Problems using Vector Extensions

[Thomas Unterthiner]

Hi there!
I'm having trouble using the vector-extensions, using GCC 4.2.1 on 
Windows XP. I have a simple class for 3D-vectors that I like to speed up 
using the extensions. The class looks like this:


*#define* USE_GCC_VECTORISATION (defined(__SSE__) && 1)

*class* Vector3D
{
    *public*:
        //*
         * ... the usual operators
         *//

    *private*:*
        #if* USE_GCC_VECTORISATION
            *typedef float* v4sf __attribute__ (( vector_size(4 * 
*sizeof*(*float*)) ));
            *typedef int* v4si __attribute__ (( vector_size(4 * 
*sizeof*(*int*)) ));
            *union*
            {
                *float* v[4];
                v4sf vec;
            };
            Vector3D(*const* v4sf& v);*
        #else
            float* v[3];*
        #endif*
};


Now, as soon as I activate the vectorization, I get segfaults whenever a 
vector-op is involved. E.g. this function always segfaults:

Triangle::Triangle(const std::string& name, const Material& material,
                  const Vector3D& v1, const Vector3D& v2, const 
Vector3D& v3)
    : TraceableObject(name, material)

{
    v[0] = v1;
    v[1] = v2;
    v[2] = v3;
    normal = crossProduct(v2 - v1, v3 - v1);   // segfault when 
computing v2 - v1
    normal.normalize();
    planeD = dotProduct(normal, v1);
}


The corresponding operator- looks like this:

*inline const* Vector3D *operator*-(*const* Vector3D& lhs, *const* 
Vector3D& rhs)
{*
#if* USE_GCC_VECTORISATION
    *return* Vector3D(lhs.vec - rhs.vec);*
#else*
    Vector3D retval(lhs);
    retval -= rhs;
    *return* retval;*
#endif*
}

More precisely, this is the disassembled code the debugger shows me:

Frame *function*: operator-(Vector3D const&, Vector3D const&) 
(src/Vector3D.cpp:138)
Frame address : 0x0022FA60
-------------------------------------------------------------------------------- 

0x4f07d0    *push*   %*ebp*
0x4f07d1    *mov*    %*esp*,%*ebp*
0x4f07d3    *push*   %*ebx*
0x4f07d4    *sub*    $0x1c,%*esp*
0x4f07d7    *mov*    0x8(%*ebp*),%*ebx*
0x4f07da    *mov*    %*ebx*,%*edx*
0x4f07dc    *mov*    0xc(%*ebp*),%*eax*
0x4f07df    movaps (%*eax*),%xmm1  /; <<<<-----the segault happens in 
this instruction/
0x4f07e2    *mov*    0x10(%*ebp*),%*eax*
0x4f07e5    movaps (%*eax*),%xmm0
0x4f07e8    movaps %xmm1,%xmm2
0x4f07eb    subps  %xmm0,%xmm2
0x4f07ee    movaps %xmm2,%xmm0
0x4f07f1    movaps %xmm0,0xffffffe8(%*ebp*)
0x4f07f5    *lea*    0xffffffe8(%*ebp*),%*eax*
0x4f07f8    *mov*    %*eax*,0x4(%*esp*)
0x4f07fc    *mov*    %*edx*,(%*esp*)
0x4f07ff    *call*   0x46cefc <Vector3D::Vector3D(float __vector const&)>
0x4f0804    *mov*    %*ebx*,%*eax*
0x4f0806    *add*    $0x1c,%*esp*
0x4f0809    *pop*    %*ebx*
0x4f080a    *pop*    %*ebp*
0x4f080b    *ret*    $0x4

At the moment of the segfault, EAX contains the address of lhs. Every 
other piece of data also seems to have the expected values in it. 
Therefore I can't see any error neither in my usage of the 
vector-extensions nor in the generated code. To further mystify things, 
the following code works like a charm (no segfaults, no nothing, 
yielding the correct output):

*#include* "_Vector3D.h_"
*int* main ( *int* argc, *char*** argv )
{
    ///  this is quite exactly what happens in my "real" program/
    Vector3D a(-4, -3, 10);
    Vector3D b(-4, 5, 10);
    Vector3D c = a - b;

    std::cout << c << std::endl;

    *return* 0;
}


So I fail to see what is really going on. Any hints?

best regards
Thomas Unterthiner