Clang's error on non-pack parameter of alias template in a dependent type

[jonathanpoelen]

It seems that clang does not consider sizeof...(xs) to be a dependent expression when the call to foo is made with a known number of arguments. The expression must absolutely contain xs....

The following code compiles with gcc, but not clang (https://godbolt.org/z/TYzvjTePT):

template<int>
struct foo_impl {
  template<class>
  using f = int;
};

template<class... xs> constexpr int sizeof_pack = sizeof...(xs);

template<class... xs>
using foo = typename foo_impl< // error: pack expansion used as argument for non-pack parameter of alias template
  sizeof...(xs) // does not work
  // sizeof_pack<xs...>  // ok
>::template f<xs...>; // f is dependent on xs

template<class... xs>
using test = list<foo<xs>...>;

// template<class... xs>
// using test = foo<xs...>; // ok with pack...

test<int> a; // ko

[llvmbot]

@llvm/issue-subscribers-clang-frontend

Author: Jonathan Poelen (jonathanpoelen)

It seems that clang does not consider `sizeof...(xs)` to be a dependent expression when the call to foo is made with a known number of arguments. The expression must absolutely contain `xs...`.

The following code compiles with gcc, but not clang (https://godbolt.org/z/TYzvjTePT):

template<int>
struct foo_impl {
  template<class>
  using f = int;
};

template<class... xs> constexpr int sizeof_pack = sizeof...(xs);

template<class... xs>
using foo = typename foo_impl< // error: pack expansion used as argument for non-pack parameter of alias template
  sizeof...(xs) // does not work
  // sizeof_pack<xs...>  // ok
>::template f<xs...>; // f is dependent on xs

template<class... xs>
using test = list<foo<xs>...>;

// template<class... xs>
// using test = foo<xs...>; // ok with pack...

test<int> a; // ko

[shafik]

Not sure who is correct: https://godbolt.org/z/6fM5do836

clang/MSVC reject
gcc/edg accept

[shafik]

Maybe this is cwg1430

Similar gcc bug: https://gcc.gnu.org/bugzilla/show_bug.cgi?id=59701

Not totally sure.

CC @zygoloid @erichkeane

[zygoloid]

While canonicalizing the RHS of test, we form foo<xs>, for which sizeof...(xs) is 1. This gives

using foo<xs> = typename foo_impl<1>::template f<{xs}...>;

... which I suppose we reject per #1430. But I don't think we should: at this point, {xs}... should have already been expanded into the single argument xs, which is not itself a pack expansion, so #1430 shouldn't apply.

[zyn0217]

The issue arises in RebuildDependentTemplateSpecializationType, where the working case using test = foo<xs...> causes name template f to be considered dependent (and thus the diagnosis in CheckTemplateArgumentList gets bypassed) while the other using test = list<foo<xs>...> resolves the name template f in some way, resulting in the diagnosis in CheckTemplateArgumentList.