[Bug tree-optimization/122103] [15 Regression] SVE vectorisation not honoring -ftrapping-math

[ktkachov at gcc dot gnu.org]

https://gcc.gnu.org/bugzilla/show_bug.cgi?id=122103

--- Comment #21 from ktkachov at gcc dot gnu.org ---
(In reply to ktkachov from comment #20)
> (In reply to Tamar Christina from comment #19)
> > (In reply to rguenther@suse.de from comment #18)
> > > On Mon, 5 Jan 2026, tnfchris at gcc dot gnu.org wrote:
> > > 
> > > > https://gcc.gnu.org/bugzilla/show_bug.cgi?id=122103
> > > > 
> > > > Tamar Christina <tnfchris at gcc dot gnu.org> changed:
> > > > 
> > > >            What    |Removed                     |Added
> > > > ----------------------------------------------------------------------------
> > > >             Summary|[15/16 Regression] SVE      |[15 Regression] SVE
> > > >                    |vectorisation not honoring  |vectorisation not honoring
> > > >                    |-ftrapping-math             |-ftrapping-math
> > > > 
> > > > --- Comment #16 from Tamar Christina <tnfchris at gcc dot gnu.org> ---
> > > > Fixed on trunk so far, not sure if backporting makes sense as it's pretty much
> > > > never been correct.
> > > 
> > > It's a bit large of a change and definitely needs soaking time, so 
> > > possibly only for GCC 15.3, but even maybe not to that.
> > 
> > Agreed. A middle ground might be to only backport the gimple_could_trap and
> > if-cvt change. This would prevent the wrong codegen since the bug was
> > reported on GCC 15.
> > 
> > But indeed probably should be revisited once we see what falls out from the
> > change.
> 
> Fixing it in GCC 15.3 would be great if the risk is acceptable. The user
> reported they can live with a GCC 16-only fix too, but it'd be good to have
> it in the 15-series too

(Thank you Tamar very much for the fix btw)