https://gcc.gnu.org/bugzilla/show_bug.cgi?id=108696
Bug ID: 108696
Summary: querying relations is slow
Product: gcc
Version: 13.0
Status: UNCONFIRMED
Severity: normal
Priority: P3
Component: tree-optimization
Assignee: unassigned at gcc dot gnu.org
Reporter: rguenth at gcc dot gnu.org
Target Milestone: ---
With the compiler.i testcase from PR26854 I notice we do quite a lot of bitmap
intersections via dom_oracle::query_relation for the case (which hits 99% of
the time) when a SSA name doesn't have any relation but equiv_set () returns a
newly allocated bitmap with just the self-equivalence recorded.
The self-equivalence case isn't used to prune the work done by query_relation
it seems. Maybe find_relation_dom does that via
// IF either name does not occur in a relation anywhere, there isnt one.
if (!bitmap_bit_p (m_relation_set, v1) || !bitmap_bit_p (m_relation_set, v2))
return VREL_VARYING;
but we still get all the way to the query_relation overload which eventually
does the intersection with m_relation_set which is a lot more expensive than
this.
A naiive
diff --git a/gcc/value-relation.cc b/gcc/value-relation.cc
index f5b1e67e420..f58d5050bdb 100644
--- a/gcc/value-relation.cc
+++ b/gcc/value-relation.cc
@@ -1373,6 +1373,10 @@ dom_oracle::query_relation (basic_block bb, tree ssa1,
tree ssa2)
unsigned v2 = SSA_NAME_VERSION (ssa2);
if (v1 == v2)
return VREL_EQ;
+
+ // IF either name does not occur in a relation anywhere, there isnt one.
+ if (!bitmap_bit_p (m_relation_set, v1) || !bitmap_bit_p (m_relation_set,
v2))
+ return VREL_VARYING;
// Check for equivalence first. They must be in each equivalency set.
const_bitmap equiv1 = equiv_set (ssa1, bb);
cuts
dominator optimization : 17.35 ( 15%)
down to
dominator optimization : 16.17 ( 14%)
note almost all of the DOM time is spent in ranger for this testcase.
