Hi,
this fixes (minor) profile updating issue which was uncovered by
somewhat devilish testcase Jakub accidentally created for branch
prediction code.
When expanding if (cond) goto xxx; gimple stmt with probability
profile_count::one into mutliple conditionals, we may end up predicting
whole chain with "guessed" quality which is unnecesary downgrade of
information we already have in CFG.
Bootstrapped/regtested x86_64-linux, comitted.
Honza
* profile-count.h (profile_count::split): Give better result when
splitting profile_probability::always.
Index: profile-count.h
===================================================================
--- profile-count.h (revision 266450)
+++ profile-count.h (working copy)
@@ -447,8 +447,12 @@ public:
{
profile_probability ret = *this * cprob;
/* The following is equivalent to:
- *this = cprob.invert () * *this / ret.invert (); */
- *this = (*this - ret) / ret.invert ();
+ *this = cprob.invert () * *this / ret.invert ();
+ Avoid scaling when overall outcome is supposed to be always.
+ Without knowing that one is inverse of toher, the result would be
+ conservative. */
+ if (!(*this == profile_probability::always ()))
+ *this = (*this - ret) / ret.invert ();
return ret;
}
