bytes,strings: use IndexByte more often in Index on AMD64

IndexByte+compare is faster than indexShortStr in good case, when
first byte is rare, but is more costly in bad cases.
Start with IndexByte and switch to indexShortStr if we encounter
false positives more often than once per 8 bytes.

Benchmark changes for package bytes:

IndexRune/4K-8                    416ns ± 0%       86ns ± 0%    -79.24%        (p=0.000 n=10+10)
IndexRune/4M-8                    413µs ± 0%      100µs ± 1%    -75.88%        (p=0.000 n=10+10)
IndexRune/64M-8                  6.73ms ± 0%     2.86ms ± 1%    -57.49%        (p=0.000 n=10+10)
Index/10-8                       8.45ns ± 0%     8.96ns ± 0%     +6.04%         (p=0.000 n=9+10)
Index/32-8                       9.64ns ± 0%     9.51ns ± 0%     -1.30%          (p=0.000 n=8+9)
Index/4K-8                       2.11µs ± 0%     2.12µs ± 0%     +0.26%        (p=0.000 n=10+10)
Index/4M-8                       3.60ms ± 5%     3.59ms ± 7%       ~            (p=0.497 n=9+10)
Index/64M-8                      57.1ms ± 3%     58.7ms ± 5%       ~            (p=0.113 n=9+10)
IndexEasy/10-8                   7.10ns ± 1%     7.71ns ± 1%     +8.60%        (p=0.000 n=10+10)
IndexEasy/32-8                   9.29ns ± 1%     9.22ns ± 0%     -0.75%         (p=0.000 n=9+10)
IndexEasy/4K-8                   1.06µs ± 0%     0.08µs ± 0%    -92.18%        (p=0.000 n=10+10)
IndexEasy/4M-8                   1.07ms ± 0%     0.10ms ± 1%    -90.74%         (p=0.000 n=9+10)
IndexEasy/64M-8                  17.3ms ± 0%      2.8ms ± 1%    -83.76%         (p=0.000 n=10+9)

IndexRune/4K-8                 9.84GB/s ± 0%  47.42GB/s ± 0%   +381.85%         (p=0.000 n=8+10)
IndexRune/4M-8                 10.1GB/s ± 0%   42.1GB/s ± 1%   +314.56%        (p=0.000 n=10+10)
IndexRune/64M-8                10.0GB/s ± 0%   23.4GB/s ± 1%   +135.25%        (p=0.000 n=10+10)
Index/10-8                     1.18GB/s ± 0%   1.12GB/s ± 0%     -5.67%         (p=0.000 n=10+9)
Index/32-8                     3.32GB/s ± 0%   3.36GB/s ± 0%     +1.27%         (p=0.000 n=10+9)
Index/4K-8                     1.94GB/s ± 0%   1.93GB/s ± 0%     -0.25%         (p=0.000 n=10+9)
Index/4M-8                     1.17GB/s ± 5%   1.17GB/s ± 7%       ~            (p=0.497 n=9+10)
Index/64M-8                    1.17GB/s ± 3%   1.15GB/s ± 6%       ~            (p=0.113 n=9+10)
IndexEasy/10-8                 1.41GB/s ± 1%   1.30GB/s ± 1%     -7.90%        (p=0.000 n=10+10)
IndexEasy/32-8                 3.45GB/s ± 1%   3.47GB/s ± 0%     +0.73%         (p=0.000 n=9+10)
IndexEasy/4K-8                 3.84GB/s ± 0%  49.16GB/s ± 0%  +1178.78%         (p=0.000 n=9+10)
IndexEasy/4M-8                 3.91GB/s ± 0%  42.19GB/s ± 1%   +980.37%         (p=0.000 n=9+10)
IndexEasy/64M-8                3.88GB/s ± 0%  23.91GB/s ± 1%   +515.76%         (p=0.000 n=10+9)

No significant changes in strings.

In regexp I see:

Match/Easy0/32-8                 536MB/s ± 1%   540MB/s ± 1%    +0.75%         (p=0.001 n=9+10)
Match/Easy0/1K-8                1.62GB/s ± 0%  4.42GB/s ± 1%  +172.48%        (p=0.000 n=10+10)
Match/Easy0/32K-8               1.87GB/s ± 0%  9.07GB/s ± 1%  +384.24%         (p=0.000 n=7+10)
Match/Easy0/1M-8                1.90GB/s ± 0%  4.83GB/s ± 0%  +154.56%         (p=0.000 n=8+10)
Match/Easy0/32M-8               1.90GB/s ± 0%  4.53GB/s ± 0%  +138.62%         (p=0.000 n=7+10)

Compared to in 1.7:

Match/Easy0/32-8                  59.5ns ± 0%    59.2ns ± 1%   -0.45%         (p=0.008 n=9+10)
Match/Easy0/1K-8                   226ns ± 1%     231ns ± 1%   +2.30%        (p=0.000 n=10+10)
Match/Easy0/32K-8                 3.73µs ± 2%    3.61µs ± 1%   -3.12%        (p=0.000 n=10+10)
Match/Easy0/1M-8                   206µs ± 1%     217µs ± 0%   +5.34%        (p=0.000 n=10+10)
Match/Easy0/32M-8                 7.03ms ± 1%    7.40ms ± 0%   +5.23%        (p=0.000 n=10+10)

Fixes #17456

Change-Id: I38b2fabcaed7119cc4bf37007ba7bfe7504c8f9f
Reviewed-on: https://go-review.googlesource.com/31690
Run-TryBot: Ilya Tocar <ilya.tocar@intel.com>
Reviewed-by: Keith Randall <khr@golang.org>

1 files changed, 36 insertions, 2 deletions

diff --git a/src/strings/strings_amd64.go b/src/strings/strings_amd64.go
index 5e26ee2c97..23a98d5945 100644
--- a/src/strings/strings_amd64.go
+++ b/src/strings/strings_amd64.go
@@ -29,8 +29,6 @@ func Index(s, sep string) int {
 		return 0
 	case n == 1:
 		return IndexByte(s, sep[0])
-	case n <= shortStringLen:
-		return indexShortStr(s, sep)
 	case n == len(s):
 		if sep == s {
 			return 0
@@ -38,6 +36,42 @@ func Index(s, sep string) int {
 		return -1
 	case n > len(s):
 		return -1
+	case n <= shortStringLen:
+		// Use brute force when s and sep both are small
+		if len(s) <= 64 {
+			return indexShortStr(s, sep)
+		}
+		c := sep[0]
+		i := 0
+		t := s[:len(s)-n+1]
+		fails := 0
+		for i < len(t) {
+			if t[i] != c {
+				// IndexByte skips 16/32 bytes per iteration,
+				// so it's faster than indexShortStr.
+				o := IndexByte(t[i:], c)
+				if o < 0 {
+					return -1
+				}
+				i += o
+			}
+			if s[i:i+n] == sep {
+				return i
+			}
+			fails++
+			i++
+			// Switch to indexShortStr when IndexByte produces too many false positives.
+			// Too many means more that 1 error per 8 characters.
+			// Allow some errors in the beginning.
+			if fails > (i+16)/8 {
+				r := indexShortStr(s[i:], sep)
+				if r >= 0 {
+					return r + i
+				}
+				return -1
+			}
+		}
+		return -1
 	}
 	// Rabin-Karp search
 	hashsep, pow := hashStr(sep)