vect-reduc-dot-s8b.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 -21856
   9
  10 signed char X[N] __attribute__ ((__aligned__(__BIGGEST_ALIGNMENT__)));
  11 signed char Y[N] __attribute__ ((__aligned__(__BIGGEST_ALIGNMENT__)));
  12
  13 /* char->short->short dot product.
  14    The dot-product pattern should be detected.
  15    The reduction is currently not vectorized becaus of the signed->unsigned->signed
  16    casts, since this patch:
  17
  18      2005-12-26  Kazu Hirata  <kazu@codesourcery.com>
  19
  20         PR tree-optimization/25125
  21
  22    When the dot-product is detected, the loop should be vectorized on vect_sdot_qi
  23    targets (targets that support dot-product of signed char).
  24    This test would currently fail to vectorize on targets that support
  25    dot-product of chars into an int accumulator.
  26    Alternatively, the loop could also be vectorized as widening-mult + summation,
  27    or with type-conversion support.
  28  */
  29 __attribute__ ((noinline)) short
  30 foo2(int len) {
  31   int i;
  32   short result = 0;
  33
  34   for (i=0; i<len; i++) {
  35     result += (X[i] * Y[i]);
  36   }
  37   return result;
  38 }
  39
  40 int main (void)
  41 {
  42   int i;
  43   short dot2;
  44
  45   check_vect ();
  46
  47   for (i=0; i<N; i++) {
  48     X[i] = i;
  49     Y[i] = 64-i;
  50     __asm__ volatile ("");
  51   }
  52
  53   dot2 = foo2 (N);
  54   if (dot2 != DOT2)
  55     abort ();
  56
  57   return 0;
  58 }
  59
  60 /* { dg-final { scan-tree-dump-times "vect_recog_dot_prod_pattern: detected" 1 "vect" { xfail *-*-* } } } */
  61 /* { dg-final { scan-tree-dump-times "vect_recog_widen_mult_pattern: detected" 1 "vect" } } */
  62
  63 /* { dg-final { scan-tree-dump-times "vectorized 1 loops" 1 "vect" { xfail *-*-* } } } */
  64
  65 /* { dg-final { cleanup-tree-dump "vect" } } */