>       register _res = 0;
> 
> That might work.

I tried that, and gcc spits out code that doesn't work then.  I'm
not sure if gcc is broken in that case, or if there is something
about that initialization that breaks the assembly that follows.

Right now I suspect GCC, as I had the assembly code in one __asm__
directive per instruction, and gcc put a branch label in between the two
asm directives, even though there is no conditional code in the inline
function.

If GCC isn't broken, it's not following the "path of least suprise"
either.

Jon