gcc/testsuite/gcc.dg/torture/pr54894.c

   1 /* { dg-do compile } */
   2
   3 typedef unsigned long long uint64_t;
   4
   5 #define n 4096
   6 double A[n][n] __attribute__((aligned(16)));
   7 double B[n][n] __attribute__((aligned(16)));
   8 double C[n][n] __attribute__((aligned(16)));
   9
  10 #define tilesize 128
  11
  12 typedef double adouble __attribute__((__aligned__(16)));
  13
  14 void foo ()
  15 {
  16   int ih, jh, kh, il, kl, jl;
  17   for (ih = 0; ih < n; ih += tilesize)
  18     for (jh = 0; jh < n; jh += tilesize)
  19       for (kh = 0; kh < n; kh += tilesize)
  20         for (il = 0; il < tilesize; ++il)
  21           {
  22             adouble *Ap = (adouble *)&A[ih+il][kh];
  23             for (kl = 0; kl < tilesize; ++kl)
  24               for (jl = 0; jl < tilesize; ++jl)
  25                 C[ih+il][jh+jl] += Ap[kl] * B[kh+kl][jh+jl];
  26           }
  27 }