C++: AssignPointerAddExpr and AssignPointerSubExpr should not be bitwise operations

[MathiasVP]

Apparently we've been considering AssignPointerAddExpr and AssignPointerSubExpr as bitwise operations for almost 4 years (since I introduced this bug in one of my very first GitHub good-first-issues: 11a545e 🙈)

[jketema]

Should we add a library change note for this?

[jketema]

I think it would also make sense to run a DCA experiment with uninterpreted queries enabled.

[MathiasVP]

Should we add a library change note for this?

Yeah, I think that's a good idea.

I think it would also make sense to run a DCA experiment with uninterpreted queries enabled.

Doh! Good point. I'll rerun DCA with uninterpreted queries enabled.

[jketema]

You might want to add data set check for full measure

[MathiasVP]

@jketema the DCA run with dataset checks have completed now. I think it LGTM. Can I get your 👀 on it as well?

[jketema]

3 hours seems suspiciously fast, and I'm missing the frontend timings for some reason.

[MathiasVP]

3 hours seems suspiciously fast, and I'm missing the frontend timings for some reason.

Oh wait. Does dataset measure require me to disable database caching? 🤔

[jketema]

You made a schema change, I you will need to run without caching.

[MathiasVP]

You made a schema change, I you will need to run without caching.

That's ... a good point. Starting another one without DB caching now, then!

[jketema]

Also, it's a bit suspect that we get a change in IR inconsistencies with caching enabled. That means the schema change somehow affects IR generation?

[MathiasVP]

Also, it's a bit suspect that we get a change in IR inconsistencies with caching enabled. That means the schema change somehow affects IR generation?

Yeah, I was wondering about that as well. I wrote it off as noise, but let me just investigate this while waiting for this (hopefully) last DCA run anyway.

[MathiasVP]

Can't reproduce the IR inconsistencies locally. Running the "count IR inconsistencies" query locally on the cached openzfs DB (on the HEAD of this PR) yields:

missingOperand	unexpectedOperand	duplicateOperand	missingPhiOperand	missingOperandType	duplicateChiOperand	sideEffectWithoutPrimary	instructionWithoutSuccessor	ambiguousSuccessors	unexplainedLoop	unnecessaryPhiInstruction	memoryOperandDefinitionIsUnmodeled	operandAcrossFunctions	containsLoopOfForwardEdges	lostReachability	backEdgeCountMismatch	useNotDominatedByDefinition	switchInstructionWithoutDefaultEdge	notMarkedAsConflated	wronglyMarkedAsConflated	invalidOverlap	nonUniqueEnclosingIRFunction	fieldAddressOnNonPointer	thisArgumentIsNonPointer	nonUniqueIRVariable
520	0	0	1	3,726	0	0	104	2,970	0	2	0	0	0	0	0	59,391	0	44	0	0	0	0	0	77,431

which is exactly what we have on the "before" side on DCA.

[MathiasVP]

DCA looks fine. The same IR inconsistencies are still present, but that may now be from nondeterministic extraction.

[jketema]

LGTM, still needs a change note though.

[MathiasVP]

LGTM, still needs a change note though.

Added in 6e385ca.