[Vm-dev] Re: WordArray at: signedness issue in COG

[Nicolas Cellier]

2013/12/27 Nicolas Cellier <nicolas.cellier.aka.nice at gmail.com>

> Hi,
> I tried to generate a recent COG (VMMaker.oscog-eem.570), but it produces
> spurious negative ints when accessing WordArrays (Bitmap and the like).
>
> It's not a pure COG, it's home brewed, but I don't think that I introduced
> the problem.
>
> I caught it after protecting Object>at:put: with
> (self class isWords and: [value isInteger and: [value between:
> -16r80000000 and: -1]]) ifTrue: [^self at:index put: (value bitAnd:
> 16rFFFFFFFF)].
> Otherwise, image is unusable.
>
> Then I can reproduce it for example while rotating a window, see
> attachment.
>
> fills is a a WordArray(4278190081 4278190081) in
> BalloonEngine>>drawGeneralBezierShape:fill:borderWidth:borderColor:transform:
>
>
> But then primAddBezierShape:segments:fill:lineWidth:lineFill: receive
> -16777215 as lineFill
> -16777215 bitAnd: 16rFFFFFFFF -> 4278190081, so this is correctly (fills
> at: 2) but with wrong sign interpretation.
>
> Curiously, fillStyle (fills at: 1) remained positive.
>
> Eliot, any idea?
>

OK, it was partly my fault, partly incorrect type simplification in
CCodeGenerator inlining.
I redefined positive32BitInteger: to take an #usqInt argument.
Thus I don't have to test >= 0, because the argument is always
re-interpreted positive.
For some reason, bytecodePrimAt is inlining like this:
sqInt result1;
...
/* Bitmap */
result1 = long32At((rcvr + BaseHeaderSize) + ((((index >> 1)) - 1) << 2));
/* begin positive32BitIntegerFor: */
if (result1 < 0x40000000) {
result = ((result1 << 1) | 1);
goto l123;
}

long32At is signed (pouah!)
Thus 16rFF000001, is interpreted < 0
Thus a (negative) SmallInteger is returned...

Why it sometimes work and sometimes fail may be related to the at cache,
but I did not try to dig further...
-------------- next part --------------
An HTML attachment was scrubbed...
URL: http://lists.squeakfoundation.org/pipermail/vm-dev/attachments/20131228/bdc914ed/attachment.htm