gcc/testsuite/gcc.target/powerpc/sse4_1-blendps.c

   1 /* { dg-do run } */
   2 /* { dg-require-effective-target p8vector_hw } */
   3 /* { dg-options "-O2 -mvsx -Wno-psabi" } */
   4 /* { dg-additional-options "-mdejagnu-cpu=power8" { target { ! has_arch_pwr8 } } } */
   5
   6 #ifndef CHECK_H
   7 #define CHECK_H "sse4_1-check.h"
   8 #endif
   9
  10 #ifndef TEST
  11 #define TEST sse4_1_test
  12 #endif
  13
  14 #include CHECK_H
  15
  16 #include <smmintrin.h>
  17 #include <string.h>
  18 #include <stdlib.h>
  19
  20 #define NUM 20
  21
  22 #ifndef MASK
  23 #define MASK 0x0f
  24 #endif
  25
  26 static void
  27 init_blendps (float *src1, float *src2)
  28 {
  29   int i, sign = 1;
  30
  31   for (i = 0; i < NUM * 4; i++)
  32     {
  33       src1[i] = i * i * sign;
  34       src2[i] = (i + 20) * sign;
  35       sign = -sign;
  36     }
  37 }
  38
  39 static int
  40 check_blendps (__m128 *dst, float *src1, float *src2)
  41 {
  42   float tmp[4];
  43   int j;
  44
  45   memcpy (&tmp[0], src1, sizeof (tmp));
  46   for (j = 0; j < 4; j++)
  47     if ((MASK & (1 << j)))
  48       tmp[j] = src2[j];
  49
  50   return memcmp (dst, &tmp[0], sizeof (tmp));
  51 }
  52
  53 static void
  54 TEST (void)
  55 {
  56   __m128 x, y;
  57   union
  58     {
  59       __m128 x[NUM];
  60       float f[NUM * 4];
  61     } dst, src1, src2;
  62   union
  63     {
  64       __m128 x;
  65       float f[4];
  66     } src3;
  67   int i;
  68
  69   init_blendps (src1.f, src2.f);
  70
  71   for (i = 0; i < 4; i++)
  72     src3.f[i] = (int) rand ();
  73
  74   /* Check blendps imm8, m128, xmm */
  75   for (i = 0; i < NUM; i++)
  76     {
  77       dst.x[i] = _mm_blend_ps (src1.x[i], src2.x[i], MASK);
  78       if (check_blendps (&dst.x[i], &src1.f[i * 4], &src2.f[i * 4]))
  79         abort ();
  80     }
  81
  82    /* Check blendps imm8, xmm, xmm */
  83   x = _mm_blend_ps (dst.x[2], src3.x, MASK);
  84   y = _mm_blend_ps (src3.x, dst.x[2], MASK);
  85
  86   if (check_blendps (&x, &dst.f[8], &src3.f[0]))
  87     abort ();
  88
  89   if (check_blendps (&y, &src3.f[0], &dst.f[8]))
  90     abort ();
  91 }