[RFC,1/3] c++: add __is_deducible trait [PR105841]

Tested x86_64-pc-linux-gnu.  Since this is fixing experimental (C++20)
functionality, I think it's reasonable to apply now; I'm interested in other
opinions, and thoughts about the user-facing functionality.  I'm thinking to
make it internal-only for GCC 13 at least by adding a space in the name, but
does this look useful to the library?

-- 8< --

C++20 class template argument deduction for an alias template involves
adding a constraint that the template arguments for the alias template can
be deduced from the return type of the deduction guide for the underlying
class template.  In the standard, this is modeled as defining a class
template with a partial specialization, but it's much more efficient to
implement with a trait that directly tries to perform the deduction.

The first argument to the trait is a template rather than a type, so various
places needed to be adjusted to accommodate that.

	PR c++/105841

gcc/ChangeLog:

	* doc/extend.texi (Type Traits):: Document __is_deducible.

gcc/cp/ChangeLog:

	* cp-trait.def (IS_DEDUCIBLE): New.
	* cxx-pretty-print.cc (pp_cxx_trait): Handle non-type.
	* parser.cc (cp_parser_trait): Likewise.
	* pt.cc (tsubst_copy_and_build): Likewise.
	(type_targs_deducible_from): New.
	(alias_ctad_tweaks): Use it.
	* semantics.cc (trait_expr_value): Handle CPTK_IS_DEDUCIBLE.
	(finish_trait_expr): Likewise.
	* constraint.cc (diagnose_trait_expr): Likewise.
	* cp-tree.h (type_targs_deducible_from): Declare.

gcc/testsuite/ChangeLog:

	* g++.dg/ext/is_deducible1.C: New test.
---
 gcc/doc/extend.texi                      |  4 +++
 gcc/cp/cp-tree.h                         |  1 +
 gcc/cp/constraint.cc                     |  3 ++
 gcc/cp/cxx-pretty-print.cc               |  5 +++-
 gcc/cp/parser.cc                         | 20 +++++++++++---
 gcc/cp/pt.cc                             | 35 +++++++++++++++++-------
 gcc/cp/semantics.cc                      | 11 ++++++++
 gcc/testsuite/g++.dg/ext/is_deducible1.C | 27 ++++++++++++++++++
 gcc/cp/cp-trait.def                      |  1 +
 9 files changed, 92 insertions(+), 15 deletions(-)
 create mode 100644 gcc/testsuite/g++.dg/ext/is_deducible1.C


base-commit: 9944ca17c0766623bce260684edc614def7ea761

On Sat, 18 Feb 2023, Jason Merrill via Gcc-patches wrote:

> Tested x86_64-pc-linux-gnu.  Since this is fixing experimental (C++20)
> functionality, I think it's reasonable to apply now; I'm interested in other
> opinions, and thoughts about the user-facing functionality.  I'm thinking to
> make it internal-only for GCC 13 at least by adding a space in the name, but
> does this look useful to the library?

IIUC this looks like a generalization of an __is_specialization_of trait
that returns whether a type is a specialization of a given class template,
which seems potentially useful for the library to me.  We already define
some ad-hoc predicates for testing this, e.g. __is_reverse_view,
__is_span etc in <ranges> as well as a more general __is_specialization_of
in <format> for templates that take only type arguments.  Using a built-in
trait should be more efficient.

> 
> -- 8< --
> 
> C++20 class template argument deduction for an alias template involves
> adding a constraint that the template arguments for the alias template can
> be deduced from the return type of the deduction guide for the underlying
> class template.  In the standard, this is modeled as defining a class
> template with a partial specialization, but it's much more efficient to
> implement with a trait that directly tries to perform the deduction.
> 
> The first argument to the trait is a template rather than a type, so various
> places needed to be adjusted to accommodate that.
> 
> 	PR c++/105841
> 
> gcc/ChangeLog:
> 
> 	* doc/extend.texi (Type Traits):: Document __is_deducible.
> 
> gcc/cp/ChangeLog:
> 
> 	* cp-trait.def (IS_DEDUCIBLE): New.
> 	* cxx-pretty-print.cc (pp_cxx_trait): Handle non-type.
> 	* parser.cc (cp_parser_trait): Likewise.
> 	* pt.cc (tsubst_copy_and_build): Likewise.
> 	(type_targs_deducible_from): New.
> 	(alias_ctad_tweaks): Use it.
> 	* semantics.cc (trait_expr_value): Handle CPTK_IS_DEDUCIBLE.
> 	(finish_trait_expr): Likewise.
> 	* constraint.cc (diagnose_trait_expr): Likewise.
> 	* cp-tree.h (type_targs_deducible_from): Declare.
> 
> gcc/testsuite/ChangeLog:
> 
> 	* g++.dg/ext/is_deducible1.C: New test.
> ---
>  gcc/doc/extend.texi                      |  4 +++
>  gcc/cp/cp-tree.h                         |  1 +
>  gcc/cp/constraint.cc                     |  3 ++
>  gcc/cp/cxx-pretty-print.cc               |  5 +++-
>  gcc/cp/parser.cc                         | 20 +++++++++++---
>  gcc/cp/pt.cc                             | 35 +++++++++++++++++-------
>  gcc/cp/semantics.cc                      | 11 ++++++++
>  gcc/testsuite/g++.dg/ext/is_deducible1.C | 27 ++++++++++++++++++
>  gcc/cp/cp-trait.def                      |  1 +
>  9 files changed, 92 insertions(+), 15 deletions(-)
>  create mode 100644 gcc/testsuite/g++.dg/ext/is_deducible1.C
> 
> diff --git a/gcc/doc/extend.texi b/gcc/doc/extend.texi
> index 1ae68b0f20a..898701424ad 100644
> --- a/gcc/doc/extend.texi
> +++ b/gcc/doc/extend.texi
> @@ -25207,6 +25207,10 @@ type.  A diagnostic is produced if this requirement is not met.
>  If @code{type} is a cv-qualified class type, and not a union type
>  ([basic.compound]) the trait is @code{true}, else it is @code{false}.
>  
> +@item __is_deducible (template, type)
> +If template arguments for @code{template} can be deduced from
> +@code{type} or obtained from default template arguments.
> +
>  @item __is_empty (type)
>  If @code{__is_class (type)} is @code{false} then the trait is @code{false}.
>  Otherwise @code{type} is considered empty if and only if: @code{type}
> diff --git a/gcc/cp/cp-tree.h b/gcc/cp/cp-tree.h
> index 5595335bbf7..e79150ca4d8 100644
> --- a/gcc/cp/cp-tree.h
> +++ b/gcc/cp/cp-tree.h
> @@ -7372,6 +7372,7 @@ extern tree fn_type_unification			(tree, tree, tree,
>  						 bool, bool);
>  extern void mark_decl_instantiated		(tree, int);
>  extern int more_specialized_fn			(tree, tree, int);
> +extern bool type_targs_deducible_from		(tree, tree);
>  extern void do_decl_instantiation		(tree, tree);
>  extern void do_type_instantiation		(tree, tree, tsubst_flags_t);
>  extern bool always_instantiate_p		(tree);
> diff --git a/gcc/cp/constraint.cc b/gcc/cp/constraint.cc
> index 9374327008b..a28c85178fe 100644
> --- a/gcc/cp/constraint.cc
> +++ b/gcc/cp/constraint.cc
> @@ -3797,6 +3797,9 @@ diagnose_trait_expr (tree expr, tree args)
>        inform (loc, "  %qT is not a reference that binds to a temporary "
>  	      "object of type %qT (copy-initialization)", t1, t2);
>        break;
> +    case CPTK_IS_DEDUCIBLE:
> +      inform (loc, "  %qD is not deducible from %qT", t1, t2);
> +      break;
>  #define DEFTRAIT_TYPE(CODE, NAME, ARITY) \
>      case CPTK_##CODE:
>  #include "cp-trait.def"
> diff --git a/gcc/cp/cxx-pretty-print.cc b/gcc/cp/cxx-pretty-print.cc
> index bea52a608f1..4ebd957decd 100644
> --- a/gcc/cp/cxx-pretty-print.cc
> +++ b/gcc/cp/cxx-pretty-print.cc
> @@ -2626,7 +2626,10 @@ pp_cxx_trait (cxx_pretty_printer *pp, tree t)
>      }
>  
>    pp_cxx_left_paren (pp);
> -  pp->type_id (type1);
> +  if (DECL_P (type1))
> +    pp->expression (type1);
> +  else
> +    pp->type_id (type1);

Since the first argument of a TRAIT_EXPR can now be a TEMPLATE_DECL, I
suppose cp_tree_equal needs to be changed too.

>    if (type2)
>      {
>        if (TREE_CODE (type2) != TREE_LIST)
> diff --git a/gcc/cp/parser.cc b/gcc/cp/parser.cc
> index 1a124f5395e..68950cace78 100644
> --- a/gcc/cp/parser.cc
> +++ b/gcc/cp/parser.cc
> @@ -10960,10 +10960,22 @@ cp_parser_trait (cp_parser* parser, enum rid keyword)
>    matching_parens parens;
>    parens.require_open (parser);
>  
> -  {
> -    type_id_in_expr_sentinel s (parser);
> -    type1 = cp_parser_type_id (parser);
> -  }
> +  if (kind == CPTK_IS_DEDUCIBLE)
> +    {
> +      const cp_token* token = cp_lexer_peek_token (parser->lexer);
> +      type1 = cp_parser_id_expression (parser,
> +				       /*template_keyword_p=*/false,
> +				       /*check_dependency_p=*/true,
> +				       nullptr,
> +				       /*declarator_p=*/false,
> +				       /*optional_p=*/false);
> +      type1 = cp_parser_lookup_name_simple (parser, type1, token->location);
> +    }
> +  else
> +    {
> +      type_id_in_expr_sentinel s (parser);
> +      type1 = cp_parser_type_id (parser);
> +    }
>  
>    if (type1 == error_mark_node)
>      return error_mark_node;
> diff --git a/gcc/cp/pt.cc b/gcc/cp/pt.cc
> index b1ac7d4beb4..2aa06557b99 100644
> --- a/gcc/cp/pt.cc
> +++ b/gcc/cp/pt.cc
> @@ -21577,8 +21577,9 @@ tsubst_copy_and_build (tree t,
>  
>      case TRAIT_EXPR:
>        {
> -	tree type1 = tsubst (TRAIT_EXPR_TYPE1 (t), args,
> -			     complain, in_decl);
> +	tree type1 = TRAIT_EXPR_TYPE1 (t);
> +	if (TREE_CODE (type1) != TEMPLATE_DECL)
> +	  type1 = tsubst (type1, args, complain, in_decl);
>  	tree type2 = tsubst (TRAIT_EXPR_TYPE2 (t), args,
>  			     complain, in_decl);
>  	RETURN (finish_trait_expr (TRAIT_EXPR_LOCATION (t),
> @@ -29979,7 +29980,7 @@ alias_ctad_tweaks (tree tmpl, tree uguides)
>    /* This implementation differs from the above in two significant ways:
>  
>       1) We include all template parameters of A, not just some.
> -     2) The added constraint is same_type instead of deducible.
> +     2) [fixed] The added constraint is same_type instead of deducible.
>  
>       I believe that while it's probably possible to construct a testcase that
>       behaves differently with this simplification, it should have the same
> @@ -30079,7 +30080,7 @@ alias_ctad_tweaks (tree tmpl, tree uguides)
>  	      /* FIXME this should mean they don't compare as equivalent.  */
>  	      || dependent_alias_template_spec_p (atype, nt_opaque))
>  	    {
> -	      tree same = finish_trait_expr (loc, CPTK_IS_SAME, atype, ret);
> +	      tree same = finish_trait_expr (loc, CPTK_IS_DEDUCIBLE, tmpl, ret);
>  	      ci = append_constraint (ci, same);
>  	    }
>  
> @@ -30093,12 +30094,7 @@ alias_ctad_tweaks (tree tmpl, tree uguides)
>  	{
>  	  /* For a non-template deduction guide, if the arguments of A aren't
>  	     deducible from the return type, don't add the candidate.  */
> -	  tree targs = make_tree_vec (natparms);
> -	  int err = unify (atparms, targs, utype, ret, UNIFY_ALLOW_NONE, false);
> -	  for (unsigned i = 0; !err && i < natparms; ++i)
> -	    if (TREE_VEC_ELT (targs, i) == NULL_TREE)
> -	      err = true;
> -	  if (err)
> +	  if (!type_targs_deducible_from (tmpl, ret))
>  	    continue;
>  	}
>  
> @@ -30108,6 +30104,25 @@ alias_ctad_tweaks (tree tmpl, tree uguides)
>    return aguides;
>  }
>  
> +/* True iff template arguments for TMPL can be deduced from TYPE.
> +   Used to implement CPTK_IS_DEDUCIBLE for alias CTAD.  */
> +
> +bool
> +type_targs_deducible_from (tree tmpl, tree type)
> +{
> +  tree tparms = DECL_INNERMOST_TEMPLATE_PARMS (tmpl);
> +  int len = TREE_VEC_LENGTH (tparms);
> +  tree targs = make_tree_vec (len);
> +  if (unify (tparms, targs, TREE_TYPE (tmpl), type,
> +	     UNIFY_ALLOW_NONE, false))
> +    return false;
> +  /* Maybe add in default template args.  */
> +  targs = coerce_template_parms (tparms, targs, tmpl, tf_none);
> +  if (targs == error_mark_node)
> +    return false;
> +  return constraints_satisfied_p (tmpl, targs);
> +}

For sake of the __is_specialization_of use case, I wonder if it'd
be possible to have a "fast path" that avoids deduction/coercion when
the given template is a class template?

> +
>  /* Return artificial deduction guides built from the constructors of class
>     template TMPL.  */
>  
> diff --git a/gcc/cp/semantics.cc b/gcc/cp/semantics.cc
> index 79b7cc72f21..9103f5de2f4 100644
> --- a/gcc/cp/semantics.cc
> +++ b/gcc/cp/semantics.cc
> @@ -12048,6 +12048,9 @@ trait_expr_value (cp_trait_kind kind, tree type1, tree type2)
>      case CPTK_REF_CONVERTS_FROM_TEMPORARY:
>        return ref_xes_from_temporary (type1, type2, /*direct_init=*/false);
>  
> +    case CPTK_IS_DEDUCIBLE:
> +      return type_targs_deducible_from (type1, type2);
> +
>  #define DEFTRAIT_TYPE(CODE, NAME, ARITY) \
>      case CPTK_##CODE:
>  #include "cp-trait.def"
> @@ -12205,6 +12208,14 @@ finish_trait_expr (location_t loc, cp_trait_kind kind, tree type1, tree type2)
>  	return error_mark_node;
>        break;
>  
> +    case CPTK_IS_DEDUCIBLE:
> +      if (!DECL_TYPE_TEMPLATE_P (type1))
> +	{
> +	  error ("%qD is not a class or alias template", type1);
> +	  return error_mark_node;
> +	}
> +      break;
> +
>  #define DEFTRAIT_TYPE(CODE, NAME, ARITY) \
>      case CPTK_##CODE:
>  #include "cp-trait.def"
> diff --git a/gcc/testsuite/g++.dg/ext/is_deducible1.C b/gcc/testsuite/g++.dg/ext/is_deducible1.C
> new file mode 100644
> index 00000000000..857f59db4c8
> --- /dev/null
> +++ b/gcc/testsuite/g++.dg/ext/is_deducible1.C
> @@ -0,0 +1,27 @@
> +// { dg-do compile { target c++20 } }
> +
> +template <class T> struct A { };
> +template <class T> struct B { };
> +
> +// Simple forms.
> +static_assert (__is_deducible (::A, A<int>));
> +static_assert (__is_deducible (B, B<int>));
> +static_assert (!__is_deducible (A, B<int>));
> +static_assert (!__is_deducible (::B, A<int>));
> +
> +// This is the interesting use case for alias CTAD.
> +template <class T> using AP = A<T*>;
> +static_assert (__is_deducible (AP, A<int*>));
> +static_assert (!__is_deducible (AP, A<int>));
> +
> +// Can't deduce a parameter not used on the RHS.
> +template <class T> using C = void;
> +static_assert (!__is_deducible (C, C<int>));
> +
> +// But a default template argument counts.
> +template <class T = void> using D = void;
> +static_assert (__is_deducible (D, D<int>));
> +
> +// We don't try to support this.
> +template <class T> void f(T);
> +bool b = __is_deducible (f, void (int)); // { dg-error "class or alias" }
> diff --git a/gcc/cp/cp-trait.def b/gcc/cp/cp-trait.def
> index 823899a26c5..e43fb464f42 100644
> --- a/gcc/cp/cp-trait.def
> +++ b/gcc/cp/cp-trait.def
> @@ -84,6 +84,7 @@ DEFTRAIT_EXPR (IS_TRIVIALLY_COPYABLE, "__is_trivially_copyable", 1)
>  DEFTRAIT_EXPR (IS_UNION, "__is_union", 1)
>  DEFTRAIT_EXPR (REF_CONSTRUCTS_FROM_TEMPORARY, "__reference_constructs_from_temporary", 2)
>  DEFTRAIT_EXPR (REF_CONVERTS_FROM_TEMPORARY, "__reference_converts_from_temporary", 2)
> +DEFTRAIT_EXPR (IS_DEDUCIBLE, "__is_deducible", 2)
>  
>  DEFTRAIT_TYPE (REMOVE_CV, "__remove_cv", 1)
>  DEFTRAIT_TYPE (REMOVE_REFERENCE, "__remove_reference", 1)
> 
> base-commit: 9944ca17c0766623bce260684edc614def7ea761
> -- 
> 2.31.1
> 
>

On 2/20/23 11:58, Patrick Palka wrote:
> On Sat, 18 Feb 2023, Jason Merrill via Gcc-patches wrote:
> 
>> Tested x86_64-pc-linux-gnu.  Since this is fixing experimental (C++20)
>> functionality, I think it's reasonable to apply now; I'm interested in other
>> opinions, and thoughts about the user-facing functionality.  I'm thinking to
>> make it internal-only for GCC 13 at least by adding a space in the name, but
>> does this look useful to the library?
> 
> IIUC this looks like a generalization of an __is_specialization_of trait
> that returns whether a type is a specialization of a given class template,
> which seems potentially useful for the library to me.  We already define
> some ad-hoc predicates for testing this, e.g. __is_reverse_view,
> __is_span etc in <ranges> as well as a more general __is_specialization_of
> in <format> for templates that take only type arguments.  Using a built-in
> trait should be more efficient.
> 
>[...]
> 
> Since the first argument of a TRAIT_EXPR can now be a TEMPLATE_DECL, I
> suppose cp_tree_equal needs to be changed too.
>
 >[...]
> 
> For sake of the __is_specialization_of use case, I wonder if it'd
> be possible to have a "fast path" that avoids deduction/coercion when
> the given template is a class template?

Thanks, done.  I've also fixed array bounds type deduction and added 
more comments about the relationship of the implementation and the 
specification in terms of partial specialization.

The second patch makes it internal-only for GCC 13; you can revert that 
if you want to experiment with using it in the library.

Tested x86_64-pc-linux-gnu, applying to trunk.