gcc/testsuite/gcc.dg/vect/bb-slp-29.c

   1 /* { dg-require-effective-target vect_int } */
   2
   3 #include <stdarg.h>
   4 #include "tree-vect.h"
   5
   6 #define A 3
   7 #define B 4
   8 #define N 256
   9
  10 short src[N], dst[N];
  11
  12 void foo (short * __restrict__ dst, short * __restrict__ src, int h,
  13           int stride)
  14 {
  15   int i;
  16   h /= 16;
  17   for (i = 0; i < h; i++)
  18     {
  19       dst[0] = A*src[0] + B*src[1];
  20       dst[1] = A*src[1] + B*src[2];
  21       dst[2] = A*src[2] + B*src[3];
  22       dst[3] = A*src[3] + B*src[4];
  23       dst[4] = A*src[4] + B*src[5];
  24       dst[5] = A*src[5] + B*src[6];
  25       dst[6] = A*src[6] + B*src[7];
  26       dst[7] = A*src[7] + B*src[8];
  27       dst += stride;
  28       src += stride;
  29       asm volatile ("" ::: "memory");
  30    }
  31 }
  32
  33
  34 int main (void)
  35 {
  36   int i;
  37
  38   check_vect ();
  39
  40   for (i = 0; i < N; i++)
  41     {
  42        dst[i] = 0;
  43        src[i] = i;
  44     }
  45
  46   foo (dst, src, N, 8);
  47
  48 #pragma GCC novector
  49   for (i = 0; i < N/2; i++)
  50     {
  51       if (dst[i] != A * src[i] + B * src[i+1])
  52         abort ();
  53     }
  54
  55   return 0;
  56 }
  57
  58 /* Exclude POWER8 (only POWER cpu for which vect_element_align is true)
  59    because loops have vectorized before SLP gets a shot.  */
  60 /* { dg-final { scan-tree-dump-times "optimized: basic block" 1 "slp1"  { target { { vect_int_mult && vect_element_align } && { ! powerpc*-*-* } } } } } */
  61