[PATCH] Fix PR tree-optimization/83195 testcase for arm

[PATCH] Fix PR tree-optimization/83195 testcase for arm

[Jakub Jelinek]

Hi!

The pr82929.c testcase uses store_merge effective target, which is
int32plus && nonstrict_align.  Unfortunately, arm is handled for
nonstrict_align wierdly, although it has STRICT_ALIGNMENT 1, it is sometimes
considered nonstrict_align (the only exception apparently).

Now, the testcase really needs a non-strict alignment target where
STRICT_ALIGNMENT is 0, otherwise the optimization it tests is not
beneficial.  So, the following patch stops testing it on arm, and adds
another test where the pointers are guaranteed to be aligned and thus we
can test for the optimization even on non-strict alignment targets.

Bootstrapped/regtested on x86_64-linux and i686-linux, tested by hand using
a cross-compiler to arm, ok for trunk?

2017-11-29  Jakub Jelinek  <jakub@redhat.com>

	PR tree-optimization/83195
	* gcc.dg/pr82929.c: Don't check for "Merging successful" on arm.
	* gcc.dg/pr82929-2.c: New test.

--- gcc/testsuite/gcc.dg/pr82929.c.jj	2017-11-10 15:42:39.000000000 +0100
+++ gcc/testsuite/gcc.dg/pr82929.c	2017-11-28 17:50:43.705221829 +0100
@@ -15,4 +15,4 @@ foo (short *p, short *q, short *r)
   p[1] = e & f;
 }
 
-/* { dg-final { scan-tree-dump-times "Merging successful" 1 "store-merging" } } */
+/* { dg-final { scan-tree-dump-times "Merging successful" 1 "store-merging" { target { ! arm*-*-* } } } } */
--- gcc/testsuite/gcc.dg/pr82929-2.c.jj	2017-11-28 17:47:41.858409094 +0100
+++ gcc/testsuite/gcc.dg/pr82929-2.c	2017-11-28 17:48:55.264526160 +0100
@@ -0,0 +1,21 @@
+/* PR tree-optimization/82929 */
+/* { dg-do compile { target store_merge } } */
+/* { dg-options "-O2 -fdump-tree-store-merging" } */
+
+void
+foo (short *p, short *q, short *r)
+{
+  p = __builtin_assume_aligned (p, __alignof__ (int));
+  q = __builtin_assume_aligned (q, __alignof__ (int));
+  r = __builtin_assume_aligned (r, __alignof__ (int));
+  short a = q[0];
+  short b = q[1];
+  short c = ~a;
+  short d = r[0];
+  short e = r[1];
+  short f = ~b;
+  p[0] = c & d;
+  p[1] = e & f;
+}
+
+/* { dg-final { scan-tree-dump-times "Merging successful" 1 "store-merging" } } */

	Jakub

Re: [PATCH] Fix PR tree-optimization/83195 testcase for arm

[Kyrill Tkachov]

Hi Jakub,

On 29/11/17 08:18, Jakub Jelinek wrote:
> Hi!
>
> The pr82929.c testcase uses store_merge effective target, which is
> int32plus && nonstrict_align.  Unfortunately, arm is handled for
> nonstrict_align wierdly, although it has STRICT_ALIGNMENT 1, it is sometimes
> considered nonstrict_align (the only exception apparently).
>
> Now, the testcase really needs a non-strict alignment target where
> STRICT_ALIGNMENT is 0, otherwise the optimization it tests is not
> beneficial.  So, the following patch stops testing it on arm, and adds
> another test where the pointers are guaranteed to be aligned and thus we
> can test for the optimization even on non-strict alignment targets.
>
> Bootstrapped/regtested on x86_64-linux and i686-linux, tested by hand using
> a cross-compiler to arm, ok for trunk?
>
> 2017-11-29  Jakub Jelinek  <jakub@redhat.com>
>
> 	PR tree-optimization/83195
> 	* gcc.dg/pr82929.c: Don't check for "Merging successful" on arm.
> 	* gcc.dg/pr82929-2.c: New test.
>
> --- gcc/testsuite/gcc.dg/pr82929.c.jj	2017-11-10 15:42:39.000000000 +0100
> +++ gcc/testsuite/gcc.dg/pr82929.c	2017-11-28 17:50:43.705221829 +0100
> @@ -15,4 +15,4 @@ foo (short *p, short *q, short *r)
>     p[1] = e & f;
>   }
>   
> -/* { dg-final { scan-tree-dump-times "Merging successful" 1 "store-merging" } } */
> +/* { dg-final { scan-tree-dump-times "Merging successful" 1 "store-merging" { target { ! arm*-*-* } } } } */
> --- gcc/testsuite/gcc.dg/pr82929-2.c.jj	2017-11-28 17:47:41.858409094 +0100
> +++ gcc/testsuite/gcc.dg/pr82929-2.c	2017-11-28 17:48:55.264526160 +0100
> @@ -0,0 +1,21 @@
> +/* PR tree-optimization/82929 */
> +/* { dg-do compile { target store_merge } } */
> +/* { dg-options "-O2 -fdump-tree-store-merging" } */
> +
> +void
> +foo (short *p, short *q, short *r)
> +{
> +  p = __builtin_assume_aligned (p, __alignof__ (int));
> +  q = __builtin_assume_aligned (q, __alignof__ (int));
> +  r = __builtin_assume_aligned (r, __alignof__ (int));
> +  short a = q[0];
> +  short b = q[1];
> +  short c = ~a;
> +  short d = r[0];
> +  short e = r[1];
> +  short f = ~b;
> +  p[0] = c & d;
> +  p[1] = e & f;
> +}
> +
> +/* { dg-final { scan-tree-dump-times "Merging successful" 1 "store-merging" } } */

Sudi has kindly tried it out and the new test passes on arm, so this 
patch is ok from an arm perspective.
Thanks for fixing this,
Kyrill

> 	Jakub