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: << 18 << 19 /* Both these macros have to start with exactl << 20 #define MOVE_BIGCHUNK(src, dst, offset, t0, t1 << 21 ldd [%src + (offset) + 0x00], %t0; << 22 ldd [%src + (offset) + 0x08], %t2; << 23 ldd [%src + (offset) + 0x10], %t4; << 24 ldd [%src + (offset) + 0x18], %t6; << 25 st %t0, [%dst + (offset) + 0x00]; << 26 st %t1, [%dst + (offset) + 0x04]; << 27 st %t2, [%dst + (offset) + 0x08]; << 28 st %t3, [%dst + (offset) + 0x0c]; << 29 st %t4, [%dst + (offset) + 0x10]; << 30 st %t5, [%dst + (offset) + 0x14]; << 31 st %t6, [%dst + (offset) + 0x18]; << 32 st %t7, [%dst + (offset) + 0x1c]; << 33 << 34 #define MOVE_BIGALIGNCHUNK(src, dst, offset, t << 35 ldd [%src + (offset) + 0x00], %t0; << 36 ldd [%src + (offset) + 0x08], %t2; << 37 ldd [%src + (offset) + 0x10], %t4; << 38 ldd [%src + (offset) + 0x18], %t6; << 39 std %t0, [%dst + (offset) + 0x00]; << 40 std %t2, [%dst + (offset) + 0x08]; << 41 std %t4, [%dst + (offset) + 0x10]; << 42 std %t6, [%dst + (offset) + 0x18]; << 43 << 44 #define MOVE_LASTCHUNK(src, dst, offset, t0, t << 45 ldd [%src - (offset) - 0x10], %t0; << 46 ldd [%src - (offset) - 0x08], %t2; << 47 st %t0, [%dst - (offset) - 0x10]; << 48 st %t1, [%dst - (offset) - 0x0c]; << 49 st %t2, [%dst - (offset) - 0x08]; << 50 st %t3, [%dst - (offset) - 0x04]; << 51 << 52 #define MOVE_LASTALIGNCHUNK(src, dst, offset, << 53 ldd [%src - (offset) - 0x10], %t0; << 54 ldd [%src - (offset) - 0x08], %t2; << 55 std %t0, [%dst - (offset) - 0x10]; << 56 std %t2, [%dst - (offset) - 0x08]; << 57 << 58 #define MOVE_SHORTCHUNK(src, dst, offset, t0, << 59 ldub [%src - (offset) - 0x02], %t0; << 60 ldub [%src - (offset) - 0x01], %t1; << 61 stb %t0, [%dst - (offset) - 0x02]; << 62 stb %t1, [%dst - (offset) - 0x01]; << 63 17 64 .text !! 18 /* 65 .align 4 !! 19 * Hack to resolve longstanding prefetch issue >> 20 * >> 21 * Prefetching may be fatal on some systems if we're prefetching beyond the >> 22 * end of memory on some systems. It's also a seriously bad idea on non >> 23 * dma-coherent systems. >> 24 */ >> 25 #ifdef CONFIG_DMA_NONCOHERENT >> 26 #undef CONFIG_CPU_HAS_PREFETCH >> 27 #endif >> 28 #ifdef CONFIG_MIPS_MALTA >> 29 #undef CONFIG_CPU_HAS_PREFETCH >> 30 #endif >> 31 #ifdef CONFIG_CPU_MIPSR6 >> 32 #undef CONFIG_CPU_HAS_PREFETCH >> 33 #endif >> 34 >> 35 #include <asm/asm.h> >> 36 #include <asm/asm-offsets.h> >> 37 #include <asm/export.h> >> 38 #include <asm/regdef.h> >> 39 >> 40 #define dst a0 >> 41 #define src a1 >> 42 #define len a2 >> 43 >> 44 /* >> 45 * Spec >> 46 * >> 47 * memcpy copies len bytes from src to dst and sets v0 to dst. >> 48 * It assumes that >> 49 * - src and dst don't overlap >> 50 * - src is readable >> 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 */ 66 69 67 FUNC(memmove) !! 70 /* 68 EXPORT_SYMBOL(memmove) !! 71 * Implementation 69 cmp %o0, %o1 !! 72 */ 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 73 125 FUNC(memcpy) /* %o0=dst %o1=src %o2=len */ !! 74 /* 126 EXPORT_SYMBOL(memcpy) !! 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 >> 85 * >> 86 * The exception handlers for stores adjust len (if necessary) and return. >> 87 * These handlers do not need to overwrite any data. >> 88 * >> 89 * For __rmemcpy and memmove an exception is always a kernel bug, therefore >> 90 * they're not protected. >> 91 */ 127 92 128 sub %o0, %o1, %o4 !! 93 /* Instruction type */ 129 mov %o0, %g7 !! 94 #define LD_INSN 1 130 9: !! 95 #define ST_INSN 2 131 andcc %o4, 3, %o5 !! 96 /* Pretech type */ 132 0: !! 97 #define SRC_PREFETCH 1 133 bne 86f !! 98 #define DST_PREFETCH 2 134 cmp %o2, 15 !! 99 #define LEGACY_MODE 1 135 !! 100 #define EVA_MODE 2 136 bleu 90f !! 101 #define USEROP 1 137 andcc %o1, 3, %g0 !! 102 #define KERNELOP 2 138 !! 103 139 bne 78b !! 104 /* 140 3: !! 105 * Wrapper to add an entry in the exception table 141 andcc %o1, 4, %g0 !! 106 * in case the insn causes a memory exception. 142 !! 107 * Arguments: 143 be 2f !! 108 * insn : Load/store instruction 144 mov %o2, %g1 !! 109 * type : Instruction type 145 !! 110 * reg : Register 146 ld [%o1], %o4 !! 111 * addr : Address 147 sub %g1, 4, %g1 !! 112 * handler : Exception handler 148 st %o4, [%o0] !! 113 */ 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: << 210 be 1f << 211 andcc %g1, 1, %g0 << 212 114 213 lduh [%o1], %g2 !! 115 #define EXC(insn, type, reg, addr, handler) \ 214 add %o1, 2, %o1 !! 116 .if \mode == LEGACY_MODE; \ 215 sth %g2, [%o0] !! 117 9: insn reg, addr; \ 216 add %o0, 2, %o0 !! 118 .section __ex_table,"a"; \ 217 1: !! 119 PTR 9b, handler; \ 218 be 1f !! 120 .previous; \ 219 nop !! 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 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 220 138 221 ldub [%o1], %g2 !! 139 /* 222 stb %g2, [%o0] !! 140 * Only on the 64-bit kernel we can made use of 64-bit registers. 223 1: !! 141 */ 224 retl !! 142 #ifdef CONFIG_64BIT 225 mov %g7, %o0 !! 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) 226 252 227 82: /* ldd_std */ !! 253 #define ADDRMASK (NBYTES-1) 228 MOVE_BIGALIGNCHUNK(o1, o0, 0x00, o2, o << 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: << 275 be 1f << 276 andcc %g1, 1, %g0 << 277 254 278 lduh [%o1], %g2 !! 255 .text 279 add %o1, 2, %o1 !! 256 .set noreorder 280 sth %g2, [%o0] !! 257 #ifndef CONFIG_CPU_DADDI_WORKAROUNDS 281 add %o0, 2, %o0 !! 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 4 282 1: 323 1: 283 be 1f !! 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 284 nop 347 nop 285 348 286 ldub [%o1], %g2 !! 349 /* 287 stb %g2, [%o0] !! 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 288 1: 482 1: 289 retl !! 483 R10KCBARRIER(0(ra)) 290 mov %g7, %o0 !! 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 291 522 292 86: /* non_aligned */ !! 523 #ifdef CONFIG_CPU_NO_LOAD_STORE_LR 293 cmp %o2, 6 !! 524 .Lcopy_unaligned_bytes\@: 294 bleu 88f << 295 nop << 296 << 297 save %sp, -96, %sp << 298 andcc %i0, 3, %g0 << 299 be 61f << 300 andcc %i0, 1, %g0 << 301 be 60f << 302 andcc %i0, 2, %g0 << 303 << 304 ldub [%i1], %g5 << 305 add %i1, 1, %i1 << 306 stb %g5, [%i0] << 307 sub %i2, 1, %i2 << 308 bne 61f << 309 add %i0, 1, %i0 << 310 60: << 311 ldub [%i1], %g3 << 312 add %i1, 2, %i1 << 313 stb %g3, [%i0] << 314 sub %i2, 2, %i2 << 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: 525 1: 409 be 1f !! 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) 410 nop 556 nop 411 ldub [%i1], %g2 !! 557 LOADK t0, THREAD_BUADDR(t0) 412 stb %g2, [%i0 + 4] << 413 1: 558 1: 414 ret !! 559 LOADB(t1, 0(src), .Ll_exc\@) 415 restore %g7, %g0, %o0 !! 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 .align 5 >> 602 LEAF(memmove) >> 603 EXPORT_SYMBOL(memmove) >> 604 ADD t0, a0, a2 >> 605 ADD t1, a1, a2 >> 606 sltu t0, a1, t0 # dst + len <= src -> memcpy >> 607 sltu t1, a0, t1 # dst >= src + len -> memcpy >> 608 and t0, t1 >> 609 beqz t0, .L__memcpy >> 610 move v0, a0 /* return value */ >> 611 beqz a2, .Lr_out >> 612 END(memmove) >> 613 >> 614 /* fall through to __rmemcpy */ >> 615 LEAF(__rmemcpy) /* a0=dst a1=src a2=len */ >> 616 sltu t0, a1, a0 >> 617 beqz t0, .Lr_end_bytes_up # src >= dst >> 618 nop >> 619 ADD a0, a2 # dst = dst + len >> 620 ADD a1, a2 # src = src + len >> 621 >> 622 .Lr_end_bytes: >> 623 R10KCBARRIER(0(ra)) >> 624 lb t0, -1(a1) >> 625 SUB a2, a2, 0x1 >> 626 sb t0, -1(a0) >> 627 SUB a1, a1, 0x1 >> 628 .set reorder /* DADDI_WAR */ >> 629 SUB a0, a0, 0x1 >> 630 bnez a2, .Lr_end_bytes >> 631 .set noreorder >> 632 >> 633 .Lr_out: >> 634 jr ra >> 635 move a2, zero >> 636 >> 637 .Lr_end_bytes_up: >> 638 R10KCBARRIER(0(ra)) >> 639 lb t0, (a1) >> 640 SUB a2, a2, 0x1 >> 641 sb t0, (a0) >> 642 ADD a1, a1, 0x1 >> 643 .set reorder /* DADDI_WAR */ >> 644 ADD a0, a0, 0x1 >> 645 bnez a2, .Lr_end_bytes_up >> 646 .set noreorder >> 647 >> 648 jr ra >> 649 move a2, zero >> 650 END(__rmemcpy) >> 651 >> 652 /* >> 653 * A combined memcpy/__copy_user >> 654 * __copy_user sets len to 0 for success; else to an upper bound of >> 655 * the number of uncopied bytes. >> 656 * memcpy sets v0 to dst. >> 657 */ >> 658 .align 5 >> 659 LEAF(memcpy) /* a0=dst a1=src a2=len */ >> 660 EXPORT_SYMBOL(memcpy) >> 661 move v0, dst /* return value */ >> 662 .L__memcpy: >> 663 FEXPORT(__copy_user) >> 664 EXPORT_SYMBOL(__copy_user) >> 665 /* Legacy Mode, user <-> user */ >> 666 __BUILD_COPY_USER LEGACY_MODE USEROP USEROP >> 667 >> 668 #ifdef CONFIG_EVA >> 669 >> 670 /* >> 671 * For EVA we need distinct symbols for reading and writing to user space. >> 672 * This is because we need to use specific EVA instructions to perform the >> 673 * virtual <-> physical translation when a virtual address is actually in user >> 674 * space >> 675 */ 416 676 417 88: /* short_end */ !! 677 /* >> 678 * __copy_from_user (EVA) >> 679 */ 418 680 419 and %o2, 0xe, %o3 !! 681 LEAF(__copy_from_user_eva) 420 20: !! 682 EXPORT_SYMBOL(__copy_from_user_eva) 421 sethi %hi(89f), %o5 !! 683 __BUILD_COPY_USER EVA_MODE USEROP KERNELOP 422 sll %o3, 3, %o4 !! 684 END(__copy_from_user_eva) 423 add %o0, %o3, %o0 << 424 sub %o5, %o4, %o5 << 425 add %o1, %o3, %o1 << 426 jmpl %o5 + %lo(89f), %g0 << 427 andcc %o2, 1, %g0 << 428 << 429 MOVE_SHORTCHUNK(o1, o0, 0x0c, g2, g3) << 430 MOVE_SHORTCHUNK(o1, o0, 0x0a, g2, g3) << 431 MOVE_SHORTCHUNK(o1, o0, 0x08, g2, g3) << 432 MOVE_SHORTCHUNK(o1, o0, 0x06, g2, g3) << 433 MOVE_SHORTCHUNK(o1, o0, 0x04, g2, g3) << 434 MOVE_SHORTCHUNK(o1, o0, 0x02, g2, g3) << 435 MOVE_SHORTCHUNK(o1, o0, 0x00, g2, g3) << 436 685 437 89: /* short_table_end */ << 438 686 439 be 1f << 440 nop << 441 687 442 ldub [%o1], %g2 !! 688 /* 443 stb %g2, [%o0] !! 689 * __copy_to_user (EVA) 444 1: !! 690 */ 445 retl << 446 mov %g7, %o0 << 447 691 448 90: /* short_aligned_end */ !! 692 LEAF(__copy_to_user_eva) 449 bne 88b !! 693 EXPORT_SYMBOL(__copy_to_user_eva) 450 andcc %o2, 8, %g0 !! 694 __BUILD_COPY_USER EVA_MODE KERNELOP USEROP 451 !! 695 END(__copy_to_user_eva) 452 be 1f !! 696 453 andcc %o2, 4, %g0 !! 697 /* 454 !! 698 * __copy_in_user (EVA) 455 ld [%o1 + 0x00], %g2 !! 699 */ 456 ld [%o1 + 0x04], %g3 !! 700 457 add %o1, 8, %o1 !! 701 LEAF(__copy_in_user_eva) 458 st %g2, [%o0 + 0x00] !! 702 EXPORT_SYMBOL(__copy_in_user_eva) 459 st %g3, [%o0 + 0x04] !! 703 __BUILD_COPY_USER EVA_MODE USEROP USEROP 460 add %o0, 8, %o0 !! 704 END(__copy_in_user_eva) 461 1: !! 705 462 b 81b !! 706 #endif 463 mov %o2, %g1 <<
Linux® is a registered trademark of Linus Torvalds in the United States and other countries.
TOMOYO® is a registered trademark of NTT DATA CORPORATION.