`fatal error in type inference (type bound)` when testing Qecsim on 1.10

[KristofferC]

PkgEval log: https://s3.amazonaws.com/julialang-reports/nanosoldier/pkgeval/by_hash/152f0e9_vs_6fc1be0/Qecsim.primary.log

Error:

qec_merge-one: Error During Test at /Users/kristoffercarlsson/PkgEvalAnalysis/dev/Qecsim/test/App.jl:336
  Got exception outside of a @test
  fatal error in type inference (type bound)
  Stacktrace:
    [1] collect(itr::Base.Generator{Tuple{Symbol, Symbol}, Qecsim.App.var"#5#11"})
      @ Base ./array.jl:841
    [2] _totuple
      @ Qecsim.App ./tuple.jl:425 [inlined]
    [3] Tuple
      @ Qecsim.App ./tuple.jl:391 [inlined]
    [4] qec_merge(::Dict{Symbol, Any}, ::Vararg{Dict{Symbol, Any}})
      @ Qecsim.App ~/PkgEvalAnalysis/dev/Qecsim/src/App.jl:376
    [5] macro expansion
      @ ~/PkgEvalAnalysis/dev/Qecsim/test/App.jl:339 [inlined]
...

Repros locally on Pkg.test. Haven't looked into it more.

[gbaraldi]

@KristofferC does this still reproduce for you?

[gbaraldi]

This was fixed by #50432. Is that change backportable @Keno ?

[Keno]

That's not really a "fixed by" so much as a "disabled by" because we thought these weren't happening anymore. We should track down what the type system bug here is and add appropriate tests.

[gbaraldi]

The error was introduced in #48421

[gbaraldi]

Short reproducer

a = [:error_weight_pvar => 1.09
    :n_logical_commutations => [2, 2]]

b = Dict(a)
c = b
vector_val_keys = (:n_logical_commutations, :custom_totals,)
_null_vec_copy(::Nothing) = nothing
_null_vec_copy(vec::AbstractVector) = deepcopy(vec)


function repro(data...)
    vector_val_keys = (:n_logical_commutations, :custom_totals,)
    for runs_data in data
        vector_vals = Tuple(_null_vec_copy(runs_data[k]) for k in vector_val_keys)
    end
end

repro(b,c)

[vtjnash]

Looks like probably an inference bug handling ccall

julia> Cthulhu.descend(Base._array_for, (Type{<:AbstractVector{T}} where T, Base.HasLength, Int))

If you inspect the return value here, it returns Array{<:AbstractVector{T},1} (note the missing where T)

This then flows through to an unexpected runtime query result during union splitting:

julia> Vector{Vector{Int}} <: (Vector{<:AbstractVector{T}} where T).body
false