[LLVMdev] New type of smart pointer for LLVM

[David Blaikie]

On Wed, Sep 24, 2014 at 7:50 PM, Justin Bogner <mail at justinbogner.com>
wrote:

> Anton Yartsev <anton.yartsev at gmail.com> writes:
> > Hello everyone,
> >
> > I bring to discussion the necessity/design of a new type of smart
> pointer.
> > r215176 and r217791 rise the problem, D5443 is devoted to the solution.
> > r215176 applies several temporary ugly fixes of memory leaks in
> TGParser.cpp
> > which would be great to be refactored using smart pointers. D5443
> demonstrates
> > how the solution with a certain type of smart pointer would look like
> (see
> > changes in TGParser::ParseDef(), TGParser::InstantiateMulticlassDef() and
> > TGParser::ParseSimpleValue()).
> >
> > Briefly:
> > consider a leaky example:
> > {
> >   T* p = new T;
> >   if (condition1) {
> >     f(p); // takes ownership of p
> >   }
> >   p->SomeMethod();
> >
> >   if (condition2) {
> >     return nullptr; // Leak!
> >   }
> >
> >   g(p); // don't take ownership of p
> >   return p;
> > }
> >
> > The preferred solution would look like:
> > {
> >   smart_ptr<T> p(new T);
> >   if (condition1) {
> >     f(p.StopOwn()); // takes ownership of p
>
> So this takes ownership, but promises not to destroy the pointee in some
> way?
>
> >   }
> >   p->SomeMethod();
> >
> >   if (condition2) {
> >     return nullptr; //
>
> I guess p is sometimes destroyed here, depending on condition1?
>
> >   }
> >
> >   g(p.Get());  // don't take ownership of p
> >   return p.StopOwn();
> > }
>
> What does it mean to stop owning the pointer twice? Doesn't this leak p
> in the case where condition1 was false?
>
> > Neither unique_ptr nor shared_ptr can be used in the place of smart_ptr
> as
> > unique_ptr sets the raw pointer to nullptr after release() (StopOwn() in
> the
> > example above) whereas shared_ptr is unable to release.
>
> I don't understand why shared_ptr wouldn't suffice for the example
> above. There are two cases in your example where you try to release the
> pointer - in one of them it seems you don't actually want to release it,
> because you continue using it after, and in the other the scope ends
> immediately after the release. The shared_ptr releases when it goes out
> of scope, so it seems to be exactly what you want here.
>
> Maybe I'm just misunderstanding the use case here, but I fear that a
> type like this one would just serve to paper over problems where the
> ownership is very unclear, without really helping much.
>

Perhaps this is a case of poor examples, I'm not sure. I certainly share
your concern that this might not be the right path forward, but it's
something that I've seen come up repeatedly when doing unique_ptr
migrations, perhaps moreso in Clang's frontend code.

Essentially there are a variety of data structures that "sometimes" own
their underlying data. Often implemented as a raw pointer and a boolean.
(sometimes I've migrated this to a raw pointer and a unique_ptr - the raw
pointer always points to the thing of interest, the unique_ptr is sometimes
non-null and points to the same thing to keep it alive (or destroy it when
it's no longer needed, more specifically))

I can go & dredge up some examples if we want to discuss the particular
merits & whether each of those cases would be better solved in some other
way, but it seemed pervasive enough in the current codebase that some
incremental improvement could be gained by replacing these cases with a
more specific tool for the job. We might still consider this tool to be
"not ideal".


>
> > Attached is a scratch that illustrates how the minimal
> > API/implementation of a desired smart pointer sufficient for
> > refactoring would look like. Any ideas and suggestions are
> > appreciated.
> >
> > //===-- CondOwnershipPtr.h - Smart ptr with conditional release -*- C++
> -*-===//
> > //
> > //                     The LLVM Compiler Infrastructure
> > //
> > // This file is distributed under the University of Illinois Open Source
> > // License. See LICENSE.TXT for details.
> >
> //===----------------------------------------------------------------------===//
> >
> > #ifndef LLVM_ADT_CONDOWNERSHIPPTR_H
> > #define LLVM_ADT_CONDOWNERSHIPPTR_H
> >
> > namespace llvm {
> >
> > template<class T>
> > class CondOwnershipPtr {
> >   T* Ptr;
> >   bool Own;
> >
> >   void Delete() {
> >     if (Ptr && !Own)
> >       delete Ptr;
> >   }
>
> This seems to delete the pointer iff we *don't* own it. I think you have
> that backwards...
>
> > public:
> >   CondOwnershipPtr() : Ptr(nullptr), Own(true) {}
> >   explicit CondOwnershipPtr(T* p) : Ptr(p), Own(true) {}
> >
> >   ~CondOwnershipPtr() {
> >     Delete();
> >   }
> >
> >   T* Get() const {
> >     return Ptr;
> >   }
> >
> >   T* StopOwn() const {
> >     Own = false;
> >     return Ptr;
> >   }
> >
> >   void Reset(T* P = nullptr) {
> >     if (P != Ptr) {
> >       Delete();
> >       Ptr = P;
> >       Own = true;
> >     }
> >   }
> >
> >   bool Owns() const {
> >     return Own;
> >   }
> >
> >   operator bool() const {
> >     return Ptr != nullptr;
> >   }
> >
> >   T& operator*() const {
> >     return *Ptr;
> >   }
> >
> >   T* operator->() const {
> >     return Ptr;
> >   }
> > };
> >
> > } // end namespace llvm
> >
> > #endif
> _______________________________________________
> 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/20140924/4cbbce90/attachment.html>