Re: C++ PATCH to fix static init with () in a template (PR c++/84582)

[Marek Polacek <polacek at redhat dot com>]

On Wed, Feb 28, 2018 at 04:50:39PM -0500, Jason Merrill wrote:
> On Wed, Feb 28, 2018 at 4:19 PM, Marek Polacek <polacek@redhat.com> wrote:
> > On Wed, Feb 28, 2018 at 10:51:17AM -0500, Jason Merrill wrote:
> >> On Wed, Feb 28, 2018 at 9:32 AM, Marek Polacek <polacek@redhat.com> wrote:
> >> > On Tue, Feb 27, 2018 at 04:16:31PM -0500, Jason Merrill wrote:
> >> >> On 02/27/2018 02:13 PM, Marek Polacek wrote:
> >> >> > My recent change introducing cxx_constant_init caused this code
> >> >> >
> >> >> > template <class> class A {
> >> >> >    static const long b = 0;
> >> >> >    static const unsigned c = (b);
> >> >> > };
> >> >> >
> >> >> > to be rejected.  The reason is that force_paren_expr turns "b" into "*(const
> >> >> > long int &) &b", where the former is not value-dependent but the latter is
> >> >> > value-dependent.  So when we get to maybe_constant_init_1:
> >> >> > 5147   if (!is_nondependent_static_init_expression (t))
> >> >> > 5148     /* Don't try to evaluate it.  */;
> >> >> > it's not evaluated and we get the non-constant initialization error.
> >> >> > (Before we'd always evaluated the expression.)
> >> >> >
> >> >> > Bootstrapped/regtested on x86_64-linux, ok for trunk?
> >> >> >
> >> >> > 2018-02-27  Marek Polacek  <polacek@redhat.com>
> >> >> >
> >> >> >     PR c++/84582
> >> >> >     * semantics.c (force_paren_expr): Avoid creating a static cast
> >> >> >     when processing a template.
> >> >> >
> >> >> >     * g++.dg/cpp1z/static1.C: New test.
> >> >> >     * g++.dg/template/static37.C: New test.
> >> >> >
> >> >> > diff --git gcc/cp/semantics.c gcc/cp/semantics.c
> >> >> > index 35569d0cb0d..b48de2df4e2 100644
> >> >> > --- gcc/cp/semantics.c
> >> >> > +++ gcc/cp/semantics.c
> >> >> > @@ -1697,7 +1697,7 @@ force_paren_expr (tree expr)
> >> >> >       expr = build1 (PAREN_EXPR, TREE_TYPE (expr), expr);
> >> >> >     else if (VAR_P (expr) && DECL_HARD_REGISTER (expr))
> >> >> >       /* We can't bind a hard register variable to a reference.  */;
> >> >> > -  else
> >> >> > +  else if (!processing_template_decl)
> >> >>
> >> >> Hmm, this means that we forget about the parentheses in a template.  I'm
> >> >> surprised that this didn't break anything in the testsuite.  In particular,
> >> >> auto-fn15.C.  I've attached an addition to auto-fn15.C to catch this issue.
> >> >
> >> > Thanks, you're right.  I'll use it.
> >> >
> >> >> Can we use PAREN_EXPR instead of the static_cast in a template?
> >> >
> >> > I don't think so, it would fix the issue you pointed out in auto-fn15.C but
> >> > it wouldn't fix the original test.  The problem with using PAREN_EXPR in a
> >> > template is that instantiate_non_dependent_expr will turn in into the
> >> > static cast anyway; tsubst_copy_and_build has
> >> >     case PAREN_EXPR:
> >> >       RETURN (finish_parenthesized_expr (RECUR (TREE_OPERAND (t, 0))));
> >> > so it calls force_paren_expr and this time we're not in a template.  And
> >> > then when calling cxx_constant_init we have the same issue.
> >>
> >> Then maybe we need something like fold_non_dependent_expr, which
> >> checks for dependency before substitution and then immediately
> >> evaluates the result.
> >
> > I hope you meant something like this.  Further testing also revealed that
> > maybe_undo_parenthesized_ref should be able to unwrap PAREN_EXPR (so that
> > (fn1)(); in paren2.C is handled correctly), and that lvalue_kind should look
> > into PAREN_EXPR so as to give the correct answer regarding lvalueness: we
> > should accept
> >
> > template<typename T>
> > void foo (int i)
> > {
> >   ++(i);
> > }
> >
> > Apologies if I'm on the wrong track.
> >
> > Bootstrapped/regtested on x86_64-linux, ok for trunk?
> >
> > 2018-02-28  Marek Polacek  <polacek@redhat.com>
> >             Jason Merrill  <jason@redhat.com>
> >
> >         PR c++/84582
> >         * semantics.c (force_paren_expr): Avoid creating the static cast
> >         when in a template.  Create a PAREN_EXPR when in a template.
> >         (maybe_undo_parenthesized_ref): Unwrap PAREN_EXPR.
> >         * typeck2.c (store_init_value): Call fold_non_dependent_expr instead
> >         of instantiate_non_dependent_expr.
> >         * tree.c (lvalue_kind): Handle PAREN_EXPR like NON_DEPENDENT_EXPR.
> >
> >         * g++.dg/cpp1y/auto-fn15.C: Extend testing.
> >         * g++.dg/cpp1z/static1.C: New test.
> >         * g++.dg/template/static37.C: New test.
> >
> > diff --git gcc/cp/semantics.c gcc/cp/semantics.c
> > index 35569d0cb0d..722e3718a14 100644
> > --- gcc/cp/semantics.c
> > +++ gcc/cp/semantics.c
> > @@ -1697,7 +1697,7 @@ force_paren_expr (tree expr)
> >      expr = build1 (PAREN_EXPR, TREE_TYPE (expr), expr);
> >    else if (VAR_P (expr) && DECL_HARD_REGISTER (expr))
> >      /* We can't bind a hard register variable to a reference.  */;
> > -  else
> > +  else if (!processing_template_decl)
> >      {
> >        cp_lvalue_kind kind = lvalue_kind (expr);
> >        if ((kind & ~clk_class) != clk_none)
> > @@ -1713,6 +1713,8 @@ force_paren_expr (tree expr)
> >             REF_PARENTHESIZED_P (expr) = true;
> >         }
> >      }
> > +  else
> > +    expr = build1 (PAREN_EXPR, TREE_TYPE (expr), expr);
> 
> There's already a branch for building PAREN_EXPR, let's just replace
> its condition.

Sure.

> > -      value = instantiate_non_dependent_expr (value);
> > +      value = fold_non_dependent_expr (value);
> 
> I was thinking that we want a parallel fold_non_dependent_init (that
> hopefully shares most of the implementation).  Then we shouldn't need
> the call to maybe_constant_init anymore.

If you mean fold_non_dependent_init that would be like fold_non_dependent_expr
but with maybe_constant_init and not maybe_constant_value, then that would
break e.g.

const double d = 9.0;   // missing constexpr
constexpr double j = d; // should give error

because maybe_constant_value checks is_nondependent_constant_expression, and
"d" in the example above is not a constant expression, so we don't evaluate,
and "d" stays "d", so require_constant_expression gives the error.  On the
other hand, maybe_constant_init checks is_nondependent_static_init_expression,
and "d" is that, so we evaluate "d" to "9.0".  Then require_constant_expression
doesn't complain.

What problem do you see with using fold_non_dependent_expr?

Thanks,

	Marek