For those who are curious why CPS is such a great compiler representation for implementing optimizations, it's because the passes require less analysis complexity. You only have to handle the "call" portion and don't have to worry separately about "return" points (though you've really just converted returns into a call to a new function!).

It sounds trivial, but in practice it dramatically reduces the size of not only optimization passes such as inlining, but it makes analyses that drive optimizations simpler as well (e.g., control-flow analysis). For an example from my experience (Manticore - http://manticore.cs.uchicago.edu/), my pretty coarse CFA analysis on our direct-style IR is about 750 lines of SML code whereas an optimized CFA with three different variants and extra support for environment analysis is only 700 lines of SML.