* [PATCH] Fix ifcvt one_cmpl_abs optimization (PR rtl-optimization/68376)
@ 2015-11-18 22:44 Jakub Jelinek
2015-11-19 0:21 ` Bernd Schmidt
0 siblings, 1 reply; 2+ messages in thread
From: Jakub Jelinek @ 2015-11-18 22:44 UTC (permalink / raw)
To: gcc-patches
Hi!
noce_try_abs optimizes these cases normal abs, which doesn't really
care if the original condition is x < 0 ? -x : x or x <= 0 ? -x : x,
but also the x < 0 ? ~x : x case. But in this case it is significant
whether for x == 0 ~x or x applies; the following patch limits the
one_cmpl optimization to those cases where the optimized expression
computes the same value.
Bootstrapped/regtested on x86_64-linux and i686-linux, ok for trunk/5.3?
2015-11-18 Jakub Jelinek <jakub@redhat.com>
PR rtl-optimization/68376
* ifcvt.c (noce_try_abs): Disable one_cmpl optimization if
encountering x <= 0 ? ~x : x or x > 0 ? ~x : x.
* gcc.c-torture/execute/pr68376-1.c: New test.
* gcc.c-torture/execute/pr68376-2.c: New test.
--- gcc/ifcvt.c.jj 2015-11-14 19:35:54.000000000 +0100
+++ gcc/ifcvt.c 2015-11-18 11:02:48.645771477 +0100
@@ -2595,12 +2595,49 @@ noce_try_abs (struct noce_if_info *if_in
/* Work around funny ideas get_condition has wrt canonicalization.
Note that these rtx constants are known to be CONST_INT, and
- therefore imply integer comparisons. */
+ therefore imply integer comparisons.
+ The one_cmpl case is more complicated, as we want to handle
+ only x < 0 ? ~x : x or x >= 0 ? ~x : x but not
+ x <= 0 ? ~x : x or x > 0 ? ~x : x, as the latter two
+ have different result for x == 0. */
if (c == constm1_rtx && GET_CODE (cond) == GT)
- ;
+ {
+ if (one_cmpl && negate)
+ return FALSE;
+ }
else if (c == const1_rtx && GET_CODE (cond) == LT)
- ;
- else if (c != CONST0_RTX (GET_MODE (b)))
+ {
+ if (one_cmpl && !negate)
+ return FALSE;
+ }
+ else if (c == CONST0_RTX (GET_MODE (b)))
+ {
+ if (one_cmpl)
+ switch (GET_CODE (cond))
+ {
+ case GT:
+ if (!negate)
+ return FALSE;
+ break;
+ case GE:
+ /* >= 0 is the same case as above > -1. */
+ if (negate)
+ return FALSE;
+ break;
+ case LT:
+ if (negate)
+ return FALSE;
+ break;
+ case LE:
+ /* <= 0 is the same case as above < 1. */
+ if (!negate)
+ return FALSE;
+ break;
+ default:
+ return FALSE;
+ }
+ }
+ else
return FALSE;
/* Determine what sort of operation this is. */
--- gcc/testsuite/gcc.c-torture/execute/pr68376-1.c.jj 2015-11-18 11:12:33.251522987 +0100
+++ gcc/testsuite/gcc.c-torture/execute/pr68376-1.c 2015-11-18 11:12:24.000000000 +0100
@@ -0,0 +1,24 @@
+/* PR rtl-optimization/68376 */
+
+int a, b, c = 1;
+signed char d;
+
+int
+main ()
+{
+ for (; a < 1; a++)
+ for (; b < 1; b++)
+ {
+ signed char e = ~d;
+ if (d < 1)
+ e = d;
+ d = e;
+ if (!c)
+ __builtin_abort ();
+ }
+
+ if (d != 0)
+ __builtin_abort ();
+
+ return 0;
+}
--- gcc/testsuite/gcc.c-torture/execute/pr68376-2.c.jj 2015-11-18 11:12:36.209481252 +0100
+++ gcc/testsuite/gcc.c-torture/execute/pr68376-2.c 2015-11-18 11:10:55.000000000 +0100
@@ -0,0 +1,41 @@
+/* PR rtl-optimization/68376 */
+
+extern void abort (void);
+
+__attribute__((noinline, noclone)) int
+f1 (int x)
+{
+ return x < 0 ? ~x : x;
+}
+
+__attribute__((noinline, noclone)) int
+f2 (int x)
+{
+ return x < 0 ? x : ~x;
+}
+
+__attribute__((noinline, noclone)) int
+f3 (int x)
+{
+ return x <= 0 ? ~x : x;
+}
+
+__attribute__((noinline, noclone)) int
+f4 (int x)
+{
+ return x <= 0 ? x : ~x;
+}
+
+int
+main ()
+{
+ if (f1 (5) != 5 || f1 (-5) != 4 || f1 (0) != 0)
+ abort ();
+ if (f2 (5) != -6 || f2 (-5) != -5 || f2 (0) != -1)
+ abort ();
+ if (f3 (5) != 5 || f3 (-5) != 4 || f3 (0) != -1)
+ abort ();
+ if (f4 (5) != -6 || f4 (-5) != -5 || f4 (0) != 0)
+ abort ();
+ return 0;
+}
Jakub
^ permalink raw reply [flat|nested] 2+ messages in thread
* Re: [PATCH] Fix ifcvt one_cmpl_abs optimization (PR rtl-optimization/68376)
2015-11-18 22:44 [PATCH] Fix ifcvt one_cmpl_abs optimization (PR rtl-optimization/68376) Jakub Jelinek
@ 2015-11-19 0:21 ` Bernd Schmidt
0 siblings, 0 replies; 2+ messages in thread
From: Bernd Schmidt @ 2015-11-19 0:21 UTC (permalink / raw)
To: Jakub Jelinek, gcc-patches
On 11/18/2015 11:44 PM, Jakub Jelinek wrote:
> noce_try_abs optimizes these cases normal abs, which doesn't really
> care if the original condition is x < 0 ? -x : x or x <= 0 ? -x : x,
> but also the x < 0 ? ~x : x case. But in this case it is significant
> whether for x == 0 ~x or x applies; the following patch limits the
> one_cmpl optimization to those cases where the optimized expression
> computes the same value.
Ok.
Bernd
^ permalink raw reply [flat|nested] 2+ messages in thread
end of thread, other threads:[~2015-11-19 0:21 UTC | newest]
Thread overview: 2+ messages (download: mbox.gz / follow: Atom feed)
-- links below jump to the message on this page --
2015-11-18 22:44 [PATCH] Fix ifcvt one_cmpl_abs optimization (PR rtl-optimization/68376) Jakub Jelinek
2015-11-19 0:21 ` Bernd Schmidt
This is a public inbox, see mirroring instructions
for how to clone and mirror all data and code used for this inbox;
as well as URLs for read-only IMAP folder(s) and NNTP newsgroup(s).