Since the annotations are processed just like all other expressions in the symbol table, the generated entries for functions etc. This would result with 

def foo():
    for number in range(5):
        foo: (yield number)
    return number

foo()

returning a generator / coroutine (depending on yield/yield from/await usage). Is this something we want to keep or maybe tweak the symbol table generator to not to handle annotations (since there are also more subtle issues regarding analysis of cell / free vars).

Ouch.  I think we should generate a SyntaxError if yield [from] is used in an annotation.

This is another side effect of processing annotations (at the symbol table construction stage) (and I would assume there are a few more cases like this);

def foo():
    outer_var = 1

    def bar():
        inner_var: outer_var = T
    
    return bar

inner = foo()
print(inner.__closure__)

In theory, there shouldn't be any cells / references to the variables from outer scope, but since we process the entry for the annotation and record `outer_var` as a free var it is listed here.

The difference is that that just causes a slight inefficiency, while processing 'yield' makes the function a generator. So I don't feel as strongly about that.

> So I don't feel as strongly about that.

Just to note, since I believe the solution for all this might be the same (not processing annotations at all, since they will be compiled to strings in the later stage). If we go down on that route, it will be simpler but we won't be able to raise SyntaxError's for "a: (yield)" / "b: (await/yield from x)". 

By the way, this is what happens if you try to use get_type_hints on a function where an argument is (yield):

>>> typing.get_type_hints(a)
Traceback (most recent call last):
  File "/usr/local/lib/python3.10/typing.py", line 537, in __init__
    code = compile(arg, '<string>', 'eval')
  File "<string>", line 1
SyntaxError: 'yield' outside function

During handling of the above exception, another exception occurred:

Traceback (most recent call last):
  File "<stdin>", line 1, in <module>
  File "/usr/local/lib/python3.10/typing.py", line 1490, in get_type_hints
    value = ForwardRef(value)
  File "/usr/local/lib/python3.10/typing.py", line 539, in __init__
    raise SyntaxError(f"Forward reference must be an expression -- got {arg!r}")
SyntaxError: Forward reference must be an expression -- got '(yield)'

If we simply ignored yield in the annotation, wouldn't we have the problem that

def f(a: (yield)): pass

silently changes from being a generator (in 3.9) to not being a generator (in 3.10)? That would be bad. I'd rather make this an error still. (But for nonlocals, not processing sounds fine.)

I concur with Guido. I feel that making this an error is the best alternative we have.

Okay, let's do it. It should probably a post-parse check (before we invoke the bytecode compiler but after we have the AST in hand).

I have a patch ready to go, but I'd prefer to block this issue until issue 42737 is resolved/decided.

One thing to note here, currently Pablo and I are trying to bring annotation unparsing from the compiler to the parser in issue 41967. If we do so, the annotations won't cause any side effects on the symbol table generation.

I've also opened #42837 which is about fixing the symbol table, so that it is correct w.r.t. to current behavior.

I'd like to fix it ASAP as the compiler should be able to rely on the symbol table being correct.

Of course, once we have decided what the behavior should be, then it may need to be updated again.

I'm inclined to agree that 'yield' in an annotation should probably be a syntax error, but I haven't put much thought into to.

Yield in an annotation should be a syntax error.

What's the process for making a decision on whether to make 'yield' in an annotation a syntax error?

As a language change it should have a PEP, IMO.
The PEP will be short, and shouldn't need a long-winded acceptance process.
I just think that a PEP is more visible to the community than the bug tracker.

I'm happy to write the PEP.

Should not "await" and "async for" (in comprehesions) and ":=" be forbidden too?

I wouldn’t have thought you’d need a PEP for this but if you want to write one that sounds like the right thing.

Re: async and walrus: I think those are different, their presence doesn’t affect the meaning of the function like yield. We can’t hope to prevent side effects syntactically.

Does not walrus affect the meaning of variable? And await affects the meaning of generator expression.

Oh, you’re right about walrus. And I don’t actually understand async generator expressions.

This suggests that we definitely need a PEP. :-)

Draft PEP here
https://github.com/markshannon/peps/blob/annotation-syntax-errors/pep-9999.rst

Guido, would you like to be co-author as it was your idea to make these things a syntax error?

It'll probably be quicker if you leave me out of the loop (feel free to quote me though). I am currently quite overwhelmed with PEP-sized discussions. I expect that the SC can rule on this quickly (just use their tracker to send them the PEP).

> The use of nonlocal variables in an annotation will be a syntax error in Python 3.10

What is the reasoning for forbidding nonlocal variables (https://bugs.python.org/msg383659), can't we just treat them like regular variables and leave the value to whom needed to deal with resolving the scope? 

Also, you should preferably clarify other cases regarding the symbol table interaction of annotations. For example this case:

Here is one;
class T:
    def foo(self):
        a: super().bar() = x
print(T.foo.__closure__)

And if we are on the road to writing a PEP, maybe we should somehow squeeze issue 42737 somewhere, since the annotations for complex targets are still evaluated and this makes everything a bit problematic and inconsistent. If you need collaboration on the PEP, just let me know (isidentical@gmail.com) (I have a patch ready to go for the symbol table to both make annotations ineffective and forbid this stuff but was waiting for issue 42737)

I'm not keen on prohibiting nonlocal variable reference in annotations,
since I can imagine uses for this (e.g. in tests). I'm not too worried
about keeping the cell alive "forever", since with Larry's PEP 649 their
lifetime would end when the object containing the annotation ends.

Someone should look into how Batuhan's super()-in-annotation example would
behave under Larry's PEP 649. (Using super() in the magical function Larry
proposes to generate might not work anyway, since super()'s own magical
powers don't work in a nested function.)

Re: issue42737, not sure what to do for that, since neither PEP 563 nor PEP
649 gives a way to access those annotations. Then again, that's also true
for annotations on local variables. So maybe we can treat them the same?
Type checkers won't care, and runtime uses of annotations (whether as type
hints or otherwise) won't have access......

I've posted an entry on python-dev to collect comments about how we should act (whether just act them as strings on the symbol table or just forbid them completely): https://mail.python.org/archives/list/python-dev@python.org/thread/5NQH5GURAYW6XINPBO6VSF45CWLQOHUQ/

Since we are almost days away from the cut, I will probably just submit a proper patch this weekend where we can debate more through reviews but first I need to know what kind of action we intend to take. (this all assumes PR 23952 is goes in).

I think stringized annotations should prohibit the same things PEP 649 prohibits: walrus, yield / yield from, and await.

This was easy in my 649 branch; walrus adds locals, and yield / yield from make it a generator.  So the code raises an error if the generated annotations code object has locals or is a generator.  I don't think I had to do anything special to prohibit await, because that's only valid in a function declared "async def", which annotations code objects are not.

> This was easy in my 649 branch; walrus adds locals, and yield / yield from make it a generator.  So the code raises an error if the generated annotations code object has locals or is a generator.  I don't think I had to do anything special to prohibit await, because that's only valid in a function declared "async def", which annotations code objects are not.

The implementation I have right now just adds a new state to the symbol table, and raises error in the visitors (like switch Yield, YieldFrom etc).

Wouldn't it be easier to just throw an exception in the stringizing code?  I note that there are dedicated functions to generate walrus, yield, yield from, and await in Python/ast_unparse.c.  In theory this is a general-purpose facility, but in practice it's not part of the limited API, and it's called from literally one place, the "stringize this annotation" call site inside compile.c.

So shouldn't we just rule out some specific bits of syntax in annotations? That can be done after the PEG grammar has accepted them, otherwise we'd end up having a whole new expression-like grammar section that excludes them. I think this is the list we should exclude, right?

- yield [from]
- walrus
- await (?)

I agree await technically doesn't need to be in this list, so maybe we shouldn't explicitly exclude it -- it's no different than writing

def f(x: open(filename).read()):
    ...

> That can be done after the PEG grammar has accepted them

But wouldn't we still end up with maintaining a custom flag to see if we are in annotation (e.g a: Something((yield) + 2)) and act upon that which would seem to do a bit messy in the grammar actions rather than just throwing syntax errors from switch/cases inside of symbol table?

I don’t really care how you do it, I just thought having a whole “expr-without-walrus-or-yield” subgrammar would be tedious.

If there is a way to have the “in-an-annotation” flag set during parsing that may be better, I don’t know.

Maybe Lysandros has an idea?

There is no “expr-without-await-or-async-for” subgrammar, but "await" and asynchronous comprehensions are invalid in synchronous functions. There should be similar flag for annotations in symtable.c.

>  I just thought having a whole “expr-without-walrus-or-yield” subgrammar would be tedious.

Yeah I have experimented with something similar in the past and this becomes out of hand super quickly and leafs to a lot of duplication.

Similarly, a flag that conditionally accept some rules can be quite tricky to hey right with the backtracking and those require some extra metagrammar so adding the flag is not super unreadable. 

In my humble opinion, I think is far simpler to reject this after the parsing is done, either in the compiler or the symbol table.

New changeset ad106c68eb00f5e4af2f937107baff6141948cee by Batuhan Taskaya in branch 'master':
bpo-42725: Render annotations effectless on symbol table with PEP 563 (GH-25583)
https://github.com/python/cpython/commit/ad106c68eb00f5e4af2f937107baff6141948cee

Batuhan, can this issue be closed?

> Batuhan, can this issue be closed?

No, I still need to take care of what's new entry.

New changeset 90d584a2ab91cc0f30e90a0a99f8c1447eb51302 by Miss Islington (bot) in branch '3.10':
bpo-42725: mention the changes on what's new (GH-25974)
https://github.com/python/cpython/commit/90d584a2ab91cc0f30e90a0a99f8c1447eb51302