On 11/22/19 6:05 PM, Uros Bizjak wrote:
> Indeed, this is a different case, an overflow test that results in one
> CMP insn. I think, we should check if the second operand is either 0
> (then proceed as it is now), or if the second operand equals first
> operand of PLUS insn, then we actually emit CMP insn (please see
> PR30315).
Here's what I committed - preapproved by Uros off-list (I forgot
Reply-All again).
Bernd
Index: gcc/ChangeLog
===================================================================
--- gcc/ChangeLog (revision 278653)
+++ gcc/ChangeLog (working copy)
@@ -1,3 +1,8 @@
+2019-11-24 Bernd Schmidt <bernds_...@t-online.de>
+
+ * config/i386/i386.c (ix86_rtx_costs): Handle care of a PLUS in a
+ COMPARE, representing an overflow detection.
+
2019-11-23 Jan Hubicka <hubi...@ucw.cz>
* cif-code.def (MAX_INLINE_INSNS_SINGLE_O2_LIMIT): Remove.
Index: gcc/config/i386/i386.c
===================================================================
--- gcc/config/i386/i386.c (revision 278653)
+++ gcc/config/i386/i386.c (working copy)
@@ -19501,6 +19501,15 @@ ix86_rtx_costs (rtx x, machine_mode mode
return true;
}
+ if (GET_CODE (XEXP (x, 0)) == PLUS
+ && rtx_equal_p (XEXP (XEXP (x, 0), 0), XEXP (x, 1)))
+ {
+ /* This is an overflow detection, count it as a normal compare. */
+ *total = rtx_cost (XEXP (x, 0), GET_MODE (XEXP (x, 0)),
+ COMPARE, 0, speed);
+ return true;
+ }
+
/* The embedded comparison operand is completely free. */
if (!general_operand (XEXP (x, 0), GET_MODE (XEXP (x, 0)))
&& XEXP (x, 1) == const0_rtx)
