1 /* SPDX-License-Identifier: GPL-2.0-only */ <<
2 /* 1 /*
3 * Copyright (c) 2010-2011, The Linux Foundati !! 2 * This file is subject to the terms and conditions of the GNU General Public
>> 3 * License. See the file "COPYING" in the main directory of this archive
>> 4 * for more details.
>> 5 *
>> 6 * Unified implementation of memcpy, memmove and the __copy_user backend.
>> 7 *
>> 8 * Copyright (C) 1998, 99, 2000, 01, 2002 Ralf Baechle (ralf@gnu.org)
>> 9 * Copyright (C) 1999, 2000, 01, 2002 Silicon Graphics, Inc.
>> 10 * Copyright (C) 2002 Broadcom, Inc.
>> 11 * memcpy/copy_user author: Mark Vandevoorde
>> 12 * Copyright (C) 2007 Maciej W. Rozycki
>> 13 * Copyright (C) 2014 Imagination Technologies Ltd.
>> 14 *
>> 15 * Mnemonic names for arguments to memcpy/__copy_user
4 */ 16 */
5 17
6 /* 18 /*
7 * Description !! 19 * Hack to resolve longstanding prefetch issue
8 * <<
9 * library function for memcpy where length <<
10 * ptr_in to ptr_out. ptr_out is returned un <<
11 * Allows any combination of alignment on in <<
12 * and length from 0 to 2^32-1 <<
13 * <<
14 * Restrictions <<
15 * The arrays should not overlap, the progra <<
16 * if they do. <<
17 * For blocks less than 16 bytes a byte by b <<
18 * 8byte alignments, and length multiples, a <<
19 * 96bytes <<
20 * History <<
21 * <<
22 * DJH 5/15/09 Initial version 1.0 <<
23 * DJH 6/ 1/09 Version 1.1 modified ABI to <<
24 * DJH 7/12/09 Version 1.2 optimized codesi <<
25 * DJH 10/14/09 Version 1.3 added special lo <<
26 * overreading bloa <<
27 * DJH 4/20/10 Version 1.4 fixed Ldword_loo <<
28 * occurring if onl <<
29 * # 3888, correcte <<
30 * 1 32byte chunk f <<
31 * loop at end to s <<
32 * over read. Fixe <<
33 * overread for blo <<
34 * codesize to 752 <<
35 * DJH 4/21/10 version 1.5 1.4 fix broke co <<
36 * aligned to dword <<
37 * byte, added dete <<
38 * little bloat. <<
39 * DJH 4/23/10 version 1.6 corrected stack <<
40 * always, fixed th <<
41 * before it was be <<
42 * Natural c model <<
43 * =============== <<
44 * void * memcpy(char * ptr_out, char * ptr_in <<
45 * int i; <<
46 * if(length) for(i=0; i < length; i++) { pt <<
47 * return(ptr_out); <<
48 * } <<
49 * <<
50 * Optimized memcpy function <<
51 * ========================= <<
52 * void * memcpy(char * ptr_out, char * ptr_in <<
53 * int i, prolog, kernel, epilog, mask; <<
54 * u8 offset; <<
55 * s64 data0, dataF8, data70; <<
56 * <<
57 * s64 * ptr8_in; <<
58 * s64 * ptr8_out; <<
59 * s32 * ptr4; <<
60 * s16 * ptr2; <<
61 * <<
62 * offset = ((int) ptr_in) & 7; <<
63 * ptr8_in = (s64 *) &ptr_in[-offset]; //r <<
64 * <<
65 * data70 = *ptr8_in++; <<
66 * dataF8 = *ptr8_in++; <<
67 * <<
68 * data0 = HEXAGON_P_valignb_PPp(dataF8, dat <<
69 * <<
70 * prolog = 32 - ((int) ptr_out); <<
71 * mask = 0x7fffffff >> HEXAGON_R_cl0_R(len <<
72 * prolog = prolog & mask; <<
73 * kernel = len - prolog; <<
74 * epilog = kernel & 0x1F; <<
75 * kernel = kernel>>5; <<
76 * <<
77 * if (prolog & 1) { ptr_out[0] = (u8) data0 <<
78 * ptr2 = (s16 *) &ptr_out[0]; <<
79 * if (prolog & 2) { ptr2[0] = (u16) data0; <<
80 * ptr4 = (s32 *) &ptr_out[0]; <<
81 * if (prolog & 4) { ptr4[0] = (u32) data0; <<
82 * <<
83 * offset = offset + (prolog & 7); <<
84 * if (offset >= 8) { <<
85 * data70 = dataF8; <<
86 * dataF8 = *ptr8_in++; <<
87 * } <<
88 * offset = offset & 0x7; <<
89 * 20 *
90 * prolog = prolog >> 3; !! 21 * Prefetching may be fatal on some systems if we're prefetching beyond the
91 * if (prolog) for (i=0; i < prolog; i++) { !! 22 * end of memory on some systems. It's also a seriously bad idea on non
92 * data0 = HEXAGON_P_valignb_PPp(dataF8, !! 23 * dma-coherent systems.
93 * ptr8_out = (s64 *) &ptr_out[0]; *ptr8 !! 24 */
94 * data70 = dataF8; !! 25 #ifdef CONFIG_DMA_NONCOHERENT
95 * dataF8 = *ptr8_in++; !! 26 #undef CONFIG_CPU_HAS_PREFETCH
96 * } !! 27 #endif
97 * if(kernel) { kernel -= 1; epilog += 32; } !! 28 #ifdef CONFIG_MIPS_MALTA
98 * if(kernel) for(i=0; i < kernel; i++) { !! 29 #undef CONFIG_CPU_HAS_PREFETCH
99 * data0 = HEXAGON_P_valignb_PPp(dataF8, !! 30 #endif
100 * ptr8_out = (s64 *) &ptr_out[0]; *ptr8 !! 31 #ifdef CONFIG_CPU_MIPSR6
101 * data70 = *ptr8_in++; !! 32 #undef CONFIG_CPU_HAS_PREFETCH
102 * !! 33 #endif
103 * data0 = HEXAGON_P_valignb_PPp(data70, !! 34
104 * ptr8_out = (s64 *) &ptr_out[0]; *ptr8 !! 35 #include <linux/export.h>
105 * dataF8 = *ptr8_in++; !! 36 #include <asm/asm.h>
106 * !! 37 #include <asm/asm-offsets.h>
107 * data0 = HEXAGON_P_valignb_PPp(dataF8, !! 38 #include <asm/regdef.h>
108 * ptr8_out = (s64 *) &ptr_out[0]; *ptr8 !! 39
109 * data70 = *ptr8_in++; !! 40 #define dst a0
110 * !! 41 #define src a1
111 * data0 = HEXAGON_P_valignb_PPp(data70, !! 42 #define len a2
112 * ptr8_out = (s64 *) &ptr_out[0]; *ptr8 !! 43
113 * dataF8 = *ptr8_in++; !! 44 /*
114 * } !! 45 * Spec
115 * epilogdws = epilog >> 3; <<
116 * if (epilogdws) for (i=0; i < epilogdws; i <<
117 * data0 = HEXAGON_P_valignb_PPp(dataF8, <<
118 * ptr8_out = (s64 *) &ptr_out[0]; *ptr8 <<
119 * data70 = dataF8; <<
120 * dataF8 = *ptr8_in++; <<
121 * } <<
122 * data0 = HEXAGON_P_valignb_PPp(dataF8, dat <<
123 * 46 *
124 * ptr4 = (s32 *) &ptr_out[0]; !! 47 * memcpy copies len bytes from src to dst and sets v0 to dst.
125 * if (epilog & 4) { ptr4[0] = (u32) data0; !! 48 * It assumes that
126 * ptr2 = (s16 *) &ptr_out[0]; !! 49 * - src and dst don't overlap
127 * if (epilog & 2) { ptr2[0] = (u16) data0; !! 50 * - src is readable
128 * if (epilog & 1) { *ptr_out++ = (u8) data0 !! 51 * - dst is writable
>> 52 * memcpy uses the standard calling convention
>> 53 *
>> 54 * __copy_user copies up to len bytes from src to dst and sets a2 (len) to
>> 55 * the number of uncopied bytes due to an exception caused by a read or write.
>> 56 * __copy_user assumes that src and dst don't overlap, and that the call is
>> 57 * implementing one of the following:
>> 58 * copy_to_user
>> 59 * - src is readable (no exceptions when reading src)
>> 60 * copy_from_user
>> 61 * - dst is writable (no exceptions when writing dst)
>> 62 * __copy_user uses a non-standard calling convention; see
>> 63 * include/asm-mips/uaccess.h
>> 64 *
>> 65 * When an exception happens on a load, the handler must
>> 66 # ensure that all of the destination buffer is overwritten to prevent
>> 67 * leaking information to user mode programs.
>> 68 */
>> 69
>> 70 /*
>> 71 * Implementation
>> 72 */
>> 73
>> 74 /*
>> 75 * The exception handler for loads requires that:
>> 76 * 1- AT contain the address of the byte just past the end of the source
>> 77 * of the copy,
>> 78 * 2- src_entry <= src < AT, and
>> 79 * 3- (dst - src) == (dst_entry - src_entry),
>> 80 * The _entry suffix denotes values when __copy_user was called.
>> 81 *
>> 82 * (1) is set up up by uaccess.h and maintained by not writing AT in copy_user
>> 83 * (2) is met by incrementing src by the number of bytes copied
>> 84 * (3) is met by not doing loads between a pair of increments of dst and src
129 * 85 *
130 * return(ptr_out - length); !! 86 * The exception handlers for stores adjust len (if necessary) and return.
131 * } !! 87 * These handlers do not need to overwrite any data.
132 * 88 *
133 * Codesize : 784 bytes !! 89 * For __rmemcpy and memmove an exception is always a kernel bug, therefore
>> 90 * they're not protected.
>> 91 */
>> 92
>> 93 /* Instruction type */
>> 94 #define LD_INSN 1
>> 95 #define ST_INSN 2
>> 96 /* Pretech type */
>> 97 #define SRC_PREFETCH 1
>> 98 #define DST_PREFETCH 2
>> 99 #define LEGACY_MODE 1
>> 100 #define EVA_MODE 2
>> 101 #define USEROP 1
>> 102 #define KERNELOP 2
>> 103
>> 104 /*
>> 105 * Wrapper to add an entry in the exception table
>> 106 * in case the insn causes a memory exception.
>> 107 * Arguments:
>> 108 * insn : Load/store instruction
>> 109 * type : Instruction type
>> 110 * reg : Register
>> 111 * addr : Address
>> 112 * handler : Exception handler
>> 113 */
>> 114
>> 115 #define EXC(insn, type, reg, addr, handler) \
>> 116 .if \mode == LEGACY_MODE; \
>> 117 9: insn reg, addr; \
>> 118 .section __ex_table,"a"; \
>> 119 PTR_WD 9b, handler; \
>> 120 .previous; \
>> 121 /* This is assembled in EVA mode */ \
>> 122 .else; \
>> 123 /* If loading from user or storing to user */ \
>> 124 .if ((\from == USEROP) && (type == LD_INSN)) || \
>> 125 ((\to == USEROP) && (type == ST_INSN)); \
>> 126 9: __BUILD_EVA_INSN(insn##e, reg, addr); \
>> 127 .section __ex_table,"a"; \
>> 128 PTR_WD 9b, handler; \
>> 129 .previous; \
>> 130 .else; \
>> 131 /* \
>> 132 * Still in EVA, but no need for \
>> 133 * exception handler or EVA insn \
>> 134 */ \
>> 135 insn reg, addr; \
>> 136 .endif; \
>> 137 .endif
>> 138
>> 139 /*
>> 140 * Only on the 64-bit kernel we can made use of 64-bit registers.
>> 141 */
>> 142 #ifdef CONFIG_64BIT
>> 143 #define USE_DOUBLE
>> 144 #endif
>> 145
>> 146 #ifdef USE_DOUBLE
>> 147
>> 148 #define LOADK ld /* No exception */
>> 149 #define LOAD(reg, addr, handler) EXC(ld, LD_INSN, reg, addr, handler)
>> 150 #define LOADL(reg, addr, handler) EXC(ldl, LD_INSN, reg, addr, handler)
>> 151 #define LOADR(reg, addr, handler) EXC(ldr, LD_INSN, reg, addr, handler)
>> 152 #define STOREL(reg, addr, handler) EXC(sdl, ST_INSN, reg, addr, handler)
>> 153 #define STORER(reg, addr, handler) EXC(sdr, ST_INSN, reg, addr, handler)
>> 154 #define STORE(reg, addr, handler) EXC(sd, ST_INSN, reg, addr, handler)
>> 155 #define ADD daddu
>> 156 #define SUB dsubu
>> 157 #define SRL dsrl
>> 158 #define SRA dsra
>> 159 #define SLL dsll
>> 160 #define SLLV dsllv
>> 161 #define SRLV dsrlv
>> 162 #define NBYTES 8
>> 163 #define LOG_NBYTES 3
>> 164
>> 165 /*
>> 166 * As we are sharing code base with the mips32 tree (which use the o32 ABI
>> 167 * register definitions). We need to redefine the register definitions from
>> 168 * the n64 ABI register naming to the o32 ABI register naming.
>> 169 */
>> 170 #undef t0
>> 171 #undef t1
>> 172 #undef t2
>> 173 #undef t3
>> 174 #define t0 $8
>> 175 #define t1 $9
>> 176 #define t2 $10
>> 177 #define t3 $11
>> 178 #define t4 $12
>> 179 #define t5 $13
>> 180 #define t6 $14
>> 181 #define t7 $15
>> 182
>> 183 #else
>> 184
>> 185 #define LOADK lw /* No exception */
>> 186 #define LOAD(reg, addr, handler) EXC(lw, LD_INSN, reg, addr, handler)
>> 187 #define LOADL(reg, addr, handler) EXC(lwl, LD_INSN, reg, addr, handler)
>> 188 #define LOADR(reg, addr, handler) EXC(lwr, LD_INSN, reg, addr, handler)
>> 189 #define STOREL(reg, addr, handler) EXC(swl, ST_INSN, reg, addr, handler)
>> 190 #define STORER(reg, addr, handler) EXC(swr, ST_INSN, reg, addr, handler)
>> 191 #define STORE(reg, addr, handler) EXC(sw, ST_INSN, reg, addr, handler)
>> 192 #define ADD addu
>> 193 #define SUB subu
>> 194 #define SRL srl
>> 195 #define SLL sll
>> 196 #define SRA sra
>> 197 #define SLLV sllv
>> 198 #define SRLV srlv
>> 199 #define NBYTES 4
>> 200 #define LOG_NBYTES 2
>> 201
>> 202 #endif /* USE_DOUBLE */
>> 203
>> 204 #define LOADB(reg, addr, handler) EXC(lb, LD_INSN, reg, addr, handler)
>> 205 #define STOREB(reg, addr, handler) EXC(sb, ST_INSN, reg, addr, handler)
>> 206
>> 207 #ifdef CONFIG_CPU_HAS_PREFETCH
>> 208 # define _PREF(hint, addr, type) \
>> 209 .if \mode == LEGACY_MODE; \
>> 210 kernel_pref(hint, addr); \
>> 211 .else; \
>> 212 .if ((\from == USEROP) && (type == SRC_PREFETCH)) || \
>> 213 ((\to == USEROP) && (type == DST_PREFETCH)); \
>> 214 /* \
>> 215 * PREFE has only 9 bits for the offset \
>> 216 * compared to PREF which has 16, so it may \
>> 217 * need to use the $at register but this \
>> 218 * register should remain intact because it's \
>> 219 * used later on. Therefore use $v1. \
>> 220 */ \
>> 221 .set at=v1; \
>> 222 user_pref(hint, addr); \
>> 223 .set noat; \
>> 224 .else; \
>> 225 kernel_pref(hint, addr); \
>> 226 .endif; \
>> 227 .endif
>> 228 #else
>> 229 # define _PREF(hint, addr, type)
>> 230 #endif
>> 231
>> 232 #define PREFS(hint, addr) _PREF(hint, addr, SRC_PREFETCH)
>> 233 #define PREFD(hint, addr) _PREF(hint, addr, DST_PREFETCH)
>> 234
>> 235 #ifdef CONFIG_CPU_LITTLE_ENDIAN
>> 236 #define LDFIRST LOADR
>> 237 #define LDREST LOADL
>> 238 #define STFIRST STORER
>> 239 #define STREST STOREL
>> 240 #define SHIFT_DISCARD SLLV
>> 241 #else
>> 242 #define LDFIRST LOADL
>> 243 #define LDREST LOADR
>> 244 #define STFIRST STOREL
>> 245 #define STREST STORER
>> 246 #define SHIFT_DISCARD SRLV
>> 247 #endif
>> 248
>> 249 #define FIRST(unit) ((unit)*NBYTES)
>> 250 #define REST(unit) (FIRST(unit)+NBYTES-1)
>> 251 #define UNIT(unit) FIRST(unit)
>> 252
>> 253 #define ADDRMASK (NBYTES-1)
>> 254
>> 255 .text
>> 256 .set noreorder
>> 257 #ifndef CONFIG_CPU_DADDI_WORKAROUNDS
>> 258 .set noat
>> 259 #else
>> 260 .set at=v1
>> 261 #endif
>> 262
>> 263 .align 5
>> 264
>> 265 /*
>> 266 * Macro to build the __copy_user common code
>> 267 * Arguments:
>> 268 * mode : LEGACY_MODE or EVA_MODE
>> 269 * from : Source operand. USEROP or KERNELOP
>> 270 * to : Destination operand. USEROP or KERNELOP
>> 271 */
>> 272 .macro __BUILD_COPY_USER mode, from, to
>> 273
>> 274 /* initialize __memcpy if this the first time we execute this macro */
>> 275 .ifnotdef __memcpy
>> 276 .set __memcpy, 1
>> 277 .hidden __memcpy /* make sure it does not leak */
>> 278 .endif
>> 279
>> 280 /*
>> 281 * Note: dst & src may be unaligned, len may be 0
>> 282 * Temps
>> 283 */
>> 284 #define rem t8
>> 285
>> 286 R10KCBARRIER(0(ra))
>> 287 /*
>> 288 * The "issue break"s below are very approximate.
>> 289 * Issue delays for dcache fills will perturb the schedule, as will
>> 290 * load queue full replay traps, etc.
>> 291 *
>> 292 * If len < NBYTES use byte operations.
>> 293 */
>> 294 PREFS( 0, 0(src) )
>> 295 PREFD( 1, 0(dst) )
>> 296 sltu t2, len, NBYTES
>> 297 and t1, dst, ADDRMASK
>> 298 PREFS( 0, 1*32(src) )
>> 299 PREFD( 1, 1*32(dst) )
>> 300 bnez t2, .Lcopy_bytes_checklen\@
>> 301 and t0, src, ADDRMASK
>> 302 PREFS( 0, 2*32(src) )
>> 303 PREFD( 1, 2*32(dst) )
>> 304 #ifndef CONFIG_CPU_NO_LOAD_STORE_LR
>> 305 bnez t1, .Ldst_unaligned\@
>> 306 nop
>> 307 bnez t0, .Lsrc_unaligned_dst_aligned\@
>> 308 #else /* CONFIG_CPU_NO_LOAD_STORE_LR */
>> 309 or t0, t0, t1
>> 310 bnez t0, .Lcopy_unaligned_bytes\@
>> 311 #endif /* CONFIG_CPU_NO_LOAD_STORE_LR */
>> 312 /*
>> 313 * use delay slot for fall-through
>> 314 * src and dst are aligned; need to compute rem
>> 315 */
>> 316 .Lboth_aligned\@:
>> 317 SRL t0, len, LOG_NBYTES+3 # +3 for 8 units/iter
>> 318 beqz t0, .Lcleanup_both_aligned\@ # len < 8*NBYTES
>> 319 and rem, len, (8*NBYTES-1) # rem = len % (8*NBYTES)
>> 320 PREFS( 0, 3*32(src) )
>> 321 PREFD( 1, 3*32(dst) )
>> 322 .align 4
>> 323 1:
>> 324 R10KCBARRIER(0(ra))
>> 325 LOAD(t0, UNIT(0)(src), .Ll_exc\@)
>> 326 LOAD(t1, UNIT(1)(src), .Ll_exc_copy\@)
>> 327 LOAD(t2, UNIT(2)(src), .Ll_exc_copy\@)
>> 328 LOAD(t3, UNIT(3)(src), .Ll_exc_copy\@)
>> 329 SUB len, len, 8*NBYTES
>> 330 LOAD(t4, UNIT(4)(src), .Ll_exc_copy\@)
>> 331 LOAD(t7, UNIT(5)(src), .Ll_exc_copy\@)
>> 332 STORE(t0, UNIT(0)(dst), .Ls_exc_p8u\@)
>> 333 STORE(t1, UNIT(1)(dst), .Ls_exc_p7u\@)
>> 334 LOAD(t0, UNIT(6)(src), .Ll_exc_copy\@)
>> 335 LOAD(t1, UNIT(7)(src), .Ll_exc_copy\@)
>> 336 ADD src, src, 8*NBYTES
>> 337 ADD dst, dst, 8*NBYTES
>> 338 STORE(t2, UNIT(-6)(dst), .Ls_exc_p6u\@)
>> 339 STORE(t3, UNIT(-5)(dst), .Ls_exc_p5u\@)
>> 340 STORE(t4, UNIT(-4)(dst), .Ls_exc_p4u\@)
>> 341 STORE(t7, UNIT(-3)(dst), .Ls_exc_p3u\@)
>> 342 STORE(t0, UNIT(-2)(dst), .Ls_exc_p2u\@)
>> 343 STORE(t1, UNIT(-1)(dst), .Ls_exc_p1u\@)
>> 344 PREFS( 0, 8*32(src) )
>> 345 PREFD( 1, 8*32(dst) )
>> 346 bne len, rem, 1b
>> 347 nop
>> 348
>> 349 /*
>> 350 * len == rem == the number of bytes left to copy < 8*NBYTES
>> 351 */
>> 352 .Lcleanup_both_aligned\@:
>> 353 beqz len, .Ldone\@
>> 354 sltu t0, len, 4*NBYTES
>> 355 bnez t0, .Lless_than_4units\@
>> 356 and rem, len, (NBYTES-1) # rem = len % NBYTES
>> 357 /*
>> 358 * len >= 4*NBYTES
>> 359 */
>> 360 LOAD( t0, UNIT(0)(src), .Ll_exc\@)
>> 361 LOAD( t1, UNIT(1)(src), .Ll_exc_copy\@)
>> 362 LOAD( t2, UNIT(2)(src), .Ll_exc_copy\@)
>> 363 LOAD( t3, UNIT(3)(src), .Ll_exc_copy\@)
>> 364 SUB len, len, 4*NBYTES
>> 365 ADD src, src, 4*NBYTES
>> 366 R10KCBARRIER(0(ra))
>> 367 STORE(t0, UNIT(0)(dst), .Ls_exc_p4u\@)
>> 368 STORE(t1, UNIT(1)(dst), .Ls_exc_p3u\@)
>> 369 STORE(t2, UNIT(2)(dst), .Ls_exc_p2u\@)
>> 370 STORE(t3, UNIT(3)(dst), .Ls_exc_p1u\@)
>> 371 .set reorder /* DADDI_WAR */
>> 372 ADD dst, dst, 4*NBYTES
>> 373 beqz len, .Ldone\@
>> 374 .set noreorder
>> 375 .Lless_than_4units\@:
>> 376 /*
>> 377 * rem = len % NBYTES
>> 378 */
>> 379 beq rem, len, .Lcopy_bytes\@
>> 380 nop
>> 381 1:
>> 382 R10KCBARRIER(0(ra))
>> 383 LOAD(t0, 0(src), .Ll_exc\@)
>> 384 ADD src, src, NBYTES
>> 385 SUB len, len, NBYTES
>> 386 STORE(t0, 0(dst), .Ls_exc_p1u\@)
>> 387 .set reorder /* DADDI_WAR */
>> 388 ADD dst, dst, NBYTES
>> 389 bne rem, len, 1b
>> 390 .set noreorder
>> 391
>> 392 #ifndef CONFIG_CPU_NO_LOAD_STORE_LR
>> 393 /*
>> 394 * src and dst are aligned, need to copy rem bytes (rem < NBYTES)
>> 395 * A loop would do only a byte at a time with possible branch
>> 396 * mispredicts. Can't do an explicit LOAD dst,mask,or,STORE
>> 397 * because can't assume read-access to dst. Instead, use
>> 398 * STREST dst, which doesn't require read access to dst.
>> 399 *
>> 400 * This code should perform better than a simple loop on modern,
>> 401 * wide-issue mips processors because the code has fewer branches and
>> 402 * more instruction-level parallelism.
>> 403 */
>> 404 #define bits t2
>> 405 beqz len, .Ldone\@
>> 406 ADD t1, dst, len # t1 is just past last byte of dst
>> 407 li bits, 8*NBYTES
>> 408 SLL rem, len, 3 # rem = number of bits to keep
>> 409 LOAD(t0, 0(src), .Ll_exc\@)
>> 410 SUB bits, bits, rem # bits = number of bits to discard
>> 411 SHIFT_DISCARD t0, t0, bits
>> 412 STREST(t0, -1(t1), .Ls_exc\@)
>> 413 jr ra
>> 414 move len, zero
>> 415 .Ldst_unaligned\@:
>> 416 /*
>> 417 * dst is unaligned
>> 418 * t0 = src & ADDRMASK
>> 419 * t1 = dst & ADDRMASK; T1 > 0
>> 420 * len >= NBYTES
>> 421 *
>> 422 * Copy enough bytes to align dst
>> 423 * Set match = (src and dst have same alignment)
>> 424 */
>> 425 #define match rem
>> 426 LDFIRST(t3, FIRST(0)(src), .Ll_exc\@)
>> 427 ADD t2, zero, NBYTES
>> 428 LDREST(t3, REST(0)(src), .Ll_exc_copy\@)
>> 429 SUB t2, t2, t1 # t2 = number of bytes copied
>> 430 xor match, t0, t1
>> 431 R10KCBARRIER(0(ra))
>> 432 STFIRST(t3, FIRST(0)(dst), .Ls_exc\@)
>> 433 beq len, t2, .Ldone\@
>> 434 SUB len, len, t2
>> 435 ADD dst, dst, t2
>> 436 beqz match, .Lboth_aligned\@
>> 437 ADD src, src, t2
>> 438
>> 439 .Lsrc_unaligned_dst_aligned\@:
>> 440 SRL t0, len, LOG_NBYTES+2 # +2 for 4 units/iter
>> 441 PREFS( 0, 3*32(src) )
>> 442 beqz t0, .Lcleanup_src_unaligned\@
>> 443 and rem, len, (4*NBYTES-1) # rem = len % 4*NBYTES
>> 444 PREFD( 1, 3*32(dst) )
>> 445 1:
>> 446 /*
>> 447 * Avoid consecutive LD*'s to the same register since some mips
>> 448 * implementations can't issue them in the same cycle.
>> 449 * It's OK to load FIRST(N+1) before REST(N) because the two addresses
>> 450 * are to the same unit (unless src is aligned, but it's not).
>> 451 */
>> 452 R10KCBARRIER(0(ra))
>> 453 LDFIRST(t0, FIRST(0)(src), .Ll_exc\@)
>> 454 LDFIRST(t1, FIRST(1)(src), .Ll_exc_copy\@)
>> 455 SUB len, len, 4*NBYTES
>> 456 LDREST(t0, REST(0)(src), .Ll_exc_copy\@)
>> 457 LDREST(t1, REST(1)(src), .Ll_exc_copy\@)
>> 458 LDFIRST(t2, FIRST(2)(src), .Ll_exc_copy\@)
>> 459 LDFIRST(t3, FIRST(3)(src), .Ll_exc_copy\@)
>> 460 LDREST(t2, REST(2)(src), .Ll_exc_copy\@)
>> 461 LDREST(t3, REST(3)(src), .Ll_exc_copy\@)
>> 462 PREFS( 0, 9*32(src) ) # 0 is PREF_LOAD (not streamed)
>> 463 ADD src, src, 4*NBYTES
>> 464 #ifdef CONFIG_CPU_SB1
>> 465 nop # improves slotting
>> 466 #endif
>> 467 STORE(t0, UNIT(0)(dst), .Ls_exc_p4u\@)
>> 468 STORE(t1, UNIT(1)(dst), .Ls_exc_p3u\@)
>> 469 STORE(t2, UNIT(2)(dst), .Ls_exc_p2u\@)
>> 470 STORE(t3, UNIT(3)(dst), .Ls_exc_p1u\@)
>> 471 PREFD( 1, 9*32(dst) ) # 1 is PREF_STORE (not streamed)
>> 472 .set reorder /* DADDI_WAR */
>> 473 ADD dst, dst, 4*NBYTES
>> 474 bne len, rem, 1b
>> 475 .set noreorder
>> 476
>> 477 .Lcleanup_src_unaligned\@:
>> 478 beqz len, .Ldone\@
>> 479 and rem, len, NBYTES-1 # rem = len % NBYTES
>> 480 beq rem, len, .Lcopy_bytes\@
>> 481 nop
>> 482 1:
>> 483 R10KCBARRIER(0(ra))
>> 484 LDFIRST(t0, FIRST(0)(src), .Ll_exc\@)
>> 485 LDREST(t0, REST(0)(src), .Ll_exc_copy\@)
>> 486 ADD src, src, NBYTES
>> 487 SUB len, len, NBYTES
>> 488 STORE(t0, 0(dst), .Ls_exc_p1u\@)
>> 489 .set reorder /* DADDI_WAR */
>> 490 ADD dst, dst, NBYTES
>> 491 bne len, rem, 1b
>> 492 .set noreorder
>> 493
>> 494 #endif /* !CONFIG_CPU_NO_LOAD_STORE_LR */
>> 495 .Lcopy_bytes_checklen\@:
>> 496 beqz len, .Ldone\@
>> 497 nop
>> 498 .Lcopy_bytes\@:
>> 499 /* 0 < len < NBYTES */
>> 500 R10KCBARRIER(0(ra))
>> 501 #define COPY_BYTE(N) \
>> 502 LOADB(t0, N(src), .Ll_exc\@); \
>> 503 SUB len, len, 1; \
>> 504 beqz len, .Ldone\@; \
>> 505 STOREB(t0, N(dst), .Ls_exc_p1\@)
>> 506
>> 507 COPY_BYTE(0)
>> 508 COPY_BYTE(1)
>> 509 #ifdef USE_DOUBLE
>> 510 COPY_BYTE(2)
>> 511 COPY_BYTE(3)
>> 512 COPY_BYTE(4)
>> 513 COPY_BYTE(5)
>> 514 #endif
>> 515 LOADB(t0, NBYTES-2(src), .Ll_exc\@)
>> 516 SUB len, len, 1
>> 517 jr ra
>> 518 STOREB(t0, NBYTES-2(dst), .Ls_exc_p1\@)
>> 519 .Ldone\@:
>> 520 jr ra
>> 521 nop
>> 522
>> 523 #ifdef CONFIG_CPU_NO_LOAD_STORE_LR
>> 524 .Lcopy_unaligned_bytes\@:
>> 525 1:
>> 526 COPY_BYTE(0)
>> 527 COPY_BYTE(1)
>> 528 COPY_BYTE(2)
>> 529 COPY_BYTE(3)
>> 530 COPY_BYTE(4)
>> 531 COPY_BYTE(5)
>> 532 COPY_BYTE(6)
>> 533 COPY_BYTE(7)
>> 534 ADD src, src, 8
>> 535 b 1b
>> 536 ADD dst, dst, 8
>> 537 #endif /* CONFIG_CPU_NO_LOAD_STORE_LR */
>> 538 .if __memcpy == 1
>> 539 END(memcpy)
>> 540 .set __memcpy, 0
>> 541 .hidden __memcpy
>> 542 .endif
>> 543
>> 544 .Ll_exc_copy\@:
>> 545 /*
>> 546 * Copy bytes from src until faulting load address (or until a
>> 547 * lb faults)
>> 548 *
>> 549 * When reached by a faulting LDFIRST/LDREST, THREAD_BUADDR($28)
>> 550 * may be more than a byte beyond the last address.
>> 551 * Hence, the lb below may get an exception.
>> 552 *
>> 553 * Assumes src < THREAD_BUADDR($28)
>> 554 */
>> 555 LOADK t0, TI_TASK($28)
>> 556 nop
>> 557 LOADK t0, THREAD_BUADDR(t0)
>> 558 1:
>> 559 LOADB(t1, 0(src), .Ll_exc\@)
>> 560 ADD src, src, 1
>> 561 sb t1, 0(dst) # can't fault -- we're copy_from_user
>> 562 .set reorder /* DADDI_WAR */
>> 563 ADD dst, dst, 1
>> 564 bne src, t0, 1b
>> 565 .set noreorder
>> 566 .Ll_exc\@:
>> 567 LOADK t0, TI_TASK($28)
>> 568 nop
>> 569 LOADK t0, THREAD_BUADDR(t0) # t0 is just past last good address
>> 570 nop
>> 571 SUB len, AT, t0 # len number of uncopied bytes
>> 572 jr ra
>> 573 nop
>> 574
>> 575 #define SEXC(n) \
>> 576 .set reorder; /* DADDI_WAR */ \
>> 577 .Ls_exc_p ## n ## u\@: \
>> 578 ADD len, len, n*NBYTES; \
>> 579 jr ra; \
>> 580 .set noreorder
>> 581
>> 582 SEXC(8)
>> 583 SEXC(7)
>> 584 SEXC(6)
>> 585 SEXC(5)
>> 586 SEXC(4)
>> 587 SEXC(3)
>> 588 SEXC(2)
>> 589 SEXC(1)
>> 590
>> 591 .Ls_exc_p1\@:
>> 592 .set reorder /* DADDI_WAR */
>> 593 ADD len, len, 1
>> 594 jr ra
>> 595 .set noreorder
>> 596 .Ls_exc\@:
>> 597 jr ra
>> 598 nop
>> 599 .endm
>> 600
>> 601 #ifndef CONFIG_HAVE_PLAT_MEMCPY
>> 602 .align 5
>> 603 LEAF(memmove)
>> 604 EXPORT_SYMBOL(memmove)
>> 605 ADD t0, a0, a2
>> 606 ADD t1, a1, a2
>> 607 sltu t0, a1, t0 # dst + len <= src -> memcpy
>> 608 sltu t1, a0, t1 # dst >= src + len -> memcpy
>> 609 and t0, t1
>> 610 beqz t0, .L__memcpy
>> 611 move v0, a0 /* return value */
>> 612 beqz a2, .Lr_out
>> 613 END(memmove)
>> 614
>> 615 /* fall through to __rmemcpy */
>> 616 LEAF(__rmemcpy) /* a0=dst a1=src a2=len */
>> 617 sltu t0, a1, a0
>> 618 beqz t0, .Lr_end_bytes_up # src >= dst
>> 619 nop
>> 620 ADD a0, a2 # dst = dst + len
>> 621 ADD a1, a2 # src = src + len
>> 622
>> 623 .Lr_end_bytes:
>> 624 R10KCBARRIER(0(ra))
>> 625 lb t0, -1(a1)
>> 626 SUB a2, a2, 0x1
>> 627 sb t0, -1(a0)
>> 628 SUB a1, a1, 0x1
>> 629 .set reorder /* DADDI_WAR */
>> 630 SUB a0, a0, 0x1
>> 631 bnez a2, .Lr_end_bytes
>> 632 .set noreorder
>> 633
>> 634 .Lr_out:
>> 635 jr ra
>> 636 move a2, zero
>> 637
>> 638 .Lr_end_bytes_up:
>> 639 R10KCBARRIER(0(ra))
>> 640 lb t0, (a1)
>> 641 SUB a2, a2, 0x1
>> 642 sb t0, (a0)
>> 643 ADD a1, a1, 0x1
>> 644 .set reorder /* DADDI_WAR */
>> 645 ADD a0, a0, 0x1
>> 646 bnez a2, .Lr_end_bytes_up
>> 647 .set noreorder
>> 648
>> 649 jr ra
>> 650 move a2, zero
>> 651 END(__rmemcpy)
>> 652
>> 653 /*
>> 654 * A combined memcpy/__copy_user
>> 655 * __copy_user sets len to 0 for success; else to an upper bound of
>> 656 * the number of uncopied bytes.
>> 657 * memcpy sets v0 to dst.
>> 658 */
>> 659 .align 5
>> 660 LEAF(memcpy) /* a0=dst a1=src a2=len */
>> 661 EXPORT_SYMBOL(memcpy)
>> 662 move v0, dst /* return value */
>> 663 .L__memcpy:
>> 664 #ifndef CONFIG_EVA
>> 665 FEXPORT(__raw_copy_from_user)
>> 666 EXPORT_SYMBOL(__raw_copy_from_user)
>> 667 FEXPORT(__raw_copy_to_user)
>> 668 EXPORT_SYMBOL(__raw_copy_to_user)
>> 669 #endif
>> 670 /* Legacy Mode, user <-> user */
>> 671 __BUILD_COPY_USER LEGACY_MODE USEROP USEROP
>> 672
>> 673 #endif
>> 674
>> 675 #ifdef CONFIG_EVA
>> 676
>> 677 /*
>> 678 * For EVA we need distinct symbols for reading and writing to user space.
>> 679 * This is because we need to use specific EVA instructions to perform the
>> 680 * virtual <-> physical translation when a virtual address is actually in user
>> 681 * space
>> 682 */
>> 683
>> 684 /*
>> 685 * __copy_from_user (EVA)
>> 686 */
>> 687
>> 688 LEAF(__raw_copy_from_user)
>> 689 EXPORT_SYMBOL(__raw_copy_from_user)
>> 690 __BUILD_COPY_USER EVA_MODE USEROP KERNELOP
>> 691 END(__raw_copy_from_user)
>> 692
>> 693
>> 694
>> 695 /*
>> 696 * __copy_to_user (EVA)
134 */ 697 */
135 698
>> 699 LEAF(__raw_copy_to_user)
>> 700 EXPORT_SYMBOL(__raw_copy_to_user)
>> 701 __BUILD_COPY_USER EVA_MODE KERNELOP USEROP
>> 702 END(__raw_copy_to_user)
136 703
137 #define ptr_out R0 /* destinatio !! 704 #endif
138 #define ptr_in R1 /* source poi <<
139 #define len R2 /* length of <<
140 <<
141 #define data70 R13:12 /* lo 8 bytes <<
142 #define dataF8 R11:10 /* hi 8 bytes <<
143 #define ldata0 R7:6 /* even 8 byt <<
144 #define ldata1 R25:24 /* odd 8 byte <<
145 #define data1 R7 /* lower 8 by <<
146 #define data0 R6 /* lower 8 by <<
147 <<
148 #define ifbyte p0 /* if transfe <<
149 #define ifhword p0 /* if transfe <<
150 #define ifword p0 /* if transfe <<
151 #define noprolog p0 /* no prolog, <<
152 #define nokernel p1 /* no 32byte <<
153 #define noepilog p0 /* no epilog, <<
154 #define align p2 /* alignment <<
155 #define kernel1 p0 /* kernel cou <<
156 <<
157 #define dalign R25 /* rel alignm <<
158 #define star3 R16 /* number byt <<
159 #define rest R8 /* length - p <<
160 #define back R7 /* nr bytes > <<
161 #define epilog R3 /* bytes in e <<
162 #define inc R15:14 /* inc kernel <<
163 #define kernel R4 /* number of <<
164 #define ptr_in_p_128 R5 /* pointer fo <<
165 #define mask R8 /* mask used <<
166 #define shift R8 /* used to wo <<
167 #define shift2 R5 /* in epilog <<
168 #define prolog R15 /* bytes in <<
169 #define epilogdws R15 /* number dwo <<
170 #define shiftb R14 /* used to ex <<
171 #define offset R9 /* same as al <<
172 #define ptr_out_p_32 R17 /* pointer to <<
173 #define align888 R14 /* if simple <<
174 #define len8 R9 /* number of <<
175 #define over R20 /* nr of byte <<
176 <<
177 #define ptr_in_p_128kernel R5:4 /* pa <<
178 <<
179 .section .text <<
180 .p2align 4 <<
181 .global memcpy <<
182 .type memcpy, @function <<
183 memcpy: <<
184 { <<
185 p2 = cmp.eq(len, #0); /* =0 <<
186 align888 = or(ptr_in, ptr_out); /* %8 <<
187 p0 = cmp.gtu(len, #23); /* %1 <<
188 p1 = cmp.eq(ptr_in, ptr_out); /* at <<
189 } <<
190 { <<
191 p1 = or(p2, p1); <<
192 p3 = cmp.gtu(len, #95); /* %8 <<
193 align888 = or(align888, len); /* %8 <<
194 len8 = lsr(len, #3); /* %8 <<
195 } <<
196 { <<
197 dcfetch(ptr_in); /* ze <<
198 p2 = bitsclr(align888, #7); /* %8 <<
199 if(p1) jumpr r31; /* =0 <<
200 } <<
201 { <<
202 p2 = and(p2,!p3); <<
203 if (p2.new) len = add(len, #-8); <<
204 if (p2.new) jump:NT .Ldwordaligned; <<
205 } <<
206 { <<
207 if(!p0) jump .Lbytes23orless; /* %1 <<
208 mask.l = #LO(0x7fffffff); <<
209 /* all bytes before line multiples of <<
210 prolog = sub(#0, ptr_out); <<
211 } <<
212 { <<
213 /* save r31 on stack, decrement sp by <<
214 allocframe(#24); <<
215 mask.h = #HI(0x7fffffff); <<
216 ptr_in_p_128 = add(ptr_in, #32); <<
217 back = cl0(len); <<
218 } <<
219 { <<
220 memd(sp+#0) = R17:16; /* sa <<
221 r31.l = #LO(.Lmemcpy_return); /* se <<
222 prolog &= lsr(mask, back); <<
223 offset = and(ptr_in, #7); <<
224 } <<
225 { <<
226 memd(sp+#8) = R25:24; /* sa <<
227 dalign = sub(ptr_out, ptr_in); <<
228 r31.h = #HI(.Lmemcpy_return); /* se <<
229 } <<
230 { <<
231 /* see if there if input buffer end i <<
232 over = add(len, ptr_in); <<
233 back = add(len, offset); <<
234 memd(sp+#16) = R21:20; /* sa <<
235 } <<
236 { <<
237 noprolog = bitsclr(prolog, #7); <<
238 prolog = and(prolog, #31); <<
239 dcfetch(ptr_in_p_128); <<
240 ptr_in_p_128 = add(ptr_in_p_128, #32); <<
241 } <<
242 { <<
243 kernel = sub(len, prolog); <<
244 shift = asl(prolog, #3); <<
245 star3 = and(prolog, #7); <<
246 ptr_in = and(ptr_in, #-8); <<
247 } <<
248 { <<
249 prolog = lsr(prolog, #3); <<
250 epilog = and(kernel, #31); <<
251 ptr_out_p_32 = add(ptr_out, prolog); <<
252 over = and(over, #7); <<
253 } <<
254 { <<
255 p3 = cmp.gtu(back, #8); <<
256 kernel = lsr(kernel, #5); <<
257 dcfetch(ptr_in_p_128); <<
258 ptr_in_p_128 = add(ptr_in_p_128, #32); <<
259 } <<
260 { <<
261 p1 = cmp.eq(prolog, #0); <<
262 if(!p1.new) prolog = add(prolog, #1); <<
263 dcfetch(ptr_in_p_128); /* reserve th <<
264 ptr_in_p_128 = add(ptr_in_p_128, #32); <<
265 } <<
266 { <<
267 nokernel = cmp.eq(kernel,#0); <<
268 dcfetch(ptr_in_p_128); /* reserve the <<
269 ptr_in_p_128 = add(ptr_in_p_128, #32); <<
270 shiftb = and(shift, #8); <<
271 } <<
272 { <<
273 dcfetch(ptr_in_p_128); /* re <<
274 ptr_in_p_128 = add(ptr_in_p_128, #32); <<
275 if(nokernel) jump .Lskip64; <<
276 p2 = cmp.eq(kernel, #1); /* sk <<
277 } <<
278 { <<
279 dczeroa(ptr_out_p_32); <<
280 /* don't advance pointer */ <<
281 if(!p2) ptr_out_p_32 = add(ptr_out_p_3 <<
282 } <<
283 { <<
284 dalign = and(dalign, #31); <<
285 dczeroa(ptr_out_p_32); <<
286 } <<
287 .Lskip64: <<
288 { <<
289 data70 = memd(ptr_in++#16); <<
290 if(p3) dataF8 = memd(ptr_in+#8); <<
291 if(noprolog) jump .Lnoprolog32; <<
292 align = offset; <<
293 } <<
294 /* upto initial 7 bytes */ <<
295 { <<
296 ldata0 = valignb(dataF8, data70, align <<
297 ifbyte = tstbit(shift,#3); <<
298 offset = add(offset, star3); <<
299 } <<
300 { <<
301 if(ifbyte) memb(ptr_out++#1) = data0; <<
302 ldata0 = lsr(ldata0, shiftb); <<
303 shiftb = and(shift, #16); <<
304 ifhword = tstbit(shift,#4); <<
305 } <<
306 { <<
307 if(ifhword) memh(ptr_out++#2) = data0; <<
308 ldata0 = lsr(ldata0, shiftb); <<
309 ifword = tstbit(shift,#5); <<
310 p2 = cmp.gtu(offset, #7); <<
311 } <<
312 { <<
313 if(ifword) memw(ptr_out++#4) = data0; <<
314 if(p2) data70 = dataF8; <<
315 if(p2) dataF8 = memd(ptr_in++#8); <<
316 align = offset; <<
317 } <<
318 .Lnoprolog32: <<
319 { <<
320 p3 = sp1loop0(.Ldword_loop_prolog, pro <<
321 rest = sub(len, star3); /* whats left <<
322 p0 = cmp.gt(over, #0); <<
323 } <<
324 if(p0) rest = add(rest, #16); <<
325 .Ldword_loop_prolog: <<
326 { <<
327 if(p3) memd(ptr_out++#8) = ldata0; <<
328 ldata0 = valignb(dataF8, data70, align <<
329 p0 = cmp.gt(rest, #16); <<
330 } <<
331 { <<
332 data70 = dataF8; <<
333 if(p0) dataF8 = memd(ptr_in++#8); <<
334 rest = add(rest, #-8); <<
335 }:endloop0 <<
336 .Lkernel: <<
337 { <<
338 /* kernel is at least 32bytes */ <<
339 p3 = cmp.gtu(kernel, #0); <<
340 /* last itn. remove edge effects */ <<
341 if(p3.new) kernel = add(kernel, #-1); <<
342 /* dealt with in last dword loop */ <<
343 if(p3.new) epilog = add(epilog, #32); <<
344 } <<
345 { <<
346 nokernel = cmp.eq(kernel, #0); <<
347 if(nokernel.new) jump:NT .Lepilog; <<
348 inc = combine(#32, #-1); <<
349 p3 = cmp.gtu(dalign, #24); <<
350 } <<
351 { <<
352 if(p3) jump .Lodd_alignment; <<
353 } <<
354 { <<
355 loop0(.Loword_loop_25to31, kernel); <<
356 kernel1 = cmp.gtu(kernel, #1); <<
357 rest = kernel; <<
358 } <<
359 .falign <<
360 .Loword_loop_25to31: <<
361 { <<
362 dcfetch(ptr_in_p_128); /* prefetch 4 <<
363 if(kernel1) ptr_out_p_32 = add(ptr_out <<
364 } <<
365 { <<
366 dczeroa(ptr_out_p_32); /* reserve th <<
367 p3 = cmp.eq(kernel, rest); <<
368 } <<
369 { <<
370 /* kernel -= 1 */ <<
371 ptr_in_p_128kernel = vaddw(ptr_in_p_12 <<
372 /* kill write on first iteration */ <<
373 if(!p3) memd(ptr_out++#8) = ldata1; <<
374 ldata1 = valignb(dataF8, data70, align <<
375 data70 = memd(ptr_in++#8); <<
376 } <<
377 { <<
378 memd(ptr_out++#8) = ldata0; <<
379 ldata0 = valignb(data70, dataF8, align <<
380 dataF8 = memd(ptr_in++#8); <<
381 } <<
382 { <<
383 memd(ptr_out++#8) = ldata1; <<
384 ldata1 = valignb(dataF8, data70, align <<
385 data70 = memd(ptr_in++#8); <<
386 } <<
387 { <<
388 memd(ptr_out++#8) = ldata0; <<
389 ldata0 = valignb(data70, dataF8, align <<
390 dataF8 = memd(ptr_in++#8); <<
391 kernel1 = cmp.gtu(kernel, #1); <<
392 }:endloop0 <<
393 { <<
394 memd(ptr_out++#8) = ldata1; <<
395 jump .Lepilog; <<
396 } <<
397 .Lodd_alignment: <<
398 { <<
399 loop0(.Loword_loop_00to24, kernel); <<
400 kernel1 = cmp.gtu(kernel, #1); <<
401 rest = add(kernel, #-1); <<
402 } <<
403 .falign <<
404 .Loword_loop_00to24: <<
405 { <<
406 dcfetch(ptr_in_p_128); /* prefetch 4 <<
407 ptr_in_p_128kernel = vaddw(ptr_in_p_12 <<
408 if(kernel1) ptr_out_p_32 = add(ptr_out <<
409 } <<
410 { <<
411 dczeroa(ptr_out_p_32); /* reserve th <<
412 } <<
413 { <<
414 memd(ptr_out++#8) = ldata0; <<
415 ldata0 = valignb(dataF8, data70, align <<
416 data70 = memd(ptr_in++#8); <<
417 } <<
418 { <<
419 memd(ptr_out++#8) = ldata0; <<
420 ldata0 = valignb(data70, dataF8, align <<
421 dataF8 = memd(ptr_in++#8); <<
422 } <<
423 { <<
424 memd(ptr_out++#8) = ldata0; <<
425 ldata0 = valignb(dataF8, data70, align <<
426 data70 = memd(ptr_in++#8); <<
427 } <<
428 { <<
429 memd(ptr_out++#8) = ldata0; <<
430 ldata0 = valignb(data70, dataF8, align <<
431 dataF8 = memd(ptr_in++#8); <<
432 kernel1 = cmp.gtu(kernel, #1); <<
433 }:endloop0 <<
434 .Lepilog: <<
435 { <<
436 noepilog = cmp.eq(epilog,#0); <<
437 epilogdws = lsr(epilog, #3); <<
438 kernel = and(epilog, #7); <<
439 } <<
440 { <<
441 if(noepilog) jumpr r31; <<
442 if(noepilog) ptr_out = sub(ptr_out, le <<
443 p3 = cmp.eq(epilogdws, #0); <<
444 shift2 = asl(epilog, #3); <<
445 } <<
446 { <<
447 shiftb = and(shift2, #32); <<
448 ifword = tstbit(epilog,#2); <<
449 if(p3) jump .Lepilog60; <<
450 if(!p3) epilog = add(epilog, #-16); <<
451 } <<
452 { <<
453 loop0(.Ldword_loop_epilog, epilogdws); <<
454 /* stop criteria is lsbs unless = 0 t <<
455 p3 = cmp.eq(kernel, #0); <<
456 if(p3.new) kernel= #8; <<
457 p1 = cmp.gt(over, #0); <<
458 } <<
459 /* if not aligned to end of buffer ex <<
460 if(p1) kernel= #0; <<
461 .Ldword_loop_epilog: <<
462 { <<
463 memd(ptr_out++#8) = ldata0; <<
464 ldata0 = valignb(dataF8, data70, align <<
465 p3 = cmp.gt(epilog, kernel); <<
466 } <<
467 { <<
468 data70 = dataF8; <<
469 if(p3) dataF8 = memd(ptr_in++#8); <<
470 epilog = add(epilog, #-8); <<
471 }:endloop0 <<
472 /* copy last 7 bytes */ <<
473 .Lepilog60: <<
474 { <<
475 if(ifword) memw(ptr_out++#4) = data0; <<
476 ldata0 = lsr(ldata0, shiftb); <<
477 ifhword = tstbit(epilog,#1); <<
478 shiftb = and(shift2, #16); <<
479 } <<
480 { <<
481 if(ifhword) memh(ptr_out++#2) = data0; <<
482 ldata0 = lsr(ldata0, shiftb); <<
483 ifbyte = tstbit(epilog,#0); <<
484 if(ifbyte.new) len = add(len, #-1); <<
485 } <<
486 { <<
487 if(ifbyte) memb(ptr_out) = data0; <<
488 ptr_out = sub(ptr_out, len); /* re <<
489 jumpr r31; <<
490 } <<
491 /* do byte copy for small n */ <<
492 .Lbytes23orless: <<
493 { <<
494 p3 = sp1loop0(.Lbyte_copy, len); <<
495 len = add(len, #-1); <<
496 } <<
497 .Lbyte_copy: <<
498 { <<
499 data0 = memb(ptr_in++#1); <<
500 if(p3) memb(ptr_out++#1) = data0; <<
501 }:endloop0 <<
502 { <<
503 memb(ptr_out) = data0; <<
504 ptr_out = sub(ptr_out, len); <<
505 jumpr r31; <<
506 } <<
507 /* do dword copies for aligned in, out and le <<
508 .Ldwordaligned: <<
509 { <<
510 p3 = sp1loop0(.Ldword_copy, len8); <<
511 } <<
512 .Ldword_copy: <<
513 { <<
514 if(p3) memd(ptr_out++#8) = ldata0; <<
515 ldata0 = memd(ptr_in++#8); <<
516 }:endloop0 <<
517 { <<
518 memd(ptr_out) = ldata0; <<
519 ptr_out = sub(ptr_out, len); <<
520 jumpr r31; /* return to function <<
521 } <<
522 .Lmemcpy_return: <<
523 r21:20 = memd(sp+#16); /* restore r2 <<
524 { <<
525 r25:24 = memd(sp+#8); /* restore r2 <<
526 r17:16 = memd(sp+#0); /* restore r1 <<
527 } <<
528 deallocframe; /* restore r31 and in <<
529 jumpr r31 <<
Linux® is a registered trademark of Linus Torvalds in the United States and other countries.
TOMOYO® is a registered trademark of NTT DATA CORPORATION.