[webkit-dev] Fix for Vector::m_inlineBuffer alignment?

[Maciej Stachowiak]

On Sep 7, 2008, at 10:20 AM, Paul Pedriana wrote:

> I agree that a union is most portable, but there are two problems  
> with it:
>
>   - The C++ Standard (section 9.5 p1) disallows union members that
>      have ctors, dtors, or assignment operators. This includes the  
> case
>      whereby the class is the only non-POD of the union.
>
>   - It would create an instance of type T. Thus an empty Vector would
>      construct an instance of T, resizing it would overwrite the  
> instance of T,
>      and on destruction it would destruct the T.

Oliver also pointed out the constructor / destructor issue, so yeah,  
this wouldn't work. We could use double to force 8-byte alignment but  
that might be wasteful for vector types where wider alignment is not  
required.

  - Maciej

>
>
> Paul
>
>
>
>> On Sep 4, 2008, at 1:20 AM, Paul Pedriana wrote:
>>
>>> Well the proposed solution in
>>> <https://bugs.webkit.org/show_bug.cgi?id=16925> doesn't work, as VC 
>>> ++
>>> doesn't accept that syntax. And the solutions in
>>> <https://bugs.webkit.org/show_bug.cgi?id=19775> whereby a uint32_t  
>>> or
>>> uint64_t buffer are made don't work for larger types and use more  
>>> memory
>>> than necessary. They might be useful as a fallback for compilers  
>>> that
>>> don't support alignment directives, though the primary  
>>> implementations
>>> of every compiler that I work with (GCC, VC++, CodeWarrior, EDG, SN,
>>> IBM) support it.
>>
>> The buffer should definitely remain char to avoid strict aliasing  
>> problems. However, I think the most portable way to ensure  
>> alignment is to union the buffer with a one-element array of the  
>> template parameter type T.
>>
>> - Maciej
>>
>>
>>> I'll make a patch and attach it to
>>> <https://bugs.webkit.org/show_bug.cgi?id=16925>, if that's OK.
>>
>>
>>> Paul
>>>
>>>> On Sep 4, 2008, at 2:05 AM, Paul Pedriana wrote:
>>>>
>>>>> I see that JavaScriptCore/wtf/Vector.h has this:
>>>>>
>>>>>  // FIXME: Nothing guarantees this buffer is appropriately  
>>>>> aligned to
>>>>> hold objects of type T.
>>>>>  char m_inlineBuffer[m_inlineBufferSize];
>>>>
>>>>
>>>> We have a bug for this:
>>>> <https://bugs.webkit.org/show_bug.cgi?id=16925>. Also,
>>>> <https://bugs.webkit.org/show_bug.cgi?id=19775>.
>>>>
>>>> There are several patches and suggestions in these bugs, which wait
>>>> for someone to submit them as clean patches.
>>>>
>>>> - WBR, Alexey Proskuryakov
>>>>
>>>>
>>>>
>>>
>>> _______________________________________________
>>> webkit-dev mailing list
>>> webkit-dev at lists.webkit.org
>>> http://lists.webkit.org/mailman/listinfo.cgi/webkit-dev
>>
>>
>