Message ID | 20160811140235.GA68714@kam.mff.cuni.cz |
---|---|
State | New |
Headers | show |
On Thu, 11 Aug 2016, Jan Hubicka wrote: > Hi, > this patch adds early jump threading pass. Jump threading is one of most > common cases where estimated profile becomes corrupted, because the branches > are predicted independently beforehand. This patch simply adds early mode to > jump threader that does not permit code growth and thus only win/win threads > are done before profile is constructed. > > Naturally this also improves IPA decisions because code size/time is estimated > more acurately. > > It is not very cool to add another pass and the jump threader is already > run 5 times. I think incrementally we could drop one of late threaders at least. > I tried to measure compile time effects but they are in wash. Moreover the patch > pays for itself in cc1plus code size: > > Before patch to tweak size estimates: 27779964 > Current mainline: 27748900 > With patch applied: 27716173 > > So despite adding few functions the code size effect is positive which I think > is quite nice. > > Given the fact that jump threading seems quite lightweit, I wonder why it is > guarded by flag_expensive_optimizations? Is it expensive in terms of code > size? > > The effectivity of individual threading passes is as follows (for tramp3d) > > mainline with patch > pass thread count profile mismatches thread count profile mismatch > early 525 > 1 1853 1900 316 337 > 2 4 812 4 288 > 3 24 1450 32 947 > 4 76 1457 75 975 > > So at least tramp3d data seems to suggest that we can drop the second occurence > of jump threading and that most of the job thread1 does can be done by the > size restricted early version (the lower absolute counts are caused by the > fact that threadable paths gets duplicated by the inliner). 50% drop in > profile distortion is not bad. I wonder why basically every threaded paths seems > to introduce a mismatch. > > The patch distorts testusite somewhat, in most cases we only want to update > dump files or disable early threading: > > +XPASS: gcc.dg/uninit-15.c (test for warnings, line 13) > +XPASS: gcc.dg/uninit-15.c (test for warnings, line 23) > +FAIL: gcc.dg/uninit-15.c (test for warnings, line 24) > +FAIL: gcc.dg/tree-ssa/pr68198.c scan-tree-dump-times thread1 "Registering FSM" 1 > +FAIL: gcc.dg/tree-ssa/pr69196-1.c scan-tree-dump thread1 "FSM did not thread around loop and would copy too many statements" > +FAIL: gcc.dg/tree-ssa/ssa-dom-thread-2b.c scan-tree-dump-times thread1 "Jumps threaded: 1" 1 > +FAIL: gcc.dg/tree-ssa/ssa-thread-13.c scan-tree-dump thread1 "FSM" > +FAIL: gcc.dg/tree-ssa/vrp01.c scan-tree-dump-times vrp1 "Folding predicate p_.*to 1" 1 > +FAIL: gcc.dg/tree-ssa/vrp56.c scan-tree-dump thread1 "Jumps threaded: 1" > +FAIL: gcc.dg/tree-ssa/vrp92.c scan-tree-dump vrp1 "res_.: \\\\[1, 1\\\\]" > > This testcase is the now probably unnecesary heuristics (it may still be > relevant in cases we do not thread because of size bounds but its effectivity > may not be worth the maintenance cost): > > +FAIL: g++.dg/predict-loop-exit-1.C -std=gnu++11 scan-tree-dump-times profile_estimate "extra loop exit heuristics of edge[^:]*:" 2 > +FAIL: g++.dg/predict-loop-exit-1.C -std=gnu++11 scan-tree-dump-times profile_estimate "loop exit heuristics of edge[^:]*:" 3 > +FAIL: g++.dg/predict-loop-exit-1.C -std=gnu++14 scan-tree-dump-times profile_estimate "extra loop exit heuristics of edge[^:]*:" 2 > +FAIL: g++.dg/predict-loop-exit-1.C -std=gnu++14 scan-tree-dump-times profile_estimate "loop exit heuristics of edge[^:]*:" 3 > +FAIL: g++.dg/predict-loop-exit-1.C -std=gnu++98 scan-tree-dump-times profile_estimate "extra loop exit heuristics of edge[^:]*:" 2 > +FAIL: g++.dg/predict-loop-exit-1.C -std=gnu++98 scan-tree-dump-times profile_estimate "loop exit heuristics of edge[^:]*:" 3 > +FAIL: g++.dg/predict-loop-exit-2.C -std=gnu++11 scan-tree-dump-times profile_estimate "extra loop exit heuristics of edge[^:]*:" 1 > +FAIL: g++.dg/predict-loop-exit-2.C -std=gnu++11 scan-tree-dump-times profile_estimate "loop exit heuristics of edge[^:]*:" 2 > +FAIL: g++.dg/predict-loop-exit-2.C -std=gnu++14 scan-tree-dump-times profile_estimate "extra loop exit heuristics of edge[^:]*:" 1 > +FAIL: g++.dg/predict-loop-exit-2.C -std=gnu++14 scan-tree-dump-times profile_estimate "loop exit heuristics of edge[^:]*:" 2 > +FAIL: g++.dg/predict-loop-exit-2.C -std=gnu++98 scan-tree-dump-times profile_estimate "extra loop exit heuristics of edge[^:]*:" 1 > +FAIL: g++.dg/predict-loop-exit-2.C -std=gnu++98 scan-tree-dump-times profile_estimate "loop exit heuristics of edge[^:]*:" 2 > +FAIL: g++.dg/predict-loop-exit-3.C -std=gnu++11 scan-tree-dump-times profile_estimate "extra loop exit heuristics of edge[^:]*:" 2 > +FAIL: g++.dg/predict-loop-exit-3.C -std=gnu++11 scan-tree-dump-times profile_estimate "loop exit heuristics of edge[^:]*:" 3 > +FAIL: g++.dg/predict-loop-exit-3.C -std=gnu++14 scan-tree-dump-times profile_estimate "extra loop exit heuristics of edge[^:]*:" 2 > +FAIL: g++.dg/predict-loop-exit-3.C -std=gnu++14 scan-tree-dump-times profile_estimate "loop exit heuristics of edge[^:]*:" 3 > +FAIL: g++.dg/predict-loop-exit-3.C -std=gnu++98 scan-tree-dump-times profile_estimate "extra loop exit heuristics of edge[^:]*:" 2 > +FAIL: g++.dg/predict-loop-exit-3.C -std=gnu++98 scan-tree-dump-times profile_estimate "loop exit heuristics of edge[^:]*:" 3 > > If the patch seems acceptable, I will do the updates. One option why I did > not do that is that it seems to be now posisble to pass parameters to passes > from passes.def, so perhaps we do not need early_thread_jumps, but doing so is > consistent with way we handle other early passes. I wonder why you choose to put the FSM threader early which only does backward threading(?!). I'd expect forward threading to be more profitable (though we don't have a separate threader for that and would need VRP or DOM - but it seems we will get an early VRP anyway). Richard. > Bootstrapped/regtested x86_64-linux > Honza > > * passes.def (pass_early_thread_jumps): Schedule after forwprop. > * tree-pass.h (make_pass_early_thread_jumps): Declare. > * tree-ssa-threadbackward.c (fsm_find_thread_path, > fsm_find_thread_path, profitable_jump_thread_path, > fsm_find_control_statement_thread_paths, > find_jump_threads_backwards): Add speed_p parameter. > (pass_data_early_thread_jumps): New pass. > (make_pass_early_thread_jumps): New function. > > Index: passes.def > =================================================================== > --- passes.def (revision 239218) > +++ passes.def (working copy) > @@ -84,6 +84,7 @@ along with GCC; see the file COPYING3. > /* After CCP we rewrite no longer addressed locals into SSA > form if possible. */ > NEXT_PASS (pass_forwprop); > + NEXT_PASS (pass_early_thread_jumps); > NEXT_PASS (pass_sra_early); > /* pass_build_ealias is a dummy pass that ensures that we > execute TODO_rebuild_alias at this point. */ > Index: tree-pass.h > =================================================================== > --- tree-pass.h (revision 239218) > +++ tree-pass.h (working copy) > @@ -399,6 +399,7 @@ extern gimple_opt_pass *make_pass_cd_dce > extern gimple_opt_pass *make_pass_call_cdce (gcc::context *ctxt); > extern gimple_opt_pass *make_pass_merge_phi (gcc::context *ctxt); > extern gimple_opt_pass *make_pass_thread_jumps (gcc::context *ctxt); > +extern gimple_opt_pass *make_pass_early_thread_jumps (gcc::context *ctxt); > extern gimple_opt_pass *make_pass_split_crit_edges (gcc::context *ctxt); > extern gimple_opt_pass *make_pass_laddress (gcc::context *ctxt); > extern gimple_opt_pass *make_pass_pre (gcc::context *ctxt); > Index: tree-ssa-threadbackward.c > =================================================================== > --- tree-ssa-threadbackward.c (revision 239219) > +++ tree-ssa-threadbackward.c (working copy) > @@ -61,12 +61,14 @@ get_gimple_control_stmt (basic_block bb) > /* Return true if the CFG contains at least one path from START_BB to END_BB. > When a path is found, record in PATH the blocks from END_BB to START_BB. > VISITED_BBS is used to make sure we don't fall into an infinite loop. Bound > - the recursion to basic blocks belonging to LOOP. */ > + the recursion to basic blocks belonging to LOOP. > + SPEED_P indicate that we could increase code size to improve the code path */ > > static bool > fsm_find_thread_path (basic_block start_bb, basic_block end_bb, > vec<basic_block, va_gc> *&path, > - hash_set<basic_block> *visited_bbs, loop_p loop) > + hash_set<basic_block> *visited_bbs, loop_p loop, > + bool speed_p) > { > if (loop != start_bb->loop_father) > return false; > @@ -82,7 +84,8 @@ fsm_find_thread_path (basic_block start_ > edge e; > edge_iterator ei; > FOR_EACH_EDGE (e, ei, start_bb->succs) > - if (fsm_find_thread_path (e->dest, end_bb, path, visited_bbs, loop)) > + if (fsm_find_thread_path (e->dest, end_bb, path, visited_bbs, loop, > + speed_p)) > { > vec_safe_push (path, start_bb); > return true; > @@ -101,11 +104,13 @@ fsm_find_thread_path (basic_block start_ > value on PATH. ARG is the value of that SSA_NAME. > > BBI will be appended to PATH when we have a profitable jump threading > - path. Callers are responsible for removing BBI from PATH in that case. */ > + path. Callers are responsible for removing BBI from PATH in that case. > + > + SPEED_P indicate that we could increase code size to improve the code path */ > > static edge > profitable_jump_thread_path (vec<basic_block, va_gc> *&path, > - basic_block bbi, tree name, tree arg) > + basic_block bbi, tree name, tree arg, bool speed_p) > { > /* Note BBI is not in the path yet, hence the +1 in the test below > to make sure BBI is accounted for in the path length test. */ > @@ -306,7 +311,7 @@ profitable_jump_thread_path (vec<basic_b > return NULL; > } > > - if (optimize_edge_for_speed_p (taken_edge)) > + if (speed_p && optimize_edge_for_speed_p (taken_edge)) > { > if (n_insns >= PARAM_VALUE (PARAM_MAX_FSM_THREAD_PATH_INSNS)) > { > @@ -421,13 +426,15 @@ convert_and_register_jump_thread_path (v > > /* We trace the value of the SSA_NAME NAME back through any phi nodes looking > for places where it gets a constant value and save the path. Stop after > - having recorded MAX_PATHS jump threading paths. */ > + having recorded MAX_PATHS jump threading paths. > + > + SPEED_P indicate that we could increase code size to improve the code path */ > > static void > fsm_find_control_statement_thread_paths (tree name, > hash_set<basic_block> *visited_bbs, > vec<basic_block, va_gc> *&path, > - bool seen_loop_phi) > + bool seen_loop_phi, bool speed_p) > { > /* If NAME appears in an abnormal PHI, then don't try to trace its > value back through PHI nodes. */ > @@ -495,7 +502,7 @@ fsm_find_control_statement_thread_paths > hash_set<basic_block> *visited_bbs = new hash_set<basic_block>; > > if (fsm_find_thread_path (var_bb, e->src, next_path, visited_bbs, > - e->src->loop_father)) > + e->src->loop_father, speed_p)) > ++e_count; > > delete visited_bbs; > @@ -561,7 +568,7 @@ fsm_find_control_statement_thread_paths > /* Recursively follow SSA_NAMEs looking for a constant > definition. */ > fsm_find_control_statement_thread_paths (arg, visited_bbs, path, > - seen_loop_phi); > + seen_loop_phi, speed_p); > > path->pop (); > continue; > @@ -572,7 +579,8 @@ fsm_find_control_statement_thread_paths > > /* If this is a profitable jump thread path, then convert it > into the canonical form and register it. */ > - edge taken_edge = profitable_jump_thread_path (path, bbi, name, arg); > + edge taken_edge = profitable_jump_thread_path (path, bbi, name, arg, > + speed_p); > if (taken_edge) > { > if (bb_loop_depth (taken_edge->src) > @@ -588,7 +596,7 @@ fsm_find_control_statement_thread_paths > > if (TREE_CODE (arg) == SSA_NAME) > fsm_find_control_statement_thread_paths (arg, visited_bbs, > - path, seen_loop_phi); > + path, seen_loop_phi, speed_p); > > else > { > @@ -598,7 +606,7 @@ fsm_find_control_statement_thread_paths > path->pop (); > > edge taken_edge = profitable_jump_thread_path (path, var_bb, > - name, arg); > + name, arg, speed_p); > if (taken_edge) > { > if (bb_loop_depth (taken_edge->src) > @@ -622,10 +630,11 @@ fsm_find_control_statement_thread_paths > is a constant. Record such paths for jump threading. > > It is assumed that BB ends with a control statement and that by > - finding a path where NAME is a constant, we can thread the path. */ > + finding a path where NAME is a constant, we can thread the path. > + SPEED_P indicate that we could increase code size to improve the code path */ > > void > -find_jump_threads_backwards (basic_block bb) > +find_jump_threads_backwards (basic_block bb, bool speed_p) > { > gimple *stmt = get_gimple_control_stmt (bb); > if (!stmt) > @@ -655,7 +664,8 @@ find_jump_threads_backwards (basic_block > hash_set<basic_block> *visited_bbs = new hash_set<basic_block>; > > max_threaded_paths = PARAM_VALUE (PARAM_MAX_FSM_THREAD_PATHS); > - fsm_find_control_statement_thread_paths (name, visited_bbs, bb_path, false); > + fsm_find_control_statement_thread_paths (name, visited_bbs, bb_path, false, > + speed_p); > > delete visited_bbs; > vec_free (bb_path); > @@ -703,7 +713,7 @@ pass_thread_jumps::execute (function *fu > FOR_EACH_BB_FN (bb, fun) > { > if (EDGE_COUNT (bb->succs) > 1) > - find_jump_threads_backwards (bb); > + find_jump_threads_backwards (bb, true); > } > thread_through_all_blocks (true); > return 0; > @@ -716,3 +726,59 @@ make_pass_thread_jumps (gcc::context *ct > { > return new pass_thread_jumps (ctxt); > } > + > +namespace { > + > +const pass_data pass_data_early_thread_jumps = > +{ > + GIMPLE_PASS, > + "early_thread", > + OPTGROUP_NONE, > + TV_TREE_SSA_THREAD_JUMPS, > + ( PROP_cfg | PROP_ssa ), > + 0, > + 0, > + 0, > + ( TODO_cleanup_cfg | TODO_update_ssa ), > +}; > + > +class pass_early_thread_jumps : public gimple_opt_pass > +{ > +public: > + pass_early_thread_jumps (gcc::context *ctxt) > + : gimple_opt_pass (pass_data_early_thread_jumps, ctxt) > + {} > + > + opt_pass * clone (void) { return new pass_early_thread_jumps (m_ctxt); } > + virtual bool gate (function *); > + virtual unsigned int execute (function *); > +}; > + > +bool > +pass_early_thread_jumps::gate (function *fun ATTRIBUTE_UNUSED) > +{ > + return true; > +} > + > + > +unsigned int > +pass_early_thread_jumps::execute (function *fun) > +{ > + /* Try to thread each block with more than one successor. */ > + basic_block bb; > + FOR_EACH_BB_FN (bb, fun) > + { > + if (EDGE_COUNT (bb->succs) > 1) > + find_jump_threads_backwards (bb, false); > + } > + thread_through_all_blocks (true); > + return 0; > +} > + > +} > + > +gimple_opt_pass * > +make_pass_early_thread_jumps (gcc::context *ctxt) > +{ > + return new pass_early_thread_jumps (ctxt); > +} > >
> On Thu, 11 Aug 2016, Jan Hubicka wrote: > > > Hi, > > this patch adds early jump threading pass. Jump threading is one of most > > common cases where estimated profile becomes corrupted, because the branches > > are predicted independently beforehand. This patch simply adds early mode to > > jump threader that does not permit code growth and thus only win/win threads > > are done before profile is constructed. > > > > Naturally this also improves IPA decisions because code size/time is estimated > > more acurately. > > > > It is not very cool to add another pass and the jump threader is already > > run 5 times. I think incrementally we could drop one of late threaders at least. > > I tried to measure compile time effects but they are in wash. Moreover the patch > > pays for itself in cc1plus code size: > > > > Before patch to tweak size estimates: 27779964 > > Current mainline: 27748900 > > With patch applied: 27716173 > > > > So despite adding few functions the code size effect is positive which I think > > is quite nice. > > > > Given the fact that jump threading seems quite lightweit, I wonder why it is > > guarded by flag_expensive_optimizations? Is it expensive in terms of code > > size? > > > > The effectivity of individual threading passes is as follows (for tramp3d) > > > > mainline with patch > > pass thread count profile mismatches thread count profile mismatch > > early 525 > > 1 1853 1900 316 337 > > 2 4 812 4 288 > > 3 24 1450 32 947 > > 4 76 1457 75 975 > > > > So at least tramp3d data seems to suggest that we can drop the second occurence > > of jump threading and that most of the job thread1 does can be done by the > > size restricted early version (the lower absolute counts are caused by the > > fact that threadable paths gets duplicated by the inliner). 50% drop in > > profile distortion is not bad. I wonder why basically every threaded paths seems > > to introduce a mismatch. > > > > The patch distorts testusite somewhat, in most cases we only want to update > > dump files or disable early threading: > > > > +XPASS: gcc.dg/uninit-15.c (test for warnings, line 13) > > +XPASS: gcc.dg/uninit-15.c (test for warnings, line 23) > > +FAIL: gcc.dg/uninit-15.c (test for warnings, line 24) > > +FAIL: gcc.dg/tree-ssa/pr68198.c scan-tree-dump-times thread1 "Registering FSM" 1 > > +FAIL: gcc.dg/tree-ssa/pr69196-1.c scan-tree-dump thread1 "FSM did not thread around loop and would copy too many statements" > > +FAIL: gcc.dg/tree-ssa/ssa-dom-thread-2b.c scan-tree-dump-times thread1 "Jumps threaded: 1" 1 > > +FAIL: gcc.dg/tree-ssa/ssa-thread-13.c scan-tree-dump thread1 "FSM" > > +FAIL: gcc.dg/tree-ssa/vrp01.c scan-tree-dump-times vrp1 "Folding predicate p_.*to 1" 1 > > +FAIL: gcc.dg/tree-ssa/vrp56.c scan-tree-dump thread1 "Jumps threaded: 1" > > +FAIL: gcc.dg/tree-ssa/vrp92.c scan-tree-dump vrp1 "res_.: \\\\[1, 1\\\\]" > > > > This testcase is the now probably unnecesary heuristics (it may still be > > relevant in cases we do not thread because of size bounds but its effectivity > > may not be worth the maintenance cost): > > > > +FAIL: g++.dg/predict-loop-exit-1.C -std=gnu++11 scan-tree-dump-times profile_estimate "extra loop exit heuristics of edge[^:]*:" 2 > > +FAIL: g++.dg/predict-loop-exit-1.C -std=gnu++11 scan-tree-dump-times profile_estimate "loop exit heuristics of edge[^:]*:" 3 > > +FAIL: g++.dg/predict-loop-exit-1.C -std=gnu++14 scan-tree-dump-times profile_estimate "extra loop exit heuristics of edge[^:]*:" 2 > > +FAIL: g++.dg/predict-loop-exit-1.C -std=gnu++14 scan-tree-dump-times profile_estimate "loop exit heuristics of edge[^:]*:" 3 > > +FAIL: g++.dg/predict-loop-exit-1.C -std=gnu++98 scan-tree-dump-times profile_estimate "extra loop exit heuristics of edge[^:]*:" 2 > > +FAIL: g++.dg/predict-loop-exit-1.C -std=gnu++98 scan-tree-dump-times profile_estimate "loop exit heuristics of edge[^:]*:" 3 > > +FAIL: g++.dg/predict-loop-exit-2.C -std=gnu++11 scan-tree-dump-times profile_estimate "extra loop exit heuristics of edge[^:]*:" 1 > > +FAIL: g++.dg/predict-loop-exit-2.C -std=gnu++11 scan-tree-dump-times profile_estimate "loop exit heuristics of edge[^:]*:" 2 > > +FAIL: g++.dg/predict-loop-exit-2.C -std=gnu++14 scan-tree-dump-times profile_estimate "extra loop exit heuristics of edge[^:]*:" 1 > > +FAIL: g++.dg/predict-loop-exit-2.C -std=gnu++14 scan-tree-dump-times profile_estimate "loop exit heuristics of edge[^:]*:" 2 > > +FAIL: g++.dg/predict-loop-exit-2.C -std=gnu++98 scan-tree-dump-times profile_estimate "extra loop exit heuristics of edge[^:]*:" 1 > > +FAIL: g++.dg/predict-loop-exit-2.C -std=gnu++98 scan-tree-dump-times profile_estimate "loop exit heuristics of edge[^:]*:" 2 > > +FAIL: g++.dg/predict-loop-exit-3.C -std=gnu++11 scan-tree-dump-times profile_estimate "extra loop exit heuristics of edge[^:]*:" 2 > > +FAIL: g++.dg/predict-loop-exit-3.C -std=gnu++11 scan-tree-dump-times profile_estimate "loop exit heuristics of edge[^:]*:" 3 > > +FAIL: g++.dg/predict-loop-exit-3.C -std=gnu++14 scan-tree-dump-times profile_estimate "extra loop exit heuristics of edge[^:]*:" 2 > > +FAIL: g++.dg/predict-loop-exit-3.C -std=gnu++14 scan-tree-dump-times profile_estimate "loop exit heuristics of edge[^:]*:" 3 > > +FAIL: g++.dg/predict-loop-exit-3.C -std=gnu++98 scan-tree-dump-times profile_estimate "extra loop exit heuristics of edge[^:]*:" 2 > > +FAIL: g++.dg/predict-loop-exit-3.C -std=gnu++98 scan-tree-dump-times profile_estimate "loop exit heuristics of edge[^:]*:" 3 > > > > If the patch seems acceptable, I will do the updates. One option why I did > > not do that is that it seems to be now posisble to pass parameters to passes > > from passes.def, so perhaps we do not need early_thread_jumps, but doing so is > > consistent with way we handle other early passes. > > I wonder why you choose to put the FSM threader early which only does > backward threading(?!). I'd expect forward threading to be more > profitable (though we don't have a separate threader for that and > would need VRP or DOM - but it seems we will get an early VRP anyway). On tramp3d all VRP passes threads together 730 branches, all DOM passes 393, so FSM threading (with 1957 branches) is the most effective one. Perhaps eventually early VRP can also do bit of work. I am not 100% sure from where "backward" is comming from. I guess is means that analysis goes backward from conditionals to definitions: it looks for conditional driven by a PHI statement that has a constant value on some paths and duplicates for those. This seems cheap and rather effective way of getting good part of the threading oppurtunities (most expensive part is probably identifying and walking paths that will not be threaded at the end). BTW I wonder if the same analysis can't be done for other instructions where constant operands are very profitable, like division or multiplication. Honza
On August 11, 2016 4:27:00 PM GMT+02:00, Jan Hubicka <hubicka@ucw.cz> wrote: >> On Thu, 11 Aug 2016, Jan Hubicka wrote: >> >> > Hi, >> > this patch adds early jump threading pass. Jump threading is one >of most >> > common cases where estimated profile becomes corrupted, because the >branches >> > are predicted independently beforehand. This patch simply adds >early mode to >> > jump threader that does not permit code growth and thus only >win/win threads >> > are done before profile is constructed. >> > >> > Naturally this also improves IPA decisions because code size/time >is estimated >> > more acurately. >> > >> > It is not very cool to add another pass and the jump threader is >already >> > run 5 times. I think incrementally we could drop one of late >threaders at least. >> > I tried to measure compile time effects but they are in wash. >Moreover the patch >> > pays for itself in cc1plus code size: >> > >> > Before patch to tweak size estimates: 27779964 >> > Current mainline: 27748900 >> > With patch applied: 27716173 >> > >> > So despite adding few functions the code size effect is positive >which I think >> > is quite nice. >> > >> > Given the fact that jump threading seems quite lightweit, I wonder >why it is >> > guarded by flag_expensive_optimizations? Is it expensive in terms >of code >> > size? >> > >> > The effectivity of individual threading passes is as follows (for >tramp3d) >> > >> > mainline with patch >> > pass thread count profile mismatches thread count profile >mismatch >> > early 525 >> > 1 1853 1900 316 337 >> > 2 4 812 4 288 >> > 3 24 1450 32 947 >> > 4 76 1457 75 975 >> > >> > So at least tramp3d data seems to suggest that we can drop the >second occurence >> > of jump threading and that most of the job thread1 does can be done >by the >> > size restricted early version (the lower absolute counts are caused >by the >> > fact that threadable paths gets duplicated by the inliner). 50% >drop in >> > profile distortion is not bad. I wonder why basically every >threaded paths seems >> > to introduce a mismatch. >> > >> > The patch distorts testusite somewhat, in most cases we only want >to update >> > dump files or disable early threading: >> > >> > +XPASS: gcc.dg/uninit-15.c (test for warnings, line 13) >> > +XPASS: gcc.dg/uninit-15.c (test for warnings, line 23) >> > +FAIL: gcc.dg/uninit-15.c (test for warnings, line 24) >> > +FAIL: gcc.dg/tree-ssa/pr68198.c scan-tree-dump-times thread1 >"Registering FSM" 1 >> > +FAIL: gcc.dg/tree-ssa/pr69196-1.c scan-tree-dump thread1 "FSM did >not thread around loop and would copy too many statements" >> > +FAIL: gcc.dg/tree-ssa/ssa-dom-thread-2b.c scan-tree-dump-times >thread1 "Jumps threaded: 1" 1 >> > +FAIL: gcc.dg/tree-ssa/ssa-thread-13.c scan-tree-dump thread1 "FSM" >> > +FAIL: gcc.dg/tree-ssa/vrp01.c scan-tree-dump-times vrp1 "Folding >predicate p_.*to 1" 1 >> > +FAIL: gcc.dg/tree-ssa/vrp56.c scan-tree-dump thread1 "Jumps >threaded: 1" >> > +FAIL: gcc.dg/tree-ssa/vrp92.c scan-tree-dump vrp1 "res_.: \\\\[1, >1\\\\]" >> > >> > This testcase is the now probably unnecesary heuristics (it may >still be >> > relevant in cases we do not thread because of size bounds but its >effectivity >> > may not be worth the maintenance cost): >> > >> > +FAIL: g++.dg/predict-loop-exit-1.C -std=gnu++11 >scan-tree-dump-times profile_estimate "extra loop exit heuristics of >edge[^:]*:" 2 >> > +FAIL: g++.dg/predict-loop-exit-1.C -std=gnu++11 >scan-tree-dump-times profile_estimate "loop exit heuristics of >edge[^:]*:" 3 >> > +FAIL: g++.dg/predict-loop-exit-1.C -std=gnu++14 >scan-tree-dump-times profile_estimate "extra loop exit heuristics of >edge[^:]*:" 2 >> > +FAIL: g++.dg/predict-loop-exit-1.C -std=gnu++14 >scan-tree-dump-times profile_estimate "loop exit heuristics of >edge[^:]*:" 3 >> > +FAIL: g++.dg/predict-loop-exit-1.C -std=gnu++98 >scan-tree-dump-times profile_estimate "extra loop exit heuristics of >edge[^:]*:" 2 >> > +FAIL: g++.dg/predict-loop-exit-1.C -std=gnu++98 >scan-tree-dump-times profile_estimate "loop exit heuristics of >edge[^:]*:" 3 >> > +FAIL: g++.dg/predict-loop-exit-2.C -std=gnu++11 >scan-tree-dump-times profile_estimate "extra loop exit heuristics of >edge[^:]*:" 1 >> > +FAIL: g++.dg/predict-loop-exit-2.C -std=gnu++11 >scan-tree-dump-times profile_estimate "loop exit heuristics of >edge[^:]*:" 2 >> > +FAIL: g++.dg/predict-loop-exit-2.C -std=gnu++14 >scan-tree-dump-times profile_estimate "extra loop exit heuristics of >edge[^:]*:" 1 >> > +FAIL: g++.dg/predict-loop-exit-2.C -std=gnu++14 >scan-tree-dump-times profile_estimate "loop exit heuristics of >edge[^:]*:" 2 >> > +FAIL: g++.dg/predict-loop-exit-2.C -std=gnu++98 >scan-tree-dump-times profile_estimate "extra loop exit heuristics of >edge[^:]*:" 1 >> > +FAIL: g++.dg/predict-loop-exit-2.C -std=gnu++98 >scan-tree-dump-times profile_estimate "loop exit heuristics of >edge[^:]*:" 2 >> > +FAIL: g++.dg/predict-loop-exit-3.C -std=gnu++11 >scan-tree-dump-times profile_estimate "extra loop exit heuristics of >edge[^:]*:" 2 >> > +FAIL: g++.dg/predict-loop-exit-3.C -std=gnu++11 >scan-tree-dump-times profile_estimate "loop exit heuristics of >edge[^:]*:" 3 >> > +FAIL: g++.dg/predict-loop-exit-3.C -std=gnu++14 >scan-tree-dump-times profile_estimate "extra loop exit heuristics of >edge[^:]*:" 2 >> > +FAIL: g++.dg/predict-loop-exit-3.C -std=gnu++14 >scan-tree-dump-times profile_estimate "loop exit heuristics of >edge[^:]*:" 3 >> > +FAIL: g++.dg/predict-loop-exit-3.C -std=gnu++98 >scan-tree-dump-times profile_estimate "extra loop exit heuristics of >edge[^:]*:" 2 >> > +FAIL: g++.dg/predict-loop-exit-3.C -std=gnu++98 >scan-tree-dump-times profile_estimate "loop exit heuristics of >edge[^:]*:" 3 >> > >> > If the patch seems acceptable, I will do the updates. One option >why I did >> > not do that is that it seems to be now posisble to pass parameters >to passes >> > from passes.def, so perhaps we do not need early_thread_jumps, but >doing so is >> > consistent with way we handle other early passes. >> >> I wonder why you choose to put the FSM threader early which only does >> backward threading(?!). I'd expect forward threading to be more >> profitable (though we don't have a separate threader for that and >> would need VRP or DOM - but it seems we will get an early VRP >anyway). > >On tramp3d all VRP passes threads together 730 branches, all DOM passes >393, so >FSM threading (with 1957 branches) is the most effective one. Perhaps >eventually >early VRP can also do bit of work. > >I am not 100% sure from where "backward" is comming from. I guess is >means that >analysis goes backward from conditionals to definitions: it looks for >conditional driven by a PHI statement that has a constant value on some >paths >and duplicates for those. This seems cheap and rather effective way of >getting >good part of the threading oppurtunities (most expensive part is >probably >identifying and walking paths that will not be threaded at the end). Ah, I thought it was exclusively dealing with threading through back edges which is sth I'd avoid doing early? >BTW I wonder if the same analysis can't be done for other instructions >where constant >operands are very profitable, like division or multiplication. No idea, but Jeff will likely know. Richard. > >Honza
On 08/11/2016 08:02 AM, Jan Hubicka wrote: > Hi, > this patch adds early jump threading pass. Jump threading is one of most > common cases where estimated profile becomes corrupted, because the branches > are predicted independently beforehand. This patch simply adds early mode to > jump threader that does not permit code growth and thus only win/win threads > are done before profile is constructed. > > Naturally this also improves IPA decisions because code size/time is estimated > more acurately. Excellent. One of the goals here was to enable you to run it early, so I'm glad to see it's working out. > > It is not very cool to add another pass and the jump threader is already > run 5 times. I think incrementally we could drop one of late threaders at least. > I tried to measure compile time effects but they are in wash. Moreover the patch > pays for itself in cc1plus code size: Most definitely we want to be dropping calls into the later passes. There's analysis to do, but the goal is to drop the old style threading from DOM/VRP completely. It may also be the case that one or more passes of the backwards/FSM threader can be avoided, but we should look at removing the DOM/VRP threading first. > > Before patch to tweak size estimates: 27779964 > Current mainline: 27748900 > With patch applied: 27716173 > > So despite adding few functions the code size effect is positive which I think > is quite nice. > > Given the fact that jump threading seems quite lightweit, I wonder why it is > guarded by flag_expensive_optimizations? Is it expensive in terms of code > size? The DOM/VRP jump threading used to be very expensive because of iteration and the ssa updates. Both of those issues have since been addressed. The backwards/FSM threader is based on an algorithm that IIRC is cubic, and we also have some implementation issues that cause it to use far more time than it should. However, in an early mode, we don't want to walk backwards through the CFG very far (if at all) and for an early mode we may not want to guard on flag_expensive_optimizations. > > The effectivity of individual threading passes is as follows (for tramp3d) > > mainline with patch > pass thread count profile mismatches thread count profile mismatch > early 525 > 1 1853 1900 316 337 > 2 4 812 4 288 > 3 24 1450 32 947 > 4 76 1457 75 975 > > So at least tramp3d data seems to suggest that we can drop the second occurence > of jump threading and that most of the job thread1 does can be done by the > size restricted early version (the lower absolute counts are caused by the > fact that threadable paths gets duplicated by the inliner). 50% drop in > profile distortion is not bad. I wonder why basically every threaded paths seems > to introduce a mismatch. I don't think the backwards/FSM threader tries to update the profile data at all right now. > > The patch distorts testusite somewhat, in most cases we only want to update > dump files or disable early threading: > > +XPASS: gcc.dg/uninit-15.c (test for warnings, line 13) > +XPASS: gcc.dg/uninit-15.c (test for warnings, line 23) > +FAIL: gcc.dg/uninit-15.c (test for warnings, line 24) > +FAIL: gcc.dg/tree-ssa/pr68198.c scan-tree-dump-times thread1 "Registering FSM" 1 > +FAIL: gcc.dg/tree-ssa/pr69196-1.c scan-tree-dump thread1 "FSM did not thread around loop and would copy too many statements" > +FAIL: gcc.dg/tree-ssa/ssa-dom-thread-2b.c scan-tree-dump-times thread1 "Jumps threaded: 1" 1 > +FAIL: gcc.dg/tree-ssa/ssa-thread-13.c scan-tree-dump thread1 "FSM" > +FAIL: gcc.dg/tree-ssa/vrp01.c scan-tree-dump-times vrp1 "Folding predicate p_.*to 1" 1 > +FAIL: gcc.dg/tree-ssa/vrp56.c scan-tree-dump thread1 "Jumps threaded: 1" > +FAIL: gcc.dg/tree-ssa/vrp92.c scan-tree-dump vrp1 "res_.: \\\\[1, 1\\\\]" > > This testcase is the now probably unnecesary heuristics (it may still be > relevant in cases we do not thread because of size bounds but its effectivity > may not be worth the maintenance cost): > > +FAIL: g++.dg/predict-loop-exit-1.C -std=gnu++11 scan-tree-dump-times profile_estimate "extra loop exit heuristics of edge[^:]*:" 2 > +FAIL: g++.dg/predict-loop-exit-1.C -std=gnu++11 scan-tree-dump-times profile_estimate "loop exit heuristics of edge[^:]*:" 3 > +FAIL: g++.dg/predict-loop-exit-1.C -std=gnu++14 scan-tree-dump-times profile_estimate "extra loop exit heuristics of edge[^:]*:" 2 > +FAIL: g++.dg/predict-loop-exit-1.C -std=gnu++14 scan-tree-dump-times profile_estimate "loop exit heuristics of edge[^:]*:" 3 > +FAIL: g++.dg/predict-loop-exit-1.C -std=gnu++98 scan-tree-dump-times profile_estimate "extra loop exit heuristics of edge[^:]*:" 2 > +FAIL: g++.dg/predict-loop-exit-1.C -std=gnu++98 scan-tree-dump-times profile_estimate "loop exit heuristics of edge[^:]*:" 3 > +FAIL: g++.dg/predict-loop-exit-2.C -std=gnu++11 scan-tree-dump-times profile_estimate "extra loop exit heuristics of edge[^:]*:" 1 > +FAIL: g++.dg/predict-loop-exit-2.C -std=gnu++11 scan-tree-dump-times profile_estimate "loop exit heuristics of edge[^:]*:" 2 > +FAIL: g++.dg/predict-loop-exit-2.C -std=gnu++14 scan-tree-dump-times profile_estimate "extra loop exit heuristics of edge[^:]*:" 1 > +FAIL: g++.dg/predict-loop-exit-2.C -std=gnu++14 scan-tree-dump-times profile_estimate "loop exit heuristics of edge[^:]*:" 2 > +FAIL: g++.dg/predict-loop-exit-2.C -std=gnu++98 scan-tree-dump-times profile_estimate "extra loop exit heuristics of edge[^:]*:" 1 > +FAIL: g++.dg/predict-loop-exit-2.C -std=gnu++98 scan-tree-dump-times profile_estimate "loop exit heuristics of edge[^:]*:" 2 > +FAIL: g++.dg/predict-loop-exit-3.C -std=gnu++11 scan-tree-dump-times profile_estimate "extra loop exit heuristics of edge[^:]*:" 2 > +FAIL: g++.dg/predict-loop-exit-3.C -std=gnu++11 scan-tree-dump-times profile_estimate "loop exit heuristics of edge[^:]*:" 3 > +FAIL: g++.dg/predict-loop-exit-3.C -std=gnu++14 scan-tree-dump-times profile_estimate "extra loop exit heuristics of edge[^:]*:" 2 > +FAIL: g++.dg/predict-loop-exit-3.C -std=gnu++14 scan-tree-dump-times profile_estimate "loop exit heuristics of edge[^:]*:" 3 > +FAIL: g++.dg/predict-loop-exit-3.C -std=gnu++98 scan-tree-dump-times profile_estimate "extra loop exit heuristics of edge[^:]*:" 2 > +FAIL: g++.dg/predict-loop-exit-3.C -std=gnu++98 scan-tree-dump-times profile_estimate "loop exit heuristics of edge[^:]*:" 3 > > If the patch seems acceptable, I will do the updates. One option why I did > not do that is that it seems to be now posisble to pass parameters to passes > from passes.def, so perhaps we do not need early_thread_jumps, but doing so is > consistent with way we handle other early passes. > > Bootstrapped/regtested x86_64-linux > Honza > > * passes.def (pass_early_thread_jumps): Schedule after forwprop. > * tree-pass.h (make_pass_early_thread_jumps): Declare. > * tree-ssa-threadbackward.c (fsm_find_thread_path, > fsm_find_thread_path, profitable_jump_thread_path, > fsm_find_control_statement_thread_paths, > find_jump_threads_backwards): Add speed_p parameter. > (pass_data_early_thread_jumps): New pass. > (make_pass_early_thread_jumps): New function. I'll try to take a look at the details shortly. jeff
On 08/11/2016 08:06 AM, Richard Biener wrote: > On Thu, 11 Aug 2016, Jan Hubicka wrote: > >> Hi, >> this patch adds early jump threading pass. Jump threading is one of most >> common cases where estimated profile becomes corrupted, because the branches >> are predicted independently beforehand. This patch simply adds early mode to >> jump threader that does not permit code growth and thus only win/win threads >> are done before profile is constructed. >> >> Naturally this also improves IPA decisions because code size/time is estimated >> more acurately. >> >> It is not very cool to add another pass and the jump threader is already >> run 5 times. I think incrementally we could drop one of late threaders at least. >> I tried to measure compile time effects but they are in wash. Moreover the patch >> pays for itself in cc1plus code size: >> >> Before patch to tweak size estimates: 27779964 >> Current mainline: 27748900 >> With patch applied: 27716173 >> >> So despite adding few functions the code size effect is positive which I think >> is quite nice. >> >> Given the fact that jump threading seems quite lightweit, I wonder why it is >> guarded by flag_expensive_optimizations? Is it expensive in terms of code >> size? >> >> The effectivity of individual threading passes is as follows (for tramp3d) >> >> mainline with patch >> pass thread count profile mismatches thread count profile mismatch >> early 525 >> 1 1853 1900 316 337 >> 2 4 812 4 288 >> 3 24 1450 32 947 >> 4 76 1457 75 975 >> >> So at least tramp3d data seems to suggest that we can drop the second occurence >> of jump threading and that most of the job thread1 does can be done by the >> size restricted early version (the lower absolute counts are caused by the >> fact that threadable paths gets duplicated by the inliner). 50% drop in >> profile distortion is not bad. I wonder why basically every threaded paths seems >> to introduce a mismatch. >> >> The patch distorts testusite somewhat, in most cases we only want to update >> dump files or disable early threading: >> >> +XPASS: gcc.dg/uninit-15.c (test for warnings, line 13) >> +XPASS: gcc.dg/uninit-15.c (test for warnings, line 23) >> +FAIL: gcc.dg/uninit-15.c (test for warnings, line 24) >> +FAIL: gcc.dg/tree-ssa/pr68198.c scan-tree-dump-times thread1 "Registering FSM" 1 >> +FAIL: gcc.dg/tree-ssa/pr69196-1.c scan-tree-dump thread1 "FSM did not thread around loop and would copy too many statements" >> +FAIL: gcc.dg/tree-ssa/ssa-dom-thread-2b.c scan-tree-dump-times thread1 "Jumps threaded: 1" 1 >> +FAIL: gcc.dg/tree-ssa/ssa-thread-13.c scan-tree-dump thread1 "FSM" >> +FAIL: gcc.dg/tree-ssa/vrp01.c scan-tree-dump-times vrp1 "Folding predicate p_.*to 1" 1 >> +FAIL: gcc.dg/tree-ssa/vrp56.c scan-tree-dump thread1 "Jumps threaded: 1" >> +FAIL: gcc.dg/tree-ssa/vrp92.c scan-tree-dump vrp1 "res_.: \\\\[1, 1\\\\]" >> >> This testcase is the now probably unnecesary heuristics (it may still be >> relevant in cases we do not thread because of size bounds but its effectivity >> may not be worth the maintenance cost): >> >> +FAIL: g++.dg/predict-loop-exit-1.C -std=gnu++11 scan-tree-dump-times profile_estimate "extra loop exit heuristics of edge[^:]*:" 2 >> +FAIL: g++.dg/predict-loop-exit-1.C -std=gnu++11 scan-tree-dump-times profile_estimate "loop exit heuristics of edge[^:]*:" 3 >> +FAIL: g++.dg/predict-loop-exit-1.C -std=gnu++14 scan-tree-dump-times profile_estimate "extra loop exit heuristics of edge[^:]*:" 2 >> +FAIL: g++.dg/predict-loop-exit-1.C -std=gnu++14 scan-tree-dump-times profile_estimate "loop exit heuristics of edge[^:]*:" 3 >> +FAIL: g++.dg/predict-loop-exit-1.C -std=gnu++98 scan-tree-dump-times profile_estimate "extra loop exit heuristics of edge[^:]*:" 2 >> +FAIL: g++.dg/predict-loop-exit-1.C -std=gnu++98 scan-tree-dump-times profile_estimate "loop exit heuristics of edge[^:]*:" 3 >> +FAIL: g++.dg/predict-loop-exit-2.C -std=gnu++11 scan-tree-dump-times profile_estimate "extra loop exit heuristics of edge[^:]*:" 1 >> +FAIL: g++.dg/predict-loop-exit-2.C -std=gnu++11 scan-tree-dump-times profile_estimate "loop exit heuristics of edge[^:]*:" 2 >> +FAIL: g++.dg/predict-loop-exit-2.C -std=gnu++14 scan-tree-dump-times profile_estimate "extra loop exit heuristics of edge[^:]*:" 1 >> +FAIL: g++.dg/predict-loop-exit-2.C -std=gnu++14 scan-tree-dump-times profile_estimate "loop exit heuristics of edge[^:]*:" 2 >> +FAIL: g++.dg/predict-loop-exit-2.C -std=gnu++98 scan-tree-dump-times profile_estimate "extra loop exit heuristics of edge[^:]*:" 1 >> +FAIL: g++.dg/predict-loop-exit-2.C -std=gnu++98 scan-tree-dump-times profile_estimate "loop exit heuristics of edge[^:]*:" 2 >> +FAIL: g++.dg/predict-loop-exit-3.C -std=gnu++11 scan-tree-dump-times profile_estimate "extra loop exit heuristics of edge[^:]*:" 2 >> +FAIL: g++.dg/predict-loop-exit-3.C -std=gnu++11 scan-tree-dump-times profile_estimate "loop exit heuristics of edge[^:]*:" 3 >> +FAIL: g++.dg/predict-loop-exit-3.C -std=gnu++14 scan-tree-dump-times profile_estimate "extra loop exit heuristics of edge[^:]*:" 2 >> +FAIL: g++.dg/predict-loop-exit-3.C -std=gnu++14 scan-tree-dump-times profile_estimate "loop exit heuristics of edge[^:]*:" 3 >> +FAIL: g++.dg/predict-loop-exit-3.C -std=gnu++98 scan-tree-dump-times profile_estimate "extra loop exit heuristics of edge[^:]*:" 2 >> +FAIL: g++.dg/predict-loop-exit-3.C -std=gnu++98 scan-tree-dump-times profile_estimate "loop exit heuristics of edge[^:]*:" 3 >> >> If the patch seems acceptable, I will do the updates. One option why I did >> not do that is that it seems to be now posisble to pass parameters to passes >> from passes.def, so perhaps we do not need early_thread_jumps, but doing so is >> consistent with way we handle other early passes. > > I wonder why you choose to put the FSM threader early which only does > backward threading(?!). I'd expect forward threading to be more > profitable (though we don't have a separate threader for that and > would need VRP or DOM - but it seems we will get an early VRP anyway). forward/backward refers to how they find threading opportunities. The backward/FSM threader walks back from a conditional through the CFG & PHI nodes. The old DOM/VRP threader utilizes state from forward walks during the CFG to try and simplify conditionals when they're encountered. Jeff
On 08/11/2016 08:27 AM, Jan Hubicka wrote: > > On tramp3d all VRP passes threads together 730 branches, all DOM passes 393, so > FSM threading (with 1957 branches) is the most effective one. Perhaps eventually > early VRP can also do bit of work. That's roughly consistent with what I've seen. I have some thoughts on what DOM's threading pass is catching that we're missing in the backward/FSM threader, but I haven't had time to see how big those effects really are. > > I am not 100% sure from where "backward" is comming from. I guess is means that > analysis goes backward from conditionals to definitions: it looks for > conditional driven by a PHI statement that has a constant value on some paths > and duplicates for those. This seems cheap and rather effective way of getting > good part of the threading oppurtunities (most expensive part is probably > identifying and walking paths that will not be threaded at the end). Correct, forward/backward is based on the direction of analysis. ie, do we start at the conditional and work backwards through the USE-DEF chain or do we build a equivalence table as we walk forward through the entire function and use the equivalence tables to simplify the conditional. > > BTW I wonder if the same analysis can't be done for other instructions where constant > operands are very profitable, like division or multiplication. Yes. There's all kinds of problems that can be solved using a backwards walk to identify useful properties on paths, then path duplication to isolate the use point and modify it. So you could (for example) walk backwards from an indirect call and identify paths through the CFG where we know the target. We then can use path duplication to isolate that path from the rest and call the target function directly. Jeff
On 08/11/2016 09:50 AM, Richard Biener wrote: > > Ah, I thought it was exclusively dealing with threading through back > edges which is sth I'd avoid doing early? No, that's one of the fundamental changes we made for gcc-6, namely using the FSM threader for general purpose threading rather than just using it for threading across loop backedges. I've got a few things to do first, but the direction I want to go is to use the FSM threader exclusively. Jeff
On August 11, 2016 8:41:53 PM GMT+02:00, Jeff Law <law@redhat.com> wrote: >On 08/11/2016 08:27 AM, Jan Hubicka wrote: >> >> On tramp3d all VRP passes threads together 730 branches, all DOM >passes 393, so >> FSM threading (with 1957 branches) is the most effective one. Perhaps >eventually >> early VRP can also do bit of work. >That's roughly consistent with what I've seen. I have some thoughts on > >what DOM's threading pass is catching that we're missing in the >backward/FSM threader, but I haven't had time to see how big those >effects really are. > >> >> I am not 100% sure from where "backward" is comming from. I guess is >means that >> analysis goes backward from conditionals to definitions: it looks for >> conditional driven by a PHI statement that has a constant value on >some paths >> and duplicates for those. This seems cheap and rather effective way >of getting >> good part of the threading oppurtunities (most expensive part is >probably >> identifying and walking paths that will not be threaded at the end). >Correct, forward/backward is based on the direction of analysis. ie, >do >we start at the conditional and work backwards through the USE-DEF >chain >or do we build a equivalence table as we walk forward through the >entire >function and use the equivalence tables to simplify the conditional. > > >> >> BTW I wonder if the same analysis can't be done for other >instructions where constant >> operands are very profitable, like division or multiplication. >Yes. There's all kinds of problems that can be solved using a >backwards >walk to identify useful properties on paths, then path duplication to >isolate the use point and modify it. > >So you could (for example) walk backwards from an indirect call and >identify paths through the CFG where we know the target. We then can >use path duplication to isolate that path from the rest and call the >target function directly. Hmm, isn't walking backwards from uses doing a lot of redundant stmt walking compared to walking stmts once in forward direction? To me it sounds like a 'local' patterns matching like optimization rather than a global one with proper data flow or a lattice? Richard. > >Jeff
On Thu, 11 Aug 2016, Jan Hubicka wrote: > Hi, > this patch adds early jump threading pass. Jump threading is one of most > common cases where estimated profile becomes corrupted, because the branches > are predicted independently beforehand. This patch simply adds early mode to > jump threader that does not permit code growth and thus only win/win threads > are done before profile is constructed. > > Naturally this also improves IPA decisions because code size/time is estimated > more acurately. > > It is not very cool to add another pass and the jump threader is already > run 5 times. I think incrementally we could drop one of late threaders at least. > I tried to measure compile time effects but they are in wash. Moreover the patch > pays for itself in cc1plus code size: > > Before patch to tweak size estimates: 27779964 > Current mainline: 27748900 > With patch applied: 27716173 > > So despite adding few functions the code size effect is positive which I think > is quite nice. > > Given the fact that jump threading seems quite lightweit, I wonder why it is > guarded by flag_expensive_optimizations? Is it expensive in terms of code > size? > > The effectivity of individual threading passes is as follows (for tramp3d) > > mainline with patch > pass thread count profile mismatches thread count profile mismatch > early 525 > 1 1853 1900 316 337 > 2 4 812 4 288 > 3 24 1450 32 947 > 4 76 1457 75 975 > > So at least tramp3d data seems to suggest that we can drop the second occurence > of jump threading and that most of the job thread1 does can be done by the > size restricted early version (the lower absolute counts are caused by the > fact that threadable paths gets duplicated by the inliner). 50% drop in > profile distortion is not bad. I wonder why basically every threaded paths seems > to introduce a mismatch. > > The patch distorts testusite somewhat, in most cases we only want to update > dump files or disable early threading: > > +XPASS: gcc.dg/uninit-15.c (test for warnings, line 13) > +XPASS: gcc.dg/uninit-15.c (test for warnings, line 23) > +FAIL: gcc.dg/uninit-15.c (test for warnings, line 24) > +FAIL: gcc.dg/tree-ssa/pr68198.c scan-tree-dump-times thread1 "Registering FSM" 1 > +FAIL: gcc.dg/tree-ssa/pr69196-1.c scan-tree-dump thread1 "FSM did not thread around loop and would copy too many statements" > +FAIL: gcc.dg/tree-ssa/ssa-dom-thread-2b.c scan-tree-dump-times thread1 "Jumps threaded: 1" 1 > +FAIL: gcc.dg/tree-ssa/ssa-thread-13.c scan-tree-dump thread1 "FSM" > +FAIL: gcc.dg/tree-ssa/vrp01.c scan-tree-dump-times vrp1 "Folding predicate p_.*to 1" 1 > +FAIL: gcc.dg/tree-ssa/vrp56.c scan-tree-dump thread1 "Jumps threaded: 1" > +FAIL: gcc.dg/tree-ssa/vrp92.c scan-tree-dump vrp1 "res_.: \\\\[1, 1\\\\]" > > This testcase is the now probably unnecesary heuristics (it may still be > relevant in cases we do not thread because of size bounds but its effectivity > may not be worth the maintenance cost): > > +FAIL: g++.dg/predict-loop-exit-1.C -std=gnu++11 scan-tree-dump-times profile_estimate "extra loop exit heuristics of edge[^:]*:" 2 > +FAIL: g++.dg/predict-loop-exit-1.C -std=gnu++11 scan-tree-dump-times profile_estimate "loop exit heuristics of edge[^:]*:" 3 > +FAIL: g++.dg/predict-loop-exit-1.C -std=gnu++14 scan-tree-dump-times profile_estimate "extra loop exit heuristics of edge[^:]*:" 2 > +FAIL: g++.dg/predict-loop-exit-1.C -std=gnu++14 scan-tree-dump-times profile_estimate "loop exit heuristics of edge[^:]*:" 3 > +FAIL: g++.dg/predict-loop-exit-1.C -std=gnu++98 scan-tree-dump-times profile_estimate "extra loop exit heuristics of edge[^:]*:" 2 > +FAIL: g++.dg/predict-loop-exit-1.C -std=gnu++98 scan-tree-dump-times profile_estimate "loop exit heuristics of edge[^:]*:" 3 > +FAIL: g++.dg/predict-loop-exit-2.C -std=gnu++11 scan-tree-dump-times profile_estimate "extra loop exit heuristics of edge[^:]*:" 1 > +FAIL: g++.dg/predict-loop-exit-2.C -std=gnu++11 scan-tree-dump-times profile_estimate "loop exit heuristics of edge[^:]*:" 2 > +FAIL: g++.dg/predict-loop-exit-2.C -std=gnu++14 scan-tree-dump-times profile_estimate "extra loop exit heuristics of edge[^:]*:" 1 > +FAIL: g++.dg/predict-loop-exit-2.C -std=gnu++14 scan-tree-dump-times profile_estimate "loop exit heuristics of edge[^:]*:" 2 > +FAIL: g++.dg/predict-loop-exit-2.C -std=gnu++98 scan-tree-dump-times profile_estimate "extra loop exit heuristics of edge[^:]*:" 1 > +FAIL: g++.dg/predict-loop-exit-2.C -std=gnu++98 scan-tree-dump-times profile_estimate "loop exit heuristics of edge[^:]*:" 2 > +FAIL: g++.dg/predict-loop-exit-3.C -std=gnu++11 scan-tree-dump-times profile_estimate "extra loop exit heuristics of edge[^:]*:" 2 > +FAIL: g++.dg/predict-loop-exit-3.C -std=gnu++11 scan-tree-dump-times profile_estimate "loop exit heuristics of edge[^:]*:" 3 > +FAIL: g++.dg/predict-loop-exit-3.C -std=gnu++14 scan-tree-dump-times profile_estimate "extra loop exit heuristics of edge[^:]*:" 2 > +FAIL: g++.dg/predict-loop-exit-3.C -std=gnu++14 scan-tree-dump-times profile_estimate "loop exit heuristics of edge[^:]*:" 3 > +FAIL: g++.dg/predict-loop-exit-3.C -std=gnu++98 scan-tree-dump-times profile_estimate "extra loop exit heuristics of edge[^:]*:" 2 > +FAIL: g++.dg/predict-loop-exit-3.C -std=gnu++98 scan-tree-dump-times profile_estimate "loop exit heuristics of edge[^:]*:" 3 > > If the patch seems acceptable, I will do the updates. One option why I did > not do that is that it seems to be now posisble to pass parameters to passes > from passes.def, so perhaps we do not need early_thread_jumps, but doing so is > consistent with way we handle other early passes. Few comments on the patch itself below. Note currently the pass is enabled at -O[s2]+ only which I think is fine. We certainly do not want it at -Og. (and nobody seems to try making -O1 sane anyway) Given your statistics we want to remove thread2 and thread3 (I guess tracer immediately before thread_jumps is odd as well, I'd at least swap them). The fact that thread4 threads more than thread3 argues for DOM being required as a FSM thread enabler rather than the other way around. So I'd try with generally moving the threader _after_ CSE-like optimizations. Richard. > Bootstrapped/regtested x86_64-linux > Honza > > * passes.def (pass_early_thread_jumps): Schedule after forwprop. > * tree-pass.h (make_pass_early_thread_jumps): Declare. > * tree-ssa-threadbackward.c (fsm_find_thread_path, > fsm_find_thread_path, profitable_jump_thread_path, > fsm_find_control_statement_thread_paths, > find_jump_threads_backwards): Add speed_p parameter. > (pass_data_early_thread_jumps): New pass. > (make_pass_early_thread_jumps): New function. > > Index: passes.def > =================================================================== > --- passes.def (revision 239218) > +++ passes.def (working copy) > @@ -84,6 +84,7 @@ along with GCC; see the file COPYING3. > /* After CCP we rewrite no longer addressed locals into SSA > form if possible. */ > NEXT_PASS (pass_forwprop); > + NEXT_PASS (pass_early_thread_jumps); What's the reason for this placement? I know Jeff argues that as jump threading helps CSE we need to place it before CSE but OTOH the FSM style threading relies on copies and redundancies being optimized already and the above has only constants and copies being propagated and forwprop left you with lots of dead code (but it should also have copies and constants propagated but it leaves PHIs alone, not propagating into them or removing degenerate ones - sth to fix I guess). So I'd be interested to see threading statistics when you place the threading pass after early FRE (or cd_dce). I guess early FRE will already handle quite some of the simplistic "threading" opportunities (it optimizes redundant checks) thus numbers may even get worse here. That said - if you put it before early FRE then I'd put it right after CCP, not after forwprop. > NEXT_PASS (pass_sra_early); > /* pass_build_ealias is a dummy pass that ensures that we > execute TODO_rebuild_alias at this point. */ > Index: tree-pass.h > =================================================================== > --- tree-pass.h (revision 239218) > +++ tree-pass.h (working copy) > @@ -399,6 +399,7 @@ extern gimple_opt_pass *make_pass_cd_dce > extern gimple_opt_pass *make_pass_call_cdce (gcc::context *ctxt); > extern gimple_opt_pass *make_pass_merge_phi (gcc::context *ctxt); > extern gimple_opt_pass *make_pass_thread_jumps (gcc::context *ctxt); > +extern gimple_opt_pass *make_pass_early_thread_jumps (gcc::context *ctxt); > extern gimple_opt_pass *make_pass_split_crit_edges (gcc::context *ctxt); > extern gimple_opt_pass *make_pass_laddress (gcc::context *ctxt); > extern gimple_opt_pass *make_pass_pre (gcc::context *ctxt); > Index: tree-ssa-threadbackward.c > =================================================================== > --- tree-ssa-threadbackward.c (revision 239219) > +++ tree-ssa-threadbackward.c (working copy) > @@ -61,12 +61,14 @@ get_gimple_control_stmt (basic_block bb) > /* Return true if the CFG contains at least one path from START_BB to END_BB. > When a path is found, record in PATH the blocks from END_BB to START_BB. > VISITED_BBS is used to make sure we don't fall into an infinite loop. Bound > - the recursion to basic blocks belonging to LOOP. */ > + the recursion to basic blocks belonging to LOOP. > + SPEED_P indicate that we could increase code size to improve the code path */ > > static bool > fsm_find_thread_path (basic_block start_bb, basic_block end_bb, > vec<basic_block, va_gc> *&path, > - hash_set<basic_block> *visited_bbs, loop_p loop) > + hash_set<basic_block> *visited_bbs, loop_p loop, > + bool speed_p) > { > if (loop != start_bb->loop_father) > return false; > @@ -82,7 +84,8 @@ fsm_find_thread_path (basic_block start_ > edge e; > edge_iterator ei; > FOR_EACH_EDGE (e, ei, start_bb->succs) > - if (fsm_find_thread_path (e->dest, end_bb, path, visited_bbs, loop)) > + if (fsm_find_thread_path (e->dest, end_bb, path, visited_bbs, loop, > + speed_p)) > { > vec_safe_push (path, start_bb); > return true; > @@ -101,11 +104,13 @@ fsm_find_thread_path (basic_block start_ > value on PATH. ARG is the value of that SSA_NAME. > > BBI will be appended to PATH when we have a profitable jump threading > - path. Callers are responsible for removing BBI from PATH in that case. */ > + path. Callers are responsible for removing BBI from PATH in that case. > + > + SPEED_P indicate that we could increase code size to improve the code path */ > > static edge > profitable_jump_thread_path (vec<basic_block, va_gc> *&path, > - basic_block bbi, tree name, tree arg) > + basic_block bbi, tree name, tree arg, bool speed_p) > { > /* Note BBI is not in the path yet, hence the +1 in the test below > to make sure BBI is accounted for in the path length test. */ > @@ -306,7 +311,7 @@ profitable_jump_thread_path (vec<basic_b > return NULL; > } > > - if (optimize_edge_for_speed_p (taken_edge)) > + if (speed_p && optimize_edge_for_speed_p (taken_edge)) > { > if (n_insns >= PARAM_VALUE (PARAM_MAX_FSM_THREAD_PATH_INSNS)) > { > @@ -421,13 +426,15 @@ convert_and_register_jump_thread_path (v > > /* We trace the value of the SSA_NAME NAME back through any phi nodes looking > for places where it gets a constant value and save the path. Stop after > - having recorded MAX_PATHS jump threading paths. */ > + having recorded MAX_PATHS jump threading paths. > + > + SPEED_P indicate that we could increase code size to improve the code path */ > > static void > fsm_find_control_statement_thread_paths (tree name, > hash_set<basic_block> *visited_bbs, > vec<basic_block, va_gc> *&path, > - bool seen_loop_phi) > + bool seen_loop_phi, bool speed_p) > { > /* If NAME appears in an abnormal PHI, then don't try to trace its > value back through PHI nodes. */ > @@ -495,7 +502,7 @@ fsm_find_control_statement_thread_paths > hash_set<basic_block> *visited_bbs = new hash_set<basic_block>; > > if (fsm_find_thread_path (var_bb, e->src, next_path, visited_bbs, > - e->src->loop_father)) > + e->src->loop_father, speed_p)) > ++e_count; > > delete visited_bbs; > @@ -561,7 +568,7 @@ fsm_find_control_statement_thread_paths > /* Recursively follow SSA_NAMEs looking for a constant > definition. */ > fsm_find_control_statement_thread_paths (arg, visited_bbs, path, > - seen_loop_phi); > + seen_loop_phi, speed_p); > > path->pop (); > continue; > @@ -572,7 +579,8 @@ fsm_find_control_statement_thread_paths > > /* If this is a profitable jump thread path, then convert it > into the canonical form and register it. */ > - edge taken_edge = profitable_jump_thread_path (path, bbi, name, arg); > + edge taken_edge = profitable_jump_thread_path (path, bbi, name, arg, > + speed_p); > if (taken_edge) > { > if (bb_loop_depth (taken_edge->src) > @@ -588,7 +596,7 @@ fsm_find_control_statement_thread_paths > > if (TREE_CODE (arg) == SSA_NAME) > fsm_find_control_statement_thread_paths (arg, visited_bbs, > - path, seen_loop_phi); > + path, seen_loop_phi, speed_p); > > else > { > @@ -598,7 +606,7 @@ fsm_find_control_statement_thread_paths > path->pop (); > > edge taken_edge = profitable_jump_thread_path (path, var_bb, > - name, arg); > + name, arg, speed_p); > if (taken_edge) > { > if (bb_loop_depth (taken_edge->src) > @@ -622,10 +630,11 @@ fsm_find_control_statement_thread_paths > is a constant. Record such paths for jump threading. > > It is assumed that BB ends with a control statement and that by > - finding a path where NAME is a constant, we can thread the path. */ > + finding a path where NAME is a constant, we can thread the path. > + SPEED_P indicate that we could increase code size to improve the code path */ > > void > -find_jump_threads_backwards (basic_block bb) > +find_jump_threads_backwards (basic_block bb, bool speed_p) > { > gimple *stmt = get_gimple_control_stmt (bb); > if (!stmt) > @@ -655,7 +664,8 @@ find_jump_threads_backwards (basic_block > hash_set<basic_block> *visited_bbs = new hash_set<basic_block>; > > max_threaded_paths = PARAM_VALUE (PARAM_MAX_FSM_THREAD_PATHS); > - fsm_find_control_statement_thread_paths (name, visited_bbs, bb_path, false); > + fsm_find_control_statement_thread_paths (name, visited_bbs, bb_path, false, > + speed_p); > > delete visited_bbs; > vec_free (bb_path); > @@ -703,7 +713,7 @@ pass_thread_jumps::execute (function *fu > FOR_EACH_BB_FN (bb, fun) > { > if (EDGE_COUNT (bb->succs) > 1) > - find_jump_threads_backwards (bb); > + find_jump_threads_backwards (bb, true); > } > thread_through_all_blocks (true); > return 0; > @@ -716,3 +726,59 @@ make_pass_thread_jumps (gcc::context *ct > { > return new pass_thread_jumps (ctxt); > } > + > +namespace { > + > +const pass_data pass_data_early_thread_jumps = > +{ > + GIMPLE_PASS, > + "early_thread", > + OPTGROUP_NONE, > + TV_TREE_SSA_THREAD_JUMPS, > + ( PROP_cfg | PROP_ssa ), > + 0, > + 0, > + 0, > + ( TODO_cleanup_cfg | TODO_update_ssa ), > +}; > + > +class pass_early_thread_jumps : public gimple_opt_pass > +{ > +public: > + pass_early_thread_jumps (gcc::context *ctxt) > + : gimple_opt_pass (pass_data_early_thread_jumps, ctxt) > + {} > + > + opt_pass * clone (void) { return new pass_early_thread_jumps (m_ctxt); } > + virtual bool gate (function *); > + virtual unsigned int execute (function *); > +}; > + > +bool > +pass_early_thread_jumps::gate (function *fun ATTRIBUTE_UNUSED) > +{ > + return true; > +} > + > + > +unsigned int > +pass_early_thread_jumps::execute (function *fun) > +{ > + /* Try to thread each block with more than one successor. */ > + basic_block bb; > + FOR_EACH_BB_FN (bb, fun) > + { > + if (EDGE_COUNT (bb->succs) > 1) > + find_jump_threads_backwards (bb, false); > + } > + thread_through_all_blocks (true); > + return 0; > +} > + > +} > + > +gimple_opt_pass * > +make_pass_early_thread_jumps (gcc::context *ctxt) > +{ > + return new pass_early_thread_jumps (ctxt); > +} > >
> > * passes.def (pass_early_thread_jumps): Schedule after forwprop. > > * tree-pass.h (make_pass_early_thread_jumps): Declare. > > * tree-ssa-threadbackward.c (fsm_find_thread_path, > > fsm_find_thread_path, profitable_jump_thread_path, > > fsm_find_control_statement_thread_paths, > > find_jump_threads_backwards): Add speed_p parameter. > > (pass_data_early_thread_jumps): New pass. > > (make_pass_early_thread_jumps): New function. > > > > Index: passes.def > > =================================================================== > > --- passes.def (revision 239218) > > +++ passes.def (working copy) > > @@ -84,6 +84,7 @@ along with GCC; see the file COPYING3. > > /* After CCP we rewrite no longer addressed locals into SSA > > form if possible. */ > > NEXT_PASS (pass_forwprop); > > + NEXT_PASS (pass_early_thread_jumps); > > What's the reason for this placement? I know Jeff argues that > as jump threading helps CSE we need to place it before CSE but > OTOH the FSM style threading relies on copies and redundancies > being optimized already and the above has only constants and copies > being propagated and forwprop left you with lots of dead code > (but it should also have copies and constants propagated but it > leaves PHIs alone, not propagating into them or removing degenerate > ones - sth to fix I guess). > > So I'd be interested to see threading statistics when you place > the threading pass after early FRE (or cd_dce). I guess early > FRE will already handle quite some of the simplistic "threading" > opportunities (it optimizes redundant checks) thus numbers may > even get worse here. > > That said - if you put it before early FRE then I'd put it > right after CCP, not after forwprop. I placed it just after forwprop becasue the pattern it handles: bb0: x = a COND b; if (x) goto ... else goto ... Will be transformed into: bb0: if (a COND b) goto ... else goto ... is probably going to help the simpistic analysis done by threadbackwards. I collected data with placement just before FRE and they were more or less identical. In general threading helps forward propagators becuase of code specialization it does. It does not like dead code (as it will get accounted and prevent duplication), degenerate PHIs (because it will do useless duplication - something to fix probably), and unpropagated temporaries (because fsm_find_control_statement_thread_paths does not look into them) Honza
On 08/12/2016 12:02 AM, Richard Biener wrote: > > Hmm, isn't walking backwards from uses doing a lot of redundant stmt > walking compared to walking stmts once in forward direction? To me > it sounds like a 'local' patterns matching like optimization rather > than a global one with proper data flow or a lattice? You end up walking the use-def chain, so you look only at the chain of feeding statements. Forward threading is actually worse because it tries to walk *past* the current point in the dominator walk. For example at a dominance frontier we have multiple paths merging -- we will walk all statements in the merge block for every incoming path as well as all statements on the outgoing paths of the merge block. We have to update the various tables, then unwind them as we work through all those paths. I have not done a full analysis, but I strongly suspect the backwards threader will ultimately end up doing less work, both in the common and pathological cases. jeff
On 08/12/2016 05:27 AM, Jan Hubicka wrote: >>> * passes.def (pass_early_thread_jumps): Schedule after forwprop. >>> * tree-pass.h (make_pass_early_thread_jumps): Declare. >>> * tree-ssa-threadbackward.c (fsm_find_thread_path, >>> fsm_find_thread_path, profitable_jump_thread_path, >>> fsm_find_control_statement_thread_paths, >>> find_jump_threads_backwards): Add speed_p parameter. >>> (pass_data_early_thread_jumps): New pass. >>> (make_pass_early_thread_jumps): New function. >>> >>> Index: passes.def >>> =================================================================== >>> --- passes.def (revision 239218) >>> +++ passes.def (working copy) >>> @@ -84,6 +84,7 @@ along with GCC; see the file COPYING3. >>> /* After CCP we rewrite no longer addressed locals into SSA >>> form if possible. */ >>> NEXT_PASS (pass_forwprop); >>> + NEXT_PASS (pass_early_thread_jumps); >> >> What's the reason for this placement? I know Jeff argues that >> as jump threading helps CSE we need to place it before CSE but >> OTOH the FSM style threading relies on copies and redundancies >> being optimized already and the above has only constants and copies >> being propagated and forwprop left you with lots of dead code >> (but it should also have copies and constants propagated but it >> leaves PHIs alone, not propagating into them or removing degenerate >> ones - sth to fix I guess). >> >> So I'd be interested to see threading statistics when you place >> the threading pass after early FRE (or cd_dce). I guess early >> FRE will already handle quite some of the simplistic "threading" >> opportunities (it optimizes redundant checks) thus numbers may >> even get worse here. >> >> That said - if you put it before early FRE then I'd put it >> right after CCP, not after forwprop. > > I placed it just after forwprop becasue the pattern it handles: > bb0: > x = a COND b; > if (x) goto ... else goto ... > > Will be transformed into: > > bb0: > if (a COND b) goto ... else goto ... Note that extending the backward threader to handle the former style sequence is relatively straightforward. In fact, building a bit of that kind of infrastructure is what I expect to be the biggest source of things we're missing relative to the forward threader. > > In general threading helps forward propagators becuase of code specialization > it does. It does not like dead code (as it will get accounted and prevent > duplication), degenerate PHIs (because it will do useless duplication - > something to fix probably), and unpropagated temporaries (because > fsm_find_control_statement_thread_paths does not look into them) Yes, avoiding useless duplication and factoring identical paths from different incoming edges are definitely on the TODO list. They're clearly a source of codesize issues when we compare the backward threader to the forward threader. Presumably for unpropagated temporaries you're referring to cases where both operands of a COND_EXPR need to be looked up? Right now we only walk one operand backward to a constant, but we really want to walk both. jeff > > Honza >
> I don't think the backwards/FSM threader tries to update the profile > data at all right now. This seems to be cause of the regression andrew is speaking about. I wrote updating of profile after threading few times and there was also Theresa's patch. I tought all the threaders use common duplication engine and that one does the updating? I will try to dive into the code, but creating basic blocks with frequency 0 across hot paths is bad idea - they will be optimized for size. Honza
On 08/16/2016 05:02 AM, Jan Hubicka wrote: >> I don't think the backwards/FSM threader tries to update the profile >> data at all right now. > > This seems to be cause of the regression andrew is speaking about. I wrote updating > of profile after threading few times and there was also Theresa's patch. > I tought all the threaders use common duplication engine and that one does the updating? > I will try to dive into the code, but creating basic blocks with frequency 0 across hot > paths is bad idea - they will be optimized for size. No, the updaters are independent. We'll need to port Theresa's work to the newer scheme. jeff
On 08/12/2016 02:02 AM, Richard Biener wrote: > On Thu, 11 Aug 2016, Jan Hubicka wrote: > >> Hi, >> this patch adds early jump threading pass. Jump threading is one of most >> common cases where estimated profile becomes corrupted, because the branches >> are predicted independently beforehand. This patch simply adds early mode to >> jump threader that does not permit code growth and thus only win/win threads >> are done before profile is constructed. >> >> Naturally this also improves IPA decisions because code size/time is estimated >> more acurately. >> >> It is not very cool to add another pass and the jump threader is already >> run 5 times. I think incrementally we could drop one of late threaders at least. >> I tried to measure compile time effects but they are in wash. Moreover the patch >> pays for itself in cc1plus code size: >> >> Before patch to tweak size estimates: 27779964 >> Current mainline: 27748900 >> With patch applied: 27716173 >> >> So despite adding few functions the code size effect is positive which I think >> is quite nice. >> >> Given the fact that jump threading seems quite lightweit, I wonder why it is >> guarded by flag_expensive_optimizations? Is it expensive in terms of code >> size? >> >> The effectivity of individual threading passes is as follows (for tramp3d) >> >> mainline with patch >> pass thread count profile mismatches thread count profile mismatch >> early 525 >> 1 1853 1900 316 337 >> 2 4 812 4 288 >> 3 24 1450 32 947 >> 4 76 1457 75 975 >> >> So at least tramp3d data seems to suggest that we can drop the second occurence >> of jump threading and that most of the job thread1 does can be done by the >> size restricted early version (the lower absolute counts are caused by the >> fact that threadable paths gets duplicated by the inliner). 50% drop in >> profile distortion is not bad. I wonder why basically every threaded paths seems >> to introduce a mismatch. >> >> The patch distorts testusite somewhat, in most cases we only want to update >> dump files or disable early threading: >> >> +XPASS: gcc.dg/uninit-15.c (test for warnings, line 13) >> +XPASS: gcc.dg/uninit-15.c (test for warnings, line 23) >> +FAIL: gcc.dg/uninit-15.c (test for warnings, line 24) >> +FAIL: gcc.dg/tree-ssa/pr68198.c scan-tree-dump-times thread1 "Registering FSM" 1 >> +FAIL: gcc.dg/tree-ssa/pr69196-1.c scan-tree-dump thread1 "FSM did not thread around loop and would copy too many statements" >> +FAIL: gcc.dg/tree-ssa/ssa-dom-thread-2b.c scan-tree-dump-times thread1 "Jumps threaded: 1" 1 >> +FAIL: gcc.dg/tree-ssa/ssa-thread-13.c scan-tree-dump thread1 "FSM" >> +FAIL: gcc.dg/tree-ssa/vrp01.c scan-tree-dump-times vrp1 "Folding predicate p_.*to 1" 1 >> +FAIL: gcc.dg/tree-ssa/vrp56.c scan-tree-dump thread1 "Jumps threaded: 1" >> +FAIL: gcc.dg/tree-ssa/vrp92.c scan-tree-dump vrp1 "res_.: \\\\[1, 1\\\\]" >> >> This testcase is the now probably unnecesary heuristics (it may still be >> relevant in cases we do not thread because of size bounds but its effectivity >> may not be worth the maintenance cost): >> >> +FAIL: g++.dg/predict-loop-exit-1.C -std=gnu++11 scan-tree-dump-times profile_estimate "extra loop exit heuristics of edge[^:]*:" 2 >> +FAIL: g++.dg/predict-loop-exit-1.C -std=gnu++11 scan-tree-dump-times profile_estimate "loop exit heuristics of edge[^:]*:" 3 >> +FAIL: g++.dg/predict-loop-exit-1.C -std=gnu++14 scan-tree-dump-times profile_estimate "extra loop exit heuristics of edge[^:]*:" 2 >> +FAIL: g++.dg/predict-loop-exit-1.C -std=gnu++14 scan-tree-dump-times profile_estimate "loop exit heuristics of edge[^:]*:" 3 >> +FAIL: g++.dg/predict-loop-exit-1.C -std=gnu++98 scan-tree-dump-times profile_estimate "extra loop exit heuristics of edge[^:]*:" 2 >> +FAIL: g++.dg/predict-loop-exit-1.C -std=gnu++98 scan-tree-dump-times profile_estimate "loop exit heuristics of edge[^:]*:" 3 >> +FAIL: g++.dg/predict-loop-exit-2.C -std=gnu++11 scan-tree-dump-times profile_estimate "extra loop exit heuristics of edge[^:]*:" 1 >> +FAIL: g++.dg/predict-loop-exit-2.C -std=gnu++11 scan-tree-dump-times profile_estimate "loop exit heuristics of edge[^:]*:" 2 >> +FAIL: g++.dg/predict-loop-exit-2.C -std=gnu++14 scan-tree-dump-times profile_estimate "extra loop exit heuristics of edge[^:]*:" 1 >> +FAIL: g++.dg/predict-loop-exit-2.C -std=gnu++14 scan-tree-dump-times profile_estimate "loop exit heuristics of edge[^:]*:" 2 >> +FAIL: g++.dg/predict-loop-exit-2.C -std=gnu++98 scan-tree-dump-times profile_estimate "extra loop exit heuristics of edge[^:]*:" 1 >> +FAIL: g++.dg/predict-loop-exit-2.C -std=gnu++98 scan-tree-dump-times profile_estimate "loop exit heuristics of edge[^:]*:" 2 >> +FAIL: g++.dg/predict-loop-exit-3.C -std=gnu++11 scan-tree-dump-times profile_estimate "extra loop exit heuristics of edge[^:]*:" 2 >> +FAIL: g++.dg/predict-loop-exit-3.C -std=gnu++11 scan-tree-dump-times profile_estimate "loop exit heuristics of edge[^:]*:" 3 >> +FAIL: g++.dg/predict-loop-exit-3.C -std=gnu++14 scan-tree-dump-times profile_estimate "extra loop exit heuristics of edge[^:]*:" 2 >> +FAIL: g++.dg/predict-loop-exit-3.C -std=gnu++14 scan-tree-dump-times profile_estimate "loop exit heuristics of edge[^:]*:" 3 >> +FAIL: g++.dg/predict-loop-exit-3.C -std=gnu++98 scan-tree-dump-times profile_estimate "extra loop exit heuristics of edge[^:]*:" 2 >> +FAIL: g++.dg/predict-loop-exit-3.C -std=gnu++98 scan-tree-dump-times profile_estimate "loop exit heuristics of edge[^:]*:" 3 >> >> If the patch seems acceptable, I will do the updates. One option why I did >> not do that is that it seems to be now posisble to pass parameters to passes >> from passes.def, so perhaps we do not need early_thread_jumps, but doing so is >> consistent with way we handle other early passes. > > Few comments on the patch itself below. Note currently the pass is > enabled at -O[s2]+ only which I think is fine. We certainly do not > want it at -Og. (and nobody seems to try making -O1 sane anyway) > > Given your statistics we want to remove thread2 and thread3 (I guess > tracer immediately before thread_jumps is odd as well, I'd at least > swap them). The fact that thread4 threads more than thread3 argues > for DOM being required as a FSM thread enabler rather than the > other way around. So I'd try with generally moving the threader > _after_ CSE-like optimizations. What we want to be doing is threading *before* CSE-like optimizations and removing the DOM/VRP jump threaders. Based on what I've seen for the last 15 years, threading helps CSE far more often than the opposite. The fact that we do CSE first, then threading is an artifact of the threading implementation wanting to exploit the CSE tables. But that's not necessary in a backwards walking threader. So what we need to be going is figuring out what key things aren't being handled so that we can drop the DOM/VRP threaders. Once that's done we will want to look at whether or not any of the standalone threading passes continue to make sense. jeff
On 08/11/2016 08:02 AM, Jan Hubicka wrote: > Hi, > this patch adds early jump threading pass. Jump threading is one of most > common cases where estimated profile becomes corrupted, because the branches > are predicted independently beforehand. This patch simply adds early mode to > jump threader that does not permit code growth and thus only win/win threads > are done before profile is constructed. > > Naturally this also improves IPA decisions because code size/time is estimated > more acurately. > > It is not very cool to add another pass and the jump threader is already > run 5 times. I think incrementally we could drop one of late threaders at least. > I tried to measure compile time effects but they are in wash. Moreover the patch > pays for itself in cc1plus code size: > > Before patch to tweak size estimates: 27779964 > Current mainline: 27748900 > With patch applied: 27716173 > > So despite adding few functions the code size effect is positive which I think > is quite nice. > > Given the fact that jump threading seems quite lightweit, I wonder why it is > guarded by flag_expensive_optimizations? Is it expensive in terms of code > size? > > The effectivity of individual threading passes is as follows (for tramp3d) > > mainline with patch > pass thread count profile mismatches thread count profile mismatch > early 525 > 1 1853 1900 316 337 > 2 4 812 4 288 So the real question here is what did VRP1 threading do between the standalone thread1/thread2 passes. ie, it may look tempting to eliminate the standalone thread2 pass, but threading done by VRP1 is expected to be pushed into the thread{1,2} passes. So removing thread2 is premature at this point. > > The patch distorts testusite somewhat, in most cases we only want to update > dump files or disable early threading: > > +XPASS: gcc.dg/uninit-15.c (test for warnings, line 13) > +XPASS: gcc.dg/uninit-15.c (test for warnings, line 23) > +FAIL: gcc.dg/uninit-15.c (test for warnings, line 24) This seems like a step backwards to me. We're warning about "j", but in the context in which its used, we really ought to be warning about "i". And we lost a later warning. > +FAIL: gcc.dg/tree-ssa/pr68198.c scan-tree-dump-times thread1 "Registering FSM" 1 Moved into early-threading, so adjusting test for that seems appropriate. > +FAIL: gcc.dg/tree-ssa/pr69196-1.c scan-tree-dump thread1 "FSM did not thread around loop and would copy too many statements" I'm not sure how to best check this one. Threading this test too aggressively results in something like a 2X bloat in the resulting code on the sparc. > +FAIL: gcc.dg/tree-ssa/ssa-dom-thread-2b.c scan-tree-dump-times thread1 "Jumps threaded: 1" 1 Moved into early threading, so adjusting the test for that seems appropriate to me. > +FAIL: gcc.dg/tree-ssa/ssa-thread-13.c scan-tree-dump thread1 "FSM" Moved into early threading, so adjusting the test for that seems appropriate to me. > +FAIL: gcc.dg/tree-ssa/vrp01.c scan-tree-dump-times vrp1 "Folding predicate p_.*to 1" 1 So the optimization moved into fre1 as a result of early jump threading which is good. We should probably check for that, even if the test is badly mis-named now. But this is showing one of the cases where the backwards threader is deficient. It's not using the conditional as an implied set. It's a known limitation and I'm hoping Andrew's work makes that an easier problem to solve. > +FAIL: gcc.dg/tree-ssa/vrp56.c scan-tree-dump thread1 "Jumps threaded: 1" Moved into early threading, so adjusting the test for that seems appropriate to me. > +FAIL: gcc.dg/tree-ssa/vrp92.c scan-tree-dump vrp1 "res_.: \\\\[1, 1\\\\]" Some results adjustment seems to be in order here. I think verifying that early threading finds the obvious jump threads, then vrp is able to eliminate the 2nd conditional and the result is a collapsed return i for the entire function. > > This testcase is the now probably unnecesary heuristics (it may still be > relevant in cases we do not thread because of size bounds but its effectivity > may not be worth the maintenance cost): > > +FAIL: g++.dg/predict-loop-exit-1.C -std=gnu++11 scan-tree-dump-times profile_estimate "extra loop exit heuristics of edge[^:]*:" 2 > +FAIL: g++.dg/predict-loop-exit-1.C -std=gnu++11 scan-tree-dump-times profile_estimate "loop exit heuristics of edge[^:]*:" 3 > +FAIL: g++.dg/predict-loop-exit-1.C -std=gnu++14 scan-tree-dump-times profile_estimate "extra loop exit heuristics of edge[^:]*:" 2 > +FAIL: g++.dg/predict-loop-exit-1.C -std=gnu++14 scan-tree-dump-times profile_estimate "loop exit heuristics of edge[^:]*:" 3 > +FAIL: g++.dg/predict-loop-exit-1.C -std=gnu++98 scan-tree-dump-times profile_estimate "extra loop exit heuristics of edge[^:]*:" 2 > +FAIL: g++.dg/predict-loop-exit-1.C -std=gnu++98 scan-tree-dump-times profile_estimate "loop exit heuristics of edge[^:]*:" 3 > +FAIL: g++.dg/predict-loop-exit-2.C -std=gnu++11 scan-tree-dump-times profile_estimate "extra loop exit heuristics of edge[^:]*:" 1 > +FAIL: g++.dg/predict-loop-exit-2.C -std=gnu++11 scan-tree-dump-times profile_estimate "loop exit heuristics of edge[^:]*:" 2 > +FAIL: g++.dg/predict-loop-exit-2.C -std=gnu++14 scan-tree-dump-times profile_estimate "extra loop exit heuristics of edge[^:]*:" 1 > +FAIL: g++.dg/predict-loop-exit-2.C -std=gnu++14 scan-tree-dump-times profile_estimate "loop exit heuristics of edge[^:]*:" 2 > +FAIL: g++.dg/predict-loop-exit-2.C -std=gnu++98 scan-tree-dump-times profile_estimate "extra loop exit heuristics of edge[^:]*:" 1 > +FAIL: g++.dg/predict-loop-exit-2.C -std=gnu++98 scan-tree-dump-times profile_estimate "loop exit heuristics of edge[^:]*:" 2 > +FAIL: g++.dg/predict-loop-exit-3.C -std=gnu++11 scan-tree-dump-times profile_estimate "extra loop exit heuristics of edge[^:]*:" 2 > +FAIL: g++.dg/predict-loop-exit-3.C -std=gnu++11 scan-tree-dump-times profile_estimate "loop exit heuristics of edge[^:]*:" 3 > +FAIL: g++.dg/predict-loop-exit-3.C -std=gnu++14 scan-tree-dump-times profile_estimate "extra loop exit heuristics of edge[^:]*:" 2 > +FAIL: g++.dg/predict-loop-exit-3.C -std=gnu++14 scan-tree-dump-times profile_estimate "loop exit heuristics of edge[^:]*:" 3 > +FAIL: g++.dg/predict-loop-exit-3.C -std=gnu++98 scan-tree-dump-times profile_estimate "extra loop exit heuristics of edge[^:]*:" 2 > +FAIL: g++.dg/predict-loop-exit-3.C -std=gnu++98 scan-tree-dump-times profile_estimate "loop exit heuristics of edge[^:]*:" 3 You're the best judge on how to deal with these. I leave that to your discretion. > If the patch seems acceptable, I will do the updates. One option why I did > not do that is that it seems to be now posisble to pass parameters to passes > from passes.def, so perhaps we do not need early_thread_jumps, but doing so is > consistent with way we handle other early passes. Your call on this. One of the things in the related literature is a limiter on how far back the use-def chains we go. I'd envisioned adding that limiter and having early jump threading use a different (much smaller) limiter. But I don't think it's critical at this point. > > Bootstrapped/regtested x86_64-linux > Honza > > * passes.def (pass_early_thread_jumps): Schedule after forwprop. > * tree-pass.h (make_pass_early_thread_jumps): Declare. > * tree-ssa-threadbackward.c (fsm_find_thread_path, > fsm_find_thread_path, profitable_jump_thread_path, > fsm_find_control_statement_thread_paths, > find_jump_threads_backwards): Add speed_p parameter. > (pass_data_early_thread_jumps): New pass. > (make_pass_early_thread_jumps): New function. LGTM. OK after fixing up the tests. Jeff
Index: passes.def =================================================================== --- passes.def (revision 239218) +++ passes.def (working copy) @@ -84,6 +84,7 @@ along with GCC; see the file COPYING3. /* After CCP we rewrite no longer addressed locals into SSA form if possible. */ NEXT_PASS (pass_forwprop); + NEXT_PASS (pass_early_thread_jumps); NEXT_PASS (pass_sra_early); /* pass_build_ealias is a dummy pass that ensures that we execute TODO_rebuild_alias at this point. */ Index: tree-pass.h =================================================================== --- tree-pass.h (revision 239218) +++ tree-pass.h (working copy) @@ -399,6 +399,7 @@ extern gimple_opt_pass *make_pass_cd_dce extern gimple_opt_pass *make_pass_call_cdce (gcc::context *ctxt); extern gimple_opt_pass *make_pass_merge_phi (gcc::context *ctxt); extern gimple_opt_pass *make_pass_thread_jumps (gcc::context *ctxt); +extern gimple_opt_pass *make_pass_early_thread_jumps (gcc::context *ctxt); extern gimple_opt_pass *make_pass_split_crit_edges (gcc::context *ctxt); extern gimple_opt_pass *make_pass_laddress (gcc::context *ctxt); extern gimple_opt_pass *make_pass_pre (gcc::context *ctxt); Index: tree-ssa-threadbackward.c =================================================================== --- tree-ssa-threadbackward.c (revision 239219) +++ tree-ssa-threadbackward.c (working copy) @@ -61,12 +61,14 @@ get_gimple_control_stmt (basic_block bb) /* Return true if the CFG contains at least one path from START_BB to END_BB. When a path is found, record in PATH the blocks from END_BB to START_BB. VISITED_BBS is used to make sure we don't fall into an infinite loop. Bound - the recursion to basic blocks belonging to LOOP. */ + the recursion to basic blocks belonging to LOOP. + SPEED_P indicate that we could increase code size to improve the code path */ static bool fsm_find_thread_path (basic_block start_bb, basic_block end_bb, vec<basic_block, va_gc> *&path, - hash_set<basic_block> *visited_bbs, loop_p loop) + hash_set<basic_block> *visited_bbs, loop_p loop, + bool speed_p) { if (loop != start_bb->loop_father) return false; @@ -82,7 +84,8 @@ fsm_find_thread_path (basic_block start_ edge e; edge_iterator ei; FOR_EACH_EDGE (e, ei, start_bb->succs) - if (fsm_find_thread_path (e->dest, end_bb, path, visited_bbs, loop)) + if (fsm_find_thread_path (e->dest, end_bb, path, visited_bbs, loop, + speed_p)) { vec_safe_push (path, start_bb); return true; @@ -101,11 +104,13 @@ fsm_find_thread_path (basic_block start_ value on PATH. ARG is the value of that SSA_NAME. BBI will be appended to PATH when we have a profitable jump threading - path. Callers are responsible for removing BBI from PATH in that case. */ + path. Callers are responsible for removing BBI from PATH in that case. + + SPEED_P indicate that we could increase code size to improve the code path */ static edge profitable_jump_thread_path (vec<basic_block, va_gc> *&path, - basic_block bbi, tree name, tree arg) + basic_block bbi, tree name, tree arg, bool speed_p) { /* Note BBI is not in the path yet, hence the +1 in the test below to make sure BBI is accounted for in the path length test. */ @@ -306,7 +311,7 @@ profitable_jump_thread_path (vec<basic_b return NULL; } - if (optimize_edge_for_speed_p (taken_edge)) + if (speed_p && optimize_edge_for_speed_p (taken_edge)) { if (n_insns >= PARAM_VALUE (PARAM_MAX_FSM_THREAD_PATH_INSNS)) { @@ -421,13 +426,15 @@ convert_and_register_jump_thread_path (v /* We trace the value of the SSA_NAME NAME back through any phi nodes looking for places where it gets a constant value and save the path. Stop after - having recorded MAX_PATHS jump threading paths. */ + having recorded MAX_PATHS jump threading paths. + + SPEED_P indicate that we could increase code size to improve the code path */ static void fsm_find_control_statement_thread_paths (tree name, hash_set<basic_block> *visited_bbs, vec<basic_block, va_gc> *&path, - bool seen_loop_phi) + bool seen_loop_phi, bool speed_p) { /* If NAME appears in an abnormal PHI, then don't try to trace its value back through PHI nodes. */ @@ -495,7 +502,7 @@ fsm_find_control_statement_thread_paths hash_set<basic_block> *visited_bbs = new hash_set<basic_block>; if (fsm_find_thread_path (var_bb, e->src, next_path, visited_bbs, - e->src->loop_father)) + e->src->loop_father, speed_p)) ++e_count; delete visited_bbs; @@ -561,7 +568,7 @@ fsm_find_control_statement_thread_paths /* Recursively follow SSA_NAMEs looking for a constant definition. */ fsm_find_control_statement_thread_paths (arg, visited_bbs, path, - seen_loop_phi); + seen_loop_phi, speed_p); path->pop (); continue; @@ -572,7 +579,8 @@ fsm_find_control_statement_thread_paths /* If this is a profitable jump thread path, then convert it into the canonical form and register it. */ - edge taken_edge = profitable_jump_thread_path (path, bbi, name, arg); + edge taken_edge = profitable_jump_thread_path (path, bbi, name, arg, + speed_p); if (taken_edge) { if (bb_loop_depth (taken_edge->src) @@ -588,7 +596,7 @@ fsm_find_control_statement_thread_paths if (TREE_CODE (arg) == SSA_NAME) fsm_find_control_statement_thread_paths (arg, visited_bbs, - path, seen_loop_phi); + path, seen_loop_phi, speed_p); else { @@ -598,7 +606,7 @@ fsm_find_control_statement_thread_paths path->pop (); edge taken_edge = profitable_jump_thread_path (path, var_bb, - name, arg); + name, arg, speed_p); if (taken_edge) { if (bb_loop_depth (taken_edge->src) @@ -622,10 +630,11 @@ fsm_find_control_statement_thread_paths is a constant. Record such paths for jump threading. It is assumed that BB ends with a control statement and that by - finding a path where NAME is a constant, we can thread the path. */ + finding a path where NAME is a constant, we can thread the path. + SPEED_P indicate that we could increase code size to improve the code path */ void -find_jump_threads_backwards (basic_block bb) +find_jump_threads_backwards (basic_block bb, bool speed_p) { gimple *stmt = get_gimple_control_stmt (bb); if (!stmt) @@ -655,7 +664,8 @@ find_jump_threads_backwards (basic_block hash_set<basic_block> *visited_bbs = new hash_set<basic_block>; max_threaded_paths = PARAM_VALUE (PARAM_MAX_FSM_THREAD_PATHS); - fsm_find_control_statement_thread_paths (name, visited_bbs, bb_path, false); + fsm_find_control_statement_thread_paths (name, visited_bbs, bb_path, false, + speed_p); delete visited_bbs; vec_free (bb_path); @@ -703,7 +713,7 @@ pass_thread_jumps::execute (function *fu FOR_EACH_BB_FN (bb, fun) { if (EDGE_COUNT (bb->succs) > 1) - find_jump_threads_backwards (bb); + find_jump_threads_backwards (bb, true); } thread_through_all_blocks (true); return 0; @@ -716,3 +726,59 @@ make_pass_thread_jumps (gcc::context *ct { return new pass_thread_jumps (ctxt); } + +namespace { + +const pass_data pass_data_early_thread_jumps = +{ + GIMPLE_PASS, + "early_thread", + OPTGROUP_NONE, + TV_TREE_SSA_THREAD_JUMPS, + ( PROP_cfg | PROP_ssa ), + 0, + 0, + 0, + ( TODO_cleanup_cfg | TODO_update_ssa ), +}; + +class pass_early_thread_jumps : public gimple_opt_pass +{ +public: + pass_early_thread_jumps (gcc::context *ctxt) + : gimple_opt_pass (pass_data_early_thread_jumps, ctxt) + {} + + opt_pass * clone (void) { return new pass_early_thread_jumps (m_ctxt); } + virtual bool gate (function *); + virtual unsigned int execute (function *); +}; + +bool +pass_early_thread_jumps::gate (function *fun ATTRIBUTE_UNUSED) +{ + return true; +} + + +unsigned int +pass_early_thread_jumps::execute (function *fun) +{ + /* Try to thread each block with more than one successor. */ + basic_block bb; + FOR_EACH_BB_FN (bb, fun) + { + if (EDGE_COUNT (bb->succs) > 1) + find_jump_threads_backwards (bb, false); + } + thread_through_all_blocks (true); + return 0; +} + +} + +gimple_opt_pass * +make_pass_early_thread_jumps (gcc::context *ctxt) +{ + return new pass_early_thread_jumps (ctxt); +}