chunks except possibly the
  38  * last decompress to a fixed number of bytes, by default 32768.  This is quite
  39  * similar to the cabinet (.cab) file format, but they are not the same.
  40  * According to the cabinet format documentation, the LZX block size is
  41  * independent from the CFDATA blocks, and an LZX block may span several CFDATA
  42  * blocks.  However, in WIMs, LZX blocks do not appear to ever span multiple WIM
  43  * chunks.  Note that this means any WIM chunk may be decompressed or compressed
  44  * independently from any other chunk, which allows random access.
  45  *
  46  * An LZX compressed WIM chunk contains one or more LZX blocks of the aligned,
  47  * verbatim, or uncompressed block types.  For aligned and verbatim blocks, the
  48  * size of the block in uncompressed bytes is specified by a bit following the 3
  49  * bits that specify the block type, possibly followed by an additional 16 bits.
  50  * '1' means to use the default block size (equal to 32768, the default size of
  51  * a WIM chunk), while '0' means that the block size is provided by the next 16
  52  * bits.
  53  *
  54  * The cabinet format, as documented, allows for the possibility that a
  55  * compressed CFDATA chunk is up to 6144 bytes larger than the data it
  56  * uncompresses to.  However, in the WIM format it appears that every chunk that
  57  * would be 32768 bytes or more when compressed is actually stored fully
  58  * uncompressed.
  59  *
  60  * The 'e8' preprocessing step that changes x86 call instructions to use
  61  * absolute offsets instead of relative offsets relies on a filesize parameter.
  62  * There is no such parameter for this in the WIM files (even though the size of
  63  * the file resource could be used for this purpose), and instead a magic file
  64  * size of 12000000 is used.  The 'e8' preprocessing is always done, and there
  65  * is no bit to indicate whether it is done or not.
  66  */
  67
  68 /*
  69  * Some more notes about errors in Microsoft's LZX documentation:
  70  *
  71  * Microsoft's LZX document and their implementation of the com.ms.util.cab Java
  72  * package do not concur.
  73  *
  74  * In the LZX document, there is a table showing the correlation between window
  75  * size and the number of position slots. It states that the 1MB window = 40
  76  * slots and the 2MB window = 42 slots. In the implementation, 1MB = 42 slots,
  77  * 2MB = 50 slots. The actual calculation is 'find the first slot whose position
  78  * base is equal to or more than the required window size'. This would explain
  79  * why other tables in the document refer to 50 slots rather than 42.
  80  *
  81  * The constant NUM_PRIMARY_LENS used in the decompression pseudocode is not
  82  * defined in the specification.
  83  *
  84  * The LZX document states that aligned offset blocks have their aligned offset
  85  * Huffman tree AFTER the main and length trees. The implementation suggests
  86  * that the aligned offset tree is BEFORE the main and length trees.
  87  *
  88  * The LZX document decoding algorithm states that, in an aligned offset block,
  89  * if an extra_bits value is 1, 2 or 3, then that number of bits should be read
  90  * and the result added to the match offset. This is correct for 1 and 2, but
  91  * not 3, where just a Huffman symbol (using the aligned tree) should be read.
  92  *
  93  * Regarding the E8 preprocessing, the LZX document states 'No translation may
  94  * be performed on the last 6 bytes of the input block'. This is correct.
  95  * However, the pseudocode provided checks for the *E8 leader* up to the last 6
  96  * bytes. If the leader appears between -10 and -7 bytes from the end, this
  97  * would cause the next four bytes to be modified, at least one of which would
  98  * be in the last 6 bytes, which is not allowed according to the spec.
  99  *
 100  * The specification states that the Huffman trees must always contain at least
 101  * one element. However, many CAB files contain blocks where the length tree is
 102  * completely empty (because there are no matches), and this is expected to
 103  * succeed.
 104  */
 105
 106 #ifdef HAVE_CONFIG_H
 107 #  include "config.h"
 108 #endif
 109
 110 #include "wimlib.h"
 111 #include "wimlib/decompressor_ops.h"
 112 #include "wimlib/decompress_common.h"
 113 #include "wimlib/lzx.h"
 114 #include "wimlib/util.h"
 115
 116 #include <string.h>
 117
 118 #ifdef __SSE2__
 119 #  include <emmintrin.h>
 120 #endif
 121
 122 /* Huffman decoding tables and maps from symbols to code lengths. */
 123 struct lzx_tables {
 124
 125         u16 maintree_decode_table[(1 << LZX_MAINCODE_TABLEBITS) +
 126                                         (LZX_MAINCODE_MAX_NUM_SYMBOLS * 2)]
 127                                         _aligned_attribute(DECODE_TABLE_ALIGNMENT);
 128         u8 maintree_lens[LZX_MAINCODE_MAX_NUM_SYMBOLS];
 129
 130
 131         u16 lentree_decode_table[(1 << LZX_LENCODE_TABLEBITS) +
 132                                         (LZX_LENCODE_NUM_SYMBOLS * 2)]
 133                                         _aligned_attribute(DECODE_TABLE_ALIGNMENT);
 134         u8 lentree_lens[LZX_LENCODE_NUM_SYMBOLS];
 135
 136
 137         u16 alignedtree_decode_table[(1 << LZX_ALIGNEDCODE_TABLEBITS) +
 138                                         (LZX_ALIGNEDCODE_NUM_SYMBOLS * 2)]
 139                                         _aligned_attribute(DECODE_TABLE_ALIGNMENT);
 140         u8 alignedtree_lens[LZX_ALIGNEDCODE_NUM_SYMBOLS];
 141 } _aligned_attribute(DECODE_TABLE_ALIGNMENT);
 142
 143 struct lzx_decompressor {
 144         u32 max_window_size;
 145         unsigned num_main_syms;
 146         struct lzx_tables tables;
 147 };
 148
 149 /*
 150  * Reads a Huffman-encoded symbol using the pre-tree.
 151  */
 152 static inline int
 153 read_huffsym_using_pretree(struct input_bitstream *istream,
 154                            const u16 pretree_decode_table[],
 155                            const u8 pretree_lens[], unsigned *n)
 156 {
 157         return read_huffsym(istream, pretree_decode_table, pretree_lens,
 158                             LZX_PRECODE_NUM_SYMBOLS, LZX_PRECODE_TABLEBITS, n,
 159                             LZX_MAX_PRE_CODEWORD_LEN);
 160 }
 161
 162 /* Reads a Huffman-encoded symbol using the main tree. */
 163 static inline int
 164 read_huffsym_using_maintree(struct input_bitstream *istream,
 165                             const struct lzx_tables *tables,
 166                             unsigned *n,
 167                             unsigned num_main_syms)
 168 {
 169         return read_huffsym(istream, tables->maintree_decode_table,
 170                             tables->maintree_lens, num_main_syms,
 171                             LZX_MAINCODE_TABLEBITS, n, LZX_MAX_MAIN_CODEWORD_LEN);
 172 }
 173
 174 /* Reads a Huffman-encoded symbol using the length tree. */
 175 static inline int
 176 read_huffsym_using_lentree(struct input_bitstream *istream,
 177                            const struct lzx_tables *tables,
 178                            unsigned *n)
 179 {
 180         return read_huffsym(istream, tables->lentree_decode_table,
 181                             tables->lentree_lens, LZX_LENCODE_NUM_SYMBOLS,
 182                             LZX_LENCODE_TABLEBITS, n, LZX_MAX_LEN_CODEWORD_LEN);
 183 }
 184
 185 /* Reads a Huffman-encoded symbol using the aligned offset tree. */
 186 static inline int
 187 read_huffsym_using_alignedtree(struct input_bitstream *istream,
 188                                const struct lzx_tables *tables,
 189                                unsigned *n)
 190 {
 191         return read_huffsym(istream, tables->alignedtree_decode_table,
 192                             tables->alignedtree_lens,
 193                             LZX_ALIGNEDCODE_NUM_SYMBOLS,
 194                             LZX_ALIGNEDCODE_TABLEBITS, n,
 195                             LZX_MAX_ALIGNED_CODEWORD_LEN);
 196 }
 197
 198 /*
 199  * Reads the pretree from the input, then uses the pretree to decode @num_lens
 200  * code length values from the input.
 201  *
 202  * @istream:    The bit stream for the input.  It is positioned on the beginning
 203  *                      of the pretree for the code length values.
 204  * @lens:       An array that contains the length values from the previous time
 205  *                      the code lengths for this Huffman tree were read, or all
 206  *                      0's if this is the first time.
 207  * @num_lens:   Number of length values to decode and return.
 208  *
 209  */
 210 static int
 211 lzx_read_code_lens(struct input_bitstream *istream, u8 lens[],
 212                    unsigned num_lens)
 213 {
 214         /* Declare the decoding table and length table for the pretree. */
 215         u16 pretree_decode_table[(1 << LZX_PRECODE_TABLEBITS) +
 216                                         (LZX_PRECODE_NUM_SYMBOLS * 2)]
 217                                         _aligned_attribute(DECODE_TABLE_ALIGNMENT);
 218         u8 pretree_lens[LZX_PRECODE_NUM_SYMBOLS];
 219         unsigned i;
 220         u32 len;
 221         int ret;
 222
 223         /* Read the code lengths of the pretree codes.  There are 20 lengths of
 224          * 4 bits each. */
 225         for (i = 0; i < LZX_PRECODE_NUM_SYMBOLS; i++) {
 226                 ret = bitstream_read_bits(istream, LZX_PRECODE_ELEMENT_SIZE,
 227                                           &len);
 228                 if (ret)
 229                         return ret;
 230                 pretree_lens[i] = len;
 231         }
 232
 233         /* Make the decoding table for the pretree. */
 234         ret = make_huffman_decode_table(pretree_decode_table,
 235                                         LZX_PRECODE_NUM_SYMBOLS,
 236                                         LZX_PRECODE_TABLEBITS,
 237                                         pretree_lens,
 238                                         LZX_MAX_PRE_CODEWORD_LEN);
 239         if (ret)
 240                 return ret;
 241
 242         /* Pointer past the last length value that needs to be filled in. */
 243         u8 *lens_end = lens + num_lens;
 244
 245         while (1) {
 246
 247                 /* Decode a symbol from the input.  If the symbol is between 0
 248                  * and 16, it is the difference from the old length.  If it is
 249                  * between 17 and 19, it is a special code that indicates that
 250                  * some number of the next lengths are all 0, or some number of
 251                  * the next lengths are all equal to the next symbol in the
 252                  * input. */
 253                 unsigned tree_code;
 254                 u32 num_zeroes;
 255                 unsigned code;
 256                 u32 num_same;
 257                 signed char value;
 258
 259                 ret = read_huffsym_using_pretree(istream, pretree_decode_table,
 260                                                  pretree_lens, &tree_code);
 261                 if (ret)
 262                         return ret;
 263                 switch (tree_code) {
 264                 case 17: /* Run of 0's */
 265                         ret = bitstream_read_bits(istream, 4, &num_zeroes);
 266                         if (ret)
 267                                 return ret;
 268                         num_zeroes += 4;
 269                         while (num_zeroes--) {
 270                                 *lens = 0;
 271                                 if (++lens == lens_end)
 272                                         return 0;
 273                         }
 274                         break;
 275                 case 18: /* Longer run of 0's */
 276                         ret = bitstream_read_bits(istream, 5, &num_zeroes);
 277                         if (ret)
 278                                 return ret;
 279                         num_zeroes += 20;
 280                         while (num_zeroes--) {
 281                                 *lens = 0;
 282                                 if (++lens == lens_end)
 283                                         return 0;
 284                         }
 285                         break;
 286                 case 19: /* Run of identical lengths */
 287                         ret = bitstream_read_bits(istream, 1, &num_same);
 288                         if (ret)
 289                                 return ret;
 290                         num_same += 4;
 291                         ret = read_huffsym_using_pretree(istream,
 292                                                          pretree_decode_table,
 293                                                          pretree_lens,
 294                                                          &code);
 295                         if (ret)
 296                                 return ret;
 297                         value = (signed char)*lens - (signed char)code;
 298                         if (value < 0)
 299                                 value += 17;
 300                         while (num_same--) {
 301                                 *lens = value;
 302                                 if (++lens == lens_end)
 303                                         return 0;
 304                         }
 305                         break;
 306                 default: /* Difference from old length. */
 307                         value = (signed char)*lens - (signed char)tree_code;
 308                         if (value < 0)
 309                                 value += 17;
 310                         *lens = value;
 311                         if (++lens == lens_end)
 312                                 return 0;
 313                         break;
 314                 }
 315         }
 316 }
 317
 318 /*
 319  * Reads the header for an LZX-compressed block.
 320  *
 321  * @istream:            The input bitstream.
 322  * @block_size_ret:     A pointer to an int into which the size of the block,
 323  *                              in bytes, will be returned.
 324  * @block_type_ret:     A pointer to an int into which the type of the block
 325  *                              (LZX_BLOCKTYPE_*) will be returned.
 326  * @tables:             A pointer to an lzx_tables structure in which the
 327  *                              main tree, the length tree, and possibly the
 328  *                              aligned offset tree will be constructed.
 329  * @queue:      A pointer to the least-recently-used queue into which
 330  *                      R0, R1, and R2 will be written (only for uncompressed
 331  *                      blocks, which contain this information in the header)
 332  */
 333 static int
 334 lzx_read_block_header(struct input_bitstream *istream,
 335                       unsigned num_main_syms,
 336                       unsigned max_window_size,
 337                       unsigned *block_size_ret,
 338                       unsigned *block_type_ret,
 339                       struct lzx_tables *tables,
 340                       struct lzx_lru_queue *queue)
 341 {
 342         int ret;
 343         unsigned block_type;
 344         unsigned block_size;
 345
 346         ret = bitstream_ensure_bits(istream, 4);
 347         if (ret)
 348                 return ret;
 349
 350         /* The first three bits tell us what kind of block it is, and are one
 351          * of the LZX_BLOCKTYPE_* values.  */
 352         block_type = bitstream_read_bits_nocheck(istream, 3);
 353
 354         /* Read the block size.  This mirrors the behavior
 355          * lzx_write_compressed_block() in lzx-compress.c; see that for more
 356          * details.  */
 357         if (bitstream_read_bits_nocheck(istream, 1)) {
 358                 block_size = LZX_DEFAULT_BLOCK_SIZE;
 359         } else {
 360                 u32 tmp;
 361                 block_size = 0;
 362
 363                 ret = bitstream_read_bits(istream, 8, &tmp);
 364                 if (ret)
 365                         return ret;
 366                 block_size |= tmp;
 367
 368                 ret = bitstream_read_bits(istream, 8, &tmp);
 369                 if (ret)
 370                         return ret;
 371                 block_size <<= 8;
 372                 block_size |= tmp;
 373
 374                 if (max_window_size >= 65536) {
 375                         ret = bitstream_read_bits(istream, 8, &tmp);
 376                         if (ret)
 377                                 return ret;
 378                         block_size <<= 8;
 379                         block_size |= tmp;
 380                 }
 381         }
 382
 383         switch (block_type) {
 384         case LZX_BLOCKTYPE_ALIGNED:
 385                 /* Read the path lengths for the elements of the aligned tree,
 386                  * then build it. */
 387
 388                 for (unsigned i = 0; i < LZX_ALIGNEDCODE_NUM_SYMBOLS; i++) {
 389                         u32 len;
 390
 391                         ret = bitstream_read_bits(istream,
 392                                                   LZX_ALIGNEDCODE_ELEMENT_SIZE,
 393                                                   &len);
 394                         if (ret)
 395                                 return ret;
 396                         tables->alignedtree_lens[i] = len;
 397                 }
 398
 399                 LZX_DEBUG("Building the aligned tree.");
 400                 ret = make_huffman_decode_table(tables->alignedtree_decode_table,
 401                                                 LZX_ALIGNEDCODE_NUM_SYMBOLS,
 402                                                 LZX_ALIGNEDCODE_TABLEBITS,
 403                                                 tables->alignedtree_lens,
 404                                                 LZX_MAX_ALIGNED_CODEWORD_LEN);
 405                 if (ret) {
 406                         LZX_DEBUG("Failed to make the decode table for the "
 407                                   "aligned offset tree");
 408                         return ret;
 409                 }
 410
 411                 /* Fall though, since the rest of the header for aligned offset
 412                  * blocks is the same as that for verbatim blocks */
 413
 414         case LZX_BLOCKTYPE_VERBATIM:
 415                 if (block_type == LZX_BLOCKTYPE_VERBATIM)
 416                         LZX_DEBUG("Found verbatim block.");
 417
 418                 LZX_DEBUG("Reading path lengths for main tree.");
 419                 /* Read the path lengths for the first 256 elements of the main
 420                  * tree. */
 421                 ret = lzx_read_code_lens(istream, tables->maintree_lens,
 422                                          LZX_NUM_CHARS);
 423                 if (ret) {
 424                         LZX_DEBUG("Failed to read the code lengths for the "
 425                                   "first 256 elements of the main tree");
 426                         return ret;
 427                 }
 428
 429                 /* Read the path lengths for the remaining elements of the main
 430                  * tree. */
 431                 LZX_DEBUG("Reading path lengths for remaining elements of "
 432                           "main tree (%d elements).",
 433                           num_main_syms - LZX_NUM_CHARS);
 434                 ret = lzx_read_code_lens(istream,
 435                                          tables->maintree_lens + LZX_NUM_CHARS,
 436                                          num_main_syms - LZX_NUM_CHARS);
 437                 if (ret) {
 438                         LZX_DEBUG("Failed to read the path lengths for the "
 439                                   "remaining elements of the main tree");
 440                         return ret;
 441                 }
 442
 443                 LZX_DEBUG("Building the Huffman decoding "
 444                           "table for the main tree.");
 445
 446                 ret = make_huffman_decode_table(tables->maintree_decode_table,
 447                                                 num_main_syms,
 448                                                 LZX_MAINCODE_TABLEBITS,
 449                                                 tables->maintree_lens,
 450                                                 LZX_MAX_MAIN_CODEWORD_LEN);
 451                 if (ret) {
 452                         LZX_DEBUG("Failed to make the decode "
 453                                   "table for the main tree");
 454                         return ret;
 455                 }
 456
 457                 LZX_DEBUG("Reading path lengths for the length tree.");
 458                 ret = lzx_read_code_lens(istream, tables->lentree_lens,
 459                                          LZX_LENCODE_NUM_SYMBOLS);
 460                 if (ret) {
 461                         LZX_DEBUG("Failed to read the path "
 462                                   "lengths for the length tree");
 463                         return ret;
 464                 }
 465
 466                 LZX_DEBUG("Building the length tree.");
 467                 ret = make_huffman_decode_table(tables->lentree_decode_table,
 468                                                 LZX_LENCODE_NUM_SYMBOLS,
 469                                                 LZX_LENCODE_TABLEBITS,
 470                                                 tables->lentree_lens,
 471                                                 LZX_MAX_LEN_CODEWORD_LEN);
 472                 if (ret) {
 473                         LZX_DEBUG("Failed to build the length Huffman tree");
 474                         return ret;
 475                 }
 476                 /* The bitstream of compressed literals and matches for this
 477                  * block directly follows and will be read in
 478                  * lzx_decompress_block(). */
 479                 break;
 480         case LZX_BLOCKTYPE_UNCOMPRESSED:
 481                 LZX_DEBUG("Found uncompressed block.");
 482                 /* Before reading the three LRU match offsets from the
 483                  * uncompressed block header, the stream needs to be aligned on
 484                  * a 16-bit boundary.  But, unexpectedly, if the stream is
 485                  * *already* aligned, the correct thing to do is to throw away
 486                  * the next 16 bits. */
 487                 if (istream->bitsleft == 0) {
 488                         if (istream->data_bytes_left < 14) {
 489                                 LZX_DEBUG("Insufficient length in "
 490                                           "uncompressed block");
 491                                 return -1;
 492                         }
 493                         istream->data += 2;
 494                         istream->data_bytes_left -= 2;
 495                 } else {
 496                         if (istream->data_bytes_left < 12) {
 497                                 LZX_DEBUG("Insufficient length in "
 498                                           "uncompressed block");
 499                                 return -1;
 500                         }
 501                         istream->bitsleft = 0;
 502                         istream->bitbuf = 0;
 503                 }
 504                 queue->R[0] = le32_to_cpu(*(le32*)(istream->data + 0));
 505                 queue->R[1] = le32_to_cpu(*(le32*)(istream->data + 4));
 506                 queue->R[2] = le32_to_cpu(*(le32*)(istream->data + 8));
 507                 istream->data += 12;
 508                 istream->data_bytes_left -= 12;
 509                 /* The uncompressed data of this block directly follows and will
 510                  * be read in lzx_decompress(). */
 511                 break;
 512         default:
 513                 LZX_DEBUG("Found invalid block");
 514                 return -1;
 515         }
 516         *block_type_ret = block_type;
 517         *block_size_ret = block_size;
 518         return 0;
 519 }
 520
 521 /*
 522  * Decodes a compressed match from a block of LZX-compressed data.  A match
 523  * refers to some match_offset to a point earlier in the window as well as some
 524  * match_len, for which the data is to be copied to the current position in the
 525  * window.
 526  *
 527  * @main_element:       The start of the match data, as decoded using the main
 528  *                      tree.
 529  *
 530  * @block_type:         The type of the block (LZX_BLOCKTYPE_ALIGNED or
 531  *                      LZX_BLOCKTYPE_VERBATIM)
 532  *
 533  * @bytes_remaining:    The amount of uncompressed data remaining to be
 534  *                      uncompressed in this block.  It is an error if the match
 535  *                      is longer than this number.
 536  *
 537  * @window:             A pointer to the window into which the uncompressed
 538  *                      data is being written.
 539  *
 540  * @window_pos:         The current byte offset in the window.
 541  *
 542  * @tables:             The Huffman decoding tables for this LZX block (main
 543  *                      code, length code, and for LZX_BLOCKTYPE_ALIGNED blocks,
 544  *                      also the aligned offset code).
 545  *
 546  * @queue:              The least-recently used queue for match offsets.
 547  *
 548  * @istream:            The input bitstream.
 549  *
 550  * Returns the length of the match, or a negative number on error.  The possible
 551  * error cases are:
 552  *      - Match would exceed the amount of data remaining to be uncompressed.
 553  *      - Match refers to data before the window.
 554  *      - The input bitstream ended unexpectedly.
 555  */
 556 static int
 557 lzx_decode_match(unsigned main_element, int block_type,
 558                  unsigned bytes_remaining, u8 *window,
 559                  unsigned window_pos,
 560                  const struct lzx_tables *tables,
 561                  struct lzx_lru_queue *queue,
 562                  struct input_bitstream *istream)
 563 {
 564         unsigned length_header;
 565         unsigned position_slot;
 566         unsigned match_len;
 567         unsigned match_offset;
 568         unsigned additional_len;
 569         unsigned num_extra_bits;
 570         u32 verbatim_bits;
 571         u32 aligned_bits;
 572         unsigned i;
 573         int ret;
 574         u8 *match_dest;
 575         u8 *match_src;
 576
 577         /* The main element is offset by 256 because values under 256 indicate a
 578          * literal value. */
 579         main_element -= LZX_NUM_CHARS;
 580
 581         /* The length header consists of the lower 3 bits of the main element.
 582          * The position slot is the rest of it. */
 583         length_header = main_element & LZX_NUM_PRIMARY_LENS;
 584         position_slot = main_element >> 3;
 585
 586         /* If the length_header is less than LZX_NUM_PRIMARY_LENS (= 7), it
 587          * gives the match length as the offset from LZX_MIN_MATCH_LEN.
 588          * Otherwise, the length is given by an additional symbol encoded using
 589          * the length tree, offset by 9 (LZX_MIN_MATCH_LEN +
 590          * LZX_NUM_PRIMARY_LENS) */
 591         match_len = LZX_MIN_MATCH_LEN + length_header;
 592         if (length_header == LZX_NUM_PRIMARY_LENS) {
 593                 ret = read_huffsym_using_lentree(istream, tables,
 594                                                  &additional_len);
 595                 if (ret)
 596                         return ret;
 597                 match_len += additional_len;
 598         }
 599
 600
 601         /* If the position_slot is 0, 1, or 2, the match offset is retrieved
 602          * from the LRU queue.  Otherwise, the match offset is not in the LRU
 603          * queue. */
 604         switch (position_slot) {
 605         case 0:
 606                 match_offset = queue->R[0];
 607                 break;
 608         case 1:
 609                 match_offset = queue->R[1];
 610                 swap(queue->R[0], queue->R[1]);
 611                 break;
 612         case 2:
 613                 /* The queue doesn't work quite the same as a real LRU queue,
 614                  * since using the R2 offset doesn't bump the R1 offset down to
 615                  * R2. */
 616                 match_offset = queue->R[2];
 617                 swap(queue->R[0], queue->R[2]);
 618                 break;
 619         default:
 620                 /* Otherwise, the offset was not encoded as one the offsets in
 621                  * the queue.  Depending on the position slot, there is a
 622                  * certain number of extra bits that need to be read to fully
 623                  * decode the match offset. */
 624
 625                 /* Look up the number of extra bits that need to be read. */
 626                 num_extra_bits = lzx_get_num_extra_bits(position_slot);
 627
 628                 /* For aligned blocks, if there are at least 3 extra bits, the
 629                  * actual number of extra bits is 3 less, and they encode a
 630                  * number of 8-byte words that are added to the offset; there
 631                  * is then an additional symbol read using the aligned tree that
 632                  * specifies the actual byte alignment. */
 633                 if (block_type == LZX_BLOCKTYPE_ALIGNED && num_extra_bits >= 3) {
 634
 635                         /* There is an error in the LZX "specification" at this
 636                          * point; it indicates that a Huffman symbol is to be
 637                          * read only if num_extra_bits is greater than 3, but
 638                          * actually it is if num_extra_bits is greater than or
 639                          * equal to 3.  (Note that in the case with
 640                          * num_extra_bits == 3, the assignment to verbatim_bits
 641                          * will just set it to 0. ) */
 642                         ret = bitstream_read_bits(istream, num_extra_bits - 3,
 643                                                   &verbatim_bits);
 644                         if (ret)
 645                                 return ret;
 646
 647                         verbatim_bits <<= 3;
 648
 649                         ret = read_huffsym_using_alignedtree(istream, tables,
 650                                                              &aligned_bits);
 651                         if (ret)
 652                                 return ret;
 653                 } else {
 654                         /* For non-aligned blocks, or for aligned blocks with
 655                          * less than 3 extra bits, the extra bits are added
 656                          * directly to the match offset, and the correction for
 657                          * the alignment is taken to be 0. */
 658                         ret = bitstream_read_bits(istream, num_extra_bits,
 659                                                   &verbatim_bits);
 660                         if (ret)
 661                                 return ret;
 662
 663                         aligned_bits = 0;
 664                 }
 665
 666                 /* Calculate the match offset. */
 667                 match_offset = lzx_position_base[position_slot] +
 668                                verbatim_bits + aligned_bits - LZX_OFFSET_OFFSET;
 669
 670                 /* Update the LRU queue. */
 671                 queue->R[2] = queue->R[1];
 672                 queue->R[1] = queue->R[0];
 673                 queue->R[0] = match_offset;
 674                 break;
 675         }
 676
 677         /* Verify that the match is in the bounds of the part of the window
 678          * currently in use, then copy the source of the match to the current
 679          * position. */
 680
 681         if (match_len > bytes_remaining) {
 682                 LZX_DEBUG("Match of length %u bytes overflows "
 683                           "uncompressed block size", match_len);
 684                 return -1;
 685         }
 686
 687         if (match_offset > window_pos) {
 688                 LZX_DEBUG("Match of length %u bytes references "
 689                           "data before window (match_offset = %u, "
 690                           "window_pos = %u)",
 691                           match_len, match_offset, window_pos);
 692                 return -1;
 693         }
 694
 695         match_dest = window + window_pos;
 696         match_src = match_dest - match_offset;
 697
 698 #if 0
 699         printf("Match: src %u, dst %u, len %u\n", match_src - window,
 700                                                 match_dest - window,
 701                                                 match_len);
 702         putchar('|');
 703         for (i = 0; i < match_len; i++) {
 704                 match_dest[i] = match_src[i];
 705                 putchar(match_src[i]);
 706         }
 707         putchar('|');
 708         putchar('\n');
 709 #else
 710         for (i = 0; i < match_len; i++)
 711                 match_dest[i] = match_src[i];
 712 #endif
 713
 714         return match_len;
 715 }
 716
 717 static void
 718 undo_call_insn_translation(u32 *call_insn_target, s32 input_pos)
 719 {
 720         s32 abs_offset;
 721         s32 rel_offset;
 722
 723         abs_offset = le32_to_cpu(*call_insn_target);
 724         if (abs_offset >= -input_pos && abs_offset < LZX_WIM_MAGIC_FILESIZE) {
 725                 if (abs_offset >= 0) {
 726                         /* "good translation" */
 727                         rel_offset = abs_offset - input_pos;
 728                 } else {
 729                         /* "compensating translation" */
 730                         rel_offset = abs_offset + LZX_WIM_MAGIC_FILESIZE;
 731                 }
 732                 *call_insn_target = cpu_to_le32(rel_offset);
 733         }
 734 }
 735
 736 /* Undo the 'E8' preprocessing, where the targets of x86 CALL instructions were
 737  * changed from relative offsets to absolute offsets.
 738  *
 739  * Note that this call instruction preprocessing can and will be used on any
 740  * data even if it is not actually x86 machine code.  In fact, this type of
 741  * preprocessing appears to always be used in LZX-compressed resources in WIM
 742  * files; there is no bit to indicate whether it is used or not, unlike in the
 743  * LZX compressed format as used in cabinet files, where a bit is reserved for
 744  * that purpose.
 745  *
 746  * Call instruction preprocessing is disabled in the last 6 bytes of the
 747  * uncompressed data, which really means the 5-byte call instruction cannot
 748  * start in the last 10 bytes of the uncompressed data.  This is one of the
 749  * errors in the LZX documentation.
 750  *
 751  * Call instruction preprocessing does not appear to be disabled after the
 752  * 32768th chunk of a WIM stream, which is apparently is yet another difference
 753  * from the LZX compression used in cabinet files.
 754  *
 755  * Call instruction processing is supposed to take the file size as a parameter,
 756  * as it is used in calculating the translated jump targets.  But in WIM files,
 757  * this file size is always the same (LZX_WIM_MAGIC_FILESIZE == 12000000).*/
 758 static void
 759 undo_call_insn_preprocessing(u8 *uncompressed_data, size_t uncompressed_size)
 760 {
 761 #ifdef __SSE2__
 762
 763         /* SSE2 vectorized implementation for x86_64.  This speeds up LZX
 764          * decompression by about 5-8% overall.  (Usually --- the performance
 765          * actually regresses slightly in the degenerate case that the data
 766          * consists entirely of 0xe8 bytes.)  */
 767         __m128i *p128 = (__m128i *)uncompressed_data;
 768         u32 valid_mask = 0xFFFFFFFF;
 769
 770         if (uncompressed_size >= 32 &&
 771             ((uintptr_t)uncompressed_data % 16 == 0))
 772         {
 773                 __m128i * const end128 = p128 + uncompressed_size / 16 - 1;
 774
 775                 /* Create a vector of all 0xe8 bytes  */
 776                 const __m128i e8_bytes = _mm_set1_epi8(0xe8);
 777
 778                 /* Iterate through the 16-byte vectors in the input.  */
 779                 do {
 780                         /* Compare the current 16-byte vector with the vector of
 781                          * all 0xe8 bytes.  This produces 0xff where the byte is
 782                          * 0xe8 and 0x00 where it is not.  */
 783                         __m128i cmpresult = _mm_cmpeq_epi8(*p128, e8_bytes);
 784
 785                         /* Map the comparison results into a single 16-bit
 786                          * number.  It will contain a 1 bit when the
 787                          * corresponding byte in the current 16-byte vector is
 788                          * an e8 byte.  Note: the low-order bit corresponds to
 789                          * the first (lowest address) byte.  */
 790                         u32 e8_mask = _mm_movemask_epi8(cmpresult);
 791
 792                         if (!e8_mask) {
 793                                 /* If e8_mask is 0, then none of these 16 bytes
 794                                  * have value 0xe8.  No e8 translation is
 795                                  * needed, and there is no restriction that
 796                                  * carries over to the next 16 bytes.  */
 797                                 valid_mask = 0xFFFFFFFF;
 798                         } else {
 799                                 /* At least one byte has value 0xe8.
 800                                  *
 801                                  * The AND with valid_mask accounts for the fact
 802                                  * that we can't start an e8 translation that
 803                                  * overlaps the previous one.  */
 804                                 while ((e8_mask &= valid_mask)) {
 805
 806                                         /* Count the number of trailing zeroes
 807                                          * in e8_mask.  This will produce the
 808                                          * index of the byte, within the 16, at
 809                                          * which the next e8 translation should
 810                                          * be done.  */
 811                                         u32 bit = __builtin_ctz(e8_mask);
 812
 813                                         /* Do the e8 translation.  */
 814                                         u8 *p8 = (u8 *)p128 + bit;
 815                                         undo_call_insn_translation((s32 *)(p8 + 1),
 816                                                                    p8 - uncompressed_data);
 817
 818                                         /* Don't start an e8 translation in the
 819                                          * next 4 bytes.  */
 820                                         valid_mask &= ~((u32)0x1F << bit);
 821                                 }
 822                                 /* Moving on to the next vector.  Shift and set
 823                                  * valid_mask accordingly.  */
 824                                 valid_mask >>= 16;
 825                                 valid_mask |= 0xFFFF0000;
 826                         }
 827                 } while (++p128 < end128);
 828         }
 829
 830         u8 *p8 = (u8 *)p128;
 831         while (!(valid_mask & 1)) {
 832                 p8++;
 833                 valid_mask >>= 1;
 834         }
 835 #else /* __SSE2__  */
 836         u8 *p8 = uncompressed_data;
 837 #endif /* !__SSE2__  */
 838
 839         if (uncompressed_size > 10) {
 840                 /* Finish any bytes that weren't processed by the vectorized
 841                  * implementation.  */
 842                 u8 *p8_end = uncompressed_data + uncompressed_size - 10;
 843                 do {
 844                         if (*p8 == 0xe8) {
 845                                 undo_call_insn_translation((s32 *)(p8 + 1),
 846                                                            p8 - uncompressed_data);
 847                                 p8 += 5;
 848                         } else {
 849                                 p8++;
 850                         }
 851                 } while (p8 < p8_end);
 852         }
 853 }
 854
 855 /*
 856  * Decompresses an LZX-compressed block of data from which the header has already
 857  * been read.
 858  *
 859  * @block_type: The type of the block (LZX_BLOCKTYPE_VERBATIM or
 860  *              LZX_BLOCKTYPE_ALIGNED)
 861  * @block_size: The size of the block, in bytes.
 862  * @num_main_syms:      Number of symbols in the main alphabet.
 863  * @window:     Pointer to the decompression window.
 864  * @window_pos: The current position in the window.  Will be 0 for the first
 865  *                      block.
 866  * @tables:     The Huffman decoding tables for the block (main, length, and
 867  *                      aligned offset, the latter only for LZX_BLOCKTYPE_ALIGNED)
 868  * @queue:      The least-recently-used queue for match offsets.
 869  * @istream:    The input bitstream for the compressed literals.
 870  */
 871 static int
 872 lzx_decompress_block(int block_type, unsigned block_size,
 873                      unsigned num_main_syms,
 874                      u8 *window,
 875                      unsigned window_pos,
 876                      const struct lzx_tables *tables,
 877                      struct lzx_lru_queue *queue,
 878                      struct input_bitstream *istream)
 879 {
 880         unsigned main_element;
 881         unsigned end;
 882         int ret;
 883         int match_len;
 884
 885         end = window_pos + block_size;
 886         while (window_pos < end) {
 887                 ret = read_huffsym_using_maintree(istream, tables,
 888                                                   &main_element,
 889                                                   num_main_syms);
 890                 if (ret)
 891                         return ret;
 892
 893                 if (main_element < LZX_NUM_CHARS) {
 894                         /* literal: 0 to LZX_NUM_CHARS - 1 */
 895                         window[window_pos++] = main_element;
 896                 } else {
 897                         /* match: LZX_NUM_CHARS to num_main_syms - 1 */
 898                         match_len = lzx_decode_match(main_element,
 899                                                      block_type,
 900                                                      end - window_pos,
 901                                                      window,
 902                                                      window_pos,
 903                                                      tables,
 904                                                      queue,
 905                                                      istream);
 906                         if (match_len < 0)
 907                                 return match_len;
 908                         window_pos += match_len;
 909                 }
 910         }
 911         return 0;
 912 }
 913
 914 static int
 915 lzx_decompress(const void *compressed_data, size_t compressed_size,
 916                void *uncompressed_data, size_t uncompressed_size,
 917                void *_ctx)
 918 {
 919         struct lzx_decompressor *ctx = _ctx;
 920         struct input_bitstream istream;
 921         struct lzx_lru_queue queue;
 922         unsigned window_pos;
 923         unsigned block_size;
 924         unsigned block_type;
 925         int ret;
 926         bool e8_preprocessing_done;
 927
 928         LZX_DEBUG("compressed_data = %p, compressed_size = %zu, "
 929                   "uncompressed_data = %p, uncompressed_size = %zu, "
 930                   "max_window_size=%u).",
 931                   compressed_data, compressed_size,
 932                   uncompressed_data, uncompressed_size,
 933                   ctx->max_window_size);
 934
 935         if (uncompressed_size > ctx->max_window_size) {
 936                 LZX_DEBUG("Uncompressed size of %zu exceeds "
 937                           "window size of %u!",
 938                           uncompressed_size, ctx->max_window_size);
 939                 return -1;
 940         }
 941
 942         memset(ctx->tables.maintree_lens, 0, sizeof(ctx->tables.maintree_lens));
 943         memset(ctx->tables.lentree_lens, 0, sizeof(ctx->tables.lentree_lens));
 944         lzx_lru_queue_init(&queue);
 945         init_input_bitstream(&istream, compressed_data, compressed_size);
 946
 947         e8_preprocessing_done = false; /* Set to true if there may be 0xe8 bytes
 948                                           in the uncompressed data. */
 949
 950         /* The compressed data will consist of one or more blocks.  The
 951          * following loop decompresses one block, and it runs until there all
 952          * the compressed data has been decompressed, so there are no more
 953          * blocks.  */
 954
 955         for (window_pos = 0;
 956              window_pos < uncompressed_size;
 957              window_pos += block_size)
 958         {
 959                 LZX_DEBUG("Reading block header.");
 960                 ret = lzx_read_block_header(&istream, ctx->num_main_syms,
 961                                             ctx->max_window_size, &block_size,
 962                                             &block_type, &ctx->tables, &queue);
 963                 if (ret)
 964                         return ret;
 965
 966                 LZX_DEBUG("block_size = %u, window_pos = %u",
 967                           block_size, window_pos);
 968
 969                 if (block_size > uncompressed_size - window_pos) {
 970                         LZX_DEBUG("Expected a block size of at "
 971                                   "most %zu bytes (found %u bytes)",
 972                                   uncompressed_size - window_pos, block_size);
 973                         return -1;
 974                 }
 975
 976                 switch (block_type) {
 977                 case LZX_BLOCKTYPE_VERBATIM:
 978                 case LZX_BLOCKTYPE_ALIGNED:
 979                         if (block_type == LZX_BLOCKTYPE_VERBATIM)
 980                                 LZX_DEBUG("LZX_BLOCKTYPE_VERBATIM");
 981                         else
 982                                 LZX_DEBUG("LZX_BLOCKTYPE_ALIGNED");
 983                         ret = lzx_decompress_block(block_type,
 984                                                    block_size,
 985                                                    ctx->num_main_syms,
 986                                                    uncompressed_data,
 987                                                    window_pos,
 988                                                    &ctx->tables,
 989                                                    &queue,
 990                                                    &istream);
 991                         if (ret)
 992                                 return ret;
 993
 994                         if (ctx->tables.maintree_lens[0xe8] != 0)
 995                                 e8_preprocessing_done = true;
 996                         break;
 997                 case LZX_BLOCKTYPE_UNCOMPRESSED:
 998                         LZX_DEBUG("LZX_BLOCKTYPE_UNCOMPRESSED");
 999                         if (istream.data_bytes_left < block_size) {
1000                                 LZX_DEBUG("Unexpected end of input when "
1001                                           "reading %u bytes from LZX bitstream "
1002                                           "(only have %u bytes left)",
1003                                           block_size, istream.data_bytes_left);
1004                                 return -1;
1005                         }
1006                         memcpy(&((u8*)uncompressed_data)[window_pos], istream.data,
1007                                block_size);
1008                         istream.data += block_size;
1009                         istream.data_bytes_left -= block_size;
1010                         /* Re-align bitstream if an odd number of bytes were
1011                          * read.  */
1012                         if (istream.data_bytes_left && (block_size & 1)) {
1013                                 istream.data_bytes_left--;
1014                                 istream.data++;
1015                         }
1016                         e8_preprocessing_done = true;
1017                         break;
1018                 }
1019         }
1020         if (e8_preprocessing_done)
1021                 undo_call_insn_preprocessing(uncompressed_data, uncompressed_size);
1022         return 0;
1023 }
1024
1025 static void
1026 lzx_free_decompressor(void *_ctx)
1027 {
1028         struct lzx_decompressor *ctx = _ctx;
1029
1030         FREE(ctx);
1031 }
1032
1033 static int
1034 lzx_create_decompressor(size_t max_window_size,
1035                         const struct wimlib_decompressor_params_header *params,
1036                         void **ctx_ret)
1037 {
1038         struct lzx_decompressor *ctx;
1039
1040         if (!lzx_window_size_valid(max_window_size))
1041                 return WIMLIB_ERR_INVALID_PARAM;
1042
1043         ctx = MALLOC(sizeof(struct lzx_decompressor));
1044         if (ctx == NULL)
1045                 return WIMLIB_ERR_NOMEM;
1046
1047         ctx->max_window_size = max_window_size;
1048         ctx->num_main_syms = lzx_get_num_main_syms(max_window_size);
1049
1050         *ctx_ret = ctx;
1051         return 0;
1052 }
1053
1054 const struct decompressor_ops lzx_decompressor_ops = {
1055         .create_decompressor = lzx_create_decompressor,
1056         .decompress          = lzx_decompress,
1057         .free_decompressor   = lzx_free_decompressor,
1058 };