From mboxrd@z Thu Jan 1 00:00:00 1970 Return-Path: Received: by sourceware.org (Postfix, from userid 48) id 625623858D33; Thu, 9 Mar 2023 12:56:23 +0000 (GMT) DKIM-Filter: OpenDKIM Filter v2.11.0 sourceware.org 625623858D33 DKIM-Signature: v=1; a=rsa-sha256; c=relaxed/relaxed; d=gcc.gnu.org; s=default; t=1678366583; bh=nrLYqmlSgeKEXpqrKTvShyYYDYIefNziJN3vss93c74=; h=From:To:Subject:Date:In-Reply-To:References:From; b=iRN9VhoB59ooARJfyCVwVAoFZDAY19SabIRToW+qm6q8R1EvL9kOiP+b+AEDdWIb5 kd//gKWdSywnjWg+5/PihApTQomko8dTr17DHSePji984Jf87ZhqaLBabh8ONx0n7C jaHUzadYGq/gvLBBTw7JAN/D0OfkNITY5gz0WSXY= From: "jakub at gcc dot 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: Thu, 09 Mar 2023 12:56:22 +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: jakub at gcc dot gnu.org X-Bugzilla-Status: RESOLVED X-Bugzilla-Resolution: FIXED 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: attachments.created Message-ID: In-Reply-To: References: 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: https://gcc.gnu.org/bugzilla/show_bug.cgi?id=3D109008 --- Comment #45 from Jakub Jelinek --- Created attachment 54623 --> https://gcc.gnu.org/bugzilla/attachment.cgi?id=3D54623&action=3Dedit gcc13-pr109008-3.patch Further incremental patch to do the .5ulp widening rather than 1ulp if not -frounding-math and not for IBM double double. So far I've tested the + and - 300000 tests, still no errors (so conservati= vely correct) and: 1 0 50 1 1 11 1 11 1 1 1 13 1 1 16 1 1 25 1 1 28 1 1 3 1 1 36 1 1 4 1 1 40 1 19 1 1 39 1 1 50 0 2 1 17 2 1 2 2 1 32 2 1 7 2 2 1 2 5 1 3 3 1 3 4 1 38 50 1 56 1 50 71 1 0 78 0 1 160 50 50 13173 0 0 130096 -1 -1 154232 1 1 so significantly better than the 1ulp version, over 50% of cases are 1ulp a= way in each direction, over 43% of other cases have really empty range, over 4% exact and rest is noise. Let me finish * and / testing. For the forward op, we'd need to do something like in the #c30 patch, which= was ugly. And regarding !MODE_HAS_INFINITIES, the min/max values really act there like infinities, so I think we should just detect that case and use infinities in the widened range even when the mode doesn't have them. But, I'm afraid I don't know how to test that...=