popcountti2.c revision 7f2d7c75e713d778106d01a54e7aef40227bbf2d
137a6a455466e5b197311771a777ab241e471ed8aEdward O'Callaghan/* ===-- popcountti2.c - Implement __popcountti2 ----------------------------===
237a6a455466e5b197311771a777ab241e471ed8aEdward O'Callaghan *
337a6a455466e5b197311771a777ab241e471ed8aEdward O'Callaghan *                     The LLVM Compiler Infrastructure
437a6a455466e5b197311771a777ab241e471ed8aEdward O'Callaghan *
59ad441ffec97db647fee3725b3424284fb913e14Howard Hinnant * This file is dual licensed under the MIT and the University of Illinois Open
69ad441ffec97db647fee3725b3424284fb913e14Howard Hinnant * Source Licenses. See LICENSE.TXT for details.
737a6a455466e5b197311771a777ab241e471ed8aEdward O'Callaghan *
837a6a455466e5b197311771a777ab241e471ed8aEdward O'Callaghan * ===----------------------------------------------------------------------===
937a6a455466e5b197311771a777ab241e471ed8aEdward O'Callaghan *
1037a6a455466e5b197311771a777ab241e471ed8aEdward O'Callaghan * This file implements __popcountti2 for the compiler_rt library.
1137a6a455466e5b197311771a777ab241e471ed8aEdward O'Callaghan *
1237a6a455466e5b197311771a777ab241e471ed8aEdward O'Callaghan * ===----------------------------------------------------------------------===
1337a6a455466e5b197311771a777ab241e471ed8aEdward O'Callaghan */
14b3a6901e66f55b35aa9e01bcb24134e6a65ea004Daniel Dunbar
15b3a6901e66f55b35aa9e01bcb24134e6a65ea004Daniel Dunbar#include "int_lib.h"
16b3a6901e66f55b35aa9e01bcb24134e6a65ea004Daniel Dunbar
177f2d7c75e713d778106d01a54e7aef40227bbf2dChandler Carruth#if __x86_64
187f2d7c75e713d778106d01a54e7aef40227bbf2dChandler Carruth
1937a6a455466e5b197311771a777ab241e471ed8aEdward O'Callaghan/* Returns: count of 1 bits */
20b3a6901e66f55b35aa9e01bcb24134e6a65ea004Daniel Dunbar
21b3a6901e66f55b35aa9e01bcb24134e6a65ea004Daniel Dunbarsi_int
22b3a6901e66f55b35aa9e01bcb24134e6a65ea004Daniel Dunbar__popcountti2(ti_int a)
23b3a6901e66f55b35aa9e01bcb24134e6a65ea004Daniel Dunbar{
24b3a6901e66f55b35aa9e01bcb24134e6a65ea004Daniel Dunbar    tu_int x3 = (tu_int)a;
25b3a6901e66f55b35aa9e01bcb24134e6a65ea004Daniel Dunbar    x3 = x3 - ((x3 >> 1) & (((tu_int)0x5555555555555555uLL << 64) |
26b3a6901e66f55b35aa9e01bcb24134e6a65ea004Daniel Dunbar                                     0x5555555555555555uLL));
2737a6a455466e5b197311771a777ab241e471ed8aEdward O'Callaghan    /* Every 2 bits holds the sum of every pair of bits (64) */
28b3a6901e66f55b35aa9e01bcb24134e6a65ea004Daniel Dunbar    x3 = ((x3 >> 2) & (((tu_int)0x3333333333333333uLL << 64) | 0x3333333333333333uLL))
29b3a6901e66f55b35aa9e01bcb24134e6a65ea004Daniel Dunbar       + (x3 & (((tu_int)0x3333333333333333uLL << 64) | 0x3333333333333333uLL));
3037a6a455466e5b197311771a777ab241e471ed8aEdward O'Callaghan    /* Every 4 bits holds the sum of every 4-set of bits (3 significant bits) (32) */
31b3a6901e66f55b35aa9e01bcb24134e6a65ea004Daniel Dunbar    x3 = (x3 + (x3 >> 4))
32b3a6901e66f55b35aa9e01bcb24134e6a65ea004Daniel Dunbar       & (((tu_int)0x0F0F0F0F0F0F0F0FuLL << 64) | 0x0F0F0F0F0F0F0F0FuLL);
3337a6a455466e5b197311771a777ab241e471ed8aEdward O'Callaghan    /* Every 8 bits holds the sum of every 8-set of bits (4 significant bits) (16) */
34b3a6901e66f55b35aa9e01bcb24134e6a65ea004Daniel Dunbar    du_int x2 = (du_int)(x3 + (x3 >> 64));
3537a6a455466e5b197311771a777ab241e471ed8aEdward O'Callaghan    /* Every 8 bits holds the sum of every 8-set of bits (5 significant bits) (8) */
36b3a6901e66f55b35aa9e01bcb24134e6a65ea004Daniel Dunbar    su_int x = (su_int)(x2 + (x2 >> 32));
3737a6a455466e5b197311771a777ab241e471ed8aEdward O'Callaghan    /* Every 8 bits holds the sum of every 8-set of bits (6 significant bits) (4) */
38b3a6901e66f55b35aa9e01bcb24134e6a65ea004Daniel Dunbar    x = x + (x >> 16);
3937a6a455466e5b197311771a777ab241e471ed8aEdward O'Callaghan    /* Every 8 bits holds the sum of every 8-set of bits (7 significant bits) (2) */
4037a6a455466e5b197311771a777ab241e471ed8aEdward O'Callaghan    /* Upper 16 bits are garbage */
4137a6a455466e5b197311771a777ab241e471ed8aEdward O'Callaghan    return (x + (x >> 8)) & 0xFF;  /* (8 significant bits) */
42b3a6901e66f55b35aa9e01bcb24134e6a65ea004Daniel Dunbar}
43b3a6901e66f55b35aa9e01bcb24134e6a65ea004Daniel Dunbar
44b3a6901e66f55b35aa9e01bcb24134e6a65ea004Daniel Dunbar#endif
45