Hi, On Mon, 22 Aug 2011, Richard Guenther wrote:
> > Some functions are extremely large though. Do you mean that MEM > > itself would be enlarged to have the MEM_ATTRS field so that one > > operand is the address, then expr, then HWI size, offset, etc.? > > Because if the mem attrs aren't shared any longer, it doesn't make > > sense to keep the indirection. I still fear we have way too many MEMs > > in RTL that this would be noticeable. > > It would be interesting to have numbers about the amount of sharing that > happens A pathetic amount. From compiling cse.c combine.c tree.c and dwarf2out.c the top 10 users of MEMs per routine are: With -O0: MEMs ATTR name 970 485 combine_simplify_rtx 1011 464 simple_cst_equal 1047 612 mem_loc_descriptor 1173 442 walk_tree_1 1296 515 loc_list_from_tree 1431 690 simplify_comparison 1911 752 substitute_placeholder_in_expr 1951 745 substitute_in_expr 2503 1532 cse_insn 3242 2206 try_combine With -O2: MEMs ATTR name 514 502 gen_tagged_type_die 701 536 simplify_comparison 743 877 find_decls_types_r 851 840 dwarf2out_finish 863 784 loc_list_from_tree 916 839 combine_simplify_rtx 978 878 gen_subprogram_die 1650 1475 cse_insn 1720 1782 mem_loc_descriptor 2336 1792 try_combine Summing doesn't make sense, but the routines with largest differences: -O0 532 force_to_mode 547 simple_cst_equal 640 simplify_shift_const_1 731 walk_tree_1 741 simplify_comparison 781 loc_list_from_tree 971 cse_insn 1036 try_combine 1159 substitute_placeholder_in_expr 1206 substitute_in_expr -O2 100 gen_subprogram_die 101 make_extraction 112 output_loc_sequence 122 if_then_else_cond 124 substitute_placeholder_in_expr 144 simplify_shift_const_1 165 simplify_comparison 175 cse_insn 205 simplify_if_then_else 544 try_combine (Using -g or not doesn't make a difference). I've counted all MEM rtx in the whole insn stream at finalization time (i.e. slightly less than potentially are actually generated during RTL passes). ATTR is the number of unique mem_attrs ever created by set_mem_attrs, reset to zero at each function start (including emptying the htab). That is, we save a whopping 48 kilobyte due to this fantastic hash table :-) (offseted by the need for a pointer in the MEM rtx) Just remove the whole thing. Same for the reg_attrs hash table (I haven't measured that one, though). > - might be not trivial though, as some re-uses would be able to > simply modify the attr inplace. Ciao, Michael.
