memcpy_neon.S revision 1.3
1 /*- 2 * Copyright (c) 2013 The NetBSD Foundation, Inc. 3 * All rights reserved. 4 * 5 * This code is derived from software contributed to The NetBSD Foundation 6 * by Matt Thomas of 3am Software Foundry. 7 * 8 * Redistribution and use in source and binary forms, with or without 9 * modification, are permitted provided that the following conditions 10 * are met: 11 * 1. Redistributions of source code must retain the above copyright 12 * notice, this list of conditions and the following disclaimer. 13 * 2. Redistributions in binary form must reproduce the above copyright 14 * notice, this list of conditions and the following disclaimer in the 15 * documentation and/or other materials provided with the distribution. 16 * 17 * THIS SOFTWARE IS PROVIDED BY THE NETBSD FOUNDATION, INC. AND CONTRIBUTORS 18 * ``AS IS'' AND ANY EXPRESS OR IMPLIED WARRANTIES, INCLUDING, BUT NOT LIMITED 19 * TO, THE IMPLIED WARRANTIES OF MERCHANTABILITY AND FITNESS FOR A PARTICULAR 20 * PURPOSE ARE DISCLAIMED. IN NO EVENT SHALL THE FOUNDATION OR CONTRIBUTORS 21 * BE LIABLE FOR ANY DIRECT, INDIRECT, INCIDENTAL, SPECIAL, EXEMPLARY, OR 22 * CONSEQUENTIAL DAMAGES (INCLUDING, BUT NOT LIMITED TO, PROCUREMENT OF 23 * SUBSTITUTE GOODS OR SERVICES; LOSS OF USE, DATA, OR PROFITS; OR BUSINESS 24 * INTERRUPTION) HOWEVER CAUSED AND ON ANY THEORY OF LIABILITY, WHETHER IN 25 * CONTRACT, STRICT LIABILITY, OR TORT (INCLUDING NEGLIGENCE OR OTHERWISE) 26 * ARISING IN ANY WAY OUT OF THE USE OF THIS SOFTWARE, EVEN IF ADVISED OF THE 27 * POSSIBILITY OF SUCH DAMAGE. 28 */ 29 30 #include <machine/asm.h> 31 32 RCSID("$NetBSD: memcpy_neon.S,v 1.3 2025/02/27 08:39:53 andvar Exp $") 33 34 .text 35 ENTRY(memcpy) 36 teq r2, #0 /* 0 length? */ 37 cmpne r0, r1 /* if not, does src == dst? */ 38 RETc(eq) /* yes, (to either) return */ 39 40 mov r3, r0 /* keep r0 unchanged */ 41 #if 0 42 cmp r2, #16 /* copy less than 8 bytes? */ 43 bhs .Ldst_aligner /* nope, do it the long way */ 44 45 1: ldrb ip, [r1], #1 /* load a byte from src */ 46 subs r2, r2, #1 /* and more to transfer? */ 47 strb ip, [r3], #1 /* save it to dst */ 48 bne 1b /* yes, do next byte */ 49 RET /* return */ 50 #endif 51 52 .Ldst_aligner: 53 tst r3, #7 /* is dst pointer word aligned? */ 54 beq .Lsrc_aligner /* yes, check src pointer */ 55 /* 56 * Until the dst pointer is word aligned, read src and dst byte by 57 * byte until it is aligned or we've copied everything. 58 */ 59 ldrb ip, [r1], #1 /* load a byte from src */ 60 strb ip, [r3], #1 /* save the byte to dst */ 61 subs r2, r2, #1 /* end of transfer? */ 62 bne .Ldst_aligner /* no, try next byte */ 63 RET /* yes, we're done! */ 64 65 .Lsrc_aligner: 66 push {r4-r5} /* save some registers */ 67 add r4, r2, r3 /* keep a pointer to the end of src */ 68 ands r5, r1, #7 /* get misalignment of src pointer */ 69 beq .Lcongruent_main /* aligned, do it the fast way */ 70 71 vdup.8 d1, r5 /* set offset for table */ 72 rsb r5, r5, #8 /* calculate leftover of each word */ 73 bic r1, r1, #7 /* dword align src pointer */ 74 75 vldr d0, .Ltbl_value /* load table value */ 76 vadd.u8 d0, d0, d1 /* add offset to it */ 77 78 vld1.64 {d1}, [r1:64]! /* load a dword from src */ 79 80 cmp r2, r5 /* do we already have enough? */ 81 bhi .Lincongruent /* no, so read more */ 82 83 .Lincongruent_finish: 84 vtbl.8 d0, {d1-d2}, d0 /* merge last dwords */ 85 cmp r2, #8 /* room for a full dword? */ 86 #ifdef __ARMEB__ 87 vrev64.32 d0, d0 /* word swap to LE */ 88 #endif 89 blo .Lfinish /* no, write final partial dword */ 90 vst1.32 {d0}, [r3:64] /* yes, write final full dword */ 91 b .Ldone /* and we're done! */ 92 93 .Lincongruent: 94 vld1.64 {d2}, [r1:64]! /* load a dword */ 95 cmp r2, #8 /* can we write a full dword? */ 96 blo .Lincongruent_finish /* no, finish it. */ 97 vtbl.8 d1, {d1-d2}, d0 /* reorder */ 98 vst1.64 {d1}, [r3:64]! /* store a dword */ 99 subs r2, r2, #8 /* have we written everything? */ 100 beq .Ldone /* yes, we're done! */ 101 vmov d1, d2 /* prepare for next dword */ 102 tst r3, #63 /* are we 64-byte aligned? */ 103 bne .Lincongruent /* no, load next dword */ 104 105 /* 106 * We are now 64-byte aligneds so all writes should fill one or more 107 * cachelines. Even if d1 has 7 bytes cached, to write 32 bytes we 108 * still need to read 4 dwords (3 full dwords and 1 dword for that 109 * last byte). 110 */ 111 cmp r2, #32 /* can we write 4 more dwords? */ 112 blo .Lincongruent_dword /* no, handle dword by dword */ 113 vld1.64 {d2-d5}, [r1:64]! /* read 4 dwords */ 114 cmp r2, #64 /* can we write 4 more dwords? */ 115 blo .Lincongruent_4dword /* no, handle it */ 116 117 1: vld1.64 {d7-d10}, [r1:64]! /* read 4 dwords */ 118 vtbl.8 d1, {d1-d2}, d0 /* reorder */ 119 vtbl.8 d2, {d2-d3}, d0 /* reorder */ 120 vtbl.8 d3, {d3-d4}, d0 /* reorder */ 121 vtbl.8 d4, {d4-d5}, d0 /* reorder */ 122 vst1.64 {d1-d4}, [r3:64]! /* write 4 dwords */ 123 vmov d6, d5 /* move out of the way the load */ 124 cmp r2, #96 /* have 8+4 dwords to write? */ 125 blo 2f /* no more data, skip the load */ 126 vld1.64 {d2-d5}, [r1:64]! /* more data, load 4 dwords */ 127 2: vtbl.8 d6, {d6-d7}, d0 /* reorder */ 128 vtbl.8 d7, {d7-d8}, d0 /* reorder */ 129 vtbl.8 d8, {d8-d9}, d0 /* reorder */ 130 vtbl.8 d9, {d9-d10}, d0 /* reorder */ 131 vst1.64 {d6-d9}, [r3:64]! /* write 4 dwords */ 132 subs r2, r2, #64 133 beq .Ldone 134 vmov d1, d10 135 cmp r2, #64 136 bhs 1b 137 138 /* 139 * we have leftovers in d1 and new untranslated date in d2-d5. 140 */ 141 .Lincongruent_4dword: 142 cmp r2, #32 143 blo .Lincongruent_dword 144 145 vtbl.8 d1, {d1-d2}, d0 /* reorder */ 146 vtbl.8 d2, {d2-d3}, d0 /* reorder */ 147 vtbl.8 d3, {d3-d4}, d0 /* reorder */ 148 vtbl.8 d4, {d4-d5}, d0 /* reorder */ 149 vst1.64 {d1-d4}, [r3:64]! /* write 4 dwords */ 150 vmov d1, d5 /* move leftovers */ 151 subs r2, r2, #32 152 beq .Ldone 153 154 .Lincongruent_dword: 155 #if 0 156 cmp r2, r5 /* enough in leftovers? */ 157 bls .Lincongruent_finish /* yes, finish it. */ 158 vld1.64 {d2}, [r1:64]! /* load a dword */ 159 cmp r2, #8 /* can we write a full dword? */ 160 blo .Lincongruent_finish /* no, finish it. */ 161 vtbl.8 d1, {d1-d2}, d0 /* reorder */ 162 vst1.64 {d1}, [r3:64]! /* store a dword */ 163 subs r2, r2, #8 /* have we written everything? */ 164 beq .Ldone /* yes, we're done! */ 165 b .Lincongruent_dword /* and go get it */ 166 #else 167 cmp r2, r5 /* are the bytes we have enough? */ 168 bls .Lincongruent_finish /* yes, finish it. */ 169 mov ip, r2 /* get remaining count */ 170 bic ip, ip, #7 /* truncate to a dword */ 171 rsb ip, ip, #32 /* subtract from 32 */ 172 ands r2, r2, #7 /* count mod 8 */ 173 add pc, pc, ip, lsl #1 /* and jump! */ 174 nop 175 vld1.64 {d2}, [r1:64]! /* load a dword */ 176 vtbl.8 d1, {d1-d2}, d0 /* reorder */ 177 vst1.64 {d1}, [r3:64]! /* store a dword */ 178 vmov d1, d2 /* prepare for next dword */ 179 vld1.64 {d2}, [r1:64]! /* load a dword */ 180 vtbl.8 d1, {d1-d2}, d0 /* reorder */ 181 vst1.64 {d1}, [r3:64]! /* store a dword */ 182 vmov d1, d2 /* prepare for next dword */ 183 vld1.64 {d2}, [r1:64]! /* load a dword */ 184 vtbl.8 d1, {d1-d2}, d0 /* reorder */ 185 vst1.64 {d1}, [r3:64]! /* store a dword */ 186 vmov d1, d2 /* prepare for next dword */ 187 vld1.64 {d2}, [r1:64]! /* load a dword */ 188 vtbl.8 d1, {d1-d2}, d0 /* reorder */ 189 vst1.64 {d1}, [r3:64]! /* store a dword */ 190 vmov d1, d2 /* prepare for next dword */ 191 beq .Ldone 192 vld1.64 {d2}, [r1:64]! /* load a dword */ 193 b .Lincongruent_finish /* write last partial dword */ 194 #endif 195 196 .Lcongruent_main: 197 vld1.32 {d0}, [r1:64]! /* load next dword */ 198 cmp r2, #8 /* compare current ptr against end */ 199 blo .Lfinish /* greater so write final dword */ 200 vst1.32 {d0}, [r3:64]! /* store dword */ 201 subs r2, r2, #8 /* compare current ptr against end */ 202 beq .Ldone /* equal? we're done! */ 203 tst r3, #63 /* have we hit a 64-byte boundary? */ 204 bne .Lcongruent_main /* no, write next word */ 205 206 cmp r2, #64 /* can we write 4 dwords? */ 207 blo .Lcongruent_loop /* no, this dword by dword */ 208 vldm r1!, {d0-d7} /* load next 7 dwords */ 209 cmp r2, #128 /* can we write 16 dwords */ 210 blo 3f /* no, then deal with 8 dwords */ 211 212 /* 213 * The following writes two 64-byte interleaving stores and loads. 214 */ 215 1: vldm r1!, {d8-d15} /* load next 8 dwords */ 216 vstm r3!, {d0-d7} /* store 8 more dwords */ 217 cmp r2, #192 /* can we write 16+8 dwords? */ 218 blo 2f /* no, don't load the next 8 dwords */ 219 vldm r1!, {d0-d7} /* yes, load next 8 dwords */ 220 2: vstm r3!, {d8-d15} /* store 8 more dwords */ 221 sub r2, r2, #128 /* we just stored 16 (8+8) dwords */ 222 beq .Ldone /* if 0, we're done! */ 223 cmp r2, #128 /* can we write 16 dwords */ 224 bhs 1b /* yes, do it again */ 225 cmp r2, #64 /* have we loaded 8 dwords? */ 226 blo .Lcongruent_loop /* no, proceed to do it dword */ 227 228 /* 229 * We now have 8 dwords we can write in d0-d7. 230 */ 231 3: vstm r3!, {d0-d7} /* store 8 more dwords */ 232 subs r2, r2, #64 /* we wrote 8 dwords */ 233 beq .Ldone /* if 0, we're done! */ 234 235 .Lcongruent_loop: 236 vld1.32 {d0}, [r1]! /* load dword from src */ 237 cmp r2, #8 /* can we write a full dword? */ 238 blo .Lfinish /* no, write last partial dword */ 239 .Lcongruent_loop_start: 240 vst1.32 {d0}, [r3]! /* store dword into dst */ 241 subs r2, r2, #8 /* subtract it from length */ 242 beq .Ldone /* if 0, we're done! */ 243 vld1.32 {d0}, [r1]! /* load dword from src */ 244 cmp r2, #8 /* can we write a full dword? */ 245 bhs .Lcongruent_loop_start /* yes, so do it */ 246 247 .Lfinish: 248 vmov r4, r5, d0 /* get last dword from NEON */ 249 tst r2, #4 /* do we have at least 4 bytes left? */ 250 strne r4, [r3], #4 /* store the 1st word */ 251 movne r4, r5 /* move 2nd word into place */ 252 tst r2, #2 /* do we have at least 2 bytes left? */ 253 #ifdef __ARMEB__ 254 movne r4, r4, ror #16 /* yes, swap halfwords */ 255 #endif 256 strneh r4, [r3], #2 /* yes, store the halfword */ 257 #ifdef __ARMEL__ 258 movne r4, r4, lsr #16 /* yes, discard just written bytes */ 259 #endif 260 tst r2, #1 /* do we have a final byte? */ 261 #ifdef __ARMEB__ 262 movne r4, r4, lsr #24 /* yes, move MSB to LSB */ 263 #endif 264 strneb r4, [r3], #1 /* yes, store it */ 265 266 .Ldone: 267 pop {r4-r5} /* restore registers */ 268 RET 269 270 .p2align 3 271 .Ltbl_value: 272 #ifdef __ARMEL__ 273 .quad 0x0706050403020100 274 #else 275 .quad 0x0001020304050607 276 #endif 277 END(memcpy) 278
Indexes created Sun Sep 21 19:09:51 GMT 2025