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

[dewar@gnat.com]

And yes, I still oppose your proposal as it is written, it fails to make
clear that positive performance data should be required to support any
transformation.