This is the mail archive of the
gcc@gcc.gnu.org
mailing list for the GCC project.
Re: GCC 8.0.0 Status Report (2018-01-15), Trunk in Regression and Documentation fixes only mode
- From: Vincent Lefevre <vincent+gcc at vinc17 dot org>
- To: gcc at gcc dot gnu dot org
- Date: Tue, 6 Feb 2018 17:28:31 +0100
- Subject: Re: GCC 8.0.0 Status Report (2018-01-15), Trunk in Regression and Documentation fixes only mode
- Authentication-results: sourceware.org; auth=none
- References: <alpine.LSU.2.20.1801150920260.32271@zhemvz.fhfr.qr> <962d449e-3ab8-f9ad-7a2d-8793ae0188d6@sky.com> <CAFiYyc0gJqTSsJmLihLozH4_DtJ24OWLm4pRzF2TMnGpcyS2tg@mail.gmail.com> <8441a3a4-de80-cf19-f9fe-2251fb9064e2@ubuntu.com> <20180206161438.GA10578@zira.vinc17.org>
On 2018-02-06 17:14:38 +0100, Vincent Lefevre wrote:
> Now, I've just found a "regression" when comparing Sipe results with
> MPFR results, but it is the Sipe result with SIPE_FLOAT equal to 1
> (float) or 2 (double) that is incorrect, 3 (long double) being OK.
> Bug triggered with -O2 -march=native on an Intel Xeon E5-2609 v3
> machine. At least from 4.9 to 8.0.1 20180124 [trunk revision 257009]
> are affected. I'll try to find a simple test case for a bug report.
> I suspect that the regression comes from the fact this is a new
> machine + the use (as in the past) of -march=native.
Well, if I add -std=c99 or -std=c11, the failure disappears. The
reason is that GCC generates a FMA (indeed, this new machine has
a FMA), but this occurs across different C statements. So, maybe
not a bug, but a bad feature, IMHO.
--
Vincent Lefèvre <vincent@vinc17.net> - Web: <https://www.vinc17.net/>
100% accessible validated (X)HTML - Blog: <https://www.vinc17.net/blog/>
Work: CR INRIA - computer arithmetic / AriC project (LIP, ENS-Lyon)