cmd/compile: teach the compiler that arguments passed to {strings|bytes}.{Index*|Count|Has*} don't escape

[valyala]

The issue

Go tip doesn't know that string / []byte arguments passed to functions like strings.IndexByte don't escape. This leads to unnecessary allocation and copy in the following code:

package stringsIndex_test

import (
        "strings"
        "testing"
)


func BenchmarkStringsIndexByte(b *testing.B) {
        bb := []byte("foobar baz")
        b.ReportAllocs()
        for i := 0; i < b.N; i++ {
                // string(bb) unnecessarily allocates new string and copies bb contents to it.
                n := strings.IndexByte(string(bb), ' ')
                Sink += n
        }
}

var Sink int

Benchmark results indicate an unnecessary allocation:

BenchmarkStringsIndexByte 	50000000	        30.5 ns/op	      16 B/op	       1 allocs/op

The solution

Mark strings and bytes functions accepting string / []byte as go:noescape in order to eliminate unnecessary allocations in the code above.

[ianlancetaylor]

CC @dr2chase

I don't think we want to start using go:noescape in places like this. But perhaps the compiler can figure it out.

[valyala]

I don't think we want to start using go:noescape in places like this. But perhaps the compiler can figure it out.

I think it will be too difficult figuring out this for functions written in assembly. Probably, it would be OK adding go:noescape for assembly-written functions?

[valyala]

This change may have significant positive impact on standard packages, since then escape analysis may prove that arguments to many functions written in Go that call strings.Index* and co don't escape.

[ianlancetaylor]

Yes, sorry, it's fine to use go:noescape for functions written in assembly.

[mundaym]

IndexByte is already marked with go:noescape:

go/src/internal/bytealg/indexbyte_native.go

Lines 9 to 10 in a2f72cc

	//go:noescape
	func IndexByte(b []byte, c byte) int

I suspect this is just a case where the string cast could be cleverer.

[randall77]

I guess we could fix this, but why are you using strings.IndexByte(string(bb), ' ') instead of bytes.IndexByte(bb, ' ')?

[valyala]

IndexByte is already marked with go:noescape

Then the []byte -> string conversion with a memory allocation and a copy for non-escaped argument looks like a bug.

why are you using strings.IndexByte(string(bb), ' ') instead of bytes.IndexByte(bb, ' ')

This is just an artificial example :)
The strings.Index(string(b), stringPrefix) is more real-life. It may be used interchangeably with bytes.Index(b, []byte(stringPrefix)), since there is no a preferred code here.

[quasilyte]

The non-escaping string(b) still can allocate if it doesn't fit the small tmp buffer.
I think #27148 (comment) can be relevant here.

bytes.Index(b, []byte(stringPrefix))

This is a preferred way for now, since prefix is usually smaller than haystack itself, and it can probably fit fixes 32-byte stack allocated buffer.

[quasilyte]

Here is a simple benchmarking code:

package foo

import (
	"bytes"
	"strings"
	"testing"
)

var haystack = bytes.Repeat([]byte("a"), 100)
var needle = "aaa"

func BenchmarkStringsIndex(b *testing.B) {
	for i := 0; i < b.N; i++ {
		_ = strings.Index(string(haystack), needle)
	}
}

func BenchmarkBytesIndex(b *testing.B) {
	for i := 0; i < b.N; i++ {
		_ = bytes.Index(haystack, []byte(needle))
	}
}

The results:

BenchmarkStringsIndex-8 20000000  73.3 ns/op 112 B/op 1 allocs/op
BenchmarkBytesIndex-8   100000000 15.1 ns/op   0 B/op 0 allocs/op

I've written a simple check that detects such calls that can be replaced with bytes.Index to aid in optimization sessions of big projects (https://go-critic.github.io/overview#indexAlloc-ref).