sse4_2-pcmpestrm-1.c   [plain text]

/* APPLE LOCAL file 5612787 mainline sse4 */
/* { dg-do run { target i?86-*-* x86_64-*-* } } */
/* { dg-require-effective-target sse4 } */
/* { dg-options "-O2 -msse4.2" } */

#include "sse4_2-check.h"
#include "sse4_2-pcmpstr.h"

#define NUM 1024

#define IMM_VAL0 \
  (SIDD_SBYTE_OPS | SIDD_CMP_RANGES | SIDD_MASKED_POSITIVE_POLARITY)
#define IMM_VAL1 \
  (SIDD_UBYTE_OPS | SIDD_CMP_EQUAL_EACH | SIDD_NEGATIVE_POLARITY \
   | SIDD_BIT_MASK)
#define IMM_VAL2 \
  (SIDD_UWORD_OPS | SIDD_CMP_EQUAL_ANY | SIDD_MASKED_NEGATIVE_POLARITY)
#define IMM_VAL3 \
  (SIDD_SWORD_OPS | SIDD_CMP_EQUAL_ORDERED \
   | SIDD_MASKED_NEGATIVE_POLARITY | SIDD_UNIT_MASK)


static void
sse4_2_test (void)
{
  union
    {
      __m128i x[NUM];
      char c[NUM *16];
    } src1, src2;
  __m128i res, correct;
  int l1, l2;
  int i;

  for (i = 0; i < NUM *16; i++)
    {
      src1.c[i] = rand ();
      src2.c[i] = rand ();
    }

  for (i = 0; i < NUM; i++)
    {
      l1 = rand () % 18;
      l2 = rand () % 18;

      switch((rand() % 4))
	{
	case 0:
	  res  = _mm_cmpestrm (src1.x[i], l1, src2.x[i], l2, IMM_VAL0);
	  correct = cmp_em (&src1.x[i], l1, &src2.x[i], l2, IMM_VAL0,
			    NULL);
	  break;

	case 1:
	  res  = _mm_cmpestrm (src1.x[i], l1, src2.x[i], l2, IMM_VAL1);
	  correct = cmp_em (&src1.x[i], l1, &src2.x[i], l2, IMM_VAL1,
			    NULL);
	  break;

	case 2:
	  res  = _mm_cmpestrm (src1.x[i], l1, src2.x[i], l2, IMM_VAL2);
	  correct = cmp_em (&src1.x[i], l1, &src2.x[i], l2, IMM_VAL2,
			    NULL);
	  break;

	default:
	  res  = _mm_cmpestrm (src1.x[i], l1, src2.x[i], l2, IMM_VAL3);
	  correct = cmp_em (&src1.x[i], l1, &src2.x[i], l2, IMM_VAL3,
			    NULL);
	  break;
        }

      if (memcmp (&correct, &res, sizeof (res)))
	abort ();
    }
}