X-Git-Url: https://git.madduck.net/etc/vim.git/blobdiff_plain/9c9f6eb6d53c4a38c46acfcffa1c90a787a72a77..bbe5ae70c17ff16801fef240f49b707535c7f64c:/black.py?ds=inline

diff --git a/black.py b/black.py
index 27691dd..69d24c5 100644
--- a/black.py
+++ b/black.py
@@ -1,237 +1,863 @@
-#!/usr/bin/env python3
+import ast
 import asyncio
-from asyncio.base_events import BaseEventLoop
 from concurrent.futures import Executor, ProcessPoolExecutor
-from functools import partial
-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, freeze_support
 import os
 from pathlib import Path
-import tokenize
+import pickle
+import regex as re
+import signal
 import sys
+import tempfile
+import tokenize
+import traceback
 from typing import (
-    Dict, Generic, Iterable, Iterator, List, Optional, Set, Tuple, TypeVar, Union
+    Any,
+    Callable,
+    Collection,
+    Dict,
+    Generator,
+    Generic,
+    Iterable,
+    Iterator,
+    List,
+    Optional,
+    Pattern,
+    Sequence,
+    Set,
+    Tuple,
+    TypeVar,
+    Union,
+    cast,
 )
+from typing_extensions import Final
+from mypy_extensions import mypyc_attr
 
-from attr import attrib, dataclass, Factory
+from appdirs import user_cache_dir
+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
 
-__version__ = "18.3a2"
+from _black_version import version as __version__
+
 DEFAULT_LINE_LENGTH = 88
+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__))
+
+
 # types
-syms = pygram.python_symbols
 FileContent = str
 Encoding = str
+NewLine = str
 Depth = int
 NodeType = int
 LeafID = int
 Priority = int
+Index = int
 LN = Union[Leaf, Node]
+SplitFunc = Callable[["Line", Collection["Feature"]], Iterator["Line"]]
+Timestamp = float
+FileSize = int
+CacheInfo = Tuple[Timestamp, FileSize]
+Cache = Dict[Path, CacheInfo]
 out = partial(click.secho, bold=True, err=True)
-err = partial(click.secho, fg='red', err=True)
+err = partial(click.secho, fg="red", err=True)
+
+pygram.initialize(CACHE_DIR)
+syms = pygram.python_symbols
 
 
 class NothingChanged(UserWarning):
-    """Raised by `format_file` when the 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."""
+
+
+class InvalidInput(ValueError):
+    """Raised when input source code fails all parse attempts."""
+
+
+class WriteBack(Enum):
+    NO = 0
+    YES = 1
+    DIFF = 2
+    CHECK = 3
+
+    @classmethod
+    def from_configuration(cls, *, check: bool, diff: bool) -> "WriteBack":
+        if check and not diff:
+            return cls.CHECK
+
+        return cls.DIFF if diff else cls.YES
+
+
+class Changed(Enum):
+    NO = 0
+    CACHED = 1
+    YES = 2
+
+
+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,
+    },
+}
+
+
+@dataclass
+class FileMode:
+    target_versions: Set[TargetVersion] = field(default_factory=set)
+    line_length: int = DEFAULT_LINE_LENGTH
+    string_normalization: bool = True
+    is_pyi: bool = False
+
+    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)
+
+
+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
+
+    if ctx.default_map is None:
+        ctx.default_map = {}
+    ctx.default_map.update(config)  # type: ignore  # bad types in .pyi
+    return value
+
 
-    Raised by `left_hand_split()` and `right_hand_split()`.
+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()
+@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',
+    "-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(
-    '--check',
+    "-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. Deprecated; use --target-version instead. "
+        "[default: per-file auto-detection]"
+    ),
+)
+@click.option(
+    "--pyi",
+    is_flag=True,
+    help=(
+        "Format all input files like typing stubs regardless of file extension "
+        "(useful when piping source on standard input)."
+    ),
+)
+@click.option(
+    "-S",
+    "--skip-string-normalization",
+    is_flag=True,
+    help="Don't normalize string quotes or prefixes.",
+)
+@click.option(
+    "--check",
     is_flag=True,
     help=(
-        "Don't write back the files, just return the status.  Return code 0 "
-        "means nothing changed.  Return code 1 means some files were "
+        "Don't write the files back, just return the status.  Return code 0 "
+        "means nothing would change.  Return code 1 means some files would be "
         "reformatted.  Return code 123 means there was an internal error."
     ),
 )
 @click.option(
-    '--fast/--safe',
+    "--diff",
+    is_flag=True,
+    help="Don't write the files back, just output a diff for each file on stdout.",
+)
+@click.option(
+    "--fast/--safe",
+    is_flag=True,
+    help="If --fast given, skip temporary sanity checks. [default: --safe]",
+)
+@click.option(
+    "--include",
+    type=str,
+    default=DEFAULT_INCLUDES,
+    help=(
+        "A regular expression that matches files and directories that should be "
+        "included on recursive searches.  An empty value means all files are "
+        "included regardless of the name.  Use forward slashes for directories on "
+        "all platforms (Windows, too).  Exclusions are calculated first, inclusions "
+        "later."
+    ),
+    show_default=True,
+)
+@click.option(
+    "--exclude",
+    type=str,
+    default=DEFAULT_EXCLUDES,
+    help=(
+        "A regular expression that matches files and directories that should be "
+        "excluded on recursive searches.  An empty value means no paths are excluded. "
+        "Use forward slashes for directories on all platforms (Windows, too).  "
+        "Exclusions are calculated first, inclusions later."
+    ),
+    show_default=True,
+)
+@click.option(
+    "-q",
+    "--quiet",
+    is_flag=True,
+    help=(
+        "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='If --fast given, skip temporary sanity checks. [default: --safe]',
+    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',
+    "src",
     nargs=-1,
-    type=click.Path(exists=True, file_okay=True, dir_okay=True, readable=True),
+    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, line_length: int, check: bool, fast: bool, src: List[str]
+    ctx: click.Context,
+    code: Optional[str],
+    line_length: int,
+    target_version: List[TargetVersion],
+    check: bool,
+    diff: bool,
+    fast: bool,
+    pyi: bool,
+    py36: bool,
+    skip_string_normalization: bool,
+    quiet: bool,
+    verbose: bool,
+    include: str,
+    exclude: 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(f"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 = FileMode(
+        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_maybe_verbose(include)
+    except re.error:
+        err(f"Invalid regular expression for include given: {include!r}")
+        ctx.exit(2)
+    try:
+        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))
-        elif p.is_file():
+            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}')
+            err(f"invalid path: {s}")
     if len(sources) == 0:
+        if verbose or not quiet:
+            out("No Python files are present to be formatted. Nothing to do 😴")
         ctx.exit(0)
-    elif len(sources) == 1:
-        p = sources[0]
-        report = Report()
-        try:
-            changed = format_file_in_place(
-                p, line_length=line_length, fast=fast, write_back=not check
-            )
-            report.done(p, changed)
-        except Exception as exc:
-            report.failed(p, str(exc))
-        ctx.exit(report.return_code)
+
+    if len(sources) == 1:
+        reformat_one(
+            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())
-        return_code = 1
-        try:
-            return_code = loop.run_until_complete(
-                schedule_formatting(
-                    sources, line_length, not check, fast, loop, executor
-                )
+        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, fast: bool, write_back: WriteBack, mode: FileMode, report: "Report"
+) -> None:
+    """Reformat a single file under `src` without spawning child processes.
+
+    `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(fast=fast, write_back=write_back, mode=mode):
+                changed = Changed.YES
+        else:
+            cache: Cache = {}
+            if write_back != WriteBack.DIFF:
+                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, fast=fast, write_back=write_back, mode=mode
+            ):
+                changed = Changed.YES
+            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: FileMode,
+    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:
-            loop.close()
-            ctx.exit(return_code)
+        )
+    finally:
+        shutdown(loop)
+        executor.shutdown()
 
 
 async def schedule_formatting(
-    sources: List[Path],
-    line_length: int,
-    write_back: bool,
+    sources: Set[Path],
     fast: bool,
-    loop: BaseEventLoop,
+    write_back: WriteBack,
+    mode: FileMode,
+    report: "Report",
+    loop: asyncio.AbstractEventLoop,
     executor: Executor,
-) -> int:
+) -> None:
+    """Run formatting of `sources` in parallel using the provided `executor`.
+
+    (Use ProcessPoolExecutors for actual parallelism.)
+
+    `write_back`, `fast`, and `mode` options are passed to
+    :func:`format_file_in_place`.
+    """
+    cache: Cache = {}
+    if write_back != WriteBack.DIFF:
+        cache = read_cache(mode)
+        sources, cached = filter_cached(cache, sources)
+        for src in sorted(cached):
+            report.done(src, Changed.CACHED)
+    if not sources:
+        return
+
+    cancelled = []
+    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 = {
-        src: loop.run_in_executor(
-            executor, format_file_in_place, src, line_length, fast, write_back
-        )
-        for src in sources
+        asyncio.ensure_future(
+            loop.run_in_executor(
+                executor, format_file_in_place, src, fast, mode, write_back, lock
+            )
+        ): src
+        for src in sorted(sources)
     }
-    await asyncio.wait(tasks.values())
-    cancelled = []
-    report = Report()
-    for src, task in tasks.items():
-        if not task.done():
-            report.failed(src, 'timed out, cancelling')
-            task.cancel()
-            cancelled.append(task)
-        elif task.exception():
-            report.failed(src, str(task.exception()))
-        else:
-            report.done(src, task.result())
+    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.wait(cancelled, timeout=2)
-    out('All done! ✨ 🍰 ✨')
-    click.echo(str(report))
-    return report.return_code
+        await asyncio.gather(*cancelled, loop=loop, return_exceptions=True)
+    if sources_to_cache:
+        write_cache(cache, sources_to_cache, mode)
 
 
 def format_file_in_place(
-    src: Path, line_length: int, fast: bool, write_back: bool = False
+    src: Path,
+    fast: bool,
+    mode: FileMode,
+    write_back: WriteBack = WriteBack.NO,
+    lock: Any = None,  # multiprocessing.Manager().Lock() is some crazy proxy
 ) -> bool:
-    """Format the file and rewrite if changed. Return True if changed."""
+    """Format file under `src` path. Return True if changed.
+
+    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 = 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:
-        contents, encoding = format_file(src, line_length=line_length, fast=fast)
+        dst_contents = format_file_contents(src_contents, fast=fast, mode=mode)
     except NothingChanged:
         return False
 
-    if write_back:
-        with open(src, "w", encoding=encoding) as f:
-            f.write(contents)
+    if write_back == WriteBack.YES:
+        with open(src, "w", encoding=encoding, newline=newline) as f:
+            f.write(dst_contents)
+    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)
+
+        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_file(
-    src: Path, line_length: int, fast: bool
-) -> Tuple[FileContent, Encoding]:
-    """Reformats a file and returns its contents and encoding."""
-    with tokenize.open(src) as src_buffer:
-        src_contents = src_buffer.read()
-    if src_contents.strip() == '':
-        raise NothingChanged(src)
+def format_stdin_to_stdout(
+    fast: bool, *, write_back: WriteBack = WriteBack.NO, mode: FileMode
+) -> bool:
+    """Format file on stdin. Return True if changed.
+
+    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`.
+    """
+    then = datetime.utcnow()
+    src, encoding, newline = decode_bytes(sys.stdin.buffer.read())
+    dst = src
+    try:
+        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:
+            f.write(dst)
+        elif write_back == WriteBack.DIFF:
+            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, *, fast: bool, mode: FileMode
+) -> 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.
+    `mode` is passed to :func:`format_str`.
+    """
+    if src_contents.strip() == "":
+        raise NothingChanged
 
-    dst_contents = format_str(src_contents, line_length=line_length)
+    dst_contents = format_str(src_contents, mode=mode)
     if src_contents == dst_contents:
-        raise NothingChanged(src)
+        raise NothingChanged
 
     if not fast:
         assert_equivalent(src_contents, dst_contents)
-        assert_stable(src_contents, dst_contents, line_length=line_length)
-    return dst_contents, src_buffer.encoding
-
-
-def format_str(src_contents: str, line_length: int) -> FileContent:
-    """Reformats a string and returns new contents."""
-    src_node = lib2to3_parse(src_contents)
-    dst_contents = ""
-    comments: List[Line] = []
-    lines = LineGenerator()
-    elt = EmptyLineTracker()
-    py36 = is_python36(src_node)
+        assert_stable(src_contents, dst_contents, mode=mode)
+    return dst_contents
+
+
+def format_str(src_contents: str, *, mode: FileMode) -> FileContent:
+    """Reformat a string and return new contents.
+
+    `mode` determines formatting options, such as how many characters per line are
+    allowed.
+    """
+    src_node = lib2to3_parse(src_contents.lstrip(), mode.target_versions)
+    dst_contents = []
+    future_imports = get_future_imports(src_node)
+    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="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=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)
-        if not current_line.is_comment:
-            for comment in comments:
-                dst_contents += str(comment)
-            comments = []
-            for line in split_line(current_line, line_length=line_length, py36=py36):
-                dst_contents += str(line)
-        else:
-            comments.append(current_line)
-    for comment in comments:
-        dst_contents += str(comment)
-    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).
 
-def lib2to3_parse(src_txt: str) -> Node:
+    `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,
+        ]
+
+    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, target_versions: Iterable[TargetVersion] = ()) -> Node:
     """Given a string with source, return the lib2to3 Node."""
-    grammar = pygram.python_grammar_no_print_statement
-    drv = driver.Driver(grammar, pytree.convert)
-    if src_txt[-1] != '\n':
-        nl = '\r\n' if '\r\n' in src_txt[:1024] else '\n'
-        src_txt += nl
-    try:
-        result = drv.parse_string(src_txt, True)
-    except ParseError as pe:
-        lineno, column = pe.context[1]
-        lines = src_txt.splitlines()
+    if src_txt[-1:] != "\n":
+        src_txt += "\n"
+
+    for grammar in get_grammars(set(target_versions)):
+        drv = driver.Driver(grammar, pytree.convert)
         try:
-            faulty_line = lines[lineno - 1]
-        except IndexError:
-            faulty_line = "<line number missing in source>"
-        raise ValueError(f"Cannot parse: {lineno}:{column}: {faulty_line}") from None
+            result = drv.parse_string(src_txt, True)
+            break
+
+        except ParseError as pe:
+            lineno, column = pe.context[1]
+            lines = src_txt.splitlines()
+            try:
+                faulty_line = lines[lineno - 1]
+            except IndexError:
+                faulty_line = "<line number missing in source>"
+            exc = InvalidInput(f"Cannot parse: {lineno}:{column}: {faulty_line}")
+    else:
+        raise exc from None
 
     if isinstance(result, Leaf):
         result = Node(syms.file_input, [result])
@@ -244,20 +870,38 @@ def lib2to3_unparse(node: Node) -> str:
     return code
 
 
-T = TypeVar('T')
+T = TypeVar("T")
 
 
 class Visitor(Generic[T]):
-    """Basic lib2to3 visitor that yields things on visiting."""
+    """Basic lib2to3 visitor that yields things of type `T` on `visit()`."""
 
     def visit(self, node: LN) -> Iterator[T]:
+        """Main method to visit `node` and its children.
+
+        It tries to find a `visit_*()` method for the given `node.type`, like
+        `visit_simple_stmt` for Node objects or `visit_INDENT` for Leaf objects.
+        If no dedicated `visit_*()` method is found, chooses `visit_default()`
+        instead.
+
+        Then yields objects of type `T` from the selected visitor.
+        """
         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`."""
         if isinstance(node, Node):
             for child in node.children:
                 yield from self.visit(child)
@@ -265,33 +909,42 @@ class Visitor(Generic[T]):
 
 @dataclass
 class DebugVisitor(Visitor[T]):
-    tree_depth: int = attrib(default=0)
+    tree_depth: int = 0
 
     def visit_default(self, node: LN) -> Iterator[T]:
-        indent = ' ' * (2 * self.tree_depth)
+        indent = " " * (2 * self.tree_depth)
         if isinstance(node, Node):
             _type = type_repr(node.type)
-            out(f'{indent}{_type}', fg='yellow')
+            out(f"{indent}{_type}", fg="yellow")
             self.tree_depth += 1
             for child in node.children:
                 yield from self.visit(child)
 
             self.tree_depth -= 1
-            out(f'{indent}/{_type}', fg='yellow', bold=False)
+            out(f"{indent}/{_type}", fg="yellow", bold=False)
         else:
             _type = token.tok_name.get(node.type, str(node.type))
-            out(f'{indent}{_type}', fg='blue', nl=False)
+            out(f"{indent}{_type}", fg="blue", nl=False)
             if node.prefix:
                 # We don't have to handle prefixes for `Node` objects since
                 # that delegates to the first child anyway.
-                out(f' {node.prefix!r}', fg='green', bold=False, nl=False)
-            out(f' {node.value!r}', fg='blue', bold=False)
+                out(f" {node.prefix!r}", fg="green", bold=False, nl=False)
+            out(f" {node.value!r}", fg="blue", bold=False)
+
+    @classmethod
+    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()
+        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,
@@ -301,9 +954,10 @@ 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,
@@ -311,136 +965,383 @@ COMPARATORS = {
     token.LESSEQUAL,
     token.GREATEREQUAL,
 }
-MATH_OPERATORS = {
+MATH_OPERATORS: Final = {
+    token.VBAR,
+    token.CIRCUMFLEX,
+    token.AMPER,
+    token.LEFTSHIFT,
+    token.RIGHTSHIFT,
     token.PLUS,
     token.MINUS,
     token.STAR,
     token.SLASH,
-    token.VBAR,
-    token.AMPER,
+    token.DOUBLESLASH,
     token.PERCENT,
-    token.CIRCUMFLEX,
-    token.LEFTSHIFT,
-    token.RIGHTSHIFT,
+    token.AT,
+    token.TILDE,
     token.DOUBLESTAR,
-    token.DOUBLESLASH,
 }
-COMPREHENSION_PRIORITY = 20
-COMMA_PRIORITY = 10
-LOGIC_PRIORITY = 5
-STRING_PRIORITY = 4
-COMPARATOR_PRIORITY = 3
-MATH_PRIORITY = 1
+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: Final = {
+    syms.atom,  # single element of a list or set literal
+    syms.dictsetmaker,
+    syms.listmaker,
+    syms.testlist_gexp,
+    syms.testlist_star_expr,
+}
+TEST_DESCENDANTS: Final = {
+    syms.test,
+    syms.lambdef,
+    syms.or_test,
+    syms.and_test,
+    syms.not_test,
+    syms.comparison,
+    syms.star_expr,
+    syms.expr,
+    syms.xor_expr,
+    syms.and_expr,
+    syms.shift_expr,
+    syms.arith_expr,
+    syms.trailer,
+    syms.term,
+    syms.power,
+}
+ASSIGNMENTS: Final = {
+    "=",
+    "+=",
+    "-=",
+    "*=",
+    "@=",
+    "/=",
+    "%=",
+    "&=",
+    "|=",
+    "^=",
+    "<<=",
+    ">>=",
+    "**=",
+    "//=",
+}
+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,
+    token.LEFTSHIFT: 6,
+    token.RIGHTSHIFT: 6,
+    token.PLUS: 5,
+    token.MINUS: 5,
+    token.STAR: 4,
+    token.SLASH: 4,
+    token.DOUBLESLASH: 4,
+    token.PERCENT: 4,
+    token.AT: 4,
+    token.TILDE: 3,
+    token.DOUBLESTAR: 2,
+}
+DOT_PRIORITY: Final = 1
 
 
 @dataclass
 class BracketTracker:
-    depth: int = attrib(default=0)
-    bracket_match: Dict[Tuple[Depth, NodeType], Leaf] = attrib(default=Factory(dict))
-    delimiters: Dict[LeafID, Priority] = attrib(default=Factory(dict))
-    previous: Optional[Leaf] = attrib(default=None)
+    """Keeps track of brackets on a line."""
+
+    depth: int = 0
+    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_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.
+
+        All leaves receive an int `bracket_depth` field that stores how deep
+        within brackets a given leaf is. 0 means there are no enclosing brackets
+        that started on this line.
+
+        If a leaf is itself a closing bracket, it receives an `opening_bracket`
+        field that it forms a pair with. This is a one-directional link to
+        avoid reference cycles.
+
+        If a leaf is a delimiter (a token on which Black can split the line if
+        needed) and it's on depth 0, its `id()` is stored in the tracker's
+        `delimiters` field.
+        """
         if leaf.type == token.COMMENT:
             return
 
+        self.maybe_decrement_after_for_loop_variable(leaf)
+        self.maybe_decrement_after_lambda_arguments(leaf)
         if leaf.type in CLOSING_BRACKETS:
             self.depth -= 1
             opening_bracket = self.bracket_match.pop((self.depth, leaf.type))
             leaf.opening_bracket = opening_bracket
         leaf.bracket_depth = self.depth
         if self.depth == 0:
-            delim = is_delimiter(leaf)
-            if delim:
-                self.delimiters[id(leaf)] = delim
-            elif self.previous is not None:
-                if leaf.type == token.STRING and self.previous.type == token.STRING:
-                    self.delimiters[id(self.previous)] = STRING_PRIORITY
-                elif (
-                    leaf.type == token.NAME and
-                    leaf.value == 'for' and
-                    leaf.parent and
-                    leaf.parent.type in {syms.comp_for, syms.old_comp_for}
-                ):
-                    self.delimiters[id(self.previous)] = COMPREHENSION_PRIORITY
-                elif (
-                    leaf.type == token.NAME and
-                    leaf.value == 'if' and
-                    leaf.parent and
-                    leaf.parent.type in {syms.comp_if, syms.old_comp_if}
-                ):
-                    self.delimiters[id(self.previous)] = COMPREHENSION_PRIORITY
+            delim = is_split_before_delimiter(leaf, self.previous)
+            if delim and self.previous is not None:
+                self.delimiters[id(self.previous)] = delim
+            else:
+                delim = is_split_after_delimiter(leaf, self.previous)
+                if delim:
+                    self.delimiters[id(leaf)] = delim
         if leaf.type in OPENING_BRACKETS:
             self.bracket_match[self.depth, BRACKET[leaf.type]] = leaf
             self.depth += 1
         self.previous = leaf
+        self.maybe_increment_lambda_arguments(leaf)
+        self.maybe_increment_for_loop_variable(leaf)
 
     def any_open_brackets(self) -> bool:
-        """Returns True if there is an yet unmatched open bracket on the line."""
+        """Return True if there is an yet unmatched open bracket on the line."""
         return bool(self.bracket_match)
 
-    def max_priority(self, exclude: Iterable[LeafID] =()) -> int:
-        """Returns the highest priority of a delimiter found on the line.
+    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_delimiter()` returns.
+        Values are consistent with what `is_split_*_delimiter()` return.
+        Raises ValueError on no delimiters.
         """
         return max(v for k, v in self.delimiters.items() if k not in exclude)
 
+    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.
+        """
+        if not self.delimiters:
+            return 0
+
+        priority = priority or self.max_delimiter_priority()
+        return sum(1 for p in self.delimiters.values() if p == priority)
+
+    def maybe_increment_for_loop_variable(self, leaf: Leaf) -> bool:
+        """In a for loop, or comprehension, the variables are often unpacks.
+
+        To avoid splitting on the comma in this situation, increase the depth of
+        tokens between `for` and `in`.
+        """
+        if leaf.type == token.NAME and leaf.value == "for":
+            self.depth += 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_depths
+            and self._for_loop_depths[-1] == self.depth
+            and leaf.type == token.NAME
+            and leaf.value == "in"
+        ):
+            self.depth -= 1
+            self._for_loop_depths.pop()
+            return True
+
+        return False
+
+    def maybe_increment_lambda_arguments(self, leaf: Leaf) -> bool:
+        """In a lambda expression, there might be more than one argument.
+
+        To avoid splitting on the comma in this situation, increase the depth of
+        tokens between `lambda` and `:`.
+        """
+        if leaf.type == token.NAME and leaf.value == "lambda":
+            self.depth += 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_argument_depths
+            and self._lambda_argument_depths[-1] == self.depth
+            and leaf.type == token.COLON
+        ):
+            self.depth -= 1
+            self._lambda_argument_depths.pop()
+            return True
+
+        return False
+
+    def get_open_lsqb(self) -> Optional[Leaf]:
+        """Return the most recent opening square bracket (if any)."""
+        return self.bracket_match.get((self.depth - 1, token.RSQB))
+
 
 @dataclass
 class Line:
-    depth: int = attrib(default=0)
-    leaves: List[Leaf] = attrib(default=Factory(list))
-    comments: Dict[LeafID, Leaf] = attrib(default=Factory(dict))
-    bracket_tracker: BracketTracker = attrib(default=Factory(BracketTracker))
-    inside_brackets: bool = attrib(default=False)
-    has_for: bool = attrib(default=False)
-    _for_loop_variable: bool = attrib(default=False, init=False)
+    """Holds leaves and comments. Can be printed with `str(line)`."""
+
+    depth: int = 0
+    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
 
     def append(self, leaf: Leaf, preformatted: bool = False) -> None:
-        has_value = leaf.value.strip()
+        """Add a new `leaf` to the end of the line.
+
+        Unless `preformatted` is True, the `leaf` will receive a new consistent
+        whitespace prefix and metadata applied by :class:`BracketTracker`.
+        Trailing commas are maybe removed, unpacked for loop variables are
+        demoted from being delimiters.
+
+        Inline comments are put aside.
+        """
+        has_value = leaf.type in BRACKETS or bool(leaf.value.strip())
         if not has_value:
             return
 
+        if token.COLON == leaf.type and self.is_class_paren_empty:
+            del self.leaves[-2:]
         if self.leaves and not preformatted:
             # Note: at this point leaf.prefix should be empty except for
             # imports, for which we only preserve newlines.
-            leaf.prefix += whitespace(leaf)
+            leaf.prefix += whitespace(
+                leaf, complex_subscript=self.is_complex_subscript(leaf)
+            )
         if self.inside_brackets or not preformatted:
-            self.maybe_decrement_after_for_loop_variable(leaf)
             self.bracket_tracker.mark(leaf)
             self.maybe_remove_trailing_comma(leaf)
-            self.maybe_increment_for_loop_variable(leaf)
-            if self.maybe_adapt_standalone_comment(leaf):
-                return
-
         if not self.append_comment(leaf):
             self.leaves.append(leaf)
 
+    def append_safe(self, leaf: Leaf, preformatted: bool = False) -> None:
+        """Like :func:`append()` but disallow invalid standalone comment structure.
+
+        Raises ValueError when any `leaf` is appended after a standalone comment
+        or when a standalone comment is not the first leaf on the line.
+        """
+        if self.bracket_tracker.depth == 0:
+            if self.is_comment:
+                raise ValueError("cannot append to standalone comments")
+
+            if self.leaves and leaf.type == STANDALONE_COMMENT:
+                raise ValueError(
+                    "cannot append standalone comments to a populated line"
+                )
+
+        self.append(leaf, preformatted=preformatted)
+
     @property
     def is_comment(self) -> bool:
-        return bool(self) and self.leaves[0].type == STANDALONE_COMMENT
+        """Is this line a standalone comment?"""
+        return len(self.leaves) == 1 and self.leaves[0].type == STANDALONE_COMMENT
 
     @property
     def is_decorator(self) -> bool:
+        """Is this line a decorator?"""
         return bool(self) and self.leaves[0].type == token.AT
 
     @property
     def is_import(self) -> bool:
+        """Is this an import line?"""
         return bool(self) and is_import(self.leaves[0])
 
     @property
     def is_class(self) -> bool:
+        """Is this line a class definition?"""
         return (
-            bool(self) and
-            self.leaves[0].type == token.NAME and
-            self.leaves[0].value == 'class'
+            bool(self)
+            and self.leaves[0].type == token.NAME
+            and self.leaves[0].value == "class"
         )
 
     @property
-    def is_def(self) -> bool:
-        """Also returns True for async defs."""
-        try:
+    def is_stub_class(self) -> bool:
+        """Is this line a class definition with a body consisting only of "..."?"""
+        return self.is_class and self.leaves[-3:] == [
+            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.)"""
+        try:
             first_leaf = self.leaves[0]
         except IndexError:
             return False
@@ -449,158 +1350,237 @@ class Line:
             second_leaf: Optional[Leaf] = self.leaves[1]
         except IndexError:
             second_leaf = None
-        return (
-            (first_leaf.type == token.NAME and first_leaf.value == 'def') or
-            (
-                first_leaf.type == token.NAME and
-                first_leaf.value == 'async' and
-                second_leaf is not None and
-                second_leaf.type == token.NAME and
-                second_leaf.value == 'def'
-            )
+        return (first_leaf.type == token.NAME and first_leaf.value == "def") or (
+            first_leaf.type == token.ASYNC
+            and second_leaf is not None
+            and second_leaf.type == token.NAME
+            and second_leaf.value == "def"
         )
 
     @property
-    def is_flow_control(self) -> bool:
+    def is_class_paren_empty(self) -> bool:
+        """Is this a class with no base classes but using parentheses?
+
+        Those are unnecessary and should be removed.
+        """
         return (
-            bool(self) and
-            self.leaves[0].type == token.NAME and
-            self.leaves[0].value in FLOW_CONTROL
+            bool(self)
+            and len(self.leaves) == 4
+            and self.is_class
+            and self.leaves[2].type == token.LPAR
+            and self.leaves[2].value == "("
+            and self.leaves[3].type == token.RPAR
+            and self.leaves[3].value == ")"
         )
 
     @property
-    def is_yield(self) -> bool:
+    def is_triple_quoted_string(self) -> bool:
+        """Is the line a triple quoted string?"""
         return (
-            bool(self) and
-            self.leaves[0].type == token.NAME and
-            self.leaves[0].value == 'yield'
+            bool(self)
+            and self.leaves[0].type == token.STRING
+            and self.leaves[0].value.startswith(('"""', "'''"))
         )
 
-    def maybe_remove_trailing_comma(self, closing: Leaf) -> bool:
-        if not (
-            self.leaves and
-            self.leaves[-1].type == token.COMMA and
-            closing.type in CLOSING_BRACKETS
-        ):
-            return False
-
-        if closing.type == token.RSQB or closing.type == token.RBRACE:
-            self.leaves.pop()
-            return True
+    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 and leaf.bracket_depth <= depth_limit:
+                return True
 
-        # For parens let's check if it's safe to remove the comma.  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
+        return False
 
-        else:
+    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
 
-        for leaf in self.leaves[_opening_index + 1:]:
-            if leaf is closing:
-                break
-
-            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
+        # 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
 
-        if commas > 1:
-            self.leaves.pop()
-            return True
+                comment_seen = True
 
         return False
 
-    def maybe_increment_for_loop_variable(self, leaf: Leaf) -> bool:
-        """In a for loop, or comprehension, the variables are often unpacks.
-
-        To avoid splitting on the comma in this situation, we will increase
-        the depth of tokens between `for` and `in`.
-        """
-        if leaf.type == token.NAME and leaf.value == 'for':
-            self.has_for = True
-            self.bracket_tracker.depth += 1
-            self._for_loop_variable = True
-            return True
-
-        return False
+    def contains_unsplittable_type_ignore(self) -> bool:
+        if not self.leaves:
+            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':
-            self.bracket_tracker.depth -= 1
-            self._for_loop_variable = False
-            return True
+        # 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
 
         return False
 
-    def maybe_adapt_standalone_comment(self, comment: Leaf) -> bool:
-        """Hack a standalone comment to act as a trailing comment for line splitting.
+    def contains_multiline_strings(self) -> bool:
+        return any(is_multiline_string(leaf) for leaf in self.leaves)
 
-        If this line has brackets and a standalone `comment`, we need to adapt
-        it to be able to still reformat the line.
+    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
 
-        This is not perfect, the line to which the standalone comment gets
-        appended will appear "too long" when splitting.
-        """
+        # We remove trailing commas only in the case of importing a
+        # single name from a module.
         if not (
-            comment.type == STANDALONE_COMMENT and
-            self.bracket_tracker.any_open_brackets()
+            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
 
-        comment.type = token.COMMENT
-        comment.prefix = '\n' + '    ' * (self.depth + 1)
-        return self.append_comment(comment)
+        self.remove_trailing_comma()
+        return True
 
     def append_comment(self, comment: Leaf) -> bool:
+        """Add an inline or standalone comment to the line."""
+        if (
+            comment.type == STANDALONE_COMMENT
+            and self.bracket_tracker.any_open_brackets()
+        ):
+            comment.prefix = ""
+            return False
+
         if comment.type != token.COMMENT:
             return False
 
-        try:
-            after = id(self.last_non_delimiter())
-        except LookupError:
+        if not self.leaves:
             comment.type = STANDALONE_COMMENT
-            comment.prefix = ''
+            comment.prefix = ""
             return False
 
-        else:
-            if after in self.comments:
-                self.comments[after].value += str(comment)
-            else:
-                self.comments[after] = comment
-            return True
+        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
+
+    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."""
+        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 = self.bracket_tracker.get_open_lsqb()
+        if open_lsqb is None:
+            return False
+
+        subscript_start = open_lsqb.next_sibling
 
-    def last_non_delimiter(self) -> Leaf:
-        for i in range(len(self.leaves)):
-            last = self.leaves[-i - 1]
-            if not is_delimiter(last):
-                return last
+        if isinstance(subscript_start, Node):
+            if subscript_start.type == syms.listmaker:
+                return False
 
-        raise LookupError("No non-delimiters found")
+            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()
+        )
 
     def __str__(self) -> str:
+        """Render the line."""
         if not self:
-            return '\n'
+            return "\n"
 
-        indent = '    ' * self.depth
+        indent = "    " * self.depth
         leaves = iter(self.leaves)
         first = next(leaves)
-        res = f'{first.prefix}{indent}{first.value}'
+        res = f"{first.prefix}{indent}{first.value}"
         for leaf in leaves:
             res += str(leaf)
-        for comment in self.comments.values():
+        for comment in itertools.chain.from_iterable(self.comments.values()):
             res += str(comment)
-        return res + '\n'
+        return res + "\n"
 
     def __bool__(self) -> bool:
+        """Return True if the line has leaves or comments."""
         return bool(self.leaves or self.comments)
 
 
@@ -609,68 +1589,117 @@ class EmptyLineTracker:
     """Provides a stateful method that returns the number of potential extra
     empty lines needed before and after the currently processed line.
 
-    Note: this tracker works on lines that haven't been split yet.
+    Note: this tracker works on lines that haven't been split yet.  It assumes
+    the prefix of the first leaf consists of optional newlines.  Those newlines
+    are consumed by `maybe_empty_lines()` and included in the computation.
     """
-    previous_line: Optional[Line] = attrib(default=None)
-    previous_after: int = attrib(default=0)
-    previous_defs: List[int] = attrib(default=Factory(list))
+
+    is_pyi: bool = False
+    previous_line: Optional[Line] = None
+    previous_after: int = 0
+    previous_defs: List[int] = field(default_factory=list)
 
     def maybe_empty_lines(self, current_line: Line) -> Tuple[int, int]:
-        """Returns the number of extra empty lines before and after the `current_line`.
+        """Return the number of extra empty lines before and after the `current_line`.
 
-        This is for separating `def`, `async def` and `class` with extra empty lines
-        (two on module-level), as well as providing an extra empty line after flow
-        control keywords to make them more prominent.
+        This is for separating `def`, `async def` and `class` with extra empty
+        lines (two on module-level).
         """
         before, after = self._maybe_empty_lines(current_line)
+        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
 
     def _maybe_empty_lines(self, current_line: Line) -> Tuple[int, int]:
-        before = 0
+        max_allowed = 1
+        if current_line.depth == 0:
+            max_allowed = 1 if self.is_pyi else 2
+        if current_line.leaves:
+            # Consume the first leaf's extra newlines.
+            first_leaf = current_line.leaves[0]
+            before = first_leaf.prefix.count("\n")
+            before = min(before, max_allowed)
+            first_leaf.prefix = ""
+        else:
+            before = 0
         depth = current_line.depth
         while self.previous_defs and self.previous_defs[-1] >= depth:
             self.previous_defs.pop()
-            before = (1 if depth else 2) - self.previous_after
-        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 and self.previous_line.is_decorator:
-                # Don't insert empty lines between decorators.
-                return 0, 0
-
-            newlines = 2
-            if current_line.depth:
-                newlines -= 1
-            newlines -= self.previous_after
-            return newlines, 0
-
-        if current_line.is_flow_control:
-            return before, 1
+            if self.is_pyi:
+                before = 0 if depth else 1
+            else:
+                before = 1 if depth else 2
+        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 and
-            self.previous_line.is_import and
-            not current_line.is_import and
-            depth == self.previous_line.depth
+            self.previous_line
+            and self.previous_line.is_import
+            and not current_line.is_import
+            and depth == self.previous_line.depth
         ):
             return (before or 1), 0
 
         if (
-            self.previous_line and
-            self.previous_line.is_yield and
-            (not current_line.is_yield or depth != self.previous_line.depth)
+            self.previous_line
+            and self.previous_line.is_class
+            and current_line.is_triple_quoted_string
         ):
-            return (before or 1), 0
+            return before, 1
 
         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]):
@@ -679,8 +1708,11 @@ class LineGenerator(Visitor[Line]):
     Note: destroys the tree it's visiting by mutating prefixes of its leaves
     in ways that will no longer stringify to valid Python code on the tree.
     """
-    current_line: Line = attrib(default=Factory(Line))
-    standalone_comments: List[Leaf] = attrib(default=Factory(list))
+
+    is_pyi: bool = False
+    normalize_strings: bool = True
+    current_line: Line = field(default_factory=Line)
+    remove_u_prefix: bool = False
 
     def line(self, indent: int = 0) -> Iterator[Line]:
         """Generate a line.
@@ -699,9 +1731,11 @@ class LineGenerator(Visitor[Line]):
         yield complete_line
 
     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()
             for comment in generate_comments(node):
-                if self.current_line.bracket_tracker.any_open_brackets():
+                if any_open_brackets:
                     # any comment within brackets is subject to splitting
                     self.current_line.append(comment)
                 elif comment.type == token.COMMENT:
@@ -710,67 +1744,95 @@ class LineGenerator(Visitor[Line]):
                     yield from self.line()
 
                 else:
-                    # regular standalone comment, to be processed later (see
-                    # docstring in `generate_comments()`
-                    self.standalone_comments.append(comment)
-            normalize_prefix(node)
-            if node.type not in WHITESPACE:
-                for comment in self.standalone_comments:
+                    # regular standalone comment
                     yield from self.line()
 
                     self.current_line.append(comment)
                     yield from self.line()
 
-                self.standalone_comments = []
+            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_suite(self, node: Node) -> Iterator[Line]:
-        """Body of a statement after a colon."""
-        children = iter(node.children)
-        # Process newline before indenting.  It might contain an inline
-        # comment that should go right after the colon.
-        newline = next(children)
-        yield from self.visit(newline)
+    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)
 
-        for child in children:
-            yield from self.visit(child)
+    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).
+        # Emit the line then.
+        yield from self.line()
 
+        # While DEDENT has no value, its prefix may contain standalone comments
+        # that belong to the current indentation level.  Get 'em.
+        yield from self.visit_default(node)
+
+        # Finally, emit the dedent.
         yield from self.line(-1)
 
-    def visit_stmt(self, node: Node, keywords: Set[str]) -> Iterator[Line]:
+    def visit_stmt(
+        self, node: Node, keywords: Set[str], parens: Set[str]
+    ) -> Iterator[Line]:
         """Visit a statement.
 
-        The relevant Python language keywords for this statement are NAME leaves
-        within it.
+        This implementation is shared for `if`, `while`, `for`, `try`, `except`,
+        `def`, `with`, `class`, `assert` and assignments.
+
+        The relevant Python language `keywords` for a given statement will be
+        NAME leaves within it. This methods puts those on a separate line.
+
+        `parens` holds a set of string leaf values immediately after which
+        invisible parens should be put.
         """
+        normalize_invisible_parens(node, parens_after=parens)
         for child in node.children:
             if child.type == token.NAME and child.value in keywords:  # type: ignore
                 yield from self.line()
 
             yield from self.visit(child)
 
+    def visit_suite(self, node: Node) -> Iterator[Line]:
+        """Visit a suite."""
+        if self.is_pyi and is_stub_suite(node):
+            yield from self.visit(node.children[2])
+        else:
+            yield from self.visit_default(node)
+
     def visit_simple_stmt(self, node: Node) -> Iterator[Line]:
-        """A statement without nested statements."""
+        """Visit a statement without nested statements."""
         is_suite_like = node.parent and node.parent.type in STATEMENT
         if is_suite_like:
-            yield from self.line(+1)
-            yield from self.visit_default(node)
-            yield from self.line(-1)
+            if self.is_pyi and is_stub_body(node):
+                yield from self.visit_default(node)
+            else:
+                yield from self.line(+1)
+                yield from self.visit_default(node)
+                yield from self.line(-1)
 
         else:
-            yield from self.line()
+            if not self.is_pyi or not node.parent or not is_stub_suite(node.parent):
+                yield from self.line()
             yield from self.visit_default(node)
 
     def visit_async_stmt(self, node: Node) -> Iterator[Line]:
+        """Visit `async def`, `async for`, `async with`."""
         yield from self.line()
 
         children = iter(node.children)
         for child in children:
             yield from self.visit(child)
 
-            if child.type == token.NAME and child.value == 'async':  # type: ignore
+            if child.type == token.ASYNC:
                 break
 
         internal_stmt = next(children)
@@ -778,44 +1840,84 @@ class LineGenerator(Visitor[Line]):
             yield from self.visit(child)
 
     def visit_decorators(self, node: Node) -> Iterator[Line]:
+        """Visit decorators."""
         for child in node.children:
             yield from self.line()
             yield from self.visit(child)
 
     def visit_SEMI(self, leaf: Leaf) -> Iterator[Line]:
+        """Remove a semicolon and put the other statement on a separate line."""
         yield from self.line()
 
     def visit_ENDMARKER(self, leaf: Leaf) -> Iterator[Line]:
+        """End of file. Process outstanding comments and end with a newline."""
         yield from self.visit_default(leaf)
         yield from self.line()
 
-    def __attrs_post_init__(self) -> None:
+    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)
+
+    def visit_factor(self, node: Node) -> Iterator[Line]:
+        """Force parentheses between a unary op and a binary power:
+
+        -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 __post_init__(self) -> None:
         """You are in a twisty little maze of passages."""
         v = self.visit_stmt
-        self.visit_if_stmt = partial(v, keywords={'if', 'else', 'elif'})
-        self.visit_while_stmt = partial(v, keywords={'while', 'else'})
-        self.visit_for_stmt = partial(v, keywords={'for', 'else'})
-        self.visit_try_stmt = partial(v, keywords={'try', 'except', 'else', 'finally'})
-        self.visit_except_clause = partial(v, keywords={'except'})
-        self.visit_funcdef = partial(v, keywords={'def'})
-        self.visit_with_stmt = partial(v, keywords={'with'})
-        self.visit_classdef = partial(v, keywords={'class'})
+        Ø: Set[str] = set()
+        self.visit_assert_stmt = partial(v, keywords={"assert"}, parens={"assert", ","})
+        self.visit_if_stmt = partial(
+            v, keywords={"if", "else", "elif"}, parens={"if", "elif"}
+        )
+        self.visit_while_stmt = partial(v, keywords={"while", "else"}, parens={"while"})
+        self.visit_for_stmt = partial(v, keywords={"for", "else"}, parens={"for", "in"})
+        self.visit_try_stmt = partial(
+            v, keywords={"try", "except", "else", "finally"}, parens=Ø
+        )
+        self.visit_except_clause = partial(v, keywords={"except"}, parens=Ø)
+        self.visit_with_stmt = partial(v, keywords={"with"}, parens=Ø)
+        self.visit_funcdef = partial(v, keywords={"def"}, parens=Ø)
+        self.visit_classdef = partial(v, keywords={"class"}, parens=Ø)
+        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
 
 
+IMPLICIT_TUPLE = {syms.testlist, syms.testlist_star_expr, syms.exprlist}
 BRACKET = {token.LPAR: token.RPAR, token.LSQB: token.RSQB, token.LBRACE: token.RBRACE}
 OPENING_BRACKETS = set(BRACKET.keys())
 CLOSING_BRACKETS = set(BRACKET.values())
 BRACKETS = OPENING_BRACKETS | CLOSING_BRACKETS
-ALWAYS_NO_SPACE = CLOSING_BRACKETS | {token.COMMA, token.COLON, STANDALONE_COMMENT}
+ALWAYS_NO_SPACE = CLOSING_BRACKETS | {token.COMMA, STANDALONE_COMMENT}
+
 
+def whitespace(leaf: Leaf, *, complex_subscript: bool) -> str:  # noqa: C901
+    """Return whitespace prefix if needed for the given `leaf`.
 
-def whitespace(leaf: Leaf) -> str:  # noqa C901
-    """Return whitespace prefix if needed for the given `leaf`."""
-    NO = ''
-    SPACE = ' '
-    DOUBLESPACE = '  '
+    `complex_subscript` signals whether the given leaf is part of a subscription
+    which has non-trivial arguments, like arithmetic expressions or function calls.
+    """
+    NO = ""
+    SPACE = " "
+    DOUBLESPACE = "  "
     t = leaf.type
     p = leaf.parent
     v = leaf.value
@@ -826,37 +1928,68 @@ def whitespace(leaf: Leaf) -> str:  # noqa C901
         return DOUBLESPACE
 
     assert p is not None, f"INTERNAL ERROR: hand-made leaf without parent: {leaf!r}"
+    if t == token.COLON and p.type not in {
+        syms.subscript,
+        syms.subscriptlist,
+        syms.sliceop,
+    }:
+        return NO
+
     prev = leaf.prev_sibling
     if not prev:
         prevp = preceding_leaf(p)
         if not prevp or prevp.type in OPENING_BRACKETS:
             return NO
 
-        if prevp.type == token.EQUAL:
-            if prevp.parent and prevp.parent.type in {
-                syms.typedargslist,
-                syms.varargslist,
-                syms.parameters,
-                syms.arglist,
-                syms.argument,
-            }:
+        if t == token.COLON:
+            if prevp.type == token.COLON:
                 return NO
 
-        elif prevp.type == token.DOUBLESTAR:
-            if prevp.parent and prevp.parent.type in {
-                syms.typedargslist,
-                syms.varargslist,
-                syms.parameters,
-                syms.arglist,
-                syms.dictsetmaker,
-            }:
+            elif prevp.type != token.COMMA and not complex_subscript:
                 return NO
 
-        elif prevp.type == token.COLON:
-            if prevp.parent and prevp.parent.type == syms.subscript:
+            return SPACE
+
+        if prevp.type == token.EQUAL:
+            if prevp.parent:
+                if prevp.parent.type in {
+                    syms.arglist,
+                    syms.argument,
+                    syms.parameters,
+                    syms.varargslist,
+                }:
+                    return NO
+
+                elif prevp.parent.type == syms.typedargslist:
+                    # A bit hacky: if the equal sign has whitespace, it means we
+                    # previously found it's a typed argument.  So, we're using
+                    # that, too.
+                    return prevp.prefix
+
+        elif prevp.type in VARARGS_SPECIALS:
+            if is_vararg(prevp, within=VARARGS_PARENTS | UNPACKING_PARENTS):
                 return NO
 
-        elif prevp.parent and prevp.parent.type in {syms.factor, syms.star_expr}:
+        elif prevp.type == token.COLON:
+            if prevp.parent and prevp.parent.type in {syms.subscript, syms.sliceop}:
+                return SPACE if complex_subscript else NO
+
+        elif (
+            prevp.parent
+            and prevp.parent.type == syms.factor
+            and prevp.type in MATH_OPERATORS
+        ):
+            return NO
+
+        elif (
+            prevp.type == token.RIGHTSHIFT
+            and prevp.parent
+            and prevp.parent.type == syms.shift_expr
+            and prevp.prev_sibling
+            and prevp.prev_sibling.type == token.NAME
+            and prevp.prev_sibling.value == "print"  # type: ignore
+        ):
+            # Python 2 print chevron
             return NO
 
     elif prev.type in OPENING_BRACKETS:
@@ -864,17 +1997,11 @@ def whitespace(leaf: Leaf) -> str:  # noqa C901
 
     if p.type in {syms.parameters, syms.arglist}:
         # untyped function signatures or calls
-        if t == token.RPAR:
-            return NO
-
         if not prev or prev.type != token.COMMA:
             return NO
 
-    if p.type == syms.varargslist:
+    elif p.type == syms.varargslist:
         # lambdas
-        if t == token.RPAR:
-            return NO
-
         if prev and prev.type != token.COMMA:
             return NO
 
@@ -929,7 +2056,7 @@ def whitespace(leaf: Leaf) -> str:  # noqa C901
             if not prevp or prevp.type == token.LPAR:
                 return NO
 
-        elif prev.type == token.EQUAL or prev.type == token.DOUBLESTAR:
+        elif prev.type in {token.EQUAL} | VARARGS_SPECIALS:
             return NO
 
     elif p.type == syms.decorator:
@@ -951,7 +2078,7 @@ def whitespace(leaf: Leaf) -> str:  # noqa C901
         if prev and prev.type == token.LPAR:
             return NO
 
-    elif p.type == syms.subscript:
+    elif p.type in {syms.subscript, syms.sliceop}:
         # indexing
         if not prev:
             assert p.parent is not None, "subscripts are always parented"
@@ -960,7 +2087,7 @@ def whitespace(leaf: Leaf) -> str:  # noqa C901
 
             return NO
 
-        elif prev.type == token.COLON:
+        elif not complex_subscript:
             return NO
 
     elif p.type == syms.atom:
@@ -968,21 +2095,9 @@ def whitespace(leaf: Leaf) -> str:  # noqa C901
             # dots, but not the first one.
             return NO
 
-    elif (
-        p.type == syms.listmaker or
-        p.type == syms.testlist_gexp or
-        p.type == syms.subscriptlist
-    ):
-        # list interior, including unpacking
-        if not prev:
-            return NO
-
     elif p.type == syms.dictsetmaker:
-        # dict and set interior, including unpacking
-        if not prev:
-            return NO
-
-        if prev.type == token.DOUBLESTAR:
+        # dict unpacking
+        if prev and prev.type == token.DOUBLESTAR:
             return NO
 
     elif p.type in {syms.factor, syms.star_expr}:
@@ -995,14 +2110,15 @@ def whitespace(leaf: Leaf) -> str:  # noqa C901
             prevp_parent = prevp.parent
             assert prevp_parent is not None
             if prevp.type == token.COLON and prevp_parent.type in {
-                syms.subscript, syms.sliceop
+                syms.subscript,
+                syms.sliceop,
             }:
                 return NO
 
             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:
@@ -1011,7 +2127,7 @@ def whitespace(leaf: Leaf) -> str:  # noqa C901
                 return NO
 
         elif t == token.NAME:
-            if v == 'import':
+            if v == "import":
                 return SPACE
 
             if prev and prev.type == token.DOT:
@@ -1024,7 +2140,7 @@ def whitespace(leaf: Leaf) -> str:  # noqa C901
 
 
 def preceding_leaf(node: Optional[LN]) -> Optional[Leaf]:
-    """Returns the first leaf that precedes `node`, if any."""
+    """Return the first leaf that precedes `node`, if any."""
     while node:
         res = node.prev_sibling
         if res:
@@ -1041,32 +2157,155 @@ def preceding_leaf(node: Optional[LN]) -> Optional[Leaf]:
     return None
 
 
-def is_delimiter(leaf: Leaf) -> int:
-    """Returns the priority of the `leaf` delimiter. Returns 0 if not delimiter.
+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: 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
+    cause a line break after themselves.
 
     Higher numbers are higher priority.
     """
     if leaf.type == token.COMMA:
         return COMMA_PRIORITY
 
-    if leaf.type == token.NAME and leaf.value in LOGIC_OPERATORS:
-        return LOGIC_PRIORITY
+    return 0
+
+
+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.
+
+    Higher numbers are higher priority.
+    """
+    if is_vararg(leaf, within=VARARGS_PARENTS | UNPACKING_PARENTS):
+        # * and ** might also be MATH_OPERATORS but in this case they are not.
+        # Don't treat them as a delimiter.
+        return 0
+
+    if (
+        leaf.type == token.DOT
+        and leaf.parent
+        and leaf.parent.type not in {syms.import_from, syms.dotted_name}
+        and (previous is None or previous.type in CLOSING_BRACKETS)
+    ):
+        return DOT_PRIORITY
+
+    if (
+        leaf.type in MATH_OPERATORS
+        and leaf.parent
+        and leaf.parent.type not in {syms.factor, syms.star_expr}
+    ):
+        return MATH_PRIORITIES[leaf.type]
 
     if leaf.type in COMPARATORS:
         return COMPARATOR_PRIORITY
 
     if (
-        leaf.type in MATH_OPERATORS and
-        leaf.parent and
-        leaf.parent.type not in {syms.factor, syms.star_expr}
+        leaf.type == token.STRING
+        and previous is not None
+        and previous.type == token.STRING
+    ):
+        return STRING_PRIORITY
+
+    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
+    ):
+        if (
+            not isinstance(leaf.prev_sibling, Leaf)
+            or leaf.prev_sibling.value != "async"
+        ):
+            return COMPREHENSION_PRIORITY
+
+    if (
+        leaf.value == "if"
+        and leaf.parent
+        and leaf.parent.type in {syms.comp_if, syms.old_comp_if}
+    ):
+        return COMPREHENSION_PRIORITY
+
+    if leaf.value in {"if", "else"} and leaf.parent and leaf.parent.type == syms.test:
+        return TERNARY_PRIORITY
+
+    if leaf.value == "is":
+        return COMPARATOR_PRIORITY
+
+    if (
+        leaf.value == "in"
+        and leaf.parent
+        and leaf.parent.type in {syms.comp_op, syms.comparison}
+        and not (
+            previous is not None
+            and previous.type == token.NAME
+            and previous.value == "not"
+        )
+    ):
+        return COMPARATOR_PRIORITY
+
+    if (
+        leaf.value == "not"
+        and leaf.parent
+        and leaf.parent.type == syms.comp_op
+        and not (
+            previous is not None
+            and previous.type == token.NAME
+            and previous.value == "is"
+        )
     ):
-        return MATH_PRIORITY
+        return COMPARATOR_PRIORITY
+
+    if leaf.value in LOGIC_OPERATORS and leaf.parent:
+        return LOGIC_PRIORITY
 
     return 0
 
 
-def generate_comments(leaf: Leaf) -> Iterator[Leaf]:
-    """Cleans the prefix of the `leaf` and generates comments from it, if any.
+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.
 
     Comments in lib2to3 are shoved into the whitespace prefix.  This happens
     in `pgen2/driver.py:Driver.parse_tokens()`.  This was a brilliant implementation
@@ -1084,74 +2323,156 @@ def generate_comments(leaf: Leaf) -> Iterator[Leaf]:
     Inline comments are emitted as regular token.COMMENT leaves.  Standalone
     are emitted with a fake STANDALONE_COMMENT token identifier.
     """
-    if not leaf.prefix:
-        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 leaf.prefix:
-        return
 
-    before_comment, content = leaf.prefix.split('#', 1)
-    content = content.rstrip()
-    if content and (content[0] not in {' ', '!', '#'}):
-        content = ' ' + content
-    is_standalone_comment = (
-        '\n' in before_comment or '\n' in content or leaf.type == token.DEDENT
-    )
-    if not is_standalone_comment:
-        # simple trailing comment
-        yield Leaf(token.COMMENT, value='#' + content)
-        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
+
 
-    for line in ('#' + content).split('\n'):
+@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
+    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.startswith('#'):
+        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
 
-        yield Leaf(STANDALONE_COMMENT, line)
+        if index == ignored_lines and not is_endmarker:
+            comment_type = token.COMMENT  # simple trailing comment
+        else:
+            comment_type = STANDALONE_COMMENT
+        comment = make_comment(line)
+        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.
+
+    If `content` didn't start with a hash sign, one is provided.
+    """
+    content = content.rstrip()
+    if not content:
+        return "#"
+
+    if content[0] == "#":
+        content = content[1:]
+    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]:
-    """Splits a `line` into potentially many lines.
+    """Split a `line` into potentially many lines.
 
     They should fit in the allotted `line_length` but might not be able to.
     `inner` signifies that there were a pair of brackets somewhere around the
     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.
     """
-    line_str = str(line).strip('\n')
-    if len(line_str) <= line_length and '\n' not in line_str:
+    if line.is_comment:
+        yield line
+        return
+
+    line_str = str(line).strip("\n")
+
+    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
 
+    split_funcs: List[SplitFunc]
     if line.is_def:
         split_funcs = [left_hand_split]
-    elif line.inside_brackets:
-        split_funcs = [delimiter_split]
-        if '\n' not in line_str:
-            # Only attempt RHS if we don't have multiline strings or comments
-            # on this line.
-            split_funcs.append(right_hand_split)
     else:
-        split_funcs = [right_hand_split]
+
+        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, features, omit=omit))
+                if is_line_short_enough(lines[0], line_length=line_length):
+                    yield from lines
+                    return
+
+            # 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.
+            # 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]
+        else:
+            split_funcs = [rhs]
     for split_func in split_funcs:
         # We are accumulating lines in `result` because we might want to abort
         # mission and return the original line in the end, or attempt a different
         # split altogether.
         result: List[Line] = []
         try:
-            for l in split_func(line, py36=py36):
-                if str(l).strip('\n') == line_str:
+            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:
@@ -1162,25 +2483,23 @@ 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.
+    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 and
-            leaf.type in CLOSING_BRACKETS and
-            leaf.opening_bracket is matching_bracket
+            current_leaves is body_leaves
+            and leaf.type in CLOSING_BRACKETS
+            and leaf.opening_bracket is matching_bracket
         ):
             current_leaves = tail_leaves if body_leaves else head_leaves
         current_leaves.append(leaf)
@@ -1188,65 +2507,121 @@ 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])
-    # 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)
-            comment_after = line.comments.get(id(leaf))
-            if comment_after:
-                result.append(comment_after, preformatted=True)
-    split_succeeded_or_raise(head, body, tail)
+    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:
             yield result
 
 
-def right_hand_split(line: Line, py36: bool = False) -> Iterator[Line]:
-    """Split line into many lines, starting with the last matching bracket pair."""
-    head = Line(depth=line.depth)
-    body = Line(depth=line.depth + 1, inside_brackets=True)
-    tail = Line(depth=line.depth)
+def right_hand_split(
+    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.
+
+    If the split was by optional parentheses, attempt splitting without them, too.
+    `omit` is a collection of closing bracket IDs that shouldn't be considered for
+    this split.
+
+    Note: running this function modifies `bracket_depth` on the leaves of `line`.
+    """
     tail_leaves: List[Leaf] = []
     body_leaves: List[Leaf] = []
     head_leaves: List[Leaf] = []
     current_leaves = tail_leaves
-    opening_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:
                 current_leaves = head_leaves if body_leaves else tail_leaves
         current_leaves.append(leaf)
         if current_leaves is tail_leaves:
-            if leaf.type in CLOSING_BRACKETS:
+            if leaf.type in CLOSING_BRACKETS and id(leaf) not in omit:
                 opening_bracket = leaf.opening_bracket
+                closing_bracket = leaf
                 current_leaves = body_leaves
+    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")
+
     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])
-    # Build the new lines.
-    for result, leaves in (
-        (head, head_leaves), (body, body_leaves), (tail, tail_leaves)
+    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)
+    if (
+        # the body shouldn't be exploded
+        not body.should_explode
+        # the opening bracket is an optional paren
+        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
+        # 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)
     ):
-        for leaf in leaves:
-            result.append(leaf, preformatted=True)
-            comment_after = line.comments.get(id(leaf))
-            if comment_after:
-                result.append(comment_after, preformatted=True)
-    split_succeeded_or_raise(head, body, tail)
+        omit = {id(closing_bracket), *omit}
+        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)
     for result in (head, body, tail):
         if result:
             yield result
 
 
-def split_succeeded_or_raise(head: Line, body: Line, tail: Line) -> None:
+def bracket_split_succeeded_or_raise(head: Line, body: Line, tail: Line) -> None:
+    """Raise :exc:`CannotSplit` if the last left- or right-hand split failed.
+
+    Do nothing otherwise.
+
+    A left- or right-hand split is based on a pair of brackets. Content before
+    (and including) the opening bracket is left on one line, content inside the
+    brackets is put on a separate line, and finally content starting with and
+    following the closing bracket is put on a separate line.
+
+    Those are called `head`, `body`, and `tail`, respectively. If the split
+    produced the same line (all content in `head`) or ended up with an empty `body`
+    and the `tail` is just the closing bracket, then it's considered failed.
+    """
     tail_len = len(str(tail).strip())
     if not body:
         if tail_len == 0:
@@ -1259,199 +2634,1125 @@ def split_succeeded_or_raise(head: Line, body: Line, tail: Line) -> None:
             )
 
 
-def delimiter_split(line: Line, py36: bool = False) -> Iterator[Line]:
+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`.
+
+    This is a decorator over relevant split functions.
+    """
+
+    @wraps(split_func)
+    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
+
+    return split_wrapper
+
+
+@dont_increase_indentation
+def delimiter_split(line: Line, features: Collection[Feature] = ()) -> Iterator[Line]:
     """Split according to delimiters of the highest priority.
 
-    This kind of split doesn't increase indentation.
-    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]
     except IndexError:
         raise CannotSplit("Line empty")
 
-    delimiters = line.bracket_tracker.delimiters
+    bt = line.bracket_tracker
     try:
-        delimiter_priority = line.bracket_tracker.max_priority(exclude={id(last_leaf)})
+        delimiter_priority = bt.max_delimiter_priority(exclude={id(last_leaf)})
     except ValueError:
         raise CannotSplit("No delimiters found")
 
+    if delimiter_priority == DOT_PRIORITY:
+        if bt.delimiter_count_with_priority(delimiter_priority) == 1:
+            raise CannotSplit("Splitting a single attribute from its owner looks wrong")
+
     current_line = Line(depth=line.depth, inside_brackets=line.inside_brackets)
     lowest_depth = sys.maxsize
     trailing_comma_safe = True
+
+    def append_to_line(leaf: Leaf) -> Iterator[Line]:
+        """Append `leaf` to current line or to new line if appending impossible."""
+        nonlocal current_line
+        try:
+            current_line.append_safe(leaf, preformatted=True)
+        except ValueError:
+            yield current_line
+
+            current_line = Line(depth=line.depth, inside_brackets=line.inside_brackets)
+            current_line.append(leaf)
+
     for leaf in line.leaves:
-        current_line.append(leaf, preformatted=True)
-        comment_after = line.comments.get(id(leaf))
-        if comment_after:
-            current_line.append(comment_after, preformatted=True)
+        yield from append_to_line(leaf)
+
+        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
-            leaf.type == token.STAR or
-            leaf.type == token.DOUBLESTAR
-        ):
-            trailing_comma_safe = trailing_comma_safe and py36
-        leaf_priority = delimiters.get(id(leaf))
+        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:
-            normalize_prefix(current_line.leaves[0])
             yield current_line
 
             current_line = Line(depth=line.depth, inside_brackets=line.inside_brackets)
     if current_line:
         if (
-            delimiter_priority == COMMA_PRIORITY and
-            current_line.leaves[-1].type != token.COMMA and
             trailing_comma_safe
+            and delimiter_priority == COMMA_PRIORITY
+            and current_line.leaves[-1].type != token.COMMA
+            and current_line.leaves[-1].type != STANDALONE_COMMENT
         ):
-            current_line.append(Leaf(token.COMMA, ','))
-        normalize_prefix(current_line.leaves[0])
+            current_line.append(Leaf(token.COMMA, ","))
+        yield current_line
+
+
+@dont_increase_indentation
+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")
+
+    current_line = Line(depth=line.depth, inside_brackets=line.inside_brackets)
+
+    def append_to_line(leaf: Leaf) -> Iterator[Line]:
+        """Append `leaf` to current line or to new line if appending impossible."""
+        nonlocal current_line
+        try:
+            current_line.append_safe(leaf, preformatted=True)
+        except ValueError:
+            yield current_line
+
+            current_line = Line(depth=line.depth, inside_brackets=line.inside_brackets)
+            current_line.append(leaf)
+
+    for leaf in line.leaves:
+        yield from append_to_line(leaf)
+
+        for comment_after in line.comments_after(leaf):
+            yield from append_to_line(comment_after)
+
+    if current_line:
         yield current_line
 
 
 def is_import(leaf: Leaf) -> bool:
-    """Returns True if the given leaf starts an import statement."""
+    """Return True if the given leaf starts an import statement."""
     p = leaf.parent
     t = leaf.type
     v = leaf.value
     return bool(
-        t == token.NAME and
-        (
-            (v == 'import' and p and p.type == syms.import_name) or
-            (v == 'from' and p and p.type == syms.import_from)
+        t == token.NAME
+        and (
+            (v == "import" and p and p.type == syms.import_name)
+            or (v == "from" and p and p.type == syms.import_from)
         )
     )
 
 
-def normalize_prefix(leaf: Leaf) -> None:
-    """Leave existing extra newlines for imports.  Remove everything else."""
-    if is_import(leaf):
-        spl = leaf.prefix.split('#', 1)
-        nl_count = spl[0].count('\n')
-        if len(spl) > 1:
-            # Skip one newline since it was for a standalone comment.
-            nl_count -= 1
-        leaf.prefix = '\n' * nl_count
+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.
+
+    Note: don't use backslashes for formatting or you'll lose your voting rights.
+    """
+    if not inside_brackets:
+        spl = leaf.prefix.split("#")
+        if "\\" not in spl[0]:
+            nl_count = spl[-1].count("\n")
+            if len(spl) > 1:
+                nl_count -= 1
+            leaf.prefix = "\n" * nl_count
+            return
+
+    leaf.prefix = ""
+
+
+def normalize_string_prefix(leaf: Leaf, remove_u_prefix: bool = False) -> None:
+    """Make all string prefixes lowercase.
+
+    If remove_u_prefix is given, also removes any u prefix from the string.
+
+    Note: Mutates its argument.
+    """
+    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()
+    if remove_u_prefix:
+        new_prefix = new_prefix.replace("u", "")
+    leaf.value = f"{new_prefix}{match.group(2)}"
+
+
+def normalize_string_quotes(leaf: Leaf) -> None:
+    """Prefer double quotes but only if it doesn't cause more escaping.
+
+    Adds or removes backslashes as appropriate. Doesn't parse and fix
+    strings nested in f-strings (yet).
+
+    Note: Mutates its argument.
+    """
+    value = leaf.value.lstrip("furbFURB")
+    if value[:3] == '"""':
         return
 
-    leaf.prefix = ''
+    elif value[:3] == "'''":
+        orig_quote = "'''"
+        new_quote = '"""'
+    elif value[0] == '"':
+        orig_quote = '"'
+        new_quote = "'"
+    else:
+        orig_quote = "'"
+        new_quote = '"'
+    first_quote_pos = leaf.value.find(orig_quote)
+    if first_quote_pos == -1:
+        return  # There's an internal error
+
+    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}")
+    body = leaf.value[first_quote_pos + len(orig_quote) : -len(orig_quote)]
+    if "r" in prefix.casefold():
+        if unescaped_new_quote.search(body):
+            # There's at least one unescaped new_quote in this raw string
+            # so converting is impossible
+            return
+
+        # Do not introduce or remove backslashes in raw strings
+        new_body = body
+    else:
+        # 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 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 "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("\\")
+    new_escape_count = new_body.count("\\")
+    if new_escape_count > orig_escape_count:
+        return  # Do not introduce more escaping
+
+    if new_escape_count == orig_escape_count and orig_quote == '"':
+        return  # Prefer double quotes
+
+    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 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.
+    """
+    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:
+                if maybe_make_parens_invisible_in_atom(child, parent=node):
+                    wrap_in_parentheses(node, child, visible=False)
+            elif is_one_tuple(child):
+                wrap_in_parentheses(node, child, visible=True)
+            elif node.type == syms.import_from:
+                # "import from" nodes store parentheses directly as part of
+                # the statement
+                if child.type == token.LPAR:
+                    # make parentheses invisible
+                    child.value = ""  # type: ignore
+                    node.children[-1].value = ""  # type: ignore
+                elif child.type != token.STAR:
+                    # insert invisible parentheses
+                    node.insert_child(index, Leaf(token.LPAR, ""))
+                    node.append_child(Leaf(token.RPAR, ""))
+                break
+
+            elif not (isinstance(child, Leaf) and is_multiline_string(child)):
+                wrap_in_parentheses(node, child, visible=False)
+
+        check_lpar = isinstance(child, Leaf) and child.value in parens_after
+
+
+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) and parent.type != syms.expr_stmt)
+        or max_delimiter_priority_in_atom(node) >= COMMA_PRIORITY
+    ):
+        return False
+
+    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
+        maybe_make_parens_invisible_in_atom(middle, parent=parent)
+
+        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:
+    """Return True if `node` holds an empty tuple."""
+    return (
+        node.type == syms.atom
+        and len(node.children) == 2
+        and node.children[0].type == token.LPAR
+        and node.children[1].type == token.RPAR
+    )
+
+
+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:
+        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
+
+    return (
+        node.type in IMPLICIT_TUPLE
+        and len(node.children) == 2
+        and node.children[1].type == token.COMMA
+    )
+
+
+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:
+        return True
+
+    if node.type == token.NAME and node.value == "yield":  # type: ignore
+        return True
+
+    if node.type != syms.atom:
+        return False
+
+    if len(node.children) != 3:
+        return False
+
+    lpar, expr, rpar = node.children
+    if lpar.type == token.LPAR and rpar.type == token.RPAR:
+        return is_yield(expr)
+
+    return False
+
+
+def is_vararg(leaf: Leaf, within: Set[NodeType]) -> bool:
+    """Return True if `leaf` is a star or double star in a vararg or kwarg.
+
+    If `within` includes VARARGS_PARENTS, this applies to function signatures.
+    If `within` includes UNPACKING_PARENTS, it applies to right hand-side
+    extended iterable unpacking (PEP 3132) and additional unpacking
+    generalizations (PEP 448).
+    """
+    if leaf.type not in VARARGS_SPECIALS or not leaf.parent:
+        return False
+
+    p = leaf.parent
+    if p.type == syms.star_expr:
+        # Star expressions are also used as assignment targets in extended
+        # iterable unpacking (PEP 3132).  See what its parent is instead.
+        if not p.parent:
+            return False
+
+        p = p.parent
+
+    return p.type in within
+
+
+def is_multiline_string(leaf: Leaf) -> bool:
+    """Return True if `leaf` is a multiline string that actually spans many lines."""
+    value = leaf.value.lstrip("furbFURB")
+    return value[:3] in {'"""', "'''"} and "\n" in value
+
+
+def is_stub_suite(node: Node) -> bool:
+    """Return True if `node` is a suite with a stub body."""
+    if (
+        len(node.children) != 4
+        or node.children[0].type != token.NEWLINE
+        or node.children[1].type != token.INDENT
+        or node.children[3].type != token.DEDENT
+    ):
+        return False
+
+    return is_stub_body(node.children[2])
+
+
+def is_stub_body(node: LN) -> bool:
+    """Return True if `node` is a simple statement containing an ellipsis."""
+    if not isinstance(node, Node) or node.type != syms.simple_stmt:
+        return False
+
+    if len(node.children) != 2:
+        return False
+
+    child = node.children[0]
+    return (
+        child.type == syms.atom
+        and len(child.children) == 3
+        and all(leaf == Leaf(token.DOT, ".") for leaf in child.children)
+    )
+
+
+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.
+    If `node` isn't an atom or there are no enclosing parentheses, returns 0.
+    """
+    if node.type != syms.atom:
+        return 0
+
+    first = node.children[0]
+    last = node.children[-1]
+    if not (first.type == token.LPAR and last.type == token.RPAR):
+        return 0
+
+    bt = BracketTracker()
+    for c in node.children[1:-1]:
+        if isinstance(c, Leaf):
+            bt.mark(c)
+        else:
+            for leaf in c.leaves():
+                bt.mark(leaf)
+    try:
+        return bt.max_delimiter_priority()
+
+    except ValueError:
+        return 0
+
+
+def ensure_visible(leaf: Leaf) -> None:
+    """Make sure parentheses are visible.
+
+    They could be invisible as part of some statements (see
+    :func:`normalize_invisible_parens` and :func:`visit_import_from`).
+    """
+    if leaf.type == token.LPAR:
+        leaf.value = "("
+    elif leaf.type == token.RPAR:
+        leaf.value = ")"
 
 
-def is_python36(node: Node) -> bool:
-    """Returns True if the current file is using Python 3.6+ features.
+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}
+        and opening_bracket.value in "[{("
+    ):
+        return False
+
+    try:
+        last_leaf = line.leaves[-1]
+        exclude = {id(last_leaf)} if last_leaf.type == token.COMMA else set()
+        max_priority = line.bracket_tracker.max_delimiter_priority(exclude=exclude)
+    except (IndexError, ValueError):
+        return False
+
+    return max_priority == COMMA_PRIORITY
+
+
+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.
+    - 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
+            if value_head in {'f"', 'F"', "f'", "F'", "rf", "fr", "RF", "FR"}:
+                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 == syms.typedargslist and
-            n.children and
-            n.children[-1].type == token.COMMA
+            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 == token.STAR or ch.type == token.DOUBLESTAR:
-                    return True
+                if ch.type in STARS:
+                    features.add(feature)
 
-    return False
+                if ch.type == syms.argument:
+                    for argch in ch.children:
+                        if argch.type in STARS:
+                            features.add(feature)
 
+    return features
 
-PYTHON_EXTENSIONS = {'.py'}
-BLACKLISTED_DIRECTORIES = {
-    'build', 'buck-out', 'dist', '_build', '.git', '.hg', '.mypy_cache', '.tox', '.venv'
-}
 
+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]]:
+    """Generate sets of closing bracket IDs that should be omitted in a RHS.
+
+    Brackets can be omitted if the entire trailer up to and including
+    a preceding closing bracket fits in one line.
+
+    Yielded sets are cumulative (contain results of previous yields, too).  First
+    set is empty.
+    """
+
+    omit: Set[LeafID] = set()
+    yield omit
+
+    length = 4 * line.depth
+    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
+        if length > line_length:
+            break
+
+        has_inline_comment = leaf_length > len(leaf.value) + len(leaf.prefix)
+        if leaf.type == STANDALONE_COMMENT or has_inline_comment:
+            break
+
+        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 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
+                # pair of brackets was good enough.
+                inner_brackets.add(id(leaf))
+                continue
+
+            if closing_bracket:
+                omit.add(id(closing_bracket))
+                omit.update(inner_brackets)
+                inner_brackets.clear()
+                yield omit
+
+            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[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.
+            if (
+                len(child.children) == 2
+                and first_child.type == token.STRING
+                and child.children[1].type == token.NEWLINE
+            ):
+                continue
+
+            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
+
+            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],
+    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.
 
-def gen_python_files_in_dir(path: Path) -> Iterator[Path]:
+    `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():
-        if child.is_dir():
-            if child.name in BLACKLISTED_DIRECTORIES:
+        # First ignore files matching .gitignore
+        if gitignore.match_file(child.as_posix()):
+            report.path_ignored(child, f"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
 
-            yield from gen_python_files_in_dir(child)
+            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, f"matches the --exclude regular expression")
+            continue
+
+        if child.is_dir():
+            yield from gen_python_files_in_dir(
+                child, root, include, exclude, report, gitignore
+            )
+
+        elif child.is_file():
+            include_match = include.search(normalized_path)
+            if include_match:
+                yield child
+
+
+@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
+    common parent.
+
+    If no directory in the tree contains a marker that would specify it's the
+    project root, the root of the file system is returned.
+    """
+    if not srcs:
+        return Path("/").resolve()
+
+    common_base = min(Path(src).resolve() for src in srcs)
+    if common_base.is_dir():
+        # 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").exists():
+            return directory
+
+        if (directory / ".hg").is_dir():
+            return directory
+
+        if (directory / "pyproject.toml").is_file():
+            return directory
 
-        elif child.suffix in PYTHON_EXTENSIONS:
-            yield child
+    return directory
 
 
 @dataclass
 class Report:
-    """Provides a reformatting counter."""
-    change_count: int = attrib(default=0)
-    same_count: int = attrib(default=0)
-    failure_count: int = attrib(default=0)
+    """Provides a reformatting counter. Can be rendered with `str(report)`."""
 
-    def done(self, src: Path, changed: bool) -> None:
+    check: bool = False
+    diff: bool = False
+    quiet: bool = False
+    verbose: bool = False
+    change_count: int = 0
+    same_count: int = 0
+    failure_count: int = 0
+
+    def done(self, src: Path, changed: Changed) -> None:
         """Increment the counter for successful reformatting. Write out a message."""
-        if changed:
-            out(f'reformatted {src}')
+        if changed is Changed.YES:
+            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:
-            out(f'{src} already well formatted, good job.', bold=False)
+            if self.verbose:
+                if changed is Changed.NO:
+                    msg = f"{src} already well formatted, good job."
+                else:
+                    msg = f"{src} wasn't modified on disk since last run."
+                out(msg, bold=False)
             self.same_count += 1
 
     def failed(self, src: Path, message: str) -> None:
         """Increment the counter for failed reformatting. Write out a message."""
-        err(f'error: cannot format {src}: {message}')
+        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:
-        """Which return code should the app use considering the current state."""
+        """Return the exit code that the app should use.
+
+        This considers the current state of changed files and failures:
+        - if there were any failures, return 123;
+        - if any files were changed and --check is being used, return 1;
+        - 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
 
-        elif self.change_count:
+        elif self.change_count and self.check:
             return 1
 
         return 0
 
     def __str__(self) -> str:
-        """A color report of the current state.
+        """Render a color report of the current state.
 
         Use `click.unstyle` to remove colors.
         """
+        if self.check or self.diff:
+            reformatted = "would be reformatted"
+            unchanged = "would be left unchanged"
+            failed = "would fail to reformat"
+        else:
+            reformatted = "reformatted"
+            unchanged = "left unchanged"
+            failed = "failed to reformat"
         report = []
         if self.change_count:
-            s = 's' if self.change_count > 1 else ''
+            s = "s" if self.change_count > 1 else ""
             report.append(
-                click.style(f'{self.change_count} file{s} reformatted', bold=True)
+                click.style(f"{self.change_count} file{s} {reformatted}", bold=True)
             )
         if self.same_count:
-            s = 's' if self.same_count > 1 else ''
-            report.append(f'{self.same_count} file{s} left unchanged')
+            s = "s" if self.same_count > 1 else ""
+            report.append(f"{self.same_count} file{s} {unchanged}")
         if self.failure_count:
-            s = 's' if self.failure_count > 1 else ''
+            s = "s" if self.failure_count > 1 else ""
             report.append(
-                click.style(
-                    f'{self.failure_count} file{s} failed to reformat', fg='red'
-                )
+                click.style(f"{self.failure_count} file{s} {failed}", fg="red")
             )
-        return ', '.join(report) + '.'
+        return ", ".join(report) + "."
 
 
-def assert_equivalent(src: str, dst: str) -> None:
-    """Raises AssertionError if `src` and `dst` aren't equivalent.
+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
 
-    This is a temporary sanity check until Black becomes stable.
-    """
+    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)
 
-    import ast
-    import traceback
+    if isinstance(node, (ast.NameConstant, ast3.NameConstant)):
+        return ast.Constant(value=node.value)
 
-    def _v(node: ast.AST, depth: int = 0) -> Iterator[str]:
+    return node
+
+
+def assert_equivalent(src: str, dst: str) -> None:
+    """Raise AssertionError if `src` and `dst` aren't equivalent."""
+
+    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:
@@ -1461,10 +3762,20 @@ 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:
@@ -1473,74 +3784,390 @@ 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:
-        raise AssertionError(f"cannot parse source: {exc}") from None
+        raise AssertionError(
+            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)
+        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
 
-    src_ast_str = '\n'.join(_v(src_ast))
-    dst_ast_str = '\n'.join(_v(dst_ast))
+    src_ast_str = "\n".join(_v(src_ast))
+    dst_ast_str = "\n".join(_v(dst_ast))
     if src_ast_str != dst_ast_str:
-        log = dump_to_file(diff(src_ast_str, dst_ast_str, 'src', 'dst'))
+        log = dump_to_file(diff(src_ast_str, dst_ast_str, "src", "dst"))
         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) -> None:
-    """Raises AssertionError if `dst` reformats differently the second time.
-
-    This is a temporary sanity check until Black becomes stable.
-    """
-    newdst = format_str(dst, line_length=line_length)
+def assert_stable(src: str, dst: str, mode: FileMode) -> None:
+    """Raise AssertionError if `dst` reformats differently the second time."""
+    newdst = format_str(dst, mode=mode)
     if dst != newdst:
         log = dump_to_file(
-            diff(src, dst, 'source', 'first pass'),
-            diff(dst, newdst, 'first pass', 'second pass'),
+            diff(src, dst, "source", "first pass"),
+            diff(dst, newdst, "first pass", "second pass"),
         )
         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:
-    """Dumps `output` to a temporary file. Returns path to the file."""
-    import tempfile
-
+    """Dump `output` to a temporary file. Return path to the file."""
     with tempfile.NamedTemporaryFile(
-        mode='w', prefix='blk_', suffix='.log', delete=False
+        mode="w", prefix="blk_", suffix=".log", delete=False, encoding="utf8"
     ) as f:
         for lines in output:
             f.write(lines)
-            f.write('\n')
+            if lines and lines[-1] != "\n":
+                f.write("\n")
     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:
-    """Returns a udiff string between strings `a` and `b`."""
+    """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')]
-    return ''.join(
+    a_lines = [line + "\n" for line in a.split("\n")]
+    b_lines = [line + "\n" for line in b.split("\n")]
+    return "".join(
         difflib.unified_diff(a_lines, b_lines, fromfile=a_name, tofile=b_name, n=5)
     )
 
 
-if __name__ == '__main__':
+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: 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 all_tasks(loop) if not task.done()]
+        if not to_cancel:
+            return
+
+        for task in to_cancel:
+            task.cancel()
+        loop.run_until_complete(
+            asyncio.gather(*to_cancel, loop=loop, return_exceptions=True)
+        )
+    finally:
+        # `concurrent.futures.Future` objects cannot be cancelled once they
+        # are already running. There might be some when the `shutdown()` happened.
+        # Silence their logger's spew about the event loop being closed.
+        cf_logger = logging.getLogger("concurrent.futures")
+        cf_logger.setLevel(logging.CRITICAL)
+        loop.close()
+
+
+def sub_twice(regex: Pattern[str], replacement: str, original: str) -> str:
+    """Replace `regex` with `replacement` twice on `original`.
+
+    This is used by string normalization to perform replaces on
+    overlapping matches.
+    """
+    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
+    for element in reversed(sequence):
+        yield (index, element)
+        index -= 1
+
+
+def enumerate_with_length(
+    line: Line, reversed: bool = False
+) -> Iterator[Tuple[Index, Leaf, int]]:
+    """Return an enumeration of leaves with their length.
+
+    Stops prematurely on multiline strings and standalone comments.
+    """
+    op = cast(
+        Callable[[Sequence[Leaf]], Iterator[Tuple[Index, Leaf]]],
+        enumerate_reversed if reversed else enumerate,
+    )
+    for index, leaf in op(line.leaves):
+        length = len(leaf.prefix) + len(leaf.value)
+        if "\n" in leaf.value:
+            return  # Multiline strings, we can't continue.
+
+        for comment in line.comments_after(leaf):
+            length += len(comment.value)
+
+        yield index, leaf, length
+
+
+def is_line_short_enough(line: Line, *, line_length: int, line_str: str = "") -> bool:
+    """Return True if `line` is no longer than `line_length`.
+
+    Uses the provided `line_str` rendering, if any, otherwise computes a new one.
+    """
+    if not line_str:
+        line_str = str(line).strip("\n")
+    return (
+        len(line_str) <= line_length
+        and "\n" not in line_str  # multiline strings
+        and not line.contains_standalone_comments()
+    )
+
+
+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?
+
+    Returns True for only a subset of potentially nice looking formattings but
+    the point is to not return false positives that end up producing lines that
+    are too long.
+    """
+    bt = line.bracket_tracker
+    if not bt.delimiters:
+        # Without delimiters the optional parentheses are useless.
+        return True
+
+    max_priority = bt.max_delimiter_priority()
+    if bt.delimiter_count_with_priority(max_priority) > 1:
+        # With more than one delimiter of a kind the optional parentheses read better.
+        return False
+
+    if max_priority == DOT_PRIORITY:
+        # A single stranded method call doesn't require optional parentheses.
+        return True
+
+    assert len(line.leaves) >= 2, "Stranded delimiter"
+
+    first = line.leaves[0]
+    second = line.leaves[1]
+    penultimate = line.leaves[-2]
+    last = line.leaves[-1]
+
+    # With a single delimiter, omit if the expression starts or ends with
+    # a bracket.
+    if first.type in OPENING_BRACKETS and second.type not in CLOSING_BRACKETS:
+        remainder = False
+        length = 4 * line.depth
+        for _index, leaf, leaf_length in enumerate_with_length(line):
+            if leaf.type in CLOSING_BRACKETS and leaf.opening_bracket is first:
+                remainder = True
+            if remainder:
+                length += leaf_length
+                if length > line_length:
+                    break
+
+                if leaf.type in OPENING_BRACKETS:
+                    # There are brackets we can further split on.
+                    remainder = False
+
+        else:
+            # checked the entire string and line length wasn't exceeded
+            if len(line.leaves) == _index + 1:
+                return True
+
+        # Note: we are not returning False here because a line might have *both*
+        # a leading opening bracket and a trailing closing bracket.  If the
+        # opening bracket doesn't match our rule, maybe the closing will.
+
+    if (
+        last.type == token.RPAR
+        or last.type == token.RBRACE
+        or (
+            # don't use indexing for omitting optional parentheses;
+            # it looks weird
+            last.type == token.RSQB
+            and last.parent
+            and last.parent.type != syms.trailer
+        )
+    ):
+        if penultimate.type in OPENING_BRACKETS:
+            # Empty brackets don't help.
+            return False
+
+        if is_multiline_string(first):
+            # Additional wrapping of a multiline string in this situation is
+            # unnecessary.
+            return True
+
+        length = 4 * line.depth
+        seen_other_brackets = False
+        for _index, leaf, leaf_length in enumerate_with_length(line):
+            length += leaf_length
+            if leaf is last.opening_bracket:
+                if seen_other_brackets or length <= line_length:
+                    return True
+
+            elif leaf.type in OPENING_BRACKETS:
+                # There are brackets we can further split on.
+                seen_other_brackets = True
+
+    return False
+
+
+def get_cache_file(mode: FileMode) -> Path:
+    return CACHE_DIR / f"cache.{mode.get_cache_key()}.pickle"
+
+
+def read_cache(mode: FileMode) -> 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(mode)
+    if not cache_file.exists():
+        return {}
+
+    with cache_file.open("rb") as fobj:
+        try:
+            cache: Cache = pickle.load(fobj)
+        except (pickle.UnpicklingError, ValueError):
+            return {}
+
+    return cache
+
+
+def get_cache_info(path: Path) -> CacheInfo:
+    """Return the information used to check if a file is already formatted or not."""
+    stat = path.stat()
+    return stat.st_mtime, stat.st_size
+
+
+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 contains paths of files that modified on disk or are not in the
+    cache. The other contains paths to non-modified files.
+    """
+    todo, done = set(), set()
+    for src in sources:
+        src = src.resolve()
+        if cache.get(src) != get_cache_info(src):
+            todo.add(src)
+        else:
+            done.add(src)
+    return todo, done
+
+
+def write_cache(cache: Cache, sources: Iterable[Path], mode: FileMode) -> None:
+    """Update the cache file."""
+    cache_file = get_cache_file(mode)
+    try:
+        CACHE_DIR.mkdir(parents=True, exist_ok=True)
+        new_cache = {**cache, **{src.resolve(): get_cache_info(src) for src in sources}}
+        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
+
+
+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()