[PATCH] D92480: [llvm] Add asserts in (ThreadSafe)?RefCountedBase destructors

[David Blaikie via Phabricator via llvm-commits]

dblaikie accepted this revision.
dblaikie added inline comments.
This revision is now accepted and ready to land.


================
Comment at: clang/lib/ASTMatchers/ASTMatchersInternal.cpp:155
 public:
-  TrueMatcherImpl() {
-    Retain(); // Reference count will never become zero.
-  }
+  TrueMatcherImpl() = default;
 
----------------
You can remove this ctor declaration, since it's the same as the implicit default.


================
Comment at: llvm/include/llvm/ADT/IntrusiveRefCntPtr.h:73
 
-public:
+protected:
   RefCountedBase() = default;
----------------
You could probably commit this separately - no need for extra review, but it's separate/independent of the rest of this change, I think?


================
Comment at: llvm/include/llvm/Support/ManagedStatic.h:25
+// that are const with no params.
+template <typename T> struct HasRetainRelease {
+private:
----------------
njames93 wrote:
> dblaikie wrote:
> > njames93 wrote:
> > > dblaikie wrote:
> > > > Are there many uses that rely on this? I don't think it's really worth all this infrastructure - compared to having it supported on an as-needed basis, such as directly in TrueMatcherImpl's ctor and dtor.
> > > It doesn't work in TrueMatcherImpl ctor/dtor due to the whole double free issue(See [[ https://reviews.llvm.org/D92480#inline-864517 | previous comment ]]). I can go back to just supporting it for TrueMatcherImpl by writing a single custom creator/deleter for that class.
> > > 
> > > This approach taken was a little overkill but likely a little more foolproof for someone using the library. I'll go ahead with whichever approach you would prefer. 
> > Ah, right - thanks for walking me through it again, now I better understand your previous comment - sorry for that erroneous suggestion/confusion.
> > 
> > Fair points all.
> > 
> > Given all that, I'm sort of leaning towards the idea that maybe the right solution here is for the `TrueMatcherInstance` bear the cost of the complexity here (if it's the only one) with something like:
> > 
> > ```
> > struct TrueMatcherImplCreator {
> >   static void *call() {
> >     return new IntrusiveRefCntPtr<TrueMatcherImpl>(new TrueMatcherImpl());  
> >   }
> > };
> > static llvm::ManagedStatic<IntrusiveRefCntPtr<TrueMatcherImpl>, TrueMatcherImplCreator> TrueMatcherInstance;
> > ```
> > 
> > I worry about creating a fairly generic extension point for customizing how elements in ManagedStatic can be constructed and destroyed via specialization rather than via explicit creator/destroyer parameters.
> > 
> > And while the custom destroyer is a bit simpler mechanically (doesn't involve dynamically allocating an IntrusiveCntPtr, which is unintuitive to say the least) - I think sticking to the "if you ever share ownership of a RefCountedBase object, you must've allocated it with 'new' and be really sharing ownership - no lies" is probably a healthier model for RefCountedBase/IntrusiveRefCntPtr.
> Decided to take a step back. This is trying to fix a problem that's only here because we are using a `ManagedStatic`. when its not needed. A function scope static has nearly the same semantics of `ManagedStatic` and as this is only used in one function it seems a much better fit.
Yeah, I'm good with that - I guess ManagedStatic has the extra "deallocate/destroy on llvm_shutdown" and this object doesn't seem big/significant enough that it'll matter to destroy them on shutdown. No doubt we've got lots of other function-local statics of similar weight/use that we don't worry about destroying on llvm_shutdown.


Repository:
  rG LLVM Github Monorepo

CHANGES SINCE LAST ACTION
  https://reviews.llvm.org/D92480/new/

https://reviews.llvm.org/D92480