1) LLVM IR is strongly typed, which makes it very challenging to implement a dynamically-typed language on top of it. AFAIK, there aren't many (any?) successful

1) LLVM IR is strongly typed, which makes it very challenging to implement a dynamically-typed language on top of it. AFAIK, there aren't many (any?) successful such implementations of dynamically-typed languages. I haven't used other backend libraries, but I do think that looser typing at the IR level would make it a lot harder to catch implementation bugs; at least with LLVM, you are likely to hit an assertion if you break typing rules in the IR. 2) The type-checker is fairly ad-hoc, and I admittedly didn't look much into the literature here. The general idea is that since variable/function/class field types are explicitly typed, every expression can be checked for violating it's expected type. Implementation-wise, this is facilitated using the visitor pattern [1]. For certain expressions, some extra flexibility is allowed - for example, assigning an object of a child class to a variable of the parent class is not a type error. [1] https://github.com/neeilan/neeilang/blob/master/src/type-checker.cc https://github.com/neeilan/neeilang/blob/master/src/type-che...