[1-3] How can I improve my Lisp programming style and coding efficiency?

There are several books about Lisp programming style, including:
   
   1. Molly M. Miller and Eric Benson
      "Lisp Style and Design"
      Digital Press, 1990. 214 pages, ISBN 1-55558-044-0, $26.95.
           How to write large Lisp programs and improve Lisp programming 
           style. Uses the development of Lucid CL as an example. 

   2. Robin Jones, Clive Maynard, and Ian Stewart.
      "The Art of Lisp Programming"
      Springer-Verlag, 1989. 169 pages, ISBN 0-387-19568-8 ($33).

   3. W. Richard Stark.
      "LISP, Lore, and Logic: An Algebraic View of LISP
       Programming, Foundations, and Applications"
      Springer-Verlag, 1990. 278 pages. ISBN 0-387-97072-X paper ($42).
           Self-modifying code, self-reproducing programs, etc.

   4. CMU CL User's Manual, Chapter 7, (talks about writing
      efficient code). It is available by anonymous ftp from any CMU CS 
      machine (e.g., ftp.cs.cmu.edu [128.2.206.173]) as the file
        /afs/cs.cmu.edu/project/clisp/docs/cmu-user/cmu-user.ps 
      [when getting this file by anonymous ftp, one must cd to 
      the directory in one atomic operation, as some of the superior
      directories on the path are protected from access by anonymous ftp.]

   5. See also Norvig's book, SICP (Abelson & Sussman), SAP
      (Springer and Friedman).

   6. Hallvard Tretteberg's Lisp Style Guide is available by anonymous
      ftp in ftp.think.com:/public/think/lisp/style-guide.text. There is
      a fair bit of overlap between Hallvard's style guide and the notes
      below and in part 3 of this FAQ.

   7. Rajeev Sangal
      "Programming Paradigms in Lisp"
      McGraw-Hill, 1991. ISBN 0-07-054666-5.

   8. Rodney A. Brooks.
      "Programming in Common Lisp"
      John Wiley & Sons, New York, 1985. 303 pages. ISBN 0-471-81888-7.
      Chapter 5 discusses Lisp programming style.

Here are some general suggestions/notes about improving Lisp
programming style, readability, correctness and efficiency:

   General Programming Style Rules:

      - Write short functions, where each function provides a single,
        well-defined operation. Small functions are easier to
        read, write, test, debug, and understand.

      - Use descriptive variable and function names. If it isn't clear
        from the name of a function or variable what its purpose is,
        document it with a documentation string and a comment. In fact,
        even if the purpose is evident from the name, it is still worth
        documenting your code.

      - Don't write Pascal (or C) code in Lisp. Use the appropriate
        predefined functions -- look in the index to CLtL2, or use the
        APROPOS and DESCRIBE functions. Don't put a close parenthesis
        on a line by itself -- this can really irritate programmers
        who grew up on Lisp. Lisp-oriented text editors include tools
        for ensuring balanced parentheses and for moving across 
        pairs of balanced parentheses. You don't need to stick
        comments on close parentheses to mark which expression they close.

      - Use proper indentation -- you should be able to understand
        the structure of your definitions without noticing the parentheses. 
        In general, the way one indents a form is controlled by the
        first symbol of the form. In DEFUNs, for example, one puts the
        symbol DEFUN, the function name, and the argument list all on
        the same line. If the argument list is too long, one can break
        it at one of the lambda keywords. Following the argument list,
        one inserts a carriage return and lists the expressions in the
        body of the definition, with each form starting on its own
        line indented three spaces relative to the open parenthesis of
        the parent (in this case the DEFUN). This general style -- of
        putting all the significant elements of a form on a single
        line, followed by a carriage return and the indented body --
        holds for many Lisp constructs. There are, of course, variations,
        such as keeping the first clause on the same line as the COND
        or CASE symbol, and the rules are relaxed in different ways to
        keep line lengths to a manageable size. If you find yourself having
        trouble fitting everything in even with line breaking and
        relaxing the rules, either your function names are too long or your
        code isn't very modular. You should perceive this as a signal that
        you need to break up your big definitions into smaller chunks, each
        with a clearly defined purpose, and possibly replace long function
        names with concise but apt shorter ones.

      - Use whitespace appropriately. Use whitespace to separate
        semantically distinct code segments, but don't use too much
        whitespace. For example,
           GOOD: 
              (defun foo (x y)
                (let ((z (+ x y 10)))
                  (* z z)))

           BAD: 
              (defun foo(x y)(let((z(+ x y 10)))(* z z)))

              (defun foo ( x  y )
                (let ( ( z (+ x y 10) ) )
                  ( * z z )
                  )
                )    
         Although the Lisp reader and compiler don't care which you
         use, most experienced Lisp programmers find the first example 
         much easier to read than the last two.

       - Don't use line lengths greater than 80 characters. People who
         write code using Zmacs on Symbolics Lisp Machines are notoriously
         guilty of violating this rule, because the CPT6 font allows
         one to squeeze a tremendous amount of code on the display,
         especially if one spreads the code out horizontally. This
         makes it more difficult to read when printed out or read on
         an 80x24 xterm window. In fact, use a line length of 72 characters
         because it leaves a strip of white space at the edge of the window.

   The following functions often abused or misunderstood by novices. 
   Think twice before using any of these functions.

      - EVAL. Novices almost always misuse EVAL. When experts use
        EVAL, they often would be better off using APPLY, FUNCALL, or
        SYMBOL-VALUE. Use of EVAL when defining a macro should set off
        a warning bell -- macro definitions are already evaluated
        during expansion. See also the answer to question 3-12.
        The general rule of thumb about EVAL is: if you think you need
        to use EVAL, you're probably wrong.

      - PROGV. PROGV binds dynamic variables and is often misused in
        conjunction with EVAL, which uses the dynamic environment. 
        In general, avoid unnecessary use of special variables.
        PROGV is mainly for writing interpreters for languages embedded
        in Lisp. If you want to bind a list of values to a list of
        lexical variables, use
            (MULTIPLE-VALUE-BIND (..) (VALUES-LIST ..) ..)
        or
            (MULTIPLE-VALUE-SETQ (..) (VALUES-LIST ..))
        instead. Most decent compilers can optimize this expression. 
        However, use of this idiom is not to be encouraged unless absolutely
        necessary.

      - CATCH and THROW. Often a named BLOCK and RETURN-FROM are
        more appropriate. Use UNWIND-PROTECT when necessary.

      - Destructive operations, such as NCONC, SORT, DELETE,
        RPLACA, and RPLACD, should be used carefully and sparingly.
        In general, trust the garbage collector: allocate new
        data structures when you need them.

   To improve the readability of your code,

      - Don't use any C{A,D}R functions with more than two
        letters between the C and the R. When nested, they become
        hard to read. If you have complex data structures, you
        are often better off describing them with a DEFSTRUCT,
        even if the type is LIST. The data abstraction afforded by
        DEFSTRUCT makes the code much more readable and its purpose
        clearer. If you must use C{A,D}R, try to use
        DESTRUCTURING-BIND instead, or at least SECOND, THIRD, 
        NTH, NTHCDR, etc.

      - Use COND instead of IF and PROGN. In general, don't use PROGN if
        there is a way to write the code within an implicit
        PROGN. For example, 
           (IF (FOO X)
               (PROGN (PRINT "hi there") 23)
               34)
        should be written using COND instead.

      - Never use a 2-argument IF or a 3-argument IF with a second
        argument of NIL unless you want to emphasize the return value;
        use WHEN and UNLESS instead. You will want to emphasize the
        return value when the IF clause is embedded within a SETQ,
        such as (SETQ X (IF (EQ Y Z) 2 NIL)). If the second argument 
        to IF is the same as the first, use OR instead: (OR P Q) rather
        than (IF P P Q). Use UNLESS instead of (WHEN (NOT ..) ..)
        but not instead of (WHEN (NULL ..) ..).

      - Use COND instead of nested IF statements. Be sure to check for
        unreachable cases, and eliminate those cond-clauses.

      - Use backquote, rather than explicit calls to LIST, CONS, and
        APPEND, whenever writing a form which produces a Lisp form, but
        not as a general substitute for LIST, CONS and APPEND. LIST, 
        CONS and APPEND usually allocate new storage, but lists produced
        by backquote may involve destructive modification (e.g., ,.).

      - Make the names of special (global) variables begin and end
        with an asterisk (*): (DEFVAR *GLOBAL-VARIABLE*)   
        Some programmers will mark the beginning and end of an internal
        global variable with a percent (%) or a period (.).
        Make the names of constants begin and end with a plus (+):
        (DEFCONSTANT +E+ 2.7182818)
        This helps distinguish them from lexical variables. Some people
        prefer to use macros to define constants, since this avoids
        the problem of accidentally trying to bind a symbol declared
        with defconstant.

      - If your program is built upon an underlying substrate which is
        implementation-dependent, consider naming those functions and
        macros in a way that visually identifies them, either by placing
        them in their own package, or prepending a character like a %, ., 
        or ! to the function name. Note that many programmers use the
        $ as a macro character for slot access, so it should be avoided
        unless you're using it for that purpose.

      - Don't use property lists. Instead, use an explicit hash table.
        This helps avoid problems caused by the symbol being in the wrong
        package, accidental reuse of property keys from other
        programs, and allows you to customize the structure of the table. 

      - Use the most specific construct that does the job. This lets
        readers of the code see what you intended when writing the code.
        For example, don't use SETF if SETQ will do (e.g., for lexical
        variables). Using SETQ will tell readers of your code that you
        aren't doing anything fancy. Likewise, don't use EQUAL where EQ
        will do. Use the most specific predicate to test your conditions. 
        
      - If you intend for a function to be a predicate, have it return T
        for true, not just non-NIL. If there is nothing worth returning
        from a function, returning T is conventional. But if a function
        is intended to be more than just a predicate, it is better to 
        return a useful value. (For example, this is one of the differences
        between MEMBER and FIND.)

      - When NIL is used as an empty list, use () in your code. When NIL
        is used as a boolean, use NIL. Similarly, use NULL to test for an
        empty list, NOT to test a logical value. Use ENDP to test for the
        end of a list, not NULL.

      - Don't use the &AUX lambda-list keyword. It is always clearer to
        define local variables using LET or LET*.

      - When using RETURN and RETURN-FROM to exit from a block, don't
        use (VALUES ..) when returning only one value, except if you
        are using it to suppress extra multiple values from the first
        argument. 

      - If you want a function to return no values (i.e., equivalent to
        VOID in C), use (VALUES) to return zero values. This signals
        to the reader that the function is used mainly for side-effects.

      - (VALUES (VALUES 1 2 3)) returns only the first value, 1.
        You can use (VALUES (some-multiple-value-function ..)) to suppress
        the extra multiple values from the function. Use MULTIPLE-VALUE-PROG1
        instead of PROG1 when the multiple values are significant.

      - When using MULTIPLE-VALUE-BIND and DESTRUCTURING-BIND, don't rely
        on the fact that NIL is used when values are missing. This is
        an error in some implementations of DESTRUCTURING-BIND. Instead,
        make sure that your function always returns the proper number of
        values.

      - Type the name of external symbols, functions, and variables
        from the COMMON-LISP package in uppercase. This will allow your
        code to work properly in a case-sensitive version of Common Lisp,
        since the print-names of symbols in the COMMON-LISP package
        are uppercase internally. (However, not everybody feels that
        being nice to case-sensitive Lisps is a requirement, so this
        isn't an absolute style rule, just a suggestion.)

    Lisp Idioms:

      - MAPCAN is used with a function to return a variable number of
        items to be included in an output list. When the function returns zero
        or one items, the function serves as a filter. For example,
           (mapcan #'(lambda (x) (when (and (numberp x) (evenp x)) (list x)))
                   '(1 2 3 4 x 5 y 6 z 7))

    Documentation:

      - Comment your code. Use three semicolons in the left margin before
        the definition for major explanations. Use two semicolons that
        float with the code to explain the routine that follows. Two
        semicolons may also be used to explain the following line when the
        comment is too long for the single semicolon treatment. Use
        a single semicolon to the right of the code to explain a particular
        line with a short comment. The number of semicolons used roughly
        corresponds with the length of the comment. Put at least one blank
        line before and after top-level expressions.

      - Include documentation strings in your code. This lets users
        get help while running your program without having to resort to
        the source code or printed documentation. 

   Issues related to macros:

      - Never use a macro instead of a function for efficiency reasons.
        Declaim the function as inline -- for example, 
          (DECLAIM (INLINE ..))
        This is *not* a magic bullet -- be forewarned that inline
        expansions can often increase the code size dramatically. INLINE
        should be used only for short functions where the tradeoff is
        likely to be worthwhile: inner loops, types that the compiler
        might do something smart with, and so on.

      - When defining a macro that provides an implicit PROGN, use the
        &BODY lambda-list keyword instead of &REST.

      - Use gensyms for bindings within a macro, unless the macro lets
        the user explicitly specify the variable. For example:
            (defmacro foo ((iter-var list) body-form &body body)
              (let ((result (gensym "RESULT")))
                `(let ((,result nil))
                   (dolist (,iter-var ,list ,result)
                     (setq ,result ,body-form)
                     (when ,result
                        ,@body)))))        
        This avoids errors caused by collisions during macro expansion
        between variable names used in the macro definition and in the
        supplied body.

      - Use a DO- prefix in the name of a macro that does some kind of
        iteration, WITH- when the macro establishes bindings, and
        DEFINE- or DEF- when the macro creates some definitions. Don't
        use the prefix MAP- in macro names, only in function names.

      - Don't create a new iteration macro when an existing function
        or macro will do.

      - Don't define a macro where a function definition will work just
        as well -- remember, you can FUNCALL or MAPCAR a function but 
        not a macro.

      - The LOOP and SERIES macros generate efficient code. If you're
        writing a new iteration macro, consider learning to use one
        of them instead.
  
   File Modularization:

      - If your program involves macros that are used in more than one
        file, it is generally a good idea to put such macros in a separate
        file that gets loaded before the other files. The same things applies
        to primitive functions. If a macro is complicated, the code that
        defines the macro should be put into a file by itself. In general, if
        a set of definitions form a cohesive and "independent" whole, they
        should be put in a file by themselves, and maybe even in their own
        package. It isn't unusual for a large Lisp program to have files named
        "site-dependent-code", "primitives.lisp", and "macros.lisp". If a file
        contains primarily macros, put "-macros" in the name of the file.

   Stylistic preferences:

      - Use (SETF (CAR ..) ..) and (SETF (CDR ..) ..) in preference to
        RPLACA and RPLACD. Likewise (SETF (GET ..) ..) instead of PUT.

      - Use INCF, DECF, PUSH and POP instead instead of the corresponding
        SETF forms.

      - Many programmers religiously avoid using CATCH, THROW, BLOCK,
        PROG, GO and TAGBODY.  Tags and go-forms should only be necessary
        to create extremely unusual and complicated iteration constructs. In
        almost every circumstance, a ready-made iteration construct or
        recursive implementation is more appropriate.

      - Don't use LET* where LET will do. Don't use LABELS where FLET
        will do. Don't use DO* where DO will do.

      - Don't use DO where DOTIMES or DOLIST will do.

      - If you like using MAPCAR instead of DO/DOLIST, use MAPC when
        no result is needed -- it's more efficient, since it doesn't
        cons up a list. If a single cumulative value is required, use
        REDUCE. If you are seeking a particular element, use FIND,
        POSITION, or MEMBER.

      - If using REMOVE and DELETE to filter a sequence, don't use the
        :test-not keyword or the REMOVE-IF-NOT or DELETE-IF-NOT functions.
        Use COMPLEMENT to complement the predicate and the REMOVE-IF
        or DELETE-IF functions instead.

      - Use complex numbers to represent points in a plane.

      - Don't use lists where vectors are more appropriate. Accessing the
        nth element of a vector is faster than finding the nth element
        of a list, since the latter requires pointer chasing while the
        former requires simple addition. Vectors also take up less space
        than lists. Use adjustable vectors with fill-pointers to
        implement a stack, instead of a list -- using a list continually
        conses and then throws away the conses.

      - When adding an entry to an association list, use ACONS, not
        two calls to CONS. This makes it clear that you're using an alist.

      - If your association list has more than about 10 entries in it,
        consider using a hash table. Hash tables are often more efficient.
        (See also [2-2].)

      - When you don't need the full power of CLOS, consider using
        structures instead. They are often faster, take up less space, and
        easier to use.

      - Use PRINT-UNREADABLE-OBJECT when writing a print-function.

      - Use WITH-OPEN-FILE instead of OPEN and CLOSE.

      - When a HANDLER-CASE clause is executed, the stack has already
        unwound, so dynamic bindings that existed when the error
        occured may no longer exist when the handler is run. Use
        HANDLER-BIND if you need this. 

      - When using CASE and TYPECASE forms, if you intend for the form
        to return NIL when all cases fail, include an explicit OTHERWISE
        clause. If it would be an error to return NIL when all cases
        fail, use ECASE, CCASE, ETYPECASE or CTYPECASE instead.

      - Use local variables in preference to global variables whenever
        possible. Do not use global variables in lieu of parameter passing.
        Global variables can be used in the following circumstances:
          *  When one function needs to affect the operation of
             another, but the second function isn't called by the first.
             (For example, *load-pathname* and *break-on-warnings*.)
          *  When a called function needs to affect the current or future
             operation of the caller, but it doesn't make sense to accomplish
             this by returning multiple values.
          *  To provide hooks into the mechanisms of the program.
             (For example, *evalhook*, *, /, and +.)
          *  Parameters which, when their value is changed, represent a
             major change to the program.
             (For example, *print-level* and *print-readably*.)
          *  For state that persists between invocations of the program.
             Also, for state which is used by more than one major program.
             (For example, *package*, *readtable*, *gensym-counter*.)
          *  To provide convenient information to the user.
             (For example, *version* and *features*.)
          *  To provide customizable defaults. 
             (For example, *default-pathname-defaults*.)
          *  When a value affects major portions of a program, and passing
             this value around would be extremely awkward. (The example
             here is output and input streams for a program. Even when
             the program passes the stream around as an argument, if you
             want to redirect all output from the program to a different
             stream, it is much easier to just rebind the global variable.)

      - Beginning students, especially ones accustomed to programming
        in C, Pascal, or Fortran, tend to use global variables to hold or pass
        information in their programs. This style is considered ugly by
        experienced Lisp programmers. Although assignment statements can't
        always be avoided in production code, good programmers take advantage
        of Lisp's functional programming style before resorting to SETF and
        SETQ. For example, they will nest function calls instead of using a
        temporary variable and use the stack to pass multiple values. When
        first learning to program in Lisp, try to avoid SETF/SETQ and their
        cousins as much as possible. And if a temporary variable is necessary,
        bind it to its first value in a LET statement, instead of letting it
        become a global variable by default. (If you see lots of compiler
        warnings about declaring variables to be special, you're probably
        making this mistake. If you intend a variable to be global, it should
        be defined with a DEFVAR or DEFPARAMETER statement, not left to the
        compiler to fix.)

   Correctness and efficiency issues:

      - In CLtL2, IN-PACKAGE does not evaluate its argument. Use defpackage
        to define a package and declare the external (exported)
        symbols from the package. 

      - The ARRAY-TOTAL-SIZE-LIMIT may be as small as 1024, and the
        CALL-ARGUMENTS-LIMIT may be as small as 50. 

      - Novices often mistakenly quote the conditions of a CASE form.
        For example, (case x ('a 3) ..) is incorrect. It would return
        3 if x were the symbol QUOTE. Use (case x (a 3) ..) instead.

      - Avoid using APPLY to flatten lists. Although 
           (apply #'append list-of-lists)
        may look like a call with only two arguments, it becomes a 
        function call to APPEND, with the LIST-OF-LISTS spread into actual
        arguments. As a result it will have as many arguments as there are
        elements in LIST-OF-LISTS, and hence may run into problems with the
        CALL-ARGUMENTS-LIMIT. Use REDUCE or MAPCAN instead:  
           (reduce #'append list-of-lists :from-end t)
           (mapcan #'copy-list list-of-lists)
        The second will often be more efficient (see note below about choosing
        the right algorithm). Beware of calls like (apply f (mapcar ..)).

      - NTH must cdr down the list to reach the elements you are
        interested in. If you don't need the structural flexibility of
        lists, try using vectors and the ELT function instead.

      - CASE statements can be vectorized if the keys are consecutive
        numbers. Such CASE statements can still have OTHERWISE clauses.
        To take advantage of this without losing readability, use #. with 
        symbolic constants:

            (eval-when (compile load eval)
               (defconstant RED 1)
               (defconstant GREEN 2)
               (defconstant BLUE 3))

            (case color
              (#.RED   ...)
              (#.GREEN ...)
              (#.BLUE  ...)
              ...)

      - Don't use quoted constants where you might later destructively
        modify them. For example, instead of writing '(c d) in
           (defun foo ()
             (let ((var '(c d)))
               ..))
        write (list 'c 'd) instead. Using a quote here can lead to
        unexpected results later. If you later destructively modify the 
        value of var, this is self-modifying code! Some Lisp compilers
        will complain about this, since they like to make constants
        read-only. Modifying constants has undefined results in ANSI CL.
        See also the answer to question [3-13].

        Similarly, beware of shared list structure arising from the use
        of backquote. Any sublist in a backquoted expression that doesn't
        contain any commas can share with the original source structure.

      - Don't proclaim unsafe optimizations, such as
           (proclaim '(optimize (safety 0) (speed 3) (space 1))) 
        since this yields a global effect. Instead, add the
        optimizations as local declarations to small pieces of
        well-tested, performance-critical code:
           (defun well-tested-function ()
              (declare (optimize (safety 0) (speed 3) (space 1)))
             ..)
        Such optimizations can remove run-time type-checking; type-checking
        is necessary unless you've very carefully checked your code
        and added all the appropriate type declarations.

      - Some programmers feel that you shouldn't add declarations to
        code until it is fully debugged, because incorrect
        declarations can be an annoying source of errors. They recommend
        using CHECK-TYPE liberally instead while you are developing the code.
        On the other hand, if you add declarations to tell the
        compiler what you think your code is doing, the compiler can
        then tell you when your assumptions are incorrect.
        Declarations also make it easier for another programmer to read
        your code. 

      - Declaring the type of variables to be FIXNUM does not
        necessarily mean that the results of arithmetic involving the 
        fixnums will be a fixnum; it could be a BIGNUM. For example,
           (declare (type fixnum x y))
           (setq z (+ (* x x) (* y y)))
        could result in z being a BIGNUM. If you know the limits of your
        numbers, use a declaration like
           (declare (type (integer 0 100) x y))
        instead, since most compilers can then do the appropriate type
        inference, leading to much faster code.

      - Don't change the compiler optimization with an OPTIMIZE
        proclamation or declaration until the code is fully debugged
        and profiled.  When first writing code you should say 
        (declare (optimize (safety 3))) regardless of the speed setting.

      - Depending on the optimization level of the compiler, type
        declarations are interpreted either as (1) a guarantee from
        you that the variable is always bound to values of that type,
        or (2) a desire that the compiler check that the variable is
        always bound to values of that type. Use CHECK-TYPE if (2) is
        your intention.

      - If you get warnings about unused variables, add IGNORE
        declarations if appropriate or fix the problem. Letting such
        warnings stand is a sloppy coding practice.

   To produce efficient code,

      - choose the right algorithm. For example, consider seven possible
        implementations of COPY-LIST:

           (defun copy-list (list)
             (let ((result nil))
               (dolist (item list result)
                 (setf result (append result (list item))))))

           (defun copy-list (list)
             (let ((result nil))
               (dolist (item list (nreverse result))
                 (push item result))))

           (defun copy-list (list)
             (mapcar #'identity list))

           (defun copy-list (list)
             (let ((result (make-list (length list))))
               (do ((original list (cdr original))
                    (new result (cdr new)))
                   ((null original) result)
                 (setf (car new) (car original)))))

           (defun copy-list (list)
             (when list
               (let* ((result (list (car list)))
                      (tail-ptr result))
                 (dolist (item (cdr list) result)
                   (setf (cdr tail-ptr) (list item))
                   (setf tail-ptr (cdr tail-ptr))))))
        
            (defun copy-list (list)
              (loop for item in list collect item))

            (defun copy-list (list)
              (if (consp list) 
                  (cons (car list)
                        (copy-list (cdr list)))
                  list))

        The first uses APPEND to tack the elements onto the end of the list.
        Since APPEND must traverse the entire partial list at each step, this
        yields a quadratic running time for the algorithm.  The second
        implementation improves on this by iterating down the list twice; once
        to build up the list in reverse order, and the second time to reverse
        it. The efficiency of the third depends on the Lisp implementation,
        but it is usually similar to the second, as is the fourth.  The fifth
        algorithm, however, iterates down the list only once. It avoids the
        extra work by keeping a pointer (reference) to the last cons of the 
        list and RPLACDing onto the end of that. Use of the fifth algorithm 
        may yield a speedup. Note that this contradicts the earlier dictum to
        avoid destructive functions. To make more efficient code one might
        selectively introduce destructive operations in critical sections of
        code. Nevertheless, the fifth implementation may be less efficient in
        Lisps with cdr-coding, since it is more expensive to RPLACD cdr-coded
        lists. Depending on the implementation of nreverse, however,
        the fifth and second implementations may be doing the same
        amount of work. The sixth example uses the Loop macro, which usually
        expands into code similar to the third. The seventh example copies
        dotted lists, and runs in linear time, but isn't tail-recursive. 

        There is a long-running discussion of whether pushing items
	onto a list and then applying NREVERSE to the result is faster or
	slower than the alternatives. According to Richard C. Waters (Lisp
	Pointers VI(4):27-34, October-December 1993), the NREVERSE strategy is
	slightly faster in most Lisp implementations. But the speed difference
	either way isn't much, so he argues that one should pursue the option
	that yields the clearest and simplest code, namely using NREVERSE.

        Here's code for a possible implementation of NREVERSE. As is
        evident, most of the alternatives to using NREVERSE involve
        essentially the same code, just reorganized. 

	   (defun nreverse (list)
	     ;; REVERSED is the partially reversed list, 
	     ;; CURRENT is the current cons cell, which will be reused, and
	     ;; REMAINING are the cons cells which have not yet been reversed.
	     (do* ((reversed nil)		
		   (current list remaining)
		   (remaining (cdr current) (cdr current)))
		 ((null current)
		  reversed)
	       ;; Reuse the cons cell at the head of the list:
	       ;;    reversed := ((car remaining) . reversed)
	       (setf (cdr current) reversed)
	       (setf reversed current)))

      - use type declarations liberally in time-critical code, but
        only if you are a seasoned Lisp programmer. Appropriate type
        declarations help the compiler generate more specific and
        optimized code. It also lets the reader know what assumptions
        were made. For example, if you only use fixnum arithmetic,
        adding declarations can lead to a significant speedup. If you
        are a novice Lisp programmer, you should use type declarations
        sparingly, as there may be no checking to see if the
        declarations are correct, and optimized code can be harder to
        debug. Wrong declarations can lead to errors in otherwise
        correct code, and can limit the reuse of code in other
        contexts. Depending on the Lisp compiler, it may also 
        be necessary to declare the type of results using THE, since
        some compilers don't deduce the result type from the inputs.

      - check the code produced by the compiler by using the
        disassemble function
Go Back Up

Go To Previous

Go To Next