- Rewritten to use SoftFloat instead of unportable native FP math
- Support added for Motorola 96-bit extended floats
- More addressing modes and conditions supported
The taitojc games function identically to the old implementation (I stepped
through the main matrix multiply in dendeg and the registers never diverged)
and speed on Core 2 Duo is a wash - sometimes softfloat's faster, sometimes
not, but the difference on -str 90 never exceeded 1%.
