Please run the benchmarks with -count=20 -cpu=1 on a quiet system (no network or browser) and pipe the output to a file then use https://godoc.org/golang.org/x/perf/cmd/benchstat to compare the results to remove some problems with outliers.

I see some variance (between var bug = false vs true) from run to rune and some sub ns differences which is not unusual even on a quiet system:
name old time/op new time/op delta

_Append  44.2ns ± 5%  44.6ns ± 6%   ~     (p=0.765 n=10+9)
_Copy    57.8ns ± 1%  57.4ns ± 2%   ~     (p=0.139 n=8+10)

Sorry but currently I can't shutdown the network and the browser.
The result of the benchmark is:

name     time/op
_Append  36.8ns ± 0%
_Copy    43.0ns ± 0%

name     alloc/op
_Append   0.00B     
_Copy     0.00B     

name     allocs/op
_Append    0.00     
_Copy      0.00

During the weekend I will try to run again the benchmark after a system reboot.

The output does not show a comparison between false and true and no multiple runs.

I used with go tip (to be 1.11 in the future):

• go test -bench=Benchmark -cpu=1 -count=10 > ~/append.false
• change bug variable to false
• go test -bench=Benchmark -cpu=1 -count=10 > ~/append.true
• benchstat ~/append.true ~/append.false

Doesnt show any significant or unusual difference between variable values.

Sorry, I have never used benchstat before; the data I posted was from benchstat append.true.

Here is the correct data:

name     old time/op    new time/op    delta
_Append    36.9ns ± 0%    43.0ns ± 0%  +16.69%  (p=0.000 n=20+17)
_Copy      43.0ns ± 0%    43.0ns ± 0%   -0.07%  (p=0.013 n=20+18)

name     old alloc/op   new alloc/op   delta
_Append     0.00B          0.00B          ~     (all equal)
_Copy       0.00B          0.00B          ~     (all equal)

name     old allocs/op  new allocs/op  delta
_Append      0.00           0.00          ~     (all equal)
_Copy        0.00           0.00          ~     (all equal)

I will try to use go tip during the weekend.

Thanks is that with bug as const or var?

With const.

Ok. I see a speed up with const bug set to true too but not for changing bug between true and false with bug as a var.

That said the loop alignment/code layout is not unexpectedly different for bug const true vs false (as the compiler can optimize some code away) which can lead to performance differences. If so i would not count this as a bug (unless the outcome is wrong) but an opportunity to investigate and change the compiler to generate better code for the const bug is true case.

Removing the "if" block below (and thereby changing the code layout) makes the benchmark a lot slower on my system:

buf := make([]byte, 0, n)
buf = append(buf, data...)
if buf == nil {
  print("true")
}

Our layout pass could use some improvements in general. (I worked on this some in 1.11 and got some improvements but never quite mailed anything.) See also #20356 and #18977 for some other discussion around layout.