fix === when encountering null pointer

[Moelf]

closes #44712

[KristofferC]

Possible to make a test for this since it is clearly something that is lacking?

[vtjnash]

julia> struct AB
       a::Some{Any}
       b::Int
       AB() = new()
       end

julia> unsafe_load(Ptr{AB}(pointer(Int[0,1]))) === unsafe_load(Ptr{AB}(pointer(Int[0,2])))
true

julia> unsafe_load(Ptr{AB}(pointer(Int[0,1]))), unsafe_load(Ptr{AB}(pointer(Int[0,2])))
(AB(#undef, 1), AB(#undef, 2))

[Moelf]

that's a good test, it returns true even in 1.7...

[Moelf]

Error seems unrelated

[KristofferC]

I rebased it to see if that makes CI happier.

[KristofferC]

Looks like there is still a bug here somewhere. Minimizing the error in CI a bit we get:

using Base.BinaryPlatforms
using Base.Experimental

begin
    Experimental.@force_compile
    p1 = Platform("x86_64", "linux"; julia_version=v"1.5.0")
    p2 = Platform("x86_64", "linux"; julia_version=v"1.6.0")
    @show tags(p1)["julia_version"]
    @show tags(p2)["julia_version"]
end

If this is run with the compilation forced it outputs:

(tags(p1))["julia_version"] = "1.5.0"
(tags(p2))["julia_version"] = "1.5.0"

otherwise

(tags(p1))["julia_version"] = "1.5.0"
(tags(p2))["julia_version"] = "1.6.0"

Breaking out the version arguments into separate variables makes the behaviour go away.

[Moelf]

Is it related to this PR somehow?

[KristofferC]

Presumably in some way since it doesn't happen on e.g the master branch.

[Moelf]

julia> begin
           Experimental.@force_compile
           p2 = Platform("x86_64", "linux"; julia_version=v"1.6.0")
           @show tags(p2)["julia_version"]
       end
(tags(p2))["julia_version"] = "1.6.0"
"1.6.0"

eh, I don't understand, feels like something to do with compilation / @pureness stuff rather than === in this PR; because if you switch order:

julia> begin
           Experimental.@force_compile
           p2 = Platform("x86_64", "linux"; julia_version=v"1.6.0")
           p1 = Platform("x86_64", "linux"; julia_version=v"1.5.0")
           @show tags(p1)["julia_version"]
           @show tags(p2)["julia_version"]
       end
(tags(p1))["julia_version"] = "1.6.0"
(tags(p2))["julia_version"] = "1.6.0"
"1.6.0"

[KristofferC]

Would be nice to have a dedicated test for the latest found problem. Maybe you can cook up one more @vtjnash?

[Moelf]

I think I can cook one up today

[Moelf]

it's harder than I thought, I try to make two immutable structs both containing non-null pointers but are different, I tried something like this but it segfaults:

julia> let a  = Int[1, 1], b = Int[2, 1]
           GC.@preserve a b begin
               unsafe_load(Ptr{N44712}(pointer(a))) === unsafe_load(Ptr{N44712}(pointer(b)))
           end
       end

[vtjnash]

neither of those contain a null pointer

[vtjnash]

What you were breaking is the much more trivial case of some pointer followed by some value:

(("", 0),) === (("", 1),)

normally we compile this, so we don't observe often if it broke

[Moelf]

The windows test failure is unrelated (I checked the CI a bunch of other PRs test also failing)