On 6/11/19 3:59 PM, Marek Polacek wrote:
On Tue, Jun 11, 2019 at 03:05:26PM -0400, Jason Merrill wrote:
On 6/11/19 2:28 PM, Marek Polacek wrote:
On Tue, Jun 11, 2019 at 08:37:27AM -0400, Jason Merrill wrote:
On 6/11/19 7:47 AM, Jakub Jelinek wrote:
On Mon, Jun 10, 2019 at 09:59:46PM -0400, Marek Polacek wrote:
This test segvs since r269078, this hunk in particular:

@@ -4581,8 +4713,9 @@ cxx_eval_constant_expression (const constexpr_ctx *ctx, 
tree t,
          break;

        case SIZEOF_EXPR:
-      r = fold_sizeof_expr (t);
-      VERIFY_CONSTANT (r);
+      r = cxx_eval_constant_expression (ctx, fold_sizeof_expr (t), lval,
+                   non_constant_p, overflow_p,
+                   jump_target);
          break;

In a template, fold_sizeof_expr will just create a new SIZEOF_EXPR, that is the
same, but not identical; see cxx_sizeof_expr.  Then cxx_eval_constant_expression

Not always, if it calls cxx_sizeof_expr, it will, but if it calls
cxx_sizeof_or_alignof_type it will only if the type is dependent or VLA.

So, I'd think you should call cxx_eval_constant_expression if TREE_CODE (r)
!= SIZEOF_EXPR, otherwise probably *non_constant_p = true; is in order,
maybe together with gcc_assert (ctx->quiet); ?  I'd hope that if we really
require a constant expression we evaluate it in !processing_template_decl
contexts.

Ok, I had been meaning to add the *non_constant_p bit but never did.  :(

Makes sense.  Also, cxx_sizeof_expr should probably only return a
SIZEOF_EXPR if the operand is instantiation-dependent.

That results in

    FAIL: g++.dg/template/incomplete6.C  -std=c++98 (internal compiler error)
    FAIL: g++.dg/template/incomplete6.C  -std=c++98 (test for excess errors)
    FAIL: g++.dg/template/overload13.C  -std=c++98 (internal compiler error)
    FAIL: g++.dg/template/overload13.C  -std=c++98 (test for excess errors)

because we trigger an assert in value_dependent_expression_p:

              /* If there are no operands, it must be an expression such
                 as "int()". This should not happen for aggregate types
                 because it would form non-constant expressions.  */
              gcc_assert (cxx_dialect >= cxx11
                          || INTEGRAL_OR_ENUMERATION_TYPE_P (type));

              return false;

and in this case we have T() where T is a class, and it's in C++98.

It's not needed to fix this PR so perhaps the following could go in,
but is there anything I should do about that?

instantiation_dependent_uneval_expression_p shouldn't have that problem.

Ah, nice.

diff --git gcc/cp/constexpr.c gcc/cp/constexpr.c
index a2f29694462..443e1c7899f 100644
--- gcc/cp/constexpr.c
+++ gcc/cp/constexpr.c
@@ -4808,9 +4808,16 @@ cxx_eval_constant_expression (const constexpr_ctx *ctx, 
tree t,
         break;
       case SIZEOF_EXPR:
-      r = cxx_eval_constant_expression (ctx, fold_sizeof_expr (t), lval,
-                                       non_constant_p, overflow_p,
-                                       jump_target);
+      r = fold_sizeof_expr (t);
+      /* In a template, fold_sizeof_expr may merely create a new SIZEOF_EXPR,
+        which could lead to an infinite recursion.  */
+      if (TREE_CODE (r) != SIZEOF_EXPR)
+       r = cxx_eval_constant_expression (ctx, r, lval,
+                                         non_constant_p, overflow_p,
+                                         jump_target);
+      else
+       *non_constant_p = true;

Let's also add the assert Jakub suggested.

Done.

Luckily, this also fixed c++/90832 so I'm adding a test for that, too.

Bootstrapped/regtested on x86_64-linux, ok for trunk?

OK.

Jason

Reply via email to