c++: ICE with invalid structured bindings [PR107276]
authorMarek Polacek <polacek@redhat.com>
Thu, 20 Oct 2022 19:55:28 +0000 (15:55 -0400)
committerMarek Polacek <polacek@redhat.com>
Mon, 24 Oct 2022 18:02:17 +0000 (14:02 -0400)
commitf7d8ccfda2d5c90dac97b1a3ede8b10391a3cc40
treeeefd4617d0158fc6e2285e20aa769c84432df5e9
parent65e3274e363cb2c6bfe6b5e648916eb7696f7e2f
c++: ICE with invalid structured bindings [PR107276]

This test ICEs in C++23 because we reach the new code in do_auto_deduction:

30468   if (cxx_dialect >= cxx23
30469       && context == adc_return_type
30470       && (!AUTO_IS_DECLTYPE (auto_node)
30471           || !unparenthesized_id_or_class_member_access_p (init))
30472       && (r = treat_lvalue_as_rvalue_p (maybe_undo_parenthesized_ref (init),
30473                                         /*return*/true)))

where 'init' is "VIEW_CONVERT_EXPR<<<< error >>>>(y)", and then the move
in treat_lvalue_as_rvalue_p returns error_mark_node whereupon
set_implicit_rvalue_p crashes.

I don't think such V_C_Es are useful so let's not create them.  But that
won't fix the ICE so I'm checking the return value of move.  A structured
bindings decl can have an error type, that is set in cp_finish_decomp:

 8908           TREE_TYPE (first) = error_mark_node;

therefore I think treat_lvalue_as_rvalue_p just needs to cope.

PR c++/107276

gcc/cp/ChangeLog:

* typeck.cc (treat_lvalue_as_rvalue_p): Check the return value of move.

gcc/ChangeLog:

* tree.cc (maybe_wrap_with_location): Don't create a location wrapper
when the type is erroneous.

gcc/testsuite/ChangeLog:

* g++.dg/cpp2a/decomp4.C: New test.
gcc/cp/typeck.cc
gcc/testsuite/g++.dg/cpp2a/decomp4.C [new file with mode: 0644]
gcc/tree.cc