This is interesting and pretty awesome. I can imagine having a lot of fun making a toy compiler that emits LLVM IR directly.

Practically speaking, is there a good reason to do this in non-toy software rather than the more-or-less standard practice of using C as an IR? That gives you, in addition to compilation via LLVM, the ability (in principle) to use gcc, icc, et al. (In practice, in many cases people end up using compiler-specific extensions in the generated C, which does impose limitations on portability to other compilers, but that's certainly avoidable.)

I suppose if you are dynamically generating things that need to be compiled as quickly as possible, eliminating an unnecessary compilation step could be a significant advantage. It doesn't seem like such a constraint would apply in most cases, though.