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 #ifndef CHECK_H
   6 #define CHECK_H "sse4_1-check.h"
   7 #endif
   8
   9 #ifndef TEST
  10 #define TEST sse4_1_test
  11 #endif
  12
  13 #include CHECK_H
  14
  15 #include <smmintrin.h>
  16 #include <string.h>
  17
  18 #define msk0 0x01
  19 #define msk1 0x10
  20 #define msk2 0x29
  21 #define msk3 0x30
  22
  23 #define msk4 0xFC
  24 #define msk5 0x05
  25 #define msk6 0x0A
  26 #define msk7 0x0F
  27
  28 static void
  29 TEST (void)
  30 {
  31   union
  32     {
  33       __m128 x;
  34       float f[4];
  35     } res[8], val1, val2, tmp;
  36   int masks[8];
  37   int i, j;
  38
  39   val2.f[0] = 55.0;
  40   val2.f[1] = 55.0;
  41   val2.f[2] = 55.0;
  42   val2.f[3] = 55.0;
  43
  44   val1.f[0] = 1.;
  45   val1.f[1] = 2.;
  46   val1.f[2] = 3.;
  47   val1.f[3] = 4.;
  48
  49   res[0].x = _mm_insert_ps (val2.x, val1.x, msk0);
  50   res[1].x = _mm_insert_ps (val2.x, val1.x, msk1);
  51   res[2].x = _mm_insert_ps (val2.x, val1.x, msk2);
  52   res[3].x = _mm_insert_ps (val2.x, val1.x, msk3);
  53
  54   masks[0] = msk0;
  55   masks[1] = msk1;
  56   masks[2] = msk2;
  57   masks[3] = msk3;
  58
  59   for (i = 0; i < 4; i++)
  60     res[i + 4].x = _mm_insert_ps (val2.x, val1.x, msk4);
  61
  62   masks[4] = msk4;
  63   masks[5] = msk4;
  64   masks[6] = msk4;
  65   masks[7] = msk4;
  66
  67   for (i=0; i < 8; i++)
  68     {
  69       tmp = val2;
  70       tmp.f[(masks[i] & 0x30) >> 4] = val1.f[(masks[i] & 0xC0) >> 6];
  71
  72       for (j = 0; j < 4; j++)
  73         if (masks[i] & (0x1 << j))
  74           tmp.f[j] = 0.f;
  75
  76       if (memcmp (&res[i], &tmp, sizeof (tmp)))
  77         abort ();
  78     }
  79 }