bzero.S revision 1.4 1 1.4 eeh /* $NetBSD: bzero.S,v 1.4 2002/03/12 22:08:22 eeh Exp $ */
2 1.1 mjl
3 1.1 mjl /*-
4 1.1 mjl * Copyright (C) 2001 Martin J. Laubach <mjl (at) netbsd.org>
5 1.1 mjl * All rights reserved.
6 1.1 mjl *
7 1.1 mjl * Redistribution and use in source and binary forms, with or without
8 1.1 mjl * modification, are permitted provided that the following conditions
9 1.1 mjl * are met:
10 1.1 mjl * 1. Redistributions of source code must retain the above copyright
11 1.1 mjl * notice, this list of conditions and the following disclaimer.
12 1.1 mjl * 2. Redistributions in binary form must reproduce the above copyright
13 1.1 mjl * notice, this list of conditions and the following disclaimer in the
14 1.1 mjl * documentation and/or other materials provided with the distribution.
15 1.1 mjl * 3. The name of the author may not be used to endorse or promote products
16 1.1 mjl * derived from this software without specific prior written permission.
17 1.1 mjl *
18 1.1 mjl * THIS SOFTWARE IS PROVIDED BY THE AUTHOR ``AS IS'' AND ANY EXPRESS OR
19 1.1 mjl * IMPLIED WARRANTIES, INCLUDING, BUT NOT LIMITED TO, THE IMPLIED WARRANTIES
20 1.1 mjl * OF MERCHANTABILITY AND FITNESS FOR A PARTICULAR PURPOSE ARE DISCLAIMED.
21 1.1 mjl * IN NO EVENT SHALL THE AUTHOR BE LIABLE FOR ANY DIRECT, INDIRECT,
22 1.1 mjl * INCIDENTAL, SPECIAL, EXEMPLARY, OR CONSEQUENTIAL DAMAGES (INCLUDING, BUT
23 1.1 mjl * NOT LIMITED TO, PROCUREMENT OF SUBSTITUTE GOODS OR SERVICES; LOSS OF USE,
24 1.1 mjl * DATA, OR PROFITS; OR BUSINESS INTERRUPTION) HOWEVER CAUSED AND ON ANY
25 1.1 mjl * THEORY OF LIABILITY, WHETHER IN CONTRACT, STRICT LIABILITY, OR TORT
26 1.1 mjl * (INCLUDING NEGLIGENCE OR OTHERWISE) ARISING IN ANY WAY OUT OF THE USE
27 1.1 mjl * OF THIS SOFTWARE, EVEN IF ADVISED OF THE POSSIBILITY OF SUCH DAMAGE.
28 1.1 mjl */
29 1.1 mjl /*----------------------------------------------------------------------*/
30 1.1 mjl
31 1.1 mjl #include <machine/asm.h>
32 1.2 mjl #ifdef _KERNEL
33 1.2 mjl #include <assym.h>
34 1.2 mjl #endif
35 1.1 mjl
36 1.1 mjl #define USE_STSWX 0 /* don't. slower than trivial copy loop */
37 1.1 mjl
38 1.1 mjl /*----------------------------------------------------------------------*/
39 1.1 mjl /*
40 1.1 mjl void bzero(void *b r3, size_t len r4);
41 1.1 mjl void * memset(void *b r3, int c r4, size_t len r5);
42 1.1 mjl */
43 1.1 mjl /*----------------------------------------------------------------------*/
44 1.1 mjl
45 1.1 mjl #define r_dst r3
46 1.1 mjl #define r_len r4
47 1.1 mjl #define r_val r0
48 1.1 mjl
49 1.1 mjl .text
50 1.1 mjl .align 4
51 1.1 mjl ENTRY(bzero)
52 1.1 mjl li r_val, 0 /* Value to stuff in */
53 1.1 mjl b cb_memset
54 1.1 mjl
55 1.1 mjl ENTRY(memset)
56 1.1 mjl cmplwi cr1, r5, 0
57 1.1 mjl mr. r0, r4
58 1.1 mjl mr r8, r3
59 1.1 mjl beqlr- cr1 /* Nothing to do */
60 1.1 mjl
61 1.1 mjl rlwimi r0, r4, 8, 16, 23 /* word extend fill value */
62 1.1 mjl rlwimi r0, r0, 16, 0, 15
63 1.1 mjl mr r4, r5
64 1.1 mjl bne- simple_fill /* =! 0, use trivial fill */
65 1.1 mjl cb_memset:
66 1.1 mjl
67 1.1 mjl /*----------------------------------------------------------------------*/
68 1.2 mjl #ifndef _KERNEL
69 1.1 mjl /* First find out cache line size */
70 1.1 mjl #ifdef PIC
71 1.1 mjl mflr r9
72 1.1 mjl bl _GLOBAL_OFFSET_TABLE_@local-4
73 1.1 mjl mflr r10
74 1.1 mjl mtlr r9
75 1.4 eeh lwz r5,cache_info@got(r10)
76 1.1 mjl #else
77 1.4 eeh lis r5,cache_info@h
78 1.4 eeh ori r5,r5,cache_info@l
79 1.1 mjl #endif
80 1.4 eeh lwz r6, 4(r5)
81 1.1 mjl cmpwi r6, -1
82 1.1 mjl bne+ cb_cacheline_known
83 1.1 mjl
84 1.1 mjl /*----------------------------------------------------------------------*/
85 1.1 mjl #define CTL_MACHDEP 7
86 1.1 mjl #define CPU_CACHELINE 1
87 1.4 eeh #define CPU_CACHEINFO 5
88 1.1 mjl
89 1.1 mjl #define STKFRAME_SZ 48
90 1.1 mjl #define MIB 8
91 1.1 mjl #define OLDPLEN 16
92 1.1 mjl #define R3_SAVE 20
93 1.1 mjl #define R4_SAVE 24
94 1.1 mjl #define R0_SAVE 28
95 1.1 mjl #define R8_SAVE 32
96 1.1 mjl
97 1.1 mjl mflr r6
98 1.1 mjl stw r6, 4(r1)
99 1.1 mjl stwu r1, -STKFRAME_SZ(r1)
100 1.1 mjl
101 1.1 mjl stw r8, R8_SAVE(r1)
102 1.1 mjl stw r3, R3_SAVE(r1)
103 1.1 mjl stw r4, R4_SAVE(r1)
104 1.1 mjl stw r0, R0_SAVE(r1)
105 1.1 mjl
106 1.4 eeh
107 1.4 eeh
108 1.4 eeh li r0, CTL_MACHDEP /* Construct MIB */
109 1.4 eeh stw r0, MIB(r1)
110 1.4 eeh li r0, CPU_CACHEINFO
111 1.4 eeh stw r0, MIB+4(r1)
112 1.4 eeh
113 1.4 eeh li r0, 4*4 /* Oldlenp := 4*4 */
114 1.4 eeh stw r0, OLDPLEN(r1)
115 1.4 eeh
116 1.4 eeh addi r3, r1, MIB
117 1.4 eeh li r4, 2 /* namelen */
118 1.4 eeh /* r5 already contains &cache_info */
119 1.4 eeh addi r6, r1, OLDPLEN
120 1.4 eeh li r7, 0
121 1.4 eeh li r8, 0
122 1.4 eeh bl PIC_PLT(_C_LABEL(sysctl))
123 1.4 eeh
124 1.4 eeh cmpwi r3, 0 /* Check result */
125 1.4 eeh beq 1f
126 1.4 eeh
127 1.4 eeh /* Failure, try older sysctl */
128 1.4 eeh
129 1.1 mjl li r0, CTL_MACHDEP /* Construct MIB */
130 1.1 mjl stw r0, MIB(r1)
131 1.1 mjl li r0, CPU_CACHELINE
132 1.1 mjl stw r0, MIB+4(r1)
133 1.1 mjl
134 1.1 mjl li r0, 4 /* Oldlenp := 4 */
135 1.1 mjl stw r0, OLDPLEN(r1)
136 1.1 mjl
137 1.1 mjl addi r3, r1, MIB
138 1.1 mjl li r4, 2 /* namelen */
139 1.4 eeh #ifdef PIC
140 1.4 eeh mflr r9
141 1.4 eeh bl _GLOBAL_OFFSET_TABLE_@local-4
142 1.4 eeh mflr r10
143 1.4 eeh mtlr r9
144 1.4 eeh lwz r5,cache_info@got(r10)
145 1.4 eeh addi r5, r5, 4
146 1.4 eeh #else
147 1.4 eeh lis r5,cache_info+4@h
148 1.4 eeh ori r5,r5,cache_info+4@l
149 1.4 eeh #endif
150 1.1 mjl addi r6, r1, OLDPLEN
151 1.1 mjl li r7, 0
152 1.1 mjl li r8, 0
153 1.1 mjl bl PIC_PLT(_C_LABEL(sysctl))
154 1.4 eeh 1:
155 1.1 mjl lwz r8, R8_SAVE(r1)
156 1.1 mjl lwz r3, R3_SAVE(r1)
157 1.1 mjl lwz r4, R4_SAVE(r1)
158 1.1 mjl lwz r0, R0_SAVE(r1)
159 1.1 mjl
160 1.1 mjl #ifdef PIC
161 1.1 mjl bl _GLOBAL_OFFSET_TABLE_@local-4
162 1.1 mjl mflr r10
163 1.4 eeh lwz r9, cache_info@got(r10)
164 1.4 eeh lwz r9, 4(r9)
165 1.1 mjl #else
166 1.4 eeh lis r5, cache_info+4@ha
167 1.4 eeh lwz r9, cache_info+4@l(r5)
168 1.1 mjl #endif
169 1.1 mjl la r1, STKFRAME_SZ(r1)
170 1.1 mjl lwz r5, 4(r1)
171 1.1 mjl mtlr r5
172 1.1 mjl
173 1.1 mjl cntlzw r6, r9 /* compute shift value */
174 1.1 mjl li r5, 31
175 1.1 mjl subf r5, r6, r5
176 1.1 mjl
177 1.1 mjl #ifdef PIC
178 1.1 mjl lwz r6, cache_sh@got(r10)
179 1.1 mjl stw r5, 0(r6)
180 1.1 mjl #else
181 1.1 mjl lis r6, cache_sh@ha
182 1.1 mjl stw r5, cache_sh@l(r6)
183 1.1 mjl #endif
184 1.1 mjl /*----------------------------------------------------------------------*/
185 1.1 mjl /* Okay, we know the cache line size (r9) and shift value (r10) */
186 1.1 mjl cb_cacheline_known:
187 1.1 mjl #ifdef PIC
188 1.4 eeh lwz r5, cache_info@got(r10)
189 1.4 eeh lwz r9, 4(r5)
190 1.1 mjl lwz r5, cache_sh@got(r10)
191 1.1 mjl lwz r10, 0(r5)
192 1.1 mjl #else
193 1.4 eeh lis r9, cache_info+4@ha
194 1.4 eeh lwz r9, cache_info+4@l(r9)
195 1.1 mjl lis r10, cache_sh@ha
196 1.1 mjl lwz r10, cache_sh@l(r10)
197 1.1 mjl #endif
198 1.2 mjl
199 1.2 mjl #else /* _KERNEL */
200 1.4 eeh #ifdef MULTIPROCESSOR
201 1.4 eeh mfspr r10, 0 /* Get cpu_info pointer */
202 1.2 mjl #else
203 1.4 eeh lis r10, cpu_info_store@ha
204 1.4 eeh addi r10, r10, cpu_info_store@l
205 1.2 mjl #endif
206 1.4 eeh lwz r9, CPU_CI+4(r10) /* Load D$ line size */
207 1.4 eeh cntlzw r10, r9 /* Calculate shift.. */
208 1.4 eeh li r6, 31
209 1.4 eeh subf r10, r10, r6
210 1.2 mjl #endif /* _KERNEL */
211 1.1 mjl /* Back in memory filling business */
212 1.1 mjl
213 1.1 mjl cmplwi cr1, r_len, 0 /* Nothing to do? */
214 1.1 mjl add r5, r9, r9
215 1.1 mjl cmplw r_len, r5 /* <= 2*CL bytes to move? */
216 1.1 mjl beqlr- cr1 /* then do nothing */
217 1.1 mjl
218 1.1 mjl blt+ simple_fill /* a trivial fill routine */
219 1.1 mjl
220 1.1 mjl /* Word align the block, fill bytewise until dst even*/
221 1.1 mjl
222 1.1 mjl andi. r5, r_dst, 0x03
223 1.1 mjl li r6, 4
224 1.1 mjl beq+ cb_aligned_w /* already aligned to word? */
225 1.1 mjl
226 1.1 mjl subf r5, r5, r6 /* bytes to fill to align4 */
227 1.1 mjl #if USE_STSWX
228 1.1 mjl mtxer r5
229 1.1 mjl stswx r0, 0, r_dst
230 1.1 mjl add r_dst, r5, r_dst
231 1.1 mjl #else
232 1.1 mjl mtctr r5
233 1.1 mjl
234 1.1 mjl subi r_dst, r_dst, 1
235 1.1 mjl 1: stbu r_val, 1(r_dst) /* Fill bytewise */
236 1.1 mjl bdnz 1b
237 1.1 mjl
238 1.1 mjl addi r_dst, r_dst, 1
239 1.1 mjl #endif
240 1.1 mjl subf r_len, r5, r_len
241 1.1 mjl
242 1.1 mjl cb_aligned_w: /* Cache block align, fill wordwise until dst aligned */
243 1.1 mjl
244 1.1 mjl /* I know I have something to do since we had > 2*CL initially */
245 1.1 mjl /* so no need to check for r_len = 0 */
246 1.1 mjl
247 1.1 mjl rlwinm. r5, r_dst, 30, 29, 31
248 1.1 mjl srwi r6, r9, 2
249 1.1 mjl beq cb_aligned_cb /* already on CL boundary? */
250 1.1 mjl
251 1.1 mjl subf r5, r5, r6 /* words to fill to alignment */
252 1.1 mjl mtctr r5
253 1.1 mjl slwi r5, r5, 2
254 1.1 mjl subf r_len, r5, r_len
255 1.1 mjl
256 1.1 mjl subi r_dst, r_dst, 4
257 1.1 mjl 1: stwu r_val, 4(r_dst) /* Fill wordwise */
258 1.1 mjl bdnz 1b
259 1.1 mjl addi r_dst, r_dst, 4
260 1.1 mjl
261 1.1 mjl cb_aligned_cb: /* no need to check r_len, see above */
262 1.1 mjl
263 1.1 mjl srw. r5, r_len, r10 /* Number of cache blocks */
264 1.1 mjl mtctr r5
265 1.1 mjl beq cblocks_done
266 1.1 mjl
267 1.1 mjl slw r5, r5, r10
268 1.1 mjl subf r_len, r5, r_len
269 1.1 mjl
270 1.1 mjl 1: dcbz 0, r_dst /* Clear blockwise */
271 1.1 mjl add r_dst, r_dst, r9
272 1.1 mjl bdnz 1b
273 1.1 mjl
274 1.1 mjl cblocks_done: /* still CL aligned, but less than CL bytes left */
275 1.1 mjl cmplwi cr1, r_len, 0
276 1.1 mjl cmplwi r_len, 8
277 1.1 mjl beq- cr1, sf_return
278 1.1 mjl
279 1.1 mjl blt- sf_bytewise /* <8 remaining? */
280 1.1 mjl b sf_aligned_w
281 1.1 mjl
282 1.1 mjl /*----------------------------------------------------------------------*/
283 1.1 mjl wbzero: li r_val, 0
284 1.1 mjl
285 1.1 mjl cmplwi r_len, 0
286 1.1 mjl beqlr- /* Nothing to do */
287 1.1 mjl
288 1.1 mjl simple_fill:
289 1.1 mjl #if USE_STSWX
290 1.1 mjl cmplwi cr1, r_len, 12 /* < 12 bytes to move? */
291 1.1 mjl #else
292 1.1 mjl cmplwi cr1, r_len, 8 /* < 8 bytes to move? */
293 1.1 mjl #endif
294 1.1 mjl andi. r5, r_dst, 0x03 /* bytes to fill to align4 */
295 1.1 mjl blt cr1, sf_bytewise /* trivial byte mover */
296 1.1 mjl
297 1.1 mjl li r6, 4
298 1.1 mjl subf r5, r5, r6
299 1.1 mjl beq+ sf_aligned_w /* dest is word aligned */
300 1.1 mjl
301 1.1 mjl #if USE_STSWX
302 1.1 mjl mtxer r5
303 1.1 mjl stswx r0, 0, r_dst
304 1.1 mjl add r_dst, r5, r_dst
305 1.1 mjl #else
306 1.1 mjl mtctr r5 /* nope, then fill bytewise */
307 1.1 mjl subi r_dst, r_dst, 1 /* until it is */
308 1.1 mjl 1: stbu r_val, 1(r_dst)
309 1.1 mjl bdnz 1b
310 1.1 mjl
311 1.1 mjl addi r_dst, r_dst, 1
312 1.1 mjl #endif
313 1.1 mjl subf r_len, r5, r_len
314 1.1 mjl
315 1.1 mjl sf_aligned_w: /* no need to check r_len since it were >= 8 bytes initially */
316 1.1 mjl #if USE_STSWX
317 1.1 mjl mr r6, r0
318 1.1 mjl mr r7, r0
319 1.1 mjl
320 1.1 mjl srwi r5, r_len, 3
321 1.1 mjl mtctr r5
322 1.1 mjl
323 1.1 mjl slwi r5, r5, 3 /* adjust len */
324 1.1 mjl subf. r_len, r5, r_len
325 1.1 mjl
326 1.1 mjl 1: stswi r6, r_dst, 8
327 1.1 mjl addi r_dst, r_dst, 8
328 1.1 mjl bdnz 1b
329 1.1 mjl #else
330 1.1 mjl srwi r5, r_len, 2 /* words to fill */
331 1.1 mjl mtctr r5
332 1.1 mjl
333 1.1 mjl slwi r5, r5, 2
334 1.1 mjl subf. r_len, r5, r_len /* adjust len for fill */
335 1.1 mjl
336 1.1 mjl subi r_dst, r_dst, 4
337 1.1 mjl 1: stwu r_val, 4(r_dst)
338 1.1 mjl bdnz 1b
339 1.1 mjl addi r_dst, r_dst, 4
340 1.1 mjl #endif
341 1.1 mjl
342 1.1 mjl sf_word_done: bne- sf_bytewise
343 1.1 mjl
344 1.1 mjl sf_return: mr r3, r8 /* restore orig ptr */
345 1.1 mjl blr /* for memset functionality */
346 1.1 mjl
347 1.1 mjl sf_bytewise:
348 1.1 mjl #if USE_STSWX
349 1.1 mjl mr r5, r0
350 1.1 mjl mr r6, r0
351 1.1 mjl mr r7, r0
352 1.1 mjl
353 1.1 mjl mtxer r_len
354 1.1 mjl stswx r5, 0, r_dst
355 1.1 mjl #else
356 1.1 mjl mtctr r_len
357 1.1 mjl
358 1.1 mjl subi r_dst, r_dst, 1
359 1.1 mjl 1: stbu r_val, 1(r_dst)
360 1.1 mjl bdnz 1b
361 1.1 mjl #endif
362 1.1 mjl mr r3, r8 /* restore orig ptr */
363 1.1 mjl blr /* for memset functionality */
364 1.1 mjl
365 1.1 mjl /*----------------------------------------------------------------------*/
366 1.3 mjl #ifndef _KERNEL
367 1.1 mjl .data
368 1.4 eeh cache_info: .long -1, -1, -1, -1
369 1.1 mjl cache_sh: .long 0
370 1.1 mjl
371 1.3 mjl #endif
372 1.1 mjl /*----------------------------------------------------------------------*/
373