2 /* { dg-require-effective-target sse4 } */
3 /* { dg-options "-O2 -msse4.1" } */
6 #define CHECK_H "sse4_1-check.h"
10 #define TEST sse4_1_test
15 #include <smmintrin.h>
49 res[0].x = _mm_dp_pd (val1.x, val2.x, HIMASK | lmskN);
50 res[1].x = _mm_dp_pd (val1.x, val2.x, HIMASK | lmsk0);
51 res[2].x = _mm_dp_pd (val1.x, val2.x, HIMASK | lmsk1);
52 res[3].x = _mm_dp_pd (val1.x, val2.x, HIMASK | lmsk01);
54 masks[0] = HIMASK | lmskN;
55 masks[1] = HIMASK | lmsk0;
56 masks[2] = HIMASK | lmsk1;
57 masks[3] = HIMASK | lmsk01;
59 for (i = 0; i < 4; i++)
63 for (j = 0; j < 2; j++)
64 if (HIMASK & (0x10 << j))
65 tmp = tmp + (val1.d[j] * val2.d[j]);
67 for (j = 0; j < 2; j++)
68 if ((masks[i] & (1 << j)) && res[i].d[j] != tmp)