* 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>
The implementation of the new backtracking logic depends heavily on deepcopying the current state of the parser before seeing one of the new keywords, which by default is an very expensive operations. On my system, formatting these 3 files takes 1.3 seconds.
```
$ touch tests/data/pattern_matching_*; time python -m black -tpy310 tests/data/pattern_matching_* 19ms
All done! ✨🍰✨
3 files left unchanged.
python -m black -tpy310 tests/data/pattern_matching_* 2,09s user 0,04s system 157% cpu 1,357 total
```
which can be optimized 3X if we integrate the existing copying logic (`clone`) to the deepcopy system;
```
$ touch tests/data/pattern_matching_*; time python -m black -tpy310 tests/data/pattern_matching_* 1ms
All done! ✨🍰✨
3 files left unchanged.
python -m black -tpy310 tests/data/pattern_matching_* 0,66s user 0,02s system 147% cpu 0,464 total
```
This still might have some potential, but that would be way trickier than this initial patch.
* Improve Python 2 only syntax detection
First of all this fixes a mistake I made in Python 2 deprecation PR
using token.* to check for print/exec statements. Turns out that
for nodes with a type value higher than 256 its numeric type isn't
guaranteed to be constant. Using syms.* instead fixes this.
Also add support for the following cases:
print "hello, world!"
exec "print('hello, world!')"
def set_position((x, y), value):
pass
try:
pass
except Exception, err:
pass
raise RuntimeError, "I feel like crashing today :p"
`wow_these_really_did_exist`
10L
* Add octal support, more test cases, and fixup long ints
Co-authored-by: Jelle Zijlstra <jelle.zijlstra@gmail.com>
Co-authored-by: Jelle Zijlstra <jelle.zijlstra@gmail.com>
Implementation stolen from PR davidhalter/parso#162. Thanks parso!
I could add support for these newer syntactical constructs in the
target version detection logic, but until I get diff-shades up
and running I don't feel very comfortable adding the code.
This fixes a bug where a trailing comma would be added to a
parenthesized return annotation changing its type to a tuple.
Here's one case where this bug shows up:
```
def spam() -> (
this_is_a_long_type_annotation_which_should_NOT_get_a_trailing_comma
):
pass
```
The root problem was that the type annotation was treated as if it was
a parameter & import list (is_body=True to linegen::bracket_split_build_line)
where a trailing comma is usually fine. Now there's another check in the
aforementioned function to make sure the body it's operating on isn't
a return annotation before truly adding a trailing comma.
we don't accidentally add backslashes to them when normalizing quotes
because that's invalid syntax!
The problem this commit fixes is that matches would eat too much
blocking important matches to occur. For example, here's one f-string
body:
{a}{b}{c}
I know there's no risk of introducing backslashes here, but the regex
already goes sideways with this. Throwing this example at regex101
I get:
{a}{b}{c} # The As and Bs are the two matches, and the upper
---- ---- # case letters are the groups with those matches.
aAaa bbBb
... we've missed the middle expression (so if any backslashes in a
more complex example were introduced there we wouldn't bail out
even though we should -- hence the bug). As it stands the regex
needs somesort of extra character (or the start/end of the body)
around the expressions but that isn't always the case as shown
above.
The fix implemented here is to turn the "eat a surrounding non-curly
bracket character" groups ie. `(?:[^{]|^)` and `(?:[^}]|$)` into
negative lookaheads and lookbehinds. This still guarantees the
already specified rules but without problematically eating extra
characters ^^
Fixes#2359.
This commit now makes Black exit with an user-friendly error message if a
.gitignore file couldn't be parsed -- a massive improvement over an opaque
traceback!
To summarise, based on what was discussed in that issue:
due to not being able to parse automagics (e.g. pip install black)
without a running IPython kernel, cells with syntax which is parseable
by neither ast.parse nor IPython will be skipped cells with multiline
magics will be skipped trailing semicolons will be preserved, as they
are often put there intentionally in Jupyter Notebooks to suppress
unnecessary output
Commit history before merge (excluding merge commits):
* wip
* fixup tests
* skip tests if no IPython
* install test requirements in ipynb tests
* if --ipynb format all as ipynb
* wip
* add some whole-notebook tests
* docstrings
* skip multiline magics
* add test for nested cell magic
* remove ipynb_test.yml, put ipynb tests in tox.ini
* add changelog entry
* typo
* make token same length as magic it replaces
* only include .ipynb by default if jupyter dependencies are found
* remove logic from const
* fixup
* fixup
* re.compile
* noop
* clear up
* new_src -> dst
* early exit for non-python notebooks
* add non-python test notebook
* add repo with many notebooks to black-primer
* install extra dependencies for black-primer
* fix planetary computer examples url
* dont run on ipynb files by default
* add scikit-lego (Expected to change) to black-primer
* add ipynb-specific diff
* fixup
* run on all (including ipynb) by default
* remove --include .ipynb from scikit-lego black-primer
* use tokenize so as to mirror the exact logic in IPython.core.displayhooks quiet
* fixup
* 🎨
* clarify docstring
* add test for when comment is after trailing semicolon
* enumerate(reversed) instead of [::-1]
* clarify docstrings
* wip
* use jupyter and no_jupyter marks
* use THIS_DIR
* windows fixup
* perform safe check cell-by-cell for ipynb
* only perform safe check in ipynb if not fast
* remove redundant Optional
* 🎨
* use typeguard
* dont process cell containing transformed magic
* require typing extensions before 3.10 so as to have TypeGuard
* use dataclasses
* mention black[jupyter] in docs as well as in README
* add faq
* add message to assertion error
* add test for indented quieted cell
* use tokenize_rt else we cant roundtrip
* fmake fronzet set for tokens to ignore when looking for trailing semicolon
* remove planetary code examples as recent commits result in changes
* use dataclasses which inherit from ast.NodeVisitor
* bump typing-extensions so that TypeGuard is available
* bump typing-extensions in Pipfile
* add test with notebook with empty metadata
* pipenv lock
* deprivative validate_cell
* Update README.md
* Update docs/getting_started.md
* dont cache notebooks if jupyter dependencies arent found
* dont write to cache if jupyter deps are not installed
* add notebook which cant be parsed
* use clirunner
* remove other subprocess calls
* add docstring
* make verbose and quiet keyword only
* 🎨
* run second many test on directory, not on file
* test for warning message when running on directory
* early return from non-python cell magics
* move NothingChanged to report to avoid circular import
* remove circular import
* reinstate --ipynb flag
Co-authored-by: Jelle Zijlstra <jelle.zijlstra@gmail.com>
`black.strings.get_string_prefix` used to lowercase the extracted
prefix before returning it. This is wrong because 1) it ignores the
fact we should leave R prefixes alone because of MagicPython, and 2)
there is dedicated prefix casing handling code that fixes issue 1.
`.lower` is too naive.
This was originally fixed in 20.8b0, but was reintroduced since 21.4b0.
I also added proper prefix normalization for docstrings by using the
`black.strings.normalize_string_prefix` helper.
Some more test strings were added to make sure strings with capitalized
prefixes aren't treated differently (actually happened with my original
patch, Jelle had to point it out to me).
This commit adds a short section discussing the non-processing of docstrings
besides spacing improvements, mentions comment moving and links to the
AST equivalence discussion. I also added a simple spacing test for good
measure.
Commit history before merge:
* Mention comment non-processing in documentation, add spacing test
* Mention special cases for comment spacing
* Add all special cases, improve wording
Not sure the fix is right. Here is what I found: issue is connected
with line
first.prefix = prefix[comment.consumed :]
in `comments.py`. `first.prefix` is a prefix of the line, that ends
with `# fmt: skip`, but `comment.consumed` is the length of the
`" # fmt: skip"` string. If prefix length is greater than 14,
`first.prefix` will grow every time we apply formatting.
Fixes#2254
PR #2286 did not fix the edge-cases (e.g. when the string is just long
enough to cause a line to be 89 characters long). This PR corrects that
mistake.
Resolves#2168 by disabling the insertion of a " " when the docstring is entirely empty.
Note that this PR is focussed only on the case of empty docstrings. In particular this does not make any changes to the behaviour that a " " is inserted if a non-empty docstring begins with the quoting character. That is, black still prefers:
""" "something" """
to:
""""something" """
and that:
""""Something""""
is not a legal docstring.
Commit history before merge:
Black now respects .gitignore files in all levels, not only root/.gitignore file
(apply .gitignore rules like git does).
* Fix: typo
* Fix: respect .gitignore files in all levels.
* Add: CHANGELOG note.
* Fix: TypeError: unsupported operand type(s) for +: 'NoneType' and 'PathSpec'
* Update docs.
* Fix: no parent .gitignore
* Add a comment since the if expression is a bit hard to understand
* Update tests - conver no parent .gitignore case.
* Use main's Pipfile.lock instead
The original changes in Pipfile.lock are whitespace only. The changes
turned the JSON's file indentation from 4 to 2. Effectively this
happened: `json.dumps(json.loads(old_pipfile_lock), indent=2) + "\n"`.
Just using main's Pipfile.lock instead of undoing the changes because
1) I don't know how to do that easily and quickly, and 2) there's a
merge conflict.
Co-authored-by: Richard Si <63936253+ichard26@users.noreply.github.com>
* Merge remote-tracking branch 'upstream/main' into i1730 …
conflicts for days ay?
Closes#2164.
Changes behavior of how .gitignore is handled. With this change, the rules in .gitignore are only used as a fallback if no exclusion rule is explicitly passed on the command line or in pyproject.toml. Previously they were used regardless if explicit exclusion rules were specified, preventing any overriding of .gitignore rules.
Those that depend only on .gitignore for their exclusion rules will not be affected. Those that use both .gitignore and exclude will find that exclude will act more like actually specifying exclude and not just another extra-excludes. If the previous behavior was desired, they should move their rules from exclude to extra-excludes.
Previously the RELAXED_DECORATOR detection would be falsely True on that
example. The problem was that an argument-less parentheses pair didn't
pass the `is_simple_decorator_trailer` check even it should. OTOH a
parentheses pair containing an argument or more passed as expected.
We're only fixing them so fuzzers don't yell at us when we break "valid"
code. I mean "valid" because some of the examples aren't even accepted by
Python.
Black would previously strip the parenthesis away from statements like this these ones:
assert (spam := 12 + 1)
return (cheese := 1 - 12)
Which happens to be invalid code. Now before making the parenthesis invisible, Black
checks if the assignment expression's parent is an assert stamtment, aborting if True.
Raise, yield, and await are already handled fine.
I added a bunch of test cases from the PEP defining asssignment expressions (PEP 572).
Fixes#1738. Fixes#1812.
Previously, Black removed leading and trailing spaces in multiline docstrings but failed to remove them from one-line docstrings.
Fixes: #1662
Work-around for https://bugs.python.org/issue2142
The test has to slightly mess with its input data, because the utility
functions default to ensuring the test data has a final newline, which
defeats the point of the test.
Signed-off-by: Paul "TBBle" Hampson <Paul.Hampson@Pobox.com>
The `fix_docstring` function expanded all tabs, which caused a
difference in the AST representation when those tabs were inline and not
leading. This changes the function to only expand leading tabs so inline
tabs are preserved.
Fixes#1601.
Black used to erroneously remove all empty lines between non-function
code and decorators when formatting typing stubs. Now a single empty
line is enforced.
I chose for putting empty lines around decorated classes that have empty
bodies since removing empty lines around such classes would cause a
formatting issue that seems to be impossible to fix.
For example:
```
class A: ...
@some_decorator
class B: ...
class C: ...
class D: ...
@some_other_decorator
def foo(): -> None: ...
```
It is easy to enforce no empty lines between class A, B, and C.
Just return 0, 0 for a line that is a decorator and precedes an stub
class. Fortunately before this commit, empty lines after that class
would be removed already.
Now let's look at the empty line between class D and function foo. In
this case, there should be an empty line there since it's class code next
to function code. The problem is that when deciding to add X empty lines
before a decorator, you can't tell whether it's before a class or a
function. If the decorator is before a function, then an empty line
is needed, while no empty lines are needed when the decorator is
before a class.
So even though I personally prefer no empty lines around decorated
classes, I had to go the other way surrounding decorated classes with
empty lines.
Co-authored-by: Jelle Zijlstra <jelle.zijlstra@gmail.com>
* Fix incorrect space before colon in if/while stmts
Previously Black would format this code
```
if (foo := True):
print(foo)
```
as
```
if (foo := True) :
print(foo)
```
adding an incorrect space after the RPAR. Buggy code in the
normalize_invisible_parens function caused the colon to be wrapped in
invisible parentheses. The LPAR of that pair was then prefixed with a
single space at the request of the whitespace function.
This commit fixes the accidental skipping of a pre-condition check
which must return True before parenthesis normalization of a specific
child Leaf or Node can happen. The pre-condition check being skipped
was why the colon was wrapped in invisible parentheses.
* Add an entry in CHANGES.md
The quotes of multiline docstrings are now only normalized when string
normalization is off, instead of the string normalization setting being
ignored and the quotes being *always* normalized.
I had to make a new test case and data file since the current pair for
docstrings only worked when there is no formatting difference between the
formatting results with string normalization on and off. I needed to add
tests for when there *are* differences between the two. So I split
test_docstring's test code when string normalization is disabled into a
new test case along with a new data file.
This addresses a few crashers, namely:
* producing non-equivalent code due to mangling escaped newlines,
* invalid hugging quote characters in the docstring body to the docstring outer
triple quotes (causing a quadruple quote which is a syntax error),
* lack of handling for docstrings that start on the same line as the `def`, and
* invalid stripping of outer triple quotes when the docstring contained
a string prefix.
As a bonus, tests now also run when string normalization is disabled.
This required some hackery. Long story short, we need to reuse the ability to
omit rightmost bracket pairs (which glues them together and splits on something
else instead), for use with pre-existing trailing commas.
This form of user-controlled formatting is brittle so we have to be careful not
to cause a scenario where Black first formats code without trailing commas in
one way, and then looks at the same file with pre-existing trailing commas
(that it itself put on the previous run) and decides to format the code again.
One particular ugly edge case here is handling of optional parentheses. In
particular, the long-standing `line_length=1` hack got in the way of
pre-existing trailing commas and had to be removed. Instead, a more
intelligent but costly solution was put in place: a "second opinion" if the
formatting that omits optional parentheses ended up causing lines to be too
long. Again, for efficiency purposes, Black reuses Leaf objects from blib2to3
and modifies them in place, which was invalid for having two separate
formattings. Line cloning was used to mitigate this.
Fixes#1619
- when a trailing comma is specified in any bracket pair, that signals to Black
that this bracket pair needs to be always exploded, e.g. presented as "one
item per line";
- this causes some changes to previously formatted code that erroneously left
trailing commas embedded into single-line expressions;
- internally, Black needs to be able to identify trailing commas that it put
itself compared to pre-existing trailing commas. We do this by using/abusing
lib2to3's `was_checked` attribute. It's True for internally generated
trailing commas and False for pre-existing ones (in fact, for all
pre-existing leaves and nodes).
Fixes#1288
* put experimental string stuff behind a flag
* update tests
* don't need an output section if it's the same as the input
* Primer: Expect no formatting changes in attrs, hypothesis and poetry with --experimental-string-processing off
Co-authored-by: Hugo van Kemenade <hugovk@users.noreply.github.com>
As discussed in #1441, Python 3.9's new parser will not parse
`(*starred)` even using `compile()` with the `PyCF_ONLY_AST`
flag (as `ast.parse()` does), it raises a `SyntaxError`. This
breaks the four tests that use this file with Python 3.9.
Upstream does not consider this to be a bug - see
https://bugs.python.org/issue40848#msg370643 - so we must
adjust the expression. As suggested by @JelleZijlstra, this just
adds a comma, which makes the new parser happy with it (the old
parser is fine with this form also).
Signed-off-by: Adam Williamson <awilliam@redhat.com>
* Re-indent the contents of docstrings when indentation changes
Keeping the contents of docstrings completely unchanged when
re-indenting (from 2-space intents to 4, for example) can cause
incorrect docstring indentation:
```
class MyClass:
"""Multiline
class docstring
"""
def method(self):
"""Multiline
method docstring
"""
pass
```
...becomes:
```
class MyClass:
"""Multiline
class docstring
"""
def method(self):
"""Multiline
method docstring
"""
pass
```
This uses the PEP 257 algorithm for determining docstring indentation,
and adjusts the contents of docstrings to match their new indentation
after `black` is applied.
A small normalization is necessary to `assert_equivalent` because the
trees are technically no longer precisely equivalent -- some constant
strings have changed. When comparing two ASTs, whitespace after
newlines within constant strings is thus folded into a single space.
Co-authored-by: Luka Zakrajšek <luka@bancek.net>
* Extract the inner `_v` method to decrease complexity
This reduces the cyclomatic complexity to a level that makes flake8 happy.
* Blacken blib2to3's docstring which had an over-indent
Co-authored-by: Luka Zakrajšek <luka@bancek.net>
Co-authored-by: Zsolt Dollenstein <zsol.zsol@gmail.com>
This pull request's main intention is to wraps long strings (as requested by #182); however, it also provides better string handling in general and, in doing so, closes the following issues:
Closes#26Closes#182Closes#933Closes#1183Closes#1243
`split("\n")` includes a final empty element `""` if the final line
ends with `\n` (as it should for POSIX-compliant text files), which
then became an extra `"\n"`.
`splitlines()` solves that, but there's a caveat, as it will split
on other types of line breaks too (like `\r`), which may not be
desired.
Fixes#526.
`type: ignore` shouldn't block collapsing a line, since it will still
apply fine to the merged line. This prevents an issue where a reformat
causes it to shift lines and then be merged on a subsequent pass.
There is a downside to this, which is that it can cause a `type:
ignore` to apply to more code than was originally intended. There
might be a way to apply this in a more limited situation, but I'm not
sure what it is.
Fixes#1061.
The old behavior would detect the existence of a `# fmt: on` in a leaf
node's comment prefix and immediately mark the node as formatting-on,
even if a subsequent `# fmt: off` in the same comment prefix would turn
it back off. This change modifies that logic to track the state through
the entire prefix and take the final state.
Note that this does not fully solve on/off behavior, since any _comment_
lines between the off/on are still formatted. We may need to add
virtual leaf nodes to truly solve that. I will leave that for a separate
commit/PR.
Fixes#1005
Fixes#1042 (and probably #1044 which looks like the same thing).
The issue with the "obviously unnecessary" parentheses that #850 removed is that sometimes they're necessary to help Black fit something in one line. I didn't see an obvious solution that still removes the parens #850 was intended to remove, so let's back out this change for now in the interest of unblocking a release.
This PR also adds a test adapted from the failing example in #1042, so that if we try to reapply the #850 change we don't break the same case again.
It turns out we also need to handle invisible *left* parens added at
the *start* of a line. Refactor `contains_unsplittable_type_ignore` to
handle this more cleanly.
* add test for special unicode symbol which usual re can not process correctly
add regex lib which supports unicode 12.1.0 standard
replace re usage in project in favor to regex
* #455 fix dependency
In #1040 I had convinced myself that the type ignore logic didn't
need anything like the ignored_ids from the type comment logic, but I
was wrong, and we do.
We hit these cases in practice a bunch.
The code introduced in #1027 to detect whether a type comment appeared
after a regular comment in a Line would spuriously misfire when a leaf
was in the comments dict but had an empty list of comments. This can
occur as an artifact of how comments on trailing commas are handled,
it seems.
(This was discovered trying to test black out on mypy.)
Type comments only apply if they are the first comment on the line,
which means that allowing them to be pushed behind a regular comment
when joining lines is a semantic change (and, indeed, one that black
catches and fails on).
