MMX Programming, part 2

[Ken g6]

The thread on SIMD and MMX programming got me wanting to try MMX, SSE2, and intrinsics. So far I'm only working with MMX (on a processor that doesn't support SSE2), and I have two compiler errors I can't get rid of.

The idea is to have a C string, and load parts of it into MMX registers. First of all, I assume that MMX can't read misaligned quadwords (or can't read them efficiently). But when I do:

static __declspec(align(8)) // Align arrays by 8
unsigned char buf[96];

I consistently get:

warning: `align' attribute directive ignored

That almost exactly matches some code I found online, so I'm stumped.

After that, I need to read an aligned chunk of 8 bytes at once. The only example code I have is for SSE2, which uses the _mm_load_si128 instruction. But there is no _mm_load_si64 instruction that I can find, and everything else I try reports:

error: can't convert value to a vector
error: can't convert between vector values of different size

Any ideas?

Thanks!

 

[nickbits]

what compiler are you using?

 

[Ken g6]

gcc (GCC) 3.4.4 (cygming special) (gdc 0.12, using dmd 0.125)
and
gcc (GCC) 3.4.5 (mingw-vista special r3)
and
gcc (4.3.2-tdm-2 mingw32) 4.3.2

All three give at least the align warning and "can't convert value to a vector".

I suppose the next question is what options I'm using: -msse -O3 -Wall
and what header files I included:
#include <stdio.h>
#include <string.h>
#include <xmmintrin.h>

(I'm using _mm_movemask_pi8 elsewhere, which is why it's xmmintrin.h.)

 

[Ken g6]

OK, I resolved the vector conversion problem, I think, with this macro:
#define ToVectorMem(x) (*((__m64*)&(x)))

I was trying to copy something from http://www.agner.org/optimize/ but I needed an extra *.

Does MMX alignment not matter, so it's not available with MMX-level intrinsics?