Revision a35f3ec7 target-i386/ops_sse.h
| b/target-i386/ops_sse.h | ||
|---|---|---|
| 1 | 1 |
/* |
| 2 |
* MMX/SSE/SSE2/PNI support
|
|
| 2 |
* MMX/3DNow!/SSE/SSE2/SSE3/PNI support
|
|
| 3 | 3 |
* |
| 4 | 4 |
* Copyright (c) 2005 Fabrice Bellard |
| 5 | 5 |
* |
| ... | ... | |
| 409 | 409 |
#define FCMPEQ(a, b) (a) == (b) ? -1 : 0 |
| 410 | 410 |
|
| 411 | 411 |
#define FMULLW(a, b) (a) * (b) |
| 412 |
#define FMULHRW(a, b) ((int16_t)(a) * (int16_t)(b) + 0x8000) >> 16 |
|
| 412 | 413 |
#define FMULHUW(a, b) (a) * (b) >> 16 |
| 413 | 414 |
#define FMULHW(a, b) (int16_t)(a) * (int16_t)(b) >> 16 |
| 414 | 415 |
|
| ... | ... | |
| 455 | 456 |
SSE_OP_L(op_pcmpeql, FCMPEQ) |
| 456 | 457 |
|
| 457 | 458 |
SSE_OP_W(op_pmullw, FMULLW) |
| 459 |
#if SHIFT == 0 |
|
| 460 |
SSE_OP_W(op_pmulhrw, FMULHRW) |
|
| 461 |
#endif |
|
| 458 | 462 |
SSE_OP_W(op_pmulhuw, FMULHUW) |
| 459 | 463 |
SSE_OP_W(op_pmulhw, FMULHW) |
| 460 | 464 |
|
| ... | ... | |
| 1383 | 1387 |
UNPCK_OP(l, 0) |
| 1384 | 1388 |
UNPCK_OP(h, 1) |
| 1385 | 1389 |
|
| 1390 |
/* 3DNow! float ops */ |
|
| 1391 |
#if SHIFT == 0 |
|
| 1392 |
void OPPROTO op_pi2fd(void) |
|
| 1393 |
{
|
|
| 1394 |
MMXReg *d = (MMXReg *)((char *)env + PARAM1); |
|
| 1395 |
MMXReg *s = (MMXReg *)((char *)env + PARAM2); |
|
| 1396 |
d->MMX_S(0) = int32_to_float32(s->MMX_L(0), &env->mmx_status); |
|
| 1397 |
d->MMX_S(1) = int32_to_float32(s->MMX_L(1), &env->mmx_status); |
|
| 1398 |
} |
|
| 1399 |
|
|
| 1400 |
void OPPROTO op_pi2fw(void) |
|
| 1401 |
{
|
|
| 1402 |
MMXReg *d = (MMXReg *)((char *)env + PARAM1); |
|
| 1403 |
MMXReg *s = (MMXReg *)((char *)env + PARAM2); |
|
| 1404 |
d->MMX_S(0) = int32_to_float32((int16_t)s->MMX_W(0), &env->mmx_status); |
|
| 1405 |
d->MMX_S(1) = int32_to_float32((int16_t)s->MMX_W(2), &env->mmx_status); |
|
| 1406 |
} |
|
| 1407 |
|
|
| 1408 |
void OPPROTO op_pf2id(void) |
|
| 1409 |
{
|
|
| 1410 |
MMXReg *d = (MMXReg *)((char *)env + PARAM1); |
|
| 1411 |
MMXReg *s = (MMXReg *)((char *)env + PARAM2); |
|
| 1412 |
d->MMX_L(0) = float32_to_int32_round_to_zero(s->MMX_S(0), &env->mmx_status); |
|
| 1413 |
d->MMX_L(1) = float32_to_int32_round_to_zero(s->MMX_S(1), &env->mmx_status); |
|
| 1414 |
} |
|
| 1415 |
|
|
| 1416 |
void OPPROTO op_pf2iw(void) |
|
| 1417 |
{
|
|
| 1418 |
MMXReg *d = (MMXReg *)((char *)env + PARAM1); |
|
| 1419 |
MMXReg *s = (MMXReg *)((char *)env + PARAM2); |
|
| 1420 |
d->MMX_L(0) = satsw(float32_to_int32_round_to_zero(s->MMX_S(0), &env->mmx_status)); |
|
| 1421 |
d->MMX_L(1) = satsw(float32_to_int32_round_to_zero(s->MMX_S(1), &env->mmx_status)); |
|
| 1422 |
} |
|
| 1423 |
|
|
| 1424 |
void OPPROTO op_pfacc(void) |
|
| 1425 |
{
|
|
| 1426 |
MMXReg *d = (MMXReg *)((char *)env + PARAM1); |
|
| 1427 |
MMXReg *s = (MMXReg *)((char *)env + PARAM2); |
|
| 1428 |
MMXReg r; |
|
| 1429 |
r.MMX_S(0) = float32_add(d->MMX_S(0), d->MMX_S(1), &env->mmx_status); |
|
| 1430 |
r.MMX_S(1) = float32_add(s->MMX_S(0), s->MMX_S(1), &env->mmx_status); |
|
| 1431 |
*d = r; |
|
| 1432 |
} |
|
| 1433 |
|
|
| 1434 |
void OPPROTO op_pfadd(void) |
|
| 1435 |
{
|
|
| 1436 |
MMXReg *d = (MMXReg *)((char *)env + PARAM1); |
|
| 1437 |
MMXReg *s = (MMXReg *)((char *)env + PARAM2); |
|
| 1438 |
d->MMX_S(0) = float32_add(d->MMX_S(0), s->MMX_S(0), &env->mmx_status); |
|
| 1439 |
d->MMX_S(1) = float32_add(d->MMX_S(1), s->MMX_S(1), &env->mmx_status); |
|
| 1440 |
} |
|
| 1441 |
|
|
| 1442 |
void OPPROTO op_pfcmpeq(void) |
|
| 1443 |
{
|
|
| 1444 |
MMXReg *d = (MMXReg *)((char *)env + PARAM1); |
|
| 1445 |
MMXReg *s = (MMXReg *)((char *)env + PARAM2); |
|
| 1446 |
d->MMX_L(0) = float32_eq(d->MMX_S(0), s->MMX_S(0), &env->mmx_status) ? -1 : 0; |
|
| 1447 |
d->MMX_L(1) = float32_eq(d->MMX_S(1), s->MMX_S(1), &env->mmx_status) ? -1 : 0; |
|
| 1448 |
} |
|
| 1449 |
|
|
| 1450 |
void OPPROTO op_pfcmpge(void) |
|
| 1451 |
{
|
|
| 1452 |
MMXReg *d = (MMXReg *)((char *)env + PARAM1); |
|
| 1453 |
MMXReg *s = (MMXReg *)((char *)env + PARAM2); |
|
| 1454 |
d->MMX_L(0) = float32_le(s->MMX_S(0), d->MMX_S(0), &env->mmx_status) ? -1 : 0; |
|
| 1455 |
d->MMX_L(1) = float32_le(s->MMX_S(1), d->MMX_S(1), &env->mmx_status) ? -1 : 0; |
|
| 1456 |
} |
|
| 1457 |
|
|
| 1458 |
void OPPROTO op_pfcmpgt(void) |
|
| 1459 |
{
|
|
| 1460 |
MMXReg *d = (MMXReg *)((char *)env + PARAM1); |
|
| 1461 |
MMXReg *s = (MMXReg *)((char *)env + PARAM2); |
|
| 1462 |
d->MMX_L(0) = float32_lt(s->MMX_S(0), d->MMX_S(0), &env->mmx_status) ? -1 : 0; |
|
| 1463 |
d->MMX_L(1) = float32_lt(s->MMX_S(1), d->MMX_S(1), &env->mmx_status) ? -1 : 0; |
|
| 1464 |
} |
|
| 1465 |
|
|
| 1466 |
void OPPROTO op_pfmax(void) |
|
| 1467 |
{
|
|
| 1468 |
MMXReg *d = (MMXReg *)((char *)env + PARAM1); |
|
| 1469 |
MMXReg *s = (MMXReg *)((char *)env + PARAM2); |
|
| 1470 |
if (float32_lt(d->MMX_S(0), s->MMX_S(0), &env->mmx_status)) |
|
| 1471 |
d->MMX_S(0) = s->MMX_S(0); |
|
| 1472 |
if (float32_lt(d->MMX_S(1), s->MMX_S(1), &env->mmx_status)) |
|
| 1473 |
d->MMX_S(1) = s->MMX_S(1); |
|
| 1474 |
} |
|
| 1475 |
|
|
| 1476 |
void OPPROTO op_pfmin(void) |
|
| 1477 |
{
|
|
| 1478 |
MMXReg *d = (MMXReg *)((char *)env + PARAM1); |
|
| 1479 |
MMXReg *s = (MMXReg *)((char *)env + PARAM2); |
|
| 1480 |
if (float32_lt(s->MMX_S(0), d->MMX_S(0), &env->mmx_status)) |
|
| 1481 |
d->MMX_S(0) = s->MMX_S(0); |
|
| 1482 |
if (float32_lt(s->MMX_S(1), d->MMX_S(1), &env->mmx_status)) |
|
| 1483 |
d->MMX_S(1) = s->MMX_S(1); |
|
| 1484 |
} |
|
| 1485 |
|
|
| 1486 |
void OPPROTO op_pfmul(void) |
|
| 1487 |
{
|
|
| 1488 |
MMXReg *d = (MMXReg *)((char *)env + PARAM1); |
|
| 1489 |
MMXReg *s = (MMXReg *)((char *)env + PARAM2); |
|
| 1490 |
d->MMX_S(0) = float32_mul(d->MMX_S(0), s->MMX_S(0), &env->mmx_status); |
|
| 1491 |
d->MMX_S(1) = float32_mul(d->MMX_S(1), s->MMX_S(1), &env->mmx_status); |
|
| 1492 |
} |
|
| 1493 |
|
|
| 1494 |
void OPPROTO op_pfnacc(void) |
|
| 1495 |
{
|
|
| 1496 |
MMXReg *d = (MMXReg *)((char *)env + PARAM1); |
|
| 1497 |
MMXReg *s = (MMXReg *)((char *)env + PARAM2); |
|
| 1498 |
MMXReg r; |
|
| 1499 |
r.MMX_S(0) = float32_sub(d->MMX_S(0), d->MMX_S(1), &env->mmx_status); |
|
| 1500 |
r.MMX_S(1) = float32_sub(s->MMX_S(0), s->MMX_S(1), &env->mmx_status); |
|
| 1501 |
*d = r; |
|
| 1502 |
} |
|
| 1503 |
|
|
| 1504 |
void OPPROTO op_pfpnacc(void) |
|
| 1505 |
{
|
|
| 1506 |
MMXReg *d = (MMXReg *)((char *)env + PARAM1); |
|
| 1507 |
MMXReg *s = (MMXReg *)((char *)env + PARAM2); |
|
| 1508 |
MMXReg r; |
|
| 1509 |
r.MMX_S(0) = float32_sub(d->MMX_S(0), d->MMX_S(1), &env->mmx_status); |
|
| 1510 |
r.MMX_S(1) = float32_add(s->MMX_S(0), s->MMX_S(1), &env->mmx_status); |
|
| 1511 |
*d = r; |
|
| 1512 |
} |
|
| 1513 |
|
|
| 1514 |
void OPPROTO op_pfrcp(void) |
|
| 1515 |
{
|
|
| 1516 |
MMXReg *d = (MMXReg *)((char *)env + PARAM1); |
|
| 1517 |
MMXReg *s = (MMXReg *)((char *)env + PARAM2); |
|
| 1518 |
d->MMX_S(0) = approx_rcp(s->MMX_S(0)); |
|
| 1519 |
d->MMX_S(1) = d->MMX_S(0); |
|
| 1520 |
} |
|
| 1521 |
|
|
| 1522 |
void OPPROTO op_pfrsqrt(void) |
|
| 1523 |
{
|
|
| 1524 |
MMXReg *d = (MMXReg *)((char *)env + PARAM1); |
|
| 1525 |
MMXReg *s = (MMXReg *)((char *)env + PARAM2); |
|
| 1526 |
d->MMX_L(1) = s->MMX_L(0) & 0x7fffffff; |
|
| 1527 |
d->MMX_S(1) = approx_rsqrt(d->MMX_S(1)); |
|
| 1528 |
d->MMX_L(1) |= s->MMX_L(0) & 0x80000000; |
|
| 1529 |
d->MMX_L(0) = d->MMX_L(1); |
|
| 1530 |
} |
|
| 1531 |
|
|
| 1532 |
void OPPROTO op_pfsub(void) |
|
| 1533 |
{
|
|
| 1534 |
MMXReg *d = (MMXReg *)((char *)env + PARAM1); |
|
| 1535 |
MMXReg *s = (MMXReg *)((char *)env + PARAM2); |
|
| 1536 |
d->MMX_S(0) = float32_sub(d->MMX_S(0), s->MMX_S(0), &env->mmx_status); |
|
| 1537 |
d->MMX_S(1) = float32_sub(d->MMX_S(1), s->MMX_S(1), &env->mmx_status); |
|
| 1538 |
} |
|
| 1539 |
|
|
| 1540 |
void OPPROTO op_pfsubr(void) |
|
| 1541 |
{
|
|
| 1542 |
MMXReg *d = (MMXReg *)((char *)env + PARAM1); |
|
| 1543 |
MMXReg *s = (MMXReg *)((char *)env + PARAM2); |
|
| 1544 |
d->MMX_S(0) = float32_sub(s->MMX_S(0), d->MMX_S(0), &env->mmx_status); |
|
| 1545 |
d->MMX_S(1) = float32_sub(s->MMX_S(1), d->MMX_S(1), &env->mmx_status); |
|
| 1546 |
} |
|
| 1547 |
|
|
| 1548 |
void OPPROTO op_pswapd(void) |
|
| 1549 |
{
|
|
| 1550 |
MMXReg *d = (MMXReg *)((char *)env + PARAM1); |
|
| 1551 |
MMXReg *s = (MMXReg *)((char *)env + PARAM2); |
|
| 1552 |
MMXReg r; |
|
| 1553 |
r.MMX_L(0) = s->MMX_L(1); |
|
| 1554 |
r.MMX_L(1) = s->MMX_L(0); |
|
| 1555 |
*d = r; |
|
| 1556 |
} |
|
| 1557 |
#endif |
|
| 1558 |
|
|
| 1386 | 1559 |
#undef SHIFT |
| 1387 | 1560 |
#undef XMM_ONLY |
| 1388 | 1561 |
#undef Reg |
Also available in: Unified diff