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.

min.S 9.4 kB

123456789101112131415161718192021222324252627282930313233343536373839404142434445464748495051525354555657585960616263646566676869707172737475767778798081828384858687888990919293949596979899100101102103104105106107108109110111112113114115116117118119120121122123124125126127128129130131132133134135136137138139140141142143144145146147148149150151152153154155156157158159160161162163164165166167168169170171172173174175176177178179180181182183184185186187188189190191192193194195196197198199200201202203204205206207208209210211212213214215216217218219220221222223224225226227228229230231232233234235236237238239240241242243244245246247248249250251252253254255256257258259260261262263264265266267268269270271272273274275276277278279280281282283284285286287288289290291292293294295296297298299300301302303304305306307308309310311312313314315316317318319320321322323324325326327328329330331332333334335336337338339340341342343344345346347348349350351352353354355356357358359360361362363364365366367368369370371372373374375376377378379380381382383384385386387388389390391392393394395396397398399400401402403404405406407408409410411412413414415416417418419420421422423424425426427428429430431432433434435436437438439440441442443444445
  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 N r3
  41. #define X r4
  42. #define INCX r5
  43. #define PREA r8
  44. #define FZERO f1
  45. #define STACKSIZE 160
  46. PROLOGUE
  47. PROFCODE
  48. addi SP, SP, -STACKSIZE
  49. li r0, 0
  50. stfd f14, 0(SP)
  51. stfd f15, 8(SP)
  52. stfd f16, 16(SP)
  53. stfd f17, 24(SP)
  54. stfd f18, 32(SP)
  55. stfd f19, 40(SP)
  56. stfd f20, 48(SP)
  57. stfd f21, 56(SP)
  58. stfd f22, 64(SP)
  59. stfd f23, 72(SP)
  60. stfd f24, 80(SP)
  61. stfd f25, 88(SP)
  62. stfd f26, 96(SP)
  63. stfd f27, 104(SP)
  64. stfd f28, 112(SP)
  65. stfd f29, 120(SP)
  66. stfd f30, 128(SP)
  67. stfd f31, 136(SP)
  68. stw r0, 144(SP)
  69. lfs FZERO,144(SP)
  70. #ifdef F_INTERFACE
  71. LDINT N, 0(N)
  72. LDINT INCX, 0(INCX)
  73. #endif
  74. slwi INCX, INCX, BASE_SHIFT
  75. li PREA, L1_PREFETCHSIZE
  76. cmpwi cr0, N, 0
  77. ble- LL(9999)
  78. cmpwi cr0, INCX, 0
  79. ble- LL(9999)
  80. LFD f1, 0 * SIZE(X)
  81. add X, X, INCX
  82. fmr f0, f1
  83. fmr f2, f1
  84. fmr f3, f1
  85. fmr f4, f1
  86. fmr f5, f1
  87. fmr f6, f1
  88. fmr f7, f1
  89. subi N, N, 1
  90. cmpwi cr0, INCX, SIZE
  91. bne- cr0, LL(100)
  92. srawi. r0, N, 4
  93. mtspr CTR, r0
  94. beq- cr0, LL(50)
  95. LFD f16, 0 * SIZE(X)
  96. LFD f17, 1 * SIZE(X)
  97. LFD f18, 2 * SIZE(X)
  98. LFD f19, 3 * SIZE(X)
  99. LFD f20, 4 * SIZE(X)
  100. LFD f21, 5 * SIZE(X)
  101. LFD f22, 6 * SIZE(X)
  102. LFD f23, 7 * SIZE(X)
  103. LFD f24, 8 * SIZE(X)
  104. LFD f25, 9 * SIZE(X)
  105. LFD f26, 10 * SIZE(X)
  106. LFD f27, 11 * SIZE(X)
  107. LFD f28, 12 * SIZE(X)
  108. LFD f29, 13 * SIZE(X)
  109. LFD f30, 14 * SIZE(X)
  110. LFD f31, 15 * SIZE(X)
  111. fsub f8, f0, f16
  112. fsub f9, f1, f17
  113. fsub f10, f2, f18
  114. fsub f11, f3, f19
  115. fsub f12, f4, f20
  116. fsub f13, f5, f21
  117. fsub f14, f6, f22
  118. fsub f15, f7, f23
  119. bdz LL(20)
  120. .align 4
  121. LL(10):
  122. fsel f0, f8, f16, f0
  123. fsub f8, f0, f24
  124. fsel f1, f9, f17, f1
  125. fsub f9, f1, f25
  126. fsel f2, f10, f18, f2
  127. fsub f10, f2, f26
  128. fsel f3, f11, f19, f3
  129. fsub f11, f3, f27
  130. LFD f16, 16 * SIZE(X)
  131. LFD f17, 17 * SIZE(X)
  132. LFD f18, 18 * SIZE(X)
  133. LFD f19, 19 * SIZE(X)
  134. fsel f4, f12, f20, f4
  135. fsub f12, f4, f28
  136. fsel f5, f13, f21, f5
  137. fsub f13, f5, f29
  138. fsel f6, f14, f22, f6
  139. fsub f14, f6, f30
  140. fsel f7, f15, f23, f7
  141. fsub f15, f7, f31
  142. LFD f20, 20 * SIZE(X)
  143. LFD f21, 21 * SIZE(X)
  144. LFD f22, 22 * SIZE(X)
  145. LFD f23, 23 * SIZE(X)
  146. fsel f0, f8, f24, f0
  147. fsub f8, f0, f16
  148. fsel f1, f9, f25, f1
  149. fsub f9, f1, f17
  150. fsel f2, f10, f26, f2
  151. fsub f10, f2, f18
  152. fsel f3, f11, f27, f3
  153. fsub f11, f3, f19
  154. LFD f24, 24 * SIZE(X)
  155. LFD f25, 25 * SIZE(X)
  156. LFD f26, 26 * SIZE(X)
  157. LFD f27, 27 * SIZE(X)
  158. fsel f4, f12, f28, f4
  159. fsub f12, f4, f20
  160. fsel f5, f13, f29, f5
  161. fsub f13, f5, f21
  162. fsel f6, f14, f30, f6
  163. fsub f14, f6, f22
  164. fsel f7, f15, f31, f7
  165. fsub f15, f7, f23
  166. LFD f28, 28 * SIZE(X)
  167. LFD f29, 29 * SIZE(X)
  168. LFD f30, 30 * SIZE(X)
  169. LFD f31, 31 * SIZE(X)
  170. #ifndef POWER6
  171. L1_PREFETCH X, PREA
  172. #endif
  173. addi X, X, 16 * SIZE
  174. #ifdef POWER6
  175. L1_PREFETCH X, PREA
  176. #endif
  177. bdnz LL(10)
  178. .align 4
  179. LL(20):
  180. fsel f0, f8, f16, f0
  181. fsub f8, f0, f24
  182. fsel f1, f9, f17, f1
  183. fsub f9, f1, f25
  184. fsel f2, f10, f18, f2
  185. fsub f10, f2, f26
  186. fsel f3, f11, f19, f3
  187. fsub f11, f3, f27
  188. fsel f4, f12, f20, f4
  189. fsub f12, f4, f28
  190. fsel f5, f13, f21, f5
  191. fsub f13, f5, f29
  192. fsel f6, f14, f22, f6
  193. fsub f14, f6, f30
  194. fsel f7, f15, f23, f7
  195. fsub f15, f7, f31
  196. fsel f0, f8, f24, f0
  197. fsel f1, f9, f25, f1
  198. fsel f2, f10, f26, f2
  199. fsel f3, f11, f27, f3
  200. fsel f4, f12, f28, f4
  201. fsel f5, f13, f29, f5
  202. fsel f6, f14, f30, f6
  203. fsel f7, f15, f31, f7
  204. addi X, X, 16 * SIZE
  205. .align 4
  206. LL(50):
  207. andi. r0, N, 15
  208. mtspr CTR, r0
  209. beq LL(999)
  210. .align 4
  211. LL(60):
  212. LFD f8, 0 * SIZE(X)
  213. addi X, X, 1 * SIZE
  214. fsub f16, f1, f8
  215. fsel f1, f16, f8, f1
  216. bdnz LL(60)
  217. b LL(999)
  218. .align 4
  219. LL(100):
  220. sub X, X, INCX
  221. srawi. r0, N, 4
  222. mtspr CTR, r0
  223. beq- LL(150)
  224. LFDUX f16, X, INCX
  225. LFDUX f17, X, INCX
  226. LFDUX f18, X, INCX
  227. LFDUX f19, X, INCX
  228. LFDUX f20, X, INCX
  229. LFDUX f21, X, INCX
  230. LFDUX f22, X, INCX
  231. LFDUX f23, X, INCX
  232. LFDUX f24, X, INCX
  233. LFDUX f25, X, INCX
  234. LFDUX f26, X, INCX
  235. LFDUX f27, X, INCX
  236. LFDUX f28, X, INCX
  237. LFDUX f29, X, INCX
  238. LFDUX f30, X, INCX
  239. LFDUX f31, X, INCX
  240. fsub f8, f0, f16
  241. fsub f9, f1, f17
  242. fsub f10, f2, f18
  243. fsub f11, f3, f19
  244. fsub f12, f4, f20
  245. fsub f13, f5, f21
  246. fsub f14, f6, f22
  247. fsub f15, f7, f23
  248. bdz LL(120)
  249. .align 4
  250. LL(110):
  251. fsel f0, f8, f16, f0
  252. fsub f8, f0, f24
  253. fsel f1, f9, f17, f1
  254. fsub f9, f1, f25
  255. fsel f2, f10, f18, f2
  256. fsub f10, f2, f26
  257. fsel f3, f11, f19, f3
  258. fsub f11, f3, f27
  259. LFDUX f16, X, INCX
  260. LFDUX f17, X, INCX
  261. LFDUX f18, X, INCX
  262. LFDUX f19, X, INCX
  263. fsel f4, f12, f20, f4
  264. fsub f12, f4, f28
  265. fsel f5, f13, f21, f5
  266. fsub f13, f5, f29
  267. fsel f6, f14, f22, f6
  268. fsub f14, f6, f30
  269. fsel f7, f15, f23, f7
  270. fsub f15, f7, f31
  271. LFDUX f20, X, INCX
  272. LFDUX f21, X, INCX
  273. LFDUX f22, X, INCX
  274. LFDUX f23, X, INCX
  275. fsel f0, f8, f24, f0
  276. fsub f8, f0, f16
  277. fsel f1, f9, f25, f1
  278. fsub f9, f1, f17
  279. fsel f2, f10, f26, f2
  280. fsub f10, f2, f18
  281. fsel f3, f11, f27, f3
  282. fsub f11, f3, f19
  283. LFDUX f24, X, INCX
  284. LFDUX f25, X, INCX
  285. LFDUX f26, X, INCX
  286. LFDUX f27, X, INCX
  287. fsel f4, f12, f28, f4
  288. fsub f12, f4, f20
  289. fsel f5, f13, f29, f5
  290. fsub f13, f5, f21
  291. fsel f6, f14, f30, f6
  292. fsub f14, f6, f22
  293. fsel f7, f15, f31, f7
  294. fsub f15, f7, f23
  295. LFDUX f28, X, INCX
  296. LFDUX f29, X, INCX
  297. LFDUX f30, X, INCX
  298. LFDUX f31, X, INCX
  299. bdnz LL(110)
  300. .align 4
  301. LL(120):
  302. fsel f0, f8, f16, f0
  303. fsub f8, f0, f24
  304. fsel f1, f9, f17, f1
  305. fsub f9, f1, f25
  306. fsel f2, f10, f18, f2
  307. fsub f10, f2, f26
  308. fsel f3, f11, f19, f3
  309. fsub f11, f3, f27
  310. fsel f4, f12, f20, f4
  311. fsub f12, f4, f28
  312. fsel f5, f13, f21, f5
  313. fsub f13, f5, f29
  314. fsel f6, f14, f22, f6
  315. fsub f14, f6, f30
  316. fsel f7, f15, f23, f7
  317. fsub f15, f7, f31
  318. fsel f0, f8, f24, f0
  319. fsel f1, f9, f25, f1
  320. fsel f2, f10, f26, f2
  321. fsel f3, f11, f27, f3
  322. fsel f4, f12, f28, f4
  323. fsel f5, f13, f29, f5
  324. fsel f6, f14, f30, f6
  325. fsel f7, f15, f31, f7
  326. .align 4
  327. LL(150):
  328. andi. r0, N, 15
  329. mtspr CTR, r0
  330. beq LL(999)
  331. .align 4
  332. LL(160):
  333. LFDUX f8, X, INCX
  334. fsub f16, f1, f8
  335. fsel f1, f16, f8, f1
  336. bdnz LL(160)
  337. .align 4
  338. LL(999):
  339. fsub f8, f0, f1
  340. fsub f9, f2, f3
  341. fsub f10, f4, f5
  342. fsub f11, f6, f7
  343. fsel f0, f8, f1, f0
  344. fsel f2, f9, f3, f2
  345. fsel f4, f10, f5, f4
  346. fsel f6, f11, f7, f6
  347. fsub f8, f0, f2
  348. fsub f9, f4, f6
  349. fsel f0, f8, f2, f0
  350. fsel f4, f9, f6, f4
  351. fsub f8, f0, f4
  352. fsel f1, f8, f4, f0
  353. .align 4
  354. LL(9999):
  355. lfd f14, 0(SP)
  356. lfd f15, 8(SP)
  357. lfd f16, 16(SP)
  358. lfd f17, 24(SP)
  359. lfd f18, 32(SP)
  360. lfd f19, 40(SP)
  361. lfd f20, 48(SP)
  362. lfd f21, 56(SP)
  363. lfd f22, 64(SP)
  364. lfd f23, 72(SP)
  365. lfd f24, 80(SP)
  366. lfd f25, 88(SP)
  367. lfd f26, 96(SP)
  368. lfd f27, 104(SP)
  369. lfd f28, 112(SP)
  370. lfd f29, 120(SP)
  371. lfd f30, 128(SP)
  372. lfd f31, 136(SP)
  373. addi SP, SP, STACKSIZE
  374. blr
  375. EPILOGUE