MemoryUsage of Arc forgets to account for ret-counts

[matklad]

loupe/crates/loupe/src/memory_usage/sync.rs

Lines 46 to 52 in 3108424

	mem::size_of_val(self)
	+ if tracker.track(Arc::as_ptr(self) as *const ()) {
	self.as_ref().size_of_val(tracker)
	} else {
	0
	}
	}

This accounts for sizeof T stored inside the Arc, as well as sizeof of the Arc itself (fat or thin pointer). However, it misses the two counters that are stored in the ArcInner. In parcicular this currently returns 8: crate::size_of_val(&Arc::new(())). I'd expect 24 instead -- one usize for the pointer + 2 * usize for strong and weak count.

[Hywan]

I'm surprised that mem::size_of_val misses the size of the ArcInner. I need to take a look at it.

[matklad]

Yeah, it's a bit tricky. The only difference between size_of and size_of_val is in handling DST.

For Arc, the size_of is just a single pointer to the ArcInner heap allocated box. There isn't a function which can tell you the size of this box itself, you'll have to hard-code it.

Well, I guess there is a way to get this size without hard-coding, by asking the allocator about it. This is what servo's analogue dose:

https://github.com/servo/servo/blob/94414dfb9aa8d29318e4825a36e4fdd188fbb892/components/malloc_size_of/lib.rs

But that obviously requires support from the allocator.

[matklad]

That's completly unrelated to the issue, but, JFYI, rust-lang/rust-analyzer#9309