Re: [boost] [config] additions for gcc 4.0.1 and IBM V11.1.0.0 duetoMSM compile errors

Subject: Re: [boost] [config] additions for gcc 4.0.1 and IBM V11.1.0.0 duetoMSM compile errors
From: Christophe Henry (christophe.j.henry_at_[hidden])
Date: 2010-09-21 16:16:26

> The compiler file for this compiler is boost/config/compiler/vacpp.hpp.  If
> in doubt about stuff like this, then go to the config tests and look at the
> config_info results for the compiler in question - it gives you this
> information and much much more:
> http://www.boost.org/development/tests/trunk/developer/output/IBM_Canada_Ltd-boost-bin-v2-libs-config-test-config_info-test-vacpp-debug.html

Ah yes now I see it. Sorry.

> The problem I have is that these compilers are passing the tests for these
> macros.

It wouldn't be the first time I see the gcc compiler supposedly
supporting a feature but breaking if the construct using it is
complicated enough. I spent days unsuccessfully trying to make the
supposedly MPL-compatible gcc 3.4 compile MSM :(

>I've just committed a stricter version of the test for
> BOOST_NO_NESTED_FRIENDSHIP in case this is an Apple GCC specific problem, as
> both the updated tests and the msm tests pass for me here with Linux
> gcc-4.0.4.

Thanks a lot! At least the problem seems to be quite limited.

>I haven't tried gcc-4.0.1 on Linux as it means yet another gcc
> build and install.... I'll leave that to you to try :-)

Is it a variant of the famous "exercise left to the reader"? ;-)

> Do you have a stricter test case than we currently use for
> BOOST_NO_MEMBER_TEMPLATE_FRIENDS that more closely represents MSM's usage?

Looking at your test case, the only possible cause I can imagine is
that I'm taking an address of a foo static method in foobar so we
could add this:
template <class T>
class foo
{
private:
   template<typename Y> friend class foobar;
   template<typename Y> friend class foo;
   template<typename Y> friend bool must_be_friend_proc(const foo<Y>& f);
   int i;
   static int my_func(){return 0;} // added
public:
   foo(){ i = 0; }
   template <class U>
   foo(const foo<U>& f){ i = f.i; }
};

template <class T>
class foobar
{
   int i;
public:
   template <class U>
   foobar(const foo<U>& f)
   { i = f.i; }
   // added
   typedef int (*foofct) ();
   void test_address() { foofct fct=&foo<int>::my_func;}
};

> Ah... just looked at the error messages and I'm not sure these are related
> to that macro or not?

It's hard to spot but I think this is pretty clear:
"../boost/msm/back/dispatch_table.hpp", line 174.40: 1540-0300
(S) The "private" member
"boost::msm::back::state_machine<<unnamed>::player_,boost::msm::back::NoHistory,boost::msm::back::favor_runtime_speed>::call_no_transition<<unnamed>::play>(library_sm
&, int, int, const play &)" cannot be accessed.

dispatch_table is made friend so this should not happen.

>You could try emailing the IBM test-runners
> directly - they're quite helpful at tracking down issues like this - just
> don't tell anyone ;-)

Of course not! It'll stay strictly between us :)
I'll take the subject to Michael then.

Thanks,

Christophe