diff --git a/conftest.py b/conftest.py
new file mode 100644
index 0000000..db21658
--- /dev/null
+++ b/conftest.py
@@ -0,0 +1,3 @@
+def pytest_report_teststatus(report, config):
+    if report.passed and report.when == "call":
+        return report.outcome, "", report.outcome.upper()
diff --git a/gbso/insn.py b/gbso/insn.py
index 8b25aa9..cbc624d 100644
--- a/gbso/insn.py
+++ b/gbso/insn.py
@@ -22,6 +22,7 @@ class LD_R_R(Insn):
 
     def exec(self, cpu: CPU) -> None:
         cpu.set_reg8(self.dst, cpu.get_reg8(self.src))
+        cpu.cycles += 4
 
     def pretty(self) -> str:
         return f"LD {self.dst}, {self.src}"
@@ -34,6 +35,7 @@ class LD_R_N8(Insn):
 
     def exec(self, cpu: CPU) -> None:
         cpu.set_reg8(self.dst, self.imm)
+        cpu.cycles += 8
 
     def pretty(self) -> str:
         return f"LD {self.dst}, {hex(self.imm)}"
@@ -45,6 +47,7 @@ class LD_R_HL(Insn):
 
     def exec(self, cpu: CPU) -> None:
         cpu.set_reg8(self.dst, cpu.get_mem8(cpu.get_reg16(R16.HL)))
+        cpu.cycles += 8
 
     def pretty(self) -> str:
         return f"LD {self.dst}, (HL)"
@@ -55,7 +58,8 @@ class LD_HL_R(Insn):
     src: R8
 
     def exec(self, cpu: CPU) -> None:
-        cpu.set_mem8(cpu.get_reg16(R16.HL), cpu.get_reg8(self.src))
+        cpu.deref_hl_set(cpu.get_reg8(self.src))
+        cpu.cycles += 8
 
     def pretty(self) -> str:
         return f"LD (HL), {self.src}"
@@ -67,6 +71,7 @@ class LD_HL_N(Insn):
 
     def exec(self, cpu: CPU) -> None:
         cpu.set_mem8(cpu.get_reg16(R16.HL), self.imm)
+        cpu.cycles += 12
 
     def pretty(self) -> str:
         return f"LD (HL), {hex(self.imm & 0xff)}"
@@ -76,6 +81,7 @@ class LD_HL_N(Insn):
 class LD_A_BC(Insn):
     def exec(self, cpu: CPU) -> None:
         cpu.set_reg8(R8.A, cpu.get_mem8(cpu.get_reg16(R16.BC)))
+        cpu.cycles += 8
 
     def pretty(self) -> str:
         return "LD A, (BC)"
@@ -85,6 +91,7 @@ class LD_A_BC(Insn):
 class LD_A_DE(Insn):
     def exec(self, cpu: CPU) -> None:
         cpu.set_reg8(R8.A, cpu.get_mem8(cpu.get_reg16(R16.DE)))
+        cpu.cycles += 8
 
     def pretty(self) -> str:
         return "LD A, (DE)"
@@ -96,6 +103,7 @@ class LD_A_NN(Insn):
 
     def exec(self, cpu: CPU) -> None:
         cpu.set_reg8(R8.A, cpu.get_mem8(self.nn))
+        cpu.cycles += 16
 
     def pretty(self) -> str:
         return f"LD A, ({hex(self.nn)})"
@@ -105,12 +113,20 @@ class LD_A_NN(Insn):
 class LD_BC_A(Insn):
     def exec(self, cpu: CPU) -> None:
         cpu.set_mem8(cpu.get_reg16(R16.BC), cpu.get_reg8(R8.A))
+        cpu.cycles += 8
+
+    def pretty(self) -> str:
+        return "LD (BC), A"
 
 
 @dataclass
 class LD_DE_A(Insn):
     def exec(self, cpu: CPU) -> None:
         cpu.set_mem8(cpu.get_reg16(R16.DE), cpu.get_reg8(R8.A))
+        cpu.cycles += 8
+
+    def pretty(self) -> str:
+        return "LD (DE), A"
 
 
 @dataclass
@@ -119,6 +135,10 @@ class LD_NN_A(Insn):
 
     def exec(self, cpu: CPU) -> None:
         cpu.set_mem8(self.nn, cpu.get_reg8(R8.A))
+        cpu.cycles += 16
+
+    def pretty(self) -> str:
+        return f"LD ({hex(self.nn)}), A"
 
 
 @dataclass
@@ -127,6 +147,10 @@ class LD_A_FF_N(Insn):
 
     def exec(self, cpu: CPU) -> None:
         cpu.set_reg8(R8.A, cpu.get_mem8(0xFF00 + self.n))
+        cpu.cycles += 12
+
+    def pretty(self) -> str:
+        return f"LD A, (0xFF00 + {hex(self.n)})"
 
 
 @dataclass
@@ -135,18 +159,30 @@ class LD_FF_N_A(Insn):
 
     def exec(self, cpu: CPU) -> None:
         cpu.set_mem8(0xFF00 + self.n, cpu.get_reg8(R8.A))
+        cpu.cycles += 12
+
+    def pretty(self) -> str:
+        return f"LD (0xFF00 + {hex(self.n)}), A"
 
 
 @dataclass
 class LD_A_FF_C(Insn):
     def exec(self, cpu: CPU) -> None:
         cpu.set_reg8(R8.A, cpu.get_mem8(0xFF00 + cpu.get_reg8(R8.C)))
+        cpu.cycles += 8
+
+    def pretty(self) -> str:
+        return "LD A, (0xFF00 + C)"
 
 
 @dataclass
 class LD_FF_C_A(Insn):
     def exec(self, cpu: CPU) -> None:
         cpu.set_mem8(0xFF00 + cpu.get_reg8(R8.C), cpu.get_reg8(R8.A))
+        cpu.cycles += 8
+
+    def pretty(self) -> str:
+        return "LD (0xFF00 + C), A"
 
 
 @dataclass
@@ -155,6 +191,10 @@ class LDI_HL_A(Insn):
         hl = cpu.get_reg16(R16.HL)
         cpu.set_mem8(hl, cpu.get_reg8(R8.A))
         cpu.set_reg16(R16.HL, hl + 1)
+        cpu.cycles += 8
+
+    def pretty(self) -> str:
+        return "LDI (HL), A"
 
 
 @dataclass
@@ -163,6 +203,10 @@ class LDI_A_HL(Insn):
         hl = cpu.get_reg16(R16.HL)
         cpu.set_reg8(R8.A, cpu.get_mem8(hl))
         cpu.set_reg16(R16.HL, hl + 1)
+        cpu.cycles += 8
+
+    def pretty(self) -> str:
+        return "LDI A, (HL)"
 
 
 @dataclass
@@ -171,6 +215,10 @@ class LDD_HL_A(Insn):
         hl = cpu.get_reg16(R16.HL)
         cpu.set_mem8(hl, cpu.get_reg8(R8.A))
         cpu.set_reg16(R16.HL, hl - 1)
+        cpu.cycles += 8
+
+    def pretty(self) -> str:
+        return "LDD (HL), A"
 
 
 @dataclass
@@ -179,6 +227,10 @@ class LDD_A_HL(Insn):
         hl = cpu.get_reg16(R16.HL)
         cpu.set_reg8(R8.A, cpu.get_mem8(hl))
         cpu.set_reg16(R16.HL, hl - 1)
+        cpu.cycles += 8
+
+    def pretty(self) -> str:
+        return "LDD A, (HL)"
 
 
 @dataclass
diff --git a/tests/test_insn.py b/tests/test_insn.py
new file mode 100644
index 0000000..244016e
--- /dev/null
+++ b/tests/test_insn.py
@@ -0,0 +1,232 @@
+import pytest
+
+from gbso.insn import *
+from gbso.regs import *
+
+
+def n8():
+    for i in range(0xFF + 1):
+        yield i
+
+
+def n16():
+    for i in range(0xFFFF + 1):
+        yield i
+
+
+@pytest.mark.parametrize("dst,src", [(x, y) for x in R8 for y in R8])
+def test_ld_r_r(dst, src):
+    cpu = CPU()
+    cpu.set_reg8(src, 0x7F)
+
+    LD_R_R(dst, src).exec(cpu)
+    assert cpu.get_reg8(src) == 0x7F
+    assert cpu.cycles == 4
+
+
+@pytest.mark.parametrize("r,imm", [(r, imm) for r in R8 for imm in n8()])
+def test_ld_r_n8(r, imm):
+    cpu = CPU()
+
+    LD_R_N8(r, 0x7F).exec(cpu)
+    assert cpu.get_reg8(r) == 0x7F
+    assert cpu.cycles == 8
+
+
+@pytest.mark.parametrize("r", R8)
+def test_ld_r_hl(r):
+    cpu = CPU()
+    cpu.set_reg16(R16.HL, 0x1234)
+    cpu.set_mem8(0x1234, 0x7F)
+
+    LD_R_HL(r).exec(cpu)
+    assert cpu.get_reg8(r) == 0x7F
+    assert cpu.cycles == 8
+
+
+@pytest.mark.parametrize("r", R8)
+def test_ld_hl_r(r):
+    cpu = CPU()
+    cpu.set_reg8(r, 0x7F)
+    cpu.set_reg16(R16.HL, 0x1234)
+
+    LD_HL_R(r).exec(cpu)
+    hl = cpu.deref_hl()
+
+    if r == R8.H:
+        assert hl == 0x12
+    elif r == R8.L:
+        assert hl == 0x34
+    else:
+        assert hl == 0x7F
+
+    assert cpu.cycles == 8
+
+
+@pytest.mark.parametrize("imm", n8())
+def test_ld_hl_n8(imm):
+    cpu = CPU()
+    cpu.set_reg16(R16.HL, 0x1234)
+
+    LD_HL_N(imm).exec(cpu)
+
+    assert cpu.deref_hl() == imm
+    assert cpu.cycles == 12
+
+
+def test_ld_a_bc():
+    cpu = CPU()
+    cpu.set_reg16(R16.BC, 0x1234)
+    cpu.set_mem8(0x1234, 0x7F)
+
+    LD_A_BC().exec(cpu)
+
+    assert cpu.get_reg8(R8.A) == 0x7F
+    assert cpu.cycles == 8
+
+
+def test_ld_a_de():
+    cpu = CPU()
+    cpu.set_reg16(R16.DE, 0x1234)
+    cpu.set_mem8(0x1234, 0x7F)
+
+    LD_A_DE().exec(cpu)
+
+    assert cpu.get_reg8(R8.A) == 0x7F
+    assert cpu.cycles == 8
+
+
+# @pytest.mark.parametrize("nn", n16())
+def test_ld_a_nn(nn=0x1234):
+    cpu = CPU()
+    cpu.set_mem8(nn, 0x7F)
+
+    LD_A_NN(nn).exec(cpu)
+
+    assert cpu.get_reg8(R8.A) == 0x7F
+    assert cpu.cycles == 16
+
+
+def test_ld_bc_a():
+    cpu = CPU()
+    cpu.set_reg16(R16.BC, 0x1234)
+    cpu.set_reg8(R8.A, 0x7F)
+
+    LD_BC_A().exec(cpu)
+
+    assert cpu.get_mem8(0x1234) == 0x7F
+    assert cpu.cycles == 8
+
+
+def test_ld_de_a():
+    cpu = CPU()
+    cpu.set_reg16(R16.DE, 0x1234)
+    cpu.set_reg8(R8.A, 0x7F)
+
+    LD_DE_A().exec(cpu)
+
+    assert cpu.get_mem8(0x1234) == 0x7F
+    assert cpu.cycles == 8
+
+
+# @pytest.mark.parametrize("nn", n16())
+def test_ld_nn_a(nn=0x1234):
+    cpu = CPU()
+    cpu.set_reg8(R8.A, 0x7F)
+
+    LD_NN_A(nn).exec(cpu)
+
+    assert cpu.get_mem8(nn) == 0x7F
+    assert cpu.cycles == 16
+
+
+@pytest.mark.parametrize("n", n8())
+def test_ld_a_ff_n(n):
+    cpu = CPU()
+    cpu.set_mem8(0xFF00 + n, 0x7F)
+
+    LD_A_FF_N(n).exec(cpu)
+
+    assert cpu.get_reg8(R8.A) == 0x7F
+    assert cpu.cycles == 12
+
+
+@pytest.mark.parametrize("n", n8())
+def test_ld_ff_n_a(n):
+    cpu = CPU()
+    cpu.set_reg8(R8.A, 0x7F)
+
+    LD_FF_N_A(n).exec(cpu)
+
+    assert cpu.get_mem8(0xFF00 + n) == 0x7F
+    assert cpu.cycles == 12
+
+
+def test_ld_a_ff_c():
+    cpu = CPU()
+    cpu.set_reg8(R8.C, 0x12)
+    cpu.set_mem8(0xFF12, 0x7F)
+
+    LD_A_FF_C().exec(cpu)
+
+    assert cpu.get_reg8(R8.A) == 0x7F
+    assert cpu.cycles == 8
+
+
+def test_ld_ff_c_a():
+    cpu = CPU()
+    cpu.set_reg8(R8.C, 0x12)
+    cpu.set_reg8(R8.A, 0x7F)
+
+    LD_FF_C_A().exec(cpu)
+
+    assert cpu.get_mem8(0xFF12) == 0x7F
+    assert cpu.cycles == 8
+
+
+def test_ldi_hl_a():
+    cpu = CPU()
+    cpu.set_reg8(R8.A, 0x7F)
+    cpu.set_reg16(R16.HL, 0x1234)
+
+    LDI_HL_A().exec(cpu)
+
+    assert cpu.get_reg16(R16.HL) == 0x1235
+    assert cpu.get_mem8(0x1234) == 0x7F
+    assert cpu.cycles == 8
+
+
+def test_ldi_a_hl():
+    cpu = CPU()
+    cpu.set_mem8(0x1234, 0x7F)
+    cpu.set_reg16(R16.HL, 0x1234)
+
+    LDI_A_HL().exec(cpu)
+
+    assert cpu.get_reg16(R16.HL) == 0x1235
+    assert cpu.get_reg8(R8.A) == 0x7F
+    assert cpu.cycles == 8
+
+
+def test_ldd_hl_a():
+    cpu = CPU()
+    cpu.set_reg8(R8.A, 0x7F)
+    cpu.set_reg16(R16.HL, 0x1234)
+
+    LDD_HL_A().exec(cpu)
+
+    assert cpu.get_reg16(R16.HL) == 0x1233
+    assert cpu.get_mem8(0x1234) == 0x7F
+    assert cpu.cycles == 8
+
+
+def test_ldd_a_hl():
+    cpu = CPU()
+    cpu.set_mem8(0x1234, 0x7F)
+    cpu.set_reg16(R16.HL, 0x1234)
+
+    LDD_A_HL().exec(cpu)
+
+    assert cpu.get_reg16(R16.HL) == 0x1233
+    assert cpu.get_reg8(R8.A) == 0x7F
+    assert cpu.cycles == 8