2 * Optimized memory copy routines.
4 * Copyright (C) 2004 Randolph Chung <tausq@debian.org>
6 * This program is free software; you can redistribute it and/or modify
7 * it under the terms of the GNU General Public License as published by
8 * the Free Software Foundation; either version 2, or (at your option)
11 * This program is distributed in the hope that it will be useful,
12 * but WITHOUT ANY WARRANTY; without even the implied warranty of
13 * MERCHANTABILITY or FITNESS FOR A PARTICULAR PURPOSE. See the
14 * GNU General Public License for more details.
16 * You should have received a copy of the GNU General Public License
17 * along with this program; if not, write to the Free Software
18 * Foundation, Inc., 675 Mass Ave, Cambridge, MA 02139, USA.
20 * Portions derived from the GNU C Library
21 * Copyright (C) 1991, 1997, 2003 Free Software Foundation, Inc.
23 * Several strategies are tried to try to get the best performance for various
24 * conditions. In the optimal case, we copy 64-bytes in an unrolled loop using
25 * fp regs. This is followed by loops that copy 32- or 16-bytes at a time using
26 * general registers. Unaligned copies are handled either by aligning the
27 * destination and then using shift-and-write method, or in a few cases by
28 * falling back to a byte-at-a-time copy.
30 * I chose to implement this in C because it is easier to maintain and debug,
31 * and in my experiments it appears that the C code generated by gcc (3.3/3.4
32 * at the time of writing) is fairly optimal. Unfortunately some of the
33 * semantics of the copy routine (exception handling) is difficult to express
34 * in C, so we have to play some tricks to get it to work.
36 * All the loads and stores are done via explicit asm() code in order to use
37 * the right space registers.
39 * Testing with various alignments and buffer sizes shows that this code is
40 * often >10x faster than a simple byte-at-a-time copy, even for strangely
41 * aligned operands. It is interesting to note that the glibc version
42 * of memcpy (written in C) is actually quite fast already. This routine is
43 * able to beat it by 30-40% for aligned copies because of the loop unrolling,
44 * but in some cases the glibc version is still slightly faster. This lends
45 * more credibility that gcc can generate very good code as long as we are
49 * - cache prefetching needs more experimentation to get optimal settings
50 * - try not to use the post-increment address modifiers; they create additional
52 * - replace byte-copy loops with stybs sequences
56 #include <linux/module.h>
57 #include <linux/compiler.h>
58 #include <asm/uaccess.h>
59 #define s_space "%%sr1"
60 #define d_space "%%sr2"
63 #define s_space "%%sr0"
64 #define d_space "%%sr0"
65 #define pa_memcpy new2_copy
68 DECLARE_PER_CPU(struct exception_data, exception_data);
70 #define preserve_branch(label) do { \
72 /* The following branch is never taken, it's just here to */ \
73 /* prevent gcc from optimizing away our exception code. */ \
74 if (unlikely(dummy != dummy)) \
78 #define get_user_space() (segment_eq(get_fs(), KERNEL_DS) ? 0 : mfsp(3))
79 #define get_kernel_space() (0)
81 #define MERGE(w0, sh_1, w1, sh_2) ({ \
85 "shrpw %1, %2, %%sar, %0\n" \
87 : "r"(w0), "r"(w1), "r"(sh_2) \
94 #define DPRINTF(fmt, args...) do { printk(KERN_DEBUG "%s:%d:%s ", __FILE__, __LINE__, __FUNCTION__ ); printk(KERN_DEBUG fmt, ##args ); } while (0)
96 #define DPRINTF(fmt, args...)
100 #define EXC_WORD ".word"
102 #define EXC_WORD ".dword"
105 #define def_load_ai_insn(_insn,_sz,_tt,_s,_a,_t,_e) \
106 __asm__ __volatile__ ( \
107 "1:\t" #_insn ",ma " #_sz "(" _s ",%1), %0\n" \
108 "\t.section __ex_table,\"aw\"\n" \
109 "\t" EXC_WORD "\t1b\n" \
110 "\t" EXC_WORD "\t" #_e "\n" \
112 : _tt(_t), "+r"(_a) \
116 #define def_store_ai_insn(_insn,_sz,_tt,_s,_a,_t,_e) \
117 __asm__ __volatile__ ( \
118 "1:\t" #_insn ",ma %1, " #_sz "(" _s ",%0)\n" \
119 "\t.section __ex_table,\"aw\"\n" \
120 "\t" EXC_WORD "\t1b\n" \
121 "\t" EXC_WORD "\t" #_e "\n" \
127 #define ldbma(_s, _a, _t, _e) def_load_ai_insn(ldbs,1,"=r",_s,_a,_t,_e)
128 #define stbma(_s, _t, _a, _e) def_store_ai_insn(stbs,1,"r",_s,_a,_t,_e)
129 #define ldwma(_s, _a, _t, _e) def_load_ai_insn(ldw,4,"=r",_s,_a,_t,_e)
130 #define stwma(_s, _t, _a, _e) def_store_ai_insn(stw,4,"r",_s,_a,_t,_e)
131 #define flddma(_s, _a, _t, _e) def_load_ai_insn(fldd,8,"=f",_s,_a,_t,_e)
132 #define fstdma(_s, _t, _a, _e) def_store_ai_insn(fstd,8,"f",_s,_a,_t,_e)
134 #define def_load_insn(_insn,_tt,_s,_o,_a,_t,_e) \
135 __asm__ __volatile__ ( \
136 "1:\t" #_insn " " #_o "(" _s ",%1), %0\n" \
137 "\t.section __ex_table,\"aw\"\n" \
138 "\t" EXC_WORD "\t1b\n" \
139 "\t" EXC_WORD "\t" #_e "\n" \
145 #define def_store_insn(_insn,_tt,_s,_t,_o,_a,_e) \
146 __asm__ __volatile__ ( \
147 "1:\t" #_insn " %0, " #_o "(" _s ",%1)\n" \
148 "\t.section __ex_table,\"aw\"\n" \
149 "\t" EXC_WORD "\t1b\n" \
150 "\t" EXC_WORD "\t" #_e "\n" \
156 #define ldw(_s,_o,_a,_t,_e) def_load_insn(ldw,"=r",_s,_o,_a,_t,_e)
157 #define stw(_s,_t,_o,_a,_e) def_store_insn(stw,"r",_s,_t,_o,_a,_e)
159 #ifdef CONFIG_PREFETCH
160 extern inline void prefetch_src(const void *addr)
162 __asm__("ldw 0(" s_space ",%0), %%r0" : : "r" (addr));
165 extern inline void prefetch_dst(const void *addr)
167 __asm__("ldd 0(" d_space ",%0), %%r0" : : "r" (addr));
170 #define prefetch_src(addr)
171 #define prefetch_dst(addr)
174 /* Copy from a not-aligned src to an aligned dst, using shifts. Handles 4 words
175 * per loop. This code is derived from glibc.
177 static inline unsigned long copy_dstaligned(unsigned long dst, unsigned long src, unsigned long len, unsigned long o_dst, unsigned long o_src, unsigned long o_len)
179 /* gcc complains that a2 and a3 may be uninitialized, but actually
180 * they cannot be. Initialize a2/a3 to shut gcc up.
182 register unsigned int a0, a1, a2 = 0, a3 = 0;
184 struct exception_data *d;
186 /* prefetch_src((const void *)src); */
188 /* Calculate how to shift a word read at the memory operation
189 aligned srcp to make it aligned for copy. */
190 sh_1 = 8 * (src % sizeof(unsigned int));
191 sh_2 = 8 * sizeof(unsigned int) - sh_1;
193 /* Make src aligned by rounding it down. */
194 src &= -sizeof(unsigned int);
199 /* a1 = ((unsigned int *) src)[0];
200 a2 = ((unsigned int *) src)[1]; */
201 ldw(s_space, 0, src, a1, cda_ldw_exc);
202 ldw(s_space, 4, src, a2, cda_ldw_exc);
203 src -= 1 * sizeof(unsigned int);
204 dst -= 3 * sizeof(unsigned int);
208 /* a0 = ((unsigned int *) src)[0];
209 a1 = ((unsigned int *) src)[1]; */
210 ldw(s_space, 0, src, a0, cda_ldw_exc);
211 ldw(s_space, 4, src, a1, cda_ldw_exc);
212 src -= 0 * sizeof(unsigned int);
213 dst -= 2 * sizeof(unsigned int);
219 /* a3 = ((unsigned int *) src)[0];
220 a0 = ((unsigned int *) src)[1]; */
221 ldw(s_space, 0, src, a3, cda_ldw_exc);
222 ldw(s_space, 4, src, a0, cda_ldw_exc);
223 src -=-1 * sizeof(unsigned int);
224 dst -= 1 * sizeof(unsigned int);
228 /* a2 = ((unsigned int *) src)[0];
229 a3 = ((unsigned int *) src)[1]; */
230 ldw(s_space, 0, src, a2, cda_ldw_exc);
231 ldw(s_space, 4, src, a3, cda_ldw_exc);
232 src -=-2 * sizeof(unsigned int);
233 dst -= 0 * sizeof(unsigned int);
237 goto do4; /* No-op. */
242 /* prefetch_src((const void *)(src + 4 * sizeof(unsigned int))); */
244 /* a0 = ((unsigned int *) src)[0]; */
245 ldw(s_space, 0, src, a0, cda_ldw_exc);
246 /* ((unsigned int *) dst)[0] = MERGE (a2, sh_1, a3, sh_2); */
247 stw(d_space, MERGE (a2, sh_1, a3, sh_2), 0, dst, cda_stw_exc);
249 /* a1 = ((unsigned int *) src)[1]; */
250 ldw(s_space, 4, src, a1, cda_ldw_exc);
251 /* ((unsigned int *) dst)[1] = MERGE (a3, sh_1, a0, sh_2); */
252 stw(d_space, MERGE (a3, sh_1, a0, sh_2), 4, dst, cda_stw_exc);
254 /* a2 = ((unsigned int *) src)[2]; */
255 ldw(s_space, 8, src, a2, cda_ldw_exc);
256 /* ((unsigned int *) dst)[2] = MERGE (a0, sh_1, a1, sh_2); */
257 stw(d_space, MERGE (a0, sh_1, a1, sh_2), 8, dst, cda_stw_exc);
259 /* a3 = ((unsigned int *) src)[3]; */
260 ldw(s_space, 12, src, a3, cda_ldw_exc);
261 /* ((unsigned int *) dst)[3] = MERGE (a1, sh_1, a2, sh_2); */
262 stw(d_space, MERGE (a1, sh_1, a2, sh_2), 12, dst, cda_stw_exc);
264 src += 4 * sizeof(unsigned int);
265 dst += 4 * sizeof(unsigned int);
271 /* ((unsigned int *) dst)[0] = MERGE (a2, sh_1, a3, sh_2); */
272 stw(d_space, MERGE (a2, sh_1, a3, sh_2), 0, dst, cda_stw_exc);
274 preserve_branch(handle_load_error);
275 preserve_branch(handle_store_error);
280 __asm__ __volatile__ ("cda_ldw_exc:\n");
281 d = &__get_cpu_var(exception_data);
282 DPRINTF("cda_ldw_exc: o_len=%lu fault_addr=%lu o_src=%lu ret=%lu\n",
283 o_len, d->fault_addr, o_src, o_len - d->fault_addr + o_src);
284 return o_len * 4 - d->fault_addr + o_src;
287 __asm__ __volatile__ ("cda_stw_exc:\n");
288 d = &__get_cpu_var(exception_data);
289 DPRINTF("cda_stw_exc: o_len=%lu fault_addr=%lu o_dst=%lu ret=%lu\n",
290 o_len, d->fault_addr, o_dst, o_len - d->fault_addr + o_dst);
291 return o_len * 4 - d->fault_addr + o_dst;
295 /* Returns 0 for success, otherwise, returns number of bytes not transferred. */
296 unsigned long pa_memcpy(void *dstp, const void *srcp, unsigned long len)
298 register unsigned long src, dst, t1, t2, t3;
299 register unsigned char *pcs, *pcd;
300 register unsigned int *pws, *pwd;
301 register double *pds, *pdd;
302 unsigned long ret = 0;
303 unsigned long o_dst, o_src, o_len;
304 struct exception_data *d;
306 src = (unsigned long)srcp;
307 dst = (unsigned long)dstp;
308 pcs = (unsigned char *)srcp;
309 pcd = (unsigned char *)dstp;
311 o_dst = dst; o_src = src; o_len = len;
313 /* prefetch_src((const void *)srcp); */
318 /* Check alignment */
320 if (unlikely(t1 & (sizeof(double)-1)))
323 /* src and dst have same alignment. */
325 /* Copy bytes till we are double-aligned. */
326 t2 = src & (sizeof(double) - 1);
327 if (unlikely(t2 != 0)) {
328 t2 = sizeof(double) - t2;
330 /* *pcd++ = *pcs++; */
331 ldbma(s_space, pcs, t3, pmc_load_exc);
333 stbma(d_space, t3, pcd, pmc_store_exc);
342 /* Copy 8 doubles at a time */
343 while (len >= 8*sizeof(double)) {
344 register double r1, r2, r3, r4, r5, r6, r7, r8;
345 /* prefetch_src((char *)pds + L1_CACHE_BYTES); */
346 flddma(s_space, pds, r1, pmc_load_exc);
347 flddma(s_space, pds, r2, pmc_load_exc);
348 flddma(s_space, pds, r3, pmc_load_exc);
349 flddma(s_space, pds, r4, pmc_load_exc);
350 fstdma(d_space, r1, pdd, pmc_store_exc);
351 fstdma(d_space, r2, pdd, pmc_store_exc);
352 fstdma(d_space, r3, pdd, pmc_store_exc);
353 fstdma(d_space, r4, pdd, pmc_store_exc);
356 if (L1_CACHE_BYTES <= 32)
357 prefetch_src((char *)pds + L1_CACHE_BYTES);
359 flddma(s_space, pds, r5, pmc_load_exc);
360 flddma(s_space, pds, r6, pmc_load_exc);
361 flddma(s_space, pds, r7, pmc_load_exc);
362 flddma(s_space, pds, r8, pmc_load_exc);
363 fstdma(d_space, r5, pdd, pmc_store_exc);
364 fstdma(d_space, r6, pdd, pmc_store_exc);
365 fstdma(d_space, r7, pdd, pmc_store_exc);
366 fstdma(d_space, r8, pdd, pmc_store_exc);
367 len -= 8*sizeof(double);
371 pws = (unsigned int *)pds;
372 pwd = (unsigned int *)pdd;
375 while (len >= 8*sizeof(unsigned int)) {
376 register unsigned int r1,r2,r3,r4,r5,r6,r7,r8;
377 /* prefetch_src((char *)pws + L1_CACHE_BYTES); */
378 ldwma(s_space, pws, r1, pmc_load_exc);
379 ldwma(s_space, pws, r2, pmc_load_exc);
380 ldwma(s_space, pws, r3, pmc_load_exc);
381 ldwma(s_space, pws, r4, pmc_load_exc);
382 stwma(d_space, r1, pwd, pmc_store_exc);
383 stwma(d_space, r2, pwd, pmc_store_exc);
384 stwma(d_space, r3, pwd, pmc_store_exc);
385 stwma(d_space, r4, pwd, pmc_store_exc);
387 ldwma(s_space, pws, r5, pmc_load_exc);
388 ldwma(s_space, pws, r6, pmc_load_exc);
389 ldwma(s_space, pws, r7, pmc_load_exc);
390 ldwma(s_space, pws, r8, pmc_load_exc);
391 stwma(d_space, r5, pwd, pmc_store_exc);
392 stwma(d_space, r6, pwd, pmc_store_exc);
393 stwma(d_space, r7, pwd, pmc_store_exc);
394 stwma(d_space, r8, pwd, pmc_store_exc);
395 len -= 8*sizeof(unsigned int);
398 while (len >= 4*sizeof(unsigned int)) {
399 register unsigned int r1,r2,r3,r4;
400 ldwma(s_space, pws, r1, pmc_load_exc);
401 ldwma(s_space, pws, r2, pmc_load_exc);
402 ldwma(s_space, pws, r3, pmc_load_exc);
403 ldwma(s_space, pws, r4, pmc_load_exc);
404 stwma(d_space, r1, pwd, pmc_store_exc);
405 stwma(d_space, r2, pwd, pmc_store_exc);
406 stwma(d_space, r3, pwd, pmc_store_exc);
407 stwma(d_space, r4, pwd, pmc_store_exc);
408 len -= 4*sizeof(unsigned int);
411 pcs = (unsigned char *)pws;
412 pcd = (unsigned char *)pwd;
416 /* *pcd++ = *pcs++; */
417 ldbma(s_space, pcs, t3, pmc_load_exc);
418 stbma(d_space, t3, pcd, pmc_store_exc);
425 /* possibly we are aligned on a word, but not on a double... */
426 if (likely(t1 & (sizeof(unsigned int)-1)) == 0) {
427 t2 = src & (sizeof(unsigned int) - 1);
429 if (unlikely(t2 != 0)) {
430 t2 = sizeof(unsigned int) - t2;
432 /* *pcd++ = *pcs++; */
433 ldbma(s_space, pcs, t3, pmc_load_exc);
434 stbma(d_space, t3, pcd, pmc_store_exc);
440 pws = (unsigned int *)pcs;
441 pwd = (unsigned int *)pcd;
445 /* Align the destination. */
446 if (unlikely((dst & (sizeof(unsigned int) - 1)) != 0)) {
447 t2 = sizeof(unsigned int) - (dst & (sizeof(unsigned int) - 1));
449 /* *pcd++ = *pcs++; */
450 ldbma(s_space, pcs, t3, pmc_load_exc);
451 stbma(d_space, t3, pcd, pmc_store_exc);
455 dst = (unsigned long)pcd;
456 src = (unsigned long)pcs;
459 ret = copy_dstaligned(dst, src, len / sizeof(unsigned int),
460 o_dst, o_src, o_len);
464 pcs += (len & -sizeof(unsigned int));
465 pcd += (len & -sizeof(unsigned int));
466 len %= sizeof(unsigned int);
468 preserve_branch(handle_load_error);
469 preserve_branch(handle_store_error);
474 __asm__ __volatile__ ("pmc_load_exc:\n");
475 d = &__get_cpu_var(exception_data);
476 DPRINTF("pmc_load_exc: o_len=%lu fault_addr=%lu o_src=%lu ret=%lu\n",
477 o_len, d->fault_addr, o_src, o_len - d->fault_addr + o_src);
478 return o_len - d->fault_addr + o_src;
481 __asm__ __volatile__ ("pmc_store_exc:\n");
482 d = &__get_cpu_var(exception_data);
483 DPRINTF("pmc_store_exc: o_len=%lu fault_addr=%lu o_dst=%lu ret=%lu\n",
484 o_len, d->fault_addr, o_dst, o_len - d->fault_addr + o_dst);
485 return o_len - d->fault_addr + o_dst;
489 unsigned long copy_to_user(void __user *dst, const void *src, unsigned long len)
491 mtsp(get_kernel_space(), 1);
492 mtsp(get_user_space(), 2);
493 return pa_memcpy((void __force *)dst, src, len);
496 unsigned long copy_from_user(void *dst, const void __user *src, unsigned long len)
498 mtsp(get_user_space(), 1);
499 mtsp(get_kernel_space(), 2);
500 return pa_memcpy(dst, (void __force *)src, len);
503 unsigned long copy_in_user(void __user *dst, const void __user *src, unsigned long len)
505 mtsp(get_user_space(), 1);
506 mtsp(get_user_space(), 2);
507 return pa_memcpy((void __force *)dst, (void __force *)src, len);
511 void * memcpy(void * dst,const void *src, size_t count)
513 mtsp(get_kernel_space(), 1);
514 mtsp(get_kernel_space(), 2);
515 pa_memcpy(dst, src, count);
519 EXPORT_SYMBOL(copy_to_user);
520 EXPORT_SYMBOL(copy_from_user);
521 EXPORT_SYMBOL(copy_in_user);
522 EXPORT_SYMBOL(memcpy);