[Bug debug/109237] csmith: another timeout with -g -O3 this time since r12-156-g8d4c374c4419a875
https://gcc.gnu.org/bugzilla/show_bug.cgi?id=109237 Richard Biener changed: What|Removed |Added Assignee|rguenth at gcc dot gnu.org |unassigned at gcc dot gnu.org Keywords|needs-reduction | Status|ASSIGNED|NEW --- Comment #16 from Richard Biener --- I wouldn't say "fixed", but at least it's as far as hotspots go.
[Bug debug/109237] csmith: another timeout with -g -O3 this time since r12-156-g8d4c374c4419a875
https://gcc.gnu.org/bugzilla/show_bug.cgi?id=109237 --- Comment #15 from Richard Biener --- Spread pretty evenly now. Everything >= 5% tree CFG cleanup : 5.25 ( 8%) 0.00 ( 0%) 5.27 ( 7%) 89k ( 0%) expand : 3.22 ( 5%) 0.18 ( 2%) 3.41 ( 4%) 2177M ( 31%) scheduling 2 : 6.93 ( 10%) 1.37 ( 18%) 8.29 ( 11%) 746M ( 11%) TOTAL : 69.08 7.60 76.70 6951M and the profile is Samples: 287K of event 'cycles:u', Event count (approx.): 376155371739 Overhead Samples Command Shared Object Symbol 3.02% 8226 cc1 cc1 [.] init_alias_analysis 3.02% 8300 cc1 cc1 [.] df_note_compute 2.65% 7299 cc1 cc1 [.] contains_no_active_insn_p 2.53% 7016 cc1 cc1 [.] delete_trivially_dead_insns 2.15% 6150 cc1 cc1 [.] get_ref_base_and_extent 1.92% 5443 cc1 cc1 [.] ggc_set_mark 1.58% 4291 cc1 cc1 [.] fast_dce 1.58% 4348 cc1 cc1 [.] gimple_purge_dead_eh_edges 1.53% 4141 cc1 cc1 [.] delete_unmarked_insns 1.52% 4146 cc1 cc1 [.] prescan_insns_for_dce 1.40% 3841 cc1 cc1 [.] cleanup_control_flow_bb 1.28% 3467 cc1 cc1 [.] rebuild_jump_labels_1 1.16% 3176 cc1 cc1 [.] df_live_bb_local_compute 1.13% 3114 cc1 cc1 [.] gimple_merge_blocks 1.09% 2977 cc1 cc1 [.] df_lr_bb_local_compute 1.05% 2962 cc1 cc1 [.] (anonymous namespace)::pass_rtl_cprop::execute 1.03% 2988 cc1 cc1 [.] for_each_inc_dec 1.01% 3168 cc1 cc1 [.] walk_tree_1 0.87% 2790 cc1 cc1 [.] ggc_internal_alloc
[Bug debug/109237] csmith: another timeout with -g -O3 this time since r12-156-g8d4c374c4419a875
https://gcc.gnu.org/bugzilla/show_bug.cgi?id=109237 --- Comment #13 from CVS Commits --- The master branch has been updated by Richard Biener : https://gcc.gnu.org/g:8f81100115f68b37fb2723e987c14a3185d1f47d commit r14-60-g8f81100115f68b37fb2723e987c14a3185d1f47d Author: Richard Biener Date: Wed Mar 22 10:05:19 2023 +0100 rtl-optimization/109237 - speedup bb_is_just_return For the testcase bb_is_just_return is on top of the profile, changing it to walk BB insns backwards puts it off the profile. That's because in the forward walk you have to process possibly many debug insns but in a backward walk you very likely run into control insns first. PR rtl-optimization/109237 * cfgcleanup.cc (bb_is_just_return): Walk insns backwards. --- Comment #14 from CVS Commits --- The master branch has been updated by Richard Biener : https://gcc.gnu.org/g:675ac8822b51a39f6a66a44858d7c31ece8700f2 commit r14-61-g675ac8822b51a39f6a66a44858d7c31ece8700f2 Author: Richard Biener Date: Wed Mar 22 09:29:49 2023 +0100 rtl-optimization/109237 - quadraticness in delete_trivially_dead_insns The following addresses quadraticness in processing debug insns in delete_trivially_dead_insns and insn_live_p by using TREE_VISITED on the INSN_VAR_LOCATION_DECL to indicate a later debug bind with the same decl and no intervening real insn or debug marker. That gets rid of the NEXT_INSN walk in insn_live_p in favor of first clearing TREE_VISITED in the first loop over insn and the book-keeping of decls we set the bit since we need to clear them when visiting a real or debug marker insn. That improves the time spent in delete_trivially_dead_insns from 10.6s to 2.2s for the testcase. PR rtl-optimization/109237 * cse.cc (insn_live_p): Remove NEXT_INSN walk, instead check TREE_VISITED on INSN_VAR_LOCATION_DECL. (delete_trivially_dead_insns): Maintain TREE_VISITED on active debug bind INSN_VAR_LOCATION_DECL.
[Bug debug/109237] csmith: another timeout with -g -O3 this time since r12-156-g8d4c374c4419a875
https://gcc.gnu.org/bugzilla/show_bug.cgi?id=109237 Richard Biener changed: What|Removed |Added Status|NEW |ASSIGNED Assignee|unassigned at gcc dot gnu.org |rguenth at gcc dot gnu.org --- Comment #12 from Richard Biener --- I have some more queued patches.
[Bug debug/109237] csmith: another timeout with -g -O3 this time since r12-156-g8d4c374c4419a875
https://gcc.gnu.org/bugzilla/show_bug.cgi?id=109237 --- Comment #11 from CVS Commits --- The master branch has been updated by Richard Biener : https://gcc.gnu.org/g:776a5bb5894315ab144dc74222fc580fde8fdd87 commit r13-6873-g776a5bb5894315ab144dc74222fc580fde8fdd87 Author: Richard Biener Date: Wed Mar 22 10:05:19 2023 +0100 rtl-optimization/109237 - speedup bb_is_just_return For the testcase bb_is_just_return is on top of the profile, changing it to walk BB insns backwards puts it off the profile. That's because in the forward walk you have to process possibly many debug insns but in a backward walk you very likely run into control insns first. PR rtl-optimization/109237 * cfgcleanup.cc (bb_is_just_return): Walk insns backwards.
[Bug debug/109237] csmith: another timeout with -g -O3 this time since r12-156-g8d4c374c4419a875
https://gcc.gnu.org/bugzilla/show_bug.cgi?id=109237 --- Comment #10 from Richard Biener --- There's similar issues with GIMPLE uses of last_stmt () which generally (but not always) want to look at a possible control transfer stmt. It looks like we're not guaranteed to have no debug stmts after such, in particular CFG cleanup seems to be ready to remove debug stmts after a noreturn call. But GIMPLE verify_flow_info should already choke on that, so it might be only a temporary thing to accept?
[Bug debug/109237] csmith: another timeout with -g -O3 this time since r12-156-g8d4c374c4419a875
https://gcc.gnu.org/bugzilla/show_bug.cgi?id=109237 --- Comment #9 from Richard Biener --- Samples: 289K of event 'cycles:u', Event count (approx.): 384226334976 Overhead Samples Command Shared Object Symbol 3.52% 9747 cc1 cc1 [.] bb_is_just_return # 2.94% 8241 cc1 cc1 [.] df_note_compute # 2.92% 8085 cc1 cc1 [.] init_alias_analysis # 2.55% 7035 cc1 cc1 [.] delete_trivially_dead_insns # 2.28% 6372 cc1 cc1 [.] contains_no_active_insn_p # 2.16% 6288 cc1 cc1 [.] get_ref_base_and_extent # 2.02% 5785 cc1 cc1 [.] ggc_set_mark # 1.55% 4308 cc1 cc1 [.] fast_dce # I see that bb_is_just_return is high in the profile and looking at its implementation I wonder whether on RTL we can scan insns backwards and stop if the last (real?) insn isn't ANY_RETURN_P ()? Using FOR_BB_INSNS_REVERSE puts it off the profile completely. Will test a patch. Similar contains_no_active_insn_p is high up in the profile and it looks like micro-optimizing it a bit would help. Using NONDEBUG_INSN_P to guard the flow_active_insn_p call doesn't seem to help (but perf is always noisy).
[Bug debug/109237] csmith: another timeout with -g -O3 this time since r12-156-g8d4c374c4419a875
https://gcc.gnu.org/bugzilla/show_bug.cgi?id=109237 Martin Liška changed: What|Removed |Added Summary|csmith: another timeout |csmith: another timeout |with -g -O3 this time |with -g -O3 this time since ||r12-156-g8d4c374c4419a875 Keywords|needs-bisection | CC||marxin at gcc dot gnu.org --- Comment #8 from Martin Liška --- I think it started with r12-156-g8d4c374c4419a875.
[Bug debug/109237] csmith: another timeout with -g -O3 this time
https://gcc.gnu.org/bugzilla/show_bug.cgi?id=109237 --- Comment #7 from Richard Biener --- Created attachment 54727 --> https://gcc.gnu.org/bugzilla/attachment.cgi?id=54727=edit patch for delete_trivially_dead_insns The attached improves that to trivially dead code: 2.16 ( 3%) I've verified the set of debug insns kept is the same as before.
[Bug debug/109237] csmith: another timeout with -g -O3 this time
https://gcc.gnu.org/bugzilla/show_bug.cgi?id=109237
--- Comment #6 from Richard Biener ---
For delete_trivially_dead_insns we have
for (insn = get_last_insn (); insn; insn = prev)
{
int live_insn = 0;
prev = PREV_INSN (insn);
if (!INSN_P (insn))
continue;
live_insn = insn_live_p (insn, counts);
/* If this is a dead insn, delete it and show registers in it aren't
being used. */
if (! live_insn && dbg_cnt (delete_trivial_dead))
{
if (DEBUG_INSN_P (insn))
...
but insn_live_p has "interesting" behavior for debug_insns:
else if (DEBUG_INSN_P (insn))
{
rtx_insn *next;
if (DEBUG_MARKER_INSN_P (insn))
return true;
for (next = NEXT_INSN (insn); next; next = NEXT_INSN (next))
if (NOTE_P (next))
continue;
else if (!DEBUG_INSN_P (next))
return true;
/* If we find an inspection point, such as a debug begin stmt,
we want to keep the earlier debug insn. */
else if (DEBUG_MARKER_INSN_P (next))
return true;
else if (INSN_VAR_LOCATION_DECL (insn) == INSN_VAR_LOCATION_DECL
(next))
return false;
return true;
that's ending up doing quadratic work that would have been better done
by keeping track of some state during the backwards walk of the insn
stream.
[Bug debug/109237] csmith: another timeout with -g -O3 this time
https://gcc.gnu.org/bugzilla/show_bug.cgi?id=109237 Richard Biener changed: What|Removed |Added Last reconfirmed||2023-03-22 Status|UNCONFIRMED |NEW Ever confirmed|0 |1 CC||rguenth at gcc dot gnu.org --- Comment #5 from Richard Biener --- Confirmed. Maybe the only suspicious thing is trivially dead code: 10.65 ( 13%) and tree CFG cleanup : 5.06 ( 6%) cfg cleanup: 5.41 ( 7%) But yes, we're likely facing a lot of debug stmts, csmith tends to leave around tons of dead code.
[Bug debug/109237] csmith: another timeout with -g -O3 this time
https://gcc.gnu.org/bugzilla/show_bug.cgi?id=109237 Andrew Pinski changed: What|Removed |Added Component|middle-end |debug --- Comment #4 from Andrew Pinski --- To give an insight into how much the verifiers give: CFG verifier : 61.10 ( 11%) 3.97 ( 13%) 65.09 ( 11%) 0 ( 0%) ... tree SSA verifier : 39.90 ( 7%) 0.02 ( 0%) 39.93 ( 7%) 0 ( 0%) tree STMT verifier : 141.76 ( 26%) 0.00 ( 0%) 141.95 ( 25%) 0 ( 0%) ... verify RTL sharing : 82.20 ( 15%) 0.00 ( 0%) 82.25 ( 14%) 0 ( 0%) Which is almost due to the number of debug statements (RTL and gimple level). The only way to reduce the cost here is reduce the number of debug statements ...
