vect-reduc-dot-u16b.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 DOT2 43680
   9
  10 unsigned short X[N] __attribute__ ((__aligned__(__BIGGEST_ALIGNMENT__)));
  11 unsigned short Y[N] __attribute__ ((__aligned__(__BIGGEST_ALIGNMENT__)));
  12
  13 /* short->int->int dot product.
  14    Currently not detected as a dot-product pattern: the multiplication
  15    promotes the ushorts to int, and then the product is promoted to unsigned
  16    int for the addition.  Which results in an int->unsigned int cast, which
  17    since no bits are modified in the cast should be trivially vectorizable.  */
  18 __attribute__ ((noinline)) unsigned int
  19 foo2(int len) {
  20   int i;
  21   unsigned int result = 0;
  22
  23   for (i=0; i<len; i++) {
  24     result += (X[i] * Y[i]);
  25   }
  26   return result;
  27 }
  28
  29
  30 int main (void)
  31 {
  32   unsigned int  dot2;
  33   int i;
  34
  35   check_vect ();
  36
  37   for (i=0; i<N; i++) {
  38     X[i] = i;
  39     Y[i] = 64-i;
  40     __asm__ volatile ("");
  41   }
  42
  43   dot2 = foo2 (N);
  44   if (dot2 != DOT2)
  45     abort ();
  46
  47   return 0;
  48 }
  49
  50 /* { dg-final { scan-tree-dump-times "vect_recog_dot_prod_pattern: detected" 1 "vect" { xfail *-*-* } } } */
  51
  52 /* Once the dot-product pattern is detected, we expect
  53    that loop to be vectorized on vect_udot_hi targets (targets that support
  54    dot-product of unsigned shorts) and targets that support widening multiplication.  */
  55 /* The induction loop in main is vectorized.  */
  56 /* { dg-final { scan-tree-dump-times "vectorized 1 loops" 2 "vect" { xfail *-*-* } } } */
  57 /* { dg-final { scan-tree-dump-times "vectorized 1 loops" 1 "vect" { target vect_pack_trunc } } } */
  58
  59 /* { dg-final { cleanup-tree-dump "vect" } } */