I ran some timing tests, partly to check out -fast-math -- my informal impression
is that it speeds up exp by a factor of about 2, but can slow down other things
(FIR filters). I didn't see any obvious improvement in -O3 over -O2. The
actual numbers are in the snd tarball, snd-edits.c -- look for "fast-math".
(The table is aimed more at comparing int/float/double operations). These
timing numbers change by a factor of, say, 20% on each run, so don't put
much weight on them.