Rework the visibility computation algorithm in preparation

for distinguishing type vs. value visibility.

The changes to the visibility of explicit specializations
are intentional.  The change to the "ugly" test case is
a consequence of a sensible implementation, and I am happy
to argue that this is better behavior.  Other changes may
or may not be intended;  it is quite difficult to divine
intent from some of the code I altered.

I've left behind a comment which I hope explains the
philosophy behind visibility computation.

llvm-svn: 175326

1 files changed, 29 insertions, 15 deletions

diff --git a/clang/test/CodeGenCXX/visibility.cpp b/clang/test/CodeGenCXX/visibility.cpp
index a196dda643d..5564dc40195 100644
--- a/clang/test/CodeGenCXX/visibility.cpp
+++ b/clang/test/CodeGenCXX/visibility.cpp
@@ -580,9 +580,7 @@ namespace PR10113 {
   };
   template class foo::bar<zed>;
   // CHECK: define weak_odr void @_ZN7PR101133foo3barINS_3zedEE3zedEv
-
-  // FIXME: This should be hidden as zed is hidden.
-  // CHECK-HIDDEN: define weak_odr void @_ZN7PR101133foo3barINS_3zedEE3zedEv
+  // CHECK-HIDDEN: define weak_odr hidden void @_ZN7PR101133foo3barINS_3zedEE3zedEv
 }
 
 namespace PR11690 {
@@ -613,9 +611,7 @@ namespace PR11690_2 {
   };
   template class foo::zed<baz>;
   // CHECK: define weak_odr void @_ZN9PR11690_23foo3zedINS_3bazENS0_3barEE3barEv
-
-  // FIXME: This should be hidden as baz is hidden.
-  // CHECK-HIDDEN: define weak_odr void @_ZN9PR11690_23foo3zedINS_3bazENS0_3barEE3barEv
+  // CHECK-HIDDEN: define weak_odr hidden void @_ZN9PR11690_23foo3zedINS_3bazENS0_3barEE3barEv
 }
 
 namespace test23 {
@@ -729,10 +725,10 @@ namespace test34 {
 
 namespace test35 {
   // This is a really ugly testcase. GCC propagates the DEFAULT in zed's
-  // definition. What we do instead is be conservative about merging
-  // implicit visibilities.
-  // FIXME: Maybe the best thing to do here is error? The test at least
-  // makes sure we don't produce a hidden symbol for foo<zed>::bar.
+  // definition. It's not really clear what we can do here, because we
+  // produce the symbols before even seeing the DEFAULT definition of zed.
+  // FIXME: Maybe the best thing to do here is error?  It's certainly hard
+  // to argue that this ought to be valid.
   template<typename T>
   struct DEFAULT foo {
     void bar() {}
@@ -742,7 +738,7 @@ namespace test35 {
   class DEFAULT zed {
   };
   // CHECK: define weak_odr void @_ZN6test353fooINS_3zedEE3barEv
-  // CHECK-HIDDEN: define weak_odr void @_ZN6test353fooINS_3zedEE3barEv
+  // CHECK-HIDDEN: define weak_odr hidden void @_ZN6test353fooINS_3zedEE3barEv
 }
 
 namespace test36 {
@@ -821,8 +817,8 @@ namespace test42 {
   };
   void bar<foo>::zed() {
   }
-  // CHECK: define hidden void @_ZN6test423barINS_3fooEE3zedEv
-  // CHECK-HIDDEN: define hidden void @_ZN6test423barINS_3fooEE3zedEv
+  // CHECK: define void @_ZN6test423barINS_3fooEE3zedEv
+  // CHECK-HIDDEN: define void @_ZN6test423barINS_3fooEE3zedEv
 }
 
 namespace test43 {
@@ -834,8 +830,8 @@ namespace test43 {
   template <>
   DEFAULT void bar<foo>() {
   }
-  // CHECK: define hidden void @_ZN6test433barINS_3fooEEEvv
-  // CHECK-HIDDEN: define hidden void @_ZN6test433barINS_3fooEEEvv
+  // CHECK: define void @_ZN6test433barINS_3fooEEEvv
+  // CHECK-HIDDEN: define void @_ZN6test433barINS_3fooEEEvv
 }
 
 namespace test44 {
@@ -1156,3 +1152,21 @@ namespace test62 {
   // gcc issues a warning about it being unused since "the type is already
   // defined". We should probably do the same.
 }
+
+namespace test63 {
+  enum HIDDEN E { E0 };
+  struct A {
+    template <E> static void foo() {}
+
+    template <E> struct B {
+      static void foo() {}
+    };
+  };
+
+  void test() {
+    A::foo<E0>();
+    A::B<E0>::foo();
+  }
+  // CHECK: define linkonce_odr hidden void @_ZN6test631A3fooILNS_1EE0EEEvv()
+  // CHECK: define linkonce_odr hidden void @_ZN6test631A1BILNS_1EE0EE3fooEv()
+}