What is acceptable for -ffast-math? A numerical viewpoint

[dewar@gnat.com]

<<3. unless 2 showed up unfixable practical problems, the optimization
is enabled under -funsafe-fast-math

Do you actively oppose to this proposal?
>>

Yes, it should also be the case that the field testing should show that
there is some gain. There is no point in having non-"correctness"
preserving transformations unless there is at least some evidence of
an advantage. Probably the best thing is to first put in the optimization
under its own special switch. Then if there is a real gain, and no
unacceptable surprises, then it can be made part of -ffast-math.