On 11/8/19 9:14 AM, Jakub Jelinek wrote:
Hi!We ICE on the following testcase, because DECL_INITIAL (decl) is error_mark_node due to previously reported error and cxx_eval_outermost_constant_expr is unhappy if ctx.ctor is not a CONSTRUCTOR, but error_mark_node. If the initializer is invalid, it should have been diagnosed already and there is no need to try to evaluate constexpr dtor on it. Bootstrapped/regtested on x86_64-linux and i686-linux, ok for trunk? 2019-11-08 Jakub Jelinek <ja...@redhat.com> PR c++/92414 * constexpr.c (cxx_constant_dtor): Don't call cxx_eval_outermost_constant_expr if DECL_INITIAL is erroneous.
Why add the error-checking here rather than in cxx_eval_outermost_constant_expr when we use DECL_INITIAL?
* g++.dg/cpp2a/constexpr-dtor4.C: New test. --- gcc/cp/constexpr.c.jj 2019-11-06 08:58:37.000000000 +0100 +++ gcc/cp/constexpr.c 2019-11-07 22:13:58.395840756 +0100 @@ -6024,7 +6024,8 @@ cxx_constant_value (tree t, tree decl) void cxx_constant_dtor (tree t, tree decl) { - cxx_eval_outermost_constant_expr (t, false, true, true, true, decl); + if (!error_operand_p (DECL_INITIAL (decl))) + cxx_eval_outermost_constant_expr (t, false, true, true, true, decl); }/* Helper routine for fold_simple function. Either return simplified--- gcc/testsuite/g++.dg/cpp2a/constexpr-dtor4.C.jj 2019-11-07 22:16:56.943181785 +0100 +++ gcc/testsuite/g++.dg/cpp2a/constexpr-dtor4.C 2019-11-07 22:18:16.752993443 +0100 @@ -0,0 +1,15 @@ +// PR c++/92414 +// { dg-do compile { target c++2a } } + +struct A { virtual void foo (); }; + +struct B : A { + constexpr B (int); // { dg-warning "used but never defined" } + constexpr ~B () { } +}; + +struct D : B { + constexpr D () : B (42) { } // { dg-error "used before its definition" } +}; + +constexpr D d; // { dg-message "in 'constexpr' expansion of" } Jakub
