gcc/testsuite/gcc.target/i386/sse4_1-dpps-2.c

   1 /* { dg-do run { target i?86-*-* x86_64-*-* } } */
   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 lmskN  0x00
  11 #define lmsk0  0x01
  12 #define lmsk1  0x02
  13 #define lmsk2  0x04
  14 #define lmsk3  0x08
  15 #define lmsk01 0x03
  16 #define lmsk02 0x05
  17 #define lmsk03 0x09
  18 #define lmsk12 0x06
  19 #define lmsk13 0x0A
  20 #define lmsk23 0x0C
  21 #define lmskA  0x0F
  22
  23 #define hmskN  0x00
  24 #define hmskA  0xF0
  25 #define hmsk0  0x10
  26 #define hmsk1  0x20
  27 #define hmsk2  0x40
  28 #define hmsk3  0x80
  29 #define hmsk01 0x30
  30 #define hmsk02 0x50
  31 #define hmsk03 0x90
  32 #define hmsk12 0x60
  33 #define hmsk13 0xA0
  34 #define hmsk23 0xC0
  35
  36 #ifndef HIMASK
  37 #define HIMASK hmskA
  38 #endif
  39
  40 #ifndef LOMASK
  41 #define LOMASK lmskA
  42 #endif
  43
  44 static void
  45 sse4_1_test (void)
  46 {
  47   union
  48     {
  49       __m128 x;
  50       float f[4];
  51     } val1[16], val2[16], res[16], chk[16];
  52   int i,j;
  53   float tmp;
  54
  55   for (i = 0; i < 16; i++)
  56     {
  57       val1[i].f[0] = 2.;
  58       val1[i].f[1] = 3.;
  59       val1[i].f[2] = 4.;
  60       val1[i].f[3] = 5.;
  61
  62       val2[i].f[0] = 10.;
  63       val2[i].f[1] = 100.;
  64       val2[i].f[2] = 1000.;
  65       val2[i].f[3] = 10000.;
  66
  67       tmp = 0.;
  68       for (j = 0; j < 4; j++)
  69         if ((HIMASK & (0x10 << j)))
  70           tmp += val1[i].f [j] * val2[i].f [j];
  71
  72       for (j = 0; j < 4; j++)
  73         if ((LOMASK & (1 << j)))
  74           chk[i].f[j] = tmp;
  75     }
  76
  77    for (i = 0; i < 16; i++)
  78      {
  79        res[i].x = _mm_dp_ps (val1[i].x, val2[i].x, HIMASK | LOMASK);
  80        if (memcmp (&res[i], &chk[i], sizeof (chk[i])))
  81          abort ();
  82      }
  83 }