[PATCH v2] powerpc: Update ULPs and output for j0 with ibm128
Tulio Magno Quites Machado Filho
tuliom@linux.ibm.com
Thu Sep 10 18:52:57 GMT 2020
Joseph Myers <joseph@codesourcery.com> writes:
> The first version is appropriate. The bigger limits for ibm128 are
> deliberate. An architecture-specific ulps update should not be rejected
> just because it increases ulps within the bounds accepted for that format
> (and you shouldn't be able to get an update that increases them outside
> those bounds unless you edit the ulps file manually, and if you do edit it
> manually then the ulps it lists outside those bounds will be ignored
> anyway).
>
> The most likely reason for rejecting an ulps update is that new ulps are
> greater than for other architectures using the same code in glibc for the
> same functions for the same format, suggesting that the update is actually
> covering up a bug somewhere else. This doesn't apply in this case.
>
> (If an ulps increase results from a change to the implementation of a libm
> functions that makes it less accurate in some cases, we need to consider
> if we want that change, but that's deciding whether we want the libm
> change, not deciding whether we want the ulps updates consequent on it.)
Ack.
Thanks for clarifying!
Carlos O'Donell via Libc-alpha <libc-alpha@sourceware.org> writes:
> It was later shown to me that >9 ULPs was acceptable for ibm128, my apologies
> for not being clearer that I was withdrawing my objection.
Actually, I should be the one apologizing.
> Patsy Griffin from my team also suggested this on September 2nd, she
> is seeing these failures in our own testing. It would be good to have
> them resolved.
Especially after Patsy's message.
> Before this patch, the following tests were failing:
>
> ppc and ppc64:
> FAIL: math/test-ldouble-j0
>
> ppc64le:
> FAIL: math/test-float128-j0
> FAIL: math/test-float64x-j0
> FAIL: math/test-ibm128-j0
> FAIL: math/test-ldouble-j0
Tested and pushed as c71d13a0984f6.
Thank you all!!
--
Tulio Magno
More information about the Libc-alpha
mailing list