This is the mail archive of the libc-alpha@sourceware.org mailing list for the glibc project.


Index Nav: [Date Index] [Subject Index] [Author Index] [Thread Index]
Message Nav: [Date Prev] [Date Next] [Thread Prev] [Thread Next]
Other format: [Raw text]

Re: Reduce various x86/x86_64 libm ulps expectations


On Wed, 29 Feb 2012, Andreas Jaeger wrote:

> Thanks for digging into this. I suggest to commit this - and if any of us that
> runs the testsuite encounters problems which then might be due to compiler or
> hardware issues, we can increase the ULPs again.

Thanks, committed.  I'd consider small ULPs increases on x86 or x86-64 (or 
m68k, given that similar issues apply there as well) to be obvious (except 
for functions that should always be correctly rounded such as rint, 
nextafter, fma, sqrt etc.).  The same probably applies where IBM long 
double is involved as well.  (For other cases, the results generally ought 
not to depend on details of the compiler or hardware so one might be more 
suspicious of an increase without any associated glibc implementation 
change to require it - but even so, if the function isn't trying to be 
correctly rounded then the increase of a few ULPs probably doesn't matter 
much.)

> Maintainers of other architectures might want to reset their ULPs files as
> well...

I've made a note to look at the larger errors listed for ARM and MIPS to 
see if they are still applicable.  Though with this patch in, the largest 
error listed for any function and target is 24ulp (there are various bugs 
filed for larger errors in various functions).

-- 
Joseph S. Myers
joseph@codesourcery.com


Index Nav: [Date Index] [Subject Index] [Author Index] [Thread Index]
Message Nav: [Date Prev] [Date Next] [Thread Prev] [Thread Next]