Skybuck Flying <skybuckflying@gmail.com> wrote:
> I discovered another dangerous one while trying to compute overlap of interval (basically ranges) and/or trying to clip/cap line segments:
>
> 0 * +infinity = -NAN
>
> This happens as the ray is on the edge of a boundary/box... cause tMinX will become -NAN.
>
> Leading to weird situations depending on how the code was written, either the wrong point will be taken or it will not clip at all.
>
....

If you are using someone else's compiled code then you have to pray
they were merciful.
There was an hacker's rule of thumb in numerical s/w -- about 2/3 of
the code deals with overflow and underflow.
One proposed solution was to abandon regular FP numbers and use a kind
of number that had a bigger range using interated logarithms.
(I.e. the exponent doesnt represent a binary exponent but the number of
times the original number was put through log2(x)).
But -- hard to believe :) -- the idea didnt catch on and the IEEE won!

Sometimes it can also be a matter of changing your "x86" hardware from
AMD to intel or vice-versa. What with pipelines and gpus not being
very well standardised some wintox exe "works" on one kind of HW and
randomly gives garbage answers on another kind.