Linus Torvalds | 1da177e | 2005-04-16 15:20:36 -0700 | [diff] [blame^] | 1 | /*#************************************************************************#*/ |
| 2 | /*#-------------------------------------------------------------------------*/ |
| 3 | /*# */ |
| 4 | /*# FUNCTION NAME: memcpy() */ |
| 5 | /*# */ |
| 6 | /*# PARAMETERS: void* dst; Destination address. */ |
| 7 | /*# void* src; Source address. */ |
| 8 | /*# int len; Number of bytes to copy. */ |
| 9 | /*# */ |
| 10 | /*# RETURNS: dst. */ |
| 11 | /*# */ |
| 12 | /*# DESCRIPTION: Copies len bytes of memory from src to dst. No guarantees */ |
| 13 | /*# about copying of overlapping memory areas. This routine is */ |
| 14 | /*# very sensitive to compiler changes in register allocation. */ |
| 15 | /*# Should really be rewritten to avoid this problem. */ |
| 16 | /*# */ |
| 17 | /*#-------------------------------------------------------------------------*/ |
| 18 | /*# */ |
| 19 | /*# HISTORY */ |
| 20 | /*# */ |
| 21 | /*# DATE NAME CHANGES */ |
| 22 | /*# ---- ---- ------- */ |
| 23 | /*# 941007 Kenny R Creation */ |
| 24 | /*# 941011 Kenny R Lots of optimizations and inlining. */ |
| 25 | /*# 941129 Ulf A Adapted for use in libc. */ |
| 26 | /*# 950216 HP N==0 forgotten if non-aligned src/dst. */ |
| 27 | /*# Added some optimizations. */ |
| 28 | /*# 001025 HP Make src and dst char *. Align dst to */ |
| 29 | /*# dword, not just word-if-both-src-and-dst- */ |
| 30 | /*# are-misaligned. */ |
| 31 | /*# */ |
| 32 | /*#-------------------------------------------------------------------------*/ |
| 33 | |
| 34 | #include <linux/types.h> |
| 35 | |
| 36 | void *memcpy(void *pdst, |
| 37 | const void *psrc, |
| 38 | size_t pn) |
| 39 | { |
| 40 | /* Ok. Now we want the parameters put in special registers. |
| 41 | Make sure the compiler is able to make something useful of this. |
| 42 | As it is now: r10 -> r13; r11 -> r11 (nop); r12 -> r12 (nop). |
| 43 | |
| 44 | If gcc was allright, it really would need no temporaries, and no |
| 45 | stack space to save stuff on. */ |
| 46 | |
| 47 | register void *return_dst __asm__ ("r10") = pdst; |
| 48 | register char *dst __asm__ ("r13") = pdst; |
| 49 | register const char *src __asm__ ("r11") = psrc; |
| 50 | register int n __asm__ ("r12") = pn; |
| 51 | |
| 52 | |
| 53 | /* When src is aligned but not dst, this makes a few extra needless |
| 54 | cycles. I believe it would take as many to check that the |
| 55 | re-alignment was unnecessary. */ |
| 56 | if (((unsigned long) dst & 3) != 0 |
| 57 | /* Don't align if we wouldn't copy more than a few bytes; so we |
| 58 | don't have to check further for overflows. */ |
| 59 | && n >= 3) |
| 60 | { |
| 61 | if ((unsigned long) dst & 1) |
| 62 | { |
| 63 | n--; |
| 64 | *(char*)dst = *(char*)src; |
| 65 | src++; |
| 66 | dst++; |
| 67 | } |
| 68 | |
| 69 | if ((unsigned long) dst & 2) |
| 70 | { |
| 71 | n -= 2; |
| 72 | *(short*)dst = *(short*)src; |
| 73 | src += 2; |
| 74 | dst += 2; |
| 75 | } |
| 76 | } |
| 77 | |
| 78 | /* Decide which copying method to use. */ |
| 79 | if (n >= 44*2) /* Break even between movem and |
| 80 | move16 is at 38.7*2, but modulo 44. */ |
| 81 | { |
| 82 | /* For large copies we use 'movem' */ |
| 83 | |
| 84 | /* It is not optimal to tell the compiler about clobbering any |
| 85 | registers; that will move the saving/restoring of those registers |
| 86 | to the function prologue/epilogue, and make non-movem sizes |
| 87 | suboptimal. |
| 88 | |
| 89 | This method is not foolproof; it assumes that the "asm reg" |
| 90 | declarations at the beginning of the function really are used |
| 91 | here (beware: they may be moved to temporary registers). |
| 92 | This way, we do not have to save/move the registers around into |
| 93 | temporaries; we can safely use them straight away. |
| 94 | |
| 95 | If you want to check that the allocation was right; then |
| 96 | check the equalities in the first comment. It should say |
| 97 | "r13=r13, r11=r11, r12=r12" */ |
| 98 | __asm__ volatile (" |
| 99 | ;; Check that the following is true (same register names on |
| 100 | ;; both sides of equal sign, as in r8=r8): |
| 101 | ;; %0=r13, %1=r11, %2=r12 |
| 102 | ;; |
| 103 | ;; Save the registers we'll use in the movem process |
| 104 | ;; on the stack. |
| 105 | subq 11*4,$sp |
| 106 | movem $r10,[$sp] |
| 107 | |
| 108 | ;; Now we've got this: |
| 109 | ;; r11 - src |
| 110 | ;; r13 - dst |
| 111 | ;; r12 - n |
| 112 | |
| 113 | ;; Update n for the first loop |
| 114 | subq 44,$r12 |
| 115 | 0: |
| 116 | movem [$r11+],$r10 |
| 117 | subq 44,$r12 |
| 118 | bge 0b |
| 119 | movem $r10,[$r13+] |
| 120 | |
| 121 | addq 44,$r12 ;; compensate for last loop underflowing n |
| 122 | |
| 123 | ;; Restore registers from stack |
| 124 | movem [$sp+],$r10" |
| 125 | |
| 126 | /* Outputs */ : "=r" (dst), "=r" (src), "=r" (n) |
| 127 | /* Inputs */ : "0" (dst), "1" (src), "2" (n)); |
| 128 | |
| 129 | } |
| 130 | |
| 131 | /* Either we directly starts copying, using dword copying |
| 132 | in a loop, or we copy as much as possible with 'movem' |
| 133 | and then the last block (<44 bytes) is copied here. |
| 134 | This will work since 'movem' will have updated src,dst,n. */ |
| 135 | |
| 136 | while ( n >= 16 ) |
| 137 | { |
| 138 | *((long*)dst)++ = *((long*)src)++; |
| 139 | *((long*)dst)++ = *((long*)src)++; |
| 140 | *((long*)dst)++ = *((long*)src)++; |
| 141 | *((long*)dst)++ = *((long*)src)++; |
| 142 | n -= 16; |
| 143 | } |
| 144 | |
| 145 | /* A switch() is definitely the fastest although it takes a LOT of code. |
| 146 | * Particularly if you inline code this. |
| 147 | */ |
| 148 | switch (n) |
| 149 | { |
| 150 | case 0: |
| 151 | break; |
| 152 | case 1: |
| 153 | *(char*)dst = *(char*)src; |
| 154 | break; |
| 155 | case 2: |
| 156 | *(short*)dst = *(short*)src; |
| 157 | break; |
| 158 | case 3: |
| 159 | *((short*)dst)++ = *((short*)src)++; |
| 160 | *(char*)dst = *(char*)src; |
| 161 | break; |
| 162 | case 4: |
| 163 | *((long*)dst)++ = *((long*)src)++; |
| 164 | break; |
| 165 | case 5: |
| 166 | *((long*)dst)++ = *((long*)src)++; |
| 167 | *(char*)dst = *(char*)src; |
| 168 | break; |
| 169 | case 6: |
| 170 | *((long*)dst)++ = *((long*)src)++; |
| 171 | *(short*)dst = *(short*)src; |
| 172 | break; |
| 173 | case 7: |
| 174 | *((long*)dst)++ = *((long*)src)++; |
| 175 | *((short*)dst)++ = *((short*)src)++; |
| 176 | *(char*)dst = *(char*)src; |
| 177 | break; |
| 178 | case 8: |
| 179 | *((long*)dst)++ = *((long*)src)++; |
| 180 | *((long*)dst)++ = *((long*)src)++; |
| 181 | break; |
| 182 | case 9: |
| 183 | *((long*)dst)++ = *((long*)src)++; |
| 184 | *((long*)dst)++ = *((long*)src)++; |
| 185 | *(char*)dst = *(char*)src; |
| 186 | break; |
| 187 | case 10: |
| 188 | *((long*)dst)++ = *((long*)src)++; |
| 189 | *((long*)dst)++ = *((long*)src)++; |
| 190 | *(short*)dst = *(short*)src; |
| 191 | break; |
| 192 | case 11: |
| 193 | *((long*)dst)++ = *((long*)src)++; |
| 194 | *((long*)dst)++ = *((long*)src)++; |
| 195 | *((short*)dst)++ = *((short*)src)++; |
| 196 | *(char*)dst = *(char*)src; |
| 197 | break; |
| 198 | case 12: |
| 199 | *((long*)dst)++ = *((long*)src)++; |
| 200 | *((long*)dst)++ = *((long*)src)++; |
| 201 | *((long*)dst)++ = *((long*)src)++; |
| 202 | break; |
| 203 | case 13: |
| 204 | *((long*)dst)++ = *((long*)src)++; |
| 205 | *((long*)dst)++ = *((long*)src)++; |
| 206 | *((long*)dst)++ = *((long*)src)++; |
| 207 | *(char*)dst = *(char*)src; |
| 208 | break; |
| 209 | case 14: |
| 210 | *((long*)dst)++ = *((long*)src)++; |
| 211 | *((long*)dst)++ = *((long*)src)++; |
| 212 | *((long*)dst)++ = *((long*)src)++; |
| 213 | *(short*)dst = *(short*)src; |
| 214 | break; |
| 215 | case 15: |
| 216 | *((long*)dst)++ = *((long*)src)++; |
| 217 | *((long*)dst)++ = *((long*)src)++; |
| 218 | *((long*)dst)++ = *((long*)src)++; |
| 219 | *((short*)dst)++ = *((short*)src)++; |
| 220 | *(char*)dst = *(char*)src; |
| 221 | break; |
| 222 | } |
| 223 | |
| 224 | return return_dst; /* destination pointer. */ |
| 225 | } /* memcpy() */ |