arch/powerpc/lib/hweight_64.S

   1 /*
   2  * This program is free software; you can redistribute it and/or modify
   3  * it under the terms of the GNU General Public License as published by
   4  * the Free Software Foundation; either version 2 of the License, or
   5  * (at your option) any later version.
   6  *
   7  * This program is distributed in the hope that it will be useful,
   8  * but WITHOUT ANY WARRANTY; without even the implied warranty of
   9  * MERCHANTABILITY or FITNESS FOR A PARTICULAR PURPOSE.  See the
  10  * GNU General Public License for more details.
  11  *
  12  * You should have received a copy of the GNU General Public License
  13  * along with this program; if not, write to the Free Software
  14  * Foundation, Inc., 59 Temple Place - Suite 330, Boston, MA 02111-1307, USA.
  15  *
  16  * Copyright (C) IBM Corporation, 2010
  17  *
  18  * Author: Anton Blanchard <anton@au.ibm.com>
  19  */
  20 #include <asm/processor.h>
  21 #include <asm/ppc_asm.h>
  22 #include <asm/export.h>
  23 #include <asm/feature-fixups.h>
  24
  25 /* Note: This code relies on -mminimal-toc */
  26
  27 _GLOBAL(__arch_hweight8)
  28 BEGIN_FTR_SECTION
  29         b __sw_hweight8
  30         nop
  31         nop
  32 FTR_SECTION_ELSE
  33         PPC_POPCNTB(R3,R3)
  34         clrldi  r3,r3,64-8
  35         blr
  36 ALT_FTR_SECTION_END_IFCLR(CPU_FTR_POPCNTB)
  37 EXPORT_SYMBOL(__arch_hweight8)
  38
  39 _GLOBAL(__arch_hweight16)
  40 BEGIN_FTR_SECTION
  41         b __sw_hweight16
  42         nop
  43         nop
  44         nop
  45         nop
  46 FTR_SECTION_ELSE
  47   BEGIN_FTR_SECTION_NESTED(50)
  48         PPC_POPCNTB(R3,R3)
  49         srdi    r4,r3,8
  50         add     r3,r4,r3
  51         clrldi  r3,r3,64-8
  52         blr
  53   FTR_SECTION_ELSE_NESTED(50)
  54         clrlwi  r3,r3,16
  55         PPC_POPCNTW(R3,R3)
  56         clrldi  r3,r3,64-8
  57         blr
  58   ALT_FTR_SECTION_END_NESTED_IFCLR(CPU_FTR_POPCNTD, 50)
  59 ALT_FTR_SECTION_END_IFCLR(CPU_FTR_POPCNTB)
  60 EXPORT_SYMBOL(__arch_hweight16)
  61
  62 _GLOBAL(__arch_hweight32)
  63 BEGIN_FTR_SECTION
  64         b __sw_hweight32
  65         nop
  66         nop
  67         nop
  68         nop
  69         nop
  70         nop
  71 FTR_SECTION_ELSE
  72   BEGIN_FTR_SECTION_NESTED(51)
  73         PPC_POPCNTB(R3,R3)
  74         srdi    r4,r3,16
  75         add     r3,r4,r3
  76         srdi    r4,r3,8
  77         add     r3,r4,r3
  78         clrldi  r3,r3,64-8
  79         blr
  80   FTR_SECTION_ELSE_NESTED(51)
  81         PPC_POPCNTW(R3,R3)
  82         clrldi  r3,r3,64-8
  83         blr
  84   ALT_FTR_SECTION_END_NESTED_IFCLR(CPU_FTR_POPCNTD, 51)
  85 ALT_FTR_SECTION_END_IFCLR(CPU_FTR_POPCNTB)
  86 EXPORT_SYMBOL(__arch_hweight32)
  87
  88 _GLOBAL(__arch_hweight64)
  89 BEGIN_FTR_SECTION
  90         b __sw_hweight64
  91         nop
  92         nop
  93         nop
  94         nop
  95         nop
  96         nop
  97         nop
  98         nop
  99 FTR_SECTION_ELSE
 100   BEGIN_FTR_SECTION_NESTED(52)
 101         PPC_POPCNTB(R3,R3)
 102         srdi    r4,r3,32
 103         add     r3,r4,r3
 104         srdi    r4,r3,16
 105         add     r3,r4,r3
 106         srdi    r4,r3,8
 107         add     r3,r4,r3
 108         clrldi  r3,r3,64-8
 109         blr
 110   FTR_SECTION_ELSE_NESTED(52)
 111         PPC_POPCNTD(R3,R3)
 112         clrldi  r3,r3,64-8
 113         blr
 114   ALT_FTR_SECTION_END_NESTED_IFCLR(CPU_FTR_POPCNTD, 52)
 115 ALT_FTR_SECTION_END_IFCLR(CPU_FTR_POPCNTB)
 116 EXPORT_SYMBOL(__arch_hweight64)