diff mbox

Fix PR c++/70332 (ICE due to aggregate initialization of NSDMI)

Message ID 1458682553-20882-1-git-send-email-patrick@parcs.ath.cx
State New
Headers show

Commit Message

Patrick Palka March 22, 2016, 9:35 p.m. UTC 
With c++14 an NSDMI no longer makes a class type non-aggregate so it's
possible to perform aggregate initialization on a class that has an
NSDMI, but tsubst_copy() currently ICEs on a use of 'this' in such
a situation.

This patch makes tsubst_copy() handle a use of 'this' in an NSDMI as
part of an aggregate initialization.  In that case current_class_ref
will be a PLACEHOLDER_EXPR (as set by get_nsdmi()) and this
PLACEHOLDER_EXPR will later get resolved to the true object by
replace_placeholders().

Does this patch look OK to commit after testing?

gcc/cp/ChangeLog:

	PR c++/70332
	* pt.c (tsubst_copy) [PARM_DECL]: Handle the use of 'this' in an
	NSDMI that's part of an aggregrate initialization.

gcc/testsuite/ChangeLog:

	PR c++/70332
	* g++.dg/cpp1y/nsdmi-aggr5.C: New test.
---
 gcc/cp/pt.c                              | 20 ++++++++++++++++----
 gcc/testsuite/g++.dg/cpp1y/nsdmi-aggr5.C | 24 ++++++++++++++++++++++++
 2 files changed, 40 insertions(+), 4 deletions(-)
 create mode 100644 gcc/testsuite/g++.dg/cpp1y/nsdmi-aggr5.C

Comments

Jason Merrill March 22, 2016, 10 p.m. UTC | #1 
On 03/22/2016 05:35 PM, Patrick Palka wrote:
> +	      if (cp_unevaluated_operand == 0

Why check this here?

Jason
Patrick Palka March 22, 2016, 10:12 p.m. UTC | #2 
On Tue, Mar 22, 2016 at 6:00 PM, Jason Merrill <jason@redhat.com> wrote:
> On 03/22/2016 05:35 PM, Patrick Palka wrote:
>>
>> +             if (cp_unevaluated_operand == 0
>
>
> Why check this here?

Just so that the change doesn't affect the behavior of tsubst_decl()
when cp_unevaluated_operand != 0.  Presumably the existing code (10
lines below) handles that case just fine.
Patrick Palka March 22, 2016, 11:12 p.m. UTC | #3 
On Tue, Mar 22, 2016 at 6:12 PM, Patrick Palka <patrick@parcs.ath.cx> wrote:
> On Tue, Mar 22, 2016 at 6:00 PM, Jason Merrill <jason@redhat.com> wrote:
>> On 03/22/2016 05:35 PM, Patrick Palka wrote:
>>>
>>> +             if (cp_unevaluated_operand == 0
>>
>>
>> Why check this here?
>
> Just so that the change doesn't affect the behavior of tsubst_decl()
> when cp_unevaluated_operand != 0.  Presumably the existing code (10
> lines below) handles that case just fine.

Turns out that without the check we can trigger the cxx_dialect >=
cxx14 assert because in c++11 mode we can reach the assert through
get_defaulted_eh_spec() which increments cp_unevaluated_operand and
then calls get_nsdmi (..., /*in_ctor=*/false) causing
current_class_ref to get set to a PLACEHOLDER_EXPR.

So for example g++.dg/cpp0x/nsdmi-template2.C regresses with an ICE.
So it seems the cp_unevaluated_operand != 0 check is necessary as long
as the assert stays.

There are no regressions if both the cp_unevaluated_operand check and
the assert are removed however.
Jason Merrill March 23, 2016, 12:58 p.m. UTC | #4 
On 03/22/2016 07:12 PM, Patrick Palka wrote:
> On Tue, Mar 22, 2016 at 6:12 PM, Patrick Palka <patrick@parcs.ath.cx> wrote:
>> On Tue, Mar 22, 2016 at 6:00 PM, Jason Merrill <jason@redhat.com> wrote:
>>> On 03/22/2016 05:35 PM, Patrick Palka wrote:
>>>>
>>>> +             if (cp_unevaluated_operand == 0
>>>
>>>
>>> Why check this here?
>>
>> Just so that the change doesn't affect the behavior of tsubst_decl()
>> when cp_unevaluated_operand != 0.  Presumably the existing code (10
>> lines below) handles that case just fine.
>
> Turns out that without the check we can trigger the cxx_dialect >=
> cxx14 assert because in c++11 mode we can reach the assert through
> get_defaulted_eh_spec() which increments cp_unevaluated_operand and
> then calls get_nsdmi (..., /*in_ctor=*/false) causing
> current_class_ref to get set to a PLACEHOLDER_EXPR.
>
> So for example g++.dg/cpp0x/nsdmi-template2.C regresses with an ICE.
> So it seems the cp_unevaluated_operand != 0 check is necessary as long
> as the assert stays.
>
> There are no regressions if both the cp_unevaluated_operand check and
> the assert are removed however.

I think that's my preference.

Jason
diff mbox

Patch

diff --git a/gcc/cp/pt.c b/gcc/cp/pt.c
index ebfc45b..49ef9d3 100644
--- a/gcc/cp/pt.c
+++ b/gcc/cp/pt.c
@@ -13878,10 +13878,22 @@  tsubst_copy (tree t, tree args, tsubst_flags_t complain, tree in_decl)
       if (r == NULL_TREE)
 	{
 	  /* We get here for a use of 'this' in an NSDMI.  */
-	  if (DECL_NAME (t) == this_identifier
-	      && current_function_decl
-	      && DECL_CONSTRUCTOR_P (current_function_decl))
-	    return current_class_ptr;
+	  if (DECL_NAME (t) == this_identifier)
+	    {
+	      /* We're processing an NSDMI as part of a constructor call.  */
+	      if (current_function_decl
+		  && DECL_CONSTRUCTOR_P (current_function_decl))
+		return current_class_ptr;
+
+	      /* Or as part of an aggregate initialization.  */
+	      if (cp_unevaluated_operand == 0
+		  && current_class_ref
+		  && TREE_CODE (current_class_ref) == PLACEHOLDER_EXPR)
+		{
+		  gcc_assert (cxx_dialect >= cxx14);
+		  return current_class_ptr;
+		}
+	    }
 
 	  /* This can happen for a parameter name used later in a function
 	     declaration (such as in a late-specified return type).  Just
diff --git a/gcc/testsuite/g++.dg/cpp1y/nsdmi-aggr5.C b/gcc/testsuite/g++.dg/cpp1y/nsdmi-aggr5.C
new file mode 100644
index 0000000..fe377c3
--- /dev/null
+++ b/gcc/testsuite/g++.dg/cpp1y/nsdmi-aggr5.C
@@ -0,0 +1,24 @@ 
+// PR c++/70332
+// { dg-do run { target c++14 } }
+
+template <class T>
+struct C
+{
+ T m;
+ T *n = &m;
+};
+
+C<int> c { };
+
+int
+main ()
+{
+  *c.n = 5;
+  if (c.m != 5)
+    __builtin_abort ();
+
+  C<int> d { 10 };
+  *d.n = *d.n + 1;
+  if (d.m != 11)
+    __builtin_abort ();
+}