Fix call graph with respect to constants.

[chriseth]

Fixes #11007

[chriseth]

@cameel can you add tests to this, please?

[cameel]

Sure.

[cameel]

Tests ready. Unfortunately they don't pass so it looks like the fix is not enough. It fixes constants but for immutables, ordinary variables or even referencing the selector in a function body an edge to the internal dispatch is still being added.

@chriseth Should I take it over and fix these cases?

[chriseth]

@cameel can you summarize the problems? Immutables should have the same behaviour as regular state variables. The expression A.g.selector should add an edge from the dispatch to g, but not to he dispatch. I think in the IR codegen, we even generated code for g

[cameel]

Sorry, it's actually an edge from the dispatch, not to the dispatch that is being added.

So anyway, the problems I was referring to were just all of these adding these edges from dispatch:

bytes4 immutable pubImmutable = Base.pub.selector;

bytes4 pubVar = Base.pub.selector;

return (Base.ext.selector, Base.pub.selector);

The functions are not being called here and I think you can't call them internally via the bytes4 variable that contains the selector? So I thought they should not be in the graph.

But you have a point about the IR codegen. I think it does add them to the dispatch when it's visiting MemberAccess. In that case I'll change the test expectations to expect the edge from dispatch.

[cameel]

@chriseth Expectations updated.

[cameel]

I have also pushed a fixup that fixes the case found by @bshastry in #11014 (comment)

[chriseth]

Looks good! Please rebaes.

[cameel]

Rebased and squashed.

[chriseth]

Approving.