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
@@ -1623,9 +1623,13 @@
     };
     GlobalizationRFI.foreachUse(SCC, CreateAA);
 
+    // Create an ExecutionDomain AA for every function and a HeapToStack AA for
+    // every function if there is a device kernel.
     for (auto *F : SCC) {
       if (!F->isDeclaration())
         A.getOrCreateAAFor<AAExecutionDomain>(IRPosition::function(*F));
+      if (!OMPInfoCache.Kernels.empty())
+        A.getOrCreateAAFor<AAHeapToStack>(IRPosition::function(*F));
     }
   }
 };
diff --git a/llvm/test/Transforms/OpenMP/gpu_state_machine_function_ptr_replacement.ll b/llvm/test/Transforms/OpenMP/gpu_state_machine_function_ptr_replacement.ll
--- a/llvm/test/Transforms/OpenMP/gpu_state_machine_function_ptr_replacement.ll
+++ b/llvm/test/Transforms/OpenMP/gpu_state_machine_function_ptr_replacement.ll
@@ -29,9 +29,9 @@
 ; CHECK-DAG:   icmp eq i8* %5, @__omp_outlined__1_wrapper.ID
 ; CHECK-DAG:   icmp eq i8* %7, @__omp_outlined__3_wrapper.ID
 
-; CHECK-DAG:   call void @__kmpc_parallel_51(%struct.ident_t* @1, i32 %1, i32 1, i32 -1, i32 -1, i8* bitcast (void (i32*, i32*)* @__omp_outlined__1 to i8*), i8* @__omp_outlined__1_wrapper.ID, i8** %2, i64 0)
+; CHECK-DAG:   call void @__kmpc_parallel_51(%struct.ident_t* noundef @1, i32 %1, i32 noundef 1, i32 noundef -1, i32 noundef -1, i8* noundef bitcast (void (i32*, i32*)* @__omp_outlined__1 to i8*), i8* noundef @__omp_outlined__1_wrapper.ID, i8** noundef %2, i64 noundef 0)
 ; CHECK-DAG:   call void @__kmpc_parallel_51(%struct.ident_t* @1, i32 %0, i32 1, i32 -1, i32 -1, i8* bitcast (void (i32*, i32*)* @__omp_outlined__2 to i8*), i8* bitcast (void (i16, i32)* @__omp_outlined__2_wrapper to i8*), i8** %1, i64 0)
-; CHECK-DAG:   call void @__kmpc_parallel_51(%struct.ident_t* @1, i32 %1, i32 1, i32 -1, i32 -1, i8* bitcast (void (i32*, i32*)* @__omp_outlined__3 to i8*), i8* @__omp_outlined__3_wrapper.ID, i8** %3, i64 0)
+; CHECK-DAG:   call void @__kmpc_parallel_51(%struct.ident_t* noundef @1, i32 %1, i32 noundef 1, i32 noundef -1, i32 noundef -1, i8* noundef bitcast (void (i32*, i32*)* @__omp_outlined__3 to i8*), i8* noundef @__omp_outlined__3_wrapper.ID, i8** noundef %3, i64 noundef 0)
 
 
 %struct.ident_t = type { i32, i32, i32, i32, i8* }
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 +1,83 @@
+; NOTE: Assertions have been autogenerated by utils/update_test_checks.py UTC_ARGS: --function-signature
+; 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*
+
+define void @kernel() {
+; CHECK-LABEL: define {{[^@]+}}@kernel() {
+; CHECK-NEXT:  entry:
+; CHECK-NEXT:    call void @foo() #[[ATTR0:[0-9]+]]
+; CHECK-NEXT:    call void @bar() #[[ATTR0]]
+; CHECK-NEXT:    ret void
+;
+entry:
+  call void @foo()
+  call void @bar()
+  ret void
+}
+
+define internal void @foo() {
+; CHECK-LABEL: define {{[^@]+}}@foo
+; CHECK-SAME: () #[[ATTR0]] {
+; CHECK-NEXT:  entry:
+; CHECK-NEXT:    [[TMP0:%.*]] = alloca i8, i64 4, align 1
+; CHECK-NEXT:    ret void
+;
+entry:
+  %0 = call i8* @__kmpc_alloc_shared(i64 4)
+  call void @use(i8* %0)
+  call void @__kmpc_free_shared(i8* %0)
+  ret void
+}
+
+define internal void @bar() {
+; CHECK-LABEL: define {{[^@]+}}@bar
+; CHECK-SAME: () #[[ATTR0]] {
+; CHECK-NEXT:  entry:
+; CHECK-NEXT:    [[TMP0:%.*]] = call i8* @__kmpc_alloc_shared(i64 noundef 4) #[[ATTR0]]
+; CHECK-NEXT:    call void @share(i8* nofree writeonly [[TMP0]]) #[[ATTR2:[0-9]+]]
+; CHECK-NEXT:    call void @__kmpc_free_shared(i8* [[TMP0]]) #[[ATTR0]]
+; CHECK-NEXT:    ret void
+;
+entry:
+  %0 = call i8* @__kmpc_alloc_shared(i64 4)
+  call void @share(i8* %0)
+  call void @__kmpc_free_shared(i8* %0)
+  ret void
+}
+
+define internal void @use(i8* %x) {
+entry:
+  ret void
+}
+
+define internal void @share(i8* %x) {
+; CHECK-LABEL: define {{[^@]+}}@share
+; CHECK-SAME: (i8* nofree writeonly [[X:%.*]]) #[[ATTR1:[0-9]+]] {
+; CHECK-NEXT:  entry:
+; CHECK-NEXT:    store i8* [[X]], i8** @S, align 8
+; CHECK-NEXT:    ret void
+;
+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*)
+
+!llvm.dbg.cu = !{!0}
+!llvm.module.flags = !{!3, !4}
+!nvvm.annotations = !{!5}
+
+!0 = distinct !DICompileUnit(language: DW_LANG_C99, file: !1, producer: "clang version 13.0.0", isOptimized: false, runtimeVersion: 0, emissionKind: FullDebug, enums: !2, splitDebugInlining: false, nameTableKind: None)
+!1 = !DIFile(filename: "remove_globalization.c", directory: "/tmp/remove_globalization.c")
+!2 = !{}
+!3 = !{i32 2, !"Debug Info Version", i32 3}
+!4 = !{i32 1, !"wchar_size", i32 4}
+!5 = !{void ()* @kernel, !"kernel", i32 1}
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() {
@@ -18,11 +20,11 @@
 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