The exact choice of memory_order_* is hard, and even for versions that
seem to work on x86_64 TSAN complains like crazy.
TODO: a version that, instead of blocking if another thread is updating,
also does the computation, with suitable protection to ensure that
pruning the tree cannot happen while anyone is computing on it.
TODO: try having AT in Lazy instead of Lazy_rep (i.e. not shared) to
avoid the indirection. We can possibly choose one or the other depending
on the type.
Marc Glisse
bf0a42d740