Method table test failures on Julia nightly #2840
Description
It looks like we have a new issue with method tables on Julia nightly. See the last CI runs on #2815
Here is a MWE that probably fails on nightly:
####################
# This file has two parts:
# 1. the functions to remove a method
# 2. the tests
####################
# 1. the functions to remove a method
####################
if VERSION < v"1.7.0-0"
@eval macro atomic(ex)
esc(ex)
end
end
const alive_world_val = methods(Base.sqrt).ms[1].deleted_world # typemax(UInt) in Julia v1.3, Int(-1) in Julia 1.0
const here = @__FILE__
isfromhere(method::Method) = here == String(method.file)
"""
Delete all methods of `f` that were defined in this notebook, and leave the ones defined in other packages, base, etc. ✂
Return whether the function has any methods left after deletion.
"""
function delete_toplevel_methods(f::Function)::Bool
# we can delete methods of functions!
# instead of deleting all methods, we only delete methods that were defined in this notebook. This is necessary when the notebook code extends a function from remote code
methods_table = typeof(f).name.mt
deleted_sigs = Set{Type}()
Base.visit(methods_table) do method # iterates through all methods of `f`, including overridden ones
if isfromhere(method) && method.deleted_world == alive_world_val
Base.delete_method(method)
delete_method_doc(method)
push!(deleted_sigs, method.sig)
end
end
# if `f` is an extension to an external function, and we defined a method that overrides a method, for example,
# we define `Base.isodd(n::Integer) = rand(Bool)`, which overrides the existing method `Base.isodd(n::Integer)`
# calling `Base.delete_method` on this method won't bring back the old method, because our new method still exists in the method table, and it has a world age which is newer than the original. (our method has a deleted_world value set, which disables it)
#
# To solve this, we iterate again, and _re-enable any methods that were hidden in this way_, by adding them again to the method table with an even newer `primary_world`.
if !isempty(deleted_sigs)
to_insert = Method[]
Base.visit(methods_table) do method
if !isfromhere(method) && method.sig ∈ deleted_sigs
push!(to_insert, method)
end
end
# separate loop to avoid visiting the recently added method
for method in Iterators.reverse(to_insert)
if VERSION >= v"1.11.0-0"
@atomic method.primary_world = one(typeof(alive_world_val)) # `1` will tell Julia to increment the world counter and set it as this function's world
@atomic method.deleted_world = alive_world_val # set the `deleted_world` property back to the 'alive' value (for Julia v1.6 and up)
else
method.primary_world = one(typeof(alive_world_val))
method.deleted_world = alive_world_val
end
ccall(:jl_method_table_insert, Cvoid, (Any, Any, Ptr{Cvoid}), methods_table, method, C_NULL) # i dont like doing this either!
end
end
return !isempty(methods(f).ms)
end
"""
delete_method_doc(m::Method)
Tries to delete the documentation for this method, this is used when methods are removed.
"""
function delete_method_doc(m::Method)
binding = Docs.Binding(m.module, m.name)
meta = Docs.meta(m.module)
if haskey(meta, binding)
method_sig = Tuple{m.sig.parameters[2:end]...}
multidoc = meta[binding]
filter!(multidoc.order) do msig
if method_sig == msig
pop!(multidoc.docs, msig)
false
else
true
end
end
end
end
####################
# 2. the tests
####################
using Test
@test tan(missing) === missing
@test tan(pi) == 0
Core.eval(Main, quote
Base.tan(::Missing) = 32
end)
@test tan(missing) == 32
@test tan(pi) == 0
delete_toplevel_methods(tan)
@test tan(missing) === missing
@test tan(pi) == 0It works on Julia 1.6, Julia 1.10, and this older Julia nightly that I have installed
➜ ~ julianightly
_
_ _ _(_)_ | Documentation: https://docs.julialang.org
(_) | (_) (_) |
_ _ _| |_ __ _ | Type "?" for help, "]?" for Pkg help.
| | | | | | |/ _` | |
| | |_| | | | (_| | | Version 1.11.0-DEV.1495 (2024-02-07)
_/ |\__'_|_|_|\__'_| | Commit bead1d32f1a (26 days old master)
|__/ |
I'm on a hotspot so I can't download the latest Julia nightly at the moment. But this should help debug the issue!