Re: [C++ PATCH] Fix ICE with return in statement expression in constexpr.c (PR c++/84192)

[Jason Merrill <jason at redhat dot com>]

OK.

On Fri, Feb 16, 2018 at 10:19 AM, Jakub Jelinek <jakub@redhat.com> wrote:
> On Fri, Feb 16, 2018 at 08:52:10AM -0500, Jason Merrill wrote:
>> On Fri, Feb 16, 2018 at 3:28 AM, Jakub Jelinek <jakub@redhat.com> wrote:
>> > pop_stmt_list, if there is just a single stmt inside statement expression
>> > moves the stmt out of the STATEMENT_LIST (and I think cp_fold does too).
>> > We only initialize jump_target to non-NULL in cxx_eval_statement_list
>> > or for calls, so before we have a chance to diagnose the error of using
>> > an expression with void type, we ICE trying to dereference NULL jump_target.
>> >
>> > This can't happen with BREAK_STMT nor CONTINUE_STMT, because they are not
>> > potential constant expressions, and I think can only happen when ctx->quiet
>> > is true, otherwise it should have been diagnosed already before.
>> > If a RETURN_EXPR (or in theory break/continue) appears in a (potential) constant
>> > expression we want to evaluate, not doing anything with jump_target if we
>> > aren't inside a statement list makes sense to me, there is no following
>> > statement to bypass.
>>
>> I think we should also set *non_constant_p.
>
> Just like this?  Tested so far just on the testcase, but given that we'd ICE
> on the *jump_target before, it can't really regress anything else (though of
> course I'll bootstrap/regtest it normally).
>
> 2018-02-16  Marek Polacek  <polacek@redhat.com>
>             Jakub Jelinek  <jakub@redhat.com>
>
>         PR c++/84192
>         * constexpr.c (cxx_eval_constant_expression) <case RETURN_EXPR>: Don't
>         set *jump_target to anything if jump_target is NULL.
>
>         * g++.dg/cpp1y/constexpr-84192.C: New test.
>
> --- gcc/cp/constexpr.c.jj       2018-02-12 19:17:37.937216029 +0100
> +++ gcc/cp/constexpr.c  2018-02-15 16:10:56.630572360 +0100
> @@ -4254,7 +4254,16 @@ cxx_eval_constant_expression (const cons
>         r = cxx_eval_constant_expression (ctx, TREE_OPERAND (t, 0),
>                                           lval,
>                                           non_constant_p, overflow_p);
> -      *jump_target = t;
> +      if (jump_target)
> +       *jump_target = t;
> +      else
> +       {
> +         /* Can happen with ({ return true; }) && false; passed to
> +            maybe_constant_value.  There is nothing to jump over in this
> +            case, and the bug will be diagnosed later.  */
> +         gcc_assert (ctx->quiet);
> +         *non_constant_p = true;
> +       }
>        break;
>
>      case SAVE_EXPR:
> --- gcc/testsuite/g++.dg/cpp1y/constexpr-84192.C.jj     2018-02-15 16:00:58.242588914 +0100
> +++ gcc/testsuite/g++.dg/cpp1y/constexpr-84192.C        2018-02-15 16:01:30.219585291 +0100
> @@ -0,0 +1,41 @@
> +// PR c++/84192
> +// { dg-do compile { target c++14 } }
> +// { dg-options "" }
> +
> +bool
> +f1 ()
> +{
> +  return ({ return true; }) && false;  // { dg-error "could not convert" }
> +}
> +
> +void
> +f2 ()
> +{
> +  for (;;)
> +    constexpr bool b = ({ break; false; }) && false;   // { dg-error "statement is not a constant expression" }
> +}
> +
> +constexpr bool
> +f3 (int n)
> +{
> +  bool b = false;
> +  for (int i = 0; i < n; i++)
> +    b = ({ break; });  // { dg-error "void value not ignored as it ought to be" }
> +  return b;
> +}
> +
> +constexpr bool b = f3 (4);
> +
> +bool
> +f4 ()
> +{
> +  constexpr bool b = ({ return true; }) && false;      // { dg-error "could not convert" }
> +  return false;
> +}
> +
> +constexpr bool
> +f5 (int x)
> +{
> +  constexpr bool b = ({ switch (x) case 0: true; }) && false;  // { dg-error "could not convert" }
> +  return false;
> +}
>
>
>         Jakub