Precondition for `advance(i, n)`

[morinmorin]

In [iterator.operations] advance does not have a requirement on iterator validity, but I think something like this is needed:

• For non-negative n, [it, itn) should be a valid range, where itn is the iterator obtained by incrementing it by n.
• For negative n, [itn, it) should be a valid range, where itn is the iterator obtained by decrementing it by |n|.

While the Standard too doesn't have this requirement, the SGI STL documentation has a similar requirement.

[CaseyCarter]

For [range.iter.op.advance]/1 (the "i, n" overload) the precondition we have now, " If I does not model BidirectionalIterator, n is not negative." is sufficient to support purely operational "evaluates these expressions and means whatever that means" semantics. But it's not terribly useful. It could be improved to:

If n > 0, then [i, n) denotes a range. If n < 0, then I models BidirectionalIterator and other stuff.

Where "other stuff" is one of:

• [j, i) denotes a range where j is the result of decrementing i by -n
• i is decrementable, as are its -n - 1 predecessors [what's a "predecessor"?]
• there exists an iterator j such that [j,-n) and [j,i) denote the same range [what does "the same range" mean?]
• there exists an iterator j such that [j,-n) denotes a range and (advance(j, -n), j) == i
• ...?

[morinmorin]

The STI doc (mirror site) does not use interval notation there and provides a concise description (which is not divided into positive and negative cases):

Preconditions

• Every iterator between i and i+n (inclusive) is nonsingular.
• …

But I think more descriptive wording (like Casey's) would be better suited for the Standard.

BTW, is there any guideline on which one to use "a valid range" or "a range"?