bytes: use Index in Count

Similar to https://go-review.googlesource.com/28586,
but for package bytes instead of strings.
This provides simpler code and some performance gain.
Also update strings.Count to use the same code.

On AMD64 with heavily optimized Index I see:

name             old time/op    new time/op     delta
Count/10-6         47.3ns ± 0%     36.8ns ± 0%    -22.35%  (p=0.000 n=10+10)
Count/32-6          286ns ± 0%       38ns ± 0%    -86.71%  (p=0.000 n=10+10)
Count/4K-6         50.1µs ± 0%      4.4µs ± 0%    -91.18%  (p=0.000 n=10+10)
Count/4M-6         48.1ms ± 1%      4.5ms ± 0%    -90.56%  (p=0.000 n=10+9)
Count/64M-6         784ms ± 0%       73ms ± 0%    -90.73%  (p=0.000 n=10+10)
CountEasy/10-6     28.4ns ± 0%     31.0ns ± 0%     +9.23%  (p=0.000 n=10+10)
CountEasy/32-6     30.6ns ± 0%     37.0ns ± 0%    +20.92%  (p=0.000 n=10+10)
CountEasy/4K-6      186ns ± 0%      198ns ± 0%     +6.45%  (p=0.000 n=9+10)
CountEasy/4M-6      233µs ± 2%      234µs ± 2%       ~     (p=0.912 n=10+10)
CountEasy/64M-6    6.70ms ± 0%     6.68ms ± 1%       ~     (p=0.762 n=8+10)

name             old speed      new speed       delta
Count/10-6        211MB/s ± 0%    272MB/s ± 0%    +28.77%  (p=0.000 n=10+9)
Count/32-6        112MB/s ± 0%    842MB/s ± 0%   +652.84%  (p=0.000 n=10+10)
Count/4K-6       81.8MB/s ± 0%  927.6MB/s ± 0%  +1033.63%  (p=0.000 n=10+9)
Count/4M-6       87.2MB/s ± 1%  924.0MB/s ± 0%   +959.25%  (p=0.000 n=10+9)
Count/64M-6      85.6MB/s ± 0%  922.9MB/s ± 0%   +978.31%  (p=0.000 n=10+10)
CountEasy/10-6    352MB/s ± 0%    322MB/s ± 0%     -8.41%  (p=0.000 n=10+10)
CountEasy/32-6   1.05GB/s ± 0%   0.87GB/s ± 0%    -17.35%  (p=0.000 n=9+10)
CountEasy/4K-6   22.0GB/s ± 0%   20.6GB/s ± 0%     -6.33%  (p=0.000 n=10+10)
CountEasy/4M-6   18.0GB/s ± 2%   18.0GB/s ± 2%       ~     (p=0.912 n=10+10)
CountEasy/64M-6  10.0GB/s ± 0%   10.0GB/s ± 1%       ~     (p=0.762 n=8+10)

On 386, without asm version of Index:

Count/10-6         57.0ns ± 0%     56.9ns ± 0%   -0.11%  (p=0.006 n=10+9)
Count/32-6          340ns ± 0%      274ns ± 0%  -19.48%  (p=0.000 n=10+9)
Count/4K-6         49.5µs ± 0%     37.1µs ± 0%  -24.96%  (p=0.000 n=10+10)
Count/4M-6         51.1ms ± 0%     38.2ms ± 0%  -25.21%  (p=0.000 n=10+10)
Count/64M-6         818ms ± 0%      613ms ± 0%  -25.07%  (p=0.000 n=8+10)
CountEasy/10-6     60.0ns ± 0%     70.4ns ± 0%  +17.34%  (p=0.000 n=10+10)
CountEasy/32-6     81.1ns ± 0%     94.0ns ± 0%  +15.97%  (p=0.000 n=9+10)
CountEasy/4K-6     4.37µs ± 0%     4.39µs ± 0%   +0.30%  (p=0.000 n=10+9)
CountEasy/4M-6     4.43ms ± 0%     4.43ms ± 0%     ~     (p=0.579 n=10+10)
CountEasy/64M-6    70.9ms ± 0%     70.9ms ± 0%     ~     (p=0.912 n=10+10)

name             old speed      new speed       delta
Count/10-6        176MB/s ± 0%    176MB/s ± 0%   +0.10%  (p=0.000 n=10+9)
Count/32-6       93.9MB/s ± 0%  116.5MB/s ± 0%  +24.06%  (p=0.000 n=10+9)
Count/4K-6       82.7MB/s ± 0%  110.3MB/s ± 0%  +33.26%  (p=0.000 n=10+10)
Count/4M-6       82.1MB/s ± 0%  109.7MB/s ± 0%  +33.70%  (p=0.000 n=10+10)
Count/64M-6      82.0MB/s ± 0%  109.5MB/s ± 0%  +33.46%  (p=0.000 n=8+10)
CountEasy/10-6    167MB/s ± 0%    142MB/s ± 0%  -14.75%  (p=0.000 n=9+10)
CountEasy/32-6    395MB/s ± 0%    340MB/s ± 0%  -13.77%  (p=0.000 n=10+10)
CountEasy/4K-6    936MB/s ± 0%    934MB/s ± 0%   -0.29%  (p=0.000 n=10+9)
CountEasy/4M-6    947MB/s ± 0%    946MB/s ± 0%     ~     (p=0.591 n=10+10)
CountEasy/64M-6   947MB/s ± 0%    947MB/s ± 0%     ~     (p=0.867 n=10+10)

Change-Id: Ia76b247372b6f5b5d23a9f10253a86536a5153b3
Reviewed-on: https://go-review.googlesource.com/36489
Reviewed-by: Brad Fitzpatrick <bradfitz@golang.org>

diff --git a/src/bytes/bytes.go b/src/bytes/bytes.go
index 406a38257a74c420266dc013782fd9c91256686f..a273a919110c8fe9a6c5be280dd3042e37ceed56 100644 (file)
--- a/src/bytes/bytes.go
+++ b/src/bytes/bytes.go
@@ -49,33 +49,19 @@ func explode(s []byte, n int) [][]byte {
 // Count counts the number of non-overlapping instances of sep in s.
 // If sep is an empty slice, Count returns 1 + the number of Unicode code points in s.
 func Count(s, sep []byte) int {
-       n := len(sep)
-       if n == 0 {
+       n := 0
+       // special case
+       if len(sep) == 0 {
                return utf8.RuneCount(s) + 1
        }
-       if n > len(s) {
-               return 0
-       }
-       count := 0
-       c := sep[0]
-       i := 0
-       t := s[:len(s)-n+1]
-       for i < len(t) {
-               if t[i] != c {
-                       o := IndexByte(t[i:], c)
-                       if o < 0 {
-                               break
-                       }
-                       i += o
-               }
-               if n == 1 || Equal(s[i:i+n], sep) {
-                       count++
-                       i += n
-                       continue
+       for {
+               i := Index(s, sep)
+               if i == -1 {
+                       return n
                }
-               i++
+               n++
+               s = s[i+len(sep):]
        }
-       return count
 }
 
 // Contains reports whether subslice is within b.

diff --git a/src/strings/strings.go b/src/strings/strings.go
index 60a281a6ac50dda9204bed74d5d510d0596a183b..2b1fbab5b2b8d2b2964d3a86b182e03687933b53 100644 (file)
--- a/src/strings/strings.go
+++ b/src/strings/strings.go
@@ -76,18 +76,17 @@ func hashStrRev(sep string) (uint32, uint32) {
 // If sep is an empty string, Count returns 1 + the number of Unicode code points in s.
 func Count(s, sep string) int {
        n := 0
-       // special cases
+       // special case
        if len(sep) == 0 {
                return utf8.RuneCountInString(s) + 1
        }
-       offset := 0
        for {
-               i := Index(s[offset:], sep)
+               i := Index(s, sep)
                if i == -1 {
                        return n
                }
                n++
-               offset += i + len(sep)
+               s = s[i+len(sep):]
        }
 }