[PATCH] D156244: [clang] Do not crash on use of a variadic overloaded operator

[Aaron Ballman via Phabricator via cfe-commits]

aaron.ballman added inline comments.


================
Comment at: clang/test/SemaCXX/overloaded-operator-decl.cpp:64
+class E {};
+void operator+(E, ...) {} // expected-error{{overloaded 'operator+' cannot be variadic}}
+void d() { E() + E(); }
----------------
cor3ntin wrote:
> aaron.ballman wrote:
> > Fznamznon wrote:
> > > aaron.ballman wrote:
> > > > I think it might make sense to extend the test coverage for the other operators you can overload, just to demonstrate we diagnose them all consistently. WDYT?
> > > Okay, while trying to add more test cases I discovered that following
> > > ```
> > > class E {};
> > > bool operator<(const E& lhs, ...);
> > > auto operator<=>(const E& lhs, ...);
> > > 
> > > void d() {
> > >   E() < E();
> > > }
> > > ```
> > > crashes even with the patch since there is code searching for best overload candidate that doesn't consider possibility for them making variadic.
> > > The code around overloading is actually pretty inconsistent, somewhere invalid candidates are considered, and somewhere not, so I spent some time not knowing what to do.
> > > I'm now inclined that we just shouldn't consider invalid candidates like @shafik 
> > > suggests. WDYY?
> > > 
> > Overload resolution doesn't need to produce a candidate that's viable to call; C++ lets you resolve to the "best viable function" only to say "and that one wasn't good enough either." e.g., http://eel.is/c++draft/over#match.general-3
> > 
> > I've not yet spotted anything in http://eel.is/c++draft/over that says invalid declarations should/should not be added to the initial candidate set. I *think* the intention is that if name lookup can find the name, it goes into the candidate set. Then that set is processed to remove functions that are not viable (http://eel.is/c++draft/over.match.viable). Then we find the best viable function from that set.
> > 
> > I think we should be keeping the function in the candidate set so long as it matches the rules in http://eel.is/c++draft/over.match.viable even if the function is otherwise not viable. Otherwise, correcting an unrelated issue might change overload resolution to find a completely different function. e.g., in my example above, we'd select `void overloaded(int);` as the best viable function, but when the user corrects the `float` function, we'd change to call that instead. I think it's easier to understand what's going on when picking the `float` overload to begin with and saying "but we can't call that because it's busted".
> > 
> > CC @cor3ntin @hubert.reinterpretcast @rsmith for some extra opinions, as I'm not certain if I'm interpreting the standard correctly or not.
> I'm not sure how much the standardese matter here as the declaration of the operators are ill-formed anyway.
> But from a QOI perspective, i think keeping in the set everything the users might reasonably expect to be considered makes sense to me, as it *should* lead to better diagnostics
Thanks for the extra perspective! Yeah, I think that's what I've been convincing myself of. Effectively:
```
int overloaded(int);
float overloaded(float) noexcept(error()); // error: invalid declaration

int main() {
  (void)overloaded(1.0f); // #1
}
```
We're always going to get the invalid declaration diagnostic. The question is whether we want a diagnostic at #1 that says "can't call this overload" or whether we want #1 to have no diagnostic (because it picked the `int` overload). From a purely diagnostic perspective, I think I can see arguments either way. But if we change it slightly:
```
template <typename Ty>
constexpr int func() { return 0; }

template <>
constexpr int func<int>() { return 1; }

template <>
constexpr int func<float>() { return 2; }

int overloaded(int);
float overloaded(float) noexcept(error()); // error: invalid declaration

static_assert(func(overloaded(1.0f)) == 2); // #1
```
we still get the invalid declaration diagnostic, but now we get a failing static assertion because we picked the `int` overload rather than the `float` overload. I think this is clearly worse behavior than if we issued an additional diagnostic at #1 saying that the overload we picked (`float`) was invalid.

https://godbolt.org/z/34aGMY43n

WDYT?


Repository:
  rG LLVM Github Monorepo

CHANGES SINCE LAST ACTION
  https://reviews.llvm.org/D156244/new/

https://reviews.llvm.org/D156244