is_primitive doesn't invalidate methods in TICache

[jumerckx]

I was looking at how is_primitive is implemented because I need similar functionality in a project.
Currently, by default, a new TapirInterpreter gets a fresh new cache.
However, when reusing the same interpreter (with the same cache), redefinitions of is_primitive rules don't invalidate the methods in the cache:

using Tapir
@inline f(x) = g(x)
@inline g(x) = h(x)
@inline h(x) = sin(x)

interp = Tapir.TapirInterpreter()
sig = Tuple{typeof(f), Float64}

Tapir.is_primitive(::Type{Tapir.DefaultCtx}, ::Type{<:Tuple{typeof(h), Any}}) = true
Base.code_ircode_by_type(sig; interp)[1][1] # h is not inlined, good

Tapir.is_primitive(::Type{Tapir.DefaultCtx}, ::Type{<:Tuple{typeof(h), Any}}) = false
Base.code_ircode_by_type(sig; interp)[1][1] # h is *still* not inlined

I believe this is unexpected but could be wrong?
The tests only check nested calls one level deep but at that point, the problem doesn't seem to occur.

This is a related slack thread where I tried getting help for my problem, with an MWE in using Julia 1.11 that doesn't depend on Tapir

[willtebbutt]

Yup, this is expected behaviour unfortunately. You need a fresh interp whenever you re-define is_primitive (and possibly anything else) because interp doesn't know about world age etc.

If you find a way to handle this in general, I would very much like to hear about it!

I'm going to close this for now because I don't anticipate that there's anything that I can do about this / it's a problem that I know about + plan to document, but please do feel free to re-open if you figure something out.