> I have this idea for almost maximally fast floating point math.
> 32 bit mantissa, and 32 bit exponent, signed int that is.

That is actually one of the internal formats that I was considering.
I just need to convince myself that the signed mantissa does not overly
complicate any of the exception handling or roundoff algorithms.

However, I think we have to respect the IEEE storage format. Once I have
a full IEEE implementation working, I want to investigate adding a "fast
and loose" option that strips out the full IEEE features and saves a lot
of time by skipping the NaN checks and so on. Then it might be possible
to cut down on the number of packing and unpacking operations, but this
is still just a sketchy idea.

What I will probably do is settle for inlining a lot of the simpler
algorithms, and try to handle the common cases in straightline assembly.
Infrequent fixup code will probably be punted to C code since their speed
doesn't matter nearly as much, and there might be other processors that
could use this.

Regardless, I will continue to maintain a "reference" C implementation
of the same algorithms which will be tested alongside the assembly versions.
In fact, I was planning on tuning that first, and then making the initial
assembly functions from the -S output of the reference C code...

Todd Whitesel
toddpw @ best.com