cmd/compile: treat "d = append(make([]T, 0, len(s)), s...)" as "d = make([]T, len(s)); copy(d, s)" #46079
Description
What version of Go are you using (go version)?
$ go version go version go1.16.3 linux/amd64
Does this issue reproduce with the latest release?
Yes.
What did you do?
package main
import (
"testing"
)
type T = int
var sMakeCopy128 []T
func Benchmark_MakeCopy_128(b *testing.B) {
s := make([]T, 128)
b.ResetTimer()
for i := 0; i < b.N; i++ {
sMakeCopy128 = make([]T, len(s))
copy(sMakeCopy128, s)
}
}
var sMakeAppend128 []T
func Benchmark_MakeAppend_128(b *testing.B) {
s := make([]T, 128)
b.ResetTimer()
for i := 0; i < b.N; i++ {
sMakeAppend128 = append(make([]T, 0, len(s)), s...)
}
}
var sMakeCopy1024 []T
func Benchmark_MakeCopy_1024(b *testing.B) {
s := make([]T, 1024)
b.ResetTimer()
for i := 0; i < b.N; i++ {
sMakeCopy1024 = make([]T, len(s))
copy(sMakeCopy1024, s)
}
}
var sMakeAppend1024 []T
func Benchmark_MakeAppend_1024(b *testing.B) {
s := make([]T, 1024)
b.ResetTimer()
for i := 0; i < b.N; i++ {
sMakeAppend1024 = append(make([]T, 0, len(s)), s...)
}
}
var sMakeCopy8192 []T
func Benchmark_MakeCopy_8192(b *testing.B) {
s := make([]T, 8192)
b.ResetTimer()
for i := 0; i < b.N; i++ {
sMakeCopy8192 = make([]T, len(s))
copy(sMakeCopy8192, s)
}
}
var sMakeAppend8192 []T
func Benchmark_MakeAppend_8192(b *testing.B) {
s := make([]T, 8192)
b.ResetTimer()
for i := 0; i < b.N; i++ {
sMakeAppend8192 = append(make([]T, 0, len(s)), s...)
}
}
var sMakeCopy65536 []T
func Benchmark_MakeCopy_65536(b *testing.B) {
s := make([]T, 65536)
b.ResetTimer()
for i := 0; i < b.N; i++ {
sMakeCopy65536 = make([]T, len(s))
copy(sMakeCopy65536, s)
}
}
var sMakeAppend65536 []T
func Benchmark_MakeAppend_65536(b *testing.B) {
s := make([]T, 65536)
b.ResetTimer()
for i := 0; i < b.N; i++ {
sMakeAppend65536 = append(make([]T, 0, len(s)), s...)
}
}What did you expect to see?
No performance difference between the two ways.
What did you see instead?
The larger the length of the slices, the bigger performance difference between the two.
$ go test -bench=. -benchmem -benchtime=3s
goos: linux
goarch: amd64
pkg: a.b/c/ttt
cpu: Intel(R) Core(TM) i5-4210U CPU @ 1.70GHz
Benchmark_MakeCopy_128-4 14257712 254.3 ns/op 1024 B/op 1 allocs/op
Benchmark_MakeAppend_128-4 13141005 268.5 ns/op 1024 B/op 1 allocs/op
Benchmark_MakeCopy_1024-4 1885544 1928 ns/op 8192 B/op 1 allocs/op
Benchmark_MakeAppend_1024-4 1663520 2369 ns/op 8192 B/op 1 allocs/op
Benchmark_MakeCopy_8192-4 268628 14040 ns/op 65536 B/op 1 allocs/op
Benchmark_MakeAppend_8192-4 196272 18757 ns/op 65536 B/op 1 allocs/op
Benchmark_MakeCopy_65536-4 36700 100680 ns/op 524291 B/op 1 allocs/op
Benchmark_MakeAppend_65536-4 24723 136324 ns/op 524288 B/op 1 allocs/op
[Edit]: it looks the allocated bytes/op is the not same for the last two benchmarks. Bug or some inaccuracy?