Re: MATH_EMULATE

["Aaron J. Grier" <agrier%poofygoof.com@localhost>]

On Sun, Jan 13, 2008 at 05:36:04PM -0500, Perry E. Metzger wrote:
> 
> Vit Herman <v.herman%sh.cvut.cz@localhost> writes:
> > I understand that, but - are we really sure that it's MATH_EMULATE
> > what is broken? If I understood it correctly, the failing ld.so.conf
> > mechanism for switching libm versions could cause this problem as
> > well. Correct me if I'm mistaken.
> 
> It is possible that this is the issue, though we haven't confirmed
> that yet.

MATH_EMULATE has never supported emulation of fsqrt, and I doubt ever
will.  the compiler has started to emit un-emulated floating point
instructions when it previously did not.  it is likely that the kernel
floating point emulation is fine.

this has been a problem since at least netbsd-3.  a fully soft-float
userland would go some way towards addressing the issue.  the other
alternative is to limit gcc to only generate floating point that the
kernel can emulate, unless explicitly directed otherwise.

-- 
  Aaron J. Grier | "Not your ordinary poofy goof." | 
agrier%poofygoof.com@localhost