&typeid(T) != &typeid( T() ) [was dynamic_cast in python module]

[Toon Knapen <toon dot knapen at fft dot be>]

Searching a bit on the problem reported in 'dynamic_cast in python 
module', I found out that in some dynamic library following two lines 
(which are directly one after the other and thus in the same dynamic 
library) print two different addresses:

1) std::cout<<&typeid(T)<<":"<<typeid(T).name()<<'\n';
2) T t ; std::cout<<&typeid(t)<<":"<<typeid(t).name()<<'\n';

(Where T is a template class derived from some non-template base-class)

So the address of the type_info objects in line 1 and 2 are different, 
the names however are identical. I have searched a lot in the archives 
but have never seen a problem like this being reported (all problems 
reported are accross dynamic library boundaries)

The above problem occurs in a python extension module. So I wanted to 
regenerate the problem in a small test-case but only succeeded 
partially. In following small test-case I got a surprising (to me) 
result too:

<test>
#include <iostream>

class A {
public:
virtual ~A() {}
} ;

template < class T > class B {} ;

void main()
{
  typedef B<int> Bint ;
  std::cout<<&typeid(Bint)<<":"<<typeid(Bint).name()<<'\n';
  std::cout<<&typeid(Bint())<<":"<<typeid(Bint()).name()<<'\n';
  Bint bint ;
  std::cout<<&typeid(bint)<<":"<<typeid(bint).name()<<'\n';
} ;
</test>

the first and third cout print the same address and name, the second 
however prints another address and a name that can't even be demangled 
by c++filt ?!

Any ideas on how to investigate further the first problem and the reason 
for the second 'surprising result' would be greatly appreciated.

I'm using gcc 3.2 and ld 2.13.90.0.2 (RH8.0) for this test