CODEWITHSUNDEEP
goassembly
chmike
chmike

Reputation: 22154

Assembly function implementation with fallback

I have a function written on Go that I would like to optimize in assembly. I currently only want to write it for amd64 which is the most frequent target. Assembly for other targets may be added later. By default it should use the Go code.

Unfortunately I didn't manage to achieve this. Looking at the documentation I found online it seam that I have to declare an external function, and provide the assembly for every possible target.

I can provide a Go function with a different name, and then in the assembly code jump to it. This looks tedious and I can't ensure I have provided the assembly file for all possible targets.

Is there a way to provide a Go function, and an alternate version to use instead when compiling for particular targets (e.g. amd64) ?

Upvotes: 4

Views: 302

Answers (1)

user142162
user142162

Reputation:

The standard way to do this is to have at least four files, three of which are protected by build constraints, either by use of the +build tag or by the filename.

  • wrapper.go:
    • Contains function and documentation for a wrapper function which calls the conditionally compiled implementation function
  • impl_generic.go
    • Go implementation for non-optimized platforms
  • impl_asm.go
    • Function declaration of any assembly implementations
  • impl_amd64.s
    • amd64 assembly implementation of your function

wrapper.go

package mypkg

import "fmt"

func Wrapper(s string) {
    fmt.Println(impl(s))
}

impl_generic.go

// +build !amd64

package mypkg

func impl(s string) string {
    // TODO: non-assembly implementation
}

impl_asm.go

// +build amd64

package mypkg

// defined in *.s

func impl(s string) string

impl_amd64.s

TEXT ·impl(SB),$0
    // TODO: AMD64 implementation
    RET

To add an assembly implementation for another architecture:

  1. Concat ,!$arch to the first impl_generic.go build constraint item
  2. Add $arch to the impl_asm.go build constraint list
  3. Write the assembly implementation in impl_$arch.s

Upvotes: 5

Related Questions