The following code crashes sometimes(!) :

const int N = 9;
double d[N];
double s = 2.0;
for (int n=0;n<N;++n) d[n]=1.0;
Map<Matrix<double,Dynamic,Dynamic,RowMajor>,Unaligned,OuterStride<> > map(d,3,3,OuterStride<>(3));
map *= s;

why ??
(the crash in somewhere in PacketMath.h)

cannot reproduce. Could you provide the backtrace and also check your program with valgrind to see if the problem does not come from another part of the program.

Tnx for the reply.

I've further investigated the issue, and found the problem:
Well, i'm using MSVC2012, on x86, according to : http://msdn.microsoft.com/en-us/library ... 49(v=vs.71).aspx
the stack on x86/windows is aligned to 4 bytes only.
My crash happened when the pointer-to-double (the d buffer) was not aligned to 8 bytes but to 4.
The loop in Assign.h (line 456, run method) assume that even if the object is "unaligned" - it's still aligned to itself size, i.e 8 bytes in my double case,
So it try to call the copyPacket method on unaligned data (mm_load_pd) and crash (when the first double is not aligned to 8 bytes - no element in the array will be aligned to 16 bytes...).
Because the default in MSVC x86 is non aligned stack, and Eigen::Map may be used for external buffers, i think this situation should be treated inside eigen.
Doesn't it?

hm, ok so the fix would be to set PacketTraits<double>::AlignedOnScalar to false on windows 32bits. It is strange that we never hit this issue. Our test platforms do include 32bits windows.

Yep, it's working...
changed arch\SSE\PacketMath.h in packet_traits<double> to:

Code: Select all

#if defined(_WIN32) && !defined(_M_X64)
 AlignedOnScalar = 0 ,
#else
 AlignedOnScalar = 1 ,
#endif


In the future, when compilers will implement alignof() as within the c++11 spec., the definition can be :

Code: Select all

 AlignedOnScalar = alignof(double)<sizeof(double)?0:1 ,


but in the current microsoft's __alignof() implementation, __alignof(double)=8 although when a double is on the stack it may not be aligned to 8bytes.
(Maybe that's the reason microsoft didn't renamed the __alignof operator to spec's alignof).
Thanks !

hm, that's strange that __alignof(double) returns 8. How are you calling the function that produces the crash? Through a pointer function? Through multi-threading?

It was a simple main().
About __alignof, I've asked it in stackoverflow forum, and didn't get a complete answer :
http://stackoverflow.com/questions/20001135/alignofdouble-in-msvc-win32