"No compiler is good enough to do what the good libraries do."

That's nonsense. And if you take a look, say, at NumPy, which uses heavily tuned BLAS/LAPACK code, and Numexpr, which uses the "compile the snippet and run it" approach, you'll see why. Libraries won't save your performance if you can't optimize across procedure and module boundaries.

Which is why C and C++ under most state of the art compilers support link time optimization.

That won't help you much, if you, say, need the fusion of matrix operations I mentioned.

Stream fusion is an optimization framework that is (to my knowledge) defined in the library, but common to most sequence types in Haskell.