Fix a crash when formatting some dicts with parenthesis-wrapped long
string keys. When LL[0] is an atom string, we need to check the atom
node's siblings instead of LL[0] itself, e.g.:
dictsetmaker
atom
STRING '"This is a really long string that can\'t be expected to fit in one line and is used as a nested dict\'s key"'
/atom
COLON ':'
atom
LSQB ' ' '['
listmaker
STRING '"value"'
COMMA ','
STRING ' ' '"value"'
/listmaker
RSQB ']'
/atom
COMMA ','
/dictsetmaker
Adds parentheses around implicit string concatenations when it's inside
a list, set, or tuple. Except when it's only element and there's no trailing
comma.
Looking at the order of the transformers here, we need to "wrap in
parens" before string_split runs. So my solution is to introduce a
"collaboration" between StringSplitter and StringParenWrapper where the
splitter "skips" the split until the wrapper adds the parens (and then
the line after the paren is split by StringSplitter) in another pass.
I have also considered an alternative approach, where I tried to add a
different "string paren wrapper" class, and it runs before string_split.
Then I found out it requires a different do_transform implementation
than StringParenWrapper.do_transform, since the later assumes it runs
after the delimiter_split transform. So I stopped researching that
route.
Originally function calls were also included in this change, but given
missing commas should usually result in a runtime error and the scary
amount of changes this cause on downstream code, they were removed in
later revisions.
Fixes#2734: a standalone comment causes strings to be merged into one far too long (and requiring two passes to do so).
Co-authored-by: Richard Si <63936253+ichard26@users.noreply.github.com>
uR is not a legal string prefix, so this test breaks (AssertionError:
cannot use --safe with this file; failed to parse source file AST:
invalid syntax) if changed to one in which the file is changed. I've
changed the last test to have u alone, and added an R to the test above
instead.
... in the middle of an expression or code block by adding a missing return.
Co-authored-by: Richard Si <63936253+ichard26@users.noreply.github.com>
Co-authored-by: Jelle Zijlstra <jelle.zijlstra@gmail.com>
When the Leaf node with `# fmt: skip` is a NEWLINE inside a `suite`
Node, the nodes to ignore should be from the siblings of the parent
`suite` Node.
There is a also a special case for the ASYNC token, where it expands
to the grandparent Node where the ASYNC token is.
This fixes GH-2646, GH-3126, GH-2680, GH-2421, GH-2339, and GH-2138.
The former was a regression I introduced a long time ago. To avoid
changing the stable style too much, the regression is only fixed if
--preview is enabled
Annoyingly enough, as we currently always enforce a second format pass if
changes were made, there's no good way to prove the existence of the
docstring quote normalization instability issue. For posterity, here's
one failing example:
--- source
+++ first pass
@@ -1,7 +1,7 @@
def some_function(self):
- ''''<text here>
+ """ '<text here>
<text here, since without another non-empty line black is stable>
- '''
+ """
pass
--- first pass
+++ second pass
@@ -1,7 +1,7 @@
def some_function(self):
- """ '<text here>
+ """'<text here>
<text here, since without another non-empty line black is stable>
"""
pass
Co-authored-by: Jelle Zijlstra <jelle.zijlstra@gmail.com>
Doing so is invalid. Note this only fixes the preview style since the
logic putting closing docstring quotes on their own line if they violate
the line length limit is quite new.
Co-authored-by: Jelle Zijlstra <jelle.zijlstra@gmail.com>
Otherwise they'd be deleted which was a regression in 22.1.0 (oops! my
bad!). Also type comments are now tracked in the AST safety check on all
compatible platforms to error out if this happens again.
Overall the line rewriting code has been rewritten to do "the right
thing (tm)", I hope this fixes other potential bugs in the code (fwiw I
got to drop the bugfix in blib2to3.pytree.Leaf.clone since now bracket
metadata is properly copied over).
Fixes#2873
This is a tricky one as await is technically an expression and therefore
in certain situations requires brackets for operator precedence.
However, the vast majority of await usage is just await some_coroutine(...)
and similar in format to return statements. Therefore this PR removes
redundant parens around these await expressions.
Co-authored-by: Jelle Zijlstra <jelle.zijlstra@gmail.com>
Co-authored-by: Richard Si <63936253+ichard26@users.noreply.github.com>
Allows us to better control placement of return annotations by:
a) removing redundant parens
b) moves very long type annotations onto their own line
Co-authored-by: Jelle Zijlstra <jelle.zijlstra@gmail.com>
It was causing stability issues because the first pass
could cause a "magic trailing comma" to appear, meaning
that the second pass might get a different result. It's
not critical.
Some things format differently (with extra parens)
Fixes#2651. Fixes#2754. Fixes#2518. Fixes#2321.
This adds a test that lists a number of cases of unstable formatting
that we have seen in the issue tracker. Checking it in will ensure
that we don't regress on these cases.
Since power operators almost always have the highest binding power in expressions, it's often more readable to hug it with its operands. The main exception to this is when its operands are non-trivial in which case the power operator will not hug, the rule for this is the following:
> For power ops, an operand is considered "simple" if it's only a NAME, numeric CONSTANT, or attribute access (chained attribute access is allowed), with or without a preceding unary operator.
Fixes GH-538.
Closes GH-2095.
diff-shades results: https://gist.github.com/ichard26/ca6c6ad4bd1de5152d95418c8645354b
Co-authored-by: Diego <dpalma@evernote.com>
Co-authored-by: Felix Hildén <felix.hilden@gmail.com>
Co-authored-by: Jelle Zijlstra <jelle.zijlstra@gmail.com>
*blib2to3's support was left untouched because: 1) I don't want to touch
parsing machinery, and 2) it'll allow us to provide a more useful error
message if someone does try to format Python 2 code.
* Treat functions/classes in blocks as if they're nested
One curveball is that we still want two preceding newlines before blocks
that are probably logically disconnected. In other words:
if condition:
def foo():
return "hi"
# <- aside: this is the goal of this commit
else:
def foo():
return "cya"
# <- the two newlines spacing here should stay
# since this probably isn't related
with open("db.json", encoding="utf-8") as f:
data = f.read()
Unfortunately that means we have to special case specific clause types
instead of just being able to just for a colon leaf. The hack used here
is to check whether we're adding preceding newlines for a standalone or
dependent clause. "Standalone" being a clause that doesn't need another
clause to be valid (eg. if) and vice versa.
Co-authored-by: Jelle Zijlstra <jelle.zijlstra@gmail.com>
