RFA (tree.c): PATCH for may_alias vs. TYPE_CANONICAL, related to c++/50800

RFA (tree.c): PATCH for may_alias vs. TYPE_CANONICAL, related to c++/50800

[Jason Merrill]

[-- Attachment #1: Type: text/plain, Size: 592 bytes --]

In general, TYPE_CANONICAL of a type strips all attributes.  An 
exception to this seems to be that TYPE_REF_CAN_ALIAS_ALL remains set on 
the TYPE_CANONICAL of a pointer/reference type even though its TREE_TYPE 
no longer has the may_alias attribute, and is inconsistent with 
"affects_type_identity" being false for may_alias.  This seems to have 
been a mistake in the patch that first introduced TYPE_CANONICAL, rather 
than a deliberate choice.

I'm also planning to fix 50800 in the front end, but this patch still 
seems like an improvement.

Tested x86-64-pc-linux-gnu, OK for trunk?

[-- Attachment #2: 50800.patch --]
[-- Type: text/x-patch, Size: 2415 bytes --]

commit 724a66ba837560ba233b00cddf5d5d93ec5888f4
Author: Jason Merrill <jason@redhat.com>
Date:   Thu Apr 16 13:57:11 2015 -0400

    	PR c++/50800
    gcc/
    	* tree.c (build_reference_type_for_mode): Don't pass can_alias_all
    	down when building TYPE_CANONICAL.
    	(build_pointer_type_for_mode): Likewise.
    gcc/cp/
    	* typeck.c (structural_comptypes): Don't check TYPE_REF_CAN_ALIAS_ALL.

diff --git a/gcc/cp/typeck.c b/gcc/cp/typeck.c
index e9d4cae..7646225 100644
--- a/gcc/cp/typeck.c
+++ b/gcc/cp/typeck.c
@@ -1310,7 +1310,6 @@ structural_comptypes (tree t1, tree t2, int strict)
 
     case POINTER_TYPE:
       if (TYPE_MODE (t1) != TYPE_MODE (t2)
-	  || TYPE_REF_CAN_ALIAS_ALL (t1) != TYPE_REF_CAN_ALIAS_ALL (t2)
 	  || !same_type_p (TREE_TYPE (t1), TREE_TYPE (t2)))
 	return false;
       break;
diff --git a/gcc/testsuite/g++.dg/ext/alias-canon2.C b/gcc/testsuite/g++.dg/ext/alias-canon2.C
index 4833db8..3806cb4 100644
--- a/gcc/testsuite/g++.dg/ext/alias-canon2.C
+++ b/gcc/testsuite/g++.dg/ext/alias-canon2.C
@@ -31,6 +31,3 @@ out_long (ui64 longVal)
         }
     }
 }
-
-void f(ui32 *) { }
-void f(ui32a *) { }
diff --git a/gcc/testsuite/g++.dg/ext/attrib50.C b/gcc/testsuite/g++.dg/ext/attrib50.C
new file mode 100644
index 0000000..a46c9f7
--- /dev/null
+++ b/gcc/testsuite/g++.dg/ext/attrib50.C
@@ -0,0 +1,11 @@
+// PR c++/50800
+
+template <typename T> struct B;
+template <typename T> struct B<T &> {
+  typedef T type;
+};
+struct A {
+  typedef int TA __attribute__((__may_alias__));
+};
+void d() { B<int &> b; }
+int main() { B<A::TA &> b; }
diff --git a/gcc/tree.c b/gcc/tree.c
index 151e3e2..43c4a36 100644
--- a/gcc/tree.c
+++ b/gcc/tree.c
@@ -7704,7 +7704,7 @@ build_pointer_type_for_mode (tree to_type, machine_mode mode,
   else if (TYPE_CANONICAL (to_type) != to_type)
     TYPE_CANONICAL (t)
       = build_pointer_type_for_mode (TYPE_CANONICAL (to_type),
-				     mode, can_alias_all);
+				     mode, false);
 
   /* Lay out the type.  This function has many callers that are concerned
      with expression-construction, and this simplifies them all.  */
@@ -7771,7 +7771,7 @@ build_reference_type_for_mode (tree to_type, machine_mode mode,
   else if (TYPE_CANONICAL (to_type) != to_type)
     TYPE_CANONICAL (t)
       = build_reference_type_for_mode (TYPE_CANONICAL (to_type),
-				       mode, can_alias_all);
+				       mode, false);
 
   layout_type (t);
 

Re: RFA (tree.c): PATCH for may_alias vs. TYPE_CANONICAL, related to c++/50800

[Richard Biener]

On Thu, Apr 23, 2015 at 5:49 PM, Jason Merrill <jason@redhat.com> wrote:
> In general, TYPE_CANONICAL of a type strips all attributes.  An exception to
> this seems to be that TYPE_REF_CAN_ALIAS_ALL remains set on the
> TYPE_CANONICAL of a pointer/reference type even though its TREE_TYPE no
> longer has the may_alias attribute, and is inconsistent with
> "affects_type_identity" being false for may_alias.  This seems to have been
> a mistake in the patch that first introduced TYPE_CANONICAL, rather than a
> deliberate choice.
>
> I'm also planning to fix 50800 in the front end, but this patch still seems
> like an improvement.
>
> Tested x86-64-pc-linux-gnu, OK for trunk?

Looks good to me.  We eventually were confused by alias.c get_alias_set
dropping to TYPE_CANONICAL (TYPE_MAIN_VARIANT (T)) before determining
the alias set of a type T.  But of course ref-all is only relevant for
the alias-set
of type *T.

Richard.