为什么切片不断从堆栈中逃脱？

Question

我正在尝试解决leetcode问题排列。但是当我用-benchmem进行测试时，我发现它分配太多了，达到1957年的分配/操作时permute([]int{1,2,3,4,5,6})

我发现它在生成子数字目标时逃逸到堆中。即使我试图分配[6]int，并使用不安全的包来构建切片，它仍然.moved to heap

我的问题是，为什么切片逃逸到堆，以及如何在堆栈上分配切片？

这是我的代码：

package main

import (

    "fmt"

    "reflect"

    "unsafe"

)

func permute(nums []int) [][]int {

    resLen := 1

    for i := 1; i<= len(nums);i ++{

        resLen *= i

    }

    // pre allocate

    res := make([][]int, resLen)

    for i := range res{

        res[i] = make([]int, 0, len(nums))

    }

    build(res, nums)

    return res

}

func build(res [][]int,targets []int){

    step := len(res) / len(targets)

    for i := range targets{

        for j := i*step; j < (i+1) * step; j ++{

            res[j] = append(res[j], targets[i])

        }

        if len(targets) != 1{

            var ab = [6]int{}

            var buff []int

            var bp  *reflect.SliceHeader

            bp = (*reflect.SliceHeader)(unsafe.Pointer(&buff))

            bp.Data = uintptr(unsafe.Pointer(&ab))

            bp.Cap = 6

            buff = append(buff, targets[:i]...)

            buff = append(buff, targets[i+1:]...)

            build(res[i*step:(i+1)*step], buff)

        }

    }

    return

}

func main() {

    nums := []int{1,2,3}

    res := permute(nums)

    fmt.Println(res)

}

build函数没有不安全，但逃逸到堆：

func build(res [][]int, targets []int) {

    step := len(res) / len(targets)

    for i := range targets {

        for j := i * step; j < (i+1)*step; j++ {

            res[j] = append(res[j], targets[i])

        }

        if len(targets) != 1 {

            buff := make([]int, 0, 6) //  make([]int, 0, 6) escapes to heap

            buff = append(buff, targets[:i]...)

            buff = append(buff, targets[i+1:]...)

            build(res[i*step:(i+1)*step], buff)

        }

    }

    return

}

还有我的测试用例：

package main

import "testing"

func Benchmark(b *testing.B){

    for i:=0;i<b.N;i++{

        permute([]int{1,2,3,4,5,6})

    }

}

当我运行时，它报告go build -gcflags="-m"./main.go:32:8: moved to heap: ab

Answer 1

尝试使用颠覆编译器只会使转义分析更难完成其工作，从而阻止切片被堆栈分配。只需分配单个切片，并在每次循环迭代中重复使用它：unsafe.Pointerfunc build(res [][]int, targets []int) {&nbsp; &nbsp; buff := make([]int, 0, 6)&nbsp; &nbsp; step := len(res) / len(targets)&nbsp; &nbsp; for i := range targets {&nbsp; &nbsp; &nbsp; &nbsp; buff = buff[:0]&nbsp; &nbsp; &nbsp; &nbsp; for j := i * step; j < (i+1)*step; j++ {&nbsp; &nbsp; &nbsp; &nbsp; &nbsp; &nbsp; res[j] = append(res[j], targets[i])&nbsp; &nbsp; &nbsp; &nbsp; }&nbsp; &nbsp; &nbsp; &nbsp; if len(targets) != 1 {&nbsp; &nbsp; &nbsp; &nbsp; &nbsp; &nbsp; buff = append(buff, targets[:i]...)&nbsp; &nbsp; &nbsp; &nbsp; &nbsp; &nbsp; buff = append(buff, targets[i+1:]...)&nbsp; &nbsp; &nbsp; &nbsp; &nbsp; &nbsp; build(res[i*step:(i+1)*step], buff)&nbsp; &nbsp; &nbsp; &nbsp; }&nbsp; &nbsp; }&nbsp; &nbsp; return}这可以通过编译器正确优化./main.go:26:17: make([]int, 0, 6) does not escape并且只会导致所需的分配：Benchmark-8&nbsp; 44607&nbsp; 26838 ns/op&nbsp; 52992 B/op&nbsp; 721 allocs/op