A real test for cgo code https://github.com/dyu/ffi-overhead

hello.go

package main

/*
#cgo CFLAGS: -Wall -O2 -fPIC
#cgo linux CFLAGS: -DLINUX=1
#cgo LDFLAGS: -Lnewplus -lnewplus -Wl,-rpath='$ORIGIN'/newplus
#include <stdlib.h>
#include "newplus/plus.h"
*/
import "C"
import "testing"

func plusOne(b *testing.B) {
	// load
	C.plusone(C.int(C.current_timestamp()))
	//start := C.current_timestamp()

	var x C.int = 0

	b.ResetTimer()

	for i := 0; i < b.N; i++ {
		x = C.plusone(x)
	}

	//fmt.Println(C.current_timestamp() - start)
}

hello_test.go

package main

import "testing"

func BenchmarkPlusOne(b *testing.B) {
	plusOne(b)
}

how-to-run.txt

ffi-overhead$ rm hello.c hello.cpp # ...to make cgo not try to compile them.

ffi-overhead$ cd newplus
newplus$ gcc -c -o libnewplus.a -Wall -O2 -fPIC plus.c
newplus$ cd ..

ffi-overhead$ go test -c

ffi-overhead$ ./ffi-overhead.test -test.bench . -test.run=xxx -test.cpuprofile=hello.pprof
goos: linux
goarch: amd64
pkg: github.com/dyu/ffi-overhead
BenchmarkPlusOne-4      20000000                81.1 ns/op
PASS

ffi-overhead$ go tool pprof hello.pprof
File: ffi-overhead.test
Build ID: 80a7ea67207dac6817521db32cea644d410500d4
Type: cpu
Time: Aug 19, 2018 at 8:50pm (MSK)
Duration: 1.90s, Total samples = 1.71s (89.81%)
Entering interactive mode (type "help" for commands, "o" for options)
(pprof) top20
Showing nodes accounting for 1.71s, 100% of 1.71s total
      flat  flat%   sum%        cum   cum%
     0.92s 53.80% 53.80%      1.64s 95.91%  runtime.cgocall
     0.25s 14.62% 68.42%      0.25s 14.62%  runtime.casgstatus
     0.19s 11.11% 79.53%      0.22s 12.87%  runtime.exitsyscallfast
     0.12s  7.02% 86.55%      0.13s  7.60%  runtime.reentersyscall
     0.11s  6.43% 92.98%      0.58s 33.92%  runtime.exitsyscall
     0.04s  2.34% 95.32%      1.70s 99.42%  github.com/dyu/ffi-overhead.plusOne
     0.03s  1.75% 97.08%      0.03s  1.75%  runtime.exitsyscallfast_reacquired
     0.02s  1.17% 98.25%      1.66s 97.08%  github.com/dyu/ffi-overhead._Cfunc_plusone
     0.01s  0.58% 98.83%      0.14s  8.19%  runtime.entersyscall
     0.01s  0.58% 99.42%      0.01s  0.58%  runtime.save
     0.01s  0.58%   100%      0.01s  0.58%  runtime.usleep
         0     0%   100%      1.70s 99.42%  github.com/dyu/ffi-overhead.BenchmarkPlusOne
         0     0%   100%      0.01s  0.58%  runtime.mstart
         0     0%   100%      0.01s  0.58%  runtime.mstart1
         0     0%   100%      0.01s  0.58%  runtime.sysmon
         0     0%   100%      1.70s 99.42%  testing.(*B).launch
         0     0%   100%      1.70s 99.42%  testing.(*B).runN
(pprof) 

# Let's see how the test function looks like:
$ go tool objdump ffi-overhead.test
...
TEXT github.com/dyu/ffi-overhead.BenchmarkPlusOne(SB) .../go/src/github.com/dyu/ffi-overhead/hello_test
  hello_test.go:5       0x4e6c30                64488b0c25f8ffffff      MOVQ FS:0xfffffff8, CX                  
  hello_test.go:5       0x4e6c39                483b6110                CMPQ 0x10(CX), SP                       
  hello_test.go:5       0x4e6c3d                7626                    JBE 0x4e6c65                            
  hello_test.go:5       0x4e6c3f                4883ec10                SUBQ $0x10, SP                          
  hello_test.go:5       0x4e6c43                48896c2408              MOVQ BP, 0x8(SP)                        
  hello_test.go:5       0x4e6c48                488d6c2408              LEAQ 0x8(SP), BP                        
  hello_test.go:5       0x4e6c4d                488b442418              MOVQ 0x18(SP), AX                       
  hello_test.go:6       0x4e6c52                48890424                MOVQ AX, 0(SP)                          
  hello_test.go:6       0x4e6c56                e825010000              CALL github.com/dyu/ffi-overhead.plusOne(SB)
  hello_test.go:7       0x4e6c5b                488b6c2408              MOVQ 0x8(SP), BP                        
  hello_test.go:7       0x4e6c60                4883c410                ADDQ $0x10, SP                          
  hello_test.go:7       0x4e6c64                c3                      RET                                     
  hello_test.go:5       0x4e6c65                e8a6c3f6ff              CALL runtime.morestack_noctxt(SB)       
  hello_test.go:5       0x4e6c6a                ebc4                    JMP github.com/dyu/ffi-overhead.BenchmarkPlusO
...

# ...as we can see, the compiler did not inline our plusOne() wrapper.