[Insight-developers] Expecting a good dashboard tomorrow

Fri Nov 29 15:36:32 EST 2013

Looks like this is compiler bug relating to SSE and alignment. The Win32
and Win64 Debug builds all use the x87 stack so do not experience this
issue.

Here is the assembly code from one of the crashing tests:

    const SpacingType spacing = outputImage->GetSpacing();
000000013FE078C9  mov         rax,qword ptr [rcx]
000000013FE078CC  call        qword ptr [rax+180h]
000000013FE078D2  movaps      xmm0,xmmword ptr
[__xmm at 3ff00000000000003ff0000000000000 (0140120C10h)]

    for ( unsigned int i = 0; i < ImageDimension; i++ )
      {
      coeffs[i] = 1.0 / static_cast< double >( spacing[i] );
000000013FE078D9  divpd       xmm0,xmmword ptr [rax]

Following it with the debugger, RAX = 0000000002162018 at the divpd
instruction.

The memory operands of divpd must be aligned on 16-byte boundary[1] but as
we can see clearly here it is only aligned on 8-byte boundary.

RAX is last set by the call to GetSpacing (itkGetConstReferenceMacro), here
is the assembly for that:

 itkGetConstReferenceMacro(Spacing, SpacingType);
000000013F58B720  lea         rax,[rcx+0B8h]
000000013F58B727  ret

RCX = 0000000002111F60

I googled to see if anyone else were having this alignment issue, and it
seems not so I wonder if we were doing something to upset the compiler's
"natural" alignment. I tried to use /Zp16 [2] compiler option but it
appears to only align to the minimum of the "natural" alignment and the Zp
parameter, so it does not work. It seems we are forced to use
__declspec(align(16)) [3] for all array declaration to fix this issue but
that still leaves user and new code unprotected.

Does anyone have any suggestions on how to fix this issue?

[1] http://n0ah.org/mirrors/siyobik.info/instruction/DIVPD.html
[2] http://msdn.microsoft.com/en-us/library/xh3e3fd0.aspx
[3] http://msdn.microsoft.com/en-us/library/83ythb65%28v=vs.80%29.aspx

Ho Cheung
hocheung20 at gmail.com
Cell: (775) 388-2368

On Wed, Nov 27, 2013 at 7:23 AM, Bradley Lowekamp <blowekamp at mail.nih.gov>wrote:

> Just to follow up. On the same system I have a clean VS12 (2013) Win32, I
> have a nightly build for Win64 with 123 test failures. So this confirms
> that it's just a Win64 issue.
>
> Brad
>
>
>
> On Nov 26, 2013, at 9:46 AM, Simon Alexander <skalexander at gmail.com>
> wrote:
>
> I have the same suspicion.  If I can find the time I will look for root
> cause and a work around if possible.
>
> Unfortunately, regressions took precedence and I haven't been able to
> rerun the dashboard.  I'll look at this again when I have a free machine.
>
>
> On Tue, Nov 26, 2013 at 8:46 AM, Bradley Lowekamp <blowekamp at mail.nih.gov>wrote:
>
>> Hello,
>>
>> Here is my nightly with VS12 32-bit:
>>
>> http://open.cdash.org/buildSummary.php?buildid=3115540
>>
>> It comes back clean.
>>
>> There reports have also been that VS12 64-bit in Debug is clean.
>>
>> Also Simon reports:
>>
>> These seem to maybe be related to copy ctor on itk::Vector types and the
>> optimizer (VS12, 64 bit)
>>
>> e.g.
>>
>>
>> code that looks like this:
>>
>>  typedef typename TOutputImage::SpacingType SpacingType;
>>  const SpacingType spacing = outputImage->GetSpacing();
>>
>> leads to a segfault as soon as spacing is accessed (in Release mode, not
>> Debug)
>>
>> but
>>
>>  typedef typename TOutputImage::SpacingType SpacingType;
>>  const SpacingType &spacing = outputImage->GetSpacing();
>>
>>
>> I am very suspicious that this is not a bug in ITK.
>>
>>
>>
>>
>>
>> On Nov 25, 2013, at 10:39 PM, Ho Cheung <hocheung20 at gmail.com> wrote:
>>
>> Here’s v4.5rc01 on VS2013 Ultimate 64-bit Release mode. Seems like still
>> quite a lot of segfaults. I’ll have a closer look when I get some time off
>> over Thanksgiving weekend if not solved by then.
>>
>> http://open.cdash.org/buildSummary.php?buildid=3114945
>>
>> On Nov 25, 2013, at 8:52 AM, Simon Alexander <skalexander at gmail.com>
>> wrote:
>>
>> Bradley,
>>
>> Thanks, missed adding the list.  My builds are all 64 bit on 64 bit OS.
>>  I will pull a new copy of master  tonight and manually run the dashboard.
>>
>>
>> On Mon, Nov 25, 2013 at 9:36 AM, Bradley Lowekamp <blowekamp at mail.nih.gov
>> > wrote:
>>
>>> Simon,
>>>
>>> I am including the developers list.
>>>
>>> My nightly is 32-bit, perhaps thats related? I'll configure up a 64-bit
>>> build nightly too, to see if that's different.
>>>
>>> You likely should start with a clean master repo, as start with a fresh
>>> CMake configured build too. Results shared on the dashboard are welcomed.
>>>
>>> Brad
>>>
>>>
>>>
>>> On Nov 25, 2013, at 9:25 AM, Simon Alexander <skalexander at gmail.com>
>>> wrote:
>>>
>>> Hi Bradley,
>>>
>>> It is interesting that the build does not show the same pattern of
>>> errors.  If it is helpful I can run the ctest suite again here (win x64, 64
>>> bit builds) but there are certainly things I can't test (like the wrapping).
>>>
>>> cheers,
>>> Simon
>>>
>>>
>>> On Sat, Nov 23, 2013 at 4:07 PM, Bradley Lowekamp <
>>> blowekamp at mail.nih.gov> wrote:
>>>
>>>> Hello,
>>>>
>>>> Thanks for all who have contributed patches and reviews in the last
>>>> couple of days in preparation of the release candidate.
>>>>
>>>> I am expecting a good dashboard tomorrow, which should have WrapITK,
>>>> and VS12 ( 2013 ) green. Here are a couple experimental build I have fired
>>>> off:
>>>>
>>>> http://open.cdash.org/buildSummary.php?buildid=3112037
>>>> http://open.cdash.org/buildSummary.php?buildid=3112076
>>>>
>>>> Interestingly, VS12 has not test failure which is different than had
>>>> been reported by others.
>>>>
>>>> Thanks,
>>>> Brad
>>>> _______________________________________________
>>>> Powered by www.kitware.com
>>>>
>>>> Visit other Kitware open-source projects at
>>>> http://www.kitware.com/opensource/opensource.html
>>>>
>>>> Kitware offers ITK Training Courses, for more information visit:
>>>> http://kitware.com/products/protraining.php
>>>>
>>>> Please keep messages on-topic and check the ITK FAQ at:
>>>> http://www.itk.org/Wiki/ITK_FAQ
>>>>
>>>> Follow this link to subscribe/unsubscribe:
>>>> http://www.itk.org/mailman/listinfo/insight-developers
>>>>
>>>
>>>
>>>
>> _______________________________________________
>> Powered by www.kitware.com
>>
>> Visit other Kitware open-source projects at
>> http://www.kitware.com/opensource/opensource.html
>>
>> Kitware offers ITK Training Courses, for more information visit:
>> http://kitware.com/products/protraining.php
>>
>> Please keep messages on-topic and check the ITK FAQ at:
>> http://www.itk.org/Wiki/ITK_FAQ
>>
>> Follow this link to subscribe/unsubscribe:
>> http://www.itk.org/mailman/listinfo/insight-developers
>>
>>
>>
>>
>
>
ue,
-------------- next part --------------
An HTML attachment was scrubbed...
URL: <http://www.itk.org/pipermail/insight-developers/attachments/20131129/df125f9e/attachment.htm>