+
+/* 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)
+{
+ 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 a valid LZX window order, return the number of symbols that will exist
+ * in the main Huffman code. */
+unsigned
+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