http://gcc.gnu.org/bugzilla/show_bug.cgi?id=45375
--- Comment #92 from Jan Hubicka <hubicka at gcc dot gnu.org> 2011-05-19 22:28:18 UTC --- decl in is now at 96 seconds. oprofile for streaming in is: 27469 9.3054 lto1 htab_find_slot_with_hash 23175 7.8508 libc-2.11.1.so _int_malloc 18044 6.1126 lto1 lto_input_tree 14823 5.0215 libc-2.11.1.so memset 14108 4.7792 lto1 gt_ggc_mx_lang_tree_node 13511 4.5770 lto1 inflate_fast 11805 3.9991 lto1 gimple_type_eq 11247 3.8100 lto1 lto_input_uleb128 11227 3.8033 lto1 ggc_set_mark 10903 3.6935 lto1 pointer_map_insert So obviously still some place for improvements for merging. I think malloc calls come mostly from SCC detection code (we create a lot of temporary obstacks and pointer maps). lto_input_tree can probably handle quite a lot of optimizations reducing amount of data we stream. Plus we don't really need to stream ulebs for everything. For whole WPA we now need about 5 minutes, the oprofile is: 152067 14.9073 lto1 decl_assembler_name_equal 48258 4.7308 lto1 htab_find_slot_with_hash 46730 4.5810 lto1 edge_badness 37954 3.7207 libc-2.11.1.so _int_malloc 36692 3.5970 lto1 pointer_map_insert 30387 2.9789 lto1 do_estimate_growth 28496 2.7935 lto1 lto_input_tree 20992 2.0579 lto1 inflate_fast 20765 2.0356 libc-2.11.1.so memset 20264 1.9865 lto1 varpool_node_for_asm 19784 1.9394 lto1 lto_output_tree 19121 1.8745 lto1 htab_hash_string 19053 1.8678 lto1 lto_input_uleb128 good news is that decl_assembler_name_equal is stupid handling of varpool aliases in varpool_node_for_asm that will go away with my alias rewrite. edge_badness is easy to track down, too, it is just inliner updating paranoia. Honza
