The only type information I can think of is invariants from type definitions and their use (e.g. this ADT contains an Integer that is never less than 5). This information can be used to (say) constant fold branches. This information would have to be reinferred by the optimiser. However, I don't think this would do anything because AFAIK LLVM can already get infer this.

This could give a speed boost: LLVM has llvm.assume with which one can annotate IR with, using this would mean that the optimizer doesn't have to infer this (and therefore save time).