MLton has been mostly unmaintained for 5+ years - GHC has a performance consortium improving it ( llvm, new code gen, new register allocator, parallel GC)... I'd expect it to be a tie now for sequential code.
I see a 2 year old llvm branch for MLton
http://mlton.org/cgi-bin/viewsvn.cgi/mlton/branches/llvm/#di...
but have no idea how functional (aargh! no not that functional) it is. There seems to be some activity on the trunk in the last 7~8 months, but possibly maintenance edits. Just putting it out there if anyone wants to poke. It would be sad for MLton to bit rot.
there's always this (#include standard disclaimer that some languages' submissions have been heavily optimized, others have not, and none of them may work like your apps; languages with modern type systems do prety well: ATS, scala, GHC, clean, ocaml, even F# under mono
http://shootout.alioth.debian.org/u32/which-programming-lang...
Haskell is usually about 1-10x slower than C for the same task, while MLton has actually outperformed C in some benchmarks.
Edit: My information is out of date -- dons is almost certainly right on this one. I'd go with his estimation.