> On Thu, May 09, 2013 at 03:52:20PM +0200, Martin Jambor wrote: > > On Tue, May 07, 2013 at 11:43:33PM +0200, Jan Hubicka wrote: > > > > Hi, > > > > > > > > the problem in PR 57084 is that late PRE devirtualization creates a > > > > direct call to a decl fro which we only have an inlined call graph > > > > mode in the given partition. I tried to find a most universal place > > > > where to fix it because this problem is not special to type-based > > > > devirtualization and in theory can be caused by any call to a decl > > > > that is grabbed from a constructor. I think the best place is the > > > > following one-liner, because all such decls should go through > > > > canonicalize_constructor_val. > > > > > > > > Bootstrapped and tested on x86_64-linux, fixes the testcase (at -m32) > > > > and I have happened to also LTO build Mozilla Firefox with it. OK for > > > > trunk? > > > > > > Is this change needed for 4.8, too? It would be OK there. > > > > Although the testcase shows the problem relies on code that is new in > > 4.9, I think that yes, even the current devirtualization by looking > > into constructors can hit the bug. So I am going to commit it to 4.8 > > after testing on that branch. > > Except that there is no such thing as > cgraph_get_create_real_symbol_node in 4.8. I suppose it's not worth > backporting it without a bug triggering there. However, this makes me
Yep, it was same situation with the PR fix I introduced cgraph_get_create_real_symbol_node for. In theory the bug exists for 4.8, too, but I do not have testcase. Lets go with bugfixes first and then I can do the renaming. Honza
