Skip to content
MELT monitor
Branch: master
Clone or download
Fetching latest commit…
Cannot retrieve the latest commit at this time.
Type Name Latest commit message Commit time
Failed to load latest commit information.

MELT monitor

This MELT monitor is an alpha stage experimental GPLv3 software (hosted on github) by Basile Starynkevitch, from France.

It is conceptually related to GCC MELT, a Lispy like domain specific language to customize the GCC compiler, implemented mostly by a GPLv3 meta-plugin (the MELT plugin for GCC) that I also developed and that is FSF copyrighted, like GCC is. Read GCC MELT sheet (two pages) for more (and, this French page (single page, in French)....)

The MELT system

The MELT system (still partially a dream) refers to a mix of this MELT monitor and GCC MELT customized compilations (including some future MELT extensions for GCC not yet developed; these extensions will communicate with this MELT monitor using some textual ad-hoc protocol on Unix sockets). The long-time goal of the MELT system is to provide a software development environment and static analysis tool for code compiled by GCC (putative free-software equivalent of non-sound static code analysis tools like Coverity, ParaSoft, etc...) and be able to help exploring and understanding and navigating source code of many software components at once (in our wildest dreams, all the free software packaged in a Debian distribution and compiled with GCC, and some large proprietary software provided by funding partners).

In my vision, a user community (e.g. a team of dozens of developers working on some mid-sized software project compiled by GCC...) would compile its software collection with a MELT enhanced GCC (this should be as simple as adding CC=gcc and CFLAGS += -fplugin=melt -fplugin-melt-arg-mode=super in their Makefile or other build machinery; of course that might slow down a bit the compilation but won't change the habits of the developers...) and use the MELT system (notably the web interface provided by the MELT monitor) to query, understand, analyze, navigate, and even discuss about the development of that software.

You can use the gcc-melt GoogleGroup to discuss technical matters related to the MELT system (so both MELT plugin and MELT monitor)

Goals of the MELT monitor

The goals of the MELT monitor includes notably

  • provide a web interface (HTML5 compliant, for recent Firefox) to the MELT system (it took years for me to admit that people want a web interface, or at least a GUI one, and that most software developers -notably those working in industry- don't use the command line anymore, like I am used to do....). And with a Web interface, even non-Linux users could use it (although the MELT system targets only Linux)... That interface would also be useful to browse (but not edit!) a large collection of user source code (compiled by a MELT enhanced gcc), and most of GCC internal representations (in its middle-end) of that user code (perhaps even explaining some of the GCC compiler's decisions, e.g. why that function was inlined here and not there...).

  • Provide a web interface to edit some abstract syntax tree (ASTs) of some future DSL (e.g. formalizing coding rules), to search into MELT system & GCC compiled software, and give him the ability to document things in a Wiki way. We don't want the user to type some DSL syntax (like the Lispy MELT DSL inside our MELT plugin) as text, we want to assist him in constructing ASTs; Old 1980s MENTOR system from INRIA (or some Smalltalk systems) is inspirational. Notice that the MELT monitor is a specialized Web server with very few simultaneous users (usually one, in the future a small team working on the same software project, e.g. a dozen of users).

  • provide a persistent machinery to the MELT system, that is a long-term store usable by MELT extensions to keep, enhance, and provide static analysis information. In other words, we need a cluster of MELT monitors which checkpoints periodically its state and which we can later restart smoothly from that persistent state: conceptually the MELT system should run during an entire software project using it (like your version control system does), because it is part of the entreprise know-how. Hence, dynamic software updating of the MELT system is a must.

  • provide a high-level declarative language framework for static analysis, above the MELT plugin, and for the MELT system ; we want some high-level rule based language (e.g. to express coding rules) ...

  • be able to take advantage of multi-core systems by using several threads and work in parallel

  • if possible, take advantage and be able to run on a small cluster or cloud (of e.g. a few dozens of Linux servers), so MELT monitors should be able to run concurrently and communicate ...

Not all of them are implemented in 2015. My roadmap includes, like I did in the MELT plugin, development of bootstrapped specific languages with a metaprogramming (and metaknowledge, inspired by J.Pitrat's work and vision approach). The MELT monitor will have some DSL and more and more of its code will become self-generated.

Current state

(I am advancing very slowly; things are difficult; this needs a Linux system to be run and used)

External dependencies

The following software components are needed (or are very likely to be needed later, I am listing -for reference- also some packages that are very likely to be useful in the future and that I have looked into). When packaged in Debian(/Sid), I just name the Debian package. See the Makefile

  • a recent GCC compiler (so gcc-5 Debian package), in particular because even when running both MELT plugin and MELT monitor are generating code.

  • the GCCJIT library, so libgccjit-5-dev Debian package

  • as for the MELT plugin, all the source dependencies of the latest GCC (so aptitude build-dep gcc-5)

  • the MELT plugin itself

  • Boehm's garbage collector thru libgc-dev Debian package; I am aware of GC techniques, but coding a multi-threaded GC is too much work (I did spend months debugging the single-thread GC in the MELT plugin, and before that I coded Qish and some proprietary GC inside Fluctuat so I consider myself experimented in garbage collection techniques.). Ravenbrook Memory Pool System might be considered as an alternative, at least after bootstrapping is completed.

  • the pkg-config utility (and pkg-config Debian package)

  • The Glib from GTK is a little bit used. so libglib2.0-dev Debian package.

  • sqlite might be used (for persistency), so libsqlite3-dev Debian package.

  • codemirror is embedded under webroot/, to nicely show some C code in a web page.

  • redis might be used (for persistency), so redis-server, redis-tools, libhiredis-dev Debian packages

  • PostGreSQL might be used (for persistency), so probably libpq-dev

  • 0mq might and probably will be used for messaging purposes, so probably libzmq5-dev or libzmq3-dev or libzmq-dev Debian package.

  • jansson library for JSON is needed (notably for AJAX), so libjansson-dev Debian package.

  • libonion is an HTTP server library that is absolutely required, but it is not packaged in Debian. So compile its latest source from David Moreno onion github; I did contribute some small patches to it (sometimes, when I need a patch not yet accepted in libonion by David Moreno, you'll need to compile my github fork of onion...). I need at least the version 0.8 (may 2016) of libonion

  • I am using some HTML5 things (e.g. JQuery, etc...) but I am copying them under my webroot/ subdirectory, which also contains some of my own code in my own source files (or generated files) there.

General features

This list only things that are (at least partly) implemented.

  • A dynamic typing framework (semantically similar to what Scheme or Javascript have). Most values are immutable (boxed strings, boxed numbers, sets or tuples of items, nodes), but items are mutable values (similar to "objects" or "structures" in Scheme or Javascript). We should later make that reflective, and generate all the C -or GCCJIT- code related to that.

  • A node value has an item connective and a sequence of son values (a bit like Prolog terms). It might have some fixed metadata (a metaitem and a metarank), which is ignored for comparison, equality, hashing. It is immutable. Nodes are also having a closure role (the connective giving the code, the sons being the closed values).

  • Items are externally identified by unique item names (internally, they are identified by their address). An item name has a radix (C-identifier like, but double or initial or terminal underscores are forbidden) and an optional suffix preceded with two underscores. For example, the_agenda or comment is an item name (without suffix). And web_state__0dmo5BBWK0mbtv is an item name of radix web_state and suffix 0dmo5BBWK0mbtv (this is an 80 bit positive random number -probably world-wide unique- nearly represented in base 60). Notice that item names are on purpose acceptable identifiers in generated C and in generated Javascript. Items have attributes (a map or "dictionnary" of some item associated to some non-nil value), components (a "vector" of values), possibly a (C or GCCJIT) function pointer, and may have some unique payload owned by the item. Each item has its mutex.

  • A persistent machinery, so the entire state of the MELT monitor, including its agenda, can be checkpointed and restarted on disk (persisted preferably in textual form for MELT monitor data, friendly to git version control, and for user data such as static analysis information, in REDIS store or PostGreSQL databases). That code should also be generated but it is not yet.

  • An agenda mechanism: the_agenda predefined item contains as its payload a queue of tasklet items. A fixed number (typically 3 to 10) of working threads are repeatedly fetching some tasklet from the agenda and running it. Of course tasklets are added (either in front, or at the end) to the_agenda's queue, either by running tasklets or by external events (such as some HTTP requests, an external process completing, etc...).

  • A very incomplete web interface. I'm struggling learning more HTML5 techniques, and most of my recent questions on StackOverflow are related to this. I'm understanding more and more that metaprogramming techniques are practically essential. HOP is inspirational.

Source files

Here is a description of the source files that I have wrote. Other files available on github are required, but are either copied from elsewhere or generated. Some generated files are mentioned, however. In the long term, most (ideally, all!) source files should be generated, but I have not yet bootstrapped my MELT monitor (like I mostly did for the MELT plugin), but it is a major subgoal (notably because that would ease adding new data structures in the MELT monitor).

  • meltmoni.h is the only (hand-written) header; it is including many other system or third-party libraries header, and some _mom*.h generated headers (in particular, the generated _mom_predef.h is listing predefined items).

  • mi19937ar.c is a near-copy of some MIT licensed Mersenne Twister pseudo-random number generator by Makoto Matsumoto and Takuji Nishimura; I modified some function names by prefixing them with momrand and making them more thread-friendly.

  • agenda.c is the agenda mechanism

  • primes.c contain many primes, and mom_prime_above (int64_t) and mom_prime_below (int64_t) functions. Prime numbers are e.g. useful as hash table sizes, etc.

  • canvedit.c, webroot/canvedit.js, webroot/canvedit.css, webroot/canbedit.html is a (probably aborted) tentative of using HTML5 canvases for web interface.

  • -and perhaps other *.mom files, is the persistent state in textual form, parsed with some stack machine algorithms.

  • hashed.c implements various hash tables, notably hashed sets, hashed associations, hashed maps, ... payloads.

  • hweb.c uses libonion to implement the web interface. Files (notably some jquery*.js) under webroot/ are statically served. Dynamic content (and AJAX) is related to the web_state and web_handlers items.

  • item.c has item related code.

  • main.c is the main program and utilities.

  • filebuf.c is for file & string buffers

  • microedit.c, webroot/microedit.css, webroot/microedit.html, webroot/microedit.js is a micro editor (HTML5 contenteditable!) to be able to edit values and use them. I am not happy with it, so I coded again the ....

  • nanoedit.c, webroot/nanoedit.css, webroot/nanoedit.html, webroot/microedit.js is a nano editor (no contenteditable)

  • modules/ should contain generated C code.

  • _mom_predef.h is generated and contains the predefined items.

  • state.c has the persistence machinery (loading and saving state notably in textual file). We'll need to persist to some REDIS store or some PostgreSQL database as soon as we'll be running in a distributed context (cloud or cluster).

  • value.c handles immutable values

It should be stressed that generating most of the above code is much simpler than hand-coding it.

Running that


Install all the required dependencies, then run make -j to build (in the source directory containing the Makefile and this ....)

Run the MELT monitor inside the source & build directory (the one containing the Makefile).

test run

It is alpha-stage software (and that is the third time I'm coding some MELT monitor). You won't be able to run it usefully in 2015. But if you want to help (e.g. to answer some questions I am asking on some forums) or see what is running, build it, then run ./monimelt --help to get some command line help. Later, try something like ./monimelt -Drun,web -W localhost.localdomain:8086/ -J 3 in the terminal and use your browser on URLs like http://localhost.localdomain:8086/hackc.html to be able to type some C code and run it, or http://localhost.localdomain:8086/miniedit.html to try the nano editor (very alpha, not working yet).

To simply dump the state and test the persistency machinery, try ./monimelt -d (perhaps also -Dload,dump). Modified files are backed up (e.g. old is backed up as, and likewise for _mom_predef.h~)

Killing the monitor with SIGTERM might trigger a dump (of the persistent state). Killing it with SIGQUIT or SIGKILL won't.

Help needed

This is currently an informal research work. Please help me to find funding for it (e.g. collaborative R&D projects where this work could fit). Otherwise, I might be asked to stop working on this. I could be interested in having my MELT system working on your proprietary software source code (with some funding available), as long as all my own (Basile Starynkevitch's) work remains free software, GPLv3 compatible. Notice that GCC runtime license strongly and legally discourage compiling proprietary code with proprietary GCC extensions.

Alternatively (but unlikely, because it is so specific), contribute some code to this. But since eventually this will be integrated into MELT -as part of GCC- (with a copyright transfered to FSF) you should follow the Contribute to GCC in particular its demanding legal prerequisites.

Since I am a near newbie in HTML5+Javascript technologies, I am sometimes asking questions on StackOverflow (and elsewhere) related to them. Please help me if you can.

Contact information

basile at starynkevitch dot net (the mailbox I read more often) for technical and scientific discussions or questions.

basile dot starynkevitch at cea dot fr (the mailbox I read at office hours only) for funding or business related discussions: collaborative R&D consortiums, industrial contracts

I (Basile Starynkevitch) am French, living near Paris, working at CEA, LIST, the IT part (nearly a thousand persons) of CEA (a 15000 persons, 4.4 billion euros/year budget, French government owned applied research agency, see Commissariat a l'Energie Atomique et aux energies alternatives wikipage). FWIW, the Frama-C, Papyrus, Unisim, Gatel (and of course this MELT monitor and the MELT plugin...) tools are developed in the same DILS department (100+ persons) where I work.

You can’t perform that action at this time.