test_vinsertps.cpp

test_vinsertps.cpp

void  test_vinsertps() {

	__m128i m128_1;
	unsigned long long allOneLL = 0x2222222211111111LL;
	unsigned long long allTwoLL = 0x4444444433333333LL;

	
	((unsigned long long*)&m128_1)[0] = allOneLL;
	((unsigned long long*)&m128_1)[1] = allTwoLL;

	__asm {
		pxor		xmm5, xmm5
		vpshufd		xmm5, m128_1, 228
		//11 10 01 00 == 228, copy m128_1 into xmm5

		pxor		xmm4, xmm4
		//vinsertps	xmm4, xmm5, m128_1, 144
		//144 == 10 01  00 00
		//source 10 = 2, because m128_1 is a memory location, source is set to 0
		//dest	 01 = 1, xmm4[63:32]
		//insert m128_1[31:0] into xmm4[63:32], and copy other bits of xmm5 into xmm4
		//xmm4 == 0x4444_4444_3333_3333_1111_1111_1111_1111

		//vinsertps	xmm4, xmm5, xmm5, 144
		//144 == 10 01 00 00
		//source 10 = 2, xmm5[95:64]
		//dest	 01 = 1, xmm4[63:32]
		//insert xmm5[95:64] into xmm4[63:32], and copy other bits from xmm5 into xmm4
		//xmm4 == 0x4444_4444_3333_3333_3333_3333_1111_1111

		//vinsertps	xmm4, xmm5, xmm5, 192
		//192 == 11 00  00 00
		//source 11 == 3, xmm5[127:96]
		//dest   00 == 1, xmm4[31:0]
		//insert xmm5[127:96] into xmm4[31:0], and copy other bits from xmm5 into xmm4
		//xmm4 == 0x4444_4444_3333_3333_2222_2222_4444_4444


		vinsertps xmm4, xmm5, xmm5, 195
		//195 == 11 00  00 11
		//source 11 == 3, xmm5[127:96]
		//dest   00 == 1, xmm4[31:0]
		//mask   0011 == xmm4[63:32] == xmm4[31:0] == 0
		//xmm4 == 0x4444_4444_3333_3333_0000_0000_0000_0000

		vmovapd		m128_1, xmm4
	}
	
	unsigned long long LL1, LL2;
	LL1 = 0LL;
	LL2 = 0LL;
	
	LL1 = ((unsigned long long*)&m128_1)[0];
	LL2 = ((unsigned long long*)&m128_1)[1];
	printf("LL1=%llx \nLL2=%llx\n", LL1, LL2);

}