gcc/testsuite/gcc.target/i386/avx512bw-vpcmpb-2.c

   1 /* { dg-do run } */
   2 /* { dg-options "-O2 -mavx512bw" } */
   3 /* { dg-require-effective-target avx512bw } */
   4
   5 #define AVX512BW
   6 #include "avx512f-helper.h"
   7
   8 #include <math.h>
   9 #define SIZE (AVX512F_LEN / 8)
  10 #include "avx512f-mask-type.h"
  11
  12 #if AVX512F_LEN == 512
  13 #undef CMP
  14 #define CMP(imm, rel)                                   \
  15     dst_ref = 0;                                        \
  16     for (i = 0; i < SIZE; i++)                          \
  17       dst_ref = ((MASK_TYPE) (rel) << i) | dst_ref;     \
  18     source1.x = _mm512_loadu_si512 (s1);                \
  19     source2.x = _mm512_loadu_si512 (s2);                \
  20     dst1 = _mm512_cmp_epi8_mask (source1.x, source2.x, imm);\
  21     dst2 = _mm512_mask_cmp_epi8_mask (mask, source1.x, source2.x, imm);\
  22     if (dst_ref != dst1) abort();                       \
  23     if ((mask & dst_ref) != dst2) abort();
  24 #endif
  25
  26 #if AVX512F_LEN == 256
  27 #undef CMP
  28 #define CMP(imm, rel)                                   \
  29     dst_ref = 0;                                        \
  30     for (i = 0; i < SIZE; i++)                          \
  31       dst_ref = ((MASK_TYPE) (rel) << i) | dst_ref;     \
  32     source1.x = _mm256_loadu_si256 ((__m256i*)s1);      \
  33     source2.x = _mm256_loadu_si256 ((__m256i*)s2);      \
  34     dst1 = _mm256_cmp_epi8_mask (source1.x, source2.x, imm);\
  35     dst2 = _mm256_mask_cmp_epi8_mask (mask, source1.x, source2.x, imm);\
  36     if (dst_ref != dst1) abort();                       \
  37     if ((mask & dst_ref) != dst2) abort();
  38 #endif
  39
  40 #if AVX512F_LEN == 128
  41 #undef CMP
  42 #define CMP(imm, rel)                                   \
  43     dst_ref = 0;                                        \
  44     for (i = 0; i < SIZE; i++)                          \
  45       dst_ref = ((MASK_TYPE) (rel) << i) | dst_ref;     \
  46     source1.x = _mm_loadu_si128 ((__m128i*)s1);         \
  47     source2.x = _mm_loadu_si128 ((__m128i*)s2);         \
  48     dst1 = _mm_cmp_epi8_mask (source1.x, source2.x, imm);\
  49     dst2 = _mm_mask_cmp_epi8_mask (mask, source1.x, source2.x, imm);\
  50     if (dst_ref != dst1) abort();                       \
  51     if ((mask & dst_ref) != dst2) abort();
  52 #endif
  53
  54 void
  55 TEST ()
  56 {
  57     char s1[64] = {34, 78,  53, 64,
  58                   1, 57, 11, 231,
  59                   14, 45, 71, 75,
  60                   55, 66, 21, 73,
  61                   34, 68, 3, 56,
  62                   1, 57, 111, 241,
  63                   14,  15, 61, 75,
  64                   55,  16, 52, 3,
  65                   34, 78,  53, 64,
  66                   1, 57, 11, 231,
  67                   14, 45, 71, 75,
  68                   45, 66, 21, 73,
  69                   34, 68, 3, 56,
  70                   1, 57, 111, 241,
  71                   14,  15, 61, 75,
  72                   55,  16, 52, 3};
  73     char s2[64] = {4, 68, 86, 8,
  74                   1, 46,  1, 1,
  75                   45, 67, 36, 3,
  76                   4, 39, 56, 56,
  77                   124, 78, 53, 56,
  78                   1, 46,  1, 12,
  79                   45, 47, 36, 13,
  80                   4, 35, 56, 67,
  81                   4, 68, 86, 8,
  82                   1, 46,  1, 1,
  83                   45, 67, 36, 3,
  84                   4, 39, 56, 56,
  85                   124, 78, 53, 56,
  86                   1, 46,  1, 12,
  87                   45, 47, 36, 13,
  88                   4, 35, 56, 67};
  89     UNION_TYPE (AVX512F_LEN, i_b) source1, source2;
  90     MASK_TYPE dst1, dst2, dst_ref;
  91     MASK_TYPE mask = MASK_VALUE;
  92     int i;
  93
  94     CMP(0x00, s1[i] == s2[i]);
  95     CMP(0x01, s1[i] < s2[i]);
  96     CMP(0x02, s1[i] <= s2[i]);
  97     CMP(0x03, 0);
  98     CMP(0x04, s1[i] != s2[i]);
  99     CMP(0x05, s1[i] >= s2[i]);
 100     CMP(0x06, s1[i] > s2[i]);
 101     CMP(0x07, 1);
 102 }