Page:AITR-474.djvu/127

117

we present here the complete working source code for RABBIT, written in SCHEME. (The listing of the code was produced by the "@" listing generator, written by Richard M. Stallman, Guy L. Steele Jr., and other contributors.)

The code is presented on successive odd-numbered pages. Commentary on the code is on the facing even-numbered page. An index appears at the end of the listing, indicating where each function is defined.

It should be emphasized that RABBIT was not written with efficiency as a particular goal. Rather, the uppermost goals were clarity, ease of debugging, and adaptability to changing algorithms during the development process Much information is generated, never used by the compilation process, and then thrown away, simply so that if some malfunction should occur it would be easier to conduct a post-mortem analysis. Information that is used for compilation is often retained longer than necessary. The overall approach is to create a big data structure and then, step by step, fill in slots, never throwing anything away, even though it may no longer be needed.

The algorithms could be increased in speed, particularly the optimizer, which often recomputes information needlessly. Determining whether or not the recomputation was necessary would have cluttered up the algorithms, however, making them harder to read and to modify, and so this was omitted. Similarly, certain improvements could dramatically' decrease the space used. The larger functions in RABBIT can just barely be compiled with a memory size of 256K words on ea PDP-10. However, it was deemed worthwhile to keep the extra information available for as long a time as possible.

The implementation of RABBIT has taken perhaps three man-months. This includes throwing away the original optimizer and rewriting it completely, and accomodating certain changes to the SCHEME language as they occurred. RABBIT was operational, without the optimizer, after about one man-month's work. The dissertation was written after the first version of the optimizer was demonstrated to work. The remaining time was spent analyzing the faults of the first optimizer, writing the second version, accomodating language changes, making performance measurements, and testing RABBIT on programs other than RABBIT itself.