bad performance of char literal to_string() vs str literal to_string()

[matthiaskrgr]

I was quite surprised when I found out that

pub fn conv() -> String {
    'a'.to_string()
}

generates several times the instruction count of

pub fn conv() -> String {
    "a".to_string()
}

(according to godbold).
Checking the (debug) build size of println!("{}", "a".to_string()); vs println!("{}", 'a'.to_string()); revealed the str variant being ~20kb lighter.

When doing some quick-and-dirty benchmarks, the str variant was more than 2x faster than the char-variant. benchmark code

test tests::bench_char_to_string ... bench:      39,890 ns/iter (+/- 42,078)
test tests::bench_str_to_string  ... bench:      14,422 ns/iter (+/- 2,427)

Is this something that could be optimized?

[alex]

The problem seems to be that char doesn't directly implement to_string, it comes from the blanket impl of Display. Adding an impl of ToString directly on char that looks like:

    let mut s = String::new();
    s.push(self);
    s

is probably the fix

[tesuji]

godbolt link: https://rust.godbolt.org/z/aYEaZr
I will add a specialization of ToString for char.

[tesuji]

Hm. {int}.to_string() is slow like char.to_string()

[pickfire]

Looks like byte and integer runs as slow.

test tests::bench_byte_to_string ... bench:      51,257 ns/iter (+/- 1,623)
test tests::bench_char_to_string ... bench:      43,246 ns/iter (+/- 7,624)
test tests::bench_int_to_string  ... bench:      52,218 ns/iter (+/- 3,077)
test tests::bench_str_to_string  ... bench:      15,701 ns/iter (+/- 380)

Benchmark code https://play.rust-lang.org/?version=nightly&mode=debug&edition=2018&gist=a440f60fb09e7e6e4c60eef0a1ebfac1

[pickfire]

Should we open another issue for byte and int?

[tesuji]

yeah, do it.