Re: [PATCH 2/2] Free multidicts from blockvector

[Tom Tromey <tom@tromey.com>]

Simon> I guess that this restriction of "objects allocated on obstack must be
Simon> trivially destructible" is a bit artificial and arbitrary?  It is
Simon> technically fine to allocate on obstack an object having a destructor,
Simon> as long as you ensure the destructor is called at some point.  I guess
Simon> the restriction was put because it's easy to forget and error prone.

I think the specific difference from ordinary 'new' is that valgrind
will detect a new-based memory leak, but not an obstack-based leak.  So,
in the heap case, the leak will also point out that the destructor isn't
being run; whereas in the obstack case it could go completely unnoticed.

Simon>  - There is a ton of blocks created, so it's nice to be able to free
Simon>    their memory in O(1).  Unfortunately, having to free
Simon>    block::m_multidict means that part is O(n), but the less we have to
Simon>    do in O(n) the better I guess?

We could also track the ownership of the dictionaries separately,
e.g. with unordered_set<mdict_unique_ptr> or similar.  This would reduce
the cost if it matters.

Tom