diff --git a/assembler.c b/assembler.c
index f0a9c25..216c0d8 100644
--- a/assembler.c
+++ b/assembler.c
@@ -45,7 +45,7 @@ Program p;
 /* clang-format off */
 
 char ops[][4] = {
-	"BRK", "NOP", "LIT", "LDR", "STR", "JMP", "JSR", "RTN", 
+	"BRK", "NOP", "LIT", "LDR", "STR", "JMP", "JSR", "---", 
 	"EQU", "NEQ", "GTH", "LTH", "AND", "XOR", "ROL", "ROR",
 	"POP", "DUP", "SWP", "OVR", "ROT", "CLN", "WSR", "RSW",
 	"ADD", "SUB", "MUL", "DIV", "---", "---", "---", "---"
@@ -155,7 +155,7 @@ findopcode(char *s)
 			continue;
 		while(s[3 + m]) {
 			if(s[3 + m] == '2') i |= (1 << 5); /* mode: short */
-			if(s[3 + m] == 'S') i |= (1 << 6); /* mode: signed */
+			if(s[3 + m] == 'r') i |= (1 << 6); /* mode: return */
 			if(s[3 + m] == '?') i |= (1 << 7); /* mode: conditional */
 			m++;
 		}
diff --git a/build.sh b/build.sh
index 539d023..9bc2dea 100755
--- a/build.sh
+++ b/build.sh
@@ -20,5 +20,5 @@ cc -std=c89 -DDEBUG -Wall -Wno-unknown-pragmas -Wpedantic -Wshadow -Wextra -Werr
 # cc uxn.c emulator.c -std=c89 -Os -DNDEBUG -g0 -s -Wall -Wno-unknown-pragmas -L/usr/local/lib -lSDL2 -o bin/emulator
 
 # run
-./bin/assembler projects/software/nasu.usm bin/boot.rom
+./bin/assembler projects/software/left.usm bin/boot.rom
 ./bin/emulator bin/boot.rom
diff --git a/projects/examples/dev.console.usm b/projects/examples/dev.console.usm
index e07d4da..89c3b90 100644
--- a/projects/examples/dev.console.usm
+++ b/projects/examples/dev.console.usm
@@ -1,5 +1,7 @@
 ( hello world )
 
+%RTN { JMP2r }
+
 @RESET 
 	
 	,text1 ,print-label JSR2
diff --git a/projects/examples/dev.mouse.usm b/projects/examples/dev.mouse.usm
index b72ebb9..51d0f25 100644
--- a/projects/examples/dev.mouse.usm
+++ b/projects/examples/dev.mouse.usm
@@ -1,5 +1,7 @@
 ( Mouse )
 
+%RTN { JMP2r }
+
 ;label { x 2 y 2 color 1 addr 2 }
 ;cat { x 2 y 2 } 
 ;pointer { x 2 y 2 }
diff --git a/projects/examples/gui.editor.usm b/projects/examples/gui.editor.usm
index 8b90c28..c6c17e0 100644
--- a/projects/examples/gui.editor.usm
+++ b/projects/examples/gui.editor.usm
@@ -1,5 +1,7 @@
 ( app/hex )      
 
+%RTN { JMP2r }
+
 ;label { x 2 y 2 color 1 addr 2 }
 ;pixel { x 1 y 1 }
 ;window { x1 2 y1 2 x2 2 y2 2 }
@@ -7,7 +9,6 @@
 ;editor { x1 2 y1 2 addr 2 }
 ;rect { x1 2 y1 2 x2 2 y2 2 }
 ;color { byte 1 }
-;addr { short 2 }
 
 |0100 @RESET
 	
diff --git a/projects/examples/gui.hover.usm b/projects/examples/gui.hover.usm
index 657f6ed..94de739 100644
--- a/projects/examples/gui.hover.usm
+++ b/projects/examples/gui.hover.usm
@@ -1,5 +1,7 @@
 ( draw routines )
 
+%RTN { JMP2r }
+
 ;color { byte 1 }
 ;pointer { x 2 y 2 sprite 2 } 
 ;rect { x1 2 y1 2 x2 2 y2 2 }
diff --git a/projects/examples/gui.label.usm b/projects/examples/gui.label.usm
index af27c33..8043a1d 100644
--- a/projects/examples/gui.label.usm
+++ b/projects/examples/gui.label.usm
@@ -1,5 +1,7 @@
 ( draw routines )
 
+%RTN { JMP2r }
+
 ;label { x 2 y 2 color 1 addr 2 }
 ;rect { x1 2 y1 2 x2 2 y2 2 }
 ;center { x 2 y 2 }
diff --git a/projects/examples/gui.picture.usm b/projects/examples/gui.picture.usm
index fc0de35..839906b 100644
--- a/projects/examples/gui.picture.usm
+++ b/projects/examples/gui.picture.usm
@@ -1,5 +1,7 @@
 ( gui/picture )
 
+%RTN { JMP2r }
+
 ;pict { x 2 y 2 width 2 height 2 color 1 addr 2 }
 
 |0100 @RESET
diff --git a/projects/examples/gui.shapes.usm b/projects/examples/gui.shapes.usm
index b7ba5c6..d361fb9 100644
--- a/projects/examples/gui.shapes.usm
+++ b/projects/examples/gui.shapes.usm
@@ -1,5 +1,7 @@
 ( draw routines )
 
+%RTN { JMP2r }
+
 ;label { x 2 y 2 color 1 addr 2 }
 ;pict { x 2 y 2 width 2 height 2 color 1 addr 2 }
 ;rect { x1 2 y1 2 x2 2 y2 2 }
diff --git a/projects/software/left.usm b/projects/software/left.usm
index 6d10d28..d96ab3d 100644
--- a/projects/software/left.usm
+++ b/projects/software/left.usm
@@ -15,6 +15,9 @@
 			- Real scrolling distance
 )  
 
+%RTN { JMP2r }
+%RTN? { JMP2r? }
+
 ;lock { byte 1 }
 ;k { byte 1 } 
 ;l { byte 1 } 
diff --git a/projects/software/nasu.usm b/projects/software/nasu.usm
index e5dd187..e9e6348 100644
--- a/projects/software/nasu.usm
+++ b/projects/software/nasu.usm
@@ -6,6 +6,9 @@
 	1 2 3  - Select brush
 )
 
+%RTN { JMP2r }
+%RTN? { JMP2r? }
+
 ;bankview { x 2 y 2 mode 1 addr 2 }
 ;tileview { x 2 y 2 addr 2 }
 ;rect { x1 2 y1 2 x2 2 y2 2 }
diff --git a/uxn.c b/uxn.c
index 14c8242..6975ff1 100644
--- a/uxn.c
+++ b/uxn.c
@@ -33,63 +33,63 @@ Uint16 peek16(Stack *s, Uint8 a) { return peek8(s, a * 2) + (peek8(s, a * 2 + 1)
 void op_brk(Uxn *u) { setflag(&u->status, FLAG_HALT, 1); }
 void op_lit(Uxn *u) { u->literal += 1; }
 void op_nop(Uxn *u) { (void)u; }
-void op_jmp(Uxn *u) { Uint8 a = pop8(&u->wst); u->ram.ptr += (Sint8)a; }
-void op_jsr(Uxn *u) { Uint8 a = pop8(&u->wst); push16(&u->rst, u->ram.ptr); u->ram.ptr += (Sint8)a; }
-void op_rtn(Uxn *u) { u->ram.ptr = pop16(&u->rst); }
-void op_ldr(Uxn *u) { Uint16 a = pop16(&u->wst); push8(&u->wst, mempeek8(u, a)); }
-void op_str(Uxn *u) { Uint16 a = pop16(&u->wst); Uint8 b = pop8(&u->wst); mempoke8(u, a, b); }
+void op_jmp(Uxn *u) { Uint8 a = pop8(u->src); u->ram.ptr += (Sint8)a; }
+void op_jsr(Uxn *u) { Uint8 a = pop8(u->src); push16(u->dst, u->ram.ptr); u->ram.ptr += (Sint8)a; }
+void op_rtn(Uxn *u) { u->ram.ptr = pop16(u->dst); }
+void op_ldr(Uxn *u) { Uint16 a = pop16(u->src); push8(u->src, mempeek8(u, a)); }
+void op_str(Uxn *u) { Uint16 a = pop16(u->src); Uint8 b = pop8(u->src); mempoke8(u, a, b); }
 /* Logic */
-void op_and(Uxn *u) { Uint8 a = pop8(&u->wst), b = pop8(&u->wst); push8(&u->wst, b & a); }
-void op_xor(Uxn *u) { Uint8 a = pop8(&u->wst), b = pop8(&u->wst); push8(&u->wst, b | a); }
-void op_rol(Uxn *u) { Uint8 a = pop8(&u->wst), b = pop8(&u->wst); push8(&u->wst, b << (a % 8)); }
-void op_ror(Uxn *u) { Uint8 a = pop8(&u->wst), b = pop8(&u->wst); push8(&u->wst, b >> (a % 8)); }
+void op_and(Uxn *u) { Uint8 a = pop8(u->src), b = pop8(u->src); push8(u->src, b & a); }
+void op_xor(Uxn *u) { Uint8 a = pop8(u->src), b = pop8(u->src); push8(u->src, b | a); }
+void op_rol(Uxn *u) { Uint8 a = pop8(u->src), b = pop8(u->src); push8(u->src, b << (a % 8)); }
+void op_ror(Uxn *u) { Uint8 a = pop8(u->src), b = pop8(u->src); push8(u->src, b >> (a % 8)); }
 /* Stack */
-void op_pop(Uxn *u) { pop8(&u->wst); }
-void op_dup(Uxn *u) { push8(&u->wst, peek8(&u->wst, 0)); }
-void op_swp(Uxn *u) { Uint8 b = pop8(&u->wst), a = pop8(&u->wst); push8(&u->wst, b); push8(&u->wst, a); }
-void op_ovr(Uxn *u) { push8(&u->wst, peek8(&u->wst, 1)); }
-void op_rot(Uxn *u) { Uint8 c = pop8(&u->wst), b = pop8(&u->wst), a = pop8(&u->wst); push8(&u->wst, b); push8(&u->wst, c); push8(&u->wst, a); }
-void op_cln(Uxn *u) { push8(&u->wst, peek8(&u->rst, 0)); }
-void op_wsr(Uxn *u) { Uint8 a = pop8(&u->wst); push8(&u->rst, a); }
-void op_rsw(Uxn *u) { Uint8 a = pop8(&u->rst); push8(&u->wst, a); }
+void op_pop(Uxn *u) { pop8(u->src); }
+void op_dup(Uxn *u) { push8(u->src, peek8(u->src, 0)); }
+void op_swp(Uxn *u) { Uint8 b = pop8(u->src), a = pop8(u->src); push8(u->src, b); push8(u->src, a); }
+void op_ovr(Uxn *u) { push8(u->src, peek8(u->src, 1)); }
+void op_rot(Uxn *u) { Uint8 c = pop8(u->src), b = pop8(u->src), a = pop8(u->src); push8(u->src, b); push8(u->src, c); push8(u->src, a); }
+void op_cln(Uxn *u) { push8(u->src, peek8(u->dst, 0)); }
+void op_wsr(Uxn *u) { Uint8 a = pop8(u->src); push8(u->dst, a); }
+void op_rsw(Uxn *u) { Uint8 a = pop8(u->dst); push8(u->src, a); }
 /* Arithmetic */
-void op_add(Uxn *u) { Uint8 a = pop8(&u->wst), b = pop8(&u->wst); push8(&u->wst, getflag(&u->status, FLAG_SIGN) ? (Sint8)b + (Sint8)a : b + a); }
-void op_sub(Uxn *u) { Uint8 a = pop8(&u->wst), b = pop8(&u->wst); push8(&u->wst, getflag(&u->status, FLAG_SIGN) ? (Sint8)b - (Sint8)a : b - a); }
-void op_mul(Uxn *u) { Uint8 a = pop8(&u->wst), b = pop8(&u->wst); push8(&u->wst, getflag(&u->status, FLAG_SIGN) ? (Sint8)b * (Sint8)a : b * a); }
-void op_div(Uxn *u) { Uint8 a = pop8(&u->wst), b = pop8(&u->wst); push8(&u->wst, getflag(&u->status, FLAG_SIGN) ? (Sint8)b / (Sint8)a : b / a); }
-void op_equ(Uxn *u) { Uint8 a = pop8(&u->wst), b = pop8(&u->wst); push8(&u->wst, getflag(&u->status, FLAG_SIGN) ? (Sint8)b == (Sint8)a : b == a); }
-void op_neq(Uxn *u) { Uint8 a = pop8(&u->wst), b = pop8(&u->wst); push8(&u->wst, getflag(&u->status, FLAG_SIGN) ? (Sint8)b != (Sint8)a : b != a); }
-void op_gth(Uxn *u) { Uint8 a = pop8(&u->wst), b = pop8(&u->wst); push8(&u->wst, getflag(&u->status, FLAG_SIGN) ? (Sint8)b > (Sint8)a : b > a); }
-void op_lth(Uxn *u) { Uint8 a = pop8(&u->wst), b = pop8(&u->wst); push8(&u->wst, getflag(&u->status, FLAG_SIGN) ? (Sint8)b < (Sint8)a : b < a); }
+void op_add(Uxn *u) { Uint8 a = pop8(u->src), b = pop8(u->src); push8(u->src, b + a); }
+void op_sub(Uxn *u) { Uint8 a = pop8(u->src), b = pop8(u->src); push8(u->src, b - a); }
+void op_mul(Uxn *u) { Uint8 a = pop8(u->src), b = pop8(u->src); push8(u->src, b * a); }
+void op_div(Uxn *u) { Uint8 a = pop8(u->src), b = pop8(u->src); push8(u->src, b / a); }
+void op_equ(Uxn *u) { Uint8 a = pop8(u->src), b = pop8(u->src); push8(u->src, b == a); }
+void op_neq(Uxn *u) { Uint8 a = pop8(u->src), b = pop8(u->src); push8(u->src, b != a); }
+void op_gth(Uxn *u) { Uint8 a = pop8(u->src), b = pop8(u->src); push8(u->src, b > a); }
+void op_lth(Uxn *u) { Uint8 a = pop8(u->src), b = pop8(u->src); push8(u->src, b < a); }
 /* --- */
 void op_lit16(Uxn *u) { u->literal += 2; }
-void op_nop16(Uxn *u) { printf("%04x\n", pop16(&u->wst)); }
-void op_jmp16(Uxn *u) { u->ram.ptr = pop16(&u->wst); }
-void op_jsr16(Uxn *u) { push16(&u->rst, u->ram.ptr); u->ram.ptr = pop16(&u->wst); }
-void op_ldr16(Uxn *u) { Uint16 a = pop16(&u->wst); push16(&u->wst, mempeek16(u, a)); }
-void op_str16(Uxn *u) { Uint16 a = pop16(&u->wst); Uint16 b = pop16(&u->wst); mempoke16(u, a, b); }
-void op_and16(Uxn *u) { Uint16 a = pop16(&u->wst), b = pop16(&u->wst); push16(&u->wst, b & a); }
-void op_xor16(Uxn *u) { Uint16 a = pop16(&u->wst), b = pop16(&u->wst); push16(&u->wst, b ^ a); }
-void op_rol16(Uxn *u) { Uint16 a = pop16(&u->wst), b = pop16(&u->wst); push16(&u->wst, b << (a % 16)); }
-void op_ror16(Uxn *u) { Uint16 a = pop16(&u->wst), b = pop16(&u->wst); push16(&u->wst, b >> (a % 16)); }
+void op_nop16(Uxn *u) { printf("%04x\n", pop16(u->src)); }
+void op_jmp16(Uxn *u) { u->ram.ptr = pop16(u->src); }
+void op_jsr16(Uxn *u) { push16(u->dst, u->ram.ptr); u->ram.ptr = pop16(u->src); }
+void op_ldr16(Uxn *u) { Uint16 a = pop16(u->src); push16(u->src, mempeek16(u, a)); }
+void op_str16(Uxn *u) { Uint16 a = pop16(u->src); Uint16 b = pop16(u->src); mempoke16(u, a, b); }
+void op_and16(Uxn *u) { Uint16 a = pop16(u->src), b = pop16(u->src); push16(u->src, b & a); }
+void op_xor16(Uxn *u) { Uint16 a = pop16(u->src), b = pop16(u->src); push16(u->src, b ^ a); }
+void op_rol16(Uxn *u) { Uint16 a = pop16(u->src), b = pop16(u->src); push16(u->src, b << (a % 16)); }
+void op_ror16(Uxn *u) { Uint16 a = pop16(u->src), b = pop16(u->src); push16(u->src, b >> (a % 16)); }
 /* Stack(16-bits) */
-void op_pop16(Uxn *u) { pop16(&u->wst); }
-void op_dup16(Uxn *u) { push16(&u->wst, peek16(&u->wst, 0)); }
-void op_swp16(Uxn *u) { Uint16 b = pop16(&u->wst), a = pop16(&u->wst); push16(&u->wst, b); push16(&u->wst, a); }
-void op_ovr16(Uxn *u) { push16(&u->wst, peek16(&u->wst, 1)); }
-void op_rot16(Uxn *u) { Uint16 c = pop16(&u->wst), b = pop16(&u->wst), a = pop16(&u->wst); push16(&u->wst, b); push16(&u->wst, c); push16(&u->wst, a); }
-void op_cln16(Uxn *u) { push16(&u->wst, peek16(&u->rst, 0)); }
-void op_wsr16(Uxn *u) { Uint16 a = pop16(&u->wst); push16(&u->rst, a); }
-void op_rsw16(Uxn *u) { Uint16 a = pop16(&u->rst); push16(&u->wst, a); }
+void op_pop16(Uxn *u) { pop16(u->src); }
+void op_dup16(Uxn *u) { push16(u->src, peek16(u->src, 0)); }
+void op_swp16(Uxn *u) { Uint16 b = pop16(u->src), a = pop16(u->src); push16(u->src, b); push16(u->src, a); }
+void op_ovr16(Uxn *u) { push16(u->src, peek16(u->src, 1)); }
+void op_rot16(Uxn *u) { Uint16 c = pop16(u->src), b = pop16(u->src), a = pop16(u->src); push16(u->src, b); push16(u->src, c); push16(u->src, a); }
+void op_cln16(Uxn *u) { push16(u->src, peek16(u->dst, 0)); }
+void op_wsr16(Uxn *u) { Uint16 a = pop16(u->src); push16(u->dst, a); }
+void op_rsw16(Uxn *u) { Uint16 a = pop16(u->dst); push16(u->src, a); }
 /* Arithmetic(16-bits) */
-void op_add16(Uxn *u) { Uint16 a = pop16(&u->wst), b = pop16(&u->wst); push16(&u->wst, getflag(&u->status, FLAG_SIGN) ? (Sint16)b + (Sint16)a : b + a); }
-void op_sub16(Uxn *u) { Uint16 a = pop16(&u->wst), b = pop16(&u->wst); push16(&u->wst, getflag(&u->status, FLAG_SIGN) ? (Sint16)b - (Sint16)a : b - a); }
-void op_mul16(Uxn *u) { Uint16 a = pop16(&u->wst), b = pop16(&u->wst); push16(&u->wst, getflag(&u->status, FLAG_SIGN) ? (Sint16)b * (Sint16)a : b * a); }
-void op_div16(Uxn *u) { Uint16 a = pop16(&u->wst), b = pop16(&u->wst); push16(&u->wst, getflag(&u->status, FLAG_SIGN) ? (Sint16)b / (Sint16)a : b / a); }
-void op_equ16(Uxn *u) { Uint16 a = pop16(&u->wst), b = pop16(&u->wst); push8(&u->wst, getflag(&u->status, FLAG_SIGN) ? (Sint16)b == (Sint16)a : b == a); }
-void op_neq16(Uxn *u) { Uint16 a = pop16(&u->wst), b = pop16(&u->wst); push8(&u->wst, getflag(&u->status, FLAG_SIGN) ? (Sint16)b != (Sint16)a : b != a); }
-void op_gth16(Uxn *u) { Uint16 a = pop16(&u->wst), b = pop16(&u->wst); push8(&u->wst, getflag(&u->status, FLAG_SIGN) ? (Sint16)b > (Sint16)a : b > a); }
-void op_lth16(Uxn *u) { Uint16 a = pop16(&u->wst), b = pop16(&u->wst); push8(&u->wst, getflag(&u->status, FLAG_SIGN) ? (Sint16)b < (Sint16)a : b < a); }
+void op_add16(Uxn *u) { Uint16 a = pop16(u->src), b = pop16(u->src); push16(u->src, b + a); }
+void op_sub16(Uxn *u) { Uint16 a = pop16(u->src), b = pop16(u->src); push16(u->src, b - a); }
+void op_mul16(Uxn *u) { Uint16 a = pop16(u->src), b = pop16(u->src); push16(u->src, b * a); }
+void op_div16(Uxn *u) { Uint16 a = pop16(u->src), b = pop16(u->src); push16(u->src, b / a); }
+void op_equ16(Uxn *u) { Uint16 a = pop16(u->src), b = pop16(u->src); push8(u->src, b == a); }
+void op_neq16(Uxn *u) { Uint16 a = pop16(u->src), b = pop16(u->src); push8(u->src, b != a); }
+void op_gth16(Uxn *u) { Uint16 a = pop16(u->src), b = pop16(u->src); push8(u->src, b > a); }
+void op_lth16(Uxn *u) { Uint16 a = pop16(u->src), b = pop16(u->src); push8(u->src, b < a); }
 
 void (*ops[])(Uxn *u) = {
 	op_brk, op_nop, op_lit, op_ldr, op_str, op_jmp, op_jsr, op_rtn, 
@@ -139,25 +139,32 @@ lituxn(Uxn *u, Uint8 instr)
 int
 opcuxn(Uxn *u, Uint8 instr)
 {
-	Uint8 op = instr & 0x1f, fcond;
+	Uint8 op = instr & 0x1f, fcond, freturn;
 	setflag(&u->status, FLAG_SHORT, (instr >> 5) & 1);
-	setflag(&u->status, FLAG_SIGN, (instr >> 6) & 1);
+	setflag(&u->status, FLAG_RETURN, (instr >> 6) & 1);
 	setflag(&u->status, FLAG_COND, (instr >> 7) & 1);
 	fcond = getflag(&u->status, FLAG_COND);
+	freturn = getflag(&u->status, FLAG_RETURN);
+	u->src = freturn ? &u->rst : &u->wst;
+	u->dst = freturn ? &u->wst : &u->rst;
 	if(getflag(&u->status, FLAG_SHORT))
 		op += 32;
-	if(u->wst.ptr < opr[op][0] || (fcond && u->wst.ptr < 1))
+	if(u->src->ptr < opr[op][0] || (fcond && u->src->ptr < 1))
 		return haltuxn(u, "Working-stack underflow", op);
-	if(u->wst.ptr + opr[op][1] - opr[op][0] >= 255)
+	if(u->src->ptr + opr[op][1] - opr[op][0] >= 255)
 		return haltuxn(u, "Working-stack overflow", instr);
-	if(u->rst.ptr < opr[op][2])
+	if(u->dst->ptr < opr[op][2])
 		return haltuxn(u, "Return-stack underflow", op);
-	if(u->rst.ptr + opr[op][3] - opr[op][2] >= 255)
+	if(u->dst->ptr + opr[op][3] - opr[op][2] >= 255)
 		return haltuxn(u, "Return-stack overflow", instr);
 	if(!fcond || (fcond && pop8(&u->wst)))
 		(*ops[op])(u);
-	else
-		u->wst.ptr -= opr[op][0] - opr[op][1];
+	else {
+		if(freturn)
+			u->src->ptr -= opr[op][2] - opr[op][3];
+		else
+			u->src->ptr -= opr[op][0] - opr[op][1];
+	}
 	return 1;
 }
 
diff --git a/uxn.h b/uxn.h
index 549528c..6b1d35a 100644
--- a/uxn.h
+++ b/uxn.h
@@ -18,7 +18,7 @@ typedef signed short Sint16;
 
 #define FLAG_HALT 0x01
 #define FLAG_SHORT 0x02
-#define FLAG_SIGN 0x04
+#define FLAG_RETURN 0x04
 #define FLAG_COND 0x08
 
 typedef struct {
@@ -39,7 +39,7 @@ typedef struct Device {
 typedef struct {
 	Uint8 literal, status, devices;
 	Uint16 counter, vreset, vframe, verror;
-	Stack wst, rst;
+	Stack wst, rst, *src, *dst;
 	Memory ram;
 	Device dev[16];
 } Uxn;