cmd/compile: skip slicebytetostring when argument doesn't escape

[bradfitz]

In this function,

func hi() string {
        b := make([]byte, 0, 100) // 1st alloc
        b = append(b, "Hello, "...)
        b = append(b, "world.\n"...)
        return string(b) // 2nd alloc, but could be removed.
}

The final line currently generates a call to runtime.slicebytetostring, causing a copy
of b to be allocated.

But the compiler already knows that b doesn't escape.  It could instead return a string
header re-using the []byte memory.

[rsc]

Comment 1:

Labels changed: added go1.3maybe.

[rsc]

Comment 2:

Labels changed: added release-none, removed go1.3maybe.

[rsc]

Comment 3:

Labels changed: added repo-main.

[josharian]

Comment 4:

I've poked at this a bit and wanted to get some feedback before I continued.
As written, hi only allocs once, since b is allocated on the stack. A quick benchmem run
bears this out. If b's cap is much larger or unknown at compile time, then yes, two
allocs occur.
I cobbled together a very crude patch that returns a string header in case cases, as
suggested. It is possibly underpowered, but it found only a few opportunities in the
stdlib for such an optimization:
strings/replace.go:502
strings/strings.go:367
strings/strings.go:430
strconv/quote.go:82
strconv/quote.go:349
encoding/base64/base64.go:121
os/env.go:26
path/filepath/path.go:314
net/url/url.go:168
net/url/url.go:211
Running the existing benchmarks, the only alloc count improvements are:
strings BenchmarkByteStringMatch: 2 -> 1
strings BenchmarkHTMLEscapeNew: 2 -> 1
strings BenchmarkCountTortureOverlapping: 4 -> 2
strconv BenchmarkUnquoteHard: 2 -> 1
net/url BenchmarkString: 44 -> 39
Extending to rune slices adds only two call sites:
compress/gzip/gunzip.go:118
testing/quick/quick.go:127
I'd need non-trivial direction and input to bring my patch up to snuff...or more likely,
rewrite it from scratch after some hints. (I am finding my way by touch; all the tests
pass, though, so that's something.)
Given all the above, should I continue to pursue this?

[gopherbot]

Comment 5 by unclejacksons:

Avoiding allocations for bytes to string conversions where possible would be very useful.
This would help in hot loops and in code where lots of items are processed. One less
allocation can make a big difference.

[OneOfOne]

Maybe this should change the milestone to 1.7?

[bradfitz]

No, this isn't critical to 1.7 or any particular release. This is not a regression or anything.

[bradfitz]

If this is fixed, revert https://golang.org/cl/25005

[kevinburke]

Is anyone working on this? I'd be interested in trying to fix it, though I've never worked on the compiler part of this project before.

[randall77]

I don't think anyone is working on it.

[aclements]

Note that "b does not escape" is not a sufficient condition. The real question is "is this the only reference to b's backing array?", which is a related but different question. For example,

func hi() string {
        b := make([]byte, 0, 100) // 1st alloc
        b = append(b, "Hello, "...)
        b = append(b, "world.\n"...)
        s := string(b)
        b[0] = 'X'
        return s
}

b does not escape here, but it's not safe to use stringtoslicebytetmp because b's backing array is still reachable from the stack.

This may be a liveness question, not an escape question. If the last point b's backing array is live is at the string conversion, then it should be safe to reuse it.

[gopherbot]

CL https://golang.org/cl/27327 mentions this issue.

[kevinburke]

I'm going to take another crack at this this week.

[rsc]

OK but too late for Go 1.8. We'll review what you send when we get to Go 1.9.

[kevinburke]

Understood! I still need to get up to speed with e.g. "how to figure out whether a variable escapes" but there has never been a better time to learn.

[bradfitz]

See also: #18990

[rkravchik]

@rsc it's late for Go 1.9 too :)

[JAicewizard]

is there any news on this?

[DmitriyMV]

@JAicewizard I think it was postponed because of the introduction of strings.Builder.

[pascaldekloe]

Seem fixed: https://play.golang.org/p/koSvIIvUnRs

[martisch]

Small fixed length slices that do not escape are allocated on the stack. I think this has been for a while. Having a variable byte slice and then a string conversion as far as I know still causes two heap allocations if the resulting string escapes even if there are no other references to the byte slice.