c++: ICE on loopy var tmpl auto deduction [PR109300]

  Now that we resolve non-dependent variable template-ids ahead of time,
cp_finish_decl needs to handle a new invalid situation: we can end up
trying to instantiate a variable template with deduced return type
before we fully parsed (and attached) its initializer.

Bootstrapped and regtested on x86_64-pc-linux-gnu, does this OK for
trunK?

	PR c++/109300

gcc/cp/ChangeLog:

	* decl.cc (cp_finish_decl): Diagnose ordinary auto deduction
	with no initializer instead of asserting.

gcc/testsuite/ChangeLog:

	* g++.dg/cpp1y/var-templ79.C: New test.
---
 gcc/cp/decl.cc                           | 15 ++++++++++++++-
 gcc/testsuite/g++.dg/cpp1y/var-templ79.C |  5 +++++
 2 files changed, 19 insertions(+), 1 deletion(-)
 create mode 100644 gcc/testsuite/g++.dg/cpp1y/var-templ79.C
  

On 3/28/23 13:37, Patrick Palka wrote:
> Now that we resolve non-dependent variable template-ids ahead of time,
> cp_finish_decl needs to handle a new invalid situation: we can end up
> trying to instantiate a variable template with deduced return type
> before we fully parsed (and attached) its initializer.
> 
> Bootstrapped and regtested on x86_64-pc-linux-gnu, does this OK for
> trunK?
> 
> 	PR c++/109300
> 
> gcc/cp/ChangeLog:
> 
> 	* decl.cc (cp_finish_decl): Diagnose ordinary auto deduction
> 	with no initializer instead of asserting.
> 
> gcc/testsuite/ChangeLog:
> 
> 	* g++.dg/cpp1y/var-templ79.C: New test.
> ---
>   gcc/cp/decl.cc                           | 15 ++++++++++++++-
>   gcc/testsuite/g++.dg/cpp1y/var-templ79.C |  5 +++++
>   2 files changed, 19 insertions(+), 1 deletion(-)
>   create mode 100644 gcc/testsuite/g++.dg/cpp1y/var-templ79.C
> 
> diff --git a/gcc/cp/decl.cc b/gcc/cp/decl.cc
> index 20b980f68c8..2c91693b99d 100644
> --- a/gcc/cp/decl.cc
> +++ b/gcc/cp/decl.cc
> @@ -8276,7 +8276,20 @@ cp_finish_decl (tree decl, tree init, bool init_const_expr_p,
>   	      return;
>   	    }
>   
> -	  gcc_assert (CLASS_PLACEHOLDER_TEMPLATE (auto_node));
> +	  if (CLASS_PLACEHOLDER_TEMPLATE (auto_node))
> +	    /* Class deduction with no initializer is OK.  */;
> +	  else
> +	    {
> +	      /* Ordinary auto deduction without an initializer, a situation
> +		 which grokdeclarator already catches and rejects for the most
> +		 part.  But we can still get here if we're instantiating a
> +		 variable template before we've fully parsed (and attached) its
> +		 initializer, e.g. template<class> auto x = x<int>;  */

In the case of recursively dependent instantiation I'd hope to have an 
error_mark_node initializer, rather than none?

> +	      error_at (DECL_SOURCE_LOCATION (decl),
> +			"declaration of %q#D has no initializer", decl);
> +	      TREE_TYPE (decl) = error_mark_node;
> +	      return;
> +	    }
>   	}
>         d_init = init;
>         if (d_init)
> diff --git a/gcc/testsuite/g++.dg/cpp1y/var-templ79.C b/gcc/testsuite/g++.dg/cpp1y/var-templ79.C
> new file mode 100644
> index 00000000000..3c0d276153a
> --- /dev/null
> +++ b/gcc/testsuite/g++.dg/cpp1y/var-templ79.C
> @@ -0,0 +1,5 @@
> +// PR c++/109300
> +// { dg-do compile { target c++14 } }
> +
> +template<class>
> +auto x = x<int>; // { dg-error "" }
  

On Wed, 29 Mar 2023, Jason Merrill wrote:

> On 3/28/23 13:37, Patrick Palka wrote:
> > Now that we resolve non-dependent variable template-ids ahead of time,
> > cp_finish_decl needs to handle a new invalid situation: we can end up
> > trying to instantiate a variable template with deduced return type
> > before we fully parsed (and attached) its initializer.
> > 
> > Bootstrapped and regtested on x86_64-pc-linux-gnu, does this OK for
> > trunK?
> > 
> > 	PR c++/109300
> > 
> > gcc/cp/ChangeLog:
> > 
> > 	* decl.cc (cp_finish_decl): Diagnose ordinary auto deduction
> > 	with no initializer instead of asserting.
> > 
> > gcc/testsuite/ChangeLog:
> > 
> > 	* g++.dg/cpp1y/var-templ79.C: New test.
> > ---
> >   gcc/cp/decl.cc                           | 15 ++++++++++++++-
> >   gcc/testsuite/g++.dg/cpp1y/var-templ79.C |  5 +++++
> >   2 files changed, 19 insertions(+), 1 deletion(-)
> >   create mode 100644 gcc/testsuite/g++.dg/cpp1y/var-templ79.C
> > 
> > diff --git a/gcc/cp/decl.cc b/gcc/cp/decl.cc
> > index 20b980f68c8..2c91693b99d 100644
> > --- a/gcc/cp/decl.cc
> > +++ b/gcc/cp/decl.cc
> > @@ -8276,7 +8276,20 @@ cp_finish_decl (tree decl, tree init, bool
> > init_const_expr_p,
> >   	      return;
> >   	    }
> >   -	  gcc_assert (CLASS_PLACEHOLDER_TEMPLATE (auto_node));
> > +	  if (CLASS_PLACEHOLDER_TEMPLATE (auto_node))
> > +	    /* Class deduction with no initializer is OK.  */;
> > +	  else
> > +	    {
> > +	      /* Ordinary auto deduction without an initializer, a situation
> > +		 which grokdeclarator already catches and rejects for the most
> > +		 part.  But we can still get here if we're instantiating a
> > +		 variable template before we've fully parsed (and attached)
> > its
> > +		 initializer, e.g. template<class> auto x = x<int>;  */
> 
> In the case of recursively dependent instantiation I'd hope to have an
> error_mark_node initializer, rather than none?

Do you mean setting the initializer to error_mark_node after the fact, e.g.

@@ -8288,7 +8297,7 @@ cp_finish_decl (tree decl, tree init, bool init_const_expr_p,
              error_at (DECL_SOURCE_LOCATION (decl),
                        "declaration of %q#D has no initializer", decl);
              TREE_TYPE (decl) = error_mark_node;
-             return;
+             init = error_mark_node;
            }
        }
       d_init = init;

or before the fact, i.e. setting DECL_INITIAL to error_mark_node as a
sentinel value for detecting recursion before we begin parsing a variable
initializer?  The former should work I suppose, but the latter is
problematic because we also call cp_finish_decl with init=error_mark_node
when the initializer is generally invalid, so by overloading the meaning
of error_mark_node here and checking for it from cp_finish_decl we would
end up emitting a bogus extra diagnostic in a bunch of cases e.g.
g++.dg/pr53055.C:

  int i = p ->* p ; // invalid initializer

I guess we would need to use a different sentinel value for detecting
recursion, or expose and inspect the 'lambda_scope' stack which already
keeps track of whether we're in the middle of a variable initializer...
Dunno if it's worth it just for sake of a better diagnostic for this
corner case, I notice e.g. Clang doesn't give a great diagnostic either:

 src/gcc/testsuite/g++.dg/cpp1y/var-templ79.C:5:6: error: declaration of variable 'x' with deduced type 'auto' requires an initializer
 auto x = x<int>; // { dg-error "" }
      ^

> 
> > +	      error_at (DECL_SOURCE_LOCATION (decl),
> > +			"declaration of %q#D has no initializer", decl);
> > +	      TREE_TYPE (decl) = error_mark_node;
> > +	      return;
> > +	    }
> >   	}
> >         d_init = init;
> >         if (d_init)
> > diff --git a/gcc/testsuite/g++.dg/cpp1y/var-templ79.C
> > b/gcc/testsuite/g++.dg/cpp1y/var-templ79.C
> > new file mode 100644
> > index 00000000000..3c0d276153a
> > --- /dev/null
> > +++ b/gcc/testsuite/g++.dg/cpp1y/var-templ79.C
> > @@ -0,0 +1,5 @@
> > +// PR c++/109300
> > +// { dg-do compile { target c++14 } }
> > +
> > +template<class>
> > +auto x = x<int>; // { dg-error "" }
> 
>
  

On 4/3/23 12:28, Patrick Palka wrote:
> On Wed, 29 Mar 2023, Jason Merrill wrote:
> 
>> On 3/28/23 13:37, Patrick Palka wrote:
>>> Now that we resolve non-dependent variable template-ids ahead of time,
>>> cp_finish_decl needs to handle a new invalid situation: we can end up
>>> trying to instantiate a variable template with deduced return type
>>> before we fully parsed (and attached) its initializer.
>>>
>>> Bootstrapped and regtested on x86_64-pc-linux-gnu, does this OK for
>>> trunK?
>>>
>>> 	PR c++/109300
>>>
>>> gcc/cp/ChangeLog:
>>>
>>> 	* decl.cc (cp_finish_decl): Diagnose ordinary auto deduction
>>> 	with no initializer instead of asserting.
>>>
>>> gcc/testsuite/ChangeLog:
>>>
>>> 	* g++.dg/cpp1y/var-templ79.C: New test.
>>> ---
>>>    gcc/cp/decl.cc                           | 15 ++++++++++++++-
>>>    gcc/testsuite/g++.dg/cpp1y/var-templ79.C |  5 +++++
>>>    2 files changed, 19 insertions(+), 1 deletion(-)
>>>    create mode 100644 gcc/testsuite/g++.dg/cpp1y/var-templ79.C
>>>
>>> diff --git a/gcc/cp/decl.cc b/gcc/cp/decl.cc
>>> index 20b980f68c8..2c91693b99d 100644
>>> --- a/gcc/cp/decl.cc
>>> +++ b/gcc/cp/decl.cc
>>> @@ -8276,7 +8276,20 @@ cp_finish_decl (tree decl, tree init, bool
>>> init_const_expr_p,
>>>    	      return;
>>>    	    }
>>>    -	  gcc_assert (CLASS_PLACEHOLDER_TEMPLATE (auto_node));
>>> +	  if (CLASS_PLACEHOLDER_TEMPLATE (auto_node))
>>> +	    /* Class deduction with no initializer is OK.  */;
>>> +	  else
>>> +	    {
>>> +	      /* Ordinary auto deduction without an initializer, a situation
>>> +		 which grokdeclarator already catches and rejects for the most
>>> +		 part.  But we can still get here if we're instantiating a
>>> +		 variable template before we've fully parsed (and attached)
>>> its
>>> +		 initializer, e.g. template<class> auto x = x<int>;  */
>>
>> In the case of recursively dependent instantiation I'd hope to have an
>> error_mark_node initializer, rather than none?
> 
> Do you mean setting the initializer to error_mark_node after the fact, e.g.
> 
> @@ -8288,7 +8297,7 @@ cp_finish_decl (tree decl, tree init, bool init_const_expr_p,
>                error_at (DECL_SOURCE_LOCATION (decl),
>                          "declaration of %q#D has no initializer", decl);
>                TREE_TYPE (decl) = error_mark_node;
> -             return;
> +             init = error_mark_node;
>              }
>          }
>         d_init = init;
> 
> or before the fact, i.e. setting DECL_INITIAL to error_mark_node as a
> sentinel value for detecting recursion before we begin parsing a variable
> initializer?  The former should work I suppose, but the latter is
> problematic because we also call cp_finish_decl with init=error_mark_node
> when the initializer is generally invalid, so by overloading the meaning
> of error_mark_node here and checking for it from cp_finish_decl we would
> end up emitting a bogus extra diagnostic in a bunch of cases e.g.
> g++.dg/pr53055.C:
> 
>    int i = p ->* p ; // invalid initializer
> 
> I guess we would need to use a different sentinel value for detecting
> recursion, or expose and inspect the 'lambda_scope' stack which already
> keeps track of whether we're in the middle of a variable initializer...
> Dunno if it's worth it just for sake of a better diagnostic for this
> corner case, I notice e.g. Clang doesn't give a great diagnostic either:
> 
>   src/gcc/testsuite/g++.dg/cpp1y/var-templ79.C:5:6: error: declaration of variable 'x' with deduced type 'auto' requires an initializer
>   auto x = x<int>; // { dg-error "" }
>        ^

Yeah, let's just go with your patch, thanks.

>>
>>> +	      error_at (DECL_SOURCE_LOCATION (decl),
>>> +			"declaration of %q#D has no initializer", decl);
>>> +	      TREE_TYPE (decl) = error_mark_node;
>>> +	      return;
>>> +	    }
>>>    	}
>>>          d_init = init;
>>>          if (d_init)
>>> diff --git a/gcc/testsuite/g++.dg/cpp1y/var-templ79.C
>>> b/gcc/testsuite/g++.dg/cpp1y/var-templ79.C
>>> new file mode 100644
>>> index 00000000000..3c0d276153a
>>> --- /dev/null
>>> +++ b/gcc/testsuite/g++.dg/cpp1y/var-templ79.C
>>> @@ -0,0 +1,5 @@
>>> +// PR c++/109300
>>> +// { dg-do compile { target c++14 } }
>>> +
>>> +template<class>
>>> +auto x = x<int>; // { dg-error "" }
>>
>>
>