[LLVMdev] ErrorOr<> conflicts with unique_ptr<>

[David Blaikie]

On Thu, Nov 21, 2013 at 4:33 PM, Nick Kledzik <kledzik at apple.com> wrote:

>
> On Nov 21, 2013, at 4:07 PM, David Blaikie <dblaikie at gmail.com> wrote:
>
>
>
>
> On Thu, Nov 21, 2013 at 3:57 PM, Nick Kledzik <kledzik at apple.com> wrote:
>
>> Michael,
>>
>> In lld, we have places that used nested a ErrorOr<std::unique_ptr<xx>>
>> and I often hit compiler errors that require breaking up expressions to
>> work around.   Do you have suggestions on how to code the following simple
>> examples to not error?  Can some of these be fixed in ErrorOr.h?  Or am I
>> totally not getting something?
>>
>> -Nick
>>
>>
>> struct Foo { void doit(); };
>>
>>
>> std::unique_ptr<Foo> factoryU() {
>>   std::unique_ptr<Foo> f(new Foo);
>>   return f;  // works as expected
>> }
>>
>> ErrorOr<Foo*> factoryE() {
>>   ErrorOr<Foo*> f = new Foo;
>>   return f;  // works as expected
>> }
>>
>> ErrorOr<std::unique_ptr<Foo>> factoryEU() {
>>   std::unique_ptr<Foo> f(new Foo);
>>   return f; // ERROR: call to implicitly-deleted copy constructor of
>> 'std::__1::unique_ptr<Foo, std::__1::default_delete<Foo> >’
>>
>
> While a local variable inside a function is implicitly moved when
> returned, that only happens when the return expression is the local
> variable and the same type. In this case you have an implicit conversion
> that would work like any other conversion of an lvalue.
>
> So you have to write return std::move(f); unfortunately. (or you could be
> more explicit/verbose and say return ErrorOr<...>(std::move(f)); )
>
> Is there no way to promote the return value automatically?
>

None that I know of. There's just one special case in the language - where
you directly return a local variable and that variable is of the return
type (no implicit conversions, etc). Might be something that could be
improved (CC'd Richard Smith in case he can comment on future work in the
C++ standard space).


>  If you make the local variable be of type Error<std::unique_ptr<Foo>>,
> you run into the errors that you can’t access its methods (below).
>
>
>
>
>> }
>>
>>
>> void sinkU(std::unique_ptr<Foo> f) {
>>   f->doit();  // works as expected
>> }
>>
>> void sinkE(ErrorOr<Foo*> f) {
>>   f->doit(); // ERROR: member reference base type 'typename
>> remove_reference<Foo *>::type' (aka 'Foo *') is not a structure or union'
>>
>
> It's questionable whether this should work. ErrorOr<T> models a pointer to
> T. So if you had ErrorOr<Foo> f you'd expect to be able to do f->doit(),
> but if it's an ErrorOr<Foo*>, jsut like if it were a Foo**, you'd expect to
> have to use (*f)->doit().
>
> If ErrorOr<T> models a pointer to T, then ErrorOr<std::unique_ptr<Foo>>
> would be modeling a pointer to a std::unique_ptr<Foo>.  But unique_ptr<>
> already adds a pointer to the type.   I thought that ErrorOr<T> did not add
> a pointer, but rather implemented operator->() to access the underlying
> type.
>

I don't really understand this last comment. Perhaps it's because I did a
poor job explaining, I'm not sure.

Let's say I have some type Foo with a member function bar.

If I have a raw T* 't', then I would expect to write "t->bar()" (or
"(*t).bar()") to call bar.

If I have an ErrorOr<T> 'e', then ErrorOr provides operator* and operator->
to access the underlying T. So I would write "e->bar()" or "(*e).bar()".
ErroOr acts as though it's a pointer to T.

If I had a T** 't', then I would expect to write "(*t)->bar()" or
"(**t).bar()".

So by analogy, if I have an ErrorOr<T*> I would expect to write
"(*e)->bar()" or "(**e).bar()".

Now replace T* with unique_ptr<T> in the above examples and everything
remains the same. ErrorOr<unique_ptr<T>> is like a pointer to a pointer to
T.


>
>
> Is there some way to do partial specialization of
> ErrorOr<std::unique_ptr<T>> to make -> see through both ErrorOr and
> unique_ptr?
>

We could, but it would be confusing, because then "e->bar()" wouldn't be
the same as "(*e).bar()" (you'd still have to write "(**e).bar()" or
"(*e)->bar()").


>
> My overall point is that unique_ptr<> is cool.  ErrorOr<> is cool.  But
> when you combine the two, the cool transparency disappears ;-(
>

I disagree that "the cool transparency disappears". unique_ptr<T> models a
pointer to T (by exposing op*/op-> to access the underlying object),
ErroOr<T> models a pointer to T (by exposing op*/op-> to access the
underlying object). So ErrorOr<unique_ptr<T>> models a pointer to pointer
to T and thus requires two dereferences.

But yes, broadly it is kind of annoying - personally I have fantasy ideas
of overloading operator '.', having a unique_ptr that was never-null and
thus didn't model a pointer (and overloaded operator '.' for access to the
underlying object) and then composing this unique_ref with ErrorOr would
just add one level of pointery-ness, not two.

- Dave


>
> -Nick
>
>
>
>
>> }
>>
>> void sinkEU(ErrorOr<std::unique_ptr<Foo>> f) {
>>   f->doit(); // ERROR: no member named 'doit' in
>> 'std::__1::unique_ptr<Foo, std::__1::default_delete<Foo> >'
>>
>
> Same here.
>
>
>> }
>>
>>
>>
>> _______________________________________________
>> LLVM Developers mailing list
>> LLVMdev at cs.uiuc.edu         http://llvm.cs.uiuc.edu
>> http://lists.cs.uiuc.edu/mailman/listinfo/llvmdev
>>
>>
>
>
-------------- next part --------------
An HTML attachment was scrubbed...
URL: <http://lists.llvm.org/pipermail/llvm-dev/attachments/20131122/fb087acd/attachment.html>