gcc/testsuite/gcc.target/i386/sse4a-insert.c

   1 /* { dg-do run } */
   2 /* { dg-require-effective-target sse4a } */
   3 /* { dg-options "-O2 -msse4a" } */
   4
   5 #include "sse4a-check.h"
   6
   7 #include <ammintrin.h>
   8
   9 typedef union
  10 {
  11   long long i[2];
  12   __m128i vec;
  13 } LI;
  14
  15 static long long
  16 sse4a_test_insert (long long in1, long long in2)
  17 {
  18   __m128i v1,v2;
  19   long long index_length, pad;
  20   LI v_out;
  21   index_length = 0x0000000000000810LL;
  22   pad = 0x0;
  23   v1 = _mm_set_epi64x (pad, in1);
  24   v2 = _mm_set_epi64x (index_length, in2);
  25   v_out.vec = _mm_insert_si64 (v1, v2);
  26   return (v_out.i[0]);
  27 }
  28
  29 static long long
  30 sse4a_test_inserti (long long in1, long long in2)
  31 {
  32   __m128i v1,v2;
  33   long long pad = 0x0;
  34   LI v_out;
  35   v1 = _mm_set_epi64x (pad, in1);
  36   v2 = _mm_set_epi64x (pad, in2);
  37   v_out.vec = _mm_inserti_si64 (v1, v2, (unsigned int) 0x10, (unsigned int) 0x08);
  38   return (v_out.i[0]);
  39 }
  40
  41 static int chk (long long i1, long long i2)
  42 {
  43   int n_fails =0;
  44   if (i1 != i2)
  45     n_fails +=1;
  46   return n_fails;
  47 }
  48
  49 long long vals_in1[5] =
  50   {
  51     0x1234567887654321LL,
  52     0x1456782093002490LL,
  53     0x2340909123990390LL,
  54     0x9595959599595999LL,
  55     0x9099038798000029LL
  56   };
  57
  58 long long vals_in2[5] =
  59   {
  60     0x9ABCDEF00FEDCBA9LL,
  61     0x234567097289672ALL,
  62     0x45476453097BD342LL,
  63     0x23569012AE586FF0LL,
  64     0x432567ABCDEF765DLL
  65   };
  66
  67 long long vals_out[5] =
  68   {
  69     0x1234567887CBA921LL,
  70     0x1456782093672A90LL,
  71     0x2340909123D34290LL,
  72     0x95959595996FF099LL,
  73     0x9099038798765D29LL
  74   };
  75
  76 static void
  77 sse4a_test (void)
  78 {
  79   int i;
  80   int fail = 0;
  81   long long out;
  82
  83   for (i = 0; i < 5; i += 1)
  84     {
  85       out = sse4a_test_insert (vals_in1[i], vals_in2[i]);
  86       fail += chk(out, vals_out[i]);
  87
  88       out = sse4a_test_inserti (vals_in1[i], vals_in2[i]);
  89       fail += chk(out, vals_out[i]);
  90     }
  91
  92   if (fail != 0)
  93     abort ();
  94 }