1 /* SPDX-License-Identifier: GPL-2.0 */ !! 1 /* 2 /* memcpy.S: Sparc optimized memcpy and memmov !! 2 * This file is subject to the terms and conditions of the GNU General Public 3 * Hand optimized from GNU libc's memcpy and m !! 3 * License. See the file "COPYING" in the main directory of this archive 4 * Copyright (C) 1991,1996 Free Software Found !! 4 * for more details. 5 * Copyright (C) 1995 Linus Torvalds (Linus.To !! 5 * 6 * Copyright (C) 1996 David S. Miller (davem@c !! 6 * Unified implementation of memcpy, memmove and the __copy_user backend. 7 * Copyright (C) 1996 Eddie C. Dost (ecd@skyne !! 7 * 8 * Copyright (C) 1996 Jakub Jelinek (jj@sunsit !! 8 * Copyright (C) 1998, 99, 2000, 01, 2002 Ralf Baechle (ralf@gnu.org) 9 */ !! 9 * Copyright (C) 1999, 2000, 01, 2002 Silicon Graphics, Inc. 10 !! 10 * Copyright (C) 2002 Broadcom, Inc. 11 #include <linux/export.h> !! 11 * memcpy/copy_user author: Mark Vandevoorde 12 !! 12 * Copyright (C) 2007 Maciej W. Rozycki 13 #define FUNC(x) \ !! 13 * Copyright (C) 2014 Imagination Technologies Ltd. 14 .globl x; \ !! 14 * 15 .type x,@function; \ !! 15 * Mnemonic names for arguments to memcpy/__copy_user 16 .align 4; \ !! 16 */ 17 x: !! 17 18 !! 18 /* 19 /* Both these macros have to start with exactl !! 19 * Hack to resolve longstanding prefetch issue 20 #define MOVE_BIGCHUNK(src, dst, offset, t0, t1 !! 20 * 21 ldd [%src + (offset) + 0x00], %t0; !! 21 * Prefetching may be fatal on some systems if we're prefetching beyond the 22 ldd [%src + (offset) + 0x08], %t2; !! 22 * end of memory on some systems. It's also a seriously bad idea on non 23 ldd [%src + (offset) + 0x10], %t4; !! 23 * dma-coherent systems. 24 ldd [%src + (offset) + 0x18], %t6; !! 24 */ 25 st %t0, [%dst + (offset) + 0x00]; !! 25 #ifdef CONFIG_DMA_NONCOHERENT 26 st %t1, [%dst + (offset) + 0x04]; !! 26 #undef CONFIG_CPU_HAS_PREFETCH 27 st %t2, [%dst + (offset) + 0x08]; !! 27 #endif 28 st %t3, [%dst + (offset) + 0x0c]; !! 28 #ifdef CONFIG_MIPS_MALTA 29 st %t4, [%dst + (offset) + 0x10]; !! 29 #undef CONFIG_CPU_HAS_PREFETCH 30 st %t5, [%dst + (offset) + 0x14]; !! 30 #endif 31 st %t6, [%dst + (offset) + 0x18]; !! 31 32 st %t7, [%dst + (offset) + 0x1c]; !! 32 #include <asm/asm.h> 33 !! 33 #include <asm/asm-offsets.h> 34 #define MOVE_BIGALIGNCHUNK(src, dst, offset, t !! 34 #include <asm/regdef.h> 35 ldd [%src + (offset) + 0x00], %t0; !! 35 36 ldd [%src + (offset) + 0x08], %t2; !! 36 #define dst a0 37 ldd [%src + (offset) + 0x10], %t4; !! 37 #define src a1 38 ldd [%src + (offset) + 0x18], %t6; !! 38 #define len a2 39 std %t0, [%dst + (offset) + 0x00]; !! 39 40 std %t2, [%dst + (offset) + 0x08]; !! 40 /* 41 std %t4, [%dst + (offset) + 0x10]; !! 41 * Spec 42 std %t6, [%dst + (offset) + 0x18]; !! 42 * 43 !! 43 * memcpy copies len bytes from src to dst and sets v0 to dst. 44 #define MOVE_LASTCHUNK(src, dst, offset, t0, t !! 44 * It assumes that 45 ldd [%src - (offset) - 0x10], %t0; !! 45 * - src and dst don't overlap 46 ldd [%src - (offset) - 0x08], %t2; !! 46 * - src is readable 47 st %t0, [%dst - (offset) - 0x10]; !! 47 * - dst is writable 48 st %t1, [%dst - (offset) - 0x0c]; !! 48 * memcpy uses the standard calling convention 49 st %t2, [%dst - (offset) - 0x08]; !! 49 * 50 st %t3, [%dst - (offset) - 0x04]; !! 50 * __copy_user copies up to len bytes from src to dst and sets a2 (len) to 51 !! 51 * the number of uncopied bytes due to an exception caused by a read or write. 52 #define MOVE_LASTALIGNCHUNK(src, dst, offset, !! 52 * __copy_user assumes that src and dst don't overlap, and that the call is 53 ldd [%src - (offset) - 0x10], %t0; !! 53 * implementing one of the following: 54 ldd [%src - (offset) - 0x08], %t2; !! 54 * copy_to_user 55 std %t0, [%dst - (offset) - 0x10]; !! 55 * - src is readable (no exceptions when reading src) 56 std %t2, [%dst - (offset) - 0x08]; !! 56 * copy_from_user 57 !! 57 * - dst is writable (no exceptions when writing dst) 58 #define MOVE_SHORTCHUNK(src, dst, offset, t0, !! 58 * __copy_user uses a non-standard calling convention; see 59 ldub [%src - (offset) - 0x02], %t0; !! 59 * include/asm-mips/uaccess.h 60 ldub [%src - (offset) - 0x01], %t1; !! 60 * 61 stb %t0, [%dst - (offset) - 0x02]; !! 61 * When an exception happens on a load, the handler must 62 stb %t1, [%dst - (offset) - 0x01]; !! 62 # ensure that all of the destination buffer is overwritten to prevent >> 63 * leaking information to user mode programs. >> 64 */ >> 65 >> 66 /* >> 67 * Implementation >> 68 */ >> 69 >> 70 /* >> 71 * The exception handler for loads requires that: >> 72 * 1- AT contain the address of the byte just past the end of the source >> 73 * of the copy, >> 74 * 2- src_entry <= src < AT, and >> 75 * 3- (dst - src) == (dst_entry - src_entry), >> 76 * The _entry suffix denotes values when __copy_user was called. >> 77 * >> 78 * (1) is set up up by uaccess.h and maintained by not writing AT in copy_user >> 79 * (2) is met by incrementing src by the number of bytes copied >> 80 * (3) is met by not doing loads between a pair of increments of dst and src >> 81 * >> 82 * The exception handlers for stores adjust len (if necessary) and return. >> 83 * These handlers do not need to overwrite any data. >> 84 * >> 85 * For __rmemcpy and memmove an exception is always a kernel bug, therefore >> 86 * they're not protected. >> 87 */ >> 88 >> 89 /* Instruction type */ >> 90 #define LD_INSN 1 >> 91 #define ST_INSN 2 >> 92 /* Pretech type */ >> 93 #define SRC_PREFETCH 1 >> 94 #define DST_PREFETCH 2 >> 95 #define LEGACY_MODE 1 >> 96 #define EVA_MODE 2 >> 97 #define USEROP 1 >> 98 #define KERNELOP 2 >> 99 >> 100 /* >> 101 * Wrapper to add an entry in the exception table >> 102 * in case the insn causes a memory exception. >> 103 * Arguments: >> 104 * insn : Load/store instruction >> 105 * type : Instruction type >> 106 * reg : Register >> 107 * addr : Address >> 108 * handler : Exception handler >> 109 */ >> 110 >> 111 #define EXC(insn, type, reg, addr, handler) \ >> 112 .if \mode == LEGACY_MODE; \ >> 113 9: insn reg, addr; \ >> 114 .section __ex_table,"a"; \ >> 115 PTR 9b, handler; \ >> 116 .previous; \ >> 117 /* This is assembled in EVA mode */ \ >> 118 .else; \ >> 119 /* If loading from user or storing to user */ \ >> 120 .if ((\from == USEROP) && (type == LD_INSN)) || \ >> 121 ((\to == USEROP) && (type == ST_INSN)); \ >> 122 9: __BUILD_EVA_INSN(insn##e, reg, addr); \ >> 123 .section __ex_table,"a"; \ >> 124 PTR 9b, handler; \ >> 125 .previous; \ >> 126 .else; \ >> 127 /* \ >> 128 * Still in EVA, but no need for \ >> 129 * exception handler or EVA insn \ >> 130 */ \ >> 131 insn reg, addr; \ >> 132 .endif; \ >> 133 .endif >> 134 >> 135 /* >> 136 * Only on the 64-bit kernel we can made use of 64-bit registers. >> 137 */ >> 138 #ifdef CONFIG_64BIT >> 139 #define USE_DOUBLE >> 140 #endif >> 141 >> 142 #ifdef USE_DOUBLE >> 143 >> 144 #define LOADK ld /* No exception */ >> 145 #define LOAD(reg, addr, handler) EXC(ld, LD_INSN, reg, addr, handler) >> 146 #define LOADL(reg, addr, handler) EXC(ldl, LD_INSN, reg, addr, handler) >> 147 #define LOADR(reg, addr, handler) EXC(ldr, LD_INSN, reg, addr, handler) >> 148 #define STOREL(reg, addr, handler) EXC(sdl, ST_INSN, reg, addr, handler) >> 149 #define STORER(reg, addr, handler) EXC(sdr, ST_INSN, reg, addr, handler) >> 150 #define STORE(reg, addr, handler) EXC(sd, ST_INSN, reg, addr, handler) >> 151 #define ADD daddu >> 152 #define SUB dsubu >> 153 #define SRL dsrl >> 154 #define SRA dsra >> 155 #define SLL dsll >> 156 #define SLLV dsllv >> 157 #define SRLV dsrlv >> 158 #define NBYTES 8 >> 159 #define LOG_NBYTES 3 >> 160 >> 161 /* >> 162 * As we are sharing code base with the mips32 tree (which use the o32 ABI >> 163 * register definitions). We need to redefine the register definitions from >> 164 * the n64 ABI register naming to the o32 ABI register naming. >> 165 */ >> 166 #undef t0 >> 167 #undef t1 >> 168 #undef t2 >> 169 #undef t3 >> 170 #define t0 $8 >> 171 #define t1 $9 >> 172 #define t2 $10 >> 173 #define t3 $11 >> 174 #define t4 $12 >> 175 #define t5 $13 >> 176 #define t6 $14 >> 177 #define t7 $15 >> 178 >> 179 #else >> 180 >> 181 #define LOADK lw /* No exception */ >> 182 #define LOAD(reg, addr, handler) EXC(lw, LD_INSN, reg, addr, handler) >> 183 #define LOADL(reg, addr, handler) EXC(lwl, LD_INSN, reg, addr, handler) >> 184 #define LOADR(reg, addr, handler) EXC(lwr, LD_INSN, reg, addr, handler) >> 185 #define STOREL(reg, addr, handler) EXC(swl, ST_INSN, reg, addr, handler) >> 186 #define STORER(reg, addr, handler) EXC(swr, ST_INSN, reg, addr, handler) >> 187 #define STORE(reg, addr, handler) EXC(sw, ST_INSN, reg, addr, handler) >> 188 #define ADD addu >> 189 #define SUB subu >> 190 #define SRL srl >> 191 #define SLL sll >> 192 #define SRA sra >> 193 #define SLLV sllv >> 194 #define SRLV srlv >> 195 #define NBYTES 4 >> 196 #define LOG_NBYTES 2 >> 197 >> 198 #endif /* USE_DOUBLE */ >> 199 >> 200 #define LOADB(reg, addr, handler) EXC(lb, LD_INSN, reg, addr, handler) >> 201 #define STOREB(reg, addr, handler) EXC(sb, ST_INSN, reg, addr, handler) >> 202 >> 203 #define _PREF(hint, addr, type) \ >> 204 .if \mode == LEGACY_MODE; \ >> 205 PREF(hint, addr); \ >> 206 .else; \ >> 207 .if ((\from == USEROP) && (type == SRC_PREFETCH)) || \ >> 208 ((\to == USEROP) && (type == DST_PREFETCH)); \ >> 209 /* \ >> 210 * PREFE has only 9 bits for the offset \ >> 211 * compared to PREF which has 16, so it may \ >> 212 * need to use the $at register but this \ >> 213 * register should remain intact because it's \ >> 214 * used later on. Therefore use $v1. \ >> 215 */ \ >> 216 .set at=v1; \ >> 217 PREFE(hint, addr); \ >> 218 .set noat; \ >> 219 .else; \ >> 220 PREF(hint, addr); \ >> 221 .endif; \ >> 222 .endif >> 223 >> 224 #define PREFS(hint, addr) _PREF(hint, addr, SRC_PREFETCH) >> 225 #define PREFD(hint, addr) _PREF(hint, addr, DST_PREFETCH) >> 226 >> 227 #ifdef CONFIG_CPU_LITTLE_ENDIAN >> 228 #define LDFIRST LOADR >> 229 #define LDREST LOADL >> 230 #define STFIRST STORER >> 231 #define STREST STOREL >> 232 #define SHIFT_DISCARD SLLV >> 233 #else >> 234 #define LDFIRST LOADL >> 235 #define LDREST LOADR >> 236 #define STFIRST STOREL >> 237 #define STREST STORER >> 238 #define SHIFT_DISCARD SRLV >> 239 #endif >> 240 >> 241 #define FIRST(unit) ((unit)*NBYTES) >> 242 #define REST(unit) (FIRST(unit)+NBYTES-1) >> 243 #define UNIT(unit) FIRST(unit) >> 244 >> 245 #define ADDRMASK (NBYTES-1) 63 246 64 .text 247 .text >> 248 .set noreorder >> 249 #ifndef CONFIG_CPU_DADDI_WORKAROUNDS >> 250 .set noat >> 251 #else >> 252 .set at=v1 >> 253 #endif >> 254 >> 255 .align 5 >> 256 >> 257 /* >> 258 * Macro to build the __copy_user common code >> 259 * Arguements: >> 260 * mode : LEGACY_MODE or EVA_MODE >> 261 * from : Source operand. USEROP or KERNELOP >> 262 * to : Destination operand. USEROP or KERNELOP >> 263 */ >> 264 .macro __BUILD_COPY_USER mode, from, to >> 265 >> 266 /* initialize __memcpy if this the first time we execute this macro */ >> 267 .ifnotdef __memcpy >> 268 .set __memcpy, 1 >> 269 .hidden __memcpy /* make sure it does not leak */ >> 270 .endif >> 271 >> 272 /* >> 273 * Note: dst & src may be unaligned, len may be 0 >> 274 * Temps >> 275 */ >> 276 #define rem t8 >> 277 >> 278 R10KCBARRIER(0(ra)) >> 279 /* >> 280 * The "issue break"s below are very approximate. >> 281 * Issue delays for dcache fills will perturb the schedule, as will >> 282 * load queue full replay traps, etc. >> 283 * >> 284 * If len < NBYTES use byte operations. >> 285 */ >> 286 PREFS( 0, 0(src) ) >> 287 PREFD( 1, 0(dst) ) >> 288 sltu t2, len, NBYTES >> 289 and t1, dst, ADDRMASK >> 290 PREFS( 0, 1*32(src) ) >> 291 PREFD( 1, 1*32(dst) ) >> 292 bnez t2, .Lcopy_bytes_checklen\@ >> 293 and t0, src, ADDRMASK >> 294 PREFS( 0, 2*32(src) ) >> 295 PREFD( 1, 2*32(dst) ) >> 296 #ifndef CONFIG_CPU_MIPSR6 >> 297 bnez t1, .Ldst_unaligned\@ >> 298 nop >> 299 bnez t0, .Lsrc_unaligned_dst_aligned\@ >> 300 #else >> 301 or t0, t0, t1 >> 302 bnez t0, .Lcopy_unaligned_bytes\@ >> 303 #endif >> 304 /* >> 305 * use delay slot for fall-through >> 306 * src and dst are aligned; need to compute rem >> 307 */ >> 308 .Lboth_aligned\@: >> 309 SRL t0, len, LOG_NBYTES+3 # +3 for 8 units/iter >> 310 beqz t0, .Lcleanup_both_aligned\@ # len < 8*NBYTES >> 311 and rem, len, (8*NBYTES-1) # rem = len % (8*NBYTES) >> 312 PREFS( 0, 3*32(src) ) >> 313 PREFD( 1, 3*32(dst) ) 65 .align 4 314 .align 4 66 << 67 FUNC(memmove) << 68 EXPORT_SYMBOL(memmove) << 69 cmp %o0, %o1 << 70 mov %o0, %g7 << 71 bleu 9f << 72 sub %o0, %o1, %o4 << 73 << 74 add %o1, %o2, %o3 << 75 cmp %o3, %o0 << 76 bleu 0f << 77 andcc %o4, 3, %o5 << 78 << 79 add %o1, %o2, %o1 << 80 add %o0, %o2, %o0 << 81 sub %o1, 1, %o1 << 82 sub %o0, 1, %o0 << 83 << 84 1: /* reverse_bytes */ << 85 << 86 ldub [%o1], %o4 << 87 subcc %o2, 1, %o2 << 88 stb %o4, [%o0] << 89 sub %o1, 1, %o1 << 90 bne 1b << 91 sub %o0, 1, %o0 << 92 << 93 retl << 94 mov %g7, %o0 << 95 << 96 /* NOTE: This code is executed just for the ca << 97 where %src (=%o1) & 3 is != 0. << 98 We need to align it to 4. So, for (%s << 99 1 we need to do ldub,lduh << 100 2 lduh << 101 3 just ldub << 102 so even if it looks weird, the branch << 103 are correct here. -jj << 104 */ << 105 78: /* dword_align */ << 106 << 107 andcc %o1, 1, %g0 << 108 be 4f << 109 andcc %o1, 2, %g0 << 110 << 111 ldub [%o1], %g2 << 112 add %o1, 1, %o1 << 113 stb %g2, [%o0] << 114 sub %o2, 1, %o2 << 115 bne 3f << 116 add %o0, 1, %o0 << 117 4: << 118 lduh [%o1], %g2 << 119 add %o1, 2, %o1 << 120 sth %g2, [%o0] << 121 sub %o2, 2, %o2 << 122 b 3f << 123 add %o0, 2, %o0 << 124 << 125 FUNC(memcpy) /* %o0=dst %o1=src %o2=len */ << 126 EXPORT_SYMBOL(memcpy) << 127 << 128 sub %o0, %o1, %o4 << 129 mov %o0, %g7 << 130 9: << 131 andcc %o4, 3, %o5 << 132 0: << 133 bne 86f << 134 cmp %o2, 15 << 135 << 136 bleu 90f << 137 andcc %o1, 3, %g0 << 138 << 139 bne 78b << 140 3: << 141 andcc %o1, 4, %g0 << 142 << 143 be 2f << 144 mov %o2, %g1 << 145 << 146 ld [%o1], %o4 << 147 sub %g1, 4, %g1 << 148 st %o4, [%o0] << 149 add %o1, 4, %o1 << 150 add %o0, 4, %o0 << 151 2: << 152 andcc %g1, 0xffffff80, %g0 << 153 be 3f << 154 andcc %o0, 4, %g0 << 155 << 156 be 82f + 4 << 157 5: << 158 MOVE_BIGCHUNK(o1, o0, 0x00, o2, o3, o4 << 159 MOVE_BIGCHUNK(o1, o0, 0x20, o2, o3, o4 << 160 MOVE_BIGCHUNK(o1, o0, 0x40, o2, o3, o4 << 161 MOVE_BIGCHUNK(o1, o0, 0x60, o2, o3, o4 << 162 sub %g1, 128, %g1 << 163 add %o1, 128, %o1 << 164 cmp %g1, 128 << 165 bge 5b << 166 add %o0, 128, %o0 << 167 3: << 168 andcc %g1, 0x70, %g4 << 169 be 80f << 170 andcc %g1, 8, %g0 << 171 << 172 sethi %hi(80f), %o5 << 173 srl %g4, 1, %o4 << 174 add %g4, %o4, %o4 << 175 add %o1, %g4, %o1 << 176 sub %o5, %o4, %o5 << 177 jmpl %o5 + %lo(80f), %g0 << 178 add %o0, %g4, %o0 << 179 << 180 79: /* memcpy_table */ << 181 << 182 MOVE_LASTCHUNK(o1, o0, 0x60, g2, g3, g << 183 MOVE_LASTCHUNK(o1, o0, 0x50, g2, g3, g << 184 MOVE_LASTCHUNK(o1, o0, 0x40, g2, g3, g << 185 MOVE_LASTCHUNK(o1, o0, 0x30, g2, g3, g << 186 MOVE_LASTCHUNK(o1, o0, 0x20, g2, g3, g << 187 MOVE_LASTCHUNK(o1, o0, 0x10, g2, g3, g << 188 MOVE_LASTCHUNK(o1, o0, 0x00, g2, g3, g << 189 << 190 80: /* memcpy_table_end */ << 191 be 81f << 192 andcc %g1, 4, %g0 << 193 << 194 ldd [%o1], %g2 << 195 add %o0, 8, %o0 << 196 st %g2, [%o0 - 0x08] << 197 add %o1, 8, %o1 << 198 st %g3, [%o0 - 0x04] << 199 << 200 81: /* memcpy_last7 */ << 201 << 202 be 1f << 203 andcc %g1, 2, %g0 << 204 << 205 ld [%o1], %g2 << 206 add %o1, 4, %o1 << 207 st %g2, [%o0] << 208 add %o0, 4, %o0 << 209 1: 315 1: 210 be 1f !! 316 R10KCBARRIER(0(ra)) 211 andcc %g1, 1, %g0 !! 317 LOAD(t0, UNIT(0)(src), .Ll_exc\@) >> 318 LOAD(t1, UNIT(1)(src), .Ll_exc_copy\@) >> 319 LOAD(t2, UNIT(2)(src), .Ll_exc_copy\@) >> 320 LOAD(t3, UNIT(3)(src), .Ll_exc_copy\@) >> 321 SUB len, len, 8*NBYTES >> 322 LOAD(t4, UNIT(4)(src), .Ll_exc_copy\@) >> 323 LOAD(t7, UNIT(5)(src), .Ll_exc_copy\@) >> 324 STORE(t0, UNIT(0)(dst), .Ls_exc_p8u\@) >> 325 STORE(t1, UNIT(1)(dst), .Ls_exc_p7u\@) >> 326 LOAD(t0, UNIT(6)(src), .Ll_exc_copy\@) >> 327 LOAD(t1, UNIT(7)(src), .Ll_exc_copy\@) >> 328 ADD src, src, 8*NBYTES >> 329 ADD dst, dst, 8*NBYTES >> 330 STORE(t2, UNIT(-6)(dst), .Ls_exc_p6u\@) >> 331 STORE(t3, UNIT(-5)(dst), .Ls_exc_p5u\@) >> 332 STORE(t4, UNIT(-4)(dst), .Ls_exc_p4u\@) >> 333 STORE(t7, UNIT(-3)(dst), .Ls_exc_p3u\@) >> 334 STORE(t0, UNIT(-2)(dst), .Ls_exc_p2u\@) >> 335 STORE(t1, UNIT(-1)(dst), .Ls_exc_p1u\@) >> 336 PREFS( 0, 8*32(src) ) >> 337 PREFD( 1, 8*32(dst) ) >> 338 bne len, rem, 1b >> 339 nop 212 340 213 lduh [%o1], %g2 !! 341 /* 214 add %o1, 2, %o1 !! 342 * len == rem == the number of bytes left to copy < 8*NBYTES 215 sth %g2, [%o0] !! 343 */ 216 add %o0, 2, %o0 !! 344 .Lcleanup_both_aligned\@: >> 345 beqz len, .Ldone\@ >> 346 sltu t0, len, 4*NBYTES >> 347 bnez t0, .Lless_than_4units\@ >> 348 and rem, len, (NBYTES-1) # rem = len % NBYTES >> 349 /* >> 350 * len >= 4*NBYTES >> 351 */ >> 352 LOAD( t0, UNIT(0)(src), .Ll_exc\@) >> 353 LOAD( t1, UNIT(1)(src), .Ll_exc_copy\@) >> 354 LOAD( t2, UNIT(2)(src), .Ll_exc_copy\@) >> 355 LOAD( t3, UNIT(3)(src), .Ll_exc_copy\@) >> 356 SUB len, len, 4*NBYTES >> 357 ADD src, src, 4*NBYTES >> 358 R10KCBARRIER(0(ra)) >> 359 STORE(t0, UNIT(0)(dst), .Ls_exc_p4u\@) >> 360 STORE(t1, UNIT(1)(dst), .Ls_exc_p3u\@) >> 361 STORE(t2, UNIT(2)(dst), .Ls_exc_p2u\@) >> 362 STORE(t3, UNIT(3)(dst), .Ls_exc_p1u\@) >> 363 .set reorder /* DADDI_WAR */ >> 364 ADD dst, dst, 4*NBYTES >> 365 beqz len, .Ldone\@ >> 366 .set noreorder >> 367 .Lless_than_4units\@: >> 368 /* >> 369 * rem = len % NBYTES >> 370 */ >> 371 beq rem, len, .Lcopy_bytes\@ >> 372 nop >> 373 1: >> 374 R10KCBARRIER(0(ra)) >> 375 LOAD(t0, 0(src), .Ll_exc\@) >> 376 ADD src, src, NBYTES >> 377 SUB len, len, NBYTES >> 378 STORE(t0, 0(dst), .Ls_exc_p1u\@) >> 379 .set reorder /* DADDI_WAR */ >> 380 ADD dst, dst, NBYTES >> 381 bne rem, len, 1b >> 382 .set noreorder >> 383 >> 384 #ifndef CONFIG_CPU_MIPSR6 >> 385 /* >> 386 * src and dst are aligned, need to copy rem bytes (rem < NBYTES) >> 387 * A loop would do only a byte at a time with possible branch >> 388 * mispredicts. Can't do an explicit LOAD dst,mask,or,STORE >> 389 * because can't assume read-access to dst. Instead, use >> 390 * STREST dst, which doesn't require read access to dst. >> 391 * >> 392 * This code should perform better than a simple loop on modern, >> 393 * wide-issue mips processors because the code has fewer branches and >> 394 * more instruction-level parallelism. >> 395 */ >> 396 #define bits t2 >> 397 beqz len, .Ldone\@ >> 398 ADD t1, dst, len # t1 is just past last byte of dst >> 399 li bits, 8*NBYTES >> 400 SLL rem, len, 3 # rem = number of bits to keep >> 401 LOAD(t0, 0(src), .Ll_exc\@) >> 402 SUB bits, bits, rem # bits = number of bits to discard >> 403 SHIFT_DISCARD t0, t0, bits >> 404 STREST(t0, -1(t1), .Ls_exc\@) >> 405 jr ra >> 406 move len, zero >> 407 .Ldst_unaligned\@: >> 408 /* >> 409 * dst is unaligned >> 410 * t0 = src & ADDRMASK >> 411 * t1 = dst & ADDRMASK; T1 > 0 >> 412 * len >= NBYTES >> 413 * >> 414 * Copy enough bytes to align dst >> 415 * Set match = (src and dst have same alignment) >> 416 */ >> 417 #define match rem >> 418 LDFIRST(t3, FIRST(0)(src), .Ll_exc\@) >> 419 ADD t2, zero, NBYTES >> 420 LDREST(t3, REST(0)(src), .Ll_exc_copy\@) >> 421 SUB t2, t2, t1 # t2 = number of bytes copied >> 422 xor match, t0, t1 >> 423 R10KCBARRIER(0(ra)) >> 424 STFIRST(t3, FIRST(0)(dst), .Ls_exc\@) >> 425 beq len, t2, .Ldone\@ >> 426 SUB len, len, t2 >> 427 ADD dst, dst, t2 >> 428 beqz match, .Lboth_aligned\@ >> 429 ADD src, src, t2 >> 430 >> 431 .Lsrc_unaligned_dst_aligned\@: >> 432 SRL t0, len, LOG_NBYTES+2 # +2 for 4 units/iter >> 433 PREFS( 0, 3*32(src) ) >> 434 beqz t0, .Lcleanup_src_unaligned\@ >> 435 and rem, len, (4*NBYTES-1) # rem = len % 4*NBYTES >> 436 PREFD( 1, 3*32(dst) ) 217 1: 437 1: 218 be 1f !! 438 /* >> 439 * Avoid consecutive LD*'s to the same register since some mips >> 440 * implementations can't issue them in the same cycle. >> 441 * It's OK to load FIRST(N+1) before REST(N) because the two addresses >> 442 * are to the same unit (unless src is aligned, but it's not). >> 443 */ >> 444 R10KCBARRIER(0(ra)) >> 445 LDFIRST(t0, FIRST(0)(src), .Ll_exc\@) >> 446 LDFIRST(t1, FIRST(1)(src), .Ll_exc_copy\@) >> 447 SUB len, len, 4*NBYTES >> 448 LDREST(t0, REST(0)(src), .Ll_exc_copy\@) >> 449 LDREST(t1, REST(1)(src), .Ll_exc_copy\@) >> 450 LDFIRST(t2, FIRST(2)(src), .Ll_exc_copy\@) >> 451 LDFIRST(t3, FIRST(3)(src), .Ll_exc_copy\@) >> 452 LDREST(t2, REST(2)(src), .Ll_exc_copy\@) >> 453 LDREST(t3, REST(3)(src), .Ll_exc_copy\@) >> 454 PREFS( 0, 9*32(src) ) # 0 is PREF_LOAD (not streamed) >> 455 ADD src, src, 4*NBYTES >> 456 #ifdef CONFIG_CPU_SB1 >> 457 nop # improves slotting >> 458 #endif >> 459 STORE(t0, UNIT(0)(dst), .Ls_exc_p4u\@) >> 460 STORE(t1, UNIT(1)(dst), .Ls_exc_p3u\@) >> 461 STORE(t2, UNIT(2)(dst), .Ls_exc_p2u\@) >> 462 STORE(t3, UNIT(3)(dst), .Ls_exc_p1u\@) >> 463 PREFD( 1, 9*32(dst) ) # 1 is PREF_STORE (not streamed) >> 464 .set reorder /* DADDI_WAR */ >> 465 ADD dst, dst, 4*NBYTES >> 466 bne len, rem, 1b >> 467 .set noreorder >> 468 >> 469 .Lcleanup_src_unaligned\@: >> 470 beqz len, .Ldone\@ >> 471 and rem, len, NBYTES-1 # rem = len % NBYTES >> 472 beq rem, len, .Lcopy_bytes\@ 219 nop 473 nop 220 << 221 ldub [%o1], %g2 << 222 stb %g2, [%o0] << 223 1: 474 1: 224 retl !! 475 R10KCBARRIER(0(ra)) 225 mov %g7, %o0 !! 476 LDFIRST(t0, FIRST(0)(src), .Ll_exc\@) >> 477 LDREST(t0, REST(0)(src), .Ll_exc_copy\@) >> 478 ADD src, src, NBYTES >> 479 SUB len, len, NBYTES >> 480 STORE(t0, 0(dst), .Ls_exc_p1u\@) >> 481 .set reorder /* DADDI_WAR */ >> 482 ADD dst, dst, NBYTES >> 483 bne len, rem, 1b >> 484 .set noreorder >> 485 >> 486 #endif /* !CONFIG_CPU_MIPSR6 */ >> 487 .Lcopy_bytes_checklen\@: >> 488 beqz len, .Ldone\@ >> 489 nop >> 490 .Lcopy_bytes\@: >> 491 /* 0 < len < NBYTES */ >> 492 R10KCBARRIER(0(ra)) >> 493 #define COPY_BYTE(N) \ >> 494 LOADB(t0, N(src), .Ll_exc\@); \ >> 495 SUB len, len, 1; \ >> 496 beqz len, .Ldone\@; \ >> 497 STOREB(t0, N(dst), .Ls_exc_p1\@) >> 498 >> 499 COPY_BYTE(0) >> 500 COPY_BYTE(1) >> 501 #ifdef USE_DOUBLE >> 502 COPY_BYTE(2) >> 503 COPY_BYTE(3) >> 504 COPY_BYTE(4) >> 505 COPY_BYTE(5) >> 506 #endif >> 507 LOADB(t0, NBYTES-2(src), .Ll_exc\@) >> 508 SUB len, len, 1 >> 509 jr ra >> 510 STOREB(t0, NBYTES-2(dst), .Ls_exc_p1\@) >> 511 .Ldone\@: >> 512 jr ra >> 513 nop 226 514 227 82: /* ldd_std */ !! 515 #ifdef CONFIG_CPU_MIPSR6 228 MOVE_BIGALIGNCHUNK(o1, o0, 0x00, o2, o !! 516 .Lcopy_unaligned_bytes\@: 229 MOVE_BIGALIGNCHUNK(o1, o0, 0x20, o2, o << 230 MOVE_BIGALIGNCHUNK(o1, o0, 0x40, o2, o << 231 MOVE_BIGALIGNCHUNK(o1, o0, 0x60, o2, o << 232 subcc %g1, 128, %g1 << 233 add %o1, 128, %o1 << 234 cmp %g1, 128 << 235 bge 82b << 236 add %o0, 128, %o0 << 237 << 238 andcc %g1, 0x70, %g4 << 239 be 84f << 240 andcc %g1, 8, %g0 << 241 << 242 sethi %hi(84f), %o5 << 243 add %o1, %g4, %o1 << 244 sub %o5, %g4, %o5 << 245 jmpl %o5 + %lo(84f), %g0 << 246 add %o0, %g4, %o0 << 247 << 248 83: /* amemcpy_table */ << 249 << 250 MOVE_LASTALIGNCHUNK(o1, o0, 0x60, g2, << 251 MOVE_LASTALIGNCHUNK(o1, o0, 0x50, g2, << 252 MOVE_LASTALIGNCHUNK(o1, o0, 0x40, g2, << 253 MOVE_LASTALIGNCHUNK(o1, o0, 0x30, g2, << 254 MOVE_LASTALIGNCHUNK(o1, o0, 0x20, g2, << 255 MOVE_LASTALIGNCHUNK(o1, o0, 0x10, g2, << 256 MOVE_LASTALIGNCHUNK(o1, o0, 0x00, g2, << 257 << 258 84: /* amemcpy_table_end */ << 259 be 85f << 260 andcc %g1, 4, %g0 << 261 << 262 ldd [%o1], %g2 << 263 add %o0, 8, %o0 << 264 std %g2, [%o0 - 0x08] << 265 add %o1, 8, %o1 << 266 85: /* amemcpy_last7 */ << 267 be 1f << 268 andcc %g1, 2, %g0 << 269 << 270 ld [%o1], %g2 << 271 add %o1, 4, %o1 << 272 st %g2, [%o0] << 273 add %o0, 4, %o0 << 274 1: 517 1: 275 be 1f !! 518 COPY_BYTE(0) 276 andcc %g1, 1, %g0 !! 519 COPY_BYTE(1) 277 !! 520 COPY_BYTE(2) 278 lduh [%o1], %g2 !! 521 COPY_BYTE(3) 279 add %o1, 2, %o1 !! 522 COPY_BYTE(4) 280 sth %g2, [%o0] !! 523 COPY_BYTE(5) 281 add %o0, 2, %o0 !! 524 COPY_BYTE(6) >> 525 COPY_BYTE(7) >> 526 ADD src, src, 8 >> 527 b 1b >> 528 ADD dst, dst, 8 >> 529 #endif /* CONFIG_CPU_MIPSR6 */ >> 530 .if __memcpy == 1 >> 531 END(memcpy) >> 532 .set __memcpy, 0 >> 533 .hidden __memcpy >> 534 .endif >> 535 >> 536 .Ll_exc_copy\@: >> 537 /* >> 538 * Copy bytes from src until faulting load address (or until a >> 539 * lb faults) >> 540 * >> 541 * When reached by a faulting LDFIRST/LDREST, THREAD_BUADDR($28) >> 542 * may be more than a byte beyond the last address. >> 543 * Hence, the lb below may get an exception. >> 544 * >> 545 * Assumes src < THREAD_BUADDR($28) >> 546 */ >> 547 LOADK t0, TI_TASK($28) >> 548 nop >> 549 LOADK t0, THREAD_BUADDR(t0) 282 1: 550 1: 283 be 1f !! 551 LOADB(t1, 0(src), .Ll_exc\@) >> 552 ADD src, src, 1 >> 553 sb t1, 0(dst) # can't fault -- we're copy_from_user >> 554 .set reorder /* DADDI_WAR */ >> 555 ADD dst, dst, 1 >> 556 bne src, t0, 1b >> 557 .set noreorder >> 558 .Ll_exc\@: >> 559 LOADK t0, TI_TASK($28) >> 560 nop >> 561 LOADK t0, THREAD_BUADDR(t0) # t0 is just past last good address >> 562 nop >> 563 SUB len, AT, t0 # len number of uncopied bytes >> 564 bnez t6, .Ldone\@ /* Skip the zeroing part if inatomic */ >> 565 /* >> 566 * Here's where we rely on src and dst being incremented in tandem, >> 567 * See (3) above. >> 568 * dst += (fault addr - src) to put dst at first byte to clear >> 569 */ >> 570 ADD dst, t0 # compute start address in a1 >> 571 SUB dst, src >> 572 /* >> 573 * Clear len bytes starting at dst. Can't call __bzero because it >> 574 * might modify len. An inefficient loop for these rare times... >> 575 */ >> 576 .set reorder /* DADDI_WAR */ >> 577 SUB src, len, 1 >> 578 beqz len, .Ldone\@ >> 579 .set noreorder >> 580 1: sb zero, 0(dst) >> 581 ADD dst, dst, 1 >> 582 #ifndef CONFIG_CPU_DADDI_WORKAROUNDS >> 583 bnez src, 1b >> 584 SUB src, src, 1 >> 585 #else >> 586 .set push >> 587 .set noat >> 588 li v1, 1 >> 589 bnez src, 1b >> 590 SUB src, src, v1 >> 591 .set pop >> 592 #endif >> 593 jr ra 284 nop 594 nop 285 595 286 ldub [%o1], %g2 << 287 stb %g2, [%o0] << 288 1: << 289 retl << 290 mov %g7, %o0 << 291 596 292 86: /* non_aligned */ !! 597 #define SEXC(n) \ 293 cmp %o2, 6 !! 598 .set reorder; /* DADDI_WAR */ \ 294 bleu 88f !! 599 .Ls_exc_p ## n ## u\@: \ 295 nop !! 600 ADD len, len, n*NBYTES; \ 296 !! 601 jr ra; \ 297 save %sp, -96, %sp !! 602 .set noreorder 298 andcc %i0, 3, %g0 !! 603 299 be 61f !! 604 SEXC(8) 300 andcc %i0, 1, %g0 !! 605 SEXC(7) 301 be 60f !! 606 SEXC(6) 302 andcc %i0, 2, %g0 !! 607 SEXC(5) 303 !! 608 SEXC(4) 304 ldub [%i1], %g5 !! 609 SEXC(3) 305 add %i1, 1, %i1 !! 610 SEXC(2) 306 stb %g5, [%i0] !! 611 SEXC(1) 307 sub %i2, 1, %i2 !! 612 308 bne 61f !! 613 .Ls_exc_p1\@: 309 add %i0, 1, %i0 !! 614 .set reorder /* DADDI_WAR */ 310 60: !! 615 ADD len, len, 1 311 ldub [%i1], %g3 !! 616 jr ra 312 add %i1, 2, %i1 !! 617 .set noreorder 313 stb %g3, [%i0] !! 618 .Ls_exc\@: 314 sub %i2, 2, %i2 !! 619 jr ra 315 ldub [%i1 - 1], %g3 << 316 add %i0, 2, %i0 << 317 stb %g3, [%i0 - 1] << 318 61: << 319 and %i1, 3, %g2 << 320 and %i2, 0xc, %g3 << 321 and %i1, -4, %i1 << 322 cmp %g3, 4 << 323 sll %g2, 3, %g4 << 324 mov 32, %g2 << 325 be 4f << 326 sub %g2, %g4, %l0 << 327 << 328 blu 3f << 329 cmp %g3, 0x8 << 330 << 331 be 2f << 332 srl %i2, 2, %g3 << 333 << 334 ld [%i1], %i3 << 335 add %i0, -8, %i0 << 336 ld [%i1 + 4], %i4 << 337 b 8f << 338 add %g3, 1, %g3 << 339 2: << 340 ld [%i1], %i4 << 341 add %i0, -12, %i0 << 342 ld [%i1 + 4], %i5 << 343 add %g3, 2, %g3 << 344 b 9f << 345 add %i1, -4, %i1 << 346 3: << 347 ld [%i1], %g1 << 348 add %i0, -4, %i0 << 349 ld [%i1 + 4], %i3 << 350 srl %i2, 2, %g3 << 351 b 7f << 352 add %i1, 4, %i1 << 353 4: << 354 ld [%i1], %i5 << 355 cmp %i2, 7 << 356 ld [%i1 + 4], %g1 << 357 srl %i2, 2, %g3 << 358 bleu 10f << 359 add %i1, 8, %i1 << 360 << 361 ld [%i1], %i3 << 362 add %g3, -1, %g3 << 363 5: << 364 sll %i5, %g4, %g2 << 365 srl %g1, %l0, %g5 << 366 or %g2, %g5, %g2 << 367 st %g2, [%i0] << 368 7: << 369 ld [%i1 + 4], %i4 << 370 sll %g1, %g4, %g2 << 371 srl %i3, %l0, %g5 << 372 or %g2, %g5, %g2 << 373 st %g2, [%i0 + 4] << 374 8: << 375 ld [%i1 + 8], %i5 << 376 sll %i3, %g4, %g2 << 377 srl %i4, %l0, %g5 << 378 or %g2, %g5, %g2 << 379 st %g2, [%i0 + 8] << 380 9: << 381 ld [%i1 + 12], %g1 << 382 sll %i4, %g4, %g2 << 383 srl %i5, %l0, %g5 << 384 addcc %g3, -4, %g3 << 385 or %g2, %g5, %g2 << 386 add %i1, 16, %i1 << 387 st %g2, [%i0 + 12] << 388 add %i0, 16, %i0 << 389 bne,a 5b << 390 ld [%i1], %i3 << 391 10: << 392 sll %i5, %g4, %g2 << 393 srl %g1, %l0, %g5 << 394 srl %l0, 3, %g3 << 395 or %g2, %g5, %g2 << 396 sub %i1, %g3, %i1 << 397 andcc %i2, 2, %g0 << 398 st %g2, [%i0] << 399 be 1f << 400 andcc %i2, 1, %g0 << 401 << 402 ldub [%i1], %g2 << 403 add %i1, 2, %i1 << 404 stb %g2, [%i0 + 4] << 405 add %i0, 2, %i0 << 406 ldub [%i1 - 1], %g2 << 407 stb %g2, [%i0 + 3] << 408 1: << 409 be 1f << 410 nop 620 nop 411 ldub [%i1], %g2 !! 621 .endm 412 stb %g2, [%i0 + 4] << 413 1: << 414 ret << 415 restore %g7, %g0, %o0 << 416 622 417 88: /* short_end */ !! 623 .align 5 >> 624 LEAF(memmove) >> 625 ADD t0, a0, a2 >> 626 ADD t1, a1, a2 >> 627 sltu t0, a1, t0 # dst + len <= src -> memcpy >> 628 sltu t1, a0, t1 # dst >= src + len -> memcpy >> 629 and t0, t1 >> 630 beqz t0, .L__memcpy >> 631 move v0, a0 /* return value */ >> 632 beqz a2, .Lr_out >> 633 END(memmove) >> 634 >> 635 /* fall through to __rmemcpy */ >> 636 LEAF(__rmemcpy) /* a0=dst a1=src a2=len */ >> 637 sltu t0, a1, a0 >> 638 beqz t0, .Lr_end_bytes_up # src >= dst >> 639 nop >> 640 ADD a0, a2 # dst = dst + len >> 641 ADD a1, a2 # src = src + len 418 642 419 and %o2, 0xe, %o3 !! 643 .Lr_end_bytes: 420 20: !! 644 R10KCBARRIER(0(ra)) 421 sethi %hi(89f), %o5 !! 645 lb t0, -1(a1) 422 sll %o3, 3, %o4 !! 646 SUB a2, a2, 0x1 423 add %o0, %o3, %o0 !! 647 sb t0, -1(a0) 424 sub %o5, %o4, %o5 !! 648 SUB a1, a1, 0x1 425 add %o1, %o3, %o1 !! 649 .set reorder /* DADDI_WAR */ 426 jmpl %o5 + %lo(89f), %g0 !! 650 SUB a0, a0, 0x1 427 andcc %o2, 1, %g0 !! 651 bnez a2, .Lr_end_bytes 428 !! 652 .set noreorder 429 MOVE_SHORTCHUNK(o1, o0, 0x0c, g2, g3) !! 653 430 MOVE_SHORTCHUNK(o1, o0, 0x0a, g2, g3) !! 654 .Lr_out: 431 MOVE_SHORTCHUNK(o1, o0, 0x08, g2, g3) !! 655 jr ra 432 MOVE_SHORTCHUNK(o1, o0, 0x06, g2, g3) !! 656 move a2, zero 433 MOVE_SHORTCHUNK(o1, o0, 0x04, g2, g3) !! 657 434 MOVE_SHORTCHUNK(o1, o0, 0x02, g2, g3) !! 658 .Lr_end_bytes_up: 435 MOVE_SHORTCHUNK(o1, o0, 0x00, g2, g3) !! 659 R10KCBARRIER(0(ra)) >> 660 lb t0, (a1) >> 661 SUB a2, a2, 0x1 >> 662 sb t0, (a0) >> 663 ADD a1, a1, 0x1 >> 664 .set reorder /* DADDI_WAR */ >> 665 ADD a0, a0, 0x1 >> 666 bnez a2, .Lr_end_bytes_up >> 667 .set noreorder >> 668 >> 669 jr ra >> 670 move a2, zero >> 671 END(__rmemcpy) 436 672 437 89: /* short_table_end */ !! 673 /* >> 674 * t6 is used as a flag to note inatomic mode. >> 675 */ >> 676 LEAF(__copy_user_inatomic) >> 677 b __copy_user_common >> 678 li t6, 1 >> 679 END(__copy_user_inatomic) >> 680 >> 681 /* >> 682 * A combined memcpy/__copy_user >> 683 * __copy_user sets len to 0 for success; else to an upper bound of >> 684 * the number of uncopied bytes. >> 685 * memcpy sets v0 to dst. >> 686 */ >> 687 .align 5 >> 688 LEAF(memcpy) /* a0=dst a1=src a2=len */ >> 689 move v0, dst /* return value */ >> 690 .L__memcpy: >> 691 FEXPORT(__copy_user) >> 692 li t6, 0 /* not inatomic */ >> 693 __copy_user_common: >> 694 /* Legacy Mode, user <-> user */ >> 695 __BUILD_COPY_USER LEGACY_MODE USEROP USEROP >> 696 >> 697 #ifdef CONFIG_EVA >> 698 >> 699 /* >> 700 * For EVA we need distinct symbols for reading and writing to user space. >> 701 * This is because we need to use specific EVA instructions to perform the >> 702 * virtual <-> physical translation when a virtual address is actually in user >> 703 * space >> 704 */ 438 705 439 be 1f !! 706 LEAF(__copy_user_inatomic_eva) 440 nop !! 707 b __copy_from_user_common >> 708 li t6, 1 >> 709 END(__copy_user_inatomic_eva) 441 710 442 ldub [%o1], %g2 !! 711 /* 443 stb %g2, [%o0] !! 712 * __copy_from_user (EVA) 444 1: !! 713 */ 445 retl << 446 mov %g7, %o0 << 447 714 448 90: /* short_aligned_end */ !! 715 LEAF(__copy_from_user_eva) 449 bne 88b !! 716 li t6, 0 /* not inatomic */ 450 andcc %o2, 8, %g0 !! 717 __copy_from_user_common: 451 !! 718 __BUILD_COPY_USER EVA_MODE USEROP KERNELOP 452 be 1f !! 719 END(__copy_from_user_eva) 453 andcc %o2, 4, %g0 !! 720 454 !! 721 455 ld [%o1 + 0x00], %g2 !! 722 456 ld [%o1 + 0x04], %g3 !! 723 /* 457 add %o1, 8, %o1 !! 724 * __copy_to_user (EVA) 458 st %g2, [%o0 + 0x00] !! 725 */ 459 st %g3, [%o0 + 0x04] !! 726 460 add %o0, 8, %o0 !! 727 LEAF(__copy_to_user_eva) 461 1: !! 728 __BUILD_COPY_USER EVA_MODE KERNELOP USEROP 462 b 81b !! 729 END(__copy_to_user_eva) 463 mov %o2, %g1 !! 730 >> 731 /* >> 732 * __copy_in_user (EVA) >> 733 */ >> 734 >> 735 LEAF(__copy_in_user_eva) >> 736 __BUILD_COPY_USER EVA_MODE USEROP USEROP >> 737 END(__copy_in_user_eva) >> 738 >> 739 #endif
Linux® is a registered trademark of Linus Torvalds in the United States and other countries.
TOMOYO® is a registered trademark of NTT DATA CORPORATION.