Fix crash on lazy initialization

[rhysd]

Fixes #334
Closes #390

Problem

Crash at

syntect/src/parsing/regex.rs

Lines 73 to 74 in 540555e

	self.regex.fill(regex).ok();
	self.regex.borrow().unwrap()

Since this rarely happens in mutli-threads environment, I couldn't trace what happened. However, logically it would be:

1. Thread1 starts .fill() call and acquires lock
2. Thread2 starts .fill() call but cannot acquire lock (since Thread1 already did it)
3. .fill() in Thread2 immediately returns Err(value) (but Thread2 ignores it by .ok())
4. Thread2 calls .borrow(). However, at this point, the .fill() call in Thread1 is not completed yet. So .borrow() in Thread2 returns None
5. The .unwrap() call crashes
6. (Thread1 finishes to set value at .fill() call)

So,

cell.fill(something).ok();
cell.borrow().unwrap();

is not crash-free actually.

Solution

Using once_cell::sync::OnceCell instead of lazycell::ActomicLazyCell solves this issue.

https://docs.rs/once_cell/1.8.0/once_cell/sync/struct.OnceCell.html#method.get_or_init

Many threads may call get_or_init concurrently with different initializing functions, but it is guaranteed that only one function will be executed.

I also read its implementation. Initializing cell is synchronized with a queue.

https://github.com/matklad/once_cell/blob/7b2943b3828867a58690165961424230ca423403/src/imp_std.rs#L86-L107

[robinst]

LGTM!

[Enselic]

I have confirmed that this change works well with syntax lazy-loading (the sequel to #382). It also looks good in and of itself.

[trishume]

Thanks for debugging this and fixing it!