Fix test failures on 1.11

[timholy]

No description provided.

[timholy]

I now have a much clearer understanding of what is going wrong, but I still lack an understanding of why.

In case later changes fix it, the report in question is https://github.com/timholy/SnoopCompile.jl/actions/runs/13989097828/job/39168795257?pr=413

Here's the logic: https://github.com/timholy/SnoopCompile.jl/blob/teh/debug_1.11/test/testmodules/Stale/StaleC/src/StaleC.jl should "heal" invalidations in the build_stale call chain. So the appearance of use_stale in the recompilation result https://github.com/timholy/SnoopCompile.jl/blob/a8f0490237f0c26ac50d35baca38f1298f532b80/test/snoop_inference.jl#L873-L877

InferenceTimingNode: 0.038472/0.054491 on Core.Compiler.Timings.ROOT() with 2 direct children
├─ InferenceTimingNode: 0.000088/0.000097 on StaleB.useA() with 1 direct children
│  └─ InferenceTimingNode: 0.000009/0.000009 on getproperty(StaleA::Module, stale::Symbol) with 0 direct children
└─ InferenceTimingNode: 0.015922/0.015922 on StaleA.use_stale(::Vector{Any}) with 0 direct children
unexpected result: 2 backedges

is surprising. "0 direct children" seems to suggest that nothing "needed" it, and a test a little higher up in the file https://github.com/timholy/SnoopCompile.jl/blob/a8f0490237f0c26ac50d35baca38f1298f532b80/test/snoop_inference.jl#L858 explicitly checks that there is a valid compiled instance. So I can't understand why this method is being recompiled. It is being triggered by the https://github.com/timholy/SnoopCompile.jl/blob/teh/debug_1.11/test/snoop_inference.jl#L875 line (if you comment it out, use_stale isn't recompiled).

Locally, this happens only with coverage=true, with coverage=false it behaves as I expect.

[timholy]

Summary of a discussion on slack: this is a consequence of avoiding cached native code:

• Use pkgimages for coverage & malloc tracking by ignoring native code in tracked packages JuliaLang/julia#52123
• staticdata: fix assert from partially disabled native code JuliaLang/julia#53439
• Prevent tainting native code loading from propagating JuliaLang/julia#53457

in conjunction with the more recent (Julia 1.11.3 and Julia 1.11.4) removal of the inferred code:

JuliaLang/julia#56749

The plan is to revert the latter on Julia 1.11, hopefully landing in Julia 1.11.5. Meanwhile, I'll edit the test to compensate.

The signature: in

    mius = only(methodinstances(StaleA.use_stale))
    cius = mius.cache
    @test cius.max_world == typemax(UInt)
    @show cius.specptr cius.invoke cius.inferred

one gets

cius.specptr = Ptr{Nothing} @0x0000000000000000
cius.invoke = Ptr{Nothing} @0x0000000000000000
cius.inferred = nothing

when coverage is on. (When coverage is off, those pointers are non-NULL.)

[timholy]

Ah, lovely.