On Mon, Sep 5, 2016 at 1:40 PM, Joseph Myers <jos...@codesourcery.com> wrote:
> On Sat, 3 Sep 2016, Andreas Schwab wrote:
>
>> On Aug 17 2016, Joseph Myers <jos...@codesourcery.com> wrote:
>>
>> > Index: gcc/testsuite/gcc.dg/torture/float32-basic.c
>> > ===================================================================
>> > --- gcc/testsuite/gcc.dg/torture/float32-basic.c (nonexistent)
>> > +++ gcc/testsuite/gcc.dg/torture/float32-basic.c (working copy)
>> > @@ -0,0 +1,9 @@
>> > +/* Test _Float32. */
>> > +/* { dg-do run } */
>> > +/* { dg-options "" } */
>> > +/* { dg-add-options float32 } */
>> > +/* { dg-require-effective-target float32_runtime } */
>> > +
>> > +#define WIDTH 32
>> > +#define EXT 0
>> > +#include "floatn-basic.h"
>>
>> This fails on powerpc32, in vafn.
>
> That seems like the alpha issue - an ABI needs to be defined and
> implemented.
Note that this makes it maybe a good idea to only enable _FloatXX support
for targets that have explicitely done so once we near the GCC 7 release.
Otherwise we risk a later ABI change if the ABI differs from GCCs implementation
from the time there was no ABI defined for those types.
Richard.
> Unfortunately the powerabi mailing list, which would have
> been the right place for ABI coordination between implementors, died
> several years ago. (The 32-bit ABI source code is available at
> <https://github.com/ryanarn/powerabi>.)
>
> If we suppose that _Float32 values should be passed in FPRs in the same
> circumstances in which float values are passed in FPRs, then
> rs6000_gimplify_va_arg would, in the case of an SFmode value coming from
> saved FPRs, need to extract a double value from the saved FPRs and then
> convert to float. (There would also be the question of later _Float32
> arguments passed on the stack in variable arguments; if those are handled
> like prototyped arguments, they would go in 4-byte stack slots, and
> probably the compiler already does that.) Of course other ABI choices are
> possible.
>
> --
> Joseph S. Myers
> jos...@codesourcery.com
