The following code fails:

    >>> lcls = {'w': 100}
    >>> eval('[w for x in ("hello", "world")]', None, lcls)
    Traceback (most recent call last):
      File "<stdin>", line 1, in <module>
      File "<string>", line 1, in <module>
      File "<string>", line 1, in <listcomp>
    NameError: name 'w' is not defined
    >>> eval('[w, w, w]', None, lcls)
    [100, 100, 100]

whereas in python2 it succeeds

    >>> lcls = {'w': 100}
    >>> eval('[w for x in ("hello", "world")]', None, lcls)
    [100, 100]

I test it on 3.7 and 3.8 and have the same problem

I don't know if this a bug for py3 or py2 because the the w variable
is not defined on the list comprehension context.

Maybe the locals on `eval` must be create w on locals context

You example is a list comprehension, but no matter.  In 3.x, the value of a comprehension is the result of calling a temporary function.  Functions access globals and nonlocals but not non-global locals of surrounding contexts.  Class locals are an example of the latter.

>>> class C:
...   w = 100
...   l = [w for x in ("hello", "world")]
...
Traceback (most recent call last):
  File "<stdin>", line 1, in <module>
  File "<stdin>", line 3, in C
  File "<stdin>", line 3, in <listcomp>
NameError: name 'w' is not defined

To see this, one must make sure that the name in question is not also in globals, as it is below.

>>> w = 50
>>> [w for x in ("hello", "world")]
[50, 50]
>>> class C:
...   w = 100
...   l = [w for x in ("hello", "world")]
...
>>> C.l
[50, 50]  # Evaluated global w, not local w.

When one calls eval with separate globals and locals, one is simulating class context.  There should be a FAQ entry about this.

More ways to be bitten by this strange behavior:

    >>> d = {'a': 1, 'b': 2}
    >>> eval('[x[k] for k in x]', {}, {'x': d})
    Traceback (most recent call last):
      File "<stdin>", line 1, in <module>
      File "<string>", line 1, in <module>
      File "<string>", line 1, in <listcomp>
    NameError: name 'x' is not defined
    >>> 
    >>> def f():
    ...     x = {'a': 1, 'b': 2}
    ...     return eval('[x[k] for k in x]')
    ... 
    >>> f()
    Traceback (most recent call last):
      File "<stdin>", line 1, in <module>
      File "<stdin>", line 3, in f
      File "<string>", line 1, in <module>
      File "<string>", line 1, in <listcomp>
    NameError: name 'x' is not defined



In both of the above cases changing

    eval('[x[k] for k in x]')

to

    eval('[v for v in x.values()]')

avoids the problem. There are no problems when using

    [x[k] for k in x]

without `eval()` either. I'd prefer this to be changed, but there should at least be a note in the documentation of `eval()` about this.

I encountered this issue because Robot Framework -- a generic Python based test automation framework -- supports evaluating Python expressions and this issue was reported for us:
https://github.com/robotframework/robotframework/issues/3207