As you can see in the patch where I wrote a program to detect FP contraction, while it is possible to detect this optimization programmatically and act accordingly, because the compiled program must run on the test target, it is nontrivial to make all the relevant tests depend on the outcome of running such a program.
Sebastian and I therefor propose that we shall have a variable that is checked, such that the variable disables the FP-dependent test by default, but it can be re-enabled by setting the appropriate variable to an appropriate value [e.g. "no"].
Didn't you mean FAST_MATH?
Also, the explanatory text above still mentions both.