On Mon, Oct 21, 2013 at 9:39 PM, Mike Stump <mikest...@comcast.net> wrote:
> Concerning:
>
> 2008-11-20 Richard Guenther <rguent...@suse.de>
>
> PR tree-optimization/37868
> * gcc.dg/torture/pr37868.c: New testcase.
> * gcc.c-torture/execute/pr38048-1.c: Likewise.
> * gcc.c-torture/execute/pr38048-2.c: Likewise.
>
> So, is there any reason why we can't make the test case portable as in the
> below instead of skipping on all the targets that blow on misaligned data?
>
> Ok?
Clearly this was an alias analysis fix so adding a union may no longer
trigger the failure mode. Please make sure to go back in time
where the testcase failed and see if your adjusted testcase fails _in
the same way_. Note that you don't adjust the dg-skip-if in your
patch and note
that actually there should be no unaligned accesses performed if you
instead do
typedef unsigned int unaligned_uint __attribute__((aligned(1)));
...
bad_bits = ((unsigned int)-1) ^ *(1+(unaligned_uint *) &x);
but double-check this adjusted testcase still fails without the fix.
Richard.
> diff --git a/gcc/testsuite/gcc.dg/torture/pr37868.c
> b/gcc/testsuite/gcc.dg/torture/pr37868.c
> index 380b089..b7ff43f 100644
> --- a/gcc/testsuite/gcc.dg/torture/pr37868.c
> +++ b/gcc/testsuite/gcc.dg/torture/pr37868.c
> @@ -24,15 +24,18 @@ struct X {
>
> int main (void)
> {
> - struct X x;
> + union {
> + struct X x;
> + unsigned int i;
> + } s;
> unsigned int bad_bits;
>
> - x.pad = -1;
> - x.a = -1;
> - x.b = -1;
> - x.c = -1;
> + s.x.pad = -1;
> + s.x.a = -1;
> + s.x.b = -1;
> + s.x.c = -1;
>
> - bad_bits = ((unsigned int)-1) ^ *(1+(unsigned int *) &x);
> + bad_bits = ((unsigned int)-1) ^ *(1+(unsigned int *) &s.x);
> if (bad_bits != 0)
> abort ();
> return 0;
>
