[Bug tree-optimization/51737] [4.6 Regression] g++ crashes (internal compiler error: Segmentation fault) when compiling quickbook

[jamborm at gcc dot gnu.org]

http://gcc.gnu.org/bugzilla/show_bug.cgi?id=51737

--- Comment #11 from Martin Jambor <jamborm at gcc dot gnu.org> 2012-01-09 22:42:00 UTC ---
The problem is much more fundamental than just a clone removal while
also walking the clones of the same function although that is the
reason why we segfault.  But that would be almost easy to solve.

The bigger problem is indeed materialization of recursive loads
itself: We have a function that calls itself twice and some of these
calls and their calls are scheduled to be recursively inlined.  When
the decisions are to be carried out, we recognize we need to save a
copy of the body to copy from somewhere and so we call
save_inline_function_body.  That function elects a clone to "own" the
body copy and calls tree_function_versioning to copy it.  And
tree_function_versioning calls the problematic
delete_unreachable_blocks_update_callgraph which analyzes CFG and
finds some BBs unreachable - I'm not quite sure what made them so
suddenly unreachable, but it almost certainly has something to do with
EH BBs that somehow are stale at that point.  And so the function
begins to nuke respective call graph edges and also call graph nodes
of respective inlined callees and their inlined callees and... and
sure enough it decides to delete the node that
save_inline_function_body chose to hold the new body, in other words
it intends to delete id->dst_node of currently running
tree_function_versioning.

Thus even if we taught delete_unreachable_blocks_update_callgraph to
be more careful about deleted nodes, tree_function_versioning would
fail nevertheless.

I'm wondering whether we could just skip
delete_unreachable_blocks_update_callgraph when
tree_function_versioning is called from within
save_inline_function_body, though I can imagine the verifier could
complain about something.  I have to run now, will experiment more
tomorrow.