rustpython-compiler 0.6.0

extern crate alloc;

use alloc::borrow::Cow;
pub use ruff_python_ast::token::{TokenKind, Tokens};
use ruff_python_parser::ParseErrorType;
use ruff_source_file::{PositionEncoding, SourceFile, SourceFileBuilder, SourceLocation};
use ruff_text_size::{Ranged, TextSize, TextSlice};
use rustpython_codegen::{compile, symboltable};
use thiserror::Error;

pub use rustpython_codegen::compile::CompileOpts;
pub use rustpython_compiler_core::{Mode, bytecode::CodeObject};

// these modules are out of repository. re-exporting them here for convenience.
pub use ruff_python_ast as ast;
pub use ruff_python_parser as parser;
pub use rustpython_codegen as codegen;
pub use rustpython_compiler_core as core;

#[derive(Error, Debug)]
pub enum CompileErrorType {
    #[error(transparent)]
    Codegen(#[from] codegen::error::CodegenErrorType),
    #[error(transparent)]
    Parse(#[from] ParseErrorType),
}

#[derive(Error, Debug)]
pub struct ParseError {
    #[source]
    pub error: ParseErrorType,
    pub raw_location: ruff_text_size::TextRange,
    pub location: SourceLocation,
    pub end_location: SourceLocation,
    pub source_path: String,
    /// Set when the error is an unclosed bracket (converted from EOF).
    pub is_unclosed_bracket: bool,
    /// Set when a string is still open at EOF and more input could close it.
    pub is_unclosed_string: bool,
}

impl ::core::fmt::Display for ParseError {
    fn fmt(&self, f: &mut ::core::fmt::Formatter<'_>) -> ::core::fmt::Result {
        self.error.fmt(f)
    }
}

#[derive(Error, Debug)]
pub enum CompileError {
    #[error(transparent)]
    Codegen(#[from] codegen::error::CodegenError),
    #[error(transparent)]
    Parse(#[from] ParseError),
}

impl CompileError {
    #[must_use]
    pub fn from_ruff_parse_error(
        error: parser::ParseError,
        source_file: &SourceFile,
        mode: Mode,
    ) -> Self {
        let raw_location = error.location;
        let diagnostic = match cpython_parse_diagnostic_override(&error, source_file, mode) {
            Some(diagnostic) => diagnostic,
            None => default_parse_diagnostic(error, source_file),
        };

        Self::Parse(ParseError {
            error: diagnostic.error,
            raw_location,
            location: diagnostic.location,
            end_location: diagnostic.end_location,
            source_path: source_file.name().to_owned(),
            is_unclosed_bracket: diagnostic.is_unclosed_bracket,
            is_unclosed_string: diagnostic.is_unclosed_string,
        })
    }

    fn from_source_error(source_file: &SourceFile, diagnostic: CpythonDiagnostic) -> Self {
        let (location, end_location) = source_locations(
            source_file,
            diagnostic.range.start(),
            diagnostic.range.end(),
        );
        Self::Parse(ParseError {
            error: parser::ParseErrorType::OtherError(diagnostic.message),
            raw_location: diagnostic.range,
            location,
            end_location,
            source_path: source_file.name().to_owned(),
            is_unclosed_bracket: diagnostic.is_unclosed_bracket,
            is_unclosed_string: diagnostic.is_unclosed_string,
        })
    }

    #[must_use]
    pub const fn location(&self) -> Option<SourceLocation> {
        match self {
            Self::Codegen(codegen_error) => codegen_error.location,
            Self::Parse(parse_error) => Some(parse_error.location),
        }
    }

    #[must_use]
    pub const fn python_location(&self) -> (usize, usize) {
        if let Some(location) = self.location() {
            (location.line.get(), location.character_offset.get())
        } else {
            (0, 0)
        }
    }

    #[must_use]
    pub fn python_end_location(&self) -> Option<(usize, usize)> {
        match self {
            Self::Codegen(codegen_error) => codegen_error
                .end_location
                .map(|end| (end.line.get(), end.character_offset.get())),
            Self::Parse(parse_error) => Some((
                parse_error.end_location.line.get(),
                parse_error.end_location.character_offset.get(),
            )),
        }
    }

    #[must_use]
    pub fn source_path(&self) -> &str {
        match self {
            Self::Codegen(codegen_error) => &codegen_error.source_path,
            Self::Parse(parse_error) => &parse_error.source_path,
        }
    }
}

// A syntax error the parser reports counts its columns in characters rather
// than in the bytes the range is measured in. An offset that lands inside a
// character walks back to where that character starts, the way decoding the
// line up to a truncated offset would.
fn source_location(source_file: &SourceFile, offset: TextSize) -> SourceLocation {
    let text = source_file.source_text();
    let mut index = offset.to_usize().min(text.len());
    while !text.is_char_boundary(index) {
        index -= 1;
    }
    source_file
        .to_source_code()
        .source_location(TextSize::new(index as u32), PositionEncoding::Utf32)
}

fn source_locations(
    source_file: &SourceFile,
    start: TextSize,
    end: TextSize,
) -> (SourceLocation, SourceLocation) {
    (
        source_location(source_file, start),
        source_location(source_file, end),
    )
}

struct NormalizedParseDiagnostic {
    error: parser::ParseErrorType,
    location: SourceLocation,
    end_location: SourceLocation,
    is_unclosed_bracket: bool,
    is_unclosed_string: bool,
}

impl NormalizedParseDiagnostic {
    const fn new(
        error: parser::ParseErrorType,
        location: SourceLocation,
        end_location: SourceLocation,
    ) -> Self {
        Self {
            error,
            location,
            end_location,
            is_unclosed_bracket: false,
            is_unclosed_string: false,
        }
    }

    fn other(source_file: &SourceFile, diagnostic: CpythonDiagnostic) -> Self {
        let (location, end_location) = source_locations(
            source_file,
            diagnostic.range.start(),
            diagnostic.range.end(),
        );
        let mut diagnostic_out = Self::new(
            parser::ParseErrorType::OtherError(diagnostic.message),
            location,
            end_location,
        );
        diagnostic_out.is_unclosed_string = diagnostic.is_unclosed_string;
        diagnostic_out.is_unclosed_bracket = diagnostic.is_unclosed_bracket;
        diagnostic_out
    }

    const fn with_unclosed_bracket(mut self, is_unclosed_bracket: bool) -> Self {
        self.is_unclosed_bracket = is_unclosed_bracket;
        self
    }
}

/// What CPython would have reported for a piece of source, before it is resolved to a line and
/// column. These are reconstructed by re-scanning after ruff's parse has already failed, so they
/// carry CPython's wording rather than a translation of ruff's own error, and they never reach
/// ruff — `NormalizedParseDiagnostic` and `CompileError` are the only things that consume one.
#[derive(Clone)]
struct CpythonDiagnostic {
    message: String,
    range: ruff_text_size::TextRange,
    is_unclosed_string: bool,
    is_unclosed_bracket: bool,
}

impl CpythonDiagnostic {
    /// The `u32` cast is ruff's invariant rather than one this adds: its `Lexer::new` asserts
    /// that the source fits in a `u32` ("Lexer only supports files with a size up to 4GB") and
    /// relies on that for its own offset arithmetic, and nothing here runs until that lexer has
    /// read the source and the parse built on it has failed.
    fn new(message: String, start: usize, end: usize) -> Self {
        Self {
            message,
            range: ruff_text_size::TextRange::new(
                TextSize::new(start as u32),
                TextSize::new(end as u32),
            ),
            is_unclosed_string: false,
            is_unclosed_bracket: false,
        }
    }

    const fn with_unclosed_string(mut self) -> Self {
        self.is_unclosed_string = true;
        self
    }

    const fn with_unclosed_bracket(mut self) -> Self {
        self.is_unclosed_bracket = true;
        self
    }
}

/// Lexer-class failures outrank print hints. Decode and f-string diagnostics
/// compete with lexer failures by offset, and an unterminated quote is only
/// a fallback when no decode/f-string diagnostic exists. Print is considered
/// against the final winner: a later 0x still beats print, but an f-string
/// that replaced that 0x must not hide an earlier print. A decode diagnostic
/// also suppresses an EOF unclosed opener.
#[derive(Clone, Copy, PartialEq, Eq)]
enum OverrideClass {
    Lexer,
    Decode,
    Print,
}

struct RankedOverride {
    diagnostic: CpythonDiagnostic,
    unclosed_bracket: bool,
    class: OverrideClass,
}

fn consider_override(
    best: &mut Option<RankedOverride>,
    diagnostic: CpythonDiagnostic,
    class: OverrideClass,
) {
    let unclosed_bracket = diagnostic.is_unclosed_bracket;
    consider_ranked(best, diagnostic, unclosed_bracket, class);
}

fn consider_ranked(
    best: &mut Option<RankedOverride>,
    diagnostic: CpythonDiagnostic,
    unclosed_bracket: bool,
    class: OverrideClass,
) {
    if best
        .as_ref()
        .is_none_or(|current| diagnostic.range.start() < current.diagnostic.range.start())
    {
        *best = Some(RankedOverride {
            diagnostic,
            unclosed_bracket,
            class,
        });
    }
}

fn cpython_parse_diagnostic_override(
    error: &parser::ParseError,
    source_file: &SourceFile,
    mode: Mode,
) -> Option<NormalizedParseDiagnostic> {
    let source_text = source_file.source_text();

    macro_rules! source_error {
        ($expr:expr) => {
            if let Some(error) = $expr {
                return Some(NormalizedParseDiagnostic::other(source_file, error));
            }
        };
    }

    let mut earliest: Option<RankedOverride> = None;
    if let Some(diagnostic) = invalid_number_literal_error(source_text) {
        consider_override(&mut earliest, diagnostic, OverrideClass::Lexer);
    }
    if let Some(diagnostic) = incompatible_string_prefix_error(source_text) {
        consider_override(&mut earliest, diagnostic, OverrideClass::Lexer);
    }
    if let Some(diagnostic) = non_printable_character_error(source_text) {
        consider_override(&mut earliest, diagnostic, OverrideClass::Lexer);
    }
    let bracket = bracket_syntax_error(source_text);
    if let Some(bracket) = bracket.as_ref() {
        // Unclosed openers are reported at the opener and only become errors
        // at EOF. A later token-time diagnostic (invalid number, prefix, …)
        // must keep winning. Mismatched closers stay in the lexer-class
        // positional ranking.
        if !bracket.unclosed {
            consider_ranked(
                &mut earliest,
                bracket.diagnostic.clone(),
                false,
                OverrideClass::Lexer,
            );
        }
    }
    let mut saw_decode = false;
    if let Some(diagnostic) = malformed_unicode_n_escape_error(source_text) {
        saw_decode = true;
        consider_override(&mut earliest, diagnostic, OverrideClass::Decode);
    }
    if let Some(diagnostic) = invalid_interpolated_string_error(source_text) {
        saw_decode = true;
        consider_override(&mut earliest, diagnostic, OverrideClass::Decode);
    }
    if let Some(diagnostic) = mixed_tstring_literal_error(error, source_text) {
        saw_decode = true;
        consider_override(&mut earliest, diagnostic, OverrideClass::Decode);
    }
    // A later ordinary unterminated quote is only a fallback. Format-spec
    // newlines and empty fields are decode diagnostics and must keep winning.
    let line_continuation = matches!(
        &error.error,
        parser::ParseErrorType::Lexical(parser::LexicalErrorType::LineContinuationError)
    );
    if !saw_decode
        && !line_continuation
        && let Some(diagnostic) = unterminated_string_error(source_text, mode)
    {
        consider_override(&mut earliest, diagnostic, OverrideClass::Lexer);
    }
    // Indentation errors outrank a print/exec missing-parentheses rewrite.
    let indent_error = matches!(
        &error.error,
        parser::ParseErrorType::Lexical(parser::LexicalErrorType::IndentationError)
            | parser::ParseErrorType::UnexpectedIndentation
    ) || expected_indented_block_error(error, source_text).is_some();
    if !indent_error
        && earliest
            .as_ref()
            .is_none_or(|current| current.class != OverrideClass::Lexer)
        && let Some(diagnostic) = invalid_legacy_statement_error(source_text)
    {
        consider_override(&mut earliest, diagnostic, OverrideClass::Print);
    }
    if !saw_decode
        && earliest
            .as_ref()
            .is_none_or(|current| current.class == OverrideClass::Print)
        && let Some(bracket) = bracket.filter(|bracket| bracket.unclosed)
    {
        consider_ranked(
            &mut earliest,
            bracket.diagnostic,
            true,
            OverrideClass::Lexer,
        );
    }
    if let Some(override_diag) = earliest {
        return Some(
            NormalizedParseDiagnostic::other(source_file, override_diag.diagnostic)
                .with_unclosed_bracket(override_diag.unclosed_bracket),
        );
    }

    if matches!(
        &error.error,
        parser::ParseErrorType::Lexical(parser::LexicalErrorType::LineContinuationError)
    ) {
        // exec input gets an implicit trailing newline, so a final `\` is a
        // continuation that then hits EOF (`E_EOF`). single/eval see `\` at
        // EOF as `E_LINECONT` instead.
        let terminal_backslash = source_text.len().checked_sub(1);
        if matches!(mode, Mode::Exec)
            && terminal_backslash == Some(error.location.start().to_usize())
        {
            let loc = source_line_end_location(source_file, error.location.start());
            return Some(NormalizedParseDiagnostic::new(
                parser::ParseErrorType::OtherError("unexpected EOF while parsing".to_owned()),
                loc,
                loc,
            ));
        }
        let loc = source_location(source_file, error.location.start() + TextSize::from(1));
        return Some(NormalizedParseDiagnostic::new(
            parser::ParseErrorType::OtherError(
                "unexpected character after line continuation character".to_owned(),
            ),
            loc,
            loc,
        ));
    }

    source_error!(unterminated_string_error(source_text, mode));
    source_error!(expected_indented_block_error(error, source_text));

    if matches!(
        &error.error,
        parser::ParseErrorType::Lexical(parser::LexicalErrorType::Eof)
    ) {
        return Some(eof_parse_diagnostic(error, source_file));
    }

    source_error!(invalid_type_param_error(source_text));
    source_error!(invalid_comprehension_error(source_text));
    source_error!(invalid_parameter_star_annotation_error(source_text));
    source_error!(invalid_parameter_list_error(source_text));
    source_error!(invalid_call_argument_error(source_text));

    if is_missing_comma_between_literals(error) {
        let (loc, end_loc) = adjusted_error_locations(source_file, error.location);
        let msg = "invalid syntax. Perhaps you forgot a comma?".into();
        return Some(NormalizedParseDiagnostic::new(
            parser::ParseErrorType::OtherError(msg),
            loc,
            end_loc,
        ));
    }

    source_error!(invalid_dict_error(source_text));
    if !matches!(mode, Mode::Eval) {
        source_error!(invalid_collection_assignment_error(source_text));
    }
    source_error!(invalid_group_error(source_text));
    source_error!(invalid_def_type_params_error(source_text));
    source_error!(invalid_expression_error(source_text));
    source_error!(invalid_named_expression_error(source_text));
    // Assignment and other statements are not expressions. Eval keeps the
    // generic parse error rather than a statement-target rewrite.
    if !matches!(mode, Mode::Eval) {
        source_error!(invalid_plain_assignment_error(source_text));
        source_error!(expression_assignment_error(source_text));
        source_error!(invalid_annotation_target_error(source_text));
        source_error!(invalid_assignment_target_error(source_text));
        source_error!(invalid_condition_assignment_error(
            source_text,
            error.location.start().to_usize()
        ));
        source_error!(invalid_augassign_target_error(source_text));
        source_error!(invalid_for_target_error(source_text));
        source_error!(invalid_with_target_error(source_text));
        source_error!(invalid_delete_target_error(source_text));
        source_error!(invalid_standalone_except_error(source_text));
        source_error!(invalid_import_statement_error(source_text));
        source_error!(invalid_import_target_error(source_text));
        source_error!(invalid_except_as_target_error(source_text));
        source_error!(invalid_match_mapping_rest_wildcard_error(source_text));
        source_error!(invalid_match_as_target_error(source_text));
        source_error!(invalid_for_if_clause_error(source_text));
        source_error!(invalid_if_expression_statement_error(source_text));
        source_error!(invalid_else_elif_error(source_text));
        source_error!(mixed_except_handlers_error(source_text));
    }

    if matches!(
        &error.error,
        parser::ParseErrorType::Lexical(parser::LexicalErrorType::InvalidByteLiteral)
    ) && let Some((start, end)) =
        bytes_literal_span(source_text, error.location.start().to_usize())
    {
        let (loc, end_loc) = source_locations(
            source_file,
            TextSize::new(start as u32),
            TextSize::new(end as u32),
        );
        return Some(NormalizedParseDiagnostic::new(
            error.error.clone(),
            loc,
            end_loc,
        ));
    }

    if matches!(
        &error.error,
        parser::ParseErrorType::Lexical(parser::LexicalErrorType::IndentationError)
    ) {
        let end_loc = source_line_end_location(source_file, error.location.start());
        return Some(NormalizedParseDiagnostic::new(
            error.error.clone(),
            end_loc,
            end_loc,
        ));
    }

    if matches!(
        &error.error,
        parser::ParseErrorType::InvalidAssignmentTarget
    ) {
        return Some(invalid_assignment_target_diagnostic(error, source_file));
    }

    if matches!(
        &error.error,
        parser::ParseErrorType::InvalidNamedAssignmentTarget
    ) {
        let (loc, end_loc) = adjusted_error_locations(source_file, error.location);
        let target = source_file.source_text().slice(error.location);
        let msg = format!("cannot use assignment expressions with {target}");
        return Some(NormalizedParseDiagnostic::new(
            parser::ParseErrorType::OtherError(msg),
            loc,
            end_loc,
        ));
    }

    // CPython's PEG parser collapses a bare "expected an expression" failure
    // into the generic "invalid syntax" message. rustpython-vm's `vm_new.rs`
    // does this same collapse for its own callers; rustpython-compiler has no
    // vm dependency, so mirror it here.
    if matches!(
        &error.error,
        parser::ParseErrorType::ExpectedExpression
            | parser::ParseErrorType::UnexpectedExpressionToken
    ) {
        let (loc, end_loc) = adjusted_error_locations(source_file, error.location);
        return Some(NormalizedParseDiagnostic::new(
            parser::ParseErrorType::OtherError("invalid syntax".into()),
            loc,
            end_loc,
        ));
    }

    None
}

fn eof_parse_diagnostic(
    error: &parser::ParseError,
    source_file: &SourceFile,
) -> NormalizedParseDiagnostic {
    let source_text = source_file.source_text();
    if let Some((bracket_char, bracket_offset)) = find_unclosed_bracket(source_text) {
        let loc = source_location(source_file, TextSize::new(bracket_offset as u32));
        let end_loc = SourceLocation {
            line: loc.line,
            character_offset: loc.character_offset.saturating_add(1),
        };
        let msg = format!("'{bracket_char}' was never closed");
        NormalizedParseDiagnostic::new(parser::ParseErrorType::OtherError(msg), loc, end_loc)
            .with_unclosed_bracket(true)
    } else {
        let end_loc = source_line_end_location(source_file, error.location.start());
        NormalizedParseDiagnostic::new(error.error.clone(), end_loc, end_loc)
    }
}

fn invalid_assignment_target_diagnostic(
    error: &parser::ParseError,
    source_file: &SourceFile,
) -> NormalizedParseDiagnostic {
    let (loc, end_loc) = adjusted_error_locations(source_file, error.location);
    let expr_str = source_file.source_text().slice(error.location);

    let msg = parser::parse_expression(expr_str).map_or_else(
        |_| match expr_str {
            "yield" => "assignment to yield expression not possible".into(),
            _ => format!("cannot assign to {expr_str}"),
        },
        |parsed| match *parsed.syntax().body {
            ast::Expr::Call(_) => "cannot assign to function call".into(),
            ast::Expr::BinOp(_) => "cannot assign to expression".into(),
            ast::Expr::If(_) => "cannot assign to conditional expression".into(),
            ast::Expr::Generator(_) => "cannot assign to generator expression".into(),
            ast::Expr::Yield(_) | ast::Expr::YieldFrom(_) => {
                "cannot assign to yield expression here. Maybe you meant '==' instead of '='?"
                    .into()
            }
            ast::Expr::FString(_) => "invalid syntax".into(),
            ast::Expr::StringLiteral(_)
            | ast::Expr::BytesLiteral(_)
            | ast::Expr::NumberLiteral(_) => {
                "cannot assign to literal here. Maybe you meant '==' instead of '='?".into()
            }
            ast::Expr::EllipsisLiteral(_) => {
                "cannot assign to ellipsis here. Maybe you meant '==' instead of '='?".into()
            }
            _ => format!("cannot assign to {expr_str}"),
        },
    );

    NormalizedParseDiagnostic::new(parser::ParseErrorType::OtherError(msg), loc, end_loc)
}

fn default_parse_diagnostic(
    error: parser::ParseError,
    source_file: &SourceFile,
) -> NormalizedParseDiagnostic {
    let (loc, end_loc) = adjusted_error_locations(source_file, error.location);
    NormalizedParseDiagnostic::new(error.error, loc, end_loc)
}

fn adjusted_error_locations(
    source_file: &SourceFile,
    range: ruff_text_size::TextRange,
) -> (SourceLocation, SourceLocation) {
    let mut locations = source_locations(source_file, range.start(), range.end());
    if locations.1.character_offset.get() == 1 && locations.1.line > locations.0.line {
        locations.1 = source_location(source_file, range.end() - TextSize::from(1));
        locations.1.character_offset = locations.1.character_offset.saturating_add(1);
    } else if range.is_empty() {
        // The parser blames a token, and the narrowest token still covers a
        // character, so an error reported between two of them spans one.
        locations.1.character_offset = locations.1.character_offset.saturating_add(1);
    }
    locations
}

fn source_line_end_location(source_file: &SourceFile, offset: TextSize) -> SourceLocation {
    let loc = source_location(source_file, offset);
    let line_idx = loc.line.to_zero_indexed();
    let line = source_file
        .source_text()
        .split('\n')
        .nth(line_idx)
        .unwrap_or("");
    let line_end_col = line.chars().count() + 1;
    SourceLocation {
        line: loc.line,
        character_offset: ruff_source_file::OneIndexed::new(line_end_col)
            .unwrap_or(loc.character_offset),
    }
}

fn is_missing_comma_between_literals(error: &parser::ParseError) -> bool {
    matches!(
        &error.error,
        parser::ParseErrorType::ExpectedToken { expected, found }
            if matches!((expected, found), (TokenKind::Comma, TokenKind::Int))
    )
}

fn is_ascii_identifier_char(byte: u8) -> bool {
    byte == b'_' || byte.is_ascii_alphanumeric()
}

fn identifier_continue_before(bytes: &[u8], index: usize) -> bool {
    if index == 0 {
        return false;
    }
    if bytes[index - 1].is_ascii() {
        return is_ascii_identifier_char(bytes[index - 1]);
    }
    let mut start = index - 1;
    while start > 0 && bytes[start] & 0b1100_0000 == 0b1000_0000 {
        start -= 1;
    }
    ::core::str::from_utf8(&bytes[start..index])
        .ok()
        .and_then(|text| text.chars().next_back())
        .is_some_and(|ch| ch == '_' || ch.is_alphanumeric())
}

fn numeric_keyword_suffix(rest: &[u8]) -> bool {
    rest.starts_with(b"and")
        || rest.starts_with(b"else")
        || rest.starts_with(b"for")
        || rest.starts_with(b"if")
        || rest.starts_with(b"in")
        || rest.starts_with(b"is")
        || rest.starts_with(b"or")
        || rest.starts_with(b"not")
}

fn consume_decimal_digits(bytes: &[u8], mut index: usize) -> usize {
    while index < bytes.len() {
        match bytes[index] {
            b'0'..=b'9' => index += 1,
            b'_' if bytes
                .get(index + 1)
                .is_some_and(|byte| byte.is_ascii_digit()) =>
            {
                index += 2;
            }
            _ => break,
        }
    }
    index
}

fn consume_radix_digits(bytes: &[u8], mut index: usize, is_digit: impl Fn(u8) -> bool) -> usize {
    while index < bytes.len() {
        if is_digit(bytes[index]) {
            index += 1;
        } else if bytes.get(index) == Some(&b'_')
            && bytes.get(index + 1).is_some_and(|&byte| is_digit(byte))
        {
            index += 2;
        } else {
            break;
        }
    }
    index
}

fn invalid_radix_literal_error(
    bytes: &[u8],
    start: usize,
    kind: &'static str,
    is_digit: impl Fn(u8) -> bool,
) -> Option<(String, usize)> {
    let mut index = start + 2;
    let mut has_digit = false;
    loop {
        let Some(&byte) = bytes.get(index) else {
            return if has_digit {
                None
            } else {
                Some((format!("invalid {kind} literal"), start + 1))
            };
        };
        if byte == b'_' {
            let Some(&next) = bytes.get(index + 1) else {
                return Some((format!("invalid {kind} literal"), index));
            };
            if is_digit(next) {
                has_digit = true;
                index += 2;
                continue;
            }
            if next.is_ascii_digit() && matches!(kind, "binary" | "octal") {
                return Some((
                    format!("invalid digit '{}' in {kind} literal", next as char),
                    index + 1,
                ));
            }
            return Some((format!("invalid {kind} literal"), index));
        }
        if is_digit(byte) {
            has_digit = true;
            index += 1;
            continue;
        }
        if byte.is_ascii_digit() && matches!(kind, "binary" | "octal") {
            return Some((
                format!("invalid digit '{}' in {kind} literal", byte as char),
                index,
            ));
        }
        if has_digit {
            return None;
        }
        return Some((format!("invalid {kind} literal"), start + 1));
    }
}

fn decimal_tail_error(bytes: &[u8], mut index: usize) -> Option<usize> {
    loop {
        while bytes.get(index).is_some_and(|byte| byte.is_ascii_digit()) {
            index += 1;
        }
        if bytes.get(index) != Some(&b'_') {
            return None;
        }
        let underscore = index;
        index += 1;
        if !bytes.get(index).is_some_and(|byte| byte.is_ascii_digit()) {
            return Some(underscore);
        }
    }
}

fn decimal_tail_end(bytes: &[u8], mut index: usize) -> usize {
    loop {
        while bytes.get(index).is_some_and(|byte| byte.is_ascii_digit()) {
            index += 1;
        }
        if bytes.get(index) == Some(&b'_')
            && bytes
                .get(index + 1)
                .is_some_and(|byte| byte.is_ascii_digit())
        {
            index += 2;
        } else {
            return index;
        }
    }
}

fn invalid_decimal_literal_error(bytes: &[u8], start: usize) -> Option<(String, usize)> {
    if bytes.get(start) == Some(&b'.') {
        return None;
    }
    let message = "invalid decimal literal".to_owned();
    if let Some(offset) = decimal_tail_error(bytes, start) {
        return Some((message, offset));
    }

    let mut index = decimal_tail_end(bytes, start);
    if bytes.get(index) == Some(&b'.') {
        if bytes.get(index + 1) == Some(&b'_') {
            return Some((message, index));
        }
        if let Some(offset) = decimal_tail_error(bytes, index + 1) {
            return Some((message, offset));
        }
        index = decimal_tail_end(bytes, index + 1);
    }
    if matches!(bytes.get(index), Some(b'e' | b'E')) {
        let exponent = index;
        index += 1;
        let sign = if matches!(bytes.get(index), Some(b'+' | b'-')) {
            let sign = index;
            index += 1;
            Some(sign)
        } else {
            None
        };
        if !bytes.get(index).is_some_and(|byte| byte.is_ascii_digit()) {
            // Without a sign the exponent letter is put back, so the position
            // is the digit before it rather than the letter.
            return Some((message, sign.unwrap_or_else(|| exponent.saturating_sub(1))));
        }
        if let Some(offset) = decimal_tail_error(bytes, index) {
            return Some((message, offset));
        }
    }
    None
}

fn leading_zero_decimal_literal_error(bytes: &[u8], start: usize) -> Option<CpythonDiagnostic> {
    if bytes.get(start) != Some(&b'0') {
        return None;
    }
    let mut index = start;
    loop {
        match bytes.get(index) {
            Some(b'0') => index += 1,
            Some(b'_')
                if bytes
                    .get(index + 1)
                    .is_some_and(|byte| byte.is_ascii_digit()) =>
            {
                index += 1;
            }
            _ => break,
        }
    }
    if bytes.get(index).is_some_and(|byte| byte.is_ascii_digit()) {
        let after_digits = decimal_tail_end(bytes, index);
        if !matches!(
            bytes.get(after_digits),
            Some(b'.' | b'e' | b'E' | b'j' | b'J')
        ) {
            return Some(CpythonDiagnostic::new(
                "leading zeros in decimal integer literals are not permitted; use an 0o prefix for octal integers"
                    .to_owned(),
                start,
                index,
            ));
        }
    }
    None
}

fn invalid_numeric_literal_error(bytes: &[u8], start: usize) -> Option<CpythonDiagnostic> {
    if bytes.get(start) == Some(&b'0') {
        let radix = match bytes.get(start + 1) {
            Some(b'x' | b'X') => invalid_radix_literal_error(bytes, start, "hexadecimal", |byte| {
                byte.is_ascii_hexdigit()
            }),
            Some(b'o' | b'O') => invalid_radix_literal_error(bytes, start, "octal", |byte| {
                matches!(byte, b'0'..=b'7')
            }),
            Some(b'b' | b'B') => invalid_radix_literal_error(bytes, start, "binary", |byte| {
                matches!(byte, b'0' | b'1')
            }),
            _ => None,
        };
        if let Some(radix) = radix {
            return Some(point_span(radix));
        }
        if let Some(err) = leading_zero_decimal_literal_error(bytes, start) {
            return Some(err);
        }
    }
    invalid_decimal_literal_error(bytes, start).map(point_span)
}

fn consume_exponent(bytes: &[u8], index: usize) -> usize {
    if !matches!(bytes.get(index), Some(b'e' | b'E')) {
        return index;
    }
    let mut cursor = index + 1;
    if matches!(bytes.get(cursor), Some(b'+' | b'-')) {
        cursor += 1;
    }
    if bytes.get(cursor).is_some_and(|byte| byte.is_ascii_digit()) {
        consume_decimal_digits(bytes, cursor)
    } else {
        index
    }
}

fn number_literal_end(bytes: &[u8], start: usize) -> Option<(&'static str, usize)> {
    if bytes.get(start) == Some(&b'.') {
        if !bytes
            .get(start + 1)
            .is_some_and(|byte| byte.is_ascii_digit())
        {
            return None;
        }
        let mut index = consume_decimal_digits(bytes, start + 1);
        index = consume_exponent(bytes, index);
        if matches!(bytes.get(index), Some(b'j' | b'J')) {
            return Some(("imaginary", index + 1));
        }
        return Some(("decimal", index));
    }

    if !bytes.get(start).is_some_and(|byte| byte.is_ascii_digit()) {
        return None;
    }

    if bytes.get(start) == Some(&b'0') {
        match bytes.get(start + 1) {
            Some(b'x' | b'X') => {
                let end = consume_radix_digits(bytes, start + 2, |byte| byte.is_ascii_hexdigit());
                return Some(("hexadecimal", end));
            }
            Some(b'o' | b'O') => {
                let end =
                    consume_radix_digits(bytes, start + 2, |byte| matches!(byte, b'0'..=b'7'));
                return Some(("octal", end));
            }
            Some(b'b' | b'B') => {
                let end =
                    consume_radix_digits(bytes, start + 2, |byte| matches!(byte, b'0' | b'1'));
                return Some(("binary", end));
            }
            _ => {}
        }
    }

    let mut index = consume_decimal_digits(bytes, start);
    if bytes.get(index) == Some(&b'.') {
        index = consume_decimal_digits(bytes, index + 1);
    }
    index = consume_exponent(bytes, index);
    if matches!(bytes.get(index), Some(b'j' | b'J')) {
        return Some(("imaginary", index + 1));
    }
    Some(("decimal", index))
}

fn quoted_string_is_closed(bytes: &[u8], start: usize) -> bool {
    let quote = bytes[start];
    let triple = bytes.get(start + 1) == Some(&quote) && bytes.get(start + 2) == Some(&quote);
    let end = skip_quoted_string(bytes, start);
    if triple {
        end >= start + 6
            && bytes[end - 3] == quote
            && bytes[end - 2] == quote
            && bytes[end - 1] == quote
    } else {
        end > start + 1 && bytes[end - 1] == quote
    }
}

fn bytes_literal_span(source: &str, error_at: usize) -> Option<(usize, usize)> {
    let bytes = source.as_bytes();
    if error_at > bytes.len() || bytes.is_empty() {
        return None;
    }
    let mut quote_idx = error_at.min(bytes.len().saturating_sub(1));
    loop {
        if matches!(bytes[quote_idx], b'\'' | b'"') {
            break;
        }
        if quote_idx == 0 {
            return None;
        }
        quote_idx -= 1;
    }
    let mut start = quote_idx;
    while start > 0 && matches!(bytes[start - 1], b'b' | b'B' | b'r' | b'R') {
        start -= 1;
    }
    if !matches!(bytes.get(start), Some(b'b' | b'B' | b'r' | b'R')) {
        return None;
    }
    let end = skip_quoted_string(bytes, quote_idx);
    Some((start, end))
}

fn skip_quoted_string(bytes: &[u8], mut index: usize) -> usize {
    let quote = bytes[index];
    let triple = bytes.get(index + 1) == Some(&quote) && bytes.get(index + 2) == Some(&quote);
    let quote_len = if triple { 3 } else { 1 };
    index += quote_len;
    while index < bytes.len() {
        if bytes[index] == b'\\' {
            index = (index + 2).min(bytes.len());
        } else if triple
            && bytes.get(index) == Some(&quote)
            && bytes.get(index + 1) == Some(&quote)
            && bytes.get(index + 2) == Some(&quote)
        {
            return index + 3;
        } else if !triple && bytes[index] == quote {
            return index + 1;
        } else {
            index += 1;
        }
    }
    index
}

// An error the tokenizer reports at a single position spans nothing.
fn point_span((message, offset): (String, usize)) -> CpythonDiagnostic {
    CpythonDiagnostic::new(message, offset, offset)
}

fn invalid_number_literal_error(source: &str) -> Option<CpythonDiagnostic> {
    let bytes = source.as_bytes();
    let mut index = 0;
    while index < bytes.len() {
        match bytes[index] {
            b'#' => {
                while index < bytes.len() && bytes[index] != b'\n' {
                    index += 1;
                }
            }
            b'\'' | b'"' => {
                index = skip_quoted_string(bytes, index);
            }
            byte if byte >= 0x80 || byte == b'_' || byte.is_ascii_alphabetic() => {
                index += 1;
                while index < bytes.len()
                    && (bytes[index] >= 0x80 || is_ascii_identifier_char(bytes[index]))
                {
                    index += 1;
                }
            }
            b'.' | b'0'..=b'9' => {
                if let Some(err) = invalid_numeric_literal_error(bytes, index) {
                    return Some(err);
                }
                let Some((kind, end)) = number_literal_end(bytes, index) else {
                    index += 1;
                    continue;
                };
                if end > index {
                    if source[end..].starts_with('⁄') {
                        return Some(CpythonDiagnostic::new(
                            "invalid character '⁄' (U+2044)".to_owned(),
                            end,
                            end,
                        ));
                    }
                    if bytes
                        .get(end)
                        .is_some_and(|byte| *byte < 128 && is_ascii_identifier_char(*byte))
                        && !numeric_keyword_suffix(&bytes[end..])
                    {
                        let offset = end.saturating_sub(1);
                        return Some(CpythonDiagnostic::new(
                            format!("invalid {kind} literal"),
                            offset,
                            offset,
                        ));
                    }
                }
                index = end.max(index + 1);
            }
            _ => index += 1,
        }
    }
    None
}

fn cpython_indented_block_clause(message: &str) -> Option<&'static str> {
    let clause = message.strip_prefix("Expected an indented block after ")?;
    Some(match clause {
        "`if` statement" => "'if' statement",
        "`elif` clause" => "'elif' statement",
        "`else` clause" => "'else' statement",
        "`for` statement" => "'for' statement",
        "`with` statement" => "'with' statement",
        "`while` statement" => "'while' statement",
        "`try` statement" => "'try' statement",
        "`except` clause" => "'except' statement",
        "`finally` clause" => "'finally' statement",
        "`match` statement" => "'match' statement",
        "`case` block" => "'case' statement",
        "`class` definition" => "class definition",
        "function definition" => "function definition",
        _ => return None,
    })
}

fn previous_non_empty_line_number(source: &str, offset: usize) -> Option<usize> {
    let bytes = source.as_bytes();
    let mut index = offset.min(bytes.len());
    while index > 0 {
        let line_end = index;
        while index > 0 && bytes[index - 1] != b'\n' {
            index -= 1;
        }
        let line_start = index;
        let content_start = skip_horizontal_whitespace(bytes, line_start);
        let mut content_end = line_end;
        while content_end > content_start
            && matches!(
                bytes.get(content_end - 1),
                Some(b' ' | b'\t' | b'\r' | b'\x0c')
            )
        {
            content_end -= 1;
        }
        if content_start < content_end {
            return Some(
                source[..line_start]
                    .bytes()
                    .filter(|byte| *byte == b'\n')
                    .count()
                    + 1,
            );
        }
        index = line_start.saturating_sub(1);
    }
    None
}

fn expected_indented_block_error(
    error: &parser::ParseError,
    source: &str,
) -> Option<CpythonDiagnostic> {
    let parser::ParseErrorType::OtherError(message) = &error.error else {
        return None;
    };
    let mut clause = cpython_indented_block_clause(message)?;
    let start = error.location.start().to_usize();
    let end = error.location.end().to_usize();
    let line = previous_non_empty_line_number(source, start)?;
    if clause == "'except' statement"
        && let Some(previous_line) = previous_non_empty_line(source, start)
        && matches!(
            previous_line.trim_start(),
            line if line.starts_with("except*") || line.starts_with("except *")
        )
    {
        clause = "'except*' statement";
    }
    Some(CpythonDiagnostic::new(
        format!("expected an indented block after {clause} on line {line}"),
        start,
        end,
    ))
}

fn previous_non_empty_line(source: &str, offset: usize) -> Option<&str> {
    let bytes = source.as_bytes();
    let mut index = offset.min(bytes.len());
    while index > 0 {
        let line_end = index;
        while index > 0 && bytes[index - 1] != b'\n' {
            index -= 1;
        }
        let line_start = index;
        let mut content_start = line_start;
        while content_start < line_end
            && matches!(bytes[content_start], b' ' | b'\t' | b'\n' | b'\r' | b'\x0c')
        {
            content_start += 1;
        }
        let mut content_end = line_end;
        while content_end > content_start
            && matches!(bytes[content_end - 1], b' ' | b'\t' | b'\r' | b'\x0c')
        {
            content_end -= 1;
        }
        if content_start < content_end {
            return source.get(line_start..line_end);
        }
        index = line_start.saturating_sub(1);
    }
    None
}

fn starts_identifier(bytes: &[u8], index: usize, word: &[u8]) -> bool {
    bytes.get(index..index + word.len()) == Some(word)
        && index
            .checked_sub(1)
            .and_then(|before| bytes.get(before))
            .is_none_or(|byte| !is_ascii_identifier_char(*byte))
        && bytes
            .get(index + word.len())
            .is_none_or(|byte| !is_ascii_identifier_char(*byte))
}

fn is_plain_assignment_operator(bytes: &[u8], index: usize) -> bool {
    bytes.get(index) == Some(&b'=')
        && bytes.get(index + 1) != Some(&b'=')
        && !matches!(
            index.checked_sub(1).and_then(|before| bytes.get(before)),
            Some(b'=' | b'!' | b'<' | b'>' | b':')
        )
}

fn is_simple_keyword_name(bytes: &[u8], mut start: usize, mut end: usize) -> bool {
    while matches!(
        bytes.get(start),
        Some(b' ' | b'\t' | b'\n' | b'\r' | b'\x0c')
    ) {
        start += 1;
    }
    while end > start
        && matches!(
            bytes.get(end - 1),
            Some(b' ' | b'\t' | b'\n' | b'\r' | b'\x0c')
        )
    {
        end -= 1;
    }
    let Some(&first) = bytes.get(start) else {
        return false;
    };
    if !(first == b'_' || first.is_ascii_alphabetic() || first >= 0x80) {
        return false;
    }
    let mut index = start + 1;
    while index < end {
        if bytes[index] < 0x80 && !is_ascii_identifier_char(bytes[index]) {
            return false;
        }
        index += 1;
    }
    true
}

fn is_function_parameter_list(bytes: &[u8], paren: usize) -> bool {
    let mut cursor = paren;
    while cursor > 0 && matches!(bytes.get(cursor - 1), Some(b' ' | b'\t' | b'\x0c')) {
        cursor -= 1;
    }
    if cursor > 0 && bytes.get(cursor - 1) == Some(&b']') {
        let mut bracket = cursor;
        let mut level = 0usize;
        while bracket > 0 {
            bracket -= 1;
            match bytes[bracket] {
                b']' => level += 1,
                b'[' => {
                    level = level.saturating_sub(1);
                    if level == 0 {
                        cursor = bracket;
                        break;
                    }
                }
                _ => {}
            }
        }
        while cursor > 0 && matches!(bytes.get(cursor - 1), Some(b' ' | b'\t' | b'\x0c')) {
            cursor -= 1;
        }
    }
    while cursor > 0
        && bytes
            .get(cursor - 1)
            .is_some_and(|byte| *byte >= 0x80 || is_ascii_identifier_char(*byte))
    {
        cursor -= 1;
    }
    while cursor > 0 && matches!(bytes.get(cursor - 1), Some(b' ' | b'\t' | b'\x0c')) {
        cursor -= 1;
    }
    cursor >= 3
        && starts_identifier(bytes, cursor - 3, b"def")
        && cursor
            .checked_sub(4)
            .and_then(|before| bytes.get(before))
            .is_none_or(|byte| !is_ascii_identifier_char(*byte))
}

#[derive(Clone, Copy)]
enum ParameterListKind {
    Function,
    Lambda,
}

fn matching_delimiter(bytes: &[u8], open: usize, close: u8) -> Option<usize> {
    let mut index = open;
    let mut level = 0usize;
    while index < bytes.len() {
        match bytes[index] {
            b'\'' | b'"' => index = skip_quoted_string(bytes, index),
            b'(' | b'[' | b'{' => {
                level += 1;
                index += 1;
            }
            byte if byte == close => {
                level = level.saturating_sub(1);
                if level == 0 {
                    return Some(index);
                }
                index += 1;
            }
            b')' | b']' | b'}' => {
                level = level.saturating_sub(1);
                index += 1;
            }
            _ => index += 1,
        }
    }
    None
}

fn find_lambda_parameter_end(bytes: &[u8], mut index: usize) -> Option<usize> {
    let mut level = 0usize;
    while index < bytes.len() {
        match bytes[index] {
            b'#' if level == 0 => return None,
            b'\'' | b'"' => index = skip_quoted_string(bytes, index),
            b'(' | b'[' | b'{' => {
                level += 1;
                index += 1;
            }
            b')' | b']' | b'}' => {
                level = level.saturating_sub(1);
                index += 1;
            }
            b':' if level == 0 => return Some(index),
            _ => index += 1,
        }
    }
    None
}

fn top_level_byte(bytes: &[u8], mut index: usize, end: usize, needle: u8) -> Option<usize> {
    let mut level = 0usize;
    while index < end {
        match bytes[index] {
            b'\'' | b'"' => index = skip_quoted_string(bytes, index),
            byte if level == 0 && byte == needle => return Some(index),
            b'(' | b'[' | b'{' => {
                level += 1;
                index += 1;
            }
            b')' | b']' | b'}' => {
                level = level.saturating_sub(1);
                index += 1;
            }
            _ => index += 1,
        }
    }
    None
}

fn identifier_end(bytes: &[u8], mut index: usize, end: usize) -> usize {
    if !bytes
        .get(index)
        .is_some_and(|byte| *byte >= 0x80 || *byte == b'_' || byte.is_ascii_alphabetic())
    {
        return index;
    }
    index += 1;
    while index < end
        && bytes
            .get(index)
            .is_some_and(|byte| *byte >= 0x80 || is_ascii_identifier_char(*byte))
    {
        index += 1;
    }
    index
}

fn expression_slice_is_tuple(source: &str, start: usize, end: usize) -> bool {
    let bytes = source.as_bytes();
    let (start, end) = trim_target_range(bytes, start, end);
    if start >= end {
        return false;
    }
    let Ok(parsed) = parser::parse(&source[start..end], parser::Mode::Expression.into()) else {
        return false;
    };
    matches!(parsed.into_syntax(), ast::Mod::Expression(expression) if matches!(*expression.body, ast::Expr::Tuple(_)))
}

fn type_param_list_open(bytes: &[u8], open: usize) -> bool {
    let mut cursor = open;
    while cursor > 0 && matches!(bytes.get(cursor - 1), Some(b' ' | b'\t' | b'\x0c')) {
        cursor -= 1;
    }
    while cursor > 0
        && bytes
            .get(cursor - 1)
            .is_some_and(|byte| *byte >= 0x80 || is_ascii_identifier_char(*byte))
    {
        cursor -= 1;
    }
    while cursor > 0 && matches!(bytes.get(cursor - 1), Some(b' ' | b'\t' | b'\x0c')) {
        cursor -= 1;
    }
    (cursor >= 3 && starts_identifier(bytes, cursor - 3, b"def"))
        || (cursor >= 5 && starts_identifier(bytes, cursor - 5, b"class"))
        || (cursor >= 4 && starts_identifier(bytes, cursor - 4, b"type"))
}

fn invalid_type_param_item_error(
    source: &str,
    start: usize,
    end: usize,
) -> Option<CpythonDiagnostic> {
    let bytes = source.as_bytes();
    let (start, end) = trim_target_range(bytes, start, end);
    if start >= end || bytes.get(start) != Some(&b'*') {
        return None;
    }
    let is_param_spec = bytes.get(start + 1) == Some(&b'*');
    let name_start = start + if is_param_spec { 2 } else { 1 };
    let name_end = identifier_end(bytes, name_start, end);
    if name_start == name_end {
        return None;
    }
    let colon = next_non_horizontal_whitespace(bytes, name_end);
    if colon >= end || bytes.get(colon) != Some(&b':') {
        return None;
    }
    let has_constraints = expression_slice_is_tuple(source, colon + 1, end);
    let message = match (is_param_spec, has_constraints) {
        (false, false) => "cannot use bound with TypeVarTuple",
        (false, true) => "cannot use constraints with TypeVarTuple",
        (true, false) => "cannot use bound with ParamSpec",
        (true, true) => "cannot use constraints with ParamSpec",
    };
    Some(CpythonDiagnostic::new(message.to_owned(), colon, colon + 1))
}

fn invalid_type_param_list_error(
    source: &str,
    open: usize,
    close: usize,
) -> Option<CpythonDiagnostic> {
    let bytes = source.as_bytes();
    let mut item_start = open + 1;
    let mut index = item_start;
    let mut level = 0usize;
    while index <= close {
        if index == close || (level == 0 && bytes.get(index) == Some(&b',')) {
            if let Some(error) = invalid_type_param_item_error(source, item_start, index) {
                return Some(error);
            }
            item_start = index + 1;
            index += 1;
            continue;
        }
        match bytes[index] {
            b'\'' | b'"' => index = skip_quoted_string(bytes, index),
            b'(' | b'[' | b'{' => {
                level += 1;
                index += 1;
            }
            b')' | b']' | b'}' => {
                level = level.saturating_sub(1);
                index += 1;
            }
            _ => index += 1,
        }
    }
    None
}

fn invalid_type_param_error(source: &str) -> Option<CpythonDiagnostic> {
    let bytes = source.as_bytes();
    let mut index = 0usize;
    while index < bytes.len() {
        match bytes[index] {
            b'#' => {
                while index < bytes.len() && bytes[index] != b'\n' {
                    index += 1;
                }
            }
            b'\'' | b'"' => index = skip_quoted_string(bytes, index),
            b'[' if type_param_list_open(bytes, index) => {
                let Some(close) = matching_delimiter(bytes, index, b']') else {
                    index += 1;
                    continue;
                };
                if let Some(error) = invalid_type_param_list_error(source, index, close) {
                    return Some(error);
                }
                index = close + 1;
            }
            _ => index += 1,
        }
    }
    None
}

fn invalid_comprehension_in_slice(
    bytes: &[u8],
    open: usize,
    close: usize,
) -> Option<CpythonDiagnostic> {
    let for_index = find_keyword_at_level(bytes, open + 1, close, b"for")?;
    let item_start = next_non_horizontal_whitespace(bytes, open + 1);
    if item_start >= for_index {
        return None;
    }
    if bytes.get(item_start..item_start + 2) == Some(b"**") && bytes.get(open) == Some(&b'{') {
        return Some(CpythonDiagnostic::new(
            "dict unpacking cannot be used in dict comprehension".to_owned(),
            item_start,
            item_start + 2,
        ));
    }
    if bytes.get(item_start..item_start + 2) == Some(b"**") && bytes.get(open) == Some(&b'(') {
        return Some(CpythonDiagnostic::new(
            "invalid syntax".to_owned(),
            for_index,
            for_index + 3,
        ));
    }
    if bytes.get(item_start) == Some(&b'*') {
        return Some(CpythonDiagnostic::new(
            "iterable unpacking cannot be used in comprehension".to_owned(),
            item_start,
            item_start + 1,
        ));
    }
    if !matches!(bytes.get(open), Some(b'[' | b'{')) {
        return None;
    }
    if top_level_colon(bytes, open + 1, for_index).is_none()
        && let Some(comma) = top_level_byte(bytes, open + 1, for_index, b',')
    {
        let (start, _) = trim_target_range(bytes, open + 1, comma);
        return Some(CpythonDiagnostic::new(
            "did you forget parentheses around the comprehension target?".to_owned(),
            start,
            comma + 1,
        ));
    }
    None
}

fn invalid_comprehension_error(source: &str) -> Option<CpythonDiagnostic> {
    let bytes = source.as_bytes();
    let mut index = 0usize;
    while index < bytes.len() {
        match bytes[index] {
            b'#' => {
                while index < bytes.len() && bytes[index] != b'\n' {
                    index += 1;
                }
            }
            b'\'' | b'"' => index = skip_quoted_string(bytes, index),
            b'(' | b'[' | b'{' => {
                let close_byte = match bytes[index] {
                    b'(' => b')',
                    b'[' => b']',
                    _ => b'}',
                };
                let Some(close) = matching_delimiter(bytes, index, close_byte) else {
                    index += 1;
                    continue;
                };
                if let Some(error) = invalid_comprehension_in_slice(bytes, index, close) {
                    return Some(error);
                }
                index = close + 1;
            }
            _ => index += 1,
        }
    }
    None
}

fn invalid_group_in_slice(bytes: &[u8], open: usize, close: usize) -> Option<CpythonDiagnostic> {
    let (item_start, item_end) = trim_target_range(bytes, open + 1, close);
    if item_start >= item_end
        || top_level_byte(bytes, item_start, item_end, b',').is_some()
        || top_level_colon(bytes, item_start, item_end).is_some()
        || find_keyword_at_level(bytes, item_start, item_end, b"for").is_some()
    {
        return None;
    }
    if bytes.get(item_start..item_start + 2) == Some(b"**") {
        return Some(CpythonDiagnostic::new(
            "cannot use double starred expression here".to_owned(),
            item_start,
            item_start + 2,
        ));
    }
    if bytes.get(item_start) == Some(&b'*') {
        return Some(CpythonDiagnostic::new(
            "cannot use starred expression here".to_owned(),
            item_start,
            item_start + 1,
        ));
    }
    None
}

fn invalid_group_error(source: &str) -> Option<CpythonDiagnostic> {
    let bytes = source.as_bytes();
    let mut index = 0usize;
    while index < bytes.len() {
        match bytes[index] {
            b'#' => {
                while index < bytes.len() && bytes[index] != b'\n' {
                    index += 1;
                }
            }
            b'\'' | b'"' => index = skip_quoted_string(bytes, index),
            b'(' => {
                let Some(close) = matching_delimiter(bytes, index, b')') else {
                    index += 1;
                    continue;
                };
                if let Some(error) = invalid_group_in_slice(bytes, index, close) {
                    return Some(error);
                }
                index = close + 1;
            }
            _ => index += 1,
        }
    }
    None
}

fn invalid_parameter_star_annotation_error(source: &str) -> Option<CpythonDiagnostic> {
    let bytes = source.as_bytes();
    let mut index = 0usize;
    while index < bytes.len() {
        match bytes[index] {
            b'#' => {
                while index < bytes.len() && bytes[index] != b'\n' {
                    index += 1;
                }
            }
            b'\'' | b'"' => index = skip_quoted_string(bytes, index),
            b'(' => {
                let Some(close) = matching_delimiter(bytes, index, b')') else {
                    index += 1;
                    continue;
                };
                let mut param_start = index + 1;
                while param_start < close {
                    let param_end =
                        find_byte_at_level(bytes, param_start, close, b',').unwrap_or(close);
                    if let Some(colon) = top_level_colon(bytes, param_start, param_end) {
                        let value_start = next_non_horizontal_whitespace(bytes, colon + 1);
                        if bytes.get(value_start) == Some(&b'*') {
                            return Some(CpythonDiagnostic::new(
                                "invalid syntax".to_owned(),
                                value_start,
                                value_start + 1,
                            ));
                        }
                    }
                    param_start = param_end.saturating_add(1);
                }
                index = close + 1;
            }
            _ => index += 1,
        }
    }
    None
}

fn invalid_def_type_params_error(source: &str) -> Option<CpythonDiagnostic> {
    let bytes = source.as_bytes();
    let mut index = 0;
    while index < bytes.len() {
        match bytes[index] {
            b'#' => {
                while index < bytes.len() && bytes[index] != b'\n' {
                    index += 1;
                }
            }
            b'\'' | b'"' => index = skip_quoted_string(bytes, index),
            _ if starts_identifier(bytes, index, b"def") => {
                let name_start = skip_horizontal_whitespace(bytes, index + 3);
                let name_end = identifier_end(bytes, name_start, bytes.len());
                let bracket = skip_horizontal_whitespace(bytes, name_end);
                if bytes.get(bracket) == Some(&b'[') {
                    let Some(close) = matching_delimiter(bytes, bracket, b']') else {
                        index = bracket + 1;
                        continue;
                    };
                    let after_close = skip_horizontal_whitespace(bytes, close + 1);
                    if bytes.get(after_close) == Some(&b'(')
                        && type_param_list_is_malformed(bytes, bracket + 1, close)
                    {
                        return Some(CpythonDiagnostic::new(
                            "expected '('".to_owned(),
                            bracket,
                            bracket + 1,
                        ));
                    }
                }
                index = name_end.max(index + 3);
            }
            _ => index += 1,
        }
    }
    None
}

fn type_param_list_is_malformed(bytes: &[u8], start: usize, end: usize) -> bool {
    let mut index = start;
    let mut expect_item = true;
    while index < end {
        index = skip_horizontal_whitespace(bytes, index);
        if index >= end {
            break;
        }
        if bytes[index] == b',' {
            if expect_item {
                return true;
            }
            expect_item = true;
            index += 1;
            continue;
        }
        if !expect_item {
            return true;
        }
        if bytes.get(index..index + 2) == Some(b"**") {
            index += 2;
        } else if bytes.get(index) == Some(&b'*') {
            index += 1;
        }
        let item_start = skip_horizontal_whitespace(bytes, index);
        let item_end = identifier_end(bytes, item_start, end);
        if item_end == item_start {
            return true;
        }
        index = item_end;
        if bytes.get(skip_horizontal_whitespace(bytes, index)) == Some(&b':') {
            index = skip_horizontal_whitespace(bytes, index) + 1;
            while index < end && bytes[index] != b',' {
                index = match bytes[index] {
                    b'\'' | b'"' => skip_quoted_string(bytes, index),
                    _ => index + 1,
                };
            }
        }
        expect_item = false;
    }
    false
}

fn invalid_parameter_list_slice_error(
    source: &str,
    start: usize,
    end: usize,
    kind: ParameterListKind,
) -> Option<CpythonDiagnostic> {
    let bytes = source.as_bytes();
    let mut index = start;
    let mut level = 0usize;
    let mut default_seen = false;
    let mut keyword_only = false;
    let mut slash_seen = false;
    let mut var_keyword_seen = false;
    while index < end {
        match bytes[index] {
            b'\'' | b'"' => index = skip_quoted_string(bytes, index),
            _ if level == 0
                && var_keyword_seen
                && bytes.get(index).is_some_and(|byte| {
                    *byte >= 0x80 || *byte == b'_' || byte.is_ascii_alphabetic()
                }) =>
            {
                let name_end = identifier_end(bytes, index, end);
                return Some(CpythonDiagnostic::new(
                    "arguments cannot follow var-keyword argument".to_owned(),
                    index,
                    name_end,
                ));
            }
            _ if level == 0
                && !keyword_only
                && bytes.get(index).is_some_and(|byte| {
                    *byte >= 0x80 || *byte == b'_' || byte.is_ascii_alphabetic()
                }) =>
            {
                let param_end = find_byte_at_level(bytes, index, end, b',')
                    .or_else(|| top_level_byte(bytes, index, end, b')'))
                    .or_else(|| {
                        matches!(kind, ParameterListKind::Lambda)
                            .then(|| top_level_byte(bytes, index, end, b':'))
                            .flatten()
                    })
                    .unwrap_or(end);
                let name_end = identifier_end(bytes, index, param_end);
                if let Some(eq) = top_level_byte(bytes, index, param_end, b'=') {
                    let value_start = next_non_horizontal_whitespace(bytes, eq + 1);
                    if value_start >= param_end
                        || matches!(bytes.get(value_start), Some(b',' | b')' | b':'))
                    {
                        let message = if matches!(kind, ParameterListKind::Lambda)
                            && matches!(bytes.get(value_start), Some(b':'))
                        {
                            "invalid syntax"
                        } else {
                            "expected default value expression"
                        };
                        return Some(CpythonDiagnostic::new(message.to_owned(), eq, eq + 1));
                    }
                    default_seen = true;
                } else if default_seen {
                    return Some(CpythonDiagnostic::new(
                        "parameter without a default follows parameter with a default".to_owned(),
                        index,
                        name_end,
                    ));
                }
                index = param_end;
            }
            b'(' if level == 0 => {
                let close = matching_delimiter(bytes, index, b')')
                    .filter(|close| *close <= end)
                    .unwrap_or(index + 1);
                let message = match kind {
                    ParameterListKind::Function => "Function parameters cannot be parenthesized",
                    ParameterListKind::Lambda => {
                        "Lambda expression parameters cannot be parenthesized"
                    }
                };
                return Some(CpythonDiagnostic::new(message.to_owned(), index, close + 1));
            }
            b'(' | b'[' | b'{' => {
                level += 1;
                index += 1;
            }
            b')' | b']' | b'}' => {
                level = level.saturating_sub(1);
                index += 1;
            }
            b'/' if level == 0 => {
                if var_keyword_seen {
                    return Some(CpythonDiagnostic::new(
                        "arguments cannot follow var-keyword argument".to_owned(),
                        index,
                        index + 1,
                    ));
                }
                if slash_seen {
                    return Some(CpythonDiagnostic::new(
                        "/ may appear only once".to_owned(),
                        index,
                        index + 1,
                    ));
                }
                slash_seen = true;
                let next = next_non_horizontal_whitespace(bytes, index + 1);
                if bytes.get(next) == Some(&b'*') {
                    return Some(CpythonDiagnostic::new(
                        "expected comma between / and *".to_owned(),
                        next,
                        next + 1,
                    ));
                }
                index += 1;
            }
            b'*' if level == 0 => {
                if var_keyword_seen {
                    return Some(CpythonDiagnostic::new(
                        "arguments cannot follow var-keyword argument".to_owned(),
                        index,
                        index + 1,
                    ));
                }
                keyword_only = true;
                let stars = usize::from(bytes.get(index + 1) == Some(&b'*')) + 1;
                let name_start = next_non_horizontal_whitespace(bytes, index + stars);
                for keyword in [b"True".as_slice(), b"False".as_slice(), b"None".as_slice()] {
                    if starts_identifier(bytes, name_start, keyword) {
                        return Some(CpythonDiagnostic::new(
                            "invalid syntax".to_owned(),
                            name_start,
                            name_start + keyword.len(),
                        ));
                    }
                }
                let param_end = find_byte_at_level(bytes, name_start, end, b',')
                    .or_else(|| top_level_byte(bytes, name_start, end, b')'))
                    .or_else(|| {
                        matches!(kind, ParameterListKind::Lambda)
                            .then(|| top_level_byte(bytes, name_start, end, b':'))
                            .flatten()
                    })
                    .unwrap_or(end);
                if stars == 1 && matches!(bytes.get(name_start), Some(b')' | b',' | b':')) {
                    return Some(CpythonDiagnostic::new(
                        "named arguments must follow bare *".to_owned(),
                        index,
                        index + 1,
                    ));
                }
                if stars == 1 && top_level_byte(bytes, name_start, param_end, b'=').is_some() {
                    return Some(CpythonDiagnostic::new(
                        "var-positional argument cannot have default value".to_owned(),
                        index,
                        index + 1,
                    ));
                }
                if stars == 2 && top_level_byte(bytes, name_start, param_end, b'=').is_some() {
                    return Some(CpythonDiagnostic::new(
                        "var-keyword argument cannot have default value".to_owned(),
                        index,
                        index + 2,
                    ));
                }
                if stars == 2 {
                    var_keyword_seen = true;
                    index = param_end;
                    continue;
                }
                index += stars;
            }
            b'=' if level == 0 => {
                let value_start = next_non_horizontal_whitespace(bytes, index + 1);
                if value_start >= end || matches!(bytes.get(value_start), Some(b',' | b')' | b':'))
                {
                    if matches!(kind, ParameterListKind::Lambda)
                        && matches!(bytes.get(value_start), Some(b':'))
                    {
                        return Some(CpythonDiagnostic::new(
                            "invalid syntax".to_owned(),
                            index,
                            index + 1,
                        ));
                    }
                    return Some(CpythonDiagnostic::new(
                        "expected default value expression".to_owned(),
                        index,
                        index + 1,
                    ));
                }
                index += 1;
            }
            _ => index += 1,
        }
    }
    None
}

fn invalid_parameter_list_error(source: &str) -> Option<CpythonDiagnostic> {
    let bytes = source.as_bytes();
    let mut index = 0usize;
    while index < bytes.len() {
        match bytes[index] {
            b'#' => {
                while index < bytes.len() && bytes[index] != b'\n' {
                    index += 1;
                }
            }
            b'\'' | b'"' => index = skip_quoted_string(bytes, index),
            _ if starts_identifier(bytes, index, b"def") => {
                let Some(paren) = top_level_byte(bytes, index + 3, bytes.len(), b'(') else {
                    index += 3;
                    continue;
                };
                let Some(close) = matching_delimiter(bytes, paren, b')') else {
                    index = paren + 1;
                    continue;
                };
                if let Some(error) = invalid_parameter_list_slice_error(
                    source,
                    paren + 1,
                    close,
                    ParameterListKind::Function,
                ) {
                    return Some(error);
                }
                index = close + 1;
            }
            _ if starts_identifier(bytes, index, b"lambda") => {
                let params_start = index + 6;
                let Some(params_end) = find_lambda_parameter_end(bytes, params_start) else {
                    index = params_start;
                    continue;
                };
                if let Some(error) = invalid_parameter_list_slice_error(
                    source,
                    params_start,
                    params_end,
                    ParameterListKind::Lambda,
                ) {
                    return Some(error);
                }
                index = params_end + 1;
            }
            _ => index += 1,
        }
    }
    None
}

#[derive(Clone, Copy)]
struct CallArgFrame {
    level: usize,
    arg_start: Option<usize>,
    in_call: bool,
}

fn next_non_horizontal_whitespace(bytes: &[u8], mut index: usize) -> usize {
    while matches!(bytes.get(index), Some(b' ' | b'\t' | b'\x0c')) {
        index += 1;
    }
    index
}

fn invalid_call_argument_assignment_error(
    source: &str,
    arg_start: usize,
    equal: usize,
) -> Option<CpythonDiagnostic> {
    let bytes = source.as_bytes();
    let start = bytes[arg_start..equal]
        .iter()
        .rposition(|byte| *byte == b'\n')
        .map_or(arg_start, |newline| arg_start + newline + 1);
    let (target_start, target_end) = trim_target_range(bytes, start, equal);
    if target_start >= target_end {
        return None;
    }
    let value_start = next_non_horizontal_whitespace(bytes, equal + 1);
    if matches!(bytes.get(value_start), None | Some(b',' | b')')) {
        return Some(CpythonDiagnostic::new(
            "expected argument value expression".to_owned(),
            target_start,
            equal + 1,
        ));
    }
    if bytes.get(target_start..target_start + 2) == Some(b"**") {
        return Some(CpythonDiagnostic::new(
            "cannot assign to keyword argument unpacking".to_owned(),
            target_start,
            value_start,
        ));
    }
    if bytes.get(target_start) == Some(&b'*') {
        return Some(CpythonDiagnostic::new(
            "cannot assign to iterable argument unpacking".to_owned(),
            target_start,
            value_start,
        ));
    }
    for keyword in [b"True".as_slice(), b"False".as_slice(), b"None".as_slice()] {
        if bytes.get(target_start..target_end) == Some(keyword) {
            let keyword = ::core::str::from_utf8(keyword).ok()?;
            return Some(CpythonDiagnostic::new(
                format!("cannot assign to {keyword}"),
                target_start,
                target_end,
            ));
        }
    }
    if is_simple_keyword_name(bytes, target_start, target_end) {
        // NAME '=' expression for_if_clauses
        if unparenthesized_comprehension(bytes, equal + 1) {
            return Some(CpythonDiagnostic::new(
                "invalid syntax. Maybe you meant '==' or ':=' instead of '='?".to_owned(),
                target_start,
                equal + 1,
            ));
        }
        return None;
    }
    Some(CpythonDiagnostic::new(
        "expression cannot contain assignment, perhaps you meant \"==\"?".to_owned(),
        target_start,
        equal,
    ))
}

fn unparenthesized_comprehension(bytes: &[u8], mut index: usize) -> bool {
    index = next_non_horizontal_whitespace(bytes, index);
    if index >= bytes.len() {
        return false;
    }
    let start = index;
    let mut level = 0usize;
    while index < bytes.len() {
        match bytes[index] {
            b'\'' | b'"' => index = skip_quoted_string(bytes, index),
            b'(' | b'[' | b'{' => {
                level += 1;
                index += 1;
            }
            b')' | b']' | b'}' => {
                if level == 0 {
                    return false;
                }
                level -= 1;
                index += 1;
            }
            b',' | b':' if level == 0 => return false,
            _ if level == 0 && index > start && starts_identifier(bytes, index, b"for") => {
                return true;
            }
            _ => index += 1,
        }
    }
    false
}

fn invalid_call_star_expression_error(
    bytes: &[u8],
    arg_start: usize,
    index: usize,
) -> Option<CpythonDiagnostic> {
    let start = next_non_horizontal_whitespace(bytes, arg_start);
    if start != index || bytes.get(index) != Some(&b'*') {
        return None;
    }
    let after_star = next_non_horizontal_whitespace(bytes, index + 1);
    if matches!(bytes.get(after_star), None | Some(b',' | b')' | b':')) {
        return Some(CpythonDiagnostic::new(
            "Invalid star expression".to_owned(),
            index,
            (index + 1).min(bytes.len()),
        ));
    }
    None
}

fn invalid_call_argument_error(source: &str) -> Option<CpythonDiagnostic> {
    let bytes = source.as_bytes();
    let mut index = 0usize;
    let mut level = 0usize;
    let mut frames: Vec<CallArgFrame> = Vec::new();
    while index < bytes.len() {
        match bytes[index] {
            b'#' => {
                while index < bytes.len() && bytes[index] != b'\n' {
                    index += 1;
                }
            }
            b'\'' | b'"' => index = skip_quoted_string(bytes, index),
            _ if starts_identifier(bytes, index, b"lambda") => {
                let params_start = index + 6;
                if let Some(params_end) = find_lambda_parameter_end(bytes, params_start) {
                    index = params_end + 1;
                } else {
                    index = params_start;
                }
            }
            b'(' => {
                level += 1;
                let in_call = opening_paren_is_call(bytes, index)
                    || frames.last().is_some_and(|frame| frame.in_call);
                frames.push(CallArgFrame {
                    level,
                    arg_start: (in_call && !is_function_parameter_list(bytes, index))
                        .then_some(index + 1),
                    in_call,
                });
                index += 1;
            }
            b')' => {
                if matches!(frames.last(), Some(frame) if frame.level == level) {
                    frames.pop();
                }
                level = level.saturating_sub(1);
                index += 1;
            }
            b'[' | b'{' => {
                level += 1;
                index += 1;
            }
            b']' | b'}' => {
                level = level.saturating_sub(1);
                index += 1;
            }
            b',' => {
                if let Some(frame) = frames.last_mut()
                    && frame.level == level
                    && frame.arg_start.is_some()
                {
                    frame.arg_start = Some(index + 1);
                }
                index += 1;
            }
            b'*' => {
                if let Some(CallArgFrame {
                    level: frame_level,
                    arg_start: Some(arg_start),
                    in_call: true,
                }) = frames.last().copied()
                    && frame_level == level
                    && let Some(error) = invalid_call_star_expression_error(bytes, arg_start, index)
                {
                    return Some(error);
                }
                index += 1;
            }
            b'=' if is_plain_assignment_operator(bytes, index) => {
                if let Some(CallArgFrame {
                    level: frame_level,
                    arg_start: Some(arg_start),
                    in_call: true,
                }) = frames.last().copied()
                    && frame_level == level
                    && let Some(error) =
                        invalid_call_argument_assignment_error(source, arg_start, index)
                {
                    return Some(error);
                }
                index += 1;
            }
            _ => index += 1,
        }
    }
    None
}

fn top_level_colon(bytes: &[u8], mut index: usize, end: usize) -> Option<usize> {
    let mut level = 0usize;
    while index < end {
        match bytes[index] {
            b'\'' | b'"' => index = skip_quoted_string(bytes, index),
            b'(' | b'[' | b'{' => {
                level += 1;
                index += 1;
            }
            b')' | b']' | b'}' => {
                level = level.saturating_sub(1);
                index += 1;
            }
            b':' if level == 0 => return Some(index),
            _ => index += 1,
        }
    }
    None
}

fn expression_slice_is_valid(source: &str, start: usize, end: usize) -> bool {
    let bytes = source.as_bytes();
    let (start, end) = trim_target_range(bytes, start, end);
    start < end
        && parser::parse(&source[start..end], parser::Mode::Expression.into())
            .is_ok_and(|parsed| matches!(parsed.into_syntax(), ast::Mod::Expression(_)))
}

fn invalid_dict_entry_error(
    source: &str,
    item_start: usize,
    item_end: usize,
    colon: Option<usize>,
    saw_dict_item: bool,
) -> Option<CpythonDiagnostic> {
    let bytes = source.as_bytes();
    let (item_start, item_end) = trim_target_range(bytes, item_start, item_end);
    if item_start >= item_end {
        return None;
    }
    if let Some(colon) = colon {
        let value_start = next_non_horizontal_whitespace(bytes, colon + 1);
        if value_start >= item_end {
            return Some(CpythonDiagnostic::new(
                "expression expected after dictionary key and ':'".to_owned(),
                colon,
                colon + 1,
            ));
        }
        if bytes.get(value_start) == Some(&b'*') {
            return Some(CpythonDiagnostic::new(
                "cannot use a starred expression in a dictionary value".to_owned(),
                value_start,
                value_start + 1,
            ));
        }
        if !expression_slice_is_valid(source, value_start, item_end) {
            return Some(CpythonDiagnostic::new(
                "invalid syntax".to_owned(),
                value_start,
                value_start,
            ));
        }
    } else if saw_dict_item {
        return Some(CpythonDiagnostic::new(
            "':' expected after dictionary key".to_owned(),
            item_end.saturating_sub(1),
            item_end,
        ));
    }
    None
}

fn invalid_dict_literal_error(
    source: &str,
    open: usize,
    close: usize,
) -> Option<CpythonDiagnostic> {
    let bytes = source.as_bytes();
    let mut item_start = open + 1;
    let mut index = item_start;
    let mut level = 0usize;
    let mut saw_dict_item = false;
    let mut item_colon = None;
    while index <= close {
        if index == close || (level == 0 && bytes.get(index) == Some(&b',')) {
            if let Some(error) =
                invalid_dict_entry_error(source, item_start, index, item_colon, saw_dict_item)
            {
                return Some(error);
            }
            saw_dict_item |= item_colon.is_some();
            item_start = index + 1;
            item_colon = None;
            index += 1;
            continue;
        }
        match bytes[index] {
            b'\'' | b'"' => index = skip_quoted_string(bytes, index),
            b'(' | b'[' | b'{' => {
                level += 1;
                index += 1;
            }
            b')' | b']' | b'}' => {
                level = level.saturating_sub(1);
                index += 1;
            }
            b':' if level == 0 && item_colon.is_none() => {
                item_colon = Some(index);
                saw_dict_item = true;
                index += 1;
            }
            _ => index += 1,
        }
    }
    None
}

fn invalid_dict_error(source: &str) -> Option<CpythonDiagnostic> {
    let bytes = source.as_bytes();
    let mut index = 0usize;
    while index < bytes.len() {
        match bytes[index] {
            b'#' => {
                while index < bytes.len() && bytes[index] != b'\n' {
                    index += 1;
                }
            }
            b'\'' | b'"' => index = skip_quoted_string(bytes, index),
            b'{' => {
                let Some(close) = matching_delimiter(bytes, index, b'}') else {
                    index += 1;
                    continue;
                };
                if top_level_colon(bytes, index + 1, close).is_some()
                    && let Some(error) = invalid_dict_literal_error(source, index, close)
                {
                    return Some(error);
                }
                index = close + 1;
            }
            _ => index += 1,
        }
    }
    None
}

fn collection_open_is_call(bytes: &[u8], open: usize) -> bool {
    if bytes.get(open) != Some(&b'(') {
        return false;
    }
    let mut cursor = open;
    while cursor > 0 && matches!(bytes.get(cursor - 1), Some(b' ' | b'\t' | b'\x0c')) {
        cursor -= 1;
    }
    matches!(
        cursor.checked_sub(1).and_then(|before| bytes.get(before)),
        Some(b')' | b']' | b'_' | b'a'..=b'z' | b'A'..=b'Z' | 0x80..=0xff)
    )
}

fn invalid_collection_assignment_in_slice(
    source: &str,
    bytes: &[u8],
    start: usize,
    end: usize,
) -> Option<CpythonDiagnostic> {
    let mut item_start = start;
    let mut index = start;
    let mut level = 0usize;
    while index < end {
        match bytes[index] {
            b'\'' | b'"' => index = skip_quoted_string(bytes, index),
            b'(' | b'[' | b'{' => {
                level += 1;
                index += 1;
            }
            b')' | b']' | b'}' => {
                level = level.saturating_sub(1);
                index += 1;
            }
            b',' if level == 0 => {
                item_start = index + 1;
                index += 1;
            }
            b'=' if level == 0 && is_plain_assignment_operator(bytes, index) => {
                if top_level_colon(bytes, item_start, index).is_none() {
                    let start = next_non_horizontal_whitespace(bytes, item_start);
                    let target_end = trim_end_horizontal_whitespace(bytes, start, index);
                    if start < target_end
                        && let Some((expr_name, expr_start, expr_end, _)) =
                            expression_name_and_range(&source[start..target_end])
                    {
                        if matches!(expr_name, "list" | "tuple") {
                            return None;
                        }
                        if matches!(expr_name, "expression" | "attribute" | "subscript") {
                            return Some(CpythonDiagnostic::new(
                                format!(
                                    "cannot assign to {expr_name} here. Maybe you meant '==' instead of '='?"
                                ),
                                start + expr_start,
                                start + expr_end,
                            ));
                        }
                    }
                    return Some(CpythonDiagnostic::new(
                        "invalid syntax. Maybe you meant '==' or ':=' instead of '='?".to_owned(),
                        start,
                        index + 1,
                    ));
                }
                index += 1;
            }
            _ => index += 1,
        }
    }
    None
}

fn invalid_collection_assignment_error(source: &str) -> Option<CpythonDiagnostic> {
    let bytes = source.as_bytes();
    let mut index = 0usize;
    while index < bytes.len() {
        match bytes[index] {
            b'#' => {
                while index < bytes.len() && bytes[index] != b'\n' {
                    index += 1;
                }
            }
            b'\'' | b'"' => index = skip_quoted_string(bytes, index),
            b'(' | b'[' | b'{' => {
                let close_byte = match bytes[index] {
                    b'(' => b')',
                    b'[' => b']',
                    _ => b'}',
                };
                let Some(close) = matching_delimiter(bytes, index, close_byte) else {
                    index += 1;
                    continue;
                };
                if !collection_open_is_call(bytes, index)
                    && let Some(error) =
                        invalid_collection_assignment_in_slice(source, bytes, index + 1, close)
                {
                    return Some(error);
                }
                index = close + 1;
            }
            _ => index += 1,
        }
    }
    None
}

fn expression_assignment_error(source: &str) -> Option<CpythonDiagnostic> {
    let bytes = source.as_bytes();
    let mut index = 0;
    let mut paren_arg_starts: Vec<(Option<usize>, bool)> = Vec::new();
    while index < bytes.len() {
        match bytes[index] {
            b'#' => {
                while index < bytes.len() && bytes[index] != b'\n' {
                    index += 1;
                }
            }
            b'\'' | b'"' => {
                index = skip_quoted_string(bytes, index);
            }
            _ if starts_identifier(bytes, index, b"lambda") => {
                let params_start = index + 6;
                if let Some(params_end) = find_lambda_parameter_end(bytes, params_start) {
                    index = params_end + 1;
                } else {
                    index = params_start;
                }
            }
            b'(' => {
                let in_call_context = opening_paren_is_call(bytes, index)
                    || paren_arg_starts.last().is_some_and(|(_, in_call)| *in_call);
                paren_arg_starts.push((
                    (!is_function_parameter_list(bytes, index)).then_some(index + 1),
                    in_call_context,
                ));
                index += 1;
            }
            b')' => {
                paren_arg_starts.pop();
                index += 1;
            }
            b',' => {
                if let Some((start, _)) = paren_arg_starts.last_mut()
                    && start.is_some()
                {
                    *start = Some(index + 1);
                }
                index += 1;
            }
            b'=' if is_plain_assignment_operator(bytes, index) => {
                if let Some((Some(start), true)) = paren_arg_starts.last().copied()
                    && !is_simple_keyword_name(bytes, start, index)
                {
                    let mut expr_start = start;
                    while matches!(bytes.get(expr_start), Some(b' ' | b'\t' | b'\x0c')) {
                        expr_start += 1;
                    }
                    return Some(CpythonDiagnostic::new(
                        "expression cannot contain assignment, perhaps you meant \"==\"?"
                            .to_owned(),
                        expr_start,
                        index,
                    ));
                }
                index += 1;
            }
            _ => index += 1,
        }
    }
    None
}

fn invalid_named_expression_error(source: &str) -> Option<CpythonDiagnostic> {
    let bytes = source.as_bytes();
    let mut index = 0;
    while index + 1 < bytes.len() {
        match bytes[index] {
            b'#' => {
                while index < bytes.len() && bytes[index] != b'\n' {
                    index += 1;
                }
            }
            b'\'' | b'"' => index = skip_quoted_string(bytes, index),
            b':' if bytes.get(index + 1) == Some(&b'=') => {
                let target_start = named_expression_target_start(bytes, index);
                let target_end = trim_end_horizontal_whitespace(bytes, target_start, index);
                if target_start < target_end
                    && let Some((expr_name, start, end, is_name)) =
                        expression_name_and_range(&source[target_start..target_end])
                    && !is_name
                {
                    return Some(CpythonDiagnostic::new(
                        format!("cannot use assignment expressions with {expr_name}"),
                        target_start + start,
                        target_start + end,
                    ));
                }
                index += 2;
            }
            _ => index += 1,
        }
    }
    None
}

#[derive(Clone, Copy)]
struct AssignmentContext {
    start: usize,
    call: bool,
}

fn invalid_plain_assignment_error(source: &str) -> Option<CpythonDiagnostic> {
    let bytes = source.as_bytes();
    let mut stack: Vec<AssignmentContext> = Vec::new();
    let mut index = 0;
    while index < bytes.len() {
        match bytes[index] {
            b'#' => {
                while index < bytes.len() && bytes[index] != b'\n' {
                    index += 1;
                }
            }
            b'\'' | b'"' => index = skip_quoted_string(bytes, index),
            b'(' | b'[' | b'{' => {
                stack.push(AssignmentContext {
                    start: index + 1,
                    call: bytes[index] == b'(' && opening_paren_is_call(bytes, index),
                });
                index += 1;
            }
            b')' | b']' | b'}' => {
                stack.pop();
                index += 1;
            }
            b',' => {
                if let Some(context) = stack.last_mut()
                    && !context.call
                {
                    context.start = index + 1;
                }
                index += 1;
            }
            b'=' if is_plain_assignment_operator(bytes, index) => {
                if let Some(context) = stack.last().copied()
                    && !context.call
                {
                    let target_start = skip_horizontal_whitespace(bytes, context.start);
                    let target_end = trim_end_horizontal_whitespace(bytes, target_start, index);
                    if target_start < target_end
                        && let Some((expr_name, start, end, _)) =
                            expression_name_and_range(&source[target_start..target_end])
                        && matches!(expr_name, "expression" | "attribute" | "subscript")
                    {
                        return Some(CpythonDiagnostic::new(
                            format!(
                                "cannot assign to {expr_name} here. Maybe you meant '==' instead of '='?"
                            ),
                            target_start + start,
                            target_start + end,
                        ));
                    }
                }
                index += 1;
            }
            _ => index += 1,
        }
    }
    None
}

fn opening_paren_is_call(bytes: &[u8], paren: usize) -> bool {
    let mut cursor = paren;
    while cursor > 0 && matches!(bytes[cursor - 1], b' ' | b'\t' | b'\x0c') {
        cursor -= 1;
    }
    cursor > 0
        && (bytes[cursor - 1] >= 0x80
            || is_ascii_identifier_char(bytes[cursor - 1])
            || matches!(bytes[cursor - 1], b')' | b']'))
}

fn named_expression_target_start(bytes: &[u8], walrus: usize) -> usize {
    let mut index = walrus;
    let mut level = 0usize;
    while index > 0 {
        index -= 1;
        match bytes[index] {
            b')' | b']' | b'}' => level += 1,
            b'(' | b'[' | b'{' if level > 0 => level -= 1,
            b'(' | b'[' | b'{' if level == 0 => return index + 1,
            b',' | b'\n' | b';' if level == 0 => return index + 1,
            _ => {}
        }
    }
    0
}

fn trim_end_horizontal_whitespace(bytes: &[u8], start: usize, mut end: usize) -> usize {
    while end > start && matches!(bytes[end - 1], b' ' | b'\t' | b'\x0c') {
        end -= 1;
    }
    end
}

fn annotation_target_error_for_slice(
    source: &str,
    start: usize,
    colon: usize,
) -> Option<CpythonDiagnostic> {
    let bytes = source.as_bytes();
    let (target_start, target_end) = trim_target_range(bytes, start, colon);
    if target_start >= target_end {
        return None;
    }
    let target_text = &source[target_start..target_end];
    let Ok(parsed) = parser::parse(target_text, parser::Mode::Expression.into()) else {
        return None;
    };
    let ast::Mod::Expression(expression) = parsed.into_syntax() else {
        return None;
    };
    match expression.body.as_ref() {
        ast::Expr::Name(_) | ast::Expr::Attribute(_) | ast::Expr::Subscript(_) => None,
        ast::Expr::List(_) => Some(CpythonDiagnostic::new(
            "only single target (not list) can be annotated".to_owned(),
            target_start,
            target_end,
        )),
        ast::Expr::Tuple(_) => Some(CpythonDiagnostic::new(
            "only single target (not tuple) can be annotated".to_owned(),
            target_start,
            target_end,
        )),
        _ => Some(CpythonDiagnostic::new(
            "illegal target for annotation".to_owned(),
            target_start,
            target_end,
        )),
    }
}

fn invalid_annotation_line_start(bytes: &[u8], line_start: usize) -> bool {
    let column = skip_horizontal_whitespace(bytes, line_start);
    for keyword in [
        b"async".as_slice(),
        b"case",
        b"class",
        b"def",
        b"elif",
        b"else",
        b"except",
        b"finally",
        b"for",
        b"if",
        b"match",
        b"try",
        b"while",
        b"with",
    ] {
        if starts_identifier(bytes, column, keyword) {
            return false;
        }
    }
    true
}

fn invalid_annotation_target_error(source: &str) -> Option<CpythonDiagnostic> {
    let bytes = source.as_bytes();
    let mut line_start = 0usize;
    for line in source.split_inclusive('\n') {
        let line_end = line_start + line.len();
        if invalid_annotation_line_start(bytes, line_start)
            && let Some(colon) = find_byte_at_level(bytes, line_start, line_end, b':')
            && bytes.get(colon + 1) != Some(&b'=')
            && colon.checked_sub(1).and_then(|before| bytes.get(before)) != Some(&b':')
            && let Some(error) = annotation_target_error_for_slice(source, line_start, colon)
        {
            return Some(error);
        }
        line_start = line_end;
    }
    None
}

fn statement_target_end(bytes: &[u8], mut index: usize) -> usize {
    let mut level = 0usize;
    while index < bytes.len() {
        match bytes[index] {
            b'#' if level == 0 => return index,
            b'\n' | b';' if level == 0 => return index,
            b'\'' | b'"' => {
                index = skip_quoted_string(bytes, index);
            }
            b'(' | b'[' | b'{' => {
                level += 1;
                index += 1;
            }
            b')' | b']' | b'}' => {
                level = level.saturating_sub(1);
                index += 1;
            }
            _ => index += 1,
        }
    }
    index
}

fn invalid_assignment_target(expression: &ast::Expr) -> Option<&ast::Expr> {
    match expression {
        ast::Expr::List(ast::ExprList { elts, .. })
        | ast::Expr::Tuple(ast::ExprTuple { elts, .. }) => {
            elts.iter().find_map(invalid_assignment_target)
        }
        ast::Expr::Starred(ast::ExprStarred { value, .. }) => invalid_assignment_target(value),
        ast::Expr::Name(_) | ast::Expr::Subscript(_) | ast::Expr::Attribute(_) => None,
        _ => Some(expression),
    }
}

fn invalid_for_target(expression: &ast::Expr) -> Option<&ast::Expr> {
    match expression {
        ast::Expr::List(ast::ExprList { elts, .. })
        | ast::Expr::Tuple(ast::ExprTuple { elts, .. }) => elts.iter().find_map(invalid_for_target),
        ast::Expr::Starred(ast::ExprStarred { value, .. }) => invalid_for_target(value),
        ast::Expr::Compare(ast::ExprCompare { left, ops, .. }) => {
            if matches!(ops.first(), Some(ast::CmpOp::In)) {
                invalid_for_target(left)
            } else {
                None
            }
        }
        ast::Expr::Name(_) | ast::Expr::Subscript(_) | ast::Expr::Attribute(_) => None,
        _ => Some(expression),
    }
}

fn invalid_delete_target(expression: &ast::Expr) -> Option<&ast::Expr> {
    match expression {
        ast::Expr::List(ast::ExprList { elts, .. })
        | ast::Expr::Tuple(ast::ExprTuple { elts, .. }) => {
            elts.iter().find_map(invalid_delete_target)
        }
        ast::Expr::Name(_) | ast::Expr::Subscript(_) | ast::Expr::Attribute(_) => None,
        ast::Expr::Starred(_) => Some(expression),
        ast::Expr::Compare(_) => Some(expression),
        _ => Some(expression),
    }
}

fn delete_target_expr_name(expression: &ast::Expr) -> &'static str {
    match expression {
        ast::Expr::Attribute(_) => "attribute",
        ast::Expr::Subscript(_) => "subscript",
        ast::Expr::Starred(_) => "starred",
        ast::Expr::Name(_) => "name",
        ast::Expr::List(_) => "list",
        ast::Expr::Tuple(_) => "tuple",
        ast::Expr::Lambda(_) => "lambda",
        ast::Expr::Call(_) => "function call",
        ast::Expr::BoolOp(_) | ast::Expr::BinOp(_) | ast::Expr::UnaryOp(_) => "expression",
        ast::Expr::Generator(_) => "generator expression",
        ast::Expr::Yield(_) | ast::Expr::YieldFrom(_) => "yield expression",
        ast::Expr::Await(_) => "await expression",
        ast::Expr::ListComp(_) => "list comprehension",
        ast::Expr::SetComp(_) => "set comprehension",
        ast::Expr::DictComp(_) => "dict comprehension",
        ast::Expr::Dict(_) => "dict literal",
        ast::Expr::Set(_) => "set display",
        ast::Expr::FString(_) => "f-string expression",
        ast::Expr::TString(_) => "t-string expression",
        ast::Expr::NumberLiteral(_) | ast::Expr::StringLiteral(_) | ast::Expr::BytesLiteral(_) => {
            "literal"
        }
        ast::Expr::Constant(expr) => match &expr.value {
            ast::ConstantValue::None => "None",
            ast::ConstantValue::Boolean(true) => "True",
            ast::ConstantValue::Boolean(false) => "False",
            ast::ConstantValue::Ellipsis => "ellipsis",
            ast::ConstantValue::Tuple(_) => "tuple",
            ast::ConstantValue::Frozenset(_) => "literal",
            ast::ConstantValue::Str(_)
            | ast::ConstantValue::Bytes(_)
            | ast::ConstantValue::Integer(_)
            | ast::ConstantValue::Float(_)
            | ast::ConstantValue::Complex { .. } => "literal",
        },
        ast::Expr::BooleanLiteral(boolean) => {
            if boolean.value {
                "True"
            } else {
                "False"
            }
        }
        ast::Expr::NoneLiteral(_) => "None",
        ast::Expr::EllipsisLiteral(_) => "ellipsis",
        ast::Expr::Compare(_) => "comparison",
        ast::Expr::If(_) => "conditional expression",
        ast::Expr::Named(_) => "named expression",
        ast::Expr::Slice(_) | ast::Expr::IpyEscapeCommand(_) => "expression",
    }
}

fn parenthesized_single_starred_delete_target(bytes: &[u8], start: usize, end: usize) -> bool {
    let mut cursor = start;
    while matches!(bytes.get(cursor), Some(b' ' | b'\t' | b'\x0c')) {
        cursor += 1;
    }
    if bytes.get(cursor) != Some(&b'(') {
        return false;
    }
    cursor += 1;
    while matches!(bytes.get(cursor), Some(b' ' | b'\t' | b'\x0c')) {
        cursor += 1;
    }
    if bytes.get(cursor) != Some(&b'*') {
        return false;
    }
    let mut level = 1usize;
    cursor += 1;
    while cursor < end {
        match bytes[cursor] {
            b'\'' | b'"' => {
                cursor = skip_quoted_string(bytes, cursor);
            }
            b'(' | b'[' | b'{' => {
                level += 1;
                cursor += 1;
            }
            b')' => {
                level = level.saturating_sub(1);
                if level == 0 {
                    cursor += 1;
                    while matches!(bytes.get(cursor), Some(b' ' | b'\t' | b'\x0c')) {
                        cursor += 1;
                    }
                    return cursor == end;
                }
                cursor += 1;
            }
            b',' if level == 1 => return false,
            b']' | b'}' => {
                level = level.saturating_sub(1);
                cursor += 1;
            }
            _ => cursor += 1,
        }
    }
    false
}

fn assignment_target_expr_range(source: &str, start: usize, end: usize) -> Option<(usize, usize)> {
    let bytes = source.as_bytes();
    let (target_start, target_end) = trim_target_range(bytes, start, end);
    if target_start >= target_end {
        return None;
    }
    if parser::parse(
        &source[target_start..target_end],
        parser::Mode::Expression.into(),
    )
    .is_ok()
    {
        return Some((target_start, target_end));
    }
    // `def f(): (yield bar)` — skip the suite header so the remaining
    // text is the assignment target expression. Other colons (`x: int += 1`)
    // are not suite headers for this diagnostic. Grouping parentheses may
    // wrap the yield expression.
    let colon = top_level_colon(bytes, target_start, target_end)?;
    let after = skip_horizontal_whitespace(bytes, colon + 1);
    let yield_at = skip_opening_parentheses(bytes, after, target_end);
    if yield_at >= target_end || !starts_identifier(bytes, yield_at, b"yield") {
        return None;
    }
    Some((after, target_end))
}

fn skip_opening_parentheses(bytes: &[u8], mut index: usize, end: usize) -> usize {
    loop {
        index = skip_horizontal_whitespace(bytes, index);
        if index >= end || bytes[index] != b'(' {
            return index;
        }
        index += 1;
    }
}

fn trim_target_range(bytes: &[u8], mut start: usize, mut end: usize) -> (usize, usize) {
    while start < end
        && matches!(
            bytes.get(start),
            Some(b' ' | b'\t' | b'\n' | b'\r' | b'\x0c')
        )
    {
        start += 1;
    }
    while end > start
        && matches!(
            bytes.get(end - 1),
            Some(b' ' | b'\t' | b'\n' | b'\r' | b'\x0c')
        )
    {
        end -= 1;
    }
    (start, end)
}

fn invalid_assignment_message(name: &'static str, top_level_bitwise: bool) -> String {
    if top_level_bitwise {
        format!("cannot assign to {name} here. Maybe you meant '==' instead of '='?")
    } else {
        format!("cannot assign to {name}")
    }
}

fn assignment_target_error_for_slice(
    source: &str,
    start: usize,
    end: usize,
) -> Option<CpythonDiagnostic> {
    let bytes = source.as_bytes();
    let (target_start, target_end) = assignment_target_expr_range(source, start, end)?;
    if starts_identifier(bytes, target_start, b"yield") {
        return Some(CpythonDiagnostic::new(
            "assignment to yield expression not possible".to_owned(),
            target_start,
            target_start + 5,
        ));
    }
    let target_text = &source[target_start..target_end];
    let Ok(parsed) = parser::parse(target_text, parser::Mode::Expression.into()) else {
        return None;
    };
    let ast::Mod::Expression(expression) = parsed.into_syntax() else {
        return None;
    };
    let invalid_target = invalid_assignment_target(&expression.body)?;
    let invalid_start = target_start + invalid_target.range().start().to_usize();
    let invalid_end = target_start + invalid_target.range().end().to_usize();
    let name = delete_target_expr_name(invalid_target);
    let top_level = invalid_target.range() == expression.body.range();
    let bitwise_like = matches!(
        invalid_target,
        ast::Expr::Call(_)
            | ast::Expr::BoolOp(_)
            | ast::Expr::BinOp(_)
            | ast::Expr::UnaryOp(_)
            | ast::Expr::NumberLiteral(_)
            | ast::Expr::StringLiteral(_)
            | ast::Expr::BytesLiteral(_)
            | ast::Expr::EllipsisLiteral(_)
            | ast::Expr::Yield(_)
            | ast::Expr::YieldFrom(_)
            | ast::Expr::Set(_)
            | ast::Expr::Dict(_)
            | ast::Expr::FString(_)
            | ast::Expr::TString(_)
    );
    Some(CpythonDiagnostic::new(
        invalid_assignment_message(name, top_level && bitwise_like),
        invalid_start,
        invalid_end,
    ))
}

fn star_target_error_for_slice(
    source: &str,
    start: usize,
    end: usize,
) -> Option<CpythonDiagnostic> {
    invalid_target_error_for_slice(source, start, end, invalid_assignment_target)
}

fn for_target_error_for_slice(source: &str, start: usize, end: usize) -> Option<CpythonDiagnostic> {
    invalid_target_error_for_slice(source, start, end, invalid_for_target)
}

fn invalid_target_error_for_slice(
    source: &str,
    start: usize,
    end: usize,
    invalid_target: for<'a> fn(&'a ast::Expr) -> Option<&'a ast::Expr>,
) -> Option<CpythonDiagnostic> {
    let bytes = source.as_bytes();
    let (target_start, target_end) = trim_target_range(bytes, start, end);
    if target_start >= target_end {
        return None;
    }
    let target_text = &source[target_start..target_end];
    let Ok(parsed) = parser::parse(target_text, parser::Mode::Expression.into()) else {
        return None;
    };
    let ast::Mod::Expression(expression) = parsed.into_syntax() else {
        return None;
    };
    let invalid_target = invalid_target(&expression.body)?;
    let name = delete_target_expr_name(invalid_target);
    let invalid_start = target_start + invalid_target.range().start().to_usize();
    let invalid_end = target_start + invalid_target.range().end().to_usize();
    Some(CpythonDiagnostic::new(
        format!("cannot assign to {name}"),
        invalid_start,
        invalid_end,
    ))
}

fn first_compare_operator_at_level(bytes: &[u8], mut index: usize, end: usize) -> Option<usize> {
    let mut level = 0usize;
    while index < end {
        match bytes[index] {
            b'\'' | b'"' => index = skip_quoted_string(bytes, index),
            b'(' | b'[' | b'{' => {
                level += 1;
                index += 1;
            }
            b')' | b']' | b'}' => {
                level = level.saturating_sub(1);
                index += 1;
            }
            b'<' | b'>' if level == 0 => return Some(index),
            b'=' if level == 0 && bytes.get(index + 1) == Some(&b'=') => return Some(index),
            b'!' if level == 0 && bytes.get(index + 1) == Some(&b'=') => return Some(index),
            _ if level == 0 && starts_identifier(bytes, index, b"is") => return Some(index),
            _ if level == 0 && starts_identifier(bytes, index, b"not") => return Some(index),
            _ => index += 1,
        }
    }
    None
}

fn non_in_compare_for_target_error(
    source: &str,
    start: usize,
    end: usize,
) -> Option<CpythonDiagnostic> {
    let bytes = source.as_bytes();
    let (target_start, target_end) = trim_target_range(bytes, start, end);
    if target_start >= target_end {
        return None;
    }
    let target_text = &source[target_start..target_end];
    let Ok(parsed) = parser::parse(target_text, parser::Mode::Expression.into()) else {
        return None;
    };
    let ast::Mod::Expression(expression) = parsed.into_syntax() else {
        return None;
    };
    let ast::Expr::Compare(ast::ExprCompare { ops, .. }) = expression.body.as_ref() else {
        return None;
    };
    if matches!(ops.first(), Some(ast::CmpOp::In)) {
        return None;
    }
    let operator = first_compare_operator_at_level(bytes, target_start, target_end)?;
    Some(CpythonDiagnostic::new(
        "invalid syntax".to_owned(),
        operator,
        (operator + 1).min(target_end),
    ))
}

fn top_level_plain_assignment_offsets(bytes: &[u8]) -> Vec<usize> {
    let mut offsets = Vec::new();
    let mut index = 0usize;
    let mut level = 0usize;
    while index < bytes.len() {
        match bytes[index] {
            b'#' if level == 0 => {
                while index < bytes.len() && bytes[index] != b'\n' {
                    index += 1;
                }
            }
            b'\'' | b'"' => index = skip_quoted_string(bytes, index),
            b'(' | b'[' | b'{' => {
                level += 1;
                index += 1;
            }
            b')' | b']' | b'}' => {
                level = level.saturating_sub(1);
                index += 1;
            }
            b'=' if level == 0 && is_plain_assignment_operator(bytes, index) => {
                offsets.push(index);
                index += 1;
            }
            _ => index += 1,
        }
    }
    offsets
}

fn invalid_condition_assignment_error(
    source: &str,
    parse_error_offset: usize,
) -> Option<CpythonDiagnostic> {
    let bytes = source.as_bytes();
    let mut index = 0usize;
    let mut line_start = 0usize;
    let mut level = 0usize;
    while index < bytes.len() {
        match bytes[index] {
            b'#' if level == 0 => {
                while index < bytes.len() && bytes[index] != b'\n' {
                    index += 1;
                }
            }
            b'\n' => {
                line_start = index + 1;
                index += 1;
            }
            b'\'' | b'"' => index = skip_quoted_string(bytes, index),
            b'(' | b'[' | b'{' => {
                level += 1;
                index += 1;
            }
            b')' | b']' | b'}' => {
                level = level.saturating_sub(1);
                index += 1;
            }
            _ if level == 0
                && index == skip_horizontal_whitespace(bytes, line_start)
                && (starts_identifier(bytes, index, b"if")
                    || starts_identifier(bytes, index, b"elif")
                    || starts_identifier(bytes, index, b"while")) =>
            {
                let keyword_len = if starts_identifier(bytes, index, b"while") {
                    5
                } else if starts_identifier(bytes, index, b"elif") {
                    4
                } else {
                    2
                };
                let cond_start = skip_horizontal_whitespace(bytes, index + keyword_len);
                let Some(colon) = top_level_colon(bytes, cond_start, bytes.len()) else {
                    index += keyword_len;
                    continue;
                };
                if parse_error_offset < index || parse_error_offset > colon {
                    index += keyword_len;
                    continue;
                }
                let Some(equal) = condition_plain_assignment(bytes, cond_start, colon) else {
                    index += keyword_len;
                    continue;
                };
                let (target_start, target_end) = trim_target_range(bytes, cond_start, equal);
                if target_start >= target_end {
                    index += keyword_len;
                    continue;
                }
                if is_simple_keyword_name(bytes, target_start, target_end) {
                    return Some(CpythonDiagnostic::new(
                        "invalid syntax. Maybe you meant '==' or ':=' instead of '='?".to_owned(),
                        target_start,
                        equal + 1,
                    ));
                }
                if let Some((expr_name, start, end, _)) =
                    expression_name_and_range(&source[target_start..target_end])
                {
                    return Some(CpythonDiagnostic::new(
                        format!(
                            "cannot assign to {expr_name} here. Maybe you meant '==' instead of '='?"
                        ),
                        target_start + start,
                        target_start + end,
                    ));
                }
                return Some(CpythonDiagnostic::new(
                    "invalid syntax. Maybe you meant '==' or ':=' instead of '='?".to_owned(),
                    target_start,
                    equal + 1,
                ));
            }
            _ => index += 1,
        }
    }
    None
}

fn condition_plain_assignment(bytes: &[u8], start: usize, end: usize) -> Option<usize> {
    let mut index = start;
    let mut nest = Vec::new();
    while index < end {
        match bytes[index] {
            b'#' => {
                while index < end && bytes[index] != b'\n' {
                    index += 1;
                }
            }
            b'\'' | b'"' => index = skip_quoted_string(bytes, index),
            b'(' => {
                nest.push(if is_call_open(bytes, start, index) {
                    b'c'
                } else {
                    b'g'
                });
                index += 1;
            }
            b'[' => {
                nest.push(b'[');
                index += 1;
            }
            b'{' => {
                nest.push(b'{');
                index += 1;
            }
            b')' | b']' | b'}' => {
                nest.pop();
                index += 1;
            }
            b':' if nest.last() == Some(&b'l') => {
                nest.pop();
                index += 1;
            }
            b'=' if is_plain_assignment_operator(bytes, index)
                && !nest.contains(&b'c')
                && !nest.contains(&b'l') =>
            {
                return Some(index);
            }
            _ if starts_identifier(bytes, index, b"lambda") => {
                nest.push(b'l');
                index += 6;
            }
            _ => index += 1,
        }
    }
    None
}

fn is_call_open(bytes: &[u8], start: usize, open: usize) -> bool {
    let mut index = open;
    while index > start {
        index -= 1;
        match bytes[index] {
            b' ' | b'\t' | b'\n' | b'\r' | b'\x0c' => {}
            b')' | b']' => return true,
            byte if byte >= 0x80 || is_ascii_identifier_char(byte) => {
                let mut ident_start = index;
                while ident_start > start
                    && bytes
                        .get(ident_start - 1)
                        .is_some_and(|b| *b >= 0x80 || is_ascii_identifier_char(*b))
                {
                    ident_start -= 1;
                }
                return !is_condition_keyword(&bytes[ident_start..=index]);
            }
            _ => return false,
        }
    }
    false
}

fn is_condition_keyword(word: &[u8]) -> bool {
    matches!(
        word,
        b"not" | b"and" | b"or" | b"in" | b"is" | b"if" | b"elif" | b"while" | b"await" | b"lambda"
    )
}

fn invalid_assignment_target_error(source: &str) -> Option<CpythonDiagnostic> {
    let bytes = source.as_bytes();
    let offsets = top_level_plain_assignment_offsets(bytes);
    if offsets.is_empty() {
        return None;
    }
    let mut start = 0usize;
    for offset in offsets {
        if let Some(error) = assignment_target_error_for_slice(source, start, offset) {
            return Some(error);
        }
        start = offset + 1;
    }
    None
}

fn top_level_augassign_offset(bytes: &[u8]) -> Option<(usize, usize)> {
    let mut index = 0usize;
    let mut level = 0usize;
    while index < bytes.len() {
        match bytes[index] {
            b'#' if level == 0 => {
                while index < bytes.len() && bytes[index] != b'\n' {
                    index += 1;
                }
            }
            b'\'' | b'"' => index = skip_quoted_string(bytes, index),
            b'(' | b'[' | b'{' => {
                level += 1;
                index += 1;
            }
            b')' | b']' | b'}' => {
                level = level.saturating_sub(1);
                index += 1;
            }
            b'+' | b'-' | b'*' | b'@' | b'/' | b'%' | b'&' | b'|' | b'^'
                if level == 0 && bytes.get(index + 1) == Some(&b'=') =>
            {
                return Some((index, 2));
            }
            b'<' | b'>'
                if level == 0
                    && bytes.get(index + 1) == Some(&bytes[index])
                    && bytes.get(index + 2) == Some(&b'=') =>
            {
                return Some((index, 3));
            }
            b'*' if level == 0
                && bytes.get(index + 1) == Some(&b'*')
                && bytes.get(index + 2) == Some(&b'=') =>
            {
                return Some((index, 3));
            }
            b'/' if level == 0
                && bytes.get(index + 1) == Some(&b'/')
                && bytes.get(index + 2) == Some(&b'=') =>
            {
                return Some((index, 3));
            }
            _ => index += 1,
        }
    }
    None
}

fn invalid_augassign_target_error(source: &str) -> Option<CpythonDiagnostic> {
    let bytes = source.as_bytes();
    let (operator, _) = top_level_augassign_offset(bytes)?;
    let (target_start, target_end) = assignment_target_expr_range(source, 0, operator)?;
    let target_text = &source[target_start..target_end];
    let Ok(parsed) = parser::parse(target_text, parser::Mode::Expression.into()) else {
        return None;
    };
    let ast::Mod::Expression(expression) = parsed.into_syntax() else {
        return None;
    };
    let name = delete_target_expr_name(&expression.body);
    Some(CpythonDiagnostic::new(
        format!("'{name}' is an illegal expression for augmented assignment"),
        target_start,
        target_end,
    ))
}

fn find_for_target_delimiter(bytes: &[u8], mut index: usize, end: usize) -> Option<usize> {
    let mut level = 0usize;
    while index < end {
        match bytes[index] {
            b'#' if level == 0 => return None,
            b'\'' | b'"' => index = skip_quoted_string(bytes, index),
            b'(' | b'[' | b'{' => {
                level += 1;
                index += 1;
            }
            b')' | b']' | b'}' => {
                level = level.saturating_sub(1);
                index += 1;
            }
            _ if level == 0 && starts_identifier(bytes, index, b"in") => return Some(index),
            b':' if level == 0 => return Some(index),
            _ => index += 1,
        }
    }
    None
}

fn invalid_for_target_error(source: &str) -> Option<CpythonDiagnostic> {
    let bytes = source.as_bytes();
    let mut index = 0usize;
    while index < bytes.len() {
        match bytes[index] {
            b'#' => {
                while index < bytes.len() && bytes[index] != b'\n' {
                    index += 1;
                }
            }
            b'\'' | b'"' => index = skip_quoted_string(bytes, index),
            _ if starts_identifier(bytes, index, b"for") => {
                let target_start = skip_horizontal_whitespace(bytes, index + 3);
                let line_end = source[index..]
                    .find('\n')
                    .map_or(bytes.len(), |newline| index + newline);
                if let Some(target_end) = find_for_target_delimiter(bytes, target_start, line_end) {
                    if let Some(error) =
                        for_target_error_for_slice(source, target_start, target_end)
                    {
                        return Some(error);
                    }
                    if let Some(error) =
                        non_in_compare_for_target_error(source, target_start, target_end)
                    {
                        return Some(error);
                    }
                }
                index = target_start.max(index + 3);
            }
            _ => index += 1,
        }
    }
    None
}

fn find_with_target_delimiter(bytes: &[u8], mut index: usize, end: usize) -> Option<usize> {
    let mut level = 0usize;
    while index < end {
        match bytes[index] {
            b'#' if level == 0 => return None,
            b'\'' | b'"' => index = skip_quoted_string(bytes, index),
            b'(' | b'[' | b'{' => {
                level += 1;
                index += 1;
            }
            b')' if level == 0 => return Some(index),
            b')' | b']' | b'}' => {
                level = level.saturating_sub(1);
                index += 1;
            }
            b',' | b':' if level == 0 => return Some(index),
            _ => index += 1,
        }
    }
    None
}

fn invalid_with_target_error(source: &str) -> Option<CpythonDiagnostic> {
    let bytes = source.as_bytes();
    let mut line_start = 0usize;
    for line in source.split_inclusive('\n') {
        let line_end = line_start + line.len();
        let mut column = skip_horizontal_whitespace(bytes, line_start);
        if starts_identifier(bytes, column, b"async") {
            column = skip_horizontal_whitespace(bytes, column + 5);
        }
        if !starts_identifier(bytes, column, b"with") {
            line_start = line_end;
            continue;
        }
        let mut index = column + 4;
        while let Some(as_index) = find_keyword_at_level(bytes, index, line_end, b"as") {
            let target_start = skip_horizontal_whitespace(bytes, as_index + 2);
            if let Some(target_end) = find_with_target_delimiter(bytes, target_start, line_end) {
                if let Some(error) = star_target_error_for_slice(source, target_start, target_end) {
                    return Some(error);
                }
                index = target_end.saturating_add(1);
            } else {
                break;
            }
        }
        line_start = line_end;
    }
    None
}

fn find_missing_in_if_keyword(bytes: &[u8], mut index: usize, end: usize) -> Option<usize> {
    let mut level = 0usize;
    while index < end {
        match bytes[index] {
            b'#' if level == 0 => return None,
            b'\'' | b'"' => index = skip_quoted_string(bytes, index),
            b'(' | b'[' | b'{' => {
                level += 1;
                index += 1;
            }
            b')' | b']' | b'}' if level == 0 => return None,
            b')' | b']' | b'}' => {
                level = level.saturating_sub(1);
                index += 1;
            }
            _ if level == 0 && starts_identifier(bytes, index, b"in") => return None,
            _ if level == 0 && starts_identifier(bytes, index, b"if") => return Some(index),
            _ => index += 1,
        }
    }
    None
}

fn invalid_for_if_clause_error(source: &str) -> Option<CpythonDiagnostic> {
    let bytes = source.as_bytes();
    let mut index = 0usize;
    let mut level = 0usize;
    while index < bytes.len() {
        match bytes[index] {
            b'#' if level == 0 => {
                while index < bytes.len() && bytes[index] != b'\n' {
                    index += 1;
                }
            }
            b'\'' | b'"' => index = skip_quoted_string(bytes, index),
            b'(' | b'[' | b'{' => {
                level += 1;
                index += 1;
            }
            b')' | b']' | b'}' => {
                level = level.saturating_sub(1);
                index += 1;
            }
            _ if level > 0 && starts_identifier(bytes, index, b"for") => {
                let target_start = skip_horizontal_whitespace(bytes, index + 3);
                let line_end = source[index..]
                    .find('\n')
                    .map_or(bytes.len(), |newline| index + newline);
                if let Some(if_index) = find_missing_in_if_keyword(bytes, target_start, line_end) {
                    return Some(CpythonDiagnostic::new(
                        "'in' expected after for-loop variables".to_owned(),
                        if_index,
                        (if_index + 2).min(line_end),
                    ));
                }
                index = target_start.max(index + 3);
            }
            _ => index += 1,
        }
    }
    None
}

fn invalid_delete_target_error(source: &str) -> Option<CpythonDiagnostic> {
    let bytes = source.as_bytes();
    let mut index = 0;
    while index < bytes.len() {
        match bytes[index] {
            b'#' => {
                while index < bytes.len() && bytes[index] != b'\n' {
                    index += 1;
                }
            }
            b'\'' | b'"' => {
                index = skip_quoted_string(bytes, index);
            }
            b'd' if starts_identifier(bytes, index, b"del") => {
                let mut target_start = index + 3;
                if !matches!(bytes.get(target_start), Some(b' ' | b'\t' | b'\x0c')) {
                    index += 3;
                    continue;
                }
                while matches!(bytes.get(target_start), Some(b' ' | b'\t' | b'\x0c')) {
                    target_start += 1;
                }
                let mut target_end = statement_target_end(bytes, target_start);
                while target_end > target_start
                    && matches!(bytes.get(target_end - 1), Some(b' ' | b'\t' | b'\x0c'))
                {
                    target_end -= 1;
                }
                if target_start >= target_end {
                    index = target_end.max(index + 3);
                    continue;
                }
                if parenthesized_single_starred_delete_target(bytes, target_start, target_end) {
                    return Some(CpythonDiagnostic::new(
                        "cannot use starred expression here".to_owned(),
                        target_start,
                        target_end,
                    ));
                }
                if bytes.get(target_start) == Some(&b'*') {
                    return Some(CpythonDiagnostic::new(
                        "cannot delete starred".to_owned(),
                        target_start,
                        (target_start + 1).min(target_end),
                    ));
                }
                let target_text = &source[target_start..target_end];
                let Ok(parsed) = parser::parse(target_text, parser::Mode::Expression.into()) else {
                    index = target_end;
                    continue;
                };
                let ast::Mod::Expression(expression) = parsed.into_syntax() else {
                    index = target_end;
                    continue;
                };
                let Some(invalid_target) = invalid_delete_target(&expression.body) else {
                    index = target_end;
                    continue;
                };
                let start = target_start + invalid_target.range().start().to_usize();
                let end = target_start + invalid_target.range().end().to_usize();
                if matches!(invalid_target, ast::Expr::FString(_)) {
                    return Some(CpythonDiagnostic::new(
                        "invalid syntax".to_owned(),
                        start,
                        end,
                    ));
                }
                let name = delete_target_expr_name(invalid_target);
                return Some(CpythonDiagnostic::new(
                    format!("cannot delete {name}"),
                    start,
                    end,
                ));
            }
            _ => index += 1,
        }
    }
    None
}

fn skip_horizontal_whitespace(bytes: &[u8], mut index: usize) -> usize {
    while matches!(bytes.get(index), Some(b' ' | b'\t' | b'\x0c')) {
        index += 1;
    }
    index
}

fn find_keyword_at_level(
    bytes: &[u8],
    mut index: usize,
    end: usize,
    keyword: &[u8],
) -> Option<usize> {
    let mut level = 0usize;
    while index < end {
        match bytes[index] {
            b'#' if level == 0 => return None,
            b'\'' | b'"' => {
                index = skip_quoted_string(bytes, index);
            }
            b'(' | b'[' | b'{' => {
                level += 1;
                index += 1;
            }
            b')' | b']' | b'}' => {
                level = level.saturating_sub(1);
                index += 1;
            }
            _ if level == 0 && starts_identifier(bytes, index, keyword) => return Some(index),
            _ => index += 1,
        }
    }
    None
}

fn find_byte_at_level(bytes: &[u8], mut index: usize, end: usize, needle: u8) -> Option<usize> {
    let mut level = 0usize;
    while index < end {
        match bytes[index] {
            b'#' if level == 0 => return None,
            b'\'' | b'"' => {
                index = skip_quoted_string(bytes, index);
            }
            b'(' | b'[' | b'{' => {
                level += 1;
                index += 1;
            }
            b')' | b']' | b'}' => {
                level = level.saturating_sub(1);
                index += 1;
            }
            byte if level == 0 && byte == needle => return Some(index),
            _ => index += 1,
        }
    }
    None
}

fn expression_name_and_range(source: &str) -> Option<(&'static str, usize, usize, bool)> {
    let parsed = parser::parse(source, parser::Mode::Expression.into()).ok()?;
    let ast::Mod::Expression(expression) = parsed.into_syntax() else {
        return None;
    };
    let is_name = matches!(expression.body.as_ref(), ast::Expr::Name(_));
    Some((
        delete_target_expr_name(&expression.body),
        expression.body.range().start().to_usize(),
        expression.body.range().end().to_usize(),
        is_name,
    ))
}

fn invalid_standalone_except_error(source: &str) -> Option<CpythonDiagnostic> {
    let bytes = source.as_bytes();
    let mut line_start = 0usize;
    let mut seen_try = false;
    for line in source.split_inclusive('\n') {
        let line_end = line_start + line.len();
        let column = skip_horizontal_whitespace(bytes, line_start);
        if column >= line_end {
            line_start = line_end;
            continue;
        }
        if starts_identifier(bytes, column, b"try") {
            seen_try = true;
        } else if (bytes.get(column..column + 7) == Some(b"except*")
            || starts_identifier(bytes, column, b"except"))
            && !seen_try
        {
            let end = if bytes.get(column..column + 7) == Some(b"except*") {
                column + 7
            } else {
                column + 6
            };
            return Some(CpythonDiagnostic::new(
                "invalid syntax".to_owned(),
                column,
                end,
            ));
        }
        line_start = line_end;
    }
    None
}

fn invalid_import_statement_error(source: &str) -> Option<CpythonDiagnostic> {
    let bytes = source.as_bytes();
    let mut line_start = 0usize;
    for line in source.split_inclusive('\n') {
        let line_end = line_start + line.len();
        let column = skip_horizontal_whitespace(bytes, line_start);
        if column < line_end
            && starts_identifier(bytes, column, b"import")
            && find_keyword_at_level(bytes, column + 6, line_end, b"from").is_some()
        {
            return Some(CpythonDiagnostic::new(
                "Did you mean to use 'from ... import ...' instead?".to_owned(),
                column,
                column + 6,
            ));
        }
        line_start = line_end;
    }
    None
}

fn import_as_target_end(bytes: &[u8], mut index: usize) -> usize {
    let mut level = 0usize;
    while index < bytes.len() {
        match bytes[index] {
            b'#' if level == 0 => return index,
            b'\'' | b'"' => index = skip_quoted_string(bytes, index),
            b'(' | b'[' | b'{' => {
                level += 1;
                index += 1;
            }
            b')' if level == 0 => return index,
            b')' | b']' | b'}' => {
                level = level.saturating_sub(1);
                index += 1;
            }
            b',' | b';' | b'\n' if level == 0 => return index,
            _ => index += 1,
        }
    }
    index
}

fn valid_import_alias_name(bytes: &[u8], mut start: usize, end: usize) -> bool {
    start = skip_horizontal_whitespace(bytes, start);
    let Some(&first) = bytes.get(start) else {
        return false;
    };
    if !(first == b'_' || first.is_ascii_alphabetic() || first >= 0x80) {
        return false;
    }
    let mut index = start + 1;
    while index < end {
        match bytes[index] {
            b' ' | b'\t' | b'\x0c' => break,
            byte if byte >= 0x80 || is_ascii_identifier_char(byte) => index += 1,
            _ => return false,
        }
    }
    let index = skip_horizontal_whitespace(bytes, index);
    matches!(
        bytes.get(index),
        None | Some(b',' | b')' | b';' | b'\n' | b'\r')
    )
}

fn import_target_error_for_slice(
    source: &str,
    start: usize,
    end: usize,
) -> Option<CpythonDiagnostic> {
    let bytes = source.as_bytes();
    let (target_start, target_end) = trim_target_range(bytes, start, end);
    if target_start >= target_end || valid_import_alias_name(bytes, target_start, target_end) {
        return None;
    }
    let parsed = parser::parse(
        &source[target_start..target_end],
        parser::Mode::Expression.into(),
    )
    .ok()?;
    let ast::Mod::Expression(expression) = parsed.into_syntax() else {
        return None;
    };
    let name = delete_target_expr_name(&expression.body);
    let start = target_start + expression.body.range().start().to_usize();
    let end = target_start + expression.body.range().end().to_usize();
    Some(CpythonDiagnostic::new(
        format!("cannot use {name} as import target"),
        start,
        end,
    ))
}

fn statement_starts_import(bytes: &[u8], line_start: usize, line_end: usize) -> bool {
    let column = skip_horizontal_whitespace(bytes, line_start);
    if starts_identifier(bytes, column, b"import") {
        return true;
    }
    starts_identifier(bytes, column, b"from")
        && find_keyword_at_level(bytes, column + 4, line_end, b"import").is_some()
}

fn invalid_import_target_error(source: &str) -> Option<CpythonDiagnostic> {
    let bytes = source.as_bytes();
    let mut line_start = 0usize;
    let mut in_parenthesized_from_import = false;
    for line in source.split_inclusive('\n') {
        let line_end = line_start + line.len();
        let starts_import = statement_starts_import(bytes, line_start, line_end);
        if starts_import && bytes[line_start..line_end].contains(&b'(') {
            in_parenthesized_from_import = true;
        }
        if starts_import || in_parenthesized_from_import {
            let mut index = line_start;
            while index < line_end {
                if starts_identifier(bytes, index, b"as") {
                    let target_start = skip_horizontal_whitespace(bytes, index + 2);
                    let target_end = import_as_target_end(bytes, target_start);
                    if let Some(error) =
                        import_target_error_for_slice(source, target_start, target_end)
                    {
                        return Some(error);
                    }
                    index = target_end.max(index + 2);
                } else {
                    index += 1;
                }
            }
        }
        if in_parenthesized_from_import && bytes[line_start..line_end].contains(&b')') {
            in_parenthesized_from_import = false;
        }
        line_start = line_end;
    }
    None
}

fn invalid_except_as_target_error(source: &str) -> Option<CpythonDiagnostic> {
    let bytes = source.as_bytes();
    let mut line_start = 0usize;
    let mut seen_try = false;
    for line in source.split_inclusive('\n') {
        let line_end = line_start + line.len();
        let mut column = skip_horizontal_whitespace(bytes, line_start);
        if column >= line_end {
            line_start = line_end;
            continue;
        }
        if starts_identifier(bytes, column, b"try") {
            seen_try = true;
            line_start = line_end;
            continue;
        }
        let (keyword_len, starred) = if bytes.get(column..column + 7) == Some(b"except*") {
            (7, true)
        } else if starts_identifier(bytes, column, b"except") {
            (6, false)
        } else {
            line_start = line_end;
            continue;
        };
        if !seen_try {
            line_start = line_end;
            continue;
        }
        column += keyword_len;
        let Some(as_index) = find_keyword_at_level(bytes, column, line_end, b"as") else {
            line_start = line_end;
            continue;
        };
        let target_start = skip_horizontal_whitespace(bytes, as_index + 2);
        let Some(delimiter) = find_byte_at_level(bytes, target_start, line_end, b':')
            .into_iter()
            .chain(find_byte_at_level(bytes, target_start, line_end, b','))
            .min()
        else {
            line_start = line_end;
            continue;
        };
        let mut target_end = delimiter;
        while target_end > target_start
            && matches!(bytes.get(target_end - 1), Some(b' ' | b'\t' | b'\x0c'))
        {
            target_end -= 1;
        }
        let Some((expr_name, start, end, is_name)) =
            expression_name_and_range(&source[target_start..target_end])
        else {
            line_start = line_end;
            continue;
        };
        if !is_name {
            let statement = if starred { "except*" } else { "except" };
            return Some(CpythonDiagnostic::new(
                format!("cannot use {statement} statement with {expr_name}"),
                target_start + start,
                target_start + end,
            ));
        }
        line_start = line_end;
    }
    None
}

fn invalid_match_as_target_error(source: &str) -> Option<CpythonDiagnostic> {
    let bytes = source.as_bytes();
    let quoted_ranges = quoted_string_ranges(bytes);
    let mut quoted_range = 0usize;
    let mut line_start = 0usize;
    for line in source.split_inclusive('\n') {
        let line_end = line_start + line.len();
        let mut column = skip_horizontal_whitespace(bytes, line_start);
        if column >= line_end
            || offset_in_ranges(&quoted_ranges, &mut quoted_range, column)
            || !starts_identifier(bytes, column, b"case")
        {
            line_start = line_end;
            continue;
        }
        column += 4;
        let Some(as_index) = find_keyword_at_level(bytes, column, line_end, b"as") else {
            line_start = line_end;
            continue;
        };
        let target_start = skip_horizontal_whitespace(bytes, as_index + 2);
        let Some(delimiter) = find_byte_at_level(bytes, target_start, line_end, b':')
            .into_iter()
            .chain(find_byte_at_level(bytes, target_start, line_end, b','))
            .min()
        else {
            line_start = line_end;
            continue;
        };
        let mut target_end = delimiter;
        while target_end > target_start
            && matches!(bytes.get(target_end - 1), Some(b' ' | b'\t' | b'\x0c'))
        {
            target_end -= 1;
        }
        if source[target_start..target_end].trim() == "_" {
            return Some(CpythonDiagnostic::new(
                "cannot use '_' as a target".to_owned(),
                target_start,
                target_end,
            ));
        }
        let Some((expr_name, start, end, is_name)) =
            expression_name_and_range(&source[target_start..target_end])
        else {
            line_start = line_end;
            continue;
        };
        if !is_name {
            if matches!(expr_name, "expression" | "subscript") {
                line_start = line_end;
                continue;
            }
            return Some(CpythonDiagnostic::new(
                format!("cannot use {expr_name} as pattern target"),
                target_start + start,
                target_start + end,
            ));
        }
        line_start = line_end;
    }
    None
}

fn quoted_string_ranges(bytes: &[u8]) -> Vec<(usize, usize)> {
    let mut ranges = Vec::new();
    let mut index = 0usize;
    while index < bytes.len() {
        match bytes[index] {
            b'#' => {
                while index < bytes.len() && bytes[index] != b'\n' {
                    index += 1;
                }
            }
            b'\'' | b'"' => {
                let end = skip_quoted_string(bytes, index);
                ranges.push((index, end));
                index = end;
            }
            _ => index += 1,
        }
    }
    ranges
}

fn offset_in_ranges(ranges: &[(usize, usize)], range_index: &mut usize, offset: usize) -> bool {
    while ranges
        .get(*range_index)
        .is_some_and(|(_, end)| *end <= offset)
    {
        *range_index += 1;
    }
    ranges
        .get(*range_index)
        .is_some_and(|(start, end)| *start <= offset && offset < *end)
}

fn invalid_match_mapping_rest_wildcard_error(source: &str) -> Option<CpythonDiagnostic> {
    let bytes = source.as_bytes();
    let next_line_end = |line_start: usize| {
        line_start
            + bytes[line_start..]
                .iter()
                .position(|byte| *byte == b'\n')
                .unwrap_or(bytes.len() - line_start)
    };
    let mut index = 0usize;
    let mut line_start = 0usize;
    let mut line_end = next_line_end(line_start);
    while index < bytes.len() {
        match bytes[index] {
            b'#' => {
                while index < bytes.len() && bytes[index] != b'\n' {
                    index += 1;
                }
            }
            b'\'' | b'"' => index = skip_quoted_string(bytes, index),
            b'\n' => {
                index += 1;
                line_start = index;
                line_end = next_line_end(line_start);
            }
            _ => {
                let column = skip_horizontal_whitespace(bytes, line_start);
                if index != column
                    || column >= line_end
                    || !starts_identifier(bytes, column, b"case")
                {
                    index += 1;
                    continue;
                }
                let mut cursor = column + 4;
                while cursor < line_end {
                    match bytes[cursor] {
                        b'#' => break,
                        b'\'' | b'"' => cursor = skip_quoted_string(bytes, cursor),
                        b'{' => {
                            let rest = next_non_horizontal_whitespace(bytes, cursor + 1);
                            if bytes.get(rest..rest + 2) == Some(b"**") {
                                let name_start = next_non_horizontal_whitespace(bytes, rest + 2);
                                let name_end = identifier_end(bytes, name_start, line_end);
                                if source.get(name_start..name_end) == Some("_") {
                                    return Some(CpythonDiagnostic::new(
                                        "invalid syntax".to_owned(),
                                        name_start,
                                        name_end,
                                    ));
                                }
                            }
                            cursor += 1;
                        }
                        _ => cursor += 1,
                    }
                }
                index = line_end;
            }
        }
    }
    None
}

fn invalid_if_expression_statement_error(source: &str) -> Option<CpythonDiagnostic> {
    let bytes = source.as_bytes();
    let mut line_start = 0usize;
    for line in source.split_inclusive('\n') {
        let line_end = line_start + line.len();
        if let Some(if_index) = find_keyword_at_level(bytes, line_start, line_end, b"if")
            && let Some((start, end)) = statement_before_if_expression(bytes, line_start, if_index)
            && find_keyword_at_level(bytes, if_index + 2, line_end, b"else").is_some()
        {
            return Some(CpythonDiagnostic::new(
                "expected expression before 'if', but statement is given".to_owned(),
                start,
                end,
            ));
        }
        if let Some(else_index) = find_keyword_at_level(bytes, line_start, line_end, b"else")
            && find_keyword_at_level(bytes, line_start, else_index, b"if").is_some()
            && let Some((start, end)) =
                statement_after_else_expression(bytes, else_index + 4, line_end)
        {
            return Some(CpythonDiagnostic::new(
                "expected expression after 'else', but statement is given".to_owned(),
                start,
                end,
            ));
        }
        line_start = line_end;
    }
    None
}

fn statement_before_if_expression(
    bytes: &[u8],
    line_start: usize,
    if_index: usize,
) -> Option<(usize, usize)> {
    let mut start = if_index;
    while start > line_start && matches!(bytes.get(start - 1), Some(b' ' | b'\t' | b'\x0c')) {
        start -= 1;
    }
    while start > line_start
        && !matches!(
            bytes.get(start - 1),
            Some(b'=' | b':' | b',' | b'(' | b'[' | b'{')
        )
    {
        start -= 1;
    }
    start = skip_horizontal_whitespace(bytes, start);
    for keyword in [b"pass".as_slice(), b"break", b"continue"] {
        if starts_identifier(bytes, start, keyword) {
            return Some((start, start + keyword.len()));
        }
    }
    None
}

fn statement_after_else_expression(
    bytes: &[u8],
    else_end: usize,
    line_end: usize,
) -> Option<(usize, usize)> {
    let start = skip_horizontal_whitespace(bytes, else_end);
    for keyword in [
        b"pass".as_slice(),
        b"return",
        b"raise",
        b"del",
        b"yield",
        b"assert",
        b"break",
        b"continue",
        b"import",
        b"from",
    ] {
        if starts_identifier(bytes, start, keyword) {
            let end = statement_target_end(bytes, start).min(line_end);
            return Some((start, end.max(start + keyword.len())));
        }
    }
    None
}

fn invalid_else_elif_error(source: &str) -> Option<CpythonDiagnostic> {
    let bytes = source.as_bytes();
    let mut line_start = 0usize;
    let mut else_indents: Vec<usize> = Vec::new();
    for line in source.split_inclusive('\n') {
        let line_end = line_start + line.len();
        let column = skip_horizontal_whitespace(bytes, line_start);
        let line_column = column.saturating_sub(line_start);
        if column >= line_end {
            line_start = line_end;
            continue;
        }
        while else_indents
            .last()
            .is_some_and(|indent| line_column < *indent)
        {
            else_indents.pop();
        }
        if starts_identifier(bytes, column, b"else")
            && find_byte_at_level(bytes, column + 4, line_end, b':').is_some()
        {
            else_indents.push(line_column);
        } else if starts_identifier(bytes, column, b"elif") && else_indents.contains(&line_column) {
            return Some(CpythonDiagnostic::new(
                "'elif' block follows an 'else' block".to_owned(),
                column,
                column + 4,
            ));
        }
        line_start = line_end;
    }
    None
}

fn mixed_except_handlers_error(source: &str) -> Option<CpythonDiagnostic> {
    let message = "cannot have both 'except' and 'except*' on the same 'try'".to_owned();
    let mut seen_except = false;
    let mut seen_except_star = false;
    let mut line_start = 0usize;
    for line in source.split_inclusive('\n') {
        let bytes = line.as_bytes();
        let mut column = 0usize;
        while matches!(bytes.get(column), Some(b' ' | b'\t' | b'\x0c')) {
            column += 1;
        }
        let token_start = line_start + column;
        if bytes.get(column..column + 7) == Some(b"except*") {
            if seen_except {
                return Some(CpythonDiagnostic::new(
                    message,
                    token_start,
                    token_start + 7,
                ));
            }
            seen_except_star = true;
        } else if starts_identifier(bytes, column, b"except") {
            if seen_except_star {
                return Some(CpythonDiagnostic::new(
                    message,
                    token_start,
                    token_start + 6,
                ));
            }
            seen_except = true;
        }
        line_start += line.len();
    }
    None
}

fn non_printable_character_error(source: &str) -> Option<CpythonDiagnostic> {
    let bytes = source.as_bytes();
    let mut index = 0;
    while index < bytes.len() {
        match bytes[index] {
            b'#' => {
                while index < bytes.len() && bytes[index] != b'\n' {
                    index += 1;
                }
            }
            b'\'' | b'"' => {
                index = skip_quoted_string(bytes, index);
            }
            byte if byte.is_ascii_control() && !matches!(byte, b'\t' | b'\n' | b'\r' | b'\x0c') => {
                return Some(CpythonDiagnostic::new(
                    format!("invalid non-printable character U+{byte:04X}"),
                    index,
                    index + 1,
                ));
            }
            byte if byte >= 0x80 => {
                let ch = source[index..].chars().next()?;
                if ch.is_control() {
                    return Some(CpythonDiagnostic::new(
                        format!("invalid non-printable character U+{:04X}", ch as u32),
                        index,
                        index + ch.len_utf8(),
                    ));
                }
                index += ch.len_utf8();
            }
            _ => index += 1,
        }
    }
    None
}

fn unterminated_string_error(source: &str, mode: Mode) -> Option<CpythonDiagnostic> {
    let bytes = source.as_bytes();
    let mut index = 0;
    let mut line = 1usize;
    while index < bytes.len() {
        match bytes[index] {
            b'#' => {
                while index < bytes.len() && bytes[index] != b'\n' {
                    index += 1;
                }
            }
            b'\n' => {
                line += 1;
                index += 1;
            }
            quote @ (b'\'' | b'"') => {
                let start = index;
                let start_line = line;
                let quote_size = if bytes.get(index + 1) == Some(&quote)
                    && bytes.get(index + 2) == Some(&quote)
                {
                    3
                } else {
                    1
                };
                index += quote_size;
                let mut has_escaped_quote = false;
                let mut ended_with_escape = false;
                let mut closed = false;
                while index < bytes.len() {
                    let c = bytes[index];
                    if c == b'\n' {
                        if quote_size == 1 {
                            // A single-quoted literal cannot span a line, so this is the same
                            // "the literal never ended" case as running out of source; CPython
                            // spells the two as one condition in lexer.c too.
                            break;
                        }
                        line += 1;
                        index += 1;
                    } else if c == quote {
                        if quote_size == 3 {
                            if bytes.get(index + 1) == Some(&quote)
                                && bytes.get(index + 2) == Some(&quote)
                            {
                                index += 3;
                                closed = true;
                                break;
                            }
                            index += 1;
                        } else {
                            index += 1;
                            closed = true;
                            break;
                        }
                    } else if c == b'\\' {
                        if bytes.get(index + 1) == Some(&quote) {
                            has_escaped_quote = true;
                        }
                        ended_with_escape = index + 1 >= bytes.len();
                        index = (index + 2).min(bytes.len());
                    } else {
                        index += 1;
                    }
                }
                if !closed {
                    if let Some(error) =
                        unclosed_replacement_field_error(bytes, start, start + quote_size, index)
                    {
                        return Some(error);
                    }
                    let detected_line = if quote_size == 3 { line } else { start_line };
                    let interpolated = interpolated_string_prefix(bytes, start);
                    let diagnostic = CpythonDiagnostic::new(
                        unterminated_string_message(
                            detected_line,
                            quote_size == 3,
                            has_escaped_quote,
                            interpolated,
                        ),
                        start,
                        start,
                    );
                    // E_EOLS is only for a plain single-quoted string at EOF.
                    // Single-quoted f/t-strings never set it. exec input appends a
                    // newline, so a single-quoted literal becomes the newline case
                    // unless a final `\` consumes that newline.
                    let exec_implicit_newline =
                        matches!(mode, Mode::Exec) && quote_size == 1 && !ended_with_escape;
                    let eval_assignment =
                        matches!(mode, Mode::Eval) && eval_has_assignment_before(bytes, start);
                    let continuable = index >= bytes.len()
                        && !(interpolated.is_some() && quote_size == 1)
                        && !exec_implicit_newline
                        && !eval_assignment;
                    return Some(if continuable {
                        diagnostic.with_unclosed_string()
                    } else {
                        diagnostic
                    });
                }
            }
            _ => index += 1,
        }
    }
    None
}

fn eval_has_assignment_before(bytes: &[u8], end: usize) -> bool {
    let mut index = 0;
    let mut level = 0usize;
    let mut in_lambda_params = false;
    while index < end {
        match bytes[index] {
            b'#' => {
                while index < end && bytes[index] != b'\n' {
                    index += 1;
                }
            }
            b'\'' | b'"' => index = skip_quoted_string(bytes, index),
            b'(' | b'[' | b'{' => {
                level += 1;
                index += 1;
            }
            b')' | b']' | b'}' => {
                level = level.saturating_sub(1);
                index += 1;
            }
            b':' if level == 0 => {
                in_lambda_params = false;
                index += 1;
            }
            b'=' if level == 0
                && !in_lambda_params
                && bytes.get(index + 1) != Some(&b'=')
                && !matches!(
                    bytes.get(index.saturating_sub(1)),
                    Some(b':' | b'!' | b'<' | b'>')
                ) =>
            {
                return true;
            }
            _ if level == 0 && starts_identifier(bytes, index, b"lambda") => {
                in_lambda_params = true;
                index += b"lambda".len();
            }
            _ => index += 1,
        }
    }
    false
}

fn invalid_interpolated_string_error(source: &str) -> Option<CpythonDiagnostic> {
    let bytes = source.as_bytes();
    let mut index = 0;
    while index < bytes.len() {
        match bytes[index] {
            b'#' => {
                while index < bytes.len() && bytes[index] != b'\n' {
                    index += 1;
                }
            }
            quote @ (b'\'' | b'"') => {
                let Some(prefix) = interpolated_string_prefix(bytes, index) else {
                    index = skip_quoted_string(bytes, index);
                    continue;
                };
                if let Some(error) =
                    single_quoted_format_spec_newline_error(bytes, index, quote, prefix)
                {
                    return Some(error);
                }
                let Some((content_start, content_end)) =
                    quoted_string_content_range(bytes, index, quote)
                else {
                    index = skip_quoted_string(bytes, index);
                    continue;
                };
                if let Some(error) =
                    invalid_replacement_field_error(bytes, content_start, content_end, prefix)
                {
                    return Some(error);
                }
                index = skip_quoted_string(bytes, index);
            }
            _ => index += 1,
        }
    }
    None
}

/// Whether the literal opened at `quote` carries a `b` in its prefix.
fn is_bytes_literal(bytes: &[u8], quote: usize) -> bool {
    let mut start = quote;
    while quote - start < 2
        && start > 0
        && matches!(
            bytes[start - 1].to_ascii_lowercase(),
            b'r' | b'b' | b'u' | b'f' | b't'
        )
    {
        start -= 1;
    }
    if start > 0 && is_ascii_identifier_char(bytes[start - 1]) {
        return false;
    }
    bytes[start..quote]
        .iter()
        .any(|byte| byte.eq_ignore_ascii_case(&b'b'))
}

/// Which of CPython's two messages a mixed literal concatenation gets. Ruff reports only the
/// bytes/non-bytes mix, so `t"x" b"y"` arrives here as a bytes error where CPython names the
/// t-string, and this scanner owns the choice between them.
///
/// CPython's `invalid_string_tstring_concat` is an `invalid_` rule, reached only on the error
/// pass, and the `strings` rule tries `(fstring|string)+` first. That alternative consumes the
/// concatenation's leading run of non-t-string literals, so a bytes/non-bytes mix among *those*
/// raises from `_PyPegen_concatenate_strings` on the first pass and the t-string rule never runs.
/// The bytes message therefore keeps precedence for `"a" b"b" t"c"` but not for `t"x" b"y"`.
fn mixed_tstring_literal_error(
    error: &parser::ParseError,
    source: &str,
) -> Option<CpythonDiagnostic> {
    let parser::ParseErrorType::OtherError(message) = &error.error else {
        return None;
    };
    if !message.eq_ignore_ascii_case("bytes literal cannot be mixed with non-bytes literals") {
        return None;
    }

    let start = error.location.start().to_usize();
    let end = error.location.end().to_usize();
    let bytes = source.as_bytes();
    if end > bytes.len() {
        return None;
    }

    let mut index = start;
    let (mut bytes_seen, mut nonbytes_seen) = (false, false);
    while index < end {
        match bytes[index] {
            b'\'' | b'"' => {
                if interpolated_string_prefix(bytes, index) == Some("t-string") {
                    let message = if bytes_seen && nonbytes_seen {
                        // Reported here rather than left to ruff's error so that the scanners
                        // further down the chain cannot claim the concatenation instead.
                        "cannot mix bytes and nonbytes literals"
                    } else {
                        "cannot mix t-string literals with string or bytes literals"
                    };
                    return Some(CpythonDiagnostic::new(message.to_owned(), start, end));
                }
                if is_bytes_literal(bytes, index) {
                    bytes_seen = true;
                } else {
                    nonbytes_seen = true;
                }
                index = skip_quoted_string(bytes, index);
            }
            _ => index += 1,
        }
    }
    None
}

fn single_quoted_format_spec_newline_error(
    bytes: &[u8],
    quote_index: usize,
    quote: u8,
    prefix: &str,
) -> Option<CpythonDiagnostic> {
    if bytes.get(quote_index + 1) == Some(&quote) && bytes.get(quote_index + 2) == Some(&quote) {
        return None;
    }

    let (content_start, content_end) = quoted_string_content_range(bytes, quote_index, quote)?;
    let mut index = content_start;
    while index < content_end {
        match bytes[index] {
            b'{' if bytes.get(index + 1) == Some(&b'{') => index += 2,
            b'}' if bytes.get(index + 1) == Some(&b'}') => index += 2,
            b'{' => {
                let expr_start = skip_ascii_whitespace(bytes, index + 1, content_end);
                if let Some(separator) = replacement_field_separator(bytes, expr_start, content_end)
                    && bytes[separator] == b':'
                {
                    let format_end =
                        replacement_field_closing_brace(bytes, separator + 1, content_end)
                            .unwrap_or(content_end);
                    if bytes[separator + 1..format_end].contains(&b'\n') {
                        return Some(CpythonDiagnostic::new(
                            format!(
                                "{prefix}: newlines are not allowed in format specifiers for single quoted {prefix}s"
                            ),
                            quote_index,
                            quote_index + 1,
                        ));
                    }
                }
                index += 1;
            }
            _ => index += 1,
        }
    }
    None
}

fn interpolated_string_prefix(bytes: &[u8], quote: usize) -> Option<&'static str> {
    let prev = quote.checked_sub(1).and_then(|index| bytes.get(index))?;
    let lower_prev = prev.to_ascii_lowercase();
    let (prefix_start, marker) = if matches!(lower_prev, b'f' | b't') {
        if quote >= 2 && bytes[quote - 2].eq_ignore_ascii_case(&b'r') {
            (quote - 2, lower_prev)
        } else {
            (quote - 1, lower_prev)
        }
    } else if lower_prev == b'r'
        && quote >= 2
        && matches!(bytes[quote - 2].to_ascii_lowercase(), b'f' | b't')
    {
        (quote - 2, bytes[quote - 2].to_ascii_lowercase())
    } else {
        return None;
    };

    if prefix_start > 0 && identifier_continue_before(bytes, prefix_start) {
        return None;
    }

    Some(if marker == b'f' {
        "f-string"
    } else {
        "t-string"
    })
}

const MAXFSTRINGLEVEL: usize = 150;

fn skip_string_token(bytes: &[u8], quote_index: usize) -> usize {
    if interpolated_string_prefix(bytes, quote_index).is_some() {
        interpolated_string_end(bytes, quote_index).unwrap_or(bytes.len())
    } else {
        skip_quoted_string(bytes, quote_index)
    }
}

fn interpolated_string_end(bytes: &[u8], quote_index: usize) -> Option<usize> {
    let quote = bytes[quote_index];
    let triple =
        bytes.get(quote_index + 1) == Some(&quote) && bytes.get(quote_index + 2) == Some(&quote);
    let quote_len = if triple { 3 } else { 1 };
    let mut index = quote_index + quote_len;
    let mut brace_depth = 0usize;
    while index < bytes.len() {
        if brace_depth == 0 {
            if bytes[index] == b'\\' {
                index = (index + 2).min(bytes.len());
                continue;
            }
            if triple
                && bytes.get(index) == Some(&quote)
                && bytes.get(index + 1) == Some(&quote)
                && bytes.get(index + 2) == Some(&quote)
            {
                return Some(index + 3);
            }
            if !triple && bytes[index] == quote {
                return Some(index + 1);
            }
            if bytes[index] == b'{' {
                if bytes.get(index + 1) == Some(&b'{') {
                    index += 2;
                } else {
                    brace_depth = 1;
                    index += 1;
                }
            } else if bytes[index] == b'}' && bytes.get(index + 1) == Some(&b'}') {
                index += 2;
            } else {
                index += 1;
            }
        } else {
            match bytes[index] {
                quote_ch @ (b'\'' | b'"') => {
                    if quoted_string_is_closed(bytes, index) {
                        index = skip_string_token(bytes, index).max(index + 1);
                    } else if !triple && quote_ch == quote {
                        return Some(index + 1);
                    } else {
                        index = skip_string_token(bytes, index).max(index + 1);
                    }
                }
                b'#' => {
                    while index < bytes.len() && bytes[index] != b'\n' {
                        index += 1;
                    }
                }
                b'{' => {
                    brace_depth += 1;
                    index += 1;
                }
                b'}' => {
                    brace_depth -= 1;
                    index += 1;
                }
                b'\\' => index = (index + 2).min(bytes.len()),
                _ => index += 1,
            }
        }
    }
    None
}

fn quote_len_at(bytes: &[u8], quote_index: usize) -> usize {
    let quote = bytes[quote_index];
    if bytes.get(quote_index + 1) == Some(&quote) && bytes.get(quote_index + 2) == Some(&quote) {
        3
    } else {
        1
    }
}

fn interpolated_string_content_range(bytes: &[u8], quote_index: usize) -> Option<(usize, usize)> {
    let end = interpolated_string_end(bytes, quote_index)?;
    let quote_len = quote_len_at(bytes, quote_index);
    Some((quote_index + quote_len, end - quote_len))
}

fn quoted_string_content_range(
    bytes: &[u8],
    quote_index: usize,
    quote: u8,
) -> Option<(usize, usize)> {
    let triple =
        bytes.get(quote_index + 1) == Some(&quote) && bytes.get(quote_index + 2) == Some(&quote);
    let quote_len = if triple { 3 } else { 1 };
    let content_start = quote_index + quote_len;
    let mut index = content_start;
    while index < bytes.len() {
        if bytes[index] == b'\\' {
            index = (index + 2).min(bytes.len());
        } else if (triple
            && bytes.get(index) == Some(&quote)
            && bytes.get(index + 1) == Some(&quote)
            && bytes.get(index + 2) == Some(&quote))
            || (!triple && bytes[index] == quote)
        {
            return Some((content_start, index));
        } else {
            index += 1;
        }
    }
    None
}

fn invalid_replacement_field_error(
    bytes: &[u8],
    start: usize,
    end: usize,
    prefix: &str,
) -> Option<CpythonDiagnostic> {
    let mut index = start;
    while index < end {
        match bytes[index] {
            b'{' if bytes.get(index + 1) == Some(&b'{') => index += 2,
            b'}' if bytes.get(index + 1) == Some(&b'}') => index += 2,
            b'{' => {
                if let Some(error) = replacement_field_error(bytes, index, end, prefix, 0) {
                    return Some(error);
                }
                // Resume in the literal text. Inside the field a `#` starts a comment and a `{`
                // opens no new field, so walking on byte by byte would misread both.
                index = replacement_field_end(bytes, index, end).unwrap_or(end);
            }
            _ => index += 1,
        }
    }
    None
}

/// Position just past a replacement field's closing brace. The expression part is code, where a
/// comment runs to the end of the line; past the separator only nested fields nest.
fn replacement_field_end(bytes: &[u8], open: usize, end: usize) -> Option<usize> {
    let expr_start = skip_replacement_field_trivia(bytes, open + 1, end);
    let separator = replacement_field_separator(bytes, expr_start, end)?;
    if bytes[separator] == b'}' {
        return Some(separator + 1);
    }
    replacement_field_closing_brace(bytes, separator + 1, end).map(|brace| brace + 1)
}

/// The one-based line holding `offset`. Counted on demand: these scanners only run once the
/// source has failed to parse, and tracking a line through the spans they skip would have to
/// re-scan them anyway.
fn line_of_offset(bytes: &[u8], offset: usize) -> usize {
    1 + bytes[..offset]
        .iter()
        .filter(|byte| **byte == b'\n')
        .count()
}

/// Brackets opened in a replacement field's expression must close before the field does. CPython
/// words these exactly as it does for brackets anywhere else, except that a closer with nothing
/// open is attributed to the literal. `start..end` must cover only the expression: a format spec
/// is plain text, so a bracket there opens nothing.
fn replacement_field_bracket_error(
    bytes: &[u8],
    start: usize,
    end: usize,
    prefix: &str,
) -> Option<CpythonDiagnostic> {
    // Each entry carries where its bracket opened, so a mismatch can name that line.
    let mut stack: Vec<(u8, usize)> = Vec::new();
    let mut index = start;
    while index < end {
        let byte = bytes[index];
        match byte {
            b'\'' | b'"' => {
                index = skip_quoted_string(bytes, index);
                continue;
            }
            b'#' => {
                index = skip_replacement_field_comment(bytes, index, end);
                continue;
            }
            b'(' | b'[' | b'{' => stack.push((byte, index)),
            b')' | b']' | b'}' => {
                let expected = expected_opening_bracket(byte as char) as u8;
                match stack.last() {
                    // The field's own closing brace: everything inside it balanced.
                    None if byte == b'}' => return None,
                    None => {
                        return Some(CpythonDiagnostic::new(
                            format!("{prefix}: unmatched '{}'", byte as char),
                            index,
                            index + 1,
                        ));
                    }
                    Some(&(opening, _)) if opening == expected => {
                        stack.pop();
                    }
                    Some(&(opening, opened_at)) => {
                        // CPython names the opening line only when it is not the closing one,
                        // comparing `parenlinenostack[level]` against the current `lineno`.
                        let opening_line = line_of_offset(bytes, opened_at);
                        let suffix = if opening_line == line_of_offset(bytes, index) {
                            String::new()
                        } else {
                            format!(" on line {opening_line}")
                        };
                        return Some(CpythonDiagnostic::new(
                            format!(
                                "closing parenthesis '{}' does not match opening parenthesis '{}'{suffix}",
                                byte as char, opening as char
                            ),
                            index,
                            index + 1,
                        ));
                    }
                }
            }
            _ => {}
        }
        index += 1;
    }
    None
}

/// A `#` in a replacement field's expression starts a comment. When no newline follows it before
/// the end of the literal, the comment swallows the closing brace and the field can never close.
/// `start..end` must cover only the expression: `#` is an ordinary character in a format spec,
/// where it selects the alternate form.
fn replacement_field_comment_error(
    bytes: &[u8],
    open: usize,
    start: usize,
    end: usize,
    literal_end: usize,
) -> Option<CpythonDiagnostic> {
    let mut index = start;
    while index < end {
        match bytes[index] {
            b'\'' | b'"' => {
                index = skip_quoted_string(bytes, index);
                continue;
            }
            b'#' => {
                if !bytes[index..literal_end].contains(&b'\n') {
                    return Some(
                        CpythonDiagnostic::new("'{' was never closed".to_owned(), open, open + 1)
                            .with_unclosed_bracket(),
                    );
                }
                index = skip_replacement_field_comment(bytes, index, end);
                continue;
            }
            _ => index += 1,
        }
    }
    None
}

/// Advances past a comment, stopping on the newline that ends it.
fn skip_replacement_field_comment(bytes: &[u8], mut index: usize, end: usize) -> usize {
    while index < end && bytes[index] != b'\n' {
        index += 1;
    }
    index
}

/// Whitespace and comments may both precede a replacement field's expression.
fn skip_replacement_field_trivia(bytes: &[u8], mut index: usize, end: usize) -> usize {
    loop {
        index = skip_ascii_whitespace(bytes, index, end);
        if index < end && bytes[index] == b'#' {
            index = skip_replacement_field_comment(bytes, index, end);
            continue;
        }
        return index;
    }
}

/// Bounds the mutual recursion between a replacement field and the format spec it contains,
/// so that pathologically nested input cannot exhaust the stack.
const MAX_REPLACEMENT_FIELD_DEPTH: usize = 32;

fn replacement_field_error(
    bytes: &[u8],
    open: usize,
    end: usize,
    prefix: &str,
    depth: usize,
) -> Option<CpythonDiagnostic> {
    let expr_start = skip_replacement_field_trivia(bytes, open + 1, end);

    // The expression runs up to the field's first top-level `=`, `!`, `:` or `}`. Everything past
    // it is a conversion or a format spec, which are text rather than code, so the checks that
    // treat the field as code have to stop here.
    let separator = replacement_field_separator(bytes, expr_start, end);
    let expr_end = separator.unwrap_or(end);

    if let Some(backslash) = replacement_field_line_continuation(bytes, expr_start, expr_end) {
        return Some(CpythonDiagnostic::new(
            "unexpected character after line continuation character".to_owned(),
            backslash + 1,
            (backslash + 2).min(end),
        ));
    }
    if let Some(quote) = unterminated_string_in_replacement_field(bytes, expr_start, expr_end) {
        return Some(CpythonDiagnostic::new(
            unterminated_string_message(1, false, false, None),
            quote,
            quote + 1,
        ));
    }
    // Brackets are reported ahead of everything else in the field, and a comment that swallows
    // the closing brace ahead of the expression itself. The comment scan starts at the brace
    // because a comment may also sit between it and the expression.
    if let Some(error) = replacement_field_bracket_error(bytes, expr_start, expr_end, prefix) {
        return Some(error);
    }
    if let Some(error) = replacement_field_comment_error(bytes, open, open + 1, expr_end, end) {
        return Some(error);
    }

    // Nothing follows `{` at all, so the field simply never closed. CPython asks for the
    // brace here; an expression that merely fails to start is reported further down.
    if expr_start >= end {
        return Some(CpythonDiagnostic::new(
            format!("{prefix}: expecting '}}'"),
            open,
            open + 1,
        ));
    }

    if let marker @ (b'=' | b'!' | b':' | b'}') = bytes[expr_start]
        && is_replacement_field_marker(bytes, expr_start)
    {
        return Some(CpythonDiagnostic::new(
            format!(
                "{prefix}: valid expression required before '{}'",
                marker as char
            ),
            expr_start,
            expr_start + 1,
        ));
    }

    // A `{` display is a valid expression, but only if something can follow it. CPython
    // reports the inner brace when it cannot, as in `f'{3:{{>10}'`.
    if bytes[expr_start] == b'{' {
        let inner = skip_ascii_whitespace(bytes, expr_start + 1, end);
        if inner < end
            && bytes[inner] != b'}'
            && invalid_replacement_expression_start(bytes, inner, end)
        {
            return Some(CpythonDiagnostic::new(
                format!("{prefix}: expecting a valid expression after '{{'"),
                expr_start,
                expr_start + 1,
            ));
        }
    }

    if starts_identifier(bytes, expr_start, b"lambda") {
        return Some(CpythonDiagnostic::new(
            format!("{prefix}: lambda expressions are not allowed without parentheses"),
            expr_start,
            expr_start + b"lambda".len(),
        ));
    }

    if invalid_replacement_expression_start(bytes, expr_start, end) {
        return Some(CpythonDiagnostic::new(
            format!("{prefix}: expecting a valid expression after '{{'"),
            open,
            open + 1,
        ));
    }

    // Adjacent atoms in a replacement field are a missing comma, not an
    // f-string "expecting '}'" at the opening brace. Prefix `yield`/`await`/
    // `not` and continuation keywords (`and`, `for`, ...) are not a second atom.
    if !starts_identifier(bytes, expr_start, b"yield")
        && !starts_identifier(bytes, expr_start, b"await")
        && !starts_identifier(bytes, expr_start, b"not")
        && let Some(atom_end) = adjacent_atom_end(bytes, expr_start)
    {
        let next = skip_ascii_whitespace(bytes, atom_end, expr_end);
        if next > atom_end
            && expression_atom_start(bytes, next)
            && !expression_continuation_keyword(bytes, next)
        {
            let second_end = adjacent_atom_end(bytes, next).unwrap_or(next + 1);
            return Some(CpythonDiagnostic::new(
                "invalid syntax. Perhaps you forgot a comma?".to_owned(),
                expr_start,
                second_end,
            ));
        }
    }

    // The expression started fine but ran into a character that cannot continue it. CPython
    // points at that character and lists the separators it wanted instead.
    if let Some(stray) = replacement_expression_stray_character(bytes, expr_start, expr_end) {
        return Some(CpythonDiagnostic::new(
            format!("{prefix}: expecting '=', or '!', or ':', or '}}'"),
            stray,
            stray + 1,
        ));
    }

    let Some(separator) = separator else {
        return Some(CpythonDiagnostic::new(
            format!("{prefix}: expecting '}}'"),
            open,
            open + 1,
        ));
    };

    // Only now that the field is known to close. A stray character above is a finished token
    // the parser rejects on lookahead, but a dangling operator makes it ask for one more, and
    // producing that token runs into the literal's closing quote: lexer.c then answers from
    // `INSIDE_FSTRING(tok)` with "%c-string: expecting '}'" before any `invalid_` rule is tried.
    if let Some(operator) = dangling_operator(bytes, expr_start, expr_end) {
        return Some(CpythonDiagnostic::new(
            format!("{prefix}: expecting '=', or '!', or ':', or '}}'"),
            operator,
            operator + 1,
        ));
    }

    if let Some(lambda_at) = top_level_lambda(bytes, expr_start, expr_end)
        && lambda_allowed_at_expression_position(bytes, expr_start, lambda_at)
    {
        return Some(CpythonDiagnostic::new(
            format!("{prefix}: lambda expressions are not allowed without parentheses"),
            lambda_at,
            lambda_at + b"lambda".len(),
        ));
    }

    if bytes[separator] == b':'
        && replacement_expression_has_parse_error(bytes, expr_start, separator)
    {
        return Some(CpythonDiagnostic::new(
            "invalid syntax".to_owned(),
            expr_start,
            separator,
        ));
    }

    match bytes[separator] {
        b'=' => invalid_debug_expression_error(bytes, separator, end, prefix, depth),
        b'!' => invalid_conversion_error(bytes, separator, end, prefix, depth),
        b':' => invalid_format_spec_error(bytes, separator, end, prefix, depth),
        b'}' => None,
        _ => unreachable!(),
    }
}

fn replacement_field_line_continuation(
    bytes: &[u8],
    mut index: usize,
    end: usize,
) -> Option<usize> {
    let mut level = 0usize;
    while index < end {
        match bytes[index] {
            b'\'' | b'"' => index = skip_quoted_string(bytes, index),
            b'#' => index = skip_replacement_field_comment(bytes, index, end),
            b'\\' => return Some(index),
            b'(' | b'[' | b'{' => {
                level += 1;
                index += 1;
            }
            b')' | b']' | b'}' if level > 0 => {
                level -= 1;
                index += 1;
            }
            b'=' | b'!' | b':' | b'}'
                if level == 0 && is_replacement_field_marker(bytes, index) =>
            {
                return None;
            }
            _ => index += 1,
        }
    }
    None
}

fn unterminated_string_in_replacement_field(
    bytes: &[u8],
    mut index: usize,
    end: usize,
) -> Option<usize> {
    while index < end {
        match bytes[index] {
            quote @ (b'\'' | b'"') => {
                let string_end = skip_quoted_string(bytes, index);
                if string_end >= end && !bytes[index + 1..end].contains(&quote) {
                    return Some(index);
                }
                index = string_end;
            }
            b'#' => index = skip_replacement_field_comment(bytes, index, end),
            _ => index += 1,
        }
    }
    None
}

fn top_level_lambda(bytes: &[u8], mut index: usize, end: usize) -> Option<usize> {
    let mut level = 0usize;
    while index < end {
        match bytes[index] {
            b'\'' | b'"' => index = skip_quoted_string(bytes, index),
            b'#' => index = skip_replacement_field_comment(bytes, index, end),
            b'(' | b'[' | b'{' => {
                level += 1;
                index += 1;
            }
            b')' | b']' | b'}' if level > 0 => {
                level -= 1;
                index += 1;
            }
            _ if level == 0 && starts_identifier(bytes, index, b"lambda") => {
                return Some(index);
            }
            _ => index += 1,
        }
    }
    None
}

/// Unparenthesized `lambda` is allowed in the same places a `lambdef` can appear:
/// at the start of the field or after a top-level comma.
fn lambda_allowed_at_expression_position(bytes: &[u8], mut index: usize, lambda_at: usize) -> bool {
    let mut after_comma = true;
    let mut level = 0usize;
    while index < lambda_at {
        match bytes[index] {
            b'\'' | b'"' => {
                after_comma = false;
                index = skip_quoted_string(bytes, index);
            }
            b'#' => index = skip_replacement_field_comment(bytes, index, lambda_at),
            b'(' | b'[' | b'{' => {
                level += 1;
                after_comma = false;
                index += 1;
            }
            b')' | b']' | b'}' if level > 0 => {
                level -= 1;
                after_comma = false;
                index += 1;
            }
            b',' if level == 0 => {
                after_comma = true;
                index += 1;
            }
            b' ' | b'\t' | b'\n' | b'\r' | b'\x0c' => index += 1,
            _ => {
                after_comma = false;
                index += 1;
            }
        }
    }
    after_comma
}

fn replacement_expression_has_parse_error(bytes: &[u8], start: usize, end: usize) -> bool {
    let Ok(expression) = ::core::str::from_utf8(&bytes[start..end]) else {
        return false;
    };
    parser::parse_expression(expression).is_err()
}

fn invalid_replacement_expression_start(bytes: &[u8], index: usize, end: usize) -> bool {
    if index >= end {
        return true;
    }

    // A leading `.` is Ellipsis or a float without an integer part, both of which start a
    // perfectly good expression.
    if bytes[index] == b'.'
        && (bytes[index..end].starts_with(b"...")
            || bytes
                .get(index + 1)
                .is_some_and(|byte| byte.is_ascii_digit()))
    {
        return false;
    }

    if matches!(
        bytes[index],
        b'.' | b',' | b'*' | b'/' | b'%' | b'&' | b'|' | b'^' | b'<' | b'>' | b'@' | b'=' | b'!'
    ) || is_stray_in_replacement_expression(bytes[index])
    {
        return true;
    }

    if matches!(bytes[index], b'+' | b'-' | b'~') {
        let operand = skip_ascii_whitespace(bytes, index + 1, end);
        if starts_identifier(bytes, operand, b"lambda") {
            return true;
        }
        return !bytes.get(operand).is_some_and(|byte| {
            *byte >= 0x80
                || *byte == b'_'
                || byte.is_ascii_alphabetic()
                || byte.is_ascii_digit()
                || matches!(*byte, b'\'' | b'"' | b'(' | b'[' | b'{')
                // `-.5` is a signed float.
                || (*byte == b'.'
                    && bytes
                        .get(operand + 1)
                        .is_some_and(|next| next.is_ascii_digit()))
        });
    }

    [
        b"and".as_slice(),
        b"as".as_slice(),
        b"else".as_slice(),
        b"for".as_slice(),
        b"if".as_slice(),
        b"in".as_slice(),
        b"is".as_slice(),
        b"or".as_slice(),
    ]
    .iter()
    .any(|keyword| starts_identifier(bytes, index, keyword))
}

/// A top-level `=` or `!` marks a debug specifier or a conversion only when it is not part of a
/// longer operator. CPython's tokenizer consumes `==`, `!=`, `<=`, `+=` and the rest as single
/// tokens before it ever considers the debug marker, so those must not end the expression here.
fn is_replacement_field_marker(bytes: &[u8], index: usize) -> bool {
    match bytes[index] {
        b'=' => {
            if bytes.get(index + 1) == Some(&b'=') {
                return false;
            }
            !index
                .checked_sub(1)
                .and_then(|previous| bytes.get(previous).copied())
                .is_some_and(precedes_equals_in_one_operator)
        }
        b'!' => bytes.get(index + 1) != Some(&b'='),
        _ => true,
    }
}

fn replacement_field_separator(bytes: &[u8], mut index: usize, end: usize) -> Option<usize> {
    let mut level = 0usize;
    while index < end {
        match bytes[index] {
            b'\'' | b'"' => index = skip_quoted_string(bytes, index),
            b'#' => index = skip_replacement_field_comment(bytes, index, end),
            b'(' | b'[' | b'{' => {
                level += 1;
                index += 1;
            }
            b')' | b']' | b'}' if level > 0 => {
                level -= 1;
                index += 1;
            }
            b'=' | b'!' | b':' | b'}'
                if level == 0 && is_replacement_field_marker(bytes, index) =>
            {
                return Some(index);
            }
            _ => index += 1,
        }
    }
    None
}

/// Characters that can neither start nor continue a replacement field expression. CPython's
/// interpolated-string tokenizer stops at them and asks for a separator instead.
const fn is_stray_in_replacement_expression(byte: u8) -> bool {
    matches!(byte, b';' | b'$' | b'?' | b'`')
}

fn replacement_expression_stray_character(
    bytes: &[u8],
    mut index: usize,
    end: usize,
) -> Option<usize> {
    while index < end {
        match bytes[index] {
            b'\'' | b'"' => index = skip_quoted_string(bytes, index),
            b'#' => index = skip_replacement_field_comment(bytes, index, end),
            byte if is_stray_in_replacement_expression(byte) => return Some(index),
            _ => index += 1,
        }
    }
    None
}

/// A byte that pairs with a following `=` to make one operator token, so that the `=` belongs
/// to it rather than opening a debug specifier: `==`, `!=`, `<=`, `>=`, the augmented
/// assignments, and the `:=` walrus. Against `is_dangling_operator_byte` below, this has `:`
/// and lacks `.` and `~`, because neither `.=` nor `~=` is a Python operator.
const fn precedes_equals_in_one_operator(byte: u8) -> bool {
    matches!(
        byte,
        b'!' | b'%'
            | b'&'
            | b'*'
            | b'+'
            | b'-'
            | b'/'
            | b':'
            | b'<'
            | b'='
            | b'>'
            | b'@'
            | b'^'
            | b'|'
    )
}

/// Characters that make up an operator which cannot end an expression. Against
/// `precedes_equals_in_one_operator` above, this has `.` and `~` for `a.b.` and `~a`, and lacks
/// `:`, which separates a replacement field's parts rather than operating on anything.
const fn is_dangling_operator_byte(byte: u8) -> bool {
    matches!(
        byte,
        b'!' | b'%'
            | b'&'
            | b'*'
            | b'+'
            | b'-'
            | b'.'
            | b'/'
            | b'<'
            | b'='
            | b'>'
            | b'@'
            | b'^'
            | b'|'
            | b'~'
    )
}

/// The identifier that ends just before `word`, skipping the whitespace between them.
fn preceding_keyword(bytes: &[u8], start: usize, word: usize) -> Option<(usize, &[u8])> {
    let mut cursor = word;
    while cursor > start && matches!(bytes[cursor - 1], b' ' | b'\t' | b'\r' | b'\n' | 0x0c) {
        cursor -= 1;
    }
    let stop = cursor;
    while cursor > start && is_ascii_identifier_char(bytes[cursor - 1]) {
        cursor -= 1;
    }
    (cursor < stop).then(|| (cursor, &bytes[cursor..stop]))
}

/// A replacement field expression cannot end on an operator that still wants an operand.
/// CPython's tokenizer then reaches the field brace with the expression unfinished and asks for
/// a separator, pointing at the operator that left it that way. Returns that operator's start.
fn dangling_operator(bytes: &[u8], start: usize, end: usize) -> Option<usize> {
    // The region cannot be read backwards: the newline that ends a comment is whitespace, so
    // trimming from `end` would step into the comment body. Walk forward instead and keep the
    // position just past the last byte that is really part of the expression.
    let mut index = start;
    let mut tail = start;
    while index < end {
        match bytes[index] {
            b'\'' | b'"' => {
                index = skip_quoted_string(bytes, index).min(end);
                tail = index;
            }
            b'#' => index = skip_replacement_field_comment(bytes, index, end),
            b' ' | b'\t' | b'\r' | b'\n' | 0x0c => index += 1,
            _ => {
                index += 1;
                tail = index;
            }
        }
    }
    if tail == start {
        return None;
    }

    // A trailing keyword operator reads back as an identifier.
    if is_ascii_identifier_char(bytes[tail - 1]) {
        let mut word = tail;
        while word > start && is_ascii_identifier_char(bytes[word - 1]) {
            word -= 1;
        }
        if ![
            b"and".as_slice(),
            b"or".as_slice(),
            b"not".as_slice(),
            b"in".as_slice(),
            b"is".as_slice(),
            b"if".as_slice(),
            b"for".as_slice(),
        ]
        .iter()
        .any(|keyword| &bytes[word..tail] == *keyword)
        {
            return None;
        }
        // `is not` and `not in` are single operators, and CPython points at their first word.
        let previous = preceding_keyword(bytes, start, word);
        return Some(match (previous, &bytes[word..tail]) {
            (Some((at, b"is")), b"not") | (Some((at, b"not")), b"in") => at,
            _ => word,
        });
    }

    if !is_dangling_operator_byte(bytes[tail - 1]) {
        return None;
    }
    let mut token = tail;
    while token > start && is_dangling_operator_byte(bytes[token - 1]) {
        token -= 1;
    }

    // `1.` is a float: that dot finishes the literal rather than dangling.
    if bytes[token..tail] == *b"." && token > start && bytes[token - 1].is_ascii_digit() {
        return None;
    }
    // `...` is Ellipsis, so consume whole triples: a leftover dot is what dangles.
    if bytes[token..tail].iter().all(|byte| *byte == b'.') {
        let leftover = (tail - token) % 3;
        return (leftover != 0).then(|| tail - leftover);
    }
    Some(token)
}

fn invalid_debug_expression_error(
    bytes: &[u8],
    equals: usize,
    end: usize,
    prefix: &str,
    depth: usize,
) -> Option<CpythonDiagnostic> {
    let next = equals + 1;
    if next >= end {
        return None;
    }
    // A conversion or format spec may follow `=`, and is validated exactly as it would be
    // when following the expression directly.
    match bytes[next] {
        b'!' => return invalid_conversion_error(bytes, next, end, prefix, depth),
        b':' => return invalid_format_spec_error(bytes, next, end, prefix, depth),
        b'}' => return None,
        _ => {}
    }
    Some(CpythonDiagnostic::new(
        format!("{prefix}: expecting '!', or ':', or '}}'"),
        next,
        next.saturating_add(1).min(end),
    ))
}

fn invalid_conversion_error(
    bytes: &[u8],
    bang: usize,
    end: usize,
    prefix: &str,
    depth: usize,
) -> Option<CpythonDiagnostic> {
    let next = bang + 1;
    if next >= end {
        return Some(CpythonDiagnostic::new(
            format!("{prefix}: expecting '}}'"),
            bang,
            bang + 1,
        ));
    }

    if bytes[next].is_ascii_whitespace() {
        let following = skip_ascii_whitespace(bytes, next, end);
        let message = if bytes
            .get(following)
            .is_some_and(|byte| byte.is_ascii_alphabetic() || *byte == b'_')
        {
            "conversion type must come right after the exclamation mark"
        } else {
            "missing conversion character"
        };
        return Some(CpythonDiagnostic::new(
            format!("{prefix}: {message}"),
            next,
            next + 1,
        ));
    }

    if matches!(bytes[next], b':' | b'}') {
        return Some(CpythonDiagnostic::new(
            format!("{prefix}: missing conversion character"),
            next,
            next + 1,
        ));
    }

    // A non-ASCII character is still a conversion character as far as CPython is concerned,
    // so it gets named in the message like any other invalid one.
    if bytes[next] >= 0x80
        && let Some(character) = ::core::str::from_utf8(&bytes[next..end])
            .ok()
            .and_then(|text| text.chars().next())
    {
        return Some(CpythonDiagnostic::new(
            format!(
                "{prefix}: invalid conversion character '{character}': expected 's', 'r', or 'a'"
            ),
            next,
            next + character.len_utf8(),
        ));
    }

    if !bytes[next].is_ascii_alphabetic() && bytes[next] != b'_' {
        return Some(CpythonDiagnostic::new(
            format!("{prefix}: invalid conversion character"),
            next,
            next + 1,
        ));
    }

    let conversion_end = identifier_end(bytes, next, end);
    let conversion = &bytes[next..conversion_end];
    if !matches!(conversion, b"s" | b"r" | b"a") {
        let conversion = ::core::str::from_utf8(conversion).unwrap_or("");
        return Some(CpythonDiagnostic::new(
            format!(
                "{prefix}: invalid conversion character '{conversion}': expected 's', 'r', or 'a'"
            ),
            next,
            conversion_end,
        ));
    }

    if conversion_end >= end {
        return None;
    }

    match bytes[conversion_end] {
        b':' => return invalid_format_spec_error(bytes, conversion_end, end, prefix, depth),
        b'}' => return None,
        _ => {}
    }

    Some(CpythonDiagnostic::new(
        format!("{prefix}: expecting ':' or '}}'"),
        conversion_end,
        conversion_end + 1,
    ))
}

fn invalid_format_spec_error(
    bytes: &[u8],
    colon: usize,
    end: usize,
    prefix: &str,
    depth: usize,
) -> Option<CpythonDiagnostic> {
    // Unlike literal text, a format spec has no `{{` escape: every `{` opens a nested
    // replacement field, so the outer scanner cannot validate these for us.
    let mut index = colon + 1;
    while index < end {
        match bytes[index] {
            b'\'' | b'"' => index = skip_quoted_string(bytes, index),
            b'{' => {
                if depth < MAX_REPLACEMENT_FIELD_DEPTH
                    && let Some(error) =
                        replacement_field_error(bytes, index, end, prefix, depth + 1)
                {
                    return Some(error);
                }
                // Resume after the nested field. Walking into it again from every enclosing
                // spec would re-scan the same bytes once per level, which costs O(2^depth).
                let Some(after) = replacement_field_end(bytes, index, end) else {
                    break;
                };
                index = after;
            }
            b'}' => return None,
            _ => index += 1,
        }
    }
    Some(CpythonDiagnostic::new(
        format!("{prefix}: expecting '}}', or format specs"),
        colon,
        colon + 1,
    ))
}

fn replacement_field_closing_brace(bytes: &[u8], mut index: usize, end: usize) -> Option<usize> {
    let mut level = 0usize;
    while index < end {
        match bytes[index] {
            b'\'' | b'"' => index = skip_quoted_string(bytes, index),
            b'{' => {
                level += 1;
                index += 1;
            }
            b'}' if level > 0 => {
                level -= 1;
                index += 1;
            }
            b'}' => return Some(index),
            _ => index += 1,
        }
    }
    None
}

fn skip_ascii_whitespace(bytes: &[u8], mut index: usize, end: usize) -> usize {
    while index < end && matches!(bytes[index], b' ' | b'\t' | b'\r' | b'\n' | 0x0c) {
        index += 1;
    }
    index
}

fn string_literal_end_at(bytes: &[u8], index: usize) -> Option<usize> {
    match bytes.get(index).copied()? {
        b'\'' | b'"' => Some(skip_quoted_string(bytes, index)),
        first if first.is_ascii_alphabetic() => {
            if matches!(bytes.get(index + 1), Some(b'\'' | b'"')) {
                return string_literal_prefix(bytes, index, index + 1)
                    .then(|| skip_quoted_string(bytes, index + 1));
            }
            if matches!(bytes.get(index + 2), Some(b'\'' | b'"')) {
                return string_literal_prefix(bytes, index, index + 2)
                    .then(|| skip_quoted_string(bytes, index + 2));
            }
            None
        }
        _ => None,
    }
}

fn string_literal_prefix(bytes: &[u8], start: usize, quote: usize) -> bool {
    let prefix = &bytes[start..quote];
    let valid = matches!(
        prefix,
        b"b" | b"B"
            | b"r"
            | b"R"
            | b"u"
            | b"U"
            | b"f"
            | b"F"
            | b"t"
            | b"T"
            | b"br"
            | b"bR"
            | b"Br"
            | b"BR"
            | b"rb"
            | b"rB"
            | b"Rb"
            | b"RB"
            | b"fr"
            | b"fR"
            | b"Fr"
            | b"FR"
            | b"rf"
            | b"rF"
            | b"Rf"
            | b"RF"
            | b"tr"
            | b"tR"
            | b"Tr"
            | b"TR"
            | b"rt"
            | b"rT"
            | b"Rt"
            | b"RT"
    );
    valid && (start == 0 || !is_ascii_identifier_char(bytes[start - 1]))
}

fn invalid_expression_error(source: &str) -> Option<CpythonDiagnostic> {
    invalid_string_expression_error(source).or_else(|| missing_comma_expression_error(source))
}

fn invalid_string_expression_error(source: &str) -> Option<CpythonDiagnostic> {
    let bytes = source.as_bytes();
    let mut index = 0;
    while index < bytes.len() {
        if let Some(first_string_end) = string_literal_end_at(bytes, index) {
            let expr_start = skip_ascii_whitespace(bytes, first_string_end, bytes.len());
            if expression_atom_start(bytes, expr_start)
                && let Some(expr_end) = adjacent_atom_end(bytes, expr_start)
            {
                let next = skip_ascii_whitespace(bytes, expr_end, bytes.len());
                if string_literal_end_at(bytes, next).is_some() {
                    return Some(CpythonDiagnostic::new(
                        "invalid syntax. Is this intended to be part of the string?".to_owned(),
                        expr_start,
                        expr_end,
                    ));
                }
            }
            index = first_string_end;
        } else {
            index += 1;
        }
    }
    None
}

fn missing_comma_expression_error(source: &str) -> Option<CpythonDiagnostic> {
    let bytes = source.as_bytes();
    let mut stack: Vec<u8> = Vec::new();
    let mut index = 0;
    while index < bytes.len() {
        if bytes[index] == b'#' {
            while index < bytes.len() && bytes[index] != b'\n' {
                index += 1;
            }
        } else if let Some(string_end) = string_literal_end_at(bytes, index) {
            index = string_end;
        } else {
            match bytes[index] {
                b'(' | b'[' | b'{' => {
                    if bytes[index] == b'[' && opening_bracket_is_class_type_params(bytes, index) {
                        let Some(close) = matching_delimiter(bytes, index, b']') else {
                            index += 1;
                            continue;
                        };
                        index = close + 1;
                        continue;
                    }
                    stack.push(bytes[index]);
                    index += 1;
                }
                b')' | b']' | b'}' => {
                    stack.pop();
                    index += 1;
                }
                _ if !stack.is_empty() && expression_continuation_keyword(bytes, index) => {
                    index = identifier_end(bytes, index, bytes.len());
                }
                byte if !stack.is_empty() && expression_atom_start_byte(byte) => {
                    // `yield x` / `await x` are prefix expressions, not two
                    // adjacent atoms missing a comma.
                    if starts_identifier(bytes, index, b"yield") {
                        index = identifier_end(bytes, index, bytes.len());
                        let after_yield = skip_ascii_whitespace(bytes, index, bytes.len());
                        if starts_identifier(bytes, after_yield, b"from") {
                            index = identifier_end(bytes, after_yield, bytes.len());
                        }
                        continue;
                    }
                    if starts_identifier(bytes, index, b"await") {
                        index = identifier_end(bytes, index, bytes.len());
                        continue;
                    }
                    let atom_end = adjacent_atom_end(bytes, index).unwrap_or(index + 1);
                    let next = skip_ascii_whitespace(bytes, atom_end, bytes.len());
                    if next > atom_end
                        && expression_atom_start(bytes, next)
                        && !expression_continuation_keyword(bytes, next)
                    {
                        // The diagnostic covers both atoms, so it ends past
                        // the whole second one.  `next + 1` ended past its
                        // first byte instead, which is only the same thing
                        // when that atom is one ASCII character.
                        let second_end = adjacent_atom_end(bytes, next).unwrap_or(next + 1);
                        return Some(CpythonDiagnostic::new(
                            "invalid syntax. Perhaps you forgot a comma?".to_owned(),
                            index,
                            second_end,
                        ));
                    }
                    index = atom_end;
                }
                _ => index += 1,
            }
        }
    }
    None
}

fn opening_bracket_is_class_type_params(bytes: &[u8], bracket: usize) -> bool {
    let mut cursor = bracket;
    while cursor > 0 && matches!(bytes[cursor - 1], b' ' | b'\t' | b'\x0c') {
        cursor -= 1;
    }
    while cursor > 0
        && bytes
            .get(cursor - 1)
            .is_some_and(|byte| *byte >= 0x80 || is_ascii_identifier_char(*byte))
    {
        cursor -= 1;
    }
    while cursor > 0 && matches!(bytes[cursor - 1], b' ' | b'\t' | b'\x0c') {
        cursor -= 1;
    }
    cursor >= 5 && starts_identifier(bytes, cursor - 5, b"class")
}

fn expression_continuation_keyword(bytes: &[u8], index: usize) -> bool {
    [
        b"and".as_slice(),
        b"else".as_slice(),
        b"for".as_slice(),
        b"if".as_slice(),
        b"in".as_slice(),
        b"is".as_slice(),
        b"not".as_slice(),
        b"or".as_slice(),
    ]
    .iter()
    .any(|keyword| starts_identifier(bytes, index, keyword))
}

fn expression_atom_start(bytes: &[u8], index: usize) -> bool {
    bytes
        .get(index)
        .is_some_and(|byte| expression_atom_start_byte(*byte))
        || string_literal_end_at(bytes, index).is_some()
}

fn expression_atom_start_byte(byte: u8) -> bool {
    byte >= 0x80
        || byte == b'_'
        || byte.is_ascii_alphabetic()
        || byte.is_ascii_digit()
        || matches!(byte, b'\'' | b'"' | b'(' | b'[' | b'{')
}

fn adjacent_atom_end(bytes: &[u8], index: usize) -> Option<usize> {
    if let Some(string_end) = string_literal_end_at(bytes, index) {
        return Some(string_end);
    }
    match bytes.get(index).copied()? {
        byte if byte >= 0x80 || byte == b'_' || byte.is_ascii_alphabetic() => {
            Some(identifier_end(bytes, index, bytes.len()))
        }
        byte if byte.is_ascii_digit() => {
            let mut end = index + 1;
            while end < bytes.len() && (bytes[end].is_ascii_alphanumeric() || bytes[end] == b'_') {
                end += 1;
            }
            Some(end)
        }
        b'(' | b'[' | b'{' => Some(index + 1),
        _ => None,
    }
}

struct OpenDelimiter {
    position: usize,
    opener: u8,
    in_format_spec: bool,
}

/// An f-string or t-string that runs off the end with a replacement field still open reports
/// the brace rather than the missing quote, matching CPython — but only while the tokenizer is
/// still reading the field's *expression*. Past the field's own `:` it is emitting FSTRING_MIDDLE
/// again, so running out of input there is an unterminated literal like any other, and
/// `f'{a:>5` gets the quote message where `f'{a` gets the brace.
fn unclosed_replacement_field_error(
    bytes: &[u8],
    quote_start: usize,
    content_start: usize,
    content_end: usize,
) -> Option<CpythonDiagnostic> {
    interpolated_string_prefix(bytes, quote_start)?;

    // Brackets and quotes are delimiters only once a field is open; in the literal's text they
    // are ordinary characters. And past a field's own format spec the tokenizer emits literal
    // text again, so a field still open at end of input is an unclosed brace only before that.
    let mut open: Vec<OpenDelimiter> = Vec::new();
    let mut index = content_start;
    while index < content_end {
        let inside_expression = matches!(
            open.last(),
            Some(OpenDelimiter {
                opener: b'{',
                in_format_spec: false,
                ..
            })
        );
        match bytes[index] {
            // `{{` and `}}` escape only in literal text, which is where a brace at depth zero is.
            b'{' if open.is_empty() && bytes.get(index + 1) == Some(&b'{') => index += 2,
            b'}' if open.is_empty() && bytes.get(index + 1) == Some(&b'}') => index += 2,
            b'{' => {
                open.push(OpenDelimiter {
                    position: index,
                    opener: b'{',
                    in_format_spec: false,
                });
                index += 1;
            }
            b'}' => {
                open.pop();
                index += 1;
            }
            byte @ (b'(' | b'[') if !open.is_empty() => {
                open.push(OpenDelimiter {
                    position: index,
                    opener: byte,
                    in_format_spec: false,
                });
                index += 1;
            }
            b')' | b']' if !open.is_empty() => {
                open.pop();
                index += 1;
            }
            // Only the field's own `:` opens a format spec — one in a slice, a display or a
            // lambda belongs to whatever bracket encloses it.
            b':' if inside_expression => {
                open.last_mut().expect("a brace is open").in_format_spec = true;
                index += 1;
            }
            b'\'' | b'"' if !open.is_empty() => {
                index = skip_quoted_string(bytes, index).min(content_end);
            }
            b'#' if inside_expression => {
                index = skip_replacement_field_comment(bytes, index, content_end);
            }
            _ => index += 1,
        }
    }

    // CPython names the innermost unclosed delimiter, so a bracket opened inside the expression
    // takes the message from the field that encloses it.
    let &OpenDelimiter {
        position,
        opener,
        in_format_spec,
    } = open.last()?;
    (!in_format_spec).then(|| {
        CpythonDiagnostic::new(
            format!("'{}' was never closed", opener as char),
            position,
            position + 1,
        )
        .with_unclosed_bracket()
    })
}

fn unterminated_string_message(
    detected_line: usize,
    triple: bool,
    has_escaped_quote: bool,
    prefix: Option<&str>,
) -> String {
    // CPython names the literal kind, e.g. "unterminated f-string literal".
    let kind = prefix.unwrap_or("string");
    if triple {
        format!("unterminated triple-quoted {kind} literal (detected at line {detected_line})")
    // The escaped-quote hint belongs to the plain-string branch of CPython's tokenizer alone.
    // Its interpolated branch has only the two forms above and below, so pairing the hint with
    // a prefix would invent a message CPython never emits.
    } else if has_escaped_quote && prefix.is_none() {
        format!(
            "unterminated {kind} literal (detected at line {detected_line}); perhaps you escaped the end quote?"
        )
    } else {
        format!("unterminated {kind} literal (detected at line {detected_line})")
    }
}

fn expected_opening_bracket(closing: char) -> char {
    match closing {
        ')' => '(',
        ']' => '[',
        '}' => '{',
        _ => unreachable!(),
    }
}

/// A bracket diagnostic, and whether it is an opener that was never closed. The caller needs
/// that apart from the message because ruff reports the unclosed case as an EOF error.
#[derive(Clone)]
struct BracketError {
    diagnostic: CpythonDiagnostic,
    unclosed: bool,
}

fn bracket_syntax_error(source: &str) -> Option<BracketError> {
    let mut stack: Vec<(char, usize, usize)> = Vec::new();
    let mut in_string = false;
    let mut string_quote = '\0';
    let mut triple_quote = false;
    let mut escape_next = false;
    let mut is_raw_string = false;
    let mut line = 1usize;

    let chars: Vec<(usize, char)> = source.char_indices().collect();
    let mut index = 0;
    while index < chars.len() {
        let (byte_offset, ch) = chars[index];

        if ch == '\n' {
            line += 1;
        }

        if escape_next {
            escape_next = false;
            index += 1;
            continue;
        }

        if in_string {
            if ch == '\\' && !is_raw_string {
                escape_next = true;
            } else if triple_quote {
                if ch == string_quote
                    && index + 2 < chars.len()
                    && chars[index + 1].1 == string_quote
                    && chars[index + 2].1 == string_quote
                {
                    in_string = false;
                    index += 3;
                    continue;
                }
            } else if ch == string_quote {
                in_string = false;
            }
            index += 1;
            continue;
        }

        if ch == '#' {
            while index < chars.len() && chars[index].1 != '\n' {
                index += 1;
            }
            continue;
        }

        if ch == '\\' {
            match chars.get(index + 1).map(|(_, next)| *next) {
                Some('\n' | '\r') => {
                    escape_next = true;
                    index += 1;
                    continue;
                }
                Some(_) => {
                    index += 2;
                    continue;
                }
                None => {
                    index += 1;
                    continue;
                }
            }
        }

        if ch == '\'' || ch == '"' {
            is_raw_string = false;
            for look_back in 1..=2.min(index) {
                let prev = chars[index - look_back].1;
                if matches!(prev, 'r' | 'R') {
                    is_raw_string = true;
                    break;
                }
                if !matches!(prev, 'b' | 'B' | 'f' | 'F' | 'u' | 'U') {
                    break;
                }
            }
            string_quote = ch;
            if index + 2 < chars.len() && chars[index + 1].1 == ch && chars[index + 2].1 == ch {
                triple_quote = true;
                in_string = true;
                index += 3;
                continue;
            }
            triple_quote = false;
            in_string = true;
            index += 1;
            continue;
        }

        match ch {
            '(' | '[' | '{' => stack.push((ch, byte_offset, line)),
            ')' | ']' | '}' => {
                let expected = expected_opening_bracket(ch);
                let Some(&(opening, _, opening_line)) = stack.last() else {
                    return Some(BracketError {
                        diagnostic: CpythonDiagnostic::new(
                            format!("unmatched '{ch}'"),
                            byte_offset,
                            byte_offset,
                        ),
                        unclosed: false,
                    });
                };
                if opening == expected {
                    stack.pop();
                } else {
                    let suffix = if opening_line != line {
                        format!(" on line {opening_line}")
                    } else {
                        String::new()
                    };
                    return Some(BracketError {
                        diagnostic: CpythonDiagnostic::new(
                            format!(
                                "closing parenthesis '{ch}' does not match opening parenthesis '{opening}'{suffix}"
                            ),
                            byte_offset,
                            byte_offset,
                        ),
                        unclosed: false,
                    });
                }
            }
            _ => {}
        }

        index += 1;
    }

    stack.last().map(|(opening, byte_offset, _)| BracketError {
        diagnostic: CpythonDiagnostic::new(
            format!("'{opening}' was never closed"),
            *byte_offset,
            *byte_offset,
        ),
        unclosed: true,
    })
}

fn is_legacy_statement_expression_start(byte: u8) -> bool {
    byte >= 0x80
        || byte == b'_'
        || byte.is_ascii_alphabetic()
        || byte.is_ascii_digit()
        || matches!(byte, b'\'' | b'"' | b'{' | b'[')
}

fn legacy_statement_container_has_invalid_attribute(bytes: &[u8], start: usize) -> bool {
    let Some(&opening) = bytes.get(start) else {
        return false;
    };
    if !matches!(opening, b'{' | b'[') {
        return false;
    }

    let mut index = start;
    let mut level = 0usize;
    while index < bytes.len() {
        match bytes[index] {
            b'#' => {
                while index < bytes.len() && bytes[index] != b'\n' {
                    index += 1;
                }
            }
            b'\n' | b';' if level == 0 => return false,
            b'\'' | b'"' => {
                index = skip_quoted_string(bytes, index);
            }
            b'(' | b'[' | b'{' => {
                level += 1;
                index += 1;
            }
            b')' | b']' | b'}' => {
                level = level.saturating_sub(1);
                index += 1;
                if level == 0 {
                    return false;
                }
            }
            b'.' => {
                let mut cursor = index + 1;
                while matches!(bytes.get(cursor), Some(b' ' | b'\t' | b'\x0c')) {
                    cursor += 1;
                }
                if matches!(bytes.get(cursor), Some(b')' | b']' | b'}')) {
                    return true;
                }
                index += 1;
            }
            _ => index += 1,
        }
    }
    false
}

fn invalid_legacy_statement_error(source: &str) -> Option<CpythonDiagnostic> {
    let bytes = source.as_bytes();
    let mut index = 0;
    while index < bytes.len() {
        match bytes[index] {
            b'#' => {
                while index < bytes.len() && bytes[index] != b'\n' {
                    index += 1;
                }
            }
            b'\'' | b'"' => {
                index = skip_quoted_string(bytes, index);
            }
            b'p' | b'e' => {
                let keyword = if starts_identifier(bytes, index, b"print") {
                    Some("print")
                } else if starts_identifier(bytes, index, b"exec") {
                    Some("exec")
                } else {
                    None
                };
                let Some(keyword) = keyword else {
                    index += 1;
                    continue;
                };
                let after_keyword = index + keyword.len();
                if !matches!(bytes.get(after_keyword), Some(b' ' | b'\t' | b'\x0c')) {
                    index = after_keyword;
                    continue;
                }
                let mut cursor = after_keyword;
                while matches!(bytes.get(cursor), Some(b' ' | b'\t' | b'\x0c')) {
                    cursor += 1;
                }
                if legacy_statement_container_has_invalid_attribute(bytes, cursor) {
                    index = after_keyword;
                    continue;
                }
                if bytes.get(cursor).is_some_and(|byte| {
                    *byte != b'(' && is_legacy_statement_expression_start(*byte)
                }) {
                    return Some(CpythonDiagnostic::new(
                        format!(
                            "Missing parentheses in call to '{keyword}'. Did you mean {keyword}(...)?"
                        ),
                        index,
                        after_keyword,
                    ));
                }
                index = after_keyword;
            }
            _ => index += 1,
        }
    }
    None
}

/// Return the syntax error for a decimal integer literal exceeding the configured limit.
///
/// The parser has already distinguished integer literals from strings, comments, floats, and
/// complex numbers. Inspecting its tokens keeps the limit consistent for every source parsing
/// entry point without reimplementing Python's lexer here.
#[must_use]
pub fn long_decimal_integer_literal_error(
    source_file: &SourceFile,
    tokens: &Tokens,
    max_str_digits: usize,
) -> Option<CompileError> {
    if max_str_digits == 0 {
        return None;
    }
    tokens.iter().find_map(|token| {
        if token.kind() != TokenKind::Int {
            return None;
        }
        let literal = source_file.source_text().slice(token.range());
        if literal
            .as_bytes()
            .get(..2)
            .is_some_and(|prefix| matches!(prefix, b"0x" | b"0X" | b"0o" | b"0O" | b"0b" | b"0B"))
        {
            return None;
        }
        let digits = literal.bytes().filter(u8::is_ascii_digit).count();
        (digits > max_str_digits).then(|| {
            let start = token.range().start().to_usize();
            CompileError::from_source_error(source_file, CpythonDiagnostic::new(format!(
                    "Exceeds the limit ({max_str_digits} digits) for integer string conversion: value has {digits} digits; use sys.set_int_max_str_digits() to increase the limit - Consider hexadecimal for huge integer literals to avoid decimal conversion limits."
                ), start, start))
        })
    })
}

fn invalid_parenthesized_import_star_error(source: &str) -> Option<CpythonDiagnostic> {
    let bytes = source.as_bytes();
    let mut index = 0;
    while index < bytes.len() {
        match bytes[index] {
            b'#' => {
                while index < bytes.len() && bytes[index] != b'\n' {
                    index += 1;
                }
            }
            b'\'' | b'"' => {
                index = skip_quoted_string(bytes, index);
            }
            b'f' if starts_identifier(bytes, index, b"from") => {
                let mut cursor = index + 4;
                while cursor < bytes.len() && !matches!(bytes[cursor], b'\n' | b';') {
                    if starts_identifier(bytes, cursor, b"import") {
                        cursor += 6;
                        while matches!(bytes.get(cursor), Some(b' ' | b'\t' | b'\r')) {
                            cursor += 1;
                        }
                        if bytes.get(cursor) == Some(&b'(') {
                            cursor += 1;
                            while cursor < bytes.len()
                                && !matches!(bytes[cursor], b')' | b'\n' | b';')
                            {
                                if bytes[cursor] == b'*' {
                                    return Some(CpythonDiagnostic::new(
                                        "invalid syntax".to_owned(),
                                        cursor,
                                        cursor + 1,
                                    ));
                                }
                                cursor += 1;
                            }
                        }
                        break;
                    }
                    cursor += 1;
                }
                index = cursor;
            }
            _ => index += 1,
        }
    }
    None
}

fn prefix_letters_before_quote(bytes: &[u8], quote_index: usize) -> Option<(usize, Vec<u8>)> {
    let mut letters = Vec::new();
    let mut index = quote_index;
    while index > 0 {
        let byte = bytes[index - 1];
        if !matches!(
            byte,
            b'r' | b'R' | b'b' | b'B' | b'u' | b'U' | b'f' | b'F' | b't' | b'T'
        ) {
            break;
        }
        letters.push(byte.to_ascii_lowercase());
        index -= 1;
    }
    if letters.is_empty() {
        return None;
    }
    if index > 0 && (bytes[index - 1] == b'_' || bytes[index - 1].is_ascii_alphabetic()) {
        return None;
    }
    letters.reverse();
    Some((index, letters))
}

fn incompatible_string_prefix_error(source: &str) -> Option<CpythonDiagnostic> {
    let bytes = source.as_bytes();
    let mut index = 0;
    while index < bytes.len() {
        match bytes[index] {
            b'#' => {
                while index < bytes.len() && bytes[index] != b'\n' {
                    index += 1;
                }
            }
            b'\'' | b'"' => {
                if let Some((start, letters)) = prefix_letters_before_quote(bytes, index)
                    && let Some(message) = incompatible_prefix_message(&letters)
                {
                    return Some(CpythonDiagnostic::new(message, start, start + 1));
                }
                index = skip_string_token(bytes, index);
            }
            _ => index += 1,
        }
    }
    None
}

fn incompatible_prefix_message(letters: &[u8]) -> Option<String> {
    if letters.len() < 2 {
        return None;
    }
    let mut seen_r = false;
    let mut seen_b = false;
    let mut seen_u = false;
    let mut seen_f = false;
    let mut seen_t = false;
    for &letter in letters {
        match letter {
            b'r' if seen_r => return None,
            b'b' if seen_b => return None,
            b'u' if seen_u => return None,
            b'f' if seen_f => return None,
            b't' if seen_t => return None,
            b'r' => seen_r = true,
            b'b' => seen_b = true,
            b'u' => seen_u = true,
            b'f' => seen_f = true,
            b't' => seen_t = true,
            _ => {}
        }
    }
    let pair = if seen_u && seen_b {
        ("u", "b")
    } else if seen_u && seen_r {
        ("u", "r")
    } else if seen_u && seen_f {
        ("u", "f")
    } else if seen_u && seen_t {
        ("u", "t")
    } else if seen_b && seen_f {
        ("b", "f")
    } else if seen_b && seen_t {
        ("b", "t")
    } else if seen_f && seen_t {
        ("f", "t")
    } else {
        return None;
    };
    Some(format!(
        "'{}' and '{}' prefixes are incompatible",
        pair.0, pair.1
    ))
}

fn malformed_unicode_n_escape_error(source: &str) -> Option<CpythonDiagnostic> {
    let bytes = source.as_bytes();
    let mut index = 0;
    while index < bytes.len() {
        match bytes[index] {
            b'#' => {
                while index < bytes.len() && bytes[index] != b'\n' {
                    index += 1;
                }
            }
            b'\'' | b'"' => {
                let quote_index = index;
                let interpolated = interpolated_string_prefix(bytes, quote_index).is_some();
                let Some((content_start, content_end)) =
                    quoted_string_content_range(bytes, quote_index, bytes[quote_index])
                else {
                    index = skip_quoted_string(bytes, quote_index);
                    continue;
                };
                if let Some(error) = malformed_unicode_n_in_content(
                    bytes,
                    content_start,
                    content_end,
                    quote_index,
                    interpolated,
                ) {
                    return Some(error);
                }
                index = skip_string_token(bytes, quote_index);
            }
            _ => index += 1,
        }
    }
    None
}

fn malformed_unicode_n_in_content(
    bytes: &[u8],
    start: usize,
    end: usize,
    quote_index: usize,
    interpolated: bool,
) -> Option<CpythonDiagnostic> {
    let mut index = start;
    let mut brace_depth = 0usize;
    while index + 1 < end {
        if interpolated && brace_depth == 0 && bytes[index] == b'{' {
            if bytes.get(index + 1) == Some(&b'{') {
                index += 2;
                continue;
            }
            brace_depth = 1;
            index += 1;
            continue;
        }
        if interpolated && brace_depth > 0 {
            match bytes[index] {
                b'\'' | b'"' => index = skip_string_token(bytes, index).max(index + 1),
                b'{' => {
                    brace_depth += 1;
                    index += 1;
                }
                b'}' => {
                    brace_depth -= 1;
                    index += 1;
                }
                _ => index += 1,
            }
            continue;
        }
        if bytes[index] == b'\\' && bytes[index + 1] == b'N' {
            let escape_start = index;
            if bytes.get(index + 2) == Some(&b'{') {
                let mut look = index + 3;
                while look < end && bytes[look] != b'}' {
                    look += 1;
                }
                if look >= end {
                    return Some(unicode_n_diagnostic(escape_start, end, start, quote_index));
                }
                index = look + 1;
                continue;
            }
            return Some(unicode_n_diagnostic(
                escape_start,
                escape_start + 2,
                start,
                quote_index,
            ));
        }
        if interpolated && bytes[index] == b'}' && bytes.get(index + 1) == Some(&b'}') {
            index += 2;
            continue;
        }
        index += 1;
    }
    None
}

fn unicode_n_diagnostic(
    escape_start: usize,
    escape_end: usize,
    content_start: usize,
    quote_index: usize,
) -> CpythonDiagnostic {
    let start = escape_start - content_start;
    let end = escape_end.saturating_sub(content_start).saturating_sub(1);
    CpythonDiagnostic::new(
        format!(
            "(unicode error) 'unicodeescape' codec can't decode bytes in position {start}-{end}: malformed \\N character escape"
        ),
        quote_index,
        quote_index + 1,
    )
}

fn too_many_nested_interpolated_strings(source: &str) -> Option<CpythonDiagnostic> {
    too_many_nested_interpolated_strings_in(source.as_bytes(), 0, source.len(), 0)
}

fn too_many_nested_interpolated_strings_in(
    bytes: &[u8],
    mut index: usize,
    end: usize,
    depth: usize,
) -> Option<CpythonDiagnostic> {
    while index < end {
        match bytes[index] {
            b'#' => {
                while index < end && bytes[index] != b'\n' {
                    index += 1;
                }
            }
            b'\'' | b'"' => {
                if interpolated_string_prefix(bytes, index).is_some() {
                    if depth + 1 >= MAXFSTRINGLEVEL {
                        return Some(CpythonDiagnostic::new(
                            "too many nested f-strings or t-strings".to_owned(),
                            index,
                            index + 1,
                        ));
                    }
                    if let Some((content_start, content_end)) =
                        interpolated_string_content_range(bytes, index)
                        && let Some(error) = too_many_nested_interpolated_strings_in(
                            bytes,
                            content_start,
                            content_end,
                            depth + 1,
                        )
                    {
                        return Some(error);
                    }
                }
                index = skip_string_token(bytes, index).max(index + 1);
            }
            _ => index += 1,
        }
    }
    None
}

fn too_many_nested_parentheses_error(source: &str) -> Option<CpythonDiagnostic> {
    const MAXLEVEL: usize = 200;

    let bytes = source.as_bytes();
    let mut index = 0;
    let mut level = 0usize;
    while index < bytes.len() {
        match bytes[index] {
            b'#' => {
                while index < bytes.len() && bytes[index] != b'\n' {
                    index += 1;
                }
            }
            b'\'' | b'"' => {
                index = skip_quoted_string(bytes, index);
            }
            b'(' | b'[' | b'{' => {
                if level >= MAXLEVEL {
                    return Some(CpythonDiagnostic::new(
                        "too many nested parentheses".to_owned(),
                        index,
                        index + 1,
                    ));
                }
                level += 1;
                index += 1;
            }
            b')' | b']' | b'}' => {
                level = level.saturating_sub(1);
                index += 1;
            }
            _ => index += 1,
        }
    }
    None
}

fn invalid_unparenthesized_yield_after_comma_error(source: &str) -> Option<CpythonDiagnostic> {
    let bytes = source.as_bytes();
    let mut index = 0;
    while index < bytes.len() {
        match bytes[index] {
            b'#' => {
                while index < bytes.len() && bytes[index] != b'\n' {
                    index += 1;
                }
            }
            b'\'' | b'"' => {
                index = skip_quoted_string(bytes, index);
            }
            b',' => {
                let mut cursor = index + 1;
                while matches!(bytes.get(cursor), Some(b' ' | b'\t' | b'\x0c')) {
                    cursor += 1;
                }
                if starts_identifier(bytes, cursor, b"yield") {
                    return Some(CpythonDiagnostic::new(
                        "invalid syntax".to_owned(),
                        cursor,
                        cursor + 5,
                    ));
                }
                index += 1;
            }
            _ => index += 1,
        }
    }
    None
}

/// Horizontal and vertical space the tokenizer skips, not `str::trim()`.
const fn is_ascii_tokenizer_whitespace(c: char) -> bool {
    matches!(c, ' ' | '\t' | '\n' | '\r' | '\x0c')
}

/// True when every line is empty or a comment after tokenizer whitespace.
#[doc(hidden)]
#[must_use]
pub fn is_blank_python_source(source: &str) -> bool {
    source.lines().all(|line| {
        let trimmed = line.trim_matches(is_ascii_tokenizer_whitespace);
        trimmed.is_empty() || trimmed.starts_with('#')
    })
}

fn single_mode_blank_source_error(source_file: &SourceFile) -> Option<CompileError> {
    let source = source_file.source_text();
    if !is_blank_python_source(source) {
        return None;
    }
    let has_indent_only_line = source.lines().any(|line| {
        line.trim_matches(is_ascii_tokenizer_whitespace).is_empty()
            && line.chars().any(|c| matches!(c, ' ' | '\t'))
    });
    if has_indent_only_line {
        let (location, end_location) =
            source_locations(source_file, TextSize::new(0), TextSize::new(0));
        return Some(CompileError::Parse(ParseError {
            error: parser::ParseErrorType::UnexpectedIndentation,
            raw_location: ruff_text_size::TextRange::new(TextSize::new(0), TextSize::new(0)),
            location,
            end_location,
            source_path: source_file.name().to_owned(),
            is_unclosed_bracket: false,
            is_unclosed_string: false,
        }));
    }
    Some(CompileError::from_source_error(
        source_file,
        CpythonDiagnostic::new("invalid syntax".to_owned(), 0, 0),
    ))
}

/// The byte-order mark is only stripped while decoding source bytes, so one
/// that survives into the text is just a non-printable character. The tokenizer
/// rejects it everywhere except at the very start of the text, which is where
/// this covers.
#[doc(hidden)]
#[must_use]
pub fn leading_byte_order_mark_error(source_file: &SourceFile) -> Option<CompileError> {
    source_file.source_text().starts_with('\u{feff}').then(|| {
        CompileError::from_source_error(
            source_file,
            CpythonDiagnostic::new("invalid non-printable character U+FEFF".to_owned(), 0, 0),
        )
    })
}

/// Source-level errors raised before parsing, as CPython's tokenizer does.
///
/// Checking after the parse would build the tree first, and one nested that
/// deep exhausts the native stack when it is dropped.
pub fn pre_parse_source_error(source_file: &SourceFile) -> Result<(), CompileError> {
    match too_many_nested_parentheses_error(source_file.source_text()) {
        Some(error) => Err(CompileError::from_source_error(source_file, error)),
        None => Ok(()),
    }
}

fn post_parse_source_error(
    source_file: &SourceFile,
    tokens: &Tokens,
    opts: &CompileOpts,
) -> Option<CompileError> {
    if let Some(error) = leading_byte_order_mark_error(source_file) {
        return Some(error);
    }
    if let Some(error) = too_many_nested_interpolated_strings(source_file.source_text()) {
        return Some(CompileError::from_source_error(source_file, error));
    }
    if let Some(error) =
        long_decimal_integer_literal_error(source_file, tokens, opts.int_max_str_digits)
    {
        return Some(error);
    }
    invalid_call_argument_error(source_file.source_text())
        .or_else(|| invalid_match_mapping_rest_wildcard_error(source_file.source_text()))
        .or_else(|| invalid_match_as_target_error(source_file.source_text()))
        .or_else(|| invalid_unparenthesized_yield_after_comma_error(source_file.source_text()))
        .or_else(|| invalid_parenthesized_import_star_error(source_file.source_text()))
        .map(|error| CompileError::from_source_error(source_file, error))
}

fn is_compound_stmt(stmt: &ast::Stmt) -> bool {
    matches!(
        stmt,
        ast::Stmt::FunctionDef(_)
            | ast::Stmt::ClassDef(_)
            | ast::Stmt::If(_)
            | ast::Stmt::For(_)
            | ast::Stmt::While(_)
            | ast::Stmt::With(_)
            | ast::Stmt::Try(_)
            | ast::Stmt::Match(_)
    )
}

/// Rejects an AST nested deeper than `limit`.
///
/// The parser grows its own stack, so it returns trees deeper than the passes
/// after it can walk, and each of those recurses without a limit of its own.
/// The symbol table applies the same limit, but `compile(..., PyCF_ONLY_AST)`
/// returns before reaching it.
pub fn too_deeply_nested_error(
    ast: &ast::Mod,
    source_file: &SourceFile,
    limit: usize,
) -> Result<(), CompileError> {
    use ast::visitor::Visitor;

    struct DepthChecker {
        depth: usize,
        limit: usize,
        too_deep: Option<ruff_text_size::TextRange>,
    }

    impl DepthChecker {
        fn descend(&mut self, range: ruff_text_size::TextRange, walk: impl FnOnce(&mut Self)) {
            if self.too_deep.is_some() {
                return;
            }
            if self.depth >= self.limit {
                self.too_deep = Some(range);
                return;
            }
            self.depth += 1;
            walk(self);
            self.depth -= 1;
        }
    }

    impl<'a> Visitor<'a> for DepthChecker {
        fn visit_stmt(&mut self, stmt: &'a ast::Stmt) {
            self.descend(stmt.range(), |checker| {
                ast::visitor::walk_stmt(checker, stmt);
            });
        }

        fn visit_expr(&mut self, expr: &'a ast::Expr) {
            self.descend(expr.range(), |checker| {
                ast::visitor::walk_expr(checker, expr);
            });
        }

        fn visit_pattern(&mut self, pattern: &'a ast::Pattern) {
            self.descend(pattern.range(), |checker| {
                ast::visitor::walk_pattern(checker, pattern);
            });
        }
    }

    let mut checker = DepthChecker {
        depth: 0,
        limit,
        too_deep: None,
    };
    match ast {
        ast::Mod::Module(module) => checker.visit_body(&module.body),
        ast::Mod::Expression(expression) => checker.visit_expr(&expression.body),
    }

    let Some(range) = checker.too_deep else {
        return Ok(());
    };
    let (location, end_location) = source_locations(source_file, range.start(), range.end());
    Err(CompileError::Codegen(codegen::error::CodegenError {
        location: Some(location),
        end_location: Some(end_location),
        error: codegen::error::CodegenErrorType::RecursionError,
        source_path: source_file.name().to_owned(),
    }))
}

/// Syntax the reference grammar has no rule for, but that this parser accepts.
///
/// A bare generator expression is one: `f(x for x in y)` is the `primary
/// genexp` alternative of a call, and a class header only takes `arguments`,
/// which has no such alternative. A format spec nested more than two deep is
/// the other; the tokenizer runs out of nesting levels for it.
#[doc(hidden)]
#[must_use]
pub fn unsupported_grammar_error(ast: &ast::Mod, source_file: &SourceFile) -> Option<CompileError> {
    use ast::visitor::Visitor;

    /// The deepest chain of format specs one string literal may hold.
    const MAX_FORMAT_SPEC_DEPTH: usize = 2;

    struct Checker<'a> {
        source_file: &'a SourceFile,
        error: Option<CompileError>,
    }

    impl Checker<'_> {
        fn fail(&mut self, message: &str, range: ruff_text_size::TextRange) {
            self.error = Some(CompileError::from_source_error(
                self.source_file,
                CpythonDiagnostic::new(
                    message.to_owned(),
                    range.start().to_usize(),
                    range.end().to_usize(),
                ),
            ));
        }

        fn check_format_specs(
            &mut self,
            kind: &str,
            elements: &ast::InterpolatedStringElements,
            depth: usize,
        ) {
            for element in elements.interpolations() {
                let Some(format_spec) = &element.format_spec else {
                    continue;
                };
                if depth == MAX_FORMAT_SPEC_DEPTH {
                    self.fail(
                        &alloc::format!("{kind}: expressions nested too deeply"),
                        format_spec.range,
                    );
                    return;
                }
                self.check_format_specs(kind, &format_spec.elements, depth + 1);
                if self.error.is_some() {
                    return;
                }
            }
        }
    }

    impl<'a> Visitor<'a> for Checker<'_> {
        fn visit_stmt(&mut self, stmt: &'a ast::Stmt) {
            if self.error.is_some() {
                return;
            }
            if let ast::Stmt::ClassDef(class_def) = stmt
                && let Some(arguments) = &class_def.arguments
                && let [ast::Expr::Generator(generator)] = &arguments.args[..]
                && !generator.parenthesized
            {
                let range = generator
                    .generators
                    .first()
                    .map_or(generator.range, |comprehension| comprehension.range);
                self.fail("invalid syntax", range);
                return;
            }
            ast::visitor::walk_stmt(self, stmt);
        }

        fn visit_expr(&mut self, expr: &'a ast::Expr) {
            if self.error.is_some() {
                return;
            }
            // Each literal counts its own nesting: one written inside a format
            // spec starts over.
            match expr {
                ast::Expr::FString(fstring) => {
                    for part in &fstring.value {
                        if let ast::FStringPart::FString(part) = part {
                            self.check_format_specs("f-string", &part.elements, 0);
                        }
                    }
                }
                ast::Expr::TString(tstring) => {
                    for part in &tstring.value {
                        self.check_format_specs("t-string", &part.elements, 0);
                    }
                }
                _ => {}
            }
            if self.error.is_some() {
                return;
            }
            ast::visitor::walk_expr(self, expr);
        }
    }

    let mut checker = Checker {
        source_file,
        error: None,
    };
    match ast {
        ast::Mod::Module(module) => checker.visit_body(&module.body),
        ast::Mod::Expression(expression) => checker.visit_expr(&expression.body),
    }
    checker.error
}

fn single_mode_body_error(body: &[ast::Stmt], source_file: &SourceFile) -> Option<CompileError> {
    let first = body.first()?;
    let source_code = source_file.to_source_code();
    let first_start = source_code.source_location(first.range().start(), PositionEncoding::Utf8);
    let first_end = source_code.source_location(first.range().end(), PositionEncoding::Utf8);

    if body.iter().skip(1).any(|stmt| {
        source_code
            .source_location(stmt.range().start(), PositionEncoding::Utf8)
            .line
            > first_start.line
    }) {
        return Some(CompileError::from_source_error(
            source_file,
            CpythonDiagnostic::new(
                "multiple statements found while compiling a single statement".to_owned(),
                first.range().end().to_usize(),
                first.range().end().to_usize(),
            ),
        ));
    }

    if is_compound_stmt(first)
        && first_start.line == first_end.line
        && !ends_with_line_break(source_file.source_text())
    {
        return Some(CompileError::from_source_error(
            source_file,
            CpythonDiagnostic::new(
                "invalid syntax".to_owned(),
                first.range().start().to_usize(),
                first.range().start().to_usize(),
            ),
        ));
    }
    None
}

fn single_mode_source_error(ast: &ast::Mod, source_file: &SourceFile) -> Option<CompileError> {
    let ast::Mod::Module(module) = ast else {
        return None;
    };
    single_mode_body_error(&module.body, source_file)
}

fn ends_with_line_break(source: &str) -> bool {
    source.ends_with('\n') || source.ends_with('\r')
}

fn ends_with_implied_dedent(source: &str) -> bool {
    let mut lexer = parser::lexer::lex(source, parser::Mode::Module);
    let mut last_kind = TokenKind::EndOfFile;
    loop {
        let kind = lexer.next_token();
        if kind.is_eof() {
            break;
        }
        last_kind = kind;
    }
    matches!(last_kind, TokenKind::Dedent)
}

/// Detect input that only parses because Ruff's lexer closes indentation at EOF.
///
/// `PyCF_DONT_IMPLY_DEDENT` is used by `codeop` and interactive compile
/// paths to keep an indented block incomplete until a terminating newline is seen.
#[must_use]
pub fn dont_imply_dedent_source_error(source_file: &SourceFile) -> Option<CompileError> {
    let source = source_file.source_text();
    if ends_with_line_break(source) || !ends_with_implied_dedent(source) {
        return None;
    }
    let eof = source.len();
    Some(CompileError::from_source_error(
        source_file,
        CpythonDiagnostic::new("incomplete input".to_owned(), eof, eof),
    ))
}

/// Find the last unclosed opening bracket in source code.
/// Returns the bracket character and its byte offset, or None if all brackets are balanced.
fn find_unclosed_bracket(source: &str) -> Option<(char, usize)> {
    let mut stack: Vec<(char, usize)> = Vec::new();
    let mut in_string = false;
    let mut string_quote = '\0';
    let mut triple_quote = false;
    let mut escape_next = false;
    let mut is_raw_string = false;

    let chars: Vec<(usize, char)> = source.char_indices().collect();
    let mut i = 0;

    while i < chars.len() {
        let (byte_offset, ch) = chars[i];

        if escape_next {
            escape_next = false;
            i += 1;
            continue;
        }

        if in_string {
            if ch == '\\' && !is_raw_string {
                escape_next = true;
            } else if triple_quote {
                if ch == string_quote
                    && i + 2 < chars.len()
                    && chars[i + 1].1 == string_quote
                    && chars[i + 2].1 == string_quote
                {
                    in_string = false;
                    i += 3;
                    continue;
                }
            } else if ch == string_quote {
                in_string = false;
            }
            i += 1;
            continue;
        }

        // Check for comments
        if ch == '#' {
            // Skip to end of line
            while i < chars.len() && chars[i].1 != '\n' {
                i += 1;
            }
            continue;
        }

        // Check for string start (with optional prefix like r, b, f, u, rb, br, etc.)
        if ch == '\'' || ch == '"' {
            // Check up to 2 characters before the quote for string prefix
            is_raw_string = false;
            for look_back in 1..=2.min(i) {
                let prev = chars[i - look_back].1;
                if matches!(prev, 'r' | 'R') {
                    is_raw_string = true;
                    break;
                }
                if !matches!(prev, 'b' | 'B' | 'f' | 'F' | 'u' | 'U') {
                    break;
                }
            }
            string_quote = ch;
            if i + 2 < chars.len() && chars[i + 1].1 == ch && chars[i + 2].1 == ch {
                triple_quote = true;
                in_string = true;
                i += 3;
                continue;
            }
            triple_quote = false;
            in_string = true;
            i += 1;
            continue;
        }

        match ch {
            '(' | '[' | '{' => stack.push((ch, byte_offset)),
            ')' | ']' | '}' => {
                let expected = match ch {
                    ')' => '(',
                    ']' => '[',
                    '}' => '{',
                    _ => unreachable!(),
                };
                if stack.last().is_some_and(|&(open, _)| open == expected) {
                    stack.pop();
                }
            }
            _ => {}
        }

        i += 1;
    }

    stack.last().copied()
}

/// Compile a given source code into a bytecode object.
pub fn compile(
    source: &str,
    mode: Mode,
    source_path: &str,
    opts: CompileOpts,
) -> Result<CodeObject, CompileError> {
    // TODO: do this less hacky; ruff's parser should translate a CRLF line
    //       break in a multiline string into just an LF in the parsed value
    #[cfg(windows)]
    let source = source.replace("\r\n", "\n");
    #[cfg(windows)]
    let source = source.as_str();

    let source_file = SourceFileBuilder::new(source_path, source).finish();
    _compile(source_file, mode, opts)
    // let index = LineIndex::from_source_text(source);
    // let source_code = SourceCode::new(source, &index);
    // let mut locator = LinearLocator::new(source);
    // let mut ast = match parser::parse(source, mode.into(), &source_path) {
    //     Ok(x) => x,
    //     Err(e) => return Err(locator.locate_error(e)),
    // };

    // TODO:
    // if opts.optimize > 0 {
    //     ast = ConstantOptimizer::new()
    //         .fold_mod(ast)
    //         .unwrap_or_else(|e| match e {});
    // }
    // let ast = locator.fold_mod(ast).unwrap_or_else(|e| match e {});
}

fn _compile(
    source_file: SourceFile,
    mode: Mode,
    opts: CompileOpts,
) -> Result<CodeObject, CompileError> {
    _compile_with_syntax_warning_handler(source_file, mode, opts, None)
}

fn _compile_with_syntax_warning_handler<'a>(
    source_file: SourceFile,
    mode: Mode,
    opts: CompileOpts,
    syntax_warning_handler: Option<&'a mut compile::SyntaxWarningHandler<'a>>,
) -> Result<CodeObject, CompileError> {
    let parser_mode = match mode {
        Mode::Exec => parser::Mode::Module,
        Mode::Eval => parser::Mode::Expression,
        // ruff does not have an interactive mode, which is fine,
        // since these are only different in terms of compilation
        Mode::Single | Mode::BlockExpr => parser::Mode::Module,
    };
    let parser_options = parser::ParseOptions::from(parser_mode);
    let barry_source = prepare_barry_as_flufl_source(
        source_file.source_text(),
        parser_options.clone(),
        opts.future_features
            .contains(core::bytecode::CodeFlags::FUTURE_BARRY_AS_BDFL),
    );
    pre_parse_source_error(&source_file)?;
    let parsed = parser::parse(barry_source.source(), parser_options);
    if let Some(error) = barry_source.diagnostic(parsed.as_ref().err(), &source_file) {
        return Err(error);
    }
    let parsed =
        parsed.map_err(|err| CompileError::from_ruff_parse_error(err, &source_file, mode))?;
    if matches!(mode, Mode::Single)
        && let Some(error) = single_mode_blank_source_error(&source_file)
    {
        return Err(error);
    }
    if opts.dont_imply_dedent
        && matches!(mode, Mode::Single)
        && let Some(error) = dont_imply_dedent_source_error(&source_file)
    {
        return Err(error);
    }
    if let Some(error) = post_parse_source_error(&source_file, parsed.tokens(), &opts) {
        return Err(error);
    }
    let ast = parsed.into_syntax();
    too_deeply_nested_error(&ast, &source_file, opts.recursion_limit)?;
    if let Some(error) = unsupported_grammar_error(&ast, &source_file) {
        return Err(error);
    }
    let single_mode_error = matches!(mode, Mode::Single)
        .then(|| single_mode_source_error(&ast, &source_file))
        .flatten();
    let code = compile::compile_top_with_syntax_warning_handler(
        ast,
        source_file,
        mode,
        opts,
        syntax_warning_handler,
    )
    .map_err(CompileError::from)?;
    if let Some(error) = single_mode_error {
        return Err(error);
    }
    Ok(code)
}

#[doc(hidden)]
pub struct BarrySource<'a> {
    source: Cow<'a, str>,
    not_equal: Option<ruff_text_size::TextRange>,
    legacy_not_equal: Vec<ruff_text_size::TextRange>,
}

impl BarrySource<'_> {
    #[must_use]
    pub fn source(&self) -> &str {
        &self.source
    }

    #[must_use]
    pub fn not_equal_before(
        &self,
        parse_error: Option<&parser::ParseError>,
    ) -> Option<ruff_text_size::TextRange> {
        self.not_equal.filter(|range| {
            parse_error.is_none_or(|error| {
                let diagnostic_start = if matches!(
                    &error.error,
                    parser::ParseErrorType::Lexical(parser::LexicalErrorType::Eof)
                ) {
                    find_unclosed_bracket(&self.source).map_or_else(
                        || error.location.start(),
                        |(_, offset)| TextSize::new(offset as u32),
                    )
                } else {
                    error.location.start()
                };
                range.start() <= diagnostic_start
            })
        })
    }

    /// The obsolete `<>` operator the parse error points at, if any. In Barry
    /// mode the operator was rewritten to `!=`, so the error lands on its
    /// start; outside Barry mode ruff lexes `<` and then an unexpected `>`, so
    /// the error lands one character in. Either way the whole operator is one
    /// token to the tokenizer, so report it as one.
    #[must_use]
    pub fn invalid_legacy_operator(
        &self,
        parse_error: &parser::ParseError,
    ) -> Option<ruff_text_size::TextRange> {
        let location = parse_error.location.start();
        self.legacy_not_equal
            .iter()
            .copied()
            .find(|range| range.contains(location) || range.start() == location)
            .filter(|range| self.outranks_unclosed_bracket(*range))
    }

    /// Whether `range` outranks an unclosed bracket. The bracket is reported
    /// at itself, so it wins over anything that starts after it.
    fn outranks_unclosed_bracket(&self, range: ruff_text_size::TextRange) -> bool {
        find_unclosed_bracket(&self.source)
            .is_none_or(|(_, offset)| range.start() <= TextSize::new(offset as u32))
    }

    /// The diagnostic for this source, if any: an obsolete `<>` the parse
    /// error points at, or -- in Barry mode only -- the first `!=`. A `<>`
    /// takes precedence over a `!=` reported later in the source.
    #[must_use]
    pub fn diagnostic(
        &self,
        parse_error: Option<&parser::ParseError>,
        source_file: &SourceFile,
    ) -> Option<CompileError> {
        if let Some(range) = parse_error.and_then(|error| self.invalid_legacy_operator(error)) {
            return Some(barry_as_flufl_invalid_legacy_operator_error(
                source_file,
                range,
            ));
        }
        self.not_equal_before(parse_error)
            .map(|range| barry_as_flufl_not_equal_error(source_file, range))
    }
}

#[doc(hidden)]
#[must_use]
pub fn barry_as_flufl_not_equal_error(
    source_file: &SourceFile,
    range: ruff_text_size::TextRange,
) -> CompileError {
    CompileError::from_source_error(
        source_file,
        CpythonDiagnostic::new(
            "with Barry as BDFL, use '<>' instead of '!='".to_owned(),
            range.start().to_usize(),
            range.end().to_usize(),
        ),
    )
}

#[doc(hidden)]
#[must_use]
pub fn barry_as_flufl_invalid_legacy_operator_error(
    source_file: &SourceFile,
    range: ruff_text_size::TextRange,
) -> CompileError {
    CompileError::from_source_error(
        source_file,
        CpythonDiagnostic::new(
            "invalid syntax".to_owned(),
            range.start().to_usize(),
            range.end().to_usize(),
        ),
    )
}

/// Every `<>` in `source`, located by plain text search. Only used where the
/// operator is not rewritten, so an occurrence inside a string or a comment
/// costs nothing: it can never coincide with the location of a parse error.
fn textual_legacy_not_equal(source: &str) -> Vec<ruff_text_size::TextRange> {
    source
        .match_indices("<>")
        .map(|(offset, matched)| {
            ruff_text_size::TextRange::at(
                TextSize::new(offset as u32),
                TextSize::new(matched.len() as u32),
            )
        })
        .collect()
}

#[doc(hidden)]
pub fn prepare_barry_as_flufl_source(
    source: &str,
    parser_options: parser::ParseOptions,
    inherited: bool,
) -> BarrySource<'_> {
    let scanned = (inherited || source.contains("barry_as_FLUFL"))
        .then(|| parser::parse_unchecked(source, parser_options));
    let enabled = scanned.as_ref().is_some_and(|scanned| {
        inherited
            || codegen::preprocess::future_features(scanned.syntax())
                .contains(core::bytecode::CodeFlags::FUTURE_BARRY_AS_BDFL)
    });
    let Some(scanned) = scanned.filter(|_| enabled) else {
        return BarrySource {
            source: Cow::Borrowed(source),
            not_equal: None,
            legacy_not_equal: textual_legacy_not_equal(source),
        };
    };

    let not_equal = scanned
        .tokens()
        .iter()
        .find(|token| token.kind() == TokenKind::NotEqual)
        .map(Ranged::range);
    let replacements = scanned
        .tokens()
        .windows(2)
        .filter_map(|tokens| {
            let [less, greater] = tokens else {
                return None;
            };
            (less.kind() == TokenKind::Less
                && greater.kind() == TokenKind::Greater
                && less.end() == greater.start())
            .then(|| ruff_text_size::TextRange::new(less.start(), greater.end()))
        })
        .collect::<Vec<_>>();

    let source = if replacements.is_empty() {
        Cow::Borrowed(source)
    } else {
        let mut rewritten = source.to_owned();
        for range in replacements.iter().rev() {
            rewritten.replace_range(range.start().to_usize()..range.end().to_usize(), "!=");
        }
        Cow::Owned(rewritten)
    };
    BarrySource {
        source,
        not_equal,
        legacy_not_equal: replacements,
    }
}

pub fn compile_with_syntax_warning_handler<'a>(
    source: &str,
    mode: Mode,
    source_path: &str,
    opts: CompileOpts,
    syntax_warning_handler: &'a mut compile::SyntaxWarningHandler<'a>,
) -> Result<CodeObject, CompileError> {
    let source = source.replace("\r\n", "\n");
    #[cfg(windows)]
    let source = source.as_str();

    let source_file = SourceFileBuilder::new(source_path, source).finish();
    _compile_with_syntax_warning_handler(source_file, mode, opts, Some(syntax_warning_handler))
}

pub fn compile_symtable(
    source: &str,
    mode: Mode,
    source_path: &str,
) -> Result<symboltable::SymbolTable, CompileError> {
    let source_file = SourceFileBuilder::new(source_path, source).finish();
    _compile_symtable(source_file, mode)
}

/// Bring a module into the shape the symbol table is built from.
fn symtable_preprocess_module(
    module: &mut ast::ModModule,
    source_file: &SourceFile,
) -> Result<(), CompileError> {
    let future_features = codegen::preprocess::checked_future_features_in_body(&module.body)
        .map_err(|error| future_feature_error(error, source_file))?;
    let future_annotations =
        future_features.contains(core::bytecode::CodeFlags::FUTURE_ANNOTATIONS);
    // Constant folding is left out; the symbol table is built from the parsed
    // program as written.
    codegen::preprocess::preprocess_statements(&mut module.body, 0, future_annotations, true);
    Ok(())
}

fn future_feature_error(
    error: codegen::preprocess::FutureFeatureError,
    source_file: &SourceFile,
) -> CompileError {
    let source_code = source_file.to_source_code();
    let location = source_code.source_location(error.range.start(), PositionEncoding::Utf8);
    let end_location = source_code.source_location(error.range.end(), PositionEncoding::Utf8);
    let error = match error.kind {
        codegen::preprocess::FutureFeatureErrorKind::InvalidFeature(feature) => {
            codegen::error::CodegenErrorType::InvalidFutureFeature(feature)
        }
        codegen::preprocess::FutureFeatureErrorKind::InvalidBraces => {
            codegen::error::CodegenErrorType::InvalidFutureBraces
        }
    };
    codegen::error::CodegenError {
        location: Some(location),
        end_location: Some(end_location),
        error,
        source_path: source_file.name().to_owned(),
    }
    .into()
}

pub fn _compile_symtable(
    source_file: SourceFile,
    mode: Mode,
) -> Result<symboltable::SymbolTable, CompileError> {
    let parser_mode = match mode {
        Mode::Exec | Mode::Single | Mode::BlockExpr => parser::Mode::Module,
        Mode::Eval => parser::Mode::Expression,
    };
    let parser_options = parser::ParseOptions::from(parser_mode);
    let barry_source =
        prepare_barry_as_flufl_source(source_file.source_text(), parser_options.clone(), false);
    let res = match mode {
        Mode::Exec | Mode::Single | Mode::BlockExpr => {
            pre_parse_source_error(&source_file)?;
            let parsed = ruff_python_parser::parse(barry_source.source(), parser_options);
            if let Some(error) = barry_source.diagnostic(parsed.as_ref().err(), &source_file) {
                return Err(error);
            }
            let ast =
                parsed.map_err(|e| CompileError::from_ruff_parse_error(e, &source_file, mode))?;
            if let Some(error) =
                post_parse_source_error(&source_file, ast.tokens(), &CompileOpts::default())
            {
                return Err(error);
            }
            let ast = ast.into_syntax();
            too_deeply_nested_error(&ast, &source_file, CompileOpts::default().recursion_limit)?;
            if let Some(error) = unsupported_grammar_error(&ast, &source_file) {
                return Err(error);
            }
            let ast = ast.expect_module();
            if matches!(mode, Mode::Single)
                && let Some(error) = single_mode_body_error(&ast.body, &source_file)
            {
                return Err(error);
            }
            let mut ast = ast;
            symtable_preprocess_module(&mut ast, &source_file)?;
            symboltable::SymbolTable::scan_program(&ast, source_file.clone())
        }
        Mode::Eval => {
            pre_parse_source_error(&source_file)?;
            let parsed = ruff_python_parser::parse(barry_source.source(), parser_options);
            if let Some(error) = barry_source.diagnostic(parsed.as_ref().err(), &source_file) {
                return Err(error);
            }
            let ast =
                parsed.map_err(|e| CompileError::from_ruff_parse_error(e, &source_file, mode))?;
            if let Some(error) =
                post_parse_source_error(&source_file, ast.tokens(), &CompileOpts::default())
            {
                return Err(error);
            }
            let ast = ast.into_syntax();
            too_deeply_nested_error(&ast, &source_file, CompileOpts::default().recursion_limit)?;
            if let Some(error) = unsupported_grammar_error(&ast, &source_file) {
                return Err(error);
            }
            let mut ast = ast;
            codegen::preprocess::preprocess_mod(&mut ast, 0, false, true);
            symboltable::SymbolTable::scan_expr(&ast.expect_expression(), source_file.clone())
        }
    };
    res.map_err(|e| e.into_codegen_error(source_file.name().to_owned()).into())
}

#[cfg(test)]
mod tests {
    use super::*;

    #[test]
    fn basic_compile() {
        let code = "x = 'abc'";
        let compiled = compile(code, Mode::Single, "<>", CompileOpts::default());
        dbg!(compiled.expect("compile error"));
    }

    #[test]
    fn empty_parameter_default_is_reported_before_later_params() {
        let err = compile(
            "def foo(a=1,d=,c):\n    pass\n",
            Mode::Exec,
            "<params>",
            CompileOpts::default(),
        )
        .expect_err("empty default should fail");
        assert!(
            err.to_string()
                .contains("expected default value expression"),
            "got {err}"
        );
    }

    #[test]
    fn too_many_nested_fstrings_match_tokenizer_limit() {
        fn nested(n: usize) -> String {
            if n == 0 {
                return "1+1".to_owned();
            }
            format!("f\"{{{}}}\"", nested(n - 1))
        }

        compile(&nested(149), Mode::Eval, "<nest>", CompileOpts::default())
            .expect("149 nested f-strings should compile");
        let err = compile(&nested(150), Mode::Eval, "<nest>", CompileOpts::default())
            .expect_err("150 nested f-strings should fail");
        assert!(
            err.to_string()
                .contains("too many nested f-strings or t-strings"),
            "got {err}"
        );
    }

    #[test]
    fn interpolated_string_diagnostics_match_cpython() {
        for (source, expected) in [
            ("f'{'", "f-string: expecting '}'"),
            ("t'{'", "t-string: expecting '}'"),
            (
                "f'{1=}{;'",
                "f-string: expecting a valid expression after '{'",
            ),
            (
                "t'{x;y}'",
                "t-string: expecting '=', or '!', or ':', or '}'",
            ),
            ("t'{x!s:'", "t-string: expecting '}', or format specs"),
            ("t'{x=!}'", "t-string: missing conversion character"),
            (
                "t'{x:{;}}'",
                "t-string: expecting a valid expression after '{'",
            ),
            ("f'{1#}'", "'{' was never closed"),
            ("t'{", "'{' was never closed"),
            // A field is only "never closed" while the tokenizer is reading its expression.
            // Past the field's own `:` it emits literal text again, so the quote is what is
            // missing — and a `:` in a slice, a display or a lambda is not the field's own.
            ("f'{a", "'{' was never closed"),
            ("f'{a!r", "'{' was never closed"),
            ("f'{a=", "'{' was never closed"),
            ("f'{ {1:2}", "'{' was never closed"),
            ("f'{d[1:2]", "'{' was never closed"),
            ("f'{(lambda x: x)", "'{' was never closed"),
            ("f'{a:{b:{c", "'{' was never closed"),
            ("f'{a:", "unterminated f-string literal"),
            ("f'{a:>5", "unterminated f-string literal"),
            ("f'{a!r:", "unterminated f-string literal"),
            ("f'{a}{b:", "unterminated f-string literal"),
            ("f'{a:{b}c", "unterminated f-string literal"),
            ("t'{a:>5", "unterminated t-string literal"),
            ("f'''{a:>5", "unterminated triple-quoted f-string literal"),
            // CPython names the innermost unclosed delimiter, not the field around it.
            ("f'{a[", "'[' was never closed"),
            ("f'{(a", "'(' was never closed"),
            ("f'{)#}'", "f-string: unmatched ')'"),
            (
                "f'{a[4)}'",
                "closing parenthesis ')' does not match opening parenthesis '['",
            ),
            ("t'", "unterminated t-string literal (detected at line 1)"),
            (
                "t'''",
                "unterminated triple-quoted t-string literal (detected at line 1)",
            ),
            (
                "t\"x\" b\"y\"",
                "Cannot mix t-string literals with string or bytes literals",
            ),
            (
                "b\"x\" t\"y\"",
                "Cannot mix t-string literals with string or bytes literals",
            ),
            // The literals ahead of the first t-string already mix, so CPython's first pass
            // raises from `_PyPegen_concatenate_strings` before the t-string rule is reached.
            (
                "\"a\" b\"b\" t\"c\"",
                "cannot mix bytes and nonbytes literals",
            ),
            // A dangling operator ends the expression before a separator could follow.
            (
                "f'{a==}'",
                "f-string: expecting '=', or '!', or ':', or '}'",
            ),
            (
                "f'{a and}'",
                "f-string: expecting '=', or '!', or ':', or '}'",
            ),
            (
                "f'{a is not}'",
                "f-string: expecting '=', or '!', or ':', or '}'",
            ),
            (
                "f'{a.b.}'",
                "f-string: expecting '=', or '!', or ':', or '}'",
            ),
            ("t'{a~}'", "t-string: expecting '=', or '!', or ':', or '}'"),
            // A doubled `=` or `!` opens no debug specifier and no conversion, so the field has
            // no expression at all rather than an empty one before a marker.
            (
                "f'{==a}'",
                "f-string: expecting a valid expression after '{'",
            ),
            (
                "f'{!=a}'",
                "f-string: expecting a valid expression after '{'",
            ),
            (
                "t'{==a}'",
                "t-string: expecting a valid expression after '{'",
            ),
            ("f'{=a}'", "f-string: valid expression required before '='"),
            ("f'{!}'", "f-string: valid expression required before '!'"),
            (
                "f'{lambda x:x}'",
                "f-string: lambda expressions are not allowed without parentheses",
            ),
            (
                "f'{1, lambda:x}'",
                "f-string: lambda expressions are not allowed without parentheses",
            ),
            (
                "f'{+ lambda:None}'",
                "f-string: expecting a valid expression after '{'",
            ),
            ("fu''", "'u' and 'f' prefixes are incompatible"),
            ("fb''", "'b' and 'f' prefixes are incompatible"),
            ("ufr''", "'u' and 'r' prefixes are incompatible"),
            (
                "(]\nbu'x'",
                "closing parenthesis ']' does not match opening parenthesis '('",
            ),
            ("0x\nbu'x'", "invalid hexadecimal literal"),
            ("(0x", "invalid hexadecimal literal"),
            ("print x; 0x", "invalid hexadecimal literal"),
            ("exec x; 0x", "invalid hexadecimal literal"),
            (
                "print x; 0x1",
                "Missing parentheses in call to 'print'. Did you mean print(...)?",
            ),
            (
                "print x; (",
                "Missing parentheses in call to 'print'. Did you mean print(...)?",
            ),
            ("print x; )", "unmatched ')'"),
            (
                "( '\\N'",
                "(unicode error) 'unicodeescape' codec can't decode bytes in position 0-1: malformed \\N character escape",
            ),
            ("(print x", "'(' was never closed"),
            (
                "(print x; '\\N'",
                "Missing parentheses in call to 'print'. Did you mean print(...)?",
            ),
            ("f'{x'; '", "f-string: expecting '}'"),
            ("f'{x'; 0x", "f-string: expecting '}'"),
            (
                "print x; f'{x'; 0x",
                "Missing parentheses in call to 'print'. Did you mean print(...)?",
            ),
            (
                concat!("f'{1:", "d\n}'"),
                "f-string: newlines are not allowed in format specifiers",
            ),
            ("f'{\n}'", "f-string: valid expression required before '}'"),
            (
                "f'''\n{\n# only a comment\n}'''",
                "f-string: valid expression required before '}'",
            ),
            ("{\\'a\\'}", "unexpected character after line continuation"),
            ("\"\\\n\"(1 for c in I,\\\n\\", "'(' was never closed"),
            (
                r"'\N'",
                "(unicode error) 'unicodeescape' codec can't decode bytes in position 0-1: malformed \\N character escape",
            ),
            (
                r"f'\N{'",
                "(unicode error) 'unicodeescape' codec can't decode bytes in position 0-2: malformed \\N character escape",
            ),
        ] {
            let err = compile(source, Mode::Eval, "<interp>", CompileOpts::default())
                .expect_err("should not compile");
            assert!(
                err.to_string().contains(expected),
                "{source:?}: expected {expected:?}, got {err}"
            );
        }
    }

    #[test]
    fn missing_indent_outranks_print_missing_parentheses() {
        for (source, expected) in [
            (
                "if True:\nprint \"No indent\"",
                "expected an indented block after 'if' statement on line 1",
            ),
            (
                "print \"old style\"",
                "Missing parentheses in call to 'print'. Did you mean print(...)?",
            ),
        ] {
            let err = compile(source, Mode::Exec, "<fragment>", CompileOpts::default())
                .expect_err("should not compile");
            assert!(
                err.to_string().contains(expected),
                "{source:?}: expected {expected:?}, got {err}"
            );
        }
    }

    #[test]
    fn unclosed_fstring_field_keeps_the_unclosed_bracket_flag() {
        let err = compile("f'{", Mode::Eval, "<interp>", CompileOpts::default())
            .expect_err("should not compile");
        let crate::CompileError::Parse(parse) = err else {
            panic!("expected a parse error, got {err}");
        };
        assert!(
            parse.is_unclosed_bracket,
            "unclosed f-string field must stay incomplete, got {parse}"
        );
        assert!(
            parse.to_string().contains("'{' was never closed"),
            "got {parse}"
        );
    }

    #[test]
    fn interpolated_literals_do_not_take_the_escaped_quote_hint() {
        // Parser/lexer/lexer.c offers "perhaps you escaped the end quote?" from its plain-string
        // branch only; the interpolated branch has just the triple-quoted and plain forms. The
        // assertion has to be exact: the hint is a suffix, so `contains` would pass either way.
        for (source, expected) in [
            (
                r"f'\'",
                "unterminated f-string literal (detected at line 1)",
            ),
            (
                r"t'\'",
                "unterminated t-string literal (detected at line 1)",
            ),
        ] {
            let err = compile(source, Mode::Eval, "<escaped>", CompileOpts::default())
                .expect_err("should not compile");
            assert_eq!(err.to_string(), expected, "{source:?}");
        }
        // A literal without a prefix still gets it.
        let err = compile(r"'\'", Mode::Eval, "<escaped>", CompileOpts::default())
            .expect_err("should not compile");
        assert_eq!(
            err.to_string(),
            "unterminated string literal (detected at line 1); perhaps you escaped the end quote?"
        );
    }

    #[test]
    fn a_field_bracket_mismatch_names_the_opening_line() {
        // CPython compares `parenlinenostack[level]` against the current `lineno`, so it names
        // the opening line only when the two brackets are not on the same one.
        for (source, expected) in [
            (
                "x = f\"\"\"{a[\n4)}\"\"\"\n",
                "closing parenthesis ')' does not match opening parenthesis '[' on line 1",
            ),
            (
                "x = f\"\"\"{a(\n\n4]}\"\"\"\n",
                "closing parenthesis ']' does not match opening parenthesis '(' on line 1",
            ),
            (
                "x = f\"{a[4)}\"\n",
                "closing parenthesis ')' does not match opening parenthesis '['",
            ),
        ] {
            let err = compile(source, Mode::Exec, "<paren>", CompileOpts::default())
                .expect_err("should not compile");
            assert_eq!(err.to_string(), expected, "{source:?}");
        }
    }

    #[test]
    fn a_dangling_operator_needs_the_field_to_close() {
        // A stray character is a finished token, so CPython's grammar rejects it on lookahead
        // and names the separators even when the field never closes. A dangling operator makes
        // the parser ask for one more token, and producing it hits the closing quote, where
        // lexer.c answers "expecting '}'" from the tokenizer instead.
        for (source, expected) in [
            ("f'{a;'", "f-string: expecting '=', or '!', or ':', or '}'"),
            ("f'{a$'", "f-string: expecting '=', or '!', or ':', or '}'"),
            ("f'{a?'", "f-string: expecting '=', or '!', or ':', or '}'"),
            ("f'{a and'", "f-string: expecting '}'"),
            ("f'{a+'", "f-string: expecting '}'"),
            ("f'{a=='", "f-string: expecting '}'"),
            ("f'{a.b.'", "f-string: expecting '}'"),
            ("f'{a is not'", "f-string: expecting '}'"),
            ("t'{a and'", "t-string: expecting '}'"),
            // With the field closed, the operator is what gets named.
            (
                "f'{a and}'",
                "f-string: expecting '=', or '!', or ':', or '}'",
            ),
            (
                "f'{a==}'",
                "f-string: expecting '=', or '!', or ':', or '}'",
            ),
        ] {
            let err = compile(source, Mode::Eval, "<dangling>", CompileOpts::default())
                .expect_err("should not compile");
            assert!(
                err.to_string().contains(expected),
                "{source:?}: expected {expected:?}, got {err}"
            );
        }
    }

    #[test]
    fn deeply_nested_format_specs_stay_linear() {
        // Each enclosing spec used to re-scan every field nested inside it, which made this
        // O(2^depth): a 110-byte source took six seconds. The scan now resumes past a nested
        // field once it has checked it, so this has to finish immediately.
        for depth in [32usize, 200] {
            let source = format!("x = f\"{}{}\"\n$", "{1:".repeat(depth), "}".repeat(depth));
            compile(&source, Mode::Exec, "<nested>", CompileOpts::default())
                .expect_err("the trailing `$` is a syntax error");
        }
    }

    #[test]
    #[expect(
        clippy::literal_string_with_formatting_args,
        reason = "these are Python format specs, not Rust format args"
    )]
    fn valid_interpolated_literals_do_not_shadow_a_later_syntax_error() {
        // A format spec is text rather than code, and a `#` in a replacement field starts a
        // comment. Mistaking either for expression syntax would blame a perfectly good literal
        // for the `$` further down, so the reported line must stay on the `$`.
        for source in [
            "x = f\"{1:(}\"\n$",
            "x = f\"{1:[}\"\n$",
            "x = f\"{1:#x}\"\n$",
            "x = f\"{1!r:#>5}\"\n$",
            "x = t\"{1:(}\"\n$",
            "x = f\"\"\"{1 # (\n}\"\"\"\n$",
            "x = f\"\"\"{1 # {\n}\"\"\"\n$",
            "x = f\"\"\"{1 # ]\n}\"\"\"\n$",
            "x = f\"\"\"{1 # a\\ b\n}\"\"\"\n$",
            // A comment tail is not the end of the expression, whatever it looks like.
            "x = f\"\"\"{a # +\n}\"\"\"\n$",
            "x = f\"\"\"{a # and\n}\"\"\"\n$",
            "x = f\"\"\"{a # ==\n}\"\"\"\n$",
            "x = t\"\"\"{a # .\n}\"\"\"\n$",
            "x = f\"\"\"{1:{a # +\n}}\"\"\"\n$",
            "x = f\"\"\"{a # +\n + b}\"\"\"\n$",
            // A signed leading-dot float is an operand.
            "x = f\"{-.5}\"\n$",
            "x = f\"{+.5}\"\n$",
            // Ellipsis and a leading-dot float start real expressions.
            "x = f\"{...}\"\n$",
            "x = f\"{.5}\"\n$",
            "x = t\"{...}\"\n$",
            // A dangling operator is reported, but a complete one is not.
            "x = f\"{a+b}\"\n$",
            "x = f\"{a.b.c}\"\n$",
            "x = f\"{1.}\"\n$",
            // A comparison operator is not a debug or conversion marker.
            "x = f\"{a==b}\"\n$",
            "x = f\"{a!=b}\"\n$",
            "x = f\"{a<=b}\"\n$",
            "x = f\"{a>=b}\"\n$",
            "x = t\"{a==b}\"\n$",
            "x = f\"{a:{b==c}}\"\n$",
        ] {
            let err = compile(source, Mode::Exec, "<interp>", CompileOpts::default())
                .expect_err("the trailing `$` is a syntax error");
            assert_eq!(
                err.python_location().0,
                source.lines().count(),
                "{source:?} reported the wrong line: {err}"
            );
        }
    }

    #[test]
    fn dont_imply_dedent_requires_terminating_newline() {
        let code = "if True:\n    pass";

        let opts = CompileOpts {
            dont_imply_dedent: true,
            ..CompileOpts::default()
        };
        let err = compile(code, Mode::Single, "<>", opts.clone()).expect_err("compile succeeded");
        assert_eq!(err.to_string(), "incomplete input");

        compile("if True:\n    pass\n", Mode::Single, "<>", opts).expect("compile error");
        compile(code, Mode::Single, "<>", CompileOpts::default()).expect("compile error");
    }

    #[test]
    fn barry_as_flufl_rewrites_legacy_not_equal_after_future_import() {
        let code = compile(
            "from __future__ import barry_as_FLUFL\nresult = 2 <> 3\n",
            Mode::Exec,
            "<barry>",
            CompileOpts::default(),
        )
        .expect("Barry comparison should compile");
        assert!(
            code.flags
                .contains(core::bytecode::CodeFlags::FUTURE_BARRY_AS_BDFL)
        );
    }

    #[test]
    fn inherited_barry_as_flufl_rewrites_legacy_not_equal() {
        let opts = CompileOpts {
            future_features: core::bytecode::CodeFlags::FUTURE_BARRY_AS_BDFL,
            ..CompileOpts::default()
        };
        compile("2 <> 3", Mode::Single, "<barry>", opts)
            .expect("inherited Barry comparison should compile");
    }

    #[test]
    fn barry_as_flufl_rejects_modern_not_equal() {
        let err = compile(
            "from __future__ import barry_as_FLUFL\n2 != 3\n",
            Mode::Exec,
            "<barry>",
            CompileOpts::default(),
        )
        .expect_err("Barry mode should reject !=");
        assert_eq!(
            err.to_string(),
            "with Barry as BDFL, use '<>' instead of '!='"
        );
        assert_eq!(err.python_location(), (2, 3));
    }

    #[test]
    fn fstring_adjacent_atoms_are_a_missing_comma() {
        let err = compile("f'{6 0}'", Mode::Exec, "<fragment>", CompileOpts::default())
            .expect_err("adjacent atoms in an f-string field are a syntax error");
        assert_eq!(
            err.to_string(),
            "invalid syntax. Perhaps you forgot a comma?"
        );
        assert_eq!(err.python_location(), (1, 4));
        assert_eq!(err.python_end_location(), Some((1, 7)));

        compile(
            "f'{not x}'",
            Mode::Exec,
            "<fragment>",
            CompileOpts::default(),
        )
        .expect("unary not is a prefix, not two atoms");
        let err = compile(
            "f'{a and}'",
            Mode::Exec,
            "<fragment>",
            CompileOpts::default(),
        )
        .expect_err("a dangling 'and' is an f-string separator error");
        assert_eq!(
            err.to_string(),
            "f-string: expecting '=', or '!', or ':', or '}'"
        );
    }

    #[test]
    fn missing_comma_diagnostic_spans_the_whole_second_atom() {
        // `(start, end)` reported as one-based character columns, matching
        // `SyntaxError.offset` / `.end_offset`.
        let span = |source: &str| {
            let err = compile(source, Mode::Eval, "<comma>", CompileOpts::default())
                .expect_err("two adjacent atoms are a syntax error");
            assert_eq!(
                err.to_string(),
                "invalid syntax. Perhaps you forgot a comma?"
            );
            (
                err.python_location().1,
                err.python_end_location().unwrap().1,
            )
        };

        // A one-character second atom is the case that already worked.
        assert_eq!(span("(a b)"), (2, 5));
        // A longer one ends where it ends, not one byte in.
        assert_eq!(span("(a bb)"), (2, 6));
        assert_eq!(span("(a bbb)"), (2, 7));
        assert_eq!(span("(1 22)"), (2, 6));
        // A non-ASCII atom is one character but several bytes, so counting
        // bytes here used to stop inside it and round back off the boundary.
        assert_eq!(span("(a \u{3b2})"), (2, 5));
        assert_eq!(span("(a \u{3b2}\u{3b2})"), (2, 6));
        assert_eq!(span("(\u{3b1}\u{3b1} \u{3b2})"), (2, 6));
        // The first atom's width was never the problem; pin it anyway.
        assert_eq!(span("(\u{3b1} b)"), (2, 5));
        // Other bracket kinds take the same path.
        assert_eq!(span("[\u{3b1} \u{3b2}]"), (2, 5));
    }

    #[test]
    fn parenthesized_yield_assignment_uses_invalid_target_message() {
        let err = compile(
            "def f(): (yield bar) = y\n",
            Mode::Exec,
            "<yield>",
            CompileOpts::default(),
        )
        .expect_err("parenthesized yield is not an assignment target");
        assert_eq!(
            err.to_string(),
            "cannot assign to yield expression here. Maybe you meant '==' instead of '='?"
        );
    }

    #[test]
    fn parenthesized_yield_augassign_uses_illegal_expression_message() {
        let err = compile(
            "def f(): (yield bar) += y\n",
            Mode::Exec,
            "<yield>",
            CompileOpts::default(),
        )
        .expect_err("parenthesized yield is not an augmented assignment target");
        assert_eq!(
            err.to_string(),
            "'yield expression' is an illegal expression for augmented assignment"
        );
    }

    #[test]
    fn kwarg_unparenthesized_genexp_uses_eq_or_walrus_message() {
        let err = compile(
            "dict(a = i for i in range(10))\n",
            Mode::Exec,
            "<kwarg>",
            CompileOpts::default(),
        )
        .expect_err("unparenthesized genexp after '=' is invalid");
        assert_eq!(
            err.to_string(),
            "invalid syntax. Maybe you meant '==' or ':=' instead of '='?"
        );
    }

    #[test]
    fn eval_fstring_assignment_keeps_invalid_syntax() {
        for source in ["f'' = 3", "f'{0}' = x", "f'{x}' = x"] {
            let err = compile(source, Mode::Eval, "<eval>", CompileOpts::default())
                .expect_err("assignment is invalid in eval");
            assert_eq!(err.to_string(), "invalid syntax", "{source}");
        }
        let err = compile("f'' = 3", Mode::Exec, "<exec>", CompileOpts::default())
            .expect_err("f-string is not an assignment target");
        assert_eq!(
            err.to_string(),
            "cannot assign to f-string expression here. Maybe you meant '==' instead of '='?"
        );
    }

    #[test]
    fn if_assignment_uses_eq_or_walrus_message() {
        let err = compile(
            "if x = 3: pass\n",
            Mode::Exec,
            "<if>",
            CompileOpts::default(),
        )
        .expect_err("assignment in if condition is invalid");
        assert_eq!(
            err.to_string(),
            "invalid syntax. Maybe you meant '==' or ':=' instead of '='?"
        );
    }

    #[test]
    fn parenthesized_if_assignment_uses_eq_or_walrus_message() {
        for source in [
            "if (x = 3): pass\n",
            "if ((x = 3)): pass\n",
            "if (x = 3) and y: pass\n",
        ] {
            let err = compile(source, Mode::Exec, "<if>", CompileOpts::default())
                .expect_err("parenthesized assignment in if condition is invalid");
            assert_eq!(
                err.to_string(),
                "invalid syntax. Maybe you meant '==' or ':=' instead of '='?"
            );
        }
    }

    #[test]
    fn earlier_syntax_error_is_not_replaced_by_later_condition() {
        let err = compile(
            "@@@\nif x = 3: pass\n",
            Mode::Exec,
            "<if>",
            CompileOpts::default(),
        )
        .expect_err("the first invalid token is the syntax error");
        assert_eq!(err.to_string(), "invalid syntax");
        assert_eq!(err.python_location().0, 1);
    }

    #[test]
    fn if_attribute_assignment_uses_invalid_target_hint() {
        let err = compile(
            "if x.a = 3: pass\n",
            Mode::Exec,
            "<if>",
            CompileOpts::default(),
        )
        .expect_err("attribute assignment in if condition is invalid");
        assert_eq!(
            err.to_string(),
            "cannot assign to attribute here. Maybe you meant '==' instead of '='?"
        );
    }

    #[test]
    fn parenthesized_yield_from_assignment_uses_invalid_target_message() {
        let err = compile(
            "def f(): (yield from value) = target\n",
            Mode::Exec,
            "<yield>",
            CompileOpts::default(),
        )
        .expect_err("parenthesized yield from is not an assignment target");
        assert_eq!(
            err.to_string(),
            "cannot assign to yield expression here. Maybe you meant '==' instead of '='?"
        );
    }

    #[test]
    fn set_display_assignment_uses_invalid_target_hint() {
        let err = compile(
            "{1, 2, 3} = 42\n",
            Mode::Exec,
            "<set>",
            CompileOpts::default(),
        )
        .expect_err("set display is not an assignment target");
        assert_eq!(
            err.to_string(),
            "cannot assign to set display here. Maybe you meant '==' instead of '='?"
        );
    }

    #[test]
    fn obsolete_not_equal_diagnostic_spans_the_whole_operator() {
        let err = compile("2 <> 3\n", Mode::Exec, "<obsolete>", CompileOpts::default())
            .expect_err("'<>' outside Barry mode is a syntax error");
        assert_eq!(err.to_string(), "invalid syntax");
        assert_eq!(err.python_location(), (1, 3));
        assert_eq!(err.python_end_location(), Some((1, 5)));

        // Only `<>` spans two characters; any other token that cannot start an
        // expression keeps its own location.
        let err = compile("2 <;\n", Mode::Exec, "<obsolete>", CompileOpts::default())
            .expect_err("'<;' is a syntax error");
        assert_eq!(err.to_string(), "invalid syntax");
        assert_eq!(err.python_location(), (1, 4));
        assert_eq!(err.python_end_location(), Some((1, 5)));

        // A `<>` that starts a statement is reported at the `<` too, where the
        // parser stops instead of one character in.
        let err = compile("<>\n", Mode::Exec, "<obsolete>", CompileOpts::default())
            .expect_err("a bare '<>' is a syntax error");
        assert_eq!(err.to_string(), "invalid syntax");
        assert_eq!(err.python_location(), (1, 1));
        assert_eq!(err.python_end_location(), Some((1, 3)));

        // A bracket left open earlier in the source outranks the operator.
        let err = compile(
            "(\n2 <> 3",
            Mode::Exec,
            "<obsolete>",
            CompileOpts::default(),
        )
        .expect_err("the bracket is never closed");
        assert_eq!(err.to_string(), "'(' was never closed");
        assert_eq!(err.python_location(), (1, 1));
    }

    #[test]
    fn barry_as_flufl_does_not_rewrite_strings_or_comments() {
        compile(
            "from __future__ import barry_as_FLUFL\nx = '<>'\n# <>\n",
            Mode::Exec,
            "<barry>",
            CompileOpts::default(),
        )
        .expect("Barry markers in strings and comments should stay untouched");
    }

    #[test]
    fn syntax_error_before_barry_not_equal_takes_precedence() {
        let err = compile(
            "from __future__ import barry_as_FLUFL\n<>\n2 != 3\n",
            Mode::Exec,
            "<barry>",
            CompileOpts::default(),
        )
        .expect_err("the earlier invalid comparison should fail");
        assert_eq!(err.to_string(), "invalid syntax");
        assert_eq!(err.python_location(), (2, 1));
    }

    #[test]
    fn unclosed_bracket_before_barry_not_equal_takes_precedence() {
        let err = compile(
            "from __future__ import barry_as_FLUFL\n(\n2 != 3",
            Mode::Exec,
            "<barry>",
            CompileOpts::default(),
        )
        .expect_err("the earlier unclosed bracket should fail");
        assert_eq!(err.to_string(), "'(' was never closed");
        assert_eq!(err.python_location(), (2, 1));
    }

    #[test]
    fn compile_phello() {
        let code = r#"
initialized = True
def main():
    print("Hello world!")
if __name__ == '__main__':
    main()
"#;
        let compiled = compile(code, Mode::Exec, "<>", CompileOpts::default());
        dbg!(compiled.expect("compile error"));
    }

    #[test]
    fn compile_if_elif_else() {
        let code = r#"
if False:
    pass
elif False:
    pass
elif False:
    pass
else:
    pass
"#;
        let compiled = compile(code, Mode::Exec, "<>", CompileOpts::default());
        dbg!(compiled.expect("compile error"));
    }

    #[test]
    fn compile_lambda() {
        let code = r#"
lambda: 'a'
"#;
        let compiled = compile(code, Mode::Exec, "<>", CompileOpts::default());
        dbg!(compiled.expect("compile error"));
    }

    #[test]
    fn compile_lambda2() {
        let code = r#"
(lambda x: f'hello, {x}')('world}')
"#;
        let compiled = compile(code, Mode::Exec, "<>", CompileOpts::default());
        dbg!(compiled.expect("compile error"));
    }

    #[test]
    fn compile_lambda3() {
        let code = r#"
def g():
    pass
def f():
    if False:
        return lambda x: g(x)
    elif False:
        return g
    else:
        return g
"#;
        let compiled = compile(code, Mode::Exec, "<>", CompileOpts::default());
        dbg!(compiled.expect("compile error"));
    }

    #[test]
    fn compile_call_arg_lambda_default() {
        let code = "signature((lambda a=10: a))";
        let compiled = compile(code, Mode::Exec, "<>", CompileOpts::default());
        dbg!(compiled.expect("compile error"));
    }

    #[test]
    fn compile_generic_function_parameter_default() {
        let code = "def __repr__[T: str](self, default: T = '') -> str: pass";
        let compiled = compile(code, Mode::Exec, "<>", CompileOpts::default());
        dbg!(compiled.expect("compile error"));
    }

    #[test]
    fn compile_int() {
        let code = r#"
a = 0xFF
"#;
        let compiled = compile(code, Mode::Exec, "<>", CompileOpts::default());
        dbg!(compiled.expect("compile error"));
    }

    #[test]
    fn compile_bigint() {
        let code = r#"
a = 0xFFFFFFFFFFFFFFFFFFFFFFFF
"#;
        let compiled = compile(code, Mode::Exec, "<>", CompileOpts::default());
        dbg!(compiled.expect("compile error"));
    }

    #[test]
    fn compile_fstring() {
        let code1 = r#"
assert f"1" == '1'
    "#;
        let compiled = compile(code1, Mode::Exec, "<>", CompileOpts::default());
        dbg!(compiled.expect("compile error"));

        let code2 = r#"
assert f"{1}" == '1'
    "#;
        let compiled = compile(code2, Mode::Exec, "<>", CompileOpts::default());
        dbg!(compiled.expect("compile error"));
        let code3 = r#"
assert f"{1+1}" == '2'
    "#;
        let compiled = compile(code3, Mode::Exec, "<>", CompileOpts::default());
        dbg!(compiled.expect("compile error"));

        let code4 = r#"
assert f"{{{(lambda: f'{1}')}" == '{1'
    "#;
        let compiled = compile(code4, Mode::Exec, "<>", CompileOpts::default());
        dbg!(compiled.expect("compile error"));

        let code5 = r#"
assert f"a{1}" == 'a1'
    "#;
        let compiled = compile(code5, Mode::Exec, "<>", CompileOpts::default());
        dbg!(compiled.expect("compile error"));

        let code6 = r#"
assert f"{{{(lambda x: f'hello, {x}')('world}')}" == '{hello, world}'
    "#;
        let compiled = compile(code6, Mode::Exec, "<>", CompileOpts::default());
        dbg!(compiled.expect("compile error"));
    }

    #[test]
    fn simple_enum() {
        let code = r#"
import enum
@enum._simple_enum(enum.IntFlag, boundary=enum.KEEP)
class RegexFlag:
    NOFLAG = 0
    DEBUG = 1
print(RegexFlag.NOFLAG & RegexFlag.DEBUG)
"#;
        let compiled = compile(code, Mode::Exec, "<string>", CompileOpts::default());
        dbg!(compiled.expect("compile error"));
    }
}