X-Git-Url: https://wimlib.net/git/?a=blobdiff_plain;f=src%2Flzx-decompress.c;h=d259eeedc38ae31b6c7156fc0b1599f3b2af8d90;hb=500874ebbf2839d2c93db2b4094817c219946f29;hp=a2a60889db7fc4c0a8f44292ac6bd84419bc0850;hpb=67c5f2c5954d07c25f57b6d9aedb73fb3720ec0b;p=wimlib diff --git a/src/lzx-decompress.c b/src/lzx-decompress.c index a2a60889..d259eeed 100644 --- a/src/lzx-decompress.c +++ b/src/lzx-decompress.c @@ -115,10 +115,6 @@ #include -#ifdef __SSE2__ -# include -#endif - /* Huffman decoding tables and maps from symbols to code lengths. */ struct lzx_tables { @@ -154,18 +150,15 @@ read_huffsym_using_pretree(struct input_bitstream *istream, const u16 pretree_decode_table[]) { return read_huffsym(istream, pretree_decode_table, - LZX_PRECODE_NUM_SYMBOLS, LZX_PRECODE_TABLEBITS, - LZX_MAX_PRE_CODEWORD_LEN); + LZX_PRECODE_TABLEBITS, LZX_MAX_PRE_CODEWORD_LEN); } /* Reads a Huffman-encoded symbol using the main tree. */ static inline u16 read_huffsym_using_maintree(struct input_bitstream *istream, - const struct lzx_tables *tables, - unsigned num_main_syms) + const struct lzx_tables *tables) { return read_huffsym(istream, tables->maintree_decode_table, - num_main_syms, LZX_MAINCODE_TABLEBITS, LZX_MAX_MAIN_CODEWORD_LEN); } @@ -175,7 +168,6 @@ read_huffsym_using_lentree(struct input_bitstream *istream, const struct lzx_tables *tables) { return read_huffsym(istream, tables->lentree_decode_table, - LZX_LENCODE_NUM_SYMBOLS, LZX_LENCODE_TABLEBITS, LZX_MAX_LEN_CODEWORD_LEN); } @@ -185,9 +177,7 @@ read_huffsym_using_alignedtree(struct input_bitstream *istream, const struct lzx_tables *tables) { return read_huffsym(istream, tables->alignedtree_decode_table, - LZX_ALIGNEDCODE_NUM_SYMBOLS, - LZX_ALIGNEDCODE_TABLEBITS, - LZX_MAX_ALIGNED_CODEWORD_LEN); + LZX_ALIGNEDCODE_TABLEBITS, LZX_MAX_ALIGNED_CODEWORD_LEN); } /* @@ -558,22 +548,15 @@ lzx_decode_match(unsigned main_element, int block_type, /* If the position_slot is 0, 1, or 2, the match offset is retrieved * from the LRU queue. Otherwise, the match offset is not in the LRU * queue. */ - switch (position_slot) { - case 0: - match_offset = queue->R[0]; - break; - case 1: - match_offset = queue->R[1]; - swap(queue->R[0], queue->R[1]); - break; - case 2: - /* The queue doesn't work quite the same as a real LRU queue, - * since using the R2 offset doesn't bump the R1 offset down to - * R2. */ - match_offset = queue->R[2]; - swap(queue->R[0], queue->R[2]); - break; - default: + if (position_slot <= 2) { + /* Note: This isn't a real LRU queue, since using the R2 offset + * doesn't bump the R1 offset down to R2. This quirk allows all + * 3 recent offsets to be handled by the same code. (For R0, + * the swap is a no-op.) */ + match_offset = queue->R[position_slot]; + queue->R[position_slot] = queue->R[0]; + queue->R[0] = match_offset; + } else { /* Otherwise, the offset was not encoded as one the offsets in * the queue. Depending on the position slot, there is a * certain number of extra bits that need to be read to fully @@ -618,7 +601,6 @@ lzx_decode_match(unsigned main_element, int block_type, queue->R[2] = queue->R[1]; queue->R[1] = queue->R[0]; queue->R[0] = match_offset; - break; } /* Verify that the match is in the bounds of the part of the window @@ -661,149 +643,6 @@ lzx_decode_match(unsigned main_element, int block_type, return match_len; } -static void -undo_call_insn_translation(u32 *call_insn_target, s32 input_pos) -{ - s32 abs_offset; - s32 rel_offset; - - abs_offset = le32_to_cpu(*call_insn_target); - if (abs_offset >= 0) { - if (abs_offset < LZX_WIM_MAGIC_FILESIZE) { - /* "good translation" */ - rel_offset = abs_offset - input_pos; - - *call_insn_target = cpu_to_le32(rel_offset); - } - } else { - if (abs_offset >= -input_pos) { - /* "compensating translation" */ - rel_offset = abs_offset + LZX_WIM_MAGIC_FILESIZE; - - *call_insn_target = cpu_to_le32(rel_offset); - } - } -} - -/* Undo the 'E8' preprocessing, where the targets of x86 CALL instructions were - * changed from relative offsets to absolute offsets. - * - * Note that this call instruction preprocessing can and will be used on any - * data even if it is not actually x86 machine code. In fact, this type of - * preprocessing appears to always be used in LZX-compressed resources in WIM - * files; there is no bit to indicate whether it is used or not, unlike in the - * LZX compressed format as used in cabinet files, where a bit is reserved for - * that purpose. - * - * Call instruction preprocessing is disabled in the last 6 bytes of the - * uncompressed data, which really means the 5-byte call instruction cannot - * start in the last 10 bytes of the uncompressed data. This is one of the - * errors in the LZX documentation. - * - * Call instruction preprocessing does not appear to be disabled after the - * 32768th chunk of a WIM stream, which is apparently is yet another difference - * from the LZX compression used in cabinet files. - * - * Call instruction processing is supposed to take the file size as a parameter, - * as it is used in calculating the translated jump targets. But in WIM files, - * this file size is always the same (LZX_WIM_MAGIC_FILESIZE == 12000000).*/ -static void -undo_call_insn_preprocessing(u8 *uncompressed_data, size_t uncompressed_size) -{ -#ifdef __SSE2__ - - /* SSE2 vectorized implementation for x86_64. This speeds up LZX - * decompression by about 5-8% overall. (Usually --- the performance - * actually regresses slightly in the degenerate case that the data - * consists entirely of 0xe8 bytes.) */ - __m128i *p128 = (__m128i *)uncompressed_data; - u32 valid_mask = 0xFFFFFFFF; - - if (uncompressed_size >= 32 && - ((uintptr_t)uncompressed_data % 16 == 0)) - { - __m128i * const end128 = p128 + uncompressed_size / 16 - 1; - - /* Create a vector of all 0xe8 bytes */ - const __m128i e8_bytes = _mm_set1_epi8(0xe8); - - /* Iterate through the 16-byte vectors in the input. */ - do { - /* Compare the current 16-byte vector with the vector of - * all 0xe8 bytes. This produces 0xff where the byte is - * 0xe8 and 0x00 where it is not. */ - __m128i cmpresult = _mm_cmpeq_epi8(*p128, e8_bytes); - - /* Map the comparison results into a single 16-bit - * number. It will contain a 1 bit when the - * corresponding byte in the current 16-byte vector is - * an e8 byte. Note: the low-order bit corresponds to - * the first (lowest address) byte. */ - u32 e8_mask = _mm_movemask_epi8(cmpresult); - - if (!e8_mask) { - /* If e8_mask is 0, then none of these 16 bytes - * have value 0xe8. No e8 translation is - * needed, and there is no restriction that - * carries over to the next 16 bytes. */ - valid_mask = 0xFFFFFFFF; - } else { - /* At least one byte has value 0xe8. - * - * The AND with valid_mask accounts for the fact - * that we can't start an e8 translation that - * overlaps the previous one. */ - while ((e8_mask &= valid_mask)) { - - /* Count the number of trailing zeroes - * in e8_mask. This will produce the - * index of the byte, within the 16, at - * which the next e8 translation should - * be done. */ - u32 bit = __builtin_ctz(e8_mask); - - /* Do the e8 translation. */ - u8 *p8 = (u8 *)p128 + bit; - undo_call_insn_translation((s32 *)(p8 + 1), - p8 - uncompressed_data); - - /* Don't start an e8 translation in the - * next 4 bytes. */ - valid_mask &= ~((u32)0x1F << bit); - } - /* Moving on to the next vector. Shift and set - * valid_mask accordingly. */ - valid_mask >>= 16; - valid_mask |= 0xFFFF0000; - } - } while (++p128 < end128); - } - - u8 *p8 = (u8 *)p128; - while (!(valid_mask & 1)) { - p8++; - valid_mask >>= 1; - } -#else /* __SSE2__ */ - u8 *p8 = uncompressed_data; -#endif /* !__SSE2__ */ - - if (uncompressed_size > 10) { - /* Finish any bytes that weren't processed by the vectorized - * implementation. */ - u8 *p8_end = uncompressed_data + uncompressed_size - 10; - do { - if (*p8 == 0xe8) { - undo_call_insn_translation((s32 *)(p8 + 1), - p8 - uncompressed_data); - p8 += 5; - } else { - p8++; - } - } while (p8 < p8_end); - } -} - /* * Decompresses an LZX-compressed block of data from which the header has already * been read. @@ -811,7 +650,6 @@ undo_call_insn_preprocessing(u8 *uncompressed_data, size_t uncompressed_size) * @block_type: The type of the block (LZX_BLOCKTYPE_VERBATIM or * LZX_BLOCKTYPE_ALIGNED) * @block_size: The size of the block, in bytes. - * @num_main_syms: Number of symbols in the main alphabet. * @window: Pointer to the decompression window. * @window_pos: The current position in the window. Will be 0 for the first * block. @@ -822,7 +660,6 @@ undo_call_insn_preprocessing(u8 *uncompressed_data, size_t uncompressed_size) */ static int lzx_decompress_block(int block_type, unsigned block_size, - unsigned num_main_syms, u8 *window, unsigned window_pos, const struct lzx_tables *tables, @@ -835,8 +672,7 @@ lzx_decompress_block(int block_type, unsigned block_size, end = window_pos + block_size; while (window_pos < end) { - main_element = read_huffsym_using_maintree(istream, tables, - num_main_syms); + main_element = read_huffsym_using_maintree(istream, tables); if (main_element < LZX_NUM_CHARS) { /* literal: 0 to LZX_NUM_CHARS - 1 */ window[window_pos++] = main_element; @@ -929,7 +765,6 @@ lzx_decompress(const void *compressed_data, size_t compressed_size, LZX_DEBUG("LZX_BLOCKTYPE_ALIGNED"); ret = lzx_decompress_block(block_type, block_size, - ctx->num_main_syms, uncompressed_data, window_pos, &ctx->tables, @@ -965,7 +800,7 @@ lzx_decompress(const void *compressed_data, size_t compressed_size, } } if (e8_preprocessing_done) - undo_call_insn_preprocessing(uncompressed_data, uncompressed_size); + lzx_undo_e8_preprocessing(uncompressed_data, uncompressed_size); return 0; } @@ -974,7 +809,7 @@ lzx_free_decompressor(void *_ctx) { struct lzx_decompressor *ctx = _ctx; - FREE(ctx); + ALIGNED_FREE(ctx); } static int @@ -987,7 +822,8 @@ lzx_create_decompressor(size_t max_window_size, if (!lzx_window_size_valid(max_window_size)) return WIMLIB_ERR_INVALID_PARAM; - ctx = MALLOC(sizeof(struct lzx_decompressor)); + ctx = ALIGNED_MALLOC(sizeof(struct lzx_decompressor), + DECODE_TABLE_ALIGNMENT); if (ctx == NULL) return WIMLIB_ERR_NOMEM;