Literal: Avoid string copies by default

tagua-vm / parser

Safe, fast and memory efficient PHP parser (lexical and syntactic analysers, and the Abstract Syntax Tree)

http://tagua.io/

119 stars 15 forks source link

Literal: Avoid string copies by default #107

Closed Hywan closed 7 years ago

Hywan commented 7 years ago

A string is no longer copied by default. A single quoted string is copied if and only if it contains an escaped character, because the resulting string must not contain the escaping character, and this is not possible to concatenate slices, so the new string is copied inside a Vec<u8>. However, if no character is escaped, or if it is a nowdoc string, there is no longer any copy.

Behind the scene, we use a clone-on-write smart pointer: std::borrow::Cow.

A benchmark needs to validate this work though, but I am confident it will improve the global performance, let's see.

Interesting commits are:

https://github.com/tagua-vm/parser/commit/a4e2688f1df002d647e36f25d06a8803c63589fd the real changeset,
https://github.com/tagua-vm/parser/commit/884b6da4a2d359eda7d0e897b0ec4ab6c1630e1b new test cases ensuring whether the string have been copied or not.

cc @jubianchi @Geal

Hywan commented 7 years ago

Bench results.

Before:

test rules::literals::tests::bench_simple  ... bench:       2,939 ns/iter (+/- 1,086)

After:

test rules::literals::tests::bench_simple  ... bench:       2,792 ns/iter (+/- 548)

tested with a string of size 5k. This is a boost of 5%.

For smaller strings of size 12 bytes, the boost is of 20%. Well, the benchmarks are less relevant in this case. I have done 10 runs, and the boost is constantly about 20%.

For escaped strings, so with a copy, strangely these is a constant boost of 1-2%. I don't know why, but I take it.