gcc/testsuite/gcc.target/i386/ssse3-phaddsw.c

   1 /* { dg-do run } */
   2 /* { dg-require-effective-target ssse3 } */
   3 /* { dg-options "-O2 -fno-strict-aliasing -mssse3" } */
   4
   5 #ifndef CHECK_H
   6 #define CHECK_H "ssse3-check.h"
   7 #endif
   8
   9 #ifndef TEST
  10 #define TEST ssse3_test
  11 #endif
  12
  13 #include CHECK_H
  14
  15 #include "ssse3-vals.h"
  16
  17 #include <tmmintrin.h>
  18
  19 #ifndef __AVX__
  20 /* Test the 64-bit form */
  21 static void
  22 ssse3_test_phaddsw (int *i1, int *i2, int *r)
  23 {
  24   __m64 t1 = *(__m64 *) i1;
  25   __m64 t2 = *(__m64 *) i2;
  26   *(__m64 *) r = _mm_hadds_pi16 (t1, t2);
  27   _mm_empty ();
  28 }
  29 #endif
  30
  31 /* Test the 128-bit form */
  32 static void
  33 ssse3_test_phaddsw128 (int *i1, int *i2, int *r)
  34 {
  35   /* Assumes incoming pointers are 16-byte aligned */
  36   __m128i t1 = *(__m128i *) i1;
  37   __m128i t2 = *(__m128i *) i2;
  38  *(__m128i *) r = _mm_hadds_epi16 (t1, t2);
  39 }
  40
  41 static short
  42 signed_saturate_to_word (int x)
  43 {
  44   if (x > (int) 0x7fff)
  45     return 0x7fff;
  46
  47   if (x < (int) 0xffff8000)
  48     return 0x8000;
  49
  50   return (short) x;
  51 }
  52
  53 /* Routine to manually compute the results */
  54 static void
  55 compute_correct_result (int *i1, int *i2, int *r)
  56 {
  57   short *s1 = (short *) i1;
  58   short *s2 = (short *) i2;
  59   short *sout = (short *) r;
  60   int i;
  61
  62   for (i = 0; i < 4; i++)
  63     sout[i] = signed_saturate_to_word(s1[2 * i] + s1[2 * i + 1]);
  64   for (i = 0; i < 4; i++)
  65     sout[i + 4] = signed_saturate_to_word(s2[2 * i] + s2[2 * i + 1]);
  66 }
  67
  68 static void
  69 TEST (void)
  70 {
  71   int i;
  72   int r [4] __attribute__ ((aligned(16)));
  73   int ck [4];
  74   int fail = 0;
  75
  76   for (i = 0; i < 256; i += 8)
  77     {
  78       /* Manually compute the result */
  79       compute_correct_result (&vals[i + 0], &vals[i + 4], ck);
  80
  81 #ifndef __AVX__
  82       /* Run the 64-bit tests */
  83       ssse3_test_phaddsw (&vals[i + 0], &vals[i + 2], &r[0]);
  84       ssse3_test_phaddsw (&vals[i + 4], &vals[i + 6], &r[2]);
  85       fail += chk_128 (ck, r);
  86 #endif
  87
  88       /* Run the 128-bit tests */
  89       ssse3_test_phaddsw128 (&vals[i + 0], &vals[i + 4], r);
  90       fail += chk_128 (ck, r);
  91     }
  92
  93   if (fail != 0)
  94     abort ();
  95 }