Top positive review
4 of 4 people found this helpful
on October 13, 2003
Advances in compiler design do not get much press these days. The reasons for this are unclear, but no doubt the perception that compilers need no further improvement has something to do with this. This book, written by one of the leading experts on compilers, certainly dispels this belief. Once readers get used to the idiosyncratic ICAN (Informal Compiler Algorithm Notation) invented by the author and used throughout the book, they get a comprehensive overview of compilers, especially that of optimization. Compilers for the SPARC, PowerPC, DEC, and Pentium architectures are treated in the book. The predominant emphasis of the book is in optimization, and so a few more recent and important topics in compiler construction, such as partial evaluation, are not discussed. Readers are expected to have a prior background in elementary compiler theory. My primary interest in reading the book was to gain insight into the compilation issues that arise in symbolic programming languages such as LISP and Prolog.
A detailed review of this book cannot be done for lack of space, but some of the helpful aspects and interesting discussions in the book include: 1. The "wrap-up" section at the end of each chapter, giving a compact summary of what was done in the chapter. 2. Generating loads and stores: The author shows how to move values to and from registers using routines more sophisticated than simply loading values into registers before using them or storing values as soon as they have been computed. 3. The main issues in the use of registers, such as variable allocation, efficiency of procedural calls, and scoping. The author lists the different categories that will result in contention for registers, such as stack, frame, and global offset table pointers and dynamic and static links. 4. The local stack frame and its uses, such as holding indexed variables (arrays, etc.) and debugging. 5. The five different parameter-passing mechanisms: call by value, call by result, call by value-result, call by reference, and call by name. A thorough discussion is given of their properties and what languages make use of them. In particular, the author notes that in the languages C and C++, call by value is the only parameter-passing mechanism, but that the address of an object may be passed, thus emulating essentially call by reference. This can be a source of confusion to those who program in C and C++. The most exotic of these mechanisms is call by name, which is a form of "lazy evaluation" in functional programming languages. The author gives a code example of the call by name parameter passing in ALGOL 60. I don't know of any modern practical programming languages that make use of call by name. 6. Shared libraries and the role of semantic linking and position independent code. 7. The compilation issues that arise in symbolic languages, such as LISP and Prolog. These languages typically have run-time type checking and function polymorphism, which gives them their power and ease of use. The author discusses how to produce efficient code for these languages. Since heap storage is utilized heavily by these languages, the allocation and recovering of it is very important. "Generation scavenging" is mentioned as the most efficient method for doing garbage collection in these languages. This method has been advertised in the literature as one that minimizes the time needed for storage reclamation in comparison with other approaches. In addition, the use of "on-the-fly" recompilation for polymorphic-language implementations is discussed. 8. Dynamic programming and its role in automatic production of code generators, as contrasted with the "greedy approach". The author explains the need for "uniform register machines" in the dynamic programming algorithm. 9. Interval analysis and its use in the analysis of control flow. This technique has been used in the field called "abstract interpretation" in recent years, the aim of which is too automatically and intelligently test program code. 10. Dependencies between dynamically allocated objects, such as links between graph structures in LISP and Prolog. The author describes the Hummel-Hendren-Nicolau technique for doing this, which involves naming schemes for locations in heap memory, a collection of axioms for characterizing aliasing locations among locations, and lastly, and most interestingly, utilizes a theorem prover to establish the properties of the data structures. The author emphasizes though that this technique, and others developed for doing dependence analysis of dynamically allocated objects, are very computationally intensive. 11. Individual optimizations, which the author divides into four groups in order of importance. 12. Induction-variable optimizations and their role in loop optimizations. The author shows how to identify induction variables, and how to transform them using various techniques, going by the name strength reduction, induction-variable removal, and linear-function test replacement. 13. Procedure integration and its role in "inlining" procedures in languages such as C++. The author emphasizes the drawbacks in using inlining, such as its impact on cache misses. 14. The trade-off between object abstraction and optimization, which occurs in object-oriented languages such as C++. The author discusses in detail the role of interprodecural optimizations in dealing with abstraction in the object-oriented modular approach to programming, particularly the identification of "side effects" in making procedure calls. 15. Code optimization that takes advantage of the memory hierarchy, such as data and instruction caches, and how to improve register allocation for arrays. The author gives a detailed and highly interesting discussion of scalar replacement for array elements. 16. Future trends and research in compiler design. The author mentions a few which he believes will dominate in the upcoming decade, such as scalar-oriented and data-cache optimizations. Scalar compilation will be he most active research area in his opinion. At the present time, there has been discussion of "intelligent compilers" that will interact with the user to develop optimal code, or even produce correct programs. These compilers will understand the intentions of the program and warn the user if these are violated, as well as reduce the time and cost needed for testing programs.