bpo-27999 invalidated test_global_err_then_warn in Lib/test/test_syntax.py:567. That test was added in bpo-763201 for testing that SyntaxWarning emitted in symtable_global() doesn't clobber a SyntaxError. In bpo-27999 a SyntaxWarning was converted to SyntaxError, and the test no longer serves its purpose.

bpo-28512 makes tests more strong (it tests the position of SyntaxError), but it can't be applied in 3.6, because _check_error() catches different SyntaxError.

Serhiy, here is a patch that might be helpful. It detects global-and-parameter errors sooner (when possible). This will cause the following:

>>> if 1:
...     def error(a):
...         global a
...     def error2():
...         b = 1
...         global b
... 
  File "<stdin>", line 3
SyntaxError: name 'a' is parameter and global

However, in more complex (nested) cases, the global-after-assign will still be detected sooner:

>>> def error(a):
...     def inner():
...         global a
...     def inner2():
...         b = 1
...         global b
... 
  File "<stdin>", line 6
SyntaxError: name 'b' is assigned to before global declaration

Maybe there is a way to delay the detection of this error until second pass in symtable.c, but don't see now how to do this.

I don't know what should we do with this. Is there a bug that needs to be fixed? Or maybe the test can be just removed?

I don't think this is a bug, but detecting first a SyntaxError that appears textually first might be seen as an improvement. I would say such behaviour seems more intuitive. A possible downside could be a (probably very minor) slow-down of compilation.

I don't have any strong opinion here.

Updated patch as proposed by Serhiy.

Do you mind to create a pull request Ivan?

OK, I made a PR with the fix and a test that checks the line number for syntax error (so that the original purpose test_global_err_then_warn is preserved).

Am I needed here?

Am I needed here?

As I understand, Serhiy is going to review the PR, so I think no.

Sorry for disturbing you Guido. I had added you as the committer of bpo-27999. Do you have thoughts about this issue?

I think this is very minor but if you two can agree that the code is right I think it's a nice little improvement, and I like that that particular test's usefulness is restored.

PS. Long-term we should really build error recovery into our antiquated parser and report as many errors as we can, without cascading. But that's probably a Python 4 project.

Is it correct that the parameter can be annotated in the function body?

def f(x):
    x: int

Or that the local variable can be annotated after assignment?

def f():
    x = 1
    x: int

Those seem things that the type checker should complain about, but I don't think Python should.

Is it correct that the parameter can be annotated in the function body?

I agree with Guido, this is rather a task for type checkers.

New changeset 8c83c23 by Serhiy Storchaka (Ivan Levkivskyi) in branch 'master':
bpo-28936: Detect lexically first syntax error first (bpo-4097)
8c83c23

Thank you for your patch Ivan.