Fourth Draft "Unsafe fp optimizations" project description.

[Stephen L Moshier]

> | The loss of accuracy is due to subtractive cancellation, an error source
> | that your document does not mention. 
> 
> Actually on x86, sin(x) just returns x for "large" values of x.

x at about 2^62 is the coprocessor's total loss of precision
threshold, where the error approaches 100% and the 8087 sets a
hardware flag.  The error increases gradually to that point, however.
The threshold, where the range reduction error due to cancellation
begins, is about x = 2^18 for double precision and x = 2^47 for single
precision.