Mangling unexpanded enclosing template parameter packs

[John McCall]

Building on my last obnoxious message to this list, I don't know how to mangle references to unexpanded template parameter packs from enclosing (class) templates.

Normally, template parameters from enclosing templates are mangled as what they expanded to (although it isn't clear what to do for the base of an unresolved-name;  see my previous email).  For the most part, there's nothing stopping this from applying to template parameter packs, either.  However, if a variadic function template is contained within a variadic class template, and the function's signature contains a pack expansion which simultaneously expands parameter packs from the function and some enclosing level, then we can't actually independently expand the packs;  or rather, there's a simultaneous expansion of both levels.

This breaks down into five different cases for the mangler.

1.  Non-type template parameters outside of an unresolved-name.

template <unsigned N...> class A {
  template <class T...> void foo(T(&param)[N]...);
};

2.  Type template parameters outside of an unresolved-name.

template <class U...> class A {
  template <class T...> void foo(T(&param)(U)...);
};

3.  Template template parameters outside of an unresolved-name.

template <template <class> class U...> class A {
  template <class T...> void foo(U<T> &param...);
};

4.  Type template parameters as the base of an unresolved-name.

template <class U...> class A {
  template <class T...> void foo(decltype(U::foo(T()) param...);
};

5.  Template template parameters as the base of an unresolved-name.

template <template <class> class U...> class A {
  template <class T...> void foo(decltype(U<T>::foo) param...);
};

(1), (2), and (3) are all basically the same problem, to which I see two families of solutions.

[A].  Invent some mangling for a reference to an enclosing template parameter pack:
[A.i]. template-param meets function-param
  enclosing-template-param ::= TL <L-1 non-negative number> __
  enclosing-template-param ::= TL <L-1 non-negative number> _ <parameter-2 non-negative number> _
[A.ii]. a pack of all the actual template args
  enclosing-template-param-pack ::= ?
[B].  Expand the pack with "holes" for the unexpanded paramters.  These unexpanded holes cannot just be mangled as template-params because of a potential ambiguity with nested expansions, but a parallel mangling would suffice.

I propose that we use [A.i].  Both [B] and (to a lesser extent) [A.ii] would significantly inflate the manglings.  [A.i] is compact, and it also gives us an obvious way to mangle (4) and (5), which we could even use for the non-variadic cases covered in my last email.

We would still use fully-expanded manglings for non-unresolved-name pack expansions that do not expand packs from the function template being mangled.

John.