https://gcc.gnu.org/bugzilla/show_bug.cgi?id=97417
--- Comment #46 from Levy <admin at levyhsu dot com> --- Looking at gcc/passed.def and gcc/config/riscv-passes.def: pass_shorten_memrefs is inserted after NEXT_PASS (pass_rtl_store_motion); NEXT_PASS (pass_rtl_store_motion); (pass_shorten_memrefs) NEXT_PASS (pass_cse_after_global_opts); NEXT_PASS (pass_rtl_ifcvt); NEXT_PASS (pass_reginfo_init); /* Perform loop optimizations. It might be better to do them a bit sooner, but we want the profile feedback to work more efficiently. */ NEXT_PASS (pass_loop2); PUSH_INSERT_PASSES_WITHIN (pass_loop2) NEXT_PASS (pass_rtl_loop_init); NEXT_PASS (pass_rtl_move_loop_invariants); NEXT_PASS (pass_rtl_unroll_loops); NEXT_PASS (pass_rtl_doloop); NEXT_PASS (pass_rtl_loop_done); POP_INSERT_PASSES () NEXT_PASS (pass_lower_subreg2); NEXT_PASS (pass_web); NEXT_PASS (pass_rtl_cprop); NEXT_PASS (pass_cse2); NEXT_PASS (pass_rtl_dse1); NEXT_PASS (pass_rtl_fwprop_addr); NEXT_PASS (pass_inc_dec); NEXT_PASS (pass_initialize_regs); NEXT_PASS (pass_ud_rtl_dce); NEXT_PASS (pass_combine); NEXT_PASS (pass_if_after_combine); NEXT_PASS (pass_jump_after_combine); NEXT_PASS (pass_partition_blocks); NEXT_PASS (pass_outof_cfg_layout_mode); NEXT_PASS (pass_split_all_insns); NEXT_PASS (pass_lower_subreg3); NEXT_PASS (pass_df_initialize_no_opt); NEXT_PASS (pass_stack_ptr_mod); NEXT_PASS (pass_mode_switching); NEXT_PASS (pass_match_asm_constraints); NEXT_PASS (pass_sms); NEXT_PASS (pass_live_range_shrinkage); NEXT_PASS (pass_sched); NEXT_PASS (pass_early_remat); NEXT_PASS (pass_ira); NEXT_PASS (pass_reload); NEXT_PASS (pass_postreload); PUSH_INSERT_PASSES_WITHIN (pass_postreload) NEXT_PASS (pass_postreload_cse); NEXT_PASS (pass_gcse2); NEXT_PASS (pass_split_after_reload); ...... After some debugging processes. it seems either: 1.The address cost info was calculated between (pass_combine) and (pass_shorten_memrefs) for patched version, then merged in the combined pass. patched one is failed to be recognized as unpathed one due to Sign/Zero extend then Subreg. This can be verified by adding -fdisable-rtl-combine option when compile, also the address_cost was not called for the whole time. 2.4 insn was determined(or say fixed?) before (pass_rtl_fwprop_addr), as for patched version, I saw forward_propagate_and_simplify() was called for 4 extra times, then pass all the way to propagate_rtx()->propagate_rtx_1()->should_replace_address()->address_cost() in fwprop.c I've also tested the (pass_postreload) as mentioned by Jim and new_address_profitable_p(). But they seem not to be the right one. Need some time to examine and trace the pass between (pass_shorten_memrefs) and (pass_rtl_fwprop_addr).
