This is the mail archive of the
libc-ports@sources.redhat.com
mailing list for the libc-ports project.
Re: [COMMITTED] hppa: Update libm-test-ulps.
- From: "Carlos O'Donell" <carlos at redhat dot com>
- To: "Joseph S. Myers" <joseph at codesourcery dot com>
- Cc: "libc-ports at sourceware dot org" <libc-ports at sourceware dot org>
- Date: Wed, 15 May 2013 17:12:47 -0400
- Subject: Re: [COMMITTED] hppa: Update libm-test-ulps.
- References: <5193B7DD dot 4090600 at redhat dot com> <Pine dot LNX dot 4 dot 64 dot 1305152018350 dot 21321 at digraph dot polyomino dot org dot uk> <Pine dot LNX dot 4 dot 64 dot 1305152055160 dot 21321 at digraph dot polyomino dot org dot uk>
On 05/15/2013 04:59 PM, Joseph S. Myers wrote:
> On Wed, 15 May 2013, Joseph S. Myers wrote:
>
>> It's not *new* here, but ulps for ceil, floor, rint, round, trunc indicate
>> you've got bugs in those functions for long double; they shouldn't have
>> any ulps at all. And the ulps for llrint, llround also shouldn't be
>> there, and include negative values, which should never happen even when
>> the functions are buggy.
>
> In fact, why are these tests being run for long double at all - why are
> there any long double ulps (presuming you truncated the file and
> regenerated from scratch, as you should at least once per release cycle)?
I've never regenerated from scratch. We have not recommended that in
Regeneration or the release process. I always regenerate incrementally
from the existing file. Though I see your point that we should probably
make this mandatory.
> You have long-double-fcts = no in your makefile. Contradicting that, you
> have #undef __NO_LONG_DOUBLE_MATH in ports/sysdeps/hppa/fpu/bits/mathdef.h
> - if you really have long double same as double, that should be defined to
> 1, not undefined. I'm not sure the combination of long-double-fcts = no
> with #undef __NO_LONG_DOUBLE_MATH is expected to do anything sensible.
Then that's a bug. We want to have long double the same as double.
> (However, I'd consider it inadvisable to fix the __NO_LONG_DOUBLE_MATH
> definition before you've set up ABI test baselines for hppa and verified
> them against old binaries of old releases - that sort of change has
> significant risk of causing unintended changes to symbols in past ABI
> versions.)
Agreed.
Cheers,
Carlos.