This patch fixes ICE seen when building spec2k17 with autofdo and enable
checking compiler. Bause we special case 0 of autofdo to be kind of 1 in IPA
scalling, we can now end up with function heving global0 profile but producing
inline clone with nonzero profile.
I think correct way is to extend auto-profile generation to merge static profile
with afdo in places afdo profile is missing and drop the autofdo 0 hacks, but
I think we need to run benchmarks first before making broader changes here.
Bootstrapped/regtested x86_64-linux, comitted.
* profile-count.cc (profile_count::to_sreal_scale): Special case 0 of
autofdo.
(profile_count::combine_with_ipa_count): If outer function has GLOBAL0
profile
but innter counter has non-zero profile, force it to be 0.
diff --git a/gcc/profile-count.cc b/gcc/profile-count.cc
index 2d9c778b375..22c109ab528 100644
--- a/gcc/profile-count.cc
+++ b/gcc/profile-count.cc
@@ -344,6 +344,10 @@ profile_count::to_sreal_scale (profile_count in, bool
*known) const
return 1;
if (!in.m_val)
return m_val * 4;
+ /* Auto-FDO 0 really just means that we have no samples.
+ Treat it as small non-zero frequency. */
+ if (!m_val && quality () == AFDO)
+ return (sreal)1 / (sreal)in.m_val;
return (sreal)m_val / (sreal)in.m_val;
}
@@ -398,7 +402,7 @@ profile_count::combine_with_ipa_count (profile_count ipa)
return this->global0adjusted ();
}
-/* Sae as profile_count::combine_with_ipa_count but within function with count
+/* Same as profile_count::combine_with_ipa_count but within function with count
IPA2. */
profile_count
profile_count::combine_with_ipa_count_within (profile_count ipa,
@@ -410,7 +414,16 @@ profile_count::combine_with_ipa_count_within
(profile_count ipa,
if (ipa2.ipa () == ipa2 && ipa.initialized_p ())
ret = ipa;
else
- ret = combine_with_ipa_count (ipa);
+ {
+ /* For inconsistent profiles we may end up having ipa2 of GLOBAL0
+ while ipa is non-zero (i.e. non-zero IPA counters within function
+ executed 0 times). Be sure we produce GLOBAL0 as well
+ so counters remain compatible. */
+ if (ipa.nonzero_p ()
+ && ipa2.ipa ().initialized_p ())
+ ipa = ipa2.ipa ();
+ ret = combine_with_ipa_count (ipa);
+ }
gcc_checking_assert (ret.compatible_p (ipa2));
return ret;
}