Parser: Do not clone on peeking

[chris-m-h]

This eliminates calls to clone() and to_owned() in the parser

• Peeking now returns reference only
• To obtain value, fetch_token needs to be called
• The parser was adapted accordingly
• Also: Pass anchor name by value to register_anchor

[hoodie]

I'd really like to see how much this impacts performance

[chris-m-h]

Using twitter.json from the JSON benchmark (http://json.rs/) I got the following bench results:

• For a parser with an EventReceiver that simply ignores the events, I got a performance gain of 22%.
• Including the creation of the YAML struct, I still got a performance gain of 15%.

I would regard that quite a difference, although I have to admit that it is less than I expected... I do not claim that the tests were performed under experimental conditions...

---

Let me add the data for parsing the yaml.org page itself - again just a very simple test...

• Parser (with Scanner) 11% faster
• Parser only (minus Scanner time): around 100% faster
• Whole process including Yaml struct generation: 12 % faster

test parser::bench::bench_parser ... bench: 132,619 ns/iter (+/- 5,744)
test scanner::bench::bench_scanner ... bench: 117,354 ns/iter (+/- 19,291)
test yaml::test::bench_yaml ... bench: 149,222 ns/iter (+/- 17,655)

test parser::bench::bench_parser ... bench: 147,476 ns/iter (+/- 4,429)
test scanner::bench::bench_scanner ... bench: 116,219 ns/iter (+/- 3,479)
test yaml::test::bench_yaml ... bench: 167,065 ns/iter (+/- 11,638)

[chris-m-h]

Any interest in this, @dtolnay or @chyh1990 ? Basically, just the internal structure of the parser functions had to be changed to stay inside the lifetime of the token references, in this way avoiding the need for cloning.

[chyh1990]

+1. Looks good to me. But it is a large patch, can we get another reviewer for it? @dtolnay

[dtolnay]

LGTM as well.