Thanks! Did you use a tool to find these spots or something else?

yeah, repro by running:
golangci-lint run ./... -E prealloc --disable-all

I don't understand why you're replacing appends with assignments to a specific index. append is the nicer way to do this, please don't change that.

I don’t have a preference, but assignments are technically faster. That said, I don’t think this matters here.

I don't understand why you're replacing appends with assignments to a specific index. append is the nicer way to do this, please don't change that.

I don’t have a preference, but assignments are technically faster. That said, I don’t think this matters here.

I'd expect it to compile to exactly the same code. I ran a benchmark, and the two perform exactly the same. If at all, appending is slightly faster, but probably not statistically significant.

goos: darwin
goarch: arm64
pkg: github.com/libp2p/go-libp2p/p2p/protocol/holepunch
BenchmarkAppend-10      24834030                48.23 ns/op
BenchmarkAssign-10      24644346                48.32 ns/op

const l = 128

func BenchmarkAppend(b *testing.B) {
	a := make([]int, 0, l)
	for i := 0; i < b.N; i++ {
		a = a[:0]
		for j := 0; j < l; j++ {
			a = append(a, j)
		}
	}
}

func BenchmarkAssign(b *testing.B) {
	a := make([]int, l)
	for i := 0; i < b.N; i++ {
		for j := 0; j < l; j++ {
			a[j] = j
		}
	}
}

The reason people tend to prefer append is that it's safer: If you get your length calculation wrong, all that happens is that you do an additional alloc. If you get your length calculation wrong and do an assignment, your process will panic.

I've always gotten better perf from assignment than appending. Same code on my machine:

goos: darwin
goarch: amd64
pkg: github.com/estensen/benchmarks/tmp
cpu: Intel(R) Core(TM) i7-9750H CPU @ 2.60GHz
BenchmarkAppend-12    	15084253	        68.56 ns/op
BenchmarkAssign-12    	17883320	        66.12 ns/op
PASS
ok  	github.com/estensen/benchmarks/tmp	2.536s

But I do agree that it is safer to use append. I'll revert it.

Thank you!

Let me preface this by saying this doesn't matter at all :)

I'd expect it to compile to exactly the same code.

I went into a rabbit hole here because this doesn't make sense to me. Append has to do more work than an assign (which should be a single instruction). So it should be slower. Sure enough if we load up the code in godbolt we see that the append has one more compare and mov than assign should have. I say should have because this benchmark actually is flawed since the compiler optimizes the assignment and makes it the same as the noop (you don't see a MOVQ in the asm output). To fix this you actually should do something like this. This optimization actually makes it a bit more surprising that append (an extra cmp and 2 movs) is as fast as a noop, but then again CPUs are pretty good at branch prediction and pipelining and parallelizing especially in this very simple case.

Also I'm impressed with how well the compiler optimized the append loop. It really keeps the hot path very small. If we run without compiler optimizations (gcflags=-N) we also see the expected speed diffs.

Results from my tests here: https://gist.github.com/MarcoPolo/18b39366c6f57270059d0ef03c29351b#file-results-md