>The compiler cannot compile libc's softfloat.c in the way that it was 
>written (it exploited a bug in gcc 2.7 to return a structure that was 
>larger than 32 bits in a register) -- I'm working on a fix for this, but 
>haven't completed it yet.

For Linux I'm currently using a later version of softfloat.c that uses `long 
long' to hold double values.  This avoids the structure return problem but 
introduces some mess of its own.

p.