Wed Jul 8 12:45:00 GMT 2009
On Tuesday 07 July 2009 05:46:34 pm Jeff Johnston wrote:
> Ken Werner wrote:
> > On Friday 03 July 2009 07:19:59 pm Jeff Johnston wrote:
> >> Hi Craig,
> >> Sorry for the delay.
> >> Patch has a problem. It will break spu math library. The spu will
> >> define isnan and isinf to not support float values which will override
> >> the definitions in math.h. Don't know why SPU has those macros in
> >> machine/ieeefp.h but they should be rewritten to match the ones in
> >> math.h or removed entirely.
> >> -- Jeff J.
> > Hi,
> > Why does this patch break the SPU math library?
> > The reason that the isinf and isnan macros are overwritten by the
> > machine/ieeefp.h is that these are evaluated at compile time and thus
> > faster than the math.h variant which calls fpclassify. They could be
> > rewritten to use fpclassify instead of __isinfd/__isnand but we don't
> > have an optimized __fpclassifyd routine like our vectorized isnan code.
> > -ken
> This breaks the SPU math library because the override macros only work
> for double arguments.
> They return 0 if the value is a float. Craig's patch has changed the
> float math routines to use isnan/isinf
> instead of __isnanf/__isinff which is fine if one uses the macro in
> math.h, but in SPU's case they will never properly detect a float nan or
> float infinity.
> -- Jeff J.
As far as I understand, the following happens when the patch is applied:
The machine/ieeefp.h header defines the isinf, isnan, __ieeefp_isnanf and
__ieeefp_isinff macros. The isinf and isnan macros expand to zero if the value
is a float. Otherwise they expand to __isinfd(__x) and __isnand(__x)) (the
prototypes are in math.h). The machine header is included by the ieeefp.h
which (unfortunately) undefines isnan and isinf and contains prototypes for the
legacy functions isnan, isinf, isnanf and isinff. If the user includes math.h
the isinf/isnan macros are used from machine/ieeefp.h.
The only concern I have is that as soon as a user includes ieeefp.h, the
legacy isnan/isinf functions will be used, and they do not return zero in case
of single precision. But this hasn't changed with Craigs patch.
More information about the Newlib