Re: Compiling smart_ptr library

From: Christopher Currie (Christopher_at_[hidden])
Date: 2003-04-30 11:14:36

Dave Abrahams wrote:
> l_at_lo wrote:
>>
>>sunpro-Link-action
>>../../../libs/smart_ptr/test/bin/shared_ptr_alloc_
>>test.test/sunpro/debug/runtime-link-dynamic/shared_ptr_alloc_test
>>Undefined                       first referenced
>> symbol                             in file
>>boost::__RTTI__1nFboostPchecked_deleter4nBZ___
>>/rest/proj/users/kercsoj/boost/boost_1_30_0/libs/s
>>mart_ptr/test/bin/shared_ptr_alloc_test.test/sunpr
>>o/debug/runtime-link-dynamic/
>>SunWS_cache/CC_obj__/_iANwWXwaw2nbnjBybyL.o
>>[Hint: static member boost::__RTTI__1nFboostPchecked_deleter4nBZ___
>>must be defined in the program]
> 
> Are you compiling with RTTI off?

I've seen this error as well. The only workaround I know of is to 
compile using static template instantiation, rather than Sun CC's 
template repository.

bjam -sTOOLS="sunpro" -sBUILD="<cxxflags>-instances=static"

It may also be necessary to delete the libs/smart_ptr/test/bin 
directory, as Sun's compiler still occasionally attempts to link in 
oudated code from the template repository.

This solution has the drawback of increasing code size due to duplicate 
template definitions, but is the only solution I've found to this 
problem. I have managed to strip down the offending code to a test case 
of about 200 lines, but as my company is dragging its feet on a support 
contract, I haven't been able to submit it as a bug report.

That said, the test suite for smart_ptr will still encounter other problems:

[ The rest of this post addresses other Sun compiler issues, feel free 
to skip it if you're not interested. I didn't cross post it, because I 
wasn't sure if it should go to boost.devel or boost.build. ]

* shared_ptr_alloc_test.cpp still triggers "Assertion: 
(../lnk/init.cc, line 991)," despite Sun's patch report claming this 
problem fixed in their most recent patch.

* shared_ptr_basic_test has 7 use count errors.

* shared_ptr_test has 3 use count errors.

* weak_ptr_test has 2 use count errors.

I've only been able to do a little investigation of the use count 
errors, but from what I've seen, they occur when the Sun compiler has to 
create a shared_ptr temporary, and the destructor for the temporary is 
delayed. I've run shared_ptr_test in purify, and no memory is being 
leaked, so I infer that the destructor is in fact being called, just not 
  before the use_count test happens:

"shared_ptr_basic_test.cpp"
// ...
// prior to this line, p5 is a shared_ptr<void> with
//   use_count == 1, weak_count == 1.

line 226:    weak_ptr<X> wp2 = static_pointer_cast<X>(p5);

// static_pointer_cast creates a shared_ptr<X> temporary:
//   use_count == 2, weak_count == 2.
// temporary is passed to weak_ptr<X>'s constructor
//   use_count == 2, weak_count == 3.
// temporary is NOT immediately destructed.

line 228:    BOOST_TEST(wp3.use_count() == 1); // this test failed

// Sometime after this point (possibly as late as the end of the
// function scope) the temporary is destructed.

This test (and the others like it) seem to assume that the temporary 
will be immediately destructed after the assignment, although AFAIK that 
is not required by the language. Sun's compiler seems to favor delayed 
destruction, although I haven't had the time to trace the debugger and 
find exactly where it does happen.

My tests were run using:

$ CC -V
CC: Forte Developer 7 C++ 5.4 Patch 111715-06 2003/3/29