C backend re-architected to SSA form.

Well, one year later, I've managed to back into the CPS output via a completely different direction.  The rewrite began as an experiment, triggered by the completely unpredictable behavior of new versions of clang when compiling Irken's output.  [I also found a gcc bug.]  I thought I'd try to generate 'real' CPS - i.e., separate C functions that only make tail calls.  This turned out to be much easier than my attempts last year starting from the other end.

After only about 3-4 days of work, we now have output like this:

(let loop ((n 1000000))
  (if (zero? n)
      "done"
      (loop (- n 1))))

static void FUN_loop_17_0 (void) {
  O r10 = ((object*) lenv) [2];
  if PXLL_IS_TRUE(PXLL_TEST(unbox(r10)==0)) {
    O r12 = (object*) &constructed_0;
    PXLL_RETURN(12);
  } else {
    O r16 = (object *) 3;
    O r17 = ((object*) lenv) [2];
    O r15 = box((pxll_int)unbox(r17)-unbox(r16));
    lenv[2] = r15;
    FUN_loop_17_0();
  }
}

A few things to note: the loop is achieved with a tail call in C.  This requires that the C compiler perform tail call optimization or the stack will explode rather quickly.

The 'registers' are now emitted as local variables, with a fresh name for every line. This output is very, very close to the SSA form needed for LLVM itself, and is practically begging for an LLVM backend.

There is no longer any dependency on C extensions - the address-of-label extension was never a happy citizen in LLVM land, anyway - and the lexical functions extension (needed by Irken circa 3 years ago) is a distant memory.

The output of the compiler runs about the same speed, but the compilation time itself is much more predictable.  clang -O3 takes about 30 seconds, and gcc -O3 around a minute.  gcc actually seems to do a better job.  I need to play around with clang and gcc's options in order to figure out the best flags for fast edit-compile-link turnaround - I'm missing the 4-second compiles I used to have with clang and the old compiler.

Other great possibilities include separate compilation, JIT, and (possibly) closure conversion / passing args in registers.  An LLVM backend might open up the possibility of efficient floating-point support.

However, the next step is probably to write a real garbage collector.  I'm seriously considering using Irken for some real server work soon, and to survive in that world a two-space collector just isn't going to cut it.  I've read over the "Beltway" paper several times and I think I grok it well enough to implement it.

LLVM, CPS

I've done quite a bit of work these past two months on writing an LLVM backend.

Rather than directly translate the current design (one giant function), I'm trying to map function to function, and use llvm's support for tail calls to get arguments-in-registers to work.

This requires a 'real' CPS transform, and one that's up near the front of the chain of passes, not the last thing before the backend.

The CPS transform introduces 'continuation functions' - these are similar to normal functions, except they take a single argument.  They are targeted by SSA 'terminators' like branch, jump, etc...

A continuation function corresponds to a 'basic block' in SSA - think of the phi insn as if it took the single argument and moved it into the correct register.

If every function were a tail call, we could emit all the code as basic blocks with branches of various kinds. Sadly, some calls are not tail calls.  This requires that we render the continuation function as a real function, and that all 'jumps' to it are tail calls.

Needless to say, this is a pretty major rewrite of Irken.  I've already made three failed swipes at it, but I hope to find the time to get it working within the next few months.

pattern matching with regexps

This is just a reminder to myself to think about extending pattern matching to regular expressions on strings.  It'd be a great feature, though with the famous caveat about regular expressions.  [I have a problem.  I think I can solve it with regular expressions.  Now I have two problems.]



(define sample
  "(?P<x>[0-9]+)+(?P<y>[0-9]+)" -> (+ (string->int x) (string->int y))
  "[^x]+(?P<nx>x+)[^x]+" -> (string-length nx)
  )


I still remember looking through the ejabberd source code, trying to find their xml parser.  When it finally dawned on me that it was automagically hidden in the pattern matching, I was impressed.

Some notes on continuation-passing style

I've been going over the CPS transform lately, and wrote some notes on how to use the transform to compile to C, stackless style.

dark.nightmare.com/rushing/factcps/

thinking about an llvm backend, again

I've been very happy with my other project that uses LLVM as the backend for a compiler.
I've also learned a lot about LLVM's IR design.

I can't help but think that it might only take a week or two to write a back end for Irken.
However, there are a few issues.

1. My CPS output assumes that each insn has a result that is put into a register.  This doesn't fit LLVM's model, which assumes that arguments to insns can be either an immediate or a register.  In fact, they do not allow you to put an immediate into a register.
2. LLVM doesn't like gcc's goto-label feature.   I think they implemented it reluctantly, because it doesn't fit the SSA model very well.  The majority of funcalls by Irken are to known functions, which translate to a branch to a known label.  However, whenever higher-level functions are called, this leads to something like "goto *r0;" in C, which turns into an indirectbr insn in LLVM.  It implements this using a gigantic phi/indirectbr table at the very end of the function.  Maybe this isn't really a problem - it just looks bad.
3. The extremely useful %cexp primitive would go away!  There's something to be said for being able to refer to OS constants, OS functions, macros, etc... with such a simple mechanism.  I'd probably just have to let it go, and force users to write stubs.

I think #1 can be dealt with by adding an extra pass that translates between the two representations.  (maybe I should revisit the CPS==SSA paper?)

Another approach might be to use an annoying alloca/store/load sequence and hope that mem2reg does a good job?

Irken status, LLVM bitcode

This is just a quick note that the Irken project has not died, only gone into a temporary hibernation.
When in the course of history it becomes necessary to earn a living...

Anyway, I have some tangentially related comments about LLVM.
My current employer is interested in maybe doing a little LLVM JIT work, wherein my experience with compilers may prove useful.  A nice side effect for me is that I finally get to play around a bit with LLVM, something which I carefully avoided while working on Irken.

My first reaction was to pick up llvm-py, which looked like a fairly complete interface to the various libraries that make up the llvm system.

Big Mistake.  It doesn't build any more.  It's only a little over a year old.  The last release of llvm-py was for LLVM 2.8rc2.  I am now running 3.0.

Here's the problem: the LLVM API's are constantly changing.  And that's a good thing, really.  But since those C++ API's are the only well-supported interface to the system, it's a moving target.

There are three main options for a compiler writer who wants to target LLVM.  The first is to use the C++ API.  (or you could use the completely undocumented and incomplete C interface).

The second option is to write llvm assembly.  This is not a bad option, but will be slower because of the need to parse the assembly.

The third is to write llvm 'bitcode'.  I thought this was probably the right approach.  Bitcode is a binary representation of the LLVM IR language, it seems that it would be likely to change much more slowly than the library interfaces.

The problem is that the person who designed the bitcode file format was on meth at the time.  This has got to be one of the most evil formats I've ever seen.  I think they were trying to make the resulting file as small as possible, but at the expense of counting every bit.  (perhaps an early LLVM project flew on a space probe?)  Just to give you an idea, 3 and 4-bit integers (not aligned on any boundary) are a common field type.  And not just that, but some of the fields are variably sized, starting at 2 bits and going up.  Symbols are represented using 6-bit characters.  Also, the concept of 'bitstream' is taken very literally, the bits in a file are viewed as if it were one large integer, starting from the LSB.  This is actually an elegant approach, but it is completely different from every other file format or network protocol.  I had to continually resist using common parsing patterns with it.   There's also a compression-like method of 'abbreviating' common record types.  And after all that work to save bits, at the end of every block the bit stream is aligned to 32 bits!

Ok, got that off my chest.

My sense is that nobody uses the bitcode format because of this (the only other implementation I could find was in Haskell, and was also impenetrable because Everything Is A Monad, You Know).  Most people just succumb to the pressure to write against the C++ API, and will then spend the rest of their lives reading the llvm-dev list to keep track of the changes.

I have mostly decoded the format, and I think I could actually output it for a compiler.
But I think I'll split the difference and start out by outputting llvm assembly to start with.  If the day comes when I need to shave some of the 10MB off the resulting .so file, and maybe speed things up, I'll revisit the bitcode issue.

In the meanwhile I've had a whack at using Cython's C++ support to make a minimal interface to LLVM where I can feed either llvm assembly or bitcode to the JIT and run it.

git, github, clang

I've imported my svn history into git, and published the whole thing on github:

https://github.com/samrushing/irken-compiler/

Still trying to grok git and github, there'll probably be a stall for a week or so.

In the meanwhile, I have interesting news:  I was able to take out all the 'nested functions' which relied on a gcc extension.  I resisted this for a long time because early testing had shown that declaring the 'registers' as local register variables in the main function was a huge win.

After a quick test the other day, I found that moving everything out to global variables has no visible impact on performance.  I don't know if this is because of improvements in gcc & llvm, or if I just wasn't trying a large enough program.

That means I'm now able to compile the output with clang, rather than relying on llvm+dragonegg, which can be much more annoying to build, install, and use.

The good news continues: although the performance of the generated code is identical (maybe a tiny bit faster), the compilation time has been cut quite significantly.  On my machine an LLVM -O2 compile dropped from 18m to 8m.  The -O0 compile time is even sweeter, about 3s, which is actually faster than Irken itself, which hovers right around 5s to self-compile.  So a roughly 10s edit-compile cycle will definitely speed the development of Irken up.

Now the bad news.  The no-nested-funs version causes 'as' on snow leopard to take a lunch break.  It actually nails the CPU for 10 or 15 minutes before finally emerging as if nothing was wrong.  No problems on FreeBSD 7.3 with gcc-4.5.3, though.  This could be a significant annoyance, since I expect many people trying out Irken would prefer to do so with a stock compiler.

More bad news:  clang-2.9 crashes when fed Irken's self-compiled output.  I was going to file a bug against it, but when I checked out the svn version and built it the problem was gone.  Sigh.

This is why I'm trying to figure out how to make branches on github, since it'd be nice to maintain both versions for a while until these issues are ironed out.  I have made a local branch called 'no-nested-funs', I just haven't figured out how to publish it on github.

nested datatypes, 2-3 trees, numerical encoding

This paper by Hinze & Paterson discusses the use of 'nested datatypes' to implement 2-3 trees, but in a way that uses the type system to automatically handle the invariants.  A blog post from Mark Chu-Carroll is a bit easier to follow.  Here's his example of a sequence encoded as a 2-3 tree (in Haskell):

data Tree t = Zero t | Succ (Tree (Node t))
data Node s = Node2 s s | Node3 s s s

Each level of the tree contains sub-trees that are 'one less' via the numeric encoding scheme.

This sent me back to Okasaki's book, where I'm finally starting to grok those last two chapters.  This bizarre mingling of numerical encoding and data structures lights a fire in the parts of my brain dedicated to data structures.  If feel as if I could automatically see the relationship between red-black trees, 2-3 trees, and binary number representations.... if only I could double my IQ.

I have this cool puzzle called the 'Spin-Out', which is actually a physical demonstration of the properties of Gray Codes.  It seems to me that Gray Codes and balanced trees should be a natural fit.

Unfortunately, I can't experiment with any of these ideas in Irken (or OCaml, or SML) because the nested datatypes require polymorphic recursion.

Type inference with polymorphic recursion is undecidable.  Haskell allows it only if the user provides a manual type signature.  I'm guessing this is impossible (i.e., unsafe) with SML/OCaml because of the presence of assignment?  Could such a feature safely be used with functions that the compiler can verify are pure?

DOOM DOOM DOOM (a coro/thread kqueue-based scheduler)

I've created a new sub-project, code-named 'doom', for the coroutine-based cooperative threading system.  It uses kqueue(), of course.  Because it's better.  It wouldn't be too hard for someone from linux-land to make a version based on epoll.

In the doom directory, you'll find a simple http demo, and the obligatory echo server.

Exception handling certainly makes the code look a lot nicer.

I'm the best at space.

exception handling, take two

A bit tricky, but I found a way to implement try/except almost completely with defmacro.  I needed to add some extra typechecking around raise and try to ensure that all exception types are monomorphic.  I achieved this by having the compiler keep a global map of exception-name => tvar.   Seems to work...

While working on this I learned that the Scheme-style call/cc is not really type safe, and I need to use the SML version.  The main difference is the protocol - SML's callcc requires the use of a throw procedure since the continuation needs to carry type information:

(define (callcc p) : (((continuation 'a) -> 'a) -> 'a)
  (p (getcc)))

(define (throw k v)
  (putcc k v))

;; -*- Mode: Irken -*-

(include "lib/core.scm")
(include "lib/random.scm")

;; We use polymorphic variants for exceptions.
;; Since we're a whole-program compiler there's no need to declare
;; them - though I might could be convinced it's still a good idea.
;;
;; Exception data must be monomorphic to preserve type safety.
;;
;; <try> and <raise> are implemented as macros, with one extra hitch:
;;  two special compiler primitives are used to check that exception
;;  types are consistent: %exn-raise and %exn-handle

(define (base-exception-handler exn) : ((rsum 'a) -> 'b)
  (error1 "uncaught exception" exn))

(define *the-exception-handler* base-exception-handler)

(defmacro raise
  (raise exn) -> (*the-exception-handler* (%exn-raise #f exn))
  )

(defmacro try
  ;; done accumulating body parts, finish up.
  (try (begin body0 ...) <except> exn-match ...)
  -> (callcc
      (lambda ($exn-k0)
        (let (($old-hand *the-exception-handler*))
          (set!
           *the-exception-handler*
           (lambda ($exn-val)
             (set! *the-exception-handler* $old-hand)
             (throw $exn-k0
              (%exn-handle #f $exn-val
               (match $exn-val with
                 exn-match ...
                 _ -> (raise $exn-val))))))
          (let (($result (begin body0 ...)))
            (set! *the-exception-handler* $old-hand)
            $result))))
  ;; accumulating body parts...
  (try (begin body0 ...) body1 body2 ...) -> (try (begin body0 ... body1) body2 ...)
  ;; begin to accumulate...
  (try body0 body1 ...)                   -> (try (begin body0) body1 ...)
  )

(define (level0 x)
  (try
   (level1 x)
   except
   (:Level0 x) -> (:pair "level 0" x)
   ))

(define (level1 x)
  (try
   (level2 x)
   except
   (:Level1 x) -> (:pair "level 1" x)
   ))

(define (level2 x)
  (try
   (level3 x)
   except
   (:Level2 x) -> (:pair "level 2" x)
   ))

(define (level3 x)
  (try
   (match x with
     0 -> (raise (:Level0 x))
     1 -> (raise (:Level1 x))
     2 -> (raise (:Level2 x))
     3 -> (raise (:Level3 x))
     4 -> (:pair "no exception!" 99)
     _ -> (raise (:Bottom x))
     )
   except
   (:Level3 x) -> (:pair "level 3" x)
   ))


(printn (level0 0))
(printn (level0 1))
(printn (level0 2))
(printn (level0 3))
(printn (level0 4))
(printn (level0 5))

exception handling with defmacro

I've started work on a coroutine/thread scheduler, code named "Doom".  Doing some socket I/O has finally  snapped my 'missing feature' threshold - I really need exception handling to do it right.

Here's my first attempt. The basic syntax is (try <body> except <exception-patterns>).

If this works, it'll be a great example of the huge advantages of the lisp syntax and macros - no changes made to the compiler to support it!

;; -*- Mode: Irken -*-

(include "lib/core.scm")
(include "lib/random.scm")

;; can we implement an exception system using the macro system?
;; XXX eventually we'll need to change call/cc to swap out exception
;;  handlers, which probably means making a 'modern' call/cc with
;;  dynamic-wind, etc.

(define (base-exception-handler exn)
  (error1 "uncaught exception" exn))

(define *the-exception-handler* base-exception-handler)

(define (raise exn)
  (*the-exception-handler* exn)
  )

;; what kind of values do we want for exceptions, and how they're caught/compared?
;; python: started with strings, went to objects of the Exception class.
;; scheme: SRFI-12 includes something that looks like property lists?
;; sml: string refs? (in sml/nj according to CwC)
;; ocaml: declared exception datatypes.
;; [alternate for ocaml: http://dutherenverseauborddelatable.wordpress.com/downloads/exception-monads-for-ocaml/
;;  sounds a little like my 'fourth idea']

;; my first temptation is to just use symbols, and stay simple

;; second idea is to use a record, which theoretically gives you the
;;   ability to attach other kinds of data, which could be very
;;   convenient. Example: (raise BadFD) => {exception='BadFD ...}

;; third idea: hardcore - define a global exception type, force the user to
;;   extend it.  [could be made easier by compiler hacks that allow you to
;;   extend the type anywhere in the source?].  Code would have to wildcard
;;   exceptions it doesn't know about?  [or is that the definition of handing
;;   it upstream?]

;; fourth idea: use records (or polymorphic variants) to define unique
;;   exception names *and* types.  For example: (raise (BadFD
;;   current-fd)) would type as {BadFD=int ...}  this would make it
;;   more like the SRFI-12 property-list thing (unless I misunderstand
;;   what SRFI-12 is all about).  Hmmm... maybe we could extend the
;;   exception handler by extending the record?

;; It'd be nice if the type system can tell us what exceptions are possible
;;   at any given point in the code... is this maybe impossible due to the
;;   dynamic (vs static/lexical) nature of exception handling?

;; the more I think about the 'fourth idea' the less reason I see to
;;  restrict the type of exceptions.  Syntax-wise, the following macro
;;  does not enforce any restrictions.  It's possible though that only
;;  the polymorphic-variant-scheme will survive type checking.

;; this is the first time I've tried to 'accumulate' pieces in a macro.
;; it's possible that there's a better idiom that I just haven't discovered yet.
;; should really look to see how the scheme folks do this (or do they just avoid
;; it all by wrapping everything in verbose s-expressions).

(defmacro try
  ;; done accumulating body parts, finish up.
  (try (begin body0 ...) <except> exn-match ...)
  -> (let (($old-hand *the-exception-handler*))
       (set!
        *the-exception-handler*
        (lambda ($exn)
          (set! *the-exception-handler* $old-hand)
          (match $exn with
            exn-match ...
            _ -> (raise $exn))))
       (let (($result (begin body0 ...)))
         (set! *the-exception-handler* $old-hand)
         $result))
  ;; accumulating body parts...
  (try (begin body0 ...) body1 body2 ...) -> (try (begin body0 ... body1) body2 ...)
  ;; begin to accumulate...
  (try body0 body1 ...)                   -> (try (begin body0) body1 ...)
  )

(define (random-barf)
  (if (= (logand (random) 1) 1)
      (raise (:OtherException 99))
      7))

(define (thing)
  (try
   (let loop ((n 0))
     (if (= n 100)
         (raise (:MyException 12))
         (begin
           (random-barf)
           (loop (+ n 1)))))
   except
   (:MyException value) -> value
   (:OtherException _)  -> 9
   ))

(thing)

Here's what the macro expansion looks like for thing:

(define (thing)
  (let (($old-hand *the-exception-handler*))
    (begin
      (set!
       *the-exception-handler*
       (lambda ($exn)
         (begin
           (set! *the-exception-handler* $old-hand)
           (%fatbar
            #f
            (%nvcase
             nil
             $exn
             (MyException OtherException)
             (1 1)
             ((let-splat
               ((m9 (%nvget (:MyException 0 1) $exn)))
               (let_subst (value m9) value))
              9)
             (%match-error #f))
            (raise $exn)))))
      (let (($result
             (letrec
                 ((loop
                   (function loop (n) (if (= n 100) (raise (:MyException 12)) (loop (+ n 1))))))
               (loop 0))))
        (begin (set! *the-exception-handler* $old-hand) $result)))))

same-fringe demo

This demonstrates how to use generators to solve the same-fringe problem.

Also, I found a minimalist way of building binary-tree literals using defmacro.

;; -*- Mode: Irken -*-

(include "lib/core.scm")

(datatype btree
  (:node (btree 'a) (btree 'a))
  (:leaf 'a))

(defmacro btree/make
  (btree/make (l r)) -> (btree:node (btree/make l) (btree/make r))
  (btree/make x)     -> (btree:leaf x))

(define t0 (literal (btree/make ((0 ((1 (2 (3 4))) 5)) (((6 7) ((8 (9 10)) 11)) ((12 (((13 14) 15) (16 17))) (18 19)))))))
(define t1 (literal (btree/make (((0 ((1 2) 3)) (((4 5) (((6 7) 8) (9 10))) ((11 ((12 13) 14)) ((15 (16 17)) 18)))) 19))))
(define t2 (literal (btree/make (((0 ((1 2) 3)) (((4 5) (((6 7) 8) (9 10))) ((88 ((12 13) 14)) ((15 (16 17)) 18)))) 19))))

(define btree/inorder
  p (btree:leaf x)   -> (begin (p x) #u)
  p (btree:node l r) -> (begin (btree/inorder p l) (btree/inorder p r) #u))

(define (btree/make-generator t)
  (make-generator
   (lambda (consumer)
     (btree/inorder
      (lambda (x) (consumer (maybe:yes x)))
      t)
     (forever (consumer (maybe:no))))))

(define (same-fringe t0 t1 =)
  (let ((g0 (btree/make-generator t0))
        (g1 (btree/make-generator t1)))
    (let loop ((m0 (g0)) (m1 (g1)))
      (match m0 m1 with
        (maybe:yes v0) (maybe:yes v1)
        -> (if (= v0 v1)
               (loop (g0) (g1))
               (print-string "NOT equal\n"))
        (maybe:no) (maybe:no)
        -> (print-string "equal\n")
        _ _ -> (print-string "unequal size\n")))))

(same-fringe t0 t1 =)
(same-fringe t0 t2 =)

win32 binary

I've tried to make the bootstrapping script work on win32, and I think for the most part I've succeeded.  There are issues with mac-specific flags making it into the bootstrap file that I still need to work out.

I used mingw, but I suspect other gcc binaries will work, too.  A 32-bit executable is now available, which may ease the bootstrapping process for some.

description of the compiler and runtime

I've started on a 'HACKING.txt' document describing the compiler and runtime.
It should be enough to get folks started on understanding the source.
Feedback appreciated!

-Sam

self-hosted distribution

I've finally put together a 'real' self-hosted distribution.

The tarball is a little bigger - even though I removed all the python code.  The reason is that I have to distribute a pre-compiled version of self/compile.c so that the compiler can be bootstrapped.

Enjoy, and feedback appreciated.

Here's the text of the README:

This is the initial release of the self-hosted compiler.

It's still in a very unpolished state, but you can use it to bootstrap itself.

Just run the script "util/bootstrap.py":

$ python util/bootstrap.py

Which does the following:
1) run gcc on the distributed version of self/compile.c
2) this binary will be used to recompile the compiler.
3) that binary will recompile the compiler again.
4) the output from steps 2 and 3 are compared, they should be identical.

If you're happy with the resulting compiler, you can compile an optimized
version of self/compile.c, but be warned - you'll need a lot of memory and
a lot of time.

I am using dragonegg for optimized builds, and that seems to take about a GB
of memory, and 18 minutes to build.  It's important to use '-O2', not '-O',
because '-O' takes 53GB of memory and hours to compile.

Very little documentation exists yet, try 'lang.html' for a brief tutorial.
The best way to get familiar with the language is to read the source code in
the 'self' directory, and browse over the files in "tests".

-Sam

C compiler issues

As I came closer to completing the Irken implementation, I noticed that my edit-compile cycle was taking longer and longer.  And by that, I don't mean a linear change.  At some point a threshold was crossed, such that compiling an optimized binary could take nearly an hour with dragonegg, and much longer with gcc, consuming over 17G of memory while at it!

After doing some tests, I've identified at least one of the causes: my varref and varset functions.

A couple of years ago, the compiler output for a varref insn looked like this:

  r0 = lenv[1][1][1][0];

Where the variable we are referencing is 3 levels up and at the 0 index.  (i.e., a De Bruijn index of (3,0)).

I noticed that I could write an inline lexical function, varref(), that did this with a loop:

  r0 = varref (3, 0);

... which is much cleaner.  With -O, gcc, llvm, and dragonegg were all unrolling the constant loop and creating code that was identical to the first version.

I didn't notice the cost of this convenient feature until my program size got large enough... the compiler sources, when using the inline functions, take 5X as long to compile -O as the first version.

Also, a 'platform' note: I work on OS X, where the stock compiler is still /usr/bin/gcc.  I did some timings for a non-optimized build and discovered that the stock gcc is over twice as fast as either my hand-built gcc-4.5.0, or dragonegg.  So for quick edit-compile cycles, I switched back to the stock version.  Though it'd be nice to know why the version from Apple is so much faster...

Self-hosted!

[assumed skynet joke here]

A few days ago, after many weeks of frenetic work, the new self-hosted Irken compiled itself.  There's still a lot of work to do, but this major milestone has been reached.  The compiler passes the stage1==stage2 test, and I'm now concentrating on various features still missing compared to the python compiler.  Also, the edit-compile cycle is still pretty slow, because both the compiler and gcc are slower than need be.  Once I have these issues solved and some rudimentary error reporting (the current error reporting might be described as medieval) I'll make an official release.

I'm also thinking about how to best re-package the system, considering orphaning the python development branch, and starting a whole new repository for the self-hosted system.

In the meanwhile, if there's anyone out there that would just like to see it compile itself, I rolled a demo tarball up.  Enjoy, and feedback is much appreciated!

progress on self-hosting

Today the self-hosted Irken compiled its first complete program, all the way through substituting into the header template file, and compiling via gcc. A nice feeling.

;; -*- Mode: Irken -*-

(datatype list
  (:nil)
  (:cons 'a (list 'a))
  )

(defmacro list
  (list)         -> (list:nil)
  (list x y ...) -> (list:cons x (list y ...)))

(define (+ a b)
  (%%cexp (int int -> int) "%0+%1" a b))

(define length
  () -> 0
  (_ . tl) -> (+ 1 (length tl)))

(length (list 1 2 3))

I also finally got around to checking in the changes from the last couple of weeks, too much slacking off there.

Y combinator in Irken

The second version requires recursive types.

;; -*- Mode: Irken -*-

;; See http://en.wikipedia.org/wiki/Fixed_point_combinator

(include "lib/core.scm")

(define (Y f)
  (lambda (x)
    ((f (Y f)) x)))

(define (factabs fact)
  (lambda (x)
    (if (= x 0)
        1
        (* x (fact (- x 1))))))

(printn ((Y factabs) 5))

;; and http://caml.inria.fr/pub/docs/u3-ocaml/ocaml-ml.html#toc5
;; let fix f' = let g f x = f' (f f) x in (g g);;
;; let fact = fix fact' in fact 5;;

(define (Y2 f1)
  ;; the '^' prefix tells the compiler not to inline this function
  ;;  which in this particular case would never terminate...
  (define (^g f)
    (lambda (x)
      (f1 (f f) x)))
  (^g ^g))

(define (fact1 fact x)
  (if (= x 0)
      1
      (* x (fact (- x 1)))))

(printn ((Y2 fact1) 5))

The function ^g has type

μ(A, (A->b))->(c->d)

recursive types, unification, sharing

I'm having trouble with recursive types in my type reconstruction (type inference) algorithm, some kind of combinatorial explosion.  I think there are multiple triggers for this: at each call to unify(), I'm running a complete apply-subst(), rather than a level at a time.  For the massive types generated by my OO code, this means lots of wasted, repeated work.

But of course there's even more going on.  I think my code that detects recursive types is not working as well as it should, and I think the cause is the lack of 'sharing'.  Remy/Pottier use a different style of unification, that involves preserving shared elements of types, and I think this where I'm screwing up: because parts of types are being copied, I'm missing links between identical parts of the graph.

I found a nice presentation from Remy describing "Core ML", that explains this in (accented) English, rather than a relentless flurry of LaTeX: "Using, Understanding, and Unraveling The OCaml Language: From Practice to Theory and vice versa".

The constraint-oriented solver gave me recursive types for free, (because its output types are naturally graphs), but I'm really hoping to avoid going back to that code, because it's much slower and I frankly don't understand it as well as I'd like.  Maybe I can pull out just the multi-equation unification stuff, and get back recursive types.