Method redefinition does not take effect

[s-broda]

Reduced from here https://discourse.julialang.org/t/julia-calls-the-old-method-after-method-is-redefined/10327/7:

Setup:

abstract type VolatilitySpec end
struct GARCH{p,q} <: VolatilitySpec end
struct ARCHModel{VS<:VolatilitySpec}
    ARCHModel{VS}() where {VS}=new()
end
ARCHModel(VS::Type{T}) where {T<:VolatilitySpec} = ARCHModel{VS}()

loglikelihood(am::ARCHModel{T}) where{T} = arch_loglik!(T)

function arch_loglik!(M::Type{GARCH{p,q}}) where {p, q}
    return 5.
end

function selectmodel()
    res = Array{ARCHModel, 2}(4, 4)
    for p = 0:3, q = 0:3
        res[p+1, q+1] = ARCHModel(GARCH{p, q})
    end
    crits = loglikelihood.(res)
end

Calling the method works as expected:

julia> arch_loglik!(GARCH{1, 1})
5.0

Now, call the function selectmodel (this step is crucial for reproducing the issue) and redefine the method (note no warning about the method redefinition):

julia> selectmodel();

julia> arch_loglik!{p, q}(M::Type{GARCH{p,q}}) = 1
arch_loglik! (generic function with 1 method)

However, the new definition doesn't take effect, despite @code_llvm showing the updated definition:

julia> arch_loglik!(GARCH{1, 1})
5.0

julia> @code_llvm arch_loglik!(GARCH{1, 1})

; Function Attrs: uwtable
define i64 @"julia_arch_loglik!_63633"(i8**) #0 !dbg !5 {
top:
  ret i64 1
}

[s-broda]

PS: This is on both 0.6.2 (Windows and Linux) and 0.7 latest nightly.

[RalphAS]

FWIW, further experimentation shows that

• there is a redefinition warning if the above is enclosed in a module other than Main.
• invokelatest seems to call the expected method in 0.7 but not 0.6.2.
• failure to get the updated method seems to depend on details of how the original was used in the selectmodel function: broadcast (as shown above), map, and certain loop expressions (even without hidden closures, AFAICT) show the problem, but scalar calls and some other loops do not.
• the method tables seem to show sensible values for world age on the two versions.

[s-broda]

Should this perhaps get a bug label? It seems that it may be related to #265 and its awesome fix by @vtjnash.

[Keno]

I can no longer reproduce this. There've been a bunch of bug fixes in this area. Can you try your original problem and see if the bug persists?

[s-broda]

My code is not compatible with 0.7 yet, so I can't test the original issue. I can confirm however that the MWE above is fixed on latest Windows nightly. Feel free to close.