I also implemented what I refer to as typed shapes, which is a mechanism to version code based on object shapes and property types that I feel meshes quite well with BBV. Generating type-specialized function entry points (passing types to callees) Type-specialized call continuations (passing return types back to callers) Shapes-changed "dirty" flag propagation to call continuations (shape mutation/aliasing info) Threading the global object shape through function calls The results so far are impressive enough that I believe BBV may actually be eliminating more type tests than what is achievable with traditional type analyses.