X-Git-Url: https://wimlib.net/git/?p=wimlib;a=blobdiff_plain;f=src%2Flzx-decompress.c;h=d3ba46af846de85ec23f9cd35743fddfda05e2a7;hp=ec8ccf25159fdc83fee6e85ddce47d64166145fe;hb=869ca1f4cc4654501d52f8c3e660c056904ccabf;hpb=35ff480326a41f7f74fe0a497c636a811922f276 diff --git a/src/lzx-decompress.c b/src/lzx-decompress.c index ec8ccf25..d3ba46af 100644 --- a/src/lzx-decompress.c +++ b/src/lzx-decompress.c @@ -7,7 +7,7 @@ */ /* - * Copyright (C) 2012, 2013 Eric Biggers + * Copyright (C) 2012, 2013, 2014 Eric Biggers * * This file is part of wimlib, a library for working with WIM files. * @@ -26,33 +26,30 @@ */ /* - * LZX is a LZ77 and Huffman-code based compression format that has many + * LZX is an LZ77 and Huffman-code based compression format that has many * similarities to the DEFLATE format used in zlib. The compression ratio is as - * good or better than DEFLATE. However, in WIM files only up to 32768 bytes of - * data can ever compressed be in the same LZX block, so a .tar.gz file could - * potentially be smaller than a WIM file that uses LZX compression because it - * can use a larger LZ77 window size. + * good or better than DEFLATE. * * Some notes on the LZX compression format as used in Windows Imaging (WIM) * files: * * A compressed WIM resource consists of a table of chunk offsets followed by * the compressed chunks themselves. All compressed chunks except possibly the - * last decompress to 32768 bytes. This is quite similar to the cabinet (.cab) - * file format, but they are not the same. According to the cabinet format - * documentation, the LZX block size is independent from the CFDATA blocks, and - * a LZX block may span several CFDATA blocks. However, in WIMs, LZX blocks do - * not appear to ever span multiple WIM chunks. Note that this means any WIM - * chunk may be decompressed or compressed independently from any other chunk, - * which is convenient. - * - * A LZX compressed WIM chunk contains one or more LZX blocks of the aligned, + * last decompress to a fixed number of bytes, by default 32768. This is quite + * similar to the cabinet (.cab) file format, but they are not the same. + * According to the cabinet format documentation, the LZX block size is + * independent from the CFDATA blocks, and an LZX block may span several CFDATA + * blocks. However, in WIMs, LZX blocks do not appear to ever span multiple WIM + * chunks. Note that this means any WIM chunk may be decompressed or compressed + * independently from any other chunk, which allows random access. + * + * An LZX compressed WIM chunk contains one or more LZX blocks of the aligned, * verbatim, or uncompressed block types. For aligned and verbatim blocks, the * size of the block in uncompressed bytes is specified by a bit following the 3 * bits that specify the block type, possibly followed by an additional 16 bits. - * '1' means to use the default block size (equal to 32768, the size of a WIM - * chunk--- and this seems to only be valid for the first LZX block in a WIM - * chunk), while '0' means that the block size is provided by the next 16 bits. + * '1' means to use the default block size (equal to 32768, the default size of + * a WIM chunk), while '0' means that the block size is provided by the next 16 + * bits. * * The cabinet format, as documented, allows for the possibility that a * compressed CFDATA chunk is up to 6144 bytes larger than the data it @@ -85,13 +82,13 @@ * defined in the specification. * * The LZX document states that aligned offset blocks have their aligned offset - * huffman tree AFTER the main and length trees. The implementation suggests + * Huffman tree AFTER the main and length trees. The implementation suggests * that the aligned offset tree is BEFORE the main and length trees. * * The LZX document decoding algorithm states that, in an aligned offset block, * if an extra_bits value is 1, 2 or 3, then that number of bits should be read * and the result added to the match offset. This is correct for 1 and 2, but - * not 3, where just a huffman symbol (using the aligned tree) should be read. + * not 3, where just a Huffman symbol (using the aligned tree) should be read. * * Regarding the E8 preprocessing, the LZX document states 'No translation may * be performed on the last 6 bytes of the input block'. This is correct. @@ -100,7 +97,7 @@ * would cause the next four bytes to be modified, at least one of which would * be in the last 6 bytes, which is not allowed according to the spec. * - * The specification states that the huffman trees must always contain at least + * The specification states that the Huffman trees must always contain at least * one element. However, many CAB files contain blocks where the length tree is * completely empty (because there are no matches), and this is expected to * succeed. @@ -111,12 +108,17 @@ #endif #include "wimlib.h" -#include "wimlib/decompress.h" +#include "wimlib/decompressor_ops.h" +#include "wimlib/decompress_common.h" #include "wimlib/lzx.h" #include "wimlib/util.h" #include +#ifdef __SSE2__ +# include +#endif + /* Huffman decoding tables and maps from symbols to code lengths. */ struct lzx_tables { @@ -138,6 +140,11 @@ struct lzx_tables { u8 alignedtree_lens[LZX_ALIGNEDCODE_NUM_SYMBOLS]; } _aligned_attribute(DECODE_TABLE_ALIGNMENT); +struct lzx_decompressor { + u32 max_window_size; + unsigned num_main_syms; + struct lzx_tables tables; +}; /* * Reads a Huffman-encoded symbol using the pre-tree. @@ -193,10 +200,10 @@ read_huffsym_using_alignedtree(struct input_bitstream *istream, * code length values from the input. * * @istream: The bit stream for the input. It is positioned on the beginning - * of the pretree for the code length values. + * of the pretree for the code length values. * @lens: An array that contains the length values from the previous time - * the code lengths for this Huffman tree were read, or all - * 0's if this is the first time. + * the code lengths for this Huffman tree were read, or all + * 0's if this is the first time. * @num_lens: Number of length values to decode and return. * */ @@ -313,15 +320,15 @@ lzx_read_code_lens(struct input_bitstream *istream, u8 lens[], * * @istream: The input bitstream. * @block_size_ret: A pointer to an int into which the size of the block, - * in bytes, will be returned. + * in bytes, will be returned. * @block_type_ret: A pointer to an int into which the type of the block - * (LZX_BLOCKTYPE_*) will be returned. - * @tables: A pointer to a lzx_tables structure in which the - * main tree, the length tree, and possibly the - * aligned offset tree will be constructed. + * (LZX_BLOCKTYPE_*) will be returned. + * @tables: A pointer to an lzx_tables structure in which the + * main tree, the length tree, and possibly the + * aligned offset tree will be constructed. * @queue: A pointer to the least-recently-used queue into which - * R0, R1, and R2 will be written (only for uncompressed - * blocks, which contain this information in the header) + * R0, R1, and R2 will be written (only for uncompressed + * blocks, which contain this information in the header) */ static int lzx_read_block_header(struct input_bitstream *istream, @@ -494,9 +501,9 @@ lzx_read_block_header(struct input_bitstream *istream, istream->bitsleft = 0; istream->bitbuf = 0; } - queue->R[0] = le32_to_cpu(*(u32*)(istream->data + 0)); - queue->R[1] = le32_to_cpu(*(u32*)(istream->data + 4)); - queue->R[2] = le32_to_cpu(*(u32*)(istream->data + 8)); + queue->R[0] = le32_to_cpu(*(le32*)(istream->data + 0)); + queue->R[1] = le32_to_cpu(*(le32*)(istream->data + 4)); + queue->R[2] = le32_to_cpu(*(le32*)(istream->data + 8)); istream->data += 12; istream->data_bytes_left -= 12; /* The uncompressed data of this block directly follows and will @@ -521,14 +528,14 @@ lzx_read_block_header(struct input_bitstream *istream, * tree. * * @block_type: The type of the block (LZX_BLOCKTYPE_ALIGNED or - * LZX_BLOCKTYPE_VERBATIM) + * LZX_BLOCKTYPE_VERBATIM) * * @bytes_remaining: The amount of uncompressed data remaining to be - * uncompressed in this block. It is an error if the match - * is longer than this number. + * uncompressed in this block. It is an error if the match + * is longer than this number. * * @window: A pointer to the window into which the uncompressed - * data is being written. + * data is being written. * * @window_pos: The current byte offset in the window. * @@ -542,9 +549,9 @@ lzx_read_block_header(struct input_bitstream *istream, * * Returns the length of the match, or a negative number on error. The possible * error cases are: - * - Match would exceed the amount of data remaining to be uncompressed. - * - Match refers to data before the window. - * - The input bitstream ended unexpectedly. + * - Match would exceed the amount of data remaining to be uncompressed. + * - Match refers to data before the window. + * - The input bitstream ended unexpectedly. */ static int lzx_decode_match(unsigned main_element, int block_type, @@ -708,20 +715,19 @@ lzx_decode_match(unsigned main_element, int block_type, } static void -undo_call_insn_translation(u32 *call_insn_target, int input_pos, - s32 file_size) +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 >= -input_pos && abs_offset < file_size) { + if (abs_offset >= -input_pos && abs_offset < LZX_WIM_MAGIC_FILESIZE) { if (abs_offset >= 0) { /* "good translation" */ rel_offset = abs_offset - input_pos; } else { /* "compensating translation" */ - rel_offset = abs_offset + file_size; + rel_offset = abs_offset + LZX_WIM_MAGIC_FILESIZE; } *call_insn_target = cpu_to_le32(rel_offset); } @@ -750,31 +756,115 @@ undo_call_insn_translation(u32 *call_insn_target, int input_pos, * 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[], int uncompressed_data_len) +undo_call_insn_preprocessing(u8 *uncompressed_data, size_t uncompressed_size) { - for (int i = 0; i < uncompressed_data_len - 10; i++) { - if (uncompressed_data[i] == 0xe8) { - undo_call_insn_translation((u32*)&uncompressed_data[i + 1], - i, - LZX_WIM_MAGIC_FILESIZE); - i += 4; - } +#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 a LZX-compressed block of data from which the header has already + * Decompresses an LZX-compressed block of data from which the header has already * been read. * * @block_type: The type of the block (LZX_BLOCKTYPE_VERBATIM or - * LZX_BLOCKTYPE_ALIGNED) + * 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. + * block. * @tables: The Huffman decoding tables for the block (main, length, and - * aligned offset, the latter only for LZX_BLOCKTYPE_ALIGNED) + * aligned offset, the latter only for LZX_BLOCKTYPE_ALIGNED) * @queue: The least-recently-used queue for match offsets. * @istream: The input bitstream for the compressed literals. */ @@ -821,46 +911,38 @@ lzx_decompress_block(int block_type, unsigned block_size, return 0; } -WIMLIBAPI int -wimlib_lzx_decompress2(const void *compressed_data, unsigned compressed_len, - void *uncompressed_data, unsigned uncompressed_len, - u32 max_window_size) +static int +lzx_decompress(const void *compressed_data, size_t compressed_size, + void *uncompressed_data, size_t uncompressed_size, + void *_ctx) { - struct lzx_tables tables; + struct lzx_decompressor *ctx = _ctx; struct input_bitstream istream; struct lzx_lru_queue queue; unsigned window_pos; unsigned block_size; unsigned block_type; - unsigned num_main_syms; int ret; bool e8_preprocessing_done; - LZX_DEBUG("compressed_data = %p, compressed_len = %u, " - "uncompressed_data = %p, uncompressed_len = %u, " + LZX_DEBUG("compressed_data = %p, compressed_size = %zu, " + "uncompressed_data = %p, uncompressed_size = %zu, " "max_window_size=%u).", - compressed_data, compressed_len, - uncompressed_data, uncompressed_len, max_window_size); + compressed_data, compressed_size, + uncompressed_data, uncompressed_size, + ctx->max_window_size); - if (!lzx_window_size_valid(max_window_size)) { - LZX_DEBUG("Window size of %u is invalid!", - max_window_size); - return -1; - } - - num_main_syms = lzx_get_num_main_syms(max_window_size); - - if (uncompressed_len > max_window_size) { - LZX_DEBUG("Uncompressed chunk size of %u exceeds " + if (uncompressed_size > ctx->max_window_size) { + LZX_DEBUG("Uncompressed size of %zu exceeds " "window size of %u!", - uncompressed_len, max_window_size); + uncompressed_size, ctx->max_window_size); return -1; } - memset(tables.maintree_lens, 0, sizeof(tables.maintree_lens)); - memset(tables.lentree_lens, 0, sizeof(tables.lentree_lens)); + memset(ctx->tables.maintree_lens, 0, sizeof(ctx->tables.maintree_lens)); + memset(ctx->tables.lentree_lens, 0, sizeof(ctx->tables.lentree_lens)); lzx_lru_queue_init(&queue); - init_input_bitstream(&istream, compressed_data, compressed_len); + init_input_bitstream(&istream, compressed_data, compressed_size); e8_preprocessing_done = false; /* Set to true if there may be 0xe8 bytes in the uncompressed data. */ @@ -871,23 +953,23 @@ wimlib_lzx_decompress2(const void *compressed_data, unsigned compressed_len, * blocks. */ for (window_pos = 0; - window_pos < uncompressed_len; + window_pos < uncompressed_size; window_pos += block_size) { LZX_DEBUG("Reading block header."); - ret = lzx_read_block_header(&istream, num_main_syms, - max_window_size, &block_size, - &block_type, &tables, &queue); + ret = lzx_read_block_header(&istream, ctx->num_main_syms, + ctx->max_window_size, &block_size, + &block_type, &ctx->tables, &queue); if (ret) return ret; LZX_DEBUG("block_size = %u, window_pos = %u", block_size, window_pos); - if (block_size > uncompressed_len - window_pos) { + if (block_size > uncompressed_size - window_pos) { LZX_DEBUG("Expected a block size of at " - "most %u bytes (found %u bytes)", - uncompressed_len - window_pos, block_size); + "most %zu bytes (found %u bytes)", + uncompressed_size - window_pos, block_size); return -1; } @@ -900,16 +982,16 @@ wimlib_lzx_decompress2(const void *compressed_data, unsigned compressed_len, LZX_DEBUG("LZX_BLOCKTYPE_ALIGNED"); ret = lzx_decompress_block(block_type, block_size, - num_main_syms, + ctx->num_main_syms, uncompressed_data, window_pos, - &tables, + &ctx->tables, &queue, &istream); if (ret) return ret; - if (tables.maintree_lens[0xe8] != 0) + if (ctx->tables.maintree_lens[0xe8] != 0) e8_preprocessing_done = true; break; case LZX_BLOCKTYPE_UNCOMPRESSED: @@ -936,16 +1018,41 @@ wimlib_lzx_decompress2(const void *compressed_data, unsigned compressed_len, } } if (e8_preprocessing_done) - undo_call_insn_preprocessing(uncompressed_data, uncompressed_len); + undo_call_insn_preprocessing(uncompressed_data, uncompressed_size); return 0; } -/* API function documented in wimlib.h */ -WIMLIBAPI int -wimlib_lzx_decompress(const void *compressed_data, unsigned compressed_len, - void *uncompressed_data, unsigned uncompressed_len) +static void +lzx_free_decompressor(void *_ctx) +{ + struct lzx_decompressor *ctx = _ctx; + + FREE(ctx); +} + +static int +lzx_create_decompressor(size_t max_window_size, + const struct wimlib_decompressor_params_header *params, + void **ctx_ret) { - return wimlib_lzx_decompress2(compressed_data, compressed_len, - uncompressed_data, uncompressed_len, - 32768); + struct lzx_decompressor *ctx; + + if (!lzx_window_size_valid(max_window_size)) + return WIMLIB_ERR_INVALID_PARAM; + + ctx = MALLOC(sizeof(struct lzx_decompressor)); + if (ctx == NULL) + return WIMLIB_ERR_NOMEM; + + ctx->max_window_size = max_window_size; + ctx->num_main_syms = lzx_get_num_main_syms(max_window_size); + + *ctx_ret = ctx; + return 0; } + +const struct decompressor_ops lzx_decompressor_ops = { + .create_decompressor = lzx_create_decompressor, + .decompress = lzx_decompress, + .free_decompressor = lzx_free_decompressor, +};