1  /* { dg-do run } */
       2  /* { dg-options "-O2 -msse2" } */
       3  /* { dg-require-effective-target sse2 } */
       4  
       5  #include "sse2-check.h"
       6  
       7  #include <mmintrin.h>
       8  
       9  #define N 4
      10  
      11  unsigned long long a[N], b[N], result[N];
      12  
      13  unsigned long long check[N] =
      14    { 0x101010101010100full,
      15      0x1010101010101010ull,
      16      0x1010101010101010ull,
      17      0x1010101010101010ull };
      18  
      19  __m64
      20  unsigned_add3 (const __m64 * a, const __m64 * b,
      21  	       __m64 * result, unsigned int count)
      22  {
      23    __m64 _a, _b, one, sum, carry, onesCarry;
      24  
      25    unsigned int i;
      26  
      27    carry = _mm_setzero_si64 ();
      28  
      29    one = _mm_cmpeq_pi8 (carry, carry);
      30    one = _mm_sub_si64 (carry, one);
      31  
      32    for (i = 0; i < count; i++)
      33      {
      34        _a = a[i];
      35        _b = b[i];
      36  
      37        sum = _mm_add_si64 (_a, _b);
      38        sum = _mm_add_si64 (sum, carry);
      39  
      40        result[i] = sum;
      41  
      42        onesCarry = _mm_and_si64 (_mm_xor_si64 (_a, _b), carry);
      43        onesCarry = _mm_or_si64 (_mm_and_si64 (_a, _b), onesCarry);
      44        onesCarry = _mm_and_si64 (onesCarry, one);
      45  
      46        _a = _mm_srli_si64 (_a, 1);
      47        _b = _mm_srli_si64 (_b, 1);
      48  
      49        carry = _mm_add_si64 (_mm_add_si64 (_a, _b), onesCarry);
      50        carry = _mm_srli_si64 (carry, 63);
      51      }
      52  
      53    return carry;
      54  }
      55  
      56  void __attribute__((noinline))
      57  sse2_test (void)
      58  {
      59    unsigned long long carry;
      60    int i;
      61  
      62    /* Really long numbers.  */
      63    a[3] = a[2] = a[1] = a[0] = 0xd3d3d3d3d3d3d3d3ull;
      64    b[3] = b[2] = b[1] = b[0] = 0x3c3c3c3c3c3c3c3cull;
      65  
      66    carry = (unsigned long long) unsigned_add3
      67      ((__m64 *)a, (__m64 *)b, (__m64 *)result, N);
      68  
      69    _mm_empty ();
      70  
      71    if (carry != 1)
      72      abort ();
      73  
      74    for (i = 0; i < N; i++)
      75      if (result [i] != check[i])
      76        abort ();
      77  }