1 /* { dg-require-effective-target vect_usad_char } */
2 
3 #include "tree-vect.h"
4 
5 typedef unsigned int uint32_t;
6 typedef unsigned short uint16_t;
7 typedef unsigned char uint8_t;
8 
9 extern int abs (int);
10 extern void abort (void);
11 
12 int __attribute__((noinline,noclone))
foo(uint8_t * pix1,uint8_t * pix2,int i_stride_pix2)13 foo (uint8_t *pix1, uint8_t *pix2, int i_stride_pix2)
14 {
15   int i_sum = 5;
16   for( int y = 0; y < 16; y++ )
17     {
18       i_sum += abs ( pix1[0] - pix2[0] );
19       i_sum += abs ( pix1[1] - pix2[1] );
20       i_sum += abs ( pix1[2] - pix2[2] );
21       i_sum += abs ( pix1[3] - pix2[3] );
22       i_sum += abs ( pix1[4] - pix2[4] );
23       i_sum += abs ( pix1[5] - pix2[5] );
24       i_sum += abs ( pix1[6] - pix2[6] );
25       i_sum += abs ( pix1[7] - pix2[7] );
26       i_sum += abs ( pix1[8] - pix2[8] );
27       i_sum += abs ( pix1[9] - pix2[9] );
28       i_sum += abs ( pix1[10] - pix2[10] );
29       i_sum += abs ( pix1[11] - pix2[11] );
30       i_sum += abs ( pix1[12] - pix2[12] );
31       i_sum += abs ( pix1[13] - pix2[13] );
32       i_sum += abs ( pix1[14] - pix2[14] );
33       i_sum += abs ( pix1[15] - pix2[15] );
34       pix1 += 16;
35       pix2 += i_stride_pix2;
36     }
37   return i_sum;
38 }
39 
40 int
main()41 main ()
42 {
43   check_vect ();
44 
45   uint8_t X[16*16];
46   uint8_t Y[16*16];
47 
48   for (int i = 0; i < 16*16; ++i)
49     {
50       X[i] = i;
51       Y[i] = 16*16 - i;
52       __asm__ volatile ("");
53     }
54 
55   if (foo (X, Y, 16) != 32512 + 5)
56     abort ();
57 
58   return 0;
59 }
60 
61 /* { dg-final { scan-tree-dump "vect_recog_sad_pattern: detected" "vect" } } */
62 /* { dg-final { scan-tree-dump "vectorizing stmts using SLP" "vect" } } */
63 /* { dg-final { scan-tree-dump-times "vectorized 1 loops" 1 "vect" } } */
64