Commit d42f183c0452a9c769b91e3911566566724983ff
1 parent
a1f6684d
Implement TCG rotation ops for x86-64
Signed-off-by: Aurelien Jarno <aurelien@aurel32.net> git-svn-id: svn://svn.savannah.nongnu.org/qemu/trunk@6795 c046a42c-6fe2-441c-8c8c-71466251a162
Showing
4 changed files
with
56 additions
and
4 deletions
tcg/tcg-op.h
| @@ -1573,6 +1573,9 @@ static inline void tcg_gen_orc_i64(TCGv_i64 ret, TCGv_i64 arg1, TCGv_i64 arg2) | @@ -1573,6 +1573,9 @@ static inline void tcg_gen_orc_i64(TCGv_i64 ret, TCGv_i64 arg1, TCGv_i64 arg2) | ||
| 1573 | 1573 | ||
| 1574 | static inline void tcg_gen_rotl_i32(TCGv_i32 ret, TCGv_i32 arg1, TCGv_i32 arg2) | 1574 | static inline void tcg_gen_rotl_i32(TCGv_i32 ret, TCGv_i32 arg1, TCGv_i32 arg2) |
| 1575 | { | 1575 | { |
| 1576 | +#ifdef TCG_TARGET_HAS_rot_i32 | ||
| 1577 | + tcg_gen_op3_i32(INDEX_op_rotl_i32, ret, arg1, arg2); | ||
| 1578 | +#else | ||
| 1576 | TCGv_i32 t0, t1; | 1579 | TCGv_i32 t0, t1; |
| 1577 | 1580 | ||
| 1578 | t0 = tcg_temp_new_i32(); | 1581 | t0 = tcg_temp_new_i32(); |
| @@ -1583,10 +1586,14 @@ static inline void tcg_gen_rotl_i32(TCGv_i32 ret, TCGv_i32 arg1, TCGv_i32 arg2) | @@ -1583,10 +1586,14 @@ static inline void tcg_gen_rotl_i32(TCGv_i32 ret, TCGv_i32 arg1, TCGv_i32 arg2) | ||
| 1583 | tcg_gen_or_i32(ret, t0, t1); | 1586 | tcg_gen_or_i32(ret, t0, t1); |
| 1584 | tcg_temp_free_i32(t0); | 1587 | tcg_temp_free_i32(t0); |
| 1585 | tcg_temp_free_i32(t1); | 1588 | tcg_temp_free_i32(t1); |
| 1589 | +#endif | ||
| 1586 | } | 1590 | } |
| 1587 | 1591 | ||
| 1588 | static inline void tcg_gen_rotl_i64(TCGv_i64 ret, TCGv_i64 arg1, TCGv_i64 arg2) | 1592 | static inline void tcg_gen_rotl_i64(TCGv_i64 ret, TCGv_i64 arg1, TCGv_i64 arg2) |
| 1589 | { | 1593 | { |
| 1594 | +#ifdef TCG_TARGET_HAS_rot_i64 | ||
| 1595 | + tcg_gen_op3_i64(INDEX_op_rotl_i64, ret, arg1, arg2); | ||
| 1596 | +#else | ||
| 1590 | TCGv_i64 t0, t1; | 1597 | TCGv_i64 t0, t1; |
| 1591 | 1598 | ||
| 1592 | t0 = tcg_temp_new_i64(); | 1599 | t0 = tcg_temp_new_i64(); |
| @@ -1597,6 +1604,7 @@ static inline void tcg_gen_rotl_i64(TCGv_i64 ret, TCGv_i64 arg1, TCGv_i64 arg2) | @@ -1597,6 +1604,7 @@ static inline void tcg_gen_rotl_i64(TCGv_i64 ret, TCGv_i64 arg1, TCGv_i64 arg2) | ||
| 1597 | tcg_gen_or_i64(ret, t0, t1); | 1604 | tcg_gen_or_i64(ret, t0, t1); |
| 1598 | tcg_temp_free_i64(t0); | 1605 | tcg_temp_free_i64(t0); |
| 1599 | tcg_temp_free_i64(t1); | 1606 | tcg_temp_free_i64(t1); |
| 1607 | +#endif | ||
| 1600 | } | 1608 | } |
| 1601 | 1609 | ||
| 1602 | static inline void tcg_gen_rotli_i32(TCGv_i32 ret, TCGv_i32 arg1, int32_t arg2) | 1610 | static inline void tcg_gen_rotli_i32(TCGv_i32 ret, TCGv_i32 arg1, int32_t arg2) |
| @@ -1605,6 +1613,11 @@ static inline void tcg_gen_rotli_i32(TCGv_i32 ret, TCGv_i32 arg1, int32_t arg2) | @@ -1605,6 +1613,11 @@ static inline void tcg_gen_rotli_i32(TCGv_i32 ret, TCGv_i32 arg1, int32_t arg2) | ||
| 1605 | if (arg2 == 0) { | 1613 | if (arg2 == 0) { |
| 1606 | tcg_gen_mov_i32(ret, arg1); | 1614 | tcg_gen_mov_i32(ret, arg1); |
| 1607 | } else { | 1615 | } else { |
| 1616 | +#ifdef TCG_TARGET_HAS_rot_i32 | ||
| 1617 | + TCGv_i32 t0 = tcg_const_i32(arg2); | ||
| 1618 | + tcg_gen_rotl_i32(ret, arg1, t0); | ||
| 1619 | + tcg_temp_free_i32(t0); | ||
| 1620 | +#else | ||
| 1608 | TCGv_i32 t0, t1; | 1621 | TCGv_i32 t0, t1; |
| 1609 | t0 = tcg_temp_new_i32(); | 1622 | t0 = tcg_temp_new_i32(); |
| 1610 | t1 = tcg_temp_new_i32(); | 1623 | t1 = tcg_temp_new_i32(); |
| @@ -1613,6 +1626,7 @@ static inline void tcg_gen_rotli_i32(TCGv_i32 ret, TCGv_i32 arg1, int32_t arg2) | @@ -1613,6 +1626,7 @@ static inline void tcg_gen_rotli_i32(TCGv_i32 ret, TCGv_i32 arg1, int32_t arg2) | ||
| 1613 | tcg_gen_or_i32(ret, t0, t1); | 1626 | tcg_gen_or_i32(ret, t0, t1); |
| 1614 | tcg_temp_free_i32(t0); | 1627 | tcg_temp_free_i32(t0); |
| 1615 | tcg_temp_free_i32(t1); | 1628 | tcg_temp_free_i32(t1); |
| 1629 | +#endif | ||
| 1616 | } | 1630 | } |
| 1617 | } | 1631 | } |
| 1618 | 1632 | ||
| @@ -1622,6 +1636,11 @@ static inline void tcg_gen_rotli_i64(TCGv_i64 ret, TCGv_i64 arg1, int64_t arg2) | @@ -1622,6 +1636,11 @@ static inline void tcg_gen_rotli_i64(TCGv_i64 ret, TCGv_i64 arg1, int64_t arg2) | ||
| 1622 | if (arg2 == 0) { | 1636 | if (arg2 == 0) { |
| 1623 | tcg_gen_mov_i64(ret, arg1); | 1637 | tcg_gen_mov_i64(ret, arg1); |
| 1624 | } else { | 1638 | } else { |
| 1639 | +#ifdef TCG_TARGET_HAS_rot_i64 | ||
| 1640 | + TCGv_i64 t0 = tcg_const_i64(arg2); | ||
| 1641 | + tcg_gen_rotl_i64(ret, arg1, t0); | ||
| 1642 | + tcg_temp_free_i64(t0); | ||
| 1643 | +#else | ||
| 1625 | TCGv_i64 t0, t1; | 1644 | TCGv_i64 t0, t1; |
| 1626 | t0 = tcg_temp_new_i64(); | 1645 | t0 = tcg_temp_new_i64(); |
| 1627 | t1 = tcg_temp_new_i64(); | 1646 | t1 = tcg_temp_new_i64(); |
| @@ -1630,11 +1649,15 @@ static inline void tcg_gen_rotli_i64(TCGv_i64 ret, TCGv_i64 arg1, int64_t arg2) | @@ -1630,11 +1649,15 @@ static inline void tcg_gen_rotli_i64(TCGv_i64 ret, TCGv_i64 arg1, int64_t arg2) | ||
| 1630 | tcg_gen_or_i64(ret, t0, t1); | 1649 | tcg_gen_or_i64(ret, t0, t1); |
| 1631 | tcg_temp_free_i64(t0); | 1650 | tcg_temp_free_i64(t0); |
| 1632 | tcg_temp_free_i64(t1); | 1651 | tcg_temp_free_i64(t1); |
| 1652 | +#endif | ||
| 1633 | } | 1653 | } |
| 1634 | } | 1654 | } |
| 1635 | 1655 | ||
| 1636 | static inline void tcg_gen_rotr_i32(TCGv_i32 ret, TCGv_i32 arg1, TCGv_i32 arg2) | 1656 | static inline void tcg_gen_rotr_i32(TCGv_i32 ret, TCGv_i32 arg1, TCGv_i32 arg2) |
| 1637 | { | 1657 | { |
| 1658 | +#ifdef TCG_TARGET_HAS_rot_i32 | ||
| 1659 | + tcg_gen_op3_i32(INDEX_op_rotr_i32, ret, arg1, arg2); | ||
| 1660 | +#else | ||
| 1638 | TCGv_i32 t0, t1; | 1661 | TCGv_i32 t0, t1; |
| 1639 | 1662 | ||
| 1640 | t0 = tcg_temp_new_i32(); | 1663 | t0 = tcg_temp_new_i32(); |
| @@ -1645,10 +1668,14 @@ static inline void tcg_gen_rotr_i32(TCGv_i32 ret, TCGv_i32 arg1, TCGv_i32 arg2) | @@ -1645,10 +1668,14 @@ static inline void tcg_gen_rotr_i32(TCGv_i32 ret, TCGv_i32 arg1, TCGv_i32 arg2) | ||
| 1645 | tcg_gen_or_i32(ret, t0, t1); | 1668 | tcg_gen_or_i32(ret, t0, t1); |
| 1646 | tcg_temp_free_i32(t0); | 1669 | tcg_temp_free_i32(t0); |
| 1647 | tcg_temp_free_i32(t1); | 1670 | tcg_temp_free_i32(t1); |
| 1671 | +#endif | ||
| 1648 | } | 1672 | } |
| 1649 | 1673 | ||
| 1650 | static inline void tcg_gen_rotr_i64(TCGv_i64 ret, TCGv_i64 arg1, TCGv_i64 arg2) | 1674 | static inline void tcg_gen_rotr_i64(TCGv_i64 ret, TCGv_i64 arg1, TCGv_i64 arg2) |
| 1651 | { | 1675 | { |
| 1676 | +#ifdef TCG_TARGET_HAS_rot_i64 | ||
| 1677 | + tcg_gen_op3_i64(INDEX_op_rotr_i64, ret, arg1, arg2); | ||
| 1678 | +#else | ||
| 1652 | TCGv_i64 t0, t1; | 1679 | TCGv_i64 t0, t1; |
| 1653 | 1680 | ||
| 1654 | t0 = tcg_temp_new_i64(); | 1681 | t0 = tcg_temp_new_i64(); |
| @@ -1659,6 +1686,7 @@ static inline void tcg_gen_rotr_i64(TCGv_i64 ret, TCGv_i64 arg1, TCGv_i64 arg2) | @@ -1659,6 +1686,7 @@ static inline void tcg_gen_rotr_i64(TCGv_i64 ret, TCGv_i64 arg1, TCGv_i64 arg2) | ||
| 1659 | tcg_gen_or_i64(ret, t0, t1); | 1686 | tcg_gen_or_i64(ret, t0, t1); |
| 1660 | tcg_temp_free_i64(t0); | 1687 | tcg_temp_free_i64(t0); |
| 1661 | tcg_temp_free_i64(t1); | 1688 | tcg_temp_free_i64(t1); |
| 1689 | +#endif | ||
| 1662 | } | 1690 | } |
| 1663 | 1691 | ||
| 1664 | static inline void tcg_gen_rotri_i32(TCGv_i32 ret, TCGv_i32 arg1, int32_t arg2) | 1692 | static inline void tcg_gen_rotri_i32(TCGv_i32 ret, TCGv_i32 arg1, int32_t arg2) |
tcg/tcg-opc.h
| @@ -67,10 +67,12 @@ DEF2(divu2_i32, 2, 3, 0, 0) | @@ -67,10 +67,12 @@ DEF2(divu2_i32, 2, 3, 0, 0) | ||
| 67 | DEF2(and_i32, 1, 2, 0, 0) | 67 | DEF2(and_i32, 1, 2, 0, 0) |
| 68 | DEF2(or_i32, 1, 2, 0, 0) | 68 | DEF2(or_i32, 1, 2, 0, 0) |
| 69 | DEF2(xor_i32, 1, 2, 0, 0) | 69 | DEF2(xor_i32, 1, 2, 0, 0) |
| 70 | -/* shifts */ | 70 | +/* shifts/rotates */ |
| 71 | DEF2(shl_i32, 1, 2, 0, 0) | 71 | DEF2(shl_i32, 1, 2, 0, 0) |
| 72 | DEF2(shr_i32, 1, 2, 0, 0) | 72 | DEF2(shr_i32, 1, 2, 0, 0) |
| 73 | DEF2(sar_i32, 1, 2, 0, 0) | 73 | DEF2(sar_i32, 1, 2, 0, 0) |
| 74 | +DEF2(rotl_i32, 1, 2, 0, 0) | ||
| 75 | +DEF2(rotr_i32, 1, 2, 0, 0) | ||
| 74 | 76 | ||
| 75 | DEF2(brcond_i32, 0, 2, 2, TCG_OPF_BB_END | TCG_OPF_SIDE_EFFECTS) | 77 | DEF2(brcond_i32, 0, 2, 2, TCG_OPF_BB_END | TCG_OPF_SIDE_EFFECTS) |
| 76 | #if TCG_TARGET_REG_BITS == 32 | 78 | #if TCG_TARGET_REG_BITS == 32 |
| @@ -120,10 +122,12 @@ DEF2(divu2_i64, 2, 3, 0, 0) | @@ -120,10 +122,12 @@ DEF2(divu2_i64, 2, 3, 0, 0) | ||
| 120 | DEF2(and_i64, 1, 2, 0, 0) | 122 | DEF2(and_i64, 1, 2, 0, 0) |
| 121 | DEF2(or_i64, 1, 2, 0, 0) | 123 | DEF2(or_i64, 1, 2, 0, 0) |
| 122 | DEF2(xor_i64, 1, 2, 0, 0) | 124 | DEF2(xor_i64, 1, 2, 0, 0) |
| 123 | -/* shifts */ | 125 | +/* shifts/rotates */ |
| 124 | DEF2(shl_i64, 1, 2, 0, 0) | 126 | DEF2(shl_i64, 1, 2, 0, 0) |
| 125 | DEF2(shr_i64, 1, 2, 0, 0) | 127 | DEF2(shr_i64, 1, 2, 0, 0) |
| 126 | DEF2(sar_i64, 1, 2, 0, 0) | 128 | DEF2(sar_i64, 1, 2, 0, 0) |
| 129 | +DEF2(rotl_i64, 1, 2, 0, 0) | ||
| 130 | +DEF2(rotr_i64, 1, 2, 0, 0) | ||
| 127 | 131 | ||
| 128 | DEF2(brcond_i64, 0, 2, 2, TCG_OPF_BB_END | TCG_OPF_SIDE_EFFECTS) | 132 | DEF2(brcond_i64, 0, 2, 2, TCG_OPF_BB_END | TCG_OPF_SIDE_EFFECTS) |
| 129 | #ifdef TCG_TARGET_HAS_ext8s_i64 | 133 | #ifdef TCG_TARGET_HAS_ext8s_i64 |
tcg/x86_64/tcg-target.c
| @@ -194,6 +194,8 @@ static inline int tcg_target_const_match(tcg_target_long val, | @@ -194,6 +194,8 @@ static inline int tcg_target_const_match(tcg_target_long val, | ||
| 194 | #define ARITH_XOR 6 | 194 | #define ARITH_XOR 6 |
| 195 | #define ARITH_CMP 7 | 195 | #define ARITH_CMP 7 |
| 196 | 196 | ||
| 197 | +#define SHIFT_ROL 0 | ||
| 198 | +#define SHIFT_ROR 1 | ||
| 197 | #define SHIFT_SHL 4 | 199 | #define SHIFT_SHL 4 |
| 198 | #define SHIFT_SHR 5 | 200 | #define SHIFT_SHR 5 |
| 199 | #define SHIFT_SAR 7 | 201 | #define SHIFT_SAR 7 |
| @@ -1049,7 +1051,13 @@ static inline void tcg_out_op(TCGContext *s, int opc, const TCGArg *args, | @@ -1049,7 +1051,13 @@ static inline void tcg_out_op(TCGContext *s, int opc, const TCGArg *args, | ||
| 1049 | case INDEX_op_sar_i32: | 1051 | case INDEX_op_sar_i32: |
| 1050 | c = SHIFT_SAR; | 1052 | c = SHIFT_SAR; |
| 1051 | goto gen_shift32; | 1053 | goto gen_shift32; |
| 1052 | - | 1054 | + case INDEX_op_rotl_i32: |
| 1055 | + c = SHIFT_ROL; | ||
| 1056 | + goto gen_shift32; | ||
| 1057 | + case INDEX_op_rotr_i32: | ||
| 1058 | + c = SHIFT_ROR; | ||
| 1059 | + goto gen_shift32; | ||
| 1060 | + | ||
| 1053 | case INDEX_op_shl_i64: | 1061 | case INDEX_op_shl_i64: |
| 1054 | c = SHIFT_SHL; | 1062 | c = SHIFT_SHL; |
| 1055 | gen_shift64: | 1063 | gen_shift64: |
| @@ -1070,7 +1078,13 @@ static inline void tcg_out_op(TCGContext *s, int opc, const TCGArg *args, | @@ -1070,7 +1078,13 @@ static inline void tcg_out_op(TCGContext *s, int opc, const TCGArg *args, | ||
| 1070 | case INDEX_op_sar_i64: | 1078 | case INDEX_op_sar_i64: |
| 1071 | c = SHIFT_SAR; | 1079 | c = SHIFT_SAR; |
| 1072 | goto gen_shift64; | 1080 | goto gen_shift64; |
| 1073 | - | 1081 | + case INDEX_op_rotl_i64: |
| 1082 | + c = SHIFT_ROL; | ||
| 1083 | + goto gen_shift64; | ||
| 1084 | + case INDEX_op_rotr_i64: | ||
| 1085 | + c = SHIFT_ROR; | ||
| 1086 | + goto gen_shift64; | ||
| 1087 | + | ||
| 1074 | case INDEX_op_brcond_i32: | 1088 | case INDEX_op_brcond_i32: |
| 1075 | tcg_out_brcond(s, args[2], args[0], args[1], const_args[1], | 1089 | tcg_out_brcond(s, args[2], args[0], args[1], const_args[1], |
| 1076 | args[3], 0); | 1090 | args[3], 0); |
| @@ -1230,6 +1244,8 @@ static const TCGTargetOpDef x86_64_op_defs[] = { | @@ -1230,6 +1244,8 @@ static const TCGTargetOpDef x86_64_op_defs[] = { | ||
| 1230 | { INDEX_op_shl_i32, { "r", "0", "ci" } }, | 1244 | { INDEX_op_shl_i32, { "r", "0", "ci" } }, |
| 1231 | { INDEX_op_shr_i32, { "r", "0", "ci" } }, | 1245 | { INDEX_op_shr_i32, { "r", "0", "ci" } }, |
| 1232 | { INDEX_op_sar_i32, { "r", "0", "ci" } }, | 1246 | { INDEX_op_sar_i32, { "r", "0", "ci" } }, |
| 1247 | + { INDEX_op_rotl_i32, { "r", "0", "ci" } }, | ||
| 1248 | + { INDEX_op_rotr_i32, { "r", "0", "ci" } }, | ||
| 1233 | 1249 | ||
| 1234 | { INDEX_op_brcond_i32, { "r", "ri" } }, | 1250 | { INDEX_op_brcond_i32, { "r", "ri" } }, |
| 1235 | 1251 | ||
| @@ -1259,6 +1275,8 @@ static const TCGTargetOpDef x86_64_op_defs[] = { | @@ -1259,6 +1275,8 @@ static const TCGTargetOpDef x86_64_op_defs[] = { | ||
| 1259 | { INDEX_op_shl_i64, { "r", "0", "ci" } }, | 1275 | { INDEX_op_shl_i64, { "r", "0", "ci" } }, |
| 1260 | { INDEX_op_shr_i64, { "r", "0", "ci" } }, | 1276 | { INDEX_op_shr_i64, { "r", "0", "ci" } }, |
| 1261 | { INDEX_op_sar_i64, { "r", "0", "ci" } }, | 1277 | { INDEX_op_sar_i64, { "r", "0", "ci" } }, |
| 1278 | + { INDEX_op_rotl_i64, { "r", "0", "ci" } }, | ||
| 1279 | + { INDEX_op_rotr_i64, { "r", "0", "ci" } }, | ||
| 1262 | 1280 | ||
| 1263 | { INDEX_op_brcond_i64, { "r", "re" } }, | 1281 | { INDEX_op_brcond_i64, { "r", "re" } }, |
| 1264 | 1282 |
tcg/x86_64/tcg-target.h
| @@ -65,6 +65,8 @@ enum { | @@ -65,6 +65,8 @@ enum { | ||
| 65 | #define TCG_TARGET_HAS_ext8s_i64 | 65 | #define TCG_TARGET_HAS_ext8s_i64 |
| 66 | #define TCG_TARGET_HAS_ext16s_i64 | 66 | #define TCG_TARGET_HAS_ext16s_i64 |
| 67 | #define TCG_TARGET_HAS_ext32s_i64 | 67 | #define TCG_TARGET_HAS_ext32s_i64 |
| 68 | +#define TCG_TARGET_HAS_rot_i32 | ||
| 69 | +#define TCG_TARGET_HAS_rot_i64 | ||
| 68 | 70 | ||
| 69 | /* Note: must be synced with dyngen-exec.h */ | 71 | /* Note: must be synced with dyngen-exec.h */ |
| 70 | #define TCG_AREG0 TCG_REG_R14 | 72 | #define TCG_AREG0 TCG_REG_R14 |