cmd/compile: optimize len([]rune(string))

Adds a new runtime function to count runes in a string.
Modifies the compiler to detect the pattern len([]rune(string))
and replaces it with the new rune counting runtime function.

RuneCount/lenruneslice/ASCII                  27.8ns ± 2%  14.5ns ± 3%  -47.70%  (p=0.000 n=10+10)
RuneCount/lenruneslice/Japanese                126ns ± 2%    60ns ± 2%  -52.03%  (p=0.000 n=10+10)
RuneCount/lenruneslice/MixedLength             104ns ± 2%    50ns ± 1%  -51.71%  (p=0.000 n=10+9)

Fixes #24923

Change-Id: Ie9c7e7391a4e2cca675c5cdcc1e5ce7d523948b9
Reviewed-on: https://go-review.googlesource.com/108985
Run-TryBot: Martin Möhrmann <[email protected]>
TryBot-Result: Gobot Gobot <[email protected]>
Reviewed-by: Josh Bleecher Snyder <[email protected]>

Author: martisch

src/cmd/compile/internal/gc/builtin.go

@@ -55,10 +55,12 @@ func slicebytetostringtmp([]byte) string
 func slicerunetostring(*[32]byte, []rune) string
 func stringtoslicebyte(*[32]byte, string) []byte
 func stringtoslicerune(*[32]rune, string) []rune
-func decoderune(string, int) (retv rune, retk int)
 func slicecopy(to any, fr any, wid uintptr) int
 func slicestringcopy(to any, fr any) int
 
+func decoderune(string, int) (retv rune, retk int)
+func countrunes(string) int
+
 // interface conversions
 func convI2I(typ *byte, elem any) (ret any)
 

src/cmd/compile/internal/gc/builtin/runtime.go

@@ -1098,7 +1098,14 @@ func (o *Order) expr(n, lhs *Node) *Node {
 		OSTRARRAYBYTE,
 		OSTRARRAYBYTETMP,
 		OSTRARRAYRUNE:
-		o.call(n)
+
+		if isRuneCount(n) {
+			// len([]rune(s)) is rewritten to runtime.countrunes(s) later.
+			n.Left.Left = o.expr(n.Left.Left, nil)
+		} else {
+			o.call(n)
+		}
+
 		if lhs == nil || lhs.Op != ONAME || instrumenting {
 			n = o.copyExpr(n, n.Type, false)
 		}

src/cmd/compile/internal/gc/order.go

@@ -538,6 +538,12 @@ opswitch:
 		n.Left = walkexpr(n.Left, init)
 
 	case OLEN, OCAP:
+		if isRuneCount(n) {
+			// Replace len([]rune(string)) with runtime.countrunes(string).
+			n = mkcall("countrunes", n.Type, init, conv(n.Left.Left, types.Types[TSTRING]))
+			break
+		}
+
 		n.Left = walkexpr(n.Left, init)
 
 		// replace len(*[10]int) with 10.
@@ -4085,3 +4091,9 @@ func canMergeLoads() bool {
 	}
 	return false
 }
+
+// isRuneCount reports whether n is of the form len([]rune(string)).
+// These are optimized into a call to runtime.runecount.
+func isRuneCount(n *Node) bool {
+	return Debug['N'] == 0 && !instrumenting && n.Op == OLEN && n.Left.Op == OSTRARRAYRUNE
+}

src/cmd/compile/internal/gc/walk.go

@@ -9,6 +9,7 @@ import (
 	"strconv"
 	"strings"
 	"testing"
+	"unicode/utf8"
 )
 
 // Strings and slices that don't escape and fit into tmpBuf are stack allocated,
@@ -110,6 +111,43 @@ var stringdata = []struct{ name, data string }{
 	{"MixedLength", "$Ѐࠀက퀀𐀀\U00040000\U0010FFFF"},
 }
 
+var sinkInt int
+
+func BenchmarkRuneCount(b *testing.B) {
+	// Each sub-benchmark counts the runes in a string in a different way.
+	b.Run("lenruneslice", func(b *testing.B) {
+		for _, sd := range stringdata {
+			b.Run(sd.name, func(b *testing.B) {
+				for i := 0; i < b.N; i++ {
+					sinkInt += len([]rune(sd.data))
+				}
+			})
+		}
+	})
+	b.Run("rangeloop", func(b *testing.B) {
+		for _, sd := range stringdata {
+			b.Run(sd.name, func(b *testing.B) {
+				for i := 0; i < b.N; i++ {
+					n := 0
+					for range sd.data {
+						n++
+					}
+					sinkInt += n
+				}
+			})
+		}
+	})
+	b.Run("utf8.RuneCountInString", func(b *testing.B) {
+		for _, sd := range stringdata {
+			b.Run(sd.name, func(b *testing.B) {
+				for i := 0; i < b.N; i++ {
+					sinkInt += utf8.RuneCountInString(sd.data)
+				}
+			})
+		}
+	})
+}
+
 func BenchmarkRuneIterate(b *testing.B) {
 	b.Run("range", func(b *testing.B) {
 		for _, sd := range stringdata {

src/runtime/string_test.go

@@ -39,6 +39,15 @@ const (
 	hicb = 0xBF // 1011 1111
 )
 
+// countrunes returns the number of runes in s.
+func countrunes(s string) int {
+	n := 0
+	for range s {
+		n++
+	}
+	return n
+}
+
 // decoderune returns the non-ASCII rune at the start of
 // s[k:] and the index after the rune in s.
 //

src/runtime/utf8.go

@@ -212,14 +212,25 @@ func TestSequencing(t *testing.T) {
 	}
 }
 
-// Check that a range loop and a []int conversion visit the same runes.
+func runtimeRuneCount(s string) int {
+	return len([]rune(s)) // Replaced by gc with call to runtime.countrunes(s).
+}
+
+// Check that a range loop, len([]rune(string)) optimization and
+// []rune conversions visit the same runes.
 // Not really a test of this package, but the assumption is used here and
-// it's good to verify
-func TestIntConversion(t *testing.T) {
+// it's good to verify.
+func TestRuntimeConversion(t *testing.T) {
 	for _, ts := range testStrings {
+		count := RuneCountInString(ts)
+		if n := runtimeRuneCount(ts); n != count {
+			t.Errorf("%q: len([]rune()) counted %d runes; got %d from RuneCountInString", ts, n, count)
+			break
+		}
+
 		runes := []rune(ts)
-		if RuneCountInString(ts) != len(runes) {
-			t.Errorf("%q: expected %d runes; got %d", ts, len(runes), RuneCountInString(ts))
+		if n := len(runes); n != count {
+			t.Errorf("%q: []rune() has length %d; got %d from RuneCountInString", ts, n, count)
 			break
 		}
 		i := 0

src/unicode/utf8/utf8_test.go

@@ -0,0 +1,15 @@
+// asmcheck
+
+// Copyright 2018 The Go Authors. All rights reserved.
+// Use of this source code is governed by a BSD-style
+// license that can be found in the LICENSE file.
+
+package codegen
+
+// This file contains code generation tests related to the handling of
+// string types.
+
+func CountRunes(s string) int { // Issue #24923
+	// amd64:`.*countrunes`
+	return len([]rune(s))
+}