On Monday, 22 June 2015 at 06:29:20 UTC, Andrei Alexandrescu
wrote:
The matter is very well understood. My point here is that
leaving it to the user to make sure which ranges are still
valid vs. not is not appropriate for D's container. -- Andrei
If we mean that we don't want undefined behavior, and we want to
do something like throw Errors when a programmer screws up and
uses an invalidated range, then fine - especially if it's
restricted to version(assert) like we've recently been doing with
our checks for calling popFront or front on empty ranges in
Phobos - then logic errors with containers can be caught more
easily. But if you mean that the programmer shouldn't have to
know what types of operations are going to invalidate a range for
a specific container type, I don't see how that makes any sense
at all. How can anyone expect to use a container without
understanding it well enough to know what's going to happen when
they start adding elements to it or removing elements from it
while iterating over it? IIRC, Java tries to be nicer than C++
and throws an exception when you screw that up rather than
treating it as undefined behavior, but even they don't try and
make it work for you. You need to understand the container types
that you're using if you expect your code to be correct or
efficient.
And remember that we're already dealing with undefined behavior
with all ranges with stuff like what happens when you call front
or popFront (or pretty much anything other than empty) on an
empty range. So, it's not like we're not going to have undefined
behavior involved here. You're just looking to avoid it in an
additional case where folks have used a range incorrectly. I'd
suggest that we do with containers what we've taken to doing with
ranges in Phobos in the last year or so and add checks in
version(assert) blocks which throw Errors when you screw up
(which in the case of ranges in general would be throwing in
functions like front or popFront when the range is empty, and in
the case of ranges for containers would mean throwing an Error
when you try and do something with them after they've been
invalidated).
But I don't think that it's at all unreasonable to consider code
that tries to iterate over an invalidated range to be a logic
error that the programmer needs to fix. Programmers should know
when what they're doing to a container is going to invalidate its
existing ranges.
- Jonathan M Davis