[Bug rtl-optimization/66351] [6 regression] r223883 miscompiles stage2 compiler on ia64

[hubicka at gcc dot gnu.org]

https://gcc.gnu.org/bugzilla/show_bug.cgi?id=66351

--- Comment #1 from Jan Hubicka <hubicka at gcc dot gnu.org> ---
Hello,
the following patch should fix the miscompilation:
Index: ipa-polymorphic-call.c
===================================================================
--- ipa-polymorphic-call.c      (revision 224053)
+++ ipa-polymorphic-call.c      (working copy)
@@ -1602,6 +1603,8 @@
                    }
                }
            }
+         if (!instance_ref)
+           return false;
        }
     }


this issue here is that for OBJ_TYPE_REF we start the alias oracle walk looking
for vtbl change from wrong instruction - it should not start by call, it should
start by load of vptr. This path is originally copied from ipa-prop code so it
was in for a while. 

Reason why i did not commit the patch to mianline yet is that I do not 100%
understand why it breaks in case of ipa-icf.c because the load should alias
with the vptr store, too. What happens is that the call is devirtualized by GVN
first, so perhaps we turn the call to pure and lose the link, but I wanted to
analyze this first.

I will bootstrap and regtest the patch on x86_64 and I guess it should go in if
it fixes the ia64 bootstrap issue (which i can't test apparently).