[v2] c++: wrong looser excep spec for dep noexcept [PR113158]

On Fri, Feb 16, 2024 at 03:58:02PM -0500, Jason Merrill wrote:
> On 2/15/24 17:17, Marek Polacek wrote:
> > Bootstrapped/regtested on x86_64-pc-linux-gnu, ok for trunk?
> > 
> > By the ??? below I mean that maybe_instantiate_noexcept could return
> > a tristate, and then maybe_check_overriding_exception_spec could check
> > 
> >    if (maybe_instantiate_noexcept ().is_unknown ())
> >      return true;
> > 
> > and we don't have to add any new checks to maybe_check_o_e_spec.
> > 
> > -- >8 --
> > Here we find ourselves in maybe_check_overriding_exception_spec in
> > a template context where we can't instantiate a dependent noexcept.
> > That's OK, but we have to defer the checking otherwise we give wrong
> > errors.
> > 
> > 	PR c++/113158
> > 
> > gcc/cp/ChangeLog:
> > 
> > 	* search.cc (maybe_check_overriding_exception_spec): Defer checking
> > 	when a noexcept couldn't be instantiated.
> > 
> > gcc/testsuite/ChangeLog:
> > 
> > 	* g++.dg/cpp0x/noexcept83.C: New test.
> > ---
> >   gcc/cp/search.cc                        |  7 +++++
> >   gcc/testsuite/g++.dg/cpp0x/noexcept83.C | 37 +++++++++++++++++++++++++
> >   2 files changed, 44 insertions(+)
> >   create mode 100644 gcc/testsuite/g++.dg/cpp0x/noexcept83.C
> > 
> > diff --git a/gcc/cp/search.cc b/gcc/cp/search.cc
> > index c948839dc53..73d254d6b84 100644
> > --- a/gcc/cp/search.cc
> > +++ b/gcc/cp/search.cc
> > @@ -1975,6 +1975,13 @@ maybe_check_overriding_exception_spec (tree overrider, tree basefn)
> >         || UNPARSED_NOEXCEPT_SPEC_P (over_throw))
> >       return true;
> > +  /* We also have to defer checking when we're in a template and couldn't
> > +     instantiate the noexcept yet.
> > +     ??? maybe_instantiate_noexcept already checked these.  Use tristate?  */
> > +  if (type_dependent_expression_p (base_throw)
> > +      || type_dependent_expression_p (over_throw))
> 
> I think we also want to avoid comparing value-dependent expressions, but
> actually checking either one seems like more work than needed here; I'd
> think we want to defer in a template if the specifiers aren't both exactly
> true or false.

Yeah, that'll work too.  So like this?

Bootstrap/regtest running; dg.exp passed.  FWIW, the new check only
triggered on the new test.

Thanks,

-- >8 --
Here we find ourselves in maybe_check_overriding_exception_spec in
a template context where we can't instantiate a dependent noexcept.
That's OK, but we have to defer the checking otherwise we give wrong
errors.

	PR c++/113158

gcc/cp/ChangeLog:

	* search.cc (maybe_check_overriding_exception_spec): Defer checking
	when a noexcept couldn't be instantiated & evaluated to false/true.

gcc/testsuite/ChangeLog:

	* g++.dg/cpp0x/noexcept83.C: New test.
---
 gcc/cp/search.cc                        | 11 ++++++++
 gcc/testsuite/g++.dg/cpp0x/noexcept83.C | 37 +++++++++++++++++++++++++
 2 files changed, 48 insertions(+)
 create mode 100644 gcc/testsuite/g++.dg/cpp0x/noexcept83.C


base-commit: 40b8d7b73ad2ce498758c1d9bd38ebdbc26b918b

On Fri, 16 Feb 2024, Marek Polacek wrote:

> On Fri, Feb 16, 2024 at 03:58:02PM -0500, Jason Merrill wrote:
> > On 2/15/24 17:17, Marek Polacek wrote:
> > > Bootstrapped/regtested on x86_64-pc-linux-gnu, ok for trunk?
> > > 
> > > By the ??? below I mean that maybe_instantiate_noexcept could return
> > > a tristate, and then maybe_check_overriding_exception_spec could check
> > > 
> > >    if (maybe_instantiate_noexcept ().is_unknown ())
> > >      return true;
> > > 
> > > and we don't have to add any new checks to maybe_check_o_e_spec.
> > > 
> > > -- >8 --
> > > Here we find ourselves in maybe_check_overriding_exception_spec in
> > > a template context where we can't instantiate a dependent noexcept.
> > > That's OK, but we have to defer the checking otherwise we give wrong
> > > errors.
> > > 
> > > 	PR c++/113158
> > > 
> > > gcc/cp/ChangeLog:
> > > 
> > > 	* search.cc (maybe_check_overriding_exception_spec): Defer checking
> > > 	when a noexcept couldn't be instantiated.
> > > 
> > > gcc/testsuite/ChangeLog:
> > > 
> > > 	* g++.dg/cpp0x/noexcept83.C: New test.
> > > ---
> > >   gcc/cp/search.cc                        |  7 +++++
> > >   gcc/testsuite/g++.dg/cpp0x/noexcept83.C | 37 +++++++++++++++++++++++++
> > >   2 files changed, 44 insertions(+)
> > >   create mode 100644 gcc/testsuite/g++.dg/cpp0x/noexcept83.C
> > > 
> > > diff --git a/gcc/cp/search.cc b/gcc/cp/search.cc
> > > index c948839dc53..73d254d6b84 100644
> > > --- a/gcc/cp/search.cc
> > > +++ b/gcc/cp/search.cc
> > > @@ -1975,6 +1975,13 @@ maybe_check_overriding_exception_spec (tree overrider, tree basefn)
> > >         || UNPARSED_NOEXCEPT_SPEC_P (over_throw))
> > >       return true;
> > > +  /* We also have to defer checking when we're in a template and couldn't
> > > +     instantiate the noexcept yet.
> > > +     ??? maybe_instantiate_noexcept already checked these.  Use tristate?  */
> > > +  if (type_dependent_expression_p (base_throw)
> > > +      || type_dependent_expression_p (over_throw))
> > 
> > I think we also want to avoid comparing value-dependent expressions, but
> > actually checking either one seems like more work than needed here; I'd
> > think we want to defer in a template if the specifiers aren't both exactly
> > true or false.
> 
> Yeah, that'll work too.  So like this?
> 
> Bootstrap/regtest running; dg.exp passed.  FWIW, the new check only
> triggered on the new test.
> 
> Thanks,
> 
> -- >8 --
> Here we find ourselves in maybe_check_overriding_exception_spec in
> a template context where we can't instantiate a dependent noexcept.
> That's OK, but we have to defer the checking otherwise we give wrong
> errors.
> 
> 	PR c++/113158
> 
> gcc/cp/ChangeLog:
> 
> 	* search.cc (maybe_check_overriding_exception_spec): Defer checking
> 	when a noexcept couldn't be instantiated & evaluated to false/true.
> 
> gcc/testsuite/ChangeLog:
> 
> 	* g++.dg/cpp0x/noexcept83.C: New test.
> ---
>  gcc/cp/search.cc                        | 11 ++++++++
>  gcc/testsuite/g++.dg/cpp0x/noexcept83.C | 37 +++++++++++++++++++++++++
>  2 files changed, 48 insertions(+)
>  create mode 100644 gcc/testsuite/g++.dg/cpp0x/noexcept83.C
> 
> diff --git a/gcc/cp/search.cc b/gcc/cp/search.cc
> index c948839dc53..554ba71f4a7 100644
> --- a/gcc/cp/search.cc
> +++ b/gcc/cp/search.cc
> @@ -1975,6 +1975,17 @@ maybe_check_overriding_exception_spec (tree overrider, tree basefn)
>        || UNPARSED_NOEXCEPT_SPEC_P (over_throw))
>      return true;
>  
> +  /* We also have to defer checking when we're in a template and couldn't
> +     instantiate & evaluate the noexcept to true/false.  */
> +  if (processing_template_decl)
> +    if ((base_throw
> +	 && (base_throw != noexcept_true_spec
> +	     || base_throw != noexcept_false_spec))

Shouldn't these innermost || be &&?

> +	|| (over_throw
> +	    && (over_throw != noexcept_true_spec
> +		|| over_throw != noexcept_false_spec)))

> +      return true;
> +
>    if (!comp_except_specs (base_throw, over_throw, ce_derived))
>      {
>        auto_diagnostic_group d;
> diff --git a/gcc/testsuite/g++.dg/cpp0x/noexcept83.C b/gcc/testsuite/g++.dg/cpp0x/noexcept83.C
> new file mode 100644
> index 00000000000..47832bbb44d
> --- /dev/null
> +++ b/gcc/testsuite/g++.dg/cpp0x/noexcept83.C
> @@ -0,0 +1,37 @@
> +// PR c++/113158
> +// { dg-do compile { target c++11 } }
> +
> +template<typename T>
> +struct V {
> +  static constexpr bool t = false;
> +};
> +struct base {
> +    virtual int f() = 0;
> +};
> +
> +template<typename T>
> +struct derived : base {
> +    int f() noexcept(V<T>::t) override;
> +};
> +
> +struct base2 {
> +    virtual int f() noexcept = 0;
> +};
> +
> +template<bool B>
> +struct W {
> +  static constexpr bool t = B;
> +};
> +
> +template<bool B>
> +struct derived2 : base2 {
> +    int f() noexcept(W<B>::t) override; // { dg-error "looser exception specification" }
> +};
> +
> +void
> +g ()
> +{
> +  derived<int> d1;
> +  derived2<false> d2; // { dg-message "required from here" }
> +  derived2<true> d3;
> +}
> 
> base-commit: 40b8d7b73ad2ce498758c1d9bd38ebdbc26b918b
> -- 
> 2.43.2
> 
>