New localopt: parallelize non-volatile loads

[r28378]

ir/ir/iropt.c

@@ -3577,6 +3577,42 @@ static ir_node *transform_node_Proj_Load(ir_node *proj)
 					return new_r_Jmp(blk);
 				}
 			}
+		} else if (get_irn_mode(proj) == mode_M) {
+			/* non-volatile loads in sequence can be parallelized (then synced) */
+			ir_node *load = get_Proj_pred(proj);
+			if (get_Load_volatility(load) == volatility_non_volatile) {
+				ir_node *pred = get_Load_mem(load);
+				if (is_Proj(pred) && get_nodes_block(pred) == get_nodes_block(proj)) {
+					ir_node *loadpred = get_Proj_pred(pred);
+					if (is_Load(loadpred)) {
+						/* duplicate load */
+						ir_node *block = get_nodes_block(proj);
+						ir_graph *irg = get_irn_irg(proj);
+						ir_node *common_mem = get_Load_mem(loadpred);
+						ir_cons_flags flags = cons_none;
+						if (get_Load_align(load) == align_non_aligned)
+							flags &= cons_unaligned;
+						if (get_irn_pinned(load) == op_pin_state_floats)
+							flags &= cons_floats;
+						ir_node *newload = new_r_Load(block, common_mem, get_Load_ptr(load), get_Load_mode(load), flags);
+						ir_node *newprojres = new_r_Proj(newload, get_Load_mode(load), pn_Load_res);
+						ir_node *newprojmem = new_r_Proj(newload, mode_M, pn_Load_M);
+
+						/* create sync */
+						ir_node *ins[2] = {pred, newprojmem};
+						ir_node *sync = new_r_Sync(block, 2, ins);
+
+						/* prepare exchange */
+						turn_into_tuple(load, pn_Load_max);
+						set_Tuple_pred(load, pn_Load_M, sync);
+						set_Tuple_pred(load, pn_Load_res, newprojres);
+						set_Tuple_pred(load, pn_Load_X_regular, new_r_Bad(irg));
+						set_Tuple_pred(load, pn_Load_X_except, new_r_Bad(irg));
+
+						return sync;
+					}
+				}
+			}
 		}
 	}
 	return proj;