From mboxrd@z Thu Jan  1 00:00:00 1970
Return-Path: <gcc-bugzilla@gcc.gnu.org>
Received: by sourceware.org (Postfix, from userid 48)
	id 6D75A3858C66; Tue,  7 Mar 2023 13:29:21 +0000 (GMT)
DKIM-Filter: OpenDKIM Filter v2.11.0 sourceware.org 6D75A3858C66
DKIM-Signature: v=1; a=rsa-sha256; c=relaxed/relaxed; d=gcc.gnu.org;
	s=default; t=1678195761;
	bh=LH3mkiYQ7jJtKw+Mi7uVOmUngcEiltVI/AHb8Ynu4q4=;
	h=From:To:Subject:Date:In-Reply-To:References:From;
	b=aQFRehxqQ1rksXYZmwy/7bI+NP1kscgym0ZLUL6onifo1w8a180fLdQeaQk+kPhMP
	 UqSmguwJCZ1z5hY3Jz0A9FVimVt+keD6+w+e+rLLMiYhQe1j2olB2GvPOo0f2JDzsB
	 W+RjZCOm7IxtrqQlpoCdl3A3sgzgVXm2q0sXb+rY=
From: "rguenth at gcc dot gnu.org" <gcc-bugzilla@gcc.gnu.org>
To: gcc-bugs@gcc.gnu.org
Subject: [Bug tree-optimization/109008] [13 Regression] Wrong code in scipy
 package since r13-3926-gd4c2f1d376da6f
Date: Tue, 07 Mar 2023 13:29:21 +0000
X-Bugzilla-Reason: CC
X-Bugzilla-Type: changed
X-Bugzilla-Watch-Reason: None
X-Bugzilla-Product: gcc
X-Bugzilla-Component: tree-optimization
X-Bugzilla-Version: 13.0
X-Bugzilla-Keywords: wrong-code
X-Bugzilla-Severity: normal
X-Bugzilla-Who: rguenth at gcc dot gnu.org
X-Bugzilla-Status: NEW
X-Bugzilla-Resolution: 
X-Bugzilla-Priority: P1
X-Bugzilla-Assigned-To: unassigned at gcc dot gnu.org
X-Bugzilla-Target-Milestone: 13.0
X-Bugzilla-Flags: 
X-Bugzilla-Changed-Fields: 
Message-ID: <bug-109008-4-0q1zyBoVKx@http.gcc.gnu.org/bugzilla/>
In-Reply-To: <bug-109008-4@http.gcc.gnu.org/bugzilla/>
References: <bug-109008-4@http.gcc.gnu.org/bugzilla/>
Content-Type: text/plain; charset="UTF-8"
Content-Transfer-Encoding: quoted-printable
X-Bugzilla-URL: http://gcc.gnu.org/bugzilla/
Auto-Submitted: auto-generated
MIME-Version: 1.0
List-Id: <gcc-bugs.sourceware.org>

https://gcc.gnu.org/bugzilla/show_bug.cgi?id=3D109008

--- Comment #20 from Richard Biener <rguenth at gcc dot gnu.org> ---
(In reply to Richard Biener from comment #19)
> I still think we should avoid iteration.  Looking at plus we have
>=20
>  x =3D y + a
>=20
> which is actually
>=20
>  x =3D y + a +- 0.5ulp (y + a)
>=20
> (0.5ulp of the y + a result), so we can compute a as
>=20
>  a =3D x - y +- 0.5ulp (y + a)
>=20
> note 0.5ulp (y + a) isn't 0.5ulp (x), not exactly at least, so we have to
> approximate it and a conservative bound is 1ulp (x) here.  We may also ha=
ve
> to make sure to round toward -Inf for the lower bound of the result range
> and +Inf of the upper bound of the result range I think.
>=20
>=20
> For multiplication it's
>=20
>  x =3D y * a +- 0.5ulp (y *  a)
>=20
>  a =3D (x +- 0.5ulp (y * a)) / y
>=20
> so it's again similar.
>=20
> So what we can indeed do is widen the LHS range by 0.5 ulp and since we
> cannot represent that and it might be imprecise if 1 ulp _after_ the
> rounding is smaller than 1 ulp _before_ the rounding operation we simply
> use 1 ulp widening of the LHS?

So in principle REAL_VALUE_TYPE would support adjusting by half an ulp,
we just cannot use real_nextafter for that but we'd have to do this
manually somehow and the real_convert after could in principle also do
the -+INF rounding (but that's not implemented yet).=