bytes,strings: in generic Index, use mix of IndexByte and Rabin-Karp

Use IndexByte first, as it allows us to skip lots of bytes quickly.
If IndexByte is generating a lot of false positives, switch over to Rabin-Karp.

Experiments for ppc64le
bytes:
name                             old time/op  new time/op  delta
IndexPeriodic/IndexPeriodic2-2   1.12ms ± 0%  0.18ms ± 0%  -83.54%  (p=0.000 n=10+9)
IndexPeriodic/IndexPeriodic4-2    635µs ± 0%   184µs ± 0%  -71.06%  (p=0.000 n=9+9)
IndexPeriodic/IndexPeriodic8-2    289µs ± 0%   184µs ± 0%  -36.51%  (p=0.000 n=10+9)
IndexPeriodic/IndexPeriodic16-2   133µs ± 0%   183µs ± 0%  +37.68%  (p=0.000 n=10+9)
IndexPeriodic/IndexPeriodic32-2  68.3µs ± 0%  70.2µs ± 0%   +2.76%  (p=0.000 n=10+10)
IndexPeriodic/IndexPeriodic64-2  35.8µs ± 0%  36.6µs ± 0%   +2.17%  (p=0.000 n=8+10)

strings:
name                             old time/op  new time/op  delta
IndexPeriodic/IndexPeriodic2-2    184µs ± 0%   184µs ± 0%   +0.11%  (p=0.029 n=4+4)
IndexPeriodic/IndexPeriodic4-2    184µs ± 0%   184µs ± 0%     ~     (p=0.886 n=4+4)
IndexPeriodic/IndexPeriodic8-2    184µs ± 0%   184µs ± 0%     ~     (p=0.486 n=4+4)
IndexPeriodic/IndexPeriodic16-2   185µs ± 1%   184µs ± 0%     ~     (p=0.343 n=4+4)
IndexPeriodic/IndexPeriodic32-2   184µs ± 0%    69µs ± 0%  -62.37%  (p=0.029 n=4+4)
IndexPeriodic/IndexPeriodic64-2   184µs ± 0%    37µs ± 0%  -80.17%  (p=0.029 n=4+4)

Fixes #22578

Change-Id: If2a4d8554cb96bfd699b58149d13ac294615f8b8
Reviewed-on: https://go-review.googlesource.com/76070
Reviewed-by: Alberto Donizetti <alb.donizetti@gmail.com>
This commit is contained in:
Keith Randall 2017-11-04 10:19:53 -07:00
parent 0ffe90b501
commit a025277505
10 changed files with 154 additions and 140 deletions

View file

@ -125,6 +125,9 @@ var indexTests = []IndexTest{
{"xx012345678901234567890123456789012345678901234567890123456789012"[:41], "0123456789012345678901234567890123456789", -1},
{"xx012345678901234567890123456789012345678901234567890123456789012", "0123456789012345678901234567890123456xxx", -1},
{"xx0123456789012345678901234567890123456789012345678901234567890120123456789012345678901234567890123456xxx", "0123456789012345678901234567890123456xxx", 65},
// test fallback to Rabin-Karp.
{"oxoxoxoxoxoxoxoxoxoxoxoy", "oy", 22},
{"oxoxoxoxoxoxoxoxoxoxoxox", "oy", -1},
}
var lastIndexTests = []IndexTest{
@ -1641,3 +1644,15 @@ func BenchmarkTrimASCII(b *testing.B) {
}
}
}
func BenchmarkIndexPeriodic(b *testing.B) {
key := "aa"
for _, skip := range [...]int{2, 4, 8, 16, 32, 64} {
b.Run(fmt.Sprintf("IndexPeriodic%d", skip), func(b *testing.B) {
s := Repeat("a"+Repeat(" ", skip-1), 1<<16/skip)
for i := 0; i < b.N; i++ {
Index(s, key)
}
})
}
}