------- Comment #50 from rguenth at gcc dot gnu dot org 2010-01-04 23:11 ------- ira_traverse_loop_tree and its callee process_bb_node_for_cost accounts for nearly as much time as reload takes on hashes.c.
Btw, I find valgrind --tool=callgrind together with kcachegrind a better tool for benchmarking this kind of stuff - obviously on the smaller testcases. It has deterministic timing at least. -- http://gcc.gnu.org/bugzilla/show_bug.cgi?id=18687
