This is the mail archive of the glibc-bugs@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]

[Bug ports/2749] powerpc32 does not build --without-fp


------- Additional Comments From joseph at codesourcery dot com  2006-09-15 21:28 -------
Subject: Re:  powerpc32 does not build --without-fp

On Fri, 15 Sep 2006, jakub at redhat dot com wrote:

> glibc on many arches doesn't have a standard conforming fma(3) - for float

Yes - that's what the Debian bug report I mentioned is about.  A new bug 
report should probably be opened in glibc Bugzilla for this issue.

> just using double multiplication and addition should be good enough, but

The standard says the result is rounded once, so double rounding isn't 
correct, but still a lot better than what we have right now.

> especially for long double the current implementation is wrong.  So, if there was
> a soft-fp FMA that could be used on all the targets that can't do this in
> hardware, it would be great.

To do it for long double you'd need to add more operations to op-8.h (and 
invent an oct-float type), since you need an intermediate type with twice 
the precision.  Also you'd need to add appropriate packing/unpacking for 
IBM long double (which soft-fp doesn't natively support).

We do have various sfp-machine.h files for targets not currently using 
soft-fp sitting around and ready to use as part of using soft-fp for 
generic fma (on all targets without fma instructions).

Now, we have the immediate problem that soft-fp patches are very hard to 
get into glibc (I have one unreviewed from June, bug 2831) and the same 
applies to other generic fixes to mathematical functions (bug 2592, 
likewise fix waiting since June) - and the glibc patches attached to this 
bug report have also mostly been ready for some time to go in as a clear 
improvement on the present state of affairs (with any subsequent fixes 
likely to be comparatively small and confined to either soft-fp or ports).  
So while I've considered generically fixing fma in glibc and adding 
appropriate testcases, I wouldn't be confident of such fixes actually 
getting anywhere.



-- 


http://sourceware.org/bugzilla/show_bug.cgi?id=2749

------- You are receiving this mail because: -------
You are on the CC list for the bug, or are watching someone who is.


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