URL https://opencores.org/ocsvn/or1k/or1k/trunk

Subversion Repositories or1k

[/] [or1k/] [trunk/] [rc203soc/] [sw/] [uClinux/] [arch/] [sparc/] [lib/] [memcpy.S] - Blame information for rev 1765

Details | Compare with Previous | View Log

Line No.	Rev	Author	Line
1	1624	jcastillo	`! Fast memmove/memcpy/bcopy`
2			`! Copyright Australian National University, 1995`
3			`! This file may be used under the terms of the GNU Public License`
4			`! Author: Paul Mackerras, September 95`
5			`! Minor beautifications David S. Miller`
6
7			`#include`
8
9			`.globl C_LABEL(bcopy)`
10			`C_LABEL(bcopy):`
11			`mov %o0,%o3`
12			`mov %o1,%o0`
13			`mov %o3,%o1`
14
15			`.globl C_LABEL(amemmove)`
16			`C_LABEL(amemmove):`
17			`.globl C_LABEL(memmove)`
18			`.globl C_LABEL(memcpy)`
19			`C_LABEL(memmove):`
20			`C_LABEL(memcpy):`
21			`save %sp,-96,%sp`
22			`mov %i0,%l7`
23
24			`cmp %i0,%i1 ! check for dest within source area`
25			`bleu,a 1f`
26			`andcc %i0,3,%l1`
27			`add %i1,%i2,%l0`
28			`cmp %i0,%l0`
29			`blu,a Lback`
30			`mov %l0,%i1`
31
32			`! copying forwards`
33			`! first get dest to be word-aligned`
34			`andcc %i0,3,%l1`
35			`1:`
36			`be,a Lwalign ! if dest already word-aligned`
37			`cmp %i2,4`
38			`mov 4,%l2`
39			`sub %l2,%l1,%l2 ! #bytes until word-aligned`
40			`subcc %i2,%l2,%i2`
41			`ble,a Lend ! not copying enough to get past word bdry`
42			`addcc %i2,%l2,%i2`
43
44			`1:`
45			`ldub [%i1],%o0 ! copy single bytes until word-aligned`
46			`add %i1,1,%i1`
47			`subcc %l2,1,%l2`
48			`stb %o0,[%i0]`
49			`bgt 1b`
50			`add %i0,1,%i0`
51			`cmp %i2,4`
52
53			`Lwalign: ! dest now word aligned`
54			`blt,a Lend`
55			`orcc %i2,%g0,%g0`
56
57			`andcc %i1,3,%l0`
58			`be,a Ldoword ! if dest word aligned wrt src`
59			`andcc %i0,4,%g0`
60
61			`! yucky cases where we have to shift`
62
63			`mov 4,%l2`
64			`sub %l2,%l0,%l2 ! address adjustment, used at Lendn`
65			`sll %l0,3,%l0 ! bit offset = shift left count`
66			`sll %l2,3,%l1 ! shift right count`
67			`add %i1,%l2,%i1 ! round up to next word`
68			`ld [%i1-4],%o0 ! get first word`
69
70			`andcc %i0,4,%g0 ! get destination double-word aligned`
71			`be,a 1f`
72			`andcc %i1,4,%g0`
73			`ld [%i1],%o1 ! by constructing and storing one word`
74			`add %i0,4,%i0`
75			`add %i1,4,%i1`
76			`sub %i2,4,%i2`
77			`sll %o0,%l0,%o0`
78			`srl %o1,%l1,%l6`
79			`or %o0,%l6,%o0`
80			`st %o0,[%i0-4]`
81			`mov %o1,%o0`
82
83			`andcc %i1,4,%g0 ! now construct & store pairs of double-words`
84			`1:`
85			`bne,a 3f ! if source now not double-word aligned`
86			`subcc %i2,4,%i2`
87			`subcc %i2,16,%i2`
88			`blt 2f`
89			`mov %o0,%o1`
90			`4:`
91			`ldd [%i1],%o2`
92			`sll %o1,%l0,%o4`
93			`ldd [%i1+8],%o0`
94			`add %i0,16,%i0`
95			`add %i1,16,%i1`
96			`subcc %i2,16,%i2`
97			`srl %o2,%l1,%l6`
98			`or %l6,%o4,%o4`
99			`sll %o2,%l0,%o5`
100			`srl %o3,%l1,%l6`
101			`or %l6,%o5,%o5`
102			`std %o4,[%i0-16]`
103			`sll %o3,%l0,%o4`
104			`srl %o0,%l1,%l6`
105			`or %l6,%o4,%o4`
106			`sll %o0,%l0,%o5`
107			`srl %o1,%l1,%l6`
108			`or %l6,%o5,%o5`
109			`bge 4b`
110			`std %o4,[%i0-8]`
111			`2:`
112			`addcc %i2,12,%i2`
113			`blt,a Lendn`
114			`addcc %i2,4,%i2`
115			`5:`
116			`ld [%i1],%o2`
117			`add %i0,4,%i0`
118			`add %i1,4,%i1`
119			`subcc %i2,4,%i2`
120			`sll %o1,%l0,%o0`
121			`srl %o2,%l1,%o1`
122			`or %o1,%o0,%o0`
123			`st %o0,[%i0-4]`
124			`bge 5b`
125			`mov %o2,%o1`
126			`ba Lendn`
127			`addcc %i2,4,%i2`
128
129			`3:`
130			`blt,a Lendn`
131			`addcc %i2,4,%i2`
132			`ld [%i1],%o1`
133			`add %i1,4,%i1`
134			`subcc %i2,16,%i2`
135			`blt,a 8f`
136			`addcc %i2,16,%i2`
137			`7:`
138			`ldd [%i1],%o2`
139			`sll %o0,%l0,%o4`
140			`srl %o1,%l1,%l6`
141			`or %l6,%o4,%o4`
142			`sll %o1,%l0,%o5`
143			`ldd [%i1+8],%o0`
144			`add %i0,16,%i0`
145			`add %i1,16,%i1`
146			`subcc %i2,16,%i2`
147			`srl %o2,%l1,%l6`
148			`or %l6,%o5,%o5`
149			`std %o4,[%i0-16]`
150			`sll %o2,%l0,%o4`
151			`srl %o3,%l1,%l6`
152			`or %l6,%o4,%o4`
153			`sll %o3,%l0,%o5`
154			`srl %o0,%l1,%l6`
155			`or %l6,%o5,%o5`
156			`bge 7b`
157			`std %o4,[%i0-8]`
158			`addcc %i2,16,%i2`
159			`8:`
160			`sll %o0,%l0,%o4`
161			`srl %o1,%l1,%l6`
162			`or %l6,%o4,%o4`
163			`st %o4,[%i0]`
164			`add %i0,4,%i0`
165			`subcc %i2,4,%i2`
166			`blt,a Lendn`
167			`addcc %i2,4,%i2`
168			`mov %o1,%o0`
169			`ld [%i1],%o1`
170			`ba 8b`
171			`add %i1,4,%i1`
172
173
174			`Ldoword:`
175			`! here both dest and src are word-aligned`
176			`! make dest double-word aligned`
177			`be,a 1f`
178			`andcc %i1,4,%g0`
179			`ld [%i1],%o0`
180			`add %i0,4,%i0`
181			`add %i1,4,%i1`
182			`sub %i2,4,%i2`
183			`st %o0,[%i0-4]`
184			`cmp %i2,4`
185			`blt,a Lend`
186			`orcc %i2,%g0,%g0`
187			`andcc %i1,4,%g0`
188
189			`1:`
190			`be,a Ldodble ! if source double-word aligned now`
191			`subcc %i2,32,%i2`
192			`ld [%i1],%o5`
193			`add %i1,4,%i1`
194			`subcc %i2,36,%i2`
195			`blt,a 3f`
196			`add %i2,32,%i2`
197			`2:`
198			`ldd [%i1],%o2`
199			`add %i1,32,%i1`
200			`subcc %i2,32,%i2`
201			`mov %o5,%o0`
202			`ldd [%i1-24],%o4`
203			`mov %o2,%o1`
204			`std %o0,[%i0]`
205			`mov %o3,%o2`
206			`ldd [%i1-16],%o0`
207			`mov %o4,%o3`
208			`std %o2,[%i0+8]`
209			`mov %o5,%o2`
210			`ldd [%i1-8],%o4`
211			`mov %o0,%o3`
212			`std %o2,[%i0+16]`
213			`mov %o1,%o0`
214			`mov %o4,%o1`
215			`std %o0,[%i0+24]`
216			`bge 2b`
217			`add %i0,32,%i0`
218			`add %i2,32,%i2`
219			`3:`
220			`st %o5,[%i0]`
221			`add %i0,4,%i0`
222			`subcc %i2,4,%i2`
223			`blt,a Lend`
224			`addcc %i2,4,%i2`
225			`ld [%i1],%o5`
226			`ba 3b`
227			`add %i1,4,%i1`
228
229			`Ldodble:`
230			`! dest and source are both double-word aligned`
231			`blt,a 2f`
232			`addcc %i2,28,%i2`
233			`1:`
234			`ldd [%i1],%o0 ! copy sets of 4 double-words`
235			`subcc %i2,32,%i2`
236			`ldd [%i1+8],%o2`
237			`add %i1,32,%i1`
238			`ldd [%i1-16],%o4`
239			`add %i0,32,%i0`
240			`std %o0,[%i0-32]`
241			`ldd [%i1-8],%o0`
242			`std %o2,[%i0-24]`
243			`std %o4,[%i0-16]`
244			`bge 1b`
245			`std %o0,[%i0-8]`
246			`addcc %i2,28,%i2`
247			`2:`
248			`blt,a Lend`
249			`addcc %i2,4,%i2`
250			`3:`
251			`ld [%i1],%o0 ! copy words`
252			`add %i1,4,%i1`
253			`add %i0,4,%i0`
254			`subcc %i2,4,%i2`
255			`bge 3b`
256			`st %o0,[%i0-4]`
257			`ba Lend`
258			`addcc %i2,4,%i2`
259
260			`Lendn:`
261			`sub %i1,%l2,%i1`
262			`Lend:`
263			`ble Lout`
264			`nop`
265			`1:`
266			`ldub [%i1],%o0`
267			`add %i1,1,%i1`
268			`subcc %i2,1,%i2`
269			`stb %o0,[%i0]`
270			`bgt 1b`
271			`add %i0,1,%i0`
272
273			`ba Lout`
274			`nop`
275
276			`Lback: ! Here we have to copy backwards`
277			`add %i0,%i2,%i0`
278			`! first get dest to be word-aligned`
279			`andcc %i0,3,%l2 ! #bytes until word-aligned`
280			`be,a Lbwal ! if dest already word-aligned`
281			`cmp %i2,4`
282			`subcc %i2,%l2,%i2`
283			`ble,a Lbend ! not copying enough to get past word bdry`
284			`addcc %i2,%l2,%i2`
285
286			`1:`
287			`ldub [%i1-1],%o0 ! copy single bytes until word-aligned`
288			`sub %i1,1,%i1`
289			`subcc %l2,1,%l2`
290			`stb %o0,[%i0-1]`
291			`bgt 1b`
292			`sub %i0,1,%i0`
293			`cmp %i2,4`
294
295			`Lbwal: ! dest now word aligned`
296			`blt,a Lbend`
297			`orcc %i2,%g0,%g0`
298
299			`andcc %i1,3,%l2`
300			`be,a Lbword ! if dest word aligned wrt src`
301			`andcc %i0,4,%g0`
302
303			`! yucky cases where we have to shift`
304			`! note %l2 used below at Lbendn`
305
306			`mov 4,%l0`
307			`sub %l0,%l2,%l0 ! # bytes to right of src in word`
308			`sll %l0,3,%l0 ! bit offset = shift right count`
309			`sll %l2,3,%l1 ! shift left count`
310			`sub %i1,%l2,%i1 ! round down to word boundary`
311			`ld [%i1],%o1 ! get first word`
312
313			`andcc %i0,4,%g0 ! get destination double-word aligned`
314			`be,a 1f`
315			`andcc %i1,4,%g0`
316			`ld [%i1-4],%o0 ! by constructing and storing one word`
317			`sub %i0,4,%i0`
318			`sub %i1,4,%i1`
319			`sub %i2,4,%i2`
320			`srl %o1,%l0,%o1`
321			`sll %o0,%l1,%l6`
322			`or %o1,%l6,%o1`
323			`st %o1,[%i0]`
324			`mov %o0,%o1`
325
326			`andcc %i1,4,%g0 ! now construct & store pairs of double-words`
327			`1:`
328			`bne,a 3f ! if source now not double-word aligned`
329			`subcc %i2,4,%i2`
330			`subcc %i2,16,%i2`
331			`blt 2f`
332			`mov %o1,%o0`
333			`4:`
334			`ldd [%i1-8],%o2`
335			`srl %o0,%l0,%o5`
336			`ldd [%i1-16],%o0`
337			`sub %i0,16,%i0`
338			`sub %i1,16,%i1`
339			`subcc %i2,16,%i2`
340			`sll %o3,%l1,%l6`
341			`or %l6,%o5,%o5`
342			`srl %o3,%l0,%o4`
343			`sll %o2,%l1,%l6`
344			`or %l6,%o4,%o4`
345			`std %o4,[%i0+8]`
346			`srl %o2,%l0,%o5`
347			`sll %o1,%l1,%l6`
348			`or %l6,%o5,%o5`
349			`srl %o1,%l0,%o4`
350			`sll %o0,%l1,%l6`
351			`or %l6,%o4,%o4`
352			`bge 4b`
353			`std %o4,[%i0]`
354			`2:`
355			`addcc %i2,12,%i2`
356			`blt,a Lbendn`
357			`addcc %i2,4,%i2`
358			`5:`
359			`ld [%i1-4],%o2`
360			`sub %i0,4,%i0`
361			`sub %i1,4,%i1`
362			`subcc %i2,4,%i2`
363			`srl %o0,%l0,%o0`
364			`sll %o2,%l1,%o1`
365			`or %o1,%o0,%o0`
366			`st %o0,[%i0]`
367			`bge 5b`
368			`mov %o2,%o0`
369			`ba Lbendn`
370			`addcc %i2,4,%i2`
371
372			`3:`
373			`blt,a Lbendn`
374			`addcc %i2,4,%i2`
375			`ld [%i1-4],%o0`
376			`sub %i1,4,%i1`
377			`subcc %i2,16,%i2`
378			`blt,a 8f`
379			`addcc %i2,16,%i2`
380			`7:`
381			`ldd [%i1-8],%o2`
382			`srl %o1,%l0,%o5`
383			`sll %o0,%l1,%l6`
384			`or %l6,%o5,%o5`
385			`srl %o0,%l0,%o4`
386			`ldd [%i1-16],%o0`
387			`sub %i0,16,%i0`
388			`sub %i1,16,%i1`
389			`subcc %i2,16,%i2`
390			`sll %o3,%l1,%l6`
391			`or %l6,%o4,%o4`
392			`std %o4,[%i0+8]`
393			`srl %o3,%l0,%o5`
394			`sll %o2,%l1,%l6`
395			`or %l6,%o5,%o5`
396			`srl %o2,%l0,%o4`
397			`sll %o1,%l1,%l6`
398			`or %l6,%o4,%o4`
399			`bge 7b`
400			`std %o4,[%i0]`
401			`addcc %i2,16,%i2`
402			`8:`
403			`srl %o1,%l0,%o5`
404			`sll %o0,%l1,%l6`
405			`or %l6,%o5,%o5`
406			`st %o5,[%i0-4]`
407			`sub %i0,4,%i0`
408			`subcc %i2,4,%i2`
409			`blt,a Lbendn`
410			`addcc %i2,4,%i2`
411			`mov %o0,%o1`
412			`ld [%i1-4],%o0`
413			`ba 8b`
414			`sub %i1,4,%i1`
415
416
417			`Lbword:`
418			`! here both dest and src are word-aligned`
419			`! make dest double-word aligned`
420			`be,a 1f`
421			`andcc %i1,4,%g0`
422			`ld [%i1-4],%o0`
423			`sub %i0,4,%i0`
424			`sub %i1,4,%i1`
425			`sub %i2,4,%i2`
426			`st %o0,[%i0]`
427			`cmp %i2,4`
428			`blt,a Lbend`
429			`orcc %i2,%g0,%g0`
430			`andcc %i1,4,%g0`
431
432			`1:`
433			`be,a Lbdble ! if source double-word aligned now`
434			`subcc %i2,32,%i2`
435			`ld [%i1-4],%o4`
436			`sub %i1,4,%i1`
437			`subcc %i2,36,%i2`
438			`blt,a 3f`
439			`add %i2,32,%i2`
440			`2:`
441			`ldd [%i1-8],%o2`
442			`sub %i1,32,%i1`
443			`subcc %i2,32,%i2`
444			`mov %o4,%o1`
445			`ldd [%i1+16],%o4`
446			`mov %o3,%o0`
447			`std %o0,[%i0-8]`
448			`mov %o2,%o3`
449			`ldd [%i1+8],%o0`
450			`mov %o5,%o2`
451			`std %o2,[%i0-16]`
452			`mov %o4,%o3`
453			`ldd [%i1],%o4`
454			`mov %o1,%o2`
455			`std %o2,[%i0-24]`
456			`mov %o0,%o1`
457			`mov %o5,%o0`
458			`std %o0,[%i0-32]`
459			`bge 2b`
460			`sub %i0,32,%i0`
461			`add %i2,32,%i2`
462			`3:`
463			`st %o4,[%i0-4]`
464			`sub %i0,4,%i0`
465			`subcc %i2,4,%i2`
466			`blt,a Lbend`
467			`addcc %i2,4,%i2`
468			`ld [%i1-4],%o4`
469			`ba 3b`
470			`sub %i1,4,%i1`
471
472			`Lbdble:`
473			`! dest and source are both double-word aligned`
474			`blt,a 2f`
475			`addcc %i2,28,%i2`
476			`1:`
477			`ldd [%i1-8],%o0 ! copy sets of 4 double-words`
478			`subcc %i2,32,%i2`
479			`ldd [%i1-16],%o2`
480			`sub %i1,32,%i1`
481			`ldd [%i1+8],%o4`
482			`sub %i0,32,%i0`
483			`std %o0,[%i0+24]`
484			`ldd [%i1],%o0`
485			`std %o2,[%i0+16]`
486			`std %o4,[%i0+8]`
487			`bge 1b`
488			`std %o0,[%i0]`
489			`addcc %i2,28,%i2`
490			`2:`
491			`blt,a Lbend`
492			`addcc %i2,4,%i2`
493			`3:`
494			`ld [%i1-4],%o0 ! copy words`
495			`sub %i1,4,%i1`
496			`sub %i0,4,%i0`
497			`subcc %i2,4,%i2`
498			`bge 3b`
499			`st %o0,[%i0]`
500			`ba Lbend`
501			`addcc %i2,4,%i2`
502
503			`Lbendn:`
504			`add %i1,%l2,%i1`
505			`Lbend:`
506			`ble Lout`
507			`nop`
508			`1:`
509			`ldub [%i1-1],%o0`
510			`sub %i1,1,%i1`
511			`subcc %i2,1,%i2`
512			`stb %o0,[%i0-1]`
513			`bgt 1b`
514			`sub %i0,1,%i0`
515
516			`Lout:`
517			`ret`
518			`restore %l7,0,%o0`
519
520

Browse

Tools

Subversion Repositories or1k

[/] [or1k/] [trunk/] [rc203soc/] [sw/] [uClinux/] [arch/] [sparc/] [lib/] [memcpy.S] - Blame information for rev 1765