|Allegro CL version 8.2|
Moderately revised from 8.1. Minimal update since 8.2 release.
This document contains the following sections:1.0 Lisp images without a compiler
The Common Lisp standard requires that a Common Lisp system have a compiler and a loader. Since running interpreted code in any Common Lisp system is slow compared to running compiled code, most users will want to compile functions after trying them out in interpreted mode. A compiled function will run many times faster than its interpreted counterpart.
There are several implementation-dependent facets to a Common Lisp compiler. These include the naming of source and object files, how declarations are handled, and how floating-point and other optimizations are performed. These issues are discussed in this document.
Building an image with build-lisp-image (see
building-images.htm) with include-compiler
nil (or include-compiler true and
discard-compiler also true) results in a compilerless Lisp
Standard runtime images (see runtime.htm) may not
have the compiler included. In such images, it is an error to call
compile or compile-file directly or indirectly. The value
further below) variable is ignored in standard runtime images and thus
does not affect whether or not an error is signaled. The condition
The remainder of this section discusses images which are not standard runtime images.
It is possible to load the compiler into a compilerless image. You
do so by evaluating
(require :compiler). However, it may
be difficult or impossible to compile everything that would have been
compiled had the compiler been present earlier (difficult or
impossible because you do not know what the things are). Therefore,
parts of Lisp will run slower than necessary after the compiler is
loaded with no space saving (since you loaded the compiler). In short,
if you intend to use the compiler, start with an image that includes
But note that you can build an image with the compiler included
during the build and discarded after the build completes. This might
be very useful when creating an application for distribution when you
are not licensed to distribute an image with the compiler. If you
t in a call to build-lisp-image or generate-application (which accepts
build-lisp-image arguments), the compiler will be present while the
image is built and discarded at the end. Thus, things like applying
advice (see fwrappers-and-advice.htm) can be done
and compiled during the build. (Note that the new fwrapper facility,
described in the same document, works with objects that can be
If you use a compilerless Lisp, the system will (in the default) print a warning every time you or the system tries to call compile. For example, consider the following script from a compilerless Lisp image:
USER(1): (defun foo nil nil) FOO USER(2): (compile 'foo) Warning: Compiler not loaded -- function left uncompiled NIL NIL NIL USER(3):
The text of the warning changes according to what the system was
unable to compile. We emphasized that compile can be called by the system since some
operations, including calling advise under certain circumstances, calling
the (obsolete) function defforeign (but not calling its
replacement macro def-foreign-call) in a
fasl file, and doing non-standard method combinations in CLOS
or MOP usage, do cause the compiler to be called and the warning to be
printed. It is possible to suppress these warnings, either globally or
locally. The variable
whether the warning is printed or not.
We do not recommend setting
nil because interpreted code will run significantly
more slowly than compiled and users should know when the compiler was
called so they can expect degraded performance.
It is possible to suppress the warning locally. The warnings
actually signal the condition
compiler-not-available-warning. Thus, it is
possible to suppress these warnings by wrapping the following code
around sections likely to produce such unwanted warnings:
(handler-bind ((excl:compiler-not-available-warning #'(lambda (c) (declare (ignore c)) (muffle-warning)))) <code-which--may-directly-or-indirectly-calls-compile>)
Note again that it is possible to build an image with the compiler included during the build and discarded at the end. Things processed during the build will be processed with the compiler but the final image will not have it (and thus suitable for delivery when the license does not permit distribution of images containing the compiler, such as standard runtime images, see runtime.htm). See the description of the include-compiler and discard-compiler arguments to build-lisp-image in building-images.htm.
The Common Lisp function compile-file is used to compile a file containing Common Lisp source code. We describe that function here to point out some Allegro-CL-specific arguments. The link above is to the ANSI description.
Arguments: input-filename &key output-file xref verbose print fasl-circle load-after-compile
This function returns three values. The first value is the truename of
the output file, or
nil if the file could not
be created. The second value is
nil if no
compiler diagnostics were issued, and true otherwise. The third value
nil if no compiler diagnostics other than
style warnings were issued. If the third returned value is
true, there were serious compiler diagnostics
issued or the other conditions of type
error or type
warning were signalled during compilation.
This function will compile the Common Lisp forms in input-filename. input-filename must be a pathname, a string, or a stream (earlier versions of Allegro CL accepted a symbol as the filename argument but this is now explicitly forbidden by the ANSI standard).
input-filename is merged with
input-filename name has no type, then
compile-file uses the types in the list
name of the output file may be specified by the output-file
keyword argument. The default type of the output file is specified by
The output file will have a text header (readable, e.g., by the Unix command head) which provides information such as the time of creation and the user who created the file.
xref keyword argument is a boolean that controls whether
cross-referencing information is stored in the output file. The
default value of the xref keyword argument is the value of
The verbose keyword argument is a boolean that controls
printing of the filename that is currently being compiled. The default
value of the verbose keyword argument is the value of
The print keyword argument is a boolean that controls whether
compile-file prints the top-level form currently
being compiled. The default value of the print keyword
argument is the value of
The fasl-circle keyword argument is an Allegro CL extension
(i.e. may not be supported in versions of Common Lisp other than
Allegro CL). If the value of this argument is
nil, there are two effects:
On the other hand, compilation speed is typically faster when this
nil. This argument defaults to
the value of
The load-after-compile keyword argument is an Allegro CL
extension (i.e. may not be supported in versions of Common Lisp other
than Allegro CL). If the value of this argument is true (default
nil), the fasl (compiled Lisp) file will
be loaded when compilation completes. See also
the :cload top-level command,
which also compiles an loads files.
If a definition of an object (with a defining macro like defpackage, defmacro, defstruct, etc.) within a file being compiled with compile-file persists in Lisp after the compilation is complete, the definition is said to be persistent. The persistence behavior prior to release 7.0 was erratic, with some defining forms persisting and others not. The situation in 7.0 is fully described in Section 2.1 Persistence of defining forms encountered by compile-file. The short story is, absent explicit direction (such as with an eval-when), the only persistent defining form is defpackage.
The :cf top-level command invokes compile-file. :cf does not accept keyword arguments but does accept multiple input-filename arguments. :cf reads its arguments as strings, so you should not surround input-filenames with double quotes. Thus
is equivalent to
Suppose in a running Lisp, the symbols
foom do not name any objects, indeed do not even
exist in any package. Consider the following file named
;; file foo.cl begin (in-package :user) (defpackage :foo (:use :cl :excl)) (defmacro foom (x) x) (defun foo (y) (cons y y)) (defclass foo () ()) (defvar foo nil) ;; file foo.cl end
We now compile the file using compile-file:
After the file compilation completes, in the running Lisp, does
foo now name a package, a function, a class, and a
global variable? Does
foom name a macro?
The answer is, only the package is now defined in the Lisp:
cl-user(16): (describe 'foo) foo is a symbol. It is unbound. It is internal in the common-lisp-user package. cl-user(17): (find-package :foo) #<The foo package> cl-user(18): (describe 'foom) foom is a symbol. It is unbound. It is internal in the common-lisp-user package. cl-user(19):
The ANSI spec does not say whether defining forms persist after a compile-file (i.e. the definition should be available in the running Lisp when compilation is complete). The behavior of Allegro CL in releases prior to 7.0 was uneven, with some defining forms persisting and others not. Starting in release 7.0, the behavior is consistent: only defpackage forms persist.
We believe this new consistency will make it easier to know what the best compilation strategy is, particularly now that we support compiler enviroments (see environments.htm). Because it is easy to ensure definitions persist by using eval-when's, we believe that the extra programming burden is slight.
Here is a list of defining macros. Those marked with (*) persisted in earlier releases but no longer persist. More complicated situations are also noted:
If you wish a defining form to persist, wrap it in an appropriate eval-when form, such as:
(eval-when (compile load eval) (defmacro foo ...))
Definitions do not persist (absent eval-when's) from one file to another within a compilation unit:
;; file foo1.cl begin (in-package :user) (defvar *foo1* nil) ;; file foo1.cl end ;; file foo2.cl begin (in-package :user) (defvar *foo2* nil) (defun bar () *foo1*) ;; file foo2.cl end cl-user(1): :cf foo1 ;;; Compiling file foo1.cl ;;; Writing fasl file foo1.fasl ;;; Fasl write complete cl-user(2): :cf foo2 ;;; Compiling file foo2.cl ; While compiling bar: Warning: Free reference to undeclared variable *foo1* assumed special. ;;; Writing fasl file foo2.fasl ;;; Fasl write complete cl-user(3): (with-compilation-unit () (compile-file "foo1.cl") (compile-file "foo2.cl")) ;;; Compiling file foo1.cl ;;; Writing fasl file foo1.fasl ;;; Fasl write complete ;;; Compiling file foo2.cl ; While compiling bar: Warning: Free reference to undeclared variable *foo1* assumed special. ;;; Writing fasl file foo2.fasl ;;; Fasl write complete #p"foo2.fasl" t t cl-user(4):
Compiled Lisp files are called fasl files. ('Fasl' stands for fast loading.) By default, compile-file creates a file with type fasl containing the compiled Lisp code. See Section 7.0 File types for information on how to change the default type of a compiled file.
Fasl files are not human readable, but they do have headers which contain readable information, including:
There are 10 readable lines at the top of a fasl file. The UNIX head() command will print the readable head of a fasl file (assuming it prints the usual 10 lines by default).
If you try to load an inappropriate fasl file into a Lisp image, an
error is signaled. One such error is
file-incompatible-fasl-error, which is signaled
when you try to load a fasl file compiled in an earlier version (so
the R/S number and the fasl version do not match). Another error is
which is signaled when an attempt is made to load a file compiled in
case-insensitive-upper (ansi) into a case-sensitive-lower (modern)
The macro ensuring-compiled-body is
like progn in that it executes
its body forms sequentially, but if encountered in interpreted code, it
wraps the body in a
lambda and passes it to
for compilation, and then funcalls the result. Thus the code will be
executed compiled in all cases. No (accessible) function is
created. Surrounding code not wrapped in ensuring-compiled-body is not compiled.
The Allegro CL compiler signals warnings when it detects potential difficulties with the code being compiled. We describe some of these warnings here. These are not errors and the compilation will complete and compiled code will be produced. The warnings simply flag potential problems, perhaps uncovering a coding error which will cause a program problem even though the code compiles.
The following condition classes are among those that might be signaled:
compiler-inconsistent-name-usage-warning: signaled whenever variable or tag names are used in a manner inconsistent with their possibly intended usage, either unused (and for variables, not declared ignore or ignorable), or, again for variables, declared ignore but used.
compiler-unreachable-code-warning: signaled whenever a cond clause or a case clause cannot be reached, presumably because of the presence of a
otherwiseclause before the unreachable clause.
compiler-undefined-functions-called-warning: signaled when the compiler notices that there are calls to not-yet-defined functions in a function being compiled. This will not be signaled when the function is defined in the same file or is defined within the body of with-compilation-unit.
compiler-no-in-package-warning: this warning is no longer used. Allegro Presto required files to have an in-package form but Allegro Presto is no longer supported. See The Allegro Presto facility has been removed in loading.htm for further information.
If you try to compile a file that contains an incomplete form
(because of, e.g., a missing closing parenthesis),
compile-file signals an error with condition
end-of-file. Consider the following file
(defun foo nil nil) (defun bar (a b) (+ a b)
The closing parenthesis is missing from the definition of bar. When Lisp tries to compile this file, it signals an error:
USER(3): :cf missing-paren.cl ; --- Compiling file /net/rubix/usr/tech/dm/missing-paren.cl --- ; Compiling FOO Error: eof encountered on stream" #<EXCL::CHARACTER-INPUT-FILE-STREAM #p"/net/rubix/usr/tech/dm/missing-paren.cl" pos 45 @ #xa99c12> starting at position 20. [condition type: END-OF-FILE] Restart actions (select using :continue): 0: retry the compilation of /net/rubix/usr/tech/dm/missing-paren.cl 1: continue compiling /net/rubix/usr/tech/dm/missing-paren.cl but generate no output file  USER(4):
Note the line:
starting at position 20.
That indicates that the incomplete form starts at position 20 in the file. Opening the file in Emacs and entering the command C-u 20 C-f (assuming standard keybindings, that means move forward 20 characters) should bring you to the beginning of the incomplete form.
*compile-verbose* provide the defaults for the
print and verbose arguments to
compile-file. Those arguments control how much
information compile-file will print out. These
variables are a standard part of common lisp.
The default source-file type in Allegro CL is cl. The default compiled-file type in Allegro CL is fasl, which is a mnemonic for fast-loadable file. The default file types may be changed to suit your preferences.
*source-file-types* is a list of pathname
types of Common Lisp source files. The initial value of this variable
is the list
("cl" "lisp" "lsp" nil)
This means that if no file type is specified for the argument of compile-file (or the top-level command :cf), files with types cl, lisp, lsp, and no type will be looked for. For example
will cause the compiler to first look for the file foo.cl,
then for foo.lisp, the foo.lsp, and finally foo. Users
may want to change the value of
*source-file-types*. Some prefer not to
allow files with no type to be looked at, since this prevents Lisp
from trying to compile inappropriate files, or even a
directory. Evaluating the form
(setq sys:*source-file-types* '("cl" "lisp" "lsp"))
will cause the compiler to look for files with file type cl, lisp and lsp. Then
will look for foo.cl, foo.lisp, and foo.lsp but not foo. The first file found will be the one compiled.
If you change
*source-file-types*, you may also wish to
that the functions load and require will look for files with the
desired file types as well. See Search lists and its subsections in
loading.htm for a description of these variables.
When a file is compiled, a new file containing the compiled Lisp
code is created. This file will have the type specified by the
*fasl-default-type*. The initial value of this
variable is "fasl". You may change its value to any string
you wish. If you change the value of this variable, you should also
modify the load and require search lists so those functions will find
the correct files. Throughout the documentation, files containing
compiled Lisp code are called fasl files and are assumed to have a
file type of fasl. You should understand that fasl
really denotes the value of
Compiling Lisp code involves many compromises. Achieving very high speed generally involves sacrificing graceful recovery from errors and even the detection of errors. The latter may cause serious problems such as wrong answers or errors that result from the propagation of the original undiscovered error. On the other hand, interpreted code is very easy to debug but is too slow for practical applications.
Fortunately, most program needs can be satisfied on some middle ground. The software development cycle generally begins with interpreted code and ends with well-optimized code. Progressing through this cycle, higher speed is achieved without significant loss of confidence because of the increase in the correctness and robustness of the Lisp code as development proceeds. (It is important to note that optimizations do not affect the behavior of correct code with expected inputs. However, it is not always easy to prove that the inputs will always be what they are expected to be or that a complex program is indeed `correct.') This section provides enough information so that a programmer may make intelligent decisions about performance compromises. Section 9.0 Pointers for choosing speed and safety values further discusses the issue.
Among specific trade-offs in compiling code are the verification of the number and types of arguments passed to functions, providing adequate information for debugging, and including code to detect interrupts. The Allegro CL compiler has been designed to be highly parameterized. Optimizations performed by the compiler may be controlled to a significant degree by the user. The compiler of course supports the primitive Common Lisp safety, space, speed, debug, and compilation-speed optimization declarations. (Allegro CL accepts integer values between 0 and 3 inclusive, higher values representing greater importance to the corresponding aspect of code generation.) More significantly, Allegro CL provides a number of optimization switches that control specific aspects of code generation. These switches are in the form of variables that are bound to functions of the five optimization parameters safety, space, speed, debug, and compilation-speed.
The initial values of the optimization qualities are set when Allegro CL is built, controlled by the build-lisp-image arguments :opt-safety, :opt-space, :opt-speed, and :opt-debug (see Arguments to build-lisp-image 2: defaults not inherited from the running image in building-images.htm).
The default values for safety, space, and speed in an image created by build-lisp-image and in prebuilt images supplied in the Allegro CL distribution is 1. The default value for debug is 2. The default value for compilation-speed is 1.
You can set the values of the four qualities in various ways. One way is globally, with proclaim as follows:
(proclaim '(optimize (safety n1) (space n2) (speed n3) (debug n4) (compilation-speed n5)))
where n1, n2, n3, n4, and n5 are integers from 0 to 3. The values may also be set with the top-level command :optimize. This command also displays the current values of the qualities (there is no portable way to access those values in Common Lisp).
The function explain-compiler-settings prints the value
that will be returned by each compiler switch if called with specific
settings of the optimization qualities. Called with no arguments, it
describes the behavior with the current settings. It takes keyword
arguments named by the optimization qualities
compilation-speed). Values specified for those
arguments causes information on the values of compiler switches using
the specified values (and the current values, where unspecified) to be
The variables discussed in the remainder of this section specify
what the various settings of safety, space, speed, debug, and
compilation-speed do. Their values are functions that
the given settings of safety, space, speed, debug, and
compilation-speed. You may change
the definitions by binding (or setting) the variables to new
You can also set (or bind) a variable to
nil and this
will be interpreted as a function that always returns
(respectively) meaning the associated switch is always on (
t) or off (
Any new function used should accept five arguments. The system calls the function with the
(lexically) current values of safety, space, speed debug, and
Following the table describing the switches, we give examples of
erroneous code run interpreted (which produces the same error behavior
as running with the switches at their safest setting) and run compiled
with the switches at their least safe setting. What you will notice in
the latter case is that erroneous code either results in a less
informative error or perhaps in a wrong answer but no error. These
examples are not intended to deter you from using compiler
optimizations. Rather, we want to make you aware of the dangers of
less safe code and show what error messages to expect when the
switches are at their high speed settings. All switches are named by
symbols in the
Note on release 8.2: the compilation-speed quality, always part
of the CL specification, was not used by Allegro CL until release
8.2. compilation-speed is currently only used (in the default)
|Switch||Description||Initial value true when:|
||When true, a format string (the second required argument to
format) is converted to a tree structure at compile time. If
||True when speed is greater than space.|
||If true, compiler will generate code that assumes the sum and
difference of declared fixnums are fixnums.
Warning: if this switch returns true during compilation but the sum or difference of two declared fixnums is not a fixnum, the compiled code will silently produce erroneous results.
See Section 8.2 Declared fixnums example below.
|True only if speed is 3 and safety is 0.|
||If true, code is added at the
beginning of the code vector
for a compiled function and at the end of a loop in a
compiled function that checks for an interrupt
(Control-C on Unix, Break/Pause on Windows).
Note that an infinite loop that does not call functions will not be interruptable except by multiple Control-C's (Unix) and using the Franz icon on the system tray (Windows). See startup.htm for more information on interrupting when all else fails.
|True unless speed is 3 and safety is 0.|
||When true, the register-linking internal optimization is performed, resulting in faster but harder to debug code.||
True when compilation-speed is less than 2 and either speed is
greater than 2 or debug less than
||When true, calls to fslot-value-typed will be opencoded if possible. See ftype.htm||True when speed is greater than safety.|
||When true, the compiler should continue to provide high optimization even if the function has become very large (and thus compilation speed is reduced).||True when compilation-speed is less that 2.|
||When true, the compiler performs peephole optimization. Peephole optimizations include removing redundant instructions, such as a jump to the immediately following location.||True when speed is greater than 0.|
||If true, argument lists for compiled functions and macros will be stored (and available to tools like arglist).||True when debug is greater than 0.|
||If true, the names of local variables will be saved in compiled code. This makes debugging easier.||True when debug is greater than 1.|
||If true, information about when local variables are alive is saved in compiled code, making debugging more easy.||True only when debug is 3.|
||If true, compiler will perform self-tail-merging (for functions which call themselves). See Section 8.6 Tail merging discussion below for more information on tail-merging.||True if speed is greater than 0.|
||If true, compiler will perform non-self-tail-merging (for functions in the tail position different than the function being executed). See Section 8.6 Tail merging discussion below for more information on tail-merging.||True if speed is greater than 1 and debug less than 2. (This is more restrictive than tail-call-self-merge-switch described just above in this table because all references to the caller function are off the stack in this case but at least one call remains on the stack in the self-merge case.)|
If true, the compiler will trust declarations in code (perhaps other than
dynamic-extent declarations -- see next entry) and produce code (when
it can) that is optimized given the declarations. These declarations
typically specify the type of values of variables. If ||True if speed is greater than safety.|
If true and if
||True if speed is greater than or equal to safety.|
||If true, the compiler will compile suitable case statements in a table-driven fashion, which is faster but less safe. See the variable description (and its links) for details.||True if speed is 3 and safety is 0.|
||If true, the compiler will add code that checks that the correct number of arguments are passed to a function.||True if speed is less than 3 or safety is greater than 0.|
||If true, code calling car
will check that the argument is appropriate (i.e. a list).
The switch is only effective on
platforms which have
||True if speed is less than 3 or safety is greater than 1.|
||If false, the compiler will compile certain calls to funcall in such a way that the call immediately jumps to the funcall'ed function's start address. This speeds up funcall'ing at the cost of harder debugging (the stepper and the runtime analyzer call-counter will not see such calls).||True (i.e. slower funcall's) if speed is less than 3 or if safety is greater than 1 or if debug is greater than 0.|
||If true, code is generated that an object of undeclared type is of the correct type when it appears as the argument to specialized functions like svref and rplaca. Thus the argument to svref must be a simple vector and if this switch is true, code will check that (and give a meaningful error message). See Section 8.4 Argument type for a specialized function example for an example. Note that generic has nothing to do with (CLOS) generic functions. The name way predates CLOS.||True for speed less than 3 or safety greater than 1.|
If true, code will be added to ensure that a symbol is bound before
the value of that symbol is used. The result (as shown in
Section 8.5 Bound symbol example below) is that an error
with an appropriate message will be signaled in code compiled with
this switch true. In code compiled with this switch
||True if speed is less than 3 or safety is greater than 1.|
This switch only affects platforms with
||True if speed is less than 3 or safety is greater than 0.|
||When true, any lambda or let bindings of the declared variable and any setqs to that variable get a runtime type check to ensure that the value is of the correct type, otherwise a continuable type-error is generated.||True if safety is greater than 1.|
This switch does not affect code speed. Instead, it tells the compiler
to save information in fasl files useful for source-level debugging
(see The source
stepper in debugging.htm) and for
coverage analysis (see with-coverage). This saved information does
not affect the actual compiled code (which will be the same whether
this switch is true or
||True if debug is greater than 2.|
This switch affects 32-bit x86 processors only. When true, flags are
set that ensure floating-point operations are done in a consistent
mode, so the same calculations have the same result from run to
||False (faster, possibly slightly different results from run to run on 32-bit x86) only when speed is 3 and safety is 0.|
The following code illustrates the effect of this switch. We define
a function that simply adds its two arguments and contains
declarations that both the arguments are fixnums. When compiled with
this switch returning
nil, the function
correctly returns the bignum resulting from the arguments in the
example. When compiled with this switch returning
t, the function returns a wrong answer.
USER(28): (defun frf (n m) (declare (fixnum n) (fixnum m)) (+ n m)) FRF USER(29): (setq bn (- most-positive-fixnum 20)) 536870891 ;; 20 less than most-positive-fixnum USER(30): (frf bn 50000) 536920891 ;; This value is a bignum USER(31): (proclaim '(optimize (safety 0) (speed 3))) T USER(32): (compile 'frf) FRF NIL NIL USER(33): (frf bn 50000) -268385477 ;; wrong answer! USER(34):
Very serious errors can occur when this switch returns
nil and the wrong number of arguments are
passed. These errors are not necessarily repeatable. We give a simple
example where you get a less than useful error message, but you should
be aware that much more serious (possibly fatal) errors can result
from code where the number of arguments are not checked. About 3
instructions (the exact number is platform dependent and ranges from 1
to 4) are added to a function call when this switch returns
USER(1): (defun foo (a b c) (+ a c)) FOO USER(2): (foo 1 2) Error: FOO got 2 args, wanted at least 3. [condition type: PROGRAM-ERROR]  USER(3): :pop USER(4): (proclaim '(optimize (speed 3) (safety 0))) T USER(5): (compile 'foo) ; While compiling FOO: Warning: variable B is never used FOO T T USER(6): (foo 1) #<unknown object of type number 4 @ #x8666c> USER(7): ;; Note you might also get an error, although not one that ;; mentions the number of arguments.
Note that no error is signaled. Note further that it is possible for a fatal garbage collection error to result from passing the wrong number of arguments.
The following examples show what happens when the switch returns
t and when it returns
USER(39): (defun foo (vec n) (svref vec n)) FOO USER(40): (setq v (list 'a 'b 'c 'd 'e)) (A B C D E) USER(41): (foo v 3) Error: Illegal vector object passed to svref: (A B C D E)  USER(42): :pop USER(43): (proclaim '(optimize (speed 3) (safety 0))) T USER(44): (compile 'foo) FOO NIL NIL USER(45): (foo v 3) Error: Received signal number 10 (Bus error) [condition type: SIMPLE-ERROR] ;; Or it might seem to work but return a bogus value.  USER(46):
The following example shows what happens if this switch returns
nil. In that case, the
symbol-value location is simply read. If the symbol is in fact unbound, an apparently
valid but in fact bogus value may be obtained. In the example, that bogus value is passed
to the function + and an error is signaled because it is not a valid argument to +.
However, it may be that no error will be signaled and computation will continue, resulting
in later confusing or uninterpretable errors or in invalid results.
USER(60): (defun foo (n) (+ n bar)) FOO USER(61): (foo 1) Error: Attempt to take the value of the unbound variable `BAR'. [condition type: UNBOUND-VARIABLE]  USER(62): :pop USER(63): (proclaim '(optimize (speed 3) (safety 0))) T USER(64): (compile 'foo) ; While compiling FOO: Warning: Symbol BAR declared special FOO T T USER(65): (foo 1) Error: (NIL) is an illegal argument to + [condition type: TYPE-ERROR] ;; You may see a different error.  USER(66):
Consider the following function definition:
(defun foo (lis) (pprint lis) (list-length lis))
When you call foo with a list as an argument, the list is pretty printed and its length is returned. But note that by the time list-length is called, no more information about foo is needed but, in the interpreter at least, the call to foo remains on the stack. The compiler can tail merge foo in such a way that the call to list-length is changed to a jump. In that case, when list-length is reached the stack looks as if list-length was called directly and foo was never called at all. The side effects of calling foo, in this case, pretty printing the list passed as an argument, have all occurred by the time list-length is called.
Unwrapping the stack in this fashion is a benefit because it saves stack space and can (under the correct circumstances) avoid creating some stack frames all together. However, it can make debugging harder (because the stack backtrace printed by :zoom will not reflect the actual sequence of function calls, but see the discussion of ghost frames in debugging.htm) and it can skew profiling data (because, looking at our example, a sample taken after list-length is called will not charge time or space to foo because foo is off the stack).
The two switches
tail-call-self-merge-switch will have an
effect when its value is true at the beginning of the function being
self-called (and no other time).
tail-call-non-self-merge-switch has effect
only when true at the point of the call.
The user may change the code which determines how any of these variables behaves by redefining the function which is the value of the variable. That function must take as arguments safety, space, speed, and debug. It is important that the function be compiled since it may be called many times during a compilation and an interpreted function will slow down the compiler. Here is the general form which modifies a switch variable. var identifies the switch variable you wish to change.
(setq var (compile nil '(lambda (safety space speed debug) form-1 ... form-n)))
t as a function of safety,
space, speed, and debug.
Note that we wrapped the function definition with compile. Note too that such a form will not affect the compilation of the file in which it is itself compiled (since the variable will not be redefined in time to affect the compile-file).
For example, if the following code is evaluated at the top-level or in an initialization file, the compiler will not save local scopes if speed is greater than 1 or if debug is less than 2 or if space is greater than 1. The value of safety has no effect on the switch.
(setq compiler:save-local-scopes-switch (compile nil '(lambda (safety space speed debug) (declare (ignore debug)) (cond ((> speed 1) nil) ((< debug 2) nil) ((> space 1) nil) (t t)))))
The initial values of safety and speed are both set during image build (with build-lisp-image) using the opt-speed, opt-safety, opt-space, and opt-debug arguments. See building-images.htm.
If you declare a variable to be of a particular type, as in
(defun foo (x) (declare (optimize (speed 3)) (fixnum x)) (1+ x))
you are making a promise about your actions: you are promising not to
call foo with a non-fixnum argument. If you do call foo
with a non-fixnum argument (we assume the
trust-declarations-switch was true, its
default behavior for speed 3), then the result might be an
But there is typically no runtime check that values of the correct type are being passed to foo. If you call foo with a float or a bignum, the possibly incorrect result is produced silently.
You can have runtime checks added for type declarations and let and
lambda bindings and setq's by compiling while
verify-type-declarations-switch switch is
true. See the description for an example. Note that with the default
verify-type-declarations-switch and the
various relevant trust-* switches are both true only for speed
The value of a compiler switch can be
nil as well as a function.
t is interpreted as a function that always return
true and so causes the switch to always be on.
nil is interpreted as a function that always returns
false and so causes the switch to always be off. These settings are
particularly useful when binding the variables during a specific
This declaration has the same meaning and syntax as the Common Lisp
type declaration, but
it will be trusted even when the compiler
nil, which means most declarations are not
trusted. This allows you to specify certain declarations as always
trusted even if you are running at high safety.
This declaration should be used sparingly if at all. It is designed for cases where code will not compile unless the type is known (which occurs, for example, during bootstrapping but rarely in normal cases).
What values should you choose for speed and safety? It is tempting to set speed to 3 and safety to 0 so that compiled code will run as fast as possible, and devil take the hindmost. Our experience is that people who do this say that Allegro CL is fast but lacks robustness, while people who use the more conservative default settings of 1 and 1 feel that Allegro CL is very robust but occasionally seems a bit sluggish. Which you prefer is for you to decide. The following points should be considered.
verify-argument-count-switch); (2) interrupt checking is disabled (see
generate-interrupt-checks-switch); (3) sums and differences of fixnums are assumed to be fixnums (see
declared-fixnums-remain-fixnums-switch); (4) case is compiled in table-driven fashioon (see
trust-table-case-argument-switch); and (5) on 32-bit x86 only, floating-point calculations are done without the flags set which guarantee that the results are the same from run to run (
generate-accurate-x86-float-code-switch). Thus you may not be able cleanly to break out of an infinite loop; passing the wrong number of arguments may cause unrepeatable, possibly fatal errors; adding fixnums whose sum is a bignum will silently produce the wrong answer; and (again 32-bit x86 only) the lower bits of floating-point calculation may vary from run to run. We recommend that those switches should only be set in an unsafe manner when compiling functions where it is very unlikely that the code is incorrect. You can set safety to 0 when compiling such functions by use of a declaration within the defun form.
nil). Look particularly at initial values, making sure the initial value is of the declared type.
nil(as appropriate) as that can ensure the switch has a known, unambiguous value unaffected by declarations within a function definition or anything else.
The compiler in Allegro CL has the ability to compile floating-point operations in-line. In order to take full advantage of this feature, the compiler must know what can be done in-line. For it to know this, the user must, through declarations, inform the compiler of the type of arguments to operations. The compiler will attempt to propagate this type information to other operations, but this is not as easy as it might seem on first glance. Because mathematical operations in Common Lisp are generic, that is they accept arguments of any type - fixed, real, complex - and produce results of the appropriate type, the compiler cannot assume results in cases when the user may think the situation is clear. For example, the user may know that only positive numbers are being passed to sqrt, but unless the compiler knows it too, it will not assume the result is real and not complex. The compiler can tell the user what it does know and what it is doing. See Section 10.0 Help with declarations: the :explain declaration below. With this information, the user can add declarations as needed to speed up the compiled code. The process should be viewed as interactive, with the user tuning code in response to what the compiler says it knows.
The compiler will expand in-line (open-code) a floating-point numeric function only if the function is one which it knows how to open-code, and the types of the function's arguments and result are known at compile time to be those for which the open-coder is appropriate. Finally, at the point of compilation the compiler switch function which is the value of comp:trust-declarations-switch must return true. The default version of this function returns true if speed is greater than safety but users can change the values for which this switch returns true. explain-compiler-settings can be called to see what the value of that switch (and all others) will be given the current values of safety, space, speed, and debug.
The floating-point functions below are subject to opencoding. In each case, the function result must be declared or derivable to be either single-float or double-float. (Note that in this implementation, the type short-float is equivalent to single-float and long-float is equivalent to double-float.) The arguments to the arithmetic and trigonometric functions must be specifically one or the other of the two floating types, or signed integer constants representable in 32 bits, or else computed integer values of fixnum range. If these conditions are not met, the function will not be open-coded and instead a normal call to the generic version of the function will be compiled. Note that it is not sufficient to declare a value to be a float, it must be declared as either a single-float or a double-float.
We are often asked exactly which functions will opencode in Allegro CL. Unfortunately, it is not easy to provide an answer. Firstly, because the compiler is different on each different architecture, the answer is different in different implementations. Secondly, the same function may opencode in one case but not in another on the same machine because of the way it is affected by surrounding code. However, the following functions are candidates to be open-coded in most platforms or in the platforms noted.
(simple-array single-float (*))
(simple-array double-float (*))
Note that the list is not exhaustive in either direction. Not all
the listed functions and operations opencode on all machines and
functions and operations not listed do opencode on some machines. The
:explain declaration described in the next section
assists with determining what did and did not opencode.
The compiler must be supplied with type declarations if it is to open-code certain common operations without type-checking. The compiler includes a type propagator that tries to derive the results of a function call (or special operator) from what it knows about its arguments (or subforms). The type propagator greatly reduces the amount of type declaration necessary to achieve opencoding. However, the programmer may need to examine the inferences made by the propagator to determine what additional declarations would increase code speed. Although supplying type declarations to the compiler is very simple, it is a task surprisingly prone to error. The usual error is that well-intentioned declarations are insufficient to tell the compiler everything it needs to know.
For example, the programmer might neglect to declare the seemingly obvious fact that the result of a certain sqrt of a double-float is also a double-float. However, sqrt in Common Lisp returns a complex if its argument is negative, so the compiler cannot assume a real result. The only impact of insufficient declarations is that some open-coders will not be invoked. However, it can be awkward for the user to determine whether or not any particular call was open-coded, and if not, why. trace and disassemble will provide the information, but these are clumsy tools for the purpose.
In order to provide better information about what the compiler is doing, a new declaration, :explain, has been added to Allegro CL.
Arguments: [:calls | (:calls t) | (:calls nil)] [:types | (:types t) | (:types nil)] [:boxing | (:boxing t) | (:boxing nil)] [:variables | (:variables t) | (:variables nil)] [:tailmerging | (:tailmerging t) | (:tailmerging nil)] [:inlining | (:inlining t) | (:inlining nil)]
This declaration instructs the compiler to report or not to report
information about argument types and non-in-line calls, boxed floats
and variables stored in registers, tail-merging, and why inlining
might not have succeeded. Reporting is enabled when a quality appears
alone or in a list with
t. It is disabled
when a quality appears in a list with
nil. Initially, no
qualities are enabled.
This declaration may be placed anywhere that a normal declaration may be, and can be proclaimed (with proclaim or declaim). The compiler will report information within the scope of the declaration. Note that results may differ from one platform to another. Again, the general form of the declaration is (these forms must be placed in a location where declares are allowed, of course):
(declare (:explain :quality ...)) ;; explanation for :quality ;; will be output ;; or (declare (:explain (:quality t) ...) ;; explanation for :quality ;; will be output ;; or (declare (:explain (:quality nil) ...) ;; explanation for :quality ;; will not be output ;; Thus (declare (:explain :types (:boxing nil) (:variables t) :tailmerging)) ;; will enable explaining for :types, :variables, and :tailmerging ;; and disable it for :boxing. Explaining for :calls is not affected ;; (off if it was already off, on if it was already on).
The arguments control various kinds of reporting the compiler will make during compilation. The arguments may appear in either order, and either may be omitted. By default, no information of either type is printed. The declaration causes the compiler to print information during compilation, and obeys normal declaration scoping rules. It has no effect on code generation. Here are the various :explain qualities:
:types: when this quality is in effect, the compiler will report for each function call it compiles (in-line or not) the types of each argument along with the result type. Further information and examples are in Calls and types explanation in compiler-explanations.htm.
:calls: when this quality is in effect, the compiler will report when it generates code for any non-in-line function call.
:callsoperate independently. Further information and examples are in Calls and types explanation in compiler-explanations.htm.
:callsshould not be used with
:inliningsince calls information is provided by the :inlining explanation.
:boxing: when this quality is in effect, the compiler will tell you when code is generated to box a number if it is possible for the code not to be boxed. A number is `boxed' when it is converted from its machine representation to the Lisp representation. For floats, the machine representation is one (for singles) or two (for doubles) words. Lisp adds an extra word, which contains a pointer and a type code. For fixnums, boxing simply involves a left shift of two bits. For bignums which are in the range of machine integers, boxing again adds an additional word. Further information and examples are in Boxing explanation in compiler-explanations.htm.
:boxingshould not be used with
:inliningsince boxing information is provided by the :inlining explanation.
:variables: when this quality is in effect, the compiler will report whether local variables (in function definitions) are being stored in registers or in memory. Since storing variables in registers results in faster code, the information printed when this variable is in effect may help in recasting function definitions to allow for more locals to be stored in registers. Further information and examples are in Variables explanation in compiler-explanations.htm.
:tailmerging: when this quality is in effect, the compiler will report why a tail-merge is or is not being done for a function in tail-position. Further information and examples are in Tail-merging explanation in compiler-explanations.htm.
:inlining: when this quality is in effect, information about why a function you might expect to be inlined was not in fact inlined by the compiler. Further information and examples are in Inlining explanation in compiler-explanations.htm. (Essentially all the information printed by the
:explain :boxingdeclarations is printed by the
:explain :inliningdeclaration, so you should use either :inlining or :boxing and :calls, but not both.)
The reports are printed during the code generation phase of compilation. Code generation occurs fairly late during compilation, after macroexpansion and other code transformations have taken place. The function calls explained by the compiler will therefore deviate in certain ways from the original user code. For example, a two-operand + operation is transformed into a call to the more efficient excl::+_2op function. In general, the user should easily be able to relate the code generator's output to the original code.
The code generator works by doing a depth-first walk of the
transformed code tree it receives from early compiler phases. In this
tree walk, the
:types printout happens during the descent
into a branch of the tree. The
:calls printout happens as
the instructions are actually generated, during the ascending return
from the branch.
:explain is implemented as a declaration
is so the user can gain fairly fine-grained control of its scope. This
can be important when tuning large functions. The tool does require
editing the source code to be compiled, but presumably the user is in
the process of editing the code to add declarations anyway. These
options may also be enabled and disabled globally by use of proclaim or declaim, although they may produce
a lot of output.
The lines of explanation output are labeled with abbreviations which identify what type of explanation is being produced and what the compiler is doing. These abbreviations are listed in the compiler-explanations.htm.
There are other declarations which affect (or in some cases do not affect) the operation of the compiler, as we describe under the next several headings.
inline declaration is ignored by the compiler, but
see define-compiler-macro, which will
define a macro for the compiler to use for an operator which is not
At appropriate settings of speed and safety, the compiler will inline
whatever it can. Only predefined system functions can be inlined. User
defined functions are never compiled inline. (The compiler will
notinline declaration, however, so you can
suppress inlining of specific functions if you want.)
We have been asked why the inline declaration is ignored. It is not that we believe that inlining user functions does not provide any advantages, it is that we believe that there are other improvements that will provide more advantages. Because we have now implemented compiler environments (see environments.htm), we are in a better position to implement inlining of user functions in a later release.
Note that inlining is not an unmixed blessing. It increases the amount of space used by a function (since both the function definition and the block to stick in when the function is inlined have to be created and stored) and it makes debugging harder (by making the compiled code diverge from the source code). It is also prone to subtle errors (on the programmers side) and bugs (on our side).
Defstruct accessors will be compiled inline under two conditions:
nil. Using the default value, that switch will return
nilwhen speed is 3 and safety is 0 or 1.
On some machines, stack consing of &rest arguments, closures, and
some lists and vectors is supported if declared as
dynamic-extent. (Stack consing means that objects are
stored on the stack and not written to memory. This can provide a
significant space saving if the objects are truly temporary, which
they often are.) Here is an example of such a declaration for an
&rest argument. The function foo checks whether
the first argument is identical to any later argument.
(defun foo (a &rest b) (declare (dynamic-extent b)) (dolist (val b) (if* (eq a val) then (return t))))
Please note that care should exercised when using stack consing and multiprocessing, since a switch between one thread/process and another causes (part of) the stack to be saved, so the stack-consed objects and values are not available to the thread/process being switched into. For a binding or object that has dynamic extent, that extent is only valid when Lisp is executing on the thread/process that creates the binding or stack-conses the object. When (on Windows, which uses OS-threads) a thread switch is executed, the extents of all dynamic-extent data and bindings are temporarily exited, to be re-established when another thread switch returns to the original thread. Since only on Windows separate Lisp lightweight processes run on separate threads, it is important on Unix that dynamic-extent data (which may be stack-consed) not be referenced while executing on a different process. On Unix, a process-wait wait-function argument should never be declared dynamic-extent, since it was funcall'ed from other stack-groups. However, on Windows, wait functions are run only in their own threads, so stack consing in wait functions should work.
Dynamic-extent declarations are only observed at values of safety,
space, speed, and debug for which
returns true and
trust-declarations-switch is also true.
If a call to make-list has a constant size, declarations are trusted, the list is made the value of a variable, and the variable is declared as dynamic-extent, then it will be stack-allocated and initialized. The initial-value keyword can be used to specify the value. An attempt to make a list of a variable size with make-list will result in heap consing.
Dynamic-extent argument properties are automatically declared on all defun forms. This gives the compiler the ability to make assumptions about the dynamic extent use of arguments passed into these functions, and to generate more efficient code. The compiler generally tracks these properties for functions that it knows about, e.g. mapcar. Dynamic-extent declarations extend this to user-defined functions and to redefinitions. Warnings are also provided for redeclared definitions and for definitions that occur after the function's first usage. To allow declaration before the first use, a new macro called defun-proto is provided.
Avoiding consing with apply using a &rest
In certain cases apply is now compiled more efficiently, to remove consing. This is the so-called applyn optimization. Consider the following code:
(defun foo (x &rest y) (apply some-fun 1 x 2 y))
&rest argument is used for nothing more than
to supply a variable length argument list to
apply. This case is now compiled in such a way that
&restargument is considered dead and as if declared ignored.
In this optimized case, the code works exactly as it did when the
&rest argument was consed, but without the
consing. Circumstances that will cause this optimization to not be
used are if the &rest argument:
Optimization hint: If you have a function like
(defun wrap-it (flag &rest x &key &allow-other-keys) (when flag (setq x (list* :new-key 10 x))) (apply 'wrapped flag x))
then the optimization will not take effect. If non-consed operation is desired, then the following modification will allow the optimization:
(defun wrap-it (flag &rest x &key &allow-other-keys) (if flag (apply 'wrapped flag :new-key 10 x) (apply 'wrapped flag x)))
The type of arrays (and thus vectors) supported in Allegro CL are discussed in Data types and array types in implementation.htm. Of those, the following types of vectors can now be stack-allocated (N/A means not stack-allocable on the indicated platform, in some cases because that type of specialized array is not supported and in other cases because that type of specialized simply cannot be allocated on the stack; * means the initial-element value is ignored when the array is stack allocated):
|element type||initializable in 32-bit Lisps? (see below)||initializable in 64-bit Lisps? (see below)|
To get a stack-allocated vector, the following coding practice should be used:
(declare (optimize <values that make
trust-declarations-switchtrue>)) ;; with initial variable value, (speed 3) (safety 1) will work ... (let ((x (make-array n <options>))) (declare (dynamic-extent x)) ...)
n is a constant integer, and options
are limited to
:initial-element, if the array is
initializable according to the table above). All
other forms might cause heap allocation of the array. Specifying an
initial-element when the array is not, according
to the table above, initializable will cause the array to be heap
allocated except in the cases denoted by a * (single-float
arrays on 32-bit and double-float arrays on 64-bit): in those cases
the array will be stack-allocated but the value specified by the
initial-element argument will be ignored.
Other functions that allow stack-consing if conditions are right are cons, list, list*, make-list, with-stack-fobject, with-stack-fobjects, with-stack-list, and with-stack-list*.
A typep-transformer allows the compiler to transform a form like
(typep x 'foo)
into a form like
(funcall some-function x)
For example, the Allegro CL compiler will already transform typep forms where the type is defined as:
(deftype foo () `(satisfies foop))
(funcall 'foop x)
The ability to add typep-transformers described here allows types defined with a more complicated syntax than (satisfies 'some-function) to be similarly transformed. The user must supply the appropriate predicate function and call the following function to associate the type with the predicate.
The function add-typep-transformer takes a type and a predicate which will allow the compiler to transform forms like
(typep x 'type)
into the form:
(funcall <predicate> x)
The function remove-typep-transformer removes the association between the type and the predicate.
For example, suppose we have defined a type called
angle, which is a normalized angular
(deftype angle () '(real #.(* -2 pi) #.(* 2 pi)))
Suppose further that we have a time critical function that takes two arguments, checks to be sure they are angles, adds them together, checks to make sure the result is an angle, and returns it:
(defun add-two-angles (a b) (declare (optimize (speed 3))) (unless (typep a 'angle) (error "~s is not an angle" a)) (unless (typep b 'angle) (error "~s is not an angle" b)) (let ((sum (+ (the angle a) (the angle b)))) (unless (typep sum 'angle) (error "Sum (~s) of angles is not an angle" sum)) sum))
As is, 10,000 calls of this function (given legal arguments) takes about 1.5 CPU seconds (on my test machine). Suppose we're unhappy with that and want to speed it up by adding a typep-transformer, without having to change the coding of add-two-angles. Further suppose that we're usually dealing with single precision floating point numbers. Here's a way we can do it. We first define our predicate function. Note that we put the most likely case (single-float) as the first choice in the typecase form.
(defun anglep (x) (declare (optimize (speed 3) (safety 0))) (typecase x (single-float (and (>= (the single-float x) #.(float (* -2 pi) 0.0s0)) (<= (the single-float x) #.(float (* 2 pi) 0.0s0)))) (fixnum (and (>= (the fixnum x) #.(truncate (* -2 pi))) (<= (the fixnum x) #.(truncate (* 2 pi))))) (double-float (and (>= (the double-float x) #.(float (* -2 pi) 0.0d0)) (<= (the double-float x) #.(float (* 2 pi) 0.0d0))))))
We then call add-typep-transformer to make the compiler aware of our predicate:
(excl:add-typep-transformer 'angle 'anglep)
Now if we recompile add-two-angles and call it another 10,000 times with the same arguments, it only takes about .25 CPU seconds, a 6 fold improvement (you may see a different speedup ratio).
See the discussion in implementation.htm and the
for information on special handling of certain top-level forms in a
file. The issue is whether the forms are treated (during the compile)
as if wrapped in an
(eval-when (compile) ...)
A top-level form in a file is one which is not a subform of anything except perhaps progn. In CLtL-1 CL, top-level forms involving calls to the following functions were treated as if wrapped in such an eval-when when compiled. In ANSI CL, they are not. You can arrange to have the CLtL-1 behavior as described in implementation.htm. The affected functions are:
make-package shadow shadowing-import export unexport require use-package unuse-package import
Consider the following definition:
(defun foo (x123 y123 z123) (declare (fixnum x123 y123 z124)) (+ x123 y123 z123))
Unless care is taken in looking at this definition, it is easy not to notice that there is a spelling error in this definition (z124 instead of z123). And yet, since declarations are optional in Common Lisp, such a spelling error would not make a functional difference in the compiled code, and so it is likely to go completely unnoticed, even when the code goes into production, unless careful optimization is done.
When compiling this function, this kind of error results in a warning. Specifically, if a type declaration finds a name which is unknown lexically, a warning will be generated. In this eample case, the warning will be for z124, which apparently should have corresponded to z123 instead.
This warning will not occur for any correct portable code. The only questionable situation is the case where the variable is truly unknown, and thus would have already been giving a warning for a variable that will be assumed to be special, as in this example:
(defun xxx () (declare (fixnum xyz)) (1+ xyz))
Now, two warnings will (likely) be given: one for the fact that
xyz is unknown in the declaration, and one for the
fact that xyz is unknown in the code and is this assumed special.
Copyright (c) 1998-2012, Franz Inc. Oakland, CA., USA. All rights reserved.
This page has had moderate revisions compared to the 8.1 page.
|Allegro CL version 8.2|
Moderately revised from 8.1. Minimal update since 8.2 release.