diff --git a/src/gbemu/cpu.c b/src/gbemu/cpu.c
index f2f2f55..89c1d22 100644
--- a/src/gbemu/cpu.c
+++ b/src/gbemu/cpu.c
@@ -170,25 +170,25 @@ static const unsigned int cb_extra_cycles[256] = { 0 };
 		(cpu)->zf = (dst) == 0;						\
 	} while (0)
 
-#define ADD_16(cpu, dst, src)							\
-	do {												\
-		uint16_t _src = src;							\
-		(cpu)->nf = 0;									\
+#define ADD_16(cpu, dst, src)								\
+	do {													\
+		uint16_t _src = src;								\
+		(cpu)->nf = 0;										\
 		(cpu)->hf = (dst & 0xfff) + (_src & 0xfff) > 0xfff;	\
-		(cpu)->cf = CALC_C_ADD_16(dst, _src);			\
-		(dst) = _src + dst;								\
-	} while (0)											\
+		(cpu)->cf = CALC_C_ADD_16(dst, _src);				\
+		(dst) = _src + dst;									\
+	} while (0)												\
 
-#define ADC_8(cpu, dst, src)									\
-	do {														\
-		uint16_t _tmp;											\
-		uint16_t _reg = src;										\
-		_tmp = _reg + dst + cpu->cf;							\
-		cpu->nf = 0;											\
-		cpu->hf = _reg + dst + cpu->cf > 0xff;					\
-		cpu->cf = !(_tmp & 0xff);								\
-		(dst) = _tmp;											\
-		cpu->zf = !(dst);										\
+#define ADC_8(cpu, dst, src)										\
+	do {															\
+		uint16_t _dst = dst;										\
+		uint16_t _src = src;										\
+		cpu->nf = 0;												\
+		cpu->hf = ((_src & 0xf) + (_dst & 0xf) + cpu->cf) > 0xf;	\
+		_dst = _src + _dst + cpu->cf;								\
+		cpu->cf = _dst > 0xff;										\
+		cpu->zf = !(_dst & 0xff);									\
+		(dst) = _dst;												\
 	} while (0)
 
 #define SUB_8(cpu, dst, src)					\
@@ -199,20 +199,21 @@ static const unsigned int cb_extra_cycles[256] = { 0 };
 		(cpu)->cf = _reg > dst;					\
 		(cpu)->nf = 1;							\
 		(cpu)->hf = CALC_H_SUB(dst, _reg);		\
-		(cpu)->zf = tmp == 0;					\
+		(cpu)->zf = !tmp;						\
 		dst = tmp;								\
 	} while (0)
 
-#define SBC_8(cpu, dst, src)					\
-	do {										\
-		int tmp;								\
-		uint8_t _reg = src;						\
-		tmp = (dst) - (_reg) - (cpu->cf);		\
-		(cpu)->cf = _reg + (cpu->cf) > dst;		\
-		(cpu)->nf = 1;							\
-		(cpu)->hf = CALC_H_ADD(_reg, dst);		\
-		(cpu)->zf = tmp == 0;					\
-		(dst) = (tmp);							\
+#define SBC_8(cpu, dst, src)								\
+	do {													\
+		int tmp;											\
+		uint8_t _reg = src;									\
+		uint8_t cf = cpu->cf;								\
+		tmp = (dst) - (_reg) - (cpu->cf);					\
+		(cpu)->cf = tmp < 0;								\
+		(cpu)->nf = 1;										\
+		(cpu)->hf = ((dst & 0xf) - (_reg & 0xf) - cf) < 0 ;	\
+		(dst) = (tmp);										\
+		(cpu)->zf = dst == 0;								\
 	} while (0)
 
 #define AND_8(cpu, dst, src)					\
@@ -330,6 +331,37 @@ static const unsigned int cb_extra_cycles[256] = { 0 };
 		reg = tmp;								\
 	} while (0)
 
+#define RLC(cpu, reg)							\
+	do {										\
+		uint8_t tmp = reg;						\
+		(cpu)->nf = 0;							\
+		(cpu)->hf = 0;							\
+		(cpu)->cf = GET_BIT(tmp, 7);			\
+		tmp = tmp << 1;							\
+		tmp |= (cpu)->cf;						\
+		(cpu)->zf = (tmp) == 0;					\
+		reg = tmp;								\
+	} while (0)
+
+#define RRC(cpu, reg)							\
+	do {										\
+		cpu->cf = GET_BIT(reg, 0);				\
+		cpu->nf = 0;							\
+		cpu->hf = 0;							\
+		(reg) >>= 1;							\
+		WRITE_BIT(reg, 7, cpu->cf);				\
+		cpu->zf = (reg == 0);					\
+	} while (0)
+
+#define SLA(cpu, reg)							\
+	do {										\
+		cpu->cf = GET_BIT(reg, 7);				\
+		reg <<= 1;								\
+		cpu->nf = 0;							\
+		cpu->zf = !reg;							\
+		cpu->hf = 0;							\
+	} while (0)
+
 #define SRL(cpu, reg)							\
 	do {										\
 		cpu->cf = GET_BIT(reg, 0);				\
@@ -350,6 +382,16 @@ static const unsigned int cb_extra_cycles[256] = { 0 };
 		cpu->hf = 0;							\
 	} while (0)
 
+#define SRA(cpu, reg)								\
+	do {											\
+		uint8_t val = GET_BIT(reg, 7);				\
+		cpu->cf = GET_BIT(reg, 0);					\
+		cpu->nf = 0;								\
+		cpu->hf = 0;								\
+		reg = (val << 7) | (reg >> 1);				\
+		cpu->zf = !reg;								\
+	} while (0)										\
+
 #define SWAP(cpu, reg)							\
 	do {										\
 		uint8_t tmp = reg << 4;					\
@@ -428,12 +470,8 @@ int lr35902_cycle(struct lr35902_state *cpu)
 		LD_D8(cpu, cpu->c);
 		break;
 	case 0x0f: /* RRCA */
-		cpu->cf = GET_BIT(cpu->a, 0);
-		cpu->nf = 0;
+		RRC(cpu, cpu->a);
 		cpu->zf = 0;
-		cpu->hf = 0;
-		cpu->a >>= 1;
-		WRITE_BIT(cpu->a, 7, cpu->cf);
 		break;
 	case 0x10: /* STOP */
 		ASSERT(0); //TODO: Implement me
@@ -473,8 +511,8 @@ int lr35902_cycle(struct lr35902_state *cpu)
 	case 0x1a: /* LD A, (DE) */
 		cpu->a = gb_mem_read(cpu->memory, cpu->de);
 		break;
-	case 0x1b: /* DEC BC */
-		cpu->bc--;
+	case 0x1b: /* DEC DE */
+		cpu->de--;
 		break;
 	case 0x1c: /* INC E */
 		INC_8(cpu, cpu->e);
@@ -585,7 +623,7 @@ int lr35902_cycle(struct lr35902_state *cpu)
 		break;
 	case 0x36: /* LD (HL), d8 */
 		val = gb_mem_read(cpu->memory, cpu->pc++);
-		gb_mem_write(cpu->memory, cpu->hl++, val);
+		gb_mem_write(cpu->memory, cpu->hl, val);
 		break;
 	case 0x37: /* SCF */
 		cpu->nf = 0;
@@ -618,6 +656,8 @@ int lr35902_cycle(struct lr35902_state *cpu)
 		LD_D8(cpu, cpu->a);
 		break;
 	case 0x3f: /* CCF */
+		cpu->nf = 0;
+		cpu->hf = 0;
 		cpu->cf = !cpu->cf;
 		break;
 	case 0x40: /* LD B,B */
@@ -830,7 +870,7 @@ int lr35902_cycle(struct lr35902_state *cpu)
 		ADD_8(cpu, cpu->a, cpu->a);
 		break;
 	case 0x88: /* ADC A,B */
-		ADC_8(cpu, cpu->a, cpu->a);
+		ADC_8(cpu, cpu->a, cpu->b);
 		break;
 	case 0x89: /* ADC A,C */
 		ADC_8(cpu, cpu->a, cpu->c);
@@ -1054,22 +1094,57 @@ int lr35902_cycle(struct lr35902_state *cpu)
 		cycles = cb_extra_cycles[inst];
 		switch (inst) {
 		case 0x00:
+			RLC(cpu, cpu->b);
+			break;
 		case 0x01:
+			RLC(cpu, cpu->c);
+			break;
 		case 0x02:
+			RLC(cpu, cpu->d);
+			break;
 		case 0x03:
+			RLC(cpu, cpu->e);
+			break;
 		case 0x04:
+			RLC(cpu, cpu->h);
+			break;
 		case 0x05:
+			RLC(cpu, cpu->l);
+			break;
 		case 0x06:
+			val = gb_mem_read(cpu->memory, cpu->hl);
+			RLC(cpu, val);
+			gb_mem_write(cpu->memory, cpu->hl, val);
+			break;
 		case 0x07:
+			RLC(cpu, cpu->a);
+			break;
 		case 0x08:
+			RRC(cpu, cpu->b);
+			break;
 		case 0x09:
+			RRC(cpu, cpu->c);
+			break;
 		case 0x0a:
+			RRC(cpu, cpu->d);
+			break;
 		case 0x0b:
+			RRC(cpu, cpu->e);
+			break;
 		case 0x0c:
+			RRC(cpu, cpu->h);
+			break;
 		case 0x0d:
+			RRC(cpu, cpu->l);
+			break;
 		case 0x0e:
+			val = gb_mem_read(cpu->memory, cpu->hl);
+			RLC(cpu, val);
+			gb_mem_write(cpu->memory, cpu->hl, val);
+			break;
 		case 0x0f:
-			ASSERT(0);
+			RRC(cpu, cpu->a);
+			break;
 		case 0x10:
 			RL(cpu, cpu->b);
 			break;
@@ -1119,21 +1194,57 @@ int lr35902_cycle(struct lr35902_state *cpu)
 			RR(cpu, cpu->a);
 			break;
 		case 0x20:
+			SLA(cpu, cpu->b);
+			break;
 		case 0x21:
+			SLA(cpu, cpu->c);
+			break;
 		case 0x22:
+			SLA(cpu, cpu->d);
+			break;
 		case 0x23:
+			SLA(cpu, cpu->e);
+			break;
 		case 0x24:
+			SLA(cpu, cpu->h);
+			break;
 		case 0x25:
+			SLA(cpu, cpu->l);
+			break;
 		case 0x26:
+			val = gb_mem_read(cpu->memory, cpu->hl);
+			SLA(cpu, val);
+			gb_mem_write(cpu->memory, cpu->hl, val);
+			break;
 		case 0x27:
+			SLA(cpu, cpu->a);
+			break;
 		case 0x28:
+			SRA(cpu, cpu->b);
+			break;
 		case 0x29:
+			SRA(cpu, cpu->c);
+			break;
 		case 0x2a:
+			SRA(cpu, cpu->d);
+			break;
 		case 0x2b:
+			SRA(cpu, cpu->e);
+			break;
 		case 0x2c:
+			SRA(cpu, cpu->h);
+			break;
 		case 0x2d:
+			SRA(cpu, cpu->l);
+			break;
 		case 0x2e:
+			val = gb_mem_read(cpu->memory, cpu->hl);
+			SRA(cpu, val);
+			gb_mem_write(cpu->memory, cpu->hl, val);
+			break;
 		case 0x2f:
+			SRA(cpu, cpu->a);
+			break;
 		case 0x30:
 			SWAP(cpu, cpu->b);
 			break;
@@ -1442,7 +1553,7 @@ int lr35902_cycle(struct lr35902_state *cpu)
 	case 0xcd: /* CALL */
 		CALL(cpu);
 		break;
-	case 0xce:
+	case 0xce: /* ADC A, d8 */
 		val = gb_mem_read(cpu->memory, cpu->pc++);
 		ADC_8(cpu, cpu->a, val);
 		break;
diff --git a/src/gbemu/video.c b/src/gbemu/video.c
index 84c248b..3b42767 100644
--- a/src/gbemu/video.c
+++ b/src/gbemu/video.c
@@ -66,7 +66,7 @@ static void gb_video_screenshot(struct gb_video *video)
 	uint8_t tile;
 	int log = id == 30;
 
-	snprintf(filename, sizeof(filename), "screenshot-%d.bmp", id++);
+	snprintf(filename, sizeof(filename), "screenshot-%05d.bmp", id++);
 
 	if (!(video->lcdcont & 1)) {
 		goto out;