[PATCH] gcc.dg/vect/vect-77a.c: xfail for lp64

[Dorit Naishlos]

Janis Johnson <janis187@us.ibm.com> wrote on 22/12/2004 19:21:34:

> On Wed, Dec 22, 2004 at 05:17:18PM +0200, Dorit Naishlos wrote:
> >
> > actually we would still fail to vectorize for sparc and alpha because
of
> > lack of alignment support. The reason this testcase gets vectorized on
> > i?86-*-* and x86_64-*-* is that this test uses -mmmx for which
misaligned
> > accesses are modeled (as opposed to -msse*). The
> > vect-[27,29,48,56,72,77]a.c tests are just a duplicate of the
> > vect-[27,29,48,56,72,77].c tests with one difference - they use -mmmx
> > instead of -msse. So, I suggest that we enable the
> > vect-[27,29,48,56,72,77]a.c tests for i?86-*-* and x86_64-*-* only, and
the
> > xfail will be empty. (Also, I don't know if you want to mention in
> > check_effective_target_vect_no_align that i?86-*-* and x86_64-*-* do
model
> > misaligned accesses for mmx, but not for sse*; the vectorizer tests
usually
> > use -msse so that's why we consider these as no_align for now).
>
> There was a discussion on IRC a couple of days ago about having the vect
> tests cycle through all of the x86 vector instruction sets and being
> compile-only for sets not supported by the test hardware.  This is now
> PR 19095.
>

thanks for the update

> Are these limitations temporary or permanent?

they are temporary

> Limitations that are
> permenent should cause tests to be UNSUPPORTED for targets where they
> are not supported, and those that are temporary should be XFAIL for
> those targets and reference a PR that explains the limitation.

I think if it xfails for targets that are vect_no_align and such, the
limitation/explanation is already within the xfail, thanks to the
descriptive new keywords you have added

> We
> could have effective-target keywords local to vect.exp that include
> all targets covered by a particular PR, where the effective target
> takes into account the instruction set being used (mmx/sse/3dNow!).
>

It will be nice if the effective-target keyword could take into account the
instruction set being used, but I don't think that assigning
effective-target keywords per PR is better than what we have now.

dorit

> Janis