Boost.Serialization test failures on AIX with vacpp

[ccambly@…]

The tests for Boost.serialization: test_polymorphic_polymorphic_text_archive test_polymorphic_polymorphic_text_warchive test_polymorphic_polymorphic_xml_archive test_polymorphic_polymorphic_xml_warchive

on AIX with vacpp all fail with the error message: (S) The call to "save" has no best match These tests work fine with address-model=64. The compilation errors are only emitted in 32bit because the macro BOOST_NO_INT64_T is not defined in 32bit.

I believe the right fix is changing boost/cstdint.hpp to define the macro BOOST_NO_INT64_T in 32bit mode on AIX with vacpp.

[Robert Ramey]

I THINK this is the correct person to deal with this. If I'm wrong about this, feel free to bounce it back to me.

Robert Ramey

[anonymous]

Whatever the cause is, I'm not convinced that this is the correct fix.

I note that with the regression tests (32-bit) run with vacpp 10.1 that the integer tests are all compiling.

Are you really saying that there is no int64_t in stdint.h? Come to that I don't see how this macro can be defined in 64-bit mode either?

I believe the problem here is in polymorphic_iarchive.hpp, rather than:

    #if !defined(BOOST_NO_INTRINSIC_INT64_T)
    virtual void load(boost::int64_t & t) = 0;
    virtual void load(boost::uint64_t & t) = 0;
    #endif

It should be:

    #if defined(BOOST_HAS_LONG_LONG)
    virtual void load(boost::long_long_type & t) = 0;
    virtual void load(boost::ulong_long_type & t) = 0;
    #elif defined(BOOST_HAS_MS_INT64)
    virtual void load(__int64 & t) = 0;
    virtual void load(unsigned __int64 & t) = 0;
    #endif

In other word if you overload on type, then overload by "long long" or equivalent as well. Alternatively overload by size: int16_t, int32_t etc (but this may miss some types out if they're the same size as each other, causing problems down the road). But whatever you do don't mix the two!

BTW long long will be part of the next standard, so we should be supporting this as a native type where available anyway.

HTH, John.

[ccambly@…]

I will retract my patch.

The problem for vacpp is that those function overloads in polymorphic_iarchive.hpp are controlled by the macro BOOST_NO_INTRINSIC_INT64_T, but inside of the tests them selves the macro used is BOOST_NO_INT64_T. It is possible to have one macro defined and not the other.

I originally was changing the tests A.ipp A.cpp A.hpp to use

#if !(defined(BOOST_NO_INT64_T)

defined(BOOST_NO_INTRINSIC_INT64_T))

instead of #ifndef BOOST_NO_INT64_T which worked as well.

I changed my mind about the fix upon seeing code for _hpux in cstdint.h doing something to control the macro BOOST_NO_INT64_T but I'm really not sure what is the real issue.

Johns change sounds reasonable.

[ccambly@…]

The section of code at the top of polymorphic_oarchive.hpp

determine if its necessary to handle (u)int64_t specifically i.e. that its not a synonym for (unsigned) long if there is no 64 bit int or if its the same as a long we shouldn't define separate functions for int64 data types. #if defined(BOOST_NO_INT64_T)

#define BOOST_NO_INTRINSIC_INT64_T

#else

#if defined(ULLONG_MAX)

#if(ULONG_MAX == 18446744073709551615ul) 264 - 1

#define BOOST_NO_INTRINSIC_INT64_T vacpp gets here

#endif

...

So we have BOOST_NO_INT64_T not defined and BOOST_NO_INTRINSIC_INT64_T defined. In A.hpp we will get the members for f, and g defined.

#ifndef BOOST_NO_INT64_T boost::int64_t f; boost::uint64_t g; #endif

but in polymorphic_oarchive.hpp we do not get the declarations for

#if !defined(BOOST_NO_INTRINSIC_INT64_T) virtual void save(const boost::int64_t t) = 0; virtual void save(const boost::uint64_t t) = 0; #endif

[Robert Ramey]

1) I've implemented John's suggestion. should show up in the trunk in a while.

2) seems like the analogous fix should be made in polymorphic_oarchive.hpp . correct?

Robert Ramey

[Robert Ramey]

looks the following headers should also get similar treatment

collection_traits.hpp polymorphic_oarchive_route.hpp