maps: Performance Regression in maps.Clone after Go 1.24 Refactor

[1911860538]

Go 1.24 refactored the map runtime using SwissTable.

When I was using Go 1.24, I found that,
compared to using a for loop to copy a map,
the performance of the built-in maps.Clone was actually worse.
At the same time, in versions prior to Go 1.24,
the performance of maps.Clone was the best.

Below is the code I used for performance testing:

package clone

import (
	"maps"
	"testing"
)

func clone1(m map[string]string) map[string]string {
	if m == nil {
		return nil
	}
	m2 := make(map[string]string, len(m))
	for k, v := range m {
		m2[k] = v
	}
	return m2
}

func clone2[M ~map[K]V, K comparable, V any](m M) M {
	if m == nil {
		return nil
	}
	m2 := make(M, len(m))
	for k, v := range m {
		m2[k] = v
	}
	return m2
}

var m = map[string]string{
	"k1": "v1",
	"k2": "v2",
	"k3": "v3",
	"k4": "v4",
}

func BenchmarkClone1(b *testing.B) {
	for i := 0; i < b.N; i++ {
		clone1(m)
	}
}

func BenchmarkClone2(b *testing.B) {
	for i := 0; i < b.N; i++ {
		clone2(m)
	}
}

func BenchmarkStdClone(b *testing.B) {
	for i := 0; i < b.N; i++ {
		maps.Clone(m)
	}
}

Performance test results for different Go versions are as follows:

Go 1.24

goos: darwin
goarch: arm64
pkg: test_code/test_map/clone
cpu: Apple M3 Pro
BenchmarkClone1
BenchmarkClone1-11      	16550962	        72.64 ns/op
BenchmarkClone2
BenchmarkClone2-11      	16820404	        71.59 ns/op
BenchmarkStdClone
BenchmarkStdClone-11    	 9009817	       134.1 ns/op
PASS

Process finished with the exit code 0

Go 1.23

goos: darwin
goarch: arm64
pkg: test_code/test_map/clone
cpu: Apple M3 Pro
BenchmarkClone1
BenchmarkClone1-11      	15726506	        77.62 ns/op
BenchmarkClone2
BenchmarkClone2-11      	14940602	        80.03 ns/op
BenchmarkStdClone
BenchmarkStdClone-11    	20451465	        59.58 ns/op
PASS

Process finished with the exit code 0

The results for Go 1.22 and Go 1.21 are similar to Go 1.23.

[gabyhelp]

Related Issues

• maps: optimized Clone with swissmaps #70836
• maps: maps.Clone reference semantics when cloning a map with large value types #64474 (closed)

Related Code Changes

• maps,runtime: improve maps.Clone

_{(Emoji vote if this was helpful or unhelpful; more detailed feedback welcome in this discussion.)}

[seankhliao]

That benchmark looks invalid, the return value is discarded and can be optimized away.
Ensuring the value is retained produces more consistent results: https://go.dev/play/p/jBtvROSBTzV

[1911860538]

That benchmark looks invalid, the return value is discarded and can be optimized away. Ensuring the value is retained produces more consistent results: https://go.dev/play/p/jBtvROSBTzV

I conducted benchmarks for Go 1.23 and Go 1.24 locally based on your code.
Below is the test code. Using the same test code, I tested it with Go 1.24 and Go 1.23 separately.
For the same maps.Clone code, the execution speed in Go 1.23 is twice that of Go 1.24.

clone_test.go

package clone2

import (
	"maps"
	"testing"
)

var g map[string]string

var m = map[string]string{
	"k1": "v1",
	"k2": "v2",
	"k3": "v3",
	"k4": "v4",
}

func BenchmarkStdClone(b *testing.B) {
	for i := 0; i < b.N; i++ {
		g = maps.Clone(m)
	}
}

Go1.24 output

GOROOT=/Users/huangzl/sdk/go1.24.0 #gosetup
GOPATH=/Users/huangzl/huangzhiwen/about_code/go/test_code/GOPATH #gosetup
GOPROXY=https://goproxy.cn,direct #gosetup
/Users/huangzl/sdk/go1.24.0/bin/go test -c -o /Users/huangzl/Library/Caches/JetBrains/GoLand2024.3/tmp/GoLand/___gobench_test_code_test_map_clone2.test test_code/test_map/clone2 #gosetup
/Users/huangzl/Library/Caches/JetBrains/GoLand2024.3/tmp/GoLand/___gobench_test_code_test_map_clone2.test -test.v -test.paniconexit0 -test.bench . -test.run ^$ #gosetup
goos: darwin
goarch: arm64
pkg: test_code/test_map/clone2
cpu: Apple M3 Pro
BenchmarkStdClone
BenchmarkStdClone-11    	 7470522	       137.7 ns/op
PASS

Process finished with the exit code 0

Go1.23 output

GOROOT=/Users/huangzl/sdk/go1.23.1 #gosetup
GOPATH=/Users/huangzl/huangzhiwen/about_code/go/test_code/GOPATH #gosetup
GOPROXY=https://goproxy.cn,direct #gosetup
/Users/huangzl/sdk/go1.23.1/bin/go test -c -o /Users/huangzl/Library/Caches/JetBrains/GoLand2024.3/tmp/GoLand/___gobench_test_code_test_map_clone2.test test_code/test_map/clone2 #gosetup
/Users/huangzl/Library/Caches/JetBrains/GoLand2024.3/tmp/GoLand/___gobench_test_code_test_map_clone2.test -test.v -test.paniconexit0 -test.bench . -test.run ^$ #gosetup
goos: darwin
goarch: arm64
pkg: test_code/test_map/clone2
cpu: Apple M3 Pro
BenchmarkStdClone
BenchmarkStdClone-11    	20376992	        58.62 ns/op
PASS

Process finished with the exit code 0

[mateusz834]

AFAIAC in go1.24 the maps.Clone is not optimized specially as it was for non-swiss maps in 1.23

(#70836)

[1911860538]

AFAIAC in go1.24 the maps.Clone is not optimized specially as it was for non-swiss maps in 1.23

(#70836)

So the performance regression of maps.Clone is a fact.