some more optimizations

arch/linux/tui/main.c

@@ -34,14 +34,19 @@ u32 syscall(u32 id, u32 size, u32 mem_ptr) {
       return 0;
     }
     case SYSCALL_CONSOLE_READ: {
+      u8 *ptr = &mem[mp];
+      mcpy(ptr, &size, sizeof(u32));
+      ptr += 4;
       for (u32 i = 0; i < size; i++) {
         u8 ch = getchar();
         if (ch == '\0')
           break;
         if (ch == '\n')
           break;
-        mem[mem_ptr + i] = ch;
+        *(ptr++) = ch;
       }
+      ptr[size] = '\0';
+      mp += 4 + size + 1;
     }
   }
 
@@ -78,7 +83,7 @@ void test_fibonacci() {
   code[cp++] = ENCODE_B(OP_LOAD_IMM, 0, 35);
   code[cp++] = ENCODE_B(OP_PUSH, 0, 0);
   code[cp++] = ENCODE_B(OP_LOAD_IMM, 1, fib);
-  code[cp++] = ENCODE_A(OP_CALL, 1, (9 * 4), (2 * 4));
+  code[cp++] = ENCODE_A(OP_CALL, 1, 9, 2);
   /* print */
   code[cp++] = ENCODE_A(OP_INT_TO_STR, 3, 2, 0);
   code[cp++] = ENCODE_A(OP_SYSCALL, SYSCALL_CONSOLE_WRITE, 1, 3);
@@ -92,11 +97,11 @@ void test_fibonacci() {
   code[cp++] = ENCODE_B(OP_LOAD_IMM, 3, 2);
   code[cp++] = ENCODE_A(OP_SUB_INT, 4, 0, 3);
   code[cp++] = ENCODE_B(OP_PUSH, 4, 0);
-  code[cp++] = ENCODE_A(OP_CALL, 8, (9 * 4), (5 * 4));
+  code[cp++] = ENCODE_A(OP_CALL, 8, 9, 5);
   code[cp++] = ENCODE_B(OP_LOAD_IMM, 3, 1);
   code[cp++] = ENCODE_A(OP_SUB_INT, 4, 0, 3);
   code[cp++] = ENCODE_B(OP_PUSH, 4, 0);
-  code[cp++] = ENCODE_A(OP_CALL, 8, (9 * 4), (6 * 4));
+  code[cp++] = ENCODE_A(OP_CALL, 8, 9, 6);
   code[cp++] = ENCODE_A(OP_ADD_INT, 7, 6, 5);
   code[cp++] = ENCODE_B(OP_RETURN, 7, 0);
   code[cp++] = ENCODE_B(OP_RETURN, 0, 0);

vm/libc.c

@@ -1,14 +1,13 @@
 #include "libc.h"
 
 void mcpy(void *to, void *from, u32 length) {
-    u32 i = 0;
     u8 *src, *dest;
     if (to == nil || from == nil) return;
 
     src = (u8 *)from; 
     dest = (u8 *)to; 
 
-    while (length > (i++)) {
+    while (length-- > 0) {
         *(dest++) = *(src++);
     }
     return;

vm/vm.c

@@ -52,13 +52,13 @@ bool step_vm() {
     /* push return address to child frame */
     (*header++) = pc;
     /* push local address to return the value to */
-    (*header++) = fp + src2;
+    (*header++) = fp + (src2 * 4);
     /* increase the mp to new size */
     mp += FRAME_HEADER_SIZE;
     /* now set the frame pointer, where the locals start */
     fp = mp;    
     /* move mp forward by count many locals */
-    mp += src1;
+    mp += (src1 * 4);
     /* jump to dest_ptr */
     pc = fn_ptr;
     return true;