aboutsummaryrefslogtreecommitdiffstats
path: root/arch/alpha/lib
diff options
context:
space:
mode:
authorMichael Cree <mcree@orcon.net.nz>2012-08-18 22:40:58 -0400
committerLinus Torvalds <torvalds@linux-foundation.org>2012-08-19 11:41:18 -0400
commitf2db633d301b4b50f5f93de0e8314cc81e9bc7de (patch)
treeaafa01b17e83221ca92ee58e66d7e6783808a224 /arch/alpha/lib
parentd8d5da129857bfd54b603771fca5409062167392 (diff)
alpha: Use new generic strncpy_from_user() and strnlen_user()
Similar to x86/sparc/powerpc implementations except: 1) we implement an extremely efficient has_zero()/find_zero() sequence with both prep_zero_mask() and create_zero_mask() no-operations. 2) Our output from prep_zero_mask() differs in that only the lowest eight bits are used to represent the zero bytes nevertheless it can be safely ORed with other similar masks from prep_zero_mask() and forms input to create_zero_mask(), the two fundamental properties prep_zero_mask() must satisfy. Tests on EV67 and EV68 CPUs revealed that the generic code is essentially as fast (to within 0.5% of CPU cycles) of the old Alpha specific code for large quadword-aligned strings, despite the 30% extra CPU instructions executed. In contrast, the generic code for unaligned strings is substantially slower (by more than a factor of 3) than the old Alpha specific code. Signed-off-by: Michael Cree <mcree@orcon.net.nz> Acked-by: Matt Turner <mattst88@gmail.com> Signed-off-by: Linus Torvalds <torvalds@linux-foundation.org>
Diffstat (limited to 'arch/alpha/lib')
-rw-r--r--arch/alpha/lib/Makefile2
-rw-r--r--arch/alpha/lib/ev6-strncpy_from_user.S424
-rw-r--r--arch/alpha/lib/ev67-strlen_user.S107
-rw-r--r--arch/alpha/lib/strlen_user.S91
-rw-r--r--arch/alpha/lib/strncpy_from_user.S339
5 files changed, 0 insertions, 963 deletions
diff --git a/arch/alpha/lib/Makefile b/arch/alpha/lib/Makefile
index c0a83ab62b78..59660743237c 100644
--- a/arch/alpha/lib/Makefile
+++ b/arch/alpha/lib/Makefile
@@ -31,8 +31,6 @@ lib-y = __divqu.o __remqu.o __divlu.o __remlu.o \
31 $(ev6-y)memchr.o \ 31 $(ev6-y)memchr.o \
32 $(ev6-y)copy_user.o \ 32 $(ev6-y)copy_user.o \
33 $(ev6-y)clear_user.o \ 33 $(ev6-y)clear_user.o \
34 $(ev6-y)strncpy_from_user.o \
35 $(ev67-y)strlen_user.o \
36 $(ev6-y)csum_ipv6_magic.o \ 34 $(ev6-y)csum_ipv6_magic.o \
37 $(ev6-y)clear_page.o \ 35 $(ev6-y)clear_page.o \
38 $(ev6-y)copy_page.o \ 36 $(ev6-y)copy_page.o \
diff --git a/arch/alpha/lib/ev6-strncpy_from_user.S b/arch/alpha/lib/ev6-strncpy_from_user.S
deleted file mode 100644
index d2e28178cacc..000000000000
--- a/arch/alpha/lib/ev6-strncpy_from_user.S
+++ /dev/null
@@ -1,424 +0,0 @@
1/*
2 * arch/alpha/lib/ev6-strncpy_from_user.S
3 * 21264 version contributed by Rick Gorton <rick.gorton@alpha-processor.com>
4 *
5 * Just like strncpy except in the return value:
6 *
7 * -EFAULT if an exception occurs before the terminator is copied.
8 * N if the buffer filled.
9 *
10 * Otherwise the length of the string is returned.
11 *
12 * Much of the information about 21264 scheduling/coding comes from:
13 * Compiler Writer's Guide for the Alpha 21264
14 * abbreviated as 'CWG' in other comments here
15 * ftp.digital.com/pub/Digital/info/semiconductor/literature/dsc-library.html
16 * Scheduling notation:
17 * E - either cluster
18 * U - upper subcluster; U0 - subcluster U0; U1 - subcluster U1
19 * L - lower subcluster; L0 - subcluster L0; L1 - subcluster L1
20 * A bunch of instructions got moved and temp registers were changed
21 * to aid in scheduling. Control flow was also re-arranged to eliminate
22 * branches, and to provide longer code sequences to enable better scheduling.
23 * A total rewrite (using byte load/stores for start & tail sequences)
24 * is desirable, but very difficult to do without a from-scratch rewrite.
25 * Save that for the future.
26 */
27
28
29#include <asm/errno.h>
30#include <asm/regdef.h>
31
32
33/* Allow an exception for an insn; exit if we get one. */
34#define EX(x,y...) \
35 99: x,##y; \
36 .section __ex_table,"a"; \
37 .long 99b - .; \
38 lda $31, $exception-99b($0); \
39 .previous
40
41
42 .set noat
43 .set noreorder
44 .text
45
46 .globl __strncpy_from_user
47 .ent __strncpy_from_user
48 .frame $30, 0, $26
49 .prologue 0
50
51 .align 4
52__strncpy_from_user:
53 and a0, 7, t3 # E : find dest misalignment
54 beq a2, $zerolength # U :
55
56 /* Are source and destination co-aligned? */
57 mov a0, v0 # E : save the string start
58 xor a0, a1, t4 # E :
59 EX( ldq_u t1, 0(a1) ) # L : Latency=3 load first quadword
60 ldq_u t0, 0(a0) # L : load first (partial) aligned dest quadword
61
62 addq a2, t3, a2 # E : bias count by dest misalignment
63 subq a2, 1, a3 # E :
64 addq zero, 1, t10 # E :
65 and t4, 7, t4 # E : misalignment between the two
66
67 and a3, 7, t6 # E : number of tail bytes
68 sll t10, t6, t10 # E : t10 = bitmask of last count byte
69 bne t4, $unaligned # U :
70 lda t2, -1 # E : build a mask against false zero
71
72 /*
73 * We are co-aligned; take care of a partial first word.
74 * On entry to this basic block:
75 * t0 == the first destination word for masking back in
76 * t1 == the first source word.
77 */
78
79 srl a3, 3, a2 # E : a2 = loop counter = (count - 1)/8
80 addq a1, 8, a1 # E :
81 mskqh t2, a1, t2 # U : detection in the src word
82 nop
83
84 /* Create the 1st output word and detect 0's in the 1st input word. */
85 mskqh t1, a1, t3 # U :
86 mskql t0, a1, t0 # U : assemble the first output word
87 ornot t1, t2, t2 # E :
88 nop
89
90 cmpbge zero, t2, t8 # E : bits set iff null found
91 or t0, t3, t0 # E :
92 beq a2, $a_eoc # U :
93 bne t8, $a_eos # U : 2nd branch in a quad. Bad.
94
95 /* On entry to this basic block:
96 * t0 == a source quad not containing a null.
97 * a0 - current aligned destination address
98 * a1 - current aligned source address
99 * a2 - count of quadwords to move.
100 * NOTE: Loop improvement - unrolling this is going to be
101 * a huge win, since we're going to stall otherwise.
102 * Fix this later. For _really_ large copies, look
103 * at using wh64 on a look-ahead basis. See the code
104 * in clear_user.S and copy_user.S.
105 * Presumably, since (a0) and (a1) do not overlap (by C definition)
106 * Lots of nops here:
107 * - Separate loads from stores
108 * - Keep it to 1 branch/quadpack so the branch predictor
109 * can train.
110 */
111$a_loop:
112 stq_u t0, 0(a0) # L :
113 addq a0, 8, a0 # E :
114 nop
115 subq a2, 1, a2 # E :
116
117 EX( ldq_u t0, 0(a1) ) # L :
118 addq a1, 8, a1 # E :
119 cmpbge zero, t0, t8 # E : Stall 2 cycles on t0
120 beq a2, $a_eoc # U :
121
122 beq t8, $a_loop # U :
123 nop
124 nop
125 nop
126
127 /* Take care of the final (partial) word store. At this point
128 * the end-of-count bit is set in t8 iff it applies.
129 *
130 * On entry to this basic block we have:
131 * t0 == the source word containing the null
132 * t8 == the cmpbge mask that found it.
133 */
134$a_eos:
135 negq t8, t12 # E : find low bit set
136 and t8, t12, t12 # E :
137
138 /* We're doing a partial word store and so need to combine
139 our source and original destination words. */
140 ldq_u t1, 0(a0) # L :
141 subq t12, 1, t6 # E :
142
143 or t12, t6, t8 # E :
144 zapnot t0, t8, t0 # U : clear src bytes > null
145 zap t1, t8, t1 # U : clear dst bytes <= null
146 or t0, t1, t0 # E :
147
148 stq_u t0, 0(a0) # L :
149 br $finish_up # L0 :
150 nop
151 nop
152
153 /* Add the end-of-count bit to the eos detection bitmask. */
154 .align 4
155$a_eoc:
156 or t10, t8, t8
157 br $a_eos
158 nop
159 nop
160
161
162/* The source and destination are not co-aligned. Align the destination
163 and cope. We have to be very careful about not reading too much and
164 causing a SEGV. */
165
166 .align 4
167$u_head:
168 /* We know just enough now to be able to assemble the first
169 full source word. We can still find a zero at the end of it
170 that prevents us from outputting the whole thing.
171
172 On entry to this basic block:
173 t0 == the first dest word, unmasked
174 t1 == the shifted low bits of the first source word
175 t6 == bytemask that is -1 in dest word bytes */
176
177 EX( ldq_u t2, 8(a1) ) # L : load second src word
178 addq a1, 8, a1 # E :
179 mskql t0, a0, t0 # U : mask trailing garbage in dst
180 extqh t2, a1, t4 # U :
181
182 or t1, t4, t1 # E : first aligned src word complete
183 mskqh t1, a0, t1 # U : mask leading garbage in src
184 or t0, t1, t0 # E : first output word complete
185 or t0, t6, t6 # E : mask original data for zero test
186
187 cmpbge zero, t6, t8 # E :
188 beq a2, $u_eocfin # U :
189 bne t8, $u_final # U : bad news - 2nd branch in a quad
190 lda t6, -1 # E : mask out the bits we have
191
192 mskql t6, a1, t6 # U : already seen
193 stq_u t0, 0(a0) # L : store first output word
194 or t6, t2, t2 # E :
195 cmpbge zero, t2, t8 # E : find nulls in second partial
196
197 addq a0, 8, a0 # E :
198 subq a2, 1, a2 # E :
199 bne t8, $u_late_head_exit # U :
200 nop
201
202 /* Finally, we've got all the stupid leading edge cases taken care
203 of and we can set up to enter the main loop. */
204
205 extql t2, a1, t1 # U : position hi-bits of lo word
206 EX( ldq_u t2, 8(a1) ) # L : read next high-order source word
207 addq a1, 8, a1 # E :
208 cmpbge zero, t2, t8 # E :
209
210 beq a2, $u_eoc # U :
211 bne t8, $u_eos # U :
212 nop
213 nop
214
215 /* Unaligned copy main loop. In order to avoid reading too much,
216 the loop is structured to detect zeros in aligned source words.
217 This has, unfortunately, effectively pulled half of a loop
218 iteration out into the head and half into the tail, but it does
219 prevent nastiness from accumulating in the very thing we want
220 to run as fast as possible.
221
222 On entry to this basic block:
223 t1 == the shifted high-order bits from the previous source word
224 t2 == the unshifted current source word
225
226 We further know that t2 does not contain a null terminator. */
227
228 /*
229 * Extra nops here:
230 * separate load quads from store quads
231 * only one branch/quad to permit predictor training
232 */
233
234 .align 4
235$u_loop:
236 extqh t2, a1, t0 # U : extract high bits for current word
237 addq a1, 8, a1 # E :
238 extql t2, a1, t3 # U : extract low bits for next time
239 addq a0, 8, a0 # E :
240
241 or t0, t1, t0 # E : current dst word now complete
242 EX( ldq_u t2, 0(a1) ) # L : load high word for next time
243 subq a2, 1, a2 # E :
244 nop
245
246 stq_u t0, -8(a0) # L : save the current word
247 mov t3, t1 # E :
248 cmpbge zero, t2, t8 # E : test new word for eos
249 beq a2, $u_eoc # U :
250
251 beq t8, $u_loop # U :
252 nop
253 nop
254 nop
255
256 /* We've found a zero somewhere in the source word we just read.
257 If it resides in the lower half, we have one (probably partial)
258 word to write out, and if it resides in the upper half, we
259 have one full and one partial word left to write out.
260
261 On entry to this basic block:
262 t1 == the shifted high-order bits from the previous source word
263 t2 == the unshifted current source word. */
264 .align 4
265$u_eos:
266 extqh t2, a1, t0 # U :
267 or t0, t1, t0 # E : first (partial) source word complete
268 cmpbge zero, t0, t8 # E : is the null in this first bit?
269 nop
270
271 bne t8, $u_final # U :
272 stq_u t0, 0(a0) # L : the null was in the high-order bits
273 addq a0, 8, a0 # E :
274 subq a2, 1, a2 # E :
275
276 .align 4
277$u_late_head_exit:
278 extql t2, a1, t0 # U :
279 cmpbge zero, t0, t8 # E :
280 or t8, t10, t6 # E :
281 cmoveq a2, t6, t8 # E :
282
283 /* Take care of a final (probably partial) result word.
284 On entry to this basic block:
285 t0 == assembled source word
286 t8 == cmpbge mask that found the null. */
287 .align 4
288$u_final:
289 negq t8, t6 # E : isolate low bit set
290 and t6, t8, t12 # E :
291 ldq_u t1, 0(a0) # L :
292 subq t12, 1, t6 # E :
293
294 or t6, t12, t8 # E :
295 zapnot t0, t8, t0 # U : kill source bytes > null
296 zap t1, t8, t1 # U : kill dest bytes <= null
297 or t0, t1, t0 # E :
298
299 stq_u t0, 0(a0) # E :
300 br $finish_up # U :
301 nop
302 nop
303
304 .align 4
305$u_eoc: # end-of-count
306 extqh t2, a1, t0 # U :
307 or t0, t1, t0 # E :
308 cmpbge zero, t0, t8 # E :
309 nop
310
311 .align 4
312$u_eocfin: # end-of-count, final word
313 or t10, t8, t8 # E :
314 br $u_final # U :
315 nop
316 nop
317
318 /* Unaligned copy entry point. */
319 .align 4
320$unaligned:
321
322 srl a3, 3, a2 # U : a2 = loop counter = (count - 1)/8
323 and a0, 7, t4 # E : find dest misalignment
324 and a1, 7, t5 # E : find src misalignment
325 mov zero, t0 # E :
326
327 /* Conditionally load the first destination word and a bytemask
328 with 0xff indicating that the destination byte is sacrosanct. */
329
330 mov zero, t6 # E :
331 beq t4, 1f # U :
332 ldq_u t0, 0(a0) # L :
333 lda t6, -1 # E :
334
335 mskql t6, a0, t6 # E :
336 nop
337 nop
338 nop
339
340 .align 4
3411:
342 subq a1, t4, a1 # E : sub dest misalignment from src addr
343 /* If source misalignment is larger than dest misalignment, we need
344 extra startup checks to avoid SEGV. */
345 cmplt t4, t5, t12 # E :
346 extql t1, a1, t1 # U : shift src into place
347 lda t2, -1 # E : for creating masks later
348
349 beq t12, $u_head # U :
350 mskqh t2, t5, t2 # U : begin src byte validity mask
351 cmpbge zero, t1, t8 # E : is there a zero?
352 nop
353
354 extql t2, a1, t2 # U :
355 or t8, t10, t5 # E : test for end-of-count too
356 cmpbge zero, t2, t3 # E :
357 cmoveq a2, t5, t8 # E : Latency=2, extra map slot
358
359 nop # E : goes with cmov
360 andnot t8, t3, t8 # E :
361 beq t8, $u_head # U :
362 nop
363
364 /* At this point we've found a zero in the first partial word of
365 the source. We need to isolate the valid source data and mask
366 it into the original destination data. (Incidentally, we know
367 that we'll need at least one byte of that original dest word.) */
368
369 ldq_u t0, 0(a0) # L :
370 negq t8, t6 # E : build bitmask of bytes <= zero
371 mskqh t1, t4, t1 # U :
372 and t6, t8, t12 # E :
373
374 subq t12, 1, t6 # E :
375 or t6, t12, t8 # E :
376 zapnot t2, t8, t2 # U : prepare source word; mirror changes
377 zapnot t1, t8, t1 # U : to source validity mask
378
379 andnot t0, t2, t0 # E : zero place for source to reside
380 or t0, t1, t0 # E : and put it there
381 stq_u t0, 0(a0) # L :
382 nop
383
384 .align 4
385$finish_up:
386 zapnot t0, t12, t4 # U : was last byte written null?
387 and t12, 0xf0, t3 # E : binary search for the address of the
388 cmovne t4, 1, t4 # E : Latency=2, extra map slot
389 nop # E : with cmovne
390
391 and t12, 0xcc, t2 # E : last byte written
392 and t12, 0xaa, t1 # E :
393 cmovne t3, 4, t3 # E : Latency=2, extra map slot
394 nop # E : with cmovne
395
396 bic a0, 7, t0
397 cmovne t2, 2, t2 # E : Latency=2, extra map slot
398 nop # E : with cmovne
399 nop
400
401 cmovne t1, 1, t1 # E : Latency=2, extra map slot
402 nop # E : with cmovne
403 addq t0, t3, t0 # E :
404 addq t1, t2, t1 # E :
405
406 addq t0, t1, t0 # E :
407 addq t0, t4, t0 # add one if we filled the buffer
408 subq t0, v0, v0 # find string length
409 ret # L0 :
410
411 .align 4
412$zerolength:
413 nop
414 nop
415 nop
416 clr v0
417
418$exception:
419 nop
420 nop
421 nop
422 ret
423
424 .end __strncpy_from_user
diff --git a/arch/alpha/lib/ev67-strlen_user.S b/arch/alpha/lib/ev67-strlen_user.S
deleted file mode 100644
index 57e0d77b81a6..000000000000
--- a/arch/alpha/lib/ev67-strlen_user.S
+++ /dev/null
@@ -1,107 +0,0 @@
1/*
2 * arch/alpha/lib/ev67-strlen_user.S
3 * 21264 version contributed by Rick Gorton <rick.gorton@api-networks.com>
4 *
5 * Return the length of the string including the NULL terminator
6 * (strlen+1) or zero if an error occurred.
7 *
8 * In places where it is critical to limit the processing time,
9 * and the data is not trusted, strnlen_user() should be used.
10 * It will return a value greater than its second argument if
11 * that limit would be exceeded. This implementation is allowed
12 * to access memory beyond the limit, but will not cross a page
13 * boundary when doing so.
14 *
15 * Much of the information about 21264 scheduling/coding comes from:
16 * Compiler Writer's Guide for the Alpha 21264
17 * abbreviated as 'CWG' in other comments here
18 * ftp.digital.com/pub/Digital/info/semiconductor/literature/dsc-library.html
19 * Scheduling notation:
20 * E - either cluster
21 * U - upper subcluster; U0 - subcluster U0; U1 - subcluster U1
22 * L - lower subcluster; L0 - subcluster L0; L1 - subcluster L1
23 * Try not to change the actual algorithm if possible for consistency.
24 */
25
26#include <asm/regdef.h>
27
28
29/* Allow an exception for an insn; exit if we get one. */
30#define EX(x,y...) \
31 99: x,##y; \
32 .section __ex_table,"a"; \
33 .long 99b - .; \
34 lda v0, $exception-99b(zero); \
35 .previous
36
37
38 .set noreorder
39 .set noat
40 .text
41
42 .globl __strlen_user
43 .ent __strlen_user
44 .frame sp, 0, ra
45
46 .align 4
47__strlen_user:
48 ldah a1, 32767(zero) # do not use plain strlen_user() for strings
49 # that might be almost 2 GB long; you should
50 # be using strnlen_user() instead
51 nop
52 nop
53 nop
54
55 .globl __strnlen_user
56
57 .align 4
58__strnlen_user:
59 .prologue 0
60 EX( ldq_u t0, 0(a0) ) # L : load first quadword (a0 may be misaligned)
61 lda t1, -1(zero) # E :
62
63 insqh t1, a0, t1 # U :
64 andnot a0, 7, v0 # E :
65 or t1, t0, t0 # E :
66 subq a0, 1, a0 # E : get our +1 for the return
67
68 cmpbge zero, t0, t1 # E : t1 <- bitmask: bit i == 1 <==> i-th byte == 0
69 subq a1, 7, t2 # E :
70 subq a0, v0, t0 # E :
71 bne t1, $found # U :
72
73 addq t2, t0, t2 # E :
74 addq a1, 1, a1 # E :
75 nop # E :
76 nop # E :
77
78 .align 4
79$loop: ble t2, $limit # U :
80 EX( ldq t0, 8(v0) ) # L :
81 nop # E :
82 nop # E :
83
84 cmpbge zero, t0, t1 # E :
85 subq t2, 8, t2 # E :
86 addq v0, 8, v0 # E : addr += 8
87 beq t1, $loop # U :
88
89$found: cttz t1, t2 # U0 :
90 addq v0, t2, v0 # E :
91 subq v0, a0, v0 # E :
92 ret # L0 :
93
94$exception:
95 nop
96 nop
97 nop
98 ret
99
100 .align 4 # currently redundant
101$limit:
102 nop
103 nop
104 subq a1, t2, v0
105 ret
106
107 .end __strlen_user
diff --git a/arch/alpha/lib/strlen_user.S b/arch/alpha/lib/strlen_user.S
deleted file mode 100644
index 508a18e96479..000000000000
--- a/arch/alpha/lib/strlen_user.S
+++ /dev/null
@@ -1,91 +0,0 @@
1/*
2 * arch/alpha/lib/strlen_user.S
3 *
4 * Return the length of the string including the NUL terminator
5 * (strlen+1) or zero if an error occurred.
6 *
7 * In places where it is critical to limit the processing time,
8 * and the data is not trusted, strnlen_user() should be used.
9 * It will return a value greater than its second argument if
10 * that limit would be exceeded. This implementation is allowed
11 * to access memory beyond the limit, but will not cross a page
12 * boundary when doing so.
13 */
14
15#include <asm/regdef.h>
16
17
18/* Allow an exception for an insn; exit if we get one. */
19#define EX(x,y...) \
20 99: x,##y; \
21 .section __ex_table,"a"; \
22 .long 99b - .; \
23 lda v0, $exception-99b(zero); \
24 .previous
25
26
27 .set noreorder
28 .set noat
29 .text
30
31 .globl __strlen_user
32 .ent __strlen_user
33 .frame sp, 0, ra
34
35 .align 3
36__strlen_user:
37 ldah a1, 32767(zero) # do not use plain strlen_user() for strings
38 # that might be almost 2 GB long; you should
39 # be using strnlen_user() instead
40
41 .globl __strnlen_user
42
43 .align 3
44__strnlen_user:
45 .prologue 0
46
47 EX( ldq_u t0, 0(a0) ) # load first quadword (a0 may be misaligned)
48 lda t1, -1(zero)
49 insqh t1, a0, t1
50 andnot a0, 7, v0
51 or t1, t0, t0
52 subq a0, 1, a0 # get our +1 for the return
53 cmpbge zero, t0, t1 # t1 <- bitmask: bit i == 1 <==> i-th byte == 0
54 subq a1, 7, t2
55 subq a0, v0, t0
56 bne t1, $found
57
58 addq t2, t0, t2
59 addq a1, 1, a1
60
61 .align 3
62$loop: ble t2, $limit
63 EX( ldq t0, 8(v0) )
64 subq t2, 8, t2
65 addq v0, 8, v0 # addr += 8
66 cmpbge zero, t0, t1
67 beq t1, $loop
68
69$found: negq t1, t2 # clear all but least set bit
70 and t1, t2, t1
71
72 and t1, 0xf0, t2 # binary search for that set bit
73 and t1, 0xcc, t3
74 and t1, 0xaa, t4
75 cmovne t2, 4, t2
76 cmovne t3, 2, t3
77 cmovne t4, 1, t4
78 addq t2, t3, t2
79 addq v0, t4, v0
80 addq v0, t2, v0
81 nop # dual issue next two on ev4 and ev5
82 subq v0, a0, v0
83$exception:
84 ret
85
86 .align 3 # currently redundant
87$limit:
88 subq a1, t2, v0
89 ret
90
91 .end __strlen_user
diff --git a/arch/alpha/lib/strncpy_from_user.S b/arch/alpha/lib/strncpy_from_user.S
deleted file mode 100644
index 73ee21160ff7..000000000000
--- a/arch/alpha/lib/strncpy_from_user.S
+++ /dev/null
@@ -1,339 +0,0 @@
1/*
2 * arch/alpha/lib/strncpy_from_user.S
3 * Contributed by Richard Henderson (rth@tamu.edu)
4 *
5 * Just like strncpy except in the return value:
6 *
7 * -EFAULT if an exception occurs before the terminator is copied.
8 * N if the buffer filled.
9 *
10 * Otherwise the length of the string is returned.
11 */
12
13
14#include <asm/errno.h>
15#include <asm/regdef.h>
16
17
18/* Allow an exception for an insn; exit if we get one. */
19#define EX(x,y...) \
20 99: x,##y; \
21 .section __ex_table,"a"; \
22 .long 99b - .; \
23 lda $31, $exception-99b($0); \
24 .previous
25
26
27 .set noat
28 .set noreorder
29 .text
30
31 .globl __strncpy_from_user
32 .ent __strncpy_from_user
33 .frame $30, 0, $26
34 .prologue 0
35
36 .align 3
37$aligned:
38 /* On entry to this basic block:
39 t0 == the first destination word for masking back in
40 t1 == the first source word. */
41
42 /* Create the 1st output word and detect 0's in the 1st input word. */
43 lda t2, -1 # e1 : build a mask against false zero
44 mskqh t2, a1, t2 # e0 : detection in the src word
45 mskqh t1, a1, t3 # e0 :
46 ornot t1, t2, t2 # .. e1 :
47 mskql t0, a1, t0 # e0 : assemble the first output word
48 cmpbge zero, t2, t8 # .. e1 : bits set iff null found
49 or t0, t3, t0 # e0 :
50 beq a2, $a_eoc # .. e1 :
51 bne t8, $a_eos # .. e1 :
52
53 /* On entry to this basic block:
54 t0 == a source word not containing a null. */
55
56$a_loop:
57 stq_u t0, 0(a0) # e0 :
58 addq a0, 8, a0 # .. e1 :
59 EX( ldq_u t0, 0(a1) ) # e0 :
60 addq a1, 8, a1 # .. e1 :
61 subq a2, 1, a2 # e0 :
62 cmpbge zero, t0, t8 # .. e1 (stall)
63 beq a2, $a_eoc # e1 :
64 beq t8, $a_loop # e1 :
65
66 /* Take care of the final (partial) word store. At this point
67 the end-of-count bit is set in t8 iff it applies.
68
69 On entry to this basic block we have:
70 t0 == the source word containing the null
71 t8 == the cmpbge mask that found it. */
72
73$a_eos:
74 negq t8, t12 # e0 : find low bit set
75 and t8, t12, t12 # e1 (stall)
76
77 /* For the sake of the cache, don't read a destination word
78 if we're not going to need it. */
79 and t12, 0x80, t6 # e0 :
80 bne t6, 1f # .. e1 (zdb)
81
82 /* We're doing a partial word store and so need to combine
83 our source and original destination words. */
84 ldq_u t1, 0(a0) # e0 :
85 subq t12, 1, t6 # .. e1 :
86 or t12, t6, t8 # e0 :
87 unop #
88 zapnot t0, t8, t0 # e0 : clear src bytes > null
89 zap t1, t8, t1 # .. e1 : clear dst bytes <= null
90 or t0, t1, t0 # e1 :
91
921: stq_u t0, 0(a0)
93 br $finish_up
94
95 /* Add the end-of-count bit to the eos detection bitmask. */
96$a_eoc:
97 or t10, t8, t8
98 br $a_eos
99
100 /*** The Function Entry Point ***/
101 .align 3
102__strncpy_from_user:
103 mov a0, v0 # save the string start
104 beq a2, $zerolength
105
106 /* Are source and destination co-aligned? */
107 xor a0, a1, t1 # e0 :
108 and a0, 7, t0 # .. e1 : find dest misalignment
109 and t1, 7, t1 # e0 :
110 addq a2, t0, a2 # .. e1 : bias count by dest misalignment
111 subq a2, 1, a2 # e0 :
112 and a2, 7, t2 # e1 :
113 srl a2, 3, a2 # e0 : a2 = loop counter = (count - 1)/8
114 addq zero, 1, t10 # .. e1 :
115 sll t10, t2, t10 # e0 : t10 = bitmask of last count byte
116 bne t1, $unaligned # .. e1 :
117
118 /* We are co-aligned; take care of a partial first word. */
119
120 EX( ldq_u t1, 0(a1) ) # e0 : load first src word
121 addq a1, 8, a1 # .. e1 :
122
123 beq t0, $aligned # avoid loading dest word if not needed
124 ldq_u t0, 0(a0) # e0 :
125 br $aligned # .. e1 :
126
127
128/* The source and destination are not co-aligned. Align the destination
129 and cope. We have to be very careful about not reading too much and
130 causing a SEGV. */
131
132 .align 3
133$u_head:
134 /* We know just enough now to be able to assemble the first
135 full source word. We can still find a zero at the end of it
136 that prevents us from outputting the whole thing.
137
138 On entry to this basic block:
139 t0 == the first dest word, unmasked
140 t1 == the shifted low bits of the first source word
141 t6 == bytemask that is -1 in dest word bytes */
142
143 EX( ldq_u t2, 8(a1) ) # e0 : load second src word
144 addq a1, 8, a1 # .. e1 :
145 mskql t0, a0, t0 # e0 : mask trailing garbage in dst
146 extqh t2, a1, t4 # e0 :
147 or t1, t4, t1 # e1 : first aligned src word complete
148 mskqh t1, a0, t1 # e0 : mask leading garbage in src
149 or t0, t1, t0 # e0 : first output word complete
150 or t0, t6, t6 # e1 : mask original data for zero test
151 cmpbge zero, t6, t8 # e0 :
152 beq a2, $u_eocfin # .. e1 :
153 bne t8, $u_final # e1 :
154
155 lda t6, -1 # e1 : mask out the bits we have
156 mskql t6, a1, t6 # e0 : already seen
157 stq_u t0, 0(a0) # e0 : store first output word
158 or t6, t2, t2 # .. e1 :
159 cmpbge zero, t2, t8 # e0 : find nulls in second partial
160 addq a0, 8, a0 # .. e1 :
161 subq a2, 1, a2 # e0 :
162 bne t8, $u_late_head_exit # .. e1 :
163
164 /* Finally, we've got all the stupid leading edge cases taken care
165 of and we can set up to enter the main loop. */
166
167 extql t2, a1, t1 # e0 : position hi-bits of lo word
168 EX( ldq_u t2, 8(a1) ) # .. e1 : read next high-order source word
169 addq a1, 8, a1 # e0 :
170 cmpbge zero, t2, t8 # e1 (stall)
171 beq a2, $u_eoc # e1 :
172 bne t8, $u_eos # e1 :
173
174 /* Unaligned copy main loop. In order to avoid reading too much,
175 the loop is structured to detect zeros in aligned source words.
176 This has, unfortunately, effectively pulled half of a loop
177 iteration out into the head and half into the tail, but it does
178 prevent nastiness from accumulating in the very thing we want
179 to run as fast as possible.
180
181 On entry to this basic block:
182 t1 == the shifted high-order bits from the previous source word
183 t2 == the unshifted current source word
184
185 We further know that t2 does not contain a null terminator. */
186
187 .align 3
188$u_loop:
189 extqh t2, a1, t0 # e0 : extract high bits for current word
190 addq a1, 8, a1 # .. e1 :
191 extql t2, a1, t3 # e0 : extract low bits for next time
192 addq a0, 8, a0 # .. e1 :
193 or t0, t1, t0 # e0 : current dst word now complete
194 EX( ldq_u t2, 0(a1) ) # .. e1 : load high word for next time
195 stq_u t0, -8(a0) # e0 : save the current word
196 mov t3, t1 # .. e1 :
197 subq a2, 1, a2 # e0 :
198 cmpbge zero, t2, t8 # .. e1 : test new word for eos
199 beq a2, $u_eoc # e1 :
200 beq t8, $u_loop # e1 :
201
202 /* We've found a zero somewhere in the source word we just read.
203 If it resides in the lower half, we have one (probably partial)
204 word to write out, and if it resides in the upper half, we
205 have one full and one partial word left to write out.
206
207 On entry to this basic block:
208 t1 == the shifted high-order bits from the previous source word
209 t2 == the unshifted current source word. */
210$u_eos:
211 extqh t2, a1, t0 # e0 :
212 or t0, t1, t0 # e1 : first (partial) source word complete
213
214 cmpbge zero, t0, t8 # e0 : is the null in this first bit?
215 bne t8, $u_final # .. e1 (zdb)
216
217 stq_u t0, 0(a0) # e0 : the null was in the high-order bits
218 addq a0, 8, a0 # .. e1 :
219 subq a2, 1, a2 # e1 :
220
221$u_late_head_exit:
222 extql t2, a1, t0 # .. e0 :
223 cmpbge zero, t0, t8 # e0 :
224 or t8, t10, t6 # e1 :
225 cmoveq a2, t6, t8 # e0 :
226 nop # .. e1 :
227
228 /* Take care of a final (probably partial) result word.
229 On entry to this basic block:
230 t0 == assembled source word
231 t8 == cmpbge mask that found the null. */
232$u_final:
233 negq t8, t6 # e0 : isolate low bit set
234 and t6, t8, t12 # e1 :
235
236 and t12, 0x80, t6 # e0 : avoid dest word load if we can
237 bne t6, 1f # .. e1 (zdb)
238
239 ldq_u t1, 0(a0) # e0 :
240 subq t12, 1, t6 # .. e1 :
241 or t6, t12, t8 # e0 :
242 zapnot t0, t8, t0 # .. e1 : kill source bytes > null
243 zap t1, t8, t1 # e0 : kill dest bytes <= null
244 or t0, t1, t0 # e1 :
245
2461: stq_u t0, 0(a0) # e0 :
247 br $finish_up
248
249$u_eoc: # end-of-count
250 extqh t2, a1, t0
251 or t0, t1, t0
252 cmpbge zero, t0, t8
253
254$u_eocfin: # end-of-count, final word
255 or t10, t8, t8
256 br $u_final
257
258 /* Unaligned copy entry point. */
259 .align 3
260$unaligned:
261
262 EX( ldq_u t1, 0(a1) ) # e0 : load first source word
263
264 and a0, 7, t4 # .. e1 : find dest misalignment
265 and a1, 7, t5 # e0 : find src misalignment
266
267 /* Conditionally load the first destination word and a bytemask
268 with 0xff indicating that the destination byte is sacrosanct. */
269
270 mov zero, t0 # .. e1 :
271 mov zero, t6 # e0 :
272 beq t4, 1f # .. e1 :
273 ldq_u t0, 0(a0) # e0 :
274 lda t6, -1 # .. e1 :
275 mskql t6, a0, t6 # e0 :
2761:
277 subq a1, t4, a1 # .. e1 : sub dest misalignment from src addr
278
279 /* If source misalignment is larger than dest misalignment, we need
280 extra startup checks to avoid SEGV. */
281
282 cmplt t4, t5, t12 # e1 :
283 extql t1, a1, t1 # .. e0 : shift src into place
284 lda t2, -1 # e0 : for creating masks later
285 beq t12, $u_head # e1 :
286
287 mskqh t2, t5, t2 # e0 : begin src byte validity mask
288 cmpbge zero, t1, t8 # .. e1 : is there a zero?
289 extql t2, a1, t2 # e0 :
290 or t8, t10, t5 # .. e1 : test for end-of-count too
291 cmpbge zero, t2, t3 # e0 :
292 cmoveq a2, t5, t8 # .. e1 :
293 andnot t8, t3, t8 # e0 :
294 beq t8, $u_head # .. e1 (zdb)
295
296 /* At this point we've found a zero in the first partial word of
297 the source. We need to isolate the valid source data and mask
298 it into the original destination data. (Incidentally, we know
299 that we'll need at least one byte of that original dest word.) */
300
301 ldq_u t0, 0(a0) # e0 :
302 negq t8, t6 # .. e1 : build bitmask of bytes <= zero
303 mskqh t1, t4, t1 # e0 :
304 and t6, t8, t12 # .. e1 :
305 subq t12, 1, t6 # e0 :
306 or t6, t12, t8 # e1 :
307
308 zapnot t2, t8, t2 # e0 : prepare source word; mirror changes
309 zapnot t1, t8, t1 # .. e1 : to source validity mask
310
311 andnot t0, t2, t0 # e0 : zero place for source to reside
312 or t0, t1, t0 # e1 : and put it there
313 stq_u t0, 0(a0) # e0 :
314
315$finish_up:
316 zapnot t0, t12, t4 # was last byte written null?
317 cmovne t4, 1, t4
318
319 and t12, 0xf0, t3 # binary search for the address of the
320 and t12, 0xcc, t2 # last byte written
321 and t12, 0xaa, t1
322 bic a0, 7, t0
323 cmovne t3, 4, t3
324 cmovne t2, 2, t2
325 cmovne t1, 1, t1
326 addq t0, t3, t0
327 addq t1, t2, t1
328 addq t0, t1, t0
329 addq t0, t4, t0 # add one if we filled the buffer
330
331 subq t0, v0, v0 # find string length
332 ret
333
334$zerolength:
335 clr v0
336$exception:
337 ret
338
339 .end __strncpy_from_user
generated by cgit v1.2.2 (git 2.25.0) at 2024-09-20 22:40:58 -0400