By only allowing the sink in the state where encryption v1 is used, we
can handle the new case where the order of attribute assignment is
flipped.
However, we get a few too many paths because we can have multiple
sources reaching the same sink... let's fix in next commit.
Since we can analyze operator.py from Python3, but not in Python 2
(since it's implemented in C), we get a difference for the index tests.
note: `operator.length_hint` is only available in Python 3.4 and later,
so would always fail under Python 2.
this illustrates that the function implementing
the comprehension does not capture `mod_local`.
We could handle this case specially, by having
a different implementation for `for`, but the
wider issue would remain.
This is a bit of an edge case, but allowed. Since we currently don't
provide information on positional only arguments, we can't do much to
solve it right now.
It's nice that it fixes the `InsecureProtocol` test-case (which maybe
should have been a test-case for the import resolution library in the
first place?)
But it's not quite right:
1. it adds spurious flow for `clashing_attr`
2. it runs into huge problems for typetracking_imports/tracked.expected
3. it runs into the problem for
https://github.com/github/codeql/pull/10176 with an `from <pkg>
import *` blocking flow from previously defined variable, that is NOT
overridden. (simplistic_reexport.bar_attr)
Just like the one added for `py/insecure-protocol` in fb425b7, but
instead added in the import-resolution tests, such that we don't have to
remember it's in a completely different directory.
It's not very useful to look at, and it's a mess when you change any
tests to see all the changes lines in the expected output that you
really do not care about!
I guess we could have done this at the very start of introducing this
test in this PR, but I think the last commit was mostly inspired from
looking at all the things that evidently was re-exported from the trace
import, even when I knew they were not available because of the
`__all__` definition.
However, we can see that `from <pkg> import *` and `import pkg` are
handled differently. Would have liked `has_defined_all_indirection` to
behave in the same way no matter how the import was made.
Notice that `has_defined_all_indirection` all have both
`all_defined_bar_copy` and `all_defined_foo_copy` marked as exported,
even though only `all_defined_foo_copy` is available.
For `from <pkg> import <attr>` we would use to treat the `<pkg>`
(ImportExpr) as a definition of the name `<attr>`.
Since this removes bad import-flow, and nothing broke, I'm guessing this
was never intentional.
