diff --git a/src/cmd/compile/internal/test/inl_test.go b/src/cmd/compile/internal/test/inl_test.go
index 201f5773e97..e59104df531 100644
--- a/src/cmd/compile/internal/test/inl_test.go
+++ b/src/cmd/compile/internal/test/inl_test.go
@@ -123,6 +123,9 @@ func TestIntendedInlining(t *testing.T) {
 			"AppendRune",
 			"ValidRune",
 		},
+		"unicode/utf16": {
+			"Decode",
+		},
 		"reflect": {
 			"Value.Bool",
 			"Value.Bytes",
diff --git a/src/unicode/utf16/utf16.go b/src/unicode/utf16/utf16.go
index 38d8be60602..1c6d2c66c30 100644
--- a/src/unicode/utf16/utf16.go
+++ b/src/unicode/utf16/utf16.go
@@ -103,23 +103,31 @@ func AppendRune(a []uint16, r rune) []uint16 {
 // Decode returns the Unicode code point sequence represented
 // by the UTF-16 encoding s.
 func Decode(s []uint16) []rune {
-	a := make([]rune, len(s))
-	n := 0
+	// Preallocate capacity to hold up to 64 runes.
+	// Decode inlines, so the allocation can live on the stack.
+	buf := make([]rune, 0, 64)
+	return decode(s, buf)
+}
+
+// decode appends to buf the Unicode code point sequence represented
+// by the UTF-16 encoding s and return the extended buffer.
+func decode(s []uint16, buf []rune) []rune {
 	for i := 0; i < len(s); i++ {
+		var ar rune
 		switch r := s[i]; {
 		case r < surr1, surr3 <= r:
 			// normal rune
-			a[n] = rune(r)
+			ar = rune(r)
 		case surr1 <= r && r < surr2 && i+1 < len(s) &&
 			surr2 <= s[i+1] && s[i+1] < surr3:
 			// valid surrogate sequence
-			a[n] = DecodeRune(rune(r), rune(s[i+1]))
+			ar = DecodeRune(rune(r), rune(s[i+1]))
 			i++
 		default:
 			// invalid surrogate sequence
-			a[n] = replacementChar
+			ar = replacementChar
 		}
-		n++
+		buf = append(buf, ar)
 	}
-	return a[:n]
+	return buf
 }
diff --git a/src/unicode/utf16/utf16_test.go b/src/unicode/utf16/utf16_test.go
index be339b1fdf1..a5a503d3874 100644
--- a/src/unicode/utf16/utf16_test.go
+++ b/src/unicode/utf16/utf16_test.go
@@ -5,6 +5,7 @@
 package utf16_test
 
 import (
+	"internal/testenv"
 	"reflect"
 	"testing"
 	"unicode"
@@ -103,6 +104,22 @@ var decodeTests = []decodeTest{
 	{[]uint16{0xdfff}, []rune{0xfffd}},
 }
 
+func TestAllocationsDecode(t *testing.T) {
+	testenv.SkipIfOptimizationOff(t)
+
+	for _, tt := range decodeTests {
+		allocs := testing.AllocsPerRun(10, func() {
+			out := Decode(tt.in)
+			if out == nil {
+				t.Errorf("Decode(%x) = nil", tt.in)
+			}
+		})
+		if allocs > 0 {
+			t.Errorf("Decode allocated %v times", allocs)
+		}
+	}
+}
+
 func TestDecode(t *testing.T) {
 	for _, tt := range decodeTests {
 		out := Decode(tt.in)