[Insight-developers] Empty FixedArray destructor: Performancehit using gcc (times 2)

[Tom Vercauteren]

Just to try and make things clearer. The current memory alignment is
not actually wrong. The computation results are indeed correct.
However the pointer is simply not aligned on 8 bytes which is thus not
optimal in terms of performance. When I said "bad alignment", I
actually meant "alignment leading to slow memory accesses".
Tom

On Thu, Jun 5, 2008 at 11:30 PM, Tom Vercauteren
<tom.vercauteren at m4x.org> wrote:
>> I am really surprised that C++ does not generate a proper alignment
>> for double. What flags are you using for your build?
>
> Nothing special. I tried
> -O0
> -O2 -g
> -O2
> and
> -O3
>
> The behavior is rather consistent across these flags. It is also
> consistent across computers. We tested on 2 ubuntu boxes with gcc 4.1
> and a mac running Mac OS 10.5 and mac gcc 4.0
>
> Could you run the test I submitted? Was it inconclusive?
>
> Tom
>
>>
>> Bill
>>
>> On Thu, Jun 5, 2008 at 3:49 PM, Tom Vercauteren <tom.vercauteren at m4x.org> wrote:
>>> Hi,
>>>
>>> As shown in the code I sent, the real performance hit does NOT come
>>> from the destructor itself but from the memory alignment.
>>>
>>> The example shows that by using the same vanilla FixedArray class
>>> (where the destructor is explicitly implemented as empty), the well
>>> aligned case is faster than the badly aligned case.
>>>
>>> What is weird is why the empty destructor leads to bad memory alignment...
>>>
>>> What is striking is that if you replace the c array of itk::FixedArray
>>> by a std::vector<itk::FixedArray>, you don't get any performance hit.
>>> I guess that STL takes care of aligning things in an optimized manner.
>>>
>>> By the way, this also implies that if ITK was using STL vector instead
>>> of c array in its image class, we wouldn't have stumbled on the issue.
>>>
>>> Anyhow, I agree with Karthik and Bill. In this specific case,
>>> explicitly implementing an empty destructor seems useless.
>>>
>>> Tom
>>>
>>> On Thu, Jun 5, 2008 at 9:02 PM, Niels Dekker <niels-xtk at xs4all.nl> wrote:
>>>> Bill Lorensen wrote:
>>>>>
>>>>> Here is an FAQ on destructors:
>>>>> http://www.informit.com/guides/content.aspx?g=cplusplus&seqNum=370
>>>>> They say that an empty destructor is pretty much useless.
>>>>
>>>> The destructor of FixedArray<TValueType, VLength> invokes the destructor of
>>>> TValueType for all of its elements.  So it may not be as empty as it
>>>> seems...  When the destructor is removed from FixedArray, it will be
>>>> implicitly defined as an inline function, as the FAQ you're referring to
>>>> says.  This might give a performance gain, but it might also produce a
>>>> larger binary.  Things might get worse (or better?) when TValueType itself
>>>> has also has an inline destructor.
>>>>
>>>> In some special cases, the destructor of a class /must/ be implemented in a
>>>> CXX file, even if it is empty. As is the case discussed at GameDev.Net,
>>>> "auto_ptr within Pimpl Class",
>>>> http://www.gamedev.net/community/forums/topic.asp?topic_id=437587
>>>>
>>>> Anyway, just my 2 cents.  It might still be okay to remove FixedArray's
>>>> destructor...
>>>>
>>>> Kind regards,
>>>> --
>>>> Niels Dekker
>>>> http://www.xs4all.nl/~nd/dekkerware
>>>> Scientific programmer at LKEB, Leiden University Medical Center
>>>>
>>>
>>
>