You can not select more than 25 topics Topics must start with a chinese character,a letter or number, can include dashes ('-') and can be up to 35 characters long.

izamin.S 16 kB

123456789101112131415161718192021222324252627282930313233343536373839404142434445464748495051525354555657585960616263646566676869707172737475767778798081828384858687888990919293949596979899100101102103104105106107108109110111112113114115116117118119120121122123124125126127128129130131132133134135136137138139140141142143144145146147148149150151152153154155156157158159160161162163164165166167168169170171172173174175176177178179180181182183184185186187188189190191192193194195196197198199200201202203204205206207208209210211212213214215216217218219220221222223224225226227228229230231232233234235236237238239240241242243244245246247248249250251252253254255256257258259260261262263264265266267268269270271272273274275276277278279280281282283284285286287288289290291292293294295296297298299300301302303304305306307308309310311312313314315316317318319320321322323324325326327328329330331332333334335336337338339340341342343344345346347348349350351352353354355356357358359360361362363364365366367368369370371372373374375376377378379380381382383384385386387388389390391392393394395396397398399400401402403404405406407408409410411412413414415416417418419420421422423424425426427428429430431432433434435436437438439440441442443444445446447448449450451452453454455456457458459460461462463464465466467468469470471472473474475476477478479480481482483484485486487488489490491492493494495496497498499500501502503504505506507508509510511512513514515516517518519520521522523524525526527528529530531532533534535536537538539540541542543544545546547548549550551552553554555556557558559560561562563564565566567568569570571572573574575576577578579580581582583584585586587588589590591592593594595596597598599600601602603604605606607608609610611612613614615616617618619620621622623624625626627628629630631632633634635636637638639640641642643644645646647648649650651652653654655656657658659660661662663664665666667668669670671672673674675676677678679680681682683684685686687688689690691692693694695696697698699700701702703704705706707708709710711712713714715716717718719720721722723724725726727728729730731732733734735736737738739740741742743744745746747748749750751752753754755756757758759760761762763764765766767768769770771772773774775776777778779780781782783784785786787788789790791792793794795796797798799800801802803804805806807808809810811812813814815816817818819820821822823824825826827828829830831832833834835836837838839840841842843844845846847848849850851852853854855856857858859860861862863864865866867868869870871872873874875876877878879880881882883884885886887888889890891892893894895896897898899900901902903904905906907908909910911912913914915916917918919920
  1. /*********************************************************************/
  2. /* Copyright 2009, 2010 The University of Texas at Austin. */
  3. /* All rights reserved. */
  4. /* */
  5. /* Redistribution and use in source and binary forms, with or */
  6. /* without modification, are permitted provided that the following */
  7. /* conditions are met: */
  8. /* */
  9. /* 1. Redistributions of source code must retain the above */
  10. /* copyright notice, this list of conditions and the following */
  11. /* disclaimer. */
  12. /* */
  13. /* 2. Redistributions in binary form must reproduce the above */
  14. /* copyright notice, this list of conditions and the following */
  15. /* disclaimer in the documentation and/or other materials */
  16. /* provided with the distribution. */
  17. /* */
  18. /* THIS SOFTWARE IS PROVIDED BY THE UNIVERSITY OF TEXAS AT */
  19. /* AUSTIN ``AS IS'' AND ANY EXPRESS OR IMPLIED WARRANTIES, */
  20. /* INCLUDING, BUT NOT LIMITED TO, THE IMPLIED WARRANTIES OF */
  21. /* MERCHANTABILITY AND FITNESS FOR A PARTICULAR PURPOSE ARE */
  22. /* DISCLAIMED. IN NO EVENT SHALL THE UNIVERSITY OF TEXAS AT */
  23. /* AUSTIN OR CONTRIBUTORS BE LIABLE FOR ANY DIRECT, INDIRECT, */
  24. /* INCIDENTAL, SPECIAL, EXEMPLARY, OR CONSEQUENTIAL DAMAGES */
  25. /* (INCLUDING, BUT NOT LIMITED TO, PROCUREMENT OF SUBSTITUTE */
  26. /* GOODS OR SERVICES; LOSS OF USE, DATA, OR PROFITS; OR */
  27. /* BUSINESS INTERRUPTION) HOWEVER CAUSED AND ON ANY THEORY OF */
  28. /* LIABILITY, WHETHER IN CONTRACT, STRICT LIABILITY, OR TORT */
  29. /* (INCLUDING NEGLIGENCE OR OTHERWISE) ARISING IN ANY WAY OUT */
  30. /* OF THE USE OF THIS SOFTWARE, EVEN IF ADVISED OF THE */
  31. /* POSSIBILITY OF SUCH DAMAGE. */
  32. /* */
  33. /* The views and conclusions contained in the software and */
  34. /* documentation are those of the authors and should not be */
  35. /* interpreted as representing official policies, either expressed */
  36. /* or implied, of The University of Texas at Austin. */
  37. /*********************************************************************/
  38. #define ASSEMBLER
  39. #include "common.h"
  40. #define RET r3
  41. #define X r4
  42. #define INCX r5
  43. #define N r6
  44. #define NN r7
  45. #define XX r8
  46. #define PREA r9
  47. #define INCXM1 r10
  48. #define FZERO f1
  49. #define STACKSIZE 160
  50. PROLOGUE
  51. PROFCODE
  52. addi SP, SP, -STACKSIZE
  53. li r0, 0
  54. stfd f14, 0(SP)
  55. stfd f15, 8(SP)
  56. stfd f16, 16(SP)
  57. stfd f17, 24(SP)
  58. stfd f18, 32(SP)
  59. stfd f19, 40(SP)
  60. stfd f20, 48(SP)
  61. stfd f21, 56(SP)
  62. stfd f22, 64(SP)
  63. stfd f23, 72(SP)
  64. stfd f24, 80(SP)
  65. stfd f25, 88(SP)
  66. stfd f26, 96(SP)
  67. stfd f27, 104(SP)
  68. stfd f28, 112(SP)
  69. stfd f29, 120(SP)
  70. stfd f30, 128(SP)
  71. stfd f31, 136(SP)
  72. stw r0, 144(SP)
  73. lfs FZERO,144(SP)
  74. #ifdef F_INTERFACE
  75. LDINT N, 0(r3)
  76. LDINT INCX, 0(INCX)
  77. #else
  78. mr N, r3
  79. #endif
  80. li RET, 0
  81. mr NN, N
  82. mr XX, X
  83. slwi INCX, INCX, ZBASE_SHIFT
  84. subi INCXM1, INCX, SIZE
  85. li PREA, L1_PREFETCHSIZE
  86. cmpwi cr0, N, 0
  87. ble- LL(9999)
  88. cmpwi cr0, INCX, 0
  89. ble- LL(9999)
  90. LFD f1, 0 * SIZE(X)
  91. LFD f2, 1 * SIZE(X)
  92. add X, X, INCX
  93. fabs f1, f1
  94. fabs f2, f2
  95. fadd f1, f1, f2
  96. fmr f0, f1
  97. fmr f2, f1
  98. fmr f3, f1
  99. subi N, N, 1
  100. cmpwi cr0, INCX, 2 * SIZE
  101. bne- cr0, LL(100)
  102. srawi. r0, N, 3
  103. mtspr CTR, r0
  104. beq- cr0, LL(50)
  105. .align 4
  106. LFD f24, 0 * SIZE(X)
  107. LFD f25, 1 * SIZE(X)
  108. LFD f26, 2 * SIZE(X)
  109. LFD f27, 3 * SIZE(X)
  110. LFD f28, 4 * SIZE(X)
  111. LFD f29, 5 * SIZE(X)
  112. LFD f30, 6 * SIZE(X)
  113. LFD f31, 7 * SIZE(X)
  114. fabs f8, f24
  115. fabs f9, f25
  116. fabs f10, f26
  117. fabs f11, f27
  118. fabs f12, f28
  119. fabs f13, f29
  120. fabs f14, f30
  121. fabs f15, f31
  122. LFD f24, 8 * SIZE(X)
  123. LFD f25, 9 * SIZE(X)
  124. LFD f26, 10 * SIZE(X)
  125. LFD f27, 11 * SIZE(X)
  126. LFD f28, 12 * SIZE(X)
  127. LFD f29, 13 * SIZE(X)
  128. LFD f30, 14 * SIZE(X)
  129. LFD f31, 15 * SIZE(X)
  130. bdz LL(20)
  131. .align 4
  132. LL(10):
  133. fadd f4, f8, f9
  134. fadd f5, f10, f11
  135. fadd f6, f12, f13
  136. fadd f7, f14, f15
  137. fabs f8, f24
  138. fabs f9, f25
  139. fabs f10, f26
  140. fabs f11, f27
  141. LFD f24, 16 * SIZE(X)
  142. LFD f25, 17 * SIZE(X)
  143. LFD f26, 18 * SIZE(X)
  144. LFD f27, 19 * SIZE(X)
  145. fabs f12, f28
  146. fabs f13, f29
  147. fabs f14, f30
  148. fabs f15, f31
  149. LFD f28, 20 * SIZE(X)
  150. LFD f29, 21 * SIZE(X)
  151. LFD f30, 22 * SIZE(X)
  152. LFD f31, 23 * SIZE(X)
  153. fsub f16, f0, f4
  154. fsub f17, f1, f5
  155. fsub f18, f2, f6
  156. fsub f19, f3, f7
  157. fadd f20, f8, f9
  158. fadd f21, f10, f11
  159. fadd f22, f12, f13
  160. fadd f23, f14, f15
  161. fabs f8, f24
  162. fabs f9, f25
  163. fabs f10, f26
  164. fabs f11, f27
  165. LFD f24, 24 * SIZE(X)
  166. LFD f25, 25 * SIZE(X)
  167. LFD f26, 26 * SIZE(X)
  168. LFD f27, 27 * SIZE(X)
  169. fsel f0, f16, f4, f0
  170. fsel f1, f17, f5, f1
  171. fsel f2, f18, f6, f2
  172. fsel f3, f19, f7, f3
  173. fabs f12, f28
  174. fabs f13, f29
  175. fabs f14, f30
  176. fabs f15, f31
  177. LFD f28, 28 * SIZE(X)
  178. LFD f29, 29 * SIZE(X)
  179. LFD f30, 30 * SIZE(X)
  180. LFD f31, 31 * SIZE(X)
  181. fsub f16, f0, f20
  182. fsub f17, f1, f21
  183. fsub f18, f2, f22
  184. fsub f19, f3, f23
  185. fsel f0, f16, f20, f0
  186. fsel f1, f17, f21, f1
  187. fsel f2, f18, f22, f2
  188. fsel f3, f19, f23, f3
  189. #ifndef POWER6
  190. L1_PREFETCH X, PREA
  191. #endif
  192. addi X, X, 16 * SIZE
  193. #ifdef POWER6
  194. L1_PREFETCH X, PREA
  195. #endif
  196. bdnz LL(10)
  197. .align 4
  198. LL(20):
  199. fadd f4, f8, f9
  200. fadd f5, f10, f11
  201. fadd f6, f12, f13
  202. fadd f7, f14, f15
  203. fabs f8, f24
  204. fabs f9, f25
  205. fabs f10, f26
  206. fabs f11, f27
  207. fabs f12, f28
  208. fabs f13, f29
  209. fabs f14, f30
  210. fabs f15, f31
  211. fsub f16, f0, f4
  212. fsub f17, f1, f5
  213. fsub f18, f2, f6
  214. fsub f19, f3, f7
  215. fadd f20, f8, f9
  216. fadd f21, f10, f11
  217. fadd f22, f12, f13
  218. fadd f23, f14, f15
  219. fsel f0, f16, f4, f0
  220. fsel f1, f17, f5, f1
  221. fsel f2, f18, f6, f2
  222. fsel f3, f19, f7, f3
  223. fsub f16, f0, f20
  224. fsub f17, f1, f21
  225. fsub f18, f2, f22
  226. fsub f19, f3, f23
  227. fsel f0, f16, f20, f0
  228. fsel f1, f17, f21, f1
  229. fsel f2, f18, f22, f2
  230. fsel f3, f19, f23, f3
  231. addi X, X, 16 * SIZE
  232. .align 4
  233. LL(50):
  234. andi. r0, N, 7
  235. mtspr CTR, r0
  236. beq LL(999)
  237. .align 4
  238. LL(60):
  239. LFD f8, 0 * SIZE(X)
  240. LFD f9, 1 * SIZE(X)
  241. addi X, X, 2 * SIZE
  242. fabs f8, f8
  243. fabs f9, f9
  244. fadd f8, f8, f9
  245. fsub f16, f1, f8
  246. fsel f1, f16, f8, f1
  247. bdnz LL(60)
  248. b LL(999)
  249. .align 4
  250. LL(100):
  251. sub X, X, INCXM1
  252. srawi. r0, N, 3
  253. mtspr CTR, r0
  254. beq- LL(150)
  255. LFDX f24, X, INCXM1
  256. LFDUX f25, X, INCX
  257. LFDX f26, X, INCXM1
  258. LFDUX f27, X, INCX
  259. LFDX f28, X, INCXM1
  260. LFDUX f29, X, INCX
  261. LFDX f30, X, INCXM1
  262. LFDUX f31, X, INCX
  263. fabs f8, f24
  264. fabs f9, f25
  265. fabs f10, f26
  266. fabs f11, f27
  267. fabs f12, f28
  268. fabs f13, f29
  269. fabs f14, f30
  270. fabs f15, f31
  271. LFDX f24, X, INCXM1
  272. LFDUX f25, X, INCX
  273. LFDX f26, X, INCXM1
  274. LFDUX f27, X, INCX
  275. LFDX f28, X, INCXM1
  276. LFDUX f29, X, INCX
  277. LFDX f30, X, INCXM1
  278. LFDUX f31, X, INCX
  279. bdz LL(120)
  280. .align 4
  281. LL(110):
  282. fadd f4, f8, f9
  283. fadd f5, f10, f11
  284. fadd f6, f12, f13
  285. fadd f7, f14, f15
  286. fabs f8, f24
  287. fabs f9, f25
  288. fabs f10, f26
  289. fabs f11, f27
  290. LFDX f24, X, INCXM1
  291. LFDUX f25, X, INCX
  292. LFDX f26, X, INCXM1
  293. LFDUX f27, X, INCX
  294. fabs f12, f28
  295. fabs f13, f29
  296. fabs f14, f30
  297. fabs f15, f31
  298. LFDX f28, X, INCXM1
  299. LFDUX f29, X, INCX
  300. LFDX f30, X, INCXM1
  301. LFDUX f31, X, INCX
  302. fsub f16, f0, f4
  303. fsub f17, f1, f5
  304. fsub f18, f2, f6
  305. fsub f19, f3, f7
  306. fadd f20, f8, f9
  307. fadd f21, f10, f11
  308. fadd f22, f12, f13
  309. fadd f23, f14, f15
  310. fabs f8, f24
  311. fabs f9, f25
  312. fabs f10, f26
  313. fabs f11, f27
  314. LFDX f24, X, INCXM1
  315. LFDUX f25, X, INCX
  316. LFDX f26, X, INCXM1
  317. LFDUX f27, X, INCX
  318. fsel f0, f16, f4, f0
  319. fsel f1, f17, f5, f1
  320. fsel f2, f18, f6, f2
  321. fsel f3, f19, f7, f3
  322. fabs f12, f28
  323. fabs f13, f29
  324. fabs f14, f30
  325. fabs f15, f31
  326. LFDX f28, X, INCXM1
  327. LFDUX f29, X, INCX
  328. LFDX f30, X, INCXM1
  329. LFDUX f31, X, INCX
  330. fsub f16, f0, f20
  331. fsub f17, f1, f21
  332. fsub f18, f2, f22
  333. fsub f19, f3, f23
  334. fsel f0, f16, f20, f0
  335. fsel f1, f17, f21, f1
  336. fsel f2, f18, f22, f2
  337. fsel f3, f19, f23, f3
  338. bdnz LL(110)
  339. .align 4
  340. LL(120):
  341. fadd f4, f8, f9
  342. fadd f5, f10, f11
  343. fadd f6, f12, f13
  344. fadd f7, f14, f15
  345. fabs f8, f24
  346. fabs f9, f25
  347. fabs f10, f26
  348. fabs f11, f27
  349. fabs f12, f28
  350. fabs f13, f29
  351. fabs f14, f30
  352. fabs f15, f31
  353. fsub f16, f0, f4
  354. fsub f17, f1, f5
  355. fsub f18, f2, f6
  356. fsub f19, f3, f7
  357. fadd f20, f8, f9
  358. fadd f21, f10, f11
  359. fadd f22, f12, f13
  360. fadd f23, f14, f15
  361. fsel f0, f16, f4, f0
  362. fsel f1, f17, f5, f1
  363. fsel f2, f18, f6, f2
  364. fsel f3, f19, f7, f3
  365. fsub f16, f0, f20
  366. fsub f17, f1, f21
  367. fsub f18, f2, f22
  368. fsub f19, f3, f23
  369. fsel f0, f16, f20, f0
  370. fsel f1, f17, f21, f1
  371. fsel f2, f18, f22, f2
  372. fsel f3, f19, f23, f3
  373. .align 4
  374. LL(150):
  375. andi. r0, N, 7
  376. mtspr CTR, r0
  377. beq LL(999)
  378. .align 4
  379. LL(160):
  380. LFDX f8, X, INCXM1
  381. LFDUX f9, X, INCX
  382. fabs f8, f8
  383. fabs f9, f9
  384. fadd f8, f8, f9
  385. fsub f16, f1, f8
  386. fsel f1, f16, f8, f1
  387. bdnz LL(160)
  388. .align 4
  389. LL(999):
  390. fsub f8, f0, f1
  391. fsub f9, f2, f3
  392. fsel f0, f8, f1, f0
  393. fsel f2, f9, f3, f2
  394. fsub f8, f0, f2
  395. fsel f1, f8, f2, f0
  396. .align 4
  397. LL(1000):
  398. cmpwi cr0, INCX, SIZE * 2
  399. bne- cr0, LL(1100)
  400. srawi. r0, NN, 3
  401. mtspr CTR, r0
  402. beq- cr0, LL(1050)
  403. LFD f24, 0 * SIZE(XX)
  404. LFD f25, 1 * SIZE(XX)
  405. LFD f26, 2 * SIZE(XX)
  406. LFD f27, 3 * SIZE(XX)
  407. LFD f28, 4 * SIZE(XX)
  408. LFD f29, 5 * SIZE(XX)
  409. LFD f30, 6 * SIZE(XX)
  410. LFD f31, 7 * SIZE(XX)
  411. bdz LL(1020)
  412. .align 4
  413. LL(1010):
  414. fabs f8, f24
  415. fabs f9, f25
  416. fabs f10, f26
  417. fabs f11, f27
  418. LFD f24, 8 * SIZE(XX)
  419. LFD f25, 9 * SIZE(XX)
  420. LFD f26, 10 * SIZE(XX)
  421. LFD f27, 11 * SIZE(XX)
  422. fabs f12, f28
  423. fabs f13, f29
  424. fabs f14, f30
  425. fabs f15, f31
  426. LFD f28, 12 * SIZE(XX)
  427. LFD f29, 13 * SIZE(XX)
  428. LFD f30, 14 * SIZE(XX)
  429. LFD f31, 15 * SIZE(XX)
  430. fadd f4, f8, f9
  431. fadd f5, f10, f11
  432. fadd f6, f12, f13
  433. fadd f7, f14, f15
  434. addi RET, RET, 1
  435. fcmpu cr0, f1, f4
  436. beq cr0, LL(9999)
  437. addi RET, RET, 1
  438. fcmpu cr0, f1, f5
  439. beq cr0, LL(9999)
  440. addi RET, RET, 1
  441. fcmpu cr0, f1, f6
  442. beq cr0, LL(9999)
  443. addi RET, RET, 1
  444. fcmpu cr0, f1, f7
  445. beq cr0, LL(9999)
  446. fabs f8, f24
  447. fabs f9, f25
  448. fabs f10, f26
  449. fabs f11, f27
  450. LFD f24, 16 * SIZE(XX)
  451. LFD f25, 17 * SIZE(XX)
  452. LFD f26, 18 * SIZE(XX)
  453. LFD f27, 19 * SIZE(XX)
  454. fabs f12, f28
  455. fabs f13, f29
  456. fabs f14, f30
  457. fabs f15, f31
  458. LFD f28, 20 * SIZE(XX)
  459. LFD f29, 21 * SIZE(XX)
  460. LFD f30, 22 * SIZE(XX)
  461. LFD f31, 23 * SIZE(XX)
  462. fadd f4, f8, f9
  463. fadd f5, f10, f11
  464. fadd f6, f12, f13
  465. fadd f7, f14, f15
  466. addi RET, RET, 1
  467. fcmpu cr0, f1, f4
  468. beq cr0, LL(9999)
  469. addi RET, RET, 1
  470. fcmpu cr0, f1, f5
  471. beq cr0, LL(9999)
  472. addi RET, RET, 1
  473. fcmpu cr0, f1, f6
  474. beq cr0, LL(9999)
  475. addi RET, RET, 1
  476. fcmpu cr0, f1, f7
  477. beq cr0, LL(9999)
  478. addi XX, XX, 16 * SIZE
  479. bdnz LL(1010)
  480. .align 4
  481. LL(1020):
  482. fabs f8, f24
  483. fabs f9, f25
  484. fabs f10, f26
  485. fabs f11, f27
  486. LFD f24, 8 * SIZE(XX)
  487. LFD f25, 9 * SIZE(XX)
  488. LFD f26, 10 * SIZE(XX)
  489. LFD f27, 11 * SIZE(XX)
  490. fabs f12, f28
  491. fabs f13, f29
  492. fabs f14, f30
  493. fabs f15, f31
  494. LFD f28, 12 * SIZE(XX)
  495. LFD f29, 13 * SIZE(XX)
  496. LFD f30, 14 * SIZE(XX)
  497. LFD f31, 15 * SIZE(XX)
  498. fadd f4, f8, f9
  499. fadd f5, f10, f11
  500. fadd f6, f12, f13
  501. fadd f7, f14, f15
  502. addi RET, RET, 1
  503. fcmpu cr0, f1, f4
  504. beq cr0, LL(9999)
  505. addi RET, RET, 1
  506. fcmpu cr0, f1, f5
  507. beq cr0, LL(9999)
  508. addi RET, RET, 1
  509. fcmpu cr0, f1, f6
  510. beq cr0, LL(9999)
  511. addi RET, RET, 1
  512. fcmpu cr0, f1, f7
  513. beq cr0, LL(9999)
  514. fabs f8, f24
  515. fabs f9, f25
  516. fabs f10, f26
  517. fabs f11, f27
  518. fabs f12, f28
  519. fabs f13, f29
  520. fabs f14, f30
  521. fabs f15, f31
  522. fadd f4, f8, f9
  523. fadd f5, f10, f11
  524. fadd f6, f12, f13
  525. fadd f7, f14, f15
  526. addi RET, RET, 1
  527. fcmpu cr0, f1, f4
  528. beq cr0, LL(9999)
  529. addi RET, RET, 1
  530. fcmpu cr0, f1, f5
  531. beq cr0, LL(9999)
  532. addi RET, RET, 1
  533. fcmpu cr0, f1, f6
  534. beq cr0, LL(9999)
  535. addi RET, RET, 1
  536. fcmpu cr0, f1, f7
  537. beq cr0, LL(9999)
  538. addi XX, XX, 16 * SIZE
  539. .align 4
  540. LL(1050):
  541. andi. r0, NN, 7
  542. mtspr CTR, r0
  543. beq LL(9999)
  544. .align 4
  545. LL(1060):
  546. LFD f8, 0 * SIZE(XX)
  547. LFD f9, 1 * SIZE(XX)
  548. addi XX, XX, 2 * SIZE
  549. fabs f8, f8
  550. fabs f9, f9
  551. fadd f8, f8, f9
  552. addi RET, RET, 1
  553. fcmpu cr0, f1, f8
  554. beq cr0, LL(9999)
  555. bdnz LL(1060)
  556. b LL(9999)
  557. .align 4
  558. LL(1100):
  559. sub XX, XX, INCXM1
  560. srawi. r0, NN, 3
  561. mtspr CTR, r0
  562. beq- LL(1150)
  563. LFDX f24, XX, INCXM1
  564. LFDUX f25, XX, INCX
  565. LFDX f26, XX, INCXM1
  566. LFDUX f27, XX, INCX
  567. LFDX f28, XX, INCXM1
  568. LFDUX f29, XX, INCX
  569. LFDX f30, XX, INCXM1
  570. LFDUX f31, XX, INCX
  571. bdz LL(1120)
  572. .align 4
  573. LL(1110):
  574. fabs f8, f24
  575. fabs f9, f25
  576. fabs f10, f26
  577. fabs f11, f27
  578. LFDX f24, XX, INCXM1
  579. LFDUX f25, XX, INCX
  580. LFDX f26, XX, INCXM1
  581. LFDUX f27, XX, INCX
  582. fabs f12, f28
  583. fabs f13, f29
  584. fabs f14, f30
  585. fabs f15, f31
  586. LFDX f28, XX, INCXM1
  587. LFDUX f29, XX, INCX
  588. LFDX f30, XX, INCXM1
  589. LFDUX f31, XX, INCX
  590. fadd f4, f8, f9
  591. fadd f5, f10, f11
  592. fadd f6, f12, f13
  593. fadd f7, f14, f15
  594. addi RET, RET, 1
  595. fcmpu cr0, f1, f4
  596. beq cr0, LL(9999)
  597. addi RET, RET, 1
  598. fcmpu cr0, f1, f5
  599. beq cr0, LL(9999)
  600. addi RET, RET, 1
  601. fcmpu cr0, f1, f6
  602. beq cr0, LL(9999)
  603. addi RET, RET, 1
  604. fcmpu cr0, f1, f7
  605. beq cr0, LL(9999)
  606. fabs f8, f24
  607. fabs f9, f25
  608. fabs f10, f26
  609. fabs f11, f27
  610. LFDX f24, XX, INCXM1
  611. LFDUX f25, XX, INCX
  612. LFDX f26, XX, INCXM1
  613. LFDUX f27, XX, INCX
  614. fabs f12, f28
  615. fabs f13, f29
  616. fabs f14, f30
  617. fabs f15, f31
  618. LFDX f28, XX, INCXM1
  619. LFDUX f29, XX, INCX
  620. LFDX f30, XX, INCXM1
  621. LFDUX f31, XX, INCX
  622. fadd f4, f8, f9
  623. fadd f5, f10, f11
  624. fadd f6, f12, f13
  625. fadd f7, f14, f15
  626. addi RET, RET, 1
  627. fcmpu cr0, f1, f4
  628. beq cr0, LL(9999)
  629. addi RET, RET, 1
  630. fcmpu cr0, f1, f5
  631. beq cr0, LL(9999)
  632. addi RET, RET, 1
  633. fcmpu cr0, f1, f6
  634. beq cr0, LL(9999)
  635. addi RET, RET, 1
  636. fcmpu cr0, f1, f7
  637. beq cr0, LL(9999)
  638. bdnz LL(1110)
  639. .align 4
  640. LL(1120):
  641. fabs f8, f24
  642. fabs f9, f25
  643. fabs f10, f26
  644. fabs f11, f27
  645. LFDX f24, XX, INCXM1
  646. LFDUX f25, XX, INCX
  647. LFDX f26, XX, INCXM1
  648. LFDUX f27, XX, INCX
  649. fabs f12, f28
  650. fabs f13, f29
  651. fabs f14, f30
  652. fabs f15, f31
  653. LFDX f28, XX, INCXM1
  654. LFDUX f29, XX, INCX
  655. LFDX f30, XX, INCXM1
  656. LFDUX f31, XX, INCX
  657. fadd f4, f8, f9
  658. fadd f5, f10, f11
  659. fadd f6, f12, f13
  660. fadd f7, f14, f15
  661. addi RET, RET, 1
  662. fcmpu cr0, f1, f4
  663. beq cr0, LL(9999)
  664. addi RET, RET, 1
  665. fcmpu cr0, f1, f5
  666. beq cr0, LL(9999)
  667. addi RET, RET, 1
  668. fcmpu cr0, f1, f6
  669. beq cr0, LL(9999)
  670. addi RET, RET, 1
  671. fcmpu cr0, f1, f7
  672. beq cr0, LL(9999)
  673. fabs f8, f24
  674. fabs f9, f25
  675. fabs f10, f26
  676. fabs f11, f27
  677. fabs f12, f28
  678. fabs f13, f29
  679. fabs f14, f30
  680. fabs f15, f31
  681. fadd f4, f8, f9
  682. fadd f5, f10, f11
  683. fadd f6, f12, f13
  684. fadd f7, f14, f15
  685. addi RET, RET, 1
  686. fcmpu cr0, f1, f4
  687. beq cr0, LL(9999)
  688. addi RET, RET, 1
  689. fcmpu cr0, f1, f5
  690. beq cr0, LL(9999)
  691. addi RET, RET, 1
  692. fcmpu cr0, f1, f6
  693. beq cr0, LL(9999)
  694. addi RET, RET, 1
  695. fcmpu cr0, f1, f7
  696. beq cr0, LL(9999)
  697. .align 4
  698. LL(1150):
  699. andi. r0, NN, 7
  700. mtspr CTR, r0
  701. beq LL(9999)
  702. .align 4
  703. LL(1160):
  704. LFDX f8, XX, INCXM1
  705. LFDUX f9, XX, INCX
  706. fabs f8, f8
  707. fabs f9, f9
  708. fadd f8, f8, f9
  709. addi RET, RET, 1
  710. fcmpu cr0, f1, f8
  711. beq cr0, LL(9999)
  712. bdnz LL(1160)
  713. .align 4
  714. LL(9999):
  715. lfd f14, 0(SP)
  716. lfd f15, 8(SP)
  717. lfd f16, 16(SP)
  718. lfd f17, 24(SP)
  719. lfd f18, 32(SP)
  720. lfd f19, 40(SP)
  721. lfd f20, 48(SP)
  722. lfd f21, 56(SP)
  723. lfd f22, 64(SP)
  724. lfd f23, 72(SP)
  725. lfd f24, 80(SP)
  726. lfd f25, 88(SP)
  727. lfd f26, 96(SP)
  728. lfd f27, 104(SP)
  729. lfd f28, 112(SP)
  730. lfd f29, 120(SP)
  731. lfd f30, 128(SP)
  732. lfd f31, 136(SP)
  733. addi SP, SP, STACKSIZE
  734. blr
  735. EPILOGUE