arch/x86/include/asm/arch_hweight.h

   1 #ifndef _ASM_X86_HWEIGHT_H
   2 #define _ASM_X86_HWEIGHT_H
   3
   4 #ifdef CONFIG_64BIT
   5 /* popcnt %edi, %eax -- redundant REX prefix for alignment */
   6 #define POPCNT32 ".byte 0xf3,0x40,0x0f,0xb8,0xc7"
   7 /* popcnt %rdi, %rax */
   8 #define POPCNT64 ".byte 0xf3,0x48,0x0f,0xb8,0xc7"
   9 #define REG_IN "D"
  10 #define REG_OUT "a"
  11 #else
  12 /* popcnt %eax, %eax */
  13 #define POPCNT32 ".byte 0xf3,0x0f,0xb8,0xc0"
  14 #define REG_IN "a"
  15 #define REG_OUT "a"
  16 #endif
  17
  18 /*
  19  * __sw_hweightXX are called from within the alternatives below
  20  * and callee-clobbered registers need to be taken care of. See
  21  * ARCH_HWEIGHT_CFLAGS in <arch/x86/Kconfig> for the respective
  22  * compiler switches.
  23  */
  24 static __always_inline unsigned int __arch_hweight32(unsigned int w)
  25 {
  26         unsigned int res = 0;
  27
  28         asm (ALTERNATIVE("call __sw_hweight32", POPCNT32, X86_FEATURE_POPCNT)
  29                      : "="REG_OUT (res)
  30                      : REG_IN (w));
  31
  32         return res;
  33 }
  34
  35 static inline unsigned int __arch_hweight16(unsigned int w)
  36 {
  37         return __arch_hweight32(w & 0xffff);
  38 }
  39
  40 static inline unsigned int __arch_hweight8(unsigned int w)
  41 {
  42         return __arch_hweight32(w & 0xff);
  43 }
  44
  45 #ifdef CONFIG_X86_32
  46 static inline unsigned long __arch_hweight64(__u64 w)
  47 {
  48         return  __arch_hweight32((u32)w) +
  49                 __arch_hweight32((u32)(w >> 32));
  50 }
  51 #else
  52 static __always_inline unsigned long __arch_hweight64(__u64 w)
  53 {
  54         unsigned long res = 0;
  55
  56         asm (ALTERNATIVE("call __sw_hweight64", POPCNT64, X86_FEATURE_POPCNT)
  57                      : "="REG_OUT (res)
  58                      : REG_IN (w));
  59
  60         return res;
  61 }
  62 #endif /* CONFIG_X86_32 */
  63
  64 #endif