#bitInvert32 broke with fast-bytes

[krono]

Squeak 64bit:

((2 raisedTo: 30) -1 ) in: [:x| 'x ', x, ' x inv ', x bitInvert32]
 'x 1073741823 x inv 3221225472'

RSqueak 64bit:

((2 raisedTo: 30) -1 ) in: [:x| 'x ', x, ' x inv ', x bitInvert32]
'x 1073741823 x inv -0000000003221225474'

Note #bitXor: is this in Integer

bitInvert32
    "Answer the 32-bit complement of the receiver."

    ^ self bitXor: 16rFFFFFFFF

This is uses excessively in BitBlt(Simulation). viz. BitBltSimulation >> #copyLoop:

copyLoop
    "…"
            destWord := (destMask bitAnd: mergeWord) bitOr:
                            (destWord bitAnd: destMask bitInvert32).
    "…"

which is translated to

// …
            destWord = (destMask & mergeWord) | (destWord & ((unsigned int)~destMask));

While this looks unproblematic on first sight (b/c uint will be 64bit on 64bit anyway), I'm unsure for all the other shifts

[krono]

Buglet:

RSqueak 64bit:

((2 raisedTo: 30) -1 ) bitInvert32. "-3221225474 [SmallInteger]"
" inlined #bitInvert32 "
((2 raisedTo: 30) -1 )  bitXor: 16rFFFFFFFF."3221225472 [SmallInteger]"

[krono]

Non-Buglet: RSqueak 32bit:

((2 raisedTo: 30) -1 ) bitInvert32. "3221225472 [SmallInteger]"
" inlined #bitInvert32 "
((2 raisedTo: 30) -1 )  bitXor: 16rFFFFFFFF."3221225472 [SmallInteger]"

[timfel]

I noticed that in the previous impl, we ignored the sign in the bit operations and forcibly interpreted the arguments as uints, whereas now we keep the sign intact (and go to rbigint, if required). One thing to try is to revert to the old behavior, and not use rbigint at all, maybe the fallback code does something that rbigint doesn't...

[krono]

Fixed via 4941d8f6d