Boost C++ Libraries: Ticket #7100: unorderd_set > > - insert does not increment count

owner, component changed

John Maddock — Sat, 07 Jul 2012 11:21:37 GMT

owner changed from John Maddock to Daniel James
component TR1 → unordered

Daniel James — Sun, 08 Jul 2012 11:55:59 GMT

(In [79358]) Unordered: Fix using a C++03 allocator with C++11 compiler.

Because the nodes had an implicit constructor, the has_construct traits was detecting that the nodes could be constructed by construction then copy, which really wasn't wanted. Also add a check that nodes aren't been copy constructed to make sure this doesn't happen again. Refs #7100.

attachment set

Mon, 09 Jul 2012 02:42:59 GMT

attachment → clang_error_nomatchingfunction.txt

clang compile error log "No Matching Function.."

Mon, 09 Jul 2012 02:46:50 GMT

Thanks for the fix, I can confirm that insert increments the ref count with boost::unordered_set<boost::shared_ptr<T< >.

However, I think it has created a problem when using std::tr1::shared_ptr. Albeit an odd setup, the following example produces a compile time error:

#include "boost/unordered_set.hpp"
#include <tr1/memory>
int main()
{
	boost::unordered_set<std::tr1::shared_ptr<int> > mySet;
	std::tr1::shared_ptr<int> intSP( new int( 1 ) );
	mySet.insert( intSP ); // <-- Error: (Semantic Issue) "no matching function for call to 'hash_value'"
    return 0;
}

I have attached the full error log from clang.

status, version, milestone changed

Daniel James — Mon, 09 Jul 2012 11:16:15 GMT

status new → assigned
version Boost 1.51.0 → Boost 1.50.0
milestone To Be Determined → Boost 1.51.0

I'm afraid that's a feature, not a bug. There's no support for std::shared_ptr or std::tr1::shared_ptr, older versions were accidentally picking up the implicit conversion to bool and hashing based on that - which is really bad as it only returns two possible hash values. So 1.51 is going to prevent that happening.

I might add support for std::shared_ptr in a future version, but if you really to use it, the best thing to do is to use a custom hash function, something like (untested):

struct hash_shared_ptr {
    std::size_t operator()(std::tr1::shared_ptr<int> const& x) {
        boost::hash<int*> hf;
        return hf(x.get());
    }
};

If you really need boost::hash to support shared_ptr, then you probably should specialize it.

Btw. you probably know this, but when you use a shared_ptr as a set element, it stores it based on the address, not the value it points to. I just thought I should mention it as people sometimes get confused.

Tue, 10 Jul 2012 01:07:08 GMT

Ah, right you are! I had forgotten about that and had a bug in my own code. That must mean that std::unordered_set is also using the implicit bool to hash? It is really nice that you can catch it, as it is almost certainly not what was intended.

Btw, yes the goal is to map unique objects based on their address. Thank you nonetheless for your kind and thorough advice!

All is working well now, cheers.

Daniel James — Tue, 10 Jul 2012 07:08:25 GMT

No, std::hash is specialized for std::shared_ptr, so I should definitely do the same. It doesn't have this problem, it's an unintended consequence of boost::hash's extensions. I think I'll add support for std::shared_ptr in 1.51, I am lagging behind a bit with C++11 support.

status changed; resolution set

Daniel James — Sun, 15 Jul 2012 23:58:05 GMT

status assigned → closed
resolution → fixed

(In [79547]) Unordered: Merge allocator fix + improved tests. Fixes #7100.