>We could provide
>users with two version of libms; one which uses 68881 instructions,
>the other which does not.  Those with FPU-less machines should use the
>latter libm, whereas those with FPU can use either one (of course, the
>first libm is faster, so no one with a real FPU would want to use the
>latter).

I think that using the complex 68881 instruction on 040 is (at least
usually) slower than using a software math lib which assumes just the
simple floating point operations which are in the 040 FPU hardware. So
what you propose might be a win for people with 040.

I suppose that for most binaries libm is shared so switching the math
library does not even mean recompilation, so I think your proposition
is very reasonable. Maybe someone could hack the loader to load the
right shared libm depending on the HW.