diff --git a/llvm/include/llvm/Frontend/OpenMP/OMPKinds.def b/llvm/include/llvm/Frontend/OpenMP/OMPKinds.def
--- a/llvm/include/llvm/Frontend/OpenMP/OMPKinds.def
+++ b/llvm/include/llvm/Frontend/OpenMP/OMPKinds.def
@@ -499,6 +499,11 @@
                                    EnumAttr(WillReturn), EnumAttr(NoFree))
                     : AttributeSet(EnumAttr(NoUnwind)))
 
+__OMP_ATTRS_SET(NoCaptureAttrs,
+                OptimisticAttributes
+                    ? AttributeSet(EnumAttr(NoCapture))
+                    : AttributeSet(EnumAttr(NoCapture)))
+
 #if 0
 __OMP_ATTRS_SET(InaccessibleOnlyAttrs,
                 OptimisticAttributes
@@ -840,6 +845,11 @@
 __OMP_RTL_ATTRS(__kmpc_doacross_fini, BarrierAttrs, AttributeSet(),
                 ParamAttrs(ReadOnlyPtrAttrs))
 
+__OMP_RTL_ATTRS(__kmpc_alloc_shared, DefaultAttrs, ReturnPtrAttrs,
+                ParamAttrs())
+__OMP_RTL_ATTRS(__kmpc_free_shared, AllocAttrs, AttributeSet(),
+                ParamAttrs(NoCaptureAttrs))
+
 __OMP_RTL_ATTRS(__kmpc_alloc, DefaultAttrs, ReturnPtrAttrs, ParamAttrs())
 __OMP_RTL_ATTRS(__kmpc_free, AllocAttrs, AttributeSet(), ParamAttrs())
 
diff --git a/llvm/lib/Transforms/IPO/OpenMPOpt.cpp b/llvm/lib/Transforms/IPO/OpenMPOpt.cpp
--- a/llvm/lib/Transforms/IPO/OpenMPOpt.cpp
+++ b/llvm/lib/Transforms/IPO/OpenMPOpt.cpp
@@ -539,6 +539,8 @@
     if (IsModulePass) {
       Changed |= runAttributor();
 
+      OMPInfoCache.recollectUses();
+
       if (remarksEnabled())
         analysisGlobalization();
     } else {
@@ -1149,28 +1151,23 @@
   }
 
   void analysisGlobalization() {
-    RuntimeFunction GlobalizationRuntimeIDs[] = {OMPRTL___kmpc_alloc_shared,
-                                                 OMPRTL___kmpc_free_shared};
-
-    for (const auto GlobalizationCallID : GlobalizationRuntimeIDs) {
-      auto &RFI = OMPInfoCache.RFIs[GlobalizationCallID];
-
-      auto CheckGlobalization = [&](Use &U, Function &Decl) {
-        if (CallInst *CI = getCallIfRegularCall(U, &RFI)) {
-          auto Remark = [&](OptimizationRemarkAnalysis ORA) {
-            return ORA
-                   << "Found thread data sharing on the GPU. "
-                   << "Expect degraded performance due to data globalization.";
-          };
-          emitRemark<OptimizationRemarkAnalysis>(CI, "OpenMPGlobalization",
-                                                 Remark);
-        }
+    auto &RFI = OMPInfoCache.RFIs[OMPRTL___kmpc_alloc_shared];
+
+    auto CheckGlobalization = [&](Use &U, Function &Decl) {
+      if (CallInst *CI = getCallIfRegularCall(U, &RFI)) {
+        auto Remark = [&](OptimizationRemarkAnalysis ORA) {
+          return ORA
+                 << "Found thread data sharing on the GPU. "
+                 << "Expect degraded performance due to data globalization.";
+        };
+        emitRemark<OptimizationRemarkAnalysis>(CI, "OpenMPGlobalization",
+                                               Remark);
+      }
 
-        return false;
-      };
+      return false;
+    };
 
-      RFI.foreachUse(SCC, CheckGlobalization);
-    }
+    RFI.foreachUse(SCC, CheckGlobalization);
   }
 
   /// Maps the values stored in the offload arrays passed as arguments to
@@ -1639,6 +1636,21 @@
       GetterRFI.foreachUse(SCC, CreateAA);
     }
 
+    // Create HeapToStack AA for all globalization calls
+    RuntimeFunction GlobalizationCallID = OMPRTL___kmpc_alloc_shared;
+    auto &GlobalizationRFI = OMPInfoCache.RFIs[GlobalizationCallID];
+    auto CreateAA = [&](Use &U, Function &Decl) {
+      CallInst *CI = OpenMPOpt::getCallIfRegularCall(U, &GlobalizationRFI);
+      if (!CI)
+        return false;
+
+      IRPosition FnPos = IRPosition::function(*CI->getFunction());
+      A.getOrCreateAAFor<AAHeapToStack>(FnPos);
+      return false;
+    };
+    GlobalizationRFI.foreachUse(SCC, CreateAA);
+
+    // Create an ExecutionDomain AA for every function
     for (auto &F : M) {
       if (!F.isDeclaration())
         A.getOrCreateAAFor<AAExecutionDomain>(IRPosition::function(F));
diff --git a/llvm/test/Transforms/OpenMP/remove_globalization.ll b/llvm/test/Transforms/OpenMP/remove_globalization.ll
new file mode 100644
--- /dev/null
+++ b/llvm/test/Transforms/OpenMP/remove_globalization.ll
@@ -0,0 +1,46 @@
+; RUN: opt -S -passes=openmp-opt  < %s | FileCheck %s
+target datalayout = "e-i64:64-i128:128-v16:16-v32:32-n16:32:64"
+target triple = "nvptx64"
+
+@S = external local_unnamed_addr global i8*
+
+; CHECK-NOT: {{.*}} = call i8* @__kmpc_alloc_shared({{.*}})
+; CHECK-NOT: call void @__kmpc_free_shared({{.*}})
+define void @foo() #0 {
+entry:
+  %0 = call i8* @__kmpc_alloc_shared(i64 4)
+  %x_on_stack = bitcast i8* %0 to i32*
+  %1 = bitcast i32* %x_on_stack to i8*
+  call void @use(i8* %1)
+  call void @__kmpc_free_shared(i8* %0)
+  ret void
+}
+
+; CHECK: {{.*}} = call i8* @__kmpc_alloc_shared({{.*}})
+; CHECK: call void @__kmpc_free_shared({{.*}})
+define void @bar() #0 {
+entry:
+  %0 = call i8* @__kmpc_alloc_shared(i64 4)
+  %x_on_stack = bitcast i8* %0 to i32*
+  %1 = bitcast i32* %x_on_stack to i8*
+  call void @share(i8* %1)
+  call void @__kmpc_free_shared(i8* %0)
+  ret void
+}
+
+define void @use(i8* %x) {
+entry:
+  ret void
+}
+
+define void @share(i8* %x) {
+entry:
+  store i8* %x, i8** @S
+  ret void
+}
+
+; CHECK: declare i8* @__kmpc_alloc_shared(i64)
+declare i8* @__kmpc_alloc_shared(i64)
+
+; CHECK: declare void @__kmpc_free_shared(i8* nocapture)
+declare void @__kmpc_free_shared(i8*)
diff --git a/llvm/test/Transforms/PhaseOrdering/openmp-opt-module.ll b/llvm/test/Transforms/PhaseOrdering/openmp-opt-module.ll
--- a/llvm/test/Transforms/PhaseOrdering/openmp-opt-module.ll
+++ b/llvm/test/Transforms/PhaseOrdering/openmp-opt-module.ll
@@ -3,6 +3,8 @@
 
 @.str = private unnamed_addr constant [13 x i8] c"Alloc Shared\00", align 1
 
+@S = external local_unnamed_addr global i8*
+
 ; MODULE: remark: openmp_opt_module.c:5:7: Found thread data sharing on the GPU. Expect degraded performance due to data globalization.
 
 define void @foo() {
@@ -17,12 +19,11 @@
 
 define void @use(i8* %0) {
 entry:
-  %.addr = alloca i8*, align 8
-  store i8* %0, i8** %.addr, align 8
+  store i8* %0, i8** @S
   ret void
 }
 
-define internal i8* @__kmpc_alloc_shared(i64 %DataSize) {
+define weak i8* @__kmpc_alloc_shared(i64 %DataSize) {
 entry:
   %call = call i8* @_Z10SafeMallocmPKc(i64 %DataSize, i8* getelementptr inbounds ([13 x i8], [13 x i8]* @.str, i64 0, i64 0)) #11
   ret i8* %call