162306a36Sopenharmony_ci/* SPDX-License-Identifier: GPL-2.0 */ 262306a36Sopenharmony_ci#include <linux/linkage.h> 362306a36Sopenharmony_ci#include <asm/export.h> 462306a36Sopenharmony_ci 562306a36Sopenharmony_ci#include <asm/asm.h> 662306a36Sopenharmony_ci 762306a36Sopenharmony_ci/* 862306a36Sopenharmony_ci * unsigned int __sw_hweight32(unsigned int w) 962306a36Sopenharmony_ci * %rdi: w 1062306a36Sopenharmony_ci */ 1162306a36Sopenharmony_ciSYM_FUNC_START(__sw_hweight32) 1262306a36Sopenharmony_ci 1362306a36Sopenharmony_ci#ifdef CONFIG_X86_64 1462306a36Sopenharmony_ci movl %edi, %eax # w 1562306a36Sopenharmony_ci#endif 1662306a36Sopenharmony_ci __ASM_SIZE(push,) %__ASM_REG(dx) 1762306a36Sopenharmony_ci movl %eax, %edx # w -> t 1862306a36Sopenharmony_ci shrl %edx # t >>= 1 1962306a36Sopenharmony_ci andl $0x55555555, %edx # t &= 0x55555555 2062306a36Sopenharmony_ci subl %edx, %eax # w -= t 2162306a36Sopenharmony_ci 2262306a36Sopenharmony_ci movl %eax, %edx # w -> t 2362306a36Sopenharmony_ci shrl $2, %eax # w_tmp >>= 2 2462306a36Sopenharmony_ci andl $0x33333333, %edx # t &= 0x33333333 2562306a36Sopenharmony_ci andl $0x33333333, %eax # w_tmp &= 0x33333333 2662306a36Sopenharmony_ci addl %edx, %eax # w = w_tmp + t 2762306a36Sopenharmony_ci 2862306a36Sopenharmony_ci movl %eax, %edx # w -> t 2962306a36Sopenharmony_ci shrl $4, %edx # t >>= 4 3062306a36Sopenharmony_ci addl %edx, %eax # w_tmp += t 3162306a36Sopenharmony_ci andl $0x0f0f0f0f, %eax # w_tmp &= 0x0f0f0f0f 3262306a36Sopenharmony_ci imull $0x01010101, %eax, %eax # w_tmp *= 0x01010101 3362306a36Sopenharmony_ci shrl $24, %eax # w = w_tmp >> 24 3462306a36Sopenharmony_ci __ASM_SIZE(pop,) %__ASM_REG(dx) 3562306a36Sopenharmony_ci RET 3662306a36Sopenharmony_ciSYM_FUNC_END(__sw_hweight32) 3762306a36Sopenharmony_ciEXPORT_SYMBOL(__sw_hweight32) 3862306a36Sopenharmony_ci 3962306a36Sopenharmony_ciSYM_FUNC_START(__sw_hweight64) 4062306a36Sopenharmony_ci#ifdef CONFIG_X86_64 4162306a36Sopenharmony_ci pushq %rdi 4262306a36Sopenharmony_ci pushq %rdx 4362306a36Sopenharmony_ci 4462306a36Sopenharmony_ci movq %rdi, %rdx # w -> t 4562306a36Sopenharmony_ci movabsq $0x5555555555555555, %rax 4662306a36Sopenharmony_ci shrq %rdx # t >>= 1 4762306a36Sopenharmony_ci andq %rdx, %rax # t &= 0x5555555555555555 4862306a36Sopenharmony_ci movabsq $0x3333333333333333, %rdx 4962306a36Sopenharmony_ci subq %rax, %rdi # w -= t 5062306a36Sopenharmony_ci 5162306a36Sopenharmony_ci movq %rdi, %rax # w -> t 5262306a36Sopenharmony_ci shrq $2, %rdi # w_tmp >>= 2 5362306a36Sopenharmony_ci andq %rdx, %rax # t &= 0x3333333333333333 5462306a36Sopenharmony_ci andq %rdi, %rdx # w_tmp &= 0x3333333333333333 5562306a36Sopenharmony_ci addq %rdx, %rax # w = w_tmp + t 5662306a36Sopenharmony_ci 5762306a36Sopenharmony_ci movq %rax, %rdx # w -> t 5862306a36Sopenharmony_ci shrq $4, %rdx # t >>= 4 5962306a36Sopenharmony_ci addq %rdx, %rax # w_tmp += t 6062306a36Sopenharmony_ci movabsq $0x0f0f0f0f0f0f0f0f, %rdx 6162306a36Sopenharmony_ci andq %rdx, %rax # w_tmp &= 0x0f0f0f0f0f0f0f0f 6262306a36Sopenharmony_ci movabsq $0x0101010101010101, %rdx 6362306a36Sopenharmony_ci imulq %rdx, %rax # w_tmp *= 0x0101010101010101 6462306a36Sopenharmony_ci shrq $56, %rax # w = w_tmp >> 56 6562306a36Sopenharmony_ci 6662306a36Sopenharmony_ci popq %rdx 6762306a36Sopenharmony_ci popq %rdi 6862306a36Sopenharmony_ci RET 6962306a36Sopenharmony_ci#else /* CONFIG_X86_32 */ 7062306a36Sopenharmony_ci /* We're getting an u64 arg in (%eax,%edx): unsigned long hweight64(__u64 w) */ 7162306a36Sopenharmony_ci pushl %ecx 7262306a36Sopenharmony_ci 7362306a36Sopenharmony_ci call __sw_hweight32 7462306a36Sopenharmony_ci movl %eax, %ecx # stash away result 7562306a36Sopenharmony_ci movl %edx, %eax # second part of input 7662306a36Sopenharmony_ci call __sw_hweight32 7762306a36Sopenharmony_ci addl %ecx, %eax # result 7862306a36Sopenharmony_ci 7962306a36Sopenharmony_ci popl %ecx 8062306a36Sopenharmony_ci RET 8162306a36Sopenharmony_ci#endif 8262306a36Sopenharmony_ciSYM_FUNC_END(__sw_hweight64) 8362306a36Sopenharmony_ciEXPORT_SYMBOL(__sw_hweight64) 84