runtime: don't allocate when converting a single byte to string

[mvdan]

For example:

var StringSink string

func Foo(bs []byte) {
    StringSink = string(bs[0])  // no allocation
    StringSink = string(bs[:1]) // no allocation
    StringSink = string(bs)     // no allocation if len(bs)==1
}

At the moment, all three allocate, easily measured with:

func BenchmarkFoo(b *testing.B) {
        bs := []byte("foobar")
        for i := 0; i < b.N; i++ {
                Foo(bs)
        }
}

There is some precedent to this, I believe. For example:

• strconv uses constant strings for small integers - https://go-review.googlesource.com/c/go/+/37963
• byte-sized values no longer allocate when put in interfaces - https://go-review.googlesource.com/c/go/+/35555

This was even briefly mentioned by @mdempsky in 2016, but as far as I can see it was never followed up on: #17725 (comment)

Another potential route would be for string(bs[0]) and string(bs[:1]) to not allocate thanks to the compiler, as it can statically know that len==1. However, that could be built on top of the runtime optimization, as a form of skipping the len check and calling the non-allocating runtime func directly.

If there is interest in investigating the performance penalty of such a change, I can work on a prototype CL to get some measurements. Unless this was already measured and discarded by someone, and I haven't been able to find a thread for it.

/cc @randall77 @josharian

[randall77]

Go ahead. This is small enough I don't think we need the proposal process for it.

[gopherbot]

Change https://golang.org/cl/97717 mentions this issue: runtime: don't allocate to build strings of length 1

[josharian]

I had a change sitting around in my tree from Sat Mar 4 16:55:03 2017 to do this. Apparently I had decided at the time that it wasn't worth it? Unclear. Anyway, I mailed it. The perf numbers may be stale.

And to echo Keith, for small- to medium-sized optimization ideas, no proposal needed in general.

[mvdan]

Oh, awesome. I used a proposal in case there was a good reason not to do this that I wasn't aware of.