[Bug tree-optimization/57539] [4.9 Regression] ice in ipa_edge_duplication_hook

[jamborm at gcc dot gnu.org]

http://gcc.gnu.org/bugzilla/show_bug.cgi?id=57539

Martin Jambor <jamborm at gcc dot gnu.org> changed:

           What    |Removed                     |Added
----------------------------------------------------------------------------
                 CC|                            |hubicka at gcc dot gnu.org

--- Comment #2 from Martin Jambor <jamborm at gcc dot gnu.org> ---
The main reason why the code does not work is that when
ipa_edge_duplication_hook is invoked, dst->caller does not yet have
its global.inlined_to set (it does not have any callers either) and so
we cannot use it to figure out which copy of rdesc in the list is the
one describing the new tree of inline clones.  This means that this
whole copying/re-mapping should be moved out of the hook into a
specialized function called from clone_inlined_nodes.

Additionally, and this has confused me for hours, we also do not even
create the rdesc we want to map a different edge to in this particular
testcase.  That is because an originally indirect edge is involved and
the call to ipa_free_edge_args_substructures in
ipa_propagate_indirect_call_infos.  That deallocated the vector of
jump functions, so the loop over it in the hook did not do anything.
That call needs to be removed.