mv compiler to compiler/

compiler/rustc_parse/src/lexer/mod.rs

@@ -0,0 +1,598 @@
+use rustc_ast::ast::AttrStyle;
+use rustc_ast::token::{self, CommentKind, Token, TokenKind};
+use rustc_data_structures::sync::Lrc;
+use rustc_errors::{error_code, Applicability, DiagnosticBuilder, FatalError};
+use rustc_lexer::Base;
+use rustc_lexer::{unescape, RawStrError};
+use rustc_session::parse::ParseSess;
+use rustc_span::symbol::{sym, Symbol};
+use rustc_span::{BytePos, Pos, Span};
+
+use std::char;
+use tracing::debug;
+
+mod tokentrees;
+mod unescape_error_reporting;
+mod unicode_chars;
+
+use rustc_lexer::{unescape::Mode, DocStyle};
+use unescape_error_reporting::{emit_unescape_error, push_escaped_char};
+
+#[derive(Clone, Debug)]
+pub struct UnmatchedBrace {
+    pub expected_delim: token::DelimToken,
+    pub found_delim: Option<token::DelimToken>,
+    pub found_span: Span,
+    pub unclosed_span: Option<Span>,
+    pub candidate_span: Option<Span>,
+}
+
+pub struct StringReader<'a> {
+    sess: &'a ParseSess,
+    /// Initial position, read-only.
+    start_pos: BytePos,
+    /// The absolute offset within the source_map of the current character.
+    pos: BytePos,
+    /// Stop reading src at this index.
+    end_src_index: usize,
+    /// Source text to tokenize.
+    src: Lrc<String>,
+    override_span: Option<Span>,
+}
+
+impl<'a> StringReader<'a> {
+    pub fn new(
+        sess: &'a ParseSess,
+        source_file: Lrc<rustc_span::SourceFile>,
+        override_span: Option<Span>,
+    ) -> Self {
+        let src = source_file.src.clone().unwrap_or_else(|| {
+            sess.span_diagnostic
+                .bug(&format!("cannot lex `source_file` without source: {}", source_file.name));
+        });
+
+        StringReader {
+            sess,
+            start_pos: source_file.start_pos,
+            pos: source_file.start_pos,
+            end_src_index: src.len(),
+            src,
+            override_span,
+        }
+    }
+
+    fn mk_sp(&self, lo: BytePos, hi: BytePos) -> Span {
+        self.override_span.unwrap_or_else(|| Span::with_root_ctxt(lo, hi))
+    }
+
+    /// Returns the next token, including trivia like whitespace or comments.
+    pub fn next_token(&mut self) -> Token {
+        let start_src_index = self.src_index(self.pos);
+        let text: &str = &self.src[start_src_index..self.end_src_index];
+
+        if text.is_empty() {
+            let span = self.mk_sp(self.pos, self.pos);
+            return Token::new(token::Eof, span);
+        }
+
+        {
+            let is_beginning_of_file = self.pos == self.start_pos;
+            if is_beginning_of_file {
+                if let Some(shebang_len) = rustc_lexer::strip_shebang(text) {
+                    let start = self.pos;
+                    self.pos = self.pos + BytePos::from_usize(shebang_len);
+
+                    let sym = self.symbol_from(start + BytePos::from_usize("#!".len()));
+                    let kind = token::Shebang(sym);
+
+                    let span = self.mk_sp(start, self.pos);
+                    return Token::new(kind, span);
+                }
+            }
+        }
+
+        let token = rustc_lexer::first_token(text);
+
+        let start = self.pos;
+        self.pos = self.pos + BytePos::from_usize(token.len);
+
+        debug!("try_next_token: {:?}({:?})", token.kind, self.str_from(start));
+
+        let kind = self.cook_lexer_token(token.kind, start);
+        let span = self.mk_sp(start, self.pos);
+        Token::new(kind, span)
+    }
+
+    /// Report a fatal lexical error with a given span.
+    fn fatal_span(&self, sp: Span, m: &str) -> FatalError {
+        self.sess.span_diagnostic.span_fatal(sp, m)
+    }
+
+    /// Report a lexical error with a given span.
+    fn err_span(&self, sp: Span, m: &str) {
+        self.sess.span_diagnostic.struct_span_err(sp, m).emit();
+    }
+
+    /// Report a fatal error spanning [`from_pos`, `to_pos`).
+    fn fatal_span_(&self, from_pos: BytePos, to_pos: BytePos, m: &str) -> FatalError {
+        self.fatal_span(self.mk_sp(from_pos, to_pos), m)
+    }
+
+    /// Report a lexical error spanning [`from_pos`, `to_pos`).
+    fn err_span_(&self, from_pos: BytePos, to_pos: BytePos, m: &str) {
+        self.err_span(self.mk_sp(from_pos, to_pos), m)
+    }
+
+    fn struct_fatal_span_char(
+        &self,
+        from_pos: BytePos,
+        to_pos: BytePos,
+        m: &str,
+        c: char,
+    ) -> DiagnosticBuilder<'a> {
+        let mut m = m.to_string();
+        m.push_str(": ");
+        push_escaped_char(&mut m, c);
+
+        self.sess.span_diagnostic.struct_span_fatal(self.mk_sp(from_pos, to_pos), &m[..])
+    }
+
+    /// Turns simple `rustc_lexer::TokenKind` enum into a rich
+    /// `librustc_ast::TokenKind`. This turns strings into interned
+    /// symbols and runs additional validation.
+    fn cook_lexer_token(&self, token: rustc_lexer::TokenKind, start: BytePos) -> TokenKind {
+        match token {
+            rustc_lexer::TokenKind::LineComment { doc_style } => {
+                match doc_style {
+                    Some(doc_style) => {
+                        // Opening delimiter of the length 3 is not included into the symbol.
+                        let content_start = start + BytePos(3);
+                        let content = self.str_from(content_start);
+
+                        self.cook_doc_comment(content_start, content, CommentKind::Line, doc_style)
+                    }
+                    None => token::Comment,
+                }
+            }
+            rustc_lexer::TokenKind::BlockComment { doc_style, terminated } => {
+                if !terminated {
+                    let msg = match doc_style {
+                        Some(_) => "unterminated block doc-comment",
+                        None => "unterminated block comment",
+                    };
+                    let last_bpos = self.pos;
+                    self.sess
+                        .span_diagnostic
+                        .struct_span_fatal_with_code(
+                            self.mk_sp(start, last_bpos),
+                            msg,
+                            error_code!(E0758),
+                        )
+                        .emit();
+                    FatalError.raise();
+                }
+                match doc_style {
+                    Some(doc_style) => {
+                        // Opening delimiter of the length 3 and closing delimiter of the length 2
+                        // are not included into the symbol.
+                        let content_start = start + BytePos(3);
+                        let content_end = self.pos - BytePos(if terminated { 2 } else { 0 });
+                        let content = self.str_from_to(content_start, content_end);
+
+                        self.cook_doc_comment(content_start, content, CommentKind::Block, doc_style)
+                    }
+                    None => token::Comment,
+                }
+            }
+            rustc_lexer::TokenKind::Whitespace => token::Whitespace,
+            rustc_lexer::TokenKind::Ident | rustc_lexer::TokenKind::RawIdent => {
+                let is_raw_ident = token == rustc_lexer::TokenKind::RawIdent;
+                let mut ident_start = start;
+                if is_raw_ident {
+                    ident_start = ident_start + BytePos(2);
+                }
+                let sym = nfc_normalize(self.str_from(ident_start));
+                let span = self.mk_sp(start, self.pos);
+                self.sess.symbol_gallery.insert(sym, span);
+                if is_raw_ident {
+                    if !sym.can_be_raw() {
+                        self.err_span(span, &format!("`{}` cannot be a raw identifier", sym));
+                    }
+                    self.sess.raw_identifier_spans.borrow_mut().push(span);
+                }
+                token::Ident(sym, is_raw_ident)
+            }
+            rustc_lexer::TokenKind::Literal { kind, suffix_start } => {
+                let suffix_start = start + BytePos(suffix_start as u32);
+                let (kind, symbol) = self.cook_lexer_literal(start, suffix_start, kind);
+                let suffix = if suffix_start < self.pos {
+                    let string = self.str_from(suffix_start);
+                    if string == "_" {
+                        self.sess
+                            .span_diagnostic
+                            .struct_span_warn(
+                                self.mk_sp(suffix_start, self.pos),
+                                "underscore literal suffix is not allowed",
+                            )
+                            .warn(
+                                "this was previously accepted by the compiler but is \
+                                   being phased out; it will become a hard error in \
+                                   a future release!",
+                            )
+                            .note(
+                                "see issue #42326 \
+                                 <https://github.com/rust-lang/rust/issues/42326> \
+                                 for more information",
+                            )
+                            .emit();
+                        None
+                    } else {
+                        Some(Symbol::intern(string))
+                    }
+                } else {
+                    None
+                };
+                token::Literal(token::Lit { kind, symbol, suffix })
+            }
+            rustc_lexer::TokenKind::Lifetime { starts_with_number } => {
+                // Include the leading `'` in the real identifier, for macro
+                // expansion purposes. See #12512 for the gory details of why
+                // this is necessary.
+                let lifetime_name = self.str_from(start);
+                if starts_with_number {
+                    self.err_span_(start, self.pos, "lifetimes cannot start with a number");
+                }
+                let ident = Symbol::intern(lifetime_name);
+                token::Lifetime(ident)
+            }
+            rustc_lexer::TokenKind::Semi => token::Semi,
+            rustc_lexer::TokenKind::Comma => token::Comma,
+            rustc_lexer::TokenKind::Dot => token::Dot,
+            rustc_lexer::TokenKind::OpenParen => token::OpenDelim(token::Paren),
+            rustc_lexer::TokenKind::CloseParen => token::CloseDelim(token::Paren),
+            rustc_lexer::TokenKind::OpenBrace => token::OpenDelim(token::Brace),
+            rustc_lexer::TokenKind::CloseBrace => token::CloseDelim(token::Brace),
+            rustc_lexer::TokenKind::OpenBracket => token::OpenDelim(token::Bracket),
+            rustc_lexer::TokenKind::CloseBracket => token::CloseDelim(token::Bracket),
+            rustc_lexer::TokenKind::At => token::At,
+            rustc_lexer::TokenKind::Pound => token::Pound,
+            rustc_lexer::TokenKind::Tilde => token::Tilde,
+            rustc_lexer::TokenKind::Question => token::Question,
+            rustc_lexer::TokenKind::Colon => token::Colon,
+            rustc_lexer::TokenKind::Dollar => token::Dollar,
+            rustc_lexer::TokenKind::Eq => token::Eq,
+            rustc_lexer::TokenKind::Bang => token::Not,
+            rustc_lexer::TokenKind::Lt => token::Lt,
+            rustc_lexer::TokenKind::Gt => token::Gt,
+            rustc_lexer::TokenKind::Minus => token::BinOp(token::Minus),
+            rustc_lexer::TokenKind::And => token::BinOp(token::And),
+            rustc_lexer::TokenKind::Or => token::BinOp(token::Or),
+            rustc_lexer::TokenKind::Plus => token::BinOp(token::Plus),
+            rustc_lexer::TokenKind::Star => token::BinOp(token::Star),
+            rustc_lexer::TokenKind::Slash => token::BinOp(token::Slash),
+            rustc_lexer::TokenKind::Caret => token::BinOp(token::Caret),
+            rustc_lexer::TokenKind::Percent => token::BinOp(token::Percent),
+
+            rustc_lexer::TokenKind::Unknown => {
+                let c = self.str_from(start).chars().next().unwrap();
+                let mut err =
+                    self.struct_fatal_span_char(start, self.pos, "unknown start of token", c);
+                // FIXME: the lexer could be used to turn the ASCII version of unicode homoglyphs,
+                // instead of keeping a table in `check_for_substitution`into the token. Ideally,
+                // this should be inside `rustc_lexer`. However, we should first remove compound
+                // tokens like `<<` from `rustc_lexer`, and then add fancier error recovery to it,
+                // as there will be less overall work to do this way.
+                let token = unicode_chars::check_for_substitution(self, start, c, &mut err)
+                    .unwrap_or_else(|| token::Unknown(self.symbol_from(start)));
+                err.emit();
+                token
+            }
+        }
+    }
+
+    fn cook_doc_comment(
+        &self,
+        content_start: BytePos,
+        content: &str,
+        comment_kind: CommentKind,
+        doc_style: DocStyle,
+    ) -> TokenKind {
+        if content.contains('\r') {
+            for (idx, _) in content.char_indices().filter(|&(_, c)| c == '\r') {
+                self.err_span_(
+                    content_start + BytePos(idx as u32),
+                    content_start + BytePos(idx as u32 + 1),
+                    match comment_kind {
+                        CommentKind::Line => "bare CR not allowed in doc-comment",
+                        CommentKind::Block => "bare CR not allowed in block doc-comment",
+                    },
+                );
+            }
+        }
+
+        let attr_style = match doc_style {
+            DocStyle::Outer => AttrStyle::Outer,
+            DocStyle::Inner => AttrStyle::Inner,
+        };
+
+        token::DocComment(comment_kind, attr_style, Symbol::intern(content))
+    }
+
+    fn cook_lexer_literal(
+        &self,
+        start: BytePos,
+        suffix_start: BytePos,
+        kind: rustc_lexer::LiteralKind,
+    ) -> (token::LitKind, Symbol) {
+        // prefix means `"` or `br"` or `r###"`, ...
+        let (lit_kind, mode, prefix_len, postfix_len) = match kind {
+            rustc_lexer::LiteralKind::Char { terminated } => {
+                if !terminated {
+                    self.sess
+                        .span_diagnostic
+                        .struct_span_fatal_with_code(
+                            self.mk_sp(start, suffix_start),
+                            "unterminated character literal",
+                            error_code!(E0762),
+                        )
+                        .emit();
+                    FatalError.raise();
+                }
+                (token::Char, Mode::Char, 1, 1) // ' '
+            }
+            rustc_lexer::LiteralKind::Byte { terminated } => {
+                if !terminated {
+                    self.sess
+                        .span_diagnostic
+                        .struct_span_fatal_with_code(
+                            self.mk_sp(start + BytePos(1), suffix_start),
+                            "unterminated byte constant",
+                            error_code!(E0763),
+                        )
+                        .emit();
+                    FatalError.raise();
+                }
+                (token::Byte, Mode::Byte, 2, 1) // b' '
+            }
+            rustc_lexer::LiteralKind::Str { terminated } => {
+                if !terminated {
+                    self.sess
+                        .span_diagnostic
+                        .struct_span_fatal_with_code(
+                            self.mk_sp(start, suffix_start),
+                            "unterminated double quote string",
+                            error_code!(E0765),
+                        )
+                        .emit();
+                    FatalError.raise();
+                }
+                (token::Str, Mode::Str, 1, 1) // " "
+            }
+            rustc_lexer::LiteralKind::ByteStr { terminated } => {
+                if !terminated {
+                    self.sess
+                        .span_diagnostic
+                        .struct_span_fatal_with_code(
+                            self.mk_sp(start + BytePos(1), suffix_start),
+                            "unterminated double quote byte string",
+                            error_code!(E0766),
+                        )
+                        .emit();
+                    FatalError.raise();
+                }
+                (token::ByteStr, Mode::ByteStr, 2, 1) // b" "
+            }
+            rustc_lexer::LiteralKind::RawStr { n_hashes, err } => {
+                self.report_raw_str_error(start, err);
+                let n = u32::from(n_hashes);
+                (token::StrRaw(n_hashes), Mode::RawStr, 2 + n, 1 + n) // r##" "##
+            }
+            rustc_lexer::LiteralKind::RawByteStr { n_hashes, err } => {
+                self.report_raw_str_error(start, err);
+                let n = u32::from(n_hashes);
+                (token::ByteStrRaw(n_hashes), Mode::RawByteStr, 3 + n, 1 + n) // br##" "##
+            }
+            rustc_lexer::LiteralKind::Int { base, empty_int } => {
+                return if empty_int {
+                    self.sess
+                        .span_diagnostic
+                        .struct_span_err_with_code(
+                            self.mk_sp(start, suffix_start),
+                            "no valid digits found for number",
+                            error_code!(E0768),
+                        )
+                        .emit();
+                    (token::Integer, sym::integer(0))
+                } else {
+                    self.validate_int_literal(base, start, suffix_start);
+                    (token::Integer, self.symbol_from_to(start, suffix_start))
+                };
+            }
+            rustc_lexer::LiteralKind::Float { base, empty_exponent } => {
+                if empty_exponent {
+                    self.err_span_(start, self.pos, "expected at least one digit in exponent");
+                }
+
+                match base {
+                    Base::Hexadecimal => self.err_span_(
+                        start,
+                        suffix_start,
+                        "hexadecimal float literal is not supported",
+                    ),
+                    Base::Octal => {
+                        self.err_span_(start, suffix_start, "octal float literal is not supported")
+                    }
+                    Base::Binary => {
+                        self.err_span_(start, suffix_start, "binary float literal is not supported")
+                    }
+                    _ => (),
+                }
+
+                let id = self.symbol_from_to(start, suffix_start);
+                return (token::Float, id);
+            }
+        };
+        let content_start = start + BytePos(prefix_len);
+        let content_end = suffix_start - BytePos(postfix_len);
+        let id = self.symbol_from_to(content_start, content_end);
+        self.validate_literal_escape(mode, content_start, content_end);
+        (lit_kind, id)
+    }
+
+    pub fn pos(&self) -> BytePos {
+        self.pos
+    }
+
+    #[inline]
+    fn src_index(&self, pos: BytePos) -> usize {
+        (pos - self.start_pos).to_usize()
+    }
+
+    /// Slice of the source text from `start` up to but excluding `self.pos`,
+    /// meaning the slice does not include the character `self.ch`.
+    fn str_from(&self, start: BytePos) -> &str {
+        self.str_from_to(start, self.pos)
+    }
+
+    /// Creates a Symbol from a given offset to the current offset.
+    fn symbol_from(&self, start: BytePos) -> Symbol {
+        debug!("taking an ident from {:?} to {:?}", start, self.pos);
+        Symbol::intern(self.str_from(start))
+    }
+
+    /// As symbol_from, with an explicit endpoint.
+    fn symbol_from_to(&self, start: BytePos, end: BytePos) -> Symbol {
+        debug!("taking an ident from {:?} to {:?}", start, end);
+        Symbol::intern(self.str_from_to(start, end))
+    }
+
+    /// Slice of the source text spanning from `start` up to but excluding `end`.
+    fn str_from_to(&self, start: BytePos, end: BytePos) -> &str {
+        &self.src[self.src_index(start)..self.src_index(end)]
+    }
+
+    fn report_raw_str_error(&self, start: BytePos, opt_err: Option<RawStrError>) {
+        match opt_err {
+            Some(RawStrError::InvalidStarter { bad_char }) => {
+                self.report_non_started_raw_string(start, bad_char)
+            }
+            Some(RawStrError::NoTerminator { expected, found, possible_terminator_offset }) => self
+                .report_unterminated_raw_string(start, expected, possible_terminator_offset, found),
+            Some(RawStrError::TooManyDelimiters { found }) => {
+                self.report_too_many_hashes(start, found)
+            }
+            None => (),
+        }
+    }
+
+    fn report_non_started_raw_string(&self, start: BytePos, bad_char: char) -> ! {
+        self.struct_fatal_span_char(
+            start,
+            self.pos,
+            "found invalid character; only `#` is allowed in raw string delimitation",
+            bad_char,
+        )
+        .emit();
+        FatalError.raise()
+    }
+
+    fn report_unterminated_raw_string(
+        &self,
+        start: BytePos,
+        n_hashes: usize,
+        possible_offset: Option<usize>,
+        found_terminators: usize,
+    ) -> ! {
+        let mut err = self.sess.span_diagnostic.struct_span_fatal_with_code(
+            self.mk_sp(start, start),
+            "unterminated raw string",
+            error_code!(E0748),
+        );
+
+        err.span_label(self.mk_sp(start, start), "unterminated raw string");
+
+        if n_hashes > 0 {
+            err.note(&format!(
+                "this raw string should be terminated with `\"{}`",
+                "#".repeat(n_hashes)
+            ));
+        }
+
+        if let Some(possible_offset) = possible_offset {
+            let lo = start + BytePos(possible_offset as u32);
+            let hi = lo + BytePos(found_terminators as u32);
+            let span = self.mk_sp(lo, hi);
+            err.span_suggestion(
+                span,
+                "consider terminating the string here",
+                "#".repeat(n_hashes),
+                Applicability::MaybeIncorrect,
+            );
+        }
+
+        err.emit();
+        FatalError.raise()
+    }
+
+    /// Note: It was decided to not add a test case, because it would be to big.
+    /// https://github.com/rust-lang/rust/pull/50296#issuecomment-392135180
+    fn report_too_many_hashes(&self, start: BytePos, found: usize) -> ! {
+        self.fatal_span_(
+            start,
+            self.pos,
+            &format!(
+                "too many `#` symbols: raw strings may be delimited \
+                by up to 65535 `#` symbols, but found {}",
+                found
+            ),
+        )
+        .raise();
+    }
+
+    fn validate_literal_escape(&self, mode: Mode, content_start: BytePos, content_end: BytePos) {
+        let lit_content = self.str_from_to(content_start, content_end);
+        unescape::unescape_literal(lit_content, mode, &mut |range, result| {
+            // Here we only check for errors. The actual unescaping is done later.
+            if let Err(err) = result {
+                let span_with_quotes =
+                    self.mk_sp(content_start - BytePos(1), content_end + BytePos(1));
+                emit_unescape_error(
+                    &self.sess.span_diagnostic,
+                    lit_content,
+                    span_with_quotes,
+                    mode,
+                    range,
+                    err,
+                );
+            }
+        });
+    }
+
+    fn validate_int_literal(&self, base: Base, content_start: BytePos, content_end: BytePos) {
+        let base = match base {
+            Base::Binary => 2,
+            Base::Octal => 8,
+            _ => return,
+        };
+        let s = self.str_from_to(content_start + BytePos(2), content_end);
+        for (idx, c) in s.char_indices() {
+            let idx = idx as u32;
+            if c != '_' && c.to_digit(base).is_none() {
+                let lo = content_start + BytePos(2 + idx);
+                let hi = content_start + BytePos(2 + idx + c.len_utf8() as u32);
+                self.err_span_(lo, hi, &format!("invalid digit for a base {} literal", base));
+            }
+        }
+    }
+}
+
+pub fn nfc_normalize(string: &str) -> Symbol {
+    use unicode_normalization::{is_nfc_quick, IsNormalized, UnicodeNormalization};
+    match is_nfc_quick(string.chars()) {
+        IsNormalized::Yes => Symbol::intern(string),
+        _ => {
+            let normalized_str: String = string.chars().nfc().collect();
+            Symbol::intern(&normalized_str)
+        }
+    }
+}

compiler/rustc_parse/src/lexer/tokentrees.rs

@@ -0,0 +1,313 @@
+use super::{StringReader, UnmatchedBrace};
+
+use rustc_ast::token::{self, DelimToken, Token};
+use rustc_ast::tokenstream::{
+    DelimSpan,
+    IsJoint::{self, *},
+    TokenStream, TokenTree, TreeAndJoint,
+};
+use rustc_ast_pretty::pprust::token_to_string;
+use rustc_data_structures::fx::FxHashMap;
+use rustc_errors::PResult;
+use rustc_span::Span;
+
+impl<'a> StringReader<'a> {
+    crate fn into_token_trees(self) -> (PResult<'a, TokenStream>, Vec<UnmatchedBrace>) {
+        let mut tt_reader = TokenTreesReader {
+            string_reader: self,
+            token: Token::dummy(),
+            joint_to_prev: Joint,
+            open_braces: Vec::new(),
+            unmatched_braces: Vec::new(),
+            matching_delim_spans: Vec::new(),
+            last_unclosed_found_span: None,
+            last_delim_empty_block_spans: FxHashMap::default(),
+            matching_block_spans: Vec::new(),
+        };
+        let res = tt_reader.parse_all_token_trees();
+        (res, tt_reader.unmatched_braces)
+    }
+}
+
+struct TokenTreesReader<'a> {
+    string_reader: StringReader<'a>,
+    token: Token,
+    joint_to_prev: IsJoint,
+    /// Stack of open delimiters and their spans. Used for error message.
+    open_braces: Vec<(token::DelimToken, Span)>,
+    unmatched_braces: Vec<UnmatchedBrace>,
+    /// The type and spans for all braces
+    ///
+    /// Used only for error recovery when arriving to EOF with mismatched braces.
+    matching_delim_spans: Vec<(token::DelimToken, Span, Span)>,
+    last_unclosed_found_span: Option<Span>,
+    /// Collect empty block spans that might have been auto-inserted by editors.
+    last_delim_empty_block_spans: FxHashMap<token::DelimToken, Span>,
+    /// Collect the spans of braces (Open, Close). Used only
+    /// for detecting if blocks are empty and only braces.
+    matching_block_spans: Vec<(Span, Span)>,
+}
+
+impl<'a> TokenTreesReader<'a> {
+    // Parse a stream of tokens into a list of `TokenTree`s, up to an `Eof`.
+    fn parse_all_token_trees(&mut self) -> PResult<'a, TokenStream> {
+        let mut buf = TokenStreamBuilder::default();
+
+        self.real_token();
+        while self.token != token::Eof {
+            buf.push(self.parse_token_tree()?);
+        }
+
+        Ok(buf.into_token_stream())
+    }
+
+    // Parse a stream of tokens into a list of `TokenTree`s, up to a `CloseDelim`.
+    fn parse_token_trees_until_close_delim(&mut self) -> TokenStream {
+        let mut buf = TokenStreamBuilder::default();
+        loop {
+            if let token::CloseDelim(..) = self.token.kind {
+                return buf.into_token_stream();
+            }
+
+            match self.parse_token_tree() {
+                Ok(tree) => buf.push(tree),
+                Err(mut e) => {
+                    e.emit();
+                    return buf.into_token_stream();
+                }
+            }
+        }
+    }
+
+    fn parse_token_tree(&mut self) -> PResult<'a, TreeAndJoint> {
+        let sm = self.string_reader.sess.source_map();
+
+        match self.token.kind {
+            token::Eof => {
+                let msg = "this file contains an unclosed delimiter";
+                let mut err =
+                    self.string_reader.sess.span_diagnostic.struct_span_err(self.token.span, msg);
+                for &(_, sp) in &self.open_braces {
+                    err.span_label(sp, "unclosed delimiter");
+                    self.unmatched_braces.push(UnmatchedBrace {
+                        expected_delim: token::DelimToken::Brace,
+                        found_delim: None,
+                        found_span: self.token.span,
+                        unclosed_span: Some(sp),
+                        candidate_span: None,
+                    });
+                }
+
+                if let Some((delim, _)) = self.open_braces.last() {
+                    if let Some((_, open_sp, close_sp)) =
+                        self.matching_delim_spans.iter().find(|(d, open_sp, close_sp)| {
+                            if let Some(close_padding) = sm.span_to_margin(*close_sp) {
+                                if let Some(open_padding) = sm.span_to_margin(*open_sp) {
+                                    return delim == d && close_padding != open_padding;
+                                }
+                            }
+                            false
+                        })
+                    // these are in reverse order as they get inserted on close, but
+                    {
+                        // we want the last open/first close
+                        err.span_label(*open_sp, "this delimiter might not be properly closed...");
+                        err.span_label(
+                            *close_sp,
+                            "...as it matches this but it has different indentation",
+                        );
+                    }
+                }
+                Err(err)
+            }
+            token::OpenDelim(delim) => {
+                // The span for beginning of the delimited section
+                let pre_span = self.token.span;
+
+                // Parse the open delimiter.
+                self.open_braces.push((delim, self.token.span));
+                self.real_token();
+
+                // Parse the token trees within the delimiters.
+                // We stop at any delimiter so we can try to recover if the user
+                // uses an incorrect delimiter.
+                let tts = self.parse_token_trees_until_close_delim();
+
+                // Expand to cover the entire delimited token tree
+                let delim_span = DelimSpan::from_pair(pre_span, self.token.span);
+
+                match self.token.kind {
+                    // Correct delimiter.
+                    token::CloseDelim(d) if d == delim => {
+                        let (open_brace, open_brace_span) = self.open_braces.pop().unwrap();
+                        let close_brace_span = self.token.span;
+
+                        if tts.is_empty() {
+                            let empty_block_span = open_brace_span.to(close_brace_span);
+                            if !sm.is_multiline(empty_block_span) {
+                                // Only track if the block is in the form of `{}`, otherwise it is
+                                // likely that it was written on purpose.
+                                self.last_delim_empty_block_spans.insert(delim, empty_block_span);
+                            }
+                        }
+
+                        match (open_brace, delim) {
+                            //only add braces
+                            (DelimToken::Brace, DelimToken::Brace) => {
+                                self.matching_block_spans.push((open_brace_span, close_brace_span));
+                            }
+                            _ => {}
+                        }
+
+                        if self.open_braces.is_empty() {
+                            // Clear up these spans to avoid suggesting them as we've found
+                            // properly matched delimiters so far for an entire block.
+                            self.matching_delim_spans.clear();
+                        } else {
+                            self.matching_delim_spans.push((
+                                open_brace,
+                                open_brace_span,
+                                close_brace_span,
+                            ));
+                        }
+                        // Parse the closing delimiter.
+                        self.real_token();
+                    }
+                    // Incorrect delimiter.
+                    token::CloseDelim(other) => {
+                        let mut unclosed_delimiter = None;
+                        let mut candidate = None;
+
+                        if self.last_unclosed_found_span != Some(self.token.span) {
+                            // do not complain about the same unclosed delimiter multiple times
+                            self.last_unclosed_found_span = Some(self.token.span);
+                            // This is a conservative error: only report the last unclosed
+                            // delimiter. The previous unclosed delimiters could actually be
+                            // closed! The parser just hasn't gotten to them yet.
+                            if let Some(&(_, sp)) = self.open_braces.last() {
+                                unclosed_delimiter = Some(sp);
+                            };
+                            if let Some(current_padding) = sm.span_to_margin(self.token.span) {
+                                for (brace, brace_span) in &self.open_braces {
+                                    if let Some(padding) = sm.span_to_margin(*brace_span) {
+                                        // high likelihood of these two corresponding
+                                        if current_padding == padding && brace == &other {
+                                            candidate = Some(*brace_span);
+                                        }
+                                    }
+                                }
+                            }
+                            let (tok, _) = self.open_braces.pop().unwrap();
+                            self.unmatched_braces.push(UnmatchedBrace {
+                                expected_delim: tok,
+                                found_delim: Some(other),
+                                found_span: self.token.span,
+                                unclosed_span: unclosed_delimiter,
+                                candidate_span: candidate,
+                            });
+                        } else {
+                            self.open_braces.pop();
+                        }
+
+                        // If the incorrect delimiter matches an earlier opening
+                        // delimiter, then don't consume it (it can be used to
+                        // close the earlier one). Otherwise, consume it.
+                        // E.g., we try to recover from:
+                        // fn foo() {
+                        //     bar(baz(
+                        // }  // Incorrect delimiter but matches the earlier `{`
+                        if !self.open_braces.iter().any(|&(b, _)| b == other) {
+                            self.real_token();
+                        }
+                    }
+                    token::Eof => {
+                        // Silently recover, the EOF token will be seen again
+                        // and an error emitted then. Thus we don't pop from
+                        // self.open_braces here.
+                    }
+                    _ => {}
+                }
+
+                Ok(TokenTree::Delimited(delim_span, delim, tts).into())
+            }
+            token::CloseDelim(delim) => {
+                // An unexpected closing delimiter (i.e., there is no
+                // matching opening delimiter).
+                let token_str = token_to_string(&self.token);
+                let msg = format!("unexpected closing delimiter: `{}`", token_str);
+                let mut err =
+                    self.string_reader.sess.span_diagnostic.struct_span_err(self.token.span, &msg);
+
+                // Braces are added at the end, so the last element is the biggest block
+                if let Some(parent) = self.matching_block_spans.last() {
+                    if let Some(span) = self.last_delim_empty_block_spans.remove(&delim) {
+                        // Check if the (empty block) is in the last properly closed block
+                        if (parent.0.to(parent.1)).contains(span) {
+                            err.span_label(
+                                span,
+                                "block is empty, you might have not meant to close it",
+                            );
+                        } else {
+                            err.span_label(parent.0, "this opening brace...");
+
+                            err.span_label(parent.1, "...matches this closing brace");
+                        }
+                    } else {
+                        err.span_label(parent.0, "this opening brace...");
+
+                        err.span_label(parent.1, "...matches this closing brace");
+                    }
+                }
+
+                err.span_label(self.token.span, "unexpected closing delimiter");
+                Err(err)
+            }
+            _ => {
+                let tt = TokenTree::Token(self.token.take());
+                self.real_token();
+                let is_joint = self.joint_to_prev == Joint && self.token.is_op();
+                Ok((tt, if is_joint { Joint } else { NonJoint }))
+            }
+        }
+    }
+
+    fn real_token(&mut self) {
+        self.joint_to_prev = Joint;
+        loop {
+            let token = self.string_reader.next_token();
+            match token.kind {
+                token::Whitespace | token::Comment | token::Shebang(_) | token::Unknown(_) => {
+                    self.joint_to_prev = NonJoint;
+                }
+                _ => {
+                    self.token = token;
+                    return;
+                }
+            }
+        }
+    }
+}
+
+#[derive(Default)]
+struct TokenStreamBuilder {
+    buf: Vec<TreeAndJoint>,
+}
+
+impl TokenStreamBuilder {
+    fn push(&mut self, (tree, joint): TreeAndJoint) {
+        if let Some((TokenTree::Token(prev_token), Joint)) = self.buf.last() {
+            if let TokenTree::Token(token) = &tree {
+                if let Some(glued) = prev_token.glue(token) {
+                    self.buf.pop();
+                    self.buf.push((TokenTree::Token(glued), joint));
+                    return;
+                }
+            }
+        }
+        self.buf.push((tree, joint))
+    }
+
+    fn into_token_stream(self) -> TokenStream {
+        TokenStream::new(self.buf)
+    }
+}

compiler/rustc_parse/src/lexer/unescape_error_reporting.rs

@@ -0,0 +1,223 @@
+//! Utilities for rendering escape sequence errors as diagnostics.
+
+use std::iter::once;
+use std::ops::Range;
+
+use rustc_errors::{Applicability, Handler};
+use rustc_lexer::unescape::{EscapeError, Mode};
+use rustc_span::{BytePos, Span};
+
+pub(crate) fn emit_unescape_error(
+    handler: &Handler,
+    // interior part of the literal, without quotes
+    lit: &str,
+    // full span of the literal, including quotes
+    span_with_quotes: Span,
+    mode: Mode,
+    // range of the error inside `lit`
+    range: Range<usize>,
+    error: EscapeError,
+) {
+    tracing::debug!(
+        "emit_unescape_error: {:?}, {:?}, {:?}, {:?}, {:?}",
+        lit,
+        span_with_quotes,
+        mode,
+        range,
+        error
+    );
+    let span = {
+        let Range { start, end } = range;
+        let (start, end) = (start as u32, end as u32);
+        let lo = span_with_quotes.lo() + BytePos(start + 1);
+        let hi = lo + BytePos(end - start);
+        span_with_quotes.with_lo(lo).with_hi(hi)
+    };
+    let last_char = || {
+        let c = lit[range.clone()].chars().rev().next().unwrap();
+        let span = span.with_lo(span.hi() - BytePos(c.len_utf8() as u32));
+        (c, span)
+    };
+    match error {
+        EscapeError::LoneSurrogateUnicodeEscape => {
+            handler
+                .struct_span_err(span, "invalid unicode character escape")
+                .help("unicode escape must not be a surrogate")
+                .emit();
+        }
+        EscapeError::OutOfRangeUnicodeEscape => {
+            handler
+                .struct_span_err(span, "invalid unicode character escape")
+                .help("unicode escape must be at most 10FFFF")
+                .emit();
+        }
+        EscapeError::MoreThanOneChar => {
+            let msg = if mode.is_bytes() {
+                "if you meant to write a byte string literal, use double quotes"
+            } else {
+                "if you meant to write a `str` literal, use double quotes"
+            };
+
+            handler
+                .struct_span_err(
+                    span_with_quotes,
+                    "character literal may only contain one codepoint",
+                )
+                .span_suggestion(
+                    span_with_quotes,
+                    msg,
+                    format!("\"{}\"", lit),
+                    Applicability::MachineApplicable,
+                )
+                .emit();
+        }
+        EscapeError::EscapeOnlyChar => {
+            let (c, _span) = last_char();
+
+            let mut msg = if mode.is_bytes() {
+                "byte constant must be escaped: "
+            } else {
+                "character constant must be escaped: "
+            }
+            .to_string();
+            push_escaped_char(&mut msg, c);
+
+            handler.span_err(span, msg.as_str())
+        }
+        EscapeError::BareCarriageReturn => {
+            let msg = if mode.in_double_quotes() {
+                "bare CR not allowed in string, use \\r instead"
+            } else {
+                "character constant must be escaped: \\r"
+            };
+            handler.span_err(span, msg);
+        }
+        EscapeError::BareCarriageReturnInRawString => {
+            assert!(mode.in_double_quotes());
+            let msg = "bare CR not allowed in raw string";
+            handler.span_err(span, msg);
+        }
+        EscapeError::InvalidEscape => {
+            let (c, span) = last_char();
+
+            let label =
+                if mode.is_bytes() { "unknown byte escape" } else { "unknown character escape" };
+            let mut msg = label.to_string();
+            msg.push_str(": ");
+            push_escaped_char(&mut msg, c);
+
+            let mut diag = handler.struct_span_err(span, msg.as_str());
+            diag.span_label(span, label);
+            if c == '{' || c == '}' && !mode.is_bytes() {
+                diag.help(
+                    "if used in a formatting string, \
+                           curly braces are escaped with `{{` and `}}`",
+                );
+            } else if c == '\r' {
+                diag.help(
+                    "this is an isolated carriage return; \
+                           consider checking your editor and version control settings",
+                );
+            }
+            diag.emit();
+        }
+        EscapeError::TooShortHexEscape => {
+            handler.span_err(span, "numeric character escape is too short")
+        }
+        EscapeError::InvalidCharInHexEscape | EscapeError::InvalidCharInUnicodeEscape => {
+            let (c, span) = last_char();
+
+            let mut msg = if error == EscapeError::InvalidCharInHexEscape {
+                "invalid character in numeric character escape: "
+            } else {
+                "invalid character in unicode escape: "
+            }
+            .to_string();
+            push_escaped_char(&mut msg, c);
+
+            handler.span_err(span, msg.as_str())
+        }
+        EscapeError::NonAsciiCharInByte => {
+            assert!(mode.is_bytes());
+            let (_c, span) = last_char();
+            handler.span_err(
+                span,
+                "byte constant must be ASCII. \
+                                    Use a \\xHH escape for a non-ASCII byte",
+            )
+        }
+        EscapeError::NonAsciiCharInByteString => {
+            assert!(mode.is_bytes());
+            let (_c, span) = last_char();
+            handler.span_err(span, "raw byte string must be ASCII")
+        }
+        EscapeError::OutOfRangeHexEscape => handler.span_err(
+            span,
+            "this form of character escape may only be used \
+                                    with characters in the range [\\x00-\\x7f]",
+        ),
+        EscapeError::LeadingUnderscoreUnicodeEscape => {
+            let (_c, span) = last_char();
+            handler.span_err(span, "invalid start of unicode escape")
+        }
+        EscapeError::OverlongUnicodeEscape => {
+            handler.span_err(span, "overlong unicode escape (must have at most 6 hex digits)")
+        }
+        EscapeError::UnclosedUnicodeEscape => {
+            handler.span_err(span, "unterminated unicode escape (needed a `}`)")
+        }
+        EscapeError::NoBraceInUnicodeEscape => {
+            let msg = "incorrect unicode escape sequence";
+            let mut diag = handler.struct_span_err(span, msg);
+
+            let mut suggestion = "\\u{".to_owned();
+            let mut suggestion_len = 0;
+            let (c, char_span) = last_char();
+            let chars = once(c).chain(lit[range.end..].chars());
+            for c in chars.take(6).take_while(|c| c.is_digit(16)) {
+                suggestion.push(c);
+                suggestion_len += c.len_utf8();
+            }
+
+            if suggestion_len > 0 {
+                suggestion.push('}');
+                let lo = char_span.lo();
+                let hi = lo + BytePos(suggestion_len as u32);
+                diag.span_suggestion(
+                    span.with_lo(lo).with_hi(hi),
+                    "format of unicode escape sequences uses braces",
+                    suggestion,
+                    Applicability::MaybeIncorrect,
+                );
+            } else {
+                diag.span_label(span, msg);
+                diag.help("format of unicode escape sequences is `\\u{...}`");
+            }
+
+            diag.emit();
+        }
+        EscapeError::UnicodeEscapeInByte => handler.span_err(
+            span,
+            "unicode escape sequences cannot be used \
+                                    as a byte or in a byte string",
+        ),
+        EscapeError::EmptyUnicodeEscape => {
+            handler.span_err(span, "empty unicode escape (must have at least 1 hex digit)")
+        }
+        EscapeError::ZeroChars => handler.span_err(span, "empty character literal"),
+        EscapeError::LoneSlash => handler.span_err(span, "invalid trailing slash in literal"),
+    }
+}
+
+/// Pushes a character to a message string for error reporting
+pub(crate) fn push_escaped_char(msg: &mut String, c: char) {
+    match c {
+        '\u{20}'..='\u{7e}' => {
+            // Don't escape \, ' or " for user-facing messages
+            msg.push(c);
+        }
+        _ => {
+            msg.extend(c.escape_default());
+        }
+    }
+}

compiler/rustc_parse/src/lexer/unicode_chars.rs

@@ -0,0 +1,392 @@
+// Characters and their corresponding confusables were collected from
+// http://www.unicode.org/Public/security/10.0.0/confusables.txt
+
+use super::StringReader;
+use crate::token;
+use rustc_errors::{Applicability, DiagnosticBuilder};
+use rustc_span::{symbol::kw, BytePos, Pos, Span};
+
+#[rustfmt::skip] // for line breaks
+const UNICODE_ARRAY: &[(char, &str, char)] = &[
+    ('
', "Line Separator", ' '),
+    ('
', "Paragraph Separator", ' '),
+    (' ', "Ogham Space mark", ' '),
+    (' ', "En Quad", ' '),
+    (' ', "Em Quad", ' '),
+    (' ', "En Space", ' '),
+    (' ', "Em Space", ' '),
+    (' ', "Three-Per-Em Space", ' '),
+    (' ', "Four-Per-Em Space", ' '),
+    (' ', "Six-Per-Em Space", ' '),
+    (' ', "Punctuation Space", ' '),
+    (' ', "Thin Space", ' '),
+    (' ', "Hair Space", ' '),
+    (' ', "Medium Mathematical Space", ' '),
+    (' ', "No-Break Space", ' '),
+    (' ', "Figure Space", ' '),
+    (' ', "Narrow No-Break Space", ' '),
+    ('　', "Ideographic Space", ' '),
+
+    ('ߺ', "Nko Lajanyalan", '_'),
+    ('﹍', "Dashed Low Line", '_'),
+    ('﹎', "Centreline Low Line", '_'),
+    ('﹏', "Wavy Low Line", '_'),
+    ('＿', "Fullwidth Low Line", '_'),
+
+    ('‐', "Hyphen", '-'),
+    ('‑', "Non-Breaking Hyphen", '-'),
+    ('‒', "Figure Dash", '-'),
+    ('–', "En Dash", '-'),
+    ('—', "Em Dash", '-'),
+    ('﹘', "Small Em Dash", '-'),
+    ('۔', "Arabic Full Stop", '-'),
+    ('⁃', "Hyphen Bullet", '-'),
+    ('˗', "Modifier Letter Minus Sign", '-'),
+    ('−', "Minus Sign", '-'),
+    ('➖', "Heavy Minus Sign", '-'),
+    ('Ⲻ', "Coptic Letter Dialect-P Ni", '-'),
+    ('ー', "Katakana-Hiragana Prolonged Sound Mark", '-'),
+    ('－', "Fullwidth Hyphen-Minus", '-'),
+    ('―', "Horizontal Bar", '-'),
+    ('─', "Box Drawings Light Horizontal", '-'),
+    ('━', "Box Drawings Heavy Horizontal", '-'),
+    ('㇐', "CJK Stroke H", '-'),
+    ('ꟷ', "Latin Epigraphic Letter Sideways I", '-'),
+    ('ᅳ', "Hangul Jungseong Eu", '-'),
+    ('ㅡ', "Hangul Letter Eu", '-'),
+    ('一', "CJK Unified Ideograph-4E00", '-'),
+    ('⼀', "Kangxi Radical One", '-'),
+
+    ('؍', "Arabic Date Separator", ','),
+    ('٫', "Arabic Decimal Separator", ','),
+    ('‚', "Single Low-9 Quotation Mark", ','),
+    ('¸', "Cedilla", ','),
+    ('ꓹ', "Lisu Letter Tone Na Po", ','),
+    ('，', "Fullwidth Comma", ','),
+
+    (';', "Greek Question Mark", ';'),
+    ('；', "Fullwidth Semicolon", ';'),
+    ('︔', "Presentation Form For Vertical Semicolon", ';'),
+
+    ('ः', "Devanagari Sign Visarga", ':'),
+    ('ઃ', "Gujarati Sign Visarga", ':'),
+    ('：', "Fullwidth Colon", ':'),
+    ('։', "Armenian Full Stop", ':'),
+    ('܃', "Syriac Supralinear Colon", ':'),
+    ('܄', "Syriac Sublinear Colon", ':'),
+    ('᛬', "Runic Multiple Punctuation", ':'),
+    ('︰', "Presentation Form For Vertical Two Dot Leader", ':'),
+    ('᠃', "Mongolian Full Stop", ':'),
+    ('᠉', "Mongolian Manchu Full Stop", ':'),
+    ('⁚', "Two Dot Punctuation", ':'),
+    ('׃', "Hebrew Punctuation Sof Pasuq", ':'),
+    ('˸', "Modifier Letter Raised Colon", ':'),
+    ('꞉', "Modifier Letter Colon", ':'),
+    ('∶', "Ratio", ':'),
+    ('ː', "Modifier Letter Triangular Colon", ':'),
+    ('ꓽ', "Lisu Letter Tone Mya Jeu", ':'),
+    ('︓', "Presentation Form For Vertical Colon", ':'),
+
+    ('！', "Fullwidth Exclamation Mark", '!'),
+    ('ǃ', "Latin Letter Retroflex Click", '!'),
+    ('ⵑ', "Tifinagh Letter Tuareg Yang", '!'),
+    ('︕', "Presentation Form For Vertical Exclamation Mark", '!'),
+
+    ('ʔ', "Latin Letter Glottal Stop", '?'),
+    ('Ɂ', "Latin Capital Letter Glottal Stop", '?'),
+    ('ॽ', "Devanagari Letter Glottal Stop", '?'),
+    ('Ꭾ', "Cherokee Letter He", '?'),
+    ('ꛫ', "Bamum Letter Ntuu", '?'),
+    ('？', "Fullwidth Question Mark", '?'),
+    ('︖', "Presentation Form For Vertical Question Mark", '?'),
+
+    ('𝅭', "Musical Symbol Combining Augmentation Dot", '.'),
+    ('․', "One Dot Leader", '.'),
+    ('܁', "Syriac Supralinear Full Stop", '.'),
+    ('܂', "Syriac Sublinear Full Stop", '.'),
+    ('꘎', "Vai Full Stop", '.'),
+    ('𐩐', "Kharoshthi Punctuation Dot", '.'),
+    ('٠', "Arabic-Indic Digit Zero", '.'),
+    ('۰', "Extended Arabic-Indic Digit Zero", '.'),
+    ('ꓸ', "Lisu Letter Tone Mya Ti", '.'),
+    ('·', "Middle Dot", '.'),
+    ('・', "Katakana Middle Dot", '.'),
+    ('･', "Halfwidth Katakana Middle Dot", '.'),
+    ('᛫', "Runic Single Punctuation", '.'),
+    ('·', "Greek Ano Teleia", '.'),
+    ('⸱', "Word Separator Middle Dot", '.'),
+    ('𐄁', "Aegean Word Separator Dot", '.'),
+    ('•', "Bullet", '.'),
+    ('‧', "Hyphenation Point", '.'),
+    ('∙', "Bullet Operator", '.'),
+    ('⋅', "Dot Operator", '.'),
+    ('ꞏ', "Latin Letter Sinological Dot", '.'),
+    ('ᐧ', "Canadian Syllabics Final Middle Dot", '.'),
+    ('ᐧ', "Canadian Syllabics Final Middle Dot", '.'),
+    ('．', "Fullwidth Full Stop", '.'),
+    ('。', "Ideographic Full Stop", '.'),
+    ('︒', "Presentation Form For Vertical Ideographic Full Stop", '.'),
+
+    ('՝', "Armenian Comma", '\''),
+    ('＇', "Fullwidth Apostrophe", '\''),
+    ('‘', "Left Single Quotation Mark", '\''),
+    ('’', "Right Single Quotation Mark", '\''),
+    ('‛', "Single High-Reversed-9 Quotation Mark", '\''),
+    ('′', "Prime", '\''),
+    ('‵', "Reversed Prime", '\''),
+    ('՚', "Armenian Apostrophe", '\''),
+    ('׳', "Hebrew Punctuation Geresh", '\''),
+    ('`', "Grave Accent", '\''),
+    ('`', "Greek Varia", '\''),
+    ('｀', "Fullwidth Grave Accent", '\''),
+    ('´', "Acute Accent", '\''),
+    ('΄', "Greek Tonos", '\''),
+    ('´', "Greek Oxia", '\''),
+    ('᾽', "Greek Koronis", '\''),
+    ('᾿', "Greek Psili", '\''),
+    ('῾', "Greek Dasia", '\''),
+    ('ʹ', "Modifier Letter Prime", '\''),
+    ('ʹ', "Greek Numeral Sign", '\''),
+    ('ˈ', "Modifier Letter Vertical Line", '\''),
+    ('ˊ', "Modifier Letter Acute Accent", '\''),
+    ('ˋ', "Modifier Letter Grave Accent", '\''),
+    ('˴', "Modifier Letter Middle Grave Accent", '\''),
+    ('ʻ', "Modifier Letter Turned Comma", '\''),
+    ('ʽ', "Modifier Letter Reversed Comma", '\''),
+    ('ʼ', "Modifier Letter Apostrophe", '\''),
+    ('ʾ', "Modifier Letter Right Half Ring", '\''),
+    ('ꞌ', "Latin Small Letter Saltillo", '\''),
+    ('י', "Hebrew Letter Yod", '\''),
+    ('ߴ', "Nko High Tone Apostrophe", '\''),
+    ('ߵ', "Nko Low Tone Apostrophe", '\''),
+    ('ᑊ', "Canadian Syllabics West-Cree P", '\''),
+    ('ᛌ', "Runic Letter Short-Twig-Sol S", '\''),
+    ('𖽑', "Miao Sign Aspiration", '\''),
+    ('𖽒', "Miao Sign Reformed Voicing", '\''),
+
+    ('᳓', "Vedic Sign Nihshvasa", '"'),
+    ('＂', "Fullwidth Quotation Mark", '"'),
+    ('“', "Left Double Quotation Mark", '"'),
+    ('”', "Right Double Quotation Mark", '"'),
+    ('‟', "Double High-Reversed-9 Quotation Mark", '"'),
+    ('″', "Double Prime", '"'),
+    ('‶', "Reversed Double Prime", '"'),
+    ('〃', "Ditto Mark", '"'),
+    ('״', "Hebrew Punctuation Gershayim", '"'),
+    ('˝', "Double Acute Accent", '"'),
+    ('ʺ', "Modifier Letter Double Prime", '"'),
+    ('˶', "Modifier Letter Middle Double Acute Accent", '"'),
+    ('˵', "Modifier Letter Middle Double Grave Accent", '"'),
+    ('ˮ', "Modifier Letter Double Apostrophe", '"'),
+    ('ײ', "Hebrew Ligature Yiddish Double Yod", '"'),
+    ('❞', "Heavy Double Comma Quotation Mark Ornament", '"'),
+    ('❝', "Heavy Double Turned Comma Quotation Mark Ornament", '"'),
+
+    ('（', "Fullwidth Left Parenthesis", '('),
+    ('❨', "Medium Left Parenthesis Ornament", '('),
+    ('﴾', "Ornate Left Parenthesis", '('),
+
+    ('）', "Fullwidth Right Parenthesis", ')'),
+    ('❩', "Medium Right Parenthesis Ornament", ')'),
+    ('﴿', "Ornate Right Parenthesis", ')'),
+
+    ('［', "Fullwidth Left Square Bracket", '['),
+    ('❲', "Light Left Tortoise Shell Bracket Ornament", '['),
+    ('「', "Left Corner Bracket", '['),
+    ('『', "Left White Corner Bracket", '['),
+    ('【', "Left Black Lenticular Bracket", '['),
+    ('〔', "Left Tortoise Shell Bracket", '['),
+    ('〖', "Left White Lenticular Bracket", '['),
+    ('〘', "Left White Tortoise Shell Bracket", '['),
+    ('〚', "Left White Square Bracket", '['),
+
+    ('］', "Fullwidth Right Square Bracket", ']'),
+    ('❳', "Light Right Tortoise Shell Bracket Ornament", ']'),
+    ('」', "Right Corner Bracket", ']'),
+    ('』', "Right White Corner Bracket", ']'),
+    ('】', "Right Black Lenticular Bracket", ']'),
+    ('〕', "Right Tortoise Shell Bracket", ']'),
+    ('〗', "Right White Lenticular Bracket", ']'),
+    ('〙', "Right White Tortoise Shell Bracket", ']'),
+    ('〛', "Right White Square Bracket", ']'),
+
+    ('❴', "Medium Left Curly Bracket Ornament", '{'),
+    ('𝄔', "Musical Symbol Brace", '{'),
+    ('｛', "Fullwidth Left Curly Bracket", '{'),
+
+    ('❵', "Medium Right Curly Bracket Ornament", '}'),
+    ('｝', "Fullwidth Right Curly Bracket", '}'),
+
+    ('⁎', "Low Asterisk", '*'),
+    ('٭', "Arabic Five Pointed Star", '*'),
+    ('∗', "Asterisk Operator", '*'),
+    ('𐌟', "Old Italic Letter Ess", '*'),
+    ('＊', "Fullwidth Asterisk", '*'),
+
+    ('᜵', "Philippine Single Punctuation", '/'),
+    ('⁁', "Caret Insertion Point", '/'),
+    ('∕', "Division Slash", '/'),
+    ('⁄', "Fraction Slash", '/'),
+    ('╱', "Box Drawings Light Diagonal Upper Right To Lower Left", '/'),
+    ('⟋', "Mathematical Rising Diagonal", '/'),
+    ('⧸', "Big Solidus", '/'),
+    ('𝈺', "Greek Instrumental Notation Symbol-47", '/'),
+    ('㇓', "CJK Stroke Sp", '/'),
+    ('〳', "Vertical Kana Repeat Mark Upper Half", '/'),
+    ('Ⳇ', "Coptic Capital Letter Old Coptic Esh", '/'),
+    ('ノ', "Katakana Letter No", '/'),
+    ('丿', "CJK Unified Ideograph-4E3F", '/'),
+    ('⼃', "Kangxi Radical Slash", '/'),
+    ('／', "Fullwidth Solidus", '/'),
+
+    ('＼', "Fullwidth Reverse Solidus", '\\'),
+    ('﹨', "Small Reverse Solidus", '\\'),
+    ('∖', "Set Minus", '\\'),
+    ('⟍', "Mathematical Falling Diagonal", '\\'),
+    ('⧵', "Reverse Solidus Operator", '\\'),
+    ('⧹', "Big Reverse Solidus", '\\'),
+    ('⧹', "Greek Vocal Notation Symbol-16", '\\'),
+    ('⧹', "Greek Instrumental Symbol-48", '\\'),
+    ('㇔', "CJK Stroke D", '\\'),
+    ('丶', "CJK Unified Ideograph-4E36", '\\'),
+    ('⼂', "Kangxi Radical Dot", '\\'),
+    ('、', "Ideographic Comma", '\\'),
+    ('ヽ', "Katakana Iteration Mark", '\\'),
+
+    ('ꝸ', "Latin Small Letter Um", '&'),
+    ('＆', "Fullwidth Ampersand", '&'),
+
+    ('᛭', "Runic Cross Punctuation", '+'),
+    ('➕', "Heavy Plus Sign", '+'),
+    ('𐊛', "Lycian Letter H", '+'),
+    ('﬩', "Hebrew Letter Alternative Plus Sign", '+'),
+    ('＋', "Fullwidth Plus Sign", '+'),
+
+    ('‹', "Single Left-Pointing Angle Quotation Mark", '<'),
+    ('❮', "Heavy Left-Pointing Angle Quotation Mark Ornament", '<'),
+    ('˂', "Modifier Letter Left Arrowhead", '<'),
+    ('𝈶', "Greek Instrumental Symbol-40", '<'),
+    ('ᐸ', "Canadian Syllabics Pa", '<'),
+    ('ᚲ', "Runic Letter Kauna", '<'),
+    ('❬', "Medium Left-Pointing Angle Bracket Ornament", '<'),
+    ('⟨', "Mathematical Left Angle Bracket", '<'),
+    ('〈', "Left-Pointing Angle Bracket", '<'),
+    ('〈', "Left Angle Bracket", '<'),
+    ('㇛', "CJK Stroke Pd", '<'),
+    ('く', "Hiragana Letter Ku", '<'),
+    ('𡿨', "CJK Unified Ideograph-21FE8", '<'),
+    ('《', "Left Double Angle Bracket", '<'),
+    ('＜', "Fullwidth Less-Than Sign", '<'),
+
+    ('᐀', "Canadian Syllabics Hyphen", '='),
+    ('⹀', "Double Hyphen", '='),
+    ('゠', "Katakana-Hiragana Double Hyphen", '='),
+    ('꓿', "Lisu Punctuation Full Stop", '='),
+    ('＝', "Fullwidth Equals Sign", '='),
+
+    ('›', "Single Right-Pointing Angle Quotation Mark", '>'),
+    ('❯', "Heavy Right-Pointing Angle Quotation Mark Ornament", '>'),
+    ('˃', "Modifier Letter Right Arrowhead", '>'),
+    ('𝈷', "Greek Instrumental Symbol-42", '>'),
+    ('ᐳ', "Canadian Syllabics Po", '>'),
+    ('𖼿', "Miao Letter Archaic Zza", '>'),
+    ('❭', "Medium Right-Pointing Angle Bracket Ornament", '>'),
+    ('⟩', "Mathematical Right Angle Bracket", '>'),
+    ('〉', "Right-Pointing Angle Bracket", '>'),
+    ('〉', "Right Angle Bracket", '>'),
+    ('》', "Right Double Angle Bracket", '>'),
+    ('＞', "Fullwidth Greater-Than Sign", '>'),
+];
+
+// FIXME: the lexer could be used to turn the ASCII version of unicode homoglyphs, instead of
+// keeping the substitution token in this table. Ideally, this should be inside `rustc_lexer`.
+// However, we should first remove compound tokens like `<<` from `rustc_lexer`, and then add
+// fancier error recovery to it, as there will be less overall work to do this way.
+const ASCII_ARRAY: &[(char, &str, Option<token::TokenKind>)] = &[
+    (' ', "Space", Some(token::Whitespace)),
+    ('_', "Underscore", Some(token::Ident(kw::Underscore, false))),
+    ('-', "Minus/Hyphen", Some(token::BinOp(token::Minus))),
+    (',', "Comma", Some(token::Comma)),
+    (';', "Semicolon", Some(token::Semi)),
+    (':', "Colon", Some(token::Colon)),
+    ('!', "Exclamation Mark", Some(token::Not)),
+    ('?', "Question Mark", Some(token::Question)),
+    ('.', "Period", Some(token::Dot)),
+    ('(', "Left Parenthesis", Some(token::OpenDelim(token::Paren))),
+    (')', "Right Parenthesis", Some(token::CloseDelim(token::Paren))),
+    ('[', "Left Square Bracket", Some(token::OpenDelim(token::Bracket))),
+    (']', "Right Square Bracket", Some(token::CloseDelim(token::Bracket))),
+    ('{', "Left Curly Brace", Some(token::OpenDelim(token::Brace))),
+    ('}', "Right Curly Brace", Some(token::CloseDelim(token::Brace))),
+    ('*', "Asterisk", Some(token::BinOp(token::Star))),
+    ('/', "Slash", Some(token::BinOp(token::Slash))),
+    ('\\', "Backslash", None),
+    ('&', "Ampersand", Some(token::BinOp(token::And))),
+    ('+', "Plus Sign", Some(token::BinOp(token::Plus))),
+    ('<', "Less-Than Sign", Some(token::Lt)),
+    ('=', "Equals Sign", Some(token::Eq)),
+    ('>', "Greater-Than Sign", Some(token::Gt)),
+    // FIXME: Literals are already lexed by this point, so we can't recover gracefully just by
+    // spitting the correct token out.
+    ('\'', "Single Quote", None),
+    ('"', "Quotation Mark", None),
+];
+
+crate fn check_for_substitution<'a>(
+    reader: &StringReader<'a>,
+    pos: BytePos,
+    ch: char,
+    err: &mut DiagnosticBuilder<'a>,
+) -> Option<token::TokenKind> {
+    let (u_name, ascii_char) = match UNICODE_ARRAY.iter().find(|&&(c, _, _)| c == ch) {
+        Some(&(_u_char, u_name, ascii_char)) => (u_name, ascii_char),
+        None => return None,
+    };
+
+    let span = Span::with_root_ctxt(pos, pos + Pos::from_usize(ch.len_utf8()));
+
+    let (ascii_name, token) = match ASCII_ARRAY.iter().find(|&&(c, _, _)| c == ascii_char) {
+        Some((_ascii_char, ascii_name, token)) => (ascii_name, token),
+        None => {
+            let msg = format!("substitution character not found for '{}'", ch);
+            reader.sess.span_diagnostic.span_bug_no_panic(span, &msg);
+            return None;
+        }
+    };
+
+    // special help suggestion for "directed" double quotes
+    if let Some(s) = peek_delimited(&reader.src[reader.src_index(pos)..], '“', '”') {
+        let msg = format!(
+            "Unicode characters '“' (Left Double Quotation Mark) and \
+             '”' (Right Double Quotation Mark) look like '{}' ({}), but are not",
+            ascii_char, ascii_name
+        );
+        err.span_suggestion(
+            Span::with_root_ctxt(
+                pos,
+                pos + Pos::from_usize('“'.len_utf8() + s.len() + '”'.len_utf8()),
+            ),
+            &msg,
+            format!("\"{}\"", s),
+            Applicability::MaybeIncorrect,
+        );
+    } else {
+        let msg = format!(
+            "Unicode character '{}' ({}) looks like '{}' ({}), but it is not",
+            ch, u_name, ascii_char, ascii_name
+        );
+        err.span_suggestion(span, &msg, ascii_char.to_string(), Applicability::MaybeIncorrect);
+    }
+    token.clone()
+}
+
+/// Extract string if found at current position with given delimiters
+fn peek_delimited(text: &str, from_ch: char, to_ch: char) -> Option<&str> {
+    let mut chars = text.chars();
+    let first_char = chars.next()?;
+    if first_char != from_ch {
+        return None;
+    }
+    let last_char_idx = chars.as_str().find(to_ch)?;
+    Some(&chars.as_str()[..last_char_idx])
+}

compiler/rustc_parse/src/lib.rs

@@ -0,0 +1,595 @@
+//! The main parser interface.
+
+#![feature(bool_to_option)]
+#![feature(crate_visibility_modifier)]
+#![feature(bindings_after_at)]
+#![feature(try_blocks)]
+#![feature(or_patterns)]
+
+use rustc_ast as ast;
+use rustc_ast::token::{self, DelimToken, Nonterminal, Token, TokenKind};
+use rustc_ast::tokenstream::{self, IsJoint, TokenStream, TokenTree};
+use rustc_ast_pretty::pprust;
+use rustc_data_structures::sync::Lrc;
+use rustc_errors::{Diagnostic, FatalError, Level, PResult};
+use rustc_session::parse::ParseSess;
+use rustc_span::{symbol::kw, FileName, SourceFile, Span, DUMMY_SP};
+
+use smallvec::SmallVec;
+use std::mem;
+use std::path::Path;
+use std::str;
+
+use tracing::{debug, info};
+
+pub const MACRO_ARGUMENTS: Option<&'static str> = Some("macro arguments");
+
+#[macro_use]
+pub mod parser;
+use parser::{emit_unclosed_delims, make_unclosed_delims_error, Parser};
+pub mod lexer;
+pub mod validate_attr;
+
+// A bunch of utility functions of the form `parse_<thing>_from_<source>`
+// where <thing> includes crate, expr, item, stmt, tts, and one that
+// uses a HOF to parse anything, and <source> includes file and
+// `source_str`.
+
+/// A variant of 'panictry!' that works on a Vec<Diagnostic> instead of a single DiagnosticBuilder.
+macro_rules! panictry_buffer {
+    ($handler:expr, $e:expr) => {{
+        use rustc_errors::FatalError;
+        use std::result::Result::{Err, Ok};
+        match $e {
+            Ok(e) => e,
+            Err(errs) => {
+                for e in errs {
+                    $handler.emit_diagnostic(&e);
+                }
+                FatalError.raise()
+            }
+        }
+    }};
+}
+
+pub fn parse_crate_from_file<'a>(input: &Path, sess: &'a ParseSess) -> PResult<'a, ast::Crate> {
+    let mut parser = new_parser_from_file(sess, input, None);
+    parser.parse_crate_mod()
+}
+
+pub fn parse_crate_attrs_from_file<'a>(
+    input: &Path,
+    sess: &'a ParseSess,
+) -> PResult<'a, Vec<ast::Attribute>> {
+    let mut parser = new_parser_from_file(sess, input, None);
+    parser.parse_inner_attributes()
+}
+
+pub fn parse_crate_from_source_str(
+    name: FileName,
+    source: String,
+    sess: &ParseSess,
+) -> PResult<'_, ast::Crate> {
+    new_parser_from_source_str(sess, name, source).parse_crate_mod()
+}
+
+pub fn parse_crate_attrs_from_source_str(
+    name: FileName,
+    source: String,
+    sess: &ParseSess,
+) -> PResult<'_, Vec<ast::Attribute>> {
+    new_parser_from_source_str(sess, name, source).parse_inner_attributes()
+}
+
+pub fn parse_stream_from_source_str(
+    name: FileName,
+    source: String,
+    sess: &ParseSess,
+    override_span: Option<Span>,
+) -> TokenStream {
+    let (stream, mut errors) =
+        source_file_to_stream(sess, sess.source_map().new_source_file(name, source), override_span);
+    emit_unclosed_delims(&mut errors, &sess);
+    stream
+}
+
+/// Creates a new parser from a source string.
+pub fn new_parser_from_source_str(sess: &ParseSess, name: FileName, source: String) -> Parser<'_> {
+    panictry_buffer!(&sess.span_diagnostic, maybe_new_parser_from_source_str(sess, name, source))
+}
+
+/// Creates a new parser from a source string. Returns any buffered errors from lexing the initial
+/// token stream.
+pub fn maybe_new_parser_from_source_str(
+    sess: &ParseSess,
+    name: FileName,
+    source: String,
+) -> Result<Parser<'_>, Vec<Diagnostic>> {
+    maybe_source_file_to_parser(sess, sess.source_map().new_source_file(name, source))
+}
+
+/// Creates a new parser, handling errors as appropriate if the file doesn't exist.
+/// If a span is given, that is used on an error as the as the source of the problem.
+pub fn new_parser_from_file<'a>(sess: &'a ParseSess, path: &Path, sp: Option<Span>) -> Parser<'a> {
+    source_file_to_parser(sess, file_to_source_file(sess, path, sp))
+}
+
+/// Creates a new parser, returning buffered diagnostics if the file doesn't exist,
+/// or from lexing the initial token stream.
+pub fn maybe_new_parser_from_file<'a>(
+    sess: &'a ParseSess,
+    path: &Path,
+) -> Result<Parser<'a>, Vec<Diagnostic>> {
+    let file = try_file_to_source_file(sess, path, None).map_err(|db| vec![db])?;
+    maybe_source_file_to_parser(sess, file)
+}
+
+/// Given a `source_file` and config, returns a parser.
+fn source_file_to_parser(sess: &ParseSess, source_file: Lrc<SourceFile>) -> Parser<'_> {
+    panictry_buffer!(&sess.span_diagnostic, maybe_source_file_to_parser(sess, source_file))
+}
+
+/// Given a `source_file` and config, return a parser. Returns any buffered errors from lexing the
+/// initial token stream.
+fn maybe_source_file_to_parser(
+    sess: &ParseSess,
+    source_file: Lrc<SourceFile>,
+) -> Result<Parser<'_>, Vec<Diagnostic>> {
+    let end_pos = source_file.end_pos;
+    let (stream, unclosed_delims) = maybe_file_to_stream(sess, source_file, None)?;
+    let mut parser = stream_to_parser(sess, stream, None);
+    parser.unclosed_delims = unclosed_delims;
+    if parser.token == token::Eof {
+        parser.token.span = Span::new(end_pos, end_pos, parser.token.span.ctxt());
+    }
+
+    Ok(parser)
+}
+
+// Must preserve old name for now, because `quote!` from the *existing*
+// compiler expands into it.
+pub fn new_parser_from_tts(sess: &ParseSess, tts: Vec<TokenTree>) -> Parser<'_> {
+    stream_to_parser(sess, tts.into_iter().collect(), crate::MACRO_ARGUMENTS)
+}
+
+// Base abstractions
+
+/// Given a session and a path and an optional span (for error reporting),
+/// add the path to the session's source_map and return the new source_file or
+/// error when a file can't be read.
+fn try_file_to_source_file(
+    sess: &ParseSess,
+    path: &Path,
+    spanopt: Option<Span>,
+) -> Result<Lrc<SourceFile>, Diagnostic> {
+    sess.source_map().load_file(path).map_err(|e| {
+        let msg = format!("couldn't read {}: {}", path.display(), e);
+        let mut diag = Diagnostic::new(Level::Fatal, &msg);
+        if let Some(sp) = spanopt {
+            diag.set_span(sp);
+        }
+        diag
+    })
+}
+
+/// Given a session and a path and an optional span (for error reporting),
+/// adds the path to the session's `source_map` and returns the new `source_file`.
+fn file_to_source_file(sess: &ParseSess, path: &Path, spanopt: Option<Span>) -> Lrc<SourceFile> {
+    match try_file_to_source_file(sess, path, spanopt) {
+        Ok(source_file) => source_file,
+        Err(d) => {
+            sess.span_diagnostic.emit_diagnostic(&d);
+            FatalError.raise();
+        }
+    }
+}
+
+/// Given a `source_file`, produces a sequence of token trees.
+pub fn source_file_to_stream(
+    sess: &ParseSess,
+    source_file: Lrc<SourceFile>,
+    override_span: Option<Span>,
+) -> (TokenStream, Vec<lexer::UnmatchedBrace>) {
+    panictry_buffer!(&sess.span_diagnostic, maybe_file_to_stream(sess, source_file, override_span))
+}
+
+/// Given a source file, produces a sequence of token trees. Returns any buffered errors from
+/// parsing the token stream.
+pub fn maybe_file_to_stream(
+    sess: &ParseSess,
+    source_file: Lrc<SourceFile>,
+    override_span: Option<Span>,
+) -> Result<(TokenStream, Vec<lexer::UnmatchedBrace>), Vec<Diagnostic>> {
+    let srdr = lexer::StringReader::new(sess, source_file, override_span);
+    let (token_trees, unmatched_braces) = srdr.into_token_trees();
+
+    match token_trees {
+        Ok(stream) => Ok((stream, unmatched_braces)),
+        Err(err) => {
+            let mut buffer = Vec::with_capacity(1);
+            err.buffer(&mut buffer);
+            // Not using `emit_unclosed_delims` to use `db.buffer`
+            for unmatched in unmatched_braces {
+                if let Some(err) = make_unclosed_delims_error(unmatched, &sess) {
+                    err.buffer(&mut buffer);
+                }
+            }
+            Err(buffer)
+        }
+    }
+}
+
+/// Given a stream and the `ParseSess`, produces a parser.
+pub fn stream_to_parser<'a>(
+    sess: &'a ParseSess,
+    stream: TokenStream,
+    subparser_name: Option<&'static str>,
+) -> Parser<'a> {
+    Parser::new(sess, stream, false, subparser_name)
+}
+
+/// Runs the given subparser `f` on the tokens of the given `attr`'s item.
+pub fn parse_in<'a, T>(
+    sess: &'a ParseSess,
+    tts: TokenStream,
+    name: &'static str,
+    mut f: impl FnMut(&mut Parser<'a>) -> PResult<'a, T>,
+) -> PResult<'a, T> {
+    let mut parser = Parser::new(sess, tts, false, Some(name));
+    let result = f(&mut parser)?;
+    if parser.token != token::Eof {
+        parser.unexpected()?;
+    }
+    Ok(result)
+}
+
+// NOTE(Centril): The following probably shouldn't be here but it acknowledges the
+// fact that architecturally, we are using parsing (read on below to understand why).
+
+pub fn nt_to_tokenstream(nt: &Nonterminal, sess: &ParseSess, span: Span) -> TokenStream {
+    // A `Nonterminal` is often a parsed AST item. At this point we now
+    // need to convert the parsed AST to an actual token stream, e.g.
+    // un-parse it basically.
+    //
+    // Unfortunately there's not really a great way to do that in a
+    // guaranteed lossless fashion right now. The fallback here is to just
+    // stringify the AST node and reparse it, but this loses all span
+    // information.
+    //
+    // As a result, some AST nodes are annotated with the token stream they
+    // came from. Here we attempt to extract these lossless token streams
+    // before we fall back to the stringification.
+    let tokens = match *nt {
+        Nonterminal::NtItem(ref item) => {
+            prepend_attrs(sess, &item.attrs, item.tokens.as_ref(), span)
+        }
+        Nonterminal::NtPat(ref pat) => pat.tokens.clone(),
+        Nonterminal::NtIdent(ident, is_raw) => {
+            Some(tokenstream::TokenTree::token(token::Ident(ident.name, is_raw), ident.span).into())
+        }
+        Nonterminal::NtLifetime(ident) => {
+            Some(tokenstream::TokenTree::token(token::Lifetime(ident.name), ident.span).into())
+        }
+        Nonterminal::NtTT(ref tt) => Some(tt.clone().into()),
+        Nonterminal::NtExpr(ref expr) => {
+            if expr.tokens.is_none() {
+                debug!("missing tokens for expr {:?}", expr);
+            }
+            prepend_attrs(sess, &expr.attrs, expr.tokens.as_ref(), span)
+        }
+        _ => None,
+    };
+
+    // FIXME(#43081): Avoid this pretty-print + reparse hack
+    let source = pprust::nonterminal_to_string(nt);
+    let filename = FileName::macro_expansion_source_code(&source);
+    let tokens_for_real = parse_stream_from_source_str(filename, source, sess, Some(span));
+
+    // During early phases of the compiler the AST could get modified
+    // directly (e.g., attributes added or removed) and the internal cache
+    // of tokens my not be invalidated or updated. Consequently if the
+    // "lossless" token stream disagrees with our actual stringification
+    // (which has historically been much more battle-tested) then we go
+    // with the lossy stream anyway (losing span information).
+    //
+    // Note that the comparison isn't `==` here to avoid comparing spans,
+    // but it *also* is a "probable" equality which is a pretty weird
+    // definition. We mostly want to catch actual changes to the AST
+    // like a `#[cfg]` being processed or some weird `macro_rules!`
+    // expansion.
+    //
+    // What we *don't* want to catch is the fact that a user-defined
+    // literal like `0xf` is stringified as `15`, causing the cached token
+    // stream to not be literal `==` token-wise (ignoring spans) to the
+    // token stream we got from stringification.
+    //
+    // Instead the "probably equal" check here is "does each token
+    // recursively have the same discriminant?" We basically don't look at
+    // the token values here and assume that such fine grained token stream
+    // modifications, including adding/removing typically non-semantic
+    // tokens such as extra braces and commas, don't happen.
+    if let Some(tokens) = tokens {
+        if tokenstream_probably_equal_for_proc_macro(&tokens, &tokens_for_real, sess) {
+            return tokens;
+        }
+        info!(
+            "cached tokens found, but they're not \"probably equal\", \
+                going with stringified version"
+        );
+        info!("cached tokens: {:?}", tokens);
+        info!("reparsed tokens: {:?}", tokens_for_real);
+    }
+    tokens_for_real
+}
+
+// See comments in `Nonterminal::to_tokenstream` for why we care about
+// *probably* equal here rather than actual equality
+//
+// This is otherwise the same as `eq_unspanned`, only recursing with a
+// different method.
+pub fn tokenstream_probably_equal_for_proc_macro(
+    first: &TokenStream,
+    other: &TokenStream,
+    sess: &ParseSess,
+) -> bool {
+    // When checking for `probably_eq`, we ignore certain tokens that aren't
+    // preserved in the AST. Because they are not preserved, the pretty
+    // printer arbitrarily adds or removes them when printing as token
+    // streams, making a comparison between a token stream generated from an
+    // AST and a token stream which was parsed into an AST more reliable.
+    fn semantic_tree(tree: &TokenTree) -> bool {
+        if let TokenTree::Token(token) = tree {
+            if let
+                // The pretty printer tends to add trailing commas to
+                // everything, and in particular, after struct fields.
+                | token::Comma
+                // The pretty printer emits `NoDelim` as whitespace.
+                | token::OpenDelim(DelimToken::NoDelim)
+                | token::CloseDelim(DelimToken::NoDelim)
+                // The pretty printer collapses many semicolons into one.
+                | token::Semi
+                // The pretty printer collapses whitespace arbitrarily and can
+                // introduce whitespace from `NoDelim`.
+                | token::Whitespace
+                // The pretty printer can turn `$crate` into `::crate_name`
+                | token::ModSep = token.kind {
+                return false;
+            }
+        }
+        true
+    }
+
+    // When comparing two `TokenStream`s, we ignore the `IsJoint` information.
+    //
+    // However, `rustc_parse::lexer::tokentrees::TokenStreamBuilder` will
+    // use `Token.glue` on adjacent tokens with the proper `IsJoint`.
+    // Since we are ignoreing `IsJoint`, a 'glued' token (e.g. `BinOp(Shr)`)
+    // and its 'split'/'unglued' compoenents (e.g. `Gt, Gt`) are equivalent
+    // when determining if two `TokenStream`s are 'probably equal'.
+    //
+    // Therefore, we use `break_two_token_op` to convert all tokens
+    // to the 'unglued' form (if it exists). This ensures that two
+    // `TokenStream`s which differ only in how their tokens are glued
+    // will be considered 'probably equal', which allows us to keep spans.
+    //
+    // This is important when the original `TokenStream` contained
+    // extra spaces (e.g. `f :: < Vec < _ > > ( ) ;'). These extra spaces
+    // will be omitted when we pretty-print, which can cause the original
+    // and reparsed `TokenStream`s to differ in the assignment of `IsJoint`,
+    // leading to some tokens being 'glued' together in one stream but not
+    // the other. See #68489 for more details.
+    fn break_tokens(tree: TokenTree) -> impl Iterator<Item = TokenTree> {
+        // In almost all cases, we should have either zero or one levels
+        // of 'unglueing'. However, in some unusual cases, we may need
+        // to iterate breaking tokens mutliple times. For example:
+        // '[BinOpEq(Shr)] => [Gt, Ge] -> [Gt, Gt, Eq]'
+        let mut token_trees: SmallVec<[_; 2]>;
+        if let TokenTree::Token(token) = &tree {
+            let mut out = SmallVec::<[_; 2]>::new();
+            out.push(token.clone());
+            // Iterate to fixpoint:
+            // * We start off with 'out' containing our initial token, and `temp` empty
+            // * If we are able to break any tokens in `out`, then `out` will have
+            //   at least one more element than 'temp', so we will try to break tokens
+            //   again.
+            // * If we cannot break any tokens in 'out', we are done
+            loop {
+                let mut temp = SmallVec::<[_; 2]>::new();
+                let mut changed = false;
+
+                for token in out.into_iter() {
+                    if let Some((first, second)) = token.kind.break_two_token_op() {
+                        temp.push(Token::new(first, DUMMY_SP));
+                        temp.push(Token::new(second, DUMMY_SP));
+                        changed = true;
+                    } else {
+                        temp.push(token);
+                    }
+                }
+                out = temp;
+                if !changed {
+                    break;
+                }
+            }
+            token_trees = out.into_iter().map(TokenTree::Token).collect();
+        } else {
+            token_trees = SmallVec::new();
+            token_trees.push(tree);
+        }
+        token_trees.into_iter()
+    }
+
+    let expand_nt = |tree: TokenTree| {
+        if let TokenTree::Token(Token { kind: TokenKind::Interpolated(nt), span }) = &tree {
+            // When checking tokenstreams for 'probable equality', we are comparing
+            // a captured (from parsing) `TokenStream` to a reparsed tokenstream.
+            // The reparsed Tokenstream will never have `None`-delimited groups,
+            // since they are only ever inserted as a result of macro expansion.
+            // Therefore, inserting a `None`-delimtied group here (when we
+            // convert a nested `Nonterminal` to a tokenstream) would cause
+            // a mismatch with the reparsed tokenstream.
+            //
+            // Note that we currently do not handle the case where the
+            // reparsed stream has a `Parenthesis`-delimited group
+            // inserted. This will cause a spurious mismatch:
+            // issue #75734 tracks resolving this.
+            nt_to_tokenstream(nt, sess, *span).into_trees()
+        } else {
+            TokenStream::new(vec![(tree, IsJoint::NonJoint)]).into_trees()
+        }
+    };
+
+    // Break tokens after we expand any nonterminals, so that we break tokens
+    // that are produced as a result of nonterminal expansion.
+    let mut t1 = first.trees().filter(semantic_tree).flat_map(expand_nt).flat_map(break_tokens);
+    let mut t2 = other.trees().filter(semantic_tree).flat_map(expand_nt).flat_map(break_tokens);
+    for (t1, t2) in t1.by_ref().zip(t2.by_ref()) {
+        if !tokentree_probably_equal_for_proc_macro(&t1, &t2, sess) {
+            return false;
+        }
+    }
+    t1.next().is_none() && t2.next().is_none()
+}
+
+// See comments in `Nonterminal::to_tokenstream` for why we care about
+// *probably* equal here rather than actual equality
+//
+// This is otherwise the same as `eq_unspanned`, only recursing with a
+// different method.
+pub fn tokentree_probably_equal_for_proc_macro(
+    first: &TokenTree,
+    other: &TokenTree,
+    sess: &ParseSess,
+) -> bool {
+    match (first, other) {
+        (TokenTree::Token(token), TokenTree::Token(token2)) => {
+            token_probably_equal_for_proc_macro(token, token2)
+        }
+        (TokenTree::Delimited(_, delim, tts), TokenTree::Delimited(_, delim2, tts2)) => {
+            delim == delim2 && tokenstream_probably_equal_for_proc_macro(&tts, &tts2, sess)
+        }
+        _ => false,
+    }
+}
+
+// See comments in `Nonterminal::to_tokenstream` for why we care about
+// *probably* equal here rather than actual equality
+fn token_probably_equal_for_proc_macro(first: &Token, other: &Token) -> bool {
+    if mem::discriminant(&first.kind) != mem::discriminant(&other.kind) {
+        return false;
+    }
+    use rustc_ast::token::TokenKind::*;
+    match (&first.kind, &other.kind) {
+        (&Eq, &Eq)
+        | (&Lt, &Lt)
+        | (&Le, &Le)
+        | (&EqEq, &EqEq)
+        | (&Ne, &Ne)
+        | (&Ge, &Ge)
+        | (&Gt, &Gt)
+        | (&AndAnd, &AndAnd)
+        | (&OrOr, &OrOr)
+        | (&Not, &Not)
+        | (&Tilde, &Tilde)
+        | (&At, &At)
+        | (&Dot, &Dot)
+        | (&DotDot, &DotDot)
+        | (&DotDotDot, &DotDotDot)
+        | (&DotDotEq, &DotDotEq)
+        | (&Comma, &Comma)
+        | (&Semi, &Semi)
+        | (&Colon, &Colon)
+        | (&ModSep, &ModSep)
+        | (&RArrow, &RArrow)
+        | (&LArrow, &LArrow)
+        | (&FatArrow, &FatArrow)
+        | (&Pound, &Pound)
+        | (&Dollar, &Dollar)
+        | (&Question, &Question)
+        | (&Whitespace, &Whitespace)
+        | (&Comment, &Comment)
+        | (&Eof, &Eof) => true,
+
+        (&BinOp(a), &BinOp(b)) | (&BinOpEq(a), &BinOpEq(b)) => a == b,
+
+        (&OpenDelim(a), &OpenDelim(b)) | (&CloseDelim(a), &CloseDelim(b)) => a == b,
+
+        (&DocComment(a1, a2, a3), &DocComment(b1, b2, b3)) => a1 == b1 && a2 == b2 && a3 == b3,
+
+        (&Shebang(a), &Shebang(b)) => a == b,
+
+        (&Literal(a), &Literal(b)) => a == b,
+
+        (&Lifetime(a), &Lifetime(b)) => a == b,
+        (&Ident(a, b), &Ident(c, d)) => {
+            b == d && (a == c || a == kw::DollarCrate || c == kw::DollarCrate)
+        }
+
+        (&Interpolated(..), &Interpolated(..)) => panic!("Unexpanded Interpolated!"),
+
+        _ => panic!("forgot to add a token?"),
+    }
+}
+
+fn prepend_attrs(
+    sess: &ParseSess,
+    attrs: &[ast::Attribute],
+    tokens: Option<&tokenstream::TokenStream>,
+    span: rustc_span::Span,
+) -> Option<tokenstream::TokenStream> {
+    let tokens = tokens?;
+    if attrs.is_empty() {
+        return Some(tokens.clone());
+    }
+    let mut builder = tokenstream::TokenStreamBuilder::new();
+    for attr in attrs {
+        assert_eq!(
+            attr.style,
+            ast::AttrStyle::Outer,
+            "inner attributes should prevent cached tokens from existing"
+        );
+
+        let source = pprust::attribute_to_string(attr);
+        let macro_filename = FileName::macro_expansion_source_code(&source);
+
+        let item = match attr.kind {
+            ast::AttrKind::Normal(ref item) => item,
+            ast::AttrKind::DocComment(..) => {
+                let stream = parse_stream_from_source_str(macro_filename, source, sess, Some(span));
+                builder.push(stream);
+                continue;
+            }
+        };
+
+        // synthesize # [ $path $tokens ] manually here
+        let mut brackets = tokenstream::TokenStreamBuilder::new();
+
+        // For simple paths, push the identifier directly
+        if item.path.segments.len() == 1 && item.path.segments[0].args.is_none() {
+            let ident = item.path.segments[0].ident;
+            let token = token::Ident(ident.name, ident.as_str().starts_with("r#"));
+            brackets.push(tokenstream::TokenTree::token(token, ident.span));
+
+        // ... and for more complicated paths, fall back to a reparse hack that
+        // should eventually be removed.
+        } else {
+            let stream = parse_stream_from_source_str(macro_filename, source, sess, Some(span));
+            brackets.push(stream);
+        }
+
+        brackets.push(item.args.outer_tokens());
+
+        // The span we list here for `#` and for `[ ... ]` are both wrong in
+        // that it encompasses more than each token, but it hopefully is "good
+        // enough" for now at least.
+        builder.push(tokenstream::TokenTree::token(token::Pound, attr.span));
+        let delim_span = tokenstream::DelimSpan::from_single(attr.span);
+        builder.push(tokenstream::TokenTree::Delimited(
+            delim_span,
+            token::DelimToken::Bracket,
+            brackets.build(),
+        ));
+    }
+    builder.push(tokens.clone());
+    Some(builder.build())
+}

compiler/rustc_parse/src/parser/attr.rs

@@ -0,0 +1,304 @@
+use super::{Parser, PathStyle};
+use rustc_ast as ast;
+use rustc_ast::attr;
+use rustc_ast::token::{self, Nonterminal};
+use rustc_ast_pretty::pprust;
+use rustc_errors::{error_code, PResult};
+use rustc_span::Span;
+
+use tracing::debug;
+
+#[derive(Debug)]
+pub(super) enum InnerAttrPolicy<'a> {
+    Permitted,
+    Forbidden { reason: &'a str, saw_doc_comment: bool, prev_attr_sp: Option<Span> },
+}
+
+const DEFAULT_UNEXPECTED_INNER_ATTR_ERR_MSG: &str = "an inner attribute is not \
+                                                     permitted in this context";
+
+pub(super) const DEFAULT_INNER_ATTR_FORBIDDEN: InnerAttrPolicy<'_> = InnerAttrPolicy::Forbidden {
+    reason: DEFAULT_UNEXPECTED_INNER_ATTR_ERR_MSG,
+    saw_doc_comment: false,
+    prev_attr_sp: None,
+};
+
+impl<'a> Parser<'a> {
+    /// Parses attributes that appear before an item.
+    pub(super) fn parse_outer_attributes(&mut self) -> PResult<'a, Vec<ast::Attribute>> {
+        let mut attrs: Vec<ast::Attribute> = Vec::new();
+        let mut just_parsed_doc_comment = false;
+        loop {
+            debug!("parse_outer_attributes: self.token={:?}", self.token);
+            if self.check(&token::Pound) {
+                let inner_error_reason = if just_parsed_doc_comment {
+                    "an inner attribute is not permitted following an outer doc comment"
+                } else if !attrs.is_empty() {
+                    "an inner attribute is not permitted following an outer attribute"
+                } else {
+                    DEFAULT_UNEXPECTED_INNER_ATTR_ERR_MSG
+                };
+                let inner_parse_policy = InnerAttrPolicy::Forbidden {
+                    reason: inner_error_reason,
+                    saw_doc_comment: just_parsed_doc_comment,
+                    prev_attr_sp: attrs.last().map(|a| a.span),
+                };
+                let attr = self.parse_attribute_with_inner_parse_policy(inner_parse_policy)?;
+                attrs.push(attr);
+                just_parsed_doc_comment = false;
+            } else if let token::DocComment(comment_kind, attr_style, data) = self.token.kind {
+                let attr = attr::mk_doc_comment(comment_kind, attr_style, data, self.token.span);
+                if attr.style != ast::AttrStyle::Outer {
+                    self.sess
+                        .span_diagnostic
+                        .struct_span_err_with_code(
+                            self.token.span,
+                            "expected outer doc comment",
+                            error_code!(E0753),
+                        )
+                        .note(
+                            "inner doc comments like this (starting with \
+                             `//!` or `/*!`) can only appear before items",
+                        )
+                        .emit();
+                }
+                attrs.push(attr);
+                self.bump();
+                just_parsed_doc_comment = true;
+            } else {
+                break;
+            }
+        }
+        Ok(attrs)
+    }
+
+    /// Matches `attribute = # ! [ meta_item ]`.
+    ///
+    /// If `permit_inner` is `true`, then a leading `!` indicates an inner
+    /// attribute.
+    pub fn parse_attribute(&mut self, permit_inner: bool) -> PResult<'a, ast::Attribute> {
+        debug!("parse_attribute: permit_inner={:?} self.token={:?}", permit_inner, self.token);
+        let inner_parse_policy =
+            if permit_inner { InnerAttrPolicy::Permitted } else { DEFAULT_INNER_ATTR_FORBIDDEN };
+        self.parse_attribute_with_inner_parse_policy(inner_parse_policy)
+    }
+
+    /// The same as `parse_attribute`, except it takes in an `InnerAttrPolicy`
+    /// that prescribes how to handle inner attributes.
+    fn parse_attribute_with_inner_parse_policy(
+        &mut self,
+        inner_parse_policy: InnerAttrPolicy<'_>,
+    ) -> PResult<'a, ast::Attribute> {
+        debug!(
+            "parse_attribute_with_inner_parse_policy: inner_parse_policy={:?} self.token={:?}",
+            inner_parse_policy, self.token
+        );
+        let lo = self.token.span;
+        let (span, item, style) = if self.eat(&token::Pound) {
+            let style =
+                if self.eat(&token::Not) { ast::AttrStyle::Inner } else { ast::AttrStyle::Outer };
+
+            self.expect(&token::OpenDelim(token::Bracket))?;
+            let item = self.parse_attr_item()?;
+            self.expect(&token::CloseDelim(token::Bracket))?;
+            let attr_sp = lo.to(self.prev_token.span);
+
+            // Emit error if inner attribute is encountered and forbidden.
+            if style == ast::AttrStyle::Inner {
+                self.error_on_forbidden_inner_attr(attr_sp, inner_parse_policy);
+            }
+
+            (attr_sp, item, style)
+        } else {
+            let token_str = pprust::token_to_string(&self.token);
+            let msg = &format!("expected `#`, found `{}`", token_str);
+            return Err(self.struct_span_err(self.token.span, msg));
+        };
+
+        Ok(attr::mk_attr_from_item(style, item, span))
+    }
+
+    pub(super) fn error_on_forbidden_inner_attr(&self, attr_sp: Span, policy: InnerAttrPolicy<'_>) {
+        if let InnerAttrPolicy::Forbidden { reason, saw_doc_comment, prev_attr_sp } = policy {
+            let prev_attr_note =
+                if saw_doc_comment { "previous doc comment" } else { "previous outer attribute" };
+
+            let mut diag = self.struct_span_err(attr_sp, reason);
+
+            if let Some(prev_attr_sp) = prev_attr_sp {
+                diag.span_label(attr_sp, "not permitted following an outer attribute")
+                    .span_label(prev_attr_sp, prev_attr_note);
+            }
+
+            diag.note(
+                "inner attributes, like `#![no_std]`, annotate the item enclosing them, \
+                and are usually found at the beginning of source files. \
+                Outer attributes, like `#[test]`, annotate the item following them.",
+            )
+            .emit();
+        }
+    }
+
+    /// Parses an inner part of an attribute (the path and following tokens).
+    /// The tokens must be either a delimited token stream, or empty token stream,
+    /// or the "legacy" key-value form.
+    ///     PATH `(` TOKEN_STREAM `)`
+    ///     PATH `[` TOKEN_STREAM `]`
+    ///     PATH `{` TOKEN_STREAM `}`
+    ///     PATH
+    ///     PATH `=` UNSUFFIXED_LIT
+    /// The delimiters or `=` are still put into the resulting token stream.
+    pub fn parse_attr_item(&mut self) -> PResult<'a, ast::AttrItem> {
+        let item = match self.token.kind {
+            token::Interpolated(ref nt) => match **nt {
+                Nonterminal::NtMeta(ref item) => Some(item.clone().into_inner()),
+                _ => None,
+            },
+            _ => None,
+        };
+        Ok(if let Some(item) = item {
+            self.bump();
+            item
+        } else {
+            let path = self.parse_path(PathStyle::Mod)?;
+            let args = self.parse_attr_args()?;
+            ast::AttrItem { path, args }
+        })
+    }
+
+    /// Parses attributes that appear after the opening of an item. These should
+    /// be preceded by an exclamation mark, but we accept and warn about one
+    /// terminated by a semicolon.
+    ///
+    /// Matches `inner_attrs*`.
+    crate fn parse_inner_attributes(&mut self) -> PResult<'a, Vec<ast::Attribute>> {
+        let mut attrs: Vec<ast::Attribute> = vec![];
+        loop {
+            // Only try to parse if it is an inner attribute (has `!`).
+            if self.check(&token::Pound) && self.look_ahead(1, |t| t == &token::Not) {
+                let attr = self.parse_attribute(true)?;
+                assert_eq!(attr.style, ast::AttrStyle::Inner);
+                attrs.push(attr);
+            } else if let token::DocComment(comment_kind, attr_style, data) = self.token.kind {
+                // We need to get the position of this token before we bump.
+                let attr = attr::mk_doc_comment(comment_kind, attr_style, data, self.token.span);
+                if attr.style == ast::AttrStyle::Inner {
+                    attrs.push(attr);
+                    self.bump();
+                } else {
+                    break;
+                }
+            } else {
+                break;
+            }
+        }
+        Ok(attrs)
+    }
+
+    crate fn parse_unsuffixed_lit(&mut self) -> PResult<'a, ast::Lit> {
+        let lit = self.parse_lit()?;
+        debug!("checking if {:?} is unusuffixed", lit);
+
+        if !lit.kind.is_unsuffixed() {
+            self.struct_span_err(lit.span, "suffixed literals are not allowed in attributes")
+                .help(
+                    "instead of using a suffixed literal (`1u8`, `1.0f32`, etc.), \
+                    use an unsuffixed version (`1`, `1.0`, etc.)",
+                )
+                .emit();
+        }
+
+        Ok(lit)
+    }
+
+    /// Parses `cfg_attr(pred, attr_item_list)` where `attr_item_list` is comma-delimited.
+    pub fn parse_cfg_attr(&mut self) -> PResult<'a, (ast::MetaItem, Vec<(ast::AttrItem, Span)>)> {
+        let cfg_predicate = self.parse_meta_item()?;
+        self.expect(&token::Comma)?;
+
+        // Presumably, the majority of the time there will only be one attr.
+        let mut expanded_attrs = Vec::with_capacity(1);
+        while self.token.kind != token::Eof {
+            let lo = self.token.span;
+            let item = self.parse_attr_item()?;
+            expanded_attrs.push((item, lo.to(self.prev_token.span)));
+            if !self.eat(&token::Comma) {
+                break;
+            }
+        }
+
+        Ok((cfg_predicate, expanded_attrs))
+    }
+
+    /// Matches `COMMASEP(meta_item_inner)`.
+    crate fn parse_meta_seq_top(&mut self) -> PResult<'a, Vec<ast::NestedMetaItem>> {
+        // Presumably, the majority of the time there will only be one attr.
+        let mut nmis = Vec::with_capacity(1);
+        while self.token.kind != token::Eof {
+            nmis.push(self.parse_meta_item_inner()?);
+            if !self.eat(&token::Comma) {
+                break;
+            }
+        }
+        Ok(nmis)
+    }
+
+    /// Matches the following grammar (per RFC 1559).
+    ///
+    ///     meta_item : PATH ( '=' UNSUFFIXED_LIT | '(' meta_item_inner? ')' )? ;
+    ///     meta_item_inner : (meta_item | UNSUFFIXED_LIT) (',' meta_item_inner)? ;
+    pub fn parse_meta_item(&mut self) -> PResult<'a, ast::MetaItem> {
+        let nt_meta = match self.token.kind {
+            token::Interpolated(ref nt) => match **nt {
+                token::NtMeta(ref e) => Some(e.clone()),
+                _ => None,
+            },
+            _ => None,
+        };
+
+        if let Some(item) = nt_meta {
+            return match item.meta(item.path.span) {
+                Some(meta) => {
+                    self.bump();
+                    Ok(meta)
+                }
+                None => self.unexpected(),
+            };
+        }
+
+        let lo = self.token.span;
+        let path = self.parse_path(PathStyle::Mod)?;
+        let kind = self.parse_meta_item_kind()?;
+        let span = lo.to(self.prev_token.span);
+        Ok(ast::MetaItem { path, kind, span })
+    }
+
+    crate fn parse_meta_item_kind(&mut self) -> PResult<'a, ast::MetaItemKind> {
+        Ok(if self.eat(&token::Eq) {
+            ast::MetaItemKind::NameValue(self.parse_unsuffixed_lit()?)
+        } else if self.check(&token::OpenDelim(token::Paren)) {
+            // Matches `meta_seq = ( COMMASEP(meta_item_inner) )`.
+            let (list, _) = self.parse_paren_comma_seq(|p| p.parse_meta_item_inner())?;
+            ast::MetaItemKind::List(list)
+        } else {
+            ast::MetaItemKind::Word
+        })
+    }
+
+    /// Matches `meta_item_inner : (meta_item | UNSUFFIXED_LIT) ;`.
+    fn parse_meta_item_inner(&mut self) -> PResult<'a, ast::NestedMetaItem> {
+        match self.parse_unsuffixed_lit() {
+            Ok(lit) => return Ok(ast::NestedMetaItem::Literal(lit)),
+            Err(ref mut err) => err.cancel(),
+        }
+
+        match self.parse_meta_item() {
+            Ok(mi) => return Ok(ast::NestedMetaItem::MetaItem(mi)),
+            Err(ref mut err) => err.cancel(),
+        }
+
+        let found = pprust::token_to_string(&self.token);
+        let msg = format!("expected unsuffixed literal or identifier, found `{}`", found);
+        Err(self.struct_span_err(self.token.span, &msg))
+    }
+}

compiler/rustc_parse/src/parser/generics.rs

@@ -0,0 +1,292 @@
+use super::Parser;
+
+use rustc_ast::token;
+use rustc_ast::{
+    self as ast, Attribute, GenericBounds, GenericParam, GenericParamKind, WhereClause,
+};
+use rustc_errors::PResult;
+use rustc_span::symbol::{kw, sym};
+
+impl<'a> Parser<'a> {
+    /// Parses bounds of a lifetime parameter `BOUND + BOUND + BOUND`, possibly with trailing `+`.
+    ///
+    /// ```text
+    /// BOUND = LT_BOUND (e.g., `'a`)
+    /// ```
+    fn parse_lt_param_bounds(&mut self) -> GenericBounds {
+        let mut lifetimes = Vec::new();
+        while self.check_lifetime() {
+            lifetimes.push(ast::GenericBound::Outlives(self.expect_lifetime()));
+
+            if !self.eat_plus() {
+                break;
+            }
+        }
+        lifetimes
+    }
+
+    /// Matches `typaram = IDENT (`?` unbound)? optbounds ( EQ ty )?`.
+    fn parse_ty_param(&mut self, preceding_attrs: Vec<Attribute>) -> PResult<'a, GenericParam> {
+        let ident = self.parse_ident()?;
+
+        // Parse optional colon and param bounds.
+        let bounds = if self.eat(&token::Colon) {
+            self.parse_generic_bounds(Some(self.prev_token.span))?
+        } else {
+            Vec::new()
+        };
+
+        let default = if self.eat(&token::Eq) { Some(self.parse_ty()?) } else { None };
+
+        Ok(GenericParam {
+            ident,
+            id: ast::DUMMY_NODE_ID,
+            attrs: preceding_attrs.into(),
+            bounds,
+            kind: GenericParamKind::Type { default },
+            is_placeholder: false,
+        })
+    }
+
+    fn parse_const_param(&mut self, preceding_attrs: Vec<Attribute>) -> PResult<'a, GenericParam> {
+        let const_span = self.token.span;
+
+        self.expect_keyword(kw::Const)?;
+        let ident = self.parse_ident()?;
+        self.expect(&token::Colon)?;
+        let ty = self.parse_ty()?;
+
+        self.sess.gated_spans.gate(sym::min_const_generics, const_span.to(self.prev_token.span));
+
+        Ok(GenericParam {
+            ident,
+            id: ast::DUMMY_NODE_ID,
+            attrs: preceding_attrs.into(),
+            bounds: Vec::new(),
+            kind: GenericParamKind::Const { ty, kw_span: const_span },
+            is_placeholder: false,
+        })
+    }
+
+    /// Parses a (possibly empty) list of lifetime and type parameters, possibly including
+    /// a trailing comma and erroneous trailing attributes.
+    pub(super) fn parse_generic_params(&mut self) -> PResult<'a, Vec<ast::GenericParam>> {
+        let mut params = Vec::new();
+        loop {
+            let attrs = self.parse_outer_attributes()?;
+            if self.check_lifetime() {
+                let lifetime = self.expect_lifetime();
+                // Parse lifetime parameter.
+                let bounds =
+                    if self.eat(&token::Colon) { self.parse_lt_param_bounds() } else { Vec::new() };
+                params.push(ast::GenericParam {
+                    ident: lifetime.ident,
+                    id: lifetime.id,
+                    attrs: attrs.into(),
+                    bounds,
+                    kind: ast::GenericParamKind::Lifetime,
+                    is_placeholder: false,
+                });
+            } else if self.check_keyword(kw::Const) {
+                // Parse const parameter.
+                params.push(self.parse_const_param(attrs)?);
+            } else if self.check_ident() {
+                // Parse type parameter.
+                params.push(self.parse_ty_param(attrs)?);
+            } else if self.token.can_begin_type() {
+                // Trying to write an associated type bound? (#26271)
+                let snapshot = self.clone();
+                match self.parse_ty_where_predicate() {
+                    Ok(where_predicate) => {
+                        self.struct_span_err(
+                            where_predicate.span(),
+                            "bounds on associated types do not belong here",
+                        )
+                        .span_label(where_predicate.span(), "belongs in `where` clause")
+                        .emit();
+                    }
+                    Err(mut err) => {
+                        err.cancel();
+                        *self = snapshot;
+                        break;
+                    }
+                }
+            } else {
+                // Check for trailing attributes and stop parsing.
+                if !attrs.is_empty() {
+                    if !params.is_empty() {
+                        self.struct_span_err(
+                            attrs[0].span,
+                            "trailing attribute after generic parameter",
+                        )
+                        .span_label(attrs[0].span, "attributes must go before parameters")
+                        .emit();
+                    } else {
+                        self.struct_span_err(attrs[0].span, "attribute without generic parameters")
+                            .span_label(
+                                attrs[0].span,
+                                "attributes are only permitted when preceding parameters",
+                            )
+                            .emit();
+                    }
+                }
+                break;
+            }
+
+            if !self.eat(&token::Comma) {
+                break;
+            }
+        }
+        Ok(params)
+    }
+
+    /// Parses a set of optional generic type parameter declarations. Where
+    /// clauses are not parsed here, and must be added later via
+    /// `parse_where_clause()`.
+    ///
+    /// matches generics = ( ) | ( < > ) | ( < typaramseq ( , )? > ) | ( < lifetimes ( , )? > )
+    ///                  | ( < lifetimes , typaramseq ( , )? > )
+    /// where   typaramseq = ( typaram ) | ( typaram , typaramseq )
+    pub(super) fn parse_generics(&mut self) -> PResult<'a, ast::Generics> {
+        let span_lo = self.token.span;
+        let (params, span) = if self.eat_lt() {
+            let params = self.parse_generic_params()?;
+            self.expect_gt()?;
+            (params, span_lo.to(self.prev_token.span))
+        } else {
+            (vec![], self.prev_token.span.shrink_to_hi())
+        };
+        Ok(ast::Generics {
+            params,
+            where_clause: WhereClause {
+                has_where_token: false,
+                predicates: Vec::new(),
+                span: self.prev_token.span.shrink_to_hi(),
+            },
+            span,
+        })
+    }
+
+    /// Parses an optional where-clause and places it in `generics`.
+    ///
+    /// ```ignore (only-for-syntax-highlight)
+    /// where T : Trait<U, V> + 'b, 'a : 'b
+    /// ```
+    pub(super) fn parse_where_clause(&mut self) -> PResult<'a, WhereClause> {
+        let mut where_clause = WhereClause {
+            has_where_token: false,
+            predicates: Vec::new(),
+            span: self.prev_token.span.shrink_to_hi(),
+        };
+
+        if !self.eat_keyword(kw::Where) {
+            return Ok(where_clause);
+        }
+        where_clause.has_where_token = true;
+        let lo = self.prev_token.span;
+
+        // We are considering adding generics to the `where` keyword as an alternative higher-rank
+        // parameter syntax (as in `where<'a>` or `where<T>`. To avoid that being a breaking
+        // change we parse those generics now, but report an error.
+        if self.choose_generics_over_qpath(0) {
+            let generics = self.parse_generics()?;
+            self.struct_span_err(
+                generics.span,
+                "generic parameters on `where` clauses are reserved for future use",
+            )
+            .span_label(generics.span, "currently unsupported")
+            .emit();
+        }
+
+        loop {
+            let lo = self.token.span;
+            if self.check_lifetime() && self.look_ahead(1, |t| !t.is_like_plus()) {
+                let lifetime = self.expect_lifetime();
+                // Bounds starting with a colon are mandatory, but possibly empty.
+                self.expect(&token::Colon)?;
+                let bounds = self.parse_lt_param_bounds();
+                where_clause.predicates.push(ast::WherePredicate::RegionPredicate(
+                    ast::WhereRegionPredicate {
+                        span: lo.to(self.prev_token.span),
+                        lifetime,
+                        bounds,
+                    },
+                ));
+            } else if self.check_type() {
+                where_clause.predicates.push(self.parse_ty_where_predicate()?);
+            } else {
+                break;
+            }
+
+            if !self.eat(&token::Comma) {
+                break;
+            }
+        }
+
+        where_clause.span = lo.to(self.prev_token.span);
+        Ok(where_clause)
+    }
+
+    fn parse_ty_where_predicate(&mut self) -> PResult<'a, ast::WherePredicate> {
+        let lo = self.token.span;
+        // Parse optional `for<'a, 'b>`.
+        // This `for` is parsed greedily and applies to the whole predicate,
+        // the bounded type can have its own `for` applying only to it.
+        // Examples:
+        // * `for<'a> Trait1<'a>: Trait2<'a /* ok */>`
+        // * `(for<'a> Trait1<'a>): Trait2<'a /* not ok */>`
+        // * `for<'a> for<'b> Trait1<'a, 'b>: Trait2<'a /* ok */, 'b /* not ok */>`
+        let lifetime_defs = self.parse_late_bound_lifetime_defs()?;
+
+        // Parse type with mandatory colon and (possibly empty) bounds,
+        // or with mandatory equality sign and the second type.
+        let ty = self.parse_ty()?;
+        if self.eat(&token::Colon) {
+            let bounds = self.parse_generic_bounds(Some(self.prev_token.span))?;
+            Ok(ast::WherePredicate::BoundPredicate(ast::WhereBoundPredicate {
+                span: lo.to(self.prev_token.span),
+                bound_generic_params: lifetime_defs,
+                bounded_ty: ty,
+                bounds,
+            }))
+        // FIXME: Decide what should be used here, `=` or `==`.
+        // FIXME: We are just dropping the binders in lifetime_defs on the floor here.
+        } else if self.eat(&token::Eq) || self.eat(&token::EqEq) {
+            let rhs_ty = self.parse_ty()?;
+            Ok(ast::WherePredicate::EqPredicate(ast::WhereEqPredicate {
+                span: lo.to(self.prev_token.span),
+                lhs_ty: ty,
+                rhs_ty,
+                id: ast::DUMMY_NODE_ID,
+            }))
+        } else {
+            self.unexpected()
+        }
+    }
+
+    pub(super) fn choose_generics_over_qpath(&self, start: usize) -> bool {
+        // There's an ambiguity between generic parameters and qualified paths in impls.
+        // If we see `<` it may start both, so we have to inspect some following tokens.
+        // The following combinations can only start generics,
+        // but not qualified paths (with one exception):
+        //     `<` `>` - empty generic parameters
+        //     `<` `#` - generic parameters with attributes
+        //     `<` (LIFETIME|IDENT) `>` - single generic parameter
+        //     `<` (LIFETIME|IDENT) `,` - first generic parameter in a list
+        //     `<` (LIFETIME|IDENT) `:` - generic parameter with bounds
+        //     `<` (LIFETIME|IDENT) `=` - generic parameter with a default
+        //     `<` const                - generic const parameter
+        // The only truly ambiguous case is
+        //     `<` IDENT `>` `::` IDENT ...
+        // we disambiguate it in favor of generics (`impl<T> ::absolute::Path<T> { ... }`)
+        // because this is what almost always expected in practice, qualified paths in impls
+        // (`impl <Type>::AssocTy { ... }`) aren't even allowed by type checker at the moment.
+        self.look_ahead(start, |t| t == &token::Lt)
+            && (self.look_ahead(start + 1, |t| t == &token::Pound || t == &token::Gt)
+                || self.look_ahead(start + 1, |t| t.is_lifetime() || t.is_ident())
+                    && self.look_ahead(start + 2, |t| {
+                        matches!(t.kind, token::Gt | token::Comma | token::Colon | token::Eq)
+                    })
+                || self.is_keyword_ahead(start + 1, &[kw::Const]))
+    }
+}

compiler/rustc_parse/src/parser/nonterminal.rs

@@ -0,0 +1,170 @@
+use rustc_ast::ptr::P;
+use rustc_ast::token::{self, Nonterminal, NonterminalKind, Token};
+use rustc_ast_pretty::pprust;
+use rustc_errors::PResult;
+use rustc_span::symbol::{kw, Ident};
+
+use crate::parser::{FollowedByType, Parser, PathStyle};
+
+impl<'a> Parser<'a> {
+    /// Checks whether a non-terminal may begin with a particular token.
+    ///
+    /// Returning `false` is a *stability guarantee* that such a matcher will *never* begin with that
+    /// token. Be conservative (return true) if not sure.
+    pub fn nonterminal_may_begin_with(kind: NonterminalKind, token: &Token) -> bool {
+        /// Checks whether the non-terminal may contain a single (non-keyword) identifier.
+        fn may_be_ident(nt: &token::Nonterminal) -> bool {
+            match *nt {
+                token::NtItem(_) | token::NtBlock(_) | token::NtVis(_) | token::NtLifetime(_) => {
+                    false
+                }
+                _ => true,
+            }
+        }
+
+        match kind {
+            NonterminalKind::Expr => {
+                token.can_begin_expr()
+                // This exception is here for backwards compatibility.
+                && !token.is_keyword(kw::Let)
+            }
+            NonterminalKind::Ty => token.can_begin_type(),
+            NonterminalKind::Ident => get_macro_ident(token).is_some(),
+            NonterminalKind::Literal => token.can_begin_literal_maybe_minus(),
+            NonterminalKind::Vis => match token.kind {
+                // The follow-set of :vis + "priv" keyword + interpolated
+                token::Comma | token::Ident(..) | token::Interpolated(..) => true,
+                _ => token.can_begin_type(),
+            },
+            NonterminalKind::Block => match token.kind {
+                token::OpenDelim(token::Brace) => true,
+                token::Interpolated(ref nt) => match **nt {
+                    token::NtItem(_)
+                    | token::NtPat(_)
+                    | token::NtTy(_)
+                    | token::NtIdent(..)
+                    | token::NtMeta(_)
+                    | token::NtPath(_)
+                    | token::NtVis(_) => false, // none of these may start with '{'.
+                    _ => true,
+                },
+                _ => false,
+            },
+            NonterminalKind::Path | NonterminalKind::Meta => match token.kind {
+                token::ModSep | token::Ident(..) => true,
+                token::Interpolated(ref nt) => match **nt {
+                    token::NtPath(_) | token::NtMeta(_) => true,
+                    _ => may_be_ident(&nt),
+                },
+                _ => false,
+            },
+            NonterminalKind::Pat => match token.kind {
+                token::Ident(..) |                  // box, ref, mut, and other identifiers (can stricten)
+                token::OpenDelim(token::Paren) |    // tuple pattern
+                token::OpenDelim(token::Bracket) |  // slice pattern
+                token::BinOp(token::And) |          // reference
+                token::BinOp(token::Minus) |        // negative literal
+                token::AndAnd |                     // double reference
+                token::Literal(..) |                // literal
+                token::DotDot |                     // range pattern (future compat)
+                token::DotDotDot |                  // range pattern (future compat)
+                token::ModSep |                     // path
+                token::Lt |                         // path (UFCS constant)
+                token::BinOp(token::Shl) => true,   // path (double UFCS)
+                token::Interpolated(ref nt) => may_be_ident(nt),
+                _ => false,
+            },
+            NonterminalKind::Lifetime => match token.kind {
+                token::Lifetime(_) => true,
+                token::Interpolated(ref nt) => match **nt {
+                    token::NtLifetime(_) | token::NtTT(_) => true,
+                    _ => false,
+                },
+                _ => false,
+            },
+            NonterminalKind::TT | NonterminalKind::Item | NonterminalKind::Stmt => match token.kind
+            {
+                token::CloseDelim(_) => false,
+                _ => true,
+            },
+        }
+    }
+
+    pub fn parse_nonterminal(&mut self, kind: NonterminalKind) -> PResult<'a, Nonterminal> {
+        // Any `Nonterminal` which stores its tokens (currently `NtItem` and `NtExpr`)
+        // needs to have them force-captured here.
+        // A `macro_rules!` invocation may pass a captured item/expr to a proc-macro,
+        // which requires having captured tokens available. Since we cannot determine
+        // in advance whether or not a proc-macro will be (transitively) invoked,
+        // we always capture tokens for any `Nonterminal` which needs them.
+        Ok(match kind {
+            NonterminalKind::Item => match self.collect_tokens(|this| this.parse_item())? {
+                (Some(mut item), tokens) => {
+                    // If we captured tokens during parsing (due to outer attributes),
+                    // use those.
+                    if item.tokens.is_none() {
+                        item.tokens = Some(tokens);
+                    }
+                    token::NtItem(item)
+                }
+                (None, _) => {
+                    return Err(self.struct_span_err(self.token.span, "expected an item keyword"));
+                }
+            },
+            NonterminalKind::Block => token::NtBlock(self.parse_block()?),
+            NonterminalKind::Stmt => match self.parse_stmt()? {
+                Some(s) => token::NtStmt(s),
+                None => return Err(self.struct_span_err(self.token.span, "expected a statement")),
+            },
+            NonterminalKind::Pat => {
+                let (mut pat, tokens) = self.collect_tokens(|this| this.parse_pat(None))?;
+                // We have have eaten an NtPat, which could already have tokens
+                if pat.tokens.is_none() {
+                    pat.tokens = Some(tokens);
+                }
+                token::NtPat(pat)
+            }
+            NonterminalKind::Expr => {
+                let (mut expr, tokens) = self.collect_tokens(|this| this.parse_expr())?;
+                // If we captured tokens during parsing (due to outer attributes),
+                // use those.
+                if expr.tokens.is_none() {
+                    expr.tokens = Some(tokens);
+                }
+                token::NtExpr(expr)
+            }
+            NonterminalKind::Literal => token::NtLiteral(self.parse_literal_maybe_minus()?),
+            NonterminalKind::Ty => token::NtTy(self.parse_ty()?),
+            // this could be handled like a token, since it is one
+            NonterminalKind::Ident => {
+                if let Some((ident, is_raw)) = get_macro_ident(&self.token) {
+                    self.bump();
+                    token::NtIdent(ident, is_raw)
+                } else {
+                    let token_str = pprust::token_to_string(&self.token);
+                    let msg = &format!("expected ident, found {}", &token_str);
+                    return Err(self.struct_span_err(self.token.span, msg));
+                }
+            }
+            NonterminalKind::Path => token::NtPath(self.parse_path(PathStyle::Type)?),
+            NonterminalKind::Meta => token::NtMeta(P(self.parse_attr_item()?)),
+            NonterminalKind::TT => token::NtTT(self.parse_token_tree()),
+            NonterminalKind::Vis => token::NtVis(self.parse_visibility(FollowedByType::Yes)?),
+            NonterminalKind::Lifetime => {
+                if self.check_lifetime() {
+                    token::NtLifetime(self.expect_lifetime().ident)
+                } else {
+                    let token_str = pprust::token_to_string(&self.token);
+                    let msg = &format!("expected a lifetime, found `{}`", &token_str);
+                    return Err(self.struct_span_err(self.token.span, msg));
+                }
+            }
+        })
+    }
+}
+
+/// The token is an identifier, but not `_`.
+/// We prohibit passing `_` to macros expecting `ident` for now.
+fn get_macro_ident(token: &Token) -> Option<(Ident, bool)> {
+    token.ident().filter(|(ident, _)| ident.name != kw::Underscore)
+}

compiler/rustc_parse/src/parser/path.rs

@@ -0,0 +1,516 @@
+use super::ty::{AllowPlus, RecoverQPath};
+use super::{Parser, TokenType};
+use crate::maybe_whole;
+use rustc_ast::ptr::P;
+use rustc_ast::token::{self, Token};
+use rustc_ast::{
+    self as ast, AngleBracketedArg, AngleBracketedArgs, GenericArg, ParenthesizedArgs,
+};
+use rustc_ast::{AnonConst, AssocTyConstraint, AssocTyConstraintKind, BlockCheckMode};
+use rustc_ast::{Path, PathSegment, QSelf};
+use rustc_errors::{pluralize, Applicability, PResult};
+use rustc_span::source_map::{BytePos, Span};
+use rustc_span::symbol::{kw, sym, Ident};
+
+use std::mem;
+use tracing::debug;
+
+/// Specifies how to parse a path.
+#[derive(Copy, Clone, PartialEq)]
+pub enum PathStyle {
+    /// In some contexts, notably in expressions, paths with generic arguments are ambiguous
+    /// with something else. For example, in expressions `segment < ....` can be interpreted
+    /// as a comparison and `segment ( ....` can be interpreted as a function call.
+    /// In all such contexts the non-path interpretation is preferred by default for practical
+    /// reasons, but the path interpretation can be forced by the disambiguator `::`, e.g.
+    /// `x<y>` - comparisons, `x::<y>` - unambiguously a path.
+    Expr,
+    /// In other contexts, notably in types, no ambiguity exists and paths can be written
+    /// without the disambiguator, e.g., `x<y>` - unambiguously a path.
+    /// Paths with disambiguators are still accepted, `x::<Y>` - unambiguously a path too.
+    Type,
+    /// A path with generic arguments disallowed, e.g., `foo::bar::Baz`, used in imports,
+    /// visibilities or attributes.
+    /// Technically, this variant is unnecessary and e.g., `Expr` can be used instead
+    /// (paths in "mod" contexts have to be checked later for absence of generic arguments
+    /// anyway, due to macros), but it is used to avoid weird suggestions about expected
+    /// tokens when something goes wrong.
+    Mod,
+}
+
+impl<'a> Parser<'a> {
+    /// Parses a qualified path.
+    /// Assumes that the leading `<` has been parsed already.
+    ///
+    /// `qualified_path = <type [as trait_ref]>::path`
+    ///
+    /// # Examples
+    /// `<T>::default`
+    /// `<T as U>::a`
+    /// `<T as U>::F::a<S>` (without disambiguator)
+    /// `<T as U>::F::a::<S>` (with disambiguator)
+    pub(super) fn parse_qpath(&mut self, style: PathStyle) -> PResult<'a, (QSelf, Path)> {
+        let lo = self.prev_token.span;
+        let ty = self.parse_ty()?;
+
+        // `path` will contain the prefix of the path up to the `>`,
+        // if any (e.g., `U` in the `<T as U>::*` examples
+        // above). `path_span` has the span of that path, or an empty
+        // span in the case of something like `<T>::Bar`.
+        let (mut path, path_span);
+        if self.eat_keyword(kw::As) {
+            let path_lo = self.token.span;
+            path = self.parse_path(PathStyle::Type)?;
+            path_span = path_lo.to(self.prev_token.span);
+        } else {
+            path_span = self.token.span.to(self.token.span);
+            path = ast::Path { segments: Vec::new(), span: path_span };
+        }
+
+        // See doc comment for `unmatched_angle_bracket_count`.
+        self.expect(&token::Gt)?;
+        if self.unmatched_angle_bracket_count > 0 {
+            self.unmatched_angle_bracket_count -= 1;
+            debug!("parse_qpath: (decrement) count={:?}", self.unmatched_angle_bracket_count);
+        }
+
+        if !self.recover_colon_before_qpath_proj() {
+            self.expect(&token::ModSep)?;
+        }
+
+        let qself = QSelf { ty, path_span, position: path.segments.len() };
+        self.parse_path_segments(&mut path.segments, style)?;
+
+        Ok((qself, Path { segments: path.segments, span: lo.to(self.prev_token.span) }))
+    }
+
+    /// Recover from an invalid single colon, when the user likely meant a qualified path.
+    /// We avoid emitting this if not followed by an identifier, as our assumption that the user
+    /// intended this to be a qualified path may not be correct.
+    ///
+    /// ```ignore (diagnostics)
+    /// <Bar as Baz<T>>:Qux
+    ///                ^ help: use double colon
+    /// ```
+    fn recover_colon_before_qpath_proj(&mut self) -> bool {
+        if self.token.kind != token::Colon
+            || self.look_ahead(1, |t| !t.is_ident() || t.is_reserved_ident())
+        {
+            return false;
+        }
+
+        self.bump(); // colon
+
+        self.diagnostic()
+            .struct_span_err(
+                self.prev_token.span,
+                "found single colon before projection in qualified path",
+            )
+            .span_suggestion(
+                self.prev_token.span,
+                "use double colon",
+                "::".to_string(),
+                Applicability::MachineApplicable,
+            )
+            .emit();
+
+        true
+    }
+
+    /// Parses simple paths.
+    ///
+    /// `path = [::] segment+`
+    /// `segment = ident | ident[::]<args> | ident[::](args) [-> type]`
+    ///
+    /// # Examples
+    /// `a::b::C<D>` (without disambiguator)
+    /// `a::b::C::<D>` (with disambiguator)
+    /// `Fn(Args)` (without disambiguator)
+    /// `Fn::(Args)` (with disambiguator)
+    pub(super) fn parse_path(&mut self, style: PathStyle) -> PResult<'a, Path> {
+        maybe_whole!(self, NtPath, |path| {
+            if style == PathStyle::Mod && path.segments.iter().any(|segment| segment.args.is_some())
+            {
+                self.struct_span_err(path.span, "unexpected generic arguments in path").emit();
+            }
+            path
+        });
+
+        let lo = self.token.span;
+        let mut segments = Vec::new();
+        let mod_sep_ctxt = self.token.span.ctxt();
+        if self.eat(&token::ModSep) {
+            segments.push(PathSegment::path_root(lo.shrink_to_lo().with_ctxt(mod_sep_ctxt)));
+        }
+        self.parse_path_segments(&mut segments, style)?;
+
+        Ok(Path { segments, span: lo.to(self.prev_token.span) })
+    }
+
+    pub(super) fn parse_path_segments(
+        &mut self,
+        segments: &mut Vec<PathSegment>,
+        style: PathStyle,
+    ) -> PResult<'a, ()> {
+        loop {
+            let segment = self.parse_path_segment(style)?;
+            if style == PathStyle::Expr {
+                // In order to check for trailing angle brackets, we must have finished
+                // recursing (`parse_path_segment` can indirectly call this function),
+                // that is, the next token must be the highlighted part of the below example:
+                //
+                // `Foo::<Bar as Baz<T>>::Qux`
+                //                      ^ here
+                //
+                // As opposed to the below highlight (if we had only finished the first
+                // recursion):
+                //
+                // `Foo::<Bar as Baz<T>>::Qux`
+                //                     ^ here
+                //
+                // `PathStyle::Expr` is only provided at the root invocation and never in
+                // `parse_path_segment` to recurse and therefore can be checked to maintain
+                // this invariant.
+                self.check_trailing_angle_brackets(&segment, &[&token::ModSep]);
+            }
+            segments.push(segment);
+
+            if self.is_import_coupler() || !self.eat(&token::ModSep) {
+                return Ok(());
+            }
+        }
+    }
+
+    pub(super) fn parse_path_segment(&mut self, style: PathStyle) -> PResult<'a, PathSegment> {
+        let ident = self.parse_path_segment_ident()?;
+
+        let is_args_start = |token: &Token| match token.kind {
+            token::Lt
+            | token::BinOp(token::Shl)
+            | token::OpenDelim(token::Paren)
+            | token::LArrow => true,
+            _ => false,
+        };
+        let check_args_start = |this: &mut Self| {
+            this.expected_tokens.extend_from_slice(&[
+                TokenType::Token(token::Lt),
+                TokenType::Token(token::OpenDelim(token::Paren)),
+            ]);
+            is_args_start(&this.token)
+        };
+
+        Ok(
+            if style == PathStyle::Type && check_args_start(self)
+                || style != PathStyle::Mod
+                    && self.check(&token::ModSep)
+                    && self.look_ahead(1, |t| is_args_start(t))
+            {
+                // We use `style == PathStyle::Expr` to check if this is in a recursion or not. If
+                // it isn't, then we reset the unmatched angle bracket count as we're about to start
+                // parsing a new path.
+                if style == PathStyle::Expr {
+                    self.unmatched_angle_bracket_count = 0;
+                    self.max_angle_bracket_count = 0;
+                }
+
+                // Generic arguments are found - `<`, `(`, `::<` or `::(`.
+                self.eat(&token::ModSep);
+                let lo = self.token.span;
+                let args = if self.eat_lt() {
+                    // `<'a, T, A = U>`
+                    let args =
+                        self.parse_angle_args_with_leading_angle_bracket_recovery(style, lo)?;
+                    self.expect_gt()?;
+                    let span = lo.to(self.prev_token.span);
+                    AngleBracketedArgs { args, span }.into()
+                } else {
+                    // `(T, U) -> R`
+                    let (inputs, _) = self.parse_paren_comma_seq(|p| p.parse_ty())?;
+                    let span = ident.span.to(self.prev_token.span);
+                    let output = self.parse_ret_ty(AllowPlus::No, RecoverQPath::No)?;
+                    ParenthesizedArgs { inputs, output, span }.into()
+                };
+
+                PathSegment { ident, args, id: ast::DUMMY_NODE_ID }
+            } else {
+                // Generic arguments are not found.
+                PathSegment::from_ident(ident)
+            },
+        )
+    }
+
+    pub(super) fn parse_path_segment_ident(&mut self) -> PResult<'a, Ident> {
+        match self.token.ident() {
+            Some((ident, false)) if ident.is_path_segment_keyword() => {
+                self.bump();
+                Ok(ident)
+            }
+            _ => self.parse_ident(),
+        }
+    }
+
+    /// Parses generic args (within a path segment) with recovery for extra leading angle brackets.
+    /// For the purposes of understanding the parsing logic of generic arguments, this function
+    /// can be thought of being the same as just calling `self.parse_angle_args()` if the source
+    /// had the correct amount of leading angle brackets.
+    ///
+    /// ```ignore (diagnostics)
+    /// bar::<<<<T as Foo>::Output>();
+    ///      ^^ help: remove extra angle brackets
+    /// ```
+    fn parse_angle_args_with_leading_angle_bracket_recovery(
+        &mut self,
+        style: PathStyle,
+        lo: Span,
+    ) -> PResult<'a, Vec<AngleBracketedArg>> {
+        // We need to detect whether there are extra leading left angle brackets and produce an
+        // appropriate error and suggestion. This cannot be implemented by looking ahead at
+        // upcoming tokens for a matching `>` character - if there are unmatched `<` tokens
+        // then there won't be matching `>` tokens to find.
+        //
+        // To explain how this detection works, consider the following example:
+        //
+        // ```ignore (diagnostics)
+        // bar::<<<<T as Foo>::Output>();
+        //      ^^ help: remove extra angle brackets
+        // ```
+        //
+        // Parsing of the left angle brackets starts in this function. We start by parsing the
+        // `<` token (incrementing the counter of unmatched angle brackets on `Parser` via
+        // `eat_lt`):
+        //
+        // *Upcoming tokens:* `<<<<T as Foo>::Output>;`
+        // *Unmatched count:* 1
+        // *`parse_path_segment` calls deep:* 0
+        //
+        // This has the effect of recursing as this function is called if a `<` character
+        // is found within the expected generic arguments:
+        //
+        // *Upcoming tokens:* `<<<T as Foo>::Output>;`
+        // *Unmatched count:* 2
+        // *`parse_path_segment` calls deep:* 1
+        //
+        // Eventually we will have recursed until having consumed all of the `<` tokens and
+        // this will be reflected in the count:
+        //
+        // *Upcoming tokens:* `T as Foo>::Output>;`
+        // *Unmatched count:* 4
+        // `parse_path_segment` calls deep:* 3
+        //
+        // The parser will continue until reaching the first `>` - this will decrement the
+        // unmatched angle bracket count and return to the parent invocation of this function
+        // having succeeded in parsing:
+        //
+        // *Upcoming tokens:* `::Output>;`
+        // *Unmatched count:* 3
+        // *`parse_path_segment` calls deep:* 2
+        //
+        // This will continue until the next `>` character which will also return successfully
+        // to the parent invocation of this function and decrement the count:
+        //
+        // *Upcoming tokens:* `;`
+        // *Unmatched count:* 2
+        // *`parse_path_segment` calls deep:* 1
+        //
+        // At this point, this function will expect to find another matching `>` character but
+        // won't be able to and will return an error. This will continue all the way up the
+        // call stack until the first invocation:
+        //
+        // *Upcoming tokens:* `;`
+        // *Unmatched count:* 2
+        // *`parse_path_segment` calls deep:* 0
+        //
+        // In doing this, we have managed to work out how many unmatched leading left angle
+        // brackets there are, but we cannot recover as the unmatched angle brackets have
+        // already been consumed. To remedy this, we keep a snapshot of the parser state
+        // before we do the above. We can then inspect whether we ended up with a parsing error
+        // and unmatched left angle brackets and if so, restore the parser state before we
+        // consumed any `<` characters to emit an error and consume the erroneous tokens to
+        // recover by attempting to parse again.
+        //
+        // In practice, the recursion of this function is indirect and there will be other
+        // locations that consume some `<` characters - as long as we update the count when
+        // this happens, it isn't an issue.
+
+        let is_first_invocation = style == PathStyle::Expr;
+        // Take a snapshot before attempting to parse - we can restore this later.
+        let snapshot = if is_first_invocation { Some(self.clone()) } else { None };
+
+        debug!("parse_generic_args_with_leading_angle_bracket_recovery: (snapshotting)");
+        match self.parse_angle_args() {
+            Ok(args) => Ok(args),
+            Err(ref mut e) if is_first_invocation && self.unmatched_angle_bracket_count > 0 => {
+                // Cancel error from being unable to find `>`. We know the error
+                // must have been this due to a non-zero unmatched angle bracket
+                // count.
+                e.cancel();
+
+                // Swap `self` with our backup of the parser state before attempting to parse
+                // generic arguments.
+                let snapshot = mem::replace(self, snapshot.unwrap());
+
+                debug!(
+                    "parse_generic_args_with_leading_angle_bracket_recovery: (snapshot failure) \
+                     snapshot.count={:?}",
+                    snapshot.unmatched_angle_bracket_count,
+                );
+
+                // Eat the unmatched angle brackets.
+                for _ in 0..snapshot.unmatched_angle_bracket_count {
+                    self.eat_lt();
+                }
+
+                // Make a span over ${unmatched angle bracket count} characters.
+                let span = lo.with_hi(lo.lo() + BytePos(snapshot.unmatched_angle_bracket_count));
+                self.struct_span_err(
+                    span,
+                    &format!(
+                        "unmatched angle bracket{}",
+                        pluralize!(snapshot.unmatched_angle_bracket_count)
+                    ),
+                )
+                .span_suggestion(
+                    span,
+                    &format!(
+                        "remove extra angle bracket{}",
+                        pluralize!(snapshot.unmatched_angle_bracket_count)
+                    ),
+                    String::new(),
+                    Applicability::MachineApplicable,
+                )
+                .emit();
+
+                // Try again without unmatched angle bracket characters.
+                self.parse_angle_args()
+            }
+            Err(e) => Err(e),
+        }
+    }
+
+    /// Parses (possibly empty) list of generic arguments / associated item constraints,
+    /// possibly including trailing comma.
+    pub(super) fn parse_angle_args(&mut self) -> PResult<'a, Vec<AngleBracketedArg>> {
+        let mut args = Vec::new();
+        while let Some(arg) = self.parse_angle_arg()? {
+            args.push(arg);
+            if !self.eat(&token::Comma) {
+                break;
+            }
+        }
+        Ok(args)
+    }
+
+    /// Parses a single argument in the angle arguments `<...>` of a path segment.
+    fn parse_angle_arg(&mut self) -> PResult<'a, Option<AngleBracketedArg>> {
+        if self.check_ident() && self.look_ahead(1, |t| matches!(t.kind, token::Eq | token::Colon))
+        {
+            // Parse associated type constraint.
+            let lo = self.token.span;
+            let ident = self.parse_ident()?;
+            let kind = if self.eat(&token::Eq) {
+                let ty = self.parse_assoc_equality_term(ident, self.prev_token.span)?;
+                AssocTyConstraintKind::Equality { ty }
+            } else if self.eat(&token::Colon) {
+                let bounds = self.parse_generic_bounds(Some(self.prev_token.span))?;
+                AssocTyConstraintKind::Bound { bounds }
+            } else {
+                unreachable!();
+            };
+
+            let span = lo.to(self.prev_token.span);
+
+            // Gate associated type bounds, e.g., `Iterator<Item: Ord>`.
+            if let AssocTyConstraintKind::Bound { .. } = kind {
+                self.sess.gated_spans.gate(sym::associated_type_bounds, span);
+            }
+
+            let constraint = AssocTyConstraint { id: ast::DUMMY_NODE_ID, ident, kind, span };
+            Ok(Some(AngleBracketedArg::Constraint(constraint)))
+        } else {
+            Ok(self.parse_generic_arg()?.map(AngleBracketedArg::Arg))
+        }
+    }
+
+    /// Parse the term to the right of an associated item equality constraint.
+    /// That is, parse `<term>` in `Item = <term>`.
+    /// Right now, this only admits types in `<term>`.
+    fn parse_assoc_equality_term(&mut self, ident: Ident, eq: Span) -> PResult<'a, P<ast::Ty>> {
+        let arg = self.parse_generic_arg()?;
+        let span = ident.span.to(self.prev_token.span);
+        match arg {
+            Some(GenericArg::Type(ty)) => return Ok(ty),
+            Some(GenericArg::Const(expr)) => {
+                self.struct_span_err(span, "cannot constrain an associated constant to a value")
+                    .span_label(ident.span, "this associated constant...")
+                    .span_label(expr.value.span, "...cannot be constrained to this value")
+                    .emit();
+            }
+            Some(GenericArg::Lifetime(lt)) => {
+                self.struct_span_err(span, "associated lifetimes are not supported")
+                    .span_label(lt.ident.span, "the lifetime is given here")
+                    .help("if you meant to specify a trait object, write `dyn Trait + 'lifetime`")
+                    .emit();
+            }
+            None => {
+                let after_eq = eq.shrink_to_hi();
+                let before_next = self.token.span.shrink_to_lo();
+                self.struct_span_err(after_eq.to(before_next), "missing type to the right of `=`")
+                    .span_suggestion(
+                        self.sess.source_map().next_point(eq).to(before_next),
+                        "to constrain the associated type, add a type after `=`",
+                        " TheType".to_string(),
+                        Applicability::HasPlaceholders,
+                    )
+                    .span_suggestion(
+                        eq.to(before_next),
+                        &format!("remove the `=` if `{}` is a type", ident),
+                        String::new(),
+                        Applicability::MaybeIncorrect,
+                    )
+                    .emit();
+            }
+        }
+        Ok(self.mk_ty(span, ast::TyKind::Err))
+    }
+
+    /// Parse a generic argument in a path segment.
+    /// This does not include constraints, e.g., `Item = u8`, which is handled in `parse_angle_arg`.
+    fn parse_generic_arg(&mut self) -> PResult<'a, Option<GenericArg>> {
+        let arg = if self.check_lifetime() && self.look_ahead(1, |t| !t.is_like_plus()) {
+            // Parse lifetime argument.
+            GenericArg::Lifetime(self.expect_lifetime())
+        } else if self.check_const_arg() {
+            // Parse const argument.
+            let expr = if let token::OpenDelim(token::Brace) = self.token.kind {
+                self.parse_block_expr(
+                    None,
+                    self.token.span,
+                    BlockCheckMode::Default,
+                    ast::AttrVec::new(),
+                )?
+            } else if self.token.is_ident() {
+                // FIXME(const_generics): to distinguish between idents for types and consts,
+                // we should introduce a GenericArg::Ident in the AST and distinguish when
+                // lowering to the HIR. For now, idents for const args are not permitted.
+                if self.token.is_bool_lit() {
+                    self.parse_literal_maybe_minus()?
+                } else {
+                    let span = self.token.span;
+                    let msg = "identifiers may currently not be used for const generics";
+                    self.struct_span_err(span, msg).emit();
+                    let block = self.mk_block_err(span);
+                    self.mk_expr(span, ast::ExprKind::Block(block, None), ast::AttrVec::new())
+                }
+            } else {
+                self.parse_literal_maybe_minus()?
+            };
+            GenericArg::Const(AnonConst { id: ast::DUMMY_NODE_ID, value: expr })
+        } else if self.check_type() {
+            // Parse type argument.
+            GenericArg::Type(self.parse_ty()?)
+        } else {
+            return Ok(None);
+        };
+        Ok(Some(arg))
+    }
+}

compiler/rustc_parse/src/parser/stmt.rs

@@ -0,0 +1,427 @@
+use super::attr::DEFAULT_INNER_ATTR_FORBIDDEN;
+use super::diagnostics::Error;
+use super::expr::LhsExpr;
+use super::pat::GateOr;
+use super::path::PathStyle;
+use super::{BlockMode, Parser, Restrictions, SemiColonMode};
+use crate::maybe_whole;
+
+use rustc_ast as ast;
+use rustc_ast::ptr::P;
+use rustc_ast::token::{self, TokenKind};
+use rustc_ast::util::classify;
+use rustc_ast::{AttrStyle, AttrVec, Attribute, MacCall, MacStmtStyle};
+use rustc_ast::{Block, BlockCheckMode, Expr, ExprKind, Local, Stmt, StmtKind, DUMMY_NODE_ID};
+use rustc_errors::{Applicability, PResult};
+use rustc_span::source_map::{BytePos, Span};
+use rustc_span::symbol::{kw, sym};
+
+use std::mem;
+
+impl<'a> Parser<'a> {
+    /// Parses a statement. This stops just before trailing semicolons on everything but items.
+    /// e.g., a `StmtKind::Semi` parses to a `StmtKind::Expr`, leaving the trailing `;` unconsumed.
+    pub(super) fn parse_stmt(&mut self) -> PResult<'a, Option<Stmt>> {
+        Ok(self.parse_stmt_without_recovery().unwrap_or_else(|mut e| {
+            e.emit();
+            self.recover_stmt_(SemiColonMode::Break, BlockMode::Ignore);
+            None
+        }))
+    }
+
+    fn parse_stmt_without_recovery(&mut self) -> PResult<'a, Option<Stmt>> {
+        maybe_whole!(self, NtStmt, |x| Some(x));
+
+        let attrs = self.parse_outer_attributes()?;
+        let lo = self.token.span;
+
+        let stmt = if self.eat_keyword(kw::Let) {
+            self.parse_local_mk(lo, attrs.into())?
+        } else if self.is_kw_followed_by_ident(kw::Mut) {
+            self.recover_stmt_local(lo, attrs.into(), "missing keyword", "let mut")?
+        } else if self.is_kw_followed_by_ident(kw::Auto) {
+            self.bump(); // `auto`
+            let msg = "write `let` instead of `auto` to introduce a new variable";
+            self.recover_stmt_local(lo, attrs.into(), msg, "let")?
+        } else if self.is_kw_followed_by_ident(sym::var) {
+            self.bump(); // `var`
+            let msg = "write `let` instead of `var` to introduce a new variable";
+            self.recover_stmt_local(lo, attrs.into(), msg, "let")?
+        } else if self.check_path() && !self.token.is_qpath_start() && !self.is_path_start_item() {
+            // We have avoided contextual keywords like `union`, items with `crate` visibility,
+            // or `auto trait` items. We aim to parse an arbitrary path `a::b` but not something
+            // that starts like a path (1 token), but it fact not a path.
+            // Also, we avoid stealing syntax from `parse_item_`.
+            self.parse_stmt_path_start(lo, attrs)?
+        } else if let Some(item) = self.parse_item_common(attrs.clone(), false, true, |_| true)? {
+            // FIXME: Bad copy of attrs
+            self.mk_stmt(lo.to(item.span), StmtKind::Item(P(item)))
+        } else if self.eat(&token::Semi) {
+            // Do not attempt to parse an expression if we're done here.
+            self.error_outer_attrs(&attrs);
+            self.mk_stmt(lo, StmtKind::Empty)
+        } else if self.token != token::CloseDelim(token::Brace) {
+            // Remainder are line-expr stmts.
+            let e = self.parse_expr_res(Restrictions::STMT_EXPR, Some(attrs.into()))?;
+            self.mk_stmt(lo.to(e.span), StmtKind::Expr(e))
+        } else {
+            self.error_outer_attrs(&attrs);
+            return Ok(None);
+        };
+        Ok(Some(stmt))
+    }
+
+    fn parse_stmt_path_start(&mut self, lo: Span, attrs: Vec<Attribute>) -> PResult<'a, Stmt> {
+        let path = self.parse_path(PathStyle::Expr)?;
+
+        if self.eat(&token::Not) {
+            return self.parse_stmt_mac(lo, attrs.into(), path);
+        }
+
+        let expr = if self.check(&token::OpenDelim(token::Brace)) {
+            self.parse_struct_expr(path, AttrVec::new())?
+        } else {
+            let hi = self.prev_token.span;
+            self.mk_expr(lo.to(hi), ExprKind::Path(None, path), AttrVec::new())
+        };
+
+        let expr = self.with_res(Restrictions::STMT_EXPR, |this| {
+            let expr = this.parse_dot_or_call_expr_with(expr, lo, attrs.into())?;
+            this.parse_assoc_expr_with(0, LhsExpr::AlreadyParsed(expr))
+        })?;
+        Ok(self.mk_stmt(lo.to(self.prev_token.span), StmtKind::Expr(expr)))
+    }
+
+    /// Parses a statement macro `mac!(args)` provided a `path` representing `mac`.
+    /// At this point, the `!` token after the path has already been eaten.
+    fn parse_stmt_mac(&mut self, lo: Span, attrs: AttrVec, path: ast::Path) -> PResult<'a, Stmt> {
+        let args = self.parse_mac_args()?;
+        let delim = args.delim();
+        let hi = self.prev_token.span;
+
+        let style =
+            if delim == token::Brace { MacStmtStyle::Braces } else { MacStmtStyle::NoBraces };
+
+        let mac = MacCall { path, args, prior_type_ascription: self.last_type_ascription };
+
+        let kind = if delim == token::Brace || self.token == token::Semi || self.token == token::Eof
+        {
+            StmtKind::MacCall(P((mac, style, attrs)))
+        } else {
+            // Since none of the above applied, this is an expression statement macro.
+            let e = self.mk_expr(lo.to(hi), ExprKind::MacCall(mac), AttrVec::new());
+            let e = self.maybe_recover_from_bad_qpath(e, true)?;
+            let e = self.parse_dot_or_call_expr_with(e, lo, attrs)?;
+            let e = self.parse_assoc_expr_with(0, LhsExpr::AlreadyParsed(e))?;
+            StmtKind::Expr(e)
+        };
+        Ok(self.mk_stmt(lo.to(hi), kind))
+    }
+
+    /// Error on outer attributes in this context.
+    /// Also error if the previous token was a doc comment.
+    fn error_outer_attrs(&self, attrs: &[Attribute]) {
+        if let [.., last] = attrs {
+            if last.is_doc_comment() {
+                self.span_fatal_err(last.span, Error::UselessDocComment).emit();
+            } else if attrs.iter().any(|a| a.style == AttrStyle::Outer) {
+                self.struct_span_err(last.span, "expected statement after outer attribute").emit();
+            }
+        }
+    }
+
+    fn recover_stmt_local(
+        &mut self,
+        lo: Span,
+        attrs: AttrVec,
+        msg: &str,
+        sugg: &str,
+    ) -> PResult<'a, Stmt> {
+        let stmt = self.parse_local_mk(lo, attrs)?;
+        self.struct_span_err(lo, "invalid variable declaration")
+            .span_suggestion(lo, msg, sugg.to_string(), Applicability::MachineApplicable)
+            .emit();
+        Ok(stmt)
+    }
+
+    fn parse_local_mk(&mut self, lo: Span, attrs: AttrVec) -> PResult<'a, Stmt> {
+        let local = self.parse_local(attrs)?;
+        Ok(self.mk_stmt(lo.to(self.prev_token.span), StmtKind::Local(local)))
+    }
+
+    /// Parses a local variable declaration.
+    fn parse_local(&mut self, attrs: AttrVec) -> PResult<'a, P<Local>> {
+        let lo = self.prev_token.span;
+        let pat = self.parse_top_pat(GateOr::Yes)?;
+
+        let (err, ty) = if self.eat(&token::Colon) {
+            // Save the state of the parser before parsing type normally, in case there is a `:`
+            // instead of an `=` typo.
+            let parser_snapshot_before_type = self.clone();
+            let colon_sp = self.prev_token.span;
+            match self.parse_ty() {
+                Ok(ty) => (None, Some(ty)),
+                Err(mut err) => {
+                    if let Ok(snip) = self.span_to_snippet(pat.span) {
+                        err.span_label(pat.span, format!("while parsing the type for `{}`", snip));
+                    }
+                    let err = if self.check(&token::Eq) {
+                        err.emit();
+                        None
+                    } else {
+                        // Rewind to before attempting to parse the type and continue parsing.
+                        let parser_snapshot_after_type =
+                            mem::replace(self, parser_snapshot_before_type);
+                        Some((parser_snapshot_after_type, colon_sp, err))
+                    };
+                    (err, None)
+                }
+            }
+        } else {
+            (None, None)
+        };
+        let init = match (self.parse_initializer(err.is_some()), err) {
+            (Ok(init), None) => {
+                // init parsed, ty parsed
+                init
+            }
+            (Ok(init), Some((_, colon_sp, mut err))) => {
+                // init parsed, ty error
+                // Could parse the type as if it were the initializer, it is likely there was a
+                // typo in the code: `:` instead of `=`. Add suggestion and emit the error.
+                err.span_suggestion_short(
+                    colon_sp,
+                    "use `=` if you meant to assign",
+                    " =".to_string(),
+                    Applicability::MachineApplicable,
+                );
+                err.emit();
+                // As this was parsed successfully, continue as if the code has been fixed for the
+                // rest of the file. It will still fail due to the emitted error, but we avoid
+                // extra noise.
+                init
+            }
+            (Err(mut init_err), Some((snapshot, _, ty_err))) => {
+                // init error, ty error
+                init_err.cancel();
+                // Couldn't parse the type nor the initializer, only raise the type error and
+                // return to the parser state before parsing the type as the initializer.
+                // let x: <parse_error>;
+                *self = snapshot;
+                return Err(ty_err);
+            }
+            (Err(err), None) => {
+                // init error, ty parsed
+                // Couldn't parse the initializer and we're not attempting to recover a failed
+                // parse of the type, return the error.
+                return Err(err);
+            }
+        };
+        let hi = if self.token == token::Semi { self.token.span } else { self.prev_token.span };
+        Ok(P(ast::Local { ty, pat, init, id: DUMMY_NODE_ID, span: lo.to(hi), attrs }))
+    }
+
+    /// Parses the RHS of a local variable declaration (e.g., '= 14;').
+    fn parse_initializer(&mut self, eq_optional: bool) -> PResult<'a, Option<P<Expr>>> {
+        let eq_consumed = match self.token.kind {
+            token::BinOpEq(..) => {
+                // Recover `let x <op>= 1` as `let x = 1`
+                self.struct_span_err(
+                    self.token.span,
+                    "can't reassign to an uninitialized variable",
+                )
+                .span_suggestion_short(
+                    self.token.span,
+                    "initialize the variable",
+                    "=".to_string(),
+                    Applicability::MaybeIncorrect,
+                )
+                .emit();
+                self.bump();
+                true
+            }
+            _ => self.eat(&token::Eq),
+        };
+
+        Ok(if eq_consumed || eq_optional { Some(self.parse_expr()?) } else { None })
+    }
+
+    /// Parses a block. No inner attributes are allowed.
+    pub(super) fn parse_block(&mut self) -> PResult<'a, P<Block>> {
+        let (attrs, block) = self.parse_inner_attrs_and_block()?;
+        if let [.., last] = &*attrs {
+            self.error_on_forbidden_inner_attr(last.span, DEFAULT_INNER_ATTR_FORBIDDEN);
+        }
+        Ok(block)
+    }
+
+    fn error_block_no_opening_brace<T>(&mut self) -> PResult<'a, T> {
+        let sp = self.token.span;
+        let tok = super::token_descr(&self.token);
+        let mut e = self.struct_span_err(sp, &format!("expected `{{`, found {}", tok));
+        let do_not_suggest_help = self.token.is_keyword(kw::In) || self.token == token::Colon;
+
+        // Check to see if the user has written something like
+        //
+        //    if (cond)
+        //      bar;
+        //
+        // which is valid in other languages, but not Rust.
+        match self.parse_stmt_without_recovery() {
+            // If the next token is an open brace (e.g., `if a b {`), the place-
+            // inside-a-block suggestion would be more likely wrong than right.
+            Ok(Some(_))
+                if self.look_ahead(1, |t| t == &token::OpenDelim(token::Brace))
+                    || do_not_suggest_help => {}
+            Ok(Some(stmt)) => {
+                let stmt_own_line = self.sess.source_map().is_line_before_span_empty(sp);
+                let stmt_span = if stmt_own_line && self.eat(&token::Semi) {
+                    // Expand the span to include the semicolon.
+                    stmt.span.with_hi(self.prev_token.span.hi())
+                } else {
+                    stmt.span
+                };
+                if let Ok(snippet) = self.span_to_snippet(stmt_span) {
+                    e.span_suggestion(
+                        stmt_span,
+                        "try placing this code inside a block",
+                        format!("{{ {} }}", snippet),
+                        // Speculative; has been misleading in the past (#46836).
+                        Applicability::MaybeIncorrect,
+                    );
+                }
+            }
+            Err(mut e) => {
+                self.recover_stmt_(SemiColonMode::Break, BlockMode::Ignore);
+                e.cancel();
+            }
+            _ => {}
+        }
+        e.span_label(sp, "expected `{`");
+        Err(e)
+    }
+
+    /// Parses a block. Inner attributes are allowed.
+    pub(super) fn parse_inner_attrs_and_block(
+        &mut self,
+    ) -> PResult<'a, (Vec<Attribute>, P<Block>)> {
+        self.parse_block_common(self.token.span, BlockCheckMode::Default)
+    }
+
+    /// Parses a block. Inner attributes are allowed.
+    pub(super) fn parse_block_common(
+        &mut self,
+        lo: Span,
+        blk_mode: BlockCheckMode,
+    ) -> PResult<'a, (Vec<Attribute>, P<Block>)> {
+        maybe_whole!(self, NtBlock, |x| (Vec::new(), x));
+
+        if !self.eat(&token::OpenDelim(token::Brace)) {
+            return self.error_block_no_opening_brace();
+        }
+
+        Ok((self.parse_inner_attributes()?, self.parse_block_tail(lo, blk_mode)?))
+    }
+
+    /// Parses the rest of a block expression or function body.
+    /// Precondition: already parsed the '{'.
+    fn parse_block_tail(&mut self, lo: Span, s: BlockCheckMode) -> PResult<'a, P<Block>> {
+        let mut stmts = vec![];
+        while !self.eat(&token::CloseDelim(token::Brace)) {
+            if self.token == token::Eof {
+                break;
+            }
+            let stmt = match self.parse_full_stmt() {
+                Err(mut err) => {
+                    self.maybe_annotate_with_ascription(&mut err, false);
+                    err.emit();
+                    self.recover_stmt_(SemiColonMode::Ignore, BlockMode::Ignore);
+                    Some(self.mk_stmt_err(self.token.span))
+                }
+                Ok(stmt) => stmt,
+            };
+            if let Some(stmt) = stmt {
+                stmts.push(stmt);
+            } else {
+                // Found only `;` or `}`.
+                continue;
+            };
+        }
+        Ok(self.mk_block(stmts, s, lo.to(self.prev_token.span)))
+    }
+
+    /// Parses a statement, including the trailing semicolon.
+    pub fn parse_full_stmt(&mut self) -> PResult<'a, Option<Stmt>> {
+        // Skip looking for a trailing semicolon when we have an interpolated statement.
+        maybe_whole!(self, NtStmt, |x| Some(x));
+
+        let mut stmt = match self.parse_stmt_without_recovery()? {
+            Some(stmt) => stmt,
+            None => return Ok(None),
+        };
+
+        let mut eat_semi = true;
+        match stmt.kind {
+            // Expression without semicolon.
+            StmtKind::Expr(ref expr)
+                if self.token != token::Eof && classify::expr_requires_semi_to_be_stmt(expr) =>
+            {
+                // Just check for errors and recover; do not eat semicolon yet.
+                if let Err(mut e) =
+                    self.expect_one_of(&[], &[token::Semi, token::CloseDelim(token::Brace)])
+                {
+                    if let TokenKind::DocComment(..) = self.token.kind {
+                        if let Ok(snippet) = self.span_to_snippet(self.token.span) {
+                            let sp = self.token.span;
+                            let marker = &snippet[..3];
+                            let (comment_marker, doc_comment_marker) = marker.split_at(2);
+
+                            e.span_suggestion(
+                                sp.with_hi(sp.lo() + BytePos(marker.len() as u32)),
+                                &format!(
+                                    "add a space before `{}` to use a regular comment",
+                                    doc_comment_marker,
+                                ),
+                                format!("{} {}", comment_marker, doc_comment_marker),
+                                Applicability::MaybeIncorrect,
+                            );
+                        }
+                    }
+                    e.emit();
+                    self.recover_stmt();
+                    // Don't complain about type errors in body tail after parse error (#57383).
+                    let sp = expr.span.to(self.prev_token.span);
+                    stmt.kind = StmtKind::Expr(self.mk_expr_err(sp));
+                }
+            }
+            StmtKind::Local(..) => {
+                self.expect_semi()?;
+                eat_semi = false;
+            }
+            StmtKind::Empty => eat_semi = false,
+            _ => {}
+        }
+
+        if eat_semi && self.eat(&token::Semi) {
+            stmt = stmt.add_trailing_semicolon();
+        }
+        stmt.span = stmt.span.to(self.prev_token.span);
+        Ok(Some(stmt))
+    }
+
+    pub(super) fn mk_block(&self, stmts: Vec<Stmt>, rules: BlockCheckMode, span: Span) -> P<Block> {
+        P(Block { stmts, id: DUMMY_NODE_ID, rules, span })
+    }
+
+    pub(super) fn mk_stmt(&self, span: Span, kind: StmtKind) -> Stmt {
+        Stmt { id: DUMMY_NODE_ID, kind, span }
+    }
+
+    fn mk_stmt_err(&self, span: Span) -> Stmt {
+        self.mk_stmt(span, StmtKind::Expr(self.mk_expr_err(span)))
+    }
+
+    pub(super) fn mk_block_err(&self, span: Span) -> P<Block> {
+        self.mk_block(vec![self.mk_stmt_err(span)], BlockCheckMode::Default, span)
+    }
+}

compiler/rustc_parse/src/parser/ty.rs

@@ -0,0 +1,631 @@
+use super::{Parser, PathStyle, TokenType};
+
+use crate::{maybe_recover_from_interpolated_ty_qpath, maybe_whole};
+
+use rustc_ast::ptr::P;
+use rustc_ast::token::{self, Token, TokenKind};
+use rustc_ast::{self as ast, BareFnTy, FnRetTy, GenericParam, Lifetime, MutTy, Ty, TyKind};
+use rustc_ast::{GenericBound, GenericBounds, MacCall, Mutability};
+use rustc_ast::{PolyTraitRef, TraitBoundModifier, TraitObjectSyntax};
+use rustc_errors::{pluralize, struct_span_err, Applicability, PResult};
+use rustc_span::source_map::Span;
+use rustc_span::symbol::{kw, sym};
+
+/// Any `?` or `?const` modifiers that appear at the start of a bound.
+struct BoundModifiers {
+    /// `?Trait`.
+    maybe: Option<Span>,
+
+    /// `?const Trait`.
+    maybe_const: Option<Span>,
+}
+
+impl BoundModifiers {
+    fn to_trait_bound_modifier(&self) -> TraitBoundModifier {
+        match (self.maybe, self.maybe_const) {
+            (None, None) => TraitBoundModifier::None,
+            (Some(_), None) => TraitBoundModifier::Maybe,
+            (None, Some(_)) => TraitBoundModifier::MaybeConst,
+            (Some(_), Some(_)) => TraitBoundModifier::MaybeConstMaybe,
+        }
+    }
+}
+
+#[derive(Copy, Clone, PartialEq)]
+pub(super) enum AllowPlus {
+    Yes,
+    No,
+}
+
+#[derive(PartialEq)]
+pub(super) enum RecoverQPath {
+    Yes,
+    No,
+}
+
+// Is `...` (`CVarArgs`) legal at this level of type parsing?
+#[derive(PartialEq)]
+enum AllowCVariadic {
+    Yes,
+    No,
+}
+
+/// Returns `true` if `IDENT t` can start a type -- `IDENT::a::b`, `IDENT<u8, u8>`,
+/// `IDENT<<u8 as Trait>::AssocTy>`.
+///
+/// Types can also be of the form `IDENT(u8, u8) -> u8`, however this assumes
+/// that `IDENT` is not the ident of a fn trait.
+fn can_continue_type_after_non_fn_ident(t: &Token) -> bool {
+    t == &token::ModSep || t == &token::Lt || t == &token::BinOp(token::Shl)
+}
+
+impl<'a> Parser<'a> {
+    /// Parses a type.
+    pub fn parse_ty(&mut self) -> PResult<'a, P<Ty>> {
+        self.parse_ty_common(AllowPlus::Yes, RecoverQPath::Yes, AllowCVariadic::No)
+    }
+
+    /// Parse a type suitable for a function or function pointer parameter.
+    /// The difference from `parse_ty` is that this version allows `...`
+    /// (`CVarArgs`) at the top level of the the type.
+    pub(super) fn parse_ty_for_param(&mut self) -> PResult<'a, P<Ty>> {
+        self.parse_ty_common(AllowPlus::Yes, RecoverQPath::Yes, AllowCVariadic::Yes)
+    }
+
+    /// Parses a type in restricted contexts where `+` is not permitted.
+    ///
+    /// Example 1: `&'a TYPE`
+    ///     `+` is prohibited to maintain operator priority (P(+) < P(&)).
+    /// Example 2: `value1 as TYPE + value2`
+    ///     `+` is prohibited to avoid interactions with expression grammar.
+    pub(super) fn parse_ty_no_plus(&mut self) -> PResult<'a, P<Ty>> {
+        self.parse_ty_common(AllowPlus::No, RecoverQPath::Yes, AllowCVariadic::No)
+    }
+
+    /// Parses an optional return type `[ -> TY ]` in a function declaration.
+    pub(super) fn parse_ret_ty(
+        &mut self,
+        allow_plus: AllowPlus,
+        recover_qpath: RecoverQPath,
+    ) -> PResult<'a, FnRetTy> {
+        Ok(if self.eat(&token::RArrow) {
+            // FIXME(Centril): Can we unconditionally `allow_plus`?
+            let ty = self.parse_ty_common(allow_plus, recover_qpath, AllowCVariadic::No)?;
+            FnRetTy::Ty(ty)
+        } else {
+            FnRetTy::Default(self.token.span.shrink_to_lo())
+        })
+    }
+
+    fn parse_ty_common(
+        &mut self,
+        allow_plus: AllowPlus,
+        recover_qpath: RecoverQPath,
+        allow_c_variadic: AllowCVariadic,
+    ) -> PResult<'a, P<Ty>> {
+        let allow_qpath_recovery = recover_qpath == RecoverQPath::Yes;
+        maybe_recover_from_interpolated_ty_qpath!(self, allow_qpath_recovery);
+        maybe_whole!(self, NtTy, |x| x);
+
+        let lo = self.token.span;
+        let mut impl_dyn_multi = false;
+        let kind = if self.check(&token::OpenDelim(token::Paren)) {
+            self.parse_ty_tuple_or_parens(lo, allow_plus)?
+        } else if self.eat(&token::Not) {
+            // Never type `!`
+            TyKind::Never
+        } else if self.eat(&token::BinOp(token::Star)) {
+            self.parse_ty_ptr()?
+        } else if self.eat(&token::OpenDelim(token::Bracket)) {
+            self.parse_array_or_slice_ty()?
+        } else if self.check(&token::BinOp(token::And)) || self.check(&token::AndAnd) {
+            // Reference
+            self.expect_and()?;
+            self.parse_borrowed_pointee()?
+        } else if self.eat_keyword_noexpect(kw::Typeof) {
+            self.parse_typeof_ty()?
+        } else if self.eat_keyword(kw::Underscore) {
+            // A type to be inferred `_`
+            TyKind::Infer
+        } else if self.check_fn_front_matter() {
+            // Function pointer type
+            self.parse_ty_bare_fn(lo, Vec::new())?
+        } else if self.check_keyword(kw::For) {
+            // Function pointer type or bound list (trait object type) starting with a poly-trait.
+            //   `for<'lt> [unsafe] [extern "ABI"] fn (&'lt S) -> T`
+            //   `for<'lt> Trait1<'lt> + Trait2 + 'a`
+            let lifetime_defs = self.parse_late_bound_lifetime_defs()?;
+            if self.check_fn_front_matter() {
+                self.parse_ty_bare_fn(lo, lifetime_defs)?
+            } else {
+                let path = self.parse_path(PathStyle::Type)?;
+                let parse_plus = allow_plus == AllowPlus::Yes && self.check_plus();
+                self.parse_remaining_bounds_path(lifetime_defs, path, lo, parse_plus)?
+            }
+        } else if self.eat_keyword(kw::Impl) {
+            self.parse_impl_ty(&mut impl_dyn_multi)?
+        } else if self.is_explicit_dyn_type() {
+            self.parse_dyn_ty(&mut impl_dyn_multi)?
+        } else if self.eat_lt() {
+            // Qualified path
+            let (qself, path) = self.parse_qpath(PathStyle::Type)?;
+            TyKind::Path(Some(qself), path)
+        } else if self.check_path() {
+            self.parse_path_start_ty(lo, allow_plus)?
+        } else if self.can_begin_bound() {
+            self.parse_bare_trait_object(lo, allow_plus)?
+        } else if self.eat(&token::DotDotDot) {
+            if allow_c_variadic == AllowCVariadic::Yes {
+                TyKind::CVarArgs
+            } else {
+                // FIXME(Centril): Should we just allow `...` syntactically
+                // anywhere in a type and use semantic restrictions instead?
+                self.error_illegal_c_varadic_ty(lo);
+                TyKind::Err
+            }
+        } else {
+            let msg = format!("expected type, found {}", super::token_descr(&self.token));
+            let mut err = self.struct_span_err(self.token.span, &msg);
+            err.span_label(self.token.span, "expected type");
+            self.maybe_annotate_with_ascription(&mut err, true);
+            return Err(err);
+        };
+
+        let span = lo.to(self.prev_token.span);
+        let ty = self.mk_ty(span, kind);
+
+        // Try to recover from use of `+` with incorrect priority.
+        self.maybe_report_ambiguous_plus(allow_plus, impl_dyn_multi, &ty);
+        self.maybe_recover_from_bad_type_plus(allow_plus, &ty)?;
+        self.maybe_recover_from_bad_qpath(ty, allow_qpath_recovery)
+    }
+
+    /// Parses either:
+    /// - `(TYPE)`, a parenthesized type.
+    /// - `(TYPE,)`, a tuple with a single field of type TYPE.
+    fn parse_ty_tuple_or_parens(&mut self, lo: Span, allow_plus: AllowPlus) -> PResult<'a, TyKind> {
+        let mut trailing_plus = false;
+        let (ts, trailing) = self.parse_paren_comma_seq(|p| {
+            let ty = p.parse_ty()?;
+            trailing_plus = p.prev_token.kind == TokenKind::BinOp(token::Plus);
+            Ok(ty)
+        })?;
+
+        if ts.len() == 1 && !trailing {
+            let ty = ts.into_iter().next().unwrap().into_inner();
+            let maybe_bounds = allow_plus == AllowPlus::Yes && self.token.is_like_plus();
+            match ty.kind {
+                // `(TY_BOUND_NOPAREN) + BOUND + ...`.
+                TyKind::Path(None, path) if maybe_bounds => {
+                    self.parse_remaining_bounds_path(Vec::new(), path, lo, true)
+                }
+                TyKind::TraitObject(bounds, TraitObjectSyntax::None)
+                    if maybe_bounds && bounds.len() == 1 && !trailing_plus =>
+                {
+                    self.parse_remaining_bounds(bounds, true)
+                }
+                // `(TYPE)`
+                _ => Ok(TyKind::Paren(P(ty))),
+            }
+        } else {
+            Ok(TyKind::Tup(ts))
+        }
+    }
+
+    fn parse_bare_trait_object(&mut self, lo: Span, allow_plus: AllowPlus) -> PResult<'a, TyKind> {
+        let lt_no_plus = self.check_lifetime() && !self.look_ahead(1, |t| t.is_like_plus());
+        let bounds = self.parse_generic_bounds_common(allow_plus, None)?;
+        if lt_no_plus {
+            self.struct_span_err(lo, "lifetime in trait object type must be followed by `+`").emit()
+        }
+        Ok(TyKind::TraitObject(bounds, TraitObjectSyntax::None))
+    }
+
+    fn parse_remaining_bounds_path(
+        &mut self,
+        generic_params: Vec<GenericParam>,
+        path: ast::Path,
+        lo: Span,
+        parse_plus: bool,
+    ) -> PResult<'a, TyKind> {
+        let poly_trait_ref = PolyTraitRef::new(generic_params, path, lo.to(self.prev_token.span));
+        let bounds = vec![GenericBound::Trait(poly_trait_ref, TraitBoundModifier::None)];
+        self.parse_remaining_bounds(bounds, parse_plus)
+    }
+
+    /// Parse the remainder of a bare trait object type given an already parsed list.
+    fn parse_remaining_bounds(
+        &mut self,
+        mut bounds: GenericBounds,
+        plus: bool,
+    ) -> PResult<'a, TyKind> {
+        assert_ne!(self.token, token::Question);
+        if plus {
+            self.eat_plus(); // `+`, or `+=` gets split and `+` is discarded
+            bounds.append(&mut self.parse_generic_bounds(Some(self.prev_token.span))?);
+        }
+        Ok(TyKind::TraitObject(bounds, TraitObjectSyntax::None))
+    }
+
+    /// Parses a raw pointer type: `*[const | mut] $type`.
+    fn parse_ty_ptr(&mut self) -> PResult<'a, TyKind> {
+        let mutbl = self.parse_const_or_mut().unwrap_or_else(|| {
+            let span = self.prev_token.span;
+            let msg = "expected mut or const in raw pointer type";
+            self.struct_span_err(span, msg)
+                .span_label(span, msg)
+                .help("use `*mut T` or `*const T` as appropriate")
+                .emit();
+            Mutability::Not
+        });
+        let ty = self.parse_ty_no_plus()?;
+        Ok(TyKind::Ptr(MutTy { ty, mutbl }))
+    }
+
+    /// Parses an array (`[TYPE; EXPR]`) or slice (`[TYPE]`) type.
+    /// The opening `[` bracket is already eaten.
+    fn parse_array_or_slice_ty(&mut self) -> PResult<'a, TyKind> {
+        let elt_ty = self.parse_ty()?;
+        let ty = if self.eat(&token::Semi) {
+            TyKind::Array(elt_ty, self.parse_anon_const_expr()?)
+        } else {
+            TyKind::Slice(elt_ty)
+        };
+        self.expect(&token::CloseDelim(token::Bracket))?;
+        Ok(ty)
+    }
+
+    fn parse_borrowed_pointee(&mut self) -> PResult<'a, TyKind> {
+        let opt_lifetime = if self.check_lifetime() { Some(self.expect_lifetime()) } else { None };
+        let mutbl = self.parse_mutability();
+        let ty = self.parse_ty_no_plus()?;
+        Ok(TyKind::Rptr(opt_lifetime, MutTy { ty, mutbl }))
+    }
+
+    // Parses the `typeof(EXPR)`.
+    // To avoid ambiguity, the type is surrounded by parenthesis.
+    fn parse_typeof_ty(&mut self) -> PResult<'a, TyKind> {
+        self.expect(&token::OpenDelim(token::Paren))?;
+        let expr = self.parse_anon_const_expr()?;
+        self.expect(&token::CloseDelim(token::Paren))?;
+        Ok(TyKind::Typeof(expr))
+    }
+
+    /// Parses a function pointer type (`TyKind::BareFn`).
+    /// ```
+    /// [unsafe] [extern "ABI"] fn (S) -> T
+    ///  ^~~~~^          ^~~~^     ^~^    ^
+    ///    |               |        |     |
+    ///    |               |        |   Return type
+    /// Function Style    ABI  Parameter types
+    /// ```
+    /// We actually parse `FnHeader FnDecl`, but we error on `const` and `async` qualifiers.
+    fn parse_ty_bare_fn(&mut self, lo: Span, params: Vec<GenericParam>) -> PResult<'a, TyKind> {
+        let ast::FnHeader { ext, unsafety, constness, asyncness } = self.parse_fn_front_matter()?;
+        let decl = self.parse_fn_decl(|_| false, AllowPlus::No)?;
+        let whole_span = lo.to(self.prev_token.span);
+        if let ast::Const::Yes(span) = constness {
+            self.error_fn_ptr_bad_qualifier(whole_span, span, "const");
+        }
+        if let ast::Async::Yes { span, .. } = asyncness {
+            self.error_fn_ptr_bad_qualifier(whole_span, span, "async");
+        }
+        Ok(TyKind::BareFn(P(BareFnTy { ext, unsafety, generic_params: params, decl })))
+    }
+
+    /// Emit an error for the given bad function pointer qualifier.
+    fn error_fn_ptr_bad_qualifier(&self, span: Span, qual_span: Span, qual: &str) {
+        self.struct_span_err(span, &format!("an `fn` pointer type cannot be `{}`", qual))
+            .span_label(qual_span, format!("`{}` because of this", qual))
+            .span_suggestion_short(
+                qual_span,
+                &format!("remove the `{}` qualifier", qual),
+                String::new(),
+                Applicability::MaybeIncorrect,
+            )
+            .emit();
+    }
+
+    /// Parses an `impl B0 + ... + Bn` type.
+    fn parse_impl_ty(&mut self, impl_dyn_multi: &mut bool) -> PResult<'a, TyKind> {
+        // Always parse bounds greedily for better error recovery.
+        let bounds = self.parse_generic_bounds(None)?;
+        *impl_dyn_multi = bounds.len() > 1 || self.prev_token.kind == TokenKind::BinOp(token::Plus);
+        Ok(TyKind::ImplTrait(ast::DUMMY_NODE_ID, bounds))
+    }
+
+    /// Is a `dyn B0 + ... + Bn` type allowed here?
+    fn is_explicit_dyn_type(&mut self) -> bool {
+        self.check_keyword(kw::Dyn)
+            && (self.token.uninterpolated_span().rust_2018()
+                || self.look_ahead(1, |t| {
+                    t.can_begin_bound() && !can_continue_type_after_non_fn_ident(t)
+                }))
+    }
+
+    /// Parses a `dyn B0 + ... + Bn` type.
+    ///
+    /// Note that this does *not* parse bare trait objects.
+    fn parse_dyn_ty(&mut self, impl_dyn_multi: &mut bool) -> PResult<'a, TyKind> {
+        self.bump(); // `dyn`
+        // Always parse bounds greedily for better error recovery.
+        let bounds = self.parse_generic_bounds(None)?;
+        *impl_dyn_multi = bounds.len() > 1 || self.prev_token.kind == TokenKind::BinOp(token::Plus);
+        Ok(TyKind::TraitObject(bounds, TraitObjectSyntax::Dyn))
+    }
+
+    /// Parses a type starting with a path.
+    ///
+    /// This can be:
+    /// 1. a type macro, `mac!(...)`,
+    /// 2. a bare trait object, `B0 + ... + Bn`,
+    /// 3. or a path, `path::to::MyType`.
+    fn parse_path_start_ty(&mut self, lo: Span, allow_plus: AllowPlus) -> PResult<'a, TyKind> {
+        // Simple path
+        let path = self.parse_path(PathStyle::Type)?;
+        if self.eat(&token::Not) {
+            // Macro invocation in type position
+            Ok(TyKind::MacCall(MacCall {
+                path,
+                args: self.parse_mac_args()?,
+                prior_type_ascription: self.last_type_ascription,
+            }))
+        } else if allow_plus == AllowPlus::Yes && self.check_plus() {
+            // `Trait1 + Trait2 + 'a`
+            self.parse_remaining_bounds_path(Vec::new(), path, lo, true)
+        } else {
+            // Just a type path.
+            Ok(TyKind::Path(None, path))
+        }
+    }
+
+    fn error_illegal_c_varadic_ty(&self, lo: Span) {
+        struct_span_err!(
+            self.sess.span_diagnostic,
+            lo.to(self.prev_token.span),
+            E0743,
+            "C-variadic type `...` may not be nested inside another type",
+        )
+        .emit();
+    }
+
+    pub(super) fn parse_generic_bounds(
+        &mut self,
+        colon_span: Option<Span>,
+    ) -> PResult<'a, GenericBounds> {
+        self.parse_generic_bounds_common(AllowPlus::Yes, colon_span)
+    }
+
+    /// Parses bounds of a type parameter `BOUND + BOUND + ...`, possibly with trailing `+`.
+    ///
+    /// See `parse_generic_bound` for the `BOUND` grammar.
+    fn parse_generic_bounds_common(
+        &mut self,
+        allow_plus: AllowPlus,
+        colon_span: Option<Span>,
+    ) -> PResult<'a, GenericBounds> {
+        let mut bounds = Vec::new();
+        let mut negative_bounds = Vec::new();
+        while self.can_begin_bound() {
+            match self.parse_generic_bound()? {
+                Ok(bound) => bounds.push(bound),
+                Err(neg_sp) => negative_bounds.push(neg_sp),
+            }
+            if allow_plus == AllowPlus::No || !self.eat_plus() {
+                break;
+            }
+        }
+
+        if !negative_bounds.is_empty() {
+            self.error_negative_bounds(colon_span, &bounds, negative_bounds);
+        }
+
+        Ok(bounds)
+    }
+
+    /// Can the current token begin a bound?
+    fn can_begin_bound(&mut self) -> bool {
+        // This needs to be synchronized with `TokenKind::can_begin_bound`.
+        self.check_path()
+        || self.check_lifetime()
+        || self.check(&token::Not) // Used for error reporting only.
+        || self.check(&token::Question)
+        || self.check_keyword(kw::For)
+        || self.check(&token::OpenDelim(token::Paren))
+    }
+
+    fn error_negative_bounds(
+        &self,
+        colon_span: Option<Span>,
+        bounds: &[GenericBound],
+        negative_bounds: Vec<Span>,
+    ) {
+        let negative_bounds_len = negative_bounds.len();
+        let last_span = *negative_bounds.last().expect("no negative bounds, but still error?");
+        let mut err = self.struct_span_err(negative_bounds, "negative bounds are not supported");
+        err.span_label(last_span, "negative bounds are not supported");
+        if let Some(bound_list) = colon_span {
+            let bound_list = bound_list.to(self.prev_token.span);
+            let mut new_bound_list = String::new();
+            if !bounds.is_empty() {
+                let mut snippets = bounds.iter().map(|bound| self.span_to_snippet(bound.span()));
+                while let Some(Ok(snippet)) = snippets.next() {
+                    new_bound_list.push_str(" + ");
+                    new_bound_list.push_str(&snippet);
+                }
+                new_bound_list = new_bound_list.replacen(" +", ":", 1);
+            }
+            err.tool_only_span_suggestion(
+                bound_list,
+                &format!("remove the bound{}", pluralize!(negative_bounds_len)),
+                new_bound_list,
+                Applicability::MachineApplicable,
+            );
+        }
+        err.emit();
+    }
+
+    /// Parses a bound according to the grammar:
+    /// ```
+    /// BOUND = TY_BOUND | LT_BOUND
+    /// ```
+    fn parse_generic_bound(&mut self) -> PResult<'a, Result<GenericBound, Span>> {
+        let anchor_lo = self.prev_token.span;
+        let lo = self.token.span;
+        let has_parens = self.eat(&token::OpenDelim(token::Paren));
+        let inner_lo = self.token.span;
+        let is_negative = self.eat(&token::Not);
+
+        let modifiers = self.parse_ty_bound_modifiers();
+        let bound = if self.token.is_lifetime() {
+            self.error_lt_bound_with_modifiers(modifiers);
+            self.parse_generic_lt_bound(lo, inner_lo, has_parens)?
+        } else {
+            self.parse_generic_ty_bound(lo, has_parens, modifiers)?
+        };
+
+        Ok(if is_negative { Err(anchor_lo.to(self.prev_token.span)) } else { Ok(bound) })
+    }
+
+    /// Parses a lifetime ("outlives") bound, e.g. `'a`, according to:
+    /// ```
+    /// LT_BOUND = LIFETIME
+    /// ```
+    fn parse_generic_lt_bound(
+        &mut self,
+        lo: Span,
+        inner_lo: Span,
+        has_parens: bool,
+    ) -> PResult<'a, GenericBound> {
+        let bound = GenericBound::Outlives(self.expect_lifetime());
+        if has_parens {
+            // FIXME(Centril): Consider not erroring here and accepting `('lt)` instead,
+            // possibly introducing `GenericBound::Paren(P<GenericBound>)`?
+            self.recover_paren_lifetime(lo, inner_lo)?;
+        }
+        Ok(bound)
+    }
+
+    /// Emits an error if any trait bound modifiers were present.
+    fn error_lt_bound_with_modifiers(&self, modifiers: BoundModifiers) {
+        if let Some(span) = modifiers.maybe_const {
+            self.struct_span_err(
+                span,
+                "`?const` may only modify trait bounds, not lifetime bounds",
+            )
+            .emit();
+        }
+
+        if let Some(span) = modifiers.maybe {
+            self.struct_span_err(span, "`?` may only modify trait bounds, not lifetime bounds")
+                .emit();
+        }
+    }
+
+    /// Recover on `('lifetime)` with `(` already eaten.
+    fn recover_paren_lifetime(&mut self, lo: Span, inner_lo: Span) -> PResult<'a, ()> {
+        let inner_span = inner_lo.to(self.prev_token.span);
+        self.expect(&token::CloseDelim(token::Paren))?;
+        let mut err = self.struct_span_err(
+            lo.to(self.prev_token.span),
+            "parenthesized lifetime bounds are not supported",
+        );
+        if let Ok(snippet) = self.span_to_snippet(inner_span) {
+            err.span_suggestion_short(
+                lo.to(self.prev_token.span),
+                "remove the parentheses",
+                snippet,
+                Applicability::MachineApplicable,
+            );
+        }
+        err.emit();
+        Ok(())
+    }
+
+    /// Parses the modifiers that may precede a trait in a bound, e.g. `?Trait` or `?const Trait`.
+    ///
+    /// If no modifiers are present, this does not consume any tokens.
+    ///
+    /// ```
+    /// TY_BOUND_MODIFIERS = "?" ["const" ["?"]]
+    /// ```
+    fn parse_ty_bound_modifiers(&mut self) -> BoundModifiers {
+        if !self.eat(&token::Question) {
+            return BoundModifiers { maybe: None, maybe_const: None };
+        }
+
+        // `? ...`
+        let first_question = self.prev_token.span;
+        if !self.eat_keyword(kw::Const) {
+            return BoundModifiers { maybe: Some(first_question), maybe_const: None };
+        }
+
+        // `?const ...`
+        let maybe_const = first_question.to(self.prev_token.span);
+        self.sess.gated_spans.gate(sym::const_trait_bound_opt_out, maybe_const);
+        if !self.eat(&token::Question) {
+            return BoundModifiers { maybe: None, maybe_const: Some(maybe_const) };
+        }
+
+        // `?const ? ...`
+        let second_question = self.prev_token.span;
+        BoundModifiers { maybe: Some(second_question), maybe_const: Some(maybe_const) }
+    }
+
+    /// Parses a type bound according to:
+    /// ```
+    /// TY_BOUND = TY_BOUND_NOPAREN | (TY_BOUND_NOPAREN)
+    /// TY_BOUND_NOPAREN = [TY_BOUND_MODIFIERS] [for<LT_PARAM_DEFS>] SIMPLE_PATH
+    /// ```
+    ///
+    /// For example, this grammar accepts `?const ?for<'a: 'b> m::Trait<'a>`.
+    fn parse_generic_ty_bound(
+        &mut self,
+        lo: Span,
+        has_parens: bool,
+        modifiers: BoundModifiers,
+    ) -> PResult<'a, GenericBound> {
+        let lifetime_defs = self.parse_late_bound_lifetime_defs()?;
+        let path = self.parse_path(PathStyle::Type)?;
+        if has_parens {
+            self.expect(&token::CloseDelim(token::Paren))?;
+        }
+
+        let modifier = modifiers.to_trait_bound_modifier();
+        let poly_trait = PolyTraitRef::new(lifetime_defs, path, lo.to(self.prev_token.span));
+        Ok(GenericBound::Trait(poly_trait, modifier))
+    }
+
+    /// Optionally parses `for<$generic_params>`.
+    pub(super) fn parse_late_bound_lifetime_defs(&mut self) -> PResult<'a, Vec<GenericParam>> {
+        if self.eat_keyword(kw::For) {
+            self.expect_lt()?;
+            let params = self.parse_generic_params()?;
+            self.expect_gt()?;
+            // We rely on AST validation to rule out invalid cases: There must not be type
+            // parameters, and the lifetime parameters must not have bounds.
+            Ok(params)
+        } else {
+            Ok(Vec::new())
+        }
+    }
+
+    pub(super) fn check_lifetime(&mut self) -> bool {
+        self.expected_tokens.push(TokenType::Lifetime);
+        self.token.is_lifetime()
+    }
+
+    /// Parses a single lifetime `'a` or panics.
+    pub(super) fn expect_lifetime(&mut self) -> Lifetime {
+        if let Some(ident) = self.token.lifetime() {
+            self.bump();
+            Lifetime { ident, id: ast::DUMMY_NODE_ID }
+        } else {
+            self.span_bug(self.token.span, "not a lifetime")
+        }
+    }
+
+    pub(super) fn mk_ty(&self, span: Span, kind: TyKind) -> P<Ty> {
+        P(Ty { kind, span, id: ast::DUMMY_NODE_ID })
+    }
+}

compiler/rustc_parse/src/validate_attr.rs

@@ -0,0 +1,163 @@
+//! Meta-syntax validation logic of attributes for post-expansion.
+
+use crate::parse_in;
+
+use rustc_ast::tokenstream::DelimSpan;
+use rustc_ast::{self as ast, Attribute, MacArgs, MacDelimiter, MetaItem, MetaItemKind};
+use rustc_errors::{Applicability, PResult};
+use rustc_feature::{AttributeTemplate, BUILTIN_ATTRIBUTE_MAP};
+use rustc_session::lint::builtin::ILL_FORMED_ATTRIBUTE_INPUT;
+use rustc_session::parse::ParseSess;
+use rustc_span::{sym, Symbol};
+
+pub fn check_meta(sess: &ParseSess, attr: &Attribute) {
+    if attr.is_doc_comment() {
+        return;
+    }
+
+    let attr_info =
+        attr.ident().and_then(|ident| BUILTIN_ATTRIBUTE_MAP.get(&ident.name)).map(|a| **a);
+
+    // Check input tokens for built-in and key-value attributes.
+    match attr_info {
+        // `rustc_dummy` doesn't have any restrictions specific to built-in attributes.
+        Some((name, _, template, _)) if name != sym::rustc_dummy => {
+            check_builtin_attribute(sess, attr, name, template)
+        }
+        _ => {
+            if let MacArgs::Eq(..) = attr.get_normal_item().args {
+                // All key-value attributes are restricted to meta-item syntax.
+                parse_meta(sess, attr)
+                    .map_err(|mut err| {
+                        err.emit();
+                    })
+                    .ok();
+            }
+        }
+    }
+}
+
+pub fn parse_meta<'a>(sess: &'a ParseSess, attr: &Attribute) -> PResult<'a, MetaItem> {
+    let item = attr.get_normal_item();
+    Ok(MetaItem {
+        span: attr.span,
+        path: item.path.clone(),
+        kind: match &item.args {
+            MacArgs::Empty => MetaItemKind::Word,
+            MacArgs::Eq(_, t) => {
+                let v = parse_in(sess, t.clone(), "name value", |p| p.parse_unsuffixed_lit())?;
+                MetaItemKind::NameValue(v)
+            }
+            MacArgs::Delimited(dspan, delim, t) => {
+                check_meta_bad_delim(sess, *dspan, *delim, "wrong meta list delimiters");
+                let nmis = parse_in(sess, t.clone(), "meta list", |p| p.parse_meta_seq_top())?;
+                MetaItemKind::List(nmis)
+            }
+        },
+    })
+}
+
+pub fn check_meta_bad_delim(sess: &ParseSess, span: DelimSpan, delim: MacDelimiter, msg: &str) {
+    if let ast::MacDelimiter::Parenthesis = delim {
+        return;
+    }
+
+    sess.span_diagnostic
+        .struct_span_err(span.entire(), msg)
+        .multipart_suggestion(
+            "the delimiters should be `(` and `)`",
+            vec![(span.open, "(".to_string()), (span.close, ")".to_string())],
+            Applicability::MachineApplicable,
+        )
+        .emit();
+}
+
+/// Checks that the given meta-item is compatible with this `AttributeTemplate`.
+fn is_attr_template_compatible(template: &AttributeTemplate, meta: &ast::MetaItemKind) -> bool {
+    match meta {
+        MetaItemKind::Word => template.word,
+        MetaItemKind::List(..) => template.list.is_some(),
+        MetaItemKind::NameValue(lit) if lit.kind.is_str() => template.name_value_str.is_some(),
+        MetaItemKind::NameValue(..) => false,
+    }
+}
+
+pub fn check_builtin_attribute(
+    sess: &ParseSess,
+    attr: &Attribute,
+    name: Symbol,
+    template: AttributeTemplate,
+) {
+    // Some special attributes like `cfg` must be checked
+    // before the generic check, so we skip them here.
+    let should_skip = |name| name == sym::cfg;
+    // Some of previously accepted forms were used in practice,
+    // report them as warnings for now.
+    let should_warn = |name| {
+        name == sym::doc
+            || name == sym::ignore
+            || name == sym::inline
+            || name == sym::link
+            || name == sym::test
+            || name == sym::bench
+    };
+
+    match parse_meta(sess, attr) {
+        Ok(meta) => {
+            if !should_skip(name) && !is_attr_template_compatible(&template, &meta.kind) {
+                let error_msg = format!("malformed `{}` attribute input", name);
+                let mut msg = "attribute must be of the form ".to_owned();
+                let mut suggestions = vec![];
+                let mut first = true;
+                if template.word {
+                    first = false;
+                    let code = format!("#[{}]", name);
+                    msg.push_str(&format!("`{}`", &code));
+                    suggestions.push(code);
+                }
+                if let Some(descr) = template.list {
+                    if !first {
+                        msg.push_str(" or ");
+                    }
+                    first = false;
+                    let code = format!("#[{}({})]", name, descr);
+                    msg.push_str(&format!("`{}`", &code));
+                    suggestions.push(code);
+                }
+                if let Some(descr) = template.name_value_str {
+                    if !first {
+                        msg.push_str(" or ");
+                    }
+                    let code = format!("#[{} = \"{}\"]", name, descr);
+                    msg.push_str(&format!("`{}`", &code));
+                    suggestions.push(code);
+                }
+                if should_warn(name) {
+                    sess.buffer_lint(
+                        &ILL_FORMED_ATTRIBUTE_INPUT,
+                        meta.span,
+                        ast::CRATE_NODE_ID,
+                        &msg,
+                    );
+                } else {
+                    sess.span_diagnostic
+                        .struct_span_err(meta.span, &error_msg)
+                        .span_suggestions(
+                            meta.span,
+                            if suggestions.len() == 1 {
+                                "must be of the form"
+                            } else {
+                                "the following are the possible correct uses"
+                            },
+                            suggestions.into_iter(),
+                            Applicability::HasPlaceholders,
+                        )
+                        .emit();
+                }
+            }
+        }
+        Err(mut err) => {
+            err.emit();
+        }
+    }
+}