Refine effects for getfield corner cases

[Keno]

See individual commits.

[Keno]

I've pushed some tiny related commits to this PR, since otherwise, this'll just end up conflicting with each other. Should all be straightforward.

[aviatesk]

I will push some changes to fix test failures.

[Keno]

thanks

[Keno]

I'm concerned about that GC root issue in llvmpasses. I also saw something similar locally on a branch that included these changes. Is that currently a known issue on master or is that new on this branch? In either case, we should try to capture it in rr and see what the missing root is.

[aviatesk]

Maybe it's the issue that Jameson pointed out about #47994 (esp. #47994 (comment))?

[Keno]

Assertions are enabled, so if that were the issue, i think we'd have seen the assertion in #48089.

[aviatesk]

(unrelated to the GC issue) So it seems like the remaining test failures (in compiler/inline and compiler/inference) are because this PR allows irinterp on more frames that are now proved as :foldable thanks to the :consistent-cy improvement of getfield.

[Keno]

The test in compiler/inline is fine, it's just getting some extra junk basic blocks:

CodeInfo(
1 ──       goto #3
2 ──       nothing::Nothing
3 ┄─ %3  = Base.getfield(x, 1)::Int64
└───       goto #4
4 ──       goto #5
5 ──       goto #7
6 ──       nothing::Nothing
7 ┄─ %8  = Base.getfield(x, 2)::Int64
└───       goto #8
8 ──       nothing::Nothing
└───       goto #10
9 ──       nothing::Nothing
10 ┄ %13 = Core.tuple(%3, %8)::Tuple{Int64, Int64}
└───       return %13
)

The test should be updated to test what it actually wants to know (two getfield calls, one tuple call, no other calls).

[aviatesk]

The test in compiler/inline is fine, it's just getting some extra junk basic blocks:

Okay, just confirmed that LLVM optimizes those junks away.

The remaining ones are related to MustAlias lattice extension. I added it for JET's use case and it's not enabled in base yet. Since JET's analyzer doesn't enable irinterp, I'm fine with marking the failing cases as broken.

[aviatesk]

Okay, fixed up all the failed cases. Should be good to go.