GCC & Alignment

[GReaper]

I want to share something i discovered yesterday.

All i wanted was to have an object aligned in memory by 16 in order to use SSE2 instructions for efficiency. But i thought that GCC alignment extends to memory other than the struct itself. How wrong i was! After using the debugger and viewing the disassembly, i can surely say that GCC doesn't respect alignment rules on the stack. Shouldn't it?

This was the code i compiled with the latest GCC and then viewed with the debugger:

#pragma pack(push, 16)
struct vec3d_t
{
    float X;
    float Y;
    float Z;
};
#pragma pack(pop)

int main()
{
    int dummy1 = 0;
    vec3d_t myVec = { 0.0f, 0.0f, 0.0f };
    int dummy2 = 0;

    int myRes = (int)(myVec.X+myVec.Y+myVec.Z+dummy1+dummy2); // Don't let the compiler optimize them away

    return myRes;
}

This was the ASM output:
( This isn't actually the version of GCC i compiled it with, as my flash drive i used right now to output this holds a little older version, but it is almost the same )

004012F0	push   %ebp
004012F1	mov    %esp,%ebp
004012F3	push   %edi
004012F4	sub    $0x44,%esp
004012F7	and    $0xfffffff0,%esp ; Stack aligned to 16
004012FA	mov    $0x0,%eax
004012FF	add    $0xf,%eax
00401302	add    $0xf,%eax
00401305	shr    $0x4,%eax
00401308	shl    $0x4,%eax
0040130B	mov    %eax,-0x3c(%ebp)
0040130E	mov    -0x3c(%ebp),%eax
00401311	call   0x401880 <_alloca>
00401316	call   0x401400 <__main>
0040131B	movl   $0x0,-0xc(%ebp)
00401322	lea    -0x28(%ebp),%edi ; Memory alignment UNKNOWN
00401325	cld
00401326	mov    $0x0,%edx
0040132B	mov    $0x3,%eax
00401330	mov    %eax,%ecx
00401332	mov    %edx,%eax
00401334	rep stos %eax,%es:(%edi)
00401336	movl   $0x0,-0x2c(%ebp)
0040133D	flds   -0x28(%ebp)
00401340	fadds  -0x24(%ebp)
00401343	fadds  -0x20(%ebp)
00401346	fildl  -0xc(%ebp)
00401349	faddp  %st,%st(1)
0040134B	fildl  -0x2c(%ebp)
0040134E	faddp  %st,%st(1)
00401350	fnstcw -0x32(%ebp)
00401353	movzwl -0x32(%ebp),%eax
00401357	or     $0xc00,%ax
0040135B	mov    %ax,-0x34(%ebp)
0040135F	fldcw  -0x34(%ebp)
00401362	fistpl -0x30(%ebp)
00401365	fldcw  -0x32(%ebp)
00401368	mov    -0x30(%ebp),%eax
0040136B	mov    -0x4(%ebp),%edi
0040136E	leave
0040136F	ret

See?! Very simple! ( Well, to me it is )

[Codeplug]

Alignment and structure packing are not the same thing.

When you pack a structure to N, the structures members will be on a boundary that is either a multiple of N, or a multiple of the size of the member, whichever is smaller.

Also, don't confuse structure alignment with the alignment of it's individual members.

http://gcc.gnu.org/onlinedocs/gcc/Ty...attribute-2583

or for MSVC:

http://msdn.microsoft.com/en-us/library/83ythb65.aspx
http://msdn.microsoft.com/en-us/library/71kf49f1.aspx

gg

[GReaper]

Thanks, Codeplug! That link solves my problem.