gcc/testsuite/gcc.dg/vect/vect-reduc-dot-u8b.c

   1 /* { dg-require-effective-target vect_int } */
   2
   3 #include <stdarg.h>
   4 #include "tree-vect.h"
   5
   6 #define N 64
   7
   8 #define DOT 43680
   9
  10 unsigned char X[N] __attribute__ ((__aligned__(16)));
  11 unsigned char Y[N] __attribute__ ((__aligned__(16)));
  12
  13 /* char->short->short dot product.
  14    Detected as a dot-product pattern.
  15    Should be vectorized on targets that support dot-product for unsigned chars,
  16    but currently this test cannot be vectorized as a dot-product on targets
  17    that support char->short->int dot-product.
  18    Alternatively, this test can be vectorized using vect_widen_mult_qi (or
  19    vect_unpack and non-widening multplication: vect_unpack && vect_short_mult).
  20    */
  21 __attribute__ ((noinline)) unsigned short
  22 foo (int len) {
  23   int i;
  24   unsigned short result = 0;
  25
  26   for (i=0; i<len; i++) {
  27     result += (unsigned short)(X[i] * Y[i]);
  28   }
  29   return result;
  30 }
  31
  32 int main (void)
  33 {
  34   unsigned short dot;
  35   int i;
  36
  37   check_vect ();
  38
  39   for (i=0; i<N; i++) {
  40     X[i] = i;
  41     Y[i] = 64-i;
  42   }
  43
  44   dot = foo (N);
  45   if (dot != DOT)
  46     abort ();
  47
  48   return 0;
  49 }
  50
  51 /* { dg-final { scan-tree-dump-times "vect_recog_dot_prod_pattern: detected" 1 "vect" } } */
  52
  53 /* When the vectorizer is enhanced to vectorize accumulation into short for
  54    targets that support accumulation into int (powerpc, ia64) we'd have:
  55 dg-final { scan-tree-dump-times "vectorized 1 loops" 1 "vect" { target vect_udot_qi || vect_widen_mult_qi_to_hi } }
  56 */
  57 /* { dg-final { scan-tree-dump-times "vectorized 1 loops" 1 "vect" {target vect_widen_mult_qi_to_hi} } } */
  58
  59 /* { dg-final { cleanup-tree-dump "vect" } } */
  60