feat(dataflow): fact subscribers

[axelcool1234]

This adds an extra field to Fact for analyses subscribed to it. This is much broader than a Fact f's dependents, which are for other Facts that need to be notified when f changes. Subscribers can be used to notify an arbitrary number of Facts of the subscribed analysis. This is used in both dead code analysis and sparse analyses like constant propagation analysis.

[math-fehr]

I am quite confused by your description, do you have an example on what would be the difference between subscribers and dependents?

[axelcool1234]

I am quite confused by your description, do you have an example on what would be the difference between subscribers and dependents?

@math-fehr I don't blame you I had trouble differentiating the two at first as well. Here's a good example: in dead code analysis, when a block goes from dead to live in the lattice, it enqueues the block and every single operation within that block for every analysis that's subscribed to it.

Now, you could have only dependents. But it's be an insane amount of dependents for some analyses. subscribers provide more flexibility and prevents this space explosion.

[math-fehr]

So what would subscribers do in that case?

[axelcool1234]

So what would subscribers do in that case?

@math-fehr for every subscribed analysis, it enqueues the pair (subscribed analysis, block) and (subscribed analysis, operation) for every operation in the block.

[axelcool1234]

So what would subscribers do in that case?

@math-fehr for every subscribed analysis, it enqueues the pair (subscribed analysis, block) and (subscribed analysis, operation) for every operation in the block.

Instead of having every single one of these pairs in dependents, which could be literally hundreds, you just have an analysis subscribe and it does this double for loop when dead code changes.

[axelcool1234]

Closed in favor of #859