gcc/testsuite/gcc.target/i386/sse4_1-insertps-1.c

   1 /* { dg-do run } */
   2 /* { dg-require-effective-target sse4 } */
   3 /* { dg-options "-O2 -msse4.1" } */
   4
   5 #include "sse4_1-check.h"
   6
   7 #include <smmintrin.h>
   8 #include <string.h>
   9
  10 #define msk0 0x01
  11 #define msk1 0x10
  12 #define msk2 0x29
  13 #define msk3 0x30
  14
  15 #define msk4 0xFC
  16 #define msk5 0x05
  17 #define msk6 0x0A
  18 #define msk7 0x0F
  19
  20 static void
  21 sse4_1_test (void)
  22 {
  23   union
  24     {
  25       __m128 x;
  26       float f[4];
  27     } res[8], val1, val2, tmp;
  28   int masks[8];
  29   int i, j;
  30
  31   val2.f[0] = 55.0;
  32   val2.f[1] = 55.0;
  33   val2.f[2] = 55.0;
  34   val2.f[3] = 55.0;
  35
  36   val1.f[0] = 1.;
  37   val1.f[1] = 2.;
  38   val1.f[2] = 3.;
  39   val1.f[3] = 4.;
  40
  41   res[0].x = _mm_insert_ps (val2.x, val1.x, msk0);
  42   res[1].x = _mm_insert_ps (val2.x, val1.x, msk1);
  43   res[2].x = _mm_insert_ps (val2.x, val1.x, msk2);
  44   res[3].x = _mm_insert_ps (val2.x, val1.x, msk3);
  45
  46   masks[0] = msk0;
  47   masks[1] = msk1;
  48   masks[2] = msk2;
  49   masks[3] = msk3;
  50
  51   for (i = 0; i < 4; i++)
  52     res[i + 4].x = _mm_insert_ps (val2.x, val1.x, msk4);
  53
  54   masks[4] = msk4;
  55   masks[5] = msk4;
  56   masks[6] = msk4;
  57   masks[7] = msk4;
  58
  59   for (i=0; i < 8; i++)
  60     {
  61       tmp = val2;
  62       tmp.f[(masks[i] & 0x30) >> 4] = val1.f[(masks[i] & 0xC0) >> 6];
  63
  64       for (j = 0; j < 4; j++)
  65         if (masks[i] & (0x1 << j))
  66           tmp.f[j] = 0.f;
  67
  68       if (memcmp (&res[i], &tmp, sizeof (tmp)))
  69         abort ();
  70     }
  71 }