Re: inlining + optimization = nuisance bugs

> [I've never understood why people are so eager to have fast wrong
> answers. -John]


Sometimes the IEEE is a little more pedantic than one wants. I don't care
if 0.0/0.0 trashes my hard disk, I won't let it happen. Similarly, if
ripple in the last digit makes a difference to my code, it's not robust
enough.


Of course correct should be the default, but I'm just making the point
that "fast wrong" should always be an option.


It *is* difficult to say how wrong, though. In some sense there's a
hierarchy like:
  undefined exception behaviour on 0/0 or sqrt(-1)
  simplifying "1.0 * x" to "x"
  transforming "(a + b) + c" into "a + (b + c)"
  transforming "vector /= 3.0" with "vector *= (1/3.0)"
  nonrepeatable behaviour
  silently treating doubles as singles.


Different people stop at different levels, and reorder the list.


--
Andrew Fitzgibbon, awf@robots.ox.ac.uk
Oxford Information Engineering Research Group +44 01865 273127
<a href=http://www.robots.ox.ac.uk/~awf> Home Page </a>
--

Return to the comp.compilers page.
Search the comp.compilers archives again.