@@ -1529,6 +1529,7 @@ static void tcg_out_setcond(TCGContext *s, int rexw, TCGCond cond,
int const_arg2)
{
bool inv = false;
+ bool cleared;
switch (cond) {
case TCG_COND_NE:
@@ -1578,9 +1579,23 @@ static void tcg_out_setcond(TCGContext *s, int rexw, TCGCond cond,
break;
}
+ /*
+ * If dest does not overlap the inputs, clearing it first is preferred.
+ * The XOR breaks any false dependency for the low-byte write to dest,
+ * and is also one byte smaller than MOVZBL.
+ */
+ cleared = false;
+ if (dest != arg1 && (const_arg2 || dest != arg2)) {
+ tgen_arithr(s, ARITH_XOR, dest, dest);
+ cleared = true;
+ }
+
tcg_out_cmp(s, rexw, arg1, arg2, const_arg2, false);
tcg_out_modrm(s, OPC_SETCC | tcg_cond_to_jcc[cond], 0, dest);
- tcg_out_ext8u(s, dest, dest);
+
+ if (!cleared) {
+ tcg_out_ext8u(s, dest, dest);
+ }
}
#if TCG_TARGET_REG_BITS == 32
Using XOR first is both smaller and more efficient, though cannot be applied if it clobbers an input. Signed-off-by: Richard Henderson <richard.henderson@linaro.org> --- tcg/i386/tcg-target.c.inc | 17 ++++++++++++++++- 1 file changed, 16 insertions(+), 1 deletion(-)