Hi,
This patch makes inliner and ipa-cp to consider optimization interesting even
in scenarios where aufdo countis 0, but scaleis high enough to make optimization
worthwhile.
Bootstrapped/regtesed x86_64-linux. Comitted.
gcc/ChangeLog:
* cgraph.cc (cgraph_edge::maybe_hot_p): For AFDO profiles force
count to be non-zero.
diff --git a/gcc/cgraph.cc b/gcc/cgraph.cc
index e5902a0fdcd..72e6f54536d 100644
--- a/gcc/cgraph.cc
+++ b/gcc/cgraph.cc
@@ -3157,9 +3157,17 @@ cgraph_edge::maybe_hot_p (sreal scale)
/* If reliable IPA count is available, just use it. */
profile_count c = count.ipa ();
- if (c.reliable_p ())
+ if (c.reliable_p ()
+ || (c.quality () == AFDO && c.nonzero_p ()))
return maybe_hot_count_p (NULL, c * scale);
+ /* In auto-FDO, count 0 may lead to hot code in case the
+ call is simply not called often enough to receive some samples. */
+ if ((c.quality () == AFDO
+ || count.quality () == GUESSED_GLOBAL0_ADJUSTED)
+ && callee && callee->count.quality () == AFDO)
+ return maybe_hot_count_p (NULL, c.force_nonzero () * scale);
+
/* See if we can determine hotness using caller frequency. */
if (caller->frequency == NODE_FREQUENCY_UNLIKELY_EXECUTED
|| (callee
