public inbox for gcc-patches@gcc.gnu.org
 help / color / mirror / Atom feed
From: Jason Merrill <jason@redhat.com>
To: Patrick Palka <ppalka@redhat.com>
Cc: gcc-patches@gcc.gnu.org
Subject: Re: [PATCH] c++: explicit spec of constrained member tmpl [PR107522]
Date: Fri, 2 Dec 2022 11:16:41 -0500	[thread overview]
Message-ID: <10b990d6-e1a1-2cbe-19f8-6f59d4e22130@redhat.com> (raw)
In-Reply-To: <fee2c84f-1609-3cae-9f4e-2d1d7d2e872d@idea>

On 12/2/22 09:30, Patrick Palka wrote:
> On Thu, 1 Dec 2022, Jason Merrill wrote:
> 
>> On 12/1/22 14:51, Patrick Palka wrote:
>>> On Thu, 1 Dec 2022, Jason Merrill wrote:
>>>
>>>> On 12/1/22 11:37, Patrick Palka wrote:
>>>>> When defining a explicit specialization of a constrained member template
>>>>> (of a class template) such as f and g in the below testcase, the
>>>>> DECL_TEMPLATE_PARMS of the corresponding TEMPLATE_DECL are partially
>>>>> instantiated, whereas its associated constraints are carried over
>>>>> from the original template and thus are in terms of the original
>>>>> DECL_TEMPLATE_PARMS.
>>>>
>>>> But why are they carried over?  We wrote a specification of the
>>>> constraints in
>>>> terms of the template parameters of the specialization, why are we
>>>> throwing
>>>> that away?
>>>
>>> Using the partially instantiated constraints would require adding a
>>> special case to satisfaction since during satisfaction we currently
>>> always use the full set of template arguments (relative to the most
>>> general template).
>>
>> But not for partial specializations, right?  It seems natural to handle this
>> explicit instantiation the way we handle partial specializations, as both have
>> their constraints written in terms of their template parameters.
> 
> True, but what about the general rule that we don't partially instantiate
> constraints outside of declaration matching?  Checking satisfaction of
> partially instantiated constraints here can introduce hard errors during
> normalization, e.g.
> 
>    template<class T>
>    concept C1 = __same_as(T, void);
> 
>    template<class T>
>    concept C2 = C1<typename T::type>;
> 
>    template<int N>
>    concept D = (N == 42);
> 
>    template<class T>
>    struct A {
>      template<int N>
>      static void f() requires C2<T> || D<N>;
>    };
> 
>    template<>
>    template<int N>
>    void A<int>::f() requires C2<int> || D<N> { }
> 
>    int main() {
>      A<int>::f<42>();
>    }
> 
> Normalization of the the partially instantiated constraints will give a
> hard error due to 'int::type' being ill-formed, whereas the uninstantiated
> constraints are fine.

Hmm, interesting point, but in this example that happens because the 
specialization is nonsensical: we wouldn't be normalizing the 
partially-instantiated constraints so much as the ones that the user 
explicitly wrote, so a hard error seems justified.

>>> For satisfaction of the partially instantiated
>>> constraints, we'd instead have to use the template arguments relative to
>>> the explicit specialization, e.g. {42} instead of {{int},{42}} for
>>> A<int>::f<42>.  Not sure if that would be preferable, but it seems
>>> doable.
>>>
>>>>
>>>>> So during normalization for such an explicit
>>>>> specialization we need to consider the (parameters of) the most general
>>>>> template, since that's what the constraints are in terms of and since we
>>>>> always use the full set of template arguments during satisfaction.
>>>>>
>>>>> Bootstrapped and regtested on x86_64-pc-linux-gnu, does this look OK for
>>>>> trunk and perhaps 12?
>>>>>
>>>>> 	PR c++/107522
>>>>>
>>>>> gcc/cp/ChangeLog:
>>>>>
>>>>> 	* constraint.cc (get_normalized_constraints_from_decl): Use the
>>>>> 	most general template for an explicit specialization of a
>>>>> 	member template.
>>>>>
>>>>> gcc/testsuite/ChangeLog:
>>>>>
>>>>> 	* g++.dg/cpp2a/concepts-explicit-spec7.C: New test.
>>>>> ---
>>>>>     gcc/cp/constraint.cc                          | 18 ++++++++---
>>>>>     .../g++.dg/cpp2a/concepts-explicit-spec7.C    | 31
>>>>> +++++++++++++++++++
>>>>>     2 files changed, 44 insertions(+), 5 deletions(-)
>>>>>     create mode 100644
>>>>> gcc/testsuite/g++.dg/cpp2a/concepts-explicit-spec7.C
>>>>>
>>>>> diff --git a/gcc/cp/constraint.cc b/gcc/cp/constraint.cc
>>>>> index ab0f66b3d7e..f1df84c2a1c 100644
>>>>> --- a/gcc/cp/constraint.cc
>>>>> +++ b/gcc/cp/constraint.cc
>>>>> @@ -973,11 +973,19 @@ get_normalized_constraints_from_decl (tree d, bool
>>>>> diag = false)
>>>>>          accepting the latter causes the template parameter level of U
>>>>>          to be reduced in a way that makes it overly difficult substitute
>>>>>          concrete arguments (i.e., eventually {int, int} during
>>>>> satisfaction.
>>>>> */
>>>>> -  if (tmpl)
>>>>> -  {
>>>>> -    if (DECL_LANG_SPECIFIC(tmpl) && !DECL_TEMPLATE_SPECIALIZATION
>>>>> (tmpl))
>>>>> -      tmpl = most_general_template (tmpl);
>>>>> -  }
>>>>> +  if (tmpl && DECL_LANG_SPECIFIC (tmpl)
>>>>> +      && (!DECL_TEMPLATE_SPECIALIZATION (tmpl)
>>>>> +	  /* DECL_TEMPLATE_SPECIALIZATION means we're dealing with either a
>>>>> +	     partial specialization or an explicit specialization of a member
>>>>> +	     template.  In the former case all is well: the constraints are in
>>>>> +	     terms in TMPL's parameters.  But in the latter case TMPL's
>>>>> +	     parameters are partially instantiated whereas its constraints
>>>>> +	     aren't, so we need to consider (the parameters of) the most
>>>>> +	     general template.  The following test distinguishes between a
>>>>> +	     partial specialization and such an explicit specialization.  */
>>>>> +	  || (TMPL_PARMS_DEPTH (DECL_TEMPLATE_PARMS (tmpl))
>>>>> +	      < TMPL_ARGS_DEPTH (DECL_TI_ARGS (tmpl)))))
>>>>> +    tmpl = most_general_template (tmpl);
>>>>>         d = tmpl ? tmpl : decl;
>>>>>     diff --git a/gcc/testsuite/g++.dg/cpp2a/concepts-explicit-spec7.C
>>>>> b/gcc/testsuite/g++.dg/cpp2a/concepts-explicit-spec7.C
>>>>> new file mode 100644
>>>>> index 00000000000..5b5a6df20ff
>>>>> --- /dev/null
>>>>> +++ b/gcc/testsuite/g++.dg/cpp2a/concepts-explicit-spec7.C
>>>>> @@ -0,0 +1,31 @@
>>>>> +// PR c++/107522
>>>>> +// { dg-do compile { target c++20 } }
>>>>> +
>>>>> +template<class T>
>>>>> +struct A
>>>>> +{
>>>>> +  template<int N>
>>>>> +  static void f() requires (N == 42);
>>>>> +
>>>>> +  template<class U>
>>>>> +  struct B {
>>>>> +    template<int N>
>>>>> +    static void g() requires (T(N) == 42);
>>>>> +  };
>>>>> +};
>>>>> +
>>>>> +template<>
>>>>> +template<int N>
>>>>> +void A<int>::f() requires (N == 42) { }
>>>>> +
>>>>> +template<>
>>>>> +template<>
>>>>> +template<int N>
>>>>> +void A<int>::B<int>::g() requires (int(N) == 42) { }
>>>>> +
>>>>> +int main() {
>>>>> +  A<int>::f<42>();
>>>>> +  A<int>::f<43>(); // { dg-error "no match" }
>>>>> +  A<int>::B<int>::g<42>();
>>>>> +  A<int>::B<int>::g<43>(); // { dg-error "no match" }
>>>>> +}
>>>>
>>>>
>>>
>>
>>
> 


      reply	other threads:[~2022-12-02 16:16 UTC|newest]

Thread overview: 6+ messages / expand[flat|nested]  mbox.gz  Atom feed  top
2022-12-01 16:37 Patrick Palka
2022-12-01 19:15 ` Jason Merrill
2022-12-01 19:51   ` Patrick Palka
2022-12-01 21:17     ` Jason Merrill
2022-12-02 14:30       ` Patrick Palka
2022-12-02 16:16         ` Jason Merrill [this message]

Reply instructions:

You may reply publicly to this message via plain-text email
using any one of the following methods:

* Save the following mbox file, import it into your mail client,
  and reply-to-all from there: mbox

  Avoid top-posting and favor interleaved quoting:
  https://en.wikipedia.org/wiki/Posting_style#Interleaved_style

* Reply using the --to, --cc, and --in-reply-to
  switches of git-send-email(1):

  git send-email \
    --in-reply-to=10b990d6-e1a1-2cbe-19f8-6f59d4e22130@redhat.com \
    --to=jason@redhat.com \
    --cc=gcc-patches@gcc.gnu.org \
    --cc=ppalka@redhat.com \
    /path/to/YOUR_REPLY

  https://kernel.org/pub/software/scm/git/docs/git-send-email.html

* If your mail client supports setting the In-Reply-To header
  via mailto: links, try the mailto: link
Be sure your reply has a Subject: header at the top and a blank line before the message body.
This is a public inbox, see mirroring instructions
for how to clone and mirror all data and code used for this inbox;
as well as URLs for read-only IMAP folder(s) and NNTP newsgroup(s).