Re: [PATCH v2] c++: ICE with invalid structured bindings [PR107276]

[Jason Merrill <jason@redhat.com>]

On 10/24/22 13:13, Marek Polacek wrote:
> On Mon, Oct 24, 2022 at 10:31:50AM -0400, Jason Merrill wrote:
>> On 10/21/22 19:29, Marek Polacek wrote:
>>> This test ICEs in C++23 because we reach the new code in do_auto_deduction:
>>>
>>> 30468   if (cxx_dialect >= cxx23
>>> 30469       && context == adc_return_type
>>> 30470       && (!AUTO_IS_DECLTYPE (auto_node)
>>> 30471           || !unparenthesized_id_or_class_member_access_p (init))
>>> 30472       && (r = treat_lvalue_as_rvalue_p (maybe_undo_parenthesized_ref (init),
>>> 30473                                         /*return*/true)))
>>>
>>> where 'init' is "VIEW_CONVERT_EXPR<<<< error >>>>(y)", and then the move
>>> in treat_lvalue_as_rvalue_p returns error_mark_node whereupon
>>> set_implicit_rvalue_p crashes.
>>>
>>> I don't think such V_C_Es are useful so let's not create them.  But that
>>> won't fix the ICE so I'm checking the return value of move.  A structured
>>> bindings decl can have an error type, that is set in cp_finish_decomp:
>>>
>>>    8908           TREE_TYPE (first) = error_mark_node;
>>>
>>> therefore I think treat_lvalue_as_rvalue_p just needs to cope.
>>>
>>> Bootstrapped/regtested on x86_64-pc-linux-gnu, ok for trunk?
>>>
>>> 	PR c++/107276
>>>
>>> gcc/cp/ChangeLog:
>>>
>>> 	* typeck.cc (treat_lvalue_as_rvalue_p): Check the return value of move.
>>>
>>> gcc/ChangeLog:
>>>
>>> 	* tree.cc (maybe_wrap_with_location): Don't create a location wrapper
>>> 	when the type is erroneous.
>>>
>>> gcc/testsuite/ChangeLog:
>>>
>>> 	* g++.dg/cpp2a/decomp4.C: New test.
>>> ---
>>>    gcc/cp/typeck.cc                     | 7 ++++++-
>>>    gcc/testsuite/g++.dg/cpp2a/decomp4.C | 8 ++++++++
>>>    gcc/tree.cc                          | 3 ++-
>>>    3 files changed, 16 insertions(+), 2 deletions(-)
>>>    create mode 100644 gcc/testsuite/g++.dg/cpp2a/decomp4.C
>>>
>>> diff --git a/gcc/cp/typeck.cc b/gcc/cp/typeck.cc
>>> index 16e7d85793d..5ca191759f6 100644
>>> --- a/gcc/cp/typeck.cc
>>> +++ b/gcc/cp/typeck.cc
>>> @@ -10726,7 +10726,12 @@ treat_lvalue_as_rvalue_p (tree expr, bool return_p)
>>>      if (DECL_CONTEXT (retval) != current_function_decl)
>>>        return NULL_TREE;
>>>      if (return_p)
>>> -    return set_implicit_rvalue_p (move (expr));
>>> +    {
>>> +      expr = move (expr);
>>> +      if (expr == error_mark_node)
>>> +	return NULL_TREE;
>>> +      return set_implicit_rvalue_p (expr);
>>> +    }
>>>      /* if the operand of a throw-expression is a (possibly parenthesized)
>>>         id-expression that names an implicitly movable entity whose scope does not
>>> diff --git a/gcc/testsuite/g++.dg/cpp2a/decomp4.C b/gcc/testsuite/g++.dg/cpp2a/decomp4.C
>>> new file mode 100644
>>> index 00000000000..28b3f172b53
>>> --- /dev/null
>>> +++ b/gcc/testsuite/g++.dg/cpp2a/decomp4.C
>>> @@ -0,0 +1,8 @@
>>> +// PR c++/107276
>>> +// { dg-do run { target c++20 } }
>>> +
>>> +auto f(auto x) {
>>> +  auto [y] = x; // { dg-error "cannot decompose" }
>>> +  return y;
>>> +}
>>> +int i = f(0);
>>> diff --git a/gcc/tree.cc b/gcc/tree.cc
>>> index 81a6ceaf181..4e5b1df4d85 100644
>>> --- a/gcc/tree.cc
>>> +++ b/gcc/tree.cc
>>> @@ -14352,7 +14352,8 @@ maybe_wrap_with_location (tree expr, location_t loc)
>>>      /* For now, don't add wrappers to exceptional tree nodes, to minimize
>>>         any impact of the wrapper nodes.  */
>>> -  if (EXCEPTIONAL_CLASS_P (expr))
>>> +  if (EXCEPTIONAL_CLASS_P (expr)
>>> +      || (TREE_TYPE (expr) && EXCEPTIONAL_CLASS_P (TREE_TYPE (expr))))
>>
>> I think check error_operand_p instead; I don't think it makes sense to look
>> for other exceptional nodes in TREE_TYPE.
> 
> Makes sense.  I don't suppose you want to replace the whole condition with
> error_operand_p, only what comes after ||.  Updated patch below, thanks.
> 
> Bootstrapped/regtested on x86_64-pc-linux-gnu, ok for trunk?

OK.

> -- >8 --
> This test ICEs in C++23 because we reach the new code in do_auto_deduction:
> 
> 30468   if (cxx_dialect >= cxx23
> 30469       && context == adc_return_type
> 30470       && (!AUTO_IS_DECLTYPE (auto_node)
> 30471           || !unparenthesized_id_or_class_member_access_p (init))
> 30472       && (r = treat_lvalue_as_rvalue_p (maybe_undo_parenthesized_ref (init),
> 30473                                         /*return*/true)))
> 
> where 'init' is "VIEW_CONVERT_EXPR<<<< error >>>>(y)", and then the move
> in treat_lvalue_as_rvalue_p returns error_mark_node whereupon
> set_implicit_rvalue_p crashes.
> 
> I don't think such V_C_Es are useful so let's not create them.  But that
> won't fix the ICE so I'm checking the return value of move.  A structured
> bindings decl can have an error type, that is set in cp_finish_decomp:
> 
>   8908           TREE_TYPE (first) = error_mark_node;
> 
> therefore I think treat_lvalue_as_rvalue_p just needs to cope.
> 
> 	PR c++/107276
> 
> gcc/cp/ChangeLog:
> 
> 	* typeck.cc (treat_lvalue_as_rvalue_p): Check the return value of move.
> 
> gcc/ChangeLog:
> 
> 	* tree.cc (maybe_wrap_with_location): Don't create a location wrapper
> 	when the type is erroneous.
> 
> gcc/testsuite/ChangeLog:
> 
> 	* g++.dg/cpp2a/decomp4.C: New test.
> ---
>   gcc/cp/typeck.cc                     | 7 ++++++-
>   gcc/testsuite/g++.dg/cpp2a/decomp4.C | 8 ++++++++
>   gcc/tree.cc                          | 2 +-
>   3 files changed, 15 insertions(+), 2 deletions(-)
>   create mode 100644 gcc/testsuite/g++.dg/cpp2a/decomp4.C
> 
> diff --git a/gcc/cp/typeck.cc b/gcc/cp/typeck.cc
> index 16e7d85793d..5ca191759f6 100644
> --- a/gcc/cp/typeck.cc
> +++ b/gcc/cp/typeck.cc
> @@ -10726,7 +10726,12 @@ treat_lvalue_as_rvalue_p (tree expr, bool return_p)
>     if (DECL_CONTEXT (retval) != current_function_decl)
>       return NULL_TREE;
>     if (return_p)
> -    return set_implicit_rvalue_p (move (expr));
> +    {
> +      expr = move (expr);
> +      if (expr == error_mark_node)
> +	return NULL_TREE;
> +      return set_implicit_rvalue_p (expr);
> +    }
>   
>     /* if the operand of a throw-expression is a (possibly parenthesized)
>        id-expression that names an implicitly movable entity whose scope does not
> diff --git a/gcc/testsuite/g++.dg/cpp2a/decomp4.C b/gcc/testsuite/g++.dg/cpp2a/decomp4.C
> new file mode 100644
> index 00000000000..d1b0c90ee26
> --- /dev/null
> +++ b/gcc/testsuite/g++.dg/cpp2a/decomp4.C
> @@ -0,0 +1,8 @@
> +// PR c++/107276
> +// { dg-do compile { target c++20 } }
> +
> +auto f(auto x) {
> +  auto [y] = x; // { dg-error "cannot decompose" }
> +  return y;
> +}
> +int i = f(0);
> diff --git a/gcc/tree.cc b/gcc/tree.cc
> index 81a6ceaf181..04603c8c902 100644
> --- a/gcc/tree.cc
> +++ b/gcc/tree.cc
> @@ -14352,7 +14352,7 @@ maybe_wrap_with_location (tree expr, location_t loc)
>   
>     /* For now, don't add wrappers to exceptional tree nodes, to minimize
>        any impact of the wrapper nodes.  */
> -  if (EXCEPTIONAL_CLASS_P (expr))
> +  if (EXCEPTIONAL_CLASS_P (expr) || error_operand_p (expr))
>       return expr;
>   
>     /* Compiler-generated temporary variables don't need a wrapper.  */
> 
> base-commit: 131d18e928a3ea1ab2d3bf61aa92d68a8a254609