Draft "Unsafe fp optimizations" project description.

[dewar@gnat.com]

<<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?
>>

Well I don't think anyone will put up with the compiler doing computations
in sp which the source program explicitly does in dp. However, in C, there
are cases where promotion is required even though in a certain sense it is
not really mandated by the plain language of the source, and I would think
that this is discussable in the -ffast-math context.