Nick Desaulniers | bce5a1e | 2022-10-18 10:21:55 -0700 | [diff] [blame] | 1 | /* SPDX-License-Identifier: GPL-2.0 */ |
| 2 | |
Masahiro Yamada | 94ea9c0 | 2023-08-06 23:59:56 +0900 | [diff] [blame] | 3 | #include <linux/export.h> |
Nick Desaulniers | bce5a1e | 2022-10-18 10:21:55 -0700 | [diff] [blame] | 4 | #include <linux/linkage.h> |
Nick Desaulniers | bce5a1e | 2022-10-18 10:21:55 -0700 | [diff] [blame] | 5 | |
| 6 | SYM_FUNC_START(memmove) |
| 7 | /* |
| 8 | * void *memmove(void *dest_in, const void *src_in, size_t n) |
| 9 | * -mregparm=3 passes these in registers: |
| 10 | * dest_in: %eax |
| 11 | * src_in: %edx |
| 12 | * n: %ecx |
| 13 | * See also: arch/x86/entry/calling.h for description of the calling convention. |
| 14 | * |
| 15 | * n can remain in %ecx, but for `rep movsl`, we'll need dest in %edi and src |
| 16 | * in %esi. |
| 17 | */ |
| 18 | .set dest_in, %eax |
| 19 | .set dest, %edi |
| 20 | .set src_in, %edx |
| 21 | .set src, %esi |
| 22 | .set n, %ecx |
| 23 | .set tmp0, %edx |
| 24 | .set tmp0w, %dx |
| 25 | .set tmp1, %ebx |
| 26 | .set tmp1w, %bx |
| 27 | .set tmp2, %eax |
| 28 | .set tmp3b, %cl |
| 29 | |
| 30 | /* |
| 31 | * Save all callee-saved registers, because this function is going to clobber |
| 32 | * all of them: |
| 33 | */ |
| 34 | pushl %ebp |
| 35 | movl %esp, %ebp // set standard frame pointer |
| 36 | |
| 37 | pushl %ebx |
| 38 | pushl %edi |
| 39 | pushl %esi |
| 40 | pushl %eax // save 'dest_in' parameter [eax] as the return value |
| 41 | |
| 42 | movl src_in, src |
| 43 | movl dest_in, dest |
| 44 | |
| 45 | /* Handle more 16 bytes in loop */ |
| 46 | cmpl $0x10, n |
| 47 | jb .Lmove_16B |
| 48 | |
| 49 | /* Decide forward/backward copy mode */ |
| 50 | cmpl dest, src |
| 51 | jb .Lbackwards_header |
| 52 | |
| 53 | /* |
| 54 | * movs instruction have many startup latency |
| 55 | * so we handle small size by general register. |
| 56 | */ |
| 57 | cmpl $680, n |
| 58 | jb .Ltoo_small_forwards |
| 59 | /* movs instruction is only good for aligned case. */ |
| 60 | movl src, tmp0 |
| 61 | xorl dest, tmp0 |
| 62 | andl $0xff, tmp0 |
| 63 | jz .Lforward_movs |
| 64 | .Ltoo_small_forwards: |
| 65 | subl $0x10, n |
| 66 | |
| 67 | /* We gobble 16 bytes forward in each loop. */ |
| 68 | .Lmove_16B_forwards_loop: |
| 69 | subl $0x10, n |
| 70 | movl 0*4(src), tmp0 |
| 71 | movl 1*4(src), tmp1 |
| 72 | movl tmp0, 0*4(dest) |
| 73 | movl tmp1, 1*4(dest) |
| 74 | movl 2*4(src), tmp0 |
| 75 | movl 3*4(src), tmp1 |
| 76 | movl tmp0, 2*4(dest) |
| 77 | movl tmp1, 3*4(dest) |
| 78 | leal 0x10(src), src |
| 79 | leal 0x10(dest), dest |
| 80 | jae .Lmove_16B_forwards_loop |
| 81 | addl $0x10, n |
| 82 | jmp .Lmove_16B |
| 83 | |
| 84 | /* Handle data forward by movs. */ |
| 85 | .p2align 4 |
| 86 | .Lforward_movs: |
| 87 | movl -4(src, n), tmp0 |
| 88 | leal -4(dest, n), tmp1 |
| 89 | shrl $2, n |
| 90 | rep movsl |
| 91 | movl tmp0, (tmp1) |
| 92 | jmp .Ldone |
| 93 | |
| 94 | /* Handle data backward by movs. */ |
| 95 | .p2align 4 |
| 96 | .Lbackwards_movs: |
| 97 | movl (src), tmp0 |
| 98 | movl dest, tmp1 |
| 99 | leal -4(src, n), src |
| 100 | leal -4(dest, n), dest |
| 101 | shrl $2, n |
| 102 | std |
| 103 | rep movsl |
| 104 | movl tmp0,(tmp1) |
| 105 | cld |
| 106 | jmp .Ldone |
| 107 | |
| 108 | /* Start to prepare for backward copy. */ |
| 109 | .p2align 4 |
| 110 | .Lbackwards_header: |
| 111 | cmpl $680, n |
| 112 | jb .Ltoo_small_backwards |
| 113 | movl src, tmp0 |
| 114 | xorl dest, tmp0 |
| 115 | andl $0xff, tmp0 |
| 116 | jz .Lbackwards_movs |
| 117 | |
| 118 | /* Calculate copy position to tail. */ |
| 119 | .Ltoo_small_backwards: |
| 120 | addl n, src |
| 121 | addl n, dest |
| 122 | subl $0x10, n |
| 123 | |
| 124 | /* We gobble 16 bytes backward in each loop. */ |
| 125 | .Lmove_16B_backwards_loop: |
| 126 | subl $0x10, n |
| 127 | |
| 128 | movl -1*4(src), tmp0 |
| 129 | movl -2*4(src), tmp1 |
| 130 | movl tmp0, -1*4(dest) |
| 131 | movl tmp1, -2*4(dest) |
| 132 | movl -3*4(src), tmp0 |
| 133 | movl -4*4(src), tmp1 |
| 134 | movl tmp0, -3*4(dest) |
| 135 | movl tmp1, -4*4(dest) |
| 136 | leal -0x10(src), src |
| 137 | leal -0x10(dest), dest |
| 138 | jae .Lmove_16B_backwards_loop |
| 139 | /* Calculate copy position to head. */ |
| 140 | addl $0x10, n |
| 141 | subl n, src |
| 142 | subl n, dest |
| 143 | |
| 144 | /* Move data from 8 bytes to 15 bytes. */ |
| 145 | .p2align 4 |
| 146 | .Lmove_16B: |
| 147 | cmpl $8, n |
| 148 | jb .Lmove_8B |
| 149 | movl 0*4(src), tmp0 |
| 150 | movl 1*4(src), tmp1 |
| 151 | movl -2*4(src, n), tmp2 |
| 152 | movl -1*4(src, n), src |
| 153 | |
| 154 | movl tmp0, 0*4(dest) |
| 155 | movl tmp1, 1*4(dest) |
| 156 | movl tmp2, -2*4(dest, n) |
| 157 | movl src, -1*4(dest, n) |
| 158 | jmp .Ldone |
| 159 | |
| 160 | /* Move data from 4 bytes to 7 bytes. */ |
| 161 | .p2align 4 |
| 162 | .Lmove_8B: |
| 163 | cmpl $4, n |
| 164 | jb .Lmove_4B |
| 165 | movl 0*4(src), tmp0 |
| 166 | movl -1*4(src, n), tmp1 |
| 167 | movl tmp0, 0*4(dest) |
| 168 | movl tmp1, -1*4(dest, n) |
| 169 | jmp .Ldone |
| 170 | |
| 171 | /* Move data from 2 bytes to 3 bytes. */ |
| 172 | .p2align 4 |
| 173 | .Lmove_4B: |
| 174 | cmpl $2, n |
| 175 | jb .Lmove_1B |
| 176 | movw 0*2(src), tmp0w |
| 177 | movw -1*2(src, n), tmp1w |
| 178 | movw tmp0w, 0*2(dest) |
| 179 | movw tmp1w, -1*2(dest, n) |
| 180 | jmp .Ldone |
| 181 | |
| 182 | /* Move data for 1 byte. */ |
| 183 | .p2align 4 |
| 184 | .Lmove_1B: |
| 185 | cmpl $1, n |
| 186 | jb .Ldone |
| 187 | movb (src), tmp3b |
| 188 | movb tmp3b, (dest) |
| 189 | .p2align 4 |
| 190 | .Ldone: |
| 191 | popl dest_in // restore 'dest_in' [eax] as the return value |
| 192 | /* Restore all callee-saved registers: */ |
| 193 | popl %esi |
| 194 | popl %edi |
| 195 | popl %ebx |
| 196 | popl %ebp |
| 197 | |
| 198 | RET |
| 199 | SYM_FUNC_END(memmove) |
| 200 | EXPORT_SYMBOL(memmove) |