C-- is used in GHC over LLVM because it (last I heard) supports tail call optimization better than LLVM does.

http://www.haskell.org/pipermail/glasgow-haskell-users/2006-...

http://www.haskell.org/pipermail/glasgow-haskell-users/2006-...

Those are rather old, though, and LLVM has changed (and improved) a lot since then. I'd also figure that LLVM supports JIT compilation much better than C--, which may not support it at all.