> I expect that relying on LLVM going forward will be the popular approach for Haskell and other languages

My vague understanding that might be outdated by a few years is that LLVM isn't a very nice target for garbage collected languages. It does have support for GC hooks, stack maps etc., but somehow people are (were?) struggling to make it all work efficiently in the real world.

There must be other factors as well, as a Haskell LLVM backend has existed for years. Here's a list of problems with LLVM for this use case pointed out two years ago: https://ghc.haskell.org/trac/ghc/wiki/ImprovedLLVMBackend