On Wed, Jan 29, 2025 at 08:01:17AM -0500, Siddhesh Poyarekar wrote:
> On 2025-01-29 07:34, Jakub Jelinek wrote:
> > > Denormal behaviour is well defined for IEEE128 long doubles, so don't
> > > XFAIL some gfortran tests on ppc64le when configured with the IEEE128
> > > long double ABI.
> >
> > If long double is just IEEE754 double, then I think denormals
> > are equally well behaved, it is solely the case of IBM double double.
> > Of course, these 3 tests won't run anyway in that case because
> > fortran_large_real effective target will be false.
>
> Yes, that's what I see in tests.
>
> > So I think you want an effective target which is solely about IBM double
> > double being the default long double. No need to have ppc in the
> > name of the effective target.
> > Now, long_double_ibm128 effective target is not what you want to use
> > because that isn't what the long double is, but what long double can be if
> > one dg-add-options long_double_ibm128.
> > So, perhaps long_double_is_ibm128 effective target with
> > #ifndef __LONG_DOUBLE_IBM128__
> > #error ...
> > #endif
> > in the test?
>
> So basically you're saying v4[1] with the name as long_double_is_ibm128
> instead of ppc_default_long_double_ibm?
And some comment adjustment.
# Check if the PPC target defaults to the IBM long double format.
changed to maybe
# Check if long double on the target defaults to the IBM extended format.
Inside of it it surely can just filter out non-powerpc targets like done in
the patch because we know no other targets uses that format for long double
(at least right now). And hopefully no target does that in the future ;)
Jakub
