You have got improvements as you compiled for hardfloat (so all the floating point in the code you're compiling uses the hardware rather than software).

You may just get a bit more of an improvement by doing any libm functions (trig, sqrt, all those sort of things) in hardware floating point rather than software too.

Yes, link against libm.a.


Simon

P.S. I should add that soft-float and hardfloat can be mixed (at the moment on the n800), which is why you can mix a hardfloat quake binary and a softfloat libm, etc.