Fair enough.

How's this?

Tested on x86-64 Linux.  LAPACK regression testing as well.

On Mon, Nov 7, 2022 at 1:56 PM Jakub Jelinek <jakub@redhat.com> wrote:
>
> On Mon, Nov 07, 2022 at 01:48:28PM +0100, Aldy Hernandez wrote:
> > On Mon, Nov 7, 2022 at 1:43 PM Jakub Jelinek <jakub@redhat.com> wrote:
> > >
> > > On Mon, Nov 07, 2022 at 01:35:35PM +0100, Aldy Hernandez wrote:
> > > > Let me see if I understand you correctly...
> > > >
> > > > real_isdenormal is always returning false for our uses in frange?  So
> > > > instead of using real_isdenormal in flush_denormals_to_zero, perhaps
> > > > we should be using:
> > > >
> > > > REAL_EXP (r) < REAL_MODE_FORMAT (mode)->emin
> > > >
> > > > ??
> > > >
> > > > Could we perhaps make real_isdenormal always work for all values?  Is
> > > > that possible?
> > >
> > > Yes.  Or have real_isdenormal_target for the uses in real.cc
> > > which would be private to real.cc and would do what real_isdenormal
> > > currently does, and then real_isdenormal with the above definition
> > > (well, r->cl == rvc_normal && ...).
> >
> > If the REAL_EXP(r)... definition works for everyone, can't we just use
> > that?  Or do you think there'd be a measurable performance impact?
>
> It doesn't work alone.
>
> Either denormals are in normalized form, then
> r->cl == rvc_normal && REAL_EXP (r) < REAL_MODE_FORMAT (mode)->emin
> is true, or they are in denormal form, then
> r->cl == rvc_normal && (r->sig[SIGSZ-1] & SIG_MSB) == 0
> is true.
>
> You could use
> r->cl == rvc_normal
> && (REAL_EXP (r) < REAL_MODE_FORMAT (mode)->emin
>     || (r->sig[SIGSZ-1] & SIG_MSB) == 0)
> but that would be waste of compile time both in real.cc and outside of
> real.cc.
>
>         Jakub
>