[LLVMdev] [RFC] Relax the rules on const auto&? (was Re: r203179 - [C++11] Replacing iterators redecls_begin() and redecls_end() ...)

[Duncan P. N. Exon Smith]

On Mar 7, 2014, at 9:15 AM, David Blaikie <dblaikie at gmail.com> wrote:

> On Fri, Mar 7, 2014 at 8:58 AM, Duncan P. N. Exon Smith
> <dexonsmith at apple.com> wrote:
>> Unless the pointers themselves are changing, the only benefit here is
>> consistency.  But ranges of iterators should use const&, since many
>> iterators are expensive to copy.
> 
> To clarify, this isn't about copying iterators

You’re right, that was a tangent.

> (iterators are
> generally cheap to copy) but copy the elements being iterated over.

The point (of my tangent) is that some iterators (scc_iterator,
po_iterator, etc.) have non-trivial state, so copies can be expensive.

>> Accidental copies can be a huge headache, since they lead to subtle
>> bugs that are hard to track down (performance implications aside).
>> 
>> I still favour applying the rule generally, since consistency matters.
>> But we could text like:
>> 
>>    Pointer ranges are a common exception.
>> 
>> Any other opinions?
> 
> If you write the exception as "auto *x" then it's unambiguous and easy
> to still look at "auto x" as being questionable/to-be-avoided.

This is obviously better.  Changed in r203254.