| Message ID | mpty31atc1z.fsf@arm.com |
|---|---|
| State | New |
| Headers | show |
| Series | Don't run DSE at -Og | expand |
On 7/7/19 3:43 AM, Richard Sandiford wrote: > This patch stops gimple and rtl DSE from running by default at -Og. > The idea is both to improve compile time and to stop us from deleting > stores that we can't track in debug info. > > We could rein this back in future for stores to local variables > with is_gimple_reg_type, but at the moment we don't have any > infrastructure for switching between binds to specific values > and binds to evolving memory locations. Even then, location > tracking only works for direct references to the variables, and doesn't > for example help with printing dereferenced pointers (see the next patch > in the series for an example). > > I'm also not sure that DSE is important enough for -Og to justify the > compile time cost -- especially in the case of RTL DSE, which is pretty > expensive. > > Tested on aarch64-linux-gnu and x86_64-linux-gnu. OK to install? > > Richard > > > 2019-07-07 Richard Sandiford <richard.sandiford@arm.com> > > gcc/ > * common.opt (Og): Change the initial value of flag_dse to 0. > * opts.c (default_options_table): Move OPT_ftree_dse from > OPT_LEVELS_1_PLUS to OPT_LEVELS_1_PLUS_NOT_DEBUG. Also add > OPT_fdse to OPT_LEVELS_1_PLUS_NOT_DEBUG. Put the OPT_ftree_pta > entry before the OPT_ftree_sra entry. > * doc/invoke.texi (Og): Add -fdse and -ftree-dse to the list > of flags disabled by Og. > > gcc/testsuite/ > * c-c++-common/guality/Og-global-dse-1.c: New test. OK. I doubt DSE is at all important if someone has asked for Og. BTW, if you're going to be poking at improving Og or debuginfo in general, you might want to loop Alex into your plans. He did a detailed evaluation of how the various options impact debugging, -Og issues, etc. He's no longer with Red Hat, but can be reached at Adacore. jeff
On Sun, Jul 7, 2019 at 9:04 PM Jeff Law <law@redhat.com> wrote: > > On 7/7/19 3:43 AM, Richard Sandiford wrote: > > This patch stops gimple and rtl DSE from running by default at -Og. > > The idea is both to improve compile time and to stop us from deleting > > stores that we can't track in debug info. > > > > We could rein this back in future for stores to local variables > > with is_gimple_reg_type, but at the moment we don't have any > > infrastructure for switching between binds to specific values > > and binds to evolving memory locations. Even then, location > > tracking only works for direct references to the variables, and doesn't > > for example help with printing dereferenced pointers (see the next patch > > in the series for an example). > > > > I'm also not sure that DSE is important enough for -Og to justify the > > compile time cost -- especially in the case of RTL DSE, which is pretty > > expensive. > > > > Tested on aarch64-linux-gnu and x86_64-linux-gnu. OK to install? > > > > Richard > > > > > > 2019-07-07 Richard Sandiford <richard.sandiford@arm.com> > > > > gcc/ > > * common.opt (Og): Change the initial value of flag_dse to 0. > > * opts.c (default_options_table): Move OPT_ftree_dse from > > OPT_LEVELS_1_PLUS to OPT_LEVELS_1_PLUS_NOT_DEBUG. Also add > > OPT_fdse to OPT_LEVELS_1_PLUS_NOT_DEBUG. Put the OPT_ftree_pta > > entry before the OPT_ftree_sra entry. > > * doc/invoke.texi (Og): Add -fdse and -ftree-dse to the list > > of flags disabled by Og. > > > > gcc/testsuite/ > > * c-c++-common/guality/Og-global-dse-1.c: New test. > OK. I doubt DSE is at all important if someone has asked for Og. Hmm. -Og also asks for runtime performance. Can we instead restrict DSE to artificial decls? I also fear that this leaves a _lot_ of abstraction for C++ code given we also do not run SRA with -Og even though that has means to preserve debug info. Richard. > BTW, if you're going to be poking at improving Og or debuginfo in > general, you might want to loop Alex into your plans. He did a detailed > evaluation of how the various options impact debugging, -Og issues, etc. > He's no longer with Red Hat, but can be reached at Adacore. > > jeff
Richard Biener <richard.guenther@gmail.com> writes: > On Sun, Jul 7, 2019 at 9:04 PM Jeff Law <law@redhat.com> wrote: >> >> On 7/7/19 3:43 AM, Richard Sandiford wrote: >> > This patch stops gimple and rtl DSE from running by default at -Og. >> > The idea is both to improve compile time and to stop us from deleting >> > stores that we can't track in debug info. >> > >> > We could rein this back in future for stores to local variables >> > with is_gimple_reg_type, but at the moment we don't have any >> > infrastructure for switching between binds to specific values >> > and binds to evolving memory locations. Even then, location >> > tracking only works for direct references to the variables, and doesn't >> > for example help with printing dereferenced pointers (see the next patch >> > in the series for an example). >> > >> > I'm also not sure that DSE is important enough for -Og to justify the >> > compile time cost -- especially in the case of RTL DSE, which is pretty >> > expensive. >> > >> > Tested on aarch64-linux-gnu and x86_64-linux-gnu. OK to install? >> > >> > Richard >> > >> > >> > 2019-07-07 Richard Sandiford <richard.sandiford@arm.com> >> > >> > gcc/ >> > * common.opt (Og): Change the initial value of flag_dse to 0. >> > * opts.c (default_options_table): Move OPT_ftree_dse from >> > OPT_LEVELS_1_PLUS to OPT_LEVELS_1_PLUS_NOT_DEBUG. Also add >> > OPT_fdse to OPT_LEVELS_1_PLUS_NOT_DEBUG. Put the OPT_ftree_pta >> > entry before the OPT_ftree_sra entry. >> > * doc/invoke.texi (Og): Add -fdse and -ftree-dse to the list >> > of flags disabled by Og. >> > >> > gcc/testsuite/ >> > * c-c++-common/guality/Og-global-dse-1.c: New test. >> OK. I doubt DSE is at all important if someone has asked for Og. > > Hmm. -Og also asks for runtime performance. Can we instead restrict > DSE to artificial decls? I also fear that this leaves a _lot_ of abstraction > for C++ code given we also do not run SRA with -Og even though that > has means to preserve debug info. Following up your question in the other patch: I see a 2.5% slowdown for cc1plus bulit at -Og when compiling tree-into-ssa.ii at -O2 -g. I think it mostly comes from removing RTL DSE. I get a very similar slowdown with gimple DSE reenabled. There again, we're not very C++-heavy, so cc1plus might not be a good test for that. Richard
Index: gcc/common.opt =================================================================== --- gcc/common.opt 2019-07-03 20:50:45.902320943 +0100 +++ gcc/common.opt 2019-07-07 10:29:23.683336097 +0100 @@ -1955,7 +1955,7 @@ Common Var(flag_dce) Init(1) Optimizatio Use the RTL dead code elimination pass. fdse -Common Var(flag_dse) Init(1) Optimization +Common Var(flag_dse) Init(0) Optimization Use the RTL dead store elimination pass. freschedule-modulo-scheduled-loops Index: gcc/opts.c =================================================================== --- gcc/opts.c 2019-06-29 17:20:49.000000000 +0100 +++ gcc/opts.c 2019-07-07 10:29:23.691336032 +0100 @@ -466,7 +466,6 @@ static const struct default_options defa { OPT_LEVELS_1_PLUS, OPT_ftree_copy_prop, NULL, 1 }, { OPT_LEVELS_1_PLUS, OPT_ftree_dce, NULL, 1 }, { OPT_LEVELS_1_PLUS, OPT_ftree_dominator_opts, NULL, 1 }, - { OPT_LEVELS_1_PLUS, OPT_ftree_dse, NULL, 1 }, { OPT_LEVELS_1_PLUS, OPT_ftree_fre, NULL, 1 }, { OPT_LEVELS_1_PLUS, OPT_ftree_sink, NULL, 1 }, { OPT_LEVELS_1_PLUS, OPT_ftree_slsr, NULL, 1 }, @@ -477,14 +476,16 @@ static const struct default_options defa #if DELAY_SLOTS { OPT_LEVELS_1_PLUS_NOT_DEBUG, OPT_fdelayed_branch, NULL, 1 }, #endif + { OPT_LEVELS_1_PLUS_NOT_DEBUG, OPT_fdse, NULL, 1 }, { OPT_LEVELS_1_PLUS_NOT_DEBUG, OPT_fif_conversion, NULL, 1 }, { 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_fssa_phiopt, NULL, 1 }, { OPT_LEVELS_1_PLUS_NOT_DEBUG, OPT_ftree_bit_ccp, NULL, 1 }, - { OPT_LEVELS_1_PLUS_NOT_DEBUG, OPT_ftree_sra, NULL, 1 }, + { OPT_LEVELS_1_PLUS_NOT_DEBUG, OPT_ftree_dse, NULL, 1 }, { OPT_LEVELS_1_PLUS_NOT_DEBUG, OPT_ftree_pta, NULL, 1 }, + { OPT_LEVELS_1_PLUS_NOT_DEBUG, OPT_ftree_sra, NULL, 1 }, /* -O2 and -Os optimizations. */ { OPT_LEVELS_2_PLUS, OPT_fcaller_saves, NULL, 1 }, Index: gcc/doc/invoke.texi =================================================================== --- gcc/doc/invoke.texi 2019-07-03 20:50:42.346350375 +0100 +++ gcc/doc/invoke.texi 2019-07-07 10:29:23.687336062 +0100 @@ -8421,10 +8421,10 @@ no effect. Otherwise @option{-Og} enabl optimization flags except for those that may interfere with debugging: @gccoptlist{-fbranch-count-reg -fdelayed-branch @gol --fif-conversion -fif-conversion2 @gol +-fdse -fif-conversion -fif-conversion2 @gol -finline-functions-called-once @gol -fmove-loop-invariants -fssa-phiopt @gol --ftree-bit-ccp -ftree-pta -ftree-sra} +-ftree-bit-ccp -ftree-dse -ftree-pta -ftree-sra} @end table Index: gcc/testsuite/c-c++-common/guality/Og-global-dse-1.c =================================================================== --- /dev/null 2019-06-14 15:59:19.298479944 +0100 +++ gcc/testsuite/c-c++-common/guality/Og-global-dse-1.c 2019-07-07 10:29:23.691336032 +0100 @@ -0,0 +1,17 @@ +/* { dg-do run } */ +/* { dg-options "-g" } */ + +struct s { int i, j; }; +struct s gs1, gs2 = { 3, 4 }; + +void __attribute__((noipa)) consume (void) {}; + +int +main (void) +{ + gs1.i = 1; + gs1.j = 2; /* { dg-final { gdb-test . "gs1.i" "1" } } */ + gs1 = gs2; /* { dg-final { gdb-test . "gs1.j" "2" } } */ + consume (); /* { dg-final { gdb-test . "gs1.i" "3" } } */ + return 0; /* { dg-final { gdb-test . "gs1.j" "4" } } */ +}