Further improve diagnostics for expressions in pattern position

[fmease]

I've had this on my TODO list for a while but never had the time to write a PR for it. Therefore I'm just going to open an issue for it instead.

Since PR #118625 we try to reparse something in pattern position as an expression under certain conditions and provide a smarter diagnostic of the form “expected a pattern, found an expression // arbitrary expressions are not allowed in patterns” which is already amazing.

However, the current heuristic has a few limitations I'd like to see addressed which should be relatively simple to implement:

1. The heuristic doesn't cover field accesses (of the form $literal.$ident and so forth) but only things that look like method calls (e.g., 0.f()). I argue we should also detect expressions like x.f and s.0. This would've prevented #104996 from getting opened!
   1. Firstly we need to remove the extra check for parentheses located after the .-“herald” check.
   2. Then, we should in my opinion check the Span to make sure that there's no whitespace/comment before & after the . to avoid overzealously emitting this diagnostic when the user accidentally typoed a comma as a dot (as in (x. y) ← (x, y)) which isn't too farfetched given the keys are right next to each other on most keyboards (just need to check sp0.hi == sp1.lo). Of course, this would rule out multi-line method / field chains but that should be rare. We could (as a follow-up) check if the would-be expression is “appropriately” indented (by utilizing SourceMap methods) to address that, too.
   3. The following is a bit opinionated: I think we should remove the capitalization check (at the moment, we don't consider x.Foo() to be a method call). If you disagree, we should at least also consider _ to be a valid start of a field or method name (x._foo() doesn't get recognized at the time of writing).
   4. We should support sequences of numeric field accesses like x.0.1 without accidentally flagging x.0e4 as an expression (x.0.1 gets lexed as [Ident, Dot, FloatLit]). See parse_expr_tuple_field_access_float for prior art.
2. We should consider the keyword as to be a “herald” next to . to recover 0 as usize etc. in pattern position
3. We should somehow recover f().g() and (0).f() meaning more method calls / field accesses where the “basis” is a more complex expression. Rephrased: Recognize more “$expr.$ident” over just “$lit_or_ident.$ident” (the latter is the status quo, roughly, if I'm not mistaken).
4. We should add a help message (structured or text-based) instructing the user to extract the expression to a const item: “extract the expression to a const and refer to that”.
5. At some point we should provide a structured multipart suggestion for wrapping expressions in pattern position in a const {} block (feature inline_const_pat). We could provide the suggestion if self.sess.unstable_features.is_nightly_build() holds but I don't know if others like the idea of that. Note that we cannot check if feature inline_const_pat is enabled from inside the parser and diagnostic stashing/stealing doesn't help here either since we want to “stash” a subdiagnostic (a structured suggestion) not true actually, should be possible to impl. A FIXME should suffice for the initial impl.

[fmease]

cc @ShE3py, @WaffleLapkin

[fmease]

I know that this is quite a wall of text and I don't expect anyone to read it. I've tagged you, ShE3py, since you might be interested in working on some of the things I listed here. Otherwise this issue just serves as a public TODO list for me.

[ShE3py]

Okay, I've somehow resurrected.

1. ii. Then, we should in my opinion check the Span to make sure that there's no whitespace/comment before & after the . to avoid overzealously emitting this diagnostic when the user accidentally typoed a comma as a dot (as in (x. y) ← (x, y)) which isn't too farfetched given the keys are right next to each other on most keyboards

It might be worth to emit both diagnostics (e.g. “unexpected field access // try introducing a const VAL: _ = x. y // you may have meant to use a comma there”).

1. iii. The following is a bit opinionated: I think we should remove the capitalization check (at the moment, we don't consider x.Foo() to be a method call).

Likewise. Especially since some scripts e.g. Arabic doesn't have cases (the check should've been something like !c.is_uppercase()).

1. iv. See parse_expr_tuple_field_access_float for prior art.

Note to self: this was inlined into parse_dot_suffix_expr in #122858.
Permalink to that PR's parent:

rust/compiler/rustc_parse/src/parser/expr.rs

Lines 1122 to 1155 in 0824b30

	fn parse_expr_tuple_field_access_float(
	&mut self,
	lo: Span,
	base: P<Expr>,
	float: Symbol,
	suffix: Option<Symbol>,
	) -> P<Expr> {
	match self.break_up_float(float, self.token.span) {
	// 1e2
	DestructuredFloat::Single(sym, _sp) => {
	self.parse_expr_tuple_field_access(lo, base, sym, suffix, None)
	}
	// 1.
	DestructuredFloat::TrailingDot(sym, ident_span, dot_span) => {
	assert!(suffix.is_none());
	self.token = Token::new(token::Ident(sym, IdentIsRaw::No), ident_span);
	let next_token = (Token::new(token::Dot, dot_span), self.token_spacing);
	self.parse_expr_tuple_field_access(lo, base, sym, None, Some(next_token))
	}
	// 1.2 | 1.2e3
	DestructuredFloat::MiddleDot(symbol1, ident1_span, dot_span, symbol2, ident2_span) => {
	self.token = Token::new(token::Ident(symbol1, IdentIsRaw::No), ident1_span);
	// This needs to be `Spacing::Alone` to prevent regressions.
	// See issue #76399 and PR #76285 for more details
	let next_token1 = (Token::new(token::Dot, dot_span), Spacing::Alone);
	let base1 =
	self.parse_expr_tuple_field_access(lo, base, symbol1, None, Some(next_token1));
	let next_token2 = Token::new(token::Ident(symbol2, IdentIsRaw::No), ident2_span);
	self.bump_with((next_token2, self.token_spacing)); // `.`
	self.parse_expr_tuple_field_access(lo, base1, symbol2, suffix, None)
	}
	DestructuredFloat::Error => base,
	}
	}

I know that this is quite a wall of text and I don't expect anyone to read it. I've tagged you, ShE3py, since you might be interested in working on some of the things I listed here. Otherwise this issue just serves as a public TODO list for me.

I don't mind reading, and I often find myself writing walls of text, albeit in French, and my English isn't so good as to write pavés in it while retaining the same meaning.

@rustbot claim