When working on dom/indexedDB, a class template that is initializable only once has shown to be useful to avoid usage errors and reduce the number of state assertions. It is based on mozilla::Maybe and exposes a small subset of its interface.

I feel this might be useful for other code as well, so it might be moved to MFBT. I can do this and add appropriate tests if this makes sense.

My initial version looks like this:

// A kind of mozilla::Maybe that can only be initialized and cleared once. It
// cannot be re-initialized. This is a more stateful than a const Maybe<T> in
// that it can be cleared, but much less stateful than a non-const Maybe<T>
// which could be reinitialized multiple times. Use with a const T to ensure
// that the contents cannot be modified either.
template <typename T, bool AllowLazy = false>
class InitializedOnce {
 public:
  template <typename Dummy = void>
  InitializedOnce(std::enable_if_t<AllowLazy, Dummy>* = nullptr) {}

  template <typename... Args>
  InitializedOnce(Args&&... aArgs)
      : mMaybe{Some(T{std::forward<Args>(aArgs)...})} {}

  InitializedOnce& operator=(const InitializedOnce&) = delete;
  InitializedOnce& operator=(InitializedOnce&&) = delete;

  template <typename... Args, typename Dummy = void>
  std::enable_if_t<AllowLazy, Dummy> init(Args&&... aArgs) {
    MOZ_ASSERT(mMaybe.isNothing());
    mMaybe.emplace(T{std::forward<Args>(aArgs)...});
  }

  explicit operator bool() const { return isSome(); }
  bool isSome() const { return mMaybe.isSome(); }
  bool isNothing() const { return mMaybe.isNothing(); }

  T& operator*() const { return *mMaybe; }
  T* operator->() const { return mMaybe.operator->(); }

  void reset() {
    MOZ_ASSERT(mMaybe.isSome());
    mMaybe.reset();
  }

 private:
  Maybe<T> mMaybe;
};

Why have AllowLazy = true if that just makes the class equivalent to Maybe with a slightly different spelling?

(In reply to Nathan Froyd [:froydnj] from comment #1)

Why have AllowLazy = true if that just makes the class equivalent to Maybe with a slightly different spelling?

With AllowLazy = true, you do not need to initialize it on construction, but it may still be initialized only once (either on construction or lazily with init). Maybe would allow to initialize it multiple times by reassigning or emplace.

What I missed until now is to disable the assignment operators.

(In reply to Simon Giesecke [:sg] [he/him] from comment #2)

(In reply to Nathan Froyd [:froydnj] from comment #1)

Why have AllowLazy = true if that just makes the class equivalent to Maybe with a slightly different spelling?

With AllowLazy = true, you do not need to initialize it on construction, but it may still be initialized only once (either on construction or lazily with init). Maybe would allow to initialize it multiple times by reassigning or emplace.

What I missed until now is to disable the assignment operators.

But you have reset, so you can just reset and then init again, no?

(In reply to Nathan Froyd [:froydnj] from comment #3)

(In reply to Simon Giesecke [:sg] [he/him] from comment #2)

(In reply to Nathan Froyd [:froydnj] from comment #1)

Why have AllowLazy = true if that just makes the class equivalent to Maybe with a slightly different spelling?

With AllowLazy = true, you do not need to initialize it on construction, but it may still be initialized only once (either on construction or lazily with init). Maybe would allow to initialize it multiple times by reassigning or emplace.

What I missed until now is to disable the assignment operators.

But you have reset, so you can just reset and then init again, no?

Mh, yes, that's true. Another flag would probably be needed to prohibit such a usage sequence as well.

FYI, this is the patch which ought to add this locally to dom/indexedDB and also uses it at a few places: https://phabricator.services.mozilla.com/D53951

It has now landed and been extended a bit to allow checking for valid values at the time of initialization: https://searchfox.org/mozilla-central/source/dom/indexedDB/InitializedOnce.h

https://hg.mozilla.org/mozilla-central/rev/711622b6327c

Depends on D65568

Backed out 10 changesets (Bug 1623278, Bug 1617170, Bug 1620273, Bug 1597954) for causing bustages.

Backout link: https://hg.mozilla.org/integration/autoland/rev/0e7e76826ace75277a1f892dc14358c675d1c911

Push with failures:
https://treeherder.mozilla.org/#/jobs?repo=autoland&resultStatus=testfailed%2Cbusted%2Cexception%2Crunnable&tochange=0e7e76826ace75277a1f892dc14358c675d1c911&fromchange=50c0e58dd6544fdcf86d5fb99e0895b68446e8c3&selectedJob=293829628

Failure log: https://treeherder.mozilla.org/logviewer.html#/jobs?job_id=293829628&repo=autoland&lineNumber=45369

https://hg.mozilla.org/mozilla-central/rev/7a219404ebe4