diff --git a/test/compiler/AbstractInterpreter.jl b/test/compiler/AbstractInterpreter.jl
index d816c3aa4647b..0d475a8259000 100644
--- a/test/compiler/AbstractInterpreter.jl
+++ b/test/compiler/AbstractInterpreter.jl
@@ -201,6 +201,23 @@ inner52938(x, types::Type, args...; kwargs...) = x
 outer52938(x) = @inline inner52938(x, Tuple{}; foo=Ref(42), bar=1)
 @test fully_eliminated(outer52938, (Any,); interp=Issue52938Interp(), retval=Argument(2))
 
+# https://github.com/JuliaGPU/CUDA.jl/issues/2241
+@newinterp Cuda2241Interp
+@MethodTable CUDA_2241_MT
+CC.method_table(interp::Cuda2241Interp) = CC.OverlayMethodTable(CC.get_inference_world(interp), CUDA_2241_MT)
+inner2241(f, types::Type, args...; kwargs...) = nothing
+function outer2241(f)
+    @inline inner2241(f, Tuple{}; foo=Ref(42), bar=1)
+    return nothing
+end
+# NOTE CUDA.jl overlays `throw_boundserror` in a way that causes effects, but these effects
+#      are ignored for this call graph at the `@assume_effects` annotation on `typejoin`.
+#      Here it's important to use `@consistent_overlay` to avoid tainting the `:nonoverlayed` bit.
+const cuda_kernel_state = Ref{Any}()
+@consistent_overlay CUDA_2241_MT @inline Base.throw_boundserror(A, I) =
+    (cuda_kernel_state[] = (A, I); error())
+@test fully_eliminated(outer2241, (Nothing,); interp=Cuda2241Interp(), retval=nothing)
+
 # Should not concrete-eval overlayed methods in semi-concrete interpretation
 @newinterp OverlaySinInterp
 @MethodTable OVERLAY_SIN_MT