Re: [PATCH][RFA] Split passes from gcse_main into separate passes

["H.J. Lu" <hjl dot tools at gmail dot com>]

On Sun, Apr 26, 2009 at 2:23 PM, Steven Bosscher <stevenb.gcc@gmail.com> wrote:
> Hi,
>
> This is the next step in the big gcse cleanup. ?This patch splits the
> uber-pass gcse_main() into individual passes: CPROP, PRE, HOIST,
> STORE_MOTION, and local CSE.
>
> After this split, we no longer have a reason to *not* alter jumps or
> bypass jumps in CPROP. Therefore, the separate jump bypassing pass is
> removed and CPROP now always alters jumps and bypasses jumps (this is
> extremely cheap anyway).
>
> Bootstrapped&tested (-m32/-m64) with c,c++,fortran on
> x86_64-unknown-linux-gnu with store motion enabled.
> Bootstrapped&tested (-m32/-m64) with c,c++,fortran on
> ia64-unknown-linux-gnu with store motion enabled.
> Bootstrap&test of all languages ongoing on ia64-unknown-linux-gnu.
>
> OK for trunk if everything passes?
>
> Ciao!
> Steven
>
>
>
>
> ? ? ? ?* dbgcnt.def (cprop1, cprop2, gcse, jump_bypass): Remove
> ? ? ? ?(cprop, hoist, pre, store_motion): New debug counters.
> ? ? ? ?* tree-pass.h (pass_tracer): Move to list of gimple passes, it
> ? ? ? ?is not an RTL pass anymore.
> ? ? ? ?(pass_profiling): Remove extern decl for pass removed in 2005.
> ? ? ? ?(pass_gcse, pass_jump_bypass): Remove.
> ? ? ? ?* final.c (rest_of_clean_state): Set flag_rerun_cse_after_global_opts
> ? ? ? ?to 0 for clean state.
> ? ? ? ?* toplev.h (flag_rerun_cse_after_global_opts): Add extern declaration.
> ? ? ? ?* cse.c (gate_handle_cse_after_global_opts,
> ? ? ? ?rest_of_handle_cse_after_global_opts): New functions.
> ? ? ? ?(pass_cse_after_global_opts): New pass, does local CSE.
> ? ? ? ?* timevar.def (TV_GCSE, TV_CPROP1, TV_CPROP2, TV_BYPASS): Remove.
> ? ? ? ?(TV_CPROP): New timevar.
> ? ? ? ?* gcse.c (flag_rerun_cse_after_global_opts): New global variable.
> ? ? ? ?(run_jump_opt_after_gcse, max_gcse_regno): Remove global vars.
> ? ? ? ?(gcse_main, recompute_all_luids): Remove.
> ? ? ? ?(compute_hash_table_work): Call max_reg_num instead of reading
> ? ? ? ?max_gcse_regno.
> ? ? ? ?(cprop_jump): Don't set run_jump_opt_after_gcse.
> ? ? ? ?(constprop_register): Always allow to alter jumps.
> ? ? ? ?(cprop_insn): Likewise.
> ? ? ? ?(do_local_cprop): Likewise.
> ? ? ? ?(local_cprop_pass): Likewise. ?Return non-zero if something changed.
> ? ? ? ?(cprop): Remove function, fold interesting bits into one_cprop_pass.
> ? ? ? ?(find_implicit_sets): Add note about missed optimization opportunity.
> ? ? ? ?(one_cprop_pass): Rewrite to be "the" CPROP pass, called from the
> ? ? ? ?pass_rtl_cprop execute function.
> ? ? ? ?Don't bother tracking the pass number, each pass gets its own dumpfile
> ? ? ? ?now anyway.
> ? ? ? ?Always allow to alter jumpsand bypass jumps.
> ? ? ? ?(bypass_block): Don't ignore regno >= max_gcse_regno, find_bypass_set
> ? ? ? ?will just find no suitable set.
> ? ? ? ?(pre_edge_insert): Fix dumping, this function is for PRE only.
> ? ? ? ?(one_pre_gcse_pass): Rewrite to be "the" PRE pass, called from the
> ? ? ? ?pass_rtl_pre execute function.
> ? ? ? ?(hoist_code): Return non-zero if something changed. ?Keep track of
> ? ? ? ?substitutions and insertions for statistics gathering similar to PRE.
> ? ? ? ?(one_code_hoisting_pass): Rewrite to be "the" code hoisting pass,
> ? ? ? ?called from the pass_rtl_hoist execute function. ?Show pass statistics.
> ? ? ? ?(compute_store_table): Use max_reg_num directly instead of using the
> ? ? ? ?formerly global max_gcse_regno.
> ? ? ? ?(build_store_vectors): Likewise.
> ? ? ? ?(replace_store_insn): Fix dumping.
> ? ? ? ?(store_motion): Rename to ...
> ? ? ? ?(one_store_motion_pass): ... this. ?Rewrite to be "the" STORE_MOTION
> ? ? ? ?pass, called from the pass_rtl_store_motion execute function. ?Keep
> ? ? ? ?track of substitutions and insertions for statistics gathering similar
> ? ? ? ?to PRE.
> ? ? ? ?(bypass_jumps): Remove, fold interesting bits into ...
> ? ? ? ?(one_cprop_pass): ... this. ?Rewrite to be "the" CPROP pass, called
> ? ? ? ?from the pass_rtl_cprop execute function.
> ? ? ? ?(gate_handle_jump_bypass, rest_of_handle_jump_bypass,
> ? ? ? ?pass_jump_bypass): Remove.
> ? ? ? ?(gate_handle_gcse, rest_of_handle_gcse): Remove.
> ? ? ? ?(gate_rtl_cprop, execute_rtl_cprop, pass_rtl_cprop): New.
> ? ? ? ?(gate_rtl_pre, execute_rtl_pre, pass_rtl_pre): New.
> ? ? ? ?(gate_rtl_hoist, execute_rtl_hoist, pass_rtl_hoist): New.
> ? ? ? ?(gate_rtl_store_motion, execute_rtl_store_motion,
> ? ? ? ?pass_rtl_store_motion): New.
> ? ? ? ?* common.opt: Remove flag_cse_skip_blocks, adjust documentation to
> ? ? ? ?make it clear that -fcse-skip-blocks is a no-op for backward compat.
> ? ? ? ?* passes.c (init_optimization_passes): Remove pass_gcse and
> ? ? ? ?pass_jump_bypass. ?Schedule cprop, pre, hoist, cprop, store_motion,
> ? ? ? ?and cse_after_global_opts in place of pass_gcse. ?Schedule cprop
> ? ? ? ?instead of pass_jump_bypass.
>

This caused:

http://gcc.gnu.org/bugzilla/show_bug.cgi?id=46777


H.J.