1  /* { dg-do compile } */
       2  /* { dg-require-effective-target size32plus } */
       3  
       4  typedef unsigned long long uint64_t;
       5  
       6  #define n 4096
       7  double A[n][n] __attribute__((aligned(16)));
       8  double B[n][n] __attribute__((aligned(16)));
       9  double C[n][n] __attribute__((aligned(16)));
      10  
      11  #define tilesize 128
      12  
      13  typedef double adouble __attribute__((__aligned__(16)));
      14  
      15  void foo ()
      16  {
      17    int ih, jh, kh, il, kl, jl;
      18    for (ih = 0; ih < n; ih += tilesize) 
      19      for (jh = 0; jh < n; jh += tilesize)                 
      20        for (kh = 0; kh < n; kh += tilesize)                      
      21  	for (il = 0; il < tilesize; ++il)
      22  	  {
      23  	    adouble *Ap = (adouble *)&A[ih+il][kh];
      24  	    for (kl = 0; kl < tilesize; ++kl)
      25  	      for (jl = 0; jl < tilesize; ++jl)
      26  		C[ih+il][jh+jl] += Ap[kl] * B[kh+kl][jh+jl];
      27  	  }
      28  }