Details | Last modification | View Log | RSS feed
| Rev | Author | Line No. | Line |
|---|---|---|---|
| 33 | pmbaty | 1 | /* |
| 2 | X86-64 inline functions for MSB(), LSB() and PopCnt(). Note |
||
| 3 | that these are 64 bit functions and they use 64 bit (quad-word) |
||
| 4 | X86-64 instructions. |
||
| 5 | */ |
||
| 6 | /* *INDENT-OFF* */ |
||
| 7 | #if defined(UNIX) |
||
| 8 | static __inline__ int MSB(uint64_t word) { |
||
| 9 | uint64_t dummy, dummy2; |
||
| 10 | |||
| 108 | pmbaty | 11 | asm(" bsrq %1, %0 " "\n\t" |
| 33 | pmbaty | 12 | : "=&r"(dummy), "=&r" (dummy2) |
| 13 | : "1"((uint64_t) (word)) |
||
| 14 | : "cc"); |
||
| 15 | return (dummy); |
||
| 16 | } |
||
| 17 | static __inline__ int LSB(uint64_t word) { |
||
| 18 | uint64_t dummy, dummy2; |
||
| 19 | |||
| 108 | pmbaty | 20 | asm(" bsfq %1, %0 " "\n\t" |
| 33 | pmbaty | 21 | : "=&r"(dummy), "=&r" (dummy2) |
| 22 | : "1"((uint64_t) (word)) |
||
| 23 | : "cc"); |
||
| 24 | return (dummy); |
||
| 25 | } |
||
| 26 | #if defined(POPCNT) |
||
| 27 | static __inline__ int PopCnt(uint64_t word) { |
||
| 28 | uint64_t dummy, dummy2; |
||
| 29 | |||
| 108 | pmbaty | 30 | asm(" popcnt %1, %0 " "\n\t" |
| 33 | pmbaty | 31 | : "=&r"(dummy), "=&r" (dummy2) |
| 32 | : "1"((uint64_t) (word)) |
||
| 33 | : "cc"); |
||
| 34 | return (dummy); |
||
| 35 | } |
||
| 36 | #else |
||
| 37 | static __inline__ int PopCnt(uint64_t word) { |
||
| 38 | uint64_t dummy, dummy2, dummy3; |
||
| 39 | |||
| 108 | pmbaty | 40 | asm(" xorq %0, %0 " "\n\t" |
| 41 | " testq %1, %1 " "\n\t" |
||
| 42 | " jz 2f " "\n\t" |
||
| 33 | pmbaty | 43 | "1: leaq -1(%1),%2 " "\n\t" |
| 108 | pmbaty | 44 | " incq %0 " "\n\t" |
| 45 | " andq %2, %1 " "\n\t" |
||
| 46 | " jnz 1b " "\n\t" |
||
| 47 | "2: " "\n\t" |
||
| 33 | pmbaty | 48 | : "=&r"(dummy), "=&r"(dummy2), "=&r"(dummy3) |
| 49 | : "1"((uint64_t) (word)) |
||
| 50 | : "cc"); |
||
| 51 | return (dummy); |
||
| 52 | } |
||
| 53 | #endif |
||
| 54 | #else |
||
| 55 | #if defined(POPCNT) |
||
| 56 | #include <nmmintrin.h> |
||
| 57 | __forceinline int PopCnt(uint64_t a) { |
||
| 58 | return _mm_popcnt_u64(a); |
||
| 59 | } |
||
| 60 | #else |
||
| 61 | |||
| 62 | __forceinline int PopCnt(uint64_t a) { |
||
| 63 | int c = 1; |
||
| 64 | |||
| 108 | pmbaty | 65 | if (!a) |
| 66 | return 0; |
||
| 33 | pmbaty | 67 | while (a &= a - 1) |
| 68 | c++; |
||
| 69 | return c; |
||
| 70 | } |
||
| 71 | #endif |
||
| 72 | |||
| 73 | __forceinline int MSB(uint64_t a) { |
||
| 74 | int v; |
||
| 75 | |||
| 76 | _BitScanReverse64(&v, a); |
||
| 77 | return v; |
||
| 78 | } |
||
| 79 | |||
| 80 | __forceinline int LSB(uint64_t a) { |
||
| 81 | int v; |
||
| 82 | |||
| 83 | _BitScanForward64(&v, a); |
||
| 84 | return v; |
||
| 85 | } |
||
| 86 | #endif |