[Bug tree-optimization/31090] Revision 121302 causes 30% performance regression
--- Comment #15 from ubizjak at gmail dot com 2007-06-23 10:00 --- (In reply to comment #14) Subject: Bug 31090 Author: dnovillo Date: Wed Apr 11 17:14:06 2007 New Revision: 123719 HJ, what is the situation w.r.t performance regression after the above patch was committed to SVN? -- http://gcc.gnu.org/bugzilla/show_bug.cgi?id=31090
[Bug tree-optimization/31090] Revision 121302 causes 30% performance regression
--- Comment #16 from hjl at lucon dot org 2007-06-23 14:46 --- HJ, what is the situation w.r.t performance regression after the above patch was committed to SVN? I think it still needs tuning. I got the following on Linux/Intel64: (r125740 - r121297)/r121297 410.bwaves -0.869565% 416.gamess -0.574713% 433.milc -0.840336% 434.zeusmp 3.37838% 435.gromacs -0.214823% 436.cactusADM-0.917431% 437.leslie3d -11.9048% 444.namd -0.671141% 447.dealII -2.33463% 450.soplex -0.675676% 453.povray 2.51256% 454.calculix 25.8675% 459.GemsFDTD 0% 465.tonto2.14286% 470.lbm -2.75862% 481.wrf 5.94059% 482.sphinx3 0.510204% SPECfp_base2006 0.757576% -- http://gcc.gnu.org/bugzilla/show_bug.cgi?id=31090
[Bug tree-optimization/31090] Revision 121302 causes 30% performance regression
--- Comment #14 from dnovillo at gcc dot gnu dot org 2007-04-11 17:14 --- Subject: Bug 31090 Author: dnovillo Date: Wed Apr 11 17:14:06 2007 New Revision: 123719 URL: http://gcc.gnu.org/viewcvs?root=gccview=revrev=123719 Log: PR 30735 PR 31090 * doc/invoke.texi: Document --params max-aliased-vops and avg-aliased-vops. * tree-ssa-operands.h (get_mpt_for, dump_memory_partitions, debug_memory_partitions): Move to tree-flow.h * params.h (AVG_ALIASED_VOPS): Define. * tree-ssa-alias.c (struct mp_info_def): Remove. Update all users. (mp_info_t): Likewise. (get_mem_sym_stats_for): New. (set_memory_partition): Move from tree-flow-inline.h. (mark_non_addressable): Only clear the set of symbols for the partition if it exists. (dump_memory_partitions): Move from tree-ssa-operands.c (debug_memory_partitions): Likewise. (need_to_partition_p): New. (dump_mem_ref_stats): New. (debug_mem_ref_stats): New. (dump_mem_sym_stats): New. (debug_mem_sym_stats): New. (update_mem_sym_stats_from_stmt): New. (compare_mp_info_entries): New. (mp_info_cmp): Call it. (sort_mp_info): Change argument to a list of mem_sym_stats_t objects. (get_mpt_for): Move from tree-ssa-operands.c. (find_partition_for): New. (create_partition_for): Remove. (estimate_vop_reduction): New. (update_reference_counts): New. (build_mp_info): New. (compute_memory_partitions): Refactor. Document new heuristic. Call build_mp_info, update_reference_counts, find_partition_for and estimate_vop_reduction. (compute_may_aliases): Populate virtual operands before calling debugging dumps. (delete_mem_sym_stats): New. (delete_mem_ref_stats): New. (init_mem_ref_stats): New. (init_alias_info): Call it. (maybe_create_global_var): Remove alias_info argument. Get number of call sites and number of pure/const call sites from gimple_mem_ref_stats(). (dump_alias_info): Call dump_memory_partitions first. (dump_points_to_info_for): Show how many times a pointer has been dereferenced. * opts.c (decode_options): For -O2 set --param max-aliased-vops to 500. For -O3 set --param max-aliased-vops to 1000 and --param avg-aliased-vops to 3. * fortran/options.c (gfc_init_options): Remove assignment to MAX_ALIASED_VOPS. * tree-flow-inline.h (gimple_mem_ref_stats): New. * tree-dfa.c (dump_variable): Dump memory reference statistics. Dump NO_ALIAS* settings. (referenced_var_lookup): Tidy. (mem_sym_stats): New. * tree-ssa-copy.c (may_propagate_copy): Return true if DEST and ORIG are different SSA names for a memory partition. * tree-ssa.c (delete_tree_ssa): Call delete_mem_ref_stats. * tree-flow.h (struct mem_sym_stats_d): Define. (mem_sym_stats_t): Define. (struct mem_ref_stats_d): Define. (struct gimple_df): Add field mem_ref_stats. (enum noalias_state): Define. (struct var_ann_d): Add bitfield noalias_state. (mem_sym_stats, delete_mem_ref_stats, dump_mem_ref_stats, debug_mem_ref_stats, debug_memory_partitions, debug_mem_sym_stats): Declare. * tree-ssa-structalias.c (update_alias_info): Update call sites, pure/const call sites and asm sites in structure returned by gimple_mem_ref_stats. Remove local variable IS_POTENTIAL_DEREF. Increase NUM_DEREFS if the memory expression is a potential dereference. Call update_mem_sym_stats_from_stmt. If the memory references memory, call update_mem_sym_stats_from_stmt for all the direct memory symbol references found. (intra_create_variable_infos): Set noalias_state field for pointer arguments according to the value of flag_argument_noalias. * tree-ssa-structalias.h (struct alias_info): Remove fields num_calls_found and num_pure_const_calls_found. (update_mem_sym_stats_from_stmt): Declare. * params.def (PARAM_MAX_ALIASED_VOPS): Change description. Set default value to 100. (PARAM_AVG_ALIASED_VOPS): Define. Modified: trunk/gcc/ChangeLog trunk/gcc/doc/invoke.texi trunk/gcc/fortran/options.c trunk/gcc/opts.c trunk/gcc/params.def trunk/gcc/params.h trunk/gcc/tree-dfa.c trunk/gcc/tree-flow-inline.h trunk/gcc/tree-flow.h trunk/gcc/tree-ssa-alias.c trunk/gcc/tree-ssa-copy.c trunk/gcc/tree-ssa-operands.c trunk/gcc/tree-ssa-operands.h trunk/gcc/tree-ssa-structalias.c trunk/gcc/tree-ssa-structalias.h trunk/gcc/tree-ssa.c trunk/gcc/tree-vrp.c --
[Bug tree-optimization/31090] Revision 121302 causes 30% performance regression
--- Comment #2 from rguenth at gcc dot gnu dot org 2007-03-09 10:32 --- It looks like a no-op change. For reference: --- ChangeLog (revision 121301) +++ ChangeLog (revision 121302) @@ -1,3 +1,41 @@ +2007-01-28 Daniel Berlin [EMAIL PROTECTED] + + * tree.h (struct tree_memory_tag): Add aliases member. + (MTAG_ALIASES): New macro. + * tree-ssa-alias.c (alias_bitmap_obstack): New variable. + (add_may_alias): Remove pointer-set. Update for may_aliases being + a bitmap. + (mark_aliases_call_clobbered): Update for may_aliases being a + bitmap. + (compute_tag_properties): Ditto. + (create_partition_for): Ditto. + (compute_memory_partitions): Ditto. + (dump_may_aliases_for): Ditto. + (is_aliased_with): Ditto. + (add_may_alias_for_new_tag): Ditto. + (rewrite_alias_set_for): Rewrite for may_aliases being a bitmap. + (compute_is_aliased): New function. + (compute_may_aliases): Call compute_is_aliased). + (init_alias_info): Initialize alias_bitmap_obstack. + (union_alias_set_into): New function. + (compute_flow_sensitive_aliasing): Use union_aliases_into. + (have_common_aliases_p): Rewrite to take two bitmaps and use + intersection. + (compute_flow_insensitive_aliasing): Stop using pointer-sets. + Update for bitmaps. + (finalize_ref_all_pointers): Update for add_may_alias changes. + (new_type_alias): Ditto. + * tree-flow-inline.h (may_aliases): Return a bitmap. + * tree-dfa.c (dump_variable): Check for MTAG_P'ness. + * tree-ssa.c (verify_flow_insensitive_alias_info): Update for + may_aliases being a bitmap. + * tree-flow.h (struct var_ann_d): Remove may_aliases member. + may_aliases now returns a bitmap. + * tree-ssa-structalias.c (merge_smts_into): Update for may_aliases + being a bitmap. + * tree-ssa-operands.c (add_virtual_operand): Update for + may_aliases being a bitmap. -- http://gcc.gnu.org/bugzilla/show_bug.cgi?id=31090
[Bug tree-optimization/31090] Revision 121302 causes 30% performance regression
--- Comment #3 from hjl at lucon dot org 2007-03-09 17:11 --- Gcc 4.3 revision 122738 has the same issue. -- http://gcc.gnu.org/bugzilla/show_bug.cgi?id=31090
[Bug tree-optimization/31090] Revision 121302 causes 30% performance regression
--- Comment #4 from hjl at lucon dot org 2007-03-09 17:40 --- Created an attachment (id=13180) -- (http://gcc.gnu.org/bugzilla/attachment.cgi?id=13180action=view) A breakdown testcase xxx.f90 is the problem. -fdump-tree-all shows that revision 121302 generates very different alias info. -- http://gcc.gnu.org/bugzilla/show_bug.cgi?id=31090
[Bug tree-optimization/31090] Revision 121302 causes 30% performance regression
--- Comment #5 from dberlin at gcc dot gnu dot org 2007-03-09 17:51 --- Subject: Re: Revision 121302 causes 30% performance regression Could you attach dumps for fdump-tree-alias-vops-details-blocks-stats (tarr'ing up the resulting dumps is fine) for before and after? I don't have a clean tree around, and it would really help. (If not, it's okay, it would just make my analysis faster). On 9 Mar 2007 17:40:50 -, hjl at lucon dot org [EMAIL PROTECTED] wrote: --- Comment #4 from hjl at lucon dot org 2007-03-09 17:40 --- Created an attachment (id=13180) -- (http://gcc.gnu.org/bugzilla/attachment.cgi?id=13180action=view) -- (http://gcc.gnu.org/bugzilla/attachment.cgi?id=13180action=view) A breakdown testcase xxx.f90 is the problem. -fdump-tree-all shows that revision 121302 generates very different alias info. -- http://gcc.gnu.org/bugzilla/show_bug.cgi?id=31090 --- You are receiving this mail because: --- You are on the CC list for the bug, or are watching someone who is. -- http://gcc.gnu.org/bugzilla/show_bug.cgi?id=31090
[Bug tree-optimization/31090] Revision 121302 causes 30% performance regression
--- Comment #6 from hjl at lucon dot org 2007-03-09 17:57 --- Created an attachment (id=13181) -- (http://gcc.gnu.org/bugzilla/attachment.cgi?id=13181action=view) Dumps of -fdump-tree-alias-vops-details-blocks-stats 121301 is from revision 121301 and 121302 is from revision 121302. -- http://gcc.gnu.org/bugzilla/show_bug.cgi?id=31090
[Bug tree-optimization/31090] Revision 121302 causes 30% performance regression
--- Comment #7 from rguenth at gcc dot gnu dot org 2007-03-09 23:00 --- The obvoious difference is more precise alias information: -bar: Maximum number of VOPS needed per statement: 80 +bar: Maximum number of VOPS needed per statement: 71 (and the resulting different partitioning). What may cause performance problems are changes like - # VUSE MPT.120_683 + # VUSE SMT.109_703, MPT.120_704 D.1662_631 = (*D.1657_625)[D.1661_630]; - # MPT.120_690 = VDEF MPT.120_683 + # SMT.109_711 = VDEF SMT.109_703 + # MPT.120_712 = VDEF MPT.120_704 (*D.1634_608)[D.1644_621] = D.1666_635; as our tree memory optimizers don't like statements with multiple virtual operands (at least store_ccp and store_copyprop don't, maybe PRE doesn't care, I'm also not sure about LIM). Other than that, more precise alias information can cause more register pressure, too. -- http://gcc.gnu.org/bugzilla/show_bug.cgi?id=31090
Re: [Bug tree-optimization/31090] Revision 121302 causes 30% performance regression
On 9 Mar 2007 23:00:55 -, rguenth at gcc dot gnu dot org [EMAIL PROTECTED] wrote: Other than that, more precise alias information can cause more register pressure, too. Fix the register allocator instead of complaining about this issue. I am sorry but if people want a compiler which works for x86, they just need to fix the RA. Now I could care less about x86. -- Pinski
[Bug tree-optimization/31090] Revision 121302 causes 30% performance regression
--- Comment #8 from pinskia at gmail dot com 2007-03-09 23:05 --- Subject: Re: Revision 121302 causes 30% performance regression On 9 Mar 2007 23:00:55 -, rguenth at gcc dot gnu dot org [EMAIL PROTECTED] wrote: Other than that, more precise alias information can cause more register pressure, too. Fix the register allocator instead of complaining about this issue. I am sorry but if people want a compiler which works for x86, they just need to fix the RA. Now I could care less about x86. -- Pinski -- http://gcc.gnu.org/bugzilla/show_bug.cgi?id=31090
[Bug tree-optimization/31090] Revision 121302 causes 30% performance regression
--- Comment #9 from rguenth at gcc dot gnu dot org 2007-03-09 23:17 --- I didn't complain about register pressure. You need to get a thicker skin. -- http://gcc.gnu.org/bugzilla/show_bug.cgi?id=31090
[Bug tree-optimization/31090] Revision 121302 causes 30% performance regression
--- Comment #10 from dberlin at gcc dot gnu dot org 2007-03-09 23:48
---
(In reply to comment #7)
The obvoious difference is more precise alias information:
-bar: Maximum number of VOPS needed per statement: 80
+bar: Maximum number of VOPS needed per statement: 71
See, it was buggy before and still getting dup'd symbols into the points-to
sets.
In particular:
80 NMT.114, UID 1780, real8[0:], is addressable, is global, call clobbered (,
is global var ), may aliases: { dyinv dxinv dzinv epsinv SMT.109 dyinv dxinv
dzinv epsinv SMT.109 dyinv dxinv dzinv epsinv SMT.109 dyinv dxinv dzinv epsinv
SMT.109 dyinv dxinv dzinv epsinv SMT.109 dyinv dxinv dzinv epsinv SMT.109 dyinv
dxinv dzinv epsinv SMT.109 dyinv dxinv dzinv epsinv SMT.109 dyinv dxinv dzinv
epsinv SMT.109 dyinv dxinv dzinv epsinv SMT.109 dyinv dxinv dzinv epsinv
SMT.109 dyinv dxinv dzinv epsinv SMT.109 dyinv dxinv dzinv epsinv SMT.109 dyinv
dxinv dzinv epsinv SMT.109 dyinv dxinv dzinv epsinv SMT.109 dyinv dxinv dzinv
epsinv SMT.109 }
Note it's the same set of symbols repeated 14 times
vs
5 NMT.114, UID 1780, real8[0:], is addressable, is global, call clobbered (,
is global var ), may aliases: { dyinv dxinv dzinv epsinv SMT.109 }
This caused different partitioning, and this caused your performance drop.
Sorry, nothing i can do about this, there is no bug here (in fact, it just
shows that the patch removed a bug)
You need Diego to look at the partitioning algorithms to get your performance
back.
--
dberlin at gcc dot gnu dot org changed:
What|Removed |Added
CC||dnovillo at gcc dot gnu dot
||org
http://gcc.gnu.org/bugzilla/show_bug.cgi?id=31090
[Bug tree-optimization/31090] Revision 121302 causes 30% performance regression
--- Comment #11 from dnovillo at gcc dot gnu dot org 2007-03-09 23:53 --- I'm already handling this family of performance problems. I need a few more days to finish the WIP patch I have. In the meantime, see if increasing --param max-aliased-vops works around the problem. -- dnovillo at gcc dot gnu dot org changed: What|Removed |Added BugsThisDependsOn||30735 http://gcc.gnu.org/bugzilla/show_bug.cgi?id=31090
[Bug tree-optimization/31090] Revision 121302 causes 30% performance regression
--- Comment #12 from hjl at lucon dot org 2007-03-10 00:04 --- --param max-aliased-vops=100 works: [EMAIL PROTECTED] 597]$ make /usr/gcc-next/bin/gfortran -c -O2 --param max-aliased-vops=100 -o 301.o test597.f90 /usr/gcc-4.3/bin/gfortran -o 301 301.o -Wl,-rpath,/usr/gcc-4.3/lib64 /usr/gcc-last/bin/gfortran -c -O2 --param max-aliased-vops=100 -o 302.o test597.f90 /usr/gcc-4.3/bin/gfortran -o 302 302.o -Wl,-rpath,/usr/gcc-4.3/lib64 /usr/gcc-last/bin/gfortran -c -O2 --param max-aliased-vops=100 -o 43.o test597.f90 /usr/gcc-4.3/bin/gfortran -o 43 43.o -Wl,-rpath,/usr/gcc-4.3/lib64 time ./301 0.86user 0.00system 0:00.87elapsed 99%CPU (0avgtext+0avgdata 0maxresident)k 0inputs+0outputs (0major+2326minor)pagefaults 0swaps time ./302 0.86user 0.00system 0:00.87elapsed 99%CPU (0avgtext+0avgdata 0maxresident)k 0inputs+0outputs (0major+2326minor)pagefaults 0swaps time ./43 0.86user 0.00system 0:00.87elapsed 99%CPU (0avgtext+0avgdata 0maxresident)k 0inputs+0outputs (0major+2327minor)pagefaults 0swaps [EMAIL PROTECTED] 597]$ -- http://gcc.gnu.org/bugzilla/show_bug.cgi?id=31090
[Bug tree-optimization/31090] Revision 121302 causes 30% performance regression
--- Comment #13 from dnovillo at redhat dot com 2007-03-10 00:07 --- Subject: Re: Revision 121302 causes 30% performance regression hjl at lucon dot org wrote on 03/09/07 19:04: --param max-aliased-vops=100 works: OK, thanks. I'll add this PR to the mix then. -- http://gcc.gnu.org/bugzilla/show_bug.cgi?id=31090
[Bug tree-optimization/31090] Revision 121302 causes 30% performance regression
--- Comment #1 from hjl at lucon dot org 2007-03-08 20:04 --- Created an attachment (id=13173) -- (http://gcc.gnu.org/bugzilla/attachment.cgi?id=13173action=view) A testcase /usr/gcc-next/bin/gfortran -c -O2 -o 301.o test597.f90 /usr/gcc-next/bin/gfortran -o 301 301.o -Wl,-rpath,/usr/gcc-4.3/lib64 /usr/gcc-last/bin/gfortran -c -O2 -o 302.o test597.f90 /usr/gcc-next/bin/gfortran -o 302 302.o -Wl,-rpath,/usr/gcc-4.3/lib64 time ./301 0.92user 0.00system 0:00.93elapsed 99%CPU (0avgtext+0avgdata 0maxresident)k 0inputs+0outputs (0major+2325minor)pagefaults 0swaps time ./302 1.24user 0.00system 0:01.24elapsed 99%CPU (0avgtext+0avgdata 0maxresident)k 0inputs+0outputs (0major+2325minor)pagefaults 0swaps [EMAIL PROTECTED] 597]$ /usr/gcc-next/bin/gfortran -v Using built-in specs. Target: x86_64-unknown-linux-gnu Configured with: /net/gnu-13/export/gnu/src/gcc-next/gcc/configure --enable-clocale=gnu --with-system-zlib --enable-decimal-float=yes --with-demangler-in-ld --enable-languages=c,fortran --enable-shared --enable-threads=posix --enable-haifa --enable-checking=assert --prefix=/usr/gcc-next --with-local-prefix=/usr/local Thread model: posix gcc version 4.3.0 20070129 (experimental) [trunk revision 121301] [EMAIL PROTECTED] 597]$ /usr/gcc-last/bin/gfortran -v Using built-in specs. Target: x86_64-unknown-linux-gnu Configured with: /net/gnu-13/export/gnu/src/gcc-last/gcc/configure --enable-clocale=gnu --with-system-zlib --enable-decimal-float=yes --with-demangler-in-ld --enable-languages=c,fortran --enable-shared --enable-threads=posix --enable-haifa --enable-checking=assert --prefix=/usr/gcc-last --with-local-prefix=/usr/local Thread model: posix gcc version 4.3.0 20070129 (experimental) [trunk revision 121302] -- http://gcc.gnu.org/bugzilla/show_bug.cgi?id=31090
