1/* $NetBSD: cpu_in_cksum.S,v 1.6 2013/08/18 15:58:19 matt Exp $ */ 2 3/* 4 * Copyright 2003 Wasabi Systems, Inc. 5 * All rights reserved. 6 * 7 * Written by Steve C. Woodford for Wasabi Systems, Inc. 8 * 9 * Redistribution and use in source and binary forms, with or without 10 * modification, are permitted provided that the following conditions 11 * are met: 12 * 1. Redistributions of source code must retain the above copyright 13 * notice, this list of conditions and the following disclaimer. 14 * 2. Redistributions in binary form must reproduce the above copyright 15 * notice, this list of conditions and the following disclaimer in the 16 * documentation and/or other materials provided with the distribution. 17 * 3. All advertising materials mentioning features or use of this software 18 * must display the following acknowledgement: 19 * This product includes software developed for the NetBSD Project by 20 * Wasabi Systems, Inc. 21 * 4. The name of Wasabi Systems, Inc. may not be used to endorse 22 * or promote products derived from this software without specific prior 23 * written permission. 24 * 25 * THIS SOFTWARE IS PROVIDED BY WASABI SYSTEMS, INC. ``AS IS'' AND 26 * ANY EXPRESS OR IMPLIED WARRANTIES, INCLUDING, BUT NOT LIMITED 27 * TO, THE IMPLIED WARRANTIES OF MERCHANTABILITY AND FITNESS FOR A PARTICULAR 28 * PURPOSE ARE DISCLAIMED. IN NO EVENT SHALL WASABI SYSTEMS, INC 29 * BE LIABLE FOR ANY DIRECT, INDIRECT, INCIDENTAL, SPECIAL, EXEMPLARY, OR 30 * CONSEQUENTIAL DAMAGES (INCLUDING, BUT NOT LIMITED TO, PROCUREMENT OF 31 * SUBSTITUTE GOODS OR SERVICES; LOSS OF USE, DATA, OR PROFITS; OR BUSINESS 32 * INTERRUPTION) HOWEVER CAUSED AND ON ANY THEORY OF LIABILITY, WHETHER IN 33 * CONTRACT, STRICT LIABILITY, OR TORT (INCLUDING NEGLIGENCE OR OTHERWISE) 34 * ARISING IN ANY WAY OUT OF THE USE OF THIS SOFTWARE, EVEN IF ADVISED OF THE 35 * POSSIBILITY OF SUCH DAMAGE. 36 */ 37 38/* 39 * Hand-optimised in_cksum() and in4_cksum() implementations for ARM/Xscale 40 */ 41 42#include <machine/asm.h> 43RCSID("$NetBSD: cpu_in_cksum.S,v 1.6 2013/08/18 15:58:19 matt Exp $") 44 45#include "assym.h" 46 47/* 48 * int cpu_in_cksum(struct mbuf *m, int len, int off, uint32_t initial_sum) 49 * 50 * Entry: 51 * r0 m 52 * r1 len 53 * r2 off 54 * r3 initial_sum 55 * 56 * Function wide register usage 57 * r8 accumulated sum 58 * r9 remaining length to parse 59 * ip pointer to next mbuf 60 */ 61/* LINTSTUB: Func: int cpu_in_cksum(struct mbuf *, int, int, uint32_t) */ 62ENTRY(cpu_in_cksum) 63 push {r4-r11,lr} 64 65 mov r8, r3 /* Accumulate sum in r8 */ 66 mov r9, r1 /* save len in r9 */ 67 mov ip, r0 /* set ip to the current mbuf */ 68 69.Lin_cksum_skip_loop: 70 ldr r1, [ip, #(M_LEN)] 71 ldr r0, [ip, #(M_DATA)] 72 ldr ip, [ip, #(M_NEXT)] 73.Lin_cksum_skip_entry: 74 subs r2, r2, r1 /* offset = offset - mbuf length */ 75 blt .Lin_cksum_skip_done /* if offset has gone negative start with this mbuf */ 76 cmp ip, #0x00 77 bne .Lin_cksum_skip_loop 78 b .Lin_cksum_whoops 79 80.Lin_cksum_skip_done: 81 add r0, r2, r0 /* data += offset (offset is < 0) */ 82 add r0, r0, r1 /* data += length of mbuf */ 83 /* data == start of data to cksum */ 84 rsb r1, r2, #0x00 /* length = remainder of mbuf to read */ 85 mov r10, #0x00 86 b .Lin_cksum_entry 87 88.Lin_cksum_loop: 89 ldr r1, [ip, #(M_LEN)] 90 ldr r0, [ip, #(M_DATA)] 91 ldr ip, [ip, #(M_NEXT)] 92.Lin_cksum_entry: 93 cmp r9, r1 94#ifdef __thumb__ 95 bge 1f 96 mov r1, r9 97#else 98 movlt r1, r9 99#endif 1001: sub r9, r9, r1 101 eor r11, r10, r0 102 add r10, r10, r1 103 adds r2, r1, #0x00 104#ifdef __thumb__ 105 it ne 106#endif 107 blne _ASM_LABEL(.L_cksumdata) 108 tst r11, #0x01 109#ifdef __thumb__ 110 it ne 111#endif 112 movne r2, r2, ror #8 113 adds r8, r8, r2 114 adc r8, r8, #0x00 115 cmp ip, #00 116 bne .Lin_cksum_loop 117 118#ifdef __thumb__ 119 mov r0, r8 120 lsls r2, r0, #16 121 adds r0, r0, r2 122 bcc 1f 123 adds r0, r0, #65536 1241: mvns r0, r0 125 lsrs r0, r0, #16 126#else 127 adds r8, r8, r8, lsl #16 128 addcs r8, r8, #65536 129 mvn r0, r8 130 lsr r0, r0, #16 131#endif 132 pop {r4-r11, lr} 133 RET 134 135.Lin_cksum_whoops: 136 adr r0, .Lin_cksum_whoops_str 137 bl _C_LABEL(panic) 138.Lin_cksum_whoops_str: 139 .asciz "in_cksum: out of mbufs\n" 140 .align 5 141END(cpu_in_cksum) 142 143 144/* 145 * The main in*_cksum() workhorse... 146 * 147 * Entry parameters: 148 * r0 Pointer to buffer 149 * r1 Buffer length 150 * lr Return address 151 * 152 * Returns: 153 * r2 Accumulated 32-bit sum 154 * 155 * Clobbers: 156 * r0-r7 157 */ 158/* LINTSTUB: Ignore */ 159ASENTRY_NP(.L_cksumdata) 160#ifdef __PROG26 161 str lr, [sp, #-4]! /* for SVC26 mode */ 162#endif 163#ifdef __XSCALE__ 164 pld [r0] /* Pre-fetch the start of the buffer */ 165#endif 166 movs r2, #0 167 168 /* We first have to word-align the buffer. */ 169 ands r7, r0, #0x03 170 beq .Lcksumdata_wordaligned 171 eors r0, r0, r7 /* r0 is word aligned */ 172 ldr r2, [r0], #0x04 173#ifdef __thumb__ 174 movs r4, r7 175 lsls r4, r4, #3 176#else 177 lsl r4, r7, #3 178#endif 179#if defined(__ARMEB__) 180 lsls r2, r2, r4 181 lsrs r2, r2, r4 182#else 183 lsrs r2, r2, r4 184 lsls r2, r2, r4 185#endif 186 rsb r7, r7, #0x04 187 subs r1, r1, r7 /* Enough bytes left to make it? */ 188 bgt .Lcksumdata_wordaligned 189#ifdef __PROG26 190 ldreq pc, [sp], #4 /* done */ 191#else 192 RETc(eq) /* done */ 193#endif 194 adds r7, r7, r1 /* undo sub */ 195 adds r7, r7, r1 /* r7 = offset + len */ 196 rsb r7, r7, #4 197 lsls r7, r7, #3 198#if defined(__ARMEB__) 199 lsrs r2, r2, r7 200 lsls r2, r2, r7 201#else 202 lsls r2, r2, r7 203 lsrs r2, r2, r7 204#endif 205#ifdef __PROG26 206 ldr pc, [sp], #4 /* done */ 207#else 208 RET /* done */ 209#endif 210 211 /* Buffer is now word aligned */ 212.Lcksumdata_wordaligned: 213#ifdef __XSCALE__ 214 cmp r1, #0x04 /* Less than 4 bytes left? */ 215 blt .Lcksumdata_endgame /* Yup */ 216 217 /* Now quad-align, if necessary */ 218 ands r7, r0, #0x04 219 ldrne r7, [r0], #0x04 220 subne r1, r1, #0x04 221 subs r1, r1, #0x40 222 blt .Lcksumdata_bigloop_end /* Note: C flag clear if branch taken */ 223 224 /* 225 * Buffer is now quad aligned. Sum 64 bytes at a time. 226 * Note: First ldrd is hoisted above the loop, together with 227 * setting r6 to zero to avoid stalling for results in the 228 * loop. (r7 is live, from above). 229 */ 230 ldrd r4, [r0], #0x08 231 mov r6, #0x00 232.Lcksumdata_bigloop: 233 pld [r0, #0x18] 234 adds r2, r2, r6 235 adcs r2, r2, r7 236 ldrd r6, [r0], #0x08 237 adcs r2, r2, r4 238 adcs r2, r2, r5 239 ldrd r4, [r0], #0x08 240 adcs r2, r2, r6 241 adcs r2, r2, r7 242 ldrd r6, [r0], #0x08 243 adcs r2, r2, r4 244 adcs r2, r2, r5 245 ldrd r4, [r0], #0x08 246 adcs r2, r2, r6 247 adcs r2, r2, r7 248 pld [r0, #0x18] 249 ldrd r6, [r0], #0x08 250 adcs r2, r2, r4 251 adcs r2, r2, r5 252 ldrd r4, [r0], #0x08 253 adcs r2, r2, r6 254 adcs r2, r2, r7 255 ldrd r6, [r0], #0x08 256 adcs r2, r2, r4 257 adcs r2, r2, r5 258 adcs r2, r2, #0x00 259 subs r1, r1, #0x40 260 ldrdge r4, [r0], #0x08 261 bge .Lcksumdata_bigloop 262 263 adds r2, r2, r6 /* r6/r7 still need summing */ 264.Lcksumdata_bigloop_end: 265 adcs r2, r2, r7 266 adcs r2, r2, #0x00 267 268#else /* !__XSCALE__ */ 269 270 subs r1, r1, #0x40 271 blt .Lcksumdata_bigloop_end 272 273.Lcksumdata_bigloop: 274 ldmia r0!, {r3, r4, r5, r6} 275 adds r2, r2, r3 276 adcs r2, r2, r4 277 adcs r2, r2, r5 278 ldmia r0!, {r3, r4, r5, r7} 279 adcs r2, r2, r6 280 adcs r2, r2, r3 281 adcs r2, r2, r4 282 adcs r2, r2, r5 283 ldmia r0!, {r3, r4, r5, r6} 284 adcs r2, r2, r7 285 adcs r2, r2, r3 286 adcs r2, r2, r4 287 adcs r2, r2, r5 288 ldmia r0!, {r3, r4, r5, r7} 289 adcs r2, r2, r6 290 adcs r2, r2, r3 291 adcs r2, r2, r4 292 adcs r2, r2, r5 293 adcs r2, r2, r7 294 adcs r2, r2, #0x00 295 subs r1, r1, #0x40 296 bge .Lcksumdata_bigloop 297.Lcksumdata_bigloop_end: 298#endif 299 300 adds r1, r1, #0x40 301#ifdef __PROG26 302 ldreq pc, [sp], #4 303#else 304 RETc(eq) 305#endif 306 cmp r1, #0x20 307 308#ifdef __XSCALE__ 309 ldrdge r4, [r0], #0x08 /* Avoid stalling pld and result */ 310 blt .Lcksumdata_less_than_32 311 pld [r0, #0x18] 312 ldrd r6, [r0], #0x08 313 adds r2, r2, r4 314 adcs r2, r2, r5 315 ldrd r4, [r0], #0x08 316 adcs r2, r2, r6 317 adcs r2, r2, r7 318 ldrd r6, [r0], #0x08 319 adcs r2, r2, r4 320 adcs r2, r2, r5 321 adcs r2, r2, r6 /* XXX: Unavoidable result stall */ 322 adcs r2, r2, r7 323#else 324 blt .Lcksumdata_less_than_32 325 ldmia r0!, {r3, r4, r5, r6} 326 adds r2, r2, r3 327 adcs r2, r2, r4 328 adcs r2, r2, r5 329 ldmia r0!, {r3, r4, r5, r7} 330 adcs r2, r2, r6 331 adcs r2, r2, r3 332 adcs r2, r2, r4 333 adcs r2, r2, r5 334 adcs r2, r2, r7 335#endif 336 adcs r2, r2, #0x00 337 subs r1, r1, #0x20 338#ifdef __PROG26 339 ldreq pc, [sp], #4 340#else 341 RETc(eq) 342#endif 343 344.Lcksumdata_less_than_32: 345 /* There are less than 32 bytes left */ 346 and r3, r1, #0x18 347 rsb r4, r3, #0x18 348 subs r1, r1, r3 349 adds r4, r4, r4, lsr #1 /* Side effect: Clear carry flag */ 350#ifdef __thumb__ 351 it ne 352#endif 353 addne pc, pc, r4 354 355/* 356 * Note: We use ldm here, even on Xscale, since the combined issue/result 357 * latencies for ldm and ldrd are the same. Using ldm avoids needless #ifdefs. 358 */ 359 /* At least 24 bytes remaining... */ 360 ldmia r0!, {r4, r5} 361 nop 362 adcs r2, r2, r4 363 adcs r2, r2, r5 364 365 /* At least 16 bytes remaining... */ 366 ldmia r0!, {r4, r5} 367 adcs r2, r2, r4 368 adcs r2, r2, r5 369 370 /* At least 8 bytes remaining... */ 371 ldmia r0!, {r4, r5} 372 adcs r2, r2, r4 373 adcs r2, r2, r5 374 375 /* Less than 8 bytes remaining... */ 376 adcs r2, r2, #0x00 377 subs r1, r1, #0x04 378 blt .Lcksumdata_lessthan4 379 380 ldr r4, [r0], #0x04 381 subs r1, r1, #0x04 382 adds r2, r2, r4 383 adcs r2, r2, #0x00 384 385 /* Deal with < 4 bytes remaining */ 386.Lcksumdata_lessthan4: 387 adds r1, r1, #0x04 388#ifdef __PROG26 389 ldreq pc, [sp], #4 390#else 391 RETc(eq) 392#endif 393 394 /* Deal with 1 to 3 remaining bytes, possibly misaligned */ 395.Lcksumdata_endgame: 396 ldr r3, [r0] /* Fetch last word */ 397 rsb r1, r1, #4 /* get discard amount */ 398 lsl r1, r1, #3 /* turn it into bits */ 399#ifdef __ARMEB__ 400 lsr r3, r3, r1 /* discard least significant bits */ 401 lsl r3, r3, r1 /* shift back filling with zeros */ 402#else 403 lsl r3, r3, r1 /* discard least significant bits */ 404 lsr r3, r3, r1 /* shift back filling with zeros */ 405#endif 406 adds r2, r2, r3 407 adcs r2, r2, #0x00 408#ifdef __PROG26 409 ldr pc, [sp], #4 410#else 411 RET 412#endif 413ASEND(.L_cksumdata) 414