Manually adding padding to `FastStr` gains 9x boost on performance, but why?

Hi, I'm the author of \FastStr\\ crate, and recently I found a wired problem that the clone cost of \FastStr\\ is really high. For example, an empty \FastStr\\ clone costs about 40ns on amd64 compared to about 4ns of a normal String.

After some time of investigation, I found that this is because the \Repr::Inline\\ part has really great affect on the performance. And after I added a padding to it(change the type of \len\\ from \u8\\ to \usize\\), the performance boosts about 9x. But the root cause is still not clear.

Furthermore, I've tried the following methods, but none helps:

change \INLINE\\_CAP\\ to 24
change \INLINE\\_CAP\\ to 22 and added a padding to the Inline variant: \Inline {\\_pad: u64,len: u8,buf: \\[u8; INLINE\\_CAP\\],},\\
change \INLINE\\_CAP\\ to 22 and add a new struct \Inline\\ without the \\\_pad\\ field

To change the \INLINE\\_CAP\\ to 22 is only for not increasing the size of \FastStr\\ itself when add an extra padding, so the performance is nothing to do with it.

Edit: rust issue: https://github.com/rust-lang/rust/issues/119247