[Numpy-discussion] Solaris Sparc build broken

Charles R Harris charlesr.harris@gmail....
Sun Nov 8 23:44:03 CST 2009


On Sun, Nov 8, 2009 at 10:04 PM, David Cournapeau <
david@ar.media.kyoto-u.ac.jp> wrote:

> Charles R Harris wrote:
> >
> > I don't remember how PPC does it's arithmetic with the two doubles,
> > but one holds small values that are added to the other double, so I
> > don't think it can be just like a double in some circumstances.
>
> That's not how it seems to work on mac os x - when I look at the
> generated code for long double, the last 8 bytes were always equal to 0.
> I cannot find a definitive reference, but it looks like it is only 16
> bytes to respect the mac os x 16 bytes alignment as required by its ABI.
>
> For other OS which run ppc with a different format for long double, I
> think we could force long double to be the same as double.
> > What's rosetta?
>
> That's the ppc emulator used in mac os x to run old powerpc-only
> applications on recent Intel mac. It is not 100 % accurate (a few unit
> tests don't pass, mostly related to FPU exceptions), but it works quite
> well for testing and is reasonably fast.
>
>
This old thread leads me to believe that OS-X on PPC (G5) is actually
different:

http://www.mail-archive.com/numpy-discussion@scipy.org/ms16428.html

However, there are enough problems with the "real" PPC long double (Nans,
etc), that maybe it isn't worth trying to make it work. It isn't ieee and it
seems to be fading away in any case.

Chuck
-------------- next part --------------
An HTML attachment was scrubbed...
URL: http://mail.scipy.org/pipermail/numpy-discussion/attachments/20091108/5465d644/attachment.html 


More information about the NumPy-Discussion mailing list