diff --git a/6502_65C02_functional_tests/6502_decimal_test.a65 b/6502_65C02_functional_tests/6502_decimal_test.a65
index 7a76ad1..3a15610 100644
--- a/6502_65C02_functional_tests/6502_decimal_test.a65
+++ b/6502_65C02_functional_tests/6502_decimal_test.a65
@@ -118,6 +118,15 @@ NEXT2   inc N2    ; [6] see text
         lda #0    ; test passed, so store 0 in ERROR
         sta ERROR
 DONE    
+        lda #1
+        cmp ERROR
+        beq _E
+        lda #'P'
+        sta $dc00
+        jmp _E2
+_E      lda #'F'
+        sta $dc00
+_E2
         end_of_test
            
 ; Calculate the actual decimal mode accumulator and flags, the accumulator
diff --git a/6502_65C02_functional_tests/6502_functional_test.a65 b/6502_65C02_functional_tests/6502_functional_test.a65
index 0c7ce35..02533b2 100644
--- a/6502_65C02_functional_tests/6502_functional_test.a65
+++ b/6502_65C02_functional_tests/6502_functional_test.a65
@@ -128,7 +128,7 @@ ram_top = -1
 
 ;disable test decimal mode ADC & SBC, 0=enable, 1=disable,
 ;2=disable including decimal flag in processor status
-disable_decimal = 1
+disable_decimal = 0
 
         noopt       ;do not take shortcuts
 
diff --git a/6502_65C02_functional_tests/report.i65 b/6502_65C02_functional_tests/report.i65
index cefc1c4..4e8d0e3 100644
--- a/6502_65C02_functional_tests/report.i65
+++ b/6502_65C02_functional_tests/report.i65
@@ -50,7 +50,6 @@ rstack  jsr rspace
         inx
         bne rstack
         jsr rcrlf   ;new line
-        jmp rend ; Skip BS
 ;show zero page workspace
         lda #0
         jsr rhex
@@ -64,6 +63,7 @@ rzp     jsr rspace
         cpx #zp_bss
         bne rzp
         jsr rcrlf
+        jmp rend ; Skip Workspace since 65k is large
 ;show absolute workspace
         lda #hi(data_segment)
         jsr rhex
diff --git a/inc/cpu.h b/inc/cpu.h
index 06f877b..595019f 100644
--- a/inc/cpu.h
+++ b/inc/cpu.h
@@ -2,7 +2,6 @@
 #define CPU_H
 
 #include <stdint.h>
-#include <unistd.h>
 
 typedef struct {
   uint16_t ProgrammeCounter;
@@ -29,6 +28,6 @@ typedef struct {
 
 void initCPU(CPU *cpu);
 void resetCPU(CPU *cpu, uint8_t *rom);
-useconds_t runCycle(CPU *cpu, uint8_t *ram, uint8_t *rom);
+uint8_t runCycle(CPU *cpu, uint8_t *ram, uint8_t *rom);
 
 #endif // CPU_H
\ No newline at end of file
diff --git a/inc/opcodes.h b/inc/opcodes.h
index f159d40..5e08cf6 100644
--- a/inc/opcodes.h
+++ b/inc/opcodes.h
@@ -123,7 +123,7 @@
 #define CLV 0xb8     // Clear overflow flag
 #define LDA_YA 0xb9  // Load A from memory, Y-indexed absolute
 #define TSX 0xba     // Transfer stack pointer to X
-#define LDY_XA 0xbb  // Load Y from memory, X-indexed absolute
+#define LDY_XA 0xbc  // Load Y from memory, X-indexed absolute
 #define LDA_XA 0xbd  // Load A from memory, X-indexed absolute
 #define LDX_YA 0xbe  // Load X from memory, Y-indexed absolute
 /* 113 */
diff --git a/src/cpu.c b/src/cpu.c
index 793b549..af10589 100644
--- a/src/cpu.c
+++ b/src/cpu.c
@@ -7,6 +7,55 @@
 #define LO_BYTE(x) ((uint8_t)((x) & 0xFF))
 #define HI_BYTE(x) ((uint8_t)(((x) >> 8) & 0xFF))
 
+static void doADC(CPU *cpu, uint8_t value) {
+  if (cpu->Flags.DecimalMode == 1) {
+    uint16_t al = (cpu->Accumulator & 0x0F) + (value & 0x0F) + cpu->Flags.CarryFlag;
+    if (al >= 0x0A) al = ((al + 0x06) & 0x0F) + 0x10;
+
+    uint16_t sum = (cpu->Accumulator & 0xF0) + (value & 0xF0) + al;
+    if (sum >= 0xA0) sum += 0x60;
+
+    // N/V/Z are hardware-quirky in decimal mode; derive from the binary sum
+    uint16_t binSum = cpu->Accumulator + value + cpu->Flags.CarryFlag;
+    cpu->Flags.OverflowFlag = ((cpu->Accumulator ^ binSum) & (value ^ binSum) & 0x80) != 0;
+    cpu->Flags.NegativeFlag = (sum & 0x80) != 0;
+    cpu->Flags.ZeroFlag = (uint8_t)binSum == 0;
+
+    cpu->Flags.CarryFlag = sum >= 0x100;
+    cpu->Accumulator = (uint8_t)sum;
+  } else {
+    uint16_t result = cpu->Accumulator + value + cpu->Flags.CarryFlag;
+    cpu->Flags.CarryFlag = result > 0xFF;
+    cpu->Flags.ZeroFlag = (uint8_t)result == 0;
+    cpu->Flags.OverflowFlag = ((cpu->Accumulator ^ result) & (value ^ result) & 0x80) != 0;
+    cpu->Flags.NegativeFlag = (result & 0x80) != 0;
+    cpu->Accumulator = (uint8_t)result;
+  }
+}
+
+static void doSBC(CPU *cpu, uint8_t value) {
+  uint8_t carryIn = cpu->Flags.CarryFlag;
+  uint8_t inverted = value ^ 0xFF;
+
+  uint16_t binResult = cpu->Accumulator + inverted + carryIn;
+  cpu->Flags.CarryFlag = binResult > 0xFF;
+  cpu->Flags.ZeroFlag = (uint8_t)binResult == 0;
+  cpu->Flags.OverflowFlag = ((cpu->Accumulator ^ binResult) & (inverted ^ binResult) & 0x80) != 0;
+  cpu->Flags.NegativeFlag = (binResult & 0x80) != 0;
+
+  if (cpu->Flags.DecimalMode == 1) {
+    int16_t al = (cpu->Accumulator & 0x0F) - (value & 0x0F) - (1 - carryIn);
+    if (al < 0) al = ((al - 0x06) & 0x0F) - 0x10;
+
+    int16_t sum = (cpu->Accumulator & 0xF0) - (value & 0xF0) + al;
+    if (sum < 0) sum -= 0x60;
+
+    cpu->Accumulator = (uint8_t)sum;
+  } else {
+    cpu->Accumulator = (uint8_t)binResult;
+  }
+}
+
 void initCPU(CPU *cpu) {
 #pragma clang diagnostic push
 #pragma clang diagnostic ignored "-Wmissing-field-initializers"
@@ -20,7 +69,7 @@ void resetCPU(CPU *cpu, uint8_t *rom) {
   cpu->ProgrammeCounter = COMBINE(high, low);
 }
 
-useconds_t runCycle(CPU *cpu, uint8_t *ram, uint8_t *rom) {
+uint8_t runCycle(CPU *cpu, uint8_t *ram, uint8_t *rom) {
   // Fetch opcode
   uint8_t opcode = readMemory(cpu->ProgrammeCounter++, ram, rom);
 
@@ -144,7 +193,7 @@ useconds_t runCycle(CPU *cpu, uint8_t *ram, uint8_t *rom) {
     }
 
     case BPL: {
-      useconds_t cycles = 2;
+      uint8_t cycles = 2;
 
       // Get offset
       int8_t offset = (int8_t)readMemory(cpu->ProgrammeCounter++, ram, rom);
@@ -399,7 +448,7 @@ useconds_t runCycle(CPU *cpu, uint8_t *ram, uint8_t *rom) {
     }
 
     case BMI: {
-      useconds_t cycles = 2;
+      uint8_t cycles = 2;
 
       // Get offset
       int8_t offset = (int8_t)readMemory(cpu->ProgrammeCounter++, ram, rom);
@@ -562,7 +611,7 @@ useconds_t runCycle(CPU *cpu, uint8_t *ram, uint8_t *rom) {
       // Perform shift
       cpu->Flags.CarryFlag = (value & 0x01) != 0;
       value = (uint8_t)(value >> 1);
-      // Set flags, write to memory
+      // Set flags and write to memory
       cpu->Flags.ZeroFlag = value == 0;
       cpu->Flags.NegativeFlag = 0;
       writeMemory(offset, value, ram, rom);
@@ -604,7 +653,17 @@ useconds_t runCycle(CPU *cpu, uint8_t *ram, uint8_t *rom) {
       return 3;
     }
     case EOR_A: {
-      break;
+      // Load address
+      uint8_t low = readMemory(cpu->ProgrammeCounter++, ram, rom);
+      uint8_t high = readMemory(cpu->ProgrammeCounter++, ram, rom);
+      // Get value
+      uint8_t value = readMemory(COMBINE(high, low), ram, rom);
+      // Perform XOR, set flags
+      cpu->Accumulator ^= value;
+      cpu->Flags.ZeroFlag = cpu->Accumulator == 0;
+      cpu->Flags.NegativeFlag = (cpu->Accumulator & 0x80) != 0;
+
+      return 4;
     }
     case LSR_A: {
       // Load address
@@ -615,7 +674,7 @@ useconds_t runCycle(CPU *cpu, uint8_t *ram, uint8_t *rom) {
       // Perform shift
       cpu->Flags.CarryFlag = (value & 0x01) != 0;
       value = (uint8_t)(value >> 1);
-      // Set flags, write to memory
+      // Set flags and write to memory
       cpu->Flags.ZeroFlag = value == 0;
       cpu->Flags.NegativeFlag = 0;
       writeMemory(COMBINE(high, low), value, ram, rom);
@@ -624,7 +683,7 @@ useconds_t runCycle(CPU *cpu, uint8_t *ram, uint8_t *rom) {
     }
 
     case BVC: {
-      useconds_t cycles = 2;
+      uint8_t cycles = 2;
 
       // Get offset
       int8_t offset = (int8_t)readMemory(cpu->ProgrammeCounter++, ram, rom);
@@ -678,7 +737,7 @@ useconds_t runCycle(CPU *cpu, uint8_t *ram, uint8_t *rom) {
       // Perform shift
       cpu->Flags.CarryFlag = (value & 0x01) != 0;
       value = (uint8_t)(value >> 1);
-      // Set flags, write to memory
+      // Set flags and write to memory
       cpu->Flags.ZeroFlag = value == 0;
       cpu->Flags.NegativeFlag = 0;
       writeMemory((uint8_t)(cpu->X + offset), value, ram, rom);
@@ -725,7 +784,7 @@ useconds_t runCycle(CPU *cpu, uint8_t *ram, uint8_t *rom) {
       // Perform shift
       cpu->Flags.CarryFlag = (value & 0x01) != 0;
       value = (uint8_t)(value >> 1);
-      // Set flags, write to memory
+      // Set flags and write to memory
       cpu->Flags.ZeroFlag = value == 0;
       cpu->Flags.NegativeFlag = 0;
       writeMemory(cpu->X + COMBINE(high, low), value, ram, rom);
@@ -752,17 +811,8 @@ useconds_t runCycle(CPU *cpu, uint8_t *ram, uint8_t *rom) {
       uint8_t high = readMemory((uint8_t)(_address + 1), ram, rom);
       // Get value
       uint8_t value = readMemory(COMBINE(high, low), ram, rom);
-      // BCD
-      if (cpu->Flags.DecimalMode == 1) {
-        break; // TODO: BCD
-      } else {
-        uint16_t result = cpu->Accumulator + value + cpu->Flags.CarryFlag;
-        cpu->Flags.CarryFlag = result > 0xFF;
-        cpu->Flags.ZeroFlag = result == 0;
-        cpu->Flags.OverflowFlag = ((cpu->Accumulator ^ result) & (value ^ result) & 0x80) != 0;
-        cpu->Flags.NegativeFlag = (result & 0x80) != 0;
-        cpu->Accumulator = (uint8_t)result;
-      }
+
+      doADC(cpu, value);
 
       return 6;
     }
@@ -771,22 +821,28 @@ useconds_t runCycle(CPU *cpu, uint8_t *ram, uint8_t *rom) {
       uint8_t offset = readMemory(cpu->ProgrammeCounter++, ram, rom);
       // Get value
       uint8_t value = readMemory(offset, ram, rom);
-      // BCD
-      if (cpu->Flags.DecimalMode == 1) {
-        break; // TODO: BCD
-      } else {
-        uint16_t result = cpu->Accumulator + value + cpu->Flags.CarryFlag;
-        cpu->Flags.CarryFlag = result > 0xFF;
-        cpu->Flags.ZeroFlag = result == 0;
-        cpu->Flags.OverflowFlag = ((cpu->Accumulator ^ result) & (value ^ result) & 0x80) != 0;
-        cpu->Flags.NegativeFlag = (result & 0x80) != 0;
-        cpu->Accumulator = (uint8_t)result;
-      }
+
+      doADC(cpu, value);
 
       return 3;
     }
     case ROR_Z: {
-      break;
+      // Get offset
+      uint8_t offset = readMemory(cpu->ProgrammeCounter++, ram, rom);
+      // Get value
+      uint8_t value = readMemory(offset, ram, rom);
+      // Save bit 0, rotate, set bit 0
+      uint8_t bit0 = value & 0x1;
+      value = value >> 1;
+      value |= (cpu->Flags.CarryFlag << 7);
+      // Set flags
+      cpu->Flags.CarryFlag = bit0 != 0;
+      cpu->Flags.ZeroFlag = value == 0;
+      cpu->Flags.NegativeFlag = (value & 0x80) != 0;
+      // Write value
+      writeMemory(offset, value, ram, rom);
+
+      return 5;
     }
     case PLA: {
       // Pull from stack
@@ -799,22 +855,22 @@ useconds_t runCycle(CPU *cpu, uint8_t *ram, uint8_t *rom) {
     case ADC_I: {
       // Get value
       uint8_t value = readMemory(cpu->ProgrammeCounter++, ram, rom);
-      // BCD
-      if (cpu->Flags.DecimalMode == 1) {
-        break; // TODO: BCD
-      } else {
-        uint16_t result = cpu->Accumulator + value + cpu->Flags.CarryFlag;
-        cpu->Flags.CarryFlag = result > 0xFF;
-        cpu->Flags.ZeroFlag = result == 0;
-        cpu->Flags.OverflowFlag = ((cpu->Accumulator ^ result) & (value ^ result) & 0x80) != 0;
-        cpu->Flags.NegativeFlag = (result & 0x80) != 0;
-        cpu->Accumulator = (uint8_t)result;
-      }
+
+      doADC(cpu, value);
 
       return 2;
     }
     case ROR: {
-      break;
+      // Save bit 0, rotate, set bit 0
+      uint8_t bit0 = cpu->Accumulator & 0x1;
+      cpu->Accumulator = cpu->Accumulator >> 1;
+      cpu->Accumulator |= (cpu->Flags.CarryFlag << 7);
+      // Set flags
+      cpu->Flags.CarryFlag = bit0 != 0;
+      cpu->Flags.ZeroFlag = cpu->Accumulator == 0;
+      cpu->Flags.NegativeFlag = (cpu->Accumulator & 0x80) != 0;
+
+      return 2;
     }
     case JMP_Ai: {
       // Load pointer address
@@ -836,26 +892,32 @@ useconds_t runCycle(CPU *cpu, uint8_t *ram, uint8_t *rom) {
       // Get value
       uint8_t value = readMemory(COMBINE(high, low), ram, rom);
 
-      // BCD
-      if (cpu->Flags.DecimalMode == 1) {
-        break; // TODO: BCD
-      } else {
-        uint16_t result = cpu->Accumulator + value + cpu->Flags.CarryFlag;
-        cpu->Flags.CarryFlag = result > 0xFF;
-        cpu->Flags.ZeroFlag = result == 0;
-        cpu->Flags.OverflowFlag = ((cpu->Accumulator ^ result) & (value ^ result) & 0x80) != 0;
-        cpu->Flags.NegativeFlag = (result & 0x80) != 0;
-        cpu->Accumulator = (uint8_t)result;
-      }
+      doADC(cpu, value);
 
       return 4;
     }
     case ROR_A: {
-      break;
+      // Get address
+      uint8_t low = readMemory(cpu->ProgrammeCounter++, ram, rom);
+      uint8_t high = readMemory(cpu->ProgrammeCounter++, ram, rom);
+      // Get value
+      uint8_t value = readMemory(COMBINE(high, low), ram, rom);
+      // Save bit 7, rotate, set bit 0
+      uint8_t bit0 = value & 0x1;
+      value = value >> 1;
+      value |= (cpu->Flags.CarryFlag << 7);
+      // Set flags
+      cpu->Flags.CarryFlag = bit0 != 0;
+      cpu->Flags.ZeroFlag = value == 0;
+      cpu->Flags.NegativeFlag = (value & 0x80) != 0;
+      // Write value
+      writeMemory(COMBINE(high, low), value, ram, rom);
+
+      return 6;
     }
 
     case BVS: {
-      useconds_t cycles = 2;
+      uint8_t cycles = 2;
 
       // Get offset
       int8_t offset = (int8_t)readMemory(cpu->ProgrammeCounter++, ram, rom);
@@ -882,17 +944,8 @@ useconds_t runCycle(CPU *cpu, uint8_t *ram, uint8_t *rom) {
       uint16_t address = cpu->Y + COMBINE(high, low);
       // Get value
       uint8_t value = readMemory(address, ram, rom);
-      // BCD
-      if (cpu->Flags.DecimalMode == 1) {
-        break; // TODO: BCD
-      } else {
-        uint16_t result = cpu->Accumulator + value + cpu->Flags.CarryFlag;
-        cpu->Flags.CarryFlag = result > 0xFF;
-        cpu->Flags.ZeroFlag = result == 0;
-        cpu->Flags.OverflowFlag = ((cpu->Accumulator ^ result) & (value ^ result) & 0x80) != 0;
-        cpu->Flags.NegativeFlag = (result & 0x80) != 0;
-        cpu->Accumulator = (uint8_t)result;
-      }
+
+      doADC(cpu, value);
 
       return 5 + (((uint16_t)low + cpu->Y) > 0xFF);
     }
@@ -901,22 +954,28 @@ useconds_t runCycle(CPU *cpu, uint8_t *ram, uint8_t *rom) {
       uint8_t offset = readMemory(cpu->ProgrammeCounter++, ram, rom);
       // Get value
       uint8_t value = readMemory((uint8_t)(cpu->X + offset), ram, rom);
-      // BCD
-      if (cpu->Flags.DecimalMode == 1) {
-        break; // TODO: BCD
-      } else {
-        uint16_t result = cpu->Accumulator + value + cpu->Flags.CarryFlag;
-        cpu->Flags.CarryFlag = result > 0xFF;
-        cpu->Flags.ZeroFlag = result == 0;
-        cpu->Flags.OverflowFlag = ((cpu->Accumulator ^ result) & (value ^ result) & 0x80) != 0;
-        cpu->Flags.NegativeFlag = (result & 0x80) != 0;
-        cpu->Accumulator = (uint8_t)result;
-      }
+
+      doADC(cpu, value);
 
       return 4;
     }
     case ROR_XZ: {
-      break;
+      // Get offset
+      uint8_t offset = readMemory(cpu->ProgrammeCounter++, ram, rom);
+      // Get value
+      uint8_t value = readMemory((uint8_t)(cpu->X + offset), ram, rom);
+      // Save bit 0, rotate, set bit 0
+      uint8_t bit0 = value & 0x1;
+      value = value >> 1;
+      value |= (cpu->Flags.CarryFlag << 7);
+      // Set flags
+      cpu->Flags.CarryFlag = bit0 != 0;
+      cpu->Flags.ZeroFlag = value == 0;
+      cpu->Flags.NegativeFlag = (value & 0x80) != 0;
+      // Write value
+      writeMemory((uint8_t)(cpu->X + offset), value, ram, rom);
+
+      return 6;
     }
     case SEI: {
       cpu->Flags.InterruptDisable = 1;
@@ -929,17 +988,8 @@ useconds_t runCycle(CPU *cpu, uint8_t *ram, uint8_t *rom) {
       uint8_t high = readMemory(cpu->ProgrammeCounter++, ram, rom);
       // Get value
       uint8_t value = readMemory(cpu->Y + COMBINE(high, low), ram, rom);
-      // BCD
-      if (cpu->Flags.DecimalMode == 1) {
-        break; // TODO: BCD
-      } else {
-        uint16_t result = cpu->Accumulator + value + cpu->Flags.CarryFlag;
-        cpu->Flags.CarryFlag = result > 0xFF;
-        cpu->Flags.ZeroFlag = result == 0;
-        cpu->Flags.OverflowFlag = ((cpu->Accumulator ^ result) & (value ^ result) & 0x80) != 0;
-        cpu->Flags.NegativeFlag = (result & 0x80) != 0;
-        cpu->Accumulator = (uint8_t)result;
-      }
+
+      doADC(cpu, value);
 
       return 4 + (((uint16_t)low + cpu->Y) > 0xFF);
     }
@@ -949,35 +999,67 @@ useconds_t runCycle(CPU *cpu, uint8_t *ram, uint8_t *rom) {
       uint8_t high = readMemory(cpu->ProgrammeCounter++, ram, rom);
       // Get value
       uint8_t value = readMemory(cpu->X + COMBINE(high, low), ram, rom);
-      // BCD
-      if (cpu->Flags.DecimalMode == 1) {
-        break; // TODO: BCD
-      } else {
-        uint16_t result = cpu->Accumulator + value + cpu->Flags.CarryFlag;
-        cpu->Flags.CarryFlag = result > 0xFF;
-        cpu->Flags.ZeroFlag = result == 0;
-        cpu->Flags.OverflowFlag = ((cpu->Accumulator ^ result) & (value ^ result) & 0x80) != 0;
-        cpu->Flags.NegativeFlag = (result & 0x80) != 0;
-        cpu->Accumulator = (uint8_t)result;
-      }
+
+      doADC(cpu, value);
 
       return 4 + (((uint16_t)low + cpu->X) > 0xFF);
     }
     case ROR_XA: {
-      break;
+      // Get address
+      uint8_t low = readMemory(cpu->ProgrammeCounter++, ram, rom);
+      uint8_t high = readMemory(cpu->ProgrammeCounter++, ram, rom);
+      // Get value
+      uint8_t value = readMemory(cpu->X + COMBINE(high, low), ram, rom);
+      // Save bit 0, rotate, set bit 0
+      uint8_t bit0 = value & 0x1;
+      value = value >> 1;
+      value |= (cpu->Flags.CarryFlag << 7);
+      // Set flags
+      cpu->Flags.CarryFlag = bit0 != 0;
+      cpu->Flags.ZeroFlag = value == 0;
+      cpu->Flags.NegativeFlag = (value & 0x80) != 0;
+      // Write value
+      writeMemory(cpu->X + COMBINE(high, low), value, ram, rom);
+
+      return 7;
     }
 
     case STA_XZi: {
-      break;
+      // Get offset
+      uint8_t offset = readMemory(cpu->ProgrammeCounter++, ram, rom);
+      // Get indirect address, discard carry
+      uint8_t _address = (uint8_t)(cpu->X + offset);
+      // Get final address, wrap around page
+      uint8_t low = readMemory(_address, ram, rom);
+      uint8_t high = readMemory((uint8_t)(_address + 1), ram, rom);
+      // Store accumulator at address
+      writeMemory(COMBINE(high, low), cpu->Accumulator, ram, rom);
+
+      return 6;
     }
     case STY_Z: {
-      break;
+      // Get offset
+      uint8_t offset = readMemory(cpu->ProgrammeCounter++, ram, rom);
+      // Store Y at address
+      writeMemory(offset, cpu->Y, ram, rom);
+
+      return 3;
     }
     case STA_Z: {
-      break;
+      // Get offset
+      uint8_t offset = readMemory(cpu->ProgrammeCounter++, ram, rom);
+      // Store accumulator at address
+      writeMemory(offset, cpu->Accumulator, ram, rom);
+
+      return 3;
     }
     case STX_Z: {
-      break;
+      // Get offset
+      uint8_t offset = readMemory(cpu->ProgrammeCounter++, ram, rom);
+      // Store X at address
+      writeMemory(offset, cpu->X, ram, rom);
+
+      return 3;
     }
     case DEY: {
       // Decrement, set flags
@@ -996,7 +1078,13 @@ useconds_t runCycle(CPU *cpu, uint8_t *ram, uint8_t *rom) {
       return 2;
     }
     case STY_A: {
-      break;
+      // Load address
+      uint8_t low = readMemory(cpu->ProgrammeCounter++, ram, rom);
+      uint8_t high = readMemory(cpu->ProgrammeCounter++, ram, rom);
+      // Store Y at address
+      writeMemory(COMBINE(high, low), cpu->Y, ram, rom);
+
+      return 4;
     }
     case STA_A: {
       // Load address
@@ -1008,11 +1096,17 @@ useconds_t runCycle(CPU *cpu, uint8_t *ram, uint8_t *rom) {
       return 4;
     }
     case STX_A: {
-      break;
+      // Load address
+      uint8_t low = readMemory(cpu->ProgrammeCounter++, ram, rom);
+      uint8_t high = readMemory(cpu->ProgrammeCounter++, ram, rom);
+      // Store X at address
+      writeMemory(COMBINE(high, low), cpu->X, ram, rom);
+
+      return 4;
     }
 
     case BCC: {
-      useconds_t cycles = 2;
+      uint8_t cycles = 2;
 
       // Get offset
       int8_t offset = (int8_t)readMemory(cpu->ProgrammeCounter++, ram, rom);
@@ -1030,16 +1124,41 @@ useconds_t runCycle(CPU *cpu, uint8_t *ram, uint8_t *rom) {
       return cycles;
     }
     case STA_ZiY: {
-      break;
+      // Get _address
+      uint8_t _address = readMemory(cpu->ProgrammeCounter++, ram, rom);
+      // Get offset
+      uint8_t low = readMemory(_address, ram, rom);
+      uint8_t high = readMemory((uint8_t)(_address + 1), ram, rom);
+      // Get address
+      uint16_t address = cpu->Y + COMBINE(high, low);
+      // Store accumulator at address
+      writeMemory(address, cpu->Accumulator, ram, rom);
+
+      return 6;
     }
     case STY_XZ: {
-      break;
+      // Get offset
+      uint8_t offset = readMemory(cpu->ProgrammeCounter++, ram, rom);
+      // Store accumulator at address
+      writeMemory((uint8_t)(cpu->X + offset), cpu->Y, ram, rom);
+
+      return 4;
     }
     case STA_XZ: {
-      break;
+      // Get offset
+      uint8_t offset = readMemory(cpu->ProgrammeCounter++, ram, rom);
+      // Store accumulator at address
+      writeMemory((uint8_t)(cpu->X + offset), cpu->Accumulator, ram, rom);
+
+      return 4;
     }
     case STX_YZ: {
-      break;
+      // Get offset
+      uint8_t offset = readMemory(cpu->ProgrammeCounter++, ram, rom);
+      // Store accumulator at address
+      writeMemory((uint8_t)(cpu->Y + offset), cpu->X, ram, rom);
+
+      return 4;
     }
     case TYA: {
       cpu->Accumulator = cpu->Y;
@@ -1050,7 +1169,13 @@ useconds_t runCycle(CPU *cpu, uint8_t *ram, uint8_t *rom) {
       return 2;
     }
     case STA_YA: {
-      break;
+      // Get address
+      uint8_t low = readMemory(cpu->ProgrammeCounter++, ram, rom);
+      uint8_t high = readMemory(cpu->ProgrammeCounter++, ram, rom);
+      // Store accumulator at address
+      writeMemory(cpu->Y + COMBINE(high, low), cpu->Accumulator, ram, rom);
+
+      return 5;
     }
     case TXS: {
       cpu->StackPointer = cpu->X;
@@ -1058,7 +1183,13 @@ useconds_t runCycle(CPU *cpu, uint8_t *ram, uint8_t *rom) {
       return 2;
     }
     case STA_XA: {
-      break;
+      // Get address
+      uint8_t low = readMemory(cpu->ProgrammeCounter++, ram, rom);
+      uint8_t high = readMemory(cpu->ProgrammeCounter++, ram, rom);
+      // Store accumulator at address
+      writeMemory(cpu->X + COMBINE(high, low), cpu->Accumulator, ram, rom);
+
+      return 5;
     }
 
     case LDY_I: {
@@ -1201,7 +1332,7 @@ useconds_t runCycle(CPU *cpu, uint8_t *ram, uint8_t *rom) {
     }
 
     case BCS: {
-      useconds_t cycles = 2;
+      uint8_t cycles = 2;
 
       // Get offset
       int8_t offset = (int8_t)readMemory(cpu->ProgrammeCounter++, ram, rom);
@@ -1393,7 +1524,18 @@ useconds_t runCycle(CPU *cpu, uint8_t *ram, uint8_t *rom) {
       return 3;
     }
     case DEC_Z: {
-      break;
+      // Get offset
+      uint8_t offset = readMemory(cpu->ProgrammeCounter++, ram, rom);
+      // Get value
+      uint8_t value = readMemory(offset, ram, rom);
+      // Decrement, set flags
+      value--;
+      cpu->Flags.ZeroFlag = value == 0;
+      cpu->Flags.NegativeFlag = (value & 0x80) != 0;
+      // Store at address
+      writeMemory(offset, value, ram, rom);
+
+      return 5;
     }
     case INY: {
       // Increment, set flags
@@ -1454,11 +1596,23 @@ useconds_t runCycle(CPU *cpu, uint8_t *ram, uint8_t *rom) {
       return 4;
     }
     case DEC_A: {
-      break;
+      // Load address
+      uint8_t low = readMemory(cpu->ProgrammeCounter++, ram, rom);
+      uint8_t high = readMemory(cpu->ProgrammeCounter++, ram, rom);
+      // Get value
+      uint8_t value = readMemory(COMBINE(high, low), ram, rom);
+      // Decrement, set flags
+      value--;
+      cpu->Flags.ZeroFlag = value == 0;
+      cpu->Flags.NegativeFlag = (value & 0x80) != 0;
+      // Store at address
+      writeMemory(COMBINE(high, low), value, ram, rom);
+
+      return 6;
     }
 
     case BNE: {
-      useconds_t cycles = 2;
+      uint8_t cycles = 2;
 
       // Get offset
       int8_t offset = (int8_t)readMemory(cpu->ProgrammeCounter++, ram, rom);
@@ -1509,7 +1663,18 @@ useconds_t runCycle(CPU *cpu, uint8_t *ram, uint8_t *rom) {
       return 4;
     }
     case DEC_XZ: {
-      break;
+      // Get offset
+      uint8_t offset = readMemory(cpu->ProgrammeCounter++, ram, rom);
+      // Get value
+      uint8_t value = readMemory((uint8_t)(cpu->X + offset), ram, rom);
+      // Decrement, set flags
+      value--;
+      cpu->Flags.ZeroFlag = value == 0;
+      cpu->Flags.NegativeFlag = (value & 0x80) != 0;
+      // Store at address
+      writeMemory((uint8_t)(cpu->X + offset), value, ram, rom);
+
+      return 6;
     }
     case CLD: {
       cpu->Flags.DecimalMode = 0;
@@ -1547,7 +1712,19 @@ useconds_t runCycle(CPU *cpu, uint8_t *ram, uint8_t *rom) {
       return 4 + (((uint16_t)low + cpu->X) > 0xFF);
     }
     case DEC_XA: {
-      break;
+      // Get address
+      uint8_t low = readMemory(cpu->ProgrammeCounter++, ram, rom);
+      uint8_t high = readMemory(cpu->ProgrammeCounter++, ram, rom);
+      // Get value
+      uint8_t value = readMemory(cpu->X + COMBINE(high, low), ram, rom);
+      // Decrement, set flags
+      value--;
+      cpu->Flags.ZeroFlag = value == 0;
+      cpu->Flags.NegativeFlag = (value & 0x80) != 0;
+      // Store at address
+      writeMemory(cpu->X + COMBINE(high, low), value, ram, rom);
+
+      return 7;
     }
 
     case CPX_I: {
@@ -1563,7 +1740,19 @@ useconds_t runCycle(CPU *cpu, uint8_t *ram, uint8_t *rom) {
       return 2;
     }
     case SBC_XZi: {
-      break;
+      // Get offset
+      uint8_t offset = readMemory(cpu->ProgrammeCounter++, ram, rom);
+      // Get indirect address, discard carry
+      uint8_t _address = (uint8_t)(cpu->X + offset);
+      // Get final address, wrap around page
+      uint8_t low = readMemory(_address, ram, rom);
+      uint8_t high = readMemory((uint8_t)(_address + 1), ram, rom);
+      // Get value
+      uint8_t value = readMemory(COMBINE(high, low), ram, rom);
+
+      doSBC(cpu, value);
+
+      return 6;
     }
     case CPX_Z: {
       // Get offset
@@ -1580,10 +1769,28 @@ useconds_t runCycle(CPU *cpu, uint8_t *ram, uint8_t *rom) {
       return 3;
     }
     case SBC_Z: {
-      break;
+      // Get offset
+      uint8_t offset = readMemory(cpu->ProgrammeCounter++, ram, rom);
+      // Get value
+      uint8_t value = readMemory(offset, ram, rom);
+
+      doSBC(cpu, value);
+
+      return 3;
     }
     case INC_Z: {
-      break;
+      // Get offset
+      uint8_t offset = readMemory(cpu->ProgrammeCounter++, ram, rom);
+      // Get value
+      uint8_t value = readMemory(offset, ram, rom);
+      // Increment, set flags
+      value++;
+      cpu->Flags.ZeroFlag = value == 0;
+      cpu->Flags.NegativeFlag = (value & 0x80) != 0;
+      // Store at address
+      writeMemory(offset, value, ram, rom);
+
+      return 5;
     }
     case INX: {
       // Increment, set flags
@@ -1594,7 +1801,12 @@ useconds_t runCycle(CPU *cpu, uint8_t *ram, uint8_t *rom) {
       return 2;
     }
     case SBC_I: {
-      break;
+      // Get value
+      uint8_t value = readMemory(cpu->ProgrammeCounter++, ram, rom);
+
+      doSBC(cpu, value);
+
+      return 2;
     }
     case NOP: {
       return 2;
@@ -1615,14 +1827,34 @@ useconds_t runCycle(CPU *cpu, uint8_t *ram, uint8_t *rom) {
       return 4;
     }
     case SBC_A: {
-      break;
+      // Load address
+      uint8_t low = readMemory(cpu->ProgrammeCounter++, ram, rom);
+      uint8_t high = readMemory(cpu->ProgrammeCounter++, ram, rom);
+      // Get value
+      uint8_t value = readMemory(COMBINE(high, low), ram, rom);
+
+      doSBC(cpu, value);
+
+      return 4;
     }
     case INC_A: {
-      break;
+      // Load address
+      uint8_t low = readMemory(cpu->ProgrammeCounter++, ram, rom);
+      uint8_t high = readMemory(cpu->ProgrammeCounter++, ram, rom);
+      // Get value
+      uint8_t value = readMemory(COMBINE(high, low), ram, rom);
+      // Increment, set flags
+      value++;
+      cpu->Flags.ZeroFlag = value == 0;
+      cpu->Flags.NegativeFlag = (value & 0x80) != 0;
+      // Store at address
+      writeMemory(COMBINE(high, low), value, ram, rom);
+
+      return 6;
     }
 
     case BEQ: {
-      useconds_t cycles = 2;
+      uint8_t cycles = 2;
 
       // Get offset
       int8_t offset = (int8_t)readMemory(cpu->ProgrammeCounter++, ram, rom);
@@ -1640,13 +1872,43 @@ useconds_t runCycle(CPU *cpu, uint8_t *ram, uint8_t *rom) {
       return cycles;
     }
     case SBC_ZiY: {
-      break;
+      // Get _address
+      uint8_t _address = readMemory(cpu->ProgrammeCounter++, ram, rom);
+      // Get offset
+      uint8_t low = readMemory(_address, ram, rom);
+      uint8_t high = readMemory((uint8_t)(_address + 1), ram, rom);
+      // Get address
+      uint16_t address = cpu->Y + COMBINE(high, low);
+      // Get value
+      uint8_t value = readMemory(address, ram, rom);
+
+      doSBC(cpu, value);
+
+      return 5 + (((uint16_t)low + cpu->Y) > 0xFF);
     }
     case SBC_XZ: {
-      break;
+      // Get offset
+      uint8_t offset = readMemory(cpu->ProgrammeCounter++, ram, rom);
+      // Get value
+      uint8_t value = readMemory((uint8_t)(cpu->X + offset), ram, rom);
+
+      doSBC(cpu, value);
+
+      return 4;
     }
     case INC_XZ: {
-      break;
+      // Get offset
+      uint8_t offset = readMemory(cpu->ProgrammeCounter++, ram, rom);
+      // Get value
+      uint8_t value = readMemory((uint8_t)(cpu->X + offset), ram, rom);
+      // Increment, set flags
+      value++;
+      cpu->Flags.ZeroFlag = value == 0;
+      cpu->Flags.NegativeFlag = (value & 0x80) != 0;
+      // Store at address
+      writeMemory((uint8_t)(cpu->X + offset), value, ram, rom);
+
+      return 6;
     }
     case SED: {
       cpu->Flags.DecimalMode = 1;
@@ -1654,13 +1916,41 @@ useconds_t runCycle(CPU *cpu, uint8_t *ram, uint8_t *rom) {
       return 2;
     }
     case SBC_YA: {
-      break;
+      // Get address
+      uint8_t low = readMemory(cpu->ProgrammeCounter++, ram, rom);
+      uint8_t high = readMemory(cpu->ProgrammeCounter++, ram, rom);
+      // Get value
+      uint8_t value = readMemory(cpu->Y + COMBINE(high, low), ram, rom);
+
+      doSBC(cpu, value);
+
+      return 4 + (((uint16_t)low + cpu->Y) > 0xFF);
     }
     case SBC_XA: {
-      break;
+      // Get address
+      uint8_t low = readMemory(cpu->ProgrammeCounter++, ram, rom);
+      uint8_t high = readMemory(cpu->ProgrammeCounter++, ram, rom);
+      // Get value
+      uint8_t value = readMemory(cpu->X + COMBINE(high, low), ram, rom);
+
+      doSBC(cpu, value);
+
+      return 4 + (((uint16_t)low + cpu->X) > 0xFF);
     }
     case INC_XA: {
-      break;
+      // Get address
+      uint8_t low = readMemory(cpu->ProgrammeCounter++, ram, rom);
+      uint8_t high = readMemory(cpu->ProgrammeCounter++, ram, rom);
+      // Get value
+      uint8_t value = readMemory(cpu->X + COMBINE(high, low), ram, rom);
+      // Increment, set flags
+      value++;
+      cpu->Flags.ZeroFlag = value == 0;
+      cpu->Flags.NegativeFlag = (value & 0x80) != 0;
+      // Store at address
+      writeMemory(cpu->X + COMBINE(high, low), value, ram, rom);
+
+      return 7;
     }
 
     default:
diff --git a/src/main.c b/src/main.c
index 08ed31d..d707c17 100644
--- a/src/main.c
+++ b/src/main.c
@@ -6,9 +6,13 @@
 #include <stdio.h>
 #include <stdlib.h>
 #include <termios.h>
+#include <time.h>
 #include <unistd.h>
 
+#define MACHINE_SPEED_HZ 1e6 // 1 MHz;
+
 static struct termios orig_termios;
+static struct timespec sleepTime;
 
 void restore_terminal(void) { tcsetattr(STDIN_FILENO, TCSANOW, &orig_termios); }
 
@@ -44,7 +48,8 @@ int main(void) {
 
   /* DEBUG */
 
-  FILE *fp = fopen("/Users/Kili2/Documents/Projekte/6502pc/6502_65C02_functional_tests/6502_functional_test.bin", "rb");
+  //FILE *fp = fopen("/Users/Kili2/Documents/Projekte/6502pc/6502_65C02_functional_tests/6502_functional_test.bin", "rb");
+  FILE *fp = fopen("/Users/Kili2/Documents/Projekte/6502pc/6502_65C02_functional_tests/6502_decimal_test.bin", "rb");
   if (!fp) {
     perror("fopen");
     return 1;
@@ -65,14 +70,15 @@ int main(void) {
   resetCPU(&cpu, rom);
 
   while (1) {
-    //printf("\tPC: 0x%x\tOP: 0x%x\n", cpu.ProgrammeCounter, ram[cpu.ProgrammeCounter]);
-    useconds_t sleep = runCycle(&cpu, ram, rom);
+    // printf("\tPC: 0x%x\tOP: 0x%x\n", cpu.ProgrammeCounter, ram[cpu.ProgrammeCounter]);
+    uint8_t sleep = runCycle(&cpu, ram, rom);
     if (sleep == 1) {
       printf("opcode missing: 0x%X\n", ram[cpu.ProgrammeCounter - 1]);
       return 1;
     }
-    fflush(stdout);
-    usleep(sleep);
+    sleepTime.tv_sec = 0;
+    sleepTime.tv_nsec = sleep * (1000 / MACHINE_SPEED_HZ);
+    nanosleep(&sleepTime, NULL);
   }
 
   /* CPU WORK */
diff --git a/src/rom.c b/src/rom.c
index 91b6051..7457b1d 100644
--- a/src/rom.c
+++ b/src/rom.c
@@ -3,6 +3,8 @@
 uint8_t rom[ROM_SIZE] = {0};
 
 void initRom(uint8_t *rom) {
+  
+
   rom[0] = 0xa9,                                             // LDA
       rom[1] = 'H',                                          // H literal
       rom[2] = 0x8d,                                         // STA
@@ -11,7 +13,7 @@ void initRom(uint8_t *rom) {
       rom[6] = 
                                                              /* Reset Vector */
       rom[0xFFFC - ROM_BASE] = 0x00,                         // Lower byte base of ROM
-      rom[0xFFFD - ROM_BASE] = 0x04,                         // Upper byte of base of ROM
+      rom[0xFFFD - ROM_BASE] = 0x02,                         // Upper byte of base of ROM
       rom[0xFFFE - ROM_BASE] = 0x00,                         // Lower byte base of ROM
       rom[0xFFFF - ROM_BASE] = 0x00;                         // Upper byte of base of ROM
 }
\ No newline at end of file