Question on 'cmd/compile: handle partially overlapping assignments' behavior on ARM64.
Josh Peterson
I stumbled across a performance regression on ARM64 for go 1.18.6 and 1.19.1 that wasn't present in earlier releases. In the following benchmark, you can see `BenchmarkSliceOfArray` experiences a 70% increase in
execution time while `BenchmarkSliceOfInt` remains unchanged.
BenchmarkSliceOfArray 413121 2855 ns/op
BenchmarkSliceOfArray 216478 4881 ns/op
Through the use of `pprof` and `GOSSAFUNC go build`, it was observed that the latest release employs `runtime.memmove` and accounts for the change in performance. I don't see the invocation of `memmove` or the perfromance degradation on AMD64.
I believe this can be traced back to - https://go-review.googlesource.com/c/go/+/425234.
Is this new behavior in this scenario correct or unindented?
Should I open an issue for this?
I am a newbie, so forgive me if there are errors in my approach.
----
go version go1.19 darwin/arm64
% GOMAXPROCS=1 go1.19 test -cpuprofile cpu.prof -bench .
goos: darwin
goarch: arm64
pkg: foo/bar
BenchmarkSliceOfInt 572226 2066 ns/op
BenchmarkSliceOfArray 413121 2855 ns/op
PASS
ok foo/bar 3.488s
----
go version go1.19.1 darwin/arm64
% GOMAXPROCS=1 go test -cpuprofile cpu.prof -bench .
goos: darwin
goarch: arm64
pkg: foo/bar
BenchmarkSliceOfInt 527084 2065 ns/op
BenchmarkSliceOfArray 216478 4881 ns/op
PASS
ok foo/bar 2.468s
ok foo/bar 3.538s
----
package datamove
type Array [2]int
func moveSliceOfArrayData(s []Array) []Array {
for i := 1; i < len(s); i++ {
s[i-1], s[i] = s[i], s[i-1]
}
return s
}
func moveSliceOfIntData(s []int) []int {
for i := 1; i < len(s); i++ {
s[i-1], s[i] = s[i], s[i-1]
}
return s
}
----
package datamove
import (
"testing"
)
var resultInt []int
var resultArray []Array
func BenchmarkSliceOfInt(b *testing.B) {
var r []int
for n := 0; n < b.N; n++ {
s := make([]int, 1000)
r = moveSliceOfIntData(s)
}
resultInt = r
}
func BenchmarkSliceOfArray(b *testing.B) {
var r []Array
for n := 0; n < b.N; n++ {
s := make([]Array, 1000)
r = moveSliceOfArrayData(s)
}
resultArray = r
}
peterGo
Run your benchmarks with option -benchmem.
Version: devel go1.20-1eeb257b88 Tue Sep 20 02:58:09 2022 +0000
Sizeof(Array{}): 16
Sizeof(int(0)): 8
goarch: amd64
BenchmarkSliceOfInt-4 391646 2583 ns/op 8192 B/op 1 allocs/op
BenchmarkSliceOfArray-4 236977 4869 ns/op 16384 B/op 1 allocs/op
Version: go1.17.13
Sizeof(Array{}): 16
Sizeof(int(0)): 8
goarch: amd64
BenchmarkSliceOfInt-4 426343 2383 ns/op 8192 B/op 1 allocs/op
BenchmarkSliceOfArray-4 233817 4771 ns/op 16384 B/op 1 allocs/op
The difference is type int is 8 bytes and type [2]int (type Array) is 16 bytes.
Change the type of Array to 8 bytes (type [1]int).
Version: devel go1.20-1eeb257b88 Tue Sep 20 02:58:09 2022 +0000
Sizeof(Array{}): 8
Sizeof(int(0)): 8
goarch: amd64
BenchmarkSliceOfInt-4 405793 2562 ns/op 8192 B/op 1 allocs/op
BenchmarkSliceOfArray-4 469999 2540 ns/op 8192 B/op 1 allocs/op
Version: go1.17.13
Sizeof(Array{}): 8
Sizeof(int(0)): 8
goarch: amd64
BenchmarkSliceOfInt-4 440340 2378 ns/op 8192 B/op 1 allocs/op
BenchmarkSliceOfArray-4 481071 2400 ns/op 8192 B/op 1 allocs/op
There is no significant difference.
peter
Brian Candler
I am unable to reproduce your results on AMD64. I don't see any regression.
peterGo
Jan Mercl
> Yes, I read that too. A fundamental tenet of science is reproducibility. I was able to reproduce the AMD64 results and documented it. A step forward.
6:02 PM (7 minutes ago) > I was able to reproduce the AMD64 results
and documented it.
peterGo
Josh Peterson
BenchmarkSliceOfArray 262800 4552 ns/op 16384 B/op 1 allocs/op
go version go1.19.1 darwin/arm64
BenchmarkSliceOfArray 229941 4607 ns/op 16384 B/op 1 allocs/op
go version go1.19 darwin/arm64
BenchmarkSliceOfArray 449023 2580 ns/op 16384 B/op 1 allocs/op
go version go1.18.6 darwin/arm64
BenchmarkSliceOfArray 239430 4377 ns/op 16384 B/op 1 allocs/op
go version go1.18.5 darwin/arm64
BenchmarkSliceOfArray 506563 2349 ns/op 16384 B/op 1 allocs/op
go version go1.18.4 darwin/arm64
BenchmarkSliceOfArray 508051 2362 ns/op 16384 B/op 1 allocs/op
go version go1.18.3 darwin/arm64
BenchmarkSliceOfArray 431294 2357 ns/op 16384 B/op 1 allocs/op
go version go1.18.2 darwin/arm64
BenchmarkSliceOfArray 425918 2715 ns/op 16384 B/op 1 allocs/op
go version go1.18.1 darwin/arm64
BenchmarkSliceOfArray 434580 2351 ns/op 16384 B/op 1 allocs/op
go version go1.18 darwin/arm64
BenchmarkSliceOfArray 415366 2482 ns/op 16384 B/op 1 allocs/op