Third Draft "Unsafe fp optimizations" project description.

[dewar@gnat.com]

<<My point is that gcc should feel free to apply transformations that
do not change results in IEEE arithmetic, even if they
might change the results in some other arithmetic (old Cray
arithmetic, say, where x / 2.0 may not have been equivalent to x * 0.5).
>>

I am not so sure that's acceptable, especially when we have quite a few
pseudo-IEEE machines around.

For sure, if we are talking about transformations that change the results
on a given  port, they come under the auspices of this document as to what
are and are not acceptable transformations.