home
***
CD-ROM
|
disk
|
FTP
|
other
***
search
/
Turnbull China Bikeride
/
Turnbull China Bikeride - Disc 2.iso
/
STUTTGART
/
LANG
/
ICON
/
ICONV8
/
Docs
/
Tr90-3
< prev
next >
Wrap
Text File
|
1990-07-19
|
13KB
|
529 lines
Personalized Interpreters for Version
8 of Icon*
Ralph E. Griswold
TR 90-3a
January 1, 1990; last revised Feburary 13, 1990
Department of Computer Science
The University of Arizona
Tucson, Arizona 85721
*This work was supported by the National Science Foundation under
Grant CCR-8713690.
Personalized Interpreters for Version
8 of Icon
1.__Introduction
Despite the fact that the Icon programming language has a
large repertoire of functions and operations for string and list
manipulation, as well as for more conventional computations
[1,2], some users need to extend that repertoire. While many
extensions can be written as procedures that build on the exist-
ing repertoire, there are some kinds of extensions for which this
approach is unacceptably inefficient, inconvenient, or simply
impractical.
Icon itself is written C and its built-in functions are writ-
ten as corresponding C functions. Thus, the natural way to extend
Icon's computational repertoire is to add new C functions to it.
The Icon system is organized so that this is comparatively
easy to do. Adding new functions does not require changes to the
Icon grammar, since all functions have a common syntactic form.
An entry must be made in an array that identifies built-in func-
tions and connects references to them to the code itself.
One method of adding new functions to Icon is to add the
corresponding C functions to the Icon system itself and to
rebuild the entire system. If the extensions are not of general
interest, it is inappropriate to include them in the public ver-
sion of Icon. On the other hand, Icon is a large and complicated
system, and having many private versions may create serious prob-
lems of maintenance and disk usage. Furthermore, rebuilding the
Icon system is time-consuming. This approach may be impractical,
for example, in a situation such as a class in which students
implement their own versions of an extension.
To remedy these problems, a mechanism for building ``personal-
ized interpreters'' is included in UNIX* implementations of Icon.
This mechanism allows a user to add C functions and to build a
corresponding interpreter quickly, easily, and without the neces-
sity to have a copy of the source code for the entire Icon sys-
tem.
To construct a personalized interpreter, the user must perform
a one-time set up that copies relevant source files to a direc-
tory specified by the user and builds the nucleus of a run-time
__________________________
*UNIX is a trademark of AT&T Bell Laboratories.
- 1 -
system. Once this is done, the user can add and modify C func-
tions and include them in the personalized run-time system with
little effort.
The modifications that can be made to Icon via a personalized
interpreter essentially are limited to the run-time system: the
addition of new functions, and modifications to existing func-
tions, operations, and support routines. There is no provision
for changing the syntax of Icon or for incorporating new opera-
tors, keywords, and control structures.
2.__Building_and_Using_a_Personalized_Interpreter
2.1__Setting_Up_a_Personalized_Interpreter_System
To set up a personalized interpreter, a new directory should
be created solely for the use of the interpreter; otherwise files
may be accidentally destroyed by the set-up process. For the
purpose of example, suppose this directory is named myicon. The
set-up consists of
mkdir myicon
cd myicon
icon_pi
Note that icon_pi must be run when in the area in which the per-
sonalized interpreter is to be built. The location of icon_pi
may vary from site to site.
The shell script icon_pi constructs four subdirectories: h,
common, std, and pi. The subdirectory h contains header files
that are needed in C routines. The subdirectory common contains
files common to several components of Icon. The subdirectory std
contains the machine-independent portions of the Icon system that
are needed to build a personalized interpreter. The subdirectory
pi contains a Makefile for building a personalized interpreter
and also is the place where source code for new C functions nor-
mally resides. Thus, work on the personalized interpreter is done
in myicon/pi.
The Makefile that is constructed by icon_pi contains two
definitions to facilitate building personalized interpreters:
OBJS a list of object modules that are to be added to or
replaced in the run-time system. OBJS initially is
empty.
LIB a list of library options that are used when the run-
time system is built. LIB initially is empty, but the
math library is loaded as a normal part of building the
run-time system.
See the listing of the generic version of this Makefile in the
- 2 -
appendix of this report.
2.2__Building_a_Personalized_Interpreter
Performing a
make pi
in myicon/pi creates two files in myicon:
picont command processor
piconx run-time system
Links to these files also are constructed in myicon/pi so that
the new personalized interpreter can be tested in the directory
in which it is made.
The user of the personalized interpreter uses picont in the
same fashion that the standard icont is used. (Note that the
accidental use of icont in place of picont may produce mysterious
results.)
The relocation bits and symbols in piconx can be removed by
make Stripx
in myicon/pi. This reduces the size of this file substantially
but may interfere with debugging.
If a make is performed in myicon/pi before any run-time files
are added or modified, the resulting personalized interpreter is
identical to the standard one. Such a make can be performed to
verify that the personalized interpreter system is performing
properly.
2.3__Version_Numbering
The Icon run-time system checks an identifying version number
to be sure the output of picont corresponds to piconx. The ver-
sion number is the string defined for IVersion in
myicon/h/version.h following the construction of a personalized
interpreter as described in Section 2.1.
In order to assure that files produced by picont can only be
run by the current versions of piconx, the value of IVersion
should be changed whenever a change is made to a personalized
interpreter. It is not important what the definition of IVersion
is, so long as it is a short string that is different from those
for other versions of Icon.
2.4__Adding_New_Functions
To add new functions to the personalized interpreter, it is
first necessary to provide the C code for them, adhering to the
- 3 -
conventions and data structures used throughout Icon [3,4]. The
directory src/iconx in the Version 8 Icon hierarchy contains the
source code for the standard functions, which can be used as
models for new ones.
Suppose that a function setenv(s) is to be added to a person-
alized interpreter.
Four things need to be done to incorporate this function in
the personalized interpreter:
1. Provide the code to the function (in, say, setenv.c).
2. Add a line consisting of
FncDef(setenv,1) # one argument
to myicon/h/fdefs.h. This adds the new function to the
Icon function repertoire.
3. Add setenv.o to the definition of OBJS in
myicon/pi/Makefile. This causes setenv.c to be compiled
and the resulting object file to be loaded with the run-
time system when a make is performed.
4. Perform a make in myicon/pi to produce new versions of
picont and piconx in myicon.
The function setenv() now can be used like any other built-in
function.
More than one function can be included in a single source
file. To incorporate these functions in a personalized inter-
preter, an FncDef entry needs to be added for each function, but
only the file need be added to the OBJS list.
2.5__Modifying_the_Existing_Run-Time_System
The use of personalized interpreters is not limited to the
addition of new functions. Any module in the standard run-time
system can be modified as well.
To modify an existing portion of the Icon run-time system,
copy the source code file from v8/src/iconx to myicon/pi.
(Source code for a few run-time routines is placed in myicon/std
when a personalized interpreter is set up. Check this directory
first and use the file there, rather than making another copy in
myicon/pi.) When a source-code file in myicon/pi has been modi-
fied, place it in the OBJS list just like a new file and perform
a make. Note that an entire module must be replaced, even if a
change is made to only one routine. Any module that is replaced
must contain all the global variables in the original module to
prevent ld(1) from also loading the original module. There is no
way to delete routines from the run-time system.
- 4 -
The directory myicon/h contains header files that are included
in various source-code files. The file myicon/h/rt.h contains
declarations and definitions that are used throughout the run-
time system. This is where information about a new type would be
placed.
Care must be taken when modifying header files so as not to
make changes that would produce inconsistencies between previ-
ously compiled components of the Icon run-time system and new
ones.
References
1. R. E. Griswold and M. T. Griswold, The Icon Programming
Language, Prentice-Hall, Inc., Englewood Cliffs, NJ, 1983.
2. R. E. Griswold, Version 8 of Icon, The Univ. of Arizona
Tech. Rep. 90-1, 1990.
3. R. E. Griswold and M. T. Griswold, The Implementation of the
Icon Programming Language, Princeton University Press, 1986.
4. R. E. Griswold, Supplementary Information for the
Implementation of Version 8 of Icon, The Univ. of Arizona
Icon Project Document IPD112, 1990.
- 5 -
Appendix - Makefile for Personalized Interpreters
Dir=`pwd`
Sdir=`pwd`/../pi
HDRS= ../h/config.h ../h/define.h
RHDRS= ../std/rproto.h ../h/rt.h $(HDRS)
ICHDRS= ../std/tproto.h ../std/general.h ../std/globals.h $(HDRS)
#
# To add or replace object files, add their names to the OBJS list below.
# For example, to add nfncs.o and iolib.o, use:
#
# OBJS=nfncs.o iolib.o # this is a sample line
#
# For each object file added to OBJS, add a dependency line to reflect files
# that are depended on. In general, new functions depend on $(RHDRS).
# For example, if nfncs.c contains new functions, use
#
# nfncs.o:$(RHDRS)
#
# Such additions to this Makefile should go at the end.
OBJS=
LIB=
ICOBJS= ../std/util.o ../std/tmain.o
RTOBJS= ../std/idata.o $(OBJS)
pi: picont piconx ../std/hdr.h
picont: ../std/icontlib.a $(ICOBJS) $(RHDRS) ../h/paths.h
rm -f ../picont picont
$(CC) $(CFLAGS) -o picont $(ICOBJS) ../std/icontlib.a
strip picont
ln picont ../picont
../std/ixhdr.o:$(ICHDRS) ../h/header.h ../h/paths.h
cd ../std; $(CC) -c $(XCFLAGS) \
-DIconx="\"$(Sdir)/piconx\"" ixhdr.c
../std/iconx.hdr:../std/ixhdr.o
$(CC) $(XLDFLAGS) ../std/ixhdr.o -o ../std/iconx.hdr
strip ../std/iconx.hdr
- 6 -
../std/hdr.h: $(ICHDRS) ../std/newhdr.c ../std/iconx.hdr
$(CC) $(XLDFLAGS) -o newhdr ../std/newhdr.c
./newhdr <../std/iconx.hdr >../std/hdr.h
rm -f newhdr iconx.hdr
piconx: ../std/rtlib.a $(RTOBJS)
rm -f ../piconx piconx
$(CC) $(LDFLAGS) -o piconx $(RTOBJS) ../std/rtlib.a \
$(LIB) -lm
ln piconx ../piconx
../std/idata.o:$(RHDRS) ../h/fdefs.h ../h/odefs.h fdefs.h
cd ../std; $(CC) -c $(CFLAGS) idata.c
../std/util.o: $(ICHDRS) ../std/link.h ../std/sizes.h ../std/trans.h \
../std/tree.h ../std/link.h ../h/fdefs.h fdefs.h
cd ../std; $(CC) -c $(CFLAGS) util.c
../std/tmain.o:$(ICHDRS) ../h/paths.h ../h/fdefs.h fdefs.h
cd ../std; $(CC) -c -DIconx="\"$(Dir)/../piconx\"" \
$(ICOBJS) $(CFLAGS) tmain.c
Stripx: piconx picont
strip piconx picont
- 7 -