[PATCH] Document arithmetic overflow semantics

[Robert Dewar]

> So what?  There's no requirement that erroneous programs produce the
> same result in different versions of GCC.


There is also no requirement that the compile speed not be multiplied by
100 in the next version of GCC. The standard is not the only source of
criteria for a usable compiler, and a compiler that takes too much advantage
of undefined behavior may be a menace even if it is conforming.