> If you think the soft-float FP is slow, emulating the FP instructions as
> we did originally was even slower.
> Certainly there is a significant speed increase that could still be
> achieved by coding an assembly version of the softfloat library.
> 				Mark

Or one could code the most used instructions as inline, such as
add, sub, and multiply. One could also make a floating point format
optimized for the StrongARM, say 32 bit exponent and 32 bit mantissa.
This could be called 'float' and have inline code, as opposed to 'double'
which would be for accuracy and not speed.

This was my project a couple of years ago, but was not done, due
to unaivailability of a compilable gcc. I also started implementing it
for acorn C++. That compiler could not handle the massive amount of
operator overloading necessary.

Anyway, to make the add,sub,&mul inline, they could be defined
as multiple word instructions, which is supported in gcc.

Kim0