/*
- * lzx-common.c - Common data for LZX compression and decompression.
+ * lzx-common.c - Common code for LZX compression and decompression.
*/
/*
- * 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.
+ * This file is free software; you can redistribute it and/or modify it under
+ * the terms of the GNU Lesser General Public License as published by the Free
+ * Software Foundation; either version 3 of the License, or (at your option) any
+ * later version.
*
- * wimlib is free software; you can redistribute it and/or modify it under the
- * terms of the GNU General Public License as published by the Free
- * Software Foundation; either version 3 of the License, or (at your option)
- * any later version.
- *
- * wimlib is distributed in the hope that it will be useful, but WITHOUT ANY
- * WARRANTY; without even the implied warranty of MERCHANTABILITY or FITNESS FOR
- * A PARTICULAR PURPOSE. See the GNU General Public License for more
+ * This file is distributed in the hope that it will be useful, but WITHOUT
+ * ANY WARRANTY; without even the implied warranty of MERCHANTABILITY or FITNESS
+ * FOR A PARTICULAR PURPOSE. See the GNU Lesser General Public License for more
* details.
*
- * You should have received a copy of the GNU General Public License
- * along with wimlib; if not, see http://www.gnu.org/licenses/.
+ * You should have received a copy of the GNU Lesser General Public License
+ * along with this file; if not, see http://www.gnu.org/licenses/.
*/
#ifdef HAVE_CONFIG_H
# include "config.h"
#endif
+#include <string.h>
+
+#include "wimlib/bitops.h"
+#include "wimlib/endianness.h"
#include "wimlib/lzx.h"
+#include "wimlib/unaligned.h"
#include "wimlib/util.h"
-/* LZX uses what it calls 'position slots' to represent match offsets.
- * What this means is that a small 'position slot' number and a small
- * offset from that slot are encoded instead of one large offset for
- * every match.
- * - lzx_position_base is an index to the position slot bases
- * - lzx_extra_bits states how many bits of offset-from-base data is needed.
- */
+#ifdef __SSE2__
+# include <emmintrin.h>
+#endif
-const u32 lzx_position_base[LZX_MAX_POSITION_SLOTS] = {
+/* Mapping: offset slot => first match offset that uses that offset slot.
+ */
+const u32 lzx_offset_slot_base[LZX_MAX_OFFSET_SLOTS] = {
0 , 1 , 2 , 3 , 4 , /* 0 --- 4 */
6 , 8 , 12 , 16 , 24 , /* 5 --- 9 */
32 , 48 , 64 , 96 , 128 , /* 10 --- 14 */
2097152 /* 50 */
};
-#ifdef USE_LZX_EXTRA_BITS_ARRAY
-const u8 lzx_extra_bits[LZX_MAX_POSITION_SLOTS] = {
+/* Mapping: offset slot => how many extra bits must be read and added to the
+ * corresponding offset slot base to decode the match offset. */
+const u8 lzx_extra_offset_bits[LZX_MAX_OFFSET_SLOTS] = {
0 , 0 , 0 , 0 , 1 ,
1 , 2 , 2 , 3 , 3 ,
4 , 4 , 5 , 5 , 6 ,
17, 17, 17, 17, 17,
17
};
-#endif
-/* LZX window size can be between 2^15 and 2^21, inclusively. */
-bool
-lzx_window_size_valid(u32 window_size)
+/* Round the specified compression block size (not LZX block size) up to the
+ * next valid LZX window size, and return its order (log2). Or, if the block
+ * size is 0 or greater than the largest valid LZX window size, return 0. */
+unsigned
+lzx_get_window_order(size_t max_block_size)
{
- if (window_size == 0)
- return false;
- u32 order = bsr32(window_size);
- if (window_size != 1U << order)
- return false;
- return (order >= LZX_MIN_WINDOW_ORDER && order <= LZX_MAX_WINDOW_ORDER);
+ unsigned order;
+
+ if (max_block_size == 0 || max_block_size > LZX_MAX_WINDOW_SIZE)
+ return 0;
+
+ order = fls32(max_block_size);
+
+ if (((u32)1 << order) != max_block_size)
+ order++;
+
+ return max(order, LZX_MIN_WINDOW_ORDER);
}
-/* Given the specified valid window size, return the number of LZX main symbols
- * that will be needed. (This depends on the number of position slots, which
- * itself depends on the window size.) */
+/* Given a valid LZX window order, return the number of symbols that will exist
+ * in the main Huffman code. */
unsigned
-lzx_get_num_main_syms(u32 window_size)
+lzx_get_num_main_syms(unsigned window_order)
+{
+ u32 window_size = (u32)1 << window_order;
+
+ /* NOTE: the calculation *should* be as follows:
+ *
+ * u32 max_offset = window_size - LZX_MIN_MATCH_LEN;
+ * u32 max_adjusted_offset = max_offset + LZX_OFFSET_OFFSET;
+ * u32 num_offset_slots = 1 + lzx_get_offset_slot_raw(max_adjusted_offset);
+ *
+ * However since LZX_MIN_MATCH_LEN == LZX_OFFSET_OFFSET, we would get
+ * max_adjusted_offset == window_size, which would bump the number of
+ * offset slots up by 1 since every valid LZX window size is equal to a
+ * offset slot base value. The format doesn't do this, and instead
+ * disallows matches with minimum length and maximum offset. This sets
+ * max_adjusted_offset = window_size - 1, so instead we must calculate:
+ *
+ * num_offset_slots = 1 + lzx_get_offset_slot_raw(window_size - 1);
+ *
+ * ... which is the same as
+ *
+ * num_offset_slots = lzx_get_offset_slot_raw(window_size);
+ *
+ * ... since every valid window size is equal to an offset base value.
+ */
+ unsigned num_offset_slots = lzx_get_offset_slot_raw(window_size);
+
+ /* Now calculate the number of main symbols as LZX_NUM_CHARS literal
+ * symbols, plus 8 symbols per offset slot (since there are 8 possible
+ * length headers, and we need all (offset slot, length header)
+ * combinations). */
+ return LZX_NUM_CHARS + (num_offset_slots << 3);
+}
+
+static void
+do_translate_target(void *target, s32 input_pos)
+{
+ s32 abs_offset, rel_offset;
+
+ rel_offset = get_unaligned_u32_le(target);
+ if (rel_offset >= -input_pos && rel_offset < LZX_WIM_MAGIC_FILESIZE) {
+ if (rel_offset < LZX_WIM_MAGIC_FILESIZE - input_pos) {
+ /* "good translation" */
+ abs_offset = rel_offset + input_pos;
+ } else {
+ /* "compensating translation" */
+ abs_offset = rel_offset - LZX_WIM_MAGIC_FILESIZE;
+ }
+ put_unaligned_u32_le(abs_offset, target);
+ }
+}
+
+static void
+undo_translate_target(void *target, s32 input_pos)
+{
+ s32 abs_offset, rel_offset;
+
+ abs_offset = get_unaligned_u32_le(target);
+ if (abs_offset >= 0) {
+ if (abs_offset < LZX_WIM_MAGIC_FILESIZE) {
+ /* "good translation" */
+ rel_offset = abs_offset - input_pos;
+ put_unaligned_u32_le(rel_offset, target);
+ }
+ } else {
+ if (abs_offset >= -input_pos) {
+ /* "compensating translation" */
+ rel_offset = abs_offset + LZX_WIM_MAGIC_FILESIZE;
+ put_unaligned_u32_le(rel_offset, target);
+ }
+ }
+}
+
+/*
+ * Do or undo the 'E8' preprocessing used in LZX. Before compression, the
+ * uncompressed data is preprocessed by changing the targets of x86 CALL
+ * instructions from relative offsets to absolute offsets. After decompression,
+ * the translation is undone by changing the targets of x86 CALL instructions
+ * from absolute offsets to relative offsets.
+ *
+ * Note that despite its intent, E8 preprocessing can be done on any data even
+ * if it is not actually x86 machine code. In fact, E8 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.
+ *
+ * E8 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.
+ *
+ * E8 preprocessing does not appear to be disabled after the 32768th chunk of a
+ * WIM resource, which apparently is another difference from the LZX compression
+ * used in cabinet files.
+ *
+ * E8 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
+lzx_e8_filter(u8 *data, u32 size, void (*process_target)(void *, s32))
+{
+
+#if !defined(__SSE2__) && !defined(__AVX2__)
+ /*
+ * A worthwhile optimization is to push the end-of-buffer check into the
+ * relatively rare E8 case. This is possible if we replace the last six
+ * bytes of data with E8 bytes; then we are guaranteed to hit an E8 byte
+ * before reaching end-of-buffer. In addition, this scheme guarantees
+ * that no translation can begin following an E8 byte in the last 10
+ * bytes because a 4-byte offset containing E8 as its high byte is a
+ * large negative number that is not valid for translation. That is
+ * exactly what we need.
+ */
+ u8 *tail;
+ u8 saved_bytes[6];
+ u8 *p;
+
+ if (size <= 10)
+ return;
+
+ tail = &data[size - 6];
+ memcpy(saved_bytes, tail, 6);
+ memset(tail, 0xE8, 6);
+ p = data;
+ for (;;) {
+ while (*p != 0xE8)
+ p++;
+ if (p >= tail)
+ break;
+ (*process_target)(p + 1, p - data);
+ p += 5;
+ }
+ memcpy(tail, saved_bytes, 6);
+#else
+ /* SSE2 or AVX-2 optimized version for x86_64 */
+
+ u8 *p = data;
+ u64 valid_mask = ~0;
+
+ if (size <= 10)
+ return;
+#ifdef __AVX2__
+# define ALIGNMENT_REQUIRED 32
+#else
+# define ALIGNMENT_REQUIRED 16
+#endif
+
+ /* Process one byte at a time until the pointer is properly aligned. */
+ while ((uintptr_t)p % ALIGNMENT_REQUIRED != 0) {
+ if (p >= data + size - 10)
+ return;
+ if (*p == 0xE8 && (valid_mask & 1)) {
+ (*process_target)(p + 1, p - data);
+ valid_mask &= ~0x1F;
+ }
+ p++;
+ valid_mask >>= 1;
+ valid_mask |= (u64)1 << 63;
+ }
+
+ if (data + size - p >= 64) {
+
+ /* Vectorized processing */
+
+ /* Note: we use a "trap" E8 byte to eliminate the need to check
+ * for end-of-buffer in the inner loop. This byte is carefully
+ * positioned so that it will never be changed by a previous
+ * translation before it is detected. */
+
+ u8 *trap = p + ((data + size - p) & ~31) - 32 + 4;
+ u8 saved_byte = *trap;
+ *trap = 0xE8;
+
+ for (;;) {
+ u32 e8_mask;
+ u8 *orig_p = p;
+ #ifdef __SSE2__
+ const __m128i e8_bytes = _mm_set1_epi8(0xE8);
+ for (;;) {
+ /* Read the next 32 bytes of data and test them
+ * for E8 bytes. */
+ __m128i bytes1 = *(const __m128i *)p;
+ __m128i bytes2 = *(const __m128i *)(p + 16);
+ __m128i cmpresult1 = _mm_cmpeq_epi8(bytes1, e8_bytes);
+ __m128i cmpresult2 = _mm_cmpeq_epi8(bytes2, e8_bytes);
+ u32 mask1 = _mm_movemask_epi8(cmpresult1);
+ u32 mask2 = _mm_movemask_epi8(cmpresult2);
+ /* The masks have a bit set for each E8 byte.
+ * We stay in this fast inner loop as long as
+ * there are no E8 bytes. */
+ if (mask1 | mask2) {
+ e8_mask = mask1 | (mask2 << 16);
+ break;
+ }
+ p += 32;
+ }
+ #else
+ /* AVX-2 */
+ const __m256i e8_bytes = _mm256_set1_epi8(0xE8);
+ for (;;) {
+ __m256i bytes = *(const __m256i *)p;
+ __m256i cmpresult = _mm256_cmpeq_epi8(bytes, e8_bytes);
+ e8_mask = _mm256_movemask_epi8(cmpresult);
+ if (e8_mask)
+ break;
+ p += 32;
+ }
+ #endif
+
+ /* Did we pass over data with no E8 bytes? */
+ if (p != orig_p)
+ valid_mask = ~0;
+
+ /* Are we nearing end-of-buffer? */
+ if (p == trap - 4)
+ break;
+
+ /* Process the E8 bytes. However, the AND with
+ * 'valid_mask' ensures we never process an E8 byte that
+ * was itself part of a translation target. */
+ while ((e8_mask &= valid_mask)) {
+ unsigned bit = ffs32(e8_mask);
+ (*process_target)(p + bit + 1, p + bit - data);
+ valid_mask &= ~((u64)0x1F << bit);
+ }
+
+ valid_mask >>= 32;
+ valid_mask |= 0xFFFFFFFF00000000;
+ p += 32;
+ }
+
+ *trap = saved_byte;
+ }
+
+ /* Approaching the end of the buffer; process one byte a time. */
+ while (p < data + size - 10) {
+ if (*p == 0xE8 && (valid_mask & 1)) {
+ (*process_target)(p + 1, p - data);
+ valid_mask &= ~0x1F;
+ }
+ p++;
+ valid_mask >>= 1;
+ valid_mask |= (u64)1 << 63;
+ }
+#endif /* __SSE2__ || __AVX2__ */
+}
+
+void
+lzx_do_e8_preprocessing(u8 *data, u32 size)
+{
+ lzx_e8_filter(data, size, do_translate_target);
+}
+
+void
+lzx_undo_e8_preprocessing(u8 *data, u32 size)
{
- unsigned num_position_slots = lzx_get_position_slot_raw(window_size);
- return LZX_NUM_CHARS + (num_position_slots << 3);
+ lzx_e8_filter(data, size, undo_translate_target);
}