+
+#ifdef __x86_64__
+static inline u8 *
+find_next_opcode_sse4_2(u8 *p)
+{
+ const __v16qi potential_opcodes = (__v16qi) {0x48, 0x4C, 0xE8, 0xE9, 0xF0, 0xFF};
+ __asm__(
+ " pcmpestri $0x0, (%[p]), %[potential_opcodes] \n"
+ " jc 2f \n"
+ "1: \n"
+ " add $0x10, %[p] \n"
+ " pcmpestri $0x0, (%[p]), %[potential_opcodes] \n"
+ " jnc 1b \n"
+ "2: \n"
+ " add %%rcx, %[p] \n"
+ : [p] "+r" (p)
+ : [potential_opcodes] "x" (potential_opcodes), "a" (6), "d" (16)
+ : "rcx", "cc"
+ );
+
+ return p;
+}
+#endif /* __x86_64__ */
+
+static inline u8 *
+find_next_opcode_default(u8 *p)
+{
+ /*
+ * The following table is used to accelerate the common case where the
+ * byte has nothing to do with x86 translation and must simply be
+ * skipped. This was faster than the following alternatives:
+ * - Jump table with 256 entries
+ * - Switch statement with default
+ */
+ static const u8 is_potential_opcode[256] = {
+ [0x48] = 1, [0x4C] = 1, [0xE8] = 1,
+ [0xE9] = 1, [0xF0] = 1, [0xFF] = 1,
+ };
+
+ for (;;) {
+ if (is_potential_opcode[*p])
+ break;
+ p++;
+ if (is_potential_opcode[*p])
+ break;
+ p++;
+ if (is_potential_opcode[*p])
+ break;
+ p++;
+ if (is_potential_opcode[*p])
+ break;
+ p++;
+ }
+ return p;
+}
+
+static inline u8 *
+translate_if_needed(u8 *data, u8 *p, s32 *last_x86_pos,
+ s32 last_target_usages[], bool undo)
+{
+ s32 max_trans_offset;
+ s32 opcode_nbytes;
+ u16 target16;
+ s32 i;
+
+ max_trans_offset = LZMS_X86_MAX_TRANSLATION_OFFSET;
+
+ /*
+ * p[0] has one of the following values:
+ * 0x48 0x4C 0xE8 0xE9 0xF0 0xFF
+ */
+
+ if (p[0] >= 0xF0) {
+ if (p[0] & 0x0F) {
+ /* 0xFF (instruction group) */
+ if (p[1] == 0x15) {
+ /* Call indirect relative */
+ opcode_nbytes = 2;
+ goto have_opcode;
+ }
+ } else {
+ /* 0xF0 (lock prefix) */
+ if (p[1] == 0x83 && p[2] == 0x05) {
+ /* Lock add relative */
+ opcode_nbytes = 3;
+ goto have_opcode;
+ }
+ }
+ } else if (p[0] <= 0x4C) {
+
+ /* 0x48 or 0x4C. In 64-bit code this is a REX prefix byte with
+ * W=1, R=[01], X=0, and B=0, and it will be followed by the
+ * actual opcode, then additional bytes depending on the opcode.
+ * We are most interested in several common instructions that
+ * access data relative to the instruction pointer. These use a
+ * 1-byte opcode, followed by a ModR/M byte, followed by a
+ * 4-byte displacement. */
+
+ /* Test: does the ModR/M byte indicate RIP-relative addressing?
+ * Note: there seems to be a mistake in the format here; the
+ * mask really should be 0xC7 instead of 0x07 so that both the
+ * MOD and R/M fields of ModR/M are tested, not just R/M. */
+ if ((p[2] & 0x07) == 0x05) {
+ /* Check for the LEA (load effective address) or MOV
+ * (move) opcodes. For MOV there are additional
+ * restrictions, although it seems they are only helpful
+ * due to the overly lax ModR/M test. */
+ if (p[1] == 0x8D ||
+ (p[1] == 0x8B && !(p[0] & 0x04) && !(p[2] & 0xF0)))
+ {
+ opcode_nbytes = 3;
+ goto have_opcode;
+ }
+ }
+ } else {
+ if (p[0] & 0x01) {
+ /* 0xE9: Jump relative. Theoretically this would be
+ * useful to translate, but in fact it's explicitly
+ * excluded. Most likely it creates too many false
+ * positives for the detection algorithm. */
+ p += 4;
+ } else {
+ /* 0xE8: Call relative. This is a common case, so it
+ * uses a reduced max_trans_offset. In other words, we
+ * have to be more confident that the data actually is
+ * x86 machine code before we'll do the translation. */
+ opcode_nbytes = 1;
+ max_trans_offset >>= 1;
+ goto have_opcode;
+ }
+ }
+
+ return p + 1;
+
+have_opcode:
+ i = p - data;
+ p += opcode_nbytes;
+ if (undo) {
+ if (i - *last_x86_pos <= max_trans_offset) {
+ u32 n = get_unaligned_u32_le(p);
+ put_unaligned_u32_le(n - i, p);
+ }
+ target16 = i + get_unaligned_u16_le(p);
+ } else {
+ target16 = i + get_unaligned_u16_le(p);
+ if (i - *last_x86_pos <= max_trans_offset) {
+ u32 n = get_unaligned_u32_le(p);
+ put_unaligned_u32_le(n + i, p);
+ }
+ }
+
+ i += opcode_nbytes + sizeof(le32) - 1;
+
+ if (i - last_target_usages[target16] <= LZMS_X86_ID_WINDOW_SIZE)
+ *last_x86_pos = i;
+
+ last_target_usages[target16] = i;
+
+ return p + sizeof(le32);
+}
+