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

[dewar@gnat.com]

<<Stop telling us what we're allowed to want.


We are just trying to understand what you want, it was you who said
that you do not want optimizations that change program results, not me!
Now you change it to "change program results enough to matter". And
that of course is the point, we have to agree on what is enough to matter.

The problem is agreeing on a criterion. As I have said many times, for
me the *primary* criterion is that the optimization actually make a real
noticable difference in performance. There is no point in allowing 
non-identity transformations that do not provide noticeable speedup.

Then if the speedup *is* significant, then it makes sense to discuss
just how safe the transformation is, and whether it meets the collected
consensus understanding of what "changing program results but not enough
to matter" might mean.