root / target-arm / op_neon.h @ 6ddbc6e4
History | View | Annotate | Download (35.5 kB)
| 1 |
/*
|
|---|---|
| 2 |
* ARM NEON vector operations.
|
| 3 |
*
|
| 4 |
* Copyright (c) 2007 CodeSourcery.
|
| 5 |
* Written by Paul Brook
|
| 6 |
*
|
| 7 |
* This code is licenced under the GPL.
|
| 8 |
*/
|
| 9 |
/* Note that for NEON an "l" prefix means it is a wide operation, unlike
|
| 10 |
scalar arm ops where it means a word size operation. */
|
| 11 |
|
| 12 |
#define SIGNBIT (uint32_t)0x80000000 |
| 13 |
/* ??? NEON ops should probably have their own float status. */
|
| 14 |
#define NFS &env->vfp.fp_status
|
| 15 |
#define NEON_OP(name) void OPPROTO op_neon_##name (void) |
| 16 |
|
| 17 |
NEON_OP(getreg_T0) |
| 18 |
{
|
| 19 |
T0 = *(uint32_t *)((char *) env + PARAM1);
|
| 20 |
} |
| 21 |
|
| 22 |
NEON_OP(getreg_T1) |
| 23 |
{
|
| 24 |
T1 = *(uint32_t *)((char *) env + PARAM1);
|
| 25 |
} |
| 26 |
|
| 27 |
NEON_OP(getreg_T2) |
| 28 |
{
|
| 29 |
T2 = *(uint32_t *)((char *) env + PARAM1);
|
| 30 |
} |
| 31 |
|
| 32 |
NEON_OP(setreg_T0) |
| 33 |
{
|
| 34 |
*(uint32_t *)((char *) env + PARAM1) = T0;
|
| 35 |
} |
| 36 |
|
| 37 |
NEON_OP(setreg_T1) |
| 38 |
{
|
| 39 |
*(uint32_t *)((char *) env + PARAM1) = T1;
|
| 40 |
} |
| 41 |
|
| 42 |
NEON_OP(setreg_T2) |
| 43 |
{
|
| 44 |
*(uint32_t *)((char *) env + PARAM1) = T2;
|
| 45 |
} |
| 46 |
|
| 47 |
#define NEON_TYPE1(name, type) \
|
| 48 |
typedef struct \ |
| 49 |
{ \
|
| 50 |
type v1; \ |
| 51 |
} neon_##name; |
| 52 |
#ifdef WORDS_BIGENDIAN
|
| 53 |
#define NEON_TYPE2(name, type) \
|
| 54 |
typedef struct \ |
| 55 |
{ \
|
| 56 |
type v2; \ |
| 57 |
type v1; \ |
| 58 |
} neon_##name; |
| 59 |
#define NEON_TYPE4(name, type) \
|
| 60 |
typedef struct \ |
| 61 |
{ \
|
| 62 |
type v4; \ |
| 63 |
type v3; \ |
| 64 |
type v2; \ |
| 65 |
type v1; \ |
| 66 |
} neon_##name; |
| 67 |
#else
|
| 68 |
#define NEON_TYPE2(name, type) \
|
| 69 |
typedef struct \ |
| 70 |
{ \
|
| 71 |
type v1; \ |
| 72 |
type v2; \ |
| 73 |
} neon_##name; |
| 74 |
#define NEON_TYPE4(name, type) \
|
| 75 |
typedef struct \ |
| 76 |
{ \
|
| 77 |
type v1; \ |
| 78 |
type v2; \ |
| 79 |
type v3; \ |
| 80 |
type v4; \ |
| 81 |
} neon_##name; |
| 82 |
#endif
|
| 83 |
|
| 84 |
NEON_TYPE4(s8, int8_t) |
| 85 |
NEON_TYPE4(u8, uint8_t) |
| 86 |
NEON_TYPE2(s16, int16_t) |
| 87 |
NEON_TYPE2(u16, uint16_t) |
| 88 |
NEON_TYPE1(s32, int32_t) |
| 89 |
NEON_TYPE1(u32, uint32_t) |
| 90 |
#undef NEON_TYPE4
|
| 91 |
#undef NEON_TYPE2
|
| 92 |
#undef NEON_TYPE1
|
| 93 |
|
| 94 |
/* Copy from a uint32_t to a vector structure type. */
|
| 95 |
#define NEON_UNPACK(vtype, dest, val) do { \ |
| 96 |
union { \
|
| 97 |
vtype v; \ |
| 98 |
uint32_t i; \ |
| 99 |
} conv_u; \ |
| 100 |
conv_u.i = (val); \ |
| 101 |
dest = conv_u.v; \ |
| 102 |
} while(0) |
| 103 |
|
| 104 |
/* Copy from a vector structure type to a uint32_t. */
|
| 105 |
#define NEON_PACK(vtype, dest, val) do { \ |
| 106 |
union { \
|
| 107 |
vtype v; \ |
| 108 |
uint32_t i; \ |
| 109 |
} conv_u; \ |
| 110 |
conv_u.v = (val); \ |
| 111 |
dest = conv_u.i; \ |
| 112 |
} while(0) |
| 113 |
|
| 114 |
#define NEON_DO1 \
|
| 115 |
NEON_FN(vdest.v1, vsrc1.v1, vsrc2.v1); |
| 116 |
#define NEON_DO2 \
|
| 117 |
NEON_FN(vdest.v1, vsrc1.v1, vsrc2.v1); \ |
| 118 |
NEON_FN(vdest.v2, vsrc1.v2, vsrc2.v2); |
| 119 |
#define NEON_DO4 \
|
| 120 |
NEON_FN(vdest.v1, vsrc1.v1, vsrc2.v1); \ |
| 121 |
NEON_FN(vdest.v2, vsrc1.v2, vsrc2.v2); \ |
| 122 |
NEON_FN(vdest.v3, vsrc1.v3, vsrc2.v3); \ |
| 123 |
NEON_FN(vdest.v4, vsrc1.v4, vsrc2.v4); |
| 124 |
|
| 125 |
#define NEON_VOP(name, vtype, n) \
|
| 126 |
NEON_OP(name) \ |
| 127 |
{ \
|
| 128 |
vtype vsrc1; \ |
| 129 |
vtype vsrc2; \ |
| 130 |
vtype vdest; \ |
| 131 |
NEON_UNPACK(vtype, vsrc1, T0); \ |
| 132 |
NEON_UNPACK(vtype, vsrc2, T1); \ |
| 133 |
NEON_DO##n; \ |
| 134 |
NEON_PACK(vtype, T0, vdest); \ |
| 135 |
FORCE_RET(); \ |
| 136 |
} |
| 137 |
|
| 138 |
#define NEON_VOP1(name, vtype, n) \
|
| 139 |
NEON_OP(name) \ |
| 140 |
{ \
|
| 141 |
vtype vsrc1; \ |
| 142 |
vtype vdest; \ |
| 143 |
NEON_UNPACK(vtype, vsrc1, T0); \ |
| 144 |
NEON_DO##n; \ |
| 145 |
NEON_PACK(vtype, T0, vdest); \ |
| 146 |
FORCE_RET(); \ |
| 147 |
} |
| 148 |
|
| 149 |
/* Pairwise operations. */
|
| 150 |
/* For 32-bit elements each segment only contains a single element, so
|
| 151 |
the elementwise and pairwise operations are the same. */
|
| 152 |
#define NEON_PDO2 \
|
| 153 |
NEON_FN(vdest.v1, vsrc1.v1, vsrc1.v2); \ |
| 154 |
NEON_FN(vdest.v2, vsrc2.v1, vsrc2.v2); |
| 155 |
#define NEON_PDO4 \
|
| 156 |
NEON_FN(vdest.v1, vsrc1.v1, vsrc1.v2); \ |
| 157 |
NEON_FN(vdest.v2, vsrc1.v3, vsrc1.v4); \ |
| 158 |
NEON_FN(vdest.v3, vsrc2.v1, vsrc2.v2); \ |
| 159 |
NEON_FN(vdest.v4, vsrc2.v3, vsrc2.v4); \ |
| 160 |
|
| 161 |
#define NEON_POP(name, vtype, n) \
|
| 162 |
NEON_OP(name) \ |
| 163 |
{ \
|
| 164 |
vtype vsrc1; \ |
| 165 |
vtype vsrc2; \ |
| 166 |
vtype vdest; \ |
| 167 |
NEON_UNPACK(vtype, vsrc1, T0); \ |
| 168 |
NEON_UNPACK(vtype, vsrc2, T1); \ |
| 169 |
NEON_PDO##n; \ |
| 170 |
NEON_PACK(vtype, T0, vdest); \ |
| 171 |
FORCE_RET(); \ |
| 172 |
} |
| 173 |
|
| 174 |
#define NEON_FN(dest, src1, src2) dest = (src1 + src2) >> 1 |
| 175 |
NEON_VOP(hadd_s8, neon_s8, 4)
|
| 176 |
NEON_VOP(hadd_u8, neon_u8, 4)
|
| 177 |
NEON_VOP(hadd_s16, neon_s16, 2)
|
| 178 |
NEON_VOP(hadd_u16, neon_u16, 2)
|
| 179 |
#undef NEON_FN
|
| 180 |
|
| 181 |
NEON_OP(hadd_s32) |
| 182 |
{
|
| 183 |
int32_t src1 = T0; |
| 184 |
int32_t src2 = T1; |
| 185 |
int32_t dest; |
| 186 |
|
| 187 |
dest = (src1 >> 1) + (src2 >> 1); |
| 188 |
if (src1 & src2 & 1) |
| 189 |
dest++; |
| 190 |
T0 = dest; |
| 191 |
FORCE_RET(); |
| 192 |
} |
| 193 |
|
| 194 |
NEON_OP(hadd_u32) |
| 195 |
{
|
| 196 |
uint32_t src1 = T0; |
| 197 |
uint32_t src2 = T1; |
| 198 |
uint32_t dest; |
| 199 |
|
| 200 |
dest = (src1 >> 1) + (src2 >> 1); |
| 201 |
if (src1 & src2 & 1) |
| 202 |
dest++; |
| 203 |
T0 = dest; |
| 204 |
FORCE_RET(); |
| 205 |
} |
| 206 |
|
| 207 |
#define NEON_FN(dest, src1, src2) dest = (src1 + src2 + 1) >> 1 |
| 208 |
NEON_VOP(rhadd_s8, neon_s8, 4)
|
| 209 |
NEON_VOP(rhadd_u8, neon_u8, 4)
|
| 210 |
NEON_VOP(rhadd_s16, neon_s16, 2)
|
| 211 |
NEON_VOP(rhadd_u16, neon_u16, 2)
|
| 212 |
#undef NEON_FN
|
| 213 |
|
| 214 |
NEON_OP(rhadd_s32) |
| 215 |
{
|
| 216 |
int32_t src1 = T0; |
| 217 |
int32_t src2 = T1; |
| 218 |
int32_t dest; |
| 219 |
|
| 220 |
dest = (src1 >> 1) + (src2 >> 1); |
| 221 |
if ((src1 | src2) & 1) |
| 222 |
dest++; |
| 223 |
T0 = dest; |
| 224 |
FORCE_RET(); |
| 225 |
} |
| 226 |
|
| 227 |
NEON_OP(rhadd_u32) |
| 228 |
{
|
| 229 |
uint32_t src1 = T0; |
| 230 |
uint32_t src2 = T1; |
| 231 |
uint32_t dest; |
| 232 |
|
| 233 |
dest = (src1 >> 1) + (src2 >> 1); |
| 234 |
if ((src1 | src2) & 1) |
| 235 |
dest++; |
| 236 |
T0 = dest; |
| 237 |
FORCE_RET(); |
| 238 |
} |
| 239 |
|
| 240 |
#define NEON_FN(dest, src1, src2) dest = (src1 - src2) >> 1 |
| 241 |
NEON_VOP(hsub_s8, neon_s8, 4)
|
| 242 |
NEON_VOP(hsub_u8, neon_u8, 4)
|
| 243 |
NEON_VOP(hsub_s16, neon_s16, 2)
|
| 244 |
NEON_VOP(hsub_u16, neon_u16, 2)
|
| 245 |
#undef NEON_FN
|
| 246 |
|
| 247 |
NEON_OP(hsub_s32) |
| 248 |
{
|
| 249 |
int32_t src1 = T0; |
| 250 |
int32_t src2 = T1; |
| 251 |
int32_t dest; |
| 252 |
|
| 253 |
dest = (src1 >> 1) - (src2 >> 1); |
| 254 |
if ((~src1) & src2 & 1) |
| 255 |
dest--; |
| 256 |
T0 = dest; |
| 257 |
FORCE_RET(); |
| 258 |
} |
| 259 |
|
| 260 |
NEON_OP(hsub_u32) |
| 261 |
{
|
| 262 |
uint32_t src1 = T0; |
| 263 |
uint32_t src2 = T1; |
| 264 |
uint32_t dest; |
| 265 |
|
| 266 |
dest = (src1 >> 1) - (src2 >> 1); |
| 267 |
if ((~src1) & src2 & 1) |
| 268 |
dest--; |
| 269 |
T0 = dest; |
| 270 |
FORCE_RET(); |
| 271 |
} |
| 272 |
|
| 273 |
/* ??? bsl, bif and bit are all the same op, just with the oparands in a
|
| 274 |
differnet order. It's currently easier to have 3 differnt ops than
|
| 275 |
rearange the operands. */
|
| 276 |
|
| 277 |
/* Bitwise Select. */
|
| 278 |
NEON_OP(bsl) |
| 279 |
{
|
| 280 |
T0 = (T0 & T2) | (T1 & ~T2); |
| 281 |
} |
| 282 |
|
| 283 |
/* Bitwise Insert If True. */
|
| 284 |
NEON_OP(bit) |
| 285 |
{
|
| 286 |
T0 = (T0 & T1) | (T2 & ~T1); |
| 287 |
} |
| 288 |
|
| 289 |
/* Bitwise Insert If False. */
|
| 290 |
NEON_OP(bif) |
| 291 |
{
|
| 292 |
T0 = (T2 & T1) | (T0 & ~T1); |
| 293 |
} |
| 294 |
|
| 295 |
#define NEON_USAT(dest, src1, src2, type) do { \ |
| 296 |
uint32_t tmp = (uint32_t)src1 + (uint32_t)src2; \ |
| 297 |
if (tmp != (type)tmp) { \
|
| 298 |
env->QF = 1; \
|
| 299 |
dest = ~0; \
|
| 300 |
} else { \
|
| 301 |
dest = tmp; \ |
| 302 |
}} while(0) |
| 303 |
#define NEON_FN(dest, src1, src2) NEON_USAT(dest, src1, src2, uint8_t)
|
| 304 |
NEON_VOP(qadd_u8, neon_u8, 4)
|
| 305 |
#undef NEON_FN
|
| 306 |
#define NEON_FN(dest, src1, src2) NEON_USAT(dest, src1, src2, uint16_t)
|
| 307 |
NEON_VOP(qadd_u16, neon_u16, 2)
|
| 308 |
#undef NEON_FN
|
| 309 |
#undef NEON_USAT
|
| 310 |
|
| 311 |
#define NEON_SSAT(dest, src1, src2, type) do { \ |
| 312 |
int32_t tmp = (uint32_t)src1 + (uint32_t)src2; \ |
| 313 |
if (tmp != (type)tmp) { \
|
| 314 |
env->QF = 1; \
|
| 315 |
if (src2 > 0) { \ |
| 316 |
tmp = (1 << (sizeof(type) * 8 - 1)) - 1; \ |
| 317 |
} else { \
|
| 318 |
tmp = 1 << (sizeof(type) * 8 - 1); \ |
| 319 |
} \ |
| 320 |
} \ |
| 321 |
dest = tmp; \ |
| 322 |
} while(0) |
| 323 |
#define NEON_FN(dest, src1, src2) NEON_SSAT(dest, src1, src2, int8_t)
|
| 324 |
NEON_VOP(qadd_s8, neon_s8, 4)
|
| 325 |
#undef NEON_FN
|
| 326 |
#define NEON_FN(dest, src1, src2) NEON_SSAT(dest, src1, src2, int16_t)
|
| 327 |
NEON_VOP(qadd_s16, neon_s16, 2)
|
| 328 |
#undef NEON_FN
|
| 329 |
#undef NEON_SSAT
|
| 330 |
|
| 331 |
#define NEON_USAT(dest, src1, src2, type) do { \ |
| 332 |
uint32_t tmp = (uint32_t)src1 - (uint32_t)src2; \ |
| 333 |
if (tmp != (type)tmp) { \
|
| 334 |
env->QF = 1; \
|
| 335 |
dest = 0; \
|
| 336 |
} else { \
|
| 337 |
dest = tmp; \ |
| 338 |
}} while(0) |
| 339 |
#define NEON_FN(dest, src1, src2) NEON_USAT(dest, src1, src2, uint8_t)
|
| 340 |
NEON_VOP(qsub_u8, neon_u8, 4)
|
| 341 |
#undef NEON_FN
|
| 342 |
#define NEON_FN(dest, src1, src2) NEON_USAT(dest, src1, src2, uint16_t)
|
| 343 |
NEON_VOP(qsub_u16, neon_u16, 2)
|
| 344 |
#undef NEON_FN
|
| 345 |
#undef NEON_USAT
|
| 346 |
|
| 347 |
#define NEON_SSAT(dest, src1, src2, type) do { \ |
| 348 |
int32_t tmp = (uint32_t)src1 - (uint32_t)src2; \ |
| 349 |
if (tmp != (type)tmp) { \
|
| 350 |
env->QF = 1; \
|
| 351 |
if (src2 < 0) { \ |
| 352 |
tmp = (1 << (sizeof(type) * 8 - 1)) - 1; \ |
| 353 |
} else { \
|
| 354 |
tmp = 1 << (sizeof(type) * 8 - 1); \ |
| 355 |
} \ |
| 356 |
} \ |
| 357 |
dest = tmp; \ |
| 358 |
} while(0) |
| 359 |
#define NEON_FN(dest, src1, src2) NEON_SSAT(dest, src1, src2, int8_t)
|
| 360 |
NEON_VOP(qsub_s8, neon_s8, 4)
|
| 361 |
#undef NEON_FN
|
| 362 |
#define NEON_FN(dest, src1, src2) NEON_SSAT(dest, src1, src2, int16_t)
|
| 363 |
NEON_VOP(qsub_s16, neon_s16, 2)
|
| 364 |
#undef NEON_FN
|
| 365 |
#undef NEON_SSAT
|
| 366 |
|
| 367 |
#define NEON_FN(dest, src1, src2) dest = (src1 > src2) ? ~0 : 0 |
| 368 |
NEON_VOP(cgt_s8, neon_s8, 4)
|
| 369 |
NEON_VOP(cgt_u8, neon_u8, 4)
|
| 370 |
NEON_VOP(cgt_s16, neon_s16, 2)
|
| 371 |
NEON_VOP(cgt_u16, neon_u16, 2)
|
| 372 |
NEON_VOP(cgt_s32, neon_s32, 1)
|
| 373 |
NEON_VOP(cgt_u32, neon_u32, 1)
|
| 374 |
#undef NEON_FN
|
| 375 |
|
| 376 |
#define NEON_FN(dest, src1, src2) dest = (src1 >= src2) ? ~0 : 0 |
| 377 |
NEON_VOP(cge_s8, neon_s8, 4)
|
| 378 |
NEON_VOP(cge_u8, neon_u8, 4)
|
| 379 |
NEON_VOP(cge_s16, neon_s16, 2)
|
| 380 |
NEON_VOP(cge_u16, neon_u16, 2)
|
| 381 |
NEON_VOP(cge_s32, neon_s32, 1)
|
| 382 |
NEON_VOP(cge_u32, neon_u32, 1)
|
| 383 |
#undef NEON_FN
|
| 384 |
|
| 385 |
#define NEON_FN(dest, src1, src2) do { \ |
| 386 |
int8_t tmp; \ |
| 387 |
tmp = (int8_t)src2; \ |
| 388 |
if (tmp < 0) { \ |
| 389 |
dest = src1 >> -tmp; \ |
| 390 |
} else { \
|
| 391 |
dest = src1 << tmp; \ |
| 392 |
}} while (0) |
| 393 |
NEON_VOP(shl_s8, neon_s8, 4)
|
| 394 |
NEON_VOP(shl_u8, neon_u8, 4)
|
| 395 |
NEON_VOP(shl_s16, neon_s16, 2)
|
| 396 |
NEON_VOP(shl_u16, neon_u16, 2)
|
| 397 |
NEON_VOP(shl_s32, neon_s32, 1)
|
| 398 |
NEON_VOP(shl_u32, neon_u32, 1)
|
| 399 |
#undef NEON_FN
|
| 400 |
|
| 401 |
NEON_OP(shl_u64) |
| 402 |
{
|
| 403 |
int8_t shift = T2; |
| 404 |
uint64_t val = T0 | ((uint64_t)T1 << 32);
|
| 405 |
if (shift < 0) { |
| 406 |
val >>= -shift; |
| 407 |
} else {
|
| 408 |
val <<= shift; |
| 409 |
} |
| 410 |
T0 = val; |
| 411 |
T1 = val >> 32;
|
| 412 |
FORCE_RET(); |
| 413 |
} |
| 414 |
|
| 415 |
NEON_OP(shl_s64) |
| 416 |
{
|
| 417 |
int8_t shift = T2; |
| 418 |
int64_t val = T0 | ((uint64_t)T1 << 32);
|
| 419 |
if (shift < 0) { |
| 420 |
val >>= -shift; |
| 421 |
} else {
|
| 422 |
val <<= shift; |
| 423 |
} |
| 424 |
T0 = val; |
| 425 |
T1 = val >> 32;
|
| 426 |
FORCE_RET(); |
| 427 |
} |
| 428 |
|
| 429 |
#define NEON_FN(dest, src1, src2) do { \ |
| 430 |
int8_t tmp; \ |
| 431 |
tmp = (int8_t)src1; \ |
| 432 |
if (tmp < 0) { \ |
| 433 |
dest = (src2 + (1 << (-1 - tmp))) >> -tmp; \ |
| 434 |
} else { \
|
| 435 |
dest = src2 << tmp; \ |
| 436 |
}} while (0) |
| 437 |
|
| 438 |
NEON_VOP(rshl_s8, neon_s8, 4)
|
| 439 |
NEON_VOP(rshl_u8, neon_u8, 4)
|
| 440 |
NEON_VOP(rshl_s16, neon_s16, 2)
|
| 441 |
NEON_VOP(rshl_u16, neon_u16, 2)
|
| 442 |
NEON_VOP(rshl_s32, neon_s32, 1)
|
| 443 |
NEON_VOP(rshl_u32, neon_u32, 1)
|
| 444 |
#undef NEON_FN
|
| 445 |
|
| 446 |
NEON_OP(rshl_u64) |
| 447 |
{
|
| 448 |
int8_t shift = T2; |
| 449 |
uint64_t val = T0 | ((uint64_t)T1 << 32);
|
| 450 |
if (shift < 0) { |
| 451 |
val = (val + ((uint64_t)1 << (-1 - shift))) >> -shift; |
| 452 |
val >>= -shift; |
| 453 |
} else {
|
| 454 |
val <<= shift; |
| 455 |
} |
| 456 |
T0 = val; |
| 457 |
T1 = val >> 32;
|
| 458 |
FORCE_RET(); |
| 459 |
} |
| 460 |
|
| 461 |
NEON_OP(rshl_s64) |
| 462 |
{
|
| 463 |
int8_t shift = T2; |
| 464 |
int64_t val = T0 | ((uint64_t)T1 << 32);
|
| 465 |
if (shift < 0) { |
| 466 |
val = (val + ((int64_t)1 << (-1 - shift))) >> -shift; |
| 467 |
} else {
|
| 468 |
val <<= shift; |
| 469 |
} |
| 470 |
T0 = val; |
| 471 |
T1 = val >> 32;
|
| 472 |
FORCE_RET(); |
| 473 |
} |
| 474 |
|
| 475 |
#define NEON_FN(dest, src1, src2) do { \ |
| 476 |
int8_t tmp; \ |
| 477 |
tmp = (int8_t)src1; \ |
| 478 |
if (tmp < 0) { \ |
| 479 |
dest = src2 >> -tmp; \ |
| 480 |
} else { \
|
| 481 |
dest = src2 << tmp; \ |
| 482 |
if ((dest >> tmp) != src2) { \
|
| 483 |
env->QF = 1; \
|
| 484 |
dest = ~0; \
|
| 485 |
} \ |
| 486 |
}} while (0) |
| 487 |
NEON_VOP(qshl_s8, neon_s8, 4)
|
| 488 |
NEON_VOP(qshl_s16, neon_s16, 2)
|
| 489 |
NEON_VOP(qshl_s32, neon_s32, 1)
|
| 490 |
#undef NEON_FN
|
| 491 |
|
| 492 |
NEON_OP(qshl_s64) |
| 493 |
{
|
| 494 |
int8_t shift = T2; |
| 495 |
int64_t val = T0 | ((uint64_t)T1 << 32);
|
| 496 |
if (shift < 0) { |
| 497 |
val >>= -shift; |
| 498 |
} else {
|
| 499 |
int64_t tmp = val; |
| 500 |
val <<= shift; |
| 501 |
if ((val >> shift) != tmp) {
|
| 502 |
env->QF = 1;
|
| 503 |
val = (tmp >> 63) ^ 0x7fffffffffffffffULL; |
| 504 |
} |
| 505 |
} |
| 506 |
T0 = val; |
| 507 |
T1 = val >> 32;
|
| 508 |
FORCE_RET(); |
| 509 |
} |
| 510 |
|
| 511 |
#define NEON_FN(dest, src1, src2) do { \ |
| 512 |
int8_t tmp; \ |
| 513 |
tmp = (int8_t)src1; \ |
| 514 |
if (tmp < 0) { \ |
| 515 |
dest = src2 >> -tmp; \ |
| 516 |
} else { \
|
| 517 |
dest = src2 << tmp; \ |
| 518 |
if ((dest >> tmp) != src2) { \
|
| 519 |
env->QF = 1; \
|
| 520 |
dest = src2 >> 31; \
|
| 521 |
} \ |
| 522 |
}} while (0) |
| 523 |
NEON_VOP(qshl_u8, neon_u8, 4)
|
| 524 |
NEON_VOP(qshl_u16, neon_u16, 2)
|
| 525 |
NEON_VOP(qshl_u32, neon_u32, 1)
|
| 526 |
#undef NEON_FN
|
| 527 |
|
| 528 |
NEON_OP(qshl_u64) |
| 529 |
{
|
| 530 |
int8_t shift = T2; |
| 531 |
uint64_t val = T0 | ((uint64_t)T1 << 32);
|
| 532 |
if (shift < 0) { |
| 533 |
val >>= -shift; |
| 534 |
} else {
|
| 535 |
uint64_t tmp = val; |
| 536 |
val <<= shift; |
| 537 |
if ((val >> shift) != tmp) {
|
| 538 |
env->QF = 1;
|
| 539 |
val = ~(uint64_t)0;
|
| 540 |
} |
| 541 |
} |
| 542 |
T0 = val; |
| 543 |
T1 = val >> 32;
|
| 544 |
FORCE_RET(); |
| 545 |
} |
| 546 |
|
| 547 |
#define NEON_FN(dest, src1, src2) do { \ |
| 548 |
int8_t tmp; \ |
| 549 |
tmp = (int8_t)src1; \ |
| 550 |
if (tmp < 0) { \ |
| 551 |
dest = (src2 + (1 << (-1 - tmp))) >> -tmp; \ |
| 552 |
} else { \
|
| 553 |
dest = src2 << tmp; \ |
| 554 |
if ((dest >> tmp) != src2) { \
|
| 555 |
dest = ~0; \
|
| 556 |
} \ |
| 557 |
}} while (0) |
| 558 |
NEON_VOP(qrshl_s8, neon_s8, 4)
|
| 559 |
NEON_VOP(qrshl_s16, neon_s16, 2)
|
| 560 |
NEON_VOP(qrshl_s32, neon_s32, 1)
|
| 561 |
#undef NEON_FN
|
| 562 |
|
| 563 |
#define NEON_FN(dest, src1, src2) do { \ |
| 564 |
int8_t tmp; \ |
| 565 |
tmp = (int8_t)src1; \ |
| 566 |
if (tmp < 0) { \ |
| 567 |
dest = (src2 + (1 << (-1 - tmp))) >> -tmp; \ |
| 568 |
} else { \
|
| 569 |
dest = src2 << tmp; \ |
| 570 |
if ((dest >> tmp) != src2) { \
|
| 571 |
env->QF = 1; \
|
| 572 |
dest = src2 >> 31; \
|
| 573 |
} \ |
| 574 |
}} while (0) |
| 575 |
NEON_VOP(qrshl_u8, neon_u8, 4)
|
| 576 |
NEON_VOP(qrshl_u16, neon_u16, 2)
|
| 577 |
NEON_VOP(qrshl_u32, neon_u32, 1)
|
| 578 |
#undef NEON_FN
|
| 579 |
|
| 580 |
#define NEON_FN(dest, src1, src2) dest = (src1 > src2) ? src1 : src2
|
| 581 |
NEON_VOP(max_s8, neon_s8, 4)
|
| 582 |
NEON_VOP(max_u8, neon_u8, 4)
|
| 583 |
NEON_VOP(max_s16, neon_s16, 2)
|
| 584 |
NEON_VOP(max_u16, neon_u16, 2)
|
| 585 |
NEON_VOP(max_s32, neon_s32, 1)
|
| 586 |
NEON_VOP(max_u32, neon_u32, 1)
|
| 587 |
NEON_POP(pmax_s8, neon_s8, 4)
|
| 588 |
NEON_POP(pmax_u8, neon_u8, 4)
|
| 589 |
NEON_POP(pmax_s16, neon_s16, 2)
|
| 590 |
NEON_POP(pmax_u16, neon_u16, 2)
|
| 591 |
#undef NEON_FN
|
| 592 |
|
| 593 |
NEON_OP(max_f32) |
| 594 |
{
|
| 595 |
float32 f0 = vfp_itos(T0); |
| 596 |
float32 f1 = vfp_itos(T1); |
| 597 |
T0 = (float32_compare_quiet(f0, f1, NFS) == 1) ? T0 : T1;
|
| 598 |
FORCE_RET(); |
| 599 |
} |
| 600 |
|
| 601 |
#define NEON_FN(dest, src1, src2) dest = (src1 < src2) ? src1 : src2
|
| 602 |
NEON_VOP(min_s8, neon_s8, 4)
|
| 603 |
NEON_VOP(min_u8, neon_u8, 4)
|
| 604 |
NEON_VOP(min_s16, neon_s16, 2)
|
| 605 |
NEON_VOP(min_u16, neon_u16, 2)
|
| 606 |
NEON_VOP(min_s32, neon_s32, 1)
|
| 607 |
NEON_VOP(min_u32, neon_u32, 1)
|
| 608 |
NEON_POP(pmin_s8, neon_s8, 4)
|
| 609 |
NEON_POP(pmin_u8, neon_u8, 4)
|
| 610 |
NEON_POP(pmin_s16, neon_s16, 2)
|
| 611 |
NEON_POP(pmin_u16, neon_u16, 2)
|
| 612 |
#undef NEON_FN
|
| 613 |
|
| 614 |
NEON_OP(min_f32) |
| 615 |
{
|
| 616 |
float32 f0 = vfp_itos(T0); |
| 617 |
float32 f1 = vfp_itos(T1); |
| 618 |
T0 = (float32_compare_quiet(f0, f1, NFS) == -1) ? T0 : T1;
|
| 619 |
FORCE_RET(); |
| 620 |
} |
| 621 |
|
| 622 |
#define NEON_FN(dest, src1, src2) \
|
| 623 |
dest = (src1 > src2) ? (src1 - src2) : (src2 - src1) |
| 624 |
NEON_VOP(abd_s8, neon_s8, 4)
|
| 625 |
NEON_VOP(abd_u8, neon_u8, 4)
|
| 626 |
NEON_VOP(abd_s16, neon_s16, 2)
|
| 627 |
NEON_VOP(abd_u16, neon_u16, 2)
|
| 628 |
NEON_VOP(abd_s32, neon_s32, 1)
|
| 629 |
NEON_VOP(abd_u32, neon_u32, 1)
|
| 630 |
#undef NEON_FN
|
| 631 |
|
| 632 |
NEON_OP(abd_f32) |
| 633 |
{
|
| 634 |
float32 f0 = vfp_itos(T0); |
| 635 |
float32 f1 = vfp_itos(T1); |
| 636 |
T0 = vfp_stoi((float32_compare_quiet(f0, f1, NFS) == 1)
|
| 637 |
? float32_sub(f0, f1, NFS) |
| 638 |
: float32_sub(f1, f0, NFS)); |
| 639 |
FORCE_RET(); |
| 640 |
} |
| 641 |
|
| 642 |
#define NEON_FN(dest, src1, src2) dest = src1 + src2
|
| 643 |
NEON_VOP(add_u8, neon_u8, 4)
|
| 644 |
NEON_VOP(add_u16, neon_u16, 2)
|
| 645 |
NEON_POP(padd_u8, neon_u8, 4)
|
| 646 |
NEON_POP(padd_u16, neon_u16, 2)
|
| 647 |
#undef NEON_FN
|
| 648 |
|
| 649 |
NEON_OP(add_f32) |
| 650 |
{
|
| 651 |
T0 = vfp_stoi(float32_add(vfp_itos(T0), vfp_itos(T1), NFS)); |
| 652 |
FORCE_RET(); |
| 653 |
} |
| 654 |
|
| 655 |
#define NEON_FN(dest, src1, src2) dest = src1 - src2
|
| 656 |
NEON_VOP(sub_u8, neon_u8, 4)
|
| 657 |
NEON_VOP(sub_u16, neon_u16, 2)
|
| 658 |
#undef NEON_FN
|
| 659 |
|
| 660 |
NEON_OP(sub_f32) |
| 661 |
{
|
| 662 |
T0 = vfp_stoi(float32_sub(vfp_itos(T0), vfp_itos(T1), NFS)); |
| 663 |
FORCE_RET(); |
| 664 |
} |
| 665 |
|
| 666 |
#define NEON_FN(dest, src1, src2) dest = src2 - src1
|
| 667 |
NEON_VOP(rsb_u8, neon_u8, 4)
|
| 668 |
NEON_VOP(rsb_u16, neon_u16, 2)
|
| 669 |
#undef NEON_FN
|
| 670 |
|
| 671 |
NEON_OP(rsb_f32) |
| 672 |
{
|
| 673 |
T0 = vfp_stoi(float32_sub(vfp_itos(T1), vfp_itos(T0), NFS)); |
| 674 |
FORCE_RET(); |
| 675 |
} |
| 676 |
|
| 677 |
#define NEON_FN(dest, src1, src2) dest = src1 * src2
|
| 678 |
NEON_VOP(mul_u8, neon_u8, 4)
|
| 679 |
NEON_VOP(mul_u16, neon_u16, 2)
|
| 680 |
#undef NEON_FN
|
| 681 |
|
| 682 |
NEON_OP(mul_f32) |
| 683 |
{
|
| 684 |
T0 = vfp_stoi(float32_mul(vfp_itos(T0), vfp_itos(T1), NFS)); |
| 685 |
FORCE_RET(); |
| 686 |
} |
| 687 |
|
| 688 |
NEON_OP(mul_p8) |
| 689 |
{
|
| 690 |
T0 = helper_neon_mul_p8(T0, T1); |
| 691 |
} |
| 692 |
|
| 693 |
#define NEON_FN(dest, src1, src2) dest = (src1 & src2) ? -1 : 0 |
| 694 |
NEON_VOP(tst_u8, neon_u8, 4)
|
| 695 |
NEON_VOP(tst_u16, neon_u16, 2)
|
| 696 |
NEON_VOP(tst_u32, neon_u32, 1)
|
| 697 |
#undef NEON_FN
|
| 698 |
|
| 699 |
#define NEON_FN(dest, src1, src2) dest = (src1 == src2) ? -1 : 0 |
| 700 |
NEON_VOP(ceq_u8, neon_u8, 4)
|
| 701 |
NEON_VOP(ceq_u16, neon_u16, 2)
|
| 702 |
NEON_VOP(ceq_u32, neon_u32, 1)
|
| 703 |
#undef NEON_FN
|
| 704 |
|
| 705 |
#define NEON_QDMULH16(dest, src1, src2, round) do { \ |
| 706 |
uint32_t tmp = (int32_t)(int16_t) src1 * (int16_t) src2; \ |
| 707 |
if ((tmp ^ (tmp << 1)) & SIGNBIT) { \ |
| 708 |
env->QF = 1; \
|
| 709 |
tmp = (tmp >> 31) ^ ~SIGNBIT; \
|
| 710 |
} \ |
| 711 |
tmp <<= 1; \
|
| 712 |
if (round) { \
|
| 713 |
int32_t old = tmp; \ |
| 714 |
tmp += 1 << 15; \ |
| 715 |
if ((int32_t)tmp < old) { \
|
| 716 |
env->QF = 1; \
|
| 717 |
tmp = SIGNBIT - 1; \
|
| 718 |
} \ |
| 719 |
} \ |
| 720 |
dest = tmp >> 16; \
|
| 721 |
} while(0) |
| 722 |
#define NEON_FN(dest, src1, src2) NEON_QDMULH16(dest, src1, src2, 0) |
| 723 |
NEON_VOP(qdmulh_s16, neon_s16, 2)
|
| 724 |
#undef NEON_FN
|
| 725 |
#define NEON_FN(dest, src1, src2) NEON_QDMULH16(dest, src1, src2, 1) |
| 726 |
NEON_VOP(qrdmulh_s16, neon_s16, 2)
|
| 727 |
#undef NEON_FN
|
| 728 |
#undef NEON_QDMULH16
|
| 729 |
|
| 730 |
#define SIGNBIT64 ((uint64_t)1 << 63) |
| 731 |
#define NEON_QDMULH32(dest, src1, src2, round) do { \ |
| 732 |
uint64_t tmp = (int64_t)(int32_t) src1 * (int32_t) src2; \ |
| 733 |
if ((tmp ^ (tmp << 1)) & SIGNBIT64) { \ |
| 734 |
env->QF = 1; \
|
| 735 |
tmp = (tmp >> 63) ^ ~SIGNBIT64; \
|
| 736 |
} else { \
|
| 737 |
tmp <<= 1; \
|
| 738 |
} \ |
| 739 |
if (round) { \
|
| 740 |
int64_t old = tmp; \ |
| 741 |
tmp += (int64_t)1 << 31; \ |
| 742 |
if ((int64_t)tmp < old) { \
|
| 743 |
env->QF = 1; \
|
| 744 |
tmp = SIGNBIT64 - 1; \
|
| 745 |
} \ |
| 746 |
} \ |
| 747 |
dest = tmp >> 32; \
|
| 748 |
} while(0) |
| 749 |
#define NEON_FN(dest, src1, src2) NEON_QDMULH32(dest, src1, src2, 0) |
| 750 |
NEON_VOP(qdmulh_s32, neon_s32, 1)
|
| 751 |
#undef NEON_FN
|
| 752 |
#define NEON_FN(dest, src1, src2) NEON_QDMULH32(dest, src1, src2, 1) |
| 753 |
NEON_VOP(qrdmulh_s32, neon_s32, 1)
|
| 754 |
#undef NEON_FN
|
| 755 |
#undef NEON_QDMULH32
|
| 756 |
|
| 757 |
NEON_OP(recps_f32) |
| 758 |
{
|
| 759 |
T0 = vfp_stoi(helper_recps_f32(vfp_itos(T0), vfp_itos(T1))); |
| 760 |
FORCE_RET(); |
| 761 |
} |
| 762 |
|
| 763 |
NEON_OP(rsqrts_f32) |
| 764 |
{
|
| 765 |
T0 = vfp_stoi(helper_rsqrts_f32(vfp_itos(T0), vfp_itos(T1))); |
| 766 |
FORCE_RET(); |
| 767 |
} |
| 768 |
|
| 769 |
/* Floating point comparisons produce an integer result. */
|
| 770 |
#define NEON_VOP_FCMP(name, cmp) \
|
| 771 |
NEON_OP(name) \ |
| 772 |
{ \
|
| 773 |
if (float32_compare_quiet(vfp_itos(T0), vfp_itos(T1), NFS) cmp 0) \ |
| 774 |
T0 = -1; \
|
| 775 |
else \
|
| 776 |
T0 = 0; \
|
| 777 |
FORCE_RET(); \ |
| 778 |
} |
| 779 |
|
| 780 |
NEON_VOP_FCMP(ceq_f32, ==) |
| 781 |
NEON_VOP_FCMP(cge_f32, >=) |
| 782 |
NEON_VOP_FCMP(cgt_f32, >) |
| 783 |
|
| 784 |
NEON_OP(acge_f32) |
| 785 |
{
|
| 786 |
float32 f0 = float32_abs(vfp_itos(T0)); |
| 787 |
float32 f1 = float32_abs(vfp_itos(T1)); |
| 788 |
T0 = (float32_compare_quiet(f0, f1,NFS) >= 0) ? -1 : 0; |
| 789 |
FORCE_RET(); |
| 790 |
} |
| 791 |
|
| 792 |
NEON_OP(acgt_f32) |
| 793 |
{
|
| 794 |
float32 f0 = float32_abs(vfp_itos(T0)); |
| 795 |
float32 f1 = float32_abs(vfp_itos(T1)); |
| 796 |
T0 = (float32_compare_quiet(f0, f1, NFS) > 0) ? -1 : 0; |
| 797 |
FORCE_RET(); |
| 798 |
} |
| 799 |
|
| 800 |
/* Narrowing instructions. The named type is the destination type. */
|
| 801 |
NEON_OP(narrow_u8) |
| 802 |
{
|
| 803 |
T0 = (T0 & 0xff) | ((T0 >> 8) & 0xff00) |
| 804 |
| ((T1 << 16) & 0xff0000) | (T1 << 24); |
| 805 |
FORCE_RET(); |
| 806 |
} |
| 807 |
|
| 808 |
NEON_OP(narrow_sat_u8) |
| 809 |
{
|
| 810 |
neon_u16 src; |
| 811 |
neon_u8 dest; |
| 812 |
#define SAT8(d, s) \
|
| 813 |
if (s > 0xff) { \ |
| 814 |
d = 0xff; \
|
| 815 |
env->QF = 1; \
|
| 816 |
} else { \
|
| 817 |
d = s; \ |
| 818 |
} |
| 819 |
|
| 820 |
NEON_UNPACK(neon_u16, src, T0); |
| 821 |
SAT8(dest.v1, src.v1); |
| 822 |
SAT8(dest.v2, src.v2); |
| 823 |
NEON_UNPACK(neon_u16, src, T1); |
| 824 |
SAT8(dest.v3, src.v1); |
| 825 |
SAT8(dest.v4, src.v2); |
| 826 |
NEON_PACK(neon_u8, T0, dest); |
| 827 |
FORCE_RET(); |
| 828 |
#undef SAT8
|
| 829 |
} |
| 830 |
|
| 831 |
NEON_OP(narrow_sat_s8) |
| 832 |
{
|
| 833 |
neon_s16 src; |
| 834 |
neon_s8 dest; |
| 835 |
#define SAT8(d, s) \
|
| 836 |
if (s != (uint8_t)s) { \
|
| 837 |
d = (s >> 15) ^ 0x7f; \ |
| 838 |
env->QF = 1; \
|
| 839 |
} else { \
|
| 840 |
d = s; \ |
| 841 |
} |
| 842 |
|
| 843 |
NEON_UNPACK(neon_s16, src, T0); |
| 844 |
SAT8(dest.v1, src.v1); |
| 845 |
SAT8(dest.v2, src.v2); |
| 846 |
NEON_UNPACK(neon_s16, src, T1); |
| 847 |
SAT8(dest.v3, src.v1); |
| 848 |
SAT8(dest.v4, src.v2); |
| 849 |
NEON_PACK(neon_s8, T0, dest); |
| 850 |
FORCE_RET(); |
| 851 |
#undef SAT8
|
| 852 |
} |
| 853 |
|
| 854 |
NEON_OP(narrow_u16) |
| 855 |
{
|
| 856 |
T0 = (T0 & 0xffff) | (T1 << 16); |
| 857 |
} |
| 858 |
|
| 859 |
NEON_OP(narrow_sat_u16) |
| 860 |
{
|
| 861 |
if (T0 > 0xffff) { |
| 862 |
T0 = 0xffff;
|
| 863 |
env->QF = 1;
|
| 864 |
} |
| 865 |
if (T1 > 0xffff) { |
| 866 |
T1 = 0xffff;
|
| 867 |
env->QF = 1;
|
| 868 |
} |
| 869 |
T0 |= T1 << 16;
|
| 870 |
FORCE_RET(); |
| 871 |
} |
| 872 |
|
| 873 |
NEON_OP(narrow_sat_s16) |
| 874 |
{
|
| 875 |
if ((int32_t)T0 != (int16_t)T0) {
|
| 876 |
T0 = ((int32_t)T0 >> 31) ^ 0x7fff; |
| 877 |
env->QF = 1;
|
| 878 |
} |
| 879 |
if ((int32_t)T1 != (int16_t) T1) {
|
| 880 |
T1 = ((int32_t)T1 >> 31) ^ 0x7fff; |
| 881 |
env->QF = 1;
|
| 882 |
} |
| 883 |
T0 = (uint16_t)T0 | (T1 << 16);
|
| 884 |
FORCE_RET(); |
| 885 |
} |
| 886 |
|
| 887 |
NEON_OP(narrow_sat_u32) |
| 888 |
{
|
| 889 |
if (T1) {
|
| 890 |
T0 = 0xffffffffu;
|
| 891 |
env->QF = 1;
|
| 892 |
} |
| 893 |
FORCE_RET(); |
| 894 |
} |
| 895 |
|
| 896 |
NEON_OP(narrow_sat_s32) |
| 897 |
{
|
| 898 |
int32_t sign = (int32_t)T1 >> 31;
|
| 899 |
|
| 900 |
if ((int32_t)T1 != sign) {
|
| 901 |
T0 = sign ^ 0x7fffffff;
|
| 902 |
env->QF = 1;
|
| 903 |
} |
| 904 |
FORCE_RET(); |
| 905 |
} |
| 906 |
|
| 907 |
/* Narrowing instructions. Named type is the narrow type. */
|
| 908 |
NEON_OP(narrow_high_u8) |
| 909 |
{
|
| 910 |
T0 = ((T0 >> 8) & 0xff) | ((T0 >> 16) & 0xff00) |
| 911 |
| ((T1 << 8) & 0xff0000) | (T1 & 0xff000000); |
| 912 |
FORCE_RET(); |
| 913 |
} |
| 914 |
|
| 915 |
NEON_OP(narrow_high_u16) |
| 916 |
{
|
| 917 |
T0 = (T0 >> 16) | (T1 & 0xffff0000); |
| 918 |
FORCE_RET(); |
| 919 |
} |
| 920 |
|
| 921 |
NEON_OP(narrow_high_round_u8) |
| 922 |
{
|
| 923 |
T0 = (((T0 + 0x80) >> 8) & 0xff) | (((T0 + 0x800000) >> 16) & 0xff00) |
| 924 |
| (((T1 + 0x80) << 8) & 0xff0000) | ((T1 + 0x800000) & 0xff000000); |
| 925 |
FORCE_RET(); |
| 926 |
} |
| 927 |
|
| 928 |
NEON_OP(narrow_high_round_u16) |
| 929 |
{
|
| 930 |
T0 = ((T0 + 0x8000) >> 16) | ((T1 + 0x8000) & 0xffff0000); |
| 931 |
FORCE_RET(); |
| 932 |
} |
| 933 |
|
| 934 |
NEON_OP(narrow_high_round_u32) |
| 935 |
{
|
| 936 |
if (T0 >= 0x80000000u) |
| 937 |
T0 = T1 + 1;
|
| 938 |
else
|
| 939 |
T0 = T1; |
| 940 |
FORCE_RET(); |
| 941 |
} |
| 942 |
|
| 943 |
/* Widening instructions. Named type is source type. */
|
| 944 |
NEON_OP(widen_s8) |
| 945 |
{
|
| 946 |
uint32_t src; |
| 947 |
|
| 948 |
src = T0; |
| 949 |
T0 = (uint16_t)(int8_t)src | ((int8_t)(src >> 8) << 16); |
| 950 |
T1 = (uint16_t)(int8_t)(src >> 16) | ((int8_t)(src >> 24) << 16); |
| 951 |
} |
| 952 |
|
| 953 |
NEON_OP(widen_u8) |
| 954 |
{
|
| 955 |
T1 = ((T0 >> 8) & 0xff0000) | ((T0 >> 16) & 0xff); |
| 956 |
T0 = ((T0 << 8) & 0xff0000) | (T0 & 0xff); |
| 957 |
} |
| 958 |
|
| 959 |
NEON_OP(widen_s16) |
| 960 |
{
|
| 961 |
int32_t src; |
| 962 |
|
| 963 |
src = T0; |
| 964 |
T0 = (int16_t)src; |
| 965 |
T1 = src >> 16;
|
| 966 |
} |
| 967 |
|
| 968 |
NEON_OP(widen_u16) |
| 969 |
{
|
| 970 |
T1 = T0 >> 16;
|
| 971 |
T0 &= 0xffff;
|
| 972 |
} |
| 973 |
|
| 974 |
NEON_OP(widen_s32) |
| 975 |
{
|
| 976 |
T1 = (int32_t)T0 >> 31;
|
| 977 |
FORCE_RET(); |
| 978 |
} |
| 979 |
|
| 980 |
NEON_OP(widen_high_u8) |
| 981 |
{
|
| 982 |
T1 = (T0 & 0xff000000) | ((T0 >> 8) & 0xff00); |
| 983 |
T0 = ((T0 << 16) & 0xff000000) | ((T0 << 8) & 0xff00); |
| 984 |
} |
| 985 |
|
| 986 |
NEON_OP(widen_high_u16) |
| 987 |
{
|
| 988 |
T1 = T0 & 0xffff0000;
|
| 989 |
T0 <<= 16;
|
| 990 |
} |
| 991 |
|
| 992 |
/* Long operations. The type is the wide type. */
|
| 993 |
NEON_OP(shll_u16) |
| 994 |
{
|
| 995 |
int shift = PARAM1;
|
| 996 |
uint32_t mask; |
| 997 |
|
| 998 |
mask = 0xffff >> (16 - shift); |
| 999 |
mask |= mask << 16;
|
| 1000 |
mask = ~mask; |
| 1001 |
|
| 1002 |
T0 = (T0 << shift) & mask; |
| 1003 |
T1 = (T1 << shift) & mask; |
| 1004 |
FORCE_RET(); |
| 1005 |
} |
| 1006 |
|
| 1007 |
NEON_OP(shll_u64) |
| 1008 |
{
|
| 1009 |
int shift = PARAM1;
|
| 1010 |
|
| 1011 |
T1 <<= shift; |
| 1012 |
T1 |= T0 >> (32 - shift);
|
| 1013 |
T0 <<= shift; |
| 1014 |
FORCE_RET(); |
| 1015 |
} |
| 1016 |
|
| 1017 |
NEON_OP(addl_u16) |
| 1018 |
{
|
| 1019 |
uint32_t tmp; |
| 1020 |
uint32_t high; |
| 1021 |
|
| 1022 |
tmp = env->vfp.scratch[0];
|
| 1023 |
high = (T0 >> 16) + (tmp >> 16); |
| 1024 |
T0 = (uint16_t)(T0 + tmp); |
| 1025 |
T0 |= (high << 16);
|
| 1026 |
tmp = env->vfp.scratch[1];
|
| 1027 |
high = (T1 >> 16) + (tmp >> 16); |
| 1028 |
T1 = (uint16_t)(T1 + tmp); |
| 1029 |
T1 |= (high << 16);
|
| 1030 |
FORCE_RET(); |
| 1031 |
} |
| 1032 |
|
| 1033 |
NEON_OP(addl_u32) |
| 1034 |
{
|
| 1035 |
T0 += env->vfp.scratch[0];
|
| 1036 |
T1 += env->vfp.scratch[1];
|
| 1037 |
FORCE_RET(); |
| 1038 |
} |
| 1039 |
|
| 1040 |
NEON_OP(addl_u64) |
| 1041 |
{
|
| 1042 |
uint64_t tmp; |
| 1043 |
tmp = T0 | ((uint64_t)T1 << 32);
|
| 1044 |
tmp += env->vfp.scratch[0];
|
| 1045 |
tmp += (uint64_t)env->vfp.scratch[1] << 32; |
| 1046 |
T0 = tmp; |
| 1047 |
T1 = tmp >> 32;
|
| 1048 |
FORCE_RET(); |
| 1049 |
} |
| 1050 |
|
| 1051 |
NEON_OP(subl_u16) |
| 1052 |
{
|
| 1053 |
uint32_t tmp; |
| 1054 |
uint32_t high; |
| 1055 |
|
| 1056 |
tmp = env->vfp.scratch[0];
|
| 1057 |
high = (T0 >> 16) - (tmp >> 16); |
| 1058 |
T0 = (uint16_t)(T0 - tmp); |
| 1059 |
T0 |= (high << 16);
|
| 1060 |
tmp = env->vfp.scratch[1];
|
| 1061 |
high = (T1 >> 16) - (tmp >> 16); |
| 1062 |
T1 = (uint16_t)(T1 - tmp); |
| 1063 |
T1 |= (high << 16);
|
| 1064 |
FORCE_RET(); |
| 1065 |
} |
| 1066 |
|
| 1067 |
NEON_OP(subl_u32) |
| 1068 |
{
|
| 1069 |
T0 -= env->vfp.scratch[0];
|
| 1070 |
T1 -= env->vfp.scratch[1];
|
| 1071 |
FORCE_RET(); |
| 1072 |
} |
| 1073 |
|
| 1074 |
NEON_OP(subl_u64) |
| 1075 |
{
|
| 1076 |
uint64_t tmp; |
| 1077 |
tmp = T0 | ((uint64_t)T1 << 32);
|
| 1078 |
tmp -= env->vfp.scratch[0];
|
| 1079 |
tmp -= (uint64_t)env->vfp.scratch[1] << 32; |
| 1080 |
T0 = tmp; |
| 1081 |
T1 = tmp >> 32;
|
| 1082 |
FORCE_RET(); |
| 1083 |
} |
| 1084 |
|
| 1085 |
#define DO_ABD(dest, x, y, type) do { \ |
| 1086 |
type tmp_x = x; \ |
| 1087 |
type tmp_y = y; \ |
| 1088 |
dest = ((tmp_x > tmp_y) ? tmp_x - tmp_y : tmp_y - tmp_x); \ |
| 1089 |
} while(0) |
| 1090 |
|
| 1091 |
NEON_OP(abdl_u16) |
| 1092 |
{
|
| 1093 |
uint32_t tmp; |
| 1094 |
uint32_t low; |
| 1095 |
uint32_t high; |
| 1096 |
|
| 1097 |
DO_ABD(low, T0, T1, uint8_t); |
| 1098 |
DO_ABD(tmp, T0 >> 8, T1 >> 8, uint8_t); |
| 1099 |
low |= tmp << 16;
|
| 1100 |
DO_ABD(high, T0 >> 16, T1 >> 16, uint8_t); |
| 1101 |
DO_ABD(tmp, T0 >> 24, T1 >> 24, uint8_t); |
| 1102 |
high |= tmp << 16;
|
| 1103 |
T0 = low; |
| 1104 |
T1 = high; |
| 1105 |
FORCE_RET(); |
| 1106 |
} |
| 1107 |
|
| 1108 |
NEON_OP(abdl_s16) |
| 1109 |
{
|
| 1110 |
uint32_t tmp; |
| 1111 |
uint32_t low; |
| 1112 |
uint32_t high; |
| 1113 |
|
| 1114 |
DO_ABD(low, T0, T1, int8_t); |
| 1115 |
DO_ABD(tmp, T0 >> 8, T1 >> 8, int8_t); |
| 1116 |
low |= tmp << 16;
|
| 1117 |
DO_ABD(high, T0 >> 16, T1 >> 16, int8_t); |
| 1118 |
DO_ABD(tmp, T0 >> 24, T1 >> 24, int8_t); |
| 1119 |
high |= tmp << 16;
|
| 1120 |
T0 = low; |
| 1121 |
T1 = high; |
| 1122 |
FORCE_RET(); |
| 1123 |
} |
| 1124 |
|
| 1125 |
NEON_OP(abdl_u32) |
| 1126 |
{
|
| 1127 |
uint32_t low; |
| 1128 |
uint32_t high; |
| 1129 |
|
| 1130 |
DO_ABD(low, T0, T1, uint16_t); |
| 1131 |
DO_ABD(high, T0 >> 16, T1 >> 16, uint16_t); |
| 1132 |
T0 = low; |
| 1133 |
T1 = high; |
| 1134 |
FORCE_RET(); |
| 1135 |
} |
| 1136 |
|
| 1137 |
NEON_OP(abdl_s32) |
| 1138 |
{
|
| 1139 |
uint32_t low; |
| 1140 |
uint32_t high; |
| 1141 |
|
| 1142 |
DO_ABD(low, T0, T1, int16_t); |
| 1143 |
DO_ABD(high, T0 >> 16, T1 >> 16, int16_t); |
| 1144 |
T0 = low; |
| 1145 |
T1 = high; |
| 1146 |
FORCE_RET(); |
| 1147 |
} |
| 1148 |
|
| 1149 |
NEON_OP(abdl_u64) |
| 1150 |
{
|
| 1151 |
DO_ABD(T0, T0, T1, uint32_t); |
| 1152 |
T1 = 0;
|
| 1153 |
} |
| 1154 |
|
| 1155 |
NEON_OP(abdl_s64) |
| 1156 |
{
|
| 1157 |
DO_ABD(T0, T0, T1, int32_t); |
| 1158 |
T1 = 0;
|
| 1159 |
} |
| 1160 |
#undef DO_ABD
|
| 1161 |
|
| 1162 |
/* Widening multiple. Named type is the source type. */
|
| 1163 |
#define DO_MULL(dest, x, y, type1, type2) do { \ |
| 1164 |
type1 tmp_x = x; \ |
| 1165 |
type1 tmp_y = y; \ |
| 1166 |
dest = (type2)((type2)tmp_x * (type2)tmp_y); \ |
| 1167 |
} while(0) |
| 1168 |
|
| 1169 |
NEON_OP(mull_u8) |
| 1170 |
{
|
| 1171 |
uint32_t tmp; |
| 1172 |
uint32_t low; |
| 1173 |
uint32_t high; |
| 1174 |
|
| 1175 |
DO_MULL(low, T0, T1, uint8_t, uint16_t); |
| 1176 |
DO_MULL(tmp, T0 >> 8, T1 >> 8, uint8_t, uint16_t); |
| 1177 |
low |= tmp << 16;
|
| 1178 |
DO_MULL(high, T0 >> 16, T1 >> 16, uint8_t, uint16_t); |
| 1179 |
DO_MULL(tmp, T0 >> 24, T1 >> 24, uint8_t, uint16_t); |
| 1180 |
high |= tmp << 16;
|
| 1181 |
T0 = low; |
| 1182 |
T1 = high; |
| 1183 |
FORCE_RET(); |
| 1184 |
} |
| 1185 |
|
| 1186 |
NEON_OP(mull_s8) |
| 1187 |
{
|
| 1188 |
uint32_t tmp; |
| 1189 |
uint32_t low; |
| 1190 |
uint32_t high; |
| 1191 |
|
| 1192 |
DO_MULL(low, T0, T1, int8_t, uint16_t); |
| 1193 |
DO_MULL(tmp, T0 >> 8, T1 >> 8, int8_t, uint16_t); |
| 1194 |
low |= tmp << 16;
|
| 1195 |
DO_MULL(high, T0 >> 16, T1 >> 16, int8_t, uint16_t); |
| 1196 |
DO_MULL(tmp, T0 >> 24, T1 >> 24, int8_t, uint16_t); |
| 1197 |
high |= tmp << 16;
|
| 1198 |
T0 = low; |
| 1199 |
T1 = high; |
| 1200 |
FORCE_RET(); |
| 1201 |
} |
| 1202 |
|
| 1203 |
NEON_OP(mull_u16) |
| 1204 |
{
|
| 1205 |
uint32_t low; |
| 1206 |
uint32_t high; |
| 1207 |
|
| 1208 |
DO_MULL(low, T0, T1, uint16_t, uint32_t); |
| 1209 |
DO_MULL(high, T0 >> 16, T1 >> 16, uint16_t, uint32_t); |
| 1210 |
T0 = low; |
| 1211 |
T1 = high; |
| 1212 |
FORCE_RET(); |
| 1213 |
} |
| 1214 |
|
| 1215 |
NEON_OP(mull_s16) |
| 1216 |
{
|
| 1217 |
uint32_t low; |
| 1218 |
uint32_t high; |
| 1219 |
|
| 1220 |
DO_MULL(low, T0, T1, int16_t, uint32_t); |
| 1221 |
DO_MULL(high, T0 >> 16, T1 >> 16, int16_t, uint32_t); |
| 1222 |
T0 = low; |
| 1223 |
T1 = high; |
| 1224 |
FORCE_RET(); |
| 1225 |
} |
| 1226 |
|
| 1227 |
NEON_OP(addl_saturate_s32) |
| 1228 |
{
|
| 1229 |
uint32_t tmp; |
| 1230 |
uint32_t res; |
| 1231 |
|
| 1232 |
tmp = env->vfp.scratch[0];
|
| 1233 |
res = T0 + tmp; |
| 1234 |
if (((res ^ T0) & SIGNBIT) && !((T0 ^ tmp) & SIGNBIT)) {
|
| 1235 |
env->QF = 1;
|
| 1236 |
T0 = (T0 >> 31) ^ 0x7fffffff; |
| 1237 |
} else {
|
| 1238 |
T0 = res; |
| 1239 |
} |
| 1240 |
tmp = env->vfp.scratch[1];
|
| 1241 |
res = T1 + tmp; |
| 1242 |
if (((res ^ T1) & SIGNBIT) && !((T1 ^ tmp) & SIGNBIT)) {
|
| 1243 |
env->QF = 1;
|
| 1244 |
T1 = (T1 >> 31) ^ 0x7fffffff; |
| 1245 |
} else {
|
| 1246 |
T1 = res; |
| 1247 |
} |
| 1248 |
FORCE_RET(); |
| 1249 |
} |
| 1250 |
|
| 1251 |
NEON_OP(addl_saturate_s64) |
| 1252 |
{
|
| 1253 |
uint64_t src1; |
| 1254 |
uint64_t src2; |
| 1255 |
uint64_t res; |
| 1256 |
|
| 1257 |
src1 = T0 + ((uint64_t)T1 << 32);
|
| 1258 |
src2 = env->vfp.scratch[0] + ((uint64_t)env->vfp.scratch[1] << 32); |
| 1259 |
res = src1 + src2; |
| 1260 |
if (((res ^ src1) & SIGNBIT64) && !((src1 ^ src2) & SIGNBIT64)) {
|
| 1261 |
env->QF = 1;
|
| 1262 |
T0 = ~(int64_t)src1 >> 63;
|
| 1263 |
T1 = T0 ^ 0x80000000;
|
| 1264 |
} else {
|
| 1265 |
T0 = res; |
| 1266 |
T1 = res >> 32;
|
| 1267 |
} |
| 1268 |
FORCE_RET(); |
| 1269 |
} |
| 1270 |
|
| 1271 |
NEON_OP(addl_saturate_u64) |
| 1272 |
{
|
| 1273 |
uint64_t src1; |
| 1274 |
uint64_t src2; |
| 1275 |
uint64_t res; |
| 1276 |
|
| 1277 |
src1 = T0 + ((uint64_t)T1 << 32);
|
| 1278 |
src2 = env->vfp.scratch[0] + ((uint64_t)env->vfp.scratch[1] << 32); |
| 1279 |
res = src1 + src2; |
| 1280 |
if (res < src1) {
|
| 1281 |
env->QF = 1;
|
| 1282 |
T0 = 0xffffffff;
|
| 1283 |
T1 = 0xffffffff;
|
| 1284 |
} else {
|
| 1285 |
T0 = res; |
| 1286 |
T1 = res >> 32;
|
| 1287 |
} |
| 1288 |
FORCE_RET(); |
| 1289 |
} |
| 1290 |
|
| 1291 |
NEON_OP(subl_saturate_s64) |
| 1292 |
{
|
| 1293 |
uint64_t src1; |
| 1294 |
uint64_t src2; |
| 1295 |
uint64_t res; |
| 1296 |
|
| 1297 |
src1 = T0 + ((uint64_t)T1 << 32);
|
| 1298 |
src2 = env->vfp.scratch[0] + ((uint64_t)env->vfp.scratch[1] << 32); |
| 1299 |
res = src1 - src2; |
| 1300 |
if (((res ^ src1) & SIGNBIT64) && ((src1 ^ src2) & SIGNBIT64)) {
|
| 1301 |
env->QF = 1;
|
| 1302 |
T0 = ~(int64_t)src1 >> 63;
|
| 1303 |
T1 = T0 ^ 0x80000000;
|
| 1304 |
} else {
|
| 1305 |
T0 = res; |
| 1306 |
T1 = res >> 32;
|
| 1307 |
} |
| 1308 |
FORCE_RET(); |
| 1309 |
} |
| 1310 |
|
| 1311 |
NEON_OP(subl_saturate_u64) |
| 1312 |
{
|
| 1313 |
uint64_t src1; |
| 1314 |
uint64_t src2; |
| 1315 |
uint64_t res; |
| 1316 |
|
| 1317 |
src1 = T0 + ((uint64_t)T1 << 32);
|
| 1318 |
src2 = env->vfp.scratch[0] + ((uint64_t)env->vfp.scratch[1] << 32); |
| 1319 |
if (src1 < src2) {
|
| 1320 |
env->QF = 1;
|
| 1321 |
T0 = 0;
|
| 1322 |
T1 = 0;
|
| 1323 |
} else {
|
| 1324 |
res = src1 - src2; |
| 1325 |
T0 = res; |
| 1326 |
T1 = res >> 32;
|
| 1327 |
} |
| 1328 |
FORCE_RET(); |
| 1329 |
} |
| 1330 |
|
| 1331 |
NEON_OP(negl_u16) |
| 1332 |
{
|
| 1333 |
uint32_t tmp; |
| 1334 |
tmp = T0 >> 16;
|
| 1335 |
tmp = -tmp; |
| 1336 |
T0 = (-T0 & 0xffff) | (tmp << 16); |
| 1337 |
tmp = T1 >> 16;
|
| 1338 |
tmp = -tmp; |
| 1339 |
T1 = (-T1 & 0xffff) | (tmp << 16); |
| 1340 |
FORCE_RET(); |
| 1341 |
} |
| 1342 |
|
| 1343 |
NEON_OP(negl_u32) |
| 1344 |
{
|
| 1345 |
T0 = -T0; |
| 1346 |
T1 = -T1; |
| 1347 |
FORCE_RET(); |
| 1348 |
} |
| 1349 |
|
| 1350 |
NEON_OP(negl_u64) |
| 1351 |
{
|
| 1352 |
uint64_t val; |
| 1353 |
|
| 1354 |
val = T0 | ((uint64_t)T1 << 32);
|
| 1355 |
val = -val; |
| 1356 |
T0 = val; |
| 1357 |
T1 = val >> 32;
|
| 1358 |
FORCE_RET(); |
| 1359 |
} |
| 1360 |
|
| 1361 |
/* Scalar operations. */
|
| 1362 |
NEON_OP(dup_low16) |
| 1363 |
{
|
| 1364 |
T0 = (T0 & 0xffff) | (T0 << 16); |
| 1365 |
FORCE_RET(); |
| 1366 |
} |
| 1367 |
|
| 1368 |
NEON_OP(dup_high16) |
| 1369 |
{
|
| 1370 |
T0 = (T0 >> 16) | (T0 & 0xffff0000); |
| 1371 |
FORCE_RET(); |
| 1372 |
} |
| 1373 |
|
| 1374 |
/* Helper for VEXT */
|
| 1375 |
NEON_OP(extract) |
| 1376 |
{
|
| 1377 |
int shift = PARAM1;
|
| 1378 |
T0 = (T0 >> shift) | (T1 << (32 - shift));
|
| 1379 |
FORCE_RET(); |
| 1380 |
} |
| 1381 |
|
| 1382 |
/* Pairwise add long. Named type is source type. */
|
| 1383 |
NEON_OP(paddl_s8) |
| 1384 |
{
|
| 1385 |
int8_t src1; |
| 1386 |
int8_t src2; |
| 1387 |
uint16_t result; |
| 1388 |
src1 = T0 >> 24;
|
| 1389 |
src2 = T0 >> 16;
|
| 1390 |
result = (uint16_t)src1 + src2; |
| 1391 |
src1 = T0 >> 8;
|
| 1392 |
src2 = T0; |
| 1393 |
T0 = (uint16_t)((uint16_t)src1 + src2) | ((uint32_t)result << 16);
|
| 1394 |
FORCE_RET(); |
| 1395 |
} |
| 1396 |
|
| 1397 |
NEON_OP(paddl_u8) |
| 1398 |
{
|
| 1399 |
uint8_t src1; |
| 1400 |
uint8_t src2; |
| 1401 |
uint16_t result; |
| 1402 |
src1 = T0 >> 24;
|
| 1403 |
src2 = T0 >> 16;
|
| 1404 |
result = (uint16_t)src1 + src2; |
| 1405 |
src1 = T0 >> 8;
|
| 1406 |
src2 = T0; |
| 1407 |
T0 = (uint16_t)((uint16_t)src1 + src2) | ((uint32_t)result << 16);
|
| 1408 |
FORCE_RET(); |
| 1409 |
} |
| 1410 |
|
| 1411 |
NEON_OP(paddl_s16) |
| 1412 |
{
|
| 1413 |
T0 = (uint32_t)(int16_t)T0 + (uint32_t)(int16_t)(T0 >> 16);
|
| 1414 |
FORCE_RET(); |
| 1415 |
} |
| 1416 |
|
| 1417 |
NEON_OP(paddl_u16) |
| 1418 |
{
|
| 1419 |
T0 = (uint32_t)(uint16_t)T0 + (uint32_t)(uint16_t)(T0 >> 16);
|
| 1420 |
FORCE_RET(); |
| 1421 |
} |
| 1422 |
|
| 1423 |
NEON_OP(paddl_s32) |
| 1424 |
{
|
| 1425 |
int64_t tmp; |
| 1426 |
tmp = (int64_t)(int32_t)T0 + (int64_t)(int32_t)T1; |
| 1427 |
T0 = tmp; |
| 1428 |
T1 = tmp >> 32;
|
| 1429 |
FORCE_RET(); |
| 1430 |
} |
| 1431 |
|
| 1432 |
NEON_OP(paddl_u32) |
| 1433 |
{
|
| 1434 |
uint64_t tmp; |
| 1435 |
tmp = (uint64_t)T0 + (uint64_t)T1; |
| 1436 |
T0 = tmp; |
| 1437 |
T1 = tmp >> 32;
|
| 1438 |
FORCE_RET(); |
| 1439 |
} |
| 1440 |
|
| 1441 |
/* Count Leading Sign/Zero Bits. */
|
| 1442 |
static inline int do_clz8(uint8_t x) |
| 1443 |
{
|
| 1444 |
int n;
|
| 1445 |
for (n = 8; x; n--) |
| 1446 |
x >>= 1;
|
| 1447 |
return n;
|
| 1448 |
} |
| 1449 |
|
| 1450 |
static inline int do_clz16(uint16_t x) |
| 1451 |
{
|
| 1452 |
int n;
|
| 1453 |
for (n = 16; x; n--) |
| 1454 |
x >>= 1;
|
| 1455 |
return n;
|
| 1456 |
} |
| 1457 |
|
| 1458 |
NEON_OP(clz_u8) |
| 1459 |
{
|
| 1460 |
uint32_t result; |
| 1461 |
uint32_t tmp; |
| 1462 |
|
| 1463 |
tmp = T0; |
| 1464 |
result = do_clz8(tmp); |
| 1465 |
result |= do_clz8(tmp >> 8) << 8; |
| 1466 |
result |= do_clz8(tmp >> 16) << 16; |
| 1467 |
result |= do_clz8(tmp >> 24) << 24; |
| 1468 |
T0 = result; |
| 1469 |
FORCE_RET(); |
| 1470 |
} |
| 1471 |
|
| 1472 |
NEON_OP(clz_u16) |
| 1473 |
{
|
| 1474 |
uint32_t result; |
| 1475 |
uint32_t tmp; |
| 1476 |
tmp = T0; |
| 1477 |
result = do_clz16(tmp); |
| 1478 |
result |= do_clz16(tmp >> 16) << 16; |
| 1479 |
T0 = result; |
| 1480 |
FORCE_RET(); |
| 1481 |
} |
| 1482 |
|
| 1483 |
NEON_OP(cls_s8) |
| 1484 |
{
|
| 1485 |
uint32_t result; |
| 1486 |
int8_t tmp; |
| 1487 |
tmp = T0; |
| 1488 |
result = do_clz8((tmp < 0) ? ~tmp : tmp) - 1; |
| 1489 |
tmp = T0 >> 8;
|
| 1490 |
result |= (do_clz8((tmp < 0) ? ~tmp : tmp) - 1) << 8; |
| 1491 |
tmp = T0 >> 16;
|
| 1492 |
result |= (do_clz8((tmp < 0) ? ~tmp : tmp) - 1) << 16; |
| 1493 |
tmp = T0 >> 24;
|
| 1494 |
result |= (do_clz8((tmp < 0) ? ~tmp : tmp) - 1) << 24; |
| 1495 |
T0 = result; |
| 1496 |
FORCE_RET(); |
| 1497 |
} |
| 1498 |
|
| 1499 |
NEON_OP(cls_s16) |
| 1500 |
{
|
| 1501 |
uint32_t result; |
| 1502 |
int16_t tmp; |
| 1503 |
tmp = T0; |
| 1504 |
result = do_clz16((tmp < 0) ? ~tmp : tmp) - 1; |
| 1505 |
tmp = T0 >> 16;
|
| 1506 |
result |= (do_clz16((tmp < 0) ? ~tmp : tmp) - 1) << 16; |
| 1507 |
T0 = result; |
| 1508 |
FORCE_RET(); |
| 1509 |
} |
| 1510 |
|
| 1511 |
NEON_OP(cls_s32) |
| 1512 |
{
|
| 1513 |
int count;
|
| 1514 |
if ((int32_t)T0 < 0) |
| 1515 |
T0 = ~T0; |
| 1516 |
for (count = 32; T0 > 0; count--) |
| 1517 |
T0 = T0 >> 1;
|
| 1518 |
T0 = count - 1;
|
| 1519 |
FORCE_RET(); |
| 1520 |
} |
| 1521 |
|
| 1522 |
/* Bit count. */
|
| 1523 |
NEON_OP(cnt_u8) |
| 1524 |
{
|
| 1525 |
T0 = (T0 & 0x55555555) + ((T0 >> 1) & 0x55555555); |
| 1526 |
T0 = (T0 & 0x33333333) + ((T0 >> 2) & 0x33333333); |
| 1527 |
T0 = (T0 & 0x0f0f0f0f) + ((T0 >> 4) & 0x0f0f0f0f); |
| 1528 |
FORCE_RET(); |
| 1529 |
} |
| 1530 |
|
| 1531 |
/* Saturnating negation. */
|
| 1532 |
/* ??? Make these use NEON_VOP1 */
|
| 1533 |
#define DO_QABS8(x) do { \ |
| 1534 |
if (x == (int8_t)0x80) { \ |
| 1535 |
x = 0x7f; \
|
| 1536 |
env->QF = 1; \
|
| 1537 |
} else if (x < 0) { \ |
| 1538 |
x = -x; \ |
| 1539 |
}} while (0) |
| 1540 |
NEON_OP(qabs_s8) |
| 1541 |
{
|
| 1542 |
neon_s8 vec; |
| 1543 |
NEON_UNPACK(neon_s8, vec, T0); |
| 1544 |
DO_QABS8(vec.v1); |
| 1545 |
DO_QABS8(vec.v2); |
| 1546 |
DO_QABS8(vec.v3); |
| 1547 |
DO_QABS8(vec.v4); |
| 1548 |
NEON_PACK(neon_s8, T0, vec); |
| 1549 |
FORCE_RET(); |
| 1550 |
} |
| 1551 |
#undef DO_QABS8
|
| 1552 |
|
| 1553 |
#define DO_QNEG8(x) do { \ |
| 1554 |
if (x == (int8_t)0x80) { \ |
| 1555 |
x = 0x7f; \
|
| 1556 |
env->QF = 1; \
|
| 1557 |
} else { \
|
| 1558 |
x = -x; \ |
| 1559 |
}} while (0) |
| 1560 |
NEON_OP(qneg_s8) |
| 1561 |
{
|
| 1562 |
neon_s8 vec; |
| 1563 |
NEON_UNPACK(neon_s8, vec, T0); |
| 1564 |
DO_QNEG8(vec.v1); |
| 1565 |
DO_QNEG8(vec.v2); |
| 1566 |
DO_QNEG8(vec.v3); |
| 1567 |
DO_QNEG8(vec.v4); |
| 1568 |
NEON_PACK(neon_s8, T0, vec); |
| 1569 |
FORCE_RET(); |
| 1570 |
} |
| 1571 |
#undef DO_QNEG8
|
| 1572 |
|
| 1573 |
#define DO_QABS16(x) do { \ |
| 1574 |
if (x == (int16_t)0x8000) { \ |
| 1575 |
x = 0x7fff; \
|
| 1576 |
env->QF = 1; \
|
| 1577 |
} else if (x < 0) { \ |
| 1578 |
x = -x; \ |
| 1579 |
}} while (0) |
| 1580 |
NEON_OP(qabs_s16) |
| 1581 |
{
|
| 1582 |
neon_s16 vec; |
| 1583 |
NEON_UNPACK(neon_s16, vec, T0); |
| 1584 |
DO_QABS16(vec.v1); |
| 1585 |
DO_QABS16(vec.v2); |
| 1586 |
NEON_PACK(neon_s16, T0, vec); |
| 1587 |
FORCE_RET(); |
| 1588 |
} |
| 1589 |
#undef DO_QABS16
|
| 1590 |
|
| 1591 |
#define DO_QNEG16(x) do { \ |
| 1592 |
if (x == (int16_t)0x8000) { \ |
| 1593 |
x = 0x7fff; \
|
| 1594 |
env->QF = 1; \
|
| 1595 |
} else { \
|
| 1596 |
x = -x; \ |
| 1597 |
}} while (0) |
| 1598 |
NEON_OP(qneg_s16) |
| 1599 |
{
|
| 1600 |
neon_s16 vec; |
| 1601 |
NEON_UNPACK(neon_s16, vec, T0); |
| 1602 |
DO_QNEG16(vec.v1); |
| 1603 |
DO_QNEG16(vec.v2); |
| 1604 |
NEON_PACK(neon_s16, T0, vec); |
| 1605 |
FORCE_RET(); |
| 1606 |
} |
| 1607 |
#undef DO_QNEG16
|
| 1608 |
|
| 1609 |
NEON_OP(qabs_s32) |
| 1610 |
{
|
| 1611 |
if (T0 == 0x80000000) { |
| 1612 |
T0 = 0x7fffffff;
|
| 1613 |
env->QF = 1;
|
| 1614 |
} else if ((int32_t)T0 < 0) { |
| 1615 |
T0 = -T0; |
| 1616 |
} |
| 1617 |
FORCE_RET(); |
| 1618 |
} |
| 1619 |
|
| 1620 |
NEON_OP(qneg_s32) |
| 1621 |
{
|
| 1622 |
if (T0 == 0x80000000) { |
| 1623 |
T0 = 0x7fffffff;
|
| 1624 |
env->QF = 1;
|
| 1625 |
} else {
|
| 1626 |
T0 = -T0; |
| 1627 |
} |
| 1628 |
FORCE_RET(); |
| 1629 |
} |
| 1630 |
|
| 1631 |
/* Unary opperations */
|
| 1632 |
#define NEON_FN(dest, src, dummy) dest = (src < 0) ? -src : src |
| 1633 |
NEON_VOP1(abs_s8, neon_s8, 4)
|
| 1634 |
NEON_VOP1(abs_s16, neon_s16, 2)
|
| 1635 |
NEON_OP(abs_s32) |
| 1636 |
{
|
| 1637 |
if ((int32_t)T0 < 0) |
| 1638 |
T0 = -T0; |
| 1639 |
FORCE_RET(); |
| 1640 |
} |
| 1641 |
#undef NEON_FN
|
| 1642 |
|
| 1643 |
/* Transpose. Argument order is rather strange to avoid special casing
|
| 1644 |
the tranlation code.
|
| 1645 |
On input T0 = rm, T1 = rd. On output T0 = rd, T1 = rm */
|
| 1646 |
NEON_OP(trn_u8) |
| 1647 |
{
|
| 1648 |
uint32_t rd; |
| 1649 |
uint32_t rm; |
| 1650 |
rd = ((T0 & 0x00ff00ff) << 8) | (T1 & 0x00ff00ff); |
| 1651 |
rm = ((T1 & 0xff00ff00) >> 8) | (T0 & 0xff00ff00); |
| 1652 |
T0 = rd; |
| 1653 |
T1 = rm; |
| 1654 |
FORCE_RET(); |
| 1655 |
} |
| 1656 |
|
| 1657 |
NEON_OP(trn_u16) |
| 1658 |
{
|
| 1659 |
uint32_t rd; |
| 1660 |
uint32_t rm; |
| 1661 |
rd = (T0 << 16) | (T1 & 0xffff); |
| 1662 |
rm = (T1 >> 16) | (T0 & 0xffff0000); |
| 1663 |
T0 = rd; |
| 1664 |
T1 = rm; |
| 1665 |
FORCE_RET(); |
| 1666 |
} |
| 1667 |
|
| 1668 |
/* Worker routines for zip and unzip. */
|
| 1669 |
NEON_OP(unzip_u8) |
| 1670 |
{
|
| 1671 |
uint32_t rd; |
| 1672 |
uint32_t rm; |
| 1673 |
rd = (T0 & 0xff) | ((T0 >> 8) & 0xff00) |
| 1674 |
| ((T1 << 16) & 0xff0000) | ((T1 << 8) & 0xff000000); |
| 1675 |
rm = ((T0 >> 8) & 0xff) | ((T0 >> 16) & 0xff00) |
| 1676 |
| ((T1 << 8) & 0xff0000) | (T1 & 0xff000000); |
| 1677 |
T0 = rd; |
| 1678 |
T1 = rm; |
| 1679 |
FORCE_RET(); |
| 1680 |
} |
| 1681 |
|
| 1682 |
NEON_OP(zip_u8) |
| 1683 |
{
|
| 1684 |
uint32_t rd; |
| 1685 |
uint32_t rm; |
| 1686 |
rd = (T0 & 0xff) | ((T1 << 8) & 0xff00) |
| 1687 |
| ((T0 << 16) & 0xff0000) | ((T1 << 24) & 0xff000000); |
| 1688 |
rm = ((T0 >> 16) & 0xff) | ((T1 >> 8) & 0xff00) |
| 1689 |
| ((T0 >> 8) & 0xff0000) | (T1 & 0xff000000); |
| 1690 |
T0 = rd; |
| 1691 |
T1 = rm; |
| 1692 |
FORCE_RET(); |
| 1693 |
} |
| 1694 |
|
| 1695 |
NEON_OP(zip_u16) |
| 1696 |
{
|
| 1697 |
uint32_t tmp; |
| 1698 |
|
| 1699 |
tmp = (T0 & 0xffff) | (T1 << 16); |
| 1700 |
T1 = (T1 & 0xffff0000) | (T0 >> 16); |
| 1701 |
T0 = tmp; |
| 1702 |
FORCE_RET(); |
| 1703 |
} |
| 1704 |
|
| 1705 |
/* Reciprocal/root estimate. */
|
| 1706 |
NEON_OP(recpe_u32) |
| 1707 |
{
|
| 1708 |
T0 = helper_recpe_u32(T0); |
| 1709 |
} |
| 1710 |
|
| 1711 |
NEON_OP(rsqrte_u32) |
| 1712 |
{
|
| 1713 |
T0 = helper_rsqrte_u32(T0); |
| 1714 |
} |
| 1715 |
|
| 1716 |
NEON_OP(recpe_f32) |
| 1717 |
{
|
| 1718 |
FT0s = helper_recpe_f32(FT0s); |
| 1719 |
} |
| 1720 |
|
| 1721 |
NEON_OP(rsqrte_f32) |
| 1722 |
{
|
| 1723 |
FT0s = helper_rsqrte_f32(FT0s); |
| 1724 |
} |
| 1725 |
|
| 1726 |
/* Table lookup. This accessed the register file directly. */
|
| 1727 |
NEON_OP(tbl) |
| 1728 |
{
|
| 1729 |
helper_neon_tbl(PARAM1, PARAM2); |
| 1730 |
} |
| 1731 |
|
| 1732 |
NEON_OP(dup_u8) |
| 1733 |
{
|
| 1734 |
T0 = (T0 >> PARAM1) & 0xff;
|
| 1735 |
T0 |= T0 << 8;
|
| 1736 |
T0 |= T0 << 16;
|
| 1737 |
FORCE_RET(); |
| 1738 |
} |
| 1739 |
|
| 1740 |
/* Helpers for element load/store. */
|
| 1741 |
NEON_OP(insert_elt) |
| 1742 |
{
|
| 1743 |
int shift = PARAM1;
|
| 1744 |
uint32_t mask = PARAM2; |
| 1745 |
T2 = (T2 & mask) | (T0 << shift); |
| 1746 |
FORCE_RET(); |
| 1747 |
} |
| 1748 |
|
| 1749 |
NEON_OP(extract_elt) |
| 1750 |
{
|
| 1751 |
int shift = PARAM1;
|
| 1752 |
uint32_t mask = PARAM2; |
| 1753 |
T0 = (T2 & mask) >> shift; |
| 1754 |
FORCE_RET(); |
| 1755 |
} |