diff --git a/maple_engine/src/invoke_method.cpp b/maple_engine/src/invoke_method.cpp
index 7bfd034e3c5dbd0528a6019ae51c86405f239b7b..946f5270547a6f1b70602a2d482a60361715c5cd 100644
--- a/maple_engine/src/invoke_method.cpp
+++ b/maple_engine/src/invoke_method.cpp
@@ -138,6 +138,7 @@ void collect_stack_refs(void* bp, std::set<void*>& refs);
 #endif // !MPLPRE_C
 
 extern "C" void MCC_DecRef_NaiveRCFast(void* obj);
+extern "C" void MCC_IncRef_NaiveRCFast(void* obj);
 
 MValue maple_invoke_method(const method_header_t* const mir_header, const MFunction *caller) {
     // Array of labels for threaded interpretion
@@ -244,7 +245,7 @@ label_OP_addrof:
         res.x.a64 = (uint8_t*)&arg.x;
     } else {
         MValue &local = MLOCALS(-idx);
-        local.ptyp = (PrimType)func.header->primtype_table[func.header->formals_num - idx];
+        local.ptyp = (PrimType)func.header->primtype_table[(func.header->formals_num - idx)*2]; // both formals and locals are 2B each
         res.x.a64 = (uint8_t*)&local.x;
     }
     MPUSH(res);
@@ -1229,9 +1230,17 @@ label_exception_handler:
         }
     }
 
-    { // Exception not handled; do reference counting cleanup for localrefvars
-      uint8_t* locals_table = (uint8_t*)&func.header->primtype_table + func.header->formals_num; // find start of table for locals; skip formals, which are 1B each
+    { // Exception not handled; do reference counting cleanup for formals and localrefvars
+      uint8_t* formals_table = (uint8_t*)&func.header->primtype_table;
+      for(int i=0; i<func.header->formals_num; i++) {
+        if(formals_table[2*i+1] == 1) { // 2B for each formal; clean up if the 2nd byte is 1.
+          MValue &arg = MARGS(i+1); // with MARGS(), formals index starts with 1
+          void* ref = (void*)arg.x.a64;
+          MCC_DecRef_NaiveRCFast(ref);
+        }
+      }
 
+      uint8_t* locals_table = (uint8_t*)&func.header->primtype_table + func.header->formals_num*2; // find start of table for locals; skip formals, which are 2B each
       for(int i=0; i<func.header->locals_num; i++) {
         if(locals_table[2*i+1] == 2 || locals_table[2*i+1] == 3) { // 2B for each local var; clean up if the 2nd byte is 2 or 3.
           if(func.operand_stack[i].ptyp != PTY_a64)
@@ -1240,6 +1249,13 @@ label_exception_handler:
           if(ref && ref != thrownval) {
             MCC_DecRef_NaiveRCFast(ref);
           }
+        } else if(locals_table[2*i+1] == 4 || locals_table[2*i+1] == 5) {
+          if(func.operand_stack[i].ptyp != PTY_a64)
+            continue;
+          void* ref = (void*)func.operand_stack[i].x.a64;
+          if(ref) {
+            MCC_IncRef_NaiveRCFast(ref);
+          }
         }
       }
     }
diff --git a/maple_engine/src/mfunction.cpp b/maple_engine/src/mfunction.cpp
index 637c3f1a6be48bcd7798c29b4f683e63e2164cb5..a79f3cfdcdf57ab6d4f65baf424bd57a53c5bc49 100644
--- a/maple_engine/src/mfunction.cpp
+++ b/maple_engine/src/mfunction.cpp
@@ -73,7 +73,7 @@ namespace maple {
                 sp = header->locals_num;
                 // for all locals, return value, throw value and evaluation stack
                 operand_stack.resize(sp + header->eval_depth + 1, {.x.i64 = 0, PTY_void});
-                var_names = (char*)(&header->primtype_table) + header->formals_num + header->locals_num*2; // locals_num*2 because each entry has 2 bytes
+                var_names = (char*)(&header->primtype_table) + header->formals_num*2 + header->locals_num*2; // *2 because formals and locals each have 2 bytes
                 if(var_names >= (char*)pc)
                     var_names = nullptr;
             }
diff --git a/maple_engine/src/shimfunction.cpp b/maple_engine/src/shimfunction.cpp
index 523d41f6c76eeaa2a678e37a1cd7b0ff421f8c83..b45474718cd564072b201f088937158b5cc88ac5 100644
--- a/maple_engine/src/shimfunction.cpp
+++ b/maple_engine/src/shimfunction.cpp
@@ -57,7 +57,7 @@ extern "C" int64_t __engine_shim(int64_t first_arg, ...) {
         uint16_t arg_idx = 0;
         while(arg_idx < arg_num) {
             // Process all other argements
-            val.ptyp = (PrimType)(header->primtype_table[arg_idx]);
+            val.ptyp = (PrimType)(header->primtype_table[arg_idx*2]);  // each argument has 2B
             switch(val.ptyp) {
                 case PTY_i8:
                     val.x.i8 = va_arg(args, int);