When the callers of a new clone already claim more count than the original
node, update_profiling_info lowers the original's profile. The comment
says to lower it to global0adjusted, or to local under partial training,
but both branches ran. make_profile_local was immediately overwritten by
make_profile_global0. Fixing it.
gcc/ChangeLog:
* ipa-cp.cc (update_profiling_info): Under
-fprofile-partial-training lower the original node's profile to
local instead of also forcing it to global0.
Signed-off-by: Kugan Vivekanandarajah <[email protected]>
---
gcc/ipa-cp.cc | 2 +-
1 file changed, 1 insertion(+), 1 deletion(-)
diff --git a/gcc/ipa-cp.cc b/gcc/ipa-cp.cc
index d0428d04684..829d9ab6361 100644
--- a/gcc/ipa-cp.cc
+++ b/gcc/ipa-cp.cc
@@ -4815,7 +4815,7 @@ update_profiling_info (struct cgraph_node *orig_node,
to global0adjusted or to local if we have partial training. */
if (opt_for_fn (orig_node->decl, flag_profile_partial_training))
orig_node->make_profile_local ();
- if (new_sum.quality () == AFDO)
+ else if (new_sum.quality () == AFDO)
orig_node->make_profile_global0 (GUESSED_GLOBAL0_AFDO);
else
orig_node->make_profile_global0 (GUESSED_GLOBAL0_ADJUSTED);
--
2.34.1
