[golang-dev] Re: code review 14419054: spec: clarify re-use of underlying arrays in slice oper... (issue 14419054)

On Wed, Oct 9, 2013 at 8:58 PM, wrote:

I don't think we are actually changing anything in the spec. If we are
changing something, then what we are changing is a requirement that the
implementation always reuse the existing backing array if there is
space. Of course in most cases the implementation will reuse the
existing array, and of course it will always be the case that appending
in a loop will be efficient. But for example if you write
a := make([]byte, 0, 1)
a = append(a, 'b')
then I think it is OK if the implementation guesses that more append
calls are coming and preemptively creates a new, larger, backing array.

On Wed, Oct 9, 2013 at 9:14 PM, wrote:

sensible implementation will use the underlying array. Note that the
previous wording didn't say re-use is required. Is was unclear.

The new wording does not require it either, but it clear about it.
Making it a requirement would be a change to the spec. That's ok, but
that's not the intent of this CL.

On 10 October 2013 06:35, wrote:

https://codereview.appspot.com/14419054/diff/13001/doc/go_spec.html
File doc/go_spec.html (right):

https://codereview.appspot.com/14419054/diff/13001/doc/go_spec.html#newcode5357
doc/go_spec.html:5357: For instance, if the capacity of <code>s</code>
is not large enough to fit the additional

The spec should not prescribe a specific implementation if not
absolutely necessary, hence the _may_. Of course any sensible
implementation _will_ (at the moment) reuse the underlying array if
there's enough capacity since that is (usually) the most efficient
approach. Furthermore, it's a very straight-forward approach, which is
another reason why there's no need to prescribe it (it's a different
story if it were difficult to achieve with so much code that relies on
it for efficiency - we don't have that case).

It's important that it's not prescribed, though. Here's a (admittedly
somewhat contrived, but illustrative) example why: A future smarter
compiler might optimize the statement

copy(dst, append(src, x))

such that the generated code directly copies the result of append(src,
x) into dst, _without_ storing that intermediate result in the
underlying array of src, even if there would be enough space for it. If
we require append to _always_ reuse the underlying array independent of
context, we give up this optimization (and others that we might not be
thinking about now). We don't want to restrict implementations this way.

Furthermore, code that _does_ rely on the result of append reusing the
underlying array if there's enough space should be explicit about it by
not using append in these cases. That code _must_ know that space is
available and so _can_ use easy alternatives to append (say, copy).

In short, requiring re-use when possible is a mistake. Any sensible
implementation should be free to choose the best possible approach.


https://codereview.appspot.com/14419054/

--

---You received this message because you are subscribed to the Google Groups
"golang-dev" group.
To unsubscribe from this group and stop receiving emails from it, send an
email to [email protected].
For more options, visit https://groups.google.com/groups/opt_out.

On Thu, Oct 10, 2013 at 3:04 AM, roger peppe wrote:

I appreciate the arguments for keeping this undefined, but I
couldn't guarantee that I haven't written some code along
these kinds of lines in the past:

For example, I might easily write this, and I doubt
I'd have called it out in a code review:

// fill populates the given slice with values from the given map. The
// slice and the map are of the same length.
func fill(xs []int, m map[string]int) {
xs := xs[:0]
for _, x := range m {
xs = append(xs, x)
}
}

rather than this:

func fill(xs []int, m map[string]int) {
i := 0
for _, x := range m {
xs[i] = x
i++
}
}

If a compiler *were* to implement different behaviour in
this respect, I suspect it could lead to some very subtle
breakages. I'd prefer to have it defined, even if it
leads to some potential inefficiencies in the future.

The copy(dst, append(src, x)) optimisation will still be
possible in some circumstances, I think.

On 10 October 2013 17:25, Robert Griesemer wrote:
I'd argue that the fill function should return the filled slice. The comment
does say that the slice and map are of the same length (capacity rather, for
the slice), but there's no check in that code that that's true. So this may
have silently failed if there's a bug elsewhere.