https://gcc.gnu.org/bugzilla/show_bug.cgi?id=64072
Bug ID: 64072
Summary: wrong cgraph node profile count
Product: gcc
Version: 5.0
Status: UNCONFIRMED
Severity: normal
Priority: P3
Component: tree-optimization
Assignee: unassigned at gcc dot gnu.org
Reporter: wmi at google dot com
CC: davidxl at gcc dot gnu.org, hubicka at gcc dot gnu.org
We have a program like this:
A() { // hot func
...
}
B() {
A(); // very hot
if (i) {
A(); // very cold
}
}
Both callsites of A will be inlined into B. In gcc func
save_inline_function_body in inline_transform stage, A's first clone
will be choosen and materialized. For our case, the clone
node choosen corresponds to the cold callsite of A.
cgraph_rebuild_references in tree_function_versioning will reset the
cgraph node count of the choosen clone to the entry bb count of func A
(A is hot). So the cgraph node count of the choosen clone becomes hot
while its inline edge count is still cold. It breaks the assumption
described here:
https://gcc.gnu.org/ml/gcc-patches/2014-05/msg01366.html:
for inline node, bb->count == edge->count == edge->callee->count
For the patch committed in the thread above (it is listed below),
cg_edge->callee->count is used for profile update to its inline
instance, which leads to a hot BB in func B which is actually very
cold. The wrong profile information causes performance regression in
one of our internal benchmarks. Our internal workround is to change
cg_edge->callee->count to MIN(cg_edge->callee->count, cg_edge->count).
Index: gcc/tree-inline.c
===================================================================
--- gcc/tree-inline.c (revision 210535)
+++ gcc/tree-inline.c (working copy)
@@ -4355,7 +4355,7 @@ expand_call_inline (basic_block bb, gimple stmt, c
function in any way before this point, as this CALL_EXPR may be
a self-referential call; if we're calling ourselves, we need to
duplicate our body before altering anything. */
- copy_body (id, bb->count,
+ copy_body (id, cg_edge->callee->count,
GCOV_COMPUTE_SCALE (cg_edge->frequency, CGRAPH_FREQ_BASE),
bb, return_block, NULL);
