On Tuesday, 10 November 2015 at 14:28:15 UTC, Shachar Shemesh
wrote:
It seems that "foreach" requires that the range be copyable, in
direct contradiction to the fact that Range is not a forward
range.
To add to what I said, whether a range is "copyable" has nothing
to do with whether it is a forward range or not. What's supposed
to be guaranteed with any and all ranges is that if you copy it,
it has the exact same state as the original had - but _not_ that
that state is separate from the original. So, if you have
auto copy = orig;
then you can pop elements off of copy like you would have popped
them off of orig, but you must on longer pop elements off of
orig. _copy_ is now the range. This is critical when you consider
that most ranges will be copied when they're passed to a
function. If they weren't copyable, they'd be almost useless.
If you couple that with the extremely amorphous definition of
"save" for forward ranges, there seems to be quite a confusion
between the two.
There is nothing amorphous about the definition of save. save
duplicates a range such that you have two ranges of the same type
which are independent of one another and which contain exactly
the same elements. save provides a way to get a copy of the
range's state regardless of whether it's value type, reference
type, or pseudo-reference type.
Consider, for instance, a range which is implemented as a class.
It's a full-on reference type, and copying it around will never
result in its state being copied. Without save, there would be no
way to duplicate it. And save provides the _only_ generic means
of duplicating a range. It's a bug any time that generic code
depends on the state of a range being duplicated when it's
copied. Unfortunately, because most ranges _do_ duplicate their
state when they're copied, there's plenty of code out there which
is buggy and relies on ranges being duplicated when they're
copied. That's why it's so critical to test algorithms with a
variety of range types.
- Jonathan M Davis
