Re: [squeak-dev] Bug in PolygonMorph>>#filledForm

 

We are overwriting upper row because taking hdir=-1 because of overlap

The problem is not that we write too far, but that we pick too early (on previous row), and it is related to ???

preload!!

To determine if we must preload, we have to take the endBits when the direction is reversed due to overlap (hDir = -1)
endBits := (sx + bbW - 1 bitAnd: pixPerM1) + 1
First error above: checkSourceOverlap has already done that:
sx := sx + bbW - 1.
So it must be:
endBits := (sx bitAnd: pixPerM1) + 1

Then we build mask1:
m1 := destMSB = (hDir > 0)
ifTrue: [AllOnes >> (32 - (startBits * destDepth))]
ifFalse: [AllOnes << (32 - (startBits * destDepth)) bitAnd: AllOnes].
Second error above: if we have reversed direction (hDir = -1), then we must build mask2:
destMSB
ifTrue: [mask2 := AllOnes << (32 - (endBits * destDepth)) bitAnd: AllOnes]
ifFalse: [mask2 := AllOnes >> (32 - (endBits * destDepth))].
Replace mask2 by m1 and endBits by startBits (because startBits contains endBits if hDir=-1)
And you see that the conditions destMSB is reversed...

Last thing: on 64bits, mask1 and mask2 have been declared int64...
But we only want 32 bits mask.
That is why I have protected with the bitAnd: AllOnes
(and thus removed the cCode:inSmalltalk: simulation guard)

Le ven. 21 févr. 2020 à 23:40, Nicolas Cellier <[hidden email]> a écrit :

Several candidates, maybe I missed some:

https://source.squeak.org/VMMaker/VMMaker.oscog-eem.2461.diff

https://source.squeak.org/VMMaker/VMMaker.oscog-eem.2460.diff (Oct 2018)

https://source.squeak.org/VMMaker/VMMaker.oscog-eem.2284.diff

https://source.squeak.org/VMMaker/VMMaker.oscog- nice.2281.diff

Le ven. 21 févr. 2020 à 23:13, Nicolas Cellier <[hidden email]> a écrit :

I do not observe the artifacts at same smear step (nor any other) with

VM [CoInterpreterPrimitives VMMaker.oscog-eem.2266] 5.0.201708312323

That's just 222 commits to bissect...

Le ven. 21 févr. 2020 à 23:00, Nicolas Cellier <[hidden email]> a écrit :

There's already a problem with end 2018 VM

VM [CoInterpreterPrimitives VMMaker.oscog-eem.2488] 5.0.201812040127

Le ven. 21 févr. 2020 à 22:48, Nicolas Cellier <[hidden email]> a écrit :

Interesting case to debug...

In 64bits VM, a problem first appears in first smear:distance: operation (direction 1@0 to the right) at skew=16...

It seems that we copy bits too far to the right (hence we overwrite the bottom rows)

Fortunately, the artefacts outside polygon bounds disappear at final bitAnd operation, but that's scary!

It could be:

- case of overlap detection failure (#checkSourceOverlap), but the code did not change

- bad initialisation in #sourceSkewAndPointerInit (the one I corrected in 2019)

 https://github.com/OpenSmalltalk/opensmalltalk-vm/issues/426

- bug in #copyLoop (which I also corrected in 2019)

https://github.com/OpenSmalltalk/opensmalltalk-vm/issues/441

Weird kind of integer overflow? (I've made most of the int unsigned, so this might not be detected by UB sanitizer).

Or could it be that the extra buffer overrun that I removed did have a side effect of filling more gaps?

It would be interesting to observe what happens at same #smear:distance: step with older VM.

Le ven. 21 févr. 2020 à 20:39, Stéphane Rollandin <[hidden email]> a écrit :

> Hmmm... the form was still filled in Squeak 5.1 (Update 16549),
> VM 201701311639 (32-bit). Not so in Squeak 5.2.

Yes, but that same 5.1 image with a current VM has the problem...

Stef