diff --git a/primitives/primproc/columncommand.cpp b/primitives/primproc/columncommand.cpp
index d1b122466..62ac4f66f 100644
--- a/primitives/primproc/columncommand.cpp
+++ b/primitives/primproc/columncommand.cpp
@@ -231,7 +231,11 @@ void ColumnCommand::loadData()
                 {
                     ByteStream::hexbyte h;
                     utils::getEmptyRowValue(colType.colDataType, colType.colWidth, (uint8_t*)&h);
-                    hPtr[idx] = h;
+                    __asm__ volatile("movups %1,%0"
+                        :"=m" ( hPtr[idx] ) // output
+                        :"v"( h ) // input
+                        : "memory" // clobbered
+                    );
                 }
             }
 
@@ -328,7 +332,18 @@ void ColumnCommand::process_OT_BOTH()
 
                 bpp->relRids[i] = *((uint16_t*) &bpp->outputMsg[pos]);
                 pos += 2;
-                wide128Values[i] = *((int128_t*) &bpp->outputMsg[pos]);
+                int128_t* int128Ptr = reinterpret_cast<int128_t*>(&bpp->outputMsg[pos]);
+                __asm__ volatile("movdqu %0,%%xmm0;"
+                    : 
+                    :"m"( *int128Ptr ) // input
+                    :"xmm0" // clobbered
+                );
+                __asm__ volatile("movups %%xmm0,%0;"
+                    : "=m" (wide128Values[i])// output
+                    : // input
+                    : "memory", "xmm0" // clobbered
+                );
+ 
                 pos += 16;
             }
 
diff --git a/utils/messageqcpp/bytestream.cpp b/utils/messageqcpp/bytestream.cpp
index c0fad8554..d42239eb0 100644
--- a/utils/messageqcpp/bytestream.cpp
+++ b/utils/messageqcpp/bytestream.cpp
@@ -240,7 +240,12 @@ ByteStream& ByteStream::operator<<(const uint128_t& o)
     if (fBuf == 0 || (fCurInPtr - fBuf + 16U > fMaxLen + ISSOverhead))
         growBuf(fMaxLen + BlockSize);
 
-    *((uint128_t*) fCurInPtr) = o;
+    __asm__ volatile("movups %1,%0;"
+        :"=m" ( *fCurInPtr ) // output
+        :"v"( o ) // input
+        : "memory" // clobbered
+    );
+ 
     fCurInPtr += 16;
 
     return *this;
@@ -251,7 +256,11 @@ ByteStream& ByteStream::operator<<(const int128_t& o)
     if (fBuf == 0 || (fCurInPtr - fBuf + 16U > fMaxLen + ISSOverhead))
         growBuf(fMaxLen + BlockSize);
 
-    *((int128_t*) fCurInPtr) = o;
+    __asm__ volatile("movups %1,%0;"
+        :"=m" ( *fCurInPtr ) // output
+        :"v"( o ) // input
+        : "memory" // clobbered
+    );
     fCurInPtr += 16;
 
     return *this;
@@ -441,7 +450,16 @@ void ByteStream::peek(uint128_t& o) const
     if (length() < 16)
         throw underflow_error("ByteStream>uint128_t: not enough data in stream to fill datatype");
 
-    o = *((uint128_t*) fCurOutPtr);
+    __asm__ volatile("movdqu %0,%%xmm0;"
+        : 
+        :"m"( *fCurOutPtr ) // input
+        :"xmm0" // clobbered
+    );
+    __asm__ volatile("movups %%xmm0,%0;"
+        : "=m" (o)// output
+        : // input
+        : "memory", "xmm0" // clobbered
+    );
 }
 
 void ByteStream::peek(int128_t& o) const
@@ -450,7 +468,16 @@ void ByteStream::peek(int128_t& o) const
     if (length() < 16)
         throw underflow_error("ByteStream>int128_t: not enough data in stream to fill datatype");
 
-    o = *((int128_t*) fCurOutPtr);
+    __asm__ volatile("movdqu %0,%%xmm0;"
+        : 
+        :"m"( *fCurOutPtr ) // input
+        :"xmm0" // clobbered
+    );
+    __asm__ volatile("movups %%xmm0,%0;"
+        : "=m" (o)// output
+        : // input
+        : "memory", "xmm0" // clobbered
+    );
 }
 
 void ByteStream::peek(string& s) const
diff --git a/utils/messageqcpp/bytestream.h b/utils/messageqcpp/bytestream.h
index 3eccb7b0f..ade220d76 100644
--- a/utils/messageqcpp/bytestream.h
+++ b/utils/messageqcpp/bytestream.h
@@ -151,6 +151,7 @@ public:
      *  push an int128_t onto the end of the stream. The byte order is whatever the native byte order is.
      */
     EXPORT ByteStream& operator<<(const int128_t& o);
+
     /**
      *	push an uint128_t onto the end of the stream. The byte order is whatever the native byte order is.
      */
diff --git a/utils/rowgroup/rowgroup.h b/utils/rowgroup/rowgroup.h
index 748bf3dfa..9064aa8b8 100644
--- a/utils/rowgroup/rowgroup.h
+++ b/utils/rowgroup/rowgroup.h
@@ -848,6 +848,22 @@ inline void Row::setBinaryField_offset<uint8_t>(uint8_t* value, uint32_t width,
    memcpy(&data[offset], value, width);
 }
 
+template<>
+inline void Row::setBinaryField_offset<int128_t>(int128_t* value, uint32_t width, uint32_t offset)
+{
+    int128_t *dst128Ptr = reinterpret_cast<int128_t*>(&data[offset]);
+    __asm__ volatile("movdqu %0,%%xmm0;"
+        : 
+        :"m"( *value ) // input
+        :"xmm0" // clobbered
+    );
+    __asm__ volatile("movups %%xmm0,%0;"
+        : "=m" (*dst128Ptr)// output
+        : // input
+        : "memory", "xmm0" // clobbered
+    );
+}
+
 inline void Row::setStringField(const uint8_t* strdata, uint32_t length, uint32_t colIndex)
 {
     uint64_t offset;