> Hi, > > until recently we did not have thunks with incoming edges in the call > graph. However, speculative devirtualization can introduce them which
We did, just not very often. Normal devirtualization code can produce them, too. > may lead to creation of new SCCs in the call graph. What is more, > ipa_reduced_postorder called with reduce parameter set to true may > decide to represent the whole SCC with the thunk which will in turn > currently cause propagate_constants_topo from IPA-CP to ignore the > whole SCC and make ipcp_verify_propagated_values fail later. This > happens for example when compiling Mozilla Firefox (I did it with LTO > but it failed even when creating a fat unit object file). I believe > Honza has this check disabled for this reason. Yes, I have this check disabled and I am waiting for this patch to enable ipa-devirt at LTO. Thanks for working on this! > > The patch below fixes propagate_constants_topo by making it check each > node in the SCC rather than just the leading one. I have also changed > it to use ipa_get_nodes_in_cycle rather than loop through dfs_infos > since that seems to be the documented way of iterating over SCCs. > (And I should probably move edge_within_scc to ipa-utils too.) Interesting, I did not know about that - all the code seems to be doing the former. I never liked the representation of dfs_infos, so it is good we have abstraction. > > 2013-09-09 Martin Jambor <mjam...@suse.cz> > > * ipa-cp.c (propagate_constants_topo): Do not ignore SCC > represented by a thunk. Patch is OK. Thanks! Honza
