You probably don't need LLVM to work with LLVM-IR. In fact, I imagine that there could be a market for a "low-level" IR with more compact implementation, sort of similar to Oberon "slim binaries" or something like that. The fact remains, though, that the "pragmatics" stuff around LLVM-IR (GC and other interfaces) seems still a little bit problematic.

But I'm a Luddite, though; I like self-hosting native compilers and I'm not terribly fond of the idea of having to pack a 20MB blob of C++ code with my code either. (There's a lot of apps that could profit from dynamic translation at run-time, but using LLVM for that feels unnecessarily heavyweight. I wish the algorithms and passes from LLVM were available in form of some high-level DSL that you'd be able to translate into whatever language you use and automatically adapt to whatever data structures you use in your code. It doesn't sound exactly impossible.)