Closed GoogleCodeExporter closed 8 years ago
Hey there,
this is a known issue actually, I've stated it in the documentation already.
You are correct in your observation that the switch statements are the main
offender.
I have been pouring over this problem for a very long time, but so far haven't
been able to tackle it. I've contacted Joa Ebert (the guy who builds Apparat)
about this and he has indicated that the opcode documentation isn't always
correct. This is probably the case with the swithc statements.
You can observe it by checking out the deserialized opcodes, when examining the
results for the switch statement in your method you will see that very probably
the last opcodes in the method body will be a 'lookupswitch', this is of course
incorrect, since the last opcode *should* be 'returnvoid'.
Clearly something fishy is going on, should you be able to figure out what is
going wrong I think a LOT of people will be eternally in your debt :)
The first place to check out is the org.as3commons.bytecode.abc.enum.Opcode
class.
good luck :)
Original comment by ihatelivelyids
on 30 Jun 2011 at 7:35
OK,
I am actually introducing a method to a class at load time.
Before introducing the method into the class my part added method body looks
like this
//maxStack=3, localCount=4, initScopeDepth=12, maxScopeDepth=13
[trace] debugline [37]
[trace] getlocal_0
[trace] pushscope
[trace] debug [1, 20, 0, 37]
[trace] debug [1, 21, 1, 37]
[trace] debugline [39]
[trace] getlocal_0
[trace] getlocal_1
[trace] getlocal_2
[trace]
callsuper [QName[Namespace[protectedNamespace::spark.components.supportClasse
s:SkinnableComponent]:partAdded], 2]
[trace] pop
[trace] jump [30]
[trace] label
[trace] debugline [49]
[trace] getlocal_0
[trace] getproperty [QName[Namespace[public]:submitButton]]
[trace] findpropstrict [QName[Namespace[public::flash.events]:MouseEvent]]
[trace] getproperty [QName[Namespace[public::flash.events]:MouseEvent]]
[trace] getproperty [QName[Namespace[public]:CLICK]]
[trace] getlocal_0
[trace]
getproperty [QName[Namespace[private::components:ChatComponent]:onSendClicked
]]
[trace] callproperty [QName[Namespace[public]:addEventListener], 2]
[trace] pop
[trace] debugline [50]
[trace] jump [51]
[trace] label
[trace] jump [46]
[trace] debugline [46]
[trace] getlocal_2
[trace] setlocal_3
[trace] debugline [48]
[trace] getlocal_0
[trace] getproperty [QName[Namespace[public]:submitButton]]
[trace] getlocal_3
[trace] ifstrictne [6]
[trace] pushshort [0]
[trace] jump [13]
[trace] pushfalse
[trace] iffalse [6]
[trace] pushshort [1]
[trace] jump [2]
[trace] pushshort [1]
[trace] kill [3]
[trace] lookupswitch [-40, 1, -65,-40]
[trace] debugline [53]
[trace] returnvoid
[trace] }
[trace]
After introducing another method the the partAdded method looks likt this
//maxStack=3, localCount=4, initScopeDepth=12, maxScopeDepth=13
[trace] debugline [37]
[trace] getlocal_0
[trace] pushscope
[trace] debug [1, 20, 0, 37]
[trace] debug [1, 21, 1, 37]
[trace] debugline [39]
[trace] getlocal_0
[trace] getlocal_1
[trace] getlocal_2
[trace]
callsuper [QName[Namespace[protectedNamespace::spark.components.supportClasse
s:SkinnableComponent]:partAdded], 2]
[trace] pop
[trace] jump [30]
[trace] label
[trace] debugline [49]
[trace] getlocal_0
[trace] getproperty [QName[Namespace[public]:submitButton]]
[trace] findpropstrict [QName[Namespace[public::flash.events]:MouseEvent]]
[trace] getproperty [QName[Namespace[public::flash.events]:MouseEvent]]
[trace] getproperty [QName[Namespace[public]:CLICK]]
[trace] getlocal_0
[trace]
getproperty [QName[Namespace[private::components:ChatComponent]:onSendClicked
]]
[trace] callproperty [QName[Namespace[public]:addEventListener], 2]
[trace] pop
[trace] debugline [50]
[trace] jump [51]
[trace] label
[trace] jump [46]
[trace] debugline [46]
[trace] getlocal_2
[trace] setlocal_3
[trace] debugline [48]
[trace] getlocal_0
[trace] getproperty [QName[Namespace[public]:submitButton]]
[trace] getlocal_3
[trace] ifstrictne [6]
[trace] pushshort [0]
[trace] jump [13]
[trace] pushfalse
[trace] iffalse [6]
[trace] pushshort [1]
[trace] jump [2]
[trace] pushshort [1]
[trace] kill [3]
[trace] lookupswitch [-98, 1, -65,-40]
[trace] debugline [53]
[trace] returnvoid
Which is identical apart from the obviously broken lookupswitch at the end. So
my theory is that this is a serialisation issue in the bytecode library. If I
replace the switch statement with an if then the class can be loaded after
introduction and works fine, which seems to indicate that I am doing the
introduction correctly - here is what I do
methodInfo.methodBody = methodBody;
methodBody.methodSignature = methodInfo;
method.traitMethod = methodInfo;
instanceInfo.addTrait( method );
abcFile.addMethodBody( methodBody );
abcFile.addMethodInfo( methodInfo );
Is that all the correct linkages for introducing a method?
(Please note that I am not creating a proxy, but introducing a method into a
class before it is loaded)
Have I got anything wrong?
Original comment by conrad.w...@gmail.com
on 30 Jun 2011 at 7:12
After more investigation:
I think there is some sort of bug in the OpCode.resolveBackPatches method. I
can see this outputting the -98 when I serialise the instanceInfo. As the
partAdded method has not changed this number should not change.
BTW here is my process - you can see that this does not involve the Flash
verifier/player at all
deserialise the swfTag of the class I am interested in
*examine the method body for partAdded (first list above)
introduce method (using the steps I mentioned above)
serialise the abcfile and make that byte array the swftag body
deserialise the new swftag
*examine the method body of partAdded (second list above)
The lookupswitch seems to be correct before resolving back patches, and seems
to be wrong afterwards. I need to understand more about this back patching
before ican see what it is doing wrong (if indeed it is)
Original comment by conrad.w...@gmail.com
on 1 Jul 2011 at 6:06
Hi there,
your process is correct indeed. The backpatching works as follows: During
deserialization all the jump opcodes (ifs, for loops, switches, etc) will have
their jump targets resolved. Which means using the jump offset the opcode that
its pointing to is looked up.
Then during serialization all the start and end positions of every opcode are
saved, afterwards all the jump opcodes are examined again and are checked to
see if their offsets are still pointing to the correct opcode, if not, the
correct opcode is looked up they the offsets are 'back patched', which means
their value is replaced in the serialized bytearray with the correct value.
From your dumps it is quite obvious that the jump targets for switch statements
are indeed not backpatched properly... :(
By the way, I see you're using a debug version of a SWF, be prepared to run
into even more weirder errors once you start deserializing/serializing a SWF
that was compiled as a release build...
Hope that helps a little, good luck!
cheers,
Roland
Original comment by rol...@stackandheap.com
on 1 Jul 2011 at 7:11
If I turn off backpatching everything works as expected :-) This is probably
because I havent mucked around with the contents of the method body. I'm busy
today and tomorrow, but will try to have a lok anyway.
Original comment by conrad.w...@gmail.com
on 2 Jul 2011 at 6:16
It seems to be that the targetOpCode for the default jump of a switch statement
is not properly populated before serialising and so the resolving of the back
patches miscalculates the jump because of this
var targetOpPos:int = (positions[targetOpcode]) ? positions[targetOpcode] : 0;
which returns 0
This means that the deserialisation of the switch statement is possibly not
correct.
Thats the next place to look :-)
Original comment by conrad.w...@gmail.com
on 2 Jul 2011 at 6:33
Roland,
I was about to tell you that there are issues with the way all jumps are
calulated in Opcode and suggest some bug fixes
LOL
I have been using 1.0RC3, and not looking at trunk. If I use trunk none of my
issues occur and I can see that the bugs I was about to report have already
been fixed.
So a bit of a red herring this - issues is closed really.
However I did learn a lot about the code :-)
Conrad
Original comment by conrad.w...@gmail.com
on 3 Jul 2011 at 8:20
OK, am now in the situation you mentioned above -
Debug build with method injection runs fine.
Release build fails with the "At least one branch target was not on a valid
instruction in the method" error.
If I deserialise the code and reserialise it without introducing the method
then it all works fine.
If I change the switch statement to an if statement and intriduce the method
then it works fine.
So it is specifically to do with switch statements and moving the bytecode
around.
I am still investigating
Original comment by conrad.w...@gmail.com
on 3 Jul 2011 at 9:51
Hi Roland - Please read this one.
I am now running non debug builds using monster debugger for getting
information. I have found the following issue
Before method introduction the partAdded routine looks like this
(String) =
protectedNamespace function QName[Namespace[protectedNamespace::components:ChatComponent]:partAdded](QName[Namespace[public]:String], QName[Namespace[public]:Object]) : QName[Namespace[public]:void]
{
//maxStack=3, localCount=4, initScopeDepth=12, maxScopeDepth=13
0:getlocal_0 :1
1:pushscope :2
2:getlocal_0 :3
3:getlocal_1 :4
4:getlocal_2 :5
5:callsupervoid [QName[Namespace[protectedNamespace::spark.components.supportClasses:SkinnableComponent]:partAdded], 2]:9
9:jump [25]:13
13:label :14
14:getlocal_0 :15
15:getproperty [QName[Namespace[public]:submitButton]]:17
17:getlex [QName[Namespace[public::flash.events]:MouseEvent]]:19
19:getproperty [QName[Namespace[public]:CLICK]]:22
22:getlocal_0 :23
23:getproperty [QName[Namespace[private::components:ChatComponent]:onSendClicked]]:26
26:callpropvoid [QName[Namespace[public]:addEventListener], 2]:30
30:jump [46]:34
34:label :35
35:jump [41]:39
39:getlocal_2 :40
40:setlocal_3 :41
41:getlocal_0 :42
42:getproperty [QName[Namespace[public]:submitButton]]:44
44:getlocal_3 :45
45:ifstrictne [6]:49
49:pushbyte [0]:51
51:jump [12]:55
55:jump [6]:59
59:pushbyte [1]:61
61:jump [2]:65
65:pushbyte [1]:67
67:kill [3]:69
69:lookupswitch [-35, 1, -55,-35]:80
80:returnvoid :81
}
and after method introduction it looks like this
(String) =
protectedNamespace function QName[Namespace[protectedNamespace::components:ChatComponent]:partAdded](QName[Namespace[public]:String], QName[Namespace[public]:Object]) : QName[Namespace[public]:void]
{
//maxStack=3, localCount=4, initScopeDepth=12, maxScopeDepth=13
0:getlocal_0 :1
1:pushscope :2
2:getlocal_0 :3
3:getlocal_1 :4
4:getlocal_2 :5
5:callsupervoid [QName[Namespace[protectedNamespace::spark.components.supportClasses:SkinnableComponent]:partAdded], 2]:9
9:jump [26]:13
13:label :14
14:getlocal_0 :15
15:getproperty [QName[Namespace[public]:submitButton]]:17
17:getlex [QName[Namespace[public::flash.events]:MouseEvent]]:19
19:getproperty [QName[Namespace[public]:CLICK]]:22
22:getlocal_0 :23
23:getproperty [QName[Namespace[private::components:ChatComponent]:onSendClicked]]:26
26:callpropvoid [QName[Namespace[public]:addEventListener], 2]:30
30:jump [46]:34
34:label :35
35:jump [41]:39
39:getlocal_2 :40
40:setlocal_3 :41
41:getlocal_0 :42
42:getproperty [QName[Namespace[public]:submitButton]]:44
44:getlocal_3 :45
45:ifstrictne [6]:49
49:pushbyte [0]:51
51:jump [12]:55
55:jump [6]:59
59:pushbyte [1]:61
61:jump [2]:65
65:pushbyte [1]:67
67:kill [3]:69
69:lookupswitch [-35, 1, -55,-35]:80
80:returnvoid :81
}
What we see here is the first jump at line 9 has changed from 25 to 26. This is
the only difference between the two (according to the debugger). The switch
statement seems to have survived the method introduction this time.
Original comment by conrad.w...@gmail.com
on 3 Jul 2011 at 1:34
Hey Conrad,
very nice catch, and incredibly weird how that jump gets resolved one bytecode
position too far... I really hope to have some spare moments in the coming
weeks to investigate this further. Thanks a lot for your time so far, this is
indeed a great help!
cheers,
Roland
Original comment by ihatelivelyids
on 5 Jul 2011 at 3:07
Hi Conrad,
I've been busy hacking away on the as3commons-bytecode again. I think I managed
to get the whole jump business sorted (all changes are in the trunk at the
moment). Could u send me the SWF that contains your example here? I'd like to
test it out. You can the file to roland [at] stackandheap dot com.
Thanks a lot!
cheers,
Roland
Original comment by ihatelivelyids
on 4 Sep 2011 at 6:49
ok, skip that, I just ran some more tests but I'm screwing up the switch jumps,
so its still not working.
I'm not sure if I have any hairs left to pull out of my head though :)
Original comment by ihatelivelyids
on 4 Sep 2011 at 7:09
Well, I believe I have things working properly, sort of. At least my unit test
that uses a method with a switch statement seems to be passing now, also in
release mode.
Yet, of course, a new error has reared its ugly head. If I use try/catch blocks
in a method body everything works fine with a debug build, yet fails with a
verify error when I use it on a release build...
Original comment by ihatelivelyids
on 4 Sep 2011 at 11:45
Original comment by mastakan...@gmail.com
on 11 Sep 2011 at 6:26
I'm setting this issue to status 'fixed' as what the title indicates, the
switch statement parsing is now working.
Original comment by ihatelivelyids
on 11 Sep 2011 at 6:38
Original issue reported on code.google.com by
conrad.w...@gmail.com
on 30 Jun 2011 at 6:10