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

[dewar@gnat.com]

Part of the trouble here is that I get the feeling that quite a few people
are discussing these issues for the first time, and reacting ab initio.
As I said earlier, these issues are very old, and it is well worth going
back to the old literature that discusses these issues. In particuloar the
discussions of floating-point evaluation in the Fortran standards are of course
highly relevant here.