Re: [C++ PATCH] Fix __builtin_constant_p constexpr handling (PR inline-asm/85172)

[Jason Merrill <jason at redhat dot com>]

OK.

On Wed, Apr 4, 2018 at 3:39 PM, Jakub Jelinek <jakub@redhat.com> wrote:
> Hi!
>
> As the following testcases show, potential_constant_expression_1 for some
> builtins returns true no matter what their arguments contain (intentionally
> so); the problem is that we call unconditionally
> cxx_eval_constant_expression on those arguments and that creates a loophole;
> normally potential_constant_expression_1 is a check what kind of expressions
> we allow in and cxx_eval* can only handle what we accept by that; through
> these builtins, anything else can appear there too.
> The following patch checks if the expression is potential constant
> expression before trying to call cxx_eval_constant_expression, but to avoid
> exponential compile time it checks it only for those builtins where
> potential_constant_expression_1 has not checked those arguments.
>
> Bootstrapped/regtested on x86_64-linux and i686-linux, ok for trunk?
>
> 2018-04-04  Jakub Jelinek  <jakub@redhat.com>
>
>         PR inline-asm/85172
>         * constexpr.c (cxx_eval_builtin_function_call): For calls to
>         builtin_valid_in_constant_expr_p functions, don't call
>         cxx_eval_constant_expression if argument is not
>         potential_constant_expression.
>
>         * g++.dg/ext/builtin13.C: New test.
>         * g++.dg/ext/atomic-4.C: New test.
>
> --- gcc/cp/constexpr.c.jj       2018-04-03 23:39:16.535665285 +0200
> +++ gcc/cp/constexpr.c  2018-04-04 12:25:32.290813343 +0200
> @@ -1189,8 +1189,14 @@ cxx_eval_builtin_function_call (const co
>    bool dummy1 = false, dummy2 = false;
>    for (i = 0; i < nargs; ++i)
>      {
> -      args[i] = cxx_eval_constant_expression (&new_ctx, CALL_EXPR_ARG (t, i),
> -                                             false, &dummy1, &dummy2);
> +      args[i] = CALL_EXPR_ARG (t, i);
> +      /* If builtin_valid_in_constant_expr_p is true,
> +        potential_constant_expression_1 has not recursed into the arguments
> +        of the builtin, verify it here.  */
> +      if (!builtin_valid_in_constant_expr_p (fun)
> +         || potential_constant_expression (args[i]))
> +       args[i] = cxx_eval_constant_expression (&new_ctx, args[i], false,
> +                                               &dummy1, &dummy2);
>        if (bi_const_p)
>         /* For __built_in_constant_p, fold all expressions with constant values
>            even if they aren't C++ constant-expressions.  */
> --- gcc/testsuite/g++.dg/ext/builtin13.C.jj     2018-04-04 12:11:03.566767129 +0200
> +++ gcc/testsuite/g++.dg/ext/builtin13.C        2018-04-04 12:11:34.380768131 +0200
> @@ -0,0 +1,9 @@
> +// PR inline-asm/85172
> +// { dg-do compile }
> +// { dg-options "" }
> +
> +int
> +foo ()
> +{
> +  return !__builtin_constant_p (({ __asm (""); 0; }));
> +}
> --- gcc/testsuite/g++.dg/ext/atomic-4.C.jj      2018-04-04 12:10:54.239766822 +0200
> +++ gcc/testsuite/g++.dg/ext/atomic-4.C 2018-04-04 12:11:56.058768833 +0200
> @@ -0,0 +1,9 @@
> +// PR inline-asm/85172
> +// { dg-do compile }
> +// { dg-options "" }
> +
> +int
> +foo (int *p)
> +{
> +  return !__atomic_always_lock_free (4, ({ __asm (""); p; }));
> +}
>
>         Jakub