Hi folks.
I just want to share my excitement about the software float code from Peter Teichmann.
http://www.peter-teichmann.de/download/fp.zip
I did some number crunching (inversion and matrix multiplication of a 300*300 float-matrix). The timing shows an average increase by factor 1.8. If I consider how much time is spent in housekeeping and index calculation I'd expect speedup of the raw math primitives of at least factor two.
The code needs a bit love to compile with EVC, also it does not do any special case handling like NANs and infinites. Rounding modes are ignored as far as I can see, so the results are a bit imprecise bit still good enough to invert a 300*300 matrix.
I think for cases where some floating point precision is needed this might be a very good alternative to the built in float emulation code.