[PATCH] Document arithmetic overflow semantics

[Robert Dewar]

> That's not really true... a straight-forward way to "fix" this problem is
> to simply initialize all variables to, for example, zero.  In almost all
> cases, optimization would DCE this initialization anyway, so there is
> really very little cost.  The only major cases I can think of that the
> initialization would not be killed would be something like this:

Hmm, I am quite dubious about the claim that this would have zero overhead
in practice, but if true, then I agree, it should be done.