[Newcompiler] How to decompile #while{True,False}

Klaus D. Witzel klaus.witzel at cobss.com
Wed Sep 5 16:00:00 UTC 2007 

Hi Math,

the only place I can see where one of {true. false} is not pushed during  
IR's work is #absorbConstantConditionalJumps: which is part of #optimize.

IRJumpIf etc must do their work so that #absorbConstantConditionalJumps:  
can really find things for optimizing them away.

Do you perhaps have a test for the cases analyzed during  
IRMethod>>#optimize, that would shed a bit more light on the issue.

Cheers
Klaus

On Wed, 05 Sep 2007 14:45:58 +0200, Mathieu Suen wrote:

> Hi Klaus
>
> Thanks
>
> On Sep 5, 2007, at 7:12 AM, Klaus D. Witzel wrote:
>
>> Hi Math,
>>
>> when the old decompiler decompiles #sweep: it finds
>
> He can find the #ifTrue:ifFalse: because he keep the push true or false.
>
>>
>> [(ind := ind + 1) > list size
>> 		ifTrue: [false]
>> 		ifFalse: [(parts at: ind) first asLowercase = projectName
>> 				and: [versions
>> 						add: (parts at: ind).
>> 					true]]] whileTrue.
>>
>> which is what can be expected, the inner #ifTrue:ifFalse: is rendered  
>> as an #and: (... ifFalse: [false]).
>>
>> The place where a #not message send is inserted is  
>> IRDecompiler>>#endAndOr2: (reducing the jump structures).
>
> Yes I know that but I don't know where the NewCompiler remove the  
> pushConstant: true.
> It should be somewhere where when the IRJumpIf is interpret but I cant  
> find it.
>
> The old compiler don't  remove the push true:
>
>> 199 <14> pushTemp: 4
>> 200 <76> pushConstant: 1
>> 201 <B0> send: +
>> 202 <81 44> storeIntoTemp: 4
>> 204 <12> pushTemp: 2
>> 205 <C2> send: size
>> 206 <B3> send: >
>> 207 <9A> jumpFalse: 211 "First ifTrue ifFalse"
>>
>> ''(ind := ind + 1) > list size  ifTrue:  [false]"
>> 208 <72> pushConstant: false
>> 209 <A4 15> jumpTo: 232
>>
>> ''(ind := ind + 1) > list size  ifFalse:  [...]"
>> 211 <13> pushTemp: 3
>> 212 <14> pushTemp: 4
>> 213 <C0> send: at:
>> 214 <83 15> send: first
>> 216 <83 14> send: asLowercase
>> 218 <16> pushTemp: 6
>> 219 <B6> send: =
>> 220 <AC 09> jumpFalse: 231   "Second ifTrue ifFalse"
>>
>> "(parts at: ind) first asLowercase = projectName ifTrue: [..]"
>> 222 <17> pushTemp: 7
>> 223 <13> pushTemp: 3
>> 224 <14> pushTemp: 4
>> 225 <C0> send: at:
>> 226 <83 37> send: add:
>> 228 <87> pop
>> 229 <71> pushConstant: true
>> 230 <90> jumpTo: 232
>>
>> "(parts at: ind) first asLowercase = projectName ifFalse: [false]"
>> 231 <72> pushConstant: false
>> 232 <99> jumpFalse: 235
>
> You can see the 208 229 and 231 bytecode. There are remove with the  
> NewCompiler.
> I don't think that is usefull to make dose optimization.
>
> 	Mth
>
>>
>> HTH.
>>
>> Cheers
>> Klaus
>>
>> On Wed, 05 Sep 2007 00:06:55 +0200, Mathieu Suen wrote:
>>
>>> Hi
>>>
>>> In a very particular case the compiler optimize the #whileTrue in a  
>>> way that the Decompiler can't find the roght code.
>>>
>>> This particular case is:
>>>
>>> [
>>> self somthings
>>>
>>> aTest
>>> 	ifTrue:[ok anOtherThings
>>> 		ifTrue:[here doSomthings. true]
>>> 		ifFalse:[false]]
>>> 	ifFalse:[false]
>>> ] whileTrue
>>>
>>> The typical exemple is from  Project class>>#sweep: :
>>>
>>>> 		…
>>>> 		[(ind _ ind + 1) > list size
>>>> 			ifFalse: [(parts at: ind) first asLowercase = projectName
>>>> 				ifTrue: [versions add: (parts at: ind).  true]
>>>> 				ifFalse: [false]]
>>>> 			ifTrue: [false]] whileTrue.
>>>> 		…
>>>
>>> The corresponding bytecode is:
>>>
>>>> 199 <14> pushTemp: 4
>>>> 200 <76> pushConstant: 1
>>>> 201 <B0> send: +
>>>> 202 <81 44> storeIntoTemp: 4
>>>> 204 <12> pushTemp: 2
>>>> 205 <C2> send: size
>>>> 206 <B3> send: >
>>>> 207 <A8 16> jumpTrue: 231
>>>>
>>>> 209 <13> pushTemp: 3
>>>> 210 <14> pushTemp: 4
>>>> 211 <C0> send: at:
>>>> 212 <83 14> send: first
>>>> 214 <83 15> send: asLowercase
>>>> 216 <16> pushTemp: 6
>>>> 217 <B6> send: =
>>>> 218 <AC 0B> jumpFalse: 231
>>>>
>>>> 220 <17> pushTemp: 7
>>>> 221 <13> pushTemp: 3
>>>> 222 <14> pushTemp: 4
>>>> 223 <C0> send: at:
>>>> 224 <83 37> send: add:
>>>> 226 <87> pop
>>>> 227 <71> pushConstant: true
>>>> 228 <99> jumpFalse: 231
>>>>
>>>> 229 <A3 E0> jumpTo: 199
>>>
>>>
>>> So the first #ifFlase:ifTrue: has been compact to a single jump:
>>> 207 <A8 16> jumpTrue: 231
>>>
>>> This avoid pushing false and make an unconditional jump to the end of  
>>> the nearly end of the block.
>>> We also have the same optimazation for the inner #iTrue:ifFalse:  
>>> message (i.e. at bytecode: 218 and: 228).
>>>
>>> We can rewrite the code to produce the same flow :
>>>
>>>> [((ind _ ind + 1) > list size) not
>>>> 			or: [(parts at: ind) first asLowercase = projectName
>>>> 				and: [versions add: (parts at: ind)]]]] whileTrue
>>>
>>> This is what the decompiler found out.
>>> But this do not produce the same bytecode:
>>>
>>>> 203 <14> pushTemp: 4
>>>> 204 <76> pushConstant: 1
>>>> 205 <B0> send: +
>>>> 206 <81 44> storeIntoTemp: 4
>>>> 208 <12> pushTemp: 2
>>>> 209 <C2> send: size
>>>> 210 <B3> send: >
>>>> 211 <83 18> send: not
>>>> 213 <A8 12> jumpTrue: 233
>>>>
>>>> 215 <13> pushTemp: 3
>>>> 216 <14> pushTemp: 4
>>>> 217 <C0> send: at:
>>>> 218 <83 14> send: first
>>>> 220 <83 15> send: asLowercase
>>>> 222 <16> pushTemp: 6
>>>> 223 <B6> send: =
>>>> 224 <AC 09> jumpFalse: 235
>>>>
>>>> 226 <17> pushTemp: 7
>>>> 227 <13> pushTemp: 3
>>>> 228 <14> pushTemp: 4
>>>> 229 <C0> send: at:
>>>> 230 <83 37> send: add:
>>>> 232 <99> jumpFalse: 235
>>>>
>>>> 233 <A3 E0> jumpTo: 203
>>>
>>> The only difference is in the bytecode 211 and 213:
>>> A "send: not" is added and the jump don't have the same destination.
>>>
>>> The old compiler dose not make this smart optimization.
>>> So I have several  question.
>>>
>>> Do we want the decompiler produce different bytecode (i.e. adding a  
>>> "send: not"…)?
>>> Should we do such optimization? (IMO I don't thinks so.)
>>> where dose the optimization take place? I can't figure out where it  
>>> is. I get lost inside ASTTranslator and IRTranslator.
>>>
>>>
>>> Thanks
>>>
>>> 	Mth
>>>
>>>
>>>
>>
>>
>
>
>
>
>
>
> ___________________________________________________________________________
> Yahoo! Mail réinvente le mail ! Découvrez le nouveau Yahoo! Mail et son  
> interface révolutionnaire.
> http://fr.mail.yahoo.com

More information about the Newcompiler mailing list