CodeGen: Correct linkage of thread_local for OS X

The backing store of thread local variables is internal for OS X and all
accesses must go through the thread wrapper.

However, individual TUs may have inlined through the thread wrapper.
To fix this, give the thread wrapper functions WeakAnyLinkage.  This
prevents them from getting inlined into call-sites.

This fixes PR19989.

llvm-svn: 210632

clang/lib/CodeGen/CodeGenModule.cpp

@@ -1903,6 +1903,16 @@ void CodeGenModule::EmitGlobalVarDefinition(const VarDecl *D) {
   // Set the llvm linkage type as appropriate.
   llvm::GlobalValue::LinkageTypes Linkage =
       getLLVMLinkageVarDefinition(D, GV->isConstant());
+
+  // On Darwin, the backing variable for a C++11 thread_local variable always
+  // has internal linkage; all accesses should just be calls to the
+  // Itanium-specified entry point, which has the normal linkage of the
+  // variable.
+  if (const auto *VD = dyn_cast<VarDecl>(D))
+    if (!VD->isStaticLocal() && VD->getTLSKind() == VarDecl::TLS_Dynamic &&
+        Context.getTargetInfo().getTriple().isMacOSX())
+      Linkage = llvm::GlobalValue::InternalLinkage;
+
   GV->setLinkage(Linkage);
   if (D->hasAttr<DLLImportAttr>())
     GV->setDLLStorageClass(llvm::GlobalVariable::DLLImportStorageClass);