old skool uncached memcpy x86

// memcpy_fast() & memset32()
// - lifted from http://www.stereopsis.com/memcpy.html, though I wrote similar ones back in the day
// - approx. twice faster than their CRT counterparts on a 2009 netbook with 1.66GHz Intel Atom CPU
// - combined writes (MOVNTQ) perform favorably for most usage patterns

void memcpy_fast(void *pDest, const void *pSrc, size_t numBytes)
{
	// numBytes must be a multiple of 64 -- use memcpy() for general purpose
	VIZ_ASSERT(!(numBytes & 63));

	// an 8-byte boundary gaurantees correctly aligned reads and writes
	VIZ_ASSERT( 0 ==  (int(pSrc) & 7) );
	VIZ_ASSERT( 0 == (int(pDest) & 7) );

	__asm
	{
		mov    edi, pDest
		mov    esi, pSrc
		mov    ecx, numBytes
		shr    ecx, 6
	_loop:
		movq   mm1, [esi]
		movq   mm2, [esi+8]
		movq   mm3, [esi+16]
		movq   mm4, [esi+24]
		movq   mm5, [esi+32]
		movq   mm6, [esi+40]
		movq   mm7, [esi+48]
		movq   mm0, [esi+56]
		movntq [edi], mm1
		movntq [edi+8], mm2
		movntq [edi+16], mm3
		movntq [edi+24], mm4
		movntq [edi+32], mm5
		movntq [edi+40], mm6
		movntq [edi+48], mm7
		movntq [edi+56], mm0
		add    esi, 64
		add    edi, 64
		dec    ecx
		jnz    _loop
	}

	_mm_empty();
}