diff -N -r -c HLib-1.3orig/Library/laplacebem.c HLib-1.3/Library/laplacebem.c
*** HLib-1.3orig/Library/laplacebem.c	2004-12-12 17:42:33.000000000 +0100
--- HLib-1.3/Library/laplacebem.c	2006-02-27 10:25:59.643972448 +0100
***************
*** 43,49 ****
--- 43,53 ----
  }
  
  #ifdef HAVE_SSE2
+ #if __GNUC__ >= 4
+ typedef double v2df __attribute__((vector_size(16)));
+ #else
  typedef double v2df __attribute__((mode(V2DF)));
+ #endif
  
  static const v2df sse_slp_factor = { 1.0 / 4.0 / M_PI, 1.0 / 4.0 / M_PI};
  static const v2df sse_one = { 1.0, 1.0 };
***************
*** 53,61 ****
  static v2df
  sse_loadsingle(double x)
  {
!   v2df xx;
!   xx = __builtin_ia32_loadsd(&x);
!   return __builtin_ia32_shufpd(xx, xx, 0);
  }
  
  static v2df
--- 57,73 ----
  static v2df
  sse_loadsingle(double x)
  {
!   double xx[2];
!   xx[0] = xx[1] = x;
!   return __builtin_ia32_loadupd(xx);
! }
! 
! static void
! sse_storesingle(double *buf, v2df x)
! {
!   double xx[2];
!   __builtin_ia32_storeupd(xx, x);
!   buf[0] = xx[0];
  }
  
  static v2df
***************
*** 465,473 ****
  	yy[1] = b0 * xj[0][1] + b1 * xj[1][1] + b2 * xj[2][1];
  	yy[2] = b0 * xj[0][2] + b1 * xj[1][2] + b2 * xj[2][2];
  
! 	__builtin_ia32_storesd(qbuf+q, 
! 			       sse_slp_kernel(xx[0], xx[1], xx[2],
! 					      yy[0], yy[1], yy[2]));
  	q++;
        }
  
--- 477,485 ----
  	yy[1] = b0 * xj[0][1] + b1 * xj[1][1] + b2 * xj[2][1];
  	yy[2] = b0 * xj[0][2] + b1 * xj[1][2] + b2 * xj[2][2];
  
! 	sse_storesingle(qbuf+q, 
! 			sse_slp_kernel(xx[0], xx[1], xx[2],
! 				       yy[0], yy[1], yy[2]));
  	q++;
        }
  
***************
*** 895,904 ****
  	yy[1] = b0 * xj[0][1] + b1 * xj[1][1] + b2 * xj[2][1];
  	yy[2] = b0 * xj[0][2] + b1 * xj[1][2] + b2 * xj[2][2];
  
! 	__builtin_ia32_storesd(qbuf+q, 
! 			       sse_dlp_kernel(xx[0], xx[1], xx[2],
! 					      yy[0], yy[1], yy[2],
! 					      nn[0], nn[1], nn[2]));
  	q++;
        }
  
--- 907,916 ----
  	yy[1] = b0 * xj[0][1] + b1 * xj[1][1] + b2 * xj[2][1];
  	yy[2] = b0 * xj[0][2] + b1 * xj[1][2] + b2 * xj[2][2];
  
! 	sse_storesingle(qbuf+q, 
! 			sse_dlp_kernel(xx[0], xx[1], xx[2],
! 				       yy[0], yy[1], yy[2],
! 				       nn[0], nn[1], nn[2]));
  	q++;
        }
  
***************
*** 1910,1916 ****
  	  xs[2] = sse_loadsingle(x2[2][0]);
  	  kslp = sse_slp_kernel(xs[0], xs[1], xs[2],
  				ys[0], ys[1], ys[2]);
! 	  __builtin_ia32_storesd(S+(i-1)+j*ldS, kslp);
  	}
  
  	j++;
--- 1922,1928 ----
  	  xs[2] = sse_loadsingle(x2[2][0]);
  	  kslp = sse_slp_kernel(xs[0], xs[1], xs[2],
  				ys[0], ys[1], ys[2]);
! 	  sse_storesingle(S+(i-1)+j*ldS, kslp);
  	}
  
  	j++;