Trailing-Edge
-
PDP-10 Archives
-
decus_20tap1_198111
-
decus/20-0004/18lisp.doc
There are no other files named 18lisp.doc in the archive.
SECTION 18
1
THE COMPILER AND ASSEMBLER
18.1 The Compiler
The compiler is available in the standard INTERLISP system. It may be
used to compile individual functions as requested or all function
definitions in a standard format LOAD file. The resulting code may be
stored as it is compiled, so as to be available for immediate use, or it
may be written onto a file for subsequent loading. The compiler in
INTERLISP-10 also provides a means of specifying sequences of machine
instructions via ASSEMBLE.
The most common way to use the compiler is to compile from a symbolic
(prettydef) file, producing a corresponding file which contains a set of
functions in compiled form which can be quickly loaded. An alternate
way of using the compiler is to compile from functions already defined
in the user's INTERLISP system. In this case, the user has the option
of specifying whether the code is to be saved on a file for subsequent
loading, or the functions redefined, or both. In either case, the
compiler will ask the user certain questions concerning the compilation.
The first question is:
LISTING?
The answer to this question controls the generation of a listing and is
explained in full below. However, for most applications, the user will
want to answer this question with either ST or F, which will also
specify an answer to the rest of the questions which would otherwise be
asked. ST means the user wants the compiler to STore the new
definitions; F means the user is only interested in compiling to a File,
and no storing of definitions is performed. In both cases, the compiler
will then ask the user one more question:
------------------------------------------------------------------------
1
The INTERLISP-10 compiler itself, i.e., the part that actually
generates code, was written and documented by, and is the
responsibility of A.K. Hartley. The user interfaces, i.e., tcompl,
recompile, bcompl, and brecompile, were written by W. Teitelman.
18.1
OUTPUT FILE:
to which the user can answer:
N or NIL no output file.
File name file is opened if not already opened, and
compiled code is written on the file.
Example:
_COMPILE((FACT FACT1 FACT2))
LISTING? ST
OUTPUT FILE: FACT.COM
(FACT COMPILING)
.
.
2
(FACT REDEFINED)
.
.
.
(FACT2 REDEFINED)
(FACT FACT1 FACT2)
_
This process caused the functions FACT, FACT1, and FACT2 to be compiled,
redefined, and the compiled definitions also written on the file
FACT.COM for subsequent loading.
18.2 Compiler Questions
The compiler uses the free variables lapflg, strf, svflg, lcfil and
lstfil which determines various modes of operation. These variables are
set by the answers to the "compset" questions. When any of the top
level compiling functions are called, the function compset is called
which asks a number of questions. Those that can be answered "yes" or
"no" can be answered with YES, Y, or T for YES; and NO, N, or NIL for
NO. The questions are:
1. LISTING?
The answer to this question controls the generation of a listing.
Possible answers are:
3
1 Prints output of pass 1, the LAP macro code.
2 Prints output of pass 2, the machine code.
------------------------------------------------------------------------
2
compiler printout and error messages are explained on page
18.34-37.
3
The LAP and machine code are usually not of interest but can be
helpful in debugging macros.
18.2
YES Prints output of both passes.
NO Prints no listings.
The variable lapflg is set to the answer. If the answer is affirmative,
compset will type FILE: to allow the user to indicate where the output
is to be written. The variable lstfil is set to the answer.
There are three other possible answers to LISTING? - each of which
specifies a complete mode for compiling. They are:
S Same as last setting.
F Compile to File (no definition of functions).
ST STore new definitions.
STF STore new definitions, Forget exprs.
Implicit in these three are the answers to the questions on disposition
of compiled code and expr's, so questions 2 and 3 would not be asked if
1 were answered with S, F, ST, or STF.
2. REDEFINE?
YES Causes each function to be redefined as it is compiled.
The compiled code is stored and the function definition
changed. The variable strf is set to T.
NO Causes function definitions to remain unchanged. The
variable strf is set to NIL.
The answer ST or STF for the first question implies YES for this
question, F implies NO, and S makes no change.
3. SAVE EXPRS?
If answered YES, svflg is set to T, and the exprs are saved on the
property list of the function name. Otherwise they are discarded. The
answer ST for the first question implies YES for this question, F or STF
implies NO, and S makes no change.
4. OUTPUT FILE:
If the compiled definitions are to be written for later loading, you
should provide the name of a file on which you wish to save the code
that is generated. If you answer T or TTY:, the output will be typed on
the teletype (not particularly useful). If you answer N, NO, or NIL,
output will not be done. If the file named is already open, it will
continue to be used. The free variable lcfil is set to the name of the
file.
18.3 Nlambdas
When compiling the call to a function, the compiler must prepare the
arguments to the function in one of three ways:
1. Evaluated (SUBR, SUBR*, EXPR, EXPR*, CEXPR, CEXPR*)
2. Unevaluated, spread (FSUBR, FEXPR, CFEXPR)
3. Unevaluated, not spread (FSUBR*, FEXPR*, CFEXPR*)
18.3
In attempting to determine which of these three is appropriate, the
compiler will first look for a definition among the functions in the
file that is being compiled. If the function is not contained there,
the compiler will look for other information which can be supplied by
the user by including nlambda nospread functions on the list nlama (for
nlambda atoms), and including nlambda spread functions on the list nlaml
4
(for nlambda list), and including lambda functions on the list lams. If
5
the function is not contained in the file, or on the list nlama, nlaml,
or lams, the compiler will look for a current definition. If the
function is defined, its function type is assumed to be the desired
type. If it is not defined, the compiler assumes that the function is
6 7
of type 1, i.e., its arguments are to be evaluated. In other words,
if there are type 2 or 3 functions called from the functions being
compiled, and they are only defined in a separate file, they must be
included on nlama or nlaml, or the compiler will incorrectly assume that
their arguments are to be evaluated, and compile the calling function
correspondingly. Note that this is only necessary if the compiler does
not "know" about the function. If the function is defined at compile
time, or is handled via a macro, or is contained in the same group of
files as the functions that call it, the compiler will automatically
handle calls to that function correctly.
18.4 Global Variables
Variables that appear on the list globalvars or have the property
------------------------------------------------------------------------
4
Including functions on lams is only necessary to override in-core
nlambda definitions, since in the absence of other information, the
compiler assumes the function is a lambda.
5
The function can be defined anywhere in any of the files given as
arguments to bcompl, tcompl, brecompile or recompile.
6
Before making this assumption, if the value of compileuserfn is not
NIL, the compiler calls (the value of) compileuserfn giving it as
arguments cdr of the form and the form itself, i.e., the compiler
does (APPLY* COMPILEUSERFN (CDR form) form). If a non-NIL value is
returned, it is compiled instead of form. If NIL is returned, the
compiler compiles the original expression as a call to a
lambda-spread that is not yet defined. CLISP (Section 23) uses
compileuserfn to tell the compiler how to compile iterative
statements, IF-THEN-ELSE statements, and pattern match constructs.
7
The names of functions so treated are added to the list alams
(for assumed lamdas). alams is not used by the compiler; it is
maintained for the user's benefit, i.e., so that the user can check
to see whether any incorrect assumptions were made.
18.4
GLOBALVAR, with value T, are called global variables. Such variables
are always accessed through their value cell when they are used freely
in a compiled funtion. In other words, a reference to the value of this
variable is equivalent to (CAR (QUOTE variable)), regardless of whether
or not it appears on the stack, i.e., the stack is not even searched for
this variable when the compiled function is entered. Similarly, (SETQ
variable value) is equivalent to (RPLACA (QUOTE variable) value); i.e.,
it sets the top-level value.
All system parameters, unless otherwise specified, are global variables,
i.e., have on their property lists the property GLOBALVAR with value T,
8
e.g., brokenfns, editmacros, #rpars, dwimflg, et al. Thus, rebinding
these variables will not affect the behavior of the system: instead, the
variables must be reset to their new values, and if they are to be
restored to their original values, reset again. For example, the user
might write ...(SETQ globalvar new-value) form
(SETQ globalvar old-value). Note that in this case, if an error
occurred during the evaluation of form, or a control-D was typed, the
global variable would not be restored to its original value. The
function resetvar (described in Section 5) provides a convenient way of
resetting global variables in such a way that their values are restored
even if an error occurred or control-D is typed.
18.5 Compiler Functions
Note: when a function is compiled from its in core definition, i.e., via
compile, recompile, or brecompile, as opposed to tcompl or bcompl (which
uses the definitions on a file), and the function has been modified by
break, trace, breakin, or advise, it is first restored to its original
state, and a message printed out, e.g., FOO UNBROKEN. If the function
is not defined as an expr, its property list is searched for the
property EXPR (see savedef, Section 8). If there is a property EXPR,
its value is used for the compilation. If there is no EXPR and the
compilation is being performed by recompile or brecompile, the
definition of the function is obtained from the file (using loadfns).
Otherwise, the compiler prints (fn NOT COMPILEABLE), and goes on to the
next function.
compile[x;flg] x is a list of functions (if atomic, list[x] is
used). compile first asks the standard compiler
questions, and then compiles each function on x,
using its in-core definition. Value is x.
If compiled definitions are being dumped to a
file, the file is closed unless flg=T.
------------------------------------------------------------------------
8
Since the stack does not have to be searched to find the values of
these variables, a considerable savings in time is achieved,
especially for deep computations.
18.5
9
compile1[name;def] compiles def, redefining name if strf=T.
compile1 is used by compile, tcompl, and
recompile. If dwimifycompflg is T, or def
contains a CLISP declaration, def is dwimified
before compiling. See Section 23.
tcompl[files] tcompl is used to "compile files", i.e., given a
symbolic load file (e.g., one created by
prettydef), it produces a "compiled file" that
contains the same S-expressions as the original
symbolic file, except that (1) a special
FILECREATED expression appears at the front of
the file which contains information used by the
file package, and which causes the message
10
COMPILED ON followed by the date, to be
printed when the file is loaded; (2) every
defineq in the symbolic file is replaced by the
corresponding compiled definitions in the
11
compiled file; and (3) expressions of the form
(DECLARE: -- DONTCOPY --) that appear in the
symbolic file are not copied to the compiled
file. This "compiled" file can be loaded into
any INTERLISP system with load.
files is a list of symbolic files to be compiled
(if atomic, list[files] is used). tcompl asks
the standard compiler questions, except for
OUTPUT FILE: Instead, the output from the
compilation of each symbolic file is written on
12
a file of the same name suffixed with COM,
e.g., tcompl[(SYM1 SYM2)] produces two files,
------------------------------------------------------------------------
9
strf is one of the variables set by compset, described earlier.
10
The actual string printed is the value of compileheader, initially
"COMPILED ON". The user can reset compileheader, for example to
distinguish between files compiled by different systems.
11
The compiled definitions appear at the front of the compiled file,
i.e., before the other expressions in the symbolic file, regardless
of where they appear in the symbolic file.
12
The actual suffix used is the value of the variable compile.ext,
which is initially COM. The user can reset compile.ext or rename the
compiled file after it has been written, without adversely affecting
any of the system packages.
18.6
13
SYM1.COM and SYM2.COM.
tcompl processes each file one at a time,
reading in the entire file. For each
FILECREATED expression, the list of functions
that were marked as changed by the file package
14
(see Section 14) is noted, and the FILECREATED
expression is written onto the output file. For
each DEFINEQ expression, tcompl adds any
15
NLAMBDA's in the DEFINEQ to nlama or laml, and
16
adds LAMBDA's to the list lams, so that calls
to these functions will be compiled correctly.
Expressions beginning with DECLARE: are
processed specially as described below. All
other expressions are collected to be
subsequently written onto the output file.
After processing the file in this fashion,
17
tcompl compiles each funtion, and writes the
compiled definition onto the output file.
tcompl then writes onto the output file the
other expressions found in the symbolic file.
The value of tcompl is a list of the names of
the output files. All files are properly
terminated and closed. If the compilation of
------------------------------------------------------------------------
13
The file name is constructed from the name field only, e.g.,
tcompl[<BOBROW>FOO.TEM;3] produces FOO.COM on the connected
directory. The version number will be the standard default.
14
for use by recompile and brecompile which use the same low level
funtions as tcompl and bcompl.
15
described earlier, page 18.4.
16
nlama, nlaml, and lams are rebound to their top level values (using
resetvar) by tcompl, recompile, bcompl, brecompile, compile, and
blockcompile, so that any additions to these lists while inside of
these functions will not propagate outside.
17
except for those functions which appear on the list dontcompilefns,
initially NIL. For example, this option might be used for functions
that compile open, since their definitions would be superfluous when
operating with the compiled file. Note that dontcompilefns can be
set via block declarations page 18.21.
18.7
any file is aborted via an error or control-D,
all files are properly closed, and the
(partially complete) compiled file is deleted.
DECLARE:
For the purposes of compilation, DECLARE: (see Section 14) has two
principal applications: (1) to specify forms that are to be evaluated at
compile time, presumably to affect the compilation, e.g., to set up
macros; and/or (2) to indicate which expressions appearing in the
symbolic file are not to be copied to the output file. (Normally,
expressions are not evaluated and are copied.) Each expression in cdr of
a DECLARE: form is either evaluated/not-evaluated and copied/not-copied
depending on the settings of two internal state variables, initially set
for copy and not-evaluate. These state variables can be reset for the
remainder of the expressions in the DECLARE: by means of the tags
DOEVAL@COMPILE (or EVAL@COMPILE) and DONTCOPY, e.g.,
(DECLARE: DOEVAL@COMPILE DONTCOPY (DEFLIST -- (QUOTE MACRO))) could be
used to set up macros at compile time.
Recompile
The purpose of recompile is to allow the user to update a compiled file
without recompiling every function in the file. Recompile does this by
using the results of a previous compilation. It produces a compiled
file similar to one that would have been produced by tcompl, but at a
considerable savings in time by compiling selected functions and copying
from an earlier tcompl or recompile file the compiled definitions for
the remainder of the functions in the file.
recompile[pfile;cfile;fns]
pfile is the name of the pretty file to be
compiled, cfile is the name of the compiled file
containing compiled definitions that may be
copied. fns indicates which functions in pfile
are to be recompiled, e.g., have been changed or
defined for the first time since cfile was made.
Note that pfile, not fns, drives recompile.
recompile asks the standard compiler questions,
except for OUTPUT FILE:. As with tcompl, the
18 19
output automatically goes to pfile.COM.
------------------------------------------------------------------------
18
or pfile.ext, where ext is the value of compile.ext.
19
In general, all constructions of the form pfile.COM, pfileCOMS,
pfileBLOCKS, etc., are performed using the name field only. For
example, if pfile=<BOBROW>FOO.TEM;3, pfile.COM means FOO.COM,
pfileCOMS means FOOCOMS, etc.
18.8
recompile process pfile the same as does tcompl
except that DEFINEQ expressions are not actually
read into core. Instead, recompile uses the
20
filemap (see Section 14) to obtain a list of
the functions contained in pfile, and simply
21
skips over the DEFINEQ's.
After this initial scan of pfile, recompile then
processes the functions defined in the file.
For each function in pfile, recompile determines
whether or not the function is to be
22
(re)compiled. A function is to be recompiled
if (1) fns is a list and the function is a
member of that list; or (2) fns=T or EXPRS and
the function is an expr; or (3) fns=CHANGES and
the function is marked as having been changed in
23
the FILECREATED expression; or (4) fns=ALL. If
a function is not to be recompiled, recompile
obtains its compiled definition from cfile, and
copies it (and all generated subfunctions) to
24
the output file, pfile.COM. Finally, after
processing all functions, recompile writes out
all other expressions that were collected in the
prescan of pfile.
------------------------------------------------------------------------
20
A map is built if the symbolic file does not already contain one,
e.g., it was written in an earlier system, or with buildmapflg=NIL.
21
The filemap enables recompile to skip over the DEFINEQ's in the file
by simply resetting the file pointer, so that in most cases the scan
of the symbolic file is very fast (the only processing required is
the reading of the non-DEFINEQ's and the processing of the DECLARE:
expressions as described earlier).
22
Functions that are members of dontcompilefns are simply ignored.
23
In this latter case, cfile is superfluous, and in fact does not have
to exist. This option is useful, for example, to compile a symbolic
file that has never been compiled before, but which has already been
loaded (since using tcompl would require reading the file in a
second time).
24
If the function does not appear on cfile, recompile simply
recompiles it.
18.9
If cfile=NIL, pfile.COM is used for copying
25
from. If both fns and cfile are NIL, fns is
26
set to T, meaning recompile all exprs.
The value of recompile is the new compiled file,
pfile.COM. If recompile is aborted due to an
error or control-D, the new (partially complete)
compiled file will be closed and deleted.
recompile is designed to allow the user to conveniently and efficiently
update a compiled file, even when the corresponding symbolic file has
not been (completely) loaded. For example, the user can perform a
27
loadfrom (Section 14) to "notice" a symbolic file, and then simply
28 29
edit the functions he wanted to change, call makefile, and then
30
perform recompile[pfile].
------------------------------------------------------------------------
25
In other words, if cfile, the file used for obtaining compiled
definitions to be copied, is NIL, pfile.COM is used, i.e., same name
as output file but a different version number (one less) than the
output file.
26
This is the most common useage. Typically, the functions the user
has changed will have been unsavedefed by the editor, and therefore
will be exprs. Thus the user can perform his edits, dump the file,
and then simply recompile[file] to update the compiled file.
27
The loadfrom would be unnecessary if the compiled file had been
previously loaded, since this would also result in the file having
been 'noticed'.
28
As described in Section 9, the editor would automatically load those
functions not already loaded.
29
As described in Section 14, makefile would copy the unchanged functions
from the symbolic file.
30
Since prettydef automatically outputs a suitable DECLARE: expression to
indicate which functions in the file (if any) are defined as
NLAMBDA's, calls to these functions will be handled correctly, even
though the NLAMBDA functions themselves may never be loaded, or even
looked at, by recompile.
18.10
18.6 Open Functions
When a function is called from a compiled function, a system routine is
invoked that sets up the parameter and control push lists as necessary
for variable bindings and return information. As a result, function
calls can take up to 350 microseconds per call. If the amount of time
spent inside the function is small, this function calling time will be a
significant percentage of the total time required to use the function.
Therefore, many "small" functions, e.g., car, cdr, eq, not, cons are
always compiled "open", i.e., they do not result in a function call.
Other larger functions such as prog, selectq, mapc, etc. are compiled
open because they are frequently used. It is useful to know exactly
which functions are compiled open in order to determine where a program
is spending its time. Therefore below is a list of those functions
which when compiled do not result in function calls. Note that the next
section tells how the user can make other functions compile open via
31
MACRO definitions.
The following functions compile open in INTERLISP-10:
AC, ADD1, AND, APPLY*, ARG, ARRAYP, ASSEMBLE, ATOM, BLKAPPLY, BLKAPPLY*,
CAR, CDR, CAAR, ... CDDDAR, CDDDDR, CLOSER, COND, CONS, EQ, ERSETQ,
EVERY, EVQ, FASSOC, FCHARACTER, FDIFFERENCE, FGTP, FIX, FIXP, FLAST,
FLENGTH, FLOAT, FLOATP, FMEMB, FMINUS, FNTH, FPLUS, FQUOTIENT, FRPLACA,
FRPLACD, FSTKARG, FSTKNTH, FTIMES, FUNCTION, GETHASH, GETPROPLIST,
GETTOPVAL, GO, IDIFFERENCE, IEQP, IGREATERP, ILESSP, IMINUS, IPLUS,
IQUOTIENT, IREMAINDER, ITIMES, LIST, LISTP, LITATOM, LLSH, LOC, LOGAND,
LOGOR, LOGXOR, LRSH, LSH, MAP, MAPC, MAPCAR, MAPCON, MAPCONC, MAPLIST,
MINUSP, NEQ, NLISTP, NLSETQ, NOT, NOTEVERY, NOTANY, NTYP, NULL, NUMBERP,
OPENR, OR, PROG, PROG1, PROGN, RESETFORM, RESETLST, RESETSAVE, RESETVAR,
RETURN, RPTQ, RSH, SELECTQ, SETARG, SETN, SETPROPLIST, SETQ, SETTOPVAL,
SMALLP, SOME, STRINGP, SUB1, SUBSET, TYPEP, UNDONLSETQ, VAG, ZEROP.
18.7 Compiler Macros
The INTERLISP compiler includes a macro capability by which the user can
affect the compiled code. Macros are defined by placing the macro
definition on the property list of the corresponding function under the
32
property MACRO. When the compiler begins compiling a form, it
retrieves a macro definition for car of the form, if any, and uses it to
------------------------------------------------------------------------
31
The user can also affect the compiled code via compileuserfn,
described in footnote on page 18.4.
32
An expression of the form (DECLARE (DEFLIST ... (QUOTE MACRO))) can
be used within a function to define a MACRO. DECLARE is defined the
same as QUOTE and thus can be placed so as to have no effect on the
running of the function.
18.11
33
direct the compilation. The three different types of macro definitions
are given below.
(1) Open macros - (LAMBDA ...) or (NLAMBDA ...)
A function can be made to compile open by giving it a macro definition
of the form (LAMBDA ...) or (NLAMBDA ...), e.g.,
(LAMBDA (X) (COND ((GREATERP X 0) X) (T (MINUS X)))) for abs. The
effect is the same as though the macro definition were written in place
of the function wherever it appears in a function being compiled, i.e.,
it compiles as an open LAMBDA or NLAMBDA expression. This saves the
time necessary to call the function at the price of more compiled code
generated.
(2) Computed macros - (atom expression)
A macro definition beginning with an atom other than LAMBDA, NLAMBDA, or
NIL, allows computation of the INTERLISP expression that is to be
compiled in place of the form. The atom which starts the macro
definition is bound to cdr of the form being compiled. The expression
following the atom is then evaluated, and the result of this evaluation
34
is compiled in place of the form. For example, list could be compiled
this way by giving it the macro definition:
[X (LIST (QUOTE CONS)
(CAR X)
(AND (CDR X)
(CONS (QUOTE LIST)
(CDR X]
This would cause (LIST X Y Z) to compile as
(CONS X (CONS Y (CONS Z NIL))). Note the recursion in the macro
35
expansion. Ersetq, nlsetq, map, mapc, mapcar, mapconc, and some, are
------------------------------------------------------------------------
33
The compiler has built into it how to compile certain basic
functions such as car, prog, etc., so that these will not be
affected by macro definitions. These functions are listed above.
However, some of them are themselves implemented via macros, so that
the user could change the way they compile.
34
In INTERLISP-10, if the result of the evaluation is the atom
INSTRUCTIONS, no code will be generated by the compiler. It is then
assumed the evaluation was done for effect and the necessary code,
if any, has been added. This is a way of giving direct instructions
to the compiler if you understand it.
35
list is actually compiled more efficiently.
18.12
compiled via macro definitions of this type.
(3) Substitution macro - (NIL expression) or (list expression)
Each argument in the form being compiled is substituted for the
corresponding atom in car of the macro definition, and the result of the
substitution is compiled instead of the form, i.e.,
(SUBPAIR (CAR macrodef) (CDR form) (CADR macrodef)). For example, the
macro definition of add1 is ((X) (IPLUS X 1)). Thus, (ADD1 (CAR Y)) is
compiled as (IPLUS (CAR Y) 1). The functions add1, sub1, neq, nlistp,
zerop, flength, fmemb, fassoc, flast, and fnth are all compiled open
using substitution macros. Note that abs could be compiled open as
shown earlier or via a substitution macro. A substitution macro,
however, would cause (ABS (FOO X)) to compile as
(COND ((GREATERP (FOO X) 0) (FOO X)) (T (MINUS (FOO X)))) and
consequently (FOO X) would be evaluated three times.
18.8 FUNCTION and Functional Arguments
Expressions that begin with FUNCTION will always be compiled as separate
36
functions named by attaching a gensym to the end of the name of the
37
function in which they appear, e.g., FOOA0003. This gensym function
will be called at run time. Thus if FOO is defined as
(LAMBDA (X) ... (FOO1 X (FUNCTION ...)) ...) and compiled, then when FOO
38
is run, FOO1 will be called with two arguments, X, and FOOA000n, and
then FOO1 will call FOOA000n each time it must use its functional
argument.
Note that a considerable savings in time could be achieved by defining
FOO1 as a computed macro of the form:
(Z (LIST (SUBST (CADADR Z) (QUOTE FN) def) (CAR Z)))
where def is the definition of FOO1 as a function of just its first
argument and FN is the name used for its functional argument in its
definition. The expression compiled contains what was previously the
functional argument to FOO1, as an open LAMBDA expression. Thus you
------------------------------------------------------------------------
36
except when they are compiled open, as is the case with most of the
mapping functions.
37
nlsetq and ersetq expressions also compile using gensym functions.
As a result, a go or return cannot be used inside of a compiled
nlsetq or ersetq if the corresponding prog is outside, i.e., above
the nlsetq or ersetq.
38
or an appropriate funarg expression, see Section 11.
18.13
save not only the function call to FOO1, but also each of the function
calls to its functional argument. For example, if FOO1 operates on a
list of length ten, eleven function calls will be saved. Of course,
this savings in time cost space, and the user must decide which is more
important.
18.9 Block Compiling
Block compiling provides a way of compiling several functions into a
single block. Function calls between the component functions of the
block are very fast, and the price of using a free variable, namely the
time required to look up its value on the stack, is paid only once -
when the block is entered. Thus, compiling a block consisting of just
a single recursive function may be yield great savings if the function
calls itself many times, e.g., equal, copy, and count are block compiled
in INTERLISP.
The output of a block compilation is a single, usually large, function.
This function looks like any other compiled function; it can be broken,
advised, printstructured, etc. Calls from within the block to functions
outside of the block look like regular function calls, except that they
are usually linked (described below). A block can be entered via
several different functions, called entries. These must be specified
39
when the block is compiled. For example, the error block has three
entries, errorx, interrupt, and fault1. Similarly, the compiler block
has nine entries.
Specvars
One savings in block compiled functions results from not having to store
on the stack the names of the variables bound within the block, since
the block functions all "know" where the variables are stored. However,
if a variable bound in a block is to be referenced outside the block, it
40
must be included on the list specvars. For example, helpclock is on
specvars, since it is rebound inside of lispxblock and editblock, but
the error functions must be able to obtain its latest value.
------------------------------------------------------------------------
39
Actually the block is entered the same as every other function,
i.e., at the top. However, the entry functions call the main block
with their name as one of its arguments, and the block dispatches on
the name, and jumps to the portion of the block corresponding to
that entry point. The effect is thus the same as though there were
several different entry points.
40
Arguments to the block that are referenced freely outside the block
must also be SPECVARS if they are reset within the block, or else
the new value will not be obtained.
18.14
Localfreevars
Localfreevars is a feature designed for those variables which are used
freely by one or more of the block functions, but which are always bound
(by some other block function) before they are referenced, i.e., their
free values above the block are never used. Normally, when a block is
entered, all variables which are used freely by any function in the
block are looked up and pointers to the bindings are stored on the
stack. When any of these variables are rebound in the block, the old
pointer is saved and a pointer to the new binding is stored in the
original stack position. It frequently happens that variables used
freely within a block are in fact always bound within the block prior to
the free reference. The unnecessary lookup of the value of the free
variable at the time of entry to the block can be avoided by putting the
variable name on the list localfreevars. If a variable is on
localfreevars, its value will not be looked up at the time of entry.
When the variable is bound, the value will be stored in the proper stack
position. Should the variable in fact be referenced before it is bound,
the program will still work correctly. Invisible to the user, a rather
time-consuming process will take place. The reference will cause a trap
which will invoke code to determine which variable was referenced and
look up the value. Future references to that variable during this call
to the block will be normal, i.e., will not cause a trap.
trapcount[x] is a function to monitor the performance of
block compiled code with respect to
localfreevars. If x is NIL, trapcount returns
the cumulative number of traps caused by
localfreevars that were not bound before use.
If x is a number, the trapcount is reset to that
number.
evq is another compiler artifice for free variables references. (EVQ X)
has the effect of (EVAL (QUOTE X)) without the call to eval (if X is an
atom). evq is intended primarily for use in conjunction with
localfreevars. For example, suppose a block consists of three
functions, FOO1, FOO2, and FOO3, with FOO1 and FOO2 being entries, and
FOO3 using X freely, where X is bound in FOO1, but not in FOO2, i.e.,
FOO1 rebinds X, but when entered via FOO2, the user intends X to be used
freely, and its higher value obtained. If X is on localfreevars, then
each time the block is entered via FOO2, a trap will occur when FOO3
first references X. In order to avoid this, the user can insert (EVQ X)
in FOO2. This will circumvent the trap by explicitly invoking the
routine that searches back up the stack for the last binding of X.
Thus, when used with localfreevars, evq does two things: it returns the
value of its argument, and also stores that value in the binding slot
for the variable so that no future references to that variable (in this
call) will cause traps. Since the time consumed by the trap can greatly
exceed the time required for a variable lookup, using evq in these
situations can result in a considerable savings.
Retfns
Another savings in block compilation arises from omitting most of the
information on the stack about internal calls between functions in the
18.15
block. However, if a function's name must be visible on the stack,
e.g., if the function is to be returned from retfrom, it must be
included on the list retfns.
Blkapplyfns
Normally, a call to apply from inside a block would be the same as a
call to any other function outside of the block. If the first argument
to apply turned out to be one of the entries to the block, the block
would have to be reentered. blkapplyfns enables a program to compute
the name of a function in the block to be called next, without the
overhead of leaving the block and reentering it. This is done by
including on the list blkapplyfns those functions which will be called
in this fashion, and by using blkapply in place of apply, and blkapply*
in place of apply*. For example, the calls to the functions handling
RI, RO, LI, LO, BI, and BO in the editor are handled this way. If
blkapply or blkapply* is given a function not on blkapplyfns, the effect
is the same as a call to apply or apply* and no error is generated.
Note however, that blkapplyfns must be set at compile time, not run
time, and furthermore, that all functions on blkapplyfns must be in the
block, or an error is generated (at compile time), NOT ON BLKFNS.
Blklibrary
Compiling a function open via a macro provides a way of eliminating a
function call. For block compiling, the same effect can be achieved by
including the function in the block. A further advantage is that the
code for this function will appear only once in the block, whereas when
a function is compiled open, its code appears at each place where it is
called.
The block library feature provides a convenient way of including
functions in a block. It is just a convenience since the user can
always achieve the same effect by specifying the function(s) in question
as one of the block functions, provided it has an expr definition at
compile time. The block library feature simply eliminates the burden of
supplying this definition.
To use the block library feature, place the names of the functions of
interest on the list blklibrary, and their EXPR definition on the
property list of the function under the property BLKLIBRARYDEF. When
the block compiler compiles a form, it first check to see if the
function being called is one of the block functions. If not, and the
function is on blklibrary, its definition is obtained from the property
value of BLKLIBRARYDEF, and it is automatically included as part of the
block. The functions assoc, equal, getp, last, length, lispxwatch,
memb, nconc1, nleft, nth, and /rplnode already have BLKLIBRARYDEF
properties.
18.10 Linked Function Calls
Conventional (non-linked) function calls from a compiled function go
through the function definition cell, i.e., the definition of the called
function is obtained from its function definition cell at call time.
Thus, when the user breaks, advises, or otherwise modifies the
18.16
definition of the function FOO, every function that subsequently calls
it instead calls the modified function. For calls from the system
functions, this is clearly not a feature. For example, the user may
wish to break on basic functions such as print, eval, rplaca, etc.,
which are used by the break package. In other words, we would like to
guarantee that the system packages will survive through user
modification (or destruction) of basic functions (unless the user
specifically requests that the system packages also be modified). This
protection is achieved by linked function calls.
For linked function calls, the definition of the called function is
obtained at link time, i.e., when the calling function is defined, and
stored in the literal table of the calling function. At call time, this
definition is retrieved from where it was stored in the literal table,
not from the function definition cell of the called function as it is
for non-linked calls. These two different types of calls are
illustrated in Figure 18-1.
Note that while function calls from block compiled functions are usually
linked, and those from standardly compiled functions are usually non-
linked, linking function calls and blockcompiling are independent
features of the INTERLISP compiler, i.e., linked function calls are
possible, and frequently employed, from standardly compiled functions.
18.17
Figure 18-1
18.18
Note that normal function calls require only the called function's name
in the literals of the compiled code, whereas a linked function call
uses two literals and hence produces slightly larger compiled functions.
The compiler's decision as to whether to link a particular function call
is determined by the variables linkfns and nolinkfns as follows:
(1) If the function appears on nolinkfns, the call is not linked;
(2) If block compiling and the function is one of the block
functions, the call is internal as described earlier;
(3) If the function appears on linkfns, the call is linked;
(4) If nolinkfns=T, the call is not linked;
(5) If block compiling, the call is linked;
(6) If linkfns=T, the call is linked;
(7) Otherwise the call is not linked.
Note that (1) takes precedence over (2), i.e., if a function appears on
nolinkfns, the call to it is not linked, even if it is one of the
functions in the block, i.e., the call will go outside of the block.
Nolinkfns is initialized to various system functions such as errorset,
break1, etc. Linkfns is initialized to NIL. Thus if the user does not
specify otherwise, all calls from a block compiled function (except for
those to functions on nolinkfns) will be linked; all calls from
standardly compiled functions will not be linked. However, when
compiling system functions such as help, error, arglist, fntyp, break1,
et al, linkfns is set to T so that even though these functions are not
block compiled, all of their calls will be linked.
If a function is not defined at link time, i.e., when an attempt is made
to link to it, it is linked instead to the function nolinkdef. When the
function is later defined, the link can be completed by relinking the
calling function using relink described below. Otherwise, if a function
is run which attempts a linked call that was not completed, nolinkdef is
called. If the function is now defined, i.e., it was defined at some
point after the attempt was made to link to it, nolinkdef will quietly
perform the link and continue the call. Otherwise, it will call
faultapply and proceed as described in Section 16.
Linked function calls are printed on the backtrace as ;fn; where fn is
the name of the function. Note that this name does not actually appear
on the stack, and that stkpos, retfrom, and the rest of the pushdown
list functions (Section 12) will not be able to find it. Functions
which must be visible on the stack should not be linked to, i.e.,
include them on nolinkfns when compiling a function that would otherwise
link its calls.
printstructure, calls, break on fn1-IN-fn2 and advise fn1-IN-fn2 all
work correctly for linked function calls, e.g., break[(FOO IN FIE)],
where FOO is called from FIE via a linked function call.
Relinking
The function relink is available for relinking a compiled function,
i.e., updating all of its linked calls so that they use the definition
extant at the time of the relink operation.
18.19
relink[fn] fn is either WORLD, the name of a function, a
list of functions, or an atom whose value is a
list of functions. relink performs the
corresponding relinking operations.
relink[WORLD] is possible because laprd
maintains on linkedfns a list of all user
functions containing any linked calls.
syslinkedfns is a list of all system functions
that have any linked calls. relink[WORLD]
performs both relink[linkedfns] and
relink[syslinkedfns].
The value of relink is fn.
It is important to stress that linking takes place when a function is
defined. Thus, if FOO calls FIE via a linked call, and a bug is found
in FIE, changing FIE is not sufficient; FOO must be relinked.
Similarly, if FOO1, FOO2, and FOO3 are defined (in that order) in a
file, and each call the others via linked calls, when a new version of
the file is loaded, FOO1 will be linked to the old FOO2 and FOO3, since
those definitions will be extant at the time it is read and defined.
Similarly, FOO2 will link to the new FOO1 and old FOO3. Only FOO3 will
link to the new FOO1 and FOO2. The user would have to perform
relink[FOOFNS] following the load.
18.11 The Block Compiler
There are three user level functions for blockcompiling, blockcompile,
bcompl, and brecompile, corresponding to compile, tcompl, and recompile.
All of them ultimately call the same low level functions in the
compiler, i.e., there is no 'blockcompiler' per se. Instead, when
blockcompiling, a flag is set to enable special treatment for specvars,
retfns, blkapplyfns, and for determining whether or not to link a
function call. Note that all of the previous remarks on macros,
globalvars, compiler messages, etc., all apply equally for block
compiling. Using block declarations described below, the user can
intermix in a single file functions compiled normally, functions
compiled normally with linked calls, and block compiled functions.
Blockcompile
blockcompile[blkname;blkfns;entries;flg]
blkfns is a list of the functions comprising the
block, blkname is the name of the block, entries a list of entries to
the block, e.g.,
_BLOCKCOMPILE(SUBPRBLOCK (SUBPAIR SUBLIS SUBPR) (SUBPAIR SUBLIS))
Each of the entries must also be on blkfns or an
18.20
41
error is generated, NOT ON BLKFNS.
If entries is NIL, list[blkname] is used, e.g.,
_BLOCKCOMPILE(COUNT (COUNT COUNT1))
If blkfns is NIL, list[blkname] is used, e.g.,
_BLOCKCOMPILE(EQUAL)
blockcompile asks the standard compiler
questions and then begins compiling. As with
compile, if the compiled code is being written
to a file, the file is closed unless flg=T. The
value of blockcompile is a list of the entries,
or if entries=NIL, the value is blkname.
The output of a call to blockcompile is one
function definition for blkname, plus
definitions for each of the functions on entries
if any. These entry functions are very short
functions which immediately call blkname.
Block Declarations
Since block compiling a file frequently involves giving the compiler a
lot of information about the nature and structure of the compilation,
e.g., block functions, entries, specvars, linking, et al, we have
implemented a special prettydef command to facilitate this
commmunication. The user includes in the third argument to prettydef a
command of the form
(BLOCKS block1 ... block2 ... blockn) where each block1 is a block
declaration. bcompl and brecompile described below are sensitive to
these declarations and take the appropriate action.
The form of a block declaration is:
(blkname blkfn1 ... blkfnm (var1 . value) ... (varn . value))
blkfn1 ... blkfnm are the functions in the block and correspond to
blkfns in the call to blockcompile. The (var . value) expressions
indicate the settings for variables affecting the compilation.
As an example, the value of editblocks is shown below. It consists of
three block declarations, editblock, editfindblock, and edit4e.
------------------------------------------------------------------------
41
If only one entry is specified, the block name can also be one of
the blkfns, e.g., BLOCKCOMPILE(FOO (FOO FIE FUM) (FOO)). However, if
more than one entry is specified, an error will be generated,
CAN'T BE BOTH AN ENTRY AND THE BLOCK NAME.
18.21
[RPAQQ EDITBLOCKS
((EDITBLOCK EDITL0 EDITL1 UNDOEDITL EDITCOM EDITCOMA EDITCOML
EDITMAC EDITCOMS EDIT]UNDO UNDOEDITCOM
UNDOEDITCOM1 EDITSMASH EDITNCONC EDIT1F EDIT2F
EDITNTH BPNT BPNT0 BPNT1 RI RO LI LO BI BO
EDITDEFAULT ## EDUP EDIT* EDOR EDRPT EDLOC EDLOCL
EDIT: EDITMBD EDITXTR EDITELT EDITCONT EDITSW
EDITMV EDITTO EDITBELOW EDITRAN TAILP EDITSAVE
EDITH (ENTRIES EDITL0 ## UNDOEDITL)
(SPECVARS L COM LCFLG #1 #2 #3 LISPXBUFS
**COMMENT**FLG PRETTYFLG UNDOLST
UNDOLST1)
(RETFNS EDITL0)
(GLOBALVARS EDITCOMSA EDITCOMSL EDITOPS
HISTORYCOMS EDITRACEFN)
(BLKAPPLYFNS RI RO LI LO BI BO EDIT: EDITMBD
EDITMV EDITXTR)
(BLKLIBRARY LENGTH NTH LAST)
(NOLINKFNS EDITRACEFN))
(EDITFINDBLOCK EDIT4E EDIT4E1 EDITQF EDIT4F EDITFPAT
EDITFPAT1 EDIT4F1 EDIT4F2 EDIT4F3 EDITSMASH
EDITFINDP EDITBF EDITBF1 ESUBST
(ENTRIES EDITQF EDIT4F EDITFPAT EDITFINDP
EDITBF ESUBST))
(EDIT4EBLOCK EDIT4E EDIT4E1 (ENTRIES EDIT4E EDIT4E1]
42
Whenever bcompl or brecompile encounter a block declaraction they
rebind retfns, specvars, localfreevars, globalvars, blklibrary,
nolinkfns, linkfns, and dontcompilefns to their top level value, bind
blkapplyfns and entries to NIL, and bind blkname to the first element of
the declaration. They then scan the rest of the declaration, gathering
up all atoms, and setting car of each nonatomic element to cdr of the
expression if atomic, e.g., (LINKFNS . T), or else to union of cdr of
43
the expressions with the current (rebound) value, e.g.,
(GLOBALVARS EDITCOMSA EDITCOMSL). When the declaration is exhausted,
the block compiler is called and given blkname, the list of block
functions, and entries.
Note that since all compiler variables are rebound for each block
declaration, the declaration only has to set those variables it wants
changed. Furthermore, setting a variable in one declaration has no
effect on the variable's value for another declaration.
After finishing all blocks, bcompl and brecompile treat any functions in
------------------------------------------------------------------------
42
The BLOCKS command outputs a DECLARE expression, which is noticed by
bcompl and brecompile.
43
Expressions of the form (var * form) will cause form to be evaluated
and the resulting list used as described above, e.g.,
(GLOBALVARS * MYGLOBALVARS).
18.22
the file that did not appear in a block declaration in the same way as
do tcompl and recompile. If the user wishes a function compiled
separately as well as in a block, or if he wishes to compile some
functions (not blockcompile), with some compiler variables changed, he
can use a special pseudo-block declaration of the form
(NIL fn1 ... fnm (var1 . value) ... (varn . value)) which means compile
fn1 ... fnm after first setting var1 ... varn as described above. For
example, (NIL CGETD FNTYP ARGLIST NARGS NCONC1 GENSYM (LINKFNS . T))
appearing as a "block declaration" will cause the six indicated
functions to be compiled while linkfns=T so that all of their calls will
be linked (except for those functions on nolinkfns).
Bcompl
bcompl[files;cfile] files is a list of symbolic files. (If atomic,
list[files] is used.) bcompl differs from tcompl
in that it compiles all of the files at once,
instead of one at a time, in order to permit one
44
block to contain functions in several files.
Output is to cfile if given, otherwise to a file
45
whose name is car[files] suffixed with COM,
e.g., bcompl[(EDIT WEDIT)] produces one file,
EDIT.COM.
bcompl asks the standard compiler questions,
except for OUTPUT FILE:, then processes each
file exactly the same as does tcompl (see page
46
18.7). Bcompl next processes the block
declarations as described above. Finally, it
compiles those functions not mentioned in one of
the block declarations, and then writes out all
other expressions.
The value of bcompl is the output file (the new
compiled file). If the compilation is aborted
due to an error or control-D, all files are
closed and the (partially complete) output file
is deleted.
------------------------------------------------------------------------
44
Thus if you have several files to be bcompled separately, you must
make several calls to bcompl.
45
or value of compile.ext, as explained earlier.
46
In fact, tcompl is defined in terms of bcompl. The only difference
is that tcompl calls bcompl with an extra argument specifying that
all block declarations are to be ignored.
18.23
Note that it is permissible to tcompl files set
up for bcompl; the block declarations will
simply have no effect. Similarly, you can
bcompl a file that does not contain any block
declarations and the result will be the same as
having tcompled it.
Brecompile
Brecompile plays the same role for bcompl that recompile plays for
tcompl: its purpose is to allow the user to update a compiled file
without requiring an entire bcompl.
brecompile[files;cfile;fns]
files is a list of symbolic files (if atomic,
list[files] is used). cfile is the compiled
file corresponding to bcompl[files] or a
previous brecompile, i.e., it contains compiled
definitions that may be copied. The
interpretation of fns is the same as with
47
recompile.
brecompile asks the standard compiler questions
except for OUTPUT FILE: As with bcompl, output
automatically goes to file.COM, where file is
the first file in files.
brecompile processes each file the same as does
recompile as described on page 18.8, then
processes each block declaration. If any of the
functions in the block are to be recompiled, the
entire block must be (is) recompiled.
Otherwise, the block is copied from cfile as
with recompile. For pseudo-block declarations
of the form (NIL fn1 ...), all variable
assignments are made, but only those functions
so indicated by fns are recompiled.
After completing the block declarations,
brecompile processes all functions that do not
appear in a block declaration, recompiling those
dictated by fns, and copying the compiled
definitions of the remaining from cfile.
Finally, brecompile writes onto the output file
the "other expressions" collected in the initial
scan of files.
------------------------------------------------------------------------
47
In fact, recompile is defined in terms of brecompile. The only
difference is that recompile calls brecompile with an extra argument
specifying that all block declarations are to be ignored.
18.24
The value of brecompile is the output file (the
new compiled file). If the compilation is
aborted due to an error or control-D, all files
are closed and the (partially complete) output
file is deleted.
48
If cfile= NIL, file.COM is used. In addition,
if fns and cfile are both NIL, fns is set to T.
18.12 Compiler Structure
The compiler has two principal passes. The first compiles its input
49
into a macro assembly language called LAP. The second pass expands the
LAP code, producing (numerical) machine language instructions. The
output of the second pass is written on a file and/or stored in binary
program space.
Input to the compiler is usually a standard INTERLISP S-expression
function definition. However, in INTERLISP-10, machine language coding
can be included within a function by the use of one or more assemble
forms. In other words, assemble allows the user to write protions of a
function in LAP. Note that assemble is only a compiler directive; it
has no independent definition. Therefore, functions which use assemble
must be compiled in order to run.
18.13 Assemble
The format of assemble is similar to that of PROG:
(ASSEMBLE V S1 S2 . . . SN). V is a list of variables to be bound
during the first pass of the compilation, not during the running of the
object code. The assemble statements S1 ... SN are compiled
sequentially, each resulting in one or more instructions of object code.
When run, the value of the assemble "form" is the contents of AC1 at the
end of the execution of the assemble instructions. Note that assemble
may appear anywhere in an INTERLISP-10 function. For example, one may
write:
(IGREATERP (IQUOTIENT (LOC (ASSEMBLE NIL
(MOVEI 1 , -5)
(JSYS 13)))
1000)
4)
------------------------------------------------------------------------
48
See footnote on page 18.8.
49
The exact form of the macro assembly language is extremely
implementation dependent, as well as being influenced by the
architecture and instruction set for the machine that will run the
compiled program. The remainder of Section 18 discusses LAP for the
INTERLISP-10.
18.25
to test if job runtime exceeds 4 seconds.
Assemble Statements
If an assemble statement is an atom, it is treated as a label
50
identifying the location of the next statement that will be assembled.
Such labels defined in an assemble form are like prog labels in that
they may be referenced from the current and lower level nested progs or
assembles.
If an assemble statement is not an atom, car of the statement must be an
atom and one of the following: (1) a number; (2) a LAP op-def (i.e., has
a property value OPD); (3) an assembler macro (i.e., has a property
value AMAC); or (4) one of the special assemble instructions given
below, e.g., C, CQ, etc. Anything else will cause the error message
OPCODE? - ASSEMBLE.
The types of assemble statements are described here in the order of
priority used in the assemble processor; that is, if an atom has both
properties OPD and AMAC, the OPD will be used. Similarly a special
assemble instruction may be redefined via an AMAC. The following
descriptions are of the first pass processing of assemble statements.
The second pass processing is described in the section on LAP, page
18.29.
(1) numbers
If car of an assemble statement is a number, the statement is not
processed in the first pass. (See page 18.29.)
(2) LAP op-defs
The property OPD is used for two different types of op-defs: PDP-10
machine instructions, and LAP macros. If the OPD definition (i.e., the
property value) is a number, the op-def is a machine instruction. When
a machine instruction, e.g., HRRZ, appears as car of an assemble
statement, the statement is not processed during the first pass but is
passed to LAP. The forms and processing of machine instructions by LAP
are described on page 18.30.
If the OPD definition is not a number, then the op-def is a LAP macro.
When a LAP macro is encountered in an assemble statement, its arguments
are evaluated and processing of the statement with evaluated arguments
is left for the second pass and LAP. For example, LDV is a LAP macro,
and (LDV (QUOTE X) SP) in assemble code results in (LDV X N) in the LAP
code, where N is the value of SP.
------------------------------------------------------------------------
50
A label can be the last thing in an assemble form, in which case it
labels the location of the first instruction after the assemble
form.
18.26
The form and processing of LAP macros are described on page 18.31.
(3) assemble macros
If car of an assemble statement has a property AMAC, the statement is an
assemble macro call. There are two types of assemble macros: lambda and
substitution. If car of the macro definition is the atom LAMBDA, the
definition will be applied to the arguments of the call and the
resulting list of statements will be assembled. For example, repeat
could be a LAMBDA macro with two arguments, n and m, which expands into
n occurrences of m, e.g., (REPEAT 3 (CAR1)) expands to
((CAR1) (CAR1) (CAR1)). The definition (i.e., value of property AMAC)
for repeat is:
(LAMBDA (N M)
(PROG (YY)
A (COND
((ILESSP N 1)
(RETURN (CAR YY)))
(T (SETQ YY (TCONC YY M))
(SETQ N (SUB1 N))
(GO A)))))
If car of the macro definition is not the atom LAMBDA, it must be a list
of dummy symbols. The arguments of the macro call will be substituted
for corresponding appearances of the dummy symbols in cdr of the
51
definition, and the resulting list of statements will be assembled.
For example, ubox could be a substitution macro which takes one
argument, a number, and expands into instructions to compile the unboxed
value of this number and put the result on the number stack.
The definition of UBOX is:
((E)
(CQ (VAG E))
(PUSH NP , 1))
Thus (UBOX (ADD1 X)) expands to:
((CQ (VAG (ADD1 X)))
(PUSH NP , 1))
(4) special assemble statements
------------------------------------------------------------------------
51
Note that assemble macros produce a list of statements to be
assembled, whereas compiler macros produce a single expression. An
assemble macro which computes a list of statements begins with
LAMBDA and may be either spread or no-spread. The analogous compiler
macro begins with an atom, (i.e., is always no-spread) and the
LAMBDA is understood.
18.27
(CQ s1 s2 ...) CQ (compile quote) takes any number of arguments
which are assumed to be regular S-expressions
and are compiled in the normal way. E.g.
(CQ (COND ((NULL Y) (SETQ Y 1)))
(SETQ X (IPLUS Y Z)))
Note: to avoid confusion, it is best to have as much of a function as
possible compiled in the normal way, e.g., to load the value of x to
AC1, (CQ X) is preferred to (LDV (QUOTE X) SP).
(C s1 s2 ...) C (compile) takes any number of arguments which
are first evaluated, then compiled in the usual
way. Both C and CQ permit the inclusion of
regular compilation within an assemble form.
(E e1 e2 ...) E (evaluate) takes any number of arguments which
are evaluated in sequence. For example, (PSTEP)
calls a function which increments the compiler
variable SP.
(SETQ var) Compiles code to set the variable var to the
contents of AC1.
(FASTCALL fn) Compiles code to call fn. Fn must be one of the
SUBR's that expects its arguments in the
accumulators, and not on the push-down stack.
Currently, these are cons, and the boxing and
52
unboxing routines.
Example:
(CQ X)
(LDV2 (QUOTE Y) SP 2)
(FASTCALL CONS)
and cons[x,y] will be in AC1.
(* ... ) * is used to indicate a comment; the statement
is ignored.
COREVALS
There are several locations in the basic machine code of INTERLISP-10
which may be referenced from compiled code. The current value of each
53
location is stored on the property list under the property COREVAL.
------------------------------------------------------------------------
52
list may also be called with fastcall by placing its arguments on
the pushdown stack, and the number of arguments in AC1.
53
The value of corevals is a list of all atoms with COREVAL
properties.
18.28
Since these locations may change in different reassemblies of INTERLISP-
10, they are written symbolically on compiled code files, i.e., the name
of the corresponding COREVAL is written, not its value. Some of the
COREVALs used frequently in assemble are:
CONS entry to function CONS
LIST entry to function LIST
KT contains (pointer to) atom T
KNIL contains (pointer to) atom NIL
MKN routine to box an integer
MKFN routine to box floating number
IUNBOX routine to unbox an integer
FUNBOX routine to unbox floating number
The index registers used for the push-down stack pointers are also
included as COREVALS. These are not expected to change, and are not
stored symbolically on compiled code files; however, they should be
referenced symbolically in assemble code. They are:
PP parameter stack
CP control stack
NP number stack
18.14 LAP
LAP (for LISP assembly Processor) expands the output of the first pass
of compilation to produce numerical machine instructions.
LAP Statements
If a LAP statement is an atom, it is treated as a label identifying the
location of the next statement to be processed. If a LAP statement is
not an atom, car of it must be an atom and one of the following: (1) a
number; (2) a machine instruction; or (3) a LAP macro.
(1) numbers
If car of a LAP statement is a number, a location containing the number
is produced in the object code.
e.g., (ADD 1 , A (1))
.
.
.
A (1)
(4)
(9)
Statements of this type are processed like machine instructions, with
the initial number serving as a 36-bit op-code.
(2) Machine Instructions
18.29
54
If car of a LAP statement has a numeric value for the property OPD,
the statement is a machine instruction. The general form of a machine
instruction is:
(opcode ac , @ address (index))
55
Opcode is any PDP-10 instruction mnemonic or INTERLISP UUO.
Ac, the accumulator field, is optional. However, if present, it must be
followed by a comma. Ac is either a number or an atom with a COREVAL
property. The low order 4 bits of the number or COREVAL are OR'd to the
AC field of the instruction.
@ may be used anywhere in the instruction to specify indirect addressing
(bit 13 set in the instruction) e.g., (HRRZ 1 , @ ' V).
Address is the address field which may be any of the following:
= constant Reference to an unboxed constant. A location
containing the unboxed constant will be created
in a region at the end of the function, and the
address of the location containing the constant
is placed in the address field of the current
instruction. The constant may be a number e.g.,
(CAME 1 , = 3596); an atom with a property
COREVAL (in which case the constant is the value
of the property, at LOAD time); any other atom
which is treated as a label (the constant is
then the address of the labeled location) e.g.,
(MOVE 1 , = TABLE) is equivalent to
(MOVEI 1 , TABLE); or an expression whose value
is a number.
' pointer The address is a reference to a INTERLISP
pointer, e.g., a list, number, string, etc. A
location containing the pointer is assembled at
the end of the function, and the current
instruction will have the address of this
location,
e.g., (HRRZ 1 , ' "IS NOT DEFINED")
(HRRZ 1 , ' (NOT FOUND))
* Specifies the current location in the compiled
function; e.g., (JRST * 2) has the same effect
as (SKIPA).
------------------------------------------------------------------------
54
The value is an 18 bit quantity (rather than 9), since some UUO's
also use the AC field of the instruction.
55
The TENEX JSYS's are not defined, that is, one must write (JSYS 107)
instead of (KFORK).
18.30
literal atom If the atom has a property COREVAL, it is a
reference to a system location, e.g.,
(SKIPA 1 , KNIL), and the address used is the
value of the coreval. Otherwise the atom is a
label referencing a location in the LAP code,
e.g., (JRST A).
number The number is the address; e.g.,
(MOVSI 1 , 400000Q)
(HLRZ 2 , 1 (1))
list The form is evaluated, and its value is the
address.
Anything else in the address field causes an error message, e.g.,
(SKIPA 1 , KNILL) - LAPERROR. A number may follow the address field and
will be added to it, e.g., (JRST A 2).
Index is denoted by a list following the address field, i.e., the
address field must be present if an index field is to be used. The
index (car of the list) must be either a number, or an atom with a
56
property COREVAL, e.g., (HRRZ 1 , 0 (1)) or (ANDM 1 , -1 (NP)).
(3) LAP macros
If car of a LAP statement is the name of a LAP macro, i.e., has the
property OPD, the statement is a macro call. The arguments of the call
follow the macro name: e.g., (LQ2 FIE 3).
LAP macro calls comprise most of the output of the first pass of the
compiler, and may also be used in assemble. The definitions of these
macros are stored on the property list under the property OPD, and like
assembler macros, may be either lambda or substitution macros. In the
first case, the macro definition is applied to the arguments of the
57
call; in the second case, the arguments of the call are substituted
for occurrences of the dummy symbols in the definition. In both cases,
the resulting list of statements is again processed, with macro
expansion continuing till the level of machine instructions is reached.
Some examples of LAP macros are shown in Figure 18-2.
------------------------------------------------------------------------
56
If assemble code is intended to be swappable (see Section 3),
indexing should not be used in instructions that refer to assemble
labels.
57
The arguments were already evaluated in the first pass, see page
18.26.
18.31
(DEFLIST(QUOTE(
(SVN ((N P) (* STORE VARIABLE NAME)
(MOVE 1 , ' N)
(HRLM 1 , P (PP))))
(SVB ((N) (* STORE VARIABLE NAME AND VALUE)
(HRL 1 , ' N)
(PUSH PP , 1)))
(LQ ((X) (* LOAD QUOTE TO AC1)
(HRRZ 1 , ' X)))
(LQ2 ((X AC) (* LOAD QUOTE TO AC)
(HRRZ AC , ' X)))
(LDV ((A SP) (* LOAD LOCAL VARIABLE TO AC1)
(HRRZ 1 , (VREF A SP))))
(STV ((A SP) (* SET LOCAL VARIABLE FROM AC1)
(HRRM 1 , (VREF A SP))))
(LDV2 ((A SP AC) (* LOAD LOCAL VARIABLE TO AC)
(HRRZ AC , (VREF A SP))))
(LDF ((A SP) (* LOAD FREE VARIABLE TO AC1)
(HRRZ 1 , (FREF A SP))))
(STF ((A SP) (* SET FREE VARIABLE FROM AC1)
(HRRM 1 , (FREF A SP))))
(LDF2 ((A SP) (* LOAD FREE VARIABLE TO AC)
(HRRZ 2 , (FREF A SP))))
(CAR1 (NIL (* CAR OF AC1 TO AC1)
(HRRZ 1 , 0 (1))))
(CDR1 (NIL (* CDR OF AC1 TO AC1)
(HLRZ 1 , 0 (1))))
(CARQ ((V) (* CAR QUOTE)
(HRRZ 1 , @ ' V)))
(CARQ2 ((V AC) (* CAR QUOTE TO AC)
(HRRZ AC , @ ' V)))
(CAR2 ((AC) (* CAR OF AC TO AC)
(HRRZ AC , 0 (AC))))
(RPQ ((V) (* RPLACA QUOTE)
(HRRM 1 , @ ' V)
(CLL ((NAM N) (* CALL FN WITH N ARGS GIVEN)
(CCALL N , ' NAM)))
(LCLL ((NAM N) (* LINKED CALL WITH N ARGS)
(LNCALL N , (MKLCL NAM))))
(STE ((TY) (* SKIP IF TYPE EQUAL)
(PSTE1 TY)))
(STN ((TY) (* SKIP IF TYPE NOT EQUAL)
(PSTN1 TY)))
(RET (NIL (* RETURN FROM FN)
(POPJ CP ,)
(PUSHP (NIL (PUSH PP , 1)))
(PUSHQ ((X) (* PUSH QUOTE)
(PUSH PP , ' X)))
))(QUOTE OPD))
Figure 18-2
Examples of LAP Macros
18.32
18.15 Using Assemble
In order to use assemble, it is helpful to know the following things
about how compiled code is run. All variable bindings and temporary
values are stored on the parameter pushdown stack. When a compiled
function is entered, the parameter pushdown list contains, in ascending
order of address:
1. bindings of arguments to the function, where each binding
occupies one word on the stack with the variable name in the
left half and the value in the right half.
2. pointers to the most recent bindings of free variables used in
the function.
The parameter push-down list pointer, index register PP, points to the
last free variable pointer on the stack.
Temporary values, PROG and LAMBDA bindings, and the arguments to
functions about to be called, are pushed on the stack following the free
variable pointers. The compiler uses the value of the variable SP to
keep track of the number of stack positions in use beyond the last free
variable pointer, so that it knows where to find the arguments and free
variable pointers. The function PSTEP adds 1 to SP, and PSTEPN(N) adds
N to SP (N can be positive or negative).
The parameter stack should only be used for storing pointers. In
addition, anything in the left half of a word on the stack is assumed to
be a variable name (see Section 12). To store unboxed numbers, use the
number stack, NP. Numbers may be PUSH'ed and POP'ed on the number
stack.
18.6 Miscellaneous
The value of a function is always returned in AC1. Therefore, the
pseudo-function, ac, is available for obtaining the current contents of
AC1. For example (CQ (FOO (AC))) compiles a call to FOO with the
current contents of AC1 as argument, and is equivalent to:
(PUSHP)
(E (PSTEP))
(CLL (QUOTE FOO) 1)
(E (PSTEPN -1))
In using ac, be sure that it appears as the first argument to be
evaluated in the expression. For example: (CQ (IPLUS (LOC (AC)) 2))
There are several ways to reference the values of variables in assemble
code. For example:
to put value of X in AC1: (CQ X)
to put value of X in AC3: (LDV2 (QUOTE X) SP 3)
to set X to contents of AC1: (SETQ X)
18.33
to set X to contents of AC2:
(E (STORIN (LIST (QUOTE HRRM) 2 (QUOTE ,)
(LIST (VARCOMP (QUOTE X))
(QUOTE X)
SP))))
to box and unbox a number:
(CQ (LOC (AC))) box contents of AC1
(FASTCALL MKN) box contents of AC1
(FASTCALL MKFN) floating box contents of AC1
(CQ (VAG X)) unboxed value of X to AC1
(FASTCALL IUNBOX) unbox contents of AC1
(FASTCALL FUNBOX) floating unbox of AC1
To call a function directly, the arguments must be pushed on the
parameter stack, and SP must be updated, and then the function called:
e.g.,
(CQ (CAR X))
(PUSHP) (* stack first argument)
(E (PSTEP))
(PUSHQ 3.14)
(E (PSTEP)) (* stack second argument)
(CLL (QUOTE FUM) 2) (* call FUM with 2 arguments)
(E (PSTEPN -2)) (* adjust stack count)
and is equivalent to:
(CQ (FUM (CAR X) 3.14))
18.17 Compiler Printout and Error Messages
For each function compiled, whether from tcomp1, recompile, or compile,
the compiler prints:
(fn COMPILING)
(fn (arg1 ... argn) (free1 ... freen))
The first message is printed when the compilation of fn begins. The
second message is printed at the beginning of the second pass of the
compilation of fn. (arg1 ... argn) is the list of arguments to fn, and
58
(free1 ... freen) the list of free variables referenced or set in fn.
The appearance of non-variables, e.g., function names, words from a
comment, etc. in (free1 ... freen) is a good indication of parenthesis
errors.
------------------------------------------------------------------------
58
Does not include global variables, see page 18.4.
18.34
If the compilation of fn causes the generation of one or more gensym
functions (see page 18.13), compiler messages will be printed for these
functions between the first message and the second message for fn, e.g.,
(FOO COMPILING)
(FOOA0027 COMPILING)
(FOOA0027 NIL (X))
(FOO (X) NIL)
The compiler output for block compilation is similar to normal
compilation. The pass one message, i.e., (fn compiling) is printed for
each function in the block. Then a second pass message is printed for
59
the entire block. Then both messages are printed for each entry to the
block.
In addition to the above output, both recompile and brecompile print the
name of each function that is being copied from the old compiled file to
the new compiled file. The normal compiler messages are printed for
each function that is actually compiled.
Compiler Error Messages
Messages describing errors in the function being compiled are also
printed on the teletype. These messages are always preceded by *****.
Unless otherwise indicated below, the compilation will continue.
((form) - NON ATOMIC CAR OF FORM)
If user intended to treat the value of form as a function, he
should use apply*. form is compiled as if apply* had been
used. See Section 8.
(fn - NO LONGER INTERPRETED AS FUNCTIONAL ARGUMENT)
The compiler has assumed fn is the name of a function. If the
user intended to treat the value of fn as a function, he must
60
use apply*. See Section 8.
(tg - MULTIPLY DEFINED TAG)
tg is a PROG label that is defined more than once in a single
PROG. The second definition is ignored.
(tg - UNDEFINED TAG)
------------------------------------------------------------------------
59
The names of the arguments to the block are generated by suffixing
"#" and a number to the block name, e.g.,
(FOOBLOCK (FOOBLOCK#0 FOOBLOCK#1) free-variables).
60
This message is printed when fn is not defined, and is also a local
variable of the function being compiled. Note that earlier versions
of the INTERLISP compiler did treat fn as a functional argument, and
compiled code to evaluate it.
18.35
tg is a PROG label that is referenced but not defined in a
PROG.
(tg - MULTIPLY DEFINED TAG, ASSEMBLE)
tg is a label that is defined more than once in an assemble
form.
(tg - UNDEFINED TAG, ASSEMBLE)
tg is a label that is referenced but not defined in an ASSEMBLE
form.
(tg - MULTIPLY DEFINED TAG, LAP)
tg is a label that was encountered twice during the second pass
of the compilation. If this error occurs with no indication of
a multiply defined tag during pass one, the tag is in a LAP
macro.
(tg - UNDEFINED TAG, LAP)
tg is a label that is referenced during the second pass of
compilation and is not defined. LAP treats tg as though it
were a coreval, and continues the compilation.
(fn - USED AS ARG TO NUMBER FN?)
The value of a predicate, such as GREATERP or EQ, is used as an
argument to a function that expects numbers, such as IPLUS.
(x - IS GLOBAL)
x is a global variable, and is also rebound in the function
being compiled, either as an argument or as a local variable.
The error message is to alert the user to the fact that other
functions will not see this binding, since x is always accessed
directly through its value cell.
(op - OPCODE? - ASSEMBLE)
op appears as car of an assemble statement, and is illegal.
See page 18.26-28 for legal assemble statements.
(blkname - USED BLKAPPLY WHEN NOT APPLICABLE)
blkapply is used in the block blkname, but there are no
blkapplyfns or entries declared for the block.
(fn - ILLEGAL RETURN)
return encountered when not in prog.
(tg - ILLEGAL GO)
go encountered when not in a prog.
(fn NOT COMPILEABLE)
An expr definition for fn could not be found. In this case, no
code is produced for fn, and the compiler proceeds to the next
function to be compiled, if any.
fn NOT COMPILEABLE.
Same as above except generates an error, thereby aborting all
compilation. For example, this error condition occurs if fn is
one of the functions in a block.
fn NOT FOUND.
18.36
Occurs when recompile or brecompile try to copy the compiled
definition of fn from cfile, and cannot find it. See page
18.36. Generates an error.
fn NOT ON BLKFNS.
fn was specified as an entry to a block, or else was on
blkapplyfns, but did not appear on the blkfns. Generates an
error.
fn CAN'T BE BOTH AN ENTRY AND THE BLOCK NAME.
Generates an error.
(fn NOT IN FILE - USING DEFINITION IN CORE)
on calls to bcompl and brecompile.
18.37