1 /* SPDX-License-Identifier: GPL-2.0 */
2 /*
3 * arch/alpha/lib/ev6-stxncpy.S
4 * 21264 version contributed by Rick Gorton <ri
5 *
6 * Copy no more than COUNT bytes of the null-t
7 * SRC to DST.
8 *
9 * This is an internal routine used by strncpy
10 * As such, it uses special linkage convention
11 * of these public functions more efficient.
12 *
13 * On input:
14 * t9 = return address
15 * a0 = DST
16 * a1 = SRC
17 * a2 = COUNT
18 *
19 * Furthermore, COUNT may not be zero.
20 *
21 * On output:
22 * t0 = last word written
23 * t10 = bitmask (with one bit set) indic
24 * the end of the range specified b
25 * t12 = bitmask (with one bit set) indic
26 * a0 = unaligned address of the last *w
27 * a2 = the number of full words left in
28 *
29 * Furthermore, v0, a3-a5, t11, and $at are un
30 *
31 * Much of the information about 21264 schedul
32 * Compiler Writer's Guide for the Alpha
33 * abbreviated as 'CWG' in other comments
34 * ftp.digital.com/pub/Digital/info/semic
35 * Scheduling notation:
36 * E - either cluster
37 * U - upper subcluster; U0 - subcl
38 * L - lower subcluster; L0 - subcl
39 * Try not to change the actual algorithm if p
40 */
41
42 #include <asm/regdef.h>
43
44 .set noat
45 .set noreorder
46
47 .text
48
49 /* There is a problem with either gdb (as of 4
50 doesn't like putting the entry point for a
51 middle of the procedure descriptor. Work a
52 aligned copy in its own procedure descripto
53
54
55 .ent stxncpy_aligned
56 .align 4
57 stxncpy_aligned:
58 .frame sp, 0, t9, 0
59 .prologue 0
60
61 /* On entry to this basic block:
62 t0 == the first destination word fo
63 t1 == the first source word. */
64
65 /* Create the 1st output word and dete
66 lda t2, -1 # E : build a
67 mskqh t2, a1, t2 # U : detect
68 mskqh t1, a1, t3 # U :
69 ornot t1, t2, t2 # E : (stall)
70
71 mskql t0, a1, t0 # U : assemble
72 cmpbge zero, t2, t8 # E : bits set
73 or t0, t3, t0 # E : (stall)
74 beq a2, $a_eoc # U :
75
76 bne t8, $a_eos # U :
77 nop
78 nop
79 nop
80
81 /* On entry to this basic block:
82 t0 == a source word not containing
83
84 /*
85 * nops here to:
86 * separate store quads from load
87 * limit of 1 bcond/quad to permi
88 */
89 $a_loop:
90 stq_u t0, 0(a0) # L :
91 addq a0, 8, a0 # E :
92 subq a2, 1, a2 # E :
93 nop
94
95 ldq_u t0, 0(a1) # L :
96 addq a1, 8, a1 # E :
97 cmpbge zero, t0, t8 # E :
98 beq a2, $a_eoc # U :
99
100 beq t8, $a_loop # U :
101 nop
102 nop
103 nop
104
105 /* Take care of the final (partial) wo
106 the end-of-count bit is set in t8 i
107
108 On entry to this basic block we hav
109 t0 == the source word containing th
110 t8 == the cmpbge mask that found it
111
112 $a_eos:
113 negq t8, t12 # E : find low
114 and t8, t12, t12 # E : (stall)
115 /* For the sake of the cache, don't re
116 if we're not going to need it. */
117 and t12, 0x80, t6 # E : (stall)
118 bne t6, 1f # U : (stall)
119
120 /* We're doing a partial word store an
121 our source and original destination
122 ldq_u t1, 0(a0) # L :
123 subq t12, 1, t6 # E :
124 or t12, t6, t8 # E : (stall)
125 zapnot t0, t8, t0 # U : clear sr
126
127 zap t1, t8, t1 # .. e1 : clea
128 or t0, t1, t0 # e1 : (sta
129 nop
130 nop
131
132 1: stq_u t0, 0(a0) # L :
133 ret (t9) # L0 : Latency
134 nop
135 nop
136
137 /* Add the end-of-count bit to the eos
138 $a_eoc:
139 or t10, t8, t8 # E :
140 br $a_eos # L0 : Latency
141 nop
142 nop
143
144 .end stxncpy_aligned
145
146 .align 4
147 .ent __stxncpy
148 .globl __stxncpy
149 __stxncpy:
150 .frame sp, 0, t9, 0
151 .prologue 0
152
153 /* Are source and destination co-align
154 xor a0, a1, t1 # E :
155 and a0, 7, t0 # E : find des
156 and t1, 7, t1 # E : (stall)
157 addq a2, t0, a2 # E : bias cou
158
159 subq a2, 1, a2 # E :
160 and a2, 7, t2 # E : (stall)
161 srl a2, 3, a2 # U : a2 = loo
162 addq zero, 1, t10 # E :
163
164 sll t10, t2, t10 # U : t10 = bi
165 bne t1, $unaligned # U :
166 /* We are co-aligned; take care of a p
167 ldq_u t1, 0(a1) # L : load fir
168 addq a1, 8, a1 # E :
169
170 beq t0, stxncpy_aligned # U :
171 ldq_u t0, 0(a0) # L :
172 nop
173 nop
174
175 br stxncpy_aligned # .. e1 :
176 nop
177 nop
178 nop
179
180
181
182 /* The source and destination are not co-align
183 and cope. We have to be very careful about
184 causing a SEGV. */
185
186 .align 4
187 $u_head:
188 /* We know just enough now to be able
189 full source word. We can still fin
190 that prevents us from outputting th
191
192 On entry to this basic block:
193 t0 == the first dest word, unmasked
194 t1 == the shifted low bits of the f
195 t6 == bytemask that is -1 in dest w
196
197 ldq_u t2, 8(a1) # L : Latency=
198 addq a1, 8, a1 # E :
199 mskql t0, a0, t0 # U : mask tra
200 extqh t2, a1, t4 # U : (3 cycle
201
202 or t1, t4, t1 # E : first al
203 mskqh t1, a0, t1 # U : mask lea
204 or t0, t1, t0 # E : first ou
205 or t0, t6, t6 # E : mask ori
206
207 cmpbge zero, t6, t8 # E :
208 beq a2, $u_eocfin # U :
209 lda t6, -1 # E :
210 nop
211
212 bne t8, $u_final # U :
213 mskql t6, a1, t6 # U : mask out
214 stq_u t0, 0(a0) # L : store fi
215 or t6, t2, t2 # E : (stall)
216
217 cmpbge zero, t2, t8 # E : find nul
218 addq a0, 8, a0 # E :
219 subq a2, 1, a2 # E :
220 bne t8, $u_late_head_exit # U :
221
222 /* Finally, we've got all the stupid l
223 of and we can set up to enter the m
224 extql t2, a1, t1 # U : position
225 beq a2, $u_eoc # U :
226 ldq_u t2, 8(a1) # L : read nex
227 addq a1, 8, a1 # E :
228
229 extqh t2, a1, t0 # U : position
230 cmpbge zero, t2, t8 # E :
231 nop
232 bne t8, $u_eos # U :
233
234 /* Unaligned copy main loop. In order
235 the loop is structured to detect ze
236 This has, unfortunately, effectivel
237 iteration out into the head and hal
238 prevent nastiness from accumulating
239 to run as fast as possible.
240
241 On entry to this basic block:
242 t0 == the shifted low-order bits fr
243 t1 == the shifted high-order bits f
244 t2 == the unshifted current source
245
246 We further know that t2 does not co
247
248 .align 4
249 $u_loop:
250 or t0, t1, t0 # E : current
251 subq a2, 1, a2 # E : decremen
252 extql t2, a1, t1 # U : extract
253 addq a0, 8, a0 # E :
254
255 stq_u t0, -8(a0) # U : save the
256 beq a2, $u_eoc # U :
257 ldq_u t2, 8(a1) # U : Latency=
258 addq a1, 8, a1 # E :
259
260 extqh t2, a1, t0 # U : extract
261 cmpbge zero, t2, t8 # E : test new
262 nop
263 beq t8, $u_loop # U :
264
265 /* We've found a zero somewhere in the
266 If it resides in the lower half, we
267 word to write out, and if it reside
268 have one full and one partial word
269
270 On entry to this basic block:
271 t0 == the shifted low-order bits fr
272 t1 == the shifted high-order bits f
273 t2 == the unshifted current source
274 $u_eos:
275 or t0, t1, t0 # E : first (p
276 nop
277 cmpbge zero, t0, t8 # E : is the n
278 bne t8, $u_final # U : (stall)
279
280 stq_u t0, 0(a0) # L : the null
281 addq a0, 8, a0 # E :
282 subq a2, 1, a2 # E :
283 nop
284
285 $u_late_head_exit:
286 extql t2, a1, t0 # U :
287 cmpbge zero, t0, t8 # E :
288 or t8, t10, t6 # E : (stall)
289 cmoveq a2, t6, t8 # E : Latency=
290
291 /* Take care of a final (probably part
292 On entry to this basic block:
293 t0 == assembled source word
294 t8 == cmpbge mask that found the nu
295 $u_final:
296 negq t8, t6 # E : isolate
297 and t6, t8, t12 # E : (stall)
298 and t12, 0x80, t6 # E : avoid de
299 bne t6, 1f # U : (stall)
300
301 ldq_u t1, 0(a0) # L :
302 subq t12, 1, t6 # E :
303 or t6, t12, t8 # E : (stall)
304 zapnot t0, t8, t0 # U : kill sou
305
306 zap t1, t8, t1 # U : kill des
307 or t0, t1, t0 # E : (stall)
308 nop
309 nop
310
311 1: stq_u t0, 0(a0) # L :
312 ret (t9) # L0 : Latency
313
314 /* Got to end-of-count before end of
315 On entry to this basic block:
316 t1 == the shifted high-order bits
317 $u_eoc:
318 and a1, 7, t6 # E : avoid fi
319 sll t10, t6, t6 # U : (stall)
320 and t6, 0xff, t6 # E : (stall)
321 bne t6, 1f # U : (stall)
322
323 ldq_u t2, 8(a1) # L : load fin
324 nop
325 extqh t2, a1, t0 # U : extract
326 or t1, t0, t1 # E : (stall)
327
328 1: cmpbge zero, t1, t8 # E :
329 mov t1, t0 # E :
330
331 $u_eocfin: # end-of-count
332 or t10, t8, t8 # E :
333 br $u_final # L0 : Latency
334
335 /* Unaligned copy entry point. */
336 .align 4
337 $unaligned:
338
339 ldq_u t1, 0(a1) # L : load fir
340 and a0, 7, t4 # E : find des
341 and a1, 7, t5 # E : find src
342 /* Conditionally load the first destin
343 with 0xff indicating that the desti
344 mov zero, t0 # E :
345
346 mov zero, t6 # E :
347 beq t4, 1f # U :
348 ldq_u t0, 0(a0) # L :
349 lda t6, -1 # E :
350
351 mskql t6, a0, t6 # U :
352 nop
353 nop
354 subq a1, t4, a1 # E : sub dest
355
356 /* If source misalignment is larger th
357 extra startup checks to avoid SEGV.
358
359 1: cmplt t4, t5, t12 # E :
360 extql t1, a1, t1 # U : shift sr
361 lda t2, -1 # E : for crea
362 beq t12, $u_head # U : (stall)
363
364 extql t2, a1, t2 # U :
365 cmpbge zero, t1, t8 # E : is there
366 andnot t2, t6, t2 # E : dest mas
367 or t8, t10, t5 # E : test for
368
369 cmpbge zero, t2, t3 # E :
370 cmoveq a2, t5, t8 # E : Latency=
371 nop # E : keep wit
372 andnot t8, t3, t8 # E : (stall)
373
374 beq t8, $u_head # U :
375 /* At this point we've found a zero in
376 the source. We need to isolate the
377 it into the original destination da
378 that we'll need at least one byte o
379 ldq_u t0, 0(a0) # L :
380 negq t8, t6 # E : build bi
381 mskqh t1, t4, t1 # U :
382
383 and t6, t8, t12 # E :
384 subq t12, 1, t6 # E : (stall)
385 or t6, t12, t8 # E : (stall)
386 zapnot t2, t8, t2 # U : prepare
387
388 zapnot t1, t8, t1 # U : to sourc
389 andnot t0, t2, t0 # E : zero pla
390 or t0, t1, t0 # E : and put
391 stq_u t0, 0(a0) # L : (stall)
392
393 ret (t9) # L0 : Latency
394 nop
395 nop
396 nop
397
398 .end __stxncpy
Linux® is a registered trademark of Linus Torvalds in the United States and other countries.
TOMOYO® is a registered trademark of NTT DATA CORPORATION.