]> git.madduck.net Git - etc/vim.git/blobdiff - black.py

madduck's git repository

Every one of the projects in this repository is available at the canonical URL git://git.madduck.net/madduck/pub/<projectpath> — see each project's metadata for the exact URL.

All patches and comments are welcome. Please squash your changes to logical commits before using git-format-patch and git-send-email to patches@git.madduck.net. If you'd read over the Git project's submission guidelines and adhered to them, I'd be especially grateful.

SSH access, as well as push access can be individually arranged.

If you use my repositories frequently, consider adding the following snippet to ~/.gitconfig and using the third clone URL listed for each project:

[url "git://git.madduck.net/madduck/"]
  insteadOf = madduck:

Small nitpicks (#1340)
[etc/vim.git] / black.py
index 0dce397e768fee2973537f157b4903c3d858390d..d9348a37b42d4c45a4bf4fd126f23c8c6626d848 100644 (file)
--- a/black.py
+++ b/black.py
@@ -1,23 +1,29 @@
+import ast
 import asyncio
-import pickle
-from asyncio.base_events import BaseEventLoop
 from concurrent.futures import Executor, ProcessPoolExecutor
-from enum import Enum, Flag
-from functools import partial, wraps
-import keyword
+from contextlib import contextmanager
+from datetime import datetime
+from enum import Enum
+from functools import lru_cache, partial, wraps
+import io
+import itertools
 import logging
-from multiprocessing import Manager
+from multiprocessing import Manager, freeze_support
 import os
 from pathlib import Path
-import re
-import tokenize
+import pickle
+import regex as re
 import signal
 import sys
+import tempfile
+import tokenize
+import traceback
 from typing import (
     Any,
     Callable,
     Collection,
     Dict,
+    Generator,
     Generic,
     Iterable,
     Iterator,
@@ -27,28 +33,31 @@ from typing import (
     Sequence,
     Set,
     Tuple,
-    Type,
     TypeVar,
     Union,
     cast,
 )
+from typing_extensions import Final
+from mypy_extensions import mypyc_attr
 
 from appdirs import user_cache_dir
-from attr import dataclass, Factory
+from dataclasses import dataclass, field, replace
 import click
+import toml
+from typed_ast import ast3, ast27
+from pathspec import PathSpec
 
 # lib2to3 fork
 from blib2to3.pytree import Node, Leaf, type_repr
 from blib2to3 import pygram, pytree
 from blib2to3.pgen2 import driver, token
+from blib2to3.pgen2.grammar import Grammar
 from blib2to3.pgen2.parse import ParseError
 
+from _black_version import version as __version__
 
-__version__ = "18.5b1"
 DEFAULT_LINE_LENGTH = 88
-DEFAULT_EXCLUDES = (
-    r"/(\.git|\.hg|\.mypy_cache|\.tox|\.venv|_build|buck-out|build|dist)/"
-)
+DEFAULT_EXCLUDES = r"/(\.eggs|\.git|\.hg|\.mypy_cache|\.nox|\.tox|\.venv|\.svn|_build|buck-out|build|dist)/"  # noqa: B950
 DEFAULT_INCLUDES = r"\.pyi?$"
 CACHE_DIR = Path(user_cache_dir("black", version=__version__))
 
@@ -56,13 +65,14 @@ CACHE_DIR = Path(user_cache_dir("black", version=__version__))
 # types
 FileContent = str
 Encoding = str
+NewLine = str
 Depth = int
 NodeType = int
 LeafID = int
 Priority = int
 Index = int
 LN = Union[Leaf, Node]
-SplitFunc = Callable[["Line", bool], Iterator["Line"]]
+SplitFunc = Callable[["Line", Collection["Feature"]], Iterator["Line"]]
 Timestamp = float
 FileSize = int
 CacheInfo = Tuple[Timestamp, FileSize]
@@ -75,80 +85,225 @@ syms = pygram.python_symbols
 
 
 class NothingChanged(UserWarning):
-    """Raised by :func:`format_file` when reformatted code is the same as source."""
+    """Raised when reformatted code is the same as source."""
 
 
 class CannotSplit(Exception):
-    """A readable split that fits the allotted line length is impossible.
+    """A readable split that fits the allotted line length is impossible."""
 
-    Raised by :func:`left_hand_split`, :func:`right_hand_split`, and
-    :func:`delimiter_split`.
-    """
 
+class InvalidInput(ValueError):
+    """Raised when input source code fails all parse attempts."""
 
-class FormatError(Exception):
-    """Base exception for `# fmt: on` and `# fmt: off` handling.
 
-    It holds the number of bytes of the prefix consumed before the format
-    control comment appeared.
-    """
+class WriteBack(Enum):
+    NO = 0
+    YES = 1
+    DIFF = 2
+    CHECK = 3
 
-    def __init__(self, consumed: int) -> None:
-        super().__init__(consumed)
-        self.consumed = consumed
+    @classmethod
+    def from_configuration(cls, *, check: bool, diff: bool) -> "WriteBack":
+        if check and not diff:
+            return cls.CHECK
 
-    def trim_prefix(self, leaf: Leaf) -> None:
-        leaf.prefix = leaf.prefix[self.consumed :]
+        return cls.DIFF if diff else cls.YES
 
-    def leaf_from_consumed(self, leaf: Leaf) -> Leaf:
-        """Returns a new Leaf from the consumed part of the prefix."""
-        unformatted_prefix = leaf.prefix[: self.consumed]
-        return Leaf(token.NEWLINE, unformatted_prefix)
 
+class Changed(Enum):
+    NO = 0
+    CACHED = 1
+    YES = 2
 
-class FormatOn(FormatError):
-    """Found a comment like `# fmt: on` in the file."""
 
+class TargetVersion(Enum):
+    PY27 = 2
+    PY33 = 3
+    PY34 = 4
+    PY35 = 5
+    PY36 = 6
+    PY37 = 7
+    PY38 = 8
+
+    def is_python2(self) -> bool:
+        return self is TargetVersion.PY27
+
+
+PY36_VERSIONS = {TargetVersion.PY36, TargetVersion.PY37, TargetVersion.PY38}
+
+
+class Feature(Enum):
+    # All string literals are unicode
+    UNICODE_LITERALS = 1
+    F_STRINGS = 2
+    NUMERIC_UNDERSCORES = 3
+    TRAILING_COMMA_IN_CALL = 4
+    TRAILING_COMMA_IN_DEF = 5
+    # The following two feature-flags are mutually exclusive, and exactly one should be
+    # set for every version of python.
+    ASYNC_IDENTIFIERS = 6
+    ASYNC_KEYWORDS = 7
+    ASSIGNMENT_EXPRESSIONS = 8
+    POS_ONLY_ARGUMENTS = 9
+
+
+VERSION_TO_FEATURES: Dict[TargetVersion, Set[Feature]] = {
+    TargetVersion.PY27: {Feature.ASYNC_IDENTIFIERS},
+    TargetVersion.PY33: {Feature.UNICODE_LITERALS, Feature.ASYNC_IDENTIFIERS},
+    TargetVersion.PY34: {Feature.UNICODE_LITERALS, Feature.ASYNC_IDENTIFIERS},
+    TargetVersion.PY35: {
+        Feature.UNICODE_LITERALS,
+        Feature.TRAILING_COMMA_IN_CALL,
+        Feature.ASYNC_IDENTIFIERS,
+    },
+    TargetVersion.PY36: {
+        Feature.UNICODE_LITERALS,
+        Feature.F_STRINGS,
+        Feature.NUMERIC_UNDERSCORES,
+        Feature.TRAILING_COMMA_IN_CALL,
+        Feature.TRAILING_COMMA_IN_DEF,
+        Feature.ASYNC_IDENTIFIERS,
+    },
+    TargetVersion.PY37: {
+        Feature.UNICODE_LITERALS,
+        Feature.F_STRINGS,
+        Feature.NUMERIC_UNDERSCORES,
+        Feature.TRAILING_COMMA_IN_CALL,
+        Feature.TRAILING_COMMA_IN_DEF,
+        Feature.ASYNC_KEYWORDS,
+    },
+    TargetVersion.PY38: {
+        Feature.UNICODE_LITERALS,
+        Feature.F_STRINGS,
+        Feature.NUMERIC_UNDERSCORES,
+        Feature.TRAILING_COMMA_IN_CALL,
+        Feature.TRAILING_COMMA_IN_DEF,
+        Feature.ASYNC_KEYWORDS,
+        Feature.ASSIGNMENT_EXPRESSIONS,
+        Feature.POS_ONLY_ARGUMENTS,
+    },
+}
 
-class FormatOff(FormatError):
-    """Found a comment like `# fmt: off` in the file."""
 
+@dataclass
+class Mode:
+    target_versions: Set[TargetVersion] = field(default_factory=set)
+    line_length: int = DEFAULT_LINE_LENGTH
+    string_normalization: bool = True
+    is_pyi: bool = False
 
-class WriteBack(Enum):
-    NO = 0
-    YES = 1
-    DIFF = 2
+    def get_cache_key(self) -> str:
+        if self.target_versions:
+            version_str = ",".join(
+                str(version.value)
+                for version in sorted(self.target_versions, key=lambda v: v.value)
+            )
+        else:
+            version_str = "-"
+        parts = [
+            version_str,
+            str(self.line_length),
+            str(int(self.string_normalization)),
+            str(int(self.is_pyi)),
+        ]
+        return ".".join(parts)
 
 
-class Changed(Enum):
-    NO = 0
-    CACHED = 1
-    YES = 2
+# Legacy name, left for integrations.
+FileMode = Mode
+
+
+def supports_feature(target_versions: Set[TargetVersion], feature: Feature) -> bool:
+    return all(feature in VERSION_TO_FEATURES[version] for version in target_versions)
+
+
+def find_pyproject_toml(path_search_start: str) -> Optional[str]:
+    """Find the absolute filepath to a pyproject.toml if it exists"""
+    path_project_root = find_project_root(path_search_start)
+    path_pyproject_toml = path_project_root / "pyproject.toml"
+    return str(path_pyproject_toml) if path_pyproject_toml.is_file() else None
+
+
+def parse_pyproject_toml(path_config: str) -> Dict[str, Any]:
+    """Parse a pyproject toml file, pulling out relevant parts for Black
+
+    If parsing fails, will raise a toml.TomlDecodeError
+    """
+    pyproject_toml = toml.load(path_config)
+    config = pyproject_toml.get("tool", {}).get("black", {})
+    return {k.replace("--", "").replace("-", "_"): v for k, v in config.items()}
+
+
+def read_pyproject_toml(
+    ctx: click.Context, param: click.Parameter, value: Union[str, int, bool, None]
+) -> Optional[str]:
+    """Inject Black configuration from "pyproject.toml" into defaults in `ctx`.
 
+    Returns the path to a successfully found and read configuration file, None
+    otherwise.
+    """
+    assert not isinstance(value, (int, bool)), "Invalid parameter type passed"
+    if not value:
+        value = find_pyproject_toml(ctx.params.get("src", ()))
+        if value is None:
+            return None
+
+    try:
+        config = parse_pyproject_toml(value)
+    except (toml.TomlDecodeError, OSError) as e:
+        raise click.FileError(
+            filename=value, hint=f"Error reading configuration file: {e}"
+        )
+
+    if not config:
+        return None
 
-class FileMode(Flag):
-    AUTO_DETECT = 0
-    PYTHON36 = 1
-    PYI = 2
-    NO_STRING_NORMALIZATION = 4
+    if ctx.default_map is None:
+        ctx.default_map = {}
+    ctx.default_map.update(config)  # type: ignore  # bad types in .pyi
+    return value
 
 
-@click.command()
+def target_version_option_callback(
+    c: click.Context, p: Union[click.Option, click.Parameter], v: Tuple[str, ...]
+) -> List[TargetVersion]:
+    """Compute the target versions from a --target-version flag.
+
+    This is its own function because mypy couldn't infer the type correctly
+    when it was a lambda, causing mypyc trouble.
+    """
+    return [TargetVersion[val.upper()] for val in v]
+
+
+@click.command(context_settings=dict(help_option_names=["-h", "--help"]))
+@click.option("-c", "--code", type=str, help="Format the code passed in as a string.")
 @click.option(
     "-l",
     "--line-length",
     type=int,
     default=DEFAULT_LINE_LENGTH,
-    help="How many character per line to allow.",
+    help="How many characters per line to allow.",
     show_default=True,
 )
+@click.option(
+    "-t",
+    "--target-version",
+    type=click.Choice([v.name.lower() for v in TargetVersion]),
+    callback=target_version_option_callback,
+    multiple=True,
+    help=(
+        "Python versions that should be supported by Black's output. [default: "
+        "per-file auto-detection]"
+    ),
+)
 @click.option(
     "--py36",
     is_flag=True,
     help=(
         "Allow using Python 3.6-only syntax on all input files.  This will put "
         "trailing commas in function signatures and calls also after *args and "
-        "**kwargs.  [default: per-file auto-detection]"
+        "**kwargs. Deprecated; use --target-version instead. "
+        "[default: per-file auto-detection]"
     ),
 )
 @click.option(
@@ -214,10 +369,19 @@ class FileMode(Flag):
     "--quiet",
     is_flag=True,
     help=(
-        "Don't emit non-error messages to stderr. Errors are still emitted, "
+        "Don't emit non-error messages to stderr. Errors are still emitted; "
         "silence those with 2>/dev/null."
     ),
 )
+@click.option(
+    "-v",
+    "--verbose",
+    is_flag=True,
+    help=(
+        "Also emit messages to stderr about files that were not changed or were "
+        "ignored due to --exclude=."
+    ),
+)
 @click.version_option(version=__version__)
 @click.argument(
     "src",
@@ -225,11 +389,23 @@ class FileMode(Flag):
     type=click.Path(
         exists=True, file_okay=True, dir_okay=True, readable=True, allow_dash=True
     ),
+    is_eager=True,
+)
+@click.option(
+    "--config",
+    type=click.Path(
+        exists=False, file_okay=True, dir_okay=False, readable=True, allow_dash=False
+    ),
+    is_eager=True,
+    callback=read_pyproject_toml,
+    help="Read configuration from PATH.",
 )
 @click.pass_context
 def main(
     ctx: click.Context,
+    code: Optional[str],
     line_length: int,
+    target_version: List[TargetVersion],
     check: bool,
     diff: bool,
     fast: bool,
@@ -237,342 +413,457 @@ def main(
     py36: bool,
     skip_string_normalization: bool,
     quiet: bool,
+    verbose: bool,
     include: str,
     exclude: str,
-    src: List[str],
+    src: Tuple[str, ...],
+    config: Optional[str],
 ) -> None:
     """The uncompromising code formatter."""
-    sources: List[Path] = []
+    write_back = WriteBack.from_configuration(check=check, diff=diff)
+    if target_version:
+        if py36:
+            err("Cannot use both --target-version and --py36")
+            ctx.exit(2)
+        else:
+            versions = set(target_version)
+    elif py36:
+        err(
+            "--py36 is deprecated and will be removed in a future version. "
+            "Use --target-version py36 instead."
+        )
+        versions = PY36_VERSIONS
+    else:
+        # We'll autodetect later.
+        versions = set()
+    mode = Mode(
+        target_versions=versions,
+        line_length=line_length,
+        is_pyi=pyi,
+        string_normalization=not skip_string_normalization,
+    )
+    if config and verbose:
+        out(f"Using configuration from {config}.", bold=False, fg="blue")
+    if code is not None:
+        print(format_str(code, mode=mode))
+        ctx.exit(0)
     try:
-        include_regex = re.compile(include)
+        include_regex = re_compile_maybe_verbose(include)
     except re.error:
         err(f"Invalid regular expression for include given: {include!r}")
         ctx.exit(2)
     try:
-        exclude_regex = re.compile(exclude)
+        exclude_regex = re_compile_maybe_verbose(exclude)
     except re.error:
         err(f"Invalid regular expression for exclude given: {exclude!r}")
         ctx.exit(2)
+    report = Report(check=check, diff=diff, quiet=quiet, verbose=verbose)
     root = find_project_root(src)
+    sources: Set[Path] = set()
+    path_empty(src, quiet, verbose, ctx)
     for s in src:
         p = Path(s)
         if p.is_dir():
-            sources.extend(
-                gen_python_files_in_dir(p, root, include_regex, exclude_regex)
+            sources.update(
+                gen_python_files_in_dir(
+                    p, root, include_regex, exclude_regex, report, get_gitignore(root)
+                )
             )
         elif p.is_file() or s == "-":
             # if a file was explicitly given, we don't care about its extension
-            sources.append(p)
+            sources.add(p)
         else:
             err(f"invalid path: {s}")
-
-    if check and not diff:
-        write_back = WriteBack.NO
-    elif diff:
-        write_back = WriteBack.DIFF
-    else:
-        write_back = WriteBack.YES
-    mode = FileMode.AUTO_DETECT
-    if py36:
-        mode |= FileMode.PYTHON36
-    if pyi:
-        mode |= FileMode.PYI
-    if skip_string_normalization:
-        mode |= FileMode.NO_STRING_NORMALIZATION
-    report = Report(check=check, quiet=quiet)
     if len(sources) == 0:
-        out("No paths given. Nothing to do 😴")
+        if verbose or not quiet:
+            out("No Python files are present to be formatted. Nothing to do 😴")
         ctx.exit(0)
-        return
 
-    elif len(sources) == 1:
+    if len(sources) == 1:
         reformat_one(
-            src=sources[0],
-            line_length=line_length,
+            src=sources.pop(),
             fast=fast,
             write_back=write_back,
             mode=mode,
             report=report,
         )
     else:
-        loop = asyncio.get_event_loop()
-        executor = ProcessPoolExecutor(max_workers=os.cpu_count())
-        try:
-            loop.run_until_complete(
-                schedule_formatting(
-                    sources=sources,
-                    line_length=line_length,
-                    fast=fast,
-                    write_back=write_back,
-                    mode=mode,
-                    report=report,
-                    loop=loop,
-                    executor=executor,
-                )
-            )
-        finally:
-            shutdown(loop)
-        if not quiet:
-            out("All done! ✨ 🍰 ✨")
-            click.echo(str(report))
+        reformat_many(
+            sources=sources, fast=fast, write_back=write_back, mode=mode, report=report
+        )
+
+    if verbose or not quiet:
+        out("Oh no! 💥 💔 💥" if report.return_code else "All done! ✨ 🍰 ✨")
+        click.secho(str(report), err=True)
     ctx.exit(report.return_code)
 
 
+def path_empty(
+    src: Tuple[str, ...], quiet: bool, verbose: bool, ctx: click.Context
+) -> None:
+    """
+    Exit if there is no `src` provided for formatting
+    """
+    if not src:
+        if verbose or not quiet:
+            out("No Path provided. Nothing to do 😴")
+            ctx.exit(0)
+
+
 def reformat_one(
-    src: Path,
-    line_length: int,
-    fast: bool,
-    write_back: WriteBack,
-    mode: FileMode,
-    report: "Report",
+    src: Path, fast: bool, write_back: WriteBack, mode: Mode, report: "Report"
 ) -> None:
     """Reformat a single file under `src` without spawning child processes.
 
-    If `quiet` is True, non-error messages are not output. `line_length`,
-    `write_back`, `fast` and `pyi` options are passed to
+    `fast`, `write_back`, and `mode` options are passed to
     :func:`format_file_in_place` or :func:`format_stdin_to_stdout`.
     """
     try:
         changed = Changed.NO
         if not src.is_file() and str(src) == "-":
-            if format_stdin_to_stdout(
-                line_length=line_length, fast=fast, write_back=write_back, mode=mode
-            ):
+            if format_stdin_to_stdout(fast=fast, write_back=write_back, mode=mode):
                 changed = Changed.YES
         else:
             cache: Cache = {}
             if write_back != WriteBack.DIFF:
-                cache = read_cache(line_length, mode)
+                cache = read_cache(mode)
                 res_src = src.resolve()
                 if res_src in cache and cache[res_src] == get_cache_info(res_src):
                     changed = Changed.CACHED
             if changed is not Changed.CACHED and format_file_in_place(
-                src,
-                line_length=line_length,
-                fast=fast,
-                write_back=write_back,
-                mode=mode,
+                src, fast=fast, write_back=write_back, mode=mode
             ):
                 changed = Changed.YES
-            if write_back == WriteBack.YES and changed is not Changed.NO:
-                write_cache(cache, [src], line_length, mode)
+            if (write_back is WriteBack.YES and changed is not Changed.CACHED) or (
+                write_back is WriteBack.CHECK and changed is Changed.NO
+            ):
+                write_cache(cache, [src], mode)
         report.done(src, changed)
     except Exception as exc:
         report.failed(src, str(exc))
 
 
+def reformat_many(
+    sources: Set[Path], fast: bool, write_back: WriteBack, mode: Mode, report: "Report"
+) -> None:
+    """Reformat multiple files using a ProcessPoolExecutor."""
+    loop = asyncio.get_event_loop()
+    worker_count = os.cpu_count()
+    if sys.platform == "win32":
+        # Work around https://bugs.python.org/issue26903
+        worker_count = min(worker_count, 61)
+    executor = ProcessPoolExecutor(max_workers=worker_count)
+    try:
+        loop.run_until_complete(
+            schedule_formatting(
+                sources=sources,
+                fast=fast,
+                write_back=write_back,
+                mode=mode,
+                report=report,
+                loop=loop,
+                executor=executor,
+            )
+        )
+    finally:
+        shutdown(loop)
+        executor.shutdown()
+
+
 async def schedule_formatting(
-    sources: List[Path],
-    line_length: int,
+    sources: Set[Path],
     fast: bool,
     write_back: WriteBack,
-    mode: FileMode,
+    mode: Mode,
     report: "Report",
-    loop: BaseEventLoop,
+    loop: asyncio.AbstractEventLoop,
     executor: Executor,
 ) -> None:
     """Run formatting of `sources` in parallel using the provided `executor`.
 
     (Use ProcessPoolExecutors for actual parallelism.)
 
-    `line_length`, `write_back`, `fast`, and `pyi` options are passed to
+    `write_back`, `fast`, and `mode` options are passed to
     :func:`format_file_in_place`.
     """
     cache: Cache = {}
     if write_back != WriteBack.DIFF:
-        cache = read_cache(line_length, mode)
+        cache = read_cache(mode)
         sources, cached = filter_cached(cache, sources)
-        for src in cached:
+        for src in sorted(cached):
             report.done(src, Changed.CACHED)
+    if not sources:
+        return
+
     cancelled = []
-    formatted = []
-    if sources:
-        lock = None
-        if write_back == WriteBack.DIFF:
-            # For diff output, we need locks to ensure we don't interleave output
-            # from different processes.
-            manager = Manager()
-            lock = manager.Lock()
-        tasks = {
+    sources_to_cache = []
+    lock = None
+    if write_back == WriteBack.DIFF:
+        # For diff output, we need locks to ensure we don't interleave output
+        # from different processes.
+        manager = Manager()
+        lock = manager.Lock()
+    tasks = {
+        asyncio.ensure_future(
             loop.run_in_executor(
-                executor,
-                format_file_in_place,
-                src,
-                line_length,
-                fast,
-                write_back,
-                mode,
-                lock,
-            ): src
-            for src in sorted(sources)
-        }
-        pending: Iterable[asyncio.Task] = tasks.keys()
-        try:
-            loop.add_signal_handler(signal.SIGINT, cancel, pending)
-            loop.add_signal_handler(signal.SIGTERM, cancel, pending)
-        except NotImplementedError:
-            # There are no good alternatives for these on Windows
-            pass
-        while pending:
-            done, _ = await asyncio.wait(pending, return_when=asyncio.FIRST_COMPLETED)
-            for task in done:
-                src = tasks.pop(task)
-                if task.cancelled():
-                    cancelled.append(task)
-                elif task.exception():
-                    report.failed(src, str(task.exception()))
-                else:
-                    formatted.append(src)
-                    report.done(src, Changed.YES if task.result() else Changed.NO)
+                executor, format_file_in_place, src, fast, mode, write_back, lock
+            )
+        ): src
+        for src in sorted(sources)
+    }
+    pending: Iterable["asyncio.Future[bool]"] = tasks.keys()
+    try:
+        loop.add_signal_handler(signal.SIGINT, cancel, pending)
+        loop.add_signal_handler(signal.SIGTERM, cancel, pending)
+    except NotImplementedError:
+        # There are no good alternatives for these on Windows.
+        pass
+    while pending:
+        done, _ = await asyncio.wait(pending, return_when=asyncio.FIRST_COMPLETED)
+        for task in done:
+            src = tasks.pop(task)
+            if task.cancelled():
+                cancelled.append(task)
+            elif task.exception():
+                report.failed(src, str(task.exception()))
+            else:
+                changed = Changed.YES if task.result() else Changed.NO
+                # If the file was written back or was successfully checked as
+                # well-formatted, store this information in the cache.
+                if write_back is WriteBack.YES or (
+                    write_back is WriteBack.CHECK and changed is Changed.NO
+                ):
+                    sources_to_cache.append(src)
+                report.done(src, changed)
     if cancelled:
         await asyncio.gather(*cancelled, loop=loop, return_exceptions=True)
-    if write_back == WriteBack.YES and formatted:
-        write_cache(cache, formatted, line_length, mode)
+    if sources_to_cache:
+        write_cache(cache, sources_to_cache, mode)
 
 
 def format_file_in_place(
     src: Path,
-    line_length: int,
     fast: bool,
+    mode: Mode,
     write_back: WriteBack = WriteBack.NO,
-    mode: FileMode = FileMode.AUTO_DETECT,
     lock: Any = None,  # multiprocessing.Manager().Lock() is some crazy proxy
 ) -> bool:
     """Format file under `src` path. Return True if changed.
 
-    If `write_back` is True, write reformatted code back to stdout.
-    `line_length` and `fast` options are passed to :func:`format_file_contents`.
+    If `write_back` is DIFF, write a diff to stdout. If it is YES, write reformatted
+    code to the file.
+    `mode` and `fast` options are passed to :func:`format_file_contents`.
     """
     if src.suffix == ".pyi":
-        mode |= FileMode.PYI
-    with tokenize.open(src) as src_buffer:
-        src_contents = src_buffer.read()
+        mode = replace(mode, is_pyi=True)
+
+    then = datetime.utcfromtimestamp(src.stat().st_mtime)
+    with open(src, "rb") as buf:
+        src_contents, encoding, newline = decode_bytes(buf.read())
     try:
-        dst_contents = format_file_contents(
-            src_contents, line_length=line_length, fast=fast, mode=mode
-        )
+        dst_contents = format_file_contents(src_contents, fast=fast, mode=mode)
     except NothingChanged:
         return False
 
-    if write_back == write_back.YES:
-        with open(src, "w", encoding=src_buffer.encoding) as f:
+    if write_back == WriteBack.YES:
+        with open(src, "w", encoding=encoding, newline=newline) as f:
             f.write(dst_contents)
-    elif write_back == write_back.DIFF:
-        src_name = f"{src}  (original)"
-        dst_name = f"{src}  (formatted)"
+    elif write_back == WriteBack.DIFF:
+        now = datetime.utcnow()
+        src_name = f"{src}\t{then} +0000"
+        dst_name = f"{src}\t{now} +0000"
         diff_contents = diff(src_contents, dst_contents, src_name, dst_name)
-        if lock:
-            lock.acquire()
-        try:
-            sys.stdout.write(diff_contents)
-        finally:
-            if lock:
-                lock.release()
+
+        with lock or nullcontext():
+            f = io.TextIOWrapper(
+                sys.stdout.buffer,
+                encoding=encoding,
+                newline=newline,
+                write_through=True,
+            )
+            f.write(diff_contents)
+            f.detach()
+
     return True
 
 
 def format_stdin_to_stdout(
-    line_length: int,
-    fast: bool,
-    write_back: WriteBack = WriteBack.NO,
-    mode: FileMode = FileMode.AUTO_DETECT,
+    fast: bool, *, write_back: WriteBack = WriteBack.NO, mode: Mode
 ) -> bool:
     """Format file on stdin. Return True if changed.
 
-    If `write_back` is True, write reformatted code back to stdout.
-    `line_length`, `fast`, `is_pyi`, and `force_py36` arguments are passed to
+    If `write_back` is YES, write reformatted code back to stdout. If it is DIFF,
+    write a diff to stdout. The `mode` argument is passed to
     :func:`format_file_contents`.
     """
-    src = sys.stdin.read()
+    then = datetime.utcnow()
+    src, encoding, newline = decode_bytes(sys.stdin.buffer.read())
     dst = src
     try:
-        dst = format_file_contents(src, line_length=line_length, fast=fast, mode=mode)
+        dst = format_file_contents(src, fast=fast, mode=mode)
         return True
 
     except NothingChanged:
         return False
 
     finally:
+        f = io.TextIOWrapper(
+            sys.stdout.buffer, encoding=encoding, newline=newline, write_through=True
+        )
         if write_back == WriteBack.YES:
-            sys.stdout.write(dst)
+            f.write(dst)
         elif write_back == WriteBack.DIFF:
-            src_name = "<stdin>  (original)"
-            dst_name = "<stdin>  (formatted)"
-            sys.stdout.write(diff(src, dst, src_name, dst_name))
+            now = datetime.utcnow()
+            src_name = f"STDIN\t{then} +0000"
+            dst_name = f"STDOUT\t{now} +0000"
+            f.write(diff(src, dst, src_name, dst_name))
+        f.detach()
 
 
-def format_file_contents(
-    src_contents: str,
-    *,
-    line_length: int,
-    fast: bool,
-    mode: FileMode = FileMode.AUTO_DETECT,
-) -> FileContent:
+def format_file_contents(src_contents: str, *, fast: bool, mode: Mode) -> FileContent:
     """Reformat contents a file and return new contents.
 
     If `fast` is False, additionally confirm that the reformatted code is
     valid by calling :func:`assert_equivalent` and :func:`assert_stable` on it.
-    `line_length` is passed to :func:`format_str`.
+    `mode` is passed to :func:`format_str`.
     """
     if src_contents.strip() == "":
         raise NothingChanged
 
-    dst_contents = format_str(src_contents, line_length=line_length, mode=mode)
+    dst_contents = format_str(src_contents, mode=mode)
     if src_contents == dst_contents:
         raise NothingChanged
 
     if not fast:
         assert_equivalent(src_contents, dst_contents)
-        assert_stable(src_contents, dst_contents, line_length=line_length, mode=mode)
+        assert_stable(src_contents, dst_contents, mode=mode)
     return dst_contents
 
 
-def format_str(
-    src_contents: str, line_length: int, *, mode: FileMode = FileMode.AUTO_DETECT
-) -> FileContent:
+def format_str(src_contents: str, *, mode: Mode) -> FileContent:
     """Reformat a string and return new contents.
 
-    `line_length` determines how many characters per line are allowed.
+    `mode` determines formatting options, such as how many characters per line are
+    allowed.  Example:
+
+    >>> import black
+    >>> print(black.format_str("def f(arg:str='')->None:...", mode=Mode()))
+    def f(arg: str = "") -> None:
+        ...
+
+    A more complex example:
+    >>> print(
+    ...   black.format_str(
+    ...     "def f(arg:str='')->None: hey",
+    ...     mode=black.Mode(
+    ...       target_versions={black.TargetVersion.PY36},
+    ...       line_length=10,
+    ...       string_normalization=False,
+    ...       is_pyi=False,
+    ...     ),
+    ...   ),
+    ... )
+    def f(
+        arg: str = '',
+    ) -> None:
+        hey
+
     """
-    src_node = lib2to3_parse(src_contents)
-    dst_contents = ""
+    src_node = lib2to3_parse(src_contents.lstrip(), mode.target_versions)
+    dst_contents = []
     future_imports = get_future_imports(src_node)
-    is_pyi = bool(mode & FileMode.PYI)
-    py36 = bool(mode & FileMode.PYTHON36) or is_python36(src_node)
-    normalize_strings = not bool(mode & FileMode.NO_STRING_NORMALIZATION)
+    if mode.target_versions:
+        versions = mode.target_versions
+    else:
+        versions = detect_target_versions(src_node)
+    normalize_fmt_off(src_node)
     lines = LineGenerator(
-        remove_u_prefix=py36 or "unicode_literals" in future_imports,
-        is_pyi=is_pyi,
-        normalize_strings=normalize_strings,
+        remove_u_prefix="unicode_literals" in future_imports
+        or supports_feature(versions, Feature.UNICODE_LITERALS),
+        is_pyi=mode.is_pyi,
+        normalize_strings=mode.string_normalization,
     )
-    elt = EmptyLineTracker(is_pyi=is_pyi)
+    elt = EmptyLineTracker(is_pyi=mode.is_pyi)
     empty_line = Line()
     after = 0
+    split_line_features = {
+        feature
+        for feature in {Feature.TRAILING_COMMA_IN_CALL, Feature.TRAILING_COMMA_IN_DEF}
+        if supports_feature(versions, feature)
+    }
     for current_line in lines.visit(src_node):
-        for _ in range(after):
-            dst_contents += str(empty_line)
+        dst_contents.append(str(empty_line) * after)
         before, after = elt.maybe_empty_lines(current_line)
-        for _ in range(before):
-            dst_contents += str(empty_line)
-        for line in split_line(current_line, line_length=line_length, py36=py36):
-            dst_contents += str(line)
-    return dst_contents
+        dst_contents.append(str(empty_line) * before)
+        for line in split_line(
+            current_line, line_length=mode.line_length, features=split_line_features
+        ):
+            dst_contents.append(str(line))
+    return "".join(dst_contents)
+
+
+def decode_bytes(src: bytes) -> Tuple[FileContent, Encoding, NewLine]:
+    """Return a tuple of (decoded_contents, encoding, newline).
 
+    `newline` is either CRLF or LF but `decoded_contents` is decoded with
+    universal newlines (i.e. only contains LF).
+    """
+    srcbuf = io.BytesIO(src)
+    encoding, lines = tokenize.detect_encoding(srcbuf.readline)
+    if not lines:
+        return "", encoding, "\n"
+
+    newline = "\r\n" if b"\r\n" == lines[0][-2:] else "\n"
+    srcbuf.seek(0)
+    with io.TextIOWrapper(srcbuf, encoding) as tiow:
+        return tiow.read(), encoding, newline
+
+
+def get_grammars(target_versions: Set[TargetVersion]) -> List[Grammar]:
+    if not target_versions:
+        # No target_version specified, so try all grammars.
+        return [
+            # Python 3.7+
+            pygram.python_grammar_no_print_statement_no_exec_statement_async_keywords,
+            # Python 3.0-3.6
+            pygram.python_grammar_no_print_statement_no_exec_statement,
+            # Python 2.7 with future print_function import
+            pygram.python_grammar_no_print_statement,
+            # Python 2.7
+            pygram.python_grammar,
+        ]
 
-GRAMMARS = [
-    pygram.python_grammar_no_print_statement_no_exec_statement,
-    pygram.python_grammar_no_print_statement,
-    pygram.python_grammar,
-]
+    if all(version.is_python2() for version in target_versions):
+        # Python 2-only code, so try Python 2 grammars.
+        return [
+            # Python 2.7 with future print_function import
+            pygram.python_grammar_no_print_statement,
+            # Python 2.7
+            pygram.python_grammar,
+        ]
+
+    # Python 3-compatible code, so only try Python 3 grammar.
+    grammars = []
+    # If we have to parse both, try to parse async as a keyword first
+    if not supports_feature(target_versions, Feature.ASYNC_IDENTIFIERS):
+        # Python 3.7+
+        grammars.append(
+            pygram.python_grammar_no_print_statement_no_exec_statement_async_keywords
+        )
+    if not supports_feature(target_versions, Feature.ASYNC_KEYWORDS):
+        # Python 3.0-3.6
+        grammars.append(pygram.python_grammar_no_print_statement_no_exec_statement)
+    # At least one of the above branches must have been taken, because every Python
+    # version has exactly one of the two 'ASYNC_*' flags
+    return grammars
 
 
-def lib2to3_parse(src_txt: str) -> Node:
+def lib2to3_parse(src_txt: str, target_versions: Iterable[TargetVersion] = ()) -> Node:
     """Given a string with source, return the lib2to3 Node."""
-    grammar = pygram.python_grammar_no_print_statement
-    if src_txt[-1] != "\n":
-        nl = "\r\n" if "\r\n" in src_txt[:1024] else "\n"
-        src_txt += nl
-    for grammar in GRAMMARS:
+    if src_txt[-1:] != "\n":
+        src_txt += "\n"
+
+    for grammar in get_grammars(set(target_versions)):
         drv = driver.Driver(grammar, pytree.convert)
         try:
             result = drv.parse_string(src_txt, True)
@@ -585,7 +876,7 @@ def lib2to3_parse(src_txt: str) -> Node:
                 faulty_line = lines[lineno - 1]
             except IndexError:
                 faulty_line = "<line number missing in source>"
-            exc = ValueError(f"Cannot parse: {lineno}:{column}: {faulty_line}")
+            exc = InvalidInput(f"Cannot parse: {lineno}:{column}: {faulty_line}")
     else:
         raise exc from None
 
@@ -619,8 +910,16 @@ class Visitor(Generic[T]):
         if node.type < 256:
             name = token.tok_name[node.type]
         else:
-            name = type_repr(node.type)
-        yield from getattr(self, f"visit_{name}", self.visit_default)(node)
+            name = str(type_repr(node.type))
+        # We explicitly branch on whether a visitor exists (instead of
+        # using self.visit_default as the default arg to getattr) in order
+        # to save needing to create a bound method object and so mypyc can
+        # generate a native call to visit_default.
+        visitf = getattr(self, f"visit_{name}", None)
+        if visitf:
+            yield from visitf(node)
+        else:
+            yield from self.visit_default(node)
 
     def visit_default(self, node: LN) -> Iterator[T]:
         """Default `visit_*()` implementation. Recurses to children of `node`."""
@@ -654,19 +953,19 @@ class DebugVisitor(Visitor[T]):
             out(f" {node.value!r}", fg="blue", bold=False)
 
     @classmethod
-    def show(cls, code: str) -> None:
+    def show(cls, code: Union[str, Leaf, Node]) -> None:
         """Pretty-print the lib2to3 AST of a given string of `code`.
 
         Convenience method for debugging.
         """
         v: DebugVisitor[None] = DebugVisitor()
-        list(v.visit(lib2to3_parse(code)))
+        if isinstance(code, str):
+            code = lib2to3_parse(code)
+        list(v.visit(code))
 
 
-KEYWORDS = set(keyword.kwlist)
-WHITESPACE = {token.DEDENT, token.INDENT, token.NEWLINE}
-FLOW_CONTROL = {"return", "raise", "break", "continue"}
-STATEMENT = {
+WHITESPACE: Final = {token.DEDENT, token.INDENT, token.NEWLINE}
+STATEMENT: Final = {
     syms.if_stmt,
     syms.while_stmt,
     syms.for_stmt,
@@ -676,9 +975,10 @@ STATEMENT = {
     syms.funcdef,
     syms.classdef,
 }
-STANDALONE_COMMENT = 153
-LOGIC_OPERATORS = {"and", "or"}
-COMPARATORS = {
+STANDALONE_COMMENT: Final = 153
+token.tok_name[STANDALONE_COMMENT] = "STANDALONE_COMMENT"
+LOGIC_OPERATORS: Final = {"and", "or"}
+COMPARATORS: Final = {
     token.LESS,
     token.GREATER,
     token.EQEQUAL,
@@ -686,7 +986,7 @@ COMPARATORS = {
     token.LESSEQUAL,
     token.GREATEREQUAL,
 }
-MATH_OPERATORS = {
+MATH_OPERATORS: Final = {
     token.VBAR,
     token.CIRCUMFLEX,
     token.AMPER,
@@ -702,21 +1002,23 @@ MATH_OPERATORS = {
     token.TILDE,
     token.DOUBLESTAR,
 }
-STARS = {token.STAR, token.DOUBLESTAR}
-VARARGS_PARENTS = {
+STARS: Final = {token.STAR, token.DOUBLESTAR}
+VARARGS_SPECIALS: Final = STARS | {token.SLASH}
+VARARGS_PARENTS: Final = {
     syms.arglist,
     syms.argument,  # double star in arglist
     syms.trailer,  # single argument to call
     syms.typedargslist,
     syms.varargslist,  # lambdas
 }
-UNPACKING_PARENTS = {
+UNPACKING_PARENTS: Final = {
     syms.atom,  # single element of a list or set literal
     syms.dictsetmaker,
     syms.listmaker,
     syms.testlist_gexp,
+    syms.testlist_star_expr,
 }
-TEST_DESCENDANTS = {
+TEST_DESCENDANTS: Final = {
     syms.test,
     syms.lambdef,
     syms.or_test,
@@ -733,7 +1035,7 @@ TEST_DESCENDANTS = {
     syms.term,
     syms.power,
 }
-ASSIGNMENTS = {
+ASSIGNMENTS: Final = {
     "=",
     "+=",
     "-=",
@@ -749,13 +1051,13 @@ ASSIGNMENTS = {
     "**=",
     "//=",
 }
-COMPREHENSION_PRIORITY = 20
-COMMA_PRIORITY = 18
-TERNARY_PRIORITY = 16
-LOGIC_PRIORITY = 14
-STRING_PRIORITY = 12
-COMPARATOR_PRIORITY = 10
-MATH_PRIORITIES = {
+COMPREHENSION_PRIORITY: Final = 20
+COMMA_PRIORITY: Final = 18
+TERNARY_PRIORITY: Final = 16
+LOGIC_PRIORITY: Final = 14
+STRING_PRIORITY: Final = 12
+COMPARATOR_PRIORITY: Final = 10
+MATH_PRIORITIES: Final = {
     token.VBAR: 9,
     token.CIRCUMFLEX: 8,
     token.AMPER: 7,
@@ -771,7 +1073,7 @@ MATH_PRIORITIES = {
     token.TILDE: 3,
     token.DOUBLESTAR: 2,
 }
-DOT_PRIORITY = 1
+DOT_PRIORITY: Final = 1
 
 
 @dataclass
@@ -779,11 +1081,11 @@ class BracketTracker:
     """Keeps track of brackets on a line."""
 
     depth: int = 0
-    bracket_match: Dict[Tuple[Depth, NodeType], Leaf] = Factory(dict)
-    delimiters: Dict[LeafID, Priority] = Factory(dict)
+    bracket_match: Dict[Tuple[Depth, NodeType], Leaf] = field(default_factory=dict)
+    delimiters: Dict[LeafID, Priority] = field(default_factory=dict)
     previous: Optional[Leaf] = None
-    _for_loop_variable: int = 0
-    _lambda_arguments: int = 0
+    _for_loop_depths: List[int] = field(default_factory=list)
+    _lambda_argument_depths: List[int] = field(default_factory=list)
 
     def mark(self, leaf: Leaf) -> None:
         """Mark `leaf` with bracket-related metadata. Keep track of delimiters.
@@ -829,7 +1131,7 @@ class BracketTracker:
         """Return True if there is an yet unmatched open bracket on the line."""
         return bool(self.bracket_match)
 
-    def max_delimiter_priority(self, exclude: Iterable[LeafID] = ()) -> int:
+    def max_delimiter_priority(self, exclude: Iterable[LeafID] = ()) -> Priority:
         """Return the highest priority of a delimiter found on the line.
 
         Values are consistent with what `is_split_*_delimiter()` return.
@@ -837,7 +1139,7 @@ class BracketTracker:
         """
         return max(v for k, v in self.delimiters.items() if k not in exclude)
 
-    def delimiter_count_with_priority(self, priority: int = 0) -> int:
+    def delimiter_count_with_priority(self, priority: Priority = 0) -> int:
         """Return the number of delimiters with the given `priority`.
 
         If no `priority` is passed, defaults to max priority on the line.
@@ -856,16 +1158,21 @@ class BracketTracker:
         """
         if leaf.type == token.NAME and leaf.value == "for":
             self.depth += 1
-            self._for_loop_variable += 1
+            self._for_loop_depths.append(self.depth)
             return True
 
         return False
 
     def maybe_decrement_after_for_loop_variable(self, leaf: Leaf) -> bool:
         """See `maybe_increment_for_loop_variable` above for explanation."""
-        if self._for_loop_variable and leaf.type == token.NAME and leaf.value == "in":
+        if (
+            self._for_loop_depths
+            and self._for_loop_depths[-1] == self.depth
+            and leaf.type == token.NAME
+            and leaf.value == "in"
+        ):
             self.depth -= 1
-            self._for_loop_variable -= 1
+            self._for_loop_depths.pop()
             return True
 
         return False
@@ -878,16 +1185,20 @@ class BracketTracker:
         """
         if leaf.type == token.NAME and leaf.value == "lambda":
             self.depth += 1
-            self._lambda_arguments += 1
+            self._lambda_argument_depths.append(self.depth)
             return True
 
         return False
 
     def maybe_decrement_after_lambda_arguments(self, leaf: Leaf) -> bool:
         """See `maybe_increment_lambda_arguments` above for explanation."""
-        if self._lambda_arguments and leaf.type == token.COLON:
+        if (
+            self._lambda_argument_depths
+            and self._lambda_argument_depths[-1] == self.depth
+            and leaf.type == token.COLON
+        ):
             self.depth -= 1
-            self._lambda_arguments -= 1
+            self._lambda_argument_depths.pop()
             return True
 
         return False
@@ -902,9 +1213,10 @@ class Line:
     """Holds leaves and comments. Can be printed with `str(line)`."""
 
     depth: int = 0
-    leaves: List[Leaf] = Factory(list)
-    comments: List[Tuple[Index, Leaf]] = Factory(list)
-    bracket_tracker: BracketTracker = Factory(BracketTracker)
+    leaves: List[Leaf] = field(default_factory=list)
+    # keys ordered like `leaves`
+    comments: Dict[LeafID, List[Leaf]] = field(default_factory=dict)
+    bracket_tracker: BracketTracker = field(default_factory=BracketTracker)
     inside_brackets: bool = False
     should_explode: bool = False
 
@@ -984,6 +1296,69 @@ class Line:
             Leaf(token.DOT, ".") for _ in range(3)
         ]
 
+    @property
+    def is_collection_with_optional_trailing_comma(self) -> bool:
+        """Is this line a collection literal with a trailing comma that's optional?
+
+        Note that the trailing comma in a 1-tuple is not optional.
+        """
+        if not self.leaves or len(self.leaves) < 4:
+            return False
+
+        # Look for and address a trailing colon.
+        if self.leaves[-1].type == token.COLON:
+            closer = self.leaves[-2]
+            close_index = -2
+        else:
+            closer = self.leaves[-1]
+            close_index = -1
+        if closer.type not in CLOSING_BRACKETS or self.inside_brackets:
+            return False
+
+        if closer.type == token.RPAR:
+            # Tuples require an extra check, because if there's only
+            # one element in the tuple removing the comma unmakes the
+            # tuple.
+            #
+            # We also check for parens before looking for the trailing
+            # comma because in some cases (eg assigning a dict
+            # literal) the literal gets wrapped in temporary parens
+            # during parsing. This case is covered by the
+            # collections.py test data.
+            opener = closer.opening_bracket
+            for _open_index, leaf in enumerate(self.leaves):
+                if leaf is opener:
+                    break
+
+            else:
+                # Couldn't find the matching opening paren, play it safe.
+                return False
+
+            commas = 0
+            comma_depth = self.leaves[close_index - 1].bracket_depth
+            for leaf in self.leaves[_open_index + 1 : close_index]:
+                if leaf.bracket_depth == comma_depth and leaf.type == token.COMMA:
+                    commas += 1
+            if commas > 1:
+                # We haven't looked yet for the trailing comma because
+                # we might also have caught noop parens.
+                return self.leaves[close_index - 1].type == token.COMMA
+
+            elif commas == 1:
+                return False  # it's either a one-tuple or didn't have a trailing comma
+
+            if self.leaves[close_index - 1].type in CLOSING_BRACKETS:
+                close_index -= 1
+                closer = self.leaves[close_index]
+                if closer.type == token.RPAR:
+                    # TODO: this is a gut feeling. Will we ever see this?
+                    return False
+
+        if self.leaves[close_index - 1].type != token.COMMA:
+            return False
+
+        return True
+
     @property
     def is_def(self) -> bool:
         """Is this a function definition? (Also returns True for async defs.)"""
@@ -1031,65 +1406,117 @@ class Line:
     def contains_standalone_comments(self, depth_limit: int = sys.maxsize) -> bool:
         """If so, needs to be split before emitting."""
         for leaf in self.leaves:
-            if leaf.type == STANDALONE_COMMENT:
-                if leaf.bracket_depth <= depth_limit:
-                    return True
+            if leaf.type == STANDALONE_COMMENT and leaf.bracket_depth <= depth_limit:
+                return True
 
         return False
 
-    def maybe_remove_trailing_comma(self, closing: Leaf) -> bool:
-        """Remove trailing comma if there is one and it's safe."""
-        if not (
-            self.leaves
-            and self.leaves[-1].type == token.COMMA
-            and closing.type in CLOSING_BRACKETS
-        ):
+    def contains_uncollapsable_type_comments(self) -> bool:
+        ignored_ids = set()
+        try:
+            last_leaf = self.leaves[-1]
+            ignored_ids.add(id(last_leaf))
+            if last_leaf.type == token.COMMA or (
+                last_leaf.type == token.RPAR and not last_leaf.value
+            ):
+                # When trailing commas or optional parens are inserted by Black for
+                # consistency, comments after the previous last element are not moved
+                # (they don't have to, rendering will still be correct).  So we ignore
+                # trailing commas and invisible.
+                last_leaf = self.leaves[-2]
+                ignored_ids.add(id(last_leaf))
+        except IndexError:
             return False
 
-        if closing.type == token.RBRACE:
-            self.remove_trailing_comma()
-            return True
+        # A type comment is uncollapsable if it is attached to a leaf
+        # that isn't at the end of the line (since that could cause it
+        # to get associated to a different argument) or if there are
+        # comments before it (since that could cause it to get hidden
+        # behind a comment.
+        comment_seen = False
+        for leaf_id, comments in self.comments.items():
+            for comment in comments:
+                if is_type_comment(comment):
+                    if comment_seen or (
+                        not is_type_comment(comment, " ignore")
+                        and leaf_id not in ignored_ids
+                    ):
+                        return True
+
+                comment_seen = True
 
-        if closing.type == token.RSQB:
-            comma = self.leaves[-1]
-            if comma.parent and comma.parent.type == syms.listmaker:
-                self.remove_trailing_comma()
-                return True
+        return False
 
-        # For parens let's check if it's safe to remove the comma.
-        # Imports are always safe.
-        if self.is_import:
-            self.remove_trailing_comma()
-            return True
+    def contains_unsplittable_type_ignore(self) -> bool:
+        if not self.leaves:
+            return False
 
-        # Otheriwsse, if the trailing one is the only one, we might mistakenly
-        # change a tuple into a different type by removing the comma.
-        depth = closing.bracket_depth + 1
-        commas = 0
-        opening = closing.opening_bracket
-        for _opening_index, leaf in enumerate(self.leaves):
-            if leaf is opening:
-                break
+        # If a 'type: ignore' is attached to the end of a line, we
+        # can't split the line, because we can't know which of the
+        # subexpressions the ignore was meant to apply to.
+        #
+        # We only want this to apply to actual physical lines from the
+        # original source, though: we don't want the presence of a
+        # 'type: ignore' at the end of a multiline expression to
+        # justify pushing it all onto one line. Thus we
+        # (unfortunately) need to check the actual source lines and
+        # only report an unsplittable 'type: ignore' if this line was
+        # one line in the original code.
+
+        # Grab the first and last line numbers, skipping generated leaves
+        first_line = next((l.lineno for l in self.leaves if l.lineno != 0), 0)
+        last_line = next((l.lineno for l in reversed(self.leaves) if l.lineno != 0), 0)
+
+        if first_line == last_line:
+            # We look at the last two leaves since a comma or an
+            # invisible paren could have been added at the end of the
+            # line.
+            for node in self.leaves[-2:]:
+                for comment in self.comments.get(id(node), []):
+                    if is_type_comment(comment, " ignore"):
+                        return True
 
-        else:
-            return False
+        return False
 
-        for leaf in self.leaves[_opening_index + 1 :]:
-            if leaf is closing:
-                break
+    def contains_multiline_strings(self) -> bool:
+        return any(is_multiline_string(leaf) for leaf in self.leaves)
 
-            bracket_depth = leaf.bracket_depth
-            if bracket_depth == depth and leaf.type == token.COMMA:
-                commas += 1
-                if leaf.parent and leaf.parent.type == syms.arglist:
-                    commas += 1
-                    break
+    def maybe_remove_trailing_comma(self, closing: Leaf) -> bool:
+        """Remove trailing comma if there is one and it's safe."""
+        if not (self.leaves and self.leaves[-1].type == token.COMMA):
+            return False
 
-        if commas > 1:
-            self.remove_trailing_comma()
-            return True
+        # We remove trailing commas only in the case of importing a
+        # single name from a module.
+        if not (
+            self.leaves
+            and self.is_import
+            and len(self.leaves) > 4
+            and self.leaves[-1].type == token.COMMA
+            and closing.type in CLOSING_BRACKETS
+            and self.leaves[-4].type == token.NAME
+            and (
+                # regular `from foo import bar,`
+                self.leaves[-4].value == "import"
+                # `from foo import (bar as baz,)
+                or (
+                    len(self.leaves) > 6
+                    and self.leaves[-6].value == "import"
+                    and self.leaves[-3].value == "as"
+                )
+                # `from foo import bar as baz,`
+                or (
+                    len(self.leaves) > 5
+                    and self.leaves[-5].value == "import"
+                    and self.leaves[-3].value == "as"
+                )
+            )
+            and closing.type == token.RPAR
+        ):
+            return False
 
-        return False
+        self.remove_trailing_comma()
+        return True
 
     def append_comment(self, comment: Leaf) -> bool:
         """Add an inline or standalone comment to the line."""
@@ -1103,56 +1530,57 @@ class Line:
         if comment.type != token.COMMENT:
             return False
 
-        after = len(self.leaves) - 1
-        if after == -1:
+        if not self.leaves:
             comment.type = STANDALONE_COMMENT
             comment.prefix = ""
             return False
 
-        else:
-            self.comments.append((after, comment))
-            return True
-
-    def comments_after(self, leaf: Leaf, _index: int = -1) -> Iterator[Leaf]:
-        """Generate comments that should appear directly after `leaf`.
-
-        Provide a non-negative leaf `_index` to speed up the function.
-        """
-        if _index == -1:
-            for _index, _leaf in enumerate(self.leaves):
-                if leaf is _leaf:
-                    break
-
-            else:
-                return
+        last_leaf = self.leaves[-1]
+        if (
+            last_leaf.type == token.RPAR
+            and not last_leaf.value
+            and last_leaf.parent
+            and len(list(last_leaf.parent.leaves())) <= 3
+            and not is_type_comment(comment)
+        ):
+            # Comments on an optional parens wrapping a single leaf should belong to
+            # the wrapped node except if it's a type comment. Pinning the comment like
+            # this avoids unstable formatting caused by comment migration.
+            if len(self.leaves) < 2:
+                comment.type = STANDALONE_COMMENT
+                comment.prefix = ""
+                return False
+
+            last_leaf = self.leaves[-2]
+        self.comments.setdefault(id(last_leaf), []).append(comment)
+        return True
 
-        for index, comment_after in self.comments:
-            if _index == index:
-                yield comment_after
+    def comments_after(self, leaf: Leaf) -> List[Leaf]:
+        """Generate comments that should appear directly after `leaf`."""
+        return self.comments.get(id(leaf), [])
 
     def remove_trailing_comma(self) -> None:
         """Remove the trailing comma and moves the comments attached to it."""
-        comma_index = len(self.leaves) - 1
-        for i in range(len(self.comments)):
-            comment_index, comment = self.comments[i]
-            if comment_index == comma_index:
-                self.comments[i] = (comma_index - 1, comment)
-        self.leaves.pop()
+        trailing_comma = self.leaves.pop()
+        trailing_comma_comments = self.comments.pop(id(trailing_comma), [])
+        self.comments.setdefault(id(self.leaves[-1]), []).extend(
+            trailing_comma_comments
+        )
 
     def is_complex_subscript(self, leaf: Leaf) -> bool:
         """Return True iff `leaf` is part of a slice with non-trivial exprs."""
-        open_lsqb = (
-            leaf if leaf.type == token.LSQB else self.bracket_tracker.get_open_lsqb()
-        )
+        open_lsqb = self.bracket_tracker.get_open_lsqb()
         if open_lsqb is None:
             return False
 
         subscript_start = open_lsqb.next_sibling
-        if (
-            isinstance(subscript_start, Node)
-            and subscript_start.type == syms.subscriptlist
-        ):
-            subscript_start = child_towards(subscript_start, leaf)
+
+        if isinstance(subscript_start, Node):
+            if subscript_start.type == syms.listmaker:
+                return False
+
+            if subscript_start.type == syms.subscriptlist:
+                subscript_start = child_towards(subscript_start, leaf)
         return subscript_start is not None and any(
             n.type in TEST_DESCENDANTS for n in subscript_start.pre_order()
         )
@@ -1168,7 +1596,7 @@ class Line:
         res = f"{first.prefix}{indent}{first.value}"
         for leaf in leaves:
             res += str(leaf)
-        for _, comment in self.comments:
+        for comment in itertools.chain.from_iterable(self.comments.values()):
             res += str(comment)
         return res + "\n"
 
@@ -1177,55 +1605,6 @@ class Line:
         return bool(self.leaves or self.comments)
 
 
-class UnformattedLines(Line):
-    """Just like :class:`Line` but stores lines which aren't reformatted."""
-
-    def append(self, leaf: Leaf, preformatted: bool = True) -> None:
-        """Just add a new `leaf` to the end of the lines.
-
-        The `preformatted` argument is ignored.
-
-        Keeps track of indentation `depth`, which is useful when the user
-        says `# fmt: on`. Otherwise, doesn't do anything with the `leaf`.
-        """
-        try:
-            list(generate_comments(leaf))
-        except FormatOn as f_on:
-            self.leaves.append(f_on.leaf_from_consumed(leaf))
-            raise
-
-        self.leaves.append(leaf)
-        if leaf.type == token.INDENT:
-            self.depth += 1
-        elif leaf.type == token.DEDENT:
-            self.depth -= 1
-
-    def __str__(self) -> str:
-        """Render unformatted lines from leaves which were added with `append()`.
-
-        `depth` is not used for indentation in this case.
-        """
-        if not self:
-            return "\n"
-
-        res = ""
-        for leaf in self.leaves:
-            res += str(leaf)
-        return res
-
-    def append_comment(self, comment: Leaf) -> bool:
-        """Not implemented in this class. Raises `NotImplementedError`."""
-        raise NotImplementedError("Unformatted lines don't store comments separately.")
-
-    def maybe_remove_trailing_comma(self, closing: Leaf) -> bool:
-        """Does nothing and returns False."""
-        return False
-
-    def maybe_increment_for_loop_variable(self, leaf: Leaf) -> bool:
-        """Does nothing and returns False."""
-        return False
-
-
 @dataclass
 class EmptyLineTracker:
     """Provides a stateful method that returns the number of potential extra
@@ -1239,7 +1618,7 @@ class EmptyLineTracker:
     is_pyi: bool = False
     previous_line: Optional[Line] = None
     previous_after: int = 0
-    previous_defs: List[int] = Factory(list)
+    previous_defs: List[int] = field(default_factory=list)
 
     def maybe_empty_lines(self, current_line: Line) -> Tuple[int, int]:
         """Return the number of extra empty lines before and after the `current_line`.
@@ -1247,11 +1626,14 @@ class EmptyLineTracker:
         This is for separating `def`, `async def` and `class` with extra empty
         lines (two on module-level).
         """
-        if isinstance(current_line, UnformattedLines):
-            return 0, 0
-
         before, after = self._maybe_empty_lines(current_line)
-        before -= self.previous_after
+        before = (
+            # Black should not insert empty lines at the beginning
+            # of the file
+            0
+            if self.previous_line is None
+            else before - self.previous_after
+        )
         self.previous_after = after
         self.previous_line = current_line
         return before, after
@@ -1275,44 +1657,8 @@ class EmptyLineTracker:
                 before = 0 if depth else 1
             else:
                 before = 1 if depth else 2
-        is_decorator = current_line.is_decorator
-        if is_decorator or current_line.is_def or current_line.is_class:
-            if not is_decorator:
-                self.previous_defs.append(depth)
-            if self.previous_line is None:
-                # Don't insert empty lines before the first line in the file.
-                return 0, 0
-
-            if self.previous_line.is_decorator:
-                return 0, 0
-
-            if self.previous_line.depth < current_line.depth and (
-                self.previous_line.is_class or self.previous_line.is_def
-            ):
-                return 0, 0
-
-            if (
-                self.previous_line.is_comment
-                and self.previous_line.depth == current_line.depth
-                and before == 0
-            ):
-                return 0, 0
-
-            if self.is_pyi:
-                if self.previous_line.depth > current_line.depth:
-                    newlines = 1
-                elif current_line.is_class or self.previous_line.is_class:
-                    if current_line.is_stub_class and self.previous_line.is_stub_class:
-                        newlines = 0
-                    else:
-                        newlines = 1
-                else:
-                    newlines = 0
-            else:
-                newlines = 2
-            if current_line.depth and newlines:
-                newlines -= 1
-            return newlines, 0
+        if current_line.is_decorator or current_line.is_def or current_line.is_class:
+            return self._maybe_empty_lines_for_class_or_def(current_line, before)
 
         if (
             self.previous_line
@@ -1331,6 +1677,50 @@ class EmptyLineTracker:
 
         return before, 0
 
+    def _maybe_empty_lines_for_class_or_def(
+        self, current_line: Line, before: int
+    ) -> Tuple[int, int]:
+        if not current_line.is_decorator:
+            self.previous_defs.append(current_line.depth)
+        if self.previous_line is None:
+            # Don't insert empty lines before the first line in the file.
+            return 0, 0
+
+        if self.previous_line.is_decorator:
+            return 0, 0
+
+        if self.previous_line.depth < current_line.depth and (
+            self.previous_line.is_class or self.previous_line.is_def
+        ):
+            return 0, 0
+
+        if (
+            self.previous_line.is_comment
+            and self.previous_line.depth == current_line.depth
+            and before == 0
+        ):
+            return 0, 0
+
+        if self.is_pyi:
+            if self.previous_line.depth > current_line.depth:
+                newlines = 1
+            elif current_line.is_class or self.previous_line.is_class:
+                if current_line.is_stub_class and self.previous_line.is_stub_class:
+                    # No blank line between classes with an empty body
+                    newlines = 0
+                else:
+                    newlines = 1
+            elif current_line.is_def and not self.previous_line.is_def:
+                # Blank line between a block of functions and a block of non-functions
+                newlines = 1
+            else:
+                newlines = 0
+        else:
+            newlines = 2
+        if current_line.depth and newlines:
+            newlines -= 1
+        return newlines, 0
+
 
 @dataclass
 class LineGenerator(Visitor[Line]):
@@ -1342,10 +1732,10 @@ class LineGenerator(Visitor[Line]):
 
     is_pyi: bool = False
     normalize_strings: bool = True
-    current_line: Line = Factory(Line)
+    current_line: Line = field(default_factory=Line)
     remove_u_prefix: bool = False
 
-    def line(self, indent: int = 0, type: Type[Line] = Line) -> Iterator[Line]:
+    def line(self, indent: int = 0) -> Iterator[Line]:
         """Generate a line.
 
         If the line is empty, only emit if it makes sense.
@@ -1354,76 +1744,50 @@ class LineGenerator(Visitor[Line]):
         If any lines were generated, set up a new current_line.
         """
         if not self.current_line:
-            if self.current_line.__class__ == type:
-                self.current_line.depth += indent
-            else:
-                self.current_line = type(depth=self.current_line.depth + indent)
+            self.current_line.depth += indent
             return  # Line is empty, don't emit. Creating a new one unnecessary.
 
         complete_line = self.current_line
-        self.current_line = type(depth=complete_line.depth + indent)
+        self.current_line = Line(depth=complete_line.depth + indent)
         yield complete_line
 
-    def visit(self, node: LN) -> Iterator[Line]:
-        """Main method to visit `node` and its children.
-
-        Yields :class:`Line` objects.
-        """
-        if isinstance(self.current_line, UnformattedLines):
-            # File contained `# fmt: off`
-            yield from self.visit_unformatted(node)
-
-        else:
-            yield from super().visit(node)
-
     def visit_default(self, node: LN) -> Iterator[Line]:
         """Default `visit_*()` implementation. Recurses to children of `node`."""
         if isinstance(node, Leaf):
             any_open_brackets = self.current_line.bracket_tracker.any_open_brackets()
-            try:
-                for comment in generate_comments(node):
-                    if any_open_brackets:
-                        # any comment within brackets is subject to splitting
-                        self.current_line.append(comment)
-                    elif comment.type == token.COMMENT:
-                        # regular trailing comment
-                        self.current_line.append(comment)
-                        yield from self.line()
-
-                    else:
-                        # regular standalone comment
-                        yield from self.line()
-
-                        self.current_line.append(comment)
-                        yield from self.line()
-
-            except FormatOff as f_off:
-                f_off.trim_prefix(node)
-                yield from self.line(type=UnformattedLines)
-                yield from self.visit(node)
-
-            except FormatOn as f_on:
-                # This only happens here if somebody says "fmt: on" multiple
-                # times in a row.
-                f_on.trim_prefix(node)
-                yield from self.visit_default(node)
+            for comment in generate_comments(node):
+                if any_open_brackets:
+                    # any comment within brackets is subject to splitting
+                    self.current_line.append(comment)
+                elif comment.type == token.COMMENT:
+                    # regular trailing comment
+                    self.current_line.append(comment)
+                    yield from self.line()
 
-            else:
-                normalize_prefix(node, inside_brackets=any_open_brackets)
-                if self.normalize_strings and node.type == token.STRING:
-                    normalize_string_prefix(node, remove_u_prefix=self.remove_u_prefix)
-                    normalize_string_quotes(node)
-                if node.type not in WHITESPACE:
-                    self.current_line.append(node)
+                else:
+                    # regular standalone comment
+                    yield from self.line()
+
+                    self.current_line.append(comment)
+                    yield from self.line()
+
+            normalize_prefix(node, inside_brackets=any_open_brackets)
+            if self.normalize_strings and node.type == token.STRING:
+                normalize_string_prefix(node, remove_u_prefix=self.remove_u_prefix)
+                normalize_string_quotes(node)
+            if node.type == token.NUMBER:
+                normalize_numeric_literal(node)
+            if node.type not in WHITESPACE:
+                self.current_line.append(node)
         yield from super().visit_default(node)
 
-    def visit_INDENT(self, node: Node) -> Iterator[Line]:
+    def visit_INDENT(self, node: Leaf) -> Iterator[Line]:
         """Increase indentation level, maybe yield a line."""
         # In blib2to3 INDENT never holds comments.
         yield from self.line(+1)
         yield from self.visit_default(node)
 
-    def visit_DEDENT(self, node: Node) -> Iterator[Line]:
+    def visit_DEDENT(self, node: Leaf) -> Iterator[Line]:
         """Decrease indentation level, maybe yield a line."""
         # The current line might still wait for trailing comments.  At DEDENT time
         # there won't be any (they would be prefixes on the preceding NEWLINE).
@@ -1511,25 +1875,29 @@ class LineGenerator(Visitor[Line]):
         yield from self.visit_default(leaf)
         yield from self.line()
 
-    def visit_unformatted(self, node: LN) -> Iterator[Line]:
-        """Used when file contained a `# fmt: off`."""
-        if isinstance(node, Node):
-            for child in node.children:
-                yield from self.visit(child)
+    def visit_STANDALONE_COMMENT(self, leaf: Leaf) -> Iterator[Line]:
+        if not self.current_line.bracket_tracker.any_open_brackets():
+            yield from self.line()
+        yield from self.visit_default(leaf)
 
-        else:
-            try:
-                self.current_line.append(node)
-            except FormatOn as f_on:
-                f_on.trim_prefix(node)
-                yield from self.line()
-                yield from self.visit(node)
+    def visit_factor(self, node: Node) -> Iterator[Line]:
+        """Force parentheses between a unary op and a binary power:
 
-            if node.type == token.ENDMARKER:
-                # somebody decided not to put a final `# fmt: on`
-                yield from self.line()
+        -2 ** 8 -> -(2 ** 8)
+        """
+        _operator, operand = node.children
+        if (
+            operand.type == syms.power
+            and len(operand.children) == 3
+            and operand.children[1].type == token.DOUBLESTAR
+        ):
+            lpar = Leaf(token.LPAR, "(")
+            rpar = Leaf(token.RPAR, ")")
+            index = operand.remove() or 0
+            node.insert_child(index, Node(syms.atom, [lpar, operand, rpar]))
+        yield from self.visit_default(node)
 
-    def __attrs_post_init__(self) -> None:
+    def __post_init__(self) -> None:
         """You are in a twisty little maze of passages."""
         v = self.visit_stmt
         Ø: Set[str] = set()
@@ -1549,6 +1917,7 @@ class LineGenerator(Visitor[Line]):
         self.visit_expr_stmt = partial(v, keywords=Ø, parens=ASSIGNMENTS)
         self.visit_return_stmt = partial(v, keywords={"return"}, parens={"return"})
         self.visit_import_from = partial(v, keywords=Ø, parens={"import"})
+        self.visit_del_stmt = partial(v, keywords=Ø, parens={"del"})
         self.visit_async_funcdef = self.visit_async_stmt
         self.visit_decorated = self.visit_decorators
 
@@ -1561,7 +1930,7 @@ BRACKETS = OPENING_BRACKETS | CLOSING_BRACKETS
 ALWAYS_NO_SPACE = CLOSING_BRACKETS | {token.COMMA, STANDALONE_COMMENT}
 
 
-def whitespace(leaf: Leaf, *, complex_subscript: bool) -> str:  # noqa C901
+def whitespace(leaf: Leaf, *, complex_subscript: bool) -> str:  # noqa: C901
     """Return whitespace prefix if needed for the given `leaf`.
 
     `complex_subscript` signals whether the given leaf is part of a subscription
@@ -1618,7 +1987,7 @@ def whitespace(leaf: Leaf, *, complex_subscript: bool) -> str:  # noqa C901
                     # that, too.
                     return prevp.prefix
 
-        elif prevp.type in STARS:
+        elif prevp.type in VARARGS_SPECIALS:
             if is_vararg(prevp, within=VARARGS_PARENTS | UNPACKING_PARENTS):
                 return NO
 
@@ -1708,7 +2077,7 @@ def whitespace(leaf: Leaf, *, complex_subscript: bool) -> str:  # noqa C901
             if not prevp or prevp.type == token.LPAR:
                 return NO
 
-        elif prev.type in {token.EQUAL} | STARS:
+        elif prev.type in {token.EQUAL} | VARARGS_SPECIALS:
             return NO
 
     elif p.type == syms.decorator:
@@ -1770,7 +2139,7 @@ def whitespace(leaf: Leaf, *, complex_subscript: bool) -> str:  # noqa C901
             elif prevp.type == token.EQUAL and prevp_parent.type == syms.argument:
                 return NO
 
-        elif t == token.NAME or t == token.NUMBER:
+        elif t in {token.NAME, token.NUMBER, token.STRING}:
             return NO
 
     elif p.type == syms.import_from:
@@ -1809,15 +2178,40 @@ def preceding_leaf(node: Optional[LN]) -> Optional[Leaf]:
     return None
 
 
-def child_towards(ancestor: Node, descendant: LN) -> Optional[LN]:
-    """Return the child of `ancestor` that contains `descendant`."""
-    node: Optional[LN] = descendant
-    while node and node.parent != ancestor:
-        node = node.parent
-    return node
+def child_towards(ancestor: Node, descendant: LN) -> Optional[LN]:
+    """Return the child of `ancestor` that contains `descendant`."""
+    node: Optional[LN] = descendant
+    while node and node.parent != ancestor:
+        node = node.parent
+    return node
+
+
+def container_of(leaf: Leaf) -> LN:
+    """Return `leaf` or one of its ancestors that is the topmost container of it.
+
+    By "container" we mean a node where `leaf` is the very first child.
+    """
+    same_prefix = leaf.prefix
+    container: LN = leaf
+    while container:
+        parent = container.parent
+        if parent is None:
+            break
+
+        if parent.children[0].prefix != same_prefix:
+            break
+
+        if parent.type == syms.file_input:
+            break
+
+        if parent.prev_sibling is not None and parent.prev_sibling.type in BRACKETS:
+            break
+
+        container = parent
+    return container
 
 
-def is_split_after_delimiter(leaf: Leaf, previous: Leaf = None) -> int:
+def is_split_after_delimiter(leaf: Leaf, previous: Optional[Leaf] = None) -> Priority:
     """Return the priority of the `leaf` delimiter, given a line break after it.
 
     The delimiter priorities returned here are from those delimiters that would
@@ -1831,8 +2225,8 @@ def is_split_after_delimiter(leaf: Leaf, previous: Leaf = None) -> int:
     return 0
 
 
-def is_split_before_delimiter(leaf: Leaf, previous: Leaf = None) -> int:
-    """Return the priority of the `leaf` delimiter, given a line before after it.
+def is_split_before_delimiter(leaf: Leaf, previous: Optional[Leaf] = None) -> Priority:
+    """Return the priority of the `leaf` delimiter, given a line break before it.
 
     The delimiter priorities returned here are from those delimiters that would
     cause a line break before themselves.
@@ -1869,15 +2263,20 @@ def is_split_before_delimiter(leaf: Leaf, previous: Leaf = None) -> int:
     ):
         return STRING_PRIORITY
 
-    if leaf.type != token.NAME:
+    if leaf.type not in {token.NAME, token.ASYNC}:
         return 0
 
     if (
         leaf.value == "for"
         and leaf.parent
         and leaf.parent.type in {syms.comp_for, syms.old_comp_for}
+        or leaf.type == token.ASYNC
     ):
-        return COMPREHENSION_PRIORITY
+        if (
+            not isinstance(leaf.prev_sibling, Leaf)
+            or leaf.prev_sibling.value != "async"
+        ):
+            return COMPREHENSION_PRIORITY
 
     if (
         leaf.value == "if"
@@ -1922,6 +2321,10 @@ def is_split_before_delimiter(leaf: Leaf, previous: Leaf = None) -> int:
     return 0
 
 
+FMT_OFF = {"# fmt: off", "# fmt:off", "# yapf: disable"}
+FMT_ON = {"# fmt: on", "# fmt:on", "# yapf: enable"}
+
+
 def generate_comments(leaf: LN) -> Iterator[Leaf]:
     """Clean the prefix of the `leaf` and generate comments from it, if any.
 
@@ -1941,49 +2344,70 @@ def generate_comments(leaf: LN) -> Iterator[Leaf]:
     Inline comments are emitted as regular token.COMMENT leaves.  Standalone
     are emitted with a fake STANDALONE_COMMENT token identifier.
     """
-    p = leaf.prefix
-    if not p:
-        return
+    for pc in list_comments(leaf.prefix, is_endmarker=leaf.type == token.ENDMARKER):
+        yield Leaf(pc.type, pc.value, prefix="\n" * pc.newlines)
 
-    if "#" not in p:
-        return
+
+@dataclass
+class ProtoComment:
+    """Describes a piece of syntax that is a comment.
+
+    It's not a :class:`blib2to3.pytree.Leaf` so that:
+
+    * it can be cached (`Leaf` objects should not be reused more than once as
+      they store their lineno, column, prefix, and parent information);
+    * `newlines` and `consumed` fields are kept separate from the `value`. This
+      simplifies handling of special marker comments like ``# fmt: off/on``.
+    """
+
+    type: int  # token.COMMENT or STANDALONE_COMMENT
+    value: str  # content of the comment
+    newlines: int  # how many newlines before the comment
+    consumed: int  # how many characters of the original leaf's prefix did we consume
+
+
+@lru_cache(maxsize=4096)
+def list_comments(prefix: str, *, is_endmarker: bool) -> List[ProtoComment]:
+    """Return a list of :class:`ProtoComment` objects parsed from the given `prefix`."""
+    result: List[ProtoComment] = []
+    if not prefix or "#" not in prefix:
+        return result
 
     consumed = 0
     nlines = 0
-    for index, line in enumerate(p.split("\n")):
+    ignored_lines = 0
+    for index, line in enumerate(prefix.split("\n")):
         consumed += len(line) + 1  # adding the length of the split '\n'
         line = line.lstrip()
         if not line:
             nlines += 1
         if not line.startswith("#"):
+            # Escaped newlines outside of a comment are not really newlines at
+            # all. We treat a single-line comment following an escaped newline
+            # as a simple trailing comment.
+            if line.endswith("\\"):
+                ignored_lines += 1
             continue
 
-        if index == 0 and leaf.type != token.ENDMARKER:
+        if index == ignored_lines and not is_endmarker:
             comment_type = token.COMMENT  # simple trailing comment
         else:
             comment_type = STANDALONE_COMMENT
         comment = make_comment(line)
-        yield Leaf(comment_type, comment, prefix="\n" * nlines)
-
-        if comment in {"# fmt: on", "# yapf: enable"}:
-            raise FormatOn(consumed)
-
-        if comment in {"# fmt: off", "# yapf: disable"}:
-            if comment_type == STANDALONE_COMMENT:
-                raise FormatOff(consumed)
-
-            prev = preceding_leaf(leaf)
-            if not prev or prev.type in WHITESPACE:  # standalone comment in disguise
-                raise FormatOff(consumed)
-
+        result.append(
+            ProtoComment(
+                type=comment_type, value=comment, newlines=nlines, consumed=consumed
+            )
+        )
         nlines = 0
+    return result
 
 
 def make_comment(content: str) -> str:
     """Return a consistently formatted comment from the given `content` string.
 
-    All comments (except for "##", "#!", "#:") should have a single space between
-    the hash sign and the content.
+    All comments (except for "##", "#!", "#:", '#'", "#%%") should have a single
+    space between the hash sign and the content.
 
     If `content` didn't start with a hash sign, one is provided.
     """
@@ -1993,13 +2417,16 @@ def make_comment(content: str) -> str:
 
     if content[0] == "#":
         content = content[1:]
-    if content and content[0] not in " !:#":
+    if content and content[0] not in " !:#'%":
         content = " " + content
     return "#" + content
 
 
 def split_line(
-    line: Line, line_length: int, inner: bool = False, py36: bool = False
+    line: Line,
+    line_length: int,
+    inner: bool = False,
+    features: Collection[Feature] = (),
 ) -> Iterator[Line]:
     """Split a `line` into potentially many lines.
 
@@ -2008,16 +2435,22 @@ def split_line(
     current `line`, possibly transitively. This means we can fallback to splitting
     by delimiters if the LHS/RHS don't yield any results.
 
-    If `py36` is True, splitting may generate syntax that is only compatible
-    with Python 3.6 and later.
+    `features` are syntactical features that may be used in the output.
     """
-    if isinstance(line, UnformattedLines) or line.is_comment:
+    if line.is_comment:
         yield line
         return
 
     line_str = str(line).strip("\n")
-    if not line.should_explode and is_line_short_enough(
-        line, line_length=line_length, line_str=line_str
+
+    if (
+        not line.contains_uncollapsable_type_comments()
+        and not line.should_explode
+        and not line.is_collection_with_optional_trailing_comma
+        and (
+            is_line_short_enough(line, line_length=line_length, line_str=line_str)
+            or line.contains_unsplittable_type_ignore()
+        )
     ):
         yield line
         return
@@ -2027,9 +2460,9 @@ def split_line(
         split_funcs = [left_hand_split]
     else:
 
-        def rhs(line: Line, py36: bool = False) -> Iterator[Line]:
+        def rhs(line: Line, features: Collection[Feature]) -> Iterator[Line]:
             for omit in generate_trailers_to_omit(line, line_length):
-                lines = list(right_hand_split(line, line_length, py36, omit=omit))
+                lines = list(right_hand_split(line, line_length, features, omit=omit))
                 if is_line_short_enough(lines[0], line_length=line_length):
                     yield from lines
                     return
@@ -2037,7 +2470,9 @@ def split_line(
             # All splits failed, best effort split with no omits.
             # This mostly happens to multiline strings that are by definition
             # reported as not fitting a single line.
-            yield from right_hand_split(line, py36)
+            # line_length=1 here was historically a bug that somehow became a feature.
+            # See #762 and #781 for the full story.
+            yield from right_hand_split(line, line_length=1, features=features)
 
         if line.inside_brackets:
             split_funcs = [delimiter_split, standalone_comment_split, rhs]
@@ -2049,14 +2484,16 @@ def split_line(
         # split altogether.
         result: List[Line] = []
         try:
-            for l in split_func(line, py36):
+            for l in split_func(line, features):
                 if str(l).strip("\n") == line_str:
                     raise CannotSplit("Split function returned an unchanged result")
 
                 result.extend(
-                    split_line(l, line_length=line_length, inner=True, py36=py36)
+                    split_line(
+                        l, line_length=line_length, inner=True, features=features
+                    )
                 )
-        except CannotSplit as cs:
+        except CannotSplit:
             continue
 
         else:
@@ -2067,21 +2504,18 @@ def split_line(
         yield line
 
 
-def left_hand_split(line: Line, py36: bool = False) -> Iterator[Line]:
+def left_hand_split(line: Line, features: Collection[Feature] = ()) -> Iterator[Line]:
     """Split line into many lines, starting with the first matching bracket pair.
 
     Note: this usually looks weird, only use this for function definitions.
     Prefer RHS otherwise.  This is why this function is not symmetrical with
     :func:`right_hand_split` which also handles optional parentheses.
     """
-    head = Line(depth=line.depth)
-    body = Line(depth=line.depth + 1, inside_brackets=True)
-    tail = Line(depth=line.depth)
     tail_leaves: List[Leaf] = []
     body_leaves: List[Leaf] = []
     head_leaves: List[Leaf] = []
     current_leaves = head_leaves
-    matching_bracket = None
+    matching_bracket: Optional[Leaf] = None
     for leaf in line.leaves:
         if (
             current_leaves is body_leaves
@@ -2094,15 +2528,12 @@ def left_hand_split(line: Line, py36: bool = False) -> Iterator[Line]:
             if leaf.type in OPENING_BRACKETS:
                 matching_bracket = leaf
                 current_leaves = body_leaves
-    # Since body is a new indent level, remove spurious leading whitespace.
-    if body_leaves:
-        normalize_prefix(body_leaves[0], inside_brackets=True)
-    # Build the new lines.
-    for result, leaves in (head, head_leaves), (body, body_leaves), (tail, tail_leaves):
-        for leaf in leaves:
-            result.append(leaf, preformatted=True)
-            for comment_after in line.comments_after(leaf):
-                result.append(comment_after, preformatted=True)
+    if not matching_bracket:
+        raise CannotSplit("No brackets found")
+
+    head = bracket_split_build_line(head_leaves, line, matching_bracket)
+    body = bracket_split_build_line(body_leaves, line, matching_bracket, is_body=True)
+    tail = bracket_split_build_line(tail_leaves, line, matching_bracket)
     bracket_split_succeeded_or_raise(head, body, tail)
     for result in (head, body, tail):
         if result:
@@ -2110,7 +2541,10 @@ def left_hand_split(line: Line, py36: bool = False) -> Iterator[Line]:
 
 
 def right_hand_split(
-    line: Line, line_length: int, py36: bool = False, omit: Collection[LeafID] = ()
+    line: Line,
+    line_length: int,
+    features: Collection[Feature] = (),
+    omit: Collection[LeafID] = (),
 ) -> Iterator[Line]:
     """Split line into many lines, starting with the last matching bracket pair.
 
@@ -2120,15 +2554,12 @@ def right_hand_split(
 
     Note: running this function modifies `bracket_depth` on the leaves of `line`.
     """
-    head = Line(depth=line.depth)
-    body = Line(depth=line.depth + 1, inside_brackets=True)
-    tail = Line(depth=line.depth)
     tail_leaves: List[Leaf] = []
     body_leaves: List[Leaf] = []
     head_leaves: List[Leaf] = []
     current_leaves = tail_leaves
-    opening_bracket = None
-    closing_bracket = None
+    opening_bracket: Optional[Leaf] = None
+    closing_bracket: Optional[Leaf] = None
     for leaf in reversed(line.leaves):
         if current_leaves is body_leaves:
             if leaf is opening_bracket:
@@ -2139,49 +2570,60 @@ def right_hand_split(
                 opening_bracket = leaf.opening_bracket
                 closing_bracket = leaf
                 current_leaves = body_leaves
-    tail_leaves.reverse()
-    body_leaves.reverse()
-    head_leaves.reverse()
-    # Since body is a new indent level, remove spurious leading whitespace.
-    if body_leaves:
-        normalize_prefix(body_leaves[0], inside_brackets=True)
-    if not head_leaves:
-        # No `head` means the split failed. Either `tail` has all content or
+    if not (opening_bracket and closing_bracket and head_leaves):
+        # If there is no opening or closing_bracket that means the split failed and
+        # all content is in the tail.  Otherwise, if `head_leaves` are empty, it means
         # the matching `opening_bracket` wasn't available on `line` anymore.
         raise CannotSplit("No brackets found")
 
-    # Build the new lines.
-    for result, leaves in (head, head_leaves), (body, body_leaves), (tail, tail_leaves):
-        for leaf in leaves:
-            result.append(leaf, preformatted=True)
-            for comment_after in line.comments_after(leaf):
-                result.append(comment_after, preformatted=True)
+    tail_leaves.reverse()
+    body_leaves.reverse()
+    head_leaves.reverse()
+    head = bracket_split_build_line(head_leaves, line, opening_bracket)
+    body = bracket_split_build_line(body_leaves, line, opening_bracket, is_body=True)
+    tail = bracket_split_build_line(tail_leaves, line, opening_bracket)
     bracket_split_succeeded_or_raise(head, body, tail)
-    assert opening_bracket and closing_bracket
     if (
+        # the body shouldn't be exploded
+        not body.should_explode
         # the opening bracket is an optional paren
-        opening_bracket.type == token.LPAR
+        and opening_bracket.type == token.LPAR
         and not opening_bracket.value
         # the closing bracket is an optional paren
         and closing_bracket.type == token.RPAR
         and not closing_bracket.value
-        # there are no standalone comments in the body
-        and not line.contains_standalone_comments(0)
-        # and it's not an import (optional parens are the only thing we can split
-        # on in this case; attempting a split without them is a waste of time)
+        # it's not an import (optional parens are the only thing we can split on
+        # in this case; attempting a split without them is a waste of time)
         and not line.is_import
+        # there are no standalone comments in the body
+        and not body.contains_standalone_comments(0)
+        # and we can actually remove the parens
+        and can_omit_invisible_parens(body, line_length)
     ):
         omit = {id(closing_bracket), *omit}
-        if can_omit_invisible_parens(body, line_length):
-            try:
-                yield from right_hand_split(line, line_length, py36=py36, omit=omit)
-                return
-            except CannotSplit:
-                pass
+        try:
+            yield from right_hand_split(line, line_length, features=features, omit=omit)
+            return
+
+        except CannotSplit:
+            if not (
+                can_be_split(body)
+                or is_line_short_enough(body, line_length=line_length)
+            ):
+                raise CannotSplit(
+                    "Splitting failed, body is still too long and can't be split."
+                )
+
+            elif head.contains_multiline_strings() or tail.contains_multiline_strings():
+                raise CannotSplit(
+                    "The current optional pair of parentheses is bound to fail to "
+                    "satisfy the splitting algorithm because the head or the tail "
+                    "contains multiline strings which by definition never fit one "
+                    "line."
+                )
 
     ensure_visible(opening_bracket)
     ensure_visible(closing_bracket)
-    body.should_explode = should_explode(body, opening_bracket)
     for result in (head, body, tail):
         if result:
             yield result
@@ -2213,6 +2655,48 @@ def bracket_split_succeeded_or_raise(head: Line, body: Line, tail: Line) -> None
             )
 
 
+def bracket_split_build_line(
+    leaves: List[Leaf], original: Line, opening_bracket: Leaf, *, is_body: bool = False
+) -> Line:
+    """Return a new line with given `leaves` and respective comments from `original`.
+
+    If `is_body` is True, the result line is one-indented inside brackets and as such
+    has its first leaf's prefix normalized and a trailing comma added when expected.
+    """
+    result = Line(depth=original.depth)
+    if is_body:
+        result.inside_brackets = True
+        result.depth += 1
+        if leaves:
+            # Since body is a new indent level, remove spurious leading whitespace.
+            normalize_prefix(leaves[0], inside_brackets=True)
+            # Ensure a trailing comma for imports and standalone function arguments, but
+            # be careful not to add one after any comments or within type annotations.
+            no_commas = (
+                original.is_def
+                and opening_bracket.value == "("
+                and not any(l.type == token.COMMA for l in leaves)
+            )
+
+            if original.is_import or no_commas:
+                for i in range(len(leaves) - 1, -1, -1):
+                    if leaves[i].type == STANDALONE_COMMENT:
+                        continue
+
+                    if leaves[i].type != token.COMMA:
+                        leaves.insert(i + 1, Leaf(token.COMMA, ","))
+                    break
+
+    # Populate the line
+    for leaf in leaves:
+        result.append(leaf, preformatted=True)
+        for comment_after in original.comments_after(leaf):
+            result.append(comment_after, preformatted=True)
+    if is_body:
+        result.should_explode = should_explode(result, opening_bracket)
+    return result
+
+
 def dont_increase_indentation(split_func: SplitFunc) -> SplitFunc:
     """Normalize prefix of the first leaf in every line returned by `split_func`.
 
@@ -2220,8 +2704,8 @@ def dont_increase_indentation(split_func: SplitFunc) -> SplitFunc:
     """
 
     @wraps(split_func)
-    def split_wrapper(line: Line, py36: bool = False) -> Iterator[Line]:
-        for l in split_func(line, py36):
+    def split_wrapper(line: Line, features: Collection[Feature] = ()) -> Iterator[Line]:
+        for l in split_func(line, features):
             normalize_prefix(l.leaves[0], inside_brackets=True)
             yield l
 
@@ -2229,11 +2713,11 @@ def dont_increase_indentation(split_func: SplitFunc) -> SplitFunc:
 
 
 @dont_increase_indentation
-def delimiter_split(line: Line, py36: bool = False) -> Iterator[Line]:
+def delimiter_split(line: Line, features: Collection[Feature] = ()) -> Iterator[Line]:
     """Split according to delimiters of the highest priority.
 
-    If `py36` is True, the split will add trailing commas also in function
-    signatures that contain `*` and `**`.
+    If the appropriate Features are given, the split will add trailing commas
+    also in function signatures and calls that contain `*` and `**`.
     """
     try:
         last_leaf = line.leaves[-1]
@@ -2259,23 +2743,29 @@ def delimiter_split(line: Line, py36: bool = False) -> Iterator[Line]:
         nonlocal current_line
         try:
             current_line.append_safe(leaf, preformatted=True)
-        except ValueError as ve:
+        except ValueError:
             yield current_line
 
             current_line = Line(depth=line.depth, inside_brackets=line.inside_brackets)
             current_line.append(leaf)
 
-    for index, leaf in enumerate(line.leaves):
+    for leaf in line.leaves:
         yield from append_to_line(leaf)
 
-        for comment_after in line.comments_after(leaf, index):
+        for comment_after in line.comments_after(leaf):
             yield from append_to_line(comment_after)
 
         lowest_depth = min(lowest_depth, leaf.bracket_depth)
-        if leaf.bracket_depth == lowest_depth and is_vararg(
-            leaf, within=VARARGS_PARENTS
-        ):
-            trailing_comma_safe = trailing_comma_safe and py36
+        if leaf.bracket_depth == lowest_depth:
+            if is_vararg(leaf, within={syms.typedargslist}):
+                trailing_comma_safe = (
+                    trailing_comma_safe and Feature.TRAILING_COMMA_IN_DEF in features
+                )
+            elif is_vararg(leaf, within={syms.arglist, syms.argument}):
+                trailing_comma_safe = (
+                    trailing_comma_safe and Feature.TRAILING_COMMA_IN_CALL in features
+                )
+
         leaf_priority = bt.delimiters.get(id(leaf))
         if leaf_priority == delimiter_priority:
             yield current_line
@@ -2293,7 +2783,9 @@ def delimiter_split(line: Line, py36: bool = False) -> Iterator[Line]:
 
 
 @dont_increase_indentation
-def standalone_comment_split(line: Line, py36: bool = False) -> Iterator[Line]:
+def standalone_comment_split(
+    line: Line, features: Collection[Feature] = ()
+) -> Iterator[Line]:
     """Split standalone comments from the rest of the line."""
     if not line.contains_standalone_comments(0):
         raise CannotSplit("Line does not have any standalone comments")
@@ -2305,16 +2797,16 @@ def standalone_comment_split(line: Line, py36: bool = False) -> Iterator[Line]:
         nonlocal current_line
         try:
             current_line.append_safe(leaf, preformatted=True)
-        except ValueError as ve:
+        except ValueError:
             yield current_line
 
             current_line = Line(depth=line.depth, inside_brackets=line.inside_brackets)
             current_line.append(leaf)
 
-    for index, leaf in enumerate(line.leaves):
+    for leaf in line.leaves:
         yield from append_to_line(leaf)
 
-        for comment_after in line.comments_after(leaf, index):
+        for comment_after in line.comments_after(leaf):
             yield from append_to_line(comment_after)
 
     if current_line:
@@ -2335,6 +2827,14 @@ def is_import(leaf: Leaf) -> bool:
     )
 
 
+def is_type_comment(leaf: Leaf, suffix: str = "") -> bool:
+    """Return True if the given leaf is a special comment.
+    Only returns true for type comments for now."""
+    t = leaf.type
+    v = leaf.value
+    return t in {token.COMMENT, STANDALONE_COMMENT} and v.startswith("# type:" + suffix)
+
+
 def normalize_prefix(leaf: Leaf, *, inside_brackets: bool) -> None:
     """Leave existing extra newlines if not `inside_brackets`. Remove everything
     else.
@@ -2363,7 +2863,7 @@ def normalize_string_prefix(leaf: Leaf, remove_u_prefix: bool = False) -> None:
     match = re.match(r"^([furbFURB]*)(.*)$", leaf.value, re.DOTALL)
     assert match is not None, f"failed to match string {leaf.value!r}"
     orig_prefix = match.group(1)
-    new_prefix = orig_prefix.lower()
+    new_prefix = orig_prefix.replace("F", "f").replace("B", "b").replace("U", "u")
     if remove_u_prefix:
         new_prefix = new_prefix.replace("u", "")
     leaf.value = f"{new_prefix}{match.group(2)}"
@@ -2396,8 +2896,8 @@ def normalize_string_quotes(leaf: Leaf) -> None:
 
     prefix = leaf.value[:first_quote_pos]
     unescaped_new_quote = re.compile(rf"(([^\\]|^)(\\\\)*){new_quote}")
-    escaped_new_quote = re.compile(rf"([^\\]|^)\\(\\\\)*{new_quote}")
-    escaped_orig_quote = re.compile(rf"([^\\]|^)\\(\\\\)*{orig_quote}")
+    escaped_new_quote = re.compile(rf"([^\\]|^)\\((?:\\\\)*){new_quote}")
+    escaped_orig_quote = re.compile(rf"([^\\]|^)\\((?:\\\\)*){orig_quote}")
     body = leaf.value[first_quote_pos + len(orig_quote) : -len(orig_quote)]
     if "r" in prefix.casefold():
         if unescaped_new_quote.search(body):
@@ -2408,15 +2908,30 @@ def normalize_string_quotes(leaf: Leaf) -> None:
         # Do not introduce or remove backslashes in raw strings
         new_body = body
     else:
-        # remove unnecessary quotes
+        # remove unnecessary escapes
         new_body = sub_twice(escaped_new_quote, rf"\1\2{new_quote}", body)
         if body != new_body:
-            # Consider the string without unnecessary quotes as the original
+            # Consider the string without unnecessary escapes as the original
             body = new_body
             leaf.value = f"{prefix}{orig_quote}{body}{orig_quote}"
         new_body = sub_twice(escaped_orig_quote, rf"\1\2{orig_quote}", new_body)
         new_body = sub_twice(unescaped_new_quote, rf"\1\\{new_quote}", new_body)
-    if new_quote == '"""' and new_body[-1] == '"':
+    if "f" in prefix.casefold():
+        matches = re.findall(
+            r"""
+            (?:[^{]|^)\{  # start of the string or a non-{ followed by a single {
+                ([^{].*?)  # contents of the brackets except if begins with {{
+            \}(?:[^}]|$)  # A } followed by end of the string or a non-}
+            """,
+            new_body,
+            re.VERBOSE,
+        )
+        for m in matches:
+            if "\\" in str(m):
+                # Do not introduce backslashes in interpolated expressions
+                return
+
+    if new_quote == '"""' and new_body[-1:] == '"':
         # edge case:
         new_body = new_body[:-1] + '\\"'
     orig_escape_count = body.count("\\")
@@ -2430,31 +2945,83 @@ def normalize_string_quotes(leaf: Leaf) -> None:
     leaf.value = f"{prefix}{new_quote}{new_body}{new_quote}"
 
 
+def normalize_numeric_literal(leaf: Leaf) -> None:
+    """Normalizes numeric (float, int, and complex) literals.
+
+    All letters used in the representation are normalized to lowercase (except
+    in Python 2 long literals).
+    """
+    text = leaf.value.lower()
+    if text.startswith(("0o", "0b")):
+        # Leave octal and binary literals alone.
+        pass
+    elif text.startswith("0x"):
+        # Change hex literals to upper case.
+        before, after = text[:2], text[2:]
+        text = f"{before}{after.upper()}"
+    elif "e" in text:
+        before, after = text.split("e")
+        sign = ""
+        if after.startswith("-"):
+            after = after[1:]
+            sign = "-"
+        elif after.startswith("+"):
+            after = after[1:]
+        before = format_float_or_int_string(before)
+        text = f"{before}e{sign}{after}"
+    elif text.endswith(("j", "l")):
+        number = text[:-1]
+        suffix = text[-1]
+        # Capitalize in "2L" because "l" looks too similar to "1".
+        if suffix == "l":
+            suffix = "L"
+        text = f"{format_float_or_int_string(number)}{suffix}"
+    else:
+        text = format_float_or_int_string(text)
+    leaf.value = text
+
+
+def format_float_or_int_string(text: str) -> str:
+    """Formats a float string like "1.0"."""
+    if "." not in text:
+        return text
+
+    before, after = text.split(".")
+    return f"{before or 0}.{after or 0}"
+
+
 def normalize_invisible_parens(node: Node, parens_after: Set[str]) -> None:
     """Make existing optional parentheses invisible or create new ones.
 
-    `parens_after` is a set of string leaf values immeditely after which parens
+    `parens_after` is a set of string leaf values immediately after which parens
     should be put.
 
     Standardizes on visible parentheses for single-element tuples, and keeps
     existing visible parentheses for other tuples and generator expressions.
     """
-    try:
-        list(generate_comments(node))
-    except FormatOff:
-        return  # This `node` has a prefix with `# fmt: off`, don't mess with parens.
-
+    for pc in list_comments(node.prefix, is_endmarker=False):
+        if pc.value in FMT_OFF:
+            # This `node` has a prefix with `# fmt: off`, don't mess with parens.
+            return
     check_lpar = False
     for index, child in enumerate(list(node.children)):
+        # Add parentheses around long tuple unpacking in assignments.
+        if (
+            index == 0
+            and isinstance(child, Node)
+            and child.type == syms.testlist_star_expr
+        ):
+            check_lpar = True
+
         if check_lpar:
+            if is_walrus_assignment(child):
+                continue
+
             if child.type == syms.atom:
-                maybe_make_parens_invisible_in_atom(child)
+                if maybe_make_parens_invisible_in_atom(child, parent=node):
+                    wrap_in_parentheses(node, child, visible=False)
             elif is_one_tuple(child):
-                # wrap child in visible parentheses
-                lpar = Leaf(token.LPAR, "(")
-                rpar = Leaf(token.RPAR, ")")
-                child.remove()
-                node.insert_child(index, Node(syms.atom, [lpar, child, rpar]))
+                wrap_in_parentheses(node, child, visible=True)
             elif node.type == syms.import_from:
                 # "import from" nodes store parentheses directly as part of
                 # the statement
@@ -2469,22 +3036,106 @@ def normalize_invisible_parens(node: Node, parens_after: Set[str]) -> None:
                 break
 
             elif not (isinstance(child, Leaf) and is_multiline_string(child)):
-                # wrap child in invisible parentheses
-                lpar = Leaf(token.LPAR, "")
-                rpar = Leaf(token.RPAR, "")
-                index = child.remove() or 0
-                node.insert_child(index, Node(syms.atom, [lpar, child, rpar]))
+                wrap_in_parentheses(node, child, visible=False)
 
         check_lpar = isinstance(child, Leaf) and child.value in parens_after
 
 
-def maybe_make_parens_invisible_in_atom(node: LN) -> bool:
-    """If it's safe, make the parens in the atom `node` invisible, recusively."""
+def normalize_fmt_off(node: Node) -> None:
+    """Convert content between `# fmt: off`/`# fmt: on` into standalone comments."""
+    try_again = True
+    while try_again:
+        try_again = convert_one_fmt_off_pair(node)
+
+
+def convert_one_fmt_off_pair(node: Node) -> bool:
+    """Convert content of a single `# fmt: off`/`# fmt: on` into a standalone comment.
+
+    Returns True if a pair was converted.
+    """
+    for leaf in node.leaves():
+        previous_consumed = 0
+        for comment in list_comments(leaf.prefix, is_endmarker=False):
+            if comment.value in FMT_OFF:
+                # We only want standalone comments. If there's no previous leaf or
+                # the previous leaf is indentation, it's a standalone comment in
+                # disguise.
+                if comment.type != STANDALONE_COMMENT:
+                    prev = preceding_leaf(leaf)
+                    if prev and prev.type not in WHITESPACE:
+                        continue
+
+                ignored_nodes = list(generate_ignored_nodes(leaf))
+                if not ignored_nodes:
+                    continue
+
+                first = ignored_nodes[0]  # Can be a container node with the `leaf`.
+                parent = first.parent
+                prefix = first.prefix
+                first.prefix = prefix[comment.consumed :]
+                hidden_value = (
+                    comment.value + "\n" + "".join(str(n) for n in ignored_nodes)
+                )
+                if hidden_value.endswith("\n"):
+                    # That happens when one of the `ignored_nodes` ended with a NEWLINE
+                    # leaf (possibly followed by a DEDENT).
+                    hidden_value = hidden_value[:-1]
+                first_idx: Optional[int] = None
+                for ignored in ignored_nodes:
+                    index = ignored.remove()
+                    if first_idx is None:
+                        first_idx = index
+                assert parent is not None, "INTERNAL ERROR: fmt: on/off handling (1)"
+                assert first_idx is not None, "INTERNAL ERROR: fmt: on/off handling (2)"
+                parent.insert_child(
+                    first_idx,
+                    Leaf(
+                        STANDALONE_COMMENT,
+                        hidden_value,
+                        prefix=prefix[:previous_consumed] + "\n" * comment.newlines,
+                    ),
+                )
+                return True
+
+            previous_consumed = comment.consumed
+
+    return False
+
+
+def generate_ignored_nodes(leaf: Leaf) -> Iterator[LN]:
+    """Starting from the container of `leaf`, generate all leaves until `# fmt: on`.
+
+    Stops at the end of the block.
+    """
+    container: Optional[LN] = container_of(leaf)
+    while container is not None and container.type != token.ENDMARKER:
+        is_fmt_on = False
+        for comment in list_comments(container.prefix, is_endmarker=False):
+            if comment.value in FMT_ON:
+                is_fmt_on = True
+            elif comment.value in FMT_OFF:
+                is_fmt_on = False
+        if is_fmt_on:
+            return
+
+        yield container
+
+        container = container.next_sibling
+
+
+def maybe_make_parens_invisible_in_atom(node: LN, parent: LN) -> bool:
+    """If it's safe, make the parens in the atom `node` invisible, recursively.
+    Additionally, remove repeated, adjacent invisible parens from the atom `node`
+    as they are redundant.
+
+    Returns whether the node should itself be wrapped in invisible parentheses.
+
+    """
     if (
         node.type != syms.atom
         or is_empty_tuple(node)
         or is_one_tuple(node)
-        or is_yield(node)
+        or (is_yield(node) and parent.type != syms.expr_stmt)
         or max_delimiter_priority_in_atom(node) >= COMMA_PRIORITY
     ):
         return False
@@ -2492,14 +3143,38 @@ def maybe_make_parens_invisible_in_atom(node: LN) -> bool:
     first = node.children[0]
     last = node.children[-1]
     if first.type == token.LPAR and last.type == token.RPAR:
+        middle = node.children[1]
         # make parentheses invisible
         first.value = ""  # type: ignore
         last.value = ""  # type: ignore
-        if len(node.children) > 1:
-            maybe_make_parens_invisible_in_atom(node.children[1])
-        return True
+        maybe_make_parens_invisible_in_atom(middle, parent=parent)
 
-    return False
+        if is_atom_with_invisible_parens(middle):
+            # Strip the invisible parens from `middle` by replacing
+            # it with the child in-between the invisible parens
+            middle.replace(middle.children[1])
+
+        return False
+
+    return True
+
+
+def is_atom_with_invisible_parens(node: LN) -> bool:
+    """Given a `LN`, determines whether it's an atom `node` with invisible
+    parens. Useful in dedupe-ing and normalizing parens.
+    """
+    if isinstance(node, Leaf) or node.type != syms.atom:
+        return False
+
+    first, last = node.children[0], node.children[-1]
+    return (
+        isinstance(first, Leaf)
+        and first.type == token.LPAR
+        and first.value == ""
+        and isinstance(last, Leaf)
+        and last.type == token.RPAR
+        and last.value == ""
+    )
 
 
 def is_empty_tuple(node: LN) -> bool:
@@ -2512,18 +3187,43 @@ def is_empty_tuple(node: LN) -> bool:
     )
 
 
+def unwrap_singleton_parenthesis(node: LN) -> Optional[LN]:
+    """Returns `wrapped` if `node` is of the shape ( wrapped ).
+
+    Parenthesis can be optional. Returns None otherwise"""
+    if len(node.children) != 3:
+        return None
+
+    lpar, wrapped, rpar = node.children
+    if not (lpar.type == token.LPAR and rpar.type == token.RPAR):
+        return None
+
+    return wrapped
+
+
+def wrap_in_parentheses(parent: Node, child: LN, *, visible: bool = True) -> None:
+    """Wrap `child` in parentheses.
+
+    This replaces `child` with an atom holding the parentheses and the old
+    child.  That requires moving the prefix.
+
+    If `visible` is False, the leaves will be valueless (and thus invisible).
+    """
+    lpar = Leaf(token.LPAR, "(" if visible else "")
+    rpar = Leaf(token.RPAR, ")" if visible else "")
+    prefix = child.prefix
+    child.prefix = ""
+    index = child.remove() or 0
+    new_child = Node(syms.atom, [lpar, child, rpar])
+    new_child.prefix = prefix
+    parent.insert_child(index, new_child)
+
+
 def is_one_tuple(node: LN) -> bool:
     """Return True if `node` holds a tuple with one element, with or without parens."""
     if node.type == syms.atom:
-        if len(node.children) != 3:
-            return False
-
-        lpar, gexp, rpar = node.children
-        if not (
-            lpar.type == token.LPAR
-            and gexp.type == syms.testlist_gexp
-            and rpar.type == token.RPAR
-        ):
+        gexp = unwrap_singleton_parenthesis(node)
+        if gexp is None or gexp.type != syms.testlist_gexp:
             return False
 
         return len(gexp.children) == 2 and gexp.children[1].type == token.COMMA
@@ -2535,6 +3235,12 @@ def is_one_tuple(node: LN) -> bool:
     )
 
 
+def is_walrus_assignment(node: LN) -> bool:
+    """Return True iff `node` is of the shape ( test := test )"""
+    inner = unwrap_singleton_parenthesis(node)
+    return inner is not None and inner.type == syms.namedexpr_test
+
+
 def is_yield(node: LN) -> bool:
     """Return True if `node` holds a `yield` or `yield from` expression."""
     if node.type == syms.yield_expr:
@@ -2564,7 +3270,7 @@ def is_vararg(leaf: Leaf, within: Set[NodeType]) -> bool:
     extended iterable unpacking (PEP 3132) and additional unpacking
     generalizations (PEP 448).
     """
-    if leaf.type not in STARS or not leaf.parent:
+    if leaf.type not in VARARGS_SPECIALS or not leaf.parent:
         return False
 
     p = leaf.parent
@@ -2614,7 +3320,7 @@ def is_stub_body(node: LN) -> bool:
     )
 
 
-def max_delimiter_priority_in_atom(node: LN) -> int:
+def max_delimiter_priority_in_atom(node: LN) -> Priority:
     """Return maximum delimiter priority inside `node`.
 
     This is specific to atoms with contents contained in a pair of parentheses.
@@ -2646,7 +3352,7 @@ def ensure_visible(leaf: Leaf) -> None:
     """Make sure parentheses are visible.
 
     They could be invisible as part of some statements (see
-    :func:`normalize_invible_parens` and :func:`visit_import_from`).
+    :func:`normalize_invisible_parens` and :func:`visit_import_from`).
     """
     if leaf.type == token.LPAR:
         leaf.value = "("
@@ -2656,6 +3362,7 @@ def ensure_visible(leaf: Leaf) -> None:
 
 def should_explode(line: Line, opening_bracket: Leaf) -> bool:
     """Should `line` immediately be split with `delimiter_split()` after RHS?"""
+
     if not (
         opening_bracket.parent
         and opening_bracket.parent.type in {syms.atom, syms.import_from}
@@ -2673,34 +3380,61 @@ def should_explode(line: Line, opening_bracket: Leaf) -> bool:
     return max_priority == COMMA_PRIORITY
 
 
-def is_python36(node: Node) -> bool:
-    """Return True if the current file is using Python 3.6+ features.
+def get_features_used(node: Node) -> Set[Feature]:
+    """Return a set of (relatively) new Python features used in this file.
 
     Currently looking for:
-    - f-strings; and
-    - trailing commas after * or ** in function signatures and calls.
+    - f-strings;
+    - underscores in numeric literals;
+    - trailing commas after * or ** in function signatures and calls;
+    - positional only arguments in function signatures and lambdas;
     """
+    features: Set[Feature] = set()
     for n in node.pre_order():
         if n.type == token.STRING:
             value_head = n.value[:2]  # type: ignore
             if value_head in {'f"', 'F"', "f'", "F'", "rf", "fr", "RF", "FR"}:
-                return True
+                features.add(Feature.F_STRINGS)
+
+        elif n.type == token.NUMBER:
+            if "_" in n.value:  # type: ignore
+                features.add(Feature.NUMERIC_UNDERSCORES)
+
+        elif n.type == token.SLASH:
+            if n.parent and n.parent.type in {syms.typedargslist, syms.arglist}:
+                features.add(Feature.POS_ONLY_ARGUMENTS)
+
+        elif n.type == token.COLONEQUAL:
+            features.add(Feature.ASSIGNMENT_EXPRESSIONS)
 
         elif (
             n.type in {syms.typedargslist, syms.arglist}
             and n.children
             and n.children[-1].type == token.COMMA
         ):
+            if n.type == syms.typedargslist:
+                feature = Feature.TRAILING_COMMA_IN_DEF
+            else:
+                feature = Feature.TRAILING_COMMA_IN_CALL
+
             for ch in n.children:
                 if ch.type in STARS:
-                    return True
+                    features.add(feature)
 
                 if ch.type == syms.argument:
                     for argch in ch.children:
                         if argch.type in STARS:
-                            return True
+                            features.add(feature)
 
-    return False
+    return features
+
+
+def detect_target_versions(node: Node) -> Set[TargetVersion]:
+    """Detect the version to target based on the nodes used."""
+    features = get_features_used(node)
+    return {
+        version for version in TargetVersion if features <= VERSION_TO_FEATURES[version]
+    }
 
 
 def generate_trailers_to_omit(line: Line, line_length: int) -> Iterator[Set[LeafID]]:
@@ -2717,9 +3451,8 @@ def generate_trailers_to_omit(line: Line, line_length: int) -> Iterator[Set[Leaf
     yield omit
 
     length = 4 * line.depth
-    opening_bracket = None
-    closing_bracket = None
-    optional_brackets: Set[LeafID] = set()
+    opening_bracket: Optional[Leaf] = None
+    closing_bracket: Optional[Leaf] = None
     inner_brackets: Set[LeafID] = set()
     for index, leaf, leaf_length in enumerate_with_length(line, reversed=True):
         length += leaf_length
@@ -2730,17 +3463,12 @@ def generate_trailers_to_omit(line: Line, line_length: int) -> Iterator[Set[Leaf
         if leaf.type == STANDALONE_COMMENT or has_inline_comment:
             break
 
-        optional_brackets.discard(id(leaf))
         if opening_bracket:
             if leaf is opening_bracket:
                 opening_bracket = None
             elif leaf.type in CLOSING_BRACKETS:
                 inner_brackets.add(id(leaf))
         elif leaf.type in CLOSING_BRACKETS:
-            if not leaf.value:
-                optional_brackets.add(id(opening_bracket))
-                continue
-
             if index > 0 and line.leaves[index - 1].type in OPENING_BRACKETS:
                 # Empty brackets would fail a split so treat them as "inner"
                 # brackets (e.g. only add them to the `omit` set if another
@@ -2748,21 +3476,43 @@ def generate_trailers_to_omit(line: Line, line_length: int) -> Iterator[Set[Leaf
                 inner_brackets.add(id(leaf))
                 continue
 
-            opening_bracket = leaf.opening_bracket
             if closing_bracket:
                 omit.add(id(closing_bracket))
                 omit.update(inner_brackets)
                 inner_brackets.clear()
                 yield omit
-            closing_bracket = leaf
+
+            if leaf.value:
+                opening_bracket = leaf.opening_bracket
+                closing_bracket = leaf
 
 
 def get_future_imports(node: Node) -> Set[str]:
     """Return a set of __future__ imports in the file."""
-    imports = set()
+    imports: Set[str] = set()
+
+    def get_imports_from_children(children: List[LN]) -> Generator[str, None, None]:
+        for child in children:
+            if isinstance(child, Leaf):
+                if child.type == token.NAME:
+                    yield child.value
+
+            elif child.type == syms.import_as_name:
+                orig_name = child.children[0]
+                assert isinstance(orig_name, Leaf), "Invalid syntax parsing imports"
+                assert orig_name.type == token.NAME, "Invalid syntax parsing imports"
+                yield orig_name.value
+
+            elif child.type == syms.import_as_names:
+                yield from get_imports_from_children(child.children)
+
+            else:
+                raise AssertionError("Invalid syntax parsing imports")
+
     for child in node.children:
         if child.type != syms.simple_stmt:
             break
+
         first_child = child.children[0]
         if isinstance(first_child, Leaf):
             # Continue looking if we see a docstring; otherwise stop.
@@ -2772,43 +3522,82 @@ def get_future_imports(node: Node) -> Set[str]:
                 and child.children[1].type == token.NEWLINE
             ):
                 continue
-            else:
-                break
+
+            break
+
         elif first_child.type == syms.import_from:
             module_name = first_child.children[1]
             if not isinstance(module_name, Leaf) or module_name.value != "__future__":
                 break
-            for import_from_child in first_child.children[3:]:
-                if isinstance(import_from_child, Leaf):
-                    if import_from_child.type == token.NAME:
-                        imports.add(import_from_child.value)
-                else:
-                    assert import_from_child.type == syms.import_as_names
-                    for leaf in import_from_child.children:
-                        if isinstance(leaf, Leaf) and leaf.type == token.NAME:
-                            imports.add(leaf.value)
+
+            imports |= set(get_imports_from_children(first_child.children[3:]))
         else:
             break
+
     return imports
 
 
+@lru_cache()
+def get_gitignore(root: Path) -> PathSpec:
+    """ Return a PathSpec matching gitignore content if present."""
+    gitignore = root / ".gitignore"
+    lines: List[str] = []
+    if gitignore.is_file():
+        with gitignore.open() as gf:
+            lines = gf.readlines()
+    return PathSpec.from_lines("gitwildmatch", lines)
+
+
 def gen_python_files_in_dir(
-    path: Path, root: Path, include: Pattern[str], exclude: Pattern[str]
+    path: Path,
+    root: Path,
+    include: Pattern[str],
+    exclude: Pattern[str],
+    report: "Report",
+    gitignore: PathSpec,
 ) -> Iterator[Path]:
     """Generate all files under `path` whose paths are not excluded by the
     `exclude` regex, but are included by the `include` regex.
+
+    Symbolic links pointing outside of the `root` directory are ignored.
+
+    `report` is where output about exclusions goes.
     """
     assert root.is_absolute(), f"INTERNAL ERROR: `root` must be absolute but is {root}"
     for child in path.iterdir():
-        normalized_path = child.resolve().relative_to(root).as_posix()
+        # First ignore files matching .gitignore
+        if gitignore.match_file(child.as_posix()):
+            report.path_ignored(child, "matches the .gitignore file content")
+            continue
+
+        # Then ignore with `exclude` option.
+        try:
+            normalized_path = "/" + child.resolve().relative_to(root).as_posix()
+        except OSError as e:
+            report.path_ignored(child, f"cannot be read because {e}")
+            continue
+
+        except ValueError:
+            if child.is_symlink():
+                report.path_ignored(
+                    child, f"is a symbolic link that points outside {root}"
+                )
+                continue
+
+            raise
+
         if child.is_dir():
             normalized_path += "/"
+
         exclude_match = exclude.search(normalized_path)
         if exclude_match and exclude_match.group(0):
+            report.path_ignored(child, "matches the --exclude regular expression")
             continue
 
         if child.is_dir():
-            yield from gen_python_files_in_dir(child, root, include, exclude)
+            yield from gen_python_files_in_dir(
+                child, root, include, exclude, report, gitignore
+            )
 
         elif child.is_file():
             include_match = include.search(normalized_path)
@@ -2816,7 +3605,8 @@ def gen_python_files_in_dir(
                 yield child
 
 
-def find_project_root(srcs: List[str]) -> Path:
+@lru_cache()
+def find_project_root(srcs: Iterable[str]) -> Path:
     """Return a directory containing .git, .hg, or pyproject.toml.
 
     That directory can be one of the directories passed in `srcs` or their
@@ -2833,7 +3623,7 @@ def find_project_root(srcs: List[str]) -> Path:
         # Append a fake file so `parents` below returns `common_base_dir`, too.
         common_base /= "fake-file"
     for directory in common_base.parents:
-        if (directory / ".git").is_dir():
+        if (directory / ".git").exists():
             return directory
 
         if (directory / ".hg").is_dir():
@@ -2850,7 +3640,9 @@ class Report:
     """Provides a reformatting counter. Can be rendered with `str(report)`."""
 
     check: bool = False
+    diff: bool = False
     quiet: bool = False
+    verbose: bool = False
     change_count: int = 0
     same_count: int = 0
     failure_count: int = 0
@@ -2858,12 +3650,12 @@ class Report:
     def done(self, src: Path, changed: Changed) -> None:
         """Increment the counter for successful reformatting. Write out a message."""
         if changed is Changed.YES:
-            reformatted = "would reformat" if self.check else "reformatted"
-            if not self.quiet:
+            reformatted = "would reformat" if self.check or self.diff else "reformatted"
+            if self.verbose or not self.quiet:
                 out(f"{reformatted} {src}")
             self.change_count += 1
         else:
-            if not self.quiet:
+            if self.verbose:
                 if changed is Changed.NO:
                     msg = f"{src} already well formatted, good job."
                 else:
@@ -2876,6 +3668,10 @@ class Report:
         err(f"error: cannot format {src}: {message}")
         self.failure_count += 1
 
+    def path_ignored(self, path: Path, message: str) -> None:
+        if self.verbose:
+            out(f"{path} ignored: {message}", bold=False)
+
     @property
     def return_code(self) -> int:
         """Return the exit code that the app should use.
@@ -2886,7 +3682,7 @@ class Report:
         - otherwise return 0.
         """
         # According to http://tldp.org/LDP/abs/html/exitcodes.html starting with
-        # 126 we have special returncodes reserved by the shell.
+        # 126 we have special return codes reserved by the shell.
         if self.failure_count:
             return 123
 
@@ -2900,7 +3696,7 @@ class Report:
 
         Use `click.unstyle` to remove colors.
         """
-        if self.check:
+        if self.check or self.diff:
             reformatted = "would be reformatted"
             unchanged = "would be left unchanged"
             failed = "would fail to reformat"
@@ -2925,17 +3721,59 @@ class Report:
         return ", ".join(report) + "."
 
 
+def parse_ast(src: str) -> Union[ast.AST, ast3.AST, ast27.AST]:
+    filename = "<unknown>"
+    if sys.version_info >= (3, 8):
+        # TODO: support Python 4+ ;)
+        for minor_version in range(sys.version_info[1], 4, -1):
+            try:
+                return ast.parse(src, filename, feature_version=(3, minor_version))
+            except SyntaxError:
+                continue
+    else:
+        for feature_version in (7, 6):
+            try:
+                return ast3.parse(src, filename, feature_version=feature_version)
+            except SyntaxError:
+                continue
+
+    return ast27.parse(src)
+
+
+def _fixup_ast_constants(
+    node: Union[ast.AST, ast3.AST, ast27.AST]
+) -> Union[ast.AST, ast3.AST, ast27.AST]:
+    """Map ast nodes deprecated in 3.8 to Constant."""
+    if isinstance(node, (ast.Str, ast3.Str, ast27.Str, ast.Bytes, ast3.Bytes)):
+        return ast.Constant(value=node.s)
+
+    if isinstance(node, (ast.Num, ast3.Num, ast27.Num)):
+        return ast.Constant(value=node.n)
+
+    if isinstance(node, (ast.NameConstant, ast3.NameConstant)):
+        return ast.Constant(value=node.value)
+
+    return node
+
+
 def assert_equivalent(src: str, dst: str) -> None:
     """Raise AssertionError if `src` and `dst` aren't equivalent."""
 
-    import ast
-    import traceback
-
-    def _v(node: ast.AST, depth: int = 0) -> Iterator[str]:
+    def _v(node: Union[ast.AST, ast3.AST, ast27.AST], depth: int = 0) -> Iterator[str]:
         """Simple visitor generating strings to compare ASTs by content."""
+
+        node = _fixup_ast_constants(node)
+
         yield f"{'  ' * depth}{node.__class__.__name__}("
 
-        for field in sorted(node._fields):
+        for field in sorted(node._fields):  # noqa: F402
+            # TypeIgnore has only one field 'lineno' which breaks this comparison
+            type_ignore_classes = (ast3.TypeIgnore, ast27.TypeIgnore)
+            if sys.version_info >= (3, 8):
+                type_ignore_classes += (ast.TypeIgnore,)
+            if isinstance(node, type_ignore_classes):
+                break
+
             try:
                 value = getattr(node, field)
             except AttributeError:
@@ -2945,10 +3783,20 @@ def assert_equivalent(src: str, dst: str) -> None:
 
             if isinstance(value, list):
                 for item in value:
-                    if isinstance(item, ast.AST):
+                    # Ignore nested tuples within del statements, because we may insert
+                    # parentheses and they change the AST.
+                    if (
+                        field == "targets"
+                        and isinstance(node, (ast.Delete, ast3.Delete, ast27.Delete))
+                        and isinstance(item, (ast.Tuple, ast3.Tuple, ast27.Tuple))
+                    ):
+                        for item in item.elts:
+                            yield from _v(item, depth + 2)
+
+                    elif isinstance(item, (ast.AST, ast3.AST, ast27.AST)):
                         yield from _v(item, depth + 2)
 
-            elif isinstance(value, ast.AST):
+            elif isinstance(value, (ast.AST, ast3.AST, ast27.AST)):
                 yield from _v(value, depth + 2)
 
             else:
@@ -2957,22 +3805,20 @@ def assert_equivalent(src: str, dst: str) -> None:
         yield f"{'  ' * depth})  # /{node.__class__.__name__}"
 
     try:
-        src_ast = ast.parse(src)
+        src_ast = parse_ast(src)
     except Exception as exc:
-        major, minor = sys.version_info[:2]
         raise AssertionError(
-            f"cannot use --safe with this file; failed to parse source file "
-            f"with Python {major}.{minor}'s builtin AST. Re-run with --fast "
-            f"or stop using deprecated Python 2 syntax. AST error message: {exc}"
+            f"cannot use --safe with this file; failed to parse source file.  "
+            f"AST error message: {exc}"
         )
 
     try:
-        dst_ast = ast.parse(dst)
+        dst_ast = parse_ast(dst)
     except Exception as exc:
         log = dump_to_file("".join(traceback.format_tb(exc.__traceback__)), dst)
         raise AssertionError(
             f"INTERNAL ERROR: Black produced invalid code: {exc}. "
-            f"Please report a bug on https://github.com/ambv/black/issues.  "
+            f"Please report a bug on https://github.com/psf/black/issues.  "
             f"This invalid output might be helpful: {log}"
         ) from None
 
@@ -2983,16 +3829,14 @@ def assert_equivalent(src: str, dst: str) -> None:
         raise AssertionError(
             f"INTERNAL ERROR: Black produced code that is not equivalent to "
             f"the source.  "
-            f"Please report a bug on https://github.com/ambv/black/issues.  "
+            f"Please report a bug on https://github.com/psf/black/issues.  "
             f"This diff might be helpful: {log}"
         ) from None
 
 
-def assert_stable(
-    src: str, dst: str, line_length: int, mode: FileMode = FileMode.AUTO_DETECT
-) -> None:
+def assert_stable(src: str, dst: str, mode: Mode) -> None:
     """Raise AssertionError if `dst` reformats differently the second time."""
-    newdst = format_str(dst, line_length=line_length, mode=mode)
+    newdst = format_str(dst, mode=mode)
     if dst != newdst:
         log = dump_to_file(
             diff(src, dst, "source", "first pass"),
@@ -3001,15 +3845,14 @@ def assert_stable(
         raise AssertionError(
             f"INTERNAL ERROR: Black produced different code on the second pass "
             f"of the formatter.  "
-            f"Please report a bug on https://github.com/ambv/black/issues.  "
+            f"Please report a bug on https://github.com/psf/black/issues.  "
             f"This diff might be helpful: {log}"
         ) from None
 
 
+@mypyc_attr(patchable=True)
 def dump_to_file(*output: str) -> str:
     """Dump `output` to a temporary file. Return path to the file."""
-    import tempfile
-
     with tempfile.NamedTemporaryFile(
         mode="w", prefix="blk_", suffix=".log", delete=False, encoding="utf8"
     ) as f:
@@ -3020,29 +3863,42 @@ def dump_to_file(*output: str) -> str:
     return f.name
 
 
+@contextmanager
+def nullcontext() -> Iterator[None]:
+    """Return an empty context manager.
+
+    To be used like `nullcontext` in Python 3.7.
+    """
+    yield
+
+
 def diff(a: str, b: str, a_name: str, b_name: str) -> str:
     """Return a unified diff string between strings `a` and `b`."""
     import difflib
 
-    a_lines = [line + "\n" for line in a.split("\n")]
-    b_lines = [line + "\n" for line in b.split("\n")]
+    a_lines = [line + "\n" for line in a.splitlines()]
+    b_lines = [line + "\n" for line in b.splitlines()]
     return "".join(
         difflib.unified_diff(a_lines, b_lines, fromfile=a_name, tofile=b_name, n=5)
     )
 
 
-def cancel(tasks: Iterable[asyncio.Task]) -> None:
+def cancel(tasks: Iterable["asyncio.Task[Any]"]) -> None:
     """asyncio signal handler that cancels all `tasks` and reports to stderr."""
     err("Aborted!")
     for task in tasks:
         task.cancel()
 
 
-def shutdown(loop: BaseEventLoop) -> None:
+def shutdown(loop: asyncio.AbstractEventLoop) -> None:
     """Cancel all pending tasks on `loop`, wait for them, and close the loop."""
     try:
+        if sys.version_info[:2] >= (3, 7):
+            all_tasks = asyncio.all_tasks
+        else:
+            all_tasks = asyncio.Task.all_tasks
         # This part is borrowed from asyncio/runners.py in Python 3.7b2.
-        to_cancel = [task for task in asyncio.Task.all_tasks(loop) if not task.done()]
+        to_cancel = [task for task in all_tasks(loop) if not task.done()]
         if not to_cancel:
             return
 
@@ -3069,6 +3925,17 @@ def sub_twice(regex: Pattern[str], replacement: str, original: str) -> str:
     return regex.sub(replacement, regex.sub(replacement, original))
 
 
+def re_compile_maybe_verbose(regex: str) -> Pattern[str]:
+    """Compile a regular expression string in `regex`.
+
+    If it contains newlines, use verbose mode.
+    """
+    if "\n" in regex:
+        regex = "(?x)" + regex
+    compiled: Pattern[str] = re.compile(regex)
+    return compiled
+
+
 def enumerate_reversed(sequence: Sequence[T]) -> Iterator[Tuple[Index, T]]:
     """Like `reversed(enumerate(sequence))` if that were possible."""
     index = len(sequence) - 1
@@ -3093,8 +3960,7 @@ def enumerate_with_length(
         if "\n" in leaf.value:
             return  # Multiline strings, we can't continue.
 
-        comment: Optional[Leaf]
-        for comment in line.comments_after(leaf, index):
+        for comment in line.comments_after(leaf):
             length += len(comment.value)
 
         yield index, leaf, length
@@ -3114,6 +3980,42 @@ def is_line_short_enough(line: Line, *, line_length: int, line_str: str = "") ->
     )
 
 
+def can_be_split(line: Line) -> bool:
+    """Return False if the line cannot be split *for sure*.
+
+    This is not an exhaustive search but a cheap heuristic that we can use to
+    avoid some unfortunate formattings (mostly around wrapping unsplittable code
+    in unnecessary parentheses).
+    """
+    leaves = line.leaves
+    if len(leaves) < 2:
+        return False
+
+    if leaves[0].type == token.STRING and leaves[1].type == token.DOT:
+        call_count = 0
+        dot_count = 0
+        next = leaves[-1]
+        for leaf in leaves[-2::-1]:
+            if leaf.type in OPENING_BRACKETS:
+                if next.type not in CLOSING_BRACKETS:
+                    return False
+
+                call_count += 1
+            elif leaf.type == token.DOT:
+                dot_count += 1
+            elif leaf.type == token.NAME:
+                if not (next.type == token.DOT or next.type in OPENING_BRACKETS):
+                    return False
+
+            elif leaf.type not in CLOSING_BRACKETS:
+                return False
+
+            if dot_count > 1 and call_count > 1:
+                return False
+
+    return True
+
+
 def can_omit_invisible_parens(line: Line, line_length: int) -> bool:
     """Does `line` have a shape safe to reformat without optional parens around it?
 
@@ -3203,28 +4105,23 @@ def can_omit_invisible_parens(line: Line, line_length: int) -> bool:
     return False
 
 
-def get_cache_file(line_length: int, mode: FileMode) -> Path:
-    pyi = bool(mode & FileMode.PYI)
-    py36 = bool(mode & FileMode.PYTHON36)
-    return (
-        CACHE_DIR
-        / f"cache.{line_length}{'.pyi' if pyi else ''}{'.py36' if py36 else ''}.pickle"
-    )
+def get_cache_file(mode: Mode) -> Path:
+    return CACHE_DIR / f"cache.{mode.get_cache_key()}.pickle"
 
 
-def read_cache(line_length: int, mode: FileMode) -> Cache:
+def read_cache(mode: Mode) -> Cache:
     """Read the cache if it exists and is well formed.
 
     If it is not well formed, the call to write_cache later should resolve the issue.
     """
-    cache_file = get_cache_file(line_length, mode)
+    cache_file = get_cache_file(mode)
     if not cache_file.exists():
         return {}
 
     with cache_file.open("rb") as fobj:
         try:
             cache: Cache = pickle.load(fobj)
-        except pickle.UnpicklingError:
+        except (pickle.UnpicklingError, ValueError):
             return {}
 
     return cache
@@ -3236,38 +4133,62 @@ def get_cache_info(path: Path) -> CacheInfo:
     return stat.st_mtime, stat.st_size
 
 
-def filter_cached(
-    cache: Cache, sources: Iterable[Path]
-) -> Tuple[List[Path], List[Path]]:
-    """Split a list of paths into two.
+def filter_cached(cache: Cache, sources: Iterable[Path]) -> Tuple[Set[Path], Set[Path]]:
+    """Split an iterable of paths in `sources` into two sets.
 
-    The first list contains paths of files that modified on disk or are not in the
-    cache. The other list contains paths to non-modified files.
+    The first contains paths of files that modified on disk or are not in the
+    cache. The other contains paths to non-modified files.
     """
-    todo, done = [], []
+    todo, done = set(), set()
     for src in sources:
         src = src.resolve()
         if cache.get(src) != get_cache_info(src):
-            todo.append(src)
+            todo.add(src)
         else:
-            done.append(src)
+            done.add(src)
     return todo, done
 
 
-def write_cache(
-    cache: Cache, sources: List[Path], line_length: int, mode: FileMode
-) -> None:
+def write_cache(cache: Cache, sources: Iterable[Path], mode: Mode) -> None:
     """Update the cache file."""
-    cache_file = get_cache_file(line_length, mode)
+    cache_file = get_cache_file(mode)
     try:
-        if not CACHE_DIR.exists():
-            CACHE_DIR.mkdir(parents=True)
+        CACHE_DIR.mkdir(parents=True, exist_ok=True)
         new_cache = {**cache, **{src.resolve(): get_cache_info(src) for src in sources}}
-        with cache_file.open("wb") as fobj:
-            pickle.dump(new_cache, fobj, protocol=pickle.HIGHEST_PROTOCOL)
+        with tempfile.NamedTemporaryFile(dir=str(cache_file.parent), delete=False) as f:
+            pickle.dump(new_cache, f, protocol=4)
+        os.replace(f.name, cache_file)
     except OSError:
         pass
 
 
-if __name__ == "__main__":
+def patch_click() -> None:
+    """Make Click not crash.
+
+    On certain misconfigured environments, Python 3 selects the ASCII encoding as the
+    default which restricts paths that it can access during the lifetime of the
+    application.  Click refuses to work in this scenario by raising a RuntimeError.
+
+    In case of Black the likelihood that non-ASCII characters are going to be used in
+    file paths is minimal since it's Python source code.  Moreover, this crash was
+    spurious on Python 3.7 thanks to PEP 538 and PEP 540.
+    """
+    try:
+        from click import core
+        from click import _unicodefun  # type: ignore
+    except ModuleNotFoundError:
+        return
+
+    for module in (core, _unicodefun):
+        if hasattr(module, "_verify_python3_env"):
+            module._verify_python3_env = lambda: None
+
+
+def patched_main() -> None:
+    freeze_support()
+    patch_click()
     main()
+
+
+if __name__ == "__main__":
+    patched_main()