slp-reduc-3.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 DOT1 21834
   9 #define DOT2 21876
  10
  11 unsigned short X[N] __attribute__ ((__aligned__(__BIGGEST_ALIGNMENT__)));
  12 unsigned short Y[N] __attribute__ ((__aligned__(__BIGGEST_ALIGNMENT__)));
  13
  14 /* short->short->int dot product.
  15    Not detected as a dot-product pattern.
  16    Requires support for non-widneing multiplication and widening-summation.
  17    Vectorized with loop-aware SLP. */
  18 __attribute__ ((noinline)) unsigned int
  19 foo1(int len, int *result1, int *result2)
  20 {
  21   int i;
  22   unsigned int res1 = 10, res2 = 20;
  23   unsigned short prod;
  24
  25   for (i=0; i<len; i++) {
  26     prod = X[2*i] * Y[2*i];
  27     res1 += prod;
  28     prod = X[2*i+1] * Y[2*i+1];
  29     res2 += prod;
  30   }
  31
  32   *result1 = res1;
  33   *result2 = res2;
  34
  35   return 0;
  36 }
  37
  38 int main (void)
  39 {
  40   unsigned int dot1, dot2;
  41   unsigned short i;
  42
  43   check_vect ();
  44
  45   for (i=0; i<N; i++) {
  46     X[i] = i;
  47     Y[i] = 64-i;
  48   }
  49
  50   foo1 (N/2, &dot1, &dot2);
  51
  52   if (dot1 != DOT1 || dot2 != DOT2)
  53     abort ();
  54
  55   return 0;
  56 }
  57
  58 /* The initialization loop in main also gets vectorized.  */
  59 /* { dg-final { scan-tree-dump-times "vect_recog_dot_prod_pattern: detected" 1 "vect" { xfail *-*-* } } } */
  60 /* { dg-final { scan-tree-dump-times "vectorized 1 loops" 2 "vect" { target { vect_short_mult && { vect_widen_sum_hi_to_si  && vect_unpack } } } } } */
  61 /* { dg-final { scan-tree-dump-times "vectorizing stmts using SLP" 1 "vect" { xfail { vect_widen_sum_hi_to_si_pattern ||  { ! vect_unpack } } } } } */
  62 /* { dg-final { cleanup-tree-dump "vect" } } */