[cxx-abi-dev] ABI modification for exception propagation

[David Vandevoorde]

On May 27, 2008, at 4:39 AM, Sebastian Redl wrote:
> David Vandevoorde wrote:
>>
>> Note that it doesn't have to copy the thrown object: It can  
>> essentially return a smartptr<__cxa_exception> (which is what  
>> std::exception_ptr would be) for the top exception.  The  
>> constructor for this smart pointer must behave a bit like  
>> __cxa_rethrow() in that it must avoid having __cxa_end_catch  
>> destroy the exception.  Instead, that destruction must now be  
>> handled by the destructor of smartptr<__cxa_exception> (when the  
>> reference count goes to zero).
> I have tried that. It doesn't work. Yes, I can make an exception_ptr  
> that refers to an existing __cxa_exception, but I can't rethrow it.  
> That would mean allowing a single __cxa_exception to be in multiple  
> throws at once, and _Unwind_Exception cannot handle that. Also,  
> __cxa_exception cannot handle having been caught by several threads  
> at once. It would mess up the nextException chain.
> I HAVE to duplicate the __cxa_exception, and as long as the  
> __cxa_exception and the exception object are in the same memory  
> block, that means copying the exception object too.
>
> You can read my rationale in the thread at the libstdc++ mailing list.
> http://gcc.gnu.org/ml/libstdc++/2008-05/msg00079.html


Thanks -- that's a nice write-up!


> Although, thinking about it some more, there might be a way. There  
> could be essentially two kinds of __cxa_exception. One looks like  
> the old one, with an added reference count. The other holds a  
> pointer to the first kind instead. All exception_ptrs refer to the  
> primary __cxa_exception, but on rethrow they allocate a secondary  
> that also refers to the first.


That's an interesting idea.


> This might work. But *only* if all modules link to the support  
> library dynamically, of course. If any of them links statically, it  
> would keep the old implementation, which would wreak havoc.


What sort of failure modes would we expect.  Could the  
"__cxa_exception by reference" be made to look like std::bad_exception  
to an old runtime support library?

> I'll look into implementing that.


FWIW, I've contacted Peter Dimov about this issue, and he seems  
agreeable to prohibit rethrowing an exception captured with  
current_exception() using "throw;" (until it has been rethrown with  
rethrow_exception, at which point it can be caught/rethrown the usual  
way).

Note that we're still dealing with a working paper, not a frozen draft  
or standard.  So if this kind of tricky issues arise, there may be  
wiggle room left to fix the spec.  (But not for long.)

	Daveed