Original email...

Date: Sat, 10 Sep 2011 13:01:20 -0500
From: Patrick Baggett
To: SDL Development List <sdl@lists.libsdl.org>
Subject: Re: [SDL] SDL_memcpyMMX uses SSE instructions

In SDL_blit_copy.c, the function SDL_memcpyMMX() actually use SSE
instructions.

It is called in this context:

#ifdef __MMX__
    if (SDL_HasMMX() &&
        !((uintptr_t) src & 7) && !(srcskip & 7) &&
        !((uintptr_t) dst & 7) && !(dstskip & 7)) {
        while (h--) {
            SDL_memcpyMMX(dst, src, w);
            src += srcskip;
            dst += dstskip;
        }
        _mm_empty();
        return;
    }
#endif

This implies that the minimum CPU features are just MMX. There is a
separate SDL_memcpySSE() function.


The SDL_memcpyMMX() function does:

#ifdef __SSE__
        _mm_prefetch(src, _MM_HINT_NTA);
#endif

...which tests at compile time if SSE intrinsics are available, not at run
time. It generates the PREFETCHNTA instruction. It also uses _mm_stream_pi()
intrinsic, which generates the MOVNTQ instruction.

If you replace the "MMX" code with:

__m64* d64 = (__m64*)dst;
__m64* s64 = (__m64*)src;
 for(i= len / 64; i--;) {
   d64[0] = s64[0];
   d64[1] = s64[1];
   d64[2] = s64[2];
   d64[3] = s64[3];
   d64[4] = s64[4];
   d64[5] = s64[5];
   d64[6] = s64[6];
   d64[7] = s64[7];
   d64 += 8;
   s64 += 8;
 }

Then MSVC generates the correct movq instructions. GCC (4.5.0) seems to
think that using 2x movl is still better, but then again, GCC isn't actually
that good at optimizing intrinsics as I've found. At least the code won't
crash on my P2 though. :)

Also, there is no requirement for MMX to be aligned to the 8th byte. I
think the author assumed that SSE's 16 byte alignment requirement must
retroactively mean that MMX requires 8 byte alignment. Attached is the full
patch.

Patrick