Le dim. 29 mars 2020 à 15:31, Thiede, Christoph < Christoph.Thiede@student.hpi.uni-potsdam.de> a écrit :
Hey Nicolas,
Seriously, did the Decompiler ever reliably produce re-generatable
parse trees in the past? But it should do so, shouldn't it? :-)
Maybe it did (see below). But I'm not sure that is was a feature... Isn't it mostly used for replacing absent source code... that will
eventually be repasrsed ? (!)
Well, it may be disputable whether decompiled trees should be optimized, but returning trees from anywhere that do not satisfy particular validity conditions (such as index being only set iff a special selector should be encoded) definitively appears wrong and buggy to me.
As you see, index i passed as argument to #code: keyword (? it's
because it's documenting the output, not the input); then code: parameter shadowing the index instance variable...
So would you agree to patch DecompilerConstructor >> #codeAnyLiteral:, too? :)
After parsing, there are other compilation phases, for analyzing
variable scope, clean blocks, etc...
Hm ... the Compiler divides the compilation phase into two main stages (see #evaluateCue:ifFail:): The first stage is actual "compilation", that is translating the source into a parse tree in the parser. The second stage is to generate a compiled method, which is done by simply passing #generate(WithTempNames) to the parse tree. For me, this appears to be a good logical separation. Things like scope analysis are done, as you say, in the second stage, of course. But I would not expect that optimizations such as special selectors are already applied in the first stage (this was also kind of confusing when I tried to debug certain optimizations such as of #caseOf:). Isn't it the general idea of a parse tree to have an intermediary representation between a primitive code string and a VM-specific set of bytecodes? Certain optimizations are not even relevant for other parser clients, for example, any code analysis tools. It would be great to decouple these stages even more - let's say, we can move all the #noteSpecialSelector: and #transform: senders and apply it directly before, or inside of #generate, only. A visitor sounds like a good pattern for this, I did not yet have many opportunities to apply this pattern in practice.
Pharo team did a complete re-engineering of compiler (OpalCompiler)
that you culd study.
Wow, I read some https://de.slideshare.net/jressia/opal-compiler slides https://de.slideshare.net/jressia/opal-compiler about OpalCompiler and it sounds great! Allow me one question, why didn't we already adapt this concept in Squeak, what are the disadvantages of this redesign? We could achieve so much more if everyone was pulling in the same direction (I know that it was the Pharo people to fork Squeak, however ...).
One disadvantage is that it's about twice slower. The is because byte code instructions are reified. Thus instead of source -> parse tree -> compiledMethod the flow is source -> parse tree -> instructions -> compiledMethod
It would be possible to make the instructions intermediate representation optional though.
The second disadvantage is that, IMO, it's a bit over engineered. One consequence is that patching the compiler for accepting methods > 15 arguments (required for Smallapack), or for accepting legacy FFI syntax took me more efforts than patching the legacy squeak compiler.
However, it's a nice piece of code. It should not be too hard to port to Squeak. Though, it relies on revamped parse tree nodes (those of refactoring browser, with slight evolutions...).
Best,
Christoph
*Von:* Squeak-dev squeak-dev-bounces@lists.squeakfoundation.org im Auftrag von Nicolas Cellier nicolas.cellier.aka.nice@gmail.com *Gesendet:* Samstag, 28. März 2020 14:09:15 *An:* The general-purpose Squeak developers list *Betreff:* Re: [squeak-dev] Decompiler buggy (was: AW: [Etoys, Compiler] Help wanted: Trying to embed SyntaxMorphs into other tiles)
Hi Christoph,
Le sam. 28 mars 2020 à 01:12, Thiede, Christoph < Christoph.Thiede@student.hpi.uni-potsdam.de> a écrit :
Hi Eliot, hi all,
ah, I finally found the bug, but this was a really hard hunt! :D
The solution is absolutely simple, again:
codeAnySelector: selector
^SelectorNode new
key: selector
- index: nil
- index: 0
type: SendType
Good find!
Seriously, did the Decompiler ever reliably produce re-generatable parse trees in the past? But it should do so, shouldn't it? :-)
Maybe it did (see below). But I'm not sure that is was a feature...
Isn't it mostly used for replacing absent source code... that will eventually be repasrsed ? (!)
Before the above patch, the following example was broken, too:
class := Object newSubclass. class compile: 'foo ^ 1 + 1'. (class >> #foo) decompile generate valueWithReceiver: class new arguments: #(). "SmallInteger does not understand #foo"
Now I'm wondering what are the actual semantics of the index variable. Its method comment about "various uses depending on the class of the receiver" is quite generic - do you know some more details about this? Should we also use nil instead of 0 in DecompilerConstructor >> # codeAnyLiteral:? At first glance, senders of #encodeLiteral: do not appear to set it to zero manually (so they leave it nil), but unless there is any documentation of the index meaning, this is speculation only, as I could not find any other example where decompilation + regeneration produce a method that cannot be executed properly.
It's very low level, some kind of reflexion of byteCode encoding.
Once upon a time (< Squeak4.0), the code was even more horrible to follow!
LeafNode>>key: object index: i type: type self key: object code: (self code: i type: type)
LeafNode>>code: index type: type index isNil ifTrue: [^type negated]. (CodeLimits at: type) > index ifTrue: [^(CodeBases at: type) + index]. ^type * 256 + index
As you see, index i passed as argument to #code: keyword (? it's because it's documenting the output, not the input); then code: parameter shadowing the index instance variable... And the index instance variable was not set... Kind of brainfuck.
We still have code:type: and index variable shadowing in current trunk...
By the way, here is another interesting one-liner:
(Object newSubclass environment: self environment; compile: 'foo ^(ObjectTracer on: nil) class'; >> #foo) decompile generate valueWithReceiver: nil arguments: #()
Interestingly, it opens a debugger - in other words, #class is sent as a regular selector. The decompiler does not know anything about special selectors at the moment. Is this desired behavior? I wonder whether it should be the parse tree's responsibility to install such kind of optimizations, rather than the responsibility of the Compiler. Because in reality, Compiler is not the only client that requests code generation from parse trees. Etoys is a good example for a client from another domain that uses this service, too. Should all these other clients be withheld these important optimizations of Smalltalk expressions?
After parsing, there are other compilation phases, for analyzing variable
scope, clean blocks, etc... It's possible to scatter the implementation of various phases in the nodes themselves, but the trend is rather to use a visitor pattern; it gather the handling in some specialized classes that hold all the states (rather than pass them as message arguments). Pharo team did a complete re-engineering of compiler (OpalCompiler) that you culd study.
Best,
Christoph
*Von:* Thiede, Christoph *Gesendet:* Freitag, 27. März 2020 23:16 Uhr *An:* The general-purpose Squeak developers list *Betreff:* AW: [squeak-dev] [Etoys, Compiler] Help wanted: Trying to embed SyntaxMorphs into other tiles
Hi Eliot,
It looks correct. Can you check it against the old bytecode set too?
We don’t want it to break old-style blocks.
Good point. I ran
(Object >> #asOrderedCollection) decompile generate valueWithReceiver: 42 arguments: #().
for both bytecode sets, and both were fine.
But:
(Collection >> #asArray) decompile generate valueWithReceiver: {42} asOrderedCollection arguments: #().
breaks - in both bytecode sets. This is weird. I will have a look into it, maybe I can discover what's wrong.
In addition, I propose to write tests for this. But it's not the goal of the decompiler to yield exactly the same parse tree or source code as the original method consisted of? In this case, we will need to write a lot of fixtures for the tests.
Best, Christoph
*Von:* Squeak-dev squeak-dev-bounces@lists.squeakfoundation.org im Auftrag von Eliot Miranda eliot.miranda@gmail.com *Gesendet:* Freitag, 27. März 2020 21:33 Uhr *An:* The general-purpose Squeak developers list *Betreff:* Re: [squeak-dev] [Etoys, Compiler] Help wanted: Trying to embed SyntaxMorphs into other tiles
Hi Christoph,
On Mar 27, 2020, at 12:45 PM, Thiede, Christoph < Christoph.Thiede@student.hpi.uni-potsdam.de> wrote:
Hi all! :-)
Just an update of the decompilation question:
Christoph Thiede wrote I don't know how to use #generate: exactly, but other senders usually appear to recompile a method before passing it to #generate:. For comparison:
[ (Collection >> #asArray) decompile generate: CompiledMethodTrailer empty ] fails, but
[ m := (Collection >> #asArray) decompile.
m := Compiler new compile: m in: Collection notifying: nil ifFail: #foo. m generate: CompiledMethodTrailer empty ] works. Why is that recompilation required but decompilation is insufficient? Is this some bug, or is it expected behavior?
The general approach seems to be correct, but I think I found an error in the decompilation of literal variables such as Array. I sent Compiler-ct.425 to the inbox which should fix this issue.
I moved this to inbox. It looks correct. Can you check it against the old bytecode set too? We don’t want it to break old-style blocks.
I am going to complete the implementation of SyntaxMorph >> #parseNode :-)
Best, Christoph
*Von:* Squeak-dev squeak-dev-bounces@lists.squeakfoundation.org im Auftrag von Thiede, Christoph *Gesendet:* Dienstag, 15. Oktober 2019 21:08:24 *An:* squeak-dev@lists.squeakfoundation.org *Betreff:* [squeak-dev] [Etoys, Compiler] Help wanted: Trying to embed SyntaxMorphs into other tiles
Hi all,
I'm currently trying to implement #parseNodeWith: on SyntaxMorph, in order to embed SyntaxMorphs into regular tiles. (Did this ever work in past?)
I'm afraid the attempt in the commit below does not work yet; you can create a script editor, but parsing is erroneous, so you cannot execute the script.
To reproduce:
Compile the following:
MyPlayer >> examplePlayerCode
self forward: 6 * 7.
self turn: (11 raisedTo: 13 modulo: 97)
and evaluate:
| e p | p := Morph new openInWorld assuredPlayer. e := (MyPlayer >> #examplePlayerCode) decompile asScriptEditorFor: p. e openInHand.
In Player>>#acceptScript:for:, #generate: is called on node, and when I decompile the result, I get a strange result:
examplePlayerCodeTest
self forward: 6 * 7.
self
forward: (#forward: forward: #forward:).
I don't know how to use #generate: exactly, but other senders usually appear to recompile a method before passing it to #generate:.
For comparison:
[ (Collection >> #asArray) decompile generate: CompiledMethodTrailer empty ] fails, but
[ m := (Collection >> #asArray) decompile. m := Compiler new compile: m in: Collection notifying: nil ifFail: #foo. m generate: CompiledMethodTrailer empty ] works.
Why is that recompilation required but decompilation is insufficient? Is this some bug, or is it expected behavior?
However, in the case of SyntaxMorph, I don't know how to recompile the node before, as a SyntaxMorph should be able to represent a node of an arbitrary type that must not be constrained to a MessageNode. So how could I solve the problem to generate code from SyntaxMorphs?
tl;dr: What is the full story of #generate: and how can it be made to work in this example?
Many thanks in advance! :-)
Best,
Christoph
*Von:* Squeak-dev squeak-dev-bounces@lists.squeakfoundation.org im Auftrag von commits@source.squeak.org commits@source.squeak.org *Gesendet:* Dienstag, 15. Oktober 2019 14:46 Uhr *An:* squeak-dev@lists.squeakfoundation.org *Betreff:* [squeak-dev] The Inbox: EToys-ct.367.mcz
A new version of EToys was added to project The Inbox: http://source.squeak.org/inbox/EToys-ct.367.mcz
==================== Summary ====================
Name: EToys-ct.367 Author: ct Time: 15 October 2019, 2:46:24.862129 pm UUID: 1394344f-b1e3-5640-a13a-70c5dffd51f4 Ancestors: EToys-mt.361
Allow for embedding SyntaxMorphs into test tiles.
=============== Diff against EToys-mt.361 ===============
Item was added:
- ----- Method: SyntaxMorph>>parseNodeWith:asStatement: (in category
'*Etoys-Squeakland-code generation') -----
- parseNodeWith: encoder asStatement: aBoolean
^ self parseNode!