Improve efficiency of some instructions

When a value is to be read and kept on the stack, instead of popping and
immediately pushing it we now use a more efficient `get` stack method.

The pop_u16 stack method has also been changed to a variation that
assembles to fewer CPU instructions.

2 files changed, 62 insertions, 46 deletions

diff --git a/src/processor.rs b/src/processor.rs
index c9cde82..9c999c0 100644
--- a/src/processor.rs
+++ b/src/processor.rs
@@ -33,6 +33,10 @@ impl<D: DeviceBus> Processor<D> {
         macro_rules! r_pop_8    { ($v:ident) => { let $v = self.rst.pop_u8();  }; }
         macro_rules! w_pop_16   { ($v:ident) => { let $v = self.wst.pop_u16(); }; }
         macro_rules! r_pop_16   { ($v:ident) => { let $v = self.rst.pop_u16(); }; }
+        macro_rules! w_get_8    { ($v:ident) => { let $v = self.wst.get_u8();  }; }
+        macro_rules! r_get_8    { ($v:ident) => { let $v = self.rst.get_u8();  }; }
+        macro_rules! w_get_16   { ($v:ident) => { let $v = self.wst.get_u16(); }; }
+        macro_rules! r_get_16   { ($v:ident) => { let $v = self.rst.get_u16(); }; }
         macro_rules! m_read_8   { ($a:expr) => { self.mem.read_u8($a)  }; }
         macro_rules! m_read_16  { ($a:expr) => { self.mem.read_u16($a) }; }
         macro_rules! m_lit_8    { ($l:ident) => { let $l = self.mem.read_u8_next();  }; }
@@ -57,25 +61,25 @@ impl<D: DeviceBus> Processor<D> {
                 /* HLT    */ 0x00 => { return Some(Signal::Halt); }
                 /* JMP    */ 0x01 => { w_pop_16!(a);                                    jump!(a);  }
                 /* JCN    */ 0x02 => { w_pop_16!(a); w_pop_8!(t);              if t!=0 {jump!(a)}; }
-                /* JKN    */ 0x03 => { w_pop_16!(a); w_pop_8!(t); w_psh_8!(t); if t!=0 {jump!(a)}; }
+                /* JKN    */ 0x03 => { w_pop_16!(a); w_get_8!(t);              if t!=0 {jump!(a)}; }
                 /* EQU    */ 0x04 => { w_pop_8!(y); w_pop_8!(x);                           w_psh_8!(truth!(x==y)); }
-                /* NKQ    */ 0x05 => { w_pop_8!(y); w_pop_8!(x); w_psh_8!(x); w_psh_8!(y); w_psh_8!(truth!(x!=y)); }
+                /* NKQ    */ 0x05 => { w_pop_8!(y); w_get_8!(x);              w_psh_8!(y); w_psh_8!(truth!(x!=y)); }
                 /* LTH    */ 0x06 => { w_pop_8!(y); w_pop_8!(x);                           w_psh_8!(truth!(x <y)); }
                 /* GTH    */ 0x07 => { w_pop_8!(y); w_pop_8!(x);                           w_psh_8!(truth!(x >y)); }
                 /* LDA    */ 0x08 => { w_pop_16!(a);               w_psh_8!(m_read_8!(a)); }
                 /* STA    */ 0x09 => { w_pop_16!(a); w_pop_8!(v);               m_write_8!(v,a); }
-                /* LKA    */ 0x0A => { w_pop_16!(a); w_psh_16!(a); w_psh_8!(m_read_8!(a)); }
-                /* SKA    */ 0x0B => { w_pop_8!(v); w_pop_16!(a); w_psh_16!(a); m_write_8!(v,a); }
+                /* LKA    */ 0x0A => { w_get_16!(a);               w_psh_8!(m_read_8!(a)); }
+                /* SKA    */ 0x0B => { w_pop_8!(v); w_get_16!(a);               m_write_8!(v,a); }
                 /* LDD    */ 0x0C => { w_pop_8!(p);              w_psh_8!(d_read_8!(p)); }
                 /* STD    */ 0x0D => { w_pop_8!(p); w_pop_8!(v);              d_write_8!(v,p); }
                 /* SHF    */ 0x0E => { w_pop_8!(y); w_pop_8!(x); w_psh_8!(shf!(x,y)); }
                 /* SHC    */ 0x0F => { w_pop_8!(y); w_pop_8!(x); w_psh_8!(shc!(x,y)); }
                 /* PSH    */ 0x10 => { r_pop_8!(x);              w_psh_8!(x); }
-                /* PSK    */ 0x11 => { r_pop_8!(x); r_psh_8!(x); w_psh_8!(x); }
+                /* PSK    */ 0x11 => { r_get_8!(x);              w_psh_8!(x); }
                 /* POP    */ 0x12 => { w_pop_8!(_x); }
                 /* SPL    */ 0x13 => { w_pop_8!(x); nyb!(x=>a,b); w_psh_8!(a); w_psh_8!(b); }
-                /* DUP    */ 0x14 => { w_pop_8!(x); w_psh_8!(x); w_psh_8!(x); }
-                /* OVR    */ 0x15 => { w_pop_8!(y); w_pop_8!(x); w_psh_8!(x); w_psh_8!(y); w_psh_8!(x); }
+                /* DUP    */ 0x14 => { w_get_8!(x);              w_psh_8!(x); }
+                /* OVR    */ 0x15 => { w_pop_8!(y); w_get_8!(x);              w_psh_8!(y); w_psh_8!(x); }
                 /* SWP    */ 0x16 => { w_pop_8!(y); w_pop_8!(x); w_psh_8!(y); w_psh_8!(x); }
                 /* ROT    */ 0x17 => { w_pop_8!(z); w_pop_8!(y); w_pop_8!(x); w_psh_8!(y); w_psh_8!(z); w_psh_8!(x); }
                 /* ADD    */ 0x18 => { w_pop_8!(y); w_pop_8!(x); w_psh_8!(x.wrapping_add(y)); }
@@ -90,25 +94,25 @@ impl<D: DeviceBus> Processor<D> {
                 /* DB1    */ 0x20 => { return Some(Signal::Debug(DebugVariant::DB1)); }
                 /* JSR    */ 0x21 => { w_pop_16!(a);                        r_psh_16!(pc!()); jump!(a);  }
                 /* JSN    */ 0x22 => { w_pop_16!(a);  w_pop_8!(t); if t!=0 {r_psh_16!(pc!()); jump!(a)}; }
-                /* JKN*   */ 0x23 => { w_pop_16!(a); w_pop_16!(t); w_psh_16!(t);     if t!=0 {jump!(a)}; }
+                /* JKN*   */ 0x23 => { w_pop_16!(a); w_get_16!(t);                   if t!=0 {jump!(a)}; }
                 /* EQU*   */ 0x24 => { w_pop_16!(y); w_pop_16!(x);                             w_psh_8!(truth!(x==y)); }
-                /* NKQ*   */ 0x25 => { w_pop_16!(y); w_pop_16!(x); w_psh_16!(x); w_psh_16!(y); w_psh_8!(truth!(x!=y)); }
+                /* NKQ*   */ 0x25 => { w_pop_16!(y); w_get_16!(x);               w_psh_16!(y); w_psh_8!(truth!(x!=y)); }
                 /* LTH*   */ 0x26 => { w_pop_16!(y); w_pop_16!(x);                             w_psh_8!(truth!(x <y)); }
                 /* GTH*   */ 0x27 => { w_pop_16!(y); w_pop_16!(x);                             w_psh_8!(truth!(x >y)); }
                 /* LDA*   */ 0x28 => { w_pop_16!(a);               w_psh_16!(m_read_16!(a)); }
                 /* STA*   */ 0x29 => { w_pop_16!(a); w_pop_16!(v);               m_write_16!(v,a); }
-                /* LKA*   */ 0x2A => { w_pop_16!(a); w_psh_16!(a); w_psh_16!(m_read_16!(a)); }
-                /* SKA*   */ 0x2B => { w_pop_16!(v); w_pop_16!(a); w_psh_16!(a); m_write_16!(v,a); }
+                /* LKA*   */ 0x2A => { w_get_16!(a);               w_psh_16!(m_read_16!(a)); }
+                /* SKA*   */ 0x2B => { w_pop_16!(v); w_get_16!(a);               m_write_16!(v,a); }
                 /* LDD*   */ 0x2C => { w_pop_8!(p);              w_psh_16!(d_read_16!(p)); }
                 /* STD*   */ 0x2D => { w_pop_8!(p); w_pop_16!(v);               d_write_16!(v,p); }
                 /* SHF*   */ 0x2E => { w_pop_8!(y); w_pop_16!(x); w_psh_16!(shf!(x,y)); }
                 /* SHC*   */ 0x2F => { w_pop_8!(y); w_pop_16!(x); w_psh_16!(shc!(x,y)); }
                 /* PSH*   */ 0x30 => { r_pop_16!(x);               w_psh_16!(x); }
-                /* PSK*   */ 0x31 => { r_pop_16!(x); r_psh_16!(x); w_psh_16!(x); }
+                /* PSK*   */ 0x31 => { r_get_16!(x);               w_psh_16!(x); }
                 /* POP*   */ 0x32 => { w_pop_16!(_x); }
                 /* SPL*   */ 0x33 => { w_pop_8!(y); w_pop_8!(x); nyb!(x=>a,b); nyb!(y=>c,d); w_psh_8!(a); w_psh_8!(b); w_psh_8!(c); w_psh_8!(d); }
-                /* DUP*   */ 0x34 => { w_pop_16!(x); w_psh_16!(x); w_psh_16!(x); }
-                /* OVR*   */ 0x35 => { w_pop_16!(y); w_pop_16!(x); w_psh_16!(x); w_psh_16!(y); w_psh_16!(x); }
+                /* DUP*   */ 0x34 => { w_get_16!(x);               w_psh_16!(x); }
+                /* OVR*   */ 0x35 => { w_pop_16!(y); w_get_16!(x);               w_psh_16!(y); w_psh_16!(x); }
                 /* SWP*   */ 0x36 => { w_pop_16!(y); w_pop_16!(x); w_psh_16!(y); w_psh_16!(x); }
                 /* ROT*   */ 0x37 => { w_pop_16!(z); w_pop_16!(y); w_pop_16!(x); w_psh_16!(y); w_psh_16!(z); w_psh_16!(x); }
                 /* ADD*   */ 0x38 => { w_pop_16!(y); w_pop_16!(x); w_psh_16!(x.wrapping_add(y)); }
@@ -123,15 +127,15 @@ impl<D: DeviceBus> Processor<D> {
                 /* DB2    */ 0x40 => { return Some(Signal::Debug(DebugVariant::DB2)); }
                 /* JMP:   */ 0x41 => { m_lit_16!(a);                                   jump!(a);  }
                 /* JCN:   */ 0x42 => { m_lit_16!(a); w_pop_8!(t);              if t!=0{jump!(a)}; }
-                /* JKN:   */ 0x43 => { m_lit_16!(a); w_pop_8!(t); w_psh_8!(t); if t!=0{jump!(a)}; }
+                /* JKN:   */ 0x43 => { m_lit_16!(a); w_get_8!(t);              if t!=0{jump!(a)}; }
                 /* EQU:   */ 0x44 => { m_lit_8!(y); w_pop_8!(x);                           w_psh_8!(truth!(x==y)); }
-                /* NKQ:   */ 0x45 => { m_lit_8!(y); w_pop_8!(x); w_psh_8!(x); w_psh_8!(y); w_psh_8!(truth!(x!=y)); }
+                /* NKQ:   */ 0x45 => { m_lit_8!(y); w_get_8!(x);              w_psh_8!(y); w_psh_8!(truth!(x!=y)); }
                 /* LTH:   */ 0x46 => { m_lit_8!(y); w_pop_8!(x);                           w_psh_8!(truth!(x <y)); }
                 /* GTH:   */ 0x47 => { m_lit_8!(y); w_pop_8!(x);                           w_psh_8!(truth!(x >y)); }
                 /* LDA:   */ 0x48 => { m_lit_16!(a);               w_psh_8!(m_read_8!(a)); }
                 /* STA:   */ 0x49 => { m_lit_16!(a); w_pop_8!(v);               m_write_8!(v,a); }
                 /* LKA:   */ 0x4A => { m_lit_16!(a); w_psh_16!(a); w_psh_8!(m_read_8!(a)); }
-                /* SKA:   */ 0x4B => { m_lit_8!(v); w_pop_16!(a); w_psh_16!(a); m_write_8!(v,a); }
+                /* SKA:   */ 0x4B => { m_lit_8!(v); w_get_16!(a);               m_write_8!(v,a); }
                 /* LDD:   */ 0x4C => { m_lit_8!(p);              w_psh_8!(d_read_8!(p)); }
                 /* STD:   */ 0x4D => { m_lit_8!(p); w_pop_8!(v);              d_write_8!(v,p); }
                 /* SHF:   */ 0x4E => { m_lit_8!(y); w_pop_8!(x); w_psh_8!(shf!(x,y)); }
@@ -141,7 +145,7 @@ impl<D: DeviceBus> Processor<D> {
                 /* POP:   */ 0x52 => { m_lit_8!(_x); }
                 /* SPL:   */ 0x53 => { m_lit_8!(x); nyb!(x=>a,b); w_psh_8!(a); w_psh_8!(b); }
                 /* DUP:   */ 0x54 => { m_lit_8!(x); w_psh_8!(x); w_psh_8!(x); }
-                /* OVR:   */ 0x55 => { m_lit_8!(y); w_pop_8!(x); w_psh_8!(x); w_psh_8!(y); w_psh_8!(x); }
+                /* OVR:   */ 0x55 => { m_lit_8!(y); w_get_8!(x);              w_psh_8!(y); w_psh_8!(x); }
                 /* SWP:   */ 0x56 => { m_lit_8!(y); w_pop_8!(x); w_psh_8!(y); w_psh_8!(x); }
                 /* ROT:   */ 0x57 => { m_lit_8!(z); w_pop_8!(y); w_pop_8!(x); w_psh_8!(y); w_psh_8!(z); w_psh_8!(x); }
                 /* ADD:   */ 0x58 => { m_lit_8!(y); w_pop_8!(x); w_psh_8!(x.wrapping_add(y)); }
@@ -156,15 +160,15 @@ impl<D: DeviceBus> Processor<D> {
                 /* DB3    */ 0x60 => { return Some(Signal::Debug(DebugVariant::DB3)); }
                 /* JSR:   */ 0x61 => { m_lit_16!(a);                        r_psh_16!(pc!()); jump!(a);  }
                 /* JSN:   */ 0x62 => { m_lit_16!(a);  w_pop_8!(t); if t!=0 {r_psh_16!(pc!()); jump!(a)}; }
-                /* JKN*:  */ 0x63 => { m_lit_16!(a); w_pop_16!(t); w_psh_16!(t);     if t!=0 {jump!(a)}; }
+                /* JKN*:  */ 0x63 => { m_lit_16!(a); w_get_16!(t);                   if t!=0 {jump!(a)}; }
                 /* EQU*:  */ 0x64 => { m_lit_16!(y); w_pop_16!(x);                             w_psh_8!(truth!(x==y)); }
-                /* NKQ*:  */ 0x65 => { m_lit_16!(y); w_pop_16!(x); w_psh_16!(x); w_psh_16!(y); w_psh_8!(truth!(x!=y)); }
+                /* NKQ*:  */ 0x65 => { m_lit_16!(y); w_get_16!(x);               w_psh_16!(y); w_psh_8!(truth!(x!=y)); }
                 /* LTH*:  */ 0x66 => { m_lit_16!(y); w_pop_16!(x);                             w_psh_8!(truth!(x <y)); }
                 /* GTH*:  */ 0x67 => { m_lit_16!(y); w_pop_16!(x);                             w_psh_8!(truth!(x >y)); }
                 /* LDA*:  */ 0x68 => { m_lit_16!(a);               w_psh_16!(m_read_16!(a)); }
                 /* STA*:  */ 0x69 => { m_lit_16!(a); w_pop_16!(v);               m_write_16!(v,a); }
                 /* LKA*:  */ 0x6A => { m_lit_16!(a); w_psh_16!(a); w_psh_16!(m_read_16!(a)); }
-                /* SKA*:  */ 0x6B => { m_lit_16!(v); w_pop_16!(a); w_psh_16!(a); m_write_16!(v,a); }
+                /* SKA*:  */ 0x6B => { m_lit_16!(v); w_get_16!(a);               m_write_16!(v,a); }
                 /* LDD*:  */ 0x6C => { m_lit_8!(p);              w_psh_16!(d_read_16!(p)); }
                 /* STD*:  */ 0x6D => { m_lit_8!(p); w_pop_16!(v);               d_write_16!(v,p); }
                 /* SHF*:  */ 0x6E => { m_lit_8!(y); w_pop_16!(x); w_psh_16!(shf!(x,y)); }
@@ -174,7 +178,7 @@ impl<D: DeviceBus> Processor<D> {
                 /* POP*:  */ 0x72 => { m_lit_16!(_x); }
                 /* SPL*:  */ 0x73 => { m_lit_8!(x); m_lit_8!(y); nyb!(x=>a,b); nyb!(y=>c,d); w_psh_8!(a); w_psh_8!(b); w_psh_8!(c); w_psh_8!(d); }
                 /* DUP*:  */ 0x74 => { m_lit_16!(x); w_psh_16!(x); w_psh_16!(x); }
-                /* OVR*:  */ 0x75 => { m_lit_16!(y); w_pop_16!(x); w_psh_16!(x); w_psh_16!(y); w_psh_16!(x); }
+                /* OVR*:  */ 0x75 => { m_lit_16!(y); w_get_16!(x);               w_psh_16!(y); w_psh_16!(x); }
                 /* SWP*:  */ 0x76 => { m_lit_16!(y); w_pop_16!(x); w_psh_16!(y); w_psh_16!(x); }
                 /* ROT*:  */ 0x77 => { m_lit_16!(z); w_pop_16!(y); w_pop_16!(x); w_psh_16!(y); w_psh_16!(z); w_psh_16!(x); }
                 /* ADD*:  */ 0x78 => { m_lit_16!(y); w_pop_16!(x); w_psh_16!(x.wrapping_add(y)); }
@@ -189,25 +193,25 @@ impl<D: DeviceBus> Processor<D> {
                 /* DB4    */ 0x80 => { return Some(Signal::Debug(DebugVariant::DB4)); }
                 /* JMPr   */ 0x81 => { r_pop_16!(a);                                   jump!(a);  }
                 /* JCNr   */ 0x82 => { r_pop_16!(a); r_pop_8!(t);              if t!=0{jump!(a)}; }
-                /* JKNr   */ 0x83 => { r_pop_16!(a); r_pop_8!(t); r_psh_8!(t); if t!=0{jump!(a)}; }
+                /* JKNr   */ 0x83 => { r_pop_16!(a); r_get_8!(t);              if t!=0{jump!(a)}; }
                 /* EQUr   */ 0x84 => { r_pop_8!(y); r_pop_8!(x);                           r_psh_8!(truth!(x==y)); }
-                /* NKQr   */ 0x85 => { r_pop_8!(y); r_pop_8!(x); r_psh_8!(x); r_psh_8!(y); r_psh_8!(truth!(x!=y)); }
+                /* NKQr   */ 0x85 => { r_pop_8!(y); r_get_8!(x);              r_psh_8!(y); r_psh_8!(truth!(x!=y)); }
                 /* LTHr   */ 0x86 => { r_pop_8!(y); r_pop_8!(x);                           r_psh_8!(truth!(x <y)); }
                 /* GTHr   */ 0x87 => { r_pop_8!(y); r_pop_8!(x);                           r_psh_8!(truth!(x >y)); }
                 /* LDAr   */ 0x88 => { r_pop_16!(a);               r_psh_8!(m_read_8!(a)); }
                 /* STAr   */ 0x89 => { r_pop_16!(a); r_pop_8!(v);               m_write_8!(v,a); }
-                /* LKAr   */ 0x8A => { r_pop_16!(a); r_psh_16!(a); r_psh_8!(m_read_8!(a)); }
-                /* SKAr   */ 0x8B => { r_pop_8!(v); r_pop_16!(a); r_psh_16!(a); m_write_8!(v,a); }
+                /* LKAr   */ 0x8A => { r_get_16!(a);             r_psh_8!(m_read_8!(a)); }
+                /* SKAr   */ 0x8B => { r_pop_8!(v); r_get_16!(a);               m_write_8!(v,a); }
                 /* LDDr   */ 0x8C => { r_pop_8!(p);              r_psh_8!(d_read_8!(p)); }
                 /* STDr   */ 0x8D => { r_pop_8!(p); r_pop_8!(v);              d_write_8!(v,p); }
                 /* SHFr   */ 0x8E => { r_pop_8!(y); r_pop_8!(x); r_psh_8!(shf!(x,y)); }
                 /* SHCr   */ 0x8F => { r_pop_8!(y); r_pop_8!(x); r_psh_8!(shc!(x,y)); }
                 /* PSHr   */ 0x90 => { w_pop_8!(x);              r_psh_8!(x); }
-                /* PSKr   */ 0x91 => { w_pop_8!(x); w_psh_8!(x); r_psh_8!(x); }
+                /* PSKr   */ 0x91 => { w_get_8!(x);              r_psh_8!(x); }
                 /* POPr   */ 0x92 => { r_pop_8!(_x); }
                 /* SPLr   */ 0x93 => { r_pop_8!(x); nyb!(x=>a,b); r_psh_8!(a); r_psh_8!(b); }
-                /* DUPr   */ 0x94 => { r_pop_8!(x); r_psh_8!(x); r_psh_8!(x); }
-                /* OVRr   */ 0x95 => { r_pop_8!(y); r_pop_8!(x); r_psh_8!(x); r_psh_8!(y); r_psh_8!(x); }
+                /* DUPr   */ 0x94 => { r_get_8!(x);              r_psh_8!(x); }
+                /* OVRr   */ 0x95 => { r_pop_8!(y); r_get_8!(x);              r_psh_8!(y); r_psh_8!(x); }
                 /* SWPr   */ 0x96 => { r_pop_8!(y); r_pop_8!(x); r_psh_8!(y); r_psh_8!(x); }
                 /* ROTr   */ 0x97 => { r_pop_8!(z); r_pop_8!(y); r_pop_8!(x); r_psh_8!(y); r_psh_8!(z); r_psh_8!(x); }
                 /* ADDr   */ 0x98 => { r_pop_8!(y); r_pop_8!(x); r_psh_8!(x.wrapping_add(y)); }
@@ -223,25 +227,25 @@ impl<D: DeviceBus> Processor<D> {
                 /* DB5    */ 0xA0 => { return Some(Signal::Debug(DebugVariant::DB5)); }
                 /* JSRr   */ 0xA1 => { r_pop_16!(a);                        w_psh_16!(pc!()); jump!(a);  }
                 /* JSNr   */ 0xA2 => { r_pop_16!(a);  r_pop_8!(t); if t!=0 {w_psh_16!(pc!()); jump!(a)}; }
-                /* JKNr*  */ 0xA3 => { r_pop_16!(a); r_pop_16!(t); r_psh_16!(t);     if t!=0 {jump!(a)}; }
+                /* JKNr*  */ 0xA3 => { r_pop_16!(a); r_get_16!(t);                   if t!=0 {jump!(a)}; }
                 /* EQUr*  */ 0xA4 => { r_pop_16!(y); r_pop_16!(x);                             r_psh_8!(truth!(x==y)); }
-                /* NKQr*  */ 0xA5 => { r_pop_16!(y); r_pop_16!(x); r_psh_16!(x); r_psh_16!(y); r_psh_8!(truth!(x!=y)); }
+                /* NKQr*  */ 0xA5 => { r_pop_16!(y); r_get_16!(x);               r_psh_16!(y); r_psh_8!(truth!(x!=y)); }
                 /* LTHr*  */ 0xA6 => { r_pop_16!(y); r_pop_16!(x);                             r_psh_8!(truth!(x <y)); }
                 /* GTHr*  */ 0xA7 => { r_pop_16!(y); r_pop_16!(x);                             r_psh_8!(truth!(x >y)); }
                 /* LDAr*  */ 0xA8 => { r_pop_16!(a);               r_psh_16!(m_read_16!(a)); }
                 /* STAr*  */ 0xA9 => { r_pop_16!(a); r_pop_16!(v);               m_write_16!(v,a); }
-                /* LKAr*  */ 0xAA => { r_pop_16!(a); r_psh_16!(a); r_psh_16!(m_read_16!(a)); }
-                /* SKAr*  */ 0xAB => { r_pop_16!(v); r_pop_16!(a); r_psh_16!(a); m_write_16!(v,a); }
+                /* LKAr*  */ 0xAA => { r_get_16!(a);               r_psh_16!(m_read_16!(a)); }
+                /* SKAr*  */ 0xAB => { r_pop_16!(v); r_get_16!(a);               m_write_16!(v,a); }
                 /* LDDr*  */ 0xAC => { r_pop_8!(p);              r_psh_16!(d_read_16!(p)); }
                 /* STDr*  */ 0xAD => { r_pop_8!(p); r_pop_16!(v);               d_write_16!(v,p); }
                 /* SHFr*  */ 0xAE => { r_pop_8!(y); r_pop_16!(x); r_psh_16!(shf!(x,y)); }
                 /* SHCr*  */ 0xAF => { r_pop_8!(y); r_pop_16!(x); r_psh_16!(shc!(x,y)); }
                 /* PSHr*  */ 0xB0 => { w_pop_16!(x);               r_psh_16!(x); }
-                /* PSKr*  */ 0xB1 => { w_pop_16!(x); w_psh_16!(x); r_psh_16!(x); }
+                /* PSKr*  */ 0xB1 => { w_get_16!(x);               r_psh_16!(x); }
                 /* POPr*  */ 0xB2 => { r_pop_16!(_x); }
                 /* SPLr*  */ 0xB3 => { r_pop_8!(y); r_pop_8!(x); nyb!(x=>a,b); nyb!(y=>c,d); r_psh_8!(a); r_psh_8!(b); r_psh_8!(c); r_psh_8!(d); }
-                /* DUPr*  */ 0xB4 => { r_pop_16!(x); r_psh_16!(x); r_psh_16!(x); }
-                /* OVRr*  */ 0xB5 => { r_pop_16!(y); r_pop_16!(x); r_psh_16!(x); r_psh_16!(y); r_psh_16!(x); }
+                /* DUPr*  */ 0xB4 => { r_get_16!(x);               r_psh_16!(x); }
+                /* OVRr*  */ 0xB5 => { r_pop_16!(y); r_get_16!(x);               r_psh_16!(y); r_psh_16!(x); }
                 /* SWPr*  */ 0xB6 => { r_pop_16!(y); r_pop_16!(x); r_psh_16!(y); r_psh_16!(x); }
                 /* ROTr*  */ 0xB7 => { r_pop_16!(z); r_pop_16!(y); r_pop_16!(x); r_psh_16!(y); r_psh_16!(z); r_psh_16!(x); }
                 /* ADDr*  */ 0xB8 => { r_pop_16!(y); r_pop_16!(x); r_psh_16!(x.wrapping_add(y)); }
@@ -256,15 +260,15 @@ impl<D: DeviceBus> Processor<D> {
                 /* DB6    */ 0xC0 => { return Some(Signal::Debug(DebugVariant::DB6)); }
                 /* JMPr:  */ 0xC1 => { m_lit_16!(a);                                   jump!(a);  }
                 /* JCNr:  */ 0xC2 => { m_lit_16!(a); r_pop_8!(t);              if t!=0{jump!(a)}; }
-                /* JKNr:  */ 0xC3 => { m_lit_16!(a); r_pop_8!(t); r_psh_8!(t); if t!=0{jump!(a)}; }
+                /* JKNr:  */ 0xC3 => { m_lit_16!(a); r_get_8!(t);              if t!=0{jump!(a)}; }
                 /* EQUr:  */ 0xC4 => { m_lit_8!(y); r_pop_8!(x);                           r_psh_8!(truth!(x==y)); }
-                /* NKQr:  */ 0xC5 => { m_lit_8!(y); r_pop_8!(x); r_psh_8!(x); r_psh_8!(y); r_psh_8!(truth!(x!=y)); }
+                /* NKQr:  */ 0xC5 => { m_lit_8!(y); r_get_8!(x);              r_psh_8!(y); r_psh_8!(truth!(x!=y)); }
                 /* LTHr:  */ 0xC6 => { m_lit_8!(y); r_pop_8!(x);                           r_psh_8!(truth!(x <y)); }
                 /* GTHr:  */ 0xC7 => { m_lit_8!(y); r_pop_8!(x);                           r_psh_8!(truth!(x >y)); }
                 /* LDAr:  */ 0xC8 => { m_lit_16!(a);               r_psh_8!(m_read_8!(a)); }
                 /* STAr:  */ 0xC9 => { m_lit_16!(a); r_pop_8!(v);               m_write_8!(v,a); }
                 /* LKAr:  */ 0xCA => { m_lit_16!(a); r_psh_16!(a); r_psh_8!(m_read_8!(a)); }
-                /* SKAr:  */ 0xCB => { m_lit_8!(v); r_pop_16!(a); r_psh_16!(a); m_write_8!(v,a); }
+                /* SKAr:  */ 0xCB => { m_lit_8!(v); r_get_16!(a);               m_write_8!(v,a); }
                 /* LDDr:  */ 0xCC => { m_lit_8!(p);              r_psh_8!(d_read_8!(p)); }
                 /* STDr:  */ 0xCD => { m_lit_8!(p); r_pop_8!(v);              d_write_8!(v,p); }
                 /* SHFr:  */ 0xCE => { m_lit_8!(y); r_pop_8!(x); r_psh_8!(shf!(x,y)); }
@@ -274,7 +278,7 @@ impl<D: DeviceBus> Processor<D> {
                 /* POPr:  */ 0xD2 => { m_lit_8!(_x); }
                 /* SPLr:  */ 0xD3 => { m_lit_8!(x); nyb!(x=>a,b); r_psh_8!(a); r_psh_8!(b); }
                 /* DUPr:  */ 0xD4 => { m_lit_8!(x); r_psh_8!(x); r_psh_8!(x); }
-                /* OVRr:  */ 0xD5 => { m_lit_8!(y); r_pop_8!(x); r_psh_8!(x); r_psh_8!(y); r_psh_8!(x); }
+                /* OVRr:  */ 0xD5 => { m_lit_8!(y); r_get_8!(x);              r_psh_8!(y); r_psh_8!(x); }
                 /* SWPr:  */ 0xD6 => { m_lit_8!(y); r_pop_8!(x); r_psh_8!(y); r_psh_8!(x); }
                 /* ROTr:  */ 0xD7 => { m_lit_8!(z); r_pop_8!(y); r_pop_8!(x); r_psh_8!(y); r_psh_8!(z); r_psh_8!(x); }
                 /* ADDr:  */ 0xD8 => { m_lit_8!(y); r_pop_8!(x); r_psh_8!(x.wrapping_add(y)); }
@@ -289,15 +293,15 @@ impl<D: DeviceBus> Processor<D> {
                 /* NOP    */ 0xE0 => { }
                 /* JSRr:  */ 0xE1 => { m_lit_16!(a);                        w_psh_16!(pc!()); jump!(a);  }
                 /* JSNr:  */ 0xE2 => { m_lit_16!(a);  r_pop_8!(t); if t!=0 {w_psh_16!(pc!()); jump!(a)}; }
-                /* JKNr*: */ 0xE3 => { m_lit_16!(a); r_pop_16!(t); r_psh_16!(t);     if t!=0 {jump!(a)}; }
+                /* JKNr*: */ 0xE3 => { m_lit_16!(a); r_get_16!(t);                   if t!=0 {jump!(a)}; }
                 /* EQUr*: */ 0xE4 => { m_lit_16!(y); r_pop_16!(x);                             r_psh_8!(truth!(x==y)); }
-                /* NKQr*: */ 0xE5 => { m_lit_16!(y); r_pop_16!(x); r_psh_16!(x); r_psh_16!(y); r_psh_8!(truth!(x!=y)); }
+                /* NKQr*: */ 0xE5 => { m_lit_16!(y); r_get_16!(x);               r_psh_16!(y); r_psh_8!(truth!(x!=y)); }
                 /* LTHr*: */ 0xE6 => { m_lit_16!(y); r_pop_16!(x);                             r_psh_8!(truth!(x <y)); }
                 /* GTHr*: */ 0xE7 => { m_lit_16!(y); r_pop_16!(x);                             r_psh_8!(truth!(x >y)); }
                 /* LDAr*: */ 0xE8 => { m_lit_16!(a);               r_psh_16!(m_read_16!(a)); }
                 /* STAr*: */ 0xE9 => { m_lit_16!(a); r_pop_16!(v);               m_write_16!(v,a); }
                 /* LKAr*: */ 0xEA => { m_lit_16!(a); r_psh_16!(a); r_psh_16!(m_read_16!(a)); }
-                /* SKAr*: */ 0xEB => { m_lit_16!(v); r_pop_16!(a); r_psh_16!(a); m_write_16!(v,a); }
+                /* SKAr*: */ 0xEB => { m_lit_16!(v); r_get_16!(a);               m_write_16!(v,a); }
                 /* LDDr*: */ 0xEC => { m_lit_8!(p);              r_psh_16!(d_read_16!(p)); }
                 /* STDr*: */ 0xED => { m_lit_8!(p); r_pop_16!(v);               d_write_16!(v,p); }
                 /* SHFr*: */ 0xEE => { m_lit_8!(y); r_pop_16!(x); r_psh_16!(shf!(x,y)); }
@@ -307,7 +311,7 @@ impl<D: DeviceBus> Processor<D> {
                 /* POPr*: */ 0xF2 => { m_lit_16!(_x); }
                 /* SPLr*: */ 0xF3 => { m_lit_8!(x); m_lit_8!(y); nyb!(x=>a,b); nyb!(y=>c,d); r_psh_8!(a); r_psh_8!(b); r_psh_8!(c); r_psh_8!(d); }
                 /* DUPr*: */ 0xF4 => { m_lit_16!(x); r_psh_16!(x); r_psh_16!(x); }
-                /* OVRr*: */ 0xF5 => { m_lit_16!(y); r_pop_16!(x); r_psh_16!(x); r_psh_16!(y); r_psh_16!(x); }
+                /* OVRr*: */ 0xF5 => { m_lit_16!(y); r_get_16!(x);               r_psh_16!(y); r_psh_16!(x); }
                 /* SWPr*: */ 0xF6 => { m_lit_16!(y); r_pop_16!(x); r_psh_16!(y); r_psh_16!(x); }
                 /* ROTr*: */ 0xF7 => { m_lit_16!(z); r_pop_16!(y); r_pop_16!(x); r_psh_16!(y); r_psh_16!(z); r_psh_16!(x); }
                 /* ADDr*: */ 0xF8 => { m_lit_16!(y); r_pop_16!(x); r_psh_16!(x.wrapping_add(y)); }
diff --git a/src/stack.rs b/src/stack.rs
index 2a0a514..14e1799 100644
--- a/src/stack.rs
+++ b/src/stack.rs
@@ -28,8 +28,20 @@ impl Stack {
     }
 
     pub fn pop_u16(&mut self) -> u16 {
-        let byte_low = self.pop_u8();
-        let byte_high = self.pop_u8();
+        self.sp = self.sp.wrapping_sub(1);
+        let byte_low = self.mem[self.sp as usize];
+        self.sp = self.sp.wrapping_sub(1);
+        let byte_high = self.mem[self.sp as usize];
+        u16::from_be_bytes([byte_high, byte_low])
+    }
+
+    pub fn get_u8(&mut self) -> u8 {
+        self.mem[self.sp.wrapping_sub(1) as usize]
+    }
+
+    pub fn get_u16(&mut self) -> u16 {
+        let byte_low = self.mem[self.sp.wrapping_sub(1) as usize];
+        let byte_high = self.mem[self.sp.wrapping_sub(2) as usize];
         u16::from_be_bytes([byte_high, byte_low])
     }