Improve syntax fixup for proc macros

[flodiebold]

To make completion work inside attribute proc macros like #[tokio::main], we take incomplete code inside the attribute input and 'fix it up' to make sure the proc macro can parse it. This already works pretty well in the common cases, but I think there are lots of edge cases missing. This usually results in completion not working inside the proc macro for a moment while typing something, and then working again once one has typed a more complete expression, so I think people are unlikely to report cases like this and will just chalk it up to "rust-analyzer being unreliable". So I think it's worth to look for edge cases that are missing and add them.

Also, while it would be nice to have some more automatic / principled way to do this syntax fixup, any work adding more cases to the current code and extending the test suite will not be wasted; the test suite is really the most important part.

The syntax fixup code lives in the fixup module:

rust-analyzer/crates/hir-expand/src/fixup.rs

Lines 1 to 3 in 96481b7

	//! To make attribute macros work reliably when typing, we need to take care to
	//! fix up syntax errors in the code we're passing to them.
	use std::mem;

New cases will need to be added to the big match_ast! in the fixup_syntax function:

rust-analyzer/crates/hir-expand/src/fixup.rs

Line 73 in 96481b7

match_ast! {

As can be seen from the existing cases, this mostly consists of checking if some part of a node is missing and inserting a simple placeholder token or two (an identifier, or some missing punctuation).

For each case, a test case should be added at the end of the module. The tests look like this:

rust-analyzer/crates/hir-expand/src/fixup.rs

Lines 288 to 300 in 96481b7

	#[test]
	fn extraneous_comma() {
	check(
	r#"
	fn foo() {
	bar(,);
	}
	"#,
	expect![[r#"
	fn foo () {__ra_fixup ;}
	"#]],
	)
	}

Test cases consist of the original incomplete code, and the "fixed up" code. The "fixed up" code can be automatically updated by running UPDATE_EXPECT=1 cargo test -p hir-expand, and the tests also check the important conditions that 1. the fixed-up code is syntactically complete, and 2. the fixups can be reversed afterwards, so if these checks succeed, the fixup is working.

Here are some cases that I think we are missing currently:

• incomplete paths like foo::
• incomplete flow control keywords like loop without a block, for missing its various components etc.

[fprasx]

@rustbot claim