@kha The Lean rbmap is 3x slower than the C++ rb_map.
I tried two different implementations: rbmap2 (Appel's Coq
implementation), and rbmap3 (tries to simulate the C++ version).
I believe rbmap3 is broken since it is too slow.
I have also identified missing reset/reuse opportunities.
The actual implementation misses a case the simple code at reuse.txt gets :(
rbmap3 also exposes the "TODO(Leo): improve this" at library/compiler/reduce_arity.cpp
