diff --git a/src/uxn.c b/src/uxn.c
index 967040d..c907e20 100644
--- a/src/uxn.c
+++ b/src/uxn.c
@@ -11,84 +11,88 @@ THE SOFTWARE IS PROVIDED "AS IS" AND THE AUTHOR DISCLAIMS ALL WARRANTIES
 WITH REGARD TO THIS SOFTWARE.
 */
 
-/*	a,b,c: general use.  bs: byte/short bool. src, dst: stack ptrs, swapped in return mode.
-	pc: program counter. sp: ptr to src stack ptr. kptr: "keep" mode copy of src stack ptr.
-	x,y: macro in params. d: macro in device. j: macro temp variables. o: macro out param. */
+/*	a,b,c: general use. m2: byte/short mode flag.
+	tsp: macro temp stack ptr. ksp: "keep" mode copy of stack ptr. sp: ptr to stack ptr.
+	x,y: macro in params. o: macro out param. */
 
 #define HALT(c) { return uxn_halt(u, instr, (c), pc - 1); }
-#define JUMP(x) { if(bs) pc = (x); else pc += (Sint8)(x); }
-#define PUSH8(s, x) { if(s->ptr == 0xff) HALT(2) s->dat[s->ptr++] = (x); }
-#define PUSH16(s, x) { if((j = s->ptr) >= 0xfe) HALT(2) k = (x); s->dat[j] = k >> 8; s->dat[j + 1] = k; s->ptr = j + 2; }
-#define PUSH(s, x) { if(bs) { PUSH16(s, (x)) } else { PUSH8(s, (x)) } }
-#define POP8(o) { if(!(j = *sp)) HALT(1) o = (Uint16)src->dat[--j]; *sp = j; }
-#define POP16(o) { if((j = *sp) <= 1) HALT(1) o = (src->dat[j - 2] << 8) + src->dat[j - 1]; *sp = j - 2; }
-#define POP(o) { if(bs) { POP16(o) } else { POP8(o) } }
-#define POKE(x, y) { if(bs) { u->ram[(x)] = (y) >> 8; u->ram[(x) + 1] = (y); } else { u->ram[(x)] = y; } }
-#define PEEK16(o, x) { o = (u->ram[(x)] << 8) + u->ram[(x) + 1]; }
-#define PEEK(o, x) { if(bs) PEEK16(o, x) else o = u->ram[(x)]; }
-#define DEVR(o, x) { o = u->dei(u, x); if (bs) o = (o << 8) + u->dei(u, (x) + 1); }
-#define DEVW(x, y) { if (bs) { u->deo(u, (x), (y) >> 8); u->deo(u, (x) + 1, (y)); } else { u->deo(u, x, (y)); } }
+#define JUMP(x) { if(m2) pc = (x); else pc += (Sint8)(x); }
+
+#define PUSH8(x) { if(s->ptr == 0xff) HALT(2) s->dat[s->ptr++] = (x); }
+#define PUSH16(x) { if((tsp = s->ptr) >= 0xfe) HALT(2) t = (x); s->dat[tsp] = t >> 8; s->dat[tsp + 1] = t; s->ptr = tsp + 2; }
+#define PUSH(x) { if(m2) { PUSH16(x) } else { PUSH8(x) } }
+
+#define POP8(o) { if(*sp == 0x00) HALT(1) o = s->dat[--*sp]; }
+#define POP16(o) { if((tsp = *sp) <= 0x01) HALT(1) o = s->dat[tsp - 1] | (s->dat[tsp - 2] << 8); *sp = tsp - 2; }
+#define POP(o) { if(m2) { POP16(o) } else { POP8(o) } }
+
+#define POKE(x, y) { if(m2) { t = (y); u->ram[(x)] = t >> 8; u->ram[(x) + 1] = t; } else { u->ram[(x)] = (y); } }
+#define PEEK16(o, x) { o = (u->ram[(x)] << 8) | u->ram[(x) + 1]; }
+#define PEEK(o, x) { if(m2) PEEK16(o, x) else o = u->ram[(x)]; }
+
+#define DEVR(o, x) { o = u->dei(u, x); if(m2) o = (o << 8) | u->dei(u, (x) + 1); }
+#define DEVW(x, y) { if(m2) { u->deo(u, (x), (y) >> 8); u->deo(u, (x) + 1, (y)); } else { u->deo(u, x, (y)); } }
 
 int
 uxn_eval(Uxn *u, Uint16 pc)
 {
-	Uint8 kptr, *sp;
-	Uint16 a, b, c, j, k, bs, instr, opcode;
-	Stack *src, *dst;
+	Uint8 instr, opcode, m2, ksp, tsp, *sp;
+	Uint16 a, b, c, t;
+	Stack *s;
 	if(!pc || u->dev[0x0f]) return 0;
 	for(;;) {
 		instr = u->ram[pc++];
-		/* Return Mode */
-		if(instr & 0x40) { src = u->rst; dst = u->wst; }
-		else { src = u->wst; dst = u->rst; }
-		/* Keep Mode */
-		if(instr & 0x80) { kptr = src->ptr; sp = &kptr; }
-		else sp = &src->ptr;
-		/* Short Mode */
-		bs = instr & 0x20;
 		opcode = instr & 0x1f;
+		/* Short Mode */
+		m2 = instr & 0x20;
+		/* Return Mode */
+		s = instr & 0x40 ? u->rst : u->wst;
+		/* Keep Mode */
+		if(instr & 0x80) { ksp = s->ptr; sp = &ksp; }
+		else sp = &s->ptr;
+		/* Opcodes */
 		switch(opcode - (!opcode * (instr >> 5))) {
-		/* Literals/Calls */
+		/* Immediate */
 		case -0x0: /* BRK */ return 1;
 		case -0x1: /* JCI */ POP8(b) if(!b) { pc += 2; break; }
 		case -0x2: /* JMI */ PEEK16(a, pc) pc += a + 2; break;
-		case -0x3: /* JSI */ PUSH16(u->rst, pc + 2) PEEK16(a, pc) pc += a + 2; break;
+		case -0x3: /* JSI */ s = u->rst; PUSH16(pc + 2) PEEK16(a, pc) pc += a + 2; break;
 		case -0x4: /* LIT */
-		case -0x6: /* LITr */ a = u->ram[pc++]; PUSH8(src, a) break;
+		case -0x6: /* LITr */ a = u->ram[pc++]; PUSH8(a) break;
 		case -0x5: /* LIT2 */
-		case -0x7: /* LIT2r */ PEEK16(a, pc) PUSH16(src, a) pc += 2; break;
+		case -0x7: /* LIT2r */ PEEK16(a, pc) PUSH16(a) pc += 2; break;
 		/* ALU */
-		case 0x01: /* INC */ POP(a) PUSH(src, a + 1) break;
+		case 0x01: /* INC */ POP(a) PUSH(a + 1) break;
 		case 0x02: /* POP */ POP(a) break;
-		case 0x03: /* NIP */ POP(a) POP(b) PUSH(src, a) break;
-		case 0x04: /* SWP */ POP(a) POP(b) PUSH(src, a) PUSH(src, b) break;
-		case 0x05: /* ROT */ POP(a) POP(b) POP(c) PUSH(src, b) PUSH(src, a) PUSH(src, c) break;
-		case 0x06: /* DUP */ POP(a) PUSH(src, a) PUSH(src, a) break;
-		case 0x07: /* OVR */ POP(a) POP(b) PUSH(src, b) PUSH(src, a) PUSH(src, b) break;
-		case 0x08: /* EQU */ POP(a) POP(b) PUSH8(src, b == a) break;
-		case 0x09: /* NEQ */ POP(a) POP(b) PUSH8(src, b != a) break;
-		case 0x0a: /* GTH */ POP(a) POP(b) PUSH8(src, b > a) break;
-		case 0x0b: /* LTH */ POP(a) POP(b) PUSH8(src, b < a) break;
+		case 0x03: /* NIP */ POP(a) POP(b) PUSH(a) break;
+		case 0x04: /* SWP */ POP(a) POP(b) PUSH(a) PUSH(b) break;
+		case 0x05: /* ROT */ POP(a) POP(b) POP(c) PUSH(b) PUSH(a) PUSH(c) break;
+		case 0x06: /* DUP */ POP(a) PUSH(a) PUSH(a) break;
+		case 0x07: /* OVR */ POP(a) POP(b) PUSH(b) PUSH(a) PUSH(b) break;
+		case 0x08: /* EQU */ POP(a) POP(b) PUSH8(b == a) break;
+		case 0x09: /* NEQ */ POP(a) POP(b) PUSH8(b != a) break;
+		case 0x0a: /* GTH */ POP(a) POP(b) PUSH8(b > a) break;
+		case 0x0b: /* LTH */ POP(a) POP(b) PUSH8(b < a) break;
 		case 0x0c: /* JMP */ POP(a) JUMP(a) break;
 		case 0x0d: /* JCN */ POP(a) POP8(b) if(b) JUMP(a) break;
-		case 0x0e: /* JSR */ POP(a) PUSH16(dst, pc) JUMP(a) break;
-		case 0x0f: /* STH */ POP(a) PUSH(dst, a) break;
-		case 0x10: /* LDZ */ POP8(a) PEEK(b, a) PUSH(src, b) break;
+		case 0x0e: /* JSR */ POP(a) s = (instr & 0x40) ? u->wst : u->rst; PUSH16(pc) JUMP(a) break;
+		case 0x0f: /* STH */ POP(a) s = (instr & 0x40) ? u->wst : u->rst; PUSH(a) break;
+		case 0x10: /* LDZ */ POP8(a) PEEK(b, a) PUSH(b) break;
 		case 0x11: /* STZ */ POP8(a) POP(b) POKE(a, b) break;
-		case 0x12: /* LDR */ POP8(a) b = pc + (Sint8)a; PEEK(c, b) PUSH(src, c) break;
-		case 0x13: /* STR */ POP8(a) POP(b) c = pc + (Sint8)a; POKE(c, b) break;
-		case 0x14: /* LDA */ POP16(a) PEEK(b, a) PUSH(src, b) break;
+		case 0x12: /* LDR */ POP8(a) PEEK(b, pc + (Sint8)a) PUSH(b) break;
+		case 0x13: /* STR */ POP8(a) POP(b) POKE(pc + (Sint8)a, b) break;
+		case 0x14: /* LDA */ POP16(a) PEEK(b, a) PUSH(b) break;
 		case 0x15: /* STA */ POP16(a) POP(b) POKE(a, b) break;
-		case 0x16: /* DEI */ POP8(a) DEVR(b, a) PUSH(src, b) break;
+		case 0x16: /* DEI */ POP8(a) DEVR(b, a) PUSH(b) break;
 		case 0x17: /* DEO */ POP8(a) POP(b) DEVW(a, b) break;
-		case 0x18: /* ADD */ POP(a) POP(b) PUSH(src, b + a) break;
-		case 0x19: /* SUB */ POP(a) POP(b) PUSH(src, b - a) break;
-		case 0x1a: /* MUL */ POP(a) POP(b) PUSH(src, (Uint32)b * a) break;
-		case 0x1b: /* DIV */ POP(a) POP(b) if(!a) HALT(3) PUSH(src, b / a) break;
-		case 0x1c: /* AND */ POP(a) POP(b) PUSH(src, b & a) break;
-		case 0x1d: /* ORA */ POP(a) POP(b) PUSH(src, b | a) break;
-		case 0x1e: /* EOR */ POP(a) POP(b) PUSH(src, b ^ a) break;
-		case 0x1f: /* SFT */ POP8(a) POP(b) PUSH(src, b >> (a & 0x0f) << ((a & 0xf0) >> 4)) break;
+		case 0x18: /* ADD */ POP(a) POP(b) PUSH(b + a) break;
+		case 0x19: /* SUB */ POP(a) POP(b) PUSH(b - a) break;
+		case 0x1a: /* MUL */ POP(a) POP(b) PUSH((Uint32)b * a) break;
+		case 0x1b: /* DIV */ POP(a) POP(b) if(!a) HALT(3) PUSH(b / a) break;
+		case 0x1c: /* AND */ POP(a) POP(b) PUSH(b & a) break;
+		case 0x1d: /* ORA */ POP(a) POP(b) PUSH(b | a) break;
+		case 0x1e: /* EOR */ POP(a) POP(b) PUSH(b ^ a) break;
+		case 0x1f: /* SFT */ POP8(a) POP(b) PUSH(b >> (a & 0x0f) << ((a & 0xf0) >> 4)) break;
 		}
 	}
 }
diff --git a/src/uxnasm.c b/src/uxnasm.c
index 6fc5cfb..644b67d 100644
--- a/src/uxnasm.c
+++ b/src/uxnasm.c
@@ -39,7 +39,7 @@ typedef struct {
 	Uint16 llen, mlen, rlen;
 	Label labels[0x400];
 	Macro macros[0x100];
-	Reference refs[0x1000];
+	Reference refs[0x400];
 	char scope[0x40];
 } Program;
 
@@ -175,27 +175,27 @@ makelabel(char *name)
 }
 
 static int
-makereference(char *scope, char *label, Uint16 addr)
+makereference(char *scope, char *label, char rune, Uint16 addr)
 {
 	char subw[0x40], parent[0x40];
 	Reference *r;
 	if(p.rlen == 0x1000)
 		return error("References limit exceeded", label);
 	r = &p.refs[p.rlen++];
-	if(label[1] == '&') {
-		if(!sublabel(subw, scope, label + 2))
+	if(label[0] == '&') {
+		if(!sublabel(subw, scope, label + 1))
 			return error("Invalid sublabel", label);
 		scpy(subw, r->name, 0x40);
 	} else {
-		int pos = spos(label + 1, '/');
+		int pos = spos(label, '/');
 		if(pos > 0) {
 			Label *l;
-			if((l = findlabel(scpy(label + 1, parent, pos))))
+			if((l = findlabel(scpy(label, parent, pos))))
 				l->refs++;
 		}
-		scpy(label + 1, r->name, 0x40);
+		scpy(label, r->name, 0x40);
 	}
-	r->rune = label[0];
+	r->rune = rune;
 	r->addr = addr;
 	return 1;
 }
@@ -305,29 +305,29 @@ parse(char *w, FILE *f)
 			return error("Invalid hex literal", w);
 		break;
 	case '_': /* raw byte relative */
-		makereference(p.scope, w, p.ptr);
+		makereference(p.scope, w + 1, w[0], p.ptr);
 		return writebyte(0xff);
 	case ',': /* literal byte relative */
-		makereference(p.scope, w, p.ptr + 1);
+		makereference(p.scope, w + 1, w[0], p.ptr + 1);
 		return writelitbyte(0xff);
 	case '-': /* raw byte absolute */
-		makereference(p.scope, w, p.ptr);
+		makereference(p.scope, w + 1, w[0], p.ptr);
 		return writebyte(0xff);
 	case '.': /* literal byte zero-page */
-		makereference(p.scope, w, p.ptr + 1);
+		makereference(p.scope, w + 1, w[0], p.ptr + 1);
 		return writelitbyte(0xff);
-	case ':': /* raw short absolute */
-	case '=':
-		makereference(p.scope, w, p.ptr);
+	case ':':
+	case '=': /* raw short absolute */
+		makereference(p.scope, w + 1, w[0], p.ptr);
 		return writeshort(0xffff, 0);
 	case ';': /* literal short absolute */
-		makereference(p.scope, w, p.ptr + 1);
+		makereference(p.scope, w + 1, w[0], p.ptr + 1);
 		return writeshort(0xffff, 1);
 	case '?': /* JCI */
-		makereference(p.scope, w, p.ptr + 1);
+		makereference(p.scope, w + 1, w[0], p.ptr + 1);
 		return writebyte(0x20) && writeshort(0xffff, 0);
 	case '!': /* JMI */
-		makereference(p.scope, w, p.ptr + 1);
+		makereference(p.scope, w + 1, w[0], p.ptr + 1);
 		return writebyte(0x40) && writeshort(0xffff, 0);
 	case '"': /* raw string */
 		i = 0;
@@ -354,7 +354,7 @@ parse(char *w, FILE *f)
 					return 0;
 			return 1;
 		} else {
-			makereference(p.scope, w - 1, p.ptr + 1);
+			makereference(p.scope, w, ' ', p.ptr + 1);
 			return writebyte(0x60) && writeshort(0xffff, 0);
 		}
 	}