[squeak-dev] Boolean over-optimization #ifTrue:/#ifFalse:

> Igor Stasenko wrote:
>  > Hello,
>  >
>  > i currently tried to implement a proxy (subclass of ProtoObject) with
>  > own IfTrue:/ifFalse: methods.
>  > What i can't understand, is why VM throws me with #mustBeBoolean
>  > message instead of doing real send if optimized pattern fails (when
>  > receiver of #IfTrue: is a non-boolean)?
>
>
> Because you cannot distinguish #ifTrue:#ifFalse: sends from #whileTrue:
>  for example, and because there are no BlockC{ontext,losure} objects for
>  the argument blocks.
>

>
>  This should also work in Squeak.
>
>
>  >  If yes, i'd like to see optimizations optional, so people can do like:
>  >
>  >  Compiler turnOffOptimizations.
>
>
> That will not only kill your performance ...
>

> On 05/03/2008, Paolo Bonzini <[hidden email]> wrote:
>> Igor Stasenko wrote:
>>> Hello,
>>>
>>> i currently tried to implement a proxy (subclass of ProtoObject)  
>>> with
>>> own IfTrue:/ifFalse: methods.
>>> What i can't understand, is why VM throws me with #mustBeBoolean
>>> message instead of doing real send if optimized pattern fails (when
>>> receiver of #IfTrue: is a non-boolean)?
>>
>>
>> Because you cannot distinguish #ifTrue:#ifFalse: sends from  
>> #whileTrue:
>> for example, and because there are no BlockC{ontext,losure} objects  
>> for
>> the argument blocks.
>>
>
> So, the only way is to compile method without optimization of  
> #ifTrue #ifFalse:,
> to make sure they do real sends?
>
> If yes, i'd like to see optimizations optional, so people can do like:
>
> Compiler turnOffOptimizations.
> MyClass compileAll.
> Compiler turnOnOptimizations.
>
>>
>> Paolo
>>
>>
>
>
> --
> Best regards,
> Igor Stasenko AKA sig.
>

> On 05/03/2008, Paolo Bonzini wrote:
>> Igor Stasenko wrote:
>>  > Hello,
>>  >
>>  > i currently tried to implement a proxy (subclass of ProtoObject) with
>>  > own IfTrue:/ifFalse: methods.
>>  > What i can't understand, is why VM throws me with #mustBeBoolean
>>  > message instead of doing real send if optimized pattern fails (when
>>  > receiver of #IfTrue: is a non-boolean)?
>>
>> Because you cannot distinguish #ifTrue:#ifFalse: sends from #whileTrue:
>>  for example, and because there are no BlockC{ontext,losure} objects for
>>  the argument blocks.
>>
>
> So, the only way is to compile method without optimization of #ifTrue  
> #ifFalse:, to make sure they do real sends?
>
> If yes, i'd like to see optimizations optional,

> so people can do like:
>
> Compiler turnOffOptimizations.
> MyClass compileAll.
> Compiler turnOnOptimizations.
>
>>
>>  Paolo
>>
>>
>
>

> On Wed, 05 Mar 2008 07:39:33 +0100, Igor Stasenko wrote:
>
>  > On 05/03/2008, Paolo Bonzini wrote:
>  >> Igor Stasenko wrote:
>  >>  > Hello,
>  >>  >
>  >>  > i currently tried to implement a proxy (subclass of ProtoObject) with
>  >>  > own IfTrue:/ifFalse: methods.
>  >>  > What i can't understand, is why VM throws me with #mustBeBoolean
>  >>  > message instead of doing real send if optimized pattern fails (when
>  >>  > receiver of #IfTrue: is a non-boolean)?
>  >>
>  >> Because you cannot distinguish #ifTrue:#ifFalse: sends from #whileTrue:
>  >>  for example, and because there are no BlockC{ontext,losure} objects for
>  >>  the argument blocks.
>  >>
>  >
>  > So, the only way is to compile method without optimization of #ifTrue
>  > #ifFalse:, to make sure they do real sends?
>  >
>  > If yes, i'd like to see optimizations optional,
>
>  For Rob Wither's project I've made a small patch
>  NoMacroCompiler-kwl.1dot5.cs which allows to specify per method (in a
>  pragma) or system-wide that macros are inlined or result in a full send.
>  Could be adapted to select what kind of macro not/to optimize. I've also
>  identified which Squeak .image methods *must* be compiled *with* macro
>  inlineing so that the .image does not crash/get stuck. And out of
>  meta-programming/memory reasons, the implementation of #whileTrue*,
>  #whileFalse* must be compiled *with* inlineing ;-)
>

>> Yes, let me take a look at it.
> i'm already using pragma <metacode> to switch between different
> simulation modes (but it's in my own domain, not related to optimized
> booleans).
>
> I'm currently doing some research, by implementing own object memory
> with a small set of classes. To make things done fast, i decided to
> not create parsers/bytecode interpreter, but instead, use squeak as
> backend for evaluating a code using proxies.
>
> My intent is to create a kind of self-hosted VM, which after some
> bootstrap can start working without the need of implementing things
> externally, like primitives/bytecode interpreter.
> A 'primitives' in this system will be replaced by a special methods,
> marked with <metacode> pragma, which should follow special rules and
> handled differently by compiler.
> But what is good, that these methods are 'live' - you can change their
> implementation at any time, in same way as any other method in
> smalltalk.
> As a first 'showcase' i like to make complete implementation of
> classes which can create objects and do smallint math without the need
> of primitives (means, all behavior implementation resides in object
> memory).
>
> To complete a full circle, the VM should be capable to compile methods
> to bare metal (CPU instructions) and can execute them without any
> external help.
>
>> /Klaus
>
> --
> Best regards,
> Igor Stasenko AKA sig.
>
>

> On Wed, 05 Mar 2008 10:09:53 +0100, Igor Stasenko wrote:
>
>
>  > On 05/03/2008, Klaus D. Witzel wrote:
>  >> On Wed, 05 Mar 2008 07:39:33 +0100, Igor Stasenko wrote:
>  >>  > On 05/03/2008, Paolo Bonzini wrote:
>  >>  >> Igor Stasenko wrote:
>  >>  >>  > Hello,
>  >>  >>  >
>  >>  >>  > i currently tried to implement a proxy (subclass of ProtoObject)
>  >>  >>  > with own IfTrue:/ifFalse: methods.
>  >>  >>  > What i can't understand, is why VM throws me with #mustBeBoolean
>  >>  >>  > message instead of doing real send if optimized pattern fails
>  >>  >>  > (when receiver of #IfTrue: is a non-boolean)?
>  >>  >>
>  >>  >> Because you cannot distinguish #ifTrue:#ifFalse: sends from
>  >> #whileTrue:
>  >>  >>  for example, and because there are no BlockC{ontext,losure}
>  >> objects for
>  >>  >>  the argument blocks.
>  >>  >
>  >>  > So, the only way is to compile method without optimization of #ifTrue
>  >>  > #ifFalse:, to make sure they do real sends?
>  >>  >
>  >>  > If yes, i'd like to see optimizations optional,
>  >>
>  >>  For Rob Wither's project I've made a small patch
>  >>  NoMacroCompiler-kwl.1dot5.cs which allows to specify per method (in a
>  >>  pragma) or system-wide that macros are inlined or result in a full
>  >> send.
>  >>  Could be adapted to select what kind of macro not/to optimize. I've
>  >> also
>  >>  identified which Squeak .image methods *must* be compiled *with* macro
>  >>  inlineing so that the .image does not crash/get stuck. And out of
>  >>  meta-programming/memory reasons, the implementation of #whileTrue*,
>  >>  #whileFalse* must be compiled *with* inlineing ;-)
>  >>
>  >
>  > The only place, where whileTrue:/whileFalse: should be _always_
>  > compiled optimized its in
>  > BlockContext/BlockClosure whileTrue:/whileFalse: methods.
>  > In rest of the code there is nothing bad in using real sends.
>
>
> Yes. And that's magically recognized by the attached.
>
>
>  >
>  >>  Let me know if that's interesting to you.
>  >
>  > Yes, let me take a look at it.
>
>
> Attached. Beware of
>
>  CommandHistory class forgetAllGrabCommandsFrom:
>  ContextPart handleSignal:
>  ContextPart restart
>  ContextPart resume:
>  Exception outer
>  ImageSegment allObjectsDo:
>  ImageSegment comeFullyUpOnReload:
>  ImageSegment rehashSets
>  ImageSegment restoreEndianness
>  ImageSegment copyFromRoots:sizeHint:areUnique:
>  ImageSegment install
>  ImageSegment readFromFile
>  PointerFinder class pointersTo:except:
>  Process debug:title:full:
>  Process debugWithTitle:
>  Process terminate
>  ProjectLoading class openName:stream:fromDirectory:withProjectView:
>  SystemNavigation allObjectsDo:
>  SystemProgressMorph nextSlotFor:
>  Utilities class pointersTo:except:
>
>  you *have* been warned. Besides that list, I found that the Squeak .image
>  can be run at the full message send level.
>

>
>  > i'm already using pragma <metacode> to switch between different
>  > simulation modes (but it's in my own domain, not related to optimized
>  > booleans).
>  >
>  > I'm currently doing some research, by implementing own object memory
>  > with a small set of classes. To make things done fast, i decided to
>  > not create parsers/bytecode interpreter, but instead, use squeak as
>  > backend for evaluating a code using proxies.
>
>
> So I don't have to develop that? Great, send me a copy please.
>

> On Mar 5, 2008, at 7:52 , Lukas Renggli wrote:
>
>  >> So, the only way is to compile method without optimization of
>  >> #ifTrue #ifFalse:,
>  >>  to make sure they do real sends?
>  >
>  > Just today in my blog-reader:
>  >
>  > http://www.cincomsmalltalk.com/userblogs/mls/blogView?
>  > showComments=true&printTitle=Boolean_Objects&entry=3382109971
>  >
>  > This should also work in Squeak.
>
>  Hehe, very cool. It works a bit differently in Squeak, where the
>  image code has to rewind the context before the jump bytecode that
>  recognized the non-boolean. Attached is a little changeset that adds
>  #asBoolean to various classes and overrides #mustBeBooleanIn: to use
>  that, provided objectsAsBooleans has been enabled. File in, then try
>  this:
>
>  Preferences enable: #objectsAsBooleans.
>  #(0 1 nil #() #(a b c) 'hi' '') collect: [:each |
>         each  -> (each ifTrue: ['yes'] ifFalse: ['no'])]
>
>  #(0->'no' 1->'yes' nil->'no' #()->'no' #(#a #b #c)->'yes' 'hi'->'yes'
>  ''->'no')
>
>  Fun indeed :)
>

> On 05/03/2008, Bert Freudenberg <[hidden email]> wrote:
>> On Mar 5, 2008, at 7:52 , Lukas Renggli wrote:
>>
>>  >> So, the only way is to compile method without optimization of
>>  >> #ifTrue #ifFalse:,
>>  >>  to make sure they do real sends?
>>  >
>>  > Just today in my blog-reader:
>>  >
>>  > http://www.cincomsmalltalk.com/userblogs/mls/blogView?
>>  > showComments=true&printTitle=Boolean_Objects&entry=3382109971
>>  >
>>  > This should also work in Squeak.
>>
>>  Hehe, very cool. It works a bit differently in Squeak, where the
>>  image code has to rewind the context before the jump bytecode that
>>  recognized the non-boolean. Attached is a little changeset that adds
>>  #asBoolean to various classes and overrides #mustBeBooleanIn: to use
>>  that, provided objectsAsBooleans has been enabled. File in, then try
>>  this:
>>
>>  Preferences enable: #objectsAsBooleans.
>>  #(0 1 nil #() #(a b c) 'hi' '') collect: [:each |
>>         each  -> (each ifTrue: ['yes'] ifFalse: ['no'])]
>>
>>  #(0->'no' 1->'yes' nil->'no' #()->'no' #(#a #b #c)->'yes' 'hi'->'yes'
>>  ''->'no')
>>
>>  Fun indeed :)
>>
>
> Yeah, rewinding a bytecode counter of parent context is a kind of
> woodo black magic. But hey! It works!

> The #compilerClass method allows you to specify the compiler to be
>  used for a class (and its subclasses).  So, one approach would be to
>  implement a subclass of Compiler (say, "ProxyCompiler") that disables
>  optimization of #ifTrue:ifFalse: etc.
>
>  I'm not sure about a clean way to disable optimizations in a subclass
>  of Compiler.  Here's a moderately horrible way to disable them
>  globally...
>         - give Compiler a boolean class variable "DisableSpecialSelectors"
>         - hack MethodNode>>noteSpecialSelector: thusly...
>
>  noteSpecialSelector: selectorSymbol
>         " special > 0 denotes specially treated messages. "
>
>         DisableSpecialSelectors ifTrue: [special := 0. ^self].
>
>         "Deconvert initial keywords from SQ2K"
>         special := #(:Test:Yes: :Test:No: :Test:Yes:No: :Test:No:Yes:

>                                 and: or:
>                                 :Until:do: :While:do: whileFalse whileTrue
>                                 :Repeat:to:do: :Repeat:to:by:do:
>                                 ) indexOf: selectorSymbol.
>         special > 0 ifTrue: [^ self].
>
>         special := MacroSelectors indexOf: selectorSymbol.
>
>
>  Then, you could override #translate:noPattern:ifFail: in
>  ProxyCompiler, as follows...
>
>  translate: aStream noPattern: noPattern ifFail: failBlock
>         [       DisableSpecialSelectors := true.
>                 ^super translate: aStream noPattern: noPattern ifFail: failBlock
>         ] ensure: [DisableSpecialSelectors := false].
>
>
>  Finally, implement #compilerClass on the class-side of your Proxy
>  class, and answer ProxyCompiler.
>
>  Heh, that was sorta fun!
>  Josh
>
>
>