gcc/testsuite/gcc.dg/vect/costmodel/x86_64/costmodel-pr69297.c

   1 /* { dg-do compile } */
   2 /* { dg-additional-options "-march=core-avx2 -fdump-tree-slp-details" } */
   3
   4 #define abs(x) (x) < 0 ? -(x) : (x)
   5 int
   6 foo (int* diff)
   7 {
   8   int k, satd = 0, m[16], d[16];
   9
  10     m[ 0] = diff[ 0] + diff[12];
  11     m[ 4] = diff[ 4] + diff[ 8];
  12     m[ 8] = diff[ 4] - diff[ 8];
  13     m[12] = diff[ 0] - diff[12];
  14     m[ 1] = diff[ 1] + diff[13];
  15     m[ 5] = diff[ 5] + diff[ 9];
  16     m[ 9] = diff[ 5] - diff[ 9];
  17     m[13] = diff[ 1] - diff[13];
  18     m[ 2] = diff[ 2] + diff[14];
  19     m[ 6] = diff[ 6] + diff[10];
  20     m[10] = diff[ 6] - diff[10];
  21     m[14] = diff[ 2] - diff[14];
  22     m[ 3] = diff[ 3] + diff[15];
  23     m[ 7] = diff[ 7] + diff[11];
  24     m[11] = diff[ 7] - diff[11];
  25     m[15] = diff[ 3] - diff[15];
  26
  27     d[ 0] = m[ 0] + m[ 4];
  28     d[ 8] = m[ 0] - m[ 4];
  29     d[ 4] = m[ 8] + m[12];
  30     d[12] = m[12] - m[ 8];
  31     d[ 1] = m[ 1] + m[ 5];
  32     d[ 9] = m[ 1] - m[ 5];
  33     d[ 5] = m[ 9] + m[13];
  34     d[13] = m[13] - m[ 9];
  35     d[ 2] = m[ 2] + m[ 6];
  36     d[10] = m[ 2] - m[ 6];
  37     d[ 6] = m[10] + m[14];
  38     d[14] = m[14] - m[10];
  39     d[ 3] = m[ 3] + m[ 7];
  40     d[11] = m[ 3] - m[ 7];
  41     d[ 7] = m[11] + m[15];
  42     d[15] = m[15] - m[11];
  43
  44     m[ 0] = d[ 0] + d[ 3];
  45     m[ 1] = d[ 1] + d[ 2];
  46     m[ 2] = d[ 1] - d[ 2];
  47     m[ 3] = d[ 0] - d[ 3];
  48     m[ 4] = d[ 4] + d[ 7];
  49     m[ 5] = d[ 5] + d[ 6];
  50     m[ 6] = d[ 5] - d[ 6];
  51     m[ 7] = d[ 4] - d[ 7];
  52     m[ 8] = d[ 8] + d[11];
  53     m[ 9] = d[ 9] + d[10];
  54     m[10] = d[ 9] - d[10];
  55     m[11] = d[ 8] - d[11];
  56     m[12] = d[12] + d[15];
  57     m[13] = d[13] + d[14];
  58     m[14] = d[13] - d[14];
  59     m[15] = d[12] - d[15];
  60
  61     d[ 0] = m[ 0] + m[ 1];
  62     d[ 1] = m[ 0] - m[ 1];
  63     d[ 2] = m[ 2] + m[ 3];
  64     d[ 3] = m[ 3] - m[ 2];
  65     d[ 4] = m[ 4] + m[ 5];
  66     d[ 5] = m[ 4] - m[ 5];
  67     d[ 6] = m[ 6] + m[ 7];
  68     d[ 7] = m[ 7] - m[ 6];
  69     d[ 8] = m[ 8] + m[ 9];
  70     d[ 9] = m[ 8] - m[ 9];
  71     d[10] = m[10] + m[11];
  72     d[11] = m[11] - m[10];
  73     d[12] = m[12] + m[13];
  74     d[13] = m[12] - m[13];
  75     d[14] = m[14] + m[15];
  76     d[15] = m[15] - m[14];
  77     for (k=0; k<16; k++)
  78       satd += abs(d[k]);
  79   return satd;
  80 }
  81
  82 /* { dg-final { scan-tree-dump "vectorization is not profitable" "slp1" } } */
  83 /* { dg-final { scan-tree-dump-not "basic block vectorized" "slp1" } } */