1.1 --- /dev/null	Thu Jan 01 00:00:00 1970 +0000
     1.2 +++ b/src/gba/arm-new.h	Sat Mar 03 10:31:27 2012 -0600
     1.3 @@ -0,0 +1,7453 @@
     1.4 +#ifdef BKPT_SUPPORT
     1.5 +#define CONSOLE_OUTPUT(a, b) \
     1.6 +    extern void (*dbgOutput)(char *, u32); \
     1.7 +    if ((opcode == 0xe0000000) && (reg[0].I == 0xC0DED00D)) { \
     1.8 +		dbgOutput((a), (b)); \
     1.9 +	}
    1.10 +#else
    1.11 +#define CONSOLE_OUTPUT(a, b)
    1.12 +#endif
    1.13 +
    1.14 +#define OP_AND \
    1.15 +    reg[dest].I = reg[(opcode >> 16) & 15].I & value; \
    1.16 +    CONSOLE_OUTPUT(NULL, reg[2].I);
    1.17 +
    1.18 +#define OP_ANDS \
    1.19 +    reg[dest].I = reg[(opcode >> 16) & 15].I & value; \
    1.20 +      \
    1.21 +    N_FLAG = (reg[dest].I & 0x80000000) ? true : false; \
    1.22 +    Z_FLAG = (reg[dest].I) ? false : true; \
    1.23 +    C_FLAG = C_OUT;
    1.24 +
    1.25 +#define OP_EOR \
    1.26 +    reg[dest].I = reg[(opcode >> 16) & 15].I ^ value;
    1.27 +
    1.28 +#define OP_EORS \
    1.29 +    reg[dest].I = reg[(opcode >> 16) & 15].I ^ value; \
    1.30 +      \
    1.31 +    N_FLAG = (reg[dest].I & 0x80000000) ? true : false; \
    1.32 +    Z_FLAG = (reg[dest].I) ? false : true; \
    1.33 +    C_FLAG = C_OUT;
    1.34 +#ifdef C_CORE
    1.35 +#define NEG(i) ((i) >> 31)
    1.36 +#define POS(i) ((~(i)) >> 31)
    1.37 +#define ADDCARRY(a, b, c) \
    1.38 +    C_FLAG = ((NEG(a) & NEG(b)) | \
    1.39 +              (NEG(a) & POS(c)) | \
    1.40 +              (NEG(b) & POS(c))) ? true : false;
    1.41 +#define ADDOVERFLOW(a, b, c) \
    1.42 +    V_FLAG = ((NEG(a) & NEG(b) & POS(c)) | \
    1.43 +              (POS(a) & POS(b) & NEG(c))) ? true : false;
    1.44 +#define SUBCARRY(a, b, c) \
    1.45 +    C_FLAG = ((NEG(a) & POS(b)) | \
    1.46 +              (NEG(a) & POS(c)) | \
    1.47 +              (POS(b) & POS(c))) ? true : false;
    1.48 +#define SUBOVERFLOW(a, b, c) \
    1.49 +    V_FLAG = ((NEG(a) & POS(b) & POS(c)) | \
    1.50 +              (POS(a) & NEG(b) & NEG(c))) ? true : false;
    1.51 +#define OP_SUB \
    1.52 +	{ \
    1.53 +		reg[dest].I = reg[base].I - value; \
    1.54 +	}
    1.55 +#define OP_SUBS \
    1.56 +	{ \
    1.57 +		u32 lhs = reg[base].I; \
    1.58 +		u32 rhs = value; \
    1.59 +		u32 res = lhs - rhs; \
    1.60 +		reg[dest].I = res; \
    1.61 +		Z_FLAG		= (res == 0) ? true : false; \
    1.62 +		N_FLAG		= NEG(res) ? true : false; \
    1.63 +		SUBCARRY(lhs, rhs, res); \
    1.64 +		SUBOVERFLOW(lhs, rhs, res); \
    1.65 +	}
    1.66 +#define OP_RSB \
    1.67 +	{ \
    1.68 +		reg[dest].I = value - reg[base].I; \
    1.69 +	}
    1.70 +#define OP_RSBS \
    1.71 +	{ \
    1.72 +		u32 lhs = reg[base].I; \
    1.73 +		u32 rhs = value; \
    1.74 +		u32 res = rhs - lhs; \
    1.75 +		reg[dest].I = res; \
    1.76 +		Z_FLAG		= (res == 0) ? true : false; \
    1.77 +		N_FLAG		= NEG(res) ? true : false; \
    1.78 +		SUBCARRY(rhs, lhs, res); \
    1.79 +		SUBOVERFLOW(rhs, lhs, res); \
    1.80 +	}
    1.81 +#define OP_ADD \
    1.82 +	{ \
    1.83 +		reg[dest].I = reg[base].I + value; \
    1.84 +	}
    1.85 +#define OP_ADDS \
    1.86 +	{ \
    1.87 +		u32 lhs = reg[base].I; \
    1.88 +		u32 rhs = value; \
    1.89 +		u32 res = lhs + rhs; \
    1.90 +		reg[dest].I = res; \
    1.91 +		Z_FLAG		= (res == 0) ? true : false; \
    1.92 +		N_FLAG		= NEG(res) ? true : false; \
    1.93 +		ADDCARRY(lhs, rhs, res); \
    1.94 +		ADDOVERFLOW(lhs, rhs, res); \
    1.95 +	}
    1.96 +#define OP_ADC \
    1.97 +	{ \
    1.98 +		reg[dest].I = reg[base].I + value + (u32)C_FLAG; \
    1.99 +	}
   1.100 +#define OP_ADCS \
   1.101 +	{ \
   1.102 +		u32 lhs = reg[base].I; \
   1.103 +		u32 rhs = value; \
   1.104 +		u32 res = lhs + rhs + (u32)C_FLAG; \
   1.105 +		reg[dest].I = res; \
   1.106 +		Z_FLAG		= (res == 0) ? true : false; \
   1.107 +		N_FLAG		= NEG(res) ? true : false; \
   1.108 +		ADDCARRY(lhs, rhs, res); \
   1.109 +		ADDOVERFLOW(lhs, rhs, res); \
   1.110 +	}
   1.111 +#define OP_SBC \
   1.112 +	{ \
   1.113 +		reg[dest].I = reg[base].I - value - !((u32)C_FLAG); \
   1.114 +	}
   1.115 +#define OP_SBCS \
   1.116 +	{ \
   1.117 +		u32 lhs = reg[base].I; \
   1.118 +		u32 rhs = value; \
   1.119 +		u32 res = lhs - rhs - !((u32)C_FLAG); \
   1.120 +		reg[dest].I = res; \
   1.121 +		Z_FLAG		= (res == 0) ? true : false; \
   1.122 +		N_FLAG		= NEG(res) ? true : false; \
   1.123 +		SUBCARRY(lhs, rhs, res); \
   1.124 +		SUBOVERFLOW(lhs, rhs, res); \
   1.125 +	}
   1.126 +#define OP_RSC \
   1.127 +	{ \
   1.128 +		reg[dest].I = value - reg[base].I - !((u32)C_FLAG); \
   1.129 +	}
   1.130 +#define OP_RSCS \
   1.131 +	{ \
   1.132 +		u32 lhs = reg[base].I; \
   1.133 +		u32 rhs = value; \
   1.134 +		u32 res = rhs - lhs - !((u32)C_FLAG); \
   1.135 +		reg[dest].I = res; \
   1.136 +		Z_FLAG		= (res == 0) ? true : false; \
   1.137 +		N_FLAG		= NEG(res) ? true : false; \
   1.138 +		SUBCARRY(rhs, lhs, res); \
   1.139 +		SUBOVERFLOW(rhs, lhs, res); \
   1.140 +	}
   1.141 +#define OP_CMP \
   1.142 +	{ \
   1.143 +		u32 lhs = reg[base].I; \
   1.144 +		u32 rhs = value; \
   1.145 +		u32 res = lhs - rhs; \
   1.146 +		Z_FLAG = (res == 0) ? true : false; \
   1.147 +		N_FLAG = NEG(res) ? true : false; \
   1.148 +		SUBCARRY(lhs, rhs, res); \
   1.149 +		SUBOVERFLOW(lhs, rhs, res); \
   1.150 +	}
   1.151 +#define OP_CMN \
   1.152 +	{ \
   1.153 +		u32 lhs = reg[base].I; \
   1.154 +		u32 rhs = value; \
   1.155 +		u32 res = lhs + rhs; \
   1.156 +		Z_FLAG = (res == 0) ? true : false; \
   1.157 +		N_FLAG = NEG(res) ? true : false; \
   1.158 +		ADDCARRY(lhs, rhs, res); \
   1.159 +		ADDOVERFLOW(lhs, rhs, res); \
   1.160 +	}
   1.161 +
   1.162 +#define LOGICAL_LSL_REG \
   1.163 +	{ \
   1.164 +		u32 v = reg[opcode & 0x0f].I; \
   1.165 +		C_OUT = (v >> (32 - shift)) & 1 ? true : false; \
   1.166 +		value = v << shift; \
   1.167 +	}
   1.168 +#define LOGICAL_LSR_REG \
   1.169 +	{ \
   1.170 +		u32 v = reg[opcode & 0x0f].I; \
   1.171 +		C_OUT = (v >> (shift - 1)) & 1 ? true : false; \
   1.172 +		value = v >> shift; \
   1.173 +	}
   1.174 +#define LOGICAL_ASR_REG \
   1.175 +	{ \
   1.176 +		u32 v = reg[opcode & 0x0f].I; \
   1.177 +		C_OUT = ((s32)v >> (int)(shift - 1)) & 1 ? true : false; \
   1.178 +		value = (s32)v >> (int)shift; \
   1.179 +	}
   1.180 +#define LOGICAL_ROR_REG \
   1.181 +	{ \
   1.182 +		u32 v = reg[opcode & 0x0f].I; \
   1.183 +		C_OUT = (v >> (shift - 1)) & 1 ? true : false; \
   1.184 +		value = ((v << (32 - shift)) | \
   1.185 +		         (v >> shift)); \
   1.186 +	}
   1.187 +#define LOGICAL_RRX_REG \
   1.188 +	{ \
   1.189 +		u32 v = reg[opcode & 0x0f].I; \
   1.190 +		shift = (int)C_FLAG; \
   1.191 +		C_OUT = (v  & 1) ? true : false; \
   1.192 +		value = ((v >> 1) | \
   1.193 +		         (shift << 31)); \
   1.194 +	}
   1.195 +#define LOGICAL_ROR_IMM \
   1.196 +	{ \
   1.197 +		u32 v = opcode & 0xff; \
   1.198 +		C_OUT = (v >> (shift - 1)) & 1 ? true : false; \
   1.199 +		value = ((v << (32 - shift)) | \
   1.200 +		         (v >> shift)); \
   1.201 +	}
   1.202 +#define ARITHMETIC_LSL_REG \
   1.203 +	{ \
   1.204 +		u32 v = reg[opcode & 0x0f].I; \
   1.205 +		value = v << shift; \
   1.206 +	}
   1.207 +#define ARITHMETIC_LSR_REG \
   1.208 +	{ \
   1.209 +		u32 v = reg[opcode & 0x0f].I; \
   1.210 +		value = v >> shift; \
   1.211 +	}
   1.212 +#define ARITHMETIC_ASR_REG \
   1.213 +	{ \
   1.214 +		u32 v = reg[opcode & 0x0f].I; \
   1.215 +		value = (s32)v >> (int)shift; \
   1.216 +	}
   1.217 +#define ARITHMETIC_ROR_REG \
   1.218 +	{ \
   1.219 +		u32 v = reg[opcode & 0x0f].I; \
   1.220 +		value = ((v << (32 - shift)) | \
   1.221 +		         (v >> shift)); \
   1.222 +	}
   1.223 +#define ARITHMETIC_RRX_REG \
   1.224 +	{ \
   1.225 +		u32 v = reg[opcode & 0x0f].I; \
   1.226 +		shift = (int)C_FLAG; \
   1.227 +		value = ((v >> 1) | \
   1.228 +		         (shift << 31)); \
   1.229 +	}
   1.230 +#define ARITHMETIC_ROR_IMM \
   1.231 +	{ \
   1.232 +		u32 v = opcode & 0xff; \
   1.233 +		value = ((v << (32 - shift)) | \
   1.234 +		         (v >> shift)); \
   1.235 +	}
   1.236 +#define ROR_IMM_MSR \
   1.237 +	{ \
   1.238 +		u32 v = opcode & 0xff; \
   1.239 +		value = ((v << (32 - shift)) | \
   1.240 +		         (v >> shift)); \
   1.241 +	}
   1.242 +#define ROR_VALUE \
   1.243 +	{ \
   1.244 +		value = ((value << (32 - shift)) | \
   1.245 +		         (value >> shift)); \
   1.246 +	}
   1.247 +#define RCR_VALUE \
   1.248 +	{ \
   1.249 +		shift = (int)C_FLAG; \
   1.250 +		value = ((value >> 1) | \
   1.251 +		         (shift << 31)); \
   1.252 +	}
   1.253 +#else
   1.254 +#ifdef __GNUC__
   1.255 +		#ifdef __POWERPC__
   1.256 +			#define OP_SUB \
   1.257 +	{ \
   1.258 +		reg[dest].I = reg[base].I - value; \
   1.259 +	}
   1.260 +			#define OP_SUBS \
   1.261 +	{ \
   1.262 +		register int Flags;                             \
   1.263 +		register int Result;                            \
   1.264 +		asm volatile ("subco. %0, %2, %3\n"              \
   1.265 +		              "mcrxr cr1\n"                       \
   1.266 +		              "mfcr %1\n"                         \
   1.267 +					  : "=r" (Result),                    \
   1.268 +		              "=r" (Flags)                      \
   1.269 +					  : "r" (reg[base].I),                \
   1.270 +		              "r" (value)                       \
   1.271 +		              );                                  \
   1.272 +		reg[dest].I = Result;                           \
   1.273 +		Z_FLAG		= (Flags >> 29) & 1;                     \
   1.274 +		N_FLAG		= (Flags >> 31) & 1;                     \
   1.275 +		C_FLAG		= (Flags >> 25) & 1;                     \
   1.276 +		V_FLAG		= (Flags >> 26) & 1;                     \
   1.277 +	}
   1.278 +			#define OP_RSB \
   1.279 +	{ \
   1.280 +		reg[dest].I = value - reg[base].I; \
   1.281 +	}
   1.282 +			#define OP_RSBS \
   1.283 +	{ \
   1.284 +		register int Flags;                             \
   1.285 +		register int Result;                            \
   1.286 +		asm volatile ("subfco. %0, %2, %3\n"             \
   1.287 +		              "mcrxr cr1\n"                       \
   1.288 +		              "mfcr %1\n"                         \
   1.289 +					  : "=r" (Result),                    \
   1.290 +		              "=r" (Flags)                      \
   1.291 +					  : "r" (reg[base].I),                \
   1.292 +		              "r" (value)                       \
   1.293 +		              );                                  \
   1.294 +		reg[dest].I = Result;                           \
   1.295 +		Z_FLAG		= (Flags >> 29) & 1;                     \
   1.296 +		N_FLAG		= (Flags >> 31) & 1;                     \
   1.297 +		C_FLAG		= (Flags >> 25) & 1;                     \
   1.298 +		V_FLAG		= (Flags >> 26) & 1;                     \
   1.299 +	}
   1.300 +			#define OP_ADD \
   1.301 +	{ \
   1.302 +		reg[dest].I = reg[base].I + value; \
   1.303 +	}
   1.304 +
   1.305 +			#define OP_ADDS \
   1.306 +	{ \
   1.307 +		register int Flags;                             \
   1.308 +		register int Result;                            \
   1.309 +		asm volatile ("addco. %0, %2, %3\n"              \
   1.310 +		              "mcrxr cr1\n"                       \
   1.311 +		              "mfcr %1\n"                         \
   1.312 +					  : "=r" (Result),                    \
   1.313 +		              "=r" (Flags)                      \
   1.314 +					  : "r" (reg[base].I),                \
   1.315 +		              "r" (value)                       \
   1.316 +		              );                                  \
   1.317 +		reg[dest].I = Result;                           \
   1.318 +		Z_FLAG		= (Flags >> 29) & 1;                     \
   1.319 +		N_FLAG		= (Flags >> 31) & 1;                     \
   1.320 +		C_FLAG		= (Flags >> 25) & 1;                     \
   1.321 +		V_FLAG		= (Flags >> 26) & 1;                     \
   1.322 +	}
   1.323 +			#define OP_ADC \
   1.324 +	{ \
   1.325 +		reg[dest].I = reg[base].I + value + (u32)C_FLAG; \
   1.326 +	}
   1.327 +			#define OP_ADCS \
   1.328 +	{ \
   1.329 +		register int Flags;                             \
   1.330 +		register int Result;                            \
   1.331 +		asm volatile ("mtspr xer, %4\n"                  \
   1.332 +		              "addeo. %0, %2, %3\n"              \
   1.333 +		              "mcrxr cr1\n"                      \
   1.334 +		              "mfcr      %1\n"                   \
   1.335 +					  : "=r" (Result),                   \
   1.336 +		              "=r" (Flags)                     \
   1.337 +					  : "r" (reg[base].I),               \
   1.338 +		              "r" (value),                     \
   1.339 +		              "r" (C_FLAG << 29)               \
   1.340 +		              );                                 \
   1.341 +		reg[dest].I = Result;                           \
   1.342 +		Z_FLAG		= (Flags >> 29) & 1;                     \
   1.343 +		N_FLAG		= (Flags >> 31) & 1;                     \
   1.344 +		C_FLAG		= (Flags >> 25) & 1;                     \
   1.345 +		V_FLAG		= (Flags >> 26) & 1;                     \
   1.346 +	}
   1.347 +			#define OP_SBC \
   1.348 +	{ \
   1.349 +		reg[dest].I = reg[base].I - value - (C_FLAG ^ 1); \
   1.350 +	}
   1.351 +			#define OP_SBCS \
   1.352 +	{ \
   1.353 +		register int Flags;                             \
   1.354 +		register int Result;                            \
   1.355 +		asm volatile ("mtspr xer, %4\n"                  \
   1.356 +		              "subfeo. %0, %3, %2\n"             \
   1.357 +		              "mcrxr cr1\n"                      \
   1.358 +		              "mfcr      %1\n"                   \
   1.359 +					  : "=r" (Result),                   \
   1.360 +		              "=r" (Flags)                     \
   1.361 +					  : "r" (reg[base].I),               \
   1.362 +		              "r" (value),                     \
   1.363 +		              "r" (C_FLAG << 29)               \
   1.364 +		              );                                 \
   1.365 +		reg[dest].I = Result;                           \
   1.366 +		Z_FLAG		= (Flags >> 29) & 1;                     \
   1.367 +		N_FLAG		= (Flags >> 31) & 1;                     \
   1.368 +		C_FLAG		= (Flags >> 25) & 1;                     \
   1.369 +		V_FLAG		= (Flags >> 26) & 1;                     \
   1.370 +	}
   1.371 +			#define OP_RSC \
   1.372 +	{ \
   1.373 +		reg[dest].I = value - reg[base].I - (C_FLAG ^ 1); \
   1.374 +	}
   1.375 +			#define OP_RSCS \
   1.376 +	{ \
   1.377 +		register int Flags;                             \
   1.378 +		register int Result;                            \
   1.379 +		asm volatile ("mtspr xer, %4\n"                  \
   1.380 +		              "subfeo. %0, %2, %3\n"             \
   1.381 +		              "mcrxr cr1\n"                      \
   1.382 +		              "mfcr      %1\n"                   \
   1.383 +					  : "=r" (Result),                   \
   1.384 +		              "=r" (Flags)                     \
   1.385 +					  : "r" (reg[base].I),               \
   1.386 +		              "r" (value),                     \
   1.387 +		              "r" (C_FLAG << 29)               \
   1.388 +		              );                                 \
   1.389 +		reg[dest].I = Result;                           \
   1.390 +		Z_FLAG		= (Flags >> 29) & 1;                     \
   1.391 +		N_FLAG		= (Flags >> 31) & 1;                     \
   1.392 +		C_FLAG		= (Flags >> 25) & 1;                     \
   1.393 +		V_FLAG		= (Flags >> 26) & 1;                     \
   1.394 +	}
   1.395 +			#define OP_CMP \
   1.396 +	{ \
   1.397 +		register int Flags;                             \
   1.398 +		register int Result;                            \
   1.399 +		asm volatile ("subco. %0, %2, %3\n"              \
   1.400 +		              "mcrxr cr1\n"                       \
   1.401 +		              "mfcr %1\n"                         \
   1.402 +					  : "=r" (Result),                    \
   1.403 +		              "=r" (Flags)                      \
   1.404 +					  : "r" (reg[base].I),                \
   1.405 +		              "r" (value)                       \
   1.406 +		              );                                  \
   1.407 +		Z_FLAG = (Flags >> 29) & 1;                     \
   1.408 +		N_FLAG = (Flags >> 31) & 1;                     \
   1.409 +		C_FLAG = (Flags >> 25) & 1;                     \
   1.410 +		V_FLAG = (Flags >> 26) & 1;                     \
   1.411 +	}
   1.412 +			#define OP_CMN \
   1.413 +	{ \
   1.414 +		register int Flags;                             \
   1.415 +		register int Result;                            \
   1.416 +		asm volatile ("addco. %0, %2, %3\n"              \
   1.417 +		              "mcrxr cr1\n"                       \
   1.418 +		              "mfcr %1\n"                         \
   1.419 +					  : "=r" (Result),                    \
   1.420 +		              "=r" (Flags)                      \
   1.421 +					  : "r" (reg[base].I),                \
   1.422 +		              "r" (value)                       \
   1.423 +		              );                                  \
   1.424 +		Z_FLAG = (Flags >> 29) & 1;                     \
   1.425 +		N_FLAG = (Flags >> 31) & 1;                     \
   1.426 +		C_FLAG = (Flags >> 25) & 1;                     \
   1.427 +		V_FLAG = (Flags >> 26) & 1;                     \
   1.428 +	}
   1.429 +
   1.430 +			#define LOGICAL_LSL_REG \
   1.431 +	{ \
   1.432 +		u32 v = reg[opcode & 0x0f].I; \
   1.433 +		C_OUT = (v >> (32 - shift)) & 1 ? true : false; \
   1.434 +		value = v << shift; \
   1.435 +	}
   1.436 +			#define LOGICAL_LSR_REG \
   1.437 +	{ \
   1.438 +		u32 v = reg[opcode & 0x0f].I; \
   1.439 +		C_OUT = (v >> (shift - 1)) & 1 ? true : false; \
   1.440 +		value = v >> shift; \
   1.441 +	}
   1.442 +			#define LOGICAL_ASR_REG \
   1.443 +	{ \
   1.444 +		u32 v = reg[opcode & 0x0f].I; \
   1.445 +		C_OUT = ((s32)v >> (int)(shift - 1)) & 1 ? true : false; \
   1.446 +		value = (s32)v >> (int)shift; \
   1.447 +	}
   1.448 +			#define LOGICAL_ROR_REG \
   1.449 +	{ \
   1.450 +		u32 v = reg[opcode & 0x0f].I; \
   1.451 +		C_OUT = (v >> (shift - 1)) & 1 ? true : false; \
   1.452 +		value = ((v << (32 - shift)) | \
   1.453 +		         (v >> shift)); \
   1.454 +	}
   1.455 +			#define LOGICAL_RRX_REG \
   1.456 +	{ \
   1.457 +		u32 v = reg[opcode & 0x0f].I; \
   1.458 +		shift = (int)C_FLAG; \
   1.459 +		C_OUT = (v  & 1) ? true : false; \
   1.460 +		value = ((v >> 1) | \
   1.461 +		         (shift << 31)); \
   1.462 +	}
   1.463 +			#define LOGICAL_ROR_IMM \
   1.464 +	{ \
   1.465 +		u32 v = opcode & 0xff; \
   1.466 +		C_OUT = (v >> (shift - 1)) & 1 ? true : false; \
   1.467 +		value = ((v << (32 - shift)) | \
   1.468 +		         (v >> shift)); \
   1.469 +	}
   1.470 +			#define ARITHMETIC_LSL_REG \
   1.471 +	{ \
   1.472 +		u32 v = reg[opcode & 0x0f].I; \
   1.473 +		value = v << shift; \
   1.474 +	}
   1.475 +			#define ARITHMETIC_LSR_REG \
   1.476 +	{ \
   1.477 +		u32 v = reg[opcode & 0x0f].I; \
   1.478 +		value = v >> shift; \
   1.479 +	}
   1.480 +			#define ARITHMETIC_ASR_REG \
   1.481 +	{ \
   1.482 +		u32 v = reg[opcode & 0x0f].I; \
   1.483 +		value = (s32)v >> (int)shift; \
   1.484 +	}
   1.485 +			#define ARITHMETIC_ROR_REG \
   1.486 +	{ \
   1.487 +		u32 v = reg[opcode & 0x0f].I; \
   1.488 +		value = ((v << (32 - shift)) | \
   1.489 +		         (v >> shift)); \
   1.490 +	}
   1.491 +			#define ARITHMETIC_RRX_REG \
   1.492 +	{ \
   1.493 +		u32 v = reg[opcode & 0x0f].I; \
   1.494 +		shift = (int)C_FLAG; \
   1.495 +		value = ((v >> 1) | \
   1.496 +		         (shift << 31)); \
   1.497 +	}
   1.498 +			#define ARITHMETIC_ROR_IMM \
   1.499 +	{ \
   1.500 +		u32 v = opcode & 0xff; \
   1.501 +		value = ((v << (32 - shift)) | \
   1.502 +		         (v >> shift)); \
   1.503 +	}
   1.504 +			#define ROR_IMM_MSR \
   1.505 +	{ \
   1.506 +		u32 v = opcode & 0xff; \
   1.507 +		value = ((v << (32 - shift)) | \
   1.508 +		         (v >> shift)); \
   1.509 +	}
   1.510 +			#define ROR_VALUE \
   1.511 +	{ \
   1.512 +		value = ((value << (32 - shift)) | \
   1.513 +		         (value >> shift)); \
   1.514 +	}
   1.515 +			#define RCR_VALUE \
   1.516 +	{ \
   1.517 +		shift = (int)C_FLAG; \
   1.518 +		value = ((value >> 1) | \
   1.519 +		         (shift << 31)); \
   1.520 +	}
   1.521 +#else
   1.522 +#define OP_SUB \
   1.523 +    asm ("sub %1, %%ebx;" \
   1.524 +		 : "=b" (reg[dest].I) \
   1.525 +		 : "r" (value), "b" (reg[base].I));
   1.526 +
   1.527 +#define OP_SUBS \
   1.528 +    asm ("sub %1, %%ebx;" \
   1.529 +         "setsb N_FLAG;" \
   1.530 +         "setzb Z_FLAG;" \
   1.531 +         "setncb C_FLAG;" \
   1.532 +         "setob V_FLAG;" \
   1.533 +		 : "=b" (reg[dest].I) \
   1.534 +		 : "r" (value), "b" (reg[base].I));
   1.535 +
   1.536 +#define OP_RSB \
   1.537 +    asm  ("sub %1, %%ebx;" \
   1.538 +		  : "=b" (reg[dest].I) \
   1.539 +		  : "r" (reg[base].I), "b" (value));
   1.540 +
   1.541 +#define OP_RSBS \
   1.542 +    asm  ("sub %1, %%ebx;" \
   1.543 +          "setsb N_FLAG;" \
   1.544 +          "setzb Z_FLAG;" \
   1.545 +          "setncb C_FLAG;" \
   1.546 +          "setob V_FLAG;" \
   1.547 +		  : "=b" (reg[dest].I) \
   1.548 +		  : "r" (reg[base].I), "b" (value));
   1.549 +
   1.550 +#define OP_ADD \
   1.551 +    asm  ("add %1, %%ebx;" \
   1.552 +		  : "=b" (reg[dest].I) \
   1.553 +		  : "r" (value), "b" (reg[base].I));
   1.554 +
   1.555 +#define OP_ADDS \
   1.556 +    asm  ("add %1, %%ebx;" \
   1.557 +          "setsb N_FLAG;" \
   1.558 +          "setzb Z_FLAG;" \
   1.559 +          "setcb C_FLAG;" \
   1.560 +          "setob V_FLAG;" \
   1.561 +		  : "=b" (reg[dest].I) \
   1.562 +		  : "r" (value), "b" (reg[base].I));
   1.563 +
   1.564 +#define OP_ADC \
   1.565 +    asm  ("bt $0, C_FLAG;" \
   1.566 +          "adc %1, %%ebx;" \
   1.567 +		  : "=b" (reg[dest].I) \
   1.568 +		  : "r" (value), "b" (reg[base].I));
   1.569 +
   1.570 +#define OP_ADCS \
   1.571 +    asm  ("bt $0, C_FLAG;" \
   1.572 +          "adc %1, %%ebx;" \
   1.573 +          "setsb N_FLAG;" \
   1.574 +          "setzb Z_FLAG;" \
   1.575 +          "setcb C_FLAG;" \
   1.576 +          "setob V_FLAG;" \
   1.577 +		  : "=b" (reg[dest].I) \
   1.578 +		  : "r" (value), "b" (reg[base].I));
   1.579 +
   1.580 +#define OP_SBC \
   1.581 +    asm  ("bt $0, C_FLAG;" \
   1.582 +          "cmc;" \
   1.583 +          "sbb %1, %%ebx;" \
   1.584 +		  : "=b" (reg[dest].I) \
   1.585 +		  : "r" (value), "b" (reg[base].I));
   1.586 +
   1.587 +#define OP_SBCS \
   1.588 +    asm  ("bt $0, C_FLAG;" \
   1.589 +          "cmc;" \
   1.590 +          "sbb %1, %%ebx;" \
   1.591 +          "setsb N_FLAG;" \
   1.592 +          "setzb Z_FLAG;" \
   1.593 +          "setncb C_FLAG;" \
   1.594 +          "setob V_FLAG;" \
   1.595 +		  : "=b" (reg[dest].I) \
   1.596 +		  : "r" (value), "b" (reg[base].I));
   1.597 +#define OP_RSC \
   1.598 +    asm  ("bt $0, C_FLAG;" \
   1.599 +          "cmc;" \
   1.600 +          "sbb %1, %%ebx;" \
   1.601 +		  : "=b" (reg[dest].I) \
   1.602 +		  : "r" (reg[base].I), "b" (value));
   1.603 +
   1.604 +#define OP_RSCS \
   1.605 +    asm  ("bt $0, C_FLAG;" \
   1.606 +          "cmc;" \
   1.607 +          "sbb %1, %%ebx;" \
   1.608 +          "setsb N_FLAG;" \
   1.609 +          "setzb Z_FLAG;" \
   1.610 +          "setncb C_FLAG;" \
   1.611 +          "setob V_FLAG;" \
   1.612 +		  : "=b" (reg[dest].I) \
   1.613 +		  : "r" (reg[base].I), "b" (value));
   1.614 +#define OP_CMP \
   1.615 +    asm  ("sub %0, %1;" \
   1.616 +          "setsb N_FLAG;" \
   1.617 +          "setzb Z_FLAG;" \
   1.618 +          "setncb C_FLAG;" \
   1.619 +          "setob V_FLAG;" \
   1.620 +		  : \
   1.621 +		  : "r" (value), "r" (reg[base].I));
   1.622 +
   1.623 +#define OP_CMN \
   1.624 +    asm  ("add %0, %1;" \
   1.625 +          "setsb N_FLAG;" \
   1.626 +          "setzb Z_FLAG;" \
   1.627 +          "setcb C_FLAG;" \
   1.628 +          "setob V_FLAG;" \
   1.629 +		  : \
   1.630 +		  : "r" (value), "r" (reg[base].I));
   1.631 +#define LOGICAL_LSL_REG \
   1.632 +    asm ("shl %%cl, %%eax;" \
   1.633 +         "setcb %%cl;" \
   1.634 +		 : "=a" (value), "=c" (C_OUT) \
   1.635 +		 : "a" (reg[opcode & 0x0f].I), "c" (shift));
   1.636 +
   1.637 +#define LOGICAL_LSR_REG \
   1.638 +    asm ("shr %%cl, %%eax;" \
   1.639 +         "setcb %%cl;" \
   1.640 +		 : "=a" (value), "=c" (C_OUT) \
   1.641 +		 : "a" (reg[opcode & 0x0f].I), "c" (shift));
   1.642 +
   1.643 +#define LOGICAL_ASR_REG \
   1.644 +    asm ("sar %%cl, %%eax;" \
   1.645 +         "setcb %%cl;" \
   1.646 +		 : "=a" (value), "=c" (C_OUT) \
   1.647 +		 : "a" (reg[opcode & 0x0f].I), "c" (shift));
   1.648 +
   1.649 +#define LOGICAL_ROR_REG \
   1.650 +    asm ("ror %%cl, %%eax;" \
   1.651 +         "setcb %%cl;" \
   1.652 +		 : "=a" (value), "=c" (C_OUT) \
   1.653 +		 : "a" (reg[opcode & 0x0f].I), "c" (shift));
   1.654 +
   1.655 +#define LOGICAL_RRX_REG \
   1.656 +    asm ("bt $0, C_FLAG;" \
   1.657 +         "rcr $1, %%eax;" \
   1.658 +         "setcb %%cl;" \
   1.659 +		 : "=a" (value), "=c" (C_OUT) \
   1.660 +		 : "a" (reg[opcode & 0x0f].I));
   1.661 +
   1.662 +#define LOGICAL_ROR_IMM \
   1.663 +    asm ("ror %%cl, %%eax;" \
   1.664 +         "setcb %%cl;" \
   1.665 +		 : "=a" (value), "=c" (C_OUT) \
   1.666 +		 : "a" (opcode & 0xff), "c" (shift));
   1.667 +#define ARITHMETIC_LSL_REG \
   1.668 +    asm ("\
   1.669 +             shl %%cl, %%eax;" \
   1.670 +		 : "=a" (value) \
   1.671 +		 : "a" (reg[opcode & 0x0f].I), "c" (shift));
   1.672 +
   1.673 +#define ARITHMETIC_LSR_REG \
   1.674 +    asm ("\
   1.675 +             shr %%cl, %%eax;" \
   1.676 +		 : "=a" (value) \
   1.677 +		 : "a" (reg[opcode & 0x0f].I), "c" (shift));
   1.678 +
   1.679 +#define ARITHMETIC_ASR_REG \
   1.680 +    asm ("\
   1.681 +             sar %%cl, %%eax;" \
   1.682 +		 : "=a" (value) \
   1.683 +		 : "a" (reg[opcode & 0x0f].I), "c" (shift));
   1.684 +
   1.685 +#define ARITHMETIC_ROR_REG \
   1.686 +    asm ("\
   1.687 +             ror %%cl, %%eax;" \
   1.688 +		 : "=a" (value) \
   1.689 +		 : "a" (reg[opcode & 0x0f].I), "c" (shift));
   1.690 +
   1.691 +#define ARITHMETIC_RRX_REG \
   1.692 +    asm ("\
   1.693 +             bt $0, C_FLAG;\
   1.694 +             rcr $1, %%eax;" \
   1.695 +		 : "=a" (value) \
   1.696 +		 : "a" (reg[opcode & 0x0f].I));
   1.697 +
   1.698 +#define ARITHMETIC_ROR_IMM \
   1.699 +    asm ("\
   1.700 +             ror %%cl, %%eax;" \
   1.701 +		 : "=a" (value) \
   1.702 +		 : "a" (opcode & 0xff), "c" (shift));
   1.703 +#define ROR_IMM_MSR \
   1.704 +    asm ("ror %%cl, %%eax;" \
   1.705 +		 : "=a" (value) \
   1.706 +		 : "a" (opcode & 0xFF), "c" (shift));
   1.707 +#define ROR_VALUE \
   1.708 +    asm ("ror %%cl, %0" \
   1.709 +		 : "=r" (value) \
   1.710 +		 : "r" (value), "c" (shift));
   1.711 +#define RCR_VALUE \
   1.712 +    asm ("bt $0, C_FLAG;" \
   1.713 +         "rcr $1, %0" \
   1.714 +		 : "=r" (value) \
   1.715 +		 : "r" (value));
   1.716 +#endif
   1.717 +#else
   1.718 +#define OP_SUB \
   1.719 +	{ \
   1.720 +		__asm mov ebx, base \
   1.721 +		__asm mov ebx, dword ptr [OFFSET reg + 4 * ebx] \
   1.722 +		__asm sub ebx, value \
   1.723 +		__asm mov eax, dest \
   1.724 +		__asm mov dword ptr [OFFSET reg + 4 * eax], ebx \
   1.725 +	}
   1.726 +
   1.727 +#define OP_SUBS \
   1.728 +	{ \
   1.729 +		__asm mov ebx, base \
   1.730 +		__asm mov ebx, dword ptr [OFFSET reg + 4 * ebx] \
   1.731 +		__asm sub ebx, value \
   1.732 +		__asm mov eax, dest \
   1.733 +		__asm mov dword ptr [OFFSET reg + 4 * eax], ebx \
   1.734 +		__asm sets byte ptr N_FLAG \
   1.735 +		__asm setz byte ptr Z_FLAG \
   1.736 +		__asm setnc byte ptr C_FLAG \
   1.737 +		__asm seto byte ptr V_FLAG \
   1.738 +	}
   1.739 +
   1.740 +#define OP_RSB \
   1.741 +	{ \
   1.742 +		__asm mov ebx, base \
   1.743 +		__asm mov ebx, dword ptr [OFFSET reg + 4 * ebx] \
   1.744 +		__asm mov eax, value \
   1.745 +		__asm sub eax, ebx \
   1.746 +		__asm mov ebx, dest \
   1.747 +		__asm mov dword ptr [OFFSET reg + 4 * ebx], eax \
   1.748 +	}
   1.749 +
   1.750 +#define OP_RSBS \
   1.751 +	{ \
   1.752 +		__asm mov ebx, base \
   1.753 +		__asm mov ebx, dword ptr [OFFSET reg + 4 * ebx] \
   1.754 +		__asm mov eax, value \
   1.755 +		__asm sub eax, ebx \
   1.756 +		__asm mov ebx, dest \
   1.757 +		__asm mov dword ptr [OFFSET reg + 4 * ebx], eax \
   1.758 +		__asm sets byte ptr N_FLAG \
   1.759 +		__asm setz byte ptr Z_FLAG \
   1.760 +		__asm setnc byte ptr C_FLAG \
   1.761 +		__asm seto byte ptr V_FLAG \
   1.762 +	}
   1.763 +
   1.764 +#define OP_ADD \
   1.765 +	{ \
   1.766 +		__asm mov ebx, base \
   1.767 +		__asm mov ebx, dword ptr [OFFSET reg + 4 * ebx] \
   1.768 +		__asm add ebx, value \
   1.769 +		__asm mov eax, dest \
   1.770 +		__asm mov dword ptr [OFFSET reg + 4 * eax], ebx \
   1.771 +	}
   1.772 +
   1.773 +#define OP_ADDS \
   1.774 +	{ \
   1.775 +		__asm mov ebx, base \
   1.776 +		__asm mov ebx, dword ptr [OFFSET reg + 4 * ebx] \
   1.777 +		__asm add ebx, value \
   1.778 +		__asm mov eax, dest \
   1.779 +		__asm mov dword ptr [OFFSET reg + 4 * eax], ebx \
   1.780 +		__asm sets byte ptr N_FLAG \
   1.781 +		__asm setz byte ptr Z_FLAG \
   1.782 +		__asm setc byte ptr C_FLAG \
   1.783 +		__asm seto byte ptr V_FLAG \
   1.784 +	}
   1.785 +
   1.786 +#define OP_ADC \
   1.787 +	{ \
   1.788 +		__asm mov ebx, base \
   1.789 +		__asm mov ebx, dword ptr [OFFSET reg + 4 * ebx] \
   1.790 +		__asm bt word ptr C_FLAG, 0 \
   1.791 +		__asm adc ebx, value \
   1.792 +		__asm mov eax, dest \
   1.793 +		__asm mov dword ptr [OFFSET reg + 4 * eax], ebx \
   1.794 +	}
   1.795 +
   1.796 +#define OP_ADCS \
   1.797 +	{ \
   1.798 +		__asm mov ebx, base \
   1.799 +		__asm mov ebx, dword ptr [OFFSET reg + 4 * ebx] \
   1.800 +		__asm bt word ptr C_FLAG, 0 \
   1.801 +		__asm adc ebx, value \
   1.802 +		__asm mov eax, dest \
   1.803 +		__asm mov dword ptr [OFFSET reg + 4 * eax], ebx \
   1.804 +		__asm sets byte ptr N_FLAG \
   1.805 +		__asm setz byte ptr Z_FLAG \
   1.806 +		__asm setc byte ptr C_FLAG \
   1.807 +		__asm seto byte ptr V_FLAG \
   1.808 +	}
   1.809 +
   1.810 +#define OP_SBC \
   1.811 +	{ \
   1.812 +		__asm mov ebx, base \
   1.813 +		__asm mov ebx, dword ptr [OFFSET reg + 4 * ebx] \
   1.814 +		__asm mov eax, value \
   1.815 +		__asm bt word ptr C_FLAG, 0 \
   1.816 +		__asm cmc \
   1.817 +		__asm sbb ebx, eax \
   1.818 +		__asm mov eax, dest \
   1.819 +		__asm mov dword ptr [OFFSET reg + 4 * eax], ebx \
   1.820 +	}
   1.821 +
   1.822 +#define OP_SBCS \
   1.823 +	{ \
   1.824 +		__asm mov ebx, base \
   1.825 +		__asm mov ebx, dword ptr [OFFSET reg + 4 * ebx] \
   1.826 +		__asm mov eax, value \
   1.827 +		__asm bt word ptr C_FLAG, 0 \
   1.828 +		__asm cmc \
   1.829 +		__asm sbb ebx, eax \
   1.830 +		__asm mov eax, dest \
   1.831 +		__asm mov dword ptr [OFFSET reg + 4 * eax], ebx \
   1.832 +		__asm sets byte ptr N_FLAG \
   1.833 +		__asm setz byte ptr Z_FLAG \
   1.834 +		__asm setnc byte ptr C_FLAG \
   1.835 +		__asm seto byte ptr V_FLAG \
   1.836 +	}
   1.837 +#define OP_RSC \
   1.838 +	{ \
   1.839 +		__asm mov ebx, value \
   1.840 +		__asm mov eax, base \
   1.841 +		__asm mov eax, dword ptr[OFFSET reg + 4 * eax] \
   1.842 +		__asm bt word ptr C_FLAG, 0 \
   1.843 +		__asm cmc \
   1.844 +		__asm sbb ebx, eax \
   1.845 +		__asm mov eax, dest \
   1.846 +		__asm mov dword ptr [OFFSET reg + 4 * eax], ebx \
   1.847 +	}
   1.848 +
   1.849 +#define OP_RSCS \
   1.850 +	{ \
   1.851 +		__asm mov ebx, value \
   1.852 +		__asm mov eax, base \
   1.853 +		__asm mov eax, dword ptr[OFFSET reg + 4 * eax] \
   1.854 +		__asm bt word ptr C_FLAG, 0 \
   1.855 +		__asm cmc \
   1.856 +		__asm sbb ebx, eax \
   1.857 +		__asm mov eax, dest \
   1.858 +		__asm mov dword ptr [OFFSET reg + 4 * eax], ebx \
   1.859 +		__asm sets byte ptr N_FLAG \
   1.860 +		__asm setz byte ptr Z_FLAG \
   1.861 +		__asm setnc byte ptr C_FLAG \
   1.862 +		__asm seto byte ptr V_FLAG \
   1.863 +	}
   1.864 +#define OP_CMP \
   1.865 +	{ \
   1.866 +		__asm mov eax, base \
   1.867 +		__asm mov ebx, dword ptr [OFFSET reg + 4 * eax] \
   1.868 +		__asm sub ebx, value \
   1.869 +		__asm sets byte ptr N_FLAG \
   1.870 +		__asm setz byte ptr Z_FLAG \
   1.871 +		__asm setnc byte ptr C_FLAG \
   1.872 +		__asm seto byte ptr V_FLAG \
   1.873 +	}
   1.874 +
   1.875 +#define OP_CMN \
   1.876 +	{ \
   1.877 +		__asm mov eax, base \
   1.878 +		__asm mov ebx, dword ptr [OFFSET reg + 4 * eax] \
   1.879 +		__asm add ebx, value \
   1.880 +		__asm sets byte ptr N_FLAG \
   1.881 +		__asm setz byte ptr Z_FLAG \
   1.882 +		__asm setc byte ptr C_FLAG \
   1.883 +		__asm seto byte ptr V_FLAG \
   1.884 +	}
   1.885 +#define LOGICAL_LSL_REG \
   1.886 +    __asm mov eax, opcode \
   1.887 +    __asm and eax, 0x0f \
   1.888 +    __asm mov eax, dword ptr [OFFSET reg + 4 * eax] \
   1.889 +    __asm mov cl, byte ptr shift \
   1.890 +    __asm shl eax, cl \
   1.891 +    __asm mov value, eax \
   1.892 +    __asm setc byte ptr C_OUT
   1.893 +
   1.894 +#define LOGICAL_LSR_REG \
   1.895 +    __asm mov eax, opcode \
   1.896 +    __asm and eax, 0x0f \
   1.897 +    __asm mov eax, dword ptr [OFFSET reg + 4 * eax] \
   1.898 +    __asm mov cl, byte ptr shift \
   1.899 +    __asm shr eax, cl \
   1.900 +    __asm mov value, eax \
   1.901 +    __asm setc byte ptr C_OUT
   1.902 +
   1.903 +#define LOGICAL_ASR_REG \
   1.904 +    __asm mov eax, opcode \
   1.905 +    __asm and eax, 0x0f \
   1.906 +    __asm mov eax, dword ptr [OFFSET reg + 4 * eax] \
   1.907 +    __asm mov cl, byte ptr shift \
   1.908 +    __asm sar eax, cl \
   1.909 +    __asm mov value, eax \
   1.910 +    __asm setc byte ptr C_OUT
   1.911 +
   1.912 +#define LOGICAL_ROR_REG \
   1.913 +    __asm mov eax, opcode \
   1.914 +    __asm and eax, 0x0F \
   1.915 +    __asm mov eax, dword ptr [OFFSET reg + 4 * eax] \
   1.916 +    __asm mov cl, byte ptr shift \
   1.917 +    __asm ror eax, cl \
   1.918 +    __asm mov value, eax \
   1.919 +    __asm setc byte ptr C_OUT
   1.920 +
   1.921 +#define LOGICAL_RRX_REG \
   1.922 +    __asm mov eax, opcode \
   1.923 +    __asm and eax, 0x0F \
   1.924 +    __asm mov eax, dword ptr [OFFSET reg + 4 * eax] \
   1.925 +    __asm bt word ptr C_OUT, 0 \
   1.926 +    __asm rcr eax, 1 \
   1.927 +    __asm mov value, eax \
   1.928 +    __asm setc byte ptr C_OUT
   1.929 +
   1.930 +#define LOGICAL_ROR_IMM \
   1.931 +    __asm mov eax, opcode \
   1.932 +    __asm and eax, 0xff \
   1.933 +    __asm mov cl, byte ptr shift \
   1.934 +    __asm ror eax, cl \
   1.935 +    __asm mov value, eax \
   1.936 +    __asm setc byte ptr C_OUT
   1.937 +#define ARITHMETIC_LSL_REG \
   1.938 +    __asm mov eax, opcode \
   1.939 +    __asm and eax, 0x0f \
   1.940 +    __asm mov eax, dword ptr [OFFSET reg + 4 * eax] \
   1.941 +    __asm mov cl, byte ptr shift \
   1.942 +    __asm shl eax, cl \
   1.943 +    __asm mov value, eax
   1.944 +
   1.945 +#define ARITHMETIC_LSR_REG \
   1.946 +    __asm mov eax, opcode \
   1.947 +    __asm and eax, 0x0f \
   1.948 +    __asm mov eax, dword ptr [OFFSET reg + 4 * eax] \
   1.949 +    __asm mov cl, byte ptr shift \
   1.950 +    __asm shr eax, cl \
   1.951 +    __asm mov value, eax
   1.952 +
   1.953 +#define ARITHMETIC_ASR_REG \
   1.954 +    __asm mov eax, opcode \
   1.955 +    __asm and eax, 0x0f \
   1.956 +    __asm mov eax, dword ptr [OFFSET reg + 4 * eax] \
   1.957 +    __asm mov cl, byte ptr shift \
   1.958 +    __asm sar eax, cl \
   1.959 +    __asm mov value, eax
   1.960 +
   1.961 +#define ARITHMETIC_ROR_REG \
   1.962 +    __asm mov eax, opcode \
   1.963 +    __asm and eax, 0x0F \
   1.964 +    __asm mov eax, dword ptr [OFFSET reg + 4 * eax] \
   1.965 +    __asm mov cl, byte ptr shift \
   1.966 +    __asm ror eax, cl \
   1.967 +    __asm mov value, eax
   1.968 +
   1.969 +#define ARITHMETIC_RRX_REG \
   1.970 +    __asm mov eax, opcode \
   1.971 +    __asm and eax, 0x0F \
   1.972 +    __asm mov eax, dword ptr [OFFSET reg + 4 * eax] \
   1.973 +    __asm bt word ptr C_FLAG, 0 \
   1.974 +    __asm rcr eax, 1 \
   1.975 +    __asm mov value, eax
   1.976 +
   1.977 +#define ARITHMETIC_ROR_IMM \
   1.978 +    __asm mov eax, opcode \
   1.979 +    __asm and eax, 0xff \
   1.980 +    __asm mov cl, byte ptr shift \
   1.981 +    __asm ror eax, cl \
   1.982 +    __asm mov value, eax
   1.983 +#define ROR_IMM_MSR \
   1.984 +	{ \
   1.985 +		__asm mov eax, opcode \
   1.986 +		          __asm and eax, 0xff \
   1.987 +		__asm mov cl, byte ptr shift \
   1.988 +		__asm ror eax, CL \
   1.989 +		__asm mov value, eax \
   1.990 +	}
   1.991 +#define ROR_VALUE \
   1.992 +	{ \
   1.993 +		__asm mov cl, byte ptr shift \
   1.994 +		__asm ror dword ptr value, cl \
   1.995 +	}
   1.996 +#define RCR_VALUE \
   1.997 +	{ \
   1.998 +		__asm mov cl, byte ptr shift \
   1.999 +		__asm bt word ptr C_FLAG, 0 \
  1.1000 +		__asm rcr dword ptr value, 1 \
  1.1001 +	}
  1.1002 +#endif
  1.1003 +#endif
  1.1004 +
  1.1005 +#define OP_TST \
  1.1006 +    u32 res = reg[base].I & value; \
  1.1007 +    N_FLAG	= (res & 0x80000000) ? true : false; \
  1.1008 +    Z_FLAG	= (res) ? false : true; \
  1.1009 +    C_FLAG	= C_OUT;
  1.1010 +
  1.1011 +#define OP_TEQ \
  1.1012 +    u32 res = reg[base].I ^ value; \
  1.1013 +    N_FLAG	= (res & 0x80000000) ? true : false; \
  1.1014 +    Z_FLAG	= (res) ? false : true; \
  1.1015 +    C_FLAG	= C_OUT;
  1.1016 +
  1.1017 +#define OP_ORR \
  1.1018 +    reg[dest].I = reg[base].I | value;
  1.1019 +
  1.1020 +#define OP_ORRS \
  1.1021 +    reg[dest].I = reg[base].I | value; \
  1.1022 +    N_FLAG		= (reg[dest].I & 0x80000000) ? true : false; \
  1.1023 +    Z_FLAG		= (reg[dest].I) ? false : true; \
  1.1024 +    C_FLAG		= C_OUT;
  1.1025 +
  1.1026 +#define OP_MOV \
  1.1027 +    reg[dest].I = value;
  1.1028 +
  1.1029 +#define OP_MOVS \
  1.1030 +    reg[dest].I = value; \
  1.1031 +    N_FLAG		= (reg[dest].I & 0x80000000) ? true : false; \
  1.1032 +    Z_FLAG		= (reg[dest].I) ? false : true; \
  1.1033 +    C_FLAG		= C_OUT;
  1.1034 +
  1.1035 +#define OP_BIC \
  1.1036 +    reg[dest].I = reg[base].I & (~value);
  1.1037 +
  1.1038 +#define OP_BICS \
  1.1039 +    reg[dest].I = reg[base].I & (~value); \
  1.1040 +    N_FLAG		= (reg[dest].I & 0x80000000) ? true : false; \
  1.1041 +    Z_FLAG		= (reg[dest].I) ? false : true; \
  1.1042 +    C_FLAG		= C_OUT;
  1.1043 +
  1.1044 +#define OP_MVN \
  1.1045 +    reg[dest].I = ~value;
  1.1046 +
  1.1047 +#define OP_MVNS \
  1.1048 +    reg[dest].I = ~value; \
  1.1049 +    N_FLAG		= (reg[dest].I & 0x80000000) ? true : false; \
  1.1050 +    Z_FLAG		= (reg[dest].I) ? false : true; \
  1.1051 +    C_FLAG		= C_OUT;
  1.1052 +
  1.1053 +#define CASE_16(BASE) \
  1.1054 +case BASE: \
  1.1055 +case BASE + 1: \
  1.1056 +case BASE + 2: \
  1.1057 +case BASE + 3: \
  1.1058 +case BASE + 4: \
  1.1059 +case BASE + 5: \
  1.1060 +case BASE + 6: \
  1.1061 +case BASE + 7: \
  1.1062 +case BASE + 8: \
  1.1063 +case BASE + 9: \
  1.1064 +case BASE + 10: \
  1.1065 +case BASE + 11: \
  1.1066 +case BASE + 12: \
  1.1067 +case BASE + 13: \
  1.1068 +case BASE + 14: \
  1.1069 +case BASE + 15:
  1.1070 +
  1.1071 +#define CASE_256(BASE) \
  1.1072 +    CASE_16(BASE) \
  1.1073 +    CASE_16(BASE + 0x10) \
  1.1074 +    CASE_16(BASE + 0x20) \
  1.1075 +    CASE_16(BASE + 0x30) \
  1.1076 +    CASE_16(BASE + 0x40) \
  1.1077 +    CASE_16(BASE + 0x50) \
  1.1078 +    CASE_16(BASE + 0x60) \
  1.1079 +    CASE_16(BASE + 0x70) \
  1.1080 +    CASE_16(BASE + 0x80) \
  1.1081 +    CASE_16(BASE + 0x90) \
  1.1082 +    CASE_16(BASE + 0xa0) \
  1.1083 +    CASE_16(BASE + 0xb0) \
  1.1084 +    CASE_16(BASE + 0xc0) \
  1.1085 +    CASE_16(BASE + 0xd0) \
  1.1086 +    CASE_16(BASE + 0xe0) \
  1.1087 +    CASE_16(BASE + 0xf0)
  1.1088 +
  1.1089 +#define LOGICAL_DATA_OPCODE(OPCODE, OPCODE2, BASE) \
  1.1090 +case BASE: \
  1.1091 +case BASE + 8: \
  1.1092 +{ \
  1.1093 +	/* OP Rd,Rb,Rm LSL # */ \
  1.1094 +	int	 base  = (opcode >> 16) & 0x0F; \
  1.1095 +	int	 shift = (opcode >> 7) & 0x1F; \
  1.1096 +	int	 dest  = (opcode >> 12) & 15; \
  1.1097 +	bool C_OUT = C_FLAG; \
  1.1098 +	u32	 value; \
  1.1099 +      \
  1.1100 +	if (shift) { \
  1.1101 +		LOGICAL_LSL_REG \
  1.1102 +	} else { \
  1.1103 +		value = reg[opcode & 0x0F].I; \
  1.1104 +	} \
  1.1105 +	if (dest == 15) { \
  1.1106 +		OPCODE2 \
  1.1107 +		/* todo */ \
  1.1108 +		if (opcode & 0x00100000) { \
  1.1109 +			clockTicks++; \
  1.1110 +			CPUSwitchMode(reg[17].I & 0x1f, false); \
  1.1111 +		} \
  1.1112 +		if (armState) { \
  1.1113 +			reg[15].I &= 0xFFFFFFFC; \
  1.1114 +			armNextPC  = reg[15].I; \
  1.1115 +			reg[15].I += 4; \
  1.1116 +		} else { \
  1.1117 +			reg[15].I &= 0xFFFFFFFE; \
  1.1118 +			armNextPC  = reg[15].I; \
  1.1119 +			reg[15].I += 2; \
  1.1120 +		} \
  1.1121 +	} else { \
  1.1122 +		OPCODE \
  1.1123 +	} \
  1.1124 +} \
  1.1125 +break; \
  1.1126 +case BASE + 2: \
  1.1127 +case BASE + 10: \
  1.1128 +{ \
  1.1129 +	/* OP Rd,Rb,Rm LSR # */ \
  1.1130 +	int	 base  = (opcode >> 16) & 0x0F; \
  1.1131 +	int	 shift = (opcode >> 7) & 0x1F; \
  1.1132 +	int	 dest  = (opcode >> 12) & 15; \
  1.1133 +	bool C_OUT = C_FLAG; \
  1.1134 +	u32	 value; \
  1.1135 +	if (shift) { \
  1.1136 +		LOGICAL_LSR_REG \
  1.1137 +	} else { \
  1.1138 +		value = 0; \
  1.1139 +		C_OUT = (reg[opcode & 0x0F].I & 0x80000000) ? true : false; \
  1.1140 +	} \
  1.1141 +      \
  1.1142 +	if (dest == 15) { \
  1.1143 +		OPCODE2 \
  1.1144 +		/* todo */ \
  1.1145 +		if (opcode & 0x00100000) { \
  1.1146 +			clockTicks++; \
  1.1147 +			CPUSwitchMode(reg[17].I & 0x1f, false); \
  1.1148 +		} \
  1.1149 +		if (armState) { \
  1.1150 +			reg[15].I &= 0xFFFFFFFC; \
  1.1151 +			armNextPC  = reg[15].I; \
  1.1152 +			reg[15].I += 4; \
  1.1153 +		} else { \
  1.1154 +			reg[15].I &= 0xFFFFFFFE; \
  1.1155 +			armNextPC  = reg[15].I; \
  1.1156 +			reg[15].I += 2; \
  1.1157 +		} \
  1.1158 +	} else { \
  1.1159 +		OPCODE \
  1.1160 +	} \
  1.1161 +} \
  1.1162 +break; \
  1.1163 +case BASE + 4: \
  1.1164 +case BASE + 12: \
  1.1165 +{ \
  1.1166 +	/* OP Rd,Rb,Rm ASR # */ \
  1.1167 +	int	 base  = (opcode >> 16) & 0x0F; \
  1.1168 +	int	 shift = (opcode >> 7) & 0x1F; \
  1.1169 +	int	 dest  = (opcode >> 12) & 15; \
  1.1170 +	bool C_OUT = C_FLAG; \
  1.1171 +	u32	 value; \
  1.1172 +	if (shift) { \
  1.1173 +		LOGICAL_ASR_REG \
  1.1174 +	} else { \
  1.1175 +		if (reg[opcode & 0x0F].I & 0x80000000) { \
  1.1176 +			value = 0xFFFFFFFF; \
  1.1177 +			C_OUT = true; \
  1.1178 +		} else { \
  1.1179 +			value = 0; \
  1.1180 +			C_OUT = false; \
  1.1181 +		}                   \
  1.1182 +	} \
  1.1183 +      \
  1.1184 +	if (dest == 15) { \
  1.1185 +		OPCODE2 \
  1.1186 +		/* todo */ \
  1.1187 +		if (opcode & 0x00100000) { \
  1.1188 +			clockTicks++; \
  1.1189 +			CPUSwitchMode(reg[17].I & 0x1f, false); \
  1.1190 +		} \
  1.1191 +		if (armState) { \
  1.1192 +			reg[15].I &= 0xFFFFFFFC; \
  1.1193 +			armNextPC  = reg[15].I; \
  1.1194 +			reg[15].I += 4; \
  1.1195 +		} else { \
  1.1196 +			reg[15].I &= 0xFFFFFFFE; \
  1.1197 +			armNextPC  = reg[15].I; \
  1.1198 +			reg[15].I += 2; \
  1.1199 +		} \
  1.1200 +	} else { \
  1.1201 +		OPCODE \
  1.1202 +	} \
  1.1203 +} \
  1.1204 +break; \
  1.1205 +case BASE + 6: \
  1.1206 +case BASE + 14: \
  1.1207 +{ \
  1.1208 +	/* OP Rd,Rb,Rm ROR # */ \
  1.1209 +	int	 base  = (opcode >> 16) & 0x0F; \
  1.1210 +	int	 shift = (opcode >> 7) & 0x1F; \
  1.1211 +	int	 dest  = (opcode >> 12) & 15; \
  1.1212 +	bool C_OUT = C_FLAG; \
  1.1213 +	u32	 value; \
  1.1214 +	if (shift) { \
  1.1215 +		LOGICAL_ROR_REG \
  1.1216 +	} else { \
  1.1217 +		LOGICAL_RRX_REG \
  1.1218 +	} \
  1.1219 +	if (dest == 15) { \
  1.1220 +		OPCODE2 \
  1.1221 +		/* todo */ \
  1.1222 +		if (opcode & 0x00100000) { \
  1.1223 +			clockTicks++; \
  1.1224 +			CPUSwitchMode(reg[17].I & 0x1f, false); \
  1.1225 +		} \
  1.1226 +		if (armState) { \
  1.1227 +			reg[15].I &= 0xFFFFFFFC; \
  1.1228 +			armNextPC  = reg[15].I; \
  1.1229 +			reg[15].I += 4; \
  1.1230 +		} else { \
  1.1231 +			reg[15].I &= 0xFFFFFFFE; \
  1.1232 +			armNextPC  = reg[15].I; \
  1.1233 +			reg[15].I += 2; \
  1.1234 +		} \
  1.1235 +	} else { \
  1.1236 +		OPCODE \
  1.1237 +	} \
  1.1238 +} \
  1.1239 +break; \
  1.1240 +case BASE + 1: \
  1.1241 +{ \
  1.1242 +	/* OP Rd,Rb,Rm LSL Rs */ \
  1.1243 +	clockTicks++; \
  1.1244 +	int	 base  = (opcode >> 16) & 0x0F; \
  1.1245 +	int	 shift = reg[(opcode >> 8) & 15].B.B0; \
  1.1246 +	int	 dest  = (opcode >> 12) & 15; \
  1.1247 +	bool C_OUT = C_FLAG; \
  1.1248 +	u32	 value; \
  1.1249 +	if (shift) { \
  1.1250 +		if (shift == 32) { \
  1.1251 +			value = 0; \
  1.1252 +			C_OUT = (reg[opcode & 0x0F].I & 1 ? true : false); \
  1.1253 +		} else if (shift < 32) { \
  1.1254 +			LOGICAL_LSL_REG \
  1.1255 +		} else { \
  1.1256 +			value = 0; \
  1.1257 +			C_OUT = false; \
  1.1258 +		} \
  1.1259 +	} else { \
  1.1260 +		value = reg[opcode & 0x0F].I; \
  1.1261 +	} \
  1.1262 +	if (dest == 15) { \
  1.1263 +		OPCODE2 \
  1.1264 +		/* todo */ \
  1.1265 +		if (opcode & 0x00100000) { \
  1.1266 +			clockTicks++; \
  1.1267 +			CPUSwitchMode(reg[17].I & 0x1f, false); \
  1.1268 +		} \
  1.1269 +		if (armState) { \
  1.1270 +			reg[15].I &= 0xFFFFFFFC; \
  1.1271 +			armNextPC  = reg[15].I; \
  1.1272 +			reg[15].I += 4; \
  1.1273 +		} else { \
  1.1274 +			reg[15].I &= 0xFFFFFFFE; \
  1.1275 +			armNextPC  = reg[15].I; \
  1.1276 +			reg[15].I += 2; \
  1.1277 +		} \
  1.1278 +	} else { \
  1.1279 +		OPCODE \
  1.1280 +	} \
  1.1281 +} \
  1.1282 +break; \
  1.1283 +case BASE + 3: \
  1.1284 +{ \
  1.1285 +	/* OP Rd,Rb,Rm LSR Rs */ \
  1.1286 +	clockTicks++; \
  1.1287 +	int	 base  = (opcode >> 16) & 0x0F; \
  1.1288 +	int	 shift = reg[(opcode >> 8) & 15].B.B0; \
  1.1289 +	int	 dest  = (opcode >> 12) & 15; \
  1.1290 +	bool C_OUT = C_FLAG; \
  1.1291 +	u32	 value; \
  1.1292 +	if (shift) { \
  1.1293 +		if (shift == 32) { \
  1.1294 +			value = 0; \
  1.1295 +			C_OUT = (reg[opcode & 0x0F].I & 0x80000000 ? true : false); \
  1.1296 +		} else if (shift < 32) { \
  1.1297 +			LOGICAL_LSR_REG \
  1.1298 +		} else { \
  1.1299 +			value = 0; \
  1.1300 +			C_OUT = false; \
  1.1301 +		} \
  1.1302 +	} else { \
  1.1303 +		value = reg[opcode & 0x0F].I; \
  1.1304 +	} \
  1.1305 +	if (dest == 15) { \
  1.1306 +		OPCODE2 \
  1.1307 +		/* todo */ \
  1.1308 +		if (opcode & 0x00100000) { \
  1.1309 +			clockTicks++; \
  1.1310 +			CPUSwitchMode(reg[17].I & 0x1f, false); \
  1.1311 +		} \
  1.1312 +		if (armState) { \
  1.1313 +			reg[15].I &= 0xFFFFFFFC; \
  1.1314 +			armNextPC  = reg[15].I; \
  1.1315 +			reg[15].I += 4; \
  1.1316 +		} else { \
  1.1317 +			reg[15].I &= 0xFFFFFFFE; \
  1.1318 +			armNextPC  = reg[15].I; \
  1.1319 +			reg[15].I += 2; \
  1.1320 +		} \
  1.1321 +	} else { \
  1.1322 +		OPCODE \
  1.1323 +	} \
  1.1324 +} \
  1.1325 +break; \
  1.1326 +case BASE + 5: \
  1.1327 +{ \
  1.1328 +	/* OP Rd,Rb,Rm ASR Rs */ \
  1.1329 +	clockTicks++; \
  1.1330 +	int	 base  = (opcode >> 16) & 0x0F; \
  1.1331 +	int	 shift = reg[(opcode >> 8) & 15].B.B0; \
  1.1332 +	int	 dest  = (opcode >> 12) & 15; \
  1.1333 +	bool C_OUT = C_FLAG; \
  1.1334 +	u32	 value; \
  1.1335 +	if (shift < 32) { \
  1.1336 +		if (shift) { \
  1.1337 +			LOGICAL_ASR_REG \
  1.1338 +		} else { \
  1.1339 +			value = reg[opcode & 0x0F].I; \
  1.1340 +		} \
  1.1341 +	} else { \
  1.1342 +		if (reg[opcode & 0x0F].I & 0x80000000) { \
  1.1343 +			value = 0xFFFFFFFF; \
  1.1344 +			C_OUT = true; \
  1.1345 +		} else { \
  1.1346 +			value = 0; \
  1.1347 +			C_OUT = false; \
  1.1348 +		} \
  1.1349 +	} \
  1.1350 +	if (dest == 15) { \
  1.1351 +		OPCODE2 \
  1.1352 +		/* todo */ \
  1.1353 +		if (opcode & 0x00100000) { \
  1.1354 +			clockTicks++; \
  1.1355 +			CPUSwitchMode(reg[17].I & 0x1f, false); \
  1.1356 +		} \
  1.1357 +		if (armState) { \
  1.1358 +			reg[15].I &= 0xFFFFFFFC; \
  1.1359 +			armNextPC  = reg[15].I; \
  1.1360 +			reg[15].I += 4; \
  1.1361 +		} else { \
  1.1362 +			reg[15].I &= 0xFFFFFFFE; \
  1.1363 +			armNextPC  = reg[15].I; \
  1.1364 +			reg[15].I += 2; \
  1.1365 +		} \
  1.1366 +	} else { \
  1.1367 +		OPCODE \
  1.1368 +	} \
  1.1369 +} \
  1.1370 +break; \
  1.1371 +case BASE + 7: \
  1.1372 +{ \
  1.1373 +	/* OP Rd,Rb,Rm ROR Rs */ \
  1.1374 +	clockTicks++; \
  1.1375 +	int	 base  = (opcode >> 16) & 0x0F; \
  1.1376 +	int	 shift = reg[(opcode >> 8) & 15].B.B0; \
  1.1377 +	int	 dest  = (opcode >> 12) & 15; \
  1.1378 +	bool C_OUT = C_FLAG; \
  1.1379 +	u32	 value; \
  1.1380 +	if (shift) { \
  1.1381 +		shift &= 0x1f; \
  1.1382 +		if (shift) { \
  1.1383 +			LOGICAL_ROR_REG \
  1.1384 +		} else { \
  1.1385 +			value = reg[opcode & 0x0F].I; \
  1.1386 +			C_OUT = (value & 0x80000000 ? true : false); \
  1.1387 +		} \
  1.1388 +	} else { \
  1.1389 +		value = reg[opcode & 0x0F].I; \
  1.1390 +		C_OUT = (value & 0x80000000 ? true : false); \
  1.1391 +	} \
  1.1392 +	if (dest == 15) { \
  1.1393 +		OPCODE2 \
  1.1394 +		/* todo */ \
  1.1395 +		if (opcode & 0x00100000) { \
  1.1396 +			clockTicks++; \
  1.1397 +			CPUSwitchMode(reg[17].I & 0x1f, false); \
  1.1398 +		} \
  1.1399 +		if (armState) { \
  1.1400 +			reg[15].I &= 0xFFFFFFFC; \
  1.1401 +			armNextPC  = reg[15].I; \
  1.1402 +			reg[15].I += 4; \
  1.1403 +		} else { \
  1.1404 +			reg[15].I &= 0xFFFFFFFE; \
  1.1405 +			armNextPC  = reg[15].I; \
  1.1406 +			reg[15].I += 2; \
  1.1407 +		} \
  1.1408 +	} else { \
  1.1409 +		OPCODE \
  1.1410 +	} \
  1.1411 +} \
  1.1412 +break; \
  1.1413 +case BASE + 0x200: \
  1.1414 +case BASE + 0x201: \
  1.1415 +case BASE + 0x202: \
  1.1416 +case BASE + 0x203: \
  1.1417 +case BASE + 0x204: \
  1.1418 +case BASE + 0x205: \
  1.1419 +case BASE + 0x206: \
  1.1420 +case BASE + 0x207: \
  1.1421 +case BASE + 0x208: \
  1.1422 +case BASE + 0x209: \
  1.1423 +case BASE + 0x20a: \
  1.1424 +case BASE + 0x20b: \
  1.1425 +case BASE + 0x20c: \
  1.1426 +case BASE + 0x20d: \
  1.1427 +case BASE + 0x20e: \
  1.1428 +case BASE + 0x20f: \
  1.1429 +{ \
  1.1430 +	int	 shift = (opcode & 0xF00) >> 7; \
  1.1431 +	int	 base  = (opcode >> 16) & 0x0F; \
  1.1432 +	int	 dest  = (opcode >> 12) & 0x0F; \
  1.1433 +	bool C_OUT = C_FLAG; \
  1.1434 +	u32	 value; \
  1.1435 +	if (shift) { \
  1.1436 +		LOGICAL_ROR_IMM \
  1.1437 +	} else { \
  1.1438 +		value = opcode & 0xff; \
  1.1439 +	} \
  1.1440 +	if (dest == 15) { \
  1.1441 +		OPCODE2 \
  1.1442 +		/* todo */ \
  1.1443 +		if (opcode & 0x00100000) { \
  1.1444 +			clockTicks++; \
  1.1445 +			CPUSwitchMode(reg[17].I & 0x1f, false); \
  1.1446 +		} \
  1.1447 +		if (armState) { \
  1.1448 +			reg[15].I &= 0xFFFFFFFC; \
  1.1449 +			armNextPC  = reg[15].I; \
  1.1450 +			reg[15].I += 4; \
  1.1451 +		} else { \
  1.1452 +			reg[15].I &= 0xFFFFFFFE; \
  1.1453 +			armNextPC  = reg[15].I; \
  1.1454 +			reg[15].I += 2; \
  1.1455 +		} \
  1.1456 +	} else { \
  1.1457 +		OPCODE \
  1.1458 +	} \
  1.1459 +} \
  1.1460 +break;
  1.1461 +
  1.1462 +#define LOGICAL_DATA_OPCODE_WITHOUT_base(OPCODE, OPCODE2, BASE) \
  1.1463 +case BASE: \
  1.1464 +case BASE + 8: \
  1.1465 +{ \
  1.1466 +	/* OP Rd,Rb,Rm LSL # */ \
  1.1467 +	int	 shift = (opcode >> 7) & 0x1F; \
  1.1468 +	int	 dest  = (opcode >> 12) & 15; \
  1.1469 +	bool C_OUT = C_FLAG; \
  1.1470 +	u32	 value; \
  1.1471 +      \
  1.1472 +	if (shift) { \
  1.1473 +		LOGICAL_LSL_REG \
  1.1474 +	} else { \
  1.1475 +		value = reg[opcode & 0x0F].I; \
  1.1476 +	} \
  1.1477 +	if (dest == 15) { \
  1.1478 +		OPCODE2 \
  1.1479 +		/* todo */ \
  1.1480 +		if (opcode & 0x00100000) { \
  1.1481 +			clockTicks++; \
  1.1482 +			CPUSwitchMode(reg[17].I & 0x1f, false); \
  1.1483 +		} \
  1.1484 +		if (armState) { \
  1.1485 +			reg[15].I &= 0xFFFFFFFC; \
  1.1486 +			armNextPC  = reg[15].I; \
  1.1487 +			reg[15].I += 4; \
  1.1488 +		} else { \
  1.1489 +			reg[15].I &= 0xFFFFFFFE; \
  1.1490 +			armNextPC  = reg[15].I; \
  1.1491 +			reg[15].I += 2; \
  1.1492 +		} \
  1.1493 +	} else { \
  1.1494 +		OPCODE \
  1.1495 +	} \
  1.1496 +} \
  1.1497 +break; \
  1.1498 +case BASE + 2: \
  1.1499 +case BASE + 10: \
  1.1500 +{ \
  1.1501 +	/* OP Rd,Rb,Rm LSR # */ \
  1.1502 +	int	 shift = (opcode >> 7) & 0x1F; \
  1.1503 +	int	 dest  = (opcode >> 12) & 15; \
  1.1504 +	bool C_OUT = C_FLAG; \
  1.1505 +	u32	 value; \
  1.1506 +	if (shift) { \
  1.1507 +		LOGICAL_LSR_REG \
  1.1508 +	} else { \
  1.1509 +		value = 0; \
  1.1510 +		C_OUT = (reg[opcode & 0x0F].I & 0x80000000) ? true : false; \
  1.1511 +	} \
  1.1512 +      \
  1.1513 +	if (dest == 15) { \
  1.1514 +		OPCODE2 \
  1.1515 +		/* todo */ \
  1.1516 +		if (opcode & 0x00100000) { \
  1.1517 +			clockTicks++; \
  1.1518 +			CPUSwitchMode(reg[17].I & 0x1f, false); \
  1.1519 +		} \
  1.1520 +		if (armState) { \
  1.1521 +			reg[15].I &= 0xFFFFFFFC; \
  1.1522 +			armNextPC  = reg[15].I; \
  1.1523 +			reg[15].I += 4; \
  1.1524 +		} else { \
  1.1525 +			reg[15].I &= 0xFFFFFFFE; \
  1.1526 +			armNextPC  = reg[15].I; \
  1.1527 +			reg[15].I += 2; \
  1.1528 +		} \
  1.1529 +	} else { \
  1.1530 +		OPCODE \
  1.1531 +	} \
  1.1532 +} \
  1.1533 +break; \
  1.1534 +case BASE + 4: \
  1.1535 +case BASE + 12: \
  1.1536 +{ \
  1.1537 +	/* OP Rd,Rb,Rm ASR # */ \
  1.1538 +	int	 shift = (opcode >> 7) & 0x1F; \
  1.1539 +	int	 dest  = (opcode >> 12) & 15; \
  1.1540 +	bool C_OUT = C_FLAG; \
  1.1541 +	u32	 value; \
  1.1542 +	if (shift) { \
  1.1543 +		LOGICAL_ASR_REG \
  1.1544 +	} else { \
  1.1545 +		if (reg[opcode & 0x0F].I & 0x80000000) { \
  1.1546 +			value = 0xFFFFFFFF; \
  1.1547 +			C_OUT = true; \
  1.1548 +		} else { \
  1.1549 +			value = 0; \
  1.1550 +			C_OUT = false; \
  1.1551 +		}                   \
  1.1552 +	} \
  1.1553 +      \
  1.1554 +	if (dest == 15) { \
  1.1555 +		OPCODE2 \
  1.1556 +		/* todo */ \
  1.1557 +		if (opcode & 0x00100000) { \
  1.1558 +			clockTicks++; \
  1.1559 +			CPUSwitchMode(reg[17].I & 0x1f, false); \
  1.1560 +		} \
  1.1561 +		if (armState) { \
  1.1562 +			reg[15].I &= 0xFFFFFFFC; \
  1.1563 +			armNextPC  = reg[15].I; \
  1.1564 +			reg[15].I += 4; \
  1.1565 +		} else { \
  1.1566 +			reg[15].I &= 0xFFFFFFFE; \
  1.1567 +			armNextPC  = reg[15].I; \
  1.1568 +			reg[15].I += 2; \
  1.1569 +		} \
  1.1570 +	} else { \
  1.1571 +		OPCODE \
  1.1572 +	} \
  1.1573 +} \
  1.1574 +break; \
  1.1575 +case BASE + 6: \
  1.1576 +case BASE + 14: \
  1.1577 +{ \
  1.1578 +	/* OP Rd,Rb,Rm ROR # */ \
  1.1579 +	int	 shift = (opcode >> 7) & 0x1F; \
  1.1580 +	int	 dest  = (opcode >> 12) & 15; \
  1.1581 +	bool C_OUT = C_FLAG; \
  1.1582 +	u32	 value; \
  1.1583 +	if (shift) { \
  1.1584 +		LOGICAL_ROR_REG \
  1.1585 +	} else { \
  1.1586 +		LOGICAL_RRX_REG \
  1.1587 +	} \
  1.1588 +	if (dest == 15) { \
  1.1589 +		OPCODE2 \
  1.1590 +		/* todo */ \
  1.1591 +		if (opcode & 0x00100000) { \
  1.1592 +			clockTicks++; \
  1.1593 +			CPUSwitchMode(reg[17].I & 0x1f, false); \
  1.1594 +		} \
  1.1595 +		if (armState) { \
  1.1596 +			reg[15].I &= 0xFFFFFFFC; \
  1.1597 +			armNextPC  = reg[15].I; \
  1.1598 +			reg[15].I += 4; \
  1.1599 +		} else { \
  1.1600 +			reg[15].I &= 0xFFFFFFFE; \
  1.1601 +			armNextPC  = reg[15].I; \
  1.1602 +			reg[15].I += 2; \
  1.1603 +		} \
  1.1604 +	} else { \
  1.1605 +		OPCODE \
  1.1606 +	} \
  1.1607 +} \
  1.1608 +break; \
  1.1609 +case BASE + 1: \
  1.1610 +{ \
  1.1611 +	/* OP Rd,Rb,Rm LSL Rs */ \
  1.1612 +	clockTicks++; \
  1.1613 +	int	 shift = reg[(opcode >> 8) & 15].B.B0; \
  1.1614 +	int	 dest  = (opcode >> 12) & 15; \
  1.1615 +	bool C_OUT = C_FLAG; \
  1.1616 +	u32	 value; \
  1.1617 +	if (shift) { \
  1.1618 +		if (shift == 32) { \
  1.1619 +			value = 0; \
  1.1620 +			C_OUT = (reg[opcode & 0x0F].I & 1 ? true : false); \
  1.1621 +		} else if (shift < 32) { \
  1.1622 +			LOGICAL_LSL_REG \
  1.1623 +		} else { \
  1.1624 +			value = 0; \
  1.1625 +			C_OUT = false; \
  1.1626 +		} \
  1.1627 +	} else { \
  1.1628 +		value = reg[opcode & 0x0F].I; \
  1.1629 +	} \
  1.1630 +	if (dest == 15) { \
  1.1631 +		OPCODE2 \
  1.1632 +		/* todo */ \
  1.1633 +		if (opcode & 0x00100000) { \
  1.1634 +			clockTicks++; \
  1.1635 +			CPUSwitchMode(reg[17].I & 0x1f, false); \
  1.1636 +		} \
  1.1637 +		if (armState) { \
  1.1638 +			reg[15].I &= 0xFFFFFFFC; \
  1.1639 +			armNextPC  = reg[15].I; \
  1.1640 +			reg[15].I += 4; \
  1.1641 +		} else { \
  1.1642 +			reg[15].I &= 0xFFFFFFFE; \
  1.1643 +			armNextPC  = reg[15].I; \
  1.1644 +			reg[15].I += 2; \
  1.1645 +		} \
  1.1646 +	} else { \
  1.1647 +		OPCODE \
  1.1648 +	} \
  1.1649 +} \
  1.1650 +break; \
  1.1651 +case BASE + 3: \
  1.1652 +{ \
  1.1653 +	/* OP Rd,Rb,Rm LSR Rs */ \
  1.1654 +	clockTicks++; \
  1.1655 +	int	 shift = reg[(opcode >> 8) & 15].B.B0; \
  1.1656 +	int	 dest  = (opcode >> 12) & 15; \
  1.1657 +	bool C_OUT = C_FLAG; \
  1.1658 +	u32	 value; \
  1.1659 +	if (shift) { \
  1.1660 +		if (shift == 32) { \
  1.1661 +			value = 0; \
  1.1662 +			C_OUT = (reg[opcode & 0x0F].I & 0x80000000 ? true : false); \
  1.1663 +		} else if (shift < 32) { \
  1.1664 +			LOGICAL_LSR_REG \
  1.1665 +		} else { \
  1.1666 +			value = 0; \
  1.1667 +			C_OUT = false; \
  1.1668 +		} \
  1.1669 +	} else { \
  1.1670 +		value = reg[opcode & 0x0F].I; \
  1.1671 +	} \
  1.1672 +	if (dest == 15) { \
  1.1673 +		OPCODE2 \
  1.1674 +		/* todo */ \
  1.1675 +		if (opcode & 0x00100000) { \
  1.1676 +			clockTicks++; \
  1.1677 +			CPUSwitchMode(reg[17].I & 0x1f, false); \
  1.1678 +		} \
  1.1679 +		if (armState) { \
  1.1680 +			reg[15].I &= 0xFFFFFFFC; \
  1.1681 +			armNextPC  = reg[15].I; \
  1.1682 +			reg[15].I += 4; \
  1.1683 +		} else { \
  1.1684 +			reg[15].I &= 0xFFFFFFFE; \
  1.1685 +			armNextPC  = reg[15].I; \
  1.1686 +			reg[15].I += 2; \
  1.1687 +		} \
  1.1688 +	} else { \
  1.1689 +		OPCODE \
  1.1690 +	} \
  1.1691 +} \
  1.1692 +break; \
  1.1693 +case BASE + 5: \
  1.1694 +{ \
  1.1695 +	/* OP Rd,Rb,Rm ASR Rs */ \
  1.1696 +	clockTicks++; \
  1.1697 +	int	 shift = reg[(opcode >> 8) & 15].B.B0; \
  1.1698 +	int	 dest  = (opcode >> 12) & 15; \
  1.1699 +	bool C_OUT = C_FLAG; \
  1.1700 +	u32	 value; \
  1.1701 +	if (shift < 32) { \
  1.1702 +		if (shift) { \
  1.1703 +			LOGICAL_ASR_REG \
  1.1704 +		} else { \
  1.1705 +			value = reg[opcode & 0x0F].I; \
  1.1706 +		} \
  1.1707 +	} else { \
  1.1708 +		if (reg[opcode & 0x0F].I & 0x80000000) { \
  1.1709 +			value = 0xFFFFFFFF; \
  1.1710 +			C_OUT = true; \
  1.1711 +		} else { \
  1.1712 +			value = 0; \
  1.1713 +			C_OUT = false; \
  1.1714 +		} \
  1.1715 +	} \
  1.1716 +	if (dest == 15) { \
  1.1717 +		OPCODE2 \
  1.1718 +		/* todo */ \
  1.1719 +		if (opcode & 0x00100000) { \
  1.1720 +			clockTicks++; \
  1.1721 +			CPUSwitchMode(reg[17].I & 0x1f, false); \
  1.1722 +		} \
  1.1723 +		if (armState) { \
  1.1724 +			reg[15].I &= 0xFFFFFFFC; \
  1.1725 +			armNextPC  = reg[15].I; \
  1.1726 +			reg[15].I += 4; \
  1.1727 +		} else { \
  1.1728 +			reg[15].I &= 0xFFFFFFFE; \
  1.1729 +			armNextPC  = reg[15].I; \
  1.1730 +			reg[15].I += 2; \
  1.1731 +		} \
  1.1732 +	} else { \
  1.1733 +		OPCODE \
  1.1734 +	} \
  1.1735 +} \
  1.1736 +break; \
  1.1737 +case BASE + 7: \
  1.1738 +{ \
  1.1739 +	/* OP Rd,Rb,Rm ROR Rs */ \
  1.1740 +	clockTicks++; \
  1.1741 +	int	 shift = reg[(opcode >> 8) & 15].B.B0; \
  1.1742 +	int	 dest  = (opcode >> 12) & 15; \
  1.1743 +	bool C_OUT = C_FLAG; \
  1.1744 +	u32	 value; \
  1.1745 +	if (shift) { \
  1.1746 +		shift &= 0x1f; \
  1.1747 +		if (shift) { \
  1.1748 +			LOGICAL_ROR_REG \
  1.1749 +		} else { \
  1.1750 +			value = reg[opcode & 0x0F].I; \
  1.1751 +			C_OUT = (value & 0x80000000 ? true : false); \
  1.1752 +		} \
  1.1753 +	} else { \
  1.1754 +		value = reg[opcode & 0x0F].I; \
  1.1755 +		C_OUT = (value & 0x80000000 ? true : false); \
  1.1756 +	} \
  1.1757 +	if (dest == 15) { \
  1.1758 +		OPCODE2 \
  1.1759 +		/* todo */ \
  1.1760 +		if (opcode & 0x00100000) { \
  1.1761 +			clockTicks++; \
  1.1762 +			CPUSwitchMode(reg[17].I & 0x1f, false); \
  1.1763 +		} \
  1.1764 +		if (armState) { \
  1.1765 +			reg[15].I &= 0xFFFFFFFC; \
  1.1766 +			armNextPC  = reg[15].I; \
  1.1767 +			reg[15].I += 4; \
  1.1768 +		} else { \
  1.1769 +			reg[15].I &= 0xFFFFFFFE; \
  1.1770 +			armNextPC  = reg[15].I; \
  1.1771 +			reg[15].I += 2; \
  1.1772 +		} \
  1.1773 +	} else { \
  1.1774 +		OPCODE \
  1.1775 +	} \
  1.1776 +} \
  1.1777 +break; \
  1.1778 +case BASE + 0x200: \
  1.1779 +case BASE + 0x201: \
  1.1780 +case BASE + 0x202: \
  1.1781 +case BASE + 0x203: \
  1.1782 +case BASE + 0x204: \
  1.1783 +case BASE + 0x205: \
  1.1784 +case BASE + 0x206: \
  1.1785 +case BASE + 0x207: \
  1.1786 +case BASE + 0x208: \
  1.1787 +case BASE + 0x209: \
  1.1788 +case BASE + 0x20a: \
  1.1789 +case BASE + 0x20b: \
  1.1790 +case BASE + 0x20c: \
  1.1791 +case BASE + 0x20d: \
  1.1792 +case BASE + 0x20e: \
  1.1793 +case BASE + 0x20f: \
  1.1794 +{ \
  1.1795 +	int	 shift = (opcode & 0xF00) >> 7; \
  1.1796 +	int	 dest  = (opcode >> 12) & 0x0F; \
  1.1797 +	bool C_OUT = C_FLAG; \
  1.1798 +	u32	 value; \
  1.1799 +	if (shift) { \
  1.1800 +		LOGICAL_ROR_IMM \
  1.1801 +	} else { \
  1.1802 +		value = opcode & 0xff; \
  1.1803 +	} \
  1.1804 +	if (dest == 15) { \
  1.1805 +		OPCODE2 \
  1.1806 +		/* todo */ \
  1.1807 +		if (opcode & 0x00100000) { \
  1.1808 +			clockTicks++; \
  1.1809 +			CPUSwitchMode(reg[17].I & 0x1f, false); \
  1.1810 +		} \
  1.1811 +		if (armState) { \
  1.1812 +			reg[15].I &= 0xFFFFFFFC; \
  1.1813 +			armNextPC  = reg[15].I; \
  1.1814 +			reg[15].I += 4; \
  1.1815 +		} else { \
  1.1816 +			reg[15].I &= 0xFFFFFFFE; \
  1.1817 +			armNextPC  = reg[15].I; \
  1.1818 +			reg[15].I += 2; \
  1.1819 +		} \
  1.1820 +	} else { \
  1.1821 +		OPCODE \
  1.1822 +	} \
  1.1823 +} \
  1.1824 +break;
  1.1825 +
  1.1826 +#define ARITHMETIC_DATA_OPCODE(OPCODE, OPCODE2, BASE) \
  1.1827 +case BASE: \
  1.1828 +case BASE + 8: \
  1.1829 +{ \
  1.1830 +	/* OP Rd,Rb,Rm LSL # */ \
  1.1831 +	int base  = (opcode >> 16) & 0x0F; \
  1.1832 +	int shift = (opcode >> 7) & 0x1F; \
  1.1833 +	int dest  = (opcode >> 12) & 15; \
  1.1834 +	u32 value; \
  1.1835 +	if (shift) { \
  1.1836 +		ARITHMETIC_LSL_REG \
  1.1837 +	} else { \
  1.1838 +		value = reg[opcode & 0x0F].I; \
  1.1839 +	} \
  1.1840 +	if (dest == 15) { \
  1.1841 +		OPCODE2 \
  1.1842 +		/* todo */ \
  1.1843 +		if (opcode & 0x00100000) { \
  1.1844 +			clockTicks++; \
  1.1845 +			CPUSwitchMode(reg[17].I & 0x1f, false); \
  1.1846 +		} \
  1.1847 +		if (armState) { \
  1.1848 +			reg[15].I &= 0xFFFFFFFC; \
  1.1849 +			armNextPC  = reg[15].I; \
  1.1850 +			reg[15].I += 4; \
  1.1851 +		} else { \
  1.1852 +			reg[15].I &= 0xFFFFFFFE; \
  1.1853 +			armNextPC  = reg[15].I; \
  1.1854 +			reg[15].I += 2; \
  1.1855 +		} \
  1.1856 +	} else { \
  1.1857 +		OPCODE \
  1.1858 +	} \
  1.1859 +} \
  1.1860 +break; \
  1.1861 +case BASE + 2: \
  1.1862 +case BASE + 10: \
  1.1863 +{ \
  1.1864 +	/* OP Rd,Rb,Rm LSR # */ \
  1.1865 +	int base  = (opcode >> 16) & 0x0F; \
  1.1866 +	int shift = (opcode >> 7) & 0x1F; \
  1.1867 +	int dest  = (opcode >> 12) & 15; \
  1.1868 +	u32 value; \
  1.1869 +	if (shift) { \
  1.1870 +		ARITHMETIC_LSR_REG \
  1.1871 +	} else { \
  1.1872 +		value = 0; \
  1.1873 +	} \
  1.1874 +	if (dest == 15) { \
  1.1875 +		OPCODE2 \
  1.1876 +		/* todo */ \
  1.1877 +		if (opcode & 0x00100000) { \
  1.1878 +			clockTicks++; \
  1.1879 +			CPUSwitchMode(reg[17].I & 0x1f, false); \
  1.1880 +		} \
  1.1881 +		if (armState) { \
  1.1882 +			reg[15].I &= 0xFFFFFFFC; \
  1.1883 +			armNextPC  = reg[15].I; \
  1.1884 +			reg[15].I += 4; \
  1.1885 +		} else { \
  1.1886 +			reg[15].I &= 0xFFFFFFFE; \
  1.1887 +			armNextPC  = reg[15].I; \
  1.1888 +			reg[15].I += 2; \
  1.1889 +		} \
  1.1890 +	} else { \
  1.1891 +		OPCODE \
  1.1892 +	} \
  1.1893 +} \
  1.1894 +break; \
  1.1895 +case BASE + 4: \
  1.1896 +case BASE + 12: \
  1.1897 +{ \
  1.1898 +	/* OP Rd,Rb,Rm ASR # */ \
  1.1899 +	int base  = (opcode >> 16) & 0x0F; \
  1.1900 +	int shift = (opcode >> 7) & 0x1F; \
  1.1901 +	int dest  = (opcode >> 12) & 15; \
  1.1902 +	u32 value; \
  1.1903 +	if (shift) { \
  1.1904 +		ARITHMETIC_ASR_REG \
  1.1905 +	} else { \
  1.1906 +		if (reg[opcode & 0x0F].I & 0x80000000) { \
  1.1907 +			value = 0xFFFFFFFF; \
  1.1908 +		} else value = 0; \
  1.1909 +	} \
  1.1910 +	if (dest == 15) { \
  1.1911 +		OPCODE2 \
  1.1912 +		/* todo */ \
  1.1913 +		if (opcode & 0x00100000) { \
  1.1914 +			clockTicks++; \
  1.1915 +			CPUSwitchMode(reg[17].I & 0x1f, false); \
  1.1916 +		} \
  1.1917 +		if (armState) { \
  1.1918 +			reg[15].I &= 0xFFFFFFFC; \
  1.1919 +			armNextPC  = reg[15].I; \
  1.1920 +			reg[15].I += 4; \
  1.1921 +		} else { \
  1.1922 +			reg[15].I &= 0xFFFFFFFE; \
  1.1923 +			armNextPC  = reg[15].I; \
  1.1924 +			reg[15].I += 2; \
  1.1925 +		} \
  1.1926 +	} else { \
  1.1927 +		OPCODE \
  1.1928 +	} \
  1.1929 +} \
  1.1930 +break; \
  1.1931 +case BASE + 6: \
  1.1932 +case BASE + 14: \
  1.1933 +{ \
  1.1934 +	/* OP Rd,Rb,Rm ROR # */ \
  1.1935 +	int base  = (opcode >> 16) & 0x0F; \
  1.1936 +	int shift = (opcode >> 7) & 0x1F; \
  1.1937 +	int dest  = (opcode >> 12) & 15; \
  1.1938 +	u32 value; \
  1.1939 +	if (shift) { \
  1.1940 +		ARITHMETIC_ROR_REG \
  1.1941 +	} else { \
  1.1942 +		ARITHMETIC_RRX_REG \
  1.1943 +	} \
  1.1944 +	if (dest == 15) { \
  1.1945 +		OPCODE2 \
  1.1946 +		/* todo */ \
  1.1947 +		if (opcode & 0x00100000) { \
  1.1948 +			clockTicks++; \
  1.1949 +			CPUSwitchMode(reg[17].I & 0x1f, false); \
  1.1950 +		} \
  1.1951 +		if (armState) { \
  1.1952 +			reg[15].I &= 0xFFFFFFFC; \
  1.1953 +			armNextPC  = reg[15].I; \
  1.1954 +			reg[15].I += 4; \
  1.1955 +		} else { \
  1.1956 +			reg[15].I &= 0xFFFFFFFE; \
  1.1957 +			armNextPC  = reg[15].I; \
  1.1958 +			reg[15].I += 2; \
  1.1959 +		} \
  1.1960 +	} else { \
  1.1961 +		OPCODE \
  1.1962 +	} \
  1.1963 +} \
  1.1964 +break; \
  1.1965 +case BASE + 1: \
  1.1966 +{ \
  1.1967 +	/* OP Rd,Rb,Rm LSL Rs */ \
  1.1968 +	clockTicks++; \
  1.1969 +	int base  = (opcode >> 16) & 0x0F; \
  1.1970 +	int shift = reg[(opcode >> 8) & 15].B.B0; \
  1.1971 +	int dest  = (opcode >> 12) & 15; \
  1.1972 +	u32 value; \
  1.1973 +	if (shift) { \
  1.1974 +		if (shift == 32) { \
  1.1975 +			value = 0; \
  1.1976 +		} else if (shift < 32) { \
  1.1977 +			ARITHMETIC_LSL_REG \
  1.1978 +		} else value = 0; \
  1.1979 +	} else { \
  1.1980 +		value = reg[opcode & 0x0F].I; \
  1.1981 +	} \
  1.1982 +	if (dest == 15) { \
  1.1983 +		OPCODE2 \
  1.1984 +		/* todo */ \
  1.1985 +		if (opcode & 0x00100000) { \
  1.1986 +			clockTicks++; \
  1.1987 +			CPUSwitchMode(reg[17].I & 0x1f, false); \
  1.1988 +		} \
  1.1989 +		if (armState) { \
  1.1990 +			reg[15].I &= 0xFFFFFFFC; \
  1.1991 +			armNextPC  = reg[15].I; \
  1.1992 +			reg[15].I += 4; \
  1.1993 +		} else { \
  1.1994 +			reg[15].I &= 0xFFFFFFFE; \
  1.1995 +			armNextPC  = reg[15].I; \
  1.1996 +			reg[15].I += 2; \
  1.1997 +		} \
  1.1998 +	} else { \
  1.1999 +		OPCODE \
  1.2000 +	} \
  1.2001 +} \
  1.2002 +break; \
  1.2003 +case BASE + 3: \
  1.2004 +{ \
  1.2005 +	/* OP Rd,Rb,Rm LSR Rs */ \
  1.2006 +	clockTicks++; \
  1.2007 +	int base  = (opcode >> 16) & 0x0F; \
  1.2008 +	int shift = reg[(opcode >> 8) & 15].B.B0; \
  1.2009 +	int dest  = (opcode >> 12) & 15; \
  1.2010 +	u32 value; \
  1.2011 +	if (shift) { \
  1.2012 +		if (shift == 32) { \
  1.2013 +			value = 0; \
  1.2014 +		} else if (shift < 32) { \
  1.2015 +			ARITHMETIC_LSR_REG \
  1.2016 +		} else value = 0; \
  1.2017 +	} else { \
  1.2018 +		value = reg[opcode & 0x0F].I; \
  1.2019 +	} \
  1.2020 +	if (dest == 15) { \
  1.2021 +		OPCODE2 \
  1.2022 +		/* todo */ \
  1.2023 +		if (opcode & 0x00100000) { \
  1.2024 +			clockTicks++; \
  1.2025 +			CPUSwitchMode(reg[17].I & 0x1f, false); \
  1.2026 +		} \
  1.2027 +		if (armState) { \
  1.2028 +			reg[15].I &= 0xFFFFFFFC; \
  1.2029 +			armNextPC  = reg[15].I; \
  1.2030 +			reg[15].I += 4; \
  1.2031 +		} else { \
  1.2032 +			reg[15].I &= 0xFFFFFFFE; \
  1.2033 +			armNextPC  = reg[15].I; \
  1.2034 +			reg[15].I += 2; \
  1.2035 +		} \
  1.2036 +	} else { \
  1.2037 +		OPCODE \
  1.2038 +	} \
  1.2039 +} \
  1.2040 +break; \
  1.2041 +case BASE + 5: \
  1.2042 +{ \
  1.2043 +	/* OP Rd,Rb,Rm ASR Rs */ \
  1.2044 +	clockTicks++; \
  1.2045 +	int base  = (opcode >> 16) & 0x0F; \
  1.2046 +	int shift = reg[(opcode >> 8) & 15].B.B0; \
  1.2047 +	int dest  = (opcode >> 12) & 15; \
  1.2048 +	u32 value; \
  1.2049 +	if (shift < 32) { \
  1.2050 +		if (shift) { \
  1.2051 +			ARITHMETIC_ASR_REG \
  1.2052 +		} else { \
  1.2053 +			value = reg[opcode & 0x0F].I; \
  1.2054 +		} \
  1.2055 +	} else { \
  1.2056 +		if (reg[opcode & 0x0F].I & 0x80000000) { \
  1.2057 +			value = 0xFFFFFFFF; \
  1.2058 +		} else value = 0; \
  1.2059 +	} \
  1.2060 +	if (dest == 15) { \
  1.2061 +		OPCODE2 \
  1.2062 +		/* todo */ \
  1.2063 +		if (opcode & 0x00100000) { \
  1.2064 +			clockTicks++; \
  1.2065 +			CPUSwitchMode(reg[17].I & 0x1f, false); \
  1.2066 +		} \
  1.2067 +		if (armState) { \
  1.2068 +			reg[15].I &= 0xFFFFFFFC; \
  1.2069 +			armNextPC  = reg[15].I; \
  1.2070 +			reg[15].I += 4; \
  1.2071 +		} else { \
  1.2072 +			reg[15].I &= 0xFFFFFFFE; \
  1.2073 +			armNextPC  = reg[15].I; \
  1.2074 +			reg[15].I += 2; \
  1.2075 +		} \
  1.2076 +	} else { \
  1.2077 +		OPCODE \
  1.2078 +	} \
  1.2079 +} \
  1.2080 +break; \
  1.2081 +case BASE + 7: \
  1.2082 +{ \
  1.2083 +	/* OP Rd,Rb,Rm ROR Rs */ \
  1.2084 +	clockTicks++; \
  1.2085 +	int base  = (opcode >> 16) & 0x0F; \
  1.2086 +	int shift = reg[(opcode >> 8) & 15].B.B0; \
  1.2087 +	int dest  = (opcode >> 12) & 15; \
  1.2088 +	u32 value; \
  1.2089 +	if (shift) { \
  1.2090 +		shift &= 0x1f; \
  1.2091 +		if (shift) { \
  1.2092 +			ARITHMETIC_ROR_REG \
  1.2093 +		} else { \
  1.2094 +			value = reg[opcode & 0x0F].I; \
  1.2095 +		} \
  1.2096 +	} else { \
  1.2097 +		value = reg[opcode & 0x0F].I; \
  1.2098 +	} \
  1.2099 +	if (dest == 15) { \
  1.2100 +		OPCODE2 \
  1.2101 +		/* todo */ \
  1.2102 +		if (opcode & 0x00100000) { \
  1.2103 +			clockTicks++; \
  1.2104 +			CPUSwitchMode(reg[17].I & 0x1f, false); \
  1.2105 +		} \
  1.2106 +		if (armState) { \
  1.2107 +			reg[15].I &= 0xFFFFFFFC; \
  1.2108 +			armNextPC  = reg[15].I; \
  1.2109 +			reg[15].I += 4; \
  1.2110 +		} else { \
  1.2111 +			reg[15].I &= 0xFFFFFFFE; \
  1.2112 +			armNextPC  = reg[15].I; \
  1.2113 +			reg[15].I += 2; \
  1.2114 +		} \
  1.2115 +	} else { \
  1.2116 +		OPCODE \
  1.2117 +	} \
  1.2118 +} \
  1.2119 +break; \
  1.2120 +case BASE + 0x200: \
  1.2121 +case BASE + 0x201: \
  1.2122 +case BASE + 0x202: \
  1.2123 +case BASE + 0x203: \
  1.2124 +case BASE + 0x204: \
  1.2125 +case BASE + 0x205: \
  1.2126 +case BASE + 0x206: \
  1.2127 +case BASE + 0x207: \
  1.2128 +case BASE + 0x208: \
  1.2129 +case BASE + 0x209: \
  1.2130 +case BASE + 0x20a: \
  1.2131 +case BASE + 0x20b: \
  1.2132 +case BASE + 0x20c: \
  1.2133 +case BASE + 0x20d: \
  1.2134 +case BASE + 0x20e: \
  1.2135 +case BASE + 0x20f: \
  1.2136 +{ \
  1.2137 +	int shift = (opcode & 0xF00) >> 7; \
  1.2138 +	int base  = (opcode >> 16) & 0x0F; \
  1.2139 +	int dest  = (opcode >> 12) & 0x0F; \
  1.2140 +	u32 value; \
  1.2141 +	{ \
  1.2142 +		ARITHMETIC_ROR_IMM \
  1.2143 +	} \
  1.2144 +	if (dest == 15) { \
  1.2145 +		OPCODE2 \
  1.2146 +		/* todo */ \
  1.2147 +		if (opcode & 0x00100000) { \
  1.2148 +			clockTicks++; \
  1.2149 +			CPUSwitchMode(reg[17].I & 0x1f, false); \
  1.2150 +		} \
  1.2151 +		if (armState) { \
  1.2152 +			reg[15].I &= 0xFFFFFFFC; \
  1.2153 +			armNextPC  = reg[15].I; \
  1.2154 +			reg[15].I += 4; \
  1.2155 +		} else { \
  1.2156 +			reg[15].I &= 0xFFFFFFFE; \
  1.2157 +			armNextPC  = reg[15].I; \
  1.2158 +			reg[15].I += 2; \
  1.2159 +		} \
  1.2160 +	} else { \
  1.2161 +		OPCODE \
  1.2162 +	} \
  1.2163 +} \
  1.2164 +break;
  1.2165 +
  1.2166 +u32 opcode = CPUReadMemoryQuick(armNextPC);
  1.2167 +
  1.2168 +clockTicks = memoryWaitFetch32[(armNextPC >> 24) & 15];
  1.2169 +
  1.2170 +#ifndef FINAL_VERSION
  1.2171 +if (armNextPC == stop)
  1.2172 +{
  1.2173 +	armNextPC++;
  1.2174 +}
  1.2175 +#endif
  1.2176 +
  1.2177 +armNextPC  = reg[15].I;
  1.2178 +reg[15].I += 4;
  1.2179 +int cond = opcode >> 28;
  1.2180 +// suggested optimization for frequent cases
  1.2181 +bool cond_res;
  1.2182 +if (cond == 0x0e)
  1.2183 +{
  1.2184 +	cond_res = true;
  1.2185 +}
  1.2186 +else
  1.2187 +{
  1.2188 +	switch (cond)
  1.2189 +	{
  1.2190 +	case 0x00: // EQ
  1.2191 +		cond_res = Z_FLAG;
  1.2192 +		break;
  1.2193 +	case 0x01: // NE
  1.2194 +		cond_res = !Z_FLAG;
  1.2195 +		break;
  1.2196 +	case 0x02: // CS
  1.2197 +		cond_res = C_FLAG;
  1.2198 +		break;
  1.2199 +	case 0x03: // CC
  1.2200 +		cond_res = !C_FLAG;
  1.2201 +		break;
  1.2202 +	case 0x04: // MI
  1.2203 +		cond_res = N_FLAG;
  1.2204 +		break;
  1.2205 +	case 0x05: // PL
  1.2206 +		cond_res = !N_FLAG;
  1.2207 +		break;
  1.2208 +	case 0x06: // VS
  1.2209 +		cond_res = V_FLAG;
  1.2210 +		break;
  1.2211 +	case 0x07: // VC
  1.2212 +		cond_res = !V_FLAG;
  1.2213 +		break;
  1.2214 +	case 0x08: // HI
  1.2215 +		cond_res = C_FLAG && !Z_FLAG;
  1.2216 +		break;
  1.2217 +	case 0x09: // LS
  1.2218 +		cond_res = !C_FLAG || Z_FLAG;
  1.2219 +		break;
  1.2220 +	case 0x0A: // GE
  1.2221 +		cond_res = N_FLAG == V_FLAG;
  1.2222 +		break;
  1.2223 +	case 0x0B: // LT
  1.2224 +		cond_res = N_FLAG != V_FLAG;
  1.2225 +		break;
  1.2226 +	case 0x0C: // GT
  1.2227 +		cond_res = !Z_FLAG && (N_FLAG == V_FLAG);
  1.2228 +		break;
  1.2229 +	case 0x0D: // LE
  1.2230 +		cond_res = Z_FLAG || (N_FLAG != V_FLAG);
  1.2231 +		break;
  1.2232 +	case 0x0E:
  1.2233 +		cond_res = true;
  1.2234 +		break;
  1.2235 +	case 0x0F:
  1.2236 +	default:
  1.2237 +		// ???
  1.2238 +		cond_res = false;
  1.2239 +		break;
  1.2240 +	}
  1.2241 +}
  1.2242 +
  1.2243 +if (cond_res)
  1.2244 +{
  1.2245 +	switch (((opcode >> 16) & 0xFF0) | ((opcode >> 4) & 0x0F))
  1.2246 +	{
  1.2247 +		LOGICAL_DATA_OPCODE_WITHOUT_base(OP_AND,  OP_AND, 0x000);
  1.2248 +		LOGICAL_DATA_OPCODE_WITHOUT_base(OP_ANDS, OP_AND, 0x010);
  1.2249 +	case 0x009:
  1.2250 +	{
  1.2251 +		// MUL Rd, Rm, Rs
  1.2252 +		int dest = (opcode >> 16) & 0x0F;
  1.2253 +		int mult = (opcode & 0x0F);
  1.2254 +		u32 rs	 = reg[(opcode >> 8) & 0x0F].I;
  1.2255 +		reg[dest].I = reg[mult].I * rs;
  1.2256 +		if (((s32)rs) < 0)
  1.2257 +			rs = ~rs;
  1.2258 +		if ((rs & 0xFFFFFF00) == 0)
  1.2259 +			clockTicks += 2;
  1.2260 +		else if ((rs & 0xFFFF0000) == 0)
  1.2261 +			clockTicks += 3;
  1.2262 +		else if ((rs & 0xFF000000) == 0)
  1.2263 +			clockTicks += 4;
  1.2264 +		else
  1.2265 +			clockTicks += 5;
  1.2266 +	}
  1.2267 +	break;
  1.2268 +	case 0x019:
  1.2269 +	{
  1.2270 +		// MULS Rd, Rm, Rs
  1.2271 +		int dest = (opcode >> 16) & 0x0F;
  1.2272 +		int mult = (opcode & 0x0F);
  1.2273 +		u32 rs	 = reg[(opcode >> 8) & 0x0F].I;
  1.2274 +		reg[dest].I = reg[mult].I * rs;
  1.2275 +		N_FLAG		= (reg[dest].I & 0x80000000) ? true : false;
  1.2276 +		Z_FLAG		= (reg[dest].I) ? false : true;
  1.2277 +		if (((s32)rs) < 0)
  1.2278 +			rs = ~rs;
  1.2279 +		if ((rs & 0xFFFFFF00) == 0)
  1.2280 +			clockTicks += 2;
  1.2281 +		else if ((rs & 0xFFFF0000) == 0)
  1.2282 +			clockTicks += 3;
  1.2283 +		else if ((rs & 0xFF000000) == 0)
  1.2284 +			clockTicks += 4;
  1.2285 +		else
  1.2286 +			clockTicks += 5;
  1.2287 +	}
  1.2288 +	break;
  1.2289 +	case 0x00b:
  1.2290 +	case 0x02b:
  1.2291 +	{
  1.2292 +		// STRH Rd, [Rn], -Rm
  1.2293 +		int base	= (opcode >> 16) & 0x0F;
  1.2294 +		int dest	= (opcode >> 12) & 0x0F;
  1.2295 +		u32 address = reg[base].I;
  1.2296 +		int offset	= reg[opcode & 0x0F].I;
  1.2297 +		clockTicks += 4 + CPUUpdateTicksAccess16(address);
  1.2298 +		CPUWriteHalfWord(address, reg[dest].W.W0);
  1.2299 +		address	   -= offset;
  1.2300 +		reg[base].I = address;
  1.2301 +	}
  1.2302 +	break;
  1.2303 +	case 0x04b:
  1.2304 +	case 0x06b:
  1.2305 +	{
  1.2306 +		// STRH Rd, [Rn], #-offset
  1.2307 +		int base	= (opcode >> 16) & 0x0F;
  1.2308 +		int dest	= (opcode >> 12) & 0x0F;
  1.2309 +		u32 address = reg[base].I;
  1.2310 +		int offset	= (opcode & 0x0F) | ((opcode >> 4) & 0xF0);
  1.2311 +		clockTicks += 4 + CPUUpdateTicksAccess16(address);
  1.2312 +		CPUWriteHalfWord(address, reg[dest].W.W0);
  1.2313 +		address	   -= offset;
  1.2314 +		reg[base].I = address;
  1.2315 +	}
  1.2316 +	break;
  1.2317 +	case 0x08b:
  1.2318 +	case 0x0ab:
  1.2319 +	{
  1.2320 +		// STRH Rd, [Rn], Rm
  1.2321 +		int base	= (opcode >> 16) & 0x0F;
  1.2322 +		int dest	= (opcode >> 12) & 0x0F;
  1.2323 +		u32 address = reg[base].I;
  1.2324 +		int offset	= reg[opcode & 0x0F].I;
  1.2325 +		clockTicks += 4 + CPUUpdateTicksAccess16(address);
  1.2326 +		CPUWriteHalfWord(address, reg[dest].W.W0);
  1.2327 +		address	   += offset;
  1.2328 +		reg[base].I = address;
  1.2329 +	}
  1.2330 +	break;
  1.2331 +	case 0x0cb:
  1.2332 +	case 0x0eb:
  1.2333 +	{
  1.2334 +		// STRH Rd, [Rn], #offset
  1.2335 +		int base	= (opcode >> 16) & 0x0F;
  1.2336 +		int dest	= (opcode >> 12) & 0x0F;
  1.2337 +		u32 address = reg[base].I;
  1.2338 +		int offset	= (opcode & 0x0F) | ((opcode >> 4) & 0xF0);
  1.2339 +		clockTicks += 4 + CPUUpdateTicksAccess16(address);
  1.2340 +		CPUWriteHalfWord(address, reg[dest].W.W0);
  1.2341 +		address	   += offset;
  1.2342 +		reg[base].I = address;
  1.2343 +	}
  1.2344 +	break;
  1.2345 +	case 0x10b:
  1.2346 +	{
  1.2347 +		// STRH Rd, [Rn, -Rm]
  1.2348 +		int base	= (opcode >> 16) & 0x0F;
  1.2349 +		int dest	= (opcode >> 12) & 0x0F;
  1.2350 +		u32 address = reg[base].I - reg[opcode & 0x0F].I;
  1.2351 +		clockTicks += 4 + CPUUpdateTicksAccess16(address);
  1.2352 +		CPUWriteHalfWord(address, reg[dest].W.W0);
  1.2353 +	}
  1.2354 +	break;
  1.2355 +	case 0x12b:
  1.2356 +	{
  1.2357 +		// STRH Rd, [Rn, -Rm]!
  1.2358 +		int base	= (opcode >> 16) & 0x0F;
  1.2359 +		int dest	= (opcode >> 12) & 0x0F;
  1.2360 +		u32 address = reg[base].I - reg[opcode & 0x0F].I;
  1.2361 +		clockTicks += 4 + CPUUpdateTicksAccess16(address);
  1.2362 +		CPUWriteHalfWord(address, reg[dest].W.W0);
  1.2363 +		reg[base].I = address;
  1.2364 +	}
  1.2365 +	break;
  1.2366 +	case 0x14b:
  1.2367 +	{
  1.2368 +		// STRH Rd, [Rn, -#offset]
  1.2369 +		int base	= (opcode >> 16) & 0x0F;
  1.2370 +		int dest	= (opcode >> 12) & 0x0F;
  1.2371 +		u32 address = reg[base].I - ((opcode & 0x0F) | ((opcode >> 4) & 0xF0));
  1.2372 +		clockTicks += 4 + CPUUpdateTicksAccess16(address);
  1.2373 +		CPUWriteHalfWord(address, reg[dest].W.W0);
  1.2374 +	}
  1.2375 +	break;
  1.2376 +	case 0x16b:
  1.2377 +	{
  1.2378 +		// STRH Rd, [Rn, -#offset]!
  1.2379 +		int base	= (opcode >> 16) & 0x0F;
  1.2380 +		int dest	= (opcode >> 12) & 0x0F;
  1.2381 +		u32 address = reg[base].I - ((opcode & 0x0F) | ((opcode >> 4) & 0xF0));
  1.2382 +		clockTicks += 4 + CPUUpdateTicksAccess16(address);
  1.2383 +		CPUWriteHalfWord(address, reg[dest].W.W0);
  1.2384 +		reg[base].I = address;
  1.2385 +	}
  1.2386 +	break;
  1.2387 +	case 0x18b:
  1.2388 +	{
  1.2389 +		// STRH Rd, [Rn, Rm]
  1.2390 +		int base	= (opcode >> 16) & 0x0F;
  1.2391 +		int dest	= (opcode >> 12) & 0x0F;
  1.2392 +		u32 address = reg[base].I + reg[opcode & 0x0F].I;
  1.2393 +		clockTicks += 4 + CPUUpdateTicksAccess16(address);
  1.2394 +		CPUWriteHalfWord(address, reg[dest].W.W0);
  1.2395 +	}
  1.2396 +	break;
  1.2397 +	case 0x1ab:
  1.2398 +	{
  1.2399 +		// STRH Rd, [Rn, Rm]!
  1.2400 +		int base	= (opcode >> 16) & 0x0F;
  1.2401 +		int dest	= (opcode >> 12) & 0x0F;
  1.2402 +		u32 address = reg[base].I + reg[opcode & 0x0F].I;
  1.2403 +		clockTicks += 4 + CPUUpdateTicksAccess16(address);
  1.2404 +		CPUWriteHalfWord(address, reg[dest].W.W0);
  1.2405 +		reg[base].I = address;
  1.2406 +	}
  1.2407 +	break;
  1.2408 +	case 0x1cb:
  1.2409 +	{
  1.2410 +		// STRH Rd, [Rn, #offset]
  1.2411 +		int base	= (opcode >> 16) & 0x0F;
  1.2412 +		int dest	= (opcode >> 12) & 0x0F;
  1.2413 +		u32 address = reg[base].I + ((opcode & 0x0F) | ((opcode >> 4) & 0xF0));
  1.2414 +		clockTicks += 4 + CPUUpdateTicksAccess16(address);
  1.2415 +		CPUWriteHalfWord(address, reg[dest].W.W0);
  1.2416 +	}
  1.2417 +	break;
  1.2418 +	case 0x1eb:
  1.2419 +	{
  1.2420 +		// STRH Rd, [Rn, #offset]!
  1.2421 +		int base	= (opcode >> 16) & 0x0F;
  1.2422 +		int dest	= (opcode >> 12) & 0x0F;
  1.2423 +		u32 address = reg[base].I + ((opcode & 0x0F) | ((opcode >> 4) & 0xF0));
  1.2424 +		clockTicks += 4 + CPUUpdateTicksAccess16(address);
  1.2425 +		CPUWriteHalfWord(address, reg[dest].W.W0);
  1.2426 +		reg[base].I = address;
  1.2427 +	}
  1.2428 +	break;
  1.2429 +	case 0x01b:
  1.2430 +	case 0x03b:
  1.2431 +	{
  1.2432 +		// LDRH Rd, [Rn], -Rm
  1.2433 +		int base	= (opcode >> 16) & 0x0F;
  1.2434 +		int dest	= (opcode >> 12) & 0x0F;
  1.2435 +		u32 address = reg[base].I;
  1.2436 +		int offset	= reg[opcode & 0x0F].I;
  1.2437 +		clockTicks += 3 + CPUUpdateTicksAccess16(address);
  1.2438 +		reg[dest].I = CPUReadHalfWord(address);
  1.2439 +		if (dest != base)
  1.2440 +		{
  1.2441 +			address	   -= offset;
  1.2442 +			reg[base].I = address;
  1.2443 +		}
  1.2444 +	}
  1.2445 +	break;
  1.2446 +	case 0x05b:
  1.2447 +	case 0x07b:
  1.2448 +	{
  1.2449 +		// LDRH Rd, [Rn], #-offset
  1.2450 +		int base	= (opcode >> 16) & 0x0F;
  1.2451 +		int dest	= (opcode >> 12) & 0x0F;
  1.2452 +		u32 address = reg[base].I;
  1.2453 +		int offset	= (opcode & 0x0F) | ((opcode >> 4) & 0xF0);
  1.2454 +		clockTicks += 3 + CPUUpdateTicksAccess16(address);
  1.2455 +		reg[dest].I = CPUReadHalfWord(address);
  1.2456 +		if (dest != base)
  1.2457 +		{
  1.2458 +			address	   -= offset;
  1.2459 +			reg[base].I = address;
  1.2460 +		}
  1.2461 +	}
  1.2462 +	break;
  1.2463 +	case 0x09b:
  1.2464 +	case 0x0bb:
  1.2465 +	{
  1.2466 +		// LDRH Rd, [Rn], Rm
  1.2467 +		int base	= (opcode >> 16) & 0x0F;
  1.2468 +		int dest	= (opcode >> 12) & 0x0F;
  1.2469 +		u32 address = reg[base].I;
  1.2470 +		int offset	= reg[opcode & 0x0F].I;
  1.2471 +		clockTicks += 3 + CPUUpdateTicksAccess16(address);
  1.2472 +		reg[dest].I = CPUReadHalfWord(address);
  1.2473 +		if (dest != base)
  1.2474 +		{
  1.2475 +			address	   += offset;
  1.2476 +			reg[base].I = address;
  1.2477 +		}
  1.2478 +	}
  1.2479 +	break;
  1.2480 +	case 0x0db:
  1.2481 +	case 0x0fb:
  1.2482 +	{
  1.2483 +		// LDRH Rd, [Rn], #offset
  1.2484 +		int base	= (opcode >> 16) & 0x0F;
  1.2485 +		int dest	= (opcode >> 12) & 0x0F;
  1.2486 +		u32 address = reg[base].I;
  1.2487 +		int offset	= (opcode & 0x0F) | ((opcode >> 4) & 0xF0);
  1.2488 +		clockTicks += 3 + CPUUpdateTicksAccess16(address);
  1.2489 +		reg[dest].I = CPUReadHalfWord(address);
  1.2490 +		if (dest != base)
  1.2491 +		{
  1.2492 +			address	   += offset;
  1.2493 +			reg[base].I = address;
  1.2494 +		}
  1.2495 +	}
  1.2496 +	break;
  1.2497 +	case 0x11b:
  1.2498 +	{
  1.2499 +		// LDRH Rd, [Rn, -Rm]
  1.2500 +		int base	= (opcode >> 16) & 0x0F;
  1.2501 +		int dest	= (opcode >> 12) & 0x0F;
  1.2502 +		u32 address = reg[base].I - reg[opcode & 0x0F].I;
  1.2503 +		clockTicks += 3 + CPUUpdateTicksAccess16(address);
  1.2504 +		reg[dest].I = CPUReadHalfWord(address);
  1.2505 +	}
  1.2506 +	break;
  1.2507 +	case 0x13b:
  1.2508 +	{
  1.2509 +		// LDRH Rd, [Rn, -Rm]!
  1.2510 +		int base	= (opcode >> 16) & 0x0F;
  1.2511 +		int dest	= (opcode >> 12) & 0x0F;
  1.2512 +		u32 address = reg[base].I - reg[opcode & 0x0F].I;
  1.2513 +		clockTicks += 3 + CPUUpdateTicksAccess16(address);
  1.2514 +		reg[dest].I = CPUReadHalfWord(address);
  1.2515 +		if (dest != base)
  1.2516 +			reg[base].I = address;
  1.2517 +	}
  1.2518 +	break;
  1.2519 +	case 0x15b:
  1.2520 +	{
  1.2521 +		// LDRH Rd, [Rn, -#offset]
  1.2522 +		int base	= (opcode >> 16) & 0x0F;
  1.2523 +		int dest	= (opcode >> 12) & 0x0F;
  1.2524 +		u32 address = reg[base].I - ((opcode & 0x0F) | ((opcode >> 4) & 0xF0));
  1.2525 +		clockTicks += 3 + CPUUpdateTicksAccess16(address);
  1.2526 +		reg[dest].I = CPUReadHalfWord(address);
  1.2527 +	}
  1.2528 +	break;
  1.2529 +	case 0x17b:
  1.2530 +	{
  1.2531 +		// LDRH Rd, [Rn, -#offset]!
  1.2532 +		int base	= (opcode >> 16) & 0x0F;
  1.2533 +		int dest	= (opcode >> 12) & 0x0F;
  1.2534 +		u32 address = reg[base].I - ((opcode & 0x0F) | ((opcode >> 4) & 0xF0));
  1.2535 +		clockTicks += 3 + CPUUpdateTicksAccess16(address);
  1.2536 +		reg[dest].I = CPUReadHalfWord(address);
  1.2537 +		if (dest != base)
  1.2538 +			reg[base].I = address;
  1.2539 +	}
  1.2540 +	break;
  1.2541 +	case 0x19b:
  1.2542 +	{
  1.2543 +		// LDRH Rd, [Rn, Rm]
  1.2544 +		int base	= (opcode >> 16) & 0x0F;
  1.2545 +		int dest	= (opcode >> 12) & 0x0F;
  1.2546 +		u32 address = reg[base].I + reg[opcode & 0x0F].I;
  1.2547 +		clockTicks += 3 + CPUUpdateTicksAccess16(address);
  1.2548 +		reg[dest].I = CPUReadHalfWord(address);
  1.2549 +	}
  1.2550 +	break;
  1.2551 +	case 0x1bb:
  1.2552 +	{
  1.2553 +		// LDRH Rd, [Rn, Rm]!
  1.2554 +		int base	= (opcode >> 16) & 0x0F;
  1.2555 +		int dest	= (opcode >> 12) & 0x0F;
  1.2556 +		u32 address = reg[base].I + reg[opcode & 0x0F].I;
  1.2557 +		clockTicks += 3 + CPUUpdateTicksAccess16(address);
  1.2558 +		reg[dest].I = CPUReadHalfWord(address);
  1.2559 +		if (dest != base)
  1.2560 +			reg[base].I = address;
  1.2561 +	}
  1.2562 +	break;
  1.2563 +	case 0x1db:
  1.2564 +	{
  1.2565 +		// LDRH Rd, [Rn, #offset]
  1.2566 +		int base	= (opcode >> 16) & 0x0F;
  1.2567 +		int dest	= (opcode >> 12) & 0x0F;
  1.2568 +		u32 address = reg[base].I + ((opcode & 0x0F) | ((opcode >> 4) & 0xF0));
  1.2569 +		clockTicks += 3 + CPUUpdateTicksAccess16(address);
  1.2570 +		reg[dest].I = CPUReadHalfWord(address);
  1.2571 +	}
  1.2572 +	break;
  1.2573 +	case 0x1fb:
  1.2574 +	{
  1.2575 +		// LDRH Rd, [Rn, #offset]!
  1.2576 +		int base	= (opcode >> 16) & 0x0F;
  1.2577 +		int dest	= (opcode >> 12) & 0x0F;
  1.2578 +		u32 address = reg[base].I + ((opcode & 0x0F) | ((opcode >> 4) & 0xF0));
  1.2579 +		clockTicks += 3 + CPUUpdateTicksAccess16(address);
  1.2580 +		reg[dest].I = CPUReadHalfWord(address);
  1.2581 +		if (dest != base)
  1.2582 +			reg[base].I = address;
  1.2583 +	}
  1.2584 +	break;
  1.2585 +	case 0x01d:
  1.2586 +	case 0x03d:
  1.2587 +	{
  1.2588 +		// LDRSB Rd, [Rn], -Rm
  1.2589 +		int base	= (opcode >> 16) & 0x0F;
  1.2590 +		int dest	= (opcode >> 12) & 0x0F;
  1.2591 +		u32 address = reg[base].I;
  1.2592 +		int offset	= reg[opcode & 0x0F].I;
  1.2593 +		clockTicks += 3 + CPUUpdateTicksAccess16(address);
  1.2594 +		reg[dest].I = (s8)CPUReadByte(address);
  1.2595 +		if (dest != base)
  1.2596 +		{
  1.2597 +			address	   -= offset;
  1.2598 +			reg[base].I = address;
  1.2599 +		}
  1.2600 +	}
  1.2601 +	break;
  1.2602 +	case 0x05d:
  1.2603 +	case 0x07d:
  1.2604 +	{
  1.2605 +		// LDRSB Rd, [Rn], #-offset
  1.2606 +		int base	= (opcode >> 16) & 0x0F;
  1.2607 +		int dest	= (opcode >> 12) & 0x0F;
  1.2608 +		u32 address = reg[base].I;
  1.2609 +		int offset	= (opcode & 0x0F) | ((opcode >> 4) & 0xF0);
  1.2610 +		clockTicks += 3 + CPUUpdateTicksAccess16(address);
  1.2611 +		reg[dest].I = (s8)CPUReadByte(address);
  1.2612 +		if (dest != base)
  1.2613 +		{
  1.2614 +			address	   -= offset;
  1.2615 +			reg[base].I = address;
  1.2616 +		}
  1.2617 +	}
  1.2618 +	break;
  1.2619 +	case 0x09d:
  1.2620 +	case 0x0bd:
  1.2621 +	{
  1.2622 +		// LDRSB Rd, [Rn], Rm
  1.2623 +		int base	= (opcode >> 16) & 0x0F;
  1.2624 +		int dest	= (opcode >> 12) & 0x0F;
  1.2625 +		u32 address = reg[base].I;
  1.2626 +		int offset	= reg[opcode & 0x0F].I;
  1.2627 +		clockTicks += 3 + CPUUpdateTicksAccess16(address);
  1.2628 +		reg[dest].I = (s8)CPUReadByte(address);
  1.2629 +		if (dest != base)
  1.2630 +		{
  1.2631 +			address	   += offset;
  1.2632 +			reg[base].I = address;
  1.2633 +		}
  1.2634 +	}
  1.2635 +	break;
  1.2636 +	case 0x0dd:
  1.2637 +	case 0x0fd:
  1.2638 +	{
  1.2639 +		// LDRSB Rd, [Rn], #offset
  1.2640 +		int base	= (opcode >> 16) & 0x0F;
  1.2641 +		int dest	= (opcode >> 12) & 0x0F;
  1.2642 +		u32 address = reg[base].I;
  1.2643 +		int offset	= (opcode & 0x0F) | ((opcode >> 4) & 0xF0);
  1.2644 +		clockTicks += 3 + CPUUpdateTicksAccess16(address);
  1.2645 +		reg[dest].I = (s8)CPUReadByte(address);
  1.2646 +		if (dest != base)
  1.2647 +		{
  1.2648 +			address	   += offset;
  1.2649 +			reg[base].I = address;
  1.2650 +		}
  1.2651 +	}
  1.2652 +	break;
  1.2653 +	case 0x11d:
  1.2654 +	{
  1.2655 +		// LDRSB Rd, [Rn, -Rm]
  1.2656 +		int base	= (opcode >> 16) & 0x0F;
  1.2657 +		int dest	= (opcode >> 12) & 0x0F;
  1.2658 +		u32 address = reg[base].I - reg[opcode & 0x0F].I;
  1.2659 +		clockTicks += 3 + CPUUpdateTicksAccess16(address);
  1.2660 +		reg[dest].I = (s8)CPUReadByte(address);
  1.2661 +	}
  1.2662 +	break;
  1.2663 +	case 0x13d:
  1.2664 +	{
  1.2665 +		// LDRSB Rd, [Rn, -Rm]!
  1.2666 +		int base	= (opcode >> 16) & 0x0F;
  1.2667 +		int dest	= (opcode >> 12) & 0x0F;
  1.2668 +		u32 address = reg[base].I - reg[opcode & 0x0F].I;
  1.2669 +		clockTicks += 3 + CPUUpdateTicksAccess16(address);
  1.2670 +		reg[dest].I = (s8)CPUReadByte(address);
  1.2671 +		if (dest != base)
  1.2672 +			reg[base].I = address;
  1.2673 +	}
  1.2674 +	break;
  1.2675 +	case 0x15d:
  1.2676 +	{
  1.2677 +		// LDRSB Rd, [Rn, -#offset]
  1.2678 +		int base	= (opcode >> 16) & 0x0F;
  1.2679 +		int dest	= (opcode >> 12) & 0x0F;
  1.2680 +		u32 address = reg[base].I - ((opcode & 0x0F) | ((opcode >> 4) & 0xF0));
  1.2681 +		clockTicks += 3 + CPUUpdateTicksAccess16(address);
  1.2682 +		reg[dest].I = (s8)CPUReadByte(address);
  1.2683 +	}
  1.2684 +	break;
  1.2685 +	case 0x17d:
  1.2686 +	{
  1.2687 +		// LDRSB Rd, [Rn, -#offset]!
  1.2688 +		int base	= (opcode >> 16) & 0x0F;
  1.2689 +		int dest	= (opcode >> 12) & 0x0F;
  1.2690 +		u32 address = reg[base].I - ((opcode & 0x0F) | ((opcode >> 4) & 0xF0));
  1.2691 +		clockTicks += 3 + CPUUpdateTicksAccess16(address);
  1.2692 +		reg[dest].I = (s8)CPUReadByte(address);
  1.2693 +		if (dest != base)
  1.2694 +			reg[base].I = address;
  1.2695 +	}
  1.2696 +	break;
  1.2697 +	case 0x19d:
  1.2698 +	{
  1.2699 +		// LDRSB Rd, [Rn, Rm]
  1.2700 +		int base	= (opcode >> 16) & 0x0F;
  1.2701 +		int dest	= (opcode >> 12) & 0x0F;
  1.2702 +		u32 address = reg[base].I + reg[opcode & 0x0F].I;
  1.2703 +		clockTicks += 3 + CPUUpdateTicksAccess16(address);
  1.2704 +		reg[dest].I = (s8)CPUReadByte(address);
  1.2705 +	}
  1.2706 +	break;
  1.2707 +	case 0x1bd:
  1.2708 +	{
  1.2709 +		// LDRSB Rd, [Rn, Rm]!
  1.2710 +		int base	= (opcode >> 16) & 0x0F;
  1.2711 +		int dest	= (opcode >> 12) & 0x0F;
  1.2712 +		u32 address = reg[base].I + reg[opcode & 0x0F].I;
  1.2713 +		clockTicks += 3 + CPUUpdateTicksAccess16(address);
  1.2714 +		reg[dest].I = (s8)CPUReadByte(address);
  1.2715 +		if (dest != base)
  1.2716 +			reg[base].I = address;
  1.2717 +	}
  1.2718 +	break;
  1.2719 +	case 0x1dd:
  1.2720 +	{
  1.2721 +		// LDRSB Rd, [Rn, #offset]
  1.2722 +		int base	= (opcode >> 16) & 0x0F;
  1.2723 +		int dest	= (opcode >> 12) & 0x0F;
  1.2724 +		u32 address = reg[base].I + ((opcode & 0x0F) | ((opcode >> 4) & 0xF0));
  1.2725 +		clockTicks += 3 + CPUUpdateTicksAccess16(address);
  1.2726 +		reg[dest].I = (s8)CPUReadByte(address);
  1.2727 +	}
  1.2728 +	break;
  1.2729 +	case 0x1fd:
  1.2730 +	{
  1.2731 +		// LDRSB Rd, [Rn, #offset]!
  1.2732 +		int base	= (opcode >> 16) & 0x0F;
  1.2733 +		int dest	= (opcode >> 12) & 0x0F;
  1.2734 +		u32 address = reg[base].I + ((opcode & 0x0F) | ((opcode >> 4) & 0xF0));
  1.2735 +		clockTicks += 3 + CPUUpdateTicksAccess16(address);
  1.2736 +		reg[dest].I = (s8)CPUReadByte(address);
  1.2737 +		if (dest != base)
  1.2738 +			reg[base].I = address;
  1.2739 +	}
  1.2740 +	break;
  1.2741 +	case 0x01f:
  1.2742 +	case 0x03f:
  1.2743 +	{
  1.2744 +		// LDRSH Rd, [Rn], -Rm
  1.2745 +		int base	= (opcode >> 16) & 0x0F;
  1.2746 +		int dest	= (opcode >> 12) & 0x0F;
  1.2747 +		u32 address = reg[base].I;
  1.2748 +		int offset	= reg[opcode & 0x0F].I;
  1.2749 +		clockTicks += 3 + CPUUpdateTicksAccess16(address);
  1.2750 +		reg[dest].I = (s16)CPUReadHalfWordSigned(address);
  1.2751 +		if (dest != base)
  1.2752 +		{
  1.2753 +			address	   -= offset;
  1.2754 +			reg[base].I = address;
  1.2755 +		}
  1.2756 +	}
  1.2757 +	break;
  1.2758 +	case 0x05f:
  1.2759 +	case 0x07f:
  1.2760 +	{
  1.2761 +		// LDRSH Rd, [Rn], #-offset
  1.2762 +		int base	= (opcode >> 16) & 0x0F;
  1.2763 +		int dest	= (opcode >> 12) & 0x0F;
  1.2764 +		u32 address = reg[base].I;
  1.2765 +		int offset	= (opcode & 0x0F) | ((opcode >> 4) & 0xF0);
  1.2766 +		clockTicks += 3 + CPUUpdateTicksAccess16(address);
  1.2767 +		reg[dest].I = (s16)CPUReadHalfWordSigned(address);
  1.2768 +		if (dest != base)
  1.2769 +		{
  1.2770 +			address	   -= offset;
  1.2771 +			reg[base].I = address;
  1.2772 +		}
  1.2773 +	}
  1.2774 +	break;
  1.2775 +	case 0x09f:
  1.2776 +	case 0x0bf:
  1.2777 +	{
  1.2778 +		// LDRSH Rd, [Rn], Rm
  1.2779 +		int base	= (opcode >> 16) & 0x0F;
  1.2780 +		int dest	= (opcode >> 12) & 0x0F;
  1.2781 +		u32 address = reg[base].I;
  1.2782 +		int offset	= reg[opcode & 0x0F].I;
  1.2783 +		clockTicks += 3 + CPUUpdateTicksAccess16(address);
  1.2784 +		reg[dest].I = (s16)CPUReadHalfWordSigned(address);
  1.2785 +		if (dest != base)
  1.2786 +		{
  1.2787 +			address	   += offset;
  1.2788 +			reg[base].I = address;
  1.2789 +		}
  1.2790 +	}
  1.2791 +	break;
  1.2792 +	case 0x0df:
  1.2793 +	case 0x0ff:
  1.2794 +	{
  1.2795 +		// LDRSH Rd, [Rn], #offset
  1.2796 +		int base	= (opcode >> 16) & 0x0F;
  1.2797 +		int dest	= (opcode >> 12) & 0x0F;
  1.2798 +		u32 address = reg[base].I;
  1.2799 +		int offset	= (opcode & 0x0F) | ((opcode >> 4) & 0xF0);
  1.2800 +		clockTicks += 3 + CPUUpdateTicksAccess16(address);
  1.2801 +		reg[dest].I = (s16)CPUReadHalfWordSigned(address);
  1.2802 +		if (dest != base)
  1.2803 +		{
  1.2804 +			address	   += offset;
  1.2805 +			reg[base].I = address;
  1.2806 +		}
  1.2807 +	}
  1.2808 +	break;
  1.2809 +	case 0x11f:
  1.2810 +	{
  1.2811 +		// LDRSH Rd, [Rn, -Rm]
  1.2812 +		int base	= (opcode >> 16) & 0x0F;
  1.2813 +		int dest	= (opcode >> 12) & 0x0F;
  1.2814 +		u32 address = reg[base].I - reg[opcode & 0x0F].I;
  1.2815 +		clockTicks += 3 + CPUUpdateTicksAccess16(address);
  1.2816 +		reg[dest].I = (s16)CPUReadHalfWordSigned(address);
  1.2817 +	}
  1.2818 +	break;
  1.2819 +	case 0x13f:
  1.2820 +	{
  1.2821 +		// LDRSH Rd, [Rn, -Rm]!
  1.2822 +		int base	= (opcode >> 16) & 0x0F;
  1.2823 +		int dest	= (opcode >> 12) & 0x0F;
  1.2824 +		u32 address = reg[base].I - reg[opcode & 0x0F].I;
  1.2825 +		clockTicks += 3 + CPUUpdateTicksAccess16(address);
  1.2826 +		reg[dest].I = (s16)CPUReadHalfWordSigned(address);
  1.2827 +		if (dest != base)
  1.2828 +			reg[base].I = address;
  1.2829 +	}
  1.2830 +	break;
  1.2831 +	case 0x15f:
  1.2832 +	{
  1.2833 +		// LDRSH Rd, [Rn, -#offset]
  1.2834 +		int base	= (opcode >> 16) & 0x0F;
  1.2835 +		int dest	= (opcode >> 12) & 0x0F;
  1.2836 +		u32 address = reg[base].I - ((opcode & 0x0F) | ((opcode >> 4) & 0xF0));
  1.2837 +		clockTicks += 3 + CPUUpdateTicksAccess16(address);
  1.2838 +		reg[dest].I = (s16)CPUReadHalfWordSigned(address);
  1.2839 +	}
  1.2840 +	break;
  1.2841 +	case 0x17f:
  1.2842 +	{
  1.2843 +		// LDRSH Rd, [Rn, -#offset]!
  1.2844 +		int base	= (opcode >> 16) & 0x0F;
  1.2845 +		int dest	= (opcode >> 12) & 0x0F;
  1.2846 +		u32 address = reg[base].I - ((opcode & 0x0F) | ((opcode >> 4) & 0xF0));
  1.2847 +		clockTicks += 3 + CPUUpdateTicksAccess16(address);
  1.2848 +		reg[dest].I = (s16)CPUReadHalfWordSigned(address);
  1.2849 +		if (dest != base)
  1.2850 +			reg[base].I = address;
  1.2851 +	}
  1.2852 +	break;
  1.2853 +	case 0x19f:
  1.2854 +	{
  1.2855 +		// LDRSH Rd, [Rn, Rm]
  1.2856 +		int base	= (opcode >> 16) & 0x0F;
  1.2857 +		int dest	= (opcode >> 12) & 0x0F;
  1.2858 +		u32 address = reg[base].I + reg[opcode & 0x0F].I;
  1.2859 +		clockTicks += 3 + CPUUpdateTicksAccess16(address);
  1.2860 +		reg[dest].I = (s16)CPUReadHalfWordSigned(address);
  1.2861 +	}
  1.2862 +	break;
  1.2863 +	case 0x1bf:
  1.2864 +	{
  1.2865 +		// LDRSH Rd, [Rn, Rm]!
  1.2866 +		int base	= (opcode >> 16) & 0x0F;
  1.2867 +		int dest	= (opcode >> 12) & 0x0F;
  1.2868 +		u32 address = reg[base].I + reg[opcode & 0x0F].I;
  1.2869 +		clockTicks += 3 + CPUUpdateTicksAccess16(address);
  1.2870 +		reg[dest].I = (s16)CPUReadHalfWordSigned(address);
  1.2871 +		if (dest != base)
  1.2872 +			reg[base].I = address;
  1.2873 +	}
  1.2874 +	break;
  1.2875 +	case 0x1df:
  1.2876 +	{
  1.2877 +		// LDRSH Rd, [Rn, #offset]
  1.2878 +		int base	= (opcode >> 16) & 0x0F;
  1.2879 +		int dest	= (opcode >> 12) & 0x0F;
  1.2880 +		u32 address = reg[base].I + ((opcode & 0x0F) | ((opcode >> 4) & 0xF0));
  1.2881 +		clockTicks += 3 + CPUUpdateTicksAccess16(address);
  1.2882 +		reg[dest].I = (s16)CPUReadHalfWordSigned(address);
  1.2883 +	}
  1.2884 +	break;
  1.2885 +	case 0x1ff:
  1.2886 +	{
  1.2887 +		// LDRSH Rd, [Rn, #offset]!
  1.2888 +		int base	= (opcode >> 16) & 0x0F;
  1.2889 +		int dest	= (opcode >> 12) & 0x0F;
  1.2890 +		u32 address = reg[base].I + ((opcode & 0x0F) | ((opcode >> 4) & 0xF0));
  1.2891 +		clockTicks += 3 + CPUUpdateTicksAccess16(address);
  1.2892 +		reg[dest].I = (s16)CPUReadHalfWordSigned(address);
  1.2893 +		if (dest != base)
  1.2894 +			reg[base].I = address;
  1.2895 +	}
  1.2896 +	break;
  1.2897 +		LOGICAL_DATA_OPCODE_WITHOUT_base(OP_EOR,  OP_EOR, 0x020);
  1.2898 +		LOGICAL_DATA_OPCODE_WITHOUT_base(OP_EORS, OP_EOR, 0x030);
  1.2899 +	case 0x029:
  1.2900 +	{
  1.2901 +		// MLA Rd, Rm, Rs, Rn
  1.2902 +		int dest = (opcode >> 16) & 0x0F;
  1.2903 +		int mult = (opcode & 0x0F);
  1.2904 +		u32 rs	 = reg[(opcode >> 8) & 0x0F].I;
  1.2905 +		reg[dest].I = reg[mult].I * rs + reg[(opcode >> 12) & 0x0f].I;
  1.2906 +		if (((s32)rs) < 0)
  1.2907 +			rs = ~rs;
  1.2908 +		if ((rs & 0xFFFFFF00) == 0)
  1.2909 +			clockTicks += 3;
  1.2910 +		else if ((rs & 0xFFFF0000) == 0)
  1.2911 +			clockTicks += 4;
  1.2912 +		else if ((rs & 0xFF000000) == 0)
  1.2913 +			clockTicks += 5;
  1.2914 +		else
  1.2915 +			clockTicks += 6;
  1.2916 +	}
  1.2917 +	break;
  1.2918 +	case 0x039:
  1.2919 +	{
  1.2920 +		// MLAS Rd, Rm, Rs, Rn
  1.2921 +		int dest = (opcode >> 16) & 0x0F;
  1.2922 +		int mult = (opcode & 0x0F);
  1.2923 +		u32 rs	 = reg[(opcode >> 8) & 0x0F].I;
  1.2924 +		reg[dest].I = reg[mult].I * rs + reg[(opcode >> 12) & 0x0f].I;
  1.2925 +		N_FLAG		= (reg[dest].I & 0x80000000) ? true : false;
  1.2926 +		Z_FLAG		= (reg[dest].I) ? false : true;
  1.2927 +		if (((s32)rs) < 0)
  1.2928 +			rs = ~rs;
  1.2929 +		if ((rs & 0xFFFFFF00) == 0)
  1.2930 +			clockTicks += 3;
  1.2931 +		else if ((rs & 0xFFFF0000) == 0)
  1.2932 +			clockTicks += 4;
  1.2933 +		else if ((rs & 0xFF000000) == 0)
  1.2934 +			clockTicks += 5;
  1.2935 +		else
  1.2936 +			clockTicks += 6;
  1.2937 +	}
  1.2938 +	break;
  1.2939 +		ARITHMETIC_DATA_OPCODE(OP_SUB,  OP_SUB, 0x040);
  1.2940 +		ARITHMETIC_DATA_OPCODE(OP_SUBS, OP_SUB, 0x050);
  1.2941 +		ARITHMETIC_DATA_OPCODE(OP_RSB,  OP_RSB, 0x060);
  1.2942 +		ARITHMETIC_DATA_OPCODE(OP_RSBS, OP_RSB, 0x070);
  1.2943 +		ARITHMETIC_DATA_OPCODE(OP_ADD,  OP_ADD, 0x080);
  1.2944 +		ARITHMETIC_DATA_OPCODE(OP_ADDS, OP_ADD, 0x090);
  1.2945 +	case 0x089:
  1.2946 +	{
  1.2947 +		// UMULL RdLo, RdHi, Rn, Rs
  1.2948 +		u32 umult	= reg[(opcode & 0x0F)].I;
  1.2949 +		u32 usource = reg[(opcode >> 8) & 0x0F].I;
  1.2950 +		int destLo	= (opcode >> 12) & 0x0F;
  1.2951 +		int destHi	= (opcode >> 16) & 0x0F;
  1.2952 +		u64 uTemp	= ((u64)umult) * ((u64)usource);
  1.2953 +		reg[destLo].I = (u32)(uTemp & 0xFFFFFFFF);
  1.2954 +		reg[destHi].I = (u32)(uTemp >> 32);
  1.2955 +		if ((usource & 0xFFFFFF00) == 0)
  1.2956 +			clockTicks += 2;
  1.2957 +		else if ((usource & 0xFFFF0000) == 0)
  1.2958 +			clockTicks += 3;
  1.2959 +		else if ((usource & 0xFF000000) == 0)
  1.2960 +			clockTicks += 4;
  1.2961 +		else
  1.2962 +			clockTicks += 5;
  1.2963 +	}
  1.2964 +	break;
  1.2965 +	case 0x099:
  1.2966 +	{
  1.2967 +		// UMULLS RdLo, RdHi, Rn, Rs
  1.2968 +		u32 umult	= reg[(opcode & 0x0F)].I;
  1.2969 +		u32 usource = reg[(opcode >> 8) & 0x0F].I;
  1.2970 +		int destLo	= (opcode >> 12) & 0x0F;
  1.2971 +		int destHi	= (opcode >> 16) & 0x0F;
  1.2972 +		u64 uTemp	= ((u64)umult) * ((u64)usource);
  1.2973 +		reg[destLo].I = (u32)(uTemp & 0xFFFFFFFF);
  1.2974 +		reg[destHi].I = (u32)(uTemp >> 32);
  1.2975 +		Z_FLAG		  = (uTemp) ? false : true;
  1.2976 +		N_FLAG		  = (reg[destHi].I & 0x80000000) ? true : false;
  1.2977 +		if ((usource & 0xFFFFFF00) == 0)
  1.2978 +			clockTicks += 2;
  1.2979 +		else if ((usource & 0xFFFF0000) == 0)
  1.2980 +			clockTicks += 3;
  1.2981 +		else if ((usource & 0xFF000000) == 0)
  1.2982 +			clockTicks += 4;
  1.2983 +		else
  1.2984 +			clockTicks += 5;
  1.2985 +	}
  1.2986 +	break;
  1.2987 +		ARITHMETIC_DATA_OPCODE(OP_ADC,  OP_ADC, 0x0a0);
  1.2988 +		ARITHMETIC_DATA_OPCODE(OP_ADCS, OP_ADC, 0x0b0);
  1.2989 +	case 0x0a9:
  1.2990 +	{
  1.2991 +		// UMLAL RdLo, RdHi, Rn, Rs
  1.2992 +		u32 umult	= reg[(opcode & 0x0F)].I;
  1.2993 +		u32 usource = reg[(opcode >> 8) & 0x0F].I;
  1.2994 +		int destLo	= (opcode >> 12) & 0x0F;
  1.2995 +		int destHi	= (opcode >> 16) & 0x0F;
  1.2996 +		u64 uTemp	= (u64)reg[destHi].I;
  1.2997 +		uTemp		<<= 32;
  1.2998 +		uTemp		 |= (u64)reg[destLo].I;
  1.2999 +		uTemp		 += ((u64)umult) * ((u64)usource);
  1.3000 +		reg[destLo].I = (u32)(uTemp & 0xFFFFFFFF);
  1.3001 +		reg[destHi].I = (u32)(uTemp >> 32);
  1.3002 +		if ((usource & 0xFFFFFF00) == 0)
  1.3003 +			clockTicks += 3;
  1.3004 +		else if ((usource & 0xFFFF0000) == 0)
  1.3005 +			clockTicks += 4;
  1.3006 +		else if ((usource & 0xFF000000) == 0)
  1.3007 +			clockTicks += 5;
  1.3008 +		else
  1.3009 +			clockTicks += 6;
  1.3010 +	}
  1.3011 +	break;
  1.3012 +	case 0x0b9:
  1.3013 +	{
  1.3014 +		// UMLALS RdLo, RdHi, Rn, Rs
  1.3015 +		u32 umult	= reg[(opcode & 0x0F)].I;
  1.3016 +		u32 usource = reg[(opcode >> 8) & 0x0F].I;
  1.3017 +		int destLo	= (opcode >> 12) & 0x0F;
  1.3018 +		int destHi	= (opcode >> 16) & 0x0F;
  1.3019 +		u64 uTemp	= (u64)reg[destHi].I;
  1.3020 +		uTemp		<<= 32;
  1.3021 +		uTemp		 |= (u64)reg[destLo].I;
  1.3022 +		uTemp		 += ((u64)umult) * ((u64)usource);
  1.3023 +		reg[destLo].I = (u32)(uTemp & 0xFFFFFFFF);
  1.3024 +		reg[destHi].I = (u32)(uTemp >> 32);
  1.3025 +		Z_FLAG		  = (uTemp) ? false : true;
  1.3026 +		N_FLAG		  = (reg[destHi].I & 0x80000000) ? true : false;
  1.3027 +		if ((usource & 0xFFFFFF00) == 0)
  1.3028 +			clockTicks += 3;
  1.3029 +		else if ((usource & 0xFFFF0000) == 0)
  1.3030 +			clockTicks += 4;
  1.3031 +		else if ((usource & 0xFF000000) == 0)
  1.3032 +			clockTicks += 5;
  1.3033 +		else
  1.3034 +			clockTicks += 6;
  1.3035 +	}
  1.3036 +	break;
  1.3037 +		ARITHMETIC_DATA_OPCODE(OP_SBC,  OP_SBC, 0x0c0);
  1.3038 +		ARITHMETIC_DATA_OPCODE(OP_SBCS, OP_SBC, 0x0d0);
  1.3039 +	case 0x0c9:
  1.3040 +	{
  1.3041 +		// SMULL RdLo, RdHi, Rm, Rs
  1.3042 +		int destLo = (opcode >> 12) & 0x0F;
  1.3043 +		int destHi = (opcode >> 16) & 0x0F;
  1.3044 +		u32 rs	   = reg[(opcode >> 8) & 0x0F].I;
  1.3045 +		s64 m	   = (s32)reg[(opcode & 0x0F)].I;
  1.3046 +		s64 s	   = (s32)rs;
  1.3047 +		s64 sTemp  = m * s;
  1.3048 +		reg[destLo].I = (u32)(sTemp & 0xFFFFFFFF);
  1.3049 +		reg[destHi].I = (u32)(sTemp >> 32);
  1.3050 +		if (((s32)rs) < 0)
  1.3051 +			rs = ~rs;
  1.3052 +		if ((rs & 0xFFFFFF00) == 0)
  1.3053 +			clockTicks += 2;
  1.3054 +		else if ((rs & 0xFFFF0000) == 0)
  1.3055 +			clockTicks += 3;
  1.3056 +		else if ((rs & 0xFF000000) == 0)
  1.3057 +			clockTicks += 4;
  1.3058 +		else
  1.3059 +			clockTicks += 5;
  1.3060 +	}
  1.3061 +	break;
  1.3062 +	case 0x0d9:
  1.3063 +	{
  1.3064 +		// SMULLS RdLo, RdHi, Rm, Rs
  1.3065 +		int destLo = (opcode >> 12) & 0x0F;
  1.3066 +		int destHi = (opcode >> 16) & 0x0F;
  1.3067 +		u32 rs	   = reg[(opcode >> 8) & 0x0F].I;
  1.3068 +		s64 m	   = (s32)reg[(opcode & 0x0F)].I;
  1.3069 +		s64 s	   = (s32)rs;
  1.3070 +		s64 sTemp  = m * s;
  1.3071 +		reg[destLo].I = (u32)(sTemp & 0xFFFFFFFF);
  1.3072 +		reg[destHi].I = (u32)(sTemp >> 32);
  1.3073 +		Z_FLAG		  = (sTemp) ? false : true;
  1.3074 +		N_FLAG		  = (sTemp < 0) ? true : false;
  1.3075 +		if (((s32)rs) < 0)
  1.3076 +			rs = ~rs;
  1.3077 +		if ((rs & 0xFFFFFF00) == 0)
  1.3078 +			clockTicks += 2;
  1.3079 +		else if ((rs & 0xFFFF0000) == 0)
  1.3080 +			clockTicks += 3;
  1.3081 +		else if ((rs & 0xFF000000) == 0)
  1.3082 +			clockTicks += 4;
  1.3083 +		else
  1.3084 +			clockTicks += 5;
  1.3085 +	}
  1.3086 +	break;
  1.3087 +		ARITHMETIC_DATA_OPCODE(OP_RSC,  OP_RSC, 0x0e0);
  1.3088 +		ARITHMETIC_DATA_OPCODE(OP_RSCS, OP_RSC, 0x0f0);
  1.3089 +	case 0x0e9:
  1.3090 +	{
  1.3091 +		// SMLAL RdLo, RdHi, Rm, Rs
  1.3092 +		int destLo = (opcode >> 12) & 0x0F;
  1.3093 +		int destHi = (opcode >> 16) & 0x0F;
  1.3094 +		u32 rs	   = reg[(opcode >> 8) & 0x0F].I;
  1.3095 +		s64 m	   = (s32)reg[(opcode & 0x0F)].I;
  1.3096 +		s64 s	   = (s32)rs;
  1.3097 +		s64 sTemp  = (u64)reg[destHi].I;
  1.3098 +		sTemp		<<= 32;
  1.3099 +		sTemp		 |= (u64)reg[destLo].I;
  1.3100 +		sTemp		 += m * s;
  1.3101 +		reg[destLo].I = (u32)(sTemp & 0xFFFFFFFF);
  1.3102 +		reg[destHi].I = (u32)(sTemp >> 32);
  1.3103 +		if (((s32)rs) < 0)
  1.3104 +			rs = ~rs;
  1.3105 +		if ((rs & 0xFFFFFF00) == 0)
  1.3106 +			clockTicks += 3;
  1.3107 +		else if ((rs & 0xFFFF0000) == 0)
  1.3108 +			clockTicks += 4;
  1.3109 +		else if ((rs & 0xFF000000) == 0)
  1.3110 +			clockTicks += 5;
  1.3111 +		else
  1.3112 +			clockTicks += 6;
  1.3113 +	}
  1.3114 +	break;
  1.3115 +	case 0x0f9:
  1.3116 +	{
  1.3117 +		// SMLALS RdLo, RdHi, Rm, Rs
  1.3118 +		int destLo = (opcode >> 12) & 0x0F;
  1.3119 +		int destHi = (opcode >> 16) & 0x0F;
  1.3120 +		u32 rs	   = reg[(opcode >> 8) & 0x0F].I;
  1.3121 +		s64 m	   = (s32)reg[(opcode & 0x0F)].I;
  1.3122 +		s64 s	   = (s32)rs;
  1.3123 +		s64 sTemp  = (u64)reg[destHi].I;
  1.3124 +		sTemp		<<= 32;
  1.3125 +		sTemp		 |= (u64)reg[destLo].I;
  1.3126 +		sTemp		 += m * s;
  1.3127 +		reg[destLo].I = (u32)(sTemp & 0xFFFFFFFF);
  1.3128 +		reg[destHi].I = (u32)(sTemp >> 32);
  1.3129 +		Z_FLAG		  = (sTemp) ? false : true;
  1.3130 +		N_FLAG		  = (sTemp < 0) ? true : false;
  1.3131 +		if (((s32)rs) < 0)
  1.3132 +			rs = ~rs;
  1.3133 +		if ((rs & 0xFFFFFF00) == 0)
  1.3134 +			clockTicks += 3;
  1.3135 +		else if ((rs & 0xFFFF0000) == 0)
  1.3136 +			clockTicks += 4;
  1.3137 +		else if ((rs & 0xFF000000) == 0)
  1.3138 +			clockTicks += 5;
  1.3139 +		else
  1.3140 +			clockTicks += 6;
  1.3141 +	}
  1.3142 +	break;
  1.3143 +		LOGICAL_DATA_OPCODE(OP_TST, OP_TST, 0x110);
  1.3144 +	case 0x100:
  1.3145 +		// MRS Rd, CPSR
  1.3146 +		// TODO: check if right instruction....
  1.3147 +		CPUUpdateCPSR();
  1.3148 +		reg[(opcode >> 12) & 0x0F].I = reg[16].I;
  1.3149 +		break;
  1.3150 +	case 0x109:
  1.3151 +	{
  1.3152 +		// SWP Rd, Rm, [Rn]
  1.3153 +		u32 address = reg[(opcode >> 16) & 15].I;
  1.3154 +		u32 temp	= CPUReadMemory(address);
  1.3155 +		CPUWriteMemory(address, reg[opcode & 15].I);
  1.3156 +		reg[(opcode >> 12) & 15].I = temp;
  1.3157 +	}
  1.3158 +	break;
  1.3159 +		LOGICAL_DATA_OPCODE(OP_TEQ, OP_TEQ, 0x130);
  1.3160 +	case 0x120:
  1.3161 +	{
  1.3162 +		// MSR CPSR_fields, Rm
  1.3163 +		CPUUpdateCPSR();
  1.3164 +		u32 value	 = reg[opcode & 15].I;
  1.3165 +		u32 newValue = reg[16].I;
  1.3166 +		if (armMode > 0x10)
  1.3167 +		{
  1.3168 +			if (opcode & 0x00010000)
  1.3169 +				newValue = (newValue & 0xFFFFFF00) | (value & 0x000000FF);
  1.3170 +			if (opcode & 0x00020000)
  1.3171 +				newValue = (newValue & 0xFFFF00FF) | (value & 0x0000FF00);
  1.3172 +			if (opcode & 0x00040000)
  1.3173 +				newValue = (newValue & 0xFF00FFFF) | (value & 0x00FF0000);
  1.3174 +		}
  1.3175 +		if (opcode & 0x00080000)
  1.3176 +			newValue = (newValue & 0x00FFFFFF) | (value & 0xFF000000);
  1.3177 +		newValue |= 0x10;
  1.3178 +		CPUSwitchMode(newValue & 0x1f, false);
  1.3179 +		reg[16].I = newValue;
  1.3180 +		CPUUpdateFlags();
  1.3181 +	}
  1.3182 +	break;
  1.3183 +	case 0x121:
  1.3184 +	{
  1.3185 +		// BX Rm
  1.3186 +		// TODO: check if right instruction...
  1.3187 +		clockTicks += 3;
  1.3188 +		int base = opcode & 0x0F;
  1.3189 +		armState = reg[base].I & 1 ? false : true;
  1.3190 +		if (armState)
  1.3191 +		{
  1.3192 +			reg[15].I  = reg[base].I & 0xFFFFFFFC;
  1.3193 +			armNextPC  = reg[15].I;
  1.3194 +			reg[15].I += 4;
  1.3195 +		}
  1.3196 +		else
  1.3197 +		{
  1.3198 +			reg[15].I  = reg[base].I & 0xFFFFFFFE;
  1.3199 +			armNextPC  = reg[15].I;
  1.3200 +			reg[15].I += 2;
  1.3201 +		}
  1.3202 +	}
  1.3203 +	break;
  1.3204 +		ARITHMETIC_DATA_OPCODE(OP_CMP, OP_CMP, 0x150);
  1.3205 +	case 0x140:
  1.3206 +		// MRS Rd, SPSR
  1.3207 +		// TODO: check if right instruction...
  1.3208 +		reg[(opcode >> 12) & 0x0F].I = reg[17].I;
  1.3209 +		break;
  1.3210 +	case 0x149:
  1.3211 +	{
  1.3212 +		// SWPB Rd, Rm, [Rn]
  1.3213 +		u32 address = reg[(opcode >> 16) & 15].I;
  1.3214 +		u32 temp	= CPUReadByte(address);
  1.3215 +		CPUWriteByte(address, reg[opcode & 15].B.B0);
  1.3216 +		reg[(opcode >> 12) & 15].I = temp;
  1.3217 +	}
  1.3218 +	break;
  1.3219 +		ARITHMETIC_DATA_OPCODE(OP_CMN, OP_CMN, 0x170);
  1.3220 +	case 0x160:
  1.3221 +	{
  1.3222 +		// MSR SPSR_fields, Rm
  1.3223 +		u32 value = reg[opcode & 15].I;
  1.3224 +		if (armMode > 0x10 && armMode < 0x1f)
  1.3225 +		{
  1.3226 +			if (opcode & 0x00010000)
  1.3227 +				reg[17].I = (reg[17].I & 0xFFFFFF00) | (value & 0x000000FF);
  1.3228 +			if (opcode & 0x00020000)
  1.3229 +				reg[17].I = (reg[17].I & 0xFFFF00FF) | (value & 0x0000FF00);
  1.3230 +			if (opcode & 0x00040000)
  1.3231 +				reg[17].I = (reg[17].I & 0xFF00FFFF) | (value & 0x00FF0000);
  1.3232 +			if (opcode & 0x00080000)
  1.3233 +				reg[17].I = (reg[17].I & 0x00FFFFFF) | (value & 0xFF000000);
  1.3234 +		}
  1.3235 +	}
  1.3236 +	break;
  1.3237 +		LOGICAL_DATA_OPCODE(OP_ORR,  OP_ORR, 0x180);
  1.3238 +		LOGICAL_DATA_OPCODE(OP_ORRS, OP_ORR, 0x190);
  1.3239 +		LOGICAL_DATA_OPCODE_WITHOUT_base(OP_MOV,  OP_MOV, 0x1a0);
  1.3240 +		LOGICAL_DATA_OPCODE_WITHOUT_base(OP_MOVS, OP_MOV, 0x1b0);
  1.3241 +		LOGICAL_DATA_OPCODE(OP_BIC,  OP_BIC, 0x1c0);
  1.3242 +		LOGICAL_DATA_OPCODE(OP_BICS, OP_BIC, 0x1d0);
  1.3243 +		LOGICAL_DATA_OPCODE_WITHOUT_base(OP_MVN,  OP_MVN, 0x1e0);
  1.3244 +		LOGICAL_DATA_OPCODE_WITHOUT_base(OP_MVNS, OP_MVN, 0x1f0);
  1.3245 +#ifdef BKPT_SUPPORT
  1.3246 +	case 0x127:
  1.3247 +	case 0x7ff: // for GDB support
  1.3248 +		extern void (*dbgSignal)(int, int);
  1.3249 +		reg[15].I -= 4;
  1.3250 +		armNextPC -= 4;
  1.3251 +		dbgSignal(5, (opcode & 0x0f) | ((opcode >> 4) & 0xfff0));
  1.3252 +		return;
  1.3253 +#endif
  1.3254 +	case 0x320:
  1.3255 +	case 0x321:
  1.3256 +	case 0x322:
  1.3257 +	case 0x323:
  1.3258 +	case 0x324:
  1.3259 +	case 0x325:
  1.3260 +	case 0x326:
  1.3261 +	case 0x327:
  1.3262 +	case 0x328:
  1.3263 +	case 0x329:
  1.3264 +	case 0x32a:
  1.3265 +	case 0x32b:
  1.3266 +	case 0x32c:
  1.3267 +	case 0x32d:
  1.3268 +	case 0x32e:
  1.3269 +	case 0x32f:
  1.3270 +	{
  1.3271 +		// MSR CPSR_fields, #
  1.3272 +		CPUUpdateCPSR();
  1.3273 +		u32 value = opcode & 0xFF;
  1.3274 +		int shift = (opcode & 0xF00) >> 7;
  1.3275 +		if (shift)
  1.3276 +		{
  1.3277 +			ROR_IMM_MSR;
  1.3278 +		}
  1.3279 +		u32 newValue = reg[16].I;
  1.3280 +		if (armMode > 0x10)
  1.3281 +		{
  1.3282 +			if (opcode & 0x00010000)
  1.3283 +				newValue = (newValue & 0xFFFFFF00) | (value & 0x000000FF);
  1.3284 +			if (opcode & 0x00020000)
  1.3285 +				newValue = (newValue & 0xFFFF00FF) | (value & 0x0000FF00);
  1.3286 +			if (opcode & 0x00040000)
  1.3287 +				newValue = (newValue & 0xFF00FFFF) | (value & 0x00FF0000);
  1.3288 +		}
  1.3289 +		if (opcode & 0x00080000)
  1.3290 +			newValue = (newValue & 0x00FFFFFF) | (value & 0xFF000000);
  1.3291 +
  1.3292 +		newValue |= 0x10;
  1.3293 +
  1.3294 +		CPUSwitchMode(newValue & 0x1f, false);
  1.3295 +		reg[16].I = newValue;
  1.3296 +		CPUUpdateFlags();
  1.3297 +	}
  1.3298 +	break;
  1.3299 +	case 0x360:
  1.3300 +	case 0x361:
  1.3301 +	case 0x362:
  1.3302 +	case 0x363:
  1.3303 +	case 0x364:
  1.3304 +	case 0x365:
  1.3305 +	case 0x366:
  1.3306 +	case 0x367:
  1.3307 +	case 0x368:
  1.3308 +	case 0x369:
  1.3309 +	case 0x36a:
  1.3310 +	case 0x36b:
  1.3311 +	case 0x36c:
  1.3312 +	case 0x36d:
  1.3313 +	case 0x36e:
  1.3314 +	case 0x36f:
  1.3315 +	{
  1.3316 +		// MSR SPSR_fields, #
  1.3317 +		if (armMode > 0x10 && armMode < 0x1f)
  1.3318 +		{
  1.3319 +			u32 value = opcode & 0xFF;
  1.3320 +			int shift = (opcode & 0xF00) >> 7;
  1.3321 +			if (shift)
  1.3322 +			{
  1.3323 +				ROR_IMM_MSR;
  1.3324 +			}
  1.3325 +			if (opcode & 0x00010000)
  1.3326 +				reg[17].I = (reg[17].I & 0xFFFFFF00) | (value & 0x000000FF);
  1.3327 +			if (opcode & 0x00020000)
  1.3328 +				reg[17].I = (reg[17].I & 0xFFFF00FF) | (value & 0x0000FF00);
  1.3329 +			if (opcode & 0x00040000)
  1.3330 +				reg[17].I = (reg[17].I & 0xFF00FFFF) | (value & 0x00FF0000);
  1.3331 +			if (opcode & 0x00080000)
  1.3332 +				reg[17].I = (reg[17].I & 0x00FFFFFF) | (value & 0xFF000000);
  1.3333 +		}
  1.3334 +	}
  1.3335 +	break;
  1.3336 +		CASE_16(0x400)
  1.3337 +		// T versions shouldn't be different on GBA
  1.3338 +		CASE_16(0x420)
  1.3339 +		{
  1.3340 +			// STR Rd, [Rn], -#
  1.3341 +			int offset	= opcode & 0xFFF;
  1.3342 +			int dest	= (opcode >> 12) & 15;
  1.3343 +			int base	= (opcode >> 16) & 15;
  1.3344 +			u32 address = reg[base].I;
  1.3345 +			CPUWriteMemory(address, reg[dest].I);
  1.3346 +			reg[base].I = address - offset;
  1.3347 +			clockTicks += 2 + CPUUpdateTicksAccess32(address);
  1.3348 +		}
  1.3349 +		break;
  1.3350 +		CASE_16(0x480)
  1.3351 +		// T versions shouldn't be different on GBA
  1.3352 +		CASE_16(0x4a0)
  1.3353 +		{
  1.3354 +			// STR Rd, [Rn], #
  1.3355 +			int offset	= opcode & 0xFFF;
  1.3356 +			int dest	= (opcode >> 12) & 15;
  1.3357 +			int base	= (opcode >> 16) & 15;
  1.3358 +			u32 address = reg[base].I;
  1.3359 +			CPUWriteMemory(address, reg[dest].I);
  1.3360 +			reg[base].I = address + offset;
  1.3361 +			clockTicks += 2 + CPUUpdateTicksAccess32(address);
  1.3362 +		}
  1.3363 +		break;
  1.3364 +		CASE_16(0x500)
  1.3365 +		{
  1.3366 +			// STR Rd, [Rn, -#]
  1.3367 +			int offset	= opcode & 0xFFF;
  1.3368 +			int dest	= (opcode >> 12) & 15;
  1.3369 +			int base	= (opcode >> 16) & 15;
  1.3370 +			u32 address = reg[base].I - offset;
  1.3371 +			CPUWriteMemory(address, reg[dest].I);
  1.3372 +			clockTicks += 2 + CPUUpdateTicksAccess32(address);
  1.3373 +		}
  1.3374 +		break;
  1.3375 +		CASE_16(0x520)
  1.3376 +		{
  1.3377 +			// STR Rd, [Rn, -#]!
  1.3378 +			int offset	= opcode & 0xFFF;
  1.3379 +			int dest	= (opcode >> 12) & 15;
  1.3380 +			int base	= (opcode >> 16) & 15;
  1.3381 +			u32 address = reg[base].I - offset;
  1.3382 +			reg[base].I = address;
  1.3383 +			CPUWriteMemory(address, reg[dest].I);
  1.3384 +			clockTicks += 2 + CPUUpdateTicksAccess32(address);
  1.3385 +		}
  1.3386 +		break;
  1.3387 +		CASE_16(0x580)
  1.3388 +		{
  1.3389 +			// STR Rd, [Rn, #]
  1.3390 +			int offset	= opcode & 0xFFF;
  1.3391 +			int dest	= (opcode >> 12) & 15;
  1.3392 +			int base	= (opcode >> 16) & 15;
  1.3393 +			u32 address = reg[base].I + offset;
  1.3394 +			CPUWriteMemory(address, reg[dest].I);
  1.3395 +			clockTicks += 2 + CPUUpdateTicksAccess32(address);
  1.3396 +		}
  1.3397 +		break;
  1.3398 +		CASE_16(0x5a0)
  1.3399 +		{
  1.3400 +			// STR Rd, [Rn, #]!
  1.3401 +			int offset	= opcode & 0xFFF;
  1.3402 +			int dest	= (opcode >> 12) & 15;
  1.3403 +			int base	= (opcode >> 16) & 15;
  1.3404 +			u32 address = reg[base].I + offset;
  1.3405 +			reg[base].I = address;
  1.3406 +			CPUWriteMemory(address, reg[dest].I);
  1.3407 +			clockTicks += 2 + CPUUpdateTicksAccess32(address);
  1.3408 +		}
  1.3409 +		break;
  1.3410 +		CASE_16(0x410)
  1.3411 +		{
  1.3412 +			// LDR Rd, [Rn], -#
  1.3413 +			int offset	= opcode & 0xFFF;
  1.3414 +			int dest	= (opcode >> 12) & 15;
  1.3415 +			int base	= (opcode >> 16) & 15;
  1.3416 +			u32 address = reg[base].I;
  1.3417 +			reg[dest].I = CPUReadMemory(address);
  1.3418 +			if (dest != base)
  1.3419 +				reg[base].I -= offset;
  1.3420 +			clockTicks += 3 + CPUUpdateTicksAccess32(address);
  1.3421 +			if (dest == 15)
  1.3422 +			{
  1.3423 +				clockTicks += 2;
  1.3424 +				reg[15].I  &= 0xFFFFFFFC;
  1.3425 +				armNextPC	= reg[15].I;
  1.3426 +				reg[15].I  += 4;
  1.3427 +			}
  1.3428 +		}
  1.3429 +		break;
  1.3430 +		CASE_16(0x430)
  1.3431 +		{
  1.3432 +			// LDRT Rd, [Rn], -#
  1.3433 +			int offset	= opcode & 0xFFF;
  1.3434 +			int dest	= (opcode >> 12) & 15;
  1.3435 +			int base	= (opcode >> 16) & 15;
  1.3436 +			u32 address = reg[base].I;
  1.3437 +			reg[dest].I = CPUReadMemory(address);
  1.3438 +			if (dest != base)
  1.3439 +				reg[base].I -= offset;
  1.3440 +			clockTicks += 3 + CPUUpdateTicksAccess32(address);
  1.3441 +		}
  1.3442 +		break;
  1.3443 +		CASE_16(0x490)
  1.3444 +		{
  1.3445 +			// LDR Rd, [Rn], #
  1.3446 +			int offset	= opcode & 0xFFF;
  1.3447 +			int dest	= (opcode >> 12) & 15;
  1.3448 +			int base	= (opcode >> 16) & 15;
  1.3449 +			u32 address = reg[base].I;
  1.3450 +			reg[dest].I = CPUReadMemory(address);
  1.3451 +			if (dest != base)
  1.3452 +				reg[base].I += offset;
  1.3453 +			clockTicks += 3 + CPUUpdateTicksAccess32(address);
  1.3454 +			if (dest == 15)
  1.3455 +			{
  1.3456 +				clockTicks += 2;
  1.3457 +				reg[15].I  &= 0xFFFFFFFC;
  1.3458 +				armNextPC	= reg[15].I;
  1.3459 +				reg[15].I  += 4;
  1.3460 +			}
  1.3461 +		}
  1.3462 +		break;
  1.3463 +		CASE_16(0x4b0)
  1.3464 +		{
  1.3465 +			// LDRT Rd, [Rn], #
  1.3466 +			int offset	= opcode & 0xFFF;
  1.3467 +			int dest	= (opcode >> 12) & 15;
  1.3468 +			int base	= (opcode >> 16) & 15;
  1.3469 +			u32 address = reg[base].I;
  1.3470 +			reg[dest].I = CPUReadMemory(address);
  1.3471 +			if (dest != base)
  1.3472 +				reg[base].I += offset;
  1.3473 +			clockTicks += 3 + CPUUpdateTicksAccess32(address);
  1.3474 +		}
  1.3475 +		break;
  1.3476 +		CASE_16(0x510)
  1.3477 +		{
  1.3478 +			// LDR Rd, [Rn, -#]
  1.3479 +			int offset	= opcode & 0xFFF;
  1.3480 +			int dest	= (opcode >> 12) & 15;
  1.3481 +			int base	= (opcode >> 16) & 15;
  1.3482 +			u32 address = reg[base].I - offset;
  1.3483 +			reg[dest].I = CPUReadMemory(address);
  1.3484 +			clockTicks += 3 + CPUUpdateTicksAccess32(address);
  1.3485 +			if (dest == 15)
  1.3486 +			{
  1.3487 +				clockTicks += 2;
  1.3488 +				reg[15].I  &= 0xFFFFFFFC;
  1.3489 +				armNextPC	= reg[15].I;
  1.3490 +				reg[15].I  += 4;
  1.3491 +			}
  1.3492 +		}
  1.3493 +		break;
  1.3494 +		CASE_16(0x530)
  1.3495 +		{
  1.3496 +			// LDR Rd, [Rn, -#]!
  1.3497 +			int offset	= opcode & 0xFFF;
  1.3498 +			int dest	= (opcode >> 12) & 15;
  1.3499 +			int base	= (opcode >> 16) & 15;
  1.3500 +			u32 address = reg[base].I - offset;
  1.3501 +			reg[dest].I = CPUReadMemory(address);
  1.3502 +			if (dest != base)
  1.3503 +				reg[base].I = address;
  1.3504 +			clockTicks += 3 + CPUUpdateTicksAccess32(address);
  1.3505 +			if (dest == 15)
  1.3506 +			{
  1.3507 +				clockTicks += 2;
  1.3508 +				reg[15].I  &= 0xFFFFFFFC;
  1.3509 +				armNextPC	= reg[15].I;
  1.3510 +				reg[15].I  += 4;
  1.3511 +			}
  1.3512 +		}
  1.3513 +		break;
  1.3514 +		CASE_16(0x590)
  1.3515 +		{
  1.3516 +			// LDR Rd, [Rn, #]
  1.3517 +			int offset	= opcode & 0xFFF;
  1.3518 +			int dest	= (opcode >> 12) & 15;
  1.3519 +			int base	= (opcode >> 16) & 15;
  1.3520 +			u32 address = reg[base].I + offset;
  1.3521 +			reg[dest].I = CPUReadMemory(address);
  1.3522 +			clockTicks += 3 + CPUUpdateTicksAccess32(address);
  1.3523 +			if (dest == 15)
  1.3524 +			{
  1.3525 +				clockTicks += 2;
  1.3526 +				reg[15].I  &= 0xFFFFFFFC;
  1.3527 +				armNextPC	= reg[15].I;
  1.3528 +				reg[15].I  += 4;
  1.3529 +			}
  1.3530 +		}
  1.3531 +		break;
  1.3532 +		CASE_16(0x5b0)
  1.3533 +		{
  1.3534 +			// LDR Rd, [Rn, #]!
  1.3535 +			int offset	= opcode & 0xFFF;
  1.3536 +			int dest	= (opcode >> 12) & 15;
  1.3537 +			int base	= (opcode >> 16) & 15;
  1.3538 +			u32 address = reg[base].I + offset;
  1.3539 +			reg[dest].I = CPUReadMemory(address);
  1.3540 +			if (dest != base)
  1.3541 +				reg[base].I = address;
  1.3542 +			clockTicks += 3 + CPUUpdateTicksAccess32(address);
  1.3543 +			if (dest == 15)
  1.3544 +			{
  1.3545 +				clockTicks += 2;
  1.3546 +				reg[15].I  &= 0xFFFFFFFC;
  1.3547 +				armNextPC	= reg[15].I;
  1.3548 +				reg[15].I  += 4;
  1.3549 +			}
  1.3550 +		}
  1.3551 +		break;
  1.3552 +		CASE_16(0x440)
  1.3553 +		// T versions shouldn't be different on GBA
  1.3554 +		CASE_16(0x460)
  1.3555 +		{
  1.3556 +			// STRB Rd, [Rn], -#
  1.3557 +			int offset	= opcode & 0xFFF;
  1.3558 +			int dest	= (opcode >> 12) & 15;
  1.3559 +			int base	= (opcode >> 16) & 15;
  1.3560 +			u32 address = reg[base].I;
  1.3561 +			CPUWriteByte(address, reg[dest].B.B0);
  1.3562 +			reg[base].I = address - offset;
  1.3563 +			clockTicks += 2 + CPUUpdateTicksAccess16(address);
  1.3564 +		}
  1.3565 +		break;
  1.3566 +		CASE_16(0x4c0)
  1.3567 +		// T versions shouldn't be different on GBA
  1.3568 +		CASE_16(0x4e0)
  1.3569 +		// STRB Rd, [Rn], #
  1.3570 +		{
  1.3571 +			int offset	= opcode & 0xFFF;
  1.3572 +			int dest	= (opcode >> 12) & 15;
  1.3573 +			int base	= (opcode >> 16) & 15;
  1.3574 +			u32 address = reg[base].I;
  1.3575 +			CPUWriteByte(address, reg[dest].B.B0);
  1.3576 +			reg[base].I = address + offset;
  1.3577 +			clockTicks += 2 + CPUUpdateTicksAccess16(address);
  1.3578 +		}
  1.3579 +		break;
  1.3580 +		CASE_16(0x540)
  1.3581 +		{
  1.3582 +			// STRB Rd, [Rn, -#]
  1.3583 +			int offset	= opcode & 0xFFF;
  1.3584 +			int dest	= (opcode >> 12) & 15;
  1.3585 +			int base	= (opcode >> 16) & 15;
  1.3586 +			u32 address = reg[base].I - offset;
  1.3587 +			CPUWriteByte(address, reg[dest].B.B0);
  1.3588 +			clockTicks += 2 + CPUUpdateTicksAccess16(address);
  1.3589 +		}
  1.3590 +		break;
  1.3591 +		CASE_16(0x560)
  1.3592 +		{
  1.3593 +			// STRB Rd, [Rn, -#]!
  1.3594 +			int offset	= opcode & 0xFFF;
  1.3595 +			int dest	= (opcode >> 12) & 15;
  1.3596 +			int base	= (opcode >> 16) & 15;
  1.3597 +			u32 address = reg[base].I - offset;
  1.3598 +			reg[base].I = address;
  1.3599 +			CPUWriteByte(address, reg[dest].B.B0);
  1.3600 +			clockTicks += 2 + CPUUpdateTicksAccess16(address);
  1.3601 +		}
  1.3602 +		break;
  1.3603 +		CASE_16(0x5c0)
  1.3604 +		{
  1.3605 +			// STRB Rd, [Rn, #]
  1.3606 +			int offset	= opcode & 0xFFF;
  1.3607 +			int dest	= (opcode >> 12) & 15;
  1.3608 +			int base	= (opcode >> 16) & 15;
  1.3609 +			u32 address = reg[base].I + offset;
  1.3610 +			CPUWriteByte(address, reg[dest].B.B0);
  1.3611 +			clockTicks += 2 + CPUUpdateTicksAccess16(address);
  1.3612 +		}
  1.3613 +		break;
  1.3614 +		CASE_16(0x5e0)
  1.3615 +		{
  1.3616 +			// STRB Rd, [Rn, #]!
  1.3617 +			int offset	= opcode & 0xFFF;
  1.3618 +			int dest	= (opcode >> 12) & 15;
  1.3619 +			int base	= (opcode >> 16) & 15;
  1.3620 +			u32 address = reg[base].I + offset;
  1.3621 +			reg[base].I = address;
  1.3622 +			CPUWriteByte(address, reg[dest].I);
  1.3623 +			clockTicks += 2 + CPUUpdateTicksAccess16(address);
  1.3624 +		}
  1.3625 +		break;
  1.3626 +		CASE_16(0x450)
  1.3627 +		// T versions shouldn't be different
  1.3628 +		CASE_16(0x470)
  1.3629 +		{
  1.3630 +			// LDRB Rd, [Rn], -#
  1.3631 +			int offset	= opcode & 0xFFF;
  1.3632 +			int dest	= (opcode >> 12) & 15;
  1.3633 +			int base	= (opcode >> 16) & 15;
  1.3634 +			u32 address = reg[base].I;
  1.3635 +			reg[dest].I = CPUReadByte(address);
  1.3636 +			if (dest != base)
  1.3637 +				reg[base].I -= offset;
  1.3638 +			clockTicks += 3 + CPUUpdateTicksAccess16(address);
  1.3639 +		}
  1.3640 +		break;
  1.3641 +		CASE_16(0x4d0)
  1.3642 +		CASE_16(0x4f0) // T versions should not be different
  1.3643 +		{
  1.3644 +			// LDRB Rd, [Rn], #
  1.3645 +			int offset	= opcode & 0xFFF;
  1.3646 +			int dest	= (opcode >> 12) & 15;
  1.3647 +			int base	= (opcode >> 16) & 15;
  1.3648 +			u32 address = reg[base].I;
  1.3649 +			reg[dest].I = CPUReadByte(address);
  1.3650 +			if (dest != base)
  1.3651 +				reg[base].I += offset;
  1.3652 +			clockTicks += 3 + CPUUpdateTicksAccess16(address);
  1.3653 +		}
  1.3654 +		break;
  1.3655 +		CASE_16(0x550)
  1.3656 +		{
  1.3657 +			// LDRB Rd, [Rn, -#]
  1.3658 +			int offset	= opcode & 0xFFF;
  1.3659 +			int dest	= (opcode >> 12) & 15;
  1.3660 +			int base	= (opcode >> 16) & 15;
  1.3661 +			u32 address = reg[base].I - offset;
  1.3662 +			reg[dest].I = CPUReadByte(address);
  1.3663 +			clockTicks += 3 + CPUUpdateTicksAccess16(address);
  1.3664 +		}
  1.3665 +		break;
  1.3666 +		CASE_16(0x570)
  1.3667 +		{
  1.3668 +			// LDRB Rd, [Rn, -#]!
  1.3669 +			int offset	= opcode & 0xFFF;
  1.3670 +			int dest	= (opcode >> 12) & 15;
  1.3671 +			int base	= (opcode >> 16) & 15;
  1.3672 +			u32 address = reg[base].I - offset;
  1.3673 +			reg[dest].I = CPUReadByte(address);
  1.3674 +			if (dest != base)
  1.3675 +				reg[base].I = address;
  1.3676 +			clockTicks += 3 + CPUUpdateTicksAccess16(address);
  1.3677 +		}
  1.3678 +		break;
  1.3679 +		CASE_16(0x5d0)
  1.3680 +		{
  1.3681 +			// LDRB Rd, [Rn, #]
  1.3682 +			int offset	= opcode & 0xFFF;
  1.3683 +			int dest	= (opcode >> 12) & 15;
  1.3684 +			int base	= (opcode >> 16) & 15;
  1.3685 +			u32 address = reg[base].I + offset;
  1.3686 +			reg[dest].I = CPUReadByte(address);
  1.3687 +			clockTicks += 3 + CPUUpdateTicksAccess16(address);
  1.3688 +		}
  1.3689 +		break;
  1.3690 +		CASE_16(0x5f0)
  1.3691 +		{
  1.3692 +			// LDRB Rd, [Rn, #]!
  1.3693 +			int offset	= opcode & 0xFFF;
  1.3694 +			int dest	= (opcode >> 12) & 15;
  1.3695 +			int base	= (opcode >> 16) & 15;
  1.3696 +			u32 address = reg[base].I + offset;
  1.3697 +			reg[dest].I = CPUReadByte(address);
  1.3698 +			if (dest != base)
  1.3699 +				reg[base].I = address;
  1.3700 +			clockTicks += 3 + CPUUpdateTicksAccess16(address);
  1.3701 +		}
  1.3702 +		break;
  1.3703 +	case 0x600:
  1.3704 +	case 0x608:
  1.3705 +	// T versions are the same
  1.3706 +	case 0x620:
  1.3707 +	case 0x628:
  1.3708 +	{
  1.3709 +		// STR Rd, [Rn], -Rm, LSL #
  1.3710 +		int offset	= reg[opcode & 15].I << ((opcode >> 7) & 31);
  1.3711 +		int dest	= (opcode >> 12) & 15;
  1.3712 +		int base	= (opcode >> 16) & 15;
  1.3713 +		u32 address = reg[base].I;
  1.3714 +		CPUWriteMemory(address, reg[dest].I);
  1.3715 +		reg[base].I = address - offset;
  1.3716 +		clockTicks += 2 + CPUUpdateTicksAccess32(address);
  1.3717 +	}
  1.3718 +	break;
  1.3719 +	case 0x602:
  1.3720 +	case 0x60a:
  1.3721 +	// T versions are the same
  1.3722 +	case 0x622:
  1.3723 +	case 0x62a:
  1.3724 +	{
  1.3725 +		// STR Rd, [Rn], -Rm, LSR #
  1.3726 +		int shift	= (opcode >> 7) & 31;
  1.3727 +		int offset	= shift ? reg[opcode & 15].I >> shift : 0;
  1.3728 +		int dest	= (opcode >> 12) & 15;
  1.3729 +		int base	= (opcode >> 16) & 15;
  1.3730 +		u32 address = reg[base].I;
  1.3731 +		CPUWriteMemory(address, reg[dest].I);
  1.3732 +		reg[base].I = address - offset;
  1.3733 +		clockTicks += 2 + CPUUpdateTicksAccess32(address);
  1.3734 +	}
  1.3735 +	break;
  1.3736 +	case 0x604:
  1.3737 +	case 0x60c:
  1.3738 +	// T versions are the same
  1.3739 +	case 0x624:
  1.3740 +	case 0x62c:
  1.3741 +	{
  1.3742 +		// STR Rd, [Rn], -Rm, ASR #
  1.3743 +		int shift = (opcode >> 7) & 31;
  1.3744 +		int offset;
  1.3745 +		if (shift)
  1.3746 +			offset = (int)((s32)reg[opcode & 15].I >> shift);
  1.3747 +		else if (reg[opcode & 15].I & 0x80000000)
  1.3748 +			offset = 0xFFFFFFFF;
  1.3749 +		else
  1.3750 +			offset = 0;
  1.3751 +		int dest	= (opcode >> 12) & 15;
  1.3752 +		int base	= (opcode >> 16) & 15;
  1.3753 +		u32 address = reg[base].I;
  1.3754 +		CPUWriteMemory(address, reg[dest].I);
  1.3755 +		reg[base].I = address - offset;
  1.3756 +		clockTicks += 2 + CPUUpdateTicksAccess32(address);
  1.3757 +	}
  1.3758 +	break;
  1.3759 +	case 0x606:
  1.3760 +	case 0x60e:
  1.3761 +	// T versions are the same
  1.3762 +	case 0x626:
  1.3763 +	case 0x62e:
  1.3764 +	{
  1.3765 +		// STR Rd, [Rn], -Rm, ROR #
  1.3766 +		int shift = (opcode >> 7) & 31;
  1.3767 +		u32 value = reg[opcode & 15].I;
  1.3768 +		if (shift)
  1.3769 +		{
  1.3770 +			ROR_VALUE;
  1.3771 +		}
  1.3772 +		else
  1.3773 +		{
  1.3774 +			RCR_VALUE;
  1.3775 +		}
  1.3776 +		int dest	= (opcode >> 12) & 15;
  1.3777 +		int base	= (opcode >> 16) & 15;
  1.3778 +		u32 address = reg[base].I;
  1.3779 +		CPUWriteMemory(address, reg[dest].I);
  1.3780 +		reg[base].I = address - value;
  1.3781 +		clockTicks += 2 + CPUUpdateTicksAccess32(address);
  1.3782 +	}
  1.3783 +	break;
  1.3784 +	case 0x680:
  1.3785 +	case 0x688:
  1.3786 +	// T versions are the same
  1.3787 +	case 0x6a0:
  1.3788 +	case 0x6a8:
  1.3789 +	{
  1.3790 +		// STR Rd, [Rn], Rm, LSL #
  1.3791 +		int offset	= reg[opcode & 15].I << ((opcode >> 7) & 31);
  1.3792 +		int dest	= (opcode >> 12) & 15;
  1.3793 +		int base	= (opcode >> 16) & 15;
  1.3794 +		u32 address = reg[base].I;
  1.3795 +		CPUWriteMemory(address, reg[dest].I);
  1.3796 +		reg[base].I = address + offset;
  1.3797 +		clockTicks += 2 + CPUUpdateTicksAccess32(address);
  1.3798 +	}
  1.3799 +	break;
  1.3800 +	case 0x682:
  1.3801 +	case 0x68a:
  1.3802 +	// T versions are the same
  1.3803 +	case 0x6a2:
  1.3804 +	case 0x6aa:
  1.3805 +	{
  1.3806 +		// STR Rd, [Rn], Rm, LSR #
  1.3807 +		int shift	= (opcode >> 7) & 31;
  1.3808 +		int offset	= shift ? reg[opcode & 15].I >> shift : 0;
  1.3809 +		int dest	= (opcode >> 12) & 15;
  1.3810 +		int base	= (opcode >> 16) & 15;
  1.3811 +		u32 address = reg[base].I;
  1.3812 +		CPUWriteMemory(address, reg[dest].I);
  1.3813 +		reg[base].I = address + offset;
  1.3814 +		clockTicks += 2 + CPUUpdateTicksAccess32(address);
  1.3815 +	}
  1.3816 +	break;
  1.3817 +	case 0x684:
  1.3818 +	case 0x68c:
  1.3819 +	// T versions are the same
  1.3820 +	case 0x6a4:
  1.3821 +	case 0x6ac:
  1.3822 +	{
  1.3823 +		// STR Rd, [Rn], Rm, ASR #
  1.3824 +		int shift = (opcode >> 7) & 31;
  1.3825 +		int offset;
  1.3826 +		if (shift)
  1.3827 +			offset = (int)((s32)reg[opcode & 15].I >> shift);
  1.3828 +		else if (reg[opcode & 15].I & 0x80000000)
  1.3829 +			offset = 0xFFFFFFFF;
  1.3830 +		else
  1.3831 +			offset = 0;
  1.3832 +		int dest	= (opcode >> 12) & 15;
  1.3833 +		int base	= (opcode >> 16) & 15;
  1.3834 +		u32 address = reg[base].I;
  1.3835 +		CPUWriteMemory(address, reg[dest].I);
  1.3836 +		reg[base].I = address + offset;
  1.3837 +		clockTicks += 2 + CPUUpdateTicksAccess32(address);
  1.3838 +	}
  1.3839 +	break;
  1.3840 +	case 0x686:
  1.3841 +	case 0x68e:
  1.3842 +	// T versions are the same
  1.3843 +	case 0x6a6:
  1.3844 +	case 0x6ae:
  1.3845 +	{
  1.3846 +		// STR Rd, [Rn], Rm, ROR #
  1.3847 +		int shift = (opcode >> 7) & 31;
  1.3848 +		u32 value = reg[opcode & 15].I;
  1.3849 +		if (shift)
  1.3850 +		{
  1.3851 +			ROR_VALUE;
  1.3852 +		}
  1.3853 +		else
  1.3854 +		{
  1.3855 +			RCR_VALUE;
  1.3856 +		}
  1.3857 +		int dest	= (opcode >> 12) & 15;
  1.3858 +		int base	= (opcode >> 16) & 15;
  1.3859 +		u32 address = reg[base].I;
  1.3860 +		CPUWriteMemory(address, reg[dest].I);
  1.3861 +		reg[base].I = address + value;
  1.3862 +		clockTicks += 2 + CPUUpdateTicksAccess32(address);
  1.3863 +	}
  1.3864 +	break;
  1.3865 +	case 0x700:
  1.3866 +	case 0x708:
  1.3867 +	{
  1.3868 +		// STR Rd, [Rn, -Rm, LSL #]
  1.3869 +		int offset	= reg[opcode & 15].I << ((opcode >> 7) & 31);
  1.3870 +		int dest	= (opcode >> 12) & 15;
  1.3871 +		int base	= (opcode >> 16) & 15;
  1.3872 +		u32 address = reg[base].I - offset;
  1.3873 +		CPUWriteMemory(address, reg[dest].I);
  1.3874 +		clockTicks += 2 + CPUUpdateTicksAccess32(address);
  1.3875 +	}
  1.3876 +	break;
  1.3877 +	case 0x702:
  1.3878 +	case 0x70a:
  1.3879 +	{
  1.3880 +		// STR Rd, [Rn, -Rm, LSR #]
  1.3881 +		int shift	= (opcode >> 7) & 31;
  1.3882 +		int offset	= shift ? reg[opcode & 15].I >> shift : 0;
  1.3883 +		int dest	= (opcode >> 12) & 15;
  1.3884 +		int base	= (opcode >> 16) & 15;
  1.3885 +		u32 address = reg[base].I - offset;
  1.3886 +		CPUWriteMemory(address, reg[dest].I);
  1.3887 +		clockTicks += 2 + CPUUpdateTicksAccess32(address);
  1.3888 +	}
  1.3889 +	break;
  1.3890 +	case 0x704:
  1.3891 +	case 0x70c:
  1.3892 +	{
  1.3893 +		// STR Rd, [Rn, -Rm, ASR #]
  1.3894 +		int shift = (opcode >> 7) & 31;
  1.3895 +		int offset;
  1.3896 +		if (shift)
  1.3897 +			offset = (int)((s32)reg[opcode & 15].I >> shift);
  1.3898 +		else if (reg[opcode & 15].I & 0x80000000)
  1.3899 +			offset = 0xFFFFFFFF;
  1.3900 +		else
  1.3901 +			offset = 0;
  1.3902 +		int dest	= (opcode >> 12) & 15;
  1.3903 +		int base	= (opcode >> 16) & 15;
  1.3904 +		u32 address = reg[base].I - offset;
  1.3905 +		CPUWriteMemory(address, reg[dest].I);
  1.3906 +		clockTicks += 2 + CPUUpdateTicksAccess32(address);
  1.3907 +	}
  1.3908 +	break;
  1.3909 +	case 0x706:
  1.3910 +	case 0x70e:
  1.3911 +	{
  1.3912 +		// STR Rd, [Rn, -Rm, ROR #]
  1.3913 +		int shift = (opcode >> 7) & 31;
  1.3914 +		u32 value = reg[opcode & 15].I;
  1.3915 +		if (shift)
  1.3916 +		{
  1.3917 +			ROR_VALUE;
  1.3918 +		}
  1.3919 +		else
  1.3920 +		{
  1.3921 +			RCR_VALUE;
  1.3922 +		}
  1.3923 +		int dest	= (opcode >> 12) & 15;
  1.3924 +		int base	= (opcode >> 16) & 15;
  1.3925 +		u32 address = reg[base].I - value;
  1.3926 +		CPUWriteMemory(address, reg[dest].I);
  1.3927 +		clockTicks += 2 + CPUUpdateTicksAccess32(address);
  1.3928 +	}
  1.3929 +	break;
  1.3930 +	case 0x720:
  1.3931 +	case 0x728:
  1.3932 +	{
  1.3933 +		// STR Rd, [Rn, -Rm, LSL #]!
  1.3934 +		int offset	= reg[opcode & 15].I << ((opcode >> 7) & 31);
  1.3935 +		int dest	= (opcode >> 12) & 15;
  1.3936 +		int base	= (opcode >> 16) & 15;
  1.3937 +		u32 address = reg[base].I - offset;
  1.3938 +		reg[base].I = address;
  1.3939 +		CPUWriteMemory(address, reg[dest].I);
  1.3940 +		clockTicks += 2 + CPUUpdateTicksAccess32(address);
  1.3941 +	}
  1.3942 +	break;
  1.3943 +	case 0x722:
  1.3944 +	case 0x72a:
  1.3945 +	{
  1.3946 +		// STR Rd, [Rn, -Rm, LSR #]!
  1.3947 +		int shift	= (opcode >> 7) & 31;
  1.3948 +		int offset	= shift ? reg[opcode & 15].I >> shift : 0;
  1.3949 +		int dest	= (opcode >> 12) & 15;
  1.3950 +		int base	= (opcode >> 16) & 15;
  1.3951 +		u32 address = reg[base].I - offset;
  1.3952 +		reg[base].I = address;
  1.3953 +		CPUWriteMemory(address, reg[dest].I);
  1.3954 +		clockTicks += 2 + CPUUpdateTicksAccess32(address);
  1.3955 +	}
  1.3956 +	break;
  1.3957 +	case 0x724:
  1.3958 +	case 0x72c:
  1.3959 +	{
  1.3960 +		// STR Rd, [Rn, -Rm, ASR #]!
  1.3961 +		int shift = (opcode >> 7) & 31;
  1.3962 +		int offset;
  1.3963 +		if (shift)
  1.3964 +			offset = (int)((s32)reg[opcode & 15].I >> shift);
  1.3965 +		else if (reg[opcode & 15].I & 0x80000000)
  1.3966 +			offset = 0xFFFFFFFF;
  1.3967 +		else
  1.3968 +			offset = 0;
  1.3969 +		int dest	= (opcode >> 12) & 15;
  1.3970 +		int base	= (opcode >> 16) & 15;
  1.3971 +		u32 address = reg[base].I - offset;
  1.3972 +		reg[base].I = address;
  1.3973 +		CPUWriteMemory(address, reg[dest].I);
  1.3974 +		clockTicks += 2 + CPUUpdateTicksAccess32(address);
  1.3975 +	}
  1.3976 +	break;
  1.3977 +	case 0x726:
  1.3978 +	case 0x72e:
  1.3979 +	{
  1.3980 +		// STR Rd, [Rn, -Rm, ROR #]!
  1.3981 +		int shift = (opcode >> 7) & 31;
  1.3982 +		u32 value = reg[opcode & 15].I;
  1.3983 +		if (shift)
  1.3984 +		{
  1.3985 +			ROR_VALUE;
  1.3986 +		}
  1.3987 +		else
  1.3988 +		{
  1.3989 +			RCR_VALUE;
  1.3990 +		}
  1.3991 +		int dest	= (opcode >> 12) & 15;
  1.3992 +		int base	= (opcode >> 16) & 15;
  1.3993 +		u32 address = reg[base].I - value;
  1.3994 +		reg[base].I = address;
  1.3995 +		CPUWriteMemory(address, reg[dest].I);
  1.3996 +		clockTicks += 2 + CPUUpdateTicksAccess32(address);
  1.3997 +	}
  1.3998 +	break;
  1.3999 +	case 0x780:
  1.4000 +	case 0x788:
  1.4001 +	{
  1.4002 +		// STR Rd, [Rn, Rm, LSL #]
  1.4003 +		int offset	= reg[opcode & 15].I << ((opcode >> 7) & 31);
  1.4004 +		int dest	= (opcode >> 12) & 15;
  1.4005 +		int base	= (opcode >> 16) & 15;
  1.4006 +		u32 address = reg[base].I + offset;
  1.4007 +		CPUWriteMemory(address, reg[dest].I);
  1.4008 +		clockTicks += 2 + CPUUpdateTicksAccess32(address);
  1.4009 +	}
  1.4010 +	break;
  1.4011 +	case 0x782:
  1.4012 +	case 0x78a:
  1.4013 +	{
  1.4014 +		// STR Rd, [Rn, Rm, LSR #]
  1.4015 +		int shift	= (opcode >> 7) & 31;
  1.4016 +		int offset	= shift ? reg[opcode & 15].I >> shift : 0;
  1.4017 +		int dest	= (opcode >> 12) & 15;
  1.4018 +		int base	= (opcode >> 16) & 15;
  1.4019 +		u32 address = reg[base].I + offset;
  1.4020 +		CPUWriteMemory(address, reg[dest].I);
  1.4021 +		clockTicks += 2 + CPUUpdateTicksAccess32(address);
  1.4022 +	}
  1.4023 +	break;
  1.4024 +	case 0x784:
  1.4025 +	case 0x78c:
  1.4026 +	{
  1.4027 +		// STR Rd, [Rn, Rm, ASR #]
  1.4028 +		int shift = (opcode >> 7) & 31;
  1.4029 +		int offset;
  1.4030 +		if (shift)
  1.4031 +			offset = (int)((s32)reg[opcode & 15].I >> shift);
  1.4032 +		else if (reg[opcode & 15].I & 0x80000000)
  1.4033 +			offset = 0xFFFFFFFF;
  1.4034 +		else
  1.4035 +			offset = 0;
  1.4036 +		int dest	= (opcode >> 12) & 15;
  1.4037 +		int base	= (opcode >> 16) & 15;
  1.4038 +		u32 address = reg[base].I + offset;
  1.4039 +		CPUWriteMemory(address, reg[dest].I);
  1.4040 +		clockTicks += 2 + CPUUpdateTicksAccess32(address);
  1.4041 +	}
  1.4042 +	break;
  1.4043 +	case 0x786:
  1.4044 +	case 0x78e:
  1.4045 +	{
  1.4046 +		// STR Rd, [Rn, Rm, ROR #]
  1.4047 +		int shift = (opcode >> 7) & 31;
  1.4048 +		u32 value = reg[opcode & 15].I;
  1.4049 +		if (shift)
  1.4050 +		{
  1.4051 +			ROR_VALUE;
  1.4052 +		}
  1.4053 +		else
  1.4054 +		{
  1.4055 +			RCR_VALUE;
  1.4056 +		}
  1.4057 +		int dest	= (opcode >> 12) & 15;
  1.4058 +		int base	= (opcode >> 16) & 15;
  1.4059 +		u32 address = reg[base].I + value;
  1.4060 +		CPUWriteMemory(address, reg[dest].I);
  1.4061 +		clockTicks += 2 + CPUUpdateTicksAccess32(address);
  1.4062 +	}
  1.4063 +	break;
  1.4064 +	case 0x7a0:
  1.4065 +	case 0x7a8:
  1.4066 +	{
  1.4067 +		// STR Rd, [Rn, Rm, LSL #]!
  1.4068 +		int offset	= reg[opcode & 15].I << ((opcode >> 7) & 31);
  1.4069 +		int dest	= (opcode >> 12) & 15;
  1.4070 +		int base	= (opcode >> 16) & 15;
  1.4071 +		u32 address = reg[base].I + offset;
  1.4072 +		reg[base].I = address;
  1.4073 +		CPUWriteMemory(address, reg[dest].I);
  1.4074 +		clockTicks += 2 + CPUUpdateTicksAccess32(address);
  1.4075 +	}
  1.4076 +	break;
  1.4077 +	case 0x7a2:
  1.4078 +	case 0x7aa:
  1.4079 +	{
  1.4080 +		// STR Rd, [Rn, Rm, LSR #]!
  1.4081 +		int shift	= (opcode >> 7) & 31;
  1.4082 +		int offset	= shift ? reg[opcode & 15].I >> shift : 0;
  1.4083 +		int dest	= (opcode >> 12) & 15;
  1.4084 +		int base	= (opcode >> 16) & 15;
  1.4085 +		u32 address = reg[base].I + offset;
  1.4086 +		reg[base].I = address;
  1.4087 +		CPUWriteMemory(address, reg[dest].I);
  1.4088 +		clockTicks += 2 + CPUUpdateTicksAccess32(address);
  1.4089 +	}
  1.4090 +	break;
  1.4091 +	case 0x7a4:
  1.4092 +	case 0x7ac:
  1.4093 +	{
  1.4094 +		// STR Rd, [Rn, Rm, ASR #]!
  1.4095 +		int shift = (opcode >> 7) & 31;
  1.4096 +		int offset;
  1.4097 +		if (shift)
  1.4098 +			offset = (int)((s32)reg[opcode & 15].I >> shift);
  1.4099 +		else if (reg[opcode & 15].I & 0x80000000)
  1.4100 +			offset = 0xFFFFFFFF;
  1.4101 +		else
  1.4102 +			offset = 0;
  1.4103 +		int dest	= (opcode >> 12) & 15;
  1.4104 +		int base	= (opcode >> 16) & 15;
  1.4105 +		u32 address = reg[base].I + offset;
  1.4106 +		reg[base].I = address;
  1.4107 +		CPUWriteMemory(address, reg[dest].I);
  1.4108 +		clockTicks += 2 + CPUUpdateTicksAccess32(address);
  1.4109 +	}
  1.4110 +	break;
  1.4111 +	case 0x7a6:
  1.4112 +	case 0x7ae:
  1.4113 +	{
  1.4114 +		// STR Rd, [Rn, Rm, ROR #]!
  1.4115 +		int shift = (opcode >> 7) & 31;
  1.4116 +		u32 value = reg[opcode & 15].I;
  1.4117 +		if (shift)
  1.4118 +		{
  1.4119 +			ROR_VALUE;
  1.4120 +		}
  1.4121 +		else
  1.4122 +		{
  1.4123 +			RCR_VALUE;
  1.4124 +		}
  1.4125 +		int dest	= (opcode >> 12) & 15;
  1.4126 +		int base	= (opcode >> 16) & 15;
  1.4127 +		u32 address = reg[base].I + value;
  1.4128 +		reg[base].I = address;
  1.4129 +		CPUWriteMemory(address, reg[dest].I);
  1.4130 +		clockTicks += 2 + CPUUpdateTicksAccess32(address);
  1.4131 +	}
  1.4132 +	break;
  1.4133 +	case 0x610:
  1.4134 +	case 0x618:
  1.4135 +	// T versions are the same
  1.4136 +	case 0x630:
  1.4137 +	case 0x638:
  1.4138 +	{
  1.4139 +		// LDR Rd, [Rn], -Rm, LSL #
  1.4140 +		int offset	= reg[opcode & 15].I << ((opcode >> 7) & 31);
  1.4141 +		int dest	= (opcode >> 12) & 15;
  1.4142 +		int base	= (opcode >> 16) & 15;
  1.4143 +		u32 address = reg[base].I;
  1.4144 +		reg[dest].I = CPUReadMemory(address);
  1.4145 +		if (dest != base)
  1.4146 +			reg[base].I = address - offset;
  1.4147 +		clockTicks += 3 + CPUUpdateTicksAccess32(address);
  1.4148 +		if (dest == 15)
  1.4149 +		{
  1.4150 +			clockTicks += 2;
  1.4151 +			reg[15].I  &= 0xFFFFFFFC;
  1.4152 +			armNextPC	= reg[15].I;
  1.4153 +			reg[15].I  += 4;
  1.4154 +		}
  1.4155 +	}
  1.4156 +	break;
  1.4157 +	case 0x612:
  1.4158 +	case 0x61a:
  1.4159 +	// T versions are the same
  1.4160 +	case 0x632:
  1.4161 +	case 0x63a:
  1.4162 +	{
  1.4163 +		// LDR Rd, [Rn], -Rm, LSR #
  1.4164 +		int shift	= (opcode >> 7) & 31;
  1.4165 +		int offset	= shift ? reg[opcode & 15].I >> shift : 0;
  1.4166 +		int dest	= (opcode >> 12) & 15;
  1.4167 +		int base	= (opcode >> 16) & 15;
  1.4168 +		u32 address = reg[base].I;
  1.4169 +		reg[dest].I = CPUReadMemory(address);
  1.4170 +		if (dest != base)
  1.4171 +			reg[base].I = address - offset;
  1.4172 +		clockTicks += 3 + CPUUpdateTicksAccess32(address);
  1.4173 +		if (dest == 15)
  1.4174 +		{
  1.4175 +			clockTicks += 2;
  1.4176 +			reg[15].I  &= 0xFFFFFFFC;
  1.4177 +			armNextPC	= reg[15].I;
  1.4178 +			reg[15].I  += 4;
  1.4179 +		}
  1.4180 +	}
  1.4181 +	break;
  1.4182 +	case 0x614:
  1.4183 +	case 0x61c:
  1.4184 +	// T versions are the same
  1.4185 +	case 0x634:
  1.4186 +	case 0x63c:
  1.4187 +	{
  1.4188 +		// LDR Rd, [Rn], -Rm, ASR #
  1.4189 +		int shift = (opcode >> 7) & 31;
  1.4190 +		int offset;
  1.4191 +		if (shift)
  1.4192 +			offset = (int)((s32)reg[opcode & 15].I >> shift);
  1.4193 +		else if (reg[opcode & 15].I & 0x80000000)
  1.4194 +			offset = 0xFFFFFFFF;
  1.4195 +		else
  1.4196 +			offset = 0;
  1.4197 +		int dest	= (opcode >> 12) & 15;
  1.4198 +		int base	= (opcode >> 16) & 15;
  1.4199 +		u32 address = reg[base].I;
  1.4200 +		reg[dest].I = CPUReadMemory(address);
  1.4201 +		if (dest != base)
  1.4202 +			reg[base].I = address - offset;
  1.4203 +		clockTicks += 3 + CPUUpdateTicksAccess32(address);
  1.4204 +		if (dest == 15)
  1.4205 +		{
  1.4206 +			clockTicks += 2;
  1.4207 +			reg[15].I  &= 0xFFFFFFFC;
  1.4208 +			armNextPC	= reg[15].I;
  1.4209 +			reg[15].I  += 4;
  1.4210 +		}
  1.4211 +	}
  1.4212 +	break;
  1.4213 +	case 0x616:
  1.4214 +	case 0x61e:
  1.4215 +	// T versions are the same
  1.4216 +	case 0x636:
  1.4217 +	case 0x63e:
  1.4218 +	{
  1.4219 +		// LDR Rd, [Rn], -Rm, ROR #
  1.4220 +		int shift = (opcode >> 7) & 31;
  1.4221 +		u32 value = reg[opcode & 15].I;
  1.4222 +		if (shift)
  1.4223 +		{
  1.4224 +			ROR_VALUE;
  1.4225 +		}
  1.4226 +		else
  1.4227 +		{
  1.4228 +			RCR_VALUE;
  1.4229 +		}
  1.4230 +		int dest	= (opcode >> 12) & 15;
  1.4231 +		int base	= (opcode >> 16) & 15;
  1.4232 +		u32 address = reg[base].I;
  1.4233 +		reg[dest].I = CPUReadMemory(address);
  1.4234 +		if (dest != base)
  1.4235 +			reg[base].I = address - value;
  1.4236 +		clockTicks += 3 + CPUUpdateTicksAccess32(address);
  1.4237 +		if (dest == 15)
  1.4238 +		{
  1.4239 +			clockTicks += 2;
  1.4240 +			reg[15].I  &= 0xFFFFFFFC;
  1.4241 +			armNextPC	= reg[15].I;
  1.4242 +			reg[15].I  += 4;
  1.4243 +		}
  1.4244 +	}
  1.4245 +	break;
  1.4246 +	case 0x690:
  1.4247 +	case 0x698:
  1.4248 +	// T versions are the same
  1.4249 +	case 0x6b0:
  1.4250 +	case 0x6b8:
  1.4251 +	{
  1.4252 +		// LDR Rd, [Rn], Rm, LSL #
  1.4253 +		int offset	= reg[opcode & 15].I << ((opcode >> 7) & 31);
  1.4254 +		int dest	= (opcode >> 12) & 15;
  1.4255 +		int base	= (opcode >> 16) & 15;
  1.4256 +		u32 address = reg[base].I;
  1.4257 +		reg[dest].I = CPUReadMemory(address);
  1.4258 +		if (dest != base)
  1.4259 +			reg[base].I = address + offset;
  1.4260 +		clockTicks += 3 + CPUUpdateTicksAccess32(address);
  1.4261 +		if (dest == 15)
  1.4262 +		{
  1.4263 +			clockTicks += 2;
  1.4264 +			reg[15].I  &= 0xFFFFFFFC;
  1.4265 +			armNextPC	= reg[15].I;
  1.4266 +			reg[15].I  += 4;
  1.4267 +		}
  1.4268 +	}
  1.4269 +	break;
  1.4270 +	case 0x692:
  1.4271 +	case 0x69a:
  1.4272 +	// T versions are the same
  1.4273 +	case 0x6b2:
  1.4274 +	case 0x6ba:
  1.4275 +	{
  1.4276 +		// LDR Rd, [Rn], Rm, LSR #
  1.4277 +		int shift	= (opcode >> 7) & 31;
  1.4278 +		int offset	= shift ? reg[opcode & 15].I >> shift : 0;
  1.4279 +		int dest	= (opcode >> 12) & 15;
  1.4280 +		int base	= (opcode >> 16) & 15;
  1.4281 +		u32 address = reg[base].I;
  1.4282 +		reg[dest].I = CPUReadMemory(address);
  1.4283 +		if (dest != base)
  1.4284 +			reg[base].I = address + offset;
  1.4285 +		clockTicks += 3 + CPUUpdateTicksAccess32(address);
  1.4286 +		if (dest == 15)
  1.4287 +		{
  1.4288 +			clockTicks += 2;
  1.4289 +			reg[15].I  &= 0xFFFFFFFC;
  1.4290 +			armNextPC	= reg[15].I;
  1.4291 +			reg[15].I  += 4;
  1.4292 +		}
  1.4293 +	}
  1.4294 +	break;
  1.4295 +	case 0x694:
  1.4296 +	case 0x69c:
  1.4297 +	// T versions are the same
  1.4298 +	case 0x6b4:
  1.4299 +	case 0x6bc:
  1.4300 +	{
  1.4301 +		// LDR Rd, [Rn], Rm, ASR #
  1.4302 +		int shift = (opcode >> 7) & 31;
  1.4303 +		int offset;
  1.4304 +		if (shift)
  1.4305 +			offset = (int)((s32)reg[opcode & 15].I >> shift);
  1.4306 +		else if (reg[opcode & 15].I & 0x80000000)
  1.4307 +			offset = 0xFFFFFFFF;
  1.4308 +		else
  1.4309 +			offset = 0;
  1.4310 +		int dest	= (opcode >> 12) & 15;
  1.4311 +		int base	= (opcode >> 16) & 15;
  1.4312 +		u32 address = reg[base].I;
  1.4313 +		reg[dest].I = CPUReadMemory(address);
  1.4314 +		if (dest != base)
  1.4315 +			reg[base].I = address + offset;
  1.4316 +		clockTicks += 3 + CPUUpdateTicksAccess32(address);
  1.4317 +		if (dest == 15)
  1.4318 +		{
  1.4319 +			clockTicks += 2;
  1.4320 +			reg[15].I  &= 0xFFFFFFFC;
  1.4321 +			armNextPC	= reg[15].I;
  1.4322 +			reg[15].I  += 4;
  1.4323 +		}
  1.4324 +	}
  1.4325 +	break;
  1.4326 +	case 0x696:
  1.4327 +	case 0x69e:
  1.4328 +	// T versions are the same
  1.4329 +	case 0x6b6:
  1.4330 +	case 0x6be:
  1.4331 +	{
  1.4332 +		// LDR Rd, [Rn], Rm, ROR #
  1.4333 +		int shift = (opcode >> 7) & 31;
  1.4334 +		u32 value = reg[opcode & 15].I;
  1.4335 +		if (shift)
  1.4336 +		{
  1.4337 +			ROR_VALUE;
  1.4338 +		}
  1.4339 +		else
  1.4340 +		{
  1.4341 +			RCR_VALUE;
  1.4342 +		}
  1.4343 +		int dest	= (opcode >> 12) & 15;
  1.4344 +		int base	= (opcode >> 16) & 15;
  1.4345 +		u32 address = reg[base].I;
  1.4346 +		reg[dest].I = CPUReadMemory(address);
  1.4347 +		if (dest != base)
  1.4348 +			reg[base].I = address + value;
  1.4349 +		clockTicks += 3 + CPUUpdateTicksAccess32(address);
  1.4350 +		if (dest == 15)
  1.4351 +		{
  1.4352 +			clockTicks += 2;
  1.4353 +			reg[15].I  &= 0xFFFFFFFC;
  1.4354 +			armNextPC	= reg[15].I;
  1.4355 +			reg[15].I  += 4;
  1.4356 +		}
  1.4357 +	}
  1.4358 +	break;
  1.4359 +	case 0x710:
  1.4360 +	case 0x718:
  1.4361 +	{
  1.4362 +		// LDR Rd, [Rn, -Rm, LSL #]
  1.4363 +		int offset	= reg[opcode & 15].I << ((opcode >> 7) & 31);
  1.4364 +		int dest	= (opcode >> 12) & 15;
  1.4365 +		int base	= (opcode >> 16) & 15;
  1.4366 +		u32 address = reg[base].I - offset;
  1.4367 +		reg[dest].I = CPUReadMemory(address);
  1.4368 +		clockTicks += 3 + CPUUpdateTicksAccess32(address);
  1.4369 +		if (dest == 15)
  1.4370 +		{
  1.4371 +			clockTicks += 2;
  1.4372 +			reg[15].I  &= 0xFFFFFFFC;
  1.4373 +			armNextPC	= reg[15].I;
  1.4374 +			reg[15].I  += 4;
  1.4375 +		}
  1.4376 +	}
  1.4377 +	break;
  1.4378 +	case 0x712:
  1.4379 +	case 0x71a:
  1.4380 +	{
  1.4381 +		// LDR Rd, [Rn, -Rm, LSR #]
  1.4382 +		int shift	= (opcode >> 7) & 31;
  1.4383 +		int offset	= shift ? reg[opcode & 15].I >> shift : 0;
  1.4384 +		int dest	= (opcode >> 12) & 15;
  1.4385 +		int base	= (opcode >> 16) & 15;
  1.4386 +		u32 address = reg[base].I - offset;
  1.4387 +		reg[dest].I = CPUReadMemory(address);
  1.4388 +		clockTicks += 3 + CPUUpdateTicksAccess32(address);
  1.4389 +		if (dest == 15)
  1.4390 +		{
  1.4391 +			clockTicks += 2;
  1.4392 +			reg[15].I  &= 0xFFFFFFFC;
  1.4393 +			armNextPC	= reg[15].I;
  1.4394 +			reg[15].I  += 4;
  1.4395 +		}
  1.4396 +	}
  1.4397 +	break;
  1.4398 +	case 0x714:
  1.4399 +	case 0x71c:
  1.4400 +	{
  1.4401 +		// LDR Rd, [Rn, -Rm, ASR #]
  1.4402 +		int shift = (opcode >> 7) & 31;
  1.4403 +		int offset;
  1.4404 +		if (shift)
  1.4405 +			offset = (int)((s32)reg[opcode & 15].I >> shift);
  1.4406 +		else if (reg[opcode & 15].I & 0x80000000)
  1.4407 +			offset = 0xFFFFFFFF;
  1.4408 +		else
  1.4409 +			offset = 0;
  1.4410 +		int dest	= (opcode >> 12) & 15;
  1.4411 +		int base	= (opcode >> 16) & 15;
  1.4412 +		u32 address = reg[base].I - offset;
  1.4413 +		reg[dest].I = CPUReadMemory(address);
  1.4414 +		clockTicks += 3 + CPUUpdateTicksAccess32(address);
  1.4415 +		if (dest == 15)
  1.4416 +		{
  1.4417 +			clockTicks += 2;
  1.4418 +			reg[15].I  &= 0xFFFFFFFC;
  1.4419 +			armNextPC	= reg[15].I;
  1.4420 +			reg[15].I  += 4;
  1.4421 +		}
  1.4422 +	}
  1.4423 +	break;
  1.4424 +	case 0x716:
  1.4425 +	case 0x71e:
  1.4426 +	{
  1.4427 +		// LDR Rd, [Rn, -Rm, ROR #]
  1.4428 +		int shift = (opcode >> 7) & 31;
  1.4429 +		u32 value = reg[opcode & 15].I;
  1.4430 +		if (shift)
  1.4431 +		{
  1.4432 +			ROR_VALUE;
  1.4433 +		}
  1.4434 +		else
  1.4435 +		{
  1.4436 +			RCR_VALUE;
  1.4437 +		}
  1.4438 +		int dest	= (opcode >> 12) & 15;
  1.4439 +		int base	= (opcode >> 16) & 15;
  1.4440 +		u32 address = reg[base].I - value;
  1.4441 +		reg[dest].I = CPUReadMemory(address);
  1.4442 +		clockTicks += 3 + CPUUpdateTicksAccess32(address);
  1.4443 +		if (dest == 15)
  1.4444 +		{
  1.4445 +			clockTicks += 2;
  1.4446 +			reg[15].I  &= 0xFFFFFFFC;
  1.4447 +			armNextPC	= reg[15].I;
  1.4448 +			reg[15].I  += 4;
  1.4449 +		}
  1.4450 +	}
  1.4451 +	break;
  1.4452 +	case 0x730:
  1.4453 +	case 0x738:
  1.4454 +	{
  1.4455 +		// LDR Rd, [Rn, -Rm, LSL #]!
  1.4456 +		int offset	= reg[opcode & 15].I << ((opcode >> 7) & 31);
  1.4457 +		int dest	= (opcode >> 12) & 15;
  1.4458 +		int base	= (opcode >> 16) & 15;
  1.4459 +		u32 address = reg[base].I - offset;
  1.4460 +		reg[dest].I = CPUReadMemory(address);
  1.4461 +		if (dest != base)
  1.4462 +			reg[base].I = address;
  1.4463 +		clockTicks += 3 + CPUUpdateTicksAccess32(address);
  1.4464 +		if (dest == 15)
  1.4465 +		{
  1.4466 +			clockTicks += 2;
  1.4467 +			reg[15].I  &= 0xFFFFFFFC;
  1.4468 +			armNextPC	= reg[15].I;
  1.4469 +			reg[15].I  += 4;
  1.4470 +		}
  1.4471 +	}
  1.4472 +	break;
  1.4473 +	case 0x732:
  1.4474 +	case 0x73a:
  1.4475 +	{
  1.4476 +		// LDR Rd, [Rn, -Rm, LSR #]!
  1.4477 +		int shift	= (opcode >> 7) & 31;
  1.4478 +		int offset	= shift ? reg[opcode & 15].I >> shift : 0;
  1.4479 +		int dest	= (opcode >> 12) & 15;
  1.4480 +		int base	= (opcode >> 16) & 15;
  1.4481 +		u32 address = reg[base].I - offset;
  1.4482 +		reg[dest].I = CPUReadMemory(address);
  1.4483 +		if (dest != base)
  1.4484 +			reg[base].I = address;
  1.4485 +		clockTicks += 3 + CPUUpdateTicksAccess32(address);
  1.4486 +		if (dest == 15)
  1.4487 +		{
  1.4488 +			clockTicks += 2;
  1.4489 +			reg[15].I  &= 0xFFFFFFFC;
  1.4490 +			armNextPC	= reg[15].I;
  1.4491 +			reg[15].I  += 4;
  1.4492 +		}
  1.4493 +	}
  1.4494 +	break;
  1.4495 +	case 0x734:
  1.4496 +	case 0x73c:
  1.4497 +	{
  1.4498 +		// LDR Rd, [Rn, -Rm, ASR #]!
  1.4499 +		int shift = (opcode >> 7) & 31;
  1.4500 +		int offset;
  1.4501 +		if (shift)
  1.4502 +			offset = (int)((s32)reg[opcode & 15].I >> shift);
  1.4503 +		else if (reg[opcode & 15].I & 0x80000000)
  1.4504 +			offset = 0xFFFFFFFF;
  1.4505 +		else
  1.4506 +			offset = 0;
  1.4507 +		int dest	= (opcode >> 12) & 15;
  1.4508 +		int base	= (opcode >> 16) & 15;
  1.4509 +		u32 address = reg[base].I - offset;
  1.4510 +		reg[dest].I = CPUReadMemory(address);
  1.4511 +		if (dest != base)
  1.4512 +			reg[base].I = address;
  1.4513 +		clockTicks += 3 + CPUUpdateTicksAccess32(address);
  1.4514 +		if (dest == 15)
  1.4515 +		{
  1.4516 +			clockTicks += 2;
  1.4517 +			reg[15].I  &= 0xFFFFFFFC;
  1.4518 +			armNextPC	= reg[15].I;
  1.4519 +			reg[15].I  += 4;
  1.4520 +		}
  1.4521 +	}
  1.4522 +	break;
  1.4523 +	case 0x736:
  1.4524 +	case 0x73e:
  1.4525 +	{
  1.4526 +		// LDR Rd, [Rn, -Rm, ROR #]!
  1.4527 +		int shift = (opcode >> 7) & 31;
  1.4528 +		u32 value = reg[opcode & 15].I;
  1.4529 +		if (shift)
  1.4530 +		{
  1.4531 +			ROR_VALUE;
  1.4532 +		}
  1.4533 +		else
  1.4534 +		{
  1.4535 +			RCR_VALUE;
  1.4536 +		}
  1.4537 +		int dest	= (opcode >> 12) & 15;
  1.4538 +		int base	= (opcode >> 16) & 15;
  1.4539 +		u32 address = reg[base].I - value;
  1.4540 +		reg[dest].I = CPUReadMemory(address);
  1.4541 +		if (dest != base)
  1.4542 +			reg[base].I = address;
  1.4543 +		clockTicks += 3 + CPUUpdateTicksAccess32(address);
  1.4544 +		if (dest == 15)
  1.4545 +		{
  1.4546 +			clockTicks += 2;
  1.4547 +			reg[15].I  &= 0xFFFFFFFC;
  1.4548 +			armNextPC	= reg[15].I;
  1.4549 +			reg[15].I  += 4;
  1.4550 +		}
  1.4551 +	}
  1.4552 +	break;
  1.4553 +	case 0x790:
  1.4554 +	case 0x798:
  1.4555 +	{
  1.4556 +		// LDR Rd, [Rn, Rm, LSL #]
  1.4557 +		int offset	= reg[opcode & 15].I << ((opcode >> 7) & 31);
  1.4558 +		int dest	= (opcode >> 12) & 15;
  1.4559 +		int base	= (opcode >> 16) & 15;
  1.4560 +		u32 address = reg[base].I + offset;
  1.4561 +		reg[dest].I = CPUReadMemory(address);
  1.4562 +		clockTicks += 3 + CPUUpdateTicksAccess32(address);
  1.4563 +		if (dest == 15)
  1.4564 +		{
  1.4565 +			clockTicks += 2;
  1.4566 +			reg[15].I  &= 0xFFFFFFFC;
  1.4567 +			armNextPC	= reg[15].I;
  1.4568 +			reg[15].I  += 4;
  1.4569 +		}
  1.4570 +	}
  1.4571 +	break;
  1.4572 +	case 0x792:
  1.4573 +	case 0x79a:
  1.4574 +	{
  1.4575 +		// LDR Rd, [Rn, Rm, LSR #]
  1.4576 +		int shift	= (opcode >> 7) & 31;
  1.4577 +		int offset	= shift ? reg[opcode & 15].I >> shift : 0;
  1.4578 +		int dest	= (opcode >> 12) & 15;
  1.4579 +		int base	= (opcode >> 16) & 15;
  1.4580 +		u32 address = reg[base].I + offset;
  1.4581 +		reg[dest].I = CPUReadMemory(address);
  1.4582 +		clockTicks += 3 + CPUUpdateTicksAccess32(address);
  1.4583 +		if (dest == 15)
  1.4584 +		{
  1.4585 +			clockTicks += 2;
  1.4586 +			reg[15].I  &= 0xFFFFFFFC;
  1.4587 +			armNextPC	= reg[15].I;
  1.4588 +			reg[15].I  += 4;
  1.4589 +		}
  1.4590 +	}
  1.4591 +	break;
  1.4592 +	case 0x794:
  1.4593 +	case 0x79c:
  1.4594 +	{
  1.4595 +		// LDR Rd, [Rn, Rm, ASR #]
  1.4596 +		int shift = (opcode >> 7) & 31;
  1.4597 +		int offset;
  1.4598 +		if (shift)
  1.4599 +			offset = (int)((s32)reg[opcode & 15].I >> shift);
  1.4600 +		else if (reg[opcode & 15].I & 0x80000000)
  1.4601 +			offset = 0xFFFFFFFF;
  1.4602 +		else
  1.4603 +			offset = 0;
  1.4604 +		int dest	= (opcode >> 12) & 15;
  1.4605 +		int base	= (opcode >> 16) & 15;
  1.4606 +		u32 address = reg[base].I + offset;
  1.4607 +		reg[dest].I = CPUReadMemory(address);
  1.4608 +		clockTicks += 3 + CPUUpdateTicksAccess32(address);
  1.4609 +		if (dest == 15)
  1.4610 +		{
  1.4611 +			clockTicks += 2;
  1.4612 +			reg[15].I  &= 0xFFFFFFFC;
  1.4613 +			armNextPC	= reg[15].I;
  1.4614 +			reg[15].I  += 4;
  1.4615 +		}
  1.4616 +	}
  1.4617 +	break;
  1.4618 +	case 0x796:
  1.4619 +	case 0x79e:
  1.4620 +	{
  1.4621 +		// LDR Rd, [Rn, Rm, ROR #]
  1.4622 +		int shift = (opcode >> 7) & 31;
  1.4623 +		u32 value = reg[opcode & 15].I;
  1.4624 +		if (shift)
  1.4625 +		{
  1.4626 +			ROR_VALUE;
  1.4627 +		}
  1.4628 +		else
  1.4629 +		{
  1.4630 +			RCR_VALUE;
  1.4631 +		}
  1.4632 +		int dest	= (opcode >> 12) & 15;
  1.4633 +		int base	= (opcode >> 16) & 15;
  1.4634 +		u32 address = reg[base].I + value;
  1.4635 +		reg[dest].I = CPUReadMemory(address);
  1.4636 +		clockTicks += 3 + CPUUpdateTicksAccess32(address);
  1.4637 +		if (dest == 15)
  1.4638 +		{
  1.4639 +			clockTicks += 2;
  1.4640 +			reg[15].I  &= 0xFFFFFFFC;
  1.4641 +			armNextPC	= reg[15].I;
  1.4642 +			reg[15].I  += 4;
  1.4643 +		}
  1.4644 +	}
  1.4645 +	break;
  1.4646 +	case 0x7b0:
  1.4647 +	case 0x7b8:
  1.4648 +	{
  1.4649 +		// LDR Rd, [Rn, Rm, LSL #]!
  1.4650 +		int offset	= reg[opcode & 15].I << ((opcode >> 7) & 31);
  1.4651 +		int dest	= (opcode >> 12) & 15;
  1.4652 +		int base	= (opcode >> 16) & 15;
  1.4653 +		u32 address = reg[base].I + offset;
  1.4654 +		reg[dest].I = CPUReadMemory(address);
  1.4655 +		if (dest != base)
  1.4656 +			reg[base].I = address;
  1.4657 +		clockTicks += 3 + CPUUpdateTicksAccess32(address);
  1.4658 +		if (dest == 15)
  1.4659 +		{
  1.4660 +			clockTicks += 2;
  1.4661 +			reg[15].I  &= 0xFFFFFFFC;
  1.4662 +			armNextPC	= reg[15].I;
  1.4663 +			reg[15].I  += 4;
  1.4664 +		}
  1.4665 +	}
  1.4666 +	break;
  1.4667 +	case 0x7b2:
  1.4668 +	case 0x7ba:
  1.4669 +	{
  1.4670 +		// LDR Rd, [Rn, Rm, LSR #]!
  1.4671 +		int shift	= (opcode >> 7) & 31;
  1.4672 +		int offset	= shift ? reg[opcode & 15].I >> shift : 0;
  1.4673 +		int dest	= (opcode >> 12) & 15;
  1.4674 +		int base	= (opcode >> 16) & 15;
  1.4675 +		u32 address = reg[base].I + offset;
  1.4676 +		reg[dest].I = CPUReadMemory(address);
  1.4677 +		if (dest != base)
  1.4678 +			reg[base].I = address;
  1.4679 +		clockTicks += 3 + CPUUpdateTicksAccess32(address);
  1.4680 +		if (dest == 15)
  1.4681 +		{
  1.4682 +			clockTicks += 2;
  1.4683 +			reg[15].I  &= 0xFFFFFFFC;
  1.4684 +			armNextPC	= reg[15].I;
  1.4685 +			reg[15].I  += 4;
  1.4686 +		}
  1.4687 +	}
  1.4688 +	break;
  1.4689 +	case 0x7b4:
  1.4690 +	case 0x7bc:
  1.4691 +	{
  1.4692 +		// LDR Rd, [Rn, Rm, ASR #]!
  1.4693 +		int shift = (opcode >> 7) & 31;
  1.4694 +		int offset;
  1.4695 +		if (shift)
  1.4696 +			offset = (int)((s32)reg[opcode & 15].I >> shift);
  1.4697 +		else if (reg[opcode & 15].I & 0x80000000)
  1.4698 +			offset = 0xFFFFFFFF;
  1.4699 +		else
  1.4700 +			offset = 0;
  1.4701 +		int dest	= (opcode >> 12) & 15;
  1.4702 +		int base	= (opcode >> 16) & 15;
  1.4703 +		u32 address = reg[base].I + offset;
  1.4704 +		reg[dest].I = CPUReadMemory(address);
  1.4705 +		if (dest != base)
  1.4706 +			reg[base].I = address;
  1.4707 +		clockTicks += 3 + CPUUpdateTicksAccess32(address);
  1.4708 +		if (dest == 15)
  1.4709 +		{
  1.4710 +			clockTicks += 2;
  1.4711 +			reg[15].I  &= 0xFFFFFFFC;
  1.4712 +			armNextPC	= reg[15].I;
  1.4713 +			reg[15].I  += 4;
  1.4714 +		}
  1.4715 +	}
  1.4716 +	break;
  1.4717 +	case 0x7b6:
  1.4718 +	case 0x7be:
  1.4719 +	{
  1.4720 +		// LDR Rd, [Rn, Rm, ROR #]!
  1.4721 +		int shift = (opcode >> 7) & 31;
  1.4722 +		u32 value = reg[opcode & 15].I;
  1.4723 +		if (shift)
  1.4724 +		{
  1.4725 +			ROR_VALUE;
  1.4726 +		}
  1.4727 +		else
  1.4728 +		{
  1.4729 +			RCR_VALUE;
  1.4730 +		}
  1.4731 +		int dest	= (opcode >> 12) & 15;
  1.4732 +		int base	= (opcode >> 16) & 15;
  1.4733 +		u32 address = reg[base].I + value;
  1.4734 +		reg[dest].I = CPUReadMemory(address);
  1.4735 +		if (dest != base)
  1.4736 +			reg[base].I = address;
  1.4737 +		clockTicks += 3 + CPUUpdateTicksAccess32(address);
  1.4738 +		if (dest == 15)
  1.4739 +		{
  1.4740 +			clockTicks += 2;
  1.4741 +			reg[15].I  &= 0xFFFFFFFC;
  1.4742 +			armNextPC	= reg[15].I;
  1.4743 +			reg[15].I  += 4;
  1.4744 +		}
  1.4745 +	}
  1.4746 +	break;
  1.4747 +	case 0x640:
  1.4748 +	case 0x648:
  1.4749 +	// T versions are the same
  1.4750 +	case 0x660:
  1.4751 +	case 0x668:
  1.4752 +	{
  1.4753 +		// STRB Rd, [Rn], -Rm, LSL #
  1.4754 +		int offset	= reg[opcode & 15].I << ((opcode >> 7) & 31);
  1.4755 +		int dest	= (opcode >> 12) & 15;
  1.4756 +		int base	= (opcode >> 16) & 15;
  1.4757 +		u32 address = reg[base].I;
  1.4758 +		CPUWriteByte(address, reg[dest].B.B0);
  1.4759 +		reg[base].I = address - offset;
  1.4760 +		clockTicks += 2 + CPUUpdateTicksAccess16(address);
  1.4761 +	}
  1.4762 +	break;
  1.4763 +	case 0x642:
  1.4764 +	case 0x64a:
  1.4765 +	// T versions are the same
  1.4766 +	case 0x662:
  1.4767 +	case 0x66a:
  1.4768 +	{
  1.4769 +		// STRB Rd, [Rn], -Rm, LSR #
  1.4770 +		int shift	= (opcode >> 7) & 31;
  1.4771 +		int offset	= shift ? reg[opcode & 15].I >> shift : 0;
  1.4772 +		int dest	= (opcode >> 12) & 15;
  1.4773 +		int base	= (opcode >> 16) & 15;
  1.4774 +		u32 address = reg[base].I;
  1.4775 +		CPUWriteByte(address, reg[dest].B.B0);
  1.4776 +		reg[base].I = address - offset;
  1.4777 +		clockTicks += 2 + CPUUpdateTicksAccess16(address);
  1.4778 +	}
  1.4779 +	break;
  1.4780 +	case 0x644:
  1.4781 +	case 0x64c:
  1.4782 +	// T versions are the same
  1.4783 +	case 0x664:
  1.4784 +	case 0x66c:
  1.4785 +	{
  1.4786 +		// STRB Rd, [Rn], -Rm, ASR #
  1.4787 +		int shift = (opcode >> 7) & 31;
  1.4788 +		int offset;
  1.4789 +		if (shift)
  1.4790 +			offset = (int)((s32)reg[opcode & 15].I >> shift);
  1.4791 +		else if (reg[opcode & 15].I & 0x80000000)
  1.4792 +			offset = 0xFFFFFFFF;
  1.4793 +		else
  1.4794 +			offset = 0;
  1.4795 +		int dest	= (opcode >> 12) & 15;
  1.4796 +		int base	= (opcode >> 16) & 15;
  1.4797 +		u32 address = reg[base].I;
  1.4798 +		CPUWriteByte(address, reg[dest].B.B0);
  1.4799 +		reg[base].I = address - offset;
  1.4800 +		clockTicks += 2 + CPUUpdateTicksAccess16(address);
  1.4801 +	}
  1.4802 +	break;
  1.4803 +	case 0x646:
  1.4804 +	case 0x64e:
  1.4805 +	// T versions are the same
  1.4806 +	case 0x666:
  1.4807 +	case 0x66e:
  1.4808 +	{
  1.4809 +		// STRB Rd, [Rn], -Rm, ROR #
  1.4810 +		int shift = (opcode >> 7) & 31;
  1.4811 +		u32 value = reg[opcode & 15].I;
  1.4812 +		if (shift)
  1.4813 +		{
  1.4814 +			ROR_VALUE;
  1.4815 +		}
  1.4816 +		else
  1.4817 +		{
  1.4818 +			RCR_VALUE;
  1.4819 +		}
  1.4820 +		int dest	= (opcode >> 12) & 15;
  1.4821 +		int base	= (opcode >> 16) & 15;
  1.4822 +		u32 address = reg[base].I;
  1.4823 +		CPUWriteByte(address, reg[dest].B.B0);
  1.4824 +		reg[base].I = address - value;
  1.4825 +		clockTicks += 2 + CPUUpdateTicksAccess16(address);
  1.4826 +	}
  1.4827 +	break;
  1.4828 +	case 0x6c0:
  1.4829 +	case 0x6c8:
  1.4830 +	// T versions are the same
  1.4831 +	case 0x6e0:
  1.4832 +	case 0x6e8:
  1.4833 +	{
  1.4834 +		// STRB Rd, [Rn], Rm, LSL #
  1.4835 +		int offset	= reg[opcode & 15].I << ((opcode >> 7) & 31);
  1.4836 +		int dest	= (opcode >> 12) & 15;
  1.4837 +		int base	= (opcode >> 16) & 15;
  1.4838 +		u32 address = reg[base].I;
  1.4839 +		CPUWriteByte(address, reg[dest].B.B0);
  1.4840 +		reg[base].I = address + offset;
  1.4841 +		clockTicks += 2 + CPUUpdateTicksAccess16(address);
  1.4842 +	}
  1.4843 +	break;
  1.4844 +	case 0x6c2:
  1.4845 +	case 0x6ca:
  1.4846 +	// T versions are the same
  1.4847 +	case 0x6e2:
  1.4848 +	case 0x6ea:
  1.4849 +	{
  1.4850 +		// STRB Rd, [Rn], Rm, LSR #
  1.4851 +		int shift	= (opcode >> 7) & 31;
  1.4852 +		int offset	= shift ? reg[opcode & 15].I >> shift : 0;
  1.4853 +		int dest	= (opcode >> 12) & 15;
  1.4854 +		int base	= (opcode >> 16) & 15;
  1.4855 +		u32 address = reg[base].I;
  1.4856 +		CPUWriteByte(address, reg[dest].B.B0);
  1.4857 +		reg[base].I = address + offset;
  1.4858 +		clockTicks += 2 + CPUUpdateTicksAccess16(address);
  1.4859 +	}
  1.4860 +	break;
  1.4861 +	case 0x6c4:
  1.4862 +	case 0x6cc:
  1.4863 +	// T versions are the same
  1.4864 +	case 0x6e4:
  1.4865 +	case 0x6ec:
  1.4866 +	{
  1.4867 +		// STR Rd, [Rn], Rm, ASR #
  1.4868 +		int shift = (opcode >> 7) & 31;
  1.4869 +		int offset;
  1.4870 +		if (shift)
  1.4871 +			offset = (int)((s32)reg[opcode & 15].I >> shift);
  1.4872 +		else if (reg[opcode & 15].I & 0x80000000)
  1.4873 +			offset = 0xFFFFFFFF;
  1.4874 +		else
  1.4875 +			offset = 0;
  1.4876 +		int dest	= (opcode >> 12) & 15;
  1.4877 +		int base	= (opcode >> 16) & 15;
  1.4878 +		u32 address = reg[base].I;
  1.4879 +		CPUWriteByte(address, reg[dest].B.B0);
  1.4880 +		reg[base].I = address + offset;
  1.4881 +		clockTicks += 2 + CPUUpdateTicksAccess16(address);
  1.4882 +	}
  1.4883 +	break;
  1.4884 +	case 0x6c6:
  1.4885 +	case 0x6ce:
  1.4886 +	// T versions are the same
  1.4887 +	case 0x6e6:
  1.4888 +	case 0x6ee:
  1.4889 +	{
  1.4890 +		// STRB Rd, [Rn], Rm, ROR #
  1.4891 +		int shift = (opcode >> 7) & 31;
  1.4892 +		u32 value = reg[opcode & 15].I;
  1.4893 +		if (shift)
  1.4894 +		{
  1.4895 +			ROR_VALUE;
  1.4896 +		}
  1.4897 +		else
  1.4898 +		{
  1.4899 +			RCR_VALUE;
  1.4900 +		}
  1.4901 +		int dest	= (opcode >> 12) & 15;
  1.4902 +		int base	= (opcode >> 16) & 15;
  1.4903 +		u32 address = reg[base].I;
  1.4904 +		CPUWriteByte(address, reg[dest].B.B0);
  1.4905 +		reg[base].I = address + value;
  1.4906 +		clockTicks += 2 + CPUUpdateTicksAccess16(address);
  1.4907 +	}
  1.4908 +	break;
  1.4909 +	case 0x740:
  1.4910 +	case 0x748:
  1.4911 +	{
  1.4912 +		// STRB Rd, [Rn, -Rm, LSL #]
  1.4913 +		int offset	= reg[opcode & 15].I << ((opcode >> 7) & 31);
  1.4914 +		int dest	= (opcode >> 12) & 15;
  1.4915 +		int base	= (opcode >> 16) & 15;
  1.4916 +		u32 address = reg[base].I - offset;
  1.4917 +		CPUWriteByte(address, reg[dest].B.B0);
  1.4918 +		clockTicks += 2 + CPUUpdateTicksAccess16(address);
  1.4919 +	}
  1.4920 +	break;
  1.4921 +	case 0x742:
  1.4922 +	case 0x74a:
  1.4923 +	{
  1.4924 +		// STRB Rd, [Rn, -Rm, LSR #]
  1.4925 +		int shift	= (opcode >> 7) & 31;
  1.4926 +		int offset	= shift ? reg[opcode & 15].I >> shift : 0;
  1.4927 +		int dest	= (opcode >> 12) & 15;
  1.4928 +		int base	= (opcode >> 16) & 15;
  1.4929 +		u32 address = reg[base].I - offset;
  1.4930 +		CPUWriteByte(address, reg[dest].B.B0);
  1.4931 +		clockTicks += 2 + CPUUpdateTicksAccess16(address);
  1.4932 +	}
  1.4933 +	break;
  1.4934 +	case 0x744:
  1.4935 +	case 0x74c:
  1.4936 +	{
  1.4937 +		// STRB Rd, [Rn, -Rm, ASR #]
  1.4938 +		int shift = (opcode >> 7) & 31;
  1.4939 +		int offset;
  1.4940 +		if (shift)
  1.4941 +			offset = (int)((s32)reg[opcode & 15].I >> shift);
  1.4942 +		else if (reg[opcode & 15].I & 0x80000000)
  1.4943 +			offset = 0xFFFFFFFF;
  1.4944 +		else
  1.4945 +			offset = 0;
  1.4946 +		int dest	= (opcode >> 12) & 15;
  1.4947 +		int base	= (opcode >> 16) & 15;
  1.4948 +		u32 address = reg[base].I - offset;
  1.4949 +		CPUWriteByte(address, reg[dest].B.B0);
  1.4950 +		clockTicks += 2 + CPUUpdateTicksAccess16(address);
  1.4951 +	}
  1.4952 +	break;
  1.4953 +	case 0x746:
  1.4954 +	case 0x74e:
  1.4955 +	{
  1.4956 +		// STRB Rd, [Rn, -Rm, ROR #]
  1.4957 +		int shift = (opcode >> 7) & 31;
  1.4958 +		u32 value = reg[opcode & 15].I;
  1.4959 +		if (shift)
  1.4960 +		{
  1.4961 +			ROR_VALUE;
  1.4962 +		}
  1.4963 +		else
  1.4964 +		{
  1.4965 +			RCR_VALUE;
  1.4966 +		}
  1.4967 +		int dest	= (opcode >> 12) & 15;
  1.4968 +		int base	= (opcode >> 16) & 15;
  1.4969 +		u32 address = reg[base].I - value;
  1.4970 +		CPUWriteByte(address, reg[dest].B.B0);
  1.4971 +		clockTicks += 2 + CPUUpdateTicksAccess16(address);
  1.4972 +	}
  1.4973 +	break;
  1.4974 +	case 0x760:
  1.4975 +	case 0x768:
  1.4976 +	{
  1.4977 +		// STRB Rd, [Rn, -Rm, LSL #]!
  1.4978 +		int offset	= reg[opcode & 15].I << ((opcode >> 7) & 31);
  1.4979 +		int dest	= (opcode >> 12) & 15;
  1.4980 +		int base	= (opcode >> 16) & 15;
  1.4981 +		u32 address = reg[base].I - offset;
  1.4982 +		reg[base].I = address;
  1.4983 +		CPUWriteByte(address, reg[dest].B.B0);
  1.4984 +		clockTicks += 2 + CPUUpdateTicksAccess16(address);
  1.4985 +	}
  1.4986 +	break;
  1.4987 +	case 0x762:
  1.4988 +	case 0x76a:
  1.4989 +	{
  1.4990 +		// STRB Rd, [Rn, -Rm, LSR #]!
  1.4991 +		int shift	= (opcode >> 7) & 31;
  1.4992 +		int offset	= shift ? reg[opcode & 15].I >> shift : 0;
  1.4993 +		int dest	= (opcode >> 12) & 15;
  1.4994 +		int base	= (opcode >> 16) & 15;
  1.4995 +		u32 address = reg[base].I - offset;
  1.4996 +		reg[base].I = address;
  1.4997 +		CPUWriteByte(address, reg[dest].B.B0);
  1.4998 +		clockTicks += 2 + CPUUpdateTicksAccess16(address);
  1.4999 +	}
  1.5000 +	break;
  1.5001 +	case 0x764:
  1.5002 +	case 0x76c:
  1.5003 +	{
  1.5004 +		// STRB Rd, [Rn, -Rm, ASR #]!
  1.5005 +		int shift = (opcode >> 7) & 31;
  1.5006 +		int offset;
  1.5007 +		if (shift)
  1.5008 +			offset = (int)((s32)reg[opcode & 15].I >> shift);
  1.5009 +		else if (reg[opcode & 15].I & 0x80000000)
  1.5010 +			offset = 0xFFFFFFFF;
  1.5011 +		else
  1.5012 +			offset = 0;
  1.5013 +		int dest	= (opcode >> 12) & 15;
  1.5014 +		int base	= (opcode >> 16) & 15;
  1.5015 +		u32 address = reg[base].I - offset;
  1.5016 +		reg[base].I = address;
  1.5017 +		CPUWriteByte(address, reg[dest].B.B0);
  1.5018 +		clockTicks += 2 + CPUUpdateTicksAccess16(address);
  1.5019 +	}
  1.5020 +	break;
  1.5021 +	case 0x766:
  1.5022 +	case 0x76e:
  1.5023 +	{
  1.5024 +		// STRB Rd, [Rn, -Rm, ROR #]!
  1.5025 +		int shift = (opcode >> 7) & 31;
  1.5026 +		u32 value = reg[opcode & 15].I;
  1.5027 +		if (shift)
  1.5028 +		{
  1.5029 +			ROR_VALUE;
  1.5030 +		}
  1.5031 +		else
  1.5032 +		{
  1.5033 +			RCR_VALUE;
  1.5034 +		}
  1.5035 +		int dest	= (opcode >> 12) & 15;
  1.5036 +		int base	= (opcode >> 16) & 15;
  1.5037 +		u32 address = reg[base].I - value;
  1.5038 +		reg[base].I = address;
  1.5039 +		CPUWriteByte(address, reg[dest].B.B0);
  1.5040 +		clockTicks += 2 + CPUUpdateTicksAccess16(address);
  1.5041 +	}
  1.5042 +	break;
  1.5043 +	case 0x7c0:
  1.5044 +	case 0x7c8:
  1.5045 +	{
  1.5046 +		// STRB Rd, [Rn, Rm, LSL #]
  1.5047 +		int offset	= reg[opcode & 15].I << ((opcode >> 7) & 31);
  1.5048 +		int dest	= (opcode >> 12) & 15;
  1.5049 +		int base	= (opcode >> 16) & 15;
  1.5050 +		u32 address = reg[base].I + offset;
  1.5051 +		CPUWriteByte(address, reg[dest].B.B0);
  1.5052 +		clockTicks += 2 + CPUUpdateTicksAccess16(address);
  1.5053 +	}
  1.5054 +	break;
  1.5055 +	case 0x7c2:
  1.5056 +	case 0x7ca:
  1.5057 +	{
  1.5058 +		// STRB Rd, [Rn, Rm, LSR #]
  1.5059 +		int shift	= (opcode >> 7) & 31;
  1.5060 +		int offset	= shift ? reg[opcode & 15].I >> shift : 0;
  1.5061 +		int dest	= (opcode >> 12) & 15;
  1.5062 +		int base	= (opcode >> 16) & 15;
  1.5063 +		u32 address = reg[base].I + offset;
  1.5064 +		CPUWriteByte(address, reg[dest].B.B0);
  1.5065 +		clockTicks += 2 + CPUUpdateTicksAccess16(address);
  1.5066 +	}
  1.5067 +	break;
  1.5068 +	case 0x7c4:
  1.5069 +	case 0x7cc:
  1.5070 +	{
  1.5071 +		// STRB Rd, [Rn, Rm, ASR #]
  1.5072 +		int shift = (opcode >> 7) & 31;
  1.5073 +		int offset;
  1.5074 +		if (shift)
  1.5075 +			offset = (int)((s32)reg[opcode & 15].I >> shift);
  1.5076 +		else if (reg[opcode & 15].I & 0x80000000)
  1.5077 +			offset = 0xFFFFFFFF;
  1.5078 +		else
  1.5079 +			offset = 0;
  1.5080 +		int dest	= (opcode >> 12) & 15;
  1.5081 +		int base	= (opcode >> 16) & 15;
  1.5082 +		u32 address = reg[base].I + offset;
  1.5083 +		CPUWriteByte(address, reg[dest].B.B0);
  1.5084 +		clockTicks += 2 + CPUUpdateTicksAccess16(address);
  1.5085 +	}
  1.5086 +	break;
  1.5087 +	case 0x7c6:
  1.5088 +	case 0x7ce:
  1.5089 +	{
  1.5090 +		// STRB Rd, [Rn, Rm, ROR #]
  1.5091 +		int shift = (opcode >> 7) & 31;
  1.5092 +		u32 value = reg[opcode & 15].I;
  1.5093 +		if (shift)
  1.5094 +		{
  1.5095 +			ROR_VALUE;
  1.5096 +		}
  1.5097 +		else
  1.5098 +		{
  1.5099 +			RCR_VALUE;
  1.5100 +		}
  1.5101 +		int dest	= (opcode >> 12) & 15;
  1.5102 +		int base	= (opcode >> 16) & 15;
  1.5103 +		u32 address = reg[base].I + value;
  1.5104 +		CPUWriteByte(address, reg[dest].B.B0);
  1.5105 +		clockTicks += 2 + CPUUpdateTicksAccess16(address);
  1.5106 +	}
  1.5107 +	break;
  1.5108 +	case 0x7e0:
  1.5109 +	case 0x7e8:
  1.5110 +	{
  1.5111 +		// STRB Rd, [Rn, Rm, LSL #]!
  1.5112 +		int offset	= reg[opcode & 15].I << ((opcode >> 7) & 31);
  1.5113 +		int dest	= (opcode >> 12) & 15;
  1.5114 +		int base	= (opcode >> 16) & 15;
  1.5115 +		u32 address = reg[base].I + offset;
  1.5116 +		reg[base].I = address;
  1.5117 +		CPUWriteByte(address, reg[dest].B.B0);
  1.5118 +		clockTicks += 2 + CPUUpdateTicksAccess16(address);
  1.5119 +	}
  1.5120 +	break;
  1.5121 +	case 0x7e2:
  1.5122 +	case 0x7ea:
  1.5123 +	{
  1.5124 +		// STRB Rd, [Rn, Rm, LSR #]!
  1.5125 +		int shift	= (opcode >> 7) & 31;
  1.5126 +		int offset	= shift ? reg[opcode & 15].I >> shift : 0;
  1.5127 +		int dest	= (opcode >> 12) & 15;
  1.5128 +		int base	= (opcode >> 16) & 15;
  1.5129 +		u32 address = reg[base].I + offset;
  1.5130 +		reg[base].I = address;
  1.5131 +		CPUWriteByte(address, reg[dest].B.B0);
  1.5132 +		clockTicks += 2 + CPUUpdateTicksAccess16(address);
  1.5133 +	}
  1.5134 +	break;
  1.5135 +	case 0x7e4:
  1.5136 +	case 0x7ec:
  1.5137 +	{
  1.5138 +		// STRB Rd, [Rn, Rm, ASR #]!
  1.5139 +		int shift = (opcode >> 7) & 31;
  1.5140 +		int offset;
  1.5141 +		if (shift)
  1.5142 +			offset = (int)((s32)reg[opcode & 15].I >> shift);
  1.5143 +		else if (reg[opcode & 15].I & 0x80000000)
  1.5144 +			offset = 0xFFFFFFFF;
  1.5145 +		else
  1.5146 +			offset = 0;
  1.5147 +		int dest	= (opcode >> 12) & 15;
  1.5148 +		int base	= (opcode >> 16) & 15;
  1.5149 +		u32 address = reg[base].I + offset;
  1.5150 +		reg[base].I = address;
  1.5151 +		CPUWriteByte(address, reg[dest].B.B0);
  1.5152 +		clockTicks += 2 + CPUUpdateTicksAccess16(address);
  1.5153 +	}
  1.5154 +	break;
  1.5155 +	case 0x7e6:
  1.5156 +	case 0x7ee:
  1.5157 +	{
  1.5158 +		// STRB Rd, [Rn, Rm, ROR #]!
  1.5159 +		int shift = (opcode >> 7) & 31;
  1.5160 +		u32 value = reg[opcode & 15].I;
  1.5161 +		if (shift)
  1.5162 +		{
  1.5163 +			ROR_VALUE;
  1.5164 +		}
  1.5165 +		else
  1.5166 +		{
  1.5167 +			RCR_VALUE;
  1.5168 +		}
  1.5169 +		int dest	= (opcode >> 12) & 15;
  1.5170 +		int base	= (opcode >> 16) & 15;
  1.5171 +		u32 address = reg[base].I + value;
  1.5172 +		reg[base].I = address;
  1.5173 +		CPUWriteByte(address, reg[dest].B.B0);
  1.5174 +		clockTicks += 2 + CPUUpdateTicksAccess16(address);
  1.5175 +	}
  1.5176 +	break;
  1.5177 +	case 0x650:
  1.5178 +	case 0x658:
  1.5179 +	// T versions are the same
  1.5180 +	case 0x670:
  1.5181 +	case 0x678:
  1.5182 +	{
  1.5183 +		// LDRB Rd, [Rn], -Rm, LSL #
  1.5184 +		int offset	= reg[opcode & 15].I << ((opcode >> 7) & 31);
  1.5185 +		int dest	= (opcode >> 12) & 15;
  1.5186 +		int base	= (opcode >> 16) & 15;
  1.5187 +		u32 address = reg[base].I;
  1.5188 +		reg[dest].I = CPUReadByte(address);
  1.5189 +		if (dest != base)
  1.5190 +			reg[base].I = address - offset;
  1.5191 +		clockTicks += 3 + CPUUpdateTicksAccess16(address);
  1.5192 +	}
  1.5193 +	break;
  1.5194 +	case 0x652:
  1.5195 +	case 0x65a:
  1.5196 +	// T versions are the same
  1.5197 +	case 0x672:
  1.5198 +	case 0x67a:
  1.5199 +	{
  1.5200 +		// LDRB Rd, [Rn], -Rm, LSR #
  1.5201 +		int shift	= (opcode >> 7) & 31;
  1.5202 +		int offset	= shift ? reg[opcode & 15].I >> shift : 0;
  1.5203 +		int dest	= (opcode >> 12) & 15;
  1.5204 +		int base	= (opcode >> 16) & 15;
  1.5205 +		u32 address = reg[base].I;
  1.5206 +		reg[dest].I = CPUReadByte(address);
  1.5207 +		if (dest != base)
  1.5208 +			reg[base].I = address - offset;
  1.5209 +		clockTicks += 3 + CPUUpdateTicksAccess16(address);
  1.5210 +	}
  1.5211 +	break;
  1.5212 +	case 0x654:
  1.5213 +	case 0x65c:
  1.5214 +	// T versions are the same
  1.5215 +	case 0x674:
  1.5216 +	case 0x67c:
  1.5217 +	{
  1.5218 +		// LDRB Rd, [Rn], -Rm, ASR #
  1.5219 +		int shift = (opcode >> 7) & 31;
  1.5220 +		int offset;
  1.5221 +		if (shift)
  1.5222 +			offset = (int)((s32)reg[opcode & 15].I >> shift);
  1.5223 +		else if (reg[opcode & 15].I & 0x80000000)
  1.5224 +			offset = 0xFFFFFFFF;
  1.5225 +		else
  1.5226 +			offset = 0;
  1.5227 +		int dest	= (opcode >> 12) & 15;
  1.5228 +		int base	= (opcode >> 16) & 15;
  1.5229 +		u32 address = reg[base].I;
  1.5230 +		reg[dest].I = CPUReadByte(address);
  1.5231 +		if (dest != base)
  1.5232 +			reg[base].I = address - offset;
  1.5233 +		clockTicks += 3 + CPUUpdateTicksAccess16(address);
  1.5234 +	}
  1.5235 +	break;
  1.5236 +	case 0x656:
  1.5237 +	case 0x65e:
  1.5238 +	// T versions are the same
  1.5239 +	case 0x676:
  1.5240 +	case 0x67e:
  1.5241 +	{
  1.5242 +		// LDRB Rd, [Rn], -Rm, ROR #
  1.5243 +		int shift = (opcode >> 7) & 31;
  1.5244 +		u32 value = reg[opcode & 15].I;
  1.5245 +		if (shift)
  1.5246 +		{
  1.5247 +			ROR_VALUE;
  1.5248 +		}
  1.5249 +		else
  1.5250 +		{
  1.5251 +			RCR_VALUE;
  1.5252 +		}
  1.5253 +		int dest	= (opcode >> 12) & 15;
  1.5254 +		int base	= (opcode >> 16) & 15;
  1.5255 +		u32 address = reg[base].I;
  1.5256 +		reg[dest].I = CPUReadByte(address);
  1.5257 +		if (dest != base)
  1.5258 +			reg[base].I = address - value;
  1.5259 +		clockTicks += 3 + CPUUpdateTicksAccess16(address);
  1.5260 +	}
  1.5261 +	break;
  1.5262 +	case 0x6d0:
  1.5263 +	case 0x6d8:
  1.5264 +	// T versions are the same
  1.5265 +	case 0x6f0:
  1.5266 +	case 0x6f8:
  1.5267 +	{
  1.5268 +		// LDRB Rd, [Rn], Rm, LSL #
  1.5269 +		int offset	= reg[opcode & 15].I << ((opcode >> 7) & 31);
  1.5270 +		int dest	= (opcode >> 12) & 15;
  1.5271 +		int base	= (opcode >> 16) & 15;
  1.5272 +		u32 address = reg[base].I;
  1.5273 +		reg[dest].I = CPUReadByte(address);
  1.5274 +		if (dest != base)
  1.5275 +			reg[base].I = address + offset;
  1.5276 +		clockTicks += 3 + CPUUpdateTicksAccess16(address);
  1.5277 +	}
  1.5278 +	break;
  1.5279 +	case 0x6d2:
  1.5280 +	case 0x6da:
  1.5281 +	// T versions are the same
  1.5282 +	case 0x6f2:
  1.5283 +	case 0x6fa:
  1.5284 +	{
  1.5285 +		// LDRB Rd, [Rn], Rm, LSR #
  1.5286 +		int shift	= (opcode >> 7) & 31;
  1.5287 +		int offset	= shift ? reg[opcode & 15].I >> shift : 0;
  1.5288 +		int dest	= (opcode >> 12) & 15;
  1.5289 +		int base	= (opcode >> 16) & 15;
  1.5290 +		u32 address = reg[base].I;
  1.5291 +		reg[dest].I = CPUReadByte(address);
  1.5292 +		if (dest != base)
  1.5293 +			reg[base].I = address + offset;
  1.5294 +		clockTicks += 3 + CPUUpdateTicksAccess16(address);
  1.5295 +	}
  1.5296 +	break;
  1.5297 +	case 0x6d4:
  1.5298 +	case 0x6dc:
  1.5299 +	// T versions are the same
  1.5300 +	case 0x6f4:
  1.5301 +	case 0x6fc:
  1.5302 +	{
  1.5303 +		// LDRB Rd, [Rn], Rm, ASR #
  1.5304 +		int shift = (opcode >> 7) & 31;
  1.5305 +		int offset;
  1.5306 +		if (shift)
  1.5307 +			offset = (int)((s32)reg[opcode & 15].I >> shift);
  1.5308 +		else if (reg[opcode & 15].I & 0x80000000)
  1.5309 +			offset = 0xFFFFFFFF;
  1.5310 +		else
  1.5311 +			offset = 0;
  1.5312 +		int dest	= (opcode >> 12) & 15;
  1.5313 +		int base	= (opcode >> 16) & 15;
  1.5314 +		u32 address = reg[base].I;
  1.5315 +		reg[dest].I = CPUReadByte(address);
  1.5316 +		if (dest != base)
  1.5317 +			reg[base].I = address + offset;
  1.5318 +		clockTicks += 3 + CPUUpdateTicksAccess16(address);
  1.5319 +	}
  1.5320 +	break;
  1.5321 +	case 0x6d6:
  1.5322 +	case 0x6de:
  1.5323 +	// T versions are the same
  1.5324 +	case 0x6f6:
  1.5325 +	case 0x6fe:
  1.5326 +	{
  1.5327 +		// LDRB Rd, [Rn], Rm, ROR #
  1.5328 +		int shift = (opcode >> 7) & 31;
  1.5329 +		u32 value = reg[opcode & 15].I;
  1.5330 +		if (shift)
  1.5331 +		{
  1.5332 +			ROR_VALUE;
  1.5333 +		}
  1.5334 +		else
  1.5335 +		{
  1.5336 +			RCR_VALUE;
  1.5337 +		}
  1.5338 +		int dest	= (opcode >> 12) & 15;
  1.5339 +		int base	= (opcode >> 16) & 15;
  1.5340 +		u32 address = reg[base].I;
  1.5341 +		reg[dest].I = CPUReadByte(address);
  1.5342 +		if (dest != base)
  1.5343 +			reg[base].I = address + value;
  1.5344 +		clockTicks += 3 + CPUUpdateTicksAccess16(address);
  1.5345 +	}
  1.5346 +	break;
  1.5347 +	case 0x750:
  1.5348 +	case 0x758:
  1.5349 +	{
  1.5350 +		// LDRB Rd, [Rn, -Rm, LSL #]
  1.5351 +		int offset	= reg[opcode & 15].I << ((opcode >> 7) & 31);
  1.5352 +		int dest	= (opcode >> 12) & 15;
  1.5353 +		int base	= (opcode >> 16) & 15;
  1.5354 +		u32 address = reg[base].I - offset;
  1.5355 +		reg[dest].I = CPUReadByte(address);
  1.5356 +		clockTicks += 3 + CPUUpdateTicksAccess16(address);
  1.5357 +	}
  1.5358 +	break;
  1.5359 +	case 0x752:
  1.5360 +	case 0x75a:
  1.5361 +	{
  1.5362 +		// LDRB Rd, [Rn, -Rm, LSR #]
  1.5363 +		int shift	= (opcode >> 7) & 31;
  1.5364 +		int offset	= shift ? reg[opcode & 15].I >> shift : 0;
  1.5365 +		int dest	= (opcode >> 12) & 15;
  1.5366 +		int base	= (opcode >> 16) & 15;
  1.5367 +		u32 address = reg[base].I - offset;
  1.5368 +		reg[dest].I = CPUReadByte(address);
  1.5369 +		clockTicks += 3 + CPUUpdateTicksAccess16(address);
  1.5370 +	}
  1.5371 +	break;
  1.5372 +	case 0x754:
  1.5373 +	case 0x75c:
  1.5374 +	{
  1.5375 +		// LDRB Rd, [Rn, -Rm, ASR #]
  1.5376 +		int shift = (opcode >> 7) & 31;
  1.5377 +		int offset;
  1.5378 +		if (shift)
  1.5379 +			offset = (int)((s32)reg[opcode & 15].I >> shift);
  1.5380 +		else if (reg[opcode & 15].I & 0x80000000)
  1.5381 +			offset = 0xFFFFFFFF;
  1.5382 +		else
  1.5383 +			offset = 0;
  1.5384 +		int dest	= (opcode >> 12) & 15;
  1.5385 +		int base	= (opcode >> 16) & 15;
  1.5386 +		u32 address = reg[base].I - offset;
  1.5387 +		reg[dest].I = CPUReadByte(address);
  1.5388 +		clockTicks += 3 + CPUUpdateTicksAccess16(address);
  1.5389 +	}
  1.5390 +	break;
  1.5391 +	case 0x756:
  1.5392 +	case 0x75e:
  1.5393 +	{
  1.5394 +		// LDRB Rd, [Rn, -Rm, ROR #]
  1.5395 +		int shift = (opcode >> 7) & 31;
  1.5396 +		u32 value = reg[opcode & 15].I;
  1.5397 +		if (shift)
  1.5398 +		{
  1.5399 +			ROR_VALUE;
  1.5400 +		}
  1.5401 +		else
  1.5402 +		{
  1.5403 +			RCR_VALUE;
  1.5404 +		}
  1.5405 +		int dest	= (opcode >> 12) & 15;
  1.5406 +		int base	= (opcode >> 16) & 15;
  1.5407 +		u32 address = reg[base].I - value;
  1.5408 +		reg[dest].I = CPUReadByte(address);
  1.5409 +		clockTicks += 3 + CPUUpdateTicksAccess16(address);
  1.5410 +	}
  1.5411 +	break;
  1.5412 +	case 0x770:
  1.5413 +	case 0x778:
  1.5414 +	{
  1.5415 +		// LDRB Rd, [Rn, -Rm, LSL #]!
  1.5416 +		int offset	= reg[opcode & 15].I << ((opcode >> 7) & 31);
  1.5417 +		int dest	= (opcode >> 12) & 15;
  1.5418 +		int base	= (opcode >> 16) & 15;
  1.5419 +		u32 address = reg[base].I - offset;
  1.5420 +		reg[dest].I = CPUReadByte(address);
  1.5421 +		if (dest != base)
  1.5422 +			reg[base].I = address;
  1.5423 +		clockTicks += 3 + CPUUpdateTicksAccess16(address);
  1.5424 +	}
  1.5425 +	break;
  1.5426 +	case 0x772:
  1.5427 +	case 0x77a:
  1.5428 +	{
  1.5429 +		// LDRB Rd, [Rn, -Rm, LSR #]!
  1.5430 +		int shift	= (opcode >> 7) & 31;
  1.5431 +		int offset	= shift ? reg[opcode & 15].I >> shift : 0;
  1.5432 +		int dest	= (opcode >> 12) & 15;
  1.5433 +		int base	= (opcode >> 16) & 15;
  1.5434 +		u32 address = reg[base].I - offset;
  1.5435 +		reg[dest].I = CPUReadByte(address);
  1.5436 +		if (dest != base)
  1.5437 +			reg[base].I = address;
  1.5438 +		clockTicks += 3 + CPUUpdateTicksAccess16(address);
  1.5439 +	}
  1.5440 +	break;
  1.5441 +	case 0x774:
  1.5442 +	case 0x77c:
  1.5443 +	{
  1.5444 +		// LDRB Rd, [Rn, -Rm, ASR #]!
  1.5445 +		int shift = (opcode >> 7) & 31;
  1.5446 +		int offset;
  1.5447 +		if (shift)
  1.5448 +			offset = (int)((s32)reg[opcode & 15].I >> shift);
  1.5449 +		else if (reg[opcode & 15].I & 0x80000000)
  1.5450 +			offset = 0xFFFFFFFF;
  1.5451 +		else
  1.5452 +			offset = 0;
  1.5453 +		int dest	= (opcode >> 12) & 15;
  1.5454 +		int base	= (opcode >> 16) & 15;
  1.5455 +		u32 address = reg[base].I - offset;
  1.5456 +		reg[dest].I = CPUReadByte(address);
  1.5457 +		if (dest != base)
  1.5458 +			reg[base].I = address;
  1.5459 +		clockTicks += 3 + CPUUpdateTicksAccess16(address);
  1.5460 +	}
  1.5461 +	break;
  1.5462 +	case 0x776:
  1.5463 +	case 0x77e:
  1.5464 +	{
  1.5465 +		// LDRB Rd, [Rn, -Rm, ROR #]!
  1.5466 +		int shift = (opcode >> 7) & 31;
  1.5467 +		u32 value = reg[opcode & 15].I;
  1.5468 +		if (shift)
  1.5469 +		{
  1.5470 +			ROR_VALUE;
  1.5471 +		}
  1.5472 +		else
  1.5473 +		{
  1.5474 +			RCR_VALUE;
  1.5475 +		}
  1.5476 +		int dest	= (opcode >> 12) & 15;
  1.5477 +		int base	= (opcode >> 16) & 15;
  1.5478 +		u32 address = reg[base].I - value;
  1.5479 +		reg[dest].I = CPUReadByte(address);
  1.5480 +		if (dest != base)
  1.5481 +			reg[base].I = address;
  1.5482 +		clockTicks += 3 + CPUUpdateTicksAccess16(address);
  1.5483 +	}
  1.5484 +	break;
  1.5485 +	case 0x7d0:
  1.5486 +	case 0x7d8:
  1.5487 +	{
  1.5488 +		// LDRB Rd, [Rn, Rm, LSL #]
  1.5489 +		int offset	= reg[opcode & 15].I << ((opcode >> 7) & 31);
  1.5490 +		int dest	= (opcode >> 12) & 15;
  1.5491 +		int base	= (opcode >> 16) & 15;
  1.5492 +		u32 address = reg[base].I + offset;
  1.5493 +		reg[dest].I = CPUReadByte(address);
  1.5494 +		clockTicks += 3 + CPUUpdateTicksAccess16(address);
  1.5495 +	}
  1.5496 +	break;
  1.5497 +	case 0x7d2:
  1.5498 +	case 0x7da:
  1.5499 +	{
  1.5500 +		// LDRB Rd, [Rn, Rm, LSR #]
  1.5501 +		int shift	= (opcode >> 7) & 31;
  1.5502 +		int offset	= shift ? reg[opcode & 15].I >> shift : 0;
  1.5503 +		int dest	= (opcode >> 12) & 15;
  1.5504 +		int base	= (opcode >> 16) & 15;
  1.5505 +		u32 address = reg[base].I + offset;
  1.5506 +		reg[dest].I = CPUReadByte(address);
  1.5507 +		clockTicks += 3 + CPUUpdateTicksAccess16(address);
  1.5508 +	}
  1.5509 +	break;
  1.5510 +	case 0x7d4:
  1.5511 +	case 0x7dc:
  1.5512 +	{
  1.5513 +		// LDRB Rd, [Rn, Rm, ASR #]
  1.5514 +		int shift = (opcode >> 7) & 31;
  1.5515 +		int offset;
  1.5516 +		if (shift)
  1.5517 +			offset = (int)((s32)reg[opcode & 15].I >> shift);
  1.5518 +		else if (reg[opcode & 15].I & 0x80000000)
  1.5519 +			offset = 0xFFFFFFFF;
  1.5520 +		else
  1.5521 +			offset = 0;
  1.5522 +		int dest	= (opcode >> 12) & 15;
  1.5523 +		int base	= (opcode >> 16) & 15;
  1.5524 +		u32 address = reg[base].I + offset;
  1.5525 +		reg[dest].I = CPUReadByte(address);
  1.5526 +		clockTicks += 3 + CPUUpdateTicksAccess16(address);
  1.5527 +	}
  1.5528 +	break;
  1.5529 +	case 0x7d6:
  1.5530 +	case 0x7de:
  1.5531 +	{
  1.5532 +		// LDRB Rd, [Rn, Rm, ROR #]
  1.5533 +		int shift = (opcode >> 7) & 31;
  1.5534 +		u32 value = reg[opcode & 15].I;
  1.5535 +		if (shift)
  1.5536 +		{
  1.5537 +			ROR_VALUE;
  1.5538 +		}
  1.5539 +		else
  1.5540 +		{
  1.5541 +			RCR_VALUE;
  1.5542 +		}
  1.5543 +		int dest	= (opcode >> 12) & 15;
  1.5544 +		int base	= (opcode >> 16) & 15;
  1.5545 +		u32 address = reg[base].I + value;
  1.5546 +		reg[dest].I = CPUReadByte(address);
  1.5547 +		clockTicks += 3 + CPUUpdateTicksAccess16(address);
  1.5548 +	}
  1.5549 +	break;
  1.5550 +	case 0x7f0:
  1.5551 +	case 0x7f8:
  1.5552 +	{
  1.5553 +		// LDRB Rd, [Rn, Rm, LSL #]!
  1.5554 +		int offset	= reg[opcode & 15].I << ((opcode >> 7) & 31);
  1.5555 +		int dest	= (opcode >> 12) & 15;
  1.5556 +		int base	= (opcode >> 16) & 15;
  1.5557 +		u32 address = reg[base].I + offset;
  1.5558 +		reg[dest].I = CPUReadByte(address);
  1.5559 +		if (dest != base)
  1.5560 +			reg[base].I = address;
  1.5561 +		clockTicks += 3 + CPUUpdateTicksAccess16(address);
  1.5562 +	}
  1.5563 +	break;
  1.5564 +	case 0x7f2:
  1.5565 +	case 0x7fa:
  1.5566 +	{
  1.5567 +		// LDRB Rd, [Rn, Rm, LSR #]!
  1.5568 +		int shift	= (opcode >> 7) & 31;
  1.5569 +		int offset	= shift ? reg[opcode & 15].I >> shift : 0;
  1.5570 +		int dest	= (opcode >> 12) & 15;
  1.5571 +		int base	= (opcode >> 16) & 15;
  1.5572 +		u32 address = reg[base].I + offset;
  1.5573 +		reg[dest].I = CPUReadByte(address);
  1.5574 +		if (dest != base)
  1.5575 +			reg[base].I = address;
  1.5576 +		clockTicks += 3 + CPUUpdateTicksAccess16(address);
  1.5577 +	}
  1.5578 +	break;
  1.5579 +	case 0x7f4:
  1.5580 +	case 0x7fc:
  1.5581 +	{
  1.5582 +		// LDRB Rd, [Rn, Rm, ASR #]!
  1.5583 +		int shift = (opcode >> 7) & 31;
  1.5584 +		int offset;
  1.5585 +		if (shift)
  1.5586 +			offset = (int)((s32)reg[opcode & 15].I >> shift);
  1.5587 +		else if (reg[opcode & 15].I & 0x80000000)
  1.5588 +			offset = 0xFFFFFFFF;
  1.5589 +		else
  1.5590 +			offset = 0;
  1.5591 +		int dest	= (opcode >> 12) & 15;
  1.5592 +		int base	= (opcode >> 16) & 15;
  1.5593 +		u32 address = reg[base].I + offset;
  1.5594 +		reg[dest].I = CPUReadByte(address);
  1.5595 +		if (dest != base)
  1.5596 +			reg[base].I = address;
  1.5597 +		clockTicks += 3 + CPUUpdateTicksAccess16(address);
  1.5598 +	}
  1.5599 +	break;
  1.5600 +	case 0x7f6:
  1.5601 +	case 0x7fe:
  1.5602 +	{
  1.5603 +		// LDRB Rd, [Rn, Rm, ROR #]!
  1.5604 +		int shift = (opcode >> 7) & 31;
  1.5605 +		u32 value = reg[opcode & 15].I;
  1.5606 +		if (shift)
  1.5607 +		{
  1.5608 +			ROR_VALUE;
  1.5609 +		}
  1.5610 +		else
  1.5611 +		{
  1.5612 +			RCR_VALUE;
  1.5613 +		}
  1.5614 +		int dest	= (opcode >> 12) & 15;
  1.5615 +		int base	= (opcode >> 16) & 15;
  1.5616 +		u32 address = reg[base].I + value;
  1.5617 +		reg[dest].I = CPUReadByte(address);
  1.5618 +		if (dest != base)
  1.5619 +			reg[base].I = address;
  1.5620 +		clockTicks += 3 + CPUUpdateTicksAccess16(address);
  1.5621 +	}
  1.5622 +	break;
  1.5623 +#define STMW_REG(val, num) \
  1.5624 +	if (opcode & (val)) { \
  1.5625 +		CPUWriteMemory(address, reg[(num)].I); \
  1.5626 +		if (!offset) { \
  1.5627 +			reg[base].I = temp; \
  1.5628 +			clockTicks += 1 + CPUUpdateTicksAccess32(address); \
  1.5629 +			offset		= 1; \
  1.5630 +		} else { \
  1.5631 +			clockTicks += 1 + CPUUpdateTicksAccessSeq32(address); \
  1.5632 +		} \
  1.5633 +		address += 4; \
  1.5634 +	}
  1.5635 +#define STM_REG(val, num) \
  1.5636 +	if (opcode & (val)) { \
  1.5637 +		CPUWriteMemory(address, reg[(num)].I); \
  1.5638 +		if (!offset) { \
  1.5639 +			clockTicks += 1 + CPUUpdateTicksAccess32(address); \
  1.5640 +			offset		= 1; \
  1.5641 +		} else { \
  1.5642 +			clockTicks += 1 + CPUUpdateTicksAccessSeq32(address); \
  1.5643 +		} \
  1.5644 +		address += 4; \
  1.5645 +	}
  1.5646 +
  1.5647 +		CASE_16(0x800)
  1.5648 +		// STMDA Rn, {Rlist}
  1.5649 +		{
  1.5650 +			int base = (opcode & 0x000F0000) >> 16;
  1.5651 +			u32 temp = reg[base].I -
  1.5652 +			           4 * (cpuBitsSet[opcode & 255] + cpuBitsSet[(opcode >> 8) & 255]);
  1.5653 +			u32 address = (temp + 4) & 0xFFFFFFFC;
  1.5654 +			clockTicks += 2;
  1.5655 +			int offset = 0;
  1.5656 +			STM_REG(1, 0);
  1.5657 +			STM_REG(2, 1);
  1.5658 +			STM_REG(4, 2);
  1.5659 +			STM_REG(8, 3);
  1.5660 +			STM_REG(16, 4);
  1.5661 +			STM_REG(32, 5);
  1.5662 +			STM_REG(64, 6);
  1.5663 +			STM_REG(128, 7);
  1.5664 +			STM_REG(256, 8);
  1.5665 +			STM_REG(512, 9);
  1.5666 +			STM_REG(1024, 10);
  1.5667 +			STM_REG(2048, 11);
  1.5668 +			STM_REG(4096, 12);
  1.5669 +			STM_REG(8192, 13);
  1.5670 +			STM_REG(16384, 14);
  1.5671 +			if (opcode & 32768)
  1.5672 +			{
  1.5673 +				CPUWriteMemory(address, reg[15].I + 4);
  1.5674 +				if (!offset)
  1.5675 +					clockTicks += 1 + CPUUpdateTicksAccess32(address);
  1.5676 +				else
  1.5677 +					clockTicks += 1 + CPUUpdateTicksAccessSeq32(address);
  1.5678 +			}
  1.5679 +		}
  1.5680 +		break;
  1.5681 +		CASE_16(0x820)
  1.5682 +		{
  1.5683 +			// STMDA Rn!, {Rlist}
  1.5684 +			int base = (opcode & 0x000F0000) >> 16;
  1.5685 +			u32 temp = reg[base].I -
  1.5686 +			           4 * (cpuBitsSet[opcode & 255] + cpuBitsSet[(opcode >> 8) & 255]);
  1.5687 +			u32 address = (temp + 4) & 0xFFFFFFFC;
  1.5688 +			clockTicks += 2;
  1.5689 +			int offset = 0;
  1.5690 +
  1.5691 +			STMW_REG(1, 0);
  1.5692 +			STMW_REG(2, 1);
  1.5693 +			STMW_REG(4, 2);
  1.5694 +			STMW_REG(8, 3);
  1.5695 +			STMW_REG(16, 4);
  1.5696 +			STMW_REG(32, 5);
  1.5697 +			STMW_REG(64, 6);
  1.5698 +			STMW_REG(128, 7);
  1.5699 +			STMW_REG(256, 8);
  1.5700 +			STMW_REG(512, 9);
  1.5701 +			STMW_REG(1024, 10);
  1.5702 +			STMW_REG(2048, 11);
  1.5703 +			STMW_REG(4096, 12);
  1.5704 +			STMW_REG(8192, 13);
  1.5705 +			STMW_REG(16384, 14);
  1.5706 +			if (opcode & 32768)
  1.5707 +			{
  1.5708 +				CPUWriteMemory(address, reg[15].I + 4);
  1.5709 +				if (!offset)
  1.5710 +					clockTicks += 1 + CPUUpdateTicksAccess32(address);
  1.5711 +				else
  1.5712 +					clockTicks += 1 + CPUUpdateTicksAccessSeq32(address);
  1.5713 +				reg[base].I = temp;
  1.5714 +			}
  1.5715 +		}
  1.5716 +		break;
  1.5717 +		CASE_16(0x840)
  1.5718 +		{
  1.5719 +			// STMDA Rn, {Rlist}^
  1.5720 +			int base = (opcode & 0x000F0000) >> 16;
  1.5721 +			u32 temp = reg[base].I -
  1.5722 +			           4 * (cpuBitsSet[opcode & 255] + cpuBitsSet[(opcode >> 8) & 255]);
  1.5723 +			u32 address = (temp + 4) & 0xFFFFFFFC;
  1.5724 +			clockTicks += 2;
  1.5725 +			int offset = 0;
  1.5726 +
  1.5727 +			STM_REG(1, 0);
  1.5728 +			STM_REG(2, 1);
  1.5729 +			STM_REG(4, 2);
  1.5730 +			STM_REG(8, 3);
  1.5731 +			STM_REG(16, 4);
  1.5732 +			STM_REG(32, 5);
  1.5733 +			STM_REG(64, 6);
  1.5734 +			STM_REG(128, 7);
  1.5735 +
  1.5736 +			if (armMode == 0x11)
  1.5737 +			{
  1.5738 +				STM_REG(256, R8_FIQ);
  1.5739 +				STM_REG(512, R9_FIQ);
  1.5740 +				STM_REG(1024, R10_FIQ);
  1.5741 +				STM_REG(2048, R11_FIQ);
  1.5742 +				STM_REG(4096, R12_FIQ);
  1.5743 +			}
  1.5744 +			else
  1.5745 +			{
  1.5746 +				STM_REG(256, 8);
  1.5747 +				STM_REG(512, 9);
  1.5748 +				STM_REG(1024, 10);
  1.5749 +				STM_REG(2048, 11);
  1.5750 +				STM_REG(4096, 12);
  1.5751 +			}
  1.5752 +
  1.5753 +			if (armMode != 0x10 && armMode != 0x1f)
  1.5754 +			{
  1.5755 +				STM_REG(8192, R13_USR);
  1.5756 +				STM_REG(16384, R14_USR);
  1.5757 +			}
  1.5758 +			else
  1.5759 +			{
  1.5760 +				STM_REG(8192, 13);
  1.5761 +				STM_REG(16384, 14);
  1.5762 +			}
  1.5763 +
  1.5764 +			if (opcode & 32768)
  1.5765 +			{
  1.5766 +				CPUWriteMemory(address, reg[15].I + 4);
  1.5767 +				if (!offset)
  1.5768 +					clockTicks += 1 + CPUUpdateTicksAccess32(address);
  1.5769 +				else
  1.5770 +					clockTicks += 1 + CPUUpdateTicksAccessSeq32(address);
  1.5771 +			}
  1.5772 +		}
  1.5773 +		break;
  1.5774 +		CASE_16(0x860)
  1.5775 +		{
  1.5776 +			// STMDA Rn!, {Rlist}^
  1.5777 +			int base = (opcode & 0x000F0000) >> 16;
  1.5778 +			u32 temp = reg[base].I -
  1.5779 +			           4 * (cpuBitsSet[opcode & 255] + cpuBitsSet[(opcode >> 8) & 255]);
  1.5780 +			u32 address = (temp + 4) & 0xFFFFFFFC;
  1.5781 +			clockTicks += 2;
  1.5782 +			int offset = 0;
  1.5783 +
  1.5784 +			STMW_REG(1, 0);
  1.5785 +			STMW_REG(2, 1);
  1.5786 +			STMW_REG(4, 2);
  1.5787 +			STMW_REG(8, 3);
  1.5788 +			STMW_REG(16, 4);
  1.5789 +			STMW_REG(32, 5);
  1.5790 +			STMW_REG(64, 6);
  1.5791 +			STMW_REG(128, 7);
  1.5792 +
  1.5793 +			if (armMode == 0x11)
  1.5794 +			{
  1.5795 +				STMW_REG(256, R8_FIQ);
  1.5796 +				STMW_REG(512, R9_FIQ);
  1.5797 +				STMW_REG(1024, R10_FIQ);
  1.5798 +				STMW_REG(2048, R11_FIQ);
  1.5799 +				STMW_REG(4096, R12_FIQ);
  1.5800 +			}
  1.5801 +			else
  1.5802 +			{
  1.5803 +				STMW_REG(256, 8);
  1.5804 +				STMW_REG(512, 9);
  1.5805 +				STMW_REG(1024, 10);
  1.5806 +				STMW_REG(2048, 11);
  1.5807 +				STMW_REG(4096, 12);
  1.5808 +			}
  1.5809 +
  1.5810 +			if (armMode != 0x10 && armMode != 0x1f)
  1.5811 +			{
  1.5812 +				STMW_REG(8192, R13_USR);
  1.5813 +				STMW_REG(16384, R14_USR);
  1.5814 +			}
  1.5815 +			else
  1.5816 +			{
  1.5817 +				STMW_REG(8192, 13);
  1.5818 +				STMW_REG(16384, 14);
  1.5819 +			}
  1.5820 +
  1.5821 +			if (opcode & 32768)
  1.5822 +			{
  1.5823 +				CPUWriteMemory(address, reg[15].I + 4);
  1.5824 +				if (!offset)
  1.5825 +					clockTicks += 1 + CPUUpdateTicksAccess32(address);
  1.5826 +				else
  1.5827 +					clockTicks += 1 + CPUUpdateTicksAccessSeq32(address);
  1.5828 +				reg[base].I = temp;
  1.5829 +			}
  1.5830 +		}
  1.5831 +		break;
  1.5832 +
  1.5833 +		CASE_16(0x880)
  1.5834 +		{
  1.5835 +			// STMIA Rn, {Rlist}
  1.5836 +			int base	= (opcode & 0x000F0000) >> 16;
  1.5837 +			u32 address = reg[base].I & 0xFFFFFFFC;
  1.5838 +			clockTicks += 2;
  1.5839 +			int offset = 0;
  1.5840 +			STM_REG(1, 0);
  1.5841 +			STM_REG(2, 1);
  1.5842 +			STM_REG(4, 2);
  1.5843 +			STM_REG(8, 3);
  1.5844 +			STM_REG(16, 4);
  1.5845 +			STM_REG(32, 5);
  1.5846 +			STM_REG(64, 6);
  1.5847 +			STM_REG(128, 7);
  1.5848 +			STM_REG(256, 8);
  1.5849 +			STM_REG(512, 9);
  1.5850 +			STM_REG(1024, 10);
  1.5851 +			STM_REG(2048, 11);
  1.5852 +			STM_REG(4096, 12);
  1.5853 +			STM_REG(8192, 13);
  1.5854 +			STM_REG(16384, 14);
  1.5855 +			if (opcode & 32768)
  1.5856 +			{
  1.5857 +				CPUWriteMemory(address, reg[15].I + 4);
  1.5858 +				if (!offset)
  1.5859 +					clockTicks += 1 + CPUUpdateTicksAccess32(address);
  1.5860 +				else
  1.5861 +					clockTicks += 1 + CPUUpdateTicksAccessSeq32(address);
  1.5862 +			}
  1.5863 +		}
  1.5864 +		break;
  1.5865 +		CASE_16(0x8a0)
  1.5866 +		{
  1.5867 +			// STMIA Rn!, {Rlist}
  1.5868 +			int base	= (opcode & 0x000F0000) >> 16;
  1.5869 +			u32 address = reg[base].I & 0xFFFFFFFC;
  1.5870 +			clockTicks += 2;
  1.5871 +			int offset = 0;
  1.5872 +			u32 temp   = reg[base].I + 4 * (cpuBitsSet[opcode & 0xFF] +
  1.5873 +			                                cpuBitsSet[(opcode >> 8) & 255]);
  1.5874 +			STMW_REG(1, 0);
  1.5875 +			STMW_REG(2, 1);
  1.5876 +			STMW_REG(4, 2);
  1.5877 +			STMW_REG(8, 3);
  1.5878 +			STMW_REG(16, 4);
  1.5879 +			STMW_REG(32, 5);
  1.5880 +			STMW_REG(64, 6);
  1.5881 +			STMW_REG(128, 7);
  1.5882 +			STMW_REG(256, 8);
  1.5883 +			STMW_REG(512, 9);
  1.5884 +			STMW_REG(1024, 10);
  1.5885 +			STMW_REG(2048, 11);
  1.5886 +			STMW_REG(4096, 12);
  1.5887 +			STMW_REG(8192, 13);
  1.5888 +			STMW_REG(16384, 14);
  1.5889 +			if (opcode & 32768)
  1.5890 +			{
  1.5891 +				CPUWriteMemory(address, reg[15].I + 4);
  1.5892 +				if (!offset)
  1.5893 +				{
  1.5894 +					reg[base].I = temp;
  1.5895 +					clockTicks += 1 + CPUUpdateTicksAccess32(address);
  1.5896 +				}
  1.5897 +				else
  1.5898 +					clockTicks += 1 + CPUUpdateTicksAccessSeq32(address);
  1.5899 +			}
  1.5900 +		}
  1.5901 +		break;
  1.5902 +		CASE_16(0x8c0)
  1.5903 +		{
  1.5904 +			// STMIA Rn, {Rlist}^
  1.5905 +			int base	= (opcode & 0x000F0000) >> 16;
  1.5906 +			u32 address = reg[base].I & 0xFFFFFFFC;
  1.5907 +			clockTicks += 2;
  1.5908 +			int offset = 0;
  1.5909 +			STM_REG(1, 0);
  1.5910 +			STM_REG(2, 1);
  1.5911 +			STM_REG(4, 2);
  1.5912 +			STM_REG(8, 3);
  1.5913 +			STM_REG(16, 4);
  1.5914 +			STM_REG(32, 5);
  1.5915 +			STM_REG(64, 6);
  1.5916 +			STM_REG(128, 7);
  1.5917 +			if (armMode == 0x11)
  1.5918 +			{
  1.5919 +				STM_REG(256, R8_FIQ);
  1.5920 +				STM_REG(512, R9_FIQ);
  1.5921 +				STM_REG(1024, R10_FIQ);
  1.5922 +				STM_REG(2048, R11_FIQ);
  1.5923 +				STM_REG(4096, R12_FIQ);
  1.5924 +			}
  1.5925 +			else
  1.5926 +			{
  1.5927 +				STM_REG(256, 8);
  1.5928 +				STM_REG(512, 9);
  1.5929 +				STM_REG(1024, 10);
  1.5930 +				STM_REG(2048, 11);
  1.5931 +				STM_REG(4096, 12);
  1.5932 +			}
  1.5933 +			if (armMode != 0x10 && armMode != 0x1f)
  1.5934 +			{
  1.5935 +				STM_REG(8192, R13_USR);
  1.5936 +				STM_REG(16384, R14_USR);
  1.5937 +			}
  1.5938 +			else
  1.5939 +			{
  1.5940 +				STM_REG(8192, 13);
  1.5941 +				STM_REG(16384, 14);
  1.5942 +			}
  1.5943 +			if (opcode & 32768)
  1.5944 +			{
  1.5945 +				CPUWriteMemory(address, reg[15].I + 4);
  1.5946 +				if (!offset)
  1.5947 +					clockTicks += 1 + CPUUpdateTicksAccess32(address);
  1.5948 +				else
  1.5949 +					clockTicks += 1 + CPUUpdateTicksAccessSeq32(address);
  1.5950 +			}
  1.5951 +		}
  1.5952 +		break;
  1.5953 +		CASE_16(0x8e0)
  1.5954 +		{
  1.5955 +			// STMIA Rn!, {Rlist}^
  1.5956 +			int base	= (opcode & 0x000F0000) >> 16;
  1.5957 +			u32 address = reg[base].I & 0xFFFFFFFC;
  1.5958 +			clockTicks += 2;
  1.5959 +			int offset = 0;
  1.5960 +			u32 temp   = reg[base].I + 4 * (cpuBitsSet[opcode & 0xFF] +
  1.5961 +			                                cpuBitsSet[(opcode >> 8) & 255]);
  1.5962 +			STMW_REG(1, 0);
  1.5963 +			STMW_REG(2, 1);
  1.5964 +			STMW_REG(4, 2);
  1.5965 +			STMW_REG(8, 3);
  1.5966 +			STMW_REG(16, 4);
  1.5967 +			STMW_REG(32, 5);
  1.5968 +			STMW_REG(64, 6);
  1.5969 +			STMW_REG(128, 7);
  1.5970 +			if (armMode == 0x11)
  1.5971 +			{
  1.5972 +				STMW_REG(256, R8_FIQ);
  1.5973 +				STMW_REG(512, R9_FIQ);
  1.5974 +				STMW_REG(1024, R10_FIQ);
  1.5975 +				STMW_REG(2048, R11_FIQ);
  1.5976 +				STMW_REG(4096, R12_FIQ);
  1.5977 +			}
  1.5978 +			else
  1.5979 +			{
  1.5980 +				STMW_REG(256, 8);
  1.5981 +				STMW_REG(512, 9);
  1.5982 +				STMW_REG(1024, 10);
  1.5983 +				STMW_REG(2048, 11);
  1.5984 +				STMW_REG(4096, 12);
  1.5985 +			}
  1.5986 +			if (armMode != 0x10 && armMode != 0x1f)
  1.5987 +			{
  1.5988 +				STMW_REG(8192, R13_USR);
  1.5989 +				STMW_REG(16384, R14_USR);
  1.5990 +			}
  1.5991 +			else
  1.5992 +			{
  1.5993 +				STMW_REG(8192, 13);
  1.5994 +				STMW_REG(16384, 14);
  1.5995 +			}
  1.5996 +			if (opcode & 32768)
  1.5997 +			{
  1.5998 +				CPUWriteMemory(address, reg[15].I + 4);
  1.5999 +				if (!offset)
  1.6000 +				{
  1.6001 +					reg[base].I = temp;
  1.6002 +					clockTicks += 1 + CPUUpdateTicksAccess32(address);
  1.6003 +				}
  1.6004 +				else
  1.6005 +					clockTicks += 1 + CPUUpdateTicksAccessSeq32(address);
  1.6006 +			}
  1.6007 +		}
  1.6008 +		break;
  1.6009 +
  1.6010 +		CASE_16(0x900)
  1.6011 +		{
  1.6012 +			// STMDB Rn, {Rlist}
  1.6013 +			int base = (opcode & 0x000F0000) >> 16;
  1.6014 +			u32 temp = reg[base].I -
  1.6015 +			           4 * (cpuBitsSet[opcode & 255] + cpuBitsSet[(opcode >> 8) & 255]);
  1.6016 +			u32 address = temp & 0xFFFFFFFC;
  1.6017 +			clockTicks += 2;
  1.6018 +			int offset = 0;
  1.6019 +			STM_REG(1, 0);
  1.6020 +			STM_REG(2, 1);
  1.6021 +			STM_REG(4, 2);
  1.6022 +			STM_REG(8, 3);
  1.6023 +			STM_REG(16, 4);
  1.6024 +			STM_REG(32, 5);
  1.6025 +			STM_REG(64, 6);
  1.6026 +			STM_REG(128, 7);
  1.6027 +			STM_REG(256, 8);
  1.6028 +			STM_REG(512, 9);
  1.6029 +			STM_REG(1024, 10);
  1.6030 +			STM_REG(2048, 11);
  1.6031 +			STM_REG(4096, 12);
  1.6032 +			STM_REG(8192, 13);
  1.6033 +			STM_REG(16384, 14);
  1.6034 +			if (opcode & 32768)
  1.6035 +			{
  1.6036 +				CPUWriteMemory(address, reg[15].I + 4);
  1.6037 +				if (!offset)
  1.6038 +					clockTicks += 1 + CPUUpdateTicksAccess32(address);
  1.6039 +				else
  1.6040 +					clockTicks += 1 + CPUUpdateTicksAccessSeq32(address);
  1.6041 +			}
  1.6042 +		}
  1.6043 +		break;
  1.6044 +		CASE_16(0x920)
  1.6045 +		{
  1.6046 +			// STMDB Rn!, {Rlist}
  1.6047 +			int base = (opcode & 0x000F0000) >> 16;
  1.6048 +			u32 temp = reg[base].I -
  1.6049 +			           4 * (cpuBitsSet[opcode & 255] + cpuBitsSet[(opcode >> 8) & 255]);
  1.6050 +			u32 address = temp & 0xFFFFFFFC;
  1.6051 +			clockTicks += 2;
  1.6052 +			int offset = 0;
  1.6053 +
  1.6054 +			STMW_REG(1, 0);
  1.6055 +			STMW_REG(2, 1);
  1.6056 +			STMW_REG(4, 2);
  1.6057 +			STMW_REG(8, 3);
  1.6058 +			STMW_REG(16, 4);
  1.6059 +			STMW_REG(32, 5);
  1.6060 +			STMW_REG(64, 6);
  1.6061 +			STMW_REG(128, 7);
  1.6062 +			STMW_REG(256, 8);
  1.6063 +			STMW_REG(512, 9);
  1.6064 +			STMW_REG(1024, 10);
  1.6065 +			STMW_REG(2048, 11);
  1.6066 +			STMW_REG(4096, 12);
  1.6067 +			STMW_REG(8192, 13);
  1.6068 +			STMW_REG(16384, 14);
  1.6069 +			if (opcode & 32768)
  1.6070 +			{
  1.6071 +				CPUWriteMemory(address, reg[15].I + 4);
  1.6072 +				if (!offset)
  1.6073 +					clockTicks += 1 + CPUUpdateTicksAccess32(address);
  1.6074 +				else
  1.6075 +					clockTicks += 1 + CPUUpdateTicksAccessSeq32(address);
  1.6076 +				reg[base].I = temp;
  1.6077 +			}
  1.6078 +		}
  1.6079 +		break;
  1.6080 +		CASE_16(0x940)
  1.6081 +		{
  1.6082 +			// STMDB Rn, {Rlist}^
  1.6083 +			int base = (opcode & 0x000F0000) >> 16;
  1.6084 +			u32 temp = reg[base].I -
  1.6085 +			           4 * (cpuBitsSet[opcode & 255] + cpuBitsSet[(opcode >> 8) & 255]);
  1.6086 +			u32 address = temp & 0xFFFFFFFC;
  1.6087 +			clockTicks += 2;
  1.6088 +			int offset = 0;
  1.6089 +
  1.6090 +			STM_REG(1, 0);
  1.6091 +			STM_REG(2, 1);
  1.6092 +			STM_REG(4, 2);
  1.6093 +			STM_REG(8, 3);
  1.6094 +			STM_REG(16, 4);
  1.6095 +			STM_REG(32, 5);
  1.6096 +			STM_REG(64, 6);
  1.6097 +			STM_REG(128, 7);
  1.6098 +
  1.6099 +			if (armMode == 0x11)
  1.6100 +			{
  1.6101 +				STM_REG(256, R8_FIQ);
  1.6102 +				STM_REG(512, R9_FIQ);
  1.6103 +				STM_REG(1024, R10_FIQ);
  1.6104 +				STM_REG(2048, R11_FIQ);
  1.6105 +				STM_REG(4096, R12_FIQ);
  1.6106 +			}
  1.6107 +			else
  1.6108 +			{
  1.6109 +				STM_REG(256, 8);
  1.6110 +				STM_REG(512, 9);
  1.6111 +				STM_REG(1024, 10);
  1.6112 +				STM_REG(2048, 11);
  1.6113 +				STM_REG(4096, 12);
  1.6114 +			}
  1.6115 +
  1.6116 +			if (armMode != 0x10 && armMode != 0x1f)
  1.6117 +			{
  1.6118 +				STM_REG(8192, R13_USR);
  1.6119 +				STM_REG(16384, R14_USR);
  1.6120 +			}
  1.6121 +			else
  1.6122 +			{
  1.6123 +				STM_REG(8192, 13);
  1.6124 +				STM_REG(16384, 14);
  1.6125 +			}
  1.6126 +
  1.6127 +			if (opcode & 32768)
  1.6128 +			{
  1.6129 +				CPUWriteMemory(address, reg[15].I + 4);
  1.6130 +				if (!offset)
  1.6131 +					clockTicks += 1 + CPUUpdateTicksAccess32(address);
  1.6132 +				else
  1.6133 +					clockTicks += 1 + CPUUpdateTicksAccessSeq32(address);
  1.6134 +			}
  1.6135 +		}
  1.6136 +		break;
  1.6137 +		CASE_16(0x960)
  1.6138 +		{
  1.6139 +			// STMDB Rn!, {Rlist}^
  1.6140 +			int base = (opcode & 0x000F0000) >> 16;
  1.6141 +			u32 temp = reg[base].I -
  1.6142 +			           4 * (cpuBitsSet[opcode & 255] + cpuBitsSet[(opcode >> 8) & 255]);
  1.6143 +			u32 address = temp & 0xFFFFFFFC;
  1.6144 +			clockTicks += 2;
  1.6145 +			int offset = 0;
  1.6146 +
  1.6147 +			STMW_REG(1, 0);
  1.6148 +			STMW_REG(2, 1);
  1.6149 +			STMW_REG(4, 2);
  1.6150 +			STMW_REG(8, 3);
  1.6151 +			STMW_REG(16, 4);
  1.6152 +			STMW_REG(32, 5);
  1.6153 +			STMW_REG(64, 6);
  1.6154 +			STMW_REG(128, 7);
  1.6155 +
  1.6156 +			if (armMode == 0x11)
  1.6157 +			{
  1.6158 +				STMW_REG(256, R8_FIQ);
  1.6159 +				STMW_REG(512, R9_FIQ);
  1.6160 +				STMW_REG(1024, R10_FIQ);
  1.6161 +				STMW_REG(2048, R11_FIQ);
  1.6162 +				STMW_REG(4096, R12_FIQ);
  1.6163 +			}
  1.6164 +			else
  1.6165 +			{
  1.6166 +				STMW_REG(256, 8);
  1.6167 +				STMW_REG(512, 9);
  1.6168 +				STMW_REG(1024, 10);
  1.6169 +				STMW_REG(2048, 11);
  1.6170 +				STMW_REG(4096, 12);
  1.6171 +			}
  1.6172 +
  1.6173 +			if (armMode != 0x10 && armMode != 0x1f)
  1.6174 +			{
  1.6175 +				STMW_REG(8192, R13_USR);
  1.6176 +				STMW_REG(16384, R14_USR);
  1.6177 +			}
  1.6178 +			else
  1.6179 +			{
  1.6180 +				STMW_REG(8192, 13);
  1.6181 +				STMW_REG(16384, 14);
  1.6182 +			}
  1.6183 +
  1.6184 +			if (opcode & 32768)
  1.6185 +			{
  1.6186 +				CPUWriteMemory(address, reg[15].I + 4);
  1.6187 +				if (!offset)
  1.6188 +					clockTicks += 1 + CPUUpdateTicksAccess32(address);
  1.6189 +				else
  1.6190 +					clockTicks += 1 + CPUUpdateTicksAccessSeq32(address);
  1.6191 +				reg[base].I = temp;
  1.6192 +			}
  1.6193 +		}
  1.6194 +		break;
  1.6195 +
  1.6196 +		CASE_16(0x980)
  1.6197 +		// STMIB Rn, {Rlist}
  1.6198 +		{
  1.6199 +			int base	= (opcode & 0x000F0000) >> 16;
  1.6200 +			u32 address = (reg[base].I + 4) & 0xFFFFFFFC;
  1.6201 +			clockTicks += 2;
  1.6202 +			int offset = 0;
  1.6203 +			STM_REG(1, 0);
  1.6204 +			STM_REG(2, 1);
  1.6205 +			STM_REG(4, 2);
  1.6206 +			STM_REG(8, 3);
  1.6207 +			STM_REG(16, 4);
  1.6208 +			STM_REG(32, 5);
  1.6209 +			STM_REG(64, 6);
  1.6210 +			STM_REG(128, 7);
  1.6211 +			STM_REG(256, 8);
  1.6212 +			STM_REG(512, 9);
  1.6213 +			STM_REG(1024, 10);
  1.6214 +			STM_REG(2048, 11);
  1.6215 +			STM_REG(4096, 12);
  1.6216 +			STM_REG(8192, 13);
  1.6217 +			STM_REG(16384, 14);
  1.6218 +			if (opcode & 32768)
  1.6219 +			{
  1.6220 +				CPUWriteMemory(address, reg[15].I + 4);
  1.6221 +				if (!offset)
  1.6222 +					clockTicks += 1 + CPUUpdateTicksAccess32(address);
  1.6223 +				else
  1.6224 +					clockTicks += 1 + CPUUpdateTicksAccessSeq32(address);
  1.6225 +			}
  1.6226 +		}
  1.6227 +		break;
  1.6228 +		CASE_16(0x9a0)
  1.6229 +		{
  1.6230 +			// STMIB Rn!, {Rlist}
  1.6231 +			int base	= (opcode & 0x000F0000) >> 16;
  1.6232 +			u32 address = (reg[base].I + 4) & 0xFFFFFFFC;
  1.6233 +			clockTicks += 2;
  1.6234 +			int offset = 0;
  1.6235 +			u32 temp   = reg[base].I + 4 * (cpuBitsSet[opcode & 0xFF] +
  1.6236 +			                                cpuBitsSet[(opcode >> 8) & 255]);
  1.6237 +			STMW_REG(1, 0);
  1.6238 +			STMW_REG(2, 1);
  1.6239 +			STMW_REG(4, 2);
  1.6240 +			STMW_REG(8, 3);
  1.6241 +			STMW_REG(16, 4);
  1.6242 +			STMW_REG(32, 5);
  1.6243 +			STMW_REG(64, 6);
  1.6244 +			STMW_REG(128, 7);
  1.6245 +			STMW_REG(256, 8);
  1.6246 +			STMW_REG(512, 9);
  1.6247 +			STMW_REG(1024, 10);
  1.6248 +			STMW_REG(2048, 11);
  1.6249 +			STMW_REG(4096, 12);
  1.6250 +			STMW_REG(8192, 13);
  1.6251 +			STMW_REG(16384, 14);
  1.6252 +			if (opcode & 32768)
  1.6253 +			{
  1.6254 +				CPUWriteMemory(address, reg[15].I + 4);
  1.6255 +				if (!offset)
  1.6256 +				{
  1.6257 +					reg[base].I = temp;
  1.6258 +					clockTicks += 1 + CPUUpdateTicksAccess32(address);
  1.6259 +				}
  1.6260 +				else
  1.6261 +					clockTicks += 1 + CPUUpdateTicksAccessSeq32(address);
  1.6262 +			}
  1.6263 +		}
  1.6264 +		break;
  1.6265 +		CASE_16(0x9c0)
  1.6266 +		{
  1.6267 +			// STMIB Rn, {Rlist}^
  1.6268 +			int base	= (opcode & 0x000F0000) >> 16;
  1.6269 +			u32 address = (reg[base].I + 4) & 0xFFFFFFFC;
  1.6270 +			clockTicks += 2;
  1.6271 +			int offset = 0;
  1.6272 +			STM_REG(1, 0);
  1.6273 +			STM_REG(2, 1);
  1.6274 +			STM_REG(4, 2);
  1.6275 +			STM_REG(8, 3);
  1.6276 +			STM_REG(16, 4);
  1.6277 +			STM_REG(32, 5);
  1.6278 +			STM_REG(64, 6);
  1.6279 +			STM_REG(128, 7);
  1.6280 +			if (armMode == 0x11)
  1.6281 +			{
  1.6282 +				STM_REG(256, R8_FIQ);
  1.6283 +				STM_REG(512, R9_FIQ);
  1.6284 +				STM_REG(1024, R10_FIQ);
  1.6285 +				STM_REG(2048, R11_FIQ);
  1.6286 +				STM_REG(4096, R12_FIQ);
  1.6287 +			}
  1.6288 +			else
  1.6289 +			{
  1.6290 +				STM_REG(256, 8);
  1.6291 +				STM_REG(512, 9);
  1.6292 +				STM_REG(1024, 10);
  1.6293 +				STM_REG(2048, 11);
  1.6294 +				STM_REG(4096, 12);
  1.6295 +			}
  1.6296 +			if (armMode != 0x10 && armMode != 0x1f)
  1.6297 +			{
  1.6298 +				STM_REG(8192, R13_USR);
  1.6299 +				STM_REG(16384, R14_USR);
  1.6300 +			}
  1.6301 +			else
  1.6302 +			{
  1.6303 +				STM_REG(8192, 13);
  1.6304 +				STM_REG(16384, 14);
  1.6305 +			}
  1.6306 +			if (opcode & 32768)
  1.6307 +			{
  1.6308 +				CPUWriteMemory(address, reg[15].I + 4);
  1.6309 +				if (!offset)
  1.6310 +					clockTicks += 1 + CPUUpdateTicksAccess32(address);
  1.6311 +				else
  1.6312 +					clockTicks += 1 + CPUUpdateTicksAccessSeq32(address);
  1.6313 +			}
  1.6314 +		}
  1.6315 +		break;
  1.6316 +		CASE_16(0x9e0)
  1.6317 +		{
  1.6318 +			// STMIB Rn!, {Rlist}^
  1.6319 +			int base	= (opcode & 0x000F0000) >> 16;
  1.6320 +			u32 address = (reg[base].I + 4) & 0xFFFFFFFC;
  1.6321 +			clockTicks += 2;
  1.6322 +			int offset = 0;
  1.6323 +			u32 temp   = reg[base].I + 4 * (cpuBitsSet[opcode & 0xFF] +
  1.6324 +			                                cpuBitsSet[(opcode >> 8) & 255]);
  1.6325 +			STMW_REG(1, 0);
  1.6326 +			STMW_REG(2, 1);
  1.6327 +			STMW_REG(4, 2);
  1.6328 +			STMW_REG(8, 3);
  1.6329 +			STMW_REG(16, 4);
  1.6330 +			STMW_REG(32, 5);
  1.6331 +			STMW_REG(64, 6);
  1.6332 +			STMW_REG(128, 7);
  1.6333 +			if (armMode == 0x11)
  1.6334 +			{
  1.6335 +				STMW_REG(256, R8_FIQ);
  1.6336 +				STMW_REG(512, R9_FIQ);
  1.6337 +				STMW_REG(1024, R10_FIQ);
  1.6338 +				STMW_REG(2048, R11_FIQ);
  1.6339 +				STMW_REG(4096, R12_FIQ);
  1.6340 +			}
  1.6341 +			else
  1.6342 +			{
  1.6343 +				STMW_REG(256, 8);
  1.6344 +				STMW_REG(512, 9);
  1.6345 +				STMW_REG(1024, 10);
  1.6346 +				STMW_REG(2048, 11);
  1.6347 +				STMW_REG(4096, 12);
  1.6348 +			}
  1.6349 +			if (armMode != 0x10 && armMode != 0x1f)
  1.6350 +			{
  1.6351 +				STMW_REG(8192, R13_USR);
  1.6352 +				STMW_REG(16384, R14_USR);
  1.6353 +			}
  1.6354 +			else
  1.6355 +			{
  1.6356 +				STMW_REG(8192, 13);
  1.6357 +				STMW_REG(16384, 14);
  1.6358 +			}
  1.6359 +			if (opcode & 32768)
  1.6360 +			{
  1.6361 +				CPUWriteMemory(address, reg[15].I + 4);
  1.6362 +				if (!offset)
  1.6363 +				{
  1.6364 +					reg[base].I = temp;
  1.6365 +					clockTicks += 1 + CPUUpdateTicksAccess32(address);
  1.6366 +				}
  1.6367 +				else
  1.6368 +					clockTicks += 1 + CPUUpdateTicksAccessSeq32(address);
  1.6369 +			}
  1.6370 +		}
  1.6371 +		break;
  1.6372 +
  1.6373 +#define LDM_REG(val, num) \
  1.6374 +	if (opcode & (val)) { \
  1.6375 +		reg[(num)].I = CPUReadMemory(address); \
  1.6376 +		if (offset) \
  1.6377 +			clockTicks += 1 + CPUUpdateTicksAccessSeq32(address); \
  1.6378 +		else { \
  1.6379 +			clockTicks += 1 + CPUUpdateTicksAccess32(address); \
  1.6380 +			offset		= 1; \
  1.6381 +		} \
  1.6382 +		address += 4; \
  1.6383 +	}
  1.6384 +
  1.6385 +		CASE_16(0x810)
  1.6386 +		{
  1.6387 +			// LDMDA Rn, {Rlist}
  1.6388 +			int base = (opcode & 0x000F0000) >> 16;
  1.6389 +			u32 temp = reg[base].I -
  1.6390 +			           4 * (cpuBitsSet[opcode & 255] + cpuBitsSet[(opcode >> 8) & 255]);
  1.6391 +			u32 address = (temp + 4) & 0xFFFFFFFC;
  1.6392 +			clockTicks += 2;
  1.6393 +			int offset = 0;
  1.6394 +			LDM_REG(1, 0);
  1.6395 +			LDM_REG(2, 1);
  1.6396 +			LDM_REG(4, 2);
  1.6397 +			LDM_REG(8, 3);
  1.6398 +			LDM_REG(16, 4);
  1.6399 +			LDM_REG(32, 5);
  1.6400 +			LDM_REG(64, 6);
  1.6401 +			LDM_REG(128, 7);
  1.6402 +			LDM_REG(256, 8);
  1.6403 +			LDM_REG(512, 9);
  1.6404 +			LDM_REG(1024, 10);
  1.6405 +			LDM_REG(2048, 11);
  1.6406 +			LDM_REG(4096, 12);
  1.6407 +			LDM_REG(8192, 13);
  1.6408 +			LDM_REG(16384, 14);
  1.6409 +			if (opcode & 32768)
  1.6410 +			{
  1.6411 +				reg[15].I = CPUReadMemory(address);
  1.6412 +				if (!offset)
  1.6413 +					clockTicks += 2 + CPUUpdateTicksAccess32(address);
  1.6414 +				else
  1.6415 +					clockTicks += 2 + CPUUpdateTicksAccessSeq32(address);
  1.6416 +				armNextPC  = reg[15].I;
  1.6417 +				reg[15].I += 4;
  1.6418 +			}
  1.6419 +		}
  1.6420 +		break;
  1.6421 +		CASE_16(0x830)
  1.6422 +		{
  1.6423 +			// LDMDA Rn!, {Rlist}
  1.6424 +			int base = (opcode & 0x000F0000) >> 16;
  1.6425 +			u32 temp = reg[base].I -
  1.6426 +			           4 * (cpuBitsSet[opcode & 255] + cpuBitsSet[(opcode >> 8) & 255]);
  1.6427 +			u32 address = (temp + 4) & 0xFFFFFFFC;
  1.6428 +			clockTicks += 2;
  1.6429 +			int offset = 0;
  1.6430 +			LDM_REG(1, 0);
  1.6431 +			LDM_REG(2, 1);
  1.6432 +			LDM_REG(4, 2);
  1.6433 +			LDM_REG(8, 3);
  1.6434 +			LDM_REG(16, 4);
  1.6435 +			LDM_REG(32, 5);
  1.6436 +			LDM_REG(64, 6);
  1.6437 +			LDM_REG(128, 7);
  1.6438 +			LDM_REG(256, 8);
  1.6439 +			LDM_REG(512, 9);
  1.6440 +			LDM_REG(1024, 10);
  1.6441 +			LDM_REG(2048, 11);
  1.6442 +			LDM_REG(4096, 12);
  1.6443 +			LDM_REG(8192, 13);
  1.6444 +			LDM_REG(16384, 14);
  1.6445 +			if (opcode & 32768)
  1.6446 +			{
  1.6447 +				reg[15].I = CPUReadMemory(address);
  1.6448 +				if (!offset)
  1.6449 +					clockTicks += 2 + CPUUpdateTicksAccess32(address);
  1.6450 +				else
  1.6451 +					clockTicks += 2 + CPUUpdateTicksAccessSeq32(address);
  1.6452 +				armNextPC  = reg[15].I;
  1.6453 +				reg[15].I += 4;
  1.6454 +			}
  1.6455 +			if (!(opcode & (1 << base)))
  1.6456 +				reg[base].I = temp;
  1.6457 +		}
  1.6458 +		break;
  1.6459 +		CASE_16(0x850)
  1.6460 +		{
  1.6461 +			// LDMDA Rn, {Rlist}^
  1.6462 +			int base = (opcode & 0x000F0000) >> 16;
  1.6463 +			u32 temp = reg[base].I -
  1.6464 +			           4 * (cpuBitsSet[opcode & 255] + cpuBitsSet[(opcode >> 8) & 255]);
  1.6465 +			u32 address = (temp + 4) & 0xFFFFFFFC;
  1.6466 +			clockTicks += 2;
  1.6467 +			int offset = 0;
  1.6468 +			if (opcode & 0x8000)
  1.6469 +			{
  1.6470 +				LDM_REG(1, 0);
  1.6471 +				LDM_REG(2, 1);
  1.6472 +				LDM_REG(4, 2);
  1.6473 +				LDM_REG(8, 3);
  1.6474 +				LDM_REG(16, 4);
  1.6475 +				LDM_REG(32, 5);
  1.6476 +				LDM_REG(64, 6);
  1.6477 +				LDM_REG(128, 7);
  1.6478 +				LDM_REG(256, 8);
  1.6479 +				LDM_REG(512, 9);
  1.6480 +				LDM_REG(1024, 10);
  1.6481 +				LDM_REG(2048, 11);
  1.6482 +				LDM_REG(4096, 12);
  1.6483 +				LDM_REG(8192, 13);
  1.6484 +				LDM_REG(16384, 14);
  1.6485 +
  1.6486 +				reg[15].I = CPUReadMemory(address);
  1.6487 +				if (!offset)
  1.6488 +					clockTicks += 2 + CPUUpdateTicksAccess32(address);
  1.6489 +				else
  1.6490 +					clockTicks += 2 + CPUUpdateTicksAccessSeq32(address);
  1.6491 +
  1.6492 +				CPUSwitchMode(reg[17].I & 0x1f, false);
  1.6493 +				if (armState)
  1.6494 +				{
  1.6495 +					armNextPC = reg[15].I & 0xFFFFFFFC;
  1.6496 +					reg[15].I = armNextPC + 4;
  1.6497 +				}
  1.6498 +				else
  1.6499 +				{
  1.6500 +					armNextPC = reg[15].I & 0xFFFFFFFE;
  1.6501 +					reg[15].I = armNextPC + 2;
  1.6502 +				}
  1.6503 +			}
  1.6504 +			else
  1.6505 +			{
  1.6506 +				LDM_REG(1, 0);
  1.6507 +				LDM_REG(2, 1);
  1.6508 +				LDM_REG(4, 2);
  1.6509 +				LDM_REG(8, 3);
  1.6510 +				LDM_REG(16, 4);
  1.6511 +				LDM_REG(32, 5);
  1.6512 +				LDM_REG(64, 6);
  1.6513 +				LDM_REG(128, 7);
  1.6514 +
  1.6515 +				if (armMode == 0x11)
  1.6516 +				{
  1.6517 +					LDM_REG(256, R8_FIQ);
  1.6518 +					LDM_REG(512, R9_FIQ);
  1.6519 +					LDM_REG(1024, R10_FIQ);
  1.6520 +					LDM_REG(2048, R11_FIQ);
  1.6521 +					LDM_REG(4096, R12_FIQ);
  1.6522 +				}
  1.6523 +				else
  1.6524 +				{
  1.6525 +					LDM_REG(256, 8);
  1.6526 +					LDM_REG(512, 9);
  1.6527 +					LDM_REG(1024, 10);
  1.6528 +					LDM_REG(2048, 11);
  1.6529 +					LDM_REG(4096, 12);
  1.6530 +				}
  1.6531 +
  1.6532 +				if (armMode != 0x10 && armMode != 0x1f)
  1.6533 +				{
  1.6534 +					LDM_REG(8192, R13_USR);
  1.6535 +					LDM_REG(16384, R14_USR);
  1.6536 +				}
  1.6537 +				else
  1.6538 +				{
  1.6539 +					LDM_REG(8192, 13);
  1.6540 +					LDM_REG(16384, 14);
  1.6541 +				}
  1.6542 +			}
  1.6543 +		}
  1.6544 +		break;
  1.6545 +		CASE_16(0x870)
  1.6546 +		{
  1.6547 +			// LDMDA Rn!, {Rlist}^
  1.6548 +			int base = (opcode & 0x000F0000) >> 16;
  1.6549 +			u32 temp = reg[base].I -
  1.6550 +			           4 * (cpuBitsSet[opcode & 255] + cpuBitsSet[(opcode >> 8) & 255]);
  1.6551 +			u32 address = (temp + 4) & 0xFFFFFFFC;
  1.6552 +			clockTicks += 2;
  1.6553 +			int offset = 0;
  1.6554 +			if (opcode & 0x8000)
  1.6555 +			{
  1.6556 +				LDM_REG(1, 0);
  1.6557 +				LDM_REG(2, 1);
  1.6558 +				LDM_REG(4, 2);
  1.6559 +				LDM_REG(8, 3);
  1.6560 +				LDM_REG(16, 4);
  1.6561 +				LDM_REG(32, 5);
  1.6562 +				LDM_REG(64, 6);
  1.6563 +				LDM_REG(128, 7);
  1.6564 +				LDM_REG(256, 8);
  1.6565 +				LDM_REG(512, 9);
  1.6566 +				LDM_REG(1024, 10);
  1.6567 +				LDM_REG(2048, 11);
  1.6568 +				LDM_REG(4096, 12);
  1.6569 +				LDM_REG(8192, 13);
  1.6570 +				LDM_REG(16384, 14);
  1.6571 +
  1.6572 +				reg[15].I = CPUReadMemory(address);
  1.6573 +				if (!offset)
  1.6574 +					clockTicks += 2 + CPUUpdateTicksAccess32(address);
  1.6575 +				else
  1.6576 +					clockTicks += 2 + CPUUpdateTicksAccessSeq32(address);
  1.6577 +
  1.6578 +				if (!(opcode & (1 << base)))
  1.6579 +					reg[base].I = temp;
  1.6580 +
  1.6581 +				CPUSwitchMode(reg[17].I & 0x1f, false);
  1.6582 +				if (armState)
  1.6583 +				{
  1.6584 +					armNextPC = reg[15].I & 0xFFFFFFFC;
  1.6585 +					reg[15].I = armNextPC + 4;
  1.6586 +				}
  1.6587 +				else
  1.6588 +				{
  1.6589 +					armNextPC = reg[15].I & 0xFFFFFFFE;
  1.6590 +					reg[15].I = armNextPC + 2;
  1.6591 +				}
  1.6592 +			}
  1.6593 +			else
  1.6594 +			{
  1.6595 +				LDM_REG(1, 0);
  1.6596 +				LDM_REG(2, 1);
  1.6597 +				LDM_REG(4, 2);
  1.6598 +				LDM_REG(8, 3);
  1.6599 +				LDM_REG(16, 4);
  1.6600 +				LDM_REG(32, 5);
  1.6601 +				LDM_REG(64, 6);
  1.6602 +				LDM_REG(128, 7);
  1.6603 +
  1.6604 +				if (armMode == 0x11)
  1.6605 +				{
  1.6606 +					LDM_REG(256, R8_FIQ);
  1.6607 +					LDM_REG(512, R9_FIQ);
  1.6608 +					LDM_REG(1024, R10_FIQ);
  1.6609 +					LDM_REG(2048, R11_FIQ);
  1.6610 +					LDM_REG(4096, R12_FIQ);
  1.6611 +				}
  1.6612 +				else
  1.6613 +				{
  1.6614 +					LDM_REG(256, 8);
  1.6615 +					LDM_REG(512, 9);
  1.6616 +					LDM_REG(1024, 10);
  1.6617 +					LDM_REG(2048, 11);
  1.6618 +					LDM_REG(4096, 12);
  1.6619 +				}
  1.6620 +
  1.6621 +				if (armMode != 0x10 && armMode != 0x1f)
  1.6622 +				{
  1.6623 +					LDM_REG(8192, R13_USR);
  1.6624 +					LDM_REG(16384, R14_USR);
  1.6625 +				}
  1.6626 +				else
  1.6627 +				{
  1.6628 +					LDM_REG(8192, 13);
  1.6629 +					LDM_REG(16384, 14);
  1.6630 +				}
  1.6631 +
  1.6632 +				if (!(opcode & (1 << base)))
  1.6633 +					reg[base].I = temp;
  1.6634 +			}
  1.6635 +		}
  1.6636 +		break;
  1.6637 +
  1.6638 +		CASE_16(0x890)
  1.6639 +		{
  1.6640 +			// LDMIA Rn, {Rlist}
  1.6641 +			int base	= (opcode & 0x000F0000) >> 16;
  1.6642 +			u32 address = reg[base].I & 0xFFFFFFFC;
  1.6643 +			clockTicks += 2;
  1.6644 +			int offset = 0;
  1.6645 +			LDM_REG(1, 0);
  1.6646 +			LDM_REG(2, 1);
  1.6647 +			LDM_REG(4, 2);
  1.6648 +			LDM_REG(8, 3);
  1.6649 +			LDM_REG(16, 4);
  1.6650 +			LDM_REG(32, 5);
  1.6651 +			LDM_REG(64, 6);
  1.6652 +			LDM_REG(128, 7);
  1.6653 +			LDM_REG(256, 8);
  1.6654 +			LDM_REG(512, 9);
  1.6655 +			LDM_REG(1024, 10);
  1.6656 +			LDM_REG(2048, 11);
  1.6657 +			LDM_REG(4096, 12);
  1.6658 +			LDM_REG(8192, 13);
  1.6659 +			LDM_REG(16384, 14);
  1.6660 +			if (opcode & 32768)
  1.6661 +			{
  1.6662 +				reg[15].I = CPUReadMemory(address);
  1.6663 +				if (!offset)
  1.6664 +					clockTicks += 2 + CPUUpdateTicksAccess32(address);
  1.6665 +				else
  1.6666 +					clockTicks += 2 + CPUUpdateTicksAccessSeq32(address);
  1.6667 +				armNextPC  = reg[15].I;
  1.6668 +				reg[15].I += 4;
  1.6669 +			}
  1.6670 +		}
  1.6671 +		break;
  1.6672 +		CASE_16(0x8b0)
  1.6673 +		{
  1.6674 +			// LDMIA Rn!, {Rlist}
  1.6675 +			int base = (opcode & 0x000F0000) >> 16;
  1.6676 +			u32 temp = reg[base].I +
  1.6677 +			           4 * (cpuBitsSet[opcode & 255] + cpuBitsSet[(opcode >> 8) & 255]);
  1.6678 +			u32 address = reg[base].I & 0xFFFFFFFC;
  1.6679 +			clockTicks += 2;
  1.6680 +			int offset = 0;
  1.6681 +			LDM_REG(1, 0);
  1.6682 +			LDM_REG(2, 1);
  1.6683 +			LDM_REG(4, 2);
  1.6684 +			LDM_REG(8, 3);
  1.6685 +			LDM_REG(16, 4);
  1.6686 +			LDM_REG(32, 5);
  1.6687 +			LDM_REG(64, 6);
  1.6688 +			LDM_REG(128, 7);
  1.6689 +			LDM_REG(256, 8);
  1.6690 +			LDM_REG(512, 9);
  1.6691 +			LDM_REG(1024, 10);
  1.6692 +			LDM_REG(2048, 11);
  1.6693 +			LDM_REG(4096, 12);
  1.6694 +			LDM_REG(8192, 13);
  1.6695 +			LDM_REG(16384, 14);
  1.6696 +			if (opcode & 32768)
  1.6697 +			{
  1.6698 +				reg[15].I = CPUReadMemory(address);
  1.6699 +				if (!offset)
  1.6700 +					clockTicks += 2 + CPUUpdateTicksAccess32(address);
  1.6701 +				else
  1.6702 +					clockTicks += 2 + CPUUpdateTicksAccessSeq32(address);
  1.6703 +				armNextPC  = reg[15].I;
  1.6704 +				reg[15].I += 4;
  1.6705 +			}
  1.6706 +			if (!(opcode & (1 << base)))
  1.6707 +				reg[base].I = temp;
  1.6708 +		}
  1.6709 +		break;
  1.6710 +		CASE_16(0x8d0)
  1.6711 +		{
  1.6712 +			// LDMIA Rn, {Rlist}^
  1.6713 +			int base	= (opcode & 0x000F0000) >> 16;
  1.6714 +			u32 address = reg[base].I & 0xFFFFFFFC;
  1.6715 +			clockTicks += 2;
  1.6716 +			int offset = 0;
  1.6717 +			if (opcode & 0x8000)
  1.6718 +			{
  1.6719 +				LDM_REG(1, 0);
  1.6720 +				LDM_REG(2, 1);
  1.6721 +				LDM_REG(4, 2);
  1.6722 +				LDM_REG(8, 3);
  1.6723 +				LDM_REG(16, 4);
  1.6724 +				LDM_REG(32, 5);
  1.6725 +				LDM_REG(64, 6);
  1.6726 +				LDM_REG(128, 7);
  1.6727 +				LDM_REG(256, 8);
  1.6728 +				LDM_REG(512, 9);
  1.6729 +				LDM_REG(1024, 10);
  1.6730 +				LDM_REG(2048, 11);
  1.6731 +				LDM_REG(4096, 12);
  1.6732 +				LDM_REG(8192, 13);
  1.6733 +				LDM_REG(16384, 14);
  1.6734 +
  1.6735 +				reg[15].I = CPUReadMemory(address);
  1.6736 +				if (!offset)
  1.6737 +					clockTicks += 2 + CPUUpdateTicksAccess32(address);
  1.6738 +				else
  1.6739 +					clockTicks += 2 + CPUUpdateTicksAccessSeq32(address);
  1.6740 +
  1.6741 +				CPUSwitchMode(reg[17].I & 0x1f, false);
  1.6742 +				if (armState)
  1.6743 +				{
  1.6744 +					armNextPC = reg[15].I & 0xFFFFFFFC;
  1.6745 +					reg[15].I = armNextPC + 4;
  1.6746 +				}
  1.6747 +				else
  1.6748 +				{
  1.6749 +					armNextPC = reg[15].I & 0xFFFFFFFE;
  1.6750 +					reg[15].I = armNextPC + 2;
  1.6751 +				}
  1.6752 +			}
  1.6753 +			else
  1.6754 +			{
  1.6755 +				LDM_REG(1, 0);
  1.6756 +				LDM_REG(2, 1);
  1.6757 +				LDM_REG(4, 2);
  1.6758 +				LDM_REG(8, 3);
  1.6759 +				LDM_REG(16, 4);
  1.6760 +				LDM_REG(32, 5);
  1.6761 +				LDM_REG(64, 6);
  1.6762 +				LDM_REG(128, 7);
  1.6763 +
  1.6764 +				if (armMode == 0x11)
  1.6765 +				{
  1.6766 +					LDM_REG(256, R8_FIQ);
  1.6767 +					LDM_REG(512, R9_FIQ);
  1.6768 +					LDM_REG(1024, R10_FIQ);
  1.6769 +					LDM_REG(2048, R11_FIQ);
  1.6770 +					LDM_REG(4096, R12_FIQ);
  1.6771 +				}
  1.6772 +				else
  1.6773 +				{
  1.6774 +					LDM_REG(256, 8);
  1.6775 +					LDM_REG(512, 9);
  1.6776 +					LDM_REG(1024, 10);
  1.6777 +					LDM_REG(2048, 11);
  1.6778 +					LDM_REG(4096, 12);
  1.6779 +				}
  1.6780 +
  1.6781 +				if (armMode != 0x10 && armMode != 0x1f)
  1.6782 +				{
  1.6783 +					LDM_REG(8192, R13_USR);
  1.6784 +					LDM_REG(16384, R14_USR);
  1.6785 +				}
  1.6786 +				else
  1.6787 +				{
  1.6788 +					LDM_REG(8192, 13);
  1.6789 +					LDM_REG(16384, 14);
  1.6790 +				}
  1.6791 +			}
  1.6792 +		}
  1.6793 +		break;
  1.6794 +		CASE_16(0x8f0)
  1.6795 +		{
  1.6796 +			// LDMIA Rn!, {Rlist}^
  1.6797 +			int base = (opcode & 0x000F0000) >> 16;
  1.6798 +			u32 temp = reg[base].I +
  1.6799 +			           4 * (cpuBitsSet[opcode & 255] + cpuBitsSet[(opcode >> 8) & 255]);
  1.6800 +			u32 address = reg[base].I & 0xFFFFFFFC;
  1.6801 +			clockTicks += 2;
  1.6802 +			int offset = 0;
  1.6803 +			if (opcode & 0x8000)
  1.6804 +			{
  1.6805 +				LDM_REG(1, 0);
  1.6806 +				LDM_REG(2, 1);
  1.6807 +				LDM_REG(4, 2);
  1.6808 +				LDM_REG(8, 3);
  1.6809 +				LDM_REG(16, 4);
  1.6810 +				LDM_REG(32, 5);
  1.6811 +				LDM_REG(64, 6);
  1.6812 +				LDM_REG(128, 7);
  1.6813 +				LDM_REG(256, 8);
  1.6814 +				LDM_REG(512, 9);
  1.6815 +				LDM_REG(1024, 10);
  1.6816 +				LDM_REG(2048, 11);
  1.6817 +				LDM_REG(4096, 12);
  1.6818 +				LDM_REG(8192, 13);
  1.6819 +				LDM_REG(16384, 14);
  1.6820 +
  1.6821 +				reg[15].I = CPUReadMemory(address);
  1.6822 +				if (!offset)
  1.6823 +					clockTicks += 2 + CPUUpdateTicksAccess32(address);
  1.6824 +				else
  1.6825 +					clockTicks += 2 + CPUUpdateTicksAccessSeq32(address);
  1.6826 +
  1.6827 +				if (!(opcode & (1 << base)))
  1.6828 +					reg[base].I = temp;
  1.6829 +
  1.6830 +				CPUSwitchMode(reg[17].I & 0x1f, false);
  1.6831 +				if (armState)
  1.6832 +				{
  1.6833 +					armNextPC = reg[15].I & 0xFFFFFFFC;
  1.6834 +					reg[15].I = armNextPC + 4;
  1.6835 +				}
  1.6836 +				else
  1.6837 +				{
  1.6838 +					armNextPC = reg[15].I & 0xFFFFFFFE;
  1.6839 +					reg[15].I = armNextPC + 2;
  1.6840 +				}
  1.6841 +			}
  1.6842 +			else
  1.6843 +			{
  1.6844 +				LDM_REG(1, 0);
  1.6845 +				LDM_REG(2, 1);
  1.6846 +				LDM_REG(4, 2);
  1.6847 +				LDM_REG(8, 3);
  1.6848 +				LDM_REG(16, 4);
  1.6849 +				LDM_REG(32, 5);
  1.6850 +				LDM_REG(64, 6);
  1.6851 +				LDM_REG(128, 7);
  1.6852 +
  1.6853 +				if (armMode == 0x11)
  1.6854 +				{
  1.6855 +					LDM_REG(256, R8_FIQ);
  1.6856 +					LDM_REG(512, R9_FIQ);
  1.6857 +					LDM_REG(1024, R10_FIQ);
  1.6858 +					LDM_REG(2048, R11_FIQ);
  1.6859 +					LDM_REG(4096, R12_FIQ);
  1.6860 +				}
  1.6861 +				else
  1.6862 +				{
  1.6863 +					LDM_REG(256, 8);
  1.6864 +					LDM_REG(512, 9);
  1.6865 +					LDM_REG(1024, 10);
  1.6866 +					LDM_REG(2048, 11);
  1.6867 +					LDM_REG(4096, 12);
  1.6868 +				}
  1.6869 +
  1.6870 +				if (armMode != 0x10 && armMode != 0x1f)
  1.6871 +				{
  1.6872 +					LDM_REG(8192, R13_USR);
  1.6873 +					LDM_REG(16384, R14_USR);
  1.6874 +				}
  1.6875 +				else
  1.6876 +				{
  1.6877 +					LDM_REG(8192, 13);
  1.6878 +					LDM_REG(16384, 14);
  1.6879 +				}
  1.6880 +
  1.6881 +				if (!(opcode & (1 << base)))
  1.6882 +					reg[base].I = temp;
  1.6883 +			}
  1.6884 +		}
  1.6885 +		break;
  1.6886 +
  1.6887 +		CASE_16(0x910)
  1.6888 +		{
  1.6889 +			// LDMDB Rn, {Rlist}
  1.6890 +			int base = (opcode & 0x000F0000) >> 16;
  1.6891 +			u32 temp = reg[base].I -
  1.6892 +			           4 * (cpuBitsSet[opcode & 255] + cpuBitsSet[(opcode >> 8) & 255]);
  1.6893 +			u32 address = temp & 0xFFFFFFFC;
  1.6894 +			clockTicks += 2;
  1.6895 +			int offset = 0;
  1.6896 +			LDM_REG(1, 0);
  1.6897 +			LDM_REG(2, 1);
  1.6898 +			LDM_REG(4, 2);
  1.6899 +			LDM_REG(8, 3);
  1.6900 +			LDM_REG(16, 4);
  1.6901 +			LDM_REG(32, 5);
  1.6902 +			LDM_REG(64, 6);
  1.6903 +			LDM_REG(128, 7);
  1.6904 +			LDM_REG(256, 8);
  1.6905 +			LDM_REG(512, 9);
  1.6906 +			LDM_REG(1024, 10);
  1.6907 +			LDM_REG(2048, 11);
  1.6908 +			LDM_REG(4096, 12);
  1.6909 +			LDM_REG(8192, 13);
  1.6910 +			LDM_REG(16384, 14);
  1.6911 +			if (opcode & 32768)
  1.6912 +			{
  1.6913 +				reg[15].I = CPUReadMemory(address);
  1.6914 +				if (!offset)
  1.6915 +					clockTicks += 2 + CPUUpdateTicksAccess32(address);
  1.6916 +				else
  1.6917 +					clockTicks += 2 + CPUUpdateTicksAccessSeq32(address);
  1.6918 +				armNextPC  = reg[15].I;
  1.6919 +				reg[15].I += 4;
  1.6920 +			}
  1.6921 +		}
  1.6922 +		break;
  1.6923 +		CASE_16(0x930)
  1.6924 +		{
  1.6925 +			// LDMDB Rn!, {Rlist}
  1.6926 +			int base = (opcode & 0x000F0000) >> 16;
  1.6927 +			u32 temp = reg[base].I -
  1.6928 +			           4 * (cpuBitsSet[opcode & 255] + cpuBitsSet[(opcode >> 8) & 255]);
  1.6929 +			u32 address = temp & 0xFFFFFFFC;
  1.6930 +			clockTicks += 2;
  1.6931 +			int offset = 0;
  1.6932 +			LDM_REG(1, 0);
  1.6933 +			LDM_REG(2, 1);
  1.6934 +			LDM_REG(4, 2);
  1.6935 +			LDM_REG(8, 3);
  1.6936 +			LDM_REG(16, 4);
  1.6937 +			LDM_REG(32, 5);
  1.6938 +			LDM_REG(64, 6);
  1.6939 +			LDM_REG(128, 7);
  1.6940 +			LDM_REG(256, 8);
  1.6941 +			LDM_REG(512, 9);
  1.6942 +			LDM_REG(1024, 10);
  1.6943 +			LDM_REG(2048, 11);
  1.6944 +			LDM_REG(4096, 12);
  1.6945 +			LDM_REG(8192, 13);
  1.6946 +			LDM_REG(16384, 14);
  1.6947 +			if (opcode & 32768)
  1.6948 +			{
  1.6949 +				reg[15].I = CPUReadMemory(address);
  1.6950 +				if (!offset)
  1.6951 +					clockTicks += 2 + CPUUpdateTicksAccess32(address);
  1.6952 +				else
  1.6953 +					clockTicks += 2 + CPUUpdateTicksAccessSeq32(address);
  1.6954 +				armNextPC  = reg[15].I;
  1.6955 +				reg[15].I += 4;
  1.6956 +			}
  1.6957 +			if (!(opcode & (1 << base)))
  1.6958 +				reg[base].I = temp;
  1.6959 +		}
  1.6960 +		break;
  1.6961 +		CASE_16(0x950)
  1.6962 +		{
  1.6963 +			// LDMDB Rn, {Rlist}^
  1.6964 +			int base = (opcode & 0x000F0000) >> 16;
  1.6965 +			u32 temp = reg[base].I -
  1.6966 +			           4 * (cpuBitsSet[opcode & 255] + cpuBitsSet[(opcode >> 8) & 255]);
  1.6967 +			u32 address = temp & 0xFFFFFFFC;
  1.6968 +			clockTicks += 2;
  1.6969 +			int offset = 0;
  1.6970 +			if (opcode & 0x8000)
  1.6971 +			{
  1.6972 +				LDM_REG(1, 0);
  1.6973 +				LDM_REG(2, 1);
  1.6974 +				LDM_REG(4, 2);
  1.6975 +				LDM_REG(8, 3);
  1.6976 +				LDM_REG(16, 4);
  1.6977 +				LDM_REG(32, 5);
  1.6978 +				LDM_REG(64, 6);
  1.6979 +				LDM_REG(128, 7);
  1.6980 +				LDM_REG(256, 8);
  1.6981 +				LDM_REG(512, 9);
  1.6982 +				LDM_REG(1024, 10);
  1.6983 +				LDM_REG(2048, 11);
  1.6984 +				LDM_REG(4096, 12);
  1.6985 +				LDM_REG(8192, 13);
  1.6986 +				LDM_REG(16384, 14);
  1.6987 +
  1.6988 +				reg[15].I = CPUReadMemory(address);
  1.6989 +				if (!offset)
  1.6990 +					clockTicks += 2 + CPUUpdateTicksAccess32(address);
  1.6991 +				else
  1.6992 +					clockTicks += 2 + CPUUpdateTicksAccessSeq32(address);
  1.6993 +
  1.6994 +				CPUSwitchMode(reg[17].I & 0x1f, false);
  1.6995 +				if (armState)
  1.6996 +				{
  1.6997 +					armNextPC = reg[15].I & 0xFFFFFFFC;
  1.6998 +					reg[15].I = armNextPC + 4;
  1.6999 +				}
  1.7000 +				else
  1.7001 +				{
  1.7002 +					armNextPC = reg[15].I & 0xFFFFFFFE;
  1.7003 +					reg[15].I = armNextPC + 2;
  1.7004 +				}
  1.7005 +			}
  1.7006 +			else
  1.7007 +			{
  1.7008 +				LDM_REG(1, 0);
  1.7009 +				LDM_REG(2, 1);
  1.7010 +				LDM_REG(4, 2);
  1.7011 +				LDM_REG(8, 3);
  1.7012 +				LDM_REG(16, 4);
  1.7013 +				LDM_REG(32, 5);
  1.7014 +				LDM_REG(64, 6);
  1.7015 +				LDM_REG(128, 7);
  1.7016 +
  1.7017 +				if (armMode == 0x11)
  1.7018 +				{
  1.7019 +					LDM_REG(256, R8_FIQ);
  1.7020 +					LDM_REG(512, R9_FIQ);
  1.7021 +					LDM_REG(1024, R10_FIQ);
  1.7022 +					LDM_REG(2048, R11_FIQ);
  1.7023 +					LDM_REG(4096, R12_FIQ);
  1.7024 +				}
  1.7025 +				else
  1.7026 +				{
  1.7027 +					LDM_REG(256, 8);
  1.7028 +					LDM_REG(512, 9);
  1.7029 +					LDM_REG(1024, 10);
  1.7030 +					LDM_REG(2048, 11);
  1.7031 +					LDM_REG(4096, 12);
  1.7032 +				}
  1.7033 +
  1.7034 +				if (armMode != 0x10 && armMode != 0x1f)
  1.7035 +				{
  1.7036 +					LDM_REG(8192, R13_USR);
  1.7037 +					LDM_REG(16384, R14_USR);
  1.7038 +				}
  1.7039 +				else
  1.7040 +				{
  1.7041 +					LDM_REG(8192, 13);
  1.7042 +					LDM_REG(16384, 14);
  1.7043 +				}
  1.7044 +			}
  1.7045 +		}
  1.7046 +		break;
  1.7047 +		CASE_16(0x970)
  1.7048 +		{
  1.7049 +			// LDMDB Rn!, {Rlist}^
  1.7050 +			int base = (opcode & 0x000F0000) >> 16;
  1.7051 +			u32 temp = reg[base].I -
  1.7052 +			           4 * (cpuBitsSet[opcode & 255] + cpuBitsSet[(opcode >> 8) & 255]);
  1.7053 +			u32 address = temp & 0xFFFFFFFC;
  1.7054 +			clockTicks += 2;
  1.7055 +			int offset = 0;
  1.7056 +			if (opcode & 0x8000)
  1.7057 +			{
  1.7058 +				LDM_REG(1, 0);
  1.7059 +				LDM_REG(2, 1);
  1.7060 +				LDM_REG(4, 2);
  1.7061 +				LDM_REG(8, 3);
  1.7062 +				LDM_REG(16, 4);
  1.7063 +				LDM_REG(32, 5);
  1.7064 +				LDM_REG(64, 6);
  1.7065 +				LDM_REG(128, 7);
  1.7066 +				LDM_REG(256, 8);
  1.7067 +				LDM_REG(512, 9);
  1.7068 +				LDM_REG(1024, 10);
  1.7069 +				LDM_REG(2048, 11);
  1.7070 +				LDM_REG(4096, 12);
  1.7071 +				LDM_REG(8192, 13);
  1.7072 +				LDM_REG(16384, 14);
  1.7073 +
  1.7074 +				reg[15].I = CPUReadMemory(address);
  1.7075 +				if (!offset)
  1.7076 +					clockTicks += 2 + CPUUpdateTicksAccess32(address);
  1.7077 +				else
  1.7078 +					clockTicks += 2 + CPUUpdateTicksAccessSeq32(address);
  1.7079 +
  1.7080 +				if (!(opcode & (1 << base)))
  1.7081 +					reg[base].I = temp;
  1.7082 +
  1.7083 +				CPUSwitchMode(reg[17].I & 0x1f, false);
  1.7084 +				if (armState)
  1.7085 +				{
  1.7086 +					armNextPC = reg[15].I & 0xFFFFFFFC;
  1.7087 +					reg[15].I = armNextPC + 4;
  1.7088 +				}
  1.7089 +				else
  1.7090 +				{
  1.7091 +					armNextPC = reg[15].I & 0xFFFFFFFE;
  1.7092 +					reg[15].I = armNextPC + 2;
  1.7093 +				}
  1.7094 +			}
  1.7095 +			else
  1.7096 +			{
  1.7097 +				LDM_REG(1, 0);
  1.7098 +				LDM_REG(2, 1);
  1.7099 +				LDM_REG(4, 2);
  1.7100 +				LDM_REG(8, 3);
  1.7101 +				LDM_REG(16, 4);
  1.7102 +				LDM_REG(32, 5);
  1.7103 +				LDM_REG(64, 6);
  1.7104 +				LDM_REG(128, 7);
  1.7105 +
  1.7106 +				if (armMode == 0x11)
  1.7107 +				{
  1.7108 +					LDM_REG(256, R8_FIQ);
  1.7109 +					LDM_REG(512, R9_FIQ);
  1.7110 +					LDM_REG(1024, R10_FIQ);
  1.7111 +					LDM_REG(2048, R11_FIQ);
  1.7112 +					LDM_REG(4096, R12_FIQ);
  1.7113 +				}
  1.7114 +				else
  1.7115 +				{
  1.7116 +					LDM_REG(256, 8);
  1.7117 +					LDM_REG(512, 9);
  1.7118 +					LDM_REG(1024, 10);
  1.7119 +					LDM_REG(2048, 11);
  1.7120 +					LDM_REG(4096, 12);
  1.7121 +				}
  1.7122 +
  1.7123 +				if (armMode != 0x10 && armMode != 0x1f)
  1.7124 +				{
  1.7125 +					LDM_REG(8192, R13_USR);
  1.7126 +					LDM_REG(16384, R14_USR);
  1.7127 +				}
  1.7128 +				else
  1.7129 +				{
  1.7130 +					LDM_REG(8192, 13);
  1.7131 +					LDM_REG(16384, 14);
  1.7132 +				}
  1.7133 +
  1.7134 +				if (!(opcode & (1 << base)))
  1.7135 +					reg[base].I = temp;
  1.7136 +			}
  1.7137 +		}
  1.7138 +		break;
  1.7139 +
  1.7140 +		CASE_16(0x990)
  1.7141 +		{
  1.7142 +			// LDMIB Rn, {Rlist}
  1.7143 +			int base	= (opcode & 0x000F0000) >> 16;
  1.7144 +			u32 address = (reg[base].I + 4) & 0xFFFFFFFC;
  1.7145 +			clockTicks += 2;
  1.7146 +			int offset = 0;
  1.7147 +			LDM_REG(1, 0);
  1.7148 +			LDM_REG(2, 1);
  1.7149 +			LDM_REG(4, 2);
  1.7150 +			LDM_REG(8, 3);
  1.7151 +			LDM_REG(16, 4);
  1.7152 +			LDM_REG(32, 5);
  1.7153 +			LDM_REG(64, 6);
  1.7154 +			LDM_REG(128, 7);
  1.7155 +			LDM_REG(256, 8);
  1.7156 +			LDM_REG(512, 9);
  1.7157 +			LDM_REG(1024, 10);
  1.7158 +			LDM_REG(2048, 11);
  1.7159 +			LDM_REG(4096, 12);
  1.7160 +			LDM_REG(8192, 13);
  1.7161 +			LDM_REG(16384, 14);
  1.7162 +			if (opcode & 32768)
  1.7163 +			{
  1.7164 +				reg[15].I = CPUReadMemory(address);
  1.7165 +				if (!offset)
  1.7166 +					clockTicks += 2 + CPUUpdateTicksAccess32(address);
  1.7167 +				else
  1.7168 +					clockTicks += 2 + CPUUpdateTicksAccessSeq32(address);
  1.7169 +				armNextPC  = reg[15].I;
  1.7170 +				reg[15].I += 4;
  1.7171 +			}
  1.7172 +		}
  1.7173 +		break;
  1.7174 +		CASE_16(0x9b0)
  1.7175 +		{
  1.7176 +			// LDMIB Rn!, {Rlist}
  1.7177 +			int base = (opcode & 0x000F0000) >> 16;
  1.7178 +			u32 temp = reg[base].I +
  1.7179 +			           4 * (cpuBitsSet[opcode & 255] + cpuBitsSet[(opcode >> 8) & 255]);
  1.7180 +			u32 address = (reg[base].I + 4) & 0xFFFFFFFC;
  1.7181 +			clockTicks += 2;
  1.7182 +			int offset = 0;
  1.7183 +			LDM_REG(1, 0);
  1.7184 +			LDM_REG(2, 1);
  1.7185 +			LDM_REG(4, 2);
  1.7186 +			LDM_REG(8, 3);
  1.7187 +			LDM_REG(16, 4);
  1.7188 +			LDM_REG(32, 5);
  1.7189 +			LDM_REG(64, 6);
  1.7190 +			LDM_REG(128, 7);
  1.7191 +			LDM_REG(256, 8);
  1.7192 +			LDM_REG(512, 9);
  1.7193 +			LDM_REG(1024, 10);
  1.7194 +			LDM_REG(2048, 11);
  1.7195 +			LDM_REG(4096, 12);
  1.7196 +			LDM_REG(8192, 13);
  1.7197 +			LDM_REG(16384, 14);
  1.7198 +			if (opcode & 32768)
  1.7199 +			{
  1.7200 +				reg[15].I = CPUReadMemory(address);
  1.7201 +				if (!offset)
  1.7202 +					clockTicks += 2 + CPUUpdateTicksAccess32(address);
  1.7203 +				else
  1.7204 +					clockTicks += 2 + CPUUpdateTicksAccessSeq32(address);
  1.7205 +				armNextPC  = reg[15].I;
  1.7206 +				reg[15].I += 4;
  1.7207 +			}
  1.7208 +			if (!(opcode & (1 << base)))
  1.7209 +				reg[base].I = temp;
  1.7210 +		}
  1.7211 +		break;
  1.7212 +		CASE_16(0x9d0)
  1.7213 +		{
  1.7214 +			// LDMIB Rn, {Rlist}^
  1.7215 +			int base	= (opcode & 0x000F0000) >> 16;
  1.7216 +			u32 address = (reg[base].I + 4) & 0xFFFFFFFC;
  1.7217 +			clockTicks += 2;
  1.7218 +			int offset = 0;
  1.7219 +			if (opcode & 0x8000)
  1.7220 +			{
  1.7221 +				LDM_REG(1, 0);
  1.7222 +				LDM_REG(2, 1);
  1.7223 +				LDM_REG(4, 2);
  1.7224 +				LDM_REG(8, 3);
  1.7225 +				LDM_REG(16, 4);
  1.7226 +				LDM_REG(32, 5);
  1.7227 +				LDM_REG(64, 6);
  1.7228 +				LDM_REG(128, 7);
  1.7229 +				LDM_REG(256, 8);
  1.7230 +				LDM_REG(512, 9);
  1.7231 +				LDM_REG(1024, 10);
  1.7232 +				LDM_REG(2048, 11);
  1.7233 +				LDM_REG(4096, 12);
  1.7234 +				LDM_REG(8192, 13);
  1.7235 +				LDM_REG(16384, 14);
  1.7236 +
  1.7237 +				reg[15].I = CPUReadMemory(address);
  1.7238 +				if (!offset)
  1.7239 +					clockTicks += 2 + CPUUpdateTicksAccess32(address);
  1.7240 +				else
  1.7241 +					clockTicks += 2 + CPUUpdateTicksAccessSeq32(address);
  1.7242 +
  1.7243 +				CPUSwitchMode(reg[17].I & 0x1f, false);
  1.7244 +				if (armState)
  1.7245 +				{
  1.7246 +					armNextPC = reg[15].I & 0xFFFFFFFC;
  1.7247 +					reg[15].I = armNextPC + 4;
  1.7248 +				}
  1.7249 +				else
  1.7250 +				{
  1.7251 +					armNextPC = reg[15].I & 0xFFFFFFFE;
  1.7252 +					reg[15].I = armNextPC + 2;
  1.7253 +				}
  1.7254 +			}
  1.7255 +			else
  1.7256 +			{
  1.7257 +				LDM_REG(1, 0);
  1.7258 +				LDM_REG(2, 1);
  1.7259 +				LDM_REG(4, 2);
  1.7260 +				LDM_REG(8, 3);
  1.7261 +				LDM_REG(16, 4);
  1.7262 +				LDM_REG(32, 5);
  1.7263 +				LDM_REG(64, 6);
  1.7264 +				LDM_REG(128, 7);
  1.7265 +
  1.7266 +				if (armMode == 0x11)
  1.7267 +				{
  1.7268 +					LDM_REG(256, R8_FIQ);
  1.7269 +					LDM_REG(512, R9_FIQ);
  1.7270 +					LDM_REG(1024, R10_FIQ);
  1.7271 +					LDM_REG(2048, R11_FIQ);
  1.7272 +					LDM_REG(4096, R12_FIQ);
  1.7273 +				}
  1.7274 +				else
  1.7275 +				{
  1.7276 +					LDM_REG(256, 8);
  1.7277 +					LDM_REG(512, 9);
  1.7278 +					LDM_REG(1024, 10);
  1.7279 +					LDM_REG(2048, 11);
  1.7280 +					LDM_REG(4096, 12);
  1.7281 +				}
  1.7282 +
  1.7283 +				if (armMode != 0x10 && armMode != 0x1f)
  1.7284 +				{
  1.7285 +					LDM_REG(8192, R13_USR);
  1.7286 +					LDM_REG(16384, R14_USR);
  1.7287 +				}
  1.7288 +				else
  1.7289 +				{
  1.7290 +					LDM_REG(8192, 13);
  1.7291 +					LDM_REG(16384, 14);
  1.7292 +				}
  1.7293 +			}
  1.7294 +		}
  1.7295 +		break;
  1.7296 +		CASE_16(0x9f0)
  1.7297 +		{
  1.7298 +			// LDMIB Rn!, {Rlist}^
  1.7299 +			int base = (opcode & 0x000F0000) >> 16;
  1.7300 +			u32 temp = reg[base].I +
  1.7301 +			           4 * (cpuBitsSet[opcode & 255] + cpuBitsSet[(opcode >> 8) & 255]);
  1.7302 +			u32 address = (reg[base].I + 4) & 0xFFFFFFFC;
  1.7303 +			clockTicks += 2;
  1.7304 +			int offset = 0;
  1.7305 +			if (opcode & 0x8000)
  1.7306 +			{
  1.7307 +				LDM_REG(1, 0);
  1.7308 +				LDM_REG(2, 1);
  1.7309 +				LDM_REG(4, 2);
  1.7310 +				LDM_REG(8, 3);
  1.7311 +				LDM_REG(16, 4);
  1.7312 +				LDM_REG(32, 5);
  1.7313 +				LDM_REG(64, 6);
  1.7314 +				LDM_REG(128, 7);
  1.7315 +				LDM_REG(256, 8);
  1.7316 +				LDM_REG(512, 9);
  1.7317 +				LDM_REG(1024, 10);
  1.7318 +				LDM_REG(2048, 11);
  1.7319 +				LDM_REG(4096, 12);
  1.7320 +				LDM_REG(8192, 13);
  1.7321 +				LDM_REG(16384, 14);
  1.7322 +
  1.7323 +				reg[15].I = CPUReadMemory(address);
  1.7324 +				if (!offset)
  1.7325 +					clockTicks += 2 + CPUUpdateTicksAccess32(address);
  1.7326 +				else
  1.7327 +					clockTicks += 2 + CPUUpdateTicksAccessSeq32(address);
  1.7328 +
  1.7329 +				if (!(opcode & (1 << base)))
  1.7330 +					reg[base].I = temp;
  1.7331 +
  1.7332 +				CPUSwitchMode(reg[17].I & 0x1f, false);
  1.7333 +				if (armState)
  1.7334 +				{
  1.7335 +					armNextPC = reg[15].I & 0xFFFFFFFC;
  1.7336 +					reg[15].I = armNextPC + 4;
  1.7337 +				}
  1.7338 +				else
  1.7339 +				{
  1.7340 +					armNextPC = reg[15].I & 0xFFFFFFFE;
  1.7341 +					reg[15].I = armNextPC + 2;
  1.7342 +				}
  1.7343 +			}
  1.7344 +			else
  1.7345 +			{
  1.7346 +				LDM_REG(1, 0);
  1.7347 +				LDM_REG(2, 1);
  1.7348 +				LDM_REG(4, 2);
  1.7349 +				LDM_REG(8, 3);
  1.7350 +				LDM_REG(16, 4);
  1.7351 +				LDM_REG(32, 5);
  1.7352 +				LDM_REG(64, 6);
  1.7353 +				LDM_REG(128, 7);
  1.7354 +
  1.7355 +				if (armMode == 0x11)
  1.7356 +				{
  1.7357 +					LDM_REG(256, R8_FIQ);
  1.7358 +					LDM_REG(512, R9_FIQ);
  1.7359 +					LDM_REG(1024, R10_FIQ);
  1.7360 +					LDM_REG(2048, R11_FIQ);
  1.7361 +					LDM_REG(4096, R12_FIQ);
  1.7362 +				}
  1.7363 +				else
  1.7364 +				{
  1.7365 +					LDM_REG(256, 8);
  1.7366 +					LDM_REG(512, 9);
  1.7367 +					LDM_REG(1024, 10);
  1.7368 +					LDM_REG(2048, 11);
  1.7369 +					LDM_REG(4096, 12);
  1.7370 +				}
  1.7371 +
  1.7372 +				if (armMode != 0x10 && armMode != 0x1f)
  1.7373 +				{
  1.7374 +					LDM_REG(8192, R13_USR);
  1.7375 +					LDM_REG(16384, R14_USR);
  1.7376 +				}
  1.7377 +				else
  1.7378 +				{
  1.7379 +					LDM_REG(8192, 13);
  1.7380 +					LDM_REG(16384, 14);
  1.7381 +				}
  1.7382 +
  1.7383 +				if (!(opcode & (1 << base)))
  1.7384 +					reg[base].I = temp;
  1.7385 +			}
  1.7386 +		}
  1.7387 +		break;
  1.7388 +		CASE_256(0xa00)
  1.7389 +		{
  1.7390 +			// B <offset>
  1.7391 +			clockTicks += 3;
  1.7392 +			int offset = opcode & 0x00FFFFFF;
  1.7393 +			if (offset & 0x00800000)
  1.7394 +			{
  1.7395 +				offset |= 0xFF000000;
  1.7396 +			}
  1.7397 +			offset	 <<= 2;
  1.7398 +			reg[15].I += offset;
  1.7399 +			armNextPC  = reg[15].I;
  1.7400 +			reg[15].I += 4;
  1.7401 +		}
  1.7402 +		break;
  1.7403 +		CASE_256(0xb00)
  1.7404 +		{
  1.7405 +			// BL <offset>
  1.7406 +			clockTicks += 3;
  1.7407 +			int offset = opcode & 0x00FFFFFF;
  1.7408 +			if (offset & 0x00800000)
  1.7409 +			{
  1.7410 +				offset |= 0xFF000000;
  1.7411 +			}
  1.7412 +			offset	 <<= 2;
  1.7413 +			reg[14].I  = reg[15].I - 4;
  1.7414 +			reg[15].I += offset;
  1.7415 +			armNextPC  = reg[15].I;
  1.7416 +			reg[15].I += 4;
  1.7417 +		}
  1.7418 +		break;
  1.7419 +		CASE_256(0xf00)
  1.7420 +		// SWI <comment>
  1.7421 +		clockTicks += 3;
  1.7422 +		CPUSoftwareInterrupt(opcode & 0x00FFFFFF);
  1.7423 +		break;
  1.7424 +#ifdef GP_SUPPORT
  1.7425 +	case 0xe11:
  1.7426 +	case 0xe13:
  1.7427 +	case 0xe15:
  1.7428 +	case 0xe17:
  1.7429 +	case 0xe19:
  1.7430 +	case 0xe1b:
  1.7431 +	case 0xe1d:
  1.7432 +	case 0xe1f:
  1.7433 +		// MRC
  1.7434 +		break;
  1.7435 +	case 0xe01:
  1.7436 +	case 0xe03:
  1.7437 +	case 0xe05:
  1.7438 +	case 0xe07:
  1.7439 +	case 0xe09:
  1.7440 +	case 0xe0b:
  1.7441 +	case 0xe0d:
  1.7442 +	case 0xe0f:
  1.7443 +		// MRC
  1.7444 +		break;
  1.7445 +#endif
  1.7446 +	default:
  1.7447 +#ifdef GBA_LOGGING
  1.7448 +		if (systemVerbose & VERBOSE_UNDEFINED)
  1.7449 +			log("Undefined ARM instruction %08x at %08x\n", opcode,
  1.7450 +			    armNextPC - 4);
  1.7451 +#endif
  1.7452 +		CPUUndefinedException();
  1.7453 +		break;
  1.7454 +		// END
  1.7455 +	}
  1.7456 +}