loading REPL causes Compiler code to invalidate due to custom REPLInterpreter

[KristofferC]

Using the following script (run from outside a REPL):

using Pkg
using SnoopCompileCore
invalidations = @snoop_invalidations using REPL
using SnoopCompile
trees = SnoopCompile.invalidation_trees(invalidations)

we can see that loading the REPL causes a lot of invalidations from defining functions on a new AbstractInterpreter (REPLInterpreter) used for the REPL tab completions:

 inserting Compiler.OptimizationParams(interp::REPL.REPLCompletions.REPLInterpreter) @ REPL.REPLCompletions ~/julia/usr/share/julia/stdlib/v1.13/REPL/src/REPLCompletions.jl:596 invalidated:
   mt_backedges:  1: signature Tuple{Type{Compiler.OptimizationParams}, Compiler.AbstractInterpreter} triggered MethodInstance for Compiler.compileable_specialization(::Core.CodeInstance, ::Compiler.Effects, ::Compiler.InliningEdgeTracker, ::Compiler.CallInfo, ::Compiler.InliningState) (0 children)
                 ...
                 52: signature Tuple{Type{Compiler.OptimizationParams}, Compiler.AbstractInterpreter} triggered MethodInstance for Compiler.compileable_specialization(::Core.CodeInstance, ::Compiler.Effects, ::Compiler.InliningEdgeTracker, ::Compiler.CallInfo, ::Compiler.InliningState) (47 children)

 inserting cache_owner(::REPL.REPLCompletions.REPLInterpreter) @ REPL.REPLCompletions ~/julia/usr/share/julia/stdlib/v1.13/REPL/src/REPLCompletions.jl:599 invalidated:
   mt_backedges:  1: signature Tuple{typeof(Compiler.cache_owner), Compiler.AbstractInterpreter} triggered MethodInstance for Compiler.edge_matches_sv(::Compiler.AbstractInterpreter, ::Compiler.InferenceState, ::Method, ::Type, ::Core.SimpleVector, ::Bool, ::Compiler.InferenceState) (0 children)
                ...
                 19: signature Tuple{typeof(Compiler.cache_owner), Compiler.AbstractInterpreter} triggered MethodInstance for Compiler.is_same_frame(::Compiler.NativeInterpreter, ::Core.MethodInstance, ::Compiler.InferenceState) (37 children)
                 20: signature Tuple{typeof(Compiler.cache_owner), Compiler.AbstractInterpreter} triggered MethodInstance for Compiler.code_cache(::Compiler.AbstractInterpreter) (64 children)

 inserting Compiler.InferenceParams(interp::REPL.REPLCompletions.REPLInterpreter) @ REPL.REPLCompletions ~/julia/usr/share/julia/stdlib/v1.13/REPL/src/REPLCompletions.jl:595 invalidated:
   mt_backedges:  1: signature Tuple{Type{Compiler.InferenceParams}, Compiler.AbstractInterpreter} triggered MethodInstance for Compiler.abstract_eval_partition_load(::Compiler.AbstractInterpreter, ::Core.BindingPartition) (0 children)
                  ...                 
                 41: signature Tuple{Type{Compiler.InferenceParams}, Compiler.AbstractInterpreter} triggered MethodInstance for Compiler.get_max_methods(::Compiler.AbstractInterpreter) (682 children)


inserting Compiler.InferenceParams(interp::REPL.REPLCompletions.REPLInterpreter) @ REPL.REPLCompletions ~/julia/usr/share/julia/stdlib/v1.13/REPL/src/REPLCompletions.jl:595 invalidated:
   mt_backedges:  1: signature Tuple{Type{Compiler.InferenceParams}, Compiler.AbstractInterpreter} triggered MethodInstance for Compiler.abstract_eval_partition_load(::Compiler.AbstractInterpreter, ::Core.BindingPartition) (0 children)
                ...
                 40: signature Tuple{Type{Compiler.InferenceParams}, Compiler.AbstractInterpreter} triggered MethodInstance for Compiler.abstract_eval_partition_load(::Compiler.AbstractInterpreter, ::Core.BindingPartition) (90 children)
                 41: signature Tuple{Type{Compiler.InferenceParams}, Compiler.AbstractInterpreter} triggered MethodInstance for Compiler.get_max_methods(::Compiler.AbstractInterpreter) (682 children)

The root cause seems to be a type instability in the Compiler code:

Looking with code_warntype:

julia> code_warntype(Core.Compiler.typeinf, Tuple{Compiler.NativeInterpreter, Compiler.InferenceState})
MethodInstance for Compiler.typeinf(::Compiler.NativeInterpreter, ::Compiler.InferenceState)
  from typeinf(interp::Compiler.AbstractInterpreter, frame::Compiler.InferenceState) @ Compiler ~/julia/usr/share/julia/Compiler/src/abstractinterpretation.jl:4363
Arguments
  #self#::Core.Const(Compiler.typeinf)
  interp@_2::Compiler.NativeInterpreter
  frame::Compiler.InferenceState
Locals
...
  interp@_20::Compiler.AbstractInterpreter <--------------------------------

This comes from the access at

julia/Compiler/src/abstractinterpretation.jl

Line 4385 in f211a77

interp = callee.interp

where is is accessing the untyped field defined at

julia/Compiler/src/inferencestate.jl

Lines 317 to 319 in f211a77

	# The interpreter that created this inference state. Not looked at by
	# NativeInterpreter. But other interpreters may use this to detect cycles
	interp::AbstractInterpreter

The comment here about "Not looked at by NativeInterpreter" is a bit odd since this is indeed code for a NativeInterpreter.

Is there anything that can be said about this interpreter from the code? Is it guaranteed to be the same type as the interp argument?