I remember reading an article somewhere on the history of the first FORTRAN compiler. The guys doing it wanted it to succeed, and they were fighting the mentality that high level languages could not possibly be as efficient as hand-coded assembly, so they put a lot of work into the optimization of the generated code. It worked so well that the results that came out of the compiler sometimes suprised the compiler writers! They then would have to dive into the compiler sources to figure out how it was done. I don't remember where I read this article. If the story rings a bell with anyone, let me know.

It was an exciting period; when later on we began to get fragments of compiled programs out of the system, we were often astonished at the surprising transformations in the indexing operations and in the arrangement of the computation which the compiler made, changes which made the object program efficient but which we would not have thought to make as programmers ourselves (even though, of course, Nelson or Ziller could figure out how the indexing worked, Sheridan could explain how an expresssion had been optimized beyond recognition, and Goldberg or Sayre could tell us how section 5 had generated additional indexing operations). Transfers of control appeared which corresponded to no source statement, expressions were radically rearranged, and the same DO statement might produce no instructions in the object program in one context, and in another it would produce many instructions in different places in the program.