diff --git a/openmp/libomptarget/DeviceRTL/CMakeLists.txt b/openmp/libomptarget/DeviceRTL/CMakeLists.txt
--- a/openmp/libomptarget/DeviceRTL/CMakeLists.txt
+++ b/openmp/libomptarget/DeviceRTL/CMakeLists.txt
@@ -108,6 +108,7 @@
   ${include_directory}/Debug.h
   ${include_directory}/Interface.h
   ${include_directory}/Mapping.h
+  ${include_directory}/Memory.h
   ${include_directory}/State.h
   ${include_directory}/Synchronization.h
   ${include_directory}/Types.h
@@ -119,6 +120,7 @@
   ${source_directory}/Debug.cpp
   ${source_directory}/Kernel.cpp
   ${source_directory}/Mapping.cpp
+  ${source_directory}/Memory.cpp
   ${source_directory}/Misc.cpp
   ${source_directory}/Parallelism.cpp
   ${source_directory}/Reduction.cpp
diff --git a/openmp/libomptarget/DeviceRTL/include/Memory.h b/openmp/libomptarget/DeviceRTL/include/Memory.h
new file mode 100644
--- /dev/null
+++ b/openmp/libomptarget/DeviceRTL/include/Memory.h
@@ -0,0 +1,23 @@
+//===--- Memory.h - OpenMP device runtime memory allocator -------- C++ -*-===//
+//
+// Part of the LLVM Project, under the Apache License v2.0 with LLVM Exceptions.
+// See https://llvm.org/LICENSE.txt for license information.
+// SPDX-License-Identifier: Apache-2.0 WITH LLVM-exception
+//
+//===----------------------------------------------------------------------===//
+//
+//
+//
+//===----------------------------------------------------------------------===//
+
+#ifndef OMPTARGET_MEMORY_H
+#define OMPTARGET_MEMORY_H
+
+#include "Types.h"
+
+extern "C" {
+__attribute__((leaf)) void *malloc(uint64_t Size);
+__attribute__((leaf)) void free(void *Ptr);
+}
+
+#endif
diff --git a/openmp/libomptarget/DeviceRTL/include/Types.h b/openmp/libomptarget/DeviceRTL/include/Types.h
--- a/openmp/libomptarget/DeviceRTL/include/Types.h
+++ b/openmp/libomptarget/DeviceRTL/include/Types.h
@@ -32,6 +32,7 @@
 using uint32_t = unsigned int;
 using int64_t = long;
 using uint64_t = unsigned long;
+using size_t = decltype(sizeof(char));
 
 static_assert(sizeof(int8_t) == 1, "type size mismatch");
 static_assert(sizeof(uint8_t) == 1, "type size mismatch");
diff --git a/openmp/libomptarget/DeviceRTL/src/CMakeLists.txt b/openmp/libomptarget/DeviceRTL/src/CMakeLists.txt
--- a/openmp/libomptarget/DeviceRTL/src/CMakeLists.txt
+++ b/openmp/libomptarget/DeviceRTL/src/CMakeLists.txt
@@ -3,6 +3,7 @@
   Debug.cpp
   Kernel.cpp
   Mapping.cpp
+  Memory.cpp
   Misc.cpp
   Parallelism.cpp
   Reduction.cpp
diff --git a/openmp/libomptarget/DeviceRTL/src/Memory.cpp b/openmp/libomptarget/DeviceRTL/src/Memory.cpp
new file mode 100644
--- /dev/null
+++ b/openmp/libomptarget/DeviceRTL/src/Memory.cpp
@@ -0,0 +1,35 @@
+//===------- Memory.cpp - OpenMP device runtime memory allocator -- C++ -*-===//
+//
+// Part of the LLVM Project, under the Apache License v2.0 with LLVM Exceptions.
+// See https://llvm.org/LICENSE.txt for license information.
+// SPDX-License-Identifier: Apache-2.0 WITH LLVM-exception
+//
+//===----------------------------------------------------------------------===//
+//
+//
+//===----------------------------------------------------------------------===//
+
+#pragma omp begin declare target device_type(nohost)
+
+#include "Memory.h"
+
+/// AMDGCN Implementation
+///
+///{
+#pragma omp begin declare variant match(device = {arch(amdgcn)})
+
+extern "C" {
+__attribute__((leaf)) unsigned long long
+__ockl_dm_alloc(unsigned long long Size);
+__attribute__((leaf)) void __ockl_dm_dealloc(unsigned long long Addr);
+
+void *malloc(size_t Size) { return (void *)__ockl_dm_alloc(Size); }
+
+void free(void *P) { __ockl_dm_dealloc((unsigned long long)P); }
+}
+
+#pragma omp end declare variant
+
+///}
+
+#pragma omp end declare target
diff --git a/openmp/libomptarget/DeviceRTL/src/State.cpp b/openmp/libomptarget/DeviceRTL/src/State.cpp
--- a/openmp/libomptarget/DeviceRTL/src/State.cpp
+++ b/openmp/libomptarget/DeviceRTL/src/State.cpp
@@ -13,6 +13,7 @@
 #include "Debug.h"
 #include "Interface.h"
 #include "Mapping.h"
+#include "Memory.h"
 #include "Synchronization.h"
 #include "Types.h"
 #include "Utils.h"
@@ -36,36 +37,6 @@
 
 namespace {
 
-/// Fallback implementations are missing to trigger a link time error.
-/// Implementations for new devices, including the host, should go into a
-/// dedicated begin/end declare variant.
-///
-///{
-
-extern "C" {
-__attribute__((leaf)) void *malloc(uint64_t Size);
-__attribute__((leaf)) void free(void *Ptr);
-}
-
-///}
-
-/// AMDGCN implementations of the shuffle sync idiom.
-///
-///{
-#pragma omp begin declare variant match(device = {arch(amdgcn)})
-
-extern "C" {
-void *malloc(uint64_t Size) {
-  // TODO: Use some preallocated space for dynamic malloc.
-  return nullptr;
-}
-
-void free(void *Ptr) {}
-}
-
-#pragma omp end declare variant
-///}
-
 /// A "smart" stack in shared memory.
 ///
 /// The stack exposes a malloc/free interface but works like a stack internally.