If it's just an inefficiency of the compiler, you can dismiss it. But if it's an inefficiency in the resulting code, then perhaps the many-pass design isn't adequate for teaching optimization.
Author of thesis is one of the authors of Blitz++: http://www.oonumerics.org/blitz/
The thesis came from his frustration when seemingly innocent changes provoke significant degradation in performance.
The transformations need to be applied together, but they can be specified together. One of the many impressive things in that thesis is the framework which lets optimizations be independently specified, but applied together. Unfortunately, the code does not seem to be available.
For a more recent system that echoes that particular aspect and does come with code there is Hoopl http://www.cs.tufts.edu/~nr/pubs/dfopt-abstract.html
It's on hackage http://hackage.haskell.org/package/hoopl
Actually, this way you will get an inferior compiler (optimizer).
http://lambda-the-ultimate.org/node/2387
The thesis referenced above argues (and provides examples) that you cannot overcome important problems by separating transformations. You need to combine them.