But, anyway, it turns out that my example above is easily transformed from a bad GHC performance story into a good one.  If you'll bear with me, I'll show how below.

   First, Manuel makes a good point about the LLVM backend.  My "6X" anecdote was from a while ago and I didn't use llvm [1].  I redid it just now with 7.4.1+LLVM, results below.  (The below table should read correctly in fixed width font, but you can also see the data in the spreadsheet here.)

                   Time (ms)   Compiled File size   Comple+Runtime (ms)

GHC 7.4.1 O0    2444        1241K

GHC 7.4.1 O2    925        1132K             1561

GHC 7.4.1 O2 llvm  931         1133K

GHC 7.0.4 O2 via-C 684         974K

So LLVM didn't help [1].  And in fact the deprecated via-C backend did the best!  

[1] P.P.S. Most concerning to me about Haskell/C++ comparisons are David Peixotto's findings that LLVM optimizations are not very effective on Haskell-generated LLVM compared with typical clang-generated LLVM.