From: Beman Dawes (bdawes_at_[hidden])
Date: 2001-06-07 10:48:49
At 06:50 AM 6/7/2001, John Maddock wrote:
>> >Same things with scoped_ptr_example_test.cpp.
>> >Is this bcb5 only bug ?
>>
>>I was able to reproduce the problem for the Borland compiler but not
some
>>of the others.
I've now verified that the Win32 GCC, Intel, Metrowerks, and Microsoft
compilers all handle this correctly, while Borland does not.
>It looks to be a Borland bug: for some reason the destructors of all
>sub-objects get instantiated when you write:
>
>T my_instance;
>
>Even if T has all constructors/destructor declared but not defined.
>
>This basically makes it impossible to implement the pimpl idon using a
>smart pointer with this compiler (or any others that have the same
>problem). Note that commenting out the static assert is *not* an option:
>the code compiles then, but issues a warning about deleting an incomplete
>type - as a result it will not behave correctly at run time - or rather
it
>*may* not, depending on which version or the smart_ptr's destructor makes
>it into the exe (there may be different versions in different translation
>units, some instantiated "prematurely", others not).
You were smart to figure that out. In my testing I just commented out the
assert and dumbly assumed everything was OK because the compile errors went
away. But I think that your analysis is correct, including the key point:
It isn't just the Boost smart pointer - any use of a templated pointer to
an incomplete type could fail if a translation unit with delete of an
incomplete type happens to be the one linked in.
This looks like a really disastrous bug in the Borland compiler.
I wonder if we've overlooked something? It seems like Borland users would
be howling by now if this bug is really biting people.
--Beman