Re: [C++ PATCH] Fix ICE with inline asm and MODIFY_EXPR/preinc/predec in output operand (PR c++/84961)

[Jason Merrill <jason at redhat dot com>]

On Tue, Mar 20, 2018 at 5:04 PM, Jakub Jelinek <jakub@redhat.com> wrote:
> Hi!
>
> While in C, x = 10 or ++x or --x aren't lvalues and so we reject such
> expressions in inline asm output operands (and inputs that only allow
> memory, not registers), in C++ they apparently are lvalues; for output
> operands we ICE in the gimplifier on this, because in the generic code
> MODIFY_EXPR or PREINCREMENT_EXPR or PREDECREMENT_EXPR aren't considered
> to be lvalues, and for "m" inputs we just reject them, but when those
> expressions are allowed on lhs of a store, they should be IMHO allowed
> as "m" inputs too.
>
> Fixed thusly, bootstrapped/regtested on x86_64-linux and i686-linux, ok for
> trunk?
>
> 2018-03-20  Jakub Jelinek  <jakub@redhat.com>
>
>         PR c++/84961
>         * semantics.c (finish_asm_stmt): Replace MODIFY_EXPR, PREINCREMENT_EXPR
>         and PREDECREMENT_EXPR in output and "m" constrained input operands with
>         COMPOUND_EXPR.  Call cxx_mark_addressable on the rightmost
>         COMPOUND_EXPR operand.
>
>         * c-c++-common/pr43690.c: Don't expect errors on "m" (--x) and
>         "m" (++x) in C++.
>         * g++.dg/torture/pr84961-1.C: New test.
>         * g++.dg/torture/pr84961-2.C: New test.
>
> --- gcc/cp/semantics.c.jj       2018-03-20 11:58:17.069356145 +0100
> +++ gcc/cp/semantics.c  2018-03-20 21:56:43.745292245 +0100
> @@ -1512,6 +1512,26 @@ finish_asm_stmt (int volatile_p, tree st
>                       && C_TYPE_FIELDS_READONLY (TREE_TYPE (operand)))))
>             cxx_readonly_error (operand, lv_asm);
>
> +         tree *op = &operand;
> +         while (TREE_CODE (*op) == COMPOUND_EXPR)
> +           op = &TREE_OPERAND (*op, 1);
> +         switch (TREE_CODE (*op))
> +           {
> +           case PREINCREMENT_EXPR:
> +           case PREDECREMENT_EXPR:
> +           case MODIFY_EXPR:
> +             if (TREE_SIDE_EFFECTS (TREE_OPERAND (*op, 0)))
> +               *op = build2 (TREE_CODE (*op), TREE_TYPE (*op),
> +                             cp_stabilize_reference (TREE_OPERAND (*op, 0)),
> +                             TREE_OPERAND (*op, 1));
> +             *op = build2 (COMPOUND_EXPR, TREE_TYPE (*op), *op,
> +                           TREE_OPERAND (*op, 0));
> +             op = &TREE_OPERAND (*op, 1);
> +             break;
> +           default:
> +             break;
> +           }

Hmm, it would be nice to share this with the similar patterns in
unary_complex_lvalue and cp_build_modify_expr.

Does COND_EXPR work without adjustment?

Jason