[PATCH] added inputs giving large errors on x86_64 for new C23 functions

[Christophe Lyon]

On Mon, 30 Sept 2024 at 13:57, Paul Zimmermann <Paul.Zimmermann@inria.fr> wrote:
>
>        Hi Christophe,
>
> I don't have access to the artifacts.

Indeed those artifacts have been cleaned...
Looking at the most recent build of
https://ci.linaro.org/view/tcwg_gnu/job/tcwg_gnu_native_check_glibc--master-arm-build/
(so "last successful artifacts" / artifacts / 00-sumfiles)
should give you access to tests.log.xz

>
> Anyway, for exp10m1, you can try the following:
Not easy for me, maybe Adhemerval can help?

Thanks,

Christophe

>
> --- a/sysdeps/arm/libm-test-ulps
> +++ b/sysdeps/arm/libm-test-ulps
> @@ -778,7 +778,7 @@ double: 2
>  float: 1
>
>  Function: "exp10m1":
> -double: 3
> +double: 4
>  float: 2
>
>  Function: "exp10m1_downward":
> @@ -790,7 +790,7 @@ double: 2
>  float: 3
>
>  Function: "exp10m1_upward":
> -double: 4
> +double: 5
>  float: 4
>
> If this solves the issue, I can submit a patch.
>
> Paul
>
> > From: Christophe Lyon <christophe.lyon@linaro.org>
> > Date: Mon, 30 Sep 2024 13:13:43 +0200
> > Cc: Adhemerval Zanella Netto <adhemerval.zanella@linaro.org>, libc-alpha@sourceware.org
> >
> > Hi!
> >
> > On Wed, 7 Aug 2024 at 14:29, Paul Zimmermann <Paul.Zimmermann@inria.fr> wrote:
> > >
> > > > LGTM, thanks.
> > > > Reviewed-by: Adhemerval Zanella  <adhemerval.zanella@linaro.org>
> > >
> > > thank you Adhemerval for your review. Patch is applied and pushed.
> > >
> >
> > According to our CI, this patch introduces regressions on arm, when
> > using trunk binutils / gcc, but still passes when using Ubuntu Jammy
> > binutils /gcc.
> > Not sure if you received a notification after your commit, but you can
> > find more info here: https://linaro.atlassian.net/browse/GNU-1309
> >
> > As of today, I'm still seeing:
> > FAIL: math/test-double-exp10m1
> > FAIL: math/test-double-log10p1
> > FAIL: math/test-float32x-exp10m1
> > FAIL: math/test-float32x-log10p1
> > FAIL: math/test-float64-exp10m1
> > FAIL: math/test-float64-log10p1
> > FAIL: math/test-ldouble-exp10m1
> > FAIL: math/test-ldouble-log10p1
> >
> > I can see these two, but they were not part of the original regression
> > report, so probably not related.
> > FAIL: math/test-double-log2p1
> > FAIL: math/test-double-j0
> > FAIL: math/test-double-log
> > FAIL: math/test-float32x-j0
> > FAIL: math/test-float32x-log
> > FAIL: math/test-float32x-log2p1
> > FAIL: math/test-float64-j0
> > FAIL: math/test-float64-log
> > FAIL: math/test-float64-log2p1
> > FAIL: math/test-ldouble-j0
> > FAIL: math/test-ldouble-log
> > FAIL: math/test-ldouble-log2p1
> >
> >
> > For the first one (math/test-double-exp10m1), the log says:
> > original exit status 1
> > testing double (without inline functions)
> > Failure: Test: exp10m1 (0x7.a20316ebef504p-4)
> > Result:
> >  is:          1.9995335758777104e+00   0x1.ffe16eb3a6114p+0
> >  should be:   1.9995335758777095e+00   0x1.ffe16eb3a6110p+0
> >  difference:  8.8817841970012523e-16   0x1.0000000000000p-50
> >  ulp       :  4.0000
> >  max.ulp   :  3.0000
> > Maximal error of `exp10m1'
> >  is      : 4 ulp
> >  accepted: 3 ulp
> > Failure: Test: exp10m1_upward (0x7.3887ffe96bb6p-4)
> > Result:
> >  is:          1.8268422546834154e+00   0x1.d3abef1ad0f0ap+0
> >  should be:   1.8268422546834143e+00   0x1.d3abef1ad0f05p+0
> >  difference:  1.1102230246251566e-15   0x1.4000000000000p-50
> >  ulp       :  5.0000
> >  max.ulp   :  4.0000
> > Maximal error of `exp10m1_upward'
> >  is      : 5 ulp
> >  accepted: 4 ulp
> >
> >
> > Does this make sense?
> >
> > Thanks,
> >
> > Christophe
> >
> >
> >
> >
> > > Paul
> > >
> >