c++: template-id ADL and partial instantiation [PR99911]

Message ID 20211110165307.4079746-1-ppalka@redhat.com
State New
Headers
Series c++: template-id ADL and partial instantiation [PR99911] |

Commit Message

Patrick Palka Nov. 10, 2021, 4:53 p.m. UTC
  Here when partially instantiating the call get<U>(T{}) with T=N::A
(for which earlier unqualified name lookup for 'get' found nothing)
the arguments after substitution are no longer dependent but the callee
still is, so perform_koenig_lookup postpones ADL.  But then we go on to
diagnose the unresolved template name anyway, as if ADL was already
performed and failed.

This patch fixes this by avoiding the error path in question when the
template arguments of an unresolved template-id are dependent, which
mirrors the dependence check in perform_koenig_lookup.  In passing, this
patch also disables the -fpermissive fallback that performs a second
unqualified lookup in the template-id ADL case; this fallback seems to be
intended for legacy code and shouldn't be used for C++20 template-id ADL.

Bootstrapped and regtested on x86_64-pc-linux-gnu, does this look OK for
trunk and perhaps 11?

	PR c++/99911

gcc/cp/ChangeLog:

	* pt.c (tsubst_copy_and_build) <case CALL_EXPR>: Don't diagnose
	name lookup failure if the arguments to an unresolved template
	name are still dependent.  Disable the -fpermissive fallback for
	template-id ADL.

gcc/testsuite/ChangeLog:

	* g++.dg/cpp2a/fn-template24.C: New test.
---
 gcc/cp/pt.c                                |  6 ++++--
 gcc/testsuite/g++.dg/cpp2a/fn-template24.C | 16 ++++++++++++++++
 2 files changed, 20 insertions(+), 2 deletions(-)
 create mode 100644 gcc/testsuite/g++.dg/cpp2a/fn-template24.C
  

Comments

Jason Merrill Nov. 18, 2021, 5:15 a.m. UTC | #1
On 11/10/21 11:53, Patrick Palka wrote:
> Here when partially instantiating the call get<U>(T{}) with T=N::A
> (for which earlier unqualified name lookup for 'get' found nothing)
> the arguments after substitution are no longer dependent but the callee
> still is, so perform_koenig_lookup postpones ADL.  But then we go on to
> diagnose the unresolved template name anyway, as if ADL was already
> performed and failed.
> 
> This patch fixes this by avoiding the error path in question when the
> template arguments of an unresolved template-id are dependent, which
> mirrors the dependence check in perform_koenig_lookup.

This change is OK.

> In passing, this
> patch also disables the -fpermissive fallback that performs a second
> unqualified lookup in the template-id ADL case; this fallback seems to be
> intended for legacy code and shouldn't be used for C++20 template-id ADL.

Why wouldn't we want the more helpful diagnostic?

> Bootstrapped and regtested on x86_64-pc-linux-gnu, does this look OK for
> trunk and perhaps 11?
> 
> 	PR c++/99911
> 
> gcc/cp/ChangeLog:
> 
> 	* pt.c (tsubst_copy_and_build) <case CALL_EXPR>: Don't diagnose
> 	name lookup failure if the arguments to an unresolved template
> 	name are still dependent.  Disable the -fpermissive fallback for
> 	template-id ADL.
> 
> gcc/testsuite/ChangeLog:
> 
> 	* g++.dg/cpp2a/fn-template24.C: New test.
> ---
>   gcc/cp/pt.c                                |  6 ++++--
>   gcc/testsuite/g++.dg/cpp2a/fn-template24.C | 16 ++++++++++++++++
>   2 files changed, 20 insertions(+), 2 deletions(-)
>   create mode 100644 gcc/testsuite/g++.dg/cpp2a/fn-template24.C
> 
> diff --git a/gcc/cp/pt.c b/gcc/cp/pt.c
> index 991a20a85d4..4beddf9caf8 100644
> --- a/gcc/cp/pt.c
> +++ b/gcc/cp/pt.c
> @@ -20427,12 +20427,14 @@ tsubst_copy_and_build (tree t,
>   	if (function != NULL_TREE
>   	    && (identifier_p (function)
>   		|| (TREE_CODE (function) == TEMPLATE_ID_EXPR
> -		    && identifier_p (TREE_OPERAND (function, 0))))
> +		    && identifier_p (TREE_OPERAND (function, 0))
> +		    && !any_dependent_template_arguments_p (TREE_OPERAND
> +							    (function, 1))))
>   	    && !any_type_dependent_arguments_p (call_args))
>   	  {
>   	    if (TREE_CODE (function) == TEMPLATE_ID_EXPR)
>   	      function = TREE_OPERAND (function, 0);
> -	    if (koenig_p && (complain & tf_warning_or_error))
> +	    else if (koenig_p && (complain & tf_warning_or_error))
>   	      {
>   		/* For backwards compatibility and good diagnostics, try
>   		   the unqualified lookup again if we aren't in SFINAE
> diff --git a/gcc/testsuite/g++.dg/cpp2a/fn-template24.C b/gcc/testsuite/g++.dg/cpp2a/fn-template24.C
> new file mode 100644
> index 00000000000..b444ac6a273
> --- /dev/null
> +++ b/gcc/testsuite/g++.dg/cpp2a/fn-template24.C
> @@ -0,0 +1,16 @@
> +// PR c++/99911
> +// { dg-do compile { target c++20 } }
> +
> +namespace N {
> +  struct A { };
> +  template<class T> void get(A);
> +};
> +
> +template<class T>
> +auto f() {
> +  return []<class U>(U) { get<U>(T{}); };
> +}
> +
> +int main() {
> +  f<N::A>()(0);
> +}
>
  
Patrick Palka Nov. 18, 2021, 2:45 p.m. UTC | #2
On Thu, 18 Nov 2021, Jason Merrill wrote:

> On 11/10/21 11:53, Patrick Palka wrote:
> > Here when partially instantiating the call get<U>(T{}) with T=N::A
> > (for which earlier unqualified name lookup for 'get' found nothing)
> > the arguments after substitution are no longer dependent but the callee
> > still is, so perform_koenig_lookup postpones ADL.  But then we go on to
> > diagnose the unresolved template name anyway, as if ADL was already
> > performed and failed.
> > 
> > This patch fixes this by avoiding the error path in question when the
> > template arguments of an unresolved template-id are dependent, which
> > mirrors the dependence check in perform_koenig_lookup.
> 
> This change is OK.
> 
> > In passing, this
> > patch also disables the -fpermissive fallback that performs a second
> > unqualified lookup in the template-id ADL case; this fallback seems to be
> > intended for legacy code and shouldn't be used for C++20 template-id ADL.
> 
> Why wouldn't we want the more helpful diagnostic?

The "no declarations were found by ADL" diagnostic is helpful, but the
backwards compatibility logic doesn't correctly handle the template-id
case.  E.g. for

  template<class T>
  void f() {
    g<int>(T{});
  }

  template<class T>
  void g(int); // #1

  int main() {
    f<int>();
  }

we get the helpful diagnostic followed by a confusing one because we
didn't incorporate the template-id's template arguments when replacing
the callee with the later-declared template #1:

<stdin>: In instantiation of ‘void f() [with T = int]’:
<stdin>:10:9:   required from here
<stdin>:3:6: error: ‘g’ was not declared in this scope, and no declarations were found by argument-dependent lookup at the point of instantiation [-fpermissive]
<stdin>:7:6: note: ‘template<class T> void g(int)’ declared here, later in the translation unit
<stdin>:3:6: error: no matching function for call to ‘g(int)’
<stdin>:7:6: note: candidate: ‘template<class T> void g(int)’
<stdin>:7:6: note:   template argument deduction/substitution failed:
<stdin>:3:6: note:   couldn’t deduce template parameter ‘T’


We also ignores template-ness of the name being looked up, so e.g. for:

  template<class T>
  void f() {
    g<>(T{});
  }

  void g(int); // #1

  int main() {
    f<int>();
  }

the secondary unqualified lookup finds the later-declared non-template #1:

<stdin>: In instantiation of ‘void f() [with T = int]’:
<stdin>:9:9:   required from here
<stdin>:3:6: error: ‘g’ was not declared in this scope, and no declarations were found by argument-dependent lookup at the point of instantiation [-fpermissive]
<stdin>:6:6: note: ‘void g(int)’ declared here, later in the translation unit

which doesn't seem right.

To fix the first issue, rather than disabling the diagnostic perhaps we
should just disable the backwards compatibility logic in the template-id
case, as in the below?

-- >8 --

diff --git a/gcc/cp/pt.c b/gcc/cp/pt.c
index 82bf7dc26f6..e2d04a52894 100644
--- a/gcc/cp/pt.c
+++ b/gcc/cp/pt.c
@@ -20428,7 +20428,8 @@ tsubst_copy_and_build (tree t,
 		    && identifier_p (TREE_OPERAND (function, 0))))
 	    && !any_type_dependent_arguments_p (call_args))
 	  {
-	    if (TREE_CODE (function) == TEMPLATE_ID_EXPR)
+	    bool template_id_p = (TREE_CODE (function) == TEMPLATE_ID_EXPR);
+	    if (template_id_p)
 	      function = TREE_OPERAND (function, 0);
 	    if (koenig_p && (complain & tf_warning_or_error))
 	      {
@@ -20443,20 +20444,21 @@ tsubst_copy_and_build (tree t,
 
 		if (unq != function)
 		  {
-		    /* In a lambda fn, we have to be careful to not
-		       introduce new this captures.  Legacy code can't
-		       be using lambdas anyway, so it's ok to be
-		       stricter.  */
-		    bool in_lambda = (current_class_type
-				      && LAMBDA_TYPE_P (current_class_type));
 		    char const *const msg
 		      = G_("%qD was not declared in this scope, "
 			   "and no declarations were found by "
 			   "argument-dependent lookup at the point "
 			   "of instantiation");
 
+		    bool in_lambda = (current_class_type
+				      && LAMBDA_TYPE_P (current_class_type));
+		    /* In a lambda fn, we have to be careful to not
+		       introduce new this captures.  Legacy code can't
+		       be using lambdas anyway, so it's ok to be
+		       stricter.  Be strict with C++20 template-id ADL too.  */
+		    bool strict = in_lambda || template_id_p;
 		    bool diag = true;
-		    if (in_lambda)
+		    if (strict)
 		      error_at (cp_expr_loc_or_input_loc (t),
 				msg, function);
 		    else
@@ -20492,7 +20494,7 @@ tsubst_copy_and_build (tree t,
 			  inform (DECL_SOURCE_LOCATION (fn),
 				  "%qD declared here, later in the "
 				  "translation unit", fn);
-			if (in_lambda)
+			if (strict)
 			  RETURN (error_mark_node);
 		      }
  
Patrick Palka Nov. 18, 2021, 3:31 p.m. UTC | #3
On Thu, 18 Nov 2021, Patrick Palka wrote:

> On Thu, 18 Nov 2021, Jason Merrill wrote:
> 
> > On 11/10/21 11:53, Patrick Palka wrote:
> > > Here when partially instantiating the call get<U>(T{}) with T=N::A
> > > (for which earlier unqualified name lookup for 'get' found nothing)
> > > the arguments after substitution are no longer dependent but the callee
> > > still is, so perform_koenig_lookup postpones ADL.  But then we go on to
> > > diagnose the unresolved template name anyway, as if ADL was already
> > > performed and failed.
> > > 
> > > This patch fixes this by avoiding the error path in question when the
> > > template arguments of an unresolved template-id are dependent, which
> > > mirrors the dependence check in perform_koenig_lookup.
> > 
> > This change is OK.
> > 
> > > In passing, this
> > > patch also disables the -fpermissive fallback that performs a second
> > > unqualified lookup in the template-id ADL case; this fallback seems to be
> > > intended for legacy code and shouldn't be used for C++20 template-id ADL.
> > 
> > Why wouldn't we want the more helpful diagnostic?
> 
> The "no declarations were found by ADL" diagnostic is helpful, but the
> backwards compatibility logic doesn't correctly handle the template-id
> case.  E.g. for
> 
>   template<class T>
>   void f() {
>     g<int>(T{});
>   }
> 
>   template<class T>
>   void g(int); // #1
> 
>   int main() {
>     f<int>();
>   }
> 
> we get the helpful diagnostic followed by a confusing one because we
> didn't incorporate the template-id's template arguments when replacing
> the callee with the later-declared template #1:
> 
> <stdin>: In instantiation of ‘void f() [with T = int]’:
> <stdin>:10:9:   required from here
> <stdin>:3:6: error: ‘g’ was not declared in this scope, and no declarations were found by argument-dependent lookup at the point of instantiation [-fpermissive]
> <stdin>:7:6: note: ‘template<class T> void g(int)’ declared here, later in the translation unit
> <stdin>:3:6: error: no matching function for call to ‘g(int)’
> <stdin>:7:6: note: candidate: ‘template<class T> void g(int)’
> <stdin>:7:6: note:   template argument deduction/substitution failed:
> <stdin>:3:6: note:   couldn’t deduce template parameter ‘T’
> 
> 
> We also ignores template-ness of the name being looked up, so e.g. for:
> 
>   template<class T>
>   void f() {
>     g<>(T{});
>   }
> 
>   void g(int); // #1
> 
>   int main() {
>     f<int>();
>   }
> 
> the secondary unqualified lookup finds the later-declared non-template #1:
> 
> <stdin>: In instantiation of ‘void f() [with T = int]’:
> <stdin>:9:9:   required from here
> <stdin>:3:6: error: ‘g’ was not declared in this scope, and no declarations were found by argument-dependent lookup at the point of instantiation [-fpermissive]
> <stdin>:6:6: note: ‘void g(int)’ declared here, later in the translation unit
> 
> which doesn't seem right.
> 
> To fix the first issue, rather than disabling the diagnostic perhaps we
> should just disable the backwards compatibility logic in the template-id
> case, as in the below?

Now in patch form:

-- >8 --

Subject: [PATCH] c++: error recovery during C++20 template-id ADL failure

When diagnosing ADL failure we perform a second unqualified lookup for
backwards compatibility with legacy code, and for better diagnostics.

For C++20 template-id ADL however, the backwards compatibility logic
causes confusing subsequent diagnostics, such as in the testcase below
where we report deduction failure following the useful "no declarations
were found by ADL" diagnostic because we've discarded the arguments of
the template-id when replacing it with the later-declared template.

So for C++20 template-id ADL, this patch just disables the backwards
compatibility code while keeping the useful diagnostic.

gcc/cp/ChangeLog:

	* pt.c (tsubst_copy_and_build) <case CALL_EXPR>: Disable the
	-fpermissive fallback for template-id ADL, but keep the
	diagnostic.

gcc/testsuite/ChangeLog:

	* g++.dg/cpp2a/fn-template25.C: New test.
---
 gcc/cp/pt.c                                | 23 +++++++++++++---------
 gcc/testsuite/g++.dg/cpp2a/fn-template25.C | 14 +++++++++++++
 2 files changed, 28 insertions(+), 9 deletions(-)
 create mode 100644 gcc/testsuite/g++.dg/cpp2a/fn-template25.C

diff --git a/gcc/cp/pt.c b/gcc/cp/pt.c
index ad51c07347b..3f1550a17ad 100644
--- a/gcc/cp/pt.c
+++ b/gcc/cp/pt.c
@@ -20439,8 +20439,12 @@ tsubst_copy_and_build (tree t,
 							    (function, 1))))
 	    && !any_type_dependent_arguments_p (call_args))
 	  {
+	    bool template_id_p = false;
 	    if (TREE_CODE (function) == TEMPLATE_ID_EXPR)
-	      function = TREE_OPERAND (function, 0);
+	      {
+		function = TREE_OPERAND (function, 0);
+		template_id_p = true;
+	      }
 	    if (koenig_p && (complain & tf_warning_or_error))
 	      {
 		/* For backwards compatibility and good diagnostics, try
@@ -20454,20 +20458,21 @@ tsubst_copy_and_build (tree t,
 
 		if (unq != function)
 		  {
-		    /* In a lambda fn, we have to be careful to not
-		       introduce new this captures.  Legacy code can't
-		       be using lambdas anyway, so it's ok to be
-		       stricter.  */
-		    bool in_lambda = (current_class_type
-				      && LAMBDA_TYPE_P (current_class_type));
 		    char const *const msg
 		      = G_("%qD was not declared in this scope, "
 			   "and no declarations were found by "
 			   "argument-dependent lookup at the point "
 			   "of instantiation");
 
+		    bool in_lambda = (current_class_type
+				      && LAMBDA_TYPE_P (current_class_type));
+		    /* In a lambda fn, we have to be careful to not
+		       introduce new this captures.  Legacy code can't
+		       be using lambdas anyway, so it's ok to be
+		       stricter.  Be strict with C++20 template-id ADL too.  */
+		    bool strict = in_lambda || template_id_p;
 		    bool diag = true;
-		    if (in_lambda)
+		    if (strict)
 		      error_at (cp_expr_loc_or_input_loc (t),
 				msg, function);
 		    else
@@ -20503,7 +20508,7 @@ tsubst_copy_and_build (tree t,
 			  inform (DECL_SOURCE_LOCATION (fn),
 				  "%qD declared here, later in the "
 				  "translation unit", fn);
-			if (in_lambda)
+			if (strict)
 			  RETURN (error_mark_node);
 		      }
 
diff --git a/gcc/testsuite/g++.dg/cpp2a/fn-template25.C b/gcc/testsuite/g++.dg/cpp2a/fn-template25.C
new file mode 100644
index 00000000000..a8888af2023
--- /dev/null
+++ b/gcc/testsuite/g++.dg/cpp2a/fn-template25.C
@@ -0,0 +1,14 @@
+// { dg-do compile { target c++20 } }
+
+template<class T>
+void f() {
+  g<int>(T{}); // { dg-error "argument-dependent lookup" }
+	       // { dg-bogus "no match" "" { target *-*-* } .-1 }
+}
+
+template<class T>
+void g(int);   // { dg-message "declared here, later" }
+
+int main() {
+  f<int>();
+}
  
Jason Merrill Nov. 18, 2021, 4:02 p.m. UTC | #4
On 11/18/21 10:31, Patrick Palka wrote:
> On Thu, 18 Nov 2021, Patrick Palka wrote:
> 
>> On Thu, 18 Nov 2021, Jason Merrill wrote:
>>
>>> On 11/10/21 11:53, Patrick Palka wrote:
>>>> Here when partially instantiating the call get<U>(T{}) with T=N::A
>>>> (for which earlier unqualified name lookup for 'get' found nothing)
>>>> the arguments after substitution are no longer dependent but the callee
>>>> still is, so perform_koenig_lookup postpones ADL.  But then we go on to
>>>> diagnose the unresolved template name anyway, as if ADL was already
>>>> performed and failed.
>>>>
>>>> This patch fixes this by avoiding the error path in question when the
>>>> template arguments of an unresolved template-id are dependent, which
>>>> mirrors the dependence check in perform_koenig_lookup.
>>>
>>> This change is OK.
>>>
>>>> In passing, this
>>>> patch also disables the -fpermissive fallback that performs a second
>>>> unqualified lookup in the template-id ADL case; this fallback seems to be
>>>> intended for legacy code and shouldn't be used for C++20 template-id ADL.
>>>
>>> Why wouldn't we want the more helpful diagnostic?
>>
>> The "no declarations were found by ADL" diagnostic is helpful, but the
>> backwards compatibility logic doesn't correctly handle the template-id
>> case.  E.g. for
>>
>>    template<class T>
>>    void f() {
>>      g<int>(T{});
>>    }
>>
>>    template<class T>
>>    void g(int); // #1
>>
>>    int main() {
>>      f<int>();
>>    }
>>
>> we get the helpful diagnostic followed by a confusing one because we
>> didn't incorporate the template-id's template arguments when replacing
>> the callee with the later-declared template #1:
>>
>> <stdin>: In instantiation of ‘void f() [with T = int]’:
>> <stdin>:10:9:   required from here
>> <stdin>:3:6: error: ‘g’ was not declared in this scope, and no declarations were found by argument-dependent lookup at the point of instantiation [-fpermissive]
>> <stdin>:7:6: note: ‘template<class T> void g(int)’ declared here, later in the translation unit
>> <stdin>:3:6: error: no matching function for call to ‘g(int)’
>> <stdin>:7:6: note: candidate: ‘template<class T> void g(int)’
>> <stdin>:7:6: note:   template argument deduction/substitution failed:
>> <stdin>:3:6: note:   couldn’t deduce template parameter ‘T’
>>
>>
>> We also ignores template-ness of the name being looked up, so e.g. for:
>>
>>    template<class T>
>>    void f() {
>>      g<>(T{});
>>    }
>>
>>    void g(int); // #1
>>
>>    int main() {
>>      f<int>();
>>    }
>>
>> the secondary unqualified lookup finds the later-declared non-template #1:
>>
>> <stdin>: In instantiation of ‘void f() [with T = int]’:
>> <stdin>:9:9:   required from here
>> <stdin>:3:6: error: ‘g’ was not declared in this scope, and no declarations were found by argument-dependent lookup at the point of instantiation [-fpermissive]
>> <stdin>:6:6: note: ‘void g(int)’ declared here, later in the translation unit
>>
>> which doesn't seem right.
>>
>> To fix the first issue, rather than disabling the diagnostic perhaps we
>> should just disable the backwards compatibility logic in the template-id
>> case, as in the below?
> 
> Now in patch form:
> 
> -- >8 --
> 
> Subject: [PATCH] c++: error recovery during C++20 template-id ADL failure
> 
> When diagnosing ADL failure we perform a second unqualified lookup for
> backwards compatibility with legacy code, and for better diagnostics.
> 
> For C++20 template-id ADL however, the backwards compatibility logic
> causes confusing subsequent diagnostics, such as in the testcase below
> where we report deduction failure following the useful "no declarations
> were found by ADL" diagnostic because we've discarded the arguments of
> the template-id when replacing it with the later-declared template.
> 
> So for C++20 template-id ADL, this patch just disables the backwards
> compatibility code while keeping the useful diagnostic.
> 
> gcc/cp/ChangeLog:
> 
> 	* pt.c (tsubst_copy_and_build) <case CALL_EXPR>: Disable the
> 	-fpermissive fallback for template-id ADL, but keep the
> 	diagnostic.
> 
> gcc/testsuite/ChangeLog:
> 
> 	* g++.dg/cpp2a/fn-template25.C: New test.
> ---
>   gcc/cp/pt.c                                | 23 +++++++++++++---------
>   gcc/testsuite/g++.dg/cpp2a/fn-template25.C | 14 +++++++++++++
>   2 files changed, 28 insertions(+), 9 deletions(-)
>   create mode 100644 gcc/testsuite/g++.dg/cpp2a/fn-template25.C
> 
> diff --git a/gcc/cp/pt.c b/gcc/cp/pt.c
> index ad51c07347b..3f1550a17ad 100644
> --- a/gcc/cp/pt.c
> +++ b/gcc/cp/pt.c
> @@ -20439,8 +20439,12 @@ tsubst_copy_and_build (tree t,
>   							    (function, 1))))
>   	    && !any_type_dependent_arguments_p (call_args))
>   	  {
> +	    bool template_id_p = false;
>   	    if (TREE_CODE (function) == TEMPLATE_ID_EXPR)
> -	      function = TREE_OPERAND (function, 0);
> +	      {
> +		function = TREE_OPERAND (function, 0);
> +		template_id_p = true;
> +	      }

I think

   bool template_id_p = TREE_CODE (function) == TEMPLATE_ID_EXPR;

would simplify this.  OK either way.

>   	    if (koenig_p && (complain & tf_warning_or_error))
>   	      {
>   		/* For backwards compatibility and good diagnostics, try
> @@ -20454,20 +20458,21 @@ tsubst_copy_and_build (tree t,
>   
>   		if (unq != function)
>   		  {
> -		    /* In a lambda fn, we have to be careful to not
> -		       introduce new this captures.  Legacy code can't
> -		       be using lambdas anyway, so it's ok to be
> -		       stricter.  */
> -		    bool in_lambda = (current_class_type
> -				      && LAMBDA_TYPE_P (current_class_type));
>   		    char const *const msg
>   		      = G_("%qD was not declared in this scope, "
>   			   "and no declarations were found by "
>   			   "argument-dependent lookup at the point "
>   			   "of instantiation");
>   
> +		    bool in_lambda = (current_class_type
> +				      && LAMBDA_TYPE_P (current_class_type));
> +		    /* In a lambda fn, we have to be careful to not
> +		       introduce new this captures.  Legacy code can't
> +		       be using lambdas anyway, so it's ok to be
> +		       stricter.  Be strict with C++20 template-id ADL too.  */
> +		    bool strict = in_lambda || template_id_p;
>   		    bool diag = true;
> -		    if (in_lambda)
> +		    if (strict)
>   		      error_at (cp_expr_loc_or_input_loc (t),
>   				msg, function);
>   		    else
> @@ -20503,7 +20508,7 @@ tsubst_copy_and_build (tree t,
>   			  inform (DECL_SOURCE_LOCATION (fn),
>   				  "%qD declared here, later in the "
>   				  "translation unit", fn);
> -			if (in_lambda)
> +			if (strict)
>   			  RETURN (error_mark_node);
>   		      }
>   
> diff --git a/gcc/testsuite/g++.dg/cpp2a/fn-template25.C b/gcc/testsuite/g++.dg/cpp2a/fn-template25.C
> new file mode 100644
> index 00000000000..a8888af2023
> --- /dev/null
> +++ b/gcc/testsuite/g++.dg/cpp2a/fn-template25.C
> @@ -0,0 +1,14 @@
> +// { dg-do compile { target c++20 } }
> +
> +template<class T>
> +void f() {
> +  g<int>(T{}); // { dg-error "argument-dependent lookup" }
> +	       // { dg-bogus "no match" "" { target *-*-* } .-1 }
> +}
> +
> +template<class T>
> +void g(int);   // { dg-message "declared here, later" }
> +
> +int main() {
> +  f<int>();
> +}
>
  

Patch

diff --git a/gcc/cp/pt.c b/gcc/cp/pt.c
index 991a20a85d4..4beddf9caf8 100644
--- a/gcc/cp/pt.c
+++ b/gcc/cp/pt.c
@@ -20427,12 +20427,14 @@  tsubst_copy_and_build (tree t,
 	if (function != NULL_TREE
 	    && (identifier_p (function)
 		|| (TREE_CODE (function) == TEMPLATE_ID_EXPR
-		    && identifier_p (TREE_OPERAND (function, 0))))
+		    && identifier_p (TREE_OPERAND (function, 0))
+		    && !any_dependent_template_arguments_p (TREE_OPERAND
+							    (function, 1))))
 	    && !any_type_dependent_arguments_p (call_args))
 	  {
 	    if (TREE_CODE (function) == TEMPLATE_ID_EXPR)
 	      function = TREE_OPERAND (function, 0);
-	    if (koenig_p && (complain & tf_warning_or_error))
+	    else if (koenig_p && (complain & tf_warning_or_error))
 	      {
 		/* For backwards compatibility and good diagnostics, try
 		   the unqualified lookup again if we aren't in SFINAE
diff --git a/gcc/testsuite/g++.dg/cpp2a/fn-template24.C b/gcc/testsuite/g++.dg/cpp2a/fn-template24.C
new file mode 100644
index 00000000000..b444ac6a273
--- /dev/null
+++ b/gcc/testsuite/g++.dg/cpp2a/fn-template24.C
@@ -0,0 +1,16 @@ 
+// PR c++/99911
+// { dg-do compile { target c++20 } }
+
+namespace N {
+  struct A { };
+  template<class T> void get(A);
+};
+
+template<class T>
+auto f() {
+  return []<class U>(U) { get<U>(T{}); };
+}
+
+int main() {
+  f<N::A>()(0);
+}