fix recursion detection in constant prop

[JeffBezanson]

#36058 was a bit too optimistic, and caused the inference stack overflow found here: JuliaGraphics/Gtk.jl#512 (comment)

This extra check fixes it while allowing the case in #36058 to still work.

[vtjnash]

Ah, right, I'd completely forgotten that the reason I had been avoiding this was due to the lack of a type-complexity limit. I'm somewhat confused though why this doesn't break the other test though.

[vtjnash]

Suggested change

	# if there might be a cycle, check to make sure we don't end up
	# calling ourselves here.
	infstate = sv
	while !(infstate === nothing)
	if method === infstate.linfo.def
	return Any
	end
	infstate = infstate.parent
	end
	# if edgecycle is true, then there is a cycle and we're calling ourself
	return Any

[JeffBezanson]

Well, that would just revert the original PR. What we need is to find out if adding the constant information breaks the cycle. Maybe we could do a full recursion check, but only look for other frames coming from constant prop?

[vtjnash]

Okay, right we could probably make edgecycle computation more lenient to allow one identical parent to have the constant-prop flag set and still be good here.

[JeffBezanson]

What is the best way to check that flag? any(result.overridden_by_const)?

[vtjnash]

Sounds about right

[vtjnash]

Ah, right, I'd completely forgotten that the reason I had been avoiding this was due to the lack of a recursion type-complexity limit. I'm somewhat confused though why this doesn't break the other test though, but I think it's because it's only looking through a subset of the parents (some of the parents are in the callers_in_cycle array, and checking both I think is equivalent to returning unconditionally here)

[JeffBezanson]

some of the parents are in the callers_in_cycle array, and checking both I think is equivalent to returning unconditionally here

I tried it, and that is correct as far as I can tell.