diff --git a/arch/linux/tui/main.c b/arch/linux/tui/main.c
index f0d85fb..bbf95b3 100644
--- a/arch/linux/tui/main.c
+++ b/arch/linux/tui/main.c
@@ -23,7 +23,8 @@ u32 syscall(u32 id, u32 size, u32 mem_ptr) {
   USED(size);
   switch(id) {
   case SYSCALL_DBG_PRINT: {
-    printf("%d\n", mem[mem_ptr]);
+    u32 val = READ_U32(mem_ptr);
+    printf("%d\n", val);
     return 0;
   }
   }
@@ -64,6 +65,7 @@ void test_fibonacci() {
   code[cp++] = ENCODE_A(OP_SYSCALL, SYSCALL_DBG_PRINT, 1, 2);
   code[cp++] = ENCODE_A(OP_HALT, 0, 0, 0);
   /* function fib (int n) int */
+  //code[cp++] = ENCODE_A(OP_SYSCALL, SYSCALL_DBG_PRINT, 1, 0);
   code[cp++] = ENCODE_B(OP_LOAD_IMM, 8, fib);
   code[cp++] = ENCODE_B(OP_LOAD_IMM, 1, 2);
   code[cp++] = ENCODE_B(OP_LOAD_IMM, 2, base_case);
diff --git a/vm/vm.c b/vm/vm.c
index ea51468..79db954 100644
--- a/vm/vm.c
+++ b/vm/vm.c
@@ -30,6 +30,7 @@ bool step_vm() {
     u32 r2 = fp + (src2 * 4);/* local child return value to*/
 
     fn_ptr = READ_U32(rd);
+    sp = 0;
 
     /* push parents frame value to reset the heap to */
     WRITE_U32(mp, fp);
@@ -41,7 +42,7 @@ bool step_vm() {
     WRITE_U32(mp, r2);
     mp += 4;
     /* now set the frame pointer, where the locals start */
-    fp = mp;
+    fp = mp;    
     /* move mp forward by count many locals */
     mp += (4 * r1);
     /* jump to dest_ptr */
@@ -96,6 +97,22 @@ bool step_vm() {
     flag = syscall(id, size, rd);
     return true;
   }
+  case OP_PUSH: {
+    DECODE_B(instruction)
+    u32 rd = fp + (dest * 4);
+    u32 val = READ_U32(rd);
+    USED(imm); 
+    WRITE_U32((mp + ((4 * sp) + FRAME_HEADER_SIZE)), val);
+    sp++;
+    return true;
+  }
+  case OP_POP: {
+    DECODE_C(instruction)
+    USED(imm); 
+    mp -= 4;
+    sp--;
+    return true;
+  }
   case OP_LOAD_IMM: {
     DECODE_B(instruction)
     u32 rd = fp + (dest * 4);
@@ -335,21 +352,6 @@ bool step_vm() {
     WRITE_U32(rd, value);
     return true;
   }
-  case OP_PUSH: {
-    DECODE_B(instruction)
-    u32 rd = fp + (dest * 4);
-    u32 val = READ_U32(rd);
-    USED(imm); 
-    WRITE_U32((mp + (4 * (sp + 3))), val);
-    sp++;
-    return true;
-  }
-  case OP_POP: {
-    DECODE_C(instruction)
-    USED(imm); 
-    mp -= 4;
-    return true;
-  }
   case OP_ADD_INT: {
     MATH_OP(i32, +);
   }