Created attachment 24260 [details]
test case, preprocessed on x86_64-linux

Created attachment 24261 [details]
test case, non-preprocessed

The temporary returned by getv() seems to be destroyed too early, before the printf that uses it

slightly reduced, but boost/foreach.hpp still includes thousands of lines

in this version vector<T> zeros its data member on destruction, so using it after destruction is obvious

#include <boost/foreach.hpp>

template<typename T>
struct vector
{
    typedef const T* const_iterator;

    ~vector() { data = T(); }

    T data;

    const_iterator begin() const { return &data; }
    const_iterator end() const { return &data + 1; }
};

vector<int> v = { 1 };

const vector<int *>
getv()
{
    vector<int *> ret = { &v.data };
    return ret;
}

int
main(int argc, char **argv)
{
    BOOST_FOREACH(int *w, getv())
        __builtin_printf("%d\n", *w);

    return 0;
}

slightly more reduced

#include <boost/foreach.hpp>

struct vector
{
    typedef int* const* const_iterator;

    ~vector() { data = 0; }

    int* data;

    const_iterator begin() const { return &data; }
    const_iterator end() const { return &data + 1; }
};

int v = 1;

const vector
getv()
{
    vector ret = { &v };
    return ret;
}

int
main(int argc, char **argv)
{
    BOOST_FOREACH(int *w, getv())
        __builtin_printf("%d\n", *w);

    return 0;
}

The regression started with r170601:

Author: jason
Date: Tue Mar  1 22:44:35 2011
New Revision: 170601

URL: http://gcc.gnu.org/viewcvs?root=gcc&view=rev&rev=170601
Log:
	PR c++/47851
	* call.c (standard_conversion): Provide requested cv-quals on
	class rvalue conversion.

well if the fix for PR 47851 made G++ correct then this could be a bug in BOOST_FOREACH, but it's hard to know because trying to reduce any MPL code is pretty tedious!

Jason, any ideas?  do we need to reduce it with delta?

I'm trying to reduce the preprocessed test case, but since I'm not a C++ person this might take some time.

(In reply to comment #8)
> I'm trying to reduce the preprocessed test case, but since I'm not a C++ person
> this might take some time.

Scratch that, the preprocessed code is way too weird for me.

(In reply to comment #9)
> (In reply to comment #8)
> > I'm trying to reduce the preprocessed test case, but since I'm not a C++ person
> > this might take some time.
> 
> Scratch that, the preprocessed code is way too weird for me.

welcome to Boost MPL!  ;)

there are huge chunks that can be removed (iostreams and locales, reverse_iterators) but reducing it manually is not easy.

See http://gcc.gnu.org/wiki/A_guide_to_testcase_reduction for a tool to do it automatically

Created attachment 24299 [details]
reduced by delta

This is a bug in BOOST_FOREACH_COMPILE_TIME_CONST_RVALUE_DETECTION.

The GCC bugfix meant that where previously

true ? boost::foreach_detail_::make_probe(getv()) : (getv())

had the type "vector", now it has type "const vector".

5.16/3 says:

Otherwise (i.e., if E1 or E2 has a nonclass type, or if they both have class types but the underlying classes are not either the same or one a base class of the other): E1 can be converted to match E2 if E1 can be implicitly converted to the type that expression E2 would have if E2 were converted to a prvalue (or the type it has, if E2 is a prvalue).

"const vector" cannot be converted to make_probe<vector>, but make_probe<vector> has an operator vector&, so it can be converted to const vector, so the type of the conditional expression is const vector.

Here's a reduced example:

template <class T>
void f (T&, int);

template <class T>
void f (T const&, ...);

struct A { };
struct B { operator A& () const; };

const A g();

int main()
{ 
  f( (true ? B() : g()), 0);
}

(In reply to comment #12)
> This is a bug in BOOST_FOREACH_COMPILE_TIME_CONST_RVALUE_DETECTION.

This was already fixed in boost/trunk, 
and the fix will be released in Boost.1.47. 

For details, please see the following ticket:
[Foreach] Compile-time const rvalue detection fails with gcc 4.6
https://svn.boost.org/trac/boost/ticket/5279