diff mbox series

C++ PATCH for c++/88184, ICE when treating name as template-name

Message ID 20181128154849.GD3269@redhat.com
State New
Headers show
Series C++ PATCH for c++/88184, ICE when treating name as template-name | expand

Commit Message

Marek Polacek Nov. 28, 2018, 3:48 p.m. UTC
Since P0846R0 was implemented, a name will be treated as a template-name when
it is an unqualified-id followed by a < and name lookup finds either one or
more functions or finds nothing, in order to potentially cause ADL to be performed.

In this case, we had

  f<T> ();

where we'd found a decl for f (not a TEMPLATE_DECL) and < follows.  But there
are no arguments, so ADL can't be performed, so I think let's diagnose this
problem sooner lest things break later in finish_call_expr.

Bootstrapped/regtested on x86_64-linux, ok for trunk?

2018-11-28  Marek Polacek  <polacek@redhat.com>

	PR c++/88184, ICE when treating name as template-name.
	* parser.c (cp_parser_postfix_expression): Give error when the
	name-as-template-name doesn't have any arguments.

	* g++.dg/cpp2a/fn-template17.C: New test.
	* g++.dg/cpp2a/fn-template18.C: New test.

Comments

Jason Merrill Nov. 29, 2018, 3:34 a.m. UTC | #1
On 11/28/18 10:48 AM, Marek Polacek wrote:
> Since P0846R0 was implemented, a name will be treated as a template-name when
> it is an unqualified-id followed by a < and name lookup finds either one or
> more functions or finds nothing, in order to potentially cause ADL to be performed.
> 
> In this case, we had
> 
>    f<T> ();
> 
> where we'd found a decl for f (not a TEMPLATE_DECL) and < follows.

 From the backtrace in the PR, it seems as though we're treating f<T> as 
non-dependent, which is wrong.  type_dependent_expression_p only looks 
at the arguments of a TEMPLATE_ID_EXPR if it has unknown_type_node, so 
we probably want to give it that type.

Jason
Marek Polacek Nov. 29, 2018, 6:28 p.m. UTC | #2
On Wed, Nov 28, 2018 at 10:34:11PM -0500, Jason Merrill wrote:
> On 11/28/18 10:48 AM, Marek Polacek wrote:
> > Since P0846R0 was implemented, a name will be treated as a template-name when
> > it is an unqualified-id followed by a < and name lookup finds either one or
> > more functions or finds nothing, in order to potentially cause ADL to be performed.
> > 
> > In this case, we had
> > 
> >    f<T> ();
> > 
> > where we'd found a decl for f (not a TEMPLATE_DECL) and < follows.
> 
> From the backtrace in the PR, it seems as though we're treating f<T> as
> non-dependent, which is wrong.  type_dependent_expression_p only looks at
> the arguments of a TEMPLATE_ID_EXPR if it has unknown_type_node, so we
> probably want to give it that type.

That was my first attempt but it was crashing everywhere, so I abandoned it.
But I was setting unknown_type_node in cp_parser_template_name whereas this
patch sets it in cp_parser_postfix_expression, which works and is arguably
better because it reuses the diagnostic in finish_call_expr.

Bootstrapped/regtested on x86_64-linux, ok for trunk?

2018-11-29  Marek Polacek  <polacek@redhat.com>

	PR c++/88184, ICE when treating name as template-name.
	* parser.c (cp_parser_postfix_expression): Set the function decl type
	to unknown type when the name-as-template-name doesn't have any
	arguments.

	* g++.dg/cpp2a/fn-template17.C: New test.
	* g++.dg/cpp2a/fn-template18.C: New test.

diff --git gcc/cp/parser.c gcc/cp/parser.c
index 3ef1eda4518..fc328c28f69 100644
--- gcc/cp/parser.c
+++ gcc/cp/parser.c
@@ -7241,6 +7241,20 @@ cp_parser_postfix_expression (cp_parser *parser, bool address_p, bool cast_p,
 						     complain);
 		      }
 		  }
+		else if (TREE_CODE (postfix_expression) == TEMPLATE_ID_EXPR
+			 && (TREE_CODE (TREE_OPERAND (postfix_expression, 0))
+			     == FUNCTION_DECL)
+			 && args->is_empty ())
+		  {
+		    /* We can get here when treating a name as a template name
+		       (because it was an unqualified-id followed by a < and
+		       name lookup found a decl), but now we see that there
+		       are no arguments, so ADL isn't possible.  Tweak the type
+		       so that the TEMPLATE_ID_EXPR is type-dependent and we get
+		       proper diagnostic down in finish_call_expr.  */
+		    TREE_TYPE (TREE_OPERAND (postfix_expression, 0))
+		      = unknown_type_node;
+		  }
 	      }
 
 	    if (TREE_CODE (postfix_expression) == COMPONENT_REF)
diff --git gcc/testsuite/g++.dg/cpp2a/fn-template17.C gcc/testsuite/g++.dg/cpp2a/fn-template17.C
new file mode 100644
index 00000000000..f0d24107682
--- /dev/null
+++ gcc/testsuite/g++.dg/cpp2a/fn-template17.C
@@ -0,0 +1,21 @@
+// PR c++/88184
+// { dg-do compile }
+// { dg-options "-std=c++2a -fchecking=2" }
+
+namespace A
+{
+  void f ();
+}
+
+using A::f;
+
+template <typename T> void g ()
+{
+  f<T> (); // { dg-error "no matching function for call" }
+}
+
+void
+fn ()
+{
+  g<int>();
+}
diff --git gcc/testsuite/g++.dg/cpp2a/fn-template18.C gcc/testsuite/g++.dg/cpp2a/fn-template18.C
new file mode 100644
index 00000000000..7fe5e89ace3
--- /dev/null
+++ gcc/testsuite/g++.dg/cpp2a/fn-template18.C
@@ -0,0 +1,23 @@
+// PR c++/88184
+// { dg-do compile }
+// { dg-options "-std=c++2a -fchecking=2" }
+
+namespace A
+{
+  void f ();
+  void f (int);
+  void f (int, int);
+}
+
+using A::f;
+
+template <typename T> void g ()
+{
+  f<T> (); // { dg-error "no matching function for call" }
+}
+
+void
+fn ()
+{
+  g<int>();
+}
Jason Merrill Nov. 29, 2018, 9:03 p.m. UTC | #3
On 11/29/18 1:28 PM, Marek Polacek wrote:
> On Wed, Nov 28, 2018 at 10:34:11PM -0500, Jason Merrill wrote:
>> On 11/28/18 10:48 AM, Marek Polacek wrote:
>>> Since P0846R0 was implemented, a name will be treated as a template-name when
>>> it is an unqualified-id followed by a < and name lookup finds either one or
>>> more functions or finds nothing, in order to potentially cause ADL to be performed.
>>>
>>> In this case, we had
>>>
>>>     f<T> ();
>>>
>>> where we'd found a decl for f (not a TEMPLATE_DECL) and < follows.
>>
>>  From the backtrace in the PR, it seems as though we're treating f<T> as
>> non-dependent, which is wrong.  type_dependent_expression_p only looks at
>> the arguments of a TEMPLATE_ID_EXPR if it has unknown_type_node, so we
>> probably want to give it that type.
> 
> That was my first attempt but it was crashing everywhere, so I abandoned it.
> But I was setting unknown_type_node in cp_parser_template_name whereas this
> patch sets it in cp_parser_postfix_expression, which works and is arguably
> better because it reuses the diagnostic in finish_call_expr.

Would it work for lookup_template_function to always use unknown_type_node?

Jason
Marek Polacek Dec. 4, 2018, 5:27 p.m. UTC | #4
On Thu, Nov 29, 2018 at 04:03:50PM -0500, Jason Merrill wrote:
> On 11/29/18 1:28 PM, Marek Polacek wrote:
> > On Wed, Nov 28, 2018 at 10:34:11PM -0500, Jason Merrill wrote:
> > > On 11/28/18 10:48 AM, Marek Polacek wrote:
> > > > Since P0846R0 was implemented, a name will be treated as a template-name when
> > > > it is an unqualified-id followed by a < and name lookup finds either one or
> > > > more functions or finds nothing, in order to potentially cause ADL to be performed.
> > > > 
> > > > In this case, we had
> > > > 
> > > >     f<T> ();
> > > > 
> > > > where we'd found a decl for f (not a TEMPLATE_DECL) and < follows.
> > > 
> > >  From the backtrace in the PR, it seems as though we're treating f<T> as
> > > non-dependent, which is wrong.  type_dependent_expression_p only looks at
> > > the arguments of a TEMPLATE_ID_EXPR if it has unknown_type_node, so we
> > > probably want to give it that type.
> > 
> > That was my first attempt but it was crashing everywhere, so I abandoned it.
> > But I was setting unknown_type_node in cp_parser_template_name whereas this
> > patch sets it in cp_parser_postfix_expression, which works and is arguably
> > better because it reuses the diagnostic in finish_call_expr.
> 
> Would it work for lookup_template_function to always use unknown_type_node?

Luckily, that seems to work!  I wish I'd thought of that, I like fixing bugs by
simplifying code.

Bootstrapped/regtested on x86_64-linux, ok for trunk?

2018-12-04  Marek Polacek  <polacek@redhat.com>

	PR c++/88184 - ICE when treating name as template-name.
	* pt.c (lookup_template_function): Always build the TEMPLATE_ID_EXPR
	with unknown_type_node.

	* g++.dg/cpp2a/fn-template17.C: New test.
	* g++.dg/cpp2a/fn-template18.C: New test.

diff --git gcc/cp/pt.c gcc/cp/pt.c
index a0d899f594b..78addc6c1c2 100644
--- gcc/cp/pt.c
+++ gcc/cp/pt.c
@@ -9068,8 +9068,6 @@ add_pending_template (tree d)
 tree
 lookup_template_function (tree fns, tree arglist)
 {
-  tree type;
-
   if (fns == error_mark_node || arglist == error_mark_node)
     return error_mark_node;
 
@@ -9090,11 +9088,7 @@ lookup_template_function (tree fns, tree arglist)
       return fns;
     }
 
-  type = TREE_TYPE (fns);
-  if (TREE_CODE (fns) == OVERLOAD || !type)
-    type = unknown_type_node;
-
-  return build2 (TEMPLATE_ID_EXPR, type, fns, arglist);
+  return build2 (TEMPLATE_ID_EXPR, unknown_type_node, fns, arglist);
 }
 
 /* Within the scope of a template class S<T>, the name S gets bound
diff --git gcc/testsuite/g++.dg/cpp2a/fn-template17.C gcc/testsuite/g++.dg/cpp2a/fn-template17.C
new file mode 100644
index 00000000000..f0d24107682
--- /dev/null
+++ gcc/testsuite/g++.dg/cpp2a/fn-template17.C
@@ -0,0 +1,21 @@
+// PR c++/88184
+// { dg-do compile }
+// { dg-options "-std=c++2a -fchecking=2" }
+
+namespace A
+{
+  void f ();
+}
+
+using A::f;
+
+template <typename T> void g ()
+{
+  f<T> (); // { dg-error "no matching function for call" }
+}
+
+void
+fn ()
+{
+  g<int>();
+}
diff --git gcc/testsuite/g++.dg/cpp2a/fn-template18.C gcc/testsuite/g++.dg/cpp2a/fn-template18.C
new file mode 100644
index 00000000000..7fe5e89ace3
--- /dev/null
+++ gcc/testsuite/g++.dg/cpp2a/fn-template18.C
@@ -0,0 +1,23 @@
+// PR c++/88184
+// { dg-do compile }
+// { dg-options "-std=c++2a -fchecking=2" }
+
+namespace A
+{
+  void f ();
+  void f (int);
+  void f (int, int);
+}
+
+using A::f;
+
+template <typename T> void g ()
+{
+  f<T> (); // { dg-error "no matching function for call" }
+}
+
+void
+fn ()
+{
+  g<int>();
+}
Jason Merrill Dec. 4, 2018, 7:09 p.m. UTC | #5
OK.
On Tue, Dec 4, 2018 at 12:27 PM Marek Polacek <polacek@redhat.com> wrote:
>
> On Thu, Nov 29, 2018 at 04:03:50PM -0500, Jason Merrill wrote:
> > On 11/29/18 1:28 PM, Marek Polacek wrote:
> > > On Wed, Nov 28, 2018 at 10:34:11PM -0500, Jason Merrill wrote:
> > > > On 11/28/18 10:48 AM, Marek Polacek wrote:
> > > > > Since P0846R0 was implemented, a name will be treated as a template-name when
> > > > > it is an unqualified-id followed by a < and name lookup finds either one or
> > > > > more functions or finds nothing, in order to potentially cause ADL to be performed.
> > > > >
> > > > > In this case, we had
> > > > >
> > > > >     f<T> ();
> > > > >
> > > > > where we'd found a decl for f (not a TEMPLATE_DECL) and < follows.
> > > >
> > > >  From the backtrace in the PR, it seems as though we're treating f<T> as
> > > > non-dependent, which is wrong.  type_dependent_expression_p only looks at
> > > > the arguments of a TEMPLATE_ID_EXPR if it has unknown_type_node, so we
> > > > probably want to give it that type.
> > >
> > > That was my first attempt but it was crashing everywhere, so I abandoned it.
> > > But I was setting unknown_type_node in cp_parser_template_name whereas this
> > > patch sets it in cp_parser_postfix_expression, which works and is arguably
> > > better because it reuses the diagnostic in finish_call_expr.
> >
> > Would it work for lookup_template_function to always use unknown_type_node?
>
> Luckily, that seems to work!  I wish I'd thought of that, I like fixing bugs by
> simplifying code.
>
> Bootstrapped/regtested on x86_64-linux, ok for trunk?
>
> 2018-12-04  Marek Polacek  <polacek@redhat.com>
>
>         PR c++/88184 - ICE when treating name as template-name.
>         * pt.c (lookup_template_function): Always build the TEMPLATE_ID_EXPR
>         with unknown_type_node.
>
>         * g++.dg/cpp2a/fn-template17.C: New test.
>         * g++.dg/cpp2a/fn-template18.C: New test.
>
> diff --git gcc/cp/pt.c gcc/cp/pt.c
> index a0d899f594b..78addc6c1c2 100644
> --- gcc/cp/pt.c
> +++ gcc/cp/pt.c
> @@ -9068,8 +9068,6 @@ add_pending_template (tree d)
>  tree
>  lookup_template_function (tree fns, tree arglist)
>  {
> -  tree type;
> -
>    if (fns == error_mark_node || arglist == error_mark_node)
>      return error_mark_node;
>
> @@ -9090,11 +9088,7 @@ lookup_template_function (tree fns, tree arglist)
>        return fns;
>      }
>
> -  type = TREE_TYPE (fns);
> -  if (TREE_CODE (fns) == OVERLOAD || !type)
> -    type = unknown_type_node;
> -
> -  return build2 (TEMPLATE_ID_EXPR, type, fns, arglist);
> +  return build2 (TEMPLATE_ID_EXPR, unknown_type_node, fns, arglist);
>  }
>
>  /* Within the scope of a template class S<T>, the name S gets bound
> diff --git gcc/testsuite/g++.dg/cpp2a/fn-template17.C gcc/testsuite/g++.dg/cpp2a/fn-template17.C
> new file mode 100644
> index 00000000000..f0d24107682
> --- /dev/null
> +++ gcc/testsuite/g++.dg/cpp2a/fn-template17.C
> @@ -0,0 +1,21 @@
> +// PR c++/88184
> +// { dg-do compile }
> +// { dg-options "-std=c++2a -fchecking=2" }
> +
> +namespace A
> +{
> +  void f ();
> +}
> +
> +using A::f;
> +
> +template <typename T> void g ()
> +{
> +  f<T> (); // { dg-error "no matching function for call" }
> +}
> +
> +void
> +fn ()
> +{
> +  g<int>();
> +}
> diff --git gcc/testsuite/g++.dg/cpp2a/fn-template18.C gcc/testsuite/g++.dg/cpp2a/fn-template18.C
> new file mode 100644
> index 00000000000..7fe5e89ace3
> --- /dev/null
> +++ gcc/testsuite/g++.dg/cpp2a/fn-template18.C
> @@ -0,0 +1,23 @@
> +// PR c++/88184
> +// { dg-do compile }
> +// { dg-options "-std=c++2a -fchecking=2" }
> +
> +namespace A
> +{
> +  void f ();
> +  void f (int);
> +  void f (int, int);
> +}
> +
> +using A::f;
> +
> +template <typename T> void g ()
> +{
> +  f<T> (); // { dg-error "no matching function for call" }
> +}
> +
> +void
> +fn ()
> +{
> +  g<int>();
> +}
diff mbox series

Patch

diff --git gcc/cp/parser.c gcc/cp/parser.c
index 3a98ed900cd..01097b57eb9 100644
--- gcc/cp/parser.c
+++ gcc/cp/parser.c
@@ -7241,6 +7241,20 @@  cp_parser_postfix_expression (cp_parser *parser, bool address_p, bool cast_p,
 						     complain);
 		      }
 		  }
+		else if (TREE_CODE (postfix_expression) == TEMPLATE_ID_EXPR
+			 && (TREE_CODE (TREE_OPERAND (postfix_expression, 0))
+			     == FUNCTION_DECL)
+			 && args->is_empty ())
+		  {
+		    /* We can get here when treating a name as a template name
+		       (because it was an unqualified-id followed by a < and
+		       name lookup found a decl), but now we see that there
+		       are no arguments, so ADL isn't possible.  */
+		    auto_diagnostic_group d;
+		    error_at (loc, "no matching function for call to %<%D()%>",
+			      postfix_expression.get_value ());
+		    postfix_expression = error_mark_node;
+		  }
 	      }
 
 	    if (TREE_CODE (postfix_expression) == COMPONENT_REF)
diff --git gcc/testsuite/g++.dg/cpp2a/fn-template17.C gcc/testsuite/g++.dg/cpp2a/fn-template17.C
new file mode 100644
index 00000000000..f0d24107682
--- /dev/null
+++ gcc/testsuite/g++.dg/cpp2a/fn-template17.C
@@ -0,0 +1,21 @@ 
+// PR c++/88184
+// { dg-do compile }
+// { dg-options "-std=c++2a -fchecking=2" }
+
+namespace A
+{
+  void f ();
+}
+
+using A::f;
+
+template <typename T> void g ()
+{
+  f<T> (); // { dg-error "no matching function for call" }
+}
+
+void
+fn ()
+{
+  g<int>();
+}
diff --git gcc/testsuite/g++.dg/cpp2a/fn-template18.C gcc/testsuite/g++.dg/cpp2a/fn-template18.C
new file mode 100644
index 00000000000..7fe5e89ace3
--- /dev/null
+++ gcc/testsuite/g++.dg/cpp2a/fn-template18.C
@@ -0,0 +1,23 @@ 
+// PR c++/88184
+// { dg-do compile }
+// { dg-options "-std=c++2a -fchecking=2" }
+
+namespace A
+{
+  void f ();
+  void f (int);
+  void f (int, int);
+}
+
+using A::f;
+
+template <typename T> void g ()
+{
+  f<T> (); // { dg-error "no matching function for call" }
+}
+
+void
+fn ()
+{
+  g<int>();
+}