Hi,
this patch fixes remaining quadratic behaviour on inlining into large functions
(in gcc the avrage function we inline into has 498 callees) by avoiding
re-checking inlinability of the call edges in the caller of inline call that
was not affected by inlining.
This is not complete no-op as can be seen, say on tramp3d. It may happen that
inlining into function may make completely unrelated call in the same function
inlinable, most commonly if that is a cold call to a function whose other calls
are hot and inlined in meantime. Then inlining last call and eliminating
offline
copy may become a win in the cost metrics.
We inline such calls later, so pracitcal difference seems to be mostly about
reordering the inline decision. Resulting tramp3d binary has same size and
runtime but different function order.
I did benchmakring on GCC and Firefox and it seems that we do not lose anying
important here. I will keep eye one performance impact of this.
Inliner now spends about 60% by updating costs and 15% by actual application of
inline decisions in callgraph. That seems pretty good and we are getting slowly
to same WPA inline time as with GCC9:
- 27.78% 0.50% lto1-wpa lto1 [.]
inline_small_functions
- 27.28% inline_small_functions
- 9.90% update_callee_keys
- 5.57% update_edge_key
4.62% edge_badness
0.62% bitmap_bit_p
- 7.20% update_caller_keys
- 6.04% want_inline_small_function_p
- 5.34% do_estimate_edge_size
- 5.15% do_estimate_edge_time
- 2.70% ipa_call_context::estimate_size_and_time
- 1.74% estimate_calls_size_and_time_1
- 1.00% estimate_edge_devirt_benefit
0.92% ipa_get_indirect_edge_target_1
- 1.73% evaluate_properties_for_edge
- 0.96% evaluate_conditions_for_known_args
0.59% fold_binary_loc
0.89% can_inline_edge_p
- 4.08% inline_call
- 1.58% clone_inlined_nodes
1.15% cgraph_node::create_clone
1.57% ipa_merge_fn_summary_after_inlining
- 1.58% fibonacci_heap<sreal, cgraph_edge>::extract_minimum_node
fibonacci_heap<sreal, cgraph_edge>::consolidate
- 0.89% want_inline_small_function_p
- 0.70% do_estimate_edge_size
0.60% do_estimate_edge_time
- 0.87% estimate_growth
- 0.85% do_estimate_growth_1
0.70% do_estimate_edge_size
0.64% edge_badness
This is quite extreme case (caused by the fact that we have a lot of
autogenerated code that calls functions which we now consider for inlining and
we are now able to propagate aggregate jump functions and figure out that
resulting code will shrink or not grow too much).
Bootstrapped/regtested x86_64-linux. Will commit it shortly.
Honza
* ipa-inline.c (update_callee_keys): Add parameter UPDATE_SINCE.
(resolve_noninline_speculation, inline_small_functions): Avoid
redundant updates.
Index: ipa-inline.c
===================================================================
--- ipa-inline.c (revision 278542)
+++ ipa-inline.c (working copy)
@@ -1481,40 +1481,63 @@ update_caller_keys (edge_heap_t *heap, s
}
}
-/* Recompute HEAP nodes for each uninlined call in NODE.
+/* Recompute HEAP nodes for each uninlined call in NODE
+ If UPDATE_SINCE is non-NULL check if edges called within that function
+ are inlinable (typically UPDATE_SINCE is the inline clone we introduced
+ where all edges have new context).
+
This is used when we know that edge badnesses are going only to increase
(we introduced new call site) and thus all we need is to insert newly
created edges into heap. */
static void
update_callee_keys (edge_heap_t *heap, struct cgraph_node *node,
+ struct cgraph_node *update_since,
bitmap updated_nodes)
{
struct cgraph_edge *e = node->callees;
+ bool check_inlinability = update_since == node;
if (!e)
return;
while (true)
if (!e->inline_failed && e->callee->callees)
- e = e->callee->callees;
+ {
+ if (e->callee == update_since)
+ check_inlinability = true;
+ e = e->callee->callees;
+ }
else
{
enum availability avail;
struct cgraph_node *callee;
+ if (!check_inlinability)
+ {
+ if (e->aux
+ && !bitmap_bit_p (updated_nodes,
+ e->callee->ultimate_alias_target
+ (&avail, e->caller)->get_uid ()))
+ update_edge_key (heap, e);
+ }
/* We do not reset callee growth cache here. Since we added a new call,
growth chould have just increased and consequentely badness metric
don't need updating. */
- if (e->inline_failed
- && (callee = e->callee->ultimate_alias_target (&avail, e->caller))
- && ipa_fn_summaries->get (callee) != NULL
- && ipa_fn_summaries->get (callee)->inlinable
- && avail >= AVAIL_AVAILABLE
- && !bitmap_bit_p (updated_nodes, callee->get_uid ()))
+ else if (e->inline_failed
+ && (callee = e->callee->ultimate_alias_target (&avail,
+ e->caller))
+ && avail >= AVAIL_AVAILABLE
+ && ipa_fn_summaries->get (callee) != NULL
+ && ipa_fn_summaries->get (callee)->inlinable
+ && !bitmap_bit_p (updated_nodes, callee->get_uid ()))
{
if (can_inline_edge_p (e, false)
&& want_inline_small_function_p (e, false)
&& can_inline_edge_by_limits_p (e, false))
- update_edge_key (heap, e);
+ {
+ gcc_checking_assert (check_inlinability || can_inline_edge_p
(e, false));
+ gcc_checking_assert (check_inlinability || e->aux);
+ update_edge_key (heap, e);
+ }
else if (e->aux)
{
report_inline_failed_reason (e);
@@ -1522,6 +1545,13 @@ update_callee_keys (edge_heap_t *heap, s
e->aux = NULL;
}
}
+ /* In case we redirected to unreachable node we only need to remove the
+ fibheap entry. */
+ else if (e->aux)
+ {
+ heap->delete_node ((edge_heap_node_t *) e->aux);
+ e->aux = NULL;
+ }
if (e->next_callee)
e = e->next_callee;
else
@@ -1530,6 +1560,8 @@ update_callee_keys (edge_heap_t *heap, s
{
if (e->caller == node)
return;
+ if (e->caller == update_since)
+ check_inlinability = false;
e = e->caller->callers;
}
while (!e->next_callee);
@@ -1820,7 +1852,7 @@ resolve_noninline_speculation (edge_heap
ipa_update_overall_fn_summary (where);
update_caller_keys (edge_heap, where,
updated_nodes, NULL);
- update_callee_keys (edge_heap, where,
+ update_callee_keys (edge_heap, where, NULL,
updated_nodes);
}
}
@@ -1993,7 +2025,7 @@ inline_small_functions (void)
reset_edge_caches (where);
update_caller_keys (&edge_heap, where,
updated_nodes, NULL);
- update_callee_keys (&edge_heap, where,
+ update_callee_keys (&edge_heap, where, NULL,
updated_nodes);
bitmap_clear (updated_nodes);
}
@@ -2124,6 +2156,8 @@ inline_small_functions (void)
continue;
}
+ profile_count old_count = callee->count;
+
/* Heuristics for inlining small functions work poorly for
recursive calls where we do effects similar to loop unrolling.
When inlining such edge seems profitable, leave decision on
@@ -2147,7 +2181,7 @@ inline_small_functions (void)
at once. Consequently we need to update all callee keys. */
if (opt_for_fn (edge->caller->decl, flag_indirect_inlining))
add_new_edges_to_heap (&edge_heap, new_indirect_edges);
- update_callee_keys (&edge_heap, where, updated_nodes);
+ update_callee_keys (&edge_heap, where, where, updated_nodes);
bitmap_clear (updated_nodes);
}
else
@@ -2197,9 +2231,12 @@ inline_small_functions (void)
/* Wrapper penalty may be non-monotonous in this respect.
Fortunately it only affects small functions. */
&& !wrapper_heuristics_may_apply (where, old_size))
- update_callee_keys (&edge_heap, edge->callee, updated_nodes);
+ update_callee_keys (&edge_heap, edge->callee, edge->callee,
+ updated_nodes);
else
- update_callee_keys (&edge_heap, where, updated_nodes);
+ update_callee_keys (&edge_heap, where,
+ edge->callee,
+ updated_nodes);
}
where = edge->caller;
if (where->inlined_to)
@@ -2214,9 +2251,10 @@ inline_small_functions (void)
update_caller_keys (&edge_heap, where, updated_nodes, NULL);
/* Offline copy count has possibly changed, recompute if profile is
available. */
- struct cgraph_node *n = cgraph_node::get (edge->callee->decl);
- if (n != edge->callee && n->analyzed && n->count.ipa ().initialized_p ())
- update_callee_keys (&edge_heap, n, updated_nodes);
+ struct cgraph_node *n
+ = cgraph_node::get (edge->callee->decl)->ultimate_alias_target ();
+ if (n != edge->callee && n->analyzed && !(n->count == old_count))
+ update_callee_keys (&edge_heap, n, NULL, updated_nodes);
bitmap_clear (updated_nodes);
if (dump_enabled_p ())
