whalesong/NOTES

Some possible optimizations with application:

    If any of the operands are constant (either by being variable
    lookups or literal constants), and if all of them are side-effect
    free, then juggle-operands might not be necessary.  I think this
    is similar to the "reorder" optimization described in casey's
    paper.
    
    In a self-application, it's not necessary to compute the operator,
    since the value is in the top control frame.  A parameterization
    can maintain the current lam in the top of the control frame.
    Given that, then there's no need to juggle operands either, since
    we can grab the operator afterwards and put it in place.

    For a kernel primitive call, if all of the operands are all
    constant, stack references, or kernel primitive calls, then
    there's no need to push for fresh stack space.




----------------------------------------------------------------------


Multiple values

There's interplay between compile-proc-appl and the linkage compiling
functions compile-linkage and compile-application-linkage.  When we
deal with multiple values, we'll have to do something here to make the
values efficient.  There's a paper by J. Michael Ashley and R. Kent
Dybvig called "An Efficient Implementation of Multiple Return Values
in Scheme" that I'll need to read.
http://citeseerx.ist.psu.edu/viewdoc/download?doi=10.1.1.39.1668&rep=rep1&type=pdf


Basic idea: each return address is actually a pair, where the
secondary address lies at a fixed offset of the first and handles
multiple value return.  Multiple values are returned back by keeping
them on the stack, and assigning argcount to the number of the
returned values.


In the context of my compiler: the compiler implicitly defines a
singleton, statement context by using next-linkage.  But some uses of
next-linkage ignore the number of values that come back, and others
should raise an error.  Here are the contexts that care:

    app
    let1
    install-value
    toplevel-set (define-values, assign)


For the contexts that don't care, we need to set up a return address
that just pops those values off.






Before introducing the multiple-value jumps
(172b1d9e5de823b53a6705fc87babfdd61152924), test-conform-browser
reports the following times:

fermi ~/work/whalesong $ racket test-conform-browser.rkt 
running test... ok (5248 milliseconds)
fermi ~/work/whalesong $ racket test-conform-browser.rkt 
running test... ok (5478 milliseconds)
fermi ~/work/whalesong $ racket test-conform-browser.rkt 
running test... ok (5501 milliseconds)
fermi ~/work/whalesong $ racket test-conform-browser.rkt 
running test... ok (5853 milliseconds)
fermi ~/work/whalesong $ racket test-conform-browser.rkt 
running test... ok (5532 milliseconds)
fermi ~/work/whalesong $ racket test-conform-browser.rkt 
running test... ok (5498 milliseconds)
fermi ~/work/whalesong $ racket test-conform-browser.rkt 
running test... ok (5351 milliseconds)
fermi ~/work/whalesong $ racket test-conform-browser.rkt 
running test... ok (5464 milliseconds)
fermi ~/work/whalesong $ racket test-conform-browser.rkt 
running test... ok (5545 milliseconds)
fermi ~/work/whalesong $ racket test-conform-browser.rkt 
running test... ok (5405 milliseconds)


After introducing the mutiple value jumps targets
(cc1c156df79bab09ca37164e75ae0afe0ac1b0d0), test-conform-browser is
reporting the following times:


running test... ok (5281 milliseconds)
fermi ~/work/whalesong $ racket test-conform-browser.rkt 
running test... ok (5554 milliseconds)
fermi ~/work/whalesong $ racket test-conform-browser.rkt 
running test... ok (5588 milliseconds)
fermi ~/work/whalesong $ racket test-conform-browser.rkt 
running test... ok (5509 milliseconds)
fermi ~/work/whalesong $ racket test-conform-browser.rkt 
running test... ok (5428 milliseconds)
fermi ~/work/whalesong $ racket test-conform-browser.rkt 
running test... ok (5387 milliseconds)
fermi ~/work/whalesong $ racket test-conform-browser.rkt 
running test... ok (5539 milliseconds)
fermi ~/work/whalesong $ racket test-conform-browser.rkt 
running test... ok (5355 milliseconds)
fermi ~/work/whalesong $ racket test-conform-browser.rkt 
running test... ok (5551 milliseconds)
fermi ~/work/whalesong $ racket test-conform-browser.rkt 
running test... ok (5331 milliseconds)



At a rough glance, I see no appreciable extra cost for this program,
since it doesn't use multiple-value-return.  Thankfully, it looks like
the JIT in JavaScript isn't significantly hurt when we set the
attribute to the procedure.




What's left to do:

    forms for using the values coming from multiple value returns
    (with-values, define-values, let-values)

    runtime error traps for contexts that must not receive multiple values.

    fixing apply definition so it doesn't return multiple values when
    given a single argument.


\\\\\\\\\\\\\\\\\\\\\\\\\\\\\\\\\\\\\\\\\\\\\\\\\\\\\\\\\\\\\\\\\\\\\\








----------------------------------------------------------------------




Open coding:

I want to be able to write the definitions of kernel primitives once,
and reuse those definitions for both the open-coding as well as the
real runtime.  I also need to be able to encode the type checks.  I
want to be able to say:


(make-kernel-primitive '+
                       (arity 0 #t)

                       (lambda (args)
			 (values (mapi (lambda (arg i)
					 (test arg i number?))
				       arg)
				 (string-join args "+"))))

and have it magically generate the definitions for the open-coding
primitive as well as:

    PRIMITIVES["+"] = function(MACHINE, arity) {
                          var result = 0;
                          for (var i = 0 ; i < arity; i++) {
                              test(isNumber(MACHINE.env[MACHINE.env.length - 1 - i]),
                                   i,
                                   "number");
                              result += MACHINE.env[MACHINE.env.length - 1 - i];
			  }
                          return result;
                      };

Is this completely unrealistic?  I have to see how Rabbit and Orbit do this.




----------------------------------------------------------------------

Runtime values and types are in in the plt.runtime namespace.  I need
to move types from WeScheme into here.


----------------------------------------------------------------------


Frames and environments.


A CallFrame consists of:

   A return address back to the caller.
   A procedure (the callee).
   A stack.
   A set of continuation marks.


A PromptFrame consists of:

   A return address back to the caller.
   A tag.
   A set of continuation marks.



On exit from a CallFrame,

    MACHINE.env = frame.env



On a regular, generic function call:

    The operator and operands are computed and placed in MACHINE.env's
    scratch space.

    A new call frame is constructed.  The frame remembers the environment.
    
    The machine jumps into the procedure entry.


On a tail call,

    The operator and operands are computed and placed in MACHINE.env's
    scratch space.

    The existing call frame is reused.
        The frame's environment consumes those elements from MACHINE.env
        MACHINE.env = the new stack segment





Optimizations with IL

The sequence PushEnvironment ... AssignImmediateStatement (EnvLexicalAddress ...)
where we're assigning directly to a spot we just allocated, can be reduced to
a single instruction.

We can do some constant folding in operands.  e.g.

    MACHINE.env[MACHINE.env.length - 1 - 3] = MACHINE.env[MACHINE.env.length - 1 - 7];

=>

    MACHINE.env[MACHINE.env.length - 4] = MACHINE.env[MACHINE.env.length - 8];






On tail calls, when we're reusing all of the arguments on the stack,
there's no need to splice, since we won't be popping anything off:

   MACHINE.env.splice(MACHINE.env.length - (MACHINE.argcount + ((10) - MACHINE.argcount)), ((10) - MACHINE.argcount));

is a no-op.








In the case where a closure has a prefix, but all the uses of the prefix are to open-coded primitives, then we don't need to close over it after all.  e.g.

    (test '(begin (letrec ([f (lambda (x) (* x x))]
                           [g (lambda (x) (* x x x))])
                    (- (g (f (+ (g 3) (f 3)))) 1)))
          2176782335
          #:debug? #t)

since (* -) are both open-coded, there's no need to capture the
prefix, and we can reduce some allocation.




I can eliminate the first instruction in the pair:


    #(struct:AssignPrimOpStatement val #(struct:GetCompiledProcedureEntry))
    #(struct:GotoStatement #(struct:Label lamEntry259))


since the val isn't even being used here...  This is the case when we
statically know the lambda target.


    - this is done now.



I can coalese

     (PushEnvironment 1 #f)
     (AssignPrimOpStatement (EnvLexicalReference 0 #f) (MakeCompiledProcedure 'lamEntry265 1 '(2 1) 'diff)

into a single statement.





If lambdas don't escape, then we can make their closures empty by
simply explicitly passing in the free arguments.





There's no good reason why the IL has both AssignImmediateStatement
and AssignPrimOpStatement.  The distinction is artificial because I'm
allowing the RHS of assignments to use arbitrary expressions, since my
runtime (JavaScript) supports it.  I should consolidate these
structures; it may allow me to remove a few more instructions (like
setting ControlLabel to 'val).







flush-output must immediately yield control to the browser, because
the browser needs control back to display changes to the dom.
Basically, we're simulating an IO interrupt here...







April 17,  2011

The dynamic recomputation for gas is only controlling one parameter:
how many times to run the trampoline before bouncing off to the
browser.  But we really have two parameters that need dynamic
computation

    *  FN: the number of function calls before invoking the trampoline.

        FN is necessarily bounded above by the browser.  The larger it
        is, the more efficient the trampoline can be.

    *  TI: the number of trampoline invokations before yielding to the browser

Both of these should be under some dynamic controller.  We want to
optimize the efficiency of the runtime.  I don't know what the
function is, but we want to optimize the parameters FN and TI such
that it maximizes FN and minimizes TI, and yet gives us the browser
reactivity we want.





April 24, 2011

The variables for linkage and target are doing double duty, which is
showing up in the defintion for compilation, since there are cases
that shouldn't exist in there.

They really should be part of the same datatype which describes,
essentially, what the code's continuation should be doing next.
Target's describing where the value needs to be installed at the end
of this, and linkage describes how to jump into the continuation.


   Return --- write value to val, pop off and jump according to
   dynamic value on control context.  Return context may be in tail
   position or not.

   Next --- write value to a particular target and continue on.

   Label --- write value to a particular target and jump
   unconditionally to labeled location.



The continuation may or may not be expecting multiple values.

   Ignore: doesn't care how many values come back.  Throw away values
   if multiple values are passed in.

   Any: receives multiple values, and ensures those values are on the
   stack.  If a single value is received, pushes it on the stack and
   sets up argcount to 1.

   N: must receive exactly N values.  If there's a mismatch, raises a
   runtime error.



Return will allow Any number of values to come back.  It doesn't need
a separate multiple-value context.

Next expects either exactly 1 value to come back, or ignores.  So it
needs an multiple-value context.

Label, too, expects exactly 1 value to come back, or ignores.  So it
needs a mulitple-value context.


When we use apply-values, it'll compile the producer expression in an
Any context.



I'm going to simplify values a bit.



----------------------------------------------------------------------

April 28


Multiple values are handled in the following way now.

In a context that expects multiple values to be returned,

   if n = 0, don't leave anything on the stack before jumping out

   if n = 1, put the single value in the 'val register

   if n > 1, put the first value in the 'val register, and leave the
   rest (the n-1 values) on the value stack.

The context is then responsible for dealing with those multiple return
values.

The contexts are now of the following types:

   'tail : keeps the values on the stack.  Used specifically for tail return.

   'drop-multiple : drops any extra values on the stack

   'keep-multiple : keeps any number of values on the stack.

   Natural : expects exactly n values.  Errors out if this can't be the case.



There appears to be a bug in compile-splice regarding multiple value
contexts.  I haven't yet fixed the bug.  I need a test case.  I need
to somehow create a splicing expression in the context of something
that expects multiple values back.  I'm not exactly sure how to create
such a context.


Ok, I think I've been able to do this successfully.  I lifted out the
code for emit-values-context-check-on-procedure-return so it's used
for both the returns from procedure call, as well as the calls from
the prompt splicing.


----------------------------------------------------------------------


May 13, 2011

begin0 is still broken; I'm a bit unsatisfied with the way that it's
coded, and I know it's not correct yet.

The other thing that's I'm sure needs to be looked at again are the
implementation of prompts.


modules are almost in place.  Needs a mechanism for loading modules on
the fly on the network, as well as some kind of predicable namespacing
mechanism.  I think the compiler will need to include something like a

    (current-module-name-canonizer)

which takes module names (symbol, path-string) and systematically
translates them to predictable identifiers.  Anything refering to a
collection should be translated to

    collects/...

Anything outside that should be given a name relative to some root.
One should be able to say:

    root the translation at "/home/dyoo/work/js-sicp-5.5/examples"

where all translated paths are either from collections, or reachable
from the root.  That way, we get predictable paths.




js-sicp-5-5 is an uninspired name for the project.  I'm renaming it to
"Whalesong".  Whale, because it's related to the Moby-Scheme project,
and song because, well, some songs can be called a "Racket".  :)



I want to use Google Closure's compiler to minify.  Closure doesn't
like some of the code, so I'll need to rename some of the identifiers
in types.js to avoid colliding with Java keywords (char, float).


----------------------------------------------------------------------



I need to be careful about targetting the right target in recursive calls to compile.
The bug I was chasing down involved using (compile subexpr target cenv), where the
target was that of the parent expression.  But this is an invalid thing to do.

The other bug I nailed down yesterday was the shared-gensym bug with
lambda entry points.  I have a single module now that distributes
gensyms for the lambda entry points.  At some point, I'd like to make
the names descriptive so it's easier to figure out the functions by
low-level inspection.



Optimizations are off at the moment until I integrate this into Moby.
Then I can start turning optimizations back on.

----------------------------------------------------------------------



May 20, 2011

I'm running my bytecode parser over the entire racket collects tree,
just to make sure the parser itself is robust.

Parsing takes milliseconds, except on Typed Racket code, which is expected.