Change single char str patterns to chars

[ljedrz]

A char is faster.

[rust-highfive]

r? @michaelwoerister

(rust_highfive has picked a reviewer for you, use r? to override)

[petrochenkov]

A char is faster

I wouldn't be so sure, see #41993
str is "search bytes in bytes", while char is "convert char to bytes, then search bytes in bytes".

[ljedrz]

Oh, ok; in that case I guess this might call for an update in clippy, as its single_char_pattern lint suggests that it's the other way round.

cc @oli-obk

[ljedrz]

@petrochenkov I just did a comparison using godbolt and the char variant produces half the str assembly; perhaps that issue is limited to starts_with? I'm not an expert, but this is confusing.

[killercup]

A char is faster.

I don't see any benchmarks? :)

[ljedrz]

@killercup see the related clippy lint.

[kennytm]

@petrochenkov#41993 only concerns about .starts_with(). The split() algorithm for char uses memchr() for a char and is way more efficient than that for a &str (which degenerates to naive search).

[killercup]

@ljedrz a piece of documentation is still no benchmark :)

Seriously, if we don't have one, we should totally write one and add it. This is easily decidable once we have the facts.

Edit: Maybe we can use the ones @kennytm wrote last year:
#41993 (comment). The most significant change to land since then was probably #46735. And if nothing else, we can link the results in the clippy docs.

Edit 2: We are currently linking to libc::memchr on most platforms. Adding a SIMD-aware Rust version might allow for better inlining. (But again, without benchmarks this is just speculation.)

[kennytm]

Since I just happened to have benchmarks from rust-lang/rfcs#2500, here's the relevant parts comparing .starts_with() and .split():

.starts_with()

test pat3_starts_with_ascii_char::long_lorem_ipsum ... bench: 174 ns/iter (+/- 12)
test pat3_starts_with_ascii_char::short_ascii ... bench: 175 ns/iter (+/- 21)
test pat3_starts_with_ascii_char::short_mixed ... bench: 174 ns/iter (+/- 9)
test pat3_starts_with_ascii_char::short_pile_of_poo ... bench: 175 ns/iter (+/- 26)
test pat3_starts_with_ascii_string::long_lorem_ipsum ... bench: 136 ns/iter (+/- 20)
test pat3_starts_with_ascii_string::short_ascii ... bench: 138 ns/iter (+/- 26)
test pat3_starts_with_ascii_string::short_mixed ... bench: 137 ns/iter (+/- 13)
test pat3_starts_with_ascii_string::short_pile_of_poo ... bench: 136 ns/iter (+/- 15)
test std_starts_with_ascii_char::long_lorem_ipsum ... bench: 339 ns/iter (+/- 29)
test std_starts_with_ascii_char::short_ascii ... bench: 338 ns/iter (+/- 21)
test std_starts_with_ascii_char::short_mixed ... bench: 844 ns/iter (+/- 42)
test std_starts_with_ascii_char::short_pile_of_poo ... bench: 930 ns/iter (+/- 91)
test std_starts_with_ascii_string::long_lorem_ipsum ... bench: 306 ns/iter (+/- 41)
test std_starts_with_ascii_string::short_ascii ... bench: 307 ns/iter (+/- 63)
test std_starts_with_ascii_string::short_mixed ... bench: 252 ns/iter (+/- 122)
test std_starts_with_ascii_string::short_pile_of_poo ... bench: 222 ns/iter (+/- 193)

Speed ups:

Method	Long ASCII	Short ASCII	Short Mixed	Short Unicode
.starts_with('/') from libstd (baseline)	0%	0%	0%	0%
.starts_with("/") from libstd	-10%	-9%	-70%	-76%
.starts_with('/') from RFC 2500	-49%	-48%	-79%	-81%
.starts_with("/") from RFC 2500	-60%	-59%	-84%	-85%

.split()

test pat3_split_a_char::long_lorem_ipsum ... bench: 3,084 ns/iter (+/- 517)
test pat3_split_a_char::short_ascii ... bench: 158 ns/iter (+/- 15)
test pat3_split_a_char::short_mixed ... bench: 102 ns/iter (+/- 27)
test pat3_split_a_char::short_pile_of_poo ... bench: 13 ns/iter (+/- 0)
test pat3_split_a_str::long_lorem_ipsum ... bench: 5,592 ns/iter (+/- 692)
test pat3_split_a_str::short_ascii ... bench: 180 ns/iter (+/- 23)
test pat3_split_a_str::short_mixed ... bench: 160 ns/iter (+/- 17)
test pat3_split_a_str::short_pile_of_poo ... bench: 109 ns/iter (+/- 9)
test std_split_a_char::long_lorem_ipsum ... bench: 3,003 ns/iter (+/- 231)
test std_split_a_char::short_ascii ... bench: 159 ns/iter (+/- 11)
test std_split_a_char::short_mixed ... bench: 108 ns/iter (+/- 2)
test std_split_a_char::short_pile_of_poo ... bench: 23 ns/iter (+/- 0)
test std_split_a_str::long_lorem_ipsum ... bench: 6,176 ns/iter (+/- 693)
test std_split_a_str::short_ascii ... bench: 194 ns/iter (+/- 7)
test std_split_a_str::short_mixed ... bench: 170 ns/iter (+/- 17)
test std_split_a_str::short_pile_of_poo ... bench: 109 ns/iter (+/- 16)

Speed ups:

Method	Long ASCII	Short ASCII	Short Mixed	Short Unicode
.split('a').count() from libstd (baseline)	0%	0%	0%	0%
.split("a").count() from libstd	+106%	+22%	+57%	+374%
.split('a').count() from RFC 2500	+3%	-1%	-6%	-43%
.split("a").count() from RFC 2500	+86%	+13%	+48%	+374%

While .starts_with('/') is much slower than .starts_with("/"), we see that the situation is reversed that .split('a').count() is much faster than .split("a").count(). This is because we did not optimize for single-byte string searching at all, but we do optimize for ASCII char searching.

[michaelwoerister]

Thanks for the PR, @ljedrz!

@kennytm's numbers suggest that this should be a win (although none of the modified code is really on the critical path).

@bors r+

[bors]

📌 Commit 49c8ba9 has been approved by michaelwoerister

[bors]

⌛ Testing commit 49c8ba9 with merge a2af866...

[bors]

☀️ Test successful - status-appveyor, status-travis
Approved by: michaelwoerister
Pushing a2af866 to master...