Improve tgamma accuracy (bug 18613) [committed]

[Szabolcs Nagy]

On 02/07/15 14:57, Joseph Myers wrote:
> On Thu, 2 Jul 2015, Szabolcs Nagy wrote:
>> On 02/07/15 13:58, Szabolcs Nagy wrote:
>>>
>>> hm i tried to verify this on aarch64 and also regenerate the libm ulps
>>>
>>> but the new ulps file somehow become much smaller than expected
>>>
>>
>> it was because builddir/math/libm-test-ulps.h was non-empty
>>
>> (it is generated at glibc build time not at make regen-ulps
>> time as i assumed)
> 
> I don't think it *should* be generated at build time (the only thing that 
> needs ulps information at build time should be the table in the manual), 
> but don't know offhand what dependency is causing it to be generated 
> early.  Also, even if generated early, it ought not to break regen-ulps 
> for the header to be already present - so that sounds like a makefile 
> issue as well.
> 

i think it is generated at build time because of

# This is needed for dependencies
before-compile += $(objpfx)libm-test.c

in math/Makefile and then libm-test.c depends on libm-test.stmp

after building glibc i renamed the libm-test-ulps in the
source dir (instead of truncating it) so make regen-ulps
could not pick up the changed time stamp i guess.

anyway, i can confirm that the tgamma_upward errno failures
happen on aarch64 too.