From: Rich Felker <dalias@libc.org>
To: musl@lists.openwall.com
Subject: Re: [PATCH] fix underflow exception in fma and fmal
Date: Sun, 19 Mar 2017 10:53:40 -0400 [thread overview]
Message-ID: <20170319145340.GS1693@brightrain.aerifal.cx> (raw)
In-Reply-To: <20170319143953.GQ2082@port70.net>
On Sun, Mar 19, 2017 at 03:39:53PM +0100, Szabolcs Nagy wrote:
> * Rich Felker <dalias@libc.org> [2017-03-19 10:12:49 -0400]:
> > On Sun, Mar 19, 2017 at 04:36:14AM +0100, Szabolcs Nagy wrote:
> > > another corner case in the freebsd fma code where signaling underflow
> > > may be missed for an inexact subnormal result.
> > > (fmaf and x86 fma are not affected)
> > > ---
> > > src/math/fma.c | 7 +++++++
> > > src/math/fmal.c | 8 ++++++++
> > > 2 files changed, 15 insertions(+)
> > >
> > > diff --git a/src/math/fma.c b/src/math/fma.c
> > > index 741ccd75..c69918d1 100644
> > > --- a/src/math/fma.c
> > > +++ b/src/math/fma.c
> > > @@ -279,6 +279,13 @@ static inline double add_and_denormalize(double a, double b, int scale)
> > > uhi.i += 1 - (((uhi.i ^ ulo.i) >> 62) & 2);
> > > sum.hi = uhi.f;
> > > }
> > > +#ifdef FE_UNDERFLOW
> > > + /*
> > > + * Raise underflow manually because scalbn won't do it if all
> > > + * lost bits are 0: fma(-0x1p-1000, 0x1.000001p-74, 0x1p-1022)
> > > + */
> > > + feraiseexcept(FE_UNDERFLOW);
> > > +#endif
> >
> > Can you explain why it should happen if all lost bits are zero
> > (usually that's an exact case). I imagine it's something specific to
> > fma or its implementation but it's not obvious to me.
> >
>
> this case is for nearest rounding mode when the
> result is in the subnormal range, at this point the
> result is represented as hi,lo,scale but the final
> returned value is computed as scalbn(hi,scale)
> (the last bits of hi are adjusted if required for
> correct rounding), however scalbn fails to raise
> underflow if lo!=0 and all lost bits of hi are 0.
>
> the example is such a case: 0x1p-1022 - 0x1.000001p-1074
> then hi=1-eps,lo=-0x1p-76,scale=-1022 or maybe with
> shifted scale and exponents, but in the end only one
> bit is lost from hi which is zero, alternatively i
> could do scalbn(lo,scale) too to raise underflow.
That makes sense. I tend to prefer the scalbn(lo,scale) approach if
there aren't good reasons (performance?) against it, simply because
it's more self-documenting and less special-cased. But whichever you
like is fine. BTW we should probably check that scalbn raises inexact
in all cases it should; I'm not sure what it (especially asm versions)
does in cases where the scale is smaller than the min exponent.
Rich
next prev parent reply other threads:[~2017-03-19 14:53 UTC|newest]
Thread overview: 5+ messages / expand[flat|nested] mbox.gz Atom feed top
2017-03-19 3:36 Szabolcs Nagy
2017-03-19 14:12 ` Rich Felker
2017-03-19 14:39 ` Szabolcs Nagy
2017-03-19 14:53 ` Rich Felker [this message]
2017-03-19 17:06 ` Szabolcs Nagy
Reply instructions:
You may reply publicly to this message via plain-text email
using any one of the following methods:
* Save the following mbox file, import it into your mail client,
and reply-to-all from there: mbox
Avoid top-posting and favor interleaved quoting:
https://en.wikipedia.org/wiki/Posting_style#Interleaved_style
* Reply using the --to, --cc, and --in-reply-to
switches of git-send-email(1):
git send-email \
--in-reply-to=20170319145340.GS1693@brightrain.aerifal.cx \
--to=dalias@libc.org \
--cc=musl@lists.openwall.com \
/path/to/YOUR_REPLY
https://kernel.org/pub/software/scm/git/docs/git-send-email.html
* If your mail client supports setting the In-Reply-To header
via mailto: links, try the mailto: link
Be sure your reply has a Subject: header at the top and a blank line
before the message body.
Code repositories for project(s) associated with this public inbox
https://git.vuxu.org/mirror/musl/
This is a public inbox, see mirroring instructions
for how to clone and mirror all data and code used for this inbox;
as well as URLs for NNTP newsgroup(s).