diff --git a/src/gbemu/cpu.c b/src/gbemu/cpu.c
index f92dd00..282d3f9 100644
--- a/src/gbemu/cpu.c
+++ b/src/gbemu/cpu.c
@@ -150,7 +150,6 @@ static const unsigned int cb_op_cycles[256] = { 0 };
 
 #define LD_D16(cpu, reg)						\
 	do {								\
-		printf("LD_D16 " #reg "\n");				\
 		(reg) = (cpu)->mem_read((cpu), (cpu)->pc++);		\
 		(reg) |= (uint16_t) (cpu)->mem_read((cpu), (cpu)->pc++) << 8; \
 	} while (0)
@@ -233,17 +232,18 @@ static const unsigned int cb_op_cycles[256] = { 0 };
 #define CP_8(cpu, dst, src)				\
 	do {						\
 		uint8_t _tmp;				\
+		uint8_t _dst = dst;			\
+		uint8_t _src = src;			\
 		(cpu)->nf = 1;				\
-		(cpu)->hf = CALC_H_SUB((dst), (src));	\
-		(cpu)->cf = CALC_C_SUB((dst), (src));	\
-		_tmp = (dst) - (src);			\
+		(cpu)->hf = CALC_H_SUB((_dst), (_src));	\
+		(cpu)->cf = CALC_C_SUB((_dst), (_src));	\
+		_tmp = (_dst) - (_src);			\
 		(cpu)->zf = ((_tmp) == 0);		\
 	} while (0)
 
-
 #define JR_8(cpu)					\
 	do {						\
-		val_16 = cpu->mem_read(cpu, cpu->pc);	\
+		val_16 = cpu->mem_read(cpu, cpu->pc++);	\
 		/* sign-extend */			\
 		val_16 |= GET_BIT(val_16, 7) ? 0xff00 : 0; \
 		cpu->pc += val_16;				\
@@ -253,7 +253,7 @@ static const unsigned int cb_op_cycles[256] = { 0 };
 #define POP_16(cpu, dst)						\
 	do {								\
 		(dst) = (cpu)->mem_read((cpu), (cpu)->sp++);		\
-		(dst) = (uint16_t) (cpu)->mem_read((cpu), (cpu)->sp++) << 8; \
+		(dst) |= (uint16_t) (cpu)->mem_read((cpu), (cpu)->sp++) << 8; \
 	} while (0)
 
 #define PUSH_16(cpu, src)					\
@@ -264,18 +264,20 @@ static const unsigned int cb_op_cycles[256] = { 0 };
 
 #define CALL(cpu)							\
 	do {								\
+		uint16_t pc_base;					\
 		cpu->mem_write(cpu, --cpu->sp, cpu->pc);		\
 		cpu->mem_write(cpu, --cpu->sp, cpu->pc >> 8);		\
-		cpu->pc = cpu->mem_read(cpu, cpu->pc++);		\
-		cpu->pc |= (uint16_t) cpu->mem_read(cpu, cpu->pc) << 8; \
-		cpu->pc++;						\
+		pc_base = cpu->mem_read(cpu, cpu->pc++);		\
+		pc_base |= (uint16_t) cpu->mem_read(cpu, cpu->pc) << 8; \
+		cpu->pc = pc_base;						\
 	} while (0)
 
 #define RET(cpu)							\
 	do {								\
 		(cpu)->pc = 0;						\
-		(cpu)->pc |= (uint16_t) (cpu)->mem_read((cpu), (cpu)->sp--) << 8; \
-		(cpu)->pc |= (cpu)->mem_read((cpu), (cpu)->sp--);	\
+		(cpu)->pc |= (uint16_t) (cpu)->mem_read((cpu), (cpu)->sp++) << 8; \
+		(cpu)->pc |= (cpu)->mem_read((cpu), (cpu)->sp++);	\
+		(cpu)->pc += 2;						\
 	} while (0)
 
 #define RST(cpu, n)							\
@@ -283,7 +285,26 @@ static const unsigned int cb_op_cycles[256] = { 0 };
 		(cpu)->mem_write((cpu), (cpu)->sp++, (cpu)->pc >> 8);	\
 		(cpu)->mem_write((cpu), (cpu)->sp++, (cpu)->pc);	\
 		(cpu)->pc = (n);					\
-	} while (0);
+	} while (0)
+
+#define BIT(cpu, reg, bit)			\
+	do {					\
+		(cpu)->zf = !GET_BIT(reg, bit);	\
+		(cpu)->nf = 0;			\
+		(cpu)->hf = 1;			\
+	} while (0)
+
+#define RL(cpu, reg)				\
+	do {					\
+		uint8_t ci = (cpu)->cf;		\
+		uint8_t tmp = reg;		\
+		(cpu)->nf = 0;			\
+		(cpu)->hf = 0;			\
+		(cpu)->cf = GET_BIT(tmp, 7);	\
+		tmp = tmp << 1;			\
+		(tmp) |= ci;			\
+		(cpu)->zf = (tmp) == 0;		\
+	} while (0)
 
 void lr35902_cycle(struct lr35902_state *cpu)
 {
@@ -292,16 +313,19 @@ void lr35902_cycle(struct lr35902_state *cpu)
 	uint8_t val;
 	uint16_t val_16;
 
+	cpu->metrics.cycles++;
+
 	if (cpu->stall_cycles > 0) {
 		cpu->stall_cycles--;
+		if (cpu->stall_cycles == 0) {
+			cpu->metrics.retired_instrs++;
+		}
 		return;
 	}
 
 	inst = cpu->mem_read(cpu, cpu->pc++);
 	cycles = op_cycles[inst];
 
-	printf("running command\n");
-
 	switch (inst) {
 	case 0x00: /* NOP */
 		break;
@@ -948,7 +972,9 @@ void lr35902_cycle(struct lr35902_state *cpu)
 			cpu->pc += 2;
 		}
 		break;
-	case 0xc5:
+	case 0xc5: /* PUSH BC */
+		PUSH_16(cpu, cpu->bc);
+		break;
 	case 0xc6:
 	case 0xc7: /* RST 00 */
 		RST(cpu, 0);
@@ -983,13 +1009,29 @@ void lr35902_cycle(struct lr35902_state *cpu)
 		case 0x0e:
 		case 0x0f:
 		case 0x10:
+			RL(cpu, cpu->b);
+			break;
 		case 0x11:
+			RL(cpu, cpu->c);
+			break;
 		case 0x12:
+			RL(cpu, cpu->d);
+			break;
 		case 0x13:
+			RL(cpu, cpu->e);
+			break;
 		case 0x14:
+			RL(cpu, cpu->h);
+			break;
 		case 0x15:
+			RL(cpu, cpu->l);
+			break;
 		case 0x16:
+			ASSERT(0);
+			break;
 		case 0x17:
+			RL(cpu, cpu->a);
+			break;
 		case 0x18:
 		case 0x19:
 		case 0x1a:
@@ -1090,10 +1132,16 @@ void lr35902_cycle(struct lr35902_state *cpu)
 		case 0x79:
 		case 0x7a:
 		case 0x7b:
+			assert(0);
 		case 0x7c:
+			BIT(cpu, cpu->h, 7);
+			break;
 		case 0x7d:
 		case 0x7e:
+			assert(0);
 		case 0x7f:
+			BIT(cpu, cpu->a, 7);
+			break;
 		case 0x80:
 		case 0x81:
 		case 0x82:
@@ -1258,7 +1306,9 @@ void lr35902_cycle(struct lr35902_state *cpu)
 			cpu->pc += 2;
 		}
 		break;
-	case 0xd5:
+	case 0xd5:/* PUSH DE */
+		PUSH_16(cpu, cpu->de);
+		break;
 	case 0xd6:
 	case 0xd7: /* RST 0x10 */
 		RST(cpu, 0x10);
@@ -1297,14 +1347,20 @@ void lr35902_cycle(struct lr35902_state *cpu)
 		break;
 	case 0xe4: /* UNDEF */
 		break;
-	case 0xe5:
+	case 0xe5: /* PUSH HL */
+		PUSH_16(cpu, cpu->hl);
+		break;
 	case 0xe6:
 	case 0xe7: /* RST 0x20 */
 		RST(cpu, 0x20);
 		break;
 	case 0xe8:
 	case 0xe9:
-	case 0xea:
+	case 0xea: /* LD (a16), A */
+		val_16 = cpu->mem_read(cpu, cpu->pc++);
+		val_16 |= cpu->mem_read(cpu, cpu->pc++) << 8;
+		cpu->mem_write(cpu, val_16, cpu->a);
+		break;
 	case 0xeb: /* UNDEF */
 		break;
 	case 0xec: /* UNDEF */
@@ -1328,7 +1384,9 @@ void lr35902_cycle(struct lr35902_state *cpu)
 	case 0xf3:
 	case 0xf4: /* UNDEF */
 		break;
-	case 0xf5:
+	case 0xf5: /* PUSH AF */
+		PUSH_16(cpu, cpu->af);
+		break;
 	case 0xf6:
 	case 0xf7: /* RST 0x30 */
 		RST(cpu, 0x30);
@@ -1341,7 +1399,9 @@ void lr35902_cycle(struct lr35902_state *cpu)
 		break;
 	case 0xfd: /* UNDEF */
 		break;
-	case 0xfe:
+	case 0xfe: /* CP d8 */
+		CP_8(cpu, cpu->a, cpu->mem_read(cpu, cpu->pc++));
+		break;
 	case 0xff: /* RST 0x38 */
 		RST(cpu, 0x38);
 		break;
diff --git a/src/gbemu/cpu.h b/src/gbemu/cpu.h
index 2187349..c2676d0 100644
--- a/src/gbemu/cpu.h
+++ b/src/gbemu/cpu.h
@@ -80,7 +80,7 @@ struct lr35902_state {
 		uint8_t regs_8[NUM_LR35902_REGS_8];
 
 		struct {
-			uint8_t b, c, d, e, h, l, a, f_dummy;
+			uint8_t c, b, e, d, l, h, f_dummy, a;
 		};
 	};