Fix PR41879 (large mgrid regression)

[Uros Bizjak]

Hello!

> It doesn't because I overwrite the options (which would have -msse2 on
> x86) with dg-options, and there's no way to just _add_ to the current
> options, except via implementing something like add_options_for_ieee (just
> for vectorizer options).  Pff, I think I'm going to checkin this patch,
> making the test active only on x86(-64), so that I can hardcode -msse2.


>	* gfortran.dg/vect/fast-math-mgrid-resid.f: Limit to x86, add
>	-msse2.

> Index: testsuite/gfortran.dg/vect/fast-math-mgrid-resid.f
> ===================================================================
> --- testsuite/gfortran.dg/vect/fast-math-mgrid-resid.f	(revision 156043)
> +++ testsuite/gfortran.dg/vect/fast-math-mgrid-resid.f	(working copy)
> @@ -1,6 +1,7 @@
> - ! { dg-do compile }
> +! { dg-do compile { target i?86-*-* x86_64-*-* } }
>  ! { dg-require-effective-target vect_double }
> -! { dg-options "-O3 -ffast-math -fpredictive-commoning -ftree-vectorize -fdump-tree-optimized" }
> +! { dg-options "-O3 -ffast-math -msse2 -fpredictive-commoning -ftree-vectorize -fdump-tree-optimized" }

Just move the test to gfortran.dg/ then, since you are not using any
of vect.exp features. x86 with -msse2 is vect_double for 32bit and
64bit by default, so no need to check. And by moving the test to
gfortran.dg/, you can also remove cleanup-tree-dump "vect" final
directive at the end of the test.

Uros.