[fix] Increment RTL return address

Fixes #20

Author: smokku

src/chips/w65c816s.h

@@ -1862,8 +1862,8 @@ uint64_t w65816_tick(w65816_t* c, uint64_t pins) {
         case (0x60<<4)|1: _SA(c->PC);break;
         case (0x60<<4)|2: _VDA(0);_SA(_SP(++_S(c)));break;
         case (0x60<<4)|3: _VDA(0);_SA(_SP(++_S(c)));c->AD=_GD();break;
-        case (0x60<<4)|4: c->PC=(_GD()<<8)|c->AD;_SA(c->PC++);break;
-        case (0x60<<4)|5: _FETCH();break;
+        case (0x60<<4)|4: _SA(_SP(_S(c)));c->PC=(_GD()<<8)|c->AD;break;
+        case (0x60<<4)|5: ++c->PC;_FETCH();break;
         case (0x60<<4)|6: assert(false);break;
         case (0x60<<4)|7: assert(false);break;
         case (0x60<<4)|8: assert(false);break;
@@ -1973,7 +1973,7 @@ uint64_t w65816_tick(w65816_t* c, uint64_t pins) {
         case (0x6B<<4)|2: _VDA(0);_SA(_SP(++_S(c)));break;
         case (0x6B<<4)|3: _VDA(0);_SA(_SP(++_S(c)));c->AD=_GD();break;
         case (0x6B<<4)|4: _VDA(0);_SA(_SP(++_S(c)));c->PC=(_GD()<<8)|c->AD;break;
-        case (0x6B<<4)|5: c->PBR=_GD();_FETCH();break;
+        case (0x6B<<4)|5: c->PBR=_GD();++c->PC;_FETCH();break;
         case (0x6B<<4)|6: assert(false);break;
         case (0x6B<<4)|7: assert(false);break;
         case (0x6B<<4)|8: assert(false);break;

src/codegen/w65c816s_gen.py

@@ -752,10 +752,10 @@ def i_rts(o):
     o.t('_VDA(0);_SA(_SP(++_S(c)));')
     # load return address high byte from stack
     o.t('_VDA(0);_SA(_SP(++_S(c)));c->AD=_GD();')
-    # put return address in PC, this is one byte before next op, do junk read from Stack
-    o.t('c->PC=(_GD()<<8)|c->AD;_SA(c->PC++);')
+    # put return address in PC, this is one byte before next op, do junk read from stack
+    o.t('_SA(_SP(_S(c)));c->PC=(_GD()<<8)|c->AD;')
     # next tick is opcode fetch
-    o.t('');
+    o.t('++c->PC;');
 
 #-------------------------------------------------------------------------------
 def i_rtl(o):
@@ -769,7 +769,7 @@ def i_rtl(o):
     # put return address in PC, read PBR from stack
     o.t('_VDA(0);_SA(_SP(++_S(c)));c->PC=(_GD()<<8)|c->AD;')
     # next tick is opcode fetch
-    o.t('c->PBR=_GD();');
+    o.t('c->PBR=_GD();++c->PC;');
 
 #-------------------------------------------------------------------------------
 def i_rti(o):