[PATCH] add -fmove-loop-stores option to control GIMPLE loop store-motion

[Martin Sebor]

On 7/2/21 5:55 AM, Richard Biener wrote:
> This adds the -fmove-loop-stores option, mainly as a way to disable
> the store-motion part of GIMPLE invariant motion (-ftree-loop-im)
> which is enabled by default.  It might be sensible to turn off
> -fmove-loop-stores at -O1 since it can result in compile-time
> as well as memory usage issues but this patch tries to preserve
> existing behavior besides introducing the new option with the
> exception of -Og where I've disabled it.
> 
> Controlling store-motion has been made easy by earlier refactoring
> for the invariant motion only use after loop interchange.
> 
> Bootstrap and regtest running on x86_64-unknown-linux-gnu.
> 
> OK?
> 
> Thanks,
> Richard.
> 
> 2021-07-02  Richard Biener  <rguenther@suse.de>
> 
> 	* doc/invoke.texi (fmove-loop-stores): Document.
> 	* common.opt (fmove-loop-stores): New option.
> 	* opts.c (default_options_table): Enable -fmove-loop-stores
> 	at -O1 but not -Og.
> 	* tree-ssa-loop-im.c (pass_lim::execute): Pass
> 	flag_move_loop_stores instead of true to
> 	loop_invariant_motion_in_fun.
> ---
>   gcc/common.opt         |  4 ++++
>   gcc/doc/invoke.texi    | 11 +++++++++--
>   gcc/opts.c             |  1 +
>   gcc/tree-ssa-loop-im.c |  2 +-
>   4 files changed, 15 insertions(+), 3 deletions(-)
> 
> diff --git a/gcc/common.opt b/gcc/common.opt
> index 5b03bbc6662..d9da1131eda 100644
> --- a/gcc/common.opt
> +++ b/gcc/common.opt
> @@ -2084,6 +2084,10 @@ fmove-loop-invariants
>   Common Var(flag_move_loop_invariants) Optimization
>   Move loop invariant computations out of loops.
>   
> +fmove-loop-stores
> +Common Var(flag_move_loop_stores) Optimization
> +Move stores out of loops.
> +
>   fdce
>   Common Var(flag_dce) Init(1) Optimization
>   Use the RTL dead code elimination pass.
> diff --git a/gcc/doc/invoke.texi b/gcc/doc/invoke.texi
> index a9fd5fdc104..7b4f5d26738 100644
> --- a/gcc/doc/invoke.texi
> +++ b/gcc/doc/invoke.texi
> @@ -528,7 +528,7 @@ Objective-C and Objective-C++ Dialects}.
>   -floop-parallelize-all  -flra-remat  -flto  -flto-compression-level @gol
>   -flto-partition=@var{alg}  -fmerge-all-constants @gol
>   -fmerge-constants  -fmodulo-sched  -fmodulo-sched-allow-regmoves @gol
> --fmove-loop-invariants  -fno-branch-count-reg @gol
> +-fmove-loop-invariants  -fmove-loop-stores  -fno-branch-count-reg @gol
>   -fno-defer-pop  -fno-fp-int-builtin-inexact  -fno-function-cse @gol
>   -fno-guess-branch-probability  -fno-inline  -fno-math-errno  -fno-peephole @gol
>   -fno-peephole2  -fno-printf-return-value  -fno-sched-interblock @gol
> @@ -10260,6 +10260,7 @@ compilation time.
>   -fipa-reference-addressable @gol
>   -fmerge-constants @gol
>   -fmove-loop-invariants @gol
> +-fmove-loop-stores@gol
>   -fomit-frame-pointer @gol
>   -freorder-blocks @gol
>   -fshrink-wrap @gol
> @@ -10403,7 +10404,7 @@ optimization flags except for those that may interfere with debugging:
>   @gccoptlist{-fbranch-count-reg  -fdelayed-branch @gol
>   -fdse  -fif-conversion  -fif-conversion2  @gol
>   -finline-functions-called-once @gol
> --fmove-loop-invariants  -fssa-phiopt @gol
> +-fmove-loop-invariants  -fmove-loop-stores  -fssa-phiopt @gol
>   -ftree-bit-ccp  -ftree-dse  -ftree-pta  -ftree-sra}
>   
>   @end table
> @@ -13011,6 +13012,12 @@ Enabled by @option{-O3}, @option{-fprofile-use}, and @option{-fauto-profile}.
>   Enables the loop invariant motion pass in the RTL loop optimizer.  Enabled
>   at level @option{-O1} and higher, except for @option{-Og}.
>   
> +@item -fmove-loop-stores
> +@opindex fmove-loop-stores
> +Enables the loop store motion pass in the GIMPLE loop optimizer.  Note for
> +this option to have an effect @code{-ftree-loop-im} has to be enabled as well.
                                  ^^^^^

The @code markup should be @option as well (same as below).

I find the brief text added to gcc/common.opt more informative than
this longer description.  Explaining what the store motion pass does
in a few words would be helpful to those not familiar with
the implementation.

Martin


> +Enabled at level @option{-O1} and higher, except for @option{-Og}.
> +
>   @item -fsplit-loops
>   @opindex fsplit-loops
>   Split a loop into two if it contains a condition that's always true
> diff --git a/gcc/opts.c b/gcc/opts.c
> index f159bb35130..25282f71a3b 100644
> --- a/gcc/opts.c
> +++ b/gcc/opts.c
> @@ -575,6 +575,7 @@ static const struct default_options default_options_table[] =
>       { OPT_LEVELS_1_PLUS_NOT_DEBUG, OPT_fif_conversion2, NULL, 1 },
>       { OPT_LEVELS_1_PLUS_NOT_DEBUG, OPT_finline_functions_called_once, NULL, 1 },
>       { OPT_LEVELS_1_PLUS_NOT_DEBUG, OPT_fmove_loop_invariants, NULL, 1 },
> +    { OPT_LEVELS_1_PLUS_NOT_DEBUG, OPT_fmove_loop_stores, NULL, 1 },
>       { OPT_LEVELS_1_PLUS_NOT_DEBUG, OPT_fssa_phiopt, NULL, 1 },
>       { OPT_LEVELS_1_PLUS_NOT_DEBUG, OPT_fipa_modref, NULL, 1 },
>       { OPT_LEVELS_1_PLUS_NOT_DEBUG, OPT_ftree_bit_ccp, NULL, 1 },
> diff --git a/gcc/tree-ssa-loop-im.c b/gcc/tree-ssa-loop-im.c
> index e7a3050ba9d..9ac390b9a4b 100644
> --- a/gcc/tree-ssa-loop-im.c
> +++ b/gcc/tree-ssa-loop-im.c
> @@ -3258,7 +3258,7 @@ pass_lim::execute (function *fun)
>   
>     if (number_of_loops (fun) <= 1)
>       return 0;
> -  unsigned int todo = loop_invariant_motion_in_fun (fun, true);
> +  unsigned int todo = loop_invariant_motion_in_fun (fun, flag_move_loop_stores);
>   
>     if (!in_loop_pipeline)
>       loop_optimizer_finalize ();
>