[COG] primitiveClosureValue assert is failing with StackVM (possible BUG in bytecodes!!)

 
Hi Mariano,

    indeed it's a bug.  Your fix (objectMemory is: rcvr instanceOf: (self splObj: ClassBlockClosure) compactClassIndex: ClassBlockClosureCompactIndex) is the correct one.

 There are a few other places where this needs to be done also.  I'll roll out new code soon.

best

Eliot

On Wed, Dec 22, 2010 at 12:03 PM, Mariano Martinez Peck <[hidden email]> wrote:

 
Hi Eliot. If I compile a StackVM, and I just press the left arrow in a code pane of a OB windows, in a pharo image, I get an asser that fails:
(oop & 1) 19202

Sorry I bother with these asserts, but I am trying to solve the crashes I have and I think starting for the asserts is a good idea.

The assert that fails is in primitiveClosureValue()  the part:

blockClosure = longAt(GIV(stackPointer) + (GIV(argumentCount) * BytesPerWord));
    /* begin quickFetchInteger:ofObject: */
    oop = longAt((blockClosure + BaseHeaderSize) + (ClosureNumArgsIndex << ShiftForWord));
    assert((oop & 1));
    numArgs = (oop >> 1);
    if (!(GIV(argumentCount) == numArgs)) {
        /* begin primitiveFail */
        if (GIV(primFailCode) == 0) {
            GIV(primFailCode) = 1;
        }
        return;
    }


numArgs results to be a large number (using #printNum:   was 718613205), and if I try to do a #printOop:  I got a " 0x55aa55aa0x55aa55aa is not on the heap"

Now, of course, the "if (!(GIV(argumentCount) == numArgs)) " fails.

What it is funny is that I did a #printOop: of blockClosure instVar, and this is what I get:  0x1f59b578: a(n) Association

Association? WTF ?   so...I guess something weird is happening. I continue debugging it..and I noticed that:

- Several classes were called in addition to Association,like Character. These seems to be compact classes and implement #value and #value:   ;)
- Both, #bytecodePrimValue and #bytecodePrimValueWithArg  were not calling #primitiveClosureValue. So, "isBlock" was false...

I tried modifying #bytecodePrimValueWithArg and #bytecodePrimValue  this line:

isBlock := objectMemory is: rcvr instanceOf: ClassBlockClosure compactClassIndex: ClassBlockClosureCompactIndex.

to this one:

isBlock := objectMemory is: rcvr instanceOf: (self splObj: ClassBlockClosure) compactClassIndex: ClassBlockClosureCompactIndex.

WIth this change, at least, the assert doesn't fail anymore. I made progress (maybe). But still, for normal closures, "isBlock" seems to be false. So.....all block values are being managed as a normal send ????

Now, I don't understand why ClassBlockClosureCompactIndex  is 0. I think this may be the problem. Why it is not 12? (in Pharo  Smalltalk compactClassesArray at: 12 ->>> BlockClosure.)
I notice that ObjectMemory >> initializeCompactClassIndices  does:

    ClassBlockClosureCompactIndex := 0 "12". "Prospective.  Currently TranslatedMethod class"


Notice that this ZERO changes the sematic of #is: oop instanceOf: classOop compactClassIndex: compactClassIndex
called from the #bytecodePrimValue and bytecodePrimValueWithArg

Finally, I changed the mentioned line to this:

isBlock := objectMemory is: rcvr instanceOf: (self splObj: ClassBlockClosure) compactClassIndex: 12.
  
and finally, with block closures, it is calling the primitive, and for Association it doesn't fail the assert.  ahh and the image doesn't crash ;)

I don't know if this is a bug, if my solution was ok, nor it impact. Please let me know.

Cheers

Mariano

 

On Wed, Dec 22, 2010 at 10:34 PM, Eliot Miranda <[hidden email]> wrote:

 
Hi Mariano,

    indeed it's a bug.  Your fix (objectMemory is: rcvr instanceOf: (self splObj: ClassBlockClosure) compactClassIndex: ClassBlockClosureCompactIndex) is the correct one.

Thanks Eliot. But with that fix, the buytecode primitives are still failing and not calling the primitive, but commonSend. So:

1) is this the expected behavior .. or

2) to fix that is not as I did it in my last fix but in some other place that you will fix?

 

 There are a few other places where this needs to be done also.  I'll roll out new code soon.

Ok. Please let me know. I have some changes for better support of SmallIntegers as methods. I can provide them if you want.

Cheers

mariano
 

best

Eliot

On Wed, Dec 22, 2010 at 12:03 PM, Mariano Martinez Peck <[hidden email]> wrote:

 
Hi Eliot. If I compile a StackVM, and I just press the left arrow in a code pane of a OB windows, in a pharo image, I get an asser that fails:
(oop & 1) 19202

Sorry I bother with these asserts, but I am trying to solve the crashes I have and I think starting for the asserts is a good idea.

The assert that fails is in primitiveClosureValue()  the part:

blockClosure = longAt(GIV(stackPointer) + (GIV(argumentCount) * BytesPerWord));
    /* begin quickFetchInteger:ofObject: */
    oop = longAt((blockClosure + BaseHeaderSize) + (ClosureNumArgsIndex << ShiftForWord));
    assert((oop & 1));
    numArgs = (oop >> 1);
    if (!(GIV(argumentCount) == numArgs)) {
        /* begin primitiveFail */
        if (GIV(primFailCode) == 0) {
            GIV(primFailCode) = 1;
        }
        return;
    }


numArgs results to be a large number (using #printNum:   was 718613205), and if I try to do a #printOop:  I got a " 0x55aa55aa0x55aa55aa is not on the heap"

Now, of course, the "if (!(GIV(argumentCount) == numArgs)) " fails.

What it is funny is that I did a #printOop: of blockClosure instVar, and this is what I get:  0x1f59b578: a(n) Association

Association? WTF ?   so...I guess something weird is happening. I continue debugging it..and I noticed that:

- Several classes were called in addition to Association,like Character. These seems to be compact classes and implement #value and #value:   ;)
- Both, #bytecodePrimValue and #bytecodePrimValueWithArg  were not calling #primitiveClosureValue. So, "isBlock" was false...

I tried modifying #bytecodePrimValueWithArg and #bytecodePrimValue  this line:

isBlock := objectMemory is: rcvr instanceOf: ClassBlockClosure compactClassIndex: ClassBlockClosureCompactIndex.

to this one:

isBlock := objectMemory is: rcvr instanceOf: (self splObj: ClassBlockClosure) compactClassIndex: ClassBlockClosureCompactIndex.

WIth this change, at least, the assert doesn't fail anymore. I made progress (maybe). But still, for normal closures, "isBlock" seems to be false. So.....all block values are being managed as a normal send ????

Now, I don't understand why ClassBlockClosureCompactIndex  is 0. I think this may be the problem. Why it is not 12? (in Pharo  Smalltalk compactClassesArray at: 12 ->>> BlockClosure.)
I notice that ObjectMemory >> initializeCompactClassIndices  does:

    ClassBlockClosureCompactIndex := 0 "12". "Prospective.  Currently TranslatedMethod class"


Notice that this ZERO changes the sematic of #is: oop instanceOf: classOop compactClassIndex: compactClassIndex
called from the #bytecodePrimValue and bytecodePrimValueWithArg

Finally, I changed the mentioned line to this:

isBlock := objectMemory is: rcvr instanceOf: (self splObj: ClassBlockClosure) compactClassIndex: 12.
  
and finally, with block closures, it is calling the primitive, and for Association it doesn't fail the assert.  ahh and the image doesn't crash ;)

I don't know if this is a bug, if my solution was ok, nor it impact. Please let me know.

Cheers

Mariano

 

On Thu, Dec 23, 2010 at 2:05 AM, Mariano Martinez Peck <[hidden email]> wrote:

 

On Wed, Dec 22, 2010 at 10:34 PM, Eliot Miranda <[hidden email]> wrote:

 
Hi Mariano,

    indeed it's a bug.  Your fix (objectMemory is: rcvr instanceOf: (self splObj: ClassBlockClosure) compactClassIndex: ClassBlockClosureCompactIndex) is the correct one.

Thanks Eliot. But with that fix, the buytecode primitives are still failing and not calling the primitive, but commonSend. So:

1) is this the expected behavior .. or

No.  The intension is for the value and value: bytecodes to directly invoke closure value.

 

2) to fix that is not as I did it in my last fix but in some other place that you will fix?

I'll try and fix once I can concentrate on Cog (I'm still focussed on Teleplace work until the week following xmas).

 

 

 There are a few other places where this needs to be done also.  I'll roll out new code soon.

Ok. Please let me know. I have some changes for better support of SmallIntegers as methods. I can provide them if you want.

Please email me the change sets.

thanks!

Eliot

 

Cheers

mariano
 

best

Eliot

On Wed, Dec 22, 2010 at 12:03 PM, Mariano Martinez Peck <[hidden email]> wrote:

 
Hi Eliot. If I compile a StackVM, and I just press the left arrow in a code pane of a OB windows, in a pharo image, I get an asser that fails:
(oop & 1) 19202

Sorry I bother with these asserts, but I am trying to solve the crashes I have and I think starting for the asserts is a good idea.

The assert that fails is in primitiveClosureValue()  the part:

blockClosure = longAt(GIV(stackPointer) + (GIV(argumentCount) * BytesPerWord));
    /* begin quickFetchInteger:ofObject: */
    oop = longAt((blockClosure + BaseHeaderSize) + (ClosureNumArgsIndex << ShiftForWord));
    assert((oop & 1));
    numArgs = (oop >> 1);
    if (!(GIV(argumentCount) == numArgs)) {
        /* begin primitiveFail */
        if (GIV(primFailCode) == 0) {
            GIV(primFailCode) = 1;
        }
        return;
    }


numArgs results to be a large number (using #printNum:   was 718613205), and if I try to do a #printOop:  I got a " 0x55aa55aa0x55aa55aa is not on the heap"

Now, of course, the "if (!(GIV(argumentCount) == numArgs)) " fails.

What it is funny is that I did a #printOop: of blockClosure instVar, and this is what I get:  0x1f59b578: a(n) Association

Association? WTF ?   so...I guess something weird is happening. I continue debugging it..and I noticed that:

- Several classes were called in addition to Association,like Character. These seems to be compact classes and implement #value and #value:   ;)
- Both, #bytecodePrimValue and #bytecodePrimValueWithArg  were not calling #primitiveClosureValue. So, "isBlock" was false...

I tried modifying #bytecodePrimValueWithArg and #bytecodePrimValue  this line:

isBlock := objectMemory is: rcvr instanceOf: ClassBlockClosure compactClassIndex: ClassBlockClosureCompactIndex.

to this one:

isBlock := objectMemory is: rcvr instanceOf: (self splObj: ClassBlockClosure) compactClassIndex: ClassBlockClosureCompactIndex.

WIth this change, at least, the assert doesn't fail anymore. I made progress (maybe). But still, for normal closures, "isBlock" seems to be false. So.....all block values are being managed as a normal send ????

Now, I don't understand why ClassBlockClosureCompactIndex  is 0. I think this may be the problem. Why it is not 12? (in Pharo  Smalltalk compactClassesArray at: 12 ->>> BlockClosure.)
I notice that ObjectMemory >> initializeCompactClassIndices  does:

    ClassBlockClosureCompactIndex := 0 "12". "Prospective.  Currently TranslatedMethod class"


Notice that this ZERO changes the sematic of #is: oop instanceOf: classOop compactClassIndex: compactClassIndex
called from the #bytecodePrimValue and bytecodePrimValueWithArg

Finally, I changed the mentioned line to this:

isBlock := objectMemory is: rcvr instanceOf: (self splObj: ClassBlockClosure) compactClassIndex: 12.
  
and finally, with block closures, it is calling the primitive, and for Association it doesn't fail the assert.  ahh and the image doesn't crash ;)

I don't know if this is a bug, if my solution was ok, nor it impact. Please let me know.

Cheers

Mariano