Draft "Unsafe fp optimizations" project description.

[Gabriel Dos_Reis]

| > Well any platform could potentially use 64-bit arithmetic in place of
| > 32-bit arithmetic, so I am not sure how you would limit the list for
| > the second case.
| 
| It wouldn't make sense on the SH3E, which has hardware single precision
| FP, but can do double precision FP only by software.
| 
| It would also be rather dubious on the SH4, where single precision is
| considerably faster than double precision.

Thanks for these data which widen the Sparc-like camp.

Now another -- some may qualify it as <censored> -- question: is it
desirable in the name of speed that the compiler chooses to use single 
precision in some of the floating-point-expression transformations?

-- Gaby