Re: [PATCH] middle-end & C-frontend: fix PR middle-end/32668

["Richard Guenther" <richard dot guenther at gmail dot com>]

On 7/8/07, Kaveh R. GHAZI <ghazi@caip.rutgers.edu> wrote:
> This patch addresses PR middle-end/32668, which is a problem with
> type-generic builtins.  Because they are implemented with variadic
> prototypes, these builtins promote float to double which can either cause
> pessimization and/or incorrect code generated from the middle-end.
>
> My solution (suggested by Joseph) was to create a new internal-only
> attribute to mark variadic builtins as "type generic" which notifies GCC
> not to do the default C promotions for these calls.  This was discussed in
> this thread: http://gcc.gnu.org/ml/gcc/2007-07/msg00221.html
>
> I didn't document the builtin because the only other internal-only builtin
> "no vops" wasn't document either AFAICT.  I could be convinced to document
> "type generic" if necessary, but I don't want to confuse users who can't
> use it anyway.
>
> Although we normally test patches in isolation, this patch needs my other
> recent work to really get exercised.  So I combined this patch with the
> other stuff I have submitted to make use of these type-generic builtins.
> It was bootstrapped along with these patches (with some minor conflicts
> resolved in builtins.def):
>
> isinf:
> http://gcc.gnu.org/ml/gcc-patches/2007-06/msg01103.html
>
> isfinite:
> http://gcc.gnu.org/ml/gcc-patches/2007-06/msg01211.html
>
> isnormal:
> http://gcc.gnu.org/ml/gcc-patches/2007-06/msg01292.html
>
> Check/test subnormals:
> http://gcc.gnu.org/ml/gcc-patches/2007-07/msg00063.html
>
> Add more testcases (& temporarily fix solaris fixincludes):
> http://gcc.gnu.org/ml/gcc-patches/2007-07/msg00583.html
>
> Remove solaris fixincludes hacks, rely only on builtins:
> http://gcc.gnu.org/ml/gcc-patches/2007-06/msg01478.html
>
>
> Bootstrap and regtest passed.  This can go in either before or after any
> of the above patches.  It's only when they're all combined that you notice
> the need for it in the isnormal builtin which gets more rigorously checked
> through my extra testcases.
>
> Okay for mainline?

The middle-end parts of this patch are ok.  The C frontend parts
need approval from a C frontend maintainer.

Thanks,
Richard.


> 2007-07-07  Kaveh R. Ghazi  <ghazi@caip.rutgers.edu>
>
>         PR middle-end/32668
>
>         * builtin-attrs.def (ATTR_TYPEGENERIC,
>         ATTR_CONST_NOTHROW_TYPEGENERIC): New.
>
>         * builtins.def (BUILT_IN_ISINF, BUILT_IN_ISNAN,
>         BUILT_IN_ISGREATER, BUILT_IN_ISGREATEREQUAL, BUILT_IN_ISLESS,
>         BUILT_IN_ISLESSEQUAL, BUILT_IN_ISLESSGREATER,
>         BUILT_IN_ISUNORDERED): Use ATTR_CONST_NOTHROW_TYPEGENERIC.
>
>         * c-common.c (handle_type_generic_attribute): New.
>         (c_common_attribute_table): Add "type generic".
>
>         * c-typeck.c (convert_arguments): Handle "type generic" functions.