> How about attached (untested) patch implementing a conservative, but > runtime-efficient approach?
This doesn't work. My understanding is that you need to recompute far more than that, in particular the points-to information for _all_ the calls in the function. I don't know enough of the machinery to tell whether this can be done incrementally, but I wouldn't think so, so passing TODO_rebuild_alias might be the only solution (in a less brutal way of course). I'll post a testcase, but it depends on some other, not yet installed changes. -- Eric Botcazou