1 |
vapier 10/09/29 23:54:55 |
2 |
|
3 |
Modified: 6027_all_alpha-fix-memchr.patch |
4 |
Added: 6028_all_alpha-fix-memchr.patch |
5 |
6029_all_alpha-fix-memchr.patch |
6 |
Log: |
7 |
grab more alpha memchr fixes from upstream |
8 |
|
9 |
Revision Changes Path |
10 |
1.2 src/patchsets/glibc/2.12.1/ports/6027_all_alpha-fix-memchr.patch |
11 |
|
12 |
file : http://sources.gentoo.org/viewvc.cgi/gentoo/src/patchsets/glibc/2.12.1/ports/6027_all_alpha-fix-memchr.patch?rev=1.2&view=markup |
13 |
plain: http://sources.gentoo.org/viewvc.cgi/gentoo/src/patchsets/glibc/2.12.1/ports/6027_all_alpha-fix-memchr.patch?rev=1.2&content-type=text/plain |
14 |
diff : http://sources.gentoo.org/viewvc.cgi/gentoo/src/patchsets/glibc/2.12.1/ports/6027_all_alpha-fix-memchr.patch?r1=1.1&r2=1.2 |
15 |
|
16 |
Index: 6027_all_alpha-fix-memchr.patch |
17 |
=================================================================== |
18 |
RCS file: /var/cvsroot/gentoo/src/patchsets/glibc/2.12.1/ports/6027_all_alpha-fix-memchr.patch,v |
19 |
retrieving revision 1.1 |
20 |
retrieving revision 1.2 |
21 |
diff -u -r1.1 -r1.2 |
22 |
--- 6027_all_alpha-fix-memchr.patch 5 Aug 2010 17:20:40 -0000 1.1 |
23 |
+++ 6027_all_alpha-fix-memchr.patch 29 Sep 2010 23:54:55 -0000 1.2 |
24 |
@@ -1,8 +1,24 @@ |
25 |
-2009-07-27 Aurelien Jarno <aurelien@×××××××.net> |
26 |
+From 200b5faee1cfac10d831e9b278ef294ca3119f53 Mon Sep 17 00:00:00 2001 |
27 |
+From: Richard Henderson <rth@×××××××.net> |
28 |
+Date: Tue, 4 May 2010 09:06:15 -0700 |
29 |
+Subject: [PATCH] alpha: fix memchr to not cause memory faults. |
30 |
|
31 |
+http://www.mail-archive.com/debian-alpha@××××××××××××.org/msg25088.html |
32 |
+ |
33 |
+2010-05-03 Aurelien Jarno <aurelien@×××××××.net> |
34 |
+ |
35 |
* sysdeps/alpha/memchr.S: Use prefetch load. |
36 |
* sysdeps/alpha/alphaev6/memchr.S: Likewise. |
37 |
|
38 |
+Signed-off-by: Matt Turner <mattst88@×××××.com> |
39 |
+--- |
40 |
+ ChangeLog.alpha | 5 +++++ |
41 |
+ sysdeps/alpha/alphaev6/memchr.S | 26 +++++++++++++------------- |
42 |
+ sysdeps/alpha/memchr.S | 22 +++++++++++----------- |
43 |
+ 3 files changed, 29 insertions(+), 24 deletions(-) |
44 |
+ |
45 |
+diff --git a/sysdeps/alpha/alphaev6/memchr.S b/sysdeps/alpha/alphaev6/memchr.S |
46 |
+index 88e91fa..fe77cd8 100644 |
47 |
--- ports/sysdeps/alpha/alphaev6/memchr.S |
48 |
+++ ports/sysdeps/alpha/alphaev6/memchr.S |
49 |
@@ -127,7 +127,7 @@ $first_quad: |
50 |
@@ -65,6 +81,8 @@ |
51 |
nop # E : |
52 |
bne $18, $last_quad # U : |
53 |
|
54 |
+diff --git a/sysdeps/alpha/memchr.S b/sysdeps/alpha/memchr.S |
55 |
+index 5d713d5..87c7fb1 100644 |
56 |
--- ports/sysdeps/alpha/memchr.S |
57 |
+++ ports/sysdeps/alpha/memchr.S |
58 |
@@ -119,7 +119,7 @@ $first_quad: |
59 |
@@ -115,3 +133,5 @@ |
60 |
bne a2, $last_quad # e1 : |
61 |
|
62 |
$not_found: |
63 |
+-- |
64 |
+1.7.3 |
65 |
|
66 |
|
67 |
|
68 |
1.1 src/patchsets/glibc/2.12.1/ports/6028_all_alpha-fix-memchr.patch |
69 |
|
70 |
file : http://sources.gentoo.org/viewvc.cgi/gentoo/src/patchsets/glibc/2.12.1/ports/6028_all_alpha-fix-memchr.patch?rev=1.1&view=markup |
71 |
plain: http://sources.gentoo.org/viewvc.cgi/gentoo/src/patchsets/glibc/2.12.1/ports/6028_all_alpha-fix-memchr.patch?rev=1.1&content-type=text/plain |
72 |
|
73 |
Index: 6028_all_alpha-fix-memchr.patch |
74 |
=================================================================== |
75 |
From 926cf114f7ca2b19116cac005303040648e17e77 Mon Sep 17 00:00:00 2001 |
76 |
From: Richard Henderson <rth@×××××××.net> |
77 |
Date: Wed, 15 Sep 2010 10:41:43 -0700 |
78 |
Subject: [PATCH] alpha: rewrite memchr. |
79 |
|
80 |
[BZ #12019] |
81 |
The new implementation does not read too much data. |
82 |
|
83 |
2010-09-23 Richard Henderson <rth@××××××.com> |
84 |
|
85 |
[BZ #12019] |
86 |
* sysdeps/alpha/alphaev6/memchr.S: Remove. |
87 |
* sysdeps/alpha/memchr.S: Remove. |
88 |
* sysdeps/alpha/memchr.c: New. |
89 |
|
90 |
--- |
91 |
ChangeLog.alpha | 9 ++- |
92 |
sysdeps/alpha/alphaev6/memchr.S | 193 --------------------------------------- |
93 |
sysdeps/alpha/memchr.S | 176 ----------------------------------- |
94 |
sysdeps/alpha/memchr.c | 175 +++++++++++++++++++++++++++++++++++ |
95 |
4 files changed, 183 insertions(+), 370 deletions(-) |
96 |
delete mode 100644 sysdeps/alpha/alphaev6/memchr.S |
97 |
delete mode 100644 sysdeps/alpha/memchr.S |
98 |
create mode 100644 sysdeps/alpha/memchr.c |
99 |
|
100 |
diff --git a/sysdeps/alpha/alphaev6/memchr.S b/sysdeps/alpha/alphaev6/memchr.S |
101 |
deleted file mode 100644 |
102 |
index fe77cd8..0000000 |
103 |
--- ports/sysdeps/alpha/alphaev6/memchr.S |
104 |
+++ /dev/null |
105 |
@@ -1,193 +0,0 @@ |
106 |
-/* Copyright (C) 2000, 2003 Free Software Foundation, Inc. |
107 |
- This file is part of the GNU C Library. |
108 |
- Contributed by David Mosberger (davidm@××××××××××.edu). |
109 |
- EV6 optimized by Rick Gorton <rick.gorton@×××××××××××××××.com>. |
110 |
- |
111 |
- The GNU C Library is free software; you can redistribute it and/or |
112 |
- modify it under the terms of the GNU Lesser General Public |
113 |
- License as published by the Free Software Foundation; either |
114 |
- version 2.1 of the License, or (at your option) any later version. |
115 |
- |
116 |
- The GNU C Library is distributed in the hope that it will be useful, |
117 |
- but WITHOUT ANY WARRANTY; without even the implied warranty of |
118 |
- MERCHANTABILITY or FITNESS FOR A PARTICULAR PURPOSE. See the GNU |
119 |
- Lesser General Public License for more details. |
120 |
- |
121 |
- You should have received a copy of the GNU Lesser General Public |
122 |
- License along with the GNU C Library; if not, write to the Free |
123 |
- Software Foundation, Inc., 59 Temple Place, Suite 330, Boston, MA |
124 |
- 02111-1307 USA. */ |
125 |
- |
126 |
-#include <sysdep.h> |
127 |
- |
128 |
- .arch ev6 |
129 |
- .set noreorder |
130 |
- .set noat |
131 |
- |
132 |
-ENTRY(__memchr) |
133 |
-#ifdef PROF |
134 |
- ldgp gp, 0(pv) |
135 |
- lda AT, _mcount |
136 |
- jsr AT, (AT), _mcount |
137 |
- .prologue 1 |
138 |
-#else |
139 |
- .prologue 0 |
140 |
-#endif |
141 |
- |
142 |
- # Hack -- if someone passes in (size_t)-1, hoping to just |
143 |
- # search til the end of the address space, we will overflow |
144 |
- # below when we find the address of the last byte. Given |
145 |
- # that we will never have a 56-bit address space, cropping |
146 |
- # the length is the easiest way to avoid trouble. |
147 |
- zap $18, 0x80, $5 # U : Bound length |
148 |
- beq $18, $not_found # U : |
149 |
- ldq_u $1, 0($16) # L : load first quadword Latency=3 |
150 |
- and $17, 0xff, $17 # E : L L U U : 00000000000000ch |
151 |
- |
152 |
- insbl $17, 1, $2 # U : 000000000000ch00 |
153 |
- cmpult $18, 9, $4 # E : small (< 1 quad) string? |
154 |
- or $2, $17, $17 # E : 000000000000chch |
155 |
- lda $3, -1($31) # E : U L L U |
156 |
- |
157 |
- sll $17, 16, $2 # U : 00000000chch0000 |
158 |
- addq $16, $5, $5 # E : Max search address |
159 |
- or $2, $17, $17 # E : 00000000chchchch |
160 |
- sll $17, 32, $2 # U : U L L U : chchchch00000000 |
161 |
- |
162 |
- or $2, $17, $17 # E : chchchchchchchch |
163 |
- extql $1, $16, $7 # U : $7 is upper bits |
164 |
- beq $4, $first_quad # U : |
165 |
- ldq_u $6, -1($5) # L : L U U L : eight or less bytes to search Latency=3 |
166 |
- |
167 |
- extqh $6, $16, $6 # U : 2 cycle stall for $6 |
168 |
- mov $16, $0 # E : |
169 |
- nop # E : |
170 |
- or $7, $6, $1 # E : L U L U $1 = quadword starting at $16 |
171 |
- |
172 |
- # Deal with the case where at most 8 bytes remain to be searched |
173 |
- # in $1. E.g.: |
174 |
- # $18 = 6 |
175 |
- # $1 = ????c6c5c4c3c2c1 |
176 |
-$last_quad: |
177 |
- negq $18, $6 # E : |
178 |
- xor $17, $1, $1 # E : |
179 |
- srl $3, $6, $6 # U : $6 = mask of $18 bits set |
180 |
- cmpbge $31, $1, $2 # E : L U L U |
181 |
- |
182 |
- nop |
183 |
- nop |
184 |
- and $2, $6, $2 # E : |
185 |
- beq $2, $not_found # U : U L U L |
186 |
- |
187 |
-$found_it: |
188 |
-#if defined(__alpha_fix__) && defined(__alpha_cix__) |
189 |
- /* |
190 |
- * Since we are guaranteed to have set one of the bits, we don't |
191 |
- * have to worry about coming back with a 0x40 out of cttz... |
192 |
- */ |
193 |
- cttz $2, $3 # U0 : |
194 |
- addq $0, $3, $0 # E : All done |
195 |
- nop # E : |
196 |
- ret # L0 : L U L U |
197 |
-#else |
198 |
- /* |
199 |
- * Slow and clunky. It can probably be improved. |
200 |
- * An exercise left for others. |
201 |
- */ |
202 |
- negq $2, $3 # E : |
203 |
- and $2, $3, $2 # E : |
204 |
- and $2, 0x0f, $1 # E : |
205 |
- addq $0, 4, $3 # E : |
206 |
- |
207 |
- cmoveq $1, $3, $0 # E : Latency 2, extra map cycle |
208 |
- nop # E : keep with cmov |
209 |
- and $2, 0x33, $1 # E : |
210 |
- addq $0, 2, $3 # E : U L U L : 2 cycle stall on $0 |
211 |
- |
212 |
- cmoveq $1, $3, $0 # E : Latency 2, extra map cycle |
213 |
- nop # E : keep with cmov |
214 |
- and $2, 0x55, $1 # E : |
215 |
- addq $0, 1, $3 # E : U L U L : 2 cycle stall on $0 |
216 |
- |
217 |
- cmoveq $1, $3, $0 # E : Latency 2, extra map cycle |
218 |
- nop |
219 |
- nop |
220 |
- ret # L0 : L U L U |
221 |
-#endif |
222 |
- |
223 |
- # Deal with the case where $18 > 8 bytes remain to be |
224 |
- # searched. $16 may not be aligned. |
225 |
- .align 4 |
226 |
-$first_quad: |
227 |
- andnot $16, 0x7, $0 # E : |
228 |
- insqh $3, $16, $2 # U : $2 = 0000ffffffffffff ($16<0:2> ff) |
229 |
- xor $1, $17, $1 # E : |
230 |
- or $1, $2, $1 # E : U L U L $1 = ====ffffffffffff |
231 |
- |
232 |
- cmpbge $31, $1, $2 # E : |
233 |
- bne $2, $found_it # U : |
234 |
- # At least one byte left to process. |
235 |
- ldq $31, 8($0) # L : |
236 |
- subq $5, 1, $18 # E : U L U L |
237 |
- |
238 |
- addq $0, 8, $0 # E : |
239 |
- # Make $18 point to last quad to be accessed (the |
240 |
- # last quad may or may not be partial). |
241 |
- andnot $18, 0x7, $18 # E : |
242 |
- cmpult $0, $18, $2 # E : |
243 |
- beq $2, $final # U : U L U L |
244 |
- |
245 |
- # At least two quads remain to be accessed. |
246 |
- |
247 |
- subq $18, $0, $4 # E : $4 <- nr quads to be processed |
248 |
- and $4, 8, $4 # E : odd number of quads? |
249 |
- bne $4, $odd_quad_count # U : |
250 |
- # At least three quads remain to be accessed |
251 |
- nop # E : L U L U : move prefetched value to correct reg |
252 |
- |
253 |
- .align 4 |
254 |
-$unrolled_loop: |
255 |
- ldq $1, 0($0) # L : load quad |
256 |
- xor $17, $1, $2 # E : |
257 |
- ldq $31, 8($0) # L : prefetch next quad |
258 |
- cmpbge $31, $2, $2 # E : U L U L |
259 |
- |
260 |
- bne $2, $found_it # U : |
261 |
- addq $0, 8, $0 # E : |
262 |
- nop # E : |
263 |
- nop # E : |
264 |
- |
265 |
-$odd_quad_count: |
266 |
- ldq $1, 0($0) # L : load quad |
267 |
- xor $17, $1, $2 # E : |
268 |
- ldq $31, 8($0) # L : prefetch $4 |
269 |
- cmpbge $31, $2, $2 # E : |
270 |
- |
271 |
- addq $0, 8, $6 # E : |
272 |
- bne $2, $found_it # U : |
273 |
- cmpult $6, $18, $6 # E : |
274 |
- addq $0, 8, $0 # E : |
275 |
- |
276 |
- bne $6, $unrolled_loop # U : |
277 |
- nop # E : |
278 |
- nop # E : |
279 |
- nop # E : |
280 |
- |
281 |
-$final: ldq $1, 0($0) # L : load last quad |
282 |
- subq $5, $0, $18 # E : $18 <- number of bytes left to do |
283 |
- nop # E : |
284 |
- bne $18, $last_quad # U : |
285 |
- |
286 |
-$not_found: |
287 |
- mov $31, $0 # E : |
288 |
- nop # E : |
289 |
- nop # E : |
290 |
- ret # L0 : |
291 |
- |
292 |
- END(__memchr) |
293 |
- |
294 |
-weak_alias (__memchr, memchr) |
295 |
-#if !__BOUNDED_POINTERS__ |
296 |
-weak_alias (__memchr, __ubp_memchr) |
297 |
-#endif |
298 |
-libc_hidden_builtin_def (memchr) |
299 |
diff --git a/sysdeps/alpha/memchr.S b/sysdeps/alpha/memchr.S |
300 |
deleted file mode 100644 |
301 |
index 87c7fb1..0000000 |
302 |
--- ports/sysdeps/alpha/memchr.S |
303 |
+++ /dev/null |
304 |
@@ -1,176 +0,0 @@ |
305 |
-/* Copyright (C) 1996, 2000, 2003 Free Software Foundation, Inc. |
306 |
- This file is part of the GNU C Library. |
307 |
- Contributed by David Mosberger (davidm@××××××××××.edu). |
308 |
- |
309 |
- The GNU C Library is free software; you can redistribute it and/or |
310 |
- modify it under the terms of the GNU Lesser General Public |
311 |
- License as published by the Free Software Foundation; either |
312 |
- version 2.1 of the License, or (at your option) any later version. |
313 |
- |
314 |
- The GNU C Library is distributed in the hope that it will be useful, |
315 |
- but WITHOUT ANY WARRANTY; without even the implied warranty of |
316 |
- MERCHANTABILITY or FITNESS FOR A PARTICULAR PURPOSE. See the GNU |
317 |
- Lesser General Public License for more details. |
318 |
- |
319 |
- You should have received a copy of the GNU Lesser General Public |
320 |
- License along with the GNU C Library; if not, write to the Free |
321 |
- Software Foundation, Inc., 59 Temple Place, Suite 330, Boston, MA |
322 |
- 02111-1307 USA. */ |
323 |
- |
324 |
-/* Finds characters in a memory area. Optimized for the Alpha: |
325 |
- |
326 |
- - memory accessed as aligned quadwords only |
327 |
- - uses cmpbge to compare 8 bytes in parallel |
328 |
- - does binary search to find 0 byte in last |
329 |
- quadword (HAKMEM needed 12 instructions to |
330 |
- do this instead of the 9 instructions that |
331 |
- binary search needs). |
332 |
- |
333 |
-For correctness consider that: |
334 |
- |
335 |
- - only minimum number of quadwords may be accessed |
336 |
- - the third argument is an unsigned long |
337 |
-*/ |
338 |
- |
339 |
-#include <sysdep.h> |
340 |
- |
341 |
- .set noreorder |
342 |
- .set noat |
343 |
- |
344 |
-ENTRY(__memchr) |
345 |
-#ifdef PROF |
346 |
- ldgp gp, 0(pv) |
347 |
- lda AT, _mcount |
348 |
- jsr AT, (AT), _mcount |
349 |
- .prologue 1 |
350 |
-#else |
351 |
- .prologue 0 |
352 |
-#endif |
353 |
- |
354 |
- # Hack -- if someone passes in (size_t)-1, hoping to just |
355 |
- # search til the end of the address space, we will overflow |
356 |
- # below when we find the address of the last byte. Given |
357 |
- # that we will never have a 56-bit address space, cropping |
358 |
- # the length is the easiest way to avoid trouble. |
359 |
- zap a2, 0x80, t4 #-e0 : |
360 |
- |
361 |
- beq a2, $not_found # .. e1 : |
362 |
- ldq_u t0, 0(a0) # e1 : load first quadword |
363 |
- insbl a1, 1, t1 # .. e0 : t1 = 000000000000ch00 |
364 |
- and a1, 0xff, a1 #-e0 : a1 = 00000000000000ch |
365 |
- cmpult a2, 9, t3 # .. e1 : |
366 |
- or t1, a1, a1 # e0 : a1 = 000000000000chch |
367 |
- lda t2, -1(zero) # .. e1 : |
368 |
- sll a1, 16, t1 #-e0 : t1 = 00000000chch0000 |
369 |
- addq a0, t4, t4 # .. e1 : |
370 |
- or t1, a1, a1 # e1 : a1 = 00000000chchchch |
371 |
- unop # : |
372 |
- sll a1, 32, t1 #-e0 : t1 = chchchch00000000 |
373 |
- or t1, a1, a1 # e1 : a1 = chchchchchchchch |
374 |
- extql t0, a0, t6 # e0 : |
375 |
- beq t3, $first_quad # .. e1 : |
376 |
- |
377 |
- ldq_u t5, -1(t4) #-e1 : eight or less bytes to search |
378 |
- extqh t5, a0, t5 # .. e0 : |
379 |
- mov a0, v0 # e0 : |
380 |
- or t6, t5, t0 # .. e1 : t0 = quadword starting at a0 |
381 |
- |
382 |
- # Deal with the case where at most 8 bytes remain to be searched |
383 |
- # in t0. E.g.: |
384 |
- # a2 = 6 |
385 |
- # t0 = ????c6c5c4c3c2c1 |
386 |
-$last_quad: |
387 |
- negq a2, t5 #-e0 : |
388 |
- xor a1, t0, t0 # .. e1 : |
389 |
- srl t2, t5, t5 # e0 : t5 = mask of a2 bits set |
390 |
- cmpbge zero, t0, t1 # .. e1 : |
391 |
- and t1, t5, t1 #-e0 : |
392 |
- beq t1, $not_found # .. e1 : |
393 |
- |
394 |
-$found_it: |
395 |
- # Now, determine which byte matched: |
396 |
- negq t1, t2 # e0 : |
397 |
- and t1, t2, t1 # e1 : |
398 |
- |
399 |
- and t1, 0x0f, t0 #-e0 : |
400 |
- addq v0, 4, t2 # .. e1 : |
401 |
- cmoveq t0, t2, v0 # e0 : |
402 |
- |
403 |
- addq v0, 2, t2 # .. e1 : |
404 |
- and t1, 0x33, t0 #-e0 : |
405 |
- cmoveq t0, t2, v0 # .. e1 : |
406 |
- |
407 |
- and t1, 0x55, t0 # e0 : |
408 |
- addq v0, 1, t2 # .. e1 : |
409 |
- cmoveq t0, t2, v0 #-e0 : |
410 |
- |
411 |
-$done: ret # .. e1 : |
412 |
- |
413 |
- # Deal with the case where a2 > 8 bytes remain to be |
414 |
- # searched. a0 may not be aligned. |
415 |
- .align 4 |
416 |
-$first_quad: |
417 |
- andnot a0, 0x7, v0 #-e1 : |
418 |
- insqh t2, a0, t1 # .. e0 : t1 = 0000ffffffffffff (a0<0:2> ff) |
419 |
- xor t0, a1, t0 # e0 : |
420 |
- or t0, t1, t0 # e1 : t0 = ====ffffffffffff |
421 |
- cmpbge zero, t0, t1 #-e0 : |
422 |
- bne t1, $found_it # .. e1 : |
423 |
- |
424 |
- # At least one byte left to process. |
425 |
- |
426 |
- ldq zero, 8(v0) # e0 : prefetch next quad |
427 |
- subq t4, 1, a2 # .. e1 : |
428 |
- addq v0, 8, v0 #-e0 : |
429 |
- |
430 |
- # Make a2 point to last quad to be accessed (the |
431 |
- # last quad may or may not be partial). |
432 |
- |
433 |
- andnot a2, 0x7, a2 # .. e1 : |
434 |
- cmpult v0, a2, t1 # e0 : |
435 |
- beq t1, $final # .. e1 : |
436 |
- |
437 |
- # At least two quads remain to be accessed. |
438 |
- |
439 |
- subq a2, v0, t3 #-e0 : t3 <- nr quads to be processed |
440 |
- and t3, 8, t3 # e1 : odd number of quads? |
441 |
- bne t3, $odd_quad_count # e1 : |
442 |
- |
443 |
- # At least three quads remain to be accessed |
444 |
- |
445 |
- .align 4 |
446 |
-$unrolled_loop: |
447 |
- ldq t0, 0(v0) # e0 : load quad |
448 |
- xor a1, t0, t1 # .. e1 : |
449 |
- ldq zero, 8(v0) # e0 : prefetch next quad |
450 |
- cmpbge zero, t1, t1 # .. e1: |
451 |
- bne t1, $found_it # e0 : |
452 |
- |
453 |
- addq v0, 8, v0 # e1 : |
454 |
-$odd_quad_count: |
455 |
- ldq t0, 0(v0) # e0 : load quad |
456 |
- xor a1, t0, t1 # .. e1 : |
457 |
- ldq zero, 8(v0) # e0 : prefetch next quad |
458 |
- cmpbge zero, t1, t1 # .. e1 : |
459 |
- addq v0, 8, t5 #-e0 : |
460 |
- bne t1, $found_it # .. e1 : |
461 |
- |
462 |
- cmpult t5, a2, t5 # e0 : |
463 |
- addq v0, 8, v0 # .. e1 : |
464 |
- bne t5, $unrolled_loop #-e1 : |
465 |
- |
466 |
-$final: ldq t0, 0(v0) # e0 : load last quad |
467 |
- subq t4, v0, a2 # .. e1 : a2 <- number of bytes left to do |
468 |
- bne a2, $last_quad # e1 : |
469 |
- |
470 |
-$not_found: |
471 |
- mov zero, v0 #-e0 : |
472 |
- ret # .. e1 : |
473 |
- |
474 |
- END(__memchr) |
475 |
- |
476 |
-weak_alias (__memchr, memchr) |
477 |
-#if !__BOUNDED_POINTERS__ |
478 |
-weak_alias (__memchr, __ubp_memchr) |
479 |
-#endif |
480 |
-libc_hidden_builtin_def (memchr) |
481 |
diff --git a/sysdeps/alpha/memchr.c b/sysdeps/alpha/memchr.c |
482 |
new file mode 100644 |
483 |
index 0000000..c52841b |
484 |
--- /dev/null |
485 |
+++ ports/sysdeps/alpha/memchr.c |
486 |
@@ -0,0 +1,175 @@ |
487 |
+/* Copyright (C) 2010 Free Software Foundation, Inc. |
488 |
+ This file is part of the GNU C Library. |
489 |
+ |
490 |
+ The GNU C Library is free software; you can redistribute it and/or |
491 |
+ modify it under the terms of the GNU Lesser General Public |
492 |
+ License as published by the Free Software Foundation; either |
493 |
+ version 2.1 of the License, or (at your option) any later version. |
494 |
+ |
495 |
+ The GNU C Library is distributed in the hope that it will be useful, |
496 |
+ but WITHOUT ANY WARRANTY; without even the implied warranty of |
497 |
+ MERCHANTABILITY or FITNESS FOR A PARTICULAR PURPOSE. See the GNU |
498 |
+ Lesser General Public License for more details. |
499 |
+ |
500 |
+ You should have received a copy of the GNU Lesser General Public |
501 |
+ License along with the GNU C Library; if not, write to the Free |
502 |
+ Software Foundation, Inc., 59 Temple Place, Suite 330, Boston, MA |
503 |
+ 02111-1307 USA. */ |
504 |
+ |
505 |
+#include <string.h> |
506 |
+ |
507 |
+typedef unsigned long word; |
508 |
+ |
509 |
+static inline word |
510 |
+ldq_u(const void *s) |
511 |
+{ |
512 |
+ return *(const word *)((word)s & -8); |
513 |
+} |
514 |
+ |
515 |
+#define unlikely(X) __builtin_expect ((X), 0) |
516 |
+#define prefetch(X) __builtin_prefetch ((void *)(X), 0) |
517 |
+ |
518 |
+#define cmpbeq0(X) __builtin_alpha_cmpbge(0, (X)) |
519 |
+#define find(X, Y) cmpbeq0 ((X) ^ (Y)) |
520 |
+ |
521 |
+/* Search no more than N bytes of S for C. */ |
522 |
+ |
523 |
+void * |
524 |
+__memchr (const void *s, int xc, size_t n) |
525 |
+{ |
526 |
+ const word *s_align; |
527 |
+ word t, current, found, mask, offset; |
528 |
+ |
529 |
+ if (unlikely (n == 0)) |
530 |
+ return 0; |
531 |
+ |
532 |
+ current = ldq_u (s); |
533 |
+ |
534 |
+ /* Replicate low byte of XC into all bytes of C. */ |
535 |
+ t = xc & 0xff; /* 0000000c */ |
536 |
+ t = (t << 8) | t; /* 000000cc */ |
537 |
+ t = (t << 16) | t; /* 0000cccc */ |
538 |
+ const word c = (t << 32) | t; /* cccccccc */ |
539 |
+ |
540 |
+ /* Align the source, and decrement the count by the number |
541 |
+ of bytes searched in the first word. */ |
542 |
+ s_align = (const word *)(s & -8); |
543 |
+ n += (s & 7); |
544 |
+ |
545 |
+ /* Deal with misalignment in the first word for the comparison. */ |
546 |
+ mask = (1ul << (s & 7)) - 1; |
547 |
+ |
548 |
+ /* If the entire string fits within one word, we may need masking |
549 |
+ at both the front and the back of the string. */ |
550 |
+ if (unlikely (n <= 8)) |
551 |
+ { |
552 |
+ mask |= -1ul << n; |
553 |
+ goto last_quad; |
554 |
+ } |
555 |
+ |
556 |
+ found = find (current, c) & ~mask; |
557 |
+ if (unlikely (found)) |
558 |
+ goto found_it; |
559 |
+ |
560 |
+ s_align++; |
561 |
+ n -= 8; |
562 |
+ |
563 |
+ /* If the block is sufficiently large, align to cacheline and prefetch. */ |
564 |
+ if (unlikely (n >= 256)) |
565 |
+ { |
566 |
+ /* Prefetch 3 cache lines beyond the one we're working on. */ |
567 |
+ prefetch (s_align + 8); |
568 |
+ prefetch (s_align + 16); |
569 |
+ prefetch (s_align + 24); |
570 |
+ |
571 |
+ while ((word)s_align & 63) |
572 |
+ { |
573 |
+ current = *s_align; |
574 |
+ found = find (current, c); |
575 |
+ if (found) |
576 |
+ goto found_it; |
577 |
+ s_align++; |
578 |
+ n -= 8; |
579 |
+ } |
580 |
+ |
581 |
+ /* Within each cacheline, advance the load for the next word |
582 |
+ before the test for the previous word is complete. This |
583 |
+ allows us to hide the 3 cycle L1 cache load latency. We |
584 |
+ only perform this advance load within a cacheline to prevent |
585 |
+ reading across page boundary. */ |
586 |
+#define CACHELINE_LOOP \ |
587 |
+ do { \ |
588 |
+ word i, next = s_align[0]; \ |
589 |
+ for (i = 0; i < 7; ++i) \ |
590 |
+ { \ |
591 |
+ current = next; \ |
592 |
+ next = s_align[1]; \ |
593 |
+ found = find (current, c); \ |
594 |
+ if (unlikely (found)) \ |
595 |
+ goto found_it; \ |
596 |
+ s_align++; \ |
597 |
+ } \ |
598 |
+ current = next; \ |
599 |
+ found = find (current, c); \ |
600 |
+ if (unlikely (found)) \ |
601 |
+ goto found_it; \ |
602 |
+ s_align++; \ |
603 |
+ n -= 64; \ |
604 |
+ } while (0) |
605 |
+ |
606 |
+ /* While there's still lots more data to potentially be read, |
607 |
+ continue issuing prefetches for the 4th cacheline out. */ |
608 |
+ while (n >= 256) |
609 |
+ { |
610 |
+ prefetch (s_align + 24); |
611 |
+ CACHELINE_LOOP; |
612 |
+ } |
613 |
+ |
614 |
+ /* Up to 3 cache lines remaining. Continue issuing advanced |
615 |
+ loads, but stop prefetching. */ |
616 |
+ while (n >= 64) |
617 |
+ CACHELINE_LOOP; |
618 |
+ |
619 |
+ /* We may have exhausted the buffer. */ |
620 |
+ if (n == 0) |
621 |
+ return NULL; |
622 |
+ } |
623 |
+ |
624 |
+ /* Quadword aligned loop. */ |
625 |
+ current = *s_align; |
626 |
+ while (n > 8) |
627 |
+ { |
628 |
+ found = find (current, c); |
629 |
+ if (unlikely (found)) |
630 |
+ goto found_it; |
631 |
+ current = *++s_align; |
632 |
+ n -= 8; |
633 |
+ } |
634 |
+ |
635 |
+ /* The last word may need masking at the tail of the compare. */ |
636 |
+ mask = -1ul << n; |
637 |
+ last_quad: |
638 |
+ found = find (current, c) & ~mask; |
639 |
+ if (found == 0) |
640 |
+ return NULL; |
641 |
+ |
642 |
+ found_it: |
643 |
+#ifdef __alpha_cix__ |
644 |
+ offset = __builtin_alpha_cttz (found); |
645 |
+#else |
646 |
+ /* Extract LSB. */ |
647 |
+ found &= -found; |
648 |
+ |
649 |
+ /* Binary search for the LSB. */ |
650 |
+ offset = (found & 0x0f ? 0 : 4); |
651 |
+ offset += (found & 0x33 ? 0 : 2); |
652 |
+ offset += (found & 0x55 ? 0 : 1); |
653 |
+#endif |
654 |
+ |
655 |
+ return (void *)((word)s_align + offset); |
656 |
+} |
657 |
+ |
658 |
+#ifdef weak_alias |
659 |
+weak_alias (__memchr, BP_SYM (memchr)) |
660 |
+#endif |
661 |
+libc_hidden_builtin_def (memchr) |
662 |
-- |
663 |
1.7.3 |
664 |
|
665 |
|
666 |
|
667 |
1.1 src/patchsets/glibc/2.12.1/ports/6029_all_alpha-fix-memchr.patch |
668 |
|
669 |
file : http://sources.gentoo.org/viewvc.cgi/gentoo/src/patchsets/glibc/2.12.1/ports/6029_all_alpha-fix-memchr.patch?rev=1.1&view=markup |
670 |
plain: http://sources.gentoo.org/viewvc.cgi/gentoo/src/patchsets/glibc/2.12.1/ports/6029_all_alpha-fix-memchr.patch?rev=1.1&content-type=text/plain |
671 |
|
672 |
Index: 6029_all_alpha-fix-memchr.patch |
673 |
=================================================================== |
674 |
From b54f998dc380ce327a7faf2c40e569fb2cf39bf0 Mon Sep 17 00:00:00 2001 |
675 |
From: Michael Cree <mcree@×××××××××.nz> |
676 |
Date: Sun, 26 Sep 2010 21:15:51 +1300 |
677 |
Subject: [PATCH] alpha: Fix compile errors in memchr |
678 |
|
679 |
Include missing header file and make some casts explicit. |
680 |
--- |
681 |
sysdeps/alpha/memchr.c | 7 ++++--- |
682 |
1 files changed, 4 insertions(+), 3 deletions(-) |
683 |
|
684 |
diff --git a/sysdeps/alpha/memchr.c b/sysdeps/alpha/memchr.c |
685 |
index c52841b..7e16f8a 100644 |
686 |
--- ports/sysdeps/alpha/memchr.c |
687 |
+++ ports/sysdeps/alpha/memchr.c |
688 |
@@ -17,6 +17,7 @@ |
689 |
02111-1307 USA. */ |
690 |
|
691 |
#include <string.h> |
692 |
+#include <bp-sym.h> |
693 |
|
694 |
typedef unsigned long word; |
695 |
|
696 |
@@ -53,11 +54,11 @@ __memchr (const void *s, int xc, size_t n) |
697 |
|
698 |
/* Align the source, and decrement the count by the number |
699 |
of bytes searched in the first word. */ |
700 |
- s_align = (const word *)(s & -8); |
701 |
- n += (s & 7); |
702 |
+ s_align = (const word *)((word)s & -8); |
703 |
+ n += ((word)s & 7); |
704 |
|
705 |
/* Deal with misalignment in the first word for the comparison. */ |
706 |
- mask = (1ul << (s & 7)) - 1; |
707 |
+ mask = (1ul << ((word)s & 7)) - 1; |
708 |
|
709 |
/* If the entire string fits within one word, we may need masking |
710 |
at both the front and the back of the string. */ |
711 |
-- |
712 |
1.7.3 |