unicode: performance improvements (API change)
*** There is an API change here: the introduction of the
LatinOffset int in the RangeTable struct. ***
* Avoid checking Latin range multiple times for non-Latin runes.
* Use linear search when it is faster than binary search.
go test -calibrate runs the calibration for where the linear/binary
crossover should be.
benchmark old MB/s new MB/s speedup
BenchmarkFields 36.27 41.43 1.14x
BenchmarkFieldsFunc 36.23 41.38 1.14x
The speedup here is evenly split between the linear scans
and the LatinOffset change. Both are about 1.07x.
R=r
CC=bradfitz, gobot, golang-dev
https://golang.org/cl/6526048
diff --git a/src/pkg/unicode/digit.go b/src/pkg/unicode/digit.go
index 4800bd6..53171b3 100644
--- a/src/pkg/unicode/digit.go
+++ b/src/pkg/unicode/digit.go
@@ -9,5 +9,5 @@
if r <= MaxLatin1 {
return '0' <= r && r <= '9'
}
- return Is(Digit, r)
+ return isExcludingLatin(Digit, r)
}
diff --git a/src/pkg/unicode/graphic.go b/src/pkg/unicode/graphic.go
index 0de90eb..1105688 100644
--- a/src/pkg/unicode/graphic.go
+++ b/src/pkg/unicode/graphic.go
@@ -78,13 +78,13 @@
if uint32(r) <= MaxLatin1 {
return properties[uint8(r)]&(pLu|pLl) != 0
}
- return Is(Letter, r)
+ return isExcludingLatin(Letter, r)
}
// IsMark reports whether the rune is a mark character (category M).
func IsMark(r rune) bool {
// There are no mark characters in Latin-1.
- return Is(Mark, r)
+ return isExcludingLatin(Mark, r)
}
// IsNumber reports whether the rune is a number (category N).
@@ -92,7 +92,7 @@
if uint32(r) <= MaxLatin1 {
return properties[uint8(r)]&pN != 0
}
- return Is(Number, r)
+ return isExcludingLatin(Number, r)
}
// IsPunct reports whether the rune is a Unicode punctuation character
@@ -119,7 +119,7 @@
}
return false
}
- return Is(White_Space, r)
+ return isExcludingLatin(White_Space, r)
}
// IsSymbol reports whether the rune is a symbolic character.
@@ -127,5 +127,5 @@
if uint32(r) <= MaxLatin1 {
return properties[uint8(r)]&pS != 0
}
- return Is(Symbol, r)
+ return isExcludingLatin(Symbol, r)
}
diff --git a/src/pkg/unicode/letter.go b/src/pkg/unicode/letter.go
index be48455..8239557 100644
--- a/src/pkg/unicode/letter.go
+++ b/src/pkg/unicode/letter.go
@@ -19,8 +19,9 @@
// The two slices must be in sorted order and non-overlapping.
// Also, R32 should contain only values >= 0x10000 (1<<16).
type RangeTable struct {
- R16 []Range16
- R32 []Range32
+ R16 []Range16
+ R32 []Range32
+ LatinOffset int // number of entries in R16 with Hi <= MaxLatin1
}
// Range16 represents of a range of 16-bit Unicode code points. The range runs from Lo to Hi
@@ -80,14 +81,31 @@
UpperLower = MaxRune + 1 // (Cannot be a valid delta.)
)
-// is16 uses binary search to test whether rune is in the specified slice of 16-bit ranges.
+// linearMax is the maximum size table for linear search for non-Latin1 rune.
+// Derived by running 'go test -calibrate'.
+const linearMax = 18
+
+// is16 reports whether r is in the sorted slice of 16-bit ranges.
func is16(ranges []Range16, r uint16) bool {
+ if len(ranges) <= linearMax || r <= MaxLatin1 {
+ for i := range ranges {
+ range_ := &ranges[i]
+ if r < range_.Lo {
+ return false
+ }
+ if r <= range_.Hi {
+ return (r-range_.Lo)%range_.Stride == 0
+ }
+ }
+ return false
+ }
+
// binary search over ranges
lo := 0
hi := len(ranges)
for lo < hi {
m := lo + (hi-lo)/2
- range_ := ranges[m]
+ range_ := &ranges[m]
if range_.Lo <= r && r <= range_.Hi {
return (r-range_.Lo)%range_.Stride == 0
}
@@ -100,8 +118,21 @@
return false
}
-// is32 uses binary search to test whether rune is in the specified slice of 32-bit ranges.
+// is32 reports whether r is in the sorted slice of 32-bit ranges.
func is32(ranges []Range32, r uint32) bool {
+ if len(ranges) <= linearMax {
+ for i := range ranges {
+ range_ := &ranges[i]
+ if r < range_.Lo {
+ return false
+ }
+ if r <= range_.Hi {
+ return (r-range_.Lo)%range_.Stride == 0
+ }
+ }
+ return false
+ }
+
// binary search over ranges
lo := 0
hi := len(ranges)
@@ -122,21 +153,6 @@
// Is tests whether rune is in the specified table of ranges.
func Is(rangeTab *RangeTable, r rune) bool {
- // common case: rune is ASCII or Latin-1.
- if uint32(r) <= MaxLatin1 {
- // Only need to check R16, since R32 is always >= 1<<16.
- r16 := uint16(r)
- for _, r := range rangeTab.R16 {
- if r16 > r.Hi {
- continue
- }
- if r16 < r.Lo {
- return false
- }
- return (r16-r.Lo)%r.Stride == 0
- }
- return false
- }
r16 := rangeTab.R16
if len(r16) > 0 && r <= rune(r16[len(r16)-1].Hi) {
return is16(r16, uint16(r))
@@ -148,13 +164,25 @@
return false
}
+func isExcludingLatin(rangeTab *RangeTable, r rune) bool {
+ r16 := rangeTab.R16
+ if off := rangeTab.LatinOffset; len(r16) > off && r <= rune(r16[len(r16)-1].Hi) {
+ return is16(r16[off:], uint16(r))
+ }
+ r32 := rangeTab.R32
+ if len(r32) > 0 && r >= rune(r32[0].Lo) {
+ return is32(r32, uint32(r))
+ }
+ return false
+}
+
// IsUpper reports whether the rune is an upper case letter.
func IsUpper(r rune) bool {
// See comment in IsGraphic.
if uint32(r) <= MaxLatin1 {
return properties[uint8(r)]&pLu != 0
}
- return Is(Upper, r)
+ return isExcludingLatin(Upper, r)
}
// IsLower reports whether the rune is a lower case letter.
@@ -163,7 +191,7 @@
if uint32(r) <= MaxLatin1 {
return properties[uint8(r)]&pLl != 0
}
- return Is(Lower, r)
+ return isExcludingLatin(Lower, r)
}
// IsTitle reports whether the rune is a title case letter.
@@ -171,7 +199,7 @@
if r <= MaxLatin1 {
return false
}
- return Is(Title, r)
+ return isExcludingLatin(Title, r)
}
// to maps the rune using the specified case mapping.
diff --git a/src/pkg/unicode/letter_test.go b/src/pkg/unicode/letter_test.go
index 2d80562..0ec25ab 100644
--- a/src/pkg/unicode/letter_test.go
+++ b/src/pkg/unicode/letter_test.go
@@ -5,6 +5,10 @@
package unicode_test
import (
+ "flag"
+ "fmt"
+ "runtime"
+ "sort"
"testing"
. "unicode"
)
@@ -427,3 +431,117 @@
}
}
}
+
+// Running 'go test -calibrate' runs the calibration to find a plausible
+// cutoff point for linear search of a range list vs. binary search.
+// We create a fake table and then time how long it takes to do a
+// sequence of searches within that table, for all possible inputs
+// relative to the ranges (something before all, in each, between each, after all).
+// This assumes that all possible runes are equally likely.
+// In practice most runes are ASCII so this is a conservative estimate
+// of an effective cutoff value. In practice we could probably set it higher
+// than what this function recommends.
+
+var calibrate = flag.Bool("calibrate", false, "compute crossover for linear vs. binary search")
+
+func TestCalibrate(t *testing.T) {
+ if !*calibrate {
+ return
+ }
+
+ if runtime.GOARCH == "amd64" {
+ fmt.Printf("warning: running calibration on %s\n", runtime.GOARCH)
+ }
+
+ // Find the point where binary search wins by more than 10%.
+ // The 10% bias gives linear search an edge when they're close,
+ // because on predominantly ASCII inputs linear search is even
+ // better than our benchmarks measure.
+ n := sort.Search(64, func(n int) bool {
+ tab := fakeTable(n)
+ blinear := func(b *testing.B) {
+ tab := tab
+ max := n*5 + 20
+ for i := 0; i < b.N; i++ {
+ for j := 0; j <= max; j++ {
+ linear(tab, uint16(j))
+ }
+ }
+ }
+ bbinary := func(b *testing.B) {
+ tab := tab
+ max := n*5 + 20
+ for i := 0; i < b.N; i++ {
+ for j := 0; j <= max; j++ {
+ binary(tab, uint16(j))
+ }
+ }
+ }
+ bmlinear := testing.Benchmark(blinear)
+ bmbinary := testing.Benchmark(bbinary)
+ fmt.Printf("n=%d: linear=%d binary=%d\n", n, bmlinear.NsPerOp(), bmbinary.NsPerOp())
+ return bmlinear.NsPerOp()*100 > bmbinary.NsPerOp()*110
+ })
+ fmt.Printf("calibration: linear cutoff = %d\n", n)
+}
+
+func fakeTable(n int) []Range16 {
+ var r16 []Range16
+ for i := 0; i < n; i++ {
+ r16 = append(r16, Range16{uint16(i*5 + 10), uint16(i*5 + 12), 1})
+ }
+ return r16
+}
+
+func linear(ranges []Range16, r uint16) bool {
+ for i := range ranges {
+ range_ := &ranges[i]
+ if r < range_.Lo {
+ return false
+ }
+ if r <= range_.Hi {
+ return (r-range_.Lo)%range_.Stride == 0
+ }
+ }
+ return false
+}
+
+func binary(ranges []Range16, r uint16) bool {
+ // binary search over ranges
+ lo := 0
+ hi := len(ranges)
+ for lo < hi {
+ m := lo + (hi-lo)/2
+ range_ := &ranges[m]
+ if range_.Lo <= r && r <= range_.Hi {
+ return (r-range_.Lo)%range_.Stride == 0
+ }
+ if r < range_.Lo {
+ hi = m
+ } else {
+ lo = m + 1
+ }
+ }
+ return false
+}
+
+func TestLatinOffset(t *testing.T) {
+ var maps = []map[string]*RangeTable{
+ Categories,
+ FoldCategory,
+ FoldScript,
+ Properties,
+ Scripts,
+ }
+ for _, m := range maps {
+ for name, tab := range m {
+ i := 0
+ for i < len(tab.R16) && tab.R16[i].Hi <= MaxLatin1 {
+ i++
+ }
+ if tab.LatinOffset != i {
+ t.Errorf("%s: LatinOffset=%d, want %d", name, tab.LatinOffset, i)
+ }
+ }
+ }
+}
diff --git a/src/pkg/unicode/maketables.go b/src/pkg/unicode/maketables.go
index fcd14fc..2ed1915 100644
--- a/src/pkg/unicode/maketables.go
+++ b/src/pkg/unicode/maketables.go
@@ -503,6 +503,7 @@
func dumpRange(header string, inCategory Op) {
fmt.Print(header)
next := rune(0)
+ latinOffset := 0
fmt.Print("\tR16: []Range16{\n")
// one Range for each iteration
count := &range16Count
@@ -546,11 +547,17 @@
break
}
}
+ if uint32(hi) <= unicode.MaxLatin1 {
+ latinOffset++
+ }
size, count = printRange(uint32(lo), uint32(hi), uint32(stride), size, count)
// next range: start looking where this range ends
next = hi + 1
}
fmt.Print("\t},\n")
+ if latinOffset > 0 {
+ fmt.Printf("\tLatinOffset: %d,\n", latinOffset)
+ }
fmt.Print("}\n\n")
}
@@ -760,14 +767,17 @@
}
ndecl++
fmt.Printf("var _%s = &RangeTable {\n", name)
- fmt.Print("\tR16: []Range16{\n")
ranges := foldAdjacent(table[name])
+ fmt.Print("\tR16: []Range16{\n")
size := 16
count := &range16Count
for _, s := range ranges {
size, count = printRange(s.Lo, s.Hi, s.Stride, size, count)
}
fmt.Print("\t},\n")
+ if off := findLatinOffset(ranges); off > 0 {
+ fmt.Printf("\tLatinOffset: %d,\n", off)
+ }
fmt.Print("}\n\n")
}
decl.Sort()
@@ -779,6 +789,14 @@
fmt.Print(")\n\n")
}
+func findLatinOffset(ranges []unicode.Range32) int {
+ i := 0
+ for i < len(ranges) && ranges[i].Hi <= unicode.MaxLatin1 {
+ i++
+ }
+ return i
+}
+
const (
CaseUpper = 1 << iota
CaseLower
diff --git a/src/pkg/unicode/tables.go b/src/pkg/unicode/tables.go
index ebd169b..859e53c 100644
--- a/src/pkg/unicode/tables.go
+++ b/src/pkg/unicode/tables.go
@@ -71,6 +71,7 @@
{0xf0000, 0xffffd, 1},
{0x100000, 0x10fffd, 1},
},
+ LatinOffset: 2,
}
var _Cc = &RangeTable{
@@ -78,6 +79,7 @@
{0x0001, 0x001f, 1},
{0x007f, 0x009f, 1},
},
+ LatinOffset: 2,
}
var _Cf = &RangeTable{
@@ -536,6 +538,7 @@
{0x2b740, 0x2b81d, 1},
{0x2f800, 0x2fa1d, 1},
},
+ LatinOffset: 6,
}
var _Ll = &RangeTable{
@@ -682,6 +685,7 @@
{0x1d7c4, 0x1d7c9, 1},
{0x1d7cb, 0x1d7cb, 1},
},
+ LatinOffset: 5,
}
var _Lm = &RangeTable{
@@ -1186,6 +1190,7 @@
{0x1d790, 0x1d7a8, 1},
{0x1d7ca, 0x1d7ca, 1},
},
+ LatinOffset: 3,
}
var _M = &RangeTable{
@@ -1769,6 +1774,7 @@
{0x1d7ce, 0x1d7ff, 1},
{0x1f100, 0x1f10a, 1},
},
+ LatinOffset: 4,
}
var _Nd = &RangeTable{
@@ -1814,6 +1820,7 @@
{0x11066, 0x1106f, 1},
{0x1d7ce, 0x1d7ff, 1},
},
+ LatinOffset: 1,
}
var _Nl = &RangeTable{
@@ -1879,6 +1886,7 @@
{0x1d360, 0x1d371, 1},
{0x1f100, 0x1f10a, 1},
},
+ LatinOffset: 3,
}
var _P = &RangeTable{
@@ -2003,6 +2011,7 @@
{0x110be, 0x110c1, 1},
{0x12470, 0x12473, 1},
},
+ LatinOffset: 10,
}
var _Pc = &RangeTable{
@@ -2053,6 +2062,7 @@
{0xff09, 0xff3d, 52},
{0xff5d, 0xff63, 3},
},
+ LatinOffset: 1,
}
var _Pf = &RangeTable{
@@ -2194,6 +2204,7 @@
{0x110be, 0x110c1, 1},
{0x12470, 0x12473, 1},
},
+ LatinOffset: 7,
}
var _Ps = &RangeTable{
@@ -2222,6 +2233,7 @@
{0xff5b, 0xff5f, 4},
{0xff62, 0xff62, 1},
},
+ LatinOffset: 1,
}
var _S = &RangeTable{
@@ -2409,6 +2421,7 @@
{0x1f680, 0x1f6c5, 1},
{0x1f700, 0x1f773, 1},
},
+ LatinOffset: 9,
}
var _Sc = &RangeTable{
@@ -2425,6 +2438,7 @@
{0xffe0, 0xffe1, 1},
{0xffe5, 0xffe6, 1},
},
+ LatinOffset: 2,
}
var _Sk = &RangeTable{
@@ -2452,6 +2466,7 @@
{0xff3e, 0xff40, 2},
{0xffe3, 0xffe3, 1},
},
+ LatinOffset: 3,
}
var _Sm = &RangeTable{
@@ -2510,6 +2525,7 @@
{0x1d76f, 0x1d789, 26},
{0x1d7a9, 0x1d7c3, 26},
},
+ LatinOffset: 5,
}
var _So = &RangeTable{
@@ -2666,6 +2682,7 @@
{0x1f680, 0x1f6c5, 1},
{0x1f700, 0x1f773, 1},
},
+ LatinOffset: 3,
}
var _Z = &RangeTable{
@@ -2677,6 +2694,7 @@
{0x202f, 0x205f, 48},
{0x3000, 0x3000, 1},
},
+ LatinOffset: 1,
}
var _Zl = &RangeTable{
@@ -2699,6 +2717,7 @@
{0x202f, 0x205f, 48},
{0x3000, 0x3000, 1},
},
+ LatinOffset: 1,
}
// These variables have type *RangeTable.
@@ -3179,6 +3198,7 @@
{0xe0001, 0xe0001, 1},
{0xe0020, 0xe007f, 1},
},
+ LatinOffset: 7,
}
var _Coptic = &RangeTable{
@@ -3649,6 +3669,7 @@
{0xff21, 0xff3a, 1},
{0xff41, 0xff5a, 1},
},
+ LatinOffset: 6,
}
var _Lepcha = &RangeTable{
@@ -4199,6 +4220,7 @@
{0x0041, 0x0046, 1},
{0x0061, 0x0066, 1},
},
+ LatinOffset: 3,
}
var _Bidi_Control = &RangeTable{
@@ -4230,6 +4252,7 @@
{0xfe63, 0xfe63, 1},
{0xff0d, 0xff0d, 1},
},
+ LatinOffset: 1,
}
var _Deprecated = &RangeTable{
@@ -4370,6 +4393,7 @@
{0x1d185, 0x1d18b, 1},
{0x1d1aa, 0x1d1ad, 1},
},
+ LatinOffset: 6,
}
var _Extender = &RangeTable{
@@ -4395,6 +4419,7 @@
{0xaadd, 0xaadd, 1},
{0xff70, 0xff70, 1},
},
+ LatinOffset: 1,
}
var _Hex_Digit = &RangeTable{
@@ -4406,6 +4431,7 @@
{0xff21, 0xff26, 1},
{0xff41, 0xff46, 1},
},
+ LatinOffset: 3,
}
var _Hyphen = &RangeTable{
@@ -4421,6 +4447,7 @@
{0xff0d, 0xff0d, 1},
{0xff65, 0xff65, 1},
},
+ LatinOffset: 2,
}
var _IDS_Binary_Operator = &RangeTable{
@@ -4695,6 +4722,7 @@
{0x1369, 0x1371, 1},
{0x19da, 0x19da, 1},
},
+ LatinOffset: 1,
}
var _Other_ID_Start = &RangeTable{
@@ -4828,6 +4856,7 @@
{0x1d7c4, 0x1d7cb, 1},
{0x1d7ce, 0x1d7ff, 1},
},
+ LatinOffset: 1,
}
var _Other_Uppercase = &RangeTable{
@@ -4868,6 +4897,7 @@
{0xfd3e, 0xfd3f, 1},
{0xfe45, 0xfe46, 1},
},
+ LatinOffset: 15,
}
var _Pattern_White_Space = &RangeTable{
@@ -4878,6 +4908,7 @@
{0x200e, 0x200f, 1},
{0x2028, 0x2029, 1},
},
+ LatinOffset: 3,
}
var _Quotation_Mark = &RangeTable{
@@ -4895,6 +4926,7 @@
{0xff07, 0xff07, 1},
{0xff62, 0xff63, 1},
},
+ LatinOffset: 4,
}
var _Radical = &RangeTable{
@@ -4957,6 +4989,7 @@
{0x11047, 0x11048, 1},
{0x110be, 0x110c1, 1},
},
+ LatinOffset: 3,
}
var _Soft_Dotted = &RangeTable{
@@ -4995,6 +5028,7 @@
{0x1d65e, 0x1d65f, 1},
{0x1d692, 0x1d693, 1},
},
+ LatinOffset: 1,
}
var _Terminal_Punctuation = &RangeTable{
@@ -5069,6 +5103,7 @@
{0x110be, 0x110c1, 1},
{0x12470, 0x12473, 1},
},
+ LatinOffset: 5,
}
var _Unified_Ideograph = &RangeTable{
@@ -5114,6 +5149,7 @@
{0x205f, 0x205f, 1},
{0x3000, 0x3000, 1},
},
+ LatinOffset: 4,
}
// These variables have type *RangeTable.
@@ -5887,6 +5923,7 @@
R32: []Range32{
{0x10400, 0x10427, 1},
},
+ LatinOffset: 3,
}
var foldLt = &RangeTable{
@@ -6001,6 +6038,7 @@
R32: []Range32{
{0x10428, 0x1044f, 1},
},
+ LatinOffset: 4,
}
var foldM = &RangeTable{
