Re: [PATCH] c++: decltype of (non-captured variable) [PR83167]

[Jason Merrill <jason@redhat.com>]

On 12/1/23 17:42, Patrick Palka wrote:
> On Fri, 1 Dec 2023, Jason Merrill wrote:
> 
>> On 12/1/23 12:32, Patrick Palka wrote:
>>> On Tue, 14 Nov 2023, Jason Merrill wrote:
>>>
>>>> On 11/14/23 11:10, Patrick Palka wrote:
>>>>> Bootstrapped and regtested on x86_64-pc-linux-gnu, does this look OK for
>>>>> trunk?
>>>>>
>>>>> -- >8 --
>>>>>
>>>>> For decltype((x)) within a lambda where x is not captured, we dubiously
>>>>> require that the lambda has a capture default, unlike for decltype(x).
>>>>> This patch fixes this inconsistency; I couldn't find a justification for
>>>>> it in the standard.
>>>>
>>>> The relevant passage seems to be
>>>>
>>>> https://eel.is/c++draft/expr.prim#id.unqual-3
>>>>
>>>> "If naming the entity from outside of an unevaluated operand within S
>>>> would
>>>> refer to an entity captured by copy in some intervening lambda-expression,
>>>> then let E be the innermost such lambda-expression.
>>>>
>>>> If there is such a lambda-expression and if P is in E's function parameter
>>>> scope but not its parameter-declaration-clause, then the type of the
>>>> expression is the type of a class member access expression ([expr.ref])
>>>> naming
>>>> the non-static data member that would be declared for such a capture in
>>>> the
>>>> object parameter ([dcl.fct]) of the function call operator of E."
>>>>
>>>> In this case I guess there is no such lambda-expression because naming x
>>>> won't
>>>> refer to a capture by copy if the lambda doesn't capture anything, so we
>>>> ignore the lambda.
>>>>
>>>> Maybe refer to that in a comment?  OK with that change.
>>>>
>>>> I'm surprised that it refers specifically to capture by copy, but I guess
>>>> a
>>>> capture by reference should have the same decltype as the captured
>>>> variable?
>>>
>>> Ah, seems like it.  So maybe we should get rid of the redundant
>>> by-reference capture-default handling, to more closely mirror the
>>> standard?
>>>
>>> Also now that r14-6026-g73e2bdbf9bed48 made capture_decltype return
>>> NULL_TREE to mean the capture is dependent, it seems we should just
>>> inline capture_decltype into finish_decltype_type rather than
>>> introducing another special return value to mean "fall back to ordinary
>>> handling".
>>>
>>> How does the following look?  Bootstrapped and regtested on
>>> x86_64-pc-linux-gnu.
>>>
>>> -- >8 --
>>>
>>> 	PR c++/83167
>>>
>>> gcc/cp/ChangeLog:
>>>
>>> 	* semantics.cc (capture_decltype): Inline into its only caller ...
>>> 	(finish_decltype_type): ... here.  Update nearby comment to refer
>>> 	to recent standard.  Restrict uncaptured variable handling to just
>>> 	lambdas with a by-copy capture-default.
>>>
>>> gcc/testsuite/ChangeLog:
>>>
>>> 	* g++.dg/cpp0x/lambda/lambda-decltype4.C: New test.
>>> ---
>>>    gcc/cp/semantics.cc                           | 107 +++++++-----------
>>>    .../g++.dg/cpp0x/lambda/lambda-decltype4.C    |  15 +++
>>>    2 files changed, 55 insertions(+), 67 deletions(-)
>>>    create mode 100644 gcc/testsuite/g++.dg/cpp0x/lambda/lambda-decltype4.C
>>>
>>> diff --git a/gcc/cp/semantics.cc b/gcc/cp/semantics.cc
>>> index fbbc18336a0..fb4c3992e34 100644
>>> --- a/gcc/cp/semantics.cc
>>> +++ b/gcc/cp/semantics.cc
>>> @@ -53,7 +53,6 @@ along with GCC; see the file COPYING3.  If not see
>>>      static tree maybe_convert_cond (tree);
>>>    static tree finalize_nrv_r (tree *, int *, void *);
>>> -static tree capture_decltype (tree);
>>>      /* Used for OpenMP non-static data member privatization.  */
>>>    @@ -11856,21 +11855,48 @@ finish_decltype_type (tree expr, bool
>>> id_expression_or_member_access_p,
>>>        }
>>>      else
>>>        {
>>> -      /* Within a lambda-expression:
>>> -
>>> -	 Every occurrence of decltype((x)) where x is a possibly
>>> -	 parenthesized id-expression that names an entity of
>>> -	 automatic storage duration is treated as if x were
>>> -	 transformed into an access to a corresponding data member
>>> -	 of the closure type that would have been declared if x
>>> -	 were a use of the denoted entity.  */
>>>          if (outer_automatic_var_p (STRIP_REFERENCE_REF (expr))
>>>    	  && current_function_decl
>>>    	  && LAMBDA_FUNCTION_P (current_function_decl))
>>>    	{
>>> -	  type = capture_decltype (STRIP_REFERENCE_REF (expr));
>>> -	  if (!type)
>>> -	    goto dependent;
>>> +	  /* [expr.prim.id.unqual]/3: If naming the entity from outside of an
>>> +	     unevaluated operand within S would refer to an entity captured by
>>> +	     copy in some intervening lambda-expression, then let E be the
>>> +	     innermost such lambda-expression.
>>> +
>>> +	     If there is such a lambda-expression and if P is in E's function
>>> +	     parameter scope but not its parameter-declaration-clause, then
>>> the
>>> +	     type of the expression is the type of a class member access
>>> +	     expression naming the non-static data member that would be
>>> declared
>>> +	     for such a capture in the object parameter of the function call
>>> +	     operator of E."  */
>>
>> Hmm, looks like this code is only checking the innermost lambda, it needs to
>> check all containing lambdas for one that would capture it by copy.
> 
> Unfortunately this seems to be a can of worms, since IIUC we also have
> to check that there's no non-default-capture lambda in the stack as
> well, e.g.
> 
>    int main() {
>      int x;
>      [] {
>        [=] {
>          using ty1 = decltype((x)); // refers to local variable despite
>                                     // innermost by-copy capture-default
>          using ty1 = int&;
>        };
>      };
>      [=] {
>        [] {
>          using ty1 = decltype((x)); // same
>          using ty1 = int&;
>        };
>      };
>      [=] {
>        [&] {
>          using ty1 = decltype((x)); // refers to hypothetical capture proxy
>          using ty1 = const int&;
>        };
>      };
>      [&] {
>        [=] {
>          using ty1 = decltype((x)); // same
>          using ty1 = const int&;
>        };
>      };
>    }
> 
> And we have to refine the logic for whether to perform the HIDDEN_LAMBDA
> name lookup (which we currently unconditionally do):
> 
>    int main() {
>      int x;
>      [x] {
>         [x] {
>           using ty1 = decltype((x)); // refers to actual capture proxy,
>                                      // found by HIDDEN_LAMBDA name lookup
>           using ty1 = const int&;
>         };
>      };
>      [x] {
>         [] {
>           using ty1 = decltype((x)); // refers to local variable,
>                                      // HIDDEN_LAMBDA name lookup not performed
>           using ty1 = int&;
>         };
>      };
>    }
> 
> These could probably be fixed locally within finish_decltype_type,
> but then there's PR86697 which basically extends all of these
> capture-related issues to 'decltype(f(x))' instead of 'decltype((x))',
> which suggests a proper fix should probably be in process_outer_var_ref
> instead of in finish_decltype_type?  Perhaps when in an unevaluated
> context, process_outer_var_ref should still rewrite uses into capture
> proxies but not actually add them to the closure or something like that?

Or remove them in prune_lambda_captures if there's no real use?

> I don't think I have the cycles to work on these issues this stage..
> Would the latest patch be OK at least?  It seems to be a strict
> improvement.

OK if you open a PR for the other cases and add a FIXME comment 
referring to it.

Jason