[Comp.Sci.Dept, Utrecht] Note from archiver<at>cs.uu.nl: This page is part of a big collection of Usenet postings, archived here for your convenience. For matters concerning the content of this page, please contact its author(s); use the source, if all else fails. For matters concerning the archive as a whole, please refer to the archive description or contact the archiver.

Subject: Modula-3 Frequently Asked Questions (FAQ)

This article was archived around: Thu, 02 May 2002 12:51:37 GMT

All FAQs in Directory: Root Directory
All FAQs posted in: comp.lang.modula3
Source: Usenet Version


Archive-name: Modula-3-faq
Michel Dagenais Michel Dagenais, GNU General Public License, 1998-2001 michel.dagenais@polymtl.ca Ecole Polytechnique C.P. 6079, Succ. Centre-Ville Montreal, Quebec, H3C 3A7 Modula-3 Frequently asked questions and answers Maintained by Michel Dagenais ([1]michel.dagenais@polymtl.ca), suggestions are most welcome. Last updated January 15 2002. The latest copy of this FAQ may be obtained from the [2]Polytechnique Montreal Modula-3 Home. Introduction What is Modula-3? Modula-3 is a systems programming language that descends from Mesa, Modula-2, Cedar, and Modula-2+. It also resembles its cousins Object Pascal, Oberon, and Euclid. The goal of Modula-3 is to be as simple and safe as it can be while meeting the needs of modern systems programmers. Instead of exploring new features, we studied the features of the Modula family of languages that have proven themselves in practice and tried to simplify them into a harmonious language. We found that most of the successful features were aimed at one of two main goals: greater robustness, and a simpler, more systematic type system. Modula-3 retains one of Modula-2's most successful features, the provision for explicit interfaces between modules. It adds objects and classes, exception handling, garbage collection, lightweight processes (or threads), and the isolation of unsafe features. Where is Modula-3 used? Is it used in industry? A number of programming teams selected Modula-3 for industrial and research projects, and for teaching. It encourages good programming practices and comes with excellent libraries for distributed programming and graphical user interfaces. A non exhaustive list is available at [3][Modula-3 at Work]. Is commercial support available? Critical Mass Corporation used to offer their own version of SRC Modula-3, CM3, an integrated development environment for Modula-3, [4]Reactor, as well as training and consulting services. Olaf Wagner from [5]Elego Software Solutions is now maintaining [6]CM3 as an open source package and offers commercial support. Where can I get documents and information on Modula-3? A [7]concise bibliography and a more complete [8]bibliography describe Modula-3 related books, technical reports, and papers. The definition of Modula-3 is contained in: [9]"System Programming with Modula-3" also known as SPwM3. Sam Harbison has written a more tutorial book titled [10]Modula3. Three main Web servers contain Modula-3 related information: [11]www.m3.org, [12]DEC SRC Modula-3 home page, and [13]Ecole Polytechnique de Montré Modula-3 home page. The Usenet newsgroup comp.lang.modula3 is the official meeting place for Modula-3 related discussions. Is Modula-3 a superset of Modula-2? No; valid Modula-2 programs are not valid Modula-3 programs. However, there is a tool to help convert Modula-2 programs to Modula-3. Comparisons between Modula-3 and other languages? From: laszlo@post.ifi.uni-klu.ac.at (Laszlo BOESZOERMENYI) A Comparison of Modula-3 and Oberon-2 by myself in Structured Programming 1993, 14:15-22 From: nayeri@gte.com Robert Henderson, Benjamin Zorn, A Comparison of Object-Oriented Programming in Four Modern Languages, Department of Computer Science, University of Colorado, Boulder, Colorado, [14]Technical Report CU-CS-641-93. The paper evaluates Oberon, Modula-3, Sather, and Self in the context of object-oriented programming. While each of these programming languages provide support for classes with inheritance, dynamic dispatch, code reuse, and information hiding, they do so in very different ways and with varying levels of efficiency and simplicity. A single application was coded in each language and the experience gained forms the foundation on which the subjective critique is based. What implementations are available, how do they compare? All implementations are based on [15]DEC SRC Modula-3. [16]Critical Mass offered an improved version with commercial support. It features incremental garbage collection on NT, and a few additional packages like ODBC database access. This is now [17]open sourced and maintained by Olaf Wagner. Ecole Polytechnique de Montréal has been maintaining an [18]updated distribution. It features integrated documentation, and NT support through the gcc cygwin compiler Can I contribute Modula-3 software? Contributions are most welcome. The primary contact to offer contributions is comp.lang.modula3. [19]The Ecole Polytechnique de Montréal Modula-3 distribution is the most regularly updated and may be a good place to submit your contribution. Why use Modula-3? Here is what John Polstra, author of the popular CVSup, replied: Subject: Re: SUP on sup.freebsd.org Date:Wed, 06 Nov 1996 12:31:26 -0800 From:John Polstra > Erhm, why on earth did you chose Modula3 ?? Modula-3 really is a different language, designed specifically for systems programming by some extremely competent and experienced people who knew what they were doing. > Oh and yes I have seen apps written in modula3, all of which was > horrible performers, and impossible to port to new platforms, so > the management decide a complete rewrite in, guess what, C! Are you sure it was Modula-3? The SRC Modula-3 compiler supports about 25 different platforms. Plenty of real world apps (*big* ones) have been written in Modula-3, and they perform pretty well. There's also the SPIN OS project [20][SPIN] at University of Washington, in which the kernel was written in Modula-3. It performs well, too. Now, you can always argue that a program would be somewhat faster and somewhat smaller if it had been written in C. Hey, guess what? I was around when Unix V6 came out, and the same stuff was written about it. Just substitute "C" for "Modula-3" and "assembly language" for "C". The answer is the same in both cases: Unix would not exist as we know it today if it had been written in assembly language. CVSup would not exist as we know it today if it had been written in C (or C++, for that matter). OK, so why on earth did I choose Modula-3? In no particular order: 1. I needed application level threads, and threads are an integral part of the Modula-3 language. About the only reasonable alternative was to use pthreads with C or C++. But pthreads was not well supported under FreeBSD at that time. 2. I needed a graphical display during development so that I could monitor the 3 client threads as they were running, debug them, appraise their relative performance, and find the bottlenecks. Modula-3 has a very nice toolkit for creating GUIs quickly and painlessly. (OK, so the scrollbars are as ugly as sin.) 3. Modula-3 is a compiled language that is reasonably efficient. 4. I needed to use some low level system functions, e.g., mapping files into memory. Modula-3 provides good access to such functions, and it is quite easy to add interfaces to foreign libraries such as "libz". 5. Modula-3 has good support for networking. 6. It is a mature and stable language that has been used in a number of serious, large projects. The language and compiler have been stable for about 5 years, which is more than you can say for C++. 7. It has nice support for object oriented programming, including a good type system, a nice exception model, and a modern high-performance garbage collector. These traits, IMHO, contribute powerfully to producing well-structured, maintainable programs. Now before you label me an unstudly OO weenie, please consider this. I've been programming in C professionally for 19 years. I made my living for many years writing C compilers and related tools such as assemblers, linkers, and disassemblers. I still use C and C++ when I feel they are appropriate for a project, not to mention when I have to because that's what the client wants to use. I have experience programming in many many different languages. Different languages are good for different things. I still like programming in C (and C++ for some things), but I'm glad I didn't use it for CVSup. 8. I had just come off a huge 3+ year C++ project. During that time, I learned just how much C++ sucks. I did not feel like doing it again right away "for fun." 9. I have spent my entire professional career getting paid to use the wrong tools, because, e.g., the manager read that C++ was "popular." For once, just once, on a _hobby_ project, I decided I was going to use the tool I felt was the best for the job at hand. I thought about it long and hard, evaluated several options (C and C++ among them), and eventually chose Modula-3. I have never regretted that decision. Any questions? :-) John - -- John Polstra jdp@polstra.com John D. Polstra & Co., Inc. Seattle, Washington USA "Self-knowledge is always bad news." -- John Barth Troubleshooting Why program receives a SEGV signal under the debugger? The garbage collector on some platforms uses the SEGV (segmentation violation) signal to detect modified portions of the dynamically allocated space. It is possible to disable this feature or to inform the debugger to let these signals propagate. See the [21]m3gdb documentation. Problems with threads, fork and VTALARM? The threads are implemented using the virtual timer interrupt. Normally, the run time environment will catch the interrupt and determine if thread switching is appropriate. However, if a new process is created with fork, it will have the virtual timer activated and no interrupt handler to receive it, resulting in a core dump. If you use the standard procedure Process.Create to fork new processes, this will be handled automatically for you. If you insist on using fork, you need to disable the timer, fork and then reenable the timer. X libraries not found? The position of X libraries is stored, for instance for pre-compiled PM3 LINUXELF binaries, in the template file m3config/src/LINUXELF as well as in X11/LINUXELF/.M3EXPORTS (m3build/templates/LINUXELF, and X11R4/LINUXELF/.M3EXPORTS for SRC-M3). Thus you may want to edit these files if your X libraries are located in an uncommon place. What means Missing RTHooks or similar messages? The standard library, libm3, is not included by default. You need in your m3makefiles to import("libm3") or to import a library which imports libm3. Otherwise, messages about run time procedures such as RTHooks not being available are produced. M3build versus Make or why m3 does not work? The Modula-3 compiler m3 does a much finer grained dependency analysis than possible with make. For this reason, a very flexible front end, m3build, reads the program description files, m3makefile, and generates the commands required to compile and link Modula-3 programs and libraries. The m3makefile content is documented in the m3build documentation. Calling the m3 compiler directly is difficult and thus not recommended, especially on PM3 where it is now merged with m3build. Why are exceptions raised by X or Network Objects applications? Graphical applications (based on Trestle/X Windows) raise the TrestleComm.Failure exception when the DISPLAY environment variable is incorrect or the X server is refusing the connection. They raise MachineIDPosix.Failure if the network configuration files are incorrectly set up, especially on LINUX; /etc/hosts must contain at least a loopback address (127.0.0.1) and the /etc/rc scripts an appropriate ifconfig command (/etc/ifconfig lo 127.0.0.1; /etc/route add 127.0.0.1). Applications with Network Objects may also raise exceptions or consume all the CPU time available when the network configuration files are incorrect. What is the story with Trestle and OpenWindows? Mark Manasse says: I think that the OpenWindows release should be enough (no need to get the MIT X release), although there are a few things in Trestle that trigger devastating bugs in OpenWindows. But the only library we depend on is Xlib, R4 or later. The main thing I know that crashes OW 2.0 is the code where we call GrabKey specifying AnyKey. You can either loop over all of the keys, or you can just comment out the call; programs won't run exactly the same, but you probably won't notice the difference. Why so many problems installing on Solaris? These notes were contributed by (simon.hood@umist.ac.uk) while installing PM3-1.1.14 on a Sun Ultra 5 running Solaris 2.8. They describe various problems and their solution or workaround. The installation of PM3 on Solaris systems is particularly prone to problems as these systems tend to be an unpredictable (from the point of view of the PM3 people) mixture of Sun and Gnu software --- Sun do not bundle a C compiler with the operating system. My machine has gcc version 2.95.2 installed; it has Sun's versions of make, ld, as and ar installed within /usr/ccs/bin; Gnu's version of these tools are not installed. My installation was successful, after a bit of fiddling around with the configuration/template files and environment variables. Some of the fixes are trivial (if you know what you are doing), while others --- for me at least --- were not --- I am a Modula 3 novice and far from experienced with Solaris. The issues that arose are: * Paths needed to be set to find tools such as make. * LD_LIBRARY_PATH needed to be set to ensure libstdc++.a.2.10.0 and/or libstdc++.so.2.10.0 were found. * A link needed to be set so that byacc points to yacc. * The configuration for linking needed to be changed since only the Sun version of ld was installed, not Gnu's. * The build of m3gdb failed to build. * The gnuemacs package failed to build. Initial Problems Paths gcc is usually installed in /usr/local/bin; on a Solaris machine, ar, as, make and ld are all in /usr/ccs/bin, by default. Hence these must both be on root's path (assuming you are installing as root). Neither were; I have not changed any paths since installation of Solaris 2.8 on a new machine a few days ago. This is in addition to /usr/local/pm3/bin, as mentioned by the PM3 installation instructions. byacc/yacc The build required byacc. yacc is installed in /usr/ccs/bin; a soft link: lrwxrwxrwx 1 root other 4 Aug 11 15:45 byacc -> yacc solved this problem. Library Paths In addition to /usr/local/pm3/lib/m3 as mentioned by the PM3 installation instructions LD_LIBRARY_PATH must include /usr/local/lib so that libstdc++ (part of the gcc distribution) can be found. CC In addition I found that the environment variable CC needed to be set to /usr/local/bin/gcc. This is of course mentioned in the PM3 installation instructions. Miscellaneous Questions Can I get Modula-3 other than by FTP or HTTP? Prime Time Freeware (PTF) includes Modula-3. PTF is a set of two ISO-9660 CDroms filled with 3GB of freeware, issued semi-annually. PTF is distributed via bookstores and mail. You can reach PTF using: Email: ptf@cfcl.com Fax: [1] (408) 738 2050 Voice: [1] (408) 738 4832 Mail: Prime Time Freeware 415-112 N. Mary Ave., Suite 50 Sunnyvale, CA 94086 USA Many Linux CDroms include a copy of the FTP site tsx-11.mit.edu which has Linux binaries for Modula-3. How to call Modula-3 procedures from a C program? Calling Modula-3 from C is tricky because M3 has a more elaborate run-time environment. The simplest solution is to make the main program M3 and then call C via EXTERNAL routines. Calling back into M3 is then relatively straightforward. Here's an example. It calls the C code to lodge the identity of the M3 procedure to be called back which avoids having to know the actual name used by the linker. First a little M3 module to be called from C (M3code), then a C module called by the M3 main and calling the M3 module (Ccode), and finally the main program (Main): (* M3code.i3 *) INTERFACE M3code; IMPORT Ctypes; PROCEDURE put (a: Ctypes.char_star); END M3code. (* M3code.m3 *) UNSAFE MODULE M3code; IMPORT Ctypes, IO, M3toC; PROCEDURE put (a: Ctypes.char_star) = BEGIN IO.Put (M3toC.StoT (a) & "\n"); END put; BEGIN END M3code. (* Ccode.i3 *) <*EXTERNAL*> INTERFACE Ccode; IMPORT Ctypes; PROCEDURE set (p: PROCEDURE (a: Ctypes.char_star)); PROCEDURE act (a: Ctypes.char_star); END Ccode. /* Ccode.c */ typedef void (*PROC)(); static PROC action; void set (p) PROC p; { action = p; /* register the M3 procedure */ } void act (a) char *a; { action (a); /* call the M3 procedure */ }; (* Main.m3 *) UNSAFE MODULE Main; IMPORT Ccode, M3code, M3toC; BEGIN Ccode.set (M3code.put); Ccode.act (M3toC.TtoS ("Hello world")); END Main. (* m3makefile *) import(libm3) interface ("Ccode") c_source ("Ccode") module ("M3code") implementation("Main") program("mixed") Can Modula-3 code call C++ and vice-versa? There is no problem to call C++ functions declared as extern C. You must use a C++ aware linker (e.g. the C++ compiler). A complete example of M3 calling C++ objects, which in turn call M3 callbacks, is available in [22]the sgml library. On some platforms, a call to get the static variables constructors called may be required: From: gwyant@cloyd.East.Sun.COM (Geoffrey Wyant - Sun Microsystems Labs BOS) You must use your C++ compiler as the linker, rather than /bin/cc or /bin/ld. You need to call the function '_main'. The easiest way to do this is to have the following set of interfaces and implementations: INTERFACE CXXMain; <*EXTERN "_main"*> CxxMain; END CXXMain; MODULE CXXMain; BEGIN CxxMain(); END; and then import CXXMain into your M3 main module. This will ensure that the C++ function _main gets called. How to copy heap objects? Deep copies are easily performed using Pickles. An object graph is Pickled to a text writer into a TEXT. Then, a copy is created by unpickling a new object graph from a text reader created from the TEXT. Shallow copies are less often needed but may be performed with the following procedure: PROCEDURE Duplicate (r: REFANY): REFANY = VAR tc := TYPECODE (r); n_dims : INTEGER; res : REFANY; shape : RTHeapRep.ArrayShape; BEGIN IF (r = NIL) THEN RETURN NIL END; (* allocate a new object of the same type (and shape) as the old one *) RTHeapRep.UnsafeGetShape (r, n_dims, shape); IF (n_dims <= 0) THEN res := RTAllocator.NewTraced (tc); ELSE res := RTAllocator.NewTracedArray (tc, SUBARRAY(shape^, 0, n_dims)); END; (* copy the old data into the new object *) RTMisc.Copy (RTHeap.GetDataAdr (r), RTHeap.GetDataAdr (res), RTHeap.GetDataSize (r)); RETURN res; END Duplicate; How to get output messages to appear immediately (flushing writers)? Modula-3 Writers are buffered. Thus, you need to issue a Wr.Flush when the output should appear immediately, for instance to prompt the user for some input. Since this can become annoying, libraries in other languages sometimes offer the option of unbuffered writes. In Modula-3, an equivalent behavior is obtained with AutoFlushWr which gets a background thread to flush a writer at a specified interval. How to read a single character as soon as typed? Characters typed on the keyboard are usually buffered. They become visible to the reading program only when the buffer is full or after, for example, a carriage return is received. This is not specific to Modula-3. To access the characters as they are typed, single character commands in a full screen editor for example, the input reader must be configured properly. From: [23]rrw1000@cus.cam.ac.uk (Richard Watts) The POSIX way of doing it is to use tcsetattr(), and here is some code that does it under Solaris 2.x (this was written for serial ports, but the same thing applies) : PROCEDURE Open(port : CHAR; timeout : INTEGER := 30) : T RAISES {Error} = VAR term : TcPosix.termios; file : TEXT; fd : T; rc : INTEGER; BEGIN (* Figure out which device we want to open : *) CASE port OF 'A' => file := "/dev/ttya"; | 'B' => file := "/dev/ttyb"; ELSE RAISE Error("Invalid port " & Fmt.Char(port) & " specified.\n"); END; (* Open it. 700 is a good default mode for serial ports. *) fd := Unix.open(M3toC.TtoS(file), Unix.O_RDWR , 8_700); IF fd = -1 THEN RAISE Error("Open() on " & file & " failed.\n"); END; (* Get the termios structure for it *) rc := TcPosix.tcgetattr(fd, ADR(term)); IF rc # 0 THEN EVAL Unix.close(fd); RAISE Error("Couldn't get terminal attributes for " & file & ".\n"); END; (* Modify the termios structure *) (* The default baud rate is right, but we'd better set it anyway in case someone left it set up wrong : *) rc := TcPosix.cfsetospeed(ADR(term), TcPosix.B9600); IF rc # 0 THEN EVAL Unix.close(fd); RAISE Error("Couldn't set output speed for " & file & "\n"); END; rc := TcPosix.cfsetispeed(ADR(term), TcPosix.B9600); IF rc # 0 THEN EVAL Unix.close(fd); RAISE Error("Couldn't set input speed for " & file & "\n"); END; (* Modify the line discipline - reset ECHO and ICANON *) term.c_lflag := Word.And( term.c_lflag, Word.Not( Word.Or(TcPosix.ICANON, TcPosix.ECHO))); term.c_cc[TcPosix.VMIN] := 0; term.c_cc[TcPosix.VTIME] := 0; (* Set up timing right *) (* Now reset the terminal attributes *) rc := TcPosix.tcsetattr(fd, TcPosix.TCSANOW, ADR(term)); IF rc # 0 THEN EVAL Unix.close(fd); RAISE Error("Can't set attributes for " & file & "\n"); END; RETURN fd; END Open; (TcPosix.i3 is one of my interfaces, not libm3's, and I'll supply it if you like, but it's just a wrapper to tcgetattr and friends. The baud rate stuff shouldn't be necessary for terminals (or serial ports..) ). You should be able to somehow get an Rd.T out of this, I think, but it may involve a bit of hacking. The University of Cambridge can't have these opinions even if it wants them. Why is Hello World larger in Modula-3 than in C? Modula-3 programs are slightly larger than C programs because the generated code includes runtime type information, and runtime checks for out-of-bound array references and NIL pointers. Many of these checks could be removed by a more sophisticated compiler. The fixed runtime is substantially larger (there is no runtime support in C). It contains a garbage collector, a thread runtime, and exception support. It is typically placed in a dynamically linked library, shared on disk and in memory between all the Modula-3 programs. What is SRC Modula-3? [24]SRC-Modula-3 was built by the DEC Systems Research Center and is freely available and redistributable, with source code. In Europe it is also available from ftp-i3.informatik.rwth-aachen.de in pub/Modula-3. The most recent version is release 3.6 The DEC SRC Modula-3 contains the following: * A native code compiler: uses the GCC backend; on machines/operating systems that have self-describing stacks, an optimized exception handling mechanism is provided, on other architectures, setjmp/longjmp is used. A very fast integrated backend is available on some platforms (currently NT386 and Linux i386). The compilation system provides for minimal recompilation. Only those units that depend on the modified interface item will be recompiled. * m3build: tool that performs dependency analysis and builds the Modula-3 programs and libraries. * m3gdb: a Modula-3 aware version of GDB. * Several tools for performance and coverage analysis. * A large standard library (libm3) providing + A multithread, incremental, generational, conservative garbage collector + Text manipulation. + Generic Containers: Lists, Sequences, Tables, SortedLists, SortedTables + Atoms and Symbolic expressions (Lisp like lists) + An extensible stream IO system + Typesafe binary object transcription (persistent objects) + Operating system interfaces + Portable interfaces to the language runtime All standard libraries are thread-friendly. Modula-3 can readily link with existing C libraries; many libraries including X11R4 and various UNIX libraries are available as part of libm3. * Several other libraries for designing graphical user interfaces and distributed applications. Why are there strange pragmas for Locking levels and other properties? The Trestle (ui library) interfaces contain Locking level pragmas. The base interfaces (libm3 library) contain SPEC pragmas. These are not processed by the compiler. Instead the Extended Static Checker, currently under development at DEC SRC, will report on problems detected based on the program content and the information specified in these pragmas [25][ESC]. The Extended Static Checker is not yet available, it may be some time in the future. Design Issues Why objects and interfaces? Allan Heydon on comp.lang.modula3, May 4th 1993: Modula-3 provides two separate mechanisms for data-hiding: one for hiding details about how interfaces are implemented, and the other for hiding details about how objects are implemented. The first data-hiding mechanism is realized by the distinction between interfaces and modules. Clients can only import interfaces, so the names declared in the modules implementing those interfaces are hidden from clients. Note that this mechanism has only two levels; a name is either declared in an interface, or it isn't. If a name is only declared in a module, it can't be used by a client. The second data-hiding mechanism is realized by opaque types and revelations. A Modula-3 interface may declare an object type to be opaque, in which case only a subset of the fields and methods of that object are revealed to clients importing the interface. Furthermore, the Modula-3 revelation mechanism allows a designer to reveal successively more fields and methods of an object in a series of interfaces. The fields and methods visible to a client then depends on which interfaces the client imports. The latter mechanism is quite flexible. As opposed to the interface/module data-hiding mechanism, opaque types allow you to define an arbitrary number of levels at which more and more information about the implementation of your object is revealed. See Sections 2.2.10, 2.4.6, and 2.4.7 of "Systems Programming with Modula-3" for more information about opaque types and about partial and complete revelations. What is the purpose of BRANDED and REVEAL? Allan Heydon writes: These two keywords are necessary because of two quite different features of the language. REVEAL is necessary because Modula-3 has opaque types and partial revelations. BRANDED is necessary because the Modula-3 type system uses structural equivalence instead of name equivalence. In Modula-3, the concrete structure of a type can be hidden from clients in an interface. A common idiom is: INTERFACE I; TYPE T <: TPublic; TPublic = OBJECT (* fields *) METHODS (* methods *) END; END I. The line "T <: TPublic" introduces the type "I.T" as an opaque subtype of the type "I.TPublic". It does not reveal any of the other details of the concrete structure of "I.T" to clients. Hence, "I.T" is said to be an opaque type. Put another way, the structure of "I.T" is only partially revealed to clients. In addition, it is possible to reveal more of "I.T"'s structure in other interfaces, like this: INTERFACE IRep; IMPORT I; TYPE TPrivate = I.TPublic OBJECT (* more fields *) METHODS (* more methods *) END; REVEAL I.T <: TPrivate; END IRep. This interface declares a type "IRep.TPrivate" that is a subtype of "I.TPublic". It also asserts that "I.T" is also a subtype of "IRep.TPrivate". A client that imports only the interface "I" has access only to the fields and methods in "I.TPublic" when accessing an object of type "I.T", but a client that imports both "I" and "IRep" also has access to the fields and methods in "IRep.TPrivate" when accessing an object of type "I.T". The "REVEAL" statement in this module simply asserts a subtype relation. Unlike type declarations, revelations introduce no new names. Hence, we could not have used the "TYPE" keyword in this case because the type "I.T" has already been declared once (albeit opaquely) in interface "I". Every opaque type must have a complete revelation. A complete revelation has the form: REVEAL T = TConcrete; The revelation specifies that "TConcrete" is the concrete type for the opaque type "T". The Modula-3 type system uses structural equivalence instead of name equivalence. This means that two types are equal iff they have the same structure. One consequence of this rule is that two types you might intend to be distinct may actually be equal. This can have unintended effects on the run-time behavior of your program. For example, if both types that you expect to be distinct are actually structurally equivalent and the two types guard two arms of a TYPECASE statement, the arm for the second type will never be taken. If you want to avoid accidental equalities between two types, you can brand one (or both) of them with the BRANDED keyword. A branded type is equivalent to no other type, even if it is structurally equivalent to some other type. In essence, the BRANDED keyword adds a bit of virtual structure to the type that guarantees it will be distinct from every other type. The Modula-3 syntax allows you to supply a text constant as a name for the brand. If you don't supply an explicit brand, the compiler will make one up; however, the implicit brand invented by the compiler is not guaranteed to be chosen deterministically. Hence, explicit brands are useful if you are communicating types from one process to another and if you want to be sure that the branded type written by one process matches the branded type read in by the other. Any two opaque types in a program must be distinct. Otherwise, it would be too easy for clients to accidentally trip over type collisions like the TYPECASE example mentioned above. To enforce the restriction that all opaque types are distinct, the language requires that the type "TConcrete" in the complete revelation above must be a branded type. Can a program recover from running out of virtual memory? No, this turns out to be quite a thorny problem. I think the best thing I can do is by attaching to this message the dialog that went on during the "beta test" of the new library interfaces (SRC Research Report 113, "Some Useful Modula-3 Interfaces). The parties are Xerox PARC's David Goldberg, Hans Boehm, Alan Demers, and David Nichols, and SRC's John DeTreville, who designed and implemented the garbage collector in SRC Modula-3. The dialog covers many of the issues, and apparently ends when the participants run out of steam. Paul McJones mcjones@src.dec.com (editor of SRC 113) RTAllocator should allow handling out of memory David Goldberg: ... there is one system problem that is not currently handled, namely running out of memory. I would very, very much like to see this handled in RTAllocator. One approach was suggested by Roy Levin a while back: Have a RegisterNoMemory(proc) routine that causes proc() to be called when memory is gone (or very low). Example of use: in the 'Solitaire' program, the 'Hint' button generates a tree of possible moves. If this tree gets very big and consumes all memory, the RegisterNoMemory proc could abandon the search down the current branch, NIL-out that branch, and ask for a garbage collection. Currently what happens is that Solitaire crashes if you bug 'Hint' and memory is low. Interface Police: Ok, make a concrete proposal and we'll talk. How low should memory be before the runtime complains? Before or after a collection? Is it ok to call your procedure from inside a runtime critical section (after all, you're probably in the middle of a NEW)? Are multiple notification procedures allowed to be registered? Shouldn't a routine that consumes arbitrary amounts of memory be coded to poll the allocator to ask how much memory is available? Hans Boehm/Alan Demers/David Goldberg/David Nichols: We believe that programs wishing to handle allocation failures will be able to do so with high (but not perfect) reliability if the interface provides two features: versions of the RTAllocator.New routines that report if the allocation is not possible (either by returning NIL or raising an exception), and a way to register a callback when memory is low. Both features are necessary. Here are two typical scenarios: * The Solitaire program. Before starting, Solitaire allocates a 'safety net' block of memory, and registers a callback. When memory is exhausted, the callback frees the safety net, sets a flag, and returns. In the Solitaire program proper, the inner loop of the move generator checks the flag immediately after allocating a new node. If the flag is set, it abandons the search. It would not work for Solitiare to allocate new tree nodes with RTAllocator.New() and check for an error: as memory gets low, a library routine in some other package could cause an allocation failure. Unfortunately, there is a race condition since another thread could run and do an allocation between the time the faulting NEW returns and all references to the search tree are NIL'ed. This can be mimimized by adding some slop to the safety net. * An editor that allocates O(n) bytes of memory when opening an n-byte file. If the users tries to open a huge file, you don't want to crash, but rather tell the user that the file can't be opened (in UNIX, the user can then kill some processes to regain swap space and try again, or in an emacs-style editor he can delete some buffers and try again). A callback won't work for this, because when attempting to open a huge file, the allocation must be aborted: there just isn't enough memory to go around. Instead an RTAllocator.New() routine should be used for this allocation. However, the editor will also want to register a callback proc to guard against NEW()s in other parts of the program that can't be satisfied. If the callback is passed the size of the memory allocation that can't be satisfied, the callback will be able to pick between two strategies. If there is a 'safety net' which is larger than the block to be allocated, the callback can free it and set a "low on memory" flag, with the editor cleaning up properly later. If the safety net is not big enough, the callback itself can attempt an emergency save before crashing. Here's a specific proposal that embodies these ideas. We're not wedded to the details. Note that RTCollector.LimitRelative is not essential: it just lifts some of the functionality currently in RTHeapPolicy. * Add the following to RTCollector.i3: PROCEDURE LimitAbsolute(n: CARDINAL); (* Don't let the heap grow beyond n bytes. The collector/allocator should observe this in all heap growth decisions. *) [Comment from Hans: I don't think there is a way to write programs that are reasonable citizens on a shared system without some such facility.] PROCEDURE LimitRelative(x: REAL); (* Advisory. Try to keep the heap size at roughly x times the amount of live data. (For ref counting it affects only the backup collector for cycles.) *) [Comments from Hans: The performance of all collectors with which I am familiar depends crucially on such a parameter. Thus it might as well be exposed in some portable interface. (The allocator should of course use less memory if it gains no time advantage from using more.) The "amount of live data" is, of course, implementation defined, as are the minimum values of x that have any chance of being observed.] * In RTAllocator.i3, add OutOfMemory to RAISES clauses of all the New routines, and add the following: EXCEPTION OutOfMemory; TYPE CallBackObj = OBJECT notify(bytes: CARDINAL) END; PROCEDURE RegisterHeapFullCallback(obj: CallBackObj); (* Add obj.notify to the list of procs to be called if an allocation request is about to fail, either because of lack of memory, or due to violation of an RTCollector.LimitAbsolute imposed limit. The notify method will be called with an argument specifying the size in bytes of the allocate call that triggered the callback. The notify method may not allocate or acquire locks, or call any procedures that do. It may be invoked synchronously with any combination of locks held. (Should there be a way to delete a registered callback?). If a garbage collection after this callback fails to reclaim enough memory to allocate the requested object, an exception will be raised if the allocation was through RTAllocator. Otherwise a checked runtime error will result. The notify proc is not called when memory fails from an RTAllocator.New call (these failures can be caught by the user). Typical actions by notify would include one of the following: 1) Clearing pointers to reduce the amount of accessible memory. 2) Calling RTCollector.LimitAbsolute with a larger limit. *) * Variations on this proposal: Might want to consider adding: PROCEDURE GetLimitAbsolute(): CARDINAL; (* Return the current absolute heap limit *) The usefulness of RTCollector.LimitAbsolute in the callback would be increased if there was a way to tell if this actually freed up any more memory. One approach would be to change CallBackObj to TYPE CallBackObj = OBJECT notify(bytes: CARDINAL; retry: BOOLEAN): BOOLEAN END; and change the action of RegisterHeapFullCallback to: (* If a garbage collection after all callbacks have been executed fails to reclaim enough memory to allocate the requested object, then any notify() procs that returned TRUE will be called again with retry := TRUE. Otherwise an exception will be raised if the allocation was through RTAllocator, or else a checked runtime error will result. *) Thus, if you wanted to first try and get more memory in the callback by calling RTCollector.LimitAbsolute, you could return TRUE and wait for a callback with retry = TRUE. If this second callback occurs, you will need to clear some pointers to free up memory. Or another variation: add PROCEDURE GetTotalBytesAllocated(): CARDINAL; (* Returns the total number of bytes allocated since the program begin. A CARDINAL may not be big enough, perhaps this should be a LONGREAL? *) Then the retry argument to the notify method can be eliminated, since a call is a retry only if GetTotalBytesAllocated() shows no additional allocations since the last callback. John DeTreville: When I read your March proposal for handling running out of memory on the traced heap, I didn't quite see how to implement the details you gave. I've been iterating to create mechanisms that are simpler and more implementable, and I've now arrived at quite a simple interface. In particular, I now believe that (almost) all the functionality you ask for is already provided by the current interface. I say "almost" because there's a few status calls to be added, and because some of the current mechanisms are clunky, but I believe I can tell a convincing story. Note that these mechanisms are or would be in SRC-specific interfaces (currently called RTAllocatorSRC, RTCollectorSRC, and RTHeapRep); I don't think we understand them well enough to put them into the public IP interfaces. Let's first distinguish VM limits from application-imposed limits. The amount of VM available to the application is a hard limit, although not one that can easily be predicted. In the current SRC M3 implementation, both the allocator and the collector allocate VM from the kernel when necessary. If the collector tries to allocate VM and fails, the program must crash: there is no way to reestablish the necessary invariants to let it continue. I propose treating VM exhaustion as a checked runtime error, in the allocator and in the collector. The goal is then to establish and maintain an application-imposed limit that is uniformly stricter than the VM limit, whatever that may be. You propose a mechanism to allow calls to the New* procedures to fail if they would exceed the application-imposed limit. Of course, only a small part of the code would take advantage of this facility. This code could equally well query the heap to determine the current size, and compare it against the limit; if the program can also predict the size of the object to be allocated, it can decide whether or not to proceed. This approach requires some collector-dependent code in the application, but I doubt that it would be very much. It also allows possible race conditions, but I believe they're not much worse in practice than in the original proposal. You also propose a mechanism to notify the program whenever the limit is about to be exceeded. It's quite complicated to get such immediate notification. First, the procedures notified can't acquire locks or call most procedures in other modules. Second, it requires a new collection to run synchronously after the procedures to see if enough space has been freed and whether some of the procedures must be called again; this causes an interruption of service. Here's a different proposal, which might not allow space bounds as tight as in the original proposal, but which seems simpler. We would add a mechanism for an application thread to wait for the next collection to finish. This mechanism could replace the current mechanisms for registering and unregistering synchronous monitors, which have numerous complex and poorly documented constraints on what actions they can perform. Each time through, the thread could compare the amount of space still reachable to the application-imposed limit, and either free some data before the next collection (the ability to hold locks would be handy here) or increase "gcRatio" to make the collector work harder and keep the total heap size under control, or both. There is still the danger that the application could allocate so rapidly that this asynchronous thread might not be able to keep up, but otherwise asynchronous actions seem a lot more reasonable than synchronous. This is one approach, and there are others. What's nice about this design is that it requires almost no changes to the interface, only better status reporting and a replacement of the mechanisms for registering and unregistering synchronous collection monitors. Maybe you could even work around the current lack of these facilities. Let me know what you think. Hans: One quick comment, without having thought much about the rest: "This code could equally well query the heap to determine the current size, and compare it against the limit; if the program can also predict the size of the object to be allocated, it can decide whether or not to proceed." Is this really true? Since the collector can't move some objects, there are presumably fragmentation issues. Am I guaranteed to be able to allocate 1 MB if the current heap size is 1 MB below the limit? This is certainly false in PCR, and I'm not sure how you could guarantee it without remapping pages. John: Hans Boehm notes that I was wrong about the client of New* being able to predict whether an allocation would succeed or fail, because of likely page-level fragmentation. This needs to be fixed in my proposal. To expand on my earlier message, let me outline a completely different approach for handling heap overflow, that perhaps has more in common with the original PARC proposal, but which seems far too complex and unwieldy to me. This complexity is why I tried to work out a simpler approach, even at the cost of providing fewer guarantees. We start by imagining that we want to be able to continue to run even if we exhaust VM. First, this means that we can never allocate VM from inside the collector. The implication is that whenever we allocate VM in the allocator, we allocate enough extra to tide us over through the next collection, no matter how much of the heap it retains. This suggests that we will significantly overallocate VM. For example, with a stop-and-copy collector and gcRatio set to 1, a program with a stably-sized reachable set currently requires 3 times as much space as the reachable set, but the "failsafe gc" would require 4 times. (Doing even this well depends crucially upon the SRC implementation detail that the current collector never copies objects bigger than one page, but leaves them in place. Otherwise, the possibility of fragmentation would make it much more difficult to determine how much memory to leave free for the collector, and in what sizes of runs of pages. It will also take some work to avoid off-by-one errors in predicting how much memory a collection could take.) Of course, if the client decreases gcRatio, or switches from sort-and-copy collection to concurrent collection, that would require allocating more VM, to ensure that the collector cannot run out of VM. That means that these operations can also fail, just like allocator operations. Only some programs will want to be able to back off when they reach VM limits. Others won't mind getting a checked runtime error; in return, they will require less VM. Therefore, we need procedures to switch back and forth between these two modes. Again, attempting to switch to failsafe mode can fail. The collector currently allocates its own space on the traced heap during collections, which will have to be moved to the untraced heap if we are to predict how much traced heap a collection can use. Note that in general, once VM is exhausted, allocations on the untraced heap may start to fail, and so programs will probably die very quickly once VM is exhausted. But let's move on. In addition to the VM limit, we also want an application-imposed limit on heap size. The allocator and collector will guarantee that the heap size will never exceed this limit. Again, we will overallocate VM in the allocator to avoid exceeding the limit in the collector. Again, setting the limit may fail. So what happens when a NEW fails, or a New*, or switching to concurrent collection, or setting the application-imposed limit, or whatever? This happens whenever performing the operation would exceed the application-imposed limit, or when attempts to allocate enough extra VM fail. Some of these can signal an error, and the client can chose to do something else instead. In some cases, such as setting gcRatio, it might make sense for the failing operation to tell the client how close to the impossible value would be possible. NEW, though, should not signal an error; this would require massive changes in all existing modules that would not be add value to most clients. In this case, I can't think of anything much better than the original proposal. Before attempting to allocate the object, the collector will try to free up some storage. First, it can perform a collection, or finish the current one. If that doesn't do it, it can call one or more procedures registered by the application to drop links to some storage, or to change collector parameters. If that doesn't do it, we can perform another collection. And so on and so on, until the procedures say to give up. Note that these collections must be synchronous, since no allocations may be possible until this mechanism completes, and the collections will therefore cause interruptions of service. Note also that the procedures cannot acquire locks, cannot allocate storage, cannot call thread primitives, and so on, and therefore cannot call into most libraries; they are essentially restricted to reading and writing data structures whose implementations they know, and changing collector parameters. This seems excessively restrictive, but also unavoidable in this approach. In short, this seems like a lot of extra mechanism to add to the allocator and collector, that doesn't seem to do quite what you want; it gives you strict limits, but at a cost. My proposal of this morning is at least much simpler, although it can give looser limits. John, continuing: Thinking a little more about the problem of running out of storage in the untraced heap, it seems that the only reasonable thing to do is to merge the implementation of the untraced heap with the traced heap. This was, untraced NEWs that fail can be handled exactly the same way as traced NEWs, with a synchronous cleanup routine that frees enough VM to proceed, or resets parameters. This means that the allocator and collector cannot use the untraced heap, but must either use a static amount of storage which they could overflow, or must allocate enough extra in response to client applications that they cannot possibly run out of space. The potential space overhead for maximum-sized stacks, for example, is huge. The more this proposal is fleshed out, it more it seems that doing a good job of recovering from heap overflows is quite tricky, which is why I suggest a lower-tech approach for now. Hans: I just went over the last few messages in this thread again. I think the bottom line is, as you say: It's hard to implement an out-of-memory call-back on top of the current collector. Given the current collector, a collector call-back that allows polling is probably the best you can reasonably do, and should certainly be provided. The remaining question, which also seems to be motivated by other concerns here, is: To what extent are you tied to this collector design? The problem here seems to be mainly caused by copying old generation objects, since you could perhaps bound the size of the young generation? My suspicion, based unfortunately only on anecdotes, is that this is not a good idea anyway, since it uses too much space, and is also fairly expensive in copying cost. (PARCPlace seems to have arrived at the same conclusion, so there's probably at least one other supporter of this position. Some recent complaints here about space usage of Modula-3 programs also point a bit in this direction.) Do you agree? If so, should the interface be designed ignoring current constraints, and should we initially accept a partial implementation? John: It's been a while, so let me recap where we are, or at least where I am. We've been discussing mechanisms for Modula-3 programs to manage their memory better. In particular, we have proposed ways that programs could bound the heap size and recover from heap overflow. I think this topic is complicated enough, and new enough, that we shouldn't try to get it into the current set of portable interfaces. The Interface Police concur. We've floated two broad (families of) proposals for attacking this problem: * Allow strong guarantees on the heap size; these guarantees would never be broken. * Allow the program to monitor its memory usage, discarding excess data as necessary. I think that the first is achievable. Adapting the current SRC Modula-3 (allocator and) collector to allow such guarantees would take a month or so. The principal problem is that the current collector tries to maximize space/time performance, and giving such guarantees will probably require extra memory to be set aside that will never be used. The collector would have two modes: with or without guarantees. Most programs would run without guarantees. I also think that a usable version of the second is possible with almost no change to the current collector. The programmer would have to do more work, and wouldn't get any strong guarantees, but this approach should work for many programs. We've also been discussing a third family of proposals, that seem to combine the worst features of the first and second: they require significant changes to the current collector, but son't give very strong guarantees. These seem much less interesting to me. Here's two pieces of opinion. First, I propose that we work out the details of #2, and you use it for a couple of programs. Get some experience with it. This could help inform a heavier-weight solution. Second, I wonder whether any of these solutions is a good candidate for a portable interface. It's one thing not to be strictly incompatible with a given collector strategy, but quite another to be easy to plug into an existing collector. Modula-3 currently doesn't require very much from its collector; making these proposals standard would significantly increase the requirements on a Modula-3 implementor. Why uppercase keywords? Some people prefer uppercase keywords others hate them. Another possibility is to accept both forms for keywords. This topic has been discussed at length and there is no solution that will completely satisfy everyone's tastes. Fortunately this is a very minor issue and you can easily have lowercase keywords automatically converted for you using an emacs macro package like [26]m3su . Why CONST Comments in Variables Declarations? John Kominek (kominek@links.uwaterloo.ca) wrote: Sprinkled throughout SRC m3 you'll find "constant" variables exported in interfaces. For instance, VAR (*CONST*) Grain: LONGREAL; where Grain is assigned during module initialization. Instead, did the modula-3 designers consider doing this. READONLY Grain: LONGREAL; Here the keyword permits only exporting modules to modify the Grain variable. Is there a problem with this proposal? The READONLY keyword is successfully used at procedure boundaries, so why not also at interface boundaries? Bill Kalsow replies: A problem with this proposal is that any module can claim to export the interface containing the variable, hence any module could modify the variable. Note that CONST says more than just READONLY. CONST implies that the variable should not be modified by clients and that once it is initialized, it won't be changed later by the implementation. READONLY would only mean that clients should not modify the variable. IMO, the "right" solution would have been to allow: INTERFACE Foo; CONST x: T; MODULE Foo; CONST x: T := <value>; In the same way it checks revelations for opaque types, the compiler could check that only one module provided a value for the constant. But, this proposal doesn't quite hang together either. Consider this example: CONST x: INTEGER; VAR v: [0..x]; The language definition says that "v"s definition is illegal if "x < 0" because its type is "empty". The system could refuse to run the program by delaying the check until it had seem the corresponding implementation module. But, I think you'll agree that it could quickly turn into a mess. The most flexible handling of opacity I've seen is in Christian Collberg's PhD Thesis, "Flexible Encapsulation". It was published Dec 5, 1992 by the CS Dept at Lund University, Lund Sweden. If I remember correctly, his system was capable of deferring all checks and decisions imposed by opaque declarations until link time. References 1. mailto:michel.dagenais@polymt.ca 2. http://m3.polymtl.ca/m3 3. http://www.cmass.com/cm3/projects.html 4. http://www.cmass.com/ 5. http://www.elego-software-solutions.com/ 6. http://www.m3.org/cm3/ 7. file://localhost/home/m3/tmp/m3/pm3/intro/src/concise-bib.html 8. file://localhost/home/m3/tmp/m3/pm3/intro/src/bib.html 9. file://localhost/home/m3/tmp/m3/pm3/intro/src/bib.html#SPwM3 10. file://localhost/home/m3/tmp/m3/pm3/intro/src/bib.html#m3-Har92 11. http://www.m3.org/ 12. http://www.research.digital.com/SRC/modula-3/html/home.html 13. http://m3.polymtl.ca/m3 14. ftp://ftp.cs.colorado.edu/pub/cs/techreports/zorn/CU-CS-641-93.ps.Z 15. http://www.research.digital.com/SRC/modula-3/html/home.html 16. http://www.cmass.com/ 17. http://www.m3.org/cm3/ 18. http://m3.polymtl.ca/m3 19. http://m3.polymtl.ca/ 20. http://www.cs.washington.edu/research/projects/spin/www/ 21. file://localhost/home/m3/tmp/m3/pm3/language/modula3/m3tools/m3gdb/src 22. file://localhost/home/m3/tmp/m3/pm3/text/sgmltools/sgml/src/nsgmls 23. mailto:rrw1000@cus.cam.ac.uk 24. file://gatekeeper.dec.com/pub/DEC/Modula-3/ 25. http://gatekeeper.dec.com/pub/misc/detlefs/escover.ps 26. ftp://pion.lcs.mit.edu/pub/m3su -- Prof. Michel Dagenais http://m3.polymtl.ca/dagenais Département de génie informatique michel.dagenais@polymtl.ca Ecole Polytechnique de Montréal tel: (514) 340-4711 ext.4029