Re: [Boost-users] [test] Malloc exceptions when using test framework on 64bit OS X 10.6.1

Subject: Re: [Boost-users] [test] Malloc exceptions when using test framework on 64bit OS X 10.6.1
From: Rush Manbert (rush_at_[hidden])
Date: 2009-09-21 17:16:41

On Sep 21, 2009, at 1:20 PM, Peter Soetens wrote:

> On Mon, Sep 21, 2009 at 22:06, Peter Klotz <peter.klotz_at_[hidden]> wrote:
>> Rush Manbert wrote:
>>>
>>> On Sep 19, 2009, at 12:42 AM, Peter Klotz wrote:
>>>
>>>> Rush Manbert wrote:
>>>>>
>>>>> On Sep 18, 2009, at 4:57 PM, Rush Manbert wrote:
>>>>>>
>>>>>> On Sep 17, 2009, at 1:02 AM, Vrai Stacey wrote:
>>>>>>
>>>>>>> Hi,
>>>>>>>
>>>>>>> I've been making my first forays in to using Boost on the new  
>>>>>>> OS X
>>>>>>> release and have encountered some strange behaviour on 64bit  
>>>>>>> machines.
>>>>>>> Testing with Boost 1.37.0 and 1.40.0 I've found that everything
>>>>>>> compiles as expected, but attempting to run any unit tests  
>>>>>>> results in
>>>>>>> an error of the form:
>>>>>>>
>>>>>>> *** No errors detected
>>>>>>> test_exceptions(92517) malloc: *** error for object  
>>>>>>> 0x3000100200460:
>>>>>>> pointer being freed was not allocated
>>>>>>> *** set a breakpoint in malloc_error_break to debug
>>>>>>>
>>>>>> <snip>
>>>>>>
>>>>>> I am seeing the same thing using Boost 1.38.0 built for 64-bit  
>>>>>> Intel on
>>>>>> OS X 10.6.1. The test program worked fine on OS X 10.5.7, but  
>>>>>> I'm seeing the
>>>>>> same malloc errors as you.
>>>>>>
>>>>>> For me, the error occurs in the file framework.ipp:
>>>>>>
>>>>>> class framework_impl : public test_tree_visitor {
>>>>>> public:
>>>>>>  framework_impl()
>>>>>>  : m_master_test_suite( 0 )
>>>>>>  , m_curr_test_case( INV_TEST_UNIT_ID )
>>>>>>  , m_next_test_case_id( MIN_TEST_CASE_ID )
>>>>>>  , m_next_test_suite_id( MIN_TEST_SUITE_ID )
>>>>>>  , m_is_initialized( false )
>>>>>>  , m_test_in_progress( false )
>>>>>>  {}
>>>>>>
>>>>>>  ~framework_impl() { clear(); }
>>>>>>
>>>>>>
>>>>>>  void            clear()
>>>>>>  {
>>>>>>      while( !m_test_units.empty() ) {
>>>>>>          test_unit_store::value_type const& tu =  
>>>>>> *m_test_units.begin();
>>>>>>
>>>>>>          // the delete will erase this element from map
>>>>>>          if( test_id_2_unit_type( tu.second->p_id ) ==  
>>>>>> tut_suite )
>>>>>>              delete  (test_suite const*)tu.second;
>>>>>>          else
>>>>>>              delete  (test_case const*)tu.second;
>>>>>> <<<<------------------------- This statement generates the error
>>>>>>      }
>>>>>>  }
>>>>>>
>>>>>> GDB won't let me put a breakpoint anywhere in this file. Maybe  
>>>>>> I can
>>>>>> hack it to see something...
>>>>>
>>>>> Okay, I was able to debug it by doing this:
>>>>> The main unit test file is called UnitTestMain.cpp, and just  
>>>>> includes
>>>>> <boost/test/included/unit_test.hpp>, and that file includes all  
>>>>> of the *.ipp
>>>>> files.
>>>>> I preprocessed UnitTestMain.cpp in Xcode and deleted all of the
>>>>> preprocessor lines with sed:
>>>>>  sed '/^#/ d'  <UnitTestMain.ii >UnitTestMain.cpp
>>>>> and I substituted the new UnitTestMain.cpp (the preprocessed  
>>>>> file) for
>>>>> the original one.
>>>>> After that I could set a breakpoint in
>>>>> framework_impl::register_test_unit(). What I see is that the  
>>>>> test case being
>>>>> registered is the "this" pointer from an object that was created  
>>>>> with new,
>>>>> so it looks like the error is a bug in Snow Leopard.
>>>>> That's about all I know right now. I have never used boost unit  
>>>>> test, so
>>>>> don't really know how to put together a reduced test case that  
>>>>> can be
>>>>> submitted to Apple. Any takers? :-)
>>>>> - Rush
>>>>
>>>>
>>>> This seems to be once again this bug in Boost.Test:
>>>>
>>>>
>>>> http://www.nabble.com/-Boost.Test--valgrind-complains-about-invalid-reads-td20500154.html
>>>>
>>>> http://www.nabble.com/-Boost.Test--Error:-Non-aligned-pointer-being-freed-td24335733.html
>>>>
>>>> Try making the test_unit destructor virtual and your problem  
>>>> should be
>>>> resolved.
>>>>
>>>
>>> Thank you, Peter. That does resolve the problem.
>>>
>>> Do you have any idea whether 1.40 still has this problem?
>>>
>>
>> Yes, 1.40.0 also suffers from this bug.
>>
>> Maybe we can finally convince Gennadiy to fix this issue in SVN.
>>
>> This would stop forcing people to use their own patches each time a  
>> new
>> version of Boost is released.
>
> +1
>
> I reported this before too. *everyone* (on 64bit?) is seeing this, and
> the fix is trivial.
>

Now I have gone back and read the original thread about valgrind  
complaints. I see that Gennadiy is rather adamant about not changing  
the code. However, under Snow Leopard the deletion of test cases fails  
completely and aborts the program. The C-style cast in  
framework_impl::clear() does not work under Snow Leopard, but it does  
work correctly if test_unit has a virtual destructor. I suspect that  
the noise level over this is going to increase rapidly.

- Rush