Patchwork [testsuite] ARM: remove -march from tests that don't need it

login
register
mail settings
Submitter Janis Johnson
Date June 21, 2011, 9:04 p.m.
Message ID <4E01074E.90407@codesourcery.com>
Download mbox | patch
Permalink /patch/101368/
State New
Headers show

Comments

Janis Johnson - June 21, 2011, 9:04 p.m.
This patch removes "-march=armv7-a" from four tests that pass with all
-march values from armv4t on.

OK for trunk?
2011-06-21  Janis Johnson  <janisjo@codesourcery.com>

	* gcc.target/arm/smlaltb-1.c: Don't specify -march.
	* gcc.target/arm/smlaltt-1.c: Likewise.
	* gcc.target/arm/smlatb-1.c: Likewise.
	* gcc.target/arm/smlatt-1.c: Likewise.
Ramana Radhakrishnan - June 21, 2011, 9:25 p.m.
On 21 June 2011 22:04, Janis Johnson <janisjo@codesourcery.com> wrote:
> This patch removes "-march=armv7-a" from four tests that pass with all
> -march values from armv4t on.
>
> OK for trunk?
>

No, that sounds like a bug - if we are generating smlaltb and friends
for v4t then the compiler is broken. This is probably because smlaltb
is a part of the file name :( .

smlaltb should be generated in ARM state for arch between v5e onwards
in Thumb state from v6t2 onwards (i.e. v6t2).

cheers
Ramana
Janis Johnson - June 21, 2011, 9:40 p.m.
On 06/21/2011 02:25 PM, Ramana Radhakrishnan wrote:
> On 21 June 2011 22:04, Janis Johnson <janisjo@codesourcery.com> wrote:
>> This patch removes "-march=armv7-a" from four tests that pass with all
>> -march values from armv4t on.
>>
>> OK for trunk?
>>
> 
> No, that sounds like a bug - if we are generating smlaltb and friends
> for v4t then the compiler is broken. This is probably because smlaltb
> is a part of the file name :( .
> 
> smlaltb should be generated in ARM state for arch between v5e onwards
> in Thumb state from v6t2 onwards (i.e. v6t2).

When I rename the files the tests fail for lots of -march values, so
renaming the tests is definitely needed.

The instructions checked for in these tests are provided in arm.md
for TARGET_DSP_MULTIPLY; is there a general way to check for that,
or should the test continue to provide a -march value that supports
it and also skip -march values from multilib flags that are known
to not support those instructions?

Janis

Patch

Index: gcc.target/arm/smlaltb-1.c
===================================================================
--- gcc.target/arm/smlaltb-1.c	(revision 175269)
+++ gcc.target/arm/smlaltb-1.c	(working copy)
@@ -1,5 +1,5 @@ 
 /* { dg-do compile } */
-/* { dg-options "-O2 -march=armv7-a" } */
+/* { dg-options "-O2" } */
 
 long long int
 foo (long long x, int in)
Index: gcc.target/arm/smlaltt-1.c
===================================================================
--- gcc.target/arm/smlaltt-1.c	(revision 175269)
+++ gcc.target/arm/smlaltt-1.c	(working copy)
@@ -1,5 +1,5 @@ 
 /* { dg-do compile } */
-/* { dg-options "-O2 -march=armv7-a" } */
+/* { dg-options "-O2" } */
 
 long long int
 foo (long long x, int in1, int in2)
Index: gcc.target/arm/smlatb-1.c
===================================================================
--- gcc.target/arm/smlatb-1.c	(revision 175269)
+++ gcc.target/arm/smlatb-1.c	(working copy)
@@ -1,5 +1,5 @@ 
 /* { dg-do compile } */
-/* { dg-options "-O2 -march=armv7-a" } */
+/* { dg-options "-O2" } */
 
 int
 foo (int x, int in)
Index: gcc.target/arm/smlatt-1.c
===================================================================
--- gcc.target/arm/smlatt-1.c	(revision 175269)
+++ gcc.target/arm/smlatt-1.c	(working copy)
@@ -1,5 +1,5 @@ 
 /* { dg-do compile } */
-/* { dg-options "-O2 -march=armv7-a" } */
+/* { dg-options "-O2" } */
 
 int
 foo (int x, int in1, int in2)