| Message ID | CAMe9rOomcm3rt-pTxZ_B_1NrQgp2kf4yfRL2A21mFpP-6ULDfA@mail.gmail.com |
|---|---|
| State | New |
| Headers | show |
> > This patch works passed profiledbootstrap with LTO as well as LTO -O3 > on 176.gcc in SPEC CPU 2000. I have to add 2 inline_edge_summary_vec > checks to avoid ICE. OK to install? Thanks, it looks good. I am just concerned about... > diff --git a/gcc/ipa-split.c b/gcc/ipa-split.c > index 33cf7d2..7a8844f 100644 > --- a/gcc/ipa-split.c > +++ b/gcc/ipa-split.c > @@ -1415,7 +1415,7 @@ execute_split_functions (void) > } > /* This can be relaxed; function might become inlinable after splitting > away the uninlinable part. */ > - if (!inline_summary (node)->inlinable) > + if (inline_edge_summary_vec && !inline_summary (node)->inlinable) .. this one. spliting is executed before free_inline_summary and thus should not be affected. Or is it because of it gets called from process_new_functions because some IPA pass adds a new function? If so, I think we need to make sure that process_new_function do not make inline summary allocated when it was previously free (and thus hunk would be OK). Honza
On Thu, Aug 2, 2012 at 3:23 AM, Jan Hubicka <hubicka@ucw.cz> wrote: >> >> This patch works passed profiledbootstrap with LTO as well as LTO -O3 >> on 176.gcc in SPEC CPU 2000. I have to add 2 inline_edge_summary_vec >> checks to avoid ICE. OK to install? > > Thanks, it looks good. I am just concerned about... >> diff --git a/gcc/ipa-split.c b/gcc/ipa-split.c >> index 33cf7d2..7a8844f 100644 >> --- a/gcc/ipa-split.c >> +++ b/gcc/ipa-split.c >> @@ -1415,7 +1415,7 @@ execute_split_functions (void) >> } >> /* This can be relaxed; function might become inlinable after splitting >> away the uninlinable part. */ >> - if (!inline_summary (node)->inlinable) >> + if (inline_edge_summary_vec && !inline_summary (node)->inlinable) > > .. this one. spliting is executed before free_inline_summary and thus should > not be affected. Or is it because of it gets called from process_new_functions > because some IPA pass adds a new function? It is called from pass_feedback_split_functions: NEXT_PASS (pass_all_early_optimizations); { ... } NEXT_PASS (pass_ipa_free_inline_summary); NEXT_PASS (pass_ipa_tree_profile); { struct opt_pass **p = &pass_ipa_tree_profile.pass.sub; NEXT_PASS (pass_feedback_split_functions); } OK to install? Thanks.
> On Thu, Aug 2, 2012 at 3:23 AM, Jan Hubicka <hubicka@ucw.cz> wrote: > >> > >> This patch works passed profiledbootstrap with LTO as well as LTO -O3 > >> on 176.gcc in SPEC CPU 2000. I have to add 2 inline_edge_summary_vec > >> checks to avoid ICE. OK to install? > > > > Thanks, it looks good. I am just concerned about... > >> diff --git a/gcc/ipa-split.c b/gcc/ipa-split.c > >> index 33cf7d2..7a8844f 100644 > >> --- a/gcc/ipa-split.c > >> +++ b/gcc/ipa-split.c > >> @@ -1415,7 +1415,7 @@ execute_split_functions (void) > >> } > >> /* This can be relaxed; function might become inlinable after splitting > >> away the uninlinable part. */ > >> - if (!inline_summary (node)->inlinable) > >> + if (inline_edge_summary_vec && !inline_summary (node)->inlinable) > > > > .. this one. spliting is executed before free_inline_summary and thus should > > not be affected. Or is it because of it gets called from process_new_functions > > because some IPA pass adds a new function? > > It is called from pass_feedback_split_functions: > > NEXT_PASS (pass_all_early_optimizations); > { > ... > } > NEXT_PASS (pass_ipa_free_inline_summary); > NEXT_PASS (pass_ipa_tree_profile); > { > struct opt_pass **p = &pass_ipa_tree_profile.pass.sub; > NEXT_PASS (pass_feedback_split_functions); > } > > OK to install? Hmm, this is an ordering issue, but minnor one. OK. As comment says, this should be relaxed anyway and in fact I have path for that in queue. Honza > > Thanks. > > -- > H.J.
diff --git a/gcc/ipa-inline-analysis.c b/gcc/ipa-inline-analysis.c index ae32131..970be1e 100644 --- a/gcc/ipa-inline-analysis.c +++ b/gcc/ipa-inline-analysis.c @@ -3242,6 +3242,8 @@ void inline_free_summary (void) { struct cgraph_node *node; + if (inline_edge_summary_vec == NULL) + return; FOR_EACH_DEFINED_FUNCTION (node) reset_inline_summary (node); if (function_insertion_hook_holder) diff --git a/gcc/ipa-split.c b/gcc/ipa-split.c index 33cf7d2..7a8844f 100644 --- a/gcc/ipa-split.c +++ b/gcc/ipa-split.c @@ -1415,7 +1415,7 @@ execute_split_functions (void) } /* This can be relaxed; function might become inlinable after splitting away the uninlinable part. */ - if (!inline_summary (node)->inlinable) + if (inline_edge_summary_vec && !inline_summary (node)->inlinable) { if (dump_file) fprintf (dump_file, "Not splitting: not inlinable.\n"); diff --git a/gcc/ipa.c b/gcc/ipa.c index 9329d9b..e270591 100644 --- a/gcc/ipa.c +++ b/gcc/ipa.c @@ -448,6 +448,11 @@ symtab_remove_unreachable_nodes (bool before_inlining_p, FILE *file) verify_symtab (); #endif + /* If we removed something, perhaps profile could be improved. */ + if (changed && optimize && inline_edge_summary_vec) + FOR_EACH_DEFINED_FUNCTION (node) + cgraph_propagate_frequency (node); + return changed; } @@ -960,6 +965,34 @@ struct simple_ipa_opt_pass pass_ipa_function_and_variable_visibility = } }; +/* Free inline summary. */ + +static unsigned +free_inline_summary (void) +{ + inline_free_summary (); + return 0; +} + +struct simple_ipa_opt_pass pass_ipa_free_inline_summary = +{ + { + SIMPLE_IPA_PASS, + "*free_inline_summary", /* name */ + NULL, /* gate */ + free_inline_summary, /* execute */ + NULL, /* sub */ + NULL, /* next */ + 0, /* static_pass_number */ + TV_IPA_FREE_INLINE_SUMMARY, /* tv_id */ + 0, /* properties_required */ + 0, /* properties_provided */ + 0, /* properties_destroyed */ + 0, /* todo_flags_start */ + TODO_ggc_collect /* todo_flags_finish */ + } +}; + /* Do not re-run on ltrans stage. */ static bool diff --git a/gcc/passes.c b/gcc/passes.c index 7709671..fc4dbc5 100644 --- a/gcc/passes.c +++ b/gcc/passes.c @@ -1324,6 +1324,7 @@ init_optimization_passes (void) NEXT_PASS (pass_rebuild_cgraph_edges); NEXT_PASS (pass_inline_parameters); } + NEXT_PASS (pass_ipa_free_inline_summary); NEXT_PASS (pass_ipa_tree_profile); { struct opt_pass **p = &pass_ipa_tree_profile.pass.sub; diff --git a/gcc/timevar.def b/gcc/timevar.def index 18d419d..8fa1a2b 100644 --- a/gcc/timevar.def +++ b/gcc/timevar.def @@ -89,6 +89,7 @@ DEFTIMEVAR (TV_IPA_PURE_CONST , "ipa pure const") DEFTIMEVAR (TV_IPA_PTA , "ipa points-to") DEFTIMEVAR (TV_IPA_SRA , "ipa SRA") DEFTIMEVAR (TV_IPA_FREE_LANG_DATA , "ipa free lang data") +DEFTIMEVAR (TV_IPA_FREE_INLINE_SUMMARY, "ipa free inline summary") /* Time spent by constructing CFG. */ DEFTIMEVAR (TV_CFG , "cfg construction") /* Time spent by cleaning up CFG. */ diff --git a/gcc/tree-pass.h b/gcc/tree-pass.h index 081e48c..5037c29 100644 --- a/gcc/tree-pass.h +++ b/gcc/tree-pass.h @@ -377,6 +377,7 @@ extern struct simple_ipa_opt_pass pass_ipa_increase_alignment; extern struct simple_ipa_opt_pass pass_ipa_matrix_reorg; extern struct ipa_opt_pass_d pass_ipa_inline; extern struct simple_ipa_opt_pass pass_ipa_free_lang_data; +extern struct simple_ipa_opt_pass pass_ipa_free_inline_summary; extern struct ipa_opt_pass_d pass_ipa_cp; extern struct ipa_opt_pass_d pass_ipa_reference;