Minic rustc's new format_args! expansion

[ShoyuVanilla]

Fixes #19929 and fixes #20051

[ChayimFriedman2]

I think it's a bit weird if activating just fmt will activate the old fmt, better for it to activate the newest.

[ShoyuVanilla]

Fair point. I'll fix those names along with the method names in expr_store/lower.rs

[ChayimFriedman2]

I'm a bit worried about the impact of searching for lang items in body lowering, as body lowering should stay fast.

I guess it's fine because lang items generally remain unchanged so they will be cached.

[ShoyuVanilla]

Yes, I agree to your both points: body lowering should be fast and in general, lang item caches persists.
But I think there would be no regression wrt this since we are already doing lang item query in current implementation:

rust-analyzer/crates/hir-def/src/expr_store/lower.rs

Lines 2866 to 2875 in 0ddaf2c

	let new_v1_formatted = LangItem::FormatArguments.ty_rel_path(
	self.db,
	self.module.krate(),
	Name::new_symbol_root(sym::new_v1_formatted),
	);
	let unsafe_arg_new = LangItem::FormatUnsafeArg.ty_rel_path(
	self.db,
	self.module.krate(),
	Name::new_symbol_root(sym::new),
	);

[ChayimFriedman2]

Maybe better to call this collect_format_args_old()? Although I'd prefer the names to tell the toolchain version, I don't want collect_format_args_really_new() if we get yet another version.

[ChayimFriedman2]

This is not good as it means they will be under an unsafe block, so will not need unsafe.

But the test testing that did not fire... So I don't know what's going on.

[ShoyuVanilla]

Oh, you're right! I missed this point because it didn't failed orphan unsafe test.
I checked why the test didn't failed and found out that the following code:

fn main() {
    let p = 3735928559 as *const i32;
    format_args!("", *p);
}

is expanded into:

fn main() {
    let p = 3735928559 as *const i32;
    {
        let args = (&*p, );
                  //^^^ Orphans are once assigned into first `arg`
        let args = [];
        unsafe {
            *p;
            builtin#lang(Arguments::new_v1_formatted)(
                &[],
                &args,
                &[],
            )
        }
    };
}

And thus, deleting those unnecessary and incorrect lines

[ShoyuVanilla]

I added a new arg orphan to check expansion behaviour or orphans

[ShoyuVanilla]

So, with the new format_args!, the completion works even with formatting with invalid matching braces, but I'm not sure whether I should take this as a regression or improvement

[ShoyuVanilla]

Those regressions are caused by failed borrowcks here, rooted from new format_args! expansions without super let or other proper alternative expansions:

rust-analyzer/crates/hir/src/term_search/tactics.rs

Lines 54 to 56 in 0ddaf2c

	ScopeDef::Local(it) => {
	if ctx.config.enable_borrowcheck {
	let borrowck = db.borrowck(it.parent).ok()?;

I'll implement super let as a followup and fix these by then