ldouble math

[Bernhard Reutner-Fischer]

On Wed, Feb 18, 2009 at 08:06:17PM +0100, Denys Vlasenko wrote:
>On Wed, Feb 18, 2009 at 5:08 PM, Carmelo AMOROSO <carmelo.amoroso at st.com> wrote:
>> -----BEGIN PGP SIGNED MESSAGE-----
>> Hash: SHA1
>>
>> Ned Ludd wrote:
>>> On Wed, 2009-02-18 at 12:07 +0100, Bernhard Reutner-Fischer wrote:
>>>> On Wed, Feb 18, 2009 at 10:07:43AM +0100, Carmelo AMOROSO wrote:
>>>>> -----BEGIN PGP SIGNED MESSAGE-----
>>>>> Hash: SHA1
>>>>>
>>>>> aldot at uclibc.org wrote:
>>>>>> Author: aldot
>>>>>> Date: 2008-10-03 07:24:28 -0700 (Fri, 03 Oct 2008)
>>>>>> New Revision: 23582
>>>>>>
>>>>>> Log:
>>>>>> - add long double math wrappers (Ned Ludd)
>>>>> Hi Berhard,
>>>>> any reason I don't know for not having include TARGET_sh among
>>>>> archs supporting long double ?
>>>> I do not immediately see why there is any arch limit.
>>>>
>>>> Solar, why was this long double math support arch specific?
>>>
>>> I'm not aware of any reason why TARGET_sh would need to be filtered. It
>>> was probably an oversight.
>>>
>>> Anything break when enabled?
>>>
>>>
>> No, it builds fine. Not all tests passes, but this is another problem.
>
>If you are talking about uclibc test/math, well, it was disabled
>in Makefile for I don't know how long. I enabled it just recently.
>It doesn't pass on x86 either (but compiles, which it wasn't doing at first).

Just once again tripped ppc32 vs. long double math.
The problem there is that we currently try to reactively disable
LONG_DOUBLE_MATH for __powerpc__ &&  __WORDSIZE == 32 (by undef'ing it
in {mathdef,wordsize}.h) and that doesn't work out too well: We end up
with half enabled ldouble support and fail on __finitel and friends.

Could LONG_DOUBLE work on ppc32 in principle? ISTR that it won't, but
can somebody confirm or decline this assumption?