Honza, You mentioned recently that you are making changes in IPA-CP wrt clones. Simon recently found a test case that I'm hoping will be handled by your changes. Essentially, the problem happens when we clone a node for IPACP whose caller is itself cloned.
In this case (attached), we have a function Baz() making two calls Bar(0) and Bar(1). Also, there are two cgraph nodes for Baz (Baz/1 and Baz/5) and 4 cgraph nodes for Bar (Bar/0, Bar/4, Bar/6 and Bar/7). All these clones were created by the inliner (which happens to run before IPA-CP when using -fwhopr). When IPA-CP decides to clone Bar/0, it creates the ipcp clone T.0 and replaces the edge Baz/1->Bar/0 with Baz/1->T.0. At the same time, it modifies the function body for Baz() to call T.0() instead of Bar(0). So, now we have updated the cgraph node Baz/1 correctly, but it is impossible to update the cgraph node Baz/5 since Baz/5 never had a reference to Bar/0. This causes a cgraph verification failure when we analyze the gimple body of Baz and find a call to T.0 but no such edge in Baz/5. The easy way out of this is to refuse to clone Bar/0. Since Baz is itself cloned, we would need to make a duplicate of Baz's body to handle this case. You can reproduce this problem by compiling with -fwhopr in the branch or by simply switching the order of pass_ipa_inline and pass_ipa_cp in mainline (though that will probably give you a different IPACP failure as we have some changes for it in the branch). Should I file a bug report or are you handling this situation already? Thanks. Diego.
