We went through this at my former employer (not MS) just before i left. In fact the politics surrounding it drove me to seek relative sanity of doing a startup. When I started working on it, there really were two camps: clang emitting house IR instead of LLVM IR, or translating LLVM IR either early (right out of clang) or late (as a target after all optimization). I was working on the prototypes for the latter pair of options before another constraint was added that invalidated all of those paths.

The problem was that clang is quite intertwined with LLVM IR. ABI is spread all over the place, both FE and BE. It was proving hard to get something the house optimizers/code generators would be happy with. So a more radical path was chosen — wiring clang's parser + ASTs to another C++ front-end's AST and then using the old IR generator to get all of the idiosyncrasies right for the target. It was wacky, but seemed to be half-way working when I left, such that major open source code was compiling.

The whole idea of taking just parts of clang/llvm reeks of a half-measure anyway. It's probably better to either keep the existing compiler stack, or cut it right out and put the secret sauce into LLVM as add-in modules.