[clang][regression][rejects-valid] clang trunk rejecting valid pack expansion

[ericniebler]

I believe the following C++20 code should compile:

template <class...>
concept True = true;

template <bool>
struct I {
  template <template <class...> class F, class... As>
  using f = F<As...>;
};

template <template <class...> class F, class... As>
using meval = I<True<As...>>::template f<F, As...>;

template <class F, class... As>
concept P = true;

template <class F, class... As>
  requires P<F, As...>
using V = void;

template <class...Ts>
auto f() -> meval<V, int, Ts...> {
}

int main() {
  f();
}

Clang 18 and earlier accepts it, as does gcc. With clang-trunk, I get:

<source>:10:29: error: pack expansion used as argument for non-pack parameter of alias template
   10 | template <template <class...> class F, class... As>
      |                             ^
1 error generated.
Compiler returned: 1

See repro here: https://godbolt.org/z/c1dG74EPd

[llvmbot]

@llvm/issue-subscribers-clang-frontend

Author: Eric Niebler (ericniebler)

I believe the following C++20 code should compile:

template <class...>
concept True = true;

template <bool>
struct I {
  template <template <class...> class F, class... As>
  using f = F<As...>;
};

template <template <class...> class F, class... As>
using meval = I<True<As...>>::template f<F, As...>;

template <class F, class... As>
concept P = true;

template <class F, class... As>
  requires P<F, As...>
using V = void;

template <class...Ts>
auto f() -> meval<V, int, Ts...> {
}

int main() {
  f();
}

Clang 18 and earlier accepts it, as does gcc. With clang-trunk, I get:

<source>:10:29: error: pack expansion used as argument for non-pack parameter of alias template
   10 | template <template <class...> class F, class... As>
      |                             ^
1 error generated.
Compiler returned: 1

See repro here: https://godbolt.org/z/c1dG74EPd

[mizvekov]

That's CWG1430

We can ask guidance from core if it should apply in the template argument deduction that happens within TTP matching (as specified to happen in P0522R0, in terms of function template rewrite).

I see no such exclusion in the proposed solution.

[ericniebler]

I'll note that this is breaking at least two popular open source libraries: stdexec and libunifex. Possibly others.

[cor3ntin]

template <template <class...> class F, class... As>
using Alias = F<As...>;

template <template <class...> class F, class... As>
using meval = Alias<F, As...>;

template <class F, class... As>
concept P = true;

template <class F, class... As>
requires P<F, As...>
using V = void;

template <class...Ts>
auto f() -> meval<V, int, Ts...> {
}

int main() {
  f();
}

After #89807,
Clang implements https://eel.is/c++draft/temp.arg.template#4, and when deducing V (afaict)
https://www.open-std.org/jtc1/sc22/wg21/docs/cwg_active.html#1430 triggers. (note that there is no wording for this issue but all compilers do seem to implement it)

It's unclear what the conforming behavior here should be. Maybe we should ignore CWG1430 during partial ordering?
@zygoloid

I'll try to raise the issue in CWG but I think this issue give us ground to temporarily revert the default of
-fno-relaxed-template-template-args until we can either resolve CWG1430 in clang in a way that does not break code,
or we can gain enough clarity to determine stdexec is not conforming and we know how to fix it.

This is disrupting everyone using stdexec, which is at least 3 vendors.

@mizvekov are you able to work on a revert (switch back the flag to be off by default + remove the depreciation warning) ?

[mizvekov]

When this sort of significant breakage happens, it's also usual that we would 'take the matter in to our own hands' and amend the rules ourselves, as we did to improve the situation with P0522.

[shafik]

Note everyone is consistent on the examples from CWG1430: https://godbolt.org/z/xavdor73e

and currently clang is the only one that rejects this example: https://godbolt.org/z/qe1ercoqj

@cor3ntin raising some questions that looks like they need resolving, so it seems like we should revert so we can get some clarity here.

[ericniebler]

Commenting out the requires clause in the definition of the alias template V causes clang to accept the code. Which doesn't make any sense, and suggests to me that this is a bug in clang rather than in the c++ standard or stdexec.

template <class F, class... As>
//requires P<F, As...>
using V = void;

[mizvekov]

It makes sense: There is a fast path that doesn't involve any template argument deduction, and it applies when the template parameter lists are close enough, but doesn't apply when constraints are involved.

If CWG1430 required us to diagnose this, then this fast path would be wrong to accept that code, but that doesn't help your problem.

[mizvekov]

After deliberation about this issue, I think we should accept the code.

The template template parameter matching rules are a duck-typed, anything goes matter, where we will figure out if the use was valid later when we instantiate the template.
If we passed a parameter with no valid uses, then it would be good we diagnose it.
But what is happening here in this example is valid for some uses, and applying a rule that requires all uses be valid would be incongruous with the whole duck-typing thing.