fastcgo

Fast Cgo calls using an assembly trampoline.

Supported

	Window	MacOS	Linux
x86_64	✅	✅	✅
ARM64	❌	✅	❓

✅ = Working and tested.
❓ = May work, unknown, no tests.
❌ = Not currently supported.

Warning

This is very unsafe and lets you call C using the current's goroutine thread's system stack internal to Go. It monopolizes the thread to itself and completly bypass the regular scheduler premptions. It also disrupts the GC while blocked inside C code. Use at your own risk.

Why?

Two reasons:

• Workaround for a scheduling issue when the call lasts longer than 20 microseconds which is causing me visible stutter when calling glfwSwapBuffers() with VSync enabled.
• Bring down the Cgo overhead from 50ns to 3ns (see below).

Benchmark

With go test -run=- -bench=. ./bench:

cpu: Intel(R) Core(TM) i7-7700K CPU @ 4.20GHz
BenchmarkGO-8           1000000000               0.4139 ns/op
BenchmarkCGO-8          23076256                 51.72 ns/op
BenchmarkFastCGO-8      414632949                2.885 ns/op

Example

package main

/*
#include <stdio.h>
void example(int x) {
	printf("Hello %d", x);
}
*/
import "C"
import "github.com/nitrix/fastcgo"

func main() {
	fastcgo.UnsafeCall1(C.example, 42)
}

API

func UnsafeCall1(fn unsafe.Pointer, arg0 uintptr)
func UnsafeCall2(fn unsafe.Pointer, arg0 uintptr, arg1 uintptr)
func UnsafeCall3(fn unsafe.Pointer, arg0 uintptr, arg1 uintptr, arg2 uintptr)
func UnsafeCall4(fn unsafe.Pointer, arg0 uintptr, arg1 uintptr, arg2 uintptr, arg3 uintptr)

func UnsafeCall1Return1(fn unsafe.Pointer, arg0 uintptr) uintptr
func UnsafeCall2Return1(fn unsafe.Pointer, arg0 uintptr, arg1 uintptr) uintptr
func UnsafeCall3Return1(fn unsafe.Pointer, arg0 uintptr, arg1 uintptr, arg2 uintptr) uintptr
func UnsafeCall4Return1(fn unsafe.Pointer, arg0 uintptr, arg1 uintptr, arg2 uintptr, arg3 uintptr) uintptr

License

This is free and unencumbered software released into the public domain. See the UNLICENSE file for more details.