cmd/compile: b.Loop un-optimizer regression heap allocate objects passed to non-leaky functions it used to stack allocate #76482
Description
Go version
go version go1.26-devel_c048a9a11f Wed Nov 26 16:20:21 2025 -0800 linux/amd64
Output of go env in your module/workspace:
.What did you do?
package main
import "testing"
type Data struct {
v [1<<10]float32
}
func NonPointerReturn() Data {
return Data{}
}
func ReturnAsInput(d *Data) {
*d = Data{}
}
func PointerReturn() *Data {
return &Data{}
}
var x float32
var n = 123
func useData(d *Data) {
x = d.v[n]
}
func Benchmark_NonPointerReturn(b *testing.B) {
for b.Loop() {
var d = NonPointerReturn()
useData(&d)
}
}
func Benchmark_NonPointerReturnB(b *testing.B) {
for b.Loop() {
var d = new(Data)
*d = NonPointerReturn()
useData(d)
}
}
func Benchmark_ReturnAsInput(b *testing.B) {
for b.Loop() {
var d = new(Data)
ReturnAsInput(d)
useData(d)
}
}
func Benchmark_PointerReturn(b *testing.B) {
for b.Loop() {
var d = PointerReturn()
useData(d)
}
}What did you see happen?
Similar benchmark results between tip and 1.25 toolchain.
What did you expect to see?
The benchmark results look quite different.
$ gotv 1.25 test -bench=. -benchmem main_test.go
[Run]: $HOME/.cache/gotv/tag_go1.25.4/bin/go test -bench=. -benchmem main_test.go
goos: linux
goarch: amd64
cpu: Intel(R) Core(TM) i5-4200M CPU @ 2.50GHz
Benchmark_NonPointerReturn-4 10438954 114.4 ns/op 0 B/op 0 allocs/op
Benchmark_NonPointerReturnB-4 5389276 222.7 ns/op 0 B/op 0 allocs/op
Benchmark_ReturnAsInput-4 10283925 109.5 ns/op 0 B/op 0 allocs/op
Benchmark_PointerReturn-4 1522304 779.8 ns/op 4096 B/op 1 allocs/op
PASS
ok command-line-arguments 4.713s
$ gotv :tip test -bench=. -benchmem main_test.go
[Run]: $HOME/.cache/gotv/bra_master/bin/go test -bench=. -benchmem main_test.go
goos: linux
goarch: amd64
cpu: Intel(R) Core(TM) i5-4200M CPU @ 2.50GHz
Benchmark_NonPointerReturn-4 567326 1982 ns/op 8192 B/op 2 allocs/op
Benchmark_NonPointerReturnB-4 11456282 103.7 ns/op 0 B/op 0 allocs/op
Benchmark_ReturnAsInput-4 10468320 107.6 ns/op 0 B/op 0 allocs/op
Benchmark_PointerReturn-4 22553568 52.91 ns/op 0 B/op 0 allocs/op
PASS
ok command-line-arguments 4.639s
It looks Go 1.26 makes some improvements, but also makes some bad things.
BTW, I expect the four benchmarks are on par (whether the size of Data is smaller than 1<<15 or not). Still not on tip.