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>

1 files changed, 27 insertions, 9 deletions

diff --git a/src/bytes/bytes_generic.go b/src/bytes/bytes_generic.go
index 32abd3b33f..b30e53bf2e 100644
--- a/src/bytes/bytes_generic.go
+++ b/src/bytes/bytes_generic.go
@@ -6,23 +6,25 @@
 
 package bytes
 
-// TODO: implements short string optimization on non amd64 platforms
-// and get rid of bytes_amd64.go
-
 // Index returns the index of the first instance of sep in s, or -1 if sep is not present in s.
 func Index(s, sep []byte) int {
 	n := len(sep)
-	if n == 0 {
+	switch {
+	case n == 0:
 		return 0
-	}
-	if n > len(s) {
+	case n == 1:
+		return IndexByte(s, sep[0])
+	case n == len(s):
+		if Equal(sep, s) {
+			return 0
+		}
+		return -1
+	case n > len(s):
 		return -1
 	}
 	c := sep[0]
-	if n == 1 {
-		return IndexByte(s, c)
-	}
 	i := 0
+	fails := 0
 	t := s[:len(s)-n+1]
 	for i < len(t) {
 		if t[i] != c {
@@ -36,6 +38,22 @@ func Index(s, sep []byte) int {
 			return i
 		}
 		i++
+		fails++
+		if fails >= 4+i>>4 && i < len(t) {
+			// Give up on IndexByte, it isn't skipping ahead
+			// far enough to be better than Rabin-Karp.
+			// Experiments (using IndexPeriodic) suggest
+			// the cutover is about 16 byte skips.
+			// TODO: if large prefixes of sep are matching
+			// we should cutover at even larger average skips,
+			// because Equal becomes that much more expensive.
+			// This code does not take that effect into account.
+			j := indexRabinKarp(s[i:], sep)
+			if j < 0 {
+				return -1
+			}
+			return i + j
+		}
 	}
 	return -1
 }