More accurate Set and Map size warnings

[meooow25]

Set and Map documentation carry warnings that say

The size of the set must not exceed maxBound::Int. Violation of this condition is not detected and if the size limit is exceeded, its behaviour is undefined.

However, we multiple sizes by delta = 3 without caring about overflow when balancing, which means we get violated invariants earlier than that.

I doubt anyone will be running into this in practice, but the limit should be document as maxBound `div` 3 perhaps. Or less, in case there are other potentially troublesome operations I missed.

[treeowl]

Good point. Is it possible to avoid the problem by being careful about how we do the comparisons? It's not an issue for 64 bit, but we still (nominally) support 32 bit, and (theoretically) 29 bit.

[meooow25]

Is it possible to avoid the problem by being careful about how we do the comparisons?

Well we could replace a * 3 > b with a > b `quot` 3. This would likely be a slowdown, benchmarks can tell.

[treeowl]

Hmph. That does sound like a slowdown. Checking for overflow would probably be a slowdown too (more work for branch prediction), but what do I know? Changing the documentation is at least easy.