| Message ID | CAO2gOZX46uovqDNSG_v-jhvP6y5zoX9QVu7VE0r3U-Xrw9AXcg@mail.gmail.com |
|---|---|
| State | New |
| Headers | show |
> Are you suggesting a patch like this: > > Index: gcc/predict.c > =================================================================== > --- gcc/predict.c (revision 189835) > +++ gcc/predict.c (working copy) > @@ -1319,6 +1319,7 @@ > tree loop_bound_var = NULL; > tree loop_iv_base = NULL; > gimple stmt = NULL; > + int header_found = 0; We should use bool these days. > > exits = get_loop_exit_edges (loop); > n_exits = VEC_length (edge, exits); > @@ -1387,9 +1388,20 @@ > > bbs = get_loop_body (loop); > > + /* Loop branch heuristics - predict an edge back to a > + loop's head as taken. */ > + if (loop->latch && loop->latch->loop_father == loop) > + { > + edge e = find_edge (loop->latch, loop->header); > + if (e) > + { > + header_found = 1; > + predict_edge_def (e, PRED_LOOP_BRANCH, TAKEN); > + } > + } > + > for (j = 0; j < loop->num_nodes; j++) > { > - int header_found = 0; > edge e; > edge_iterator ei; > > @@ -1402,21 +1414,9 @@ > if (predicted_by_p (bb, PRED_CONTINUE)) > continue; > > - /* Loop branch heuristics - predict an edge back to a > - loop's head as taken. */ > - if (bb == loop->latch) > - { > - e = find_edge (loop->latch, loop->header); > - if (e) > - { > - header_found = 1; > - predict_edge_def (e, PRED_LOOP_BRANCH, TAKEN); > - } > - } > - > /* Loop exit heuristics - predict an edge exiting the loop if the > conditional has no loop header successors as not taken. */ > - if (!header_found > + if (!(header_found && bb == loop->latch) Yes, this seems resonable to me. > > which also makes header_found initialized before we visit loop blocks. > > > > Instead the code > > > > /* Loop exit heuristics - predict an edge exiting the loop if the > > conditional has no loop header successors as not taken. */ > > if (!header_found > > /* If we already used more reliable loop exit predictors, do not > > bother with PRED_LOOP_EXIT. */ > > ... > > FOR_EACH_EDGE (e, ei, bb->succs) > > if (e->dest->index < NUM_FIXED_BLOCKS > > || !flow_bb_inside_loop_p (loop, e->dest)) > > predict_edge (e, PRED_LOOP_EXIT, probability); > > > > looks wrong for bb's that are parts of an inner loop of loop - assuming we > > only want to predicate exits from loop and not exits from an inner loop > > that also happen to exit loop (we will do that when predicating the inner loop). > > You are right. And if we want to change this, we'd also need to modify > get_loop_exit_edges to only count edges whose src is in the same loop > level. However, this is relatively minor issue because it only > predicts inaccurate bias ratio, while in the testcase I gave, > LOOP_BRANCH is predicting in the wrong direction. Indeed, it is not the most important thing around. Patch seems OK. We should get the statistic about branch prediction effectivity working again. It is broken since we moved tree-profile early. I will try to look into it. Honza > > Thanks, > Dehao > > > > > > Is that what you experienced? > > > > Thanks, > > Richard.
On Tue, Jul 31, 2012 at 12:20 PM, Dehao Chen <dehao@google.com> wrote: > Are you suggesting a patch like this: > > Index: gcc/predict.c > =================================================================== > --- gcc/predict.c (revision 189835) > +++ gcc/predict.c (working copy) > @@ -1319,6 +1319,7 @@ > tree loop_bound_var = NULL; > tree loop_iv_base = NULL; > gimple stmt = NULL; > + int header_found = 0; > > exits = get_loop_exit_edges (loop); > n_exits = VEC_length (edge, exits); > @@ -1387,9 +1388,20 @@ > > bbs = get_loop_body (loop); > > + /* Loop branch heuristics - predict an edge back to a > + loop's head as taken. */ > + if (loop->latch && loop->latch->loop_father == loop) Hmm, so the issue is that loop->latch does not belong to loop? That looks like a bogus loop structure. Indeed we have the loop header of the inner loop as latch of the outer loop. It still looks ok to predict this as unlikely as the edge is not only the latch edge of the outer loop but also an exit of the inner loop. Easier for profile would be to force canonicalization via loop_optimizer_init (LOOPS_NORMAL); instead of loop_optimizer_init (0); if (dump_file && (dump_flags & TDF_DETAILS)) flow_loops_dump (dump_file, NULL, 0); mark_irreducible_loops (); > + { > + edge e = find_edge (loop->latch, loop->header); > + if (e) > + { > + header_found = 1; > + predict_edge_def (e, PRED_LOOP_BRANCH, TAKEN); > + } > + } > + > for (j = 0; j < loop->num_nodes; j++) > { > - int header_found = 0; > edge e; > edge_iterator ei; > > @@ -1402,21 +1414,9 @@ > if (predicted_by_p (bb, PRED_CONTINUE)) > continue; > > - /* Loop branch heuristics - predict an edge back to a > - loop's head as taken. */ > - if (bb == loop->latch) > - { > - e = find_edge (loop->latch, loop->header); > - if (e) > - { > - header_found = 1; > - predict_edge_def (e, PRED_LOOP_BRANCH, TAKEN); > - } > - } > - Yes until here, > /* Loop exit heuristics - predict an edge exiting the loop if the > conditional has no loop header successors as not taken. */ > - if (!header_found > + if (!(header_found && bb == loop->latch) here instead !header_found && bb->loop_father == loop > /* If we already used more reliable loop exit predictors, do not > bother with PRED_LOOP_EXIT. */ > && !predicted_by_p (bb, PRED_LOOP_ITERATIONS_GUESSED) > > On Tue, Jul 31, 2012 at 5:18 PM, Richard Guenther > <richard.guenther@gmail.com> wrote: >> On Tue, Jul 31, 2012 at 5:17 AM, Dehao Chen <dehao@google.com> wrote: >>> Hi, >>> >>> This patch fixed the problem when a LOOP_EXIT edge for the inner loop >>> happened to target at the LOOP_LATCH of the outer loop. As the outer >>> loop is processed first, the LOOP_BRANCH heuristic is honored >>> (first_match), thus the inner loop's trip count is 0. (The attached >>> unittest demonstrates this). >>> >>> Bootstrapped and passed gcc regression test. >>> >>> Is it ok for trunk? >>> >>> Thanks, >>> Dehao >>> >>> gcc/ChangeLog >>> >>> 2012-07-30 Dehao Chen <dehao@google.com> >>> >>> * predict.c (predict_loops): Fix the prediction of LOOP_BRANCH. >>> >>> gcc/testsuite/ChangeLog >>> >>> 2012-07-31 Dehao Chen <dehao@google.com> >>> >>> * gcc.dg/predict-7.c: New test. >>> >>> Index: gcc/testsuite/gcc.dg/predict-7.c >>> =================================================================== >>> --- gcc/testsuite/gcc.dg/predict-7.c (revision 0) >>> +++ gcc/testsuite/gcc.dg/predict-7.c (revision 0) >>> @@ -0,0 +1,17 @@ >>> +/* { dg-do compile } */ >>> +/* { dg-options "-O2 -fdump-tree-profile_estimate" } */ >>> + >>> +extern int global; >>> + >>> +int bar (int); >>> + >>> +void foo (int base) >>> +{ >>> + int i; >>> + while (global < 10) >>> + for (i = base; i < 10; i++) >>> + bar (i); >>> +} >>> + >>> +/* { dg-final { scan-tree-dump-times "loop branch heuristics" 0 >>> "profile_estimate"} } */ >>> +/* { dg-final { cleanup-tree-dump "profile_estimate" } } */ >>> Index: gcc/predict.c >>> =================================================================== >>> --- gcc/predict.c (revision 189835) >>> +++ gcc/predict.c (working copy) >>> @@ -1404,7 +1404,7 @@ >>> >>> /* Loop branch heuristics - predict an edge back to a >>> loop's head as taken. */ >>> - if (bb == loop->latch) >>> + if (bb == loop->latch && bb->loop_father == loop) >>> { >>> e = find_edge (loop->latch, loop->header); >>> if (e) >> >> I think this heuristic should instead move out of the loop iterating over loop >> nodes and be done before like >> >> if (loop->latch) >> { >> e = find_edge (loop->latch, loop->header); >> ... >> } >> >> which also makes header_found initialized before we visit loop blocks. >> >> Instead the code >> >> /* Loop exit heuristics - predict an edge exiting the loop if the >> conditional has no loop header successors as not taken. */ >> if (!header_found >> /* If we already used more reliable loop exit predictors, do not >> bother with PRED_LOOP_EXIT. */ >> ... >> FOR_EACH_EDGE (e, ei, bb->succs) >> if (e->dest->index < NUM_FIXED_BLOCKS >> || !flow_bb_inside_loop_p (loop, e->dest)) >> predict_edge (e, PRED_LOOP_EXIT, probability); >> >> looks wrong for bb's that are parts of an inner loop of loop - assuming we >> only want to predicate exits from loop and not exits from an inner loop >> that also happen to exit loop (we will do that when predicating the inner loop). > > You are right. And if we want to change this, we'd also need to modify > get_loop_exit_edges to only count edges whose src is in the same loop > level. However, this is relatively minor issue because it only > predicts inaccurate bias ratio, while in the testcase I gave, > LOOP_BRANCH is predicting in the wrong direction. > > Thanks, > Dehao > > >> >> Is that what you experienced? >> >> Thanks, >> Richard.
> On Tue, Jul 31, 2012 at 12:20 PM, Dehao Chen <dehao@google.com> wrote: > > Are you suggesting a patch like this: > > > > Index: gcc/predict.c > > =================================================================== > > --- gcc/predict.c (revision 189835) > > +++ gcc/predict.c (working copy) > > @@ -1319,6 +1319,7 @@ > > tree loop_bound_var = NULL; > > tree loop_iv_base = NULL; > > gimple stmt = NULL; > > + int header_found = 0; > > > > exits = get_loop_exit_edges (loop); > > n_exits = VEC_length (edge, exits); > > @@ -1387,9 +1388,20 @@ > > > > bbs = get_loop_body (loop); > > > > + /* Loop branch heuristics - predict an edge back to a > > + loop's head as taken. */ > > + if (loop->latch && loop->latch->loop_father == loop) > > Hmm, so the issue is that loop->latch does not belong to loop? That looks > like a bogus loop structure. Indeed we have the loop header of the inner > loop as latch of the outer loop. It still looks ok to predict this as unlikely > as the edge is not only the latch edge of the outer loop but also an exit > of the inner loop. > > Easier for profile would be to force canonicalization via > > loop_optimizer_init (LOOPS_NORMAL); > > instead of > > loop_optimizer_init (0); > if (dump_file && (dump_flags & TDF_DETAILS)) > flow_loops_dump (dump_file, NULL, 0); > > mark_irreducible_loops (); Yeah, this may also work. The reason it is not done is that 1) it seemed expensive to force CFG changes just to compute profile decade ago 2) cfgcleanup afterwards will anyway remove the headers again. So I originally hoped to do the right thing without normalization. Honza
On Tue, Jul 31, 2012 at 12:38 PM, Jan Hubicka <hubicka@ucw.cz> wrote: >> On Tue, Jul 31, 2012 at 12:20 PM, Dehao Chen <dehao@google.com> wrote: >> > Are you suggesting a patch like this: >> > >> > Index: gcc/predict.c >> > =================================================================== >> > --- gcc/predict.c (revision 189835) >> > +++ gcc/predict.c (working copy) >> > @@ -1319,6 +1319,7 @@ >> > tree loop_bound_var = NULL; >> > tree loop_iv_base = NULL; >> > gimple stmt = NULL; >> > + int header_found = 0; >> > >> > exits = get_loop_exit_edges (loop); >> > n_exits = VEC_length (edge, exits); >> > @@ -1387,9 +1388,20 @@ >> > >> > bbs = get_loop_body (loop); >> > >> > + /* Loop branch heuristics - predict an edge back to a >> > + loop's head as taken. */ >> > + if (loop->latch && loop->latch->loop_father == loop) >> >> Hmm, so the issue is that loop->latch does not belong to loop? That looks >> like a bogus loop structure. Indeed we have the loop header of the inner >> loop as latch of the outer loop. It still looks ok to predict this as unlikely >> as the edge is not only the latch edge of the outer loop but also an exit >> of the inner loop. >> >> Easier for profile would be to force canonicalization via >> >> loop_optimizer_init (LOOPS_NORMAL); >> >> instead of >> >> loop_optimizer_init (0); >> if (dump_file && (dump_flags & TDF_DETAILS)) >> flow_loops_dump (dump_file, NULL, 0); >> >> mark_irreducible_loops (); > > Yeah, this may also work. The reason it is not done is that > 1) it seemed expensive to force CFG changes just to compute profile decade ago > 2) cfgcleanup afterwards will anyway remove the headers again. > So I originally hoped to do the right thing without normalization. Ok ... then you should pass AVOID_CFG_MODIFICATIONS instead. And be prepared for odd situations like this ;) > Honza
On Tue, Jul 31, 2012 at 12:43 PM, Richard Guenther <richard.guenther@gmail.com> wrote: > On Tue, Jul 31, 2012 at 12:38 PM, Jan Hubicka <hubicka@ucw.cz> wrote: >>> On Tue, Jul 31, 2012 at 12:20 PM, Dehao Chen <dehao@google.com> wrote: >>> > Are you suggesting a patch like this: >>> > >>> > Index: gcc/predict.c >>> > =================================================================== >>> > --- gcc/predict.c (revision 189835) >>> > +++ gcc/predict.c (working copy) >>> > @@ -1319,6 +1319,7 @@ >>> > tree loop_bound_var = NULL; >>> > tree loop_iv_base = NULL; >>> > gimple stmt = NULL; >>> > + int header_found = 0; >>> > >>> > exits = get_loop_exit_edges (loop); >>> > n_exits = VEC_length (edge, exits); >>> > @@ -1387,9 +1388,20 @@ >>> > >>> > bbs = get_loop_body (loop); >>> > >>> > + /* Loop branch heuristics - predict an edge back to a >>> > + loop's head as taken. */ >>> > + if (loop->latch && loop->latch->loop_father == loop) >>> >>> Hmm, so the issue is that loop->latch does not belong to loop? That looks >>> like a bogus loop structure. Indeed we have the loop header of the inner >>> loop as latch of the outer loop. It still looks ok to predict this as unlikely >>> as the edge is not only the latch edge of the outer loop but also an exit >>> of the inner loop. >>> >>> Easier for profile would be to force canonicalization via >>> >>> loop_optimizer_init (LOOPS_NORMAL); >>> >>> instead of >>> >>> loop_optimizer_init (0); >>> if (dump_file && (dump_flags & TDF_DETAILS)) >>> flow_loops_dump (dump_file, NULL, 0); >>> >>> mark_irreducible_loops (); >> >> Yeah, this may also work. The reason it is not done is that >> 1) it seemed expensive to force CFG changes just to compute profile decade ago >> 2) cfgcleanup afterwards will anyway remove the headers again. >> So I originally hoped to do the right thing without normalization. > > Ok ... then you should pass AVOID_CFG_MODIFICATIONS instead. And be > prepared for odd situations like this ;) In which case the bug looks like that we predict the inner loop exit as unlikely but not the outer loop exit which should compensate things and not end up predicting zero iterations? That is, all patches seem to paper over a real issue elsewhere. Richard. >> Honza
> > > > Yeah, this may also work. The reason it is not done is that > > 1) it seemed expensive to force CFG changes just to compute profile decade ago > > 2) cfgcleanup afterwards will anyway remove the headers again. > > So I originally hoped to do the right thing without normalization. > > Ok ... then you should pass AVOID_CFG_MODIFICATIONS instead. And be > prepared for odd situations like this ;) Well, I guess we could do the extra work to avoid strange side cases like this. Does normalization fix the testcase, too? Honza > > > Honza
On Tue, Jul 31, 2012 at 6:56 PM, Jan Hubicka <hubicka@ucw.cz> wrote: >> > >> > Yeah, this may also work. The reason it is not done is that >> > 1) it seemed expensive to force CFG changes just to compute profile decade ago >> > 2) cfgcleanup afterwards will anyway remove the headers again. >> > So I originally hoped to do the right thing without normalization. >> >> Ok ... then you should pass AVOID_CFG_MODIFICATIONS instead. And be >> prepared for odd situations like this ;) > > Well, I guess we could do the extra work to avoid strange side cases like this. > Does normalization fix the testcase, too? Normalization indeed fixed this issue too. So what shall we do about this patch? Shall we simply change to use normalization instead? Thanks, Dehao > > Honza >> >> > Honza
> On Tue, Jul 31, 2012 at 6:56 PM, Jan Hubicka <hubicka@ucw.cz> wrote: > >> > > >> > Yeah, this may also work. The reason it is not done is that > >> > 1) it seemed expensive to force CFG changes just to compute profile decade ago > >> > 2) cfgcleanup afterwards will anyway remove the headers again. > >> > So I originally hoped to do the right thing without normalization. > >> > >> Ok ... then you should pass AVOID_CFG_MODIFICATIONS instead. And be > >> prepared for odd situations like this ;) > > > > Well, I guess we could do the extra work to avoid strange side cases like this. > > Does normalization fix the testcase, too? > > Normalization indeed fixed this issue too. So what shall we do about > this patch? Shall we simply change to use normalization instead? Yes, I think it is not _that_ expensive and we do some relatively tricky CFG analysis there these days that should also get bit better. (the code was written at a time CFG was new citizen to GCC and changes in CFG was hard and considered harmful :) Honza > > Thanks, > Dehao > > > > > Honza > >> > >> > Honza
Index: gcc/predict.c =================================================================== --- gcc/predict.c (revision 189835) +++ gcc/predict.c (working copy) @@ -1319,6 +1319,7 @@ tree loop_bound_var = NULL; tree loop_iv_base = NULL; gimple stmt = NULL; + int header_found = 0; exits = get_loop_exit_edges (loop); n_exits = VEC_length (edge, exits); @@ -1387,9 +1388,20 @@ bbs = get_loop_body (loop); + /* Loop branch heuristics - predict an edge back to a + loop's head as taken. */ + if (loop->latch && loop->latch->loop_father == loop) + { + edge e = find_edge (loop->latch, loop->header); + if (e) + { + header_found = 1; + predict_edge_def (e, PRED_LOOP_BRANCH, TAKEN); + } + } + for (j = 0; j < loop->num_nodes; j++) { - int header_found = 0; edge e; edge_iterator ei; @@ -1402,21 +1414,9 @@ if (predicted_by_p (bb, PRED_CONTINUE)) continue; - /* Loop branch heuristics - predict an edge back to a - loop's head as taken. */ - if (bb == loop->latch) - { - e = find_edge (loop->latch, loop->header); - if (e) - { - header_found = 1; - predict_edge_def (e, PRED_LOOP_BRANCH, TAKEN); - } - } - /* Loop exit heuristics - predict an edge exiting the loop if the conditional has no loop header successors as not taken. */ - if (!header_found + if (!(header_found && bb == loop->latch) /* If we already used more reliable loop exit predictors, do not bother with PRED_LOOP_EXIT. */ && !predicted_by_p (bb, PRED_LOOP_ITERATIONS_GUESSED)