[swift-evolution] Contiguous Memory and the Effect of Borrowing on Safety

[Dave Abrahams]

on Thu Nov 10 2016, John McCall <rjmccall-AT-apple.com> wrote:

>> On Nov 10, 2016, at 9:31 AM, Joe Groff <jgroff at apple.com> wrote:
>>> On Nov 8, 2016, at 9:29 AM, John McCall <rjmccall at apple.com> wrote:
>>> 
>>>> On Nov 8, 2016, at 7:44 AM, Joe Groff via swift-evolution <swift-evolution at swift.org> wrote:
>>>>> On Nov 7, 2016, at 3:55 PM, Dave Abrahams via swift-evolution <swift-evolution at swift.org>
> wrote:
>
>>>>> 
>>>>> 
>>>>> on Mon Nov 07 2016, John McCall <swift-evolution at swift.org> wrote:
>>>>> 
>>>>>>> On Nov 6, 2016, at 1:20 PM, Dave Abrahams via swift-evolution <swift-evolution at swift.org> wrote:
>>>>>>> 
>>>>>>> 
>>>>>>> Given that we're headed for ABI (and thus stdlib API) stability, I've
>>>>>>> been giving lots of thought to the bottom layer of our collection
>>>>>> 
>>>>>>> abstraction and how it may limit our potential for efficiency.  In
>>>>>>> particular, I want to keep the door open for optimizations that work on
>>>>>>> contiguous memory regions.  Every cache-friendly data structure, even if
>>>>>>> it is not an array, contains contiguous memory regions over which
>>>>>>> operations can often be vectorized, that should define boundaries for
>>>>>>> parallelism, etc.  Throughout Cocoa you can find patterns designed to
>>>>>>> exploit this fact when possible (NSFastEnumeration).  Posix I/O bottoms
>>>>>>> out in readv/writev, and MPI datatypes essentially boil down to
>>>>>>> identifying the contiguous parts of data structures.  My point is that
>>>>>>> this is an important class of optimization, with numerous real-world
>>>>>>> examples.
>>>>>>> 
>>>>>>> If you think about what it means to build APIs for contiguous memory
>>>>>>> into abstractions like Sequence or Collection, at least without
>>>>>>> penalizing the lowest-level code, it means exposing UnsafeBufferPointers
>>>>>>> as a first-class part of the protocols, which is really
>>>>>>> unappealing... unless you consider that *borrowed* UnsafeBufferPointers
>>>>>>> can be made safe.  
>>>>>>> 
>>>>>>> [Well, it's slightly more complicated than that because
>>>>>>> UnsafeBufferPointer is designed to bypass bounds checking in release
>>>>>>> builds, and to ensure safety you'd need a BoundsCheckedBuffer—or
>>>>>>> something—that checks bounds unconditionally... but] the point remains
>>>>>>> that
>>>>>>> 
>>>>>>> A thing that is unsafe when it's arbitrarily copied can become safe if
>>>>>>> you ensure that it's only borrowed (in accordance with well-understood
>>>>>>> lifetime rules).
>>>>>> 
>>>>>> UnsafeBufferPointer today is a copyable type.  Having a borrowed value
>>>>>> doesn't prevent you from making your own copy, which could then escape
>>>>>> the scope that was guaranteeing safety.
>>>>>> 
>>>>>> This is fixable, of course, but it's a more significant change to the
>>>>>> type and how it would be used.
>>>>> 
>>>>> It sounds like you're saying that, to get static safety benefits from
>>>>> ownership, we'll need a whole parallel universe of safe move-only
>>>>> types. Seems a cryin' shame.
>>>> 
>>>> We've discussed the possibility of types being able to control
>>>> their "borrowed" representation. Even if this isn't something we
>>>> generalize, arrays and contiguous buffers might be important
>>>> enough to the language that your safe BufferPointer could be
>>>> called 'borrowed ArraySlice<T>', with the owner backreference
>>>> optimized out of the borrowed representation. Perhaps Array's own
>>>> borrowed representation would benefit from acting like a slice
>>>> rather than a whole-buffer borrow too.
>>> 
>>> The disadvantage of doing this is that it much more heavily
>>> penalizes the case where we actually do a copy from a borrowed
>>> reference — it becomes an actual array copy, not just a reference
>>> bump.
>> 
>> Fair point, though the ArraySlice/Array dichotomy strikes me as
>> already kind of encouraging this—you might pass ArraySlices down
>> into your algorithm, but we encourage people to use Array at storage
>> and API boundaries, forcing copies.
>
> Fair point.  In practice, though, I think most algorithms won't need
> to "escape" that array slice.

I disagree. I'm working on some generic matching algorithms (to lay the
foundation for String search and regexes).  There's going to be a broad
category of functions in this area that work on Strings and return
SubStrings, or work on Collections and return slices thereof.  Often
they'll be called from a context where the resultant slices don't
outlive the collection, but they still do need to be returned.

>> From a philosophical perspective of making systems Swift feel like
>> "the same language" as Swift today, it feels better to me to try to
>> express this as making our high-level safe abstractions efficient
>> rather than making our low-level unsafe abstractions safe. Given our
>> short-term goals for the borrow model as I understand them, I don't
>> think we can really make a BufferPointer-like type safe in the way
>> Dave is suggesting, since the pointer fields *inside* the struct
>> need to be first class lifetime-qualified rather than the value of
>> the struct itself. Since Array and ArraySlice already communicate an
>> ownership stake in the memory they reference, a borrowed Array or
>> ArraySlice value *would* safely and efficiently provide access to
>> contiguous memory with only support for first-order
>> borrowed/consumed property declarations and not full first class
>> lifetime support.
>
> I agree.
>
> John.

-- 
-Dave