numeric fdlibm

[Ian Lance Taylor]

"David G Hough at validlab.com" <validlabavalidlab.com> writes:

> A couple of years ago Stephen Moshier wrote:
>  
>  It is well known that fdlibm as originally posted was riddled with
>  pointer aliasing bugs and other portability problems.  You can get the
>  benefit of other people's debugging efforts by looking at the Cygnus,
>  DJGPP, linux glibc, or perhaps BSD versions.  The Cygnus one is at
>    ftp://sourceware.cygnus.com/pub/newlib/
>  DJGPP is
>    http://www.delorie.com
>  Both of these mentioned come with very permissive license boilerplate.
>  
> Is this still the current status?    I'm looking for a version of fdlibm
> that preserves the original numerical algorithms but
> compiles correctly with full optimization on a variety of platforms
> with current versions of GCC (e.g. 3.1.1).   The original libfdm problem
> with optimizers -
> finding an efficient way to access doubles as pairs of 32-bit ints -
> has been said to be insoluble since the C standard doesn't GUARANTEE any
> of the usual ways of doing so by pointers or unions.     Even if that be
> true, I'd be happy for a version that works anyway at least on SPARC.

I did a bunch of work on eliminating this non-standard C usage from
fdlibm back in 1994 or so.  I also added single precision versions of
the routines, but I didn't change the double precision routines.  This
work can be found at
    http://sources.redhat.com/newlib
in the newlib/libm/math directory.  (sources.redhat.com is the
successor to sourceware.cygnus.com).  I don't know how much the code
may have been changed in the past couple of years, though.

Ian