TLDR; I took a lot of complex Java source code for a complete R5RS Scheme interpreter and converted it to C# and it worked! It was much harder than this little TLDR makes it sound.
I’ve been on a Scheme learning kick lately because it was something I always wanted to do for many various reasons. My original goal was to just read and work through the classic SICP book. I’m not very far along on that front because at some point I needed a Scheme system and I got sidetracked and sucked into the totally ridiculous idea of porting a Java-based R5RS compliant Scheme interpreter (SISC) to C# and .NET. I’m not really sure why, it just sounded challenging and interesting to me and I got sucked in by the somewhat naïve and seemingly possible notion that it could mostly be done mechanically. Plus there really are no full-featured Scheme interpreters or compilers for .NET* – the few that exist leave out features like support for full continuations and tail calls - stuff that would be useful for really learning Scheme. Although it hasn’t been updated for a few years, SISC claims to be a fully R5RS compliant Scheme interpreter.
I started with a machine translation of the java source code (using Sharpen). Even getting Sharpen to work was a unusual chore (my fault mostly) and what C# it split out would not build. SISC is comprised of 203 java class files – 18,000+ lines of actual source statements. What Sharpen gave me is currently much more C# code because it uses a lot of custom classes and shims to match Java’s intentions, functionality, etc.… It’s (hopefully) a lot of extra layers that I will be able to mostly remove with careful, targeted re-writes to native C#.
Getting it to build wasn’t too hard, but it took time. Sharpen’s conversion doesn’t have a shim for everything so sometimes the code it translates calls class methods that don’t exist in .NET. Other times it entirely skips and omits lines of Java code altogether (I’m still manually checking for those cases here and there). So I had to write more shim code and hand translate what was missing. Getting it from simply building to actually working though was far harder.
Oh, did I mention that I am not really a Java programmer? I’ve dabbled here and there before but nothing as complex as this and many of Java’s conventions and idioms were less than obvious to me.
The hardest part was that the SISC code base is complex and hard to understand. It seems to be very well written though. There are two primary ways to write a Scheme system: an interpreter or a compiler (either to native code, a byte code VM, or cross compilation to another language like C). Both are complex things to write and understand, especially if you did not write them. SISC is an interpreter which means that much of the code is just plain Java. But that code looks up symbols, builds expression trees, applies procedures, manages registers and call stacks – all abstractions of the Scheme language it is interpreting. In doing so the source code for SISC leverages some features of Java that have to be carefully understood when translating to another language.
First off, Java has some different notions about things like equality from .NET. For instance, in Java if you have two Lists classes and use equals() to compare them, the contents of the lists are checked to determine equality. In .NET this is rarely the case as Equals() usually just settles for reference equality. So more shim code to fix that.
Collection classes can also be pretty different, with Java supporting things like mutable views into Lists instances via subList(). Some Java collection classes have no equivalents in .NET. The Java SISC source was also not written to use generics and instead liberally uses Object which makes it hard to tell what is intended to go into any collection instance (and often it is a mixed bag).
Then there’s things like WeakHashMap or ReferenceQueue which not only do not exist in .NET, they are near impossible to accomplish since the CLR doesn’t publically expose the garbage collector hooks necessary to write .NET equivalents that follow Java semantics. To create similar types of things in .NET you have to liberally use WeakReferences together with polling and/or class finalizers. It works but is not ideal.
Finally there’s exceptions, which normally wouldn’t have been a big deal but the SISC code base unfortunately relies on using exceptions for non-exceptional things, like rewinding the interpreter’s execution when failing to convert a piece of text that might be a number (or a symbol that just looks like a number) into a number. Catching exceptions happens polymorphically which means that if your exceptions class hierarchies are out of sync it’s going to be a world of hurt. Telling the “normal” exceptions apart from the true failure exceptions was tough. Luckily, it turned out to be not as bad as it could have been but I had to spend considerable time studying both Java’s and .NET’s exception class hierarchies to see where there were out of sync and fix up the converted code so that the original intentions were preserved. I am sure there are more similar Java to C# gotcha’s waiting for me in this converted code base that I have yet to uncover.
The final complication came in the form of Scheme itself. The core of an R5RS Scheme language is a very small, simple language containing just 23 constructs, 11 of which can be built from combinations of the other 12. The entire R5RS language spec is less than 50 pages. To have a functioning Scheme language that does something interesting though requires much more, like a set of primitive functions (math, string functions, basic I/O APIs, etc.…) and a Scheme library so that it can interface with input and output devices, files systems, etc... Most of these primitives are contained within the SISC interpreter as Java (and now C#) native code but much of the library is actual Scheme code that uses these lower level native primitives to build a more complete and useful library. This Scheme library code amounts to more that 70,000 lines of Scheme code that gets built into a heap image during the build process. The trick is you need a functioning Scheme system to build this Scheme heap. However, SISC was designed to be self-bootstrapping, which meant that it could load and parse a small set of special Scheme source files and split out a binary heap image to a file that could then be reloaded when running the full interpreter.
All throughout this project I had the nagging question of “how will I know when it’s really working?” It’s an interpreter so unless it is interpreting lots of code it would never be tested very well. However at this point I could barely understand Scheme source much less write any. The surprise answer came when I finally understood that the self-bootstrapping process was in fact executing lots and lots of actual Scheme code, parsing it and building internal representations of Scheme s-expressions. Moving beyond having it simply not blow up, verifying that the heap image it was creating turned out to be very difficult.
The heap is structured as a memory image of serialized s-expressions with lots of pointer indirection in the form of dynamically generated integer IDs that index into tables of class types, expressions, data fragments, and BER encoded integers. To make matters worst, most of the internal data structures that get serialized are stored internally in hash tables which means that when serializing these sets they usually get written out in a slightly different order each time (i.e enumerating hash set entries with dynamically generated symbol names is not guaranteed to be deterministically ordered from run to run). In the end I got the bugs squashed by modifying both code bases to be completely deterministic when writing heap images and output copious amounts of trace logs. This way I could do a binary comparison of the heap image files and logs as well as step through each code base side-by-side to see where they differed when reloading the heap. it was the most frustrating part of the project but peace of mind in the end to know that internally, they were producing the exact same set of data structures. This also meant that they were interpreting lots of Scheme code identically too to produce those data structures. This was as close to a testing framework as I was likely to get, comparing one working Java version to a converted C# version when processing Scheme source code.
All in all it took a few weeks but I got it working – not complete, but complete enough to be useful (some parts of the library are not ported yet, like networking functions, the FFI, etc.…). It was one of those things where it seemed possible at first glance, then impossible once I got into it, then suddenly and surprisingly possible again the next moment – flip-flop, flip-flop with each little bit of progress I made. I came close to stopping several times but then I’d have a little breakthrough and it would seem like the finish line was very near (only to be not). In the end is just suddenly mostly started working with minor little bugs here and there. I can’t quite prove that it completely works though but I’m able to learn Scheme and use little bits of Scheme snippets and test cases I find around the net. I still don’t fully understand how the interpreter works though – it is very hard to mentally trace through the code.
All so I could learn stuff like this: http://www.willdonnelly.net/blog/scheme-syntax-rules/
Scheme’s ability to extend the language syntax is fascinating to me. You can write a whole new language right inside of Scheme with new syntax rules and everything.
Unfortunately, SISC.net is a ways from being put out there for anyone else to use but someday I hope to do so. The issue is, it contains some source code that is Ok for me to use but not so Ok for me to redistribute (Reflector *cough* *cough*). When I’ll be able to do so I don’t quite know. This was my fun little fall learning project and I’m done for a while now. If you’re a Seasoned Schemer and this is something that would be useful to you though, knowing full well that it is experimental at best, leave me a comment to prod me along to get it into a more stable release state.
P.S. I also wrote a Forth threaded interpreter in .NET but it was much, much simpler than any of this and I had an old but good reference for how to do so. It’s the project I never completed from my early days of computing. It’s not quite finished yet either (it needs a bit of a re-write to solve some issues) but I’m working on getting it out as well. Forth is a magical little language.
* The exception is the Common Larceny project but its future does not look too bright: https://lists.ccs.neu.edu/pipermail/larceny-users/2010-October/000829.html