Data flow: Use pruning to reduce call contexts

[hvitved]

https://jenkins.internal.semmle.com/job/Changes/job/CSharp-Differences/1232/
https://jenkins.internal.semmle.com/job/Changes/job/CPP-Differences/2186/
https://jenkins.internal.semmle.com/job/Changes/job/Java-Differences/1531/
https://jenkins.internal.semmle.com/job/Changes/job/Python-Differences/644/

https://jenkins.internal.semmle.com/job/Changes/job/CSharp-Differences/1234/
https://jenkins.internal.semmle.com/job/Changes/job/CPP-Differences/2189/
https://jenkins.internal.semmle.com/job/Changes/job/Java-Differences/1533/
https://jenkins.internal.semmle.com/job/Changes/job/Python-Differences/647/

[aschackmull]

What's the motivation for this? The if recordDataFlowCallSite(call, c) then ought to be doing a two-column join acting as a filter and the same goes for if reducedViableImplInReturn(c, call) then, so that ought to be cheap. Or do we have a bad join-order lurking?
Ahh, I see it goes a little beyond that by removing the call context if it isn't going to be needed. I guess you have some motivating example where this made a big difference?

[hvitved]

Ahh, I see it goes a little beyond that by removing the call context if it isn't going to be needed. I guess you have some motivating example where this made a big difference?

The motivation comes from the fact that I would like to experiment with adding call contexts earlier in the pruning stages (possibly as a new stage in between what is currently stages 2 and 3). Other than that no motivation, except it felt like a "free lunch".

[aschackmull]

I'm not sure I see the relation to "adding call contexts earlier in the pruning stages"? That should be orthogonal to this, right? (Might even be easier without this change?).

[aschackmull]

Is this motivated by lambda-dispatch in higher-order summaries? If so, then I have some other ideas.

[hvitved]

I'm not sure I see the relation to "adding call contexts earlier in the pruning stages"? That should be orthogonal to this, right? (Might even be easier without this change?).

What I meant was to apply the same trick when adding call contexts earlier, to reduce tuple counts.

[aschackmull]

The Differences jobs don't really look convincing.

[hvitved]

The Differences jobs don't really look convincing.

No, it looks like maybe there is a bad join-order in the newly introduced predicates.

[tausbn]

This PR appears to have gone stale, so I've taken the liberty of closing it (in lieu of asking "hey can I close this?" and waiting for someone else to take appropriate action).

Feel free to reopen it if it's still relevant! 🙂