1.35 -> 1.36 regression: && incorrectly considered to be unable to start expressions.

[eddyb]

This example compiles on 1.35 but not 1.36 or later:

pub fn foo() -> &'static &'static bool {
    {} // equivalent to {}; or ();
    &&false
}

(The example above was reduced from one involving match ... {...} && false, found by @digama0)

This regression was likely introduced by #60188, due to the definition of can_continue_expr_unambiguously:

rust/src/librustc_ast/util/parser.rs

Lines 209 to 232 in daecab3

	/// This operator could be used to follow a block unambiguously.
	///
	/// This is used for error recovery at the moment, providing a suggestion to wrap blocks with
	/// parentheses while having a high degree of confidence on the correctness of the suggestion.
	pub fn can_continue_expr_unambiguously(&self) -> bool {
	use AssocOp::*;
	match self {
	BitXor | // `{ 42 } ^ 3`
	Assign | // `{ 42 } = { 42 }`
	Divide | // `{ 42 } / 42`
	Modulus | // `{ 42 } % 2`
	ShiftRight | // `{ 42 } >> 2`
	LessEqual | // `{ 42 } <= 3`
	Greater | // `{ 42 } > 3`
	GreaterEqual | // `{ 42 } >= 3`
	AssignOp(_) | // `{ 42 } +=`
	LAnd | // `{ 42 } &&foo`
	As | // `{ 42 } as usize`
	// Equal | // `{ 42 } == { 42 }` Accepting these here would regress incorrect
	// NotEqual | // `{ 42 } != { 42 } struct literals parser recovery.
	Colon => true, // `{ 42 }: usize`
	_ => false,
	}
	}

That function lists all of the operators that can't start an expression, but only continue it.
LAnd (&&) is incorrectly included, as &&expr is parsed the same as & &expr at the start of an expression.

cc @estebank

[eddyb]

I've left some comments on #61500 (a partial fix for a regression that seems similar to this, but more specific), e.g.: #61500 (comment)

Looking again at this, I don't think this PR was necessary, it seems to have worked around the bug without fully fixing it.
As per #74233, can_continue_expr_unambiguously should've never returned true for LAnd, preventing the regression.

[spastorino]

Assigning P-high as discussed as part of the Prioritization Working Group procedure and removing I-prioritize.