diff --git a/rocclr/runtime/device/rocm/rockernel.hpp b/rocclr/runtime/device/rocm/rockernel.hpp
index 96455a2f0f..7ca68edb98 100644
--- a/rocclr/runtime/device/rocm/rockernel.hpp
+++ b/rocclr/runtime/device/rocm/rockernel.hpp
@@ -128,7 +128,7 @@ public:
     bool init_LC();
 #endif // defined(WITH_LIGHTNING_COMPILER)
 
-    const HSAILProgram* program() {
+    const HSAILProgram* program() const {
         return static_cast<const HSAILProgram*>(program_);
     }
 
diff --git a/rocclr/runtime/device/rocm/rocprogram.cpp b/rocclr/runtime/device/rocm/rocprogram.cpp
index bf66c20e82..c09c5faa80 100644
--- a/rocclr/runtime/device/rocm/rocprogram.cpp
+++ b/rocclr/runtime/device/rocm/rocprogram.cpp
@@ -109,6 +109,8 @@ HSAILProgram::HSAILProgram(roc::NullDevice& device)
     hsaProgramHandle_.handle = 0;
     hsaExecutable_.handle = 0;
 
+    hasGlobalStores_ = false;
+
 #if defined(WITH_LIGHTNING_COMPILER)
     metadata_ = NULL;
 #endif // defined(WITH_LIGHTNING_COMPILER)
@@ -886,6 +888,8 @@ HSAILProgram::setKernels_LC(amd::option::Options *options, void* binary, size_t
     }
 
     size_t progvarsTotalSize = 0;
+    size_t dynamicSize = 0;
+    size_t progvarsWriteSize = 0;
 
     // Begin the Elf image from memory
     Elf* e = elf_memory((char*) binary, binSize, NULL);
@@ -936,9 +940,16 @@ HSAILProgram::setKernels_LC(amd::option::Options *options, void* binary, size_t
             }
         }
         // Accumulate the size of R & !X loadable segments
-        else if (pHdr.p_type == PT_LOAD
-                 && (pHdr.p_flags & PF_R) && !(pHdr.p_flags & PF_X)) {
-            progvarsTotalSize += pHdr.p_memsz;
+        else if (pHdr.p_type == PT_LOAD && !(pHdr.p_flags & PF_X)) {
+            if (pHdr.p_flags & PF_R) {
+                progvarsTotalSize += pHdr.p_memsz;
+            }
+            if (pHdr.p_flags & PF_W) {
+                progvarsWriteSize += pHdr.p_memsz;
+            }
+        }
+        else if (pHdr.p_type == PT_DYNAMIC) {
+            dynamicSize += pHdr.p_memsz;
         }
     }
 
@@ -950,6 +961,10 @@ HSAILProgram::setKernels_LC(amd::option::Options *options, void* binary, size_t
         return false;
     }
 
+    if (progvarsWriteSize != dynamicSize) {
+        hasGlobalStores_ = true;
+    }
+    progvarsTotalSize -= dynamicSize;
     setGlobalVariableTotalSize(progvarsTotalSize);
 
     saveBinaryAndSetType(TYPE_EXECUTABLE, binary, binSize);
@@ -1037,6 +1052,11 @@ HSAILProgram::setKernels_LC(amd::option::Options *options, void* binary, size_t
             return false;
         }
 
+        // FIME_lmoriche: the compiler should set the kernarg alignment based
+        // on the alignment requirement of the parameters. For now, bump it to
+        // the worse case: 128byte aligned.
+        kernargSegmentAlignment = std::max(kernargSegmentAlignment, 128u);
+
         Kernel *aKernel = new roc::Kernel(
             kernelName,
             this,
diff --git a/rocclr/runtime/device/rocm/rocprogram.hpp b/rocclr/runtime/device/rocm/rocprogram.hpp
index 02cc1193bb..d273a982cc 100644
--- a/rocclr/runtime/device/rocm/rocprogram.hpp
+++ b/rocclr/runtime/device/rocm/rocprogram.hpp
@@ -58,6 +58,8 @@ public:
         return dev().getBackendDevice();
     }
 
+    bool hasGlobalStores() const { return hasGlobalStores_; }
+
 protected:
     //! pre-compile setup for GPU
     virtual bool initBuild(amd::option::Options* options);
@@ -160,6 +162,7 @@ private:
     // aclBinary and aclCompiler - for the compiler library
     aclBinary*      binaryElf_; //!< Binary for the new compiler library
     aclBinaryOptions binOpts_; //!< Binary options to create aclBinary
+    bool hasGlobalStores_; //!< program has writable program scope variables
 
     /* HSA executable */
     hsa_ext_program_t hsaProgramHandle_; //!< Handle to HSA runtime program
diff --git a/rocclr/runtime/device/rocm/rocvirtual.cpp b/rocclr/runtime/device/rocm/rocvirtual.cpp
index 431cdf53f3..0b48c25af7 100644
--- a/rocclr/runtime/device/rocm/rocvirtual.cpp
+++ b/rocclr/runtime/device/rocm/rocvirtual.cpp
@@ -315,6 +315,14 @@ VirtualGPU::processMemObjects(
         }
     }
 
+    if (hsaKernel.program()->hasGlobalStores()) {
+        // Sync AQL packets
+        setAqlHeader(kDispatchPacketHeader);
+        // Clear memory dependency state
+        const static bool All = true;
+        memoryDependency().clear(!All);
+    }
+
     return true;
 }