diff options
author | Andrew Mahone <andrew.mahone@gmail.com> | 2010-02-01 01:36:46 +0000 |
---|---|---|
committer | Andrew Mahone <andrew.mahone@gmail.com> | 2010-02-01 01:36:46 +0000 |
commit | bff5a35c3c51ebe1fe72ee20147b16ede847971d (patch) | |
tree | 549fc29d08c298462b91b9d1cfc054d704faa4bd /firmware | |
parent | 40a3e806974ff482006d4ee99b718a52d1e85224 (diff) | |
download | rockbox-bff5a35c3c51ebe1fe72ee20147b16ede847971d.tar.gz rockbox-bff5a35c3c51ebe1fe72ee20147b16ede847971d.zip |
FS#10943, optimized division and clz routines to replace libgcc routines for ARM. Replaces libgcc support functions for unsigned and signed 32-bit division on ARMv4 and up, and leading-zero count on ARMv4.
git-svn-id: svn://svn.rockbox.org/rockbox/trunk@24432 a1c6a512-1295-4272-9138-f99709370657
Diffstat (limited to 'firmware')
-rw-r--r-- | firmware/SOURCES | 1 | ||||
-rw-r--r-- | firmware/target/arm/support-arm.S | 699 |
2 files changed, 700 insertions, 0 deletions
diff --git a/firmware/SOURCES b/firmware/SOURCES index 56ab680417..e60bf63419 100644 --- a/firmware/SOURCES +++ b/firmware/SOURCES | |||
@@ -375,6 +375,7 @@ target/coldfire/i2c-coldfire.c | |||
375 | 375 | ||
376 | #elif defined(CPU_PP) || defined(CPU_ARM) | 376 | #elif defined(CPU_PP) || defined(CPU_ARM) |
377 | /* CPU_PP => CPU_ARM, CPU_ARM !=> CPU_PP */ | 377 | /* CPU_PP => CPU_ARM, CPU_ARM !=> CPU_PP */ |
378 | target/arm/support-arm.S | ||
378 | target/arm/memcpy-arm.S | 379 | target/arm/memcpy-arm.S |
379 | target/arm/memmove-arm.S | 380 | target/arm/memmove-arm.S |
380 | common/strlen.c | 381 | common/strlen.c |
diff --git a/firmware/target/arm/support-arm.S b/firmware/target/arm/support-arm.S new file mode 100644 index 0000000000..8703dd5b0a --- /dev/null +++ b/firmware/target/arm/support-arm.S | |||
@@ -0,0 +1,699 @@ | |||
1 | /*************************************************************************** | ||
2 | * __________ __ ___. | ||
3 | * Open \______ \ ____ ____ | | _\_ |__ _______ ___ | ||
4 | * Source | _// _ \_/ ___\| |/ /| __ \ / _ \ \/ / | ||
5 | * Jukebox | | ( <_> ) \___| < | \_\ ( <_> > < < | ||
6 | * Firmware |____|_ /\____/ \___ >__|_ \|___ /\____/__/\_ \ | ||
7 | * \/ \/ \/ \/ \/ | ||
8 | * $Id$ | ||
9 | * | ||
10 | * Copyright (C) 2008 by Jens Arnold | ||
11 | * Copyright (C) 2009 by Andrew Mahone | ||
12 | * | ||
13 | * Optimised replacements for libgcc functions | ||
14 | * | ||
15 | * Based on: libgcc routines for ARM cpu, additional algorithms from ARM System | ||
16 | * Developer's Guide | ||
17 | * Division routines, written by Richard Earnshaw, (rearnsha@armltd.co.uk) | ||
18 | * Copyright 1995, 1996, 1998, 1999, 2000, 2003, 2004, 2005 | ||
19 | * Free Software Foundation, Inc. | ||
20 | * | ||
21 | * This program is free software; you can redistribute it and/or | ||
22 | * modify it under the terms of the GNU General Public License | ||
23 | * as published by the Free Software Foundation; either version 2 | ||
24 | * of the License, or (at your option) any later version. | ||
25 | * | ||
26 | * This software is distributed on an "AS IS" basis, WITHOUT WARRANTY OF ANY | ||
27 | * KIND, either express or implied. | ||
28 | * | ||
29 | ****************************************************************************/ | ||
30 | |||
31 | #include <config.h> | ||
32 | |||
33 | .macro ARM_SDIV32_PRE numerator, divisor, sign | ||
34 | /* sign[31] = divisor sign */ | ||
35 | ands \sign, \divisor, #1<<31 | ||
36 | rsbeq \divisor, \divisor, #0 | ||
37 | /* sign[31] = result sign, sign[0:30], C = numerator sign */ | ||
38 | eors \sign, \sign, \numerator, asr #32 | ||
39 | rsbcs \numerator, \numerator, #0 | ||
40 | .endm | ||
41 | |||
42 | .macro ARM_SDIV32_POST quotient, remainder, sign | ||
43 | movs \sign, \sign, lsl #1 | ||
44 | .ifnc "", "\quotient" | ||
45 | rsbcs \quotient, \quotient, #0 | ||
46 | .endif | ||
47 | .ifnc "", "\remainder" | ||
48 | rsbmi \remainder, \remainder, #0 | ||
49 | .endif | ||
50 | .endm | ||
51 | |||
52 | #if ARM_ARCH < 5 | ||
53 | .macro ARMV4_UDIV32_BODY numerator, divisor, quotient, remainder, tmp, bits, div0label, return | ||
54 | .ifnc "", "\div0label" | ||
55 | rsbs \divisor, \divisor, #0 | ||
56 | beq \div0label | ||
57 | .else | ||
58 | rsb \divisor, \divisor, #0 | ||
59 | .endif | ||
60 | /* This SWAR divider requires a numerator less than 1<<31, because it must | ||
61 | be able to shift the remainder left at each step without shifting out | ||
62 | topmost bit. Since a shift might be needed for the aligned remainder to | ||
63 | exceed the divisor, the topmost bit must be unset at the start to avoid | ||
64 | this overflow case. The original numerator is saved so that the result | ||
65 | can be corrected after the reduced division completes. */ | ||
66 | cmn \numerator, \divisor | ||
67 | .ifc "", "\quotient" | ||
68 | .ifc "\numerator", "\remainder" | ||
69 | .if \return | ||
70 | bxcc lr | ||
71 | .else | ||
72 | b 99f | ||
73 | .endif | ||
74 | .else | ||
75 | bcc 20f | ||
76 | .endif | ||
77 | .else | ||
78 | bcc 20f | ||
79 | .endif | ||
80 | movs \tmp, \numerator | ||
81 | movmi \numerator, \numerator, lsr #1 | ||
82 | mov \bits, #30 | ||
83 | .set shift, 16 | ||
84 | .rept 5 | ||
85 | cmn \divisor, \numerator, lsr #shift | ||
86 | subcs \bits, \bits, #shift | ||
87 | movcs \divisor, \divisor, lsl #shift | ||
88 | .set shift, shift >> 1 | ||
89 | .endr | ||
90 | adds \numerator, \numerator, \divisor | ||
91 | subcc \numerator, \numerator, \divisor | ||
92 | add pc, pc, \bits, lsl #3 | ||
93 | nop | ||
94 | .rept 30 | ||
95 | adcs \numerator, \divisor, \numerator, lsl #1 | ||
96 | subcc \numerator, \numerator, \divisor | ||
97 | .endr | ||
98 | adc \numerator, \numerator, \numerator | ||
99 | movs \tmp, \tmp, asr #1 | ||
100 | rsb \bits, \bits, #31 | ||
101 | bmi 10f | ||
102 | .ifc "", "\quotient" | ||
103 | mov \remainder, \numerator, lsr \bits | ||
104 | .else | ||
105 | .ifc "", "\remainder" | ||
106 | mov \divisor, \numerator, lsr \bits | ||
107 | eor \quotient, \numerator, \divisor, lsl \bits | ||
108 | .else | ||
109 | mov \remainder, \numerator, lsr \bits | ||
110 | eor \quotient, \numerator, \remainder, lsl \bits | ||
111 | .endif | ||
112 | .endif | ||
113 | .ifne \return | ||
114 | bx lr | ||
115 | .else | ||
116 | b 99f | ||
117 | .endif | ||
118 | 10: | ||
119 | mov \tmp, \numerator, lsr \bits | ||
120 | eor \numerator, \numerator, \tmp, lsl \bits | ||
121 | sub \bits, \bits, #1 | ||
122 | adc \tmp, \tmp, \tmp | ||
123 | adds \tmp, \tmp, \divisor, asr \bits | ||
124 | .ifnc "", "\quotient" | ||
125 | adc \quotient, \numerator, \numerator | ||
126 | .endif | ||
127 | .ifnc "", "\remainder" | ||
128 | subcc \remainder, \tmp, \divisor, asr \bits | ||
129 | movcs \remainder, \tmp | ||
130 | .endif | ||
131 | .ifne \return | ||
132 | bx lr | ||
133 | .else | ||
134 | b 99f | ||
135 | .endif | ||
136 | 20: | ||
137 | .ifnc "", "\remainder" | ||
138 | .ifnc "\remainder", "\numerator" | ||
139 | mov \remainder, \numerator | ||
140 | .endif | ||
141 | .endif | ||
142 | .ifnc "", "\quotient" | ||
143 | mov \quotient, #0 | ||
144 | .endif | ||
145 | .ifne \return | ||
146 | bx lr | ||
147 | .else | ||
148 | 99: | ||
149 | .endif | ||
150 | .endm | ||
151 | |||
152 | .macro ARMV4_SDIV32_BODY numerator, divisor, quotient, remainder, bits, sign, div0label, return | ||
153 | /* When this is wrapped for signed division, the wrapper code will handle | ||
154 | inverting the divisor, and also the zero divisor test. */ | ||
155 | ARM_SDIV32_PRE \numerator, \divisor, \sign | ||
156 | .ifnc "", "\div0label" | ||
157 | tst \divisor, \divisor | ||
158 | beq \div0label | ||
159 | .endif | ||
160 | /* This SWAR divider requires a numerator less than 1<<31, because it must | ||
161 | be able to shift the remainder left at each step without shifting out | ||
162 | topmost bit. With signed inputs, whose absolute value may not exceed | ||
163 | 1<<31,this may be accomplished simply by subtracting the divisor before | ||
164 | beginning division, and adding 1 to the quotient. */ | ||
165 | adds \numerator, \numerator, \divisor | ||
166 | bcc 20f | ||
167 | mov \bits, #30 | ||
168 | .set shift, 16 | ||
169 | .rept 5 | ||
170 | cmn \divisor, \numerator, lsr #shift | ||
171 | subcs \bits, \bits, #shift | ||
172 | movcs \divisor, \divisor, lsl #shift | ||
173 | .set shift, shift >> 1 | ||
174 | .endr | ||
175 | adds \numerator, \numerator, \divisor | ||
176 | subcc \numerator, \numerator, \divisor | ||
177 | add pc, pc, \bits, lsl #3 | ||
178 | nop | ||
179 | .rept 30 | ||
180 | adcs \numerator, \divisor, \numerator, lsl #1 | ||
181 | subcc \numerator, \numerator, \divisor | ||
182 | .endr | ||
183 | rsb \bits, \bits, #31 | ||
184 | adc \numerator, \numerator, \numerator | ||
185 | .ifc "", "\quotient" | ||
186 | mov \remainder, \numerator, lsr \bits | ||
187 | .else | ||
188 | .ifc "", "\remainder" | ||
189 | mov \divisor, \numerator, lsr \bits | ||
190 | add \numerator, \numerator, #1 | ||
191 | sub \quotient, \numerator, \divisor, lsl \bits | ||
192 | .else | ||
193 | mov \remainder, \numerator, lsr \bits | ||
194 | add \numerator, \numerator, #1 | ||
195 | sub \quotient, \numerator, \remainder, lsl \bits | ||
196 | .endif | ||
197 | .endif | ||
198 | .ifne \return | ||
199 | ARM_SDIV32_POST \quotient, \remainder, \sign | ||
200 | bx lr | ||
201 | .else | ||
202 | b 99f | ||
203 | .endif | ||
204 | 20: | ||
205 | .ifnc "", "\remainder" | ||
206 | sub \remainder, \numerator, \divisor | ||
207 | .endif | ||
208 | .ifnc "", "\quotient" | ||
209 | mov \quotient, #0 | ||
210 | .endif | ||
211 | .ifne \return | ||
212 | ARM_SDIV32_POST "", \remainder, \sign | ||
213 | bx lr | ||
214 | .else | ||
215 | 99: | ||
216 | ARM_SDIV32_POST \quotient, \remainder, \sign | ||
217 | .endif | ||
218 | .endm | ||
219 | |||
220 | #else | ||
221 | .macro ARMV5_UDIV32_BODY numerator, divisor, quotient, remainder, bits, inv, neg, div0label, return | ||
222 | cmp \numerator, \divisor | ||
223 | clz \bits, \divisor | ||
224 | bcc 30f | ||
225 | mov \inv, \divisor, lsl \bits | ||
226 | add \neg, pc, \inv, lsr #25 | ||
227 | /* Test whether divisor is 2^N */ | ||
228 | cmp \inv, #1<<31 | ||
229 | /* Load approximate reciprocal */ | ||
230 | ldrhib \inv, [\neg, #.L_udiv_est_table-.-64] | ||
231 | bls 20f | ||
232 | subs \bits, \bits, #7 | ||
233 | rsb \neg, \divisor, #0 | ||
234 | /* Scale approximate reciprocal, or else branch to large-divisor path */ | ||
235 | movpl \divisor, \inv, lsl \bits | ||
236 | bmi 10f | ||
237 | /* Newton-Raphson iteration to improve reciprocal accuracy */ | ||
238 | mul \inv, \divisor, \neg | ||
239 | smlawt \divisor, \divisor, \inv, \divisor | ||
240 | mul \inv, \divisor, \neg | ||
241 | /* Complete N-R math and produce approximate quotient. Use smmla/smmul on | ||
242 | ARMv6. */ | ||
243 | #if ARM_ARCH >= 6 | ||
244 | tst \numerator, \numerator | ||
245 | smmla \divisor, \divisor, \inv, \divisor | ||
246 | /* Branch to large-numerator handler, or else use smmul if sign bit is not | ||
247 | set. */ | ||
248 | bmi 40f | ||
249 | smmul \inv, \numerator, \divisor | ||
250 | #else | ||
251 | /* ARMv5e lacks smmul, so always uses umull. */ | ||
252 | mov \bits, #0 | ||
253 | smlal \bits, \divisor, \inv, \divisor | ||
254 | umull \bits, \inv, \numerator, \divisor | ||
255 | #endif | ||
256 | /* Calculate remainder and correct result. */ | ||
257 | add \numerator, \numerator, \neg | ||
258 | .ifnc "", "\remainder" | ||
259 | mla \remainder, \inv, \neg, \numerator | ||
260 | .ifnc "", "\quotient" | ||
261 | mov \quotient, \inv | ||
262 | cmn \remainder, \neg | ||
263 | subcs \remainder, \remainder, \neg | ||
264 | addpl \remainder, \remainder, \neg, lsl #1 | ||
265 | addcc \quotient, \quotient, #1 | ||
266 | addpl \quotient, \quotient, #2 | ||
267 | .else | ||
268 | cmn \remainder, \neg | ||
269 | subcs \remainder, \remainder, \neg | ||
270 | addpl \remainder, \remainder, \neg, lsl #1 | ||
271 | .endif | ||
272 | .else | ||
273 | mla \divisor, \inv, \neg, \numerator | ||
274 | mov \quotient, \inv | ||
275 | cmn \divisor, \neg | ||
276 | addcc \quotient, \quotient, #1 | ||
277 | addpl \quotient, \quotient, #2 | ||
278 | .endif | ||
279 | .if \return | ||
280 | bx lr | ||
281 | .else | ||
282 | b 99f | ||
283 | .endif | ||
284 | 10: | ||
285 | /* Very large divisors can be handled without further improving the | ||
286 | reciprocal. First the reciprocal must be reduced to ensure that it | ||
287 | underestimates the correct value. */ | ||
288 | rsb \bits, \bits, #0 | ||
289 | sub \inv, \inv, #4 | ||
290 | mov \divisor, \inv, lsr \bits | ||
291 | /* Calculate approximate quotient and remainder */ | ||
292 | umull \bits, \inv, \numerator, \divisor | ||
293 | /* Correct quotient and remainder */ | ||
294 | .ifnc "", "\remainder" | ||
295 | mla \remainder, \inv, \neg, \numerator | ||
296 | .ifnc "", "\quotient" | ||
297 | mov \quotient, \inv | ||
298 | cmn \neg, \remainder, lsr #1 | ||
299 | addcs \remainder, \remainder, \neg, lsl #1 | ||
300 | addcs \quotient, \quotient, #2 | ||
301 | cmn \neg, \remainder | ||
302 | addcs \remainder, \remainder, \neg | ||
303 | addcs \quotient, \quotient, #1 | ||
304 | .else | ||
305 | cmn \neg, \remainder, lsr #1 | ||
306 | addcs \remainder, \remainder, \neg, lsl #1 | ||
307 | cmn \neg, \remainder | ||
308 | addcs \remainder, \remainder, \neg | ||
309 | .endif | ||
310 | .else | ||
311 | mla \divisor, \inv, \neg, \numerator | ||
312 | mov \quotient, \inv | ||
313 | cmn \neg, \divisor, lsr #1 | ||
314 | addcs \divisor, \divisor, \neg, lsl #1 | ||
315 | addcs \quotient, \quotient, #2 | ||
316 | cmn \neg, \divisor | ||
317 | addcs \quotient, \quotient, #1 | ||
318 | .endif | ||
319 | .if \return | ||
320 | bx lr | ||
321 | .else | ||
322 | b 99f | ||
323 | .endif | ||
324 | 20: | ||
325 | /* Handle division by powers of two by shifting right. Mod is handled | ||
326 | by using divisor-1 as a bitmask. */ | ||
327 | .ifnc "", "\remainder" | ||
328 | .ifnc "", "\div0label" | ||
329 | bne \div0label | ||
330 | .endif | ||
331 | .ifnc "", "\quotient" | ||
332 | sub \divisor, \divisor, #1 | ||
333 | rsb \bits, \bits, #31 | ||
334 | and \remainder, \numerator, \divisor | ||
335 | mov \quotient, \numerator, lsr \bits | ||
336 | .else | ||
337 | sub \divisor, \divisor, #1 | ||
338 | and \remainder, \numerator, \divisor | ||
339 | .endif | ||
340 | .else | ||
341 | rsb \bits, \bits, #31 | ||
342 | .ifnc "", "\div0label" | ||
343 | bne \div0label | ||
344 | .endif | ||
345 | mov \quotient, \numerator, lsr \bits | ||
346 | .endif | ||
347 | .if \return | ||
348 | bx lr | ||
349 | .else | ||
350 | b 99f | ||
351 | .endif | ||
352 | 30: | ||
353 | /* Handle numerator < divisor - quotient is zero, remainder is numerator, | ||
354 | which must be restored to its original value on ARMv6. */ | ||
355 | .ifnc "", "\remainder" | ||
356 | mov \remainder, \numerator | ||
357 | .endif | ||
358 | .ifnc "", "\quotient" | ||
359 | mov \quotient, #0 | ||
360 | .endif | ||
361 | .if \return | ||
362 | bx lr | ||
363 | .endif | ||
364 | #if ARM_ARCH >= 6 | ||
365 | 40: | ||
366 | /* Handle large (sign bit set) numerators. Works exactly as the ARMv5e code | ||
367 | above 10:. */ | ||
368 | umull \bits, \inv, \numerator, \divisor | ||
369 | add \numerator, \numerator, \neg | ||
370 | .ifnc "", "\remainder" | ||
371 | mla \remainder, \inv, \neg, \numerator | ||
372 | .ifnc "", "\quotient" | ||
373 | mla \remainder, \inv, \neg, \numerator | ||
374 | mov \quotient, \inv | ||
375 | cmn \remainder, \neg | ||
376 | subcs \remainder, \remainder, \neg | ||
377 | addpl \remainder, \remainder, \neg, lsl #1 | ||
378 | addcc \quotient, \quotient, #1 | ||
379 | addpl \quotient, \quotient, #2 | ||
380 | .else | ||
381 | cmn \remainder, \neg | ||
382 | subcs \remainder, \remainder, \neg | ||
383 | addpl \remainder, \remainder, \neg, lsl #1 | ||
384 | .endif | ||
385 | .else | ||
386 | mla \divisor, \inv, \neg, \numerator | ||
387 | mov \quotient, \inv | ||
388 | cmn \divisor, \neg | ||
389 | addcc \quotient, \quotient, #1 | ||
390 | addpl \quotient, \quotient, #2 | ||
391 | .endif | ||
392 | .if \return | ||
393 | bx lr | ||
394 | .else | ||
395 | b 99f | ||
396 | .endif | ||
397 | #endif | ||
398 | 99: | ||
399 | .endm | ||
400 | |||
401 | .macro ARMV5_SDIV32_BODY numerator, divisor, quotient, remainder, bits, inv, neg, sign, div0label, return | ||
402 | /* sign[31] = divisor sign */ | ||
403 | ands \sign, \divisor, #1<<31 | ||
404 | rsbne \divisor, \divisor, #0 | ||
405 | /* sign[31] = result sign, sign[0:30], C = numerator sign */ | ||
406 | eors \sign, \sign, \numerator, asr #32 | ||
407 | clz \bits, \divisor | ||
408 | rsbcs \numerator, \numerator, #0 | ||
409 | /* On ARMv6, subtract divisor before performing division, which ensures | ||
410 | numerator sign bit is clear and smmul may be used in place of umull. The | ||
411 | fixup for the results can be fit entirely into existing delay slots on | ||
412 | the main division paths. It costs 1c in the num<div path if the | ||
413 | the remainder is to be produced in the numerator's register, and 1c in | ||
414 | the power-of-2-divisor path only if producing both remainder and | ||
415 | quotient. */ | ||
416 | #if ARM_ARCH >= 6 | ||
417 | subs \numerator, \numerator, \divisor | ||
418 | #else | ||
419 | cmp \numerator, \divisor | ||
420 | #endif | ||
421 | movcs \inv, \divisor, lsl \bits | ||
422 | bcc 30f | ||
423 | /* Test whether divisor is 2^N */ | ||
424 | cmp \inv, #1<<31 | ||
425 | add \inv, pc, \inv, lsr #25 | ||
426 | bls 20f | ||
427 | /* Load approximate reciprocal */ | ||
428 | ldrb \inv, [\inv, #.L_udiv_est_table-.-64] | ||
429 | subs \bits, \bits, #7 | ||
430 | rsb \neg, \divisor, #0 | ||
431 | /* Scale approximate reciprocal, or else branch to large-divisor path */ | ||
432 | movpl \divisor, \inv, lsl \bits | ||
433 | bmi 10f | ||
434 | /* Newton-Raphson iteration to improve reciprocal accuracy */ | ||
435 | mul \inv, \divisor, \neg | ||
436 | smlawt \divisor, \divisor, \inv, \divisor | ||
437 | mul \inv, \divisor, \neg | ||
438 | /* Complete N-R math and produce approximate quotient. Use smmla/smmul on | ||
439 | ARMv6. */ | ||
440 | #if ARM_ARCH >= 6 | ||
441 | smmla \divisor, \divisor, \inv, \divisor | ||
442 | smmul \inv, \numerator, \divisor | ||
443 | #else | ||
444 | mov \bits, #0 | ||
445 | smlal \bits, \divisor, \inv, \divisor | ||
446 | umull \bits, \inv, \numerator, \divisor | ||
447 | #endif | ||
448 | /* Calculate remainder and correct quotient. */ | ||
449 | add \numerator, \numerator, \neg | ||
450 | .ifnc "", "\remainder" | ||
451 | mla \remainder, \inv, \neg, \numerator | ||
452 | .ifnc "", "\quotient" | ||
453 | #if ARM_ARCH >= 6 | ||
454 | add \quotient, \inv, #1 | ||
455 | #else | ||
456 | mov \quotient, \inv | ||
457 | #endif | ||
458 | cmn \remainder, \neg | ||
459 | subcs \remainder, \remainder, \neg | ||
460 | addpl \remainder, \remainder, \neg, lsl #1 | ||
461 | addcc \quotient, \quotient, #1 | ||
462 | addpl \quotient, \quotient, #2 | ||
463 | .else | ||
464 | cmn \remainder, \neg | ||
465 | subcs \remainder, \remainder, \neg | ||
466 | addpl \remainder, \remainder, \neg, lsl #1 | ||
467 | .endif | ||
468 | .else | ||
469 | mla \divisor, \inv, \neg, \numerator | ||
470 | #if ARM_ARCH >= 6 | ||
471 | add \quotient, \inv, #1 | ||
472 | #else | ||
473 | mov \quotient, \inv | ||
474 | #endif | ||
475 | cmn \divisor, \neg | ||
476 | addcc \quotient, \quotient, #1 | ||
477 | addpl \quotient, \quotient, #2 | ||
478 | .endif | ||
479 | ARM_SDIV32_POST \quotient, \remainder, \sign | ||
480 | .ifnc "", "\return" | ||
481 | \return | ||
482 | .else | ||
483 | b 99f | ||
484 | .endif | ||
485 | 10: | ||
486 | /* Very large divisors can be handled without further improving the | ||
487 | reciprocal. First the reciprocal must be reduced to ensure that it | ||
488 | underestimates the correct value. */ | ||
489 | rsb \bits, \bits, #0 | ||
490 | sub \inv, \inv, #4 | ||
491 | mov \divisor, \inv, lsr \bits | ||
492 | /* Calculate approximate quotient and remainder */ | ||
493 | #if ARM_ARCH >= 6 | ||
494 | smmul \inv, \numerator, \divisor | ||
495 | #else | ||
496 | umull \bits, \inv, \numerator, \divisor | ||
497 | #endif | ||
498 | /* Correct quotient and remainder */ | ||
499 | .ifnc "", "\remainder" | ||
500 | mla \remainder, \inv, \neg, \numerator | ||
501 | .ifnc "", "\quotient" | ||
502 | #if ARM_ARCH >= 6 | ||
503 | add \quotient, \inv, #1 | ||
504 | #else | ||
505 | mov \quotient, \inv | ||
506 | #endif | ||
507 | cmn \neg, \remainder, lsr #1 | ||
508 | addcs \remainder, \remainder, \neg, lsl #1 | ||
509 | addcs \quotient, \quotient, #2 | ||
510 | cmn \neg, \remainder | ||
511 | addcs \remainder, \remainder, \neg | ||
512 | addcs \quotient, \quotient, #1 | ||
513 | .else | ||
514 | cmn \neg, \remainder, lsr #1 | ||
515 | addcs \remainder, \remainder, \neg, lsl #1 | ||
516 | cmn \neg, \remainder | ||
517 | addcs \remainder, \remainder, \neg | ||
518 | .endif | ||
519 | .else | ||
520 | mla \divisor, \inv, \neg, \numerator | ||
521 | #if ARM_ARCH >= 6 | ||
522 | add \quotient, \inv, #1 | ||
523 | #else | ||
524 | mov \quotient, \inv | ||
525 | #endif | ||
526 | cmn \neg, \divisor, lsr #1 | ||
527 | addcs \divisor, \divisor, \neg, lsl #1 | ||
528 | addcs \quotient, \quotient, #2 | ||
529 | cmn \neg, \divisor | ||
530 | addcs \quotient, \quotient, #1 | ||
531 | .endif | ||
532 | ARM_SDIV32_POST \quotient, \remainder, \sign | ||
533 | .ifnc "", "\return" | ||
534 | \return | ||
535 | .else | ||
536 | b 99f | ||
537 | .endif | ||
538 | 20: | ||
539 | /* Handle division by powers of two by shifting right. Mod is handled | ||
540 | by using divisor-1 as a bitmask. */ | ||
541 | .ifnc "", "\div0label" | ||
542 | bne \div0label | ||
543 | .endif | ||
544 | .ifnc "", "\remainder" | ||
545 | .ifnc "", "\quotient" | ||
546 | rsb \bits, \bits, #31 | ||
547 | #if ARM_ARCH >= 6 | ||
548 | add \numerator, \numerator, \divisor | ||
549 | #endif | ||
550 | sub \divisor, \divisor, #1 | ||
551 | and \remainder, \numerator, \divisor | ||
552 | mov \quotient, \numerator, lsr \bits | ||
553 | .else | ||
554 | sub \divisor, \divisor, #1 | ||
555 | and \remainder, \numerator, \divisor | ||
556 | .endif | ||
557 | .else | ||
558 | rsb \bits, \bits, #31 | ||
559 | #if ARM_ARCH >= 6 | ||
560 | add \numerator, \numerator, \divisor | ||
561 | #endif | ||
562 | mov \quotient, \numerator, lsr \bits | ||
563 | .endif | ||
564 | ARM_SDIV32_POST \quotient, \remainder, \sign | ||
565 | .ifnc "", "\return" | ||
566 | \return | ||
567 | .else | ||
568 | b 99f | ||
569 | .endif | ||
570 | 30: | ||
571 | /* Handle numerator < divisor - quotient is zero, remainder is numerator, | ||
572 | which must be restored to its original value on ARMv6. */ | ||
573 | .ifnc "", "\remainder" | ||
574 | #if ARM_ARCH >= 6 | ||
575 | add \remainder, \numerator, \divisor | ||
576 | #else | ||
577 | .ifnc "\remainder", "\numerator" | ||
578 | mov \remainder, \numerator | ||
579 | .endif | ||
580 | #endif | ||
581 | .endif | ||
582 | .ifnc "", "\quotient" | ||
583 | mov \quotient, #0 | ||
584 | .endif | ||
585 | ARM_SDIV32_POST "", \remainder, \sign | ||
586 | .ifnc "", "\return" | ||
587 | \return | ||
588 | .endif | ||
589 | 99: | ||
590 | .endm | ||
591 | #endif | ||
592 | |||
593 | .section .text | ||
594 | |||
595 | __div0_wrap_s: | ||
596 | sub sp, sp, #4 | ||
597 | mov r0, #0 | ||
598 | b __div0 | ||
599 | .size __div0_wrap_s, . - __div0_wrap_s | ||
600 | |||
601 | __div0_wrap: | ||
602 | str lr, [sp, #-4]! | ||
603 | mov r0, #0 | ||
604 | b __div0 | ||
605 | .size __div0_wrap, . - __div0_wrap | ||
606 | |||
607 | #ifndef __ARM_EABI__ | ||
608 | .global __divsi3 | ||
609 | .type __divsi3,%function | ||
610 | .global __udivsi3 | ||
611 | .type __udivsi3,%function | ||
612 | .global __udivsi3 | ||
613 | .type __udivsi3,%function | ||
614 | #else | ||
615 | /* The div+mod averagess a fraction of a cycle worse for signed values, and | ||
616 | slightly better for unsigned, so just alias div to divmod. */ | ||
617 | .global __aeabi_uidivmod | ||
618 | .type __aeabi_uidivmod,%function | ||
619 | .global __aeabi_uidiv | ||
620 | .type __aeabi_uidiv,%function | ||
621 | .set __aeabi_uidiv,__aeabi_uidivmod | ||
622 | .global __aeabi_idivmod | ||
623 | .type __aeabi_idivmod,%function | ||
624 | .global __aeabi_idiv | ||
625 | .type __aeabi_idiv,%function | ||
626 | .set __aeabi_idiv,__aeabi_idivmod | ||
627 | #endif | ||
628 | |||
629 | |||
630 | #if ARM_ARCH < 5 | ||
631 | .global __clzsi2 | ||
632 | .type __clzsi2, %function | ||
633 | |||
634 | __clzsi2: | ||
635 | orr r0, r0, r0, lsr #8 | ||
636 | orr r0, r0, r0, lsr #4 | ||
637 | orr r0, r0, r0, lsr #2 | ||
638 | orr r0, r0, r0, lsr #1 | ||
639 | bic r0, r0, r0, lsr #16 | ||
640 | rsb r0, r0, r0, lsl #14 | ||
641 | rsb r0, r0, r0, lsl #11 | ||
642 | rsb r0, r0, r0, lsl #9 | ||
643 | ldrb r0, [pc, r0, lsr #26] | ||
644 | bx lr | ||
645 | .byte 32, 20, 19, 0, 0, 18, 0, 7, 10, 17, 0, 0, 14, 0, 6, 0 | ||
646 | .byte 0, 9, 0, 16, 0, 0, 1, 26, 0, 13, 0, 0, 24, 5, 0, 0 | ||
647 | .byte 0, 21, 0, 8, 11, 0, 15, 0, 0, 0, 0, 2, 27, 0, 25, 0 | ||
648 | .byte 22, 0, 12, 0, 0, 3, 28, 0, 23, 0, 4, 29, 0, 0, 30, 31 | ||
649 | .size __clzsi2, .-__clzsi2 | ||
650 | |||
651 | #ifndef __ARM_EABI__ | ||
652 | __udivsi3: | ||
653 | ARMV4_UDIV32_BODY r0, r1, r0, "", r2, r3, __div0_wrap, 1 | ||
654 | .size __udivsi3, . - __udivsi3 | ||
655 | |||
656 | __divsi3: | ||
657 | ARMV4_SDIV32_BODY r0, r1, r0, "", r2, r3, __div0_wrap, 1 | ||
658 | .size __divsi3, . - __divsi3 | ||
659 | |||
660 | #else | ||
661 | __aeabi_uidivmod: | ||
662 | ARMV4_UDIV32_BODY r0, r1, r0, r1, r2, r3, __div0_wrap, 1 | ||
663 | .size __aeabi_uidivmod, . - __aeabi_uidivmo | ||
664 | |||
665 | __aeabi_idivmod | ||
666 | ARMV4_SDIV32_BODY r0, r1, r0, r1, r2, r3, __div0_wrap, 1 | ||
667 | .size __aeabi_idivmod, . - __aeabi_idivmod | ||
668 | #endif | ||
669 | |||
670 | #else | ||
671 | #ifndef __ARM_EABI__ | ||
672 | __udivsi3: | ||
673 | ARMV5_UDIV32_BODY r0, r1, r0, "", r2, r3, ip, __div0_wrap, 1 | ||
674 | .size __udivsi3, . - __udivsi3 | ||
675 | |||
676 | __divsi3: | ||
677 | ARMV5_SDIV32_BODY r0, r1, r0, "", r2, lr, ip, r3, __div0_wrap_s, "ldr pc, [sp, #-4]" | ||
678 | .size __divsi3, . - __divsi3 | ||
679 | |||
680 | #else | ||
681 | __aeabi_uidivmod: | ||
682 | ARMV5_UDIV32_BODY r0, r1, r0, r1, r2, r3, ip, __div0_wrap, 1 | ||
683 | .size __aeabi_uidivmod, . - __aeabi_uidivmo | ||
684 | |||
685 | __aeabi_idivmod | ||
686 | ARMV5_SDIV32_BODY r0, r1, r0, r1, r2, lr, ip, r3, __div0_wrap_s, "ldr pc, [sp, #-4]" | ||
687 | .size __aeabi_idivmod, . - __aeabi_idivmod | ||
688 | #endif | ||
689 | |||
690 | .L_udiv_est_table: | ||
691 | .byte 0xff, 0xfc, 0xf8, 0xf4, 0xf0, 0xed, 0xea, 0xe6 | ||
692 | .byte 0xe3, 0xe0, 0xdd, 0xda, 0xd7, 0xd4, 0xd2, 0xcf | ||
693 | .byte 0xcc, 0xca, 0xc7, 0xc5, 0xc3, 0xc0, 0xbe, 0xbc | ||
694 | .byte 0xba, 0xb8, 0xb6, 0xb4, 0xb2, 0xb0, 0xae, 0xac | ||
695 | .byte 0xaa, 0xa8, 0xa7, 0xa5, 0xa3, 0xa2, 0xa0, 0x9f | ||
696 | .byte 0x9d, 0x9c, 0x9a, 0x99, 0x97, 0x96, 0x94, 0x93 | ||
697 | .byte 0x92, 0x90, 0x8f, 0x8e, 0x8d, 0x8c, 0x8a, 0x89 | ||
698 | .byte 0x88, 0x87, 0x86, 0x85, 0x84, 0x83, 0x82, 0x81 | ||
699 | #endif | ||