bjh21@netbsd.org said:
> I was referring to the case where the code uses -msoft-float, but libc
> implements the soft-float routines in terms of FPA instructions.  In
> this case, every FP op gets its arguments from integer registers and
> puts its results in them, so the added overhead of word-swapping
> doubles is small.

Ah!  Misunderstood what you were saying.  Yes, this is quite feasible, 
especially since all those functions could easily be written as assembler 
stubs.

> If you're going to arrange to make it possible to link -msoft-float
> and -mhard-float code together then things will obviously get painful.
>  This won't work at present (they maintain FP exception stuff in
> different places if nothing else), and I don't think it's a terribly
> important thing to have. 

Good. I'm happy to agree on that.

R.