Mypy: Crash with global type aliasing

[ethanhs]

This was originally discovered through the connection package in cpython, but I have a more minimal repro here:

def foo():
    global myglobal
    from typing import Any  # can be any type, as long as the next line is a type alias
    myglobal = Any

It seems that when we look it up, it is in the global declarations, but not not found in the globals.

Traceback:

test.py:4: error: INTERNAL ERROR -- please report a bug at https://github.com/python/mypy/issues version: 0.600+dev-36b036bd966896f9c8ab9773f89763766e485792
Traceback (most recent call last):
  File "C:\Python36\lib\runpy.py", line 193, in _run_module_as_main
    "__main__", mod_spec)
  File "C:\Python36\lib\runpy.py", line 85, in _run_code
    exec(code, run_globals)
  File "C:\Users\ethanhs\Documents\mypy\mypy\__main__.py", line 11, in <module>
    main(None)
  File "C:\Users\ethanhs\Documents\mypy\mypy\main.py", line 91, in main
    res = type_check_only(sources, bin_dir, options, flush_errors, fscache)  # noqa
  File "C:\Users\ethanhs\Documents\mypy\mypy\main.py", line 148, in type_check_only
    fscache=fscache)
  File "C:\Users\ethanhs\Documents\mypy\mypy\build.py", line 185, in build
    flush_errors, fscache)
  File "C:\Users\ethanhs\Documents\mypy\mypy\build.py", line 285, in _build
    graph = dispatch(sources, manager)
  File "C:\Users\ethanhs\Documents\mypy\mypy\build.py", line 2420, in dispatch
    process_graph(graph, manager)
  File "C:\Users\ethanhs\Documents\mypy\mypy\build.py", line 2713, in process_graph
    process_stale_scc(graph, scc, manager)
  File "C:\Users\ethanhs\Documents\mypy\mypy\build.py", line 2823, in process_stale_scc
    graph[id].semantic_analysis()
  File "C:\Users\ethanhs\Documents\mypy\mypy\build.py", line 2038, in semantic_analysis
    self.manager.semantic_analyzer.visit_file(self.tree, self.xpath, self.options, patches)
  File "C:\Users\ethanhs\Documents\mypy\mypy\semanal.py", line 297, in visit_file
    self.accept(d)
  File "C:\Users\ethanhs\Documents\mypy\mypy\semanal.py", line 3235, in accept
    node.accept(self)
  File "C:\Users\ethanhs\Documents\mypy\mypy\nodes.py", line 583, in accept
    return visitor.visit_func_def(self)
  File "C:\Users\ethanhs\Documents\mypy\mypy\semanal.py", line 366, in visit_func_def
    self._visit_func_def(defn)
  File "C:\Users\ethanhs\Documents\mypy\mypy\semanal.py", line 427, in _visit_func_def
    self.analyze_function(defn)
  File "C:\Users\ethanhs\Documents\mypy\mypy\semanal.py", line 652, in analyze_function
    defn.body.accept(self)
  File "C:\Users\ethanhs\Documents\mypy\mypy\nodes.py", line 865, in accept
    return visitor.visit_block(self)
  File "C:\Users\ethanhs\Documents\mypy\mypy\semanal.py", line 1537, in visit_block
    self.accept(s)
  File "C:\Users\ethanhs\Documents\mypy\mypy\semanal.py", line 3235, in accept
    node.accept(self)
  File "C:\Users\ethanhs\Documents\mypy\mypy\nodes.py", line 913, in accept
    return visitor.visit_assignment_stmt(self)
  File "C:\Users\ethanhs\Documents\mypy\mypy\semanal.py", line 1616, in visit_assignment_stmt
    self.check_and_set_up_type_alias(s)
  File "C:\Users\ethanhs\Documents\mypy\mypy\semanal.py", line 1737, in check_and_set_up_type_alias
    assert node is not None
AssertionError:
test.py:4: : note: use --pdb to drop into pdb

[onlined]

According to your example, should myglobal an alias of Any from the start of the file, or after the first call to foo?

[ethanhs]

Well, I think the global statement wouldn't be executed until foo is called, so I would say the name gets added to the global namespace after the first call to foo. But I could be wrong.

[gvanrossum]

I'd say what happens at runtime is irrelevant. I think mypy should most likely not ever consider this as a definition of an alias (global or otherwise); it should perhaps flag this as an error (though arguably it should just leave it alone); and it should definitely not crash.

[onlined]

So I think that we (actually not we, the lead developers) should agree on a
decision, and then this crash can be fixed according to that decision.

By the way, I agree with @gvanrossum. It should be ignored. It should cause
neither an error nor a crash.

[gvanrossum]

OK, so extending the example:

def foo():
    global myglobal
    from typing import Any  # can be any type, as long as the next line is a type alias
    myglobal = Any  # NO ERROR HERE
    x: myglobal  # THIS WOULD BE AN ERROR

y: myglobal  # THIS WOULD ALSO BE AN ERROR

I can live with that. It's somewhat similar to this example (which passes without errors):

class C:
      x: object

C.x = Any

[onlined]

I think that we should do the same with nonlocal.

[onlined]

I think #5732 is related to this issue.

[onlined]

Since #5732, @gvanrossum's example actually turns into this:

def foo():
    global myglobal
    from typing import Any  # can be any type, as long as the next line is a type alias
    myglobal = Any  # THIS WOULD BE AN ERROR TOO
    x: myglobal  # THIS WOULD BE AN ERROR

y: myglobal  # THIS WOULD ALSO BE AN ERROR

How should we proceed? Fixing the other issue before this, or fixing this with accepting the behaviour mentioned in the other one?

[JukkaL]

This seems to work now. It may have been fixed by the new semantic analyzer.