gcc/testsuite/gcc.target/i386/sse4_1-blendps.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 #include <stdlib.h>
  18
  19 /* mingw runtime don't provide random().  */
  20 #ifdef __MINGW32__
  21 #define random rand
  22 #endif
  23
  24 #define NUM 20
  25
  26 #ifndef MASK
  27 #define MASK 0x0f
  28 #endif
  29
  30 static void
  31 init_blendps (float *src1, float *src2)
  32 {
  33   int i, sign = 1;
  34
  35   for (i = 0; i < NUM * 4; i++)
  36     {
  37       src1[i] = i * i * sign;
  38       src2[i] = (i + 20) * sign;
  39       sign = -sign;
  40     }
  41 }
  42
  43 static int
  44 check_blendps (__m128 *dst, float *src1, float *src2)
  45 {
  46   float tmp[4];
  47   int j;
  48
  49   memcpy (&tmp[0], src1, sizeof (tmp));
  50   for (j = 0; j < 4; j++)
  51     if ((MASK & (1 << j)))
  52       tmp[j] = src2[j];
  53
  54   return memcmp (dst, &tmp[0], sizeof (tmp));
  55 }
  56
  57 static void
  58 TEST (void)
  59 {
  60   __m128 x, y;
  61   union
  62     {
  63       __m128 x[NUM];
  64       float f[NUM * 4];
  65     } dst, src1, src2;
  66   union
  67     {
  68       __m128 x;
  69       float f[4];
  70     } src3;
  71   int i;
  72
  73   init_blendps (src1.f, src2.f);
  74
  75   for (i = 0; i < 4; i++)
  76     src3.f[i] = (int) random ();
  77
  78   /* Check blendps imm8, m128, xmm */
  79   for (i = 0; i < NUM; i++)
  80     {
  81       dst.x[i] = _mm_blend_ps (src1.x[i], src2.x[i], MASK);
  82       if (check_blendps (&dst.x[i], &src1.f[i * 4], &src2.f[i * 4]))
  83         abort ();
  84     }
  85
  86    /* Check blendps imm8, xmm, xmm */
  87   x = _mm_blend_ps (dst.x[2], src3.x, MASK);
  88   y = _mm_blend_ps (src3.x, dst.x[2], MASK);
  89
  90   if (check_blendps (&x, &dst.f[8], &src3.f[0]))
  91     abort ();
  92
  93   if (check_blendps (&y, &src3.f[0], &dst.f[8]))
  94     abort ();
  95 }