On Mon, Jul 27, 2009 at 7:29 PM, Matthew Scouten
(TT)<matthew.scou...@tradingtechnologies.com> wrote:
> The documentation points out (with very good reason) that just using
> "::with_policies(bp::return_internal_reference)" can cause major
> problems if the container in question reallocates things between the
> time that you extract the object and the time you try to access it.
>
> v = VectorofFoo()
> v.extend([Foo() for x in range(10)])
> f = v[5] #returns a reference to a foo on the heap
> f.extend([Foo() for x in range(1000)]) #almost certainly causes
> reallocation of underlying array in c++
>
> f.function() #accesses a no longer valid reference, sky falls.
>
> The standard indexing suite uses clever proxy magic to work around this.
>
Okey, that is why v2 provides container_proxy class. There is an
advantage of not mixing "proxy magic" with container suite: simple
implementation.
>From my experience, this limitation was never a problem. After all,
users understands that they deal with wrapper and should be careful.
>
> On second Look, it looks like the intended usage of container_proxy is
> something like this:
>
> class_< container_proxy <std::vector<foo> > >("class_< std::vector<foo>
>>("VectorOfFoo") )
> .def( indexing::container_suite< container_proxy <std::vector<foo> >
>>() )
> ;
>
> Which seems kind of clunky to me.
I don't think the author had a choice here. Can you propose an
alternative interface?
One possible solution is to add helper function: ( pseudo code)
template< class TContainer >
void register_container_proxy( const char* name, boost::type< TContainer > ){
class_< container_proxy < TContainer > >( name )
.def( indexing::container_suite< container_proxy < TContainer > >>() ) ;
}
> I think defaulting to by-value is the wrong behavior. The idiom
> "v[n].some_attribute = blah" works seamlessly on python lists, and c++
> vectors. The indexing suite V1 is the default behavior, and can be
> disabled if needed.
May be it is a little bit confusing, but
>> The documentation points out (with very good reason) that just using
>> "::with_policies(bp::return_internal_reference)" can cause major
>> problems if the container in question reallocates things between the
>> time that you extract the object and the time you try to access it.
makes it reasonable and consistent.
>
> I am using container_suite on a "std::map<std::string,
> std::vector<std::string> >" and the only problem I have is a compiler
> warning about a very long decorated name. using foo_vector_exposer_t
> directly seems kind of clunky to me.
You can write the code as you wish. I like this way, mainly because it
is easier to understand a compiler errors. At least it points you
exactly to the problematic line. The syntax
class_ < ... >
.def (...)
.def (...);
is cool but it has its price.
> While I was trying to use indexing suite v2 to wrap a non-STL but
> STL-like container class, I noticed that the documentation for doing so
> is badly out of date. This is the main (almost the only) advantage that
> V2 has over V1, so these docs badly need to be updated.
You are welcome to contribute. My English and explanation skills are
not as good as I want.
--
Roman Yakovenko
C++ Python language binding
http://www.language-binding.net/
_______________________________________________
Cplusplus-sig mailing list
Cplusplus-sig@python.org
http://mail.python.org/mailman/listinfo/cplusplus-sig
