v1.9 regression: calling function from within function is not same as calling from REPL

[odow]

We have a very weird bug in the new Julia 1.9 release with InfiniteOpt: infiniteopt/InfiniteOpt.jl#316.

cc @pulsipher

The behavior is different when a function from JuMP is called within a function, or when called by the REPL.

Here's the a reproducible example. You'll need:

import Pkg; Pkg.add("InfiniteOpt")

(ifo) pkg> st
Status `/private/tmp/ifo/Project.toml`
  [20393b10] InfiniteOpt v0.5.7

julia> versioninfo()
Julia Version 1.9.0
Commit 8e630552924 (2023-05-07 11:25 UTC)
Platform Info:
  OS: macOS (x86_64-apple-darwin22.4.0)
  CPU: 8 × Intel(R) Core(TM) i5-8259U CPU @ 2.30GHz
  WORD_SIZE: 64
  LIBM: libopenlibm
  LLVM: libLLVM-14.0.6 (ORCJIT, skylake)
  Threads: 1 on 8 virtual cores

julia> using InfiniteOpt

julia> function my_add_variable(model, v)
           y = InfiniteOpt.dispatch_variable_ref(v.infinite_variable_ref)
           JuMP.check_belongs_to_model(y, model)
           return
       end
my_add_variable (generic function with 1 method)

julia> m1 = InfiniteOpt.InfiniteModel();

julia> InfiniteOpt.@infinite_parameter(m1, p in [0, 1]);

julia> InfiniteOpt.@variable(m1, x, Infinite(p));

julia> v = JuMP.build_variable(
           error, 
           JuMP.VariableInfo(false, 0.0, false, 0.0, false, 0.0, false, 0.0, false, false), 
           InfiniteOpt.Point(x, 0.5),
       );

julia> m2 = InfiniteOpt.InfiniteModel();

julia> my_add_variable(m2, v)  # INCORRECT: When called via a function, does not error

julia> y = dispatch_variable_ref(v.infinite_variable_ref)
x(p)

julia> JuMP.check_belongs_to_model(y, m2)  # CORRECT: When called via REPL, does error
ERROR: VariableNotOwned{InfiniteVariableRef}(x(p))
Stacktrace:
 [1] check_belongs_to_model(v::InfiniteVariableRef, model::InfiniteModel)
   @ JuMP ~/.julia/packages/JuMP/AKvOr/src/variables.jl:282
 [2] top-level scope
   @ REPL[12]:1

I can't seem to reduce this down to a simpler example.

JuMP.check_belongs_to_model is https://github.com/jump-dev/JuMP.jl/blob/dd66846cd28b9cc65b82b8d6bb8de67162c83f8b/src/variables.jl#L280-L284
which is basically just

if owner_model(v) !== model
    throw(VariableNotOwned(v))
end

but if you replace the JuMP.check_belongs_to_model call with owner_model(v) !== model then things are the same.

[KristofferC]

Bisected to b3f25d7

cc @Keno

[odow]

This wasn't reported earlier because despite InfiniteOpt testing on nightly, recent commits were passing:
https://github.com/infiniteopt/InfiniteOpt.jl/actions/runs/4878443048/jobs/8704079025

But I don't know if this was something that was fixed on master that needs back porting to 1.9, or if the bug is still there and it just doesn't trigger on CI for some reason.

On a recent nightly, I still get the bug:

julia> using InfiniteOpt

julia> function my_add_variable(model, v)
           y = InfiniteOpt.dispatch_variable_ref(v.infinite_variable_ref)
           JuMP.check_belongs_to_model(y, model)
           return
       end
my_add_variable (generic function with 1 method)

julia> m1 = InfiniteOpt.InfiniteModel();

julia> InfiniteOpt.@infinite_parameter(m1, p in [0, 1]);

julia> InfiniteOpt.@variable(m1, x, Infinite(p));

julia> v = JuMP.build_variable(
           error, 
           JuMP.VariableInfo(false, 0.0, false, 0.0, false, 0.0, false, 0.0, false, false), 
           InfiniteOpt.Point(x, 0.5),
       );

julia> m2 = InfiniteOpt.InfiniteModel();

julia> my_add_variable(m2, v)  # INCORRECT: When called via a function, does not error

julia> y = dispatch_variable_ref(v.infinite_variable_ref)
x(p)

julia> JuMP.check_belongs_to_model(y, m2)  # CORRECT: When called via REPL, does error
ERROR: VariableNotOwned{InfiniteVariableRef}(x(p))
Stacktrace:
 [1] check_belongs_to_model(v::InfiniteVariableRef, model::InfiniteModel)
   @ JuMP ~/.julia/packages/JuMP/AKvOr/src/variables.jl:282
 [2] top-level scope
   @ REPL[13]:1

julia> versioninfo()
Julia Version 1.10.0-DEV.1151
Commit b4d15302517 (2023-04-29 23:21 UTC)
Platform Info:
  OS: macOS (x86_64-apple-darwin21.5.0)
  CPU: 8 × Intel(R) Core(TM) i5-8259U CPU @ 2.30GHz
  WORD_SIZE: 64
  LIBM: libopenlibm
  LLVM: libLLVM-14.0.6 (ORCJIT, skylake)
  Threads: 1 on 8 virtual cores

[vtjnash]

Correctly fails in v1.11. There have been many fixes to the effects system, particular for errors

julia> my_add_variable(m2, v)  # INCORRECT: When called via a function, does not error
ERROR: VariableNotOwned{InfiniteVariableRef}(x(p)): the variable x(p) cannot be used in this model because
it belongs to a different model.