Merge branch 'dynarec' into performanceperformance

16 files changed, 4601 insertions, 0 deletions

diff --git a/Makefile b/Makefile
index f834e3b..23b8979 100644
--- a/Makefile
+++ b/Makefile
@@ -293,6 +293,7 @@ else ifeq ($(platform), trimui)
 	CC = $(CROSS_COMPILE)gcc
 	AR = $(CROSS_COMPILE)ar
 	ARM_ASM := 1
+	USE_DYNAREC := 1
 	SHARED := -shared -Wl,--version-script=link.T -Wl,--no-undefined
 	CFLAGS += -fsingle-precision-constant -fno-PIC -flto
 	CFLAGS += -DLSB_FIRST -DFAST_ALIGNED_LSB_WORD_ACCESS -DRIGHTSHIFT_IS_SAR
@@ -631,9 +632,11 @@ ifeq ($(platform), psp1)
 endif
 
 OBJECTS := $(SOURCES_C:.c=.o)
+OBJECTS := $(OBJECTS:.S=.o)
 
 CXXFLAGS += $(FLAGS)
 CFLAGS += $(FLAGS)
+ASFLAGS += $(FLAGS)
 
 OBJOUT   = -o
 LINKOUT  = -o 
diff --git a/Makefile.common b/Makefile.common
index 2c015a8..15c4a8e 100644
--- a/Makefile.common
+++ b/Makefile.common
@@ -52,6 +52,14 @@ ifeq ($(ARM_ASM),1)
 FLAGS += -DARM_ASM
 endif
 
+ifeq ($(USE_DYNAREC),1)
+SOURCES_C += \
+	$(CORE_DIR)/arm_dynarec/dynaexec.c \
+	$(CORE_DIR)/arm_dynarec/map_rwx.c \
+	$(CORE_DIR)/arm_dynarec/armfn.S
+FLAGS += -DUSE_DYNAREC
+endif
+
 ifeq ($(LOAD_FROM_MEMORY),1)
 	FLAGS += -DLOAD_FROM_MEMORY
 else ifneq ($(STATIC_LINKING), 1)
diff --git a/libretro.c b/libretro.c
index 5e4dc31..2370f4c 100644
--- a/libretro.c
+++ b/libretro.c
@@ -12,6 +12,10 @@
 #include "srtc.h"
 #include "sa1.h"
 
+#ifdef USE_DYNAREC
+#include "arm_dynarec/dynaexec.h"
+#endif
+
 #ifdef PSP
 #include <pspkernel.h>
 #include <pspgu.h>
@@ -456,6 +460,12 @@ void retro_init(void)
 #endif
    CPU.SaveStateVersion = 0;
 
+#ifdef USE_DYNAREC
+   /* Init before loading ROM to have a better chance of mmaping close to code */
+   if (DynaInit())
+      Settings.EnableDynarec = false;
+#endif
+
    if (environ_cb(RETRO_ENVIRONMENT_GET_INPUT_BITMASKS, NULL))
       libretro_supports_bitmasks = true;
 }
@@ -535,6 +545,7 @@ static void check_variables(bool first_run)
 {
    struct retro_variable var;
    bool prev_frameskip_type;
+   bool prev_dynarec;
 
    if (first_run)
    {
@@ -632,6 +643,27 @@ static void check_variables(bool first_run)
       if (strcmp(var.value, "enabled") == 0)
          reduce_sprite_flicker = true;
 
+#ifdef USE_DYNAREC
+   prev_dynarec = Settings.EnableDynarec;
+
+   var.key = "snes9x_2005_dynarec";
+   var.value = NULL;
+
+   if (environ_cb(RETRO_ENVIRONMENT_GET_VARIABLE, &var) && var.value)
+   {
+      if (strcmp(var.value, "enabled") == 0)
+         Settings.EnableDynarec = true;
+      else
+         Settings.EnableDynarec = false;
+   }
+
+   if (Settings.EnableDynarec && Settings.EnableDynarec != prev_dynarec) {
+      if (DynaReset())
+         Settings.EnableDynarec = false;
+   }
+
+#endif /* USE_DYNAREC */
+
    /* Reinitialise frameskipping, if required */
    if (!first_run &&
        (frameskip_type != prev_frameskip_type))
@@ -827,6 +859,13 @@ void retro_reset(void)
 {
    CPU.Flags = 0;
    S9xSoftReset();
+#ifdef USE_DYNAREC
+   if (Settings.EnableDynarec)
+   {
+      if (DynaReset())
+         Settings.EnableDynarec = false;
+   }
+#endif /* USE_DYNAREC */
 }
 
 size_t retro_serialize_size(void)
diff --git a/libretro_core_options.h b/libretro_core_options.h
index 409590c..154dc74 100644
--- a/libretro_core_options.h
+++ b/libretro_core_options.h
@@ -221,6 +221,22 @@ struct retro_core_option_v2_definition option_defs_us[] = {
       },
       "disabled"
    },
+#ifdef USE_DYNAREC
+   {
+      "snes9x_2005_dynarec",
+      "Dynamic Recompiler",
+      NULL,
+      "Enables experimental dynamic recompiler.",
+      NULL,
+      NULL,
+      {
+         { "disabled", NULL },
+         { "enabled",  NULL },
+         { NULL, NULL },
+      },
+      "disabled"
+   },
+#endif
    { NULL, NULL, NULL, NULL, NULL, NULL, {{0}}, NULL },
 };
 
diff --git a/source/arm_dynarec/armfn.S b/source/arm_dynarec/armfn.S
new file mode 100644
index 0000000..c5d1b3f
--- /dev/null
+++ b/source/arm_dynarec/armfn.S
@@ -0,0 +1,501 @@
+#define RegA      r4
+#define RegX      r5
+#define RegY      r6
+#define RegP      r7
+#define RegCPU    r8
+#define RegCPUPC  r9
+#define RegCycles r10
+#define RegChecks ip
+
+#ifndef NDEBUG
+#define METRICS_GETSETS_OFF     28
+#define METRICS_SLOWGETSETS_OFF 32
+#endif
+
+#define CPU_FLAGS_OFF              0
+#define CPU_PC_OFF                 12
+#define CPU_PCBASE_OFF             16
+#define CPU_PC_AT_OPCODE_START_OFF 20
+#define CPU_WAIT_ADDRESS_OFF       24
+#define CPU_CYCLES_OFF             32
+#define CPU_NEXT_EVENT_OFF         36
+#define CPU_ICPU_OFF               0x80
+#define CPU_MEMORY_MAP_OFF         0x100
+
+#define ICPU_S9X_OPCODES_OFF (CPU_ICPU_OFF + 4)
+#define ICPU_REG_PB_OFF      (CPU_ICPU_OFF + 8)
+#define ICPU_REG_DB_OFF      (CPU_ICPU_OFF + 9)
+#define ICPU_REG_P_OFF       (CPU_ICPU_OFF + 10)
+#define ICPU_REG_A_OFF       (CPU_ICPU_OFF + 12)
+#define ICPU_REG_D_OFF       (CPU_ICPU_OFF + 14)
+#define ICPU_REG_S_OFF       (CPU_ICPU_OFF + 16)
+#define ICPU_REG_X_OFF       (CPU_ICPU_OFF + 18)
+#define ICPU_REG_Y_OFF       (CPU_ICPU_OFF + 20)
+#define ICPU_REG_PC_OFF      (CPU_ICPU_OFF + 22)
+#define ICPU_CARRY_OFF       (CPU_ICPU_OFF + 24)
+#define ICPU_ZERO_OFF        (CPU_ICPU_OFF + 25)
+#define ICPU_NEGATIVE_OFF    (CPU_ICPU_OFF + 26)
+#define ICPU_OVERFLOW_OFF    (CPU_ICPU_OFF + 27)
+
+#define MEMORY_MAP_OFF                    0x0024
+#define MEMORY_MAP_WRITE_MAP_OFF          0x4000
+#define MEMORY_MAP_SPEED_OFF              0x8000
+#define MEMORY_MAP_BLOCK_IS_RAM_OFF       0x9000
+#define MEMORY_WRITE_MAP_SPEED_OFF        0x4000
+#define MEMORY_WRITE_MAP_BLOCK_IS_RAM_OFF 0x5000
+
+#define SA1_OPCODES_OFF   0
+#define SA1_EXECUTING_OFF 24
+#define SA1_WAIT_COUNTER_OFF 52
+#define SA1_WAIT_BYTE_ADDRESS_1_OFF 56
+#define SA1_WAIT_BYTE_ADDRESS_2_OFF 60
+
+#define FLAG_DECIMAL   0x08
+#define FLAG_INDEX     0x10
+#define FLAG_MEMORY    0x20
+#define FLAG_EMULATION 0x100
+
+#define FUNC(name) \
+  .global name ;   \
+name:
+
+.macro S9xCall func
+  push { ip, lr }
+  bl S9xRegsToMem
+  bl \func
+  bl S9xMemToRegs
+  pop { ip, pc }
+.endm
+
+  .text
+  .align 2
+
+#ifndef NDEBUG
+.macro MetricsIncOffset off
+  push { r0, r1 }
+  ldr r0, = Metrics
+  ldr r1, [r0, #\off]
+  add r1, r1, #1
+  str r1, [r0, #\off]
+  pop { r0, r1 }
+.endm
+
+FUNC(DynaBreak)
+  bx lr
+#endif
+
+FUNC(BlockEnter)
+  push { r4-r11, ip, lr }
+  ldr RegCPU, = CPU
+  bl S9xMemToRegs
+  bx r0
+
+FUNC(BlockReturn)
+  bl S9xRegsToMem
+  pop { r4-r11, ip, pc }
+
+.macro LoadChecks rs
+  ldr \rs, [RegCPU, #CPU_NEXT_EVENT_OFF]
+  ldr RegChecks, [RegCPU, #CPU_FLAGS_OFF]
+  orr RegChecks, RegChecks, \rs, lsl #16
+.endm
+
+.macro PushCPUState
+  str RegCPUPC, [RegCPU, #CPU_PC_OFF]
+  str RegCycles, [RegCPU, #CPU_CYCLES_OFF]
+.endm
+
+.macro PopCPUState rs
+  ldr RegCPUPC, [RegCPU, #CPU_PC_OFF]
+  ldr RegCycles, [RegCPU, #CPU_CYCLES_OFF]
+  LoadChecks \rs
+.endm
+
+.macro S9xUnpackStatusFast
+  strh RegP, [RegCPU, #ICPU_REG_P_OFF]
+  mov r3, #1
+  and r2, r3, RegP, lsr #1
+  sub r2, r2, #1
+  strb r2, [RegCPU, #ICPU_ZERO_OFF]
+  and r2, RegP, #0x80
+  strb r2, [RegCPU, #ICPU_NEGATIVE_OFF]
+  and r2, RegP, r3
+  strb r2, [RegCPU, #ICPU_CARRY_OFF]
+  and r2, r3, RegP, lsr #6
+  strb r2, [RegCPU, #ICPU_OVERFLOW_OFF]
+.endm
+
+.macro S9xPackStatusFast
+  ldrh RegP, [RegCPU, #ICPU_REG_P_OFF]
+  ldrb r2, [RegCPU, #ICPU_ZERO_OFF]
+  ldrb r3, [RegCPU, #ICPU_NEGATIVE_OFF]
+  bic RegP, RegP, #0xc3
+  cmp r2, #0
+  ldrb r2, [RegCPU, #ICPU_CARRY_OFF]
+  orreq RegP, RegP, #0x2
+  tst r3, #0x80
+  ldrb r3, [RegCPU, #ICPU_OVERFLOW_OFF]
+  orrne RegP, RegP, #0x80
+  orr RegP, RegP, r2
+  orr RegP, RegP, r3, lsl #6
+  strh RegP, [RegCPU, #ICPU_REG_P_OFF]
+.endm
+
+  @@ Callers assume r0 is not touched
+FUNC(S9xCallUnpackStatusFast)
+  S9xUnpackStatusFast
+  bx lr
+
+  @@ Preserves scratch and r0, can be used as a wrapper
+FUNC(S9xRegsToMem)
+  push { r0, r1, r2, r3, ip, lr }
+  str RegCPUPC, [RegCPU, #CPU_PC_OFF]
+  str RegCycles, [RegCPU, #CPU_CYCLES_OFF]
+  strh RegA, [RegCPU, #ICPU_REG_A_OFF]
+  strh RegX, [RegCPU, #ICPU_REG_X_OFF]
+  strh RegY, [RegCPU, #ICPU_REG_Y_OFF]
+  S9xUnpackStatusFast
+  pop { r0, r1, r2, r3, ip, pc }
+
+
+  @@ Preserves scratch and r0, can be used as a wrapper
+FUNC(S9xMemToRegs)
+  push { r0, r1, r2, r3, ip, lr }
+  ldr RegCPUPC, [RegCPU, #CPU_PC_OFF]
+  ldr RegCycles, [RegCPU, #CPU_CYCLES_OFF]
+  S9xPackStatusFast
+  ldrh RegA, [RegCPU, #ICPU_REG_A_OFF]
+  ldrh RegX, [RegCPU, #ICPU_REG_X_OFF]
+  ldrh RegY, [RegCPU, #ICPU_REG_Y_OFF]
+  pop { r0, r1, r2, r3, ip, lr }
+  LoadChecks r2
+  bx lr
+
+FUNC(UpdateSection)
+  push { RegP, RegCPU, lr }
+  ldr RegCPU, =CPU
+  ldrh RegP, [RegCPU, #ICPU_REG_P_OFF]
+  bl S9xCallUpdateSection
+  pop { RegP, RegCPU, pc }
+
+FUNC(S9xCallUpdateSection)
+  ldr r1, =CacheSection
+  ldrb r0, [RegCPU, #ICPU_REG_PB_OFF]
+  tst RegP, #FLAG_DECIMAL
+  lsl r0, #4
+  orrne r0, r0, #0x8000
+  tst RegP, #FLAG_EMULATION
+  orrne r0, r0, #0x4000
+  bxne lr
+  tst RegP, #FLAG_INDEX
+  orrne r0, r0, #0x1000
+  tst RegP, #FLAG_MEMORY
+  orrne r0, r0, #0x2000
+  str r0, [r1]
+  bx lr
+
+FUNC(LoadChecks)
+  LoadChecks r2
+  bx lr
+
+FUNC(S9xCallSA1MainLoop)
+  push { ip, lr }
+  bl S9xSA1MainLoop
+  pop { ip, pc }
+
+FUNC(S9xCallSetPCBase)
+  push { ip, lr }
+  bl S9xSetPCBase
+  pop { ip, pc }
+
+.macro S9xCallHandleHBlank type
+FUNC(S9xCallHandleHBlank\type)
+  @ changes cycles, maybe PC in IRQ
+  push { ip, lr }
+  str RegCPUPC, [RegCPU, #CPU_PC_OFF]
+  str RegCycles, [RegCPU, #CPU_CYCLES_OFF]
+  bl HandleHBlank\type
+  ldr RegCPUPC, [RegCPU, #CPU_PC_OFF]
+  ldr RegCycles, [RegCPU, #CPU_CYCLES_OFF]
+  pop { ip, lr }
+  LoadChecks r2
+  bx lr
+.endm
+
+S9xCallHandleHBlank SFX
+S9xCallHandleHBlank NoSFX
+
+.macro S9xCallHandleFlags WFI
+FUNC(S9xCallHandleFlags\WFI)
+  @ needs S, changes flags (unpack status needed), changes PC, changes cycles
+  push { ip, lr }
+  str RegCPUPC, [RegCPU, #CPU_PC_OFF]
+  str RegCycles, [RegCPU, #CPU_CYCLES_OFF]
+  S9xUnpackStatusFast
+  bl HandleFlags\WFI
+  ldr RegCPUPC, [RegCPU, #CPU_PC_OFF]
+  ldr RegCycles, [RegCPU, #CPU_CYCLES_OFF]
+  S9xPackStatusFast
+  pop { ip, lr }
+  LoadChecks r2
+  bx lr
+.endm
+
+S9xCallHandleFlags WFI
+S9xCallHandleFlags NoWFI
+
+.macro S9xCallHandleChecks WFI, SFX
+FUNC(S9xCallHandleChecks\WFI\SFX)
+  push { r1, lr }
+  mov r0, #0
+  cmp RegCycles, RegChecks, lsr #16
+  blhs S9xCallHandleHBlank\SFX
+  cmp r0, #0
+  popne { r1, lr }
+  bne BlockReturn
+
+  tst RegChecks, #0x880
+  blne S9xCallHandleFlags\WFI
+  cmp r0, #0
+  movne RegCPUPC, r0
+  popne { r1, lr }
+  bne JumpIndirect
+
+  pop { r1, pc }
+.endm
+
+S9xCallHandleChecks WFI, SFX
+S9xCallHandleChecks WFI, NoSFX
+S9xCallHandleChecks NoWFI, SFX
+S9xCallHandleChecks NoWFI, NoSFX
+
+#ifndef NDEBUG
+FUNC(S9xCallCheckInstruction)
+  S9xCall CheckInstruction
+#endif
+
+FUNC(JumpIndirect)
+  push { r1, ip }
+  ldr r1, [RegCPU, #CPU_PCBASE_OFF]
+  sub r0, RegCPUPC, r1
+  bl FetchBlock
+  pop { r1, ip }
+  ldr r0, [r0]
+  bx r0
+
+FUNC(JumpDirectChecks)
+  push { r1, ip }
+  mov r0, RegCPUPC
+  sub r1, lr, #4                @ Patch the opcode at the address of this caller
+  bl PatchJumpDirectChecks
+  pop { r1, ip }
+  bx r0
+
+FUNC(JumpDirect)
+  push { r1, ip }
+  mov r0, RegCPUPC
+  sub r1, lr, #4                @ Patch the opcode at the address of this caller
+  bl PatchJumpDirect
+  pop { r1, ip }
+  bx r0
+
+.macro GetMemFast Type
+FUNC(S9xGet\Type\()Fast)
+  push { r0, r3, ip, lr }
+#ifndef NDEBUG
+  MetricsIncOffset METRICS_GETSETS_OFF
+#endif
+  ldr r2, =0x0fff               @ MEMMAP_MASK
+  .ifc \Type,Word
+  and r3, r0, r2
+  cmp r3, r2
+  beq 9f
+  .endif
+
+  add ip, RegCPU, #CPU_MEMORY_MAP_OFF
+  and r3, r2, r0, lsr #12       @ MEMMAP_SHIFT
+  ldr r2, [ip, r3, lsl #2]
+  add lr, ip, #MEMORY_MAP_BLOCK_IS_RAM_OFF
+  cmp r2, #18                   @ MAP_LAST
+  blo 9f
+
+  ldrb lr, [lr, r3]
+  add ip, ip, #MEMORY_MAP_SPEED_OFF
+  ldrb ip, [ip, r3]
+  cmp lr, #1
+  streq r1, [RegCPU, #CPU_WAIT_ADDRESS_OFF]
+
+  mov r1, r0, lsl #16
+  mov r1, r1, lsr #16
+  ldrb r0, [r2, r1]
+  .ifc \Type,Word
+  add r1, r1, #1
+  ldrb r1, [r2, r1]
+  .endif
+  .ifc \Type,Byte
+  add RegCycles, RegCycles, ip
+  .else
+  add RegCycles, RegCycles, ip, lsl #1 @ * 2 for GetWord
+  .endif
+  .ifc \Type,Word
+  orr r0, r0, r1, lsl #8
+  .endif
+  pop { r1, r3, ip, pc }
+9:
+#ifndef NDEBUG
+  MetricsIncOffset METRICS_SLOWGETSETS_OFF
+#endif
+  str r1, [RegCPU, #CPU_PC_AT_OPCODE_START_OFF]
+  PushCPUState
+  bl S9xGet\Type
+  pop { r1, r3, ip, lr }
+  PopCPUState r2
+  bx lr
+.endm
+
+GetMemFast Byte
+GetMemFast Word
+
+  .macro SetMemFast Type, SA1
+FUNC(S9xSet\Type\()Fast\SA1)
+  push { r1, r3, ip, lr}
+#ifndef NDEBUG
+  MetricsIncOffset METRICS_GETSETS_OFF
+#endif
+  ldr r2, =0x0fff               @ MEMMAP_MASK
+  mov ip, #0
+  .ifc \Type,Word
+  and r3, r1, r2
+  cmp r3, r2
+  beq 9f
+  .endif
+
+  str ip, [RegCPU, #CPU_WAIT_ADDRESS_OFF]
+
+  and r3, r2, r1, lsr #12       @ MEMMAP_SHIFT
+  add ip, RegCPU, #(CPU_MEMORY_MAP_OFF + MEMORY_MAP_WRITE_MAP_OFF)
+  ldr r2, [ip, r3, lsl #2]
+  cmp r2, #18                   @ MAP_LAST
+  blo 9f
+
+  add ip, ip, #MEMORY_WRITE_MAP_SPEED_OFF
+  ldrb ip, [ip, r3]
+  .ifc \Type,Byte
+  add RegCycles, RegCycles, ip
+  .else
+  add RegCycles, RegCycles, ip, lsl #1 @ * 2 for GetWord
+  .endif
+
+
+  mov r1, r1, lsl #16
+  mov r1, r1, lsr #16
+
+  @@ Handle SA1 WaitAddress if needed
+  .ifc \SA1,SA1
+  push { r1 }
+  ldr ip, =SA1
+  add r1, r1, r2
+  ldr r3, [ip, #SA1_WAIT_BYTE_ADDRESS_1_OFF]
+  cmp r3, r1
+  beq 8f
+  ldr r3, [ip, #SA1_WAIT_BYTE_ADDRESS_2_OFF]
+  cmp r3, r1
+  beq 8f
+2:
+  pop { r1 }
+  .endif
+
+  strb r0, [r2, r1]
+  .ifc \Type,Word
+  add r1, r1, #1
+  mov r0, r0, lsr #8
+  strb r0, [r2, r1]
+  .endif
+  pop { r1, r3, ip, pc }
+
+  .ifc \SA1,SA1
+8:
+  ldr r1, [ip, #SA1_OPCODES_OFF]
+  mov r3, #0
+  str r3, [ip, #SA1_WAIT_COUNTER_OFF]
+  cmp r1, #0
+  movne r3, #1
+  strb r3, [ip, #SA1_EXECUTING_OFF]
+  b 2b
+  .endif
+9:
+#ifndef NDEBUG
+  MetricsIncOffset METRICS_SLOWGETSETS_OFF
+#endif
+  PushCPUState
+  .ifc \Type,Word
+  mov r0, r0, lsl #16
+  mov r0, r0, lsr #16
+  .else
+  and r0, r0, #0xFF
+  .endif
+  bl S9xSet\Type
+  pop { r1, r3, ip, lr }
+  PopCPUState r2
+  bx lr
+.endm
+
+SetMemFast Byte, NoSA1
+SetMemFast Word, NoSA1
+
+SetMemFast Byte, SA1
+SetMemFast Word, SA1
+
+FUNC(S9xCallFixCycles)
+  tst RegP, #FLAG_EMULATION
+  ldrne r0, =S9xOpcodesE1
+  bne 9f
+
+  tst RegP, #FLAG_MEMORY
+  beq 2f
+  tst RegP, #FLAG_INDEX
+  ldrne r0, =S9xOpcodesM1X1
+  ldreq r0, =S9xOpcodesM1X0
+  b 9f
+2:
+  tst RegP, #FLAG_INDEX
+  ldrne r0, =S9xOpcodesM0X1
+  ldreq r0, =S9xOpcodesM0X0
+9:
+  str r0, [RegCPU, #ICPU_S9X_OPCODES_OFF]
+  bx lr
+
+FUNC(S9xCallCPUShutdown)
+  push { ip, lr }
+  str RegCycles, [RegCPU, #CPU_CYCLES_OFF]
+  bl DynaCPUShutdown
+  ldr RegCycles, [RegCPU, #CPU_CYCLES_OFF]
+  pop { ip, pc }
+
+FUNC(S9xCallWAIShutdown)
+  push { ip, lr }
+  str RegCycles, [RegCPU, #CPU_CYCLES_OFF]
+  bl DynaWAIShutdown
+  ldr RegCycles, [RegCPU, #CPU_CYCLES_OFF]
+  pop { ip, pc }
+
+  .bss
+  .align 4
+  .global CPU
+CPU:
+  .space 0x80 @ Actual: 0x50
+
+  .global ICPU
+ICPU:
+  .space 0x40 @ Actual: 0x34
+
+  .global OpenBus
+OpenBus:
+  .space 0x1
+
+  @ padding so Memory.Map is at 0x100
+  .space 0x3F - 0x24
+
+  .global Memory
+Memory:
+  .space 0xb468
diff --git a/source/arm_dynarec/armfn.h b/source/arm_dynarec/armfn.h
new file mode 100644
index 0000000..4b859f9
--- /dev/null
+++ b/source/arm_dynarec/armfn.h
@@ -0,0 +1,50 @@
+#ifndef ARMFN_H
+#define ARMFN_H
+
+#include <stdint.h>
+
+int BlockEnter(uint8_t *addr);
+void UpdateSection(void);
+
+/* Below can only be called from generated code */
+
+void S9xRegsToMem(void);
+void S9xMemToRegs(void);
+
+void JumpIndirect(uint8_t *pc);
+void JumpDirect(uint8_t *pc);
+void JumpDirectChecks(uint8_t *pc);
+
+uint16_t S9xGetByteFast(uint32_t Address, uint8_t *StartPC);
+uint16_t S9xGetWordFast(uint32_t Address, uint8_t *StartPC);
+void S9xSetByteFastSA1(uint16_t Word, uint32_t Address);
+void S9xSetWordFastSA1(uint16_t Word, uint32_t Address);
+void S9xSetByteFastNoSA1(uint16_t Word, uint32_t Address);
+void S9xSetWordFastNoSA1(uint16_t Word, uint32_t Address);
+
+#define S9xSetByteFast (Settings.SA1 ? S9xSetByteFastSA1 : S9xSetByteFastNoSA1)
+#define S9xSetWordFast (Settings.SA1 ? S9xSetWordFastSA1 : S9xSetWordFastNoSA1)
+
+void S9xCallUnpackStatusFast(void);
+void S9xCallUpdateSection(void);
+
+void S9xCallSA1MainLoop(void);
+void S9xCallSetPCBase(uint32_t Address);
+
+void S9xCallHandleChecksWFISFX(void);
+void S9xCallHandleChecksWFINoSFX(void);
+void S9xCallHandleChecksNoWFISFX(void);
+void S9xCallHandleChecksNoWFINoSFX(void);
+
+#define S9xCallHandleChecksWFI (Settings.SuperFX ? S9xCallHandleChecksWFISFX : S9xCallHandleChecksWFINoSFX)
+#define S9xCallHandleChecksNoWFI (Settings.SuperFX ? S9xCallHandleChecksNoWFISFX : S9xCallHandleChecksNoWFINoSFX)
+
+void S9xCallFixCycles(void);
+void S9xCallCPUShutdown(void);
+void S9xCallWAIShutdown(void);
+
+#ifndef NDEBUG
+void S9xCallCheckInstruction(uint8_t opcode, uint8_t *pc);
+#endif
+
+#endif  /* ARMFN_H */
diff --git a/source/arm_dynarec/armgen.h b/source/arm_dynarec/armgen.h
new file mode 100644
index 0000000..bec7a58
--- /dev/null
+++ b/source/arm_dynarec/armgen.h
@@ -0,0 +1,576 @@
+#include <stdbool.h>
+#include <stdint.h>
+#include <stdlib.h>
+
+static uint8_t *CachePtr;
+
+enum ARMCond
+{
+   CondEQ,
+   CondNE,
+   CondCS,
+   CondHS = CondCS,
+   CondCC,
+   CondLO = CondCC,
+   CondMI,
+   CondPL,
+   CondVS,
+   CondVC,
+   CondHI,
+   CondLS,
+   CondGE,
+   CondLT,
+   CondGT,
+   CondLE,
+   CondAL,
+};
+
+enum ARMShift
+{
+   ShiftLSL,
+   ShiftLSR,
+   ShiftASR,
+   ShiftROR,
+   ShiftRRX
+};
+
+enum ARMReg
+{
+   RegR0,
+   RegR1,
+   RegR2,
+   RegR3,
+   RegR4,
+   RegR5,
+   RegR6,
+   RegR7,
+   RegR8,
+   RegR9,
+   RegR10,
+   RegR11,
+   RegR12,
+   RegR13,
+   RegR14,
+   RegR15,
+
+   RegA1 = RegR0,
+   RegA2,
+   RegA3,
+   RegA4,
+   RegV1 = RegR4,
+   RegV2,
+   RegV3,
+   RegV4,
+   RegV5,
+   RegV6,
+   RegV7,
+   RegV8,
+   RegIP = RegR12,
+   RegSP,
+   RegLR,
+   RegPC,
+
+   RegA = RegR4,
+   RegX = RegR5,
+   RegY = RegR6,
+   RegP = RegR7,
+   RegCPU = RegR8,
+   RegCPUPC = RegR9,
+   RegCycles = RegR10,
+
+   RegChecks = RegR12,
+};
+
+#define ASCOND(cond) (cond << 28)
+#define ASLIST(reg) (1 << reg)
+#define PCREL(pc, target) ((intptr_t)(target) - (intptr_t)(pc) - 8)
+#define PCOFFSET(function) ((PCREL(CachePtr, function) >> 2) & 0xFFFFFF)
+#define ENCODE_IMM(imm, shl) (((((32 - (shl)) & 0x1F) >> 1) << 8) | ((imm) & 0xFF))
+#define DECODE_IMM(imm) (((imm) & 0xFF) << (32 - (((imm) >> 8) << 1)));
+#define ENCODE_SHIFT(type, imm) ((type) | (((imm) >= 32 ? 0 : (imm)) << 2))
+
+#define CPU_ICPU_OFFSET     ((uintptr_t)&ICPU - (uintptr_t)&CPU)
+#define CPU_OPEN_BUS_OFFSET ((uintptr_t)&OpenBus - (uintptr_t)&CPU)
+
+#define CPU_OFFSET(field) (offsetof(__typeof__(CPU), field))
+#define ICPU_OFFSET(field) (CPU_ICPU_OFFSET + offsetof(__typeof__(ICPU), field))
+
+static size_t DecomposeImm32(uint32_t imm32, uint32_t *result)
+{
+   uint8_t shift = 0;
+   size_t count = 0;
+
+   if (!imm32) {
+      result[0] = 0;
+      return 1;
+   }
+
+   while (1)
+   {
+      while (shift < 32 && !(imm32 >> shift & 0x03))
+         shift += 2;
+
+      if (shift >= 32)
+         break;
+
+      if (shift > 24)
+      {
+         uint8_t remaining = (1 << (shift - 24));
+         uint32_t firstStore = count == 0 ? 0 : DECODE_IMM(result[0]);
+
+         if (firstStore && firstStore < remaining)
+         {
+            uint8_t bottom = imm32 >> shift;
+            uint8_t top = (imm32 & (remaining - 1)) << (32 - shift);
+            result[0] = ENCODE_IMM(bottom | top, shift);
+            break;
+         }
+      }
+
+      result[count++] = ENCODE_IMM(imm32 >> shift, shift);
+      shift += 8;
+   }
+
+   return count;
+}
+
+static inline void ARMEmit(uint32_t value)
+{
+   *(uint32_t *)CachePtr = value;
+   CachePtr += 4;
+}
+
+enum ARMDPIOp
+{
+   DPIOpAND,
+   DPIOpEOR,
+   DPIOpSUB,
+   DPIOpRSB,
+   DPIOpADD,
+   DPIOpADC,
+   DPIOpSBC,
+   DPIOpRSC,
+   DPIOpTST,
+   DPIOpTEQ,
+   DPIOpCMP,
+   DPIOpCMN,
+   DPIOpORR,
+   DPIOpMOV,
+   DPIOpBIC,
+   DPIOpNOT,
+};
+
+static void ARMEmitDPI_Imm(enum ARMCond cond, uint8_t op, bool s, enum ARMReg rd, enum ARMReg rn, uint16_t imm)
+{
+   ARMEmit(
+      ASCOND(cond) |
+      0 << 26 | /* 0b00 */
+      1 << 25 |
+      op << 21 |
+      s << 20 |
+      rn << 16 |
+      rd << 12 |
+      imm
+   );
+}
+
+static void ARMEmitDPI_Reg(enum ARMCond cond, uint8_t op, bool s, enum ARMReg rd, enum ARMReg rn, uint8_t shift, enum ARMReg rm)
+{
+   ARMEmit(
+      ASCOND(cond) |
+      0 << 26 | /* 0b00 */
+      0 << 25 |
+      op << 21 |
+      s << 20 |
+      rn << 16 |
+      rd << 12 |
+      shift << 5 |
+      0 << 4 |
+      rm
+   );
+}
+
+static void ARMEmitDPI_RegShift(enum ARMCond cond, uint8_t op, bool s, enum ARMReg rd, enum ARMReg rn, enum ARMReg rs, uint8_t type, enum ARMReg rm)
+{
+   ARMEmit(
+      ASCOND(cond) |
+      0 << 26 | /* 0b00 */
+      0 << 25 |
+      op << 21 |
+      s << 20 |
+      rn << 16 |
+      rd << 12 |
+      rs << 8 |
+      0 << 7 |
+      type << 5 |
+      1 << 4 |
+      rm
+   );
+}
+
+static void GenARM_CMP_Imm(enum ARMCond cond, enum ARMReg reg, uint16_t imm)
+{
+   ARMEmitDPI_Imm(cond, DPIOpCMP, 1, 0, reg, imm);
+}
+
+static void GenARM_CMP_Reg(enum ARMCond cond, enum ARMReg rn, uint8_t shift, enum ARMReg rm)
+{
+   ARMEmitDPI_Reg(cond, DPIOpCMP, 1, 0, rn, shift, rm);
+}
+
+static void GenARM_SUB_Imm(enum ARMCond cond, enum ARMReg rd, enum ARMReg rn, uint16_t imm)
+{
+   ARMEmitDPI_Imm(cond, DPIOpSUB, 0, rd, rn, imm);
+}
+
+static void GenARM_SUBS_Imm(enum ARMCond cond, enum ARMReg rd, enum ARMReg rn, uint16_t imm)
+{
+   ARMEmitDPI_Imm(cond, DPIOpSUB, 1, rd, rn, imm);
+}
+
+static void GenARM_SUB_Reg(enum ARMCond cond, enum ARMReg rd, enum ARMReg rn, uint8_t shift, enum ARMReg rm)
+{
+   ARMEmitDPI_Reg(cond, DPIOpSUB, 0, rd, rn, shift, rm);
+}
+
+static void GenARM_SUBS_Reg(enum ARMCond cond, enum ARMReg rd, enum ARMReg rn, uint8_t shift, enum ARMReg rm)
+{
+   ARMEmitDPI_Reg(cond, DPIOpSUB, 1, rd, rn, shift, rm);
+}
+
+static void GenARM_ADD_Imm(enum ARMCond cond, enum ARMReg rd, enum ARMReg rn, uint16_t imm)
+{
+   ARMEmitDPI_Imm(cond, DPIOpADD, 0, rd, rn, imm);
+}
+
+static void GenARM_ADD_Reg(enum ARMCond cond, enum ARMReg rd, enum ARMReg rn, uint8_t shift, enum ARMReg rm)
+{
+   ARMEmitDPI_Reg(cond, DPIOpADD, 0, rd, rn, shift, rm);
+}
+
+static void GenARM_ADDS_Reg(enum ARMCond cond, enum ARMReg rd, enum ARMReg rn, uint8_t shift, enum ARMReg rm)
+{
+   ARMEmitDPI_Reg(cond, DPIOpADD, 1, rd, rn, shift, rm);
+}
+
+static void GenARM_MOV_Imm(enum ARMCond cond, enum ARMReg rd, uint16_t imm)
+{
+   ARMEmitDPI_Imm(cond, DPIOpMOV, 0, rd, 0, imm);
+}
+
+static void GenARM_MOV_Reg(enum ARMCond cond, enum ARMReg rd, uint8_t shift, enum ARMReg rm)
+{
+   ARMEmitDPI_Reg(cond, DPIOpMOV, 0, rd, 0, shift, rm);
+}
+
+static void GenARM_MOVS_Reg(enum ARMCond cond, enum ARMReg rd, uint8_t shift, enum ARMReg rm)
+{
+   ARMEmitDPI_Reg(cond, DPIOpMOV, 1, rd, 0, shift, rm);
+}
+
+static void GenARM_AND_Imm(enum ARMCond cond, enum ARMReg rd, enum ARMReg rn, uint16_t imm)
+{
+   ARMEmitDPI_Imm(cond, DPIOpAND, 0, rd, rn, imm);
+}
+
+static void GenARM_ANDS_Imm(enum ARMCond cond, enum ARMReg rd, enum ARMReg rn, uint16_t imm)
+{
+   ARMEmitDPI_Imm(cond, DPIOpAND, 1, rd, rn, imm);
+}
+
+static void GenARM_AND_Reg(enum ARMCond cond, enum ARMReg rd, enum ARMReg rn, uint8_t shift, enum ARMReg rm)
+{
+   ARMEmitDPI_Reg(cond, DPIOpAND, 0, rd, rn, shift, rm);
+}
+
+static void GenARM_ANDS_Reg(enum ARMCond cond, enum ARMReg rd, enum ARMReg rn, uint8_t shift, enum ARMReg rm)
+{
+   ARMEmitDPI_Reg(cond, DPIOpAND, 1, rd, rn, shift, rm);
+}
+
+static void GenARM_EOR_Imm(enum ARMCond cond, enum ARMReg rd, enum ARMReg rn, uint16_t imm)
+{
+   ARMEmitDPI_Imm(cond, DPIOpEOR, 0, rd, rn, imm);
+}
+
+static void GenARM_EORS_Imm(enum ARMCond cond, enum ARMReg rd, enum ARMReg rn, uint16_t imm)
+{
+   ARMEmitDPI_Imm(cond, DPIOpEOR, 1, rd, rn, imm);
+}
+
+static void GenARM_EOR_Reg(enum ARMCond cond, enum ARMReg rd, enum ARMReg rn, uint8_t shift, enum ARMReg rm)
+{
+   ARMEmitDPI_Reg(cond, DPIOpEOR, 0, rd, rn, shift, rm);
+}
+
+static void GenARM_EORS_Reg(enum ARMCond cond, enum ARMReg rd, enum ARMReg rn, uint8_t shift, enum ARMReg rm)
+{
+   ARMEmitDPI_Reg(cond, DPIOpEOR, 1, rd, rn, shift, rm);
+}
+
+static void GenARM_ORR_Imm(enum ARMCond cond, enum ARMReg rd, enum ARMReg rn, uint16_t imm)
+{
+   ARMEmitDPI_Imm(cond, DPIOpORR, 0, rd, rn, imm);
+}
+
+static void GenARM_ORRS_Imm(enum ARMCond cond, enum ARMReg rd, enum ARMReg rn, uint16_t imm)
+{
+   ARMEmitDPI_Imm(cond, DPIOpORR, 1, rd, rn, imm);
+}
+
+static void GenARM_ORR_Reg(enum ARMCond cond, enum ARMReg rd, enum ARMReg rn, uint8_t shift, enum ARMReg rm)
+{
+   ARMEmitDPI_Reg(cond, DPIOpORR, 0, rd, rn, shift, rm);
+}
+
+static void GenARM_ORRS_Reg(enum ARMCond cond, enum ARMReg rd, enum ARMReg rn, uint8_t shift, enum ARMReg rm)
+{
+   ARMEmitDPI_Reg(cond, DPIOpORR, 1, rd, rn, shift, rm);
+}
+
+static void GenARM_TST_Imm(enum ARMCond cond, enum ARMReg rn, uint16_t imm)
+{
+   ARMEmitDPI_Imm(cond, DPIOpTST, 1, 0, rn, imm);
+}
+
+static void GenARM_TST_Reg(enum ARMCond cond, enum ARMReg rn, uint8_t shift, enum ARMReg rm)
+{
+   ARMEmitDPI_Reg(cond, DPIOpTST, 1, 0, rn, shift, rm);
+}
+
+static void GenARM_BIC_Imm(enum ARMCond cond, enum ARMReg rd, enum ARMReg rn, uint16_t imm)
+{
+   ARMEmitDPI_Imm(cond, DPIOpBIC, 0, rd, rn, imm);
+}
+
+static void GenARM_MVN_Reg(enum ARMCond cond, enum ARMReg rd, uint8_t shift, enum ARMReg rm)
+{
+   ARMEmitDPI_Reg(cond, DPIOpNOT, 0, rd, 0, shift, rm);
+}
+
+static void ARMEmitLS_Imm(enum ARMCond cond, bool p, bool w, uint8_t op, enum ARMReg rt, enum ARMReg rn, int16_t imm)
+{
+   bool u = imm >= 0 ? 1 : 0;
+   imm = (u ? imm : -imm) & 0xFFF;
+
+   ARMEmit(
+      ASCOND(cond) |
+      1 << 26 | /* 0b01 */
+      0 << 25 |
+      p << 24 |
+      u << 23 |
+      w << 21 |
+      op << 20 |
+      rn << 16 |
+      rt << 12 |
+      imm
+   );
+}
+
+static void ARMEmitLS_Reg(enum ARMCond cond, bool p, bool u, bool w, uint8_t op, enum ARMReg rt, enum ARMReg rn, uint8_t shift, enum ARMReg rm)
+{
+   ARMEmit(
+      ASCOND(cond) |
+      1 << 26 | /* 0b01 */
+      1 << 25 |
+      p << 24 |
+      u << 23 |
+      w << 21 |
+      op << 20 |
+      rn << 16 |
+      rt << 12 |
+      shift << 5 |
+      0 << 4 |
+      rm
+   );
+}
+
+static void ARMEmitLSH_Imm(enum ARMCond cond, bool p, bool w, uint8_t op, enum ARMReg rt, enum ARMReg rn, int16_t imm)
+{
+   bool u = imm >= 0 ? 1 : 0;
+   imm = (u ? imm : -imm) & 0xFF;
+
+   ARMEmit(
+      ASCOND(cond) |
+      0 << 25 | /* 0b000 */
+      p << 24 |
+      u << 23 |
+      1 << 22 |
+      w << 21 |
+      op << 20 |
+      rn << 16 |
+      rt << 12 |
+      ((imm & 0xF0) << 4) |
+      0xb << 4 | /* 0b1011 */
+      (imm & 0x0F)
+   );
+}
+
+static void ARMEmitLSB_Imm(enum ARMCond cond, bool p, bool w, uint8_t op, enum ARMReg rt, enum ARMReg rn, int16_t imm)
+{
+   bool u = imm >= 0 ? 1 : 0;
+   imm = (u ? imm : -imm) & 0xFF;
+
+   ARMEmit(
+      ASCOND(cond) |
+      0x2 << 25 | /* 0b010 */
+      p << 24 |
+      u << 23 |
+      1 << 22 |
+      w << 21 |
+      op << 20 |
+      rn << 16 |
+      rt << 12 |
+      imm
+   );
+}
+
+static void GenARM_LDR_Imm(enum ARMCond cond, enum ARMReg rt, enum ARMReg rn, int16_t imm)
+{
+   ARMEmitLS_Imm(cond, 1, 0, 1, rt, rn, imm);
+}
+
+static void GenARM_LDR_Reg(enum ARMCond cond, enum ARMReg rt, enum ARMReg rn, bool add, uint8_t shift, enum ARMReg rm)
+{
+   ARMEmitLS_Reg(cond, 1, add, 0, 1, rt, rn, shift, rm);
+}
+
+static void GenARM_LDRH_Imm(enum ARMCond cond, enum ARMReg rt, enum ARMReg rn, int16_t imm)
+{
+   ARMEmitLSH_Imm(cond, 1, 0, 1, rt, rn, imm);
+}
+
+static void GenARM_LDRB_Imm(enum ARMCond cond, enum ARMReg rt, enum ARMReg rn, int16_t imm)
+{
+   ARMEmitLSB_Imm(cond, 1, 0, 1, rt, rn, imm);
+}
+
+static void GenARM_STR_Imm(enum ARMCond cond, enum ARMReg rt, enum ARMReg rn, int16_t imm)
+{
+   ARMEmitLS_Imm(cond, 1, 0, 0, rt, rn, imm);
+}
+
+static void GenARM_STRH_Imm(enum ARMCond cond, enum ARMReg rt, enum ARMReg rn, int16_t imm)
+{
+   ARMEmitLSH_Imm(cond, 1, 0, 0, rt, rn, imm);
+}
+
+static void GenARM_STRB_Imm(enum ARMCond cond, enum ARMReg rt, enum ARMReg rn, int16_t imm)
+{
+   ARMEmitLSB_Imm(cond, 1, 0, 0, rt, rn, imm);
+}
+
+static void GenARM_STR_Reg(enum ARMCond cond, enum ARMReg rt, enum ARMReg rn, bool add, uint8_t shift, enum ARMReg rm)
+{
+   ARMEmitLS_Reg(cond, 1, add, 0, 0, rt, rn, shift, rm);
+}
+
+static void GenARM_PUSH(uint16_t regs)
+{
+   ARMEmit(
+      ASCOND(CondAL) |
+      (0x92 << 20) | /* 0b10010010 */
+      (RegSP << 16) |
+      (regs)
+   );
+}
+
+static void GenARM_POP(uint16_t regs)
+{
+   ARMEmit(
+      ASCOND(CondAL) |
+      (0x8B << 20) | /* 0b10001011 */
+      (RegSP << 16) |
+      (regs)
+   );
+}
+
+static void GenARM_B(enum ARMCond cond, uint32_t offset)
+{
+   ARMEmit(
+      ASCOND(cond) |
+      (0xA << 24) | /* 0b1010 */
+      (offset)
+   );
+}
+
+static void GenARM_BL(enum ARMCond cond, uint32_t offset)
+{
+   ARMEmit(
+      ASCOND(cond) |
+      (0xB << 24) | /* 0b1011 */
+      (offset)
+   );
+}
+
+static void GenARM_BX_Reg(enum ARMCond cond, enum ARMReg reg)
+{
+   ARMEmit(
+      ASCOND(cond) |
+      (0x12 << 20) | /* 0b 0001 0010 */
+      (0xFFF1 << 4) | /* 0b 1111 1111 1111 0001 */
+      (reg)
+    );
+}
+
+static void GenARM_BLX_Reg(enum ARMCond cond, enum ARMReg reg)
+{
+   ARMEmit(
+      ASCOND(cond) |
+      (0x12 << 20) | /* 0b 0001 0010 */
+      (0xFFF3 << 4) | /* 0b 1111 1111 1111 0011 */
+      (reg)
+    );
+}
+
+static void GenARM_UXTB_Reg(enum ARMCond cond, enum ARMReg rd, enum ARMReg rm)
+{
+   GenARM_AND_Imm(cond, rd, rm, ENCODE_IMM(0xFF, 0));
+}
+
+static void GenARM_MOVB_Reg(enum ARMCond cond, enum ARMReg rd, enum ARMReg rm)
+{
+   GenARM_MOV_Reg(cond, rd, ENCODE_SHIFT(ShiftLSR, 8), rd);
+   GenARM_ORR_Reg(cond, rd, rd, ENCODE_SHIFT(ShiftLSL, 24), rm);
+   GenARM_MOV_Reg(cond, rd, ENCODE_SHIFT(ShiftROR, 24), rd);
+}
+
+static void GenARM_MOVBS_Reg(enum ARMCond cond, enum ARMReg rd, enum ARMReg scratch, enum ARMReg rm)
+{
+   GenARM_AND_Imm(cond, rd, rd, ENCODE_IMM(0xFF, 8));
+   GenARM_ANDS_Imm(CondAL, scratch, rm, ENCODE_IMM(0xFF, 0));
+   GenARM_ORR_Reg(CondAL, rd, rd, 0, scratch);
+}
+
+static void GenARM_UXTHS_Reg_(enum ARMCond cond, bool s, enum ARMReg rd, enum ARMReg rm)
+{
+   GenARM_MOV_Reg(cond, rd, ENCODE_SHIFT(ShiftLSL, 16), rm);
+   if (s)
+      GenARM_MOVS_Reg(cond, rd, ENCODE_SHIFT(ShiftLSR, 16), rd);
+   else
+      GenARM_MOV_Reg(cond, rd, ENCODE_SHIFT(ShiftLSR, 16), rd);
+}
+
+static void GenARM_UXTH_Reg(enum ARMCond cond, enum ARMReg rd, enum ARMReg rm)
+{
+   GenARM_UXTHS_Reg_(cond, 0, rd, rm);
+}
+
+static void GenARM_UXTHS_Reg(enum ARMCond cond, enum ARMReg rd, enum ARMReg rm)
+{
+   GenARM_UXTHS_Reg_(cond, 1, rd, rm);
+}
+
+static void GenARM_MOV32_Imm(enum ARMCond cond, enum ARMReg reg, uint32_t imm)
+{
+   uint32_t values[4];
+   uint32_t count = DecomposeImm32(imm, values);
+   uint32_t i;
+
+   GenARM_MOV_Imm(cond, reg, values[0]);
+
+   for (i = 1; i < count; i++)
+   {
+      GenARM_ORR_Imm(cond, reg, reg, values[i]);
+   }
+}
diff --git a/source/arm_dynarec/dynaexec.c b/source/arm_dynarec/dynaexec.c
new file mode 100644
index 0000000..d63d710
--- /dev/null
+++ b/source/arm_dynarec/dynaexec.c
@@ -0,0 +1,3049 @@
+#include <stdbool.h>
+#include <stdint.h>
+#include <stdio.h>
+#include <stdlib.h>
+
+#include "snes9x.h"
+#include "cpuexec.h"
+#include "sa1.h"
+#include "apu.h"
+#include "arm_dynarec/armfn.h"
+#include "arm_dynarec/armgen.h"
+#include "arm_dynarec/dynaexec.h"
+
+#define BUFFER_SIZE (6 << 20)
+#define BUFFER_EXTRA (1 << 20)
+#define BLOCK_SIZE 0x1000
+
+#ifndef NDEBUG
+int trace[128];
+int traceidx;
+int BreakPC;
+int BreakOpcode;
+#endif
+
+enum SNESRegArg
+{
+   ArgNULL,
+   ArgA,
+   ArgX,
+   ArgY,
+   ArgS,
+   ArgD,
+   ArgDB,
+   ArgPB,
+   ArgP,
+   ArgZ,
+};
+
+enum OpcodeFlag
+{
+   OFNone          = 0,
+   OFBreak         = 1 << 0,
+   OFSectionUpdate = 1 << 1,
+};
+
+#define V(val) val##8, val##16
+
+enum OpAddrMode
+{
+   AddrNone,
+   V(AddrRegister),
+   V(AddrImmediate),
+   AddrImmediate24,
+   V(AddrZeroPage),
+   V(AddrIndirect),
+   V(AddrIndirectX),
+   V(AddrIndirectY),
+   V(AddrIndirectS),
+   V(AddrIndirectFar),
+   V(AddrAbsolute),
+   V(AddrLong),
+};
+
+enum OpFunction
+{
+   V(FnMOV),
+   V(FnMOVSP),
+   V(FnLD),
+   V(FnST),
+   V(FnPUSH),
+   FnPEA,
+   FnPER,
+   V(FnPOP),
+   FnMVN,
+   FnMVP,
+   V(FnOR),
+   V(FnAND),
+   V(FnEOR),
+   V(FnADC),
+   V(FnSBC),
+   V(FnCMP),
+   V(FnBIT),
+   V(FnINC),
+   V(FnDEC),
+   V(FnTSB),
+   V(FnTRB),
+   V(FnASL),
+   V(FnLSR),
+   V(FnROL),
+   V(FnROR),
+   FnBRA,
+   FnBRL,
+   FnJMP,
+   FnJML,
+   FnJSR,
+   FnJSL,
+   FnRTI,
+   FnRTL,
+   FnRTS,
+   FnBPL,
+   FnBMI,
+   FnBVC,
+   FnBVS,
+   FnBCC,
+   FnBCS,
+   FnBZC,
+   FnBZS,
+   FnBRK,
+   FnCOP,
+   FnCLC,
+   FnCLI,
+   FnCLD,
+   FnCLV,
+   FnSEC,
+   FnSEI,
+   FnSED,
+   FnREP,
+   FnSEP,
+   FnXCE,
+   FnSTP,
+   FnXBA,
+   FnWAI,
+   FnWDM,
+   FnNOP,
+};
+
+#undef V
+
+#define CHECK_8_16(field, value) (in->field == value##8 || in->field == value##16 )
+#define CHECK_FIELD(field, value) (in->field == value)
+#define IS_VOP(value) (CHECK_8_16(Op, Fn##value))
+#define IS_OP(value) (CHECK_FIELD(Op, Fn##value))
+#define IS_VADDRMODE(value) (CHECK_8_16(AddrMode, Addr##value))
+#define IS_ADDRMODE(value) (CHECK_FIELD(AddrMode, Addr##value))
+
+#define CASE_VOP(value) case Fn##value##8: case Fn##value##16
+#define CASE_OP(value) case Fn##value
+
+enum ChecksFlags
+{
+   CheckWFI = 1,
+};
+
+struct Instruction;
+
+typedef struct Instruction {
+   void (*S9xOpcode)(void);
+   void (*Emitter)(struct Instruction *);
+   uint8_t *PC;
+   uint8_t Opcode;
+   enum OpAddrMode AddrMode;
+   enum OpFunction Op;
+   enum SNESRegArg Arg1;
+   enum SNESRegArg Arg2;
+   uint16_t OutFlags;
+   uint16_t InFlags;
+   bool SectionUpdate;
+   bool ShouldBreak;
+} Instruction;
+
+uint32_t CacheSection;
+static uint8_t *CachePtr;
+static uint8_t *Cache;
+static uint8_t **Sections[13 * 0x1000]; /* Max: Decimal | Emulation = 13 */
+
+#ifndef NDEBUG
+struct
+{
+   uint32_t Emits;
+   uint32_t Finds;
+   uint32_t InterpretedFinds;
+   bool     InterpretedBlock;
+   uint32_t Ops;
+   uint32_t InterpretedOps;
+   uint32_t ContextSwitches;
+   uint32_t GetSets;
+   uint32_t SlowGetSets;
+   uint32_t OpCounts[0x100];
+   uint32_t InterpretedOpCounts[0x100];
+   uint32_t InterpretedBlockStarts[0x100];
+} Metrics;
+
+void MetricsReset(void)
+{
+   memset(&Metrics, 0, sizeof(Metrics));
+}
+
+void MetricsPrint(void)
+{
+   int i;
+   printf("Cache hit rate: %f\n", (float)(Metrics.Finds - Metrics.Emits) / Metrics.Finds);
+   printf("Avg. instructions per find: %f\n", (float)Metrics.Ops / Metrics.Finds);
+
+   printf("Counts by instruction: \n");
+   for (i = 0; i < sizeof(Metrics.OpCounts) / sizeof(Metrics.OpCounts[0]); i++)
+   {
+      printf("0x%X: %d\n", i, Metrics.OpCounts[i]);
+   }
+
+   printf("Interpreted counts by instruction: \n");
+   for (i = 0; i < sizeof(Metrics.InterpretedOpCounts) / sizeof(Metrics.InterpretedOpCounts[0]); i++)
+   {
+      if (Metrics.InterpretedOpCounts[i] > 0)
+         printf("0x%X: %d\n", i, Metrics.InterpretedOpCounts[i]);
+   }
+
+   printf("Interpreted block starts by instruction: \n");
+   for (i = 0; i < sizeof(Metrics.InterpretedBlockStarts) / sizeof(Metrics.InterpretedBlockStarts[0]); i++)
+   {
+      if (Metrics.InterpretedBlockStarts[i] > 0)
+         printf("0x%X: %d\n", i, Metrics.InterpretedBlockStarts[i]);
+   }
+}
+#endif
+
+uint8_t **FetchBlock(uint16_t pc);
+
+static void *CacheAlloc(size_t size)
+{
+   uint8_t *prev = CachePtr;
+   CachePtr += size;
+   return (void *)prev;
+}
+
+static void CacheEmpty(void)
+{
+   memset(Cache, 0, BUFFER_SIZE + BUFFER_EXTRA);
+   memset(Sections, 0, sizeof(Sections));
+   CachePtr = Cache;
+}
+
+static uint8_t **FindBlock(uint16_t pc)
+{
+   /* Section is D/E/X/M flags, PB, and top 4 bits of PC */
+   uint32_t section = (CacheSection | ((pc & 0xF000) >> 12));
+   uint8_t **block = Sections[section];
+   pc &= ~0xF000;
+
+   if (!block || !block[pc])
+   {
+         if (CachePtr >= Cache + BUFFER_SIZE)
+         {
+            CacheEmpty();
+            printf("Reset translation cache\n");
+            block = NULL;
+         }
+
+         if (!block)
+            block = Sections[section] = CacheAlloc(BLOCK_SIZE * sizeof(uint8_t *));
+   }
+
+   return &block[pc];
+}
+
+
+#ifndef NDEBUG
+
+void CheckInstruction(uint8_t opcode, uint8_t *pc, bool interpreted)
+{
+   Metrics.Ops++;
+   Metrics.OpCounts[opcode]++;
+
+   if (interpreted)
+   {
+      Metrics.InterpretedOps++;
+      Metrics.InterpretedOpCounts[opcode]++;
+      if (!Metrics.InterpretedBlock)
+      {
+         Metrics.InterpretedBlock = true;
+         Metrics.InterpretedFinds++;
+         Metrics.InterpretedBlockStarts[opcode]++;
+      }
+   }
+
+   if (CPU.PC - CPU.PCBase == (BreakPC & 0xffff) && ICPU.Registers.PB == ((BreakPC >> 16) & 0xFF))
+      DynaBreak();
+
+   if (*CPU.PC == BreakOpcode)
+      DynaBreak();
+
+   if (pc != CPU.PC)
+   {
+      fprintf(stderr, "Incorrect PC: Expected 0x%X, was 0x%X\n", (uintptr_t)pc, (uintptr_t)CPU.PC);
+      DynaBreak();
+   }
+
+   if (opcode != *CPU.PC)
+   {
+      fprintf(stderr, "Incorrect opcode: Expected 0x%X, was 0x%X\n", opcode, *CPU.PC);
+      DynaBreak();
+   }
+}
+#endif
+
+static void EmitChecks(enum ChecksFlags flags)
+{
+   if (Settings.SA1)
+   {
+      GenARM_MOV32_Imm(CondAL, RegR0, (uintptr_t)&SA1);
+      GenARM_LDRB_Imm(CondAL, RegR0, RegR0, offsetof(__typeof__(SA1), Executing));
+      GenARM_CMP_Imm(CondAL, RegR0, ENCODE_IMM(0, 0));
+      GenARM_BL(CondNE, PCOFFSET(S9xCallSA1MainLoop));
+   }
+
+   GenARM_ANDS_Imm(CondAL, RegR0, RegChecks, ENCODE_IMM(0x88, 4)); /* 0x880 = NMI_FLAG | IRQ_PENDING_FLAG */
+   GenARM_CMP_Reg(CondEQ, RegCycles, ENCODE_SHIFT(ShiftLSR, 16), RegChecks);
+   GenARM_ORRS_Imm(CondHS, RegR0, RegR0, ENCODE_IMM(1, 0));
+
+   if (flags & CheckWFI)
+      GenARM_BL(CondNE, PCOFFSET(S9xCallHandleChecksWFI));
+   else
+      GenARM_BL(CondNE, PCOFFSET(S9xCallHandleChecksNoWFI));
+}
+
+static void PatchJump(uint32_t *source, uint8_t *target)
+{
+   *source = ((*source & 0xF0000000) | (0xA << 24) | ((PCREL(source, target) >> 2) & 0xFFFFFF));
+}
+
+static uint8_t *HandleFlags(bool checkWFI)
+{
+   uint8_t *NewPC = NULL;
+
+   if (CPU.Flags & NMI_FLAG)
+   {
+      if (--CPU.NMICycleCount == 0)
+      {
+         CPU.Flags &= ~NMI_FLAG;
+         if (checkWFI && CPU.WaitingForInterrupt)
+         {
+            CPU.WaitingForInterrupt = false;
+            CPU.PC++;
+         }
+         S9xOpcode_NMI();
+         UpdateSection();
+         NewPC = CPU.PC;
+      }
+   }
+
+   if (CPU.Flags & IRQ_PENDING_FLAG)
+   {
+      if (CPU.IRQCycleCount == 0)
+      {
+         if (checkWFI && CPU.WaitingForInterrupt)
+         {
+            CPU.WaitingForInterrupt = false;
+            NewPC = CPU.PC++;
+         }
+         if (CPU.IRQActive && !Settings.DisableIRQ)
+         {
+            if (!CheckFlag(IRQ))
+            {
+               S9xOpcode_IRQ();
+               UpdateSection();
+               NewPC = CPU.PC;
+            }
+         }
+         else
+            CPU.Flags &= ~IRQ_PENDING_FLAG;
+      }
+      else if (--CPU.IRQCycleCount == 0 && CheckFlag(IRQ))
+         CPU.IRQCycleCount = 1;
+   }
+
+   return NewPC;
+}
+
+uint8_t *HandleFlagsNoWFI(void)
+{
+   return HandleFlags(false);
+}
+
+uint8_t *HandleFlagsWFI(void)
+{
+   return HandleFlags(true);
+}
+
+static int HandleHBlank(bool SFX)
+{
+#ifndef NDEBUG
+   Metrics.ContextSwitches++;
+#endif
+
+#ifndef USE_BLARGG_APU
+   if (CPU.WhichEvent == HBLANK_END_EVENT)
+      while (IAPU.APUExecuting && APU.Cycles <= CPU.Cycles)
+         APU_EXECUTE1();
+#endif
+
+   if (SFX)
+      S9xDoHBlankProcessing_SFX();
+   else
+      S9xDoHBlankProcessing_NoSFX();
+
+#ifdef LAGFIX
+   if(finishedFrame)
+      return 3;
+#endif
+   if (CPU.Flags & SCAN_KEYS_FLAG)
+      return 1;
+
+   return 0;
+}
+
+int HandleHBlankSFX(void)
+{
+   return HandleHBlank(true);
+}
+
+int HandleHBlankNoSFX(void)
+{
+   return HandleHBlank(false);
+}
+
+uint8_t *PatchJumpDirectChecks(uint8_t *PC, uint32_t *source)
+{
+   uint8_t *dest = *FetchBlock(PC - CPU.PCBase);
+   uint8_t *checks = CachePtr;
+   EmitChecks(0);
+   GenARM_B(CondAL, PCOFFSET(dest));
+   __clear_cache(checks, CachePtr);
+   PatchJump(source, checks);
+   __clear_cache(source, source+1);
+   return checks;
+}
+
+uint8_t *PatchJumpDirect(uint8_t *PC, uint32_t *source)
+{
+   uint8_t *dest = *FetchBlock(PC - CPU.PCBase);
+   PatchJump(source, dest);
+   __clear_cache(source, source+1);
+   return dest;
+}
+
+void DynaCPUShutdown(void)
+{
+   /* Don't skip cycles with a pending NMI or IRQ - could cause delayed
+    * interrupt. Interrupts are delayed for a few cycles already, but
+    * the delay could allow the shutdown code to cycle skip again.
+    * Was causing screen flashing on Top Gear 3000. */
+   if (CPU.WaitCounter == 0 && !(CPU.Flags & (IRQ_PENDING_FLAG | NMI_FLAG)))
+   {
+      CPU.WaitAddress = NULL;
+#ifndef USE_BLARGG_APU
+      CPU.Cycles = CPU.NextEvent;
+      if (IAPU.APUExecuting)
+      {
+         ICPU.CPUExecuting = false;
+         do
+         {
+            APU_EXECUTE1();
+         } while (APU.Cycles < CPU.NextEvent);
+         ICPU.CPUExecuting = true;
+      }
+#endif
+   }
+   else if (CPU.WaitCounter >= 2)
+      CPU.WaitCounter = 1;
+   else
+      CPU.WaitCounter--;
+}
+
+void DynaWAIShutdown(void)
+{
+   CPU.Cycles = CPU.NextEvent;
+#ifndef USE_BLARGG_APU
+   if (IAPU.APUExecuting)
+   {
+      ICPU.CPUExecuting = false;
+      do
+      {
+         APU_EXECUTE1();
+      } while (APU.Cycles < CPU.NextEvent);
+      ICPU.CPUExecuting = true;
+   }
+#endif
+}
+
+static void EmitFlushCyclesPC(int *dpc, int *dcycles)
+{
+   if (*dpc != 0)
+      GenARM_ADD_Imm(CondAL, RegCPUPC, RegCPUPC, ENCODE_IMM(*dpc, 0));
+
+   if (*dcycles != 0)
+      GenARM_ADD_Imm(CondAL, RegCycles, RegCycles, ENCODE_IMM(*dcycles, 0));
+
+   dpc = 0;
+   dcycles = 0;
+}
+
+/* a1 = value, a4 = sp, preserves a4 */
+static void EmitPush(Instruction *in, bool n8)
+{
+   enum ARMReg RegS = RegA4;    /* Preserved during memory ops */
+
+   if (n8)
+      GenARM_SUB_Imm(CondAL, RegS, RegS, ENCODE_IMM(1, 0));
+   else
+      GenARM_SUB_Imm(CondAL, RegS, RegS, ENCODE_IMM(2, 0));
+
+   if (CheckEmulation())
+   {
+      GenARM_BIC_Imm(CondAL, RegS, RegS, ENCODE_IMM(0xFF, 8));
+      GenARM_ORR_Imm(CondAL, RegS, RegS, ENCODE_IMM(1, 8));
+   }
+
+   GenARM_ADD_Imm(CondAL, RegA2, RegS, ENCODE_IMM(1, 0));
+   GenARM_BIC_Imm(CondAL, RegA2, RegA2, ENCODE_IMM(0xFF, 16));
+
+   if (n8)
+   {
+      GenARM_BL(CondAL, PCOFFSET(S9xSetByteFast));
+   }
+   else
+   {
+      GenARM_BL(CondAL, PCOFFSET(S9xSetWordFast));
+   }
+}
+
+/* a1 = value, a4 = sp, preserves a4 */
+static void EmitPop(Instruction *in, bool n8)
+{
+   enum ARMReg RegS = RegA4;    /* Preserved during memory ops */
+   GenARM_MOV_Reg(CondAL, RegA2, 0, RegCPUPC);
+
+   if (n8)
+      GenARM_ADD_Imm(CondAL, RegS, RegS, ENCODE_IMM(1, 0));
+   else
+      GenARM_ADD_Imm(CondAL, RegS, RegS, ENCODE_IMM(2, 0));
+
+   if (CheckEmulation())
+   {
+      GenARM_BIC_Imm(CondAL, RegS, RegS, ENCODE_IMM(0xFF, 8));
+      GenARM_ORR_Imm(CondAL, RegS, RegS, ENCODE_IMM(1, 8));
+   }
+
+   if (n8)
+   {
+      GenARM_MOV_Reg(CondAL, RegA1, 0, RegS);
+      GenARM_BL(CondAL, PCOFFSET(S9xGetByteFast));
+   }
+   else
+   {
+      GenARM_SUB_Imm(CondAL, RegA1, RegS, ENCODE_IMM(1, 0));
+      GenARM_BL(CondAL, PCOFFSET(S9xGetWordFast));
+   }
+}
+
+static enum ARMReg EmitLoadRegister(Instruction *in, enum SNESRegArg arg, enum ARMReg RegMEM, bool read)
+{
+   enum ARMReg reg;
+
+   switch(arg)
+   {
+   case ArgA:
+      reg = RegA;
+      break;
+   case ArgX:
+      reg = RegX;
+      break;
+   case ArgY:
+      reg = RegY;
+      break;
+   case ArgP:
+      reg = RegP;
+      break;
+   case ArgD:
+      if (read)
+         GenARM_LDRH_Imm(CondAL, RegMEM, RegCPU, ICPU_OFFSET(Registers.D));
+      reg = RegMEM;
+      break;
+   case ArgS:
+      if (read)
+         GenARM_LDRH_Imm(CondAL, RegMEM, RegCPU, ICPU_OFFSET(Registers.S));
+      reg = RegMEM;
+      break;
+   case ArgDB:
+      if (read)
+         GenARM_LDRB_Imm(CondAL, RegMEM, RegCPU, ICPU_OFFSET(Registers.DB));
+      reg = RegMEM;
+      break;
+   case ArgPB:
+      if (read)
+         GenARM_LDRB_Imm(CondAL, RegMEM, RegCPU, ICPU_OFFSET(Registers.PB));
+      reg = RegMEM;
+      break;
+   default:
+      fprintf(stderr, "Unhandled SNES Register: %d\n", arg);
+      reg = -1;
+      break;
+   }
+
+   return reg;
+}
+
+/* Preserves reg, modifies R3 */
+static void EmitStoreRegister(Instruction *in, enum SNESRegArg arg, enum ARMReg reg, bool n8)
+{
+   switch(arg)
+   {
+   case ArgA:
+      if (reg != RegA)
+      {
+         if (n8)
+            GenARM_MOVB_Reg(CondAL, RegA, reg);
+         else
+            GenARM_UXTH_Reg(CondAL, RegA, reg);
+      }
+      break;
+   case ArgX:
+      if (reg != RegX)
+      {
+         if (n8)
+            GenARM_MOVB_Reg(CondAL, RegX, reg);
+         else
+            GenARM_UXTH_Reg(CondAL, RegX, reg);
+      }
+      break;
+   case ArgY:
+      if (reg != RegY)
+      {
+         if (n8)
+            GenARM_MOVB_Reg(CondAL, RegY, reg);
+         else
+            GenARM_UXTH_Reg(CondAL, RegY, reg);
+      }
+      break;
+   case ArgP:
+   {
+      uint8_t *branch;
+      GenARM_EOR_Reg(CondAL, RegP, RegP, 0, reg);
+      GenARM_EOR_Reg(CondAL, reg, RegP, 0, reg);
+      GenARM_EOR_Reg(CondAL, RegP, RegP, 0, reg);
+
+      GenARM_TST_Imm(CondAL, RegP, ENCODE_IMM(Emulation >> 8, 8));
+      GenARM_ORR_Imm(CondNE, RegP, RegP, ENCODE_IMM(MemoryFlag | IndexFlag, 0));
+      GenARM_BL(CondAL, PCOFFSET(S9xCallUnpackStatusFast));
+
+      GenARM_TST_Imm(CondAL, RegP, ENCODE_IMM(IndexFlag, 0));
+      GenARM_BIC_Imm(CondNE, RegX, RegX, ENCODE_IMM(0xFF, 8));
+      GenARM_BIC_Imm(CondNE, RegY, RegY, ENCODE_IMM(0xFF, 8));
+
+      /* Update interpreter lookup table and CacheSection if processor flags changed */
+      GenARM_EOR_Reg(CondAL, reg, reg, 0, RegP);
+      GenARM_TST_Imm(CondAL, reg, ENCODE_IMM((Emulation | MemoryFlag | IndexFlag | Decimal) >> 2, 2));
+      branch = CachePtr;
+      GenARM_B(CondEQ, PCOFFSET(branch));
+      GenARM_BL(CondAL, PCOFFSET(S9xCallUpdateSection));
+      GenARM_BL(CondAL, PCOFFSET(S9xCallFixCycles));
+      PatchJump((uint32_t *)branch, CachePtr);
+      break;
+   }
+   case ArgD:
+      GenARM_STRH_Imm(CondAL, reg, RegCPU, ICPU_OFFSET(Registers.D));
+      break;
+   case ArgDB:
+      GenARM_MOV_Reg(CondAL, RegA3, ENCODE_SHIFT(ShiftLSL, 16), reg);
+      GenARM_STRB_Imm(CondAL, reg, RegCPU, ICPU_OFFSET(Registers.DB));
+      GenARM_STR_Imm(CondAL, RegA3, RegCPU, ICPU_OFFSET(ShiftedDB));
+      break;
+   case ArgPB:
+      GenARM_MOV_Reg(CondAL, RegA3, ENCODE_SHIFT(ShiftLSL, 16), reg);
+      GenARM_STRB_Imm(CondAL, reg, RegCPU, ICPU_OFFSET(Registers.PB));
+      GenARM_STR_Imm(CondAL, RegA3, RegCPU, ICPU_OFFSET(ShiftedPB));
+      break;
+   case ArgS:
+      if (CheckEmulation())
+      {
+         GenARM_BIC_Imm(CondAL, RegA3, reg, ENCODE_IMM(0xFF, 8));
+         GenARM_ORR_Imm(CondAL, RegA3, RegA3, ENCODE_IMM(1, 8));
+         GenARM_STRH_Imm(CondAL, RegA3, RegCPU, ICPU_OFFSET(Registers.S));
+      }
+      else
+      {
+         GenARM_STRH_Imm(CondAL, reg, RegCPU, ICPU_OFFSET(Registers.S));
+      }
+
+      break;
+   default:
+      fprintf(stderr, "Unhandled SNES Register: %d\n", arg);
+      break;
+   }
+}
+
+static uint16_t EmitLoadLiteral(Instruction *in, bool n8, int *dpc, int *dcycles)
+{
+   uint16_t literal;
+   *dcycles += CPU.MemSpeed;
+   *dpc += 1;
+   literal = in->PC[1];
+
+   if (!n8)
+   {
+      *dcycles += CPU.MemSpeed;
+      *dpc += 1;
+      literal |= in->PC[2] << 8;
+   }
+
+   return literal;
+}
+
+static int32_t EmitLoadRelative(Instruction *in, bool n8, int *dpc, int *dcycles)
+{
+   int32_t disp;
+   int32_t pc;
+   int32_t target;
+   bool overflow;
+
+   if (n8)
+   {
+      disp = (int8_t)in->PC[1];
+      pc = (int32_t)((in->PC + 2) - CPU.PCBase);
+      (*dpc)++;
+      *dcycles += CPU.MemSpeed;
+   }
+   else
+   {
+      disp = (int16_t)(in->PC[1] | (in->PC[2] << 8));
+      pc = (int32_t)((in->PC + 3) - CPU.PCBase);
+      *dpc += 2;
+      *dcycles += CPU.MemSpeedx2;
+   }
+
+   target = (pc + disp) & 0xFFFF;
+   overflow = target != (pc + disp);
+
+   if (overflow)
+      disp = (target - pc);
+
+   return disp;
+}
+
+/* Stores result in A1 for reads, A2 for writes, because that is where
+ * S9XRead / S9xWrite look for address. Modifies the other. Reads also
+ * write over A4. */
+static void EmitAddrZeroPage(bool read, Instruction *in, enum SNESRegArg offsetReg, int *dpc, int *dcycles)
+{
+   enum ARMReg rd = read ? RegA1 : RegA2;
+   enum ARMReg scratch = read ? RegA2 : RegA1;
+
+   *dpc += 1;
+   *dcycles += CPU.MemSpeed;
+   GenARM_MOV_Imm(CondAL, rd, ENCODE_IMM(in->PC[1], 0));
+
+   if (offsetReg == ArgS)
+      GenARM_LDRH_Imm(CondAL, scratch, RegCPU, ICPU_OFFSET(Registers.S));
+   else
+      GenARM_LDRH_Imm(CondAL, scratch, RegCPU, ICPU_OFFSET(Registers.D));
+
+   if (offsetReg == ArgX)
+      GenARM_ADD_Reg(CondAL, rd, rd, 0, RegX);
+   else if (offsetReg == ArgY)
+      GenARM_ADD_Reg(CondAL, rd, rd, 0, RegY);
+
+   if (offsetReg == ArgX || offsetReg == ArgY || offsetReg == ArgS)
+      *dcycles += ONE_CYCLE;
+
+   if (read)
+   {
+      /* Set OpenBus to in->PC[1] */
+      GenARM_MOV_Imm(CondAL, RegA4, ENCODE_IMM(in->PC[1], 0));
+      GenARM_STRB_Imm(CondAL, RegA4, RegCPU, CPU_OPEN_BUS_OFFSET);
+   }
+
+   GenARM_ADD_Reg(CondAL, rd, rd, 0, scratch);
+
+   if (CheckEmulation() && (offsetReg == ArgX || offsetReg == ArgY))
+   {
+      GenARM_UXTB_Reg(CondAL, rd, rd);
+      *dcycles += ONE_CYCLE;
+   }
+   else
+   {
+      GenARM_UXTH_Reg(CondAL, rd, rd);
+   }
+}
+
+/* Stores result in A1 for reads, A2 for writes, because that is where
+ * S9XRead / S9xWrite look for address. Modifies the other. Reads also
+ * write over A4. */
+static void EmitAddrAbsolute(bool read, Instruction *in, enum SNESRegArg offsetReg, int *dpc, int *dcycles)
+{
+   enum ARMReg rd = read ? RegA1 : RegA2;
+   enum ARMReg scratch = read ? RegA2 : RegA1;
+   *dpc += 2;
+   *dcycles += CPU.MemSpeedx2;
+   GenARM_MOV32_Imm(CondAL, scratch, in->PC[1] + (in->PC[2] << 8));
+   GenARM_LDR_Imm(CondAL, rd, RegCPU, ICPU_OFFSET(ShiftedDB));
+
+   if (offsetReg == ArgX)
+      GenARM_ADD_Reg(CondAL, scratch, scratch, 0, RegX);
+   else if (offsetReg == ArgY)
+      GenARM_ADD_Reg(CondAL, scratch, scratch, 0, RegY);
+
+   if (read)
+   {
+      /* Set OpenBus to in->PC[2] */
+      GenARM_MOV_Imm(CondAL, RegA4, ENCODE_IMM(in->PC[2], 0));
+      GenARM_STRB_Imm(CondAL, RegA4, RegCPU, CPU_OPEN_BUS_OFFSET);
+   }
+
+   GenARM_ADD_Reg(CondAL, rd, rd, 0, scratch);
+}
+
+/* Stores result in A1 for reads, A2 for writes, because that's where
+ * S9XRead / S9xWrite look for address. Reads also write over A4. */
+static void EmitAddrLong(bool read, Instruction *in, enum SNESRegArg offsetReg, int *dpc, int *dcycles)
+{
+   enum ARMReg rd = read ? RegA1 : RegA2;
+   *dpc += 3;
+   *dcycles += CPU.MemSpeedx2 + CPU.MemSpeed;
+
+   GenARM_MOV32_Imm(CondAL, rd, in->PC[1] + (in->PC[2] << 8) + (in->PC[3] << 16));
+
+   if (offsetReg == ArgX)
+   {
+      GenARM_ADD_Reg(CondAL, rd, rd, 0, RegX);
+      GenARM_BIC_Imm(CondAL, rd, rd, ENCODE_IMM(0xFF, 24));
+   }
+
+   if (read)
+   {
+      /* Set OpenBus to in->PC[2] */
+      GenARM_MOV_Imm(CondAL, RegA4, ENCODE_IMM(in->PC[2], 0));
+      GenARM_STRB_Imm(CondAL, RegA4, RegCPU, CPU_OPEN_BUS_OFFSET);
+   }
+}
+
+/* Stores result in A1 for reads, A2 for writes, because that's where
+ * S9XRead / S9xWrite look for address. Writes over A4. */
+static void EmitAddrIndirect(bool read, Instruction *in, enum SNESRegArg offsetReg, int *dpc, int *dcycles)
+{
+   enum ARMReg rd = read ? RegA1 : RegA2;
+   enum ARMReg scratch = read ? RegA2 : RegA1;
+   enum SNESRegArg ZPArg = ArgNULL;
+
+   if (offsetReg == ArgX)
+      ZPArg = ArgX;
+   else if (IS_VADDRMODE(IndirectS))
+      ZPArg = ArgS;
+
+   EmitAddrZeroPage(true, in, ZPArg, dpc, dcycles);
+
+   GenARM_MOV_Reg(CondAL, RegA2, 0, RegCPUPC);
+   GenARM_BL(CondAL, PCOFFSET(S9xGetWordFast));
+
+   if (rd != RegA1)
+      GenARM_MOV_Reg(CondAL, rd, 0, RegA1);
+
+   if (offsetReg == ArgY)
+   {
+      GenARM_ADD_Reg(CondAL, rd, rd, 0, RegY);
+   }
+
+   if (IS_VADDRMODE(IndirectS))
+      *dcycles += ONE_CYCLE;
+
+   GenARM_LDR_Imm(CondAL, scratch, RegCPU, ICPU_OFFSET(ShiftedDB));
+   GenARM_ADD_Reg(CondAL, rd, rd, 0, scratch);
+
+   if (read)
+   {
+      /* Set OpenBus to last byte fetched */
+      GenARM_MOV_Reg(CondAL, RegA4, ENCODE_SHIFT(ShiftLSR, 8), rd);
+      GenARM_STRB_Imm(CondAL, RegA4, RegCPU, CPU_OPEN_BUS_OFFSET);
+   }
+}
+
+/* Stores result in A1 for reads, A2 for writes, because that's where
+ * S9XRead / S9xWrite look for address. Writes over A4. */
+static void EmitAddrIndirectFar(bool read, Instruction *in, enum SNESRegArg offsetReg, int *dpc, int *dcycles)
+{
+   enum ARMReg rd = read ? RegA1 : RegA2;
+
+   EmitAddrZeroPage(true, in, ArgNULL, dpc, dcycles);
+
+   GenARM_MOV_Reg(CondAL, RegA2, 0, RegCPUPC);
+   GenARM_BL(CondAL, PCOFFSET(S9xGetWordFast));
+
+   /* GetFast preserves A4 */
+   GenARM_MOV_Reg(CondAL, RegA4, 0, RegA1);
+   GenARM_ADD_Imm(CondAL, RegA1, RegA2, ENCODE_IMM(2, 0));
+   GenARM_MOV_Reg(CondAL, RegA2, 0, RegCPUPC);
+   GenARM_BL(CondAL, PCOFFSET(S9xGetByteFast));
+
+   if (offsetReg == ArgY)
+   {
+      GenARM_ADD_Reg(CondAL, RegA4, RegA4, 0, RegY);
+   }
+
+   if (read)
+   {
+      /* Set OpenBus to last byte fetched */
+      GenARM_STRB_Imm(CondAL, RegA1, RegCPU, CPU_OPEN_BUS_OFFSET);
+   }
+
+   GenARM_ADD_Reg(CondAL, rd, RegA4, ENCODE_SHIFT(ShiftLSL, 16), RegA1);
+}
+
+static void EmitGetAddress(bool read, Instruction *in, enum SNESRegArg offsetReg, int *dpc, int *dcycles)
+{
+   switch(in->AddrMode)
+   {
+   case AddrZeroPage8:
+   case AddrZeroPage16:
+      EmitAddrZeroPage(read, in, offsetReg, dpc, dcycles);
+      break;
+   case AddrAbsolute8:
+   case AddrAbsolute16:
+      EmitAddrAbsolute(read, in, offsetReg, dpc, dcycles);
+      break;
+   case AddrLong8:
+   case AddrLong16:
+      EmitAddrLong(read, in, offsetReg, dpc, dcycles);
+      break;
+   case AddrIndirect8:
+   case AddrIndirect16:
+   case AddrIndirectX8:
+   case AddrIndirectX16:
+   case AddrIndirectY8:
+   case AddrIndirectY16:
+   case AddrIndirectS8:
+   case AddrIndirectS16:
+      EmitAddrIndirect(read, in, offsetReg, dpc, dcycles);
+      break;
+   case AddrIndirectFar8:
+   case AddrIndirectFar16:
+      EmitAddrIndirectFar(read, in, offsetReg, dpc, dcycles);
+      break;
+   default:
+      break;
+   }
+}
+
+/* Returns loaded register, leaves address in RegA2 for storing */
+static enum ARMReg EmitLoadMemRegister(Instruction *in, enum ARMReg RegMEM, bool n8, int *dpc, int *dcycles)
+{
+   EmitGetAddress(true, in, in->Arg2, dpc, dcycles);
+   GenARM_MOV_Reg(CondAL, RegA2, 0, RegCPUPC);
+
+   if (n8)
+      GenARM_BL(CondAL, PCOFFSET(S9xGetByteFast));
+   else
+      GenARM_BL(CondAL, PCOFFSET(S9xGetWordFast));
+
+   GenARM_MOV_Reg(CondAL, RegMEM, 0, RegA1);
+
+   return RegMEM;
+}
+
+/* Assumes address is in RegA2 */
+static void EmitStoreMemRegister(enum ARMReg RegMEM, bool n8)
+{
+   GenARM_MOV_Reg(CondAL, RegA1, 0, RegMEM);
+
+   if (n8)
+      GenARM_BL(CondAL, PCOFFSET(S9xSetByteFast));
+   else
+      GenARM_BL(CondAL, PCOFFSET(S9xSetWordFast));
+}
+
+/* Expects Z to be set in ARM flags, HS to be set in ARM if CarryHS is true (SBC, CMP) */
+static void EmitSetFlags(enum ARMReg reg, uint32_t flags, bool n8, bool CarryHS)
+{
+   if (flags == 0)
+      return;
+
+   GenARM_BIC_Imm(CondAL, RegP, RegP, flags);
+
+   if (flags & Zero)
+      GenARM_ORR_Imm(CondEQ, RegP, RegP, Zero);
+
+   if ((flags & Carry) && CarryHS)
+      GenARM_ORR_Imm(CondHS, RegP, RegP, Carry);
+
+   if (n8)
+   {
+      if (flags & Negative)
+      {
+         GenARM_TST_Imm(CondAL, reg, ENCODE_IMM(2, 6));
+         GenARM_ORR_Imm(CondNE, RegP, RegP, Negative);
+      }
+
+      if ((flags & Carry) && !CarryHS)
+      {
+         GenARM_TST_Imm(CondAL, reg, ENCODE_IMM(1, 8));
+         GenARM_ORR_Imm(CondNE, RegP, RegP, Carry);
+      }
+   }
+   else
+   {
+      if (flags & Negative)
+      {
+         GenARM_TST_Imm(CondAL, reg, ENCODE_IMM(2, 14));
+         GenARM_ORR_Imm(CondNE, RegP, RegP, Negative);
+      }
+
+      if ((flags & Carry) && !CarryHS)
+      {
+         GenARM_TST_Imm(CondAL, reg, ENCODE_IMM(1, 16));
+         GenARM_ORR_Imm(CondNE, RegP, RegP, Carry);
+      }
+   }
+}
+
+static void EmitLD(enum ARMReg reg, bool n8, int *dpc, int *dcycles)
+{
+   /* Update cycles / cpu for slow lookup */
+   GenARM_MOV_Reg(CondAL, RegA2, 0, RegCPUPC);
+   EmitFlushCyclesPC(dpc, dcycles);
+
+   if (n8)
+   {
+      GenARM_BL(CondAL, PCOFFSET(S9xGetByteFast));
+      GenARM_MOVBS_Reg(CondAL, reg, RegR0, RegR0);
+   }
+   else
+   {
+      GenARM_BL(CondAL, PCOFFSET(S9xGetWordFast));
+      GenARM_MOVS_Reg(CondAL, reg, 0, RegR0);
+   }
+}
+
+static void EmitST(enum SNESRegArg reg, bool n8)
+{
+   switch(reg)
+   {
+   case ArgA:
+      GenARM_MOV_Reg(CondAL, RegA1, 0, RegA);
+      break;
+   case ArgX:
+      GenARM_MOV_Reg(CondAL, RegA1, 0, RegX);
+      break;
+   case ArgY:
+      GenARM_MOV_Reg(CondAL, RegA1, 0, RegY);
+      break;
+   case ArgZ:
+      GenARM_MOV_Imm(CondAL, RegA1, ENCODE_IMM(0, 0));
+      break;
+   default:
+      break;
+   }
+
+   if (n8)
+   {
+      GenARM_BL(CondAL, PCOFFSET(S9xSetByteFast));
+   }
+   else
+   {
+      GenARM_BL(CondAL, PCOFFSET(S9xSetWordFast));
+   }
+}
+
+static void EmitBranch(Instruction *in, uint32_t flag, bool set, int *dpc, int *dcycles)
+{
+   int32_t disp = EmitLoadRelative(in, true, dpc, dcycles);
+
+   EmitFlushCyclesPC(dpc, dcycles);
+
+   /* Interpreter runs BranchCheck here. Only when APU is disabled
+    * until next reset. So cost of the load seems not worth it, unless
+    * games break. */
+
+   GenARM_TST_Imm(CondAL, RegP, ENCODE_IMM(flag, 0));
+   /* If condition doesn't match, jump to next instruction */
+   if (set)
+      GenARM_BL(CondEQ, PCOFFSET(JumpDirectChecks));
+   else
+      GenARM_BL(CondNE, PCOFFSET(JumpDirectChecks));
+
+   GenARM_ADD_Imm(CondAL, RegCycles, RegCycles, ENCODE_IMM(ONE_CYCLE, 0));
+
+   if (Settings.Shutdown)
+   {
+      GenARM_LDR_Imm(CondAL, RegA1, RegCPU, CPU_OFFSET(WaitAddress));
+   }
+
+   if (disp > 0)
+   {
+      if (disp & 0xFF)
+         GenARM_ADD_Imm(CondAL, RegCPUPC, RegCPUPC, ENCODE_IMM(disp & 0xFF, 0));
+
+      if (disp & 0xFF00)
+         GenARM_ADD_Imm(CondAL, RegCPUPC, RegCPUPC, ENCODE_IMM(disp >> 8 & 0xFF, 8));
+   }
+   else if (disp < 0)
+   {
+      disp = -disp;
+      if (disp & 0xFF)
+         GenARM_SUB_Imm(CondAL, RegCPUPC, RegCPUPC, ENCODE_IMM(disp & 0xFF, 0));
+
+      if (disp & 0xFF00)
+         GenARM_SUB_Imm(CondAL, RegCPUPC, RegCPUPC, ENCODE_IMM(disp >> 8 & 0xFF, 8));
+   }
+
+   if (Settings.Shutdown) {
+      GenARM_CMP_Reg(CondAL, RegA1, 0, RegCPUPC);
+      GenARM_BL(CondEQ, PCOFFSET(S9xCallCPUShutdown));
+   }
+
+   EmitChecks(0);
+   GenARM_BL(CondAL, PCOFFSET(JumpDirect));
+}
+
+static void EmitNativeCall(Instruction *in)
+{
+   int dpc = 1;
+   int dcycles = CPU.MemSpeed;
+   bool didBreak = false;
+   bool n8;
+
+   switch(in->Op)
+   {
+   CASE_VOP(MOV):
+   CASE_VOP(MOVSP):
+      {
+         enum ARMReg RegSRC;
+         enum ARMReg RegDST;
+         n8 = (in->Op == FnMOV8) || (in->Op == FnMOVSP8);
+
+         dcycles += ONE_CYCLE;
+         RegDST = EmitLoadRegister(in, in->Arg1, RegA2, false);
+         RegSRC = EmitLoadRegister(in, in->Arg2, RegA3, true);
+
+         EmitFlushCyclesPC(&dpc, &dcycles);
+
+         if (n8)
+         {
+            GenARM_MOVBS_Reg(CondAL, RegDST, RegA1, RegSRC);
+         }
+         else
+         {
+            GenARM_UXTHS_Reg(CondAL, RegDST, RegSRC);
+         }
+
+         EmitSetFlags(RegDST, in->OutFlags & (Zero | Negative), n8, false);
+         EmitStoreRegister(in, in->Arg1, RegDST, n8);
+         EmitChecks(0);
+         break;
+      }
+   CASE_VOP(LD):
+      {
+         enum ARMReg RegDST = EmitLoadRegister(in, in->Arg1, RegR0, true);
+         n8 = in->Op == FnLD8;
+
+         if (IS_VADDRMODE(Immediate))
+         {
+            uint16_t literal = EmitLoadLiteral(in, n8, &dpc, &dcycles);
+            EmitFlushCyclesPC(&dpc, &dcycles);
+
+            if (n8)
+            {
+               GenARM_BIC_Imm(CondAL, RegDST, RegDST, ENCODE_IMM(0xFF, 0));
+               GenARM_ORR_Imm(CondAL, RegDST, RegDST, ENCODE_IMM(literal, 0));
+
+               if (in->OutFlags & Zero)
+                  GenARM_ANDS_Imm(CondAL, RegA1, RegDST, ENCODE_IMM(literal, 0));
+            }
+            else
+            {
+               GenARM_MOV32_Imm(CondAL, RegDST, literal);
+               if (in->OutFlags & Zero)
+                  GenARM_ANDS_Reg(CondAL, RegDST, RegDST, 0, RegDST);
+            }
+         }
+         else
+         {
+            EmitGetAddress(true, in, in->Arg2, &dpc, &dcycles);
+            EmitLD(RegDST, n8, &dpc, &dcycles);
+         }
+
+         EmitSetFlags(RegDST, in->OutFlags & (Zero | Negative), n8, false);
+         EmitChecks(0);
+         break;
+      }
+   CASE_VOP(ST):
+      {
+         n8 = in->Op == FnST8;
+
+         EmitGetAddress(false, in, in->Arg2, &dpc, &dcycles);
+         EmitFlushCyclesPC(&dpc, &dcycles);
+         EmitST(in->Arg1, n8);
+         EmitChecks(0);
+         break;
+      }
+   CASE_VOP(PUSH):
+      {
+         enum ARMReg RegSRC;
+         n8 = in->Op == FnPUSH8;
+
+         if (in->Arg1 == ArgNULL)
+         {
+            RegSRC = EmitLoadMemRegister(in, RegA1, n8, &dpc, &dcycles);
+         }
+         else
+         {
+            RegSRC = EmitLoadRegister(in, in->Arg1, RegA1, true);
+            dcycles += ONE_CYCLE;
+         }
+
+         GenARM_LDRH_Imm(CondAL, RegA4, RegCPU, ICPU_OFFSET(Registers.S));
+
+         if (RegSRC != RegA1)
+            GenARM_MOV_Reg(CondAL, RegA1, 0, RegSRC);
+
+         EmitFlushCyclesPC(&dpc, &dcycles);
+         EmitPush(in, n8);
+
+         GenARM_STRH_Imm(CondAL, RegA4, RegCPU, ICPU_OFFSET(Registers.S));
+
+         EmitChecks(0);
+         break;
+      }
+   CASE_OP(PEA):
+      {
+         uint16_t literal = EmitLoadLiteral(in, false, &dpc, &dcycles);
+
+         GenARM_LDRH_Imm(CondAL, RegA4, RegCPU, ICPU_OFFSET(Registers.S));
+         GenARM_MOV32_Imm(CondAL, RegA1, literal);
+
+         EmitFlushCyclesPC(&dpc, &dcycles);
+         EmitPush(in, false);
+
+         GenARM_STRH_Imm(CondAL, RegA4, RegCPU, ICPU_OFFSET(Registers.S));
+
+         EmitChecks(0);
+         break;
+      }
+   CASE_OP(PER):
+      {
+         uint16_t literal = (in->PC + 3 - CPU.PCBase + in->PC[1] + (in->PC[2] << 8)) & 0xffff;
+         dcycles += CPU.MemSpeedx2 + ONE_CYCLE;
+         dpc += 2;
+
+         GenARM_LDRH_Imm(CondAL, RegA4, RegCPU, ICPU_OFFSET(Registers.S));
+         GenARM_MOV32_Imm(CondAL, RegA1, literal);
+
+         EmitFlushCyclesPC(&dpc, &dcycles);
+         EmitPush(in, false);
+
+         GenARM_STRH_Imm(CondAL, RegA4, RegCPU, ICPU_OFFSET(Registers.S));
+
+         EmitChecks(0);
+         break;
+      }
+   CASE_VOP(POP):
+      {
+         n8 = in->Op == FnPOP8;
+         dcycles += TWO_CYCLES;
+
+         GenARM_LDRH_Imm(CondAL, RegA4, RegCPU, ICPU_OFFSET(Registers.S));
+
+         EmitPop(in, n8);
+         EmitFlushCyclesPC(&dpc, &dcycles);
+
+         if (in->OutFlags & Zero)
+            GenARM_MOVS_Reg(CondAL, RegA1, 0, RegA1); /* Set ZF on A1 */
+
+         GenARM_STRH_Imm(CondAL, RegA4, RegCPU, ICPU_OFFSET(Registers.S));
+
+         EmitSetFlags(RegA1, in->OutFlags & (Zero | Negative), n8, false);
+         EmitStoreRegister(in, in->Arg1, RegA1, n8);
+         EmitChecks(0);
+         break;
+      }
+   CASE_OP(MVP):
+   CASE_OP(MVN):
+      {
+         uint8_t *start;
+         dpc += 2;
+         dcycles += CPU.MemSpeedx2 + TWO_CYCLES;
+
+         /* Load DD */
+         /* Set DB / ShiftedDB */
+         GenARM_MOV_Imm(CondAL, RegA3, ENCODE_IMM(in->PC[1], 0));
+         GenARM_STRB_Imm(CondAL, RegA3, RegCPU, ICPU_OFFSET(Registers.DB));
+         GenARM_MOV_Reg(CondAL, RegA3, ENCODE_SHIFT(ShiftLSL, 16), RegA3);
+         GenARM_STR_Imm(CondAL, RegA3, RegCPU, ICPU_OFFSET(ShiftedDB));
+
+         /* A3 = DD << 16 */
+
+         /* Load SS */
+         /* Set OpenBus to SS */
+         GenARM_MOV_Imm(CondAL, RegA4, ENCODE_IMM(in->PC[2], 0));
+         GenARM_STRB_Imm(CondAL, RegA4, RegCPU, CPU_OPEN_BUS_OFFSET);
+
+         /* A4 = SS << 16 */
+         start = CachePtr;
+
+         GenARM_MOV_Imm(CondAL, RegA1, ENCODE_IMM(in->PC[2], 16));
+         GenARM_ORR_Reg(CondAL, RegA1, RegA1, 0, RegX);
+
+         /* Update cycles / cpu for slow lookup */
+         GenARM_MOV_Reg(CondAL, RegA2, 0, RegCPUPC);
+         EmitFlushCyclesPC(&dpc, &dcycles);
+
+         GenARM_BL(CondAL, PCOFFSET(S9xGetByteFast));
+
+         /* SetByte -- A1 val, A2 addr */
+         GenARM_MOV_Imm(CondAL, RegA2, ENCODE_IMM(in->PC[1], 16));
+         GenARM_ORR_Reg(CondAL, RegA2, RegA2, 0, RegY);
+         GenARM_BL(CondAL, PCOFFSET(S9xSetByteFast));
+
+         if (in->Op == FnMVN)
+         {
+            if (CheckIndex())
+            {
+               GenARM_ADD_Imm(CondAL, RegA1, RegX, ENCODE_IMM(1, 0));
+               GenARM_MOVB_Reg(CondAL, RegX, RegA1);
+               GenARM_ADD_Imm(CondAL, RegA1, RegY, ENCODE_IMM(1, 0));
+               GenARM_MOVB_Reg(CondAL, RegY, RegA1);
+            }
+            else
+            {
+               GenARM_ADD_Imm(CondAL, RegA1, RegX, ENCODE_IMM(1, 0));
+               GenARM_UXTH_Reg(CondAL, RegX, RegA1);
+               GenARM_ADD_Imm(CondAL, RegA1, RegY, ENCODE_IMM(1, 0));
+               GenARM_UXTH_Reg(CondAL, RegY, RegA1);
+            }
+         }
+         else
+         {
+            if (CheckIndex())
+            {
+               GenARM_SUB_Imm(CondAL, RegA1, RegX, ENCODE_IMM(1, 0));
+               GenARM_MOVB_Reg(CondAL, RegX, RegA1);
+               GenARM_SUB_Imm(CondAL, RegA1, RegY, ENCODE_IMM(1, 0));
+               GenARM_MOVB_Reg(CondAL, RegY, RegA1);
+            }
+            else
+            {
+               GenARM_SUB_Imm(CondAL, RegA1, RegX, ENCODE_IMM(1, 0));
+               GenARM_UXTH_Reg(CondAL, RegX, RegA1);
+               GenARM_SUB_Imm(CondAL, RegA1, RegY, ENCODE_IMM(1, 0));
+               GenARM_UXTH_Reg(CondAL, RegY, RegA1);
+            }
+         }
+
+         GenARM_CMP_Imm(CondAL, RegA, ENCODE_IMM(0, 0));
+         GenARM_SUB_Imm(CondAL, RegA1, RegA, ENCODE_IMM(1, 0));
+         GenARM_UXTH_Reg(CondAL, RegA, RegA1);
+         GenARM_BL(CondEQ, PCOFFSET(JumpDirectChecks));
+
+         GenARM_SUB_Imm(CondAL, RegCPUPC, RegCPUPC, ENCODE_IMM(dpc, 0));
+         EmitChecks(0);
+         GenARM_B(CondAL, PCOFFSET(start));
+         didBreak = true;
+         break;
+      }
+   CASE_VOP(OR):
+      {
+         n8 = in->Op == FnOR8;
+         if (IS_VADDRMODE(Immediate))
+         {
+            uint16_t literal = EmitLoadLiteral(in, n8, &dpc, &dcycles);
+            EmitFlushCyclesPC(&dpc, &dcycles);
+
+            if (n8)
+            {
+               GenARM_ORR_Imm(CondAL, RegA, RegA, ENCODE_IMM(literal, 0));
+               GenARM_ANDS_Imm(CondAL, RegA2, RegA, ENCODE_IMM(0xFF, 0));
+               EmitSetFlags(RegA2, in->OutFlags & (Zero | Negative), n8, false);
+            }
+            else
+            {
+               GenARM_MOV32_Imm(CondAL, RegR0, literal);
+               GenARM_ORRS_Reg(CondAL, RegA, RegA, 0, RegR0);
+               EmitSetFlags(RegA, in->OutFlags & (Zero | Negative), n8, false);
+            }
+         }
+         else
+         {
+            EmitGetAddress(true, in, in->Arg2, &dpc, &dcycles);
+
+            /* Update cycles / cpu for slow lookup */
+            GenARM_MOV_Reg(CondAL, RegA2, 0, RegCPUPC);
+            EmitFlushCyclesPC(&dpc, &dcycles);
+
+            if (n8)
+            {
+               GenARM_BL(CondAL, PCOFFSET(S9xGetByteFast));
+               GenARM_ORR_Reg(CondAL, RegA, RegA, 0, RegR0);
+               GenARM_ANDS_Imm(CondAL, RegA2, RegA, ENCODE_IMM(0xFF, 0));
+               EmitSetFlags(RegA2, in->OutFlags & (Zero | Negative), n8, false);
+            }
+            else
+            {
+               GenARM_BL(CondAL, PCOFFSET(S9xGetWordFast));
+               GenARM_ORRS_Reg(CondAL, RegA, RegA, 0, RegR0);
+               EmitSetFlags(RegA, in->OutFlags & (Zero | Negative), n8, false);
+            }
+         }
+
+         EmitChecks(0);
+         break;
+      }
+   CASE_VOP(AND):
+      {
+         n8 = in->Op == FnAND8;
+         if (IS_VADDRMODE(Immediate))
+         {
+            uint16_t literal = EmitLoadLiteral(in, n8, &dpc, &dcycles);
+            EmitFlushCyclesPC(&dpc, &dcycles);
+
+            if (n8)
+            {
+               GenARM_ANDS_Imm(CondAL, RegA1, RegA, ENCODE_IMM(literal, 0));
+               GenARM_AND_Imm(CondAL, RegA, RegA, ENCODE_IMM(0xFF, 8));
+               GenARM_ORR_Reg(CondAL, RegA, RegA, 0, RegA1);
+            }
+            else
+            {
+               GenARM_MOV32_Imm(CondAL, RegA1, literal);
+               GenARM_ANDS_Reg(CondAL, RegA, RegA, 0, RegA1);
+            }
+         }
+         else
+         {
+            EmitGetAddress(true, in, in->Arg2, &dpc, &dcycles);
+
+            /* Update cycles / cpu for slow lookup */
+            GenARM_MOV_Reg(CondAL, RegA2, 0, RegCPUPC);
+            EmitFlushCyclesPC(&dpc, &dcycles);
+
+            if (n8)
+            {
+               GenARM_BL(CondAL, PCOFFSET(S9xGetByteFast));
+               GenARM_ANDS_Reg(CondAL, RegA1, RegA, 0, RegA1);
+               GenARM_AND_Imm(CondAL, RegA, RegA, ENCODE_IMM(0xFF, 8));
+               GenARM_ORR_Reg(CondAL, RegA, RegA, 0, RegA1);
+            }
+            else
+            {
+               GenARM_BL(CondAL, PCOFFSET(S9xGetWordFast));
+               GenARM_ANDS_Reg(CondAL, RegA, RegA, 0, RegA1);
+            }
+         }
+
+         EmitSetFlags(RegA, in->OutFlags & (Zero | Negative), n8, false);
+         EmitChecks(0);
+         break;
+      }
+   CASE_VOP(EOR):
+      {
+         n8 = in->Op == FnEOR8;
+         if (IS_VADDRMODE(Immediate))
+         {
+            uint16_t literal = EmitLoadLiteral(in, n8, &dpc, &dcycles);
+            EmitFlushCyclesPC(&dpc, &dcycles);
+
+            if (n8)
+            {
+               GenARM_AND_Imm(CondAL, RegA1, RegA, ENCODE_IMM(0xFF, 0));
+               GenARM_EORS_Imm(CondAL, RegA1, RegA1, ENCODE_IMM(literal, 0));
+               GenARM_AND_Imm(CondAL, RegA, RegA, ENCODE_IMM(0xFF, 8));
+               GenARM_ORR_Reg(CondAL, RegA, RegA, 0, RegA1);
+            }
+            else
+            {
+               GenARM_MOV32_Imm(CondAL, RegA1, literal);
+               GenARM_EORS_Reg(CondAL, RegA, RegA, 0, RegA1);
+            }
+         }
+         else
+         {
+            EmitGetAddress(true, in, in->Arg2, &dpc, &dcycles);
+
+            /* Update cycles / cpu for slow lookup */
+            GenARM_MOV_Reg(CondAL, RegA2, 0, RegCPUPC);
+            EmitFlushCyclesPC(&dpc, &dcycles);
+
+            if (n8)
+            {
+               GenARM_BL(CondAL, PCOFFSET(S9xGetByteFast));
+               GenARM_AND_Imm(CondAL, RegA2, RegA, ENCODE_IMM(0xFF, 0));
+               GenARM_EORS_Reg(CondAL, RegA1, RegA1, 0, RegA2);
+               GenARM_AND_Imm(CondAL, RegA, RegA, ENCODE_IMM(0xFF, 8));
+               GenARM_ORR_Reg(CondAL, RegA, RegA, 0, RegA1);
+            }
+            else
+            {
+               GenARM_BL(CondAL, PCOFFSET(S9xGetWordFast));
+               GenARM_EORS_Reg(CondAL, RegA, RegA, 0, RegA1);
+            }
+         }
+
+         EmitSetFlags(RegA, in->OutFlags & (Zero | Negative), n8, false);
+         EmitChecks(0);
+         break;
+      }
+   CASE_VOP(ADC):
+      {
+         enum ARMReg RegDST = EmitLoadRegister(in, in->Arg1, RegR0, true);
+         n8 = in->Op == FnADC8;
+
+         if (IS_VADDRMODE(Immediate))
+         {
+            uint16_t literal = EmitLoadLiteral(in, n8, &dpc, &dcycles);
+            EmitFlushCyclesPC(&dpc, &dcycles);
+
+            GenARM_MOV32_Imm(CondAL, RegA1, literal);
+         }
+         else
+         {
+            EmitGetAddress(true, in, in->Arg2, &dpc, &dcycles);
+            /* Update cycles / cpu for slow lookup */
+            GenARM_MOV_Reg(CondAL, RegA2, 0, RegCPUPC);
+            EmitFlushCyclesPC(&dpc, &dcycles);
+
+            if (n8)
+               GenARM_BL(CondAL, PCOFFSET(S9xGetByteFast));
+            else
+               GenARM_BL(CondAL, PCOFFSET(S9xGetWordFast));
+         }
+
+         GenARM_AND_Imm(CondAL, RegA3, RegP, ENCODE_IMM(Carry, 0));
+
+         if (CheckDecimal())
+         {
+            uint32_t shift;
+
+            for (shift = 0; shift < (n8 ? 8 : 16) ; shift += 4)
+            {
+               GenARM_AND_Imm(CondAL, RegA4, RegA1, ENCODE_IMM(0xF, shift));
+               GenARM_ADD_Reg(CondAL, RegA3, RegA3, 0, RegA4);
+               GenARM_AND_Imm(CondAL, RegA4, RegDST, ENCODE_IMM(0xF, shift));
+               GenARM_ADD_Reg(CondAL, RegA3, RegA3, 0, RegA4);
+               GenARM_CMP_Imm(CondAL, RegA3, ENCODE_IMM(0xA, shift));
+               GenARM_ADD_Imm(CondHS, RegA3, RegA3, ENCODE_IMM(0x6, shift));
+            }
+
+            if (n8)
+            {
+               GenARM_AND_Imm(CondAL, RegA2, RegDST, ENCODE_IMM(0xFF, 0));
+               GenARM_MOVBS_Reg(CondAL, RegDST, RegA4, RegA3);
+            }
+            else
+            {
+               GenARM_MOV_Reg(CondAL, RegA2, 0, RegDST);
+               GenARM_UXTHS_Reg(CondAL, RegDST, RegA3);
+            }
+         }
+         else
+         {
+            GenARM_ADD_Reg(CondAL, RegA3, RegA3, 0, RegA1);
+
+            if (n8)
+            {
+               GenARM_AND_Imm(CondAL, RegA2, RegDST, ENCODE_IMM(0xFF, 0));
+               GenARM_ADD_Reg(CondAL, RegA3, RegA3, 0, RegA2);
+               GenARM_MOVBS_Reg(CondAL, RegDST, RegA4, RegA3);
+            }
+            else
+            {
+               GenARM_MOV_Reg(CondAL, RegA2, 0, RegDST);
+               GenARM_ADD_Reg(CondAL, RegA3, RegA3, 0, RegDST);
+               GenARM_UXTHS_Reg(CondAL, RegDST, RegA3);
+            }
+         }
+
+         EmitSetFlags(RegA3, in->OutFlags & (Zero | Negative | Carry | Overflow), n8, false);
+
+         if (in->OutFlags & Overflow)
+         {
+            /* Calculate overflow flag */
+            if (CheckDecimal()) {
+               GenARM_CMP_Imm(CondAL, RegA3, ENCODE_IMM(1, n8 ? 8 : 16));
+               GenARM_SUB_Imm(CondGE, RegA3, RegA3, ENCODE_IMM(0x60, n8 ? 0 : 8));
+            }
+
+            GenARM_EOR_Reg(CondAL, RegA1, RegA2, 0, RegA1);
+            GenARM_MVN_Reg(CondAL, RegA1, 0, RegA1);
+            GenARM_EOR_Reg(CondAL, RegA2, RegA3, 0, RegA2);
+            GenARM_AND_Reg(CondAL, RegA1, RegA1, 0, RegA2);
+            if (n8)
+            {
+               GenARM_TST_Imm(CondAL, RegA1, ENCODE_IMM(0x80, 0));
+            }
+            else
+            {
+               GenARM_TST_Imm(CondAL, RegA1, ENCODE_IMM(0x80, 8));
+            }
+            GenARM_ORR_Imm(CondNE, RegP, RegP, ENCODE_IMM(Overflow, 0));
+         }
+
+         EmitChecks(0);
+         break;
+      }
+   CASE_VOP(SBC):
+      {
+         enum ARMReg RegDST = EmitLoadRegister(in, in->Arg1, RegR0, true);
+         n8 = in->Op == FnSBC8;
+         if (IS_VADDRMODE(Immediate))
+         {
+            uint16_t literal = EmitLoadLiteral(in, n8, &dpc, &dcycles);
+            EmitFlushCyclesPC(&dpc, &dcycles);
+
+            GenARM_MOV32_Imm(CondAL, RegA1, literal);
+         }
+         else
+         {
+            EmitGetAddress(true, in, in->Arg2, &dpc, &dcycles);
+            /* Update cycles / cpu for slow lookup */
+            GenARM_MOV_Reg(CondAL, RegA2, 0, RegCPUPC);
+            EmitFlushCyclesPC(&dpc, &dcycles);
+
+            if (n8)
+               GenARM_BL(CondAL, PCOFFSET(S9xGetByteFast));
+            else
+               GenARM_BL(CondAL, PCOFFSET(S9xGetWordFast));
+         }
+
+         GenARM_MVN_Reg(CondAL, RegA1, 0, RegA1);
+         if (n8)
+            GenARM_UXTB_Reg(CondAL, RegA1, RegA1);
+         else
+            GenARM_UXTH_Reg(CondAL, RegA1, RegA1);
+
+         GenARM_AND_Imm(CondAL, RegA3, RegP, ENCODE_IMM(Carry, 0));
+
+         if (CheckDecimal())
+         {
+            uint32_t shift;
+
+            for (shift = 0; shift < (n8 ? 8 : 16) ; shift += 4)
+            {
+               GenARM_AND_Imm(CondAL, RegA4, RegA1, ENCODE_IMM(0xF, shift));
+               GenARM_ADD_Reg(CondAL, RegA3, RegA3, 0, RegA4);
+               GenARM_AND_Imm(CondAL, RegA4, RegDST, ENCODE_IMM(0xF, shift));
+               GenARM_ADD_Reg(CondAL, RegA3, RegA3, 0, RegA4);
+               GenARM_CMP_Imm(CondAL, RegA3, ENCODE_IMM(0x10, shift));
+               GenARM_SUB_Imm(CondLO, RegA3, RegA3, ENCODE_IMM(0x6, shift));
+            }
+
+            if (n8)
+            {
+               GenARM_AND_Imm(CondAL, RegA2, RegDST, ENCODE_IMM(0xFF, 0));
+               GenARM_MOVBS_Reg(CondAL, RegDST, RegA4, RegA3);
+            }
+            else
+            {
+               GenARM_MOV_Reg(CondAL, RegA2, 0, RegDST);
+               GenARM_UXTHS_Reg(CondAL, RegDST, RegA3);
+            }
+         }
+         else
+         {
+            GenARM_ADD_Reg(CondAL, RegA3, RegA3, 0, RegA1);
+
+            if (n8)
+            {
+               GenARM_AND_Imm(CondAL, RegA2, RegDST, ENCODE_IMM(0xFF, 0));
+               GenARM_ADD_Reg(CondAL, RegA3, RegA3, 0, RegA2);
+               GenARM_MOVBS_Reg(CondAL, RegDST, RegA4, RegA3);
+            }
+            else
+            {
+               GenARM_MOV_Reg(CondAL, RegA2, 0, RegDST);
+               GenARM_ADD_Reg(CondAL, RegA3, RegA3, 0, RegDST);
+               GenARM_UXTHS_Reg(CondAL, RegDST, RegA3);
+            }
+         }
+
+         EmitSetFlags(RegA3, in->OutFlags & (Zero | Negative | Overflow), n8, false);
+
+         if (in->OutFlags & Carry)
+         {
+            GenARM_BIC_Imm(CondAL, RegP, RegP, ENCODE_IMM(Carry, 0));
+            if (n8)
+               GenARM_CMP_Imm(CondAL, RegA3, ENCODE_IMM(1, 8));
+            else
+               GenARM_CMP_Imm(CondAL, RegA3, ENCODE_IMM(1, 16));
+            GenARM_ORR_Imm(CondGE, RegP, RegP, ENCODE_IMM(Carry, 0));
+         }
+
+         if (in->OutFlags & Overflow)
+         {
+            /* Calculate overflow flag */
+            GenARM_EOR_Reg(CondAL, RegA1, RegA2, 0, RegA1);
+            GenARM_MVN_Reg(CondAL, RegA1, 0, RegA1);
+            GenARM_EOR_Reg(CondAL, RegA2, RegA3, 0, RegA2);
+            GenARM_AND_Reg(CondAL, RegA1, RegA1, 0, RegA2);
+            if (n8)
+            {
+               GenARM_TST_Imm(CondAL, RegA1, ENCODE_IMM(0x80, 0));
+            }
+            else
+            {
+               GenARM_TST_Imm(CondAL, RegA1, ENCODE_IMM(0x80, 8));
+            }
+            GenARM_ORR_Imm(CondNE, RegP, RegP, ENCODE_IMM(Overflow, 0));
+         }
+
+         EmitChecks(0);
+         break;
+      }
+   CASE_VOP(CMP):
+      {
+         enum ARMReg RegW = EmitLoadRegister(in, in->Arg1, RegR0, true);
+         n8 = in->Op == FnCMP8;
+
+         if (IS_VADDRMODE(Immediate))
+         {
+            uint16_t literal = EmitLoadLiteral(in, n8, &dpc, &dcycles);
+            EmitFlushCyclesPC(&dpc, &dcycles);
+
+            if (n8)
+            {
+               GenARM_AND_Imm(CondAL, RegR1, RegW, ENCODE_IMM(0xFF, 0));
+               GenARM_SUBS_Imm(CondAL, RegR0, RegR1, ENCODE_IMM(literal, 0));
+            }
+            else
+            {
+               GenARM_MOV32_Imm(CondAL, RegR0, literal);
+               GenARM_SUBS_Reg(CondAL, RegR0, RegW, 0, RegR0);
+            }
+         }
+         else
+         {
+            EmitGetAddress(true, in, in->Arg2, &dpc, &dcycles);
+
+            /* Update cycles / cpu for slow lookup */
+            GenARM_MOV_Reg(CondAL, RegA2, 0, RegCPUPC);
+            EmitFlushCyclesPC(&dpc, &dcycles);
+
+            if (n8)
+            {
+               GenARM_BL(CondAL, PCOFFSET(S9xGetByteFast));
+               GenARM_AND_Imm(CondAL, RegR1, RegW, ENCODE_IMM(0xFF, 0));
+               GenARM_SUBS_Reg(CondAL, RegR0, RegR1, 0, RegR0);
+            }
+            else
+            {
+               GenARM_BL(CondAL, PCOFFSET(S9xGetWordFast));
+               GenARM_SUBS_Reg(CondAL, RegR0, RegW, 0, RegR0);
+            }
+         }
+
+         EmitSetFlags(RegR0, in->OutFlags & (Carry | Zero | Negative), n8, true);
+         EmitChecks(0);
+         break;
+      }
+   CASE_VOP(BIT):
+      {
+         n8 = in->Op == FnBIT8;
+
+         if (IS_VADDRMODE(Immediate))
+         {
+            uint16_t literal = EmitLoadLiteral(in, n8, &dpc, &dcycles);
+            EmitFlushCyclesPC(&dpc, &dcycles);
+
+            if (n8)
+            {
+               GenARM_TST_Imm(CondAL, RegA, ENCODE_IMM(literal, 0));
+            }
+            else
+            {
+               GenARM_MOV32_Imm(CondAL, RegA1, literal);
+               GenARM_TST_Reg(CondAL, RegA, 0, RegA1);
+            }
+
+            EmitSetFlags(RegA, in->OutFlags & (Zero), n8, false);
+         }
+         else
+         {
+            uint32_t flags;
+            EmitGetAddress(true, in, in->Arg2, &dpc, &dcycles);
+
+            /* Update cycles / cpu for slow lookup */
+            GenARM_MOV_Reg(CondAL, RegA2, 0, RegCPUPC);
+            EmitFlushCyclesPC(&dpc, &dcycles);
+
+            if (n8)
+            {
+               GenARM_BL(CondAL, PCOFFSET(S9xGetByteFast));
+               GenARM_TST_Reg(CondAL, RegA, 0, RegA1);
+            }
+            else
+            {
+               GenARM_BL(CondAL, PCOFFSET(S9xGetWordFast));
+               GenARM_TST_Reg(CondAL, RegA, 0, RegA1);
+            }
+
+            flags = in->OutFlags & (Zero | Negative | Overflow);
+
+            if (flags == 0)
+               return;
+
+            GenARM_BIC_Imm(CondAL, RegP, RegP, flags);
+
+            if (flags & Zero)
+               GenARM_ORR_Imm(CondEQ, RegP, RegP, Zero);
+
+            if (n8)
+            {
+               if (flags & Negative)
+               {
+                  GenARM_TST_Imm(CondAL, RegA1, ENCODE_IMM(0x80, 0));
+                  GenARM_ORR_Imm(CondNE, RegP, RegP, Negative);
+               }
+
+               if ((flags & Overflow))
+               {
+                  GenARM_TST_Imm(CondAL, RegA1, ENCODE_IMM(0x40, 0));
+                  GenARM_ORR_Imm(CondNE, RegP, RegP, Overflow);
+               }
+            }
+            else
+            {
+               if (flags & Negative)
+               {
+                  GenARM_TST_Imm(CondAL, RegA1, ENCODE_IMM(0x80, 8));
+                  GenARM_ORR_Imm(CondNE, RegP, RegP, Negative);
+               }
+
+               if ((flags & Overflow))
+               {
+                  GenARM_TST_Imm(CondAL, RegA1, ENCODE_IMM(0x40, 8));
+                  GenARM_ORR_Imm(CondNE, RegP, RegP, Overflow);
+               }
+            }
+         }
+
+         EmitChecks(0);
+         break;
+      }
+   CASE_VOP(INC):
+   CASE_VOP(DEC):
+      {
+         enum ARMReg RegW;
+         n8 = in->Op == FnINC8 || in->Op == FnDEC8;
+
+         if (in->Arg1 == ArgNULL)
+            RegW = EmitLoadMemRegister(in, RegA4, n8, &dpc, &dcycles);
+         else
+            RegW = EmitLoadRegister(in, in->Arg1, RegA2, true);
+
+         /* Add cycles */
+         dcycles += ONE_CYCLE;
+
+         /* Clear WaitAddress */
+         GenARM_MOV_Imm(CondAL, RegA1, ENCODE_SHIFT(0, 0));
+         GenARM_STR_Imm(CondAL, RegA1, RegCPU, CPU_OFFSET(WaitAddress));
+
+         EmitFlushCyclesPC(&dpc, &dcycles);
+
+         if (IS_VOP(INC))
+            GenARM_ADD_Imm(CondAL, RegA1, RegW, ENCODE_IMM(1, 0));
+         else
+            GenARM_SUB_Imm(CondAL, RegA1, RegW, ENCODE_IMM(1, 0));
+
+         if (n8)
+         {
+            GenARM_MOVBS_Reg(CondAL, RegW, RegA1, RegA1);
+         }
+         else
+         {
+            GenARM_UXTHS_Reg(CondAL, RegW, RegA1);
+         }
+
+         EmitSetFlags(RegA1, in->OutFlags & (Zero | Negative), n8, false);
+
+         if (in->Arg1 == ArgNULL)
+            EmitStoreMemRegister(RegW, n8);
+
+         EmitChecks(0);
+         break;
+      }
+   CASE_VOP(TSB):
+   CASE_VOP(TRB):
+      {
+         n8 = in->Op == FnTSB8 || in->Op == FnTRB8;
+         EmitGetAddress(true, in, in->Arg2, &dpc, &dcycles);
+
+         dcycles += ONE_CYCLE;
+
+         /* Update cycles / cpu for slow lookup */
+         GenARM_MOV_Reg(CondAL, RegA2, 0, RegCPUPC);
+         EmitFlushCyclesPC(&dpc, &dcycles);
+
+         if (n8)
+         {
+            GenARM_BL(CondAL, PCOFFSET(S9xGetByteFast));
+            GenARM_AND_Imm(CondAL, RegA4, RegA, ENCODE_IMM(0xFF, 0));
+
+            GenARM_TST_Reg(CondAL, RegA4, 0, RegA1);
+            EmitSetFlags(RegA4, in->OutFlags & (Zero), n8, false);
+
+            if (IS_VOP(TSB))
+            {
+               GenARM_ORR_Reg(CondAL, RegA1, RegA4, 0, RegA1);
+            }
+            else
+            {
+               GenARM_MVN_Reg(CondAL, RegA4, 0, RegA4);
+               GenARM_AND_Reg(CondAL, RegA1, RegA4, 0, RegA1);
+            }
+            GenARM_BL(CondAL, PCOFFSET(S9xSetByteFast));
+         }
+         else
+         {
+            GenARM_BL(CondAL, PCOFFSET(S9xGetWordFast));
+            GenARM_TST_Reg(CondAL, RegA, 0, RegA1);
+            EmitSetFlags(RegA, in->OutFlags & (Zero), n8, false);
+
+            if (IS_VOP(TSB))
+            {
+               GenARM_ORR_Reg(CondAL, RegA1, RegA, 0, RegA1);
+            }
+            else
+            {
+               GenARM_MVN_Reg(CondAL, RegA4, 0, RegA);
+               GenARM_AND_Reg(CondAL, RegA1, RegA4, 0, RegA1);
+            }
+            GenARM_BL(CondAL, PCOFFSET(S9xSetWordFast));
+         }
+
+         EmitChecks(0);
+         break;
+      }
+   CASE_VOP(ASL):
+      {
+         enum ARMReg RegW;
+         n8 = in->Op == FnASL8;
+
+         if (in->Arg1 == ArgNULL)
+            RegW = EmitLoadMemRegister(in, RegA4, n8, &dpc, &dcycles);
+         else
+            RegW = EmitLoadRegister(in, in->Arg1, RegA4, true);
+
+         /* Add cycles */
+         dcycles += ONE_CYCLE;
+         EmitFlushCyclesPC(&dpc, &dcycles);
+
+         GenARM_MOV_Reg(CondAL, RegA1, ENCODE_SHIFT(ShiftLSL, 1), RegW);
+
+         if (n8)
+         {
+            GenARM_MOVBS_Reg(CondAL, RegW, RegA4, RegA1);
+         }
+         else
+         {
+            GenARM_UXTHS_Reg(CondAL, RegW, RegA1);
+         }
+
+         EmitSetFlags(RegA1, in->OutFlags & (Carry | Zero | Negative), CheckMemory(), false);
+
+         if (in->Arg1 == ArgNULL)
+            EmitStoreMemRegister(RegW, n8);
+
+         EmitChecks(0);
+         break;
+      }
+   CASE_VOP(LSR):
+      {
+         enum ARMReg RegW;
+         n8 = in->Op == FnLSR8;
+
+         if (in->Arg1 == ArgNULL)
+            RegW = EmitLoadMemRegister(in, RegA3, n8, &dpc, &dcycles);
+         else
+            RegW = EmitLoadRegister(in, in->Arg1, RegA3, true);
+
+         /* Add cycles */
+         dcycles += ONE_CYCLE;
+         EmitFlushCyclesPC(&dpc, &dcycles);
+
+         if (n8)
+         {
+            GenARM_AND_Imm(CondAL, RegA1, RegW, ENCODE_IMM(0xFF, 0));
+            GenARM_MOVS_Reg(CondAL, RegA1, ENCODE_SHIFT(ShiftLSR, 1), RegA1);
+            GenARM_BIC_Imm(CondAL, RegW, RegW, ENCODE_IMM(0xFF, 0));
+            GenARM_ORR_Reg(CondAL, RegW, RegW, 0, RegA1);
+         }
+         else
+         {
+            GenARM_MOVS_Reg(CondAL, RegW, ENCODE_SHIFT(ShiftLSR, 1), RegW);
+         }
+
+         /* Flags easier to set inline here */
+         GenARM_BIC_Imm(CondAL, RegP, RegP, in->OutFlags & (Carry | Zero | Negative));
+
+         /* Negative is always unset */
+         if (in->OutFlags & Zero)
+            GenARM_ORR_Imm(CondEQ, RegP, RegP, Zero);
+
+         if (in->OutFlags & Carry)
+            GenARM_ORR_Imm(CondCS, RegP, RegP, Carry);
+
+         if (in->Arg1 == ArgNULL)
+            EmitStoreMemRegister(RegW, n8);
+
+         EmitChecks(0);
+         break;
+      }
+   CASE_VOP(ROL):
+      {
+         enum ARMReg RegW;
+         n8 = in->Op == FnROL8;
+
+         if (in->Arg1 == ArgNULL)
+            RegW = EmitLoadMemRegister(in, RegA4, n8, &dpc, &dcycles);
+         else
+            RegW = EmitLoadRegister(in, in->Arg1, RegA4, true);
+
+         /* Add cycles */
+         dcycles += ONE_CYCLE;
+         EmitFlushCyclesPC(&dpc, &dcycles);
+
+         GenARM_MOV_Reg(CondAL, RegA1, ENCODE_SHIFT(ShiftLSL, 1), RegW);
+         GenARM_TST_Imm(CondAL, RegP, ENCODE_IMM(Carry, 0));
+         GenARM_ORR_Imm(CondNE, RegA1, RegA1, ENCODE_IMM(1, 0));
+
+         if (in->OutFlags & Carry)
+         {
+            GenARM_BIC_Imm(CondAL, RegP, RegP, ENCODE_IMM(Carry, 0));
+            GenARM_TST_Imm(CondAL, RegA1, ENCODE_IMM(1, n8 ? 8 : 16));
+            GenARM_ORR_Imm(CondNE, RegP, RegP, ENCODE_IMM(Carry, 0));
+         }
+
+         if (n8)
+         {
+            GenARM_MOVBS_Reg(CondAL, RegW, RegA4, RegA1);
+         }
+         else
+         {
+            GenARM_UXTHS_Reg(CondAL, RegW, RegA1);
+         }
+
+         EmitSetFlags(RegA1, in->OutFlags & (Zero | Negative), n8, false);
+
+         if (in->Arg1 == ArgNULL)
+            EmitStoreMemRegister(RegW, n8);
+
+         EmitChecks(0);
+         break;
+      }
+   CASE_VOP(ROR):
+      {
+         enum ARMReg RegW;
+         n8 = in->Op == FnROR8;
+
+         if (in->Arg1 == ArgNULL)
+            RegW = EmitLoadMemRegister(in, RegA4, n8, &dpc, &dcycles);
+         else
+            RegW = EmitLoadRegister(in, in->Arg1, RegA4, true);
+
+         /* Add cycles */
+         dcycles += ONE_CYCLE;
+         EmitFlushCyclesPC(&dpc, &dcycles);
+
+         if (n8)
+         {
+            GenARM_AND_Imm(CondAL, RegA1, RegW, ENCODE_IMM(0xFF, 0));
+            GenARM_TST_Imm(CondAL, RegP, ENCODE_IMM(Carry, 0));
+            GenARM_ORR_Imm(CondNE, RegA1, RegA1, ENCODE_IMM(1, 8));
+            GenARM_MOVS_Reg(CondAL, RegA1, ENCODE_SHIFT(ShiftLSR, 1), RegA1);
+            GenARM_MOVB_Reg(CondAL, RegW, RegA1);
+         }
+         else
+         {
+            GenARM_TST_Imm(CondAL, RegP, ENCODE_IMM(Carry, 0));
+            GenARM_ORR_Imm(CondNE, RegW, RegW, ENCODE_IMM(1, 16));
+            GenARM_MOVS_Reg(CondAL, RegW, ENCODE_SHIFT(ShiftLSR, 1), RegW);
+         }
+
+         if (in->OutFlags & Carry)
+         {
+            GenARM_BIC_Imm(CondAL, RegP, RegP, ENCODE_IMM(Carry, 0));
+            GenARM_ORR_Imm(CondCS, RegP, RegP, ENCODE_IMM(Carry, 0));
+         }
+
+         EmitSetFlags(n8 ? RegA1 : RegW, in->OutFlags & (Zero | Negative), n8, false);
+
+         if (in->Arg1 == ArgNULL)
+            EmitStoreMemRegister(RegW, n8);
+
+         EmitChecks(0);
+         break;
+      }
+   CASE_OP(BRA):
+   CASE_OP(BRL):
+      {
+         int32_t disp;
+         n8 = in->Op == FnBRA;
+         disp = EmitLoadRelative(in, n8, &dpc, &dcycles);
+         dcycles += ONE_CYCLE;
+         EmitFlushCyclesPC(&dpc, &dcycles);
+
+         if (in->Op == FnBRA && Settings.Shutdown)
+         {
+            GenARM_LDR_Imm(CondAL, RegA1, RegCPU, CPU_OFFSET(WaitAddress));
+         }
+
+         if (disp > 0)
+         {
+            if (disp & 0xFF)
+               GenARM_ADD_Imm(CondAL, RegCPUPC, RegCPUPC, ENCODE_IMM(disp & 0xFF, 0));
+
+            if (disp & 0xFF00)
+               GenARM_ADD_Imm(CondAL, RegCPUPC, RegCPUPC, ENCODE_IMM(disp >> 8 & 0xFF, 8));
+         }
+         else if (disp < 0)
+         {
+            disp = -disp;
+            if (disp & 0xFF)
+               GenARM_SUB_Imm(CondAL, RegCPUPC, RegCPUPC, ENCODE_IMM(disp & 0xFF, 0));
+
+            if (disp & 0xFF00)
+               GenARM_SUB_Imm(CondAL, RegCPUPC, RegCPUPC, ENCODE_IMM(disp >> 8 & 0xFF, 8));
+         }
+
+         if (in->Op == FnBRA && Settings.Shutdown) {
+            GenARM_CMP_Reg(CondAL, RegA1, 0, RegCPUPC);
+            GenARM_BL(CondEQ, PCOFFSET(S9xCallCPUShutdown));
+         }
+
+         EmitChecks(0);
+         GenARM_BL(CondAL, PCOFFSET(JumpDirect));
+         didBreak = true;
+         break;
+      }
+
+   CASE_OP(JMP):
+   CASE_OP(JML):
+      {
+         switch(in->AddrMode)
+         {
+         case AddrImmediate16:
+            dcycles += CPU.MemSpeedx2;
+            GenARM_MOV32_Imm(CondAL, RegA1, in->PC[1] | (in->PC[2] << 8) | ICPU.ShiftedPB);
+
+            /* Only flush cycles because PC will change */
+            GenARM_ADD_Imm(CondAL, RegCycles, RegCycles, ENCODE_IMM(dcycles, 0));
+
+            break;
+         case AddrImmediate24:
+            dcycles += CPU.MemSpeedx2 + CPU.MemSpeed;
+            GenARM_MOV32_Imm(CondAL, RegA1, in->PC[1] | (in->PC[2] << 8));
+            GenARM_MOV_Imm(CondAL, RegA2, ENCODE_IMM(in->PC[3], 0));
+            EmitStoreRegister(in, ArgPB, RegA2, true);
+            GenARM_ORR_Imm(CondAL, RegA1, RegA1, ENCODE_IMM(in->PC[3], 16));
+
+            /* Only flush cycles because PC will change */
+            GenARM_ADD_Imm(CondAL, RegCycles, RegCycles, ENCODE_IMM(dcycles, 0));
+
+            break;
+         case AddrAbsolute16:
+            dcycles += CPU.MemSpeedx2;
+            dpc += 2;
+            GenARM_MOV32_Imm(CondAL, RegA1, in->PC[1] | (in->PC[2] << 8));
+
+            if (in->Arg2 == ArgX)
+            {
+               GenARM_ADD_Reg(CondAL, RegA1, RegA1, 0, RegX);
+               GenARM_BIC_Imm(CondAL, RegA1, RegA1, ENCODE_IMM(0xFF, 16));
+               GenARM_ORR_Imm(CondAL, RegA1, RegA1, ENCODE_IMM(ICPU.Registers.PB, 16));
+               dcycles += ONE_CYCLE;
+            }
+
+            /* Update cycles / cpu for slow lookup */
+            GenARM_MOV_Reg(CondAL, RegA2, 0, RegCPUPC);
+            EmitFlushCyclesPC(&dpc, &dcycles);
+
+            GenARM_BL(CondAL, PCOFFSET(S9xGetWordFast));
+
+            if (in->Op == FnJMP)
+            {
+               GenARM_ORR_Imm(CondAL, RegA1, RegA1, ENCODE_IMM(ICPU.Registers.PB, 16));
+            }
+            else
+            {
+               GenARM_MOV_Reg(CondAL, RegA4, 0, RegA1);
+               GenARM_ADD_Imm(CondAL, RegA1, RegA2, ENCODE_IMM(2, 0));
+               GenARM_SUB_Imm(CondAL, RegA2, RegCPUPC, ENCODE_IMM(3, 0));
+               GenARM_BL(CondAL, PCOFFSET(S9xGetByteFast));
+               GenARM_MOV_Reg(CondAL, RegA2, ENCODE_SHIFT(ShiftLSL, 16), RegA1);
+               GenARM_STRB_Imm(CondAL, RegA1, RegCPU, ICPU_OFFSET(Registers.PB));
+               GenARM_STR_Imm(CondAL, RegA2, RegCPU, ICPU_OFFSET(ShiftedPB));
+               GenARM_ORR_Reg(CondAL, RegA1, RegA4, 0, RegA2);
+            }
+
+            break;
+         default:
+            break;
+         }
+
+         GenARM_BL(CondAL, PCOFFSET(S9xCallSetPCBase));
+         GenARM_LDR_Imm(CondAL, RegCPUPC, RegCPU, CPU_OFFSET(PC));
+
+         if (in->AddrMode == AddrImmediate24 || in->Op == FnJML)
+            GenARM_BL(CondAL, PCOFFSET(S9xCallUpdateSection));
+
+         if (in->AddrMode == AddrImmediate16 && Settings.Shutdown)
+         {
+            GenARM_LDR_Imm(CondAL, RegA1, RegCPU, CPU_OFFSET(WaitAddress));
+            GenARM_CMP_Reg(CondAL, RegA1, 0, RegCPUPC);
+            GenARM_BL(CondEQ, PCOFFSET(S9xCallCPUShutdown));
+         }
+
+         EmitChecks(0);
+
+         if (in->AddrMode == AddrImmediate16 || in->AddrMode == AddrImmediate24)
+            GenARM_BL(CondAL, PCOFFSET(JumpDirect));
+         else
+            GenARM_BL(CondAL, PCOFFSET(JumpIndirect));
+
+         didBreak = true;
+         break;
+      }
+   CASE_OP(JSR):
+   CASE_OP(JSL):
+      {
+         if (in->Op == FnJSL)
+         {
+            GenARM_MOV_Imm(CondAL, RegA1, ENCODE_IMM(ICPU.Registers.PB, 0));
+            GenARM_LDRH_Imm(CondAL, RegA4, RegCPU, ICPU_OFFSET(Registers.S));
+
+            dcycles += CPU.MemSpeed * 3;
+            dpc += 3;
+            /* Only flush cycles because PC will change */
+            GenARM_ADD_Imm(CondAL, RegCycles, RegCycles, ENCODE_IMM(dcycles, 0));
+
+            EmitPush(in, true);
+         }
+         else
+         {
+            dcycles += CPU.MemSpeedx2 + ONE_CYCLE;
+            dpc += 2;
+            /* Only flush cycles because PC will change */
+            GenARM_ADD_Imm(CondAL, RegCycles, RegCycles, ENCODE_IMM(dcycles, 0));
+            GenARM_LDRH_Imm(CondAL, RegA4, RegCPU, ICPU_OFFSET(Registers.S));
+         }
+
+         GenARM_LDR_Imm(CondAL, RegA1, RegCPU, CPU_OFFSET(PCBase));
+         GenARM_ADD_Imm(CondAL, RegA3, RegCPUPC, ENCODE_IMM(dpc - 1, 0));
+         GenARM_SUB_Reg(CondAL, RegA1, RegA3, 0, RegA1);
+
+         EmitPush(in, false);
+
+         /* Load new PC */
+         if (in->Op == FnJSL)
+         {
+            GenARM_MOV_Imm(CondAL, RegA2, ENCODE_IMM(in->PC[3], 0));
+            GenARM_STRB_Imm(CondAL, RegA2, RegCPU, ICPU_OFFSET(Registers.PB));
+            GenARM_MOV_Reg(CondAL, RegA2, ENCODE_SHIFT(ShiftLSL, 16), RegA2);
+            GenARM_STR_Imm(CondAL, RegA2, RegCPU, ICPU_OFFSET(ShiftedPB));
+
+            GenARM_MOV32_Imm(CondAL, RegA1, in->PC[1] + (in->PC[2] << 8) + (in->PC[3] << 16));
+         }
+         else if (in->Arg2 == ArgNULL)
+         {
+            GenARM_MOV32_Imm(CondAL, RegA2, in->PC[1] + (in->PC[2] << 8));
+            GenARM_ADD_Imm(CondAL, RegA1, RegA2, ENCODE_IMM(ICPU.Registers.PB, 16));
+         }
+         else
+         {
+            GenARM_MOV32_Imm(CondAL, RegA1, in->PC[1] | (in->PC[2] << 8));
+            /* Set OpenBus to PC[2] */
+
+            GenARM_ADD_Reg(CondAL, RegA1, RegA1, 0, RegX);
+            GenARM_BIC_Imm(CondAL, RegA1, RegA1, ENCODE_IMM(0xFF, 16));
+            GenARM_ORR_Imm(CondAL, RegA1, RegA1, ENCODE_IMM(ICPU.Registers.PB, 16));
+
+            /* Update cycles / cpu for slow lookup */
+            GenARM_MOV_Reg(CondAL, RegA2, 0, RegCPUPC);
+            EmitFlushCyclesPC(&dpc, &dcycles);
+
+            GenARM_BL(CondAL, PCOFFSET(S9xGetWordFast));
+            GenARM_ORR_Imm(CondAL, RegA1, RegA1, ENCODE_IMM(ICPU.Registers.PB, 16));
+         }
+
+         GenARM_STRH_Imm(CondAL, RegA4, RegCPU, ICPU_OFFSET(Registers.S));
+
+         GenARM_BL(CondAL, PCOFFSET(S9xCallSetPCBase));
+         GenARM_LDR_Imm(CondAL, RegCPUPC, RegCPU, CPU_OFFSET(PC));
+
+         if (in->Op == FnJSL)
+            GenARM_BL(CondAL, PCOFFSET(S9xCallUpdateSection));
+
+         EmitChecks(0);
+         if (in->Arg2 == ArgNULL)
+            GenARM_BL(CondAL, PCOFFSET(JumpDirect));
+         else
+            GenARM_BL(CondAL, PCOFFSET(JumpIndirect));
+         didBreak = true;
+         break;
+      }
+   CASE_OP(RTI):
+      {
+         uint8_t *branch1;
+         dcycles += ONE_CYCLE * 2;
+
+         GenARM_LDRH_Imm(CondAL, RegA4, RegCPU, ICPU_OFFSET(Registers.S));
+
+         GenARM_ADD_Imm(CondAL, RegCycles, RegCycles, ENCODE_IMM(dcycles, 0));
+         EmitPop(in, true);
+         GenARM_STRH_Imm(CondAL, RegA4, RegCPU, ICPU_OFFSET(Registers.S));
+
+         EmitStoreRegister(in, ArgP, RegA1, false);
+
+         GenARM_LDRH_Imm(CondAL, RegA4, RegCPU, ICPU_OFFSET(Registers.S));
+
+         EmitPop(in, false);
+         GenARM_UXTH_Reg(CondAL, RegCPUPC, RegA1); /* RegCPU will be overwritten later */
+
+         /* Ignore pop PB if Emulation is set */
+         GenARM_TST_Imm(CondAL, RegP, ENCODE_IMM(Emulation >> 8, 8));
+         branch1 = CachePtr;
+         GenARM_B(CondNE, PCOFFSET(branch1));
+
+         EmitPop(in, true);
+
+         GenARM_MOV_Reg(CondAL, RegA2, ENCODE_SHIFT(ShiftLSL, 16), RegA1);
+         GenARM_STRB_Imm(CondAL, RegA1, RegCPU, ICPU_OFFSET(Registers.PB));
+         GenARM_STR_Imm(CondAL, RegA2, RegCPU, ICPU_OFFSET(ShiftedPB));
+
+         PatchJump((uint32_t *)branch1, CachePtr);
+
+         GenARM_LDR_Imm(CondNE, RegA2, RegCPU, ICPU_OFFSET(ShiftedPB));
+         GenARM_ADD_Reg(CondAL, RegA1, RegCPUPC, 0, RegA2);
+
+         GenARM_STRH_Imm(CondAL, RegA4, RegCPU, ICPU_OFFSET(Registers.S));
+
+         GenARM_BL(CondAL, PCOFFSET(S9xCallSetPCBase));
+         GenARM_LDR_Imm(CondAL, RegCPUPC, RegCPU, CPU_OFFSET(PC));
+
+         GenARM_BL(CondAL, PCOFFSET(S9xCallUpdateSection));
+
+         EmitChecks(0);
+         GenARM_B(CondAL, PCOFFSET(JumpIndirect));
+         didBreak = true;
+         break;
+      }
+   CASE_OP(RTL):
+   CASE_OP(RTS):
+      {
+         dcycles += ONE_CYCLE * 2;
+         if (in->Op == FnRTS)
+            dcycles += ONE_CYCLE;
+
+         GenARM_LDRH_Imm(CondAL, RegA4, RegCPU, ICPU_OFFSET(Registers.S));
+
+         GenARM_ADD_Imm(CondAL, RegCycles, RegCycles, ENCODE_IMM(dcycles, 0));
+         EmitPop(in, false);
+
+         GenARM_ADD_Imm(CondAL, RegA1, RegA1, ENCODE_IMM(1, 0));
+         GenARM_UXTH_Reg(CondAL, RegCPUPC, RegA1); /* No need to preserve RegCPU. It will be overwritten later, and this should never be a WaitAddress */
+
+         if (in->Op == FnRTL)
+         {
+            EmitPop(in, true);
+            GenARM_MOV_Reg(CondAL, RegA2, ENCODE_SHIFT(ShiftLSL, 16), RegA1);
+            GenARM_STRB_Imm(CondAL, RegA1, RegCPU, ICPU_OFFSET(Registers.PB));
+            GenARM_STR_Imm(CondAL, RegA2, RegCPU, ICPU_OFFSET(ShiftedPB));
+            GenARM_ADD_Reg(CondAL, RegA1, RegCPUPC, 0, RegA2);
+         }
+         else
+         {
+            GenARM_ADD_Imm(CondAL, RegA1, RegCPUPC, ENCODE_IMM(ICPU.Registers.PB, 16));
+         }
+
+         GenARM_STRH_Imm(CondAL, RegA4, RegCPU, ICPU_OFFSET(Registers.S));
+
+         GenARM_BL(CondAL, PCOFFSET(S9xCallSetPCBase));
+         GenARM_LDR_Imm(CondAL, RegCPUPC, RegCPU, CPU_OFFSET(PC));
+
+         if (in->Op == FnRTL)
+            GenARM_BL(CondAL, PCOFFSET(S9xCallUpdateSection));
+
+         EmitChecks(0);
+         GenARM_B(CondAL, PCOFFSET(JumpIndirect));
+         didBreak = true;
+         break;
+      }
+   CASE_OP(BPL):
+      EmitBranch(in, Negative, false, &dpc, &dcycles);
+      didBreak = true;
+      break;
+   CASE_OP(BMI):
+      EmitBranch(in, Negative, true, &dpc, &dcycles);
+      didBreak = true;
+      break;
+   CASE_OP(BVC):
+      EmitBranch(in, Overflow, false, &dpc, &dcycles);
+      didBreak = true;
+      break;
+   CASE_OP(BVS):
+      EmitBranch(in, Overflow, true, &dpc, &dcycles);
+      didBreak = true;
+      break;
+   CASE_OP(BCC):
+      EmitBranch(in, Carry, false, &dpc, &dcycles);
+      didBreak = true;
+      break;
+   CASE_OP(BCS):
+      EmitBranch(in, Carry, true, &dpc, &dcycles);
+      didBreak = true;
+      break;
+   CASE_OP(BZC):
+      EmitBranch(in, Zero, false, &dpc, &dcycles);
+      didBreak = true;
+      break;
+   CASE_OP(BZS):
+      EmitBranch(in, Zero, true, &dpc, &dcycles);
+      didBreak = true;
+      break;
+   CASE_OP(BRK):
+   CASE_OP(COP):
+      {
+         uint32_t destPC;
+         dcycles += ONE_CYCLE;
+
+         GenARM_LDRH_Imm(CondAL, RegA4, RegCPU, ICPU_OFFSET(Registers.S));
+
+         if (!CheckEmulation())
+         {
+            GenARM_LDRB_Imm(CondAL, RegA1, RegCPU, ICPU_OFFSET(Registers.PB));
+            EmitPush(in, true);
+            dcycles += ONE_CYCLE;
+         }
+         EmitFlushCyclesPC(&dpc, &dcycles);
+
+         GenARM_LDR_Imm(CondAL, RegA1, RegCPU, CPU_OFFSET(PCBase));
+
+         if (!CheckEmulation())
+         {
+            GenARM_ADD_Imm(CondAL, RegA3, RegCPUPC, ENCODE_IMM(1, 0));
+         }
+
+         GenARM_SUB_Reg(CondAL, RegA1, RegA3, 0, RegA1);
+
+         EmitPush(in, false);
+
+         GenARM_AND_Imm(CondAL, RegA1, RegP, ENCODE_IMM(0xFF, 0));
+         EmitPush(in, true);
+         GenARM_STRH_Imm(CondAL, RegA4, RegCPU, ICPU_OFFSET(Registers.S));
+
+         GenARM_STRB_Imm(CondAL, RegA1, RegCPU, CPU_OPEN_BUS_OFFSET);
+
+         GenARM_BIC_Imm(CondAL, RegP, RegP, ENCODE_IMM(Decimal, 0));
+         GenARM_ORR_Imm(CondAL, RegP, RegP, ENCODE_IMM(IRQ, 0));
+         GenARM_STRH_Imm(CondAL, RegP, RegCPU, ICPU_OFFSET(Registers.P));
+
+         GenARM_MOV_Imm(CondAL, RegA1, ENCODE_IMM(0, 0));
+         GenARM_STRB_Imm(CondAL, RegA1, RegCPU, ICPU_OFFSET(Registers.PB));
+         GenARM_STR_Imm(CondAL, RegA1, RegCPU, ICPU_OFFSET(ShiftedPB));
+
+         if (in->Op == FnBRK)
+            destPC = CheckEmulation() ? 0xFFFE : 0xFFE6;
+         else
+            destPC = CheckEmulation() ? 0xFFF4 : 0xFFE4;
+
+         GenARM_MOV32_Imm(CondAL, RegA1, destPC);
+         GenARM_SUB_Imm(CondAL, RegA2, RegCPUPC, ENCODE_IMM(1, 0));
+         GenARM_BL(CondAL, PCOFFSET(S9xGetWordFast));
+
+         GenARM_BL(CondAL, PCOFFSET(S9xCallSetPCBase));
+         GenARM_LDR_Imm(CondAL, RegCPUPC, RegCPU, CPU_OFFSET(PC));
+
+         GenARM_BL(CondAL, PCOFFSET(S9xCallUpdateSection));
+
+         EmitChecks(0);
+         GenARM_BL(CondAL, PCOFFSET(JumpIndirect));
+         didBreak = true;
+         break;
+      }
+   CASE_OP(CLC):
+   CASE_OP(CLI):
+   CASE_OP(CLD):
+   CASE_OP(CLV):
+   CASE_OP(SEC):
+   CASE_OP(SEI):
+   CASE_OP(SED):
+      {
+         uint16_t flag;
+         bool update = false;
+
+         switch(in->Op)
+         {
+         CASE_OP(SEC):
+         CASE_OP(CLC):
+            flag = Carry;
+            break;
+         CASE_OP(SEI):
+         CASE_OP(CLI):
+            flag = IRQ;
+            update = true;
+            break;
+         CASE_OP(SED):
+         CASE_OP(CLD):
+            flag = Decimal;
+            update = true;
+            break;
+         CASE_OP(CLV):
+            flag = Overflow;
+            break;
+         default:
+            fprintf(stderr, "Invalid opcode: %X\n", in->Opcode);
+            break;
+         }
+
+         dcycles += ONE_CYCLE;
+         EmitFlushCyclesPC(&dpc, &dcycles);
+
+         if (in->Op == FnCLC || in->Op == FnCLD || in->Op == FnCLI || in->Op == FnCLV)
+            GenARM_BIC_Imm(CondAL, RegP, RegP, ENCODE_IMM(flag, 0));
+         else
+            GenARM_ORR_Imm(CondAL, RegP, RegP, ENCODE_IMM(flag, 0));
+
+         if (update)
+            GenARM_STRH_Imm(CondAL, RegP, RegCPU, ICPU_OFFSET(Registers.P));
+
+         EmitChecks(0);
+
+         if (in->SectionUpdate)
+            GenARM_BL(CondAL, PCOFFSET(S9xCallUpdateSection));
+
+         if (in->ShouldBreak)
+         {
+            GenARM_BL(CondAL, PCOFFSET(JumpDirect));
+            didBreak = true;
+         }
+         break;
+      }
+   CASE_OP(SEP):
+   CASE_OP(REP):
+      {
+         dpc++;
+         dcycles += CPU.MemSpeed + ONE_CYCLE;
+         EmitFlushCyclesPC(&dpc, &dcycles);
+
+         if (IS_OP(REP))
+            GenARM_BIC_Imm(CondAL, RegA1, RegP, ENCODE_IMM(in->PC[1], 0));
+         else
+            GenARM_ORR_Imm(CondAL, RegA1, RegP, ENCODE_IMM(in->PC[1], 0));
+
+         EmitStoreRegister(in, ArgP, RegA1, false);
+
+         EmitChecks(0);
+         GenARM_BL(CondAL, PCOFFSET(JumpDirect));
+         didBreak = true;
+         break;
+      }
+   CASE_OP(XCE):
+      {
+         dcycles += ONE_CYCLE;
+         EmitFlushCyclesPC(&dpc, &dcycles);
+
+         GenARM_BIC_Imm(CondAL, RegA2, RegP, ENCODE_IMM(Emulation >> 8, 8));
+         GenARM_BIC_Imm(CondAL, RegA2, RegA2, ENCODE_IMM(Carry, 0));
+         GenARM_ORR_Reg(CondAL, RegA2, RegA2, ENCODE_SHIFT(ShiftLSR, 8), RegP);
+         GenARM_TST_Imm(CondAL, RegP, ENCODE_IMM(Carry, 0));
+         GenARM_ORR_Imm(CondNE, RegA2, RegA2, ENCODE_IMM(Emulation >> 8, 8));
+
+         EmitStoreRegister(in, ArgP, RegA2, false);
+         EmitChecks(0);
+         GenARM_BL(CondAL, PCOFFSET(JumpIndirect));
+         didBreak = true;
+         break;
+      }
+   CASE_OP(STP):
+      {
+#ifdef NO_SPEEDHACKS
+         dpc--;
+         GenARM_LDR_Imm(CondAL, RegA1, RegCPU, CPU_OFFSET(Flags));
+         GenARM_ORR_Imm(CondAL, RegA1, RegA1, ENCODE_IMM(DEBUG_MODE_FLAG, 0));
+         GenARM_STR_Imm(CondAL, RegA1, RegCPU, CPU_OFFSET(Flags));
+#else
+         int8_t disp = (in->PC[1] & 0x7F) | ((in->PC[1] & 0x40) << 1);
+         uint32_t pc = in->PC + 2 - CPU.PCBase;
+         uint32_t target = (pc + disp) & 0xFFFF;
+         bool overflow = target != (pc + disp);
+         dpc++;
+
+         if (overflow)
+            disp = (target - pc);
+
+         GenARM_MOV_Imm(CondAL, RegA1, ENCODE_IMM(0, 0));
+         GenARM_STR_Imm(CondAL, RegA1, RegCPU, CPU_OFFSET(WaitAddress));
+         EmitFlushCyclesPC(&dpc, &dcycles);
+
+         GenARM_BL(CondAL, PCOFFSET(S9xCallWAIShutdown));
+
+         /* Interpreter runs BranchCheck here. Only when APU is disabled
+          * until next reset. So cost of the load seems not worth it, unless
+          * games break. */
+
+         GenARM_TST_Imm(CondAL, RegP, ENCODE_IMM(Zero, 0));
+
+         if (in->PC[1] & 0x80)
+            GenARM_BL(CondEQ, PCOFFSET(JumpDirectChecks));
+         else
+            GenARM_BL(CondNE, PCOFFSET(JumpDirectChecks));
+
+         GenARM_ADD_Imm(CondAL, RegCycles, RegCycles, ENCODE_IMM(ONE_CYCLE, 0));
+
+         if (Settings.Shutdown)
+         {
+            GenARM_LDR_Imm(CondAL, RegA1, RegCPU, CPU_OFFSET(WaitAddress));
+         }
+
+         if (disp > 0)
+         {
+            if (disp & 0xFF)
+               GenARM_ADD_Imm(CondAL, RegCPUPC, RegCPUPC, ENCODE_IMM(disp & 0xFF, 0));
+
+            if (disp & 0xFF00)
+               GenARM_ADD_Imm(CondAL, RegCPUPC, RegCPUPC, ENCODE_IMM(disp >> 8 & 0xFF, 8));
+         }
+         else if (disp < 0)
+         {
+            disp = -disp;
+            if (disp & 0xFF)
+               GenARM_SUB_Imm(CondAL, RegCPUPC, RegCPUPC, ENCODE_IMM(disp & 0xFF, 0));
+
+            if (disp & 0xFF00)
+               GenARM_SUB_Imm(CondAL, RegCPUPC, RegCPUPC, ENCODE_IMM(disp >> 8 & 0xFF, 8));
+         }
+
+         if (Settings.Shutdown) {
+            GenARM_CMP_Reg(CondAL, RegA1, 0, RegCPUPC);
+            GenARM_BL(CondEQ, PCOFFSET(S9xCallCPUShutdown));
+         }
+#endif
+         EmitChecks(0);
+         GenARM_BL(CondAL, PCOFFSET(JumpDirect));
+         didBreak = true;
+         break;
+      }
+   CASE_OP(XBA):
+      {
+         dcycles += TWO_CYCLES;
+         EmitFlushCyclesPC(&dpc, &dcycles);
+         GenARM_MOVS_Reg(CondAL, RegA1, ENCODE_SHIFT(ShiftLSR, 8), RegA);
+         GenARM_MOV_Reg(CondAL, RegA, ENCODE_SHIFT(ShiftLSL, 24), RegA);
+         GenARM_ORR_Reg(CondAL, RegA, RegA1, ENCODE_SHIFT(ShiftLSR, 16), RegA);
+
+         EmitSetFlags(RegA1, in->OutFlags & (Zero | Negative), true, false);
+         EmitChecks(0);
+         break;
+      }
+   CASE_OP(WAI):
+      {
+         uint8_t *loop;
+         GenARM_MOV_Imm(CondAL, RegA1, ENCODE_IMM(1, 0));
+         GenARM_STRB_Imm(CondAL, RegA1, RegCPU, CPU_OFFSET(WaitingForInterrupt));
+         dpc--;
+         loop = CachePtr;
+         EmitFlushCyclesPC(&dpc, &dcycles);
+
+         if (Settings.Shutdown) {
+            GenARM_BL(CondAL, PCOFFSET(S9xCallWAIShutdown));
+         }
+
+         EmitChecks(CheckWFI);
+         GenARM_B(CondAL, PCOFFSET(loop));
+         break;
+      }
+   CASE_OP(WDM):
+      {
+#ifndef NO_SPEEDHACKS
+         int8_t disp = 0xF0 | (in->PC[1] & 0xF);
+         uint32_t pc = in->PC + 2 - CPU.PCBase;
+         uint32_t target = (pc + disp) & 0xFFFF;
+         bool overflow = target != (pc + disp);
+         uint8_t flag;
+         uint8_t skip;
+         dpc++;
+
+         if (overflow)
+            disp = (target - pc);
+
+         GenARM_MOV_Imm(CondAL, RegA1, ENCODE_IMM(0, 0));
+         GenARM_STR_Imm(CondAL, RegA1, RegCPU, CPU_OFFSET(WaitAddress));
+         EmitFlushCyclesPC(&dpc, &dcycles);
+
+         GenARM_BL(CondAL, PCOFFSET(S9xCallWAIShutdown));
+
+         switch(in->PC[1] & 0xF0)
+         {
+         case 0x10:                /* BPL */
+            flag = Negative;
+            skip = CondNE;
+            break;
+         case 0x30:                /* BMI */
+            flag = Negative;
+            skip = CondEQ;
+            break;
+         case 0x50:                /* BVC */
+            flag = Overflow;
+            skip = CondNE;
+            break;
+         case 0x70:                /* BVS */
+            flag = Overflow;
+            skip = CondEQ;
+            break;
+         case 0x80:                /* BRA */
+            flag = 0;
+            break;
+         case 0x90:                /* BCC */
+            flag = Carry;
+            skip = CondNE;
+            break;
+         case 0xB0:                /* BCS */
+            flag = Carry;
+            skip = CondEQ;
+            break;
+         case 0xD0:                /* BNE */
+            flag = Zero;
+            skip = CondNE;
+            break;
+         case 0xF0:                /* BEQ */
+            flag = Zero;
+            skip = CondEQ;
+            break;
+         default:
+            didBreak = true;
+            break;
+         }
+
+         if (!didBreak)
+         {
+            /* Interpreter runs BranchCheck here. Only when APU is disabled
+             * until next reset. So cost of the load seems not worth it, unless
+             * games break. */
+
+            if (flag)
+            {
+               GenARM_TST_Imm(CondAL, RegP, ENCODE_IMM(flag, 0));
+               GenARM_BL(skip, PCOFFSET(JumpDirectChecks));
+            }
+
+            GenARM_ADD_Imm(CondAL, RegCycles, RegCycles, ENCODE_IMM(ONE_CYCLE, 0));
+
+            if (Settings.Shutdown)
+            {
+               GenARM_LDR_Imm(CondAL, RegA1, RegCPU, CPU_OFFSET(WaitAddress));
+            }
+
+            if (disp > 0)
+            {
+               if (disp & 0xFF)
+                  GenARM_ADD_Imm(CondAL, RegCPUPC, RegCPUPC, ENCODE_IMM(disp & 0xFF, 0));
+
+               if (disp & 0xFF00)
+                  GenARM_ADD_Imm(CondAL, RegCPUPC, RegCPUPC, ENCODE_IMM(disp >> 8 & 0xFF, 8));
+            }
+            else if (disp < 0)
+            {
+               disp = -disp;
+               if (disp & 0xFF)
+                  GenARM_SUB_Imm(CondAL, RegCPUPC, RegCPUPC, ENCODE_IMM(disp & 0xFF, 0));
+
+               if (disp & 0xFF00)
+                  GenARM_SUB_Imm(CondAL, RegCPUPC, RegCPUPC, ENCODE_IMM(disp >> 8 & 0xFF, 8));
+            }
+
+            if (Settings.Shutdown) {
+               GenARM_CMP_Reg(CondAL, RegA1, 0, RegCPUPC);
+               GenARM_BL(CondEQ, PCOFFSET(S9xCallCPUShutdown));
+            }
+         }
+
+#endif
+         EmitChecks(0);
+         GenARM_BL(CondAL, PCOFFSET(JumpDirect));
+         didBreak = true;
+         break;
+      }
+   CASE_OP(NOP):
+      {
+         dcycles += ONE_CYCLE;
+         EmitFlushCyclesPC(&dpc, &dcycles);
+         EmitChecks(0);
+         break;
+      }
+   }
+
+   if (in->ShouldBreak && !didBreak)
+   {
+      GenARM_B(CondAL, PCOFFSET(JumpIndirect));
+   }
+}
+
+static void EmitS9xCall(Instruction *in)
+{
+   /* CPU.PCAtOpcodeStart = CPU.PC; */
+   /* CPU.Cycles += CPU.MemSpeed; */
+   /* CPU.PC++; */
+
+   GenARM_STR_Imm(CondAL, RegCPUPC, RegCPU, CPU_OFFSET(PCAtOpcodeStart));
+   GenARM_ADD_Imm(CondAL, RegCPUPC, RegCPUPC, ENCODE_IMM(1, 0));
+   GenARM_ADD_Imm(CondAL, RegCycles, RegCycles, ENCODE_IMM(CPU.MemSpeed, 0));
+
+   GenARM_BL(CondAL, PCOFFSET(S9xRegsToMem));
+
+   GenARM_BL(CondAL, PCOFFSET((uintptr_t)in->S9xOpcode));
+
+   if (in->SectionUpdate)
+      GenARM_BL(CondAL, PCOFFSET(S9xCallUpdateSection));
+
+   GenARM_BL(CondAL, PCOFFSET(S9xMemToRegs));
+
+   EmitChecks(0);
+
+   if (in->ShouldBreak)
+   {
+      GenARM_B(CondAL, PCOFFSET(JumpIndirect));
+   }
+}
+
+static void EmitInterpreterLoop(Instruction *in)
+{
+   /* CPU.PCAtOpcodeStart = CPU.PC; */
+   /* CPU.Cycles += CPU.MemSpeed; */
+   /* CPU.PC++; */
+   uint8_t *start = CachePtr;
+
+   GenARM_LDRB_Imm(CondAL, RegA1, RegCPUPC, 0);
+   GenARM_STR_Imm(CondAL, RegCPUPC, RegCPU, CPU_OFFSET(PCAtOpcodeStart));
+   GenARM_LDR_Imm(CondAL, RegA2, RegCPU, ICPU_OFFSET(S9xOpcodes));
+   GenARM_ADD_Imm(CondAL, RegCPUPC, RegCPUPC, ENCODE_IMM(1, 0));
+   GenARM_ADD_Imm(CondAL, RegCycles, RegCycles, ENCODE_IMM(CPU.MemSpeed, 0));
+   GenARM_LDR_Reg(CondAL, RegA2, RegA2, 1, ENCODE_SHIFT(ShiftLSL, 2), RegA1);
+
+   GenARM_BL(CondAL, PCOFFSET(S9xRegsToMem));
+
+   GenARM_BLX_Reg(CondAL, RegA2);
+   GenARM_BL(CondAL, PCOFFSET(S9xCallUpdateSection));
+
+   GenARM_BL(CondAL, PCOFFSET(S9xMemToRegs));
+
+   EmitChecks(0);
+
+   GenARM_B(CondAL, PCOFFSET(start));
+}
+
+static void EmitOne(Instruction *in)
+{
+
+#ifndef NDEBUG
+   GenARM_MOV_Imm(CondAL, RegA1, ENCODE_IMM(in->Opcode, 0));
+   GenARM_MOV32_Imm(CondAL, RegA2, (uintptr_t)in->PC);
+   GenARM_MOV_Imm(CondAL, RegA3, ENCODE_IMM(in->Emitter == EmitS9xCall, 0));
+   GenARM_BL(CondAL, PCOFFSET(S9xCallCheckInstruction));
+#endif
+
+   /* Call Opcode */
+   in->Emitter(in);
+}
+
+#define SizeNone            0
+#define SizeRegister8       0
+#define SizeRegister16      0
+#define SizeImmediate8      1
+#define SizeImmediate16     2
+#define SizeImmediate24     3
+#define SizeZeroPage8       1
+#define SizeZeroPage16      1
+#define SizeIndirect8       1
+#define SizeIndirect16      1
+#define SizeIndirectX8      1
+#define SizeIndirectX16     1
+#define SizeIndirectY8      1
+#define SizeIndirectY16     1
+#define SizeIndirectS8      1
+#define SizeIndirectS16     1
+#define SizeIndirectFar8    1
+#define SizeIndirectFar16   1
+#define SizeAbsolute8       2
+#define SizeAbsolute16      2
+#define SizeLong8           3
+#define SizeLong16          3
+
+#define OPSIZE(cond, t, f) cond ? Size##t : Size##f
+#define OPADDRMODE(cond, t, f) cond ? Addr##t : Addr##f
+#define OPFN(cond, t, f) cond ? Fn##t : Fn##f
+
+#define OP(opcode, addrmode, op, arg1, arg2, gen, used, flags) \
+   case 0x##opcode: pc += OPSIZE(addrmode); in.Emitter = EmitNativeCall; in.ShouldBreak = (flags) & OFBreak; in.SectionUpdate = (flags) & OFSectionUpdate; in.Op = OPFN(op); in.AddrMode = OPADDRMODE(addrmode); in.Arg1 = Arg##arg1; in.Arg2 = Arg##arg2; in.OutFlags = gen; in.InFlags = used; break;
+
+#define OPD(opcode, addrmode, op, arg1, arg2, gen, used, flags) \
+   case 0x##opcode: pc += OPSIZE(addrmode); in.Emitter = EmitS9xCall; in.S9xOpcode = ICPU.S9xOpcodes[0x##opcode].S9xOpcode; in.SectionUpdate = (flags) & OFSectionUpdate; in.ShouldBreak = (flags) & OFBreak; in.OutFlags = gen; in.InFlags = used; break;
+
+static size_t ParseInstructions(uint8_t *pc, Instruction *instructions, size_t max)
+{
+   bool shouldBreak;
+   int i;
+
+   for (i = 0, shouldBreak = false; i < max && !shouldBreak; i++)
+   {
+      uint8_t opcode;
+      Instruction in = {0};
+
+      in.PC = pc;
+
+#include "arm_dynarec/opgen.h"
+
+      in.Opcode = opcode;
+
+      instructions[i] = in;
+      shouldBreak = in.ShouldBreak;
+   }
+
+   if (i > 0)
+      instructions[i - 1].ShouldBreak = true;
+
+   return i;
+}
+
+static void ScanInstructions(Instruction *instructions, size_t length)
+{
+   uint16_t CurrentFlags = (Negative | Zero | Carry | Overflow);
+   uint16_t i;
+
+   if (length < 2)
+      return;
+
+   /* Dead flag elimination */
+   for (i = length - 1; i > 0; i--)
+   {
+      Instruction *in = &instructions[i - 1];
+      Instruction *next = &instructions[i];
+
+      CurrentFlags = CurrentFlags | next->InFlags;
+      in->OutFlags = in->OutFlags & CurrentFlags;
+      CurrentFlags = CurrentFlags & ~in->OutFlags;
+   }
+}
+
+static void Emit(uint16_t pc, uint8_t **block)
+{
+   uint8_t *start = CachePtr;
+   uint8_t *StartPC = pc + CPU.PCBase;
+   uint32_t address = pc | ICPU.ShiftedPB;
+   int32_t MemoryBlock = (address >> MEMMAP_SHIFT) & MEMMAP_MASK;
+
+   Instruction instructions[100];
+   size_t length;
+   int i;
+
+   if (Memory.BlockIsRAM[MemoryBlock])
+   {
+      /* Code in RAM, for now just run interpreter until interrupt */
+      Instruction in = { 0 };
+      instructions[0] = in;
+      EmitInterpreterLoop(&instructions[0]);
+   }
+   else
+   {
+      length = ParseInstructions(StartPC, instructions, sizeof(instructions) / sizeof(instructions[0]));
+      ScanInstructions(instructions, length);
+
+      for (i = 0; i < length; i++)
+         EmitOne(&instructions[i]);
+   }
+
+   __clear_cache(start, CachePtr);
+   *block = start;
+}
+
+uint8_t **FetchBlock(uint16_t pc)
+{
+   uint8_t **block = FindBlock(pc);
+#ifndef NDEBUG
+   Metrics.Finds++;
+   Metrics.InterpretedBlock = false;
+
+   if (!*block)
+      Metrics.Emits++;
+#endif
+   if (!*block)
+      Emit(pc, block);
+
+#ifndef NDEBUG
+   trace[traceidx++] = (CacheSection << 16) | pc;
+   trace[traceidx++] = (uintptr_t)*block;
+   traceidx %= 128;
+#endif
+
+   return block;
+}
+
+void S9xMainLoop_Dyna(void)
+{
+   UpdateSection();
+
+#ifdef LAGFIX
+   do
+   {
+#endif
+      uint8_t **block = FetchBlock(CPU.PC - CPU.PCBase);
+      BlockEnter(*block);
+
+      ICPU.Registers.PC = CPU.PC - CPU.PCBase;
+#ifndef USE_BLARGG_APU
+      IAPU.Registers.PC = IAPU.PC - IAPU.RAM;
+#endif
+
+#ifdef LAGFIX
+      if(!finishedFrame)
+      {
+#endif
+         S9xPackStatus();
+#ifndef USE_BLARGG_APU
+         S9xAPUPackStatus();
+#endif
+         CPU.Flags &= ~SCAN_KEYS_FLAG;
+#ifdef LAGFIX
+      }
+      else
+      {
+         finishedFrame = false;
+         break;
+      }
+   } while(!finishedFrame);
+#endif
+}
+
+int DynaInit(void)
+{
+   static bool inited = false;
+
+   if (!inited)
+   {
+      inited = true;
+#ifndef NDEBUG
+      DynaBreak();
+#endif
+
+      Cache = MapRWX((void *)((uintptr_t)DynaInit), BUFFER_SIZE + BUFFER_EXTRA);
+
+      if (!Cache)
+         return -1;
+   }
+
+   return 0;
+}
+
+int DynaReset(void)
+{
+   if (DynaInit())
+      return -1;
+
+   printf("Starting dynarec\n");
+   CacheEmpty();
+   return 0;
+}
diff --git a/source/arm_dynarec/dynaexec.h b/source/arm_dynarec/dynaexec.h
new file mode 100644
index 0000000..b8a807d
--- /dev/null
+++ b/source/arm_dynarec/dynaexec.h
@@ -0,0 +1,14 @@
+#ifndef DYNAEXEC_H
+#define DYNAEXEC_H
+
+#include <stdint.h>
+#include <stddef.h>
+
+void S9xMainLoop_Dyna();
+int DynaInit(void);
+int DynaReset(void);
+void DynaBreak(void);
+
+void *MapRWX(void *target, size_t size);
+
+#endif  /* DYNAEXEC_H */
diff --git a/source/arm_dynarec/map_rwx.c b/source/arm_dynarec/map_rwx.c
new file mode 100644
index 0000000..d4373f7
--- /dev/null
+++ b/source/arm_dynarec/map_rwx.c
@@ -0,0 +1,16 @@
+#define _GNU_SOURCE             /* MAP_ANONYMOUS */
+
+#include <sys/mman.h>
+#include <stdlib.h>
+
+void *MapRWX(void *target, size_t size) {
+	void *buf = mmap(target, size, PROT_READ | PROT_WRITE | PROT_EXEC, MAP_ANON | MAP_PRIVATE, -1, 0);
+
+	if (buf == MAP_FAILED)
+	{
+		perror("mmap");
+		return NULL;
+	}
+
+	return buf;
+}
diff --git a/source/arm_dynarec/opdef.h b/source/arm_dynarec/opdef.h
new file mode 100644
index 0000000..80606d8
--- /dev/null
+++ b/source/arm_dynarec/opdef.h
@@ -0,0 +1,281 @@
+OP (A8, XF(Register),        XF(MOV),    Y,    A,    NZ,   0,    OFNone);
+OP (AA, XF(Register),        XF(MOV),    X,    A,    NZ,   0,    OFNone);
+OP (BA, XF(Register),        XF(MOV),    X,    S,    NZ,   0,    OFNone);
+OP (98, MF(Register),        MF(MOV),    A,    Y,    NZ,   0,    OFNone);
+OP (8A, MF(Register),        MF(MOV),    A,    X,    NZ,   0,    OFNone);
+OP (9A, EF(Register),        EF(MOVSP),  S,    X,    0,    0,    OFNone);
+OP (9B, XF(Register),        XF(MOV),    Y,    X,    NZ,   0,    OFNone);
+OP (BB, XF(Register),        XF(MOV),    X,    Y,    NZ,   0,    OFNone);
+OP (7B, NF16(Register),      NF16(MOV),  A,    D,    NZ,   0,    OFNone);
+OP (5B, NF16(Register),      NF16(MOV),  D,    A,    NZ,   0,    OFNone);
+OP (3B, NF16(Register),      NF16(MOV),  A,    S,    NZ,   0,    OFNone);
+OP (1B, EF(Register),        EF(MOVSP),  S,    A,    0,    0,    OFNone);
+
+OP (A9, MF(Immediate),       MF(LD),     A,    NULL, NZ,   0,    OFNone);
+OP (A5, MF(ZeroPage),        MF(LD),     A,    NULL, NZ,   0,    OFNone);
+OP (B5, MF(ZeroPage),        MF(LD),     A,    X,    NZ,   0,    OFNone);
+OP (A3, MF(ZeroPage),        MF(LD),     A,    S,    NZ,   0,    OFNone);
+OP (AD, MF(Absolute),        MF(LD),     A,    NULL, NZ,   0,    OFNone);
+OP (BD, MF(Absolute),        MF(LD),     A,    X,    NZ,   0,    OFNone);
+OP (B9, MF(Absolute),        MF(LD),     A,    Y,    NZ,   0,    OFNone);
+OP (AF, MF(Long),            MF(LD),     A,    NULL, NZ,   0,    OFNone);
+OP (BF, MF(Long),            MF(LD),     A,    X,    NZ,   0,    OFNone);
+OP (B2, MF(Indirect),        MF(LD),     A,    NULL, NZ,   0,    OFNone);
+OP (A1, MF(IndirectX),       MF(LD),     A,    X,    NZ,   0,    OFNone);
+OP (B1, MF(IndirectY),       MF(LD),     A,    Y,    NZ,   0,    OFNone);
+OP (B3, MF(IndirectS),       MF(LD),     A,    Y,    NZ,   0,    OFNone);
+OP (A7, MF(IndirectFar),     MF(LD),     A,    NULL, NZ,   0,    OFNone);
+OP (B7, MF(IndirectFar),     MF(LD),     A,    Y,    NZ,   0,    OFNone);
+OP (A2, XF(Immediate),       XF(LD),     X,    NULL, NZ,   0,    OFNone);
+OP (A6, XF(ZeroPage),        XF(LD),     X,    NULL, NZ,   0,    OFNone);
+OP (B6, XF(ZeroPage),        XF(LD),     X,    Y,    NZ,   0,    OFNone);
+OP (AE, XF(Absolute),        XF(LD),     X,    NULL, NZ,   0,    OFNone);
+OP (BE, XF(Absolute),        XF(LD),     X,    Y,    NZ,   0,    OFNone);
+OP (A0, XF(Immediate),       XF(LD),     Y,    NULL, NZ,   0,    OFNone);
+OP (A4, XF(ZeroPage),        XF(LD),     Y,    NULL, NZ,   0,    OFNone);
+OP (B4, XF(ZeroPage),        XF(LD),     Y,    X,    NZ,   0,    OFNone);
+OP (AC, XF(Absolute),        XF(LD),     Y,    NULL, NZ,   0,    OFNone);
+OP (BC, XF(Absolute),        XF(LD),     Y,    X,    NZ,   0,    OFNone);
+
+OP (64, MF(ZeroPage),        MF(ST),     Z,    NULL, 0,    0,    OFNone);
+OP (74, MF(ZeroPage),        MF(ST),     Z,    X,    0,    0,    OFNone);
+OP (9C, MF(Absolute),        MF(ST),     Z,    NULL, 0,    0,    OFNone);
+OP (9E, MF(Absolute),        MF(ST),     Z,    X,    0,    0,    OFNone);
+OP (85, MF(ZeroPage),        MF(ST),     A,    NULL, 0,    0,    OFNone);
+OP (95, MF(ZeroPage),        MF(ST),     A,    X,    0,    0,    OFNone);
+OP (83, MF(ZeroPage),        MF(ST),     A,    S,    0,    0,    OFNone);
+OP (8D, MF(Absolute),        MF(ST),     A,    NULL, 0,    0,    OFNone);
+OP (9D, MF(Absolute),        MF(ST),     A,    X,    0,    0,    OFNone);
+OP (99, MF(Absolute),        MF(ST),     A,    Y,    0,    0,    OFNone);
+OP (8F, MF(Long),            MF(ST),     A,    NULL, 0,    0,    OFNone);
+OP (9F, MF(Long),            MF(ST),     A,    X,    0,    0,    OFNone);
+OP (81, MF(IndirectX),       MF(ST),     A,    X,    0,    0,    OFNone);
+OP (91, MF(IndirectY),       MF(ST),     A,    Y,    0,    0,    OFNone);
+OP (92, MF(Indirect),        MF(ST),     A,    NULL, 0,    0,    OFNone);
+OP (93, MF(IndirectS),       MF(ST),     A,    Y,    0,    0,    OFNone);
+OP (87, MF(IndirectFar),     MF(ST),     A,    NULL, 0,    0,    OFNone);
+OP (97, MF(IndirectFar),     MF(ST),     A,    Y,    0,    0,    OFNone);
+OP (86, XF(ZeroPage),        XF(ST),     X,    NULL, 0,    0,    OFNone);
+OP (96, XF(ZeroPage),        XF(ST),     X,    Y,    0,    0,    OFNone);
+OP (8E, XF(Absolute),        XF(ST),     X,    NULL, 0,    0,    OFNone);
+OP (84, XF(ZeroPage),        XF(ST),     Y,    NULL, 0,    0,    OFNone);
+OP (94, XF(ZeroPage),        XF(ST),     Y,    X,    0,    0,    OFNone);
+OP (8C, XF(Absolute),        XF(ST),     Y,    NULL, 0,    0,    OFNone);
+
+OP (48, MF(Register),        MF(PUSH),   A,    NULL, 0,    0,    OFNone);
+OP (DA, XF(Register),        XF(PUSH),   X,    NULL, 0,    0,    OFNone);
+OP (5A, XF(Register),        XF(PUSH),   Y,    NULL, 0,    0,    OFNone);
+OP (08, NF8(Register),       NF8(PUSH),  P,    NULL, 0,    NZCV, OFNone);
+OP (8B, NF8(Register),       NF8(PUSH),  DB,   NULL, 0,    0,    OFNone);
+OP (4B, NF8(Register),       NF8(PUSH),  PB,   NULL, 0,    0,    OFNone);
+OP (0B, NF16(Register),      NF16(PUSH), D,    NULL, 0,    0,    OFNone);
+OP (D4, NF16(ZeroPage),      NF16(PUSH), NULL, NULL, 0,    0,    OFNone);
+OP (F4, NF16(Immediate),     NF(PEA),    NULL, NULL, 0,    0,    OFNone);
+OP (62, NF16(Immediate),     NF(PER),    NULL, NULL, 0,    0,    OFNone);
+OP (68, MF(Register),        MF(POP),    A,    NULL, NZ,   0,    OFNone);
+OP (FA, XF(Register),        XF(POP),    X,    NULL, NZ,   0,    OFNone);
+OP (7A, XF(Register),        XF(POP),    Y,    NULL, NZ,   0,    OFNone);
+OP (2B, NF16(Register),      NF16(POP),  D,    NULL, NZ,   0,    OFNone);
+OP (AB, NF8(Register),       NF8(POP),   DB,   NULL, NZ,   0,    OFNone);
+OP (28, NF8(Register),       NF8(POP),   P,    NULL, NZCV, 0,    OFBreak | OFSectionUpdate);
+
+OP (44, NF16(Immediate),     NF(MVP),    NULL, NULL, 0,    NZCV, OFBreak);
+OP (54, NF16(Immediate),     NF(MVN),    NULL, NULL, 0,    NZCV, OFBreak);
+
+OP (09, MF(Immediate),       MF(OR),     A,    NULL, NZ,   0,    OFNone);
+OP (05, MF(ZeroPage),        MF(OR),     A,    NULL, NZ,   0,    OFNone);
+OP (15, MF(ZeroPage),        MF(OR),     A,    X,    NZ,   0,    OFNone);
+OP (0D, MF(Absolute),        MF(OR),     A,    NULL, NZ,   0,    OFNone);
+OP (1D, MF(Absolute),        MF(OR),     A,    X,    NZ,   0,    OFNone);
+OP (19, MF(Absolute),        MF(OR),     A,    Y,    NZ,   0,    OFNone);
+OP (01, MF(IndirectX),       MF(OR),     A,    X,    NZ,   0,    OFNone);
+OP (11, MF(IndirectY),       MF(OR),     A,    Y,    NZ,   0,    OFNone);
+OP (12, MF(Indirect),        MF(OR),     A,    NULL, NZ,   0,    OFNone);
+OP (03, MF(ZeroPage),        MF(OR),     A,    S,    NZ,   0,    OFNone);
+OP (13, MF(IndirectS),       MF(OR),     A,    Y,    NZ,   0,    OFNone);
+OP (07, MF(IndirectFar),     MF(OR),     A,    NULL, NZ,   0,    OFNone);
+OP (17, MF(IndirectFar),     MF(OR),     A,    Y,    NZ,   0,    OFNone);
+OP (0F, MF(Long),            MF(OR),     A,    NULL, NZ,   0,    OFNone);
+OP (1F, MF(Long),            MF(OR),     A,    X,    NZ,   0,    OFNone);
+
+OP (29, MF(Immediate),       MF(AND),    A,    NULL, NZ,   0,    OFNone);
+OP (25, MF(ZeroPage),        MF(AND),    A,    NULL, NZ,   0,    OFNone);
+OP (35, MF(ZeroPage),        MF(AND),    A,    X,    NZ,   0,    OFNone);
+OP (2D, MF(Absolute),        MF(AND),    A,    NULL, NZ,   0,    OFNone);
+OP (3D, MF(Absolute),        MF(AND),    A,    X,    NZ,   0,    OFNone);
+OP (39, MF(Absolute),        MF(AND),    A,    Y,    NZ,   0,    OFNone);
+OP (21, MF(IndirectX),       MF(AND),    A,    X,    NZ,   0,    OFNone);
+OP (31, MF(IndirectY),       MF(AND),    A,    Y,    NZ,   0,    OFNone);
+OP (32, MF(Indirect),        MF(AND),    A,    NULL, NZ,   0,    OFNone);
+OP (23, MF(ZeroPage),        MF(AND),    A,    S,    NZ,   0,    OFNone);
+OP (33, MF(IndirectS),       MF(AND),    A,    Y,    NZ,   0,    OFNone);
+OP (27, MF(IndirectFar),     MF(AND),    A,    NULL, NZ,   0,    OFNone);
+OP (37, MF(IndirectFar),     MF(AND),    A,    Y,    NZ,   0,    OFNone);
+OP (2F, MF(Long),            MF(AND),    A,    NULL, NZ,   0,    OFNone);
+OP (3F, MF(Long),            MF(AND),    A,    X,    NZ,   0,    OFNone);
+
+OP (49, MF(Immediate),       MF(EOR),    A,    NULL, NZ,   0,    OFNone);
+OP (45, MF(ZeroPage),        MF(EOR),    A,    NULL, NZ,   0,    OFNone);
+OP (55, MF(ZeroPage),        MF(EOR),    A,    X,    NZ,   0,    OFNone);
+OP (4D, MF(Absolute),        MF(EOR),    A,    NULL, NZ,   0,    OFNone);
+OP (5D, MF(Absolute),        MF(EOR),    A,    X,    NZ,   0,    OFNone);
+OP (59, MF(Absolute),        MF(EOR),    A,    Y,    NZ,   0,    OFNone);
+OP (41, MF(IndirectX),       MF(EOR),    A,    X,    NZ,   0,    OFNone);
+OP (51, MF(IndirectY),       MF(EOR),    A,    Y,    NZ,   0,    OFNone);
+OP (52, MF(Indirect),        MF(EOR),    A,    NULL, NZ,   0,    OFNone);
+OP (43, MF(ZeroPage),        MF(EOR),    A,    S,    NZ,   0,    OFNone);
+OP (53, MF(IndirectS),       MF(EOR),    A,    Y,    NZ,   0,    OFNone);
+OP (47, MF(IndirectFar),     MF(EOR),    A,    NULL, NZ,   0,    OFNone);
+OP (57, MF(IndirectFar),     MF(EOR),    A,    Y,    NZ,   0,    OFNone);
+OP (4F, MF(Long),            MF(EOR),    A,    NULL, NZ,   0,    OFNone);
+OP (5F, MF(Long),            MF(EOR),    A,    X,    NZ,   0,    OFNone);
+
+OP (69, MF(Immediate),       MF(ADC),    A,    NULL, NZCV, C,    OFNone);
+OP (65, MF(ZeroPage),        MF(ADC),    A,    NULL, NZCV, C,    OFNone);
+OP (75, MF(ZeroPage),        MF(ADC),    A,    X,    NZCV, C,    OFNone);
+OP (6D, MF(Absolute),        MF(ADC),    A,    NULL, NZCV, C,    OFNone);
+OP (7D, MF(Absolute),        MF(ADC),    A,    X,    NZCV, C,    OFNone);
+OP (79, MF(Absolute),        MF(ADC),    A,    Y,    NZCV, C,    OFNone);
+OP (61, MF(IndirectX),       MF(ADC),    A,    X,    NZCV, C,    OFNone);
+OP (71, MF(IndirectY),       MF(ADC),    A,    Y,    NZCV, C,    OFNone);
+OP (72, MF(Indirect),        MF(ADC),    A,    NULL, NZCV, C,    OFNone);
+OP (63, MF(ZeroPage),        MF(ADC),    A,    S,    NZCV, C,    OFNone);
+OP (73, MF(IndirectS),       MF(ADC),    A,    Y,    NZCV, C,    OFNone);
+OP (67, MF(IndirectFar),     MF(ADC),    A,    NULL, NZCV, C,    OFNone);
+OP (77, MF(IndirectFar),     MF(ADC),    A,    Y,    NZCV, C,    OFNone);
+OP (6F, MF(Long),            MF(ADC),    A,    NULL, NZCV, C,    OFNone);
+OP (7F, MF(Long),            MF(ADC),    A,    X,    NZCV, C,    OFNone);
+
+OP (E9, MF(Immediate),       MF(SBC),    A,    NULL, NZCV, C,    OFNone);
+OP (E5, MF(ZeroPage),        MF(SBC),    A,    NULL, NZCV, C,    OFNone);
+OP (F5, MF(ZeroPage),        MF(SBC),    A,    X,    NZCV, C,    OFNone);
+OP (ED, MF(Absolute),        MF(SBC),    A,    NULL, NZCV, C,    OFNone);
+OP (FD, MF(Absolute),        MF(SBC),    A,    X,    NZCV, C,    OFNone);
+OP (F9, MF(Absolute),        MF(SBC),    A,    Y,    NZCV, C,    OFNone);
+OP (E1, MF(IndirectX),       MF(SBC),    A,    X,    NZCV, C,    OFNone);
+OP (F1, MF(IndirectY),       MF(SBC),    A,    Y,    NZCV, C,    OFNone);
+OP (F2, MF(Indirect),        MF(SBC),    A,    NULL, NZCV, C,    OFNone);
+OP (E3, MF(ZeroPage),        MF(SBC),    A,    S,    NZCV, C,    OFNone);
+OP (F3, MF(IndirectS),       MF(SBC),    A,    Y,    NZCV, C,    OFNone);
+OP (E7, MF(IndirectFar),     MF(SBC),    A,    NULL, NZCV, C,    OFNone);
+OP (F7, MF(IndirectFar),     MF(SBC),    A,    Y,    NZCV, C,    OFNone);
+OP (EF, MF(Long),            MF(SBC),    A,    NULL, NZCV, C,    OFNone);
+OP (FF, MF(Long),            MF(SBC),    A,    X,    NZCV, C,    OFNone);
+
+OP (C9, MF(Immediate),       MF(CMP),    A,    NULL, NZC,  0,    OFNone);
+OP (C5, MF(ZeroPage),        MF(CMP),    A,    NULL, NZC,  0,    OFNone);
+OP (D5, MF(ZeroPage),        MF(CMP),    A,    X,    NZC,  0,    OFNone);
+OP (CD, MF(Absolute),        MF(CMP),    A,    NULL, NZC,  0,    OFNone);
+OP (DD, MF(Absolute),        MF(CMP),    A,    X,    NZC,  0,    OFNone);
+OP (D9, MF(Absolute),        MF(CMP),    A,    Y,    NZC,  0,    OFNone);
+OP (C1, MF(IndirectX),       MF(CMP),    A,    X,    NZC,  0,    OFNone);
+OP (D1, MF(IndirectY),       MF(CMP),    A,    Y,    NZC,  0,    OFNone);
+OP (D2, MF(Indirect),        MF(CMP),    A,    NULL, NZC,  0,    OFNone);
+OP (C3, MF(ZeroPage),        MF(CMP),    A,    S,    NZC,  0,    OFNone);
+OP (D3, MF(IndirectS),       MF(CMP),    A,    Y,    NZC,  0,    OFNone);
+OP (C7, MF(IndirectFar),     MF(CMP),    A,    NULL, NZC,  0,    OFNone);
+OP (D7, MF(IndirectFar),     MF(CMP),    A,    Y,    NZC,  0,    OFNone);
+OP (CF, MF(Long),            MF(CMP),    A,    NULL, NZC,  0,    OFNone);
+OP (DF, MF(Long),            MF(CMP),    A,    X,    NZC,  0,    OFNone);
+
+OP (E0, XF(Immediate),       XF(CMP),    X,    NULL, NZC,  0,    OFNone);
+OP (E4, XF(ZeroPage),        XF(CMP),    X,    NULL, NZC,  0,    OFNone);
+OP (EC, XF(Absolute),        XF(CMP),    X,    NULL, NZC,  0,    OFNone);
+
+OP (C0, XF(Immediate),       XF(CMP),    Y,    NULL, NZC,  0,    OFNone);
+OP (C4, XF(ZeroPage),        XF(CMP),    Y,    NULL, NZC,  0,    OFNone);
+OP (CC, XF(Absolute),        XF(CMP),    Y,    NULL, NZC,  0,    OFNone);
+
+OP (24, MF(ZeroPage),        MF(BIT),    A,    NULL, NZV,  0,    OFNone);
+OP (2C, MF(Absolute),        MF(BIT),    A,    NULL, NZV,  0,    OFNone);
+OP (34, MF(ZeroPage),        MF(BIT),    A,    X,    NZV,  0,    OFNone);
+OP (3C, MF(Absolute),        MF(BIT),    A,    X,    NZV,  0,    OFNone);
+OP (89, MF(Immediate),       MF(BIT),    A,    NULL, Z,    0,    OFNone);
+
+OP (E6, MF(ZeroPage),        MF(INC),    NULL, NULL, NZ,   0,    OFNone);
+OP (F6, MF(ZeroPage),        MF(INC),    NULL, X,    NZ,   0,    OFNone);
+OP (EE, MF(Absolute),        MF(INC),    NULL, NULL, NZ,   0,    OFNone);
+OP (FE, MF(Absolute),        MF(INC),    NULL, X,    NZ,   0,    OFNone);
+OP (E8, XF(Register),        XF(INC),    X,    NULL, NZ,   0,    OFNone);
+OP (C8, XF(Register),        XF(INC),    Y,    NULL, NZ,   0,    OFNone);
+OP (1A, MF(Register),        MF(INC),    A,    NULL, NZ,   0,    OFNone);
+
+OP (C6, MF(ZeroPage),        MF(DEC),    NULL, NULL, NZ,   0,    OFNone);
+OP (D6, MF(ZeroPage),        MF(DEC),    NULL, X,    NZ,   0,    OFNone);
+OP (CE, MF(Absolute),        MF(DEC),    NULL, NULL, NZ,   0,    OFNone);
+OP (DE, MF(Absolute),        MF(DEC),    NULL, X,    NZ,   0,    OFNone);
+OP (CA, XF(Register),        XF(DEC),    X,    NULL, NZ,   0,    OFNone);
+OP (88, XF(Register),        XF(DEC),    Y,    NULL, NZ,   0,    OFNone);
+OP (3A, MF(Register),        MF(DEC),    A,    NULL, NZ,   0,    OFNone);
+
+OP (04, MF(ZeroPage),        MF(TSB),    A,    NULL, Z,    0,    OFNone);
+OP (0C, MF(Absolute),        MF(TSB),    A,    NULL, Z,    0,    OFNone);
+OP (14, MF(ZeroPage),        MF(TRB),    A,    NULL, Z,    0,    OFNone);
+OP (1C, MF(Absolute),        MF(TRB),    A,    NULL, Z,    0,    OFNone);
+
+OP (0A, MF(Register),        MF(ASL),    A,    NULL, NZC,  0,    OFNone);
+OP (06, MF(ZeroPage),        MF(ASL),    NULL, NULL, NZC,  0,    OFNone);
+OP (16, MF(ZeroPage),        MF(ASL),    NULL, X,    NZC,  0,    OFNone);
+OP (0E, MF(Absolute),        MF(ASL),    NULL, NULL, NZC,  0,    OFNone);
+OP (1E, MF(Absolute),        MF(ASL),    NULL, X,    NZC,  0,    OFNone);
+
+OP (4A, MF(Register),        MF(LSR),    A,    NULL, NZC,  0,    OFNone);
+OP (46, MF(ZeroPage),        MF(LSR),    NULL, NULL, NZC,  0,    OFNone);
+OP (56, MF(ZeroPage),        MF(LSR),    NULL, X,    NZC,  0,    OFNone);
+OP (4E, MF(Absolute),        MF(LSR),    NULL, NULL, NZC,  0,    OFNone);
+OP (5E, MF(Absolute),        MF(LSR),    NULL, X,    NZC,  0,    OFNone);
+
+OP (2A, MF(Register),        MF(ROL),    A,    NULL, NZC,  C,    OFNone);
+OP (26, MF(ZeroPage),        MF(ROL),    NULL, NULL, NZC,  C,    OFNone);
+OP (36, MF(ZeroPage),        MF(ROL),    NULL, X,    NZC,  C,    OFNone);
+OP (2E, MF(Absolute),        MF(ROL),    NULL, NULL, NZC,  C,    OFNone);
+OP (3E, MF(Absolute),        MF(ROL),    NULL, X,    NZC,  C,    OFNone);
+
+OP (6A, MF(Register),        MF(ROR),    A,    NULL, NZC,  C,    OFNone);
+OP (66, MF(ZeroPage),        MF(ROR),    NULL, NULL, NZC,  C,    OFNone);
+OP (76, MF(ZeroPage),        MF(ROR),    NULL, X,    NZC,  C,    OFNone);
+OP (6E, MF(Absolute),        MF(ROR),    NULL, NULL, NZC,  C,    OFNone);
+OP (7E, MF(Absolute),        MF(ROR),    NULL, X,    NZC,  C,    OFNone);
+
+OP (80, NF8(Immediate),      NF(BRA),    NULL, NULL, 0,    NZCV, OFBreak);
+OP (82, NF16(Immediate),     NF(BRL),    NULL, NULL, 0,    NZCV, OFBreak);
+OP (4C, NF16(Immediate),     NF(JMP),    NULL, NULL, 0,    NZCV, OFBreak);
+OP (5C, NF(Immediate24),     NF(JMP),    NULL, NULL, 0,    NZCV, OFBreak | OFSectionUpdate);
+OP (6C, NF16(Absolute),      NF(JMP),    NULL, NULL, 0,    NZCV, OFBreak);
+OP (7C, NF16(Absolute),      NF(JMP),    NULL, X,    0,    NZCV, OFBreak);
+OP (DC, NF16(Absolute),      NF(JML),    NULL, NULL, 0,    NZCV, OFBreak | OFSectionUpdate);
+OP (20, NF16(Immediate),     NF(JSR),    NULL, NULL, 0,    NZCV, OFBreak);
+OP (22, NF(Immediate24),     NF(JSL),    NULL, NULL, 0,    NZCV, OFBreak | OFSectionUpdate);
+OP (FC, NF16(IndirectX),     NF(JSR),    NULL, X,    0,    NZCV, OFBreak);
+OP (40, NF(None),            NF(RTI),    NULL, NULL, NZCV, NZCV, OFBreak | OFSectionUpdate);
+OP (6B, NF(None),            NF(RTL),    NULL, NULL, 0,    NZCV, OFBreak | OFSectionUpdate);
+OP (60, NF(None),            NF(RTS),    NULL, NULL, 0,    NZCV, OFBreak);
+
+OP (10, NF8(Immediate),      NF(BPL),    NULL, NULL, 0,    NZCV, OFBreak);
+OP (30, NF8(Immediate),      NF(BMI),    NULL, NULL, 0,    NZCV, OFBreak);
+OP (50, NF8(Immediate),      NF(BVC),    NULL, NULL, 0,    NZCV, OFBreak);
+OP (70, NF8(Immediate),      NF(BVS),    NULL, NULL, 0,    NZCV, OFBreak);
+OP (90, NF8(Immediate),      NF(BCC),    NULL, NULL, 0,    NZCV, OFBreak);
+OP (B0, NF8(Immediate),      NF(BCS),    NULL, NULL, 0,    NZCV, OFBreak);
+OP (D0, NF8(Immediate),      NF(BZC),    NULL, NULL, 0,    NZCV, OFBreak);
+OP (F0, NF8(Immediate),      NF(BZS),    NULL, NULL, 0,    NZCV, OFBreak);
+
+OP (00, NF(None),            NF(BRK),    NULL, NULL, 0,    NZCV, OFBreak | OFSectionUpdate);
+OP (02, NF(None),            NF(COP),    NULL, NULL, 0,    NZCV, OFBreak | OFSectionUpdate);
+
+OP (18, NF(None),            NF(CLC),    NULL, NULL, C,    0,    OFNone);
+OP (58, NF(None),            NF(CLI),    NULL, NULL, 0,    0,    OFNone);
+OP (D8, NF(None),            NF(CLD),    NULL, NULL, 0,    NZCV, OFBreak | OFSectionUpdate);
+OP (B8, NF(None),            NF(CLV),    NULL, NULL, V,    0,    OFNone);
+OP (38, NF(None),            NF(SEC),    NULL, NULL, C,    0,    OFNone);
+OP (78, NF(None),            NF(SEI),    NULL, NULL, 0,    0,    OFNone);
+OP (F8, NF(None),            NF(SED),    NULL, NULL, 0,    NZCV, OFBreak | OFSectionUpdate);
+OP (C2, NF8(Immediate),      NF(REP),    NULL, NULL, NZCV, NZCV, OFBreak | OFSectionUpdate);
+OP (E2, NF8(Immediate),      NF(SEP),    NULL, NULL, NZCV, NZCV, OFBreak | OFSectionUpdate);
+OP (FB, NF(None),            NF(XCE),    NULL, NULL, C,    NZCV, OFBreak | OFSectionUpdate);
+
+OP (DB, NF(None),            NF(STP),    NULL, NULL, 0,    NZCV, OFBreak);
+OP (EB, NF(None),            NF(XBA),    NULL, NULL, NZ,   0,    OFNone);
+OP (CB, NF(None),            NF(WAI),    NULL, NULL, 0,    NZCV, OFBreak);
+OP (42, NF8(Immediate),      NF(WDM),    NULL, NULL, 0,    NZCV, OFBreak); /* Speedhacks, might change PC */
+OP (EA, NF(None),            NF(NOP),    NULL, NULL, 0,    0,    OFNone);
diff --git a/source/arm_dynarec/opgen.h b/source/arm_dynarec/opgen.h
new file mode 100644
index 0000000..aaa5556
--- /dev/null
+++ b/source/arm_dynarec/opgen.h
@@ -0,0 +1,26 @@
+/* Macros turn opcode tables into code */
+
+#define F8(F) F##8
+#define F16(F) F##16
+#define EF(F) 0, F16(F), F16(F)
+#define NF(F) 0, F, F
+#define NF8(F) 0, F8(F), F8(F)
+#define NF16(F) 0, F16(F), F16(F)
+#define MF(F) (CheckEmulation() || CheckMemory()), F8(F), F16(F)
+#define XF(F) (CheckEmulation() || CheckIndex()), F8(F), F16(F)
+
+#define C (Carry)
+#define Z (Zero)
+#define V (Overflow)
+#define NZ (Negative | Zero)
+#define NZC (Negative | Zero | Carry)
+#define NZV (Negative | Zero | Overflow)
+#define NZCV (Negative | Zero | Carry | Overflow)
+
+switch(opcode = *pc++) {
+
+#include "opdef.h"
+
+default:                                              \
+   printf("Invalid opcode : 0x%X\n", opcode);         \
+}
diff --git a/source/cpuexec.c b/source/cpuexec.c
index 2bbf90a..15c9157 100644
--- a/source/cpuexec.c
+++ b/source/cpuexec.c
@@ -11,6 +11,7 @@
 #include "fxemu.h"
 #include "sa1.h"
 #include "spc7110.h"
+#include "arm_dynarec/dynaexec.h"
 
 void S9xMainLoop_SA1_SFX(void);
 void S9xMainLoop_SA1_NoSFX(void);
@@ -28,6 +29,13 @@ void S9xMainLoop_NoSA1_NoSFX(void);
  */
 void S9xMainLoop()
 {
+#ifdef USE_DYNAREC
+   if (Settings.EnableDynarec)
+   {
+      S9xMainLoop_Dyna();
+   }
+   else
+#endif
    if (Settings.SA1)
    {
       if (Settings.SuperFX)
diff --git a/source/globals.c b/source/globals.c
index f7cdb85..3ccbd96 100644
--- a/source/globals.c
+++ b/source/globals.c
@@ -16,8 +16,10 @@
 
 char String[513];
 
+#ifndef USE_DYNAREC
 SICPU ICPU;
 SCPUState CPU;
+#endif
 
 #ifndef USE_BLARGG_APU
 SAPU APU;
@@ -35,9 +37,11 @@ SnesModel* Model = &M1SNES;
 
 int32_t OpAddress = 0;
 
+#ifndef USE_DYNAREC
 CMemory Memory;
 
 uint8_t OpenBus = 0;
+#endif
 
 FxInit_s SuperFX;
 
diff --git a/source/ppu.c b/source/ppu.c
index 1f5df0b..caf47ed 100644
--- a/source/ppu.c
+++ b/source/ppu.c
@@ -629,6 +629,11 @@ void S9xSetPPU(uint8_t Byte, uint16_t Address)
       case 0x217e:
       case 0x217f:
 #ifndef USE_BLARGG_APU
+#ifdef USE_DYNAREC
+         if (Settings.EnableDynarec)
+            while (APU.Cycles <= CPU.Cycles)
+               APU_EXECUTE1();
+#endif
          Memory.FillRAM [Address] = Byte;
          IAPU.RAM [(Address & 3) + 0xf4] = Byte;
          IAPU.APUExecuting = Settings.APUEnabled;
@@ -889,6 +894,10 @@ uint8_t S9xGetPPU(uint16_t Address)
       case 0x217e:
       case 0x217f:
 #ifndef USE_BLARGG_APU
+#ifdef USE_DYNAREC
+         if (Settings.EnableDynarec)
+            APU_EXECUTE();
+#endif
          IAPU.APUExecuting = Settings.APUEnabled;
          IAPU.WaitCounter++;
 
diff --git a/source/snes9x.h b/source/snes9x.h
index 6dd2e8f..9ce602a 100644
--- a/source/snes9x.h
+++ b/source/snes9x.h
@@ -214,6 +214,7 @@ typedef struct
    bool     SecondJustifier;
    int8_t   SETA;
    bool     HardDisableAudio;
+   bool     EnableDynarec;
 } SSettings;
 
 extern SSettings Settings;