New compiler pushed to master

Hi guys!

I finally pushed the new compiler in master, it is now used by default.
This compiler offers several improvements compared to the old one:

- Object-oriented
- Semantic analysis
- Intermediate Representation (flatten, no specific semantic aspects of
  Smalltalk, etc) that is easy to optimize
- Properly tested (although some more tests are needed)
- Smarter inlining

The new compiler toolchain compiles from source code in 5 steps (the
inlining is optional):

1. AST
2. Semantic analysis (AST annotation + validation)
3. IR
4. IR with inlining
5. JS output

The semantic analysis gives us interesting informations:
- variable scope
- variable shadowing
- block/method scopes
- unknown variables

as well as semantic validation.

The Intermediate Representation is much simpler and closer to JS
compared to the Smalltalk AST, allowing us to perform better inlining:

- Inlined blocks (ifTrue: ifNil:, etc)
- inlined non local returns
- move assignments, returns, etc. when required

Inlining examples:

The new compiler can perform inlinings like the following (note inlined
returns, etc):

true ifTrue: [ ^ 1 ]

if(smalltalk.assert(true) {
  return 1;
}

--

a := b ifTrue: [ 2 ]

if(smalltalk.assert(b) {
  a = 2;
}

--

^ b ifFalse: [ self ]

if(!smalltalk.assert(b) {
  return self;
}

I would like to thank Herby for his work on inlining that I
integrated here (the implementation turns out to be different though).

--
Nicolas Petton
http://nicolas-petton.fr

Nicolas Petton wrote:
I should get accustomed to it... when I saw #visitInlinedIfTrue: or
something like that, I facepalmed and stop to watch it. It just seemed
unbearable to me to include visitor for every inlined selector... but I
may have misunderstood something.

I think this misses 'else { a = nil; }' ...
>
> --
>
> ^ b ifFalse: [ self ]
>
> if(!smalltalk.assert(b) {
>    return self;
> }

...  and this misses 'else { return nil; }'. I hope it is just the error
written in this mail and the code in fact generates the missing branches.

> I would like to thank Herby for his work on inlining that I
> integrated here (the implementation turns out to be different though).

Thanks.

>

Herby

On Wed, 26 Sep 2012 14:41:45 +0200, Herby Vojčík <[hidden email]> wrote:
> I should get accustomed to it... when I saw #visitInlinedIfTrue: or
> something like that, I facepalmed and stop to watch it. It just seemed
> unbearable to me to include visitor for every inlined selector... but I
> may have misunderstood something.

Well, we can talk about it of course, but isn't this better than
manipulating strings around with a stateful & error prone  do-it-all compiler?


> > ^ b ifFalse: [ self ]
> >
> > if(!smalltalk.assert(b) {
> >    return self;
> > }
>
> ...  and this misses 'else { return nil; }'. I hope it is just the error
> written in this mail and the code in fact generates the missing branches.

Well, yes :) This is just pseudo code written to explain the idea, not
actual code generated by the compiler.

Nico

--
Nicolas Petton
http://nicolas-petton.fr

Nicolas Petton wrote:
> On Wed, 26 Sep 2012 14:41:45 +0200, Herby Vojčík<[hidden email]>  wrote:
>> I should get accustomed to it... when I saw #visitInlinedIfTrue: or
>> something like that, I facepalmed and stop to watch it. It just seemed
>> unbearable to me to include visitor for every inlined selector... but I
>> may have misunderstood something.
>
> Well, we can talk about it of course, but isn't this better than
> manipulating strings around with a stateful&  error prone  do-it-all compiler?

Definitely, I just think it went a bit overboard, but I may be wrong,
since I did not thoroughly study the solution.

Ok. Thanks that it is in master already.

Is the whole Amber rebuilt (or tried if it works when it is completely
rebuilt)?

> Nico

Herby

On Wed, 26 Sep 2012 15:21:06 +0200, Herby Vojčík <[hidden email]> wrote:
>
>
> Nicolas Petton wrote:
> Is the whole Amber rebuilt (or tried if it works when it is completely
> rebuilt)?

Yep, it's completely rebuilt.

Nico

--
Nicolas Petton
http://nicolas-petton.fr