Re: [Boost-users] dereferencing type-punned pointer in function_base under gcc-4.3

From: Douglas Gregor (dgregor_at_[hidden])
Date: 2007-10-22 14:06:39

On Oct 21, 2007, at 1:34 PM, Benoit Hudson wrote:
> I'm trying to make sure my code compiles under gcc-4.3 before that
> compiler comes out.  I'm currently stuck on an error it reports in
> boost, version 1.34.1 (which I downloaded seconds ago; error was also
> present in 1.34.0):
>
> In file included from ../boost/function/detail/prologue.hpp:16,
>                  from ../boost/function.hpp:22,
>                  from ../boost/iterator/transform_iterator.hpp:10,
>                  from [my code]:
> ../boost/function/function_base.hpp: In member function 'const
> std::type_info& boost::function_base::target_type() const':
> ../boost/function/function_base.hpp:485: error: dereferencing
> type-punned pointer will break strict-aliasing rules
>
>
> I don't understand how there's any type-punning going on at line 485,
> although to my untrained eye it looks a lot like it's returning a ref
> to a temporary.  A couple lines later looks only very slightly
> fishier, but still shouldn't be breaking any aliasing rules.  So I
> leave it to the experts to figure out what's going on here.

I'm pretty certain the compiler is wrong here, although we would need  
to distill this into a much simpler test case before reporting it to  
GCC.

There is no type-punning going on, because the "manager" will always  
set the const_obj_ptr to a pointer to a const std::type_info object,  
and then we extract the pointer as a std::type_info const*.

There also is no temporary involved here, because typeid always  
returns an lvalue. The object that the lvalue refers to lives until  
the end of the program.

I think we might be able to help the compiler by adding

        type.const_obj_ptr = 0;

prior to the call to vtable->manager in the target_type function.

        - Doug