| Borislav Petkov | f596710 | 2016-05-30 12:56:27 +0200 | [diff] [blame] | 1 | #include <linux/linkage.h> |
| 2 | |
| 3 | #include <asm/asm.h> |
| 4 | |
| 5 | /* |
| 6 | * unsigned int __sw_hweight32(unsigned int w) |
| 7 | * %rdi: w |
| 8 | */ |
| 9 | ENTRY(__sw_hweight32) |
| 10 | |
| 11 | #ifdef CONFIG_X86_64 |
| 12 | movl %edi, %eax # w |
| 13 | #endif |
| 14 | __ASM_SIZE(push,) %__ASM_REG(dx) |
| 15 | movl %eax, %edx # w -> t |
| 16 | shrl %edx # t >>= 1 |
| 17 | andl $0x55555555, %edx # t &= 0x55555555 |
| 18 | subl %edx, %eax # w -= t |
| 19 | |
| 20 | movl %eax, %edx # w -> t |
| 21 | shrl $2, %eax # w_tmp >>= 2 |
| 22 | andl $0x33333333, %edx # t &= 0x33333333 |
| 23 | andl $0x33333333, %eax # w_tmp &= 0x33333333 |
| 24 | addl %edx, %eax # w = w_tmp + t |
| 25 | |
| 26 | movl %eax, %edx # w -> t |
| 27 | shrl $4, %edx # t >>= 4 |
| 28 | addl %edx, %eax # w_tmp += t |
| 29 | andl $0x0f0f0f0f, %eax # w_tmp &= 0x0f0f0f0f |
| 30 | imull $0x01010101, %eax, %eax # w_tmp *= 0x01010101 |
| 31 | shrl $24, %eax # w = w_tmp >> 24 |
| 32 | __ASM_SIZE(pop,) %__ASM_REG(dx) |
| 33 | ret |
| 34 | ENDPROC(__sw_hweight32) |
| 35 | |
| 36 | ENTRY(__sw_hweight64) |
| 37 | #ifdef CONFIG_X86_64 |
| 38 | pushq %rdx |
| 39 | |
| 40 | movq %rdi, %rdx # w -> t |
| 41 | movabsq $0x5555555555555555, %rax |
| 42 | shrq %rdx # t >>= 1 |
| 43 | andq %rdx, %rax # t &= 0x5555555555555555 |
| 44 | movabsq $0x3333333333333333, %rdx |
| 45 | subq %rax, %rdi # w -= t |
| 46 | |
| 47 | movq %rdi, %rax # w -> t |
| 48 | shrq $2, %rdi # w_tmp >>= 2 |
| 49 | andq %rdx, %rax # t &= 0x3333333333333333 |
| 50 | andq %rdi, %rdx # w_tmp &= 0x3333333333333333 |
| 51 | addq %rdx, %rax # w = w_tmp + t |
| 52 | |
| 53 | movq %rax, %rdx # w -> t |
| 54 | shrq $4, %rdx # t >>= 4 |
| 55 | addq %rdx, %rax # w_tmp += t |
| 56 | movabsq $0x0f0f0f0f0f0f0f0f, %rdx |
| 57 | andq %rdx, %rax # w_tmp &= 0x0f0f0f0f0f0f0f0f |
| 58 | movabsq $0x0101010101010101, %rdx |
| 59 | imulq %rdx, %rax # w_tmp *= 0x0101010101010101 |
| 60 | shrq $56, %rax # w = w_tmp >> 56 |
| 61 | |
| 62 | popq %rdx |
| 63 | ret |
| 64 | #else /* CONFIG_X86_32 */ |
| 65 | /* We're getting an u64 arg in (%eax,%edx): unsigned long hweight64(__u64 w) */ |
| 66 | pushl %ecx |
| 67 | |
| 68 | call __sw_hweight32 |
| 69 | movl %eax, %ecx # stash away result |
| 70 | movl %edx, %eax # second part of input |
| 71 | call __sw_hweight32 |
| 72 | addl %ecx, %eax # result |
| 73 | |
| 74 | popl %ecx |
| 75 | ret |
| 76 | #endif |
| 77 | ENDPROC(__sw_hweight64) |