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

madduck's git repository

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

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

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

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

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

Add underscores to numeric literals with more than six digits (#529)
[etc/vim.git] / black.py
index 9fbacc1048cd15fd7816ccdfb8dd3f6dd7252b9c..56765313aa85b0571c7831f209573bd2b3fbb955 100644 (file)
--- a/black.py
+++ b/black.py
@@ -1,23 +1,26 @@
 import asyncio
-import pickle
 from asyncio.base_events import BaseEventLoop
 from concurrent.futures import Executor, ProcessPoolExecutor
-from enum import Enum
-from functools import partial, wraps
+from datetime import datetime
+from enum import Enum, Flag
+from functools import lru_cache, partial, wraps
+import io
 import keyword
 import logging
 from multiprocessing import Manager
 import os
 from pathlib import Path
+import pickle
 import re
-import tokenize
 import signal
 import sys
+import tokenize
 from typing import (
     Any,
     Callable,
     Collection,
     Dict,
+    Generator,
     Generic,
     Iterable,
     Iterator,
@@ -27,7 +30,6 @@ from typing import (
     Sequence,
     Set,
     Tuple,
-    Type,
     TypeVar,
     Union,
     cast,
@@ -36,6 +38,7 @@ from typing import (
 from appdirs import user_cache_dir
 from attr import dataclass, Factory
 import click
+import toml
 
 # lib2to3 fork
 from blib2to3.pytree import Node, Leaf, type_repr
@@ -44,14 +47,19 @@ from blib2to3.pgen2 import driver, token
 from blib2to3.pgen2.parse import ParseError
 
 
-__version__ = "18.5b0"
+__version__ = "18.6b4"
 DEFAULT_LINE_LENGTH = 88
+DEFAULT_EXCLUDES = (
+    r"/(\.git|\.hg|\.mypy_cache|\.nox|\.tox|\.venv|_build|buck-out|build|dist)/"
+)
+DEFAULT_INCLUDES = r"\.pyi?$"
 CACHE_DIR = Path(user_cache_dir("black", version=__version__))
 
 
 # types
 FileContent = str
 Encoding = str
+NewLine = str
 Depth = int
 NodeType = int
 LeafID = int
@@ -71,49 +79,29 @@ syms = pygram.python_symbols
 
 
 class NothingChanged(UserWarning):
-    """Raised by :func:`format_file` when reformatted code is the same as source."""
+    """Raised when reformatted code is the same as source."""
 
 
 class CannotSplit(Exception):
-    """A readable split that fits the allotted line length is impossible.
-
-    Raised by :func:`left_hand_split`, :func:`right_hand_split`, and
-    :func:`delimiter_split`.
-    """
-
-
-class FormatError(Exception):
-    """Base exception for `# fmt: on` and `# fmt: off` handling.
-
-    It holds the number of bytes of the prefix consumed before the format
-    control comment appeared.
-    """
-
-    def __init__(self, consumed: int) -> None:
-        super().__init__(consumed)
-        self.consumed = consumed
-
-    def trim_prefix(self, leaf: Leaf) -> None:
-        leaf.prefix = leaf.prefix[self.consumed :]
+    """A readable split that fits the allotted line length is impossible."""
 
-    def leaf_from_consumed(self, leaf: Leaf) -> Leaf:
-        """Returns a new Leaf from the consumed part of the prefix."""
-        unformatted_prefix = leaf.prefix[: self.consumed]
-        return Leaf(token.NEWLINE, unformatted_prefix)
 
-
-class FormatOn(FormatError):
-    """Found a comment like `# fmt: on` in the file."""
-
-
-class FormatOff(FormatError):
-    """Found a comment like `# fmt: off` in the file."""
+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):
@@ -122,15 +110,106 @@ class Changed(Enum):
     YES = 2
 
 
-@click.command()
+class FileMode(Flag):
+    AUTO_DETECT = 0
+    PYTHON36 = 1
+    PYI = 2
+    NO_STRING_NORMALIZATION = 4
+    NO_NUMERIC_UNDERSCORE_NORMALIZATION = 8
+
+    @classmethod
+    def from_configuration(
+        cls,
+        *,
+        py36: bool,
+        pyi: bool,
+        skip_string_normalization: bool,
+        skip_numeric_underscore_normalization: bool,
+    ) -> "FileMode":
+        mode = cls.AUTO_DETECT
+        if py36:
+            mode |= cls.PYTHON36
+        if pyi:
+            mode |= cls.PYI
+        if skip_string_normalization:
+            mode |= cls.NO_STRING_NORMALIZATION
+        if skip_numeric_underscore_normalization:
+            mode |= cls.NO_NUMERIC_UNDERSCORE_NORMALIZATION
+        return mode
+
+
+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:
+        root = find_project_root(ctx.params.get("src", ()))
+        path = root / "pyproject.toml"
+        if path.is_file():
+            value = str(path)
+        else:
+            return None
+
+    try:
+        pyproject_toml = toml.load(value)
+        config = pyproject_toml.get("tool", {}).get("black", {})
+    except (toml.TomlDecodeError, OSError) as e:
+        raise click.BadOptionUsage(f"Error reading configuration file: {e}", ctx)
+
+    if not config:
+        return None
+
+    if ctx.default_map is None:
+        ctx.default_map = {}
+    ctx.default_map.update(  # type: ignore  # bad types in .pyi
+        {k.replace("--", "").replace("-", "_"): v for k, v in config.items()}
+    )
+    return value
+
+
+@click.command(context_settings=dict(help_option_names=["-h", "--help"]))
 @click.option(
     "-l",
     "--line-length",
     type=int,
     default=DEFAULT_LINE_LENGTH,
-    help="How many character per line to allow.",
+    help="How many characters per line to allow.",
     show_default=True,
 )
+@click.option(
+    "--py36",
+    is_flag=True,
+    help=(
+        "Allow using Python 3.6-only syntax on all input files.  This will put "
+        "trailing commas in function signatures and calls also after *args and "
+        "**kwargs.  [default: per-file auto-detection]"
+    ),
+)
+@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(
+    "-N",
+    "--skip-numeric-underscore-normalization",
+    is_flag=True,
+    help="Don't normalize underscores in numeric literals.",
+)
 @click.option(
     "--check",
     is_flag=True,
@@ -150,6 +229,31 @@ class Changed(Enum):
     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",
@@ -159,6 +263,15 @@ class Changed(Enum):
         "silence those with 2>/dev/null."
     ),
 )
+@click.option(
+    "-v",
+    "--verbose",
+    is_flag=True,
+    help=(
+        "Also emit messages to stderr about files that were not changed or were "
+        "ignored due to --exclude=."
+    ),
+)
 @click.version_option(version=__version__)
 @click.argument(
     "src",
@@ -166,6 +279,16 @@ class Changed(Enum):
     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(
@@ -174,92 +297,141 @@ def main(
     check: bool,
     diff: bool,
     fast: bool,
+    pyi: bool,
+    py36: bool,
+    skip_string_normalization: bool,
+    skip_numeric_underscore_normalization: bool,
     quiet: bool,
-    src: List[str],
+    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)
+    mode = FileMode.from_configuration(
+        py36=py36,
+        pyi=pyi,
+        skip_string_normalization=skip_string_normalization,
+        skip_numeric_underscore_normalization=skip_numeric_underscore_normalization,
+    )
+    if config and verbose:
+        out(f"Using configuration from {config}.", bold=False, fg="blue")
+    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, quiet=quiet, verbose=verbose)
+    root = find_project_root(src)
+    sources: Set[Path] = set()
     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)
+            )
+        elif p.is_file() or s == "-":
             # if a file was explicitly given, we don't care about its extension
-            sources.append(p)
-        elif s == "-":
-            sources.append(Path("-"))
+            sources.add(p)
         else:
             err(f"invalid path: {s}")
-
-    if check and not diff:
-        write_back = WriteBack.NO
-    elif diff:
-        write_back = WriteBack.DIFF
-    else:
-        write_back = WriteBack.YES
-    report = Report(check=check, quiet=quiet)
     if len(sources) == 0:
-        out("No paths given. Nothing to do 😴")
+        if verbose or not quiet:
+            out("No paths given. Nothing to do 😴")
         ctx.exit(0)
-        return
 
-    elif len(sources) == 1:
-        reformat_one(sources[0], line_length, fast, write_back, report)
+    if len(sources) == 1:
+        reformat_one(
+            src=sources.pop(),
+            line_length=line_length,
+            fast=fast,
+            write_back=write_back,
+            mode=mode,
+            report=report,
+        )
     else:
         loop = asyncio.get_event_loop()
         executor = ProcessPoolExecutor(max_workers=os.cpu_count())
         try:
             loop.run_until_complete(
                 schedule_formatting(
-                    sources, line_length, fast, write_back, report, loop, executor
+                    sources=sources,
+                    line_length=line_length,
+                    fast=fast,
+                    write_back=write_back,
+                    mode=mode,
+                    report=report,
+                    loop=loop,
+                    executor=executor,
                 )
             )
         finally:
             shutdown(loop)
-        if not quiet:
-            out("All done! ✨ 🍰 ✨")
-            click.echo(str(report))
+    if verbose or not quiet:
+        bang = "💥 💔 💥" if report.return_code else "✨ 🍰 ✨"
+        out(f"All done! {bang}")
+        click.secho(str(report), err=True)
     ctx.exit(report.return_code)
 
 
 def reformat_one(
-    src: Path, line_length: int, fast: bool, write_back: WriteBack, report: "Report"
+    src: Path,
+    line_length: int,
+    fast: bool,
+    write_back: WriteBack,
+    mode: FileMode,
+    report: "Report",
 ) -> None:
     """Reformat a single file under `src` without spawning child processes.
 
     If `quiet` is True, non-error messages are not output. `line_length`,
-    `write_back`, and `fast` options are passed to :func:`format_file_in_place`.
+    `write_back`, `fast` and `pyi` options are passed to
+    :func:`format_file_in_place` or :func:`format_stdin_to_stdout`.
     """
     try:
         changed = Changed.NO
         if not src.is_file() and str(src) == "-":
             if format_stdin_to_stdout(
-                line_length=line_length, fast=fast, write_back=write_back
+                line_length=line_length, fast=fast, write_back=write_back, mode=mode
             ):
                 changed = Changed.YES
         else:
             cache: Cache = {}
             if write_back != WriteBack.DIFF:
-                cache = read_cache(line_length)
-                src = src.resolve()
-                if src in cache and cache[src] == get_cache_info(src):
+                cache = read_cache(line_length, mode)
+                res_src = src.resolve()
+                if res_src in cache and cache[res_src] == get_cache_info(res_src):
                     changed = Changed.CACHED
             if changed is not Changed.CACHED and format_file_in_place(
-                src, line_length=line_length, fast=fast, write_back=write_back
+                src,
+                line_length=line_length,
+                fast=fast,
+                write_back=write_back,
+                mode=mode,
             ):
                 changed = Changed.YES
-            if write_back == WriteBack.YES and changed is not Changed.NO:
-                write_cache(cache, [src], line_length)
+            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], line_length, mode)
         report.done(src, changed)
     except Exception as exc:
         report.failed(src, str(exc))
 
 
 async def schedule_formatting(
-    sources: List[Path],
+    sources: Set[Path],
     line_length: int,
     fast: bool,
     write_back: WriteBack,
+    mode: FileMode,
     report: "Report",
     loop: BaseEventLoop,
     executor: Executor,
@@ -268,52 +440,67 @@ async def schedule_formatting(
 
     (Use ProcessPoolExecutors for actual parallelism.)
 
-    `line_length`, `write_back`, and `fast` options are passed to
+    `line_length`, `write_back`, `fast`, and `pyi` options are passed to
     :func:`format_file_in_place`.
     """
     cache: Cache = {}
     if write_back != WriteBack.DIFF:
-        cache = read_cache(line_length)
+        cache = read_cache(line_length, mode)
         sources, cached = filter_cached(cache, sources)
-        for src in cached:
+        for src in sorted(cached):
             report.done(src, Changed.CACHED)
+    if not sources:
+        return
+
     cancelled = []
-    formatted = []
-    if sources:
-        lock = None
-        if write_back == WriteBack.DIFF:
-            # For diff output, we need locks to ensure we don't interleave output
-            # from different processes.
-            manager = Manager()
-            lock = manager.Lock()
-        tasks = {
-            loop.run_in_executor(
-                executor, format_file_in_place, src, line_length, fast, write_back, lock
-            ): src
-            for src in sorted(sources)
-        }
-        pending: Iterable[asyncio.Task] = tasks.keys()
-        try:
-            loop.add_signal_handler(signal.SIGINT, cancel, pending)
-            loop.add_signal_handler(signal.SIGTERM, cancel, pending)
-        except NotImplementedError:
-            # There are no good alternatives for these on Windows
-            pass
-        while pending:
-            done, _ = await asyncio.wait(pending, return_when=asyncio.FIRST_COMPLETED)
-            for task in done:
-                src = tasks.pop(task)
-                if task.cancelled():
-                    cancelled.append(task)
-                elif task.exception():
-                    report.failed(src, str(task.exception()))
-                else:
-                    formatted.append(src)
-                    report.done(src, Changed.YES if task.result() else Changed.NO)
+    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 = {
+        loop.run_in_executor(
+            executor,
+            format_file_in_place,
+            src,
+            line_length,
+            fast,
+            write_back,
+            mode,
+            lock,
+        ): src
+        for src in sorted(sources)
+    }
+    pending: Iterable[asyncio.Task] = tasks.keys()
+    try:
+        loop.add_signal_handler(signal.SIGINT, cancel, pending)
+        loop.add_signal_handler(signal.SIGTERM, cancel, pending)
+    except NotImplementedError:
+        # There are no good alternatives for these on Windows.
+        pass
+    while pending:
+        done, _ = await asyncio.wait(pending, return_when=asyncio.FIRST_COMPLETED)
+        for task in done:
+            src = tasks.pop(task)
+            if task.cancelled():
+                cancelled.append(task)
+            elif task.exception():
+                report.failed(src, str(task.exception()))
+            else:
+                changed = Changed.YES if task.result() else Changed.NO
+                # If the file was written back or was successfully checked as
+                # well-formatted, store this information in the cache.
+                if write_back is WriteBack.YES or (
+                    write_back is WriteBack.CHECK and changed is Changed.NO
+                ):
+                    sources_to_cache.append(src)
+                report.done(src, changed)
     if cancelled:
         await asyncio.gather(*cancelled, loop=loop, return_exceptions=True)
-    if write_back == WriteBack.YES and formatted:
-        write_cache(cache, formatted, line_length)
+    if sources_to_cache:
+        write_cache(cache, sources_to_cache, line_length, mode)
 
 
 def format_file_in_place(
@@ -321,35 +508,47 @@ def format_file_in_place(
     line_length: int,
     fast: bool,
     write_back: WriteBack = WriteBack.NO,
+    mode: FileMode = FileMode.AUTO_DETECT,
     lock: Any = None,  # multiprocessing.Manager().Lock() is some crazy proxy
 ) -> bool:
     """Format file under `src` path. Return True if changed.
 
-    If `write_back` is True, write reformatted code back to stdout.
+    If `write_back` is DIFF, write a diff to stdout. If it is YES, write reformatted
+    code to the file.
     `line_length` and `fast` options are passed to :func:`format_file_contents`.
     """
-    is_pyi = src.suffix == ".pyi"
+    if src.suffix == ".pyi":
+        mode |= FileMode.PYI
 
-    with tokenize.open(src) as src_buffer:
-        src_contents = src_buffer.read()
+    then = datetime.utcfromtimestamp(src.stat().st_mtime)
+    with open(src, "rb") as buf:
+        src_contents, encoding, newline = decode_bytes(buf.read())
     try:
         dst_contents = format_file_contents(
-            src_contents, line_length=line_length, fast=fast, is_pyi=is_pyi
+            src_contents, line_length=line_length, fast=fast, mode=mode
         )
     except NothingChanged:
         return False
 
     if write_back == write_back.YES:
-        with open(src, "w", encoding=src_buffer.encoding) as f:
+        with open(src, "w", encoding=encoding, newline=newline) as f:
             f.write(dst_contents)
     elif write_back == write_back.DIFF:
-        src_name = f"{src}  (original)"
-        dst_name = f"{src}  (formatted)"
+        now = datetime.utcnow()
+        src_name = f"{src}\t{then} +0000"
+        dst_name = f"{src}\t{now} +0000"
         diff_contents = diff(src_contents, dst_contents, src_name, dst_name)
         if lock:
             lock.acquire()
         try:
-            sys.stdout.write(diff_contents)
+            f = io.TextIOWrapper(
+                sys.stdout.buffer,
+                encoding=encoding,
+                newline=newline,
+                write_through=True,
+            )
+            f.write(diff_contents)
+            f.detach()
         finally:
             if lock:
                 lock.release()
@@ -357,33 +556,48 @@ def format_file_in_place(
 
 
 def format_stdin_to_stdout(
-    line_length: int, fast: bool, write_back: WriteBack = WriteBack.NO
+    line_length: int,
+    fast: bool,
+    write_back: WriteBack = WriteBack.NO,
+    mode: FileMode = FileMode.AUTO_DETECT,
 ) -> bool:
     """Format file on stdin. Return True if changed.
 
-    If `write_back` is True, write reformatted code back to stdout.
-    `line_length` and `fast` arguments are passed to :func:`format_file_contents`.
+    If `write_back` is YES, write reformatted code back to stdout. If it is DIFF,
+    write a diff to stdout.
+    `line_length`, `fast`, `is_pyi`, and `force_py36` arguments are passed to
+    :func:`format_file_contents`.
     """
-    src = sys.stdin.read()
+    then = datetime.utcnow()
+    src, encoding, newline = decode_bytes(sys.stdin.buffer.read())
     dst = src
     try:
-        dst = format_file_contents(src, line_length=line_length, fast=fast)
+        dst = format_file_contents(src, line_length=line_length, fast=fast, mode=mode)
         return True
 
     except NothingChanged:
         return False
 
     finally:
+        f = io.TextIOWrapper(
+            sys.stdout.buffer, encoding=encoding, newline=newline, write_through=True
+        )
         if write_back == WriteBack.YES:
-            sys.stdout.write(dst)
+            f.write(dst)
         elif write_back == WriteBack.DIFF:
-            src_name = "<stdin>  (original)"
-            dst_name = "<stdin>  (formatted)"
-            sys.stdout.write(diff(src, dst, src_name, dst_name))
+            now = datetime.utcnow()
+            src_name = f"STDIN\t{then} +0000"
+            dst_name = f"STDOUT\t{now} +0000"
+            f.write(diff(src, dst, src_name, dst_name))
+        f.detach()
 
 
 def format_file_contents(
-    src_contents: str, *, line_length: int, fast: bool, is_pyi: bool = False
+    src_contents: str,
+    *,
+    line_length: int,
+    fast: bool,
+    mode: FileMode = FileMode.AUTO_DETECT,
 ) -> FileContent:
     """Reformat contents a file and return new contents.
 
@@ -394,20 +608,18 @@ def format_file_contents(
     if src_contents.strip() == "":
         raise NothingChanged
 
-    dst_contents = format_str(src_contents, line_length=line_length, is_pyi=is_pyi)
+    dst_contents = format_str(src_contents, line_length=line_length, mode=mode)
     if src_contents == dst_contents:
         raise NothingChanged
 
     if not fast:
         assert_equivalent(src_contents, dst_contents)
-        assert_stable(
-            src_contents, dst_contents, line_length=line_length, is_pyi=is_pyi
-        )
+        assert_stable(src_contents, dst_contents, line_length=line_length, mode=mode)
     return dst_contents
 
 
 def format_str(
-    src_contents: str, line_length: int, *, is_pyi: bool = False
+    src_contents: str, line_length: int, *, mode: FileMode = FileMode.AUTO_DETECT
 ) -> FileContent:
     """Reformat a string and return new contents.
 
@@ -416,11 +628,18 @@ def format_str(
     src_node = lib2to3_parse(src_contents)
     dst_contents = ""
     future_imports = get_future_imports(src_node)
-    elt = EmptyLineTracker(is_pyi=is_pyi)
-    py36 = is_python36(src_node)
+    is_pyi = bool(mode & FileMode.PYI)
+    py36 = bool(mode & FileMode.PYTHON36) or is_python36(src_node)
+    normalize_strings = not bool(mode & FileMode.NO_STRING_NORMALIZATION)
+    normalize_fmt_off(src_node)
     lines = LineGenerator(
-        remove_u_prefix=py36 or "unicode_literals" in future_imports, is_pyi=is_pyi
+        remove_u_prefix=py36 or "unicode_literals" in future_imports,
+        is_pyi=is_pyi,
+        normalize_strings=normalize_strings,
+        allow_underscores=py36
+        and not bool(mode & FileMode.NO_NUMERIC_UNDERSCORE_NORMALIZATION),
     )
+    elt = EmptyLineTracker(is_pyi=is_pyi)
     empty_line = Line()
     after = 0
     for current_line in lines.visit(src_node):
@@ -434,6 +653,23 @@ def format_str(
     return dst_contents
 
 
+def decode_bytes(src: bytes) -> Tuple[FileContent, Encoding, NewLine]:
+    """Return a tuple of (decoded_contents, encoding, newline).
+
+    `newline` is either CRLF or LF but `decoded_contents` is decoded with
+    universal newlines (i.e. only contains LF).
+    """
+    srcbuf = io.BytesIO(src)
+    encoding, lines = tokenize.detect_encoding(srcbuf.readline)
+    if not lines:
+        return "", encoding, "\n"
+
+    newline = "\r\n" if b"\r\n" == lines[0][-2:] else "\n"
+    srcbuf.seek(0)
+    with io.TextIOWrapper(srcbuf, encoding) as tiow:
+        return tiow.read(), encoding, newline
+
+
 GRAMMARS = [
     pygram.python_grammar_no_print_statement_no_exec_statement,
     pygram.python_grammar_no_print_statement,
@@ -444,9 +680,8 @@ GRAMMARS = [
 def lib2to3_parse(src_txt: str) -> Node:
     """Given a string with source, return the lib2to3 Node."""
     grammar = pygram.python_grammar_no_print_statement
-    if src_txt[-1] != "\n":
-        nl = "\r\n" if "\r\n" in src_txt[:1024] else "\n"
-        src_txt += nl
+    if src_txt[-1:] != "\n":
+        src_txt += "\n"
     for grammar in GRAMMARS:
         drv = driver.Driver(grammar, pytree.convert)
         try:
@@ -460,7 +695,7 @@ def lib2to3_parse(src_txt: str) -> Node:
                 faulty_line = lines[lineno - 1]
             except IndexError:
                 faulty_line = "<line number missing in source>"
-            exc = ValueError(f"Cannot parse: {lineno}:{column}: {faulty_line}")
+            exc = InvalidInput(f"Cannot parse: {lineno}:{column}: {faulty_line}")
     else:
         raise exc from None
 
@@ -529,13 +764,15 @@ class DebugVisitor(Visitor[T]):
             out(f" {node.value!r}", fg="blue", bold=False)
 
     @classmethod
-    def show(cls, code: str) -> None:
+    def show(cls, code: Union[str, Leaf, Node]) -> None:
         """Pretty-print the lib2to3 AST of a given string of `code`.
 
         Convenience method for debugging.
         """
         v: DebugVisitor[None] = DebugVisitor()
-        list(v.visit(lib2to3_parse(code)))
+        if isinstance(code, str):
+            code = lib2to3_parse(code)
+        list(v.visit(code))
 
 
 KEYWORDS = set(keyword.kwlist)
@@ -552,6 +789,7 @@ STATEMENT = {
     syms.classdef,
 }
 STANDALONE_COMMENT = 153
+token.tok_name[STANDALONE_COMMENT] = "STANDALONE_COMMENT"
 LOGIC_OPERATORS = {"and", "or"}
 COMPARATORS = {
     token.LESS,
@@ -590,6 +828,7 @@ UNPACKING_PARENTS = {
     syms.dictsetmaker,
     syms.listmaker,
     syms.testlist_gexp,
+    syms.testlist_star_expr,
 }
 TEST_DESCENDANTS = {
     syms.test,
@@ -657,8 +896,8 @@ class BracketTracker:
     bracket_match: Dict[Tuple[Depth, NodeType], Leaf] = Factory(dict)
     delimiters: Dict[LeafID, Priority] = Factory(dict)
     previous: Optional[Leaf] = None
-    _for_loop_variable: int = 0
-    _lambda_arguments: int = 0
+    _for_loop_depths: List[int] = Factory(list)
+    _lambda_argument_depths: List[int] = Factory(list)
 
     def mark(self, leaf: Leaf) -> None:
         """Mark `leaf` with bracket-related metadata. Keep track of delimiters.
@@ -731,16 +970,21 @@ class BracketTracker:
         """
         if leaf.type == token.NAME and leaf.value == "for":
             self.depth += 1
-            self._for_loop_variable += 1
+            self._for_loop_depths.append(self.depth)
             return True
 
         return False
 
     def maybe_decrement_after_for_loop_variable(self, leaf: Leaf) -> bool:
         """See `maybe_increment_for_loop_variable` above for explanation."""
-        if self._for_loop_variable and leaf.type == token.NAME and leaf.value == "in":
+        if (
+            self._for_loop_depths
+            and self._for_loop_depths[-1] == self.depth
+            and leaf.type == token.NAME
+            and leaf.value == "in"
+        ):
             self.depth -= 1
-            self._for_loop_variable -= 1
+            self._for_loop_depths.pop()
             return True
 
         return False
@@ -753,16 +997,20 @@ class BracketTracker:
         """
         if leaf.type == token.NAME and leaf.value == "lambda":
             self.depth += 1
-            self._lambda_arguments += 1
+            self._lambda_argument_depths.append(self.depth)
             return True
 
         return False
 
     def maybe_decrement_after_lambda_arguments(self, leaf: Leaf) -> bool:
         """See `maybe_increment_lambda_arguments` above for explanation."""
-        if self._lambda_arguments and leaf.type == token.COLON:
+        if (
+            self._lambda_argument_depths
+            and self._lambda_argument_depths[-1] == self.depth
+            and leaf.type == token.COLON
+        ):
             self.depth -= 1
-            self._lambda_arguments -= 1
+            self._lambda_argument_depths.pop()
             return True
 
         return False
@@ -878,27 +1126,6 @@ class Line:
             and second_leaf.value == "def"
         )
 
-    @property
-    def is_flow_control(self) -> bool:
-        """Is this line a flow control statement?
-
-        Those are `return`, `raise`, `break`, and `continue`.
-        """
-        return (
-            bool(self)
-            and self.leaves[0].type == token.NAME
-            and self.leaves[0].value in FLOW_CONTROL
-        )
-
-    @property
-    def is_yield(self) -> bool:
-        """Is this line a yield statement?"""
-        return (
-            bool(self)
-            and self.leaves[0].type == token.NAME
-            and self.leaves[0].value == "yield"
-        )
-
     @property
     def is_class_paren_empty(self) -> bool:
         """Is this a class with no base classes but using parentheses?
@@ -915,6 +1142,15 @@ class Line:
             and self.leaves[3].value == ")"
         )
 
+    @property
+    def is_triple_quoted_string(self) -> bool:
+        """Is the line a triple quoted string?"""
+        return (
+            bool(self)
+            and self.leaves[0].type == token.STRING
+            and self.leaves[0].value.startswith(('"""', "'''"))
+        )
+
     def contains_standalone_comments(self, depth_limit: int = sys.maxsize) -> bool:
         """If so, needs to be split before emitting."""
         for leaf in self.leaves:
@@ -924,6 +1160,13 @@ class Line:
 
         return False
 
+    def contains_multiline_strings(self) -> bool:
+        for leaf in self.leaves:
+            if is_multiline_string(leaf):
+                return True
+
+        return False
+
     def maybe_remove_trailing_comma(self, closing: Leaf) -> bool:
         """Remove trailing comma if there is one and it's safe."""
         if not (
@@ -949,7 +1192,7 @@ class Line:
             self.remove_trailing_comma()
             return True
 
-        # Otheriwsse, if the trailing one is the only one, we might mistakenly
+        # Otherwise, 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
@@ -1005,6 +1248,9 @@ class Line:
 
         Provide a non-negative leaf `_index` to speed up the function.
         """
+        if not self.comments:
+            return
+
         if _index == -1:
             for _index, _leaf in enumerate(self.leaves):
                 if leaf is _leaf:
@@ -1028,18 +1274,18 @@ class Line:
 
     def is_complex_subscript(self, leaf: Leaf) -> bool:
         """Return True iff `leaf` is part of a slice with non-trivial exprs."""
-        open_lsqb = (
-            leaf if leaf.type == token.LSQB else self.bracket_tracker.get_open_lsqb()
-        )
+        open_lsqb = self.bracket_tracker.get_open_lsqb()
         if open_lsqb is None:
             return False
 
         subscript_start = open_lsqb.next_sibling
-        if (
-            isinstance(subscript_start, Node)
-            and subscript_start.type == syms.subscriptlist
-        ):
-            subscript_start = child_towards(subscript_start, leaf)
+
+        if isinstance(subscript_start, Node):
+            if subscript_start.type == syms.listmaker:
+                return False
+
+            if subscript_start.type == syms.subscriptlist:
+                subscript_start = child_towards(subscript_start, leaf)
         return subscript_start is not None and any(
             n.type in TEST_DESCENDANTS for n in subscript_start.pre_order()
         )
@@ -1064,55 +1310,6 @@ class Line:
         return bool(self.leaves or self.comments)
 
 
-class UnformattedLines(Line):
-    """Just like :class:`Line` but stores lines which aren't reformatted."""
-
-    def append(self, leaf: Leaf, preformatted: bool = True) -> None:
-        """Just add a new `leaf` to the end of the lines.
-
-        The `preformatted` argument is ignored.
-
-        Keeps track of indentation `depth`, which is useful when the user
-        says `# fmt: on`. Otherwise, doesn't do anything with the `leaf`.
-        """
-        try:
-            list(generate_comments(leaf))
-        except FormatOn as f_on:
-            self.leaves.append(f_on.leaf_from_consumed(leaf))
-            raise
-
-        self.leaves.append(leaf)
-        if leaf.type == token.INDENT:
-            self.depth += 1
-        elif leaf.type == token.DEDENT:
-            self.depth -= 1
-
-    def __str__(self) -> str:
-        """Render unformatted lines from leaves which were added with `append()`.
-
-        `depth` is not used for indentation in this case.
-        """
-        if not self:
-            return "\n"
-
-        res = ""
-        for leaf in self.leaves:
-            res += str(leaf)
-        return res
-
-    def append_comment(self, comment: Leaf) -> bool:
-        """Not implemented in this class. Raises `NotImplementedError`."""
-        raise NotImplementedError("Unformatted lines don't store comments separately.")
-
-    def maybe_remove_trailing_comma(self, closing: Leaf) -> bool:
-        """Does nothing and returns False."""
-        return False
-
-    def maybe_increment_for_loop_variable(self, leaf: Leaf) -> bool:
-        """Does nothing and returns False."""
-        return False
-
-
 @dataclass
 class EmptyLineTracker:
     """Provides a stateful method that returns the number of potential extra
@@ -1122,6 +1319,7 @@ class EmptyLineTracker:
     the prefix of the first leaf consists of optional newlines.  Those newlines
     are consumed by `maybe_empty_lines()` and included in the computation.
     """
+
     is_pyi: bool = False
     previous_line: Optional[Line] = None
     previous_after: int = 0
@@ -1131,12 +1329,8 @@ class EmptyLineTracker:
         """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.
+        lines (two on module-level).
         """
-        if isinstance(current_line, UnformattedLines):
-            return 0, 0
-
         before, after = self._maybe_empty_lines(current_line)
         before -= self.previous_after
         self.previous_after = after
@@ -1162,39 +1356,8 @@ class EmptyLineTracker:
                 before = 0 if depth else 1
             else:
                 before = 1 if depth else 2
-        is_decorator = current_line.is_decorator
-        if is_decorator or current_line.is_def or current_line.is_class:
-            if not is_decorator:
-                self.previous_defs.append(depth)
-            if self.previous_line is None:
-                # Don't insert empty lines before the first line in the file.
-                return 0, 0
-
-            if self.previous_line.is_decorator:
-                return 0, 0
-
-            if (
-                self.previous_line.is_comment
-                and self.previous_line.depth == current_line.depth
-                and before == 0
-            ):
-                return 0, 0
-
-            if self.is_pyi:
-                if self.previous_line.depth > current_line.depth:
-                    newlines = 1
-                elif current_line.is_class or self.previous_line.is_class:
-                    if current_line.is_stub_class and self.previous_line.is_stub_class:
-                        newlines = 0
-                    else:
-                        newlines = 1
-                else:
-                    newlines = 0
-            else:
-                newlines = 2
-            if current_line.depth and newlines:
-                newlines -= 1
-            return newlines, 0
+        if current_line.is_decorator or current_line.is_def or current_line.is_class:
+            return self._maybe_empty_lines_for_class_or_def(current_line, before)
 
         if (
             self.previous_line
@@ -1204,8 +1367,59 @@ class EmptyLineTracker:
         ):
             return (before or 1), 0
 
+        if (
+            self.previous_line
+            and self.previous_line.is_class
+            and current_line.is_triple_quoted_string
+        ):
+            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]):
@@ -1214,11 +1428,14 @@ 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.
     """
+
     is_pyi: bool = False
+    normalize_strings: bool = True
     current_line: Line = Factory(Line)
     remove_u_prefix: bool = False
+    allow_underscores: bool = False
 
-    def line(self, indent: int = 0, type: Type[Line] = Line) -> Iterator[Line]:
+    def line(self, indent: int = 0) -> Iterator[Line]:
         """Generate a line.
 
         If the line is empty, only emit if it makes sense.
@@ -1227,67 +1444,41 @@ class LineGenerator(Visitor[Line]):
         If any lines were generated, set up a new current_line.
         """
         if not self.current_line:
-            if self.current_line.__class__ == type:
-                self.current_line.depth += indent
-            else:
-                self.current_line = type(depth=self.current_line.depth + indent)
+            self.current_line.depth += indent
             return  # Line is empty, don't emit. Creating a new one unnecessary.
 
         complete_line = self.current_line
-        self.current_line = type(depth=complete_line.depth + indent)
+        self.current_line = Line(depth=complete_line.depth + indent)
         yield complete_line
 
-    def visit(self, node: LN) -> Iterator[Line]:
-        """Main method to visit `node` and its children.
-
-        Yields :class:`Line` objects.
-        """
-        if isinstance(self.current_line, UnformattedLines):
-            # File contained `# fmt: off`
-            yield from self.visit_unformatted(node)
-
-        else:
-            yield from super().visit(node)
-
     def visit_default(self, node: LN) -> Iterator[Line]:
         """Default `visit_*()` implementation. Recurses to children of `node`."""
         if isinstance(node, Leaf):
             any_open_brackets = self.current_line.bracket_tracker.any_open_brackets()
-            try:
-                for comment in generate_comments(node):
-                    if any_open_brackets:
-                        # any comment within brackets is subject to splitting
-                        self.current_line.append(comment)
-                    elif comment.type == token.COMMENT:
-                        # regular trailing comment
-                        self.current_line.append(comment)
-                        yield from self.line()
-
-                    else:
-                        # regular standalone comment
-                        yield from self.line()
-
-                        self.current_line.append(comment)
-                        yield from self.line()
-
-            except FormatOff as f_off:
-                f_off.trim_prefix(node)
-                yield from self.line(type=UnformattedLines)
-                yield from self.visit(node)
-
-            except FormatOn as f_on:
-                # This only happens here if somebody says "fmt: on" multiple
-                # times in a row.
-                f_on.trim_prefix(node)
-                yield from self.visit_default(node)
+            for comment in generate_comments(node):
+                if any_open_brackets:
+                    # any comment within brackets is subject to splitting
+                    self.current_line.append(comment)
+                elif comment.type == token.COMMENT:
+                    # regular trailing comment
+                    self.current_line.append(comment)
+                    yield from self.line()
 
-            else:
-                normalize_prefix(node, inside_brackets=any_open_brackets)
-                if node.type == token.STRING:
-                    normalize_string_prefix(node, remove_u_prefix=self.remove_u_prefix)
-                    normalize_string_quotes(node)
-                if node.type not in WHITESPACE:
-                    self.current_line.append(node)
+                else:
+                    # regular standalone comment
+                    yield from self.line()
+
+                    self.current_line.append(comment)
+                    yield from self.line()
+
+            normalize_prefix(node, inside_brackets=any_open_brackets)
+            if self.normalize_strings and node.type == token.STRING:
+                normalize_string_prefix(node, remove_u_prefix=self.remove_u_prefix)
+                normalize_string_quotes(node)
+            if node.type == token.NUMBER:
+                normalize_numeric_literal(node, self.allow_underscores)
+            if node.type not in WHITESPACE:
+                self.current_line.append(node)
         yield from super().visit_default(node)
 
     def visit_INDENT(self, node: Node) -> Iterator[Line]:
@@ -1375,32 +1566,6 @@ class LineGenerator(Visitor[Line]):
             yield from self.line()
             yield from self.visit(child)
 
-    def visit_import_from(self, node: Node) -> Iterator[Line]:
-        """Visit import_from and maybe put invisible parentheses.
-
-        This is separate from `visit_stmt` because import statements don't
-        support arbitrary atoms and thus handling of parentheses is custom.
-        """
-        check_lpar = False
-        for index, child in enumerate(node.children):
-            if check_lpar:
-                if child.type == token.LPAR:
-                    # make parentheses invisible
-                    child.value = ""  # type: ignore
-                    node.children[-1].value = ""  # type: ignore
-                else:
-                    # insert invisible parentheses
-                    node.insert_child(index, Leaf(token.LPAR, ""))
-                    node.append_child(Leaf(token.RPAR, ""))
-                break
-
-            check_lpar = (
-                child.type == token.NAME and child.value == "import"  # type: ignore
-            )
-
-        for child in node.children:
-            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()
@@ -1410,23 +1575,10 @@ class LineGenerator(Visitor[Line]):
         yield from self.visit_default(leaf)
         yield from self.line()
 
-    def visit_unformatted(self, node: LN) -> Iterator[Line]:
-        """Used when file contained a `# fmt: off`."""
-        if isinstance(node, Node):
-            for child in node.children:
-                yield from self.visit(child)
-
-        else:
-            try:
-                self.current_line.append(node)
-            except FormatOn as f_on:
-                f_on.trim_prefix(node)
-                yield from self.line()
-                yield from self.visit(node)
-
-            if node.type == token.ENDMARKER:
-                # somebody decided not to put a final `# fmt: on`
-                yield from self.line()
+    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 __attrs_post_init__(self) -> None:
         """You are in a twisty little maze of passages."""
@@ -1447,6 +1599,7 @@ class LineGenerator(Visitor[Line]):
         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_async_funcdef = self.visit_async_stmt
         self.visit_decorated = self.visit_decorators
 
@@ -1668,7 +1821,7 @@ def whitespace(leaf: Leaf, *, complex_subscript: bool) -> str:  # noqa C901
             elif prevp.type == token.EQUAL and prevp_parent.type == syms.argument:
                 return NO
 
-        elif t == token.NAME or t == token.NUMBER:
+        elif t in {token.NAME, token.NUMBER, token.STRING}:
             return NO
 
     elif p.type == syms.import_from:
@@ -1715,6 +1868,31 @@ def child_towards(ancestor: Node, descendant: LN) -> Optional[LN]:
     return node
 
 
+def container_of(leaf: Leaf) -> LN:
+    """Return `leaf` or one of its ancestors that is the topmost container of it.
+
+    By "container" we mean a node where `leaf` is the very first child.
+    """
+    same_prefix = leaf.prefix
+    container: LN = leaf
+    while container:
+        parent = container.parent
+        if parent is None:
+            break
+
+        if parent.children[0].prefix != same_prefix:
+            break
+
+        if parent.type == syms.file_input:
+            break
+
+        if parent.prev_sibling is not None and parent.prev_sibling.type in BRACKETS:
+            break
+
+        container = parent
+    return container
+
+
 def is_split_after_delimiter(leaf: Leaf, previous: Leaf = None) -> int:
     """Return the priority of the `leaf` delimiter, given a line break after it.
 
@@ -1730,7 +1908,7 @@ def is_split_after_delimiter(leaf: Leaf, previous: Leaf = None) -> int:
 
 
 def is_split_before_delimiter(leaf: Leaf, previous: Leaf = None) -> int:
-    """Return the priority of the `leaf` delimiter, given a line before after it.
+    """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.
@@ -1767,15 +1945,20 @@ def is_split_before_delimiter(leaf: Leaf, previous: Leaf = None) -> int:
     ):
         return STRING_PRIORITY
 
-    if leaf.type != token.NAME:
+    if leaf.type not in {token.NAME, token.ASYNC}:
         return 0
 
     if (
         leaf.value == "for"
         and leaf.parent
         and leaf.parent.type in {syms.comp_for, syms.old_comp_for}
+        or leaf.type == token.ASYNC
     ):
-        return COMPREHENSION_PRIORITY
+        if (
+            not isinstance(leaf.prev_sibling, Leaf)
+            or leaf.prev_sibling.value != "async"
+        ):
+            return COMPREHENSION_PRIORITY
 
     if (
         leaf.value == "if"
@@ -1820,6 +2003,10 @@ def is_split_before_delimiter(leaf: Leaf, previous: Leaf = None) -> int:
     return 0
 
 
+FMT_OFF = {"# fmt: off", "# fmt:off", "# yapf: disable"}
+FMT_ON = {"# fmt: on", "# fmt:on", "# yapf: enable"}
+
+
 def generate_comments(leaf: LN) -> Iterator[Leaf]:
     """Clean the prefix of the `leaf` and generate comments from it, if any.
 
@@ -1839,16 +2026,38 @@ def generate_comments(leaf: LN) -> Iterator[Leaf]:
     Inline comments are emitted as regular token.COMMENT leaves.  Standalone
     are emitted with a fake STANDALONE_COMMENT token identifier.
     """
-    p = leaf.prefix
-    if not p:
-        return
+    for pc in list_comments(leaf.prefix, is_endmarker=leaf.type == token.ENDMARKER):
+        yield Leaf(pc.type, pc.value, prefix="\n" * pc.newlines)
 
-    if "#" not in p:
-        return
+
+@dataclass
+class ProtoComment:
+    """Describes a piece of syntax that is a comment.
+
+    It's not a :class:`blib2to3.pytree.Leaf` so that:
+
+    * it can be cached (`Leaf` objects should not be reused more than once as
+      they store their lineno, column, prefix, and parent information);
+    * `newlines` and `consumed` fields are kept separate from the `value`. This
+      simplifies handling of special marker comments like ``# fmt: off/on``.
+    """
+
+    type: int  # token.COMMENT or STANDALONE_COMMENT
+    value: str  # content of the comment
+    newlines: int  # how many newlines before the comment
+    consumed: int  # how many characters of the original leaf's prefix did we consume
+
+
+@lru_cache(maxsize=4096)
+def list_comments(prefix: str, *, is_endmarker: bool) -> List[ProtoComment]:
+    """Return a list of :class:`ProtoComment` objects parsed from the given `prefix`."""
+    result: List[ProtoComment] = []
+    if not prefix or "#" not in prefix:
+        return result
 
     consumed = 0
     nlines = 0
-    for index, line in enumerate(p.split("\n")):
+    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:
@@ -1856,25 +2065,18 @@ def generate_comments(leaf: LN) -> Iterator[Leaf]:
         if not line.startswith("#"):
             continue
 
-        if index == 0 and leaf.type != token.ENDMARKER:
+        if index == 0 and not is_endmarker:
             comment_type = token.COMMENT  # simple trailing comment
         else:
             comment_type = STANDALONE_COMMENT
         comment = make_comment(line)
-        yield Leaf(comment_type, comment, prefix="\n" * nlines)
-
-        if comment in {"# fmt: on", "# yapf: enable"}:
-            raise FormatOn(consumed)
-
-        if comment in {"# fmt: off", "# yapf: disable"}:
-            if comment_type == STANDALONE_COMMENT:
-                raise FormatOff(consumed)
-
-            prev = preceding_leaf(leaf)
-            if not prev or prev.type in WHITESPACE:  # standalone comment in disguise
-                raise FormatOff(consumed)
-
+        result.append(
+            ProtoComment(
+                type=comment_type, value=comment, newlines=nlines, consumed=consumed
+            )
+        )
         nlines = 0
+    return result
 
 
 def make_comment(content: str) -> str:
@@ -1909,7 +2111,7 @@ def split_line(
     If `py36` is True, splitting may generate syntax that is only compatible
     with Python 3.6 and later.
     """
-    if isinstance(line, UnformattedLines) or line.is_comment:
+    if line.is_comment:
         yield line
         return
 
@@ -2054,32 +2256,50 @@ def right_hand_split(
             result.append(leaf, preformatted=True)
             for comment_after in line.comments_after(leaf):
                 result.append(comment_after, preformatted=True)
-    bracket_split_succeeded_or_raise(head, body, tail)
     assert opening_bracket and closing_bracket
+    body.should_explode = should_explode(body, 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
-        opening_bracket.type == token.LPAR
+        and opening_bracket.type == token.LPAR
         and not opening_bracket.value
         # the closing bracket is an optional paren
         and closing_bracket.type == token.RPAR
         and not closing_bracket.value
-        # there are no standalone comments in the body
-        and not line.contains_standalone_comments(0)
-        # and it's not an import (optional parens are the only thing we can split
-        # on in this case; attempting a split without them is a waste of time)
+        # it's not an import (optional parens are the only thing we can split on
+        # in this case; attempting a split without them is a waste of time)
         and not line.is_import
+        # there are no standalone comments in the body
+        and not body.contains_standalone_comments(0)
+        # and we can actually remove the parens
+        and can_omit_invisible_parens(body, line_length)
     ):
         omit = {id(closing_bracket), *omit}
-        if can_omit_invisible_parens(body, line_length):
-            try:
-                yield from right_hand_split(line, line_length, py36=py36, omit=omit)
-                return
-            except CannotSplit:
-                pass
+        try:
+            yield from right_hand_split(line, line_length, py36=py36, omit=omit)
+            return
+
+        except CannotSplit:
+            if not (
+                can_be_split(body)
+                or is_line_short_enough(body, line_length=line_length)
+            ):
+                raise CannotSplit(
+                    "Splitting failed, body is still too long and can't be split."
+                )
+
+            elif head.contains_multiline_strings() or tail.contains_multiline_strings():
+                raise CannotSplit(
+                    "The current optional pair of parentheses is bound to fail to "
+                    "satisfy the splitting algorithm because the head or the tail "
+                    "contains multiline strings which by definition never fit one "
+                    "line."
+                )
 
     ensure_visible(opening_bracket)
     ensure_visible(closing_bracket)
-    body.should_explode = should_explode(body, opening_bracket)
     for result in (head, body, tail):
         if result:
             yield result
@@ -2294,8 +2514,8 @@ def normalize_string_quotes(leaf: Leaf) -> None:
 
     prefix = leaf.value[:first_quote_pos]
     unescaped_new_quote = re.compile(rf"(([^\\]|^)(\\\\)*){new_quote}")
-    escaped_new_quote = re.compile(rf"([^\\]|^)\\(\\\\)*{new_quote}")
-    escaped_orig_quote = re.compile(rf"([^\\]|^)\\(\\\\)*{orig_quote}")
+    escaped_new_quote = re.compile(rf"([^\\]|^)\\((?:\\\\)*){new_quote}")
+    escaped_orig_quote = re.compile(rf"([^\\]|^)\\((?:\\\\)*){orig_quote}")
     body = leaf.value[first_quote_pos + len(orig_quote) : -len(orig_quote)]
     if "r" in prefix.casefold():
         if unescaped_new_quote.search(body):
@@ -2306,15 +2526,21 @@ def normalize_string_quotes(leaf: Leaf) -> None:
         # Do not introduce or remove backslashes in raw strings
         new_body = body
     else:
-        # remove unnecessary quotes
+        # remove unnecessary escapes
         new_body = sub_twice(escaped_new_quote, rf"\1\2{new_quote}", body)
         if body != new_body:
-            # Consider the string without unnecessary quotes as the original
+            # Consider the string without unnecessary escapes as the original
             body = new_body
             leaf.value = f"{prefix}{orig_quote}{body}{orig_quote}"
         new_body = sub_twice(escaped_orig_quote, rf"\1\2{orig_quote}", new_body)
         new_body = sub_twice(unescaped_new_quote, rf"\1\\{new_quote}", new_body)
-    if new_quote == '"""' and new_body[-1] == '"':
+    if "f" in prefix.casefold():
+        matches = re.findall(r"[^{]\{(.*?)\}[^}]", new_body)
+        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("\\")
@@ -2328,6 +2554,83 @@ def normalize_string_quotes(leaf: Leaf) -> None:
     leaf.value = f"{prefix}{new_quote}{new_body}{new_quote}"
 
 
+def normalize_numeric_literal(leaf: Leaf, allow_underscores: bool) -> None:
+    """Normalizes numeric (float, int, and complex) literals.
+
+    All letters used in the representation are normalized to lowercase (except
+    in Python 2 long literals), and long number literals are split using underscores.
+    """
+    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, allow_underscores)
+        after = format_int_string(after, allow_underscores)
+        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, allow_underscores)}{suffix}"
+    else:
+        text = format_float_or_int_string(text, allow_underscores)
+    leaf.value = text
+
+
+def format_float_or_int_string(text: str, allow_underscores: bool) -> str:
+    """Formats a float string like "1.0"."""
+    if "." not in text:
+        return format_int_string(text, allow_underscores)
+
+    before, after = text.split(".")
+    before = format_int_string(before, allow_underscores) if before else "0"
+    if after:
+        after = format_int_string(after, allow_underscores, count_from_end=False)
+    else:
+        after = "0"
+    return f"{before}.{after}"
+
+
+def format_int_string(
+    text: str, allow_underscores: bool, count_from_end: bool = True
+) -> str:
+    """Normalizes underscores in a string to e.g. 1_000_000.
+
+    Input must be a string of digits and optional underscores.
+    If count_from_end is False, we add underscores after groups of three digits
+    counting from the beginning instead of the end of the strings. This is used
+    for the fractional part of float literals.
+    """
+    if not allow_underscores:
+        return text
+
+    text = text.replace("_", "")
+    if len(text) <= 5:
+        # No underscores for numbers <= 5 digits long.
+        return text
+
+    if count_from_end:
+        # Avoid removing leading zeros, which are important if we're formatting
+        # part of a number like "0.001".
+        return format(int("1" + text), "3_")[1:].lstrip("_")
+    else:
+        return "_".join(text[i : i + 3] for i in range(0, len(text), 3))
+
+
 def normalize_invisible_parens(node: Node, parens_after: Set[str]) -> None:
     """Make existing optional parentheses invisible or create new ones.
 
@@ -2337,22 +2640,39 @@ def normalize_invisible_parens(node: Node, parens_after: Set[str]) -> None:
     Standardizes on visible parentheses for single-element tuples, and keeps
     existing visible parentheses for other tuples and generator expressions.
     """
-    try:
-        list(generate_comments(node))
-    except FormatOff:
-        return  # This `node` has a prefix with `# fmt: off`, don't mess with parens.
+    for pc in list_comments(node.prefix, is_endmarker=False):
+        if pc.value in FMT_OFF:
+            # This `node` has a prefix with `# fmt: off`, don't mess with parens.
+            return
 
     check_lpar = False
-    for child in list(node.children):
+    for index, child in enumerate(list(node.children)):
         if check_lpar:
             if child.type == syms.atom:
-                maybe_make_parens_invisible_in_atom(child)
+                if maybe_make_parens_invisible_in_atom(child):
+                    lpar = Leaf(token.LPAR, "")
+                    rpar = Leaf(token.RPAR, "")
+                    index = child.remove() or 0
+                    node.insert_child(index, Node(syms.atom, [lpar, child, rpar]))
             elif is_one_tuple(child):
                 # wrap child in visible parentheses
                 lpar = Leaf(token.LPAR, "(")
                 rpar = Leaf(token.RPAR, ")")
-                index = child.remove() or 0
+                child.remove()
                 node.insert_child(index, Node(syms.atom, [lpar, child, rpar]))
+            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 child in invisible parentheses
                 lpar = Leaf(token.LPAR, "")
@@ -2363,8 +2683,89 @@ def normalize_invisible_parens(node: Node, parens_after: Set[str]) -> None:
         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 = 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:
+        for comment in list_comments(container.prefix, is_endmarker=False):
+            if comment.value in FMT_ON:
+                return
+
+        yield container
+
+        container = container.next_sibling
+
+
 def maybe_make_parens_invisible_in_atom(node: LN) -> bool:
-    """If it's safe, make the parens in the atom `node` invisible, recusively."""
+    """If it's safe, make the parens in the atom `node` invisible, recursively.
+
+    Returns whether the node should itself be wrapped in invisible parentheses.
+
+    """
     if (
         node.type != syms.atom
         or is_empty_tuple(node)
@@ -2382,9 +2783,9 @@ def maybe_make_parens_invisible_in_atom(node: LN) -> bool:
         last.value = ""  # type: ignore
         if len(node.children) > 1:
             maybe_make_parens_invisible_in_atom(node.children[1])
-        return True
+        return False
 
-    return False
+    return True
 
 
 def is_empty_tuple(node: LN) -> bool:
@@ -2562,7 +2963,8 @@ def is_python36(node: Node) -> bool:
     """Return True if the current file is using Python 3.6+ features.
 
     Currently looking for:
-    - f-strings; and
+    - f-strings;
+    - underscores in numeric literals; and
     - trailing commas after * or ** in function signatures and calls.
     """
     for n in node.pre_order():
@@ -2571,6 +2973,10 @@ def is_python36(node: Node) -> bool:
             if value_head in {'f"', 'F"', "f'", "F'", "rf", "fr", "RF", "FR"}:
                 return True
 
+        elif n.type == token.NUMBER:
+            if "_" in n.value:  # type: ignore
+                return True
+
         elif (
             n.type in {syms.typedargslist, syms.arglist}
             and n.children
@@ -2604,27 +3010,22 @@ def generate_trailers_to_omit(line: Line, line_length: int) -> Iterator[Set[Leaf
     length = 4 * line.depth
     opening_bracket = None
     closing_bracket = None
-    optional_brackets: Set[LeafID] = set()
     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
 
-        if leaf.type == STANDALONE_COMMENT:
+        has_inline_comment = leaf_length > len(leaf.value) + len(leaf.prefix)
+        if leaf.type == STANDALONE_COMMENT or has_inline_comment:
             break
 
-        optional_brackets.discard(id(leaf))
         if opening_bracket:
             if leaf is opening_bracket:
                 opening_bracket = None
             elif leaf.type in CLOSING_BRACKETS:
                 inner_brackets.add(id(leaf))
         elif leaf.type in CLOSING_BRACKETS:
-            if not leaf.value:
-                optional_brackets.add(id(opening_bracket))
-                continue
-
             if index > 0 and line.leaves[index - 1].type in OPENING_BRACKETS:
                 # Empty brackets would fail a split so treat them as "inner"
                 # brackets (e.g. only add them to the `omit` set if another
@@ -2632,18 +3033,36 @@ def generate_trailers_to_omit(line: Line, line_length: int) -> Iterator[Set[Leaf
                 inner_brackets.add(id(leaf))
                 continue
 
-            opening_bracket = leaf.opening_bracket
             if closing_bracket:
                 omit.add(id(closing_bracket))
                 omit.update(inner_brackets)
                 inner_brackets.clear()
                 yield omit
-            closing_bracket = leaf
+
+            if leaf.value:
+                opening_bracket = leaf.opening_bracket
+                closing_bracket = leaf
 
 
 def get_future_imports(node: Node) -> Set[str]:
     """Return a set of __future__ imports in the file."""
-    imports = set()
+    imports: Set[str] = set()
+
+    def get_imports_from_children(children: List[LN]) -> Generator[str, None, None]:
+        for child in children:
+            if isinstance(child, Leaf):
+                if child.type == token.NAME:
+                    yield child.value
+            elif child.type == syms.import_as_name:
+                orig_name = child.children[0]
+                assert isinstance(orig_name, Leaf), "Invalid syntax parsing imports"
+                assert orig_name.type == token.NAME, "Invalid syntax parsing imports"
+                yield orig_name.value
+            elif child.type == syms.import_as_names:
+                yield from get_imports_from_children(child.children)
+            else:
+                assert False, "Invalid syntax parsing imports"
+
     for child in node.children:
         if child.type != syms.simple_stmt:
             break
@@ -2662,54 +3081,92 @@ def get_future_imports(node: Node) -> Set[str]:
             module_name = first_child.children[1]
             if not isinstance(module_name, Leaf) or module_name.value != "__future__":
                 break
-            for import_from_child in first_child.children[3:]:
-                if isinstance(import_from_child, Leaf):
-                    if import_from_child.type == token.NAME:
-                        imports.add(import_from_child.value)
-                else:
-                    assert import_from_child.type == syms.import_as_names
-                    for leaf in import_from_child.children:
-                        if isinstance(leaf, Leaf) and leaf.type == token.NAME:
-                            imports.add(leaf.value)
+            imports |= set(get_imports_from_children(first_child.children[3:]))
         else:
             break
     return imports
 
 
-PYTHON_EXTENSIONS = {".py", ".pyi"}
-BLACKLISTED_DIRECTORIES = {
-    "build",
-    "buck-out",
-    "dist",
-    "_build",
-    ".git",
-    ".hg",
-    ".mypy_cache",
-    ".tox",
-    ".venv",
-}
+def gen_python_files_in_dir(
+    path: Path,
+    root: Path,
+    include: Pattern[str],
+    exclude: Pattern[str],
+    report: "Report",
+) -> 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]:
-    """Generate all files under `path` which aren't under BLACKLISTED_DIRECTORIES
-    and have one of the PYTHON_EXTENSIONS.
+    `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:
+        try:
+            normalized_path = "/" + child.resolve().relative_to(root).as_posix()
+        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)
+
+        elif child.is_file():
+            include_match = include.search(normalized_path)
+            if include_match:
+                yield child
 
-        elif child.is_file() and child.suffix in PYTHON_EXTENSIONS:
-            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").is_dir():
+            return directory
+
+        if (directory / ".hg").is_dir():
+            return directory
+
+        if (directory / "pyproject.toml").is_file():
+            return directory
+
+    return directory
 
 
 @dataclass
 class Report:
     """Provides a reformatting counter. Can be rendered with `str(report)`."""
+
     check: bool = False
     quiet: bool = False
+    verbose: bool = False
     change_count: int = 0
     same_count: int = 0
     failure_count: int = 0
@@ -2718,11 +3175,11 @@ class Report:
         """Increment the counter for successful reformatting. Write out a message."""
         if changed is Changed.YES:
             reformatted = "would reformat" if self.check else "reformatted"
-            if not self.quiet:
+            if self.verbose or not self.quiet:
                 out(f"{reformatted} {src}")
             self.change_count += 1
         else:
-            if not self.quiet:
+            if self.verbose:
                 if changed is Changed.NO:
                     msg = f"{src} already well formatted, good job."
                 else:
@@ -2735,6 +3192,10 @@ class Report:
         err(f"error: cannot format {src}: {message}")
         self.failure_count += 1
 
+    def path_ignored(self, path: Path, message: str) -> None:
+        if self.verbose:
+            out(f"{path} ignored: {message}", bold=False)
+
     @property
     def return_code(self) -> int:
         """Return the exit code that the app should use.
@@ -2745,7 +3206,7 @@ class Report:
         - otherwise return 0.
         """
         # According to http://tldp.org/LDP/abs/html/exitcodes.html starting with
-        # 126 we have special returncodes reserved by the shell.
+        # 126 we have special return codes reserved by the shell.
         if self.failure_count:
             return 123
 
@@ -2847,9 +3308,11 @@ def assert_equivalent(src: str, dst: str) -> None:
         ) from None
 
 
-def assert_stable(src: str, dst: str, line_length: int, is_pyi: bool = False) -> None:
+def assert_stable(
+    src: str, dst: str, line_length: int, mode: FileMode = FileMode.AUTO_DETECT
+) -> None:
     """Raise AssertionError if `dst` reformats differently the second time."""
-    newdst = format_str(dst, line_length=line_length, is_pyi=is_pyi)
+    newdst = format_str(dst, line_length=line_length, mode=mode)
     if dst != newdst:
         log = dump_to_file(
             diff(src, dst, "source", "first pass"),
@@ -2926,6 +3389,16 @@ def sub_twice(regex: Pattern[str], replacement: str, original: str) -> str:
     return regex.sub(replacement, regex.sub(replacement, original))
 
 
+def re_compile_maybe_verbose(regex: str) -> Pattern[str]:
+    """Compile a regular expression string in `regex`.
+
+    If it contains newlines, use verbose mode.
+    """
+    if "\n" in regex:
+        regex = "(?x)" + regex
+    return re.compile(regex)
+
+
 def enumerate_reversed(sequence: Sequence[T]) -> Iterator[Tuple[Index, T]]:
     """Like `reversed(enumerate(sequence))` if that were possible."""
     index = len(sequence) - 1
@@ -2952,9 +3425,6 @@ def enumerate_with_length(
 
         comment: Optional[Leaf]
         for comment in line.comments_after(leaf, index):
-            if "\n" in comment.prefix:
-                return  # Oops, standalone comment!
-
             length += len(comment.value)
 
         yield index, leaf, length
@@ -2974,6 +3444,42 @@ def is_line_short_enough(line: Line, *, line_length: int, line_str: str = "") ->
     )
 
 
+def can_be_split(line: Line) -> bool:
+    """Return False if the line cannot be split *for sure*.
+
+    This is not an exhaustive search but a cheap heuristic that we can use to
+    avoid some unfortunate formattings (mostly around wrapping unsplittable code
+    in unnecessary parentheses).
+    """
+    leaves = line.leaves
+    if len(leaves) < 2:
+        return False
+
+    if leaves[0].type == token.STRING and leaves[1].type == token.DOT:
+        call_count = 0
+        dot_count = 0
+        next = leaves[-1]
+        for leaf in leaves[-2::-1]:
+            if leaf.type in OPENING_BRACKETS:
+                if next.type not in CLOSING_BRACKETS:
+                    return False
+
+                call_count += 1
+            elif leaf.type == token.DOT:
+                dot_count += 1
+            elif leaf.type == token.NAME:
+                if not (next.type == token.DOT or next.type in OPENING_BRACKETS):
+                    return False
+
+            elif leaf.type not in CLOSING_BRACKETS:
+                return False
+
+            if dot_count > 1 and call_count > 1:
+                return False
+
+    return True
+
+
 def can_omit_invisible_parens(line: Line, line_length: int) -> bool:
     """Does `line` have a shape safe to reformat without optional parens around it?
 
@@ -3063,16 +3569,16 @@ def can_omit_invisible_parens(line: Line, line_length: int) -> bool:
     return False
 
 
-def get_cache_file(line_length: int) -> Path:
-    return CACHE_DIR / f"cache.{line_length}.pickle"
+def get_cache_file(line_length: int, mode: FileMode) -> Path:
+    return CACHE_DIR / f"cache.{line_length}.{mode.value}.pickle"
 
 
-def read_cache(line_length: int) -> Cache:
+def read_cache(line_length: int, 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(line_length)
+    cache_file = get_cache_file(line_length, mode)
     if not cache_file.exists():
         return {}
 
@@ -3091,27 +3597,27 @@ def get_cache_info(path: Path) -> CacheInfo:
     return stat.st_mtime, stat.st_size
 
 
-def filter_cached(
-    cache: Cache, sources: Iterable[Path]
-) -> Tuple[List[Path], List[Path]]:
-    """Split a list of paths into two.
+def filter_cached(cache: Cache, sources: Iterable[Path]) -> Tuple[Set[Path], Set[Path]]:
+    """Split an iterable of paths in `sources` into two sets.
 
-    The first list contains paths of files that modified on disk or are not in the
-    cache. The other list contains paths to non-modified files.
+    The first contains paths of files that modified on disk or are not in the
+    cache. The other contains paths to non-modified files.
     """
-    todo, done = [], []
+    todo, done = set(), set()
     for src in sources:
         src = src.resolve()
         if cache.get(src) != get_cache_info(src):
-            todo.append(src)
+            todo.add(src)
         else:
-            done.append(src)
+            done.add(src)
     return todo, done
 
 
-def write_cache(cache: Cache, sources: List[Path], line_length: int) -> None:
+def write_cache(
+    cache: Cache, sources: Iterable[Path], line_length: int, mode: FileMode
+) -> None:
     """Update the cache file."""
-    cache_file = get_cache_file(line_length)
+    cache_file = get_cache_file(line_length, mode)
     try:
         if not CACHE_DIR.exists():
             CACHE_DIR.mkdir(parents=True)
@@ -3122,5 +3628,28 @@ def write_cache(cache: Cache, sources: List[Path], line_length: int) -> None:
         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
+
+
 if __name__ == "__main__":
+    patch_click()
     main()