LLVM IR is machine-independent, so the NIH accusation is not really fair.
MDIL is described to be the platform's native machine code + some extension tokens (to avoid direct encoding of pointers, for example).
A similar LLVM IR derivative would no longer be by a long shot something resembling LLVM IR, so the property of being "derivative" would not buy you anything.
LLVM IR is machine-independent? What? There's plenty that isn't machine-independent, from the obvious (e.g., x86_fp80), to the less obvious (to create IR in the first place one has to have knowledge of the target ABI), what integer types are supported (i64 isn't supported everywhere!), etc.
While true in a narrow sense, it does so by targeting the lowest common practical denominator—in practice, a chip that looks a lot like a 386 or 32-bit ARM with no SIMD pipeline. That puts it as little more than a better GNU Lightning.
MDIL is described to be the platform's native machine code + some extension tokens (to avoid direct encoding of pointers, for example).
A similar LLVM IR derivative would no longer be by a long shot something resembling LLVM IR, so the property of being "derivative" would not buy you anything.