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

[Gabriel Dos Reis]

Wolfgang Bangerth <wolfgang.bangerth@iwr.uni-heidelberg.de> writes:

| I belong to the latter group and can only speak for them.

I do extensive numerics (for living) in C++, and clearly you're not
speaking for me.

| - denormals: if you run into the risk of getting NaNs one way but not
|   the other, you most certainly already have gone beyond the limits of
|   your program

Untrue.  Using Graeffe's iteration (a stable algorithm to find the
greatest or smallest root radius) to isolate polynomial roots, one
can get NaNs for specific reasons without necessarily using an
unstable formaltation of the problem -- that was the whole point of a
subproject of the FIRSCO project.

-- Gaby