gcc/testsuite/gcc.target/powerpc/ssse3-phaddd.c

   1 /* { dg-do run } */
   2 /* { dg-options "-O3 -mvsx -Wno-psabi" } */
   3 /* { dg-additional-options "-mdejagnu-cpu=power8" { target { ! has_arch_pwr8 } } } */
   4 /* { dg-require-effective-target p8vector_hw } */
   5
   6 #ifndef CHECK_H
   7 #define CHECK_H "ssse3-check.h"
   8 #endif
   9
  10 #ifndef TEST
  11 #define TEST ssse3_test
  12 #endif
  13
  14 #include CHECK_H
  15
  16 #include "ssse3-vals.h"
  17
  18 #include <tmmintrin.h>
  19
  20 #ifndef __AVX__
  21 /* Test the 64-bit form */
  22 static void
  23 ssse3_test_phaddd (__m64 *i1, __m64 *i2, __m64 *r)
  24 {
  25   *r = _mm_hadd_pi32 (*i1, *i2);
  26   _mm_empty();
  27 }
  28 #endif
  29
  30 /* Test the 128-bit form */
  31 static void
  32 ssse3_test_phaddd128 (__m128i *i1, __m128i *i2, __m128i *r)
  33 {
  34   /* Assumes incoming pointers are 16-byte aligned */
  35   *r = _mm_hadd_epi32 (*i1, *i2);
  36 }
  37
  38 /* Routine to manually compute the results */
  39 static void
  40 compute_correct_result(int *i1, int *i2, int *r)
  41 {
  42   int i;
  43
  44   for (i = 0; i < 2; i++)
  45     r[i] = i1[2 * i] + i1[2 * i + 1];
  46   for (i = 0; i < 2; i++)
  47     r[i + 2] = i2[2 * i] + i2[2 * i + 1];
  48 }
  49
  50 static void
  51 TEST (void)
  52 {
  53   int i;
  54   union data r __attribute__ ((aligned(16)));
  55   union data ck;
  56   int fail = 0;
  57
  58   for (i = 0; i < ARRAY_SIZE (vals) - 1; i++)
  59     {
  60       /* Manually compute the result */
  61       compute_correct_result (&vals[i + 0].w[0], &vals[i + 1].w[0], &ck.w[0]);
  62
  63 #ifndef __AVX__
  64       /* Run the 64-bit tests */
  65       ssse3_test_phaddd (&vals[i + 0].ll[0], &vals[i + 0].ll[1], &r.ll[0]);
  66       ssse3_test_phaddd (&vals[i + 1].ll[0], &vals[i + 1].ll[1], &r.ll[1]);
  67       fail += chk_128 (ck.m[0], r.m[0]);
  68 #endif
  69
  70       /* Run the 128-bit tests */
  71       ssse3_test_phaddd128 (&vals[i + 0].m[0], &vals[i + 1].m[0], &r.m[0]);
  72       fail += chk_128 (ck.m[0], r.m[0]);
  73     }
  74
  75   if (fail != 0)
  76     abort ();
  77 }