This is the mail archive of the
gcc-patches@gcc.gnu.org
mailing list for the GCC project.
Re: [PATCH] Don't fold nextafter/nexttoward if -ftrapping-math or -fmath-errno if they produce denormal results (PR c/86420)
- From: Richard Biener <rguenther at suse dot de>
- To: gcc-patches at gcc dot gnu dot org
- Cc: Jakub Jelinek <jakub at redhat dot com>
- Date: Mon, 9 Jul 2018 11:49:01 +0200 (CEST)
- Subject: Re: [PATCH] Don't fold nextafter/nexttoward if -ftrapping-math or -fmath-errno if they produce denormal results (PR c/86420)
- References: <20180707080605.GP7166@tucnak> <alpine.DEB.2.21.1807071132260.19203@stedding.saclay.inria.fr>
On Sat, 7 Jul 2018, Marc Glisse wrote:
> On Sat, 7 Jul 2018, Jakub Jelinek wrote:
>
> > 2018-07-07 Jakub Jelinek <jakub@redhat.com>
> >
> > PR c/86420
> > * real.c (real_nextafter): Return true if result is denormal.
>
> I have a question on the side: would it be hard / useful, in cases where
> nextafter may set errno or some exception flag, to fold the result to a
> constant while keeping the function call (ignoring the value it returns)? To
> clarify, I mean replace
>
> _2 = nextafter(DBL_DENORM_MIN, 0);
>
> with
>
> nextafter(DBL_DENORM_MIN, 0);
> _2 = 0;
>
> I think we already do that for some other calls, although I can't remember
> where. The point would be that we have the value of _2 and can keep folding
> its uses.
There's tree-call-dce.c which is doing a related (more complex) transform
but nothing doing constant propagation through calls. I think it would
be more useful for the target C library to expose sth like
set_errno (...) so we can constant fold the errno setting as well.
Maybe there's some cheap non-DCEable math function we could abuse...
Richard.