-#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. Also, this optimization affects
- * compression as well, but the percentage improvement is less because
- * LZX compression is much slower than LZX decompression. ) */
- __m128i *p128 = (__m128i *)data;
- u32 valid_mask = 0xFFFFFFFF;
-
- if (size >= 32 && (uintptr_t)data % 16 == 0) {
- __m128i * const end128 = p128 + 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 (or undo) the e8 translation. */
- u8 *p8 = (u8 *)p128 + bit;
- (*process_target)((s32 *)(p8 + 1),
- p8 - 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);
+
+#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;