On Fri, Aug 29, 2025 at 10:56 AM Matteo Nicoli <matteo.nicoli...@gmail.com> wrote: > > Dear Richard, > > > It can trap with sNaN ± 0.0. ±Inf ± 0.0 = ±Inf, so it does not raise an > FE_OVERFLOW (because there’s no overflow of a finite quantity), and qNaN does > not raise an FE_INVALID because it’s quiet. > > > There’s already a check for sNaN in fold-const.cc > > > /* Don't allow the fold with -fsignaling-nans. */ > if (arg ? tree_expr_maybe_signaling_nan_p (arg) : HONOR_SNANS (type)) > > return false; > > > As far as I know, the purpose of this bug fix was to suppress this specific > optimization when the program is compiled with -ftrapping-math flag.
So we've discussed on IRC and the conclusion was that with default exception handling !HONOR_SNANS would be sufficient but alternate exception handling might trap on x + 0.0 when x is denormal. Ideally we'd have a separate flag for non-default exception handling, but as-is we don't. Please add a comment like /* With non-default exception handling denormal + 0.0 might trap. Otherwise !HONOR_SNANS would be sufficient here. */ The patch is OK with that change. Richard. > > Best regards, > > Matteo > > > > > On Aug 28, 2025, at 10:43 AM, Richard Biener <richard.guent...@gmail.com> > wrote: > > On Sat, Aug 23, 2025 at 11:56 PM Matteo Nicoli > <matteo.nicoli...@gmail.com> wrote: > > > Dear reviewers, > > I attached a patch for bug 121595 > (https://gcc.gnu.org/bugzilla/show_bug.cgi?id=121595). I signed it, and added > `Reviewed-by: Andrew Pinski <andrew.pin...@oss.qualcomm.com>` (here in CC). > > > +/* fabs(x + 0.0) -> fabs(x), safe even with signed zeros when > -fno-trapping-math. */ > +(for op (plus minus) > + (simplify > + (abs (op @0 real_zerop@1)) > + (if (!flag_trapping_math) > + (abs @0)))) > > so forgive my ignorance, possibly IEEE abs() never raises FP exceptions > (unless operating on sNaN?)? But does Inf + 0.0 raise FE_OVERFLOW? > Does NaN + 0.0 raise FE_INVALID? > > So what I wonder is whether !HONOR_SNANS (@0) would be enough to check? > > I refrained from trusting AI on those questions ... > > Richard. > > Best regards, > Matteo > > >
