Borrowing degradations

Iterations are not the only place where translation degradations occur. There are also a couple of extra points where you can sometimes see that the same code performs much worse in Rust than in C/C++. One of these points is reference handling. Due to borrow checker constraints, you can do three things with variables when passing them to a function: send a reference (borrow), give the new function control of the variable (own), or copy/clone the variable to send it to a function. It seems easy to decide, right? If you do not require the variable anymore, let the function own your variable. If you require it, send the reference, and if you require it and the API only accepts ownership, clone it.

Well, it actually isn't so simple. For example, integers are faster to copy than to reference, and so are small structures. The rule of thumb is, if it's smaller than or equal to usize, copy, always. If it's somewhere between usize and 10 times that size, it's probably better to copy. If it's bigger, it's probably better to reference. If the structure has a heap allocation (such as a Vec or a Box), it's usually better to send a reference.

There are some cases, though, when you cannot decide what happens to the variable. In a macro, for example, the variable is passed as is, and the macro decides what to do with it. For example, the println! macro gets all elements by reference, since it does not require more. The problem is that if you are trying to print an integer, for example, a bottleneck appears. That's what happened some time ago to Robert Grosse, and he wrote an article about it.

Long story short, he had to force the copying of the integer. How did he do that? Well, it's as simple as creating a scope that will return that integer. Since integers implement Copy, the integer will be copied to the scope and then returned, effectively copying it to the macro:

let my_int = 76_u32;
println!("{}", {my_int});

For normal prints, this is not usually necessary, but if you need to quickly print thousands or millions of integers, you will not avoid the I/O interface, but you can at least avoid this bottleneck.