src/test/ui/asm/sym.rs makes incorrect use of `#[track_caller]`

[anp]

In this test:

rust/src/test/ui/asm/sym.rs

Lines 12 to 16 in 16957bd

	// The compiler will generate a shim to hide the caller location parameter.
	#[track_caller]
	fn f2() -> i32 {
	222
	}

I don't see anything in the current RFC text to suggest that the sym usage here should create a call shim, as they're only created when taking function pointers. Is there a function pointer being created in this test that I'm missing?

cc @Amanieu @eddyb

[Amanieu]

I did it this way because it is the only reasonable way to handle #[track_caller] with asm. The only other alternative would be to disallow the use of #[track_caller] for sym operands.

[eddyb]

I think I was too quick to dismiss the implementation - this looks like it does reify: https://godbolt.org/z/AnWrTP.
But I didn't actually see where that happens, I'll have to take another look.

EDIT: ugh, bool args strike again:

rust/src/librustc_mir/monomorphize/collector.rs

Line 663 in d462551

visit_fn_use(self.tcx, fn_ty, false, source, &mut self.output);

rust/src/librustc_mir/monomorphize/collector.rs

Line 715 in d462551

is_direct_call: bool,

I could've sworn that was is_indirect (maybe it's miri which has that?)

[eddyb]

That said, maybe f2 should return something like the line number? So that if the direct version was called, it would crash/return an incorrect value.

Also, ideally we'd test stuff like this with .stdout files, so we don't have to trust rustc to not optimize out assert_eqs that involve UB, but sadly I don't think those are set up to test runtime output yet.