Sink CodeInfo transformation into transform_result_for_cache

[Keno]

Several downstream consumers want to cache IRCode rather than CodeInfo. Right now they need to override both finish! and transform_result_for_cache. With this change, only overriding the latter should be sufficient. As a nice bonus, we can avoid doing the work of converting to CodeInfo for interpreters that don't cache at all such as the REPL interpreter.

[aviatesk]

This change removes the assignments to caller.src and result.src, but opt.src should already be identical to their CodeInfo. My understanding is that it is destructively updated by ir_to_codeinf!.

[Keno]

Yes, the code path that goes through finish! twice (although I'm not sure why that happens - possibly a bug?) takes the ir === nothing path now.

[vtjnash]

although I'm not sure why that happens - possibly a bug?

Somewhat of an implementation defect, though one needed to ensure the soundness of the caller mode currently. This is the reason for #56880 which deletes that call, so we could likely just merge that PR first. I think it will have some superficial conflicts with this PR, but nothing significant and VSCode merge tool might be able to figure it out. (the tests on that PR are hitting a different bug in Windows, apparently by changing the probability of encountering it on that particular test for the specific issue, so that PR is almost ready to merge, but I need some time to fix the other bug)

[serenity4]

After taking a look, it seems that result.src must be filled with a CodeInfo. It seems to be captured by this test:

julia/Compiler/test/inline.jl

Lines 590 to 600 in 99fd5d9

	@noinline fresh_edge_noinlined(a::Integer) = unresolvable(a)
	let src = code_typed1((Integer,)) do x
	@inline fresh_edge_noinlined(x)
	end
	@test count(iscall((src, fresh_edge_noinlined)), src.code) == 0
	end
	let src = code_typed1((Integer,)) do x
	@inline fresh_edge_noinlined(x)
	end
	@test count(iscall((src, fresh_edge_noinlined)), src.code) == 0 # should be idempotent
	end

which, if it passes, solves most (all?) other tests. The failure is that on the second code_typed, fresh_edge_noinlined is no longer inlined, which makes me believe that inference results are not properly cached in this case.

Adding the following to finish!(interp::AbstractInterpreter, caller::InferenceState) solves these issues, but I'm not sure that's the way to go:

    ...
    if !isdefined(result, :ci)
        if isa(opt, OptimizationState)
            result.src = ir_to_codeinf!(opt)
        end
    else
    ...
        if inferred_result isa CodeInfo
            result.src = inferred_result
            ...
        end
    end

Any ideas on how to address this?

[serenity4]

We can close this in favor of #57375