| 1 | /* memset/bzero with unaligned store and rep stosb |
| 2 | Copyright (C) 2016-2017 Free Software Foundation, Inc. |
| 3 | This file is part of the GNU C Library. |
| 4 | |
| 5 | The GNU C Library is free software; you can redistribute it and/or |
| 6 | modify it under the terms of the GNU Lesser General Public |
| 7 | License as published by the Free Software Foundation; either |
| 8 | version 2.1 of the License, or (at your option) any later version. |
| 9 | |
| 10 | The GNU C Library is distributed in the hope that it will be useful, |
| 11 | but WITHOUT ANY WARRANTY; without even the implied warranty of |
| 12 | MERCHANTABILITY or FITNESS FOR A PARTICULAR PURPOSE. See the GNU |
| 13 | Lesser General Public License for more details. |
| 14 | |
| 15 | You should have received a copy of the GNU Lesser General Public |
| 16 | License along with the GNU C Library; if not, see |
| 17 | <http://www.gnu.org/licenses/>. */ |
| 18 | |
| 19 | /* memset is implemented as: |
| 20 | 1. Use overlapping store to avoid branch. |
| 21 | 2. If size is less than VEC, use integer register stores. |
| 22 | 3. If size is from VEC_SIZE to 2 * VEC_SIZE, use 2 VEC stores. |
| 23 | 4. If size is from 2 * VEC_SIZE to 4 * VEC_SIZE, use 4 VEC stores. |
| 24 | 5. If size is more to 4 * VEC_SIZE, align to 4 * VEC_SIZE with |
| 25 | 4 VEC stores and store 4 * VEC at a time until done. */ |
| 26 | |
| 27 | #include <sysdep.h> |
| 28 | |
| 29 | #ifndef MEMSET_CHK_SYMBOL |
| 30 | # define MEMSET_CHK_SYMBOL(p,s) MEMSET_SYMBOL(p, s) |
| 31 | #endif |
| 32 | |
| 33 | #ifndef WMEMSET_CHK_SYMBOL |
| 34 | # define WMEMSET_CHK_SYMBOL(p,s) WMEMSET_SYMBOL(p, s) |
| 35 | #endif |
| 36 | |
| 37 | #ifndef VZEROUPPER |
| 38 | # if VEC_SIZE > 16 |
| 39 | # define VZEROUPPER vzeroupper |
| 40 | # else |
| 41 | # define VZEROUPPER |
| 42 | # endif |
| 43 | #endif |
| 44 | |
| 45 | #ifndef VZEROUPPER_SHORT_RETURN |
| 46 | # if VEC_SIZE > 16 |
| 47 | # define VZEROUPPER_SHORT_RETURN vzeroupper |
| 48 | # else |
| 49 | # define VZEROUPPER_SHORT_RETURN rep |
| 50 | # endif |
| 51 | #endif |
| 52 | |
| 53 | #ifndef MOVQ |
| 54 | # if VEC_SIZE > 16 |
| 55 | # define MOVQ vmovq |
| 56 | # else |
| 57 | # define MOVQ movq |
| 58 | # endif |
| 59 | #endif |
| 60 | |
| 61 | /* Threshold to use Enhanced REP STOSB. Since there is overhead to set |
| 62 | up REP STOSB operation, REP STOSB isn't faster on short data. The |
| 63 | memset micro benchmark in glibc shows that 2KB is the approximate |
| 64 | value above which REP STOSB becomes faster on processors with |
| 65 | Enhanced REP STOSB. Since the stored value is fixed, larger register |
| 66 | size has minimal impact on threshold. */ |
| 67 | #ifndef REP_STOSB_THRESHOLD |
| 68 | # define REP_STOSB_THRESHOLD 2048 |
| 69 | #endif |
| 70 | |
| 71 | #ifndef SECTION |
| 72 | # error SECTION is not defined! |
| 73 | #endif |
| 74 | |
| 75 | .section SECTION(.text),"ax" ,@progbits |
| 76 | #if VEC_SIZE == 16 && IS_IN (libc) |
| 77 | ENTRY (__bzero) |
| 78 | movq %rdi, %rax /* Set return value. */ |
| 79 | movq %rsi, %rdx /* Set n. */ |
| 80 | pxor %xmm0, %xmm0 |
| 81 | jmp L(entry_from_bzero) |
| 82 | END (__bzero) |
| 83 | weak_alias (__bzero, bzero) |
| 84 | #endif |
| 85 | |
| 86 | #if IS_IN (libc) |
| 87 | # if defined SHARED |
| 88 | ENTRY_CHK (WMEMSET_CHK_SYMBOL (__wmemset_chk, unaligned)) |
| 89 | cmpq %rdx, %rcx |
| 90 | jb HIDDEN_JUMPTARGET (__chk_fail) |
| 91 | END_CHK (WMEMSET_CHK_SYMBOL (__wmemset_chk, unaligned)) |
| 92 | # endif |
| 93 | |
| 94 | ENTRY (WMEMSET_SYMBOL (__wmemset, unaligned)) |
| 95 | shlq $2, %rdx |
| 96 | WMEMSET_VDUP_TO_VEC0_AND_SET_RETURN (%esi, %rdi) |
| 97 | jmp L(entry_from_bzero) |
| 98 | END (WMEMSET_SYMBOL (__wmemset, unaligned)) |
| 99 | #endif |
| 100 | |
| 101 | #if defined SHARED && IS_IN (libc) |
| 102 | ENTRY_CHK (MEMSET_CHK_SYMBOL (__memset_chk, unaligned)) |
| 103 | cmpq %rdx, %rcx |
| 104 | jb HIDDEN_JUMPTARGET (__chk_fail) |
| 105 | END_CHK (MEMSET_CHK_SYMBOL (__memset_chk, unaligned)) |
| 106 | #endif |
| 107 | |
| 108 | ENTRY (MEMSET_SYMBOL (__memset, unaligned)) |
| 109 | MEMSET_VDUP_TO_VEC0_AND_SET_RETURN (%esi, %rdi) |
| 110 | L(entry_from_bzero): |
| 111 | cmpq $VEC_SIZE, %rdx |
| 112 | jb L(less_vec) |
| 113 | cmpq $(VEC_SIZE * 2), %rdx |
| 114 | ja L(more_2x_vec) |
| 115 | /* From VEC and to 2 * VEC. No branch when size == VEC_SIZE. */ |
| 116 | VMOVU %VEC(0), -VEC_SIZE(%rdi,%rdx) |
| 117 | VMOVU %VEC(0), (%rdi) |
| 118 | VZEROUPPER |
| 119 | ret |
| 120 | #if defined USE_MULTIARCH && IS_IN (libc) |
| 121 | END (MEMSET_SYMBOL (__memset, unaligned)) |
| 122 | |
| 123 | # if VEC_SIZE == 16 |
| 124 | ENTRY (__memset_chk_erms) |
| 125 | cmpq %rdx, %rcx |
| 126 | jb HIDDEN_JUMPTARGET (__chk_fail) |
| 127 | END (__memset_chk_erms) |
| 128 | |
| 129 | /* Only used to measure performance of REP STOSB. */ |
| 130 | ENTRY (__memset_erms) |
| 131 | # else |
| 132 | /* Provide a symbol to debugger. */ |
| 133 | ENTRY (MEMSET_SYMBOL (__memset, erms)) |
| 134 | # endif |
| 135 | L(stosb): |
| 136 | /* Issue vzeroupper before rep stosb. */ |
| 137 | VZEROUPPER |
| 138 | movq %rdx, %rcx |
| 139 | movzbl %sil, %eax |
| 140 | movq %rdi, %rdx |
| 141 | rep stosb |
| 142 | movq %rdx, %rax |
| 143 | ret |
| 144 | # if VEC_SIZE == 16 |
| 145 | END (__memset_erms) |
| 146 | # else |
| 147 | END (MEMSET_SYMBOL (__memset, erms)) |
| 148 | # endif |
| 149 | |
| 150 | # if defined SHARED && IS_IN (libc) |
| 151 | ENTRY_CHK (MEMSET_CHK_SYMBOL (__memset_chk, unaligned_erms)) |
| 152 | cmpq %rdx, %rcx |
| 153 | jb HIDDEN_JUMPTARGET (__chk_fail) |
| 154 | END_CHK (MEMSET_CHK_SYMBOL (__memset_chk, unaligned_erms)) |
| 155 | # endif |
| 156 | |
| 157 | ENTRY (MEMSET_SYMBOL (__memset, unaligned_erms)) |
| 158 | MEMSET_VDUP_TO_VEC0_AND_SET_RETURN (%esi, %rdi) |
| 159 | cmpq $VEC_SIZE, %rdx |
| 160 | jb L(less_vec) |
| 161 | cmpq $(VEC_SIZE * 2), %rdx |
| 162 | ja L(stosb_more_2x_vec) |
| 163 | /* From VEC and to 2 * VEC. No branch when size == VEC_SIZE. */ |
| 164 | VMOVU %VEC(0), -VEC_SIZE(%rdi,%rdx) |
| 165 | VMOVU %VEC(0), (%rdi) |
| 166 | VZEROUPPER |
| 167 | ret |
| 168 | |
| 169 | L(stosb_more_2x_vec): |
| 170 | cmpq $REP_STOSB_THRESHOLD, %rdx |
| 171 | ja L(stosb) |
| 172 | #endif |
| 173 | L(more_2x_vec): |
| 174 | cmpq $(VEC_SIZE * 4), %rdx |
| 175 | ja L(loop_start) |
| 176 | VMOVU %VEC(0), (%rdi) |
| 177 | VMOVU %VEC(0), VEC_SIZE(%rdi) |
| 178 | VMOVU %VEC(0), -VEC_SIZE(%rdi,%rdx) |
| 179 | VMOVU %VEC(0), -(VEC_SIZE * 2)(%rdi,%rdx) |
| 180 | L(return): |
| 181 | VZEROUPPER |
| 182 | ret |
| 183 | |
| 184 | L(loop_start): |
| 185 | leaq (VEC_SIZE * 4)(%rdi), %rcx |
| 186 | VMOVU %VEC(0), (%rdi) |
| 187 | andq $-(VEC_SIZE * 4), %rcx |
| 188 | VMOVU %VEC(0), -VEC_SIZE(%rdi,%rdx) |
| 189 | VMOVU %VEC(0), VEC_SIZE(%rdi) |
| 190 | VMOVU %VEC(0), -(VEC_SIZE * 2)(%rdi,%rdx) |
| 191 | VMOVU %VEC(0), (VEC_SIZE * 2)(%rdi) |
| 192 | VMOVU %VEC(0), -(VEC_SIZE * 3)(%rdi,%rdx) |
| 193 | VMOVU %VEC(0), (VEC_SIZE * 3)(%rdi) |
| 194 | VMOVU %VEC(0), -(VEC_SIZE * 4)(%rdi,%rdx) |
| 195 | addq %rdi, %rdx |
| 196 | andq $-(VEC_SIZE * 4), %rdx |
| 197 | cmpq %rdx, %rcx |
| 198 | je L(return) |
| 199 | L(loop): |
| 200 | VMOVA %VEC(0), (%rcx) |
| 201 | VMOVA %VEC(0), VEC_SIZE(%rcx) |
| 202 | VMOVA %VEC(0), (VEC_SIZE * 2)(%rcx) |
| 203 | VMOVA %VEC(0), (VEC_SIZE * 3)(%rcx) |
| 204 | addq $(VEC_SIZE * 4), %rcx |
| 205 | cmpq %rcx, %rdx |
| 206 | jne L(loop) |
| 207 | VZEROUPPER_SHORT_RETURN |
| 208 | ret |
| 209 | L(less_vec): |
| 210 | /* Less than 1 VEC. */ |
| 211 | # if VEC_SIZE != 16 && VEC_SIZE != 32 && VEC_SIZE != 64 |
| 212 | # error Unsupported VEC_SIZE! |
| 213 | # endif |
| 214 | # if VEC_SIZE > 32 |
| 215 | cmpb $32, %dl |
| 216 | jae L(between_32_63) |
| 217 | # endif |
| 218 | # if VEC_SIZE > 16 |
| 219 | cmpb $16, %dl |
| 220 | jae L(between_16_31) |
| 221 | # endif |
| 222 | MOVQ %xmm0, %rcx |
| 223 | cmpb $8, %dl |
| 224 | jae L(between_8_15) |
| 225 | cmpb $4, %dl |
| 226 | jae L(between_4_7) |
| 227 | cmpb $1, %dl |
| 228 | ja L(between_2_3) |
| 229 | jb 1f |
| 230 | movb %cl, (%rdi) |
| 231 | 1: |
| 232 | VZEROUPPER |
| 233 | ret |
| 234 | # if VEC_SIZE > 32 |
| 235 | /* From 32 to 63. No branch when size == 32. */ |
| 236 | L(between_32_63): |
| 237 | vmovdqu %ymm0, -32(%rdi,%rdx) |
| 238 | vmovdqu %ymm0, (%rdi) |
| 239 | VZEROUPPER |
| 240 | ret |
| 241 | # endif |
| 242 | # if VEC_SIZE > 16 |
| 243 | /* From 16 to 31. No branch when size == 16. */ |
| 244 | L(between_16_31): |
| 245 | vmovdqu %xmm0, -16(%rdi,%rdx) |
| 246 | vmovdqu %xmm0, (%rdi) |
| 247 | VZEROUPPER |
| 248 | ret |
| 249 | # endif |
| 250 | /* From 8 to 15. No branch when size == 8. */ |
| 251 | L(between_8_15): |
| 252 | movq %rcx, -8(%rdi,%rdx) |
| 253 | movq %rcx, (%rdi) |
| 254 | VZEROUPPER |
| 255 | ret |
| 256 | L(between_4_7): |
| 257 | /* From 4 to 7. No branch when size == 4. */ |
| 258 | movl %ecx, -4(%rdi,%rdx) |
| 259 | movl %ecx, (%rdi) |
| 260 | VZEROUPPER |
| 261 | ret |
| 262 | L(between_2_3): |
| 263 | /* From 2 to 3. No branch when size == 2. */ |
| 264 | movw %cx, -2(%rdi,%rdx) |
| 265 | movw %cx, (%rdi) |
| 266 | VZEROUPPER |
| 267 | ret |
| 268 | END (MEMSET_SYMBOL (__memset, unaligned_erms)) |
| 269 | |