Arc Forumnew | comments | leaders | submitlogin
1 point by almkglor 6072 days ago | link | parent

I assume that for code blocks 3 and 4, 'let is the Lisp 'let, not the Arc one? (let ((var val)) ...) not (let var val ...)?

Also on code block 1 you show (prn (f 30)) but in code block 2 you seem to pass (f <continuation> 20)? I assume this is just a mistake?

I managed to understand the transformation up to code block 3, but not code block 4; what does '%closure do? Allocate heap memory for the closure? Also, what's the syntax for "goto"?

Finally: how about the equivalent C code?

LOL, maybe I should just wait for you to push it on the git and experiment with it myself.



2 points by almkglor 6072 days ago | link

Aruu, okay okay I finally actually looked at the presentation docs, which I probably should have looked at first. One of the things that threw me off was 'self - I was thinking of Arc's sense of 'self as in 'afn!

The output C code looks suspiciously like assembly language to me. Perhaps we can also target a temporary assembly syntax so that we can do minimal peephole opts, such as convert stuff like PUSH(x); y = TOS(); to MOVE(x,y);. Can't wait to actually see this code on the git ^^.

P.S. Given that the transformations are (gasp!) syntactic, it might actually be possible to implement the entire compiler as (gasp gasp!) a treeparse parser (or at the very least a piped chain of treeparse parsers) ^^.

-----

2 points by sacado 6072 days ago | link

Maybe treeparse would be the right thing to use... The code is getting uglier everytime I try to add a new primitive / special form... I dunno...

As for the generated code, yes, it's a lot like a portable assembly code. There are certainly easy optimizations to perform on it, but as for now, it's working and that's a lot :)

And yes, let is the traditional one -- with tons of parens everywhere.

-----

2 points by almkglor 6072 days ago | link

I'll go through the code later to see what can be done. Certainly the AST looks representable as plain lists to me, although I haven't fully analyzed it yet.

As an aside compile-file could be restructured like the following:

  (def compile-file (filename)
    (compile-ast (parse-file filename) (+ (strip-ext filename) ".c")))
  ; to allow programmatic access
  (def compile-ast (ast dest)
    ; chain of conversions
    (let chain
         (list
           (list cps-convert "CPS-CONVERSION")
           (list closure-convert "CLOSURE-CONVERSION"))
    ; do reduction
    (let final-ast
         (reduce
           (fn (ast (f desc))
             (let new-ast (f ast)
               (prn "----------------- AST after " desc)
              (prn (source new-ast))
               new-ast))
           chain ast)
        (prn "-------------------- C Code:")
        (w/outfile f dest
          (w/stdout f
            (prn:liststr:code-generate final-code))))))
This should allow easy insertion of any steps (e.g. optimization steps) along the way.

In fact the chain list should probably be better use `(,), so that we can support flags or suchlike for optimizations:

   `(
       (,cps-convert "CPS-CONVERSION")
       ,@(if optimize
             `((,optimize-after-cps "CPS-OPTIMIZATION")))
       (,closure-convert "CLOSURE-CONVERSION")
       ,@(if optimize
             `((,optimize-after-closure "CLOSURE-OPTIMIZATION"))))

-----

2 points by binx 6071 days ago | link

Just leave the peephole stuff to gcc, it almost always does better than handcoded optimizer.

The CPS transformed code can be arbitrarily inlined, so a simple inliner without flow analysis can give you much efficiency for free.

-----

2 points by almkglor 6070 days ago | link

And if the target isn't gcc?

For that matter my concern is with the expansion of PUSH and POP:

   PUSH(x); y = POP();

   =>

   *sp++ = x; y = *--sp;
Can gcc peephole the above, completely eliminating the assignment to the stack (which is unnecessary in our case after all)?

   y = x; //desired target
Without somehow informing gcc to the contrary, gcc will assume that writing to * sp is significant, even though in our "higher-level" view the assignment to * sp is secondary to transferring the data between two locations.

-----

4 points by sacado 6070 days ago | link

Actually, I tried the above (tuning generated code so as to change something like :

  BEGIN_JUMP(3); PUSH(LOCAL(5)); PUSH(LOCAL(6)); PUSH(LOCAL(7)); END_JUMP(3);
to its semantic but obviously much faster equivalent :

  memcpy (stack, stack + 5, sizeof(obj) * 3); sp = stack + 3; END_JUMP(3);
and

  PUSH(x); if(POP)
to

  if (x)
Well, with full optimizations on gcc (-O3), it doesn't change anything (at least in execution time, I didn't compare generated machine codes). Wow, gcc seems really clever. Now that I know how hard it is to implement a compiler, I can only applaud :)

-----

1 point by almkglor 6070 days ago | link

WOW. gcc must be real good then ^^.

-----