Mysterious inference failure

[nalimilan]

This bug has appeared in the last few days on master. Depending on the order in which you call f and g, you get different results. In all cases, the Any result is unexpected and didn't happen before.

First call order:

julia> f{S,T}(x::S, y::T) = Base.promote_op(^, S, T)
f (generic function with 1 method)

julia> f(Nullable(0.0f0), Nullable(BigInt(0)))
Nullable{Any}

julia> g{S,T}(x::Nullable{S}, y::Nullable{T}) = Base.promote_op(^, S, T)
g (generic function with 1 method)

julia> g(Nullable(0.0f0), Nullable(BigInt(0)))
Float32

Second call order:

julia> g{S,T}(x::Nullable{S}, y::Nullable{T}) = Base.promote_op(^, S, T)
g (generic function with 1 method)

julia> g(Nullable(0.0f0), Nullable(BigInt(0)))
Any

julia> f{S,T}(x::S, y::T) = Base.promote_op(^, S, T)
f (generic function with 1 method)

julia> f(Nullable(0.0f0), Nullable(BigInt(0)))
Nullable{Float32}

[pabloferz]

Sorry, my fault. We need promote_op inlined in some places and I mistakenly marked as pure instead of forcing inlining. See #19273