gcc/testsuite/gcc.target/powerpc/ssse3-phsubsw.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_phsubsw (__m64 *i1, __m64 *i2, __m64 *r)
  24 {
  25   *(__m64 *) r = _mm_hsubs_pi16 (*i1, *i2);
  26   _mm_empty ();
  27 }
  28 #endif
  29
  30 /* Test the 128-bit form */
  31 static void
  32 ssse3_test_phsubsw128 (__m128i *i1, __m128i *i2, __m128i *r)
  33 {
  34   /* Assumes incoming pointers are 16-byte aligned */
  35   *r = _mm_hsubs_epi16 (*i1, *i2);
  36 }
  37
  38 static short
  39 signed_saturate_to_word (int x)
  40 {
  41   if (x > (int )0x7fff)
  42     return 0x7fff;
  43
  44   if (x < (int) 0xffff8000)
  45     return 0x8000;
  46
  47   return (short)x;
  48 }
  49
  50 /* Routine to manually compute the results */
  51 static void
  52 compute_correct_result (short *i1, short *i2, short *r)
  53 {
  54   int i;
  55
  56   for (i = 0; i < 4; i++)
  57     r[i] = signed_saturate_to_word (i1[2 * i] - i1[2 * i + 1]);
  58
  59   for (i = 0; i < 4; i++)
  60     r[i + 4] = signed_saturate_to_word (i2[2 * i] - i2[2 * i + 1]);
  61 }
  62
  63 static void
  64 TEST (void)
  65 {
  66   int i;
  67   union data r __attribute__ ((aligned(16)));
  68   union data ck;
  69   int fail = 0;
  70
  71   for (i = 0; i < ARRAY_SIZE (vals) - 1; i++)
  72     {
  73       /* Manually compute the result */
  74       compute_correct_result (&vals[i + 0].h[0], &vals[i + 1].h[0], &ck.h[0]);
  75
  76 #ifndef __AVX__
  77       /* Run the 64-bit tests */
  78       ssse3_test_phsubsw (&vals[i + 0].ll[0], &vals[i + 0].ll[1], &r.ll[0]);
  79       ssse3_test_phsubsw (&vals[i + 1].ll[0], &vals[i + 1].ll[1], &r.ll[1]);
  80       fail += chk_128 (ck.m[0], r.m[0]);
  81 #endif
  82
  83       /* Run the 128-bit tests */
  84       ssse3_test_phsubsw128 (&vals[i + 0].m[0], &vals[i + 1].m[0], &r.m[0]);
  85       fail += chk_128 (ck.m[0], r.m[0]);
  86     }
  87
  88   if (fail != 0)
  89     abort ();
  90 }