Comment by Jeaye

> i mean it's as prone to error as any other thing that relies on string munging.

This is misleading. Having done a great deal of both (as jank also supports C++ codegen as an alternative to IR), if the input is a fully analyzed AST, generating IR is significantly more error prone than generating C++. Why? Well, C++ is statically typed and one can enable warnings and errors for all sorts of issues. LLVM IR has a verifier, but it doesn't check that much. Handling references, pointers, closures, ABI issues, and so many more things ends up being a huge effort for IR.

For example, want to access the `foo.bar` member of a struct? In IR, you'll need to access foo, which may require loading it if it's a reference. You'll need to calculate the offset to `bar`, using GEP. You'll need to then determine if you're returning a reference to `bar` or if a copy is happening. Referencing will require storing a pointer, whereas copying may involve a lot more code. If we're generating C++, though, we just take `foo` and add a `.bar`. The C++ compiler handles the rest and will tell us if we messed anything up.

If you're going to hand wave and say anything that's building strings is error prone and unsafe, regardless of how richly typed and thoroughly analyzed the input is, the stance feels much less genuine.