[RFC, PATCH] LTO: IPA inline speed up for large apps (Chrome)
Martin Liška
mliska@suse.cz
Wed Feb 18 14:13:00 GMT 2015
On 02/18/2015 02:58 PM, Martin Liška wrote:
> On 02/17/2015 10:03 PM, Jan Hubicka wrote:
>> Hi,
>> this patch should chase away the expensive thunks and aliases walks from most
>> of analysis code. I think only real use left is local_p predicate that needs to
>> stay because i386 expect local flag to match between caller and callee when
>> expanding assembler thunk. I at least optimized it by first moving the walk to
>> be conditional for nonlocal functions only and then reorganizing
>> call_for_symbol_thunks_and_aliases to first inspect aliases (that is cheap) and
>> only then work on thunks. Most likely this will find the non-local thunk/alias
>> faster. Other cases was leftovers from the conversion of thunks from aliases
>> to functions.
>>
>> I also noticed a bug in ipa-profile that does not disable all the
>> transofrms with !ipa_profile_flag used on OPTIMIZTION_NODE and fixed it.
>>
>> Bootstrapped/regtested x86_64-linux, comitted. I would be interested to
>> know if the call_for_symbol_thunks_and_aliases is now off your oprofiles
>> (sorry, easier to type than perf-profiles)
>>
>> Honza
>>
>> * ipa-visibility.c (function_and_variable_visibility): Only
>> check locality if node is not already local.
>> * ipa-inline.c (want_inline_function_to_all_callers_p): Use
>> call_for_symbol_and_aliases instead of
>> call_for_symbol_thunks_and_aliases.
>> (ipa_inline): Likewise.
>> * cgraph.c (cgraph_node::call_for_symbol_thunks_and_aliases):
>> first walk aliases.
>> * ipa.c (symbol_table::remove_unreachable_nodes): Use
>> call_for_symbol_and_aliases.
>> * ipa-profile.c (ipa_propagate_frequency_data): Add function_symbol.
>> (ipa_propagate_frequency_1): Use it; use opt_for_fn
>> (ipa_propagate_frequency): Update.
>> (ipa_profile): Add opt_for_fn gueards.
>> Index: ipa-visibility.c
>> ===================================================================
>> --- ipa-visibility.c (revision 220741)
>> +++ ipa-visibility.c (working copy)
>> @@ -595,7 +595,8 @@ function_and_variable_visibility (bool w
>> }
>> FOR_EACH_DEFINED_FUNCTION (node)
>> {
>> - node->local.local |= node->local_p ();
>> + if (!node->local.local)
>> + node->local.local |= node->local_p ();
>>
>> /* If we know that function can not be overwritten by a different semantics
>> and moreover its section can not be discarded, replace all direct calls
>> Index: ipa-inline.c
>> ===================================================================
>> --- ipa-inline.c (revision 220741)
>> +++ ipa-inline.c (working copy)
>> @@ -975,14 +975,14 @@ want_inline_function_to_all_callers_p (s
>> if (node->global.inlined_to)
>> return false;
>> /* Does it have callers? */
>> - if (!node->call_for_symbol_thunks_and_aliases (has_caller_p, NULL, true))
>> + if (!node->call_for_symbol_and_aliases (has_caller_p, NULL, true))
>> return false;
>> /* Inlining into all callers would increase size? */
>> if (estimate_growth (node) > 0)
>> return false;
>> /* All inlines must be possible. */
>> - if (node->call_for_symbol_thunks_and_aliases (check_callers, &has_hot_call,
>> - true))
>> + if (node->call_for_symbol_and_aliases (check_callers, &has_hot_call,
>> + true))
>> return false;
>> if (!cold && !has_hot_call)
>> return false;
>> @@ -2359,9 +2359,9 @@ ipa_inline (void)
>> if (want_inline_function_to_all_callers_p (node, cold))
>> {
>> int num_calls = 0;
>> - node->call_for_symbol_thunks_and_aliases (sum_callers, &num_calls,
>> - true);
>> - while (node->call_for_symbol_thunks_and_aliases
>> + node->call_for_symbol_and_aliases (sum_callers, &num_calls,
>> + true);
>> + while (node->call_for_symbol_and_aliases
>> (inline_to_all_callers, &num_calls, true))
>> ;
>> remove_functions = true;
>> Index: cgraph.c
>> ===================================================================
>> --- cgraph.c (revision 220741)
>> +++ cgraph.c (working copy)
>> @@ -2191,6 +2191,16 @@ cgraph_node::call_for_symbol_thunks_and_
>>
>> if (callback (this, data))
>> return true;
>> + FOR_EACH_ALIAS (this, ref)
>> + {
>> + cgraph_node *alias = dyn_cast <cgraph_node *> (ref->referring);
>> + if (include_overwritable
>> + || alias->get_availability () > AVAIL_INTERPOSABLE)
>> + if (alias->call_for_symbol_thunks_and_aliases (callback, data,
>> + include_overwritable,
>> + exclude_virtual_thunks))
>> + return true;
>> + }
>> for (e = callers; e; e = e->next_caller)
>> if (e->caller->thunk.thunk_p
>> && (include_overwritable
>> @@ -2202,16 +2212,6 @@ cgraph_node::call_for_symbol_thunks_and_
>> exclude_virtual_thunks))
>> return true;
>>
>> - FOR_EACH_ALIAS (this, ref)
>> - {
>> - cgraph_node *alias = dyn_cast <cgraph_node *> (ref->referring);
>> - if (include_overwritable
>> - || alias->get_availability () > AVAIL_INTERPOSABLE)
>> - if (alias->call_for_symbol_thunks_and_aliases (callback, data,
>> - include_overwritable,
>> - exclude_virtual_thunks))
>> - return true;
>> - }
>> return false;
>> }
>>
>> Index: ipa.c
>> ===================================================================
>> --- ipa.c (revision 220741)
>> +++ ipa.c (working copy)
>> @@ -661,7 +661,7 @@ symbol_table::remove_unreachable_nodes (
>> if (node->address_taken
>> && !node->used_from_other_partition)
>> {
>> - if (!node->call_for_symbol_thunks_and_aliases
>> + if (!node->call_for_symbol_and_aliases
>> (has_addr_references_p, NULL, true)
>> && (!node->instrumentation_clone
>> || !node->instrumented_version
>> Index: ipa-profile.c
>> ===================================================================
>> --- ipa-profile.c (revision 220741)
>> +++ ipa-profile.c (working copy)
>> @@ -322,6 +322,7 @@ ipa_profile_read_summary (void)
>>
>> struct ipa_propagate_frequency_data
>> {
>> + cgraph_node *function_symbol;
>> bool maybe_unlikely_executed;
>> bool maybe_executed_once;
>> bool only_called_at_startup;
>> @@ -342,7 +343,7 @@ ipa_propagate_frequency_1 (struct cgraph
>> || d->only_called_at_startup || d->only_called_at_exit);
>> edge = edge->next_caller)
>> {
>> - if (edge->caller != node)
>> + if (edge->caller != d->function_symbol)
>> {
>> d->only_called_at_startup &= edge->caller->only_called_at_startup;
>> /* It makes sense to put main() together with the static constructors.
>> @@ -358,7 +359,11 @@ ipa_propagate_frequency_1 (struct cgraph
>> errors can make us to push function into unlikely section even when
>> it is executed by the train run. Transfer the function only if all
>> callers are unlikely executed. */
>> - if (profile_info && flag_branch_probabilities
>> + if (profile_info
>> + && opt_for_fn (d->function_symbol->decl, flag_branch_probabilities)
>> + /* Thunks are not profiled. This is more or less implementation
>> + bug. */
>> + && !d->function_symbol->thunk.thunk_p
>> && (edge->caller->frequency != NODE_FREQUENCY_UNLIKELY_EXECUTED
>> || (edge->caller->global.inlined_to
>> && edge->caller->global.inlined_to->frequency
>> @@ -418,7 +423,7 @@ contains_hot_call_p (struct cgraph_node
>> bool
>> ipa_propagate_frequency (struct cgraph_node *node)
>> {
>> - struct ipa_propagate_frequency_data d = {true, true, true, true};
>> + struct ipa_propagate_frequency_data d = {node, true, true, true, true};
>> bool changed = false;
>>
>> /* We can not propagate anything useful about externally visible functions
>> @@ -432,8 +437,8 @@ ipa_propagate_frequency (struct cgraph_n
>> if (dump_file && (dump_flags & TDF_DETAILS))
>> fprintf (dump_file, "Processing frequency %s\n", node->name ());
>>
>> - node->call_for_symbol_thunks_and_aliases (ipa_propagate_frequency_1, &d,
>> - true);
>> + node->call_for_symbol_and_aliases (ipa_propagate_frequency_1, &d,
>> + true);
>>
>> if ((d.only_called_at_startup && !d.only_called_at_exit)
>> && !node->only_called_at_startup)
>> @@ -597,6 +602,9 @@ ipa_profile (void)
>> {
>> bool update = false;
>>
>> + if (!opt_for_fn (n->decl, flag_ipa_profile))
>> + continue;
>> +
>> for (e = n->indirect_calls; e; e = e->next_callee)
>> {
>> if (n->count)
>> @@ -697,7 +705,9 @@ ipa_profile (void)
>> order_pos = ipa_reverse_postorder (order);
>> for (i = order_pos - 1; i >= 0; i--)
>> {
>> - if (order[i]->local.local && ipa_propagate_frequency (order[i]))
>> + if (order[i]->local.local
>> + && opt_for_fn (order[i]->decl, flag_ipa_profile)
>> + && ipa_propagate_frequency (order[i]))
>> {
>> for (e = order[i]->callees; e; e = e->next_callee)
>> if (e->callee->local.local && !e->callee->aux)
>> @@ -714,7 +724,9 @@ ipa_profile (void)
>> something_changed = false;
>> for (i = order_pos - 1; i >= 0; i--)
>> {
>> - if (order[i]->aux && ipa_propagate_frequency (order[i]))
>> + if (order[i]->aux
>> + && opt_for_fn (order[i]->decl, flag_ipa_profile)
>> + && ipa_propagate_frequency (order[i]))
>> {
>> for (e = order[i]->callees; e; e = e->next_callee)
>> if (e->callee->local.local && !e->callee->aux)
>>
>
> Hi.
>
> There's perf report and -ftime report of WPA phase.
>
> Martin
Hm, using the same compiler, Firefox LTO time statistics and perf report and very different.
I'm wondering how can be that possible?
Martin
-------------- next part --------------
Execution times (seconds)
phase setup : 0.00 ( 0%) usr 0.00 ( 0%) sys 0.01 ( 0%) wall 1988 kB ( 0%) ggc
phase opt and generate : 42.32 (70%) usr 0.85 (56%) sys 43.16 (69%) wall 1387464 kB (28%) ggc
phase stream in : 18.50 (30%) usr 0.68 (44%) sys 19.17 (31%) wall 3528077 kB (72%) ggc
garbage collection : 2.24 ( 4%) usr 0.00 ( 0%) sys 2.24 ( 4%) wall 0 kB ( 0%) ggc
callgraph optimization : 0.37 ( 1%) usr 0.00 ( 0%) sys 0.37 ( 1%) wall 38 kB ( 0%) ggc
ipa dead code removal : 3.06 ( 5%) usr 0.01 ( 1%) sys 2.88 ( 5%) wall 0 kB ( 0%) ggc
ipa virtual call target : 5.72 ( 9%) usr 0.06 ( 4%) sys 5.87 ( 9%) wall 0 kB ( 0%) ggc
ipa devirtualization : 0.18 ( 0%) usr 0.00 ( 0%) sys 0.23 ( 0%) wall 22382 kB ( 0%) ggc
ipa cp : 2.88 ( 5%) usr 0.09 ( 6%) sys 2.97 ( 5%) wall 515623 kB (10%) ggc
ipa inlining heuristics : 13.96 (23%) usr 0.13 ( 8%) sys 14.12 (23%) wall 471848 kB (10%) ggc
ipa comdats : 0.12 ( 0%) usr 0.00 ( 0%) sys 0.12 ( 0%) wall 0 kB ( 0%) ggc
ipa lto gimple in : 2.54 ( 4%) usr 0.48 (31%) sys 3.23 ( 5%) wall 645652 kB (13%) ggc
ipa lto decl in : 12.64 (21%) usr 0.37 (24%) sys 13.01 (21%) wall 2592737 kB (53%) ggc
ipa lto constructors in : 0.17 ( 0%) usr 0.01 ( 1%) sys 0.20 ( 0%) wall 16493 kB ( 0%) ggc
ipa lto cgraph I/O : 0.58 ( 1%) usr 0.09 ( 6%) sys 0.67 ( 1%) wall 437504 kB ( 9%) ggc
ipa lto decl merge : 1.90 ( 3%) usr 0.00 ( 0%) sys 1.90 ( 3%) wall 8191 kB ( 0%) ggc
ipa lto cgraph merge : 1.30 ( 2%) usr 0.00 ( 0%) sys 1.29 ( 2%) wall 14989 kB ( 0%) ggc
whopr wpa : 0.91 ( 1%) usr 0.00 ( 0%) sys 0.88 ( 1%) wall 2 kB ( 0%) ggc
whopr partitioning : 2.66 ( 4%) usr 0.00 ( 0%) sys 2.67 ( 4%) wall 6081 kB ( 0%) ggc
ipa reference : 1.38 ( 2%) usr 0.01 ( 1%) sys 1.40 ( 2%) wall 0 kB ( 0%) ggc
ipa profile : 0.21 ( 0%) usr 0.01 ( 1%) sys 0.21 ( 0%) wall 0 kB ( 0%) ggc
ipa pure const : 1.61 ( 3%) usr 0.01 ( 1%) sys 1.61 ( 3%) wall 0 kB ( 0%) ggc
ipa icf : 4.99 ( 8%) usr 0.06 ( 4%) sys 5.00 ( 8%) wall 1120 kB ( 0%) ggc
tree SSA rewrite : 0.12 ( 0%) usr 0.02 ( 1%) sys 0.12 ( 0%) wall 23170 kB ( 0%) ggc
tree SSA incremental : 0.23 ( 0%) usr 0.05 ( 3%) sys 0.21 ( 0%) wall 14434 kB ( 0%) ggc
tree operand scan : 0.14 ( 0%) usr 0.03 ( 2%) sys 0.22 ( 0%) wall 145252 kB ( 3%) ggc
dominance frontiers : 0.04 ( 0%) usr 0.00 ( 0%) sys 0.01 ( 0%) wall 0 kB ( 0%) ggc
dominance computation : 0.14 ( 0%) usr 0.05 ( 3%) sys 0.11 ( 0%) wall 0 kB ( 0%) ggc
varconst : 0.01 ( 0%) usr 0.02 ( 1%) sys 0.03 ( 0%) wall 0 kB ( 0%) ggc
loop fini : 0.07 ( 0%) usr 0.00 ( 0%) sys 0.03 ( 0%) wall 0 kB ( 0%) ggc
unaccounted todo : 0.62 ( 1%) usr 0.00 ( 0%) sys 0.65 ( 1%) wall 0 kB ( 0%) ggc
TOTAL : 60.82 1.53 62.34 4917531 kB
[ perf record: Woken up 59 times to write data ]
[ perf record: Captured and wrote 14.722 MB perf.data (~643202 samples) ]
marxin@marxinbox:~/Programming/gecko-dev/obj-x86_64-unknown-linux-gnu/toolkit/library> perf report
marxin@marxinbox:~/Programming/gecko-dev/obj-x86_64-unknown-linux-gnu/toolkit/library> gcc -v
Using built-in specs.
COLLECT_GCC=gcc
COLLECT_LTO_WRAPPER=/home/marxin/Programming/bin/gcc2/lib/gcc/x86_64-unknown-linux-gnu/5.0.0/lto-wrapper
Target: x86_64-unknown-linux-gnu
Configured with: ../configure --enable-languages=c,c++ --disable-libsanitizer --prefix=/home/marxin/Programming/bin/gcc2 --disable-bootstrap --enable-checking=release
Thread model: posix
gcc version 5.0.0 20150218 (experimental) (GCC)
marxin@marxinbox:~/Programming/gecko-dev/obj-x86_64-unknown-linux-gnu/toolkit/library> perf report
marxin@marxinbox:~/Programming/gecko-dev/obj-x86_64-unknown-linux-gnu/toolkit/library> perf report --stdio | sed 's/\ *$//' | head -n50
# To display the perf.data header info, please use --header/--header-only options.
#
# Samples: 245K of event 'cycles'
# Event count (approx.): 216467422123
#
# Overhead Command Shared Object
# ........ ........ ................. ..................................................................................................................................................................................................................................................................................................
#
4.97% lto1-wpa lto1 [.] inflate_fast
2.78% lto1-wpa lto1 [.] symbol_table::remove_unreachable_nodes(_IO_FILE*)
2.37% lto1-wpa libc-2.19.so [.] _int_malloc
1.77% lto1-wpa lto1 [.] record_target_from_binfo(vec<cgraph_node*, va_heap, vl_ptr>&, vec<tree_node*, va_heap, vl_ptr>*, tree_node*, tree_node*, vec<tree_node*, va_heap, vl_ptr>&, long, tree_node*, long, hash_set<tree_node*, default_hashset_traits>*, hash_set<tree_node*, default_hashset_traits>*, bool, bool*)
1.57% lto1-wpa lto1 [.] ht_lookup_with_hash(ht*, unsigned char const*, unsigned long, unsigned int, ht_lookup_option)
1.56% lto1-wpa lto1 [.] streamer_read_uhwi(lto_input_block*)
1.48% lto1-wpa lto1 [.] estimate_calls_size_and_time(cgraph_node*, int*, int*, int*, int*, unsigned int, vec<tree_node*, va_heap, vl_ptr>, vec<ipa_polymorphic_call_context, va_heap, vl_ptr>, vec<ipa_agg_jump_function*, va_heap, vl_ptr>) [clone .isra.129]
1.48% lto1-wpa lto1 [.] unify_scc(streamer_tree_cache_d*, unsigned int, unsigned int, unsigned int, unsigned int)
1.40% lto1-wpa lto1 [.] lto_cgraph_replace_node(cgraph_node*, cgraph_node*)
1.38% lto1-wpa lto1 [.] ggc_set_mark(void const*)
1.30% lto1-wpa libc-2.19.so [.] malloc_consolidate
1.28% lto1-wpa lto1 [.] htab_hash_string
1.25% lto1-wpa lto1 [.] compare_tree_sccs_1(tree_node*, tree_node*, tree_node***)
1.23% lto1-wpa lto1 [.] fibonacci_heap<sreal, cgraph_edge>::consolidate()
1.19% lto1-wpa lto1 [.] splay_tree_splay
1.15% lto1-wpa lto1 [.] can_inline_edge_p(cgraph_edge*, bool, bool)
1.14% lto1-wpa lto1 [.] cgraph_node::get_availability()
1.14% lto1-wpa lto1 [.] evaluate_properties_for_edge(cgraph_edge*, bool, unsigned int*, vec<tree_node*, va_heap, vl_ptr>*, vec<ipa_polymorphic_call_context, va_heap, vl_ptr>*, vec<ipa_agg_jump_function*, va_heap, vl_ptr>*) [clone .constprop.131]
1.13% lto1-wpa lto1 [.] gimple_get_virt_method_for_vtable(long, tree_node*, unsigned long, bool*)
1.10% lto1-wpa lto1 [.] types_same_for_odr(tree_node const*, tree_node const*)
1.08% lto1-wpa lto1 [.] gt_ggc_mx_lang_tree_node(void*)
1.05% lto1-wpa lto1 [.] streamer_read_tree_bitfields(lto_input_block*, data_in*, tree_node*)
0.99% lto1-wpa lto1 [.] type_in_anonymous_namespace_p(tree_node const*)
0.99% lto1-wpa lto1 [.] gimple_has_body_p(tree_node*)
0.95% lto1-wpa lto1 [.] decl_assembler_name(tree_node*)
0.93% lto1-wpa lto1 [.] do_per_function(void (*)(function*, void*), void*)
0.82% lto1-wpa libc-2.19.so [.] _int_free
0.81% lto1-wpa lto1 [.] possible_polymorphic_call_targets_1(vec<cgraph_node*, va_heap, vl_ptr>&, hash_set<tree_node*, default_hashset_traits>*, hash_set<tree_node*, default_hashset_traits>*, tree_node*, odr_type_d*, long, tree_node*, long, bool*, vec<tree_node*, va_heap, vl_ptr>&, bool)
0.81% lto1-wpa lto1 [.] searchc(searchc_env*, cgraph_node*, bool (*)(cgraph_edge*))
0.80% lto1-wpa lto1 [.] streamer_get_pickled_tree(lto_input_block*, data_in*)
0.78% lto1-wpa lto1 [.] edge_badness(cgraph_edge*, bool)
0.77% lto1-wpa lto1 [.] hash_table<asmname_hasher, xcallocator, true>::find_slot_with_hash(tree_node const* const&, unsigned int, insert_option)
0.77% lto1-wpa lto1 [.] update_callee_keys(fibonacci_heap<sreal, cgraph_edge>*, cgraph_node*, bitmap_head*)
0.76% lto1-wpa lto1 [.] ggc_internal_alloc(unsigned long, void (*)(void*), unsigned long, unsigned long)
0.75% lto1-wpa lto1 [.] fibonacci_heap<sreal, cgraph_edge>::extract_minimum_node()
0.75% lto1-wpa lto1 [.] execute_one_pass(opt_pass*)
0.74% lto1-wpa lto1 [.] inflate
0.71% lto1-wpa lto1 [.] contains_polymorphic_type_p(tree_node const*)
0.67% lto1-wpa lto1 [.] get_binfo_at_offset(tree_node*, long, tree_node*)
0.64% lto1-wpa lto1 [.] symbol_table::decl_assembler_name_equal(tree_node*, tree_node const*)
0.61% lto1-wpa lto1 [.] lto_balanced_map(int)
0.61% lto1-wpa lto1 [.] ipa_icf::sem_item_optimizer::do_congruence_step_for_index(ipa_icf::congruence_class*, unsigned int)
More information about the Gcc-patches
mailing list